1 /* 2 * SPDX-License-Identifier: MIT 3 * 4 * Copyright © 2014-2016 Intel Corporation 5 */ 6 7 #include <linux/pagevec.h> 8 #include <linux/swap.h> 9 10 #include "gem/i915_gem_region.h" 11 #include "i915_drv.h" 12 #include "i915_gemfs.h" 13 #include "i915_gem_object.h" 14 #include "i915_scatterlist.h" 15 #include "i915_trace.h" 16 17 /* 18 * Move pages to appropriate lru and release the pagevec, decrementing the 19 * ref count of those pages. 20 */ 21 static void check_release_pagevec(struct pagevec *pvec) 22 { 23 check_move_unevictable_pages(pvec); 24 __pagevec_release(pvec); 25 cond_resched(); 26 } 27 28 static int shmem_get_pages(struct drm_i915_gem_object *obj) 29 { 30 struct drm_i915_private *i915 = to_i915(obj->base.dev); 31 struct intel_memory_region *mem = obj->mm.region; 32 const unsigned long page_count = obj->base.size / PAGE_SIZE; 33 unsigned long i; 34 struct address_space *mapping; 35 struct sg_table *st; 36 struct scatterlist *sg; 37 struct sgt_iter sgt_iter; 38 struct page *page; 39 unsigned long last_pfn = 0; /* suppress gcc warning */ 40 unsigned int max_segment = i915_sg_segment_size(); 41 unsigned int sg_page_sizes; 42 gfp_t noreclaim; 43 int ret; 44 45 /* 46 * Assert that the object is not currently in any GPU domain. As it 47 * wasn't in the GTT, there shouldn't be any way it could have been in 48 * a GPU cache 49 */ 50 GEM_BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS); 51 GEM_BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS); 52 53 /* 54 * If there's no chance of allocating enough pages for the whole 55 * object, bail early. 56 */ 57 if (obj->base.size > resource_size(&mem->region)) 58 return -ENOMEM; 59 60 st = kmalloc(sizeof(*st), GFP_KERNEL); 61 if (!st) 62 return -ENOMEM; 63 64 rebuild_st: 65 if (sg_alloc_table(st, page_count, GFP_KERNEL)) { 66 kfree(st); 67 return -ENOMEM; 68 } 69 70 /* 71 * Get the list of pages out of our struct file. They'll be pinned 72 * at this point until we release them. 73 * 74 * Fail silently without starting the shrinker 75 */ 76 mapping = obj->base.filp->f_mapping; 77 mapping_set_unevictable(mapping); 78 noreclaim = mapping_gfp_constraint(mapping, ~__GFP_RECLAIM); 79 noreclaim |= __GFP_NORETRY | __GFP_NOWARN; 80 81 sg = st->sgl; 82 st->nents = 0; 83 sg_page_sizes = 0; 84 for (i = 0; i < page_count; i++) { 85 const unsigned int shrink[] = { 86 I915_SHRINK_BOUND | I915_SHRINK_UNBOUND, 87 0, 88 }, *s = shrink; 89 gfp_t gfp = noreclaim; 90 91 do { 92 cond_resched(); 93 page = shmem_read_mapping_page_gfp(mapping, i, gfp); 94 if (!IS_ERR(page)) 95 break; 96 97 if (!*s) { 98 ret = PTR_ERR(page); 99 goto err_sg; 100 } 101 102 i915_gem_shrink(NULL, i915, 2 * page_count, NULL, *s++); 103 104 /* 105 * We've tried hard to allocate the memory by reaping 106 * our own buffer, now let the real VM do its job and 107 * go down in flames if truly OOM. 108 * 109 * However, since graphics tend to be disposable, 110 * defer the oom here by reporting the ENOMEM back 111 * to userspace. 112 */ 113 if (!*s) { 114 /* reclaim and warn, but no oom */ 115 gfp = mapping_gfp_mask(mapping); 116 117 /* 118 * Our bo are always dirty and so we require 119 * kswapd to reclaim our pages (direct reclaim 120 * does not effectively begin pageout of our 121 * buffers on its own). However, direct reclaim 122 * only waits for kswapd when under allocation 123 * congestion. So as a result __GFP_RECLAIM is 124 * unreliable and fails to actually reclaim our 125 * dirty pages -- unless you try over and over 126 * again with !