1 /* 2 * SPDX-License-Identifier: MIT 3 * 4 * Copyright © 2008-2018 Intel Corporation 5 */ 6 7 #include <linux/sched/mm.h> 8 #include <linux/stop_machine.h> 9 10 #include "display/intel_overlay.h" 11 12 #include "gem/i915_gem_context.h" 13 14 #include "i915_drv.h" 15 #include "i915_gpu_error.h" 16 #include "i915_irq.h" 17 #include "intel_engine_pm.h" 18 #include "intel_gt_pm.h" 19 #include "intel_reset.h" 20 21 #include "intel_guc.h" 22 23 #define RESET_MAX_RETRIES 3 24 25 /* XXX How to handle concurrent GGTT updates using tiling registers? */ 26 #define RESET_UNDER_STOP_MACHINE 0 27 28 static void rmw_set(struct intel_uncore *uncore, i915_reg_t reg, u32 set) 29 { 30 intel_uncore_rmw(uncore, reg, 0, set); 31 } 32 33 static void rmw_clear(struct intel_uncore *uncore, i915_reg_t reg, u32 clr) 34 { 35 intel_uncore_rmw(uncore, reg, clr, 0); 36 } 37 38 static void rmw_set_fw(struct intel_uncore *uncore, i915_reg_t reg, u32 set) 39 { 40 intel_uncore_rmw_fw(uncore, reg, 0, set); 41 } 42 43 static void rmw_clear_fw(struct intel_uncore *uncore, i915_reg_t reg, u32 clr) 44 { 45 intel_uncore_rmw_fw(uncore, reg, clr, 0); 46 } 47 48 static void engine_skip_context(struct i915_request *rq) 49 { 50 struct intel_engine_cs *engine = rq->engine; 51 struct i915_gem_context *hung_ctx = rq->gem_context; 52 53 lockdep_assert_held(&engine->active.lock); 54 55 if (!i915_request_is_active(rq)) 56 return; 57 58 list_for_each_entry_continue(rq, &engine->active.requests, sched.link) 59 if (rq->gem_context == hung_ctx) 60 i915_request_skip(rq, -EIO); 61 } 62 63 static void client_mark_guilty(struct drm_i915_file_private *file_priv, 64 const struct i915_gem_context *ctx) 65 { 66 unsigned int score; 67 unsigned long prev_hang; 68 69 if (i915_gem_context_is_banned(ctx)) 70 score = I915_CLIENT_SCORE_CONTEXT_BAN; 71 else 72 score = 0; 73 74 prev_hang = xchg(&file_priv->hang_timestamp, jiffies); 75 if (time_before(jiffies, prev_hang + I915_CLIENT_FAST_HANG_JIFFIES)) 76 score += I915_CLIENT_SCORE_HANG_FAST; 77 78 if (score) { 79 atomic_add(score, &file_priv->ban_score); 80 81 DRM_DEBUG_DRIVER("client %s: gained %u ban score, now %u\n", 82 ctx->name, score, 83 atomic_read(&file_priv->ban_score)); 84 } 85 } 86 87 static bool context_mark_guilty(struct i915_gem_context *ctx) 88 { 89 unsigned long prev_hang; 90 bool banned; 91 int i; 92 93 atomic_inc(&ctx->guilty_count); 94 95 /* Cool contexts are too cool to be banned! (Used for reset testing.) */ 96 if (!i915_gem_context_is_bannable(ctx)) 97 return false; 98 99 /* Record the timestamp for the last N hangs */ 100 prev_hang = ctx->hang_timestamp[0]; 101 for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp) - 1; i++) 102 ctx->hang_timestamp[i] = ctx->hang_timestamp[i + 1]; 103 ctx->hang_timestamp[i] = jiffies; 104 105 /* If we have hung N+1 times in rapid succession, we ban the context! */ 106 banned = !i915_gem_context_is_recoverable(ctx); 107 if (time_before(jiffies, prev_hang + CONTEXT_FAST_HANG_JIFFIES)) 108 banned = true; 109 if (banned) { 110 DRM_DEBUG_DRIVER("context %s: guilty %d, banned\n", 111 ctx->name, atomic_read(&ctx->guilty_count)); 112 i915_gem_context_set_banned(ctx); 113 } 114 115 if (!IS_ERR_OR_NULL(ctx->file_priv)) 116 client_mark_guilty(ctx->file_priv, ctx); 117 118 return banned; 119 } 120 121 static void context_mark_innocent(struct i915_gem_context *ctx) 122 { 123 atomic_inc(&ctx->active_count); 124 } 125 126 void i915_reset_request(struct i915_request *rq, bool guilty) 127 { 128 GEM_TRACE("%s rq=%llx:%lld, guilty? %s\n", 129 rq->engine->name, 130 rq->fence.context, 131 rq->fence.seqno, 132 yesno(guilty)); 133 134 lockdep_assert_held(&rq->engine->active.lock); 135 GEM_BUG_ON(i915_request_completed(rq)); 136 137 if (guilty) { 138 i915_request_skip(rq, -EIO); 139 if (context_mark_guilty(rq->gem_context)) 140 engine_skip_context(rq); 141 } else { 142 dma_fence_set_error(&rq->fence, -EAGAIN); 143 context_mark_innocent(rq->gem_context); 144 } 145 } 146 147 static void gen3_stop_engine(struct intel_engine_cs *engine) 148 { 149 struct intel_uncore *uncore = engine->uncore; 150 const u32 base = engine->mmio_base; 151 152 GEM_TRACE("%s\n", engine->name); 153 154 if (intel_engine_stop_cs(engine)) 155 GEM_TRACE("%s: timed out on STOP_RING\n", engine->name); 156 157 intel_uncore_write_fw(uncore, 158 RING_HEAD(base), 159 intel_uncore_read_fw(uncore, RING_TAIL(base))); 160 intel_uncore_posting_read_fw(uncore, RING_HEAD(base)); /* paranoia */ 161 162 intel_uncore_write_fw(uncore, RING_HEAD(base), 0); 163 intel_uncore_write_fw(uncore, RING_TAIL(base), 0); 164 intel_uncore_posting_read_fw(uncore, RING_TAIL(base)); 165 166 /* The ring must be empty before it is disabled */ 167 intel_uncore_write_fw(uncore, RING_CTL(base), 0); 168 169 /* Check acts as a post */ 170 if (intel_uncore_read_fw(uncore, RING_HEAD(base))) 171 GEM_TRACE("%s: ring head [%x] not parked\n", 172 engine->name, 173 intel_uncore_read_fw(uncore, RING_HEAD(base))); 174 } 175 176 static void i915_stop_engines(struct drm_i915_private *i915, 177 intel_engine_mask_t engine_mask) 178 { 179 struct intel_engine_cs *engine; 180 intel_engine_mask_t tmp; 181 182 if (INTEL_GEN(i915) < 3) 183 return; 184 185 for_each_engine_masked(engine, i915, engine_mask, tmp) 186 gen3_stop_engine(engine); 187 } 188 189 static bool i915_in_reset(struct pci_dev *pdev) 190 { 191 u8 gdrst; 192 193 pci_read_config_byte(pdev, I915_GDRST, &gdrst); 194 return gdrst & GRDOM_RESET_STATUS; 