1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2024 Intel Corporation 4 */ 5 6 #include <drm/drm_drv.h> 7 8 #include "xe_bo.h" 9 #include "xe_exec_queue_types.h" 10 #include "xe_gt_stats.h" 11 #include "xe_migrate.h" 12 #include "xe_module.h" 13 #include "xe_pm.h" 14 #include "xe_pt.h" 15 #include "xe_svm.h" 16 #include "xe_tile.h" 17 #include "xe_ttm_vram_mgr.h" 18 #include "xe_vm.h" 19 #include "xe_vm_types.h" 20 #include "xe_vram_types.h" 21 22 static bool xe_svm_range_in_vram(struct xe_svm_range *range) 23 { 24 /* 25 * Advisory only check whether the range is currently backed by VRAM 26 * memory. 27 */ 28 29 struct drm_gpusvm_pages_flags flags = { 30 /* Pairs with WRITE_ONCE in drm_gpusvm.c */ 31 .__flags = READ_ONCE(range->base.pages.flags.__flags), 32 }; 33 34 return flags.has_devmem_pages; 35 } 36 37 static bool xe_svm_range_has_vram_binding(struct xe_svm_range *range) 38 { 39 /* Not reliable without notifier lock */ 40 return xe_svm_range_in_vram(range) && range->tile_present; 41 } 42 43 static struct xe_vm *gpusvm_to_vm(struct drm_gpusvm *gpusvm) 44 { 45 return container_of(gpusvm, struct xe_vm, svm.gpusvm); 46 } 47 48 static struct xe_vm *range_to_vm(struct drm_gpusvm_range *r) 49 { 50 return gpusvm_to_vm(r->gpusvm); 51 } 52 53 #define range_debug(r__, operation__) \ 54 vm_dbg(&range_to_vm(&(r__)->base)->xe->drm, \ 55 "%s: asid=%u, gpusvm=%p, vram=%d,%d, seqno=%lu, " \ 56 "start=0x%014lx, end=0x%014lx, size=%lu", \ 57 (operation__), range_to_vm(&(r__)->base)->usm.asid, \ 58 (r__)->base.gpusvm, \ 59 xe_svm_range_in_vram((r__)) ? 1 : 0, \ 60 xe_svm_range_has_vram_binding((r__)) ? 1 : 0, \ 61 (r__)->base.pages.notifier_seq, \ 62 xe_svm_range_start((r__)), xe_svm_range_end((r__)), \ 63 xe_svm_range_size((r__))) 64 65 void xe_svm_range_debug(struct xe_svm_range *range, const char *operation) 66 { 67 range_debug(range, operation); 68 } 69 70 static struct drm_gpusvm_range * 71 xe_svm_range_alloc(struct drm_gpusvm *gpusvm) 72 { 73 struct xe_svm_range *range; 74 75 range = kzalloc(sizeof(*range), GFP_KERNEL); 76 if (!range) 77 return NULL; 78 79 INIT_LIST_HEAD(&range->garbage_collector_link); 80 xe_vm_get(gpusvm_to_vm(gpusvm)); 81 82 return &range->base; 83 } 84 85 static void xe_svm_range_free(struct drm_gpusvm_range *range) 86 { 87 xe_vm_put(range_to_vm(range)); 88 kfree(range); 89 } 90 91 static void 92 xe_svm_garbage_collector_add_range(struct xe_vm *vm, struct xe_svm_range *range, 93 const struct mmu_notifier_range *mmu_range) 94 { 95 struct xe_device *xe = vm->xe; 96 97 range_debug(range, "GARBAGE COLLECTOR ADD"); 98 99 drm_gpusvm_range_set_unmapped(&range->base, mmu_range); 100 101 spin_lock(&vm->svm.garbage_collector.lock); 102 if (list_empty(&range->garbage_collector_link)) 103 list_add_tail(&range->garbage_collector_link, 104 &vm->svm.garbage_collector.range_list); 105 spin_unlock(&vm->svm.garbage_collector.lock); 106 107 queue_work(xe_device_get_root_tile(xe)->primary_gt->usm.pf_wq, 108 &vm->svm.garbage_collector.work); 109 } 110 111 static void xe_svm_tlb_inval_count_stats_incr(struct xe_gt *gt) 112 { 113 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_TLB_INVAL_COUNT, 1); 114 } 115 116 static u8 117 xe_svm_range_notifier_event_begin(struct xe_vm *vm, struct drm_gpusvm_range *r, 118 const struct mmu_notifier_range *mmu_range, 119 u64 *adj_start, u64 *adj_end) 120 { 121 struct xe_svm_range *range = to_xe_range(r); 122 struct xe_device *xe = vm->xe; 123 struct xe_tile *tile; 124 u8 tile_mask = 0; 125 u8 id; 126 127 xe_svm_assert_in_notifier(vm); 128 129 range_debug(range, "NOTIFIER"); 130 131 /* Skip if already unmapped or if no binding exist */ 132 if (range->base.pages.flags.unmapped || !range->tile_present) 133 return 0; 134 135 range_debug(range, "NOTIFIER - EXECUTE"); 136 137 /* Adjust invalidation to range boundaries */ 138 *adj_start = min(xe_svm_range_start(range), mmu_range->start); 139 *adj_end = max(xe_svm_range_end(range), mmu_range->end); 140 141 /* 142 * XXX: Ideally would zap PTEs in one shot in xe_svm_invalidate but the 143 * invalidation code can't correctly cope with sparse ranges or 144 * invalidations spanning multiple ranges. 145 */ 146 for_each_tile(tile, xe, id) 147 if (xe_pt_zap_ptes_range(tile, vm, range)) { 148 /* 149 * WRITE_ONCE pairs with READ_ONCE in 150 * xe_vm_has_valid_gpu_mapping() 151 */ 152 WRITE_ONCE(range->tile_invalidated, 153 range->tile_invalidated | BIT(id)); 154 155 if (!(tile_mask & BIT(id))) { 156 xe_svm_tlb_inval_count_stats_incr(tile->primary_gt); 157 if (tile->media_gt) 158 xe_svm_tlb_inval_count_stats_incr(tile->media_gt); 159 tile_mask |= BIT(id); 160 } 161 } 162 163 return tile_mask; 164 } 165 166 static void 167 xe_svm_range_notifier_event_end(struct xe_vm *vm, struct drm_gpusvm_range *r, 168 const struct mmu_notifier_range *mmu_range) 169 { 170 struct drm_gpusvm_ctx ctx = { .in_notifier = true, }; 171 172 xe_svm_assert_in_notifier(vm); 173 174 drm_gpusvm_range_unmap_pages(&vm->svm.gpusvm, r, &ctx); 175 if (!xe_vm_is_closed(vm) && mmu_range->event == MMU_NOTIFY_UNMAP) 176 xe_svm_garbage_collector_add_range(vm, to_xe_range(r), 177 mmu_range); 178 } 179 180 static s64 xe_svm_stats_ktime_us_delta(ktime_t start) 181 { 182 return IS_ENABLED(CONFIG_DEBUG_FS) ? 