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> 882508de2SJani Nikula #include <linux/shmem_fs.h> 98475355fSChris Wilson #include <linux/swap.h> 108475355fSChris Wilson 115f2ec909SJani Nikula #include <drm/drm_cache.h> 125f2ec909SJani Nikula 13da1184cdSMatthew Auld #include "gem/i915_gem_region.h" 148475355fSChris Wilson #include "i915_drv.h" 15da1184cdSMatthew Auld #include "i915_gemfs.h" 168475355fSChris Wilson #include "i915_gem_object.h" 1737d63f8fSChris Wilson #include "i915_scatterlist.h" 18a09d9a80SJani Nikula #include "i915_trace.h" 198475355fSChris Wilson 208475355fSChris Wilson /* 218475355fSChris Wilson * Move pages to appropriate lru and release the pagevec, decrementing the 228475355fSChris Wilson * ref count of those pages. 238475355fSChris Wilson */ 248475355fSChris Wilson static void check_release_pagevec(struct pagevec *pvec) 258475355fSChris Wilson { 268475355fSChris Wilson check_move_unevictable_pages(pvec); 278475355fSChris Wilson __pagevec_release(pvec); 288475355fSChris Wilson cond_resched(); 298475355fSChris Wilson } 308475355fSChris Wilson 31cad7109aSThomas Hellström void shmem_sg_free_table(struct sg_table *st, struct address_space *mapping, 32f05b985eSThomas Hellström bool dirty, bool backup) 338475355fSChris Wilson { 34f05b985eSThomas Hellström struct sgt_iter sgt_iter; 35f05b985eSThomas Hellström struct pagevec pvec; 36f05b985eSThomas Hellström struct page *page; 37f05b985eSThomas Hellström 38f05b985eSThomas Hellström mapping_clear_unevictable(mapping); 39f05b985eSThomas Hellström 40f05b985eSThomas Hellström pagevec_init(&pvec); 41f05b985eSThomas Hellström for_each_sgt_page(page, sgt_iter, st) { 42f05b985eSThomas Hellström if (dirty) 43f05b985eSThomas Hellström set_page_dirty(page); 44f05b985eSThomas Hellström 45f05b985eSThomas Hellström if (backup) 46f05b985eSThomas Hellström mark_page_accessed(page); 47f05b985eSThomas Hellström 48f05b985eSThomas Hellström if (!pagevec_add(&pvec, page)) 49f05b985eSThomas Hellström check_release_pagevec(&pvec); 50f05b985eSThomas Hellström } 51f05b985eSThomas Hellström if (pagevec_count(&pvec)) 52f05b985eSThomas Hellström check_release_pagevec(&pvec); 53f05b985eSThomas Hellström 54f05b985eSThomas Hellström sg_free_table(st); 55f05b985eSThomas Hellström } 56f05b985eSThomas Hellström 57cad7109aSThomas Hellström int shmem_sg_alloc_table(struct drm_i915_private *i915, struct sg_table *st, 58f05b985eSThomas Hellström size_t size, struct intel_memory_region *mr, 59f05b985eSThomas Hellström struct address_space *mapping, 60f05b985eSThomas Hellström unsigned int max_segment) 61f05b985eSThomas Hellström { 62f05b985eSThomas Hellström const unsigned long page_count = size / PAGE_SIZE; 638475355fSChris Wilson unsigned long i; 648475355fSChris Wilson struct scatterlist *sg; 658475355fSChris Wilson struct page *page; 668475355fSChris Wilson unsigned long last_pfn = 0; /* suppress gcc warning */ 678475355fSChris Wilson gfp_t noreclaim; 688475355fSChris Wilson int ret; 698475355fSChris Wilson 708475355fSChris Wilson /* 718475355fSChris Wilson * If there's no chance of allocating enough pages for the whole 728475355fSChris Wilson * object, bail early. 738475355fSChris Wilson */ 74f05b985eSThomas Hellström if (size > resource_size(&mr->region)) 75cad7109aSThomas Hellström return -ENOMEM; 768475355fSChris Wilson 77cad7109aSThomas Hellström if (sg_alloc_table(st, page_count, GFP_KERNEL)) 78cad7109aSThomas Hellström return -ENOMEM; 798475355fSChris Wilson 808475355fSChris Wilson /* 818475355fSChris Wilson * Get the list of pages out of our struct file. They'll be pinned 828475355fSChris Wilson * at this point until we release them. 838475355fSChris Wilson * 848475355fSChris Wilson * Fail silently without starting the shrinker 858475355fSChris Wilson */ 868475355fSChris Wilson mapping_set_unevictable(mapping); 878475355fSChris Wilson noreclaim = mapping_gfp_constraint(mapping, ~__GFP_RECLAIM); 888475355fSChris Wilson noreclaim |= __GFP_NORETRY | __GFP_NOWARN; 898475355fSChris Wilson 908475355fSChris Wilson sg = st->sgl; 918475355fSChris Wilson st->nents = 0; 928475355fSChris Wilson for (i = 0; i < page_count; i++) { 938475355fSChris Wilson const unsigned int shrink[] = { 943b4fa964SChris Wilson I915_SHRINK_BOUND | I915_SHRINK_UNBOUND, 958475355fSChris Wilson 0, 968475355fSChris Wilson }, *s = shrink; 978475355fSChris Wilson gfp_t gfp = noreclaim; 988475355fSChris Wilson 998475355fSChris Wilson do { 1008475355fSChris Wilson cond_resched(); 1018475355fSChris Wilson page = shmem_read_mapping_page_gfp(mapping, i, gfp); 1028475355fSChris Wilson if (!IS_ERR(page)) 1038475355fSChris Wilson break; 1048475355fSChris Wilson 1058475355fSChris Wilson if (!