1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2019 Intel Corporation 4 */ 5 6 #include <drm/drm_managed.h> 7 #include <drm/intel-gtt.h> 8 9 #include "gem/i915_gem_internal.h" 10 #include "gem/i915_gem_lmem.h" 11 12 #include "i915_drv.h" 13 #include "i915_perf_oa_regs.h" 14 #include "i915_reg.h" 15 #include "intel_context.h" 16 #include "intel_engine_pm.h" 17 #include "intel_engine_regs.h" 18 #include "intel_ggtt_gmch.h" 19 #include "intel_gt.h" 20 #include "intel_gt_buffer_pool.h" 21 #include "intel_gt_clock_utils.h" 22 #include "intel_gt_debugfs.h" 23 #include "intel_gt_mcr.h" 24 #include "intel_gt_pm.h" 25 #include "intel_gt_print.h" 26 #include "intel_gt_regs.h" 27 #include "intel_gt_requests.h" 28 #include "intel_migrate.h" 29 #include "intel_mocs.h" 30 #include "intel_pci_config.h" 31 #include "intel_pm.h" 32 #include "intel_rc6.h" 33 #include "intel_renderstate.h" 34 #include "intel_rps.h" 35 #include "intel_sa_media.h" 36 #include "intel_gt_sysfs.h" 37 #include "intel_uncore.h" 38 #include "shmem_utils.h" 39 40 void intel_gt_common_init_early(struct intel_gt *gt) 41 { 42 spin_lock_init(gt->irq_lock); 43 44 INIT_LIST_HEAD(>->closed_vma); 45 spin_lock_init(>->closed_lock); 46 47 init_llist_head(>->watchdog.list); 48 INIT_WORK(>->watchdog.work, intel_gt_watchdog_work); 49 50 intel_gt_init_buffer_pool(gt); 51 intel_gt_init_reset(gt); 52 intel_gt_init_requests(gt); 53 intel_gt_init_timelines(gt); 54 mutex_init(>->tlb.invalidate_lock); 55 seqcount_mutex_init(>->tlb.seqno, >->tlb.invalidate_lock); 56 intel_gt_pm_init_early(gt); 57 58 intel_wopcm_init_early(>->wopcm); 59 intel_uc_init_early(>->uc); 60 intel_rps_init_early(>->rps); 61 } 62 63 /* Preliminary initialization of Tile 0 */ 64 int intel_root_gt_init_early(struct drm_i915_private *i915) 65 { 66 struct intel_gt *gt = to_gt(i915); 67 68 gt->i915 = i915; 69 gt->uncore = &i915->uncore; 70 gt->irq_lock = drmm_kzalloc(&i915->drm, sizeof(*gt->irq_lock), GFP_KERNEL); 71 if (!gt->irq_lock) 72 return -ENOMEM; 73 74 intel_gt_common_init_early(gt); 75 76 return 0; 77 } 78 79 static int intel_gt_probe_lmem(struct intel_gt *gt) 80 { 81 struct drm_i915_private *i915 = gt->i915; 82 unsigned int instance = gt->info.id; 83 int id = INTEL_REGION_LMEM_0 + instance; 84 struct intel_memory_region *mem; 85 int err; 86 87 mem = intel_gt_setup_lmem(gt); 88 if (IS_ERR(mem)) { 89 err = PTR_ERR(mem); 90 if (err == -ENODEV) 91 return 0; 92 93 gt_err(gt, "Failed to setup region(%d) type=%d\n", 94 err, INTEL_MEMORY_LOCAL); 95 return err; 96 } 97 98 mem->id = id; 99 mem->instance = instance; 100 101 intel_memory_region_set_name(mem, "local%u", mem->instance); 102 103 GEM_BUG_ON(!HAS_REGION(i915, id)); 104 GEM_BUG_ON(i915->mm.regions[id]); 105 i915->mm.regions[id] = mem; 106 107 return 0; 108 } 109 110 int intel_gt_assign_ggtt(struct intel_gt *gt) 111 { 112 /* Media GT shares primary GT's GGTT */ 113 if (gt->type == GT_MEDIA) { 114 gt->ggtt = to_gt(gt->i915)->ggtt; 115 } else { 116 gt->ggtt = i915_ggtt_create(gt->i915); 117 if (IS_ERR(gt->ggtt)) 118 return PTR_ERR(gt->ggtt); 119 } 120 121 list_add_tail(>->ggtt_link, >->ggtt->gt_list); 122 123 return 0; 124 } 125 126 int intel_gt_init_mmio(struct intel_gt *gt) 127 { 128 intel_gt_init_clock_frequency(gt); 129 130 intel_uc_init_mmio(>->uc); 131 intel_sseu_info_init(gt); 132 intel_gt_mcr_init(gt); 133 134 return intel_engines_init_mmio(gt); 135 } 136 137 static void init_unused_ring(struct intel_gt *gt, u32 base) 138 { 139 struct intel_uncore *uncore = gt->uncore; 140 141 intel_uncore_write(uncore, RING_CTL(base), 0); 142 intel_uncore_write(uncore, RING_HEAD(base), 0); 143 intel_uncore_write(uncore, RING_TAIL(base), 0); 144 intel_uncore_write(uncore, RING_START(base), 0); 145 } 146 147 static void init_unused_rings(struct intel_gt *gt) 148 { 149 struct drm_i915_private *i915 = gt->i915; 150 151 if (IS_I830(i915)) { 152 init_unused_ring(gt, PRB1_BASE); 153 init_unused_ring(gt, SRB0_BASE); 154 init_unused_ring(gt, SRB1_BASE); 155 init_unused_ring(gt, SRB2_BASE); 156 init_unused_ring(gt, SRB3_BASE); 157 } else if (GRAPHICS_VER(i915) == 2) { 158 init_unused_ring(gt, SRB0_BASE); 