| File: | dev/pci/drm/i915/gem/i915_gem_shmem.c |
| Warning: | line 670, column 3 Value stored to 'mask' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | /* |
| 2 | * SPDX-License-Identifier: MIT |
| 3 | * |
| 4 | * Copyright © 2014-2016 Intel Corporation |
| 5 | */ |
| 6 | |
| 7 | #include <linux/pagevec.h> |
| 8 | #include <linux/shmem_fs.h> |
| 9 | #include <linux/swap.h> |
| 10 | |
| 11 | #include <drm/drm_cache.h> |
| 12 | |
| 13 | #include "gem/i915_gem_region.h" |
| 14 | #include "i915_drv.h" |
| 15 | #include "i915_gem_object.h" |
| 16 | #include "i915_gem_tiling.h" |
| 17 | #include "i915_gemfs.h" |
| 18 | #include "i915_scatterlist.h" |
| 19 | #include "i915_trace.h" |
| 20 | |
| 21 | /* |
| 22 | * Move pages to appropriate lru and release the pagevec, decrementing the |
| 23 | * ref count of those pages. |
| 24 | */ |
| 25 | static void check_release_pagevec(struct pagevec *pvec) |
| 26 | { |
| 27 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
| 28 | #ifdef notyet |
| 29 | check_move_unevictable_pages(pvec); |
| 30 | #endif |
| 31 | __pagevec_release(pvec); |
| 32 | cond_resched()do { if (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_schedstate.spc_schedflags & 0x0002) yield (); } while (0); |
| 33 | } |
| 34 | |
| 35 | void shmem_sg_free_table(struct sg_table *st, struct address_space *mapping, |
| 36 | bool_Bool dirty, bool_Bool backup, |
| 37 | struct drm_i915_gem_object *obj) |
| 38 | { |
| 39 | struct sgt_iter sgt_iter; |
| 40 | struct pagevec pvec; |
| 41 | struct vm_page *page; |
| 42 | |
| 43 | #ifdef __linux__ |
| 44 | mapping_clear_unevictable(mapping); |
| 45 | #endif |
| 46 | |
| 47 | pagevec_init(&pvec); |
| 48 | for_each_sgt_page(page, sgt_iter, st)for ((sgt_iter) = __sgt_iter((st)->sgl, 0); ((page) = (sgt_iter ).pfn == 0 ? ((void *)0) : (PHYS_TO_VM_PAGE(((paddr_t)((sgt_iter ).pfn + ((sgt_iter).curr >> 12)) << 12)))); (((sgt_iter ).curr += (1 << 12)) >= (sgt_iter).max) ? (sgt_iter) = __sgt_iter(__sg_next((sgt_iter).sgp), 0), 0 : 0) { |
| 49 | if (dirty) |
| 50 | set_page_dirty(page)x86_atomic_clearbits_u32(&page->pg_flags, 0x00000008); |
| 51 | |
| 52 | #ifdef __linux__ |
| 53 | if (backup) |
| 54 | mark_page_accessed(page); |
| 55 | |
| 56 | if (!pagevec_add(&pvec, page)) |
| 57 | check_release_pagevec(&pvec); |
| 58 | #endif |
| 59 | } |
| 60 | #ifdef __linux__ |
| 61 | if (pagevec_count(&pvec)) |
| 62 | check_release_pagevec(&pvec); |
| 63 | #else |
| 64 | uvm_obj_unwire(obj->base.uao, 0, obj->base.size); |
| 65 | #endif |
| 66 | |
| 67 | sg_free_table(st); |
| 68 | } |
| 69 | |
| 70 | int shmem_sg_alloc_table(struct drm_i915_privateinteldrm_softc *i915, struct sg_table *st, |
| 71 | size_t size, struct intel_memory_region *mr, |
| 72 | struct address_space *mapping, |
| 73 | unsigned int max_segment, |
| 74 | struct drm_i915_gem_object *obj) |
| 75 | { |
| 76 | const unsigned long page_count = size / PAGE_SIZE(1 << 12); |
| 77 | unsigned long i; |
| 78 | struct scatterlist *sg; |
| 79 | struct vm_page *page; |
| 80 | unsigned long last_pfn = 0; /* suppress gcc warning */ |
| 81 | gfp_t noreclaim; |
| 82 | int ret; |
| 83 | struct pglist plist; |
| 84 | |
| 85 | /* |
| 86 | * If there's no chance of allocating enough pages for the whole |
| 87 | * object, bail early. |
| 88 | */ |
| 89 | if (size > resource_size(&mr->region)) |
| 90 | return -ENOMEM12; |
| 91 | |
| 92 | if (sg_alloc_table(st, page_count, GFP_KERNEL(0x0001 | 0x0004) | __GFP_NOWARN0)) |
| 93 | return -ENOMEM12; |
| 94 | #ifdef __linux__ |
| 95 | |
| 96 | /* |
| 97 | * Get the list of pages out of our struct file. They'll be pinned |
| 98 | * at this point until we release them. |
| 99 | * |
| 100 | * Fail silently without starting the shrinker |
| 101 | */ |
| 102 | mapping_set_unevictable(mapping); |
| 103 | noreclaim = mapping_gfp_constraint(mapping, ~__GFP_RECLAIM); |
| 104 | noreclaim |= __GFP_NORETRY0 | __GFP_NOWARN0; |
| 105 | |
| 106 | sg = st->sgl; |
| 107 | st->nents = 0; |
| 108 | for (i = 0; i < page_count; i++) { |
