drivers/ddr/altera/sequencer.c

   1 /*
   2  * Copyright Altera Corporation (C) 2012-2015
   3  *
   4  * SPDX-License-Identifier:    BSD-3-Clause
   5  */
   6
   7 #include <common.h>
   8 #include <asm/io.h>
   9 #include <asm/arch/sdram.h>
  10 #include <errno.h>
  11 #include "sequencer.h"
  12 #include "sequencer_auto.h"
  13 #include "sequencer_auto_ac_init.h"
  14 #include "sequencer_auto_inst_init.h"
  15 #include "sequencer_defines.h"
  16
  17 static struct socfpga_sdr_rw_load_manager *sdr_rw_load_mgr_regs =
  18         (struct socfpga_sdr_rw_load_manager *)(SDR_PHYGRP_RWMGRGRP_ADDRESS | 0x800);
  19
  20 static struct socfpga_sdr_rw_load_jump_manager *sdr_rw_load_jump_mgr_regs =
  21         (struct socfpga_sdr_rw_load_jump_manager *)(SDR_PHYGRP_RWMGRGRP_ADDRESS | 0xC00);
  22
  23 static struct socfpga_sdr_reg_file *sdr_reg_file =
  24         (struct socfpga_sdr_reg_file *)SDR_PHYGRP_REGFILEGRP_ADDRESS;
  25
  26 static struct socfpga_sdr_scc_mgr *sdr_scc_mgr =
  27         (struct socfpga_sdr_scc_mgr *)(SDR_PHYGRP_SCCGRP_ADDRESS | 0xe00);
  28
  29 static struct socfpga_phy_mgr_cmd *phy_mgr_cmd =
  30         (struct socfpga_phy_mgr_cmd *)SDR_PHYGRP_PHYMGRGRP_ADDRESS;
  31
  32 static struct socfpga_phy_mgr_cfg *phy_mgr_cfg =
  33         (struct socfpga_phy_mgr_cfg *)(SDR_PHYGRP_PHYMGRGRP_ADDRESS | 0x40);
  34
  35 static struct socfpga_data_mgr *data_mgr =
  36         (struct socfpga_data_mgr *)SDR_PHYGRP_DATAMGRGRP_ADDRESS;
  37
  38 static struct socfpga_sdr_ctrl *sdr_ctrl =
  39         (struct socfpga_sdr_ctrl *)SDR_CTRLGRP_ADDRESS;
  40
  41 #define DELTA_D         1
  42
  43 /*
  44  * In order to reduce ROM size, most of the selectable calibration steps are
  45  * decided at compile time based on the user's calibration mode selection,
  46  * as captured by the STATIC_CALIB_STEPS selection below.
  47  *
  48  * However, to support simulation-time selection of fast simulation mode, where
  49  * we skip everything except the bare minimum, we need a few of the steps to
  50  * be dynamic.  In those cases, we either use the DYNAMIC_CALIB_STEPS for the
  51  * check, which is based on the rtl-supplied value, or we dynamically compute
  52  * the value to use based on the dynamically-chosen calibration mode
  53  */
  54
  55 #define DLEVEL 0
  56 #define STATIC_IN_RTL_SIM 0
  57 #define STATIC_SKIP_DELAY_LOOPS 0
  58
  59 #define STATIC_CALIB_STEPS (STATIC_IN_RTL_SIM | CALIB_SKIP_FULL_TEST | \
  60         STATIC_SKIP_DELAY_LOOPS)
  61
  62 /* calibration steps requested by the rtl */
  63 uint16_t dyn_calib_steps;
  64
  65 /*
  66  * To make CALIB_SKIP_DELAY_LOOPS a dynamic conditional option
  67  * instead of static, we use boolean logic to select between
  68  * non-skip and skip values
  69  *
  70  * The mask is set to include all bits when not-skipping, but is
  71  * zero when skipping
  72  */
  73
  74 uint16_t skip_delay_mask;       /* mask off bits when skipping/not-skipping */
  75
  76 #define SKIP_DELAY_LOOP_VALUE_OR_ZERO(non_skip_value) \
  77         ((non_skip_value) & skip_delay_mask)
  78
  79 struct gbl_type *gbl;
  80 struct param_type *param;
  81 uint32_t curr_shadow_reg;
  82
  83 static uint32_t rw_mgr_mem_calibrate_write_test(uint32_t rank_bgn,
  84         uint32_t write_group, uint32_t use_dm,
  85         uint32_t all_correct, uint32_t *bit_chk, uint32_t all_ranks);
  86
  87 static void set_failing_group_stage(uint32_t group, uint32_t stage,
  88         uint32_t substage)
  89 {
  90         /*
  91          * Only set the global stage if there was not been any other
  92          * failing group
  93          */
  94         if (gbl->error_stage == CAL_STAGE_NIL)  {
  95                 gbl->error_substage = substage;
  96                 gbl->error_stage = stage;
  97                 gbl->error_group = group;
  98         }
  99 }
 100
 101 static void reg_file_set_group(u16 set_group)
 102 {
 103         clrsetbits_le32(&sdr_reg_file->cur_stage, 0xffff0000, set_group << 16);
 104 }
 105
 106 static void reg_file_set_stage(u8 set_stage)
 107 {
 108         clrsetbits_le32(&sdr_reg_file->cur_stage, 0xffff, set_stage & 0xff);
 109 }
 110
 111 static void reg_file_set_sub_stage(u8 set_sub_stage)
 112 {
 113         set_sub_stage &= 0xff;
 114         clrsetbits_le32(&sdr_reg_file->cur_stage, 0xff00, set_sub_stage << 8);
 115 }
 116
 117 /**
 118  * phy_mgr_initialize() - Initialize PHY Manager
 119  *
 120  * Initialize PHY Manager.
 121  */
 122 static void phy_mgr_initialize(void)
 123 {
 124         u32 ratio;
 125
 126         debug("%s:%d\n", __func__, __LINE__);
 127         /* Calibration has control over path to memory */
 128         /*
 129          * In Hard PHY this is a 2-bit control:
 130          * 0: AFI Mux Select
 131          * 1: DDIO Mux Select
 132          */
 133         writel(0x3, &phy_mgr_cfg->mux_sel);
 134
 135         /* USER memory clock is not stable we begin initialization  */
 136         writel(0, &phy_mgr_cfg->reset_mem_stbl);
 137
 138         /* USER calibration status all set to zero */
 139         writel(0, &phy_mgr_cfg->cal_status);
 140
 141         writel(0, &phy_mgr_cfg->cal_debug_info);
 142
 143         /* Init params only if we do NOT skip calibration. */
 144         if ((dyn_calib_steps & CALIB_SKIP_ALL) == CALIB_SKIP_ALL)
 145                 return;
 146
 147         ratio = RW_MGR_MEM_DQ_PER_READ_DQS /
 148                 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS;
 149         param->read_correct_mask_vg = (1 << ratio) - 1;
 150         param->write_correct_mask_vg = (1 << ratio) - 1;
 151         param->read_correct_mask = (1 << RW_MGR_MEM_DQ_PER_READ_DQS) - 1;
 152         param->write_correct_mask = (1 << RW_MGR_MEM_DQ_PER_WRITE_DQS) - 1;
 153         ratio = RW_MGR_MEM_DATA_WIDTH /
 154                 RW_MGR_MEM_DATA_MASK_WIDTH;
 155         param->dm_correct_mask = (1 << ratio) - 1;
 156 }
 157
 158 /**
 159  * set_rank_and_odt_mask() - Set Rank and ODT mask
 160  * @rank:       Rank mask
 161  * @odt_mode:   ODT mode, OFF or READ_WRITE
 162  *
 163  * Set Rank and ODT mask (On-Die Termination).
 164  */
 165 static void set_rank_and_odt_mask(const u32 rank, const u32 odt_mode)
 166 {
 167         u32 odt_mask_0 = 0;
 168         u32 odt_mask_1 = 0;
 169         u32 cs_and_odt_mask;
 170
 171         if (odt_mode == RW_MGR_ODT_MODE_OFF) {
 172                 odt_mask_0 = 0x0;
 173                 odt_mask_1 = 0x0;
 174         } else {        /* RW_MGR_ODT_MODE_READ_WRITE */
 175                 switch (RW_MGR_MEM_NUMBER_OF_RANKS) {
 176                 case 1: /* 1 Rank */
 177                         /* Read: ODT = 0 ; Write: ODT = 1 */
 178                         odt_mask_0 = 0x0;
 179                         odt_mask_1 = 0x1;
 180                         break;
 181                 case 2: /* 2 Ranks */
 182                         if (RW_MGR_MEM_NUMBER_OF_CS_PER_DIMM == 1) {
 183                                 /*
 184                                  * - Dual-Slot , Single-Rank (1 CS per DIMM)
 185                                  *   OR
 186                                  * - RDIMM, 4 total CS (2 CS per DIMM, 2 DIMM)
 187                                  *
 188                                  * Since MEM_NUMBER_OF_RANKS is 2, they
 189                                  * are both single rank with 2 CS each
 190                                  * (special for RDIMM).
 191                                  *
 192                                  * Read: Turn on ODT on the opposite rank
 193                                  * Write: Turn on ODT on all ranks
 194                                  */
 195                                 odt_mask_0 = 0x3 & ~(1 << rank);
 196                                 odt_mask_1 = 0x3;
 197                         } else {
 198                                 /*
 199                                  * - Single-Slot , Dual-Rank (2 CS per DIMM)
 200                                  *
 201                                  * Read: Turn on ODT off on all ranks
 202                                  * Write: Turn on ODT on active rank
 203                                  */
 204                                 odt_mask_0 = 0x0;
 205                                 odt_mask_1 = 0x3 & (1 << rank);
 206                         }
 207                         break;
 208                 case 4: /* 4 Ranks */
 209                         /* Read:
 210                          * ----------+-----------------------+
 211                          *           |         ODT           |
 212                          * Read From +-----------------------+
 213                          *   Rank    |  3  |  2  |  1  |  0  |
 214                          * ----------+-----+-----+-----+-----+
 215                          *     0     |  0  |  1  |  0  |  0  |
 216                          *     1     |  1  |  0  |  0  |  0  |
 217                          *     2     |  0  |  0  |  0  |  1  |
 218                          *     3     |  0  |  0  |  1  |  0  |
 219                          * ----------+-----+-----+-----+-----+
 220                          *
 221                          * Write:
 222                          * ----------+-----------------------+
 223                          *           |         ODT           |
 224                          * Write To  +-----------------------+
 225                          *   Rank    |  3  |  2  |  1  |  0  |
 226                          * ----------+-----+-----+-----+-----+
 227                          *     0     |  0  |  1  |  0  |  1  |
 228                          *     1     |  1  |  0  |  1  |  0  |
 229                          *     2     |  0  |  1  |  0  |  1  |
 230                          *     3     |  1  |  0  |  1  |  0  |
 231                          * ----------+-----+-----+-----+-----+
 232                          */
 233                         switch (rank) {
 234                         case 0:
 235                                 odt_mask_0 = 0x4;
 236                                 odt_mask_1 = 0x5;
 237                                 break;
 238                         case 1:
 239                                 odt_mask_0 = 0x8;
 240                                 odt_mask_1 = 0xA;
 241                                 break;
 242                         case 2:
 243                                 odt_mask_0 = 0x1;
 244                                 odt_mask_1 = 0x5;
 245                                 break;
 246                         case 3:
 247                                 odt_mask_0 = 0x2;
 248                                 odt_mask_1 = 0xA;
 249                                 break;
 250                         }
 251                         break;
 252                 }
 253         }
 254
 255         cs_and_odt_mask = (0xFF & ~(1 << rank)) |
 256                           ((0xFF & odt_mask_0) << 8) |
 257                           ((0xFF & odt_mask_1) << 16);
 258         writel(cs_and_odt_mask, SDR_PHYGRP_RWMGRGRP_ADDRESS |
 259                                 RW_MGR_SET_CS_AND_ODT_MASK_OFFSET);
 260 }
 261
 262 /**
 263  * scc_mgr_set() - Set SCC Manager register
 264  * @off:        Base offset in SCC Manager space
 265  * @grp:        Read/Write group
 266  * @val:        Value to be set
 267  *
 268  * This function sets the SCC Manager (Scan Chain Control Manager) register.
 269  */
 270 static void scc_mgr_set(u32 off, u32 grp, u32 val)
 271 {
 272         writel(val, SDR_PHYGRP_SCCGRP_ADDRESS | off | (grp << 2));
 273 }
 274
 275 /**
 276  * scc_mgr_initialize() - Initialize SCC Manager registers
 277  *
 278  * Initialize SCC Manager registers.
 279  */
 280 static void scc_mgr_initialize(void)
 281 {
 282         /*
 283          * Clear register file for HPS. 16 (2^4) is the size of the
 284          * full register file in the scc mgr:
 285          *      RFILE_DEPTH = 1 + log2(MEM_DQ_PER_DQS + 1 + MEM_DM_PER_DQS +
 286          *                             MEM_IF_READ_DQS_WIDTH - 1);
 287          */
 288         int i;
 289
 290         for (i = 0; i < 16; i++) {
 291                 debug_cond(DLEVEL == 1, "%s:%d: Clearing SCC RFILE index %u\n",
 292                            __func__, __LINE__, i);
 293                 scc_mgr_set(SCC_MGR_HHP_RFILE_OFFSET, 0, i);
 294         }
 295 }
 296
 297 static void scc_mgr_set_dqdqs_output_phase(uint32_t write_group, uint32_t phase)
 298 {
 299         scc_mgr_set(SCC_MGR_DQDQS_OUT_PHASE_OFFSET, write_group, phase);
 300 }
 301
 302 static void scc_mgr_set_dqs_bus_in_delay(uint32_t read_group, uint32_t delay)
 303 {
 304         scc_mgr_set(SCC_MGR_DQS_IN_DELAY_OFFSET, read_group, delay);
 305 }
 306
 307 static void scc_mgr_set_dqs_en_phase(uint32_t read_group, uint32_t phase)
 308 {
 309         scc_mgr_set(SCC_MGR_DQS_EN_PHASE_OFFSET, read_group, phase);
 310 }
 311
 312 static void scc_mgr_set_dqs_en_delay(uint32_t read_group, uint32_t delay)
 313 {
 314         scc_mgr_set(SCC_MGR_DQS_EN_DELAY_OFFSET, read_group, delay);
 315 }
 316
 317 static void scc_mgr_set_dqs_io_in_delay(uint32_t delay)
 318 {
 319         scc_mgr_set(SCC_MGR_IO_IN_DELAY_OFFSET, RW_MGR_MEM_DQ_PER_WRITE_DQS,
 320                     delay);
 321 }
 322
 323 static void scc_mgr_set_dq_in_delay(uint32_t dq_in_group, uint32_t delay)
 324 {
 325         scc_mgr_set(SCC_MGR_IO_IN_DELAY_OFFSET, dq_in_group, delay);
 326 }
 327
 328 static void scc_mgr_set_dq_out1_delay(uint32_t dq_in_group, uint32_t delay)
 329 {
 330         scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET, dq_in_group, delay);
 331 }
 332
 333 static void scc_mgr_set_dqs_out1_delay(uint32_t delay)
 334 {
 335         scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET, RW_MGR_MEM_DQ_PER_WRITE_DQS,
 336                     delay);
 337 }
 338
 339 static void scc_mgr_set_dm_out1_delay(uint32_t dm, uint32_t delay)
 340 {
 341         scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET,
 342                     RW_MGR_MEM_DQ_PER_WRITE_DQS + 1 + dm,
 343                     delay);
 344 }
 345
 346 /* load up dqs config settings */
 347 static void scc_mgr_load_dqs(uint32_t dqs)
 348 {
 349         writel(dqs, &sdr_scc_mgr->dqs_ena);
 350 }
 351
 352 /* load up dqs io config settings */
 353 static void scc_mgr_load_dqs_io(void)
 354 {
 355         writel(0, &sdr_scc_mgr->dqs_io_ena);
 356 }
 357
 358 /* load up dq config settings */
 359 static void scc_mgr_load_dq(uint32_t dq_in_group)
 360 {
 361         writel(dq_in_group, &sdr_scc_mgr->dq_ena);
 362 }
 363
 364 /* load up dm config settings */
 365 static void scc_mgr_load_dm(uint32_t dm)
 366 {
 367         writel(dm, &sdr_scc_mgr->dm_ena);
 368 }
 369
 370 /**
 371  * scc_mgr_set_all_ranks() - Set SCC Manager register for all ranks
 372  * @off:        Base offset in SCC Manager space
 373  * @grp:        Read/Write group
 374  * @val:        Value to be set
 375  * @update:     If non-zero, trigger SCC Manager update for all ranks
 376  *
 377  * This function sets the SCC Manager (Scan Chain Control Manager) register
 378  * and optionally triggers the SCC update for all ranks.
 379  */
 380 static void scc_mgr_set_all_ranks(const u32 off, const u32 grp, const u32 val,
 381                                   const int update)
 382 {
 383         u32 r;
 384
 385         for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
 386              r += NUM_RANKS_PER_SHADOW_REG) {
 387                 scc_mgr_set(off, grp, val);
 388
 389                 if (update || (r == 0)) {
 390                         writel(grp, &sdr_scc_mgr->dqs_ena);
 391                         writel(0, &sdr_scc_mgr->update);
 392                 }
 393         }
 394 }
 395
 396 static void scc_mgr_set_dqs_en_phase_all_ranks(u32 read_group, u32 phase)
 397 {
 398         /*
 399          * USER although the h/w doesn't support different phases per
 400          * shadow register, for simplicity our scc manager modeling
 401          * keeps different phase settings per shadow reg, and it's
 402          * important for us to keep them in sync to match h/w.
 403          * for efficiency, the scan chain update should occur only
 404          * once to sr0.
 405          */
 406         scc_mgr_set_all_ranks(SCC_MGR_DQS_EN_PHASE_OFFSET,
 407                               read_group, phase, 0);
 408 }
 409
 410 static void scc_mgr_set_dqdqs_output_phase_all_ranks(uint32_t write_group,
 411                                                      uint32_t phase)
 412 {
 413         /*
 414          * USER although the h/w doesn't support different phases per
 415          * shadow register, for simplicity our scc manager modeling
 416          * keeps different phase settings per shadow reg, and it's
 417          * important for us to keep them in sync to match h/w.
 418          * for efficiency, the scan chain update should occur only
 419          * once to sr0.
 420          */
 421         scc_mgr_set_all_ranks(SCC_MGR_DQDQS_OUT_PHASE_OFFSET,
 422                               write_group, phase, 0);
 423 }
 424
 425 static void scc_mgr_set_dqs_en_delay_all_ranks(uint32_t read_group,
 426                                                uint32_t delay)
 427 {
 428         /*
 429          * In shadow register mode, the T11 settings are stored in
 430          * registers in the core, which are updated by the DQS_ENA
 431          * signals. Not issuing the SCC_MGR_UPD command allows us to
 432          * save lots of rank switching overhead, by calling
 433          * select_shadow_regs_for_update with update_scan_chains
 434          * set to 0.
 435          */
 436         scc_mgr_set_all_ranks(SCC_MGR_DQS_EN_DELAY_OFFSET,
 437                               read_group, delay, 1);
 438         writel(0, &sdr_scc_mgr->update);
 439 }
 440
 441 /**
 442  * scc_mgr_set_oct_out1_delay() - Set OCT output delay
 443  * @write_group:        Write group
 444  * @delay:              Delay value
 445  *
 446  * This function sets the OCT output delay in SCC manager.
 447  */
 448 static void scc_mgr_set_oct_out1_delay(const u32 write_group, const u32 delay)
 449 {
 450         const int ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
 451                           RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
 452         const int base = write_group * ratio;
 453         int i;
 454         /*
 455          * Load the setting in the SCC manager
 456          * Although OCT affects only write data, the OCT delay is controlled
 457          * by the DQS logic block which is instantiated once per read group.
