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 "gt/intel_gt.h" 29 #include "i915_drv.h" 30 #include "i915_gem_clflush.h" 31 #include "i915_gem_context.h" 32 #include "i915_gem_mman.h" 33 #include "i915_gem_object.h" 34 #include "i915_globals.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) 64 { 65 __mutex_init(&obj->mm.lock, ops->name ?: "obj->mm.lock", key); 66 67 spin_lock_init(&obj->vma.lock); 68 INIT_LIST_HEAD(&obj->vma.list); 69 70 INIT_LIST_HEAD(&obj->mm.link); 71 72 INIT_LIST_HEAD(&obj->lut_list); 73 spin_lock_init(&obj->lut_lock); 74 75 spin_lock_init(&obj->mmo.lock); 76 obj->mmo.offsets = RB_ROOT; 77 78 init_rcu_head(&obj->rcu); 79 80 obj->ops = ops; 81 82 obj->mm.madv = I915_MADV_WILLNEED; 83 INIT_RADIX_TREE(&obj->mm.get_page.radix, GFP_KERNEL | __GFP_NOWARN); 84 mutex_init(&obj->mm.get_page.lock); 85 INIT_RADIX_TREE(&obj->mm.get_dma_page.radix, GFP_KERNEL | __GFP_NOWARN); 86 mutex_init(&obj->mm.get_dma_page.lock); 87 88 if (IS_ENABLED(CONFIG_LOCKDEP) && i915_gem_object_is_shrinkable(obj)) 89 i915_gem_shrinker_taints_mutex(to_i915(obj->base.dev), 90 &obj->mm.lock); 91 } 92 93 /** 94 * Mark up the object's coherency levels for a given cache_level 95 * @obj: #drm_i915_gem_object 96 * @cache_level: cache level 97 */ 98 void i915_gem_object_set_cache_coherency(struct drm_i915_gem_object *obj, 99 unsigned int cache_level) 100 { 101 obj->cache_level = cache_level; 102 103 if (cache_level != I915_CACHE_NONE) 104 obj->cache_coherent = (I915_BO_CACHE_COHERENT_FOR_READ | 105 I915_BO_CACHE_COHERENT_FOR_WRITE); 106 else if (HAS_LLC(to_i915(obj->base.dev))) 107 obj->cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ; 108 else 109 obj->cache_coherent = 0; 110 111 obj->cache_dirty = 112 !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE); 113 } 114 115 static void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file) 116 { 117 struct drm_i915_gem_object *obj = to_intel_bo(gem); 118 struct drm_i915_file_private *fpriv = file->driver_priv; 119 struct i915_lut_handle bookmark = {}; 120 struct i915_mmap_offset *mmo, *mn; 121 struct i915_lut_handle *lut, *ln; 122 LIST_HEAD(close); 123 124 spin_lock(&obj->lut_lock); 125 list_for_each_entry_safe(lut, ln, &obj->lut_list, obj_link) { 126 struct i915_gem_context *ctx = lut->ctx; 127 128 if (ctx && ctx->file_priv == fpriv) { 129 i915_gem_context_get(ctx); 130 list_move(&lut->obj_link, &close); 131 } 132 133 /* Break long locks, and carefully continue on from this spot */ 134 if (&ln->obj_link != &obj->lut_list) { 135 list_add_tail(&bookmark.obj_link, &ln->obj_link); 136 if (cond_resched_lock(&obj->lut_lock)) 137 list_safe_reset_next(&bookmark, ln, obj_link); 138 __list_del_entry(&bookmark.obj_link); 139 } 140 } 141 spin_unlock(&obj->lut_lock); 142 143 spin_lock(&obj->mmo.lock); 144 rbtree_postorder_for_each_entry_safe(mmo, mn, &obj->mmo.offsets, offset) 145 drm_vma_node_revoke(&mmo->vma_node, file); 146 spin_unlock(&obj->mmo.lock); 147 148 list_for_each_entry_safe(lut, ln, &close, obj_link) { 149 struct i915_gem_context *ctx = lut->ctx; 150 struct i915_vma *vma; 151 152 /* 153 * We allow the process to have multiple handles to the same 154 * vma, in the same fd namespace, by virtue of flink/open. 155 */ 156 157 mutex_lock(&ctx->lut_mutex); 158 vma = radix_tree_delete(&ctx->handles_vma, lut->handle); 159 if (vma) { 160 GEM_BUG_ON(vma->obj != obj); 161 GEM_BUG_ON(!atomic_read(&vma->open_count)); 162 i915_vma_close(vma); 163 } 164 mutex_unlock(&ctx->lut_mutex); 165 166 i915_gem_context_put(lut->ctx); 167 i915_lut_handle_free(lut); 168 i915_gem_object_put(obj); 169 } 170 } 171 172 static void __i915_gem_free_object_rcu(struct