__GFP_NORETRY. However, we still 127 * want to fail this allocation rather than 128 * trigger the out-of-memory killer and for 129 * this we want __GFP_RETRY_MAYFAIL. 130 */ 131 gfp |= __GFP_RETRY_MAYFAIL; 132 } 133 } while (1); 134 135 if (!i || 136 sg->length >= max_segment || 137 page_to_pfn(page) != last_pfn + 1) { 138 if (i) { 139 sg_page_sizes |= sg->length; 140 sg = sg_next(sg); 141 } 142 st->nents++; 143 sg_set_page(sg, page, PAGE_SIZE, 0); 144 } else { 145 sg->length += PAGE_SIZE; 146 } 147 last_pfn = page_to_pfn(page); 148 149 /* Check that the i965g/gm workaround works. */ 150 GEM_BUG_ON(gfp & __GFP_DMA32 && last_pfn >= 0x00100000UL); 151 } 152 if (sg) { /* loop terminated early; short sg table */ 153 sg_page_sizes |= sg->length; 154 sg_mark_end(sg); 155 } 156 157 /* Trim unused sg entries to avoid wasting memory. */ 158 i915_sg_trim(st); 159 160 ret = i915_gem_gtt_prepare_pages(obj, st); 161 if (ret) { 162 /* 163 * DMA remapping failed? One possible cause is that 164 * it could not reserve enough large entries, asking 165 * for PAGE_SIZE chunks instead may be helpful. 166 */ 167 if (max_segment > PAGE_SIZE) { 168 for_each_sgt_page(page, sgt_iter, st) 169 put_page(page); 170 sg_free_table(st); 171 172 max_segment = PAGE_SIZE; 173 goto rebuild_st; 174 } else { 175 dev_warn(i915->drm.dev, 176 "Failed to DMA remap %lu pages\n", 177 page_count); 178 goto err_pages; 179 } 180 } 181 182 if (i915_gem_object_needs_bit17_swizzle(obj)) 183 i915_gem_object_do_bit_17_swizzle(obj, st); 184 185 if (i915_gem_object_can_bypass_llc(obj)) 186 obj->cache_dirty = true; 187 188 __i915_gem_object_set_pages(obj, st, sg_page_sizes); 189 190 return 0; 191 192 err_sg: 193 sg_mark_end(sg); 194 err_pages: 195 mapping_clear_unevictable(mapping); 196 if (sg != st->sgl) { 197 struct pagevec pvec; 198 199 pagevec_init(&pvec); 200 for_each_sgt_page(page, sgt_iter, st) { 201 if (!pagevec_add(&pvec, page)) 202 check_release_pagevec(&pvec); 203 } 204 if (pagevec_count(&pvec)) 205 check_release_pagevec(&pvec); 206 } 207 sg_free_table(st); 208 kfree(st); 209 210 /* 211 * shmemfs first checks if there is enough memory to allocate the page 212 * and reports ENOSPC should there be insufficient, along with the usual 213 * ENOMEM for a genuine allocation failure. 214 * 215 * We use ENOSPC in our driver to mean that we have run out of aperture 216 * space and so want to translate the error from shmemfs back to our 217 * usual understanding of ENOMEM. 218 */ 219 if (ret == -ENOSPC) 220 ret = -ENOMEM; 221 222 return ret; 223 } 224 225 static void 226 shmem_truncate(struct drm_i915_gem_object *obj) 227 { 228 /* 229 * Our goal here is to return as much of the memory as 230 * is possible back to the system as we are called from OOM. 231 * To do this we must instruct the shmfs to drop all of its 232 * backing pages, *now*. 233 */ 234 shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1); 235 obj->mm.madv = __I915_MADV_PURGED; 236 obj->mm.pages = ERR_PTR(-EFAULT); 237 } 238 239 static void 240 shmem_writeback(struct drm_i915_gem_object *obj) 241 { 242 struct address_space *mapping; 243 struct writeback_control wbc = { 244 .sync_mode = WB_SYNC_NONE, 245 .nr_to_write = SWAP_CLUSTER_MAX, 246 .range_start = 0, 247 .range_end = LLONG_MAX, 248 .for_reclaim = 1, 249 }; 250 unsigned long i; 251 252 /* 253 * Leave mmapings intact (GTT will have been revoked on unbinding, 254 * leaving only CPU mmapings around) and add those pages to the LRU 255 * instead of invoking writeback so they are aged and paged out 256 * as normal. 