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 10*5f2ec909SJani Nikula #include <drm/drm_cache.h> 11*5f2ec909SJani Nikula 12da1184cdSMatthew Auld #include "gem/i915_gem_region.h" 138475355fSChris Wilson #include "i915_drv.h" 14da1184cdSMatthew Auld #include "i915_gemfs.h" 158475355fSChris Wilson #include "i915_gem_object.h" 1637d63f8fSChris Wilson #include "i915_scatterlist.h" 17a09d9a80SJani Nikula #include "i915_trace.h" 188475355fSChris Wilson 198475355fSChris Wilson /* 208475355fSChris Wilson * Move pages to appropriate lru and release the pagevec, decrementing the 218475355fSChris Wilson * ref count of those pages. 228475355fSChris Wilson */ 238475355fSChris Wilson static void check_release_pagevec(struct pagevec *pvec) 248475355fSChris Wilson { 258475355fSChris Wilson check_move_unevictable_pages(pvec); 268475355fSChris Wilson __pagevec_release(pvec); 278475355fSChris Wilson cond_resched(); 288475355fSChris Wilson } 298475355fSChris Wilson 30cad7109aSThomas Hellström void shmem_sg_free_table(struct sg_table *st, struct address_space *mapping, 31f05b985eSThomas Hellström bool dirty, bool backup) 328475355fSChris Wilson { 33f05b985eSThomas Hellström struct sgt_iter sgt_iter; 34f05b985eSThomas Hellström struct pagevec pvec; 35f05b985eSThomas Hellström struct page *page; 36f05b985eSThomas Hellström 37f05b985eSThomas Hellström mapping_clear_unevictable(mapping); 38f05b985eSThomas Hellström 39f05b985eSThomas Hellström pagevec_init(&pvec); 40f05b985eSThomas Hellström for_each_sgt_page(page, sgt_iter, st) { 41f05b985eSThomas Hellström if (dirty) 42f05b985eSThomas Hellström set_page_dirty(page); 43f05b985eSThomas Hellström 44f05b985eSThomas Hellström if (backup) 45f05b985eSThomas Hellström mark_page_accessed(page); 46f05b985eSThomas Hellström 47f05b985eSThomas Hellström if (!pagevec_add(&pvec, page)) 48f05b985eSThomas Hellström check_release_pagevec(&pvec); 49f05b985eSThomas Hellström } 50f05b985eSThomas Hellström if (pagevec_count(&pvec)) 51f05b985eSThomas Hellström check_release_pagevec(&pvec); 52f05b985eSThomas Hellström 53f05b985eSThomas Hellström sg_free_table(st); 54f05b985eSThomas Hellström } 55f05b985eSThomas Hellström 56cad7109aSThomas Hellström int shmem_sg_alloc_table(struct drm_i915_private *i915, struct sg_table *st, 57f05b985eSThomas Hellström size_t size, struct intel_memory_region *mr, 58f05b985eSThomas Hellström struct address_space *mapping, 59f05b985eSThomas Hellström unsigned int max_segment) 60f05b985eSThomas Hellström { 61f05b985eSThomas Hellström const unsigned long page_count = size / PAGE_SIZE; 628475355fSChris Wilson unsigned long i; 638475355fSChris Wilson struct scatterlist *sg; 648475355fSChris Wilson struct page *page; 658475355fSChris Wilson unsigned long last_pfn = 0; /* suppress gcc warning */ 668475355fSChris Wilson gfp_t noreclaim; 678475355fSChris Wilson int ret; 688475355fSChris Wilson 698475355fSChris Wilson /* 708475355fSChris Wilson * If there's no chance of allocating enough pages for the whole 718475355fSChris Wilson * object, bail early. 728475355fSChris Wilson */ 73f05b985eSThomas Hellström if (size > resource_size(&mr->region)) 74cad7109aSThomas Hellström return -ENOMEM; 758475355fSChris Wilson 76cad7109aSThomas Hellström if (sg_alloc_table(st, page_count, GFP_KERNEL)) 77cad7109aSThomas Hellström return -ENOMEM; 788475355fSChris Wilson 798475355fSChris Wilson /* 808475355fSChris Wilson * Get the list of pages out of our struct file. They'll be pinned 818475355fSChris Wilson * at this point until we release them. 828475355fSChris Wilson * 838475355fSChris Wilson * Fail silently without starting the shrinker 848475355fSChris Wilson */ 858475355fSChris Wilson mapping_set_unevictable(mapping); 868475355fSChris Wilson noreclaim = mapping_gfp_constraint(mapping, ~__GFP_RECLAIM); 878475355fSChris Wilson noreclaim |= __GFP_NORETRY | __GFP_NOWARN; 888475355fSChris Wilson 898475355fSChris Wilson sg = st->sgl; 908475355fSChris Wilson st->nents = 0; 918475355fSChris Wilson for (i = 0; i < page_count; i++) { 928475355fSChris Wilson const unsigned int shrink[] = { 933b4fa964SChris Wilson I915_SHRINK_BOUND | I915_SHRINK_UNBOUND, 948475355fSChris Wilson 0, 958475355fSChris Wilson }, *s = shrink; 968475355fSChris Wilson gfp_t gfp = noreclaim; 978475355fSChris Wilson 988475355fSChris Wilson do { 998475355fSChris Wilson cond_resched(); 1008475355fSChris Wilson page = shmem_read_mapping_page_gfp(mapping, i, gfp); 1018475355fSChris Wilson if (!IS_ERR(page)) 1028475355fSChris Wilson break; 1038475355fSChris Wilson 1048475355fSChris Wilson if (!