159 init_unused_ring(gt, SRB1_BASE); 160 } else if (GRAPHICS_VER(i915) == 3) { 161 init_unused_ring(gt, PRB1_BASE); 162 init_unused_ring(gt, PRB2_BASE); 163 } 164 } 165 166 int intel_gt_init_hw(struct intel_gt *gt) 167 { 168 struct drm_i915_private *i915 = gt->i915; 169 struct intel_uncore *uncore = gt->uncore; 170 int ret; 171 172 gt->last_init_time = ktime_get(); 173 174 /* Double layer security blanket, see i915_gem_init() */ 175 intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL); 176 177 if (HAS_EDRAM(i915) && GRAPHICS_VER(i915) < 9) 178 intel_uncore_rmw(uncore, HSW_IDICR, 0, IDIHASHMSK(0xf)); 179 180 if (IS_HASWELL(i915)) 181 intel_uncore_write(uncore, 182 HSW_MI_PREDICATE_RESULT_2, 183 IS_HSW_GT3(i915) ? 184 LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED); 185 186 /* Apply the GT workarounds... */ 187 intel_gt_apply_workarounds(gt); 188 /* ...and determine whether they are sticking. */ 189 intel_gt_verify_workarounds(gt, "init"); 190 191 intel_gt_init_swizzling(gt); 192 193 /* 194 * At least 830 can leave some of the unused rings 195 * "active" (ie. head != tail) after resume which 196 * will prevent c3 entry. Makes sure all unused rings 197 * are totally idle. 198 */ 199 init_unused_rings(gt); 200 201 ret = i915_ppgtt_init_hw(gt); 202 if (ret) { 203 gt_err(gt, "Enabling PPGTT failed (%d)\n", ret); 204 goto out; 205 } 206 207 /* We can't enable contexts until all firmware is loaded */ 208 ret = intel_uc_init_hw(>->uc); 209 if (ret) { 210 gt_probe_error(gt, "Enabling uc failed (%d)\n", ret); 211 goto out; 212 } 213 214 intel_mocs_init(gt); 215 216 out: 217 intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL); 218 return ret; 219 } 220 221 static void gen6_clear_engine_error_register(struct intel_engine_cs *engine) 222 { 223 GEN6_RING_FAULT_REG_RMW(engine, RING_FAULT_VALID, 0); 224 GEN6_RING_FAULT_REG_POSTING_READ(engine); 225 } 226 227 i915_reg_t intel_gt_perf_limit_reasons_reg(struct intel_gt *gt) 228 { 229 /* GT0_PERF_LIMIT_REASONS is available only for Gen11+ */ 230 if (GRAPHICS_VER(gt->i915) < 11) 231 return INVALID_MMIO_REG; 232 233 return gt->type == GT_MEDIA ? 234 MTL_MEDIA_PERF_LIMIT_REASONS : GT0_PERF_LIMIT_REASONS; 235 } 236 237 void 238 intel_gt_clear_error_registers(struct intel_gt *gt, 239 intel_engine_mask_t engine_mask) 240 { 241 struct drm_i915_private *i915 = gt->i915; 242 struct intel_uncore *uncore = gt->uncore; 243 u32 eir; 244 245 if (GRAPHICS_VER(i915) != 2) 246 intel_uncore_write(uncore, PGTBL_ER, 0); 247 248 if (GRAPHICS_VER(i915) < 4) 249 intel_uncore_write(uncore, IPEIR(RENDER_RING_BASE), 0); 250 else 251 intel_uncore_write(uncore, IPEIR_I965, 0); 252 253 intel_uncore_write(uncore, EIR, 0); 254 eir = intel_uncore_read(uncore, EIR); 255 if (eir) { 256 /* 257 * some errors might have become stuck, 258 * mask them. 259 */ 260 gt_dbg(gt, "EIR stuck: 0x%08x, masking\n", eir); 261 intel_uncore_rmw(uncore, EMR, 0, eir); 262 intel_uncore_write(uncore, GEN2_IIR, 263 I915_MASTER_ERROR_INTERRUPT); 264 } 265 266 if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) { 267 intel_gt_mcr_multicast_rmw(gt, XEHP_RING_FAULT_REG, 268 RING_FAULT_VALID, 0); 269 intel_gt_mcr_read_any(gt, XEHP_RING_FAULT_REG); 270 } else if (GRAPHICS_VER(i915) >= 12) { 271 intel_uncore_rmw(uncore, GEN12_RING_FAULT_REG, RING_FAULT_VALID, 0); 272 intel_uncore_posting_read(uncore, GEN12_RING_FAULT_REG); 273 } else if (GRAPHICS_VER(i915) >= 8) { 274 intel_uncore_rmw(uncore, GEN8_RING_FAULT_REG, RING_FAULT_VALID, 0); 275 intel_uncore_posting_read(uncore, GEN8_RING_FAULT_REG); 276 } else if (GRAPHICS_VER(i915) >= 6) { 277 struct intel_engine_cs *engine; 278 enum intel_engine_id id; 279 280 for_each_engine_masked(engine, gt, engine_mask, id) 281 gen6_clear_engine_error_register(engine); 282 } 283 } 284 285 static void gen6_check_faults(struct intel_gt *gt) 286 { 287 struct intel_engine_cs *engine; 288 enum intel_engine_id id; 289 u32 fault; 290 291 for_each_engine(engine, gt, id) { 292 fault = GEN6_RING_FAULT_REG_READ(engine); 293 if (fault & RING_FAULT_VALID) { 294 gt_dbg(gt, "Unexpected fault\n" 295 "\tAddr: 0x%08lx\n" 296 "\tAddress space: %s\n" 297 "\tSource ID: %d\n" 298 "\tType: %d\n", 299 fault & PAGE_MASK, 300 fault & RING_FAULT_GTTSEL_MASK ? 