drivers/gpu/drm/i915/gt/intel_reset.c

   1 // SPDX-License-Identifier: MIT
   2 /*
   3  * Copyright © 2008-2018 Intel Corporation
   4  */
   5
   6 #include <linux/sched/mm.h>
   7 #include <linux/stop_machine.h>
   8 #include <linux/string_helpers.h>
   9
  10 #include "display/intel_display_reset.h"
  11 #include "display/intel_overlay.h"
  12
  13 #include "gem/i915_gem_context.h"
  14
  15 #include "gt/intel_gt_regs.h"
  16
  17 #include "gt/uc/intel_gsc_fw.h"
  18
  19 #include "i915_drv.h"
  20 #include "i915_file_private.h"
  21 #include "i915_gpu_error.h"
  22 #include "i915_irq.h"
  23 #include "i915_reg.h"
  24 #include "intel_breadcrumbs.h"
  25 #include "intel_engine_pm.h"
  26 #include "intel_engine_regs.h"
  27 #include "intel_gt.h"
  28 #include "intel_gt_pm.h"
  29 #include "intel_gt_print.h"
  30 #include "intel_gt_requests.h"
  31 #include "intel_mchbar_regs.h"
  32 #include "intel_pci_config.h"
  33 #include "intel_reset.h"
  34
  35 #include "uc/intel_guc.h"
  36
  37 #define RESET_MAX_RETRIES 3
  38
  39 static void client_mark_guilty(struct i915_gem_context *ctx, bool banned)
  40 {
  41         struct drm_i915_file_private *file_priv = ctx->file_priv;
  42         unsigned long prev_hang;
  43         unsigned int score;
  44
  45         if (IS_ERR_OR_NULL(file_priv))
  46                 return;
  47
  48         score = 0;
  49         if (banned)
  50                 score = I915_CLIENT_SCORE_CONTEXT_BAN;
  51
  52         prev_hang = xchg(&file_priv->hang_timestamp, jiffies);
  53         if (time_before(jiffies, prev_hang + I915_CLIENT_FAST_HANG_JIFFIES))
  54                 score += I915_CLIENT_SCORE_HANG_FAST;
  55
  56         if (score) {
  57                 atomic_add(score, &file_priv->ban_score);
  58
  59                 drm_dbg(&ctx->i915->drm,
  60                         "client %s: gained %u ban score, now %u\n",
  61                         ctx->name, score,
  62                         atomic_read(&file_priv->ban_score));
  63         }
  64 }
  65
  66 static bool mark_guilty(struct i915_request *rq)
  67 {
  68         struct i915_gem_context *ctx;
  69         unsigned long prev_hang;
  70         bool banned;
  71         int i;
  72
  73         if (intel_context_is_closed(rq->context))
  74                 return true;
  75
  76         rcu_read_lock();
  77         ctx = rcu_dereference(rq->context->gem_context);
  78         if (ctx && !kref_get_unless_zero(&ctx->ref))
  79                 ctx = NULL;
  80         rcu_read_unlock();
  81         if (!ctx)
  82                 return intel_context_is_banned(rq->context);
  83
  84         atomic_inc(&ctx->guilty_count);
  85
  86         /* Cool contexts are too cool to be banned! (Used for reset testing.) */
  87         if (!i915_gem_context_is_bannable(ctx)) {
  88                 banned = false;
  89                 goto out;
  90         }
  91
  92         drm_notice(&ctx->i915->drm,
  93                    "%s context reset due to GPU hang\n",
  94                    ctx->name);
  95
  96         /* Record the timestamp for the last N hangs */
  97         prev_hang = ctx->hang_timestamp[0];
  98         for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp) - 1; i++)
  99                 ctx->hang_timestamp[i] = ctx->hang_timestamp[i + 1];
 100         ctx->hang_timestamp[i] = jiffies;
 101
 102         /* If we have hung N+1 times in rapid succession, we ban the context! */
 103         banned = !i915_gem_context_is_recoverable(ctx);
 104         if (time_before(jiffies, prev_hang + CONTEXT_FAST_HANG_JIFFIES))
 105                 banned = true;
 106         if (banned)
 107                 drm_dbg(&ctx->i915->drm, "context %s: guilty %d, banned\n",
 108                         ctx->name, atomic_read(&ctx->guilty_count));
 109
 110         client_mark_guilty(ctx, banned);
 111
 112 out:
 113         i915_gem_context_put(ctx);
 114         return banned;
 115 }
 116
 117 static void mark_innocent(struct i915_request *rq)
 118 {
 119         struct i915_gem_context *ctx;
 120
 121         rcu_read_lock();
 122         ctx = rcu_dereference(rq->context->gem_context);
 123         if (ctx)
 124                 atomic_inc(&ctx->active_count);
 125         rcu_read_unlock();
 126 }
 127
 128 void __i915_request_reset(struct i915_request *rq, bool guilty)
 129 {
 130         bool banned = false;
 131
 132         RQ_TRACE(rq, "guilty? %s\n", str_yes_no(guilty));
 133         GEM_BUG_ON(__i915_request_is_complete(rq));
 134
 135         rcu_read_lock(); /* protect the GEM context */
 136         if (guilty) {
 137                 i915_request_set_error_once(rq, -EIO);
 138                 __i915_request_skip(rq);
 139                 banned = mark_guilty(rq);
 140         } else {
 141                 i915_request_set_error_once(rq, -EAGAIN);
 142                 mark_innocent(rq);
 143         }
 144         rcu_read_unlock();
 145
 146         if (banned)
 147                 intel_context_ban(rq->context, rq);
 148 }
 149
 150 static bool i915_in_reset(struct pci_dev *pdev)
 151 {
 152         u8 gdrst;
 153
 154         pci_read_config_byte(pdev, I915_GDRST, &gdrst);
 155         return gdrst & GRDOM_RESET_STATUS;
 156 }
 157
 158 static int i915_do_reset(struct intel_gt *gt,
 159                          intel_engine_mask_t engine_mask,
 160                          unsigned int retry)
 161 {
 162         struct pci_dev *pdev = to_pci_dev(gt->i915->drm.dev);
 163         int err;
 164
 165         /* Assert reset for at least 50 usec, and wait for acknowledgement. */
 166         pci_write_config_byte(pdev, I915_GDRST, GRDOM_RESET_ENABLE);
 167         udelay(50);
 168         err = _wait_for_atomic(i915_in_reset(pdev), 50000, 0);
 169
 170         /* Clear the reset request. */
 171         pci_write_config_byte(pdev, I915_GDRST, 0);
 172         udelay(50);
 173         if (!err)
 174                 err = _wait_for_atomic(!i915_in_reset(pdev), 50000, 0);
 175
 176         return err;
 177 }
 178
 179 static bool g4x_reset_complete(struct pci_dev *pdev)
 180 {
 181         u8 gdrst;
 182
 183         pci_read_config_byte(pdev, I915_GDRST, &gdrst);
 184         return (gdrst & GRDOM_RESET_ENABLE) == 0;
 185 }
 186
 187 static int g33_do_reset(struct intel_gt *gt,
 188                         intel_engine_mask_t engine_mask,
 189                         unsigned int retry)
 190 {
 191         struct pci_dev *pdev = to_pci_dev(gt->i915->drm.dev);
 192
 193         pci_write_config_byte(pdev, I915_GDRST, GRDOM_RESET_ENABLE);
 194         return _wait_for_atomic(g4x_reset_complete(pdev), 50000, 0);
 195 }
 196
 197 static int g4x_do_reset(struct intel_gt *gt,
 198                         intel_engine_mask_t engine_mask,
 199                         unsigned int retry)
 200 {
 201         struct pci_dev *pdev = to_pci_dev(gt->i915->drm.dev);
 202         struct intel_uncore *uncore = gt->uncore;
 203         int ret;
 204
 205         /* WaVcpClkGateDisableForMediaReset:ctg,elk */
 206         intel_uncore_rmw_fw(uncore, VDECCLK_GATE_D, 0, VCP_UNIT_CLOCK_GATE_DISABLE);
 207         intel_uncore_posting_read_fw(uncore, VDECCLK_GATE_D);
 208
 209         pci_write_config_byte(pdev, I915_GDRST,
 210                               GRDOM_MEDIA | GRDOM_RESET_ENABLE);
 211         ret =  _wait_for_atomic(g4x_reset_complete(pdev), 50000, 0);
 212         if (ret) {
 213                 GT_TRACE(gt, "Wait for media reset failed\n");
 214                 goto out;
 215         }
 216
 217         pci_write_config_byte(pdev, I915_GDRST,
 218                               GRDOM_RENDER | GRDOM_RESET_ENABLE);
 219         ret =  _wait_for_atomic(g4x_reset_complete(pdev), 50000, 0);
 220         if (ret) {
 221                 GT_TRACE(gt, "Wait for render reset failed\n");
 222                 goto out;
 223         }
 224
 225 out:
 226         pci_write_config_byte(pdev, I915_GDRST, 0);
 227
 228         intel_uncore_rmw_fw(uncore, VDECCLK_GATE_D, VCP_UNIT_CLOCK_GATE_DISABLE, 0);
 229         intel_uncore_posting_read_fw(uncore, VDECCLK_GATE_D);
 230
 231         return ret;
 232 }
 233
 234 static int ilk_do_reset(struct intel_gt *gt, intel_engine_mask_t engine_mask,
 235                         unsigned int retry)
 236 {
 237         struct intel_uncore *uncore = gt->uncore;
 238         int ret;
 239
 240         intel_uncore_write_fw(uncore, ILK_GDSR,
 241                               ILK_GRDOM_RENDER | ILK_GRDOM_RESET_ENABLE);
 242         ret = __intel_wait_for_register_fw(uncore, ILK_GDSR,