183 ktime_us_delta(ktime_get(), start) : 0; 184 } 185 186 static void xe_svm_tlb_inval_us_stats_incr(struct xe_gt *gt, ktime_t start) 187 { 188 s64 us_delta = xe_svm_stats_ktime_us_delta(start); 189 190 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_TLB_INVAL_US, us_delta); 191 } 192 193 static ktime_t xe_svm_stats_ktime_get(void) 194 { 195 return IS_ENABLED(CONFIG_DEBUG_FS) ? ktime_get() : 0; 196 } 197 198 static void xe_svm_invalidate(struct drm_gpusvm *gpusvm, 199 struct drm_gpusvm_notifier *notifier, 200 const struct mmu_notifier_range *mmu_range) 201 { 202 struct xe_vm *vm = gpusvm_to_vm(gpusvm); 203 struct xe_device *xe = vm->xe; 204 struct drm_gpusvm_range *r, *first; 205 struct xe_tile *tile; 206 ktime_t start = xe_svm_stats_ktime_get(); 207 u64 adj_start = mmu_range->start, adj_end = mmu_range->end; 208 u8 tile_mask = 0, id; 209 long err; 210 211 xe_svm_assert_in_notifier(vm); 212 213 vm_dbg(&gpusvm_to_vm(gpusvm)->xe->drm, 214 "INVALIDATE: asid=%u, gpusvm=%p, seqno=%lu, start=0x%016lx, end=0x%016lx, event=%d", 215 vm->usm.asid, gpusvm, notifier->notifier.invalidate_seq, 216 mmu_range->start, mmu_range->end, mmu_range->event); 217 218 /* Adjust invalidation to notifier boundaries */ 219 adj_start = max(drm_gpusvm_notifier_start(notifier), adj_start); 220 adj_end = min(drm_gpusvm_notifier_end(notifier), adj_end); 221 222 first = drm_gpusvm_range_find(notifier, adj_start, adj_end); 223 if (!first) 224 return; 225 226 /* 227 * PTs may be getting destroyed so not safe to touch these but PT should 228 * be invalidated at this point in time. Regardless we still need to 229 * ensure any dma mappings are unmapped in the here. 230 */ 231 if (xe_vm_is_closed(vm)) 232 goto range_notifier_event_end; 233 234 /* 235 * XXX: Less than ideal to always wait on VM's resv slots if an 236 * invalidation is not required. Could walk range list twice to figure 237 * out if an invalidations is need, but also not ideal. 238 */ 239 err = dma_resv_wait_timeout(xe_vm_resv(vm), 240 DMA_RESV_USAGE_BOOKKEEP, 241 false, MAX_SCHEDULE_TIMEOUT); 242 XE_WARN_ON(err <= 0); 243 244 r = first; 245 drm_gpusvm_for_each_range(r, notifier, adj_start, adj_end) 246 tile_mask |= xe_svm_range_notifier_event_begin(vm, r, mmu_range, 247 &adj_start, 248 &adj_end); 249 if (!tile_mask) 250 goto range_notifier_event_end; 251 252 xe_device_wmb(xe); 253 254 err = xe_vm_range_tilemask_tlb_inval(vm, adj_start, adj_end, tile_mask); 255 WARN_ON_ONCE(err); 256 257 range_notifier_event_end: 258 r = first; 259 drm_gpusvm_for_each_range(r, notifier, adj_start, adj_end) 260 xe_svm_range_notifier_event_end(vm, r, mmu_range); 261 for_each_tile(tile, xe, id) { 262 if (tile_mask & BIT(id)) { 263 xe_svm_tlb_inval_us_stats_incr(tile->primary_gt, start); 264 if (tile->media_gt) 265 xe_svm_tlb_inval_us_stats_incr(tile->media_gt, start); 266 } 267 } 268 } 269 270 static int __xe_svm_garbage_collector(struct xe_vm *vm, 271 struct xe_svm_range *range) 272 { 273 struct dma_fence *fence; 274 275 range_debug(range, "GARBAGE COLLECTOR"); 276 277 xe_vm_lock(vm, false); 278 fence = xe_vm_range_unbind(vm, range); 279 xe_vm_unlock(vm); 280 if (IS_ERR(fence)) 281 return PTR_ERR(fence); 282 dma_fence_put(fence); 283 284 drm_gpusvm_range_remove(&vm->svm.gpusvm, &range->base); 285 286 return 0; 287 } 288 289 static int xe_svm_range_set_default_attr(struct xe_vm *vm, u64 range_start, u64 range_end) 290 { 291 struct xe_vma *vma; 292 struct xe_vma_mem_attr default_attr = { 293 .preferred_loc = { 294 .devmem_fd = DRM_XE_PREFERRED_LOC_DEFAULT_DEVICE, 295 .migration_policy = DRM_XE_MIGRATE_ALL_PAGES, 296 }, 297 .atomic_access = DRM_XE_ATOMIC_UNDEFINED, 298 }; 299 int err = 0; 300 301 vma = xe_vm_find_vma_by_addr(vm, range_start); 302 if (!vma) 303 return -EINVAL; 304 305 if (!(vma->gpuva.flags & XE_VMA_MADV_AUTORESET)) { 306 drm_dbg(&vm->xe->drm, "Skipping madvise reset for vma.\n"); 307 return 0; 308 } 309 310 if (xe_vma_has_default_mem_attrs(vma)) 311 return 0; 312 313 vm_dbg(&vm->xe->drm, "Existing VMA start=0x%016llx, vma_end=0x%016llx", 314 xe_vma_start(vma), xe_vma_end(vma)); 315 316 if (xe_vma_start(vma) == range_start && xe_vma_end(vma) == range_end) { 317 default_attr.pat_index = vma->attr.default_pat_index; 318 default_attr.default_pat_index = vma->attr.default_pat_index; 319 vma->attr = default_attr; 320 } else { 321 vm_dbg(&vm->xe->drm, "Split VMA start=0x%016llx, vma_end=0x%016llx", 322 range_start, range_end); 323 err = xe_vm_alloc_cpu_addr_mirror_vma(vm, range_start, range_end - range_start); 324 if (err) { 325 drm_warn(&vm->xe->drm, "VMA SPLIT failed: %pe\n", ERR_PTR(err)); 326 xe_vm_kill(vm, true); 327 return err; 328 } 329 } 330 331 /* 332 * On call from xe_svm_handle_pagefault original VMA might be changed 333 * signal this to lookup for VMA again. 334 */ 335 return -EAGAIN; 336 } 337 338 static int xe_svm_garbage_collector(struct xe_vm *vm) 339 { 340 struct xe_svm_range *range; 341 u64 range_start; 342 u64 range_end; 343 int err, ret = 0; 344 345 lockdep_assert_held_write(&vm->lock); 346 347 if (xe_vm_is_closed_or_banned(vm)) 348 return -ENOENT; 349 350 for (;;) { 351 spin_lock(&vm->svm.garbage_collector.lock); 352 range = list_first_entry_or_null(&vm->svm.garbage_collector.range_list, 353 typeof(*range), 354 garbage_collector_link); 355 if (!range) 356 break; 357 358 range_start = xe_svm_range_start(range); 359 range_end = xe_svm_range_end(range); 360 361 list_del(&range->garbage_collector_link); 362 spin_unlock(&vm->svm.garbage_collector.lock); 363 364 err = __xe_svm_garbage_collector(vm, range); 365 if (err) { 366 drm_warn(&vm->xe->drm, 367 "Garbage collection failed: %pe\n", 368 ERR_PTR(err)); 369 xe_vm_kill(vm, true); 370 return err; 371 } 372 373 err = xe_svm_range_set_default_attr(vm, range_start, range_end); 374 if (err) { 375 if (err == -EAGAIN) 376 ret = -EAGAIN; 377 else 378 return err; 379 } 380 } 381 spin_unlock(&vm->svm.garbage_collector.lock); 382 383 return ret; 384 } 385 386 static void xe_svm_garbage_collector_work_func(struct work_struct *w) 387 { 388 struct xe_vm *vm = container_of(w, struct xe_vm, 389 svm.garbage_collector.work); 390 391 down_write(&vm->lock); 392 xe_svm_garbage_collector(vm); 393 up_write(&vm->lock); 394 } 395 396 #if IS_ENABLED(CONFIG_DRM_XE_PAGEMAP) 397 398 static struct xe_vram_region *page_to_vr(struct page *page) 399 { 400 return container_of(page_pgmap(page), struct xe_vram_region, pagemap); 401 } 402 403 static u64 xe_vram_region_page_to_dpa(struct