drivers/gpu/drm/radeon/radeon_kfd.c

   1 /*
   2  * Copyright 2014 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
  23 #include <linux/module.h>
  24 #include <linux/fdtable.h>
  25 #include <linux/uaccess.h>
  26 #include <drm/drmP.h>
  27 #include "radeon.h"
  28 #include "cikd.h"
  29 #include "cik_reg.h"
  30 #include "radeon_kfd.h"
  31 #include "radeon_ucode.h"
  32 #include <linux/firmware.h>
  33 #include "cik_structs.h"
  34
  35 #define CIK_PIPE_PER_MEC        (4)
  36
  37 static const uint32_t watchRegs[MAX_WATCH_ADDRESSES * ADDRESS_WATCH_REG_MAX] = {
  38         TCP_WATCH0_ADDR_H, TCP_WATCH0_ADDR_L, TCP_WATCH0_CNTL,
  39         TCP_WATCH1_ADDR_H, TCP_WATCH1_ADDR_L, TCP_WATCH1_CNTL,
  40         TCP_WATCH2_ADDR_H, TCP_WATCH2_ADDR_L, TCP_WATCH2_CNTL,
  41         TCP_WATCH3_ADDR_H, TCP_WATCH3_ADDR_L, TCP_WATCH3_CNTL
  42 };
  43
  44 struct kgd_mem {
  45         struct radeon_bo *bo;
  46         uint64_t gpu_addr;
  47         void *cpu_ptr;
  48 };
  49
  50
  51 static int alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
  52                         void **mem_obj, uint64_t *gpu_addr,
  53                         void **cpu_ptr);
  54
  55 static void free_gtt_mem(struct kgd_dev *kgd, void *mem_obj);
  56
  57 static uint64_t get_vmem_size(struct kgd_dev *kgd);
  58 static uint64_t get_gpu_clock_counter(struct kgd_dev *kgd);
  59
  60 static uint32_t get_max_engine_clock_in_mhz(struct kgd_dev *kgd);
  61 static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type);
  62
  63 /*
  64  * Register access functions
  65  */
  66
  67 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
  68                 uint32_t sh_mem_config, uint32_t sh_mem_ape1_base,
  69                 uint32_t sh_mem_ape1_limit, uint32_t sh_mem_bases);
  70
  71 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
  72                                         unsigned int vmid);
  73
  74 static int kgd_init_pipeline(struct kgd_dev *kgd, uint32_t pipe_id,
  75                                 uint32_t hpd_size, uint64_t hpd_gpu_addr);
  76 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
  77 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
  78                         uint32_t queue_id, uint32_t __user *wptr);
  79 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd);
  80 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
  81                                 uint32_t pipe_id, uint32_t queue_id);
  82
  83 static int kgd_hqd_destroy(struct kgd_dev *kgd, uint32_t reset_type,
  84                                 unsigned int timeout, uint32_t pipe_id,
  85                                 uint32_t queue_id);
  86 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
  87 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
  88                                 unsigned int timeout);
  89 static int kgd_address_watch_disable(struct kgd_dev *kgd);
  90 static int kgd_address_watch_execute(struct kgd_dev *kgd,
  91                                         unsigned int watch_point_id,
  92                                         uint32_t cntl_val,
  93                                         uint32_t addr_hi,
  94                                         uint32_t addr_lo);
  95 static int kgd_wave_control_execute(struct kgd_dev *kgd,
  96                                         uint32_t gfx_index_val,
  97                                         uint32_t sq_cmd);
  98 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
  99                                         unsigned int watch_point_id,
 100                                         unsigned int reg_offset);
 101
 102 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd, uint8_t vmid);
 103 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
