18475355fSChris Wilson /* 28475355fSChris Wilson * SPDX-License-Identifier: MIT 38475355fSChris Wilson * 48475355fSChris Wilson * Copyright © 2014-2016 Intel Corporation 58475355fSChris Wilson */ 68475355fSChris Wilson 78475355fSChris Wilson #include <linux/pagevec.h> 88475355fSChris Wilson #include <linux/swap.h> 98475355fSChris Wilson 10da1184cdSMatthew Auld #include "gem/i915_gem_region.h" 118475355fSChris Wilson #include "i915_drv.h" 12da1184cdSMatthew Auld #include "i915_gemfs.h" 138475355fSChris Wilson #include "i915_gem_object.h" 1437d63f8fSChris Wilson #include "i915_scatterlist.h" 15a09d9a80SJani Nikula #include "i915_trace.h" 168475355fSChris Wilson 178475355fSChris Wilson /* 188475355fSChris Wilson * Move pages to appropriate lru and release the pagevec, decrementing the 198475355fSChris Wilson * ref count of those pages. 208475355fSChris Wilson */ 218475355fSChris Wilson static void check_release_pagevec(struct pagevec *pvec) 228475355fSChris Wilson { 238475355fSChris Wilson check_move_unevictable_pages(pvec); 248475355fSChris Wilson __pagevec_release(pvec); 258475355fSChris Wilson cond_resched(); 268475355fSChris Wilson } 278475355fSChris Wilson 288475355fSChris Wilson static int shmem_get_pages(struct drm_i915_gem_object *obj) 298475355fSChris Wilson { 308475355fSChris Wilson struct drm_i915_private *i915 = to_i915(obj->base.dev); 31da1184cdSMatthew Auld struct intel_memory_region *mem = obj->mm.region; 328475355fSChris Wilson const unsigned long page_count = obj->base.size / PAGE_SIZE; 338475355fSChris Wilson unsigned long i; 348475355fSChris Wilson struct address_space *mapping; 358475355fSChris Wilson struct sg_table *st; 368475355fSChris Wilson struct scatterlist *sg; 378475355fSChris Wilson struct sgt_iter sgt_iter; 388475355fSChris Wilson struct page *page; 398475355fSChris Wilson unsigned long last_pfn = 0; /* suppress gcc warning */ 408475355fSChris Wilson unsigned int max_segment = i915_sg_segment_size(); 418475355fSChris Wilson unsigned int sg_page_sizes; 428475355fSChris Wilson struct pagevec pvec; 438475355fSChris Wilson gfp_t noreclaim; 448475355fSChris Wilson int ret; 458475355fSChris Wilson 468475355fSChris Wilson /* 478475355fSChris Wilson * Assert that the object is not currently in any GPU domain. As it 488475355fSChris Wilson * wasn't in the GTT, there shouldn't be any way it could have been in 498475355fSChris Wilson * a GPU cache 508475355fSChris Wilson */ 518475355fSChris Wilson GEM_BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS); 528475355fSChris Wilson GEM_BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS); 538475355fSChris Wilson 548475355fSChris Wilson /* 558475355fSChris Wilson * If there's no chance of allocating enough pages for the whole 568475355fSChris Wilson * object, bail early. 578475355fSChris Wilson */ 58da1184cdSMatthew Auld if (obj->base.size > resource_size(&mem->region)) 598475355fSChris Wilson return -ENOMEM; 608475355fSChris Wilson 618475355fSChris Wilson st = kmalloc(sizeof(*st), GFP_KERNEL); 628475355fSChris Wilson if (!st) 638475355fSChris Wilson return -ENOMEM; 648475355fSChris Wilson 658475355fSChris Wilson rebuild_st: 668475355fSChris Wilson if (sg_alloc_table(st, page_count, GFP_KERNEL)) { 678475355fSChris Wilson kfree(st); 688475355fSChris Wilson return -ENOMEM; 698475355fSChris Wilson } 708475355fSChris Wilson 718475355fSChris Wilson /* 728475355fSChris Wilson * Get the list of pages out of our struct file. They'll be pinned 738475355fSChris Wilson * at this point until we release them. 748475355fSChris Wilson * 758475355fSChris Wilson * Fail silently without starting the shrinker 768475355fSChris Wilson */ 778475355fSChris Wilson mapping = obj->base.filp->f_mapping; 