| 109 | const unsigned int shrink[] = { |
| 110 | I915_SHRINK_BOUND(1UL << (1)) | I915_SHRINK_UNBOUND(1UL << (0)), |
| 111 | 0, |
| 112 | }, *s = shrink; |
| 113 | gfp_t gfp = noreclaim; |
| 114 | |
| 115 | do { |
| 116 | cond_resched()do { if (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_schedstate.spc_schedflags & 0x0002) yield (); } while (0); |
| 117 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); |
| 118 | if (!IS_ERR(page)) |
| 119 | break; |
| 120 | |
| 121 | if (!*s) { |
| 122 | ret = PTR_ERR(page); |
| 123 | goto err_sg; |
| 124 | } |
| 125 | |
| 126 | i915_gem_shrink(NULL((void *)0), i915, 2 * page_count, NULL((void *)0), *s++); |
| 127 | |
| 128 | /* |
| 129 | * We've tried hard to allocate the memory by reaping |
| 130 | * our own buffer, now let the real VM do its job and |
| 131 | * go down in flames if truly OOM. |
| 132 | * |
| 133 | * However, since graphics tend to be disposable, |
| 134 | * defer the oom here by reporting the ENOMEM back |
| 135 | * to userspace. |
| 136 | */ |
| 137 | if (!*s) { |
| 138 | /* reclaim and warn, but no oom */ |
| 139 | gfp = mapping_gfp_mask(mapping); |
| 140 | |
| 141 | /* |
| 142 | * Our bo are always dirty and so we require |
| 143 | * kswapd to reclaim our pages (direct reclaim |
| 144 | * does not effectively begin pageout of our |
| 145 | * buffers on its own). However, direct reclaim |
| 146 | * only waits for kswapd when under allocation |
| 147 | * congestion. So as a result __GFP_RECLAIM is |
| 148 | * unreliable and fails to actually reclaim our |
| 149 | * dirty pages -- unless you try over and over |
| 150 | * again with !__GFP_NORETRY. However, we still |
| 151 | * want to fail this allocation rather than |
| 152 | * trigger the out-of-memory killer and for |
| 153 | * this we want __GFP_RETRY_MAYFAIL. |
| 154 | */ |
| 155 | gfp |= __GFP_RETRY_MAYFAIL0 | __GFP_NOWARN0; |
| 156 | } |
| 157 | } while (1); |
| 158 | |
| 159 | if (!i || |
| 160 | sg->length >= max_segment || |
| 161 | page_to_pfn(page)(((page)->phys_addr) / (1 << 12)) != last_pfn + 1) { |
| 162 | if (i) |
| 163 | sg = sg_next(sg); |
| 164 | |
| 165 | st->nents++; |
| 166 | sg_set_page(sg, page, PAGE_SIZE(1 << 12), 0); |
| 167 | } else { |
| 168 | sg->length += PAGE_SIZE(1 << 12); |
| 169 | } |
| 170 | last_pfn = page_to_pfn(page)(((page)->phys_addr) / (1 << 12)); |
| 171 | |
| 172 | /* Check that the i965g/gm workaround works. */ |
| 173 | GEM_BUG_ON(gfp & __GFP_DMA32 && last_pfn >= 0x00100000UL)((void)0); |
| 174 | } |
| 175 | #else |
| 176 | sg = st->sgl; |
| 177 | st->nents = 0; |
| 178 | |
| 179 | TAILQ_INIT(&plist)do { (&plist)->tqh_first = ((void *)0); (&plist)-> tqh_last = &(&plist)->tqh_first; } while (0); |
| 180 | if (uvm_obj_wire(obj->base.uao, 0, obj->base.size, &plist)) { |
| 181 | sg_free_table(st); |
| 182 | kfree(st); |
| 183 | return -ENOMEM12; |
| 184 | } |
| 185 | |
| 186 | i = 0; |
| 187 | TAILQ_FOREACH(page, &plist, pageq)for((page) = ((&plist)->tqh_first); (page) != ((void * )0); (page) = ((page)->pageq.tqe_next)) { |
| 188 | if (i) |
| 189 | sg = sg_next(sg); |
| 190 | st->nents++; |
| 191 | sg_set_page(sg, page, PAGE_SIZE(1 << 12), 0); |
| 192 | i++; |
| 193 | } |
| 194 | #endif |
| 195 | if (sg) /* loop terminated early; short sg table */ |
| 196 | sg_mark_end(sg); |
| 197 | |
| 198 | /* Trim unused sg entries to avoid wasting memory. */ |
| 199 | i915_sg_trim(st); |
| 200 | |
| 201 | return 0; |
| 202 | #ifdef notyet |
| 203 | err_sg: |
| 204 | sg_mark_end(sg); |
| 205 | if (sg != st->sgl) { |
| 206 | shmem_sg_free_table(st, mapping, false0, false0); |
| 207 | } else { |
| 208 | mapping_clear_unevictable(mapping); |
| 209 | sg_free_table(st); |
| 210 | } |
| 211 | |
| 212 | /* |
| 213 | * shmemfs first checks if there is enough memory to allocate the page |
| 214 | * and reports ENOSPC should there be insufficient, along with the usual |
| 215 | * ENOMEM for a genuine allocation failure. |
| 216 | * |