301 "GGTT" : "PPGTT", 302 RING_FAULT_SRCID(fault), 303 RING_FAULT_FAULT_TYPE(fault)); 304 } 305 } 306 } 307 308 static void xehp_check_faults(struct intel_gt *gt) 309 { 310 u32 fault; 311 312 /* 313 * Although the fault register now lives in an MCR register range, 314 * the GAM registers are special and we only truly need to read 315 * the "primary" GAM instance rather than handling each instance 316 * individually. intel_gt_mcr_read_any() will automatically steer 317 * toward the primary instance. 318 */ 319 fault = intel_gt_mcr_read_any(gt, XEHP_RING_FAULT_REG); 320 if (fault & RING_FAULT_VALID) { 321 u32 fault_data0, fault_data1; 322 u64 fault_addr; 323 324 fault_data0 = intel_gt_mcr_read_any(gt, XEHP_FAULT_TLB_DATA0); 325 fault_data1 = intel_gt_mcr_read_any(gt, XEHP_FAULT_TLB_DATA1); 326 327 fault_addr = ((u64)(fault_data1 & FAULT_VA_HIGH_BITS) << 44) | 328 ((u64)fault_data0 << 12); 329 330 gt_dbg(gt, "Unexpected fault\n" 331 "\tAddr: 0x%08x_%08x\n" 332 "\tAddress space: %s\n" 333 "\tEngine ID: %d\n" 334 "\tSource ID: %d\n" 335 "\tType: %d\n", 336 upper_32_bits(fault_addr), lower_32_bits(fault_addr), 337 fault_data1 & FAULT_GTT_SEL ? "GGTT" : "PPGTT", 338 GEN8_RING_FAULT_ENGINE_ID(fault), 339 RING_FAULT_SRCID(fault), 340 RING_FAULT_FAULT_TYPE(fault)); 341 } 342 } 343 344 static void gen8_check_faults(struct intel_gt *gt) 345 { 346 struct intel_uncore *uncore = gt->uncore; 347 i915_reg_t fault_reg, fault_data0_reg, fault_data1_reg; 348 u32 fault; 349 350 if (GRAPHICS_VER(gt->i915) >= 12) { 351 fault_reg = GEN12_RING_FAULT_REG; 352 fault_data0_reg = GEN12_FAULT_TLB_DATA0; 353 fault_data1_reg = GEN12_FAULT_TLB_DATA1; 354 } else { 355 fault_reg = GEN8_RING_FAULT_REG; 356 fault_data0_reg = GEN8_FAULT_TLB_DATA0; 357 fault_data1_reg = GEN8_FAULT_TLB_DATA1; 358 } 359 360 fault = intel_uncore_read(uncore, fault_reg); 361 if (fault & RING_FAULT_VALID) { 362 u32 fault_data0, fault_data1; 363 u64 fault_addr; 364 365 fault_data0 = intel_uncore_read(uncore, fault_data0_reg); 366 fault_data1 = intel_uncore_read(uncore, fault_data1_reg); 367 368 fault_addr = ((u64)(fault_data1 & FAULT_VA_HIGH_BITS) << 44) | 369 ((u64)fault_data0 << 12); 370 371 gt_dbg(gt, "Unexpected fault\n" 372 "\tAddr: 0x%08x_%08x\n" 373 "\tAddress space: %s\n" 374 "\tEngine ID: %d\n" 375 "\tSource ID: %d\n" 376 "\tType: %d\n", 377 upper_32_bits(fault_addr), lower_32_bits(fault_addr), 378 fault_data1 & FAULT_GTT_SEL ? "GGTT" : "PPGTT", 379 GEN8_RING_FAULT_ENGINE_ID(fault), 380 RING_FAULT_SRCID(fault), 381 RING_FAULT_FAULT_TYPE(fault)); 382 } 383 } 384 385 void intel_gt_check_and_clear_faults(struct intel_gt *gt) 386 { 387 struct drm_i915_private *i915 = gt->i915; 388 389 /* From GEN8 onwards we only have one 'All Engine Fault Register' */ 390 if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) 391 xehp_check_faults(gt); 392 else if (GRAPHICS_VER(i915) >= 8) 393 gen8_check_faults(gt); 394 else if (GRAPHICS_VER(i915) >= 6) 395 gen6_check_faults(gt); 396 else 397 return; 398 399 intel_gt_clear_error_registers(gt, ALL_ENGINES); 400 } 401 402 void intel_gt_flush_ggtt_writes(struct intel_gt *gt) 403 { 404 struct intel_uncore *uncore = gt->uncore; 405 intel_wakeref_t wakeref; 406 407 /* 408 * No actual flushing is required for the GTT write domain for reads 409 * from the GTT domain. Writes to it "immediately" go to main memory 410 * as far as we know, so there's no chipset flush. It also doesn't 411 * land in the GPU render cache. 412 * 413 * However, we do have to enforce the order so that all writes through 414 * the GTT land before any writes to the device, such as updates to 415 * the GATT itself. 416 * 417 * We also have to wait a bit for the writes to land from the GTT. 418 * An uncached read (i.e. mmio) seems to be ideal for the round-trip 419 * timing. This issue has only been observed when switching quickly 420 * between GTT writes and CPU reads from inside the kernel on recent hw, 421 * and it appears to only affect discrete GTT blocks (i.e. on LLC 422 * system agents we cannot reproduce this behaviour, until Cannonlake 423 * that was!). 