drivers/gpu/drm/radeon/cik.c

   1 /*
   2  * Copyright 2012 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  * Authors: Alex Deucher
  23  */
  24 #include <linux/firmware.h>
  25 #include <linux/slab.h>
  26 #include <linux/module.h>
  27 #include "drmP.h"
  28 #include "radeon.h"
  29 #include "radeon_asic.h"
  30 #include "cikd.h"
  31 #include "atom.h"
  32 #include "cik_blit_shaders.h"
  33 #include "radeon_ucode.h"
  34 #include "clearstate_ci.h"
  35
  36 MODULE_FIRMWARE("radeon/BONAIRE_pfp.bin");
  37 MODULE_FIRMWARE("radeon/BONAIRE_me.bin");
  38 MODULE_FIRMWARE("radeon/BONAIRE_ce.bin");
  39 MODULE_FIRMWARE("radeon/BONAIRE_mec.bin");
  40 MODULE_FIRMWARE("radeon/BONAIRE_mc.bin");
  41 MODULE_FIRMWARE("radeon/BONAIRE_mc2.bin");
  42 MODULE_FIRMWARE("radeon/BONAIRE_rlc.bin");
  43 MODULE_FIRMWARE("radeon/BONAIRE_sdma.bin");
  44 MODULE_FIRMWARE("radeon/BONAIRE_smc.bin");
  45
  46 MODULE_FIRMWARE("radeon/bonaire_pfp.bin");
  47 MODULE_FIRMWARE("radeon/bonaire_me.bin");
  48 MODULE_FIRMWARE("radeon/bonaire_ce.bin");
  49 MODULE_FIRMWARE("radeon/bonaire_mec.bin");
  50 MODULE_FIRMWARE("radeon/bonaire_mc.bin");
  51 MODULE_FIRMWARE("radeon/bonaire_rlc.bin");
  52 MODULE_FIRMWARE("radeon/bonaire_sdma.bin");
  53 MODULE_FIRMWARE("radeon/bonaire_smc.bin");
  54
  55 MODULE_FIRMWARE("radeon/HAWAII_pfp.bin");
  56 MODULE_FIRMWARE("radeon/HAWAII_me.bin");
  57 MODULE_FIRMWARE("radeon/HAWAII_ce.bin");
  58 MODULE_FIRMWARE("radeon/HAWAII_mec.bin");
  59 MODULE_FIRMWARE("radeon/HAWAII_mc.bin");
  60 MODULE_FIRMWARE("radeon/HAWAII_mc2.bin");
  61 MODULE_FIRMWARE("radeon/HAWAII_rlc.bin");
  62 MODULE_FIRMWARE("radeon/HAWAII_sdma.bin");
  63 MODULE_FIRMWARE("radeon/HAWAII_smc.bin");
  64
  65 MODULE_FIRMWARE("radeon/hawaii_pfp.bin");
  66 MODULE_FIRMWARE("radeon/hawaii_me.bin");
  67 MODULE_FIRMWARE("radeon/hawaii_ce.bin");
  68 MODULE_FIRMWARE("radeon/hawaii_mec.bin");
  69 MODULE_FIRMWARE("radeon/hawaii_mc.bin");
  70 MODULE_FIRMWARE("radeon/hawaii_rlc.bin");
  71 MODULE_FIRMWARE("radeon/hawaii_sdma.bin");
  72 MODULE_FIRMWARE("radeon/hawaii_smc.bin");
  73
  74 MODULE_FIRMWARE("radeon/KAVERI_pfp.bin");
  75 MODULE_FIRMWARE("radeon/KAVERI_me.bin");
  76 MODULE_FIRMWARE("radeon/KAVERI_ce.bin");
  77 MODULE_FIRMWARE("radeon/KAVERI_mec.bin");
  78 MODULE_FIRMWARE("radeon/KAVERI_rlc.bin");
  79 MODULE_FIRMWARE("radeon/KAVERI_sdma.bin");
  80
  81 MODULE_FIRMWARE("radeon/kaveri_pfp.bin");
  82 MODULE_FIRMWARE("radeon/kaveri_me.bin");
  83 MODULE_FIRMWARE("radeon/kaveri_ce.bin");
  84 MODULE_FIRMWARE("radeon/kaveri_mec.bin");
  85 MODULE_FIRMWARE("radeon/kaveri_mec2.bin");
  86 MODULE_FIRMWARE("radeon/kaveri_rlc.bin");
  87 MODULE_FIRMWARE("radeon/kaveri_sdma.bin");
  88
  89 MODULE_FIRMWARE("radeon/KABINI_pfp.bin");
  90 MODULE_FIRMWARE("radeon/KABINI_me.bin");
  91 MODULE_FIRMWARE("radeon/KABINI_ce.bin");
  92 MODULE_FIRMWARE("radeon/KABINI_mec.bin");
  93 MODULE_FIRMWARE("radeon/KABINI_rlc.bin");
  94 MODULE_FIRMWARE("radeon/KABINI_sdma.bin");
  95
  96 MODULE_FIRMWARE("radeon/kabini_pfp.bin");
  97 MODULE_FIRMWARE("radeon/kabini_me.bin");
  98 MODULE_FIRMWARE("radeon/kabini_ce.bin");
  99 MODULE_FIRMWARE("radeon/kabini_mec.bin");
 100 MODULE_FIRMWARE("radeon/kabini_rlc.bin");
 101 MODULE_FIRMWARE("radeon/kabini_sdma.bin");
 102
 103 MODULE_FIRMWARE("radeon/MULLINS_pfp.bin");
 104 MODULE_FIRMWARE("radeon/MULLINS_me.bin");
 105 MODULE_FIRMWARE("radeon/MULLINS_ce.bin");
 106 MODULE_FIRMWARE("radeon/MULLINS_mec.bin");
 107 MODULE_FIRMWARE("radeon/MULLINS_rlc.bin");
 108 MODULE_FIRMWARE("radeon/MULLINS_sdma.bin");
 109
 110 MODULE_FIRMWARE("radeon/mullins_pfp.bin");
 111 MODULE_FIRMWARE("radeon/mullins_me.bin");
 112 MODULE_FIRMWARE("radeon/mullins_ce.bin");
 113 MODULE_FIRMWARE("radeon/mullins_mec.bin");
 114 MODULE_FIRMWARE("radeon/mullins_rlc.bin");
 115 MODULE_FIRMWARE("radeon/mullins_sdma.bin");
 116
 117 extern int r600_ih_ring_alloc(struct radeon_device *rdev);
 118 extern void r600_ih_ring_fini(struct radeon_device *rdev);
 119 extern void evergreen_mc_stop(struct radeon_device *rdev, struct evergreen_mc_save *save);
 120 extern void evergreen_mc_resume(struct radeon_device *rdev, struct evergreen_mc_save *save);
 121 extern bool evergreen_is_display_hung(struct radeon_device *rdev);
 122 extern void sumo_rlc_fini(struct radeon_device *rdev);
 123 extern int sumo_rlc_init(struct radeon_device *rdev);
 124 extern void si_vram_gtt_location(struct radeon_device *rdev, struct radeon_mc *mc);
 125 extern void si_rlc_reset(struct radeon_device *rdev);
 126 extern void si_init_uvd_internal_cg(struct radeon_device *rdev);
 127 static u32 cik_get_cu_active_bitmap(struct radeon_device *rdev, u32 se, u32 sh);
 128 extern int cik_sdma_resume(struct radeon_device *rdev);
 129 extern void cik_sdma_enable(struct radeon_device *rdev, bool enable);
 130 extern void cik_sdma_fini(struct radeon_device *rdev);
 131 extern void vce_v2_0_enable_mgcg(struct radeon_device *rdev, bool enable);
 132 static void cik_rlc_stop(struct radeon_device *rdev);
 133 static void cik_pcie_gen3_enable(struct radeon_device *rdev);
 134 static void cik_program_aspm(struct radeon_device *rdev);
 135 static void cik_init_pg(struct radeon_device *rdev);
 136 static void cik_init_cg(struct radeon_device *rdev);
 137 static void cik_fini_pg(struct radeon_device *rdev);
 138 static void cik_fini_cg(struct radeon_device *rdev);
 139 static void cik_enable_gui_idle_interrupt(struct radeon_device *rdev,
 140                                           bool enable);
 141
 142 /* get temperature in millidegrees */
 143 int ci_get_temp(struct radeon_device *rdev)
 144 {
 145         u32 temp;
 146         int actual_temp = 0;
 147
 148         temp = (RREG32_SMC(CG_MULT_THERMAL_STATUS) & CTF_TEMP_MASK) >>
 149                 CTF_TEMP_SHIFT;
 150
 151         if (temp & 0x200)
 152                 actual_temp = 255;
 153         else
 154                 actual_temp = temp & 0x1ff;
 155
 156         actual_temp = actual_temp * 1000;
 157
 158         return actual_temp;
 159 }
 160
 161 /* get temperature in millidegrees */
 162 int kv_get_temp(struct radeon_device *rdev)
 163 {
 164         u32 temp;
 165         int actual_temp = 0;
 166
 167         temp = RREG32_SMC(0xC0300E0C);
 168
 169         if (temp)
 170                 actual_temp = (temp / 8) - 49;
 171         else
 172                 actual_temp = 0;
 173
 174         actual_temp = actual_temp * 1000;
 175
 176         return actual_temp;
 177 }
 178
 179 /*
 180  * Indirect registers accessor
 181  */
 182 u32 cik_pciep_rreg(struct radeon_device *rdev, u32 reg)
 183 {
 184         unsigned long flags;
 185         u32 r;
 186
 187         spin_lock_irqsave(&rdev->pciep_idx_lock, flags);
 188         WREG32(PCIE_INDEX, reg);
 189         (void)RREG32(PCIE_INDEX);
 190         r = RREG32(PCIE_DATA);
 191         spin_unlock_irqrestore(&rdev->pciep_idx_lock, flags);
 192         return r;
 193 }
 194
 195 void cik_pciep_wreg(struct radeon_device *rdev, u32 reg, u32 v)
 196 {
 197         unsigned long flags;
 198
 199         spin_lock_irqsave(&rdev->pciep_idx_lock, flags);
 200         WREG32(PCIE_INDEX, reg);
 201         (void)RREG32(PCIE_INDEX);
 202         WREG32(PCIE_DATA, v);
 203         (void)RREG32(PCIE_DATA);
 204         spin_unlock_irqrestore(&rdev->pciep_idx_lock, flags);
 205 }
 206
 207 static const u32 spectre_rlc_save_restore_register_list[] =
 208 {
 209         (0x0e00 << 16) | (0xc12c >> 2),
 210         0x00000000,
 211         (0x0e00 << 16) | (0xc140 >> 2),
 212         0x00000000,
 213         (0x0e00 << 16) | (0xc150 >> 2),
 214         0x00000000,
 215         (0x0e00 << 16) | (0xc15c >> 2),
 216         0x00000000,
 217         (0x0e00 << 16) | (0xc168 >> 2),
 218         0x00000000,
 219         (0x0e00 << 16) | (0xc170 >> 2),
 220         0x00000000,
 221         (0x0e00 << 16) | (0xc178 >> 2),
 222         0x00000000,
 223         (0x0e00 << 16) | (0xc204 >> 2),
 224         0x00000000,
 225         (0x0e00 << 16) | (0xc2b4 >> 2),
 226         0x00000000,
 227         (0x0e00 << 16) | (0xc2b8 >> 2),
 228         0x00000000,
 229         (0x0e00 << 16) | (0xc2bc >> 2),
 230         0x00000000,
 231         (0x0e00 << 16) | (0xc2c0 >> 2),
 232         0x00000000,
 233         (0x0e00 << 16) | (0x8228 >> 2),
 234         0x00000000,
 235         (0x0e00 << 16) | (0x829c >> 2),
 236         0x00000000,
 237         (0x0e00 << 16) | (0x869c >> 2),
 238         0x00000000,
 239         (0x0600 << 16) | (0x98f4 >> 2),
 240         0x00000000,
 241         (0x0e00 << 16) | (0x98f8 >> 2),
 242         0x00000000,
 243         (0x0e00 << 16) | (0x9900 >> 2),
 244         0x00000000,
 245         (0x0e00 << 16) | (0xc260 >> 2),
 246         0x00000000,
 247         (0x0e00 << 16) | (0x90e8 >> 2),
 248         0x00000000,
 249         (0x0e00 << 16) | (0x3c000 >> 2),
 250         0x00000000,
 251         (0x0e00 << 16) | (0x3c00c >> 2),
 252         0x00000000,
 253         (0x0e00 << 16) | (0x8c1c >> 2),
 254         0x00000000,
 255         (0x0e00 << 16) | (0x9700 >> 2),
 256         0x00000000,
 257         (0x0e00 << 16) | (0xcd20 >> 2),
 258         0x00000000,
 259         (0x4e00 << 16) | (0xcd20 >> 2),
 260         0x00000000,
 261         (0x5e00 << 16) | (0xcd20 >> 2),
 262         0x00000000,
 263         (0x6e00 << 16) | (0xcd20 >> 2),
 264         0x00000000,
 265         (0x7e00 << 16) | (0xcd20 >> 2),
 266         0x00000000,
 267         (0x8e00 << 16) | (0xcd20 >> 2),
 268         0x00000000,
 269         (0x9e00 << 16) | (0xcd20 >> 2),
 270         0x00000000,
 271         (0xae00 << 16) | (0xcd20 >> 2),
 272         0x00000000,
 273         (0xbe00 << 16) | (0xcd20 >> 2),
 274         0x00000000,
 275         (0x0e00 << 16) | (0x89bc >> 2),
 276         0x00000000,
 277         (0x0e00 << 16) | (0x8900 >> 2),
 278         0x00000000,
 279         0x3,
 280         (0x0e00 << 16) | (0xc130 >> 2),
 281         0x00000000,
 282         (0x0e00 << 16) | (0xc134 >> 2),
 283         0x00000000,
 284         (0x0e00 << 16) | (0xc1fc >> 2),
 285         0x00000000,
 286         (0x0e00 << 16) | (0xc208 >> 2),
 287         0x00000000,
 288         (0x0e00 << 16) | (0xc264 >> 2),
 289         0x00000000,
 290         (0x0e00 << 16) | (0xc268 >> 2),
 291         0x00000000,
 292         (0x0e00 << 16) | (0xc26c >> 2),
 293         0x00000000,
 294         (0x0e00 << 16) | (0xc270 >> 2),
 295         0x00000000,
 296         (0x0e00 << 16) | (0xc274 >> 2),
 297         0x00000000,
 298         (0x0e00 << 16) | (0xc278 >> 2),
 299         0x00000000,
 300         (0x0e00 << 16) | (0xc27c >> 2),
 301         0x00000000,
 302         (0x0e00 << 16) | (0xc280 >> 2),
 303         0x00000000,
 304         (0x0e00 << 16) | (0xc284 >> 2),
 305         0x00000000,
 306         (0x0e00 << 16) | (0xc288 >> 2),
 307         0x00000000,
 308         (0x0e00 << 16) | (0xc28c >> 2),
 309         0x00000000,
 310         (0x0e00 << 16) | (0xc290 >> 2),
 311         0x00000000,
 312         (0x0e00 << 16) | (0xc294 >> 2),
 313         0x00000000,
 314         (0x0e00 << 16) | (0xc298 >> 2),
 315         0x00000000,
 316         (0x0e00 << 16) | (0xc29c >> 2),
 317         0x00000000,
 318         (0x0e00 << 16) | (0xc2a0 >> 2),
 319         0x00000000,
 320         (0x0e00 << 16) | (0xc2a4 >> 2),
 321         0x00000000,
 322         (0x0e00 << 16) | (0xc2a8 >> 2),
 323         0x00000000,
 324         (0x0e00 << 16) | (0xc2ac  >> 2),
 325         0x00000000,
 326         (0x0e00 << 16) | (0xc2b0 >> 2),
 327         0x00000000,
 328         (0x0e00 << 16) | (0x301d0 >> 2),
 329         0x00000000,
 330         (0x0e00 << 16) | (0x30238 >> 2),
 331         0x00000000,
 332         (0x0e00 << 16) | (0x30250 >> 2),
 333         0x00000000,
 334         (0x0e00 << 16) | (0x30254 >> 2),
 335         0x00000000,
 336         (0x0e00 << 16) | (0x30258 >> 2),
 337         0x00000000,
 338         (0x0e00 << 16) | (0x3025c >> 2),
 339         0x00000000,
 340         (0x4e00 << 16) | (0xc900 >> 2),
 341         0x00000000,
 342         (0x5e00 << 16) | (0xc900 >> 2),
 343         0x00000000,
 344         (0x6e00 << 16) | (0xc900 >> 2),
 345         0x00000000,
 346         (0x7e00 << 16) | (0xc900 >> 2),
 347         0x00000000,
 348         (0x8e00 << 16) | (0xc900 >> 2),
 349         0x00000000,
 350         (0x9e00 << 16) | (0xc900 >> 2),
 351         0x00000000,
 352         (0xae00 << 16) | (0xc900 >> 2),
 353         0x00000000,
 354         (0xbe00 << 16) | (0xc900 >> 2),
 355         0x00000000,
 356         (0x4e00 << 16) | (0xc904 >> 2),
 357         0x00000000,
 358         (0x5e00 << 16) | (0xc904 >> 2),
 359         0x00000000,
 360         (0x6e00 << 16) | (0xc904 >> 2),
 361         0x00000000,
 362         (0x7e00 << 16) | (0xc904 >> 2),
 363         0x00000000,
 364         (0x8e00 << 16) | (0xc904 >> 2),
 365         0x00000000,
 366         (0x9e00 << 16) | (0xc904 >> 2),
 367         0x00000000,
 368         (0xae00 << 16) | (0xc904 >> 2),
 369         0x00000000,
 370         (0xbe00 << 16) | (0xc904 >> 2),
 371         0x00000000,
 372         (0x4e00 << 16) | (0xc908 >> 2),
 373         0x00000000,
 374         (0x5e00 << 16) | (0xc908 >> 2),
 375         0x00000000,
 376         (0x6e00 << 16) | (0xc908 >> 2),
 377         0x00000000,
 378         (0x7e00 << 16) | (0xc908 >> 2),
 379         0x00000000,
 380         (0x8e00 << 16) | (0xc908 >> 2),
 381         0x00000000,
 382         (0x9e00 << 16) | (0xc908 >> 2),
 383         0x00000000,
 384         (0xae00 << 16) | (0xc908 >> 2),
 385         0x00000000,
 386         (0xbe00 << 16) | (0xc908 >> 2),
 387         0x00000000,
 388         (0x4e00 << 16) | (0xc90c >> 2),
 389         0x00000000,
 390         (0x5e00 << 16) | (0xc90c >> 2),
 391         0x00000000,
 392         (0x6e00 << 16) | (0xc90c >> 2),
 393         0x00000000,
 394         (0x7e00 << 16) | (0xc90c >> 2),
 395         0x00000000,
 396         (0x8e00 << 16) | (0xc90c >> 2),
 397         0x00000000,
 398         (0x9e00 << 16) | (0xc90c >> 2),
 399         0x00000000,
 400         (0xae00 << 16) | (0xc90c >> 2),
 401         0x00000000,
 402         (0xbe00 << 16) | (0xc90c >> 2),
 403         0x00000000,
 404         (0x4e00 << 16) | (0xc910 >> 2),
 405         0x00000000,
 406         (0x5e00 << 16) | (0xc910 >> 2),
 407         0x00000000,
 408         (0x6e00 << 16) | (0xc910 >> 2),
 409         0x00000000,
 410         (0x7e00 << 16) | (0xc910 >> 2),
 411         0x00000000,
 412         (0x8e00 << 16) | (0xc910 >> 2),
 413         0x00000000,
 414         (0x9e00 << 16) | (0xc910 >> 2),
 415         0x00000000,
 416         (0xae00 << 16) | (0xc910 >> 2),
 417         0x00000000,
 418         (0xbe00 << 16) | (0xc910 >> 2),
 419         0x00000000,
 420         (0x0e00 << 16) | (0xc99c >> 2),
 421         0x00000000,
 422         (0x0e00 << 16) | (0x9834 >> 2),
 423         0x00000000,
 424         (0x0000 << 16) | (0x30f00 >> 2),
 425         0x00000000,
 426         (0x0001 << 16) | (0x30f00 >> 2),
 427         0x00000000,
 428         (0x0000 << 16) | (0x30f04 >> 2),
 429         0x00000000,
 430         (0x0001 << 16) | (0x30f04 >> 2),
 431         0x00000000,
 432         (0x0000 << 16) | (0x30f08 >> 2),
 433         0x00000000,
 434         (0x0001 << 16) | (0x30f08 >> 2),
 435         0x00000000,
 436         (0x0000 << 16) | (0x30f0c >> 2),
 437         0x00000000,
 438         (0x0001 << 16) | (0x30f0c >> 2),
 439         0x00000000,
 440         (0x0600 << 16) | (0x9b7c >> 2),
 441         0x00000000,
 442         (0x0e00 << 16) | (0x8a14 >> 2),
 443         0x00000000,
 444         (0x0e00 << 16) | (0x8a18 >> 2),
 445         0x00000000,
 446         (0x0600 << 16) | (0x30a00 >> 2),
 447         0x00000000,
 448         (0x0e00 << 16) | (0x8bf0 >> 2),
 449         0x00000000,
 450         (0x0e00 << 16) | (0x8bcc >> 2),
 451         0x00000000,
 452         (0x0e00 << 16) | (0x8b24 >> 2),
 453         0x00000000,
 454         (0x0e00 << 16) | (0x30a04 >> 2),
 455         0x00000000,
 456         (0x0600 << 16) | (0x30a10 >> 2),
 457         0x00000000,
 458         (0x0600 << 16) | (0x30a14 >> 2),
 459         0x00000000,
 460         (0x0600 << 16) | (0x30a18 >> 2),
 461         0x00000000,
 462         (0x0600 << 16) | (0x30a2c >> 2),
 463         0x00000000,
 464         (0x0e00 << 16) | (0xc700 >> 2),
 465         0x00000000,
 466         (0x0e00 << 16) | (0xc704 >> 2),
 467         0x00000000,
 468         (0x0e00 << 16) | (0xc708 >> 2),
 469         0x00000000,
 470         (0x0e00 << 16) | (0xc768 >> 2),
 471         0x00000000,
 472         (0x0400 << 16) | (0xc770 >> 2),
 473         0x00000000,
 474         (0x0400 << 16) | (0xc774 >> 2),
 475         0x00000000,
 476         (0x0400 << 16) | (0xc778 >> 2),
 477         0x00000000,
 478         (0x0400 << 16) | (0xc77c >> 2),
 479         0x00000000,
 480         (0x0400 << 16) | (0xc780 >> 2),
 481         0x00000000,
 482         (0x0400 << 16) | (0xc784 >> 2),
 483         0x00000000,
 484         (0x0400 << 16) | (0xc788 >> 2),
 485         0x00000000,
 486         (0x0400 << 16) | (0xc78c >> 2),
 487         0x00000000,
 488         (0x0400 << 16) | (0xc798 >> 2),
 489         0x00000000,
 490         (0x0400 << 16) | (0xc79c >> 2),
 491         0x00000000,
 492         (0x0400 << 16) | (0xc7a0 >> 2),
 493         0x00000000,
 494         (0x0400 << 16) | (0xc7a4 >> 2),
 495         0x00000000,
 496         (0x0400 << 16) | (0xc7a8 >> 2),
 497         0x00000000,
 498         (0x0400 << 16) | (0xc7ac >> 2),
 499         0x00000000,
 500         (0x0400 << 16) | (0xc7b0 >> 2),
 501         0x00000000,
 502         (0x0400 << 16) | (0xc7b4 >> 2),
 503         0x00000000,
 504         (0x0e00 << 16) | (0x9100 >> 2),
 505         0x00000000,
 506         (0x0e00 << 16) | (0x3c010 >> 2),
 507         0x00000000,
 508         (0x0e00 << 16) | (0x92a8 >> 2),
 509         0x00000000,
 510         (0x0e00 << 16) | (0x92ac >> 2),
 511         0x00000000,
 512         (0x0e00 << 16) | (0x92b4 >> 2),
 513         0x00000000,
 514         (0x0e00 << 16) | (0x92b8 >> 2),
 515         0x00000000,
 516         (0x0e00 << 16) | (0x92bc >> 2),
 517         0x00000000,
 518         (0x0e00 << 16) | (0x92c0 >> 2),
 519         0x00000000,
 520         (0x0e00 << 16) | (0x92c4 >> 2),
 521         0x00000000,
 522         (0x0e00 << 16) | (0x92c8 >> 2),
 523         0x00000000,
 524         (0x0e00 << 16) | (0x92cc >> 2),
 525         0x00000000,
 526         (0x0e00 << 16) | (0x92d0 >> 2),
 527         0x00000000,
 528         (0x0e00 << 16) | (0x8c00 >> 2),
 529         0x00000000,
 530         (0x0e00 << 16) | (0x8c04 >> 2),
 531         0x00000000,
 532         (0x0e00 << 16) | (0x8c20 >> 2),
 533         0x00000000,
 534         (0x0e00 << 16) | (0x8c38 >> 2),
 535         0x00000000,
 536         (0x0e00 << 16) | (0x8c3c >> 2),
 537         0x00000000,
 538         (0x0e00 << 16) | (0xae00 >> 2),
 539         0x00000000,
 540         (0x0e00 << 16) | (0x9604 >> 2),
 541         0x00000000,
 542         (0x0e00 << 16) | (0xac08 >> 2),
 543         0x00000000,
 544         (0x0e00 << 16) | (0xac0c >> 2),
 545         0x00000000,
 546         (0x0e00 << 16) | (0xac10 >> 2),
 547         0x00000000,
 548         (0x0e00 << 16) | (0xac14 >> 2),
 549         0x00000000,
 550         (0x0e00 << 16) | (0xac58 >> 2),
 551         0x00000000,
 552         (0x0e00 << 16) | (0xac68 >> 2),
 553         0x00000000,
 554         (0x0e00 << 16) | (0xac6c >> 2),
 555         0x00000000,
 556         (0x0e00 << 16) | (0xac70 >> 2),
 557         0x00000000,
 558         (0x0e00 << 16) | (0xac74 >> 2),
 559         0x00000000,
 560         (0x0e00 << 16) | (0xac78 >> 2),
 561         0x00000000,
 562         (0x0e00 << 16) | (0xac7c >> 2),
 563         0x00000000,
 564         (0x0e00 << 16) | (0xac80 >> 2),
 565         0x00000000,
 566         (0x0e00 << 16) | (0xac84 >> 2),
 567         0x00000000,
 568         (0x0e00 << 16) | (0xac88 >> 2),
 569         0x00000000,
 570         (0x0e00 << 16) | (0xac8c >> 2),
 571         0x00000000,
 572         (0x0e00 << 16) | (0x970c >> 2),
 573         0x00000000,
 574         (0x0e00 << 16) | (0x9714 >> 2),
 575         0x00000000,
 576         (0x0e00 << 16) | (0x9718 >> 2),
 577         0x00000000,
 578         (0x0e00 << 16) | (0x971c >> 2),
 579         0x00000000,
 580         (0x0e00 << 16) | (0x31068 >> 2),
 581         0x00000000,
 582         (0x4e00 << 16) | (0x31068 >> 2),
 583         0x00000000,
 584         (0x5e00 << 16) | (0x31068 >> 2),
 585         0x00000000,
 586         (0x6e00 << 16) | (0x31068 >> 2),
 587         0x00000000,
 588         (0x7e00 << 16) | (0x31068 >> 2),
 589         0x00000000,
 590         (0x8e00 << 16) | (0x31068 >> 2),
 591         0x00000000,
 592         (0x9e00 << 16) | (0x31068 >> 2),
 593         0x00000000,
 594         (0xae00 << 16) | (0x31068 >> 2),
 595         0x00000000,
 596         (0xbe00 << 16) | (0x31068 >> 2),
 597         0x00000000,
 598         (0x0e00 << 16) | (0xcd10 >> 2),
 599         0x00000000,
 600         (0x0e00 << 16) | (0xcd14 >> 2),
 601         0x00000000,
 602         (0x0e00 << 16) | (0x88b0 >> 2),
 603         0x00000000,
 604         (0x0e00 << 16) | (0x88b4 >> 2),
 605         0x00000000,
 606         (0x0e00 << 16) | (0x88b8 >> 2),
 607         0x00000000,
 608         (0x0e00 << 16) | (0x88bc >> 2),
 609         0x00000000,
 610         (0x0400 << 16) | (0x89c0 >> 2),
 611         0x00000000,
 612         (0x0e00 << 16) | (0x88c4 >> 2),
 613         0x00000000,
 614         (0x0e00 << 16) | (0x88c8 >> 2),
 615         0x00000000,
 616         (0x0e00 << 16) | (0x88d0 >> 2),
 617         0x00000000,
 618         (0x0e00 << 16) | (0x88d4 >> 2),
 619         0x00000000,
 620         (0x0e00 << 16) | (0x88d8 >> 2),
 621         0x00000000,
 622         (0x0e00 << 16) | (0x8980 >> 2),
 623         0x00000000,
 624         (0x0e00 << 16) | (0x30938 >> 2),
 625         0x00000000,
 626         (0x0e00 << 16) | (0x3093c >> 2),
 627         0x00000000,
 628         (0x0e00 << 16) | (0x30940 >> 2),
 629         0x00000000,
 630         (0x0e00 << 16) | (0x89a0 >> 2),
 631         0x00000000,
 632         (0x0e00 << 16) | (0x30900 >> 2),
 633         0x00000000,
 634         (0x0e00 << 16) | (0x30904 >> 2),
 635         0x00000000,
 636         (0x0e00 << 16) | (0x89b4 >> 2),
 637         0x00000000,
 638         (0x0e00 << 16) | (0x3c210 >> 2),
 639         0x00000000,
 640         (0x0e00 << 16) | (0x3c214 >> 2),
 641         0x00000000,
 642         (0x0e00 << 16) | (0x3c218 >> 2),
 643         0x00000000,
 644         (0x0e00 << 16) | (0x8904 >> 2),
 645         0x00000000,
 646         0x5,
 647         (0x0e00 << 16) | (0x8c28 >> 2),
 648         (0x0e00 << 16) | (0x8c2c >> 2),
 649         (0x0e00 << 16) | (0x8c30 >> 2),
 650         (0x0e00 << 16) | (0x8c34 >> 2),
 651         (0x0e00 << 16) | (0x9600 >> 2),
 652 };
 653
 654 static const u32 kalindi_rlc_save_restore_register_list[] =
 655 {
 656         (0x0e00 << 16) | (0xc12c >> 2),
 657         0x00000000,
 658         (0x0e00 << 16) | (0xc140 >> 2),
 659         0x00000000,
 660         (0x0e00 << 16) | (0xc150 >> 2),
 661         0x00000000,
 662         (0x0e00 << 16) | (0xc15c >> 2),
 663         0x00000000,
 664         (0x0e00 << 16) | (0xc168 >> 2),
 665         0x00000000,
 666         (0x0e00 << 16) | (0xc170 >> 2),
 667         0x00000000,
 668         (0x0e00 << 16) | (0xc204 >> 2),
 669         0x00000000,
 670         (0x0e00 << 16) | (0xc2b4 >> 2),
 671         0x00000000,
 672         (0x0e00 << 16) | (0xc2b8 >> 2),
 673         0x00000000,
 674         (0x0e00 << 16) | (0xc2bc >> 2),
 675         0x00000000,
 676         (0x0e00 << 16) | (0xc2c0 >> 2),
 677         0x00000000,
 678         (0x0e00 << 16) | (0x8228 >> 2),
 679         0x00000000,
 680         (0x0e00 << 16) | (0x829c >> 2),
 681         0x00000000,
 682         (0x0e00 << 16) | (0x869c >> 2),
 683         0x00000000,
 684         (0x0600 << 16) | (0x98f4 >> 2),
 685         0x00000000,
 686         (0x0e00 << 16) | (0x98f8 >> 2),
 687         0x00000000,
 688         (0x0e00 << 16) | (0x9900 >> 2),
 689         0x00000000,
 690         (0x0e00 << 16) | (0xc260 >> 2),
 691         0x00000000,
 692         (0x0e00 << 16) | (0x90e8 >> 2),
 693         0x00000000,
 694         (0x0e00 << 16) | (0x3c000 >> 2),
 695         0x00000000,
 696         (0x0e00 << 16) | (0x3c00c >> 2),
 697         0x00000000,
 698         (0x0e00 << 16) | (0x8c1c >> 2),
 699         0x00000000,
 700         (0x0e00 << 16) | (0x9700 >> 2),
 701         0x00000000,
 702         (0x0e00 << 16) | (0xcd20 >> 2),
 703         0x00000000,
 704         (0x4e00 << 16) | (0xcd20 >> 2),
 705         0x00000000,
 706         (0x5e00 << 16) | (0xcd20 >> 2),
 707         0x00000000,
 708         (0x6e00 << 16) | (0xcd20 >> 2),
 709         0x00000000,
 710         (0x7e00 << 16) | (0xcd20 >> 2),
 711         0x00000000,
 712         (0x0e00 << 16) | (0x89bc >> 2),
 713         0x00000000,
 714         (0x0e00 << 16) | (0x8900 >> 2),
 715         0x00000000,
 716         0x3,
 717         (0x0e00 << 16) | (0xc130 >> 2),
 718         0x00000000,
 719         (0x0e00 << 16) | (0xc134 >> 2),
 720         0x00000000,
 721         (0x0e00 << 16) | (0xc1fc >> 2),
 722         0x00000000,
 723         (0x0e00 << 16) | (0xc208 >> 2),
 724         0x00000000,
 725         (0x0e00 << 16) | (0xc264 >> 2),
 726         0x00000000,
 727         (0x0e00 << 16) | (0xc268 >> 2),
 728         0x00000000,
 729         (0x0e00 << 16) | (0xc26c >> 2),
 730         0x00000000,
 731         (0x0e00 << 16) | (0xc270 >> 2),
 732         0x00000000,
 733         (0x0e00 << 16) | (0xc274 >> 2),
 734         0x00000000,
 735         (0x0e00 << 16) | (0xc28c >> 2),
 736         0x00000000,
 737         (0x0e00 << 16) | (0xc290 >> 2),
 738         0x00000000,
 739         (0x0e00 << 16) | (0xc294 >> 2),
 740         0x00000000,
 741         (0x0e00 << 16) | (0xc298 >> 2),
 742         0x00000000,
 743         (0x0e00 << 16) | (0xc2a0 >> 2),
 744         0x00000000,
 745         (0x0e00 << 16) | (0xc2a4 >> 2),
 746         0x00000000,
 747         (0x0e00 << 16) | (0xc2a8 >> 2),
 748         0x00000000,
 749         (0x0e00 << 16) | (0xc2ac >> 2),
 750         0x00000000,
 751         (0x0e00 << 16) | (0x301d0 >> 2),
 752         0x00000000,
 753         (0x0e00 << 16) | (0x30238 >> 2),
 754         0x00000000,
 755         (0x0e00 << 16) | (0x30250 >> 2),
 756         0x00000000,
 757         (0x0e00 << 16) | (0x30254 >> 2),
 758         0x00000000,
 759         (0x0e00 << 16) | (0x30258 >> 2),
 760         0x00000000,
 761         (0x0e00 << 16) | (0x3025c >> 2),
 762         0x00000000,
 763         (0x4e00 << 16) | (0xc900 >> 2),
 764         0x00000000,
 765         (0x5e00 << 16) | (0xc900 >> 2),
 766         0x00000000,
 767         (0x6e00 << 16) | (0xc900 >> 2),
 768         0x00000000,
 769         (0x7e00 << 16) | (0xc900 >> 2),
 770         0x00000000,
 771         (0x4e00 << 16) | (0xc904 >> 2),
 772         0x00000000,
 773         (0x5e00 << 16) | (0xc904 >> 2),
 774         0x00000000,
 775         (0x6e00 << 16) | (0xc904 >> 2),
 776         0x00000000,
 777         (0x7e00 << 16) | (0xc904 >> 2),
 778         0x00000000,
 779         (0x4e00 << 16) | (0xc908 >> 2),
 780         0x00000000,
 781         (0x5e00 << 16) | (0xc908 >> 2),
 782         0x00000000,
 783         (0x6e00 << 16) | (0xc908 >> 2),
 784         0x00000000,
 785         (0x7e00 << 16) | (0xc908 >> 2),
 786         0x00000000,
 787         (0x4e00 << 16) | (0xc90c >> 2),
 788         0x00000000,
 789         (0x5e00 << 16) | (0xc90c >> 2),
 790         0x00000000,
 791         (0x6e00 << 16) | (0xc90c >> 2),
 792         0x00000000,
 793         (0x7e00 << 16) | (0xc90c >> 2),
 794         0x00000000,
 795         (0x4e00 << 16) | (0xc910 >> 2),
 796         0x00000000,
 797         (0x5e00 << 16) | (0xc910 >> 2),
 798         0x00000000,
 799         (0x6e00 << 16) | (0xc910 >> 2),
 800         0x00000000,
 801         (0x7e00 << 16) | (0xc910 >> 2),
 802         0x00000000,
 803         (0x0e00 << 16) | (0xc99c >> 2),
 804         0x00000000,
 805         (0x0e00 << 16) | (0x9834 >> 2),
 806         0x00000000,
 807         (0x0000 << 16) | (0x30f00 >> 2),
 808         0x00000000,
 809         (0x0000 << 16) | (0x30f04 >> 2),
 810         0x00000000,
 811         (0x0000 << 16) | (0x30f08 >> 2),
 812         0x00000000,
 813         (0x0000 << 16) | (0x30f0c >> 2),
 814         0x00000000,
 815         (0x0600 << 16) | (0x9b7c >> 2),
 816         0x00000000,
 817         (0x0e00 << 16) | (0x8a14 >> 2),
 818         0x00000000,
 819         (0x0e00 << 16) | (0x8a18 >> 2),
 820         0x00000000,
 821         (0x0600 << 16) | (0x30a00 >> 2),
 822         0x00000000,
 823         (0x0e00 << 16) | (0x8bf0 >> 2),
 824         0x00000000,
 825         (0x0e00 << 16) | (0x8bcc >> 2),
 826         0x00000000,
 827         (0x0e00 << 16) | (0x8b24 >> 2),
 828         0x00000000,
 829         (0x0e00 << 16) | (0x30a04 >> 2),
 830         0x00000000,
 831         (0x0600 << 16) | (0x30a10 >> 2),
 832         0x00000000,
 833         (0x0600 << 16) | (0x30a14 >> 2),
 834         0x00000000,
 835         (0x0600 << 16) | (0x30a18 >> 2),
 836         0x00000000,
 837         (0x0600 << 16) | (0x30a2c >> 2),
 838         0x00000000,
 839         (0x0e00 << 16) | (0xc700 >> 2),
 840         0x00000000,
 841         (0x0e00 << 16) | (0xc704 >> 2),
 842         0x00000000,
 843         (0x0e00 << 16) | (0xc708 >> 2),
 844         0x00000000,
 845         (0x0e00 << 16) | (0xc768 >> 2),
 846         0x00000000,
 847         (0x0400 << 16) | (0xc770 >> 2),
 848         0x00000000,
 849         (0x0400 << 16) | (0xc774 >> 2),
 850         0x00000000,
 851         (0x0400 << 16) | (0xc798 >> 2),
 852         0x00000000,
 853         (0x0400 << 16) | (0xc79c >> 2),
 854         0x00000000,
 855         (0x0e00 << 16) | (0x9100 >> 2),
 856         0x00000000,
 857         (0x0e00 << 16) | (0x3c010 >> 2),
 858         0x00000000,
 859         (0x0e00 << 16) | (0x8c00 >> 2),
 860         0x00000000,
 861         (0x0e00 << 16) | (0x8c04 >> 2),
 862         0x00000000,
 863         (0x0e00 << 16) | (0x8c20 >> 2),
 864         0x00000000,
 865         (0x0e00 << 16) | (0x8c38 >> 2),
 866         0x00000000,
 867         (0x0e00 << 16) | (0x8c3c >> 2),
 868         0x00000000,
 869         (0x0e00 << 16) | (0xae00 >> 2),
 870         0x00000000,
 871         (0x0e00 << 16) | (0x9604 >> 2),
 872         0x00000000,
 873         (0x0e00 << 16) | (0xac08 >> 2),
 874         0x00000000,
 875         (0x0e00 << 16) | (0xac0c >> 2),
 876         0x00000000,
 877         (0x0e00 << 16) | (0xac10 >> 2),
 878         0x00000000,
 879         (0x0e00 << 16) | (0xac14 >> 2),
 880         0x00000000,
 881         (0x0e00 << 16) | (0xac58 >> 2),
 882         0x00000000,
 883         (0x0e00 << 16) | (0xac68 >> 2),
 884         0x00000000,
 885         (0x0e00 << 16) | (0xac6c >> 2),
 886         0x00000000,
 887         (0x0e00 << 16) | (0xac70 >> 2),
 888         0x00000000,
 889         (0x0e00 << 16) | (0xac74 >> 2),
 890         0x00000000,
 891         (0x0e00 << 16) | (0xac78 >> 2),
 892         0x00000000,
 893         (0x0e00 << 16) | (0xac7c >> 2),
 894         0x00000000,
 895         (0x0e00 << 16) | (0xac80 >> 2),
 896         0x00000000,
 897         (0x0e00 << 16) | (0xac84 >> 2),
 898         0x00000000,
 899         (0x0e00 << 16) | (0xac88 >> 2),
 900         0x00000000,
 901         (0x0e00 << 16) | (0xac8c >> 2),
 902         0x00000000,
 903         (0x0e00 << 16) | (0x970c >> 2),
 904         0x00000000,
 905         (0x0e00 << 16) | (0x9714 >> 2),
 906         0x00000000,
 907         (0x0e00 << 16) | (0x9718 >> 2),
 908         0x00000000,
 909         (0x0e00 << 16) | (0x971c >> 2),
 910         0x00000000,
 911         (0x0e00 << 16) | (0x31068 >> 2),
 912         0x00000000,
 913         (0x4e00 << 16) | (0x31068 >> 2),
 914         0x00000000,
 915         (0x5e00 << 16) | (0x31068 >> 2),
 916         0x00000000,
 917         (0x6e00 << 16) | (0x31068 >> 2),
 918         0x00000000,
 919         (0x7e00 << 16) | (0x31068 >> 2),
 920         0x00000000,
 921         (0x0e00 << 16) | (0xcd10 >> 2),
 922         0x00000000,
 923         (0x0e00 << 16) | (0xcd14 >> 2),
 924         0x00000000,
 925         (0x0e00 << 16) | (0x88b0 >> 2),
 926         0x00000000,
 927         (0x0e00 << 16) | (0x88b4 >> 2),
 928         0x00000000,
 929         (0x0e00 << 16) | (0x88b8 >> 2),
 930         0x00000000,
 931         (0x0e00 << 16) | (0x88bc >> 2),
 932         0x00000000,
 933         (0x0400 << 16) | (0x89c0 >> 2),
 934         0x00000000,
 935         (0x0e00 << 16) | (0x88c4 >> 2),
 936         0x00000000,
 937         (0x0e00 << 16) | (0x88c8 >> 2),
 938         0x00000000,
 939         (0x0e00 << 16) | (0x88d0 >> 2),
 940         0x00000000,
 941         (0x0e00 << 16) | (0x88d4 >> 2),
 942         0x00000000,
 943         (0x0e00 << 16) | (0x88d8 >> 2),
 944         0x00000000,
 945         (0x0e00 << 16) | (0x8980 >> 2),
 946         0x00000000,
 947         (0x0e00 << 16) | (0x30938 >> 2),
 948         0x00000000,
 949         (0x0e00 << 16) | (0x3093c >> 2),
 950         0x00000000,
 951         (0x0e00 << 16) | (0x30940 >> 2),
 952         0x00000000,
 953         (0x0e00 << 16) | (0x89a0 >> 2),
 954         0x00000000,
 955         (0x0e00 << 16) | (0x30900 >> 2),
 956         0x00000000,
 957         (0x0e00 << 16) | (0x30904 >> 2),
 958         0x00000000,
 959         (0x0e00 << 16) | (0x89b4 >> 2),
 960         0x00000000,
 961         (0x0e00 << 16) | (0x3e1fc >> 2),
 962         0x00000000,
 963         (0x0e00 << 16) | (0x3c210 >> 2),
 964         0x00000000,
 965         (0x0e00 << 16) | (0x3c214 >> 2),
 966         0x00000000,
 967         (0x0e00 << 16) | (0x3c218 >> 2),
 968         0x00000000,
 969         (0x0e00 << 16) | (0x8904 >> 2),
 970         0x00000000,
 971         0x5,
 972         (0x0e00 << 16) | (0x8c28 >> 2),
 973         (0x0e00 << 16) | (0x8c2c >> 2),
 974         (0x0e00 << 16) | (0x8c30 >> 2),
 975         (0x0e00 << 16) | (0x8c34 >> 2),
 976         (0x0e00 << 16) | (0x9600 >> 2),
 977 };
 978
 979 static const u32 bonaire_golden_spm_registers[] =
 980 {
 981         0x30800, 0xe0ffffff, 0xe0000000
 982 };
 983
 984 static const u32 bonaire_golden_common_registers[] =
 985 {
 986         0xc770, 0xffffffff, 0x00000800,
 987         0xc774, 0xffffffff, 0x00000800,
 988         0xc798, 0xffffffff, 0x00007fbf,
 989         0xc79c, 0xffffffff, 0x00007faf
 990 };
 991
 992 static const u32 bonaire_golden_registers[] =
 993 {
 994         0x3354, 0x00000333, 0x00000333,
 995         0x3350, 0x000c0fc0, 0x00040200,
 996         0x9a10, 0x00010000, 0x00058208,
 997         0x3c000, 0xffff1fff, 0x00140000,
 998         0x3c200, 0xfdfc0fff, 0x00000100,
 999         0x3c234, 0x40000000, 0x40000200,
1000         0x9830, 0xffffffff, 0x00000000,
1001         0x9834, 0xf00fffff, 0x00000400,
1002         0x9838, 0x0002021c, 0x00020200,
1003         0xc78, 0x00000080, 0x00000000,
1004         0x5bb0, 0x000000f0, 0x00000070,
1005         0x5bc0, 0xf0311fff, 0x80300000,
1006         0x98f8, 0x73773777, 0x12010001,
1007         0x350c, 0x00810000, 0x408af000,
1008         0x7030, 0x31000111, 0x00000011,
1009         0x2f48, 0x73773777, 0x12010001,
1010         0x220c, 0x00007fb6, 0x0021a1b1,
1011         0x2210, 0x00007fb6, 0x002021b1,
1012         0x2180, 0x00007fb6, 0x00002191,
1013         0x2218, 0x00007fb6, 0x002121b1,
1014         0x221c, 0x00007fb6, 0x002021b1,
1015         0x21dc, 0x00007fb6, 0x00002191,
1016         0x21e0, 0x00007fb6, 0x00002191,
1017         0x3628, 0x0000003f, 0x0000000a,
1018         0x362c, 0x0000003f, 0x0000000a,
1019         0x2ae4, 0x00073ffe, 0x000022a2,
1020         0x240c, 0x000007ff, 0x00000000,
1021         0x8a14, 0xf000003f, 0x00000007,
1022         0x8bf0, 0x00002001, 0x00000001,
1023         0x8b24, 0xffffffff, 0x00ffffff,
1024         0x30a04, 0x0000ff0f, 0x00000000,
1025         0x28a4c, 0x07ffffff, 0x06000000,
1026         0x4d8, 0x00000fff, 0x00000100,
1027         0x3e78, 0x00000001, 0x00000002,
1028         0x9100, 0x03000000, 0x0362c688,
1029         0x8c00, 0x000000ff, 0x00000001,
1030         0xe40, 0x00001fff, 0x00001fff,
1031         0x9060, 0x0000007f, 0x00000020,
1032         0x9508, 0x00010000, 0x00010000,
1033         0xac14, 0x000003ff, 0x000000f3,
1034         0xac0c, 0xffffffff, 0x00001032
1035 };
1036
1037 static const u32 bonaire_mgcg_cgcg_init[] =
1038 {
1039         0xc420, 0xffffffff, 0xfffffffc,
1040         0x30800, 0xffffffff, 0xe0000000,
1041         0x3c2a0, 0xffffffff, 0x00000100,
1042         0x3c208, 0xffffffff, 0x00000100,
1043         0x3c2c0, 0xffffffff, 0xc0000100,
1044         0x3c2c8, 0xffffffff, 0xc0000100,
1045         0x3c2c4, 0xffffffff, 0xc0000100,
1046         0x55e4, 0xffffffff, 0x00600100,
1047         0x3c280, 0xffffffff, 0x00000100,
1048         0x3c214, 0xffffffff, 0x06000100,
1049         0x3c220, 0xffffffff, 0x00000100,
1050         0x3c218, 0xffffffff, 0x06000100,
1051         0x3c204, 0xffffffff, 0x00000100,
1052         0x3c2e0, 0xffffffff, 0x00000100,
1053         0x3c224, 0xffffffff, 0x00000100,
1054         0x3c200, 0xffffffff, 0x00000100,
1055         0x3c230, 0xffffffff, 0x00000100,
1056         0x3c234, 0xffffffff, 0x00000100,
1057         0x3c250, 0xffffffff, 0x00000100,
1058         0x3c254, 0xffffffff, 0x00000100,
1059         0x3c258, 0xffffffff, 0x00000100,
1060         0x3c25c, 0xffffffff, 0x00000100,
1061         0x3c260, 0xffffffff, 0x00000100,
1062         0x3c27c, 0xffffffff, 0x00000100,
1063         0x3c278, 0xffffffff, 0x00000100,
1064         0x3c210, 0xffffffff, 0x06000100,
1065         0x3c290, 0xffffffff, 0x00000100,
1066         0x3c274, 0xffffffff, 0x00000100,
1067         0x3c2b4, 0xffffffff, 0x00000100,
1068         0x3c2b0, 0xffffffff, 0x00000100,
1069         0x3c270, 0xffffffff, 0x00000100,
1070         0x30800, 0xffffffff, 0xe0000000,
1071         0x3c020, 0xffffffff, 0x00010000,
1072         0x3c024, 0xffffffff, 0x00030002,
1073         0x3c028, 0xffffffff, 0x00040007,
1074         0x3c02c, 0xffffffff, 0x00060005,
1075         0x3c030, 0xffffffff, 0x00090008,
1076         0x3c034, 0xffffffff, 0x00010000,
1077         0x3c038, 0xffffffff, 0x00030002,
1078         0x3c03c, 0xffffffff, 0x00040007,
1079         0x3c040, 0xffffffff, 0x00060005,
1080         0x3c044, 0xffffffff, 0x00090008,
1081         0x3c048, 0xffffffff, 0x00010000,
1082         0x3c04c, 0xffffffff, 0x00030002,
1083         0x3c050, 0xffffffff, 0x00040007,
1084         0x3c054, 0xffffffff, 0x00060005,
1085         0x3c058, 0xffffffff, 0x00090008,
1086         0x3c05c, 0xffffffff, 0x00010000,
1087         0x3c060, 0xffffffff, 0x00030002,
1088         0x3c064, 0xffffffff, 0x00040007,
1089         0x3c068, 0xffffffff, 0x00060005,
1090         0x3c06c, 0xffffffff, 0x00090008,
1091         0x3c070, 0xffffffff, 0x00010000,
1092         0x3c074, 0xffffffff, 0x00030002,
1093         0x3c078, 0xffffffff, 0x00040007,
1094         0x3c07c, 0xffffffff, 0x00060005,
1095         0x3c080, 0xffffffff, 0x00090008,
1096         0x3c084, 0xffffffff, 0x00010000,
1097         0x3c088, 0xffffffff, 0x00030002,
1098         0x3c08c, 0xffffffff, 0x00040007,
1099         0x3c090, 0xffffffff, 0x00060005,
1100         0x3c094, 0xffffffff, 0x00090008,
1101         0x3c098, 0xffffffff, 0x00010000,
1102         0x3c09c, 0xffffffff, 0x00030002,
1103         0x3c0a0, 0xffffffff, 0x00040007,
1104         0x3c0a4, 0xffffffff, 0x00060005,
1105         0x3c0a8, 0xffffffff, 0x00090008,
1106         0x3c000, 0xffffffff, 0x96e00200,
1107         0x8708, 0xffffffff, 0x00900100,
1108         0xc424, 0xffffffff, 0x0020003f,
1109         0x38, 0xffffffff, 0x0140001c,
1110         0x3c, 0x000f0000, 0x000f0000,
1111         0x220, 0xffffffff, 0xC060000C,
1112         0x224, 0xc0000fff, 0x00000100,
1113         0xf90, 0xffffffff, 0x00000100,
1114         0xf98, 0x00000101, 0x00000000,
1115         0x20a8, 0xffffffff, 0x00000104,
1116         0x55e4, 0xff000fff, 0x00000100,
1117         0x30cc, 0xc0000fff, 0x00000104,
1118         0xc1e4, 0x00000001, 0x00000001,
1119         0xd00c, 0xff000ff0, 0x00000100,
1120         0xd80c, 0xff000ff0, 0x00000100
1121 };
1122
1123 static const u32 spectre_golden_spm_registers[] =
1124 {
1125         0x30800, 0xe0ffffff, 0xe0000000
1126 };
1127
1128 static const u32 spectre_golden_common_registers[] =
1129 {
1130         0xc770, 0xffffffff, 0x00000800,
1131         0xc774, 0xffffffff, 0x00000800,
1132         0xc798, 0xffffffff, 0x00007fbf,
1133         0xc79c, 0xffffffff, 0x00007faf
1134 };
1135
1136 static const u32 spectre_golden_registers[] =
1137 {
1138         0x3c000, 0xffff1fff, 0x96940200,
1139         0x3c00c, 0xffff0001, 0xff000000,
1140         0x3c200, 0xfffc0fff, 0x00000100,
1141         0x6ed8, 0x00010101, 0x00010000,
1142         0x9834, 0xf00fffff, 0x00000400,
1143         0x9838, 0xfffffffc, 0x00020200,
1144         0x5bb0, 0x000000f0, 0x00000070,
1145         0x5bc0, 0xf0311fff, 0x80300000,
1146         0x98f8, 0x73773777, 0x12010001,
1147         0x9b7c, 0x00ff0000, 0x00fc0000,
1148         0x2f48, 0x73773777, 0x12010001,
1149         0x8a14, 0xf000003f, 0x00000007,
1150         0x8b24, 0xffffffff, 0x00ffffff,
1151         0x28350, 0x3f3f3fff, 0x00000082,
1152         0x28354, 0x0000003f, 0x00000000,
1153         0x3e78, 0x00000001, 0x00000002,
1154         0x913c, 0xffff03df, 0x00000004,
1155         0xc768, 0x00000008, 0x00000008,
1156         0x8c00, 0x000008ff, 0x00000800,
1157         0x9508, 0x00010000, 0x00010000,
1158         0xac0c, 0xffffffff, 0x54763210,
1159         0x214f8, 0x01ff01ff, 0x00000002,
1160         0x21498, 0x007ff800, 0x00200000,
1161         0x2015c, 0xffffffff, 0x00000f40,
1162         0x30934, 0xffffffff, 0x00000001
1163 };
1164
1165 static const u32 spectre_mgcg_cgcg_init[] =
1166 {
1167         0xc420, 0xffffffff, 0xfffffffc,
1168         0x30800, 0xffffffff, 0xe0000000,
1169         0x3c2a0, 0xffffffff, 0x00000100,
1170         0x3c208, 0xffffffff, 0x00000100,
1171         0x3c2c0, 0xffffffff, 0x00000100,
1172         0x3c2c8, 0xffffffff, 0x00000100,
1173         0x3c2c4, 0xffffffff, 0x00000100,
1174         0x55e4, 0xffffffff, 0x00600100,
1175         0x3c280, 0xffffffff, 0x00000100,
1176         0x3c214, 0xffffffff, 0x06000100,
1177         0x3c220, 0xffffffff, 0x00000100,
1178         0x3c218, 0xffffffff, 0x06000100,
1179         0x3c204, 0xffffffff, 0x00000100,
1180         0x3c2e0, 0xffffffff, 0x00000100,
1181         0x3c224, 0xffffffff, 0x00000100,
1182         0x3c200, 0xffffffff, 0x00000100,
1183         0x3c230, 0xffffffff, 0x00000100,
1184         0x3c234, 0xffffffff, 0x00000100,
1185         0x3c250, 0xffffffff, 0x00000100,
1186         0x3c254, 0xffffffff, 0x00000100,
1187         0x3c258, 0xffffffff, 0x00000100,
1188         0x3c25c, 0xffffffff, 0x00000100,
1189         0x3c260, 0xffffffff, 0x00000100,
1190         0x3c27c, 0xffffffff, 0x00000100,
1191         0x3c278, 0xffffffff, 0x00000100,
1192         0x3c210, 0xffffffff, 0x06000100,
1193         0x3c290, 0xffffffff, 0x00000100,
1194         0x3c274, 0xffffffff, 0x00000100,
1195         0x3c2b4, 0xffffffff, 0x00000100,
1196         0x3c2b0, 0xffffffff, 0x00000100,
1197         0x3c270, 0xffffffff, 0x00000100,
1198         0x30800, 0xffffffff, 0xe0000000,
1199         0x3c020, 0xffffffff, 0x00010000,
1200         0x3c024, 0xffffffff, 0x00030002,
1201         0x3c028, 0xffffffff, 0x00040007,
1202         0x3c02c, 0xffffffff, 0x00060005,
1203         0x3c030, 0xffffffff, 0x00090008,
1204         0x3c034, 0xffffffff, 0x00010000,
1205         0x3c038, 0xffffffff, 0x00030002,
1206         0x3c03c, 0xffffffff, 0x00040007,
1207         0x3c040, 0xffffffff, 0x00060005,
1208         0x3c044, 0xffffffff, 0x00090008,
1209         0x3c048, 0xffffffff, 0x00010000,
1210         0x3c04c, 0xffffffff, 0x00030002,
1211         0x3c050, 0xffffffff, 0x00040007,
1212         0x3c054, 0xffffffff, 0x00060005,
1213         0x3c058, 0xffffffff, 0x00090008,
1214         0x3c05c, 0xffffffff, 0x00010000,
1215         0x3c060, 0xffffffff, 0x00030002,
1216         0x3c064, 0xffffffff, 0x00040007,
1217         0x3c068, 0xffffffff, 0x00060005,
1218         0x3c06c, 0xffffffff, 0x00090008,
1219         0x3c070, 0xffffffff, 0x00010000,
1220         0x3c074, 0xffffffff, 0x00030002,
1221         0x3c078, 0xffffffff, 0x00040007,
1222         0x3c07c, 0xffffffff, 0x00060005,
1223         0x3c080, 0xffffffff, 0x00090008,
1224         0x3c084, 0xffffffff, 0x00010000,
1225         0x3c088, 0xffffffff, 0x00030002,
1226         0x3c08c, 0xffffffff, 0x00040007,
1227         0x3c090, 0xffffffff, 0x00060005,
1228         0x3c094, 0xffffffff, 0x00090008,
1229         0x3c098, 0xffffffff, 0x00010000,
1230         0x3c09c, 0xffffffff, 0x00030002,
1231         0x3c0a0, 0xffffffff, 0x00040007,
1232         0x3c0a4, 0xffffffff, 0x00060005,
1233         0x3c0a8, 0xffffffff, 0x00090008,
1234         0x3c0ac, 0xffffffff, 0x00010000,
1235         0x3c0b0, 0xffffffff, 0x00030002,
1236         0x3c0b4, 0xffffffff, 0x00040007,
1237         0x3c0b8, 0xffffffff, 0x00060005,
1238         0x3c0bc, 0xffffffff, 0x00090008,
1239         0x3c000, 0xffffffff, 0x96e00200,
1240         0x8708, 0xffffffff, 0x00900100,
1241         0xc424, 0xffffffff, 0x0020003f,
1242         0x38, 0xffffffff, 0x0140001c,
1243         0x3c, 0x000f0000, 0x000f0000,
1244         0x220, 0xffffffff, 0xC060000C,
1245         0x224, 0xc0000fff, 0x00000100,
1246         0xf90, 0xffffffff, 0x00000100,
1247         0xf98, 0x00000101, 0x00000000,
1248         0x20a8, 0xffffffff, 0x00000104,
1249         0x55e4, 0xff000fff, 0x00000100,