| 217 | * We use ENOSPC in our driver to mean that we have run out of aperture |
| 218 | * space and so want to translate the error from shmemfs back to our |
| 219 | * usual understanding of ENOMEM. |
| 220 | */ |
| 221 | if (ret == -ENOSPC28) |
| 222 | ret = -ENOMEM12; |
| 223 | |
| 224 | return ret; |
| 225 | #endif |
| 226 | } |
| 227 | |
| 228 | static int shmem_get_pages(struct drm_i915_gem_object *obj) |
| 229 | { |
| 230 | struct drm_i915_privateinteldrm_softc *i915 = to_i915(obj->base.dev); |
| 231 | struct intel_memory_region *mem = obj->mm.region; |
| 232 | #ifdef __linux__ |
| 233 | struct address_space *mapping = obj->base.filp->f_mapping; |
| 234 | #endif |
| 235 | const unsigned long page_count = obj->base.size / PAGE_SIZE(1 << 12); |
| 236 | unsigned int max_segment = i915_sg_segment_size(i915->drm.dev); |
| 237 | struct sg_table *st; |
| 238 | struct sgt_iter sgt_iter; |
| 239 | struct vm_page *page; |
| 240 | int ret; |
| 241 | |
| 242 | /* |
| 243 | * Assert that the object is not currently in any GPU domain. As it |
| 244 | * wasn't in the GTT, there shouldn't be any way it could have been in |
| 245 | * a GPU cache |
| 246 | */ |
| 247 | GEM_BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS)((void)0); |
| 248 | GEM_BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS)((void)0); |
| 249 | |
| 250 | rebuild_st: |
| 251 | st = kmalloc(sizeof(*st), GFP_KERNEL(0x0001 | 0x0004) | __GFP_NOWARN0); |
| 252 | if (!st) |
| 253 | return -ENOMEM12; |
| 254 | |
| 255 | #ifdef __linux__ |
| 256 | ret = shmem_sg_alloc_table(i915, st, obj->base.size, mem, mapping, |
| 257 | max_segment); |
| 258 | #else |
| 259 | ret = shmem_sg_alloc_table(i915, st, obj->base.size, mem, NULL((void *)0), |
| 260 | max_segment, obj); |
| 261 | #endif |
| 262 | if (ret) |
| 263 | goto err_st; |
| 264 | |
| 265 | ret = i915_gem_gtt_prepare_pages(obj, st); |
| 266 | if (ret) { |
| 267 | /* |
| 268 | * DMA remapping failed? One possible cause is that |
| 269 | * it could not reserve enough large entries, asking |
| 270 | * for PAGE_SIZE chunks instead may be helpful. |
| 271 | */ |
| 272 | if (max_segment > PAGE_SIZE(1 << 12)) { |
| 273 | #ifdef __linux__ |
| 274 | for_each_sgt_page(page, sgt_iter, st)for ((sgt_iter) = __sgt_iter((st)->sgl, 0); ((page) = (sgt_iter ).pfn == 0 ? ((void *)0) : (PHYS_TO_VM_PAGE(((paddr_t)((sgt_iter ).pfn + ((sgt_iter).curr >> 12)) << 12)))); (((sgt_iter ).curr += (1 << 12)) >= (sgt_iter).max) ? (sgt_iter) = __sgt_iter(__sg_next((sgt_iter).sgp), 0), 0 : 0) |
| 275 | put_page(page); |
| 276 | #else |
| 277 | uvm_obj_unwire(obj->base.uao, 0, obj->base.size); |
| 278 | #endif |
| 279 | sg_free_table(st); |
| 280 | kfree(st); |
| 281 | |
| 282 | max_segment = PAGE_SIZE(1 << 12); |
| 283 | goto rebuild_st; |
| 284 | } else { |
| 285 | dev_warn(i915->drm.dev,printf("drm:pid%d:%s *WARNING* " "Failed to DMA remap %lu pages\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , page_count ) |
| 286 | "Failed to DMA remap %lu pages\n",printf("drm:pid%d:%s *WARNING* " "Failed to DMA remap %lu pages\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , page_count ) |
| 287 | page_count)printf("drm:pid%d:%s *WARNING* " "Failed to DMA remap %lu pages\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , page_count ); |
| 288 | goto err_pages; |
| 289 | } |
| 290 | } |
| 291 | |
| 292 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
| 293 | i915_gem_object_do_bit_17_swizzle(obj, st); |
| 294 | |
| 295 | if (i915_gem_object_can_bypass_llc(obj)) |
| 296 | obj->cache_dirty = true1; |
| 297 | |
| 298 | __i915_gem_object_set_pages(obj, st, i915_sg_dma_sizes(st->sgl)); |
| 299 | |
| 300 | return 0; |
| 301 | |
| 302 | err_pages: |
| 303 | #ifdef __linux__ |
| 304 | shmem_sg_free_table(st, mapping, false0, false0); |
| 305 | #else |
| 306 | shmem_sg_free_table(st, NULL((void *)0), false0, false0, obj); |
| 307 | #endif |
| 308 | /* |
| 309 | * shmemfs first checks if there is enough memory to allocate the page |
| 310 | * and reports ENOSPC should there be insufficient, along with the usual |