*s) { 1058475355fSChris Wilson ret = PTR_ERR(page); 1068475355fSChris Wilson goto err_sg; 1078475355fSChris Wilson } 1088475355fSChris Wilson 109cf41a8f1SMaarten Lankhorst i915_gem_shrink(NULL, i915, 2 * page_count, NULL, *s++); 1108475355fSChris Wilson 1118475355fSChris Wilson /* 1128475355fSChris Wilson * We've tried hard to allocate the memory by reaping 1138475355fSChris Wilson * our own buffer, now let the real VM do its job and 1148475355fSChris Wilson * go down in flames if truly OOM. 1158475355fSChris Wilson * 1168475355fSChris Wilson * However, since graphics tend to be disposable, 1178475355fSChris Wilson * defer the oom here by reporting the ENOMEM back 1188475355fSChris Wilson * to userspace. 1198475355fSChris Wilson */ 1208475355fSChris Wilson if (!*s) { 1218475355fSChris Wilson /* reclaim and warn, but no oom */ 1228475355fSChris Wilson gfp = mapping_gfp_mask(mapping); 1238475355fSChris Wilson 1248475355fSChris Wilson /* 1258475355fSChris Wilson * Our bo are always dirty and so we require 1268475355fSChris Wilson * kswapd to reclaim our pages (direct reclaim 1278475355fSChris Wilson * does not effectively begin pageout of our 1288475355fSChris Wilson * buffers on its own). However, direct reclaim 1298475355fSChris Wilson * only waits for kswapd when under allocation 1308475355fSChris Wilson * congestion. So as a result __GFP_RECLAIM is 1318475355fSChris Wilson * unreliable and fails to actually reclaim our 1328475355fSChris Wilson * dirty pages -- unless you try over and over 1338475355fSChris Wilson * again with !__GFP_NORETRY. However, we still 1348475355fSChris Wilson * want to fail this allocation rather than 1358475355fSChris Wilson * trigger the out-of-memory killer and for 1368475355fSChris Wilson * this we want __GFP_RETRY_MAYFAIL. 1378475355fSChris Wilson */ 1388475355fSChris Wilson gfp |= __GFP_RETRY_MAYFAIL; 1398475355fSChris Wilson } 1408475355fSChris Wilson } while (1); 1418475355fSChris Wilson 1428475355fSChris Wilson if (!i || 1438475355fSChris Wilson sg->length >= max_segment || 1448475355fSChris Wilson page_to_pfn(page) != last_pfn + 1) { 145f05b985eSThomas Hellström if (i) 1468475355fSChris Wilson sg = sg_next(sg); 147f05b985eSThomas Hellström 1488475355fSChris Wilson st->nents++; 1498475355fSChris Wilson sg_set_page(sg, page, PAGE_SIZE, 0); 1508475355fSChris Wilson } else { 1518475355fSChris Wilson sg->length += PAGE_SIZE; 1528475355fSChris Wilson } 1538475355fSChris Wilson last_pfn = page_to_pfn(page); 1548475355fSChris Wilson 1558475355fSChris Wilson /* Check that the i965g/gm workaround works. */ 156ea97c4caSChris Wilson GEM_BUG_ON(gfp & __GFP_DMA32 && last_pfn >= 0x00100000UL); 1578475355fSChris Wilson } 158f05b985eSThomas Hellström if (sg) /* loop terminated early; short sg table */ 1598475355fSChris Wilson sg_mark_end(sg); 1608475355fSChris Wilson 1618475355fSChris Wilson /* Trim unused sg entries to avoid wasting memory. */ 1628475355fSChris Wilson i915_sg_trim(st); 1638475355fSChris Wilson 164cad7109aSThomas Hellström return 0; 165f05b985eSThomas Hellström err_sg: 166f05b985eSThomas Hellström sg_mark_end(sg); 167f05b985eSThomas Hellström if (sg != st->sgl) { 168cad7109aSThomas Hellström shmem_sg_free_table(st, mapping, false, false); 169f05b985eSThomas Hellström } else { 170f05b985eSThomas Hellström mapping_clear_unevictable(mapping); 171f05b985eSThomas Hellström sg_free_table(st); 172f05b985eSThomas Hellström } 173f05b985eSThomas Hellström 174f05b985eSThomas Hellström /* 175f05b985eSThomas Hellström * shmemfs first checks if there is enough memory to allocate the page 176f05b985eSThomas Hellström * and reports ENOSPC should there be insufficient, along with the usual 177f05b985eSThomas Hellström * ENOMEM for a genuine allocation failure. 178f05b985eSThomas Hellström * 179f05b985eSThomas Hellström * We use ENOSPC in our driver to mean that we have run out of aperture 180f05b985eSThomas Hellström * space and so want to translate the error from shmemfs back to our 181f05b985eSThomas Hellström * usual understanding of ENOMEM. 