424 */ 425 426 wmb(); 427 428 if (INTEL_INFO(gt->i915)->has_coherent_ggtt) 429 return; 430 431 intel_gt_chipset_flush(gt); 432 433 with_intel_runtime_pm_if_in_use(uncore->rpm, wakeref) { 434 unsigned long flags; 435 436 spin_lock_irqsave(&uncore->lock, flags); 437 intel_uncore_posting_read_fw(uncore, 438 RING_HEAD(RENDER_RING_BASE)); 439 spin_unlock_irqrestore(&uncore->lock, flags); 440 } 441 } 442 443 void intel_gt_chipset_flush(struct intel_gt *gt) 444 { 445 wmb(); 446 if (GRAPHICS_VER(gt->i915) < 6) 447 intel_ggtt_gmch_flush(); 448 } 449 450 void intel_gt_driver_register(struct intel_gt *gt) 451 { 452 intel_gsc_init(>->gsc, gt->i915); 453 454 intel_rps_driver_register(>->rps); 455 456 intel_gt_debugfs_register(gt); 457 intel_gt_sysfs_register(gt); 458 } 459 460 static int intel_gt_init_scratch(struct intel_gt *gt, unsigned int size) 461 { 462 struct drm_i915_private *i915 = gt->i915; 463 struct drm_i915_gem_object *obj; 464 struct i915_vma *vma; 465 int ret; 466 467 obj = i915_gem_object_create_lmem(i915, size, 468 I915_BO_ALLOC_VOLATILE | 469 I915_BO_ALLOC_GPU_ONLY); 470 if (IS_ERR(obj)) 471 obj = i915_gem_object_create_stolen(i915, size); 472 if (IS_ERR(obj)) 473 obj = i915_gem_object_create_internal(i915, size); 474 if (IS_ERR(obj)) { 475 gt_err(gt, "Failed to allocate scratch page\n"); 476 return PTR_ERR(obj); 477 } 478 479 vma = i915_vma_instance(obj, >->ggtt->vm, NULL); 480 if (IS_ERR(vma)) { 481 ret = PTR_ERR(vma); 482 goto err_unref; 483 } 484 485 ret = i915_ggtt_pin(vma, NULL, 0, PIN_HIGH); 486 if (ret) 487 goto err_unref; 488 489 gt->scratch = i915_vma_make_unshrinkable(vma); 490 491 return 0; 492 493 err_unref: 494 i915_gem_object_put(obj); 495 return ret; 496 } 497 498 static void intel_gt_fini_scratch(struct intel_gt *gt) 499 { 500 i915_vma_unpin_and_release(>->scratch, 0); 501 } 502 503 static struct i915_address_space *kernel_vm(struct intel_gt *gt) 504 { 505 if (INTEL_PPGTT(gt->i915) > INTEL_PPGTT_ALIASING) 506 return &i915_ppgtt_create(gt, I915_BO_ALLOC_PM_EARLY)->vm; 507 else 508 return i915_vm_get(>->ggtt->vm); 509 } 510 511 static int __engines_record_defaults(struct intel_gt *gt) 512 { 513 struct i915_request *requests[I915_NUM_ENGINES] = {}; 514 struct intel_engine_cs *engine; 515 enum intel_engine_id id; 516 int err = 0; 517 518 /* 519 * As we reset the gpu during very early sanitisation, the current 520 * register state on the GPU should reflect its defaults values. 521 * We load a context onto the hw (with restore-inhibit), then switch 522 * over to a second context to save that default register state. We 523 * can then prime every new context with that state so they all start 524 * from the same default HW values. 525 */ 526 527 for_each_engine(engine, gt, id) { 528 struct intel_renderstate so; 529 struct intel_context *ce; 530 struct i915_request *rq; 531 532 /* We must be able to switch to something! */ 533 GEM_BUG_ON(!engine->kernel_context); 534 535 ce = intel_context_create(engine); 536 if (IS_ERR(ce)) { 537 err = PTR_ERR(ce); 538 goto out; 539 } 540 541 err = intel_renderstate_init(&so, ce); 542 if (err) 543 goto err; 544 545 rq = i915_request_create(ce); 546 if (IS_ERR(rq)) { 547 err = PTR_ERR(rq); 548 goto err_fini; 549 } 550 551 err = intel_engine_emit_ctx_wa(rq); 552 if (err) 553 goto err_rq; 554 555 err = intel_renderstate_emit(&so, rq); 556 if (err) 557 goto err_rq; 558 559 err_rq: 560 requests[id] = i915_request_get(rq); 561 i915_request_add(rq); 562 err_fini: 563 intel_renderstate_fini(&so, ce); 564 err: 565 if (err) { 566 intel_context_put(ce); 567 goto out; 568 } 569 } 570 571 /* Flush the default context image to memory, and enable powersaving. */ 572 if (intel_gt_wait_for_idle(gt, I915_GEM_IDLE_TIMEOUT) == -ETIME) { 573 err = -EIO; 574 goto out; 575 } 576 577 for (id = 0; id < ARRAY_SIZE(requests); id++) { 578 struct i915_request *rq; 579 struct file *state; 580 581 rq = requests[id]; 582 if (!rq) 583 continue; 584 585 if (rq->fence.error) { 586 err = -EIO; 587 goto out; 588 } 589 590 GEM_BUG_ON(!test_bit(CONTEXT_ALLOC_BIT, &rq->context->flags)); 591 if (!rq->context->state) 592 continue; 593 594 /* Keep a copy of the state's backing pages; free the obj */ 595 state = shmem_create_from_object(rq->context->state->obj); 596 if (IS_ERR(state)) { 597 err = PTR_ERR(state); 598 goto out; 599 } 600 rq->engine->default_state = state; 601 } 602 603 out: 604 /* 605 * If we have to abandon now, we expect the engines to be idle 606 * and ready to be torn-down. The quickest way we can accomplish 607 * this is by declaring ourselves wedged. 