| 311 | * ENOMEM for a genuine allocation failure. |
| 312 | * |
| 313 | * We use ENOSPC in our driver to mean that we have run out of aperture |
| 314 | * space and so want to translate the error from shmemfs back to our |
| 315 | * usual understanding of ENOMEM. |
| 316 | */ |
| 317 | err_st: |
| 318 | if (ret == -ENOSPC28) |
| 319 | ret = -ENOMEM12; |
| 320 | |
| 321 | kfree(st); |
| 322 | |
| 323 | return ret; |
| 324 | } |
| 325 | |
| 326 | static int |
| 327 | shmem_truncate(struct drm_i915_gem_object *obj) |
| 328 | { |
| 329 | /* |
| 330 | * Our goal here is to return as much of the memory as |
| 331 | * is possible back to the system as we are called from OOM. |
| 332 | * To do this we must instruct the shmfs to drop all of its |
| 333 | * backing pages, *now*. |
| 334 | */ |
| 335 | #ifdef __linux__ |
| 336 | shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1); |
| 337 | #else |
| 338 | rw_enter(obj->base.uao->vmobjlock, RW_WRITE0x0001UL); |
| 339 | obj->base.uao->pgops->pgo_flush(obj->base.uao, 0, obj->base.size, |
| 340 | PGO_ALLPAGES0x010 | PGO_FREE0x008); |
| 341 | rw_exit(obj->base.uao->vmobjlock); |
| 342 | #endif |
| 343 | obj->mm.madv = __I915_MADV_PURGED2; |
| 344 | obj->mm.pages = ERR_PTR(-EFAULT14); |
| 345 | |
| 346 | return 0; |
| 347 | } |
| 348 | |
| 349 | void __shmem_writeback(size_t size, struct address_space *mapping) |
| 350 | { |
| 351 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
| 352 | #ifdef notyet |
| 353 | struct writeback_control wbc = { |
| 354 | .sync_mode = WB_SYNC_NONE, |
| 355 | .nr_to_write = SWAP_CLUSTER_MAX, |
| 356 | .range_start = 0, |
| 357 | .range_end = LLONG_MAX0x7fffffffffffffffLL, |
| 358 | .for_reclaim = 1, |
| 359 | }; |
| 360 | unsigned long i; |
| 361 | |
| 362 | /* |
| 363 | * Leave mmapings intact (GTT will have been revoked on unbinding, |
| 364 | * leaving only CPU mmapings around) and add those pages to the LRU |
| 365 | * instead of invoking writeback so they are aged and paged out |
| 366 | * as normal. |
| 367 | */ |
| 368 | |
| 369 | /* Begin writeback on each dirty page */ |
| 370 | for (i = 0; i < size >> PAGE_SHIFT12; i++) { |
| 371 | struct vm_page *page; |
| 372 | |
| 373 | page = find_lock_page(mapping, i); |
| 374 | if (!page) |
| 375 | continue; |
| 376 | |
| 377 | if (!page_mapped(page) && clear_page_dirty_for_io(page)) { |
| 378 | int ret; |
| 379 | |
| 380 | SetPageReclaim(page); |
| 381 | ret = mapping->a_ops->writepage(page, &wbc); |
| 382 | if (!PageWriteback(page)) |
| 383 | ClearPageReclaim(page); |
| 384 | if (!ret) |
| 385 | goto put; |
| 386 | } |
| 387 | unlock_page(page); |
| 388 | put: |
| 389 | put_page(page); |
| 390 | } |
| 391 | #endif |
| 392 | } |
| 393 | |
| 394 | static void |
| 395 | shmem_writeback(struct drm_i915_gem_object *obj) |
| 396 | { |
| 397 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
| 398 | #ifdef notyet |
| 399 | __shmem_writeback(obj->base.size, obj->base.filp->f_mapping); |
| 400 | #endif |
| 401 | } |
| 402 | |
| 403 | static int shmem_shrink(struct drm_i915_gem_object *obj, unsigned int flags) |
| 404 | { |
| 405 | switch (obj->mm.madv) { |
| 406 | case I915_MADV_DONTNEED1: |
| 407 | return i915_gem_object_truncate(obj); |
| 408 | case __I915_MADV_PURGED2: |
| 409 | return 0; |
| 410 | } |
| 411 | |
| 412 | if (flags & I915_GEM_OBJECT_SHRINK_WRITEBACK(1UL << (0))) |
| 413 | shmem_writeback(obj); |
| 414 | |
| 415 | return 0; |
| 416 | } |
| 417 | |
| 418 | void |
| 419 | __i915_gem_object_release_shmem(struct drm_i915_gem_object *obj, |
| 420 | struct sg_table *pages, |
| 421 | bool_Bool needs_clflush) |
| 422 | { |
| 423 | struct drm_i915_privateinteldrm_softc *i915 = to_i915(obj->base.dev); |
| 424 | |
| 425 | GEM_BUG_ON(obj->mm.madv == __I915_MADV_PURGED)((void)0); |
| 426 | |
| 427 | if (obj->mm.madv == I915_MADV_DONTNEED1) |
| 428 | obj->mm.dirty = false0; |
| 429 | |
| 430 | if (needs_clflush && |
| 431 | (obj->read_domains & I915_GEM_DOMAIN_CPU0x00000001) == 0 && |
