1 /* 2 * Copyright © 2016 Intel Corporation 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 21 * IN THE SOFTWARE. 22 * 23 */ 24 25 #include <linux/list_sort.h> 26 #include <linux/prime_numbers.h> 27 28 #include "gem/i915_gem_context.h" 29 #include "gem/selftests/mock_context.h" 30 #include "gt/intel_context.h" 31 #include "gt/intel_gpu_commands.h" 32 33 #include "i915_random.h" 34 #include "i915_selftest.h" 35 36 #include "mock_drm.h" 37 #include "mock_gem_device.h" 38 #include "mock_gtt.h" 39 #include "igt_flush_test.h" 40 41 static void cleanup_freed_objects(struct drm_i915_private *i915) 42 { 43 i915_gem_drain_freed_objects(i915); 44 } 45 46 static void fake_free_pages(struct drm_i915_gem_object *obj, 47 struct sg_table *pages) 48 { 49 sg_free_table(pages); 50 kfree(pages); 51 } 52 53 static int fake_get_pages(struct drm_i915_gem_object *obj) 54 { 55 #define GFP (GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY) 56 #define PFN_BIAS 0x1000 57 struct sg_table *pages; 58 struct scatterlist *sg; 59 unsigned int sg_page_sizes; 60 typeof(obj->base.size) rem; 61 62 pages = kmalloc(sizeof(*pages), GFP); 63 if (!pages) 64 return -ENOMEM; 65 66 rem = round_up(obj->base.size, BIT(31)) >> 31; 67 if (sg_alloc_table(pages, rem, GFP)) { 68 kfree(pages); 69 return -ENOMEM; 70 } 71 72 sg_page_sizes = 0; 73 rem = obj->base.size; 74 for (sg = pages->sgl; sg; sg = sg_next(sg)) { 75 unsigned long len = min_t(typeof(rem), rem, BIT(31)); 76 77 GEM_BUG_ON(!len); 78 sg_set_page(sg, pfn_to_page(PFN_BIAS), len, 0); 79 sg_dma_address(sg) = page_to_phys(sg_page(sg)); 80 sg_dma_len(sg) = len; 81 sg_page_sizes |= len; 82 83 rem -= len; 84 } 85 GEM_BUG_ON(rem); 86 87 __i915_gem_object_set_pages(obj, pages, sg_page_sizes); 88 89 return 0; 90 #undef GFP 91 } 92 93 static void fake_put_pages(struct drm_i915_gem_object *obj, 94 struct sg_table *pages) 95 { 96 fake_free_pages(obj, pages); 97 obj->mm.dirty = false; 98 } 99 100 static const struct drm_i915_gem_object_ops fake_ops = { 101 .name = "fake-gem", 102 .flags = I915_GEM_OBJECT_IS_SHRINKABLE, 103 .get_pages = fake_get_pages, 104 .put_pages = fake_put_pages, 105 }; 106 107 static struct drm_i915_gem_object * 108 fake_dma_object(struct drm_i915_private *i915, u64 size) 109 { 110 static struct lock_class_key lock_class; 111 struct drm_i915_gem_object *obj; 112 113 GEM_BUG_ON(!size); 114 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE)); 115 116 if (overflows_type(size, obj->base.size)) 117 return ERR_PTR(-E2BIG); 118 119 obj = i915_gem_object_alloc(); 120 if (!obj) 121 goto err; 122 123 drm_gem_private_object_init(&i915->drm, &obj->base, size); 124 i915_gem_object_init(obj, &fake_ops, &lock_class, 0); 125 126 i915_gem_object_set_volatile(obj); 127 128 obj->write_domain = I915_GEM_DOMAIN_CPU; 129 obj->read_domains = I915_GEM_DOMAIN_CPU; 130 obj->cache_level = I915_CACHE_NONE; 131 132 /* Preallocate the "backing storage" */ 133 if (i915_gem_object_pin_pages_unlocked(obj)) 134 goto err_obj; 135 136 i915_gem_object_unpin_pages(obj); 137 return obj; 138 139 err_obj: 140 i915_gem_object_put(obj); 141 err: 142 return ERR_PTR(-ENOMEM); 143 } 144 145 static int igt_ppgtt_alloc(void *arg) 146 { 147 struct drm_i915_private *dev_priv = arg; 148 struct i915_ppgtt *ppgtt; 149 struct i915_gem_ww_ctx ww; 150 u64 size, last, limit; 151 int err = 0; 152 153 /* Allocate a ppggt and try to fill the entire range */ 154 155 if (!HAS_PPGTT(dev_priv)) 156 return 0; 157 158 ppgtt = i915_ppgtt_create(&dev_priv->gt); 159 if (IS_ERR(ppgtt)) 160 return PTR_ERR(ppgtt); 161 162 if (!ppgtt->vm.allocate_va_range) 163 goto err_ppgtt_cleanup; 164 165 /* 166 * While we only allocate the page tables here and so we could 167 * address a much larger GTT than we could actually fit into 168 * RAM, a practical limit is the amount of physical pages in the system. 169 * This should ensure that we do not run into the oomkiller during 170 * the test and take down the machine wilfully. 171 */ 172 limit = totalram_pages() << PAGE_SHIFT; 173 limit = min(ppgtt->vm.total, limit); 174 175 i915_gem_ww_ctx_init(&ww, false); 176 retry: 177 err = i915_vm_lock_objects(&ppgtt->vm, &ww); 178 if (err) 179 goto err_ppgtt_cleanup; 180 181 /* Check we can allocate the entire range */ 182 for (size = 4096; size <= limit; size <<= 2) { 183 struct i915_vm_pt_stash stash = {}; 184 185 err = i915_vm_alloc_pt_stash(&ppgtt->vm, &stash, size); 186 if (err) 187 goto err_ppgtt_cleanup; 188 189 err = i915_vm_pin_pt_stash(&ppgtt->vm, &stash); 190 if (err) { 191 i915_vm_free_pt_stash(&ppgtt->vm, &stash); 192 goto err_ppgtt_cleanup; 193 } 194 195 ppgtt->vm.allocate_va_range(&ppgtt->vm, &stash, 0, size); 196 cond_resched(); 197 198 ppgtt->vm.clear_range(&ppgtt->vm, 0, size); 199 200 i915_vm_free_pt_stash(&ppgtt->vm, &stash); 201 } 202 203 /* Check we can incrementally allocate the entire range */ 204 for (last = 0, size = 4096; size <= limit; last = size, size <<= 2) { 205 struct i915_vm_pt_stash stash = {}; 206 207 err = i915_vm_alloc_pt_stash(&ppgtt->vm, &stash, size - last); 208 if (err) 209 goto err_ppgtt_cleanup; 210 211 err = i915_vm_pin_pt_stash(&ppgtt->vm, &stash); 212 if (err) { 213 i915_vm_free_pt_stash(&ppgtt->vm, &stash); 214 goto err_ppgtt_cleanup; 215 } 216 217 ppgtt->vm.allocate_va_range(&ppgtt->vm, &stash, 218 last, size - last); 219 cond_resched(); 220 221 i915_vm_free_pt_stash(&ppgtt->vm, &stash); 222 } 223 224 err_ppgtt_cleanup: 225 if (err == -EDEADLK) { 226 err = i915_gem_ww_ctx_backoff(&ww); 227 if (!err) 228 goto retry; 229 } 230 i915_gem_ww_ctx_fini(&ww); 231 232 i915_vm_put(&ppgtt->vm); 233 return err; 234 } 235 236 static int lowlevel_hole(struct i915_address_space *vm, 237 u64 hole_start, u64 hole_end, 238 unsigned long end_time) 239 { 240 I915_RND_STATE(seed_prng); 241 struct i915_vma *mock_vma; 242 unsigned int size; 243 244 mock_vma = kzalloc(sizeof(*mock_vma), GFP_KERNEL); 245 if (!mock_vma) 246 return -ENOMEM; 247 248 /* Keep creating larger objects until one cannot fit into the hole */ 249 for (size = 12; (hole_end - hole_start) >> size; size++) { 250 I915_RND_SUBSTATE(prng, seed_prng); 251 struct drm_i915_gem_object *obj; 252 unsigned int *order, count, n; 253 u64 hole_size; 254 255 hole_size = (hole_end - hole_start) >> size; 256 if (hole_size > KMALLOC_MAX_SIZE / sizeof(u32)) 257 hole_size = KMALLOC_MAX_SIZE / sizeof(u32); 258 count = hole_size >> 1; 259 if (!count) { 260 pr_debug("%s: hole is too small [%llx - %llx] >> %d: %lld\n", 261 __func__, hole_start, hole_end, size, hole_size); 262 break; 263 } 264 265 do { 266 order = i915_random_order(count, &prng); 267 if (order) 268 break; 269 } while (count >>= 1); 270 if (!count) { 271 kfree(mock_vma); 272 return -ENOMEM; 273 } 274 GEM_BUG_ON(!order); 275 276 GEM_BUG_ON(count * BIT_ULL(size) > vm->total); 277 GEM_BUG_ON(hole_start + count * BIT_ULL(size) > hole_end); 278 279 /* Ignore allocation failures (i.e. don't report them as 280 * a test failure) as we are purposefully allocating very 281 * large objects without checking that we have sufficient 282 * memory. We expect to hit -ENOMEM. 