 104                                                         uint8_t vmid);
 105 static void write_vmid_invalidate_request(struct kgd_dev *kgd, uint8_t vmid);
 106
 107 static const struct kfd2kgd_calls kfd2kgd = {
 108         .init_gtt_mem_allocation = alloc_gtt_mem,
 109         .free_gtt_mem = free_gtt_mem,
 110         .get_vmem_size = get_vmem_size,
 111         .get_gpu_clock_counter = get_gpu_clock_counter,
 112         .get_max_engine_clock_in_mhz = get_max_engine_clock_in_mhz,
 113         .program_sh_mem_settings = kgd_program_sh_mem_settings,
 114         .set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
 115         .init_pipeline = kgd_init_pipeline,
 116         .init_interrupts = kgd_init_interrupts,
 117         .hqd_load = kgd_hqd_load,
 118         .hqd_sdma_load = kgd_hqd_sdma_load,
 119         .hqd_is_occupied = kgd_hqd_is_occupied,
 120         .hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
 121         .hqd_destroy = kgd_hqd_destroy,
 122         .hqd_sdma_destroy = kgd_hqd_sdma_destroy,
 123         .address_watch_disable = kgd_address_watch_disable,
 124         .address_watch_execute = kgd_address_watch_execute,
 125         .wave_control_execute = kgd_wave_control_execute,
 126         .address_watch_get_offset = kgd_address_watch_get_offset,
 127         .get_atc_vmid_pasid_mapping_pasid = get_atc_vmid_pasid_mapping_pasid,
 128         .get_atc_vmid_pasid_mapping_valid = get_atc_vmid_pasid_mapping_valid,
 129         .write_vmid_invalidate_request = write_vmid_invalidate_request,
 130         .get_fw_version = get_fw_version
 131 };
 132
 133 static const struct kgd2kfd_calls *kgd2kfd;
 134
 135 int radeon_kfd_init(void)
 136 {
 137         int ret;
 138
 139 #if defined(CONFIG_HSA_AMD_MODULE)
 140         int (*kgd2kfd_init_p)(unsigned, const struct kgd2kfd_calls**);
 141
 142         kgd2kfd_init_p = symbol_request(kgd2kfd_init);
 143
 144         if (kgd2kfd_init_p == NULL)
 145                 return -ENOENT;
 146
 147         ret = kgd2kfd_init_p(KFD_INTERFACE_VERSION, &kgd2kfd);
 148         if (ret) {
 149                 symbol_put(kgd2kfd_init);
 150                 kgd2kfd = NULL;
 151         }
 152
 153 #elif defined(CONFIG_HSA_AMD)
 154         ret = kgd2kfd_init(KFD_INTERFACE_VERSION, &kgd2kfd);
 155         if (ret)
 156                 kgd2kfd = NULL;
 157
 158 #else
 159         ret = -ENOENT;
 160 #endif
 161
 162         return ret;
 163 }
 164
 165 void radeon_kfd_fini(void)
 166 {
 167         if (kgd2kfd) {
 168                 kgd2kfd->exit();
 169                 symbol_put(kgd2kfd_init);
 170         }
 171 }
 172
 173 void radeon_kfd_device_probe(struct radeon_device *rdev)
 174 {
 175         if (kgd2kfd)
 176                 rdev->kfd = kgd2kfd->probe((struct kgd_dev *)rdev,
 177                         rdev->pdev, &kfd2kgd);
 178 }
 179
 180 void radeon_kfd_device_init(struct radeon_device *rdev)
 181 {
 182         int i, queue, pipe, mec;
 183
 184         if (rdev->kfd) {
 185                 struct kgd2kfd_shared_resources gpu_resources = {
 186                         .compute_vmid_bitmap = 0xFF00,
 187                         .num_mec = 1,
 188                         .num_pipe_per_mec = 4,
 189                         .num_queue_per_pipe = 8
 190                 };
 191
 192                 bitmap_zero(gpu_resources.queue_bitmap, KGD_MAX_QUEUES);
 193
 194                 for (i = 0; i < KGD_MAX_QUEUES; ++i) {
 195                         queue = i % gpu_resources.num_queue_per_pipe;
 196                         pipe = (i / gpu_resources.num_queue_per_pipe)
 197                                 % gpu_resources.num_pipe_per_mec;
 198                         mec = (i / gpu_resources.num_queue_per_pipe)
 199                                 / gpu_resources.num_pipe_per_mec;
 200
