1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2016 Intel Corporation 4 */ 5 6 #include <linux/string_helpers.h> 7 8 #include <drm/drm_print.h> 9 10 #include "gem/i915_gem_context.h" 11 #include "gem/i915_gem_internal.h" 12 #include "gt/intel_gt_print.h" 13 #include "gt/intel_gt_regs.h" 14 15 #include "i915_cmd_parser.h" 16 #include "i915_drv.h" 17 #include "i915_irq.h" 18 #include "i915_reg.h" 19 #include "intel_breadcrumbs.h" 20 #include "intel_context.h" 21 #include "intel_engine.h" 22 #include "intel_engine_pm.h" 23 #include "intel_engine_regs.h" 24 #include "intel_engine_user.h" 25 #include "intel_execlists_submission.h" 26 #include "intel_gt.h" 27 #include "intel_gt_mcr.h" 28 #include "intel_gt_pm.h" 29 #include "intel_gt_requests.h" 30 #include "intel_lrc.h" 31 #include "intel_lrc_reg.h" 32 #include "intel_reset.h" 33 #include "intel_ring.h" 34 #include "uc/intel_guc_submission.h" 35 36 /* Haswell does have the CXT_SIZE register however it does not appear to be 37 * valid. Now, docs explain in dwords what is in the context object. The full 38 * size is 70720 bytes, however, the power context and execlist context will 39 * never be saved (power context is stored elsewhere, and execlists don't work 40 * on HSW) - so the final size, including the extra state required for the 41 * Resource Streamer, is 66944 bytes, which rounds to 17 pages. 42 */ 43 #define HSW_CXT_TOTAL_SIZE (17 * PAGE_SIZE) 44 45 #define DEFAULT_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE) 46 #define GEN8_LR_CONTEXT_RENDER_SIZE (20 * PAGE_SIZE) 47 #define GEN9_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE) 48 #define GEN11_LR_CONTEXT_RENDER_SIZE (14 * PAGE_SIZE) 49 50 #define GEN8_LR_CONTEXT_OTHER_SIZE (2 * PAGE_SIZE) 51 52 #define MAX_MMIO_BASES 3 53 struct engine_info { 54 u8 class; 55 u8 instance; 56 /* mmio bases table *must* be sorted in reverse graphics_ver order */ 57 struct engine_mmio_base { 58 u32 graphics_ver : 8; 59 u32 base : 24; 60 } mmio_bases[MAX_MMIO_BASES]; 61 }; 62 63 static const struct engine_info intel_engines[] = { 64 [RCS0] = { 65 .class = RENDER_CLASS, 66 .instance = 0, 67 .mmio_bases = { 68 { .graphics_ver = 1, .base = RENDER_RING_BASE } 69 }, 70 }, 71 [BCS0] = { 72 .class = COPY_ENGINE_CLASS, 73 .instance = 0, 74 .mmio_bases = { 75 { .graphics_ver = 6, .base = BLT_RING_BASE } 76 }, 77 }, 78 [BCS1] = { 79 .class = COPY_ENGINE_CLASS, 80 .instance = 1, 81 .mmio_bases = { 82 { .graphics_ver = 12, .base = XEHPC_BCS1_RING_BASE } 83 }, 84 }, 85 [BCS2] = { 86 .class = COPY_ENGINE_CLASS, 87 .instance = 2, 88 .mmio_bases = { 89 { .graphics_ver = 12, .base = XEHPC_BCS2_RING_BASE } 90 }, 91 }, 92 [BCS3] = { 93 .class = COPY_ENGINE_CLASS, 94 .instance = 3, 95 .mmio_bases = { 96 { .graphics_ver = 12, .base = XEHPC_BCS3_RING_BASE } 97 }, 98 }, 99 [BCS4] = { 100 .class = COPY_ENGINE_CLASS, 101 .instance = 4, 102 .mmio_bases = { 103 { .graphics_ver = 12, .base = XEHPC_BCS4_RING_BASE } 104 }, 105 }, 106 [BCS5] = { 107 .class = COPY_ENGINE_CLASS, 108 .instance = 5, 109 .mmio_bases = { 110 { .graphics_ver = 12, .base = XEHPC_BCS5_RING_BASE } 111 }, 112 }, 113 [BCS6] = { 114 .class = COPY_ENGINE_CLASS, 115 .instance = 6, 116 .mmio_bases = { 117 { .graphics_ver = 12, .base = XEHPC_BCS6_RING_BASE } 118 }, 119 }, 120 [BCS7] = { 121 .class = COPY_ENGINE_CLASS, 122 .instance = 7, 123 .mmio_bases = { 124 { .graphics_ver = 12, .base = XEHPC_BCS7_RING_BASE } 125 }, 126 }, 127 [BCS8] = { 128 .class = COPY_ENGINE_CLASS, 129 .instance = 8, 130 .mmio_bases = { 131 { .graphics_ver = 12, .base = XEHPC_BCS8_RING_BASE } 132 }, 133 }, 134 [VCS0] = { 135 .class = VIDEO_DECODE_CLASS, 136 .instance = 0, 137 .mmio_bases = { 138 { .graphics_ver = 11, .base = GEN11_BSD_RING_BASE }, 139 { .graphics_ver = 6, .base = GEN6_BSD_RING_BASE }, 140 { .graphics_ver = 4, .base = BSD_RING_BASE } 141 }, 142 }, 143 [VCS1] = { 144 .class = VIDEO_DECODE_CLASS, 145 .instance = 1, 146 .mmio_bases = { 147 { .graphics_ver = 11, .base = GEN11_BSD2_RING_BASE }, 148 { .graphics_ver = 8, .base = GEN8_BSD2_RING_BASE } 149 }, 150 }, 151 [VCS2] = { 152 .class = VIDEO_DECODE_CLASS, 153 .instance = 2, 154 .mmio_bases = { 155 { .graphics_ver = 11, .base = GEN11_BSD3_RING_BASE } 156 }, 157 }, 158 [VCS3] = { 159 .class = VIDEO_DECODE_CLASS, 160 .instance = 3, 161 .mmio_bases = { 162 { .graphics_ver = 11, .base = GEN11_BSD4_RING_BASE } 163 }, 164 }, 165 [VCS4] = { 166 .class = VIDEO_DECODE_CLASS, 167 .instance = 4, 168 .mmio_bases = { 169 { .graphics_ver = 12, .base = XEHP_BSD5_RING_BASE } 170 }, 171 }, 172 [VCS5] = { 173 .class = VIDEO_DECODE_CLASS, 174 .instance = 5, 175 .mmio_bases = { 176 { .graphics_ver = 12, .base = XEHP_BSD6_RING_BASE } 177 }, 178 }, 179 [VCS6] = { 180 .class = VIDEO_DECODE_CLASS, 181 .instance = 6, 182 .mmio_bases = { 183 { .graphics_ver = 12, .base = XEHP_BSD7_RING_BASE } 184 }, 185 }, 186 [VCS7] = { 187 .class = VIDEO_DECODE_CLASS, 188 .instance = 7, 189 .mmio_bases = { 190 { .graphics_ver = 12, .base = XEHP_BSD8_RING_BASE } 191 }, 192 }, 193 [VECS0] = { 194 .class = VIDEO_ENHANCEMENT_CLASS, 195 .instance = 0, 196 .mmio_bases = { 197 { .graphics_ver = 11, .base = GEN11_VEBOX_RING_BASE }, 198 { .graphics_ver = 7, .base = VEBOX_RING_BASE } 199 }, 200 }, 201 [VECS1] = { 202 .class = VIDEO_ENHANCEMENT_CLASS, 203 .instance = 1, 204 .mmio_bases = { 205 { .graphics_ver = 11, .base = GEN11_VEBOX2_RING_BASE } 206 }, 207 }, 208 [VECS2] = { 209 .class = VIDEO_ENHANCEMENT_CLASS, 210 .instance = 2, 211 .mmio_bases = { 212 { .graphics_ver = 12, .base = XEHP_VEBOX3_RING_BASE } 213 }, 214 }, 215 [VECS3] = { 216 .class = VIDEO_ENHANCEMENT_CLASS, 217 .instance = 3, 218 .mmio_bases = { 219 { .graphics_ver = 12, .base = XEHP_VEBOX4_RING_BASE } 220 }, 221 }, 222 [CCS0] = { 223 .class = COMPUTE_CLASS, 224 .instance = 0, 225 .mmio_bases = { 226 { .graphics_ver = 12, .base = GEN12_COMPUTE0_RING_BASE } 227 } 228 }, 229 [CCS1] = { 230 .class = COMPUTE_CLASS, 231 .instance = 1, 232 .mmio_bases = { 233 { .graphics_ver = 12, .base = GEN12_COMPUTE1_RING_BASE } 234 } 235 }, 236 [CCS2] = { 237 .class = COMPUTE_CLASS, 238 .instance = 2, 239 .mmio_bases = { 240 { .graphics_ver = 12, .base = GEN12_COMPUTE2_RING_BASE } 241 } 242 }, 243 [CCS3] = { 244 .class = COMPUTE_CLASS, 245 .instance = 3, 246 .mmio_bases = { 247 { .graphics_ver = 12, .base = GEN12_COMPUTE3_RING_BASE } 248 } 249 }, 250 [GSC0] = { 251 .class = OTHER_CLASS, 252 .instance = OTHER_GSC_INSTANCE, 253 .mmio_bases = { 254 { .graphics_ver = 12, .base = MTL_GSC_RING_BASE } 255 } 256 }, 257 }; 258 259 /** 260 * intel_engine_context_size() - return the size of the context for an engine 261 * @gt: the gt 262 * @class: engine class 263 * 264 * Each engine class may require a different amount of space for a context 265 * image. 266 * 267 * Return: size (in bytes) of an engine class specific context image 268 * 269 * Note: this size includes the HWSP, which is part of the context image 270 * in LRC mode, but does not include the "shared data page" used with 271 * GuC submission. The caller should account for this if using the GuC. 272 */ 273 u32 intel_engine_context_size(struct intel_gt *gt, u8 class) 274 { 275 struct intel_uncore *uncore = gt->uncore; 276 u32 cxt_size; 277 278 BUILD_BUG_ON(I915_GTT_PAGE_SIZE != PAGE_SIZE); 279 280 switch (class) { 281 case COMPUTE_CLASS: 282 fallthrough; 283 case RENDER_CLASS: 284 switch (GRAPHICS_VER(gt->i915)) { 285 default: 286 MISSING_CASE(GRAPHICS_VER(gt->i915)); 287 return DEFAULT_LR_CONTEXT_RENDER_SIZE; 288 case 12: 289 case 11: 290 return GEN11_LR_CONTEXT_RENDER_SIZE; 291 case 9: 292 return GEN9_LR_CONTEXT_RENDER_SIZE; 293 case 8: 294 return GEN8_LR_CONTEXT_RENDER_SIZE; 295 case 7: 296 if (IS_HASWELL(gt->i915)) 297 return HSW_CXT_TOTAL_SIZE; 298 299 cxt_size = intel_uncore_read(uncore, GEN7_CXT_SIZE); 300 return round_up(GEN7_CXT_TOTAL_SIZE(cxt_size) * 64, 301 PAGE_SIZE); 302 case 6: 303 cxt_size = intel_uncore_read(uncore, CXT_SIZE); 304 return round_up(GEN6_CXT_TOTAL_SIZE(cxt_size) * 64, 305 PAGE_SIZE); 306 case 5: 307 case 4: 308 /* 309 * There is a discrepancy here between the size reported 310 * by the register and the size of the context layout 311 * in the docs. Both are described as authorative! 312 * 313 * The discrepancy is on the order of a few cachelines, 314 * but the total is under one page (4k), which is our 315 * minimum allocation anyway so it should all come 316 * out in the wash. 317 */ 318 cxt_size = intel_uncore_read(uncore, CXT_SIZE) + 1; 319 gt_dbg(gt, "graphics_ver = %d CXT_SIZE = %d bytes [0x%08x]\n", 320 GRAPHICS_VER(gt->i915), cxt_size * 64, 321 cxt_size - 1); 322 return round_up(cxt_size * 64, PAGE_SIZE); 323 case 3: 324 case 2: 325 /* For the special day when i810 gets merged. */ 326 case 1: 327 return 0; 328 } 329 break; 330 default: 331 MISSING_CASE(class); 332 fallthrough; 333 case VIDEO_DECODE_CLASS: 334 case VIDEO_ENHANCEMENT_CLASS: 335 case COPY_ENGINE_CLASS: 336 case OTHER_CLASS: 337 if (GRAPHICS_VER(gt->i915) < 8) 338 return 0; 339 return GEN8_LR_CONTEXT_OTHER_SIZE; 340 } 341 } 342 343 static u32 __engine_mmio_base(struct drm_i915_private *i915, 344 const struct engine_mmio_base *bases) 345 { 346 int i; 347 348 for (i = 0; i < MAX_MMIO_BASES; i++) 349 if (GRAPHICS_VER(i915) >= bases[i].graphics_ver) 350 break; 351 352 GEM_BUG_ON(i == MAX_MMIO_BASES); 353 GEM_BUG_ON(!bases[i].base); 354 355 return bases[i].base; 356 } 357 358 static void __sprint_engine_name(struct intel_engine_cs *engine) 359 { 360 /* 361 * Before we know what the uABI name for this engine will be, 362 * we still would like to keep track of this engine in the debug logs. 363 * We throw in a ' here as a reminder that this isn't its final name. 364 */ 365 GEM_WARN_ON(snprintf(engine->name, sizeof(engine->name), "%s'%u", 366 intel_engine_class_repr(engine->class), 367 engine->instance) >= sizeof(engine->name)); 368 } 369 370 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask) 371 { 372 /* 373 * Though they added more rings on g4x/ilk, they did not add 374 * per-engine HWSTAM until gen6. 