283 */ 284 285 obj = fake_dma_object(vm->i915, BIT_ULL(size)); 286 if (IS_ERR(obj)) { 287 kfree(order); 288 break; 289 } 290 291 GEM_BUG_ON(obj->base.size != BIT_ULL(size)); 292 293 if (i915_gem_object_pin_pages_unlocked(obj)) { 294 i915_gem_object_put(obj); 295 kfree(order); 296 break; 297 } 298 299 for (n = 0; n < count; n++) { 300 u64 addr = hole_start + order[n] * BIT_ULL(size); 301 intel_wakeref_t wakeref; 302 303 GEM_BUG_ON(addr + BIT_ULL(size) > vm->total); 304 305 if (igt_timeout(end_time, 306 "%s timed out before %d/%d\n", 307 __func__, n, count)) { 308 hole_end = hole_start; /* quit */ 309 break; 310 } 311 312 if (vm->allocate_va_range) { 313 struct i915_vm_pt_stash stash = {}; 314 struct i915_gem_ww_ctx ww; 315 int err; 316 317 i915_gem_ww_ctx_init(&ww, false); 318 retry: 319 err = i915_vm_lock_objects(vm, &ww); 320 if (err) 321 goto alloc_vm_end; 322 323 err = -ENOMEM; 324 if (i915_vm_alloc_pt_stash(vm, &stash, 325 BIT_ULL(size))) 326 goto alloc_vm_end; 327 328 err = i915_vm_pin_pt_stash(vm, &stash); 329 if (!err) 330 vm->allocate_va_range(vm, &stash, 331 addr, BIT_ULL(size)); 332 333 i915_vm_free_pt_stash(vm, &stash); 334 alloc_vm_end: 335 if (err == -EDEADLK) { 336 err = i915_gem_ww_ctx_backoff(&ww); 337 if (!err) 338 goto retry; 339 } 340 i915_gem_ww_ctx_fini(&ww); 341 342 if (err) 343 break; 344 } 345 346 mock_vma->pages = obj->mm.pages; 347 mock_vma->node.size = BIT_ULL(size); 348 mock_vma->node.start = addr; 349 350 with_intel_runtime_pm(vm->gt->uncore->rpm, wakeref) 351 vm->insert_entries(vm, mock_vma, 352 I915_CACHE_NONE, 0); 353 } 354 count = n; 355 356 i915_random_reorder(order, count, &prng); 357 for (n = 0; n < count; n++) { 358 u64 addr = hole_start + order[n] * BIT_ULL(size); 359 intel_wakeref_t wakeref; 360 361 GEM_BUG_ON(addr + BIT_ULL(size) > vm->total); 362 with_intel_runtime_pm(vm->gt->uncore->rpm, wakeref) 363 vm->clear_range(vm, addr, BIT_ULL(size)); 364 } 365 366 i915_gem_object_unpin_pages(obj); 367 i915_gem_object_put(obj); 368 369 kfree(order); 370 371 cleanup_freed_objects(vm->i915); 372 } 373 374 kfree(mock_vma); 375 return 0; 376 } 377 378 static void close_object_list(struct list_head *objects, 379 struct i915_address_space *vm) 380 { 381 struct drm_i915_gem_object *obj, *on; 382 int ignored; 383 384 list_for_each_entry_safe(obj, on, objects, st_link) { 385 struct i915_vma *vma; 386 387 vma = i915_vma_instance(obj, vm, NULL); 388 if (!IS_ERR(vma)) 389 ignored = i915_vma_unbind(vma); 390 391 list_del(&obj->st_link); 392 i915_gem_object_put(obj); 393 } 394 } 395 396 static int fill_hole(struct i915_address_space *vm, 397 u64 hole_start, u64 hole_end, 398 unsigned long end_time) 399 { 400 const u64 hole_size = hole_end - hole_start; 401 struct drm_i915_gem_object *obj; 402 const unsigned long max_pages = 403 min_t(u64, ULONG_MAX - 1, hole_size/2 >> PAGE_SHIFT); 404 const unsigned long max_step = max(int_sqrt(max_pages), 2UL); 405 unsigned long npages, prime, flags; 406 struct i915_vma *vma; 407 LIST_HEAD(objects); 408 int err; 409 410 /* Try binding many VMA working inwards from either edge */ 411 412 flags = PIN_OFFSET_FIXED | PIN_USER; 413 if (i915_is_ggtt(vm)) 414 flags |= PIN_GLOBAL; 415 416 for_each_prime_number_from(prime, 2, max_step) { 417 for (npages = 1; npages <= max_pages; npages *= prime) { 418 const u64 full_size = npages << PAGE_SHIFT; 419 const struct { 420 const char *name; 421 u64 offset; 422 int step; 423 } phases[] = { 424 { "top-down", hole_end, -1, }, 425 { "bottom-up", hole_start, 1, }, 426 { } 427 }, *p; 428 429 obj = fake_dma_object(vm->i915, full_size); 430 if (IS_ERR(obj)) 431 break; 432 433 list_add(&obj->st_link, &objects); 434 435 /* Align differing sized objects against the edges, and 436 * check we don't walk off into the void when binding 437 * them into the GTT. 438 */ 439 for (p = phases; p->name; p++) { 440 u64 offset; 441 442 offset = p->offset; 443 list_for_each_entry(obj, &objects, st_link) { 444 vma = i915_vma_instance(obj, vm, NULL); 445 if (IS_ERR(vma)) 446 continue; 447 448 if (p->step < 0) { 449 if (offset < hole_start + obj->base.size) 450 break; 451 offset -= obj->base.size; 452 } 453 454 err = i915_vma_pin(vma, 0, 0, offset | flags); 455 if (err) { 456 pr_err("%s(%s) pin (forward) failed with err=%d on size=%lu pages (prime=%lu), offset=%llx\n", 457 __func__, p->name, err, npages, prime, offset); 458 goto err; 459 } 460 461 if (!drm_mm_node_allocated(&vma->node) || 462 i915_vma_misplaced(vma, 0, 0, offset | flags)) { 463 pr_err("%s(%s) (forward) insert failed: vma.node=%llx + %llx [allocated? %d], expected offset %llx\n", 464 __func__, p->name, vma->node.start, vma->node.size, drm_mm_node_allocated(&vma->node), 465 offset); 466 err = -EINVAL; 467 goto err; 468 } 469 470 i915_vma_unpin(vma); 471 472 if (p->step > 0) { 473 if (offset + obj->base.size > hole_end) 474 break; 475 offset += obj->base.size; 476 } 477 } 478 479 offset = p->offset; 480 list_for_each_entry(obj, &objects, st_link) { 481 vma = i915_vma_instance(obj, vm, NULL); 482 if (IS_ERR(vma)) 483 continue; 484 485 if (p->step < 0) { 486 if (offset < hole_start + obj->base.size) 487 break; 488 offset -= obj->base.size; 489 } 490 491 if (!drm_mm_node_allocated(&vma->node) || 492 i915_vma_misplaced(vma, 0, 0, offset | flags)) { 493 pr_err("%s(%s) (forward) moved vma.node=%llx + %llx, expected offset %llx\n", 494 __func__, p->name, vma->node.start, vma->node.size, 495 offset); 496 err = -EINVAL; 497 goto err; 498 } 499 500 err = i915_vma_unbind(vma); 501 if (err) { 502 pr_err("%s(%s) (forward) unbind of vma.node=%llx + %llx failed with err=%d\n", 503 __func__, p->name, vma->node.start, vma->node.size, 504 err); 505 goto err; 506 } 507 508 if (p->step > 0) { 509 if (offset + obj->base.size > hole_end) 510 break; 511 offset += obj->base.size; 512 } 513 } 514 515 offset = p->offset; 516 list_for_each_entry_reverse(obj, &objects, st_link) { 517 vma = i915_vma_instance(obj, vm, NULL); 518 if (IS_ERR(vma)) 519 continue; 520 521 if (p->step < 0) { 522 if (offset < hole_start + obj->base.size) 523 break; 524 offset -= obj->base.size; 525 } 526 527 err = i915_vma_pin(vma, 0, 0, offset | flags); 528 if (err) { 529 pr_err("%s(%s) pin (backward) failed with err=%d on size=%lu pages (prime=%lu), offset=%llx\n", 530 __func__, p->name, err, npages, prime, offset); 531 