1 /* 2 * Copyright © 2017 Intel Corporation 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 21 * IN THE SOFTWARE. 22 * 23 */ 24 25 #include <linux/sched/mm.h> 26 27 #include "display/intel_frontbuffer.h" 28 #include "i915_drv.h" 29 #include "i915_gem_clflush.h" 30 #include "i915_gem_context.h" 31 #include "i915_gem_mman.h" 32 #include "i915_gem_object.h" 33 #include "i915_globals.h" 34 #include "i915_memcpy.h" 35 #include "i915_trace.h" 36 37 static struct i915_global_object { 38 struct i915_global base; 39 struct kmem_cache *slab_objects; 40 } global; 41 42 static const struct drm_gem_object_funcs i915_gem_object_funcs; 43 44 struct drm_i915_gem_object *i915_gem_object_alloc(void) 45 { 46 struct drm_i915_gem_object *obj; 47 48 obj = kmem_cache_zalloc(global.slab_objects, GFP_KERNEL); 49 if (!obj) 50 return NULL; 51 obj->base.funcs = &i915_gem_object_funcs; 52 53 return obj; 54 } 55 56 void i915_gem_object_free(struct drm_i915_gem_object *obj) 57 { 58 return kmem_cache_free(global.slab_objects, obj); 59 } 60 61 void i915_gem_object_init(struct drm_i915_gem_object *obj, 62 const struct drm_i915_gem_object_ops *ops, 63 struct lock_class_key *key, unsigned flags) 64 { 65 /* 66 * A gem object is embedded both in a struct ttm_buffer_object :/ and 67 * in a drm_i915_gem_object. Make sure they are aliased. 68 */ 69 BUILD_BUG_ON(offsetof(typeof(*obj), base) != 70 offsetof(typeof(*obj), __do_not_access.base)); 71 72 spin_lock_init(&obj->vma.lock); 73 INIT_LIST_HEAD(&obj->vma.list); 74 75 INIT_LIST_HEAD(&obj->mm.link); 76 77 INIT_LIST_HEAD(&obj->lut_list); 78 spin_lock_init(&obj->lut_lock); 79 80 spin_lock_init(&obj->mmo.lock); 81 obj->mmo.offsets = RB_ROOT; 82 83 init_rcu_head(&obj->rcu); 84 85 obj->ops = ops; 86 GEM_BUG_ON(flags & ~I915_BO_ALLOC_FLAGS); 87 obj->flags = flags; 88 89 obj->mm.madv = I915_MADV_WILLNEED; 90 INIT_RADIX_TREE(&obj->mm.get_page.radix, GFP_KERNEL | __GFP_NOWARN); 91 mutex_init(&obj->mm.get_page.lock); 92 INIT_RADIX_TREE(&obj->mm.get_dma_page.radix, GFP_KERNEL | __GFP_NOWARN); 93 mutex_init(&obj->mm.get_dma_page.lock); 94 } 95 96 /** 97 * Mark up the object's coherency levels for a given cache_level 98 * @obj: #drm_i915_gem_object 99 * @cache_level: cache level 100 */ 101 void i915_gem_object_set_cache_coherency(struct drm_i915_gem_object *obj, 102 unsigned int cache_level) 103 { 104 obj->cache_level = cache_level; 105 106 if (cache_level != I915_CACHE_NONE) 107 obj->cache_coherent = (I915_BO_CACHE_COHERENT_FOR_READ | 108 I915_BO_CACHE_COHERENT_FOR_WRITE); 109 else if (HAS_LLC(to_i915(obj->base.dev))) 110 obj->cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ; 111 else 112 obj->cache_coherent = 0; 113 114 obj->cache_dirty = 115 !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE); 116 } 117 118 static void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file) 119 { 120 struct drm_i915_gem_object *obj = to_intel_bo(gem); 121 struct drm_i915_file_private *fpriv = file->driver_priv; 122 struct i915_lut_handle bookmark = {}; 123 struct i915_mmap_offset *mmo, *mn; 124 struct i915_lut_handle *lut, *ln; 125 LIST_HEAD(close); 126 127 spin_lock(&obj->lut_lock); 128 list_for_each_entry_safe(lut, ln, &obj->lut_list, obj_link) { 129 struct i915_gem_context *ctx = lut->ctx; 130 131 if (ctx && ctx->file_priv == fpriv) { 132 i915_gem_context_get(ctx); 133 list_move(&lut->obj_link, &close); 134 } 135 136 /* Break long locks, and carefully continue on from this spot */ 137 if (&ln->obj_link != &obj->lut_list) { 138 list_add_tail(&bookmark.obj_link, &ln->obj_link); 139 if (cond_resched_lock(&obj->lut_lock)) 140 list_safe_reset_next(&bookmark, ln, obj_link); 141 __list_del_entry(&bookmark.obj_link); 142 } 143 } 144 spin_unlock(&obj->lut_lock); 145 146 spin_lock(&obj->mmo.lock); 147 rbtree_postorder_for_each_entry_safe(mmo, mn, &obj->mmo.offsets, offset) 148 drm_vma_node_revoke(&mmo->vma_node, file); 149 spin_unlock(&obj->mmo.lock); 150 151 list_for_each_entry_safe(lut, ln, &close, obj_link) { 152 struct i915_gem_context *ctx = lut->ctx; 153 struct i915_vma *vma; 154 155 /* 156 * We allow the process to have multiple handles to the same 157 * vma, in the same fd namespace, by virtue of flink/open. 158 */ 159 160 mutex_lock(&ctx->lut_mutex); 161 vma = radix_tree_delete(&ctx->handles_vma, lut->handle); 162 if (vma) { 163 GEM_BUG_ON(vma->obj != obj); 164 GEM_BUG_ON(!atomic_read(&vma->open_count)); 165 i915_vma_close(vma); 166 } 167 mutex_unlock(&ctx->lut_mutex); 168 169 i915_gem_context_put(lut->ctx); 170 i915_lut_handle_free(lut); 171 i915_gem_object_put(obj); 172 } 173 } 174 175 static void __i915_gem_free_object_rcu(struct rcu_head *head) 176 { 177 struct drm_i915_gem_object *obj = 178 container_of(head, typeof(*obj), rcu); 179 struct drm_i915_private *i915 = to_i915(obj->base.dev); 180 181 dma_resv_fini(&obj->base._resv); 182 i915_gem_object_free(obj); 183 184 GEM_BUG_ON(!atomic_read(&i915->mm.free_count)); 185 atomic_dec(&i915->mm.free_count); 186 } 187 188 static void __i915_gem_object_free_mmaps(struct drm_i915_gem_object *obj) 189 { 190 /* Skip serialisation and waking the device if known to be not used. */ 191 192 if (obj->userfault_count) 193 i915_gem_object_release_mmap_gtt(obj); 194 195 if (!RB_EMPTY_ROOT(&obj->mmo.offsets)) { 196 struct i915_mmap_offset *mmo, *mn; 197 198 i915_gem_object_release_mmap_offset(obj); 199 200 rbtree_postorder_for_each_entry_safe(mmo, mn, 201 &obj->mmo.offsets, 202 offset) { 203 drm_vma_offset_remove(obj->base.dev->vma_offset_manager, 204 &mmo->vma_node); 205 kfree(mmo); 206 } 207 obj->mmo.offsets = RB_ROOT; 208 } 209 } 210 211 static void __i915_gem_free_objects(struct drm_i915_private *i915, 212 struct llist_node *freed) 213 { 214 struct drm_i915_gem_object *obj, *on; 215 216 llist_for_each_entry_safe(obj, on, freed, freed) { 217 trace_i915_gem_object_destroy(obj); 218 219 if (!list_empty(&obj->vma.list)) { 220 struct i915_vma *vma; 221 222 /* 223 * Note that the vma keeps an object reference while 224 * it is active, so it *should* not sleep while we 225 * destroy it. Our debug code errs insits it *might*. 226 * For the moment, play along. 