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_regs.h" 13 14 #include "i915_cmd_parser.h" 15 #include "i915_drv.h" 16 #include "i915_irq.h" 17 #include "i915_reg.h" 18 #include "intel_breadcrumbs.h" 19 #include "intel_context.h" 20 #include "intel_engine.h" 21 #include "intel_engine_pm.h" 22 #include "intel_engine_regs.h" 23 #include "intel_engine_user.h" 24 #include "intel_execlists_submission.h" 25 #include "intel_gt.h" 26 #include "intel_gt_mcr.h" 27 #include "intel_gt_pm.h" 28 #include "intel_gt_requests.h" 29 #include "intel_lrc.h" 30 #include "intel_lrc_reg.h" 31 #include "intel_reset.h" 32 #include "intel_ring.h" 33 #include "uc/intel_guc_submission.h" 34 35 /* Haswell does have the CXT_SIZE register however it does not appear to be 36 * valid. Now, docs explain in dwords what is in the context object. The full 37 * size is 70720 bytes, however, the power context and execlist context will 38 * never be saved (power context is stored elsewhere, and execlists don't work 39 * on HSW) - so the final size, including the extra state required for the 40 * Resource Streamer, is 66944 bytes, which rounds to 17 pages. 41 */ 42 #define HSW_CXT_TOTAL_SIZE (17 * PAGE_SIZE) 43 44 #define DEFAULT_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE) 45 #define GEN8_LR_CONTEXT_RENDER_SIZE (20 * PAGE_SIZE) 46 #define GEN9_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE) 47 #define GEN11_LR_CONTEXT_RENDER_SIZE (14 * PAGE_SIZE) 48 49 #define GEN8_LR_CONTEXT_OTHER_SIZE ( 2 * PAGE_SIZE) 50 51 #define MAX_MMIO_BASES 3 52 struct engine_info { 53 u8 class; 54 u8 instance; 55 /* mmio bases table *must* be sorted in reverse graphics_ver order */ 56 struct engine_mmio_base { 57 u32 graphics_ver : 8; 58 u32 base : 24; 59 } mmio_bases[MAX_MMIO_BASES]; 60 }; 61 62 static const struct engine_info intel_engines[] = { 63 [RCS0] = { 64 .class = RENDER_CLASS, 65 .instance = 0, 66 .mmio_bases = { 67 { .graphics_ver = 1, .base = RENDER_RING_BASE } 68 }, 69 }, 70 [BCS0] = { 71 .class = COPY_ENGINE_CLASS, 72 .instance = 0, 73 .mmio_bases = { 74 { .graphics_ver = 6, .base = BLT_RING_BASE } 75 }, 76 }, 77 [BCS1] = { 78 .class = COPY_ENGINE_CLASS, 79 .instance = 1, 80 .mmio_bases = { 81 { .graphics_ver = 12, .base = XEHPC_BCS1_RING_BASE } 82 }, 83 }, 84 [BCS2] = { 85 .class = COPY_ENGINE_CLASS, 86 .instance = 2, 87 .mmio_bases = { 88 { .graphics_ver = 12, .base = XEHPC_BCS2_RING_BASE } 89 }, 90 }, 91 [BCS3] = { 92 .class = COPY_ENGINE_CLASS, 93 .instance = 3, 94 .mmio_bases = { 95 { .graphics_ver = 12, .base = XEHPC_BCS3_RING_BASE } 96 }, 97 }, 98 [BCS4] = { 99 .class = COPY_ENGINE_CLASS, 100 .instance = 4, 101 .mmio_bases = { 102 { .graphics_ver = 12, .base = XEHPC_BCS4_RING_BASE } 103 }, 104 }, 105 [BCS5] = { 106 .class = COPY_ENGINE_CLASS, 107 .instance = 5, 108 .mmio_bases = { 109 { .graphics_ver = 12, .base = XEHPC_BCS5_RING_BASE } 110 }, 111 }, 112 [BCS6] = { 113 .class = COPY_ENGINE_CLASS, 114 .instance = 6, 115 .mmio_bases = { 116 { .graphics_ver = 12, .base = XEHPC_BCS6_RING_BASE } 117 }, 118 }, 119 [BCS7] = { 120 .class = COPY_ENGINE_CLASS, 121 .instance = 7, 122 .mmio_bases = { 123 { .graphics_ver = 12, .base = XEHPC_BCS7_RING_BASE } 124 }, 125 }, 126 [BCS8] = { 127 .class = COPY_ENGINE_CLASS, 128 .instance = 8, 129 .mmio_bases = { 130 { .graphics_ver = 12, .base = XEHPC_BCS8_RING_BASE } 131 }, 132 }, 133 [VCS0] = { 134 .class = VIDEO_DECODE_CLASS, 135 .instance = 0, 136 .mmio_bases = { 137 { .graphics_ver = 11, .base = GEN11_BSD_RING_BASE }, 138 { .graphics_ver = 6, .base = GEN6_BSD_RING_BASE }, 139 { .graphics_ver = 4, .base = BSD_RING_BASE } 140 }, 141 }, 142 [VCS1] = { 143 .class = VIDEO_DECODE_CLASS, 144 .instance = 1, 145 .mmio_bases = { 146 { .graphics_ver = 11, .base = GEN11_BSD2_RING_BASE }, 147 { .graphics_ver = 8, .base = GEN8_BSD2_RING_BASE } 148 }, 149 }, 150 [VCS2] = { 151 .class = VIDEO_DECODE_CLASS, 152 .instance = 2, 153 .mmio_bases = { 154 { .graphics_ver = 11, .base = GEN11_BSD3_RING_BASE } 155 }, 156 }, 157 [VCS3] = { 158 .class = VIDEO_DECODE_CLASS, 159 .instance = 3, 160 .mmio_bases = { 161 { .graphics_ver = 11, .base = GEN11_BSD4_RING_BASE } 162 }, 163 }, 164 [VCS4] = { 165 .class = VIDEO_DECODE_CLASS, 166 .instance = 4, 167 .mmio_bases = { 168 { .graphics_ver = 12, .base = XEHP_BSD5_RING_BASE } 169 }, 170 }, 171 [VCS5] = { 172 .class = VIDEO_DECODE_CLASS, 173 .instance = 5, 174 .mmio_bases = { 175 { .graphics_ver = 12, .base = XEHP_BSD6_RING_BASE } 176 }, 177 }, 178 [VCS6] = { 179 .class = VIDEO_DECODE_CLASS, 180 .instance = 6, 181 .mmio_bases = { 182 { .graphics_ver = 12, .base = XEHP_BSD7_RING_BASE } 183 }, 184 }, 185 [VCS7] = { 186 .class = VIDEO_DECODE_CLASS, 187 .instance = 7, 188 .mmio_bases = { 189 { .graphics_ver = 12, .base = XEHP_BSD8_RING_BASE } 190 }, 191 }, 192 [VECS0] = { 193 .class = VIDEO_ENHANCEMENT_CLASS, 194 .instance = 0, 195 .mmio_bases = { 196 { .graphics_ver = 11, .base = GEN11_VEBOX_RING_BASE }, 197 { .graphics_ver = 7, .base = VEBOX_RING_BASE } 198 }, 199 }, 200 [VECS1] = { 201 .class = VIDEO_ENHANCEMENT_CLASS, 202 .instance = 1, 203 .mmio_bases = { 204 { .graphics_ver = 11, .base = GEN11_VEBOX2_RING_BASE } 205 }, 206 }, 207 [VECS2] = { 208 .class = VIDEO_ENHANCEMENT_CLASS, 209 .instance = 2, 210 .mmio_bases = { 211 { .graphics_ver = 12, .base = XEHP_VEBOX3_RING_BASE } 212 }, 213 }, 214 [VECS3] = { 215 .class = VIDEO_ENHANCEMENT_CLASS, 216 .instance = 3, 217 .mmio_bases = { 218 { .graphics_ver = 12, .base = XEHP_VEBOX4_RING_BASE } 219 }, 220 }, 221 [CCS0] = { 222 .class = COMPUTE_CLASS, 223 .instance = 0, 224 .mmio_bases = { 225 { .graphics_ver = 12, .base = GEN12_COMPUTE0_RING_BASE } 226 } 227 }, 228 [CCS1] = { 229 .class = COMPUTE_CLASS, 230 .instance = 1, 231 .mmio_bases = { 232 { .graphics_ver = 12, .base = GEN12_COMPUTE1_RING_BASE } 233 } 234 }, 235 [CCS2] = { 236 .class = COMPUTE_CLASS, 237 .instance = 2, 238 .mmio_bases = { 239 { .graphics_ver = 12, .base = GEN12_COMPUTE2_RING_BASE } 240 } 241 }, 242 [CCS3] = { 243 .class = COMPUTE_CLASS, 244 .instance = 3, 245 .mmio_bases = { 246 { .graphics_ver = 12, .base = GEN12_COMPUTE3_RING_BASE } 247 } 248 }, 249 [GSC0] = { 250 .class = OTHER_CLASS, 251 .instance = OTHER_GSC_INSTANCE, 252 .mmio_bases = { 253 { .graphics_ver = 12, .base = MTL_GSC_RING_BASE } 254 } 255 }, 256 }; 257 258 /** 259 * intel_engine_context_size() - return the size of the context for an engine 260 * @gt: the gt 261 * @class: engine class 262 * 263 * Each engine class may require a different amount of space for a context 264 * image. 265 * 266 * Return: size (in bytes) of an engine class specific context image 267 * 268 * Note: this size includes the HWSP, which is part of the context image 269 * in LRC mode, but does not include the "shared data page" used with 270 * GuC submission. The caller should account for this if using the GuC. 271 */ 272 u32 intel_engine_context_size(struct intel_gt *gt, u8 class) 273 { 274 struct intel_uncore *uncore = gt->uncore; 275 u32 cxt_size; 276 277 BUILD_BUG_ON(I915_GTT_PAGE_SIZE != PAGE_SIZE); 278 279 switch (class) { 280 case COMPUTE_CLASS: 281 fallthrough; 282 case RENDER_CLASS: 283 switch (GRAPHICS_VER(gt->i915)) { 284 default: 285 MISSING_CASE(GRAPHICS_VER(gt->i915)); 286 return DEFAULT_LR_CONTEXT_RENDER_SIZE; 287 case 12: 288 case 11: 289 return GEN11_LR_CONTEXT_RENDER_SIZE; 290 case 9: 291 return GEN9_LR_CONTEXT_RENDER_SIZE; 292 case 8: 293 return GEN8_LR_CONTEXT_RENDER_SIZE; 294 case 7: 295 if (IS_HASWELL(gt->i915)) 296 return HSW_CXT_TOTAL_SIZE; 297 298 cxt_size = intel_uncore_read(uncore, GEN7_CXT_SIZE); 299 return round_up(GEN7_CXT_TOTAL_SIZE(cxt_size) * 64, 300 PAGE_SIZE); 301 case 6: 302 cxt_size = intel_uncore_read(uncore, CXT_SIZE); 303 return round_up(GEN6_CXT_TOTAL_SIZE(cxt_size) * 64, 304 PAGE_SIZE); 305 case 5: 306 case 4: 307 /* 308 * There is a discrepancy here between the size reported 309 * by the register and the size of the context layout 310 * in the docs. Both are described as authorative! 