195 } 196 197 static int i915_do_reset(struct drm_i915_private *i915, 198 intel_engine_mask_t engine_mask, 199 unsigned int retry) 200 { 201 struct pci_dev *pdev = i915->drm.pdev; 202 int err; 203 204 /* Assert reset for at least 20 usec, and wait for acknowledgement. */ 205 pci_write_config_byte(pdev, I915_GDRST, GRDOM_RESET_ENABLE); 206 udelay(50); 207 err = wait_for_atomic(i915_in_reset(pdev), 50); 208 209 /* Clear the reset request. */ 210 pci_write_config_byte(pdev, I915_GDRST, 0); 211 udelay(50); 212 if (!err) 213 err = wait_for_atomic(!i915_in_reset(pdev), 50); 214 215 return err; 216 } 217 218 static bool g4x_reset_complete(struct pci_dev *pdev) 219 { 220 u8 gdrst; 221 222 pci_read_config_byte(pdev, I915_GDRST, &gdrst); 223 return (gdrst & GRDOM_RESET_ENABLE) == 0; 224 } 225 226 static int g33_do_reset(struct drm_i915_private *i915, 227 intel_engine_mask_t engine_mask, 228 unsigned int retry) 229 { 230 struct pci_dev *pdev = i915->drm.pdev; 231 232 pci_write_config_byte(pdev, I915_GDRST, GRDOM_RESET_ENABLE); 233 return wait_for_atomic(g4x_reset_complete(pdev), 50); 234 } 235 236 static int g4x_do_reset(struct drm_i915_private *i915, 237 intel_engine_mask_t engine_mask, 238 unsigned int retry) 239 { 240 struct pci_dev *pdev = i915->drm.pdev; 241 struct intel_uncore *uncore = &i915->uncore; 242 int ret; 243 244 /* WaVcpClkGateDisableForMediaReset:ctg,elk */ 245 rmw_set_fw(uncore, VDECCLK_GATE_D, VCP_UNIT_CLOCK_GATE_DISABLE); 246 intel_uncore_posting_read_fw(uncore, VDECCLK_GATE_D); 247 248 pci_write_config_byte(pdev, I915_GDRST, 249 GRDOM_MEDIA | GRDOM_RESET_ENABLE); 250 ret = wait_for_atomic(g4x_reset_complete(pdev), 50); 251 if (ret) { 252 DRM_DEBUG_DRIVER("Wait for media reset failed\n"); 253 goto out; 254 } 255 256 pci_write_config_byte(pdev, I915_GDRST, 257 GRDOM_RENDER | GRDOM_RESET_ENABLE); 258 ret = wait_for_atomic(g4x_reset_complete(pdev), 50); 259 if (ret) { 260 DRM_DEBUG_DRIVER("Wait for render reset failed\n"); 261 goto out; 262 } 263 264 out: 265 pci_write_config_byte(pdev, I915_GDRST, 0); 266 267 rmw_clear_fw(uncore, VDECCLK_GATE_D, VCP_UNIT_CLOCK_GATE_DISABLE); 268 intel_uncore_posting_read_fw(uncore, VDECCLK_GATE_D); 269 270 return ret; 271 } 272 273 static int ironlake_do_reset(struct drm_i915_private *i915, 274 intel_engine_mask_t engine_mask, 275 unsigned int retry) 276 { 277 struct intel_uncore *uncore = &i915->uncore; 278 int ret; 279 280 intel_uncore_write_fw(uncore, ILK_GDSR, 281 ILK_GRDOM_RENDER | ILK_GRDOM_RESET_ENABLE); 282 ret = __intel_wait_for_register_fw(uncore, ILK_GDSR, 283 ILK_GRDOM_RESET_ENABLE, 0, 284 5000, 0, 285 NULL); 286 if (ret) { 287 DRM_DEBUG_DRIVER("Wait for render reset failed\n"); 288 goto out; 289 } 290 291 intel_uncore_write_fw(uncore, ILK_GDSR, 292 ILK_GRDOM_MEDIA | ILK_GRDOM_RESET_ENABLE); 293 ret = __intel_wait_for_register_fw(uncore, ILK_GDSR, 294 ILK_GRDOM_RESET_ENABLE, 0, 295 5000, 0, 296 NULL); 297 if (ret) { 298 DRM_DEBUG_DRIVER("Wait for media reset failed\n"); 299 goto out; 300 } 301 302 out: 303 intel_uncore_write_fw(uncore, ILK_GDSR, 0); 304 intel_uncore_posting_read_fw(uncore, ILK_GDSR); 305 return ret; 306 } 307 308 /* Reset the hardware domains (GENX_GRDOM_*) specified by mask */ 309 static int gen6_hw_domain_reset(struct drm_i915_private *i915, 310 u32 hw_domain_mask) 311 { 312 struct intel_uncore *uncore = &i915->uncore; 313 int err; 314 315 /* 316 * GEN6_GDRST is not in the gt power well, no need to check 317 * for fifo space for the write or forcewake the chip for 318 * the read 319 */ 320 intel_uncore_write_fw(uncore, GEN6_GDRST, hw_domain_mask); 321 322 /* Wait for the device to ack the reset requests */ 323 err = __intel_wait_for_register_fw(uncore, 324 GEN6_GDRST, hw_domain_mask, 0, 325 500, 0, 326 NULL); 327 if (err) 328 DRM_DEBUG_DRIVER("Wait for 0x%08x engines reset failed\n", 329 hw_domain_mask); 330 331 return err; 332 } 333 334 static int gen6_reset_engines(struct drm_i915_private *i915, 335 intel_engine_mask_t engine_mask, 336 unsigned int retry) 337 { 338 struct intel_engine_cs *engine; 339 const u32 hw_engine_mask[] = { 340 [RCS0] = GEN6_GRDOM_RENDER, 341 [BCS0] = GEN6_GRDOM_BLT, 342 [VCS0] = GEN6_GRDOM_MEDIA, 343 [VCS1] = GEN8_GRDOM_MEDIA2, 344 [VECS0] = GEN6_GRDOM_VECS, 345 }; 346 u32 hw_mask; 347 348 if (engine_mask == ALL_ENGINES) { 349 hw_mask = GEN6_GRDOM_FULL; 350 } else { 351 intel_engine_mask_t tmp; 352 353 hw_mask = 0; 354 for_each_engine_masked(engine, i915, engine_mask, tmp) { 355 GEM_BUG_ON(engine->id >= ARRAY_SIZE(hw_engine_mask)); 356 hw_mask |= hw_engine_mask[engine->id]; 357 } 358 } 359 360 return gen6_hw_domain_reset(i915, hw_mask); 361 } 362 363 static u32 gen11_lock_sfc(struct intel_engine_cs *engine) 364 { 365 struct intel_uncore *uncore = engine->uncore; 366 u8 vdbox_sfc_access = RUNTIME_INFO(engine->i915)->vdbox_sfc_access; 367 i915_reg_t sfc_forced_lock, sfc_forced_lock_ack; 368 u32 sfc_forced_lock_bit, sfc_forced_lock_ack_bit; 369 i915_reg_t sfc_usage; 370 u32 sfc_usage_bit; 371 u32 sfc_reset_bit; 372 373 switch (engine->class) { 374 case VIDEO_DECODE_CLASS: 375 if ((BIT(engine->instance) & vdbox_sfc_access) == 0) 376 return 0; 377 378 sfc_forced_lock = GEN11_VCS_SFC_FORCED_LOCK(engine); 379 sfc_forced_lock_bit = GEN11_VCS_SFC_FORCED_LOCK_BIT; 380 381 sfc_forced_lock_ack = GEN11_VCS_SFC_LOCK_STATUS(engine); 382 sfc_forced_lock_ack_bit = GEN11_VCS_SFC_LOCK_ACK_BIT; 383 384 sfc_usage = GEN11_VCS_SFC_LOCK_STATUS(engine); 