goto err; 532 } 533 534 if (!drm_mm_node_allocated(&vma->node) || 535 i915_vma_misplaced(vma, 0, 0, offset | flags)) { 536 pr_err("%s(%s) (backward) insert failed: vma.node=%llx + %llx [allocated? %d], expected offset %llx\n", 537 __func__, p->name, vma->node.start, vma->node.size, drm_mm_node_allocated(&vma->node), 538 offset); 539 err = -EINVAL; 540 goto err; 541 } 542 543 i915_vma_unpin(vma); 544 545 if (p->step > 0) { 546 if (offset + obj->base.size > hole_end) 547 break; 548 offset += obj->base.size; 549 } 550 } 551 552 offset = p->offset; 553 list_for_each_entry_reverse(obj, &objects, st_link) { 554 vma = i915_vma_instance(obj, vm, NULL); 555 if (IS_ERR(vma)) 556 continue; 557 558 if (p->step < 0) { 559 if (offset < hole_start + obj->base.size) 560 break; 561 offset -= obj->base.size; 562 } 563 564 if (!drm_mm_node_allocated(&vma->node) || 565 i915_vma_misplaced(vma, 0, 0, offset | flags)) { 566 pr_err("%s(%s) (backward) moved vma.node=%llx + %llx [allocated? %d], expected offset %llx\n", 567 __func__, p->name, vma->node.start, vma->node.size, drm_mm_node_allocated(&vma->node), 568 offset); 569 err = -EINVAL; 570 goto err; 571 } 572 573 err = i915_vma_unbind(vma); 574 if (err) { 575 pr_err("%s(%s) (backward) unbind of vma.node=%llx + %llx failed with err=%d\n", 576 __func__, p->name, vma->node.start, vma->node.size, 577 err); 578 goto err; 579 } 580 581 if (p->step > 0) { 582 if (offset + obj->base.size > hole_end) 583 break; 584 offset += obj->base.size; 585 } 586 } 587 } 588 589 if (igt_timeout(end_time, "%s timed out (npages=%lu, prime=%lu)\n", 590 __func__, npages, prime)) { 591 err = -EINTR; 592 goto err; 593 } 594 } 595 596 close_object_list(&objects, vm); 597 cleanup_freed_objects(vm->i915); 598 } 599 600 return 0; 601 602 err: 603 close_object_list(&objects, vm); 604 return err; 605 } 606 607 static int walk_hole(struct i915_address_space *vm, 608 u64 hole_start, u64 hole_end, 609 unsigned long end_time) 610 { 611 const u64 hole_size = hole_end - hole_start; 612 const unsigned long max_pages = 613 min_t(u64, ULONG_MAX - 1, hole_size >> PAGE_SHIFT); 614 unsigned long flags; 615 u64 size; 616 617 /* Try binding a single VMA in different positions within the hole */ 618 619 flags = PIN_OFFSET_FIXED | PIN_USER; 620 if (i915_is_ggtt(vm)) 621 flags |= PIN_GLOBAL; 622 623 for_each_prime_number_from(size, 1, max_pages) { 624 struct drm_i915_gem_object *obj; 625 struct i915_vma *vma; 626 u64 addr; 627 int err = 0; 628 629 obj = fake_dma_object(vm->i915, size << PAGE_SHIFT); 630 if (IS_ERR(obj)) 631 break; 632 633 vma = i915_vma_instance(obj, vm, NULL); 634 if (IS_ERR(vma)) { 635 err = PTR_ERR(vma); 636 goto err_put; 637 } 638 639 for (addr = hole_start; 640 addr + obj->base.size < hole_end; 641 addr += obj->base.size) { 642 err = i915_vma_pin(vma, 0, 0, addr | flags); 643 if (err) { 644 pr_err("%s bind failed at %llx + %llx [hole %llx- %llx] with err=%d\n", 645 __func__, addr, vma->size, 646 hole_start, hole_end, err); 647 goto err_put; 648 } 649 i915_vma_unpin(vma); 650 651 if (!drm_mm_node_allocated(&vma->node) || 652 i915_vma_misplaced(vma, 0, 0, addr | flags)) { 653 pr_err("%s incorrect at %llx + %llx\n", 654 __func__, addr, vma->size); 655 err = -EINVAL; 656 goto err_put; 657 } 658 659 err = i915_vma_unbind(vma); 660 if (err) { 661 pr_err("%s unbind failed at %llx + %llx with err=%d\n", 662 __func__, addr, vma->size, err); 663 goto err_put; 664 } 665 666 GEM_BUG_ON(drm_mm_node_allocated(&vma->node)); 667 668 if (igt_timeout(end_time, 669 "%s timed out at %llx\n", 670 __func__, addr)) { 671 err = -EINTR; 672 goto err_put; 673 } 674 } 675 676 err_put: 677 i915_gem_object_put(obj); 678 if (err) 679 return err; 680 681 cleanup_freed_objects(vm->i915); 682 } 683 684 return 0; 685 } 686 687 static int pot_hole(struct i915_address_space *vm, 688 u64 hole_start, u64 hole_end, 689 unsigned long end_time) 690 { 691 struct drm_i915_gem_object *obj; 692 struct i915_vma *vma; 693 unsigned long flags; 694 unsigned int pot; 695 int err = 0; 696 697 flags = PIN_OFFSET_FIXED | PIN_USER; 698 if (i915_is_ggtt(vm)) 699 flags |= PIN_GLOBAL; 700 701 obj = i915_gem_object_create_internal(vm->i915, 2 * I915_GTT_PAGE_SIZE); 702 if (IS_ERR(obj)) 703 return PTR_ERR(obj); 704 705 vma = i915_vma_instance(obj, vm, NULL); 706 if (IS_ERR(vma)) { 707 err = PTR_ERR(vma); 708 goto err_obj; 709 } 710 711 /* Insert a pair of pages across every pot boundary within the hole */ 712 for (pot = fls64(hole_end - 1) - 1; 713 pot > ilog2(2 * I915_GTT_PAGE_SIZE); 714 pot--) { 715 u64 step = BIT_ULL(pot); 716 u64 addr; 717 718 for (addr = round_up(hole_start + I915_GTT_PAGE_SIZE, step) - I915_GTT_PAGE_SIZE; 719 addr <= round_down(hole_end - 2*I915_GTT_PAGE_SIZE, step) - I915_GTT_PAGE_SIZE; 720 addr += step) { 721 err = i915_vma_pin(vma, 0, 0, addr | flags); 722 if (err) { 723 pr_err("%s failed to pin object at %llx in hole [%llx - %llx], with err=%d\n", 724 __func__, 725 addr, 726 hole_start, hole_end, 727 err); 728 goto err_obj; 729 } 730 731 if (!drm_mm_node_allocated(&vma->node) || 732 i915_vma_misplaced(vma, 0, 0, addr | flags)) { 733 pr_err("%s incorrect at %llx + %llx\n", 734 __func__, addr, vma->size); 735 i915_vma_unpin(vma); 736 err = i915_vma_unbind(vma); 737 err = -EINVAL; 738 goto err_obj; 739 } 740 741 i915_vma_unpin(vma); 742 err = i915_vma_unbind(vma); 743 GEM_BUG_ON(err); 744 } 745 746 if (igt_timeout(end_time, 747 "%s timed out after %d/%d\n", 748 __func__, pot, fls64(hole_end - 1) - 1)) { 749 err = -EINTR; 750 goto err_obj; 751 } 752 } 753 754 err_obj: 755 i915_gem_object_put(obj); 756 return err; 757 } 758 759 static int drunk_hole(struct i915_address_space *vm, 760 u64 hole_start, u64 hole_end, 761 unsigned long end_time) 762 { 763 I915_RND_STATE(prng); 764 unsigned int size; 765 unsigned long flags; 766 767 flags = PIN_OFFSET_FIXED | PIN_USER; 768 if (i915_is_ggtt(vm)) 769 flags |= PIN_GLOBAL; 770 771 /* Keep creating larger objects until one cannot fit into the hole */ 772 for (size = 12; (hole_end - hole_start) >> size; size++) { 773 struct drm_i915_gem_object *obj; 774 unsigned int *order, count, n; 775 struct i915_vma *vma; 776 u64 hole_size; 777 int err = -ENODEV; 778 779 hole_size = (hole_end - hole_start) >> size; 780 if (hole_size > KMALLOC_MAX_SIZE / sizeof(u32)) 781 hole_size = KMALLOC_MAX_SIZE / sizeof(u32); 782 count = hole_size >> 1; 783 if (!count) { 784 pr_debug("%s: hole is too small [%llx - %llx] >> %d: %lld\n", 785 __func__, hole_start, hole_end, size, hole_size); 786 break; 787 } 788 789 do { 790 order = i915_random_order(count, &prng); 791 if (order) 792 break; 793 } while (count >>= 1); 794 if (!count) 795 return -ENOMEM; 796 GEM_BUG_ON(!order); 797 798 /* Ignore allocation failures (i.e. don't report them as 799 * a test failure) as we are purposefully allocating very 800 * large objects without checking that we have sufficient 801 * memory. We expect to hit -ENOMEM. 