311 * 312 * The discrepancy is on the order of a few cachelines, 313 * but the total is under one page (4k), which is our 314 * minimum allocation anyway so it should all come 315 * out in the wash. 316 */ 317 cxt_size = intel_uncore_read(uncore, CXT_SIZE) + 1; 318 drm_dbg(>->i915->drm, 319 "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 && !RCS_MASK(engine->gt) && 501 __ffs(CCS_MASK(engine->gt)) == engine->instance) || 502 engine->class == RENDER_CLASS) 503 engine->flags |= I915_ENGINE_FIRST_RENDER_COMPUTE; 504 505 /* features common between engines sharing EUs */ 506 if (engine->class == RENDER_CLASS || engine->class == COMPUTE_CLASS) { 507 engine->flags |= I915_ENGINE_HAS_RCS_REG_STATE; 508 engine->flags |= I915_ENGINE_HAS_EU_PRIORITY; 509 } 510 511 engine->props.heartbeat_interval_ms = 512 CONFIG_DRM_I915_HEARTBEAT_INTERVAL; 513 engine->props.max_busywait_duration_ns = 514 CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT; 515 engine->props.preempt_timeout_ms = 516 CONFIG_DRM_I915_PREEMPT_TIMEOUT; 517 engine->props.stop_timeout_ms = 518 CONFIG_DRM_I915_STOP_TIMEOUT; 519 engine->props.timeslice_duration_ms = 520 CONFIG_DRM_I915_TIMESLICE_DURATION; 521 522 /* 523 * Mid-thread pre-emption is not available in Gen12. Unfortunately, 524 * some compute workloads run quite long threads. That means they get 525 * reset due to not pre-empting in a timely manner. So, bump the 526 * pre-emption timeout value to be much higher for compute engines. 527 */ 528 if (GRAPHICS_VER(i915) == 12 && (engine->flags & I915_ENGINE_HAS_RCS_REG_STATE)) 529 engine->props.preempt_timeout_ms = CONFIG_DRM_I915_PREEMPT_TIMEOUT_COMPUTE; 530 531 /* Cap properties according to any system limits */ 532 #define CLAMP_PROP(field) \ 533 do { \ 534 u64 clamp = intel_clamp_##field(engine, engine->props.field); \ 535 if (clamp != engine->props.field) { \ 536 drm_notice(&engine->i915->drm, \ 537 "Warning, clamping %s to %lld to prevent overflow\n", \ 538 #field, clamp); \ 539 engine->props.field = clamp; \ 540 } \ 541 } while (0) 542 543 CLAMP_PROP(heartbeat_interval_ms); 544 CLAMP_PROP(max_busywait_duration_ns); 545 CLAMP_PROP(preempt_timeout_ms); 546 CLAMP_PROP(stop_timeout_ms); 547 CLAMP_PROP(timeslice_duration_ms); 548 549 #undef CLAMP_PROP 550 551 engine->defaults = engine->props; /* never to change again */ 552 553 engine->context_size = intel_engine_context_size(gt, engine->class); 554 if (WARN_ON(engine->context_size > BIT(20))) 555 engine->context_size = 0; 556 if (engine->context_size) 557 DRIVER_CAPS(i915)->has_logical_contexts = true; 558 559 ewma__engine_latency_init(&engine->latency); 560 seqcount_init(&engine->stats.execlists.lock); 561 562 ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier); 563 564 /* Scrub mmio state on takeover */ 565 intel_engine_sanitize_mmio(engine); 566 567 gt->engine_class[info->class][info->instance] = engine; 568 gt->engine[id] = engine; 569 570 return 0; 571 } 572 573 u64 intel_clamp_heartbeat_interval_ms(struct intel_engine_cs *engine, u64 value) 574 { 575 value = min_t(u64, value, jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT)); 576 577 return value; 578 } 579 580 u64 intel_clamp_max_busywait_duration_ns(struct intel_engine_cs *engine, u64 value) 581 { 582 value = min(value, jiffies_to_nsecs(2)); 583 584 return value; 585 } 586 587 u64 intel_clamp_preempt_timeout_ms(struct intel_engine_cs *engine, u64 value) 588 { 589 /* 590 * NB: The GuC API only supports 32bit values. However, the limit is further 591 * reduced due to internal calculations which would otherwise overflow. 592 */ 593 if (intel_guc_submission_is_wanted(&engine->gt->uc.guc)) 594 value = min_t(u64, value, guc_policy_max_preempt_timeout_ms()); 595 596 value = min_t(u64, value, jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT)); 597 598 return value; 599 } 600 601 u64 intel_clamp_stop_timeout_ms(struct intel_engine_cs *engine, u64 value) 602 { 603 value = min_t(u64, value, jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT)); 604 605 return value; 606 } 607 608 u64 intel_clamp_timeslice_duration_ms(struct intel_engine_cs *engine, u64 value) 609 { 610 /* 611 * NB: The GuC API only supports 32bit values. However, the limit is further 612 * reduced due to internal calculations which would otherwise overflow. 613 */ 614 if (intel_guc_submission_is_wanted(&engine->gt->uc.guc)) 615 value = min_t(u64, value, guc_policy_max_exec_quantum_ms()); 616 617 value = min_t(u64, value, jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT)); 618 619 return value; 620 } 621 622 static void __setup_engine_capabilities(struct intel_engine_cs *engine) 623 { 624 struct drm_i915_private *i915 = engine->i915; 625 626 if (engine->class == VIDEO_DECODE_CLASS) { 627 /* 628 * HEVC support is present on first engine instance 629 * before Gen11 and on all instances afterwards. 630 */ 631 if (GRAPHICS_VER(i915) >= 11 || 632 (GRAPHICS_VER(i915) >= 9 && engine->instance == 0)) 633 engine->uabi_capabilities |= 634 I915_VIDEO_CLASS_CAPABILITY_HEVC; 635 636 /* 637 * SFC block is present only on even logical engine 638 * instances. 639 */ 640 if ((GRAPHICS_VER(i915) >= 11 && 641 (engine->gt->info.vdbox_sfc_access & 642 BIT(engine->instance))) || 643 (GRAPHICS_VER(i915) >= 9 && engine->instance == 0)) 644 engine->uabi_capabilities |= 645 I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC; 646 } else if (engine->class == VIDEO_ENHANCEMENT_CLASS) { 647 if (GRAPHICS_VER(i915) >= 9 && 648 engine->gt->info.sfc_mask & BIT(engine->instance)) 649 engine->uabi_capabilities |= 650 I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC; 651 } 652 } 653 654 static void intel_setup_engine_capabilities(struct intel_gt *gt) 655 { 656 struct intel_engine_cs *engine; 657 enum intel_engine_id id; 658 659 for_each_engine(engine, gt, id) 660 __setup_engine_capabilities(engine); 661 } 662 663 /** 664 * intel_engines_release() - free the resources allocated for Command Streamers 665 * @gt: pointer to struct intel_gt 666 */ 667 void intel_engines_release(struct intel_gt *gt) 668 { 669 struct intel_engine_cs *engine; 670 enum intel_engine_id id; 671 672 /* 673 * Before we release the resources held by engine, we must be certain 674 * that the HW is no longer accessing them -- having the GPU scribble 675 * to or read from a page being used for something else causes no end 676 * of fun. 677 * 678 * The GPU should be reset by this point, but assume the worst just 679 * in case we aborted before completely initialising the engines. 680 */ 681 GEM_BUG_ON(intel_gt_pm_is_awake(gt)); 682 if (!INTEL_INFO(gt->i915)->gpu_reset_clobbers_display) 683 __intel_gt_reset(gt, ALL_ENGINES); 684 685 /* Decouple the backend; but keep the layout for late GPU resets */ 686 for_each_engine(engine, gt, id) { 687 if (!engine->release) 688 continue; 689 690 intel_wakeref_wait_for_idle(&engine->wakeref); 691 GEM_BUG_ON(intel_engine_pm_is_awake(engine)); 692 693 engine->release(engine); 694 engine->release = NULL; 695 696 memset(&engine->reset, 0, sizeof(engine->reset)); 697 } 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, 50)) 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, 50)) { 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 drm_dbg(&i915->drm, "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 drm_dbg(&i915->drm, "vdbox enable: %04x, instances: %04lx\n", 800 vdbox_mask, VDBOX_MASK(gt)); 801 GEM_BUG_ON(vdbox_mask != VDBOX_MASK(gt)); 802 803 for (i = 0; i < I915_MAX_VECS; i++) { 804 if (!HAS_ENGINE(gt, _VECS(i))) { 805 vebox_mask &= ~BIT(i); 806 continue; 807 } 808 809 if (!(BIT(i) & vebox_mask)) { 810 gt->info.engine_mask &= ~BIT(_VECS(i)); 811 drm_dbg(&i915->drm, "vecs%u fused off\n", i); 812 } 813 } 814 drm_dbg(&i915->drm, "vebox enable: %04x, instances: %04lx\n", 815 vebox_mask, VEBOX_MASK(gt)); 816 GEM_BUG_ON(vebox_mask != VEBOX_MASK(gt)); 817 } 818 819 static void engine_mask_apply_compute_fuses(struct intel_gt *gt) 820 { 821 struct drm_i915_private *i915 = gt->i915; 822 struct intel_gt_info *info = >->info; 823 int ss_per_ccs = info->sseu.max_subslices / I915_MAX_CCS; 824 unsigned long ccs_mask; 825 unsigned int i; 826 827 if (GRAPHICS_VER(i915) < 11) 828 return; 829 830 if (hweight32(CCS_MASK(gt)) <= 1) 831 return; 832 833 ccs_mask = intel_slicemask_from_xehp_dssmask(info->sseu.compute_subslice_mask, 834 ss_per_ccs); 835 /* 836 * If all DSS in a quadrant are fused off, the corresponding CCS 837 * engine is not available for use. 838 */ 839 for_each_clear_bit(i, &ccs_mask, I915_MAX_CCS) { 840 info->engine_mask &= ~BIT(_CCS(i)); 841 drm_dbg(&i915->drm, "ccs%u fused off\n", i); 842 } 843 } 844 845 static void engine_mask_apply_copy_fuses(struct intel_gt *gt) 846 { 847 struct drm_i915_private *i915 = gt->i915; 848 struct intel_gt_info *info = >->info; 849 unsigned long meml3_mask; 850 unsigned long quad; 851 852 if (!