| 432 | !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ(1UL << (0)))) |
| 433 | drm_clflush_sg(pages); |
| 434 | |
| 435 | __start_cpu_write(obj); |
| 436 | /* |
| 437 | * On non-LLC igfx platforms, force the flush-on-acquire if this is ever |
| 438 | * swapped-in. Our async flush path is not trust worthy enough yet(and |
| 439 | * happens in the wrong order), and with some tricks it's conceivable |
| 440 | * for userspace to change the cache-level to I915_CACHE_NONE after the |
| 441 | * pages are swapped-in, and since execbuf binds the object before doing |
| 442 | * the async flush, we have a race window. |
| 443 | */ |
| 444 | if (!HAS_LLC(i915)((&(i915)->__info)->has_llc) && !IS_DGFX(i915)((&(i915)->__info)->is_dgfx)) |
| 445 | obj->cache_dirty = true1; |
| 446 | } |
| 447 | |
| 448 | void i915_gem_object_put_pages_shmem(struct drm_i915_gem_object *obj, struct sg_table *pages) |
| 449 | { |
| 450 | __i915_gem_object_release_shmem(obj, pages, true1); |
| 451 | |
| 452 | i915_gem_gtt_finish_pages(obj, pages); |
| 453 | |
| 454 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
| 455 | i915_gem_object_save_bit_17_swizzle(obj, pages); |
| 456 | |
| 457 | #ifdef __linux__ |
| 458 | shmem_sg_free_table(pages, file_inode(obj->base.filp)->i_mapping, |
| 459 | obj->mm.dirty, obj->mm.madv == I915_MADV_WILLNEED0); |
| 460 | #else |
| 461 | shmem_sg_free_table(pages, NULL((void *)0), |
| 462 | obj->mm.dirty, obj->mm.madv == I915_MADV_WILLNEED0, obj); |
| 463 | #endif |
| 464 | kfree(pages); |
| 465 | obj->mm.dirty = false0; |
| 466 | } |
| 467 | |
| 468 | static void |
| 469 | shmem_put_pages(struct drm_i915_gem_object *obj, struct sg_table *pages) |
| 470 | { |
| 471 | if (likely(i915_gem_object_has_struct_page(obj))__builtin_expect(!!(i915_gem_object_has_struct_page(obj)), 1)) |
| 472 | i915_gem_object_put_pages_shmem(obj, pages); |
| 473 | else |
| 474 | i915_gem_object_put_pages_phys(obj, pages); |
| 475 | } |
| 476 | |
| 477 | static int |
| 478 | shmem_pwrite(struct drm_i915_gem_object *obj, |
| 479 | const struct drm_i915_gem_pwrite *arg) |
| 480 | { |
| 481 | #ifdef __linux__ |
| 482 | struct address_space *mapping = obj->base.filp->f_mapping; |
| 483 | const struct address_space_operations *aops = mapping->a_ops; |
| 484 | #endif |
| 485 | char __user *user_data = u64_to_user_ptr(arg->data_ptr)((void *)(uintptr_t)(arg->data_ptr)); |
| 486 | u64 remain, offset; |
| 487 | unsigned int pg; |
| 488 | |
| 489 | /* Caller already validated user args */ |
| 490 | GEM_BUG_ON(!access_ok(user_data, arg->size))((void)0); |
| 491 | |
| 492 | if (!i915_gem_object_has_struct_page(obj)) |
| 493 | return i915_gem_object_pwrite_phys(obj, arg); |
| 494 | |
| 495 | /* |
| 496 | * Before we instantiate/pin the backing store for our use, we |
| 497 | * can prepopulate the shmemfs filp efficiently using a write into |
| 498 | * the pagecache. We avoid the penalty of instantiating all the |
| 499 | * pages, important if the user is just writing to a few and never |
| 500 | * uses the object on the GPU, and using a direct write into shmemfs |
| 501 | * allows it to avoid the cost of retrieving a page (either swapin |
| 502 | * or clearing-before-use) before it is overwritten. |
| 503 | */ |
| 504 | if (i915_gem_object_has_pages(obj)) |
| 505 | return -ENODEV19; |
| 506 | |
| 507 | if (obj->mm.madv != I915_MADV_WILLNEED0) |
| 508 | return -EFAULT14; |
| 509 | |
| 510 | /* |
| 511 | * Before the pages are instantiated the object is treated as being |
| 512 | * in the CPU domain. The pages will be clflushed as required before |
| 513 | * use, and we can freely write into the pages directly. If userspace |
| 514 | * races pwrite with any other operation; corruption will ensue - |
| 515 | * that is userspace's prerogative! |
| 516 | */ |
| 517 | |
| 518 | remain = arg->size; |
| 519 | offset = arg->offset; |
| 520 | pg = offset_in_page(offset)((vaddr_t)(offset) & ((1 << 12) - 1)); |
| 521 | |
| 522 | do { |