xe_vram_region *vr, 404 struct page *page) 405 { 406 u64 dpa; 407 u64 pfn = page_to_pfn(page); 408 u64 offset; 409 410 xe_assert(vr->xe, is_device_private_page(page)); 411 xe_assert(vr->xe, (pfn << PAGE_SHIFT) >= vr->hpa_base); 412 413 offset = (pfn << PAGE_SHIFT) - vr->hpa_base; 414 dpa = vr->dpa_base + offset; 415 416 return dpa; 417 } 418 419 enum xe_svm_copy_dir { 420 XE_SVM_COPY_TO_VRAM, 421 XE_SVM_COPY_TO_SRAM, 422 }; 423 424 static void xe_svm_copy_kb_stats_incr(struct xe_gt *gt, 425 const enum xe_svm_copy_dir dir, 426 int kb) 427 { 428 if (dir == XE_SVM_COPY_TO_VRAM) 429 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_DEVICE_COPY_KB, kb); 430 else 431 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_CPU_COPY_KB, kb); 432 } 433 434 static void xe_svm_copy_us_stats_incr(struct xe_gt *gt, 435 const enum xe_svm_copy_dir dir, 436 unsigned long npages, 437 ktime_t start) 438 { 439 s64 us_delta = xe_svm_stats_ktime_us_delta(start); 440 441 if (dir == XE_SVM_COPY_TO_VRAM) { 442 switch (npages) { 443 case 1: 444 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_4K_DEVICE_COPY_US, 445 us_delta); 446 break; 447 case 16: 448 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_64K_DEVICE_COPY_US, 449 us_delta); 450 break; 451 case 512: 452 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_2M_DEVICE_COPY_US, 453 us_delta); 454 break; 455 } 456 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_DEVICE_COPY_US, 457 us_delta); 458 } else { 459 switch (npages) { 460 case 1: 461 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_4K_CPU_COPY_US, 462 us_delta); 463 break; 464 case 16: 465 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_64K_CPU_COPY_US, 466 us_delta); 467 break; 468 case 512: 469 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_2M_CPU_COPY_US, 470 us_delta); 471 break; 472 } 473 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_CPU_COPY_US, 474 us_delta); 475 } 476 } 477 478 static int xe_svm_copy(struct page **pages, 479 struct drm_pagemap_addr *pagemap_addr, 480 unsigned long npages, const enum xe_svm_copy_dir dir) 481 { 482 struct xe_vram_region *vr = NULL; 483 struct xe_gt *gt = NULL; 484 struct xe_device *xe; 485 struct dma_fence *fence = NULL; 486 unsigned long i; 487 #define XE_VRAM_ADDR_INVALID ~0x0ull 488 u64 vram_addr = XE_VRAM_ADDR_INVALID; 489 int err = 0, pos = 0; 490 bool sram = dir == XE_SVM_COPY_TO_SRAM; 491 ktime_t start = xe_svm_stats_ktime_get(); 492 493 /* 494 * This flow is complex: it locates physically contiguous device pages, 495 * derives the starting physical address, and performs a single GPU copy 496 * to for every 8M chunk in a DMA address array. Both device pages and 497 * DMA addresses may be sparsely populated. If either is NULL, a copy is 498 * triggered based on the current search state. The last GPU copy is 499 * waited on to ensure all copies are complete. 500 */ 501 502 for (i = 0; i < npages; ++i) { 503 struct page *spage = pages[i]; 504 struct dma_fence *__fence; 505 u64 __vram_addr; 506 bool match = false, chunk, last; 507 508 #define XE_MIGRATE_CHUNK_SIZE SZ_8M 509 chunk = (i - pos) == (XE_MIGRATE_CHUNK_SIZE / PAGE_SIZE); 510 last = (i + 1) == npages; 511 512 /* No CPU page and no device pages queue'd to copy */ 513 if (!pagemap_addr[i].addr && vram_addr == XE_VRAM_ADDR_INVALID) 514 continue; 515 516 if (!vr && spage) { 517 vr = page_to_vr(spage); 518 gt = xe_migrate_exec_queue(vr->migrate)->gt; 519 xe = vr->xe; 520 } 521 XE_WARN_ON(spage && page_to_vr(spage) != vr); 522 523 /* 524 * CPU page and device page valid, capture physical address on 525 * first device page, check if physical contiguous on subsequent 526 * device pages. 527 */ 528 if (pagemap_addr[i].addr && spage) { 529 __vram_addr = xe_vram_region_page_to_dpa(vr, spage); 530 if (vram_addr == XE_VRAM_ADDR_INVALID) { 531 vram_addr = __vram_addr; 532 pos = i; 533 } 534 535 match = vram_addr + PAGE_SIZE * (i - pos) == __vram_addr; 536 /* Expected with contiguous memory */ 537 xe_assert(vr->xe, match); 538 539 if (pagemap_addr[i].order) { 540 i += NR_PAGES(pagemap_addr[i].order) - 1; 541 chunk = (i - pos) == (XE_MIGRATE_CHUNK_SIZE / PAGE_SIZE); 542 last = (i + 1) == npages; 543 } 544 } 545 546 /* 547 * Mismatched physical address, 8M copy chunk, or last page - 548 * trigger a copy. 549 */ 550 if (!match || chunk || last) { 551 /* 552 * Extra page for first copy if last page and matching 553 * physical address. 554 */ 555 int incr = (match && last) ? 1 : 0; 556 557 if (vram_addr != XE_VRAM_ADDR_INVALID) { 558 xe_svm_copy_kb_stats_incr(gt, dir, 559 (i - pos + incr) * 560 (PAGE_SIZE / SZ_1K)); 561 if (sram) { 562 vm_dbg(&xe->drm, 563 "COPY TO SRAM - 0x%016llx -> 0x%016llx, NPAGES=%ld", 564 vram_addr, 565 (u64)pagemap_addr[pos].addr, i - pos + incr); 566 __fence = xe_migrate_from_vram(vr->migrate, 567 i - pos + incr, 568 vram_addr, 569 &pagemap_addr[pos]); 570 } else { 571 vm_dbg(&xe->drm, 572 "COPY TO VRAM - 0x%016llx -> 0x%016llx, NPAGES=%ld", 573 (u64)pagemap_addr[pos].addr, vram_addr, 574 i - pos + incr); 575 __fence = xe_migrate_to_vram(vr->migrate, 576 i - pos + incr, 577 &pagemap_addr[pos], 578 vram_addr); 579 } 580 if (IS_ERR(__fence)) { 581 err = PTR_ERR(__fence); 582 goto err_out; 583 } 584 585 dma_fence_put(fence); 586 fence = __fence; 587 } 588 589 /* Setup physical address of next device page */ 590 if (pagemap_addr[i].addr && spage) { 591 vram_addr = __vram_addr; 592 pos = i; 593 } else { 594 vram_addr = XE_VRAM_ADDR_INVALID; 595 } 596 597 /* Extra mismatched device page, copy it */ 598 if (!match && last && vram_addr != XE_VRAM_ADDR_INVALID) { 599 xe_svm_copy_kb_stats_incr(gt, dir, 600 (PAGE_SIZE / SZ_1K)); 601 if (sram) { 602 vm_dbg(&xe->drm, 603 "COPY TO SRAM - 0x%016llx -> 0x%016llx, NPAGES=%d", 604 vram_addr, (u64)pagemap_addr[pos].addr, 1); 605 __fence = xe_migrate_from_vram(vr->migrate, 1, 606 vram_addr, 607 &pagemap_addr[pos]); 608 } else { 609 vm_dbg(&xe->drm, 610 "COPY TO VRAM - 0x%016llx -> 0x%016llx, NPAGES=%d", 611 (u64)pagemap_addr[pos].addr, vram_addr, 1); 612 __fence = xe_migrate_to_vram(vr->migrate, 1, 613 &pagemap_addr[pos], 614 vram_addr); 615 } 616 if (IS_ERR(__fence)) { 617 err = PTR_ERR(__fence); 618 goto err_out; 619 } 620 621 dma_fence_put(fence); 622 fence = __fence; 623 } 624 } 625 } 626 627 err_out: 628 /* Wait for all copies to complete */ 629 if (fence) { 630 dma_fence_wait(fence, false); 631 dma_fence_put(fence); 632 } 633 634 /* 635 * XXX: We can't derive the GT here (or anywhere in this functions, but 636 * compute always uses the primary GT so accumlate stats on the likely 637 * GT of the fault. 