rcu_head *head) 173 { 174 struct drm_i915_gem_object *obj = 175 container_of(head, typeof(*obj), rcu); 176 struct drm_i915_private *i915 = to_i915(obj->base.dev); 177 178 dma_resv_fini(&obj->base._resv); 179 i915_gem_object_free(obj); 180 181 GEM_BUG_ON(!atomic_read(&i915->mm.free_count)); 182 atomic_dec(&i915->mm.free_count); 183 } 184 185 static void __i915_gem_object_free_mmaps(struct drm_i915_gem_object *obj) 186 { 187 /* Skip serialisation and waking the device if known to be not used. */ 188 189 if (obj->userfault_count) 190 i915_gem_object_release_mmap_gtt(obj); 191 192 if (!RB_EMPTY_ROOT(&obj->mmo.offsets)) { 193 struct i915_mmap_offset *mmo, *mn; 194 195 i915_gem_object_release_mmap_offset(obj); 196 197 rbtree_postorder_for_each_entry_safe(mmo, mn, 198 &obj->mmo.offsets, 199 offset) { 200 drm_vma_offset_remove(obj->base.dev->vma_offset_manager, 201 &mmo->vma_node); 202 kfree(mmo); 203 } 204 obj->mmo.offsets = RB_ROOT; 205 } 206 } 207 208 static void __i915_gem_free_objects(struct drm_i915_private *i915, 209 struct llist_node *freed) 210 { 211 struct drm_i915_gem_object *obj, *on; 212 213 llist_for_each_entry_safe(obj, on, freed, freed) { 214 trace_i915_gem_object_destroy(obj); 215 216 if (!list_empty(&obj->vma.list)) { 217 struct i915_vma *vma; 218 219 /* 220 * Note that the vma keeps an object reference while 221 * it is active, so it *should* not sleep while we 222 * destroy it. Our debug code errs insits it *might*. 223 * For the moment, play along. 224 */ 225 spin_lock(&obj->vma.lock); 226 while ((vma = list_first_entry_or_null(&obj->vma.list, 227 struct i915_vma, 228 obj_link))) { 229 GEM_BUG_ON(vma->obj != obj); 230 spin_unlock(&obj->vma.lock); 231 232 __i915_vma_put(vma); 233 234 spin_lock(&obj->vma.lock); 235 } 236 spin_unlock(&obj->vma.lock); 237 } 238 239 __i915_gem_object_free_mmaps(obj); 240 241 GEM_BUG_ON(!list_empty(&obj->lut_list)); 242 243 atomic_set(&obj->mm.pages_pin_count, 0); 244 __i915_gem_object_put_pages(obj); 245 GEM_BUG_ON(i915_gem_object_has_pages(obj)); 246 bitmap_free(obj->bit_17); 247 248 if (obj->base.import_attach) 249 drm_prime_gem_destroy(&obj->base, NULL); 250 251 drm_gem_free_mmap_offset(&obj->base); 252 253 if (obj->ops->release) 254 obj->ops->release(obj); 255 256 /* But keep the pointer alive for RCU-protected lookups */ 257 call_rcu(&obj->rcu, __i915_gem_free_object_rcu); 258 cond_resched(); 259 } 260 } 261 262 void i915_gem_flush_free_objects(struct drm_i915_private *i915) 263 { 264 struct llist_node *freed = llist_del_all(&i915->mm.free_list); 265 266 if (unlikely(freed)) 267 __i915_gem_free_objects(i915, freed); 268 } 269 270 static void __i915_gem_free_work(struct work_struct *work) 271 { 272 struct drm_i915_private *i915 = 273 container_of(work, struct drm_i915_private, mm.free_work); 274 275 i915_gem_flush_free_objects(i915); 276 } 277 278 static void i915_gem_free_object(struct drm_gem_object *gem_obj) 279 { 280 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj); 281 struct drm_i915_private *i915 = to_i915(obj->base.dev); 282 283 GEM_BUG_ON(i915_gem_object_is_framebuffer(obj)); 284 285 /* 286 * Before we free the object, make sure any pure RCU-only 287 * read-side critical sections are complete, e.g. 288 * i915_gem_busy_ioctl(). For the corresponding synchronized 289 * lookup see i915_gem_object_lookup_rcu(). 290 */ 291 atomic_inc(&i915->mm.free_count); 292 293 /* 294 * This serializes freeing with the shrinker. Since the free 295 * is delayed, first by RCU then by the workqueue, we want the 296 * shrinker to be able to free pages of unreferenced objects, 297 * or else we may oom whilst there are plenty of deferred 298 * freed objects. 