| 523 | unsigned int len, unwritten; |
| 524 | struct vm_page *page; |
| 525 | void *data, *vaddr; |
| 526 | int err; |
| 527 | char c; |
| 528 | |
| 529 | len = PAGE_SIZE(1 << 12) - pg; |
| 530 | if (len > remain) |
| 531 | len = remain; |
| 532 | |
| 533 | /* Prefault the user page to reduce potential recursion */ |
| 534 | err = __get_user(c, user_data)-copyin((user_data), &((c)), sizeof((c))); |
| 535 | if (err) |
| 536 | return err; |
| 537 | |
| 538 | err = __get_user(c, user_data + len - 1)-copyin((user_data + len - 1), &((c)), sizeof((c))); |
| 539 | if (err) |
| 540 | return err; |
| 541 | |
| 542 | #ifdef __linux__ |
| 543 | err = aops->write_begin(obj->base.filp, mapping, offset, len, |
| 544 | &page, &data); |
| 545 | if (err < 0) |
| 546 | return err; |
| 547 | #else |
| 548 | struct pglist plist; |
| 549 | TAILQ_INIT(&plist)do { (&plist)->tqh_first = ((void *)0); (&plist)-> tqh_last = &(&plist)->tqh_first; } while (0); |
| 550 | if (uvm_obj_wire(obj->base.uao, trunc_page(offset)((offset) & ~((1 << 12) - 1)), |
| 551 | trunc_page(offset)((offset) & ~((1 << 12) - 1)) + PAGE_SIZE(1 << 12), &plist)) { |
| 552 | return -ENOMEM12; |
| 553 | } |
| 554 | page = TAILQ_FIRST(&plist)((&plist)->tqh_first); |
| 555 | #endif |
| 556 | |
| 557 | vaddr = kmap_atomic(page); |
| 558 | unwritten = __copy_from_user_inatomic(vaddr + pg, |
| 559 | user_data, |
| 560 | len); |
| 561 | kunmap_atomic(vaddr); |
| 562 | |
| 563 | #ifdef __linux__ |
| 564 | err = aops->write_end(obj->base.filp, mapping, offset, len, |
| 565 | len - unwritten, page, data); |
| 566 | if (err < 0) |
| 567 | return err; |
| 568 | #else |
| 569 | uvm_obj_unwire(obj->base.uao, trunc_page(offset)((offset) & ~((1 << 12) - 1)), |
| 570 | trunc_page(offset)((offset) & ~((1 << 12) - 1)) + PAGE_SIZE(1 << 12)); |
| 571 | #endif |
| 572 | |
| 573 | /* We don't handle -EFAULT, leave it to the caller to check */ |
| 574 | if (unwritten) |
| 575 | return -ENODEV19; |
| 576 | |
| 577 | remain -= len; |
| 578 | user_data += len; |
| 579 | offset += len; |
| 580 | pg = 0; |
| 581 | } while (remain); |
| 582 | |
| 583 | return 0; |
| 584 | } |
| 585 | |
| 586 | static int |
| 587 | shmem_pread(struct drm_i915_gem_object *obj, |
| 588 | const struct drm_i915_gem_pread *arg) |
| 589 | { |
| 590 | if (!i915_gem_object_has_struct_page(obj)) |
| 591 | return i915_gem_object_pread_phys(obj, arg); |
| 592 | |
| 593 | return -ENODEV19; |
| 594 | } |
| 595 | |
| 596 | static void shmem_release(struct drm_i915_gem_object *obj) |
| 597 | { |
| 598 | if (i915_gem_object_has_struct_page(obj)) |
| 599 | i915_gem_object_release_memory_region(obj); |
| 600 | |
| 601 | #ifdef __linux__ |
| 602 | fput(obj->base.filp); |
| 603 | #endif |
| 604 | } |
| 605 | |
| 606 | const struct drm_i915_gem_object_ops i915_gem_shmem_ops = { |
| 607 | .name = "i915_gem_object_shmem", |
| 608 | .flags = I915_GEM_OBJECT_IS_SHRINKABLE(1UL << (1)), |
| 609 | |
| 610 | .get_pages = shmem_get_pages, |
| 611 | .put_pages = shmem_put_pages, |
| 612 | .truncate = shmem_truncate, |
| 613 | .shrink = shmem_shrink, |
| 614 | |
| 615 | .pwrite = shmem_pwrite, |
| 616 | .pread = shmem_pread, |
| 617 | |
| 618 | .release = shmem_release, |
| 619 | }; |
| 620 | |
| 621 | #ifdef __linux__ |
| 622 | static int __create_shmem(struct drm_i915_privateinteldrm_softc *i915, |
| 623 | struct drm_gem_object *obj, |
| 624 | resource_size_t size) |
| 625 | { |
| 626 | unsigned long flags = VM_NORESERVE; |
| 627 | struct file *filp; |
| 628 | |
| 629 | drm_gem_private_object_init(&i915->drm, obj, size); |
| 630 | |
| 631 | if (i915->mm.gemfs) |
| 632 | filp = shmem_file_setup_with_mnt(i915->mm.gemfs, "i915", size, |
| 633 | flags); |
| 634 | else |
| 635 | filp = shmem_file_setup("i915", size, flags); |
| 636 | if (IS_ERR(filp)) |
| 637 | return PTR_ERR(filp); |
| 638 | |
| 639 | obj->filp = filp; |
| 640 | return 0; |
| 641 | } |
| 642 | #endif |
| 643 | |
| 644 | static int shmem_object_init(struct intel_memory_region *mem, |