227 */ 228 spin_lock(&obj->vma.lock); 229 while ((vma = list_first_entry_or_null(&obj->vma.list, 230 struct i915_vma, 231 obj_link))) { 232 GEM_BUG_ON(vma->obj != obj); 233 spin_unlock(&obj->vma.lock); 234 235 __i915_vma_put(vma); 236 237 spin_lock(&obj->vma.lock); 238 } 239 spin_unlock(&obj->vma.lock); 240 } 241 242 __i915_gem_object_free_mmaps(obj); 243 244 GEM_BUG_ON(!list_empty(&obj->lut_list)); 245 246 atomic_set(&obj->mm.pages_pin_count, 0); 247 __i915_gem_object_put_pages(obj); 248 GEM_BUG_ON(i915_gem_object_has_pages(obj)); 249 bitmap_free(obj->bit_17); 250 251 if (obj->base.import_attach) 252 drm_prime_gem_destroy(&obj->base, NULL); 253 254 drm_gem_free_mmap_offset(&obj->base); 255 256 if (obj->ops->release) 257 obj->ops->release(obj); 258 259 if (obj->mm.n_placements > 1) 260 kfree(obj->mm.placements); 261 262 if (obj->shares_resv_from) 263 i915_vm_resv_put(obj->shares_resv_from); 264 265 /* But keep the pointer alive for RCU-protected lookups */ 266 call_rcu(&obj->rcu, __i915_gem_free_object_rcu); 267 cond_resched(); 268 } 269 } 270 271 void i915_gem_flush_free_objects(struct drm_i915_private *i915) 272 { 273 struct llist_node *freed = llist_del_all(&i915->mm.free_list); 274 275 if (unlikely(freed)) 276 __i915_gem_free_objects(i915, freed); 277 } 278 279 static void __i915_gem_free_work(struct work_struct *work) 280 { 281 struct drm_i915_private *i915 = 282 container_of(work, struct drm_i915_private, mm.free_work); 283 284 i915_gem_flush_free_objects(i915); 285 } 286 287 static void i915_gem_free_object(struct drm_gem_object *gem_obj) 288 { 289 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj); 290 struct drm_i915_private *i915 = to_i915(obj->base.dev); 291 292 GEM_BUG_ON(i915_gem_object_is_framebuffer(obj)); 293 294 /* 295 * Before we free the object, make sure any pure RCU-only 296 * read-side critical sections are complete, e.g. 297 * i915_gem_busy_ioctl(). For the corresponding synchronized 298 * lookup see i915_gem_object_lookup_rcu(). 299 */ 300 atomic_inc(&i915->mm.free_count); 301 302 /* 303 * This serializes freeing with the shrinker. Since the free 304 * is delayed, first by RCU then by the workqueue, we want the 305 * shrinker to be able to free pages of unreferenced objects, 306 * or else we may oom whilst there are plenty of deferred 307 * freed objects. 308 */ 309 i915_gem_object_make_unshrinkable(obj); 310 311 /* 312 * Since we require blocking on struct_mutex to unbind the freed 313 * object from the GPU before releasing resources back to the 314 * system, we can not do that directly from the RCU callback (which may 315 * be a softirq context), but must instead then defer that work onto a 316 * kthread. We use the RCU callback rather than move the freed object 317 * directly onto the work queue so that we can mix between using the 318 * worker and performing frees directly from subsequent allocations for 319 * crude but effective memory throttling. 320 */ 321 if (llist_add(&obj->freed, &i915->mm.free_list)) 322 queue_work(i915->wq, &i915->mm.free_work); 323 } 324 325 void __i915_gem_object_flush_frontbuffer(struct drm_i915_gem_object *obj, 326 enum fb_op_origin origin) 327 { 328 struct intel_frontbuffer *front; 329 330 front = __intel_frontbuffer_get(obj); 331 if (front) { 332 intel_frontbuffer_flush(front, origin); 333 intel_frontbuffer_put(front); 334 } 335 } 336 337 void __i915_gem_object_invalidate_frontbuffer(struct drm_i915_gem_object *obj, 338 enum fb_op_origin origin) 339 { 340 struct intel_frontbuffer *front; 341 342 front = __intel_frontbuffer_get(obj); 343 if (front) { 344 intel_frontbuffer_invalidate(front, origin); 345 intel_frontbuffer_put(front); 346 } 347 } 348 349 static void 350 i915_gem_object_read_from_page_kmap(struct drm_i915_gem_object *obj, u64 offset, void *dst, int size) 351 { 352 void *src_map; 353 void *src_ptr; 354 355 src_map = kmap_atomic(i915_gem_object_get_page(obj, offset >> PAGE_SHIFT)); 356 357 src_ptr = src_map + offset_in_page(offset); 358 if (!