802 */ 803 804 obj = fake_dma_object(vm->i915, BIT_ULL(size)); 805 if (IS_ERR(obj)) { 806 kfree(order); 807 break; 808 } 809 810 vma = i915_vma_instance(obj, vm, NULL); 811 if (IS_ERR(vma)) { 812 err = PTR_ERR(vma); 813 goto err_obj; 814 } 815 816 GEM_BUG_ON(vma->size != BIT_ULL(size)); 817 818 for (n = 0; n < count; n++) { 819 u64 addr = hole_start + order[n] * BIT_ULL(size); 820 821 err = i915_vma_pin(vma, 0, 0, addr | flags); 822 if (err) { 823 pr_err("%s failed to pin object at %llx + %llx in hole [%llx - %llx], with err=%d\n", 824 __func__, 825 addr, BIT_ULL(size), 826 hole_start, hole_end, 827 err); 828 goto err_obj; 829 } 830 831 if (!drm_mm_node_allocated(&vma->node) || 832 i915_vma_misplaced(vma, 0, 0, addr | flags)) { 833 pr_err("%s incorrect at %llx + %llx\n", 834 __func__, addr, BIT_ULL(size)); 835 i915_vma_unpin(vma); 836 err = i915_vma_unbind(vma); 837 err = -EINVAL; 838 goto err_obj; 839 } 840 841 i915_vma_unpin(vma); 842 err = i915_vma_unbind(vma); 843 GEM_BUG_ON(err); 844 845 if (igt_timeout(end_time, 846 "%s timed out after %d/%d\n", 847 __func__, n, count)) { 848 err = -EINTR; 849 goto err_obj; 850 } 851 } 852 853 err_obj: 854 i915_gem_object_put(obj); 855 kfree(order); 856 if (err) 857 return err; 858 859 cleanup_freed_objects(vm->i915); 860 } 861 862 return 0; 863 } 864 865 static int __shrink_hole(struct i915_address_space *vm, 866 u64 hole_start, u64 hole_end, 867 unsigned long end_time) 868 { 869 struct drm_i915_gem_object *obj; 870 unsigned long flags = PIN_OFFSET_FIXED | PIN_USER; 871 unsigned int order = 12; 872 LIST_HEAD(objects); 873 int err = 0; 874 u64 addr; 875 876 /* Keep creating larger objects until one cannot fit into the hole */ 877 for (addr = hole_start; addr < hole_end; ) { 878 struct i915_vma *vma; 879 u64 size = BIT_ULL(order++); 880 881 size = min(size, hole_end - addr); 882 obj = fake_dma_object(vm->i915, size); 883 if (IS_ERR(obj)) { 884 err = PTR_ERR(obj); 885 break; 886 } 887 888 list_add(&obj->st_link, &objects); 889 890 vma = i915_vma_instance(obj, vm, NULL); 891 if (IS_ERR(vma)) { 892 err = PTR_ERR(vma); 893 break; 894 } 895 896 GEM_BUG_ON(vma->size != size); 897 898 err = i915_vma_pin(vma, 0, 0, addr | flags); 899 if (err) { 900 pr_err("%s failed to pin object at %llx + %llx in hole [%llx - %llx], with err=%d\n", 901 __func__, addr, size, hole_start, hole_end, err); 902 break; 903 } 904 905 if (!drm_mm_node_allocated(&vma->node) || 906 i915_vma_misplaced(vma, 0, 0, addr | flags)) { 907 pr_err("%s incorrect at %llx + %llx\n", 908 __func__, addr, size); 909 i915_vma_unpin(vma); 910 err = i915_vma_unbind(vma); 911 err = -EINVAL; 912 break; 913 } 914 915 i915_vma_unpin(vma); 916 addr += size; 917 918 /* 919 * Since we are injecting allocation faults at random intervals, 920 * wait for this allocation to complete before we change the 921 * faultinjection. 922 */ 923 err = i915_vma_sync(vma); 924 if (err) 925 break; 926 927 if (igt_timeout(end_time, 928 "%s timed out at ofset %llx [%llx - %llx]\n", 929 __func__, addr, hole_start, hole_end)) { 930 err = -EINTR; 931 break; 932 } 933 } 934 935 close_object_list(&objects, vm); 936 cleanup_freed_objects(vm->i915); 937 return err; 938 } 939 940 static int shrink_hole(struct i915_address_space *vm, 941 u64 hole_start, u64 hole_end, 942 unsigned long end_time) 943 { 944 unsigned long prime; 945 int err; 946 947 vm->fault_attr.probability = 999; 948 atomic_set(&vm->fault_attr.times, -1); 949 950 for_each_prime_number_from(prime, 0, ULONG_MAX - 1) { 951 vm->fault_attr.interval = prime; 952 err = __shrink_hole(vm, hole_start, hole_end, end_time); 953 if (err) 954 break; 955 } 956 957 memset(&vm->fault_attr, 0, sizeof(vm->fault_attr)); 958 959 return err; 960 } 961 962 static int shrink_boom(struct i915_address_space *vm, 963 u64 hole_start, u64 hole_end, 964 unsigned long end_time) 965 { 966 unsigned int sizes[] = { SZ_2M, SZ_1G }; 967 struct drm_i915_gem_object *purge; 968 struct drm_i915_gem_object *explode; 969 int err; 970 int i; 971 972 /* 973 * Catch the case which shrink_hole seems to miss. The setup here 974 * requires invoking the shrinker as we do the alloc_pt/alloc_pd, while 975 * ensuring that all vma assiocated with the respective pd/pdp are 976 * unpinned at the time. 977 */ 978 979 for (i = 0; i < ARRAY_SIZE(sizes); ++i) { 980 unsigned int flags = PIN_USER | PIN_OFFSET_FIXED; 981 unsigned int size = sizes[i]; 982 struct i915_vma *vma; 983 984 purge = fake_dma_object(vm->i915, size); 985 if (IS_ERR(purge)) 986 return PTR_ERR(purge); 987 988 vma = i915_vma_instance(purge, vm, NULL); 989 if (IS_ERR(vma)) { 990 err = PTR_ERR(vma); 991 goto err_purge; 992 } 993 994 err = i915_vma_pin(vma, 0, 0, flags); 995 if (err) 996 goto err_purge; 997 998 /* Should now be ripe for purging */ 999 i915_vma_unpin(vma); 1000 1001 explode = fake_dma_object(vm->i915, size); 1002 if (IS_ERR(explode)) { 1003 err = PTR_ERR(explode); 1004 goto err_purge; 1005 } 1006 1007 vm->fault_attr.probability = 100; 1008 vm->fault_attr.interval = 1; 1009 atomic_set(&vm->fault_attr.times, -1); 1010 1011 vma = i915_vma_instance(explode, vm, NULL); 1012 if (IS_ERR(vma)) { 1013 err = PTR_ERR(vma); 1014 goto err_explode; 1015 } 1016 1017 err = i915_vma_pin(vma, 0, 0, flags | size); 1018 if (err) 1019 goto err_explode; 1020 1021 i915_vma_unpin(vma); 1022 1023 i915_gem_object_put(purge); 1024 i915_gem_object_put(explode); 1025 1026 memset(&vm->fault_attr, 0, sizeof(vm->fault_attr)); 1027 cleanup_freed_objects(vm->i915); 1028 } 1029 1030 return 0; 1031 1032 err_explode: 1033 i915_gem_object_put(explode); 1034 err_purge: 1035 i915_gem_object_put(purge); 1036 memset(&vm->fault_attr, 0, sizeof(vm->fault_attr)); 1037 return err; 1038 } 1039 1040 static int exercise_ppgtt(struct drm_i915_private *dev_priv, 1041 int (*func)(struct i915_address_space *vm, 1042 u64 hole_start, u64 hole_end, 1043 unsigned long end_time)) 1044 { 1045 struct i915_ppgtt *ppgtt; 1046 IGT_TIMEOUT(end_time); 1047 struct file *file; 1048 int err; 1049 1050 if (!HAS_FULL_PPGTT(dev_priv)) 1051 return 0; 1052 1053 file = mock_file(dev_priv); 1054 if (IS_ERR(file)) 1055 return PTR_ERR(file); 1056 1057 ppgtt = i915_ppgtt_create(&dev_priv->gt); 1058 if (IS_ERR(ppgtt)) { 1059 err = PTR_ERR(ppgtt); 1060 goto out_free; 1061 } 1062 GEM_BUG_ON(offset_in_page(ppgtt->vm.total)); 1063 GEM_BUG_ON(!atomic_read(&ppgtt->vm.open)); 1064 1065 err = func(&ppgtt->vm, 0, ppgtt->vm.total, end_time); 1066 1067 i915_vm_put(&ppgtt->vm); 1068 1069 out_free: 1070 