 201                         if (mec == 0 && pipe > 0)
 202                                 set_bit(i, gpu_resources.queue_bitmap);
 203                 }
 204
 205                 radeon_doorbell_get_kfd_info(rdev,
 206                                 &gpu_resources.doorbell_physical_address,
 207                                 &gpu_resources.doorbell_aperture_size,
 208                                 &gpu_resources.doorbell_start_offset);
 209
 210                 kgd2kfd->device_init(rdev->kfd, &gpu_resources);
 211         }
 212 }
 213
 214 void radeon_kfd_device_fini(struct radeon_device *rdev)
 215 {
 216         if (rdev->kfd) {
 217                 kgd2kfd->device_exit(rdev->kfd);
 218                 rdev->kfd = NULL;
 219         }
 220 }
 221
 222 void radeon_kfd_interrupt(struct radeon_device *rdev, const void *ih_ring_entry)
 223 {
 224         if (rdev->kfd)
 225                 kgd2kfd->interrupt(rdev->kfd, ih_ring_entry);
 226 }
 227
 228 void radeon_kfd_suspend(struct radeon_device *rdev)
 229 {
 230         if (rdev->kfd)
 231                 kgd2kfd->suspend(rdev->kfd);
 232 }
 233
 234 int radeon_kfd_resume(struct radeon_device *rdev)
 235 {
 236         int r = 0;
 237
 238         if (rdev->kfd)
 239                 r = kgd2kfd->resume(rdev->kfd);
 240
 241         return r;
 242 }
 243
 244 static int alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
 245                         void **mem_obj, uint64_t *gpu_addr,
 246                         void **cpu_ptr)
 247 {
 248         struct radeon_device *rdev = (struct radeon_device *)kgd;
 249         struct kgd_mem **mem = (struct kgd_mem **) mem_obj;
 250         int r;
 251
 252         BUG_ON(kgd == NULL);
 253         BUG_ON(gpu_addr == NULL);
 254         BUG_ON(cpu_ptr == NULL);
 255
 256         *mem = kmalloc(sizeof(struct kgd_mem), GFP_KERNEL);
 257         if ((*mem) == NULL)
 258                 return -ENOMEM;
 259
 260         r = radeon_bo_create(rdev, size, PAGE_SIZE, true, RADEON_GEM_DOMAIN_GTT,
 261                                 RADEON_GEM_GTT_WC, NULL, NULL, &(*mem)->bo);
 262         if (r) {
 263                 dev_err(rdev->dev,
 264                         "failed to allocate BO for amdkfd (%d)\n", r);
 265                 return r;
 266         }
 267
 268         /* map the buffer */
 269         r = radeon_bo_reserve((*mem)->bo, true);
 270         if (r) {
 271                 dev_err(rdev->dev, "(%d) failed to reserve bo for amdkfd\n", r);
 272                 goto allocate_mem_reserve_bo_failed;
 273         }
 274
 275         r = radeon_bo_pin((*mem)->bo, RADEON_GEM_DOMAIN_GTT,
 276                                 &(*mem)->gpu_addr);
 277         if (r) {
 278                 dev_err(rdev->dev, "(%d) failed to pin bo for amdkfd\n", r);
 279                 goto allocate_mem_pin_bo_failed;
 280         }
 281         *gpu_addr = (*mem)->gpu_addr;
 282
 283         r = radeon_bo_kmap((*mem)->bo, &(*mem)->cpu_ptr);
 284         if (r) {
 285                 dev_err(rdev->dev,
 286                         "(%d) failed to map bo to kernel for amdkfd\n", r);
 287                 goto allocate_mem_kmap_bo_failed;
 288         }
 289         *cpu_ptr = (*mem)->cpu_ptr;
 290
 291         radeon_bo_unreserve((*mem)->bo);
 292
 293         return 0;
 294
 295 allocate_mem_kmap_bo_failed:
 296         radeon_bo_unpin((*mem)->bo);
 297 allocate_mem_pin_bo_failed:
 298         radeon_bo_unreserve((*mem)->bo);
 299 allocate_mem_reserve_bo_failed:
 300         radeon_bo_unref(&(*mem)->bo);
 301
 302         return r;
 303 }
 304
 305 static void free_gtt_mem(struct kgd_dev *kgd, void *mem_obj)
 306 {
 307         struct kgd_mem *mem = (struct kgd_mem *) mem_obj;
 308
 309         BUG_ON(mem == NULL);
 310
 311         radeon_bo_reserve(mem->bo, true);