 243                                            ILK_GRDOM_RESET_ENABLE, 0,
 244                                            5000, 0,
 245                                            NULL);
 246         if (ret) {
 247                 GT_TRACE(gt, "Wait for render reset failed\n");
 248                 goto out;
 249         }
 250
 251         intel_uncore_write_fw(uncore, ILK_GDSR,
 252                               ILK_GRDOM_MEDIA | ILK_GRDOM_RESET_ENABLE);
 253         ret = __intel_wait_for_register_fw(uncore, ILK_GDSR,
 254                                            ILK_GRDOM_RESET_ENABLE, 0,
 255                                            5000, 0,
 256                                            NULL);
 257         if (ret) {
 258                 GT_TRACE(gt, "Wait for media reset failed\n");
 259                 goto out;
 260         }
 261
 262 out:
 263         intel_uncore_write_fw(uncore, ILK_GDSR, 0);
 264         intel_uncore_posting_read_fw(uncore, ILK_GDSR);
 265         return ret;
 266 }
 267
 268 /* Reset the hardware domains (GENX_GRDOM_*) specified by mask */
 269 static int gen6_hw_domain_reset(struct intel_gt *gt, u32 hw_domain_mask)
 270 {
 271         struct intel_uncore *uncore = gt->uncore;
 272         int loops;
 273         int err;
 274
 275         /*
 276          * On some platforms, e.g. Jasperlake, we see that the engine register
 277          * state is not cleared until shortly after GDRST reports completion,
 278          * causing a failure as we try to immediately resume while the internal
 279          * state is still in flux. If we immediately repeat the reset, the
 280          * second reset appears to serialise with the first, and since it is a
 281          * no-op, the registers should retain their reset value. However, there
 282          * is still a concern that upon leaving the second reset, the internal
 283          * engine state is still in flux and not ready for resuming.
 284          *
 285          * Starting on MTL, there are some prep steps that we need to do when
 286          * resetting some engines that need to be applied every time we write to
 287          * GEN6_GDRST. As those are time consuming (tens of ms), we don't want
 288          * to perform that twice, so, since the Jasperlake issue hasn't been
 289          * observed on MTL, we avoid repeating the reset on newer platforms.
 290          */
 291         loops = GRAPHICS_VER_FULL(gt->i915) < IP_VER(12, 70) ? 2 : 1;
 292
 293         /*
 294          * GEN6_GDRST is not in the gt power well, no need to check
 295          * for fifo space for the write or forcewake the chip for
 296          * the read
 297          */
 298         do {
 299                 intel_uncore_write_fw(uncore, GEN6_GDRST, hw_domain_mask);
 300
 301                 /* Wait for the device to ack the reset requests. */
 302                 err = __intel_wait_for_register_fw(uncore, GEN6_GDRST,
 303                                                    hw_domain_mask, 0,
 304                                                    2000, 0,
 305                                                    NULL);
 306         } while (err == 0 && --loops);
 307         if (err)
 308                 GT_TRACE(gt,
 309                          "Wait for 0x%08x engines reset failed\n",
 310                          hw_domain_mask);
 311
 312         /*
 313          * As we have observed that the engine state is still volatile
 314          * after GDRST is acked, impose a small delay to let everything settle.
 315          */
 316         udelay(50);
 317
 318         return err;
 319 }
 320
 321 static int __gen6_reset_engines(struct intel_gt *gt,
 322                                 intel_engine_mask_t engine_mask,
 323                                 unsigned int retry)
 324 {
 325         struct intel_engine_cs *engine;
 326         u32 hw_mask;
 327
 328         if (engine_mask == ALL_ENGINES) {
 329                 hw_mask = GEN6_GRDOM_FULL;
 330         } else {
 331                 intel_engine_mask_t tmp;
 332
 333                 hw_mask = 0;
 334                 for_each_engine_masked(engine, gt, engine_mask, tmp) {
 335                         hw_mask |= engine->reset_domain;
 336                 }
 337         }
 338
 339         return gen6_hw_domain_reset(gt, hw_mask);
 340 }
 341
 342 static int gen6_reset_engines(struct intel_gt *gt,
 343                               intel_engine_mask_t engine_mask,
 344                               unsigned int retry)
 345 {
 346         unsigned long flags;
 347         int ret;
 348
 349         spin_lock_irqsave(&gt->uncore->lock, flags);
 350         ret = __gen6_reset_engines(gt, engine_mask, retry);
 351         spin_unlock_irqrestore(&gt->uncore->lock, flags);
 352
 353         return ret;
 354 }
 355
 356 static struct intel_engine_cs *find_sfc_paired_vecs_engine(struct intel_engine_cs *engine)
 357 {
 358         int vecs_id;
 359
 360         GEM_BUG_ON(engine->class != VIDEO_DECODE_CLASS);
 361
 362         vecs_id = _VECS((engine->instance) / 2);
 363
 364         return engine->gt->engine[vecs_id];
 365 }
 366
 367 struct sfc_lock_data {
 368         i915_reg_t lock_reg;
 369         i915_reg_t ack_reg;
 370         i915_reg_t usage_reg;
 371         u32 lock_bit;
 372         u32 ack_bit;
 373         u32 usage_bit;
 374         u32 reset_bit;
 375 };
 376
 377 static void get_sfc_forced_lock_data(struct intel_engine_cs *engine,
 378                                      struct sfc_lock_data *sfc_lock)
 379 {
 380         switch (engine->class) {
 381         default:
 382                 MISSING_CASE(engine->class);
 383                 fallthrough;
 384         case VIDEO_DECODE_CLASS:
 385                 sfc_lock->lock_reg = GEN11_VCS_SFC_FORCED_LOCK(engine->mmio_base);
 386                 sfc_lock->lock_bit = GEN11_VCS_SFC_FORCED_LOCK_BIT;
 387
 388                 sfc_lock->ack_reg = GEN11_VCS_SFC_LOCK_STATUS(engine->mmio_base);
 389                 sfc_lock->ack_bit  = GEN11_VCS_SFC_LOCK_ACK_BIT;
 390
 391                 sfc_lock->usage_reg = GEN11_VCS_SFC_LOCK_STATUS(engine->mmio_base);
 392                 sfc_lock->usage_bit = GEN11_VCS_SFC_USAGE_BIT;
 393                 sfc_lock->reset_bit = GEN11_VCS_SFC_RESET_BIT(engine->instance);
 394
 395                 break;
 396         case VIDEO_ENHANCEMENT_CLASS:
 397                 sfc_lock->lock_reg = GEN11_VECS_SFC_FORCED_LOCK(engine->mmio_base);
 398                 sfc_lock->lock_bit = GEN11_VECS_SFC_FORCED_LOCK_BIT;
 399
 400                 sfc_lock->ack_reg = GEN11_VECS_SFC_LOCK_ACK(engine->mmio_base);
 401                 sfc_lock->ack_bit  = GEN11_VECS_SFC_LOCK_ACK_BIT;
 402
 403                 sfc_lock->usage_reg = GEN11_VECS_SFC_USAGE(engine->mmio_base);
 404                 sfc_lock->usage_bit = GEN11_VECS_SFC_USAGE_BIT;
 405                 sfc_lock->reset_bit = GEN11_VECS_SFC_RESET_BIT(engine->instance);
 406
 407                 break;
 408         }
 409 }
 410
 411 static int gen11_lock_sfc(struct intel_engine_cs *engine,
 412                           u32 *reset_mask,
 413                           u32 *unlock_mask)
 414 {
 415         struct intel_uncore *uncore = engine->uncore;
 416         u8 vdbox_sfc_access = engine->gt->info.vdbox_sfc_access;
 417         struct sfc_lock_data sfc_lock;
 418         bool lock_obtained, lock_to_other = false;
 419         int ret;
 420
 421         switch (engine->class) {
 422         case VIDEO_DECODE_CLASS:
 423                 if ((BIT(engine->instance) & vdbox_sfc_access) == 0)
 424                         return 0;
 425
 426                 fallthrough;
 427         case VIDEO_ENHANCEMENT_CLASS:
 428                 get_sfc_forced_lock_data(engine, &sfc_lock);
 429
 430                 break;
 431         default:
 432                 return 0;
 433         }
 434
 435         if (!(intel_uncore_read_fw(uncore, sfc_lock.usage_reg) & sfc_lock.usage_bit)) {
 436                 struct intel_engine_cs *paired_vecs;
 437
 438                 if (engine->class != VIDEO_DECODE_CLASS ||
 439                     GRAPHICS_VER(engine->i915) != 12)
 440                         return 0;
 441
 442                 /*
 443                  * Wa_14010733141
 444                  *
 445                  * If the VCS-MFX isn't using the SFC, we also need to check
 446                  * whether VCS-HCP is using it.  If so, we need to issue a *VE*
 447                  * forced lock on the VE engine that shares the same SFC.