*s) { 1068475355fSChris Wilson ret = PTR_ERR(page); 1078475355fSChris Wilson goto err_sg; 1088475355fSChris Wilson } 1098475355fSChris Wilson 110cf41a8f1SMaarten Lankhorst i915_gem_shrink(NULL, i915, 2 * page_count, NULL, *s++); 1118475355fSChris Wilson 1128475355fSChris Wilson /* 1138475355fSChris Wilson * We've tried hard to allocate the memory by reaping 1148475355fSChris Wilson * our own buffer, now let the real VM do its job and 1158475355fSChris Wilson * go down in flames if truly OOM. 1168475355fSChris Wilson * 1178475355fSChris Wilson * However, since graphics tend to be disposable, 1188475355fSChris Wilson * defer the oom here by reporting the ENOMEM back 1198475355fSChris Wilson * to userspace. 1208475355fSChris Wilson */ 1218475355fSChris Wilson if (!*s) { 1228475355fSChris Wilson /* reclaim and warn, but no oom */ 1238475355fSChris Wilson gfp = mapping_gfp_mask(mapping); 1248475355fSChris Wilson 1258475355fSChris Wilson /* 1268475355fSChris Wilson * Our bo are always dirty and so we require 1278475355fSChris Wilson * kswapd to reclaim our pages (direct reclaim 1288475355fSChris Wilson * does not effectively begin pageout of our 1298475355fSChris Wilson * buffers on its own). However, direct reclaim 1308475355fSChris Wilson * only waits for kswapd when under allocation 1318475355fSChris Wilson * congestion. So as a result __GFP_RECLAIM is 1328475355fSChris Wilson * unreliable and fails to actually reclaim our 1338475355fSChris Wilson * dirty pages -- unless you try over and over 1348475355fSChris Wilson * again with !__GFP_NORETRY. However, we still 1358475355fSChris Wilson * want to fail this allocation rather than 1368475355fSChris Wilson * trigger the out-of-memory killer and for 1378475355fSChris Wilson * this we want __GFP_RETRY_MAYFAIL. 1388475355fSChris Wilson */ 1398475355fSChris Wilson gfp |= __GFP_RETRY_MAYFAIL; 1408475355fSChris Wilson } 1418475355fSChris Wilson } while (1); 1428475355fSChris Wilson 1438475355fSChris Wilson if (!i || 1448475355fSChris Wilson sg->length >= max_segment || 1458475355fSChris Wilson page_to_pfn(page) != last_pfn + 1) { 146f05b985eSThomas Hellström if (i) 1478475355fSChris Wilson sg = sg_next(sg); 148f05b985eSThomas Hellström 1498475355fSChris Wilson st->nents++; 1508475355fSChris Wilson sg_set_page(sg, page, PAGE_SIZE, 0); 1518475355fSChris Wilson } else { 1528475355fSChris Wilson sg->length += PAGE_SIZE; 1538475355fSChris Wilson } 1548475355fSChris Wilson last_pfn = page_to_pfn(page); 1558475355fSChris Wilson 1568475355fSChris Wilson /* Check that the i965g/gm workaround works. */ 157ea97c4caSChris Wilson GEM_BUG_ON(gfp & __GFP_DMA32 && last_pfn >= 0x00100000UL); 1588475355fSChris Wilson } 159f05b985eSThomas Hellström if (sg) /* loop terminated early; short sg table */ 1608475355fSChris Wilson sg_mark_end(sg); 1618475355fSChris Wilson 1628475355fSChris Wilson /* Trim unused sg entries to avoid wasting memory. */ 1638475355fSChris Wilson i915_sg_trim(st); 1648475355fSChris Wilson 165cad7109aSThomas Hellström return 0; 166f05b985eSThomas Hellström err_sg: 167f05b985eSThomas Hellström sg_mark_end(sg); 168f05b985eSThomas Hellström if (sg != st->sgl) { 169cad7109aSThomas Hellström shmem_sg_free_table(st, mapping, false, false); 170f05b985eSThomas Hellström } else { 171f05b985eSThomas Hellström mapping_clear_unevictable(mapping); 172f05b985eSThomas Hellström sg_free_table(st); 173f05b985eSThomas Hellström } 174f05b985eSThomas Hellström 175f05b985eSThomas Hellström /* 176f05b985eSThomas Hellström * shmemfs first checks if there is enough memory to allocate the page 177f05b985eSThomas Hellström * and reports ENOSPC should there be insufficient, along with the usual 178f05b985eSThomas Hellström * ENOMEM for a genuine allocation failure. 179f05b985eSThomas Hellström * 180f05b985eSThomas Hellström * We use ENOSPC in our driver to mean that we have run out of aperture 181f05b985eSThomas Hellström * space and so want to translate the error from shmemfs back to our 182f05b985eSThomas Hellström * usual understanding of ENOMEM. 