 312         radeon_bo_kunmap(mem->bo);
 313         radeon_bo_unpin(mem->bo);
 314         radeon_bo_unreserve(mem->bo);
 315         radeon_bo_unref(&(mem->bo));
 316         kfree(mem);
 317 }
 318
 319 static uint64_t get_vmem_size(struct kgd_dev *kgd)
 320 {
 321         struct radeon_device *rdev = (struct radeon_device *)kgd;
 322
 323         BUG_ON(kgd == NULL);
 324
 325         return rdev->mc.real_vram_size;
 326 }
 327
 328 static uint64_t get_gpu_clock_counter(struct kgd_dev *kgd)
 329 {
 330         struct radeon_device *rdev = (struct radeon_device *)kgd;
 331
 332         return rdev->asic->get_gpu_clock_counter(rdev);
 333 }
 334
 335 static uint32_t get_max_engine_clock_in_mhz(struct kgd_dev *kgd)
 336 {
 337         struct radeon_device *rdev = (struct radeon_device *)kgd;
 338
 339         /* The sclk is in quantas of 10kHz */
 340         return rdev->pm.dpm.dyn_state.max_clock_voltage_on_ac.sclk / 100;
 341 }
 342
 343 static inline struct radeon_device *get_radeon_device(struct kgd_dev *kgd)
 344 {
 345         return (struct radeon_device *)kgd;
 346 }
 347
 348 static void write_register(struct kgd_dev *kgd, uint32_t offset, uint32_t value)
 349 {
 350         struct radeon_device *rdev = get_radeon_device(kgd);
 351
 352         writel(value, (void __iomem *)(rdev->rmmio + offset));
 353 }
 354
 355 static uint32_t read_register(struct kgd_dev *kgd, uint32_t offset)
 356 {
 357         struct radeon_device *rdev = get_radeon_device(kgd);
 358
 359         return readl((void __iomem *)(rdev->rmmio + offset));
 360 }
 361
 362 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
 363                         uint32_t queue, uint32_t vmid)
 364 {
 365         struct radeon_device *rdev = get_radeon_device(kgd);
 366         uint32_t value = PIPEID(pipe) | MEID(mec) | VMID(vmid) | QUEUEID(queue);
 367
 368         mutex_lock(&rdev->srbm_mutex);
 369         write_register(kgd, SRBM_GFX_CNTL, value);
 370 }
 371
 372 static void unlock_srbm(struct kgd_dev *kgd)
 373 {
 374         struct radeon_device *rdev = get_radeon_device(kgd);
 375
 376         write_register(kgd, SRBM_GFX_CNTL, 0);
 377         mutex_unlock(&rdev->srbm_mutex);
 378 }
 379
 380 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
 381                                 uint32_t queue_id)
 382 {
 383         uint32_t mec = (++pipe_id / CIK_PIPE_PER_MEC) + 1;
 384         uint32_t pipe = (pipe_id % CIK_PIPE_PER_MEC);
 385
 386         lock_srbm(kgd, mec, pipe, queue_id, 0);
 387 }
 388
 389 static void release_queue(struct kgd_dev *kgd)
 390 {
 391         unlock_srbm(kgd);
 392 }
 393
 394 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
 395                                         uint32_t sh_mem_config,
 396                                         uint32_t sh_mem_ape1_base,
 397                                         uint32_t sh_mem_ape1_limit,
 398                                         uint32_t sh_mem_bases)
 399 {
 400         lock_srbm(kgd, 0, 0, 0, vmid);
 401
 402         write_register(kgd, SH_MEM_CONFIG, sh_mem_config);
 403         write_register(kgd, SH_MEM_APE1_BASE, sh_mem_ape1_base);
 404         write_register(kgd, SH_MEM_APE1_LIMIT, sh_mem_ape1_limit);
 405         write_register(kgd, SH_MEM_BASES, sh_mem_bases);
 406
 407         unlock_srbm(kgd);
 408 }
 409
 410 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
 411                                         unsigned int vmid)
 412 {
 413         /*
 414          * We have to assume that there is no outstanding mapping.
 415          * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0
 416          * because a mapping is in progress or because a mapping finished and
 417          * the SW cleared it.
 418          * So the protocol is to always wait & clear.