 448                  */
 449                 if (!(intel_uncore_read_fw(uncore,
 450                                            GEN12_HCP_SFC_LOCK_STATUS(engine->mmio_base)) &
 451                       GEN12_HCP_SFC_USAGE_BIT))
 452                         return 0;
 453
 454                 paired_vecs = find_sfc_paired_vecs_engine(engine);
 455                 get_sfc_forced_lock_data(paired_vecs, &sfc_lock);
 456                 lock_to_other = true;
 457                 *unlock_mask |= paired_vecs->mask;
 458         } else {
 459                 *unlock_mask |= engine->mask;
 460         }
 461
 462         /*
 463          * If the engine is using an SFC, tell the engine that a software reset
 464          * is going to happen. The engine will then try to force lock the SFC.
 465          * If SFC ends up being locked to the engine we want to reset, we have
 466          * to reset it as well (we will unlock it once the reset sequence is
 467          * completed).
 468          */
 469         intel_uncore_rmw_fw(uncore, sfc_lock.lock_reg, 0, sfc_lock.lock_bit);
 470
 471         ret = __intel_wait_for_register_fw(uncore,
 472                                            sfc_lock.ack_reg,
 473                                            sfc_lock.ack_bit,
 474                                            sfc_lock.ack_bit,
 475                                            1000, 0, NULL);
 476
 477         /*
 478          * Was the SFC released while we were trying to lock it?
 479          *
 480          * We should reset both the engine and the SFC if:
 481          *  - We were locking the SFC to this engine and the lock succeeded
 482          *       OR
 483          *  - We were locking the SFC to a different engine (Wa_14010733141)
 484          *    but the SFC was released before the lock was obtained.
 485          *
 486          * Otherwise we need only reset the engine by itself and we can
 487          * leave the SFC alone.
 488          */
 489         lock_obtained = (intel_uncore_read_fw(uncore, sfc_lock.usage_reg) &
 490                         sfc_lock.usage_bit) != 0;
 491         if (lock_obtained == lock_to_other)
 492                 return 0;
 493
 494         if (ret) {
 495                 ENGINE_TRACE(engine, "Wait for SFC forced lock ack failed\n");
 496                 return ret;
 497         }
 498
 499         *reset_mask |= sfc_lock.reset_bit;
 500         return 0;
 501 }
 502
 503 static void gen11_unlock_sfc(struct intel_engine_cs *engine)
 504 {
 505         struct intel_uncore *uncore = engine->uncore;
 506         u8 vdbox_sfc_access = engine->gt->info.vdbox_sfc_access;
 507         struct sfc_lock_data sfc_lock = {};
 508
 509         if (engine->class != VIDEO_DECODE_CLASS &&
 510             engine->class != VIDEO_ENHANCEMENT_CLASS)
 511                 return;
 512
 513         if (engine->class == VIDEO_DECODE_CLASS &&
 514             (BIT(engine->instance) & vdbox_sfc_access) == 0)
 515                 return;
 516
 517         get_sfc_forced_lock_data(engine, &sfc_lock);
 518
 519         intel_uncore_rmw_fw(uncore, sfc_lock.lock_reg, sfc_lock.lock_bit, 0);
 520 }
 521
 522 static int __gen11_reset_engines(struct intel_gt *gt,
 523                                  intel_engine_mask_t engine_mask,
 524                                  unsigned int retry)
 525 {
 526         struct intel_engine_cs *engine;
 527         intel_engine_mask_t tmp;
 528         u32 reset_mask, unlock_mask = 0;
 529         int ret;
 530
 531         if (engine_mask == ALL_ENGINES) {
 532                 reset_mask = GEN11_GRDOM_FULL;
 533         } else {
 534                 reset_mask = 0;
 535                 for_each_engine_masked(engine, gt, engine_mask, tmp) {
 536                         reset_mask |= engine->reset_domain;
 537                         ret = gen11_lock_sfc(engine, &reset_mask, &unlock_mask);
 538                         if (ret)
 539                                 goto sfc_unlock;
 540                 }
 541         }
 542
 543         ret = gen6_hw_domain_reset(gt, reset_mask);
 544
 545 sfc_unlock:
 546         /*
 547          * We unlock the SFC based on the lock status and not the result of
 548          * gen11_lock_sfc to make sure that we clean properly if something
 549          * wrong happened during the lock (e.g. lock acquired after timeout
 550          * expiration).
 551          *
 552          * Due to Wa_14010733141, we may have locked an SFC to an engine that
 553          * wasn't being reset.  So instead of calling gen11_unlock_sfc()
 554          * on engine_mask, we instead call it on the mask of engines that our
 555          * gen11_lock_sfc() calls told us actually had locks attempted.
