drivers/gpu/drm/i915/gem/i915_gem_shmem.c

   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 "i915_drv.h"
  11 #include "i915_gem_object.h"
  12 #include "i915_scatterlist.h"
  13
  14 /*
  15  * Move pages to appropriate lru and release the pagevec, decrementing the
  16  * ref count of those pages.
  17  */
  18 static void check_release_pagevec(struct pagevec *pvec)
  19 {
  20         check_move_unevictable_pages(pvec);
  21         __pagevec_release(pvec);
  22         cond_resched();
  23 }
  24
  25 static int shmem_get_pages(struct drm_i915_gem_object *obj)
  26 {
  27         struct drm_i915_private *i915 = to_i915(obj->base.dev);
  28         const unsigned long page_count = obj->base.size / PAGE_SIZE;
  29         unsigned long i;
  30         struct address_space *mapping;
  31         struct sg_table *st;
  32         struct scatterlist *sg;
  33         struct sgt_iter sgt_iter;
  34         struct page *page;
  35         unsigned long last_pfn = 0;     /* suppress gcc warning */
  36         unsigned int max_segment = i915_sg_segment_size();
  37         unsigned int sg_page_sizes;
  38         struct pagevec pvec;
  39         gfp_t noreclaim;
  40         int ret;
  41
  42         /*
  43          * Assert that the object is not currently in any GPU domain. As it
  44          * wasn't in the GTT, there shouldn't be any way it could have been in
  45          * a GPU cache
  46          */
  47         GEM_BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS);
  48         GEM_BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS);
  49
  50         /*
  51          * If there's no chance of allocating enough pages for the whole
  52          * object, bail early.
  53          */
  54         if (page_count > totalram_pages())
  55                 return -ENOMEM;
  56
  57         st = kmalloc(sizeof(*st), GFP_KERNEL);
  58         if (!st)
  59                 return -ENOMEM;
  60
  61 rebuild_st:
  62         if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
  63                 kfree(st);
  64                 return -ENOMEM;
  65         }
  66
  67         /*
  68          * Get the list of pages out of our struct file.  They'll be pinned
  69          * at this point until we release them.
  70          *
  71          * Fail silently without starting the shrinker
  72          */
  73         mapping = obj->base.filp->f_mapping;
  74         mapping_set_unevictable(mapping);
  75         noreclaim = mapping_gfp_constraint(mapping, ~__GFP_RECLAIM);
  76         noreclaim |= __GFP_NORETRY | __GFP_NOWARN;
  77
  78         sg = st->sgl;
  79         st->nents = 0;
  80         sg_page_sizes = 0;
  81         for (i = 0; i < page_count; i++) {
  82                 const unsigned int shrink[] = {
  83                         I915_SHRINK_BOUND | I915_SHRINK_UNBOUND,
  84                         0,
  85                 }, *s = shrink;
  86                 gfp_t gfp = noreclaim;
  87
  88                 do {
  89                         cond_resched();
  90                         page = shmem_read_mapping_page_gfp(mapping, i, gfp);
  91                         if (!IS_ERR(page))
  92                                 break;
  93
  94                         if (!*s) {
  95                                 ret = PTR_ERR(page);
  96                                 goto err_sg;
  97                         }
  98
  99                         i915_gem_shrink(i915, 2 * page_count, NULL, *s++);
 100
 101                         /*
 102                          * We've tried hard to allocate the memory by reaping
 103                          * our own buffer, now let the real VM do its job and
 104                          * go down in flames if truly OOM.
 105                          *
 106                          * However, since graphics tend to be disposable,
 107                          * defer the oom here by reporting the ENOMEM back
 108                          * to userspace.
 109                          */
 110                         if (!*s) {
 111                                 /* reclaim and warn, but no oom */
 112                                 gfp = mapping_gfp_mask(mapping);
 113
 114                                 /*
 115                                  * Our bo are always dirty and so we require
 116                                  * kswapd to reclaim our pages (direct reclaim
 117                                  * does not effectively begin pageout of our
 118                                  * buffers on its own). However, direct reclaim
 119                                  * only waits for kswapd when under allocation
 120                                  * congestion. So as a result __GFP_RECLAIM is
 121                                  * unreliable and fails to actually reclaim our
 122                                  * dirty pages -- unless you try over and over
 123                                  * again with !__GFP_NORETRY. However, we still
 124                                  * want to fail this allocation rather than
 125                                  * trigger the out-of-memory killer and for
 126                                  * this we want __GFP_RETRY_MAYFAIL.