 458          * For protocols where a write group consists of multiple read groups,
 459          * the setting must be set multiple times.
 460          */
 461         for (i = 0; i < ratio; i++)
 462                 scc_mgr_set(SCC_MGR_OCT_OUT1_DELAY_OFFSET, base + i, delay);
 463 }
 464
 465 /**
 466  * scc_mgr_set_hhp_extras() - Set HHP extras.
 467  *
 468  * Load the fixed setting in the SCC manager HHP extras.
 469  */
 470 static void scc_mgr_set_hhp_extras(void)
 471 {
 472         /*
 473          * Load the fixed setting in the SCC manager
 474          * bits: 0:0 = 1'b1     - DQS bypass
 475          * bits: 1:1 = 1'b1     - DQ bypass
 476          * bits: 4:2 = 3'b001   - rfifo_mode
 477          * bits: 6:5 = 2'b01    - rfifo clock_select
 478          * bits: 7:7 = 1'b0     - separate gating from ungating setting
 479          * bits: 8:8 = 1'b0     - separate OE from Output delay setting
 480          */
 481         const u32 value = (0 << 8) | (0 << 7) | (1 << 5) |
 482                           (1 << 2) | (1 << 1) | (1 << 0);
 483         const u32 addr = SDR_PHYGRP_SCCGRP_ADDRESS |
 484                          SCC_MGR_HHP_GLOBALS_OFFSET |
 485                          SCC_MGR_HHP_EXTRAS_OFFSET;
 486
 487         debug_cond(DLEVEL == 1, "%s:%d Setting HHP Extras\n",
 488                    __func__, __LINE__);
 489         writel(value, addr);
 490         debug_cond(DLEVEL == 1, "%s:%d Done Setting HHP Extras\n",
 491                    __func__, __LINE__);
 492 }
 493
 494 /**
 495  * scc_mgr_zero_all() - Zero all DQS config
 496  *
 497  * Zero all DQS config.
 498  */
 499 static void scc_mgr_zero_all(void)
 500 {
 501         int i, r;
 502
 503         /*
 504          * USER Zero all DQS config settings, across all groups and all
 505          * shadow registers
 506          */
 507         for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
 508              r += NUM_RANKS_PER_SHADOW_REG) {
 509                 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
 510                         /*
 511                          * The phases actually don't exist on a per-rank basis,
 512                          * but there's no harm updating them several times, so
 513                          * let's keep the code simple.
 514                          */
 515                         scc_mgr_set_dqs_bus_in_delay(i, IO_DQS_IN_RESERVE);
 516                         scc_mgr_set_dqs_en_phase(i, 0);
 517                         scc_mgr_set_dqs_en_delay(i, 0);
 518                 }
 519
 520                 for (i = 0; i < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; i++) {
 521                         scc_mgr_set_dqdqs_output_phase(i, 0);
 522                         /* Arria V/Cyclone V don't have out2. */
 523                         scc_mgr_set_oct_out1_delay(i, IO_DQS_OUT_RESERVE);
 524                 }
 525         }
 526
 527         /* Multicast to all DQS group enables. */
 528         writel(0xff, &sdr_scc_mgr->dqs_ena);
 529         writel(0, &sdr_scc_mgr->update);
 530 }
 531
 532 /**
 533  * scc_set_bypass_mode() - Set bypass mode and trigger SCC update
 534  * @write_group:        Write group
 535  *
 536  * Set bypass mode and trigger SCC update.
 537  */
 538 static void scc_set_bypass_mode(const u32 write_group)
 539 {
 540         /* Multicast to all DQ enables. */
 541         writel(0xff, &sdr_scc_mgr->dq_ena);
 542         writel(0xff, &sdr_scc_mgr->dm_ena);
 543
 544         /* Update current DQS IO enable. */
 545         writel(0, &sdr_scc_mgr->dqs_io_ena);
 546
 547         /* Update the DQS logic. */
 548         writel(write_group, &sdr_scc_mgr->dqs_ena);
 549
 550         /* Hit update. */
 551         writel(0, &sdr_scc_mgr->update);
 552 }
 553
 554 /**
 555  * scc_mgr_load_dqs_for_write_group() - Load DQS settings for Write Group
 556  * @write_group:        Write group
 557  *
 558  * Load DQS settings for Write Group, do not trigger SCC update.
 559  */
 560 static void scc_mgr_load_dqs_for_write_group(const u32 write_group)
 561 {
 562         const int ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
 563                           RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
 564         const int base = write_group * ratio;
 565         int i;
 566         /*
 567          * Load the setting in the SCC manager
 568          * Although OCT affects only write data, the OCT delay is controlled
 569          * by the DQS logic block which is instantiated once per read group.
 570          * For protocols where a write group consists of multiple read groups,
 571          * the setting must be set multiple times.
 572          */
 573         for (i = 0; i < ratio; i++)
 574                 writel(base + i, &sdr_scc_mgr->dqs_ena);
 575 }
 576
 577 /**
 578  * scc_mgr_zero_group() - Zero all configs for a group
 579  *
 580  * Zero DQ, DM, DQS and OCT configs for a group.
 581  */
 582 static void scc_mgr_zero_group(const u32 write_group, const int out_only)
 583 {
 584         int i, r;
 585
 586         for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
 587              r += NUM_RANKS_PER_SHADOW_REG) {
 588                 /* Zero all DQ config settings. */
 589                 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
 590                         scc_mgr_set_dq_out1_delay(i, 0);
 591                         if (!out_only)
 592                                 scc_mgr_set_dq_in_delay(i, 0);
 593                 }
 594
 595                 /* Multicast to all DQ enables. */
 596                 writel(0xff, &sdr_scc_mgr->dq_ena);
 597
 598                 /* Zero all DM config settings. */
 599                 for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++)
 600                         scc_mgr_set_dm_out1_delay(i, 0);
 601
 602                 /* Multicast to all DM enables. */
 603                 writel(0xff, &sdr_scc_mgr->dm_ena);
 604
 605                 /* Zero all DQS IO settings. */
 606                 if (!out_only)
 607                         scc_mgr_set_dqs_io_in_delay(0);
 608
 609                 /* Arria V/Cyclone V don't have out2. */
 610                 scc_mgr_set_dqs_out1_delay(IO_DQS_OUT_RESERVE);
 611                 scc_mgr_set_oct_out1_delay(write_group, IO_DQS_OUT_RESERVE);
 612                 scc_mgr_load_dqs_for_write_group(write_group);
 613
 614                 /* Multicast to all DQS IO enables (only 1 in total). */
 615                 writel(0, &sdr_scc_mgr->dqs_io_ena);
 616
 617                 /* Hit update to zero everything. */
 618                 writel(0, &sdr_scc_mgr->update);
 619         }
 620 }
 621
 622 /*
 623  * apply and load a particular input delay for the DQ pins in a group
 624  * group_bgn is the index of the first dq pin (in the write group)
 625  */
 626 static void scc_mgr_apply_group_dq_in_delay(uint32_t group_bgn, uint32_t delay)
 627 {
 628         uint32_t i, p;
 629
 630         for (i = 0, p = group_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++) {
 631                 scc_mgr_set_dq_in_delay(p, delay);
 632                 scc_mgr_load_dq(p);
 633         }
 634 }
 635
 636 /**
 637  * scc_mgr_apply_group_dq_out1_delay() - Apply and load an output delay for the DQ pins in a group
 638  * @delay:              Delay value
 639  *
 640  * Apply and load a particular output delay for the DQ pins in a group.
 641  */
 642 static void scc_mgr_apply_group_dq_out1_delay(const u32 delay)
 643 {
 644         int i;
 645
 646         for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
 647                 scc_mgr_set_dq_out1_delay(i, delay);
 648                 scc_mgr_load_dq(i);
 649         }
 650 }
 651
 652 /* apply and load a particular output delay for the DM pins in a group */
 653 static void scc_mgr_apply_group_dm_out1_delay(uint32_t delay1)
 654 {
 655         uint32_t i;
 656
 657         for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++) {
 658                 scc_mgr_set_dm_out1_delay(i, delay1);
 659                 scc_mgr_load_dm(i);
 660         }
 661 }
 662
 663
 664 /* apply and load delay on both DQS and OCT out1 */
 665 static void scc_mgr_apply_group_dqs_io_and_oct_out1(uint32_t write_group,
 666                                                     uint32_t delay)
 667 {
 668         scc_mgr_set_dqs_out1_delay(delay);
 669         scc_mgr_load_dqs_io();
 670
 671         scc_mgr_set_oct_out1_delay(write_group, delay);
 672         scc_mgr_load_dqs_for_write_group(write_group);
 673 }
 674
 675 /**
 676  * scc_mgr_apply_group_all_out_delay_add() - Apply a delay to the entire output side: DQ, DM, DQS, OCT
 677  * @write_group:        Write group
 678  * @delay:              Delay value
 679  *
 680  * Apply a delay to the entire output side: DQ, DM, DQS, OCT.
 681  */
 682 static void scc_mgr_apply_group_all_out_delay_add(const u32 write_group,
 683                                                   const u32 delay)
 684 {
 685         u32 i, new_delay;
 686
 687         /* DQ shift */
 688         for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++)
 689                 scc_mgr_load_dq(i);
 690
 691         /* DM shift */
 692         for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++)
 693                 scc_mgr_load_dm(i);
 694
 695         /* DQS shift */
 696         new_delay = READ_SCC_DQS_IO_OUT2_DELAY + delay;
 697         if (new_delay > IO_IO_OUT2_DELAY_MAX) {
 698                 debug_cond(DLEVEL == 1,
 699                            "%s:%d (%u, %u) DQS: %u > %d; adding %u to OUT1\n",
 700                            __func__, __LINE__, write_group, delay, new_delay,
 701                            IO_IO_OUT2_DELAY_MAX,
 702                            new_delay - IO_IO_OUT2_DELAY_MAX);
 703                 new_delay -= IO_IO_OUT2_DELAY_MAX;
 704                 scc_mgr_set_dqs_out1_delay(new_delay);
 705         }
 706
 707         scc_mgr_load_dqs_io();
 708
 709         /* OCT shift */
 710         new_delay = READ_SCC_OCT_OUT2_DELAY + delay;
 711         if (new_delay > IO_IO_OUT2_DELAY_MAX) {
 712                 debug_cond(DLEVEL == 1,
 713                            "%s:%d (%u, %u) DQS: %u > %d; adding %u to OUT1\n",
 714                            __func__, __LINE__, write_group, delay,
 715                            new_delay, IO_IO_OUT2_DELAY_MAX,
 716                            new_delay - IO_IO_OUT2_DELAY_MAX);
 717                 new_delay -= IO_IO_OUT2_DELAY_MAX;
 718                 scc_mgr_set_oct_out1_delay(write_group, new_delay);
 719         }
 720
 721         scc_mgr_load_dqs_for_write_group(write_group);
 722 }
 723
 724 /**
 725  * scc_mgr_apply_group_all_out_delay_add() - Apply a delay to the entire output side to all ranks
 726  * @write_group:        Write group
 727  * @delay:              Delay value
 728  *
 729  * Apply a delay to the entire output side (DQ, DM, DQS, OCT) to all ranks.
 730  */
 731 static void
 732 scc_mgr_apply_group_all_out_delay_add_all_ranks(const u32 write_group,
 733                                                 const u32 delay)
 734 {
 735         int r;
 736
 737         for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
 738              r += NUM_RANKS_PER_SHADOW_REG) {
 739                 scc_mgr_apply_group_all_out_delay_add(write_group, delay);
 740                 writel(0, &sdr_scc_mgr->update);
 741         }
 742 }
 743
 744 /**
 745  * set_jump_as_return() - Return instruction optimization
 746  *
 747  * Optimization used to recover some slots in ddr3 inst_rom could be
 748  * applied to other protocols if we wanted to
 749  */
 750 static void set_jump_as_return(void)
 751 {
 752         /*
 753          * To save space, we replace return with jump to special shared
 754          * RETURN instruction so we set the counter to large value so that
 755          * we always jump.
 756          */
 757         writel(0xff, &sdr_rw_load_mgr_regs->load_cntr0);
 758         writel(RW_MGR_RETURN, &sdr_rw_load_jump_mgr_regs->load_jump_add0);
 759 }
 760
 761 /*
 762  * should always use constants as argument to ensure all computations are
 763  * performed at compile time
 764  */
 765 static void delay_for_n_mem_clocks(const uint32_t clocks)
 766 {
 767         uint32_t afi_clocks;
 768         uint8_t inner = 0;
 769         uint8_t outer = 0;
 770         uint16_t c_loop = 0;
 771
 772         debug("%s:%d: clocks=%u ... start\n", __func__, __LINE__, clocks);
 773
 774
 775         afi_clocks = (clocks + AFI_RATE_RATIO-1) / AFI_RATE_RATIO;
 776         /* scale (rounding up) to get afi clocks */
 777
 778         /*
 779          * Note, we don't bother accounting for being off a little bit
 780          * because of a few extra instructions in outer loops
 781          * Note, the loops have a test at the end, and do the test before
 782          * the decrement, and so always perform the loop
 783          * 1 time more than the counter value
 784          */
 785         if (afi_clocks == 0) {
 786                 ;
 787         } else if (afi_clocks <= 0x100) {
 788                 inner = afi_clocks-1;
 789                 outer = 0;
 790                 c_loop = 0;
 791         } else if (afi_clocks <= 0x10000) {
 792                 inner = 0xff;
 793                 outer = (afi_clocks-1) >> 8;
 794                 c_loop = 0;
 795         } else {
 796                 inner = 0xff;
 797                 outer = 0xff;
 798                 c_loop = (afi_clocks-1) >> 16;
 799         }
 800
 801         /*
 802          * rom instructions are structured as follows:
 803          *
 804          *    IDLE_LOOP2: jnz cntr0, TARGET_A
 805          *    IDLE_LOOP1: jnz cntr1, TARGET_B
 806          *                return
 807          *
 808          * so, when doing nested loops, TARGET_A is set to IDLE_LOOP2, and
 809          * TARGET_B is set to IDLE_LOOP2 as well
 810          *
 811          * if we have no outer loop, though, then we can use IDLE_LOOP1 only,
 812          * and set TARGET_B to IDLE_LOOP1 and we skip IDLE_LOOP2 entirely
 813          *
 814          * a little confusing, but it helps save precious space in the inst_rom
 815          * and sequencer rom and keeps the delays more accurate and reduces
 816          * overhead
 817          */
 818         if (afi_clocks <= 0x100) {
 819                 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(inner),
 820                         &sdr_rw_load_mgr_regs->load_cntr1);
 821
 822                 writel(RW_MGR_IDLE_LOOP1,
 823                         &sdr_rw_load_jump_mgr_regs->load_jump_add1);
 824
 825                 writel(RW_MGR_IDLE_LOOP1, SDR_PHYGRP_RWMGRGRP_ADDRESS |
 826                                           RW_MGR_RUN_SINGLE_GROUP_OFFSET);
 827         } else {
 828                 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(inner),
 829                         &sdr_rw_load_mgr_regs->load_cntr0);
 830
 831                 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(outer),
 832                         &sdr_rw_load_mgr_regs->load_cntr1);
 833
 834                 writel(RW_MGR_IDLE_LOOP2,
 835                         &sdr_rw_load_jump_mgr_regs->load_jump_add0);
 836
 837                 writel(RW_MGR_IDLE_LOOP2,
 838                         &sdr_rw_load_jump_mgr_regs->load_jump_add1);
 839
 840                 /* hack to get around compiler not being smart enough */
 841                 if (afi_clocks <= 0x10000) {
 842                         /* only need to run once */
 843                         writel(RW_MGR_IDLE_LOOP2, SDR_PHYGRP_RWMGRGRP_ADDRESS |
 844                                                   RW_MGR_RUN_SINGLE_GROUP_OFFSET);
 845                 } else {
 846                         do {
 847                                 writel(RW_MGR_IDLE_LOOP2,
 848                                         SDR_PHYGRP_RWMGRGRP_ADDRESS |
 849                                         RW_MGR_RUN_SINGLE_GROUP_OFFSET);
 850                         } while (c_loop-- != 0);
 851                 }
 852         }
 853         debug("%s:%d clocks=%u ... end\n", __func__, __LINE__, clocks);
 854 }
 855
 856 /**
 857  * rw_mgr_mem_init_load_regs() - Load instruction registers
 858  * @cntr0:      Counter 0 value
 859  * @cntr1:      Counter 1 value
 860  * @cntr2:      Counter 2 value
 861  * @jump:       Jump instruction value
 862  *
 863  * Load instruction registers.
 864  */
 865 static void rw_mgr_mem_init_load_regs(u32 cntr0, u32 cntr1, u32 cntr2, u32 jump)
 866 {
 867         uint32_t grpaddr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
 868                            RW_MGR_RUN_SINGLE_GROUP_OFFSET;
 869
 870         /* Load counters */
 871         writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(cntr0),
 872                &sdr_rw_load_mgr_regs->load_cntr0);
 873         writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(cntr1),
 874                &sdr_rw_load_mgr_regs->load_cntr1);
 875         writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(cntr2),
 876                &sdr_rw_load_mgr_regs->load_cntr2);
 877
 878         /* Load jump address */
 879         writel(jump, &sdr_rw_load_jump_mgr_regs->load_jump_add0);
 880         writel(jump, &sdr_rw_load_jump_mgr_regs->load_jump_add1);
 881         writel(jump, &sdr_rw_load_jump_mgr_regs->load_jump_add2);
 882
 883         /* Execute count instruction */
 884         writel(jump, grpaddr);
 885 }
 886
 887 /**
 888  * rw_mgr_mem_load_user() - Load user calibration values
 889  * @fin1:       Final instruction 1
 890  * @fin2:       Final instruction 2
 891  * @precharge:  If 1, precharge the banks at the end
 892  *
 893  * Load user calibration values and optionally precharge the banks.
 894  */
 895 static void rw_mgr_mem_load_user(const u32 fin1, const u32 fin2,
 896                                  const int precharge)
 897 {
 898         u32 grpaddr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
 899                       RW_MGR_RUN_SINGLE_GROUP_OFFSET;
 900         u32 r;
 901
 902         for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
 903                 if (param->skip_ranks[r]) {
 904                         /* request to skip the rank */
 905                         continue;
 906                 }
 907
 908                 /* set rank */
 909                 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);
 910
 911                 /* precharge all banks ... */
 912                 if (precharge)
 913                         writel(RW_MGR_PRECHARGE_ALL, grpaddr);
 914
 915                 /*
 916                  * USER Use Mirror-ed commands for odd ranks if address
 917                  * mirrorring is on
 918                  */
 919                 if ((RW_MGR_MEM_ADDRESS_MIRRORING >> r) & 0x1) {
 920                         set_jump_as_return();
 921                         writel(RW_MGR_MRS2_MIRR, grpaddr);
 922                         delay_for_n_mem_clocks(4);
 923                         set_jump_as_return();
 924                         writel(RW_MGR_MRS3_MIRR, grpaddr);
 925                         delay_for_n_mem_clocks(4);
 926                         set_jump_as_return();
 927                         writel(RW_MGR_MRS1_MIRR, grpaddr);
 928                         delay_for_n_mem_clocks(4);
 929                         set_jump_as_return();
 930                         writel(fin1, grpaddr);
 931                 } else {
 932                         set_jump_as_return();
 933                         writel(RW_MGR_MRS2, grpaddr);
 934                         delay_for_n_mem_clocks(4);
 935                         set_jump_as_return();
 936                         writel(RW_MGR_MRS3, grpaddr);
 937                         delay_for_n_mem_clocks(4);
 938                         set_jump_as_return();
 939                         writel(RW_MGR_MRS1, grpaddr);
 940                         set_jump_as_return();
 941                         writel(fin2, grpaddr);
 942                 }
 943
 944                 if (precharge)
 945                         continue;
 946
 947                 set_jump_as_return();
 948                 writel(RW_MGR_ZQCL, grpaddr);
 949
 950                 /* tZQinit = tDLLK = 512 ck cycles */
 951                 delay_for_n_mem_clocks(512);
 952         }
 953 }
 954
 955 /**
 956  * rw_mgr_mem_initialize() - Initialize RW Manager
 957  *
 958  * Initialize RW Manager.