375 */ 376 if (GRAPHICS_VER(engine->i915) < 6 && engine->class != RENDER_CLASS) 377 return; 378 379 if (GRAPHICS_VER(engine->i915) >= 3) 380 ENGINE_WRITE(engine, RING_HWSTAM, mask); 381 else 382 ENGINE_WRITE16(engine, RING_HWSTAM, mask); 383 } 384 385 static void intel_engine_sanitize_mmio(struct intel_engine_cs *engine) 386 { 387 /* Mask off all writes into the unknown HWSP */ 388 intel_engine_set_hwsp_writemask(engine, ~0u); 389 } 390 391 static void nop_irq_handler(struct intel_engine_cs *engine, u16 iir) 392 { 393 GEM_DEBUG_WARN_ON(iir); 394 } 395 396 static u32 get_reset_domain(u8 ver, enum intel_engine_id id) 397 { 398 u32 reset_domain; 399 400 if (ver >= 11) { 401 static const u32 engine_reset_domains[] = { 402 [RCS0] = GEN11_GRDOM_RENDER, 403 [BCS0] = GEN11_GRDOM_BLT, 404 [BCS1] = XEHPC_GRDOM_BLT1, 405 [BCS2] = XEHPC_GRDOM_BLT2, 406 [BCS3] = XEHPC_GRDOM_BLT3, 407 [BCS4] = XEHPC_GRDOM_BLT4, 408 [BCS5] = XEHPC_GRDOM_BLT5, 409 [BCS6] = XEHPC_GRDOM_BLT6, 410 [BCS7] = XEHPC_GRDOM_BLT7, 411 [BCS8] = XEHPC_GRDOM_BLT8, 412 [VCS0] = GEN11_GRDOM_MEDIA, 413 [VCS1] = GEN11_GRDOM_MEDIA2, 414 [VCS2] = GEN11_GRDOM_MEDIA3, 415 [VCS3] = GEN11_GRDOM_MEDIA4, 416 [VCS4] = GEN11_GRDOM_MEDIA5, 417 [VCS5] = GEN11_GRDOM_MEDIA6, 418 [VCS6] = GEN11_GRDOM_MEDIA7, 419 [VCS7] = GEN11_GRDOM_MEDIA8, 420 [VECS0] = GEN11_GRDOM_VECS, 421 [VECS1] = GEN11_GRDOM_VECS2, 422 [VECS2] = GEN11_GRDOM_VECS3, 423 [VECS3] = GEN11_GRDOM_VECS4, 424 [CCS0] = GEN11_GRDOM_RENDER, 425 [CCS1] = GEN11_GRDOM_RENDER, 426 [CCS2] = GEN11_GRDOM_RENDER, 427 [CCS3] = GEN11_GRDOM_RENDER, 428 [GSC0] = GEN12_GRDOM_GSC, 429 }; 430 GEM_BUG_ON(id >= ARRAY_SIZE(engine_reset_domains) || 431 !engine_reset_domains[id]); 432 reset_domain = engine_reset_domains[id]; 433 } else { 434 static const u32 engine_reset_domains[] = { 435 [RCS0] = GEN6_GRDOM_RENDER, 436 [BCS0] = GEN6_GRDOM_BLT, 437 [VCS0] = GEN6_GRDOM_MEDIA, 438 [VCS1] = GEN8_GRDOM_MEDIA2, 439 [VECS0] = GEN6_GRDOM_VECS, 440 }; 441 GEM_BUG_ON(id >= ARRAY_SIZE(engine_reset_domains) || 442 !engine_reset_domains[id]); 443 reset_domain = engine_reset_domains[id]; 444 } 445 446 return reset_domain; 447 } 448 449 static int intel_engine_setup(struct intel_gt *gt, enum intel_engine_id id, 450 u8 logical_instance) 451 { 452 const struct engine_info *info = &intel_engines[id]; 453 struct drm_i915_private *i915 = gt->i915; 454 struct intel_engine_cs *engine; 455 u8 guc_class; 456 457 BUILD_BUG_ON(MAX_ENGINE_CLASS >= BIT(GEN11_ENGINE_CLASS_WIDTH)); 458 BUILD_BUG_ON(MAX_ENGINE_INSTANCE >= BIT(GEN11_ENGINE_INSTANCE_WIDTH)); 459 BUILD_BUG_ON(I915_MAX_VCS > (MAX_ENGINE_INSTANCE + 1)); 460 BUILD_BUG_ON(I915_MAX_VECS > (MAX_ENGINE_INSTANCE + 1)); 461 462 if (GEM_DEBUG_WARN_ON(id >= ARRAY_SIZE(gt->engine))) 463 return -EINVAL; 464 465 if (GEM_DEBUG_WARN_ON(info->class > MAX_ENGINE_CLASS)) 466 return -EINVAL; 467 468 if (GEM_DEBUG_WARN_ON(info->instance > MAX_ENGINE_INSTANCE)) 469 return -EINVAL; 470 471 if (GEM_DEBUG_WARN_ON(gt->engine_class[info->class][info->instance])) 472 return -EINVAL; 473 474 engine = kzalloc(sizeof(*engine), GFP_KERNEL); 475 if (!engine) 476 return -ENOMEM; 477 478 BUILD_BUG_ON(BITS_PER_TYPE(engine->mask) < I915_NUM_ENGINES); 479 480 INIT_LIST_HEAD(&engine->pinned_contexts_list); 481 engine->id = id; 482 engine->legacy_idx = INVALID_ENGINE; 483 engine->mask = BIT(id); 484 engine->reset_domain = get_reset_domain(GRAPHICS_VER(gt->i915), 485 id); 486 engine->i915 = i915; 487 engine->gt = gt; 488 engine->uncore = gt->uncore; 489 guc_class = engine_class_to_guc_class(info->class); 490 engine->guc_id = MAKE_GUC_ID(guc_class, info->instance); 491 engine->mmio_base = __engine_mmio_base(i915, info->mmio_bases); 492 493 engine->irq_handler = nop_irq_handler; 494 495 engine->class = info->class; 496 engine->instance = info->instance; 497 engine->logical_mask = BIT(logical_instance); 498 __sprint_engine_name(engine); 499 500 if ((engine->class == COMPUTE_CLASS || engine->class == RENDER_CLASS) && 501 __ffs(CCS_MASK(engine->gt) | RCS_MASK(engine->gt)) == engine->instance) 502 engine->flags |= I915_ENGINE_FIRST_RENDER_COMPUTE; 503 504 /* features common between engines sharing EUs */ 505 if (engine->class == RENDER_CLASS || engine->class == COMPUTE_CLASS) { 506 engine->flags |= I915_ENGINE_HAS_RCS_REG_STATE; 507 engine->flags |= I915_ENGINE_HAS_EU_PRIORITY; 508 } 509 510 engine->props.heartbeat_interval_ms = 511 CONFIG_DRM_I915_HEARTBEAT_INTERVAL; 512 engine->props.max_busywait_duration_ns = 513 CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT; 514 engine->props.preempt_timeout_ms = 515 CONFIG_DRM_I915_PREEMPT_TIMEOUT; 516 engine->props.stop_timeout_ms = 517 CONFIG_DRM_I915_STOP_TIMEOUT; 518 engine->props.timeslice_duration_ms = 519 CONFIG_DRM_I915_TIMESLICE_DURATION; 520 521 /* 522 * Mid-thread pre-emption is not available in Gen12. Unfortunately, 523 * some compute workloads run quite long threads. That means they get 524 * reset due to not pre-empting in a timely manner. So, bump the 525 * pre-emption timeout value to be much higher for compute engines. 526 */ 527 if (GRAPHICS_VER(i915) == 12 && (engine->flags & I915_ENGINE_HAS_RCS_REG_STATE)) 528 engine->props.preempt_timeout_ms = CONFIG_DRM_I915_PREEMPT_TIMEOUT_COMPUTE; 529 530 /* Cap properties according to any system limits */ 531 #define CLAMP_PROP(field) \ 532 do { \ 533 u64 clamp = intel_clamp_##field(engine, engine->props.field); \ 534 if (clamp != engine->props.field) { \ 535 drm_notice(&engine->i915->drm, \ 536 "Warning, clamping %s to %lld to prevent overflow\n", \ 537 #field, clamp); \ 538 engine->props.field = clamp; \ 539 } \ 540 } while (0) 541 542 CLAMP_PROP(heartbeat_interval_ms); 543 CLAMP_PROP(max_busywait_duration_ns); 544 CLAMP_PROP(preempt_timeout_ms); 545 CLAMP_PROP(stop_timeout_ms); 546 CLAMP_PROP(timeslice_duration_ms); 547 548 #undef CLAMP_PROP 549 550 engine->defaults = engine->props; /* never to change again */ 551 552 engine->context_size = intel_engine_context_size(gt, engine->class); 553 if (WARN_ON(engine->context_size > BIT(20))) 554 engine->context_size = 0; 555 if (engine->context_size) 556 DRIVER_CAPS(i915)->has_logical_contexts = true; 557 558 ewma__engine_latency_init(&engine->latency); 559 560 ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier); 561 562 /* Scrub mmio state on takeover */ 563 intel_engine_sanitize_mmio(engine); 564 565 gt->engine_class[info->class][info->instance] = engine; 566 gt->engine[id] = engine; 567 568 return 0; 569 } 570 571 u64 intel_clamp_heartbeat_interval_ms(struct intel_engine_cs *engine, u64 value) 572 { 573 value = min_t(u64, value, jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT)); 574 575 return value; 576 } 577 578 u64 intel_clamp_max_busywait_duration_ns(struct intel_engine_cs *engine, u64 value) 579 { 580 value = min(value, jiffies_to_nsecs(2)); 581 582 return value; 583 } 584 585 u64 intel_clamp_preempt_timeout_ms(struct intel_engine_cs *engine, u64 value) 586 { 587 /* 588 * NB: The GuC API only supports 32bit values. However, the limit is further 589 * reduced due to internal calculations which would otherwise overflow. 590 */ 591 if (intel_guc_submission_is_wanted(gt_to_guc(engine->gt))) 592 value = min_t(u64, value, guc_policy_max_preempt_timeout_ms()); 593 594 value = min_t(u64, value, jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT)); 595 596 return value; 597 } 598 599 u64 intel_clamp_stop_timeout_ms(struct intel_engine_cs *engine, u64 value) 600 { 601 value = min_t(u64, value, jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT)); 602 603 return value; 604 } 605 606 u64 intel_clamp_timeslice_duration_ms(struct intel_engine_cs *engine, u64 value) 607 { 608 /* 609 * NB: The GuC API only supports 32bit values. However, the limit is further 610 * reduced due to internal calculations which would otherwise overflow. 611 */ 612 if (intel_guc_submission_is_wanted(gt_to_guc(engine->gt))) 613 value = min_t(u64, value, guc_policy_max_exec_quantum_ms()); 614 615 value = min_t(u64, value, jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT)); 616 617 return value; 618 } 619 620 static void __setup_engine_capabilities(struct intel_engine_cs *engine) 621 { 622 struct drm_i915_private *i915 = engine->i915; 623 624 if (engine->class == VIDEO_DECODE_CLASS) { 625 /* 626 * HEVC support is present on first engine instance 627 * before Gen11 and on all instances afterwards. 628 */ 629 if (GRAPHICS_VER(i915) >= 11 || 630 (GRAPHICS_VER(i915) >= 9 && engine->instance == 0)) 631 engine->uabi_capabilities |= 632 I915_VIDEO_CLASS_CAPABILITY_HEVC; 633 634 /* 635 * SFC block is present only on even logical engine 636 * instances. 637 */ 638 if ((GRAPHICS_VER(i915) >= 11 && 639 (engine->gt->info.vdbox_sfc_access & 640 BIT(engine->instance))) || 641 (GRAPHICS_VER(i915) >= 9 && engine->instance == 0)) 642 engine->uabi_capabilities |= 643 I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC; 644 } else if (engine->class == VIDEO_ENHANCEMENT_CLASS) { 645 if (GRAPHICS_VER(i915) >= 9 && 646 engine->gt->info.sfc_mask & BIT(engine->instance)) 647 engine->uabi_capabilities |= 648 I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC; 649 } 650 } 651 652 static void intel_setup_engine_capabilities(struct intel_gt *gt) 653 { 654 struct intel_engine_cs *engine; 655 enum intel_engine_id id; 656 657 for_each_engine(engine, gt, id) 658 __setup_engine_capabilities(engine); 659 } 660 661 /** 662 * intel_engines_release() - free the resources allocated for Command Streamers 663 * @gt: pointer to struct intel_gt 664 */ 665 void intel_engines_release(struct intel_gt *gt) 666 { 667 struct intel_engine_cs *engine; 668 enum intel_engine_id id; 669 670 /* 671 * Before we release the resources held by engine, we must be certain 672 * that the HW is no longer accessing them -- having the GPU scribble 673 * to or read from a page being used for something else causes no end 674 * of fun. 675 * 676 * The GPU should be reset by this point, but assume the worst just 677 * in case we aborted before completely initialising the engines. 