788475355fSChris Wilson mapping_set_unevictable(mapping); 798475355fSChris Wilson noreclaim = mapping_gfp_constraint(mapping, ~__GFP_RECLAIM); 808475355fSChris Wilson noreclaim |= __GFP_NORETRY | __GFP_NOWARN; 818475355fSChris Wilson 828475355fSChris Wilson sg = st->sgl; 838475355fSChris Wilson st->nents = 0; 848475355fSChris Wilson sg_page_sizes = 0; 858475355fSChris Wilson for (i = 0; i < page_count; i++) { 868475355fSChris Wilson const unsigned int shrink[] = { 873b4fa964SChris Wilson I915_SHRINK_BOUND | I915_SHRINK_UNBOUND, 888475355fSChris Wilson 0, 898475355fSChris Wilson }, *s = shrink; 908475355fSChris Wilson gfp_t gfp = noreclaim; 918475355fSChris Wilson 928475355fSChris Wilson do { 938475355fSChris Wilson cond_resched(); 948475355fSChris Wilson page = shmem_read_mapping_page_gfp(mapping, i, gfp); 958475355fSChris Wilson if (!IS_ERR(page)) 968475355fSChris Wilson break; 978475355fSChris Wilson 988475355fSChris Wilson if (!*s) { 998475355fSChris Wilson ret = PTR_ERR(page); 1008475355fSChris Wilson goto err_sg; 1018475355fSChris Wilson } 1028475355fSChris Wilson 1038475355fSChris Wilson i915_gem_shrink(i915, 2 * page_count, NULL, *s++); 1048475355fSChris Wilson 1058475355fSChris Wilson /* 1068475355fSChris Wilson * We've tried hard to allocate the memory by reaping 1078475355fSChris Wilson * our own buffer, now let the real VM do its job and 1088475355fSChris Wilson * go down in flames if truly OOM. 1098475355fSChris Wilson * 1108475355fSChris Wilson * However, since graphics tend to be disposable, 1118475355fSChris Wilson * defer the oom here by reporting the ENOMEM back 1128475355fSChris Wilson * to userspace. 1138475355fSChris Wilson */ 1148475355fSChris Wilson if (!*s) { 1158475355fSChris Wilson /* reclaim and warn, but no oom */ 1168475355fSChris Wilson gfp = mapping_gfp_mask(mapping); 1178475355fSChris Wilson 1188475355fSChris Wilson /* 1198475355fSChris Wilson * Our bo are always dirty and so we require 1208475355fSChris Wilson * kswapd to reclaim our pages (direct reclaim 1218475355fSChris Wilson * does not effectively begin pageout of our 1228475355fSChris Wilson * buffers on its own). However, direct reclaim 1238475355fSChris Wilson * only waits for kswapd when under allocation 1248475355fSChris Wilson * congestion. So as a result __GFP_RECLAIM is 1258475355fSChris Wilson * unreliable and fails to actually reclaim our 1268475355fSChris Wilson * dirty pages -- unless you try over and over 1278475355fSChris Wilson * again with !__GFP_NORETRY. However, we still 1288475355fSChris Wilson * want to fail this allocation rather than 1298475355fSChris Wilson * trigger the out-of-memory killer and for 1308475355fSChris Wilson * this we want __GFP_RETRY_MAYFAIL. 1318475355fSChris Wilson */ 1328475355fSChris Wilson gfp |= __GFP_RETRY_MAYFAIL; 1338475355fSChris Wilson } 1348475355fSChris Wilson } while (1); 1358475355fSChris Wilson 1368475355fSChris Wilson if (!i || 1378475355fSChris Wilson sg->length >= max_segment || 1388475355fSChris Wilson page_to_pfn(page) != last_pfn + 1) { 1398475355fSChris Wilson if (i) { 1408475355fSChris Wilson sg_page_sizes |= sg->length; 1418475355fSChris Wilson sg = sg_next(sg); 1428475355fSChris Wilson } 1438475355fSChris Wilson st->nents++; 1448475355fSChris Wilson sg_set_page(sg, page, PAGE_SIZE, 0); 1458475355fSChris Wilson } else { 1468475355fSChris Wilson sg->length += PAGE_SIZE; 1478475355fSChris Wilson } 1488475355fSChris Wilson last_pfn = page_to_pfn(page); 1498475355fSChris Wilson 1508475355fSChris Wilson /* Check that the i965g/gm workaround works. */ 1518475355fSChris Wilson WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL)); 1528475355fSChris Wilson } 1538475355fSChris Wilson if (sg) { /* loop terminated early; short sg table */ 1548475355fSChris Wilson sg_page_sizes |= sg->length; 1558475355fSChris Wilson sg_mark_end(sg); 1568475355fSChris Wilson } 1578475355fSChris Wilson 1588475355fSChris Wilson /* Trim unused sg entries to avoid wasting memory. */ 1598475355fSChris Wilson i915_sg_trim(st); 1608475355fSChris Wilson 1618475355fSChris Wilson ret = i915_gem_gtt_prepare_pages(obj, st); 1628475355fSChris Wilson if (ret) { 1638475355fSChris Wilson /* 1648475355fSChris Wilson * DMA remapping failed? One possible cause is that 1658475355fSChris Wilson * it could not reserve enough large entries, asking 1668475355fSChris Wilson * for PAGE_SIZE chunks instead may be helpful. 1678475355fSChris Wilson */ 1688475355fSChris Wilson if (max_segment > PAGE_SIZE) { 1698475355fSChris Wilson for_each_sgt_page(page, sgt_iter, st) 1708475355fSChris Wilson put_page(page); 1718475355fSChris Wilson sg_free_table(st); 1728475355fSChris Wilson 1738475355fSChris Wilson max_segment = PAGE_SIZE; 1748475355fSChris Wilson goto rebuild_st; 1758475355fSChris Wilson } else { 1768475355fSChris Wilson dev_warn(&i915->drm.pdev->dev, 1778475355fSChris Wilson "Failed to DMA remap %lu pages\n", 1788475355fSChris Wilson page_count); 1798475355fSChris Wilson goto err_pages; 1808475355fSChris Wilson } 1818475355fSChris Wilson } 1828475355fSChris Wilson 1838475355fSChris Wilson if (i915_gem_object_needs_bit17_swizzle(obj)) 1848475355fSChris Wilson i915_gem_object_do_bit_17_swizzle(obj, st); 1858475355fSChris Wilson 1868475355fSChris Wilson __i915_gem_object_set_pages(obj, st, sg_page_sizes); 1878475355fSChris Wilson 1888475355fSChris Wilson return 0; 1898475355fSChris Wilson 1908475355fSChris Wilson err_sg: 1918475355fSChris Wilson sg_mark_end(sg); 1928475355fSChris Wilson err_pages: 1938475355fSChris Wilson mapping_clear_unevictable(mapping); 1948475355fSChris Wilson pagevec_init(&pvec); 1958475355fSChris Wilson for_each_sgt_page(page, sgt_iter, st) { 1968475355fSChris Wilson if (!pagevec_add(&pvec, page)) 1978475355fSChris Wilson check_release_pagevec(&pvec); 1988475355fSChris Wilson } 1998475355fSChris Wilson if (pagevec_count(&pvec)) 2008475355fSChris Wilson check_release_pagevec(&pvec); 2018475355fSChris Wilson sg_free_table(st); 2028475355fSChris Wilson kfree(st); 2038475355fSChris Wilson 2048475355fSChris Wilson /* 2058475355fSChris Wilson * shmemfs first checks if there is enough memory to allocate the page 2068475355fSChris Wilson * and reports ENOSPC should there be insufficient, along with the usual 2078475355fSChris Wilson * ENOMEM for a genuine allocation failure. 2088475355fSChris Wilson * 2098475355fSChris Wilson * We use ENOSPC in our driver to mean that we have run out of aperture 2108475355fSChris Wilson * space and so want to translate the error from shmemfs back to our 2118475355fSChris Wilson * usual understanding of ENOMEM. 2128475355fSChris Wilson */ 2138475355fSChris Wilson if (ret == -ENOSPC) 2148475355fSChris Wilson ret = -ENOMEM; 2158475355fSChris Wilson 2168475355fSChris Wilson return ret; 2178475355fSChris Wilson } 2188475355fSChris Wilson 219f033428dSChris Wilson static void 220f033428dSChris Wilson shmem_truncate(struct drm_i915_gem_object *obj) 221f033428dSChris Wilson { 222f033428dSChris Wilson /* 223f033428dSChris Wilson * Our goal here is to return as much of the memory as 224f033428dSChris Wilson * is possible back to the system as we are called from OOM. 225f033428dSChris Wilson * To do this we must instruct the shmfs to drop all of its 226f033428dSChris Wilson * backing pages, *now*. 