 959  */
 960 static void rw_mgr_mem_initialize(void)
 961 {
 962         debug("%s:%d\n", __func__, __LINE__);
 963
 964         /* The reset / cke part of initialization is broadcasted to all ranks */
 965         writel(RW_MGR_RANK_ALL, SDR_PHYGRP_RWMGRGRP_ADDRESS |
 966                                 RW_MGR_SET_CS_AND_ODT_MASK_OFFSET);
 967
 968         /*
 969          * Here's how you load register for a loop
 970          * Counters are located @ 0x800
 971          * Jump address are located @ 0xC00
 972          * For both, registers 0 to 3 are selected using bits 3 and 2, like
 973          * in 0x800, 0x804, 0x808, 0x80C and 0xC00, 0xC04, 0xC08, 0xC0C
 974          * I know this ain't pretty, but Avalon bus throws away the 2 least
 975          * significant bits
 976          */
 977
 978         /* Start with memory RESET activated */
 979
 980         /* tINIT = 200us */
 981
 982         /*
 983          * 200us @ 266MHz (3.75 ns) ~ 54000 clock cycles
 984          * If a and b are the number of iteration in 2 nested loops
 985          * it takes the following number of cycles to complete the operation:
 986          * number_of_cycles = ((2 + n) * a + 2) * b
 987          * where n is the number of instruction in the inner loop
 988          * One possible solution is n = 0 , a = 256 , b = 106 => a = FF,
 989          * b = 6A
 990          */
 991         rw_mgr_mem_init_load_regs(SEQ_TINIT_CNTR0_VAL, SEQ_TINIT_CNTR1_VAL,
 992                                   SEQ_TINIT_CNTR2_VAL,
 993                                   RW_MGR_INIT_RESET_0_CKE_0);
 994
 995         /* Indicate that memory is stable. */
 996         writel(1, &phy_mgr_cfg->reset_mem_stbl);
 997
 998         /*
 999          * transition the RESET to high
1000          * Wait for 500us
1001          */
1002
1003         /*
1004          * 500us @ 266MHz (3.75 ns) ~ 134000 clock cycles
1005          * If a and b are the number of iteration in 2 nested loops
1006          * it takes the following number of cycles to complete the operation
1007          * number_of_cycles = ((2 + n) * a + 2) * b
1008          * where n is the number of instruction in the inner loop
1009          * One possible solution is n = 2 , a = 131 , b = 256 => a = 83,
1010          * b = FF
1011          */
1012         rw_mgr_mem_init_load_regs(SEQ_TRESET_CNTR0_VAL, SEQ_TRESET_CNTR1_VAL,
1013                                   SEQ_TRESET_CNTR2_VAL,
1014                                   RW_MGR_INIT_RESET_1_CKE_0);
1015
1016         /* Bring up clock enable. */
1017
1018         /* tXRP < 250 ck cycles */
1019         delay_for_n_mem_clocks(250);
1020
1021         rw_mgr_mem_load_user(RW_MGR_MRS0_DLL_RESET_MIRR, RW_MGR_MRS0_DLL_RESET,
1022                              0);
1023 }
1024
1025 /*
1026  * At the end of calibration we have to program the user settings in, and
1027  * USER  hand off the memory to the user.
1028  */
1029 static void rw_mgr_mem_handoff(void)
1030 {
1031         rw_mgr_mem_load_user(RW_MGR_MRS0_USER_MIRR, RW_MGR_MRS0_USER, 1);
1032         /*
1033          * USER  need to wait tMOD (12CK or 15ns) time before issuing
1034          * other commands, but we will have plenty of NIOS cycles before
1035          * actual handoff so its okay.
1036          */
1037 }
1038
1039 /**
1040  * rw_mgr_mem_calibrate_read_test_patterns() - Read back test patterns
1041  * @rank_bgn:   Rank number
1042  * @group:      Read/Write Group
1043  * @all_ranks:  Test all ranks
1044  *
1045  * Performs a guaranteed read on the patterns we are going to use during a
1046  * read test to ensure memory works.
1047  */
1048 static int
1049 rw_mgr_mem_calibrate_read_test_patterns(const u32 rank_bgn, const u32 group,
1050                                         const u32 all_ranks)
1051 {
1052         const u32 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
1053                          RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1054         const u32 addr_offset =
1055                          (group * RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS) << 2;
1056         const u32 rank_end = all_ranks ?
1057                                 RW_MGR_MEM_NUMBER_OF_RANKS :
1058                                 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
1059         const u32 shift_ratio = RW_MGR_MEM_DQ_PER_READ_DQS /
1060                                 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS;
1061         const u32 correct_mask_vg = param->read_correct_mask_vg;
1062
1063         u32 tmp_bit_chk, base_rw_mgr, bit_chk;
1064         int vg, r;
1065         int ret = 0;
1066
1067         bit_chk = param->read_correct_mask;
1068
1069         for (r = rank_bgn; r < rank_end; r++) {
1070                 /* Request to skip the rank */
1071                 if (param->skip_ranks[r])
1072                         continue;
1073
1074                 /* Set rank */
1075                 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
1076
1077                 /* Load up a constant bursts of read commands */
1078                 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr0);
1079                 writel(RW_MGR_GUARANTEED_READ,
1080                         &sdr_rw_load_jump_mgr_regs->load_jump_add0);
1081
1082                 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr1);
1083                 writel(RW_MGR_GUARANTEED_READ_CONT,
1084                         &sdr_rw_load_jump_mgr_regs->load_jump_add1);
1085
1086                 tmp_bit_chk = 0;
1087                 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS - 1;
1088                      vg >= 0; vg--) {
1089                         /* Reset the FIFOs to get pointers to known state. */
1090                         writel(0, &phy_mgr_cmd->fifo_reset);
1091                         writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
1092                                   RW_MGR_RESET_READ_DATAPATH_OFFSET);
1093                         writel(RW_MGR_GUARANTEED_READ,
1094                                addr + addr_offset + (vg << 2));
1095
1096                         base_rw_mgr = readl(SDR_PHYGRP_RWMGRGRP_ADDRESS);
1097                         tmp_bit_chk <<= shift_ratio;
1098                         tmp_bit_chk |= correct_mask_vg & ~base_rw_mgr;
1099                 }
1100
1101                 bit_chk &= tmp_bit_chk;
1102         }
1103
1104         writel(RW_MGR_CLEAR_DQS_ENABLE, addr + (group << 2));
1105
1106         set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1107
1108         if (bit_chk != param->read_correct_mask)
1109                 ret = -EIO;
1110
1111         debug_cond(DLEVEL == 1,
1112                    "%s:%d test_load_patterns(%u,ALL) => (%u == %u) => %i\n",
1113                    __func__, __LINE__, group, bit_chk,
1114                    param->read_correct_mask, ret);
1115
1116         return ret;
1117 }
1118
1119 /**
1120  * rw_mgr_mem_calibrate_read_load_patterns() - Load up the patterns for read test
1121  * @rank_bgn:   Rank number
1122  * @all_ranks:  Test all ranks
1123  *
1124  * Load up the patterns we are going to use during a read test.
1125  */
1126 static void rw_mgr_mem_calibrate_read_load_patterns(const u32 rank_bgn,
1127                                                     const int all_ranks)
1128 {
1129         const u32 rank_end = all_ranks ?
1130                         RW_MGR_MEM_NUMBER_OF_RANKS :
1131                         (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
1132         u32 r;
1133
1134         debug("%s:%d\n", __func__, __LINE__);
1135
1136         for (r = rank_bgn; r < rank_end; r++) {
1137                 if (param->skip_ranks[r])
1138                         /* request to skip the rank */
1139                         continue;
1140
1141                 /* set rank */
1142                 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
1143
1144                 /* Load up a constant bursts */
1145                 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr0);
1146
1147                 writel(RW_MGR_GUARANTEED_WRITE_WAIT0,
1148                         &sdr_rw_load_jump_mgr_regs->load_jump_add0);
1149
1150                 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr1);
1151
1152                 writel(RW_MGR_GUARANTEED_WRITE_WAIT1,
1153                         &sdr_rw_load_jump_mgr_regs->load_jump_add1);
1154
1155                 writel(0x04, &sdr_rw_load_mgr_regs->load_cntr2);
1156
1157                 writel(RW_MGR_GUARANTEED_WRITE_WAIT2,
1158                         &sdr_rw_load_jump_mgr_regs->load_jump_add2);
1159
1160                 writel(0x04, &sdr_rw_load_mgr_regs->load_cntr3);
1161
1162                 writel(RW_MGR_GUARANTEED_WRITE_WAIT3,
1163                         &sdr_rw_load_jump_mgr_regs->load_jump_add3);
1164
1165                 writel(RW_MGR_GUARANTEED_WRITE, SDR_PHYGRP_RWMGRGRP_ADDRESS |
1166                                                 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
1167         }
1168
1169         set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1170 }
1171
1172 /**
1173  * rw_mgr_mem_calibrate_read_test() - Perform READ test on single rank
1174  * @rank_bgn:           Rank number
1175  * @group:              Read/Write group
1176  * @num_tries:          Number of retries of the test
1177  * @all_correct:        All bits must be correct in the mask
1178  * @bit_chk:            Resulting bit mask after the test
1179  * @all_groups:         Test all R/W groups
1180  * @all_ranks:          Test all ranks
1181  *
1182  * Try a read and see if it returns correct data back. Test has dummy reads
1183  * inserted into the mix used to align DQS enable. Test has more thorough
1184  * checks than the regular read test.
1185  */
1186 static int
1187 rw_mgr_mem_calibrate_read_test(const u32 rank_bgn, const u32 group,
1188                                const u32 num_tries, const u32 all_correct,
1189                                u32 *bit_chk,
1190                                const u32 all_groups, const u32 all_ranks)
1191 {
1192         const u32 rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
1193                 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
1194         const u32 quick_read_mode =
1195                 ((STATIC_CALIB_STEPS & CALIB_SKIP_DELAY_SWEEPS) &&
1196                  ENABLE_SUPER_QUICK_CALIBRATION);
1197         u32 correct_mask_vg = param->read_correct_mask_vg;
1198         u32 tmp_bit_chk;
1199         u32 base_rw_mgr;
1200         u32 addr;
1201
1202         int r, vg, ret;
1203
1204         *bit_chk = param->read_correct_mask;
1205
1206         for (r = rank_bgn; r < rank_end; r++) {
1207                 if (param->skip_ranks[r])
1208                         /* request to skip the rank */
1209                         continue;
1210
1211                 /* set rank */
1212                 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
1213
1214                 writel(0x10, &sdr_rw_load_mgr_regs->load_cntr1);
1215
1216                 writel(RW_MGR_READ_B2B_WAIT1,
1217                         &sdr_rw_load_jump_mgr_regs->load_jump_add1);
1218
1219                 writel(0x10, &sdr_rw_load_mgr_regs->load_cntr2);
1220                 writel(RW_MGR_READ_B2B_WAIT2,
1221                         &sdr_rw_load_jump_mgr_regs->load_jump_add2);
1222
1223                 if (quick_read_mode)
1224                         writel(0x1, &sdr_rw_load_mgr_regs->load_cntr0);
1225                         /* need at least two (1+1) reads to capture failures */
1226                 else if (all_groups)
1227                         writel(0x06, &sdr_rw_load_mgr_regs->load_cntr0);
1228                 else
1229                         writel(0x32, &sdr_rw_load_mgr_regs->load_cntr0);
1230
1231                 writel(RW_MGR_READ_B2B,
1232                         &sdr_rw_load_jump_mgr_regs->load_jump_add0);
1233                 if (all_groups)
1234                         writel(RW_MGR_MEM_IF_READ_DQS_WIDTH *
1235                                RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS - 1,
1236                                &sdr_rw_load_mgr_regs->load_cntr3);
1237                 else
1238                         writel(0x0, &sdr_rw_load_mgr_regs->load_cntr3);
1239
1240                 writel(RW_MGR_READ_B2B,
1241                         &sdr_rw_load_jump_mgr_regs->load_jump_add3);
1242
1243                 tmp_bit_chk = 0;
1244                 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS - 1; vg >= 0;
1245                      vg--) {
1246                         /* Reset the FIFOs to get pointers to known state. */
1247                         writel(0, &phy_mgr_cmd->fifo_reset);
1248                         writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
1249                                   RW_MGR_RESET_READ_DATAPATH_OFFSET);
1250
1251                         if (all_groups) {
1252                                 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
1253                                        RW_MGR_RUN_ALL_GROUPS_OFFSET;
1254                         } else {
1255                                 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
1256                                        RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1257                         }
1258
1259                         writel(RW_MGR_READ_B2B, addr +
1260                                ((group * RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS +
1261                                vg) << 2));
1262
1263                         base_rw_mgr = readl(SDR_PHYGRP_RWMGRGRP_ADDRESS);
1264                         tmp_bit_chk <<= RW_MGR_MEM_DQ_PER_READ_DQS /
1265                                         RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS;
1266                         tmp_bit_chk |= correct_mask_vg & ~(base_rw_mgr);
1267                 }
1268
1269                 *bit_chk &= tmp_bit_chk;
1270         }
1271
1272         addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1273         writel(RW_MGR_CLEAR_DQS_ENABLE, addr + (group << 2));
1274
1275         set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1276
1277         if (all_correct) {
1278                 ret = (*bit_chk == param->read_correct_mask);
1279                 debug_cond(DLEVEL == 2,
1280                            "%s:%d read_test(%u,ALL,%u) => (%u == %u) => %i\n",
1281                            __func__, __LINE__, group, all_groups, *bit_chk,
1282                            param->read_correct_mask, ret);
1283         } else  {
1284                 ret = (*bit_chk != 0x00);
1285                 debug_cond(DLEVEL == 2,
1286                            "%s:%d read_test(%u,ONE,%u) => (%u != %u) => %i\n",
1287                            __func__, __LINE__, group, all_groups, *bit_chk,
1288                            0, ret);
1289         }
1290
1291         return ret;
1292 }
1293
1294 /**
1295  * rw_mgr_mem_calibrate_read_test_all_ranks() - Perform READ test on all ranks
1296  * @grp:                Read/Write group
1297  * @num_tries:          Number of retries of the test
1298  * @all_correct:        All bits must be correct in the mask
1299  * @all_groups:         Test all R/W groups
1300  *
1301  * Perform a READ test across all memory ranks.
1302  */
1303 static int
1304 rw_mgr_mem_calibrate_read_test_all_ranks(const u32 grp, const u32 num_tries,
1305                                          const u32 all_correct,
1306                                          const u32 all_groups)
1307 {
1308         u32 bit_chk;
1309         return rw_mgr_mem_calibrate_read_test(0, grp, num_tries, all_correct,
1310                                               &bit_chk, all_groups, 1);
1311 }
1312
1313 /**
1314  * rw_mgr_incr_vfifo() - Increase VFIFO value
1315  * @grp:        Read/Write group
1316  *
1317  * Increase VFIFO value.
1318  */
1319 static void rw_mgr_incr_vfifo(const u32 grp)
1320 {
1321         writel(grp, &phy_mgr_cmd->inc_vfifo_hard_phy);
1322 }
1323
1324 /**
1325  * rw_mgr_decr_vfifo() - Decrease VFIFO value
1326  * @grp:        Read/Write group
1327  *
1328  * Decrease VFIFO value.
1329  */
1330 static void rw_mgr_decr_vfifo(const u32 grp)
1331 {
1332         u32 i;
1333
1334         for (i = 0; i < VFIFO_SIZE - 1; i++)
1335                 rw_mgr_incr_vfifo(grp);
1336 }
1337
1338 /**
1339  * find_vfifo_failing_read() - Push VFIFO to get a failing read
1340  * @grp:        Read/Write group
1341  *
1342  * Push VFIFO until a failing read happens.
1343  */
1344 static int find_vfifo_failing_read(const u32 grp)
1345 {
1346         u32 v, ret, fail_cnt = 0;
1347
1348         for (v = 0; v < VFIFO_SIZE; v++) {
1349                 debug_cond(DLEVEL == 2, "%s:%d: vfifo %u\n",
1350                            __func__, __LINE__, v);
1351                 ret = rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1352                                                 PASS_ONE_BIT, 0);
1353                 if (!ret) {
1354                         fail_cnt++;
1355
1356                         if (fail_cnt == 2)
1357                                 return v;
1358                 }
1359
1360                 /* Fiddle with FIFO. */
1361                 rw_mgr_incr_vfifo(grp);
1362         }
1363
1364         /* No failing read found! Something must have gone wrong. */
1365         debug_cond(DLEVEL == 2, "%s:%d: vfifo failed\n", __func__, __LINE__);
1366         return 0;
1367 }
1368
1369 /**
1370  * sdr_find_phase_delay() - Find DQS enable phase or delay
1371  * @working:    If 1, look for working phase/delay, if 0, look for non-working
1372  * @delay:      If 1, look for delay, if 0, look for phase
1373  * @grp:        Read/Write group
1374  * @work:       Working window position
1375  * @work_inc:   Working window increment
1376  * @pd:         DQS Phase/Delay Iterator
1377  *
1378  * Find working or non-working DQS enable phase setting.
1379  */
1380 static int sdr_find_phase_delay(int working, int delay, const u32 grp,
1381                                 u32 *work, const u32 work_inc, u32 *pd)
1382 {
1383         const u32 max = delay ? IO_DQS_EN_DELAY_MAX : IO_DQS_EN_PHASE_MAX;
1384         u32 ret;
1385
1386         for (; *pd <= max; (*pd)++) {
1387                 if (delay)
1388                         scc_mgr_set_dqs_en_delay_all_ranks(grp, *pd);
1389                 else
1390                         scc_mgr_set_dqs_en_phase_all_ranks(grp, *pd);
1391
1392                 ret = rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1393                                         PASS_ONE_BIT, 0);
1394                 if (!working)
1395                         ret = !ret;
1396
1397                 if (ret)
1398                         return 0;
1399
1400                 if (work)
1401                         *work += work_inc;
1402         }
1403
1404         return -EINVAL;
1405 }
1406 /**
1407  * sdr_find_phase() - Find DQS enable phase
1408  * @working:    If 1, look for working phase, if 0, look for non-working phase
1409  * @grp:        Read/Write group
1410  * @work:       Working window position
1411  * @i:          Iterator
1412  * @p:          DQS Phase Iterator
1413  *
1414  * Find working or non-working DQS enable phase setting.
1415  */
1416 static int sdr_find_phase(int working, const u32 grp, u32 *work,
1417                           u32 *i, u32 *p)
1418 {
1419         const u32 end = VFIFO_SIZE + (working ? 0 : 1);
1420         int ret;
1421
1422         for (; *i < end; (*i)++) {
1423                 if (working)
1424                         *p = 0;
1425
1426                 ret = sdr_find_phase_delay(working, 0, grp, work,
1427                                            IO_DELAY_PER_OPA_TAP, p);
1428                 if (!ret)
1429                         return 0;
1430
1431                 if (*p > IO_DQS_EN_PHASE_MAX) {
1432                         /* Fiddle with FIFO. */
1433                         rw_mgr_incr_vfifo(grp);
1434                         if (!working)
1435                                 *p = 0;
1436                 }
1437         }
1438
1439         return -EINVAL;
1440 }
1441
1442 /**
1443  * sdr_working_phase() - Find working DQS enable phase
1444  * @grp:        Read/Write group
1445  * @work_bgn:   Working window start position
1446  * @d:          dtaps output value
1447  * @p:          DQS Phase Iterator
1448  * @i:          Iterator
1449  *
1450  * Find working DQS enable phase setting.
