1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2022 Intel Corporation 4 */ 5 6 #include "xe_gt.h" 7 8 #include <linux/minmax.h> 9 10 #include <drm/drm_managed.h> 11 #include <uapi/drm/xe_drm.h> 12 13 #include <generated/xe_wa_oob.h> 14 15 #include "instructions/xe_alu_commands.h" 16 #include "instructions/xe_gfxpipe_commands.h" 17 #include "instructions/xe_mi_commands.h" 18 #include "regs/xe_engine_regs.h" 19 #include "regs/xe_gt_regs.h" 20 #include "xe_assert.h" 21 #include "xe_bb.h" 22 #include "xe_bo.h" 23 #include "xe_device.h" 24 #include "xe_eu_stall.h" 25 #include "xe_exec_queue.h" 26 #include "xe_execlist.h" 27 #include "xe_force_wake.h" 28 #include "xe_ggtt.h" 29 #include "xe_gsc.h" 30 #include "xe_gt_ccs_mode.h" 31 #include "xe_gt_clock.h" 32 #include "xe_gt_freq.h" 33 #include "xe_gt_idle.h" 34 #include "xe_gt_mcr.h" 35 #include "xe_gt_pagefault.h" 36 #include "xe_gt_printk.h" 37 #include "xe_gt_sriov_pf.h" 38 #include "xe_gt_sriov_vf.h" 39 #include "xe_gt_sysfs.h" 40 #include "xe_gt_tlb_invalidation.h" 41 #include "xe_gt_topology.h" 42 #include "xe_guc_exec_queue_types.h" 43 #include "xe_guc_pc.h" 44 #include "xe_guc_submit.h" 45 #include "xe_hw_fence.h" 46 #include "xe_hw_engine_class_sysfs.h" 47 #include "xe_irq.h" 48 #include "xe_lmtt.h" 49 #include "xe_lrc.h" 50 #include "xe_map.h" 51 #include "xe_migrate.h" 52 #include "xe_mmio.h" 53 #include "xe_pat.h" 54 #include "xe_pm.h" 55 #include "xe_mocs.h" 56 #include "xe_reg_sr.h" 57 #include "xe_ring_ops.h" 58 #include "xe_sa.h" 59 #include "xe_sched_job.h" 60 #include "xe_sriov.h" 61 #include "xe_tuning.h" 62 #include "xe_uc.h" 63 #include "xe_uc_fw.h" 64 #include "xe_vm.h" 65 #include "xe_wa.h" 66 #include "xe_wopcm.h" 67 68 static void gt_fini(struct drm_device *drm, void *arg) 69 { 70 struct xe_gt *gt = arg; 71 72 destroy_workqueue(gt->ordered_wq); 73 } 74 75 struct xe_gt *xe_gt_alloc(struct xe_tile *tile) 76 { 77 struct xe_gt *gt; 78 int err; 79 80 gt = drmm_kzalloc(&tile_to_xe(tile)->drm, sizeof(*gt), GFP_KERNEL); 81 if (!gt) 82 return ERR_PTR(-ENOMEM); 83 84 gt->tile = tile; 85 gt->ordered_wq = alloc_ordered_workqueue("gt-ordered-wq", 86 WQ_MEM_RECLAIM); 87 88 err = drmm_add_action_or_reset(>_to_xe(gt)->drm, gt_fini, gt); 89 if (err) 90 return ERR_PTR(err); 91 92 return gt; 93 } 94 95 void xe_gt_sanitize(struct xe_gt *gt) 96 { 97 /* 98 * FIXME: if xe_uc_sanitize is called here, on TGL driver will not 99 * reload 100 */ 101 xe_guc_submit_disable(>->uc.guc); 102 } 103 104 static void xe_gt_enable_host_l2_vram(struct xe_gt *gt) 105 { 106 unsigned int fw_ref; 107 u32 reg; 108 109 if (!XE_WA(gt, 16023588340)) 110 return; 111 112 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT); 113 if (!fw_ref) 114 return; 115 116 if (xe_gt_is_main_type(gt)) { 117 reg = xe_gt_mcr_unicast_read_any(gt, XE2_GAMREQSTRM_CTRL); 118 reg |= CG_DIS_CNTLBUS; 119 xe_gt_mcr_multicast_write(gt, XE2_GAMREQSTRM_CTRL, reg); 120 } 121 122 xe_gt_mcr_multicast_write(gt, XEHPC_L3CLOS_MASK(3), 0xF); 123 xe_force_wake_put(gt_to_fw(gt), fw_ref); 124 } 125 126 static void xe_gt_disable_host_l2_vram(struct xe_gt *gt) 127 { 128 unsigned int