 556          */
 557         for_each_engine_masked(engine, gt, unlock_mask, tmp)
 558                 gen11_unlock_sfc(engine);
 559
 560         return ret;
 561 }
 562
 563 static int gen8_engine_reset_prepare(struct intel_engine_cs *engine)
 564 {
 565         struct intel_uncore *uncore = engine->uncore;
 566         const i915_reg_t reg = RING_RESET_CTL(engine->mmio_base);
 567         u32 request, mask, ack;
 568         int ret;
 569
 570         if (I915_SELFTEST_ONLY(should_fail(&engine->reset_timeout, 1)))
 571                 return -ETIMEDOUT;
 572
 573         ack = intel_uncore_read_fw(uncore, reg);
 574         if (ack & RESET_CTL_CAT_ERROR) {
 575                 /*
 576                  * For catastrophic errors, ready-for-reset sequence
 577                  * needs to be bypassed: HAS#396813
 578                  */
 579                 request = RESET_CTL_CAT_ERROR;
 580                 mask = RESET_CTL_CAT_ERROR;
 581
 582                 /* Catastrophic errors need to be cleared by HW */
 583                 ack = 0;
 584         } else if (!(ack & RESET_CTL_READY_TO_RESET)) {
 585                 request = RESET_CTL_REQUEST_RESET;
 586                 mask = RESET_CTL_READY_TO_RESET;
 587                 ack = RESET_CTL_READY_TO_RESET;
 588         } else {
 589                 return 0;
 590         }
 591
 592         intel_uncore_write_fw(uncore, reg, _MASKED_BIT_ENABLE(request));
 593         ret = __intel_wait_for_register_fw(uncore, reg, mask, ack,
 594                                            700, 0, NULL);
 595         if (ret)
 596                 gt_err(engine->gt,
 597                        "%s reset request timed out: {request: %08x, RESET_CTL: %08x}\n",
 598                        engine->name, request,
 599                        intel_uncore_read_fw(uncore, reg));
 600
 601         return ret;
 602 }
 603
 604 static void gen8_engine_reset_cancel(struct intel_engine_cs *engine)
 605 {
 606         intel_uncore_write_fw(engine->uncore,
 607                               RING_RESET_CTL(engine->mmio_base),
 608                               _MASKED_BIT_DISABLE(RESET_CTL_REQUEST_RESET));
 609 }
 610
 611 static int gen8_reset_engines(struct intel_gt *gt,
 612                               intel_engine_mask_t engine_mask,
 613                               unsigned int retry)
 614 {
 615         struct intel_engine_cs *engine;
 616         const bool reset_non_ready = retry >= 1;
 617         intel_engine_mask_t tmp;
 618         unsigned long flags;
 619         int ret;
 620
 621         spin_lock_irqsave(&gt->uncore->lock, flags);
 622
 623         for_each_engine_masked(engine, gt, engine_mask, tmp) {
 624                 ret = gen8_engine_reset_prepare(engine);
 625                 if (ret && !reset_non_ready)
 626                         goto skip_reset;
 627
 628                 /*
 629                  * If this is not the first failed attempt to prepare,
 630                  * we decide to proceed anyway.
 631                  *
 632                  * By doing so we risk context corruption and with
 633                  * some gens (kbl), possible system hang if reset
 634                  * happens during active bb execution.
 635                  *
 636                  * We rather take context corruption instead of
 637                  * failed reset with a wedged driver/gpu. And
 638                  * active bb execution case should be covered by
 639                  * stop_engines() we have before the reset.
 640                  */
 641         }
 642
 643         /*
 644          * Wa_22011100796:dg2, whenever Full soft reset is required,
 645          * reset all individual engines firstly, and then do a full soft reset.
 646          *
 647          * This is best effort, so ignore any error from the initial reset.
 648          */
 649         if (IS_DG2(gt->i915) && engine_mask == ALL_ENGINES)
 650                 __gen11_reset_engines(gt, gt->info.engine_mask, 0);
 651
 652         if (GRAPHICS_VER(gt->i915) >= 11)
 653                 ret = __gen11_reset_engines(gt, engine_mask, retry);
 654         else
 655                 ret = __gen6_reset_engines(gt, engine_mask, retry);
 656
 657 skip_reset:
 658         for_each_engine_masked(engine, gt, engine_mask, tmp)
 659                 gen8_engine_reset_cancel(engine);
 660
 661         spin_unlock_irqrestore(&gt->uncore->lock, flags);
 662
 663         return ret;
 664 }
 665
 666 static int mock_reset(struct intel_gt *gt,
 667                       intel_engine_mask_t mask,
 668                       unsigned int retry)
 669 {
 670         return 0;
 671 }
 672
 673 typedef int (*reset_func)(struct intel_gt *,
 674                           intel_engine_mask_t engine_mask,
 675                           unsigned int retry);
 676
 677 static reset_func intel_get_gpu_reset(const struct intel_gt *gt)
 678 {
 679         struct drm_i915_private *i915 = gt->i915;
 680
 681         if (is_mock_gt(gt))
 682                 return mock_reset;
 683         else if (GRAPHICS_VER(i915) >= 8)
 684                 return gen8_reset_engines;
 685         else if (GRAPHICS_VER(i915) >= 6)
 686                 return gen6_reset_engines;
 687         else if (GRAPHICS_VER(i915) >= 5)
 688                 return ilk_do_reset;
 689         else if (IS_G4X(i915))
 690                 return g4x_do_reset;
 691         else if (IS_G33(i915) || IS_PINEVIEW(i915))
 692                 return g33_do_reset;
 693         else if (GRAPHICS_VER(i915) >= 3)
 694                 return i915_do_reset;
 695         else
 696                 return NULL;
 697 }
 698
 699 static int __reset_guc(struct intel_gt *gt)
 700 {
 701         u32 guc_domain =
 702                 GRAPHICS_VER(gt->i915) >= 11 ? GEN11_GRDOM_GUC : GEN9_GRDOM_GUC;
 703
 704         return gen6_hw_domain_reset(gt, guc_domain);
 705 }
 706
 707 static bool needs_wa_14015076503(struct intel_gt *gt, intel_engine_mask_t engine_mask)
 708 {
 709         if (MEDIA_VER_FULL(gt->i915) != IP_VER(13, 0) || !HAS_ENGINE(gt, GSC0))
 710                 return false;
 711
 712         if (!__HAS_ENGINE(engine_mask, GSC0))
 713                 return false;
 714
 715         return intel_gsc_uc_fw_init_done(&gt->uc.gsc);
 716 }
 717
 718 static intel_engine_mask_t
 719 wa_14015076503_start(struct intel_gt *gt, intel_engine_mask_t engine_mask, bool first)
 720 {
 721         if (!needs_wa_14015076503(gt, engine_mask))
 722                 return engine_mask;
 723
 724         /*
 725          * wa_14015076503: if the GSC FW is loaded, we need to alert it that
 726          * we're going to do a GSC engine reset and then wait for 200ms for the
 727          * FW to get ready for it. However, if this is the first ALL_ENGINES
 728          * reset attempt and the GSC is not busy, we can try to instead reset
 729          * the GuC and all the other engines individually to avoid the 200ms
 730          * wait.
 731          * Skipping the GSC engine is safe because, differently from other
 732          * engines, the GSCCS only role is to forward the commands to the GSC
 733          * FW, so it doesn't have any HW outside of the CS itself and therefore
 734          * it has no state that we don't explicitly re-init on resume or on
 735          * context switch LRC or power context). The HW for the GSC uC is
 736          * managed by the GSC FW so we don't need to care about that.
 737          */
 738         if (engine_mask == ALL_ENGINES && first && intel_engine_is_idle(gt->engine[GSC0])) {
 739                 __reset_guc(gt);
 740                 engine_mask = gt->info.engine_mask & ~BIT(GSC0);
 741         } else {
 742                 intel_uncore_rmw(gt->uncore,
 743                                  HECI_H_GS1(MTL_GSC_HECI2_BASE),
 744                                  0, HECI_H_GS1_ER_PREP);
 745
 746                 /* make sure the reset bit is clear when writing the CSR reg */
 747                 intel_uncore_rmw(gt->uncore,
 748                                  HECI_H_CSR(MTL_GSC_HECI2_BASE),
 749                                  HECI_H_CSR_RST, HECI_H_CSR_IG);
 750                 msleep(200);
 751         }
 752
 753         return engine_mask;
 754 }
 755
 756 static void
 757 wa_14015076503_end(struct intel_gt *gt, intel_engine_mask_t engine_mask)
 758 {
 759         if (!needs_wa_14015076503(gt, engine_mask))
 760                 return;
 761
 762         intel_uncore_rmw(gt->uncore,
 763                          HECI_H_GS1(MTL_GSC_HECI2_BASE),
 764                          HECI_H_GS1_ER_PREP, 0);
 765 }
 766
 767 int __intel_gt_reset(struct intel_gt *gt, intel_engine_mask_t engine_mask)
 768 {
 769         const int retries = engine_mask == ALL_ENGINES ? RESET_MAX_RETRIES : 1;
 770         reset_func reset;
 771         int ret = -ETIMEDOUT;
 772         int retry;
 773
 774         reset = intel_get_gpu_reset(gt);
 775         if (!reset)
 776                 return -ENODEV;
 777
 778         /*
 779          * If the power well sleeps during the reset, the reset
 780          * request may be dropped and never completes (causing -EIO).