(GRAPHICS_VER_FULL(i915) >= IP_VER(12, 60) && 853 GRAPHICS_VER_FULL(i915) < IP_VER(12, 70))) 854 return; 855 856 meml3_mask = intel_uncore_read(gt->uncore, GEN10_MIRROR_FUSE3); 857 meml3_mask = REG_FIELD_GET(GEN12_MEML3_EN_MASK, meml3_mask); 858 859 /* 860 * Link Copy engines may be fused off according to meml3_mask. Each 861 * bit is a quad that houses 2 Link Copy and two Sub Copy engines. 862 */ 863 for_each_clear_bit(quad, &meml3_mask, GEN12_MAX_MSLICES) { 864 unsigned int instance = quad * 2 + 1; 865 intel_engine_mask_t mask = GENMASK(_BCS(instance + 1), 866 _BCS(instance)); 867 868 if (mask & info->engine_mask) { 869 drm_dbg(&i915->drm, "bcs%u fused off\n", instance); 870 drm_dbg(&i915->drm, "bcs%u fused off\n", instance + 1); 871 872 info->engine_mask &= ~mask; 873 } 874 } 875 } 876 877 /* 878 * Determine which engines are fused off in our particular hardware. 879 * Note that we have a catch-22 situation where we need to be able to access 880 * the blitter forcewake domain to read the engine fuses, but at the same time 881 * we need to know which engines are available on the system to know which 882 * forcewake domains are present. We solve this by intializing the forcewake 883 * domains based on the full engine mask in the platform capabilities before 884 * calling this function and pruning the domains for fused-off engines 885 * afterwards. 886 */ 887 static intel_engine_mask_t init_engine_mask(struct intel_gt *gt) 888 { 889 struct intel_gt_info *info = >->info; 890 891 GEM_BUG_ON(!info->engine_mask); 892 893 engine_mask_apply_media_fuses(gt); 894 engine_mask_apply_compute_fuses(gt); 895 engine_mask_apply_copy_fuses(gt); 896 897 /* 898 * The only use of the GSC CS is to load and communicate with the GSC 899 * FW, so we have no use for it if we don't have the FW. 900 * 901 * IMPORTANT: in cases where we don't have the GSC FW, we have a 902 * catch-22 situation that breaks media C6 due to 2 requirements: 903 * 1) once turned on, the GSC power well will not go to sleep unless the 904 * GSC FW is loaded. 905 * 2) to enable idling (which is required for media C6) we need to 906 * initialize the IDLE_MSG register for the GSC CS and do at least 1 907 * submission, which will wake up the GSC power well. 908 */ 909 if (__HAS_ENGINE(info->engine_mask, GSC0) && !intel_uc_wants_gsc_uc(>->uc)) { 910 drm_notice(>->i915->drm, 911 "No GSC FW selected, disabling GSC CS and media C6\n"); 912 info->engine_mask &= ~BIT(GSC0); 913 } 914 915 return info->engine_mask; 916 } 917 918 static void populate_logical_ids(struct intel_gt *gt, u8 *logical_ids, 919 u8 class, const u8 *map, u8 num_instances) 920 { 921 int i, j; 922 u8 current_logical_id = 0; 923 924 for (j = 0; j < num_instances; ++j) { 925 for (i = 0; i < ARRAY_SIZE(intel_engines); ++i) { 926 if (!HAS_ENGINE(gt, i) || 927 intel_engines[i].class != class) 928 continue; 929 930 if (intel_engines[i].instance == map[j]) { 931 logical_ids[intel_engines[i].instance] = 932 current_logical_id++; 933 break; 934 } 935 } 936 } 937 } 938 939 static void setup_logical_ids(struct intel_gt *gt, u8 *logical_ids, u8 class) 940 { 941 /* 942 * Logical to physical mapping is needed for proper support 943 * to split-frame feature. 944 */ 945 if (MEDIA_VER(gt->i915) >= 11 && class == VIDEO_DECODE_CLASS) { 946 const u8 map[] = { 0, 2, 4, 6, 1, 3, 5, 7 }; 947 948 populate_logical_ids(gt, logical_ids, class, 949 map, ARRAY_SIZE(map)); 950 } else { 951 int i; 952 u8 map[MAX_ENGINE_INSTANCE + 1]; 953 954 for (i = 0; i < MAX_ENGINE_INSTANCE + 1; ++i) 955 map[i] = i; 956 populate_logical_ids(gt, logical_ids, class, 957 map, ARRAY_SIZE(map)); 958 } 959 } 960 961 /** 962 * intel_engines_init_mmio() - allocate and prepare the Engine Command Streamers 963 * @gt: pointer to struct intel_gt 964 * 965 * Return: non-zero if the initialization failed. 966 */ 967 int intel_engines_init_mmio(struct intel_gt *gt) 968 { 969 struct drm_i915_private *i915 = gt->i915; 970 const unsigned int engine_mask = init_engine_mask(gt); 971 unsigned int mask = 0; 972 unsigned int i, class; 973 u8 logical_ids[MAX_ENGINE_INSTANCE + 1]; 974 int err; 975 976 drm_WARN_ON(&i915->drm, engine_mask == 0); 977 drm_WARN_ON(&i915->drm, engine_mask & 978 GENMASK(BITS_PER_TYPE(mask) - 1, I915_NUM_ENGINES)); 979 980 if (i915_inject_probe_failure(i915)) 981 return -ENODEV; 982 983 for (class = 0; class < MAX_ENGINE_CLASS + 1; ++class) { 984 setup_logical_ids(gt, logical_ids, class); 985 986 for (i = 0; i < ARRAY_SIZE(intel_engines); ++i) { 987 u8 instance = intel_engines[i].instance; 988 989 if (intel_engines[i].class != class || 990 !HAS_ENGINE(gt, i)) 991 continue; 992 993 err = intel_engine_setup(gt, i, 994 logical_ids[instance]); 995 if (err) 996 goto cleanup; 997 998 mask |= BIT(i); 999 } 1000 } 1001 1002 /* 1003 * Catch failures to update intel_engines table when the new engines 1004 * are added to the driver by a warning and disabling the forgotten 1005 * engines. 1006 */ 1007 if (drm_WARN_ON(&i915->drm, mask != engine_mask)) 1008 gt->info.engine_mask = mask; 1009 1010 gt->info.num_engines = hweight32(mask); 1011 1012 intel_gt_check_and_clear_faults(gt); 1013 1014 intel_setup_engine_capabilities(gt); 1015 1016 intel_uncore_prune_engine_fw_domains(gt->uncore, gt); 1017 1018 return 0; 1019 1020 cleanup: 1021 intel_engines_free(gt); 1022 return err; 1023 } 1024 1025 void intel_engine_init_execlists(struct intel_engine_cs *engine) 1026 { 1027 struct intel_engine_execlists * const execlists = &engine->execlists; 1028 1029 execlists->port_mask = 1; 1030 GEM_BUG_ON(!is_power_of_2(execlists_num_ports(execlists))); 1031 GEM_BUG_ON(execlists_num_ports(execlists) > EXECLIST_MAX_PORTS); 1032 1033 memset(execlists->pending, 0, sizeof(execlists->pending)); 1034 execlists->active = 1035 memset(execlists->inflight, 0, sizeof(execlists->inflight)); 1036 } 1037 1038 static void cleanup_status_page(struct intel_engine_cs *engine) 1039 { 1040 struct i915_vma *vma; 1041 1042 /* Prevent writes into HWSP after returning the page to the system */ 1043 intel_engine_set_hwsp_writemask(engine, ~0u); 1044 1045 vma = fetch_and_zero(&engine->status_page.vma); 1046 if (!vma) 1047 return; 1048 1049 if (!HWS_NEEDS_PHYSICAL(engine->i915)) 1050 i915_vma_unpin(vma); 1051 1052 i915_gem_object_unpin_map(vma->obj); 1053 i915_gem_object_put(vma->obj); 1054 } 1055 1056 static int pin_ggtt_status_page(struct intel_engine_cs *engine, 1057 struct i915_gem_ww_ctx *ww, 1058 struct i915_vma *vma) 1059 { 1060 unsigned int flags; 1061 1062 if (!HAS_LLC(engine->i915) && i915_ggtt_has_aperture(engine->gt->ggtt)) 1063 /* 1064 * On g33, we cannot place HWS above 256MiB, so 1065 * restrict its pinning to the low mappable arena. 1066 * Though this restriction is not documented for 1067 * gen4, gen5, or byt, they also behave similarly 1068 * and hang if the HWS is placed at the top of the 1069 * GTT. To generalise, it appears that all !llc 1070 * platforms have issues with us placing the HWS 1071 * above the mappable region (even though we never 1072 * actually map it). 