183f05b985eSThomas Hellström */ 184f05b985eSThomas Hellström if (ret == -ENOSPC) 185f05b985eSThomas Hellström ret = -ENOMEM; 186f05b985eSThomas Hellström 187cad7109aSThomas Hellström return ret; 188f05b985eSThomas Hellström } 189f05b985eSThomas Hellström 190f05b985eSThomas Hellström static int shmem_get_pages(struct drm_i915_gem_object *obj) 191f05b985eSThomas Hellström { 192f05b985eSThomas Hellström struct drm_i915_private *i915 = to_i915(obj->base.dev); 193f05b985eSThomas Hellström struct intel_memory_region *mem = obj->mm.region; 194f05b985eSThomas Hellström struct address_space *mapping = obj->base.filp->f_mapping; 195f05b985eSThomas Hellström const unsigned long page_count = obj->base.size / PAGE_SIZE; 196f05b985eSThomas Hellström unsigned int max_segment = i915_sg_segment_size(); 197f05b985eSThomas Hellström struct sg_table *st; 198f05b985eSThomas Hellström struct sgt_iter sgt_iter; 199f05b985eSThomas Hellström struct page *page; 200f05b985eSThomas Hellström int ret; 201f05b985eSThomas Hellström 202f05b985eSThomas Hellström /* 203f05b985eSThomas Hellström * Assert that the object is not currently in any GPU domain. As it 204f05b985eSThomas Hellström * wasn't in the GTT, there shouldn't be any way it could have been in 205f05b985eSThomas Hellström * a GPU cache 206f05b985eSThomas Hellström */ 207f05b985eSThomas Hellström GEM_BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS); 208f05b985eSThomas Hellström GEM_BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS); 209f05b985eSThomas Hellström 210f05b985eSThomas Hellström rebuild_st: 211cad7109aSThomas Hellström st = kmalloc(sizeof(*st), GFP_KERNEL); 212cad7109aSThomas Hellström if (!st) 213cad7109aSThomas Hellström return -ENOMEM; 214cad7109aSThomas Hellström 215cad7109aSThomas Hellström ret = shmem_sg_alloc_table(i915, st, obj->base.size, mem, mapping, 216cad7109aSThomas Hellström max_segment); 217cad7109aSThomas Hellström if (ret) 218f05b985eSThomas Hellström goto err_st; 219f05b985eSThomas Hellström 2208475355fSChris Wilson ret = i915_gem_gtt_prepare_pages(obj, st); 2218475355fSChris Wilson if (ret) { 2228475355fSChris Wilson /* 2238475355fSChris Wilson * DMA remapping failed? One possible cause is that 2248475355fSChris Wilson * it could not reserve enough large entries, asking 2258475355fSChris Wilson * for PAGE_SIZE chunks instead may be helpful. 2268475355fSChris Wilson */ 2278475355fSChris Wilson if (max_segment > PAGE_SIZE) { 2288475355fSChris Wilson for_each_sgt_page(page, sgt_iter, st) 2298475355fSChris Wilson put_page(page); 2308475355fSChris Wilson sg_free_table(st); 231f05b985eSThomas Hellström kfree(st); 2328475355fSChris Wilson 2338475355fSChris Wilson max_segment = PAGE_SIZE; 2348475355fSChris Wilson goto rebuild_st; 2358475355fSChris Wilson } else { 2368ff5446aSThomas Zimmermann dev_warn(i915->drm.dev, 2378475355fSChris Wilson "Failed to DMA remap %lu pages\n", 2388475355fSChris Wilson page_count); 2398475355fSChris Wilson goto err_pages; 2408475355fSChris Wilson } 2418475355fSChris Wilson } 2428475355fSChris Wilson 2438475355fSChris Wilson if (i915_gem_object_needs_bit17_swizzle(obj)) 2448475355fSChris Wilson i915_gem_object_do_bit_17_swizzle(obj, st); 2458475355fSChris Wilson 24630f1dccdSMatthew Auld if (i915_gem_object_can_bypass_llc(obj)) 24713d29c82SMatthew Auld obj->cache_dirty = true; 24813d29c82SMatthew Auld 249f05b985eSThomas Hellström __i915_gem_object_set_pages(obj, st, i915_sg_dma_sizes(st->sgl)); 2508475355fSChris Wilson 2518475355fSChris Wilson return 0; 2528475355fSChris Wilson 2538475355fSChris Wilson err_pages: 254cad7109aSThomas Hellström shmem_sg_free_table(st, mapping, false, false); 2558475355fSChris Wilson /* 2568475355fSChris Wilson * shmemfs first checks if there is enough memory to allocate the page 2578475355fSChris Wilson * and reports ENOSPC should there be insufficient, along with the usual 2588475355fSChris Wilson * ENOMEM for a genuine allocation failure. 2598475355fSChris Wilson * 2608475355fSChris Wilson * We use ENOSPC in our driver to mean that we have run out of aperture 2618475355fSChris Wilson * space and so want to translate the error from shmemfs back to our 2628475355fSChris Wilson * usual understanding of ENOMEM. 