182f05b985eSThomas Hellström */ 183f05b985eSThomas Hellström if (ret == -ENOSPC) 184f05b985eSThomas Hellström ret = -ENOMEM; 185f05b985eSThomas Hellström 186cad7109aSThomas Hellström return ret; 187f05b985eSThomas Hellström } 188f05b985eSThomas Hellström 189f05b985eSThomas Hellström static int shmem_get_pages(struct drm_i915_gem_object *obj) 190f05b985eSThomas Hellström { 191f05b985eSThomas Hellström struct drm_i915_private *i915 = to_i915(obj->base.dev); 192f05b985eSThomas Hellström struct intel_memory_region *mem = obj->mm.region; 193f05b985eSThomas Hellström struct address_space *mapping = obj->base.filp->f_mapping; 194f05b985eSThomas Hellström const unsigned long page_count = obj->base.size / PAGE_SIZE; 195f05b985eSThomas Hellström unsigned int max_segment = i915_sg_segment_size(); 196f05b985eSThomas Hellström struct sg_table *st; 197f05b985eSThomas Hellström struct sgt_iter sgt_iter; 198f05b985eSThomas Hellström struct page *page; 199f05b985eSThomas Hellström int ret; 200f05b985eSThomas Hellström 201f05b985eSThomas Hellström /* 202f05b985eSThomas Hellström * Assert that the object is not currently in any GPU domain. As it 203f05b985eSThomas Hellström * wasn't in the GTT, there shouldn't be any way it could have been in 204f05b985eSThomas Hellström * a GPU cache 205f05b985eSThomas Hellström */ 206f05b985eSThomas Hellström GEM_BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS); 207f05b985eSThomas Hellström GEM_BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS); 208f05b985eSThomas Hellström 209f05b985eSThomas Hellström rebuild_st: 210cad7109aSThomas Hellström st = kmalloc(sizeof(*st), GFP_KERNEL); 211cad7109aSThomas Hellström if (!st) 212cad7109aSThomas Hellström return -ENOMEM; 213cad7109aSThomas Hellström 214cad7109aSThomas Hellström ret = shmem_sg_alloc_table(i915, st, obj->base.size, mem, mapping, 215cad7109aSThomas Hellström max_segment); 216cad7109aSThomas Hellström if (ret) 217f05b985eSThomas Hellström goto err_st; 218f05b985eSThomas Hellström 2198475355fSChris Wilson ret = i915_gem_gtt_prepare_pages(obj, st); 2208475355fSChris Wilson if (ret) { 2218475355fSChris Wilson /* 2228475355fSChris Wilson * DMA remapping failed? One possible cause is that 2238475355fSChris Wilson * it could not reserve enough large entries, asking 2248475355fSChris Wilson * for PAGE_SIZE chunks instead may be helpful. 2258475355fSChris Wilson */ 2268475355fSChris Wilson if (max_segment > PAGE_SIZE) { 2278475355fSChris Wilson for_each_sgt_page(page, sgt_iter, st) 2288475355fSChris Wilson put_page(page); 2298475355fSChris Wilson sg_free_table(st); 230f05b985eSThomas Hellström kfree(st); 2318475355fSChris Wilson 2328475355fSChris Wilson max_segment = PAGE_SIZE; 2338475355fSChris Wilson goto rebuild_st; 2348475355fSChris Wilson } else { 2358ff5446aSThomas Zimmermann dev_warn(i915->drm.dev, 2368475355fSChris Wilson "Failed to DMA remap %lu pages\n", 2378475355fSChris Wilson page_count); 2388475355fSChris Wilson goto err_pages; 2398475355fSChris Wilson } 2408475355fSChris Wilson } 2418475355fSChris Wilson 2428475355fSChris Wilson if (i915_gem_object_needs_bit17_swizzle(obj)) 2438475355fSChris Wilson i915_gem_object_do_bit_17_swizzle(obj, st); 2448475355fSChris Wilson 24530f1dccdSMatthew Auld if (i915_gem_object_can_bypass_llc(obj)) 24613d29c82SMatthew Auld obj->cache_dirty = true; 24713d29c82SMatthew Auld 248f05b985eSThomas Hellström __i915_gem_object_set_pages(obj, st, i915_sg_dma_sizes(st->sgl)); 2498475355fSChris Wilson 2508475355fSChris Wilson return 0; 2518475355fSChris Wilson 2528475355fSChris Wilson err_pages: 253cad7109aSThomas Hellström shmem_sg_free_table(st, mapping, false, false); 2548475355fSChris Wilson /* 2558475355fSChris Wilson * shmemfs first checks if there is enough memory to allocate the page 2568475355fSChris Wilson * and reports ENOSPC should there be insufficient, along with the usual 2578475355fSChris Wilson * ENOMEM for a genuine allocation failure. 2588475355fSChris Wilson * 2598475355fSChris Wilson * We use ENOSPC in our driver to mean that we have run out of aperture 2608475355fSChris Wilson * space and so want to translate the error from shmemfs back to our 2618475355fSChris Wilson * usual understanding of ENOMEM. 