385 sfc_usage_bit = GEN11_VCS_SFC_USAGE_BIT; 386 sfc_reset_bit = GEN11_VCS_SFC_RESET_BIT(engine->instance); 387 break; 388 389 case VIDEO_ENHANCEMENT_CLASS: 390 sfc_forced_lock = GEN11_VECS_SFC_FORCED_LOCK(engine); 391 sfc_forced_lock_bit = GEN11_VECS_SFC_FORCED_LOCK_BIT; 392 393 sfc_forced_lock_ack = GEN11_VECS_SFC_LOCK_ACK(engine); 394 sfc_forced_lock_ack_bit = GEN11_VECS_SFC_LOCK_ACK_BIT; 395 396 sfc_usage = GEN11_VECS_SFC_USAGE(engine); 397 sfc_usage_bit = GEN11_VECS_SFC_USAGE_BIT; 398 sfc_reset_bit = GEN11_VECS_SFC_RESET_BIT(engine->instance); 399 break; 400 401 default: 402 return 0; 403 } 404 405 /* 406 * Tell the engine that a software reset is going to happen. The engine 407 * will then try to force lock the SFC (if currently locked, it will 408 * remain so until we tell the engine it is safe to unlock; if currently 409 * unlocked, it will ignore this and all new lock requests). If SFC 410 * ends up being locked to the engine we want to reset, we have to reset 411 * it as well (we will unlock it once the reset sequence is completed). 412 */ 413 rmw_set_fw(uncore, sfc_forced_lock, sfc_forced_lock_bit); 414 415 if (__intel_wait_for_register_fw(uncore, 416 sfc_forced_lock_ack, 417 sfc_forced_lock_ack_bit, 418 sfc_forced_lock_ack_bit, 419 1000, 0, NULL)) { 420 DRM_DEBUG_DRIVER("Wait for SFC forced lock ack failed\n"); 421 return 0; 422 } 423 424 if (intel_uncore_read_fw(uncore, sfc_usage) & sfc_usage_bit) 425 return sfc_reset_bit; 426 427 return 0; 428 } 429 430 static void gen11_unlock_sfc(struct intel_engine_cs *engine) 431 { 432 struct intel_uncore *uncore = engine->uncore; 433 u8 vdbox_sfc_access = RUNTIME_INFO(engine->i915)->vdbox_sfc_access; 434 i915_reg_t sfc_forced_lock; 435 u32 sfc_forced_lock_bit; 436 437 switch (engine->class) { 438 case VIDEO_DECODE_CLASS: 439 if ((BIT(engine->instance) & vdbox_sfc_access) == 0) 440 return; 441 442 sfc_forced_lock = GEN11_VCS_SFC_FORCED_LOCK(engine); 443 sfc_forced_lock_bit = GEN11_VCS_SFC_FORCED_LOCK_BIT; 444 break; 445 446 case VIDEO_ENHANCEMENT_CLASS: 447 sfc_forced_lock = GEN11_VECS_SFC_FORCED_LOCK(engine); 448 sfc_forced_lock_bit = GEN11_VECS_SFC_FORCED_LOCK_BIT; 449 break; 450 451 default: 452 return; 453 } 454 455 rmw_clear_fw(uncore, sfc_forced_lock, sfc_forced_lock_bit); 456 } 457 458 static int gen11_reset_engines(struct drm_i915_private *i915, 459 intel_engine_mask_t engine_mask, 460 unsigned int retry) 461 { 462 const u32 hw_engine_mask[] = { 463 [RCS0] = GEN11_GRDOM_RENDER, 464 [BCS0] = GEN11_GRDOM_BLT, 465 [VCS0] = GEN11_GRDOM_MEDIA, 466 [VCS1] = GEN11_GRDOM_MEDIA2, 467 [VCS2] = GEN11_GRDOM_MEDIA3, 468 [VCS3] = GEN11_GRDOM_MEDIA4, 469 [VECS0] = GEN11_GRDOM_VECS, 470 [VECS1] = GEN11_GRDOM_VECS2, 471 }; 472 struct intel_engine_cs *engine; 473 intel_engine_mask_t tmp; 474 u32 hw_mask; 475 int ret; 476 477 if (engine_mask == ALL_ENGINES) { 478 hw_mask = GEN11_GRDOM_FULL; 479 } else { 480 hw_mask = 0; 481 for_each_engine_masked(engine, i915, engine_mask, tmp) { 482 GEM_BUG_ON(engine->id >= ARRAY_SIZE(hw_engine_mask)); 483 hw_mask |= hw_engine_mask[engine->id]; 484 hw_mask |= gen11_lock_sfc(engine); 485 } 486 } 487 488 ret = gen6_hw_domain_reset(i915, hw_mask); 489 490 if (engine_mask != ALL_ENGINES) 491 for_each_engine_masked(engine, i915, engine_mask, tmp) 492 gen11_unlock_sfc(engine); 493 494 return ret; 495 } 496 497 static int gen8_engine_reset_prepare(struct intel_engine_cs *engine) 498 { 499 struct intel_uncore *uncore = engine->uncore; 500 const i915_reg_t reg = RING_RESET_CTL(engine->mmio_base); 501 u32 request, mask, ack; 502 int ret; 503 504 ack = intel_uncore_read_fw(uncore, reg); 505 if (ack & RESET_CTL_CAT_ERROR) { 506 /* 507 * For catastrophic errors, ready-for-reset sequence 508 * needs to be bypassed: HAS#396813 509 */ 510 request = RESET_CTL_CAT_ERROR; 511 mask = RESET_CTL_CAT_ERROR; 512 513 /* Catastrophic errors need to be cleared by HW */ 514 ack = 0; 515 } else if (!(ack & RESET_CTL_READY_TO_RESET)) { 516 request = RESET_CTL_REQUEST_RESET; 517 mask = RESET_CTL_READY_TO_RESET; 518 ack = RESET_CTL_READY_TO_RESET; 519 } else { 520 return 0; 521 } 522 523 intel_uncore_write_fw(uncore, reg, _MASKED_BIT_ENABLE(request)); 524 ret = __intel_wait_for_register_fw(uncore, reg, mask, ack, 525 700, 0, NULL); 526 if (ret) 527 DRM_ERROR("%s reset request timed out: {request: %08x, RESET_CTL: %08x}\n", 528 engine->name, request, 529 intel_uncore_read_fw(uncore, reg)); 530 531 return ret; 532 } 533 534 static void gen8_engine_reset_cancel(struct intel_engine_cs *engine) 535 { 536 intel_uncore_write_fw(engine->uncore, 537 RING_RESET_CTL(engine->mmio_base), 538 _MASKED_BIT_DISABLE(RESET_CTL_REQUEST_RESET)); 539 } 540 541 static int gen8_reset_engines(struct drm_i915_private *i915, 542 intel_engine_mask_t engine_mask, 543 unsigned int retry) 544 { 545 struct intel_engine_cs *engine; 546 const bool reset_non_ready = retry >= 1; 547 intel_engine_mask_t tmp; 548 int ret; 549 550 for_each_engine_masked(engine, i915, engine_mask, tmp) { 551 ret = gen8_engine_reset_prepare(engine); 552 if (ret && !reset_non_ready) 553 goto skip_reset; 554 555 /* 556 * If this is not the first failed attempt to prepare, 557 * we decide to proceed anyway. 558 * 559 * By doing so we risk context corruption and with 560 * some gens (kbl), possible system hang if reset 561 * happens during active bb execution. 