608 */ 609 if (err) 610 intel_gt_set_wedged(gt); 611 612 for (id = 0; id < ARRAY_SIZE(requests); id++) { 613 struct intel_context *ce; 614 struct i915_request *rq; 615 616 rq = requests[id]; 617 if (!rq) 618 continue; 619 620 ce = rq->context; 621 i915_request_put(rq); 622 intel_context_put(ce); 623 } 624 return err; 625 } 626 627 static int __engines_verify_workarounds(struct intel_gt *gt) 628 { 629 struct intel_engine_cs *engine; 630 enum intel_engine_id id; 631 int err = 0; 632 633 if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)) 634 return 0; 635 636 for_each_engine(engine, gt, id) { 637 if (intel_engine_verify_workarounds(engine, "load")) 638 err = -EIO; 639 } 640 641 /* Flush and restore the kernel context for safety */ 642 if (intel_gt_wait_for_idle(gt, I915_GEM_IDLE_TIMEOUT) == -ETIME) 643 err = -EIO; 644 645 return err; 646 } 647 648 static void __intel_gt_disable(struct intel_gt *gt) 649 { 650 intel_gt_set_wedged_on_fini(gt); 651 652 intel_gt_suspend_prepare(gt); 653 intel_gt_suspend_late(gt); 654 655 GEM_BUG_ON(intel_gt_pm_is_awake(gt)); 656 } 657 658 int intel_gt_wait_for_idle(struct intel_gt *gt, long timeout) 659 { 660 long remaining_timeout; 661 662 /* If the device is asleep, we have no requests outstanding */ 663 if (!intel_gt_pm_is_awake(gt)) 664 return 0; 665 666 while ((timeout = intel_gt_retire_requests_timeout(gt, timeout, 667 &remaining_timeout)) > 0) { 668 cond_resched(); 669 if (signal_pending(current)) 670 return -EINTR; 671 } 672 673 if (timeout) 674 return timeout; 675 676 if (remaining_timeout < 0) 677 remaining_timeout = 0; 678 679 return intel_uc_wait_for_idle(>->uc, remaining_timeout); 680 } 681 682 int intel_gt_init(struct intel_gt *gt) 683 { 684 int err; 685 686 err = i915_inject_probe_error(gt->i915, -ENODEV); 687 if (err) 688 return err; 689 690 intel_gt_init_workarounds(gt); 691 692 /* 693 * This is just a security blanket to placate dragons. 694 * On some systems, we very sporadically observe that the first TLBs 695 * used by the CS may be stale, despite us poking the TLB reset. If 696 * we hold the forcewake during initialisation these problems 697 * just magically go away. 698 */ 699 intel_uncore_forcewake_get(gt->uncore, FORCEWAKE_ALL); 700 701 err = intel_gt_init_scratch(gt, 702 GRAPHICS_VER(gt->i915) == 2 ? SZ_256K : SZ_4K); 703 if (err) 704 goto out_fw; 705 706 intel_gt_pm_init(gt); 707 708 gt->vm = kernel_vm(gt); 709 if (!gt->vm) { 710 err = -ENOMEM; 711 goto err_pm; 712 } 713 714 intel_set_mocs_index(gt); 715 716 err = intel_engines_init(gt); 717 if (err) 718 goto err_engines; 719 720 err = intel_uc_init(>->uc); 721 if (err) 722 goto err_engines; 723 724 err = intel_gt_resume(gt); 725 if (err) 726 goto err_uc_init; 727 728 err = intel_gt_init_hwconfig(gt); 729 if (err) 730 gt_err(gt, "Failed to retrieve hwconfig table: %pe\n", ERR_PTR(err)); 731 732 err = __engines_record_defaults(gt); 733 if (err) 734 goto err_gt; 735 736 err = __engines_verify_workarounds(gt); 737 if (err) 738 goto err_gt; 739 740 err = i915_inject_probe_error(gt->i915, -EIO); 741 if (err) 742 goto err_gt; 743 744 intel_uc_init_late(>->uc); 745 746 intel_migrate_init(>->migrate, gt); 747 748 goto out_fw; 749 err_gt: 750 __intel_gt_disable(gt); 751 intel_uc_fini_hw(>->uc); 752 err_uc_init: 753 intel_uc_fini(>->uc); 754 err_engines: 755 intel_engines_release(gt); 756 i915_vm_put(fetch_and_zero(>->vm)); 757 err_pm: 758 intel_gt_pm_fini(gt); 759 intel_gt_fini_scratch(gt); 760 out_fw: 761 if (err) 762 intel_gt_set_wedged_on_init(gt); 763 intel_uncore_forcewake_put(gt->uncore, FORCEWAKE_ALL); 764 return err; 765 } 766 767 void intel_gt_driver_remove(struct intel_gt *gt) 768 { 769 __intel_gt_disable(gt); 770 771 intel_migrate_fini(>->migrate); 772 intel_uc_driver_remove(>->uc); 773 774 intel_engines_release(gt); 775 776 intel_gt_flush_buffer_pool(gt); 777 } 778 779 void intel_gt_driver_unregister(struct intel_gt *gt) 780 { 781 intel_wakeref_t wakeref; 782 783 intel_gt_sysfs_unregister(gt); 784 intel_rps_driver_unregister(>->rps); 785 intel_gsc_fini(>->gsc); 786 787 /* 788 * Upon unregistering the device to prevent any new users, cancel 789 * all in-flight requests so that we can quickly unbind the active 790 * resources. 