1073 */ 1074 flags = PIN_MAPPABLE; 1075 else 1076 flags = PIN_HIGH; 1077 1078 return i915_ggtt_pin(vma, ww, 0, flags); 1079 } 1080 1081 static int init_status_page(struct intel_engine_cs *engine) 1082 { 1083 struct drm_i915_gem_object *obj; 1084 struct i915_gem_ww_ctx ww; 1085 struct i915_vma *vma; 1086 void *vaddr; 1087 int ret; 1088 1089 INIT_LIST_HEAD(&engine->status_page.timelines); 1090 1091 /* 1092 * Though the HWS register does support 36bit addresses, historically 1093 * we have had hangs and corruption reported due to wild writes if 1094 * the HWS is placed above 4G. We only allow objects to be allocated 1095 * in GFP_DMA32 for i965, and no earlier physical address users had 1096 * access to more than 4G. 1097 */ 1098 obj = i915_gem_object_create_internal(engine->i915, PAGE_SIZE); 1099 if (IS_ERR(obj)) { 1100 drm_err(&engine->i915->drm, 1101 "Failed to allocate status page\n"); 1102 return PTR_ERR(obj); 1103 } 1104 1105 i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC); 1106 1107 vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL); 1108 if (IS_ERR(vma)) { 1109 ret = PTR_ERR(vma); 1110 goto err_put; 1111 } 1112 1113 i915_gem_ww_ctx_init(&ww, true); 1114 retry: 1115 ret = i915_gem_object_lock(obj, &ww); 1116 if (!ret && !HWS_NEEDS_PHYSICAL(engine->i915)) 1117 ret = pin_ggtt_status_page(engine, &ww, vma); 1118 if (ret) 1119 goto err; 1120 1121 vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB); 1122 if (IS_ERR(vaddr)) { 1123 ret = PTR_ERR(vaddr); 1124 goto err_unpin; 1125 } 1126 1127 engine->status_page.addr = memset(vaddr, 0, PAGE_SIZE); 1128 engine->status_page.vma = vma; 1129 1130 err_unpin: 1131 if (ret) 1132 i915_vma_unpin(vma); 1133 err: 1134 if (ret == -EDEADLK) { 1135 ret = i915_gem_ww_ctx_backoff(&ww); 1136 if (!ret) 1137 goto retry; 1138 } 1139 i915_gem_ww_ctx_fini(&ww); 1140 err_put: 1141 if (ret) 1142 i915_gem_object_put(obj); 1143 return ret; 1144 } 1145 1146 static int engine_setup_common(struct intel_engine_cs *engine) 1147 { 1148 int err; 1149 1150 init_llist_head(&engine->barrier_tasks); 1151 1152 err = init_status_page(engine); 1153 if (err) 1154 return err; 1155 1156 engine->breadcrumbs = intel_breadcrumbs_create(engine); 1157 if (!engine->breadcrumbs) { 1158 err = -ENOMEM; 1159 goto err_status; 1160 } 1161 1162 engine->sched_engine = i915_sched_engine_create(ENGINE_PHYSICAL); 1163 if (!engine->sched_engine) { 1164 err = -ENOMEM; 1165 goto err_sched_engine; 1166 } 1167 engine->sched_engine->private_data = engine; 1168 1169 err = intel_engine_init_cmd_parser(engine); 1170 if (err) 1171 goto err_cmd_parser; 1172 1173 intel_engine_init_execlists(engine); 1174 intel_engine_init__pm(engine); 1175 intel_engine_init_retire(engine); 1176 1177 /* Use the whole device by default */ 1178 engine->sseu = 1179 intel_sseu_from_device_info(&engine->gt->info.sseu); 1180 1181 intel_engine_init_workarounds(engine); 1182 intel_engine_init_whitelist(engine); 1183 intel_engine_init_ctx_wa(engine); 1184 1185 if (GRAPHICS_VER(engine->i915) >= 12) 1186 engine->flags |= I915_ENGINE_HAS_RELATIVE_MMIO; 1187 1188 return 0; 1189 1190 err_cmd_parser: 1191 i915_sched_engine_put(engine->sched_engine); 1192 err_sched_engine: 1193 intel_breadcrumbs_put(engine->breadcrumbs); 1194 err_status: 1195 cleanup_status_page(engine); 1196 return err; 1197 } 1198 1199 struct measure_breadcrumb { 1200 struct i915_request rq; 1201 struct intel_ring ring; 1202 u32 cs[2048]; 1203 }; 1204 1205 static int measure_breadcrumb_dw(struct intel_context *ce) 1206 { 1207 struct intel_engine_cs *engine = ce->engine; 1208 struct measure_breadcrumb *frame; 1209 int dw; 1210 1211 GEM_BUG_ON(!engine->gt->scratch); 1212 1213 frame = kzalloc(sizeof(*frame), GFP_KERNEL); 1214 if (!frame) 1215 return -ENOMEM; 1216 1217 frame->rq.engine = engine; 1218 frame->rq.context = ce; 1219 rcu_assign_pointer(frame->rq.timeline, ce->timeline); 1220 frame->rq.hwsp_seqno = ce->timeline->hwsp_seqno; 1221 1222 frame->ring.vaddr = frame->cs; 1223 frame->ring.size = sizeof(frame->cs); 1224 frame->ring.wrap = 1225 BITS_PER_TYPE(frame->ring.size) - ilog2(frame->ring.size); 1226 frame->ring.effective_size = frame->ring.size; 1227 intel_ring_update_space(&frame->ring); 1228 frame->rq.ring = &frame->ring; 1229 1230 mutex_lock(&ce->timeline->mutex); 1231 spin_lock_irq(&engine->sched_engine->lock); 1232 1233 dw = engine->emit_fini_breadcrumb(&frame->rq, frame->cs) - frame->cs; 1234 1235 spin_unlock_irq(&engine->sched_engine->lock); 1236 mutex_unlock(&ce->timeline->mutex); 1237 1238 GEM_BUG_ON(dw & 1); /* RING_TAIL must be qword aligned */ 1239 1240 kfree(frame); 1241 return dw; 1242 } 1243 1244 struct intel_context * 1245 intel_engine_create_pinned_context(struct intel_engine_cs *engine, 1246 struct i915_address_space *vm, 1247 unsigned int ring_size, 1248 unsigned int hwsp, 1249 struct lock_class_key *key, 1250 const char *name) 1251 { 1252 struct intel_context *ce; 1253 int err; 1254 1255 ce = intel_context_create(engine); 1256 if (IS_ERR(ce)) 1257 return ce; 1258 1259 __set_bit(CONTEXT_BARRIER_BIT, &ce->flags); 1260 ce->timeline = page_pack_bits(NULL, hwsp); 1261 ce->ring = NULL; 1262 ce->ring_size = ring_size; 1263 1264 i915_vm_put(ce->vm); 1265 ce->vm = i915_vm_get(vm); 1266 1267 err = intel_context_pin(ce); /* perma-pin so it is always available */ 1268 if (err) { 1269 intel_context_put(ce); 1270 return ERR_PTR(err); 1271 } 1272 1273 list_add_tail(&ce->pinned_contexts_link, &engine->pinned_contexts_list); 1274 1275 /* 1276 * Give our perma-pinned kernel timelines a separate lockdep class, 1277 * so that we can use them from within the normal user timelines 1278 * should we need to inject GPU operations during their request 1279 * construction. 1280 */ 1281 lockdep_set_class_and_name(&ce->timeline->mutex, key, name); 1282 1283 return ce; 1284 } 1285 1286 void intel_engine_destroy_pinned_context(struct intel_context *ce) 1287 { 1288 struct intel_engine_cs *engine = ce->engine; 1289 struct i915_vma *hwsp = engine->status_page.vma; 1290 1291 GEM_BUG_ON(ce->timeline->hwsp_ggtt != hwsp); 1292 1293 mutex_lock(&hwsp->vm->mutex); 1294 list_del(&ce->timeline->engine_link); 1295 mutex_unlock(&hwsp->vm->mutex); 1296 1297 list_del(&ce->pinned_contexts_link); 1298 intel_context_unpin(ce); 1299 intel_context_put(ce); 1300 } 1301 1302 static struct intel_context * 1303 create_kernel_context(struct intel_engine_cs *engine) 1304 { 1305 static struct lock_class_key kernel; 1306 1307 return intel_engine_create_pinned_context(engine, engine->gt->vm, SZ_4K, 1308 I915_GEM_HWS_SEQNO_ADDR, 1309 &kernel, "kernel_context"); 1310 } 1311 1312 /** 1313 * intel_engines_init_common - initialize cengine state which might require hw access 1314 * @engine: Engine to initialize. 1315 * 1316 * Initializes @engine@ structure members shared between legacy and execlists 1317 * submission modes which do require hardware access. 1318 * 1319 * Typcally done at later stages of submission mode specific engine setup. 1320 * 1321 * Returns zero on success or an error code on failure. 1322 */ 1323 static int engine_init_common(struct intel_engine_cs *engine) 1324 { 1325 struct intel_context *ce; 1326 int ret; 1327 1328 engine->set_default_submission(engine); 1329 1330 /* 1331 * We may need to do things with the shrinker which 1332 * require us to immediately switch back to the default 1333 * context. This can cause a problem as pinning the 1334 * default context also requires GTT space which may not 1335 * be available. To avoid this we always pin the default 1336 * context. 1337 */ 1338 ce = create_kernel_context(engine); 1339 if (IS_ERR(ce)) 1340 return PTR_ERR(ce); 1341 1342 ret = measure_breadcrumb_dw(ce); 1343 if (ret < 0) 1344 goto err_context; 1345 1346 engine->emit_fini_breadcrumb_dw = ret; 1347 engine->kernel_context = ce; 1348 1349 return 0; 1350 1351 err_context: 1352 intel_engine_destroy_pinned_context(ce); 1353 return ret; 1354 } 1355 1356 int intel_engines_init(struct intel_gt *gt) 1357 { 1358 int (*setup)(struct intel_engine_cs *engine); 1359 struct intel_engine_cs *engine; 1360 enum intel_engine_id id; 1361 int err; 1362 1363 if (intel_uc_uses_guc_submission(>->uc)) { 1364 gt->submission_method = INTEL_SUBMISSION_GUC; 1365 setup = intel_guc_submission_setup; 1366 } else if (HAS_EXECLISTS(gt->i915)) { 1367 gt->submission_method = INTEL_SUBMISSION_ELSP; 1368 setup = intel_execlists_submission_setup; 1369 } else { 1370 gt->submission_method = INTEL_SUBMISSION_RING; 1371 setup = intel_ring_submission_setup; 1372 } 1373 1374 for_each_engine(engine, gt, id) { 1375 err = engine_setup_common(engine); 1376 if (err) 1377 return err; 1378 1379 err = setup(engine); 1380 if (err) { 1381 intel_engine_cleanup_common(engine); 1382 return err; 1383 } 1384 1385 /* The backend should now be responsible for cleanup */ 1386 GEM_BUG_ON(engine->release == NULL); 1387 1388 err = engine_init_common(engine); 1389 if (err) 1390 return err; 1391 1392 intel_engine_add_user(engine); 1393 } 1394 1395 return 0; 1396 } 1397 1398 /** 1399 * intel_engines_cleanup_common - cleans up the engine state created by 1400 * the common initiailizers. 