fw_ref; 129 u32 reg; 130 131 if (!XE_WA(gt, 16023588340)) 132 return; 133 134 if (xe_gt_is_media_type(gt)) 135 return; 136 137 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT); 138 if (!fw_ref) 139 return; 140 141 reg = xe_gt_mcr_unicast_read_any(gt, XE2_GAMREQSTRM_CTRL); 142 reg &= ~CG_DIS_CNTLBUS; 143 xe_gt_mcr_multicast_write(gt, XE2_GAMREQSTRM_CTRL, reg); 144 145 xe_force_wake_put(gt_to_fw(gt), fw_ref); 146 } 147 148 static void gt_reset_worker(struct work_struct *w); 149 150 static int emit_job_sync(struct xe_exec_queue *q, struct xe_bb *bb, 151 long timeout_jiffies) 152 { 153 struct xe_sched_job *job; 154 struct dma_fence *fence; 155 long timeout; 156 157 job = xe_bb_create_job(q, bb); 158 if (IS_ERR(job)) 159 return PTR_ERR(job); 160 161 xe_sched_job_arm(job); 162 fence = dma_fence_get(&job->drm.s_fence->finished); 163 xe_sched_job_push(job); 164 165 timeout = dma_fence_wait_timeout(fence, false, timeout_jiffies); 166 dma_fence_put(fence); 167 if (timeout < 0) 168 return timeout; 169 else if (!timeout) 170 return -ETIME; 171 172 return 0; 173 } 174 175 static int emit_nop_job(struct xe_gt *gt, struct xe_exec_queue *q) 176 { 177 struct xe_bb *bb; 178 int ret; 179 180 bb = xe_bb_new(gt, 4, false); 181 if (IS_ERR(bb)) 182 return PTR_ERR(bb); 183 184 ret = emit_job_sync(q, bb, HZ); 185 xe_bb_free(bb, NULL); 186 187 return ret; 188 } 189 190 static int emit_wa_job(struct xe_gt *gt, struct xe_exec_queue *q) 191 { 192 struct xe_reg_sr *sr = &q->hwe->reg_lrc; 193 struct xe_reg_sr_entry *entry; 194 int count_rmw = 0, count = 0, ret; 195 unsigned long idx; 196 struct xe_bb *bb; 197 size_t bb_len = 0; 198 u32 *cs; 199 200 /* count RMW registers as those will be handled separately */ 201 xa_for_each(&sr->xa, idx, entry) { 202 if (entry->reg.masked || entry->clr_bits == ~0) 203 ++count; 204 else 205 ++count_rmw; 206 } 207 208 if (count) 209 bb_len += count * 2 + 1; 210 211 if (count_rmw) 212 bb_len += count_rmw * 20 + 7; 213 214 if (q->hwe->class == XE_ENGINE_CLASS_RENDER) 215 /* 216 * Big enough to emit all of the context's 3DSTATE via 217 * xe_lrc_emit_hwe_state_instructions() 218 */ 219 bb_len += xe_gt_lrc_size(gt, q->hwe->class) / sizeof(u32); 220 221 xe_gt_dbg(gt, "LRC %s WA job: %zu dwords\n", q->hwe->name, bb_len); 222 223 bb = xe_bb_new(gt, bb_len, false); 224 if (IS_ERR(bb)) 225 return PTR_ERR(bb); 226 227 cs = bb->cs; 228 229 if (count) { 230 /* 231 * Emit single LRI with all non RMW regs: 1 leading dw + 2dw per 232 * reg + 1 233 */ 234 235 *cs++ = MI_LOAD_REGISTER_IMM | MI_LRI_NUM_REGS(count); 236 237 xa_for_each(&sr->xa, idx, entry) { 238 struct xe_reg reg = entry->reg; 239 u32 val; 240 241 if (reg.masked) 242 val = entry->clr_bits << 16; 243 else if (entry->clr_bits == ~0) 244 val = 0; 245 else 246 continue; 247 248 val |= entry->set_bits; 249 250 *cs++ = reg.addr; 251 *cs++ = val; 252 xe_gt_dbg(gt, "REG[0x%x] = 0x%08x", reg.addr, val); 253 } 254 } 255 256 if (count_rmw) { 257 /* Emit MI_MATH for each RMW reg: 20dw per reg + 7 trailing dw */ 258 259 xa_for_each(&sr->xa, idx, entry) { 260 if (entry->reg.masked || entry->clr_bits == ~0) 261 continue; 262 263 *cs++ = MI_LOAD_REGISTER_REG | MI_LRR_DST_CS_MMIO; 264 *cs++ = entry->reg.addr; 265 *cs++ = CS_GPR_REG(0, 0).addr; 266 267 *cs++ = MI_LOAD_REGISTER_IMM | MI_LRI_NUM_REGS(2) | 268 MI_LRI_LRM_CS_MMIO; 269 *cs++ = CS_GPR_REG(0, 1).addr; 270 *cs++ = entry->clr_bits; 271 *cs++ = CS_GPR_REG(0, 2).addr; 272 *cs++ = entry->set_bits; 273 274 *cs++ = MI_MATH(8); 275 *cs++ = CS_ALU_INSTR_LOAD(SRCA, REG0); 276 *cs++ = CS_ALU_INSTR_LOADINV(SRCB, REG1); 277 *cs++ = CS_ALU_INSTR_AND; 278 *cs++ = CS_ALU_INSTR_STORE(REG0, ACCU); 279 *cs++ = CS_ALU_INSTR_LOAD(SRCA, REG0); 280 *cs++ = CS_ALU_INSTR_LOAD(SRCB, REG2); 281 *cs++ = CS_ALU_INSTR_OR; 282 *cs++ = CS_ALU_INSTR_STORE(REG0, ACCU); 283 284 *cs++ = MI_LOAD_REGISTER_REG | MI_LRR_SRC_CS_MMIO; 285 *cs++ = CS_GPR_REG(0, 0).addr; 286 *cs++ = entry->reg.addr; 287 288 xe_gt_dbg(gt, "REG[%#x] = ~%#x|%#x\n", 289 entry->reg.addr, entry->clr_bits, entry->set_bits); 290 } 291 292 /* reset used GPR */ 293 *cs++ = MI_LOAD_REGISTER_IMM | MI_LRI_NUM_REGS(3) | 294 MI_LRI_LRM_CS_MMIO; 295 *cs++ = CS_GPR_REG(0, 0).addr; 296 *cs++ = 0; 297 *cs++ = CS_GPR_REG(0, 1).addr; 298 *cs++ = 0; 299 *cs++ = CS_GPR_REG(0, 2).addr; 300 *cs++ = 0; 301 } 302 303 cs = xe_lrc_emit_hwe_state_instructions(q, cs); 304 305 bb->len = cs - bb->cs; 306 307 ret = emit_job_sync(q, bb, HZ); 308 309 xe_bb_free(bb, NULL); 310 311 return ret; 312 } 313 314 int xe_gt_record_default_lrcs(struct xe_gt *gt) 315 { 316 struct xe_device *xe = gt_to_xe(gt); 317 struct xe_hw_engine *hwe; 318 enum xe_hw_engine_id id; 319 int err = 0; 320 321 for_each_hw_engine(hwe, gt, id) { 322 struct xe_exec_queue *q, *nop_q; 323 void *default_lrc; 324 325 if (gt->default_lrc[hwe->class]) 326 continue; 327 328 xe_reg_sr_init(&hwe->reg_lrc, hwe->name, xe); 329 xe_wa_process_lrc(hwe); 330 xe_hw_engine_setup_default_lrc_state(hwe); 331 xe_tuning_process_lrc(hwe); 332 333 default_lrc = drmm_kzalloc(&xe->drm, 334 xe_gt_lrc_size(gt, hwe->class), 335 GFP_KERNEL); 336 if (!default_lrc) 337 return -ENOMEM; 338 339 q = xe_exec_queue_create(xe, NULL, BIT(hwe->logical_instance), 1, 340 hwe, EXEC_QUEUE_FLAG_KERNEL, 0); 341 if (IS_ERR(q)) { 342 err = PTR_ERR(q); 343 xe_gt_err(gt, "hwe %s: xe_exec_queue_create failed (%pe)\n", 344 hwe->name, q); 345 return err; 346 } 347 348 /* Prime golden LRC with known good state */ 349 err = emit_wa_job(gt, q); 350 if (err) { 351 xe_gt_err(gt, "hwe %s: emit_wa_job failed (%pe) guc_id=%u\n", 352 hwe->name, ERR_PTR(err), q->guc->id); 353 goto put_exec_queue; 354 } 355 356 nop_q = xe_exec_queue_create(xe, NULL, BIT(hwe->logical_instance), 357 1, hwe, EXEC_QUEUE_FLAG_KERNEL, 0); 358 if (IS_ERR(nop_q)) { 359 err = PTR_ERR(nop_q); 360 xe_gt_err(gt, "hwe %s: nop xe_exec_queue_create failed (%pe)\n", 361 hwe->name, nop_q); 362 goto put_exec_queue; 363 } 364 365 /* Switch to different LRC */ 366 err = emit_nop_job(gt, nop_q); 367 if (err) { 368 xe_gt_err(gt, "hwe %s: nop emit_nop_job failed (%pe) guc_id=%u\n", 369 hwe->name, ERR_PTR(err), nop_q->guc->id); 370 goto put_nop_q; 371 } 372 373 xe_map_memcpy_from(xe, default_lrc, 374 &q->lrc[0]->bo->vmap, 375 xe_lrc_pphwsp_offset(q->lrc[0]), 376 xe_gt_lrc_size(gt, hwe->class)); 