| 645 | struct drm_i915_gem_object *obj, |
| 646 | resource_size_t offset, |
| 647 | resource_size_t size, |
| 648 | resource_size_t page_size, |
| 649 | unsigned int flags) |
| 650 | { |
| 651 | static struct lock_class_key lock_class; |
| 652 | struct drm_i915_privateinteldrm_softc *i915 = mem->i915; |
| 653 | struct address_space *mapping; |
| 654 | unsigned int cache_level; |
| 655 | gfp_t mask; |
| 656 | int ret; |
| 657 | |
| 658 | #ifdef __linux__ |
| 659 | ret = __create_shmem(i915, &obj->base, size); |
| 660 | #else |
| 661 | ret = drm_gem_object_init(&i915->drm, &obj->base, size); |
| 662 | #endif |
| 663 | if (ret) |
| 664 | return ret; |
| 665 | |
| 666 | mask = GFP_HIGHUSER0 | __GFP_RECLAIMABLE0; |
| 667 | if (IS_I965GM(i915)IS_PLATFORM(i915, INTEL_I965GM) || IS_I965G(i915)IS_PLATFORM(i915, INTEL_I965G)) { |
| 668 | /* 965gm cannot relocate objects above 4GiB. */ |
| 669 | mask &= ~__GFP_HIGHMEM0; |
| 670 | mask |= __GFP_DMA320x00010000; |
Value stored to 'mask' is never read | |
| 671 | } |
| 672 | |
| 673 | #ifdef __linux__ |
| 674 | mapping = obj->base.filp->f_mapping; |
| 675 | mapping_set_gfp_mask(mapping, mask); |
| 676 | GEM_BUG_ON(!(mapping_gfp_mask(mapping) & __GFP_RECLAIM))((void)0); |
| 677 | #endif |
| 678 | |
| 679 | i915_gem_object_init(obj, &i915_gem_shmem_ops, &lock_class, flags); |
| 680 | obj->mem_flags |= I915_BO_FLAG_STRUCT_PAGE(1UL << (0)); |
| 681 | obj->write_domain = I915_GEM_DOMAIN_CPU0x00000001; |
| 682 | obj->read_domains = I915_GEM_DOMAIN_CPU0x00000001; |
| 683 | |
| 684 | if (HAS_LLC(i915)((&(i915)->__info)->has_llc)) |
| 685 | /* On some devices, we can have the GPU use the LLC (the CPU |
| 686 | * cache) for about a 10% performance improvement |
| 687 | * compared to uncached. Graphics requests other than |
| 688 | * display scanout are coherent with the CPU in |
| 689 | * accessing this cache. This means in this mode we |
| 690 | * don't need to clflush on the CPU side, and on the |
| 691 | * GPU side we only need to flush internal caches to |
| 692 | * get data visible to the CPU. |
| 693 | * |
| 694 | * However, we maintain the display planes as UC, and so |
| 695 | * need to rebind when first used as such. |
| 696 | */ |
| 697 | cache_level = I915_CACHE_LLC; |
| 698 | else |
| 699 | cache_level = I915_CACHE_NONE; |
| 700 | |
| 701 | i915_gem_object_set_cache_coherency(obj, cache_level); |
| 702 | |
| 703 | i915_gem_object_init_memory_region(obj, mem); |
| 704 | |
| 705 | return 0; |
| 706 | } |
| 707 | |
| 708 | struct drm_i915_gem_object * |
| 709 | i915_gem_object_create_shmem(struct drm_i915_privateinteldrm_softc *i915, |
| 710 | resource_size_t size) |
| 711 | { |
| 712 | return i915_gem_object_create_region(i915->mm.regions[INTEL_REGION_SMEM], |
| 713 | size, 0, 0); |
| 714 | } |
| 715 | |
| 716 | /* Allocate a new GEM object and fill it with the supplied data */ |
| 717 | #ifdef __linux__ |
| 718 | struct drm_i915_gem_object * |
| 719 | i915_gem_object_create_shmem_from_data(struct drm_i915_privateinteldrm_softc *dev_priv, |
| 720 | const void *data, resource_size_t size) |
| 721 | { |
| 722 | struct drm_i915_gem_object *obj; |
| 723 | struct file *file; |
| 724 | const struct address_space_operations *aops; |
| 725 | resource_size_t offset; |
| 726 | int err; |
| 727 | |
| 728 | GEM_WARN_ON(IS_DGFX(dev_priv))({ __builtin_expect(!!(!!(((&(dev_priv)->__info)->is_dgfx ))), 0); }); |
| 729 | obj = i915_gem_object_create_shmem(dev_priv, round_up(size, PAGE_SIZE)((((size) + (((1 << 12)) - 1)) / ((1 << 12))) * ( (1 << 12)))); |
| 730 | if (IS_ERR(obj)) |
| 731 | return obj; |
| 732 | |
| 733 | GEM_BUG_ON(obj->write_domain != I915_GEM_DOMAIN_CPU)((void)0); |
| 734 | |
| 735 | file = obj->base.filp; |
| 736 | aops = file->f_mapping->a_ops; |
| 737 | offset = 0; |
| 738 | do { |
| 739 | unsigned int len = min_t(typeof(size), size, PAGE_SIZE)({ typeof(size) __min_a = (size); typeof(size) __min_b = ((1 << 12)); __min_a < __min_b ? __min_a : __min_b; }); |
| 740 | struct vm_page *page; |