1250         0x30cc, 0xc0000fff, 0x00000104,
1251         0xc1e4, 0x00000001, 0x00000001,
1252         0xd00c, 0xff000ff0, 0x00000100,
1253         0xd80c, 0xff000ff0, 0x00000100
1254 };
1255
1256 static const u32 kalindi_golden_spm_registers[] =
1257 {
1258         0x30800, 0xe0ffffff, 0xe0000000
1259 };
1260
1261 static const u32 kalindi_golden_common_registers[] =
1262 {
1263         0xc770, 0xffffffff, 0x00000800,
1264         0xc774, 0xffffffff, 0x00000800,
1265         0xc798, 0xffffffff, 0x00007fbf,
1266         0xc79c, 0xffffffff, 0x00007faf
1267 };
1268
1269 static const u32 kalindi_golden_registers[] =
1270 {
1271         0x3c000, 0xffffdfff, 0x6e944040,
1272         0x55e4, 0xff607fff, 0xfc000100,
1273         0x3c220, 0xff000fff, 0x00000100,
1274         0x3c224, 0xff000fff, 0x00000100,
1275         0x3c200, 0xfffc0fff, 0x00000100,
1276         0x6ed8, 0x00010101, 0x00010000,
1277         0x9830, 0xffffffff, 0x00000000,
1278         0x9834, 0xf00fffff, 0x00000400,
1279         0x5bb0, 0x000000f0, 0x00000070,
1280         0x5bc0, 0xf0311fff, 0x80300000,
1281         0x98f8, 0x73773777, 0x12010001,
1282         0x98fc, 0xffffffff, 0x00000010,
1283         0x9b7c, 0x00ff0000, 0x00fc0000,
1284         0x8030, 0x00001f0f, 0x0000100a,
1285         0x2f48, 0x73773777, 0x12010001,
1286         0x2408, 0x000fffff, 0x000c007f,
1287         0x8a14, 0xf000003f, 0x00000007,
1288         0x8b24, 0x3fff3fff, 0x00ffcfff,
1289         0x30a04, 0x0000ff0f, 0x00000000,
1290         0x28a4c, 0x07ffffff, 0x06000000,
1291         0x4d8, 0x00000fff, 0x00000100,
1292         0x3e78, 0x00000001, 0x00000002,
1293         0xc768, 0x00000008, 0x00000008,
1294         0x8c00, 0x000000ff, 0x00000003,
1295         0x214f8, 0x01ff01ff, 0x00000002,
1296         0x21498, 0x007ff800, 0x00200000,
1297         0x2015c, 0xffffffff, 0x00000f40,
1298         0x88c4, 0x001f3ae3, 0x00000082,
1299         0x88d4, 0x0000001f, 0x00000010,
1300         0x30934, 0xffffffff, 0x00000000
1301 };
1302
1303 static const u32 kalindi_mgcg_cgcg_init[] =
1304 {
1305         0xc420, 0xffffffff, 0xfffffffc,
1306         0x30800, 0xffffffff, 0xe0000000,
1307         0x3c2a0, 0xffffffff, 0x00000100,
1308         0x3c208, 0xffffffff, 0x00000100,
1309         0x3c2c0, 0xffffffff, 0x00000100,
1310         0x3c2c8, 0xffffffff, 0x00000100,
1311         0x3c2c4, 0xffffffff, 0x00000100,
1312         0x55e4, 0xffffffff, 0x00600100,
1313         0x3c280, 0xffffffff, 0x00000100,
1314         0x3c214, 0xffffffff, 0x06000100,
1315         0x3c220, 0xffffffff, 0x00000100,
1316         0x3c218, 0xffffffff, 0x06000100,
1317         0x3c204, 0xffffffff, 0x00000100,
1318         0x3c2e0, 0xffffffff, 0x00000100,
1319         0x3c224, 0xffffffff, 0x00000100,
1320         0x3c200, 0xffffffff, 0x00000100,
1321         0x3c230, 0xffffffff, 0x00000100,
1322         0x3c234, 0xffffffff, 0x00000100,
1323         0x3c250, 0xffffffff, 0x00000100,
1324         0x3c254, 0xffffffff, 0x00000100,
1325         0x3c258, 0xffffffff, 0x00000100,
1326         0x3c25c, 0xffffffff, 0x00000100,
1327         0x3c260, 0xffffffff, 0x00000100,
1328         0x3c27c, 0xffffffff, 0x00000100,
1329         0x3c278, 0xffffffff, 0x00000100,
1330         0x3c210, 0xffffffff, 0x06000100,
1331         0x3c290, 0xffffffff, 0x00000100,
1332         0x3c274, 0xffffffff, 0x00000100,
1333         0x3c2b4, 0xffffffff, 0x00000100,
1334         0x3c2b0, 0xffffffff, 0x00000100,
1335         0x3c270, 0xffffffff, 0x00000100,
1336         0x30800, 0xffffffff, 0xe0000000,
1337         0x3c020, 0xffffffff, 0x00010000,
1338         0x3c024, 0xffffffff, 0x00030002,
1339         0x3c028, 0xffffffff, 0x00040007,
1340         0x3c02c, 0xffffffff, 0x00060005,
1341         0x3c030, 0xffffffff, 0x00090008,
1342         0x3c034, 0xffffffff, 0x00010000,
1343         0x3c038, 0xffffffff, 0x00030002,
1344         0x3c03c, 0xffffffff, 0x00040007,
1345         0x3c040, 0xffffffff, 0x00060005,
1346         0x3c044, 0xffffffff, 0x00090008,
1347         0x3c000, 0xffffffff, 0x96e00200,
1348         0x8708, 0xffffffff, 0x00900100,
1349         0xc424, 0xffffffff, 0x0020003f,
1350         0x38, 0xffffffff, 0x0140001c,
1351         0x3c, 0x000f0000, 0x000f0000,
1352         0x220, 0xffffffff, 0xC060000C,
1353         0x224, 0xc0000fff, 0x00000100,
1354         0x20a8, 0xffffffff, 0x00000104,
1355         0x55e4, 0xff000fff, 0x00000100,
1356         0x30cc, 0xc0000fff, 0x00000104,
1357         0xc1e4, 0x00000001, 0x00000001,
1358         0xd00c, 0xff000ff0, 0x00000100,
1359         0xd80c, 0xff000ff0, 0x00000100
1360 };
1361
1362 static const u32 hawaii_golden_spm_registers[] =
1363 {
1364         0x30800, 0xe0ffffff, 0xe0000000
1365 };
1366
1367 static const u32 hawaii_golden_common_registers[] =
1368 {
1369         0x30800, 0xffffffff, 0xe0000000,
1370         0x28350, 0xffffffff, 0x3a00161a,
1371         0x28354, 0xffffffff, 0x0000002e,
1372         0x9a10, 0xffffffff, 0x00018208,
1373         0x98f8, 0xffffffff, 0x12011003
1374 };
1375
1376 static const u32 hawaii_golden_registers[] =
1377 {
1378         0x3354, 0x00000333, 0x00000333,
1379         0x9a10, 0x00010000, 0x00058208,
1380         0x9830, 0xffffffff, 0x00000000,
1381         0x9834, 0xf00fffff, 0x00000400,
1382         0x9838, 0x0002021c, 0x00020200,
1383         0xc78, 0x00000080, 0x00000000,
1384         0x5bb0, 0x000000f0, 0x00000070,
1385         0x5bc0, 0xf0311fff, 0x80300000,
1386         0x350c, 0x00810000, 0x408af000,
1387         0x7030, 0x31000111, 0x00000011,
1388         0x2f48, 0x73773777, 0x12010001,
1389         0x2120, 0x0000007f, 0x0000001b,
1390         0x21dc, 0x00007fb6, 0x00002191,
1391         0x3628, 0x0000003f, 0x0000000a,
1392         0x362c, 0x0000003f, 0x0000000a,
1393         0x2ae4, 0x00073ffe, 0x000022a2,
1394         0x240c, 0x000007ff, 0x00000000,
1395         0x8bf0, 0x00002001, 0x00000001,
1396         0x8b24, 0xffffffff, 0x00ffffff,
1397         0x30a04, 0x0000ff0f, 0x00000000,
1398         0x28a4c, 0x07ffffff, 0x06000000,
1399         0x3e78, 0x00000001, 0x00000002,
1400         0xc768, 0x00000008, 0x00000008,
1401         0xc770, 0x00000f00, 0x00000800,
1402         0xc774, 0x00000f00, 0x00000800,
1403         0xc798, 0x00ffffff, 0x00ff7fbf,
1404         0xc79c, 0x00ffffff, 0x00ff7faf,
1405         0x8c00, 0x000000ff, 0x00000800,
1406         0xe40, 0x00001fff, 0x00001fff,
1407         0x9060, 0x0000007f, 0x00000020,
1408         0x9508, 0x00010000, 0x00010000,
1409         0xae00, 0x00100000, 0x000ff07c,
1410         0xac14, 0x000003ff, 0x0000000f,
1411         0xac10, 0xffffffff, 0x7564fdec,
1412         0xac0c, 0xffffffff, 0x3120b9a8,
1413         0xac08, 0x20000000, 0x0f9c0000
1414 };
1415
1416 static const u32 hawaii_mgcg_cgcg_init[] =
1417 {
1418         0xc420, 0xffffffff, 0xfffffffd,
1419         0x30800, 0xffffffff, 0xe0000000,
1420         0x3c2a0, 0xffffffff, 0x00000100,
1421         0x3c208, 0xffffffff, 0x00000100,
1422         0x3c2c0, 0xffffffff, 0x00000100,
1423         0x3c2c8, 0xffffffff, 0x00000100,
1424         0x3c2c4, 0xffffffff, 0x00000100,
1425         0x55e4, 0xffffffff, 0x00200100,
1426         0x3c280, 0xffffffff, 0x00000100,
1427         0x3c214, 0xffffffff, 0x06000100,
1428         0x3c220, 0xffffffff, 0x00000100,
1429         0x3c218, 0xffffffff, 0x06000100,
1430         0x3c204, 0xffffffff, 0x00000100,
1431         0x3c2e0, 0xffffffff, 0x00000100,
1432         0x3c224, 0xffffffff, 0x00000100,
1433         0x3c200, 0xffffffff, 0x00000100,
1434         0x3c230, 0xffffffff, 0x00000100,
1435         0x3c234, 0xffffffff, 0x00000100,
1436         0x3c250, 0xffffffff, 0x00000100,
1437         0x3c254, 0xffffffff, 0x00000100,
1438         0x3c258, 0xffffffff, 0x00000100,
1439         0x3c25c, 0xffffffff, 0x00000100,
1440         0x3c260, 0xffffffff, 0x00000100,
1441         0x3c27c, 0xffffffff, 0x00000100,
1442         0x3c278, 0xffffffff, 0x00000100,
1443         0x3c210, 0xffffffff, 0x06000100,
1444         0x3c290, 0xffffffff, 0x00000100,
1445         0x3c274, 0xffffffff, 0x00000100,
1446         0x3c2b4, 0xffffffff, 0x00000100,
1447         0x3c2b0, 0xffffffff, 0x00000100,
1448         0x3c270, 0xffffffff, 0x00000100,
1449         0x30800, 0xffffffff, 0xe0000000,
1450         0x3c020, 0xffffffff, 0x00010000,
1451         0x3c024, 0xffffffff, 0x00030002,
1452         0x3c028, 0xffffffff, 0x00040007,
1453         0x3c02c, 0xffffffff, 0x00060005,
1454         0x3c030, 0xffffffff, 0x00090008,
1455         0x3c034, 0xffffffff, 0x00010000,
1456         0x3c038, 0xffffffff, 0x00030002,
1457         0x3c03c, 0xffffffff, 0x00040007,
1458         0x3c040, 0xffffffff, 0x00060005,
1459         0x3c044, 0xffffffff, 0x00090008,
1460         0x3c048, 0xffffffff, 0x00010000,
1461         0x3c04c, 0xffffffff, 0x00030002,
1462         0x3c050, 0xffffffff, 0x00040007,
1463         0x3c054, 0xffffffff, 0x00060005,
1464         0x3c058, 0xffffffff, 0x00090008,
1465         0x3c05c, 0xffffffff, 0x00010000,
1466         0x3c060, 0xffffffff, 0x00030002,
1467         0x3c064, 0xffffffff, 0x00040007,
1468         0x3c068, 0xffffffff, 0x00060005,
1469         0x3c06c, 0xffffffff, 0x00090008,
1470         0x3c070, 0xffffffff, 0x00010000,
1471         0x3c074, 0xffffffff, 0x00030002,
1472         0x3c078, 0xffffffff, 0x00040007,
1473         0x3c07c, 0xffffffff, 0x00060005,
1474         0x3c080, 0xffffffff, 0x00090008,
1475         0x3c084, 0xffffffff, 0x00010000,
1476         0x3c088, 0xffffffff, 0x00030002,
1477         0x3c08c, 0xffffffff, 0x00040007,
1478         0x3c090, 0xffffffff, 0x00060005,
1479         0x3c094, 0xffffffff, 0x00090008,
1480         0x3c098, 0xffffffff, 0x00010000,
1481         0x3c09c, 0xffffffff, 0x00030002,
1482         0x3c0a0, 0xffffffff, 0x00040007,
1483         0x3c0a4, 0xffffffff, 0x00060005,
1484         0x3c0a8, 0xffffffff, 0x00090008,
1485         0x3c0ac, 0xffffffff, 0x00010000,
1486         0x3c0b0, 0xffffffff, 0x00030002,
1487         0x3c0b4, 0xffffffff, 0x00040007,
1488         0x3c0b8, 0xffffffff, 0x00060005,
1489         0x3c0bc, 0xffffffff, 0x00090008,
1490         0x3c0c0, 0xffffffff, 0x00010000,
1491         0x3c0c4, 0xffffffff, 0x00030002,
1492         0x3c0c8, 0xffffffff, 0x00040007,
1493         0x3c0cc, 0xffffffff, 0x00060005,
1494         0x3c0d0, 0xffffffff, 0x00090008,
1495         0x3c0d4, 0xffffffff, 0x00010000,
1496         0x3c0d8, 0xffffffff, 0x00030002,
1497         0x3c0dc, 0xffffffff, 0x00040007,
1498         0x3c0e0, 0xffffffff, 0x00060005,
1499         0x3c0e4, 0xffffffff, 0x00090008,
1500         0x3c0e8, 0xffffffff, 0x00010000,
1501         0x3c0ec, 0xffffffff, 0x00030002,
1502         0x3c0f0, 0xffffffff, 0x00040007,
1503         0x3c0f4, 0xffffffff, 0x00060005,
1504         0x3c0f8, 0xffffffff, 0x00090008,
1505         0xc318, 0xffffffff, 0x00020200,
1506         0x3350, 0xffffffff, 0x00000200,
1507         0x15c0, 0xffffffff, 0x00000400,
1508         0x55e8, 0xffffffff, 0x00000000,
1509         0x2f50, 0xffffffff, 0x00000902,
1510         0x3c000, 0xffffffff, 0x96940200,
1511         0x8708, 0xffffffff, 0x00900100,
1512         0xc424, 0xffffffff, 0x0020003f,
1513         0x38, 0xffffffff, 0x0140001c,
1514         0x3c, 0x000f0000, 0x000f0000,
1515         0x220, 0xffffffff, 0xc060000c,
1516         0x224, 0xc0000fff, 0x00000100,
1517         0xf90, 0xffffffff, 0x00000100,
1518         0xf98, 0x00000101, 0x00000000,
1519         0x20a8, 0xffffffff, 0x00000104,
1520         0x55e4, 0xff000fff, 0x00000100,
1521         0x30cc, 0xc0000fff, 0x00000104,
1522         0xc1e4, 0x00000001, 0x00000001,
1523         0xd00c, 0xff000ff0, 0x00000100,
1524         0xd80c, 0xff000ff0, 0x00000100
1525 };
1526
1527 static const u32 godavari_golden_registers[] =
1528 {
1529         0x55e4, 0xff607fff, 0xfc000100,
1530         0x6ed8, 0x00010101, 0x00010000,
1531         0x9830, 0xffffffff, 0x00000000,
1532         0x98302, 0xf00fffff, 0x00000400,
1533         0x6130, 0xffffffff, 0x00010000,
1534         0x5bb0, 0x000000f0, 0x00000070,
1535         0x5bc0, 0xf0311fff, 0x80300000,
1536         0x98f8, 0x73773777, 0x12010001,
1537         0x98fc, 0xffffffff, 0x00000010,
1538         0x8030, 0x00001f0f, 0x0000100a,
1539         0x2f48, 0x73773777, 0x12010001,
1540         0x2408, 0x000fffff, 0x000c007f,
1541         0x8a14, 0xf000003f, 0x00000007,
1542         0x8b24, 0xffffffff, 0x00ff0fff,
1543         0x30a04, 0x0000ff0f, 0x00000000,
1544         0x28a4c, 0x07ffffff, 0x06000000,
1545         0x4d8, 0x00000fff, 0x00000100,
1546         0xd014, 0x00010000, 0x00810001,
1547         0xd814, 0x00010000, 0x00810001,
1548         0x3e78, 0x00000001, 0x00000002,
1549         0xc768, 0x00000008, 0x00000008,
1550         0xc770, 0x00000f00, 0x00000800,
1551         0xc774, 0x00000f00, 0x00000800,
1552         0xc798, 0x00ffffff, 0x00ff7fbf,
1553         0xc79c, 0x00ffffff, 0x00ff7faf,
1554         0x8c00, 0x000000ff, 0x00000001,
1555         0x214f8, 0x01ff01ff, 0x00000002,
1556         0x21498, 0x007ff800, 0x00200000,
1557         0x2015c, 0xffffffff, 0x00000f40,
1558         0x88c4, 0x001f3ae3, 0x00000082,
1559         0x88d4, 0x0000001f, 0x00000010,
1560         0x30934, 0xffffffff, 0x00000000
1561 };
1562
1563
1564 static void cik_init_golden_registers(struct radeon_device *rdev)
1565 {
1566         switch (rdev->family) {
1567         case CHIP_BONAIRE:
1568                 radeon_program_register_sequence(rdev,
1569                                                  bonaire_mgcg_cgcg_init,
1570                                                  (const u32)ARRAY_SIZE(bonaire_mgcg_cgcg_init));
1571                 radeon_program_register_sequence(rdev,
1572                                                  bonaire_golden_registers,
1573                                                  (const u32)ARRAY_SIZE(bonaire_golden_registers));
1574                 radeon_program_register_sequence(rdev,
1575                                                  bonaire_golden_common_registers,
1576                                                  (const u32)ARRAY_SIZE(bonaire_golden_common_registers));
1577                 radeon_program_register_sequence(rdev,
1578                                                  bonaire_golden_spm_registers,
1579                                                  (const u32)ARRAY_SIZE(bonaire_golden_spm_registers));
1580                 break;
1581         case CHIP_KABINI:
1582                 radeon_program_register_sequence(rdev,
1583                                                  kalindi_mgcg_cgcg_init,
1584                                                  (const u32)ARRAY_SIZE(kalindi_mgcg_cgcg_init));
1585                 radeon_program_register_sequence(rdev,
1586                                                  kalindi_golden_registers,
1587                                                  (const u32)ARRAY_SIZE(kalindi_golden_registers));
1588                 radeon_program_register_sequence(rdev,
1589                                                  kalindi_golden_common_registers,
1590                                                  (const u32)ARRAY_SIZE(kalindi_golden_common_registers));
1591                 radeon_program_register_sequence(rdev,
1592                                                  kalindi_golden_spm_registers,
1593                                                  (const u32)ARRAY_SIZE(kalindi_golden_spm_registers));
1594                 break;
1595         case CHIP_MULLINS:
1596                 radeon_program_register_sequence(rdev,
1597                                                  kalindi_mgcg_cgcg_init,
1598                                                  (const u32)ARRAY_SIZE(kalindi_mgcg_cgcg_init));
1599                 radeon_program_register_sequence(rdev,
1600                                                  godavari_golden_registers,
1601                                                  (const u32)ARRAY_SIZE(godavari_golden_registers));
1602                 radeon_program_register_sequence(rdev,
1603                                                  kalindi_golden_common_registers,
1604                                                  (const u32)ARRAY_SIZE(kalindi_golden_common_registers));
1605                 radeon_program_register_sequence(rdev,
1606                                                  kalindi_golden_spm_registers,
1607                                                  (const u32)ARRAY_SIZE(kalindi_golden_spm_registers));
1608                 break;
1609         case CHIP_KAVERI:
1610                 radeon_program_register_sequence(rdev,
1611                                                  spectre_mgcg_cgcg_init,
1612                                                  (const u32)ARRAY_SIZE(spectre_mgcg_cgcg_init));
1613                 radeon_program_register_sequence(rdev,
1614                                                  spectre_golden_registers,
1615                                                  (const u32)ARRAY_SIZE(spectre_golden_registers));
1616                 radeon_program_register_sequence(rdev,
1617                                                  spectre_golden_common_registers,
1618                                                  (const u32)ARRAY_SIZE(spectre_golden_common_registers));
1619                 radeon_program_register_sequence(rdev,
1620                                                  spectre_golden_spm_registers,
1621                                                  (const u32)ARRAY_SIZE(spectre_golden_spm_registers));
1622                 break;
1623         case CHIP_HAWAII:
1624                 radeon_program_register_sequence(rdev,
1625                                                  hawaii_mgcg_cgcg_init,
1626                                                  (const u32)ARRAY_SIZE(hawaii_mgcg_cgcg_init));
1627                 radeon_program_register_sequence(rdev,
1628                                                  hawaii_golden_registers,
1629                                                  (const u32)ARRAY_SIZE(hawaii_golden_registers));
1630                 radeon_program_register_sequence(rdev,
1631                                                  hawaii_golden_common_registers,
1632                                                  (const u32)ARRAY_SIZE(hawaii_golden_common_registers));
1633                 radeon_program_register_sequence(rdev,
1634                                                  hawaii_golden_spm_registers,
1635                                                  (const u32)ARRAY_SIZE(hawaii_golden_spm_registers));
1636                 break;
1637         default:
1638                 break;
1639         }
1640 }
1641
1642 /**
1643  * cik_get_xclk - get the xclk
1644  *
1645  * @rdev: radeon_device pointer
1646  *
1647  * Returns the reference clock used by the gfx engine
1648  * (CIK).
1649  */
1650 u32 cik_get_xclk(struct radeon_device *rdev)
1651 {
1652         u32 reference_clock = rdev->clock.spll.reference_freq;
1653
1654         if (rdev->flags & RADEON_IS_IGP) {
1655                 if (RREG32_SMC(GENERAL_PWRMGT) & GPU_COUNTER_CLK)
1656                         return reference_clock / 2;
1657         } else {
1658                 if (RREG32_SMC(CG_CLKPIN_CNTL) & XTALIN_DIVIDE)
1659                         return reference_clock / 4;
1660         }
1661         return reference_clock;
1662 }
1663
1664 /**
1665  * cik_mm_rdoorbell - read a doorbell dword
1666  *
1667  * @rdev: radeon_device pointer
1668  * @index: doorbell index
1669  *
1670  * Returns the value in the doorbell aperture at the
1671  * requested doorbell index (CIK).
1672  */
1673 u32 cik_mm_rdoorbell(struct radeon_device *rdev, u32 index)
1674 {
1675         if (index < rdev->doorbell.num_doorbells) {
1676                 return readl(rdev->doorbell.ptr + index);
1677         } else {
1678                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
1679                 return 0;
1680         }
1681 }
1682
1683 /**
1684  * cik_mm_wdoorbell - write a doorbell dword
1685  *
1686  * @rdev: radeon_device pointer
1687  * @index: doorbell index
1688  * @v: value to write
1689  *
1690  * Writes @v to the doorbell aperture at the
1691  * requested doorbell index (CIK).
1692  */
1693 void cik_mm_wdoorbell(struct radeon_device *rdev, u32 index, u32 v)
1694 {
1695         if (index < rdev->doorbell.num_doorbells) {
1696                 writel(v, rdev->doorbell.ptr + index);
1697         } else {
1698                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
1699         }
1700 }
1701
1702 #define BONAIRE_IO_MC_REGS_SIZE 36
1703
1704 static const u32 bonaire_io_mc_regs[BONAIRE_IO_MC_REGS_SIZE][2] =
1705 {
1706         {0x00000070, 0x04400000},
1707         {0x00000071, 0x80c01803},
1708         {0x00000072, 0x00004004},
1709         {0x00000073, 0x00000100},
1710         {0x00000074, 0x00ff0000},
1711         {0x00000075, 0x34000000},
1712         {0x00000076, 0x08000014},
1713         {0x00000077, 0x00cc08ec},
1714         {0x00000078, 0x00000400},
1715         {0x00000079, 0x00000000},
1716         {0x0000007a, 0x04090000},
1717         {0x0000007c, 0x00000000},
1718         {0x0000007e, 0x4408a8e8},
1719         {0x0000007f, 0x00000304},
1720         {0x00000080, 0x00000000},
1721         {0x00000082, 0x00000001},
1722         {0x00000083, 0x00000002},
1723         {0x00000084, 0xf3e4f400},
1724         {0x00000085, 0x052024e3},
1725         {0x00000087, 0x00000000},
1726         {0x00000088, 0x01000000},
1727         {0x0000008a, 0x1c0a0000},
1728         {0x0000008b, 0xff010000},
1729         {0x0000008d, 0xffffefff},
1730         {0x0000008e, 0xfff3efff},
1731         {0x0000008f, 0xfff3efbf},
1732         {0x00000092, 0xf7ffffff},
1733         {0x00000093, 0xffffff7f},
1734         {0x00000095, 0x00101101},
1735         {0x00000096, 0x00000fff},
1736         {0x00000097, 0x00116fff},
1737         {0x00000098, 0x60010000},
1738         {0x00000099, 0x10010000},
1739         {0x0000009a, 0x00006000},
1740         {0x0000009b, 0x00001000},
1741         {0x0000009f, 0x00b48000}
1742 };
1743
1744 #define HAWAII_IO_MC_REGS_SIZE 22
1745
1746 static const u32 hawaii_io_mc_regs[HAWAII_IO_MC_REGS_SIZE][2] =
1747 {
1748         {0x0000007d, 0x40000000},
1749         {0x0000007e, 0x40180304},
1750         {0x0000007f, 0x0000ff00},
1751         {0x00000081, 0x00000000},
1752         {0x00000083, 0x00000800},
1753         {0x00000086, 0x00000000},
1754         {0x00000087, 0x00000100},
1755         {0x00000088, 0x00020100},
1756         {0x00000089, 0x00000000},
1757         {0x0000008b, 0x00040000},
1758         {0x0000008c, 0x00000100},
1759         {0x0000008e, 0xff010000},
1760         {0x00000090, 0xffffefff},
1761         {0x00000091, 0xfff3efff},
1762         {0x00000092, 0xfff3efbf},
1763         {0x00000093, 0xf7ffffff},
1764         {0x00000094, 0xffffff7f},
1765         {0x00000095, 0x00000fff},
1766         {0x00000096, 0x00116fff},
1767         {0x00000097, 0x60010000},
1768         {0x00000098, 0x10010000},
1769         {0x0000009f, 0x00c79000}
1770 };
1771
1772
1773 /**
1774  * cik_srbm_select - select specific register instances
1775  *
1776  * @rdev: radeon_device pointer
1777  * @me: selected ME (micro engine)
1778  * @pipe: pipe
1779  * @queue: queue
1780  * @vmid: VMID
1781  *
1782  * Switches the currently active registers instances.  Some
1783  * registers are instanced per VMID, others are instanced per
1784  * me/pipe/queue combination.
1785  */
1786 static void cik_srbm_select(struct radeon_device *rdev,
1787                             u32 me, u32 pipe, u32 queue, u32 vmid)
1788 {
1789         u32 srbm_gfx_cntl = (PIPEID(pipe & 0x3) |
1790                              MEID(me & 0x3) |
1791                              VMID(vmid & 0xf) |
1792                              QUEUEID(queue & 0x7));
1793         WREG32(SRBM_GFX_CNTL, srbm_gfx_cntl);
1794 }
1795
1796 /* ucode loading */
1797 /**
1798  * ci_mc_load_microcode - load MC ucode into the hw
1799  *
1800  * @rdev: radeon_device pointer
1801  *
1802  * Load the GDDR MC ucode into the hw (CIK).
1803  * Returns 0 on success, error on failure.
1804  */
1805 int ci_mc_load_microcode(struct radeon_device *rdev)
1806 {
1807         const __be32 *fw_data = NULL;
1808         const __le32 *new_fw_data = NULL;
1809         u32 running, blackout = 0;
1810         u32 *io_mc_regs = NULL;
1811         const __le32 *new_io_mc_regs = NULL;
1812         int i, regs_size, ucode_size;
1813
1814         if (!rdev->mc_fw)
1815                 return -EINVAL;
1816
1817         if (rdev->new_fw) {
1818                 const struct mc_firmware_header_v1_0 *hdr =
1819                         (const struct mc_firmware_header_v1_0 *)rdev->mc_fw->data;
1820
1821                 radeon_ucode_print_mc_hdr(&hdr->header);
1822
1823                 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
1824                 new_io_mc_regs = (const __le32 *)
1825                         (rdev->mc_fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
1826                 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
1827                 new_fw_data = (const __le32 *)
1828                         (rdev->mc_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1829         } else {
1830                 ucode_size = rdev->mc_fw->size / 4;
1831
1832                 switch (rdev->family) {
1833                 case CHIP_BONAIRE:
1834                         io_mc_regs = (u32 *)&bonaire_io_mc_regs;
1835                         regs_size = BONAIRE_IO_MC_REGS_SIZE;
1836                         break;
1837                 case CHIP_HAWAII:
1838                         io_mc_regs = (u32 *)&hawaii_io_mc_regs;
1839                         regs_size = HAWAII_IO_MC_REGS_SIZE;
1840                         break;
1841                 default:
1842                         return -EINVAL;
1843                 }
1844                 fw_data = (const __be32 *)rdev->mc_fw->data;
1845         }
1846
1847         running = RREG32(MC_SEQ_SUP_CNTL) & RUN_MASK;
1848
1849         if (running == 0) {
1850                 if (running) {
1851                         blackout = RREG32(MC_SHARED_BLACKOUT_CNTL);
1852                         WREG32(MC_SHARED_BLACKOUT_CNTL, blackout | 1);
1853                 }
1854
1855                 /* reset the engine and set to writable */
1856                 WREG32(MC_SEQ_SUP_CNTL, 0x00000008);
1857                 WREG32(MC_SEQ_SUP_CNTL, 0x00000010);
1858
1859                 /* load mc io regs */
1860                 for (i = 0; i < regs_size; i++) {
1861                         if (rdev->new_fw) {
1862                                 WREG32(MC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(new_io_mc_regs++));
1863                                 WREG32(MC_SEQ_IO_DEBUG_DATA, le32_to_cpup(new_io_mc_regs++));
1864                         } else {
1865                                 WREG32(MC_SEQ_IO_DEBUG_INDEX, io_mc_regs[(i << 1)]);
1866                                 WREG32(MC_SEQ_IO_DEBUG_DATA, io_mc_regs[(i << 1) + 1]);
1867                         }
1868                 }
1869                 /* load the MC ucode */
1870                 for (i = 0; i < ucode_size; i++) {
1871                         if (rdev->new_fw)
1872                                 WREG32(MC_SEQ_SUP_PGM, le32_to_cpup(new_fw_data++));
1873                         else
1874                                 WREG32(MC_SEQ_SUP_PGM, be32_to_cpup(fw_data++));
1875                 }
1876
1877                 /* put the engine back into the active state */
1878                 WREG32(MC_SEQ_SUP_CNTL, 0x00000008);
1879                 WREG32(MC_SEQ_SUP_CNTL, 0x00000004);
1880                 WREG32(MC_SEQ_SUP_CNTL, 0x00000001);
1881
1882                 /* wait for training to complete */
1883                 for (i = 0; i < rdev->usec_timeout; i++) {
1884                         if (RREG32(MC_SEQ_TRAIN_WAKEUP_CNTL) & TRAIN_DONE_D0)
1885                                 break;
1886                         udelay(1);
1887                 }
1888                 for (i = 0; i < rdev->usec_timeout; i++) {
1889                         if (RREG32(MC_SEQ_TRAIN_WAKEUP_CNTL) & TRAIN_DONE_D1)
1890                                 break;
1891                         udelay(1);
1892                 }
1893
1894                 if (running)
1895                         WREG32(MC_SHARED_BLACKOUT_CNTL, blackout);
1896         }
1897
1898         return 0;
1899 }
1900
1901 /**
1902  * cik_init_microcode - load ucode images from disk
1903  *
1904  * @rdev: radeon_device pointer
1905  *
1906  * Use the firmware interface to load the ucode images into
1907  * the driver (not loaded into hw).
1908  * Returns 0 on success, error on failure.
1909  */
1910 static int cik_init_microcode(struct radeon_device *rdev)
1911 {
1912         const char *chip_name;
1913         const char *new_chip_name;
1914         size_t pfp_req_size, me_req_size, ce_req_size,
1915                 mec_req_size, rlc_req_size, mc_req_size = 0,
1916                 sdma_req_size, smc_req_size = 0, mc2_req_size = 0;
1917         char fw_name[30];
1918         int new_fw = 0;
1919         int err;
1920         int num_fw;
1921
1922         DRM_DEBUG("\n");
1923
1924         switch (rdev->family) {
1925         case CHIP_BONAIRE:
1926                 chip_name = "BONAIRE";
1927                 new_chip_name = "bonaire";
1928                 pfp_req_size = CIK_PFP_UCODE_SIZE * 4;
1929                 me_req_size = CIK_ME_UCODE_SIZE * 4;
1930                 ce_req_size = CIK_CE_UCODE_SIZE * 4;
1931                 mec_req_size = CIK_MEC_UCODE_SIZE * 4;
1932                 rlc_req_size = BONAIRE_RLC_UCODE_SIZE * 4;
1933                 mc_req_size = BONAIRE_MC_UCODE_SIZE * 4;
1934                 mc2_req_size = BONAIRE_MC2_UCODE_SIZE * 4;
1935                 sdma_req_size = CIK_SDMA_UCODE_SIZE * 4;
1936                 smc_req_size = ALIGN(BONAIRE_SMC_UCODE_SIZE, 4);
1937                 num_fw = 8;
1938                 break;
1939         case CHIP_HAWAII:
1940                 chip_name = "HAWAII";
1941                 new_chip_name = "hawaii";
1942                 pfp_req_size = CIK_PFP_UCODE_SIZE * 4;
1943                 me_req_size = CIK_ME_UCODE_SIZE * 4;
1944                 ce_req_size = CIK_CE_UCODE_SIZE * 4;
1945                 mec_req_size = CIK_MEC_UCODE_SIZE * 4;
1946                 rlc_req_size = BONAIRE_RLC_UCODE_SIZE * 4;
1947                 mc_req_size = HAWAII_MC_UCODE_SIZE * 4;
1948                 mc2_req_size = HAWAII_MC2_UCODE_SIZE * 4;
1949                 sdma_req_size = CIK_SDMA_UCODE_SIZE * 4;
1950                 smc_req_size = ALIGN(HAWAII_SMC_UCODE_SIZE, 4);
1951                 num_fw = 8;
1952                 break;
1953         case CHIP_KAVERI:
1954                 chip_name = "KAVERI";
1955                 new_chip_name = "kaveri";
1956                 pfp_req_size = CIK_PFP_UCODE_SIZE * 4;
1957                 me_req_size = CIK_ME_UCODE_SIZE * 4;
1958                 ce_req_size = CIK_CE_UCODE_SIZE * 4;
1959                 mec_req_size = CIK_MEC_UCODE_SIZE * 4;
1960                 rlc_req_size = KV_RLC_UCODE_SIZE * 4;
1961                 sdma_req_size = CIK_SDMA_UCODE_SIZE * 4;
1962                 num_fw = 7;
1963                 break;
1964         case CHIP_KABINI:
1965                 chip_name = "KABINI";
1966                 new_chip_name = "kabini";
1967                 pfp_req_size = CIK_PFP_UCODE_SIZE * 4;
1968                 me_req_size = CIK_ME_UCODE_SIZE * 4;
1969                 ce_req_size = CIK_CE_UCODE_SIZE * 4;
1970                 mec_req_size = CIK_MEC_UCODE_SIZE * 4;
1971                 rlc_req_size = KB_RLC_UCODE_SIZE * 4;
1972                 sdma_req_size = CIK_SDMA_UCODE_SIZE * 4;
1973                 num_fw = 6;
1974                 break;
1975         case CHIP_MULLINS:
1976                 chip_name = "MULLINS";
1977                 new_chip_name = "mullins";
1978                 pfp_req_size = CIK_PFP_UCODE_SIZE * 4;
1979                 me_req_size = CIK_ME_UCODE_SIZE * 4;
1980                 ce_req_size = CIK_CE_UCODE_SIZE * 4;
1981                 mec_req_size = CIK_MEC_UCODE_SIZE * 4;
1982                 rlc_req_size = ML_RLC_UCODE_SIZE * 4;
1983                 sdma_req_size = CIK_SDMA_UCODE_SIZE * 4;
1984                 num_fw = 6;
1985                 break;
1986         default: BUG();
1987         }
1988
1989         DRM_INFO("Loading %s Microcode\n", new_chip_name);
1990
1991         snprintf(fw_name, sizeof(fw_name), "radeon/%s_pfp.bin", new_chip_name);
1992         err = request_firmware(&rdev->pfp_fw, fw_name, rdev->dev);
1993         if (err) {
1994                 snprintf(fw_name, sizeof(fw_name), "radeon/%s_pfp.bin", chip_name);
1995                 err = request_firmware(&rdev->pfp_fw, fw_name, rdev->dev);
1996                 if (err)
1997                         goto out;
1998                 if (rdev->pfp_fw->size != pfp_req_size) {
1999                         printk(KERN_ERR
2000                                "cik_cp: Bogus length %zu in firmware \"%s\"\n",
2001                                rdev->pfp_fw->size, fw_name);
2002                         err = -EINVAL;
2003                         goto out;
2004                 }
2005         } else {
2006                 err = radeon_ucode_validate(rdev->pfp_fw);
2007                 if (err) {
2008                         printk(KERN_ERR
2009                                "cik_fw: validation failed for firmware \"%s\"\n",
2010                                fw_name);
2011                         goto out;
2012                 } else {
2013                         new_fw++;
2014                 }
2015         }
2016
2017         snprintf(fw_name, sizeof(fw_name), "radeon/%s_me.bin", new_chip_name);
2018         err = request_firmware(&rdev->me_fw, fw_name, rdev->dev);
2019         if (err) {
2020                 snprintf(fw_name, sizeof(fw_name), "radeon/%s_me.bin", chip_name);
2021                 err = request_firmware(&rdev->me_fw, fw_name, rdev->dev);
2022                 if (err)
2023                         goto out;
2024                 if (rdev->me_fw->size != me_req_size) {
2025                         printk(KERN_ERR
2026                                "cik_cp: Bogus length %zu in firmware \"%s\"\n",
2027                                rdev->me_fw->size, fw_name);
2028                         err = -EINVAL;
2029                 }
2030         } else {
2031                 err = radeon_ucode_validate(rdev->me_fw);
2032                 if (err) {
2033                         printk(KERN_ERR
2034                                "cik_fw: validation failed for firmware \"%s\"\n",
2035                                fw_name);
2036                         goto out;
2037                 } else {
2038                         new_fw++;
2039                 }
2040         }
2041
2042         snprintf(fw_name, sizeof(fw_name), "radeon/%s_ce.bin", new_chip_name);
2043         err = request_firmware(&rdev->ce_fw, fw_name, rdev->dev);
2044         if (err) {
2045                 snprintf(fw_name, sizeof(fw_name), "radeon/%s_ce.bin", chip_name);
2046                 err = request_firmware(&rdev->ce_fw, fw_name, rdev->dev);
2047                 if (err)
2048                         goto out;
2049                 if (rdev->ce_fw->size != ce_req_size) {
2050                         printk(KERN_ERR
2051                                "cik_cp: Bogus length %zu in firmware \"%s\"\n",
2052                                rdev->ce_fw->size, fw_name);
2053                         err = -EINVAL;
2054                 }
2055         } else {
2056                 err = radeon_ucode_validate(rdev->ce_fw);
2057                 if (err) {
2058                         printk(KERN_ERR
2059                                "cik_fw: validation failed for firmware \"%s\"\n",
2060                                fw_name);
2061                         goto out;
2062                 } else {
2063                         new_fw++;
2064                 }
2065         }
2066
2067         snprintf(fw_name, sizeof(fw_name), "radeon/%s_mec.bin", new_chip_name);
2068         err = request_firmware(&rdev->mec_fw, fw_name, rdev->dev);
2069         if (err) {
2070                 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mec.bin", chip_name);
2071                 err = request_firmware(&rdev->mec_fw, fw_name, rdev->dev);
2072                 if (err)
2073                         goto out;
2074                 if (rdev->mec_fw->size != mec_req_size) {
2075                         printk(KERN_ERR
2076                                "cik_cp: Bogus length %zu in firmware \"%s\"\n",
2077                                rdev->mec_fw->size, fw_name);
2078                         err = -EINVAL;
2079                 }
2080         } else {
2081                 err = radeon_ucode_validate(rdev->mec_fw);
2082                 if (err) {
2083                         printk(KERN_ERR
2084                                "cik_fw: validation failed for firmware \"%s\"\n",
2085                                fw_name);
2086                         goto out;
2087                 } else {
2088                         new_fw++;
2089                 }
2090         }
2091
2092         if (rdev->family == CHIP_KAVERI) {
2093                 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mec2.bin", new_chip_name);
2094                 err = request_firmware(&rdev->mec2_fw, fw_name, rdev->dev);
2095                 if (err) {
2096                         goto out;
2097                 } else {
2098                         err = radeon_ucode_validate(rdev->mec2_fw);
2099                         if (err) {
2100                                 goto out;
2101                         } else {
2102                                 new_fw++;
2103                         }
2104                 }
2105         }
2106
2107         snprintf(fw_name, sizeof(fw_name), "radeon/%s_rlc.bin", new_chip_name);
2108         err = request_firmware(&rdev->rlc_fw, fw_name, rdev->dev);
2109         if (err) {
2110                 snprintf(fw_name, sizeof(fw_name), "radeon/%s_rlc.bin", chip_name);
2111                 err = request_firmware(&rdev->rlc_fw, fw_name, rdev->dev);
2112                 if (err)
2113                         goto out;
2114                 if (rdev->rlc_fw->size != rlc_req_size) {
2115                         printk(KERN_ERR
2116                                "cik_rlc: Bogus length %zu in firmware \"%s\"\n",
2117                                rdev->rlc_fw->size, fw_name);
2118                         err = -EINVAL;
2119                 }
2120         } else {
2121                 err = radeon_ucode_validate(rdev->rlc_fw);
2122                 if (err) {
2123                         printk(KERN_ERR
2124                                "cik_fw: validation failed for firmware \"%s\"\n",
2125                                fw_name);
2126                         goto out;
2127                 } else {
2128                         new_fw++;
2129                 }
2130         }
2131
2132         snprintf(fw_name, sizeof(fw_name), "radeon/%s_sdma.bin", new_chip_name);
2133         err = request_firmware(&rdev->sdma_fw, fw_name, rdev->dev);
2134         if (err) {
2135                 snprintf(fw_name, sizeof(fw_name), "radeon/%s_sdma.bin", chip_name);
2136                 err = request_firmware(&rdev->sdma_fw, fw_name, rdev->dev);
2137                 if (err)
2138                         goto out;
2139                 if (rdev->sdma_fw->size != sdma_req_size) {
2140                         printk(KERN_ERR
2141                                "cik_sdma: Bogus length %zu in firmware \"%s\"\n",
2142                                rdev->sdma_fw->size, fw_name);
2143                         err = -EINVAL;
2144                 }
2145         } else {
2146                 err = radeon_ucode_validate(rdev->sdma_fw);
2147                 if (err) {
2148                         printk(KERN_ERR
2149                                "cik_fw: validation failed for firmware \"%s\"\n",
2150                                fw_name);
2151                         goto out;
2152                 } else {
2153                         new_fw++;
2154                 }
2155         }
2156
2157         /* No SMC, MC ucode on APUs */
2158         if (!(rdev->flags & RADEON_IS_IGP)) {
2159                 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mc.bin", new_chip_name);
2160                 err = request_firmware(&rdev->mc_fw, fw_name, rdev->dev);
2161                 if (err) {
2162                         snprintf(fw_name, sizeof(fw_name), "radeon/%s_mc2.bin", chip_name);
2163                         err = request_firmware(&rdev->mc_fw, fw_name, rdev->dev);
2164                         if (err) {
2165                                 snprintf(fw_name, sizeof(fw_name), "radeon/%s_mc.bin", chip_name);
2166                                 err = request_firmware(&rdev->mc_fw, fw_name, rdev->dev);
2167                                 if (err)
2168                                         goto out;
2169                         }
2170                         if ((rdev->mc_fw->size != mc_req_size) &&
2171                             (rdev->mc_fw->size != mc2_req_size)){
2172                                 printk(KERN_ERR
2173                                        "cik_mc: Bogus length %zu in firmware \"%s\"\n",
2174                                        rdev->mc_fw->size, fw_name);
2175                                 err = -EINVAL;
2176                         }
2177                         DRM_INFO("%s: %zu bytes\n", fw_name, rdev->mc_fw->size);
2178                 } else {
2179                         err = radeon_ucode_validate(rdev->mc_fw);
2180                         if (err) {
2181                                 printk(KERN_ERR
2182                                        "cik_fw: validation failed for firmware \"%s\"\n",
2183                                        fw_name);
2184                                 goto out;
2185                         } else {
2186                                 new_fw++;
2187                         }
2188                 }
2189
2190                 snprintf(fw_name, sizeof(fw_name), "radeon/%s_smc.bin", new_chip_name);
2191                 err = request_firmware(&rdev->smc_fw, fw_name, rdev->dev);
2192                 if (err) {
2193                         snprintf(fw_name, sizeof(fw_name), "radeon/%s_smc.bin", chip_name);
2194                         err = request_firmware(&rdev->smc_fw, fw_name, rdev->dev);
2195                         if (err) {
2196                                 printk(KERN_ERR
2197                                        "smc: error loading firmware \"%s\"\n",
2198                                        fw_name);
2199                                 release_firmware(rdev->smc_fw);
2200                                 rdev->smc_fw = NULL;
2201                                 err = 0;
2202                         } else if (rdev->smc_fw->size != smc_req_size) {
2203                                 printk(KERN_ERR
2204                                        "cik_smc: Bogus length %zu in firmware \"%s\"\n",
2205                                        rdev->smc_fw->size, fw_name);
2206                                 err = -EINVAL;
2207                         }
2208                 } else {
2209                         err = radeon_ucode_validate(rdev->smc_fw);
2210                         if (err) {
2211                                 printk(KERN_ERR
2212                                        "cik_fw: validation failed for firmware \"%s\"\n",
2213                                        fw_name);
2214                                 goto out;
2215                         } else {
2216                                 new_fw++;
2217                         }
2218                 }
2219         }
2220
2221         if (new_fw == 0) {
2222                 rdev->new_fw = false;
2223         } else if (new_fw < num_fw) {
2224                 printk(KERN_ERR "ci_fw: mixing new and old firmware!\n");
2225                 err = -EINVAL;
2226         } else {
2227                 rdev->new_fw = true;
2228         }
2229
2230 out:
2231         if (err) {
2232                 if (err != -EINVAL)
2233                         printk(KERN_ERR
2234                                "cik_cp: Failed to load firmware \"%s\"\n",
2235                                fw_name);
2236                 release_firmware(rdev->pfp_fw);
2237                 rdev->pfp_fw = NULL;
2238                 release_firmware(rdev->me_fw);
2239                 rdev->me_fw = NULL;
2240                 release_firmware(rdev->ce_fw);
2241                 rdev->ce_fw = NULL;
2242                 release_firmware(rdev->mec_fw);
2243                 rdev->mec_fw = NULL;
2244                 release_firmware(rdev->mec2_fw);
2245                 rdev->mec2_fw = NULL;
2246                 release_firmware(rdev->rlc_fw);
2247                 rdev->rlc_fw = NULL;
2248                 release_firmware(rdev->sdma_fw);
2249                 rdev->sdma_fw = NULL;
2250                 release_firmware(rdev->mc_fw);
2251                 rdev->mc_fw = NULL;
2252                 release_firmware(rdev->smc_fw);
2253                 rdev->smc_fw = NULL;
2254         }
2255         return err;
2256 }
2257
2258 /*
2259  * Core functions
2260  */
2261 /**
2262  * cik_tiling_mode_table_init - init the hw tiling table
2263  *
2264  * @rdev: radeon_device pointer
2265  *
2266  * Starting with SI, the tiling setup is done globally in a
2267  * set of 32 tiling modes.  Rather than selecting each set of
2268  * parameters per surface as on older asics, we just select
2269  * which index in the tiling table we want to use, and the
2270  * surface uses those parameters (CIK).