2638475355fSChris Wilson */ 264f05b985eSThomas Hellström err_st: 2658475355fSChris Wilson if (ret == -ENOSPC) 2668475355fSChris Wilson ret = -ENOMEM; 2678475355fSChris Wilson 268cad7109aSThomas Hellström kfree(st); 269cad7109aSThomas Hellström 2708475355fSChris Wilson return ret; 2718475355fSChris Wilson } 2728475355fSChris Wilson 2737ae03459SMatthew Auld static int 274f033428dSChris Wilson shmem_truncate(struct drm_i915_gem_object *obj) 275f033428dSChris Wilson { 276f033428dSChris Wilson /* 277f033428dSChris Wilson * Our goal here is to return as much of the memory as 278f033428dSChris Wilson * is possible back to the system as we are called from OOM. 279f033428dSChris Wilson * To do this we must instruct the shmfs to drop all of its 280f033428dSChris Wilson * backing pages, *now*. 281f033428dSChris Wilson */ 282f033428dSChris Wilson shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1); 283f033428dSChris Wilson obj->mm.madv = __I915_MADV_PURGED; 284f033428dSChris Wilson obj->mm.pages = ERR_PTR(-EFAULT); 2857ae03459SMatthew Auld 2867ae03459SMatthew Auld return 0; 287f033428dSChris Wilson } 288f033428dSChris Wilson 2897ae03459SMatthew Auld void __shmem_writeback(size_t size, struct address_space *mapping) 290f033428dSChris Wilson { 291f033428dSChris Wilson struct writeback_control wbc = { 292f033428dSChris Wilson .sync_mode = WB_SYNC_NONE, 293f033428dSChris Wilson .nr_to_write = SWAP_CLUSTER_MAX, 294f033428dSChris Wilson .range_start = 0, 295f033428dSChris Wilson .range_end = LLONG_MAX, 296f033428dSChris Wilson .for_reclaim = 1, 297f033428dSChris Wilson }; 298f033428dSChris Wilson unsigned long i; 299f033428dSChris Wilson 300f033428dSChris Wilson /* 301f033428dSChris Wilson * Leave mmapings intact (GTT will have been revoked on unbinding, 302f033428dSChris Wilson * leaving only CPU mmapings around) and add those pages to the LRU 303f033428dSChris Wilson * instead of invoking writeback so they are aged and paged out 304f033428dSChris Wilson * as normal. 305f033428dSChris Wilson */ 306f033428dSChris Wilson 307f033428dSChris Wilson /* Begin writeback on each dirty page */ 308f05b985eSThomas Hellström for (i = 0; i < size >> PAGE_SHIFT; i++) { 309f033428dSChris Wilson struct page *page; 310f033428dSChris Wilson 3119dfc8ff3SMatthew Wilcox (Oracle) page = find_lock_page(mapping, i); 3129dfc8ff3SMatthew Wilcox (Oracle) if (!page) 313f033428dSChris Wilson continue; 314f033428dSChris Wilson 315f033428dSChris Wilson if (!page_mapped(page) && clear_page_dirty_for_io(page)) { 316f033428dSChris Wilson int ret; 317f033428dSChris Wilson 318f033428dSChris Wilson SetPageReclaim(page); 319f033428dSChris Wilson ret = mapping->a_ops->writepage(page, &wbc); 320f033428dSChris Wilson if (!PageWriteback(page)) 321f033428dSChris Wilson ClearPageReclaim(page); 322f033428dSChris Wilson if (!ret) 323f033428dSChris Wilson goto put; 324f033428dSChris Wilson } 325f033428dSChris Wilson unlock_page(page); 326f033428dSChris Wilson put: 327f033428dSChris Wilson put_page(page); 328f033428dSChris Wilson } 329f033428dSChris Wilson } 330f033428dSChris Wilson 331f05b985eSThomas Hellström static void 332f05b985eSThomas Hellström shmem_writeback(struct drm_i915_gem_object *obj) 333f05b985eSThomas Hellström { 334f05b985eSThomas Hellström __shmem_writeback(obj->base.size, obj->base.filp->f_mapping); 335f05b985eSThomas Hellström } 336f05b985eSThomas Hellström 337ffa3fe08SMatthew Auld static int shmem_shrink(struct drm_i915_gem_object *obj, unsigned int flags) 33893544177SMatthew Auld { 33993544177SMatthew Auld switch (obj->mm.madv) { 34093544177SMatthew Auld case I915_MADV_DONTNEED: 34193544177SMatthew Auld return i915_gem_object_truncate(obj); 34293544177SMatthew Auld case __I915_MADV_PURGED: 34393544177SMatthew Auld return 0; 34493544177SMatthew Auld } 34593544177SMatthew Auld 346ffa3fe08SMatthew Auld if (flags & I915_GEM_OBJECT_SHRINK_WRITEBACK) 34793544177SMatthew Auld shmem_writeback(obj); 34893544177SMatthew Auld 34993544177SMatthew Auld return 0; 35093544177SMatthew Auld } 35193544177SMatthew Auld 3528475355fSChris Wilson void 3538475355fSChris Wilson __i915_gem_object_release_shmem(struct drm_i915_gem_object *obj, 3548475355fSChris Wilson struct sg_table *pages, 3558475355fSChris Wilson bool needs_clflush) 3568475355fSChris Wilson { 357d70af579SMatthew Auld struct drm_i915_private *i915 = to_i915(obj->base.dev); 358d70af579SMatthew Auld 3598475355fSChris Wilson GEM_BUG_ON(obj->mm.madv == __I915_MADV_PURGED); 3608475355fSChris Wilson 3618475355fSChris Wilson if (obj->mm.madv == I915_MADV_DONTNEED) 3628475355fSChris Wilson obj->mm.dirty = false; 3638475355fSChris Wilson 3648475355fSChris Wilson if (needs_clflush && 3658475355fSChris Wilson (obj->read_domains & I915_GEM_DOMAIN_CPU) == 0 && 3668475355fSChris Wilson !