377 378 gt->default_lrc[hwe->class] = default_lrc; 379 put_nop_q: 380 xe_exec_queue_put(nop_q); 381 put_exec_queue: 382 xe_exec_queue_put(q); 383 if (err) 384 break; 385 } 386 387 return err; 388 } 389 390 int xe_gt_init_early(struct xe_gt *gt) 391 { 392 unsigned int fw_ref; 393 int err; 394 395 if (IS_SRIOV_PF(gt_to_xe(gt))) { 396 err = xe_gt_sriov_pf_init_early(gt); 397 if (err) 398 return err; 399 } 400 401 xe_reg_sr_init(>->reg_sr, "GT", gt_to_xe(gt)); 402 403 err = xe_wa_init(gt); 404 if (err) 405 return err; 406 407 err = xe_tuning_init(gt); 408 if (err) 409 return err; 410 411 xe_wa_process_oob(gt); 412 413 xe_force_wake_init_gt(gt, gt_to_fw(gt)); 414 spin_lock_init(>->global_invl_lock); 415 416 err = xe_gt_tlb_invalidation_init_early(gt); 417 if (err) 418 return err; 419 420 xe_mocs_init_early(gt); 421 422 /* 423 * Only after this point can GT-specific MMIO operations 424 * (including things like communication with the GuC) 425 * be performed. 426 */ 427 xe_gt_mmio_init(gt); 428 429 err = xe_uc_init_noalloc(>->uc); 430 if (err) 431 return err; 432 433 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT); 434 if (!fw_ref) 435 return -ETIMEDOUT; 436 437 xe_gt_mcr_init_early(gt); 438 xe_pat_init(gt); 439 xe_force_wake_put(gt_to_fw(gt), fw_ref); 440 441 return 0; 442 } 443 444 static void dump_pat_on_error(struct xe_gt *gt) 445 { 446 struct drm_printer p; 447 char prefix[32]; 448 449 snprintf(prefix, sizeof(prefix), "[GT%u Error]", gt->info.id); 450 p = drm_dbg_printer(>_to_xe(gt)->drm, DRM_UT_DRIVER, prefix); 451 452 xe_pat_dump(gt, &p); 453 } 454 455 static int gt_init_with_gt_forcewake(struct xe_gt *gt) 456 { 457 unsigned int fw_ref; 458 int err; 459 460 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT); 461 if (!fw_ref) 462 return -ETIMEDOUT; 463 464 err = xe_uc_init(>->uc); 465 if (err) 466 goto err_force_wake; 467 468 xe_gt_topology_init(gt); 469 xe_gt_mcr_init(gt); 470 xe_gt_enable_host_l2_vram(gt); 471 472 if (xe_gt_is_main_type(gt)) { 473 err = xe_ggtt_init(gt_to_tile(gt)->mem.ggtt); 474 if (err) 475 goto err_force_wake; 476 if (IS_SRIOV_PF(gt_to_xe(gt))) 477 xe_lmtt_init(>_to_tile(gt)->sriov.pf.lmtt); 478 } 479 480 /* Enable per hw engine IRQs */ 481 xe_irq_enable_hwe(gt); 482 483 /* Rerun MCR init as we now have hw engine list */ 484 xe_gt_mcr_init(gt); 485 486 err = xe_hw_engines_init_early(gt); 487 if (err) { 488 dump_pat_on_error(gt); 489 goto err_force_wake; 490 } 491 492 err = xe_hw_engine_class_sysfs_init(gt); 493 if (err) 494 goto err_force_wake; 495 496 /* Initialize CCS mode sysfs after early initialization of HW engines */ 497 err = xe_gt_ccs_mode_sysfs_init(gt); 498 if (err) 499 goto err_force_wake; 500 501 /* 502 * Stash hardware-reported version. Since this register does not exist 503 * on pre-MTL platforms, reading it there will (correctly) return 0. 