2271  */
2272 static void cik_tiling_mode_table_init(struct radeon_device *rdev)
2273 {
2274         const u32 num_tile_mode_states = 32;
2275         const u32 num_secondary_tile_mode_states = 16;
2276         u32 reg_offset, gb_tile_moden, split_equal_to_row_size;
2277         u32 num_pipe_configs;
2278         u32 num_rbs = rdev->config.cik.max_backends_per_se *
2279                 rdev->config.cik.max_shader_engines;
2280
2281         switch (rdev->config.cik.mem_row_size_in_kb) {
2282         case 1:
2283                 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_1KB;
2284                 break;
2285         case 2:
2286         default:
2287                 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_2KB;
2288                 break;
2289         case 4:
2290                 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_4KB;
2291                 break;
2292         }
2293
2294         num_pipe_configs = rdev->config.cik.max_tile_pipes;
2295         if (num_pipe_configs > 8)
2296                 num_pipe_configs = 16;
2297
2298         if (num_pipe_configs == 16) {
2299                 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
2300                         switch (reg_offset) {
2301                         case 0:
2302                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2303                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2304                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2305                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B));
2306                                 break;
2307                         case 1:
2308                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2309                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2310                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2311                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B));
2312                                 break;
2313                         case 2:
2314                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2315                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2316                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2317                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B));
2318                                 break;
2319                         case 3:
2320                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2321                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2322                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2323                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B));
2324                                 break;
2325                         case 4:
2326                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2327                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2328                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2329                                                  TILE_SPLIT(split_equal_to_row_size));
2330                                 break;
2331                         case 5:
2332                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2333                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2334                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
2335                                 break;
2336                         case 6:
2337                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2338                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2339                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2340                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B));
2341                                 break;
2342                         case 7:
2343                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2344                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2345                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2346                                                  TILE_SPLIT(split_equal_to_row_size));
2347                                 break;
2348                         case 8:
2349                                 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
2350                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16));
2351                                 break;
2352                         case 9:
2353                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2354                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2355                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING));
2356                                 break;
2357                         case 10:
2358                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2359                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
2360                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2361                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2362                                 break;
2363                         case 11:
2364                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
2365                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
2366                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_8x16) |
2367                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2368                                 break;
2369                         case 12:
2370                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2371                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
2372                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2373                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2374                                 break;
2375                         case 13:
2376                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2377                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2378                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING));
2379                                 break;
2380                         case 14:
2381                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2382                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
2383                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2384                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2385                                 break;
2386                         case 16:
2387                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
2388                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
2389                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_8x16) |
2390                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2391                                 break;
2392                         case 17:
2393                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2394                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
2395                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2396                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2397                                 break;
2398                         case 27:
2399                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2400                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2401                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING));
2402                                 break;
2403                         case 28:
2404                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2405                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
2406                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2407                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2408                                 break;
2409                         case 29:
2410                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
2411                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
2412                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_8x16) |
2413                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2414                                 break;
2415                         case 30:
2416                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2417                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
2418                                                  PIPE_CONFIG(ADDR_SURF_P16_32x32_16x16) |
2419                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2420                                 break;
2421                         default:
2422                                 gb_tile_moden = 0;
2423                                 break;
2424                         }
2425                         rdev->config.cik.tile_mode_array[reg_offset] = gb_tile_moden;
2426                         WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
2427                 }
2428                 for (reg_offset = 0; reg_offset < num_secondary_tile_mode_states; reg_offset++) {
2429                         switch (reg_offset) {
2430                         case 0:
2431                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2432                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2433                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
2434                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2435                                 break;
2436                         case 1:
2437                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2438                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2439                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
2440                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2441                                 break;
2442                         case 2:
2443                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2444                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2445                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2446                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2447                                 break;
2448                         case 3:
2449                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2450                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2451                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2452                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2453                                 break;
2454                         case 4:
2455                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2456                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2457                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2458                                                  NUM_BANKS(ADDR_SURF_8_BANK));
2459                                 break;
2460                         case 5:
2461                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2462                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2463                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2464                                                  NUM_BANKS(ADDR_SURF_4_BANK));
2465                                 break;
2466                         case 6:
2467                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2468                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2469                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2470                                                  NUM_BANKS(ADDR_SURF_2_BANK));
2471                                 break;
2472                         case 8:
2473                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2474                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2475                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
2476                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2477                                 break;
2478                         case 9:
2479                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2480                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2481                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
2482                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2483                                 break;
2484                         case 10:
2485                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2486                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2487                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2488                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2489                                 break;
2490                         case 11:
2491                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2492                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2493                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2494                                                  NUM_BANKS(ADDR_SURF_8_BANK));
2495                                 break;
2496                         case 12:
2497                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2498                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2499                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2500                                                  NUM_BANKS(ADDR_SURF_4_BANK));
2501                                 break;
2502                         case 13:
2503                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2504                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2505                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2506                                                  NUM_BANKS(ADDR_SURF_2_BANK));
2507                                 break;
2508                         case 14:
2509                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2510                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2511                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2512                                                  NUM_BANKS(ADDR_SURF_2_BANK));
2513                                 break;
2514                         default:
2515                                 gb_tile_moden = 0;
2516                                 break;
2517                         }
2518                         rdev->config.cik.macrotile_mode_array[reg_offset] = gb_tile_moden;
2519                         WREG32(GB_MACROTILE_MODE0 + (reg_offset * 4), gb_tile_moden);
2520                 }
2521         } else if (num_pipe_configs == 8) {
2522                 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
2523                         switch (reg_offset) {
2524                         case 0:
2525                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2526                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2527                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2528                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B));
2529                                 break;
2530                         case 1:
2531                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2532                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2533                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2534                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B));
2535                                 break;
2536                         case 2:
2537                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2538                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2539                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2540                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B));
2541                                 break;
2542                         case 3:
2543                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2544                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2545                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2546                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B));
2547                                 break;
2548                         case 4:
2549                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2550                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2551                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2552                                                  TILE_SPLIT(split_equal_to_row_size));
2553                                 break;
2554                         case 5:
2555                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2556                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2557                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
2558                                 break;
2559                         case 6:
2560                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2561                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2562                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2563                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B));
2564                                 break;
2565                         case 7:
2566                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2567                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2568                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2569                                                  TILE_SPLIT(split_equal_to_row_size));
2570                                 break;
2571                         case 8:
2572                                 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
2573                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16));
2574                                 break;
2575                         case 9:
2576                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2577                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2578                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING));
2579                                 break;
2580                         case 10:
2581                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2582                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
2583                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2584                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2585                                 break;
2586                         case 11:
2587                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
2588                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
2589                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2590                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2591                                 break;
2592                         case 12:
2593                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2594                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
2595                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2596                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2597                                 break;
2598                         case 13:
2599                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2600                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2601                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING));
2602                                 break;
2603                         case 14:
2604                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2605                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
2606                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2607                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2608                                 break;
2609                         case 16:
2610                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
2611                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
2612                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2613                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2614                                 break;
2615                         case 17:
2616                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2617                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
2618                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2619                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2620                                 break;
2621                         case 27:
2622                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2623                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2624                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING));
2625                                 break;
2626                         case 28:
2627                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2628                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
2629                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2630                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2631                                 break;
2632                         case 29:
2633                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
2634                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
2635                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
2636                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2637                                 break;
2638                         case 30:
2639                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2640                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
2641                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_16x16) |
2642                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2643                                 break;
2644                         default:
2645                                 gb_tile_moden = 0;
2646                                 break;
2647                         }
2648                         rdev->config.cik.tile_mode_array[reg_offset] = gb_tile_moden;
2649                         WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
2650                 }
2651                 for (reg_offset = 0; reg_offset < num_secondary_tile_mode_states; reg_offset++) {
2652                         switch (reg_offset) {
2653                         case 0:
2654                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2655                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2656                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
2657                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2658                                 break;
2659                         case 1:
2660                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2661                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2662                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
2663                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2664                                 break;
2665                         case 2:
2666                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2667                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2668                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
2669                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2670                                 break;
2671                         case 3:
2672                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2673                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2674                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
2675                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2676                                 break;
2677                         case 4:
2678                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2679                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2680                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2681                                                  NUM_BANKS(ADDR_SURF_8_BANK));
2682                                 break;
2683                         case 5:
2684                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2685                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2686                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2687                                                  NUM_BANKS(ADDR_SURF_4_BANK));
2688                                 break;
2689                         case 6:
2690                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2691                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2692                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2693                                                  NUM_BANKS(ADDR_SURF_2_BANK));
2694                                 break;
2695                         case 8:
2696                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2697                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_8) |
2698                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
2699                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2700                                 break;
2701                         case 9:
2702                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2703                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
2704                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
2705                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2706                                 break;
2707                         case 10:
2708                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2709                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
2710                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
2711                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2712                                 break;
2713                         case 11:
2714                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2715                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2716                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
2717                                                  NUM_BANKS(ADDR_SURF_16_BANK));
2718                                 break;
2719                         case 12:
2720                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2721                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2722                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2723                                                  NUM_BANKS(ADDR_SURF_8_BANK));
2724                                 break;
2725                         case 13:
2726                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2727                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2728                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2729                                                  NUM_BANKS(ADDR_SURF_4_BANK));
2730                                 break;
2731                         case 14:
2732                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
2733                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
2734                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
2735                                                  NUM_BANKS(ADDR_SURF_2_BANK));
2736                                 break;
2737                         default:
2738                                 gb_tile_moden = 0;
2739                                 break;
2740                         }
2741                         rdev->config.cik.macrotile_mode_array[reg_offset] = gb_tile_moden;
2742                         WREG32(GB_MACROTILE_MODE0 + (reg_offset * 4), gb_tile_moden);
2743                 }
2744         } else if (num_pipe_configs == 4) {
2745                 if (num_rbs == 4) {
2746                         for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
2747                                 switch (reg_offset) {
2748                                 case 0:
2749                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2750                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2751                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2752                                                          TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B));
2753                                         break;
2754                                 case 1:
2755                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2756                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2757                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2758                                                          TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B));
2759                                         break;
2760                                 case 2:
2761                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2762                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2763                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2764                                                          TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B));
2765                                         break;
2766                                 case 3:
2767                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2768                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2769                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2770                                                          TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B));
2771                                         break;
2772                                 case 4:
2773                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2774                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2775                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2776                                                          TILE_SPLIT(split_equal_to_row_size));
2777                                         break;
2778                                 case 5:
2779                                         gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2780                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2781                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
2782                                         break;
2783                                 case 6:
2784                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2785                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2786                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2787                                                          TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B));
2788                                         break;
2789                                 case 7:
2790                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2791                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2792                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2793                                                          TILE_SPLIT(split_equal_to_row_size));
2794                                         break;
2795                                 case 8:
2796                                         gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
2797                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16));
2798                                         break;
2799                                 case 9:
2800                                         gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2801                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2802                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING));
2803                                         break;
2804                                 case 10:
2805                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2806                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
2807                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2808                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2809                                         break;
2810                                 case 11:
2811                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
2812                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
2813                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2814                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2815                                         break;
2816                                 case 12:
2817                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2818                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
2819                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2820                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2821                                         break;
2822                                 case 13:
2823                                         gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2824                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2825                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING));
2826                                         break;
2827                                 case 14:
2828                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2829                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
2830                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2831                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2832                                         break;
2833                                 case 16:
2834                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
2835                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
2836                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2837                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2838                                         break;
2839                                 case 17:
2840                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2841                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
2842                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2843                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2844                                         break;
2845                                 case 27:
2846                                         gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2847                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2848                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING));
2849                                         break;
2850                                 case 28:
2851                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2852                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
2853                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2854                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2855                                         break;
2856                                 case 29:
2857                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
2858                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
2859                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2860                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2861                                         break;
2862                                 case 30:
2863                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2864                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
2865                                                          PIPE_CONFIG(ADDR_SURF_P4_16x16) |
2866                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2867                                         break;
2868                                 default:
2869                                         gb_tile_moden = 0;
2870                                         break;
2871                                 }
2872                                 rdev->config.cik.tile_mode_array[reg_offset] = gb_tile_moden;
2873                                 WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
2874                         }
2875                 } else if (num_rbs < 4) {
2876                         for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
2877                                 switch (reg_offset) {
2878                                 case 0:
2879                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2880                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2881                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2882                                                          TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B));
2883                                         break;
2884                                 case 1:
2885                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2886                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2887                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2888                                                          TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B));
2889                                         break;
2890                                 case 2:
2891                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2892                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2893                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2894                                                          TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B));
2895                                         break;
2896                                 case 3:
2897                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2898                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2899                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2900                                                          TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B));
2901                                         break;
2902                                 case 4:
2903                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2904                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2905                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2906                                                          TILE_SPLIT(split_equal_to_row_size));
2907                                         break;
2908                                 case 5:
2909                                         gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2910                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2911                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
2912                                         break;
2913                                 case 6:
2914                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2915                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2916                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2917                                                          TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B));
2918                                         break;
2919                                 case 7:
2920                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2921                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
2922                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2923                                                          TILE_SPLIT(split_equal_to_row_size));
2924                                         break;
2925                                 case 8:
2926                                         gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
2927                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16));
2928                                         break;
2929                                 case 9:
2930                                         gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2931                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2932                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING));
2933                                         break;
2934                                 case 10:
2935                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2936                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
2937                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2938                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2939                                         break;
2940                                 case 11:
2941                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
2942                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
2943                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2944                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2945                                         break;
2946                                 case 12:
2947                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2948                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
2949                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2950                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2951                                         break;
2952                                 case 13:
2953                                         gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2954                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2955                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING));
2956                                         break;
2957                                 case 14:
2958                                         gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
2959                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
2960                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2961                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2962                                         break;
2963                                 case 16:
2964                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
2965                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
2966                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2967                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2968                                         break;
2969                                 case 17:
2970                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2971                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
2972                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2973                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2974                                         break;
2975                                 case 27:
2976                                         gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
2977                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2978                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING));
2979                                         break;
2980                                 case 28:
2981                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2982                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
2983                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2984                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2985                                         break;
2986                                 case 29:
2987                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
2988                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
2989                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2990                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2991                                         break;
2992                                 case 30:
2993                                         gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
2994                                                          MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
2995                                                          PIPE_CONFIG(ADDR_SURF_P4_8x16) |
2996                                                          SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
2997                                         break;
2998                                 default:
2999                                         gb_tile_moden = 0;
3000                                         break;
3001                                 }
3002                                 rdev->config.cik.tile_mode_array[reg_offset] = gb_tile_moden;
3003                                 WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
3004                         }
3005                 }
3006                 for (reg_offset = 0; reg_offset < num_secondary_tile_mode_states; reg_offset++) {
3007                         switch (reg_offset) {
3008                         case 0:
3009                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3010                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
3011                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3012                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3013                                 break;
3014                         case 1:
3015                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3016                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
3017                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3018                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3019                                 break;
3020                         case 2:
3021                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3022                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3023                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
3024                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3025                                 break;
3026                         case 3:
3027                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3028                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3029                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
3030                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3031                                 break;
3032                         case 4:
3033                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3034                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3035                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
3036                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3037                                 break;
3038                         case 5:
3039                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3040                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3041                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
3042                                                  NUM_BANKS(ADDR_SURF_8_BANK));
3043                                 break;
3044                         case 6:
3045                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3046                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3047                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
3048                                                  NUM_BANKS(ADDR_SURF_4_BANK));
3049                                 break;
3050                         case 8:
3051                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
3052                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_8) |
3053                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3054                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3055                                 break;
3056                         case 9:
3057                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
3058                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
3059                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3060                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3061                                 break;
3062                         case 10:
3063                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3064                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
3065                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3066                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3067                                 break;
3068                         case 11:
3069                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3070                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
3071                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3072                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3073                                 break;
3074                         case 12:
3075                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3076                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3077                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
3078                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3079                                 break;
3080                         case 13:
3081                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3082                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3083                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
3084                                                  NUM_BANKS(ADDR_SURF_8_BANK));
3085                                 break;
3086                         case 14:
3087                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3088                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3089                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1) |
3090                                                  NUM_BANKS(ADDR_SURF_4_BANK));
3091                                 break;
3092                         default:
3093                                 gb_tile_moden = 0;
3094                                 break;
3095                         }
3096                         rdev->config.cik.macrotile_mode_array[reg_offset] = gb_tile_moden;
3097                         WREG32(GB_MACROTILE_MODE0 + (reg_offset * 4), gb_tile_moden);
3098                 }
3099         } else if (num_pipe_configs == 2) {
3100                 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
3101                         switch (reg_offset) {
3102                         case 0:
3103                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
3104                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
3105                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3106                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B));
3107                                 break;
3108                         case 1:
3109                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
3110                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
3111                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3112                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B));
3113                                 break;
3114                         case 2:
3115                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
3116                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
3117                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3118                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B));
3119                                 break;
3120                         case 3:
3121                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
3122                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
3123                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3124                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B));
3125                                 break;
3126                         case 4:
3127                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
3128                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
3129                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3130                                                  TILE_SPLIT(split_equal_to_row_size));
3131                                 break;
3132                         case 5:
3133                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
3134                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3135                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING));
3136                                 break;
3137                         case 6:
3138                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
3139                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
3140                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3141                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B));
3142                                 break;
3143                         case 7:
3144                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
3145                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DEPTH_MICRO_TILING) |
3146                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3147                                                  TILE_SPLIT(split_equal_to_row_size));
3148                                 break;
3149                         case 8:
3150                                 gb_tile_moden = ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
3151                                                 PIPE_CONFIG(ADDR_SURF_P2);
3152                                 break;
3153                         case 9:
3154                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
3155                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
3156                                                  PIPE_CONFIG(ADDR_SURF_P2));
3157                                 break;
3158                         case 10:
3159                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
3160                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
3161                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3162                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
3163                                 break;
3164                         case 11:
3165                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
3166                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
3167                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3168                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
3169                                 break;
3170                         case 12:
3171                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
3172                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_DISPLAY_MICRO_TILING) |
3173                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3174                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
3175                                 break;
3176                         case 13:
3177                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
3178                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3179                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING));
3180                                 break;
3181                         case 14:
3182                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
3183                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
3184                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3185                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
3186                                 break;
3187                         case 16:
3188                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
3189                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
3190                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3191                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
3192                                 break;
3193                         case 17:
3194                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
3195                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_THIN_MICRO_TILING) |
3196                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3197                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
3198                                 break;
3199                         case 27:
3200                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
3201                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
3202                                                  PIPE_CONFIG(ADDR_SURF_P2));
3203                                 break;
3204                         case 28:
3205                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
3206                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
3207                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3208                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
3209                                 break;
3210                         case 29:
3211                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_TILED_THIN1) |
3212                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
3213                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3214                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
3215                                 break;
3216                         case 30:
3217                                 gb_tile_moden = (ARRAY_MODE(ARRAY_PRT_2D_TILED_THIN1) |
3218                                                  MICRO_TILE_MODE_NEW(ADDR_SURF_ROTATED_MICRO_TILING) |
3219                                                  PIPE_CONFIG(ADDR_SURF_P2) |
3220                                                  SAMPLE_SPLIT(ADDR_SURF_SAMPLE_SPLIT_2));
3221                                 break;
3222                         default:
3223                                 gb_tile_moden = 0;
3224                                 break;
3225                         }
3226                         rdev->config.cik.tile_mode_array[reg_offset] = gb_tile_moden;
3227                         WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
3228                 }
3229                 for (reg_offset = 0; reg_offset < num_secondary_tile_mode_states; reg_offset++) {
3230                         switch (reg_offset) {
3231                         case 0:
3232                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
3233                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
3234                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3235                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3236                                 break;
3237                         case 1:
3238                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
3239                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
3240                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3241                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3242                                 break;
3243                         case 2:
3244                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3245                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
3246                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3247                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3248                                 break;
3249                         case 3:
3250                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3251                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3252                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3253                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3254                                 break;
3255                         case 4:
3256                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3257                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3258                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3259                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3260                                 break;
3261                         case 5:
3262                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3263                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3264                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3265                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3266                                 break;
3267                         case 6:
3268                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3269                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3270                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
3271                                                  NUM_BANKS(ADDR_SURF_8_BANK));
3272                                 break;
3273                         case 8:
3274                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_4) |
3275                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_8) |
3276                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3277                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3278                                 break;
3279                         case 9:
3280                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_4) |
3281                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
3282                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3283                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3284                                 break;
3285                         case 10:
3286                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
3287                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
3288                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3289                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3290                                 break;
3291                         case 11:
3292                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
3293                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
3294                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3295                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3296                                 break;
3297                         case 12:
3298                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3299                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
3300                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3301                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3302                                 break;
3303                         case 13:
3304                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3305                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3306                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4) |
3307                                                  NUM_BANKS(ADDR_SURF_16_BANK));
3308                                 break;
3309                         case 14:
3310                                 gb_tile_moden = (BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
3311                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
3312                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2) |
3313                                                  NUM_BANKS(ADDR_SURF_8_BANK));
3314                                 break;
3315                         default:
3316                                 gb_tile_moden = 0;
3317                                 break;
3318                         }
3319                         rdev->config.cik.macrotile_mode_array[reg_offset] = gb_tile_moden;
3320                         WREG32(GB_MACROTILE_MODE0 + (reg_offset * 4), gb_tile_moden);
3321                 }
3322         } else
3323                 DRM_ERROR("unknown num pipe config: 0x%x\n", num_pipe_configs);
3324 }
3325
3326 /**
3327  * cik_select_se_sh - select which SE, SH to address
3328  *
3329  * @rdev: radeon_device pointer
3330  * @se_num: shader engine to address
3331  * @sh_num: sh block to address
3332  *
3333  * Select which SE, SH combinations to address. Certain
3334  * registers are instanced per SE or SH.  0xffffffff means
3335  * broadcast to all SEs or SHs (CIK).
3336  */
3337 static void cik_select_se_sh(struct radeon_device *rdev,
3338                              u32 se_num, u32 sh_num)
3339 {
3340         u32 data = INSTANCE_BROADCAST_WRITES;
3341
3342         if ((se_num == 0xffffffff) && (sh_num == 0xffffffff))
3343                 data |= SH_BROADCAST_WRITES | SE_BROADCAST_WRITES;
3344         else if (se_num == 0xffffffff)
3345                 data |= SE_BROADCAST_WRITES | SH_INDEX(sh_num);
3346         else if (sh_num == 0xffffffff)
3347                 data |= SH_BROADCAST_WRITES | SE_INDEX(se_num);
3348         else
3349                 data |= SH_INDEX(sh_num) | SE_INDEX(se_num);
3350         WREG32(GRBM_GFX_INDEX, data);
3351 }
3352
3353 /**
3354  * cik_create_bitmask - create a bitmask
3355  *
3356  * @bit_width: length of the mask
3357  *
3358  * create a variable length bit mask (CIK).
3359  * Returns the bitmask.
3360  */
3361 static u32 cik_create_bitmask(u32 bit_width)
3362 {
3363         u32 i, mask = 0;
3364
3365         for (i = 0; i < bit_width; i++) {
3366                 mask <<= 1;
3367                 mask |= 1;
3368         }
3369         return mask;
3370 }
3371
3372 /**
3373  * cik_get_rb_disabled - computes the mask of disabled RBs
3374  *
3375  * @rdev: radeon_device pointer
3376  * @max_rb_num: max RBs (render backends) for the asic
3377  * @se_num: number of SEs (shader engines) for the asic
3378  * @sh_per_se: number of SH blocks per SE for the asic
3379  *
3380  * Calculates the bitmask of disabled RBs (CIK).
3381  * Returns the disabled RB bitmask.
3382  */
3383 static u32 cik_get_rb_disabled(struct radeon_device *rdev,
3384                               u32 max_rb_num_per_se,
3385                               u32 sh_per_se)
3386 {
3387         u32 data, mask;
3388
3389         data = RREG32(CC_RB_BACKEND_DISABLE);
3390         if (data & 1)
3391                 data &= BACKEND_DISABLE_MASK;
3392         else
3393                 data = 0;
3394         data |= RREG32(GC_USER_RB_BACKEND_DISABLE);
3395
3396         data >>= BACKEND_DISABLE_SHIFT;
3397
3398         mask = cik_create_bitmask(max_rb_num_per_se / sh_per_se);
3399
3400         return data & mask;
3401 }
3402
3403 /**
3404  * cik_setup_rb - setup the RBs on the asic
3405  *
3406  * @rdev: radeon_device pointer
3407  * @se_num: number of SEs (shader engines) for the asic
3408  * @sh_per_se: number of SH blocks per SE for the asic
3409  * @max_rb_num: max RBs (render backends) for the asic
3410  *
3411  * Configures per-SE/SH RB registers (CIK).
3412  */
3413 static void cik_setup_rb(struct radeon_device *rdev,
3414                          u32 se_num, u32 sh_per_se,
3415                          u32 max_rb_num_per_se)
3416 {
3417         int i, j;
3418         u32 data, mask;
3419         u32 disabled_rbs = 0;
3420         u32 enabled_rbs = 0;
3421
3422         for (i = 0; i < se_num; i++) {
3423                 for (j = 0; j < sh_per_se; j++) {
3424                         cik_select_se_sh(rdev, i, j);
3425                         data = cik_get_rb_disabled(rdev, max_rb_num_per_se, sh_per_se);
3426                         if (rdev->family == CHIP_HAWAII)
3427                                 disabled_rbs |= data << ((i * sh_per_se + j) * HAWAII_RB_BITMAP_WIDTH_PER_SH);
3428                         else
3429                                 disabled_rbs |= data << ((i * sh_per_se + j) * CIK_RB_BITMAP_WIDTH_PER_SH);
3430                 }
3431         }
3432         cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
3433
3434         mask = 1;
3435         for (i = 0; i < max_rb_num_per_se * se_num; i++) {
3436                 if (!(disabled_rbs & mask))
3437                         enabled_rbs |= mask;
3438                 mask <<= 1;
3439         }
3440
3441         rdev->config.cik.backend_enable_mask = enabled_rbs;
3442
3443         for (i = 0; i < se_num; i++) {
3444                 cik_select_se_sh(rdev, i, 0xffffffff);
3445                 data = 0;
3446                 for (j = 0; j < sh_per_se; j++) {
3447                         switch (enabled_rbs & 3) {
3448                         case 0:
3449                                 if (j == 0)
3450                                         data |= PKR_MAP(RASTER_CONFIG_RB_MAP_3);
3451                                 else
3452                                         data |= PKR_MAP(RASTER_CONFIG_RB_MAP_0);
3453                                 break;
3454                         case 1:
3455                                 data |= (RASTER_CONFIG_RB_MAP_0 << (i * sh_per_se + j) * 2);
3456                                 break;
3457                         case 2:
3458                                 data |= (RASTER_CONFIG_RB_MAP_3 << (i * sh_per_se + j) * 2);
3459                                 break;
3460                         case 3:
3461                         default:
3462                                 data |= (RASTER_CONFIG_RB_MAP_2 << (i * sh_per_se + j) * 2);
3463                                 break;
3464                         }
3465                         enabled_rbs >>= 2;
3466                 }
3467                 WREG32(PA_SC_RASTER_CONFIG, data);
3468         }
3469         cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
3470 }
3471
3472 /**
3473  * cik_gpu_init - setup the 3D engine
3474  *
3475  * @rdev: radeon_device pointer
3476  *
3477  * Configures the 3D engine and tiling configuration
3478  * registers so that the 3D engine is usable.
3479  */
3480 static void cik_gpu_init(struct radeon_device *rdev)
3481 {
3482         u32 gb_addr_config = RREG32(GB_ADDR_CONFIG);
3483         u32 mc_shared_chmap, mc_arb_ramcfg;
3484         u32 hdp_host_path_cntl;
3485         u32 tmp;
3486         int i, j;
3487
3488         switch (rdev->family) {
3489         case CHIP_BONAIRE:
3490                 rdev->config.cik.max_shader_engines = 2;
3491                 rdev->config.cik.max_tile_pipes = 4;
3492                 rdev->config.cik.max_cu_per_sh = 7;
3493                 rdev->config.cik.max_sh_per_se = 1;
3494                 rdev->config.cik.max_backends_per_se = 2;
3495                 rdev->config.cik.max_texture_channel_caches = 4;
3496                 rdev->config.cik.max_gprs = 256;
3497                 rdev->config.cik.max_gs_threads = 32;
3498                 rdev->config.cik.max_hw_contexts = 8;
3499
3500                 rdev->config.cik.sc_prim_fifo_size_frontend = 0x20;
3501                 rdev->config.cik.sc_prim_fifo_size_backend = 0x100;
3502                 rdev->config.cik.sc_hiz_tile_fifo_size = 0x30;
3503                 rdev->config.cik.sc_earlyz_tile_fifo_size = 0x130;
3504                 gb_addr_config = BONAIRE_GB_ADDR_CONFIG_GOLDEN;
3505                 break;
3506         case CHIP_HAWAII:
3507                 rdev->config.cik.max_shader_engines = 4;
3508                 rdev->config.cik.max_tile_pipes = 16;
3509                 rdev->config.cik.max_cu_per_sh = 11;
3510                 rdev->config.cik.max_sh_per_se = 1;
3511                 rdev->config.cik.max_backends_per_se = 4;
3512                 rdev->config.cik.max_texture_channel_caches = 16;
3513                 rdev->config.cik.max_gprs = 256;
3514                 rdev->config.cik.max_gs_threads = 32;
3515                 rdev->config.cik.max_hw_contexts = 8;
3516
3517                 rdev->config.cik.sc_prim_fifo_size_frontend = 0x20;
3518                 rdev->config.cik.sc_prim_fifo_size_backend = 0x100;
3519                 rdev->config.cik.sc_hiz_tile_fifo_size = 0x30;
3520                 rdev->config.cik.sc_earlyz_tile_fifo_size = 0x130;
3521                 gb_addr_config = HAWAII_GB_ADDR_CONFIG_GOLDEN;
3522                 break;
3523         case CHIP_KAVERI:
3524                 rdev->config.cik.max_shader_engines = 1;
3525                 rdev->config.cik.max_tile_pipes = 4;
3526                 if ((rdev->pdev->device == 0x1304) ||
3527                     (rdev->pdev->device == 0x1305) ||
3528                     (rdev->pdev->device == 0x130C) ||
3529                     (rdev->pdev->device == 0x130F) ||
3530                     (rdev->pdev->device == 0x1310) ||
3531                     (rdev->pdev->device == 0x1311) ||
3532                     (rdev->pdev->device == 0x131C)) {
3533                         rdev->config.cik.max_cu_per_sh = 8;
3534                         rdev->config.cik.max_backends_per_se = 2;
3535                 } else if ((rdev->pdev->device == 0x1309) ||
3536                            (rdev->pdev->device == 0x130A) ||
3537                            (rdev->pdev->device == 0x130D) ||
3538                            (rdev->pdev->device == 0x1313) ||
3539                            (rdev->pdev->device == 0x131D)) {
3540                         rdev->config.cik.max_cu_per_sh = 6;
3541                         rdev->config.cik.max_backends_per_se = 2;
3542                 } else if ((rdev->pdev->device == 0x1306) ||
3543                            (rdev->pdev->device == 0x1307) ||
3544                            (rdev->pdev->device == 0x130B) ||
3545                            (rdev->pdev->device == 0x130E) ||
3546                            (rdev->pdev->device == 0x1315) ||
3547                            (rdev->pdev->device == 0x1318) ||
3548                            (rdev->pdev->device == 0x131B)) {
3549                         rdev->config.cik.max_cu_per_sh = 4;
3550                         rdev->config.cik.max_backends_per_se = 1;
3551                 } else {
3552                         rdev->config.cik.max_cu_per_sh = 3;
3553                         rdev->config.cik.max_backends_per_se = 1;
3554                 }
3555                 rdev->config.cik.max_sh_per_se = 1;
3556                 rdev->config.cik.max_texture_channel_caches = 4;
3557                 rdev->config.cik.max_gprs = 256;
3558                 rdev->config.cik.max_gs_threads = 16;
3559                 rdev->config.cik.max_hw_contexts = 8;
3560
3561                 rdev->config.cik.sc_prim_fifo_size_frontend = 0x20;
3562                 rdev->config.cik.sc_prim_fifo_size_backend = 0x100;
3563                 rdev->config.cik.sc_hiz_tile_fifo_size = 0x30;
3564                 rdev->config.cik.sc_earlyz_tile_fifo_size = 0x130;
3565                 gb_addr_config = BONAIRE_GB_ADDR_CONFIG_GOLDEN;
3566                 break;
3567         case CHIP_KABINI:
3568         case CHIP_MULLINS:
3569         default:
3570                 rdev->config.cik.max_shader_engines = 1;
3571                 rdev->config.cik.max_tile_pipes = 2;
3572                 rdev->config.cik.max_cu_per_sh = 2;
3573                 rdev->config.cik.max_sh_per_se = 1;
3574                 rdev->config.cik.max_backends_per_se = 1;
3575                 rdev->config.cik.max_texture_channel_caches = 2;
3576                 rdev->config.cik.max_gprs = 256;
3577                 rdev->config.cik.max_gs_threads = 16;
3578                 rdev->config.cik.max_hw_contexts = 8;
3579
3580                 rdev->config.cik.sc_prim_fifo_size_frontend = 0x20;
3581                 rdev->config.cik.sc_prim_fifo_size_backend = 0x100;
3582                 rdev->config.cik.sc_hiz_tile_fifo_size = 0x30;
3583                 rdev->config.cik.sc_earlyz_tile_fifo_size = 0x130;
3584                 gb_addr_config = BONAIRE_GB_ADDR_CONFIG_GOLDEN;
3585                 break;
3586         }
3587
3588         /* Initialize HDP */
3589         for (i = 0, j = 0; i < 32; i++, j += 0x18) {
3590                 WREG32((0x2c14 + j), 0x00000000);
3591                 WREG32((0x2c18 + j), 0x00000000);
3592                 WREG32((0x2c1c + j), 0x00000000);
3593                 WREG32((0x2c20 + j), 0x00000000);
3594                 WREG32((0x2c24 + j), 0x00000000);
3595         }
3596
3597         WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
3598
3599         WREG32(BIF_FB_EN, FB_READ_EN | FB_WRITE_EN);
3600
3601         mc_shared_chmap = RREG32(MC_SHARED_CHMAP);
3602         mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
3603
3604         rdev->config.cik.num_tile_pipes = rdev->config.cik.max_tile_pipes;
3605         rdev->config.cik.mem_max_burst_length_bytes = 256;
3606         tmp = (mc_arb_ramcfg & NOOFCOLS_MASK) >> NOOFCOLS_SHIFT;
3607         rdev->config.cik.mem_row_size_in_kb = (4 * (1 << (8 + tmp))) / 1024;
3608         if (rdev->config.cik.mem_row_size_in_kb > 4)
3609                 rdev->config.cik.mem_row_size_in_kb = 4;
3610         /* XXX use MC settings? */
3611         rdev->config.cik.shader_engine_tile_size = 32;
3612         rdev->config.cik.num_gpus = 1;
3613         rdev->config.cik.multi_gpu_tile_size = 64;
3614
3615         /* fix up row size */
3616         gb_addr_config &= ~ROW_SIZE_MASK;
3617         switch (rdev->config.cik.mem_row_size_in_kb) {
3618         case 1:
3619         default:
3620                 gb_addr_config |= ROW_SIZE(0);
3621                 break;
3622         case 2:
3623                 gb_addr_config |= ROW_SIZE(1);
3624                 break;
3625         case 4:
3626                 gb_addr_config |= ROW_SIZE(2);
3627                 break;
3628         }
3629
3630         /* setup tiling info dword.  gb_addr_config is not adequate since it does
3631          * not have bank info, so create a custom tiling dword.
3632          * bits 3:0   num_pipes
3633          * bits 7:4   num_banks
3634          * bits 11:8  group_size
3635          * bits 15:12 row_size
3636          */
3637         rdev->config.cik.tile_config = 0;
3638         switch (rdev->config.cik.num_tile_pipes) {
3639         case 1:
3640                 rdev->config.cik.tile_config |= (0 << 0);
3641                 break;
3642         case 2:
3643                 rdev->config.cik.tile_config |= (1 << 0);
3644                 break;
3645         case 4:
3646                 rdev->config.cik.tile_config |= (2 << 0);
3647                 break;
3648         case 8:
3649         default:
3650                 /* XXX what about 12? */
3651                 rdev->config.cik.tile_config |= (3 << 0);
3652                 break;
3653         }
3654         rdev->config.cik.tile_config |=
3655                 ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT) << 4;
3656         rdev->config.cik.tile_config |=
3657                 ((gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >> PIPE_INTERLEAVE_SIZE_SHIFT) << 8;
3658         rdev->config.cik.tile_config |=
3659                 ((gb_addr_config & ROW_SIZE_MASK) >> ROW_SIZE_SHIFT) << 12;
3660
3661         WREG32(GB_ADDR_CONFIG, gb_addr_config);
3662         WREG32(HDP_ADDR_CONFIG, gb_addr_config);
3663         WREG32(DMIF_ADDR_CALC, gb_addr_config);
3664         WREG32(SDMA0_TILING_CONFIG + SDMA0_REGISTER_OFFSET, gb_addr_config & 0x70);
3665         WREG32(SDMA0_TILING_CONFIG + SDMA1_REGISTER_OFFSET, gb_addr_config & 0x70);
3666         WREG32(UVD_UDEC_ADDR_CONFIG, gb_addr_config);
3667         WREG32(UVD_UDEC_DB_ADDR_CONFIG, gb_addr_config);
3668         WREG32(UVD_UDEC_DBW_ADDR_CONFIG, gb_addr_config);
3669
3670         cik_tiling_mode_table_init(rdev);
3671
3672         cik_setup_rb(rdev, rdev->config.cik.max_shader_engines,
3673                      rdev->config.cik.max_sh_per_se,
3674                      rdev->config.cik.max_backends_per_se);
3675
3676         rdev->config.cik.active_cus = 0;
3677         for (i = 0; i < rdev->config.cik.max_shader_engines; i++) {
3678                 for (j = 0; j < rdev->config.cik.max_sh_per_se; j++) {
3679                         rdev->config.cik.active_cus +=
3680                                 hweight32(cik_get_cu_active_bitmap(rdev, i, j));
3681                 }
3682         }
3683
3684         /* set HW defaults for 3D engine */
3685         WREG32(CP_MEQ_THRESHOLDS, MEQ1_START(0x30) | MEQ2_START(0x60));
3686
3687         WREG32(SX_DEBUG_1, 0x20);
3688
3689         WREG32(TA_CNTL_AUX, 0x00010000);
3690
3691         tmp = RREG32(SPI_CONFIG_CNTL);
3692         tmp |= 0x03000000;
3693         WREG32(SPI_CONFIG_CNTL, tmp);
3694
3695         WREG32(SQ_CONFIG, 1);
3696
3697         WREG32(DB_DEBUG, 0);
3698
3699         tmp = RREG32(DB_DEBUG2) & ~0xf00fffff;
3700         tmp |= 0x00000400;
3701         WREG32(DB_DEBUG2, tmp);
3702
3703         tmp = RREG32(DB_DEBUG3) & ~0x0002021c;
3704         tmp |= 0x00020200;
3705         WREG32(DB_DEBUG3, tmp);
3706
3707         tmp = RREG32(CB_HW_CONTROL) & ~0x00010000;
3708         tmp |= 0x00018208;
3709         WREG32(CB_HW_CONTROL, tmp);
3710
3711         WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4));
3712
3713         WREG32(PA_SC_FIFO_SIZE, (SC_FRONTEND_PRIM_FIFO_SIZE(rdev->config.cik.sc_prim_fifo_size_frontend) |
3714                                  SC_BACKEND_PRIM_FIFO_SIZE(rdev->config.cik.sc_prim_fifo_size_backend) |
3715                                  SC_HIZ_TILE_FIFO_SIZE(rdev->config.cik.sc_hiz_tile_fifo_size) |
3716                                  SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.cik.sc_earlyz_tile_fifo_size)));
3717
3718         WREG32(VGT_NUM_INSTANCES, 1);
3719
3720         WREG32(CP_PERFMON_CNTL, 0);
3721
3722         WREG32(SQ_CONFIG, 0);
3723
3724         WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
3725                                           FORCE_EOV_MAX_REZ_CNT(255)));
3726
3727         WREG32(VGT_CACHE_INVALIDATION, CACHE_INVALIDATION(VC_AND_TC) |
3728                AUTO_INVLD_EN(ES_AND_GS_AUTO));
3729
3730         WREG32(VGT_GS_VERTEX_REUSE, 16);
3731         WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
3732
3733         tmp = RREG32(HDP_MISC_CNTL);
3734         tmp |= HDP_FLUSH_INVALIDATE_CACHE;
3735         WREG32(HDP_MISC_CNTL, tmp);
3736
3737         hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
3738         WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
3739
3740         WREG32(PA_CL_ENHANCE, CLIP_VTX_REORDER_ENA | NUM_CLIP_SEQ(3));
3741         WREG32(PA_SC_ENHANCE, ENABLE_PA_SC_OUT_OF_ORDER);
3742
3743         udelay(50);
3744 }
3745
3746 /*
3747  * GPU scratch registers helpers function.
3748  */
3749 /**
3750  * cik_scratch_init - setup driver info for CP scratch regs
3751  *
3752  * @rdev: radeon_device pointer
3753  *
3754  * Set up the number and offset of the CP scratch registers.
3755  * NOTE: use of CP scratch registers is a legacy inferface and
3756  * is not used by default on newer asics (r6xx+).  On newer asics,
3757  * memory buffers are used for fences rather than scratch regs.
3758  */
3759 static void cik_scratch_init(struct radeon_device *rdev)
3760 {
3761         int i;
3762
3763         rdev->scratch.num_reg = 7;
3764         rdev->scratch.reg_base = SCRATCH_REG0;
3765         for (i = 0; i < rdev->scratch.num_reg; i++) {
3766                 rdev->scratch.free[i] = true;
3767                 rdev->scratch.reg[i] = rdev->scratch.reg_base + (i * 4);
3768         }
3769 }
3770
3771 /**
3772  * cik_ring_test - basic gfx ring test
3773  *
3774  * @rdev: radeon_device pointer
3775  * @ring: radeon_ring structure holding ring information
3776  *
3777  * Allocate a scratch register and write to it using the gfx ring (CIK).
3778  * Provides a basic gfx ring test to verify that the ring is working.
3779  * Used by cik_cp_gfx_resume();
3780  * Returns 0 on success, error on failure.
3781  */
3782 int cik_ring_test(struct radeon_device *rdev, struct radeon_ring *ring)
3783 {
3784         uint32_t scratch;
3785         uint32_t tmp = 0;
3786         unsigned i;
3787         int r;
3788
3789         r = radeon_scratch_get(rdev, &scratch);
3790         if (r) {
3791                 DRM_ERROR("radeon: cp failed to get scratch reg (%d).\n", r);
3792                 return r;
3793         }
3794         WREG32(scratch, 0xCAFEDEAD);
3795         r = radeon_ring_lock(rdev, ring, 3);
3796         if (r) {
3797                 DRM_ERROR("radeon: cp failed to lock ring %d (%d).\n", ring->idx, r);
3798                 radeon_scratch_free(rdev, scratch);
3799                 return r;
3800         }
3801         radeon_ring_write(ring, PACKET3(PACKET3_SET_UCONFIG_REG, 1));
3802         radeon_ring_write(ring, ((scratch - PACKET3_SET_UCONFIG_REG_START) >> 2));
3803         radeon_ring_write(ring, 0xDEADBEEF);
3804         radeon_ring_unlock_commit(rdev, ring, false);
3805
3806         for (i = 0; i < rdev->usec_timeout; i++) {
3807                 tmp = RREG32(scratch);
3808                 if (tmp == 0xDEADBEEF)
3809                         break;
3810                 DRM_UDELAY(1);
3811         }
3812         if (i < rdev->usec_timeout) {
3813                 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
3814         } else {
3815                 DRM_ERROR("radeon: ring %d test failed (scratch(0x%04X)=0x%08X)\n",
3816                           ring->idx, scratch, tmp);
3817                 r = -EINVAL;
3818         }
3819         radeon_scratch_free(rdev, scratch);
3820         return r;
3821 }
3822
3823 /**
3824  * cik_hdp_flush_cp_ring_emit - emit an hdp flush on the cp
3825  *
3826  * @rdev: radeon_device pointer
3827  * @ridx: radeon ring index
3828  *
3829  * Emits an hdp flush on the cp.
3830  */
3831 static void cik_hdp_flush_cp_ring_emit(struct radeon_device *rdev,
3832                                        int ridx)
3833 {
3834         struct radeon_ring *ring = &rdev->ring[ridx];
3835         u32 ref_and_mask;
3836
3837         switch (ring->idx) {
3838         case CAYMAN_RING_TYPE_CP1_INDEX:
3839         case CAYMAN_RING_TYPE_CP2_INDEX:
3840         default:
3841                 switch (ring->me) {
3842                 case 0:
3843                         ref_and_mask = CP2 << ring->pipe;
3844                         break;
3845                 case 1:
3846                         ref_and_mask = CP6 << ring->pipe;
3847                         break;
3848                 default:
3849                         return;
3850                 }
3851                 break;
3852         case RADEON_RING_TYPE_GFX_INDEX:
3853                 ref_and_mask = CP0;
3854                 break;
3855         }
3856
3857         radeon_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5));
3858         radeon_ring_write(ring, (WAIT_REG_MEM_OPERATION(1) | /* write, wait, write */
3859                                  WAIT_REG_MEM_FUNCTION(3) |  /* == */
3860                                  WAIT_REG_MEM_ENGINE(1)));   /* pfp */
3861         radeon_ring_write(ring, GPU_HDP_FLUSH_REQ >> 2);
3862         radeon_ring_write(ring, GPU_HDP_FLUSH_DONE >> 2);
3863         radeon_ring_write(ring, ref_and_mask);
3864         radeon_ring_write(ring, ref_and_mask);
3865         radeon_ring_write(ring, 0x20); /* poll interval */
3866 }
3867
3868 /**
3869  * cik_fence_gfx_ring_emit - emit a fence on the gfx ring
3870  *
3871  * @rdev: radeon_device pointer
3872  * @fence: radeon fence object
3873  *
3874  * Emits a fence sequnce number on the gfx ring and flushes
3875  * GPU caches.