678 */ 679 GEM_BUG_ON(intel_gt_pm_is_awake(gt)); 680 if (!INTEL_INFO(gt->i915)->gpu_reset_clobbers_display) 681 intel_gt_reset_all_engines(gt); 682 683 /* Decouple the backend; but keep the layout for late GPU resets */ 684 for_each_engine(engine, gt, id) { 685 if (!engine->release) 686 continue; 687 688 intel_wakeref_wait_for_idle(&engine->wakeref); 689 GEM_BUG_ON(intel_engine_pm_is_awake(engine)); 690 691 engine->release(engine); 692 engine->release = NULL; 693 694 memset(&engine->reset, 0, sizeof(engine->reset)); 695 } 696 697 llist_del_all(>->i915->uabi_engines_llist); 698 } 699 700 void intel_engine_free_request_pool(struct intel_engine_cs *engine) 701 { 702 if (!engine->request_pool) 703 return; 704 705 kmem_cache_free(i915_request_slab_cache(), engine->request_pool); 706 } 707 708 void intel_engines_free(struct intel_gt *gt) 709 { 710 struct intel_engine_cs *engine; 711 enum intel_engine_id id; 712 713 /* Free the requests! dma-resv keeps fences around for an eternity */ 714 rcu_barrier(); 715 716 for_each_engine(engine, gt, id) { 717 intel_engine_free_request_pool(engine); 718 kfree(engine); 719 gt->engine[id] = NULL; 720 } 721 } 722 723 static 724 bool gen11_vdbox_has_sfc(struct intel_gt *gt, 725 unsigned int physical_vdbox, 726 unsigned int logical_vdbox, u16 vdbox_mask) 727 { 728 struct drm_i915_private *i915 = gt->i915; 729 730 /* 731 * In Gen11, only even numbered logical VDBOXes are hooked 732 * up to an SFC (Scaler & Format Converter) unit. 733 * In Gen12, Even numbered physical instance always are connected 734 * to an SFC. Odd numbered physical instances have SFC only if 735 * previous even instance is fused off. 736 * 737 * Starting with Xe_HP, there's also a dedicated SFC_ENABLE field 738 * in the fuse register that tells us whether a specific SFC is present. 739 */ 740 if ((gt->info.sfc_mask & BIT(physical_vdbox / 2)) == 0) 741 return false; 742 else if (MEDIA_VER(i915) >= 12) 743 return (physical_vdbox % 2 == 0) || 744 !(BIT(physical_vdbox - 1) & vdbox_mask); 745 else if (MEDIA_VER(i915) == 11) 746 return logical_vdbox % 2 == 0; 747 748 return false; 749 } 750 751 static void engine_mask_apply_media_fuses(struct intel_gt *gt) 752 { 753 struct drm_i915_private *i915 = gt->i915; 754 unsigned int logical_vdbox = 0; 755 unsigned int i; 756 u32 media_fuse, fuse1; 757 u16 vdbox_mask; 758 u16 vebox_mask; 759 760 if (MEDIA_VER(gt->i915) < 11) 761 return; 762 763 /* 764 * On newer platforms the fusing register is called 'enable' and has 765 * enable semantics, while on older platforms it is called 'disable' 766 * and bits have disable semantices. 767 */ 768 media_fuse = intel_uncore_read(gt->uncore, GEN11_GT_VEBOX_VDBOX_DISABLE); 769 if (MEDIA_VER_FULL(i915) < IP_VER(12, 55)) 770 media_fuse = ~media_fuse; 771 772 vdbox_mask = media_fuse & GEN11_GT_VDBOX_DISABLE_MASK; 773 vebox_mask = (media_fuse & GEN11_GT_VEBOX_DISABLE_MASK) >> 774 GEN11_GT_VEBOX_DISABLE_SHIFT; 775 776 if (MEDIA_VER_FULL(i915) >= IP_VER(12, 55)) { 777 fuse1 = intel_uncore_read(gt->uncore, HSW_PAVP_FUSE1); 778 gt->info.sfc_mask = REG_FIELD_GET(XEHP_SFC_ENABLE_MASK, fuse1); 779 } else { 780 gt->info.sfc_mask = ~0; 781 } 782 783 for (i = 0; i < I915_MAX_VCS; i++) { 784 if (!HAS_ENGINE(gt, _VCS(i))) { 785 vdbox_mask &= ~BIT(i); 786 continue; 787 } 788 789 if (!(BIT(i) & vdbox_mask)) { 790 gt->info.engine_mask &= ~BIT(_VCS(i)); 791 gt_dbg(gt, "vcs%u fused off\n", i); 792 continue; 793 } 794 795 if (gen11_vdbox_has_sfc(gt, i, logical_vdbox, vdbox_mask)) 796 gt->info.vdbox_sfc_access |= BIT(i); 797 logical_vdbox++; 798 } 799 gt_dbg(gt, "vdbox enable: %04x, instances: %04lx\n", vdbox_mask, VDBOX_MASK(gt)); 800 GEM_BUG_ON(vdbox_mask != VDBOX_MASK(gt)); 801 802 for (i = 0; i < I915_MAX_VECS; i++) { 803 if (!HAS_ENGINE(gt, _VECS(i))) { 804 vebox_mask &= ~BIT(i); 805 continue; 806 } 807 808 if (!(BIT(i) & vebox_mask)) { 809 gt->info.engine_mask &= ~BIT(_VECS(i)); 810 gt_dbg(gt, "vecs%u fused off\n", i); 811 } 812 } 813 gt_dbg(gt, "vebox enable: %04x, instances: %04lx\n", vebox_mask, VEBOX_MASK(gt)); 814 GEM_BUG_ON(vebox_mask != VEBOX_MASK(gt)); 815 } 816 817 static void engine_mask_apply_compute_fuses(struct intel_gt *gt) 818 { 819 struct drm_i915_private *i915 = gt->i915; 820 struct intel_gt_info *info = >->info; 821 int ss_per_ccs = info->sseu.max_subslices / I915_MAX_CCS; 822 unsigned long ccs_mask; 823 unsigned int i; 824 825 if (GRAPHICS_VER(i915) < 11) 826 return; 827 828 if (hweight32(CCS_MASK(gt)) <= 1) 829 return; 830 831 ccs_mask = intel_slicemask_from_xehp_dssmask(info->sseu.compute_subslice_mask, 832 ss_per_ccs); 833 /* 834 * If all DSS in a quadrant are fused off, the corresponding CCS 835 * engine is not available for use. 836 */ 837 for_each_clear_bit(i, &ccs_mask, I915_MAX_CCS) { 838 info->engine_mask &= ~BIT(_CCS(i)); 839 gt_dbg(gt, "ccs%u fused off\n", i); 840 } 841 } 842 843 /* 844 * Determine which engines are fused off in our particular hardware. 845 * Note that we have a catch-22 situation where we need to be able to access 846 * the blitter forcewake domain to read the engine fuses, but at the same time 847 * we need to know which engines are available on the system to know which 848 * forcewake domains are present. We solve this by intializing the forcewake 849 * domains based on the full engine mask in the platform capabilities before 850 * calling this function and pruning the domains for fused-off engines 851 * afterwards. 852 */ 853 static intel_engine_mask_t init_engine_mask(struct intel_gt *gt) 854 { 855 struct intel_gt_info *info = >->info; 856 857 GEM_BUG_ON(!info->engine_mask); 858 859 engine_mask_apply_media_fuses(gt); 860 engine_mask_apply_compute_fuses(gt); 861 862 /* 863 * The only use of the GSC CS is to load and communicate with the GSC 864 * FW, so we have no use for it if we don't have the FW. 865 * 866 * IMPORTANT: in cases where we don't have the GSC FW, we have a 867 * catch-22 situation that breaks media C6 due to 2 requirements: 868 * 1) once turned on, the GSC power well will not go to sleep unless the 869 * GSC FW is loaded. 870 * 2) to enable idling (which is required for media C6) we need to 871 * initialize the IDLE_MSG register for the GSC CS and do at least 1 872 * submission, which will wake up the GSC power well. 873 */ 874 if (__HAS_ENGINE(info->engine_mask, GSC0) && !intel_uc_wants_gsc_uc(>->uc)) { 875 gt_notice(gt, "No GSC FW selected, disabling GSC CS and media C6\n"); 876 info->engine_mask &= ~BIT(GSC0); 877 } 878 879 /* 880 * Do not create the command streamer for CCS slices beyond the first. 881 * All the workload submitted to the first engine will be shared among 882 * all the slices. 883 * 884 * Once the user will be allowed to customize the CCS mode, then this 885 * check needs to be removed. 886 */ 887 if (IS_DG2(gt->i915)) { 888 u8 first_ccs = __ffs(CCS_MASK(gt)); 889 890 /* 891 * Store the number of active cslices before 892 * changing the CCS engine configuration 893 */ 894 gt->ccs.cslices = CCS_MASK(gt); 895 896 /* Mask off all the CCS engine */ 897 info->engine_mask &= ~GENMASK(CCS3, CCS0); 898 /* Put back in the first CCS engine */ 899 info->engine_mask |= BIT(_CCS(first_ccs)); 900 } 901 902 return info->engine_mask; 903 } 904 905 static void populate_logical_ids(struct intel_gt *gt, u8 *logical_ids, 906 u8 class, const u8 *map, u8 num_instances) 907 { 908 int i, j; 909 u8 current_logical_id = 0; 910 911 for (j = 0; j < num_instances; ++j) { 912 for (i = 0; i < ARRAY_SIZE(intel_engines); ++i) { 913 if (!HAS_ENGINE(gt, i) || 914 intel_engines[i].class != class) 915 continue; 916 917 if (intel_engines[i].instance == map[j]) { 918 logical_ids[intel_engines[i].instance] = 919 current_logical_id++; 920 break; 921 } 922 } 923 } 924 } 925 926 static void setup_logical_ids(struct intel_gt *gt, u8 *logical_ids, u8 class) 927 { 928 /* 929 * Logical to physical mapping is needed for proper support 930 * to split-frame feature. 931 */ 932 if (MEDIA_VER(gt->i915) >= 11 && class == VIDEO_DECODE_CLASS) { 933 const u8 map[] = { 0, 2, 4, 6, 1, 3, 5, 7 }; 934 935 populate_logical_ids(gt, logical_ids, class, 936 map, ARRAY_SIZE(map)); 937 } else { 938 int i; 939 u8 map[MAX_ENGINE_INSTANCE + 1]; 940 941 for (i = 0; i < MAX_ENGINE_INSTANCE + 1; ++i) 942 map[i] = i; 943 populate_logical_ids(gt, logical_ids, class, 944 map, ARRAY_SIZE(map)); 945 } 946 } 947 948 /** 949 * intel_engines_init_mmio() - allocate and prepare the Engine Command Streamers 950 * @gt: pointer to struct intel_gt 951 * 952 * Return: non-zero if the initialization failed. 953 */ 954 int intel_engines_init_mmio(struct intel_gt *gt) 955 { 956 struct drm_i915_private *i915 = gt->i915; 957 const unsigned int engine_mask = init_engine_mask(gt); 958 unsigned int mask = 0; 959 unsigned int i, class; 960 u8 logical_ids[MAX_ENGINE_INSTANCE + 1]; 961 int err; 962 963 drm_WARN_ON(&i915->drm, engine_mask == 0); 964 drm_WARN_ON(&i915->drm, engine_mask & 965 GENMASK(BITS_PER_TYPE(mask) - 1, I915_NUM_ENGINES)); 966 967 if (i915_inject_probe_failure(i915)) 968 return -ENODEV; 969 970 for (class = 0; class < MAX_ENGINE_CLASS + 1; ++class) { 971 setup_logical_ids(gt, logical_ids, class); 972 973 for (i = 0; i < ARRAY_SIZE(intel_engines); ++i) { 974 u8 instance = intel_engines[i].instance; 975 976 if (intel_engines[i].class != class || 977 !HAS_ENGINE(gt, i)) 978 continue; 979 980 err = intel_engine_setup(gt, i, 981 logical_ids[instance]); 982 if (err) 983 goto cleanup; 984 985 mask |= BIT(i); 986 } 987 } 988 989 /* 990 * Catch failures to update intel_engines table when the new engines 991 * are added to the driver by a warning and disabling the forgotten 992 * engines. 