504 */ 505 gt->info.gmdid = xe_mmio_read32(>->mmio, GMD_ID); 506 507 xe_force_wake_put(gt_to_fw(gt), fw_ref); 508 return 0; 509 510 err_force_wake: 511 xe_force_wake_put(gt_to_fw(gt), fw_ref); 512 513 return err; 514 } 515 516 static int gt_init_with_all_forcewake(struct xe_gt *gt) 517 { 518 unsigned int fw_ref; 519 int err; 520 521 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FORCEWAKE_ALL); 522 if (!xe_force_wake_ref_has_domain(fw_ref, XE_FORCEWAKE_ALL)) { 523 err = -ETIMEDOUT; 524 goto err_force_wake; 525 } 526 527 xe_gt_mcr_set_implicit_defaults(gt); 528 xe_wa_process_gt(gt); 529 xe_tuning_process_gt(gt); 530 xe_reg_sr_apply_mmio(>->reg_sr, gt); 531 532 err = xe_gt_clock_init(gt); 533 if (err) 534 goto err_force_wake; 535 536 xe_mocs_init(gt); 537 err = xe_execlist_init(gt); 538 if (err) 539 goto err_force_wake; 540 541 err = xe_hw_engines_init(gt); 542 if (err) 543 goto err_force_wake; 544 545 err = xe_uc_init_post_hwconfig(>->uc); 546 if (err) 547 goto err_force_wake; 548 549 if (xe_gt_is_main_type(gt)) { 550 /* 551 * USM has its only SA pool to non-block behind user operations 552 */ 553 if (gt_to_xe(gt)->info.has_usm) { 554 struct xe_device *xe = gt_to_xe(gt); 555 556 gt->usm.bb_pool = xe_sa_bo_manager_init(gt_to_tile(gt), 557 IS_DGFX(xe) ? SZ_1M : SZ_512K, 16); 558 if (IS_ERR(gt->usm.bb_pool)) { 559 err = PTR_ERR(gt->usm.bb_pool); 560 goto err_force_wake; 561 } 562 } 563 } 564 565 if (xe_gt_is_main_type(gt)) { 566 struct xe_tile *tile = gt_to_tile(gt); 567 568 tile->migrate = xe_migrate_init(tile); 569 if (IS_ERR(tile->migrate)) { 570 err = PTR_ERR(tile->migrate); 571 goto err_force_wake; 572 } 573 } 574 575 err = xe_uc_load_hw(>->uc); 576 if (err) 577 goto err_force_wake; 578 579 /* Configure default CCS mode of 1 engine with all resources */ 580 if (xe_gt_ccs_mode_enabled(gt)) { 581 gt->ccs_mode = 1; 582 xe_gt_apply_ccs_mode(gt); 583 } 584 585 if (IS_SRIOV_PF(gt_to_xe(gt)) && xe_gt_is_main_type(gt)) 586 xe_lmtt_init_hw(>_to_tile(gt)->sriov.pf.lmtt); 587 588 if (IS_SRIOV_PF(gt_to_xe(gt))) { 589 xe_gt_sriov_pf_init(gt); 590 xe_gt_sriov_pf_init_hw(gt); 591 } 592 593 xe_force_wake_put(gt_to_fw(gt), fw_ref); 594 595 return 0; 596 597 err_force_wake: 598 xe_force_wake_put(gt_to_fw(gt), fw_ref); 599 600 return err; 601 } 602 603 static void xe_gt_fini(void *arg) 604 { 605 struct xe_gt *gt = arg; 606 int i; 607 608 for (i = 0; i < XE_ENGINE_CLASS_MAX; ++i) 609 xe_hw_fence_irq_finish(>->fence_irq[i]); 610 611 xe_gt_disable_host_l2_vram(gt); 612 } 613 614 int xe_gt_init(struct xe_gt *gt) 615 { 616 int err; 617 int i; 618 619 INIT_WORK(>->reset.worker, gt_reset_worker); 620 621 for (i = 0; i < XE_ENGINE_CLASS_MAX; ++i) { 622 gt->ring_ops[i] = xe_ring_ops_get(gt, i); 623 xe_hw_fence_irq_init(>->fence_irq[i]); 624 } 625 626 err = devm_add_action_or_reset(gt_to_xe(gt)->drm.dev, xe_gt_fini, gt); 627 if (err) 628 return err; 629 630 err = xe_gt_sysfs_init(gt); 631 if (err) 632 return err; 633 634 err = gt_init_with_gt_forcewake(gt); 635 if (err) 636 return err; 637 638 err = xe_gt_pagefault_init(gt); 639 if (err) 640 return err; 641 642 err = xe_gt_idle_init(>->gtidle); 643 if (err) 644 return err; 645 646 err = xe_gt_freq_init(gt); 647 if (err) 648 return err; 649 650 xe_force_wake_init_engines(gt, gt_to_fw(gt)); 651 652 err = gt_init_with_all_forcewake(gt); 653 if (err) 654 return err; 655 656 xe_gt_record_user_engines(gt); 657 658 err = xe_eu_stall_init(gt); 