299 */ 300 i915_gem_object_make_unshrinkable(obj); 301 302 /* 303 * Since we require blocking on struct_mutex to unbind the freed 304 * object from the GPU before releasing resources back to the 305 * system, we can not do that directly from the RCU callback (which may 306 * be a softirq context), but must instead then defer that work onto a 307 * kthread. We use the RCU callback rather than move the freed object 308 * directly onto the work queue so that we can mix between using the 309 * worker and performing frees directly from subsequent allocations for 310 * crude but effective memory throttling. 311 */ 312 if (llist_add(&obj->freed, &i915->mm.free_list)) 313 queue_work(i915->wq, &i915->mm.free_work); 314 } 315 316 static bool gpu_write_needs_clflush(struct drm_i915_gem_object *obj) 317 { 318 return !(obj->cache_level == I915_CACHE_NONE || 319 obj->cache_level == I915_CACHE_WT); 320 } 321 322 void 323 i915_gem_object_flush_write_domain(struct drm_i915_gem_object *obj, 324 unsigned int flush_domains) 325 { 326 struct i915_vma *vma; 327 328 assert_object_held(obj); 329 330 if (!(obj->write_domain & flush_domains)) 331 return; 332 333 switch (obj->write_domain) { 334 case I915_GEM_DOMAIN_GTT: 335 spin_lock(&obj->vma.lock); 336 for_each_ggtt_vma(vma, obj) { 337 if (i915_vma_unset_ggtt_write(vma)) 338 intel_gt_flush_ggtt_writes(vma->vm->gt); 339 } 340 spin_unlock(&obj->vma.lock); 341 342 i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU); 343 break; 344 345 case I915_GEM_DOMAIN_WC: 346 wmb(); 347 break; 348 349 case I915_GEM_DOMAIN_CPU: 350 i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC); 351 break; 352 353 case I915_GEM_DOMAIN_RENDER: 354 if (gpu_write_needs_clflush(obj)) 355 obj->cache_dirty = true; 356 break; 357 } 358 359 obj->write_domain = 0; 360 } 361 362 void __i915_gem_object_flush_frontbuffer(struct drm_i915_gem_object *obj, 363 enum fb_op_origin origin) 364 { 365 struct intel_frontbuffer *front; 366 367 front = __intel_frontbuffer_get(obj); 368 if (front) { 369 intel_frontbuffer_flush(front, origin); 370 intel_frontbuffer_put(front); 371 } 372 } 373 374 void __i915_gem_object_invalidate_frontbuffer(struct drm_i915_gem_object *obj, 375 enum fb_op_origin origin) 376 { 377 struct intel_frontbuffer *front; 378 379 front = __intel_frontbuffer_get(obj); 380 if (front) { 381 intel_frontbuffer_invalidate(front, origin); 382 intel_frontbuffer_put(front); 383 } 384 } 385 386 void i915_gem_init__objects(struct drm_i915_private *i915) 387 { 388 INIT_WORK(&i915->mm.free_work, __i915_gem_free_work); 389 } 390 391 static void i915_global_objects_shrink(void) 392 { 393 kmem_cache_shrink(global.slab_objects); 394 } 395 396 static void i915_global_objects_exit(void) 397 { 398 kmem_cache_destroy(global.slab_objects); 399 } 400 401 static struct i915_global_object global = { { 402 .shrink = i915_global_objects_shrink, 403 .exit = i915_global_objects_exit, 404 } }; 405 406 int __init i915_global_objects_init(void) 407 { 408 global.slab_objects = 409 KMEM_CACHE(drm_i915_gem_object, SLAB_HWCACHE_ALIGN); 410 if (!global.slab_objects) 411 return -ENOMEM; 412 413 i915_global_register(&global.base); 414 return 0; 415 } 416 417 static const struct drm_gem_object_funcs i915_gem_object_funcs = { 418 .free = i915_gem_free_object, 419 .close = i915_gem_close_object, 420 .export = i915_gem_prime_export, 421 }; 422 423 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 424 #include "selftests/huge_gem_object.c" 425 #include "selftests/huge_pages.c" 426 #include "selftests/i915_gem_object.c" 427 #include "selftests/i915_gem_coherency.c" 428 #endif 429