 127                                  */
 128                                 gfp |= __GFP_RETRY_MAYFAIL;
 129                         }
 130                 } while (1);
 131
 132                 if (!i ||
 133                     sg->length >= max_segment ||
 134                     page_to_pfn(page) != last_pfn + 1) {
 135                         if (i) {
 136                                 sg_page_sizes |= sg->length;
 137                                 sg = sg_next(sg);
 138                         }
 139                         st->nents++;
 140                         sg_set_page(sg, page, PAGE_SIZE, 0);
 141                 } else {
 142                         sg->length += PAGE_SIZE;
 143                 }
 144                 last_pfn = page_to_pfn(page);
 145
 146                 /* Check that the i965g/gm workaround works. */
 147                 WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
 148         }
 149         if (sg) { /* loop terminated early; short sg table */
 150                 sg_page_sizes |= sg->length;
 151                 sg_mark_end(sg);
 152         }
 153
 154         /* Trim unused sg entries to avoid wasting memory. */
 155         i915_sg_trim(st);
 156
 157         ret = i915_gem_gtt_prepare_pages(obj, st);
 158         if (ret) {
 159                 /*
 160                  * DMA remapping failed? One possible cause is that
 161                  * it could not reserve enough large entries, asking
 162                  * for PAGE_SIZE chunks instead may be helpful.
 163                  */
 164                 if (max_segment > PAGE_SIZE) {
 165                         for_each_sgt_page(page, sgt_iter, st)
 166                                 put_page(page);
 167                         sg_free_table(st);
 168
 169                         max_segment = PAGE_SIZE;
 170                         goto rebuild_st;
 171                 } else {
 172                         dev_warn(&i915->drm.pdev->dev,
 173                                  "Failed to DMA remap %lu pages\n",
 174                                  page_count);
 175                         goto err_pages;
 176                 }
 177         }
 178
 179         if (i915_gem_object_needs_bit17_swizzle(obj))
 180                 i915_gem_object_do_bit_17_swizzle(obj, st);
 181
 182         __i915_gem_object_set_pages(obj, st, sg_page_sizes);
 183
 184         return 0;
 185
 186 err_sg:
 187         sg_mark_end(sg);
 188 err_pages:
 189         mapping_clear_unevictable(mapping);
 190         pagevec_init(&pvec);
 191         for_each_sgt_page(page, sgt_iter, st) {
 192                 if (!pagevec_add(&pvec, page))
 193                         check_release_pagevec(&pvec);
 194         }
 195         if (pagevec_count(&pvec))
 196                 check_release_pagevec(&pvec);
 197         sg_free_table(st);
 198         kfree(st);
 199
 200         /*
 201          * shmemfs first checks if there is enough memory to allocate the page
 202          * and reports ENOSPC should there be insufficient, along with the usual
 203          * ENOMEM for a genuine allocation failure.
 204          *
 205          * We use ENOSPC in our driver to mean that we have run out of aperture
 206          * space and so want to translate the error from shmemfs back to our
 207          * usual understanding of ENOMEM.
 208          */
 209         if (ret == -ENOSPC)
 210                 ret = -ENOMEM;
 211
 212         return ret;
 213 }
 214
 215 static void
 216 shmem_truncate(struct drm_i915_gem_object *obj)
 217 {
 218         /*
 219          * Our goal here is to return as much of the memory as
 220          * is possible back to the system as we are called from OOM.
 221          * To do this we must instruct the shmfs to drop all of its
 222          * backing pages, *now*.