257 */ 258 mapping = obj->base.filp->f_mapping; 259 260 /* Begin writeback on each dirty page */ 261 for (i = 0; i < obj->base.size >> PAGE_SHIFT; i++) { 262 struct page *page; 263 264 page = find_lock_page(mapping, i); 265 if (!page) 266 continue; 267 268 if (!page_mapped(page) && clear_page_dirty_for_io(page)) { 269 int ret; 270 271 SetPageReclaim(page); 272 ret = mapping->a_ops->writepage(page, &wbc); 273 if (!PageWriteback(page)) 274 ClearPageReclaim(page); 275 if (!ret) 276 goto put; 277 } 278 unlock_page(page); 279 put: 280 put_page(page); 281 } 282 } 283 284 void 285 __i915_gem_object_release_shmem(struct drm_i915_gem_object *obj, 286 struct sg_table *pages, 287 bool needs_clflush) 288 { 289 GEM_BUG_ON(obj->mm.madv == __I915_MADV_PURGED); 290 291 if (obj->mm.madv == I915_MADV_DONTNEED) 292 obj->mm.dirty = false; 293 294 if (needs_clflush && 295 (obj->read_domains & I915_GEM_DOMAIN_CPU) == 0 && 296 !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ)) 297 drm_clflush_sg(pages); 298 299 __start_cpu_write(obj); 300 } 301 302 void i915_gem_object_put_pages_shmem(struct drm_i915_gem_object *obj, struct sg_table *pages) 303 { 304 struct sgt_iter sgt_iter; 305 struct pagevec pvec; 306 struct page *page; 307 308 GEM_WARN_ON(IS_DGFX(to_i915(obj->base.dev))); 309 __i915_gem_object_release_shmem(obj, pages, true); 310 311 i915_gem_gtt_finish_pages(obj, pages); 312 313 if (i915_gem_object_needs_bit17_swizzle(obj)) 314 i915_gem_object_save_bit_17_swizzle(obj, pages); 315 316 mapping_clear_unevictable(file_inode(obj->base.filp)->i_mapping); 317 318 pagevec_init(&pvec); 319 for_each_sgt_page(page, sgt_iter, pages) { 320 if (obj->mm.dirty) 321 set_page_dirty(page); 322 323 if (obj->mm.madv == I915_MADV_WILLNEED) 324 mark_page_accessed(page); 325 326 if (!pagevec_add(&pvec, page)) 327 check_release_pagevec(&pvec); 328 } 329 if (pagevec_count(&pvec)) 330 check_release_pagevec(&pvec); 331 obj->mm.dirty = false; 332 333 sg_free_table(pages); 334 kfree(pages); 335 } 336 337 static void 338 shmem_put_pages(struct drm_i915_gem_object *obj, struct sg_table *pages) 339 { 340 if (likely(i915_gem_object_has_struct_page(obj))) 341 i915_gem_object_put_pages_shmem(obj, pages); 342 else 343 i915_gem_object_put_pages_phys(obj, pages); 344 } 345 346 static int 347 shmem_pwrite(struct drm_i915_gem_object *obj, 348 const struct drm_i915_gem_pwrite *arg) 349 { 350 struct address_space *mapping = obj->base.filp->f_mapping; 351 char __user *user_data = u64_to_user_ptr(arg->data_ptr); 352 u64 remain, offset; 353 unsigned int pg; 354 355 /* Caller already validated user args */ 356 GEM_BUG_ON(!access_ok(user_data, arg->size)); 357 358 if (!i915_gem_object_has_struct_page(obj)) 359 return i915_gem_object_pwrite_phys(obj, arg); 360 361 /* 362 * Before we instantiate/pin the backing store for our use, we 363 * can prepopulate the shmemfs filp efficiently using a write into 364 * the pagecache. We avoid the penalty of instantiating all the 365 * pages, important if the user is just writing to a few and never 366 * uses the object on the GPU, and using a direct write into shmemfs 367 * allows it to avoid the cost of retrieving a page (either swapin 368 * or clearing-before-use) before it is overwritten. 369 */ 370 if (i915_gem_object_has_pages(obj)) 371 return -ENODEV; 372 373 if (obj->mm.madv != I915_MADV_WILLNEED) 374 return -EFAULT; 375 376 /* 377 * Before the pages are instantiated the object is treated as being 378 * in the CPU domain. The pages will be clflushed as required before 379 * use, and we can freely write into the pages directly. If userspace 380 * races pwrite with any other operation; corruption will ensue - 381 * that is userspace's prerogative! 