562 * 563 * We rather take context corruption instead of 564 * failed reset with a wedged driver/gpu. And 565 * active bb execution case should be covered by 566 * i915_stop_engines we have before the reset. 567 */ 568 } 569 570 if (INTEL_GEN(i915) >= 11) 571 ret = gen11_reset_engines(i915, engine_mask, retry); 572 else 573 ret = gen6_reset_engines(i915, engine_mask, retry); 574 575 skip_reset: 576 for_each_engine_masked(engine, i915, engine_mask, tmp) 577 gen8_engine_reset_cancel(engine); 578 579 return ret; 580 } 581 582 typedef int (*reset_func)(struct drm_i915_private *, 583 intel_engine_mask_t engine_mask, 584 unsigned int retry); 585 586 static reset_func intel_get_gpu_reset(struct drm_i915_private *i915) 587 { 588 if (INTEL_GEN(i915) >= 8) 589 return gen8_reset_engines; 590 else if (INTEL_GEN(i915) >= 6) 591 return gen6_reset_engines; 592 else if (INTEL_GEN(i915) >= 5) 593 return ironlake_do_reset; 594 else if (IS_G4X(i915)) 595 return g4x_do_reset; 596 else if (IS_G33(i915) || IS_PINEVIEW(i915)) 597 return g33_do_reset; 598 else if (INTEL_GEN(i915) >= 3) 599 return i915_do_reset; 600 else 601 return NULL; 602 } 603 604 int intel_gpu_reset(struct drm_i915_private *i915, 605 intel_engine_mask_t engine_mask) 606 { 607 const int retries = engine_mask == ALL_ENGINES ? RESET_MAX_RETRIES : 1; 608 reset_func reset; 609 int ret = -ETIMEDOUT; 610 int retry; 611 612 reset = intel_get_gpu_reset(i915); 613 if (!reset) 614 return -ENODEV; 615 616 /* 617 * If the power well sleeps during the reset, the reset 618 * request may be dropped and never completes (causing -EIO). 619 */ 620 intel_uncore_forcewake_get(&i915->uncore, FORCEWAKE_ALL); 621 for (retry = 0; ret == -ETIMEDOUT && retry < retries; retry++) { 622 /* 623 * We stop engines, otherwise we might get failed reset and a 624 * dead gpu (on elk). Also as modern gpu as kbl can suffer 625 * from system hang if batchbuffer is progressing when 626 * the reset is issued, regardless of READY_TO_RESET ack. 627 * Thus assume it is best to stop engines on all gens 628 * where we have a gpu reset. 629 * 630 * WaKBLVECSSemaphoreWaitPoll:kbl (on ALL_ENGINES) 631 * 632 * WaMediaResetMainRingCleanup:ctg,elk (presumably) 633 * 634 * FIXME: Wa for more modern gens needs to be validated 635 */ 636 if (retry) 637 i915_stop_engines(i915, engine_mask); 638 639 GEM_TRACE("engine_mask=%x\n", engine_mask); 640 preempt_disable(); 641 ret = reset(i915, engine_mask, retry); 642 preempt_enable(); 643 } 644 intel_uncore_forcewake_put(&i915->uncore, FORCEWAKE_ALL); 645 646 return ret; 647 } 648 649 bool intel_has_gpu_reset(struct drm_i915_private *i915) 650 { 651 if (!i915_modparams.reset) 652 return NULL; 653 654 return intel_get_gpu_reset(i915); 655 } 656 657 bool intel_has_reset_engine(struct drm_i915_private *i915) 658 { 659 return INTEL_INFO(i915)->has_reset_engine && i915_modparams.reset >= 2; 660 } 661 662 int intel_reset_guc(struct drm_i915_private *i915) 663 { 664 u32 guc_domain = 665 INTEL_GEN(i915) >= 11 ? GEN11_GRDOM_GUC : GEN9_GRDOM_GUC; 666 int ret; 667 668 GEM_BUG_ON(!HAS_GUC(i915)); 669 670 intel_uncore_forcewake_get(&i915->uncore, FORCEWAKE_ALL); 671 ret = gen6_hw_domain_reset(i915, guc_domain); 672 intel_uncore_forcewake_put(&i915->uncore, FORCEWAKE_ALL); 673 674 return ret; 675 } 676 677 /* 678 * Ensure irq handler finishes, and not run again. 679 * Also return the active request so that we only search for it once. 680 */ 681 static void reset_prepare_engine(struct intel_engine_cs *engine) 682 { 683 /* 684 * During the reset sequence, we must prevent the engine from 685 * entering RC6. As the context state is undefined until we restart 686 * the engine, if it does enter RC6 during the reset, the state 687 * written to the powercontext is undefined and so we may lose 688 * GPU state upon resume, i.e. fail to restart after a reset. 689 */ 690 intel_engine_pm_get(engine); 691 intel_uncore_forcewake_get(engine->uncore, FORCEWAKE_ALL); 692 engine->reset.prepare(engine); 693 } 694 695 static void revoke_mmaps(struct drm_i915_private *i915) 696 { 697 int i; 698 699 for (i = 0; i < i915->ggtt.num_fences; i++) { 700 struct drm_vma_offset_node *node; 701 struct i915_vma *vma; 702 u64 vma_offset; 703 704 vma = READ_ONCE(i915->ggtt.fence_regs[i].vma); 705 if (!vma) 706 continue; 707 708 if (!i915_vma_has_userfault(vma)) 709 continue; 710 711 GEM_BUG_ON(vma->fence != &i915->ggtt.fence_regs[i]); 712 node = &vma->obj->base.vma_node; 713 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT; 714 unmap_mapping_range(i915->drm.anon_inode->i_mapping, 715 drm_vma_node_offset_addr(node) + vma_offset, 716 vma->size, 717 1); 718 } 719 } 720 721 static void reset_prepare(struct drm_i915_private *i915) 722 { 723 struct intel_engine_cs *engine; 724 enum intel_engine_id id; 725 726 intel_gt_pm_get(i915); 727 for_each_engine(engine, i915, id) 728 reset_prepare_engine(engine); 729 730 intel_uc_reset_prepare(i915); 731 } 732 733 static void gt_revoke(struct drm_i915_private *i915) 734 { 735 revoke_mmaps(i915); 736 } 737 738 static int gt_reset(struct drm_i915_private *i915, 739 intel_engine_mask_t stalled_mask) 740 { 741 struct intel_engine_cs *engine; 742 enum intel_engine_id id; 743 int err; 744 745 /* 746 * Everything depends on having the GTT running, so we need to start 747 * there. 