fput(file); 1071 return err; 1072 } 1073 1074 static int igt_ppgtt_fill(void *arg) 1075 { 1076 return exercise_ppgtt(arg, fill_hole); 1077 } 1078 1079 static int igt_ppgtt_walk(void *arg) 1080 { 1081 return exercise_ppgtt(arg, walk_hole); 1082 } 1083 1084 static int igt_ppgtt_pot(void *arg) 1085 { 1086 return exercise_ppgtt(arg, pot_hole); 1087 } 1088 1089 static int igt_ppgtt_drunk(void *arg) 1090 { 1091 return exercise_ppgtt(arg, drunk_hole); 1092 } 1093 1094 static int igt_ppgtt_lowlevel(void *arg) 1095 { 1096 return exercise_ppgtt(arg, lowlevel_hole); 1097 } 1098 1099 static int igt_ppgtt_shrink(void *arg) 1100 { 1101 return exercise_ppgtt(arg, shrink_hole); 1102 } 1103 1104 static int igt_ppgtt_shrink_boom(void *arg) 1105 { 1106 return exercise_ppgtt(arg, shrink_boom); 1107 } 1108 1109 static int sort_holes(void *priv, struct list_head *A, struct list_head *B) 1110 { 1111 struct drm_mm_node *a = list_entry(A, typeof(*a), hole_stack); 1112 struct drm_mm_node *b = list_entry(B, typeof(*b), hole_stack); 1113 1114 if (a->start < b->start) 1115 return -1; 1116 else 1117 return 1; 1118 } 1119 1120 static int exercise_ggtt(struct drm_i915_private *i915, 1121 int (*func)(struct i915_address_space *vm, 1122 u64 hole_start, u64 hole_end, 1123 unsigned long end_time)) 1124 { 1125 struct i915_ggtt *ggtt = &i915->ggtt; 1126 u64 hole_start, hole_end, last = 0; 1127 struct drm_mm_node *node; 1128 IGT_TIMEOUT(end_time); 1129 int err = 0; 1130 1131 restart: 1132 list_sort(NULL, &ggtt->vm.mm.hole_stack, sort_holes); 1133 drm_mm_for_each_hole(node, &ggtt->vm.mm, hole_start, hole_end) { 1134 if (hole_start < last) 1135 continue; 1136 1137 if (ggtt->vm.mm.color_adjust) 1138 ggtt->vm.mm.color_adjust(node, 0, 1139 &hole_start, &hole_end); 1140 if (hole_start >= hole_end) 1141 continue; 1142 1143 err = func(&ggtt->vm, hole_start, hole_end, end_time); 1144 if (err) 1145 break; 1146 1147 /* As we have manipulated the drm_mm, the list may be corrupt */ 1148 last = hole_end; 1149 goto restart; 1150 } 1151 1152 return err; 1153 } 1154 1155 static int igt_ggtt_fill(void *arg) 1156 { 1157 return exercise_ggtt(arg, fill_hole); 1158 } 1159 1160 static int igt_ggtt_walk(void *arg) 1161 { 1162 return exercise_ggtt(arg, walk_hole); 1163 } 1164 1165 static int igt_ggtt_pot(void *arg) 1166 { 1167 return exercise_ggtt(arg, pot_hole); 1168 } 1169 1170 static int igt_ggtt_drunk(void *arg) 1171 { 1172 return exercise_ggtt(arg, drunk_hole); 1173 } 1174 1175 static int igt_ggtt_lowlevel(void *arg) 1176 { 1177 return exercise_ggtt(arg, lowlevel_hole); 1178 } 1179 1180 static int igt_ggtt_page(void *arg) 1181 { 1182 const unsigned int count = PAGE_SIZE/sizeof(u32); 1183 I915_RND_STATE(prng); 1184 struct drm_i915_private *i915 = arg; 1185 struct i915_ggtt *ggtt = &i915->ggtt; 1186 struct drm_i915_gem_object *obj; 1187 intel_wakeref_t wakeref; 1188 struct drm_mm_node tmp; 1189 unsigned int *order, n; 1190 int err; 1191 1192 if (!i915_ggtt_has_aperture(ggtt)) 1193 return 0; 1194 1195 obj = i915_gem_object_create_internal(i915, PAGE_SIZE); 1196 if (IS_ERR(obj)) 1197 return PTR_ERR(obj); 1198 1199 err = i915_gem_object_pin_pages_unlocked(obj); 1200 if (err) 1201 goto out_free; 1202 1203 memset(&tmp, 0, sizeof(tmp)); 1204 mutex_lock(&ggtt->vm.mutex); 1205 err = drm_mm_insert_node_in_range(&ggtt->vm.mm, &tmp, 1206 count * PAGE_SIZE, 0, 1207 I915_COLOR_UNEVICTABLE, 1208 0, ggtt->mappable_end, 1209 DRM_MM_INSERT_LOW); 1210 mutex_unlock(&ggtt->vm.mutex); 1211 if (err) 1212 goto out_unpin; 1213 1214 wakeref = intel_runtime_pm_get(&i915->runtime_pm); 1215 1216 for (n = 0; n < count; n++) { 1217 u64 offset = tmp.start + n * PAGE_SIZE; 1218 1219 ggtt->vm.insert_page(&ggtt->vm, 1220 i915_gem_object_get_dma_address(obj, 0), 1221 offset, I915_CACHE_NONE, 0); 1222 } 1223 1224 order = i915_random_order(count, &prng); 1225 if (!order) { 1226 err = -ENOMEM; 1227 goto out_remove; 1228 } 1229 1230 for (n = 0; n < count; n++) { 1231 u64 offset = tmp.start + order[n] * PAGE_SIZE; 1232 u32 __iomem *vaddr; 1233 1234 vaddr = io_mapping_map_atomic_wc(&ggtt->iomap, offset); 1235 iowrite32(n, vaddr + n); 1236 io_mapping_unmap_atomic(vaddr); 1237 } 1238 intel_gt_flush_ggtt_writes(ggtt->vm.gt); 1239 1240 i915_random_reorder(order, count, &prng); 1241 for (n = 0; n < count; n++) { 1242 u64 offset = tmp.start + order[n] * PAGE_SIZE; 1243 u32 __iomem *vaddr; 1244 u32 val; 1245 1246 vaddr = io_mapping_map_atomic_wc(&ggtt->iomap, offset); 1247 val = ioread32(vaddr + n); 1248 io_mapping_unmap_atomic(vaddr); 1249 1250 if (val != n) { 1251 pr_err("insert page failed: found %d, expected %d\n", 1252 val, n); 1253 err = -EINVAL; 1254 break; 1255 } 1256 } 1257 1258 kfree(order); 1259 out_remove: 1260 ggtt->vm.clear_range(&ggtt->vm, tmp.start, tmp.size); 1261 intel_runtime_pm_put(&i915->runtime_pm, wakeref); 1262 mutex_lock(&ggtt->vm.mutex); 1263 drm_mm_remove_node(&tmp); 1264 mutex_unlock(&ggtt->vm.mutex); 1265 out_unpin: 1266 i915_gem_object_unpin_pages(obj); 1267 out_free: 1268 i915_gem_object_put(obj); 1269 return err; 1270 } 1271 1272 static void track_vma_bind(struct i915_vma *vma) 1273 { 1274 struct drm_i915_gem_object *obj = vma->obj; 1275 1276 __i915_gem_object_pin_pages(obj); 1277 1278 GEM_BUG_ON(vma->pages); 1279 atomic_set(&vma->pages_count, I915_VMA_PAGES_ACTIVE); 1280 __i915_gem_object_pin_pages(obj); 1281 vma->pages = obj->mm.pages; 1282 1283 mutex_lock(&vma->vm->mutex); 1284 list_add_tail(&vma->vm_link, &vma->vm->bound_list); 1285 mutex_unlock(&vma->vm->mutex); 1286 } 1287 1288 static int exercise_mock(struct drm_i915_private *i915, 1289 int (*func)(struct i915_address_space *vm, 1290 u64 hole_start, u64 hole_end, 1291 unsigned long end_time)) 1292 { 1293 const u64 limit = totalram_pages() << PAGE_SHIFT; 1294 struct i915_address_space *vm; 1295 struct i915_gem_context *ctx; 1296 IGT_TIMEOUT(end_time); 1297 int err; 1298 1299 ctx = mock_context(i915, "mock"); 1300 if (!ctx) 1301 return -ENOMEM; 1302 1303 vm = i915_gem_context_get_vm_rcu(ctx); 1304 err = func(vm, 0, min(vm->total, limit), end_time); 1305 i915_vm_put(vm); 1306 1307 mock_context_close(ctx); 1308 return err; 1309 } 1310 1311 static int igt_mock_fill(void *arg) 1312 { 1313 struct i915_ggtt *ggtt = arg; 1314 1315 return exercise_mock(ggtt->vm.i915, fill_hole); 1316 } 1317 1318 static int igt_mock_walk(void *arg) 1319 { 1320 struct i915_ggtt *ggtt = arg; 1321 1322 return exercise_mock(ggtt->vm.i915, walk_hole); 1323 } 1324 1325 static int igt_mock_pot(void *arg) 1326 { 1327 struct i915_ggtt *ggtt = arg; 1328 1329 return exercise_mock(ggtt->vm.i915, pot_hole); 1330 } 1331 1332 static int igt_mock_drunk(void *arg) 1333 { 1334 struct i915_ggtt *ggtt = arg; 1335 1336 return exercise_mock(ggtt->vm.i915, drunk_hole); 1337 } 1338 1339 static int igt_gtt_reserve(void *arg) 1340 { 1341 struct i915_ggtt *ggtt = arg; 1342 struct drm_i915_gem_object *obj, *on; 1343 I915_RND_STATE(prng); 1344 LIST_HEAD(objects); 1345 u64 total; 1346 int err = -ENODEV; 1347 1348 /* i915_gem_gtt_reserve() tries to reserve the precise range 1349 * for the node, and evicts if it has to. So our test checks that 1350 * it can give us the requsted space and prevent overlaps. 