 781          */
 782         intel_uncore_forcewake_get(gt->uncore, FORCEWAKE_ALL);
 783         for (retry = 0; ret == -ETIMEDOUT && retry < retries; retry++) {
 784                 intel_engine_mask_t reset_mask;
 785
 786                 reset_mask = wa_14015076503_start(gt, engine_mask, !retry);
 787
 788                 GT_TRACE(gt, "engine_mask=%x\n", reset_mask);
 789                 ret = reset(gt, reset_mask, retry);
 790
 791                 wa_14015076503_end(gt, reset_mask);
 792         }
 793         intel_uncore_forcewake_put(gt->uncore, FORCEWAKE_ALL);
 794
 795         return ret;
 796 }
 797
 798 bool intel_has_gpu_reset(const struct intel_gt *gt)
 799 {
 800         if (!gt->i915->params.reset)
 801                 return NULL;
 802
 803         return intel_get_gpu_reset(gt);
 804 }
 805
 806 bool intel_has_reset_engine(const struct intel_gt *gt)
 807 {
 808         if (gt->i915->params.reset < 2)
 809                 return false;
 810
 811         return INTEL_INFO(gt->i915)->has_reset_engine;
 812 }
 813
 814 int intel_reset_guc(struct intel_gt *gt)
 815 {
 816         int ret;
 817
 818         GEM_BUG_ON(!HAS_GT_UC(gt->i915));
 819
 820         intel_uncore_forcewake_get(gt->uncore, FORCEWAKE_ALL);
 821         ret = __reset_guc(gt);
 822         intel_uncore_forcewake_put(gt->uncore, FORCEWAKE_ALL);
 823
 824         return ret;
 825 }
 826
 827 /*
 828  * Ensure irq handler finishes, and not run again.
 829  * Also return the active request so that we only search for it once.
 830  */
 831 static void reset_prepare_engine(struct intel_engine_cs *engine)
 832 {
 833         /*
 834          * During the reset sequence, we must prevent the engine from
 835          * entering RC6. As the context state is undefined until we restart
 836          * the engine, if it does enter RC6 during the reset, the state
 837          * written to the powercontext is undefined and so we may lose
 838          * GPU state upon resume, i.e. fail to restart after a reset.
 839          */
 840         intel_uncore_forcewake_get(engine->uncore, FORCEWAKE_ALL);
 841         if (engine->reset.prepare)
 842                 engine->reset.prepare(engine);
 843 }
 844
 845 static void revoke_mmaps(struct intel_gt *gt)
 846 {
 847         int i;
 848
 849         for (i = 0; i < gt->ggtt->num_fences; i++) {
 850                 struct drm_vma_offset_node *node;
 851                 struct i915_vma *vma;
 852                 u64 vma_offset;
 853
 854                 vma = READ_ONCE(gt->ggtt->fence_regs[i].vma);
 855                 if (!vma)
 856                         continue;
 857
 858                 if (!i915_vma_has_userfault(vma))
 859                         continue;
 860
 861                 GEM_BUG_ON(vma->fence != &gt->ggtt->fence_regs[i]);
 862
 863                 if (!vma->mmo)
 864                         continue;
 865
 866                 node = &vma->mmo->vma_node;
 867                 vma_offset = vma->gtt_view.partial.offset << PAGE_SHIFT;
 868
 869                 unmap_mapping_range(gt->i915->drm.anon_inode->i_mapping,
 870                                     drm_vma_node_offset_addr(node) + vma_offset,
 871                                     vma->size,
 872                                     1);
 873         }
 874 }
 875
 876 static intel_engine_mask_t reset_prepare(struct intel_gt *gt)
 877 {
 878         struct intel_engine_cs *engine;
 879         intel_engine_mask_t awake = 0;
 880         enum intel_engine_id id;
 881
 882         /* For GuC mode, ensure submission is disabled before stopping ring */
 883         intel_uc_reset_prepare(&gt->uc);
 884
 885         for_each_engine(engine, gt, id) {
 886                 if (intel_engine_pm_get_if_awake(engine))
 887                         awake |= engine->mask;
 888                 reset_prepare_engine(engine);
 889         }
 890
 891         return awake;
 892 }
 893
 894 static void gt_revoke(struct intel_gt *gt)
 895 {
 896         revoke_mmaps(gt);
 897 }
 898
 899 static int gt_reset(struct intel_gt *gt, intel_engine_mask_t stalled_mask)
 900 {
 901         struct intel_engine_cs *engine;
 902         enum intel_engine_id id;
 903         int err;
 904
 905         /*
 906          * Everything depends on having the GTT running, so we need to start
 907          * there.
 908          */
 909         err = i915_ggtt_enable_hw(gt->i915);
 910         if (err)
 911                 return err;
 912
 913         local_bh_disable();
 914         for_each_engine(engine, gt, id)
 915                 __intel_engine_reset(engine, stalled_mask & engine->mask);
 916         local_bh_enable();
 917
 918         intel_uc_reset(&gt->uc, ALL_ENGINES);
 919
 920         intel_ggtt_restore_fences(gt->ggtt);
 921
 922         return err;
 923 }
 924
 925 static void reset_finish_engine(struct intel_engine_cs *engine)
 926 {
 927         if (engine->reset.finish)
 928                 engine->reset.finish(engine);
 929         intel_uncore_forcewake_put(engine->uncore, FORCEWAKE_ALL);
 930
 931         intel_engine_signal_breadcrumbs(engine);
 932 }
 933
 934 static void reset_finish(struct intel_gt *gt, intel_engine_mask_t awake)
 935 {
 936         struct intel_engine_cs *engine;
 937         enum intel_engine_id id;
 938
 939         for_each_engine(engine, gt, id) {
 940                 reset_finish_engine(engine);
 941                 if (awake & engine->mask)
 942                         intel_engine_pm_put(engine);
 943         }
 944
 945         intel_uc_reset_finish(&gt->uc);
 946 }
 947
 948 static void nop_submit_request(struct i915_request *request)
 949 {
 950         RQ_TRACE(request, "-EIO\n");
 951
 952         request = i915_request_mark_eio(request);
 953         if (request) {
 954                 i915_request_submit(request);
 955                 intel_engine_signal_breadcrumbs(request->engine);
 956
 957                 i915_request_put(request);
 958         }
 959 }
 960
 961 static void __intel_gt_set_wedged(struct intel_gt *gt)
 962 {
 963         struct intel_engine_cs *engine;
 964         intel_engine_mask_t awake;
 965         enum intel_engine_id id;
 966
 967         if (test_bit(I915_WEDGED, &gt->reset.flags))
 968                 return;
 969
 970         GT_TRACE(gt, "start\n");
 971
 972         /*
 973          * First, stop submission to hw, but do not yet complete requests by
 974          * rolling the global seqno forward (since this would complete requests
 975          * for which we haven't set the fence error to EIO yet).
 976          */
 977         awake = reset_prepare(gt);
 978
 979         /* Even if the GPU reset fails, it should still stop the engines */
 980         if (!INTEL_INFO(gt->i915)->gpu_reset_clobbers_display)
 981                 __intel_gt_reset(gt, ALL_ENGINES);
 982
 983         for_each_engine(engine, gt, id)
 984                 engine->submit_request = nop_submit_request;
 985
 986         /*
 987          * Make sure no request can slip through without getting completed by
 988          * either this call here to intel_engine_write_global_seqno, or the one
 989          * in nop_submit_request.