3876  */
3877 void cik_fence_gfx_ring_emit(struct radeon_device *rdev,
3878                              struct radeon_fence *fence)
3879 {
3880         struct radeon_ring *ring = &rdev->ring[fence->ring];
3881         u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
3882
3883         /* EVENT_WRITE_EOP - flush caches, send int */
3884         radeon_ring_write(ring, PACKET3(PACKET3_EVENT_WRITE_EOP, 4));
3885         radeon_ring_write(ring, (EOP_TCL1_ACTION_EN |
3886                                  EOP_TC_ACTION_EN |
3887                                  EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) |
3888                                  EVENT_INDEX(5)));
3889         radeon_ring_write(ring, addr & 0xfffffffc);
3890         radeon_ring_write(ring, (upper_32_bits(addr) & 0xffff) | DATA_SEL(1) | INT_SEL(2));
3891         radeon_ring_write(ring, fence->seq);
3892         radeon_ring_write(ring, 0);
3893 }
3894
3895 /**
3896  * cik_fence_compute_ring_emit - emit a fence on the compute ring
3897  *
3898  * @rdev: radeon_device pointer
3899  * @fence: radeon fence object
3900  *
3901  * Emits a fence sequnce number on the compute ring and flushes
3902  * GPU caches.
3903  */
3904 void cik_fence_compute_ring_emit(struct radeon_device *rdev,
3905                                  struct radeon_fence *fence)
3906 {
3907         struct radeon_ring *ring = &rdev->ring[fence->ring];
3908         u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
3909
3910         /* RELEASE_MEM - flush caches, send int */
3911         radeon_ring_write(ring, PACKET3(PACKET3_RELEASE_MEM, 5));
3912         radeon_ring_write(ring, (EOP_TCL1_ACTION_EN |
3913                                  EOP_TC_ACTION_EN |
3914                                  EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) |
3915                                  EVENT_INDEX(5)));
3916         radeon_ring_write(ring, DATA_SEL(1) | INT_SEL(2));
3917         radeon_ring_write(ring, addr & 0xfffffffc);
3918         radeon_ring_write(ring, upper_32_bits(addr));
3919         radeon_ring_write(ring, fence->seq);
3920         radeon_ring_write(ring, 0);
3921 }
3922
3923 /**
3924  * cik_semaphore_ring_emit - emit a semaphore on the CP ring
3925  *
3926  * @rdev: radeon_device pointer
3927  * @ring: radeon ring buffer object
3928  * @semaphore: radeon semaphore object
3929  * @emit_wait: Is this a sempahore wait?
3930  *
3931  * Emits a semaphore signal/wait packet to the CP ring and prevents the PFP
3932  * from running ahead of semaphore waits.
3933  */
3934 bool cik_semaphore_ring_emit(struct radeon_device *rdev,
3935                              struct radeon_ring *ring,
3936                              struct radeon_semaphore *semaphore,
3937                              bool emit_wait)
3938 {
3939         uint64_t addr = semaphore->gpu_addr;
3940         unsigned sel = emit_wait ? PACKET3_SEM_SEL_WAIT : PACKET3_SEM_SEL_SIGNAL;
3941
3942         radeon_ring_write(ring, PACKET3(PACKET3_MEM_SEMAPHORE, 1));
3943         radeon_ring_write(ring, lower_32_bits(addr));
3944         radeon_ring_write(ring, (upper_32_bits(addr) & 0xffff) | sel);
3945
3946         if (emit_wait && ring->idx == RADEON_RING_TYPE_GFX_INDEX) {
3947                 /* Prevent the PFP from running ahead of the semaphore wait */
3948                 radeon_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
3949                 radeon_ring_write(ring, 0x0);
3950         }
3951
3952         return true;
3953 }
3954
3955 /**
3956  * cik_copy_cpdma - copy pages using the CP DMA engine
3957  *
3958  * @rdev: radeon_device pointer
3959  * @src_offset: src GPU address
3960  * @dst_offset: dst GPU address
3961  * @num_gpu_pages: number of GPU pages to xfer
3962  * @resv: reservation object to sync to
3963  *
3964  * Copy GPU paging using the CP DMA engine (CIK+).
3965  * Used by the radeon ttm implementation to move pages if
3966  * registered as the asic copy callback.
3967  */
3968 struct radeon_fence *cik_copy_cpdma(struct radeon_device *rdev,
3969                                     uint64_t src_offset, uint64_t dst_offset,
3970                                     unsigned num_gpu_pages,
3971                                     struct reservation_object *resv)
3972 {
3973         struct radeon_semaphore *sem = NULL;
3974         struct radeon_fence *fence;
3975         int ring_index = rdev->asic->copy.blit_ring_index;
3976         struct radeon_ring *ring = &rdev->ring[ring_index];
3977         u32 size_in_bytes, cur_size_in_bytes, control;
3978         int i, num_loops;
3979         int r = 0;
3980
3981         r = radeon_semaphore_create(rdev, &sem);
3982         if (r) {
3983                 DRM_ERROR("radeon: moving bo (%d).\n", r);
3984                 return ERR_PTR(r);
3985         }
3986
3987         size_in_bytes = (num_gpu_pages << RADEON_GPU_PAGE_SHIFT);
3988         num_loops = DIV_ROUND_UP(size_in_bytes, 0x1fffff);
3989         r = radeon_ring_lock(rdev, ring, num_loops * 7 + 18);
3990         if (r) {
3991                 DRM_ERROR("radeon: moving bo (%d).\n", r);
3992                 radeon_semaphore_free(rdev, &sem, NULL);
3993                 return ERR_PTR(r);
3994         }
3995
3996         radeon_semaphore_sync_resv(rdev, sem, resv, false);
3997         radeon_semaphore_sync_rings(rdev, sem, ring->idx);
3998
3999         for (i = 0; i < num_loops; i++) {
4000                 cur_size_in_bytes = size_in_bytes;
4001                 if (cur_size_in_bytes > 0x1fffff)
4002                         cur_size_in_bytes = 0x1fffff;
4003                 size_in_bytes -= cur_size_in_bytes;
4004                 control = 0;
4005                 if (size_in_bytes == 0)
4006                         control |= PACKET3_DMA_DATA_CP_SYNC;
4007                 radeon_ring_write(ring, PACKET3(PACKET3_DMA_DATA, 5));
4008                 radeon_ring_write(ring, control);
4009                 radeon_ring_write(ring, lower_32_bits(src_offset));
4010                 radeon_ring_write(ring, upper_32_bits(src_offset));
4011                 radeon_ring_write(ring, lower_32_bits(dst_offset));
4012                 radeon_ring_write(ring, upper_32_bits(dst_offset));
4013                 radeon_ring_write(ring, cur_size_in_bytes);
4014                 src_offset += cur_size_in_bytes;
4015                 dst_offset += cur_size_in_bytes;
4016         }
4017
4018         r = radeon_fence_emit(rdev, &fence, ring->idx);
4019         if (r) {
4020                 radeon_ring_unlock_undo(rdev, ring);
4021                 radeon_semaphore_free(rdev, &sem, NULL);
4022                 return ERR_PTR(r);
4023         }
4024
4025         radeon_ring_unlock_commit(rdev, ring, false);
4026         radeon_semaphore_free(rdev, &sem, fence);
4027
4028         return fence;
4029 }
4030
4031 /*
4032  * IB stuff
4033  */
4034 /**
4035  * cik_ring_ib_execute - emit an IB (Indirect Buffer) on the gfx ring
4036  *
4037  * @rdev: radeon_device pointer
4038  * @ib: radeon indirect buffer object
4039  *
4040  * Emits an DE (drawing engine) or CE (constant engine) IB
4041  * on the gfx ring.  IBs are usually generated by userspace
4042  * acceleration drivers and submitted to the kernel for
4043  * sheduling on the ring.  This function schedules the IB
4044  * on the gfx ring for execution by the GPU.
4045  */
4046 void cik_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
4047 {
4048         struct radeon_ring *ring = &rdev->ring[ib->ring];
4049         u32 header, control = INDIRECT_BUFFER_VALID;
4050
4051         if (ib->is_const_ib) {
4052                 /* set switch buffer packet before const IB */
4053                 radeon_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
4054                 radeon_ring_write(ring, 0);
4055
4056                 header = PACKET3(PACKET3_INDIRECT_BUFFER_CONST, 2);
4057         } else {
4058                 u32 next_rptr;
4059                 if (ring->rptr_save_reg) {
4060                         next_rptr = ring->wptr + 3 + 4;
4061                         radeon_ring_write(ring, PACKET3(PACKET3_SET_UCONFIG_REG, 1));
4062                         radeon_ring_write(ring, ((ring->rptr_save_reg -
4063                                                   PACKET3_SET_UCONFIG_REG_START) >> 2));
4064                         radeon_ring_write(ring, next_rptr);
4065                 } else if (rdev->wb.enabled) {
4066                         next_rptr = ring->wptr + 5 + 4;
4067                         radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4068                         radeon_ring_write(ring, WRITE_DATA_DST_SEL(1));
4069                         radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
4070                         radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr));
4071                         radeon_ring_write(ring, next_rptr);
4072                 }
4073
4074                 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2);
4075         }
4076
4077         control |= ib->length_dw |
4078                 (ib->vm ? (ib->vm->id << 24) : 0);
4079
4080         radeon_ring_write(ring, header);
4081         radeon_ring_write(ring,
4082 #ifdef __BIG_ENDIAN
4083                           (2 << 0) |
4084 #endif
4085                           (ib->gpu_addr & 0xFFFFFFFC));
4086         radeon_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xFFFF);
4087         radeon_ring_write(ring, control);
4088 }
4089
4090 /**
4091  * cik_ib_test - basic gfx ring IB test
4092  *
4093  * @rdev: radeon_device pointer
4094  * @ring: radeon_ring structure holding ring information
4095  *
4096  * Allocate an IB and execute it on the gfx ring (CIK).
4097  * Provides a basic gfx ring test to verify that IBs are working.
4098  * Returns 0 on success, error on failure.
4099  */
4100 int cik_ib_test(struct radeon_device *rdev, struct radeon_ring *ring)
4101 {
4102         struct radeon_ib ib;
4103         uint32_t scratch;
4104         uint32_t tmp = 0;
4105         unsigned i;
4106         int r;
4107
4108         r = radeon_scratch_get(rdev, &scratch);
4109         if (r) {
4110                 DRM_ERROR("radeon: failed to get scratch reg (%d).\n", r);
4111                 return r;
4112         }
4113         WREG32(scratch, 0xCAFEDEAD);
4114         r = radeon_ib_get(rdev, ring->idx, &ib, NULL, 256);
4115         if (r) {
4116                 DRM_ERROR("radeon: failed to get ib (%d).\n", r);
4117                 radeon_scratch_free(rdev, scratch);
4118                 return r;
4119         }
4120         ib.ptr[0] = PACKET3(PACKET3_SET_UCONFIG_REG, 1);
4121         ib.ptr[1] = ((scratch - PACKET3_SET_UCONFIG_REG_START) >> 2);
4122         ib.ptr[2] = 0xDEADBEEF;
4123         ib.length_dw = 3;
4124         r = radeon_ib_schedule(rdev, &ib, NULL, false);
4125         if (r) {
4126                 radeon_scratch_free(rdev, scratch);
4127                 radeon_ib_free(rdev, &ib);
4128                 DRM_ERROR("radeon: failed to schedule ib (%d).\n", r);
4129                 return r;
4130         }
4131         r = radeon_fence_wait(ib.fence, false);
4132         if (r) {
4133                 DRM_ERROR("radeon: fence wait failed (%d).\n", r);
4134                 radeon_scratch_free(rdev, scratch);
4135                 radeon_ib_free(rdev, &ib);
4136                 return r;
4137         }
4138         for (i = 0; i < rdev->usec_timeout; i++) {
4139                 tmp = RREG32(scratch);
4140                 if (tmp == 0xDEADBEEF)
4141                         break;
4142                 DRM_UDELAY(1);
4143         }
4144         if (i < rdev->usec_timeout) {
4145                 DRM_INFO("ib test on ring %d succeeded in %u usecs\n", ib.fence->ring, i);
4146         } else {
4147                 DRM_ERROR("radeon: ib test failed (scratch(0x%04X)=0x%08X)\n",
4148                           scratch, tmp);
4149                 r = -EINVAL;
4150         }
4151         radeon_scratch_free(rdev, scratch);
4152         radeon_ib_free(rdev, &ib);
4153         return r;
4154 }
4155
4156 /*
4157  * CP.
4158  * On CIK, gfx and compute now have independant command processors.
4159  *
4160  * GFX
4161  * Gfx consists of a single ring and can process both gfx jobs and
4162  * compute jobs.  The gfx CP consists of three microengines (ME):
4163  * PFP - Pre-Fetch Parser
4164  * ME - Micro Engine
4165  * CE - Constant Engine
4166  * The PFP and ME make up what is considered the Drawing Engine (DE).
4167  * The CE is an asynchronous engine used for updating buffer desciptors
4168  * used by the DE so that they can be loaded into cache in parallel
4169  * while the DE is processing state update packets.
4170  *
4171  * Compute
4172  * The compute CP consists of two microengines (ME):
4173  * MEC1 - Compute MicroEngine 1
4174  * MEC2 - Compute MicroEngine 2
4175  * Each MEC supports 4 compute pipes and each pipe supports 8 queues.
4176  * The queues are exposed to userspace and are programmed directly
4177  * by the compute runtime.
4178  */
4179 /**
4180  * cik_cp_gfx_enable - enable/disable the gfx CP MEs
4181  *
4182  * @rdev: radeon_device pointer
4183  * @enable: enable or disable the MEs
4184  *
4185  * Halts or unhalts the gfx MEs.
4186  */
4187 static void cik_cp_gfx_enable(struct radeon_device *rdev, bool enable)
4188 {
4189         if (enable)
4190                 WREG32(CP_ME_CNTL, 0);
4191         else {
4192                 if (rdev->asic->copy.copy_ring_index == RADEON_RING_TYPE_GFX_INDEX)
4193                         radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
4194                 WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT));
4195                 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
4196         }
4197         udelay(50);
4198 }
4199
4200 /**
4201  * cik_cp_gfx_load_microcode - load the gfx CP ME ucode
4202  *
4203  * @rdev: radeon_device pointer
4204  *
4205  * Loads the gfx PFP, ME, and CE ucode.
4206  * Returns 0 for success, -EINVAL if the ucode is not available.
4207  */
4208 static int cik_cp_gfx_load_microcode(struct radeon_device *rdev)
4209 {
4210         int i;
4211
4212         if (!rdev->me_fw || !rdev->pfp_fw || !rdev->ce_fw)
4213                 return -EINVAL;
4214
4215         cik_cp_gfx_enable(rdev, false);
4216
4217         if (rdev->new_fw) {
4218                 const struct gfx_firmware_header_v1_0 *pfp_hdr =
4219                         (const struct gfx_firmware_header_v1_0 *)rdev->pfp_fw->data;
4220                 const struct gfx_firmware_header_v1_0 *ce_hdr =
4221                         (const struct gfx_firmware_header_v1_0 *)rdev->ce_fw->data;
4222                 const struct gfx_firmware_header_v1_0 *me_hdr =
4223                         (const struct gfx_firmware_header_v1_0 *)rdev->me_fw->data;
4224                 const __le32 *fw_data;
4225                 u32 fw_size;
4226
4227                 radeon_ucode_print_gfx_hdr(&pfp_hdr->header);
4228                 radeon_ucode_print_gfx_hdr(&ce_hdr->header);
4229                 radeon_ucode_print_gfx_hdr(&me_hdr->header);
4230
4231                 /* PFP */
4232                 fw_data = (const __le32 *)
4233                         (rdev->pfp_fw->data + le32_to_cpu(pfp_hdr->header.ucode_array_offset_bytes));
4234                 fw_size = le32_to_cpu(pfp_hdr->header.ucode_size_bytes) / 4;
4235                 WREG32(CP_PFP_UCODE_ADDR, 0);
4236                 for (i = 0; i < fw_size; i++)
4237                         WREG32(CP_PFP_UCODE_DATA, le32_to_cpup(fw_data++));
4238                 WREG32(CP_PFP_UCODE_ADDR, le32_to_cpu(pfp_hdr->header.ucode_version));
4239
4240                 /* CE */
4241                 fw_data = (const __le32 *)
4242                         (rdev->ce_fw->data + le32_to_cpu(ce_hdr->header.ucode_array_offset_bytes));
4243                 fw_size = le32_to_cpu(ce_hdr->header.ucode_size_bytes) / 4;
4244                 WREG32(CP_CE_UCODE_ADDR, 0);
4245                 for (i = 0; i < fw_size; i++)
4246                         WREG32(CP_CE_UCODE_DATA, le32_to_cpup(fw_data++));
4247                 WREG32(CP_CE_UCODE_ADDR, le32_to_cpu(ce_hdr->header.ucode_version));
4248
4249                 /* ME */
4250                 fw_data = (const __be32 *)
4251                         (rdev->me_fw->data + le32_to_cpu(me_hdr->header.ucode_array_offset_bytes));
4252                 fw_size = le32_to_cpu(me_hdr->header.ucode_size_bytes) / 4;
4253                 WREG32(CP_ME_RAM_WADDR, 0);
4254                 for (i = 0; i < fw_size; i++)
4255                         WREG32(CP_ME_RAM_DATA, le32_to_cpup(fw_data++));
4256                 WREG32(CP_ME_RAM_WADDR, le32_to_cpu(me_hdr->header.ucode_version));
4257                 WREG32(CP_ME_RAM_RADDR, le32_to_cpu(me_hdr->header.ucode_version));
4258         } else {
4259                 const __be32 *fw_data;
4260
4261                 /* PFP */
4262                 fw_data = (const __be32 *)rdev->pfp_fw->data;
4263                 WREG32(CP_PFP_UCODE_ADDR, 0);
4264                 for (i = 0; i < CIK_PFP_UCODE_SIZE; i++)
4265                         WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
4266                 WREG32(CP_PFP_UCODE_ADDR, 0);
4267
4268                 /* CE */
4269                 fw_data = (const __be32 *)rdev->ce_fw->data;
4270                 WREG32(CP_CE_UCODE_ADDR, 0);
4271                 for (i = 0; i < CIK_CE_UCODE_SIZE; i++)
4272                         WREG32(CP_CE_UCODE_DATA, be32_to_cpup(fw_data++));
4273                 WREG32(CP_CE_UCODE_ADDR, 0);
4274
4275                 /* ME */
4276                 fw_data = (const __be32 *)rdev->me_fw->data;
4277                 WREG32(CP_ME_RAM_WADDR, 0);
4278                 for (i = 0; i < CIK_ME_UCODE_SIZE; i++)
4279                         WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
4280                 WREG32(CP_ME_RAM_WADDR, 0);
4281         }
4282
4283         return 0;
4284 }
4285
4286 /**
4287  * cik_cp_gfx_start - start the gfx ring
4288  *
4289  * @rdev: radeon_device pointer
4290  *
4291  * Enables the ring and loads the clear state context and other
4292  * packets required to init the ring.
4293  * Returns 0 for success, error for failure.
4294  */
4295 static int cik_cp_gfx_start(struct radeon_device *rdev)
4296 {
4297         struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
4298         int r, i;
4299
4300         /* init the CP */
4301         WREG32(CP_MAX_CONTEXT, rdev->config.cik.max_hw_contexts - 1);
4302         WREG32(CP_ENDIAN_SWAP, 0);
4303         WREG32(CP_DEVICE_ID, 1);
4304
4305         cik_cp_gfx_enable(rdev, true);
4306
4307         r = radeon_ring_lock(rdev, ring, cik_default_size + 17);
4308         if (r) {
4309                 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
4310                 return r;
4311         }
4312
4313         /* init the CE partitions.  CE only used for gfx on CIK */
4314         radeon_ring_write(ring, PACKET3(PACKET3_SET_BASE, 2));
4315         radeon_ring_write(ring, PACKET3_BASE_INDEX(CE_PARTITION_BASE));
4316         radeon_ring_write(ring, 0xc000);
4317         radeon_ring_write(ring, 0xc000);
4318
4319         /* setup clear context state */
4320         radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
4321         radeon_ring_write(ring, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
4322
4323         radeon_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1));
4324         radeon_ring_write(ring, 0x80000000);
4325         radeon_ring_write(ring, 0x80000000);
4326
4327         for (i = 0; i < cik_default_size; i++)
4328                 radeon_ring_write(ring, cik_default_state[i]);
4329
4330         radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
4331         radeon_ring_write(ring, PACKET3_PREAMBLE_END_CLEAR_STATE);
4332
4333         /* set clear context state */
4334         radeon_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0));
4335         radeon_ring_write(ring, 0);
4336
4337         radeon_ring_write(ring, PACKET3(PACKET3_SET_CONTEXT_REG, 2));
4338         radeon_ring_write(ring, 0x00000316);
4339         radeon_ring_write(ring, 0x0000000e); /* VGT_VERTEX_REUSE_BLOCK_CNTL */
4340         radeon_ring_write(ring, 0x00000010); /* VGT_OUT_DEALLOC_CNTL */
4341
4342         radeon_ring_unlock_commit(rdev, ring, false);
4343
4344         return 0;
4345 }
4346
4347 /**
4348  * cik_cp_gfx_fini - stop the gfx ring
4349  *
4350  * @rdev: radeon_device pointer
4351  *
4352  * Stop the gfx ring and tear down the driver ring
4353  * info.
4354  */
4355 static void cik_cp_gfx_fini(struct radeon_device *rdev)
4356 {
4357         cik_cp_gfx_enable(rdev, false);
4358         radeon_ring_fini(rdev, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX]);
4359 }
4360
4361 /**
4362  * cik_cp_gfx_resume - setup the gfx ring buffer registers
4363  *
4364  * @rdev: radeon_device pointer
4365  *
4366  * Program the location and size of the gfx ring buffer
4367  * and test it to make sure it's working.
4368  * Returns 0 for success, error for failure.
4369  */
4370 static int cik_cp_gfx_resume(struct radeon_device *rdev)
4371 {
4372         struct radeon_ring *ring;
4373         u32 tmp;
4374         u32 rb_bufsz;
4375         u64 rb_addr;
4376         int r;
4377
4378         WREG32(CP_SEM_WAIT_TIMER, 0x0);
4379         if (rdev->family != CHIP_HAWAII)
4380                 WREG32(CP_SEM_INCOMPLETE_TIMER_CNTL, 0x0);
4381
4382         /* Set the write pointer delay */
4383         WREG32(CP_RB_WPTR_DELAY, 0);
4384
4385         /* set the RB to use vmid 0 */
4386         WREG32(CP_RB_VMID, 0);
4387
4388         WREG32(SCRATCH_ADDR, ((rdev->wb.gpu_addr + RADEON_WB_SCRATCH_OFFSET) >> 8) & 0xFFFFFFFF);
4389
4390         /* ring 0 - compute and gfx */
4391         /* Set ring buffer size */
4392         ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
4393         rb_bufsz = order_base_2(ring->ring_size / 8);
4394         tmp = (order_base_2(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
4395 #ifdef __BIG_ENDIAN
4396         tmp |= BUF_SWAP_32BIT;
4397 #endif
4398         WREG32(CP_RB0_CNTL, tmp);
4399
4400         /* Initialize the ring buffer's read and write pointers */
4401         WREG32(CP_RB0_CNTL, tmp | RB_RPTR_WR_ENA);
4402         ring->wptr = 0;
4403         WREG32(CP_RB0_WPTR, ring->wptr);
4404
4405         /* set the wb address wether it's enabled or not */
4406         WREG32(CP_RB0_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFFFFFFFC);
4407         WREG32(CP_RB0_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFF);
4408
4409         /* scratch register shadowing is no longer supported */
4410         WREG32(SCRATCH_UMSK, 0);
4411
4412         if (!rdev->wb.enabled)
4413                 tmp |= RB_NO_UPDATE;
4414
4415         mdelay(1);
4416         WREG32(CP_RB0_CNTL, tmp);
4417
4418         rb_addr = ring->gpu_addr >> 8;
4419         WREG32(CP_RB0_BASE, rb_addr);
4420         WREG32(CP_RB0_BASE_HI, upper_32_bits(rb_addr));
4421
4422         /* start the ring */
4423         cik_cp_gfx_start(rdev);
4424         rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = true;
4425         r = radeon_ring_test(rdev, RADEON_RING_TYPE_GFX_INDEX, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX]);
4426         if (r) {
4427                 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
4428                 return r;
4429         }
4430
4431         if (rdev->asic->copy.copy_ring_index == RADEON_RING_TYPE_GFX_INDEX)
4432                 radeon_ttm_set_active_vram_size(rdev, rdev->mc.real_vram_size);
4433
4434         return 0;
4435 }
4436
4437 u32 cik_gfx_get_rptr(struct radeon_device *rdev,
4438                      struct radeon_ring *ring)
4439 {
4440         u32 rptr;
4441
4442         if (rdev->wb.enabled)
4443                 rptr = rdev->wb.wb[ring->rptr_offs/4];
4444         else
4445                 rptr = RREG32(CP_RB0_RPTR);
4446
4447         return rptr;
4448 }
4449
4450 u32 cik_gfx_get_wptr(struct radeon_device *rdev,
4451                      struct radeon_ring *ring)
4452 {
4453         u32 wptr;
4454
4455         wptr = RREG32(CP_RB0_WPTR);
4456
4457         return wptr;
4458 }
4459
4460 void cik_gfx_set_wptr(struct radeon_device *rdev,
4461                       struct radeon_ring *ring)
4462 {
4463         WREG32(CP_RB0_WPTR, ring->wptr);
4464         (void)RREG32(CP_RB0_WPTR);
4465 }
4466
4467 u32 cik_compute_get_rptr(struct radeon_device *rdev,
4468                          struct radeon_ring *ring)
4469 {
4470         u32 rptr;
4471
4472         if (rdev->wb.enabled) {
4473                 rptr = rdev->wb.wb[ring->rptr_offs/4];
4474         } else {
4475                 mutex_lock(&rdev->srbm_mutex);
4476                 cik_srbm_select(rdev, ring->me, ring->pipe, ring->queue, 0);
4477                 rptr = RREG32(CP_HQD_PQ_RPTR);
4478                 cik_srbm_select(rdev, 0, 0, 0, 0);
4479                 mutex_unlock(&rdev->srbm_mutex);
4480         }
4481
4482         return rptr;
4483 }
4484
4485 u32 cik_compute_get_wptr(struct radeon_device *rdev,
4486                          struct radeon_ring *ring)
4487 {
4488         u32 wptr;
4489
4490         if (rdev->wb.enabled) {
4491                 /* XXX check if swapping is necessary on BE */
4492                 wptr = rdev->wb.wb[ring->wptr_offs/4];
4493         } else {
4494                 mutex_lock(&rdev->srbm_mutex);
4495                 cik_srbm_select(rdev, ring->me, ring->pipe, ring->queue, 0);
4496                 wptr = RREG32(CP_HQD_PQ_WPTR);
4497                 cik_srbm_select(rdev, 0, 0, 0, 0);
4498                 mutex_unlock(&rdev->srbm_mutex);
4499         }
4500
4501         return wptr;
4502 }
4503
4504 void cik_compute_set_wptr(struct radeon_device *rdev,
4505                           struct radeon_ring *ring)
4506 {
4507         /* XXX check if swapping is necessary on BE */
4508         rdev->wb.wb[ring->wptr_offs/4] = ring->wptr;
4509         WDOORBELL32(ring->doorbell_index, ring->wptr);
4510 }
4511
4512 /**
4513  * cik_cp_compute_enable - enable/disable the compute CP MEs
4514  *
4515  * @rdev: radeon_device pointer
4516  * @enable: enable or disable the MEs
4517  *
4518  * Halts or unhalts the compute MEs.
4519  */
4520 static void cik_cp_compute_enable(struct radeon_device *rdev, bool enable)
4521 {
4522         if (enable)
4523                 WREG32(CP_MEC_CNTL, 0);
4524         else {
4525                 WREG32(CP_MEC_CNTL, (MEC_ME1_HALT | MEC_ME2_HALT));
4526                 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
4527                 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
4528         }
4529         udelay(50);
4530 }
4531
4532 /**
4533  * cik_cp_compute_load_microcode - load the compute CP ME ucode
4534  *
4535  * @rdev: radeon_device pointer
4536  *
4537  * Loads the compute MEC1&2 ucode.
4538  * Returns 0 for success, -EINVAL if the ucode is not available.
4539  */
4540 static int cik_cp_compute_load_microcode(struct radeon_device *rdev)
4541 {
4542         int i;
4543
4544         if (!rdev->mec_fw)
4545                 return -EINVAL;
4546
4547         cik_cp_compute_enable(rdev, false);
4548
4549         if (rdev->new_fw) {
4550                 const struct gfx_firmware_header_v1_0 *mec_hdr =
4551                         (const struct gfx_firmware_header_v1_0 *)rdev->mec_fw->data;
4552                 const __le32 *fw_data;
4553                 u32 fw_size;
4554
4555                 radeon_ucode_print_gfx_hdr(&mec_hdr->header);
4556
4557                 /* MEC1 */
4558                 fw_data = (const __le32 *)
4559                         (rdev->mec_fw->data + le32_to_cpu(mec_hdr->header.ucode_array_offset_bytes));
4560                 fw_size = le32_to_cpu(mec_hdr->header.ucode_size_bytes) / 4;
4561                 WREG32(CP_MEC_ME1_UCODE_ADDR, 0);
4562                 for (i = 0; i < fw_size; i++)
4563                         WREG32(CP_MEC_ME1_UCODE_DATA, le32_to_cpup(fw_data++));
4564                 WREG32(CP_MEC_ME1_UCODE_ADDR, le32_to_cpu(mec_hdr->header.ucode_version));
4565
4566                 /* MEC2 */
4567                 if (rdev->family == CHIP_KAVERI) {
4568                         const struct gfx_firmware_header_v1_0 *mec2_hdr =
4569                                 (const struct gfx_firmware_header_v1_0 *)rdev->mec2_fw->data;
4570
4571                         fw_data = (const __le32 *)
4572                                 (rdev->mec2_fw->data +
4573                                  le32_to_cpu(mec2_hdr->header.ucode_array_offset_bytes));
4574                         fw_size = le32_to_cpu(mec2_hdr->header.ucode_size_bytes) / 4;
4575                         WREG32(CP_MEC_ME2_UCODE_ADDR, 0);
4576                         for (i = 0; i < fw_size; i++)
4577                                 WREG32(CP_MEC_ME2_UCODE_DATA, le32_to_cpup(fw_data++));
4578                         WREG32(CP_MEC_ME2_UCODE_ADDR, le32_to_cpu(mec2_hdr->header.ucode_version));
4579                 }
4580         } else {
4581                 const __be32 *fw_data;
4582
4583                 /* MEC1 */
4584                 fw_data = (const __be32 *)rdev->mec_fw->data;
4585                 WREG32(CP_MEC_ME1_UCODE_ADDR, 0);
4586                 for (i = 0; i < CIK_MEC_UCODE_SIZE; i++)
4587                         WREG32(CP_MEC_ME1_UCODE_DATA, be32_to_cpup(fw_data++));
4588                 WREG32(CP_MEC_ME1_UCODE_ADDR, 0);
4589
4590                 if (rdev->family == CHIP_KAVERI) {
4591                         /* MEC2 */
4592                         fw_data = (const __be32 *)rdev->mec_fw->data;
4593                         WREG32(CP_MEC_ME2_UCODE_ADDR, 0);
4594                         for (i = 0; i < CIK_MEC_UCODE_SIZE; i++)
4595                                 WREG32(CP_MEC_ME2_UCODE_DATA, be32_to_cpup(fw_data++));
4596                         WREG32(CP_MEC_ME2_UCODE_ADDR, 0);
4597                 }
4598         }
4599
4600         return 0;
4601 }
4602
4603 /**
4604  * cik_cp_compute_start - start the compute queues
4605  *
4606  * @rdev: radeon_device pointer
4607  *
4608  * Enable the compute queues.
4609  * Returns 0 for success, error for failure.
4610  */
4611 static int cik_cp_compute_start(struct radeon_device *rdev)
4612 {
4613         cik_cp_compute_enable(rdev, true);
4614
4615         return 0;
4616 }
4617
4618 /**
4619  * cik_cp_compute_fini - stop the compute queues
4620  *
4621  * @rdev: radeon_device pointer
4622  *
4623  * Stop the compute queues and tear down the driver queue
4624  * info.
4625  */
4626 static void cik_cp_compute_fini(struct radeon_device *rdev)
4627 {
4628         int i, idx, r;
4629
4630         cik_cp_compute_enable(rdev, false);
4631
4632         for (i = 0; i < 2; i++) {
4633                 if (i == 0)
4634                         idx = CAYMAN_RING_TYPE_CP1_INDEX;
4635                 else
4636                         idx = CAYMAN_RING_TYPE_CP2_INDEX;
4637
4638                 if (rdev->ring[idx].mqd_obj) {
4639                         r = radeon_bo_reserve(rdev->ring[idx].mqd_obj, false);
4640                         if (unlikely(r != 0))
4641                                 dev_warn(rdev->dev, "(%d) reserve MQD bo failed\n", r);
4642
4643                         radeon_bo_unpin(rdev->ring[idx].mqd_obj);
4644                         radeon_bo_unreserve(rdev->ring[idx].mqd_obj);
4645
4646                         radeon_bo_unref(&rdev->ring[idx].mqd_obj);
4647                         rdev->ring[idx].mqd_obj = NULL;
4648                 }
4649         }
4650 }
4651
4652 static void cik_mec_fini(struct radeon_device *rdev)
4653 {
4654         int r;
4655
4656         if (rdev->mec.hpd_eop_obj) {
4657                 r = radeon_bo_reserve(rdev->mec.hpd_eop_obj, false);
4658                 if (unlikely(r != 0))
4659                         dev_warn(rdev->dev, "(%d) reserve HPD EOP bo failed\n", r);
4660                 radeon_bo_unpin(rdev->mec.hpd_eop_obj);
4661                 radeon_bo_unreserve(rdev->mec.hpd_eop_obj);
4662
4663                 radeon_bo_unref(&rdev->mec.hpd_eop_obj);
4664                 rdev->mec.hpd_eop_obj = NULL;
4665         }
4666 }
4667
4668 #define MEC_HPD_SIZE 2048
4669
4670 static int cik_mec_init(struct radeon_device *rdev)
4671 {
4672         int r;
4673         u32 *hpd;
4674
4675         /*
4676          * KV:    2 MEC, 4 Pipes/MEC, 8 Queues/Pipe - 64 Queues total
4677          * CI/KB: 1 MEC, 4 Pipes/MEC, 8 Queues/Pipe - 32 Queues total
4678          */
4679         if (rdev->family == CHIP_KAVERI)
4680                 rdev->mec.num_mec = 2;
4681         else
4682                 rdev->mec.num_mec = 1;
4683         rdev->mec.num_pipe = 4;
4684         rdev->mec.num_queue = rdev->mec.num_mec * rdev->mec.num_pipe * 8;
4685
4686         if (rdev->mec.hpd_eop_obj == NULL) {
4687                 r = radeon_bo_create(rdev,
4688                                      rdev->mec.num_mec *rdev->mec.num_pipe * MEC_HPD_SIZE * 2,
4689                                      PAGE_SIZE, true,
4690                                      RADEON_GEM_DOMAIN_GTT, 0, NULL, NULL,
4691                                      &rdev->mec.hpd_eop_obj);
4692                 if (r) {
4693                         dev_warn(rdev->dev, "(%d) create HDP EOP bo failed\n", r);
4694                         return r;
4695                 }
4696         }
4697
4698         r = radeon_bo_reserve(rdev->mec.hpd_eop_obj, false);
4699         if (unlikely(r != 0)) {
4700                 cik_mec_fini(rdev);
4701                 return r;
4702         }
4703         r = radeon_bo_pin(rdev->mec.hpd_eop_obj, RADEON_GEM_DOMAIN_GTT,
4704                           &rdev->mec.hpd_eop_gpu_addr);
4705         if (r) {
4706                 dev_warn(rdev->dev, "(%d) pin HDP EOP bo failed\n", r);
4707                 cik_mec_fini(rdev);
4708                 return r;
4709         }
4710         r = radeon_bo_kmap(rdev->mec.hpd_eop_obj, (void **)&hpd);
4711         if (r) {
4712                 dev_warn(rdev->dev, "(%d) map HDP EOP bo failed\n", r);
4713                 cik_mec_fini(rdev);
4714                 return r;
4715         }
4716
4717         /* clear memory.  Not sure if this is required or not */
4718         memset(hpd, 0, rdev->mec.num_mec *rdev->mec.num_pipe * MEC_HPD_SIZE * 2);
4719
4720         radeon_bo_kunmap(rdev->mec.hpd_eop_obj);
4721         radeon_bo_unreserve(rdev->mec.hpd_eop_obj);
4722
4723         return 0;
4724 }
4725
4726 struct hqd_registers
4727 {
4728         u32 cp_mqd_base_addr;
4729         u32 cp_mqd_base_addr_hi;
4730         u32 cp_hqd_active;
4731         u32 cp_hqd_vmid;
4732         u32 cp_hqd_persistent_state;
4733         u32 cp_hqd_pipe_priority;
4734         u32 cp_hqd_queue_priority;
4735         u32 cp_hqd_quantum;
4736         u32 cp_hqd_pq_base;
4737         u32 cp_hqd_pq_base_hi;
4738         u32 cp_hqd_pq_rptr;
4739         u32 cp_hqd_pq_rptr_report_addr;
4740         u32 cp_hqd_pq_rptr_report_addr_hi;
4741         u32 cp_hqd_pq_wptr_poll_addr;
4742         u32 cp_hqd_pq_wptr_poll_addr_hi;
4743         u32 cp_hqd_pq_doorbell_control;
4744         u32 cp_hqd_pq_wptr;
4745         u32 cp_hqd_pq_control;
4746         u32 cp_hqd_ib_base_addr;
4747         u32 cp_hqd_ib_base_addr_hi;
4748         u32 cp_hqd_ib_rptr;
4749         u32 cp_hqd_ib_control;
4750         u32 cp_hqd_iq_timer;
4751         u32 cp_hqd_iq_rptr;
4752         u32 cp_hqd_dequeue_request;
4753         u32 cp_hqd_dma_offload;
4754         u32 cp_hqd_sema_cmd;
4755         u32 cp_hqd_msg_type;
4756         u32 cp_hqd_atomic0_preop_lo;
4757         u32 cp_hqd_atomic0_preop_hi;
4758         u32 cp_hqd_atomic1_preop_lo;
4759         u32 cp_hqd_atomic1_preop_hi;
4760         u32 cp_hqd_hq_scheduler0;
4761         u32 cp_hqd_hq_scheduler1;
4762         u32 cp_mqd_control;
4763 };
4764
4765 struct bonaire_mqd
4766 {
4767         u32 header;
4768         u32 dispatch_initiator;
4769         u32 dimensions[3];
4770         u32 start_idx[3];
4771         u32 num_threads[3];
4772         u32 pipeline_stat_enable;
4773         u32 perf_counter_enable;
4774         u32 pgm[2];
4775         u32 tba[2];
4776         u32 tma[2];
4777         u32 pgm_rsrc[2];
4778         u32 vmid;
4779         u32 resource_limits;
4780         u32 static_thread_mgmt01[2];
4781         u32 tmp_ring_size;
4782         u32 static_thread_mgmt23[2];
4783         u32 restart[3];
4784         u32 thread_trace_enable;
4785         u32 reserved1;
4786         u32 user_data[16];
4787         u32 vgtcs_invoke_count[2];
4788         struct hqd_registers queue_state;
4789         u32 dequeue_cntr;
4790         u32 interrupt_queue[64];
4791 };
4792
4793 /**
4794  * cik_cp_compute_resume - setup the compute queue registers
4795  *
4796  * @rdev: radeon_device pointer
4797  *
4798  * Program the compute queues and test them to make sure they
4799  * are working.
4800  * Returns 0 for success, error for failure.
4801  */
4802 static int cik_cp_compute_resume(struct radeon_device *rdev)
4803 {
4804         int r, i, idx;
4805         u32 tmp;
4806         bool use_doorbell = true;
4807         u64 hqd_gpu_addr;
4808         u64 mqd_gpu_addr;
4809         u64 eop_gpu_addr;
4810         u64 wb_gpu_addr;
4811         u32 *buf;
4812         struct bonaire_mqd *mqd;
4813
4814         r = cik_cp_compute_start(rdev);
4815         if (r)
4816                 return r;
4817
4818         /* fix up chicken bits */
4819         tmp = RREG32(CP_CPF_DEBUG);
4820         tmp |= (1 << 23);
4821         WREG32(CP_CPF_DEBUG, tmp);
4822
4823         /* init the pipes */
4824         mutex_lock(&rdev->srbm_mutex);
4825         for (i = 0; i < (rdev->mec.num_pipe * rdev->mec.num_mec); i++) {
4826                 int me = (i < 4) ? 1 : 2;
4827                 int pipe = (i < 4) ? i : (i - 4);
4828
4829                 eop_gpu_addr = rdev->mec.hpd_eop_gpu_addr + (i * MEC_HPD_SIZE * 2);
4830
4831                 cik_srbm_select(rdev, me, pipe, 0, 0);
4832
4833                 /* write the EOP addr */
4834                 WREG32(CP_HPD_EOP_BASE_ADDR, eop_gpu_addr >> 8);
4835                 WREG32(CP_HPD_EOP_BASE_ADDR_HI, upper_32_bits(eop_gpu_addr) >> 8);
4836
4837                 /* set the VMID assigned */
4838                 WREG32(CP_HPD_EOP_VMID, 0);
4839
4840                 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
4841                 tmp = RREG32(CP_HPD_EOP_CONTROL);
4842                 tmp &= ~EOP_SIZE_MASK;
4843                 tmp |= order_base_2(MEC_HPD_SIZE / 8);
4844                 WREG32(CP_HPD_EOP_CONTROL, tmp);
4845         }
4846         cik_srbm_select(rdev, 0, 0, 0, 0);
4847         mutex_unlock(&rdev->srbm_mutex);
4848
4849         /* init the queues.  Just two for now. */
4850         for (i = 0; i < 2; i++) {
4851                 if (i == 0)
4852                         idx = CAYMAN_RING_TYPE_CP1_INDEX;
4853                 else
4854                         idx = CAYMAN_RING_TYPE_CP2_INDEX;
4855
4856                 if (rdev->ring[idx].mqd_obj == NULL) {
4857                         r = radeon_bo_create(rdev,
4858                                              sizeof(struct bonaire_mqd),
4859                                              PAGE_SIZE, true,
4860                                              RADEON_GEM_DOMAIN_GTT, 0, NULL,
4861                                              NULL, &rdev->ring[idx].mqd_obj);
4862                         if (r) {
4863                                 dev_warn(rdev->dev, "(%d) create MQD bo failed\n", r);
4864                                 return r;
4865                         }
4866                 }
4867
4868                 r = radeon_bo_reserve(rdev->ring[idx].mqd_obj, false);
4869                 if (unlikely(r != 0)) {
4870                         cik_cp_compute_fini(rdev);
4871                         return r;
4872                 }
4873                 r = radeon_bo_pin(rdev->ring[idx].mqd_obj, RADEON_GEM_DOMAIN_GTT,
4874                                   &mqd_gpu_addr);
4875                 if (r) {
4876                         dev_warn(rdev->dev, "(%d) pin MQD bo failed\n", r);
4877                         cik_cp_compute_fini(rdev);
4878                         return r;
4879                 }
4880                 r = radeon_bo_kmap(rdev->ring[idx].mqd_obj, (void **)&buf);
4881                 if (r) {
4882                         dev_warn(rdev->dev, "(%d) map MQD bo failed\n", r);
4883                         cik_cp_compute_fini(rdev);
4884                         return r;
4885                 }
4886
4887                 /* init the mqd struct */
4888                 memset(buf, 0, sizeof(struct bonaire_mqd));
4889
4890                 mqd = (struct bonaire_mqd *)buf;
4891                 mqd->header = 0xC0310800;
4892                 mqd->static_thread_mgmt01[0] = 0xffffffff;
4893                 mqd->static_thread_mgmt01[1] = 0xffffffff;
4894                 mqd->static_thread_mgmt23[0] = 0xffffffff;
4895                 mqd->static_thread_mgmt23[1] = 0xffffffff;
4896
4897                 mutex_lock(&rdev->srbm_mutex);
4898                 cik_srbm_select(rdev, rdev->ring[idx].me,
4899                                 rdev->ring[idx].pipe,
4900                                 rdev->ring[idx].queue, 0);
4901
4902                 /* disable wptr polling */
4903                 tmp = RREG32(CP_PQ_WPTR_POLL_CNTL);
4904                 tmp &= ~WPTR_POLL_EN;
4905                 WREG32(CP_PQ_WPTR_POLL_CNTL, tmp);
4906
4907                 /* enable doorbell? */
4908                 mqd->queue_state.cp_hqd_pq_doorbell_control =
4909                         RREG32(CP_HQD_PQ_DOORBELL_CONTROL);
4910                 if (use_doorbell)
4911                         mqd->queue_state.cp_hqd_pq_doorbell_control |= DOORBELL_EN;
4912                 else
4913                         mqd->queue_state.cp_hqd_pq_doorbell_control &= ~DOORBELL_EN;
4914                 WREG32(CP_HQD_PQ_DOORBELL_CONTROL,
4915                        mqd->queue_state.cp_hqd_pq_doorbell_control);
4916
4917                 /* disable the queue if it's active */
4918                 mqd->queue_state.cp_hqd_dequeue_request = 0;
4919                 mqd->queue_state.cp_hqd_pq_rptr = 0;
4920                 mqd->queue_state.cp_hqd_pq_wptr= 0;
4921                 if (RREG32(CP_HQD_ACTIVE) & 1) {
4922                         WREG32(CP_HQD_DEQUEUE_REQUEST, 1);
4923                         for (i = 0; i < rdev->usec_timeout; i++) {
4924                                 if (!(RREG32(CP_HQD_ACTIVE) & 1))
4925                                         break;
4926                                 udelay(1);
4927                         }
4928                         WREG32(CP_HQD_DEQUEUE_REQUEST, mqd->queue_state.cp_hqd_dequeue_request);
4929                         WREG32(CP_HQD_PQ_RPTR, mqd->queue_state.cp_hqd_pq_rptr);
4930                         WREG32(CP_HQD_PQ_WPTR, mqd->queue_state.cp_hqd_pq_wptr);
4931                 }
4932
4933                 /* set the pointer to the MQD */
4934                 mqd->queue_state.cp_mqd_base_addr = mqd_gpu_addr & 0xfffffffc;
4935                 mqd->queue_state.cp_mqd_base_addr_hi = upper_32_bits(mqd_gpu_addr);
4936                 WREG32(CP_MQD_BASE_ADDR, mqd->queue_state.cp_mqd_base_addr);
4937                 WREG32(CP_MQD_BASE_ADDR_HI, mqd->queue_state.cp_mqd_base_addr_hi);
4938                 /* set MQD vmid to 0 */
4939                 mqd->queue_state.cp_mqd_control = RREG32(CP_MQD_CONTROL);
4940                 mqd->queue_state.cp_mqd_control &= ~MQD_VMID_MASK;
4941                 WREG32(CP_MQD_CONTROL, mqd->queue_state.cp_mqd_control);
4942
4943                 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
4944                 hqd_gpu_addr = rdev->ring[idx].gpu_addr >> 8;
4945                 mqd->queue_state.cp_hqd_pq_base = hqd_gpu_addr;
4946                 mqd->queue_state.cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
4947                 WREG32(CP_HQD_PQ_BASE, mqd->queue_state.cp_hqd_pq_base);
4948                 WREG32(CP_HQD_PQ_BASE_HI, mqd->queue_state.cp_hqd_pq_base_hi);
4949
4950                 /* set up the HQD, this is similar to CP_RB0_CNTL */
4951                 mqd->queue_state.cp_hqd_pq_control = RREG32(CP_HQD_PQ_CONTROL);
4952                 mqd->queue_state.cp_hqd_pq_control &=
4953                         ~(QUEUE_SIZE_MASK | RPTR_BLOCK_SIZE_MASK);
4954
4955                 mqd->queue_state.cp_hqd_pq_control |=
4956                         order_base_2(rdev->ring[idx].ring_size / 8);
4957                 mqd->queue_state.cp_hqd_pq_control |=
4958                         (order_base_2(RADEON_GPU_PAGE_SIZE/8) << 8);
4959 #ifdef __BIG_ENDIAN
4960                 mqd->queue_state.cp_hqd_pq_control |= BUF_SWAP_32BIT;
4961 #endif
4962                 mqd->queue_state.cp_hqd_pq_control &=
4963                         ~(UNORD_DISPATCH | ROQ_PQ_IB_FLIP | PQ_VOLATILE);
4964                 mqd->queue_state.cp_hqd_pq_control |=
4965                         PRIV_STATE | KMD_QUEUE; /* assuming kernel queue control */
4966                 WREG32(CP_HQD_PQ_CONTROL, mqd->queue_state.cp_hqd_pq_control);
4967
4968                 /* only used if CP_PQ_WPTR_POLL_CNTL.WPTR_POLL_EN=1 */
4969                 if (i == 0)
4970                         wb_gpu_addr = rdev->wb.gpu_addr + CIK_WB_CP1_WPTR_OFFSET;
4971                 else
4972                         wb_gpu_addr = rdev->wb.gpu_addr + CIK_WB_CP2_WPTR_OFFSET;
4973                 mqd->queue_state.cp_hqd_pq_wptr_poll_addr = wb_gpu_addr & 0xfffffffc;
4974                 mqd->queue_state.cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
4975                 WREG32(CP_HQD_PQ_WPTR_POLL_ADDR, mqd->queue_state.cp_hqd_pq_wptr_poll_addr);
4976                 WREG32(CP_HQD_PQ_WPTR_POLL_ADDR_HI,
4977                        mqd->queue_state.cp_hqd_pq_wptr_poll_addr_hi);
4978
4979                 /* set the wb address wether it's enabled or not */
4980                 if (i == 0)
4981                         wb_gpu_addr = rdev->wb.gpu_addr + RADEON_WB_CP1_RPTR_OFFSET;
4982                 else
4983                         wb_gpu_addr = rdev->wb.gpu_addr + RADEON_WB_CP2_RPTR_OFFSET;
4984                 mqd->queue_state.cp_hqd_pq_rptr_report_addr = wb_gpu_addr & 0xfffffffc;
4985                 mqd->queue_state.cp_hqd_pq_rptr_report_addr_hi =
4986                         upper_32_bits(wb_gpu_addr) & 0xffff;
4987                 WREG32(CP_HQD_PQ_RPTR_REPORT_ADDR,
4988                        mqd->queue_state.cp_hqd_pq_rptr_report_addr);
4989                 WREG32(CP_HQD_PQ_RPTR_REPORT_ADDR_HI,
4990                        mqd->queue_state.cp_hqd_pq_rptr_report_addr_hi);
4991
4992                 /* enable the doorbell if requested */
4993                 if (use_doorbell) {
4994                         mqd->queue_state.cp_hqd_pq_doorbell_control =
4995                                 RREG32(CP_HQD_PQ_DOORBELL_CONTROL);
4996                         mqd->queue_state.cp_hqd_pq_doorbell_control &= ~DOORBELL_OFFSET_MASK;
4997                         mqd->queue_state.cp_hqd_pq_doorbell_control |=
4998                                 DOORBELL_OFFSET(rdev->ring[idx].doorbell_index);
4999                         mqd->queue_state.cp_hqd_pq_doorbell_control |= DOORBELL_EN;
5000                         mqd->queue_state.cp_hqd_pq_doorbell_control &=
5001                                 ~(DOORBELL_SOURCE | DOORBELL_HIT);
5002
5003                 } else {
5004                         mqd->queue_state.cp_hqd_pq_doorbell_control = 0;
5005                 }
5006                 WREG32(CP_HQD_PQ_DOORBELL_CONTROL,
5007                        mqd->queue_state.cp_hqd_pq_doorbell_control);
5008
5009                 /* read and write pointers, similar to CP_RB0_WPTR/_RPTR */
5010                 rdev->ring[idx].wptr = 0;
5011                 mqd->queue_state.cp_hqd_pq_wptr = rdev->ring[idx].wptr;
5012                 WREG32(CP_HQD_PQ_WPTR, mqd->queue_state.cp_hqd_pq_wptr);
5013                 mqd->queue_state.cp_hqd_pq_rptr = RREG32(CP_HQD_PQ_RPTR);
5014
5015                 /* set the vmid for the queue */
5016                 mqd->queue_state.cp_hqd_vmid = 0;
5017                 WREG32(CP_HQD_VMID, mqd->queue_state.cp_hqd_vmid);
5018
5019                 /* activate the queue */
5020                 mqd->queue_state.cp_hqd_active = 1;
5021                 WREG32(CP_HQD_ACTIVE, mqd->queue_state.cp_hqd_active);
5022
5023                 cik_srbm_select(rdev, 0, 0, 0, 0);
5024                 mutex_unlock(&rdev->srbm_mutex);
5025
5026                 radeon_bo_kunmap(rdev->ring[idx].mqd_obj);
5027                 radeon_bo_unreserve(rdev->ring[idx].mqd_obj);
5028
5029                 rdev->ring[idx].ready = true;
5030                 r = radeon_ring_test(rdev, idx, &rdev->ring[idx]);
5031                 if (r)
5032                         rdev->ring[idx].ready = false;
5033         }
5034
5035         return 0;
5036 }
5037
5038 static void cik_cp_enable(struct radeon_device *rdev, bool enable)
5039 {
5040         cik_cp_gfx_enable(rdev, enable);
5041         cik_cp_compute_enable(rdev, enable);
5042 }
5043
5044 static int cik_cp_load_microcode(struct radeon_device *rdev)
5045 {
5046         int r;
5047
5048         r = cik_cp_gfx_load_microcode(rdev);
5049         if (r)
5050                 return r;
5051         r = cik_cp_compute_load_microcode(rdev);
5052         if (r)
5053                 return r;
5054
5055         return 0;
5056 }
5057
5058 static void cik_cp_fini(struct radeon_device *rdev)
5059 {
5060         cik_cp_gfx_fini(rdev);
5061         cik_cp_compute_fini(rdev);
5062 }
5063
5064 static int cik_cp_resume(struct radeon_device *rdev)
5065 {
5066         int r;
5067
5068         cik_enable_gui_idle_interrupt(rdev, false);
5069
5070         r = cik_cp_load_microcode(rdev);
5071         if (r)
5072                 return r;
5073
5074         r = cik_cp_gfx_resume(rdev);
5075         if (r)
5076                 return r;
5077         r = cik_cp_compute_resume(rdev);
5078         if (r)
5079                 return r;
5080
5081         cik_enable_gui_idle_interrupt(rdev, true);
5082
5083         return 0;
5084 }
5085
5086 static void cik_print_gpu_status_regs(struct radeon_device *rdev)
5087 {
5088         dev_info(rdev->dev, "  GRBM_STATUS=0x%08X\n",
5089                 RREG32(GRBM_STATUS));
5090         dev_info(rdev->dev, "  GRBM_STATUS2=0x%08X\n",
5091                 RREG32(GRBM_STATUS2));
5092         dev_info(rdev->dev, "  GRBM_STATUS_SE0=0x%08X\n",
5093                 RREG32(GRBM_STATUS_SE0));
5094         dev_info(rdev->dev, "  GRBM_STATUS_SE1=0x%08X\n",
5095                 RREG32(GRBM_STATUS_SE1));
5096         dev_info(rdev->dev, "  GRBM_STATUS_SE2=0x%08X\n",
5097                 RREG32(GRBM_STATUS_SE2));
5098         dev_info(rdev->dev, "  GRBM_STATUS_SE3=0x%08X\n",
5099                 RREG32(GRBM_STATUS_SE3));
5100         dev_info(rdev->dev, "  SRBM_STATUS=0x%08X\n",
5101                 RREG32(SRBM_STATUS));
5102         dev_info(rdev->dev, "  SRBM_STATUS2=0x%08X\n",
5103                 RREG32(SRBM_STATUS2));
5104         dev_info(rdev->dev, "  SDMA0_STATUS_REG   = 0x%08X\n",
5105                 RREG32(SDMA0_STATUS_REG + SDMA0_REGISTER_OFFSET));
5106         dev_info(rdev->dev, "  SDMA1_STATUS_REG   = 0x%08X\n",
5107                  RREG32(SDMA0_STATUS_REG + SDMA1_REGISTER_OFFSET));
5108         dev_info(rdev->dev, "  CP_STAT = 0x%08x\n", RREG32(CP_STAT));
5109         dev_info(rdev->dev, "  CP_STALLED_STAT1 = 0x%08x\n",
5110                  RREG32(CP_STALLED_STAT1));
5111         dev_info(rdev->dev, "  CP_STALLED_STAT2 = 0x%08x\n",
5112                  RREG32(CP_STALLED_STAT2));
5113         dev_info(rdev->dev, "  CP_STALLED_STAT3 = 0x%08x\n",
5114                  RREG32(CP_STALLED_STAT3));
5115         dev_info(rdev->dev, "  CP_CPF_BUSY_STAT = 0x%08x\n",
5116                  RREG32(CP_CPF_BUSY_STAT));
5117         dev_info(rdev->dev, "  CP_CPF_STALLED_STAT1 = 0x%08x\n",
5118                  RREG32(CP_CPF_STALLED_STAT1));
5119         dev_info(rdev->dev, "  CP_CPF_STATUS = 0x%08x\n", RREG32(CP_CPF_STATUS));
5120         dev_info(rdev->dev, "  CP_CPC_BUSY_STAT = 0x%08x\n", RREG32(CP_CPC_BUSY_STAT));
5121         dev_info(rdev->dev, "  CP_CPC_STALLED_STAT1 = 0x%08x\n",
5122                  RREG32(CP_CPC_STALLED_STAT1));
5123         dev_info(rdev->dev, "  CP_CPC_STATUS = 0x%08x\n", RREG32(CP_CPC_STATUS));
5124 }
5125
5126 /**
5127  * cik_gpu_check_soft_reset - check which blocks are busy
5128  *
5129  * @rdev: radeon_device pointer
5130  *
5131  * Check which blocks are busy and return the relevant reset
5132  * mask to be used by cik_gpu_soft_reset().