 419          */
 420         uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
 421                                         ATC_VMID_PASID_MAPPING_VALID_MASK;
 422
 423         write_register(kgd, ATC_VMID0_PASID_MAPPING + vmid*sizeof(uint32_t),
 424                         pasid_mapping);
 425
 426         while (!(read_register(kgd, ATC_VMID_PASID_MAPPING_UPDATE_STATUS) &
 427                                                                 (1U << vmid)))
 428                 cpu_relax();
 429         write_register(kgd, ATC_VMID_PASID_MAPPING_UPDATE_STATUS, 1U << vmid);
 430
 431         /* Mapping vmid to pasid also for IH block */
 432         write_register(kgd, IH_VMID_0_LUT + vmid * sizeof(uint32_t),
 433                         pasid_mapping);
 434
 435         return 0;
 436 }
 437
 438 static int kgd_init_pipeline(struct kgd_dev *kgd, uint32_t pipe_id,
 439                                 uint32_t hpd_size, uint64_t hpd_gpu_addr)
 440 {
 441         /* nothing to do here */
 442         return 0;
 443 }
 444
 445 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
 446 {
 447         uint32_t mec;
 448         uint32_t pipe;
 449
 450         mec = (pipe_id / CIK_PIPE_PER_MEC) + 1;
 451         pipe = (pipe_id % CIK_PIPE_PER_MEC);
 452
 453         lock_srbm(kgd, mec, pipe, 0, 0);
 454
 455         write_register(kgd, CPC_INT_CNTL,
 456                         TIME_STAMP_INT_ENABLE | OPCODE_ERROR_INT_ENABLE);
 457
 458         unlock_srbm(kgd);
 459
 460         return 0;
 461 }
 462
 463 static inline uint32_t get_sdma_base_addr(struct cik_sdma_rlc_registers *m)
 464 {
 465         uint32_t retval;
 466
 467         retval = m->sdma_engine_id * SDMA1_REGISTER_OFFSET +
 468                         m->sdma_queue_id * KFD_CIK_SDMA_QUEUE_OFFSET;
 469
 470         pr_debug("kfd: sdma base address: 0x%x\n", retval);
 471
 472         return retval;
 473 }
 474
 475 static inline struct cik_mqd *get_mqd(void *mqd)
 476 {
 477         return (struct cik_mqd *)mqd;
 478 }
 479
 480 static inline struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd)
 481 {
 482         return (struct cik_sdma_rlc_registers *)mqd;
 483 }
 484
 485 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
 486                         uint32_t queue_id, uint32_t __user *wptr)
 487 {
 488         uint32_t wptr_shadow, is_wptr_shadow_valid;
 489         struct cik_mqd *m;
 490
 491         m = get_mqd(mqd);
 492
 493         is_wptr_shadow_valid = !get_user(wptr_shadow, wptr);
 494
 495         acquire_queue(kgd, pipe_id, queue_id);
 496         write_register(kgd, CP_MQD_BASE_ADDR, m->cp_mqd_base_addr_lo);
 497         write_register(kgd, CP_MQD_BASE_ADDR_HI, m->cp_mqd_base_addr_hi);
 498         write_register(kgd, CP_MQD_CONTROL, m->cp_mqd_control);
 499
 500         write_register(kgd, CP_HQD_PQ_BASE, m->cp_hqd_pq_base_lo);
 501         write_register(kgd, CP_HQD_PQ_BASE_HI, m->cp_hqd_pq_base_hi);
 502         write_register(kgd, CP_HQD_PQ_CONTROL, m->cp_hqd_pq_control);
 503
 504         write_register(kgd, CP_HQD_IB_CONTROL, m->cp_hqd_ib_control);
 505         write_register(kgd, CP_HQD_IB_BASE_ADDR, m->cp_hqd_ib_base_addr_lo);
 506         write_register(kgd, CP_HQD_IB_BASE_ADDR_HI, m->cp_hqd_ib_base_addr_hi);
 507
 508         write_register(kgd, CP_HQD_IB_RPTR, m->cp_hqd_ib_rptr);
 509
 510         write_register(kgd, CP_HQD_PERSISTENT_STATE,
 511                         m->cp_hqd_persistent_state);
 512         write_register(kgd, CP_HQD_SEMA_CMD, m->cp_hqd_sema_cmd);
 513         write_register(kgd, CP_HQD_MSG_TYPE, m->cp_hqd_msg_type);
 514
 515         write_register(kgd, CP_HQD_ATOMIC0_PREOP_LO,
 516                         m->cp_hqd_atomic0_preop_lo);
 517
 518         write_register(kgd, CP_HQD_ATOMIC0_PREOP_HI,
 519                         m->cp_hqd_atomic0_preop_hi);
 520
 521         write_register(kgd, CP_HQD_ATOMIC1_PREOP_LO,
 522                         m->cp_hqd_atomic1_preop_lo);
 523
 524         write_register(kgd, CP_HQD_ATOMIC1_PREOP_HI,
 525                         m->cp_hqd_atomic1_preop_hi);
 526
 527         write_register(kgd, CP_HQD_PQ_RPTR_REPORT_ADDR,