638 */ 639 if (gt) 640 xe_svm_copy_us_stats_incr(gt, dir, npages, start); 641 642 return err; 643 #undef XE_MIGRATE_CHUNK_SIZE 644 #undef XE_VRAM_ADDR_INVALID 645 } 646 647 static int xe_svm_copy_to_devmem(struct page **pages, 648 struct drm_pagemap_addr *pagemap_addr, 649 unsigned long npages) 650 { 651 return xe_svm_copy(pages, pagemap_addr, npages, XE_SVM_COPY_TO_VRAM); 652 } 653 654 static int xe_svm_copy_to_ram(struct page **pages, 655 struct drm_pagemap_addr *pagemap_addr, 656 unsigned long npages) 657 { 658 return xe_svm_copy(pages, pagemap_addr, npages, XE_SVM_COPY_TO_SRAM); 659 } 660 661 static struct xe_bo *to_xe_bo(struct drm_pagemap_devmem *devmem_allocation) 662 { 663 return container_of(devmem_allocation, struct xe_bo, devmem_allocation); 664 } 665 666 static void xe_svm_devmem_release(struct drm_pagemap_devmem *devmem_allocation) 667 { 668 struct xe_bo *bo = to_xe_bo(devmem_allocation); 669 struct xe_device *xe = xe_bo_device(bo); 670 671 xe_bo_put_async(bo); 672 xe_pm_runtime_put(xe); 673 } 674 675 static u64 block_offset_to_pfn(struct xe_vram_region *vr, u64 offset) 676 { 677 return PHYS_PFN(offset + vr->hpa_base); 678 } 679 680 static struct drm_buddy *vram_to_buddy(struct xe_vram_region *vram) 681 { 682 return &vram->ttm.mm; 683 } 684 685 static int xe_svm_populate_devmem_pfn(struct drm_pagemap_devmem *devmem_allocation, 686 unsigned long npages, unsigned long *pfn) 687 { 688 struct xe_bo *bo = to_xe_bo(devmem_allocation); 689 struct ttm_resource *res = bo->ttm.resource; 690 struct list_head *blocks = &to_xe_ttm_vram_mgr_resource(res)->blocks; 691 struct drm_buddy_block *block; 692 int j = 0; 693 694 list_for_each_entry(block, blocks, link) { 695 struct xe_vram_region *vr = block->private; 696 struct drm_buddy *buddy = vram_to_buddy(vr); 697 u64 block_pfn = block_offset_to_pfn(vr, drm_buddy_block_offset(block)); 698 int i; 699 700 for (i = 0; i < drm_buddy_block_size(buddy, block) >> PAGE_SHIFT; ++i) 701 pfn[j++] = block_pfn + i; 702 } 703 704 return 0; 705 } 706 707 static const struct drm_pagemap_devmem_ops dpagemap_devmem_ops = { 708 .devmem_release = xe_svm_devmem_release, 709 .populate_devmem_pfn = xe_svm_populate_devmem_pfn, 710 .copy_to_devmem = xe_svm_copy_to_devmem, 711 .copy_to_ram = xe_svm_copy_to_ram, 712 }; 713 714 #endif 715 716 static const struct drm_gpusvm_ops gpusvm_ops = { 717 .range_alloc = xe_svm_range_alloc, 718 .range_free = xe_svm_range_free, 719 .invalidate = xe_svm_invalidate, 720 }; 721 722 static const unsigned long fault_chunk_sizes[] = { 723 SZ_2M, 724 SZ_64K, 725 SZ_4K, 726 }; 727 728 /** 729 * xe_svm_init() - SVM initialize 730 * @vm: The VM. 731 * 732 * Initialize SVM state which is embedded within the VM. 733 * 734 * Return: 0 on success, negative error code on error. 735 */ 736 int xe_svm_init(struct xe_vm *vm) 737 { 738 int err; 739 740 if (vm->flags & XE_VM_FLAG_FAULT_MODE) { 741 spin_lock_init(&vm->svm.garbage_collector.lock); 742 INIT_LIST_HEAD(&vm->svm.garbage_collector.range_list); 743 INIT_WORK(&vm->svm.garbage_collector.work, 744 xe_svm_garbage_collector_work_func); 745 746 err = drm_gpusvm_init(&vm->svm.gpusvm, "Xe SVM", &vm->xe->drm, 747 current->mm, 0, vm->size, 748 xe_modparam.svm_notifier_size * SZ_1M, 749 &gpusvm_ops, fault_chunk_sizes, 750 ARRAY_SIZE(fault_chunk_sizes)); 751 drm_gpusvm_driver_set_lock(&vm->svm.gpusvm, &vm->lock); 752 } else { 753 err = drm_gpusvm_init(&vm->svm.gpusvm, "Xe SVM (simple)", 754 &vm->xe->drm, NULL, 0, 0, 0, NULL, 755 NULL, 0); 756 } 757 758 return err; 759 } 760 761 /** 762 * xe_svm_close() - SVM close 763 * @vm: The VM. 764 * 765 * Close SVM state (i.e., stop and flush all SVM actions). 766 */ 767 void xe_svm_close(struct xe_vm *vm) 768 { 769 xe_assert(vm->xe, xe_vm_is_closed(vm)); 770 flush_work(&vm->svm.garbage_collector.work); 771 } 772 773 /** 774 * xe_svm_fini() - SVM finalize 775 * @vm: The VM. 776 * 777 * Finalize SVM state which is embedded within the VM. 778 */ 779 void xe_svm_fini(struct xe_vm *vm) 780 { 781 xe_assert(vm->xe, xe_vm_is_closed(vm)); 782 783 drm_gpusvm_fini(&vm->svm.gpusvm); 784 } 785 786 static bool xe_svm_range_is_valid(struct xe_svm_range *range, 787 struct xe_tile *tile, 788 bool devmem_only) 789 { 790 return (xe_vm_has_valid_gpu_mapping(tile, range->tile_present, 791 range->tile_invalidated) && 792 (!devmem_only || xe_svm_range_in_vram(range))); 793 } 794 795 /** xe_svm_range_migrate_to_smem() - Move range pages from VRAM to SMEM 796 * @vm: xe_vm pointer 797 * @range: Pointer to the SVM range structure 798 * 799 * The xe_svm_range_migrate_to_smem() checks range has pages in VRAM 800 * and migrates them to SMEM 801 */ 802 void xe_svm_range_migrate_to_smem(struct xe_vm *vm, struct xe_svm_range *range) 803 { 804 if (xe_svm_range_in_vram(range)) 805 drm_gpusvm_range_evict(&vm->svm.gpusvm, &range->base); 806 } 807 808 /** 809 * xe_svm_range_validate() - Check if the SVM range is valid 810 * @vm: xe_vm pointer 811 * @range: Pointer to the SVM range structure 812 * @tile_mask: Mask representing the tiles to be checked 813 * @devmem_preferred : if true range needs to be in devmem 814 * 815 * The xe_svm_range_validate() function checks if a range is 816 * valid and located in the desired memory region. 