382 */ 383 384 remain = arg->size; 385 offset = arg->offset; 386 pg = offset_in_page(offset); 387 388 do { 389 unsigned int len, unwritten; 390 struct page *page; 391 void *data, *vaddr; 392 int err; 393 char c; 394 395 len = PAGE_SIZE - pg; 396 if (len > remain) 397 len = remain; 398 399 /* Prefault the user page to reduce potential recursion */ 400 err = __get_user(c, user_data); 401 if (err) 402 return err; 403 404 err = __get_user(c, user_data + len - 1); 405 if (err) 406 return err; 407 408 err = pagecache_write_begin(obj->base.filp, mapping, 409 offset, len, 0, 410 &page, &data); 411 if (err < 0) 412 return err; 413 414 vaddr = kmap_atomic(page); 415 unwritten = __copy_from_user_inatomic(vaddr + pg, 416 user_data, 417 len); 418 kunmap_atomic(vaddr); 419 420 err = pagecache_write_end(obj->base.filp, mapping, 421 offset, len, len - unwritten, 422 page, data); 423 if (err < 0) 424 return err; 425 426 /* We don't handle -EFAULT, leave it to the caller to check */ 427 if (unwritten) 428 return -ENODEV; 429 430 remain -= len; 431 user_data += len; 432 offset += len; 433 pg = 0; 434 } while (remain); 435 436 return 0; 437 } 438 439 static int 440 shmem_pread(struct drm_i915_gem_object *obj, 441 const struct drm_i915_gem_pread *arg) 442 { 443 if (!i915_gem_object_has_struct_page(obj)) 444 return i915_gem_object_pread_phys(obj, arg); 445 446 return -ENODEV; 447 } 448 449 static void shmem_release(struct drm_i915_gem_object *obj) 450 { 451 if (i915_gem_object_has_struct_page(obj)) 452 i915_gem_object_release_memory_region(obj); 453 454 fput(obj->base.filp); 455 } 456 457 const struct drm_i915_gem_object_ops i915_gem_shmem_ops = { 458 .name = "i915_gem_object_shmem", 459 .flags = I915_GEM_OBJECT_IS_SHRINKABLE, 460 461 .get_pages = shmem_get_pages, 462 .put_pages = shmem_put_pages, 463 .truncate = shmem_truncate, 464 .writeback = shmem_writeback, 465 466 .pwrite = shmem_pwrite, 467 .pread = shmem_pread, 468 469 .release = shmem_release, 470 }; 471 472 static int __create_shmem(struct drm_i915_private *i915, 473 struct drm_gem_object *obj, 474 resource_size_t size) 475 { 476 unsigned long flags = VM_NORESERVE; 477 struct file *filp; 478 479 drm_gem_private_object_init(&i915->drm, obj, size); 480 481 if (i915->mm.gemfs) 482 filp = shmem_file_setup_with_mnt(i915->mm.gemfs, "i915", size, 483 flags); 484 else 485 filp = shmem_file_setup("i915", size, flags); 486 if (IS_ERR(filp)) 487 return PTR_ERR(filp); 488 489 obj->filp = filp; 490 return 0; 491 } 492 493 static int shmem_object_init(struct intel_memory_region *mem, 494 struct drm_i915_gem_object *obj, 495 resource_size_t size, 496 resource_size_t page_size, 497 unsigned int flags) 498 { 499 static struct lock_class_key lock_class; 500 struct drm_i915_private *i915 = mem->i915; 501 struct address_space *mapping; 502 unsigned int cache_level; 503 gfp_t mask; 504 int ret; 505 506 ret = __create_shmem(i915, &obj->base, size); 507 if (ret) 508 return ret; 509 510 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE; 511 if (IS_I965GM(i915) || IS_I965G(i915)) { 512 /* 965gm cannot relocate objects above 4GiB. */ 513 mask &= ~__GFP_HIGHMEM; 514 mask |= __GFP_DMA32; 515 } 516 517 mapping = obj->base.filp->f_mapping; 518 mapping_set_gfp_mask(mapping, mask); 519 GEM_BUG_ON(!