5133  * Returns a mask of the blocks to be reset.
5134  */
5135 u32 cik_gpu_check_soft_reset(struct radeon_device *rdev)
5136 {
5137         u32 reset_mask = 0;
5138         u32 tmp;
5139
5140         /* GRBM_STATUS */
5141         tmp = RREG32(GRBM_STATUS);
5142         if (tmp & (PA_BUSY | SC_BUSY |
5143                    BCI_BUSY | SX_BUSY |
5144                    TA_BUSY | VGT_BUSY |
5145                    DB_BUSY | CB_BUSY |
5146                    GDS_BUSY | SPI_BUSY |
5147                    IA_BUSY | IA_BUSY_NO_DMA))
5148                 reset_mask |= RADEON_RESET_GFX;
5149
5150         if (tmp & (CP_BUSY | CP_COHERENCY_BUSY))
5151                 reset_mask |= RADEON_RESET_CP;
5152
5153         /* GRBM_STATUS2 */
5154         tmp = RREG32(GRBM_STATUS2);
5155         if (tmp & RLC_BUSY)
5156                 reset_mask |= RADEON_RESET_RLC;
5157
5158         /* SDMA0_STATUS_REG */
5159         tmp = RREG32(SDMA0_STATUS_REG + SDMA0_REGISTER_OFFSET);
5160         if (!(tmp & SDMA_IDLE))
5161                 reset_mask |= RADEON_RESET_DMA;
5162
5163         /* SDMA1_STATUS_REG */
5164         tmp = RREG32(SDMA0_STATUS_REG + SDMA1_REGISTER_OFFSET);
5165         if (!(tmp & SDMA_IDLE))
5166                 reset_mask |= RADEON_RESET_DMA1;
5167
5168         /* SRBM_STATUS2 */
5169         tmp = RREG32(SRBM_STATUS2);
5170         if (tmp & SDMA_BUSY)
5171                 reset_mask |= RADEON_RESET_DMA;
5172
5173         if (tmp & SDMA1_BUSY)
5174                 reset_mask |= RADEON_RESET_DMA1;
5175
5176         /* SRBM_STATUS */
5177         tmp = RREG32(SRBM_STATUS);
5178
5179         if (tmp & IH_BUSY)
5180                 reset_mask |= RADEON_RESET_IH;
5181
5182         if (tmp & SEM_BUSY)
5183                 reset_mask |= RADEON_RESET_SEM;
5184
5185         if (tmp & GRBM_RQ_PENDING)
5186                 reset_mask |= RADEON_RESET_GRBM;
5187
5188         if (tmp & VMC_BUSY)
5189                 reset_mask |= RADEON_RESET_VMC;
5190
5191         if (tmp & (MCB_BUSY | MCB_NON_DISPLAY_BUSY |
5192                    MCC_BUSY | MCD_BUSY))
5193                 reset_mask |= RADEON_RESET_MC;
5194
5195         if (evergreen_is_display_hung(rdev))
5196                 reset_mask |= RADEON_RESET_DISPLAY;
5197
5198         /* Skip MC reset as it's mostly likely not hung, just busy */
5199         if (reset_mask & RADEON_RESET_MC) {
5200                 DRM_DEBUG("MC busy: 0x%08X, clearing.\n", reset_mask);
5201                 reset_mask &= ~RADEON_RESET_MC;
5202         }
5203
5204         return reset_mask;
5205 }
5206
5207 /**
5208  * cik_gpu_soft_reset - soft reset GPU
5209  *
5210  * @rdev: radeon_device pointer
5211  * @reset_mask: mask of which blocks to reset
5212  *
5213  * Soft reset the blocks specified in @reset_mask.
5214  */
5215 static void cik_gpu_soft_reset(struct radeon_device *rdev, u32 reset_mask)
5216 {
5217         struct evergreen_mc_save save;
5218         u32 grbm_soft_reset = 0, srbm_soft_reset = 0;
5219         u32 tmp;
5220
5221         if (reset_mask == 0)
5222                 return;
5223
5224         dev_info(rdev->dev, "GPU softreset: 0x%08X\n", reset_mask);
5225
5226         cik_print_gpu_status_regs(rdev);
5227         dev_info(rdev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_ADDR   0x%08X\n",
5228                  RREG32(VM_CONTEXT1_PROTECTION_FAULT_ADDR));
5229         dev_info(rdev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
5230                  RREG32(VM_CONTEXT1_PROTECTION_FAULT_STATUS));
5231
5232         /* disable CG/PG */
5233         cik_fini_pg(rdev);
5234         cik_fini_cg(rdev);
5235
5236         /* stop the rlc */
5237         cik_rlc_stop(rdev);
5238
5239         /* Disable GFX parsing/prefetching */
5240         WREG32(CP_ME_CNTL, CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT);
5241
5242         /* Disable MEC parsing/prefetching */
5243         WREG32(CP_MEC_CNTL, MEC_ME1_HALT | MEC_ME2_HALT);
5244
5245         if (reset_mask & RADEON_RESET_DMA) {
5246                 /* sdma0 */
5247                 tmp = RREG32(SDMA0_ME_CNTL + SDMA0_REGISTER_OFFSET);
5248                 tmp |= SDMA_HALT;
5249                 WREG32(SDMA0_ME_CNTL + SDMA0_REGISTER_OFFSET, tmp);
5250         }
5251         if (reset_mask & RADEON_RESET_DMA1) {
5252                 /* sdma1 */
5253                 tmp = RREG32(SDMA0_ME_CNTL + SDMA1_REGISTER_OFFSET);
5254                 tmp |= SDMA_HALT;
5255                 WREG32(SDMA0_ME_CNTL + SDMA1_REGISTER_OFFSET, tmp);
5256         }
5257
5258         evergreen_mc_stop(rdev, &save);
5259         if (evergreen_mc_wait_for_idle(rdev)) {
5260                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
5261         }
5262
5263         if (reset_mask & (RADEON_RESET_GFX | RADEON_RESET_COMPUTE | RADEON_RESET_CP))
5264                 grbm_soft_reset = SOFT_RESET_CP | SOFT_RESET_GFX;
5265
5266         if (reset_mask & RADEON_RESET_CP) {
5267                 grbm_soft_reset |= SOFT_RESET_CP;
5268
5269                 srbm_soft_reset |= SOFT_RESET_GRBM;
5270         }
5271
5272         if (reset_mask & RADEON_RESET_DMA)
5273                 srbm_soft_reset |= SOFT_RESET_SDMA;
5274
5275         if (reset_mask & RADEON_RESET_DMA1)
5276                 srbm_soft_reset |= SOFT_RESET_SDMA1;
5277
5278         if (reset_mask & RADEON_RESET_DISPLAY)
5279                 srbm_soft_reset |= SOFT_RESET_DC;
5280
5281         if (reset_mask & RADEON_RESET_RLC)
5282                 grbm_soft_reset |= SOFT_RESET_RLC;
5283
5284         if (reset_mask & RADEON_RESET_SEM)
5285                 srbm_soft_reset |= SOFT_RESET_SEM;
5286
5287         if (reset_mask & RADEON_RESET_IH)
5288                 srbm_soft_reset |= SOFT_RESET_IH;
5289
5290         if (reset_mask & RADEON_RESET_GRBM)
5291                 srbm_soft_reset |= SOFT_RESET_GRBM;
5292
5293         if (reset_mask & RADEON_RESET_VMC)
5294                 srbm_soft_reset |= SOFT_RESET_VMC;
5295
5296         if (!(rdev->flags & RADEON_IS_IGP)) {
5297                 if (reset_mask & RADEON_RESET_MC)
5298                         srbm_soft_reset |= SOFT_RESET_MC;
5299         }
5300
5301         if (grbm_soft_reset) {
5302                 tmp = RREG32(GRBM_SOFT_RESET);
5303                 tmp |= grbm_soft_reset;
5304                 dev_info(rdev->dev, "GRBM_SOFT_RESET=0x%08X\n", tmp);
5305                 WREG32(GRBM_SOFT_RESET, tmp);
5306                 tmp = RREG32(GRBM_SOFT_RESET);
5307
5308                 udelay(50);
5309
5310                 tmp &= ~grbm_soft_reset;
5311                 WREG32(GRBM_SOFT_RESET, tmp);
5312                 tmp = RREG32(GRBM_SOFT_RESET);
5313         }
5314
5315         if (srbm_soft_reset) {
5316                 tmp = RREG32(SRBM_SOFT_RESET);
5317                 tmp |= srbm_soft_reset;
5318                 dev_info(rdev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
5319                 WREG32(SRBM_SOFT_RESET, tmp);
5320                 tmp = RREG32(SRBM_SOFT_RESET);
5321
5322                 udelay(50);
5323
5324                 tmp &= ~srbm_soft_reset;
5325                 WREG32(SRBM_SOFT_RESET, tmp);
5326                 tmp = RREG32(SRBM_SOFT_RESET);
5327         }
5328
5329         /* Wait a little for things to settle down */
5330         udelay(50);
5331
5332         evergreen_mc_resume(rdev, &save);
5333         udelay(50);
5334
5335         cik_print_gpu_status_regs(rdev);
5336 }
5337
5338 struct kv_reset_save_regs {
5339         u32 gmcon_reng_execute;
5340         u32 gmcon_misc;
5341         u32 gmcon_misc3;
5342 };
5343
5344 static void kv_save_regs_for_reset(struct radeon_device *rdev,
5345                                    struct kv_reset_save_regs *save)
5346 {
5347         save->gmcon_reng_execute = RREG32(GMCON_RENG_EXECUTE);
5348         save->gmcon_misc = RREG32(GMCON_MISC);
5349         save->gmcon_misc3 = RREG32(GMCON_MISC3);
5350
5351         WREG32(GMCON_RENG_EXECUTE, save->gmcon_reng_execute & ~RENG_EXECUTE_ON_PWR_UP);
5352         WREG32(GMCON_MISC, save->gmcon_misc & ~(RENG_EXECUTE_ON_REG_UPDATE |
5353                                                 STCTRL_STUTTER_EN));
5354 }
5355
5356 static void kv_restore_regs_for_reset(struct radeon_device *rdev,
5357                                       struct kv_reset_save_regs *save)
5358 {
5359         int i;
5360
5361         WREG32(GMCON_PGFSM_WRITE, 0);
5362         WREG32(GMCON_PGFSM_CONFIG, 0x200010ff);
5363
5364         for (i = 0; i < 5; i++)
5365                 WREG32(GMCON_PGFSM_WRITE, 0);
5366
5367         WREG32(GMCON_PGFSM_WRITE, 0);
5368         WREG32(GMCON_PGFSM_CONFIG, 0x300010ff);
5369
5370         for (i = 0; i < 5; i++)
5371                 WREG32(GMCON_PGFSM_WRITE, 0);
5372
5373         WREG32(GMCON_PGFSM_WRITE, 0x210000);
5374         WREG32(GMCON_PGFSM_CONFIG, 0xa00010ff);
5375
5376         for (i = 0; i < 5; i++)
5377                 WREG32(GMCON_PGFSM_WRITE, 0);
5378
5379         WREG32(GMCON_PGFSM_WRITE, 0x21003);
5380         WREG32(GMCON_PGFSM_CONFIG, 0xb00010ff);
5381
5382         for (i = 0; i < 5; i++)
5383                 WREG32(GMCON_PGFSM_WRITE, 0);
5384
5385         WREG32(GMCON_PGFSM_WRITE, 0x2b00);
5386         WREG32(GMCON_PGFSM_CONFIG, 0xc00010ff);
5387
5388         for (i = 0; i < 5; i++)
5389                 WREG32(GMCON_PGFSM_WRITE, 0);
5390
5391         WREG32(GMCON_PGFSM_WRITE, 0);
5392         WREG32(GMCON_PGFSM_CONFIG, 0xd00010ff);
5393
5394         for (i = 0; i < 5; i++)
5395                 WREG32(GMCON_PGFSM_WRITE, 0);
5396
5397         WREG32(GMCON_PGFSM_WRITE, 0x420000);
5398         WREG32(GMCON_PGFSM_CONFIG, 0x100010ff);
5399
5400         for (i = 0; i < 5; i++)
5401                 WREG32(GMCON_PGFSM_WRITE, 0);
5402
5403         WREG32(GMCON_PGFSM_WRITE, 0x120202);
5404         WREG32(GMCON_PGFSM_CONFIG, 0x500010ff);
5405
5406         for (i = 0; i < 5; i++)
5407                 WREG32(GMCON_PGFSM_WRITE, 0);
5408
5409         WREG32(GMCON_PGFSM_WRITE, 0x3e3e36);
5410         WREG32(GMCON_PGFSM_CONFIG, 0x600010ff);
5411
5412         for (i = 0; i < 5; i++)
5413                 WREG32(GMCON_PGFSM_WRITE, 0);
5414
5415         WREG32(GMCON_PGFSM_WRITE, 0x373f3e);
5416         WREG32(GMCON_PGFSM_CONFIG, 0x700010ff);
5417
5418         for (i = 0; i < 5; i++)
5419                 WREG32(GMCON_PGFSM_WRITE, 0);
5420
5421         WREG32(GMCON_PGFSM_WRITE, 0x3e1332);
5422         WREG32(GMCON_PGFSM_CONFIG, 0xe00010ff);
5423
5424         WREG32(GMCON_MISC3, save->gmcon_misc3);
5425         WREG32(GMCON_MISC, save->gmcon_misc);
5426         WREG32(GMCON_RENG_EXECUTE, save->gmcon_reng_execute);
5427 }
5428
5429 static void cik_gpu_pci_config_reset(struct radeon_device *rdev)
5430 {
5431         struct evergreen_mc_save save;
5432         struct kv_reset_save_regs kv_save = { 0 };
5433         u32 tmp, i;
5434
5435         dev_info(rdev->dev, "GPU pci config reset\n");
5436
5437         /* disable dpm? */
5438
5439         /* disable cg/pg */
5440         cik_fini_pg(rdev);
5441         cik_fini_cg(rdev);
5442
5443         /* Disable GFX parsing/prefetching */
5444         WREG32(CP_ME_CNTL, CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT);
5445
5446         /* Disable MEC parsing/prefetching */
5447         WREG32(CP_MEC_CNTL, MEC_ME1_HALT | MEC_ME2_HALT);
5448
5449         /* sdma0 */
5450         tmp = RREG32(SDMA0_ME_CNTL + SDMA0_REGISTER_OFFSET);
5451         tmp |= SDMA_HALT;
5452         WREG32(SDMA0_ME_CNTL + SDMA0_REGISTER_OFFSET, tmp);
5453         /* sdma1 */
5454         tmp = RREG32(SDMA0_ME_CNTL + SDMA1_REGISTER_OFFSET);
5455         tmp |= SDMA_HALT;
5456         WREG32(SDMA0_ME_CNTL + SDMA1_REGISTER_OFFSET, tmp);
5457         /* XXX other engines? */
5458
5459         /* halt the rlc, disable cp internal ints */
5460         cik_rlc_stop(rdev);
5461
5462         udelay(50);
5463
5464         /* disable mem access */
5465         evergreen_mc_stop(rdev, &save);
5466         if (evergreen_mc_wait_for_idle(rdev)) {
5467                 dev_warn(rdev->dev, "Wait for MC idle timed out !\n");
5468         }
5469
5470         if (rdev->flags & RADEON_IS_IGP)
5471                 kv_save_regs_for_reset(rdev, &kv_save);
5472
5473         /* disable BM */
5474         pci_clear_master(rdev->pdev);
5475         /* reset */
5476         radeon_pci_config_reset(rdev);
5477
5478         udelay(100);
5479
5480         /* wait for asic to come out of reset */
5481         for (i = 0; i < rdev->usec_timeout; i++) {
5482                 if (RREG32(CONFIG_MEMSIZE) != 0xffffffff)
5483                         break;
5484                 udelay(1);
5485         }
5486
5487         /* does asic init need to be run first??? */
5488         if (rdev->flags & RADEON_IS_IGP)
5489                 kv_restore_regs_for_reset(rdev, &kv_save);
5490 }
5491
5492 /**
5493  * cik_asic_reset - soft reset GPU
5494  *
5495  * @rdev: radeon_device pointer
5496  *
5497  * Look up which blocks are hung and attempt
5498  * to reset them.
5499  * Returns 0 for success.
5500  */
5501 int cik_asic_reset(struct radeon_device *rdev)
5502 {
5503         u32 reset_mask;
5504
5505         reset_mask = cik_gpu_check_soft_reset(rdev);
5506
5507         if (reset_mask)
5508                 r600_set_bios_scratch_engine_hung(rdev, true);
5509
5510         /* try soft reset */
5511         cik_gpu_soft_reset(rdev, reset_mask);
5512
5513         reset_mask = cik_gpu_check_soft_reset(rdev);
5514
5515         /* try pci config reset */
5516         if (reset_mask && radeon_hard_reset)
5517                 cik_gpu_pci_config_reset(rdev);
5518
5519         reset_mask = cik_gpu_check_soft_reset(rdev);
5520
5521         if (!reset_mask)
5522                 r600_set_bios_scratch_engine_hung(rdev, false);
5523
5524         return 0;
5525 }
5526
5527 /**
5528  * cik_gfx_is_lockup - check if the 3D engine is locked up
5529  *
5530  * @rdev: radeon_device pointer
5531  * @ring: radeon_ring structure holding ring information
5532  *
5533  * Check if the 3D engine is locked up (CIK).
5534  * Returns true if the engine is locked, false if not.
5535  */
5536 bool cik_gfx_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
5537 {
5538         u32 reset_mask = cik_gpu_check_soft_reset(rdev);
5539
5540         if (!(reset_mask & (RADEON_RESET_GFX |
5541                             RADEON_RESET_COMPUTE |
5542                             RADEON_RESET_CP))) {
5543                 radeon_ring_lockup_update(rdev, ring);
5544                 return false;
5545         }
5546         return radeon_ring_test_lockup(rdev, ring);
5547 }
5548
5549 /* MC */
5550 /**
5551  * cik_mc_program - program the GPU memory controller
5552  *
5553  * @rdev: radeon_device pointer
5554  *
5555  * Set the location of vram, gart, and AGP in the GPU's
5556  * physical address space (CIK).
5557  */
5558 static void cik_mc_program(struct radeon_device *rdev)
5559 {
5560         struct evergreen_mc_save save;
5561         u32 tmp;
5562         int i, j;
5563
5564         /* Initialize HDP */
5565         for (i = 0, j = 0; i < 32; i++, j += 0x18) {
5566                 WREG32((0x2c14 + j), 0x00000000);
5567                 WREG32((0x2c18 + j), 0x00000000);
5568                 WREG32((0x2c1c + j), 0x00000000);
5569                 WREG32((0x2c20 + j), 0x00000000);
5570                 WREG32((0x2c24 + j), 0x00000000);
5571         }
5572         WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0);
5573
5574         evergreen_mc_stop(rdev, &save);
5575         if (radeon_mc_wait_for_idle(rdev)) {
5576                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
5577         }
5578         /* Lockout access through VGA aperture*/
5579         WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
5580         /* Update configuration */
5581         WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
5582                rdev->mc.vram_start >> 12);
5583         WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
5584                rdev->mc.vram_end >> 12);
5585         WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR,
5586                rdev->vram_scratch.gpu_addr >> 12);
5587         tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
5588         tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF);
5589         WREG32(MC_VM_FB_LOCATION, tmp);
5590         /* XXX double check these! */
5591         WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8));
5592         WREG32(HDP_NONSURFACE_INFO, (2 << 7) | (1 << 30));
5593         WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF);
5594         WREG32(MC_VM_AGP_BASE, 0);
5595         WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF);
5596         WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF);
5597         if (radeon_mc_wait_for_idle(rdev)) {
5598                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
5599         }
5600         evergreen_mc_resume(rdev, &save);
5601         /* we need to own VRAM, so turn off the VGA renderer here
5602          * to stop it overwriting our objects */
5603         rv515_vga_render_disable(rdev);
5604 }
5605
5606 /**
5607  * cik_mc_init - initialize the memory controller driver params
5608  *
5609  * @rdev: radeon_device pointer
5610  *
5611  * Look up the amount of vram, vram width, and decide how to place
5612  * vram and gart within the GPU's physical address space (CIK).
5613  * Returns 0 for success.
5614  */
5615 static int cik_mc_init(struct radeon_device *rdev)
5616 {
5617         u32 tmp;
5618         int chansize, numchan;
5619
5620         /* Get VRAM informations */
5621         rdev->mc.vram_is_ddr = true;
5622         tmp = RREG32(MC_ARB_RAMCFG);
5623         if (tmp & CHANSIZE_MASK) {
5624                 chansize = 64;
5625         } else {
5626                 chansize = 32;
5627         }
5628         tmp = RREG32(MC_SHARED_CHMAP);
5629         switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
5630         case 0:
5631         default:
5632                 numchan = 1;
5633                 break;
5634         case 1:
5635                 numchan = 2;
5636                 break;
5637         case 2:
5638                 numchan = 4;
5639                 break;
5640         case 3:
5641                 numchan = 8;
5642                 break;
5643         case 4:
5644                 numchan = 3;
5645                 break;
5646         case 5:
5647                 numchan = 6;
5648                 break;
5649         case 6:
5650                 numchan = 10;
5651                 break;
5652         case 7:
5653                 numchan = 12;
5654                 break;
5655         case 8:
5656                 numchan = 16;
5657                 break;
5658         }
5659         rdev->mc.vram_width = numchan * chansize;
5660         /* Could aper size report 0 ? */
5661         rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
5662         rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
5663         /* size in MB on si */
5664         rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE) * 1024ULL * 1024ULL;
5665         rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE) * 1024ULL * 1024ULL;
5666         rdev->mc.visible_vram_size = rdev->mc.aper_size;
5667         si_vram_gtt_location(rdev, &rdev->mc);
5668         radeon_update_bandwidth_info(rdev);
5669
5670         return 0;
5671 }
5672
5673 /*
5674  * GART
5675  * VMID 0 is the physical GPU addresses as used by the kernel.
5676  * VMIDs 1-15 are used for userspace clients and are handled
5677  * by the radeon vm/hsa code.
5678  */
5679 /**
5680  * cik_pcie_gart_tlb_flush - gart tlb flush callback
5681  *
5682  * @rdev: radeon_device pointer
5683  *
5684  * Flush the TLB for the VMID 0 page table (CIK).
5685  */
5686 void cik_pcie_gart_tlb_flush(struct radeon_device *rdev)
5687 {
5688         /* flush hdp cache */
5689         WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0);
5690
5691         /* bits 0-15 are the VM contexts0-15 */
5692         WREG32(VM_INVALIDATE_REQUEST, 0x1);
5693 }
5694
5695 /**
5696  * cik_pcie_gart_enable - gart enable
5697  *
5698  * @rdev: radeon_device pointer
5699  *
5700  * This sets up the TLBs, programs the page tables for VMID0,
5701  * sets up the hw for VMIDs 1-15 which are allocated on
5702  * demand, and sets up the global locations for the LDS, GDS,
5703  * and GPUVM for FSA64 clients (CIK).
5704  * Returns 0 for success, errors for failure.
5705  */
5706 static int cik_pcie_gart_enable(struct radeon_device *rdev)
5707 {
5708         int r, i;
5709
5710         if (rdev->gart.robj == NULL) {
5711                 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
5712                 return -EINVAL;
5713         }
5714         r = radeon_gart_table_vram_pin(rdev);
5715         if (r)
5716                 return r;
5717         /* Setup TLB control */
5718         WREG32(MC_VM_MX_L1_TLB_CNTL,
5719                (0xA << 7) |
5720                ENABLE_L1_TLB |
5721                ENABLE_L1_FRAGMENT_PROCESSING |
5722                SYSTEM_ACCESS_MODE_NOT_IN_SYS |
5723                ENABLE_ADVANCED_DRIVER_MODEL |
5724                SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
5725         /* Setup L2 cache */
5726         WREG32(VM_L2_CNTL, ENABLE_L2_CACHE |
5727                ENABLE_L2_FRAGMENT_PROCESSING |
5728                ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
5729                ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
5730                EFFECTIVE_L2_QUEUE_SIZE(7) |
5731                CONTEXT1_IDENTITY_ACCESS_MODE(1));
5732         WREG32(VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS | INVALIDATE_L2_CACHE);
5733         WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
5734                BANK_SELECT(4) |
5735                L2_CACHE_BIGK_FRAGMENT_SIZE(4));
5736         /* setup context0 */
5737         WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
5738         WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
5739         WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
5740         WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
5741                         (u32)(rdev->dummy_page.addr >> 12));
5742         WREG32(VM_CONTEXT0_CNTL2, 0);
5743         WREG32(VM_CONTEXT0_CNTL, (ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
5744                                   RANGE_PROTECTION_FAULT_ENABLE_DEFAULT));
5745
5746         WREG32(0x15D4, 0);
5747         WREG32(0x15D8, 0);
5748         WREG32(0x15DC, 0);
5749
5750         /* restore context1-15 */
5751         /* set vm size, must be a multiple of 4 */
5752         WREG32(VM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
5753         WREG32(VM_CONTEXT1_PAGE_TABLE_END_ADDR, rdev->vm_manager.max_pfn);
5754         for (i = 1; i < 16; i++) {
5755                 if (i < 8)
5756                         WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (i << 2),
5757                                rdev->vm_manager.saved_table_addr[i]);
5758                 else
5759                         WREG32(VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((i - 8) << 2),
5760                                rdev->vm_manager.saved_table_addr[i]);
5761         }
5762
5763         /* enable context1-15 */
5764         WREG32(VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
5765                (u32)(rdev->dummy_page.addr >> 12));
5766         WREG32(VM_CONTEXT1_CNTL2, 4);
5767         WREG32(VM_CONTEXT1_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(1) |
5768                                 PAGE_TABLE_BLOCK_SIZE(radeon_vm_block_size - 9) |
5769                                 RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT |
5770                                 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT |
5771                                 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT |
5772                                 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT |
5773                                 PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT |
5774                                 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT |
5775                                 VALID_PROTECTION_FAULT_ENABLE_INTERRUPT |
5776                                 VALID_PROTECTION_FAULT_ENABLE_DEFAULT |
5777                                 READ_PROTECTION_FAULT_ENABLE_INTERRUPT |
5778                                 READ_PROTECTION_FAULT_ENABLE_DEFAULT |
5779                                 WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT |
5780                                 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT);
5781
5782         if (rdev->family == CHIP_KAVERI) {
5783                 u32 tmp = RREG32(CHUB_CONTROL);
5784                 tmp &= ~BYPASS_VM;
5785                 WREG32(CHUB_CONTROL, tmp);
5786         }
5787
5788         /* XXX SH_MEM regs */
5789         /* where to put LDS, scratch, GPUVM in FSA64 space */
5790         mutex_lock(&rdev->srbm_mutex);
5791         for (i = 0; i < 16; i++) {
5792                 cik_srbm_select(rdev, 0, 0, 0, i);
5793                 /* CP and shaders */
5794                 WREG32(SH_MEM_CONFIG, 0);
5795                 WREG32(SH_MEM_APE1_BASE, 1);
5796                 WREG32(SH_MEM_APE1_LIMIT, 0);
5797                 WREG32(SH_MEM_BASES, 0);
5798                 /* SDMA GFX */
5799                 WREG32(SDMA0_GFX_VIRTUAL_ADDR + SDMA0_REGISTER_OFFSET, 0);
5800                 WREG32(SDMA0_GFX_APE1_CNTL + SDMA0_REGISTER_OFFSET, 0);
5801                 WREG32(SDMA0_GFX_VIRTUAL_ADDR + SDMA1_REGISTER_OFFSET, 0);
5802                 WREG32(SDMA0_GFX_APE1_CNTL + SDMA1_REGISTER_OFFSET, 0);
5803                 /* XXX SDMA RLC - todo */
5804         }
5805         cik_srbm_select(rdev, 0, 0, 0, 0);
5806         mutex_unlock(&rdev->srbm_mutex);
5807
5808         cik_pcie_gart_tlb_flush(rdev);
5809         DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
5810                  (unsigned)(rdev->mc.gtt_size >> 20),
5811                  (unsigned long long)rdev->gart.table_addr);
5812         rdev->gart.ready = true;
5813         return 0;
5814 }
5815
5816 /**
5817  * cik_pcie_gart_disable - gart disable
5818  *
5819  * @rdev: radeon_device pointer
5820  *
5821  * This disables all VM page table (CIK).
5822  */
5823 static void cik_pcie_gart_disable(struct radeon_device *rdev)
5824 {
5825         unsigned i;
5826
5827         for (i = 1; i < 16; ++i) {
5828                 uint32_t reg;
5829                 if (i < 8)
5830                         reg = VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (i << 2);
5831                 else
5832                         reg = VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((i - 8) << 2);
5833                 rdev->vm_manager.saved_table_addr[i] = RREG32(reg);
5834         }
5835
5836         /* Disable all tables */
5837         WREG32(VM_CONTEXT0_CNTL, 0);
5838         WREG32(VM_CONTEXT1_CNTL, 0);
5839         /* Setup TLB control */
5840         WREG32(MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE_NOT_IN_SYS |
5841                SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
5842         /* Setup L2 cache */
5843         WREG32(VM_L2_CNTL,
5844                ENABLE_L2_FRAGMENT_PROCESSING |
5845                ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
5846                ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
5847                EFFECTIVE_L2_QUEUE_SIZE(7) |
5848                CONTEXT1_IDENTITY_ACCESS_MODE(1));
5849         WREG32(VM_L2_CNTL2, 0);
5850         WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
5851                L2_CACHE_BIGK_FRAGMENT_SIZE(6));
5852         radeon_gart_table_vram_unpin(rdev);
5853 }
5854
5855 /**
5856  * cik_pcie_gart_fini - vm fini callback
5857  *
5858  * @rdev: radeon_device pointer
5859  *
5860  * Tears down the driver GART/VM setup (CIK).
5861  */
5862 static void cik_pcie_gart_fini(struct radeon_device *rdev)
5863 {
5864         cik_pcie_gart_disable(rdev);
5865         radeon_gart_table_vram_free(rdev);
5866         radeon_gart_fini(rdev);
5867 }
5868
5869 /* vm parser */
5870 /**
5871  * cik_ib_parse - vm ib_parse callback
5872  *
5873  * @rdev: radeon_device pointer
5874  * @ib: indirect buffer pointer
5875  *
5876  * CIK uses hw IB checking so this is a nop (CIK).
5877  */
5878 int cik_ib_parse(struct radeon_device *rdev, struct radeon_ib *ib)
5879 {
5880         return 0;
5881 }
5882
5883 /*
5884  * vm
5885  * VMID 0 is the physical GPU addresses as used by the kernel.
5886  * VMIDs 1-15 are used for userspace clients and are handled
5887  * by the radeon vm/hsa code.
5888  */
5889 /**
5890  * cik_vm_init - cik vm init callback
5891  *
5892  * @rdev: radeon_device pointer
5893  *
5894  * Inits cik specific vm parameters (number of VMs, base of vram for
5895  * VMIDs 1-15) (CIK).
5896  * Returns 0 for success.
5897  */
5898 int cik_vm_init(struct radeon_device *rdev)
5899 {
5900         /* number of VMs */
5901         rdev->vm_manager.nvm = 16;
5902         /* base offset of vram pages */
5903         if (rdev->flags & RADEON_IS_IGP) {
5904                 u64 tmp = RREG32(MC_VM_FB_OFFSET);
5905                 tmp <<= 22;
5906                 rdev->vm_manager.vram_base_offset = tmp;
5907         } else
5908                 rdev->vm_manager.vram_base_offset = 0;
5909
5910         return 0;
5911 }
5912
5913 /**
5914  * cik_vm_fini - cik vm fini callback
5915  *
5916  * @rdev: radeon_device pointer
5917  *
5918  * Tear down any asic specific VM setup (CIK).
5919  */
5920 void cik_vm_fini(struct radeon_device *rdev)
5921 {
5922 }
5923
5924 /**
5925  * cik_vm_decode_fault - print human readable fault info
5926  *
5927  * @rdev: radeon_device pointer
5928  * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value
5929  * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value
5930  *
5931  * Print human readable fault information (CIK).
5932  */
5933 static void cik_vm_decode_fault(struct radeon_device *rdev,
5934                                 u32 status, u32 addr, u32 mc_client)
5935 {
5936         u32 mc_id;
5937         u32 vmid = (status & FAULT_VMID_MASK) >> FAULT_VMID_SHIFT;
5938         u32 protections = (status & PROTECTIONS_MASK) >> PROTECTIONS_SHIFT;
5939         char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff,
5940                 (mc_client >> 8) & 0xff, mc_client & 0xff, 0 };
5941
5942         if (rdev->family == CHIP_HAWAII)
5943                 mc_id = (status & HAWAII_MEMORY_CLIENT_ID_MASK) >> MEMORY_CLIENT_ID_SHIFT;
5944         else
5945                 mc_id = (status & MEMORY_CLIENT_ID_MASK) >> MEMORY_CLIENT_ID_SHIFT;
5946
5947         printk("VM fault (0x%02x, vmid %d) at page %u, %s from '%s' (0x%08x) (%d)\n",
5948                protections, vmid, addr,
5949                (status & MEMORY_CLIENT_RW_MASK) ? "write" : "read",
5950                block, mc_client, mc_id);
5951 }
5952
5953 /**
5954  * cik_vm_flush - cik vm flush using the CP
5955  *
5956  * @rdev: radeon_device pointer
5957  *
5958  * Update the page table base and flush the VM TLB
5959  * using the CP (CIK).
5960  */
5961 void cik_vm_flush(struct radeon_device *rdev, int ridx, struct radeon_vm *vm)
5962 {
5963         struct radeon_ring *ring = &rdev->ring[ridx];
5964         int usepfp = (ridx == RADEON_RING_TYPE_GFX_INDEX);
5965
5966         if (vm == NULL)
5967                 return;
5968
5969         radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
5970         radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(usepfp) |
5971                                  WRITE_DATA_DST_SEL(0)));
5972         if (vm->id < 8) {
5973                 radeon_ring_write(ring,
5974                                   (VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2)) >> 2);
5975         } else {
5976                 radeon_ring_write(ring,
5977                                   (VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((vm->id - 8) << 2)) >> 2);
5978         }
5979         radeon_ring_write(ring, 0);
5980         radeon_ring_write(ring, vm->pd_gpu_addr >> 12);
5981
5982         /* update SH_MEM_* regs */
5983         radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
5984         radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(usepfp) |
5985                                  WRITE_DATA_DST_SEL(0)));
5986         radeon_ring_write(ring, SRBM_GFX_CNTL >> 2);
5987         radeon_ring_write(ring, 0);
5988         radeon_ring_write(ring, VMID(vm->id));
5989
5990         radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 6));
5991         radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(usepfp) |
5992                                  WRITE_DATA_DST_SEL(0)));
5993         radeon_ring_write(ring, SH_MEM_BASES >> 2);
5994         radeon_ring_write(ring, 0);
5995
5996         radeon_ring_write(ring, 0); /* SH_MEM_BASES */
5997         radeon_ring_write(ring, 0); /* SH_MEM_CONFIG */
5998         radeon_ring_write(ring, 1); /* SH_MEM_APE1_BASE */
5999         radeon_ring_write(ring, 0); /* SH_MEM_APE1_LIMIT */
6000
6001         radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
6002         radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(usepfp) |
6003                                  WRITE_DATA_DST_SEL(0)));
6004         radeon_ring_write(ring, SRBM_GFX_CNTL >> 2);
6005         radeon_ring_write(ring, 0);
6006         radeon_ring_write(ring, VMID(0));
6007
6008         /* HDP flush */
6009         cik_hdp_flush_cp_ring_emit(rdev, ridx);
6010
6011         /* bits 0-15 are the VM contexts0-15 */
6012         radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
6013         radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(usepfp) |
6014                                  WRITE_DATA_DST_SEL(0)));
6015         radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
6016         radeon_ring_write(ring, 0);
6017         radeon_ring_write(ring, 1 << vm->id);
6018
6019         /* compute doesn't have PFP */
6020         if (usepfp) {
6021                 /* sync PFP to ME, otherwise we might get invalid PFP reads */
6022                 radeon_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
6023                 radeon_ring_write(ring, 0x0);
6024         }
6025 }
6026
6027 /*
6028  * RLC
6029  * The RLC is a multi-purpose microengine that handles a
6030  * variety of functions, the most important of which is
6031  * the interrupt controller.
6032  */
6033 static void cik_enable_gui_idle_interrupt(struct radeon_device *rdev,
6034                                           bool enable)
6035 {
6036         u32 tmp = RREG32(CP_INT_CNTL_RING0);
6037
6038         if (enable)
6039                 tmp |= (CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
6040         else
6041                 tmp &= ~(CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
6042         WREG32(CP_INT_CNTL_RING0, tmp);
6043 }
6044
6045 static void cik_enable_lbpw(struct radeon_device *rdev, bool enable)
6046 {
6047         u32 tmp;
6048
6049         tmp = RREG32(RLC_LB_CNTL);
6050         if (enable)
6051                 tmp |= LOAD_BALANCE_ENABLE;
6052         else
6053                 tmp &= ~LOAD_BALANCE_ENABLE;
6054         WREG32(RLC_LB_CNTL, tmp);
6055 }
6056
6057 static void cik_wait_for_rlc_serdes(struct radeon_device *rdev)
6058 {
6059         u32 i, j, k;
6060         u32 mask;
6061
6062         for (i = 0; i < rdev->config.cik.max_shader_engines; i++) {
6063                 for (j = 0; j < rdev->config.cik.max_sh_per_se; j++) {
6064                         cik_select_se_sh(rdev, i, j);
6065                         for (k = 0; k < rdev->usec_timeout; k++) {
6066                                 if (RREG32(RLC_SERDES_CU_MASTER_BUSY) == 0)
6067                                         break;
6068                                 udelay(1);
6069                         }
6070                 }
6071         }
6072         cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
6073
6074         mask = SE_MASTER_BUSY_MASK | GC_MASTER_BUSY | TC0_MASTER_BUSY | TC1_MASTER_BUSY;
6075         for (k = 0; k < rdev->usec_timeout; k++) {
6076                 if ((RREG32(RLC_SERDES_NONCU_MASTER_BUSY) & mask) == 0)
6077                         break;
6078                 udelay(1);
6079         }
6080 }
6081
6082 static void cik_update_rlc(struct radeon_device *rdev, u32 rlc)
6083 {
6084         u32 tmp;
6085
6086         tmp = RREG32(RLC_CNTL);
6087         if (tmp != rlc)
6088                 WREG32(RLC_CNTL, rlc);
6089 }
6090
6091 static u32 cik_halt_rlc(struct radeon_device *rdev)
6092 {
6093         u32 data, orig;
6094
6095         orig = data = RREG32(RLC_CNTL);
6096
6097         if (data & RLC_ENABLE) {
6098                 u32 i;
6099
6100                 data &= ~RLC_ENABLE;
6101                 WREG32(RLC_CNTL, data);
6102
6103                 for (i = 0; i < rdev->usec_timeout; i++) {
6104                         if ((RREG32(RLC_GPM_STAT) & RLC_GPM_BUSY) == 0)
6105                                 break;
6106                         udelay(1);
6107                 }
6108
6109                 cik_wait_for_rlc_serdes(rdev);
6110         }
6111
6112         return orig;
6113 }
6114
6115 void cik_enter_rlc_safe_mode(struct radeon_device *rdev)
6116 {
6117         u32 tmp, i, mask;
6118
6119         tmp = REQ | MESSAGE(MSG_ENTER_RLC_SAFE_MODE);
6120         WREG32(RLC_GPR_REG2, tmp);
6121
6122         mask = GFX_POWER_STATUS | GFX_CLOCK_STATUS;
6123         for (i = 0; i < rdev->usec_timeout; i++) {
6124                 if ((RREG32(RLC_GPM_STAT) & mask) == mask)
6125                         break;
6126                 udelay(1);
6127         }
6128
6129         for (i = 0; i < rdev->usec_timeout; i++) {
6130                 if ((RREG32(RLC_GPR_REG2) & REQ) == 0)
6131                         break;
6132                 udelay(1);
6133         }
6134 }
6135
6136 void cik_exit_rlc_safe_mode(struct radeon_device *rdev)
6137 {
6138         u32 tmp;
6139
6140         tmp = REQ | MESSAGE(MSG_EXIT_RLC_SAFE_MODE);
6141         WREG32(RLC_GPR_REG2, tmp);
6142 }
6143
6144 /**
6145  * cik_rlc_stop - stop the RLC ME
6146  *
6147  * @rdev: radeon_device pointer
6148  *
6149  * Halt the RLC ME (MicroEngine) (CIK).
6150  */
6151 static void cik_rlc_stop(struct radeon_device *rdev)
6152 {
6153         WREG32(RLC_CNTL, 0);
6154
6155         cik_enable_gui_idle_interrupt(rdev, false);
6156
6157         cik_wait_for_rlc_serdes(rdev);
6158 }
6159
6160 /**
6161  * cik_rlc_start - start the RLC ME
6162  *
6163  * @rdev: radeon_device pointer
6164  *
6165  * Unhalt the RLC ME (MicroEngine) (CIK).
6166  */
6167 static void cik_rlc_start(struct radeon_device *rdev)
6168 {
6169         WREG32(RLC_CNTL, RLC_ENABLE);
6170
6171         cik_enable_gui_idle_interrupt(rdev, true);
6172
6173         udelay(50);
6174 }
6175
6176 /**
6177  * cik_rlc_resume - setup the RLC hw
6178  *
6179  * @rdev: radeon_device pointer
6180  *
6181  * Initialize the RLC registers, load the ucode,
6182  * and start the RLC (CIK).
6183  * Returns 0 for success, -EINVAL if the ucode is not available.