(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ)) 359 drm_clflush_virt_range(src_ptr, size); 360 memcpy(dst, src_ptr, size); 361 362 kunmap_atomic(src_map); 363 } 364 365 static void 366 i915_gem_object_read_from_page_iomap(struct drm_i915_gem_object *obj, u64 offset, void *dst, int size) 367 { 368 void __iomem *src_map; 369 void __iomem *src_ptr; 370 dma_addr_t dma = i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT); 371 372 src_map = io_mapping_map_wc(&obj->mm.region->iomap, 373 dma - obj->mm.region->region.start, 374 PAGE_SIZE); 375 376 src_ptr = src_map + offset_in_page(offset); 377 if (!i915_memcpy_from_wc(dst, (void __force *)src_ptr, size)) 378 memcpy_fromio(dst, src_ptr, size); 379 380 io_mapping_unmap(src_map); 381 } 382 383 /** 384 * i915_gem_object_read_from_page - read data from the page of a GEM object 385 * @obj: GEM object to read from 386 * @offset: offset within the object 387 * @dst: buffer to store the read data 388 * @size: size to read 389 * 390 * Reads data from @obj at the specified offset. The requested region to read 391 * from can't cross a page boundary. The caller must ensure that @obj pages 392 * are pinned and that @obj is synced wrt. any related writes. 393 * 394 * Returns 0 on success or -ENODEV if the type of @obj's backing store is 395 * unsupported. 396 */ 397 int i915_gem_object_read_from_page(struct drm_i915_gem_object *obj, u64 offset, void *dst, int size) 398 { 399 GEM_BUG_ON(offset >= obj->base.size); 400 GEM_BUG_ON(offset_in_page(offset) > PAGE_SIZE - size); 401 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj)); 402 403 if (i915_gem_object_has_struct_page(obj)) 404 i915_gem_object_read_from_page_kmap(obj, offset, dst, size); 405 else if (i915_gem_object_has_iomem(obj)) 406 i915_gem_object_read_from_page_iomap(obj, offset, dst, size); 407 else 408 return -ENODEV; 409 410 return 0; 411 } 412 413 void i915_gem_init__objects(struct drm_i915_private *i915) 414 { 415 INIT_WORK(&i915->mm.free_work, __i915_gem_free_work); 416 } 417 418 static void i915_global_objects_shrink(void) 419 { 420 kmem_cache_shrink(global.slab_objects); 421 } 422 423 static void i915_global_objects_exit(void) 424 { 425 kmem_cache_destroy(global.slab_objects); 426 } 427 428 static struct i915_global_object global = { { 429 .shrink = i915_global_objects_shrink, 430 .exit = i915_global_objects_exit, 431 } }; 432 433 int __init i915_global_objects_init(void) 434 { 435 global.slab_objects = 436 KMEM_CACHE(drm_i915_gem_object, SLAB_HWCACHE_ALIGN); 437 if (!global.slab_objects) 438 return -ENOMEM; 439 440 i915_global_register(&global.base); 441 return 0; 442 } 443 444 static const struct drm_gem_object_funcs i915_gem_object_funcs = { 445 .free = i915_gem_free_object, 446 .close = i915_gem_close_object, 447 .export = i915_gem_prime_export, 448 }; 449 450 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 451 #include "selftests/huge_gem_object.c" 452 #include "selftests/huge_pages.c" 453 #include "selftests/i915_gem_object.c" 454 #include "selftests/i915_gem_coherency.c" 455 #endif 456