1401 * @engine: Engine to cleanup. 1402 * 1403 * This cleans up everything created by the common helpers. 1404 */ 1405 void intel_engine_cleanup_common(struct intel_engine_cs *engine) 1406 { 1407 GEM_BUG_ON(!list_empty(&engine->sched_engine->requests)); 1408 1409 i915_sched_engine_put(engine->sched_engine); 1410 intel_breadcrumbs_put(engine->breadcrumbs); 1411 1412 intel_engine_fini_retire(engine); 1413 intel_engine_cleanup_cmd_parser(engine); 1414 1415 if (engine->default_state) 1416 fput(engine->default_state); 1417 1418 if (engine->kernel_context) 1419 intel_engine_destroy_pinned_context(engine->kernel_context); 1420 1421 GEM_BUG_ON(!llist_empty(&engine->barrier_tasks)); 1422 cleanup_status_page(engine); 1423 1424 intel_wa_list_free(&engine->ctx_wa_list); 1425 intel_wa_list_free(&engine->wa_list); 1426 intel_wa_list_free(&engine->whitelist); 1427 } 1428 1429 /** 1430 * intel_engine_resume - re-initializes the HW state of the engine 1431 * @engine: Engine to resume. 1432 * 1433 * Returns zero on success or an error code on failure. 1434 */ 1435 int intel_engine_resume(struct intel_engine_cs *engine) 1436 { 1437 intel_engine_apply_workarounds(engine); 1438 intel_engine_apply_whitelist(engine); 1439 1440 return engine->resume(engine); 1441 } 1442 1443 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine) 1444 { 1445 struct drm_i915_private *i915 = engine->i915; 1446 1447 u64 acthd; 1448 1449 if (GRAPHICS_VER(i915) >= 8) 1450 acthd = ENGINE_READ64(engine, RING_ACTHD, RING_ACTHD_UDW); 1451 else if (GRAPHICS_VER(i915) >= 4) 1452 acthd = ENGINE_READ(engine, RING_ACTHD); 1453 else 1454 acthd = ENGINE_READ(engine, ACTHD); 1455 1456 return acthd; 1457 } 1458 1459 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine) 1460 { 1461 u64 bbaddr; 1462 1463 if (GRAPHICS_VER(engine->i915) >= 8) 1464 bbaddr = ENGINE_READ64(engine, RING_BBADDR, RING_BBADDR_UDW); 1465 else 1466 bbaddr = ENGINE_READ(engine, RING_BBADDR); 1467 1468 return bbaddr; 1469 } 1470 1471 static unsigned long stop_timeout(const struct intel_engine_cs *engine) 1472 { 1473 if (in_atomic() || irqs_disabled()) /* inside atomic preempt-reset? */ 1474 return 0; 1475 1476 /* 1477 * If we are doing a normal GPU reset, we can take our time and allow 1478 * the engine to quiesce. We've stopped submission to the engine, and 1479 * if we wait long enough an innocent context should complete and 1480 * leave the engine idle. So they should not be caught unaware by 1481 * the forthcoming GPU reset (which usually follows the stop_cs)! 1482 */ 1483 return READ_ONCE(engine->props.stop_timeout_ms); 1484 } 1485 1486 static int __intel_engine_stop_cs(struct intel_engine_cs *engine, 1487 int fast_timeout_us, 1488 int slow_timeout_ms) 1489 { 1490 struct intel_uncore *uncore = engine->uncore; 1491 const i915_reg_t mode = RING_MI_MODE(engine->mmio_base); 1492 int err; 1493 1494 intel_uncore_write_fw(uncore, mode, _MASKED_BIT_ENABLE(STOP_RING)); 1495 1496 /* 1497 * Wa_22011802037: Prior to doing a reset, ensure CS is 1498 * stopped, set ring stop bit and prefetch disable bit to halt CS 1499 */ 1500 if (IS_MTL_GRAPHICS_STEP(engine->i915, M, STEP_A0, STEP_B0) || 1501 (GRAPHICS_VER(engine->i915) >= 11 && 1502 GRAPHICS_VER_FULL(engine->i915) < IP_VER(12, 70))) 1503 intel_uncore_write_fw(uncore, RING_MODE_GEN7(engine->mmio_base), 1504 _MASKED_BIT_ENABLE(GEN12_GFX_PREFETCH_DISABLE)); 1505 1506 err = __intel_wait_for_register_fw(engine->uncore, mode, 1507 MODE_IDLE, MODE_IDLE, 1508 fast_timeout_us, 1509 slow_timeout_ms, 1510 NULL); 1511 1512 /* A final mmio read to let GPU writes be hopefully flushed to memory */ 1513 intel_uncore_posting_read_fw(uncore, mode); 1514 return err; 1515 } 1516 1517 int intel_engine_stop_cs(struct intel_engine_cs *engine) 1518 { 1519 int err = 0; 1520 1521 if (GRAPHICS_VER(engine->i915) < 3) 1522 return -ENODEV; 1523 1524 ENGINE_TRACE(engine, "\n"); 1525 /* 1526 * TODO: Find out why occasionally stopping the CS times out. Seen 1527 * especially with gem_eio tests. 1528 * 1529 * Occasionally trying to stop the cs times out, but does not adversely 1530 * affect functionality. The timeout is set as a config parameter that 1531 * defaults to 100ms. In most cases the follow up operation is to wait 1532 * for pending MI_FORCE_WAKES. The assumption is that this timeout is 1533 * sufficient for any pending MI_FORCEWAKEs to complete. Once root 1534 * caused, the caller must check and handle the return from this 1535 * function. 1536 */ 1537 if (__intel_engine_stop_cs(engine, 1000, stop_timeout(engine))) { 1538 ENGINE_TRACE(engine, 1539 "timed out on STOP_RING -> IDLE; HEAD:%04x, TAIL:%04x\n", 1540 ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR, 1541 ENGINE_READ_FW(engine, RING_TAIL) & TAIL_ADDR); 1542 1543 /* 1544 * Sometimes we observe that the idle flag is not 1545 * set even though the ring is empty. So double 1546 * check before giving up. 1547 */ 1548 if ((ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR) != 1549 (ENGINE_READ_FW(engine, RING_TAIL) & TAIL_ADDR)) 1550 err = -ETIMEDOUT; 1551 } 1552 1553 return err; 1554 } 1555 1556 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine) 1557 { 1558 ENGINE_TRACE(engine, "\n"); 1559 1560 ENGINE_WRITE_FW(engine, RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING)); 1561 } 1562 1563 static u32 __cs_pending_mi_force_wakes(struct intel_engine_cs *engine) 1564 { 1565 static const i915_reg_t _reg[I915_NUM_ENGINES] = { 1566 [RCS0] = MSG_IDLE_CS, 1567 [BCS0] = MSG_IDLE_BCS, 1568 [VCS0] = MSG_IDLE_VCS0, 1569 [VCS1] = MSG_IDLE_VCS1, 1570 [VCS2] = MSG_IDLE_VCS2, 1571 [VCS3] = MSG_IDLE_VCS3, 1572 [VCS4] = MSG_IDLE_VCS4, 1573 [VCS5] = MSG_IDLE_VCS5, 1574 [VCS6] = MSG_IDLE_VCS6, 1575 [VCS7] = MSG_IDLE_VCS7, 1576 [VECS0] = MSG_IDLE_VECS0, 1577 [VECS1] = MSG_IDLE_VECS1, 1578 [VECS2] = MSG_IDLE_VECS2, 1579 [VECS3] = MSG_IDLE_VECS3, 1580 [CCS0] = MSG_IDLE_CS, 1581 [CCS1] = MSG_IDLE_CS, 1582 [CCS2] = MSG_IDLE_CS, 1583 [CCS3] = MSG_IDLE_CS, 1584 }; 1585 u32 val; 1586 1587 if (!_reg[engine->id].reg) 1588 return 0; 1589 1590 val = intel_uncore_read(engine->uncore, _reg[engine->id]); 1591 1592 /* bits[29:25] & bits[13:9] >> shift */ 1593 return (val & (val >> 16) & MSG_IDLE_FW_MASK) >> MSG_IDLE_FW_SHIFT; 1594 } 1595 1596 static void __gpm_wait_for_fw_complete(struct intel_gt *gt, u32 fw_mask) 1597 { 1598 int ret; 1599 1600 /* Ensure GPM receives fw up/down after CS is stopped */ 1601 udelay(1); 1602 1603 /* Wait for forcewake request to complete in GPM */ 1604 ret = __intel_wait_for_register_fw(gt->uncore, 1605 GEN9_PWRGT_DOMAIN_STATUS, 1606 fw_mask, fw_mask, 5000, 0, NULL); 1607 1608 /* Ensure CS receives fw ack from GPM */ 1609 udelay(1); 1610 1611 if (ret) 1612 GT_TRACE(gt, "Failed to complete pending forcewake %d\n", ret); 1613 } 1614 1615 /* 1616 * Wa_22011802037:gen12: In addition to stopping the cs, we need to wait for any 1617 * pending MI_FORCE_WAKEUP requests that the CS has initiated to complete. The 1618 * pending status is indicated by bits[13:9] (masked by bits[29:25]) in the 1619 * MSG_IDLE register. There's one MSG_IDLE register per reset domain. Since we 1620 * are concerned only with the gt reset here, we use a logical OR of pending 1621 * forcewakeups from all reset domains and then wait for them to complete by 1622 * querying PWRGT_DOMAIN_STATUS. 