817 * 818 * Return: true if the range is valid, false otherwise 819 */ 820 bool xe_svm_range_validate(struct xe_vm *vm, 821 struct xe_svm_range *range, 822 u8 tile_mask, bool devmem_preferred) 823 { 824 bool ret; 825 826 xe_svm_notifier_lock(vm); 827 828 ret = (range->tile_present & ~range->tile_invalidated & tile_mask) == tile_mask && 829 (devmem_preferred == range->base.pages.flags.has_devmem_pages); 830 831 xe_svm_notifier_unlock(vm); 832 833 return ret; 834 } 835 836 /** 837 * xe_svm_find_vma_start - Find start of CPU VMA 838 * @vm: xe_vm pointer 839 * @start: start address 840 * @end: end address 841 * @vma: Pointer to struct xe_vma 842 * 843 * 844 * This function searches for a cpu vma, within the specified 845 * range [start, end] in the given VM. It adjusts the range based on the 846 * xe_vma start and end addresses. If no cpu VMA is found, it returns ULONG_MAX. 847 * 848 * Return: The starting address of the VMA within the range, 849 * or ULONG_MAX if no VMA is found 850 */ 851 u64 xe_svm_find_vma_start(struct xe_vm *vm, u64 start, u64 end, struct xe_vma *vma) 852 { 853 return drm_gpusvm_find_vma_start(&vm->svm.gpusvm, 854 max(start, xe_vma_start(vma)), 855 min(end, xe_vma_end(vma))); 856 } 857 858 #if IS_ENABLED(CONFIG_DRM_XE_PAGEMAP) 859 static int xe_drm_pagemap_populate_mm(struct drm_pagemap *dpagemap, 860 unsigned long start, unsigned long end, 861 struct mm_struct *mm, 862 unsigned long timeslice_ms) 863 { 864 struct xe_vram_region *vr = container_of(dpagemap, typeof(*vr), dpagemap); 865 struct xe_device *xe = vr->xe; 866 struct device *dev = xe->drm.dev; 867 struct drm_buddy_block *block; 868 struct xe_validation_ctx vctx; 869 struct list_head *blocks; 870 struct drm_exec exec; 871 struct xe_bo *bo; 872 int err = 0, idx; 873 874 if (!drm_dev_enter(&xe->drm, &idx)) 875 return -ENODEV; 876 877 xe_pm_runtime_get(xe); 878 879 xe_validation_guard(&vctx, &xe->val, &exec, (struct xe_val_flags) {}, err) { 880 bo = xe_bo_create_locked(xe, NULL, NULL, end - start, 881 ttm_bo_type_device, 882 (IS_DGFX(xe) ? XE_BO_FLAG_VRAM(vr) : XE_BO_FLAG_SYSTEM) | 883 XE_BO_FLAG_CPU_ADDR_MIRROR, &exec); 884 drm_exec_retry_on_contention(&exec); 885 if (IS_ERR(bo)) { 886 err = PTR_ERR(bo); 887 xe_validation_retry_on_oom(&vctx, &err); 888 break; 889 } 890 891 drm_pagemap_devmem_init(&bo->devmem_allocation, dev, mm, 892 &dpagemap_devmem_ops, dpagemap, end - start); 893 894 blocks = &to_xe_ttm_vram_mgr_resource(bo->ttm.resource)->blocks; 895 list_for_each_entry(block, blocks, link) 896 block->private = vr; 897 898 xe_bo_get(bo); 899 900 /* Ensure the device has a pm ref while there are device pages active. */ 901 xe_pm_runtime_get_noresume(xe); 902 err = drm_pagemap_migrate_to_devmem(&bo->devmem_allocation, mm, 903 start, end, timeslice_ms, 904 xe_svm_devm_owner(xe)); 905 if (err) 906 xe_svm_devmem_release(&bo->devmem_allocation); 907 xe_bo_unlock(bo); 908 xe_bo_put(bo); 909 } 910 xe_pm_runtime_put(xe); 911 drm_dev_exit(idx); 912 913 return err; 914 } 915 #endif 916 917 static bool supports_4K_migration(struct xe_device *xe) 918 { 919 if (xe->info.vram_flags & XE_VRAM_FLAGS_NEED64K) 920 return false; 921 922 return true; 923 } 924 925 /** 926 * xe_svm_range_needs_migrate_to_vram() - SVM range needs migrate to VRAM or not 927 * @range: SVM range for which migration needs to be decided 928 * @vma: vma which has range 929 * @preferred_region_is_vram: preferred region for range is vram 930 * 931 * Return: True for range needing migration and migration is supported else false 932 */ 933 bool xe_svm_range_needs_migrate_to_vram(struct xe_svm_range *range, struct xe_vma *vma, 934 bool preferred_region_is_vram) 935 { 936 struct xe_vm *vm = range_to_vm(&range->base); 937 u64 range_size = xe_svm_range_size(range); 938 939 if (!range->base.pages.flags.migrate_devmem || !preferred_region_is_vram) 940 return false; 941 942 xe_assert(vm->xe, IS_DGFX(vm->xe)); 943 944 if (xe_svm_range_in_vram(range)) { 945 drm_info(&vm->xe->drm, "Range is already in VRAM\n"); 946 return false; 947 } 948 949 if (range_size < SZ_64K && !supports_4K_migration(vm->xe)) { 950 drm_dbg(&vm->xe->drm, "Platform doesn't support SZ_4K range migration\n"); 951 return false; 952 } 953 954 return true; 955 } 956 957 #define DECL_SVM_RANGE_COUNT_STATS(elem, stat) \ 958 static void xe_svm_range_##elem##_count_stats_incr(struct xe_gt *gt, \ 959 struct xe_svm_range *range) \ 960 { \ 961 switch (xe_svm_range_size(range)) { \ 962 case SZ_4K: \ 963 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_4K_##stat##_COUNT, 1); \ 964 break; \ 965 case SZ_64K: \ 966 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_64K_##stat##_COUNT, 1); \ 967 break; \ 968 case SZ_2M: \ 969 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_2M_##stat##_COUNT, 1); \ 970 break; \ 971 } \ 972 } \ 973 974 DECL_SVM_RANGE_COUNT_STATS(fault, PAGEFAULT) 975 DECL_SVM_RANGE_COUNT_STATS(valid_fault, VALID_PAGEFAULT) 976 DECL_SVM_RANGE_COUNT_STATS(migrate, MIGRATE) 977 978 #define DECL_SVM_RANGE_US_STATS(elem, stat) \ 979 static void xe_svm_range_##elem##_us_stats_incr(struct xe_gt *gt, \ 980 struct xe_svm_range *range, \ 981 ktime_t start) \ 982 { \ 983 s64 us_delta = xe_svm_stats_ktime_us_delta(start); \ 984 \ 985 switch (xe_svm_range_size(range)) { \ 986 case SZ_4K: \ 987 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_4K_##stat##_US, \ 988 us_delta); \ 989 break; \ 990 case SZ_64K: \ 991 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_64K_##stat##_US, \ 992 us_delta); \ 993 break; \ 994 case SZ_2M: \ 995 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_2M_##stat##_US, \ 996 us_delta); \ 997 break; \ 998 } \ 999 } \ 1000 1001 DECL_SVM_RANGE_US_STATS(migrate, MIGRATE) 1002 DECL_SVM_RANGE_US_STATS(get_pages, GET_PAGES) 1003 DECL_SVM_RANGE_US_STATS(bind, BIND) 1004 DECL_SVM_RANGE_US_STATS(fault, PAGEFAULT) 1005 1006 static int __xe_svm_handle_pagefault(struct xe_vm *vm, struct xe_vma *vma, 1007 struct xe_gt *gt, u64 fault_addr, 1008 bool need_vram) 1009 { 1010 int devmem_possible = IS_DGFX(vm->xe) && 1011 IS_ENABLED(CONFIG_DRM_XE_PAGEMAP); 1012 struct drm_gpusvm_ctx ctx = { 1013 .read_only = xe_vma_read_only(vma), 1014 .devmem_possible = devmem_possible, 1015 .check_pages_threshold = devmem_possible ? SZ_64K : 0, 1016 .devmem_only = need_vram && devmem_possible, 1017 .timeslice_ms = need_vram && devmem_possible ? 