2628475355fSChris Wilson */ 263f05b985eSThomas Hellström err_st: 2648475355fSChris Wilson if (ret == -ENOSPC) 2658475355fSChris Wilson ret = -ENOMEM; 2668475355fSChris Wilson 267cad7109aSThomas Hellström kfree(st); 268cad7109aSThomas Hellström 2698475355fSChris Wilson return ret; 2708475355fSChris Wilson } 2718475355fSChris Wilson 2727ae03459SMatthew Auld static int 273f033428dSChris Wilson shmem_truncate(struct drm_i915_gem_object *obj) 274f033428dSChris Wilson { 275f033428dSChris Wilson /* 276f033428dSChris Wilson * Our goal here is to return as much of the memory as 277f033428dSChris Wilson * is possible back to the system as we are called from OOM. 278f033428dSChris Wilson * To do this we must instruct the shmfs to drop all of its 279f033428dSChris Wilson * backing pages, *now*. 280f033428dSChris Wilson */ 281f033428dSChris Wilson shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1); 282f033428dSChris Wilson obj->mm.madv = __I915_MADV_PURGED; 283f033428dSChris Wilson obj->mm.pages = ERR_PTR(-EFAULT); 2847ae03459SMatthew Auld 2857ae03459SMatthew Auld return 0; 286f033428dSChris Wilson } 287f033428dSChris Wilson 2887ae03459SMatthew Auld void __shmem_writeback(size_t size, struct address_space *mapping) 289f033428dSChris Wilson { 290f033428dSChris Wilson struct writeback_control wbc = { 291f033428dSChris Wilson .sync_mode = WB_SYNC_NONE, 292f033428dSChris Wilson .nr_to_write = SWAP_CLUSTER_MAX, 293f033428dSChris Wilson .range_start = 0, 294f033428dSChris Wilson .range_end = LLONG_MAX, 295f033428dSChris Wilson .for_reclaim = 1, 296f033428dSChris Wilson }; 297f033428dSChris Wilson unsigned long i; 298f033428dSChris Wilson 299f033428dSChris Wilson /* 300f033428dSChris Wilson * Leave mmapings intact (GTT will have been revoked on unbinding, 301f033428dSChris Wilson * leaving only CPU mmapings around) and add those pages to the LRU 302f033428dSChris Wilson * instead of invoking writeback so they are aged and paged out 303f033428dSChris Wilson * as normal. 304f033428dSChris Wilson */ 305f033428dSChris Wilson 306f033428dSChris Wilson /* Begin writeback on each dirty page */ 307f05b985eSThomas Hellström for (i = 0; i < size >> PAGE_SHIFT; i++) { 308f033428dSChris Wilson struct page *page; 309f033428dSChris Wilson 3109dfc8ff3SMatthew Wilcox (Oracle) page = find_lock_page(mapping, i); 3119dfc8ff3SMatthew Wilcox (Oracle) if (!page) 312f033428dSChris Wilson continue; 313f033428dSChris Wilson 314f033428dSChris Wilson if (!page_mapped(page) && clear_page_dirty_for_io(page)) { 315f033428dSChris Wilson int ret; 316f033428dSChris Wilson 317f033428dSChris Wilson SetPageReclaim(page); 318f033428dSChris Wilson ret = mapping->a_ops->writepage(page, &wbc); 319f033428dSChris Wilson if (!PageWriteback(page)) 320f033428dSChris Wilson ClearPageReclaim(page); 321f033428dSChris Wilson if (!ret) 322f033428dSChris Wilson goto put; 323f033428dSChris Wilson } 324f033428dSChris Wilson unlock_page(page); 325f033428dSChris Wilson put: 326f033428dSChris Wilson put_page(page); 327f033428dSChris Wilson } 328f033428dSChris Wilson } 329f033428dSChris Wilson 330f05b985eSThomas Hellström static void 331f05b985eSThomas Hellström shmem_writeback(struct drm_i915_gem_object *obj) 332f05b985eSThomas Hellström { 333f05b985eSThomas Hellström __shmem_writeback(obj->base.size, obj->base.filp->f_mapping); 334f05b985eSThomas Hellström } 335f05b985eSThomas Hellström 3368475355fSChris Wilson void 3378475355fSChris Wilson __i915_gem_object_release_shmem(struct drm_i915_gem_object *obj, 3388475355fSChris Wilson struct sg_table *pages, 3398475355fSChris Wilson bool needs_clflush) 3408475355fSChris Wilson { 341d70af579SMatthew Auld struct drm_i915_private *i915 = to_i915(obj->base.dev); 342d70af579SMatthew Auld 3438475355fSChris Wilson GEM_BUG_ON(obj->mm.madv == __I915_MADV_PURGED); 3448475355fSChris Wilson 3458475355fSChris Wilson if (obj->mm.madv == I915_MADV_DONTNEED) 3468475355fSChris Wilson obj->mm.dirty = false; 3478475355fSChris Wilson 3488475355fSChris Wilson if (needs_clflush && 3498475355fSChris Wilson (obj->read_domains & I915_GEM_DOMAIN_CPU) == 0 && 3508475355fSChris Wilson !