6184  */
6185 static int cik_rlc_resume(struct radeon_device *rdev)
6186 {
6187         u32 i, size, tmp;
6188
6189         if (!rdev->rlc_fw)
6190                 return -EINVAL;
6191
6192         cik_rlc_stop(rdev);
6193
6194         /* disable CG */
6195         tmp = RREG32(RLC_CGCG_CGLS_CTRL) & 0xfffffffc;
6196         WREG32(RLC_CGCG_CGLS_CTRL, tmp);
6197
6198         si_rlc_reset(rdev);
6199
6200         cik_init_pg(rdev);
6201
6202         cik_init_cg(rdev);
6203
6204         WREG32(RLC_LB_CNTR_INIT, 0);
6205         WREG32(RLC_LB_CNTR_MAX, 0x00008000);
6206
6207         cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
6208         WREG32(RLC_LB_INIT_CU_MASK, 0xffffffff);
6209         WREG32(RLC_LB_PARAMS, 0x00600408);
6210         WREG32(RLC_LB_CNTL, 0x80000004);
6211
6212         WREG32(RLC_MC_CNTL, 0);
6213         WREG32(RLC_UCODE_CNTL, 0);
6214
6215         if (rdev->new_fw) {
6216                 const struct rlc_firmware_header_v1_0 *hdr =
6217                         (const struct rlc_firmware_header_v1_0 *)rdev->rlc_fw->data;
6218                 const __le32 *fw_data = (const __le32 *)
6219                         (rdev->rlc_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
6220
6221                 radeon_ucode_print_rlc_hdr(&hdr->header);
6222
6223                 size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
6224                 WREG32(RLC_GPM_UCODE_ADDR, 0);
6225                 for (i = 0; i < size; i++)
6226                         WREG32(RLC_GPM_UCODE_DATA, le32_to_cpup(fw_data++));
6227                 WREG32(RLC_GPM_UCODE_ADDR, le32_to_cpu(hdr->header.ucode_version));
6228         } else {
6229                 const __be32 *fw_data;
6230
6231                 switch (rdev->family) {
6232                 case CHIP_BONAIRE:
6233                 case CHIP_HAWAII:
6234                 default:
6235                         size = BONAIRE_RLC_UCODE_SIZE;
6236                         break;
6237                 case CHIP_KAVERI:
6238                         size = KV_RLC_UCODE_SIZE;
6239                         break;
6240                 case CHIP_KABINI:
6241                         size = KB_RLC_UCODE_SIZE;
6242                         break;
6243                 case CHIP_MULLINS:
6244                         size = ML_RLC_UCODE_SIZE;
6245                         break;
6246                 }
6247
6248                 fw_data = (const __be32 *)rdev->rlc_fw->data;
6249                 WREG32(RLC_GPM_UCODE_ADDR, 0);
6250                 for (i = 0; i < size; i++)
6251                         WREG32(RLC_GPM_UCODE_DATA, be32_to_cpup(fw_data++));
6252                 WREG32(RLC_GPM_UCODE_ADDR, 0);
6253         }
6254
6255         /* XXX - find out what chips support lbpw */
6256         cik_enable_lbpw(rdev, false);
6257
6258         if (rdev->family == CHIP_BONAIRE)
6259                 WREG32(RLC_DRIVER_DMA_STATUS, 0);
6260
6261         cik_rlc_start(rdev);
6262
6263         return 0;
6264 }
6265
6266 static void cik_enable_cgcg(struct radeon_device *rdev, bool enable)
6267 {
6268         u32 data, orig, tmp, tmp2;
6269
6270         orig = data = RREG32(RLC_CGCG_CGLS_CTRL);
6271
6272         if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_CGCG)) {
6273                 cik_enable_gui_idle_interrupt(rdev, true);
6274
6275                 tmp = cik_halt_rlc(rdev);
6276
6277                 cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
6278                 WREG32(RLC_SERDES_WR_CU_MASTER_MASK, 0xffffffff);
6279                 WREG32(RLC_SERDES_WR_NONCU_MASTER_MASK, 0xffffffff);
6280                 tmp2 = BPM_ADDR_MASK | CGCG_OVERRIDE_0 | CGLS_ENABLE;
6281                 WREG32(RLC_SERDES_WR_CTRL, tmp2);
6282
6283                 cik_update_rlc(rdev, tmp);
6284
6285                 data |= CGCG_EN | CGLS_EN;
6286         } else {
6287                 cik_enable_gui_idle_interrupt(rdev, false);
6288
6289                 RREG32(CB_CGTT_SCLK_CTRL);
6290                 RREG32(CB_CGTT_SCLK_CTRL);
6291                 RREG32(CB_CGTT_SCLK_CTRL);
6292                 RREG32(CB_CGTT_SCLK_CTRL);
6293
6294                 data &= ~(CGCG_EN | CGLS_EN);
6295         }
6296
6297         if (orig != data)
6298                 WREG32(RLC_CGCG_CGLS_CTRL, data);
6299
6300 }
6301
6302 static void cik_enable_mgcg(struct radeon_device *rdev, bool enable)
6303 {
6304         u32 data, orig, tmp = 0;
6305
6306         if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_MGCG)) {
6307                 if (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_MGLS) {
6308                         if (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_CP_LS) {
6309                                 orig = data = RREG32(CP_MEM_SLP_CNTL);
6310                                 data |= CP_MEM_LS_EN;
6311                                 if (orig != data)
6312                                         WREG32(CP_MEM_SLP_CNTL, data);
6313                         }
6314                 }
6315
6316                 orig = data = RREG32(RLC_CGTT_MGCG_OVERRIDE);
6317                 data &= 0xfffffffd;
6318                 if (orig != data)
6319                         WREG32(RLC_CGTT_MGCG_OVERRIDE, data);
6320
6321                 tmp = cik_halt_rlc(rdev);
6322
6323                 cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
6324                 WREG32(RLC_SERDES_WR_CU_MASTER_MASK, 0xffffffff);
6325                 WREG32(RLC_SERDES_WR_NONCU_MASTER_MASK, 0xffffffff);
6326                 data = BPM_ADDR_MASK | MGCG_OVERRIDE_0;
6327                 WREG32(RLC_SERDES_WR_CTRL, data);
6328
6329                 cik_update_rlc(rdev, tmp);
6330
6331                 if (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_CGTS) {
6332                         orig = data = RREG32(CGTS_SM_CTRL_REG);
6333                         data &= ~SM_MODE_MASK;
6334                         data |= SM_MODE(0x2);
6335                         data |= SM_MODE_ENABLE;
6336                         data &= ~CGTS_OVERRIDE;
6337                         if ((rdev->cg_flags & RADEON_CG_SUPPORT_GFX_MGLS) &&
6338                             (rdev->cg_flags & RADEON_CG_SUPPORT_GFX_CGTS_LS))
6339                                 data &= ~CGTS_LS_OVERRIDE;
6340                         data &= ~ON_MONITOR_ADD_MASK;
6341                         data |= ON_MONITOR_ADD_EN;
6342                         data |= ON_MONITOR_ADD(0x96);
6343                         if (orig != data)
6344                                 WREG32(CGTS_SM_CTRL_REG, data);
6345                 }
6346         } else {
6347                 orig = data = RREG32(RLC_CGTT_MGCG_OVERRIDE);
6348                 data |= 0x00000002;
6349                 if (orig != data)
6350                         WREG32(RLC_CGTT_MGCG_OVERRIDE, data);
6351
6352                 data = RREG32(RLC_MEM_SLP_CNTL);
6353                 if (data & RLC_MEM_LS_EN) {
6354                         data &= ~RLC_MEM_LS_EN;
6355                         WREG32(RLC_MEM_SLP_CNTL, data);
6356                 }
6357
6358                 data = RREG32(CP_MEM_SLP_CNTL);
6359                 if (data & CP_MEM_LS_EN) {
6360                         data &= ~CP_MEM_LS_EN;
6361                         WREG32(CP_MEM_SLP_CNTL, data);
6362                 }
6363
6364                 orig = data = RREG32(CGTS_SM_CTRL_REG);
6365                 data |= CGTS_OVERRIDE | CGTS_LS_OVERRIDE;
6366                 if (orig != data)
6367                         WREG32(CGTS_SM_CTRL_REG, data);
6368
6369                 tmp = cik_halt_rlc(rdev);
6370
6371                 cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
6372                 WREG32(RLC_SERDES_WR_CU_MASTER_MASK, 0xffffffff);
6373                 WREG32(RLC_SERDES_WR_NONCU_MASTER_MASK, 0xffffffff);
6374                 data = BPM_ADDR_MASK | MGCG_OVERRIDE_1;
6375                 WREG32(RLC_SERDES_WR_CTRL, data);
6376
6377                 cik_update_rlc(rdev, tmp);
6378         }
6379 }
6380
6381 static const u32 mc_cg_registers[] =
6382 {
6383         MC_HUB_MISC_HUB_CG,
6384         MC_HUB_MISC_SIP_CG,
6385         MC_HUB_MISC_VM_CG,
6386         MC_XPB_CLK_GAT,
6387         ATC_MISC_CG,
6388         MC_CITF_MISC_WR_CG,
6389         MC_CITF_MISC_RD_CG,
6390         MC_CITF_MISC_VM_CG,
6391         VM_L2_CG,
6392 };
6393
6394 static void cik_enable_mc_ls(struct radeon_device *rdev,
6395                              bool enable)
6396 {
6397         int i;
6398         u32 orig, data;
6399
6400         for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
6401                 orig = data = RREG32(mc_cg_registers[i]);
6402                 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_MC_LS))
6403                         data |= MC_LS_ENABLE;
6404                 else
6405                         data &= ~MC_LS_ENABLE;
6406                 if (data != orig)
6407                         WREG32(mc_cg_registers[i], data);
6408         }
6409 }
6410
6411 static void cik_enable_mc_mgcg(struct radeon_device *rdev,
6412                                bool enable)
6413 {
6414         int i;
6415         u32 orig, data;
6416
6417         for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
6418                 orig = data = RREG32(mc_cg_registers[i]);
6419                 if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_MC_MGCG))
6420                         data |= MC_CG_ENABLE;
6421                 else
6422                         data &= ~MC_CG_ENABLE;
6423                 if (data != orig)
6424                         WREG32(mc_cg_registers[i], data);
6425         }
6426 }
6427
6428 static void cik_enable_sdma_mgcg(struct radeon_device *rdev,
6429                                  bool enable)
6430 {
6431         u32 orig, data;
6432
6433         if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_SDMA_MGCG)) {
6434                 WREG32(SDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, 0x00000100);
6435                 WREG32(SDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, 0x00000100);
6436         } else {
6437                 orig = data = RREG32(SDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET);
6438                 data |= 0xff000000;
6439                 if (data != orig)
6440                         WREG32(SDMA0_CLK_CTRL + SDMA0_REGISTER_OFFSET, data);
6441
6442                 orig = data = RREG32(SDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET);
6443                 data |= 0xff000000;
6444                 if (data != orig)
6445                         WREG32(SDMA0_CLK_CTRL + SDMA1_REGISTER_OFFSET, data);
6446         }
6447 }
6448
6449 static void cik_enable_sdma_mgls(struct radeon_device *rdev,
6450                                  bool enable)
6451 {
6452         u32 orig, data;
6453
6454         if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_SDMA_LS)) {
6455                 orig = data = RREG32(SDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
6456                 data |= 0x100;
6457                 if (orig != data)
6458                         WREG32(SDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
6459
6460                 orig = data = RREG32(SDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
6461                 data |= 0x100;
6462                 if (orig != data)
6463                         WREG32(SDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
6464         } else {
6465                 orig = data = RREG32(SDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET);
6466                 data &= ~0x100;
6467                 if (orig != data)
6468                         WREG32(SDMA0_POWER_CNTL + SDMA0_REGISTER_OFFSET, data);
6469
6470                 orig = data = RREG32(SDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET);
6471                 data &= ~0x100;
6472                 if (orig != data)
6473                         WREG32(SDMA0_POWER_CNTL + SDMA1_REGISTER_OFFSET, data);
6474         }
6475 }
6476
6477 static void cik_enable_uvd_mgcg(struct radeon_device *rdev,
6478                                 bool enable)
6479 {
6480         u32 orig, data;
6481
6482         if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_UVD_MGCG)) {
6483                 data = RREG32_UVD_CTX(UVD_CGC_MEM_CTRL);
6484                 data = 0xfff;
6485                 WREG32_UVD_CTX(UVD_CGC_MEM_CTRL, data);
6486
6487                 orig = data = RREG32(UVD_CGC_CTRL);
6488                 data |= DCM;
6489                 if (orig != data)
6490                         WREG32(UVD_CGC_CTRL, data);
6491         } else {
6492                 data = RREG32_UVD_CTX(UVD_CGC_MEM_CTRL);
6493                 data &= ~0xfff;
6494                 WREG32_UVD_CTX(UVD_CGC_MEM_CTRL, data);
6495
6496                 orig = data = RREG32(UVD_CGC_CTRL);
6497                 data &= ~DCM;
6498                 if (orig != data)
6499                         WREG32(UVD_CGC_CTRL, data);
6500         }
6501 }
6502
6503 static void cik_enable_bif_mgls(struct radeon_device *rdev,
6504                                bool enable)
6505 {
6506         u32 orig, data;
6507
6508         orig = data = RREG32_PCIE_PORT(PCIE_CNTL2);
6509
6510         if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_BIF_LS))
6511                 data |= SLV_MEM_LS_EN | MST_MEM_LS_EN |
6512                         REPLAY_MEM_LS_EN | SLV_MEM_AGGRESSIVE_LS_EN;
6513         else
6514                 data &= ~(SLV_MEM_LS_EN | MST_MEM_LS_EN |
6515                           REPLAY_MEM_LS_EN | SLV_MEM_AGGRESSIVE_LS_EN);
6516
6517         if (orig != data)
6518                 WREG32_PCIE_PORT(PCIE_CNTL2, data);
6519 }
6520
6521 static void cik_enable_hdp_mgcg(struct radeon_device *rdev,
6522                                 bool enable)
6523 {
6524         u32 orig, data;
6525
6526         orig = data = RREG32(HDP_HOST_PATH_CNTL);
6527
6528         if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_HDP_MGCG))
6529                 data &= ~CLOCK_GATING_DIS;
6530         else
6531                 data |= CLOCK_GATING_DIS;
6532
6533         if (orig != data)
6534                 WREG32(HDP_HOST_PATH_CNTL, data);
6535 }
6536
6537 static void cik_enable_hdp_ls(struct radeon_device *rdev,
6538                               bool enable)
6539 {
6540         u32 orig, data;
6541
6542         orig = data = RREG32(HDP_MEM_POWER_LS);
6543
6544         if (enable && (rdev->cg_flags & RADEON_CG_SUPPORT_HDP_LS))
6545                 data |= HDP_LS_ENABLE;
6546         else
6547                 data &= ~HDP_LS_ENABLE;
6548
6549         if (orig != data)
6550                 WREG32(HDP_MEM_POWER_LS, data);
6551 }
6552
6553 void cik_update_cg(struct radeon_device *rdev,
6554                    u32 block, bool enable)
6555 {
6556
6557         if (block & RADEON_CG_BLOCK_GFX) {
6558                 cik_enable_gui_idle_interrupt(rdev, false);
6559                 /* order matters! */
6560                 if (enable) {
6561                         cik_enable_mgcg(rdev, true);
6562                         cik_enable_cgcg(rdev, true);
6563                 } else {
6564                         cik_enable_cgcg(rdev, false);
6565                         cik_enable_mgcg(rdev, false);
6566                 }
6567                 cik_enable_gui_idle_interrupt(rdev, true);
6568         }
6569
6570         if (block & RADEON_CG_BLOCK_MC) {
6571                 if (!(rdev->flags & RADEON_IS_IGP)) {
6572                         cik_enable_mc_mgcg(rdev, enable);
6573                         cik_enable_mc_ls(rdev, enable);
6574                 }
6575         }
6576
6577         if (block & RADEON_CG_BLOCK_SDMA) {
6578                 cik_enable_sdma_mgcg(rdev, enable);
6579                 cik_enable_sdma_mgls(rdev, enable);
6580         }
6581
6582         if (block & RADEON_CG_BLOCK_BIF) {
6583                 cik_enable_bif_mgls(rdev, enable);
6584         }
6585
6586         if (block & RADEON_CG_BLOCK_UVD) {
6587                 if (rdev->has_uvd)
6588                         cik_enable_uvd_mgcg(rdev, enable);
6589         }
6590
6591         if (block & RADEON_CG_BLOCK_HDP) {
6592                 cik_enable_hdp_mgcg(rdev, enable);
6593                 cik_enable_hdp_ls(rdev, enable);
6594         }
6595
6596         if (block & RADEON_CG_BLOCK_VCE) {
6597                 vce_v2_0_enable_mgcg(rdev, enable);
6598         }
6599 }
6600
6601 static void cik_init_cg(struct radeon_device *rdev)
6602 {
6603
6604         cik_update_cg(rdev, RADEON_CG_BLOCK_GFX, true);
6605
6606         if (rdev->has_uvd)
6607                 si_init_uvd_internal_cg(rdev);
6608
6609         cik_update_cg(rdev, (RADEON_CG_BLOCK_MC |
6610                              RADEON_CG_BLOCK_SDMA |
6611                              RADEON_CG_BLOCK_BIF |
6612                              RADEON_CG_BLOCK_UVD |
6613                              RADEON_CG_BLOCK_HDP), true);
6614 }
6615
6616 static void cik_fini_cg(struct radeon_device *rdev)
6617 {
6618         cik_update_cg(rdev, (RADEON_CG_BLOCK_MC |
6619                              RADEON_CG_BLOCK_SDMA |
6620                              RADEON_CG_BLOCK_BIF |
6621                              RADEON_CG_BLOCK_UVD |
6622                              RADEON_CG_BLOCK_HDP), false);
6623
6624         cik_update_cg(rdev, RADEON_CG_BLOCK_GFX, false);
6625 }
6626
6627 static void cik_enable_sck_slowdown_on_pu(struct radeon_device *rdev,
6628                                           bool enable)
6629 {
6630         u32 data, orig;
6631
6632         orig = data = RREG32(RLC_PG_CNTL);
6633         if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_RLC_SMU_HS))
6634                 data |= SMU_CLK_SLOWDOWN_ON_PU_ENABLE;
6635         else
6636                 data &= ~SMU_CLK_SLOWDOWN_ON_PU_ENABLE;
6637         if (orig != data)
6638                 WREG32(RLC_PG_CNTL, data);
6639 }
6640
6641 static void cik_enable_sck_slowdown_on_pd(struct radeon_device *rdev,
6642                                           bool enable)
6643 {
6644         u32 data, orig;
6645
6646         orig = data = RREG32(RLC_PG_CNTL);
6647         if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_RLC_SMU_HS))
6648                 data |= SMU_CLK_SLOWDOWN_ON_PD_ENABLE;
6649         else
6650                 data &= ~SMU_CLK_SLOWDOWN_ON_PD_ENABLE;
6651         if (orig != data)
6652                 WREG32(RLC_PG_CNTL, data);
6653 }
6654
6655 static void cik_enable_cp_pg(struct radeon_device *rdev, bool enable)
6656 {
6657         u32 data, orig;
6658
6659         orig = data = RREG32(RLC_PG_CNTL);
6660         if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_CP))
6661                 data &= ~DISABLE_CP_PG;
6662         else
6663                 data |= DISABLE_CP_PG;
6664         if (orig != data)
6665                 WREG32(RLC_PG_CNTL, data);
6666 }
6667
6668 static void cik_enable_gds_pg(struct radeon_device *rdev, bool enable)
6669 {
6670         u32 data, orig;
6671
6672         orig = data = RREG32(RLC_PG_CNTL);
6673         if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_GDS))
6674                 data &= ~DISABLE_GDS_PG;
6675         else
6676                 data |= DISABLE_GDS_PG;
6677         if (orig != data)
6678                 WREG32(RLC_PG_CNTL, data);
6679 }
6680
6681 #define CP_ME_TABLE_SIZE    96
6682 #define CP_ME_TABLE_OFFSET  2048
6683 #define CP_MEC_TABLE_OFFSET 4096
6684
6685 void cik_init_cp_pg_table(struct radeon_device *rdev)
6686 {
6687         volatile u32 *dst_ptr;
6688         int me, i, max_me = 4;
6689         u32 bo_offset = 0;
6690         u32 table_offset, table_size;
6691
6692         if (rdev->family == CHIP_KAVERI)
6693                 max_me = 5;
6694
6695         if (rdev->rlc.cp_table_ptr == NULL)
6696                 return;
6697
6698         /* write the cp table buffer */
6699         dst_ptr = rdev->rlc.cp_table_ptr;
6700         for (me = 0; me < max_me; me++) {
6701                 if (rdev->new_fw) {
6702                         const __le32 *fw_data;
6703                         const struct gfx_firmware_header_v1_0 *hdr;
6704
6705                         if (me == 0) {
6706                                 hdr = (const struct gfx_firmware_header_v1_0 *)rdev->ce_fw->data;
6707                                 fw_data = (const __le32 *)
6708                                         (rdev->ce_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
6709                                 table_offset = le32_to_cpu(hdr->jt_offset);
6710                                 table_size = le32_to_cpu(hdr->jt_size);
6711                         } else if (me == 1) {
6712                                 hdr = (const struct gfx_firmware_header_v1_0 *)rdev->pfp_fw->data;
6713                                 fw_data = (const __le32 *)
6714                                         (rdev->pfp_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
6715                                 table_offset = le32_to_cpu(hdr->jt_offset);
6716                                 table_size = le32_to_cpu(hdr->jt_size);
6717                         } else if (me == 2) {
6718                                 hdr = (const struct gfx_firmware_header_v1_0 *)rdev->me_fw->data;
6719                                 fw_data = (const __le32 *)
6720                                         (rdev->me_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
6721                                 table_offset = le32_to_cpu(hdr->jt_offset);
6722                                 table_size = le32_to_cpu(hdr->jt_size);
6723                         } else if (me == 3) {
6724                                 hdr = (const struct gfx_firmware_header_v1_0 *)rdev->mec_fw->data;
6725                                 fw_data = (const __le32 *)
6726                                         (rdev->mec_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
6727                                 table_offset = le32_to_cpu(hdr->jt_offset);
6728                                 table_size = le32_to_cpu(hdr->jt_size);
6729                         } else {
6730                                 hdr = (const struct gfx_firmware_header_v1_0 *)rdev->mec2_fw->data;
6731                                 fw_data = (const __le32 *)
6732                                         (rdev->mec2_fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
6733                                 table_offset = le32_to_cpu(hdr->jt_offset);
6734                                 table_size = le32_to_cpu(hdr->jt_size);
6735                         }
6736
6737                         for (i = 0; i < table_size; i ++) {
6738                                 dst_ptr[bo_offset + i] =
6739                                         cpu_to_le32(le32_to_cpu(fw_data[table_offset + i]));
6740                         }
6741                         bo_offset += table_size;
6742                 } else {
6743                         const __be32 *fw_data;
6744                         table_size = CP_ME_TABLE_SIZE;
6745
6746                         if (me == 0) {
6747                                 fw_data = (const __be32 *)rdev->ce_fw->data;
6748                                 table_offset = CP_ME_TABLE_OFFSET;
6749                         } else if (me == 1) {
6750                                 fw_data = (const __be32 *)rdev->pfp_fw->data;
6751                                 table_offset = CP_ME_TABLE_OFFSET;
6752                         } else if (me == 2) {
6753                                 fw_data = (const __be32 *)rdev->me_fw->data;
6754                                 table_offset = CP_ME_TABLE_OFFSET;
6755                         } else {
6756                                 fw_data = (const __be32 *)rdev->mec_fw->data;
6757                                 table_offset = CP_MEC_TABLE_OFFSET;
6758                         }
6759
6760                         for (i = 0; i < table_size; i ++) {
6761                                 dst_ptr[bo_offset + i] =
6762                                         cpu_to_le32(be32_to_cpu(fw_data[table_offset + i]));
6763                         }
6764                         bo_offset += table_size;
6765                 }
6766         }
6767 }
6768
6769 static void cik_enable_gfx_cgpg(struct radeon_device *rdev,
6770                                 bool enable)
6771 {
6772         u32 data, orig;
6773
6774         if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_PG)) {
6775                 orig = data = RREG32(RLC_PG_CNTL);
6776                 data |= GFX_PG_ENABLE;
6777                 if (orig != data)
6778                         WREG32(RLC_PG_CNTL, data);
6779
6780                 orig = data = RREG32(RLC_AUTO_PG_CTRL);
6781                 data |= AUTO_PG_EN;
6782                 if (orig != data)
6783                         WREG32(RLC_AUTO_PG_CTRL, data);
6784         } else {
6785                 orig = data = RREG32(RLC_PG_CNTL);
6786                 data &= ~GFX_PG_ENABLE;
6787                 if (orig != data)
6788                         WREG32(RLC_PG_CNTL, data);
6789
6790                 orig = data = RREG32(RLC_AUTO_PG_CTRL);
6791                 data &= ~AUTO_PG_EN;
6792                 if (orig != data)
6793                         WREG32(RLC_AUTO_PG_CTRL, data);
6794
6795                 data = RREG32(DB_RENDER_CONTROL);
6796         }
6797 }
6798
6799 static u32 cik_get_cu_active_bitmap(struct radeon_device *rdev, u32 se, u32 sh)
6800 {
6801         u32 mask = 0, tmp, tmp1;
6802         int i;
6803
6804         cik_select_se_sh(rdev, se, sh);
6805         tmp = RREG32(CC_GC_SHADER_ARRAY_CONFIG);
6806         tmp1 = RREG32(GC_USER_SHADER_ARRAY_CONFIG);
6807         cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
6808
6809         tmp &= 0xffff0000;
6810
6811         tmp |= tmp1;
6812         tmp >>= 16;
6813
6814         for (i = 0; i < rdev->config.cik.max_cu_per_sh; i ++) {
6815                 mask <<= 1;
6816                 mask |= 1;
6817         }
6818
6819         return (~tmp) & mask;
6820 }
6821
6822 static void cik_init_ao_cu_mask(struct radeon_device *rdev)
6823 {
6824         u32 i, j, k, active_cu_number = 0;
6825         u32 mask, counter, cu_bitmap;
6826         u32 tmp = 0;
6827
6828         for (i = 0; i < rdev->config.cik.max_shader_engines; i++) {
6829                 for (j = 0; j < rdev->config.cik.max_sh_per_se; j++) {
6830                         mask = 1;
6831                         cu_bitmap = 0;
6832                         counter = 0;
6833                         for (k = 0; k < rdev->config.cik.max_cu_per_sh; k ++) {
6834                                 if (cik_get_cu_active_bitmap(rdev, i, j) & mask) {
6835                                         if (counter < 2)
6836                                                 cu_bitmap |= mask;
6837                                         counter ++;
6838                                 }
6839                                 mask <<= 1;
6840                         }
6841
6842                         active_cu_number += counter;
6843                         tmp |= (cu_bitmap << (i * 16 + j * 8));
6844                 }
6845         }
6846
6847         WREG32(RLC_PG_AO_CU_MASK, tmp);
6848
6849         tmp = RREG32(RLC_MAX_PG_CU);
6850         tmp &= ~MAX_PU_CU_MASK;
6851         tmp |= MAX_PU_CU(active_cu_number);
6852         WREG32(RLC_MAX_PG_CU, tmp);
6853 }
6854
6855 static void cik_enable_gfx_static_mgpg(struct radeon_device *rdev,
6856                                        bool enable)
6857 {
6858         u32 data, orig;
6859
6860         orig = data = RREG32(RLC_PG_CNTL);
6861         if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_SMG))
6862                 data |= STATIC_PER_CU_PG_ENABLE;
6863         else
6864                 data &= ~STATIC_PER_CU_PG_ENABLE;
6865         if (orig != data)
6866                 WREG32(RLC_PG_CNTL, data);
6867 }
6868
6869 static void cik_enable_gfx_dynamic_mgpg(struct radeon_device *rdev,
6870                                         bool enable)
6871 {
6872         u32 data, orig;
6873
6874         orig = data = RREG32(RLC_PG_CNTL);
6875         if (enable && (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_DMG))
6876                 data |= DYN_PER_CU_PG_ENABLE;
6877         else
6878                 data &= ~DYN_PER_CU_PG_ENABLE;
6879         if (orig != data)
6880                 WREG32(RLC_PG_CNTL, data);
6881 }
6882
6883 #define RLC_SAVE_AND_RESTORE_STARTING_OFFSET 0x90
6884 #define RLC_CLEAR_STATE_DESCRIPTOR_OFFSET    0x3D
6885
6886 static void cik_init_gfx_cgpg(struct radeon_device *rdev)
6887 {
6888         u32 data, orig;
6889         u32 i;
6890
6891         if (rdev->rlc.cs_data) {
6892                 WREG32(RLC_GPM_SCRATCH_ADDR, RLC_CLEAR_STATE_DESCRIPTOR_OFFSET);
6893                 WREG32(RLC_GPM_SCRATCH_DATA, upper_32_bits(rdev->rlc.clear_state_gpu_addr));
6894                 WREG32(RLC_GPM_SCRATCH_DATA, lower_32_bits(rdev->rlc.clear_state_gpu_addr));
6895                 WREG32(RLC_GPM_SCRATCH_DATA, rdev->rlc.clear_state_size);
6896         } else {
6897                 WREG32(RLC_GPM_SCRATCH_ADDR, RLC_CLEAR_STATE_DESCRIPTOR_OFFSET);
6898                 for (i = 0; i < 3; i++)
6899                         WREG32(RLC_GPM_SCRATCH_DATA, 0);
6900         }
6901         if (rdev->rlc.reg_list) {
6902                 WREG32(RLC_GPM_SCRATCH_ADDR, RLC_SAVE_AND_RESTORE_STARTING_OFFSET);
6903                 for (i = 0; i < rdev->rlc.reg_list_size; i++)
6904                         WREG32(RLC_GPM_SCRATCH_DATA, rdev->rlc.reg_list[i]);
6905         }
6906
6907         orig = data = RREG32(RLC_PG_CNTL);
6908         data |= GFX_PG_SRC;
6909         if (orig != data)
6910                 WREG32(RLC_PG_CNTL, data);
6911
6912         WREG32(RLC_SAVE_AND_RESTORE_BASE, rdev->rlc.save_restore_gpu_addr >> 8);
6913         WREG32(RLC_CP_TABLE_RESTORE, rdev->rlc.cp_table_gpu_addr >> 8);
6914
6915         data = RREG32(CP_RB_WPTR_POLL_CNTL);
6916         data &= ~IDLE_POLL_COUNT_MASK;
6917         data |= IDLE_POLL_COUNT(0x60);
6918         WREG32(CP_RB_WPTR_POLL_CNTL, data);
6919
6920         data = 0x10101010;
6921         WREG32(RLC_PG_DELAY, data);
6922
6923         data = RREG32(RLC_PG_DELAY_2);
6924         data &= ~0xff;
6925         data |= 0x3;
6926         WREG32(RLC_PG_DELAY_2, data);
6927
6928         data = RREG32(RLC_AUTO_PG_CTRL);
6929         data &= ~GRBM_REG_SGIT_MASK;
6930         data |= GRBM_REG_SGIT(0x700);
6931         WREG32(RLC_AUTO_PG_CTRL, data);
6932
6933 }
6934
6935 static void cik_update_gfx_pg(struct radeon_device *rdev, bool enable)
6936 {
6937         cik_enable_gfx_cgpg(rdev, enable);
6938         cik_enable_gfx_static_mgpg(rdev, enable);
6939         cik_enable_gfx_dynamic_mgpg(rdev, enable);
6940 }
6941
6942 u32 cik_get_csb_size(struct radeon_device *rdev)
6943 {
6944         u32 count = 0;
6945         const struct cs_section_def *sect = NULL;
6946         const struct cs_extent_def *ext = NULL;
6947
6948         if (rdev->rlc.cs_data == NULL)
6949                 return 0;
6950
6951         /* begin clear state */
6952         count += 2;
6953         /* context control state */
6954         count += 3;
6955
6956         for (sect = rdev->rlc.cs_data; sect->section != NULL; ++sect) {
6957                 for (ext = sect->section; ext->extent != NULL; ++ext) {
6958                         if (sect->id == SECT_CONTEXT)
6959                                 count += 2 + ext->reg_count;
6960                         else
6961                                 return 0;
6962                 }
6963         }
6964         /* pa_sc_raster_config/pa_sc_raster_config1 */
6965         count += 4;
6966         /* end clear state */
6967         count += 2;
6968         /* clear state */
6969         count += 2;
6970
6971         return count;
6972 }
6973
6974 void cik_get_csb_buffer(struct radeon_device *rdev, volatile u32 *buffer)
6975 {
6976         u32 count = 0, i;
6977         const struct cs_section_def *sect = NULL;
6978         const struct cs_extent_def *ext = NULL;
6979
6980         if (rdev->rlc.cs_data == NULL)
6981                 return;
6982         if (buffer == NULL)
6983                 return;
6984
6985         buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
6986         buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
6987
6988         buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CONTEXT_CONTROL, 1));
6989         buffer[count++] = cpu_to_le32(0x80000000);
6990         buffer[count++] = cpu_to_le32(0x80000000);
6991
6992         for (sect = rdev->rlc.cs_data; sect->section != NULL; ++sect) {
6993                 for (ext = sect->section; ext->extent != NULL; ++ext) {
6994                         if (sect->id == SECT_CONTEXT) {
6995                                 buffer[count++] =
6996                                         cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, ext->reg_count));
6997                                 buffer[count++] = cpu_to_le32(ext->reg_index - 0xa000);
6998                                 for (i = 0; i < ext->reg_count; i++)
6999                                         buffer[count++] = cpu_to_le32(ext->extent[i]);
7000                         } else {
7001                                 return;
7002                         }
7003                 }
7004         }
7005
7006         buffer[count++] = cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, 2));
7007         buffer[count++] = cpu_to_le32(PA_SC_RASTER_CONFIG - PACKET3_SET_CONTEXT_REG_START);
7008         switch (rdev->family) {
7009         case CHIP_BONAIRE:
7010                 buffer[count++] = cpu_to_le32(0x16000012);
7011                 buffer[count++] = cpu_to_le32(0x00000000);
7012                 break;
7013         case CHIP_KAVERI:
7014                 buffer[count++] = cpu_to_le32(0x00000000); /* XXX */
7015                 buffer[count++] = cpu_to_le32(0x00000000);
7016                 break;
7017         case CHIP_KABINI:
7018         case CHIP_MULLINS:
7019                 buffer[count++] = cpu_to_le32(0x00000000); /* XXX */
7020                 buffer[count++] = cpu_to_le32(0x00000000);
7021                 break;
7022         case CHIP_HAWAII:
7023                 buffer[count++] = cpu_to_le32(0x3a00161a);
7024                 buffer[count++] = cpu_to_le32(0x0000002e);
7025                 break;
7026         default:
7027                 buffer[count++] = cpu_to_le32(0x00000000);
7028                 buffer[count++] = cpu_to_le32(0x00000000);
7029                 break;
7030         }
7031
7032         buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
7033         buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_END_CLEAR_STATE);
7034
7035         buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CLEAR_STATE, 0));
7036         buffer[count++] = cpu_to_le32(0);
7037 }
7038
7039 static void cik_init_pg(struct radeon_device *rdev)
7040 {
7041         if (rdev->pg_flags) {
7042                 cik_enable_sck_slowdown_on_pu(rdev, true);
7043                 cik_enable_sck_slowdown_on_pd(rdev, true);
7044                 if (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_PG) {
7045                         cik_init_gfx_cgpg(rdev);
7046                         cik_enable_cp_pg(rdev, true);
7047                         cik_enable_gds_pg(rdev, true);
7048                 }
7049                 cik_init_ao_cu_mask(rdev);
7050                 cik_update_gfx_pg(rdev, true);
7051         }
7052 }
7053
7054 static void cik_fini_pg(struct radeon_device *rdev)
7055 {
7056         if (rdev->pg_flags) {
7057                 cik_update_gfx_pg(rdev, false);
7058                 if (rdev->pg_flags & RADEON_PG_SUPPORT_GFX_PG) {
7059                         cik_enable_cp_pg(rdev, false);
7060                         cik_enable_gds_pg(rdev, false);
7061                 }
7062         }
7063 }
7064
7065 /*
7066  * Interrupts
7067  * Starting with r6xx, interrupts are handled via a ring buffer.
7068  * Ring buffers are areas of GPU accessible memory that the GPU
7069  * writes interrupt vectors into and the host reads vectors out of.
7070  * There is a rptr (read pointer) that determines where the
7071  * host is currently reading, and a wptr (write pointer)
7072  * which determines where the GPU has written.  When the
7073  * pointers are equal, the ring is idle.  When the GPU
7074  * writes vectors to the ring buffer, it increments the
7075  * wptr.  When there is an interrupt, the host then starts
7076  * fetching commands and processing them until the pointers are
7077  * equal again at which point it updates the rptr.
7078  */
7079
7080 /**
7081  * cik_enable_interrupts - Enable the interrupt ring buffer
7082  *
7083  * @rdev: radeon_device pointer
7084  *
7085  * Enable the interrupt ring buffer (CIK).
7086  */
7087 static void cik_enable_interrupts(struct radeon_device *rdev)
7088 {
7089         u32 ih_cntl = RREG32(IH_CNTL);
7090         u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
7091
7092         ih_cntl |= ENABLE_INTR;
7093         ih_rb_cntl |= IH_RB_ENABLE;
7094         WREG32(IH_CNTL, ih_cntl);
7095         WREG32(IH_RB_CNTL, ih_rb_cntl);
7096         rdev->ih.enabled = true;
7097 }
7098
7099 /**
7100  * cik_disable_interrupts - Disable the interrupt ring buffer
7101  *
7102  * @rdev: radeon_device pointer
7103  *
7104  * Disable the interrupt ring buffer (CIK).
7105  */
7106 static void cik_disable_interrupts(struct radeon_device *rdev)
7107 {
7108         u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
7109         u32 ih_cntl = RREG32(IH_CNTL);
7110
7111         ih_rb_cntl &= ~IH_RB_ENABLE;
7112         ih_cntl &= ~ENABLE_INTR;
7113         WREG32(IH_RB_CNTL, ih_rb_cntl);
7114         WREG32(IH_CNTL, ih_cntl);
7115         /* set rptr, wptr to 0 */
7116         WREG32(IH_RB_RPTR, 0);
7117         WREG32(IH_RB_WPTR, 0);
7118         rdev->ih.enabled = false;
7119         rdev->ih.rptr = 0;
7120 }
7121
7122 /**
7123  * cik_disable_interrupt_state - Disable all interrupt sources
7124  *
7125  * @rdev: radeon_device pointer
7126  *
7127  * Clear all interrupt enable bits used by the driver (CIK).
7128  */
7129 static void cik_disable_interrupt_state(struct radeon_device *rdev)
7130 {
7131         u32 tmp;
7132
7133         /* gfx ring */
7134         tmp = RREG32(CP_INT_CNTL_RING0) &
7135                 (CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
7136         WREG32(CP_INT_CNTL_RING0, tmp);
7137         /* sdma */
7138         tmp = RREG32(SDMA0_CNTL + SDMA0_REGISTER_OFFSET) & ~TRAP_ENABLE;
7139         WREG32(SDMA0_CNTL + SDMA0_REGISTER_OFFSET, tmp);
7140         tmp = RREG32(SDMA0_CNTL + SDMA1_REGISTER_OFFSET) & ~TRAP_ENABLE;
7141         WREG32(SDMA0_CNTL + SDMA1_REGISTER_OFFSET, tmp);
7142         /* compute queues */
7143         WREG32(CP_ME1_PIPE0_INT_CNTL, 0);
7144         WREG32(CP_ME1_PIPE1_INT_CNTL, 0);
7145         WREG32(CP_ME1_PIPE2_INT_CNTL, 0);
7146         WREG32(CP_ME1_PIPE3_INT_CNTL, 0);
7147         WREG32(CP_ME2_PIPE0_INT_CNTL, 0);
7148         WREG32(CP_ME2_PIPE1_INT_CNTL, 0);
7149         WREG32(CP_ME2_PIPE2_INT_CNTL, 0);
7150         WREG32(CP_ME2_PIPE3_INT_CNTL, 0);
7151         /* grbm */
7152         WREG32(GRBM_INT_CNTL, 0);
7153         /* vline/vblank, etc. */
7154         WREG32(LB_INTERRUPT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
7155         WREG32(LB_INTERRUPT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
7156         if (rdev->num_crtc >= 4) {
7157                 WREG32(LB_INTERRUPT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
7158                 WREG32(LB_INTERRUPT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
7159         }
7160         if (rdev->num_crtc >= 6) {
7161                 WREG32(LB_INTERRUPT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
7162                 WREG32(LB_INTERRUPT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
7163         }
7164         /* pflip */
7165         if (rdev->num_crtc >= 2) {
7166                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
7167                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
7168         }
7169         if (rdev->num_crtc >= 4) {
7170                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
7171                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
7172         }
7173         if (rdev->num_crtc >= 6) {
7174                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
7175                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
7176         }
7177
7178         /* dac hotplug */
7179         WREG32(DAC_AUTODETECT_INT_CONTROL, 0);
7180
7181         /* digital hotplug */
7182         tmp = RREG32(DC_HPD1_INT_CONTROL) & DC_HPDx_INT_POLARITY;
7183         WREG32(DC_HPD1_INT_CONTROL, tmp);
7184         tmp = RREG32(DC_HPD2_INT_CONTROL) & DC_HPDx_INT_POLARITY;
7185         WREG32(DC_HPD2_INT_CONTROL, tmp);
7186         tmp = RREG32(DC_HPD3_INT_CONTROL) & DC_HPDx_INT_POLARITY;
7187         WREG32(DC_HPD3_INT_CONTROL, tmp);
7188         tmp = RREG32(DC_HPD4_INT_CONTROL) & DC_HPDx_INT_POLARITY;
7189         WREG32(DC_HPD4_INT_CONTROL, tmp);
7190         tmp = RREG32(DC_HPD5_INT_CONTROL) & DC_HPDx_INT_POLARITY;
7191         WREG32(DC_HPD5_INT_CONTROL, tmp);
7192         tmp = RREG32(DC_HPD6_INT_CONTROL) & DC_HPDx_INT_POLARITY;
7193         WREG32(DC_HPD6_INT_CONTROL, tmp);
7194
7195 }
7196
7197 /**
7198  * cik_irq_init - init and enable the interrupt ring
7199  *
7200  * @rdev: radeon_device pointer
7201  *
7202  * Allocate a ring buffer for the interrupt controller,
7203  * enable the RLC, disable interrupts, enable the IH
7204  * ring buffer and enable it (CIK).
7205  * Called at device load and reume.
7206  * Returns 0 for success, errors for failure.
7207  */
7208 static int cik_irq_init(struct radeon_device *rdev)
7209 {
7210         int ret = 0;
7211         int rb_bufsz;
7212         u32 interrupt_cntl, ih_cntl, ih_rb_cntl;
7213
7214         /* allocate ring */
7215         ret = r600_ih_ring_alloc(rdev);
7216         if (ret)
7217                 return ret;
7218
7219         /* disable irqs */
7220         cik_disable_interrupts(rdev);
7221
7222         /* init rlc */
7223         ret = cik_rlc_resume(rdev);
7224         if (ret) {
7225                 r600_ih_ring_fini(rdev);
7226                 return ret;
7227         }
7228
7229         /* setup interrupt control */
7230         /* XXX this should actually be a bus address, not an MC address. same on older asics */
7231         WREG32(INTERRUPT_CNTL2, rdev->ih.gpu_addr >> 8);
7232         interrupt_cntl = RREG32(INTERRUPT_CNTL);
7233         /* IH_DUMMY_RD_OVERRIDE=0 - dummy read disabled with msi, enabled without msi
7234          * IH_DUMMY_RD_OVERRIDE=1 - dummy read controlled by IH_DUMMY_RD_EN
7235          */
7236         interrupt_cntl &= ~IH_DUMMY_RD_OVERRIDE;
7237         /* IH_REQ_NONSNOOP_EN=1 if ring is in non-cacheable memory, e.g., vram */
7238         interrupt_cntl &= ~IH_REQ_NONSNOOP_EN;
7239         WREG32(INTERRUPT_CNTL, interrupt_cntl);
7240
7241         WREG32(IH_RB_BASE, rdev->ih.gpu_addr >> 8);
7242         rb_bufsz = order_base_2(rdev->ih.ring_size / 4);
7243
7244         ih_rb_cntl = (IH_WPTR_OVERFLOW_ENABLE |
7245                       IH_WPTR_OVERFLOW_CLEAR |
7246                       (rb_bufsz << 1));
7247
7248         if (rdev->wb.enabled)
7249                 ih_rb_cntl |= IH_WPTR_WRITEBACK_ENABLE;
7250
7251         /* set the writeback address whether it's enabled or not */
7252         WREG32(IH_RB_WPTR_ADDR_LO, (rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFFFFFFFC);
7253         WREG32(IH_RB_WPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFF);
7254
7255         WREG32(IH_RB_CNTL, ih_rb_cntl);
7256
7257         /* set rptr, wptr to 0 */
7258         WREG32(IH_RB_RPTR, 0);
7259         WREG32(IH_RB_WPTR, 0);
7260
7261         /* Default settings for IH_CNTL (disabled at first) */
7262         ih_cntl = MC_WRREQ_CREDIT(0x10) | MC_WR_CLEAN_CNT(0x10) | MC_VMID(0);
7263         /* RPTR_REARM only works if msi's are enabled */
7264         if (rdev->msi_enabled)
7265                 ih_cntl |= RPTR_REARM;
7266         WREG32(IH_CNTL, ih_cntl);
7267
7268         /* force the active interrupt state to all disabled */
7269         cik_disable_interrupt_state(rdev);
7270
7271         pci_set_master(rdev->pdev);
7272
7273         /* enable irqs */
7274         cik_enable_interrupts(rdev);
7275
7276         return ret;
7277 }
7278
7279 /**
7280  * cik_irq_set - enable/disable interrupt sources
7281  *
7282  * @rdev: radeon_device pointer
7283  *
7284  * Enable interrupt sources on the GPU (vblanks, hpd,
7285  * etc.) (CIK).
7286  * Returns 0 for success, errors for failure.
7287  */
7288 int cik_irq_set(struct radeon_device *rdev)
7289 {
7290         u32 cp_int_cntl;
7291         u32 cp_m1p0, cp_m1p1, cp_m1p2, cp_m1p3;
7292         u32 cp_m2p0, cp_m2p1, cp_m2p2, cp_m2p3;
7293         u32 crtc1 = 0, crtc2 = 0, crtc3 = 0, crtc4 = 0, crtc5 = 0, crtc6 = 0;
7294         u32 hpd1, hpd2, hpd3, hpd4, hpd5, hpd6;
7295         u32 grbm_int_cntl = 0;
7296         u32 dma_cntl, dma_cntl1;
7297         u32 thermal_int;
7298
7299         if (!rdev->irq.installed) {
7300                 WARN(1, "Can't enable IRQ/MSI because no handler is installed\n");
7301                 return -EINVAL;
7302         }
7303         /* don't enable anything if the ih is disabled */
7304         if (!rdev->ih.enabled) {
7305                 cik_disable_interrupts(rdev);
7306                 /* force the active interrupt state to all disabled */
7307                 cik_disable_interrupt_state(rdev);
7308                 return 0;
7309         }
7310
7311         cp_int_cntl = RREG32(CP_INT_CNTL_RING0) &
7312                 (CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
7313         cp_int_cntl |= PRIV_INSTR_INT_ENABLE | PRIV_REG_INT_ENABLE;
7314
7315         hpd1 = RREG32(DC_HPD1_INT_CONTROL) & ~DC_HPDx_INT_EN;
7316         hpd2 = RREG32(DC_HPD2_INT_CONTROL) & ~DC_HPDx_INT_EN;
7317         hpd3 = RREG32(DC_HPD3_INT_CONTROL) & ~DC_HPDx_INT_EN;
7318         hpd4 = RREG32(DC_HPD4_INT_CONTROL) & ~DC_HPDx_INT_EN;
7319         hpd5 = RREG32(DC_HPD5_INT_CONTROL) & ~DC_HPDx_INT_EN;
7320         hpd6 = RREG32(DC_HPD6_INT_CONTROL) & ~DC_HPDx_INT_EN;
7321
7322         dma_cntl = RREG32(SDMA0_CNTL + SDMA0_REGISTER_OFFSET) & ~TRAP_ENABLE;
7323         dma_cntl1 = RREG32(SDMA0_CNTL + SDMA1_REGISTER_OFFSET) & ~TRAP_ENABLE;
7324
7325         cp_m1p0 = RREG32(CP_ME1_PIPE0_INT_CNTL) & ~TIME_STAMP_INT_ENABLE;
7326         cp_m1p1 = RREG32(CP_ME1_PIPE1_INT_CNTL) & ~TIME_STAMP_INT_ENABLE;
7327         cp_m1p2 = RREG32(CP_ME1_PIPE2_INT_CNTL) & ~TIME_STAMP_INT_ENABLE;
7328         cp_m1p3 = RREG32(CP_ME1_PIPE3_INT_CNTL) & ~TIME_STAMP_INT_ENABLE;
7329         cp_m2p0 = RREG32(CP_ME2_PIPE0_INT_CNTL) & ~TIME_STAMP_INT_ENABLE;
7330         cp_m2p1 = RREG32(CP_ME2_PIPE1_INT_CNTL) & ~TIME_STAMP_INT_ENABLE;
7331         cp_m2p2 = RREG32(CP_ME2_PIPE2_INT_CNTL) & ~TIME_STAMP_INT_ENABLE;
7332         cp_m2p3 = RREG32(CP_ME2_PIPE3_INT_CNTL) & ~TIME_STAMP_INT_ENABLE;
7333
7334         if (rdev->flags & RADEON_IS_IGP)
7335                 thermal_int = RREG32_SMC(CG_THERMAL_INT_CTRL) &
7336                         ~(THERM_INTH_MASK | THERM_INTL_MASK);
7337         else
7338                 thermal_int = RREG32_SMC(CG_THERMAL_INT) &
7339                         ~(THERM_INT_MASK_HIGH | THERM_INT_MASK_LOW);
7340
7341         /* enable CP interrupts on all rings */
7342         if (atomic_read(&rdev->irq.ring_int[RADEON_RING_TYPE_GFX_INDEX])) {
7343                 DRM_DEBUG("cik_irq_set: sw int gfx\n");
7344                 cp_int_cntl |= TIME_STAMP_INT_ENABLE;
7345         }
7346         if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_CP1_INDEX])) {
7347                 struct radeon_ring *ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
7348                 DRM_DEBUG("si_irq_set: sw int cp1\n");
7349                 if (ring->me == 1) {
7350                         switch (ring->pipe) {
7351                         case 0:
7352                                 cp_m1p0 |= TIME_STAMP_INT_ENABLE;
7353                                 break;
7354                         case 1:
7355                                 cp_m1p1 |= TIME_STAMP_INT_ENABLE;
7356                                 break;
7357                         case 2:
7358                                 cp_m1p2 |= TIME_STAMP_INT_ENABLE;
7359                                 break;
7360                         case 3:
7361                                 cp_m1p2 |= TIME_STAMP_INT_ENABLE;
7362                                 break;
7363                         default:
7364                                 DRM_DEBUG("si_irq_set: sw int cp1 invalid pipe %d\n", ring->pipe);
7365                                 break;
7366                         }
7367                 } else if (ring->me == 2) {
7368                         switch (ring->pipe) {
7369                         case 0:
7370                                 cp_m2p0 |= TIME_STAMP_INT_ENABLE;
7371                                 break;
7372                         case 1:
7373                                 cp_m2p1 |= TIME_STAMP_INT_ENABLE;
7374                                 break;
7375                         case 2:
7376                                 cp_m2p2 |= TIME_STAMP_INT_ENABLE;
7377                                 break;
7378                         case 3:
7379                                 cp_m2p2 |= TIME_STAMP_INT_ENABLE;
7380                                 break;
7381                         default:
7382                                 DRM_DEBUG("si_irq_set: sw int cp1 invalid pipe %d\n", ring->pipe);
7383                                 break;
7384                         }
7385                 } else {
7386                         DRM_DEBUG("si_irq_set: sw int cp1 invalid me %d\n", ring->me);
7387                 }
7388         }
7389         if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_CP2_INDEX])) {
7390                 struct radeon_ring *ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
7391                 DRM_DEBUG("si_irq_set: sw int cp2\n");
7392                 if (ring->me == 1) {
7393                         switch (ring->pipe) {
7394                         case 0:
7395                                 cp_m1p0 |= TIME_STAMP_INT_ENABLE;
7396                                 break;
7397                         case 1:
7398                                 cp_m1p1 |= TIME_STAMP_INT_ENABLE;
7399                                 break;
7400                         case 2:
7401                                 cp_m1p2 |= TIME_STAMP_INT_ENABLE;
7402                                 break;
7403                         case 3:
7404                                 cp_m1p2 |= TIME_STAMP_INT_ENABLE;
7405                                 break;
7406                         default:
7407                                 DRM_DEBUG("si_irq_set: sw int cp2 invalid pipe %d\n", ring->pipe);
7408                                 break;
7409                         }
7410                 } else if (ring->me == 2) {
7411                         switch (ring->pipe) {
7412                         case 0:
7413                                 cp_m2p0 |= TIME_STAMP_INT_ENABLE;
7414                                 break;
7415                         case 1:
7416                                 cp_m2p1 |= TIME_STAMP_INT_ENABLE;
7417                                 break;
7418                         case 2:
7419                                 cp_m2p2 |= TIME_STAMP_INT_ENABLE;
7420                                 break;
7421                         case 3:
7422                                 cp_m2p2 |= TIME_STAMP_INT_ENABLE;
7423                                 break;
7424                         default:
7425                                 DRM_DEBUG("si_irq_set: sw int cp2 invalid pipe %d\n", ring->pipe);
7426                                 break;
7427                         }
7428                 } else {
7429                         DRM_DEBUG("si_irq_set: sw int cp2 invalid me %d\n", ring->me);
7430                 }
7431         }
7432
7433         if (atomic_read(&rdev->irq.ring_int[R600_RING_TYPE_DMA_INDEX])) {
7434                 DRM_DEBUG("cik_irq_set: sw int dma\n");
7435                 dma_cntl |= TRAP_ENABLE;
7436         }
7437
7438         if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_DMA1_INDEX])) {
7439                 DRM_DEBUG("cik_irq_set: sw int dma1\n");
7440                 dma_cntl1 |= TRAP_ENABLE;
7441         }
7442
7443         if (rdev->irq.crtc_vblank_int[0] ||
7444             atomic_read(&rdev->irq.pflip[0])) {
7445                 DRM_DEBUG("cik_irq_set: vblank 0\n");
7446                 crtc1 |= VBLANK_INTERRUPT_MASK;
7447         }
7448         if (rdev->irq.crtc_vblank_int[1] ||
7449             atomic_read(&rdev->irq.pflip[1])) {
7450                 DRM_DEBUG("cik_irq_set: vblank 1\n");
7451                 crtc2 |= VBLANK_INTERRUPT_MASK;
7452         }
7453         if (rdev->irq.crtc_vblank_int[2] ||
7454             atomic_read(&rdev->irq.pflip[2])) {
7455                 DRM_DEBUG("cik_irq_set: vblank 2\n");
7456                 crtc3 |= VBLANK_INTERRUPT_MASK;
7457         }
7458         if (rdev->irq.crtc_vblank_int[3] ||
7459             atomic_read(&rdev->irq.pflip[3])) {
7460                 DRM_DEBUG("cik_irq_set: vblank 3\n");
7461                 crtc4 |= VBLANK_INTERRUPT_MASK;
7462         }
7463         if (rdev->irq.crtc_vblank_int[4] ||
7464             atomic_read(&rdev->irq.pflip[4])) {
7465                 DRM_DEBUG("cik_irq_set: vblank 4\n");
7466                 crtc5 |= VBLANK_INTERRUPT_MASK;
7467         }
7468         if (rdev->irq.crtc_vblank_int[5] ||
7469             atomic_read(&rdev->irq.pflip[5])) {
7470                 DRM_DEBUG("cik_irq_set: vblank 5\n");
7471                 crtc6 |= VBLANK_INTERRUPT_MASK;
7472         }
7473         if (rdev->irq.hpd[0]) {
7474                 DRM_DEBUG("cik_irq_set: hpd 1\n");
7475                 hpd1 |= DC_HPDx_INT_EN;
7476         }
7477         if (rdev->irq.hpd[1]) {
7478                 DRM_DEBUG("cik_irq_set: hpd 2\n");
7479                 hpd2 |= DC_HPDx_INT_EN;
7480         }
7481         if (rdev->irq.hpd[2]) {
7482                 DRM_DEBUG("cik_irq_set: hpd 3\n");
7483                 hpd3 |= DC_HPDx_INT_EN;
7484         }
7485         if (rdev->irq.hpd[3]) {
7486                 DRM_DEBUG("cik_irq_set: hpd 4\n");
7487                 hpd4 |= DC_HPDx_INT_EN;
7488         }
7489         if (rdev->irq.hpd[4]) {
7490                 DRM_DEBUG("cik_irq_set: hpd 5\n");
7491                 hpd5 |= DC_HPDx_INT_EN;
7492         }
7493         if (rdev->irq.hpd[5]) {
7494                 DRM_DEBUG("cik_irq_set: hpd 6\n");
7495                 hpd6 |= DC_HPDx_INT_EN;
7496         }
7497
7498         if (rdev->irq.dpm_thermal) {
7499                 DRM_DEBUG("dpm thermal\n");
7500                 if (rdev->flags & RADEON_IS_IGP)
7501                         thermal_int |= THERM_INTH_MASK | THERM_INTL_MASK;
7502                 else
7503                         thermal_int |= THERM_INT_MASK_HIGH | THERM_INT_MASK_LOW;
7504         }
7505
7506         WREG32(CP_INT_CNTL_RING0, cp_int_cntl);
7507
7508         WREG32(SDMA0_CNTL + SDMA0_REGISTER_OFFSET, dma_cntl);
7509         WREG32(SDMA0_CNTL + SDMA1_REGISTER_OFFSET, dma_cntl1);
7510
7511         WREG32(CP_ME1_PIPE0_INT_CNTL, cp_m1p0);
7512         WREG32(CP_ME1_PIPE1_INT_CNTL, cp_m1p1);
7513         WREG32(CP_ME1_PIPE2_INT_CNTL, cp_m1p2);
7514         WREG32(CP_ME1_PIPE3_INT_CNTL, cp_m1p3);
7515         WREG32(CP_ME2_PIPE0_INT_CNTL, cp_m2p0);
7516         WREG32(CP_ME2_PIPE1_INT_CNTL, cp_m2p1);
7517         WREG32(CP_ME2_PIPE2_INT_CNTL, cp_m2p2);
7518         WREG32(CP_ME2_PIPE3_INT_CNTL, cp_m2p3);
7519
7520         WREG32(GRBM_INT_CNTL, grbm_int_cntl);
7521
7522         WREG32(LB_INTERRUPT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, crtc1);
7523         WREG32(LB_INTERRUPT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, crtc2);
7524         if (rdev->num_crtc >= 4) {
7525                 WREG32(LB_INTERRUPT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, crtc3);
7526                 WREG32(LB_INTERRUPT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, crtc4);
7527         }
7528         if (rdev->num_crtc >= 6) {
7529                 WREG32(LB_INTERRUPT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, crtc5);
7530                 WREG32(LB_INTERRUPT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, crtc6);
7531         }
7532
7533         if (rdev->num_crtc >= 2) {
7534                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET,
7535                        GRPH_PFLIP_INT_MASK);
7536                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET,
7537                        GRPH_PFLIP_INT_MASK);
7538         }
7539         if (rdev->num_crtc >= 4) {
7540                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET,
7541                        GRPH_PFLIP_INT_MASK);
7542                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET,
7543                        GRPH_PFLIP_INT_MASK);
7544         }
7545         if (rdev->num_crtc >= 6) {
7546                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET,
7547                        GRPH_PFLIP_INT_MASK);
7548                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET,
7549                        GRPH_PFLIP_INT_MASK);
7550         }
7551
7552         WREG32(DC_HPD1_INT_CONTROL, hpd1);
7553         WREG32(DC_HPD2_INT_CONTROL, hpd2);
7554         WREG32(DC_HPD3_INT_CONTROL, hpd3);
7555         WREG32(DC_HPD4_INT_CONTROL, hpd4);
7556         WREG32(DC_HPD5_INT_CONTROL, hpd5);
7557         WREG32(DC_HPD6_INT_CONTROL, hpd6);
7558
7559         if (rdev->flags & RADEON_IS_IGP)
7560                 WREG32_SMC(CG_THERMAL_INT_CTRL, thermal_int);
7561         else
7562                 WREG32_SMC(CG_THERMAL_INT, thermal_int);
7563
7564         return 0;
7565 }
7566
7567 /**
7568  * cik_irq_ack - ack interrupt sources
7569  *
7570  * @rdev: radeon_device pointer
7571  *
7572  * Ack interrupt sources on the GPU (vblanks, hpd,
7573  * etc.) (CIK).  Certain interrupts sources are sw
7574  * generated and do not require an explicit ack.