 990          */
 991         synchronize_rcu_expedited();
 992         set_bit(I915_WEDGED, &gt->reset.flags);
 993
 994         /* Mark all executing requests as skipped */
 995         local_bh_disable();
 996         for_each_engine(engine, gt, id)
 997                 if (engine->reset.cancel)
 998                         engine->reset.cancel(engine);
 999         intel_uc_cancel_requests(&gt->uc);
1000         local_bh_enable();
1001
1002         reset_finish(gt, awake);
1003
1004         GT_TRACE(gt, "end\n");
1005 }
1006
1007 void intel_gt_set_wedged(struct intel_gt *gt)
1008 {
1009         intel_wakeref_t wakeref;
1010
1011         if (test_bit(I915_WEDGED, &gt->reset.flags))
1012                 return;
1013
1014         wakeref = intel_runtime_pm_get(gt->uncore->rpm);
1015         mutex_lock(&gt->reset.mutex);
1016
1017         if (GEM_SHOW_DEBUG()) {
1018                 struct drm_printer p = drm_debug_printer(__func__);
1019                 struct intel_engine_cs *engine;
1020                 enum intel_engine_id id;
1021
1022                 drm_printf(&p, "called from %pS\n", (void *)_RET_IP_);
1023                 for_each_engine(engine, gt, id) {
1024                         if (intel_engine_is_idle(engine))
1025                                 continue;
1026
1027                         intel_engine_dump(engine, &p, "%s\n", engine->name);
1028                 }
1029         }
1030
1031         __intel_gt_set_wedged(gt);
1032
1033         mutex_unlock(&gt->reset.mutex);
1034         intel_runtime_pm_put(gt->uncore->rpm, wakeref);
1035 }
1036
1037 static bool __intel_gt_unset_wedged(struct intel_gt *gt)
1038 {
1039         struct intel_gt_timelines *timelines = &gt->timelines;
1040         struct intel_timeline *tl;
1041         bool ok;
1042
1043         if (!test_bit(I915_WEDGED, &gt->reset.flags))
1044                 return true;
1045
1046         /* Never fully initialised, recovery impossible */
1047         if (intel_gt_has_unrecoverable_error(gt))
1048                 return false;
1049
1050         GT_TRACE(gt, "start\n");
1051
1052         /*
1053          * Before unwedging, make sure that all pending operations
1054          * are flushed and errored out - we may have requests waiting upon
1055          * third party fences. We marked all inflight requests as EIO, and
1056          * every execbuf since returned EIO, for consistency we want all
1057          * the currently pending requests to also be marked as EIO, which
1058          * is done inside our nop_submit_request - and so we must wait.
1059          *
1060          * No more can be submitted until we reset the wedged bit.
1061          */
1062         spin_lock(&timelines->lock);
1063         list_for_each_entry(tl, &timelines->active_list, link) {
1064                 struct dma_fence *fence;
1065
1066                 fence = i915_active_fence_get(&tl->last_request);
1067                 if (!fence)
1068                         continue;
1069
1070                 spin_unlock(&timelines->lock);
1071
1072                 /*
1073                  * All internal dependencies (i915_requests) will have
1074                  * been flushed by the set-wedge, but we may be stuck waiting
1075                  * for external fences. These should all be capped to 10s
1076                  * (I915_FENCE_TIMEOUT) so this wait should not be unbounded
1077                  * in the worst case.
1078                  */
1079                 dma_fence_default_wait(fence, false, MAX_SCHEDULE_TIMEOUT);
1080                 dma_fence_put(fence);
1081
1082                 /* Restart iteration after droping lock */
1083                 spin_lock(&timelines->lock);
1084                 tl = list_entry(&timelines->active_list, typeof(*tl), link);
1085         }
1086         spin_unlock(&timelines->lock);
1087
1088         /* We must reset pending GPU events before restoring our submission */
1089         ok = !HAS_EXECLISTS(gt->i915); /* XXX better agnosticism desired */
1090         if (!INTEL_INFO(gt->i915)->gpu_reset_clobbers_display)
1091                 ok = __intel_gt_reset(gt, ALL_ENGINES) == 0;
1092         if (!ok) {
1093                 /*
1094                  * Warn CI about the unrecoverable wedged condition.
1095                  * Time for a reboot.
1096                  */
1097                 add_taint_for_CI(gt->i915, TAINT_WARN);
1098                 return false;
1099         }
1100
1101         /*
1102          * Undo nop_submit_request. We prevent all new i915 requests from
1103          * being queued (by disallowing execbuf whilst wedged) so having
1104          * waited for all active requests above, we know the system is idle
1105          * and do not have to worry about a thread being inside
1106          * engine->submit_request() as we swap over. So unlike installing
1107          * the nop_submit_request on reset, we can do this from normal
1108          * context and do not require stop_machine().
1109          */
1110         intel_engines_reset_default_submission(gt);
1111
1112         GT_TRACE(gt, "end\n");
1113
1114         smp_mb__before_atomic(); /* complete takeover before enabling execbuf */
1115         clear_bit(I915_WEDGED, &gt->reset.flags);
1116
1117         return true;
1118 }
1119
1120 bool intel_gt_unset_wedged(struct intel_gt *gt)
1121 {
1122         bool result;
1123
1124         mutex_lock(&gt->reset.mutex);
1125         result = __intel_gt_unset_wedged(gt);
1126         mutex_unlock(&gt->reset.mutex);
1127
1128         return result;
1129 }
1130
1131 static int do_reset(struct intel_gt *gt, intel_engine_mask_t stalled_mask)
1132 {
1133         int err, i;
1134
1135         err = __intel_gt_reset(gt, ALL_ENGINES);
1136         for (i = 0; err && i < RESET_MAX_RETRIES; i++) {
1137                 msleep(10 * (i + 1));
1138                 err = __intel_gt_reset(gt, ALL_ENGINES);
1139         }
1140         if (err)
1141                 return err;
1142
1143         return gt_reset(gt, stalled_mask);
1144 }
1145
1146 static int resume(struct intel_gt *gt)
1147 {
1148         struct intel_engine_cs *engine;
1149         enum intel_engine_id id;
1150         int ret;
1151
1152         for_each_engine(engine, gt, id) {
1153                 ret = intel_engine_resume(engine);
1154                 if (ret)
1155                         return ret;
1156         }
1157
1158         return 0;
1159 }
1160
1161 /**
1162  * intel_gt_reset - reset chip after a hang
1163  * @gt: #intel_gt to reset
1164  * @stalled_mask: mask of the stalled engines with the guilty requests
1165  * @reason: user error message for why we are resetting
1166  *
1167  * Reset the chip.  Useful if a hang is detected. Marks the device as wedged
1168  * on failure.
1169  *
1170  * Procedure is fairly simple:
1171  *   - reset the chip using the reset reg
1172  *   - re-init context state
1173  *   - re-init hardware status page
1174  *   - re-init ring buffer
1175  *   - re-init interrupt state
1176  *   - re-init display
1177  */
1178 void intel_gt_reset(struct intel_gt *gt,
1179                     intel_engine_mask_t stalled_mask,
1180                     const char *reason)
1181 {
1182         intel_engine_mask_t awake;
1183         int ret;
1184
1185         GT_TRACE(gt, "flags=%lx\n", gt->reset.flags);
1186
1187         might_sleep();
1188         GEM_BUG_ON(!test_bit(I915_RESET_BACKOFF, &gt->reset.flags));
1189
1190         /*
1191          * FIXME: Revoking cpu mmap ptes cannot be done from a dma_fence
1192          * critical section like gpu reset.
1193          */
1194         gt_revoke(gt);
1195
1196         mutex_lock(&gt->reset.mutex);
1197
1198         /* Clear any previous failed attempts at recovery. Time to try again. */
1199         if (!__intel_gt_unset_wedged(gt))
1200                 goto unlock;
1201
1202         if (reason)
1203                 gt_notice(gt, "Resetting chip for %s\n", reason);
1204         atomic_inc(&gt->i915->gpu_error.reset_count);
1205
1206         awake = reset_prepare(gt);
1207
1208         if (!intel_has_gpu_reset(gt)) {
1209                 if (gt->i915->params.reset)
1210                         gt_err(gt, "GPU reset not supported\n");
1211                 else
1212                         gt_dbg(gt, "GPU reset disabled\n");
1213                 goto error;
1214         }
1215
1216         if (INTEL_INFO(gt->i915)->gpu_reset_clobbers_display)
1217                 intel_runtime_pm_disable_interrupts(gt->i915);
1218
1219         if (do_reset(gt, stalled_mask)) {
1220                 gt_err(gt, "Failed to reset chip\n");
1221                 goto taint;
1222         }
1223
1224         if (INTEL_INFO(gt->i915)->gpu_reset_clobbers_display)
1225                 intel_runtime_pm_enable_interrupts(gt->i915);
1226
1227         intel_overlay_reset(gt->i915);
1228
1229         /*
1230          * Next we need to restore the context, but we don't use those
1231          * yet either...
1232          *
1233          * Ring buffer needs to be re-initialized in the KMS case, or if X
1234          * was running at the time of the reset (i.e. we weren't VT
1235          * switched away).