791 */ 792 intel_gt_set_wedged_on_fini(gt); 793 794 /* Scrub all HW state upon release */ 795 with_intel_runtime_pm(gt->uncore->rpm, wakeref) 796 __intel_gt_reset(gt, ALL_ENGINES); 797 } 798 799 void intel_gt_driver_release(struct intel_gt *gt) 800 { 801 struct i915_address_space *vm; 802 803 vm = fetch_and_zero(>->vm); 804 if (vm) /* FIXME being called twice on error paths :( */ 805 i915_vm_put(vm); 806 807 intel_wa_list_free(>->wa_list); 808 intel_gt_pm_fini(gt); 809 intel_gt_fini_scratch(gt); 810 intel_gt_fini_buffer_pool(gt); 811 intel_gt_fini_hwconfig(gt); 812 } 813 814 void intel_gt_driver_late_release_all(struct drm_i915_private *i915) 815 { 816 struct intel_gt *gt; 817 unsigned int id; 818 819 /* We need to wait for inflight RCU frees to release their grip */ 820 rcu_barrier(); 821 822 for_each_gt(gt, i915, id) { 823 intel_uc_driver_late_release(>->uc); 824 intel_gt_fini_requests(gt); 825 intel_gt_fini_reset(gt); 826 intel_gt_fini_timelines(gt); 827 mutex_destroy(>->tlb.invalidate_lock); 828 intel_engines_free(gt); 829 } 830 } 831 832 static int intel_gt_tile_setup(struct intel_gt *gt, phys_addr_t phys_addr) 833 { 834 int ret; 835 836 if (!gt_is_root(gt)) { 837 struct intel_uncore *uncore; 838 spinlock_t *irq_lock; 839 840 uncore = drmm_kzalloc(>->i915->drm, sizeof(*uncore), GFP_KERNEL); 841 if (!uncore) 842 return -ENOMEM; 843 844 irq_lock = drmm_kzalloc(>->i915->drm, sizeof(*irq_lock), GFP_KERNEL); 845 if (!irq_lock) 846 return -ENOMEM; 847 848 gt->uncore = uncore; 849 gt->irq_lock = irq_lock; 850 851 intel_gt_common_init_early(gt); 852 } 853 854 intel_uncore_init_early(gt->uncore, gt); 855 856 ret = intel_uncore_setup_mmio(gt->uncore, phys_addr); 857 if (ret) 858 return ret; 859 860 gt->phys_addr = phys_addr; 861 862 return 0; 863 } 864 865 int intel_gt_probe_all(struct drm_i915_private *i915) 866 { 867 struct pci_dev *pdev = to_pci_dev(i915->drm.dev); 868 struct intel_gt *gt = &i915->gt0; 869 const struct intel_gt_definition *gtdef; 870 phys_addr_t phys_addr; 871 unsigned int mmio_bar; 872 unsigned int i; 873 int ret; 874 875 mmio_bar = intel_mmio_bar(GRAPHICS_VER(i915)); 876 phys_addr = pci_resource_start(pdev, mmio_bar); 877 878 /* 879 * We always have at least one primary GT on any device 880 * and it has been already initialized early during probe 881 * in i915_driver_probe() 882 */ 883 gt->i915 = i915; 884 gt->name = "Primary GT"; 885 gt->info.engine_mask = RUNTIME_INFO(i915)->platform_engine_mask; 886 887 gt_dbg(gt, "Setting up %s\n", gt->name); 888 ret = intel_gt_tile_setup(gt, phys_addr); 889 if (ret) 890 return ret; 891 892 i915->gt[0] = gt; 893 894 if (!HAS_EXTRA_GT_LIST(i915)) 895 return 0; 896 897 for (i = 1, gtdef = &INTEL_INFO(i915)->extra_gt_list[i - 1]; 898 gtdef->name != NULL; 899 i++, gtdef = &INTEL_INFO(i915)->extra_gt_list[i - 1]) { 900 gt = drmm_kzalloc(&i915->drm, sizeof(*gt), GFP_KERNEL); 901 if (!gt) { 902 ret = -ENOMEM; 903 goto err; 904 } 905 906 gt->i915 = i915; 907 gt->name = gtdef->name; 908 gt->type = gtdef->type; 909 gt->info.engine_mask = gtdef->engine_mask; 910 gt->info.id = i; 911 912 gt_dbg(gt, "Setting up %s\n", gt->name); 913 if (GEM_WARN_ON(range_overflows_t(resource_size_t, 914 gtdef->mapping_base, 915 SZ_16M, 916 pci_resource_len(pdev, mmio_bar)))) { 917 ret = -ENODEV; 918 goto err; 919 } 920 921 switch (gtdef->type) { 922 case GT_TILE: 923 ret = intel_gt_tile_setup(gt, phys_addr + gtdef->mapping_base); 924 break; 925 926 case GT_MEDIA: 927 ret = intel_sa_mediagt_setup(gt, phys_addr + gtdef->mapping_base, 