7575  */
7576 static inline void cik_irq_ack(struct radeon_device *rdev)
7577 {
7578         u32 tmp;
7579
7580         rdev->irq.stat_regs.cik.disp_int = RREG32(DISP_INTERRUPT_STATUS);
7581         rdev->irq.stat_regs.cik.disp_int_cont = RREG32(DISP_INTERRUPT_STATUS_CONTINUE);
7582         rdev->irq.stat_regs.cik.disp_int_cont2 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE2);
7583         rdev->irq.stat_regs.cik.disp_int_cont3 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE3);
7584         rdev->irq.stat_regs.cik.disp_int_cont4 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE4);
7585         rdev->irq.stat_regs.cik.disp_int_cont5 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE5);
7586         rdev->irq.stat_regs.cik.disp_int_cont6 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE6);
7587
7588         rdev->irq.stat_regs.cik.d1grph_int = RREG32(GRPH_INT_STATUS +
7589                 EVERGREEN_CRTC0_REGISTER_OFFSET);
7590         rdev->irq.stat_regs.cik.d2grph_int = RREG32(GRPH_INT_STATUS +
7591                 EVERGREEN_CRTC1_REGISTER_OFFSET);
7592         if (rdev->num_crtc >= 4) {
7593                 rdev->irq.stat_regs.cik.d3grph_int = RREG32(GRPH_INT_STATUS +
7594                         EVERGREEN_CRTC2_REGISTER_OFFSET);
7595                 rdev->irq.stat_regs.cik.d4grph_int = RREG32(GRPH_INT_STATUS +
7596                         EVERGREEN_CRTC3_REGISTER_OFFSET);
7597         }
7598         if (rdev->num_crtc >= 6) {
7599                 rdev->irq.stat_regs.cik.d5grph_int = RREG32(GRPH_INT_STATUS +
7600                         EVERGREEN_CRTC4_REGISTER_OFFSET);
7601                 rdev->irq.stat_regs.cik.d6grph_int = RREG32(GRPH_INT_STATUS +
7602                         EVERGREEN_CRTC5_REGISTER_OFFSET);
7603         }
7604
7605         if (rdev->irq.stat_regs.cik.d1grph_int & GRPH_PFLIP_INT_OCCURRED)
7606                 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET,
7607                        GRPH_PFLIP_INT_CLEAR);
7608         if (rdev->irq.stat_regs.cik.d2grph_int & GRPH_PFLIP_INT_OCCURRED)
7609                 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET,
7610                        GRPH_PFLIP_INT_CLEAR);
7611         if (rdev->irq.stat_regs.cik.disp_int & LB_D1_VBLANK_INTERRUPT)
7612                 WREG32(LB_VBLANK_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VBLANK_ACK);
7613         if (rdev->irq.stat_regs.cik.disp_int & LB_D1_VLINE_INTERRUPT)
7614                 WREG32(LB_VLINE_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VLINE_ACK);
7615         if (rdev->irq.stat_regs.cik.disp_int_cont & LB_D2_VBLANK_INTERRUPT)
7616                 WREG32(LB_VBLANK_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VBLANK_ACK);
7617         if (rdev->irq.stat_regs.cik.disp_int_cont & LB_D2_VLINE_INTERRUPT)
7618                 WREG32(LB_VLINE_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VLINE_ACK);
7619
7620         if (rdev->num_crtc >= 4) {
7621                 if (rdev->irq.stat_regs.cik.d3grph_int & GRPH_PFLIP_INT_OCCURRED)
7622                         WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET,
7623                                GRPH_PFLIP_INT_CLEAR);
7624                 if (rdev->irq.stat_regs.cik.d4grph_int & GRPH_PFLIP_INT_OCCURRED)
7625                         WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET,
7626                                GRPH_PFLIP_INT_CLEAR);
7627                 if (rdev->irq.stat_regs.cik.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT)
7628                         WREG32(LB_VBLANK_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VBLANK_ACK);
7629                 if (rdev->irq.stat_regs.cik.disp_int_cont2 & LB_D3_VLINE_INTERRUPT)
7630                         WREG32(LB_VLINE_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VLINE_ACK);
7631                 if (rdev->irq.stat_regs.cik.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT)
7632                         WREG32(LB_VBLANK_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VBLANK_ACK);
7633                 if (rdev->irq.stat_regs.cik.disp_int_cont3 & LB_D4_VLINE_INTERRUPT)
7634                         WREG32(LB_VLINE_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VLINE_ACK);
7635         }
7636
7637         if (rdev->num_crtc >= 6) {
7638                 if (rdev->irq.stat_regs.cik.d5grph_int & GRPH_PFLIP_INT_OCCURRED)
7639                         WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET,
7640                                GRPH_PFLIP_INT_CLEAR);
7641                 if (rdev->irq.stat_regs.cik.d6grph_int & GRPH_PFLIP_INT_OCCURRED)
7642                         WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET,
7643                                GRPH_PFLIP_INT_CLEAR);
7644                 if (rdev->irq.stat_regs.cik.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT)
7645                         WREG32(LB_VBLANK_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VBLANK_ACK);
7646                 if (rdev->irq.stat_regs.cik.disp_int_cont4 & LB_D5_VLINE_INTERRUPT)
7647                         WREG32(LB_VLINE_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VLINE_ACK);
7648                 if (rdev->irq.stat_regs.cik.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT)
7649                         WREG32(LB_VBLANK_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VBLANK_ACK);
7650                 if (rdev->irq.stat_regs.cik.disp_int_cont5 & LB_D6_VLINE_INTERRUPT)
7651                         WREG32(LB_VLINE_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VLINE_ACK);
7652         }
7653
7654         if (rdev->irq.stat_regs.cik.disp_int & DC_HPD1_INTERRUPT) {
7655                 tmp = RREG32(DC_HPD1_INT_CONTROL);
7656                 tmp |= DC_HPDx_INT_ACK;
7657                 WREG32(DC_HPD1_INT_CONTROL, tmp);
7658         }
7659         if (rdev->irq.stat_regs.cik.disp_int_cont & DC_HPD2_INTERRUPT) {
7660                 tmp = RREG32(DC_HPD2_INT_CONTROL);
7661                 tmp |= DC_HPDx_INT_ACK;
7662                 WREG32(DC_HPD2_INT_CONTROL, tmp);
7663         }
7664         if (rdev->irq.stat_regs.cik.disp_int_cont2 & DC_HPD3_INTERRUPT) {
7665                 tmp = RREG32(DC_HPD3_INT_CONTROL);
7666                 tmp |= DC_HPDx_INT_ACK;
7667                 WREG32(DC_HPD3_INT_CONTROL, tmp);
7668         }
7669         if (rdev->irq.stat_regs.cik.disp_int_cont3 & DC_HPD4_INTERRUPT) {
7670                 tmp = RREG32(DC_HPD4_INT_CONTROL);
7671                 tmp |= DC_HPDx_INT_ACK;
7672                 WREG32(DC_HPD4_INT_CONTROL, tmp);
7673         }
7674         if (rdev->irq.stat_regs.cik.disp_int_cont4 & DC_HPD5_INTERRUPT) {
7675                 tmp = RREG32(DC_HPD5_INT_CONTROL);
7676                 tmp |= DC_HPDx_INT_ACK;
7677                 WREG32(DC_HPD5_INT_CONTROL, tmp);
7678         }
7679         if (rdev->irq.stat_regs.cik.disp_int_cont5 & DC_HPD6_INTERRUPT) {
7680                 tmp = RREG32(DC_HPD5_INT_CONTROL);
7681                 tmp |= DC_HPDx_INT_ACK;
7682                 WREG32(DC_HPD6_INT_CONTROL, tmp);
7683         }
7684 }
7685
7686 /**
7687  * cik_irq_disable - disable interrupts
7688  *
7689  * @rdev: radeon_device pointer
7690  *
7691  * Disable interrupts on the hw (CIK).
7692  */
7693 static void cik_irq_disable(struct radeon_device *rdev)
7694 {
7695         cik_disable_interrupts(rdev);
7696         /* Wait and acknowledge irq */
7697         mdelay(1);
7698         cik_irq_ack(rdev);
7699         cik_disable_interrupt_state(rdev);
7700 }
7701
7702 /**
7703  * cik_irq_disable - disable interrupts for suspend
7704  *
7705  * @rdev: radeon_device pointer
7706  *
7707  * Disable interrupts and stop the RLC (CIK).
7708  * Used for suspend.
7709  */
7710 static void cik_irq_suspend(struct radeon_device *rdev)
7711 {
7712         cik_irq_disable(rdev);
7713         cik_rlc_stop(rdev);
7714 }
7715
7716 /**
7717  * cik_irq_fini - tear down interrupt support
7718  *
7719  * @rdev: radeon_device pointer
7720  *
7721  * Disable interrupts on the hw and free the IH ring
7722  * buffer (CIK).
7723  * Used for driver unload.
7724  */
7725 static void cik_irq_fini(struct radeon_device *rdev)
7726 {
7727         cik_irq_suspend(rdev);
7728         r600_ih_ring_fini(rdev);
7729 }
7730
7731 /**
7732  * cik_get_ih_wptr - get the IH ring buffer wptr
7733  *
7734  * @rdev: radeon_device pointer
7735  *
7736  * Get the IH ring buffer wptr from either the register
7737  * or the writeback memory buffer (CIK).  Also check for
7738  * ring buffer overflow and deal with it.
7739  * Used by cik_irq_process().
7740  * Returns the value of the wptr.
7741  */
7742 static inline u32 cik_get_ih_wptr(struct radeon_device *rdev)
7743 {
7744         u32 wptr, tmp;
7745
7746         if (rdev->wb.enabled)
7747                 wptr = le32_to_cpu(rdev->wb.wb[R600_WB_IH_WPTR_OFFSET/4]);
7748         else
7749                 wptr = RREG32(IH_RB_WPTR);
7750
7751         if (wptr & RB_OVERFLOW) {
7752                 /* When a ring buffer overflow happen start parsing interrupt
7753                  * from the last not overwritten vector (wptr + 16). Hopefully
7754                  * this should allow us to catchup.
7755                  */
7756                 dev_warn(rdev->dev, "IH ring buffer overflow (0x%08X, %d, %d)\n",
7757                         wptr, rdev->ih.rptr, (wptr + 16) + rdev->ih.ptr_mask);
7758                 rdev->ih.rptr = (wptr + 16) & rdev->ih.ptr_mask;
7759                 tmp = RREG32(IH_RB_CNTL);
7760                 tmp |= IH_WPTR_OVERFLOW_CLEAR;
7761                 WREG32(IH_RB_CNTL, tmp);
7762                 wptr &= ~RB_OVERFLOW;
7763         }
7764         return (wptr & rdev->ih.ptr_mask);
7765 }
7766
7767 /*        CIK IV Ring
7768  * Each IV ring entry is 128 bits:
7769  * [7:0]    - interrupt source id
7770  * [31:8]   - reserved
7771  * [59:32]  - interrupt source data
7772  * [63:60]  - reserved
7773  * [71:64]  - RINGID
7774  *            CP:
7775  *            ME_ID [1:0], PIPE_ID[1:0], QUEUE_ID[2:0]
7776  *            QUEUE_ID - for compute, which of the 8 queues owned by the dispatcher
7777  *                     - for gfx, hw shader state (0=PS...5=LS, 6=CS)
7778  *            ME_ID - 0 = gfx, 1 = first 4 CS pipes, 2 = second 4 CS pipes
7779  *            PIPE_ID - ME0 0=3D
7780  *                    - ME1&2 compute dispatcher (4 pipes each)
7781  *            SDMA:
7782  *            INSTANCE_ID [1:0], QUEUE_ID[1:0]
7783  *            INSTANCE_ID - 0 = sdma0, 1 = sdma1
7784  *            QUEUE_ID - 0 = gfx, 1 = rlc0, 2 = rlc1
7785  * [79:72]  - VMID
7786  * [95:80]  - PASID
7787  * [127:96] - reserved
7788  */
7789 /**
7790  * cik_irq_process - interrupt handler
7791  *
7792  * @rdev: radeon_device pointer
7793  *
7794  * Interrupt hander (CIK).  Walk the IH ring,
7795  * ack interrupts and schedule work to handle
7796  * interrupt events.
7797  * Returns irq process return code.
7798  */
7799 int cik_irq_process(struct radeon_device *rdev)
7800 {
7801         struct radeon_ring *cp1_ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
7802         struct radeon_ring *cp2_ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
7803         u32 wptr;
7804         u32 rptr;
7805         u32 src_id, src_data, ring_id;
7806         u8 me_id, pipe_id, queue_id;
7807         u32 ring_index;
7808         bool queue_hotplug = false;
7809         bool queue_reset = false;
7810         u32 addr, status, mc_client;
7811         bool queue_thermal = false;
7812
7813         if (!rdev->ih.enabled || rdev->shutdown)
7814                 return IRQ_NONE;
7815
7816         wptr = cik_get_ih_wptr(rdev);
7817
7818 restart_ih:
7819         /* is somebody else already processing irqs? */
7820         if (atomic_xchg(&rdev->ih.lock, 1))
7821                 return IRQ_NONE;
7822
7823         rptr = rdev->ih.rptr;
7824         DRM_DEBUG("cik_irq_process start: rptr %d, wptr %d\n", rptr, wptr);
7825
7826         /* Order reading of wptr vs. reading of IH ring data */
7827         rmb();
7828
7829         /* display interrupts */
7830         cik_irq_ack(rdev);
7831
7832         while (rptr != wptr) {
7833                 /* wptr/rptr are in bytes! */
7834                 ring_index = rptr / 4;
7835                 src_id =  le32_to_cpu(rdev->ih.ring[ring_index]) & 0xff;
7836                 src_data = le32_to_cpu(rdev->ih.ring[ring_index + 1]) & 0xfffffff;
7837                 ring_id = le32_to_cpu(rdev->ih.ring[ring_index + 2]) & 0xff;
7838
7839                 switch (src_id) {
7840                 case 1: /* D1 vblank/vline */
7841                         switch (src_data) {
7842                         case 0: /* D1 vblank */
7843                                 if (rdev->irq.stat_regs.cik.disp_int & LB_D1_VBLANK_INTERRUPT) {
7844                                         if (rdev->irq.crtc_vblank_int[0]) {
7845                                                 drm_handle_vblank(rdev->ddev, 0);
7846                                                 rdev->pm.vblank_sync = true;
7847                                                 wake_up(&rdev->irq.vblank_queue);
7848                                         }
7849                                         if (atomic_read(&rdev->irq.pflip[0]))
7850                                                 radeon_crtc_handle_vblank(rdev, 0);
7851                                         rdev->irq.stat_regs.cik.disp_int &= ~LB_D1_VBLANK_INTERRUPT;
7852                                         DRM_DEBUG("IH: D1 vblank\n");
7853                                 }
7854                                 break;
7855                         case 1: /* D1 vline */
7856                                 if (rdev->irq.stat_regs.cik.disp_int & LB_D1_VLINE_INTERRUPT) {
7857                                         rdev->irq.stat_regs.cik.disp_int &= ~LB_D1_VLINE_INTERRUPT;
7858                                         DRM_DEBUG("IH: D1 vline\n");
7859                                 }
7860                                 break;
7861                         default:
7862                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
7863                                 break;
7864                         }
7865                         break;
7866                 case 2: /* D2 vblank/vline */
7867                         switch (src_data) {
7868                         case 0: /* D2 vblank */
7869                                 if (rdev->irq.stat_regs.cik.disp_int_cont & LB_D2_VBLANK_INTERRUPT) {
7870                                         if (rdev->irq.crtc_vblank_int[1]) {
7871                                                 drm_handle_vblank(rdev->ddev, 1);
7872                                                 rdev->pm.vblank_sync = true;
7873                                                 wake_up(&rdev->irq.vblank_queue);
7874                                         }
7875                                         if (atomic_read(&rdev->irq.pflip[1]))
7876                                                 radeon_crtc_handle_vblank(rdev, 1);
7877                                         rdev->irq.stat_regs.cik.disp_int_cont &= ~LB_D2_VBLANK_INTERRUPT;
7878                                         DRM_DEBUG("IH: D2 vblank\n");
7879                                 }
7880                                 break;
7881                         case 1: /* D2 vline */
7882                                 if (rdev->irq.stat_regs.cik.disp_int_cont & LB_D2_VLINE_INTERRUPT) {
7883                                         rdev->irq.stat_regs.cik.disp_int_cont &= ~LB_D2_VLINE_INTERRUPT;
7884                                         DRM_DEBUG("IH: D2 vline\n");
7885                                 }
7886                                 break;
7887                         default:
7888                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
7889                                 break;
7890                         }
7891                         break;
7892                 case 3: /* D3 vblank/vline */
7893                         switch (src_data) {
7894                         case 0: /* D3 vblank */
7895                                 if (rdev->irq.stat_regs.cik.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT) {
7896                                         if (rdev->irq.crtc_vblank_int[2]) {
7897                                                 drm_handle_vblank(rdev->ddev, 2);
7898                                                 rdev->pm.vblank_sync = true;
7899                                                 wake_up(&rdev->irq.vblank_queue);
7900                                         }
7901                                         if (atomic_read(&rdev->irq.pflip[2]))
7902                                                 radeon_crtc_handle_vblank(rdev, 2);
7903                                         rdev->irq.stat_regs.cik.disp_int_cont2 &= ~LB_D3_VBLANK_INTERRUPT;
7904                                         DRM_DEBUG("IH: D3 vblank\n");
7905                                 }
7906                                 break;
7907                         case 1: /* D3 vline */
7908                                 if (rdev->irq.stat_regs.cik.disp_int_cont2 & LB_D3_VLINE_INTERRUPT) {
7909                                         rdev->irq.stat_regs.cik.disp_int_cont2 &= ~LB_D3_VLINE_INTERRUPT;
7910                                         DRM_DEBUG("IH: D3 vline\n");
7911                                 }
7912                                 break;
7913                         default:
7914                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
7915                                 break;
7916                         }
7917                         break;
7918                 case 4: /* D4 vblank/vline */
7919                         switch (src_data) {
7920                         case 0: /* D4 vblank */
7921                                 if (rdev->irq.stat_regs.cik.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT) {
7922                                         if (rdev->irq.crtc_vblank_int[3]) {
7923                                                 drm_handle_vblank(rdev->ddev, 3);
7924                                                 rdev->pm.vblank_sync = true;
7925                                                 wake_up(&rdev->irq.vblank_queue);
7926                                         }
7927                                         if (atomic_read(&rdev->irq.pflip[3]))
7928                                                 radeon_crtc_handle_vblank(rdev, 3);
7929                                         rdev->irq.stat_regs.cik.disp_int_cont3 &= ~LB_D4_VBLANK_INTERRUPT;
7930                                         DRM_DEBUG("IH: D4 vblank\n");
7931                                 }
7932                                 break;
7933                         case 1: /* D4 vline */
7934                                 if (rdev->irq.stat_regs.cik.disp_int_cont3 & LB_D4_VLINE_INTERRUPT) {
7935                                         rdev->irq.stat_regs.cik.disp_int_cont3 &= ~LB_D4_VLINE_INTERRUPT;
7936                                         DRM_DEBUG("IH: D4 vline\n");
7937                                 }
7938                                 break;
7939                         default:
7940                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
7941                                 break;
7942                         }
7943                         break;
7944                 case 5: /* D5 vblank/vline */
7945                         switch (src_data) {
7946                         case 0: /* D5 vblank */
7947                                 if (rdev->irq.stat_regs.cik.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT) {
7948                                         if (rdev->irq.crtc_vblank_int[4]) {
7949                                                 drm_handle_vblank(rdev->ddev, 4);
7950                                                 rdev->pm.vblank_sync = true;
7951                                                 wake_up(&rdev->irq.vblank_queue);
7952                                         }
7953                                         if (atomic_read(&rdev->irq.pflip[4]))
7954                                                 radeon_crtc_handle_vblank(rdev, 4);
7955                                         rdev->irq.stat_regs.cik.disp_int_cont4 &= ~LB_D5_VBLANK_INTERRUPT;
7956                                         DRM_DEBUG("IH: D5 vblank\n");
7957                                 }
7958                                 break;
7959                         case 1: /* D5 vline */
7960                                 if (rdev->irq.stat_regs.cik.disp_int_cont4 & LB_D5_VLINE_INTERRUPT) {
7961                                         rdev->irq.stat_regs.cik.disp_int_cont4 &= ~LB_D5_VLINE_INTERRUPT;
7962                                         DRM_DEBUG("IH: D5 vline\n");
7963                                 }
7964                                 break;
7965                         default:
7966                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
7967                                 break;
7968                         }
7969                         break;
7970                 case 6: /* D6 vblank/vline */
7971                         switch (src_data) {
7972                         case 0: /* D6 vblank */
7973                                 if (rdev->irq.stat_regs.cik.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT) {
7974                                         if (rdev->irq.crtc_vblank_int[5]) {
7975                                                 drm_handle_vblank(rdev->ddev, 5);
7976                                                 rdev->pm.vblank_sync = true;
7977                                                 wake_up(&rdev->irq.vblank_queue);
7978                                         }
7979                                         if (atomic_read(&rdev->irq.pflip[5]))
7980                                                 radeon_crtc_handle_vblank(rdev, 5);
7981                                         rdev->irq.stat_regs.cik.disp_int_cont5 &= ~LB_D6_VBLANK_INTERRUPT;
7982                                         DRM_DEBUG("IH: D6 vblank\n");
7983                                 }
7984                                 break;
7985                         case 1: /* D6 vline */
7986                                 if (rdev->irq.stat_regs.cik.disp_int_cont5 & LB_D6_VLINE_INTERRUPT) {
7987                                         rdev->irq.stat_regs.cik.disp_int_cont5 &= ~LB_D6_VLINE_INTERRUPT;
7988                                         DRM_DEBUG("IH: D6 vline\n");
7989                                 }
7990                                 break;
7991                         default:
7992                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
7993                                 break;
7994                         }
7995                         break;
7996                 case 8: /* D1 page flip */
7997                 case 10: /* D2 page flip */
7998                 case 12: /* D3 page flip */
7999                 case 14: /* D4 page flip */
8000                 case 16: /* D5 page flip */
8001                 case 18: /* D6 page flip */
8002                         DRM_DEBUG("IH: D%d flip\n", ((src_id - 8) >> 1) + 1);
8003                         if (radeon_use_pflipirq > 0)
8004                                 radeon_crtc_handle_flip(rdev, (src_id - 8) >> 1);
8005                         break;
8006                 case 42: /* HPD hotplug */
8007                         switch (src_data) {
8008                         case 0:
8009                                 if (rdev->irq.stat_regs.cik.disp_int & DC_HPD1_INTERRUPT) {
8010                                         rdev->irq.stat_regs.cik.disp_int &= ~DC_HPD1_INTERRUPT;
8011                                         queue_hotplug = true;
8012                                         DRM_DEBUG("IH: HPD1\n");
8013                                 }
8014                                 break;
8015                         case 1:
8016                                 if (rdev->irq.stat_regs.cik.disp_int_cont & DC_HPD2_INTERRUPT) {
8017                                         rdev->irq.stat_regs.cik.disp_int_cont &= ~DC_HPD2_INTERRUPT;
8018                                         queue_hotplug = true;
8019                                         DRM_DEBUG("IH: HPD2\n");
8020                                 }
8021                                 break;
8022                         case 2:
8023                                 if (rdev->irq.stat_regs.cik.disp_int_cont2 & DC_HPD3_INTERRUPT) {
8024                                         rdev->irq.stat_regs.cik.disp_int_cont2 &= ~DC_HPD3_INTERRUPT;
8025                                         queue_hotplug = true;
8026                                         DRM_DEBUG("IH: HPD3\n");
8027                                 }
8028                                 break;
8029                         case 3:
8030                                 if (rdev->irq.stat_regs.cik.disp_int_cont3 & DC_HPD4_INTERRUPT) {
8031                                         rdev->irq.stat_regs.cik.disp_int_cont3 &= ~DC_HPD4_INTERRUPT;
8032                                         queue_hotplug = true;
8033                                         DRM_DEBUG("IH: HPD4\n");
8034                                 }
8035                                 break;
8036                         case 4:
8037                                 if (rdev->irq.stat_regs.cik.disp_int_cont4 & DC_HPD5_INTERRUPT) {
8038                                         rdev->irq.stat_regs.cik.disp_int_cont4 &= ~DC_HPD5_INTERRUPT;
8039                                         queue_hotplug = true;
8040                                         DRM_DEBUG("IH: HPD5\n");
8041                                 }
8042                                 break;
8043                         case 5:
8044                                 if (rdev->irq.stat_regs.cik.disp_int_cont5 & DC_HPD6_INTERRUPT) {
8045                                         rdev->irq.stat_regs.cik.disp_int_cont5 &= ~DC_HPD6_INTERRUPT;
8046                                         queue_hotplug = true;
8047                                         DRM_DEBUG("IH: HPD6\n");
8048                                 }
8049                                 break;
8050                         default:
8051                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
8052                                 break;
8053                         }
8054                         break;
8055                 case 124: /* UVD */
8056                         DRM_DEBUG("IH: UVD int: 0x%08x\n", src_data);
8057                         radeon_fence_process(rdev, R600_RING_TYPE_UVD_INDEX);
8058                         break;
8059                 case 146:
8060                 case 147:
8061                         addr = RREG32(VM_CONTEXT1_PROTECTION_FAULT_ADDR);
8062                         status = RREG32(VM_CONTEXT1_PROTECTION_FAULT_STATUS);
8063                         mc_client = RREG32(VM_CONTEXT1_PROTECTION_FAULT_MCCLIENT);
8064                         /* reset addr and status */
8065                         WREG32_P(VM_CONTEXT1_CNTL2, 1, ~1);
8066                         if (addr == 0x0 && status == 0x0)
8067                                 break;
8068                         dev_err(rdev->dev, "GPU fault detected: %d 0x%08x\n", src_id, src_data);
8069                         dev_err(rdev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_ADDR   0x%08X\n",
8070                                 addr);
8071                         dev_err(rdev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
8072                                 status);
8073                         cik_vm_decode_fault(rdev, status, addr, mc_client);
8074                         break;
8075                 case 167: /* VCE */
8076                         DRM_DEBUG("IH: VCE int: 0x%08x\n", src_data);
8077                         switch (src_data) {
8078                         case 0:
8079                                 radeon_fence_process(rdev, TN_RING_TYPE_VCE1_INDEX);
8080                                 break;
8081                         case 1:
8082                                 radeon_fence_process(rdev, TN_RING_TYPE_VCE2_INDEX);
8083                                 break;
8084                         default:
8085                                 DRM_ERROR("Unhandled interrupt: %d %d\n", src_id, src_data);
8086                                 break;
8087                         }
8088                         break;
8089                 case 176: /* GFX RB CP_INT */
8090                 case 177: /* GFX IB CP_INT */
8091                         radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
8092                         break;
8093                 case 181: /* CP EOP event */
8094                         DRM_DEBUG("IH: CP EOP\n");
8095                         /* XXX check the bitfield order! */
8096                         me_id = (ring_id & 0x60) >> 5;
8097                         pipe_id = (ring_id & 0x18) >> 3;
8098                         queue_id = (ring_id & 0x7) >> 0;
8099                         switch (me_id) {
8100                         case 0:
8101                                 radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
8102                                 break;
8103                         case 1:
8104                         case 2:
8105                                 if ((cp1_ring->me == me_id) & (cp1_ring->pipe == pipe_id))
8106                                         radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
8107                                 if ((cp2_ring->me == me_id) & (cp2_ring->pipe == pipe_id))
8108                                         radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
8109                                 break;
8110                         }
8111                         break;
8112                 case 184: /* CP Privileged reg access */
8113                         DRM_ERROR("Illegal register access in command stream\n");
8114                         /* XXX check the bitfield order! */
8115                         me_id = (ring_id & 0x60) >> 5;
8116                         pipe_id = (ring_id & 0x18) >> 3;
8117                         queue_id = (ring_id & 0x7) >> 0;
8118                         switch (me_id) {
8119                         case 0:
8120                                 /* This results in a full GPU reset, but all we need to do is soft
8121                                  * reset the CP for gfx
8122                                  */
8123                                 queue_reset = true;
8124                                 break;
8125                         case 1:
8126                                 /* XXX compute */
8127                                 queue_reset = true;
8128                                 break;
8129                         case 2:
8130                                 /* XXX compute */
8131                                 queue_reset = true;
8132                                 break;
8133                         }
8134                         break;
8135                 case 185: /* CP Privileged inst */
8136                         DRM_ERROR("Illegal instruction in command stream\n");
8137                         /* XXX check the bitfield order! */
8138                         me_id = (ring_id & 0x60) >> 5;
8139                         pipe_id = (ring_id & 0x18) >> 3;
8140                         queue_id = (ring_id & 0x7) >> 0;
8141                         switch (me_id) {
8142                         case 0:
8143                                 /* This results in a full GPU reset, but all we need to do is soft
8144                                  * reset the CP for gfx
8145                                  */
8146                                 queue_reset = true;
8147                                 break;
8148                         case 1:
8149                                 /* XXX compute */
8150                                 queue_reset = true;
8151                                 break;
8152                         case 2:
8153                                 /* XXX compute */
8154                                 queue_reset = true;
8155                                 break;
8156                         }
8157                         break;
8158                 case 224: /* SDMA trap event */
8159                         /* XXX check the bitfield order! */
8160                         me_id = (ring_id & 0x3) >> 0;
8161                         queue_id = (ring_id & 0xc) >> 2;
8162                         DRM_DEBUG("IH: SDMA trap\n");
8163                         switch (me_id) {
8164                         case 0:
8165                                 switch (queue_id) {
8166                                 case 0:
8167                                         radeon_fence_process(rdev, R600_RING_TYPE_DMA_INDEX);
8168                                         break;
8169                                 case 1:
8170                                         /* XXX compute */
8171                                         break;
8172                                 case 2:
8173                                         /* XXX compute */
8174                                         break;
8175                                 }
8176                                 break;
8177                         case 1:
8178                                 switch (queue_id) {
8179                                 case 0:
8180                                         radeon_fence_process(rdev, CAYMAN_RING_TYPE_DMA1_INDEX);
8181                                         break;
8182                                 case 1:
8183                                         /* XXX compute */
8184                                         break;
8185                                 case 2:
8186                                         /* XXX compute */
8187                                         break;
8188                                 }
8189                                 break;
8190                         }
8191                         break;
8192                 case 230: /* thermal low to high */
8193                         DRM_DEBUG("IH: thermal low to high\n");
8194                         rdev->pm.dpm.thermal.high_to_low = false;
8195                         queue_thermal = true;
8196                         break;
8197                 case 231: /* thermal high to low */
8198                         DRM_DEBUG("IH: thermal high to low\n");
8199                         rdev->pm.dpm.thermal.high_to_low = true;
8200                         queue_thermal = true;
8201                         break;
8202                 case 233: /* GUI IDLE */
8203                         DRM_DEBUG("IH: GUI idle\n");
8204                         break;
8205                 case 241: /* SDMA Privileged inst */
8206                 case 247: /* SDMA Privileged inst */
8207                         DRM_ERROR("Illegal instruction in SDMA command stream\n");
8208                         /* XXX check the bitfield order! */
8209                         me_id = (ring_id & 0x3) >> 0;
8210                         queue_id = (ring_id & 0xc) >> 2;
8211                         switch (me_id) {
8212                         case 0:
8213                                 switch (queue_id) {
8214                                 case 0:
8215                                         queue_reset = true;
8216                                         break;
8217                                 case 1:
8218                                         /* XXX compute */
8219                                         queue_reset = true;
8220                                         break;
8221                                 case 2:
8222                                         /* XXX compute */
8223                                         queue_reset = true;
8224                                         break;
8225                                 }
8226                                 break;
8227                         case 1:
8228                                 switch (queue_id) {
8229                                 case 0:
8230                                         queue_reset = true;
8231                                         break;
8232                                 case 1:
8233                                         /* XXX compute */
8234                                         queue_reset = true;
8235                                         break;
8236                                 case 2:
8237                                         /* XXX compute */
8238                                         queue_reset = true;
8239                                         break;
8240                                 }
8241                                 break;
8242                         }
8243                         break;
8244                 default:
8245                         DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
8246                         break;
8247                 }
8248
8249                 /* wptr/rptr are in bytes! */
8250                 rptr += 16;
8251                 rptr &= rdev->ih.ptr_mask;
8252         }
8253         if (queue_hotplug)
8254                 schedule_work(&rdev->hotplug_work);
8255         if (queue_reset) {
8256                 rdev->needs_reset = true;
8257                 wake_up_all(&rdev->fence_queue);
8258         }
8259         if (queue_thermal)
8260                 schedule_work(&rdev->pm.dpm.thermal.work);
8261         rdev->ih.rptr = rptr;
8262         WREG32(IH_RB_RPTR, rdev->ih.rptr);
8263         atomic_set(&rdev->ih.lock, 0);
8264
8265         /* make sure wptr hasn't changed while processing */
8266         wptr = cik_get_ih_wptr(rdev);
8267         if (wptr != rptr)
8268                 goto restart_ih;
8269
8270         return IRQ_HANDLED;
8271 }
8272
8273 /*
8274  * startup/shutdown callbacks
8275  */
8276 /**
8277  * cik_startup - program the asic to a functional state
8278  *
8279  * @rdev: radeon_device pointer
8280  *
8281  * Programs the asic to a functional state (CIK).
8282  * Called by cik_init() and cik_resume().
8283  * Returns 0 for success, error for failure.
8284  */
8285 static int cik_startup(struct radeon_device *rdev)
8286 {
8287         struct radeon_ring *ring;
8288         u32 nop;
8289         int r;
8290
8291         /* enable pcie gen2/3 link */
8292         cik_pcie_gen3_enable(rdev);
8293         /* enable aspm */
8294         cik_program_aspm(rdev);
8295
8296         /* scratch needs to be initialized before MC */
8297         r = r600_vram_scratch_init(rdev);
8298         if (r)
8299                 return r;
8300
8301         cik_mc_program(rdev);
8302
8303         if (!(rdev->flags & RADEON_IS_IGP) && !rdev->pm.dpm_enabled) {
8304                 r = ci_mc_load_microcode(rdev);
8305                 if (r) {
8306                         DRM_ERROR("Failed to load MC firmware!\n");
8307                         return r;
8308                 }
8309         }
8310
8311         r = cik_pcie_gart_enable(rdev);
8312         if (r)
8313                 return r;
8314         cik_gpu_init(rdev);
8315
8316         /* allocate rlc buffers */
8317         if (rdev->flags & RADEON_IS_IGP) {
8318                 if (rdev->family == CHIP_KAVERI) {
8319                         rdev->rlc.reg_list = spectre_rlc_save_restore_register_list;
8320                         rdev->rlc.reg_list_size =
8321                                 (u32)ARRAY_SIZE(spectre_rlc_save_restore_register_list);
8322                 } else {
8323                         rdev->rlc.reg_list = kalindi_rlc_save_restore_register_list;
8324                         rdev->rlc.reg_list_size =
8325                                 (u32)ARRAY_SIZE(kalindi_rlc_save_restore_register_list);
8326                 }
8327         }
8328         rdev->rlc.cs_data = ci_cs_data;
8329         rdev->rlc.cp_table_size = CP_ME_TABLE_SIZE * 5 * 4;
8330         r = sumo_rlc_init(rdev);
8331         if (r) {
8332                 DRM_ERROR("Failed to init rlc BOs!\n");
8333                 return r;
8334         }
8335
8336         /* allocate wb buffer */
8337         r = radeon_wb_init(rdev);
8338         if (r)
8339                 return r;
8340
8341         /* allocate mec buffers */
8342         r = cik_mec_init(rdev);
8343         if (r) {
8344                 DRM_ERROR("Failed to init MEC BOs!\n");
8345                 return r;
8346         }
8347
8348         r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
8349         if (r) {
8350                 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
8351                 return r;
8352         }
8353
8354         r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
8355         if (r) {
8356                 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
8357                 return r;
8358         }
8359
8360         r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
8361         if (r) {
8362                 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
8363                 return r;
8364         }
8365
8366         r = radeon_fence_driver_start_ring(rdev, R600_RING_TYPE_DMA_INDEX);
8367         if (r) {
8368                 dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
8369                 return r;
8370         }
8371
8372         r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_DMA1_INDEX);
8373         if (r) {
8374                 dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
8375                 return r;
8376         }
8377
8378         r = radeon_uvd_resume(rdev);
8379         if (!r) {
8380                 r = uvd_v4_2_resume(rdev);
8381                 if (!r) {
8382                         r = radeon_fence_driver_start_ring(rdev,
8383                                                            R600_RING_TYPE_UVD_INDEX);
8384                         if (r)
8385                                 dev_err(rdev->dev, "UVD fences init error (%d).\n", r);
8386                 }
8387         }
8388         if (r)
8389                 rdev->ring[R600_RING_TYPE_UVD_INDEX].ring_size = 0;
8390
8391         r = radeon_vce_resume(rdev);
8392         if (!r) {
8393                 r = vce_v2_0_resume(rdev);
8394                 if (!r)
8395                         r = radeon_fence_driver_start_ring(rdev,
8396                                                            TN_RING_TYPE_VCE1_INDEX);
8397                 if (!r)
8398                         r = radeon_fence_driver_start_ring(rdev,
8399                                                            TN_RING_TYPE_VCE2_INDEX);
8400         }
8401         if (r) {
8402                 dev_err(rdev->dev, "VCE init error (%d).\n", r);
8403                 rdev->ring[TN_RING_TYPE_VCE1_INDEX].ring_size = 0;
8404                 rdev->ring[TN_RING_TYPE_VCE2_INDEX].ring_size = 0;
8405         }
8406
8407         /* Enable IRQ */
8408         if (!rdev->irq.installed) {
8409                 r = radeon_irq_kms_init(rdev);
8410                 if (r)
8411                         return r;
8412         }
8413
8414         r = cik_irq_init(rdev);
8415         if (r) {
8416                 DRM_ERROR("radeon: IH init failed (%d).\n", r);
8417                 radeon_irq_kms_fini(rdev);
8418                 return r;
8419         }
8420         cik_irq_set(rdev);
8421
8422         if (rdev->family == CHIP_HAWAII) {
8423                 if (rdev->new_fw)
8424                         nop = PACKET3(PACKET3_NOP, 0x3FFF);
8425                 else
8426                         nop = RADEON_CP_PACKET2;
8427         } else {
8428                 nop = PACKET3(PACKET3_NOP, 0x3FFF);
8429         }
8430
8431         ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
8432         r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP_RPTR_OFFSET,
8433                              nop);
8434         if (r)
8435                 return r;
8436
8437         /* set up the compute queues */
8438         /* type-2 packets are deprecated on MEC, use type-3 instead */
8439         ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
8440         r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP1_RPTR_OFFSET,
8441                              nop);
8442         if (r)
8443                 return r;
8444         ring->me = 1; /* first MEC */
8445         ring->pipe = 0; /* first pipe */
8446         ring->queue = 0; /* first queue */
8447         ring->wptr_offs = CIK_WB_CP1_WPTR_OFFSET;
8448
8449         /* type-2 packets are deprecated on MEC, use type-3 instead */
8450         ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
8451         r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP2_RPTR_OFFSET,
8452                              nop);
8453         if (r)
8454                 return r;
8455         /* dGPU only have 1 MEC */
8456         ring->me = 1; /* first MEC */
8457         ring->pipe = 0; /* first pipe */
8458         ring->queue = 1; /* second queue */
8459         ring->wptr_offs = CIK_WB_CP2_WPTR_OFFSET;
8460
8461         ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
8462         r = radeon_ring_init(rdev, ring, ring->ring_size, R600_WB_DMA_RPTR_OFFSET,
8463                              SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0));
8464         if (r)
8465                 return r;
8466
8467         ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
8468         r = radeon_ring_init(rdev, ring, ring->ring_size, CAYMAN_WB_DMA1_RPTR_OFFSET,
8469                              SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0));
8470         if (r)
8471                 return r;
8472
8473         r = cik_cp_resume(rdev);
8474         if (r)
8475                 return r;
8476
8477         r = cik_sdma_resume(rdev);
8478         if (r)
8479                 return r;
8480
8481         ring = &rdev->ring[R600_RING_TYPE_UVD_INDEX];
8482         if (ring->ring_size) {
8483                 r = radeon_ring_init(rdev, ring, ring->ring_size, 0,
8484                                      RADEON_CP_PACKET2);
8485                 if (!r)
8486                         r = uvd_v1_0_init(rdev);
8487                 if (r)
8488                         DRM_ERROR("radeon: failed initializing UVD (%d).\n", r);
8489         }
8490
8491         r = -ENOENT;
8492
8493         ring = &rdev->ring[TN_RING_TYPE_VCE1_INDEX];
8494         if (ring->ring_size)
8495                 r = radeon_ring_init(rdev, ring, ring->ring_size, 0,
8496                                      VCE_CMD_NO_OP);
8497
8498         ring = &rdev->ring[TN_RING_TYPE_VCE2_INDEX];
8499         if (ring->ring_size)
8500                 r = radeon_ring_init(rdev, ring, ring->ring_size, 0,
8501                                      VCE_CMD_NO_OP);
8502
8503         if (!r)
8504                 r = vce_v1_0_init(rdev);
8505         else if (r != -ENOENT)
8506                 DRM_ERROR("radeon: failed initializing VCE (%d).\n", r);
8507
8508         r = radeon_ib_pool_init(rdev);
8509         if (r) {
8510                 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
8511                 return r;
8512         }
8513
8514         r = radeon_vm_manager_init(rdev);
8515         if (r) {
8516                 dev_err(rdev->dev, "vm manager initialization failed (%d).\n", r);
8517                 return r;
8518         }
8519
8520         r = dce6_audio_init(rdev);
8521         if (r)
8522                 return r;
8523
8524         return 0;
8525 }
8526
8527 /**
8528  * cik_resume - resume the asic to a functional state
8529  *
8530  * @rdev: radeon_device pointer
8531  *
8532  * Programs the asic to a functional state (CIK).
8533  * Called at resume.
8534  * Returns 0 for success, error for failure.
8535  */
8536 int cik_resume(struct radeon_device *rdev)
8537 {
8538         int r;
8539
8540         /* post card */
8541         atom_asic_init(rdev->mode_info.atom_context);
8542
8543         /* init golden registers */
8544         cik_init_golden_registers(rdev);
8545
8546         if (rdev->pm.pm_method == PM_METHOD_DPM)
8547                 radeon_pm_resume(rdev);
8548
8549         rdev->accel_working = true;
8550         r = cik_startup(rdev);
8551         if (r) {
8552                 DRM_ERROR("cik startup failed on resume\n");
8553                 rdev->accel_working = false;
8554                 return r;
8555         }
8556
8557         return r;
8558
8559 }
8560
8561 /**
8562  * cik_suspend - suspend the asic
8563  *
8564  * @rdev: radeon_device pointer
8565  *
8566  * Bring the chip into a state suitable for suspend (CIK).
8567  * Called at suspend.
8568  * Returns 0 for success.
8569  */
8570 int cik_suspend(struct radeon_device *rdev)
8571 {
8572         radeon_pm_suspend(rdev);
8573         dce6_audio_fini(rdev);
8574         radeon_vm_manager_fini(rdev);
8575         cik_cp_enable(rdev, false);
8576         cik_sdma_enable(rdev, false);
8577         uvd_v1_0_fini(rdev);
8578         radeon_uvd_suspend(rdev);
8579         radeon_vce_suspend(rdev);
8580         cik_fini_pg(rdev);
8581         cik_fini_cg(rdev);
8582         cik_irq_suspend(rdev);
8583         radeon_wb_disable(rdev);
8584         cik_pcie_gart_disable(rdev);
8585         return 0;
8586 }
8587
8588 /* Plan is to move initialization in that function and use
8589  * helper function so that radeon_device_init pretty much
8590  * do nothing more than calling asic specific function. This
8591  * should also allow to remove a bunch of callback function
8592  * like vram_info.
8593  */
8594 /**
8595  * cik_init - asic specific driver and hw init
8596  *
8597  * @rdev: radeon_device pointer
8598  *
8599  * Setup asic specific driver variables and program the hw
8600  * to a functional state (CIK).
8601  * Called at driver startup.
8602  * Returns 0 for success, errors for failure.
8603  */
8604 int cik_init(struct radeon_device *rdev)
8605 {
8606         struct radeon_ring *ring;
8607         int r;
8608
8609         /* Read BIOS */
8610         if (!radeon_get_bios(rdev)) {
8611                 if (ASIC_IS_AVIVO(rdev))
8612                         return -EINVAL;
8613         }
8614         /* Must be an ATOMBIOS */
8615         if (!rdev->is_atom_bios) {
8616                 dev_err(rdev->dev, "Expecting atombios for cayman GPU\n");
8617                 return -EINVAL;
8618         }
8619         r = radeon_atombios_init(rdev);
8620         if (r)
8621                 return r;
8622
8623         /* Post card if necessary */
8624         if (!radeon_card_posted(rdev)) {
8625                 if (!rdev->bios) {
8626                         dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
8627                         return -EINVAL;
8628                 }
8629                 DRM_INFO("GPU not posted. posting now...\n");
8630                 atom_asic_init(rdev->mode_info.atom_context);
8631         }
8632         /* init golden registers */
8633         cik_init_golden_registers(rdev);
8634         /* Initialize scratch registers */
8635         cik_scratch_init(rdev);
8636         /* Initialize surface registers */
8637         radeon_surface_init(rdev);
8638         /* Initialize clocks */
8639         radeon_get_clock_info(rdev->ddev);
8640
8641         /* Fence driver */
8642         r = radeon_fence_driver_init(rdev);
8643         if (r)
8644                 return r;
8645
8646         /* initialize memory controller */
8647         r = cik_mc_init(rdev);
8648         if (r)
8649                 return r;
8650         /* Memory manager */
8651         r = radeon_bo_init(rdev);
8652         if (r)
8653                 return r;
8654
8655         if (rdev->flags & RADEON_IS_IGP) {
8656                 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->ce_fw ||
8657                     !rdev->mec_fw || !rdev->sdma_fw || !rdev->rlc_fw) {
8658                         r = cik_init_microcode(rdev);
8659                         if (r) {
8660                                 DRM_ERROR("Failed to load firmware!\n");
8661                                 return r;
8662                         }
8663                 }
8664         } else {
8665                 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->ce_fw ||
8666                     !rdev->mec_fw || !rdev->sdma_fw || !rdev->rlc_fw ||
8667                     !rdev->mc_fw) {
8668                         r = cik_init_microcode(rdev);
8669                         if (r) {
8670                                 DRM_ERROR("Failed to load firmware!\n");
8671                                 return r;
8672                         }
8673                 }
8674         }
8675
8676         /* Initialize power management */
8677         radeon_pm_init(rdev);
8678
8679         ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
8680         ring->ring_obj = NULL;
8681         r600_ring_init(rdev, ring, 1024 * 1024);
8682
8683         ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
8684         ring->ring_obj = NULL;
8685         r600_ring_init(rdev, ring, 1024 * 1024);
8686         r = radeon_doorbell_get(rdev, &ring->doorbell_index);
8687         if (r)
8688                 return r;
8689
8690         ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
8691         ring->ring_obj = NULL;
8692         r600_ring_init(rdev, ring, 1024 * 1024);
8693         r = radeon_doorbell_get(rdev, &ring->doorbell_index);
8694         if (r)
8695                 return r;
8696
8697         ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
8698         ring->ring_obj = NULL;
8699         r600_ring_init(rdev, ring, 256 * 1024);
8700
8701         ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
8702         ring->ring_obj = NULL;
8703         r600_ring_init(rdev, ring, 256 * 1024);
8704
8705         r = radeon_uvd_init(rdev);
8706         if (!r) {
8707                 ring = &rdev->ring[R600_RING_TYPE_UVD_INDEX];
8708                 ring->ring_obj = NULL;
8709                 r600_ring_init(rdev, ring, 4096);
8710         }
8711
8712         r = radeon_vce_init(rdev);
8713         if (!r) {
8714                 ring = &rdev->ring[TN_RING_TYPE_VCE1_INDEX];
8715                 ring->ring_obj = NULL;
8716                 r600_ring_init(rdev, ring, 4096);
8717
8718                 ring = &rdev->ring[TN_RING_TYPE_VCE2_INDEX];
8719                 ring->ring_obj = NULL;
8720                 r600_ring_init(rdev, ring, 4096);
8721         }
8722
8723         rdev->ih.ring_obj = NULL;
8724         r600_ih_ring_init(rdev, 64 * 1024);
8725
8726         r = r600_pcie_gart_init(rdev);
8727         if (r)
8728                 return r;
8729
8730         rdev->accel_working = true;
8731         r = cik_startup(rdev);
8732         if (r) {
8733                 dev_err(rdev->dev, "disabling GPU acceleration\n");
8734                 cik_cp_fini(rdev);
8735                 cik_sdma_fini(rdev);
8736                 cik_irq_fini(rdev);
8737                 sumo_rlc_fini(rdev);
8738                 cik_mec_fini(rdev);
8739                 radeon_wb_fini(rdev);
8740                 radeon_ib_pool_fini(rdev);
8741                 radeon_vm_manager_fini(rdev);
8742                 radeon_irq_kms_fini(rdev);
8743                 cik_pcie_gart_fini(rdev);
8744                 rdev->accel_working = false;
8745         }
8746
8747         /* Don't start up if the MC ucode is missing.