1351 */ 1352 1353 /* Start by filling the GGTT */ 1354 for (total = 0; 1355 total + 2 * I915_GTT_PAGE_SIZE <= ggtt->vm.total; 1356 total += 2 * I915_GTT_PAGE_SIZE) { 1357 struct i915_vma *vma; 1358 1359 obj = i915_gem_object_create_internal(ggtt->vm.i915, 1360 2 * PAGE_SIZE); 1361 if (IS_ERR(obj)) { 1362 err = PTR_ERR(obj); 1363 goto out; 1364 } 1365 1366 err = i915_gem_object_pin_pages_unlocked(obj); 1367 if (err) { 1368 i915_gem_object_put(obj); 1369 goto out; 1370 } 1371 1372 list_add(&obj->st_link, &objects); 1373 1374 vma = i915_vma_instance(obj, &ggtt->vm, NULL); 1375 if (IS_ERR(vma)) { 1376 err = PTR_ERR(vma); 1377 goto out; 1378 } 1379 1380 mutex_lock(&ggtt->vm.mutex); 1381 err = i915_gem_gtt_reserve(&ggtt->vm, &vma->node, 1382 obj->base.size, 1383 total, 1384 obj->cache_level, 1385 0); 1386 mutex_unlock(&ggtt->vm.mutex); 1387 if (err) { 1388 pr_err("i915_gem_gtt_reserve (pass 1) failed at %llu/%llu with err=%d\n", 1389 total, ggtt->vm.total, err); 1390 goto out; 1391 } 1392 track_vma_bind(vma); 1393 1394 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 1395 if (vma->node.start != total || 1396 vma->node.size != 2*I915_GTT_PAGE_SIZE) { 1397 pr_err("i915_gem_gtt_reserve (pass 1) placement failed, found (%llx + %llx), expected (%llx + %llx)\n", 1398 vma->node.start, vma->node.size, 1399 total, 2*I915_GTT_PAGE_SIZE); 1400 err = -EINVAL; 1401 goto out; 1402 } 1403 } 1404 1405 /* Now we start forcing evictions */ 1406 for (total = I915_GTT_PAGE_SIZE; 1407 total + 2 * I915_GTT_PAGE_SIZE <= ggtt->vm.total; 1408 total += 2 * I915_GTT_PAGE_SIZE) { 1409 struct i915_vma *vma; 1410 1411 obj = i915_gem_object_create_internal(ggtt->vm.i915, 1412 2 * PAGE_SIZE); 1413 if (IS_ERR(obj)) { 1414 err = PTR_ERR(obj); 1415 goto out; 1416 } 1417 1418 err = i915_gem_object_pin_pages_unlocked(obj); 1419 if (err) { 1420 i915_gem_object_put(obj); 1421 goto out; 1422 } 1423 1424 list_add(&obj->st_link, &objects); 1425 1426 vma = i915_vma_instance(obj, &ggtt->vm, NULL); 1427 if (IS_ERR(vma)) { 1428 err = PTR_ERR(vma); 1429 goto out; 1430 } 1431 1432 mutex_lock(&ggtt->vm.mutex); 1433 err = i915_gem_gtt_reserve(&ggtt->vm, &vma->node, 1434 obj->base.size, 1435 total, 1436 obj->cache_level, 1437 0); 1438 mutex_unlock(&ggtt->vm.mutex); 1439 if (err) { 1440 pr_err("i915_gem_gtt_reserve (pass 2) failed at %llu/%llu with err=%d\n", 1441 total, ggtt->vm.total, err); 1442 goto out; 1443 } 1444 track_vma_bind(vma); 1445 1446 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 1447 if (vma->node.start != total || 1448 vma->node.size != 2*I915_GTT_PAGE_SIZE) { 1449 pr_err("i915_gem_gtt_reserve (pass 2) placement failed, found (%llx + %llx), expected (%llx + %llx)\n", 1450 vma->node.start, vma->node.size, 1451 total, 2*I915_GTT_PAGE_SIZE); 1452 err = -EINVAL; 1453 goto out; 1454 } 1455 } 1456 1457 /* And then try at random */ 1458 list_for_each_entry_safe(obj, on, &objects, st_link) { 1459 struct i915_vma *vma; 1460 u64 offset; 1461 1462 vma = i915_vma_instance(obj, &ggtt->vm, NULL); 1463 if (IS_ERR(vma)) { 1464 err = PTR_ERR(vma); 1465 goto out; 1466 } 1467 1468 err = i915_vma_unbind(vma); 1469 if (err) { 1470 pr_err("i915_vma_unbind failed with err=%d!\n", err); 1471 goto out; 1472 } 1473 1474 offset = igt_random_offset(&prng, 1475 0, ggtt->vm.total, 1476 2 * I915_GTT_PAGE_SIZE, 1477 I915_GTT_MIN_ALIGNMENT); 1478 1479 mutex_lock(&ggtt->vm.mutex); 1480 err = i915_gem_gtt_reserve(&ggtt->vm, &vma->node, 1481 obj->base.size, 1482 offset, 1483 obj->cache_level, 1484 0); 1485 mutex_unlock(&ggtt->vm.mutex); 1486 if (err) { 1487 pr_err("i915_gem_gtt_reserve (pass 3) failed at %llu/%llu with err=%d\n", 1488 total, ggtt->vm.total, err); 1489 goto out; 1490 } 1491 track_vma_bind(vma); 1492 1493 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 1494 if (vma->node.start != offset || 1495 vma->node.size != 2*I915_GTT_PAGE_SIZE) { 1496 pr_err("i915_gem_gtt_reserve (pass 3) placement failed, found (%llx + %llx), expected (%llx + %llx)\n", 1497 vma->node.start, vma->node.size, 1498 offset, 2*I915_GTT_PAGE_SIZE); 1499 err = -EINVAL; 1500 goto out; 1501 } 1502 } 1503 1504 out: 1505 list_for_each_entry_safe(obj, on, &objects, st_link) { 1506 i915_gem_object_unpin_pages(obj); 1507 i915_gem_object_put(obj); 1508 } 1509 return err; 1510 } 1511 1512 static int igt_gtt_insert(void *arg) 1513 { 1514 struct i915_ggtt *ggtt = arg; 1515 struct drm_i915_gem_object *obj, *on; 1516 struct drm_mm_node tmp = {}; 1517 const struct invalid_insert { 1518 u64 size; 1519 u64 alignment; 1520 u64 start, end; 1521 } invalid_insert[] = { 1522 { 1523 ggtt->vm.total + I915_GTT_PAGE_SIZE, 0, 1524 0, ggtt->vm.total, 1525 }, 1526 { 1527 2*I915_GTT_PAGE_SIZE, 0, 1528 0, I915_GTT_PAGE_SIZE, 1529 }, 1530 { 1531 -(u64)I915_GTT_PAGE_SIZE, 0, 1532 0, 4*I915_GTT_PAGE_SIZE, 1533 }, 1534 { 1535 -(u64)2*I915_GTT_PAGE_SIZE, 2*I915_GTT_PAGE_SIZE, 1536 0, 4*I915_GTT_PAGE_SIZE, 1537 }, 1538 { 1539 I915_GTT_PAGE_SIZE, I915_GTT_MIN_ALIGNMENT << 1, 1540 I915_GTT_MIN_ALIGNMENT, I915_GTT_MIN_ALIGNMENT << 1, 1541 }, 1542 {} 1543 }, *ii; 1544 LIST_HEAD(objects); 1545 u64 total; 1546 int err = -ENODEV; 1547 1548 /* i915_gem_gtt_insert() tries to allocate some free space in the GTT 1549 * to the node, evicting if required. 