1451  */
1452 static int sdr_working_phase(const u32 grp, u32 *work_bgn, u32 *d,
1453                              u32 *p, u32 *i)
1454 {
1455         const u32 dtaps_per_ptap = IO_DELAY_PER_OPA_TAP /
1456                                    IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1457         int ret;
1458
1459         *work_bgn = 0;
1460
1461         for (*d = 0; *d <= dtaps_per_ptap; (*d)++) {
1462                 *i = 0;
1463                 scc_mgr_set_dqs_en_delay_all_ranks(grp, *d);
1464                 ret = sdr_find_phase(1, grp, work_bgn, i, p);
1465                 if (!ret)
1466                         return 0;
1467                 *work_bgn += IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1468         }
1469
1470         /* Cannot find working solution */
1471         debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: no vfifo/ptap/dtap\n",
1472                    __func__, __LINE__);
1473         return -EINVAL;
1474 }
1475
1476 /**
1477  * sdr_backup_phase() - Find DQS enable backup phase
1478  * @grp:        Read/Write group
1479  * @work_bgn:   Working window start position
1480  * @p:          DQS Phase Iterator
1481  *
1482  * Find DQS enable backup phase setting.
1483  */
1484 static void sdr_backup_phase(const u32 grp, u32 *work_bgn, u32 *p)
1485 {
1486         u32 tmp_delay, d;
1487         int ret;
1488
1489         /* Special case code for backing up a phase */
1490         if (*p == 0) {
1491                 *p = IO_DQS_EN_PHASE_MAX;
1492                 rw_mgr_decr_vfifo(grp);
1493         } else {
1494                 (*p)--;
1495         }
1496         tmp_delay = *work_bgn - IO_DELAY_PER_OPA_TAP;
1497         scc_mgr_set_dqs_en_phase_all_ranks(grp, *p);
1498
1499         for (d = 0; d <= IO_DQS_EN_DELAY_MAX && tmp_delay < *work_bgn; d++) {
1500                 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1501
1502                 ret = rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1503                                         PASS_ONE_BIT, 0);
1504                 if (ret) {
1505                         *work_bgn = tmp_delay;
1506                         break;
1507                 }
1508
1509                 tmp_delay += IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1510         }
1511
1512         /* Restore VFIFO to old state before we decremented it (if needed). */
1513         (*p)++;
1514         if (*p > IO_DQS_EN_PHASE_MAX) {
1515                 *p = 0;
1516                 rw_mgr_incr_vfifo(grp);
1517         }
1518
1519         scc_mgr_set_dqs_en_delay_all_ranks(grp, 0);
1520 }
1521
1522 /**
1523  * sdr_nonworking_phase() - Find non-working DQS enable phase
1524  * @grp:        Read/Write group
1525  * @work_end:   Working window end position
1526  * @p:          DQS Phase Iterator
1527  * @i:          Iterator
1528  *
1529  * Find non-working DQS enable phase setting.
1530  */
1531 static int sdr_nonworking_phase(const u32 grp, u32 *work_end, u32 *p, u32 *i)
1532 {
1533         int ret;
1534
1535         (*p)++;
1536         *work_end += IO_DELAY_PER_OPA_TAP;
1537         if (*p > IO_DQS_EN_PHASE_MAX) {
1538                 /* Fiddle with FIFO. */
1539                 *p = 0;
1540                 rw_mgr_incr_vfifo(grp);
1541         }
1542
1543         ret = sdr_find_phase(0, grp, work_end, i, p);
1544         if (ret) {
1545                 /* Cannot see edge of failing read. */
1546                 debug_cond(DLEVEL == 2, "%s:%d: end: failed\n",
1547                            __func__, __LINE__);
1548         }
1549
1550         return ret;
1551 }
1552
1553 /**
1554  * sdr_find_window_center() - Find center of the working DQS window.
1555  * @grp:        Read/Write group
1556  * @work_bgn:   First working settings
1557  * @work_end:   Last working settings
1558  *
1559  * Find center of the working DQS enable window.
1560  */
1561 static int sdr_find_window_center(const u32 grp, const u32 work_bgn,
1562                                   const u32 work_end)
1563 {
1564         u32 work_mid;
1565         int tmp_delay = 0;
1566         int i, p, d;
1567
1568         work_mid = (work_bgn + work_end) / 2;
1569
1570         debug_cond(DLEVEL == 2, "work_bgn=%d work_end=%d work_mid=%d\n",
1571                    work_bgn, work_end, work_mid);
1572         /* Get the middle delay to be less than a VFIFO delay */
1573         tmp_delay = (IO_DQS_EN_PHASE_MAX + 1) * IO_DELAY_PER_OPA_TAP;
1574
1575         debug_cond(DLEVEL == 2, "vfifo ptap delay %d\n", tmp_delay);
1576         work_mid %= tmp_delay;
1577         debug_cond(DLEVEL == 2, "new work_mid %d\n", work_mid);
1578
1579         tmp_delay = rounddown(work_mid, IO_DELAY_PER_OPA_TAP);
1580         if (tmp_delay > IO_DQS_EN_PHASE_MAX * IO_DELAY_PER_OPA_TAP)
1581                 tmp_delay = IO_DQS_EN_PHASE_MAX * IO_DELAY_PER_OPA_TAP;
1582         p = tmp_delay / IO_DELAY_PER_OPA_TAP;
1583
1584         debug_cond(DLEVEL == 2, "new p %d, tmp_delay=%d\n", p, tmp_delay);
1585
1586         d = DIV_ROUND_UP(work_mid - tmp_delay, IO_DELAY_PER_DQS_EN_DCHAIN_TAP);
1587         if (d > IO_DQS_EN_DELAY_MAX)
1588                 d = IO_DQS_EN_DELAY_MAX;
1589         tmp_delay += d * IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1590
1591         debug_cond(DLEVEL == 2, "new d %d, tmp_delay=%d\n", d, tmp_delay);
1592
1593         scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1594         scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1595
1596         /*
1597          * push vfifo until we can successfully calibrate. We can do this
1598          * because the largest possible margin in 1 VFIFO cycle.
1599          */
1600         for (i = 0; i < VFIFO_SIZE; i++) {
1601                 debug_cond(DLEVEL == 2, "find_dqs_en_phase: center\n");
1602                 if (rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1603                                                              PASS_ONE_BIT,
1604                                                              0)) {
1605                         debug_cond(DLEVEL == 2,
1606                                    "%s:%d center: found: ptap=%u dtap=%u\n",
1607                                    __func__, __LINE__, p, d);
1608                         return 0;
1609                 }
1610
1611                 /* Fiddle with FIFO. */
1612                 rw_mgr_incr_vfifo(grp);
1613         }
1614
1615         debug_cond(DLEVEL == 2, "%s:%d center: failed.\n",
1616                    __func__, __LINE__);
1617         return -EINVAL;
1618 }
1619
1620 /**
1621  * rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase() - Find a good DQS enable to use
1622  * @grp:        Read/Write Group
1623  *
1624  * Find a good DQS enable to use.
1625  */
1626 static int rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(const u32 grp)
1627 {
1628         u32 d, p, i;
1629         u32 dtaps_per_ptap;
1630         u32 work_bgn, work_end;
1631         u32 found_passing_read, found_failing_read, initial_failing_dtap;
1632         int ret;
1633
1634         debug("%s:%d %u\n", __func__, __LINE__, grp);
1635
1636         reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);
1637
1638         scc_mgr_set_dqs_en_delay_all_ranks(grp, 0);
1639         scc_mgr_set_dqs_en_phase_all_ranks(grp, 0);
1640
1641         /* Step 0: Determine number of delay taps for each phase tap. */
1642         dtaps_per_ptap = IO_DELAY_PER_OPA_TAP / IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1643
1644         /* Step 1: First push vfifo until we get a failing read. */
1645         find_vfifo_failing_read(grp);
1646
1647         /* Step 2: Find first working phase, increment in ptaps. */
1648         work_bgn = 0;
1649         ret = sdr_working_phase(grp, &work_bgn, &d, &p, &i);
1650         if (ret)
1651                 return ret;
1652
1653         work_end = work_bgn;
1654
1655         /*
1656          * If d is 0 then the working window covers a phase tap and we can
1657          * follow the old procedure. Otherwise, we've found the beginning
1658          * and we need to increment the dtaps until we find the end.
1659          */
1660         if (d == 0) {
1661                 /*
1662                  * Step 3a: If we have room, back off by one and
1663                  *          increment in dtaps.
1664                  */
1665                 sdr_backup_phase(grp, &work_bgn, &p);
1666
1667                 /*
1668                  * Step 4a: go forward from working phase to non working
1669                  * phase, increment in ptaps.
1670                  */
1671                 ret = sdr_nonworking_phase(grp, &work_end, &p, &i);
1672                 if (ret)
1673                         return ret;
1674
1675                 /* Step 5a: Back off one from last, increment in dtaps. */
1676
1677                 /* Special case code for backing up a phase */
1678                 if (p == 0) {
1679                         p = IO_DQS_EN_PHASE_MAX;
1680                         rw_mgr_decr_vfifo(grp);
1681                 } else {
1682                         p = p - 1;
1683                 }
1684
1685                 work_end -= IO_DELAY_PER_OPA_TAP;
1686                 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1687
1688                 d = 0;
1689
1690                 debug_cond(DLEVEL == 2, "%s:%d p: ptap=%u\n",
1691                            __func__, __LINE__, p);
1692         }
1693
1694         /* The dtap increment to find the failing edge is done here. */
1695         sdr_find_phase_delay(0, 1, grp, &work_end,
1696                              IO_DELAY_PER_DQS_EN_DCHAIN_TAP, &d);
1697
1698         /* Go back to working dtap */
1699         if (d != 0)
1700                 work_end -= IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1701
1702         debug_cond(DLEVEL == 2,
1703                    "%s:%d p/d: ptap=%u dtap=%u end=%u\n",
1704                    __func__, __LINE__, p, d - 1, work_end);
1705
1706         if (work_end < work_bgn) {
1707                 /* nil range */
1708                 debug_cond(DLEVEL == 2, "%s:%d end-2: failed\n",
1709                            __func__, __LINE__);
1710                 return -EINVAL;
1711         }
1712
1713         debug_cond(DLEVEL == 2, "%s:%d found range [%u,%u]\n",
1714                    __func__, __LINE__, work_bgn, work_end);
1715
1716         /*
1717          * We need to calculate the number of dtaps that equal a ptap.
1718          * To do that we'll back up a ptap and re-find the edge of the
1719          * window using dtaps
1720          */
1721         debug_cond(DLEVEL == 2, "%s:%d calculate dtaps_per_ptap for tracking\n",
1722                    __func__, __LINE__);
1723
1724         /* Special case code for backing up a phase */
1725         if (p == 0) {
1726                 p = IO_DQS_EN_PHASE_MAX;
1727                 rw_mgr_decr_vfifo(grp);
1728                 debug_cond(DLEVEL == 2, "%s:%d backedup cycle/phase: p=%u\n",
1729                            __func__, __LINE__, p);
1730         } else {
1731                 p = p - 1;
1732                 debug_cond(DLEVEL == 2, "%s:%d backedup phase only: p=%u",
1733                            __func__, __LINE__, p);
1734         }
1735
1736         scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1737
1738         /*
1739          * Increase dtap until we first see a passing read (in case the
1740          * window is smaller than a ptap), and then a failing read to
1741          * mark the edge of the window again.
1742          */
1743
1744         /* Find a passing read. */
1745         debug_cond(DLEVEL == 2, "%s:%d find passing read\n",
1746                    __func__, __LINE__);
1747
1748         initial_failing_dtap = d;
1749
1750         found_passing_read = !sdr_find_phase_delay(1, 1, grp, NULL, 0, &d);
1751         if (found_passing_read) {
1752                 /* Find a failing read. */
1753                 debug_cond(DLEVEL == 2, "%s:%d find failing read\n",
1754                            __func__, __LINE__);
1755                 d++;
1756                 found_failing_read = !sdr_find_phase_delay(0, 1, grp, NULL, 0,
1757                                                            &d);
1758         } else {
1759                 debug_cond(DLEVEL == 1,
1760                            "%s:%d failed to calculate dtaps per ptap. Fall back on static value\n",
1761                            __func__, __LINE__);
1762         }
1763
1764         /*
1765          * The dynamically calculated dtaps_per_ptap is only valid if we
1766          * found a passing/failing read. If we didn't, it means d hit the max
1767          * (IO_DQS_EN_DELAY_MAX). Otherwise, dtaps_per_ptap retains its
1768          * statically calculated value.
1769          */
1770         if (found_passing_read && found_failing_read)
1771                 dtaps_per_ptap = d - initial_failing_dtap;
1772
1773         writel(dtaps_per_ptap, &sdr_reg_file->dtaps_per_ptap);
1774         debug_cond(DLEVEL == 2, "%s:%d dtaps_per_ptap=%u - %u = %u",
1775                    __func__, __LINE__, d, initial_failing_dtap, dtaps_per_ptap);
1776
1777         /* Step 6: Find the centre of the window. */
1778         ret = sdr_find_window_center(grp, work_bgn, work_end);
1779
1780         return ret;
1781 }
1782
1783 /* per-bit deskew DQ and center */
1784 static uint32_t rw_mgr_mem_calibrate_vfifo_center(uint32_t rank_bgn,
1785         uint32_t write_group, uint32_t read_group, uint32_t test_bgn,
1786         uint32_t use_read_test, uint32_t update_fom)
1787 {
1788         uint32_t i, p, d, min_index;
1789         /*
1790          * Store these as signed since there are comparisons with
1791          * signed numbers.
1792          */
1793         uint32_t bit_chk;
1794         uint32_t sticky_bit_chk;
1795         int32_t left_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
1796         int32_t right_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
1797         int32_t final_dq[RW_MGR_MEM_DQ_PER_READ_DQS];
1798         int32_t mid;
1799         int32_t orig_mid_min, mid_min;
1800         int32_t new_dqs, start_dqs, start_dqs_en, shift_dq, final_dqs,
1801                 final_dqs_en;
1802         int32_t dq_margin, dqs_margin;
1803         uint32_t stop;
1804         uint32_t temp_dq_in_delay1, temp_dq_in_delay2;
1805         uint32_t addr;
1806
1807         debug("%s:%d: %u %u", __func__, __LINE__, read_group, test_bgn);
1808
1809         addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_DQS_IN_DELAY_OFFSET;
1810         start_dqs = readl(addr + (read_group << 2));
1811         if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
1812                 start_dqs_en = readl(addr + ((read_group << 2)
1813                                      - IO_DQS_EN_DELAY_OFFSET));
1814
1815         /* set the left and right edge of each bit to an illegal value */
1816         /* use (IO_IO_IN_DELAY_MAX + 1) as an illegal value */
1817         sticky_bit_chk = 0;
1818         for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1819                 left_edge[i]  = IO_IO_IN_DELAY_MAX + 1;
1820                 right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1821         }
1822
1823         /* Search for the left edge of the window for each bit */
1824         for (d = 0; d <= IO_IO_IN_DELAY_MAX; d++) {
1825                 scc_mgr_apply_group_dq_in_delay(write_group, test_bgn, d);
1826
1827                 writel(0, &sdr_scc_mgr->update);
1828
1829                 /*
1830                  * Stop searching when the read test doesn't pass AND when
1831                  * we've seen a passing read on every bit.
1832                  */
1833                 if (use_read_test) {
1834                         stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
1835                                 read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
1836                                 &bit_chk, 0, 0);
1837                 } else {
1838                         rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
1839                                                         0, PASS_ONE_BIT,
1840                                                         &bit_chk, 0);
1841                         bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
1842                                 (read_group - (write_group *
1843                                         RW_MGR_MEM_IF_READ_DQS_WIDTH /
1844                                         RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
1845                         stop = (bit_chk == 0);
1846                 }
1847                 sticky_bit_chk = sticky_bit_chk | bit_chk;
1848                 stop = stop && (sticky_bit_chk == param->read_correct_mask);
1849                 debug_cond(DLEVEL == 2, "%s:%d vfifo_center(left): dtap=%u => %u == %u \
1850                            && %u", __func__, __LINE__, d,
1851                            sticky_bit_chk,
1852                         param->read_correct_mask, stop);
1853
1854                 if (stop == 1) {
1855                         break;
1856                 } else {
1857                         for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1858                                 if (bit_chk & 1) {
1859                                         /* Remember a passing test as the
1860                                         left_edge */
1861                                         left_edge[i] = d;
1862                                 } else {
1863                                         /* If a left edge has not been seen yet,
1864                                         then a future passing test will mark
1865                                         this edge as the right edge */
1866                                         if (left_edge[i] ==
1867                                                 IO_IO_IN_DELAY_MAX + 1) {
1868                                                 right_edge[i] = -(d + 1);
1869                                         }
1870                                 }
1871                                 bit_chk = bit_chk >> 1;
1872                         }
1873                 }
1874         }
1875
1876         /* Reset DQ delay chains to 0 */
1877         scc_mgr_apply_group_dq_in_delay(test_bgn, 0);
1878         sticky_bit_chk = 0;
1879         for (i = RW_MGR_MEM_DQ_PER_READ_DQS - 1;; i--) {
1880                 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
1881                            %d right_edge[%u]: %d\n", __func__, __LINE__,
1882                            i, left_edge[i], i, right_edge[i]);
1883
1884                 /*
1885                  * Check for cases where we haven't found the left edge,
1886                  * which makes our assignment of the the right edge invalid.
1887                  * Reset it to the illegal value.
1888                  */
1889                 if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) && (
1890                         right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
1891                         right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1892                         debug_cond(DLEVEL == 2, "%s:%d vfifo_center: reset \
1893                                    right_edge[%u]: %d\n", __func__, __LINE__,
1894                                    i, right_edge[i]);
1895                 }
1896
1897                 /*
1898                  * Reset sticky bit (except for bits where we have seen
1899                  * both the left and right edge).
1900                  */
1901                 sticky_bit_chk = sticky_bit_chk << 1;
1902                 if ((left_edge[i] != IO_IO_IN_DELAY_MAX + 1) &&
1903                     (right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
1904                         sticky_bit_chk = sticky_bit_chk | 1;
1905                 }
1906
1907                 if (i == 0)
1908                         break;
1909         }
1910
1911         /* Search for the right edge of the window for each bit */
1912         for (d = 0; d <= IO_DQS_IN_DELAY_MAX - start_dqs; d++) {
1913                 scc_mgr_set_dqs_bus_in_delay(read_group, d + start_dqs);
1914                 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
1915                         uint32_t delay = d + start_dqs_en;
1916                         if (delay > IO_DQS_EN_DELAY_MAX)
1917                                 delay = IO_DQS_EN_DELAY_MAX;
1918                         scc_mgr_set_dqs_en_delay(read_group, delay);
1919                 }
1920                 scc_mgr_load_dqs(read_group);
1921
1922                 writel(0, &sdr_scc_mgr->update);
1923
1924                 /*
1925                  * Stop searching when the read test doesn't pass AND when
1926                  * we've seen a passing read on every bit.