748 */ 749 err = i915_ggtt_enable_hw(i915); 750 if (err) 751 return err; 752 753 for_each_engine(engine, i915, id) 754 intel_engine_reset(engine, stalled_mask & engine->mask); 755 756 i915_gem_restore_fences(i915); 757 758 return err; 759 } 760 761 static void reset_finish_engine(struct intel_engine_cs *engine) 762 { 763 engine->reset.finish(engine); 764 intel_engine_pm_put(engine); 765 intel_uncore_forcewake_put(engine->uncore, FORCEWAKE_ALL); 766 } 767 768 static void reset_finish(struct drm_i915_private *i915) 769 { 770 struct intel_engine_cs *engine; 771 enum intel_engine_id id; 772 773 for_each_engine(engine, i915, id) { 774 reset_finish_engine(engine); 775 intel_engine_signal_breadcrumbs(engine); 776 } 777 intel_gt_pm_put(i915); 778 } 779 780 static void nop_submit_request(struct i915_request *request) 781 { 782 struct intel_engine_cs *engine = request->engine; 783 unsigned long flags; 784 785 GEM_TRACE("%s fence %llx:%lld -> -EIO\n", 786 engine->name, request->fence.context, request->fence.seqno); 787 dma_fence_set_error(&request->fence, -EIO); 788 789 spin_lock_irqsave(&engine->active.lock, flags); 790 __i915_request_submit(request); 791 i915_request_mark_complete(request); 792 spin_unlock_irqrestore(&engine->active.lock, flags); 793 794 intel_engine_queue_breadcrumbs(engine); 795 } 796 797 static void __i915_gem_set_wedged(struct drm_i915_private *i915) 798 { 799 struct i915_gpu_error *error = &i915->gpu_error; 800 struct intel_engine_cs *engine; 801 enum intel_engine_id id; 802 803 if (test_bit(I915_WEDGED, &error->flags)) 804 return; 805 806 if (GEM_SHOW_DEBUG() && !intel_engines_are_idle(i915)) { 807 struct drm_printer p = drm_debug_printer(__func__); 808 809 for_each_engine(engine, i915, id) 810 intel_engine_dump(engine, &p, "%s\n", engine->name); 811 } 812 813 GEM_TRACE("start\n"); 814 815 /* 816 * First, stop submission to hw, but do not yet complete requests by 817 * rolling the global seqno forward (since this would complete requests 818 * for which we haven't set the fence error to EIO yet). 819 */ 820 reset_prepare(i915); 821 822 /* Even if the GPU reset fails, it should still stop the engines */ 823 if (!INTEL_INFO(i915)->gpu_reset_clobbers_display) 824 intel_gpu_reset(i915, ALL_ENGINES); 825 826 for_each_engine(engine, i915, id) { 827 engine->submit_request = nop_submit_request; 828 engine->schedule = NULL; 829 } 830 i915->caps.scheduler = 0; 831 832 /* 833 * Make sure no request can slip through without getting completed by 834 * either this call here to intel_engine_write_global_seqno, or the one 835 * in nop_submit_request. 836 */ 837 synchronize_rcu_expedited(); 838 set_bit(I915_WEDGED, &error->flags); 839 840 /* Mark all executing requests as skipped */ 841 for_each_engine(engine, i915, id) 842 engine->cancel_requests(engine); 843 844 reset_finish(i915); 845 846 GEM_TRACE("end\n"); 847 } 848 849 void i915_gem_set_wedged(struct drm_i915_private *i915) 850 { 851 struct i915_gpu_error *error = &i915->gpu_error; 852 intel_wakeref_t wakeref; 853 854 mutex_lock(&error->wedge_mutex); 855 with_intel_runtime_pm(&i915->runtime_pm, wakeref) 856 __i915_gem_set_wedged(i915); 857 mutex_unlock(&error->wedge_mutex); 858 } 859 860 static bool __i915_gem_unset_wedged(struct drm_i915_private *i915) 861 { 862 struct i915_gpu_error *error = &i915->gpu_error; 863 struct i915_timeline *tl; 864 865 if (!test_bit(I915_WEDGED, &error->flags)) 866 return true; 867 868 if (!i915->gt.scratch) /* Never full initialised, recovery impossible */ 869 return false; 870 871 GEM_TRACE("start\n"); 872 873 /* 874 * Before unwedging, make sure that all pending operations 875 * are flushed and errored out - we may have requests waiting upon 876 * third party fences. We marked all inflight requests as EIO, and 877 * every execbuf since returned EIO, for consistency we want all 878 * the currently pending requests to also be marked as EIO, which 879 * is done inside our nop_submit_request - and so we must wait. 880 * 881 * No more can be submitted until we reset the wedged bit. 882 */ 883 mutex_lock(&i915->gt.timelines.mutex); 884 list_for_each_entry(tl, &i915->gt.timelines.active_list, link) { 885 struct i915_request *rq; 886 887 rq = i915_active_request_get_unlocked(&tl->last_request); 888 if (!rq) 889 continue; 890 891 /* 892 * All internal dependencies (i915_requests) will have 893 * been flushed by the set-wedge, but we may be stuck waiting 894 * for external fences. These should all be capped to 10s 895 * (I915_FENCE_TIMEOUT) so this wait should not be unbounded 896 * in the worst case. 897 */ 898 dma_fence_default_wait(&rq->fence, false, MAX_SCHEDULE_TIMEOUT); 899 i915_request_put(rq); 900 } 901 mutex_unlock(&i915->gt.timelines.mutex); 902 903 intel_gt_sanitize(i915, false); 904 905 /* 906 * Undo nop_submit_request. We prevent all new i915 requests from 907 * being queued (by disallowing execbuf whilst wedged) so having 908 * waited for all active requests above, we know the system is idle 909 * and do not have to worry about a thread being inside 910 * engine->submit_request() as we swap over. So unlike installing 911 * the nop_submit_request on reset, we can do this from normal 912 * context and do not require stop_machine(). 