1018 vm->xe->atomic_svm_timeslice_ms : 0, 1019 .device_private_page_owner = xe_svm_devm_owner(vm->xe), 1020 }; 1021 struct xe_validation_ctx vctx; 1022 struct drm_exec exec; 1023 struct xe_svm_range *range; 1024 struct dma_fence *fence; 1025 struct drm_pagemap *dpagemap; 1026 struct xe_tile *tile = gt_to_tile(gt); 1027 int migrate_try_count = ctx.devmem_only ? 3 : 1; 1028 ktime_t start = xe_svm_stats_ktime_get(), bind_start, get_pages_start; 1029 int err; 1030 1031 lockdep_assert_held_write(&vm->lock); 1032 xe_assert(vm->xe, xe_vma_is_cpu_addr_mirror(vma)); 1033 1034 xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_PAGEFAULT_COUNT, 1); 1035 1036 retry: 1037 /* Always process UNMAPs first so view SVM ranges is current */ 1038 err = xe_svm_garbage_collector(vm); 1039 if (err) 1040 return err; 1041 1042 dpagemap = xe_vma_resolve_pagemap(vma, tile); 1043 if (!dpagemap && !ctx.devmem_only) 1044 ctx.device_private_page_owner = NULL; 1045 range = xe_svm_range_find_or_insert(vm, fault_addr, vma, &ctx); 1046 1047 if (IS_ERR(range)) 1048 return PTR_ERR(range); 1049 1050 xe_svm_range_fault_count_stats_incr(gt, range); 1051 1052 if (ctx.devmem_only && !range->base.pages.flags.migrate_devmem) { 1053 err = -EACCES; 1054 goto out; 1055 } 1056 1057 if (xe_svm_range_is_valid(range, tile, ctx.devmem_only)) { 1058 xe_svm_range_valid_fault_count_stats_incr(gt, range); 1059 range_debug(range, "PAGE FAULT - VALID"); 1060 goto out; 1061 } 1062 1063 range_debug(range, "PAGE FAULT"); 1064 1065 if (--migrate_try_count >= 0 && 1066 xe_svm_range_needs_migrate_to_vram(range, vma, !!dpagemap || ctx.devmem_only)) { 1067 ktime_t migrate_start = xe_svm_stats_ktime_get(); 1068 1069 /* TODO : For multi-device dpagemap will be used to find the 1070 * remote tile and remote device. Will need to modify 1071 * xe_svm_alloc_vram to use dpagemap for future multi-device 1072 * support. 1073 */ 1074 xe_svm_range_migrate_count_stats_incr(gt, range); 1075 err = xe_svm_alloc_vram(tile, range, &ctx); 1076 xe_svm_range_migrate_us_stats_incr(gt, range, migrate_start); 1077 ctx.timeslice_ms <<= 1; /* Double timeslice if we have to retry */ 1078 if (err) { 1079 if (migrate_try_count || !ctx.devmem_only) { 1080 drm_dbg(&vm->xe->drm, 1081 "VRAM allocation failed, falling back to retrying fault, asid=%u, errno=%pe\n", 1082 vm->usm.asid, ERR_PTR(err)); 1083 1084 /* 1085 * In the devmem-only case, mixed mappings may 1086 * be found. The get_pages function will fix 1087 * these up to a single location, allowing the 1088 * page fault handler to make forward progress. 1089 */ 1090 if (ctx.devmem_only) 1091 goto get_pages; 1092 else 1093 goto retry; 1094 } else { 1095 drm_err(&vm->xe->drm, 1096 "VRAM allocation failed, retry count exceeded, asid=%u, errno=%pe\n", 1097 vm->usm.asid, ERR_PTR(err)); 1098 return err; 1099 } 1100 } 1101 } 1102 1103 get_pages: 1104 get_pages_start = xe_svm_stats_ktime_get(); 1105 1106 range_debug(range, "GET PAGES"); 1107 err = xe_svm_range_get_pages(vm, range, &ctx); 1108 /* Corner where CPU mappings have changed */ 1109 if (err == -EOPNOTSUPP || err == -EFAULT || err == -EPERM) { 1110 ctx.timeslice_ms <<= 1; /* Double timeslice if we have to retry */ 1111 if (migrate_try_count > 0 || !ctx.devmem_only) { 1112 drm_dbg(&vm->xe->drm, 1113 "Get pages failed, falling back to retrying, asid=%u, gpusvm=%p, errno=%pe\n", 1114 vm->usm.asid, &vm->svm.gpusvm, ERR_PTR(err)); 1115 range_debug(range, "PAGE FAULT - RETRY PAGES"); 1116 goto retry; 1117 } else { 1118 drm_err(&vm->xe->drm, 1119 "Get pages failed, retry count exceeded, asid=%u, gpusvm=%p, errno=%pe\n", 1120 vm->usm.asid, &vm->svm.gpusvm, ERR_PTR(err)); 1121 } 1122 } 1123 if (err) { 1124 range_debug(range, "PAGE FAULT - FAIL PAGE COLLECT"); 1125 goto out; 1126 } 1127 1128 xe_svm_range_get_pages_us_stats_incr(gt, range, get_pages_start); 1129 range_debug(range, "PAGE FAULT - BIND"); 1130 1131 bind_start = xe_svm_stats_ktime_get(); 1132 xe_validation_guard(&vctx, &vm->xe->val, &exec, (struct xe_val_flags) {}, err) { 1133 err = xe_vm_drm_exec_lock(vm, &exec); 1134 drm_exec_retry_on_contention(&exec); 1135 1136 xe_vm_set_validation_exec(vm, &exec); 1137 fence = xe_vm_range_rebind(vm, vma, range, BIT(tile->id)); 1138 xe_vm_set_validation_exec(vm, NULL); 1139 if (IS_ERR(fence)) { 1140 drm_exec_retry_on_contention(&exec); 1141 err = PTR_ERR(fence); 1142 xe_validation_retry_on_oom(&vctx, &err); 1143 xe_svm_range_bind_us_stats_incr(gt, range, bind_start); 1144 break; 1145 } 1146 } 1147 if (err) 1148 goto err_out; 1149 1150 dma_fence_wait(fence, false); 1151 dma_fence_put(fence); 1152 xe_svm_range_bind_us_stats_incr(gt, range, bind_start); 1153 1154 out: 1155 xe_svm_range_fault_us_stats_incr(gt, range, start); 1156 return 0; 1157 1158 err_out: 1159 if (err == -EAGAIN) { 1160 ctx.timeslice_ms <<= 1; /* Double timeslice if we have to retry */ 1161 range_debug(range, "PAGE FAULT - RETRY BIND"); 1162 goto retry; 1163 } 1164 1165 return err; 1166 } 1167 1168 /** 1169 * xe_svm_handle_pagefault() - SVM handle page fault 1170 * @vm: The VM. 1171 * @vma: The CPU address mirror VMA. 1172 * @gt: The gt upon the fault occurred. 1173 * @fault_addr: The GPU fault address. 