659 if (err) 660 return err; 661 662 return 0; 663 } 664 665 /** 666 * xe_gt_mmio_init() - Initialize GT's MMIO access 667 * @gt: the GT object 668 * 669 * Initialize GT's MMIO accessor, which will be used to access registers inside 670 * this GT. 671 */ 672 void xe_gt_mmio_init(struct xe_gt *gt) 673 { 674 struct xe_tile *tile = gt_to_tile(gt); 675 struct xe_device *xe = tile_to_xe(tile); 676 677 xe_mmio_init(>->mmio, tile, tile->mmio.regs, tile->mmio.regs_size); 678 679 if (gt->info.type == XE_GT_TYPE_MEDIA) { 680 gt->mmio.adj_offset = MEDIA_GT_GSI_OFFSET; 681 gt->mmio.adj_limit = MEDIA_GT_GSI_LENGTH; 682 } else { 683 gt->mmio.adj_offset = 0; 684 gt->mmio.adj_limit = 0; 685 } 686 687 if (IS_SRIOV_VF(xe)) 688 gt->mmio.sriov_vf_gt = gt; 689 } 690 691 void xe_gt_record_user_engines(struct xe_gt *gt) 692 { 693 struct xe_hw_engine *hwe; 694 enum xe_hw_engine_id id; 695 696 gt->user_engines.mask = 0; 697 memset(gt->user_engines.instances_per_class, 0, 698 sizeof(gt->user_engines.instances_per_class)); 699 700 for_each_hw_engine(hwe, gt, id) { 701 if (xe_hw_engine_is_reserved(hwe)) 702 continue; 703 704 gt->user_engines.mask |= BIT_ULL(id); 705 gt->user_engines.instances_per_class[hwe->class]++; 706 } 707 708 xe_gt_assert(gt, (gt->user_engines.mask | gt->info.engine_mask) 709 == gt->info.engine_mask); 710 } 711 712 static int do_gt_reset(struct xe_gt *gt) 713 { 714 int err; 715 716 if (IS_SRIOV_VF(gt_to_xe(gt))) 717 return xe_gt_sriov_vf_reset(gt); 718 719 xe_gsc_wa_14015076503(gt, true); 720 721 xe_mmio_write32(>->mmio, GDRST, GRDOM_FULL); 722 err = xe_mmio_wait32(>->mmio, GDRST, GRDOM_FULL, 0, 5000, NULL, false); 723 if (err) 724 xe_gt_err(gt, "failed to clear GRDOM_FULL (%pe)\n", 725 ERR_PTR(err)); 726 727 xe_gsc_wa_14015076503(gt, false); 728 729 return err; 730 } 731 732 static int vf_gt_restart(struct xe_gt *gt) 733 { 734 int err; 735 736 err = xe_uc_sanitize_reset(>->uc); 737 if (err) 738 return err; 739 740 err = xe_uc_load_hw(>->uc); 741 if (err) 742 return err; 743 744 err = xe_uc_start(>->uc); 745 if (err) 746 return err; 747 748 return 0; 749 } 750 751 static int do_gt_restart(struct xe_gt *gt) 752 { 753 struct xe_hw_engine *hwe; 754 enum xe_hw_engine_id id; 755 int err; 756 757 if (IS_SRIOV_VF(gt_to_xe(gt))) 758 return vf_gt_restart(gt); 759 760 xe_pat_init(gt); 761 762 xe_gt_enable_host_l2_vram(gt); 763 764 xe_gt_mcr_set_implicit_defaults(gt); 765 xe_reg_sr_apply_mmio(>->reg_sr, gt); 766 767 err = xe_wopcm_init(>->uc.wopcm); 768 if (err) 769 return err; 770 771 for_each_hw_engine(hwe, gt, id) 772 xe_hw_engine_enable_ring(hwe); 773 774 err = xe_uc_sanitize_reset(>->uc); 775 if (err) 776 return err; 777 778 err = xe_uc_load_hw(>->uc); 779 if (err) 780 return err; 781 782 if (IS_SRIOV_PF(gt_to_xe(gt)) && xe_gt_is_main_type(gt)) 783 xe_lmtt_init_hw(>_to_tile(gt)->sriov.pf.lmtt); 784 785 if (IS_SRIOV_PF(gt_to_xe(gt))) 786 xe_gt_sriov_pf_init_hw(gt); 787 788 xe_mocs_init(gt); 789 err = xe_uc_start(>->uc); 790 if (err) 791 return err; 792 793 for_each_hw_engine(hwe, gt, id) 794 xe_reg_sr_apply_mmio(&hwe->reg_sr, gt); 795 796 /* Get CCS mode in sync between sw/hw */ 797 