1927                  */
1928                 if (use_read_test) {
1929                         stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
1930                                 read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
1931                                 &bit_chk, 0, 0);
1932                 } else {
1933                         rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
1934                                                         0, PASS_ONE_BIT,
1935                                                         &bit_chk, 0);
1936                         bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
1937                                 (read_group - (write_group *
1938                                         RW_MGR_MEM_IF_READ_DQS_WIDTH /
1939                                         RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
1940                         stop = (bit_chk == 0);
1941                 }
1942                 sticky_bit_chk = sticky_bit_chk | bit_chk;
1943                 stop = stop && (sticky_bit_chk == param->read_correct_mask);
1944
1945                 debug_cond(DLEVEL == 2, "%s:%d vfifo_center(right): dtap=%u => %u == \
1946                            %u && %u", __func__, __LINE__, d,
1947                            sticky_bit_chk, param->read_correct_mask, stop);
1948
1949                 if (stop == 1) {
1950                         break;
1951                 } else {
1952                         for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1953                                 if (bit_chk & 1) {
1954                                         /* Remember a passing test as
1955                                         the right_edge */
1956                                         right_edge[i] = d;
1957                                 } else {
1958                                         if (d != 0) {
1959                                                 /* If a right edge has not been
1960                                                 seen yet, then a future passing
1961                                                 test will mark this edge as the
1962                                                 left edge */
1963                                                 if (right_edge[i] ==
1964                                                 IO_IO_IN_DELAY_MAX + 1) {
1965                                                         left_edge[i] = -(d + 1);
1966                                                 }
1967                                         } else {
1968                                                 /* d = 0 failed, but it passed
1969                                                 when testing the left edge,
1970                                                 so it must be marginal,
1971                                                 set it to -1 */
1972                                                 if (right_edge[i] ==
1973                                                         IO_IO_IN_DELAY_MAX + 1 &&
1974                                                         left_edge[i] !=
1975                                                         IO_IO_IN_DELAY_MAX
1976                                                         + 1) {
1977                                                         right_edge[i] = -1;
1978                                                 }
1979                                                 /* If a right edge has not been
1980                                                 seen yet, then a future passing
1981                                                 test will mark this edge as the
1982                                                 left edge */
1983                                                 else if (right_edge[i] ==
1984                                                         IO_IO_IN_DELAY_MAX +
1985                                                         1) {
1986                                                         left_edge[i] = -(d + 1);
1987                                                 }
1988                                         }
1989                                 }
1990
1991                                 debug_cond(DLEVEL == 2, "%s:%d vfifo_center[r,\
1992                                            d=%u]: ", __func__, __LINE__, d);
1993                                 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d ",
1994                                            (int)(bit_chk & 1), i, left_edge[i]);
1995                                 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
1996                                            right_edge[i]);
1997                                 bit_chk = bit_chk >> 1;
1998                         }
1999                 }
2000         }
2001
2002         /* Check that all bits have a window */
2003         for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
2004                 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
2005                            %d right_edge[%u]: %d", __func__, __LINE__,
2006                            i, left_edge[i], i, right_edge[i]);
2007                 if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) || (right_edge[i]
2008                         == IO_IO_IN_DELAY_MAX + 1)) {
2009                         /*
2010                          * Restore delay chain settings before letting the loop
2011                          * in rw_mgr_mem_calibrate_vfifo to retry different
2012                          * dqs/ck relationships.
2013                          */
2014                         scc_mgr_set_dqs_bus_in_delay(read_group, start_dqs);
2015                         if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2016                                 scc_mgr_set_dqs_en_delay(read_group,
2017                                                          start_dqs_en);
2018                         }
2019                         scc_mgr_load_dqs(read_group);
2020                         writel(0, &sdr_scc_mgr->update);
2021
2022                         debug_cond(DLEVEL == 1, "%s:%d vfifo_center: failed to \
2023                                    find edge [%u]: %d %d", __func__, __LINE__,
2024                                    i, left_edge[i], right_edge[i]);
2025                         if (use_read_test) {
2026                                 set_failing_group_stage(read_group *
2027                                         RW_MGR_MEM_DQ_PER_READ_DQS + i,
2028                                         CAL_STAGE_VFIFO,
2029                                         CAL_SUBSTAGE_VFIFO_CENTER);
2030                         } else {
2031                                 set_failing_group_stage(read_group *
2032                                         RW_MGR_MEM_DQ_PER_READ_DQS + i,
2033                                         CAL_STAGE_VFIFO_AFTER_WRITES,
2034                                         CAL_SUBSTAGE_VFIFO_CENTER);
2035                         }
2036                         return 0;
2037                 }
2038         }
2039
2040         /* Find middle of window for each DQ bit */
2041         mid_min = left_edge[0] - right_edge[0];
2042         min_index = 0;
2043         for (i = 1; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
2044                 mid = left_edge[i] - right_edge[i];
2045                 if (mid < mid_min) {
2046                         mid_min = mid;
2047                         min_index = i;
2048                 }
2049         }
2050
2051         /*
2052          * -mid_min/2 represents the amount that we need to move DQS.
2053          * If mid_min is odd and positive we'll need to add one to
2054          * make sure the rounding in further calculations is correct
2055          * (always bias to the right), so just add 1 for all positive values.
2056          */
2057         if (mid_min > 0)
2058                 mid_min++;
2059
2060         mid_min = mid_min / 2;
2061
2062         debug_cond(DLEVEL == 1, "%s:%d vfifo_center: mid_min=%d (index=%u)\n",
2063                    __func__, __LINE__, mid_min, min_index);
2064
2065         /* Determine the amount we can change DQS (which is -mid_min) */
2066         orig_mid_min = mid_min;
2067         new_dqs = start_dqs - mid_min;
2068         if (new_dqs > IO_DQS_IN_DELAY_MAX)
2069                 new_dqs = IO_DQS_IN_DELAY_MAX;
2070         else if (new_dqs < 0)
2071                 new_dqs = 0;
2072
2073         mid_min = start_dqs - new_dqs;
2074         debug_cond(DLEVEL == 1, "vfifo_center: new mid_min=%d new_dqs=%d\n",
2075                    mid_min, new_dqs);
2076
2077         if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2078                 if (start_dqs_en - mid_min > IO_DQS_EN_DELAY_MAX)
2079                         mid_min += start_dqs_en - mid_min - IO_DQS_EN_DELAY_MAX;
2080                 else if (start_dqs_en - mid_min < 0)
2081                         mid_min += start_dqs_en - mid_min;
2082         }
2083         new_dqs = start_dqs - mid_min;
2084
2085         debug_cond(DLEVEL == 1, "vfifo_center: start_dqs=%d start_dqs_en=%d \
2086                    new_dqs=%d mid_min=%d\n", start_dqs,
2087                    IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS ? start_dqs_en : -1,
2088                    new_dqs, mid_min);
2089
2090         /* Initialize data for export structures */
2091         dqs_margin = IO_IO_IN_DELAY_MAX + 1;
2092         dq_margin  = IO_IO_IN_DELAY_MAX + 1;
2093
2094         /* add delay to bring centre of all DQ windows to the same "level" */
2095         for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++) {
2096                 /* Use values before divide by 2 to reduce round off error */
2097                 shift_dq = (left_edge[i] - right_edge[i] -
2098                         (left_edge[min_index] - right_edge[min_index]))/2  +
2099                         (orig_mid_min - mid_min);
2100
2101                 debug_cond(DLEVEL == 2, "vfifo_center: before: \
2102                            shift_dq[%u]=%d\n", i, shift_dq);
2103
2104                 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_IN_DELAY_OFFSET;
2105                 temp_dq_in_delay1 = readl(addr + (p << 2));
2106                 temp_dq_in_delay2 = readl(addr + (i << 2));
2107
2108                 if (shift_dq + (int32_t)temp_dq_in_delay1 >
2109                         (int32_t)IO_IO_IN_DELAY_MAX) {
2110                         shift_dq = (int32_t)IO_IO_IN_DELAY_MAX - temp_dq_in_delay2;
2111                 } else if (shift_dq + (int32_t)temp_dq_in_delay1 < 0) {
2112                         shift_dq = -(int32_t)temp_dq_in_delay1;
2113                 }
2114                 debug_cond(DLEVEL == 2, "vfifo_center: after: \
2115                            shift_dq[%u]=%d\n", i, shift_dq);
2116                 final_dq[i] = temp_dq_in_delay1 + shift_dq;
2117                 scc_mgr_set_dq_in_delay(p, final_dq[i]);
2118                 scc_mgr_load_dq(p);
2119
2120                 debug_cond(DLEVEL == 2, "vfifo_center: margin[%u]=[%d,%d]\n", i,
2121                            left_edge[i] - shift_dq + (-mid_min),
2122                            right_edge[i] + shift_dq - (-mid_min));
2123                 /* To determine values for export structures */
2124                 if (left_edge[i] - shift_dq + (-mid_min) < dq_margin)
2125                         dq_margin = left_edge[i] - shift_dq + (-mid_min);
2126
2127                 if (right_edge[i] + shift_dq - (-mid_min) < dqs_margin)
2128                         dqs_margin = right_edge[i] + shift_dq - (-mid_min);
2129         }
2130
2131         final_dqs = new_dqs;
2132         if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
2133                 final_dqs_en = start_dqs_en - mid_min;
2134
2135         /* Move DQS-en */
2136         if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2137                 scc_mgr_set_dqs_en_delay(read_group, final_dqs_en);
2138                 scc_mgr_load_dqs(read_group);
2139         }
2140
2141         /* Move DQS */
2142         scc_mgr_set_dqs_bus_in_delay(read_group, final_dqs);
2143         scc_mgr_load_dqs(read_group);
2144         debug_cond(DLEVEL == 2, "%s:%d vfifo_center: dq_margin=%d \
2145                    dqs_margin=%d", __func__, __LINE__,
2146                    dq_margin, dqs_margin);
2147
2148         /*
2149          * Do not remove this line as it makes sure all of our decisions
2150          * have been applied. Apply the update bit.
2151          */
2152         writel(0, &sdr_scc_mgr->update);
2153
2154         return (dq_margin >= 0) && (dqs_margin >= 0);
2155 }
2156
2157 /**
2158  * rw_mgr_mem_calibrate_guaranteed_write() - Perform guaranteed write into the device
2159  * @rw_group:   Read/Write Group
2160  * @phase:      DQ/DQS phase
2161  *
2162  * Because initially no communication ca be reliably performed with the memory
2163  * device, the sequencer uses a guaranteed write mechanism to write data into
2164  * the memory device.
2165  */
2166 static int rw_mgr_mem_calibrate_guaranteed_write(const u32 rw_group,
2167                                                  const u32 phase)
2168 {
2169         int ret;
2170
2171         /* Set a particular DQ/DQS phase. */
2172         scc_mgr_set_dqdqs_output_phase_all_ranks(rw_group, phase);
2173
2174         debug_cond(DLEVEL == 1, "%s:%d guaranteed write: g=%u p=%u\n",
2175                    __func__, __LINE__, rw_group, phase);
2176
2177         /*
2178          * Altera EMI_RM 2015.05.04 :: Figure 1-25
2179          * Load up the patterns used by read calibration using the
2180          * current DQDQS phase.
2181          */
2182         rw_mgr_mem_calibrate_read_load_patterns(0, 1);
2183
2184         if (gbl->phy_debug_mode_flags & PHY_DEBUG_DISABLE_GUARANTEED_READ)
2185                 return 0;
2186
2187         /*
2188          * Altera EMI_RM 2015.05.04 :: Figure 1-26
2189          * Back-to-Back reads of the patterns used for calibration.
2190          */
2191         ret = rw_mgr_mem_calibrate_read_test_patterns(0, rw_group, 1);
2192         if (ret)
2193                 debug_cond(DLEVEL == 1,
2194                            "%s:%d Guaranteed read test failed: g=%u p=%u\n",
2195                            __func__, __LINE__, rw_group, phase);
2196         return ret;
2197 }
2198
2199 /**
2200  * rw_mgr_mem_calibrate_dqs_enable_calibration() - DQS Enable Calibration
2201  * @rw_group:   Read/Write Group
2202  * @test_bgn:   Rank at which the test begins
2203  *
2204  * DQS enable calibration ensures reliable capture of the DQ signal without
2205  * glitches on the DQS line.
2206  */
2207 static int rw_mgr_mem_calibrate_dqs_enable_calibration(const u32 rw_group,
2208                                                        const u32 test_bgn)
2209 {
2210         /*
2211          * Altera EMI_RM 2015.05.04 :: Figure 1-27
2212          * DQS and DQS Eanble Signal Relationships.
2213          */
2214
2215         /* We start at zero, so have one less dq to devide among */
2216         const u32 delay_step = IO_IO_IN_DELAY_MAX /
2217                                (RW_MGR_MEM_DQ_PER_READ_DQS - 1);
2218         int ret;
2219         u32 i, p, d, r;
2220
2221         debug("%s:%d (%u,%u)\n", __func__, __LINE__, rw_group, test_bgn);
2222
2223         /* Try different dq_in_delays since the DQ path is shorter than DQS. */
2224         for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
2225              r += NUM_RANKS_PER_SHADOW_REG) {
2226                 for (i = 0, p = test_bgn, d = 0;
2227                      i < RW_MGR_MEM_DQ_PER_READ_DQS;
2228                      i++, p++, d += delay_step) {
2229                         debug_cond(DLEVEL == 1,
2230                                    "%s:%d: g=%u r=%u i=%u p=%u d=%u\n",
2231                                    __func__, __LINE__, rw_group, r, i, p, d);
2232
2233                         scc_mgr_set_dq_in_delay(p, d);
2234                         scc_mgr_load_dq(p);
2235                 }
2236
2237                 writel(0, &sdr_scc_mgr->update);
2238         }
2239
2240         /*
2241          * Try rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase across different
2242          * dq_in_delay values
2243          */
2244         ret = rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(rw_group);
2245
2246         debug_cond(DLEVEL == 1,
2247                    "%s:%d: g=%u found=%u; Reseting delay chain to zero\n",
2248                    __func__, __LINE__, rw_group, !ret);
2249
2250         for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
2251              r += NUM_RANKS_PER_SHADOW_REG) {
2252                 scc_mgr_apply_group_dq_in_delay(test_bgn, 0);
2253                 writel(0, &sdr_scc_mgr->update);
2254         }
2255
2256         return ret;
2257 }
2258
2259 /**
2260  * rw_mgr_mem_calibrate_dq_dqs_centering() - Centering DQ/DQS
2261  * @rw_group:           Read/Write Group
2262  * @test_bgn:           Rank at which the test begins
2263  * @use_read_test:      Perform a read test
2264  * @update_fom:         Update FOM
2265  *
2266  * The centerin DQ/DQS stage attempts to align DQ and DQS signals on reads
2267  * within a group.
2268  */
2269 static int
2270 rw_mgr_mem_calibrate_dq_dqs_centering(const u32 rw_group, const u32 test_bgn,
2271                                       const int use_read_test,
2272                                       const int update_fom)
2273
2274 {
2275         int ret, grp_calibrated;
2276         u32 rank_bgn, sr;
2277
2278         /*
2279          * Altera EMI_RM 2015.05.04 :: Figure 1-28
2280          * Read per-bit deskew can be done on a per shadow register basis.
2281          */
2282         grp_calibrated = 1;
2283         for (rank_bgn = 0, sr = 0;
2284              rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
2285              rank_bgn += NUM_RANKS_PER_SHADOW_REG, sr++) {
2286                 /* Check if this set of ranks should be skipped entirely. */
2287                 if (param->skip_shadow_regs[sr])
2288                         continue;
2289
2290                 ret = rw_mgr_mem_calibrate_vfifo_center(rank_bgn, rw_group,
2291                                                         rw_group, test_bgn,
2292                                                         use_read_test,
2293                                                         update_fom);
2294                 if (ret)
2295                         continue;
2296
2297                 grp_calibrated = 0;
2298         }
2299
2300         if (!grp_calibrated)
2301                 return -EIO;
2302
2303         return 0;
2304 }
2305
2306 /**
2307  * rw_mgr_mem_calibrate_vfifo() - Calibrate the read valid prediction FIFO
2308  * @rw_group:           Read/Write Group
2309  * @test_bgn:           Rank at which the test begins
2310  *
2311  * Stage 1: Calibrate the read valid prediction FIFO.
2312  *
2313  * This function implements UniPHY calibration Stage 1, as explained in
2314  * detail in Altera EMI_RM 2015.05.04 , "UniPHY Calibration Stages".
2315  *
2316  * - read valid prediction will consist of finding:
2317  *   - DQS enable phase and DQS enable delay (DQS Enable Calibration)
2318  *   - DQS input phase  and DQS input delay (DQ/DQS Centering)
2319  *  - we also do a per-bit deskew on the DQ lines.
2320  */
2321 static int rw_mgr_mem_calibrate_vfifo(const u32 rw_group, const u32 test_bgn)
2322 {
2323         uint32_t p, d;
2324         uint32_t dtaps_per_ptap;
2325         uint32_t failed_substage;
2326
2327         int ret;
2328
2329         debug("%s:%d: %u %u\n", __func__, __LINE__, rw_group, test_bgn);
2330
2331         /* Update info for sims */
2332         reg_file_set_group(rw_group);
2333         reg_file_set_stage(CAL_STAGE_VFIFO);
2334         reg_file_set_sub_stage(CAL_SUBSTAGE_GUARANTEED_READ);
2335
2336         failed_substage = CAL_SUBSTAGE_GUARANTEED_READ;
2337
2338         /* USER Determine number of delay taps for each phase tap. */
2339         dtaps_per_ptap = DIV_ROUND_UP(IO_DELAY_PER_OPA_TAP,
2340                                       IO_DELAY_PER_DQS_EN_DCHAIN_TAP) - 1;
2341
2342         for (d = 0; d <= dtaps_per_ptap; d += 2) {
2343                 /*
2344                  * In RLDRAMX we may be messing the delay of pins in
2345                  * the same write rw_group but outside of the current read
2346                  * the rw_group, but that's ok because we haven't calibrated
2347                  * output side yet.
2348                  */
2349                 if (d > 0) {
2350                         scc_mgr_apply_group_all_out_delay_add_all_ranks(
2351                                                                 rw_group, d);
2352                 }
2353
2354                 for (p = 0; p <= IO_DQDQS_OUT_PHASE_MAX; p++) {
2355                         /* 1) Guaranteed Write */
2356                         ret = rw_mgr_mem_calibrate_guaranteed_write(rw_group, p);
2357                         if (ret)
2358                                 break;
2359
2360                         /* 2) DQS Enable Calibration */
2361                         ret = rw_mgr_mem_calibrate_dqs_enable_calibration(rw_group,
2362                                                                           test_bgn);
2363                         if (ret) {
2364                                 failed_substage = CAL_SUBSTAGE_DQS_EN_PHASE;
2365                                 continue;
2366                         }
2367
2368                         /* 3) Centering DQ/DQS */
2369                         /*
2370                          * If doing read after write calibration, do not update
2371                          * FOM now. Do it then.
2372                          */
2373                         ret = rw_mgr_mem_calibrate_dq_dqs_centering(rw_group,
2374                                                                 test_bgn, 1, 0);
2375                         if (ret) {
2376                                 failed_substage = CAL_SUBSTAGE_VFIFO_CENTER;
2377                                 continue;
2378                         }
2379
2380                         /* All done. */
2381                         goto cal_done_ok;
2382                 }
2383         }
2384
2385         /* Calibration Stage 1 failed. */
2386         set_failing_group_stage(rw_group, CAL_STAGE_VFIFO, failed_substage);
2387         return 0;
2388
2389         /* Calibration Stage 1 completed OK. */
2390 cal_done_ok:
2391         /*
2392          * Reset the delay chains back to zero if they have moved > 1
2393          * (check for > 1 because loop will increase d even when pass in
2394          * first case).