8748          * The default clocks and voltages before the MC ucode
8749          * is loaded are not suffient for advanced operations.
8750          */
8751         if (!rdev->mc_fw && !(rdev->flags & RADEON_IS_IGP)) {
8752                 DRM_ERROR("radeon: MC ucode required for NI+.\n");
8753                 return -EINVAL;
8754         }
8755
8756         return 0;
8757 }
8758
8759 /**
8760  * cik_fini - asic specific driver and hw fini
8761  *
8762  * @rdev: radeon_device pointer
8763  *
8764  * Tear down the asic specific driver variables and program the hw
8765  * to an idle state (CIK).
8766  * Called at driver unload.
8767  */
8768 void cik_fini(struct radeon_device *rdev)
8769 {
8770         radeon_pm_fini(rdev);
8771         cik_cp_fini(rdev);
8772         cik_sdma_fini(rdev);
8773         cik_fini_pg(rdev);
8774         cik_fini_cg(rdev);
8775         cik_irq_fini(rdev);
8776         sumo_rlc_fini(rdev);
8777         cik_mec_fini(rdev);
8778         radeon_wb_fini(rdev);
8779         radeon_vm_manager_fini(rdev);
8780         radeon_ib_pool_fini(rdev);
8781         radeon_irq_kms_fini(rdev);
8782         uvd_v1_0_fini(rdev);
8783         radeon_uvd_fini(rdev);
8784         radeon_vce_fini(rdev);
8785         cik_pcie_gart_fini(rdev);
8786         r600_vram_scratch_fini(rdev);
8787         radeon_gem_fini(rdev);
8788         radeon_fence_driver_fini(rdev);
8789         radeon_bo_fini(rdev);
8790         radeon_atombios_fini(rdev);
8791         kfree(rdev->bios);
8792         rdev->bios = NULL;
8793 }
8794
8795 void dce8_program_fmt(struct drm_encoder *encoder)
8796 {
8797         struct drm_device *dev = encoder->dev;
8798         struct radeon_device *rdev = dev->dev_private;
8799         struct radeon_encoder *radeon_encoder = to_radeon_encoder(encoder);
8800         struct radeon_crtc *radeon_crtc = to_radeon_crtc(encoder->crtc);
8801         struct drm_connector *connector = radeon_get_connector_for_encoder(encoder);
8802         int bpc = 0;
8803         u32 tmp = 0;
8804         enum radeon_connector_dither dither = RADEON_FMT_DITHER_DISABLE;
8805
8806         if (connector) {
8807                 struct radeon_connector *radeon_connector = to_radeon_connector(connector);
8808                 bpc = radeon_get_monitor_bpc(connector);
8809                 dither = radeon_connector->dither;
8810         }
8811
8812         /* LVDS/eDP FMT is set up by atom */
8813         if (radeon_encoder->devices & ATOM_DEVICE_LCD_SUPPORT)
8814                 return;
8815
8816         /* not needed for analog */
8817         if ((radeon_encoder->encoder_id == ENCODER_OBJECT_ID_INTERNAL_KLDSCP_DAC1) ||
8818             (radeon_encoder->encoder_id == ENCODER_OBJECT_ID_INTERNAL_KLDSCP_DAC2))
8819                 return;
8820
8821         if (bpc == 0)
8822                 return;
8823
8824         switch (bpc) {
8825         case 6:
8826                 if (dither == RADEON_FMT_DITHER_ENABLE)
8827                         /* XXX sort out optimal dither settings */
8828                         tmp |= (FMT_FRAME_RANDOM_ENABLE | FMT_HIGHPASS_RANDOM_ENABLE |
8829                                 FMT_SPATIAL_DITHER_EN | FMT_SPATIAL_DITHER_DEPTH(0));
8830                 else
8831                         tmp |= (FMT_TRUNCATE_EN | FMT_TRUNCATE_DEPTH(0));
8832                 break;
8833         case 8:
8834                 if (dither == RADEON_FMT_DITHER_ENABLE)
8835                         /* XXX sort out optimal dither settings */
8836                         tmp |= (FMT_FRAME_RANDOM_ENABLE | FMT_HIGHPASS_RANDOM_ENABLE |
8837                                 FMT_RGB_RANDOM_ENABLE |
8838                                 FMT_SPATIAL_DITHER_EN | FMT_SPATIAL_DITHER_DEPTH(1));
8839                 else
8840                         tmp |= (FMT_TRUNCATE_EN | FMT_TRUNCATE_DEPTH(1));
8841                 break;
8842         case 10:
8843                 if (dither == RADEON_FMT_DITHER_ENABLE)
8844                         /* XXX sort out optimal dither settings */
8845                         tmp |= (FMT_FRAME_RANDOM_ENABLE | FMT_HIGHPASS_RANDOM_ENABLE |
8846                                 FMT_RGB_RANDOM_ENABLE |
8847                                 FMT_SPATIAL_DITHER_EN | FMT_SPATIAL_DITHER_DEPTH(2));
8848                 else
8849                         tmp |= (FMT_TRUNCATE_EN | FMT_TRUNCATE_DEPTH(2));
8850                 break;
8851         default:
8852                 /* not needed */
8853                 break;
8854         }
8855
8856         WREG32(FMT_BIT_DEPTH_CONTROL + radeon_crtc->crtc_offset, tmp);
8857 }
8858
8859 /* display watermark setup */
8860 /**
8861  * dce8_line_buffer_adjust - Set up the line buffer
8862  *
8863  * @rdev: radeon_device pointer
8864  * @radeon_crtc: the selected display controller
8865  * @mode: the current display mode on the selected display
8866  * controller
8867  *
8868  * Setup up the line buffer allocation for
8869  * the selected display controller (CIK).
8870  * Returns the line buffer size in pixels.
8871  */
8872 static u32 dce8_line_buffer_adjust(struct radeon_device *rdev,
8873                                    struct radeon_crtc *radeon_crtc,
8874                                    struct drm_display_mode *mode)
8875 {
8876         u32 tmp, buffer_alloc, i;
8877         u32 pipe_offset = radeon_crtc->crtc_id * 0x20;
8878         /*
8879          * Line Buffer Setup
8880          * There are 6 line buffers, one for each display controllers.
8881          * There are 3 partitions per LB. Select the number of partitions
8882          * to enable based on the display width.  For display widths larger
8883          * than 4096, you need use to use 2 display controllers and combine
8884          * them using the stereo blender.
8885          */
8886         if (radeon_crtc->base.enabled && mode) {
8887                 if (mode->crtc_hdisplay < 1920) {
8888                         tmp = 1;
8889                         buffer_alloc = 2;
8890                 } else if (mode->crtc_hdisplay < 2560) {
8891                         tmp = 2;
8892                         buffer_alloc = 2;
8893                 } else if (mode->crtc_hdisplay < 4096) {
8894                         tmp = 0;
8895                         buffer_alloc = (rdev->flags & RADEON_IS_IGP) ? 2 : 4;
8896                 } else {
8897                         DRM_DEBUG_KMS("Mode too big for LB!\n");
8898                         tmp = 0;
8899                         buffer_alloc = (rdev->flags & RADEON_IS_IGP) ? 2 : 4;
8900                 }
8901         } else {
8902                 tmp = 1;
8903                 buffer_alloc = 0;
8904         }
8905
8906         WREG32(LB_MEMORY_CTRL + radeon_crtc->crtc_offset,
8907                LB_MEMORY_CONFIG(tmp) | LB_MEMORY_SIZE(0x6B0));
8908
8909         WREG32(PIPE0_DMIF_BUFFER_CONTROL + pipe_offset,
8910                DMIF_BUFFERS_ALLOCATED(buffer_alloc));
8911         for (i = 0; i < rdev->usec_timeout; i++) {
8912                 if (RREG32(PIPE0_DMIF_BUFFER_CONTROL + pipe_offset) &
8913                     DMIF_BUFFERS_ALLOCATED_COMPLETED)
8914                         break;
8915                 udelay(1);
8916         }
8917
8918         if (radeon_crtc->base.enabled && mode) {
8919                 switch (tmp) {
8920                 case 0:
8921                 default:
8922                         return 4096 * 2;
8923                 case 1:
8924                         return 1920 * 2;
8925                 case 2:
8926                         return 2560 * 2;
8927                 }
8928         }
8929
8930         /* controller not enabled, so no lb used */
8931         return 0;
8932 }
8933
8934 /**
8935  * cik_get_number_of_dram_channels - get the number of dram channels
8936  *
8937  * @rdev: radeon_device pointer
8938  *
8939  * Look up the number of video ram channels (CIK).
8940  * Used for display watermark bandwidth calculations
8941  * Returns the number of dram channels
8942  */
8943 static u32 cik_get_number_of_dram_channels(struct radeon_device *rdev)
8944 {
8945         u32 tmp = RREG32(MC_SHARED_CHMAP);
8946
8947         switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
8948         case 0:
8949         default:
8950                 return 1;
8951         case 1:
8952                 return 2;
8953         case 2:
8954                 return 4;
8955         case 3:
8956                 return 8;
8957         case 4:
8958                 return 3;
8959         case 5:
8960                 return 6;
8961         case 6:
8962                 return 10;
8963         case 7:
8964                 return 12;
8965         case 8:
8966                 return 16;
8967         }
8968 }
8969
8970 struct dce8_wm_params {
8971         u32 dram_channels; /* number of dram channels */
8972         u32 yclk;          /* bandwidth per dram data pin in kHz */
8973         u32 sclk;          /* engine clock in kHz */
8974         u32 disp_clk;      /* display clock in kHz */
8975         u32 src_width;     /* viewport width */
8976         u32 active_time;   /* active display time in ns */
8977         u32 blank_time;    /* blank time in ns */
8978         bool interlaced;    /* mode is interlaced */
8979         fixed20_12 vsc;    /* vertical scale ratio */
8980         u32 num_heads;     /* number of active crtcs */
8981         u32 bytes_per_pixel; /* bytes per pixel display + overlay */
8982         u32 lb_size;       /* line buffer allocated to pipe */
8983         u32 vtaps;         /* vertical scaler taps */
8984 };
8985
8986 /**
8987  * dce8_dram_bandwidth - get the dram bandwidth
8988  *
8989  * @wm: watermark calculation data
8990  *
8991  * Calculate the raw dram bandwidth (CIK).
8992  * Used for display watermark bandwidth calculations
8993  * Returns the dram bandwidth in MBytes/s
8994  */
8995 static u32 dce8_dram_bandwidth(struct dce8_wm_params *wm)
8996 {
8997         /* Calculate raw DRAM Bandwidth */
8998         fixed20_12 dram_efficiency; /* 0.7 */
8999         fixed20_12 yclk, dram_channels, bandwidth;
9000         fixed20_12 a;
9001
9002         a.full = dfixed_const(1000);
9003         yclk.full = dfixed_const(wm->yclk);
9004         yclk.full = dfixed_div(yclk, a);
9005         dram_channels.full = dfixed_const(wm->dram_channels * 4);
9006         a.full = dfixed_const(10);
9007         dram_efficiency.full = dfixed_const(7);
9008         dram_efficiency.full = dfixed_div(dram_efficiency, a);
9009         bandwidth.full = dfixed_mul(dram_channels, yclk);
9010         bandwidth.full = dfixed_mul(bandwidth, dram_efficiency);
9011
9012         return dfixed_trunc(bandwidth);
9013 }
9014
9015 /**
9016  * dce8_dram_bandwidth_for_display - get the dram bandwidth for display
9017  *
9018  * @wm: watermark calculation data
9019  *
9020  * Calculate the dram bandwidth used for display (CIK).
9021  * Used for display watermark bandwidth calculations
9022  * Returns the dram bandwidth for display in MBytes/s
9023  */
9024 static u32 dce8_dram_bandwidth_for_display(struct dce8_wm_params *wm)
9025 {
9026         /* Calculate DRAM Bandwidth and the part allocated to display. */
9027         fixed20_12 disp_dram_allocation; /* 0.3 to 0.7 */
9028         fixed20_12 yclk, dram_channels, bandwidth;
9029         fixed20_12 a;
9030
9031         a.full = dfixed_const(1000);
9032         yclk.full = dfixed_const(wm->yclk);
9033         yclk.full = dfixed_div(yclk, a);
9034         dram_channels.full = dfixed_const(wm->dram_channels * 4);
9035         a.full = dfixed_const(10);
9036         disp_dram_allocation.full = dfixed_const(3); /* XXX worse case value 0.3 */
9037         disp_dram_allocation.full = dfixed_div(disp_dram_allocation, a);
9038         bandwidth.full = dfixed_mul(dram_channels, yclk);
9039         bandwidth.full = dfixed_mul(bandwidth, disp_dram_allocation);
9040
9041         return dfixed_trunc(bandwidth);
9042 }
9043
9044 /**
9045  * dce8_data_return_bandwidth - get the data return bandwidth
9046  *
9047  * @wm: watermark calculation data
9048  *
9049  * Calculate the data return bandwidth used for display (CIK).
9050  * Used for display watermark bandwidth calculations
9051  * Returns the data return bandwidth in MBytes/s
9052  */
9053 static u32 dce8_data_return_bandwidth(struct dce8_wm_params *wm)
9054 {
9055         /* Calculate the display Data return Bandwidth */
9056         fixed20_12 return_efficiency; /* 0.8 */
9057         fixed20_12 sclk, bandwidth;
9058         fixed20_12 a;
9059
9060         a.full = dfixed_const(1000);
9061         sclk.full = dfixed_const(wm->sclk);
9062         sclk.full = dfixed_div(sclk, a);
9063         a.full = dfixed_const(10);
9064         return_efficiency.full = dfixed_const(8);
9065         return_efficiency.full = dfixed_div(return_efficiency, a);
9066         a.full = dfixed_const(32);
9067         bandwidth.full = dfixed_mul(a, sclk);
9068         bandwidth.full = dfixed_mul(bandwidth, return_efficiency);
9069
9070         return dfixed_trunc(bandwidth);
9071 }
9072
9073 /**
9074  * dce8_dmif_request_bandwidth - get the dmif bandwidth
9075  *
9076  * @wm: watermark calculation data
9077  *
9078  * Calculate the dmif bandwidth used for display (CIK).
9079  * Used for display watermark bandwidth calculations
9080  * Returns the dmif bandwidth in MBytes/s
9081  */
9082 static u32 dce8_dmif_request_bandwidth(struct dce8_wm_params *wm)
9083 {
9084         /* Calculate the DMIF Request Bandwidth */
9085         fixed20_12 disp_clk_request_efficiency; /* 0.8 */
9086         fixed20_12 disp_clk, bandwidth;
9087         fixed20_12 a, b;
9088
9089         a.full = dfixed_const(1000);
9090         disp_clk.full = dfixed_const(wm->disp_clk);
9091         disp_clk.full = dfixed_div(disp_clk, a);
9092         a.full = dfixed_const(32);
9093         b.full = dfixed_mul(a, disp_clk);
9094
9095         a.full = dfixed_const(10);
9096         disp_clk_request_efficiency.full = dfixed_const(8);
9097         disp_clk_request_efficiency.full = dfixed_div(disp_clk_request_efficiency, a);
9098
9099         bandwidth.full = dfixed_mul(b, disp_clk_request_efficiency);
9100
9101         return dfixed_trunc(bandwidth);
9102 }
9103
9104 /**
9105  * dce8_available_bandwidth - get the min available bandwidth
9106  *
9107  * @wm: watermark calculation data
9108  *
9109  * Calculate the min available bandwidth used for display (CIK).
9110  * Used for display watermark bandwidth calculations
9111  * Returns the min available bandwidth in MBytes/s
9112  */
9113 static u32 dce8_available_bandwidth(struct dce8_wm_params *wm)
9114 {
9115         /* Calculate the Available bandwidth. Display can use this temporarily but not in average. */
9116         u32 dram_bandwidth = dce8_dram_bandwidth(wm);
9117         u32 data_return_bandwidth = dce8_data_return_bandwidth(wm);
9118         u32 dmif_req_bandwidth = dce8_dmif_request_bandwidth(wm);
9119
9120         return min(dram_bandwidth, min(data_return_bandwidth, dmif_req_bandwidth));
9121 }
9122
9123 /**
9124  * dce8_average_bandwidth - get the average available bandwidth
9125  *
9126  * @wm: watermark calculation data
9127  *
9128  * Calculate the average available bandwidth used for display (CIK).
9129  * Used for display watermark bandwidth calculations
9130  * Returns the average available bandwidth in MBytes/s
9131  */
9132 static u32 dce8_average_bandwidth(struct dce8_wm_params *wm)
9133 {
9134         /* Calculate the display mode Average Bandwidth
9135          * DisplayMode should contain the source and destination dimensions,
9136          * timing, etc.
9137          */
9138         fixed20_12 bpp;
9139         fixed20_12 line_time;
9140         fixed20_12 src_width;
9141         fixed20_12 bandwidth;
9142         fixed20_12 a;
9143
9144         a.full = dfixed_const(1000);
9145         line_time.full = dfixed_const(wm->active_time + wm->blank_time);
9146         line_time.full = dfixed_div(line_time, a);
9147         bpp.full = dfixed_const(wm->bytes_per_pixel);
9148         src_width.full = dfixed_const(wm->src_width);
9149         bandwidth.full = dfixed_mul(src_width, bpp);
9150         bandwidth.full = dfixed_mul(bandwidth, wm->vsc);
9151         bandwidth.full = dfixed_div(bandwidth, line_time);
9152
9153         return dfixed_trunc(bandwidth);
9154 }
9155
9156 /**
9157  * dce8_latency_watermark - get the latency watermark
9158  *
9159  * @wm: watermark calculation data
9160  *
9161  * Calculate the latency watermark (CIK).
9162  * Used for display watermark bandwidth calculations
9163  * Returns the latency watermark in ns
9164  */
9165 static u32 dce8_latency_watermark(struct dce8_wm_params *wm)
9166 {
9167         /* First calculate the latency in ns */
9168         u32 mc_latency = 2000; /* 2000 ns. */
9169         u32 available_bandwidth = dce8_available_bandwidth(wm);
9170         u32 worst_chunk_return_time = (512 * 8 * 1000) / available_bandwidth;
9171         u32 cursor_line_pair_return_time = (128 * 4 * 1000) / available_bandwidth;
9172         u32 dc_latency = 40000000 / wm->disp_clk; /* dc pipe latency */
9173         u32 other_heads_data_return_time = ((wm->num_heads + 1) * worst_chunk_return_time) +
9174                 (wm->num_heads * cursor_line_pair_return_time);
9175         u32 latency = mc_latency + other_heads_data_return_time + dc_latency;
9176         u32 max_src_lines_per_dst_line, lb_fill_bw, line_fill_time;
9177         u32 tmp, dmif_size = 12288;
9178         fixed20_12 a, b, c;
9179
9180         if (wm->num_heads == 0)
9181                 return 0;
9182
9183         a.full = dfixed_const(2);
9184         b.full = dfixed_const(1);
9185         if ((wm->vsc.full > a.full) ||
9186             ((wm->vsc.full > b.full) && (wm->vtaps >= 3)) ||
9187             (wm->vtaps >= 5) ||
9188             ((wm->vsc.full >= a.full) && wm->interlaced))
9189                 max_src_lines_per_dst_line = 4;
9190         else
9191                 max_src_lines_per_dst_line = 2;
9192
9193         a.full = dfixed_const(available_bandwidth);
9194         b.full = dfixed_const(wm->num_heads);
9195         a.full = dfixed_div(a, b);
9196
9197         b.full = dfixed_const(mc_latency + 512);
9198         c.full = dfixed_const(wm->disp_clk);
9199         b.full = dfixed_div(b, c);
9200
9201         c.full = dfixed_const(dmif_size);
9202         b.full = dfixed_div(c, b);
9203
9204         tmp = min(dfixed_trunc(a), dfixed_trunc(b));
9205
9206         b.full = dfixed_const(1000);
9207         c.full = dfixed_const(wm->disp_clk);
9208         b.full = dfixed_div(c, b);
9209         c.full = dfixed_const(wm->bytes_per_pixel);
9210         b.full = dfixed_mul(b, c);
9211
9212         lb_fill_bw = min(tmp, dfixed_trunc(b));
9213
9214         a.full = dfixed_const(max_src_lines_per_dst_line * wm->src_width * wm->bytes_per_pixel);
9215         b.full = dfixed_const(1000);
9216         c.full = dfixed_const(lb_fill_bw);
9217         b.full = dfixed_div(c, b);
9218         a.full = dfixed_div(a, b);
9219         line_fill_time = dfixed_trunc(a);
9220
9221         if (line_fill_time < wm->active_time)
9222                 return latency;
9223         else
9224                 return latency + (line_fill_time - wm->active_time);
9225
9226 }
9227
9228 /**
9229  * dce8_average_bandwidth_vs_dram_bandwidth_for_display - check
9230  * average and available dram bandwidth
9231  *
9232  * @wm: watermark calculation data
9233  *
9234  * Check if the display average bandwidth fits in the display
9235  * dram bandwidth (CIK).
9236  * Used for display watermark bandwidth calculations
9237  * Returns true if the display fits, false if not.
9238  */
9239 static bool dce8_average_bandwidth_vs_dram_bandwidth_for_display(struct dce8_wm_params *wm)
9240 {
9241         if (dce8_average_bandwidth(wm) <=
9242             (dce8_dram_bandwidth_for_display(wm) / wm->num_heads))
9243                 return true;
9244         else
9245                 return false;
9246 }
9247
9248 /**
9249  * dce8_average_bandwidth_vs_available_bandwidth - check
9250  * average and available bandwidth
9251  *
9252  * @wm: watermark calculation data
9253  *
9254  * Check if the display average bandwidth fits in the display
9255  * available bandwidth (CIK).
9256  * Used for display watermark bandwidth calculations
9257  * Returns true if the display fits, false if not.
9258  */
9259 static bool dce8_average_bandwidth_vs_available_bandwidth(struct dce8_wm_params *wm)
9260 {
9261         if (dce8_average_bandwidth(wm) <=
9262             (dce8_available_bandwidth(wm) / wm->num_heads))
9263                 return true;
9264         else
9265                 return false;
9266 }
9267
9268 /**
9269  * dce8_check_latency_hiding - check latency hiding
9270  *
9271  * @wm: watermark calculation data
9272  *
9273  * Check latency hiding (CIK).
9274  * Used for display watermark bandwidth calculations
9275  * Returns true if the display fits, false if not.
9276  */
9277 static bool dce8_check_latency_hiding(struct dce8_wm_params *wm)
9278 {
9279         u32 lb_partitions = wm->lb_size / wm->src_width;
9280         u32 line_time = wm->active_time + wm->blank_time;
9281         u32 latency_tolerant_lines;
9282         u32 latency_hiding;
9283         fixed20_12 a;
9284
9285         a.full = dfixed_const(1);
9286         if (wm->vsc.full > a.full)
9287                 latency_tolerant_lines = 1;
9288         else {
9289                 if (lb_partitions <= (wm->vtaps + 1))
9290                         latency_tolerant_lines = 1;
9291                 else
9292                         latency_tolerant_lines = 2;
9293         }
9294
9295         latency_hiding = (latency_tolerant_lines * line_time + wm->blank_time);
9296
9297         if (dce8_latency_watermark(wm) <= latency_hiding)
9298                 return true;
9299         else
9300                 return false;
9301 }
9302
9303 /**
9304  * dce8_program_watermarks - program display watermarks
9305  *
9306  * @rdev: radeon_device pointer
9307  * @radeon_crtc: the selected display controller
9308  * @lb_size: line buffer size
9309  * @num_heads: number of display controllers in use
9310  *
9311  * Calculate and program the display watermarks for the
9312  * selected display controller (CIK).
9313  */
9314 static void dce8_program_watermarks(struct radeon_device *rdev,
9315                                     struct radeon_crtc *radeon_crtc,
9316                                     u32 lb_size, u32 num_heads)
9317 {
9318         struct drm_display_mode *mode = &radeon_crtc->base.mode;
9319         struct dce8_wm_params wm_low, wm_high;
9320         u32 pixel_period;
9321         u32 line_time = 0;
9322         u32 latency_watermark_a = 0, latency_watermark_b = 0;
9323         u32 tmp, wm_mask;
9324
9325         if (radeon_crtc->base.enabled && num_heads && mode) {
9326                 pixel_period = 1000000 / (u32)mode->clock;
9327                 line_time = min((u32)mode->crtc_htotal * pixel_period, (u32)65535);
9328
9329                 /* watermark for high clocks */
9330                 if ((rdev->pm.pm_method == PM_METHOD_DPM) &&
9331                     rdev->pm.dpm_enabled) {
9332                         wm_high.yclk =
9333                                 radeon_dpm_get_mclk(rdev, false) * 10;
9334                         wm_high.sclk =
9335                                 radeon_dpm_get_sclk(rdev, false) * 10;
9336                 } else {
9337                         wm_high.yclk = rdev->pm.current_mclk * 10;
9338                         wm_high.sclk = rdev->pm.current_sclk * 10;
9339                 }
9340
9341                 wm_high.disp_clk = mode->clock;
9342                 wm_high.src_width = mode->crtc_hdisplay;
9343                 wm_high.active_time = mode->crtc_hdisplay * pixel_period;
9344                 wm_high.blank_time = line_time - wm_high.active_time;
9345                 wm_high.interlaced = false;
9346                 if (mode->flags & DRM_MODE_FLAG_INTERLACE)
9347                         wm_high.interlaced = true;
9348                 wm_high.vsc = radeon_crtc->vsc;
9349                 wm_high.vtaps = 1;
9350                 if (radeon_crtc->rmx_type != RMX_OFF)
9351                         wm_high.vtaps = 2;
9352                 wm_high.bytes_per_pixel = 4; /* XXX: get this from fb config */
9353                 wm_high.lb_size = lb_size;
9354                 wm_high.dram_channels = cik_get_number_of_dram_channels(rdev);
9355                 wm_high.num_heads = num_heads;
9356
9357                 /* set for high clocks */
9358                 latency_watermark_a = min(dce8_latency_watermark(&wm_high), (u32)65535);
9359
9360                 /* possibly force display priority to high */
9361                 /* should really do this at mode validation time... */
9362                 if (!dce8_average_bandwidth_vs_dram_bandwidth_for_display(&wm_high) ||
9363                     !dce8_average_bandwidth_vs_available_bandwidth(&wm_high) ||
9364                     !dce8_check_latency_hiding(&wm_high) ||
9365                     (rdev->disp_priority == 2)) {
9366                         DRM_DEBUG_KMS("force priority to high\n");
9367                 }
9368
9369                 /* watermark for low clocks */
9370                 if ((rdev->pm.pm_method == PM_METHOD_DPM) &&
9371                     rdev->pm.dpm_enabled) {
9372                         wm_low.yclk =
9373                                 radeon_dpm_get_mclk(rdev, true) * 10;
9374                         wm_low.sclk =
9375                                 radeon_dpm_get_sclk(rdev, true) * 10;
9376                 } else {
9377                         wm_low.yclk = rdev->pm.current_mclk * 10;
9378                         wm_low.sclk = rdev->pm.current_sclk * 10;
9379                 }
9380
9381                 wm_low.disp_clk = mode->clock;
9382                 wm_low.src_width = mode->crtc_hdisplay;
9383                 wm_low.active_time = mode->crtc_hdisplay * pixel_period;
9384                 wm_low.blank_time = line_time - wm_low.active_time;
9385                 wm_low.interlaced = false;
9386                 if (mode->flags & DRM_MODE_FLAG_INTERLACE)
9387                         wm_low.interlaced = true;
9388                 wm_low.vsc = radeon_crtc->vsc;
9389                 wm_low.vtaps = 1;
9390                 if (radeon_crtc->rmx_type != RMX_OFF)
9391                         wm_low.vtaps = 2;
9392                 wm_low.bytes_per_pixel = 4; /* XXX: get this from fb config */
9393                 wm_low.lb_size = lb_size;
9394                 wm_low.dram_channels = cik_get_number_of_dram_channels(rdev);
9395                 wm_low.num_heads = num_heads;
9396
9397                 /* set for low clocks */
9398                 latency_watermark_b = min(dce8_latency_watermark(&wm_low), (u32)65535);
9399
9400                 /* possibly force display priority to high */
9401                 /* should really do this at mode validation time... */
9402                 if (!dce8_average_bandwidth_vs_dram_bandwidth_for_display(&wm_low) ||
9403                     !dce8_average_bandwidth_vs_available_bandwidth(&wm_low) ||
9404                     !dce8_check_latency_hiding(&wm_low) ||
9405                     (rdev->disp_priority == 2)) {
9406                         DRM_DEBUG_KMS("force priority to high\n");
9407                 }
9408         }
9409
9410         /* select wm A */
9411         wm_mask = RREG32(DPG_WATERMARK_MASK_CONTROL + radeon_crtc->crtc_offset);
9412         tmp = wm_mask;
9413         tmp &= ~LATENCY_WATERMARK_MASK(3);
9414         tmp |= LATENCY_WATERMARK_MASK(1);
9415         WREG32(DPG_WATERMARK_MASK_CONTROL + radeon_crtc->crtc_offset, tmp);
9416         WREG32(DPG_PIPE_LATENCY_CONTROL + radeon_crtc->crtc_offset,
9417                (LATENCY_LOW_WATERMARK(latency_watermark_a) |
9418                 LATENCY_HIGH_WATERMARK(line_time)));
9419         /* select wm B */
9420         tmp = RREG32(DPG_WATERMARK_MASK_CONTROL + radeon_crtc->crtc_offset);
9421         tmp &= ~LATENCY_WATERMARK_MASK(3);
9422         tmp |= LATENCY_WATERMARK_MASK(2);
9423         WREG32(DPG_WATERMARK_MASK_CONTROL + radeon_crtc->crtc_offset, tmp);
9424         WREG32(DPG_PIPE_LATENCY_CONTROL + radeon_crtc->crtc_offset,
9425                (LATENCY_LOW_WATERMARK(latency_watermark_b) |
9426                 LATENCY_HIGH_WATERMARK(line_time)));
9427         /* restore original selection */
9428         WREG32(DPG_WATERMARK_MASK_CONTROL + radeon_crtc->crtc_offset, wm_mask);
9429
9430         /* save values for DPM */
9431         radeon_crtc->line_time = line_time;
9432         radeon_crtc->wm_high = latency_watermark_a;
9433         radeon_crtc->wm_low = latency_watermark_b;
9434 }
9435
9436 /**
9437  * dce8_bandwidth_update - program display watermarks
9438  *
9439  * @rdev: radeon_device pointer
9440  *
9441  * Calculate and program the display watermarks and line
9442  * buffer allocation (CIK).
9443  */
9444 void dce8_bandwidth_update(struct radeon_device *rdev)
9445 {
9446         struct drm_display_mode *mode = NULL;
9447         u32 num_heads = 0, lb_size;
9448         int i;
9449
9450         radeon_update_display_priority(rdev);
9451
9452         for (i = 0; i < rdev->num_crtc; i++) {
9453                 if (rdev->mode_info.crtcs[i]->base.enabled)
9454                         num_heads++;
9455         }
9456         for (i = 0; i < rdev->num_crtc; i++) {
9457                 mode = &rdev->mode_info.crtcs[i]->base.mode;
9458                 lb_size = dce8_line_buffer_adjust(rdev, rdev->mode_info.crtcs[i], mode);
9459                 dce8_program_watermarks(rdev, rdev->mode_info.crtcs[i], lb_size, num_heads);
9460         }
9461 }
9462
9463 /**
9464  * cik_get_gpu_clock_counter - return GPU clock counter snapshot
9465  *
9466  * @rdev: radeon_device pointer
9467  *
9468  * Fetches a GPU clock counter snapshot (SI).
9469  * Returns the 64 bit clock counter snapshot.
9470  */
9471 uint64_t cik_get_gpu_clock_counter(struct radeon_device *rdev)
9472 {
9473         uint64_t clock;
9474
9475         mutex_lock(&rdev->gpu_clock_mutex);
9476         WREG32(RLC_CAPTURE_GPU_CLOCK_COUNT, 1);
9477         clock = (uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_LSB) |
9478                 ((uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_MSB) << 32ULL);
9479         mutex_unlock(&rdev->gpu_clock_mutex);
9480         return clock;
9481 }
9482
9483 static int cik_set_uvd_clock(struct radeon_device *rdev, u32 clock,
9484                               u32 cntl_reg, u32 status_reg)
9485 {
9486         int r, i;
9487         struct atom_clock_dividers dividers;
9488         uint32_t tmp;
9489
9490         r = radeon_atom_get_clock_dividers(rdev, COMPUTE_GPUCLK_INPUT_FLAG_DEFAULT_GPUCLK,
9491                                            clock, false, &dividers);
9492         if (r)
9493                 return r;
9494
9495         tmp = RREG32_SMC(cntl_reg);
9496         tmp &= ~(DCLK_DIR_CNTL_EN|DCLK_DIVIDER_MASK);
9497         tmp |= dividers.post_divider;
9498         WREG32_SMC(cntl_reg, tmp);
9499
9500         for (i = 0; i < 100; i++) {
9501                 if (RREG32_SMC(status_reg) & DCLK_STATUS)
9502                         break;
9503                 mdelay(10);
9504         }
9505         if (i == 100)
9506                 return -ETIMEDOUT;
9507
9508         return 0;
9509 }
9510
9511 int cik_set_uvd_clocks(struct radeon_device *rdev, u32 vclk, u32 dclk)
9512 {
9513         int r = 0;
9514
9515         r = cik_set_uvd_clock(rdev, vclk, CG_VCLK_CNTL, CG_VCLK_STATUS);
9516         if (r)
9517                 return r;
9518
9519         r = cik_set_uvd_clock(rdev, dclk, CG_DCLK_CNTL, CG_DCLK_STATUS);
9520         return r;
9521 }
9522
9523 int cik_set_vce_clocks(struct radeon_device *rdev, u32 evclk, u32 ecclk)
9524 {
9525         int r, i;
9526         struct atom_clock_dividers dividers;
9527         u32 tmp;
9528
9529         r = radeon_atom_get_clock_dividers(rdev, COMPUTE_GPUCLK_INPUT_FLAG_DEFAULT_GPUCLK,
9530                                            ecclk, false, &dividers);
9531         if (r)
9532                 return r;
9533
9534         for (i = 0; i < 100; i++) {
9535                 if (RREG32_SMC(CG_ECLK_STATUS) & ECLK_STATUS)
9536                         break;
9537                 mdelay(10);
9538         }
9539         if (i == 100)
9540                 return -ETIMEDOUT;
9541
9542         tmp = RREG32_SMC(CG_ECLK_CNTL);
9543         tmp &= ~(ECLK_DIR_CNTL_EN|ECLK_DIVIDER_MASK);
9544         tmp |= dividers.post_divider;
9545         WREG32_SMC(CG_ECLK_CNTL, tmp);
9546
9547         for (i = 0; i < 100; i++) {
9548                 if (RREG32_SMC(CG_ECLK_STATUS) & ECLK_STATUS)
9549                         break;
9550                 mdelay(10);
9551         }
9552         if (i == 100)
9553                 return -ETIMEDOUT;
9554
9555         return 0;
9556 }
9557
9558 static void cik_pcie_gen3_enable(struct radeon_device *rdev)
9559 {
9560         struct pci_dev *root = rdev->pdev->bus->self;
9561         int bridge_pos, gpu_pos;
9562         u32 speed_cntl, mask, current_data_rate;
9563         int ret, i;
9564         u16 tmp16;
9565
9566         if (pci_is_root_bus(rdev->pdev->bus))
9567                 return;
9568
9569         if (radeon_pcie_gen2 == 0)
9570                 return;
9571
9572         if (rdev->flags & RADEON_IS_IGP)
9573                 return;
9574
9575         if (!(rdev->flags & RADEON_IS_PCIE))
9576                 return;
9577
9578         ret = drm_pcie_get_speed_cap_mask(rdev->ddev, &mask);
9579         if (ret != 0)
9580                 return;
9581
9582         if (!(mask & (DRM_PCIE_SPEED_50 | DRM_PCIE_SPEED_80)))
9583                 return;
9584
9585         speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
9586         current_data_rate = (speed_cntl & LC_CURRENT_DATA_RATE_MASK) >>
9587                 LC_CURRENT_DATA_RATE_SHIFT;
9588         if (mask & DRM_PCIE_SPEED_80) {
9589                 if (current_data_rate == 2) {
9590                         DRM_INFO("PCIE gen 3 link speeds already enabled\n");
9591                         return;
9592                 }
9593                 DRM_INFO("enabling PCIE gen 3 link speeds, disable with radeon.pcie_gen2=0\n");
9594         } else if (mask & DRM_PCIE_SPEED_50) {
9595                 if (current_data_rate == 1) {
9596                         DRM_INFO("PCIE gen 2 link speeds already enabled\n");
9597                         return;
9598                 }
9599                 DRM_INFO("enabling PCIE gen 2 link speeds, disable with radeon.pcie_gen2=0\n");
9600         }
9601
9602         bridge_pos = pci_pcie_cap(root);
9603         if (!bridge_pos)
9604                 return;
9605
9606         gpu_pos = pci_pcie_cap(rdev->pdev);
9607         if (!gpu_pos)
9608                 return;
9609
9610         if (mask & DRM_PCIE_SPEED_80) {
9611                 /* re-try equalization if gen3 is not already enabled */
9612                 if (current_data_rate != 2) {
9613                         u16 bridge_cfg, gpu_cfg;
9614                         u16 bridge_cfg2, gpu_cfg2;
9615                         u32 max_lw, current_lw, tmp;
9616
9617                         pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &bridge_cfg);
9618                         pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &gpu_cfg);
9619
9620                         tmp16 = bridge_cfg | PCI_EXP_LNKCTL_HAWD;
9621                         pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL, tmp16);
9622
9623                         tmp16 = gpu_cfg | PCI_EXP_LNKCTL_HAWD;
9624                         pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, tmp16);
9625
9626                         tmp = RREG32_PCIE_PORT(PCIE_LC_STATUS1);
9627                         max_lw = (tmp & LC_DETECTED_LINK_WIDTH_MASK) >> LC_DETECTED_LINK_WIDTH_SHIFT;
9628                         current_lw = (tmp & LC_OPERATING_LINK_WIDTH_MASK) >> LC_OPERATING_LINK_WIDTH_SHIFT;
9629
9630                         if (current_lw < max_lw) {
9631                                 tmp = RREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL);
9632                                 if (tmp & LC_RENEGOTIATION_SUPPORT) {
9633                                         tmp &= ~(LC_LINK_WIDTH_MASK | LC_UPCONFIGURE_DIS);
9634                                         tmp |= (max_lw << LC_LINK_WIDTH_SHIFT);
9635                                         tmp |= LC_UPCONFIGURE_SUPPORT | LC_RENEGOTIATE_EN | LC_RECONFIG_NOW;
9636                                         WREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL, tmp);
9637                                 }
9638                         }
9639
9640                         for (i = 0; i < 10; i++) {
9641                                 /* check status */
9642                                 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_DEVSTA, &tmp16);
9643                                 if (tmp16 & PCI_EXP_DEVSTA_TRPND)
9644                                         break;
9645
9646                                 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &bridge_cfg);
9647                                 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &gpu_cfg);
9648
9649                                 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, &bridge_cfg2);
9650                                 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &gpu_cfg2);
9651
9652                                 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
9653                                 tmp |= LC_SET_QUIESCE;
9654                                 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
9655
9656                                 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
9657                                 tmp |= LC_REDO_EQ;
9658                                 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
9659
9660                                 mdelay(100);
9661
9662                                 /* linkctl */
9663                                 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL, &tmp16);
9664                                 tmp16 &= ~PCI_EXP_LNKCTL_HAWD;
9665                                 tmp16 |= (bridge_cfg & PCI_EXP_LNKCTL_HAWD);
9666                                 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL, tmp16);
9667
9668                                 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, &tmp16);
9669                                 tmp16 &= ~PCI_EXP_LNKCTL_HAWD;
9670                                 tmp16 |= (gpu_cfg & PCI_EXP_LNKCTL_HAWD);
9671                                 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL, tmp16);
9672
9673                                 /* linkctl2 */
9674                                 pci_read_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, &tmp16);
9675                                 tmp16 &= ~((1 << 4) | (7 << 9));
9676                                 tmp16 |= (bridge_cfg2 & ((1 << 4) | (7 << 9)));
9677                                 pci_write_config_word(root, bridge_pos + PCI_EXP_LNKCTL2, tmp16);
9678
9679                                 pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &tmp16);
9680                                 tmp16 &= ~((1 << 4) | (7 << 9));
9681                                 tmp16 |= (gpu_cfg2 & ((1 << 4) | (7 << 9)));
9682                                 pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, tmp16);
9683
9684                                 tmp = RREG32_PCIE_PORT(PCIE_LC_CNTL4);
9685                                 tmp &= ~LC_SET_QUIESCE;
9686                                 WREG32_PCIE_PORT(PCIE_LC_CNTL4, tmp);
9687                         }
9688                 }
9689         }
9690
9691         /* set the link speed */
9692         speed_cntl |= LC_FORCE_EN_SW_SPEED_CHANGE | LC_FORCE_DIS_HW_SPEED_CHANGE;
9693         speed_cntl &= ~LC_FORCE_DIS_SW_SPEED_CHANGE;
9694         WREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL, speed_cntl);
9695
9696         pci_read_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, &tmp16);
9697         tmp16 &= ~0xf;
9698         if (mask & DRM_PCIE_SPEED_80)
9699                 tmp16 |= 3; /* gen3 */
9700         else if (mask & DRM_PCIE_SPEED_50)
9701                 tmp16 |= 2; /* gen2 */
9702         else
9703                 tmp16 |= 1; /* gen1 */
9704         pci_write_config_word(rdev->pdev, gpu_pos + PCI_EXP_LNKCTL2, tmp16);
9705
9706         speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
9707         speed_cntl |= LC_INITIATE_LINK_SPEED_CHANGE;
9708         WREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL, speed_cntl);
9709
9710         for (i = 0; i < rdev->usec_timeout; i++) {
9711                 speed_cntl = RREG32_PCIE_PORT(PCIE_LC_SPEED_CNTL);
9712                 if ((speed_cntl & LC_INITIATE_LINK_SPEED_CHANGE) == 0)
9713                         break;
9714                 udelay(1);
9715         }
9716 }
9717
9718 static void cik_program_aspm(struct radeon_device *rdev)
9719 {
9720         u32 data, orig;
9721         bool disable_l0s = false, disable_l1 = false, disable_plloff_in_l1 = false;
9722         bool disable_clkreq = false;
9723
9724         if (radeon_aspm == 0)
9725                 return;
9726
9727         /* XXX double check IGPs */
9728         if (rdev->flags & RADEON_IS_IGP)
9729                 return;
9730
9731         if (!(rdev->flags & RADEON_IS_PCIE))
9732                 return;
9733
9734         orig = data = RREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL);
9735         data &= ~LC_XMIT_N_FTS_MASK;
9736         data |= LC_XMIT_N_FTS(0x24) | LC_XMIT_N_FTS_OVERRIDE_EN;
9737         if (orig != data)
9738                 WREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL, data);
9739
9740         orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL3);
9741         data |= LC_GO_TO_RECOVERY;
9742         if (orig != data)
9743                 WREG32_PCIE_PORT(PCIE_LC_CNTL3, data);
9744
9745         orig = data = RREG32_PCIE_PORT(PCIE_P_CNTL);
9746         data |= P_IGNORE_EDB_ERR;
9747         if (orig != data)
9748                 WREG32_PCIE_PORT(PCIE_P_CNTL, data);
9749
9750         orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL);
9751         data &= ~(LC_L0S_INACTIVITY_MASK | LC_L1_INACTIVITY_MASK);
9752         data |= LC_PMI_TO_L1_DIS;
9753         if (!disable_l0s)
9754                 data |= LC_L0S_INACTIVITY(7);
9755
9756         if (!disable_l1) {
9757                 data |= LC_L1_INACTIVITY(7);
9758                 data &= ~LC_PMI_TO_L1_DIS;
9759                 if (orig != data)
9760                         WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
9761
9762                 if (!disable_plloff_in_l1) {
9763                         bool clk_req_support;
9764
9765                         orig = data = RREG32_PCIE_PORT(PB0_PIF_PWRDOWN_0);
9766                         data &= ~(PLL_POWER_STATE_IN_OFF_0_MASK | PLL_POWER_STATE_IN_TXS2_0_MASK);
9767                         data |= PLL_POWER_STATE_IN_OFF_0(7) | PLL_POWER_STATE_IN_TXS2_0(7);
9768                         if (orig != data)
9769                                 WREG32_PCIE_PORT(PB0_PIF_PWRDOWN_0, data);
9770
9771                         orig = data = RREG32_PCIE_PORT(PB0_PIF_PWRDOWN_1);
9772                         data &= ~(PLL_POWER_STATE_IN_OFF_1_MASK | PLL_POWER_STATE_IN_TXS2_1_MASK);
9773                         data |= PLL_POWER_STATE_IN_OFF_1(7) | PLL_POWER_STATE_IN_TXS2_1(7);
9774                         if (orig != data)
9775                                 WREG32_PCIE_PORT(PB0_PIF_PWRDOWN_1, data);
9776
9777                         orig = data = RREG32_PCIE_PORT(PB1_PIF_PWRDOWN_0);
9778                         data &= ~(PLL_POWER_STATE_IN_OFF_0_MASK | PLL_POWER_STATE_IN_TXS2_0_MASK);
9779                         data |= PLL_POWER_STATE_IN_OFF_0(7) | PLL_POWER_STATE_IN_TXS2_0(7);
9780                         if (orig != data)
9781                                 WREG32_PCIE_PORT(PB1_PIF_PWRDOWN_0, data);
9782
9783                         orig = data = RREG32_PCIE_PORT(PB1_PIF_PWRDOWN_1);
9784                         data &= ~(PLL_POWER_STATE_IN_OFF_1_MASK | PLL_POWER_STATE_IN_TXS2_1_MASK);
9785                         data |= PLL_POWER_STATE_IN_OFF_1(7) | PLL_POWER_STATE_IN_TXS2_1(7);
9786                         if (orig != data)
9787                                 WREG32_PCIE_PORT(PB1_PIF_PWRDOWN_1, data);
9788
9789                         orig = data = RREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL);
9790                         data &= ~LC_DYN_LANES_PWR_STATE_MASK;
9791                         data |= LC_DYN_LANES_PWR_STATE(3);
9792                         if (orig != data)
9793                                 WREG32_PCIE_PORT(PCIE_LC_LINK_WIDTH_CNTL, data);
9794
9795                         if (!disable_clkreq &&
9796                             !pci_is_root_bus(rdev->pdev->bus)) {
9797                                 struct pci_dev *root = rdev->pdev->bus->self;
9798                                 u32 lnkcap;
9799
9800                                 clk_req_support = false;
9801                                 pcie_capability_read_dword(root, PCI_EXP_LNKCAP, &lnkcap);
9802                                 if (lnkcap & PCI_EXP_LNKCAP_CLKPM)
9803                                         clk_req_support = true;
9804                         } else {
9805                                 clk_req_support = false;
9806                         }
9807
9808                         if (clk_req_support) {
9809                                 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL2);
9810                                 data |= LC_ALLOW_PDWN_IN_L1 | LC_ALLOW_PDWN_IN_L23;
9811                                 if (orig != data)
9812                                         WREG32_PCIE_PORT(PCIE_LC_CNTL2, data);
9813
9814                                 orig = data = RREG32_SMC(THM_CLK_CNTL);
9815                                 data &= ~(CMON_CLK_SEL_MASK | TMON_CLK_SEL_MASK);
9816                                 data |= CMON_CLK_SEL(1) | TMON_CLK_SEL(1);
9817                                 if (orig != data)
9818                                         WREG32_SMC(THM_CLK_CNTL, data);
9819
9820                                 orig = data = RREG32_SMC(MISC_CLK_CTRL);
9821                                 data &= ~(DEEP_SLEEP_CLK_SEL_MASK | ZCLK_SEL_MASK);
9822                                 data |= DEEP_SLEEP_CLK_SEL(1) | ZCLK_SEL(1);
9823                                 if (orig != data)
9824                                         WREG32_SMC(MISC_CLK_CTRL, data);
9825
9826                                 orig = data = RREG32_SMC(CG_CLKPIN_CNTL);
9827                                 data &= ~BCLK_AS_XCLK;
9828                                 if (orig != data)
9829                                         WREG32_SMC(CG_CLKPIN_CNTL, data);
9830
9831                                 orig = data = RREG32_SMC(CG_CLKPIN_CNTL_2);
9832                                 data &= ~FORCE_BIF_REFCLK_EN;
9833                                 if (orig != data)
9834                                         WREG32_SMC(CG_CLKPIN_CNTL_2, data);
9835
9836                                 orig = data = RREG32_SMC(MPLL_BYPASSCLK_SEL);
9837                                 data &= ~MPLL_CLKOUT_SEL_MASK;
9838                                 data |= MPLL_CLKOUT_SEL(4);
9839                                 if (orig != data)
9840                                         WREG32_SMC(MPLL_BYPASSCLK_SEL, data);
9841                         }
9842                 }
9843         } else {
9844                 if (orig != data)
9845                         WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
9846         }
9847
9848         orig = data = RREG32_PCIE_PORT(PCIE_CNTL2);
9849         data |= SLV_MEM_LS_EN | MST_MEM_LS_EN | REPLAY_MEM_LS_EN;
9850         if (orig != data)
9851                 WREG32_PCIE_PORT(PCIE_CNTL2, data);
9852
9853         if (!disable_l0s) {
9854                 data = RREG32_PCIE_PORT(PCIE_LC_N_FTS_CNTL);
9855                 if((data & LC_N_FTS_MASK) == LC_N_FTS_MASK) {
9856                         data = RREG32_PCIE_PORT(PCIE_LC_STATUS1);
9857                         if ((data & LC_REVERSE_XMIT) && (data & LC_REVERSE_RCVR)) {
9858                                 orig = data = RREG32_PCIE_PORT(PCIE_LC_CNTL);
9859                                 data &= ~LC_L0S_INACTIVITY_MASK;
9860                                 if (orig != data)
9861                                         WREG32_PCIE_PORT(PCIE_LC_CNTL, data);
9862                         }
9863                 }
9864         }
9865 }