993 */ 994 if (drm_WARN_ON(&i915->drm, mask != engine_mask)) 995 gt->info.engine_mask = mask; 996 997 gt->info.num_engines = hweight32(mask); 998 999 intel_gt_check_and_clear_faults(gt); 1000 1001 intel_setup_engine_capabilities(gt); 1002 1003 intel_uncore_prune_engine_fw_domains(gt->uncore, gt); 1004 1005 return 0; 1006 1007 cleanup: 1008 intel_engines_free(gt); 1009 return err; 1010 } 1011 1012 void intel_engine_init_execlists(struct intel_engine_cs *engine) 1013 { 1014 struct intel_engine_execlists * const execlists = &engine->execlists; 1015 1016 execlists->port_mask = 1; 1017 GEM_BUG_ON(!is_power_of_2(execlists_num_ports(execlists))); 1018 GEM_BUG_ON(execlists_num_ports(execlists) > EXECLIST_MAX_PORTS); 1019 1020 memset(execlists->pending, 0, sizeof(execlists->pending)); 1021 execlists->active = 1022 memset(execlists->inflight, 0, sizeof(execlists->inflight)); 1023 } 1024 1025 static void cleanup_status_page(struct intel_engine_cs *engine) 1026 { 1027 struct i915_vma *vma; 1028 1029 /* Prevent writes into HWSP after returning the page to the system */ 1030 intel_engine_set_hwsp_writemask(engine, ~0u); 1031 1032 vma = fetch_and_zero(&engine->status_page.vma); 1033 if (!vma) 1034 return; 1035 1036 if (!HWS_NEEDS_PHYSICAL(engine->i915)) 1037 i915_vma_unpin(vma); 1038 1039 i915_gem_object_unpin_map(vma->obj); 1040 i915_gem_object_put(vma->obj); 1041 } 1042 1043 static int pin_ggtt_status_page(struct intel_engine_cs *engine, 1044 struct i915_gem_ww_ctx *ww, 1045 struct i915_vma *vma) 1046 { 1047 unsigned int flags; 1048 1049 if (!HAS_LLC(engine->i915) && i915_ggtt_has_aperture(engine->gt->ggtt)) 1050 /* 1051 * On g33, we cannot place HWS above 256MiB, so 1052 * restrict its pinning to the low mappable arena. 1053 * Though this restriction is not documented for 1054 * gen4, gen5, or byt, they also behave similarly 1055 * and hang if the HWS is placed at the top of the 1056 * GTT. To generalise, it appears that all !llc 1057 * platforms have issues with us placing the HWS 1058 * above the mappable region (even though we never 1059 * actually map it). 1060 */ 1061 flags = PIN_MAPPABLE; 1062 else 1063 flags = PIN_HIGH; 1064 1065 return i915_ggtt_pin(vma, ww, 0, flags); 1066 } 1067 1068 static int init_status_page(struct intel_engine_cs *engine) 1069 { 1070 struct drm_i915_gem_object *obj; 1071 struct i915_gem_ww_ctx ww; 1072 struct i915_vma *vma; 1073 void *vaddr; 1074 int ret; 1075 1076 INIT_LIST_HEAD(&engine->status_page.timelines); 1077 1078 /* 1079 * Though the HWS register does support 36bit addresses, historically 1080 * we have had hangs and corruption reported due to wild writes if 1081 * the HWS is placed above 4G. We only allow objects to be allocated 1082 * in GFP_DMA32 for i965, and no earlier physical address users had 1083 * access to more than 4G. 1084 */ 1085 obj = i915_gem_object_create_internal(engine->i915, PAGE_SIZE); 1086 if (IS_ERR(obj)) { 1087 gt_err(engine->gt, "Failed to allocate status page\n"); 1088 return PTR_ERR(obj); 1089 } 1090 1091 i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC); 1092 1093 vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL); 1094 if (IS_ERR(vma)) { 1095 ret = PTR_ERR(vma); 1096 goto err_put; 1097 } 1098 1099 i915_gem_ww_ctx_init(&ww, true); 1100 retry: 1101 ret = i915_gem_object_lock(obj, &ww); 1102 if (!ret && !HWS_NEEDS_PHYSICAL(engine->i915)) 1103 ret = pin_ggtt_status_page(engine, &ww, vma); 1104 if (ret) 1105 goto err; 1106 1107 vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB); 1108 if (IS_ERR(vaddr)) { 1109 ret = PTR_ERR(vaddr); 1110 goto err_unpin; 1111 } 1112 1113 engine->status_page.addr = memset(vaddr, 0, PAGE_SIZE); 1114 engine->status_page.vma = vma; 1115 1116 err_unpin: 1117 if (ret) 1118 i915_vma_unpin(vma); 1119 err: 1120 if (ret == -EDEADLK) { 1121 ret = i915_gem_ww_ctx_backoff(&ww); 1122 if (!ret) 1123 goto retry; 1124 } 1125 i915_gem_ww_ctx_fini(&ww); 1126 err_put: 1127 if (ret) 1128 i915_gem_object_put(obj); 1129 return ret; 1130 } 1131 1132 static int intel_engine_init_tlb_invalidation(struct intel_engine_cs *engine) 1133 { 1134 static const union intel_engine_tlb_inv_reg gen8_regs[] = { 1135 [RENDER_CLASS].reg = GEN8_RTCR, 1136 [VIDEO_DECODE_CLASS].reg = GEN8_M1TCR, /* , GEN8_M2TCR */ 1137 [VIDEO_ENHANCEMENT_CLASS].reg = GEN8_VTCR, 1138 [COPY_ENGINE_CLASS].reg = GEN8_BTCR, 1139 }; 1140 static const union intel_engine_tlb_inv_reg gen12_regs[] = { 1141 [RENDER_CLASS].reg = GEN12_GFX_TLB_INV_CR, 1142 [VIDEO_DECODE_CLASS].reg = GEN12_VD_TLB_INV_CR, 1143 [VIDEO_ENHANCEMENT_CLASS].reg = GEN12_VE_TLB_INV_CR, 1144 [COPY_ENGINE_CLASS].reg = GEN12_BLT_TLB_INV_CR, 1145 [COMPUTE_CLASS].reg = GEN12_COMPCTX_TLB_INV_CR, 1146 }; 1147 static const union intel_engine_tlb_inv_reg xehp_regs[] = { 1148 [RENDER_CLASS].mcr_reg = XEHP_GFX_TLB_INV_CR, 1149 [VIDEO_DECODE_CLASS].mcr_reg = XEHP_VD_TLB_INV_CR, 1150 [VIDEO_ENHANCEMENT_CLASS].mcr_reg = XEHP_VE_TLB_INV_CR, 1151 [COPY_ENGINE_CLASS].mcr_reg = XEHP_BLT_TLB_INV_CR, 1152 [COMPUTE_CLASS].mcr_reg = XEHP_COMPCTX_TLB_INV_CR, 1153 }; 1154 static const union intel_engine_tlb_inv_reg xelpmp_regs[] = { 1155 [VIDEO_DECODE_CLASS].reg = GEN12_VD_TLB_INV_CR, 1156 [VIDEO_ENHANCEMENT_CLASS].reg = GEN12_VE_TLB_INV_CR, 1157 [OTHER_CLASS].reg = XELPMP_GSC_TLB_INV_CR, 1158 }; 1159 struct drm_i915_private *i915 = engine->i915; 1160 const unsigned int instance = engine->instance; 1161 const unsigned int class = engine->class; 1162 const union intel_engine_tlb_inv_reg *regs; 1163 union intel_engine_tlb_inv_reg reg; 1164 unsigned int num = 0; 1165 u32 val; 1166 1167 /* 1168 * New platforms should not be added with catch-all-newer (>=) 1169 * condition so that any later platform added triggers the below warning 1170 * and in turn mandates a human cross-check of whether the invalidation 1171 * flows have compatible semantics. 1172 * 1173 * For instance with the 11.00 -> 12.00 transition three out of five 1174 * respective engine registers were moved to masked type. Then after the 1175 * 12.00 -> 12.50 transition multi cast handling is required too. 1176 */ 1177 1178 if (engine->gt->type == GT_MEDIA) { 1179 if (MEDIA_VER_FULL(i915) == IP_VER(13, 0)) { 1180 regs = xelpmp_regs; 1181 num = ARRAY_SIZE(xelpmp_regs); 1182 } 1183 } else { 1184 if (GRAPHICS_VER_FULL(i915) == IP_VER(12, 74) || 1185 GRAPHICS_VER_FULL(i915) == IP_VER(12, 71) || 1186 GRAPHICS_VER_FULL(i915) == IP_VER(12, 70) || 1187 GRAPHICS_VER_FULL(i915) == IP_VER(12, 55)) { 1188 regs = xehp_regs; 1189 num = ARRAY_SIZE(xehp_regs); 1190 } else if (GRAPHICS_VER_FULL(i915) == IP_VER(12, 0) || 1191 GRAPHICS_VER_FULL(i915) == IP_VER(12, 10)) { 1192 regs = gen12_regs; 1193 num = ARRAY_SIZE(gen12_regs); 1194 } else if (GRAPHICS_VER(i915) >= 8 && GRAPHICS_VER(i915) <= 11) { 1195 regs = gen8_regs; 1196 num = ARRAY_SIZE(gen8_regs); 1197 } else if (GRAPHICS_VER(i915) < 8) { 1198 return 0; 1199 } 1200 } 1201 1202 if (gt_WARN_ONCE(engine->gt, !num, 1203 "Platform does not implement TLB invalidation!")) 1204 return -ENODEV; 1205 1206 if (gt_WARN_ON_ONCE(engine->gt, 1207 class >= num || 1208 (!regs[class].reg.reg && 1209 !regs[class].mcr_reg.reg))) 1210 return -ERANGE; 1211 1212 reg = regs[class]; 1213 1214 if (regs == xelpmp_regs && class == OTHER_CLASS) { 1215 /* 1216 * There's only a single GSC instance, but it uses register bit 1217 * 1 instead of either 0 or OTHER_GSC_INSTANCE. 1218 */ 1219 GEM_WARN_ON(instance != OTHER_GSC_INSTANCE); 1220 val = 1; 1221 } else if (regs == gen8_regs && class == VIDEO_DECODE_CLASS && instance == 1) { 1222 reg.reg = GEN8_M2TCR; 1223 val = 0; 1224 } else { 1225 val = instance; 1226 } 1227 1228 val = BIT(val); 1229 1230 engine->tlb_inv.mcr = regs == xehp_regs; 1231 engine->tlb_inv.reg = reg; 1232 engine->tlb_inv.done = val; 1233 1234 if (GRAPHICS_VER(i915) >= 12 && 1235 (engine->class == VIDEO_DECODE_CLASS || 1236 engine->class == VIDEO_ENHANCEMENT_CLASS || 1237 engine->class == COMPUTE_CLASS || 1238 engine->class == OTHER_CLASS)) 1239 engine->tlb_inv.request = _MASKED_BIT_ENABLE(val); 1240 else 1241 engine->tlb_inv.request = val; 1242 1243 return 0; 1244 } 1245 1246 static int engine_setup_common(struct intel_engine_cs *engine) 1247 { 1248 int err; 1249 1250 init_llist_head(&engine->barrier_tasks); 1251 1252 err = intel_engine_init_tlb_invalidation(engine); 1253 if (err) 1254 return err; 1255 1256 err = init_status_page(engine); 1257 if (err) 1258 return err; 1259 1260 engine->breadcrumbs = intel_breadcrumbs_create(engine); 1261 if (!engine->breadcrumbs) { 1262 err = -ENOMEM; 1263 goto err_status; 1264 } 1265 1266 engine->sched_engine = i915_sched_engine_create(ENGINE_PHYSICAL); 1267 if (!engine->sched_engine) { 1268 err = -ENOMEM; 1269 goto err_sched_engine; 1270 } 1271 engine->sched_engine->private_data = engine; 1272 1273 err = intel_engine_init_cmd_parser(engine); 1274 if (err) 1275 goto err_cmd_parser; 1276 1277 intel_engine_init_execlists(engine); 1278 intel_engine_init__pm(engine); 1279 intel_engine_init_retire(engine); 1280 1281 /* Use the whole device by default */ 1282 engine->sseu = 1283 intel_sseu_from_device_info(&engine->gt->info.sseu); 1284 1285 intel_engine_init_workarounds(engine); 1286 intel_engine_init_whitelist(engine); 1287 intel_engine_init_ctx_wa(engine); 1288 1289 if (GRAPHICS_VER(engine->i915) >= 12) 1290 engine->flags |= I915_ENGINE_HAS_RELATIVE_MMIO; 1291 1292 return 0; 1293 1294 err_cmd_parser: 1295 i915_sched_engine_put(engine->sched_engine); 1296 err_sched_engine: 1297 intel_breadcrumbs_put(engine->breadcrumbs); 1298 err_status: 1299 cleanup_status_page(engine); 1300 return err; 1301 } 1302 1303 struct measure_breadcrumb { 1304 struct i915_request rq; 1305 struct intel_ring ring; 