| 741 | void *pgdata, *vaddr; |
| 742 | |
| 743 | err = aops->write_begin(file, file->f_mapping, offset, len, |
| 744 | &page, &pgdata); |
| 745 | if (err < 0) |
| 746 | goto fail; |
| 747 | |
| 748 | vaddr = kmap(page); |
| 749 | memcpy(vaddr, data, len)__builtin_memcpy((vaddr), (data), (len)); |
| 750 | kunmap(page); |
| 751 | |
| 752 | err = aops->write_end(file, file->f_mapping, offset, len, len, |
| 753 | page, pgdata); |
| 754 | if (err < 0) |
| 755 | goto fail; |
| 756 | |
| 757 | size -= len; |
| 758 | data += len; |
| 759 | offset += len; |
| 760 | } while (size); |
| 761 | |
| 762 | return obj; |
| 763 | |
| 764 | fail: |
| 765 | i915_gem_object_put(obj); |
| 766 | return ERR_PTR(err); |
| 767 | } |
| 768 | #else /* !__linux__ */ |
| 769 | struct drm_i915_gem_object * |
| 770 | i915_gem_object_create_shmem_from_data(struct drm_i915_privateinteldrm_softc *dev_priv, |
| 771 | const void *data, resource_size_t size) |
| 772 | { |
| 773 | struct drm_i915_gem_object *obj; |
| 774 | struct uvm_object *uao; |
| 775 | resource_size_t offset; |
| 776 | int err; |
| 777 | |
| 778 | GEM_WARN_ON(IS_DGFX(dev_priv))({ __builtin_expect(!!(!!(((&(dev_priv)->__info)->is_dgfx ))), 0); }); |
| 779 | obj = i915_gem_object_create_shmem(dev_priv, round_up(size, PAGE_SIZE)((((size) + (((1 << 12)) - 1)) / ((1 << 12))) * ( (1 << 12)))); |
| 780 | if (IS_ERR(obj)) |
| 781 | return obj; |
| 782 | |
| 783 | GEM_BUG_ON(obj->write_domain != I915_GEM_DOMAIN_CPU)((void)0); |
| 784 | |
| 785 | uao = obj->base.uao; |
| 786 | offset = 0; |
| 787 | do { |
| 788 | unsigned int len = min_t(typeof(size), size, PAGE_SIZE)({ typeof(size) __min_a = (size); typeof(size) __min_b = ((1 << 12)); __min_a < __min_b ? __min_a : __min_b; }); |
| 789 | struct vm_page *page; |
| 790 | void *pgdata, *vaddr; |
| 791 | struct pglist plist; |
| 792 | |
| 793 | TAILQ_INIT(&plist)do { (&plist)->tqh_first = ((void *)0); (&plist)-> tqh_last = &(&plist)->tqh_first; } while (0); |
| 794 | if (uvm_obj_wire(uao, trunc_page(offset)((offset) & ~((1 << 12) - 1)), |
| 795 | trunc_page(offset)((offset) & ~((1 << 12) - 1)) + PAGE_SIZE(1 << 12), &plist)) { |
| 796 | err = -ENOMEM12; |
| 797 | goto fail; |
| 798 | } |
| 799 | page = TAILQ_FIRST(&plist)((&plist)->tqh_first); |
| 800 | |
| 801 | vaddr = kmap(page); |
| 802 | memcpy(vaddr, data, len)__builtin_memcpy((vaddr), (data), (len)); |
| 803 | kunmap_va(vaddr); |
| 804 | |
| 805 | uvm_obj_unwire(uao, trunc_page(offset)((offset) & ~((1 << 12) - 1)), |
| 806 | trunc_page(offset)((offset) & ~((1 << 12) - 1)) + PAGE_SIZE(1 << 12)); |
| 807 | |
| 808 | size -= len; |
| 809 | data += len; |
| 810 | offset += len; |
| 811 | } while (size); |
| 812 | |
| 813 | return obj; |
| 814 | |
| 815 | fail: |
| 816 | i915_gem_object_put(obj); |
| 817 | return ERR_PTR(err); |
| 818 | } |
| 819 | #endif |
| 820 | |
| 821 | static int init_shmem(struct intel_memory_region *mem) |
| 822 | { |
| 823 | i915_gemfs_init(mem->i915); |
| 824 | intel_memory_region_set_name(mem, "system"); |
| 825 | |
| 826 | return 0; /* We have fallback to the kernel mnt if gemfs init failed. */ |
| 827 | } |
| 828 | |
| 829 | static int release_shmem(struct intel_memory_region *mem) |
| 830 | { |
| 831 | i915_gemfs_fini(mem->i915); |
| 832 | return 0; |
| 833 | } |
| 834 | |
| 835 | static const struct intel_memory_region_ops shmem_region_ops = { |
| 836 | .init = init_shmem, |
| 837 | .release = release_shmem, |
| 838 | .init_object = shmem_object_init, |
| 839 | }; |
| 840 | |
| 841 | struct intel_memory_region *i915_gem_shmem_setup(struct drm_i915_privateinteldrm_softc *i915, |
| 842 | u16 type, u16 instance) |
| 843 | { |
| 844 | return intel_memory_region_create(i915, 0, |
| 845 | totalram_pages() << PAGE_SHIFT12, |
| 846 | PAGE_SIZE(1 << 12), 0, 0, |
| 847 | type, instance, |
| 848 | &shmem_region_ops); |
| 849 | } |
| 850 | |
| 851 | bool_Bool i915_gem_object_is_shmem(const struct drm_i915_gem_object *obj) |
| 852 | { |
| 853 | return obj->ops == &i915_gem_shmem_ops; |
| 854 | } |