2395          */
2396         if (d > 2)
2397                 scc_mgr_zero_group(rw_group, 1);
2398
2399         return 1;
2400 }
2401
2402 /* VFIFO Calibration -- Read Deskew Calibration after write deskew */
2403 static uint32_t rw_mgr_mem_calibrate_vfifo_end(uint32_t read_group,
2404                                                uint32_t test_bgn)
2405 {
2406         uint32_t rank_bgn, sr;
2407         uint32_t grp_calibrated;
2408         uint32_t write_group;
2409
2410         debug("%s:%d %u %u", __func__, __LINE__, read_group, test_bgn);
2411
2412         /* update info for sims */
2413
2414         reg_file_set_stage(CAL_STAGE_VFIFO_AFTER_WRITES);
2415         reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);
2416
2417         write_group = read_group;
2418
2419         /* update info for sims */
2420         reg_file_set_group(read_group);
2421
2422         grp_calibrated = 1;
2423         /* Read per-bit deskew can be done on a per shadow register basis */
2424         for (rank_bgn = 0, sr = 0; rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
2425                 rank_bgn += NUM_RANKS_PER_SHADOW_REG, ++sr) {
2426                 /* Determine if this set of ranks should be skipped entirely */
2427                 if (!param->skip_shadow_regs[sr]) {
2428                 /* This is the last calibration round, update FOM here */
2429                         if (!rw_mgr_mem_calibrate_vfifo_center(rank_bgn,
2430                                                                 write_group,
2431                                                                 read_group,
2432                                                                 test_bgn, 0,
2433                                                                 1)) {
2434                                 grp_calibrated = 0;
2435                         }
2436                 }
2437         }
2438
2439
2440         if (grp_calibrated == 0) {
2441                 set_failing_group_stage(write_group,
2442                                         CAL_STAGE_VFIFO_AFTER_WRITES,
2443                                         CAL_SUBSTAGE_VFIFO_CENTER);
2444                 return 0;
2445         }
2446
2447         return 1;
2448 }
2449
2450 /* Calibrate LFIFO to find smallest read latency */
2451 static uint32_t rw_mgr_mem_calibrate_lfifo(void)
2452 {
2453         uint32_t found_one;
2454
2455         debug("%s:%d\n", __func__, __LINE__);
2456
2457         /* update info for sims */
2458         reg_file_set_stage(CAL_STAGE_LFIFO);
2459         reg_file_set_sub_stage(CAL_SUBSTAGE_READ_LATENCY);
2460
2461         /* Load up the patterns used by read calibration for all ranks */
2462         rw_mgr_mem_calibrate_read_load_patterns(0, 1);
2463         found_one = 0;
2464
2465         do {
2466                 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
2467                 debug_cond(DLEVEL == 2, "%s:%d lfifo: read_lat=%u",
2468                            __func__, __LINE__, gbl->curr_read_lat);
2469
2470                 if (!rw_mgr_mem_calibrate_read_test_all_ranks(0,
2471                                                               NUM_READ_TESTS,
2472                                                               PASS_ALL_BITS,
2473                                                               1)) {
2474                         break;
2475                 }
2476
2477                 found_one = 1;
2478                 /* reduce read latency and see if things are working */
2479                 /* correctly */
2480                 gbl->curr_read_lat--;
2481         } while (gbl->curr_read_lat > 0);
2482
2483         /* reset the fifos to get pointers to known state */
2484
2485         writel(0, &phy_mgr_cmd->fifo_reset);
2486
2487         if (found_one) {
2488                 /* add a fudge factor to the read latency that was determined */
2489                 gbl->curr_read_lat += 2;
2490                 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
2491                 debug_cond(DLEVEL == 2, "%s:%d lfifo: success: using \
2492                            read_lat=%u\n", __func__, __LINE__,
2493                            gbl->curr_read_lat);
2494                 return 1;
2495         } else {
2496                 set_failing_group_stage(0xff, CAL_STAGE_LFIFO,
2497                                         CAL_SUBSTAGE_READ_LATENCY);
2498
2499                 debug_cond(DLEVEL == 2, "%s:%d lfifo: failed at initial \
2500                            read_lat=%u\n", __func__, __LINE__,
2501                            gbl->curr_read_lat);
2502                 return 0;
2503         }
2504 }
2505
2506 /*
2507  * issue write test command.
2508  * two variants are provided. one that just tests a write pattern and
2509  * another that tests datamask functionality.
2510  */
2511 static void rw_mgr_mem_calibrate_write_test_issue(uint32_t group,
2512                                                   uint32_t test_dm)
2513 {
2514         uint32_t mcc_instruction;
2515         uint32_t quick_write_mode = (((STATIC_CALIB_STEPS) & CALIB_SKIP_WRITES) &&
2516                 ENABLE_SUPER_QUICK_CALIBRATION);
2517         uint32_t rw_wl_nop_cycles;
2518         uint32_t addr;
2519
2520         /*
2521          * Set counter and jump addresses for the right
2522          * number of NOP cycles.
2523          * The number of supported NOP cycles can range from -1 to infinity
2524          * Three different cases are handled:
2525          *
2526          * 1. For a number of NOP cycles greater than 0, the RW Mgr looping
2527          *    mechanism will be used to insert the right number of NOPs
2528          *
2529          * 2. For a number of NOP cycles equals to 0, the micro-instruction
2530          *    issuing the write command will jump straight to the
2531          *    micro-instruction that turns on DQS (for DDRx), or outputs write
2532          *    data (for RLD), skipping
2533          *    the NOP micro-instruction all together
2534          *
2535          * 3. A number of NOP cycles equal to -1 indicates that DQS must be
2536          *    turned on in the same micro-instruction that issues the write
2537          *    command. Then we need
2538          *    to directly jump to the micro-instruction that sends out the data
2539          *
2540          * NOTE: Implementing this mechanism uses 2 RW Mgr jump-counters
2541          *       (2 and 3). One jump-counter (0) is used to perform multiple
2542          *       write-read operations.
2543          *       one counter left to issue this command in "multiple-group" mode
2544          */
2545
2546         rw_wl_nop_cycles = gbl->rw_wl_nop_cycles;
2547
2548         if (rw_wl_nop_cycles == -1) {
2549                 /*
2550                  * CNTR 2 - We want to execute the special write operation that
2551                  * turns on DQS right away and then skip directly to the
2552                  * instruction that sends out the data. We set the counter to a
2553                  * large number so that the jump is always taken.
2554                  */
2555                 writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);
2556
2557                 /* CNTR 3 - Not used */
2558                 if (test_dm) {
2559                         mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0_WL_1;
2560                         writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DATA,
2561                                &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2562                         writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
2563                                &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2564                 } else {
2565                         mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0_WL_1;
2566                         writel(RW_MGR_LFSR_WR_RD_BANK_0_DATA,
2567                                 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2568                         writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
2569                                 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2570                 }
2571         } else if (rw_wl_nop_cycles == 0) {
2572                 /*
2573                  * CNTR 2 - We want to skip the NOP operation and go straight
2574                  * to the DQS enable instruction. We set the counter to a large
2575                  * number so that the jump is always taken.
2576                  */
2577                 writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);
2578
2579                 /* CNTR 3 - Not used */
2580                 if (test_dm) {
2581                         mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
2582                         writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DQS,
2583                                &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2584                 } else {
2585                         mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
2586                         writel(RW_MGR_LFSR_WR_RD_BANK_0_DQS,
2587                                 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2588                 }
2589         } else {
2590                 /*
2591                  * CNTR 2 - In this case we want to execute the next instruction
2592                  * and NOT take the jump. So we set the counter to 0. The jump
2593                  * address doesn't count.
2594                  */
2595                 writel(0x0, &sdr_rw_load_mgr_regs->load_cntr2);
2596                 writel(0x0, &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2597
2598                 /*
2599                  * CNTR 3 - Set the nop counter to the number of cycles we
2600                  * need to loop for, minus 1.
2601                  */
2602                 writel(rw_wl_nop_cycles - 1, &sdr_rw_load_mgr_regs->load_cntr3);
2603                 if (test_dm) {
2604                         mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
2605                         writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
2606                                 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2607                 } else {
2608                         mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
2609                         writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
2610                                 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2611                 }
2612         }
2613
2614         writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
2615                   RW_MGR_RESET_READ_DATAPATH_OFFSET);
2616
2617         if (quick_write_mode)
2618                 writel(0x08, &sdr_rw_load_mgr_regs->load_cntr0);
2619         else
2620                 writel(0x40, &sdr_rw_load_mgr_regs->load_cntr0);
2621
2622         writel(mcc_instruction, &sdr_rw_load_jump_mgr_regs->load_jump_add0);
2623
2624         /*
2625          * CNTR 1 - This is used to ensure enough time elapses
2626          * for read data to come back.
2627          */
2628         writel(0x30, &sdr_rw_load_mgr_regs->load_cntr1);
2629
2630         if (test_dm) {
2631                 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_WAIT,
2632                         &sdr_rw_load_jump_mgr_regs->load_jump_add1);
2633         } else {
2634                 writel(RW_MGR_LFSR_WR_RD_BANK_0_WAIT,
2635                         &sdr_rw_load_jump_mgr_regs->load_jump_add1);
2636         }
2637
2638         addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
2639         writel(mcc_instruction, addr + (group << 2));
2640 }
2641
2642 /* Test writes, can check for a single bit pass or multiple bit pass */
2643 static uint32_t rw_mgr_mem_calibrate_write_test(uint32_t rank_bgn,
2644         uint32_t write_group, uint32_t use_dm, uint32_t all_correct,
2645         uint32_t *bit_chk, uint32_t all_ranks)
2646 {
2647         uint32_t r;
2648         uint32_t correct_mask_vg;
2649         uint32_t tmp_bit_chk;
2650         uint32_t vg;
2651         uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
2652                 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
2653         uint32_t addr_rw_mgr;
2654         uint32_t base_rw_mgr;
2655
2656         *bit_chk = param->write_correct_mask;
2657         correct_mask_vg = param->write_correct_mask_vg;
2658
2659         for (r = rank_bgn; r < rank_end; r++) {
2660                 if (param->skip_ranks[r]) {
2661                         /* request to skip the rank */
2662                         continue;
2663                 }
2664
2665                 /* set rank */
2666                 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
2667
2668                 tmp_bit_chk = 0;
2669                 addr_rw_mgr = SDR_PHYGRP_RWMGRGRP_ADDRESS;
2670                 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS-1; ; vg--) {
2671                         /* reset the fifos to get pointers to known state */
2672                         writel(0, &phy_mgr_cmd->fifo_reset);
2673
2674                         tmp_bit_chk = tmp_bit_chk <<
2675                                 (RW_MGR_MEM_DQ_PER_WRITE_DQS /
2676                                 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS);
2677                         rw_mgr_mem_calibrate_write_test_issue(write_group *
2678                                 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS+vg,
2679                                 use_dm);
2680
2681                         base_rw_mgr = readl(addr_rw_mgr);
2682                         tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & ~(base_rw_mgr));
2683                         if (vg == 0)
2684                                 break;
2685                 }
2686                 *bit_chk &= tmp_bit_chk;
2687         }
2688
2689         if (all_correct) {
2690                 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
2691                 debug_cond(DLEVEL == 2, "write_test(%u,%u,ALL) : %u == \
2692                            %u => %lu", write_group, use_dm,
2693                            *bit_chk, param->write_correct_mask,
2694                            (long unsigned int)(*bit_chk ==
2695                            param->write_correct_mask));
2696                 return *bit_chk == param->write_correct_mask;
2697         } else {
2698                 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
2699                 debug_cond(DLEVEL == 2, "write_test(%u,%u,ONE) : %u != ",
2700                        write_group, use_dm, *bit_chk);
2701                 debug_cond(DLEVEL == 2, "%lu" " => %lu", (long unsigned int)0,
2702                         (long unsigned int)(*bit_chk != 0));
2703                 return *bit_chk != 0x00;
2704         }
2705 }
2706
2707 /*
2708  * center all windows. do per-bit-deskew to possibly increase size of
2709  * certain windows.
2710  */
2711 static uint32_t rw_mgr_mem_calibrate_writes_center(uint32_t rank_bgn,
2712         uint32_t write_group, uint32_t test_bgn)
2713 {
2714         uint32_t i, p, min_index;
2715         int32_t d;
2716         /*
2717          * Store these as signed since there are comparisons with
2718          * signed numbers.
2719          */
2720         uint32_t bit_chk;
2721         uint32_t sticky_bit_chk;
2722         int32_t left_edge[RW_MGR_MEM_DQ_PER_WRITE_DQS];
2723         int32_t right_edge[RW_MGR_MEM_DQ_PER_WRITE_DQS];
2724         int32_t mid;
2725         int32_t mid_min, orig_mid_min;
2726         int32_t new_dqs, start_dqs, shift_dq;
2727         int32_t dq_margin, dqs_margin, dm_margin;
2728         uint32_t stop;
2729         uint32_t temp_dq_out1_delay;
2730         uint32_t addr;
2731
2732         debug("%s:%d %u %u", __func__, __LINE__, write_group, test_bgn);
2733
2734         dm_margin = 0;
2735
2736         addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
2737         start_dqs = readl(addr +
2738                           (RW_MGR_MEM_DQ_PER_WRITE_DQS << 2));
2739
2740         /* per-bit deskew */
2741
2742         /*
2743          * set the left and right edge of each bit to an illegal value
2744          * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value.
2745          */
2746         sticky_bit_chk = 0;
2747         for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2748                 left_edge[i]  = IO_IO_OUT1_DELAY_MAX + 1;
2749                 right_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2750         }
2751
2752         /* Search for the left edge of the window for each bit */
2753         for (d = 0; d <= IO_IO_OUT1_DELAY_MAX; d++) {
2754                 scc_mgr_apply_group_dq_out1_delay(write_group, d);
2755
2756                 writel(0, &sdr_scc_mgr->update);
2757
2758                 /*
2759                  * Stop searching when the read test doesn't pass AND when
2760                  * we've seen a passing read on every bit.
2761                  */
2762                 stop = !rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
2763                         0, PASS_ONE_BIT, &bit_chk, 0);
2764                 sticky_bit_chk = sticky_bit_chk | bit_chk;
2765                 stop = stop && (sticky_bit_chk == param->write_correct_mask);
2766                 debug_cond(DLEVEL == 2, "write_center(left): dtap=%d => %u \
2767                            == %u && %u [bit_chk= %u ]\n",
2768                         d, sticky_bit_chk, param->write_correct_mask,
2769                         stop, bit_chk);
2770
2771                 if (stop == 1) {
2772                         break;
2773                 } else {
2774                         for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2775                                 if (bit_chk & 1) {
2776                                         /*
2777                                          * Remember a passing test as the
2778                                          * left_edge.
2779                                          */
2780                                         left_edge[i] = d;
2781                                 } else {
2782                                         /*
2783                                          * If a left edge has not been seen
2784                                          * yet, then a future passing test will
2785                                          * mark this edge as the right edge.
2786                                          */
2787                                         if (left_edge[i] ==
2788                                                 IO_IO_OUT1_DELAY_MAX + 1) {
2789                                                 right_edge[i] = -(d + 1);
2790                                         }
2791                                 }
2792                                 debug_cond(DLEVEL == 2, "write_center[l,d=%d):", d);
2793                                 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d",
2794                                            (int)(bit_chk & 1), i, left_edge[i]);
2795                                 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2796                                        right_edge[i]);
2797                                 bit_chk = bit_chk >> 1;
2798                         }
2799                 }
2800         }
2801
2802         /* Reset DQ delay chains to 0 */
2803         scc_mgr_apply_group_dq_out1_delay(0);
2804         sticky_bit_chk = 0;
2805         for (i = RW_MGR_MEM_DQ_PER_WRITE_DQS - 1;; i--) {
2806                 debug_cond(DLEVEL == 2, "%s:%d write_center: left_edge[%u]: \
2807                            %d right_edge[%u]: %d\n", __func__, __LINE__,
2808                            i, left_edge[i], i, right_edge[i]);
2809
2810                 /*
2811                  * Check for cases where we haven't found the left edge,
2812                  * which makes our assignment of the the right edge invalid.
2813                  * Reset it to the illegal value.
2814                  */
2815                 if ((left_edge[i] == IO_IO_OUT1_DELAY_MAX + 1) &&
2816                     (right_edge[i] != IO_IO_OUT1_DELAY_MAX + 1)) {
2817                         right_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2818                         debug_cond(DLEVEL == 2, "%s:%d write_center: reset \
2819                                    right_edge[%u]: %d\n", __func__, __LINE__,
2820                                    i, right_edge[i]);
2821                 }
2822
2823                 /*
2824                  * Reset sticky bit (except for bits where we have
2825                  * seen the left edge).
2826                  */
2827                 sticky_bit_chk = sticky_bit_chk << 1;
2828                 if ((left_edge[i] != IO_IO_OUT1_DELAY_MAX + 1))
2829                         sticky_bit_chk = sticky_bit_chk | 1;
2830
2831                 if (i == 0)
2832                         break;
2833         }
2834
2835         /* Search for the right edge of the window for each bit */
2836         for (d = 0; d <= IO_IO_OUT1_DELAY_MAX - start_dqs; d++) {
2837                 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group,
2838                                                         d + start_dqs);
2839
2840                 writel(0, &sdr_scc_mgr->update);
2841
2842                 /*
2843                  * Stop searching when the read test doesn't pass AND when
2844                  * we've seen a passing read on every bit.
2845                  */
2846                 stop = !rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
2847                         0, PASS_ONE_BIT, &bit_chk, 0);
2848
2849                 sticky_bit_chk = sticky_bit_chk | bit_chk;
2850                 stop = stop && (sticky_bit_chk == param->write_correct_mask);
2851
2852                 debug_cond(DLEVEL == 2, "write_center (right): dtap=%u => %u == \
2853                            %u && %u\n", d, sticky_bit_chk,
2854                            param->write_correct_mask, stop);
2855
2856                 if (stop == 1) {
2857                         if (d == 0) {
2858                                 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS;
2859                                         i++) {
2860                                         /* d = 0 failed, but it passed when
2861                                         testing the left edge, so it must be
2862                                         marginal, set it to -1 */
2863                                         if (right_edge[i] ==
2864                                                 IO_IO_OUT1_DELAY_MAX + 1 &&
2865                                                 left_edge[i] !=
2866                                                 IO_IO_OUT1_DELAY_MAX + 1) {
2867                                                 right_edge[i] = -1;
2868                                         }
2869                                 }
2870                         }
2871                         break;
2872                 } else {
2873                         for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2874                                 if (bit_chk & 1) {
2875                                         /*
2876                                          * Remember a passing test as
2877                                          * the right_edge.
2878                                          */
2879                                         right_edge[i] = d;
2880                                 } else {
2881                                         if (d != 0) {
2882                                                 /*
2883                                                  * If a right edge has not
2884                                                  * been seen yet, then a future
2885                                                  * passing test will mark this
2886                                                  * edge as the left edge.
2887                                                  */
2888                                                 if (right_edge[i] ==
2889                                                     IO_IO_OUT1_DELAY_MAX + 1)
2890                                                         left_edge[i] = -(d + 1);
2891                                         } else {
2892                                                 /*
2893                                                  * d = 0 failed, but it passed
2894                                                  * when testing the left edge,
2895                                                  * so it must be marginal, set
2896                                                  * it to -1.
2897                                                  */
2898                                                 if (right_edge[i] ==
2899                                                     IO_IO_OUT1_DELAY_MAX + 1 &&
2900                                                     left_edge[i] !=
2901                                                     IO_IO_OUT1_DELAY_MAX + 1)
2902                                                         right_edge[i] = -1;
2903                                                 /*
2904                                                  * If a right edge has not been
2905                                                  * seen yet, then a future
2906                                                  * passing test will mark this
2907                                                  * edge as the left edge.