 528                         m->cp_hqd_pq_rptr_report_addr_lo);
 529
 530         write_register(kgd, CP_HQD_PQ_RPTR_REPORT_ADDR_HI,
 531                         m->cp_hqd_pq_rptr_report_addr_hi);
 532
 533         write_register(kgd, CP_HQD_PQ_RPTR, m->cp_hqd_pq_rptr);
 534
 535         write_register(kgd, CP_HQD_PQ_WPTR_POLL_ADDR,
 536                         m->cp_hqd_pq_wptr_poll_addr_lo);
 537
 538         write_register(kgd, CP_HQD_PQ_WPTR_POLL_ADDR_HI,
 539                         m->cp_hqd_pq_wptr_poll_addr_hi);
 540
 541         write_register(kgd, CP_HQD_PQ_DOORBELL_CONTROL,
 542                         m->cp_hqd_pq_doorbell_control);
 543
 544         write_register(kgd, CP_HQD_VMID, m->cp_hqd_vmid);
 545
 546         write_register(kgd, CP_HQD_QUANTUM, m->cp_hqd_quantum);
 547
 548         write_register(kgd, CP_HQD_PIPE_PRIORITY, m->cp_hqd_pipe_priority);
 549         write_register(kgd, CP_HQD_QUEUE_PRIORITY, m->cp_hqd_queue_priority);
 550
 551         write_register(kgd, CP_HQD_IQ_RPTR, m->cp_hqd_iq_rptr);
 552
 553         if (is_wptr_shadow_valid)
 554                 write_register(kgd, CP_HQD_PQ_WPTR, wptr_shadow);
 555
 556         write_register(kgd, CP_HQD_ACTIVE, m->cp_hqd_active);
 557         release_queue(kgd);
 558
 559         return 0;
 560 }
 561
 562 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd)
 563 {
 564         struct cik_sdma_rlc_registers *m;
 565         uint32_t sdma_base_addr;
 566
 567         m = get_sdma_mqd(mqd);
 568         sdma_base_addr = get_sdma_base_addr(m);
 569
 570         write_register(kgd,
 571                         sdma_base_addr + SDMA0_RLC0_VIRTUAL_ADDR,
 572                         m->sdma_rlc_virtual_addr);
 573
 574         write_register(kgd,
 575                         sdma_base_addr + SDMA0_RLC0_RB_BASE,
 576                         m->sdma_rlc_rb_base);
 577
 578         write_register(kgd,
 579                         sdma_base_addr + SDMA0_RLC0_RB_BASE_HI,
 580                         m->sdma_rlc_rb_base_hi);
 581
 582         write_register(kgd,
 583                         sdma_base_addr + SDMA0_RLC0_RB_RPTR_ADDR_LO,
 584                         m->sdma_rlc_rb_rptr_addr_lo);
 585
 586         write_register(kgd,
 587                         sdma_base_addr + SDMA0_RLC0_RB_RPTR_ADDR_HI,
 588                         m->sdma_rlc_rb_rptr_addr_hi);
 589
 590         write_register(kgd,
 591                         sdma_base_addr + SDMA0_RLC0_DOORBELL,
 592                         m->sdma_rlc_doorbell);
 593
 594         write_register(kgd,
 595                         sdma_base_addr + SDMA0_RLC0_RB_CNTL,
 596                         m->sdma_rlc_rb_cntl);
 597
 598         return 0;
 599 }
 600
 601 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
 602                                 uint32_t pipe_id, uint32_t queue_id)
 603 {
 604         uint32_t act;
 605         bool retval = false;
 606         uint32_t low, high;
 607
 608         acquire_queue(kgd, pipe_id, queue_id);
 609         act = read_register(kgd, CP_HQD_ACTIVE);
 610         if (act) {
 611                 low = lower_32_bits(queue_address >> 8);
 612                 high = upper_32_bits(queue_address >> 8);
 613
 614                 if (low == read_register(kgd, CP_HQD_PQ_BASE) &&
 615                                 high == read_register(kgd, CP_HQD_PQ_BASE_HI))
 616                         retval = true;
 617         }
 618         release_queue(kgd);
 619         return retval;
 620 }
 621
 622 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
 623 {
 624         struct cik_sdma_rlc_registers *m;
 625         uint32_t sdma_base_addr;
 626         uint32_t sdma_rlc_rb_cntl;
 627
 628         m = get_sdma_mqd(mqd);
 629         sdma_base_addr = get_sdma_base_addr(m);
 630
 631         sdma_rlc_rb_cntl = read_register(kgd,
 632                                         sdma_base_addr + SDMA0_RLC0_RB_CNTL);
 633
 634         if (sdma_rlc_rb_cntl & SDMA_RB_ENABLE)
 635                 return true;
 636
 637         return false;
 638 }
 639
 640 static int kgd_hqd_destroy(struct kgd_dev *kgd, uint32_t reset_type,
 641                                 unsigned int timeout, uint32_t pipe_id,