 223          */
 224         shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1);
 225         obj->mm.madv = __I915_MADV_PURGED;
 226         obj->mm.pages = ERR_PTR(-EFAULT);
 227 }
 228
 229 static void
 230 shmem_writeback(struct drm_i915_gem_object *obj)
 231 {
 232         struct address_space *mapping;
 233         struct writeback_control wbc = {
 234                 .sync_mode = WB_SYNC_NONE,
 235                 .nr_to_write = SWAP_CLUSTER_MAX,
 236                 .range_start = 0,
 237                 .range_end = LLONG_MAX,
 238                 .for_reclaim = 1,
 239         };
 240         unsigned long i;
 241
 242         /*
 243          * Leave mmapings intact (GTT will have been revoked on unbinding,
 244          * leaving only CPU mmapings around) and add those pages to the LRU
 245          * instead of invoking writeback so they are aged and paged out
 246          * as normal.
 247          */
 248         mapping = obj->base.filp->f_mapping;
 249
 250         /* Begin writeback on each dirty page */
 251         for (i = 0; i < obj->base.size >> PAGE_SHIFT; i++) {
 252                 struct page *page;
 253
 254                 page = find_lock_entry(mapping, i);
 255                 if (!page || xa_is_value(page))
 256                         continue;
 257
 258                 if (!page_mapped(page) && clear_page_dirty_for_io(page)) {
 259                         int ret;
 260
 261                         SetPageReclaim(page);
 262                         ret = mapping->a_ops->writepage(page, &wbc);
 263                         if (!PageWriteback(page))
 264                                 ClearPageReclaim(page);
 265                         if (!ret)
 266                                 goto put;
 267                 }
 268                 unlock_page(page);
 269 put:
 270                 put_page(page);
 271         }
 272 }
 273
 274 void
 275 __i915_gem_object_release_shmem(struct drm_i915_gem_object *obj,
 276                                 struct sg_table *pages,
 277                                 bool needs_clflush)
 278 {
 279         GEM_BUG_ON(obj->mm.madv == __I915_MADV_PURGED);
 280
 281         if (obj->mm.madv == I915_MADV_DONTNEED)
 282                 obj->mm.dirty = false;
 283
 284         if (needs_clflush &&
 285             (obj->read_domains & I915_GEM_DOMAIN_CPU) == 0 &&
 286             !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ))
 287                 drm_clflush_sg(pages);
 288
 289         __start_cpu_write(obj);
 290 }
 291
 292 static void
 293 shmem_put_pages(struct drm_i915_gem_object *obj, struct sg_table *pages)
 294 {
 295         struct sgt_iter sgt_iter;
 296         struct pagevec pvec;
 297         struct page *page;
 298
 299         __i915_gem_object_release_shmem(obj, pages, true);
 300
 301         i915_gem_gtt_finish_pages(obj, pages);
 302
 303         if (i915_gem_object_needs_bit17_swizzle(obj))
 304                 i915_gem_object_save_bit_17_swizzle(obj, pages);
 305
 306         mapping_clear_unevictable(file_inode(obj->base.filp)->i_mapping);
 307
 308         pagevec_init(&pvec);
 309         for_each_sgt_page(page, sgt_iter, pages) {
 310                 if (obj->mm.dirty)
 311                         set_page_dirty(page);
 312
 313                 if (obj->mm.madv == I915_MADV_WILLNEED)
 314                         mark_page_accessed(page);
 315
 316                 if (!pagevec_add(&pvec, page))
 317                         check_release_pagevec(&pvec);
 318         }
 319         if (pagevec_count(&pvec))
 320                 check_release_pagevec(&pvec);
 321         obj->mm.dirty = false;
 322
 323         sg_free_table(pages);
 324         kfree(pages);
 325 }
 326
 327 static int
 328 shmem_pwrite(struct drm_i915_gem_object *obj,
 329              const struct drm_i915_gem_pwrite *arg)
 330 {
 331         struct address_space *mapping = obj->base.filp->f_mapping;
 332         char __user *user_data = u64_to_user_ptr(arg->data_ptr);
 333         u64 remain, offset;
 334         unsigned int pg;
 335
 336         /* Caller already validated user args */
 337         GEM_BUG_ON(!access_ok(user_data, arg->size));
 338
 339         /*
 340          * Before we instantiate/pin the backing store for our use, we
 341          * can prepopulate the shmemfs filp efficiently using a write into
 342          * the pagecache. We avoid the penalty of instantiating all the
 343          * pages, important if the user is just writing to a few and never
 344          * uses the object on the GPU, and using a direct write into shmemfs
 345          * allows it to avoid the cost of retrieving a page (either swapin
 346          * or clearing-before-use) before it is overwritten.