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ)) 3678475355fSChris Wilson drm_clflush_sg(pages); 3688475355fSChris Wilson 3698475355fSChris Wilson __start_cpu_write(obj); 370d70af579SMatthew Auld /* 371d70af579SMatthew Auld * On non-LLC platforms, force the flush-on-acquire if this is ever 372d70af579SMatthew Auld * swapped-in. Our async flush path is not trust worthy enough yet(and 373d70af579SMatthew Auld * happens in the wrong order), and with some tricks it's conceivable 374d70af579SMatthew Auld * for userspace to change the cache-level to I915_CACHE_NONE after the 375d70af579SMatthew Auld * pages are swapped-in, and since execbuf binds the object before doing 376d70af579SMatthew Auld * the async flush, we have a race window. 377d70af579SMatthew Auld */ 378d70af579SMatthew Auld if (!HAS_LLC(i915)) 379d70af579SMatthew Auld obj->cache_dirty = true; 3808475355fSChris Wilson } 3818475355fSChris Wilson 382a85fffe3SMaarten Lankhorst void i915_gem_object_put_pages_shmem(struct drm_i915_gem_object *obj, struct sg_table *pages) 3838475355fSChris Wilson { 3848475355fSChris Wilson __i915_gem_object_release_shmem(obj, pages, true); 3858475355fSChris Wilson 3868475355fSChris Wilson i915_gem_gtt_finish_pages(obj, pages); 3878475355fSChris Wilson 3888475355fSChris Wilson if (i915_gem_object_needs_bit17_swizzle(obj)) 3898475355fSChris Wilson i915_gem_object_save_bit_17_swizzle(obj, pages); 3908475355fSChris Wilson 391cad7109aSThomas Hellström shmem_sg_free_table(pages, file_inode(obj->base.filp)->i_mapping, 392f05b985eSThomas Hellström obj->mm.dirty, obj->mm.madv == I915_MADV_WILLNEED); 393cad7109aSThomas Hellström kfree(pages); 3948475355fSChris Wilson obj->mm.dirty = false; 3958475355fSChris Wilson } 3968475355fSChris Wilson 397a85fffe3SMaarten Lankhorst static void 398a85fffe3SMaarten Lankhorst shmem_put_pages(struct drm_i915_gem_object *obj, struct sg_table *pages) 399a85fffe3SMaarten Lankhorst { 400a85fffe3SMaarten Lankhorst if (likely(i915_gem_object_has_struct_page(obj))) 401a85fffe3SMaarten Lankhorst i915_gem_object_put_pages_shmem(obj, pages); 402a85fffe3SMaarten Lankhorst else 403a85fffe3SMaarten Lankhorst i915_gem_object_put_pages_phys(obj, pages); 404a85fffe3SMaarten Lankhorst } 405a85fffe3SMaarten Lankhorst 4068475355fSChris Wilson static int 4078475355fSChris Wilson shmem_pwrite(struct drm_i915_gem_object *obj, 4088475355fSChris Wilson const struct drm_i915_gem_pwrite *arg) 4098475355fSChris Wilson { 4108475355fSChris Wilson struct address_space *mapping = obj->base.filp->f_mapping; 411*c5edd542SMatthew Wilcox (Oracle) const struct address_space_operations *aops = mapping->a_ops; 4128475355fSChris Wilson char __user *user_data = u64_to_user_ptr(arg->data_ptr); 4138475355fSChris Wilson u64 remain, offset; 4148475355fSChris Wilson unsigned int pg; 4158475355fSChris Wilson 4168475355fSChris Wilson /* Caller already validated user args */ 4178475355fSChris Wilson GEM_BUG_ON(!access_ok(user_data, arg->size)); 4188475355fSChris Wilson 419a6117097SMaarten Lankhorst if (!i915_gem_object_has_struct_page(obj)) 420a6117097SMaarten Lankhorst return i915_gem_object_pwrite_phys(obj, arg); 421a6117097SMaarten Lankhorst 4228475355fSChris Wilson /* 4238475355fSChris Wilson * Before we instantiate/pin the backing store for our use, we 4248475355fSChris Wilson * can prepopulate the shmemfs filp efficiently using a write into 4258475355fSChris Wilson * the pagecache. We avoid the penalty of instantiating all the 4268475355fSChris Wilson * pages, important if the user is just writing to a few and never 4278475355fSChris Wilson * uses the object on the GPU, and using a direct write into shmemfs 4288475355fSChris Wilson * allows it to avoid the cost of retrieving a page (either swapin 4298475355fSChris Wilson * or clearing-before-use) before it is overwritten. 