1306 u32 cs[2048]; 1307 }; 1308 1309 static int measure_breadcrumb_dw(struct intel_context *ce) 1310 { 1311 struct intel_engine_cs *engine = ce->engine; 1312 struct measure_breadcrumb *frame; 1313 int dw; 1314 1315 GEM_BUG_ON(!engine->gt->scratch); 1316 1317 frame = kzalloc(sizeof(*frame), GFP_KERNEL); 1318 if (!frame) 1319 return -ENOMEM; 1320 1321 frame->rq.i915 = engine->i915; 1322 frame->rq.engine = engine; 1323 frame->rq.context = ce; 1324 rcu_assign_pointer(frame->rq.timeline, ce->timeline); 1325 frame->rq.hwsp_seqno = ce->timeline->hwsp_seqno; 1326 1327 frame->ring.vaddr = frame->cs; 1328 frame->ring.size = sizeof(frame->cs); 1329 frame->ring.wrap = 1330 BITS_PER_TYPE(frame->ring.size) - ilog2(frame->ring.size); 1331 frame->ring.effective_size = frame->ring.size; 1332 intel_ring_update_space(&frame->ring); 1333 frame->rq.ring = &frame->ring; 1334 1335 mutex_lock(&ce->timeline->mutex); 1336 spin_lock_irq(&engine->sched_engine->lock); 1337 1338 dw = engine->emit_fini_breadcrumb(&frame->rq, frame->cs) - frame->cs; 1339 1340 spin_unlock_irq(&engine->sched_engine->lock); 1341 mutex_unlock(&ce->timeline->mutex); 1342 1343 GEM_BUG_ON(dw & 1); /* RING_TAIL must be qword aligned */ 1344 1345 kfree(frame); 1346 return dw; 1347 } 1348 1349 struct intel_context * 1350 intel_engine_create_pinned_context(struct intel_engine_cs *engine, 1351 struct i915_address_space *vm, 1352 unsigned int ring_size, 1353 unsigned int hwsp, 1354 struct lock_class_key *key, 1355 const char *name) 1356 { 1357 struct intel_context *ce; 1358 int err; 1359 1360 ce = intel_context_create(engine); 1361 if (IS_ERR(ce)) 1362 return ce; 1363 1364 __set_bit(CONTEXT_BARRIER_BIT, &ce->flags); 1365 ce->timeline = page_pack_bits(NULL, hwsp); 1366 ce->ring = NULL; 1367 ce->ring_size = ring_size; 1368 1369 i915_vm_put(ce->vm); 1370 ce->vm = i915_vm_get(vm); 1371 1372 err = intel_context_pin(ce); /* perma-pin so it is always available */ 1373 if (err) { 1374 intel_context_put(ce); 1375 return ERR_PTR(err); 1376 } 1377 1378 list_add_tail(&ce->pinned_contexts_link, &engine->pinned_contexts_list); 1379 1380 /* 1381 * Give our perma-pinned kernel timelines a separate lockdep class, 1382 * so that we can use them from within the normal user timelines 1383 * should we need to inject GPU operations during their request 1384 * construction. 1385 */ 1386 lockdep_set_class_and_name(&ce->timeline->mutex, key, name); 1387 1388 return ce; 1389 } 1390 1391 void intel_engine_destroy_pinned_context(struct intel_context *ce) 1392 { 1393 struct intel_engine_cs *engine = ce->engine; 1394 struct i915_vma *hwsp = engine->status_page.vma; 1395 1396 GEM_BUG_ON(ce->timeline->hwsp_ggtt != hwsp); 1397 1398 mutex_lock(&hwsp->vm->mutex); 1399 list_del(&ce->timeline->engine_link); 1400 mutex_unlock(&hwsp->vm->mutex); 1401 1402 list_del(&ce->pinned_contexts_link); 1403 intel_context_unpin(ce); 1404 intel_context_put(ce); 1405 } 1406 1407 static struct intel_context * 1408 create_ggtt_bind_context(struct intel_engine_cs *engine) 1409 { 1410 static struct lock_class_key kernel; 1411 1412 /* 1413 * MI_UPDATE_GTT can insert up to 511 PTE entries and there could be multiple 1414 * bind requets at a time so get a bigger ring. 1415 */ 1416 return intel_engine_create_pinned_context(engine, engine->gt->vm, SZ_512K, 1417 I915_GEM_HWS_GGTT_BIND_ADDR, 1418 &kernel, "ggtt_bind_context"); 1419 } 1420 1421 static struct intel_context * 1422 create_kernel_context(struct intel_engine_cs *engine) 1423 { 1424 static struct lock_class_key kernel; 1425 1426 return intel_engine_create_pinned_context(engine, engine->gt->vm, SZ_4K, 1427 I915_GEM_HWS_SEQNO_ADDR, 1428 &kernel, "kernel_context"); 1429 } 1430 1431 /* 1432 * engine_init_common - initialize engine state which might require hw access 1433 * @engine: Engine to initialize. 1434 * 1435 * Initializes @engine@ structure members shared between legacy and execlists 1436 * submission modes which do require hardware access. 1437 * 1438 * Typcally done at later stages of submission mode specific engine setup. 1439 * 1440 * Returns zero on success or an error code on failure. 1441 */ 1442 static int engine_init_common(struct intel_engine_cs *engine) 1443 { 1444 struct intel_context *ce, *bce = NULL; 1445 int ret; 1446 1447 engine->set_default_submission(engine); 1448 1449 /* 1450 * We may need to do things with the shrinker which 1451 * require us to immediately switch back to the default 1452 * context. This can cause a problem as pinning the 1453 * default context also requires GTT space which may not 1454 * be available. To avoid this we always pin the default 1455 * context. 1456 */ 1457 ce = create_kernel_context(engine); 1458 if (IS_ERR(ce)) 1459 return PTR_ERR(ce); 1460 /* 1461 * Create a separate pinned context for GGTT update with blitter engine 1462 * if a platform require such service. MI_UPDATE_GTT works on other 1463 * engines as well but BCS should be less busy engine so pick that for 1464 * GGTT updates. 1465 */ 1466 if (i915_ggtt_require_binder(engine->i915) && engine->id == BCS0) { 1467 bce = create_ggtt_bind_context(engine); 1468 if (IS_ERR(bce)) { 1469 ret = PTR_ERR(bce); 1470 goto err_ce_context; 1471 } 1472 } 1473 1474 ret = measure_breadcrumb_dw(ce); 1475 if (ret < 0) 1476 goto err_bce_context; 1477 1478 engine->emit_fini_breadcrumb_dw = ret; 1479 engine->kernel_context = ce; 1480 engine->bind_context = bce; 1481 1482 return 0; 1483 1484 err_bce_context: 1485 if (bce) 1486 intel_engine_destroy_pinned_context(bce); 1487 err_ce_context: 1488 intel_engine_destroy_pinned_context(ce); 1489 return ret; 1490 } 1491 1492 int intel_engines_init(struct intel_gt *gt) 1493 { 1494 int (*setup)(struct intel_engine_cs *engine); 1495 struct intel_engine_cs *engine; 1496 enum intel_engine_id id; 1497 int err; 1498 1499 if (intel_uc_uses_guc_submission(>->uc)) { 1500 gt->submission_method = INTEL_SUBMISSION_GUC; 1501 setup = intel_guc_submission_setup; 1502 } else if (HAS_EXECLISTS(gt->i915)) { 1503 gt->submission_method = INTEL_SUBMISSION_ELSP; 1504 setup = intel_execlists_submission_setup; 1505 } else { 1506 gt->submission_method = INTEL_SUBMISSION_RING; 1507 setup = intel_ring_submission_setup; 1508 } 1509 1510 for_each_engine(engine, gt, id) { 1511 err = engine_setup_common(engine); 1512 if (err) 1513 return err; 1514 1515 err = setup(engine); 1516 if (err) { 1517 intel_engine_cleanup_common(engine); 1518 return err; 1519 } 1520 1521 /* The backend should now be responsible for cleanup */ 1522 GEM_BUG_ON(engine->release == NULL); 1523 1524 err = engine_init_common(engine); 1525 if (err) 1526 return err; 1527 1528 intel_engine_add_user(engine); 1529 } 1530 1531 return 0; 1532 } 1533 1534 /** 1535 * intel_engine_cleanup_common - cleans up the engine state created by 1536 * the common initiailizers. 1537 * @engine: Engine to cleanup. 1538 * 1539 * This cleans up everything created by the common helpers. 1540 */ 1541 void intel_engine_cleanup_common(struct intel_engine_cs *engine) 1542 { 1543 GEM_BUG_ON(!list_empty(&engine->sched_engine->requests)); 1544 1545 i915_sched_engine_put(engine->sched_engine); 1546 intel_breadcrumbs_put(engine->breadcrumbs); 1547 1548 intel_engine_fini_retire(engine); 1549 intel_engine_cleanup_cmd_parser(engine); 1550 1551 if (engine->default_state) 1552 fput(engine->default_state); 1553 1554 if (engine->kernel_context) 1555 intel_engine_destroy_pinned_context(engine->kernel_context); 1556 1557 if (engine->bind_context) 1558 intel_engine_destroy_pinned_context(engine->bind_context); 1559 1560 1561 GEM_BUG_ON(!llist_empty(&engine->barrier_tasks)); 1562 cleanup_status_page(engine); 1563 1564 intel_wa_list_free(&engine->ctx_wa_list); 1565 intel_wa_list_free(&engine->wa_list); 1566 intel_wa_list_free(&engine->whitelist); 1567 } 1568 1569 /** 1570 * intel_engine_resume - re-initializes the HW state of the engine 1571 * @engine: Engine to resume. 1572 * 1573 * Returns zero on success or an error code on failure. 1574 */ 1575 int intel_engine_resume(struct intel_engine_cs *engine) 1576 { 1577 intel_engine_apply_workarounds(engine); 1578 intel_engine_apply_whitelist(engine); 1579 1580 return engine->resume(engine); 1581 } 1582 1583 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine) 1584 { 1585 struct drm_i915_private *i915 = engine->i915; 1586 1587 u64 acthd; 1588 1589 if (GRAPHICS_VER(i915) >= 8) 1590 acthd = ENGINE_READ64(engine, RING_ACTHD, RING_ACTHD_UDW); 1591 else if (GRAPHICS_VER(i915) >= 4) 1592 acthd = ENGINE_READ(engine, RING_ACTHD); 1593 else 1594 acthd = ENGINE_READ(engine, ACTHD); 1595 1596 return acthd; 1597 } 1598 1599 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine) 1600 { 1601 u64 bbaddr; 1602 1603 if (GRAPHICS_VER(engine->i915) >= 8) 1604 bbaddr = ENGINE_READ64(engine, RING_BBADDR, RING_BBADDR_UDW); 1605 else 1606 bbaddr = ENGINE_READ(engine, RING_BBADDR); 1607 1608 return bbaddr; 1609 } 1610 1611 static unsigned long stop_timeout(const struct intel_engine_cs *engine) 1612 { 1613 if (in_atomic() || irqs_disabled()) /* inside atomic preempt-reset? */ 1614 return 0; 1615 1616 /* 1617 * If we are doing a normal GPU reset, we can take our time and allow 1618 * the engine to quiesce. We've stopped submission to the engine, and 1619 * if we wait long enough an innocent context should complete and 1620 * leave the engine idle. So they should not be caught unaware by 1621 * the forthcoming GPU reset (which usually follows the stop_cs)! 