xe_gt_apply_ccs_mode(gt); 798 799 /* Restore GT freq to expected values */ 800 xe_gt_sanitize_freq(gt); 801 802 if (IS_SRIOV_PF(gt_to_xe(gt))) 803 xe_gt_sriov_pf_restart(gt); 804 805 return 0; 806 } 807 808 static int gt_reset(struct xe_gt *gt) 809 { 810 unsigned int fw_ref; 811 int err; 812 813 if (xe_device_wedged(gt_to_xe(gt))) 814 return -ECANCELED; 815 816 /* We only support GT resets with GuC submission */ 817 if (!xe_device_uc_enabled(gt_to_xe(gt))) 818 return -ENODEV; 819 820 xe_gt_info(gt, "reset started\n"); 821 822 xe_pm_runtime_get(gt_to_xe(gt)); 823 824 if (xe_fault_inject_gt_reset()) { 825 err = -ECANCELED; 826 goto err_fail; 827 } 828 829 xe_gt_sanitize(gt); 830 831 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FORCEWAKE_ALL); 832 if (!xe_force_wake_ref_has_domain(fw_ref, XE_FORCEWAKE_ALL)) { 833 err = -ETIMEDOUT; 834 goto err_out; 835 } 836 837 if (IS_SRIOV_PF(gt_to_xe(gt))) 838 xe_gt_sriov_pf_stop_prepare(gt); 839 840 xe_uc_gucrc_disable(>->uc); 841 xe_uc_stop_prepare(>->uc); 842 xe_gt_pagefault_reset(gt); 843 844 xe_uc_stop(>->uc); 845 846 xe_gt_tlb_invalidation_reset(gt); 847 848 err = do_gt_reset(gt); 849 if (err) 850 goto err_out; 851 852 err = do_gt_restart(gt); 853 if (err) 854 goto err_out; 855 856 xe_force_wake_put(gt_to_fw(gt), fw_ref); 857 xe_pm_runtime_put(gt_to_xe(gt)); 858 859 xe_gt_info(gt, "reset done\n"); 860 861 return 0; 862 863 err_out: 864 xe_force_wake_put(gt_to_fw(gt), fw_ref); 865 XE_WARN_ON(xe_uc_start(>->uc)); 866 err_fail: 867 xe_gt_err(gt, "reset failed (%pe)\n", ERR_PTR(err)); 868 869 xe_device_declare_wedged(gt_to_xe(gt)); 870 xe_pm_runtime_put(gt_to_xe(gt)); 871 872 return err; 873 } 874 875 static void gt_reset_worker(struct work_struct *w) 876 { 877 struct xe_gt *gt = container_of(w, typeof(*gt), reset.worker); 878 879 gt_reset(gt); 880 } 881 882 void xe_gt_reset_async(struct xe_gt *gt) 883 { 884 xe_gt_info(gt, "trying reset from %ps\n", __builtin_return_address(0)); 885 886 /* Don't do a reset while one is already in flight */ 887 if (!xe_fault_inject_gt_reset() && xe_uc_reset_prepare(>->uc)) 888 return; 889 890 xe_gt_info(gt, "reset queued\n"); 891 queue_work(gt->ordered_wq, >->reset.worker); 892 } 893 894 void xe_gt_suspend_prepare(struct xe_gt *gt) 895 { 896 unsigned int fw_ref; 897 898 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FORCEWAKE_ALL); 899 900 xe_uc_suspend_prepare(>->uc); 901 902 xe_force_wake_put(gt_to_fw(gt), fw_ref); 903 } 904 905 int xe_gt_suspend(struct xe_gt *gt) 906 { 907 unsigned int fw_ref; 908 int err; 909 910 xe_gt_dbg(gt, "suspending\n"); 911 xe_gt_sanitize(gt); 912 913 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FORCEWAKE_ALL); 914 if (!xe_force_wake_ref_has_domain(fw_ref, XE_FORCEWAKE_ALL)) 915 goto err_msg; 916 917 err = xe_uc_suspend(>->uc); 918 if (err) 919 goto err_force_wake; 920 921 xe_gt_idle_disable_pg(gt); 922 923 xe_gt_disable_host_l2_vram(gt); 924 925 xe_force_wake_put(gt_to_fw(gt), fw_ref); 926 xe_gt_dbg(gt, "suspended\n"); 927 928 return 0; 929 930 err_msg: 931 err = -ETIMEDOUT; 932 err_force_wake: 933 xe_force_wake_put(gt_to_fw(gt), fw_ref); 934 xe_gt_err(gt, "suspend failed (%pe)\n", ERR_PTR(err)); 935 936 return err; 937 } 938 939 void xe_gt_shutdown(struct xe_gt *gt) 940 { 941 unsigned int fw_ref; 942 943 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FORCEWAKE_ALL); 944 do_gt_reset(gt); 945 xe_force_wake_put(gt_to_fw(gt), fw_ref); 946 } 947 948 /** 949 * xe_gt_sanitize_freq() - Restore saved frequencies if necessary. 