 642                                 uint32_t queue_id)
 643 {
 644         uint32_t temp;
 645
 646         acquire_queue(kgd, pipe_id, queue_id);
 647         write_register(kgd, CP_HQD_PQ_DOORBELL_CONTROL, 0);
 648
 649         write_register(kgd, CP_HQD_DEQUEUE_REQUEST, reset_type);
 650
 651         while (true) {
 652                 temp = read_register(kgd, CP_HQD_ACTIVE);
 653                 if (temp & 0x1)
 654                         break;
 655                 if (timeout == 0) {
 656                         pr_err("kfd: cp queue preemption time out (%dms)\n",
 657                                 temp);
 658                         release_queue(kgd);
 659                         return -ETIME;
 660                 }
 661                 msleep(20);
 662                 timeout -= 20;
 663         }
 664
 665         release_queue(kgd);
 666         return 0;
 667 }
 668
 669 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
 670                                 unsigned int timeout)
 671 {
 672         struct cik_sdma_rlc_registers *m;
 673         uint32_t sdma_base_addr;
 674         uint32_t temp;
 675
 676         m = get_sdma_mqd(mqd);
 677         sdma_base_addr = get_sdma_base_addr(m);
 678
 679         temp = read_register(kgd, sdma_base_addr + SDMA0_RLC0_RB_CNTL);
 680         temp = temp & ~SDMA_RB_ENABLE;
 681         write_register(kgd, sdma_base_addr + SDMA0_RLC0_RB_CNTL, temp);
 682
 683         while (true) {
 684                 temp = read_register(kgd, sdma_base_addr +
 685                                                 SDMA0_RLC0_CONTEXT_STATUS);
 686                 if (temp & SDMA_RLC_IDLE)
 687                         break;
 688                 if (timeout == 0)
 689                         return -ETIME;
 690                 msleep(20);
 691                 timeout -= 20;
 692         }
 693
 694         write_register(kgd, sdma_base_addr + SDMA0_RLC0_DOORBELL, 0);
 695         write_register(kgd, sdma_base_addr + SDMA0_RLC0_RB_RPTR, 0);
 696         write_register(kgd, sdma_base_addr + SDMA0_RLC0_RB_WPTR, 0);
 697         write_register(kgd, sdma_base_addr + SDMA0_RLC0_RB_BASE, 0);
 698
 699         return 0;
 700 }
 701
 702 static int kgd_address_watch_disable(struct kgd_dev *kgd)
 703 {
 704         union TCP_WATCH_CNTL_BITS cntl;
 705         unsigned int i;
 706
 707         cntl.u32All = 0;
 708
 709         cntl.bitfields.valid = 0;
 710         cntl.bitfields.mask = ADDRESS_WATCH_REG_CNTL_DEFAULT_MASK;
 711         cntl.bitfields.atc = 1;
 712
 713         /* Turning off this address until we set all the registers */
 714         for (i = 0; i < MAX_WATCH_ADDRESSES; i++)
 715                 write_register(kgd,
 716                                 watchRegs[i * ADDRESS_WATCH_REG_MAX +
 717                                         ADDRESS_WATCH_REG_CNTL],
 718                                 cntl.u32All);
 719
 720         return 0;
 721 }
 722
 723 static int kgd_address_watch_execute(struct kgd_dev *kgd,
 724                                         unsigned int watch_point_id,
 725                                         uint32_t cntl_val,
 726                                         uint32_t addr_hi,
 727                                         uint32_t addr_lo)
 728 {
 729         union TCP_WATCH_CNTL_BITS cntl;
 730
 731         cntl.u32All = cntl_val;
 732
 733         /* Turning off this watch point until we set all the registers */
 734         cntl.bitfields.valid = 0;
 735         write_register(kgd,
 736                         watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX +
 737                                 ADDRESS_WATCH_REG_CNTL],
 738                         cntl.u32All);
 739
 740         write_register(kgd,
 741                         watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX +
 742                                 ADDRESS_WATCH_REG_ADDR_HI],
 743                         addr_hi);
 744
 745         write_register(kgd,
 746                         watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX +
 747                                 ADDRESS_WATCH_REG_ADDR_LO],
 748                         addr_lo);