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ)) 3518475355fSChris Wilson drm_clflush_sg(pages); 3528475355fSChris Wilson 3538475355fSChris Wilson __start_cpu_write(obj); 354d70af579SMatthew Auld /* 355d70af579SMatthew Auld * On non-LLC platforms, force the flush-on-acquire if this is ever 356d70af579SMatthew Auld * swapped-in. Our async flush path is not trust worthy enough yet(and 357d70af579SMatthew Auld * happens in the wrong order), and with some tricks it's conceivable 358d70af579SMatthew Auld * for userspace to change the cache-level to I915_CACHE_NONE after the 359d70af579SMatthew Auld * pages are swapped-in, and since execbuf binds the object before doing 360d70af579SMatthew Auld * the async flush, we have a race window. 361d70af579SMatthew Auld */ 362d70af579SMatthew Auld if (!HAS_LLC(i915)) 363d70af579SMatthew Auld obj->cache_dirty = true; 3648475355fSChris Wilson } 3658475355fSChris Wilson 366a85fffe3SMaarten Lankhorst void i915_gem_object_put_pages_shmem(struct drm_i915_gem_object *obj, struct sg_table *pages) 3678475355fSChris Wilson { 3688475355fSChris Wilson __i915_gem_object_release_shmem(obj, pages, true); 3698475355fSChris Wilson 3708475355fSChris Wilson i915_gem_gtt_finish_pages(obj, pages); 3718475355fSChris Wilson 3728475355fSChris Wilson if (i915_gem_object_needs_bit17_swizzle(obj)) 3738475355fSChris Wilson i915_gem_object_save_bit_17_swizzle(obj, pages); 3748475355fSChris Wilson 375cad7109aSThomas Hellström shmem_sg_free_table(pages, file_inode(obj->base.filp)->i_mapping, 376f05b985eSThomas Hellström obj->mm.dirty, obj->mm.madv == I915_MADV_WILLNEED); 377cad7109aSThomas Hellström kfree(pages); 3788475355fSChris Wilson obj->mm.dirty = false; 3798475355fSChris Wilson } 3808475355fSChris Wilson 381a85fffe3SMaarten Lankhorst static void 382a85fffe3SMaarten Lankhorst shmem_put_pages(struct drm_i915_gem_object *obj, struct sg_table *pages) 383a85fffe3SMaarten Lankhorst { 384a85fffe3SMaarten Lankhorst if (likely(i915_gem_object_has_struct_page(obj))) 385a85fffe3SMaarten Lankhorst i915_gem_object_put_pages_shmem(obj, pages); 386a85fffe3SMaarten Lankhorst else 387a85fffe3SMaarten Lankhorst i915_gem_object_put_pages_phys(obj, pages); 388a85fffe3SMaarten Lankhorst } 389a85fffe3SMaarten Lankhorst 3908475355fSChris Wilson static int 3918475355fSChris Wilson shmem_pwrite(struct drm_i915_gem_object *obj, 3928475355fSChris Wilson const struct drm_i915_gem_pwrite *arg) 3938475355fSChris Wilson { 3948475355fSChris Wilson struct address_space *mapping = obj->base.filp->f_mapping; 3958475355fSChris Wilson char __user *user_data = u64_to_user_ptr(arg->data_ptr); 3968475355fSChris Wilson u64 remain, offset; 3978475355fSChris Wilson unsigned int pg; 3988475355fSChris Wilson 3998475355fSChris Wilson /* Caller already validated user args */ 4008475355fSChris Wilson GEM_BUG_ON(!access_ok(user_data, arg->size)); 4018475355fSChris Wilson 402a6117097SMaarten Lankhorst if (!i915_gem_object_has_struct_page(obj)) 403a6117097SMaarten Lankhorst return i915_gem_object_pwrite_phys(obj, arg); 404a6117097SMaarten Lankhorst 4058475355fSChris Wilson /* 4068475355fSChris Wilson * Before we instantiate/pin the backing store for our use, we 4078475355fSChris Wilson * can prepopulate the shmemfs filp efficiently using a write into 4088475355fSChris Wilson * the pagecache. We avoid the penalty of instantiating all the 4098475355fSChris Wilson * pages, important if the user is just writing to a few and never 4108475355fSChris Wilson * uses the object on the GPU, and using a direct write into shmemfs 4118475355fSChris Wilson * allows it to avoid the cost of retrieving a page (either swapin 4128475355fSChris Wilson * or clearing-before-use) before it is overwritten. 4138475355fSChris Wilson */ 4148475355fSChris Wilson if (i915_gem_object_has_pages(obj)) 4158475355fSChris Wilson return -ENODEV; 4168475355fSChris Wilson 4178475355fSChris Wilson if (obj->mm.madv != I915_MADV_WILLNEED) 4188475355fSChris Wilson return -EFAULT; 4198475355fSChris Wilson 4208475355fSChris Wilson /* 4218475355fSChris Wilson * Before the pages are instantiated the object is treated as being 4228475355fSChris Wilson * in the CPU domain. The pages will be clflushed as required before 4238475355fSChris Wilson * use, and we can freely write into the pages directly. If userspace 4248475355fSChris Wilson * races pwrite with any other operation; corruption will ensue - 4258475355fSChris Wilson * that is userspace's prerogative! 