950 * @gt: the GT object 951 * 952 * Called after driver init/GSC load completes to restore GT frequencies if we 953 * limited them for any WAs. 954 */ 955 int xe_gt_sanitize_freq(struct xe_gt *gt) 956 { 957 int ret = 0; 958 959 if ((!xe_uc_fw_is_available(>->uc.gsc.fw) || 960 xe_uc_fw_is_loaded(>->uc.gsc.fw) || 961 xe_uc_fw_is_in_error_state(>->uc.gsc.fw)) && 962 XE_WA(gt, 22019338487)) 963 ret = xe_guc_pc_restore_stashed_freq(>->uc.guc.pc); 964 965 return ret; 966 } 967 968 int xe_gt_resume(struct xe_gt *gt) 969 { 970 unsigned int fw_ref; 971 int err; 972 973 xe_gt_dbg(gt, "resuming\n"); 974 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FORCEWAKE_ALL); 975 if (!xe_force_wake_ref_has_domain(fw_ref, XE_FORCEWAKE_ALL)) 976 goto err_msg; 977 978 err = do_gt_restart(gt); 979 if (err) 980 goto err_force_wake; 981 982 xe_gt_idle_enable_pg(gt); 983 984 xe_force_wake_put(gt_to_fw(gt), fw_ref); 985 xe_gt_dbg(gt, "resumed\n"); 986 987 return 0; 988 989 err_msg: 990 err = -ETIMEDOUT; 991 err_force_wake: 992 xe_force_wake_put(gt_to_fw(gt), fw_ref); 993 xe_gt_err(gt, "resume failed (%pe)\n", ERR_PTR(err)); 994 995 return err; 996 } 997 998 struct xe_hw_engine *xe_gt_hw_engine(struct xe_gt *gt, 999 enum xe_engine_class class, 1000 u16 instance, bool logical) 1001 { 1002 struct xe_hw_engine *hwe; 1003 enum xe_hw_engine_id id; 1004 1005 for_each_hw_engine(hwe, gt, id) 1006 if (hwe->class == class && 1007 ((!logical && hwe->instance == instance) || 1008 (logical && hwe->logical_instance == instance))) 1009 return hwe; 1010 1011 return NULL; 1012 } 1013 1014 struct xe_hw_engine *xe_gt_any_hw_engine_by_reset_domain(struct xe_gt *gt, 1015 enum xe_engine_class class) 1016 { 1017 struct xe_hw_engine *hwe; 1018 enum xe_hw_engine_id id; 1019 1020 for_each_hw_engine(hwe, gt, id) { 1021 switch (class) { 1022 case XE_ENGINE_CLASS_RENDER: 1023 case XE_ENGINE_CLASS_COMPUTE: 1024 if (hwe->class == XE_ENGINE_CLASS_RENDER || 1025 hwe->class == XE_ENGINE_CLASS_COMPUTE) 1026 return hwe; 1027 break; 1028 default: 1029 if (hwe->class == class) 1030 return hwe; 1031 } 1032 } 1033 1034 return NULL; 1035 } 1036 1037 struct xe_hw_engine *xe_gt_any_hw_engine(struct xe_gt *gt) 1038 { 1039 struct xe_hw_engine *hwe; 1040 enum xe_hw_engine_id id; 1041 1042 for_each_hw_engine(hwe, gt, id) 1043 return hwe; 1044 1045 return NULL; 1046 } 1047 1048 /** 1049 * xe_gt_declare_wedged() - Declare GT wedged 1050 * @gt: the GT object 1051 * 1052 * Wedge the GT which stops all submission, saves desired debug state, and 1053 * cleans up anything which could timeout. 1054 */ 1055 void xe_gt_declare_wedged(struct xe_gt *gt) 1056 { 1057 xe_gt_assert(gt, gt_to_xe(gt)->wedged.mode); 1058 1059 xe_uc_declare_wedged(>->uc); 1060 xe_gt_tlb_invalidation_reset(gt); 1061 } 1062