227f033428dSChris Wilson */ 228f033428dSChris Wilson shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1); 229f033428dSChris Wilson obj->mm.madv = __I915_MADV_PURGED; 230f033428dSChris Wilson obj->mm.pages = ERR_PTR(-EFAULT); 231f033428dSChris Wilson } 232f033428dSChris Wilson 233f033428dSChris Wilson static void 234f033428dSChris Wilson shmem_writeback(struct drm_i915_gem_object *obj) 235f033428dSChris Wilson { 236f033428dSChris Wilson struct address_space *mapping; 237f033428dSChris Wilson struct writeback_control wbc = { 238f033428dSChris Wilson .sync_mode = WB_SYNC_NONE, 239f033428dSChris Wilson .nr_to_write = SWAP_CLUSTER_MAX, 240f033428dSChris Wilson .range_start = 0, 241f033428dSChris Wilson .range_end = LLONG_MAX, 242f033428dSChris Wilson .for_reclaim = 1, 243f033428dSChris Wilson }; 244f033428dSChris Wilson unsigned long i; 245f033428dSChris Wilson 246f033428dSChris Wilson /* 247f033428dSChris Wilson * Leave mmapings intact (GTT will have been revoked on unbinding, 248f033428dSChris Wilson * leaving only CPU mmapings around) and add those pages to the LRU 249f033428dSChris Wilson * instead of invoking writeback so they are aged and paged out 250f033428dSChris Wilson * as normal. 251f033428dSChris Wilson */ 252f033428dSChris Wilson mapping = obj->base.filp->f_mapping; 253f033428dSChris Wilson 254f033428dSChris Wilson /* Begin writeback on each dirty page */ 255f033428dSChris Wilson for (i = 0; i < obj->base.size >> PAGE_SHIFT; i++) { 256f033428dSChris Wilson struct page *page; 257f033428dSChris Wilson 258f033428dSChris Wilson page = find_lock_entry(mapping, i); 259f033428dSChris Wilson if (!page || xa_is_value(page)) 260f033428dSChris Wilson continue; 261f033428dSChris Wilson 262f033428dSChris Wilson if (!page_mapped(page) && clear_page_dirty_for_io(page)) { 263f033428dSChris Wilson int ret; 264f033428dSChris Wilson 265f033428dSChris Wilson SetPageReclaim(page); 266f033428dSChris Wilson ret = mapping->a_ops->writepage(page, &wbc); 267f033428dSChris Wilson if (!PageWriteback(page)) 268f033428dSChris Wilson ClearPageReclaim(page); 269f033428dSChris Wilson if (!ret) 270f033428dSChris Wilson goto put; 271f033428dSChris Wilson } 272f033428dSChris Wilson unlock_page(page); 273f033428dSChris Wilson put: 274f033428dSChris Wilson put_page(page); 275f033428dSChris Wilson } 276f033428dSChris Wilson } 277f033428dSChris Wilson 2788475355fSChris Wilson void 2798475355fSChris Wilson __i915_gem_object_release_shmem(struct drm_i915_gem_object *obj, 2808475355fSChris Wilson struct sg_table *pages, 2818475355fSChris Wilson bool needs_clflush) 2828475355fSChris Wilson { 2838475355fSChris Wilson GEM_BUG_ON(obj->mm.madv == __I915_MADV_PURGED); 2848475355fSChris Wilson 2858475355fSChris Wilson if (obj->mm.madv == I915_MADV_DONTNEED) 2868475355fSChris Wilson obj->mm.dirty = false; 2878475355fSChris Wilson 2888475355fSChris Wilson if (needs_clflush && 2898475355fSChris Wilson (obj->read_domains & I915_GEM_DOMAIN_CPU) == 0 && 2908475355fSChris Wilson !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ)) 2918475355fSChris Wilson drm_clflush_sg(pages); 2928475355fSChris Wilson 2938475355fSChris Wilson __start_cpu_write(obj); 2948475355fSChris Wilson } 2958475355fSChris Wilson 2968475355fSChris Wilson static void 2978475355fSChris Wilson shmem_put_pages(struct drm_i915_gem_object *obj, struct sg_table *pages) 2988475355fSChris Wilson { 2998475355fSChris Wilson struct sgt_iter sgt_iter; 3008475355fSChris Wilson struct pagevec pvec; 3018475355fSChris Wilson struct page *page; 3028475355fSChris Wilson 3038475355fSChris Wilson __i915_gem_object_release_shmem(obj, pages, true); 3048475355fSChris Wilson 3058475355fSChris Wilson i915_gem_gtt_finish_pages(obj, pages); 3068475355fSChris Wilson 3078475355fSChris Wilson if (i915_gem_object_needs_bit17_swizzle(obj)) 3088475355fSChris Wilson