4268475355fSChris Wilson */ 4278475355fSChris Wilson 4288475355fSChris Wilson remain = arg->size; 4298475355fSChris Wilson offset = arg->offset; 4308475355fSChris Wilson pg = offset_in_page(offset); 4318475355fSChris Wilson 4328475355fSChris Wilson do { 4338475355fSChris Wilson unsigned int len, unwritten; 4348475355fSChris Wilson struct page *page; 4358475355fSChris Wilson void *data, *vaddr; 4368475355fSChris Wilson int err; 4378475355fSChris Wilson char c; 4388475355fSChris Wilson 4398475355fSChris Wilson len = PAGE_SIZE - pg; 4408475355fSChris Wilson if (len > remain) 4418475355fSChris Wilson len = remain; 4428475355fSChris Wilson 4438475355fSChris Wilson /* Prefault the user page to reduce potential recursion */ 4448475355fSChris Wilson err = __get_user(c, user_data); 4458475355fSChris Wilson if (err) 4468475355fSChris Wilson return err; 4478475355fSChris Wilson 4488475355fSChris Wilson err = __get_user(c, user_data + len - 1); 4498475355fSChris Wilson if (err) 4508475355fSChris Wilson return err; 4518475355fSChris Wilson 4528475355fSChris Wilson err = pagecache_write_begin(obj->base.filp, mapping, 4538475355fSChris Wilson offset, len, 0, 4548475355fSChris Wilson &page, &data); 4558475355fSChris Wilson if (err < 0) 4568475355fSChris Wilson return err; 4578475355fSChris Wilson 4588475355fSChris Wilson vaddr = kmap_atomic(page); 4598475355fSChris Wilson unwritten = __copy_from_user_inatomic(vaddr + pg, 4608475355fSChris Wilson user_data, 4618475355fSChris Wilson len); 4628475355fSChris Wilson kunmap_atomic(vaddr); 4638475355fSChris Wilson 4648475355fSChris Wilson err = pagecache_write_end(obj->base.filp, mapping, 4658475355fSChris Wilson offset, len, len - unwritten, 4668475355fSChris Wilson page, data); 4678475355fSChris Wilson if (err < 0) 4688475355fSChris Wilson return err; 4698475355fSChris Wilson 4708475355fSChris Wilson /* We don't handle -EFAULT, leave it to the caller to check */ 4718475355fSChris Wilson if (unwritten) 4728475355fSChris Wilson return -ENODEV; 4738475355fSChris Wilson 4748475355fSChris Wilson remain -= len; 4758475355fSChris Wilson user_data += len; 4768475355fSChris Wilson offset += len; 4778475355fSChris Wilson pg = 0; 4788475355fSChris Wilson } while (remain); 4798475355fSChris Wilson 4808475355fSChris Wilson return 0; 4818475355fSChris Wilson } 4828475355fSChris Wilson 483a6117097SMaarten Lankhorst static int 484a6117097SMaarten Lankhorst shmem_pread(struct drm_i915_gem_object *obj, 485a6117097SMaarten Lankhorst const struct drm_i915_gem_pread *arg) 486a6117097SMaarten Lankhorst { 487a6117097SMaarten Lankhorst if (!i915_gem_object_has_struct_page(obj)) 488a6117097SMaarten Lankhorst return i915_gem_object_pread_phys(obj, arg); 489a6117097SMaarten Lankhorst 490a6117097SMaarten Lankhorst return -ENODEV; 491a6117097SMaarten Lankhorst } 492a6117097SMaarten Lankhorst 4930c159ffeSChris Wilson static void shmem_release(struct drm_i915_gem_object *obj) 4940c159ffeSChris Wilson { 4950ff37575SThomas Hellström if (i915_gem_object_has_struct_page(obj)) 496da1184cdSMatthew Auld i915_gem_object_release_memory_region(obj); 497da1184cdSMatthew Auld 4980c159ffeSChris Wilson fput(obj->base.filp); 4990c159ffeSChris Wilson } 5000c159ffeSChris Wilson 5018475355fSChris Wilson const struct drm_i915_gem_object_ops i915_gem_shmem_ops = { 5027d192daaSChris Wilson .name = "i915_gem_object_shmem", 503c471748dSMaarten Lankhorst .flags = I915_GEM_OBJECT_IS_SHRINKABLE, 5048475355fSChris Wilson 5058475355fSChris Wilson .get_pages = shmem_get_pages, 5068475355fSChris Wilson .put_pages = shmem_put_pages, 507f033428dSChris Wilson .truncate = shmem_truncate, 508f033428dSChris Wilson .writeback = shmem_writeback, 5098475355fSChris Wilson 5108475355fSChris Wilson .pwrite = shmem_pwrite, 511a6117097SMaarten Lankhorst .pread = shmem_pread, 5120c159ffeSChris Wilson 5130c159ffeSChris Wilson .release = shmem_release, 5148475355fSChris Wilson }; 5158475355fSChris Wilson 516da1184cdSMatthew Auld static int __create_shmem(struct drm_i915_private *i915, 5178475355fSChris Wilson struct drm_gem_object *obj, 518da1184cdSMatthew Auld resource_size_t size) 5198475355fSChris