1622 */ 1623 return READ_ONCE(engine->props.stop_timeout_ms); 1624 } 1625 1626 static int __intel_engine_stop_cs(struct intel_engine_cs *engine, 1627 int fast_timeout_us, 1628 int slow_timeout_ms) 1629 { 1630 struct intel_uncore *uncore = engine->uncore; 1631 const i915_reg_t mode = RING_MI_MODE(engine->mmio_base); 1632 int err; 1633 1634 intel_uncore_write_fw(uncore, mode, _MASKED_BIT_ENABLE(STOP_RING)); 1635 1636 /* 1637 * Wa_22011802037: Prior to doing a reset, ensure CS is 1638 * stopped, set ring stop bit and prefetch disable bit to halt CS 1639 */ 1640 if (intel_engine_reset_needs_wa_22011802037(engine->gt)) 1641 intel_uncore_write_fw(uncore, RING_MODE_GEN7(engine->mmio_base), 1642 _MASKED_BIT_ENABLE(GEN12_GFX_PREFETCH_DISABLE)); 1643 1644 err = __intel_wait_for_register_fw(engine->uncore, mode, 1645 MODE_IDLE, MODE_IDLE, 1646 fast_timeout_us, 1647 slow_timeout_ms, 1648 NULL); 1649 1650 /* A final mmio read to let GPU writes be hopefully flushed to memory */ 1651 intel_uncore_posting_read_fw(uncore, mode); 1652 return err; 1653 } 1654 1655 int intel_engine_stop_cs(struct intel_engine_cs *engine) 1656 { 1657 int err = 0; 1658 1659 if (GRAPHICS_VER(engine->i915) < 3) 1660 return -ENODEV; 1661 1662 ENGINE_TRACE(engine, "\n"); 1663 /* 1664 * TODO: Find out why occasionally stopping the CS times out. Seen 1665 * especially with gem_eio tests. 1666 * 1667 * Occasionally trying to stop the cs times out, but does not adversely 1668 * affect functionality. The timeout is set as a config parameter that 1669 * defaults to 100ms. In most cases the follow up operation is to wait 1670 * for pending MI_FORCE_WAKES. The assumption is that this timeout is 1671 * sufficient for any pending MI_FORCEWAKEs to complete. Once root 1672 * caused, the caller must check and handle the return from this 1673 * function. 1674 */ 1675 if (__intel_engine_stop_cs(engine, 1000, stop_timeout(engine))) { 1676 ENGINE_TRACE(engine, 1677 "timed out on STOP_RING -> IDLE; HEAD:%04x, TAIL:%04x\n", 1678 ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR, 1679 ENGINE_READ_FW(engine, RING_TAIL) & TAIL_ADDR); 1680 1681 /* 1682 * Sometimes we observe that the idle flag is not 1683 * set even though the ring is empty. So double 1684 * check before giving up. 1685 */ 1686 if ((ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR) != 1687 (ENGINE_READ_FW(engine, RING_TAIL) & TAIL_ADDR)) 1688 err = -ETIMEDOUT; 1689 } 1690 1691 return err; 1692 } 1693 1694 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine) 1695 { 1696 ENGINE_TRACE(engine, "\n"); 1697 1698 ENGINE_WRITE_FW(engine, RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING)); 1699 } 1700 1701 static u32 __cs_pending_mi_force_wakes(struct intel_engine_cs *engine) 1702 { 1703 static const i915_reg_t _reg[I915_NUM_ENGINES] = { 1704 [RCS0] = MSG_IDLE_CS, 1705 [BCS0] = MSG_IDLE_BCS, 1706 [VCS0] = MSG_IDLE_VCS0, 1707 [VCS1] = MSG_IDLE_VCS1, 1708 [VCS2] = MSG_IDLE_VCS2, 1709 [VCS3] = MSG_IDLE_VCS3, 1710 [VCS4] = MSG_IDLE_VCS4, 1711 [VCS5] = MSG_IDLE_VCS5, 1712 [VCS6] = MSG_IDLE_VCS6, 1713 [VCS7] = MSG_IDLE_VCS7, 1714 [VECS0] = MSG_IDLE_VECS0, 1715 [VECS1] = MSG_IDLE_VECS1, 1716 [VECS2] = MSG_IDLE_VECS2, 1717 [VECS3] = MSG_IDLE_VECS3, 1718 [CCS0] = MSG_IDLE_CS, 1719 [CCS1] = MSG_IDLE_CS, 1720 [CCS2] = MSG_IDLE_CS, 1721 [CCS3] = MSG_IDLE_CS, 1722 }; 1723 u32 val; 1724 1725 if (!_reg[engine->id].reg) 1726 return 0; 1727 1728 val = intel_uncore_read(engine->uncore, _reg[engine->id]); 1729 1730 /* bits[29:25] & bits[13:9] >> shift */ 1731 return (val & (val >> 16) & MSG_IDLE_FW_MASK) >> MSG_IDLE_FW_SHIFT; 1732 } 1733 1734 static void __gpm_wait_for_fw_complete(struct intel_gt *gt, u32 fw_mask) 1735 { 1736 int ret; 1737 1738 /* Ensure GPM receives fw up/down after CS is stopped */ 1739 udelay(1); 1740 1741 /* Wait for forcewake request to complete in GPM */ 1742 ret = __intel_wait_for_register_fw(gt->uncore, 1743 GEN9_PWRGT_DOMAIN_STATUS, 1744 fw_mask, fw_mask, 5000, 0, NULL); 1745 1746 /* Ensure CS receives fw ack from GPM */ 1747 udelay(1); 1748 1749 if (ret) 1750 GT_TRACE(gt, "Failed to complete pending forcewake %d\n", ret); 1751 } 1752 1753 /* 1754 * Wa_22011802037:gen12: In addition to stopping the cs, we need to wait for any 1755 * pending MI_FORCE_WAKEUP requests that the CS has initiated to complete. The 1756 * pending status is indicated by bits[13:9] (masked by bits[29:25]) in the 1757 * MSG_IDLE register. There's one MSG_IDLE register per reset domain. Since we 1758 * are concerned only with the gt reset here, we use a logical OR of pending 1759 * forcewakeups from all reset domains and then wait for them to complete by 1760 * querying PWRGT_DOMAIN_STATUS. 1761 */ 1762 void intel_engine_wait_for_pending_mi_fw(struct intel_engine_cs *engine) 1763 { 1764 u32 fw_pending = __cs_pending_mi_force_wakes(engine); 1765 1766 if (fw_pending) 1767 __gpm_wait_for_fw_complete(engine->gt, fw_pending); 1768 } 1769 1770 /* NB: please notice the memset */ 1771 void intel_engine_get_instdone(const struct intel_engine_cs *engine, 1772 struct intel_instdone *instdone) 1773 { 1774 struct drm_i915_private *i915 = engine->i915; 1775 struct intel_uncore *uncore = engine->uncore; 1776 u32 mmio_base = engine->mmio_base; 1777 int slice; 1778 int subslice; 1779 int iter; 1780 1781 memset(instdone, 0, sizeof(*instdone)); 1782 1783 if (GRAPHICS_VER(i915) >= 8) { 1784 instdone->instdone = 1785 intel_uncore_read(uncore, RING_INSTDONE(mmio_base)); 1786 1787 if (engine->id != RCS0) 1788 return; 1789 1790 instdone->slice_common = 1791 intel_uncore_read(uncore, GEN7_SC_INSTDONE); 1792 if (GRAPHICS_VER(i915) >= 12) { 1793 instdone->slice_common_extra[0] = 1794 intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA); 1795 instdone->slice_common_extra[1] = 1796 intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA2); 1797 } 1798 1799 for_each_ss_steering(iter, engine->gt, slice, subslice) { 1800 instdone->sampler[slice][subslice] = 1801 intel_gt_mcr_read(engine->gt, 1802 GEN8_SAMPLER_INSTDONE, 1803 slice, subslice); 1804 instdone->row[slice][subslice] = 1805 intel_gt_mcr_read(engine->gt, 1806 GEN8_ROW_INSTDONE, 1807 slice, subslice); 1808 } 1809 1810 if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 55)) { 1811 for_each_ss_steering(iter, engine->gt, slice, subslice) 1812 instdone->geom_svg[slice][subslice] = 1813 intel_gt_mcr_read(engine->gt, 1814 XEHPG_INSTDONE_GEOM_SVG, 1815 slice, subslice); 1816 } 1817 } else if (GRAPHICS_VER(i915) >= 7) { 1818 instdone->instdone = 1819 intel_uncore_read(uncore, RING_INSTDONE(mmio_base)); 1820 1821 if (engine->id != RCS0) 1822 return; 1823 1824 instdone->slice_common = 1825 intel_uncore_read(uncore, GEN7_SC_INSTDONE); 1826 instdone->sampler[0][0] = 1827 intel_uncore_read(uncore, GEN7_SAMPLER_INSTDONE); 1828 instdone->row[0][0] = 1829 intel_uncore_read(uncore, GEN7_ROW_INSTDONE); 1830 } else if (GRAPHICS_VER(i915) >= 4) { 1831 instdone->instdone = 1832 intel_uncore_read(uncore, RING_INSTDONE(mmio_base)); 1833 if (engine->id == RCS0) 1834 /* HACK: Using the wrong struct member */ 1835 instdone->slice_common = 1836 intel_uncore_read(uncore, GEN4_INSTDONE1); 1837 } else { 1838 instdone->instdone = intel_uncore_read(uncore, GEN2_INSTDONE); 1839 } 1840 } 1841 1842 static bool ring_is_idle(struct intel_engine_cs *engine) 1843 { 1844 bool idle = true; 1845 1846 if (I915_SELFTEST_ONLY(!engine->mmio_base)) 1847 return true; 1848 1849 if (!intel_engine_pm_get_if_awake(engine)) 1850 return true; 1851 1852 /* First check that no commands are left in the ring */ 1853 if ((ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) != 1854 (ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR)) 1855 idle = false; 1856 1857 /* No bit for gen2, so assume the CS parser is idle */ 1858 if (GRAPHICS_VER(engine->i915) > 2 && 1859 !(ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE)) 1860 idle = false; 1861 1862 intel_engine_pm_put(engine); 1863 1864 return idle; 1865 } 1866 1867 void __intel_engine_flush_submission(struct intel_engine_cs *engine, bool sync) 1868 { 1869 struct tasklet_struct *t = &engine->sched_engine->tasklet; 1870 1871 if (!t->callback) 1872 return; 1873 1874 local_bh_disable(); 1875 if (tasklet_trylock(t)) { 1876 /* Must wait for any GPU reset in progress. */ 1877 if (__tasklet_is_enabled(t)) 1878 t->callback(t); 1879 tasklet_unlock(t); 1880 } 1881 local_bh_enable(); 1882 1883 /* Synchronise and wait for the tasklet on another CPU */ 1884 if (sync) 1885 tasklet_unlock_wait(t); 1886 } 1887 1888 /** 1889 * intel_engine_is_idle() - Report if the engine has finished process all work 1890 * @engine: the intel_engine_cs 1891 * 1892 * Return true if there are no requests pending, nothing left to be submitted 1893 * to hardware, and that the engine is idle. 1894 */ 1895 bool intel_engine_is_idle(struct intel_engine_cs *engine) 1896 { 1897 /* More white lies, if wedged, hw state is inconsistent */ 1898 if (intel_gt_is_wedged(engine->gt)) 1899 return true; 1900 1901 if (!intel_engine_pm_is_awake(engine)) 1902 return true; 1903 1904 /* Waiting to drain ELSP? */ 1905 intel_synchronize_hardirq(engine->i915); 1906 intel_engine_flush_submission(engine); 1907 1908 /* ELSP is empty, but there are ready requests? E.g. after reset */ 1909 if (!i915_sched_engine_is_empty(engine->sched_engine)) 1910 return false; 1911 1912 /* Ring stopped? */ 1913 return ring_is_idle(engine); 1914 } 1915 1916 bool intel_engines_are_idle(struct intel_gt *gt) 1917 { 1918 struct intel_engine_cs *engine; 1919 enum intel_engine_id id; 1920 1921 /* 1922 * If the driver is wedged, HW state may be very inconsistent and 1923 * report that it is still busy, even though we have stopped using it. 1924 */ 1925 if (intel_gt_is_wedged(gt)) 1926 return true; 1927 1928 /* Already parked (and passed an idleness test); must still be idle */ 1929 if (!READ_ONCE(gt->awake)) 1930 return true; 1931 1932 for_each_engine(engine, gt, id) { 1933 if (!intel_engine_is_idle(engine)) 1934 return false; 1935 } 1936 1937 return true; 1938 } 1939 1940 bool intel_engine_irq_enable(struct intel_engine_cs *engine) 1941 { 1942 if (!engine->irq_enable) 1943 return false; 1944 1945 /* Caller disables interrupts */ 1946 spin_lock(engine->gt->irq_lock); 1947 engine->irq_enable(engine); 1948 spin_unlock(engine->gt->irq_lock); 1949 1950 return true; 1951 } 1952 1953 void intel_engine_irq_disable(struct intel_engine_cs *engine) 1954 { 1955 if (!engine->irq_disable) 1956 return; 1957 1958 /* Caller disables interrupts */ 1959 spin_lock(engine->gt->irq_lock); 1960 engine->irq_disable(engine); 1961 spin_unlock(engine->gt->irq_lock); 1962 } 1963 1964 void intel_engines_reset_default_submission(struct intel_gt *gt) 1965 { 1966 struct intel_engine_cs *engine; 1967 enum intel_engine_id id; 1968 1969 for_each_engine(engine, gt, id) { 1970 if (engine->sanitize) 1971 engine->sanitize(engine); 1972 1973 engine->set_default_submission(engine); 1974 } 1975 } 1976 1977 bool intel_engine_can_store_dword(struct intel_engine_cs *engine) 1978 { 1979 switch (GRAPHICS_VER(engine->i915)) { 1980 case 2: 1981 return false; /* uses physical not virtual addresses */ 1982 case 3: 1983 /* maybe only uses physical not virtual addresses */ 1984 return !