1623 */ 1624 void intel_engine_wait_for_pending_mi_fw(struct intel_engine_cs *engine) 1625 { 1626 u32 fw_pending = __cs_pending_mi_force_wakes(engine); 1627 1628 if (fw_pending) 1629 __gpm_wait_for_fw_complete(engine->gt, fw_pending); 1630 } 1631 1632 /* NB: please notice the memset */ 1633 void intel_engine_get_instdone(const struct intel_engine_cs *engine, 1634 struct intel_instdone *instdone) 1635 { 1636 struct drm_i915_private *i915 = engine->i915; 1637 struct intel_uncore *uncore = engine->uncore; 1638 u32 mmio_base = engine->mmio_base; 1639 int slice; 1640 int subslice; 1641 int iter; 1642 1643 memset(instdone, 0, sizeof(*instdone)); 1644 1645 if (GRAPHICS_VER(i915) >= 8) { 1646 instdone->instdone = 1647 intel_uncore_read(uncore, RING_INSTDONE(mmio_base)); 1648 1649 if (engine->id != RCS0) 1650 return; 1651 1652 instdone->slice_common = 1653 intel_uncore_read(uncore, GEN7_SC_INSTDONE); 1654 if (GRAPHICS_VER(i915) >= 12) { 1655 instdone->slice_common_extra[0] = 1656 intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA); 1657 instdone->slice_common_extra[1] = 1658 intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA2); 1659 } 1660 1661 for_each_ss_steering(iter, engine->gt, slice, subslice) { 1662 instdone->sampler[slice][subslice] = 1663 intel_gt_mcr_read(engine->gt, 1664 GEN8_SAMPLER_INSTDONE, 1665 slice, subslice); 1666 instdone->row[slice][subslice] = 1667 intel_gt_mcr_read(engine->gt, 1668 GEN8_ROW_INSTDONE, 1669 slice, subslice); 1670 } 1671 1672 if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 55)) { 1673 for_each_ss_steering(iter, engine->gt, slice, subslice) 1674 instdone->geom_svg[slice][subslice] = 1675 intel_gt_mcr_read(engine->gt, 1676 XEHPG_INSTDONE_GEOM_SVG, 1677 slice, subslice); 1678 } 1679 } else if (GRAPHICS_VER(i915) >= 7) { 1680 instdone->instdone = 1681 intel_uncore_read(uncore, RING_INSTDONE(mmio_base)); 1682 1683 if (engine->id != RCS0) 1684 return; 1685 1686 instdone->slice_common = 1687 intel_uncore_read(uncore, GEN7_SC_INSTDONE); 1688 instdone->sampler[0][0] = 1689 intel_uncore_read(uncore, GEN7_SAMPLER_INSTDONE); 1690 instdone->row[0][0] = 1691 intel_uncore_read(uncore, GEN7_ROW_INSTDONE); 1692 } else if (GRAPHICS_VER(i915) >= 4) { 1693 instdone->instdone = 1694 intel_uncore_read(uncore, RING_INSTDONE(mmio_base)); 1695 if (engine->id == RCS0) 1696 /* HACK: Using the wrong struct member */ 1697 instdone->slice_common = 1698 intel_uncore_read(uncore, GEN4_INSTDONE1); 1699 } else { 1700 instdone->instdone = intel_uncore_read(uncore, GEN2_INSTDONE); 1701 } 1702 } 1703 1704 static bool ring_is_idle(struct intel_engine_cs *engine) 1705 { 1706 bool idle = true; 1707 1708 if (I915_SELFTEST_ONLY(!engine->mmio_base)) 1709 return true; 1710 1711 if (!intel_engine_pm_get_if_awake(engine)) 1712 return true; 1713 1714 /* First check that no commands are left in the ring */ 1715 if ((ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) != 1716 (ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR)) 1717 idle = false; 1718 1719 /* No bit for gen2, so assume the CS parser is idle */ 1720 if (GRAPHICS_VER(engine->i915) > 2 && 1721 !(ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE)) 1722 idle = false; 1723 1724 intel_engine_pm_put(engine); 1725 1726 return idle; 1727 } 1728 1729 void __intel_engine_flush_submission(struct intel_engine_cs *engine, bool sync) 1730 { 1731 struct tasklet_struct *t = &engine->sched_engine->tasklet; 1732 1733 if (!t->callback) 1734 return; 1735 1736 local_bh_disable(); 1737 if (tasklet_trylock(t)) { 1738 /* Must wait for any GPU reset in progress. */ 1739 if (__tasklet_is_enabled(t)) 1740 t->callback(t); 1741 tasklet_unlock(t); 1742 } 1743 local_bh_enable(); 1744 1745 /* Synchronise and wait for the tasklet on another CPU */ 1746 if (sync) 1747 tasklet_unlock_wait(t); 1748 } 1749 1750 /** 1751 * intel_engine_is_idle() - Report if the engine has finished process all work 1752 * @engine: the intel_engine_cs 1753 * 1754 * Return true if there are no requests pending, nothing left to be submitted 1755 * to hardware, and that the engine is idle. 1756 */ 1757 bool intel_engine_is_idle(struct intel_engine_cs *engine) 1758 { 1759 /* More white lies, if wedged, hw state is inconsistent */ 1760 if (intel_gt_is_wedged(engine->gt)) 1761 return true; 1762 1763 if (!intel_engine_pm_is_awake(engine)) 1764 return true; 1765 1766 /* Waiting to drain ELSP? */ 1767 intel_synchronize_hardirq(engine->i915); 1768 intel_engine_flush_submission(engine); 1769 1770 /* ELSP is empty, but there are ready requests? E.g. after reset */ 1771 if (!i915_sched_engine_is_empty(engine->sched_engine)) 1772 return false; 1773 1774 /* Ring stopped? */ 1775 return ring_is_idle(engine); 1776 } 1777 1778 bool intel_engines_are_idle(struct intel_gt *gt) 1779 { 1780 struct intel_engine_cs *engine; 1781 enum intel_engine_id id; 1782 1783 /* 1784 * If the driver is wedged, HW state may be very inconsistent and 1785 * report that it is still busy, even though we have stopped using it. 1786 */ 1787 if (intel_gt_is_wedged(gt)) 1788 return true; 1789 1790 /* Already parked (and passed an idleness test); must still be idle */ 1791 if (!READ_ONCE(gt->awake)) 1792 return true; 1793 1794 for_each_engine(engine, gt, id) { 1795 if (!intel_engine_is_idle(engine)) 1796 return false; 1797 } 1798 1799 return true; 1800 } 1801 1802 bool intel_engine_irq_enable(struct intel_engine_cs *engine) 1803 { 1804 if (!engine->irq_enable) 1805 return false; 1806 1807 /* Caller disables interrupts */ 1808 spin_lock(engine->gt->irq_lock); 1809 engine->irq_enable(engine); 1810 spin_unlock(engine->gt->irq_lock); 1811 1812 return true; 1813 } 1814 1815 void intel_engine_irq_disable(struct intel_engine_cs *engine) 1816 { 1817 if (!engine->irq_disable) 1818 return; 1819 1820 /* Caller disables interrupts */ 1821 spin_lock(engine->gt->irq_lock); 1822 engine->irq_disable(engine); 1823 spin_unlock(engine->gt->irq_lock); 1824 } 1825 1826 void intel_engines_reset_default_submission(struct intel_gt *gt) 1827 { 1828 struct intel_engine_cs *engine; 1829 enum intel_engine_id id; 1830 1831 for_each_engine(engine, gt, id) { 1832 if (engine->sanitize) 1833 engine->sanitize(engine); 1834 1835 engine->set_default_submission(engine); 1836 } 1837 } 1838 1839 bool intel_engine_can_store_dword(struct intel_engine_cs *engine) 1840 { 1841 switch (GRAPHICS_VER(engine->i915)) { 1842 case 2: 1843 return false; /* uses physical not virtual addresses */ 1844 case 3: 1845 /* maybe only uses physical not virtual addresses */ 1846 return !(IS_I915G(engine->i915) || IS_I915GM(engine->i915)); 1847 case 4: 1848 return !IS_I965G(engine->i915); /* who knows! */ 1849 case 6: 1850 return engine->class != VIDEO_DECODE_CLASS; /* b0rked */ 1851 default: 1852 return true; 1853 } 1854 } 1855 1856 static struct intel_timeline *get_timeline(struct i915_request *rq) 1857 { 1858 struct intel_timeline *tl; 1859 1860 /* 1861 * Even though we are holding the engine->sched_engine->lock here, there 1862 * is no control over the submission queue per-se and we are 1863 * inspecting the active state at a random point in time, with an 1864 * unknown queue. Play safe and make sure the timeline remains valid. 1865 * (Only being used for pretty printing, one extra kref shouldn't 1866 * cause a camel stampede!) 1867 */ 1868 rcu_read_lock(); 1869 tl = rcu_dereference(rq->timeline); 1870 if (!kref_get_unless_zero(&tl->kref)) 1871 tl = NULL; 1872 rcu_read_unlock(); 1873 1874 return tl; 1875 } 1876 1877 static int print_ring(char *buf, int sz, struct i915_request *rq) 1878 { 1879 int len = 0; 1880 1881 if (!i915_request_signaled(rq)) { 1882 struct intel_timeline *tl = get_timeline(rq); 1883 1884 len = scnprintf(buf, sz, 1885 "ring:{start:%08x, hwsp:%08x, seqno:%08x, runtime:%llums}, ", 1886 i915_ggtt_offset(rq->ring->vma), 1887 tl ? tl->hwsp_offset : 0, 1888 hwsp_seqno(rq), 1889 DIV_ROUND_CLOSEST_ULL(intel_context_get_total_runtime_ns(rq->context), 1890 1000 * 1000)); 1891 1892 if (tl) 1893 intel_timeline_put(tl); 1894 } 1895 1896 return len; 1897 } 1898 1899 static void hexdump(struct drm_printer *m, const void *buf, size_t len) 1900 { 1901 const size_t rowsize = 8 * sizeof(u32); 1902 const void *prev = NULL; 1903 bool skip = false; 1904 size_t pos; 1905 1906 for (pos = 0; pos < len; pos += rowsize) { 1907 char line[128]; 1908 1909 if (prev && !memcmp(prev, buf + pos, rowsize)) { 1910 if (!skip) { 1911 drm_printf(m, "*\n"); 1912 skip = true; 1913 } 1914 continue; 1915 } 1916 1917 WARN_ON_ONCE(hex_dump_to_buffer(buf + pos, len - pos, 1918 rowsize, sizeof(u32), 1919 line, sizeof(line), 1920 false) >= sizeof(line)); 1921 drm_printf(m, "[%04zx] %s\n", pos, line); 1922 1923 prev = buf + pos; 1924 skip = false; 1925 } 1926 } 1927 1928 static const char *repr_timer(const struct timer_list *t) 1929 { 1930 if (!READ_ONCE(t->expires)) 1931 return "inactive"; 1932 1933 if (timer_pending(t)) 1934 return "active"; 1935 1936 return "expired"; 1937 } 1938 1939 static void intel_engine_print_registers(struct intel_engine_cs *engine, 1940 struct drm_printer *m) 1941 { 1942 struct drm_i915_private *dev_priv = engine->i915; 1943 struct intel_engine_execlists * const execlists = &engine->execlists; 1944 u64 addr; 1945 1946 if (engine->id == RENDER_CLASS && IS_GRAPHICS_VER(dev_priv, 4, 7)) 1947 drm_printf(m, "\tCCID: 0x%08x\n", ENGINE_READ(engine, CCID)); 1948 if (HAS_EXECLISTS(dev_priv)) { 1949 drm_printf(m, "\tEL_STAT_HI: 0x%08x\n", 1950 ENGINE_READ(engine, RING_EXECLIST_STATUS_HI)); 1951 drm_printf(m, "\tEL_STAT_LO: 0x%08x\n", 1952 ENGINE_READ(engine, RING_EXECLIST_STATUS_LO)); 1953 } 1954 drm_printf(m, "\tRING_START: 0x%08x\n", 1955 ENGINE_READ(engine, RING_START)); 1956 drm_printf(m, "\tRING_HEAD: 0x%08x\n", 1957 ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR); 1958 drm_printf(m, "\tRING_TAIL: 0x%08x\n", 1959 ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR); 1960 drm_printf(m, "\tRING_CTL: 0x%08x%s\n", 1961 ENGINE_READ(engine, RING_CTL), 1962 ENGINE_READ(engine, RING_CTL) & (RING_WAIT | RING_WAIT_SEMAPHORE) ? " [waiting]" : ""); 1963 if (GRAPHICS_VER(engine->i915) > 2) { 1964 drm_printf(m, "\tRING_MODE: 0x%08x%s\n", 1965 ENGINE_READ(engine, RING_MI_MODE), 1966 ENGINE_READ(engine, RING_MI_MODE) & (MODE_IDLE) ? " [idle]" : ""); 1967 } 1968 1969 if (GRAPHICS_VER(dev_priv) >= 6) { 1970 drm_printf(m, "\tRING_IMR: 0x%08x\n", 1971 ENGINE_READ(engine, RING_IMR)); 1972 drm_printf(m, "\tRING_ESR: 0x%08x\n", 1973 ENGINE_READ(engine, RING_ESR)); 1974 drm_printf(m, "\tRING_EMR: 0x%08x\n", 1975 ENGINE_READ(engine, RING_EMR)); 1976 drm_printf(m, "\tRING_EIR: 0x%08x\n", 1977 ENGINE_READ(engine, RING_EIR)); 1978 } 1979 1980 addr = intel_engine_get_active_head(engine); 1981 drm_printf(m, "\tACTHD: 0x%08x_%08x\n", 1982 upper_32_bits(addr), lower_32_bits(addr)); 1983 addr = intel_engine_get_last_batch_head(engine); 1984 drm_printf(m, "\tBBADDR: 0x%08x_%08x\n", 1985 upper_32_bits(addr), lower_32_bits(addr)); 1986 if (GRAPHICS_VER(dev_priv) >= 8) 1987 addr = ENGINE_READ64(engine, RING_DMA_FADD, RING_DMA_FADD_UDW); 1988 else if (GRAPHICS_VER(dev_priv) >= 4) 1989 addr = ENGINE_READ(engine, RING_DMA_FADD); 1990 else 1991 addr = ENGINE_READ(engine, DMA_FADD_I8XX); 1992 drm_printf(m, "\tDMA_FADDR: 0x%08x_%08x\n", 1993 upper_32_bits(addr), lower_32_bits(addr)); 1994 if (GRAPHICS_VER(dev_priv) >= 4) { 1995 drm_printf(m, "\tIPEIR: 0x%08x\n", 1996 ENGINE_READ(engine, RING_IPEIR)); 1997 drm_printf(m, "\tIPEHR: 0x%08x\n", 1998 ENGINE_READ(engine, RING_IPEHR)); 1999 } else { 2000 drm_printf(m, "\tIPEIR: 0x%08x\n", ENGINE_READ(engine, IPEIR)); 2001 drm_printf(m, "\tIPEHR: 0x%08x\n", ENGINE_READ(engine, IPEHR)); 2002 } 2003 2004 if (HAS_EXECLISTS(dev_priv) && !intel_engine_uses_guc(engine)) { 2005 struct i915_request * const *port, *rq; 2006 const u32 *hws = 2007 &engine->status_page.addr[I915_HWS_CSB_BUF0_INDEX]; 2008 const u8 num_entries = execlists->csb_size; 2009 unsigned int idx; 2010 u8 read, write; 2011 2012 drm_printf(m, "\tExeclist tasklet queued? %s (%s), preempt? %s, timeslice? %s\n", 2013 str_yes_no(test_bit(TASKLET_STATE_SCHED, &engine->sched_engine->tasklet.state)), 2014 str_enabled_disabled(!atomic_read(&engine->sched_engine->tasklet.count)), 2015 repr_timer(&engine->execlists.preempt), 2016 repr_timer(&engine->execlists.timer)); 2017 2018 read = execlists->csb_head; 2019 write = READ_ONCE(*execlists->csb_write); 2020 2021 drm_printf(m, "\tExeclist status: 0x%08x %08x; CSB read:%d, write:%d, entries:%d\n", 2022 ENGINE_READ(engine, RING_EXECLIST_STATUS_LO), 2023 ENGINE_READ(engine, RING_EXECLIST_STATUS_HI), 2024 read, write, num_entries); 2025 2026 if (read >= num_entries) 2027 read = 0; 2028 if (write >= num_entries) 2029 write = 0; 2030 if (read > write) 2031 write += num_entries; 2032 while (read < write) { 2033 idx = ++read % num_entries; 2034 drm_printf(m, "\tExeclist CSB[%d]: 0x%08x, context: %d\n", 2035 idx, hws[idx * 2], hws[idx * 2 + 1]); 2036 } 2037 2038 i915_sched_engine_active_lock_bh(engine->sched_engine); 2039 rcu_read_lock(); 2040 for (port = execlists->active; (rq = *port); port++) { 2041 char hdr[160]; 2042 int len; 2043 2044 len = scnprintf(hdr, sizeof(hdr), 2045 "\t\tActive[%d]: ccid:%08x%s%s, ", 2046 (int)(port - execlists->active), 2047 rq->context->lrc.ccid, 2048 intel_context_is_closed(rq->context) ? "!" : "", 2049 intel_context_is_banned(rq->context) ? "*" : ""); 2050 len += print_ring(hdr + len, sizeof(hdr) - len, rq); 2051 scnprintf(hdr + len, sizeof(hdr) - len, "rq: "); 2052 i915_request_show(m, rq, hdr, 0); 2053 } 2054 for (port = execlists->pending; (rq = *port); port++) { 2055 char hdr[160]; 2056 int len; 2057 2058 len = scnprintf(hdr, sizeof(hdr), 2059 "\t\tPending[%d]: ccid:%08x%s%s, ", 2060 (int)(port - execlists->pending), 2061 rq->context->lrc.ccid, 2062 intel_context_is_closed(rq->context) ? "!" : "", 2063 intel_context_is_banned(rq->context) ? "*" : ""); 2064 len += print_ring(hdr + len, sizeof(hdr) - len, rq); 2065 scnprintf(hdr + len, sizeof(hdr) - len, "rq: "); 2066 i915_request_show(m, rq, hdr, 0); 2067 } 2068 rcu_read_unlock(); 2069 i915_sched_engine_active_unlock_bh(engine->sched_engine); 2070 } else if (GRAPHICS_VER(dev_priv) > 6) { 2071 drm_printf(m, "\tPP_DIR_BASE: 0x%08x\n", 2072 ENGINE_READ(engine, RING_PP_DIR_BASE)); 2073 drm_printf(m, "\tPP_DIR_BASE_READ: 0x%08x\n", 2074 ENGINE_READ(engine, RING_PP_DIR_BASE_READ)); 2075 drm_printf(m, "\tPP_DIR_DCLV: 0x%08x\n", 2076 ENGINE_READ(engine, RING_PP_DIR_DCLV)); 2077 } 2078 } 2079 2080 static void print_request_ring(struct drm_printer *m, struct i915_request *rq) 2081 { 2082 struct i915_vma_resource *vma_res = rq->batch_res; 2083 void *ring; 2084 int size; 2085 2086 drm_printf(m, 2087 "[head %04x, postfix %04x, tail %04x, batch 0x%08x_%08x]:\n", 2088 rq->head, rq->postfix, rq->tail, 2089 vma_res ? upper_32_bits(vma_res->start) : ~0u, 2090 vma_res ? lower_32_bits(vma_res->start) : ~0u); 2091 2092 size = rq->tail - rq->head; 2093 if (rq->tail < rq->head) 2094 size += rq->ring->size; 2095 2096 ring = kmalloc(size, GFP_ATOMIC); 2097 if (ring) { 2098 const void *vaddr = rq->ring->vaddr; 2099 unsigned int head = rq->head; 2100 unsigned int len = 0; 2101 2102 if (rq->tail < head) { 2103 len = rq->ring->size - head; 2104 memcpy(ring, vaddr + head, len); 2105 head = 0; 2106 } 2107 memcpy(ring + len, vaddr + head, size - len); 2108 2109 hexdump(m, ring, size); 2110 kfree(ring); 2111 } 2112 } 2113 2114 static unsigned long read_ul(void *p, size_t x) 2115 { 2116 return *(unsigned long *)(p + x); 2117 } 2118 2119 static void print_properties(struct intel_engine_cs *engine, 2120 struct drm_printer *m) 2121 { 2122 static const struct pmap { 2123 size_t offset; 2124 const char *name; 2125 } props[] = { 2126 #define P(x) { \ 2127 .offset = offsetof(typeof(engine->props), x), \ 2128 .name = #x \ 2129 } 2130 P(heartbeat_interval_ms), 2131 P(max_busywait_duration_ns), 2132 P(preempt_timeout_ms), 2133 P(stop_timeout_ms), 2134 P(timeslice_duration_ms), 2135 2136 {}, 2137 #undef P 2138 }; 2139 const struct pmap *p; 2140 2141 drm_printf(m, "\tProperties:\n"); 2142 for (p = props; p->name; p++) 2143 drm_printf(m, "\t\t%s: %lu [default %lu]\n", 2144 p->name, 2145 read_ul(&engine->props, p->offset), 2146 read_ul(&engine->defaults, p->offset)); 2147 } 2148 2149 static void engine_dump_request(struct i915_request *rq, struct drm_printer *m, const char *msg) 2150 { 2151 struct intel_timeline *tl = get_timeline(rq); 2152 2153 i915_request_show(m, rq, msg, 0); 2154 2155 drm_printf(m, "\t\tring->start: 0x%08x\n", 2156 i915_ggtt_offset(rq->ring->vma)); 2157 drm_printf(m, "\t\tring->head: 0x%08x\n", 2158 rq->ring->head); 2159 drm_printf(m, "\t\tring->tail: 0x%08x\n", 2160 rq->ring->tail); 2161 drm_printf(m, "\t\tring->emit: 0x%08x\n", 2162 rq->ring->emit); 2163 drm_printf(m, "\t\tring->space: 0x%08x\n", 2164 rq->ring->space); 2165 2166 if (tl) { 2167 drm_printf(m, "\t\tring->hwsp: 0x%08x\n", 2168 tl->hwsp_offset); 2169 intel_timeline_put(tl); 2170 } 2171 2172 print_request_ring(m, rq); 2173 2174 if (rq->context->lrc_reg_state) { 2175 drm_printf(m, "Logical Ring Context:\n"); 2176 hexdump(m, rq->context->lrc_reg_state, PAGE_SIZE); 2177 } 2178 } 2179 2180 void intel_engine_dump_active_requests(struct list_head *requests, 2181 struct i915_request *hung_rq, 2182 struct drm_printer *m) 2183 { 2184 struct i915_request *rq; 2185 const char *msg; 2186 enum i915_request_state state; 2187 2188 list_for_each_entry(rq, requests, sched.link) { 2189 if (rq == hung_rq) 2190 continue; 2191 2192 state = i915_test_request_state(rq); 2193 if (state < I915_REQUEST_QUEUED) 2194 continue; 2195 2196 if (state == I915_REQUEST_ACTIVE) 2197 msg = "\t\tactive on engine"; 2198 else 2199 msg = "\t\tactive in queue"; 2200 2201 engine_dump_request(rq, m, msg); 2202 } 2203 } 2204 2205 static void engine_dump_active_requests(struct intel_engine_cs *engine, 2206 struct drm_printer *m) 2207 { 2208 struct intel_context *hung_ce = NULL; 2209 struct i915_request *hung_rq = NULL; 2210 2211 /* 2212 * No need for an engine->irq_seqno_barrier() before the seqno reads. 