Wilson { 5208475355fSChris Wilson unsigned long flags = VM_NORESERVE; 5218475355fSChris Wilson struct file *filp; 5228475355fSChris Wilson 5238475355fSChris Wilson drm_gem_private_object_init(&i915->drm, obj, size); 5248475355fSChris Wilson 5258475355fSChris Wilson if (i915->mm.gemfs) 5268475355fSChris Wilson filp = shmem_file_setup_with_mnt(i915->mm.gemfs, "i915", size, 5278475355fSChris Wilson flags); 5288475355fSChris Wilson else 5298475355fSChris Wilson filp = shmem_file_setup("i915", size, flags); 5308475355fSChris Wilson if (IS_ERR(filp)) 5318475355fSChris Wilson return PTR_ERR(filp); 5328475355fSChris Wilson 5338475355fSChris Wilson obj->filp = filp; 5348475355fSChris Wilson return 0; 5358475355fSChris Wilson } 5368475355fSChris Wilson 53797d55396SMatthew Auld static int shmem_object_init(struct intel_memory_region *mem, 53897d55396SMatthew Auld struct drm_i915_gem_object *obj, 539da1184cdSMatthew Auld resource_size_t size, 540d22632c8SMatthew Auld resource_size_t page_size, 541da1184cdSMatthew Auld unsigned int flags) 5428475355fSChris Wilson { 5437867d709SChris Wilson static struct lock_class_key lock_class; 544da1184cdSMatthew Auld struct drm_i915_private *i915 = mem->i915; 5458475355fSChris Wilson struct address_space *mapping; 5468475355fSChris Wilson unsigned int cache_level; 5478475355fSChris Wilson gfp_t mask; 5488475355fSChris Wilson int ret; 5498475355fSChris Wilson 550da1184cdSMatthew Auld ret = __create_shmem(i915, &obj->base, size); 5518475355fSChris Wilson if (ret) 55297d55396SMatthew Auld return ret; 5538475355fSChris Wilson 5548475355fSChris Wilson mask = GFP_HIGHUSER | __GFP_RECLAIMABLE; 5558475355fSChris Wilson if (IS_I965GM(i915) || IS_I965G(i915)) { 5568475355fSChris Wilson /* 965gm cannot relocate objects above 4GiB. */ 5578475355fSChris Wilson mask &= ~__GFP_HIGHMEM; 5588475355fSChris Wilson mask |= __GFP_DMA32; 5598475355fSChris Wilson } 5608475355fSChris Wilson 5618475355fSChris Wilson mapping = obj->base.filp->f_mapping; 5628475355fSChris Wilson mapping_set_gfp_mask(mapping, mask); 5638475355fSChris Wilson GEM_BUG_ON(!(mapping_gfp_mask(mapping) & __GFP_RECLAIM)); 5648475355fSChris Wilson 5650ff37575SThomas Hellström i915_gem_object_init(obj, &i915_gem_shmem_ops, &lock_class, 0); 5660ff37575SThomas Hellström obj->mem_flags |= I915_BO_FLAG_STRUCT_PAGE; 5678475355fSChris Wilson obj->write_domain = I915_GEM_DOMAIN_CPU; 5688475355fSChris Wilson obj->read_domains = I915_GEM_DOMAIN_CPU; 5698475355fSChris Wilson 5708475355fSChris Wilson if (HAS_LLC(i915)) 5718475355fSChris Wilson /* On some devices, we can have the GPU use the LLC (the CPU 5728475355fSChris Wilson * cache) for about a 10% performance improvement 5738475355fSChris Wilson * compared to uncached. Graphics requests other than 5748475355fSChris Wilson * display scanout are coherent with the CPU in 5758475355fSChris Wilson * accessing this cache. This means in this mode we 5768475355fSChris Wilson * don't need to clflush on the CPU side, and on the 5778475355fSChris Wilson * GPU side we only need to flush internal caches to 5788475355fSChris Wilson * get data visible to the CPU. 5798475355fSChris Wilson * 5808475355fSChris Wilson * However, we maintain the display planes as UC, and so 5818475355fSChris Wilson * need to rebind when first used as such. 5828475355fSChris Wilson */ 5838475355fSChris Wilson cache_level = I915_CACHE_LLC; 5848475355fSChris Wilson else 5858475355fSChris Wilson cache_level = I915_CACHE_NONE; 5868475355fSChris Wilson 5878475355fSChris Wilson i915_gem_object_set_cache_coherency(obj, cache_level); 5888475355fSChris Wilson 589c471748dSMaarten Lankhorst i915_gem_object_init_memory_region(obj, mem); 5908475355fSChris Wilson 59197d55396SMatthew Auld return 0; 5928475355fSChris Wilson } 5938475355fSChris Wilson 594da1184cdSMatthew Auld struct drm_i915_gem_object * 595da1184cdSMatthew Auld i915_gem_object_create_shmem(struct drm_i915_private *i915, 596da1184cdSMatthew Auld resource_size_t size) 597da1184cdSMatthew Auld { 598da1184cdSMatthew Auld return i915_gem_object_create_region(i915->mm.regions[INTEL_REGION_SMEM], 599d22632c8SMatthew Auld size, 0, 0); 600da1184cdSMatthew Auld } 601da1184cdSMatthew Auld 6028475355fSChris Wilson /* Allocate a new GEM object and