 749
 750         /* Enable the watch point */
 751         cntl.bitfields.valid = 1;
 752
 753         write_register(kgd,
 754                         watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX +
 755                                 ADDRESS_WATCH_REG_CNTL],
 756                         cntl.u32All);
 757
 758         return 0;
 759 }
 760
 761 static int kgd_wave_control_execute(struct kgd_dev *kgd,
 762                                         uint32_t gfx_index_val,
 763                                         uint32_t sq_cmd)
 764 {
 765         struct radeon_device *rdev = get_radeon_device(kgd);
 766         uint32_t data;
 767
 768         mutex_lock(&rdev->grbm_idx_mutex);
 769
 770         write_register(kgd, GRBM_GFX_INDEX, gfx_index_val);
 771         write_register(kgd, SQ_CMD, sq_cmd);
 772
 773         /*  Restore the GRBM_GFX_INDEX register  */
 774
 775         data = INSTANCE_BROADCAST_WRITES | SH_BROADCAST_WRITES |
 776                 SE_BROADCAST_WRITES;
 777
 778         write_register(kgd, GRBM_GFX_INDEX, data);
 779
 780         mutex_unlock(&rdev->grbm_idx_mutex);
 781
 782         return 0;
 783 }
 784
 785 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
 786                                         unsigned int watch_point_id,
 787                                         unsigned int reg_offset)
 788 {
 789         return watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX + reg_offset];
 790 }
 791
 792 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd, uint8_t vmid)
 793 {
 794         uint32_t reg;
 795         struct radeon_device *rdev = (struct radeon_device *) kgd;
 796
 797         reg = RREG32(ATC_VMID0_PASID_MAPPING + vmid*4);
 798         return reg & ATC_VMID_PASID_MAPPING_VALID_MASK;
 799 }
 800
 801 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
 802                                                         uint8_t vmid)
 803 {
 804         uint32_t reg;
 805         struct radeon_device *rdev = (struct radeon_device *) kgd;
 806
 807         reg = RREG32(ATC_VMID0_PASID_MAPPING + vmid*4);
 808         return reg & ATC_VMID_PASID_MAPPING_PASID_MASK;
 809 }
 810
 811 static void write_vmid_invalidate_request(struct kgd_dev *kgd, uint8_t vmid)
 812 {
 813         struct radeon_device *rdev = (struct radeon_device *) kgd;
 814
 815         return WREG32(VM_INVALIDATE_REQUEST, 1 << vmid);
 816 }
 817
 818 static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type)
 819 {
 820         struct radeon_device *rdev = (struct radeon_device *) kgd;
 821         const union radeon_firmware_header *hdr;
 822
 823         BUG_ON(kgd == NULL || rdev->mec_fw == NULL);
 824
 825         switch (type) {
 826         case KGD_ENGINE_PFP:
 827                 hdr = (const union radeon_firmware_header *) rdev->pfp_fw->data;
 828                 break;
 829
 830         case KGD_ENGINE_ME:
 831                 hdr = (const union radeon_firmware_header *) rdev->me_fw->data;
 832                 break;
 833
 834         case KGD_ENGINE_CE:
 835                 hdr = (const union radeon_firmware_header *) rdev->ce_fw->data;
 836                 break;
 837
 838         case KGD_ENGINE_MEC1:
 839                 hdr = (const union radeon_firmware_header *) rdev->mec_fw->data;
 840                 break;
 841
 842         case KGD_ENGINE_MEC2:
 843                 hdr = (const union radeon_firmware_header *)
 844                                                         rdev->mec2_fw->data;
 845                 break;
 846
 847         case KGD_ENGINE_RLC:
 848                 hdr = (const union radeon_firmware_header *) rdev->rlc_fw->data;
 849                 break;
 850
 851         case KGD_ENGINE_SDMA1:
 852         case KGD_ENGINE_SDMA2:
 853                 hdr = (const union radeon_firmware_header *)
 854                                                         rdev->sdma_fw->data;
 855                 break;
 856
 857         default:
 858                 return 0;
 859         }
 860
 861         if (hdr == NULL)
 862                 return 0;
 863
 864         /* Only 12 bit in use*/
 865         return hdr->common.ucode_version;
 866 }