2908                                                  */
2909                                                 else if (right_edge[i] ==
2910                                                         IO_IO_OUT1_DELAY_MAX +
2911                                                         1)
2912                                                         left_edge[i] = -(d + 1);
2913                                         }
2914                                 }
2915                                 debug_cond(DLEVEL == 2, "write_center[r,d=%d):", d);
2916                                 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d",
2917                                            (int)(bit_chk & 1), i, left_edge[i]);
2918                                 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2919                                            right_edge[i]);
2920                                 bit_chk = bit_chk >> 1;
2921                         }
2922                 }
2923         }
2924
2925         /* Check that all bits have a window */
2926         for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2927                 debug_cond(DLEVEL == 2, "%s:%d write_center: left_edge[%u]: \
2928                            %d right_edge[%u]: %d", __func__, __LINE__,
2929                            i, left_edge[i], i, right_edge[i]);
2930                 if ((left_edge[i] == IO_IO_OUT1_DELAY_MAX + 1) ||
2931                     (right_edge[i] == IO_IO_OUT1_DELAY_MAX + 1)) {
2932                         set_failing_group_stage(test_bgn + i,
2933                                                 CAL_STAGE_WRITES,
2934                                                 CAL_SUBSTAGE_WRITES_CENTER);
2935                         return 0;
2936                 }
2937         }
2938
2939         /* Find middle of window for each DQ bit */
2940         mid_min = left_edge[0] - right_edge[0];
2941         min_index = 0;
2942         for (i = 1; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2943                 mid = left_edge[i] - right_edge[i];
2944                 if (mid < mid_min) {
2945                         mid_min = mid;
2946                         min_index = i;
2947                 }
2948         }
2949
2950         /*
2951          * -mid_min/2 represents the amount that we need to move DQS.
2952          * If mid_min is odd and positive we'll need to add one to
2953          * make sure the rounding in further calculations is correct
2954          * (always bias to the right), so just add 1 for all positive values.
2955          */
2956         if (mid_min > 0)
2957                 mid_min++;
2958         mid_min = mid_min / 2;
2959         debug_cond(DLEVEL == 1, "%s:%d write_center: mid_min=%d\n", __func__,
2960                    __LINE__, mid_min);
2961
2962         /* Determine the amount we can change DQS (which is -mid_min) */
2963         orig_mid_min = mid_min;
2964         new_dqs = start_dqs;
2965         mid_min = 0;
2966         debug_cond(DLEVEL == 1, "%s:%d write_center: start_dqs=%d new_dqs=%d \
2967                    mid_min=%d\n", __func__, __LINE__, start_dqs, new_dqs, mid_min);
2968         /* Initialize data for export structures */
2969         dqs_margin = IO_IO_OUT1_DELAY_MAX + 1;
2970         dq_margin  = IO_IO_OUT1_DELAY_MAX + 1;
2971
2972         /* add delay to bring centre of all DQ windows to the same "level" */
2973         for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++, p++) {
2974                 /* Use values before divide by 2 to reduce round off error */
2975                 shift_dq = (left_edge[i] - right_edge[i] -
2976                         (left_edge[min_index] - right_edge[min_index]))/2  +
2977                 (orig_mid_min - mid_min);
2978
2979                 debug_cond(DLEVEL == 2, "%s:%d write_center: before: shift_dq \
2980                            [%u]=%d\n", __func__, __LINE__, i, shift_dq);
2981
2982                 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
2983                 temp_dq_out1_delay = readl(addr + (i << 2));
2984                 if (shift_dq + (int32_t)temp_dq_out1_delay >
2985                         (int32_t)IO_IO_OUT1_DELAY_MAX) {
2986                         shift_dq = (int32_t)IO_IO_OUT1_DELAY_MAX - temp_dq_out1_delay;
2987                 } else if (shift_dq + (int32_t)temp_dq_out1_delay < 0) {
2988                         shift_dq = -(int32_t)temp_dq_out1_delay;
2989                 }
2990                 debug_cond(DLEVEL == 2, "write_center: after: shift_dq[%u]=%d\n",
2991                            i, shift_dq);
2992                 scc_mgr_set_dq_out1_delay(i, temp_dq_out1_delay + shift_dq);
2993                 scc_mgr_load_dq(i);
2994
2995                 debug_cond(DLEVEL == 2, "write_center: margin[%u]=[%d,%d]\n", i,
2996                            left_edge[i] - shift_dq + (-mid_min),
2997                            right_edge[i] + shift_dq - (-mid_min));
2998                 /* To determine values for export structures */
2999                 if (left_edge[i] - shift_dq + (-mid_min) < dq_margin)
3000                         dq_margin = left_edge[i] - shift_dq + (-mid_min);
3001
3002                 if (right_edge[i] + shift_dq - (-mid_min) < dqs_margin)
3003                         dqs_margin = right_edge[i] + shift_dq - (-mid_min);
3004         }
3005
3006         /* Move DQS */
3007         scc_mgr_apply_group_dqs_io_and_oct_out1(write_group, new_dqs);
3008         writel(0, &sdr_scc_mgr->update);
3009
3010         /* Centre DM */
3011         debug_cond(DLEVEL == 2, "%s:%d write_center: DM\n", __func__, __LINE__);
3012
3013         /*
3014          * set the left and right edge of each bit to an illegal value,
3015          * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value,
3016          */
3017         left_edge[0]  = IO_IO_OUT1_DELAY_MAX + 1;
3018         right_edge[0] = IO_IO_OUT1_DELAY_MAX + 1;
3019         int32_t bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3020         int32_t end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3021         int32_t bgn_best = IO_IO_OUT1_DELAY_MAX + 1;
3022         int32_t end_best = IO_IO_OUT1_DELAY_MAX + 1;
3023         int32_t win_best = 0;
3024
3025         /* Search for the/part of the window with DM shift */
3026         for (d = IO_IO_OUT1_DELAY_MAX; d >= 0; d -= DELTA_D) {
3027                 scc_mgr_apply_group_dm_out1_delay(d);
3028                 writel(0, &sdr_scc_mgr->update);
3029
3030                 if (rw_mgr_mem_calibrate_write_test(rank_bgn, write_group, 1,
3031                                                     PASS_ALL_BITS, &bit_chk,
3032                                                     0)) {
3033                         /* USE Set current end of the window */
3034                         end_curr = -d;
3035                         /*
3036                          * If a starting edge of our window has not been seen
3037                          * this is our current start of the DM window.
3038                          */
3039                         if (bgn_curr == IO_IO_OUT1_DELAY_MAX + 1)
3040                                 bgn_curr = -d;
3041
3042                         /*
3043                          * If current window is bigger than best seen.
3044                          * Set best seen to be current window.
3045                          */
3046                         if ((end_curr-bgn_curr+1) > win_best) {
3047                                 win_best = end_curr-bgn_curr+1;
3048                                 bgn_best = bgn_curr;
3049                                 end_best = end_curr;
3050                         }
3051                 } else {
3052                         /* We just saw a failing test. Reset temp edge */
3053                         bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3054                         end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3055                         }
3056                 }
3057
3058
3059         /* Reset DM delay chains to 0 */
3060         scc_mgr_apply_group_dm_out1_delay(0);
3061
3062         /*
3063          * Check to see if the current window nudges up aganist 0 delay.
3064          * If so we need to continue the search by shifting DQS otherwise DQS
3065          * search begins as a new search. */
3066         if (end_curr != 0) {
3067                 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3068                 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3069         }
3070
3071         /* Search for the/part of the window with DQS shifts */
3072         for (d = 0; d <= IO_IO_OUT1_DELAY_MAX - new_dqs; d += DELTA_D) {
3073                 /*
3074                  * Note: This only shifts DQS, so are we limiting ourselve to
3075                  * width of DQ unnecessarily.
3076                  */
3077                 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group,
3078                                                         d + new_dqs);
3079
3080                 writel(0, &sdr_scc_mgr->update);
3081                 if (rw_mgr_mem_calibrate_write_test(rank_bgn, write_group, 1,
3082                                                     PASS_ALL_BITS, &bit_chk,
3083                                                     0)) {
3084                         /* USE Set current end of the window */
3085                         end_curr = d;
3086                         /*
3087                          * If a beginning edge of our window has not been seen
3088                          * this is our current begin of the DM window.
3089                          */
3090                         if (bgn_curr == IO_IO_OUT1_DELAY_MAX + 1)
3091                                 bgn_curr = d;
3092
3093                         /*
3094                          * If current window is bigger than best seen. Set best
3095                          * seen to be current window.
3096                          */
3097                         if ((end_curr-bgn_curr+1) > win_best) {
3098                                 win_best = end_curr-bgn_curr+1;
3099                                 bgn_best = bgn_curr;
3100                                 end_best = end_curr;
3101                         }
3102                 } else {
3103                         /* We just saw a failing test. Reset temp edge */
3104                         bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3105                         end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3106
3107                         /* Early exit optimization: if ther remaining delay
3108                         chain space is less than already seen largest window
3109                         we can exit */
3110                         if ((win_best-1) >
3111                                 (IO_IO_OUT1_DELAY_MAX - new_dqs - d)) {
3112                                         break;
3113                                 }
3114                         }
3115                 }
3116
3117         /* assign left and right edge for cal and reporting; */
3118         left_edge[0] = -1*bgn_best;
3119         right_edge[0] = end_best;
3120
3121         debug_cond(DLEVEL == 2, "%s:%d dm_calib: left=%d right=%d\n", __func__,
3122                    __LINE__, left_edge[0], right_edge[0]);
3123
3124         /* Move DQS (back to orig) */
3125         scc_mgr_apply_group_dqs_io_and_oct_out1(write_group, new_dqs);
3126
3127         /* Move DM */
3128
3129         /* Find middle of window for the DM bit */
3130         mid = (left_edge[0] - right_edge[0]) / 2;
3131
3132         /* only move right, since we are not moving DQS/DQ */
3133         if (mid < 0)
3134                 mid = 0;
3135
3136         /* dm_marign should fail if we never find a window */
3137         if (win_best == 0)
3138                 dm_margin = -1;
3139         else
3140                 dm_margin = left_edge[0] - mid;
3141
3142         scc_mgr_apply_group_dm_out1_delay(mid);
3143         writel(0, &sdr_scc_mgr->update);
3144
3145         debug_cond(DLEVEL == 2, "%s:%d dm_calib: left=%d right=%d mid=%d \
3146                    dm_margin=%d\n", __func__, __LINE__, left_edge[0],
3147                    right_edge[0], mid, dm_margin);
3148         /* Export values */
3149         gbl->fom_out += dq_margin + dqs_margin;
3150
3151         debug_cond(DLEVEL == 2, "%s:%d write_center: dq_margin=%d \
3152                    dqs_margin=%d dm_margin=%d\n", __func__, __LINE__,
3153                    dq_margin, dqs_margin, dm_margin);
3154
3155         /*
3156          * Do not remove this line as it makes sure all of our
3157          * decisions have been applied.
3158          */
3159         writel(0, &sdr_scc_mgr->update);
3160         return (dq_margin >= 0) && (dqs_margin >= 0) && (dm_margin >= 0);
3161 }
3162
3163 /**
3164  * rw_mgr_mem_calibrate_writes() - Write Calibration Part One
3165  * @rank_bgn:           Rank number
3166  * @group:              Read/Write Group
3167  * @test_bgn:           Rank at which the test begins
3168  *
3169  * Stage 2: Write Calibration Part One.
3170  *
3171  * This function implements UniPHY calibration Stage 2, as explained in
3172  * detail in Altera EMI_RM 2015.05.04 , "UniPHY Calibration Stages".
3173  */
3174 static int rw_mgr_mem_calibrate_writes(const u32 rank_bgn, const u32 group,
3175                                        const u32 test_bgn)
3176 {
3177         int ret;
3178
3179         /* Update info for sims */
3180         debug("%s:%d %u %u\n", __func__, __LINE__, group, test_bgn);
3181
3182         reg_file_set_group(group);
3183         reg_file_set_stage(CAL_STAGE_WRITES);
3184         reg_file_set_sub_stage(CAL_SUBSTAGE_WRITES_CENTER);
3185
3186         ret = rw_mgr_mem_calibrate_writes_center(rank_bgn, group, test_bgn);
3187         if (!ret) {
3188                 set_failing_group_stage(group, CAL_STAGE_WRITES,
3189                                         CAL_SUBSTAGE_WRITES_CENTER);
3190                 return -EIO;
3191         }
3192
3193         return 0;
3194 }
3195
3196 /**
3197  * mem_precharge_and_activate() - Precharge all banks and activate
3198  *
3199  * Precharge all banks and activate row 0 in bank "000..." and bank "111...".
3200  */
3201 static void mem_precharge_and_activate(void)
3202 {
3203         int r;
3204
3205         for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
3206                 /* Test if the rank should be skipped. */
3207                 if (param->skip_ranks[r])
3208                         continue;
3209
3210                 /* Set rank. */
3211                 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);
3212
3213                 /* Precharge all banks. */
3214                 writel(RW_MGR_PRECHARGE_ALL, SDR_PHYGRP_RWMGRGRP_ADDRESS |
3215                                              RW_MGR_RUN_SINGLE_GROUP_OFFSET);
3216
3217                 writel(0x0F, &sdr_rw_load_mgr_regs->load_cntr0);
3218                 writel(RW_MGR_ACTIVATE_0_AND_1_WAIT1,
3219                         &sdr_rw_load_jump_mgr_regs->load_jump_add0);
3220
3221                 writel(0x0F, &sdr_rw_load_mgr_regs->load_cntr1);
3222                 writel(RW_MGR_ACTIVATE_0_AND_1_WAIT2,
3223                         &sdr_rw_load_jump_mgr_regs->load_jump_add1);
3224
3225                 /* Activate rows. */
3226                 writel(RW_MGR_ACTIVATE_0_AND_1, SDR_PHYGRP_RWMGRGRP_ADDRESS |
3227                                                 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
3228         }
3229 }
3230
3231 /**
3232  * mem_init_latency() - Configure memory RLAT and WLAT settings
3233  *
3234  * Configure memory RLAT and WLAT parameters.
3235  */
3236 static void mem_init_latency(void)
3237 {
3238         /*
3239          * For AV/CV, LFIFO is hardened and always runs at full rate
3240          * so max latency in AFI clocks, used here, is correspondingly
3241          * smaller.
3242          */
3243         const u32 max_latency = (1 << MAX_LATENCY_COUNT_WIDTH) - 1;
3244         u32 rlat, wlat;
3245
3246         debug("%s:%d\n", __func__, __LINE__);
3247
3248         /*
3249          * Read in write latency.
3250          * WL for Hard PHY does not include additive latency.
3251          */
3252         wlat = readl(&data_mgr->t_wl_add);
3253         wlat += readl(&data_mgr->mem_t_add);
3254
3255         gbl->rw_wl_nop_cycles = wlat - 1;
3256
3257         /* Read in readl latency. */
3258         rlat = readl(&data_mgr->t_rl_add);
3259
3260         /* Set a pretty high read latency initially. */
3261         gbl->curr_read_lat = rlat + 16;
3262         if (gbl->curr_read_lat > max_latency)
3263                 gbl->curr_read_lat = max_latency;
3264
3265         writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
3266
3267         /* Advertise write latency. */
3268         writel(wlat, &phy_mgr_cfg->afi_wlat);
3269 }
3270
3271 /**
3272  * @mem_skip_calibrate() - Set VFIFO and LFIFO to instant-on settings
3273  *
3274  * Set VFIFO and LFIFO to instant-on settings in skip calibration mode.
3275  */
3276 static void mem_skip_calibrate(void)
3277 {
3278         uint32_t vfifo_offset;
3279         uint32_t i, j, r;
3280
3281         debug("%s:%d\n", __func__, __LINE__);
3282         /* Need to update every shadow register set used by the interface */
3283         for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
3284              r += NUM_RANKS_PER_SHADOW_REG) {
3285                 /*
3286                  * Set output phase alignment settings appropriate for
3287                  * skip calibration.
3288                  */
3289                 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3290                         scc_mgr_set_dqs_en_phase(i, 0);
3291 #if IO_DLL_CHAIN_LENGTH == 6
3292                         scc_mgr_set_dqdqs_output_phase(i, 6);
3293 #else
3294                         scc_mgr_set_dqdqs_output_phase(i, 7);
3295 #endif
3296                         /*
3297                          * Case:33398
3298                          *
3299                          * Write data arrives to the I/O two cycles before write
3300                          * latency is reached (720 deg).
3301                          *   -> due to bit-slip in a/c bus
3302                          *   -> to allow board skew where dqs is longer than ck
3303                          *      -> how often can this happen!?
3304                          *      -> can claim back some ptaps for high freq
3305                          *       support if we can relax this, but i digress...
3306                          *
3307                          * The write_clk leads mem_ck by 90 deg
3308                          * The minimum ptap of the OPA is 180 deg
3309                          * Each ptap has (360 / IO_DLL_CHAIN_LENGH) deg of delay
3310                          * The write_clk is always delayed by 2 ptaps
3311                          *
3312                          * Hence, to make DQS aligned to CK, we need to delay
3313                          * DQS by:
3314                          *    (720 - 90 - 180 - 2 * (360 / IO_DLL_CHAIN_LENGTH))
3315                          *
3316                          * Dividing the above by (360 / IO_DLL_CHAIN_LENGTH)
3317                          * gives us the number of ptaps, which simplies to:
3318                          *
3319                          *    (1.25 * IO_DLL_CHAIN_LENGTH - 2)
3320                          */
3321                         scc_mgr_set_dqdqs_output_phase(i,
3322                                         1.25 * IO_DLL_CHAIN_LENGTH - 2);
3323                 }
3324                 writel(0xff, &sdr_scc_mgr->dqs_ena);
3325                 writel(0xff, &sdr_scc_mgr->dqs_io_ena);
3326
3327                 for (i = 0; i < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; i++) {
3328                         writel(i, SDR_PHYGRP_SCCGRP_ADDRESS |
3329                                   SCC_MGR_GROUP_COUNTER_OFFSET);
3330                 }
3331                 writel(0xff, &sdr_scc_mgr->dq_ena);
3332                 writel(0xff, &sdr_scc_mgr->dm_ena);
3333                 writel(0, &sdr_scc_mgr->update);
3334         }
3335
3336         /* Compensate for simulation model behaviour */
3337         for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3338                 scc_mgr_set_dqs_bus_in_delay(i, 10);
3339                 scc_mgr_load_dqs(i);
3340         }
3341         writel(0, &sdr_scc_mgr->update);
3342
3343         /*
3344          * ArriaV has hard FIFOs that can only be initialized by incrementing
3345          * in sequencer.
3346          */
3347         vfifo_offset = CALIB_VFIFO_OFFSET;
3348         for (j = 0; j < vfifo_offset; j++)
3349                 writel(0xff, &phy_mgr_cmd->inc_vfifo_hard_phy);
3350         writel(0, &phy_mgr_cmd->fifo_reset);
3351
3352         /*
3353          * For Arria V and Cyclone V with hard LFIFO, we get the skip-cal
3354          * setting from generation-time constant.
3355          */
3356         gbl->curr_read_lat = CALIB_LFIFO_OFFSET;
3357         writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
3358 }
3359
3360 /**
3361  * mem_calibrate() - Memory calibration entry point.
3362  *
3363  * Perform memory calibration.