(IS_I915G(engine->i915) || IS_I915GM(engine->i915)); 1985 case 4: 1986 return !IS_I965G(engine->i915); /* who knows! */ 1987 case 6: 1988 return engine->class != VIDEO_DECODE_CLASS; /* b0rked */ 1989 default: 1990 return true; 1991 } 1992 } 1993 1994 static struct intel_timeline *get_timeline(struct i915_request *rq) 1995 { 1996 struct intel_timeline *tl; 1997 1998 /* 1999 * Even though we are holding the engine->sched_engine->lock here, there 2000 * is no control over the submission queue per-se and we are 2001 * inspecting the active state at a random point in time, with an 2002 * unknown queue. Play safe and make sure the timeline remains valid. 2003 * (Only being used for pretty printing, one extra kref shouldn't 2004 * cause a camel stampede!) 2005 */ 2006 rcu_read_lock(); 2007 tl = rcu_dereference(rq->timeline); 2008 if (!kref_get_unless_zero(&tl->kref)) 2009 tl = NULL; 2010 rcu_read_unlock(); 2011 2012 return tl; 2013 } 2014 2015 static int print_ring(char *buf, int sz, struct i915_request *rq) 2016 { 2017 int len = 0; 2018 2019 if (!i915_request_signaled(rq)) { 2020 struct intel_timeline *tl = get_timeline(rq); 2021 2022 len = scnprintf(buf, sz, 2023 "ring:{start:%08x, hwsp:%08x, seqno:%08x, runtime:%llums}, ", 2024 i915_ggtt_offset(rq->ring->vma), 2025 tl ? tl->hwsp_offset : 0, 2026 hwsp_seqno(rq), 2027 DIV_ROUND_CLOSEST_ULL(intel_context_get_total_runtime_ns(rq->context), 2028 1000 * 1000)); 2029 2030 if (tl) 2031 intel_timeline_put(tl); 2032 } 2033 2034 return len; 2035 } 2036 2037 static void hexdump(struct drm_printer *m, const void *buf, size_t len) 2038 { 2039 const size_t rowsize = 8 * sizeof(u32); 2040 const void *prev = NULL; 2041 bool skip = false; 2042 size_t pos; 2043 2044 for (pos = 0; pos < len; pos += rowsize) { 2045 char line[128]; 2046 2047 if (prev && !memcmp(prev, buf + pos, rowsize)) { 2048 if (!skip) { 2049 drm_printf(m, "*\n"); 2050 skip = true; 2051 } 2052 continue; 2053 } 2054 2055 WARN_ON_ONCE(hex_dump_to_buffer(buf + pos, len - pos, 2056 rowsize, sizeof(u32), 2057 line, sizeof(line), 2058 false) >= sizeof(line)); 2059 drm_printf(m, "[%04zx] %s\n", pos, line); 2060 2061 prev = buf + pos; 2062 skip = false; 2063 } 2064 } 2065 2066 static const char *repr_timer(const struct timer_list *t) 2067 { 2068 if (!READ_ONCE(t->expires)) 2069 return "inactive"; 2070 2071 if (timer_pending(t)) 2072 return "active"; 2073 2074 return "expired"; 2075 } 2076 2077 static void intel_engine_print_registers(struct intel_engine_cs *engine, 2078 struct drm_printer *m) 2079 { 2080 struct drm_i915_private *i915 = engine->i915; 2081 struct intel_engine_execlists * const execlists = &engine->execlists; 2082 u64 addr; 2083 2084 if (engine->id == RENDER_CLASS && IS_GRAPHICS_VER(i915, 4, 7)) 2085 drm_printf(m, "\tCCID: 0x%08x\n", ENGINE_READ(engine, CCID)); 2086 if (HAS_EXECLISTS(i915)) { 2087 drm_printf(m, "\tEL_STAT_HI: 0x%08x\n", 2088 ENGINE_READ(engine, RING_EXECLIST_STATUS_HI)); 2089 drm_printf(m, "\tEL_STAT_LO: 0x%08x\n", 2090 ENGINE_READ(engine, RING_EXECLIST_STATUS_LO)); 2091 } 2092 drm_printf(m, "\tRING_START: 0x%08x\n", 2093 ENGINE_READ(engine, RING_START)); 2094 drm_printf(m, "\tRING_HEAD: 0x%08x\n", 2095 ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR); 2096 drm_printf(m, "\tRING_TAIL: 0x%08x\n", 2097 ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR); 2098 drm_printf(m, "\tRING_CTL: 0x%08x%s\n", 2099 ENGINE_READ(engine, RING_CTL), 2100 ENGINE_READ(engine, RING_CTL) & (RING_WAIT | RING_WAIT_SEMAPHORE) ? " [waiting]" : ""); 2101 if (GRAPHICS_VER(engine->i915) > 2) { 2102 drm_printf(m, "\tRING_MODE: 0x%08x%s\n", 2103 ENGINE_READ(engine, RING_MI_MODE), 2104 ENGINE_READ(engine, RING_MI_MODE) & (MODE_IDLE) ? " [idle]" : ""); 2105 } 2106 2107 if (GRAPHICS_VER(i915) >= 6) { 2108 drm_printf(m, "\tRING_IMR: 0x%08x\n", 2109 ENGINE_READ(engine, RING_IMR)); 2110 drm_printf(m, "\tRING_ESR: 0x%08x\n", 2111 ENGINE_READ(engine, RING_ESR)); 2112 drm_printf(m, "\tRING_EMR: 0x%08x\n", 2113 ENGINE_READ(engine, RING_EMR)); 2114 drm_printf(m, "\tRING_EIR: 0x%08x\n", 2115 ENGINE_READ(engine, RING_EIR)); 2116 } 2117 2118 addr = intel_engine_get_active_head(engine); 2119 drm_printf(m, "\tACTHD: 0x%08x_%08x\n", 2120 upper_32_bits(addr), lower_32_bits(addr)); 2121 addr = intel_engine_get_last_batch_head(engine); 2122 drm_printf(m, "\tBBADDR: 0x%08x_%08x\n", 2123 upper_32_bits(addr), lower_32_bits(addr)); 2124 if (GRAPHICS_VER(i915) >= 8) 2125 addr = ENGINE_READ64(engine, RING_DMA_FADD, RING_DMA_FADD_UDW); 2126 else if (GRAPHICS_VER(i915) >= 4) 2127 addr = ENGINE_READ(engine, RING_DMA_FADD); 2128 else 2129 addr = ENGINE_READ(engine, DMA_FADD_I8XX); 2130 drm_printf(m, "\tDMA_FADDR: 0x%08x_%08x\n", 2131 upper_32_bits(addr), lower_32_bits(addr)); 2132 if (GRAPHICS_VER(i915) >= 4) { 2133 drm_printf(m, "\tIPEIR: 0x%08x\n", 2134 ENGINE_READ(engine, RING_IPEIR)); 2135 drm_printf(m, "\tIPEHR: 0x%08x\n", 2136 ENGINE_READ(engine, RING_IPEHR)); 2137 } else { 2138 drm_printf(m, "\tIPEIR: 0x%08x\n", ENGINE_READ(engine, IPEIR)); 2139 drm_printf(m, "\tIPEHR: 0x%08x\n", ENGINE_READ(engine, IPEHR)); 2140 } 2141 2142 if (HAS_EXECLISTS(i915) && !intel_engine_uses_guc(engine)) { 2143 struct i915_request * const *port, *rq; 2144 const u32 *hws = 2145 &engine->status_page.addr[I915_HWS_CSB_BUF0_INDEX]; 2146 const u8 num_entries = execlists->csb_size; 2147 unsigned int idx; 2148 u8 read, write; 2149 2150 drm_printf(m, "\tExeclist tasklet queued? %s (%s), preempt? %s, timeslice? %s\n", 2151 str_yes_no(test_bit(TASKLET_STATE_SCHED, &engine->sched_engine->tasklet.state)), 2152 str_enabled_disabled(!atomic_read(&engine->sched_engine->tasklet.count)), 2153 repr_timer(&engine->execlists.preempt), 2154 repr_timer(&engine->execlists.timer)); 2155 2156 read = execlists->csb_head; 2157 write = READ_ONCE(*execlists->csb_write); 2158 2159 drm_printf(m, "\tExeclist status: 0x%08x %08x; CSB read:%d, write:%d, entries:%d\n", 2160 ENGINE_READ(engine, RING_EXECLIST_STATUS_LO), 2161 ENGINE_READ(engine, RING_EXECLIST_STATUS_HI), 2162 read, write, num_entries); 2163 2164 if (read >= num_entries) 2165 read = 0; 2166 if (write >= num_entries) 2167 write = 0; 2168 if (read > write) 2169 write += num_entries; 2170 while (read < write) { 2171 idx = ++read % num_entries; 2172 drm_printf(m, "\tExeclist CSB[%d]: 0x%08x, context: %d\n", 2173 idx, hws[idx * 2], hws[idx * 2 + 1]); 2174 } 2175 2176 i915_sched_engine_active_lock_bh(engine->sched_engine); 2177 rcu_read_lock(); 2178 for (port = execlists->active; (rq = *port); port++) { 2179 char hdr[160]; 2180 int len; 2181 2182 len = scnprintf(hdr, sizeof(hdr), 2183 "\t\tActive[%d]: ccid:%08x%s%s, ", 2184 (int)(port - execlists->active), 2185 rq->context->lrc.ccid, 2186 intel_context_is_closed(rq->context) ? "!" : "", 2187 intel_context_is_banned(rq->context) ? "*" : ""); 2188 len += print_ring(hdr + len, sizeof(hdr) - len, rq); 2189 scnprintf(hdr + len, sizeof(hdr) - len, "rq: "); 2190 i915_request_show(m, rq, hdr, 0); 2191 } 2192 for (port = execlists->pending; (rq = *port); port++) { 2193 char hdr[160]; 2194 int len; 2195 2196 len = scnprintf(hdr, sizeof(hdr), 2197 "\t\tPending[%d]: ccid:%08x%s%s, ", 2198 (int)(port - execlists->pending), 2199 rq->context->lrc.ccid, 2200 intel_context_is_closed(rq->context) ? "!" : "", 2201 intel_context_is_banned(rq->context) ? "*" : ""); 2202 len += print_ring(hdr + len, sizeof(hdr) - len, rq); 2203 scnprintf(hdr + len, sizeof(hdr) - len, "rq: "); 2204 i915_request_show(m, rq, hdr, 0); 2205 } 2206 rcu_read_unlock(); 2207 i915_sched_engine_active_unlock_bh(engine->sched_engine); 2208 } else if (GRAPHICS_VER(i915) > 6) { 2209 drm_printf(m, "\tPP_DIR_BASE: 0x%08x\n", 2210 ENGINE_READ(engine, RING_PP_DIR_BASE)); 2211 drm_printf(m, "\tPP_DIR_BASE_READ: 0x%08x\n", 2212 ENGINE_READ(engine, RING_PP_DIR_BASE_READ)); 2213 drm_printf(m, "\tPP_DIR_DCLV: 0x%08x\n", 2214 ENGINE_READ(engine, RING_PP_DIR_DCLV)); 2215 } 2216 } 2217 2218 static void print_request_ring(struct drm_printer *m, struct i915_request *rq) 2219 { 2220 struct i915_vma_resource *vma_res = rq->batch_res; 2221 void *ring; 2222 int size; 2223 2224 drm_printf(m, 2225 "[head %04x, postfix %04x, tail %04x, batch 0x%08x_%08x]:\n", 2226 rq->head, rq->postfix, rq->tail, 2227 vma_res ? upper_32_bits(vma_res->start) : ~0u, 2228 vma_res ? lower_32_bits(vma_res->start) : ~0u); 2229 2230 size = rq->tail - rq->head; 2231 if (rq->tail < rq->head) 2232 size += rq->ring->size; 2233 2234 ring = kmalloc(size, GFP_ATOMIC); 2235 if (ring) { 2236 const void *vaddr = rq->ring->vaddr; 2237 unsigned int head = rq->head; 2238 unsigned int len = 0; 2239 2240 if (rq->tail < head) { 2241 len = rq->ring->size - head; 2242 memcpy(ring, vaddr + head, len); 2243 head = 0; 2244 } 2245 memcpy(ring + len, vaddr + head, size - len); 2246 2247 hexdump(m, ring, size); 2248 kfree(ring); 2249 } 2250 } 2251 2252 static unsigned long read_ul(void *p, size_t x) 2253 { 