 347          */
 348         if (i915_gem_object_has_pages(obj))
 349                 return -ENODEV;
 350
 351         if (obj->mm.madv != I915_MADV_WILLNEED)
 352                 return -EFAULT;
 353
 354         /*
 355          * Before the pages are instantiated the object is treated as being
 356          * in the CPU domain. The pages will be clflushed as required before
 357          * use, and we can freely write into the pages directly. If userspace
 358          * races pwrite with any other operation; corruption will ensue -
 359          * that is userspace's prerogative!
 360          */
 361
 362         remain = arg->size;
 363         offset = arg->offset;
 364         pg = offset_in_page(offset);
 365
 366         do {
 367                 unsigned int len, unwritten;
 368                 struct page *page;
 369                 void *data, *vaddr;
 370                 int err;
 371                 char c;
 372
 373                 len = PAGE_SIZE - pg;
 374                 if (len > remain)
 375                         len = remain;
 376
 377                 /* Prefault the user page to reduce potential recursion */
 378                 err = __get_user(c, user_data);
 379                 if (err)
 380                         return err;
 381
 382                 err = __get_user(c, user_data + len - 1);
 383                 if (err)
 384                         return err;
 385
 386                 err = pagecache_write_begin(obj->base.filp, mapping,
 387                                             offset, len, 0,
 388                                             &page, &data);
 389                 if (err < 0)
 390                         return err;
 391
 392                 vaddr = kmap_atomic(page);
 393                 unwritten = __copy_from_user_inatomic(vaddr + pg,
 394                                                       user_data,
 395                                                       len);
 396                 kunmap_atomic(vaddr);
 397
 398                 err = pagecache_write_end(obj->base.filp, mapping,
 399                                           offset, len, len - unwritten,
 400                                           page, data);
 401                 if (err < 0)
 402                         return err;
 403
 404                 /* We don't handle -EFAULT, leave it to the caller to check */
 405                 if (unwritten)
 406                         return -ENODEV;
 407
 408                 remain -= len;
 409                 user_data += len;
 410                 offset += len;
 411                 pg = 0;
 412         } while (remain);
 413
 414         return 0;
 415 }
 416
 417 const struct drm_i915_gem_object_ops i915_gem_shmem_ops = {
 418         .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
 419                  I915_GEM_OBJECT_IS_SHRINKABLE,
 420
 421         .get_pages = shmem_get_pages,
 422         .put_pages = shmem_put_pages,
 423         .truncate = shmem_truncate,
 424         .writeback = shmem_writeback,
 425
 426         .pwrite = shmem_pwrite,
 427 };
 428
 429 static int create_shmem(struct drm_i915_private *i915,
 430                         struct drm_gem_object *obj,
 431                         size_t size)
 432 {
 433         unsigned long flags = VM_NORESERVE;
 434         struct file *filp;
 435
 436         drm_gem_private_object_init(&i915->drm, obj, size);
 437
 438         if (i915->mm.gemfs)
 439                 filp = shmem_file_setup_with_mnt(i915->mm.gemfs, "i915", size,
 440                                                  flags);
 441         else
 442                 filp = shmem_file_setup("i915", size, flags);
 443         if (IS_ERR(filp))
 444                 return PTR_ERR(filp);
 445
 446         obj->filp = filp;
 447         return 0;
 448 }
 449
 450 struct drm_i915_gem_object *
 451 i915_gem_object_create_shmem(struct drm_i915_private *i915, u64 size)
 452 {
 453         struct drm_i915_gem_object *obj;
 454         struct address_space *mapping;
 455         unsigned int cache_level;
 456         gfp_t mask;
 457         int ret;
 458
 459         /* There is a prevalence of the assumption that we fit the object's
 460          * page count inside a 32bit _signed_ variable. Let's document this and
 461          * catch if we ever need to fix it. In the meantime, if you do spot
 462          * such a local variable, please consider fixing!