i915_gem_object_save_bit_17_swizzle(obj, pages); 3098475355fSChris Wilson 3108475355fSChris Wilson mapping_clear_unevictable(file_inode(obj->base.filp)->i_mapping); 3118475355fSChris Wilson 3128475355fSChris Wilson pagevec_init(&pvec); 3138475355fSChris Wilson for_each_sgt_page(page, sgt_iter, pages) { 3148475355fSChris Wilson if (obj->mm.dirty) 3158475355fSChris Wilson set_page_dirty(page); 3168475355fSChris Wilson 3178475355fSChris Wilson if (obj->mm.madv == I915_MADV_WILLNEED) 3188475355fSChris Wilson mark_page_accessed(page); 3198475355fSChris Wilson 3208475355fSChris Wilson if (!pagevec_add(&pvec, page)) 3218475355fSChris Wilson check_release_pagevec(&pvec); 3228475355fSChris Wilson } 3238475355fSChris Wilson if (pagevec_count(&pvec)) 3248475355fSChris Wilson check_release_pagevec(&pvec); 3258475355fSChris Wilson obj->mm.dirty = false; 3268475355fSChris Wilson 3278475355fSChris Wilson sg_free_table(pages); 3288475355fSChris Wilson kfree(pages); 3298475355fSChris Wilson } 3308475355fSChris Wilson 3318475355fSChris Wilson static int 3328475355fSChris Wilson shmem_pwrite(struct drm_i915_gem_object *obj, 3338475355fSChris Wilson const struct drm_i915_gem_pwrite *arg) 3348475355fSChris Wilson { 3358475355fSChris Wilson struct address_space *mapping = obj->base.filp->f_mapping; 3368475355fSChris Wilson char __user *user_data = u64_to_user_ptr(arg->data_ptr); 3378475355fSChris Wilson u64 remain, offset; 3388475355fSChris Wilson unsigned int pg; 3398475355fSChris Wilson 3408475355fSChris Wilson /* Caller already validated user args */ 3418475355fSChris Wilson GEM_BUG_ON(!access_ok(user_data, arg->size)); 3428475355fSChris Wilson 3438475355fSChris Wilson /* 3448475355fSChris Wilson * Before we instantiate/pin the backing store for our use, we 3458475355fSChris Wilson * can prepopulate the shmemfs filp efficiently using a write into 3468475355fSChris Wilson * the pagecache. We avoid the penalty of instantiating all the 3478475355fSChris Wilson * pages, important if the user is just writing to a few and never 3488475355fSChris Wilson * uses the object on the GPU, and using a direct write into shmemfs 3498475355fSChris Wilson * allows it to avoid the cost of retrieving a page (either swapin 3508475355fSChris Wilson * or clearing-before-use) before it is overwritten. 3518475355fSChris Wilson */ 3528475355fSChris Wilson if (i915_gem_object_has_pages(obj)) 3538475355fSChris Wilson return -ENODEV; 3548475355fSChris Wilson 3558475355fSChris Wilson if (obj->mm.madv != I915_MADV_WILLNEED) 3568475355fSChris Wilson return -EFAULT; 3578475355fSChris Wilson 3588475355fSChris Wilson /* 3598475355fSChris Wilson * Before the pages are instantiated the object is treated as being 3608475355fSChris Wilson * in the CPU domain. The pages will be clflushed as required before 3618475355fSChris Wilson * use, and we can freely write into the pages directly. If userspace 3628475355fSChris Wilson * races pwrite with any other operation; corruption will ensue - 3638475355fSChris Wilson * that is userspace's prerogative! 3648475355fSChris Wilson */ 3658475355fSChris Wilson 3668475355fSChris Wilson remain = arg->size; 3678475355fSChris Wilson offset = arg->offset; 3688475355fSChris Wilson pg = offset_in_page(offset); 3698475355fSChris Wilson 3708475355fSChris Wilson do { 3718475355fSChris Wilson unsigned int len, unwritten; 3728475355fSChris Wilson struct page *page; 3738475355fSChris Wilson void *data, *vaddr; 3748475355fSChris Wilson int err; 3758475355fSChris Wilson char c; 3768475355fSChris Wilson 3778475355fSChris Wilson len = PAGE_SIZE - pg; 3788475355fSChris Wilson if (len > remain) 3798475355fSChris Wilson len = remain; 3808475355fSChris Wilson 3818475355fSChris Wilson /* Prefault the