4308475355fSChris Wilson */ 4318475355fSChris Wilson if (i915_gem_object_has_pages(obj)) 4328475355fSChris Wilson return -ENODEV; 4338475355fSChris Wilson 4348475355fSChris Wilson if (obj->mm.madv != I915_MADV_WILLNEED) 4358475355fSChris Wilson return -EFAULT; 4368475355fSChris Wilson 4378475355fSChris Wilson /* 4388475355fSChris Wilson * Before the pages are instantiated the object is treated as being 4398475355fSChris Wilson * in the CPU domain. The pages will be clflushed as required before 4408475355fSChris Wilson * use, and we can freely write into the pages directly. If userspace 4418475355fSChris Wilson * races pwrite with any other operation; corruption will ensue - 4428475355fSChris Wilson * that is userspace's prerogative! 4438475355fSChris Wilson */ 4448475355fSChris Wilson 4458475355fSChris Wilson remain = arg->size; 4468475355fSChris Wilson offset = arg->offset; 4478475355fSChris Wilson pg = offset_in_page(offset); 4488475355fSChris Wilson 4498475355fSChris Wilson do { 4508475355fSChris Wilson unsigned int len, unwritten; 4518475355fSChris Wilson struct page *page; 4528475355fSChris Wilson void *data, *vaddr; 4538475355fSChris Wilson int err; 4548475355fSChris Wilson char c; 4558475355fSChris Wilson 4568475355fSChris Wilson len = PAGE_SIZE - pg; 4578475355fSChris Wilson if (len > remain) 4588475355fSChris Wilson len = remain; 4598475355fSChris Wilson 4608475355fSChris Wilson /* Prefault the user page to reduce potential recursion */ 4618475355fSChris Wilson err = __get_user(c, user_data); 4628475355fSChris Wilson if (err) 4638475355fSChris Wilson return err; 4648475355fSChris Wilson 4658475355fSChris Wilson err = __get_user(c, user_data + len - 1); 4668475355fSChris Wilson if (err) 4678475355fSChris Wilson return err; 4688475355fSChris Wilson 469*c5edd542SMatthew Wilcox (Oracle) err = aops->write_begin(obj->base.filp, mapping, offset, len, 4708475355fSChris Wilson &page, &data); 4718475355fSChris Wilson if (err < 0) 4728475355fSChris Wilson return err; 4738475355fSChris Wilson 4748475355fSChris Wilson vaddr = kmap_atomic(page); 4758475355fSChris Wilson unwritten = __copy_from_user_inatomic(vaddr + pg, 4768475355fSChris Wilson user_data, 4778475355fSChris Wilson len); 4788475355fSChris Wilson kunmap_atomic(vaddr); 4798475355fSChris Wilson 480*c5edd542SMatthew Wilcox (Oracle) err = aops->write_end(obj->base.filp, mapping, offset, len, 481*c5edd542SMatthew Wilcox (Oracle) len - unwritten, page, data); 4828475355fSChris Wilson if (err < 0) 4838475355fSChris Wilson return err; 4848475355fSChris Wilson 4858475355fSChris Wilson /* We don't handle -EFAULT, leave it to the caller to check */ 4868475355fSChris Wilson if (unwritten) 4878475355fSChris Wilson return -ENODEV; 4888475355fSChris Wilson 4898475355fSChris Wilson remain -= len; 4908475355fSChris Wilson user_data += len; 4918475355fSChris Wilson offset += len; 4928475355fSChris Wilson pg = 0; 4938475355fSChris Wilson } while (remain); 4948475355fSChris Wilson 4958475355fSChris Wilson return 0; 4968475355fSChris Wilson } 4978475355fSChris Wilson 498a6117097SMaarten Lankhorst static int 499a6117097SMaarten Lankhorst shmem_pread(struct drm_i915_gem_object *obj, 500a6117097SMaarten Lankhorst const struct drm_i915_gem_pread *arg) 501a6117097SMaarten Lankhorst { 502a6117097SMaarten Lankhorst if (!i915_gem_object_has_struct_page(obj)) 503a6117097SMaarten Lankhorst return i915_gem_object_pread_phys(obj, arg); 504a6117097SMaarten Lankhorst 505a6117097SMaarten Lankhorst return -ENODEV; 506a6117097SMaarten Lankhorst } 507a6117097SMaarten Lankhorst 5080c159ffeSChris Wilson static void shmem_release(struct drm_i915_gem_object *obj) 5090c159ffeSChris Wilson { 5100ff37575SThomas Hellström if (i915_gem_object_has_struct_page(obj)) 511da1184cdSMatthew Auld i915_gem_object_release_memory_region(obj); 512da1184cdSMatthew Auld 5130c159ffeSChris Wilson fput(obj->base.filp); 5140c159ffeSChris Wilson } 5150c159ffeSChris Wilson 5168475355fSChris Wilson const struct drm_i915_gem_object_ops i915_gem_shmem_ops = { 5177d192daaSChris Wilson .name = "i915_gem_object_shmem", 518c471748dSMaarten Lankhorst .flags = I915_GEM_OBJECT_IS_SHRINKABLE, 5198475355fSChris Wilson 5208475355fSChris Wilson .get_pages = shmem_get_pages, 5218475355fSChris Wilson .put_pages = shmem_put_pages, 522f033428dSChris Wilson .truncate = shmem_truncate, 523ffa3fe08SMatthew Auld .shrink = shmem_shrink, 5248475355fSChris Wilson 5258475355fSChris Wilson .pwrite = shmem_pwrite, 526a6117097SMaarten Lankhorst .pread = shmem_pread, 5270c159ffeSChris Wilson 5280c159ffeSChris Wilson .release = shmem_release, 5298475355fSChris Wilson }; 5308475355fSChris Wilson 531da1184cdSMatthew Auld static int __create_shmem(struct drm_i915_private *i915, 5328475355fSChris Wilson struct drm_gem_object *obj, 533da1184cdSMatthew Auld resource_size_t size) 5348475355fSChris Wilson { 5358475355fSChris Wilson unsigned long flags = VM_NORESERVE; 5368475355fSChris Wilson struct file *filp; 5378475355fSChris Wilson 5388475355fSChris Wilson drm_gem_private_object_init(&i915->drm, obj, size); 5398475355fSChris Wilson 5408475355fSChris Wilson if (i915->mm.gemfs) 5418475355fSChris Wilson filp = shmem_file_setup_with_mnt(i915->mm.gemfs, "i915", size, 5428475355fSChris Wilson flags); 5438475355fSChris Wilson else 5448475355fSChris Wilson filp = shmem_file_setup("i915", size, flags); 5458475355fSChris Wilson if (IS_ERR(filp)) 5468475355fSChris Wilson return PTR_ERR(filp); 5478475355fSChris Wilson 5488475355fSChris Wilson obj->filp = filp; 5498475355fSChris Wilson return 0; 5508475355fSChris Wilson } 5518475355fSChris Wilson 55297d55396SMatthew Auld static int shmem_object_init(struct intel_memory_region *mem, 55397d55396SMatthew Auld struct drm_i915_gem_object *obj, 554da1184cdSMatthew Auld resource_size_t size, 555d22632c8SMatthew Auld resource_size_t page_size, 556da1184cdSMatthew Auld unsigned int flags) 5578475355fSChris Wilson { 5587867d709SChris Wilson static struct lock_class_key lock_class; 559da1184cdSMatthew Auld struct drm_i915_private *i915 = mem->i915; 5608475355fSChris Wilson struct address_space *mapping; 5618475355fSChris Wilson unsigned int cache_level; 5628475355fSChris Wilson gfp_t mask; 5638475355fSChris Wilson int ret; 5648475355fSChris Wilson 565da1184cdSMatthew Auld ret = __create_shmem(i915, &obj->base, size); 5668475355fSChris Wilson if (ret) 56797d55396SMatthew Auld return ret; 5688475355fSChris Wilson 5698475355fSChris Wilson mask = GFP_HIGHUSER | __GFP_RECLAIMABLE; 5708475355fSChris Wilson if (IS_I965GM(i915) || IS_I965G(i915)) { 5718475355fSChris Wilson /* 965gm cannot relocate objects above 4GiB. */ 5728475355fSChris Wilson mask &= ~__GFP_HIGHMEM; 5738475355fSChris Wilson mask |= __GFP_DMA32; 5748475355fSChris Wilson } 5758475355fSChris Wilson 5768475355fSChris Wilson mapping = obj->base.filp->f_mapping; 5778475355fSChris Wilson mapping_set_gfp_mask(mapping, mask); 5788475355fSChris Wilson GEM_BUG_ON(!(mapping_gfp_mask(mapping) & __GFP_RECLAIM)); 5798475355fSChris Wilson 5800ff37575SThomas Hellström i915_gem_object_init(obj, &i915_gem_shmem_ops, &lock_class, 0); 5810ff37575SThomas Hellström obj->mem_flags |= I915_BO_FLAG_STRUCT_PAGE; 5828475355fSChris Wilson obj->write_domain = I915_GEM_DOMAIN_CPU; 5838475355fSChris Wilson obj->read_domains = I915_GEM_DOMAIN_CPU; 5848475355fSChris Wilson 5858475355fSChris Wilson if (HAS_LLC(i915)) 5868475355fSChris Wilson /* On some devices, we can have the GPU use the LLC (the CPU 5878475355fSChris Wilson * cache) for about a 10% performance improvement 5888475355fSChris Wilson * compared to uncached. Graphics requests other than 5898475355fSChris Wilson * display scanout are coherent with the CPU in 5908475355fSChris Wilson * accessing this cache. This means in this mode we 5918475355fSChris Wilson * don't need to clflush on the CPU side, and on the 5928475355fSChris Wilson * GPU side we only need to flush internal caches to 5938475355fSChris Wilson * get data visible to the CPU. 5948475355fSChris Wilson * 5958475355fSChris Wilson * However, we maintain the display planes as UC, and so 5968475355fSChris Wilson * need to rebind when first used as such. 5978475355fSChris Wilson */ 5988475355fSChris Wilson cache_level = I915_CACHE_LLC; 5998475355fSChris Wilson else 6008475355fSChris Wilson cache_level = I915_CACHE_NONE; 6018475355fSChris Wilson 6028475355fSChris Wilson i915_gem_object_set_cache_coherency(obj, cache_level); 6038475355fSChris Wilson 604c471748dSMaarten Lankhorst i915_gem_object_init_memory_region(obj, mem); 6058475355fSChris Wilson 60697d55396SMatthew Auld return 0; 6078475355fSChris Wilson } 6088475355fSChris Wilson 609da1184cdSMatthew Auld struct drm_i915_gem_object * 610da1184cdSMatthew Auld i915_gem_object_create_shmem(struct drm_i915_private *i915, 611da1184cdSMatthew Auld resource_size_t size) 612da1184cdSMatthew Auld { 613da1184cdSMatthew Auld return i915_gem_object_create_region(i915->mm.regions[INTEL_REGION_SMEM], 614d22632c8SMatthew Auld size, 0, 0); 615da1184cdSMatthew Auld } 616da1184cdSMatthew Auld 6178475355fSChris Wilson /* Allocate a new GEM object