(mapping_gfp_mask(mapping) & __GFP_RECLAIM)); 520 521 i915_gem_object_init(obj, &i915_gem_shmem_ops, &lock_class, 0); 522 obj->mem_flags |= I915_BO_FLAG_STRUCT_PAGE; 523 obj->write_domain = I915_GEM_DOMAIN_CPU; 524 obj->read_domains = I915_GEM_DOMAIN_CPU; 525 526 if (HAS_LLC(i915)) 527 /* On some devices, we can have the GPU use the LLC (the CPU 528 * cache) for about a 10% performance improvement 529 * compared to uncached. Graphics requests other than 530 * display scanout are coherent with the CPU in 531 * accessing this cache. This means in this mode we 532 * don't need to clflush on the CPU side, and on the 533 * GPU side we only need to flush internal caches to 534 * get data visible to the CPU. 535 * 536 * However, we maintain the display planes as UC, and so 537 * need to rebind when first used as such. 538 */ 539 cache_level = I915_CACHE_LLC; 540 else 541 cache_level = I915_CACHE_NONE; 542 543 i915_gem_object_set_cache_coherency(obj, cache_level); 544 545 i915_gem_object_init_memory_region(obj, mem); 546 547 return 0; 548 } 549 550 struct drm_i915_gem_object * 551 i915_gem_object_create_shmem(struct drm_i915_private *i915, 552 resource_size_t size) 553 { 554 return i915_gem_object_create_region(i915->mm.regions[INTEL_REGION_SMEM], 555 size, 0, 0); 556 } 557 558 /* Allocate a new GEM object and fill it with the supplied data */ 559 struct drm_i915_gem_object * 560 i915_gem_object_create_shmem_from_data(struct drm_i915_private *dev_priv, 561 const void *data, resource_size_t size) 562 { 563 struct drm_i915_gem_object *obj; 564 struct file *file; 565 resource_size_t offset; 566 int err; 567 568 GEM_WARN_ON(IS_DGFX(dev_priv)); 569 obj = i915_gem_object_create_shmem(dev_priv, round_up(size, PAGE_SIZE)); 570 if (IS_ERR(obj)) 571 return obj; 572 573 GEM_BUG_ON(obj->write_domain != I915_GEM_DOMAIN_CPU); 574 575 file = obj->base.filp; 576 offset = 0; 577 do { 578 unsigned int len = min_t(typeof(size), size, PAGE_SIZE); 579 struct page *page; 580 void *pgdata, *vaddr; 581 582 err = pagecache_write_begin(file, file->f_mapping, 583 offset, len, 0, 584 &page, &pgdata); 585 if (err < 0) 586 goto fail; 587 588 vaddr = kmap(page); 589 memcpy(vaddr, data, len); 590 kunmap(page); 591 592 err = pagecache_write_end(file, file->f_mapping, 593 offset, len, len, 594 page, pgdata); 595 if (err < 0) 596 goto fail; 597 598 size -= len; 599 data += len; 600 offset += len; 601 } while (size); 602 603 return obj; 604 605 fail: 606 i915_gem_object_put(obj); 607 return ERR_PTR(err); 608 } 609 610 static int init_shmem(struct intel_memory_region *mem) 611 { 612 int err; 613 614 err = i915_gemfs_init(mem->i915); 615 if (err) { 616 DRM_NOTE("Unable to create a private tmpfs mount, hugepage support will be disabled(%d).\n", 617 err); 618 } 619 620 intel_memory_region_set_name(mem, "system"); 621 622 return 0; /* Don't error, we can simply fallback to the kernel mnt */ 623 } 624 625 static void release_shmem(struct intel_memory_region *mem) 626 { 627 i915_gemfs_fini(mem->i915); 628 } 629 630 static const struct intel_memory_region_ops shmem_region_ops = { 631 .init = init_shmem, 632 .release = release_shmem, 633 .init_object = shmem_object_init, 634 }; 635 636 struct intel_memory_region *i915_gem_shmem_setup(struct drm_i915_private *i915, 637 u16 type, u16 instance) 638 { 639 return intel_memory_region_create(i915, 0, 640 totalram_pages() << PAGE_SHIFT, 641 PAGE_SIZE, 0, 642 type, instance, 643 &shmem_region_ops); 644 } 645 646 bool i915_gem_object_is_shmem(const struct drm_i915_gem_object *obj) 647 { 648 return obj->ops == &i915_gem_shmem_ops; 649 } 650