1174 * @atomic: The fault atomic access bit. 1175 * 1176 * Create GPU bindings for a SVM page fault. Optionally migrate to device 1177 * memory. 1178 * 1179 * Return: 0 on success, negative error code on error. 1180 */ 1181 int xe_svm_handle_pagefault(struct xe_vm *vm, struct xe_vma *vma, 1182 struct xe_gt *gt, u64 fault_addr, 1183 bool atomic) 1184 { 1185 int need_vram, ret; 1186 retry: 1187 need_vram = xe_vma_need_vram_for_atomic(vm->xe, vma, atomic); 1188 if (need_vram < 0) 1189 return need_vram; 1190 1191 ret = __xe_svm_handle_pagefault(vm, vma, gt, fault_addr, 1192 need_vram ? true : false); 1193 if (ret == -EAGAIN) { 1194 /* 1195 * Retry once on -EAGAIN to re-lookup the VMA, as the original VMA 1196 * may have been split by xe_svm_range_set_default_attr. 1197 */ 1198 vma = xe_vm_find_vma_by_addr(vm, fault_addr); 1199 if (!vma) 1200 return -EINVAL; 1201 1202 goto retry; 1203 } 1204 return ret; 1205 } 1206 1207 /** 1208 * xe_svm_has_mapping() - SVM has mappings 1209 * @vm: The VM. 1210 * @start: Start address. 1211 * @end: End address. 1212 * 1213 * Check if an address range has SVM mappings. 1214 * 1215 * Return: True if address range has a SVM mapping, False otherwise 1216 */ 1217 bool xe_svm_has_mapping(struct xe_vm *vm, u64 start, u64 end) 1218 { 1219 return drm_gpusvm_has_mapping(&vm->svm.gpusvm, start, end); 1220 } 1221 1222 /** 1223 * xe_svm_unmap_address_range - UNMAP SVM mappings and ranges 1224 * @vm: The VM 1225 * @start: start addr 1226 * @end: end addr 1227 * 1228 * This function UNMAPS svm ranges if start or end address are inside them. 1229 */ 1230 void xe_svm_unmap_address_range(struct xe_vm *vm, u64 start, u64 end) 1231 { 1232 struct drm_gpusvm_notifier *notifier, *next; 1233 1234 lockdep_assert_held_write(&vm->lock); 1235 1236 drm_gpusvm_for_each_notifier_safe(notifier, next, &vm->svm.gpusvm, start, end) { 1237 struct drm_gpusvm_range *range, *__next; 1238 1239 drm_gpusvm_for_each_range_safe(range, __next, notifier, start, end) { 1240 if (start > drm_gpusvm_range_start(range) || 1241 end < drm_gpusvm_range_end(range)) { 1242 if (IS_DGFX(vm->xe) && xe_svm_range_in_vram(to_xe_range(range))) 1243 drm_gpusvm_range_evict(&vm->svm.gpusvm, range); 1244 drm_gpusvm_range_get(range); 1245 __xe_svm_garbage_collector(vm, to_xe_range(range)); 1246 if (!list_empty(&to_xe_range(range)->garbage_collector_link)) { 1247 spin_lock(&vm->svm.garbage_collector.lock); 1248 list_del(&to_xe_range(range)->garbage_collector_link); 1249 spin_unlock(&vm->svm.garbage_collector.lock); 1250 } 1251 drm_gpusvm_range_put(range); 1252 } 1253 } 1254 } 1255 } 1256 1257 /** 1258 * xe_svm_bo_evict() - SVM evict BO to system memory 1259 * @bo: BO to evict 1260 * 1261 * SVM evict BO to system memory. GPU SVM layer ensures all device pages 1262 * are evicted before returning. 1263 * 1264 * Return: 0 on success standard error code otherwise 1265 */ 1266 int xe_svm_bo_evict(struct xe_bo *bo) 1267 { 1268 return drm_pagemap_evict_to_ram(&bo->devmem_allocation); 1269 } 1270 1271 /** 1272 * xe_svm_range_find_or_insert- Find or insert GPU SVM range 1273 * @vm: xe_vm pointer 1274 * @addr: address for which range needs to be found/inserted 1275 * @vma: Pointer to struct xe_vma which mirrors CPU 1276 * @ctx: GPU SVM context 1277 * 1278 * This function finds or inserts a newly allocated a SVM range based on the 1279 * address. 1280 * 1281 * Return: Pointer to the SVM range on success, ERR_PTR() on failure. 1282 */ 1283 struct xe_svm_range *xe_svm_range_find_or_insert(struct xe_vm *vm, u64 addr, 1284 struct xe_vma *vma, struct drm_gpusvm_ctx *ctx) 1285 { 1286 struct drm_gpusvm_range *r; 1287 1288 r = drm_gpusvm_range_find_or_insert(&vm->svm.gpusvm, max(addr, xe_vma_start(vma)), 1289 xe_vma_start(vma), xe_vma_end(vma), ctx); 1290 if (IS_ERR(r)) 1291 return ERR_CAST(r); 1292 1293 return to_xe_range(r); 1294 } 1295 1296 /** 1297 * xe_svm_range_get_pages() - Get pages for a SVM range 1298 * @vm: Pointer to the struct xe_vm 1299 * @range: Pointer to the xe SVM range structure 1300 * @ctx: GPU SVM context 1301 * 1302 * This function gets pages for a SVM range and ensures they are mapped for 1303 * DMA access. In case of failure with -EOPNOTSUPP, it evicts the range. 1304 * 1305 * Return: 0 on success, negative error code on failure. 1306 */ 1307 int xe_svm_range_get_pages(struct xe_vm *vm, struct xe_svm_range *range, 1308 struct drm_gpusvm_ctx *ctx) 1309 { 1310 int err = 0; 1311 1312 err = drm_gpusvm_range_get_pages(&vm->svm.gpusvm, &range->base, ctx); 1313 if (err == -EOPNOTSUPP) { 1314 range_debug(range, "PAGE FAULT - EVICT PAGES"); 1315 drm_gpusvm_range_evict(&vm->svm.gpusvm, &range->base); 1316 } 1317 1318 return err; 1319 } 1320 1321 /** 1322 * xe_svm_ranges_zap_ptes_in_range - clear ptes of svm ranges in input range 1323 * @vm: Pointer to the xe_vm structure 1324 * @start: Start of the input range 1325 * @end: End of the input range 1326 * 1327 * This function removes the page table entries (PTEs) associated 1328 * with the svm ranges within the given input start and end 1329 * 1330 * Return: tile_mask for which gt's need to be tlb invalidated. 1331 */ 1332 u8 xe_svm_ranges_zap_ptes_in_range(struct xe_vm *vm, u64 start, u64 end) 1333 { 1334 struct drm_gpusvm_notifier *notifier; 1335 struct xe_svm_range *range; 1336 u64 adj_start, adj_end; 1337 struct xe_tile *tile; 1338 u8 tile_mask = 0; 1339 u8 id; 1340 1341 lockdep_assert(lockdep_is_held_type(&vm->svm.gpusvm.notifier_lock, 1) && 1342 lockdep_is_held_type(&vm->lock, 0)); 1343 1344 drm_gpusvm_for_each_notifier(notifier, &vm->svm.gpusvm, start, end) { 1345 struct drm_gpusvm_range *r = NULL; 1346 1347 adj_start = max(start, drm_gpusvm_notifier_start(notifier)); 1348 adj_end = min(end, drm_gpusvm_notifier_end(notifier)); 1349 drm_gpusvm_for_each_range(r, notifier, adj_start, adj_end) { 1350 range = to_xe_range(r); 1351 for_each_tile(tile, vm->xe, id) { 1352 if (xe_pt_zap_ptes_range(tile, vm, range)) { 1353 tile_mask |= BIT(id); 1354 /* 1355 * WRITE_ONCE pairs with READ_ONCE in 1356 * xe_vm_has_valid_gpu_mapping(). 