1236          */
1237         ret = intel_gt_init_hw(gt);
1238         if (ret) {
1239                 gt_err(gt, "Failed to initialise HW following reset (%d)\n", ret);
1240                 goto taint;
1241         }
1242
1243         ret = resume(gt);
1244         if (ret)
1245                 goto taint;
1246
1247 finish:
1248         reset_finish(gt, awake);
1249 unlock:
1250         mutex_unlock(&gt->reset.mutex);
1251         return;
1252
1253 taint:
1254         /*
1255          * History tells us that if we cannot reset the GPU now, we
1256          * never will. This then impacts everything that is run
1257          * subsequently. On failing the reset, we mark the driver
1258          * as wedged, preventing further execution on the GPU.
1259          * We also want to go one step further and add a taint to the
1260          * kernel so that any subsequent faults can be traced back to
1261          * this failure. This is important for CI, where if the
1262          * GPU/driver fails we would like to reboot and restart testing
1263          * rather than continue on into oblivion. For everyone else,
1264          * the system should still plod along, but they have been warned!
1265          */
1266         add_taint_for_CI(gt->i915, TAINT_WARN);
1267 error:
1268         __intel_gt_set_wedged(gt);
1269         goto finish;
1270 }
1271
1272 static int intel_gt_reset_engine(struct intel_engine_cs *engine)
1273 {
1274         return __intel_gt_reset(engine->gt, engine->mask);
1275 }
1276
1277 int __intel_engine_reset_bh(struct intel_engine_cs *engine, const char *msg)
1278 {
1279         struct intel_gt *gt = engine->gt;
1280         int ret;
1281
1282         ENGINE_TRACE(engine, "flags=%lx\n", gt->reset.flags);
1283         GEM_BUG_ON(!test_bit(I915_RESET_ENGINE + engine->id, &gt->reset.flags));
1284
1285         if (intel_engine_uses_guc(engine))
1286                 return -ENODEV;
1287
1288         if (!intel_engine_pm_get_if_awake(engine))
1289                 return 0;
1290
1291         reset_prepare_engine(engine);
1292
1293         if (msg)
1294                 drm_notice(&engine->i915->drm,
1295                            "Resetting %s for %s\n", engine->name, msg);
1296         i915_increase_reset_engine_count(&engine->i915->gpu_error, engine);
1297
1298         ret = intel_gt_reset_engine(engine);
1299         if (ret) {
1300                 /* If we fail here, we expect to fallback to a global reset */
1301                 ENGINE_TRACE(engine, "Failed to reset %s, err: %d\n", engine->name, ret);
1302                 goto out;
1303         }
1304
1305         /*
1306          * The request that caused the hang is stuck on elsp, we know the
1307          * active request and can drop it, adjust head to skip the offending
1308          * request to resume executing remaining requests in the queue.
1309          */
1310         __intel_engine_reset(engine, true);
1311
1312         /*
1313          * The engine and its registers (and workarounds in case of render)
1314          * have been reset to their default values. Follow the init_ring
1315          * process to program RING_MODE, HWSP and re-enable submission.
1316          */
1317         ret = intel_engine_resume(engine);
1318
1319 out:
1320         intel_engine_cancel_stop_cs(engine);
1321         reset_finish_engine(engine);
1322         intel_engine_pm_put_async(engine);
1323         return ret;
1324 }
1325
1326 /**
1327  * intel_engine_reset - reset GPU engine to recover from a hang
1328  * @engine: engine to reset
1329  * @msg: reason for GPU reset; or NULL for no drm_notice()
1330  *
1331  * Reset a specific GPU engine. Useful if a hang is detected.
1332  * Returns zero on successful reset or otherwise an error code.
1333  *
1334  * Procedure is:
1335  *  - identifies the request that caused the hang and it is dropped
1336  *  - reset engine (which will force the engine to idle)
1337  *  - re-init/configure engine
1338  */
1339 int intel_engine_reset(struct intel_engine_cs *engine, const char *msg)
1340 {
1341         int err;
1342
1343         local_bh_disable();
1344         err = __intel_engine_reset_bh(engine, msg);
1345         local_bh_enable();
1346
1347         return err;
1348 }
1349
1350 static void intel_gt_reset_global(struct intel_gt *gt,
1351                                   u32 engine_mask,
1352                                   const char *reason)
1353 {
1354         struct kobject *kobj = &gt->i915->drm.primary->kdev->kobj;
1355         char *error_event[] = { I915_ERROR_UEVENT "=1", NULL };
1356         char *reset_event[] = { I915_RESET_UEVENT "=1", NULL };
1357         char *reset_done_event[] = { I915_ERROR_UEVENT "=0", NULL };
1358         struct intel_wedge_me w;
1359
1360         kobject_uevent_env(kobj, KOBJ_CHANGE, error_event);
1361
1362         GT_TRACE(gt, "resetting chip, engines=%x\n", engine_mask);
1363         kobject_uevent_env(kobj, KOBJ_CHANGE, reset_event);
1364
1365         /* Use a watchdog to ensure that our reset completes */
1366         intel_wedge_on_timeout(&w, gt, 60 * HZ) {
1367                 intel_display_reset_prepare(gt->i915);
1368
1369                 intel_gt_reset(gt, engine_mask, reason);
1370
1371                 intel_display_reset_finish(gt->i915);
1372         }
1373
1374         if (!test_bit(I915_WEDGED, &gt->reset.flags))
1375                 kobject_uevent_env(kobj, KOBJ_CHANGE, reset_done_event);
1376 }
1377
1378 /**
1379  * intel_gt_handle_error - handle a gpu error
1380  * @gt: the intel_gt
1381  * @engine_mask: mask representing engines that are hung
1382  * @flags: control flags
1383  * @fmt: Error message format string
1384  *
1385  * Do some basic checking of register state at error time and
1386  * dump it to the syslog.  Also call i915_capture_error_state() to make
1387  * sure we get a record and make it available in debugfs.  Fire a uevent
1388  * so userspace knows something bad happened (should trigger collection
1389  * of a ring dump etc.).
1390  */
1391 void intel_gt_handle_error(struct intel_gt *gt,
1392                            intel_engine_mask_t engine_mask,
1393                            unsigned long flags,
1394                            const char *fmt, ...)
1395 {
1396         struct intel_engine_cs *engine;
1397         intel_wakeref_t wakeref;
1398         intel_engine_mask_t tmp;
1399         char error_msg[80];
1400         char *msg = NULL;
1401
1402         if (fmt) {
1403                 va_list args;
1404
1405                 va_start(args, fmt);
1406                 vscnprintf(error_msg, sizeof(error_msg), fmt, args);
1407                 va_end(args);
1408
1409                 msg = error_msg;
1410         }
1411
1412         /*
1413          * In most cases it's guaranteed that we get here with an RPM
1414          * reference held, for example because there is a pending GPU
1415          * request that won't finish until the reset is done. This
1416          * isn't the case at least when we get here by doing a
1417          * simulated reset via debugfs, so get an RPM reference.
1418          */
1419         wakeref = intel_runtime_pm_get(gt->uncore->rpm);
1420
1421         engine_mask &= gt->info.engine_mask;
1422
1423         if (flags & I915_ERROR_CAPTURE) {
1424                 i915_capture_error_state(gt, engine_mask, CORE_DUMP_FLAG_NONE);
1425                 intel_gt_clear_error_registers(gt, engine_mask);
1426         }
1427
1428         /*
1429          * Try engine reset when available. We fall back to full reset if
1430          * single reset fails.