1550 */ 1551 1552 /* Check a couple of obviously invalid requests */ 1553 for (ii = invalid_insert; ii->size; ii++) { 1554 mutex_lock(&ggtt->vm.mutex); 1555 err = i915_gem_gtt_insert(&ggtt->vm, &tmp, 1556 ii->size, ii->alignment, 1557 I915_COLOR_UNEVICTABLE, 1558 ii->start, ii->end, 1559 0); 1560 mutex_unlock(&ggtt->vm.mutex); 1561 if (err != -ENOSPC) { 1562 pr_err("Invalid i915_gem_gtt_insert(.size=%llx, .alignment=%llx, .start=%llx, .end=%llx) succeeded (err=%d)\n", 1563 ii->size, ii->alignment, ii->start, ii->end, 1564 err); 1565 return -EINVAL; 1566 } 1567 } 1568 1569 /* Start by filling the GGTT */ 1570 for (total = 0; 1571 total + I915_GTT_PAGE_SIZE <= ggtt->vm.total; 1572 total += I915_GTT_PAGE_SIZE) { 1573 struct i915_vma *vma; 1574 1575 obj = i915_gem_object_create_internal(ggtt->vm.i915, 1576 I915_GTT_PAGE_SIZE); 1577 if (IS_ERR(obj)) { 1578 err = PTR_ERR(obj); 1579 goto out; 1580 } 1581 1582 err = i915_gem_object_pin_pages_unlocked(obj); 1583 if (err) { 1584 i915_gem_object_put(obj); 1585 goto out; 1586 } 1587 1588 list_add(&obj->st_link, &objects); 1589 1590 vma = i915_vma_instance(obj, &ggtt->vm, NULL); 1591 if (IS_ERR(vma)) { 1592 err = PTR_ERR(vma); 1593 goto out; 1594 } 1595 1596 mutex_lock(&ggtt->vm.mutex); 1597 err = i915_gem_gtt_insert(&ggtt->vm, &vma->node, 1598 obj->base.size, 0, obj->cache_level, 1599 0, ggtt->vm.total, 1600 0); 1601 mutex_unlock(&ggtt->vm.mutex); 1602 if (err == -ENOSPC) { 1603 /* maxed out the GGTT space */ 1604 i915_gem_object_put(obj); 1605 break; 1606 } 1607 if (err) { 1608 pr_err("i915_gem_gtt_insert (pass 1) failed at %llu/%llu with err=%d\n", 1609 total, ggtt->vm.total, err); 1610 goto out; 1611 } 1612 track_vma_bind(vma); 1613 __i915_vma_pin(vma); 1614 1615 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 1616 } 1617 1618 list_for_each_entry(obj, &objects, st_link) { 1619 struct i915_vma *vma; 1620 1621 vma = i915_vma_instance(obj, &ggtt->vm, NULL); 1622 if (IS_ERR(vma)) { 1623 err = PTR_ERR(vma); 1624 goto out; 1625 } 1626 1627 if (!drm_mm_node_allocated(&vma->node)) { 1628 pr_err("VMA was unexpectedly evicted!\n"); 1629 err = -EINVAL; 1630 goto out; 1631 } 1632 1633 __i915_vma_unpin(vma); 1634 } 1635 1636 /* If we then reinsert, we should find the same hole */ 1637 list_for_each_entry_safe(obj, on, &objects, st_link) { 1638 struct i915_vma *vma; 1639 u64 offset; 1640 1641 vma = i915_vma_instance(obj, &ggtt->vm, NULL); 1642 if (IS_ERR(vma)) { 1643 err = PTR_ERR(vma); 1644 goto out; 1645 } 1646 1647 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 1648 offset = vma->node.start; 1649 1650 err = i915_vma_unbind(vma); 1651 if (err) { 1652 pr_err("i915_vma_unbind failed with err=%d!\n", err); 1653 goto out; 1654 } 1655 1656 mutex_lock(&ggtt->vm.mutex); 1657 err = i915_gem_gtt_insert(&ggtt->vm, &vma->node, 1658 obj->base.size, 0, obj->cache_level, 1659 0, ggtt->vm.total, 1660 0); 1661 mutex_unlock(&ggtt->vm.mutex); 1662 if (err) { 1663 pr_err("i915_gem_gtt_insert (pass 2) failed at %llu/%llu with err=%d\n", 1664 total, ggtt->vm.total, err); 1665 goto out; 1666 } 1667 track_vma_bind(vma); 1668 1669 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 1670 if (vma->node.start != offset) { 1671 pr_err("i915_gem_gtt_insert did not return node to its previous location (the only hole), expected address %llx, found %llx\n", 1672 offset, vma->node.start); 1673 err = -EINVAL; 1674 goto out; 1675 } 1676 } 1677 1678 /* And then force evictions */ 1679 for (total = 0; 1680 total + 2 * I915_GTT_PAGE_SIZE <= ggtt->vm.total; 1681 total += 2 * I915_GTT_PAGE_SIZE) { 1682 struct i915_vma *vma; 1683 1684 obj = i915_gem_object_create_internal(ggtt->vm.i915, 1685 2 * I915_GTT_PAGE_SIZE); 1686 if (IS_ERR(obj)) { 1687 err = PTR_ERR(obj); 1688 goto out; 1689 } 1690 1691 err = i915_gem_object_pin_pages_unlocked(obj); 1692 if (err) { 1693 i915_gem_object_put(obj); 1694 goto out; 1695 } 1696 1697 list_add(&obj->st_link, &objects); 1698 1699 vma = i915_vma_instance(obj, &ggtt->vm, NULL); 1700 if (IS_ERR(vma)) { 1701 err = PTR_ERR(vma); 1702 goto out; 1703 } 1704 1705 mutex_lock(&ggtt->vm.mutex); 1706 err = i915_gem_gtt_insert(&ggtt->vm, &vma->node, 1707 obj->base.size, 0, obj->cache_level, 1708 0, ggtt->vm.total, 1709 0); 1710 mutex_unlock(&ggtt->vm.mutex); 1711 if (err) { 1712 pr_err("i915_gem_gtt_insert (pass 3) failed at %llu/%llu with err=%d\n", 1713 total, ggtt->vm.total, err); 1714 goto out; 1715 } 1716 track_vma_bind(vma); 1717 1718 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 1719 } 1720 1721 out: 1722 list_for_each_entry_safe(obj, on, &objects, st_link) { 1723 i915_gem_object_unpin_pages(obj); 1724 i915_gem_object_put(obj); 1725 } 1726 return err; 1727 } 1728 1729 int i915_gem_gtt_mock_selftests(void) 1730 { 1731 static const struct i915_subtest tests[] = { 1732 SUBTEST(igt_mock_drunk), 1733 SUBTEST(igt_mock_walk), 1734 SUBTEST(igt_mock_pot), 1735 SUBTEST(igt_mock_fill), 1736 SUBTEST(igt_gtt_reserve), 1737 SUBTEST(igt_gtt_insert), 1738 }; 1739 struct drm_i915_private *i915; 1740 struct i915_ggtt *ggtt; 1741 int err; 1742 1743 i915 = mock_gem_device(); 1744 if (!i915) 1745 return -ENOMEM; 1746 1747 ggtt = kmalloc(sizeof(*ggtt), GFP_KERNEL); 1748 if (!ggtt) { 1749 err = -ENOMEM; 1750 goto out_put; 1751 } 1752 mock_init_ggtt(i915, ggtt); 1753 1754 err = i915_subtests(tests, ggtt); 1755 1756 mock_device_flush(i915); 1757 i915_gem_drain_freed_objects(i915); 1758 mock_fini_ggtt(ggtt); 1759 kfree(ggtt); 1760 out_put: 1761 mock_destroy_device(i915); 1762 return err; 1763 } 1764 1765 static int context_sync(struct intel_context *ce) 1766 { 1767 struct i915_request *rq; 1768 long timeout; 1769 1770 rq = intel_context_create_request(ce); 1771 if (IS_ERR(rq)) 1772 return PTR_ERR(rq); 1773 1774 i915_request_get(rq); 1775 i915_request_add(rq); 1776 1777 timeout = i915_request_wait(rq, 0, HZ / 5); 1778 i915_request_put(rq); 1779 1780 return timeout < 0 ? -EIO : 0; 1781 } 1782 1783 static struct i915_request * 1784 submit_batch(struct intel_context *ce, u64 addr) 1785 { 1786 struct i915_request *rq; 1787 int err; 1788 1789 rq = intel_context_create_request(ce); 1790 if (IS_ERR(rq)) 1791 return rq; 1792 1793 err = 0; 1794 if (rq->engine->emit_init_breadcrumb) /* detect a hang */ 1795 err = rq->engine->emit_init_breadcrumb(rq); 1796 if (err == 0) 1797 err = rq->engine->emit_bb_start(rq, addr, 0, 0); 1798 1799 if (err == 0) 1800 i915_request_get(rq); 1801 i915_request_add(rq); 1802 1803 return err ? ERR_PTR(err) : rq; 1804 } 1805 1806 static u32 *spinner(u32 *batch, int i) 1807 { 1808 return batch + i * 64 / sizeof(*batch) + 4; 1809 } 1810 1811 static void end_spin(u32 *batch, int i) 1812 { 1813 *spinner(batch, i) = MI_BATCH_BUFFER_END; 1814 wmb(); 1815 } 1816 1817 static int igt_cs_tlb(void *arg) 1818 { 1819 const unsigned int count = PAGE_SIZE / 64; 1820 const unsigned int chunk_size = count * PAGE_SIZE; 1821 struct drm_i915_private *i915 = arg; 1822 struct drm_i915_gem_object *bbe, *act, *out; 1823 struct i915_gem_engines_iter it; 1824 struct i915_address_space *vm; 1825 struct i915_gem_context *ctx; 1826 struct intel_context *ce; 1827 struct i915_vma *vma; 1828 I915_RND_STATE(prng); 1829 struct file *file; 1830 unsigned int i; 1831 u32 *result; 1832 u32 *batch; 1833 int err = 0; 1834 1835 /* 1836 * Our mission here is to fool the hardware to execute something 1837 * from scratch as it has not seen the batch move (due to missing 1838 * the TLB invalidate). 