913 */ 914 intel_engines_reset_default_submission(i915); 915 916 GEM_TRACE("end\n"); 917 918 smp_mb__before_atomic(); /* complete takeover before enabling execbuf */ 919 clear_bit(I915_WEDGED, &i915->gpu_error.flags); 920 921 return true; 922 } 923 924 bool i915_gem_unset_wedged(struct drm_i915_private *i915) 925 { 926 struct i915_gpu_error *error = &i915->gpu_error; 927 bool result; 928 929 mutex_lock(&error->wedge_mutex); 930 result = __i915_gem_unset_wedged(i915); 931 mutex_unlock(&error->wedge_mutex); 932 933 return result; 934 } 935 936 static int do_reset(struct drm_i915_private *i915, 937 intel_engine_mask_t stalled_mask) 938 { 939 int err, i; 940 941 gt_revoke(i915); 942 943 err = intel_gpu_reset(i915, ALL_ENGINES); 944 for (i = 0; err && i < RESET_MAX_RETRIES; i++) { 945 msleep(10 * (i + 1)); 946 err = intel_gpu_reset(i915, ALL_ENGINES); 947 } 948 if (err) 949 return err; 950 951 return gt_reset(i915, stalled_mask); 952 } 953 954 /** 955 * i915_reset - reset chip after a hang 956 * @i915: #drm_i915_private to reset 957 * @stalled_mask: mask of the stalled engines with the guilty requests 958 * @reason: user error message for why we are resetting 959 * 960 * Reset the chip. Useful if a hang is detected. Marks the device as wedged 961 * on failure. 962 * 963 * Procedure is fairly simple: 964 * - reset the chip using the reset reg 965 * - re-init context state 966 * - re-init hardware status page 967 * - re-init ring buffer 968 * - re-init interrupt state 969 * - re-init display 970 */ 971 void i915_reset(struct drm_i915_private *i915, 972 intel_engine_mask_t stalled_mask, 973 const char *reason) 974 { 975 struct i915_gpu_error *error = &i915->gpu_error; 976 int ret; 977 978 GEM_TRACE("flags=%lx\n", error->flags); 979 980 might_sleep(); 981 GEM_BUG_ON(!test_bit(I915_RESET_BACKOFF, &error->flags)); 982 mutex_lock(&error->wedge_mutex); 983 984 /* Clear any previous failed attempts at recovery. Time to try again. */ 985 if (!__i915_gem_unset_wedged(i915)) 986 goto unlock; 987 988 if (reason) 989 dev_notice(i915->drm.dev, "Resetting chip for %s\n", reason); 990 error->reset_count++; 991 992 reset_prepare(i915); 993 994 if (!intel_has_gpu_reset(i915)) { 995 if (i915_modparams.reset) 996 dev_err(i915->drm.dev, "GPU reset not supported\n"); 997 else 998 DRM_DEBUG_DRIVER("GPU reset disabled\n"); 999 goto error; 1000 } 1001 1002 if (INTEL_INFO(i915)->gpu_reset_clobbers_display) 1003 intel_runtime_pm_disable_interrupts(i915); 1004 1005 if (do_reset(i915, stalled_mask)) { 1006 dev_err(i915->drm.dev, "Failed to reset chip\n"); 1007 goto taint; 1008 } 1009 1010 if (INTEL_INFO(i915)->gpu_reset_clobbers_display) 1011 intel_runtime_pm_enable_interrupts(i915); 1012 1013 intel_overlay_reset(i915); 1014 1015 /* 1016 * Next we need to restore the context, but we don't use those 1017 * yet either... 1018 * 1019 * Ring buffer needs to be re-initialized in the KMS case, or if X 1020 * was running at the time of the reset (i.e. we weren't VT 1021 * switched away). 1022 */ 1023 ret = i915_gem_init_hw(i915); 1024 if (ret) { 1025 DRM_ERROR("Failed to initialise HW following reset (%d)\n", 1026 ret); 1027 goto error; 1028 } 1029 1030 i915_queue_hangcheck(i915); 1031 1032 finish: 1033 reset_finish(i915); 1034 unlock: 1035 mutex_unlock(&error->wedge_mutex); 1036 return; 1037 1038 taint: 1039 /* 1040 * History tells us that if we cannot reset the GPU now, we 1041 * never will. This then impacts everything that is run 1042 * subsequently. On failing the reset, we mark the driver 1043 * as wedged, preventing further execution on the GPU. 1044 * We also want to go one step further and add a taint to the 1045 * kernel so that any subsequent faults can be traced back to 1046 * this failure. This is important for CI, where if the 1047 * GPU/driver fails we would like to reboot and restart testing 1048 * rather than continue on into oblivion. For everyone else, 1049 * the system should still plod along, but they have been warned! 1050 */ 1051 add_taint_for_CI(TAINT_WARN); 1052 error: 1053 __i915_gem_set_wedged(i915); 1054 goto finish; 1055 } 1056 1057 static inline int intel_gt_reset_engine(struct drm_i915_private *i915, 1058 struct intel_engine_cs *engine) 1059 { 1060 return intel_gpu_reset(i915, engine->mask); 1061 } 1062 1063 /** 1064 * i915_reset_engine - reset GPU engine to recover from a hang 1065 * @engine: engine to reset 1066 * @msg: reason for GPU reset; or NULL for no dev_notice() 1067 * 1068 * Reset a specific GPU engine. Useful if a hang is detected. 1069 * Returns zero on successful reset or otherwise an error code. 1070 * 1071 * Procedure is: 1072 * - identifies the request that caused the hang and it is dropped 1073 * - reset engine (which will force the engine to idle) 1074 * - re-init/configure engine 1075 */ 1076 int i915_reset_engine(struct intel_engine_cs *engine, const char *msg) 1077 { 1078 struct i915_gpu_error *error = &engine->i915->gpu_error; 1079 int ret; 1080 1081 GEM_TRACE("%s flags=%lx\n", engine->name, error->flags); 1082 GEM_BUG_ON(!test_bit(I915_RESET_ENGINE + engine->id, &error->flags)); 1083 1084 if (!intel_wakeref_active(&engine->wakeref)) 1085 return 0; 1086 1087 reset_prepare_engine(engine); 1088 1089 if (msg) 1090 dev_notice(engine->i915->drm.dev, 1091 "Resetting %s for %s\n", engine->name, msg); 1092 error->reset_engine_count[engine->id]++; 1093 1094 if (!engine->i915->guc.execbuf_client) 1095 ret = intel_gt_reset_engine(engine->i915, engine); 1096 else 1097 ret = intel_guc_reset_engine(&engine->i915->guc, engine); 1098 if (ret) { 1099 /* If we fail here, we expect to fallback to a global reset */ 1100 DRM_DEBUG_DRIVER("%sFailed to reset %s, ret=%d\n", 1101 engine->i915->guc.execbuf_client ? "GuC " : "", 1102 engine->name, ret); 1103 goto out; 1104 } 1105 1106 /* 1107 * The request that caused the hang is stuck on elsp, we know the 1108 * active request and can drop it, adjust head to skip the offending 1109 * request to resume executing remaining requests in the queue. 