928 gtdef->gsi_offset); 929 break; 930 931 case GT_PRIMARY: 932 /* Primary GT should not appear in extra GT list */ 933 default: 934 MISSING_CASE(gtdef->type); 935 ret = -ENODEV; 936 } 937 938 if (ret) 939 goto err; 940 941 i915->gt[i] = gt; 942 } 943 944 return 0; 945 946 err: 947 i915_probe_error(i915, "Failed to initialize %s! (%d)\n", gtdef->name, ret); 948 intel_gt_release_all(i915); 949 950 return ret; 951 } 952 953 int intel_gt_tiles_init(struct drm_i915_private *i915) 954 { 955 struct intel_gt *gt; 956 unsigned int id; 957 int ret; 958 959 for_each_gt(gt, i915, id) { 960 ret = intel_gt_probe_lmem(gt); 961 if (ret) 962 return ret; 963 } 964 965 return 0; 966 } 967 968 void intel_gt_release_all(struct drm_i915_private *i915) 969 { 970 struct intel_gt *gt; 971 unsigned int id; 972 973 for_each_gt(gt, i915, id) 974 i915->gt[id] = NULL; 975 } 976 977 void intel_gt_info_print(const struct intel_gt_info *info, 978 struct drm_printer *p) 979 { 980 drm_printf(p, "available engines: %x\n", info->engine_mask); 981 982 intel_sseu_dump(&info->sseu, p); 983 } 984 985 struct reg_and_bit { 986 union { 987 i915_reg_t reg; 988 i915_mcr_reg_t mcr_reg; 989 }; 990 u32 bit; 991 }; 992 993 static struct reg_and_bit 994 get_reg_and_bit(const struct intel_engine_cs *engine, const bool gen8, 995 const i915_reg_t *regs, const unsigned int num) 996 { 997 const unsigned int class = engine->class; 998 struct reg_and_bit rb = { }; 999 1000 if (gt_WARN_ON_ONCE(engine->gt, class >= num || !regs[class].reg)) 1001 return rb; 1002 1003 rb.reg = regs[class]; 1004 if (gen8 && class == VIDEO_DECODE_CLASS) 1005 rb.reg.reg += 4 * engine->instance; /* GEN8_M2TCR */ 1006 else 1007 rb.bit = engine->instance; 1008 1009 rb.bit = BIT(rb.bit); 1010 1011 return rb; 1012 } 1013 1014 /* 1015 * HW architecture suggest typical invalidation time at 40us, 1016 * with pessimistic cases up to 100us and a recommendation to 1017 * cap at 1ms. We go a bit higher just in case. 1018 */ 1019 #define TLB_INVAL_TIMEOUT_US 100 1020 #define TLB_INVAL_TIMEOUT_MS 4 1021 1022 /* 1023 * On Xe_HP the TLB invalidation registers are located at the same MMIO offsets 1024 * but are now considered MCR registers. Since they exist within a GAM range, 1025 * the primary instance of the register rolls up the status from each unit. 1026 */ 1027 static int wait_for_invalidate(struct intel_gt *gt, struct reg_and_bit rb) 1028 { 1029 if (GRAPHICS_VER_FULL(gt->i915) >= IP_VER(12, 50)) 1030 return intel_gt_mcr_wait_for_reg(gt, rb.mcr_reg, rb.bit, 0, 1031 TLB_INVAL_TIMEOUT_US, 1032 TLB_INVAL_TIMEOUT_MS); 1033 else 1034 return __intel_wait_for_register_fw(gt->uncore, rb.reg, rb.bit, 0, 1035 TLB_INVAL_TIMEOUT_US, 1036 TLB_INVAL_TIMEOUT_MS, 1037 NULL); 1038 } 1039 1040 static void mmio_invalidate_full(struct intel_gt *gt) 1041 { 1042 static const i915_reg_t gen8_regs[] = { 1043 [RENDER_CLASS] = GEN8_RTCR, 1044 [VIDEO_DECODE_CLASS] = GEN8_M1TCR, /* , GEN8_M2TCR */ 1045 [VIDEO_ENHANCEMENT_CLASS] = GEN8_VTCR, 1046 [COPY_ENGINE_CLASS] = GEN8_BTCR, 1047 }; 1048 static const i915_reg_t gen12_regs[] = { 1049 [RENDER_CLASS] = GEN12_GFX_TLB_INV_CR, 1050 [VIDEO_DECODE_CLASS] = GEN12_VD_TLB_INV_CR, 1051 [VIDEO_ENHANCEMENT_CLASS] = GEN12_VE_TLB_INV_CR, 1052 [COPY_ENGINE_CLASS] = GEN12_BLT_TLB_INV_CR, 1053 [COMPUTE_CLASS] = GEN12_COMPCTX_TLB_INV_CR, 1054 }; 1055 static const i915_mcr_reg_t xehp_regs[] = { 1056 [RENDER_CLASS] = XEHP_GFX_TLB_INV_CR, 1057 [VIDEO_DECODE_CLASS] = XEHP_VD_TLB_INV_CR, 1058 [VIDEO_ENHANCEMENT_CLASS] = XEHP_VE_TLB_INV_CR, 1059 [COPY_ENGINE_CLASS] = XEHP_BLT_TLB_INV_CR, 1060 [COMPUTE_CLASS] = XEHP_COMPCTX_TLB_INV_CR, 1061 }; 1062 struct drm_i915_private *i915 = gt->i915; 1063 struct intel_uncore *uncore = gt->uncore; 1064 struct intel_engine_cs *engine; 1065 intel_engine_mask_t awake, tmp; 1066 enum intel_engine_id id; 1067 const i915_reg_t *regs; 1068 unsigned int num = 0; 1069 unsigned long flags; 1070 1071 /* 1072 * New platforms should not be added with catch-all-newer (>=) 1073 * condition so that any later platform added triggers the below warning 1074 * and in turn mandates a human cross-check of whether the invalidation 1075 * flows have compatible semantics. 