fill it with the supplied data */ 6038475355fSChris Wilson struct drm_i915_gem_object * 6048475355fSChris Wilson i915_gem_object_create_shmem_from_data(struct drm_i915_private *dev_priv, 605da1184cdSMatthew Auld const void *data, resource_size_t size) 6068475355fSChris Wilson { 6078475355fSChris Wilson struct drm_i915_gem_object *obj; 6088475355fSChris Wilson struct file *file; 609da1184cdSMatthew Auld resource_size_t offset; 6108475355fSChris Wilson int err; 6118475355fSChris Wilson 61232b7cf51SThomas Hellström GEM_WARN_ON(IS_DGFX(dev_priv)); 6138475355fSChris Wilson obj = i915_gem_object_create_shmem(dev_priv, round_up(size, PAGE_SIZE)); 6148475355fSChris Wilson if (IS_ERR(obj)) 6158475355fSChris Wilson return obj; 6168475355fSChris Wilson 6178475355fSChris Wilson GEM_BUG_ON(obj->write_domain != I915_GEM_DOMAIN_CPU); 6188475355fSChris Wilson 6198475355fSChris Wilson file = obj->base.filp; 6208475355fSChris Wilson offset = 0; 6218475355fSChris Wilson do { 6228475355fSChris Wilson unsigned int len = min_t(typeof(size), size, PAGE_SIZE); 6238475355fSChris Wilson struct page *page; 6248475355fSChris Wilson void *pgdata, *vaddr; 6258475355fSChris Wilson 6268475355fSChris Wilson err = pagecache_write_begin(file, file->f_mapping, 6278475355fSChris Wilson offset, len, 0, 6288475355fSChris Wilson &page, &pgdata); 6298475355fSChris Wilson if (err < 0) 6308475355fSChris Wilson goto fail; 6318475355fSChris Wilson 6328475355fSChris Wilson vaddr = kmap(page); 6338475355fSChris Wilson memcpy(vaddr, data, len); 6348475355fSChris Wilson kunmap(page); 6358475355fSChris Wilson 6368475355fSChris Wilson err = pagecache_write_end(file, file->f_mapping, 6378475355fSChris Wilson offset, len, len, 6388475355fSChris Wilson page, pgdata); 6398475355fSChris Wilson if (err < 0) 6408475355fSChris Wilson goto fail; 6418475355fSChris Wilson 6428475355fSChris Wilson size -= len; 6438475355fSChris Wilson data += len; 6448475355fSChris Wilson offset += len; 6458475355fSChris Wilson } while (size); 6468475355fSChris Wilson 6478475355fSChris Wilson return obj; 6488475355fSChris Wilson 6498475355fSChris Wilson fail: 6508475355fSChris Wilson i915_gem_object_put(obj); 6518475355fSChris Wilson return ERR_PTR(err); 6528475355fSChris Wilson } 653da1184cdSMatthew Auld 654da1184cdSMatthew Auld static int init_shmem(struct intel_memory_region *mem) 655da1184cdSMatthew Auld { 656da1184cdSMatthew Auld int err; 657da1184cdSMatthew Auld 658da1184cdSMatthew Auld err = i915_gemfs_init(mem->i915); 659da1184cdSMatthew Auld if (err) { 660da1184cdSMatthew Auld DRM_NOTE("Unable to create a private tmpfs mount, hugepage support will be disabled(%d).\n", 661da1184cdSMatthew Auld err); 662da1184cdSMatthew Auld } 663da1184cdSMatthew Auld 66438f1cb68SLukasz Fiedorowicz intel_memory_region_set_name(mem, "system"); 66538f1cb68SLukasz Fiedorowicz 666da1184cdSMatthew Auld return 0; /* Don't error, we can simply fallback to the kernel mnt */ 667da1184cdSMatthew Auld } 668da1184cdSMatthew Auld 6698b1f7f92SThomas Hellström static int release_shmem(struct intel_memory_region *mem) 670da1184cdSMatthew Auld { 671da1184cdSMatthew Auld i915_gemfs_fini(mem->i915); 6728b1f7f92SThomas Hellström return 0; 673da1184cdSMatthew Auld } 674da1184cdSMatthew Auld 675da1184cdSMatthew Auld static const struct intel_memory_region_ops shmem_region_ops = { 676da1184cdSMatthew Auld .init = init_shmem, 677da1184cdSMatthew Auld .release = release_shmem, 67897d55396SMatthew Auld .init_object = shmem_object_init, 679da1184cdSMatthew Auld }; 680da1184cdSMatthew Auld 681d1487389SThomas Hellström struct intel_memory_region *i915_gem_shmem_setup(struct drm_i915_private *i915, 682d1487389SThomas Hellström u16 type, u16 instance) 683da1184cdSMatthew Auld { 684da1184cdSMatthew Auld return intel_memory_region_create(i915, 0, 685da1184cdSMatthew Auld totalram_pages() << PAGE_SHIFT, 686da1184cdSMatthew Auld PAGE_SIZE, 0, 687d1487389SThomas Hellström type, instance, 688da1184cdSMatthew Auld &shmem_region_ops); 689da1184cdSMatthew Auld } 69041a9c75dSChris Wilson 69141a9c75dSChris Wilson bool i915_gem_object_is_shmem(const struct drm_i915_gem_object *obj) 69241a9c75dSChris Wilson { 69341a9c75dSChris Wilson return obj->ops == &i915_gem_shmem_ops; 69441a9c75dSChris Wilson } 695