user page to reduce potential recursion */ 3828475355fSChris Wilson err = __get_user(c, user_data); 3838475355fSChris Wilson if (err) 3848475355fSChris Wilson return err; 3858475355fSChris Wilson 3868475355fSChris Wilson err = __get_user(c, user_data + len - 1); 3878475355fSChris Wilson if (err) 3888475355fSChris Wilson return err; 3898475355fSChris Wilson 3908475355fSChris Wilson err = pagecache_write_begin(obj->base.filp, mapping, 3918475355fSChris Wilson offset, len, 0, 3928475355fSChris Wilson &page, &data); 3938475355fSChris Wilson if (err < 0) 3948475355fSChris Wilson return err; 3958475355fSChris Wilson 3968475355fSChris Wilson vaddr = kmap_atomic(page); 3978475355fSChris Wilson unwritten = __copy_from_user_inatomic(vaddr + pg, 3988475355fSChris Wilson user_data, 3998475355fSChris Wilson len); 4008475355fSChris Wilson kunmap_atomic(vaddr); 4018475355fSChris Wilson 4028475355fSChris Wilson err = pagecache_write_end(obj->base.filp, mapping, 4038475355fSChris Wilson offset, len, len - unwritten, 4048475355fSChris Wilson page, data); 4058475355fSChris Wilson if (err < 0) 4068475355fSChris Wilson return err; 4078475355fSChris Wilson 4088475355fSChris Wilson /* We don't handle -EFAULT, leave it to the caller to check */ 4098475355fSChris Wilson if (unwritten) 4108475355fSChris Wilson return -ENODEV; 4118475355fSChris Wilson 4128475355fSChris Wilson remain -= len; 4138475355fSChris Wilson user_data += len; 4148475355fSChris Wilson offset += len; 4158475355fSChris Wilson pg = 0; 4168475355fSChris Wilson } while (remain); 4178475355fSChris Wilson 4188475355fSChris Wilson return 0; 4198475355fSChris Wilson } 4208475355fSChris Wilson 4210c159ffeSChris Wilson static void shmem_release(struct drm_i915_gem_object *obj) 4220c159ffeSChris Wilson { 423da1184cdSMatthew Auld i915_gem_object_release_memory_region(obj); 424da1184cdSMatthew Auld 4250c159ffeSChris Wilson fput(obj->base.filp); 4260c159ffeSChris Wilson } 4270c159ffeSChris Wilson 4288475355fSChris Wilson const struct drm_i915_gem_object_ops i915_gem_shmem_ops = { 4298475355fSChris Wilson .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE | 4308475355fSChris Wilson I915_GEM_OBJECT_IS_SHRINKABLE, 4318475355fSChris Wilson 4328475355fSChris Wilson .get_pages = shmem_get_pages, 4338475355fSChris Wilson .put_pages = shmem_put_pages, 434f033428dSChris Wilson .truncate = shmem_truncate, 435f033428dSChris Wilson .writeback = shmem_writeback, 4368475355fSChris Wilson 4378475355fSChris Wilson .pwrite = shmem_pwrite, 4380c159ffeSChris Wilson 4390c159ffeSChris Wilson .release = shmem_release, 4408475355fSChris Wilson }; 4418475355fSChris Wilson 442da1184cdSMatthew Auld static int __create_shmem(struct drm_i915_private *i915, 4438475355fSChris Wilson struct drm_gem_object *obj, 444da1184cdSMatthew Auld resource_size_t size) 4458475355fSChris Wilson { 4468475355fSChris Wilson unsigned long flags = VM_NORESERVE; 4478475355fSChris Wilson struct file *filp; 4488475355fSChris Wilson 4498475355fSChris Wilson drm_gem_private_object_init(&i915->drm, obj, size); 4508475355fSChris Wilson 4518475355fSChris Wilson if (i915->mm.gemfs) 4528475355fSChris Wilson filp = shmem_file_setup_with_mnt(i915->mm.gemfs, "i915", size, 4538475355fSChris Wilson flags); 4548475355fSChris Wilson else 4558475355fSChris Wilson filp = shmem_file_setup("i915", size, flags); 4568475355fSChris Wilson if (IS_ERR(filp)) 4578475355fSChris Wilson return PTR_ERR(filp); 4588475355fSChris Wilson 4598475355fSChris Wilson obj->filp = filp; 4608475355fSChris Wilson return 0; 4618475355fSChris Wilson } 4628475355fSChris Wilson 463da1184cdSMatthew Auld static struct drm_i915_gem_object * 464da1184cdSMatthew Auld create_shmem(struct intel_memory_region *mem, 465da1184cdSMatthew Auld resource_size_t size, 466da1184cdSMatthew Auld unsigned int flags) 4678475355fSChris Wilson { 4687867d709SChris Wilson static struct lock_class_key lock_class; 469da1184cdSMatthew Auld struct drm_i915_private *i915 = mem->i915; 4708475355fSChris Wilson struct drm_i915_gem_object *obj; 4718475355fSChris Wilson struct address_space *mapping; 4728475355fSChris Wilson unsigned int cache_level; 4738475355fSChris Wilson gfp_t mask; 4748475355fSChris Wilson int ret; 4758475355fSChris Wilson 4768475355fSChris Wilson obj = i915_gem_object_alloc(); 4778475355fSChris Wilson if (!obj) 4788475355fSChris Wilson return ERR_PTR(-ENOMEM); 4798475355fSChris Wilson 480da1184cdSMatthew Auld ret = __create_shmem(i915, &obj->base, size); 4818475355fSChris Wilson if (ret) 4828475355fSChris Wilson goto fail; 4838475355fSChris Wilson 4848475355fSChris Wilson mask = GFP_HIGHUSER | __GFP_RECLAIMABLE; 4858475355fSChris Wilson if (IS_I965GM(i915) || IS_I965G(i915)) { 4868475355fSChris Wilson /* 965gm cannot relocate objects above 4GiB. */ 4878475355fSChris Wilson mask &= ~__GFP_HIGHMEM; 4888475355fSChris Wilson mask |= __GFP_DMA32; 4898475355fSChris Wilson } 4908475355fSChris Wilson 4918475355fSChris Wilson mapping = obj->base.filp->f_mapping; 4928475355fSChris Wilson mapping_set_gfp_mask(mapping, mask); 4938475355fSChris Wilson GEM_BUG_ON(!(mapping_gfp_mask(mapping) & __GFP_RECLAIM)); 4948475355fSChris Wilson 4957867d709SChris Wilson i915_gem_object_init(obj, &i915_gem_shmem_ops, &lock_class); 4968475355fSChris Wilson 4978475355fSChris Wilson obj->write_domain = I915_GEM_DOMAIN_CPU; 4988475355fSChris Wilson obj->read_domains = I915_GEM_DOMAIN_CPU; 4998475355fSChris Wilson 5008475355fSChris Wilson if (HAS_LLC(i915)) 5018475355fSChris Wilson /* On some devices, we can have the GPU use the LLC (the CPU 5028475355fSChris Wilson * cache) for about a 10% performance improvement 5038475355fSChris Wilson * compared to uncached. Graphics requests other than 5048475355fSChris Wilson * display scanout are coherent with the CPU in 5058475355fSChris Wilson * accessing this cache. This means in this mode we 5068475355fSChris Wilson * don't need to clflush on the CPU side, and on the 5078475355fSChris Wilson * GPU side we only need to flush internal caches to 5088475355fSChris Wilson * get data visible to the CPU. 5098475355fSChris Wilson * 5108475355fSChris Wilson * However, we maintain the display planes as UC, and so 5118475355fSChris Wilson * need to rebind when first used as such. 5128475355fSChris Wilson */ 5138475355fSChris Wilson cache_level = I915_CACHE_LLC; 5148475355fSChris Wilson else 5158475355fSChris Wilson cache_level = I915_CACHE_NONE; 5168475355fSChris Wilson 5178475355fSChris Wilson i915_gem_object_set_cache_coherency(obj, cache_level); 5188475355fSChris Wilson 519da1184cdSMatthew Auld i915_gem_object_init_memory_region(obj, mem, 0); 5208475355fSChris Wilson 5218475355fSChris Wilson return obj; 5228475355fSChris Wilson 5238475355fSChris Wilson fail: 5248475355fSChris Wilson i915_gem_object_free(obj); 5258475355fSChris Wilson return ERR_PTR(ret); 5268475355fSChris Wilson } 5278475355fSChris Wilson 528da1184cdSMatthew Auld struct drm_i915_gem_object * 529da1184cdSMatthew Auld i915_gem_object_create_shmem(struct drm_i915_private *i915, 530da1184cdSMatthew Auld resource_size_t size) 531da1184cdSMatthew Auld { 532da1184cdSMatthew Auld return i915_gem_object_create_region(i915->mm.regions[INTEL_REGION_SMEM], 533da1184cdSMatthew Auld size, 0); 534da1184cdSMatthew Auld } 535da1184cdSMatthew Auld 5368475355fSChris Wilson /* Allocate a new GEM object and fill it with the supplied data */ 5378475355fSChris Wilson struct drm_i915_gem_object * 