1357 * Must not fail after setting 1358 * tile_invalidated and before 1359 * TLB invalidation. 1360 */ 1361 WRITE_ONCE(range->tile_invalidated, 1362 range->tile_invalidated | BIT(id)); 1363 } 1364 } 1365 } 1366 } 1367 1368 return tile_mask; 1369 } 1370 1371 #if IS_ENABLED(CONFIG_DRM_XE_PAGEMAP) 1372 1373 static struct drm_pagemap *tile_local_pagemap(struct xe_tile *tile) 1374 { 1375 return &tile->mem.vram->dpagemap; 1376 } 1377 1378 /** 1379 * xe_vma_resolve_pagemap - Resolve the appropriate DRM pagemap for a VMA 1380 * @vma: Pointer to the xe_vma structure containing memory attributes 1381 * @tile: Pointer to the xe_tile structure used as fallback for VRAM mapping 1382 * 1383 * This function determines the correct DRM pagemap to use for a given VMA. 1384 * It first checks if a valid devmem_fd is provided in the VMA's preferred 1385 * location. If the devmem_fd is negative, it returns NULL, indicating no 1386 * pagemap is available and smem to be used as preferred location. 1387 * If the devmem_fd is equal to the default faulting 1388 * GT identifier, it returns the VRAM pagemap associated with the tile. 1389 * 1390 * Future support for multi-device configurations may use drm_pagemap_from_fd() 1391 * to resolve pagemaps from arbitrary file descriptors. 1392 * 1393 * Return: A pointer to the resolved drm_pagemap, or NULL if none is applicable. 1394 */ 1395 struct drm_pagemap *xe_vma_resolve_pagemap(struct xe_vma *vma, struct xe_tile *tile) 1396 { 1397 s32 fd = (s32)vma->attr.preferred_loc.devmem_fd; 1398 1399 if (fd == DRM_XE_PREFERRED_LOC_DEFAULT_SYSTEM) 1400 return NULL; 1401 1402 if (fd == DRM_XE_PREFERRED_LOC_DEFAULT_DEVICE) 1403 return IS_DGFX(tile_to_xe(tile)) ? tile_local_pagemap(tile) : NULL; 1404 1405 /* TODO: Support multi-device with drm_pagemap_from_fd(fd) */ 1406 return NULL; 1407 } 1408 1409 /** 1410 * xe_svm_alloc_vram()- Allocate device memory pages for range, 1411 * migrating existing data. 1412 * @tile: tile to allocate vram from 1413 * @range: SVM range 1414 * @ctx: DRM GPU SVM context 1415 * 1416 * Return: 0 on success, error code on failure. 1417 */ 1418 int xe_svm_alloc_vram(struct xe_tile *tile, struct xe_svm_range *range, 1419 const struct drm_gpusvm_ctx *ctx) 1420 { 1421 struct drm_pagemap *dpagemap; 1422 1423 xe_assert(tile_to_xe(tile), range->base.pages.flags.migrate_devmem); 1424 range_debug(range, "ALLOCATE VRAM"); 1425 1426 dpagemap = tile_local_pagemap(tile); 1427 return drm_pagemap_populate_mm(dpagemap, xe_svm_range_start(range), 1428 xe_svm_range_end(range), 1429 range->base.gpusvm->mm, 1430 ctx->timeslice_ms); 1431 } 1432 1433 static struct drm_pagemap_addr 1434 xe_drm_pagemap_device_map(struct drm_pagemap *dpagemap, 1435 struct device *dev, 1436 struct page *page, 1437 unsigned int order, 1438 enum dma_data_direction dir) 1439 { 1440 struct device *pgmap_dev = dpagemap->dev; 1441 enum drm_interconnect_protocol prot; 1442 dma_addr_t addr; 1443 1444 if (pgmap_dev == dev) { 1445 addr = xe_vram_region_page_to_dpa(page_to_vr(page), page); 1446 prot = XE_INTERCONNECT_VRAM; 1447 } else { 1448 addr = DMA_MAPPING_ERROR; 1449 prot = 0; 1450 } 1451 1452 return drm_pagemap_addr_encode(addr, prot, order, dir); 1453 } 1454 1455 static const struct drm_pagemap_ops xe_drm_pagemap_ops = { 1456 .device_map = xe_drm_pagemap_device_map, 1457 .populate_mm = xe_drm_pagemap_populate_mm, 1458 }; 1459 1460 /** 1461 * xe_devm_add: Remap and provide memmap backing for device memory 1462 * @tile: tile that the memory region belongs to 1463 * @vr: vram memory region to remap 1464 * 1465 * This remap device memory to host physical address space and create 1466 * struct page to back device memory 1467 * 1468 * Return: 0 on success standard error code otherwise 1469 */ 1470 int xe_devm_add(struct xe_tile *tile, struct xe_vram_region *vr) 1471 { 1472 struct xe_device *xe = tile_to_xe(tile); 1473 struct device *dev = &to_pci_dev(xe->drm.dev)->dev; 1474 struct resource *res; 1475 void *addr; 1476 int ret; 1477 1478 res = devm_request_free_mem_region(dev, &iomem_resource, 1479 vr->usable_size); 1480 if (IS_ERR(res)) { 1481 ret = PTR_ERR(res); 1482 return ret; 1483 } 1484 1485 vr->pagemap.type = MEMORY_DEVICE_PRIVATE; 1486 vr->pagemap.range.start = res->start; 1487 vr->pagemap.range.end = res->end; 1488 vr->pagemap.nr_range = 1; 1489 vr->pagemap.ops = drm_pagemap_pagemap_ops_get(); 1490 vr->pagemap.owner = xe_svm_devm_owner(xe); 1491 addr = devm_memremap_pages(dev, &vr->pagemap); 1492 1493 vr->dpagemap.dev = dev; 1494 vr->dpagemap.ops = &xe_drm_pagemap_ops; 1495 1496 if (IS_ERR(addr)) { 1497 devm_release_mem_region(dev, res->start, resource_size(res)); 1498 ret = PTR_ERR(addr); 1499 drm_err(&xe->drm, "Failed to remap tile %d memory, errno %pe\n", 1500 tile->id, ERR_PTR(ret)); 1501 return ret; 1502 } 1503 vr->hpa_base = res->start; 1504 1505 drm_dbg(&xe->drm, "Added tile %d memory [%llx-%llx] to devm, remapped to %pr\n", 1506 tile->id, vr->io_start, vr->io_start + vr->usable_size, res); 1507 return 0; 1508 } 1509 #else 1510 int xe_svm_alloc_vram(struct xe_tile *tile, 1511 struct xe_svm_range *range, 1512 const struct drm_gpusvm_ctx *ctx) 1513 { 1514 return -EOPNOTSUPP; 1515 } 1516 1517 int xe_devm_add(struct xe_tile *tile, struct xe_vram_region *vr) 1518 { 1519 return 0; 1520 } 1521 1522 struct drm_pagemap *xe_vma_resolve_pagemap(struct xe_vma *vma, struct xe_tile *tile) 1523 { 1524 return NULL; 1525 } 1526 #endif 1527 1528 /** 1529 * xe_svm_flush() - SVM flush 1530 * @vm: The VM. 1531 * 1532 * Flush all SVM actions. 1533 */ 1534 void xe_svm_flush(struct xe_vm *vm) 1535 { 1536 if (xe_vm_in_fault_mode(vm)) 1537 flush_work(&vm->svm.garbage_collector.work); 1538 } 1539