1431          */
1432         if (!intel_uc_uses_guc_submission(&gt->uc) &&
1433             intel_has_reset_engine(gt) && !intel_gt_is_wedged(gt)) {
1434                 local_bh_disable();
1435                 for_each_engine_masked(engine, gt, engine_mask, tmp) {
1436                         BUILD_BUG_ON(I915_RESET_MODESET >= I915_RESET_ENGINE);
1437                         if (test_and_set_bit(I915_RESET_ENGINE + engine->id,
1438                                              &gt->reset.flags))
1439                                 continue;
1440
1441                         if (__intel_engine_reset_bh(engine, msg) == 0)
1442                                 engine_mask &= ~engine->mask;
1443
1444                         clear_and_wake_up_bit(I915_RESET_ENGINE + engine->id,
1445                                               &gt->reset.flags);
1446                 }
1447                 local_bh_enable();
1448         }
1449
1450         if (!engine_mask)
1451                 goto out;
1452
1453         /* Full reset needs the mutex, stop any other user trying to do so. */
1454         if (test_and_set_bit(I915_RESET_BACKOFF, &gt->reset.flags)) {
1455                 wait_event(gt->reset.queue,
1456                            !test_bit(I915_RESET_BACKOFF, &gt->reset.flags));
1457                 goto out; /* piggy-back on the other reset */
1458         }
1459
1460         /* Make sure i915_reset_trylock() sees the I915_RESET_BACKOFF */
1461         synchronize_rcu_expedited();
1462
1463         /*
1464          * Prevent any other reset-engine attempt. We don't do this for GuC
1465          * submission the GuC owns the per-engine reset, not the i915.
1466          */
1467         if (!intel_uc_uses_guc_submission(&gt->uc)) {
1468                 for_each_engine(engine, gt, tmp) {
1469                         while (test_and_set_bit(I915_RESET_ENGINE + engine->id,
1470                                                 &gt->reset.flags))
1471                                 wait_on_bit(&gt->reset.flags,
1472                                             I915_RESET_ENGINE + engine->id,
1473                                             TASK_UNINTERRUPTIBLE);
1474                 }
1475         }
1476
1477         /* Flush everyone using a resource about to be clobbered */
1478         synchronize_srcu_expedited(&gt->reset.backoff_srcu);
1479
1480         intel_gt_reset_global(gt, engine_mask, msg);
1481
1482         if (!intel_uc_uses_guc_submission(&gt->uc)) {
1483                 for_each_engine(engine, gt, tmp)
1484                         clear_bit_unlock(I915_RESET_ENGINE + engine->id,
1485                                          &gt->reset.flags);
1486         }
1487         clear_bit_unlock(I915_RESET_BACKOFF, &gt->reset.flags);
1488         smp_mb__after_atomic();
1489         wake_up_all(&gt->reset.queue);
1490
1491 out:
1492         intel_runtime_pm_put(gt->uncore->rpm, wakeref);
1493 }
1494
1495 static int _intel_gt_reset_lock(struct intel_gt *gt, int *srcu, bool retry)
1496 {
1497         might_lock(&gt->reset.backoff_srcu);
1498         if (retry)
1499                 might_sleep();
1500
1501         rcu_read_lock();
1502         while (test_bit(I915_RESET_BACKOFF, &gt->reset.flags)) {
1503                 rcu_read_unlock();
1504
1505                 if (!retry)
1506                         return -EBUSY;
1507
1508                 if (wait_event_interruptible(gt->reset.queue,
1509                                              !test_bit(I915_RESET_BACKOFF,
1510                                                        &gt->reset.flags)))
1511                         return -EINTR;
1512
1513                 rcu_read_lock();
1514         }
1515         *srcu = srcu_read_lock(&gt->reset.backoff_srcu);
1516         rcu_read_unlock();
1517
1518         return 0;
1519 }
1520
1521 int intel_gt_reset_trylock(struct intel_gt *gt, int *srcu)
1522 {
1523         return _intel_gt_reset_lock(gt, srcu, false);
1524 }
1525
1526 int intel_gt_reset_lock_interruptible(struct intel_gt *gt, int *srcu)
1527 {
1528         return _intel_gt_reset_lock(gt, srcu, true);
1529 }
1530
1531 void intel_gt_reset_unlock(struct intel_gt *gt, int tag)
1532 __releases(&gt->reset.backoff_srcu)
1533 {
1534         srcu_read_unlock(&gt->reset.backoff_srcu, tag);
1535 }
1536
1537 int intel_gt_terminally_wedged(struct intel_gt *gt)
1538 {
1539         might_sleep();
1540
1541         if (!intel_gt_is_wedged(gt))
1542                 return 0;
1543
1544         if (intel_gt_has_unrecoverable_error(gt))
1545                 return -EIO;
1546
1547         /* Reset still in progress? Maybe we will recover? */
1548         if (wait_event_interruptible(gt->reset.queue,
1549                                      !test_bit(I915_RESET_BACKOFF,
1550                                                &gt->reset.flags)))
1551                 return -EINTR;
1552
1553         return intel_gt_is_wedged(gt) ? -EIO : 0;
1554 }
1555
1556 void intel_gt_set_wedged_on_init(struct intel_gt *gt)
1557 {
1558         BUILD_BUG_ON(I915_RESET_ENGINE + I915_NUM_ENGINES >
1559                      I915_WEDGED_ON_INIT);
1560         intel_gt_set_wedged(gt);
1561         i915_disable_error_state(gt->i915, -ENODEV);
1562         set_bit(I915_WEDGED_ON_INIT, &gt->reset.flags);
1563
1564         /* Wedged on init is non-recoverable */
1565         add_taint_for_CI(gt->i915, TAINT_WARN);
1566 }
1567
1568 void intel_gt_set_wedged_on_fini(struct intel_gt *gt)
1569 {
1570         intel_gt_set_wedged(gt);
1571         i915_disable_error_state(gt->i915, -ENODEV);
1572         set_bit(I915_WEDGED_ON_FINI, &gt->reset.flags);
1573         intel_gt_retire_requests(gt); /* cleanup any wedged requests */
1574 }
1575
1576 void intel_gt_init_reset(struct intel_gt *gt)
1577 {
1578         init_waitqueue_head(&gt->reset.queue);
1579         mutex_init(&gt->reset.mutex);
1580         init_srcu_struct(&gt->reset.backoff_srcu);
1581
1582         /*
1583          * While undesirable to wait inside the shrinker, complain anyway.
1584          *
1585          * If we have to wait during shrinking, we guarantee forward progress
1586          * by forcing the reset. Therefore during the reset we must not
1587          * re-enter the shrinker. By declaring that we take the reset mutex
1588          * within the shrinker, we forbid ourselves from performing any
1589          * fs-reclaim or taking related locks during reset.
1590          */
1591         i915_gem_shrinker_taints_mutex(gt->i915, &gt->reset.mutex);
1592
1593         /* no GPU until we are ready! */
1594         __set_bit(I915_WEDGED, &gt->reset.flags);
1595 }
1596
1597 void intel_gt_fini_reset(struct intel_gt *gt)
1598 {
1599         cleanup_srcu_struct(&gt->reset.backoff_srcu);
1600 }
1601
1602 static void intel_wedge_me(struct work_struct *work)
1603 {
1604         struct intel_wedge_me *w = container_of(work, typeof(*w), work.work);
1605
1606         gt_err(w->gt, "%s timed out, cancelling all in-flight rendering.\n", w->name);
1607         intel_gt_set_wedged(w->gt);
1608 }
1609
1610 void __intel_init_wedge(struct intel_wedge_me *w,
1611                         struct intel_gt *gt,
1612                         long timeout,
1613                         const char *name)
1614 {
1615         w->gt = gt;
1616         w->name = name;
1617
1618         INIT_DELAYED_WORK_ONSTACK(&w->work, intel_wedge_me);
1619         queue_delayed_work(gt->i915->unordered_wq, &w->work, timeout);
1620 }
1621
1622 void __intel_fini_wedge(struct intel_wedge_me *w)
1623 {
1624         cancel_delayed_work_sync(&w->work);
1625         destroy_delayed_work_on_stack(&w->work);
1626         w->gt = NULL;
1627 }
1628
1629 /*
1630  * Wa_22011802037 requires that we (or the GuC) ensure that no command
1631  * streamers are executing MI_FORCE_WAKE while an engine reset is initiated.
1632  */
1633 bool intel_engine_reset_needs_wa_22011802037(struct intel_gt *gt)
1634 {
1635         if (GRAPHICS_VER(gt->i915) < 11)
1636                 return false;
1637
1638         if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0))
1639                 return true;
1640
1641         if (GRAPHICS_VER_FULL(gt->i915) >= IP_VER(12, 70))
1642                 return false;
1643
1644         return true;
1645 }
1646
1647 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1648 #include "selftest_reset.c"
1649 #include "selftest_hangcheck.c"
1650 #endif