2213 * The GPU is still running so requests are still executing and any 2214 * hardware reads will be out of date by the time they are reported. 2215 * But the intention here is just to report an instantaneous snapshot 2216 * so that's fine. 2217 */ 2218 intel_engine_get_hung_entity(engine, &hung_ce, &hung_rq); 2219 2220 drm_printf(m, "\tRequests:\n"); 2221 2222 if (hung_rq) 2223 engine_dump_request(hung_rq, m, "\t\thung"); 2224 else if (hung_ce) 2225 drm_printf(m, "\t\tGot hung ce but no hung rq!\n"); 2226 2227 if (intel_uc_uses_guc_submission(&engine->gt->uc)) 2228 intel_guc_dump_active_requests(engine, hung_rq, m); 2229 else 2230 intel_execlists_dump_active_requests(engine, hung_rq, m); 2231 2232 if (hung_rq) 2233 i915_request_put(hung_rq); 2234 } 2235 2236 void intel_engine_dump(struct intel_engine_cs *engine, 2237 struct drm_printer *m, 2238 const char *header, ...) 2239 { 2240 struct i915_gpu_error * const error = &engine->i915->gpu_error; 2241 struct i915_request *rq; 2242 intel_wakeref_t wakeref; 2243 ktime_t dummy; 2244 2245 if (header) { 2246 va_list ap; 2247 2248 va_start(ap, header); 2249 drm_vprintf(m, header, &ap); 2250 va_end(ap); 2251 } 2252 2253 if (intel_gt_is_wedged(engine->gt)) 2254 drm_printf(m, "*** WEDGED ***\n"); 2255 2256 drm_printf(m, "\tAwake? %d\n", atomic_read(&engine->wakeref.count)); 2257 drm_printf(m, "\tBarriers?: %s\n", 2258 str_yes_no(!llist_empty(&engine->barrier_tasks))); 2259 drm_printf(m, "\tLatency: %luus\n", 2260 ewma__engine_latency_read(&engine->latency)); 2261 if (intel_engine_supports_stats(engine)) 2262 drm_printf(m, "\tRuntime: %llums\n", 2263 ktime_to_ms(intel_engine_get_busy_time(engine, 2264 &dummy))); 2265 drm_printf(m, "\tForcewake: %x domains, %d active\n", 2266 engine->fw_domain, READ_ONCE(engine->fw_active)); 2267 2268 rcu_read_lock(); 2269 rq = READ_ONCE(engine->heartbeat.systole); 2270 if (rq) 2271 drm_printf(m, "\tHeartbeat: %d ms ago\n", 2272 jiffies_to_msecs(jiffies - rq->emitted_jiffies)); 2273 rcu_read_unlock(); 2274 drm_printf(m, "\tReset count: %d (global %d)\n", 2275 i915_reset_engine_count(error, engine), 2276 i915_reset_count(error)); 2277 print_properties(engine, m); 2278 2279 engine_dump_active_requests(engine, m); 2280 2281 drm_printf(m, "\tMMIO base: 0x%08x\n", engine->mmio_base); 2282 wakeref = intel_runtime_pm_get_if_in_use(engine->uncore->rpm); 2283 if (wakeref) { 2284 intel_engine_print_registers(engine, m); 2285 intel_runtime_pm_put(engine->uncore->rpm, wakeref); 2286 } else { 2287 drm_printf(m, "\tDevice is asleep; skipping register dump\n"); 2288 } 2289 2290 intel_execlists_show_requests(engine, m, i915_request_show, 8); 2291 2292 drm_printf(m, "HWSP:\n"); 2293 hexdump(m, engine->status_page.addr, PAGE_SIZE); 2294 2295 drm_printf(m, "Idle? %s\n", str_yes_no(intel_engine_is_idle(engine))); 2296 2297 intel_engine_print_breadcrumbs(engine, m); 2298 } 2299 2300 /** 2301 * intel_engine_get_busy_time() - Return current accumulated engine busyness 2302 * @engine: engine to report on 2303 * @now: monotonic timestamp of sampling 2304 * 2305 * Returns accumulated time @engine was busy since engine stats were enabled. 2306 */ 2307 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine, ktime_t *now) 2308 { 2309 return engine->busyness(engine, now); 2310 } 2311 2312 struct intel_context * 2313 intel_engine_create_virtual(struct intel_engine_cs **siblings, 2314 unsigned int count, unsigned long flags) 2315 { 2316 if (count == 0) 2317 return ERR_PTR(-EINVAL); 2318 2319 if (count == 1 && !(flags & FORCE_VIRTUAL)) 2320 return intel_context_create(siblings[0]); 2321 2322 GEM_BUG_ON(!siblings[0]->cops->create_virtual); 2323 return siblings[0]->cops->create_virtual(siblings, count, flags); 2324 } 2325 2326 static struct i915_request *engine_execlist_find_hung_request(struct intel_engine_cs *engine) 2327 { 2328 struct i915_request *request, *active = NULL; 2329 2330 /* 2331 * This search does not work in GuC submission mode. However, the GuC 2332 * will report the hanging context directly to the driver itself. So 2333 * the driver should never get here when in GuC mode. 2334 */ 2335 GEM_BUG_ON(intel_uc_uses_guc_submission(&engine->gt->uc)); 2336 2337 /* 2338 * We are called by the error capture, reset and to dump engine 2339 * state at random points in time. In particular, note that neither is 2340 * crucially ordered with an interrupt. After a hang, the GPU is dead 2341 * and we assume that no more writes can happen (we waited long enough 2342 * for all writes that were in transaction to be flushed) - adding an 2343 * extra delay for a recent interrupt is pointless. Hence, we do 2344 * not need an engine->irq_seqno_barrier() before the seqno reads. 2345 * At all other times, we must assume the GPU is still running, but 2346 * we only care about the snapshot of this moment. 2347 */ 2348 lockdep_assert_held(&engine->sched_engine->lock); 2349 2350 rcu_read_lock(); 2351 request = execlists_active(&engine->execlists); 2352 if (request) { 2353 struct intel_timeline *tl = request->context->timeline; 2354 2355 list_for_each_entry_from_reverse(request, &tl->requests, link) { 2356 if (__i915_request_is_complete(request)) 2357 break; 2358 2359 active = request; 2360 } 2361 } 2362 rcu_read_unlock(); 2363 if (active) 2364 return active; 2365 2366 list_for_each_entry(request, &engine->sched_engine->requests, 2367 sched.link) { 2368 if (i915_test_request_state(request) != I915_REQUEST_ACTIVE) 2369 continue; 2370 2371 active = request; 2372 break; 2373 } 2374 2375 return active; 2376 } 2377 2378 void intel_engine_get_hung_entity(struct intel_engine_cs *engine, 2379 struct intel_context **ce, struct i915_request **rq) 2380 { 2381 unsigned long flags; 2382 2383 *ce = intel_engine_get_hung_context(engine); 2384 if (*ce) { 2385 intel_engine_clear_hung_context(engine); 2386 2387 *rq = intel_context_get_active_request(*ce); 2388 return; 2389 } 2390 2391 /* 2392 * Getting here with GuC enabled means it is a forced error capture 2393 * with no actual hang. So, no need to attempt the execlist search. 2394 */ 2395 if (intel_uc_uses_guc_submission(&engine->gt->uc)) 2396 return; 2397 2398 spin_lock_irqsave(&engine->sched_engine->lock, flags); 2399 *rq = engine_execlist_find_hung_request(engine); 2400 if (*rq) 2401 *rq = i915_request_get_rcu(*rq); 2402 spin_unlock_irqrestore(&engine->sched_engine->lock, flags); 2403 } 2404 2405 void xehp_enable_ccs_engines(struct intel_engine_cs *engine) 2406 { 2407 /* 2408 * If there are any non-fused-off CCS engines, we need to enable CCS 2409 * support in the RCU_MODE register. This only needs to be done once, 2410 * so for simplicity we'll take care of this in the RCS engine's 2411 * resume handler; since the RCS and all CCS engines belong to the 2412 * same reset domain and are reset together, this will also take care 2413 * of re-applying the setting after i915-triggered resets. 2414 */ 2415 if (!CCS_MASK(engine->gt)) 2416 return; 2417 2418 intel_uncore_write(engine->uncore, GEN12_RCU_MODE, 2419 _MASKED_BIT_ENABLE(GEN12_RCU_MODE_CCS_ENABLE)); 2420 } 2421 2422 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 2423 #include "mock_engine.c" 2424 #include "selftest_engine.c" 2425 #include "selftest_engine_cs.c" 2426 #endif 2427