1076 * 1077 * For instance with the 11.00 -> 12.00 transition three out of five 1078 * respective engine registers were moved to masked type. Then after the 1079 * 12.00 -> 12.50 transition multi cast handling is required too. 1080 */ 1081 1082 if (GRAPHICS_VER_FULL(i915) == IP_VER(12, 50) || 1083 GRAPHICS_VER_FULL(i915) == IP_VER(12, 55)) { 1084 regs = NULL; 1085 num = ARRAY_SIZE(xehp_regs); 1086 } else if (GRAPHICS_VER_FULL(i915) == IP_VER(12, 0) || 1087 GRAPHICS_VER_FULL(i915) == IP_VER(12, 10)) { 1088 regs = gen12_regs; 1089 num = ARRAY_SIZE(gen12_regs); 1090 } else if (GRAPHICS_VER(i915) >= 8 && GRAPHICS_VER(i915) <= 11) { 1091 regs = gen8_regs; 1092 num = ARRAY_SIZE(gen8_regs); 1093 } else if (GRAPHICS_VER(i915) < 8) { 1094 return; 1095 } 1096 1097 if (gt_WARN_ONCE(gt, !num, "Platform does not implement TLB invalidation!")) 1098 return; 1099 1100 intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL); 1101 1102 intel_gt_mcr_lock(gt, &flags); 1103 spin_lock(&uncore->lock); /* serialise invalidate with GT reset */ 1104 1105 awake = 0; 1106 for_each_engine(engine, gt, id) { 1107 struct reg_and_bit rb; 1108 1109 if (!intel_engine_pm_is_awake(engine)) 1110 continue; 1111 1112 if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) { 1113 u32 val = BIT(engine->instance); 1114 1115 if (engine->class == VIDEO_DECODE_CLASS || 1116 engine->class == VIDEO_ENHANCEMENT_CLASS || 1117 engine->class == COMPUTE_CLASS) 1118 val = _MASKED_BIT_ENABLE(val); 1119 intel_gt_mcr_multicast_write_fw(gt, 1120 xehp_regs[engine->class], 1121 val); 1122 } else { 1123 rb = get_reg_and_bit(engine, regs == gen8_regs, regs, num); 1124 if (!i915_mmio_reg_offset(rb.reg)) 1125 continue; 1126 1127 if (GRAPHICS_VER(i915) == 12 && (engine->class == VIDEO_DECODE_CLASS || 1128 engine->class == VIDEO_ENHANCEMENT_CLASS || 1129 engine->class == COMPUTE_CLASS)) 1130 rb.bit = _MASKED_BIT_ENABLE(rb.bit); 1131 1132 intel_uncore_write_fw(uncore, rb.reg, rb.bit); 1133 } 1134 awake |= engine->mask; 1135 } 1136 1137 GT_TRACE(gt, "invalidated engines %08x\n", awake); 1138 1139 /* Wa_2207587034:tgl,dg1,rkl,adl-s,adl-p */ 1140 if (awake && 1141 (IS_TIGERLAKE(i915) || 1142 IS_DG1(i915) || 1143 IS_ROCKETLAKE(i915) || 1144 IS_ALDERLAKE_S(i915) || 1145 IS_ALDERLAKE_P(i915))) 1146 intel_uncore_write_fw(uncore, GEN12_OA_TLB_INV_CR, 1); 1147 1148 spin_unlock(&uncore->lock); 1149 intel_gt_mcr_unlock(gt, flags); 1150 1151 for_each_engine_masked(engine, gt, awake, tmp) { 1152 struct reg_and_bit rb; 1153 1154 if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) { 1155 rb.mcr_reg = xehp_regs[engine->class]; 1156 rb.bit = BIT(engine->instance); 1157 } else { 1158 rb = get_reg_and_bit(engine, regs == gen8_regs, regs, num); 1159 } 1160 1161 if (wait_for_invalidate(gt, rb)) 1162 gt_err_ratelimited(gt, "%s TLB invalidation did not complete in %ums!\n", 1163 engine->name, TLB_INVAL_TIMEOUT_MS); 1164 } 1165 1166 /* 1167 * Use delayed put since a) we mostly expect a flurry of TLB 1168 * invalidations so it is good to avoid paying the forcewake cost and 1169 * b) it works around a bug in Icelake which cannot cope with too rapid 1170 * transitions. 1171 */ 1172 intel_uncore_forcewake_put_delayed(uncore, FORCEWAKE_ALL); 1173 } 1174 1175 static bool tlb_seqno_passed(const struct intel_gt *gt, u32 seqno) 1176 { 1177 u32 cur = intel_gt_tlb_seqno(gt); 1178 1179 /* Only skip if a *full* TLB invalidate barrier has passed */ 1180 return (s32)(cur - ALIGN(seqno, 2)) > 0; 1181 } 1182 1183 void intel_gt_invalidate_tlb(struct intel_gt *gt, u32 seqno) 1184 { 1185 intel_wakeref_t wakeref; 1186 1187 if (I915_SELFTEST_ONLY(gt->awake == -ENODEV)) 1188 return; 1189 1190 if (intel_gt_is_wedged(gt)) 1191 return; 1192 1193 if (tlb_seqno_passed(gt, seqno)) 1194 return; 1195 1196 with_intel_gt_pm_if_awake(gt, wakeref) { 1197 mutex_lock(>->tlb.invalidate_lock); 1198 if (tlb_seqno_passed(gt, seqno)) 1199 goto unlock; 1200 1201 mmio_invalidate_full(gt); 1202 1203 write_seqcount_invalidate(>->tlb.seqno); 1204 unlock: 1205 mutex_unlock(>->tlb.invalidate_lock); 1206 } 1207 } 1208