3364  */
3365 static uint32_t mem_calibrate(void)
3366 {
3367         uint32_t i;
3368         uint32_t rank_bgn, sr;
3369         uint32_t write_group, write_test_bgn;
3370         uint32_t read_group, read_test_bgn;
3371         uint32_t run_groups, current_run;
3372         uint32_t failing_groups = 0;
3373         uint32_t group_failed = 0;
3374
3375         const u32 rwdqs_ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
3376                                 RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
3377
3378         debug("%s:%d\n", __func__, __LINE__);
3379
3380         /* Initialize the data settings */
3381         gbl->error_substage = CAL_SUBSTAGE_NIL;
3382         gbl->error_stage = CAL_STAGE_NIL;
3383         gbl->error_group = 0xff;
3384         gbl->fom_in = 0;
3385         gbl->fom_out = 0;
3386
3387         /* Initialize WLAT and RLAT. */
3388         mem_init_latency();
3389
3390         /* Initialize bit slips. */
3391         mem_precharge_and_activate();
3392
3393         for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3394                 writel(i, SDR_PHYGRP_SCCGRP_ADDRESS |
3395                           SCC_MGR_GROUP_COUNTER_OFFSET);
3396                 /* Only needed once to set all groups, pins, DQ, DQS, DM. */
3397                 if (i == 0)
3398                         scc_mgr_set_hhp_extras();
3399
3400                 scc_set_bypass_mode(i);
3401         }
3402
3403         /* Calibration is skipped. */
3404         if ((dyn_calib_steps & CALIB_SKIP_ALL) == CALIB_SKIP_ALL) {
3405                 /*
3406                  * Set VFIFO and LFIFO to instant-on settings in skip
3407                  * calibration mode.
3408                  */
3409                 mem_skip_calibrate();
3410
3411                 /*
3412                  * Do not remove this line as it makes sure all of our
3413                  * decisions have been applied.
3414                  */
3415                 writel(0, &sdr_scc_mgr->update);
3416                 return 1;
3417         }
3418
3419         /* Calibration is not skipped. */
3420         for (i = 0; i < NUM_CALIB_REPEAT; i++) {
3421                 /*
3422                  * Zero all delay chain/phase settings for all
3423                  * groups and all shadow register sets.
3424                  */
3425                 scc_mgr_zero_all();
3426
3427                 run_groups = ~param->skip_groups;
3428
3429                 for (write_group = 0, write_test_bgn = 0; write_group
3430                         < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; write_group++,
3431                         write_test_bgn += RW_MGR_MEM_DQ_PER_WRITE_DQS) {
3432
3433                         /* Initialize the group failure */
3434                         group_failed = 0;
3435
3436                         current_run = run_groups & ((1 <<
3437                                 RW_MGR_NUM_DQS_PER_WRITE_GROUP) - 1);
3438                         run_groups = run_groups >>
3439                                 RW_MGR_NUM_DQS_PER_WRITE_GROUP;
3440
3441                         if (current_run == 0)
3442                                 continue;
3443
3444                         writel(write_group, SDR_PHYGRP_SCCGRP_ADDRESS |
3445                                             SCC_MGR_GROUP_COUNTER_OFFSET);
3446                         scc_mgr_zero_group(write_group, 0);
3447
3448                         for (read_group = write_group * rwdqs_ratio,
3449                              read_test_bgn = 0;
3450                              read_group < (write_group + 1) * rwdqs_ratio;
3451                              read_group++,
3452                              read_test_bgn += RW_MGR_MEM_DQ_PER_READ_DQS) {
3453                                 if (STATIC_CALIB_STEPS & CALIB_SKIP_VFIFO)
3454                                         continue;
3455
3456                                 /* Calibrate the VFIFO */
3457                                 if (rw_mgr_mem_calibrate_vfifo(read_group,
3458                                                                read_test_bgn))
3459                                         continue;
3460
3461                                 if (!(gbl->phy_debug_mode_flags & PHY_DEBUG_SWEEP_ALL_GROUPS))
3462                                         return 0;
3463
3464                                 /* The group failed, we're done. */
3465                                 goto grp_failed;
3466                         }
3467
3468                         /* Calibrate the output side */
3469                         for (rank_bgn = 0, sr = 0;
3470                              rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
3471                              rank_bgn += NUM_RANKS_PER_SHADOW_REG, sr++) {
3472                                 if (STATIC_CALIB_STEPS & CALIB_SKIP_WRITES)
3473                                         continue;
3474
3475                                 /* Not needed in quick mode! */
3476                                 if (STATIC_CALIB_STEPS & CALIB_SKIP_DELAY_SWEEPS)
3477                                         continue;
3478
3479                                 /*
3480                                  * Determine if this set of ranks
3481                                  * should be skipped entirely.
3482                                  */
3483                                 if (param->skip_shadow_regs[sr])
3484                                         continue;
3485
3486                                 /* Calibrate WRITEs */
3487                                 if (!rw_mgr_mem_calibrate_writes(rank_bgn,
3488                                                 write_group, write_test_bgn))
3489                                         continue;
3490
3491                                 group_failed = 1;
3492                                 if (!(gbl->phy_debug_mode_flags & PHY_DEBUG_SWEEP_ALL_GROUPS))
3493                                         return 0;
3494                         }
3495
3496                         /* Some group failed, we're done. */
3497                         if (group_failed)
3498                                 goto grp_failed;
3499
3500                         for (read_group = write_group * rwdqs_ratio,
3501                              read_test_bgn = 0;
3502                              read_group < (write_group + 1) * rwdqs_ratio;
3503                              read_group++,
3504                              read_test_bgn += RW_MGR_MEM_DQ_PER_READ_DQS) {
3505                                 if (STATIC_CALIB_STEPS & CALIB_SKIP_WRITES)
3506                                         continue;
3507
3508                                 if (rw_mgr_mem_calibrate_vfifo_end(read_group,
3509                                                                 read_test_bgn))
3510                                         continue;
3511
3512                                 if (!(gbl->phy_debug_mode_flags & PHY_DEBUG_SWEEP_ALL_GROUPS))
3513                                         return 0;
3514
3515                                 /* The group failed, we're done. */
3516                                 goto grp_failed;
3517                         }
3518
3519                         /* No group failed, continue as usual. */
3520                         continue;
3521
3522 grp_failed:             /* A group failed, increment the counter. */
3523                         failing_groups++;
3524                 }
3525
3526                 /*
3527                  * USER If there are any failing groups then report
3528                  * the failure.
3529                  */
3530                 if (failing_groups != 0)
3531                         return 0;
3532
3533                 if (STATIC_CALIB_STEPS & CALIB_SKIP_LFIFO)
3534                         continue;
3535
3536                 /*
3537                  * If we're skipping groups as part of debug,
3538                  * don't calibrate LFIFO.
3539                  */
3540                 if (param->skip_groups != 0)
3541                         continue;
3542
3543                 /* Calibrate the LFIFO */
3544                 if (!rw_mgr_mem_calibrate_lfifo())
3545                         return 0;
3546         }
3547
3548         /*
3549          * Do not remove this line as it makes sure all of our decisions
3550          * have been applied.
3551          */
3552         writel(0, &sdr_scc_mgr->update);
3553         return 1;
3554 }
3555
3556 /**
3557  * run_mem_calibrate() - Perform memory calibration
3558  *
3559  * This function triggers the entire memory calibration procedure.
3560  */
3561 static int run_mem_calibrate(void)
3562 {
3563         int pass;
3564
3565         debug("%s:%d\n", __func__, __LINE__);
3566
3567         /* Reset pass/fail status shown on afi_cal_success/fail */
3568         writel(PHY_MGR_CAL_RESET, &phy_mgr_cfg->cal_status);
3569
3570         /* Stop tracking manager. */
3571         clrbits_le32(&sdr_ctrl->ctrl_cfg, 1 << 22);
3572
3573         phy_mgr_initialize();
3574         rw_mgr_mem_initialize();
3575
3576         /* Perform the actual memory calibration. */
3577         pass = mem_calibrate();
3578
3579         mem_precharge_and_activate();
3580         writel(0, &phy_mgr_cmd->fifo_reset);
3581
3582         /* Handoff. */
3583         rw_mgr_mem_handoff();
3584         /*
3585          * In Hard PHY this is a 2-bit control:
3586          * 0: AFI Mux Select
3587          * 1: DDIO Mux Select
3588          */
3589         writel(0x2, &phy_mgr_cfg->mux_sel);
3590
3591         /* Start tracking manager. */
3592         setbits_le32(&sdr_ctrl->ctrl_cfg, 1 << 22);
3593
3594         return pass;
3595 }
3596
3597 /**
3598  * debug_mem_calibrate() - Report result of memory calibration
3599  * @pass:       Value indicating whether calibration passed or failed
3600  *
3601  * This function reports the results of the memory calibration
3602  * and writes debug information into the register file.
3603  */
3604 static void debug_mem_calibrate(int pass)
3605 {
3606         uint32_t debug_info;
3607
3608         if (pass) {
3609                 printf("%s: CALIBRATION PASSED\n", __FILE__);
3610
3611                 gbl->fom_in /= 2;
3612                 gbl->fom_out /= 2;
3613
3614                 if (gbl->fom_in > 0xff)
3615                         gbl->fom_in = 0xff;
3616
3617                 if (gbl->fom_out > 0xff)
3618                         gbl->fom_out = 0xff;
3619
3620                 /* Update the FOM in the register file */
3621                 debug_info = gbl->fom_in;
3622                 debug_info |= gbl->fom_out << 8;
3623                 writel(debug_info, &sdr_reg_file->fom);
3624
3625                 writel(debug_info, &phy_mgr_cfg->cal_debug_info);
3626                 writel(PHY_MGR_CAL_SUCCESS, &phy_mgr_cfg->cal_status);
3627         } else {
3628                 printf("%s: CALIBRATION FAILED\n", __FILE__);
3629
3630                 debug_info = gbl->error_stage;
3631                 debug_info |= gbl->error_substage << 8;
3632                 debug_info |= gbl->error_group << 16;
3633
3634                 writel(debug_info, &sdr_reg_file->failing_stage);
3635                 writel(debug_info, &phy_mgr_cfg->cal_debug_info);
3636                 writel(PHY_MGR_CAL_FAIL, &phy_mgr_cfg->cal_status);
3637
3638                 /* Update the failing group/stage in the register file */
3639                 debug_info = gbl->error_stage;
3640                 debug_info |= gbl->error_substage << 8;
3641                 debug_info |= gbl->error_group << 16;
3642                 writel(debug_info, &sdr_reg_file->failing_stage);
3643         }
3644
3645         printf("%s: Calibration complete\n", __FILE__);
3646 }
3647
3648 /**
3649  * hc_initialize_rom_data() - Initialize ROM data
3650  *
3651  * Initialize ROM data.
3652  */
3653 static void hc_initialize_rom_data(void)
3654 {
3655         u32 i, addr;
3656
3657         addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_INST_ROM_WRITE_OFFSET;
3658         for (i = 0; i < ARRAY_SIZE(inst_rom_init); i++)
3659                 writel(inst_rom_init[i], addr + (i << 2));
3660
3661         addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_AC_ROM_WRITE_OFFSET;
3662         for (i = 0; i < ARRAY_SIZE(ac_rom_init); i++)
3663                 writel(ac_rom_init[i], addr + (i << 2));
3664 }
3665
3666 /**
3667  * initialize_reg_file() - Initialize SDR register file
3668  *
3669  * Initialize SDR register file.
3670  */
3671 static void initialize_reg_file(void)
3672 {
3673         /* Initialize the register file with the correct data */
3674         writel(REG_FILE_INIT_SEQ_SIGNATURE, &sdr_reg_file->signature);
3675         writel(0, &sdr_reg_file->debug_data_addr);
3676         writel(0, &sdr_reg_file->cur_stage);
3677         writel(0, &sdr_reg_file->fom);
3678         writel(0, &sdr_reg_file->failing_stage);
3679         writel(0, &sdr_reg_file->debug1);
3680         writel(0, &sdr_reg_file->debug2);
3681 }
3682
3683 /**
3684  * initialize_hps_phy() - Initialize HPS PHY
3685  *
3686  * Initialize HPS PHY.
3687  */
3688 static void initialize_hps_phy(void)
3689 {
3690         uint32_t reg;
3691         /*
3692          * Tracking also gets configured here because it's in the
3693          * same register.
3694          */
3695         uint32_t trk_sample_count = 7500;
3696         uint32_t trk_long_idle_sample_count = (10 << 16) | 100;
3697         /*
3698          * Format is number of outer loops in the 16 MSB, sample
3699          * count in 16 LSB.
3700          */
3701
3702         reg = 0;
3703         reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_ACDELAYEN_SET(2);
3704         reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQDELAYEN_SET(1);
3705         reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQSDELAYEN_SET(1);
3706         reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQSLOGICDELAYEN_SET(1);
3707         reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_RESETDELAYEN_SET(0);
3708         reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_LPDDRDIS_SET(1);
3709         /*
3710          * This field selects the intrinsic latency to RDATA_EN/FULL path.
3711          * 00-bypass, 01- add 5 cycles, 10- add 10 cycles, 11- add 15 cycles.
3712          */
3713         reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_ADDLATSEL_SET(0);
3714         reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_SAMPLECOUNT_19_0_SET(
3715                 trk_sample_count);
3716         writel(reg, &sdr_ctrl->phy_ctrl0);
3717
3718         reg = 0;
3719         reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_SAMPLECOUNT_31_20_SET(
3720                 trk_sample_count >>
3721                 SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_SAMPLECOUNT_19_0_WIDTH);
3722         reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_LONGIDLESAMPLECOUNT_19_0_SET(
3723                 trk_long_idle_sample_count);
3724         writel(reg, &sdr_ctrl->phy_ctrl1);
3725
3726         reg = 0;
3727         reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_2_LONGIDLESAMPLECOUNT_31_20_SET(
3728                 trk_long_idle_sample_count >>
3729                 SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_LONGIDLESAMPLECOUNT_19_0_WIDTH);
3730         writel(reg, &sdr_ctrl->phy_ctrl2);
3731 }
3732
3733 /**
3734  * initialize_tracking() - Initialize tracking
3735  *
3736  * Initialize the register file with usable initial data.
3737  */
3738 static void initialize_tracking(void)
3739 {
3740         /*
3741          * Initialize the register file with the correct data.
3742          * Compute usable version of value in case we skip full
3743          * computation later.
3744          */
3745         writel(DIV_ROUND_UP(IO_DELAY_PER_OPA_TAP, IO_DELAY_PER_DCHAIN_TAP) - 1,
3746                &sdr_reg_file->dtaps_per_ptap);
3747
3748         /* trk_sample_count */
3749         writel(7500, &sdr_reg_file->trk_sample_count);
3750
3751         /* longidle outer loop [15:0] */
3752         writel((10 << 16) | (100 << 0), &sdr_reg_file->trk_longidle);
3753
3754         /*
3755          * longidle sample count [31:24]
3756          * trfc, worst case of 933Mhz 4Gb [23:16]
3757          * trcd, worst case [15:8]
3758          * vfifo wait [7:0]
3759          */
3760         writel((243 << 24) | (14 << 16) | (10 << 8) | (4 << 0),
3761                &sdr_reg_file->delays);
3762
3763         /* mux delay */
3764         writel((RW_MGR_IDLE << 24) | (RW_MGR_ACTIVATE_1 << 16) |
3765                (RW_MGR_SGLE_READ << 8) | (RW_MGR_PRECHARGE_ALL << 0),
3766                &sdr_reg_file->trk_rw_mgr_addr);
3767
3768         writel(RW_MGR_MEM_IF_READ_DQS_WIDTH,
3769                &sdr_reg_file->trk_read_dqs_width);
3770
3771         /* trefi [7:0] */
3772         writel((RW_MGR_REFRESH_ALL << 24) | (1000 << 0),
3773                &sdr_reg_file->trk_rfsh);
3774 }
3775
3776 int sdram_calibration_full(void)
3777 {
3778         struct param_type my_param;
3779         struct gbl_type my_gbl;
3780         uint32_t pass;
3781
3782         memset(&my_param, 0, sizeof(my_param));
3783         memset(&my_gbl, 0, sizeof(my_gbl));
3784
3785         param = &my_param;
3786         gbl = &my_gbl;
3787
3788         /* Set the calibration enabled by default */
3789         gbl->phy_debug_mode_flags |= PHY_DEBUG_ENABLE_CAL_RPT;
3790         /*
3791          * Only sweep all groups (regardless of fail state) by default
3792          * Set enabled read test by default.
3793          */
3794 #if DISABLE_GUARANTEED_READ
3795         gbl->phy_debug_mode_flags |= PHY_DEBUG_DISABLE_GUARANTEED_READ;
3796 #endif
3797         /* Initialize the register file */
3798         initialize_reg_file();
3799
3800         /* Initialize any PHY CSR */
3801         initialize_hps_phy();
3802
3803         scc_mgr_initialize();
3804
3805         initialize_tracking();
3806
3807         printf("%s: Preparing to start memory calibration\n", __FILE__);
3808
3809         debug("%s:%d\n", __func__, __LINE__);
3810         debug_cond(DLEVEL == 1,
3811                    "DDR3 FULL_RATE ranks=%u cs/dimm=%u dq/dqs=%u,%u vg/dqs=%u,%u ",
3812                    RW_MGR_MEM_NUMBER_OF_RANKS, RW_MGR_MEM_NUMBER_OF_CS_PER_DIMM,
3813                    RW_MGR_MEM_DQ_PER_READ_DQS, RW_MGR_MEM_DQ_PER_WRITE_DQS,
3814                    RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS,
3815                    RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS);
3816         debug_cond(DLEVEL == 1,
3817                    "dqs=%u,%u dq=%u dm=%u ptap_delay=%u dtap_delay=%u ",
3818                    RW_MGR_MEM_IF_READ_DQS_WIDTH, RW_MGR_MEM_IF_WRITE_DQS_WIDTH,
3819                    RW_MGR_MEM_DATA_WIDTH, RW_MGR_MEM_DATA_MASK_WIDTH,
3820                    IO_DELAY_PER_OPA_TAP, IO_DELAY_PER_DCHAIN_TAP);
3821         debug_cond(DLEVEL == 1, "dtap_dqsen_delay=%u, dll=%u",
3822                    IO_DELAY_PER_DQS_EN_DCHAIN_TAP, IO_DLL_CHAIN_LENGTH);
3823         debug_cond(DLEVEL == 1, "max values: en_p=%u dqdqs_p=%u en_d=%u dqs_in_d=%u ",
3824                    IO_DQS_EN_PHASE_MAX, IO_DQDQS_OUT_PHASE_MAX,
3825                    IO_DQS_EN_DELAY_MAX, IO_DQS_IN_DELAY_MAX);
3826         debug_cond(DLEVEL == 1, "io_in_d=%u io_out1_d=%u io_out2_d=%u ",
3827                    IO_IO_IN_DELAY_MAX, IO_IO_OUT1_DELAY_MAX,
3828                    IO_IO_OUT2_DELAY_MAX);
3829         debug_cond(DLEVEL == 1, "dqs_in_reserve=%u dqs_out_reserve=%u\n",
3830                    IO_DQS_IN_RESERVE, IO_DQS_OUT_RESERVE);
3831
3832         hc_initialize_rom_data();
3833
3834         /* update info for sims */
3835         reg_file_set_stage(CAL_STAGE_NIL);
3836         reg_file_set_group(0);
3837
3838         /*
3839          * Load global needed for those actions that require
3840          * some dynamic calibration support.
3841          */
3842         dyn_calib_steps = STATIC_CALIB_STEPS;
3843         /*
3844          * Load global to allow dynamic selection of delay loop settings
3845          * based on calibration mode.
3846          */
3847         if (!(dyn_calib_steps & CALIB_SKIP_DELAY_LOOPS))
3848                 skip_delay_mask = 0xff;
3849         else
3850                 skip_delay_mask = 0x0;
3851
3852         pass = run_mem_calibrate();
3853         debug_mem_calibrate(pass);
3854         return pass;
3855 }