1839 */ 1840 1841 file = mock_file(i915); 1842 if (IS_ERR(file)) 1843 return PTR_ERR(file); 1844 1845 ctx = live_context(i915, file); 1846 if (IS_ERR(ctx)) { 1847 err = PTR_ERR(ctx); 1848 goto out_unlock; 1849 } 1850 1851 vm = i915_gem_context_get_vm_rcu(ctx); 1852 if (i915_is_ggtt(vm)) 1853 goto out_vm; 1854 1855 /* Create two pages; dummy we prefill the TLB, and intended */ 1856 bbe = i915_gem_object_create_internal(i915, PAGE_SIZE); 1857 if (IS_ERR(bbe)) { 1858 err = PTR_ERR(bbe); 1859 goto out_vm; 1860 } 1861 1862 batch = i915_gem_object_pin_map_unlocked(bbe, I915_MAP_WC); 1863 if (IS_ERR(batch)) { 1864 err = PTR_ERR(batch); 1865 goto out_put_bbe; 1866 } 1867 memset32(batch, MI_BATCH_BUFFER_END, PAGE_SIZE / sizeof(u32)); 1868 i915_gem_object_flush_map(bbe); 1869 i915_gem_object_unpin_map(bbe); 1870 1871 act = i915_gem_object_create_internal(i915, PAGE_SIZE); 1872 if (IS_ERR(act)) { 1873 err = PTR_ERR(act); 1874 goto out_put_bbe; 1875 } 1876 1877 /* Track the execution of each request by writing into different slot */ 1878 batch = i915_gem_object_pin_map_unlocked(act, I915_MAP_WC); 1879 if (IS_ERR(batch)) { 1880 err = PTR_ERR(batch); 1881 goto out_put_act; 1882 } 1883 for (i = 0; i < count; i++) { 1884 u32 *cs = batch + i * 64 / sizeof(*cs); 1885 u64 addr = (vm->total - PAGE_SIZE) + i * sizeof(u32); 1886 1887 GEM_BUG_ON(INTEL_GEN(i915) < 6); 1888 cs[0] = MI_STORE_DWORD_IMM_GEN4; 1889 if (INTEL_GEN(i915) >= 8) { 1890 cs[1] = lower_32_bits(addr); 1891 cs[2] = upper_32_bits(addr); 1892 cs[3] = i; 1893 cs[4] = MI_NOOP; 1894 cs[5] = MI_BATCH_BUFFER_START_GEN8; 1895 } else { 1896 cs[1] = 0; 1897 cs[2] = lower_32_bits(addr); 1898 cs[3] = i; 1899 cs[4] = MI_NOOP; 1900 cs[5] = MI_BATCH_BUFFER_START; 1901 } 1902 } 1903 1904 out = i915_gem_object_create_internal(i915, PAGE_SIZE); 1905 if (IS_ERR(out)) { 1906 err = PTR_ERR(out); 1907 goto out_put_batch; 1908 } 1909 i915_gem_object_set_cache_coherency(out, I915_CACHING_CACHED); 1910 1911 vma = i915_vma_instance(out, vm, NULL); 1912 if (IS_ERR(vma)) { 1913 err = PTR_ERR(vma); 1914 goto out_put_out; 1915 } 1916 1917 err = i915_vma_pin(vma, 0, 0, 1918 PIN_USER | 1919 PIN_OFFSET_FIXED | 1920 (vm->total - PAGE_SIZE)); 1921 if (err) 1922 goto out_put_out; 1923 GEM_BUG_ON(vma->node.start != vm->total - PAGE_SIZE); 1924 1925 result = i915_gem_object_pin_map_unlocked(out, I915_MAP_WB); 1926 if (IS_ERR(result)) { 1927 err = PTR_ERR(result); 1928 goto out_put_out; 1929 } 1930 1931 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) { 1932 IGT_TIMEOUT(end_time); 1933 unsigned long pass = 0; 1934 1935 if (!intel_engine_can_store_dword(ce->engine)) 1936 continue; 1937 1938 while (!__igt_timeout(end_time, NULL)) { 1939 struct i915_vm_pt_stash stash = {}; 1940 struct i915_request *rq; 1941 struct i915_gem_ww_ctx ww; 1942 u64 offset; 1943 1944 offset = igt_random_offset(&prng, 1945 0, vm->total - PAGE_SIZE, 1946 chunk_size, PAGE_SIZE); 1947 1948 memset32(result, STACK_MAGIC, PAGE_SIZE / sizeof(u32)); 1949 1950 vma = i915_vma_instance(bbe, vm, NULL); 1951 if (IS_ERR(vma)) { 1952 err = PTR_ERR(vma); 1953 goto end; 1954 } 1955 1956 err = vma->ops->set_pages(vma); 1957 if (err) 1958 goto end; 1959 1960 i915_gem_ww_ctx_init(&ww, false); 1961 retry: 1962 err = i915_vm_lock_objects(vm, &ww); 1963 if (err) 1964 goto end_ww; 1965 1966 err = i915_vm_alloc_pt_stash(vm, &stash, chunk_size); 1967 if (err) 1968 goto end_ww; 1969 1970 err = i915_vm_pin_pt_stash(vm, &stash); 1971 if (!err) 1972 vm->allocate_va_range(vm, &stash, offset, chunk_size); 1973 1974 i915_vm_free_pt_stash(vm, &stash); 1975 end_ww: 1976 if (err == -EDEADLK) { 1977 err = i915_gem_ww_ctx_backoff(&ww); 1978 if (!err) 1979 goto retry; 1980 } 1981 i915_gem_ww_ctx_fini(&ww); 1982 if (err) 1983 goto end; 1984 1985 /* Prime the TLB with the dummy pages */ 1986 for (i = 0; i < count; i++) { 1987 vma->node.start = offset + i * PAGE_SIZE; 1988 vm->insert_entries(vm, vma, I915_CACHE_NONE, 0); 1989 1990 rq = submit_batch(ce, vma->node.start); 1991 if (IS_ERR(rq)) { 1992 err = PTR_ERR(rq); 1993 goto end; 1994 } 1995 i915_request_put(rq); 1996 } 1997 1998 vma->ops->clear_pages(vma); 1999 2000 err = context_sync(ce); 2001 if (err) { 2002 pr_err("%s: dummy setup timed out\n", 2003 ce->engine->name); 2004 goto end; 2005 } 2006 2007 vma = i915_vma_instance(act, vm, NULL); 2008 if (IS_ERR(vma)) { 2009 err = PTR_ERR(vma); 2010 goto end; 2011 } 2012 2013 err = vma->ops->set_pages(vma); 2014 if (err) 2015 goto end; 2016 2017 /* Replace the TLB with target batches */ 2018 for (i = 0; i < count; i++) { 2019 struct i915_request *rq; 2020 u32 *cs = batch + i * 64 / sizeof(*cs); 2021 u64 addr; 2022 2023 vma->node.start = offset + i * PAGE_SIZE; 2024 vm->insert_entries(vm, vma, I915_CACHE_NONE, 0); 2025 2026 addr = vma->node.start + i * 64; 2027 cs[4] = MI_NOOP; 2028 cs[6] = lower_32_bits(addr); 2029 cs[7] = upper_32_bits(addr); 2030 wmb(); 2031 2032 rq = submit_batch(ce, addr); 2033 if (IS_ERR(rq)) { 2034 err = PTR_ERR(rq); 2035 goto end; 2036 } 2037 2038 /* Wait until the context chain has started */ 2039 if (i == 0) { 2040 while (READ_ONCE(result[i]) && 2041 !i915_request_completed(rq)) 2042 cond_resched(); 2043 } else { 2044 end_spin(batch, i - 1); 2045 } 2046 2047 i915_request_put(rq); 2048 } 2049 end_spin(batch, count - 1); 2050 2051 vma->ops->clear_pages(vma); 2052 2053 err = context_sync(ce); 2054 if (err) { 2055 pr_err("%s: writes timed out\n", 2056 ce->engine->name); 2057 goto end; 2058 } 2059 2060 for (i = 0; i < count; i++) { 2061 if (result[i] != i) { 2062 pr_err("%s: Write lost on pass %lu, at offset %llx, index %d, found %x, expected %x\n", 2063 ce->engine->name, pass, 2064 offset, i, result[i], i); 2065 err = -EINVAL; 2066 goto end; 2067 } 2068 } 2069 2070 vm->clear_range(vm, offset, chunk_size); 2071 pass++; 2072 } 2073 } 2074 end: 2075 if (igt_flush_test(i915)) 2076 err = -EIO; 2077 i915_gem_context_unlock_engines(ctx); 2078 i915_gem_object_unpin_map(out); 2079 out_put_out: 2080 i915_gem_object_put(out); 2081 out_put_batch: 2082 i915_gem_object_unpin_map(act); 2083 out_put_act: 2084 i915_gem_object_put(act); 2085 out_put_bbe: 2086 i915_gem_object_put(bbe); 2087 out_vm: 2088 i915_vm_put(vm); 2089 out_unlock: 2090 fput(file); 2091 return err; 2092 } 2093 2094 int i915_gem_gtt_live_selftests(struct drm_i915_private *i915) 2095 { 2096 static const struct i915_subtest tests[] = { 2097 SUBTEST(igt_ppgtt_alloc), 2098 SUBTEST(igt_ppgtt_lowlevel), 2099 SUBTEST(igt_ppgtt_drunk), 2100 SUBTEST(igt_ppgtt_walk), 2101 SUBTEST(igt_ppgtt_pot), 2102 SUBTEST(igt_ppgtt_fill), 2103 SUBTEST(igt_ppgtt_shrink), 2104 SUBTEST(igt_ppgtt_shrink_boom), 2105 SUBTEST(igt_ggtt_lowlevel), 2106 SUBTEST(igt_ggtt_drunk), 2107 SUBTEST(igt_ggtt_walk), 2108 SUBTEST(igt_ggtt_pot), 2109 SUBTEST(igt_ggtt_fill), 2110 SUBTEST(igt_ggtt_page), 2111 SUBTEST(igt_cs_tlb), 2112 }; 2113 2114 GEM_BUG_ON(offset_in_page(i915->ggtt.vm.total)); 2115 2116 return i915_subtests(tests, i915); 2117 } 2118