5388475355fSChris Wilson i915_gem_object_create_shmem_from_data(struct drm_i915_private *dev_priv, 539da1184cdSMatthew Auld const void *data, resource_size_t size) 5408475355fSChris Wilson { 5418475355fSChris Wilson struct drm_i915_gem_object *obj; 5428475355fSChris Wilson struct file *file; 543da1184cdSMatthew Auld resource_size_t offset; 5448475355fSChris Wilson int err; 5458475355fSChris Wilson 5468475355fSChris Wilson obj = i915_gem_object_create_shmem(dev_priv, round_up(size, PAGE_SIZE)); 5478475355fSChris Wilson if (IS_ERR(obj)) 5488475355fSChris Wilson return obj; 5498475355fSChris Wilson 5508475355fSChris Wilson GEM_BUG_ON(obj->write_domain != I915_GEM_DOMAIN_CPU); 5518475355fSChris Wilson 5528475355fSChris Wilson file = obj->base.filp; 5538475355fSChris Wilson offset = 0; 5548475355fSChris Wilson do { 5558475355fSChris Wilson unsigned int len = min_t(typeof(size), size, PAGE_SIZE); 5568475355fSChris Wilson struct page *page; 5578475355fSChris Wilson void *pgdata, *vaddr; 5588475355fSChris Wilson 5598475355fSChris Wilson err = pagecache_write_begin(file, file->f_mapping, 5608475355fSChris Wilson offset, len, 0, 5618475355fSChris Wilson &page, &pgdata); 5628475355fSChris Wilson if (err < 0) 5638475355fSChris Wilson goto fail; 5648475355fSChris Wilson 5658475355fSChris Wilson vaddr = kmap(page); 5668475355fSChris Wilson memcpy(vaddr, data, len); 5678475355fSChris Wilson kunmap(page); 5688475355fSChris Wilson 5698475355fSChris Wilson err = pagecache_write_end(file, file->f_mapping, 5708475355fSChris Wilson offset, len, len, 5718475355fSChris Wilson page, pgdata); 5728475355fSChris Wilson if (err < 0) 5738475355fSChris Wilson goto fail; 5748475355fSChris Wilson 5758475355fSChris Wilson size -= len; 5768475355fSChris Wilson data += len; 5778475355fSChris Wilson offset += len; 5788475355fSChris Wilson } while (size); 5798475355fSChris Wilson 5808475355fSChris Wilson return obj; 5818475355fSChris Wilson 5828475355fSChris Wilson fail: 5838475355fSChris Wilson i915_gem_object_put(obj); 5848475355fSChris Wilson return ERR_PTR(err); 5858475355fSChris Wilson } 586da1184cdSMatthew Auld 587da1184cdSMatthew Auld static int init_shmem(struct intel_memory_region *mem) 588da1184cdSMatthew Auld { 589da1184cdSMatthew Auld int err; 590da1184cdSMatthew Auld 591da1184cdSMatthew Auld err = i915_gemfs_init(mem->i915); 592da1184cdSMatthew Auld if (err) { 593da1184cdSMatthew Auld DRM_NOTE("Unable to create a private tmpfs mount, hugepage support will be disabled(%d).\n", 594da1184cdSMatthew Auld err); 595da1184cdSMatthew Auld } 596da1184cdSMatthew Auld 59738f1cb68SLukasz Fiedorowicz intel_memory_region_set_name(mem, "system"); 59838f1cb68SLukasz Fiedorowicz 599da1184cdSMatthew Auld return 0; /* Don't error, we can simply fallback to the kernel mnt */ 600da1184cdSMatthew Auld } 601da1184cdSMatthew Auld 602da1184cdSMatthew Auld static void release_shmem(struct intel_memory_region *mem) 603da1184cdSMatthew Auld { 604da1184cdSMatthew Auld i915_gemfs_fini(mem->i915); 605da1184cdSMatthew Auld } 606da1184cdSMatthew Auld 607da1184cdSMatthew Auld static const struct intel_memory_region_ops shmem_region_ops = { 608da1184cdSMatthew Auld .init = init_shmem, 609da1184cdSMatthew Auld .release = release_shmem, 610da1184cdSMatthew Auld .create_object = create_shmem, 611da1184cdSMatthew Auld }; 612da1184cdSMatthew Auld 613da1184cdSMatthew Auld struct intel_memory_region *i915_gem_shmem_setup(struct drm_i915_private *i915) 614da1184cdSMatthew Auld { 615da1184cdSMatthew Auld return intel_memory_region_create(i915, 0, 616da1184cdSMatthew Auld totalram_pages() << PAGE_SHIFT, 617da1184cdSMatthew Auld PAGE_SIZE, 0, 618da1184cdSMatthew Auld &shmem_region_ops); 619da1184cdSMatthew Auld } 620