 463          */
 464         if (size >> PAGE_SHIFT > INT_MAX)
 465                 return ERR_PTR(-E2BIG);
 466
 467         if (overflows_type(size, obj->base.size))
 468                 return ERR_PTR(-E2BIG);
 469
 470         obj = i915_gem_object_alloc();
 471         if (!obj)
 472                 return ERR_PTR(-ENOMEM);
 473
 474         ret = create_shmem(i915, &obj->base, size);
 475         if (ret)
 476                 goto fail;
 477
 478         mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
 479         if (IS_I965GM(i915) || IS_I965G(i915)) {
 480                 /* 965gm cannot relocate objects above 4GiB. */
 481                 mask &= ~__GFP_HIGHMEM;
 482                 mask |= __GFP_DMA32;
 483         }
 484
 485         mapping = obj->base.filp->f_mapping;
 486         mapping_set_gfp_mask(mapping, mask);
 487         GEM_BUG_ON(!(mapping_gfp_mask(mapping) & __GFP_RECLAIM));
 488
 489         i915_gem_object_init(obj, &i915_gem_shmem_ops);
 490
 491         obj->write_domain = I915_GEM_DOMAIN_CPU;
 492         obj->read_domains = I915_GEM_DOMAIN_CPU;
 493
 494         if (HAS_LLC(i915))
 495                 /* On some devices, we can have the GPU use the LLC (the CPU
 496                  * cache) for about a 10% performance improvement
 497                  * compared to uncached.  Graphics requests other than
 498                  * display scanout are coherent with the CPU in
 499                  * accessing this cache.  This means in this mode we
 500                  * don't need to clflush on the CPU side, and on the
 501                  * GPU side we only need to flush internal caches to
 502                  * get data visible to the CPU.
 503                  *
 504                  * However, we maintain the display planes as UC, and so
 505                  * need to rebind when first used as such.
 506                  */
 507                 cache_level = I915_CACHE_LLC;
 508         else
 509                 cache_level = I915_CACHE_NONE;
 510
 511         i915_gem_object_set_cache_coherency(obj, cache_level);
 512
 513         trace_i915_gem_object_create(obj);
 514
 515         return obj;
 516
 517 fail:
 518         i915_gem_object_free(obj);
 519         return ERR_PTR(ret);
 520 }
 521
 522 /* Allocate a new GEM object and fill it with the supplied data */
 523 struct drm_i915_gem_object *
 524 i915_gem_object_create_shmem_from_data(struct drm_i915_private *dev_priv,
 525                                        const void *data, size_t size)
 526 {
 527         struct drm_i915_gem_object *obj;
 528         struct file *file;
 529         size_t offset;
 530         int err;
 531
 532         obj = i915_gem_object_create_shmem(dev_priv, round_up(size, PAGE_SIZE));
 533         if (IS_ERR(obj))
 534                 return obj;
 535
 536         GEM_BUG_ON(obj->write_domain != I915_GEM_DOMAIN_CPU);
 537
 538         file = obj->base.filp;
 539         offset = 0;
 540         do {
 541                 unsigned int len = min_t(typeof(size), size, PAGE_SIZE);
 542                 struct page *page;
 543                 void *pgdata, *vaddr;
 544
 545                 err = pagecache_write_begin(file, file->f_mapping,
 546                                             offset, len, 0,
 547                                             &page, &pgdata);
 548                 if (err < 0)
 549                         goto fail;
 550
 551                 vaddr = kmap(page);
 552                 memcpy(vaddr, data, len);
 553                 kunmap(page);
 554
 555                 err = pagecache_write_end(file, file->f_mapping,
 556                                           offset, len, len,
 557                                           page, pgdata);
 558                 if (err < 0)
 559                         goto fail;
 560
 561                 size -= len;
 562                 data += len;
 563                 offset += len;
 564         } while (size);
 565
 566         return obj;
 567
 568 fail:
 569         i915_gem_object_put(obj);
 570         return ERR_PTR(err);
 571 }