2254 return *(unsigned long *)(p + x); 2255 } 2256 2257 static void print_properties(struct intel_engine_cs *engine, 2258 struct drm_printer *m) 2259 { 2260 static const struct pmap { 2261 size_t offset; 2262 const char *name; 2263 } props[] = { 2264 #define P(x) { \ 2265 .offset = offsetof(typeof(engine->props), x), \ 2266 .name = #x \ 2267 } 2268 P(heartbeat_interval_ms), 2269 P(max_busywait_duration_ns), 2270 P(preempt_timeout_ms), 2271 P(stop_timeout_ms), 2272 P(timeslice_duration_ms), 2273 2274 {}, 2275 #undef P 2276 }; 2277 const struct pmap *p; 2278 2279 drm_printf(m, "\tProperties:\n"); 2280 for (p = props; p->name; p++) 2281 drm_printf(m, "\t\t%s: %lu [default %lu]\n", 2282 p->name, 2283 read_ul(&engine->props, p->offset), 2284 read_ul(&engine->defaults, p->offset)); 2285 } 2286 2287 static void engine_dump_request(struct i915_request *rq, struct drm_printer *m, const char *msg) 2288 { 2289 struct intel_timeline *tl = get_timeline(rq); 2290 2291 i915_request_show(m, rq, msg, 0); 2292 2293 drm_printf(m, "\t\tring->start: 0x%08x\n", 2294 i915_ggtt_offset(rq->ring->vma)); 2295 drm_printf(m, "\t\tring->head: 0x%08x\n", 2296 rq->ring->head); 2297 drm_printf(m, "\t\tring->tail: 0x%08x\n", 2298 rq->ring->tail); 2299 drm_printf(m, "\t\tring->emit: 0x%08x\n", 2300 rq->ring->emit); 2301 drm_printf(m, "\t\tring->space: 0x%08x\n", 2302 rq->ring->space); 2303 2304 if (tl) { 2305 drm_printf(m, "\t\tring->hwsp: 0x%08x\n", 2306 tl->hwsp_offset); 2307 intel_timeline_put(tl); 2308 } 2309 2310 print_request_ring(m, rq); 2311 2312 if (rq->context->lrc_reg_state) { 2313 drm_printf(m, "Logical Ring Context:\n"); 2314 hexdump(m, rq->context->lrc_reg_state, PAGE_SIZE); 2315 } 2316 } 2317 2318 void intel_engine_dump_active_requests(struct list_head *requests, 2319 struct i915_request *hung_rq, 2320 struct drm_printer *m) 2321 { 2322 struct i915_request *rq; 2323 const char *msg; 2324 enum i915_request_state state; 2325 2326 list_for_each_entry(rq, requests, sched.link) { 2327 if (rq == hung_rq) 2328 continue; 2329 2330 state = i915_test_request_state(rq); 2331 if (state < I915_REQUEST_QUEUED) 2332 continue; 2333 2334 if (state == I915_REQUEST_ACTIVE) 2335 msg = "\t\tactive on engine"; 2336 else 2337 msg = "\t\tactive in queue"; 2338 2339 engine_dump_request(rq, m, msg); 2340 } 2341 } 2342 2343 static void engine_dump_active_requests(struct intel_engine_cs *engine, 2344 struct drm_printer *m) 2345 { 2346 struct intel_context *hung_ce = NULL; 2347 struct i915_request *hung_rq = NULL; 2348 2349 /* 2350 * No need for an engine->irq_seqno_barrier() before the seqno reads. 2351 * The GPU is still running so requests are still executing and any 2352 * hardware reads will be out of date by the time they are reported. 2353 * But the intention here is just to report an instantaneous snapshot 2354 * so that's fine. 2355 */ 2356 intel_engine_get_hung_entity(engine, &hung_ce, &hung_rq); 2357 2358 drm_printf(m, "\tRequests:\n"); 2359 2360 if (hung_rq) 2361 engine_dump_request(hung_rq, m, "\t\thung"); 2362 else if (hung_ce) 2363 drm_printf(m, "\t\tGot hung ce but no hung rq!\n"); 2364 2365 if (intel_uc_uses_guc_submission(&engine->gt->uc)) 2366 intel_guc_dump_active_requests(engine, hung_rq, m); 2367 else 2368 intel_execlists_dump_active_requests(engine, hung_rq, m); 2369 2370 if (hung_rq) 2371 i915_request_put(hung_rq); 2372 } 2373 2374 void intel_engine_dump(struct intel_engine_cs *engine, 2375 struct drm_printer *m, 2376 const char *header, ...) 2377 { 2378 struct i915_gpu_error * const error = &engine->i915->gpu_error; 2379 struct i915_request *rq; 2380 intel_wakeref_t wakeref; 2381 ktime_t dummy; 2382 2383 if (header) { 2384 va_list ap; 2385 2386 va_start(ap, header); 2387 drm_vprintf(m, header, &ap); 2388 va_end(ap); 2389 } 2390 2391 if (intel_gt_is_wedged(engine->gt)) 2392 drm_printf(m, "*** WEDGED ***\n"); 2393 2394 drm_printf(m, "\tAwake? %d\n", atomic_read(&engine->wakeref.count)); 2395 drm_printf(m, "\tBarriers?: %s\n", 2396 str_yes_no(!llist_empty(&engine->barrier_tasks))); 2397 drm_printf(m, "\tLatency: %luus\n", 2398 ewma__engine_latency_read(&engine->latency)); 2399 if (intel_engine_supports_stats(engine)) 2400 drm_printf(m, "\tRuntime: %llums\n", 2401 ktime_to_ms(intel_engine_get_busy_time(engine, 2402 &dummy))); 2403 drm_printf(m, "\tForcewake: %x domains, %d active\n", 2404 engine->fw_domain, READ_ONCE(engine->fw_active)); 2405 2406 rcu_read_lock(); 2407 rq = READ_ONCE(engine->heartbeat.systole); 2408 if (rq) 2409 drm_printf(m, "\tHeartbeat: %d ms ago\n", 2410 jiffies_to_msecs(jiffies - rq->emitted_jiffies)); 2411 rcu_read_unlock(); 2412 drm_printf(m, "\tReset count: %d (global %d)\n", 2413 i915_reset_engine_count(error, engine), 2414 i915_reset_count(error)); 2415 print_properties(engine, m); 2416 2417 engine_dump_active_requests(engine, m); 2418 2419 drm_printf(m, "\tMMIO base: 0x%08x\n", engine->mmio_base); 2420 wakeref = intel_runtime_pm_get_if_in_use(engine->uncore->rpm); 2421 if (wakeref) { 2422 intel_engine_print_registers(engine, m); 2423 intel_runtime_pm_put(engine->uncore->rpm, wakeref); 2424 } else { 2425 drm_printf(m, "\tDevice is asleep; skipping register dump\n"); 2426 } 2427 2428 intel_execlists_show_requests(engine, m, i915_request_show, 8); 2429 2430 drm_printf(m, "HWSP:\n"); 2431 hexdump(m, engine->status_page.addr, PAGE_SIZE); 2432 2433 drm_printf(m, "Idle? %s\n", str_yes_no(intel_engine_is_idle(engine))); 2434 2435 intel_engine_print_breadcrumbs(engine, m); 2436 } 2437 2438 /** 2439 * intel_engine_get_busy_time() - Return current accumulated engine busyness 2440 * @engine: engine to report on 2441 * @now: monotonic timestamp of sampling 2442 * 2443 * Returns accumulated time @engine was busy since engine stats were enabled. 2444 */ 2445 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine, ktime_t *now) 2446 { 2447 return engine->busyness(engine, now); 2448 } 2449 2450 struct intel_context * 2451 intel_engine_create_virtual(struct intel_engine_cs **siblings, 2452 unsigned int count, unsigned long flags) 2453 { 2454 if (count == 0) 2455 return ERR_PTR(-EINVAL); 2456 2457 if (count == 1 && !(flags & FORCE_VIRTUAL)) 2458 return intel_context_create(siblings[0]); 2459 2460 GEM_BUG_ON(!siblings[0]->cops->create_virtual); 2461 return siblings[0]->cops->create_virtual(siblings, count, flags); 2462 } 2463 2464 static struct i915_request *engine_execlist_find_hung_request(struct intel_engine_cs *engine) 2465 { 2466 struct i915_request *request, *active = NULL; 2467 2468 /* 2469 * This search does not work in GuC submission mode. However, the GuC 2470 * will report the hanging context directly to the driver itself. So 2471 * the driver should never get here when in GuC mode. 2472 */ 2473 GEM_BUG_ON(intel_uc_uses_guc_submission(&engine->gt->uc)); 2474 2475 /* 2476 * We are called by the error capture, reset and to dump engine 2477 * state at random points in time. In particular, note that neither is 2478 * crucially ordered with an interrupt. After a hang, the GPU is dead 2479 * and we assume that no more writes can happen (we waited long enough 2480 * for all writes that were in transaction to be flushed) - adding an 2481 * extra delay for a recent interrupt is pointless. Hence, we do 2482 * not need an engine->irq_seqno_barrier() before the seqno reads. 2483 * At all other times, we must assume the GPU is still running, but 2484 * we only care about the snapshot of this moment. 2485 */ 2486 lockdep_assert_held(&engine->sched_engine->lock); 2487 2488 rcu_read_lock(); 2489 request = execlists_active(&engine->execlists); 2490 if (request) { 2491 struct intel_timeline *tl = request->context->timeline; 2492 2493 list_for_each_entry_from_reverse(request, &tl->requests, link) { 2494 if (__i915_request_is_complete(request)) 2495 break; 2496 2497 active = request; 2498 } 2499 } 2500 rcu_read_unlock(); 2501 if (active) 2502 return active; 2503 2504 list_for_each_entry(request, &engine->sched_engine->requests, 2505 sched.link) { 2506 if (i915_test_request_state(request) != I915_REQUEST_ACTIVE) 2507 continue; 2508 2509 active = request; 2510 break; 2511 } 2512 2513 return active; 2514 } 2515 2516 void intel_engine_get_hung_entity(struct intel_engine_cs *engine, 2517 struct intel_context **ce, struct i915_request **rq) 2518 { 2519 unsigned long flags; 2520 2521 *ce = intel_engine_get_hung_context(engine); 2522 if (*ce) { 2523 intel_engine_clear_hung_context(engine); 2524 2525 *rq = intel_context_get_active_request(*ce); 2526 return; 2527 } 2528 2529 /* 2530 * Getting here with GuC enabled means it is a forced error capture 2531 * with no actual hang. So, no need to attempt the execlist search. 2532 */ 2533 if (intel_uc_uses_guc_submission(&engine->gt->uc)) 2534 return; 2535 2536 spin_lock_irqsave(&engine->sched_engine->lock, flags); 2537 *rq = engine_execlist_find_hung_request(engine); 2538 if (*rq) 2539 *rq = i915_request_get_rcu(*rq); 2540 spin_unlock_irqrestore(&engine->sched_engine->lock, flags); 2541 } 2542 2543 void xehp_enable_ccs_engines(struct intel_engine_cs *engine) 2544 { 2545 /* 2546 * If there are any non-fused-off CCS engines, we need to enable CCS 2547 * support in the RCU_MODE register. This only needs to be done once, 2548 * so for simplicity we'll take care of this in the RCS engine's 2549 * resume handler; since the RCS and all CCS engines belong to the 2550 * same reset domain and are reset together, this will also take care 2551 * of re-applying the setting after i915-triggered resets. 2552 */ 2553 if (!CCS_MASK(engine->gt)) 2554 return; 2555 2556 intel_uncore_write(engine->uncore, GEN12_RCU_MODE, 2557 _MASKED_BIT_ENABLE(GEN12_RCU_MODE_CCS_ENABLE)); 2558 } 2559 2560 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 2561 #include "mock_engine.c" 2562 #include "selftest_engine.c" 2563 #include "selftest_engine_cs.c" 2564 #endif 2565