and fill it with the supplied data */ 6188475355fSChris Wilson struct drm_i915_gem_object * 6198475355fSChris Wilson i915_gem_object_create_shmem_from_data(struct drm_i915_private *dev_priv, 620da1184cdSMatthew Auld const void *data, resource_size_t size) 6218475355fSChris Wilson { 6228475355fSChris Wilson struct drm_i915_gem_object *obj; 6238475355fSChris Wilson struct file *file; 624*c5edd542SMatthew Wilcox (Oracle) const struct address_space_operations *aops; 625da1184cdSMatthew Auld resource_size_t offset; 6268475355fSChris Wilson int err; 6278475355fSChris Wilson 62832b7cf51SThomas Hellström GEM_WARN_ON(IS_DGFX(dev_priv)); 6298475355fSChris Wilson obj = i915_gem_object_create_shmem(dev_priv, round_up(size, PAGE_SIZE)); 6308475355fSChris Wilson if (IS_ERR(obj)) 6318475355fSChris Wilson return obj; 6328475355fSChris Wilson 6338475355fSChris Wilson GEM_BUG_ON(obj->write_domain != I915_GEM_DOMAIN_CPU); 6348475355fSChris Wilson 6358475355fSChris Wilson file = obj->base.filp; 636*c5edd542SMatthew Wilcox (Oracle) aops = file->f_mapping->a_ops; 6378475355fSChris Wilson offset = 0; 6388475355fSChris Wilson do { 6398475355fSChris Wilson unsigned int len = min_t(typeof(size), size, PAGE_SIZE); 6408475355fSChris Wilson struct page *page; 6418475355fSChris Wilson void *pgdata, *vaddr; 6428475355fSChris Wilson 643*c5edd542SMatthew Wilcox (Oracle) err = aops->write_begin(file, file->f_mapping, offset, len, 6448475355fSChris Wilson &page, &pgdata); 6458475355fSChris Wilson if (err < 0) 6468475355fSChris Wilson goto fail; 6478475355fSChris Wilson 6488475355fSChris Wilson vaddr = kmap(page); 6498475355fSChris Wilson memcpy(vaddr, data, len); 6508475355fSChris Wilson kunmap(page); 6518475355fSChris Wilson 652*c5edd542SMatthew Wilcox (Oracle) err = aops->write_end(file, file->f_mapping, offset, len, len, 6538475355fSChris Wilson page, pgdata); 6548475355fSChris Wilson if (err < 0) 6558475355fSChris Wilson goto fail; 6568475355fSChris Wilson 6578475355fSChris Wilson size -= len; 6588475355fSChris Wilson data += len; 6598475355fSChris Wilson offset += len; 6608475355fSChris Wilson } while (size); 6618475355fSChris Wilson 6628475355fSChris Wilson return obj; 6638475355fSChris Wilson 6648475355fSChris Wilson fail: 6658475355fSChris Wilson i915_gem_object_put(obj); 6668475355fSChris Wilson return ERR_PTR(err); 6678475355fSChris Wilson } 668da1184cdSMatthew Auld 669da1184cdSMatthew Auld static int init_shmem(struct intel_memory_region *mem) 670da1184cdSMatthew Auld { 671da1184cdSMatthew Auld int err; 672da1184cdSMatthew Auld 673da1184cdSMatthew Auld err = i915_gemfs_init(mem->i915); 674da1184cdSMatthew Auld if (err) { 675da1184cdSMatthew Auld DRM_NOTE("Unable to create a private tmpfs mount, hugepage support will be disabled(%d).\n", 676da1184cdSMatthew Auld err); 677da1184cdSMatthew Auld } 678da1184cdSMatthew Auld 67938f1cb68SLukasz Fiedorowicz intel_memory_region_set_name(mem, "system"); 68038f1cb68SLukasz Fiedorowicz 681da1184cdSMatthew Auld return 0; /* Don't error, we can simply fallback to the kernel mnt */ 682da1184cdSMatthew Auld } 683da1184cdSMatthew Auld 6848b1f7f92SThomas Hellström static int release_shmem(struct intel_memory_region *mem) 685da1184cdSMatthew Auld { 686da1184cdSMatthew Auld i915_gemfs_fini(mem->i915); 6878b1f7f92SThomas Hellström return 0; 688da1184cdSMatthew Auld } 689da1184cdSMatthew Auld 690da1184cdSMatthew Auld static const struct intel_memory_region_ops shmem_region_ops = { 691da1184cdSMatthew Auld .init = init_shmem, 692da1184cdSMatthew Auld .release = release_shmem, 69397d55396SMatthew Auld .init_object = shmem_object_init, 694da1184cdSMatthew Auld }; 695da1184cdSMatthew Auld 696d1487389SThomas Hellström struct intel_memory_region *i915_gem_shmem_setup(struct drm_i915_private *i915, 697d1487389SThomas Hellström u16 type, u16 instance) 698da1184cdSMatthew Auld { 699da1184cdSMatthew Auld return intel_memory_region_create(i915, 0, 700da1184cdSMatthew Auld totalram_pages() << PAGE_SHIFT, 701235582caSMatthew Auld PAGE_SIZE, 0, 0, 702d1487389SThomas Hellström type, instance, 703da1184cdSMatthew Auld &shmem_region_ops); 704da1184cdSMatthew Auld } 70541a9c75dSChris Wilson 70641a9c75dSChris Wilson bool i915_gem_object_is_shmem(const struct drm_i915_gem_object *obj) 70741a9c75dSChris Wilson { 70841a9c75dSChris Wilson return obj->ops == &i915_gem_shmem_ops; 70941a9c75dSChris Wilson } 710