1110 */ 1111 intel_engine_reset(engine, true); 1112 1113 /* 1114 * The engine and its registers (and workarounds in case of render) 1115 * have been reset to their default values. Follow the init_ring 1116 * process to program RING_MODE, HWSP and re-enable submission. 1117 */ 1118 ret = engine->resume(engine); 1119 if (ret) 1120 goto out; 1121 1122 out: 1123 intel_engine_cancel_stop_cs(engine); 1124 reset_finish_engine(engine); 1125 return ret; 1126 } 1127 1128 static void i915_reset_device(struct drm_i915_private *i915, 1129 u32 engine_mask, 1130 const char *reason) 1131 { 1132 struct i915_gpu_error *error = &i915->gpu_error; 1133 struct kobject *kobj = &i915->drm.primary->kdev->kobj; 1134 char *error_event[] = { I915_ERROR_UEVENT "=1", NULL }; 1135 char *reset_event[] = { I915_RESET_UEVENT "=1", NULL }; 1136 char *reset_done_event[] = { I915_ERROR_UEVENT "=0", NULL }; 1137 struct i915_wedge_me w; 1138 1139 kobject_uevent_env(kobj, KOBJ_CHANGE, error_event); 1140 1141 DRM_DEBUG_DRIVER("resetting chip\n"); 1142 kobject_uevent_env(kobj, KOBJ_CHANGE, reset_event); 1143 1144 /* Use a watchdog to ensure that our reset completes */ 1145 i915_wedge_on_timeout(&w, i915, 5 * HZ) { 1146 intel_prepare_reset(i915); 1147 1148 /* Flush everyone using a resource about to be clobbered */ 1149 synchronize_srcu_expedited(&error->reset_backoff_srcu); 1150 1151 i915_reset(i915, engine_mask, reason); 1152 1153 intel_finish_reset(i915); 1154 } 1155 1156 if (!test_bit(I915_WEDGED, &error->flags)) 1157 kobject_uevent_env(kobj, KOBJ_CHANGE, reset_done_event); 1158 } 1159 1160 static void clear_register(struct intel_uncore *uncore, i915_reg_t reg) 1161 { 1162 intel_uncore_rmw(uncore, reg, 0, 0); 1163 } 1164 1165 static void gen8_clear_engine_error_register(struct intel_engine_cs *engine) 1166 { 1167 GEN6_RING_FAULT_REG_RMW(engine, RING_FAULT_VALID, 0); 1168 GEN6_RING_FAULT_REG_POSTING_READ(engine); 1169 } 1170 1171 static void clear_error_registers(struct drm_i915_private *i915, 1172 intel_engine_mask_t engine_mask) 1173 { 1174 struct intel_uncore *uncore = &i915->uncore; 1175 u32 eir; 1176 1177 if (!IS_GEN(i915, 2)) 1178 clear_register(uncore, PGTBL_ER); 1179 1180 if (INTEL_GEN(i915) < 4) 1181 clear_register(uncore, IPEIR(RENDER_RING_BASE)); 1182 else 1183 clear_register(uncore, IPEIR_I965); 1184 1185 clear_register(uncore, EIR); 1186 eir = intel_uncore_read(uncore, EIR); 1187 if (eir) { 1188 /* 1189 * some errors might have become stuck, 1190 * mask them. 1191 */ 1192 DRM_DEBUG_DRIVER("EIR stuck: 0x%08x, masking\n", eir); 1193 rmw_set(uncore, EMR, eir); 1194 intel_uncore_write(uncore, GEN2_IIR, 1195 I915_MASTER_ERROR_INTERRUPT); 1196 } 1197 1198 if (INTEL_GEN(i915) >= 8) { 1199 rmw_clear(uncore, GEN8_RING_FAULT_REG, RING_FAULT_VALID); 1200 intel_uncore_posting_read(uncore, GEN8_RING_FAULT_REG); 1201 } else if (INTEL_GEN(i915) >= 6) { 1202 struct intel_engine_cs *engine; 1203 enum intel_engine_id id; 1204 1205 for_each_engine_masked(engine, i915, engine_mask, id) 1206 gen8_clear_engine_error_register(engine); 1207 } 1208 } 1209 1210 static void gen6_check_faults(struct drm_i915_private *dev_priv) 1211 { 1212 struct intel_engine_cs *engine; 1213 enum intel_engine_id id; 1214 u32 fault; 1215 1216 for_each_engine(engine, dev_priv, id) { 1217 fault = GEN6_RING_FAULT_REG_READ(engine); 1218 if (fault & RING_FAULT_VALID) { 1219 DRM_DEBUG_DRIVER("Unexpected fault\n" 1220 "\tAddr: 0x%08lx\n" 1221 "\tAddress space: %s\n" 1222 "\tSource ID: %d\n" 1223 "\tType: %d\n", 1224 fault & PAGE_MASK, 1225 fault & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT", 1226 RING_FAULT_SRCID(fault), 1227 RING_FAULT_FAULT_TYPE(fault)); 1228 } 1229 } 1230 } 1231 1232 static void gen8_check_faults(struct drm_i915_private *dev_priv) 1233 { 1234 u32 fault = I915_READ(GEN8_RING_FAULT_REG); 1235 1236 if (fault & RING_FAULT_VALID) { 1237 u32 fault_data0, fault_data1; 1238 u64 fault_addr; 1239 1240 fault_data0 = I915_READ(GEN8_FAULT_TLB_DATA0); 1241 fault_data1 = I915_READ(GEN8_FAULT_TLB_DATA1); 1242 fault_addr = ((u64)(fault_data1 & FAULT_VA_HIGH_BITS) << 44) | 1243 ((u64)fault_data0 << 12); 1244 1245 DRM_DEBUG_DRIVER("Unexpected fault\n" 1246 "\tAddr: 0x%08x_%08x\n" 1247 "\tAddress space: %s\n" 1248 "\tEngine ID: %d\n" 1249 "\tSource ID: %d\n" 1250 "\tType: %d\n", 1251 upper_32_bits(fault_addr), 1252 lower_32_bits(fault_addr), 1253 fault_data1 & FAULT_GTT_SEL ? "GGTT" : "PPGTT", 1254 GEN8_RING_FAULT_ENGINE_ID(fault), 1255 RING_FAULT_SRCID(fault), 1256 RING_FAULT_FAULT_TYPE(fault)); 1257 } 1258 } 1259 1260 void i915_check_and_clear_faults(struct drm_i915_private *i915) 1261 { 1262 /* From GEN8 onwards we only have one 'All Engine Fault Register' */ 1263 if (INTEL_GEN(i915) >= 8) 1264 gen8_check_faults(i915); 1265 else if (INTEL_GEN(i915) >= 6) 1266 gen6_check_faults(i915); 1267 else 1268 return; 1269 1270 clear_error_registers(i915, ALL_ENGINES); 1271 } 1272 1273 /** 1274 * i915_handle_error - handle a gpu error 1275 * @i915: i915 device private 1276 * @engine_mask: mask representing engines that are hung 1277 * @flags: control flags 1278 * @fmt: Error message format string 1279 * 1280 * Do some basic checking of register state at error time and 1281 * dump it to the syslog. Also call i915_capture_error_state() to make 1282 * sure we get a record and make it available in debugfs. Fire a uevent 1283 * so userspace knows something bad happened (should trigger collection 1284 * of a ring dump etc.). 1285 */ 1286 void i915_handle_error(struct drm_i915_private *i915, 1287 intel_engine_mask_t engine_mask, 1288 unsigned long flags, 1289 const char *fmt, ...) 1290 { 1291 struct i915_gpu_error *error = &i915->gpu_error; 1292 struct intel_engine_cs *engine; 1293 intel_wakeref_t wakeref; 1294 intel_engine_mask_t tmp; 1295 char error_msg[80]; 1296 char *msg = NULL; 1297 1298 if (fmt) { 1299 va_list args; 1300 1301 va_start(args, fmt); 1302 vscnprintf(error_msg, sizeof(error_msg), fmt, args); 1303 va_end(args); 1304 1305 msg = error_msg; 1306 } 1307 1308 /* 1309 * In most cases it's guaranteed that we get here with an RPM 1310 * reference held, for example because there is a pending GPU 1311 * request that won't finish until the reset is done. This 1312 * isn't the case at least when we get here by doing a 1313 * simulated reset via debugfs, so get an RPM reference. 1314 */ 1315 wakeref = intel_runtime_pm_get(&i915->runtime_pm); 1316 1317 engine_mask &= INTEL_INFO(i915)->engine_mask; 1318 1319 if (flags & I915_ERROR_CAPTURE) { 1320 i915_capture_error_state(i915, engine_mask, msg); 1321 clear_error_registers(i915, engine_mask); 1322 } 1323 1324 /* 1325 * Try engine reset when available. We fall back to full reset if 1326 * single reset fails. 1327 */ 1328 if (intel_has_reset_engine(i915) && !__i915_wedged(error)) { 1329 for_each_engine_masked(engine, i915, engine_mask, tmp) { 1330 BUILD_BUG_ON(I915_RESET_MODESET >= I915_RESET_ENGINE); 1331 if (test_and_set_bit(I915_RESET_ENGINE + engine->id, 1332 &error->flags)) 1333 continue; 1334 1335 if (i915_reset_engine(engine, msg) == 0) 1336 engine_mask &= ~engine->mask; 1337 1338 clear_bit(I915_RESET_ENGINE + engine->id, 1339 &error->flags); 1340 wake_up_bit(&error->flags, 1341 I915_RESET_ENGINE + engine->id); 1342 } 1343 } 1344 1345 if (!engine_mask) 1346 goto out; 1347 1348 /* Full reset needs the mutex, stop any other user trying to do so. */ 1349 if (test_and_set_bit(I915_RESET_BACKOFF, &error->flags)) { 1350 wait_event(error->reset_queue, 1351 !test_bit(I915_RESET_BACKOFF, &error->flags)); 1352 goto out; /* piggy-back on the other reset */ 1353 } 1354 1355 /* Make sure i915_reset_trylock() sees the I915_RESET_BACKOFF */ 1356 synchronize_rcu_expedited(); 1357 1358 /* Prevent any other reset-engine attempt. */ 1359 for_each_engine(engine, i915, tmp) { 1360 while (test_and_set_bit(I915_RESET_ENGINE + engine->id, 1361 &error->flags)) 1362 wait_on_bit(&error->flags, 1363 I915_RESET_ENGINE + engine->id, 1364 TASK_UNINTERRUPTIBLE); 1365 } 1366 1367 i915_reset_device(i915, engine_mask, msg); 1368 1369 for_each_engine(engine, i915, tmp) { 1370 clear_bit(I915_RESET_ENGINE + engine->id, 1371 &error->flags); 1372 } 1373 1374 clear_bit(I915_RESET_BACKOFF, &error->flags); 1375 wake_up_all(&error->reset_queue); 1376 1377 out: 1378 intel_runtime_pm_put(&i915->runtime_pm, wakeref); 1379 } 1380 1381 int i915_reset_trylock(struct drm_i915_private *i915) 1382 { 1383 struct i915_gpu_error *error = &i915->gpu_error; 1384 int srcu; 1385 1386 might_lock(&error->reset_backoff_srcu); 1387 might_sleep(); 1388 1389 rcu_read_lock(); 1390 while (test_bit(I915_RESET_BACKOFF, &error->flags)) { 1391 rcu_read_unlock(); 1392 1393 if (wait_event_interruptible(error->reset_queue, 1394 !test_bit(I915_RESET_BACKOFF, 1395 &error->flags))) 1396 return -EINTR; 1397 1398 rcu_read_lock(); 1399 } 1400 srcu = srcu_read_lock(&error->reset_backoff_srcu); 1401 rcu_read_unlock(); 1402 1403 return srcu; 1404 } 1405 1406 void i915_reset_unlock(struct drm_i915_private *i915, int tag) 1407 __releases(&i915->gpu_error.reset_backoff_srcu) 1408 { 1409 struct i915_gpu_error *error = &i915->gpu_error; 1410 1411 srcu_read_unlock(&error->reset_backoff_srcu, tag); 1412 } 1413 1414 int i915_terminally_wedged(struct drm_i915_private *i915) 1415 { 1416 struct i915_gpu_error *error = &i915->gpu_error; 1417 1418 might_sleep(); 1419 1420 if (!__i915_wedged(error)) 1421 return 0; 1422 1423 /* Reset still in progress? Maybe we will recover? */ 1424 if (!test_bit(I915_RESET_BACKOFF, &error->flags)) 1425 return -EIO; 1426 1427 /* XXX intel_reset_finish() still takes struct_mutex!!! */ 1428 if (mutex_is_locked(&i915->drm.struct_mutex)) 1429 return -EAGAIN; 1430 1431 if (wait_event_interruptible(error->reset_queue, 1432 !test_bit(I915_RESET_BACKOFF, 1433 &error->flags))) 1434 return -EINTR; 1435 1436 return __i915_wedged(error) ? -EIO : 0; 1437 } 1438 1439 static void i915_wedge_me(struct work_struct *work) 1440 { 1441 struct i915_wedge_me *w = container_of(work, typeof(*w), work.work); 1442 1443 dev_err(w->i915->drm.dev, 1444 "%s timed out, cancelling all in-flight rendering.\n", 1445 w->name); 1446 i915_gem_set_wedged(w->i915); 1447 } 1448 1449 void __i915_init_wedge(struct i915_wedge_me *w, 1450 struct drm_i915_private *i915, 1451 long timeout, 1452 const char *name) 1453 { 1454 w->i915 = i915; 1455 w->name = name; 1456 1457 INIT_DELAYED_WORK_ONSTACK(&w->work, i915_wedge_me); 1458 schedule_delayed_work(&w->work, timeout); 1459 } 1460 1461 void __i915_fini_wedge(struct i915_wedge_me *w) 1462 { 1463 cancel_delayed_work_sync(&w->work); 1464 destroy_delayed_work_on_stack(&w->work); 1465 w->i915 = NULL; 1466 } 1467 1468 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 1469 #include "selftest_reset.c" 1470 #endif 1471