gem/i915_gem_object.c

   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 struct drm_i915_gem_object *i915_gem_object_alloc(void)
  43 {
  44         return kmem_cache_zalloc(global.slab_objects, GFP_KERNEL);
  45 }
  46
  47 void i915_gem_object_free(struct drm_i915_gem_object *obj)
  48 {
  49         return kmem_cache_free(global.slab_objects, obj);
  50 }
  51
  52 void i915_gem_object_init(struct drm_i915_gem_object *obj,
  53                           const struct drm_i915_gem_object_ops *ops,
  54                           struct lock_class_key *key)
  55 {
  56         __mutex_init(&obj->mm.lock, ops->name ?: "obj->mm.lock", key);
  57
  58         spin_lock_init(&obj->vma.lock);
  59         INIT_LIST_HEAD(&obj->vma.list);
  60
  61         INIT_LIST_HEAD(&obj->mm.link);
  62
  63         INIT_LIST_HEAD(&obj->lut_list);
  64         spin_lock_init(&obj->lut_lock);
  65
  66         spin_lock_init(&obj->mmo.lock);
  67         obj->mmo.offsets = RB_ROOT;
  68
  69         init_rcu_head(&obj->rcu);
  70
  71         obj->ops = ops;
  72
  73         obj->mm.madv = I915_MADV_WILLNEED;
  74         INIT_RADIX_TREE(&obj->mm.get_page.radix, GFP_KERNEL | __GFP_NOWARN);
  75         mutex_init(&obj->mm.get_page.lock);
  76
  77         if (IS_ENABLED(CONFIG_LOCKDEP) && i915_gem_object_is_shrinkable(obj))
  78                 i915_gem_shrinker_taints_mutex(to_i915(obj->base.dev),
  79                                                &obj->mm.lock);
  80 }
  81
  82 /**
  83  * Mark up the object's coherency levels for a given cache_level
  84  * @obj: #drm_i915_gem_object
  85  * @cache_level: cache level
  86  */
  87 void i915_gem_object_set_cache_coherency(struct drm_i915_gem_object *obj,
  88                                          unsigned int cache_level)
  89 {
  90         obj->cache_level = cache_level;
  91
  92         if (cache_level != I915_CACHE_NONE)
  93                 obj->cache_coherent = (I915_BO_CACHE_COHERENT_FOR_READ |
  94                                        I915_BO_CACHE_COHERENT_FOR_WRITE);
  95         else if (HAS_LLC(to_i915(obj->base.dev)))
  96                 obj->cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ;
  97         else
  98                 obj->cache_coherent = 0;
  99
 100         obj->cache_dirty =
 101                 !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE);
 102 }
 103
 104 void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file)
 105 {
 106         struct drm_i915_gem_object *obj = to_intel_bo(gem);
 107         struct drm_i915_file_private *fpriv = file->driver_priv;
 108         struct i915_lut_handle bookmark = {};
 109         struct i915_mmap_offset *mmo, *mn;
 110         struct i915_lut_handle *lut, *ln;
 111         LIST_HEAD(close);
 112
 113         spin_lock(&obj->lut_lock);
 114         list_for_each_entry_safe(lut, ln, &obj->lut_list, obj_link) {
 115                 struct i915_gem_context *ctx = lut->ctx;
 116
 117                 if (ctx && ctx->file_priv == fpriv) {
 118                         i915_gem_context_get(ctx);
 119                         list_move(&lut->obj_link, &close);
 120                 }
 121
 122                 /* Break long locks, and carefully continue on from this spot */
 123                 if (&ln->obj_link != &obj->lut_list) {
 124                         list_add_tail(&bookmark.obj_link, &ln->obj_link);
 125                         if (cond_resched_lock(&obj->lut_lock))
 126                                 list_safe_reset_next(&bookmark, ln, obj_link);
 127                         __list_del_entry(&bookmark.obj_link);
 128                 }
 129         }
 130         spin_unlock(&obj->lut_lock);
 131
 132         spin_lock(&obj->mmo.lock);
 133         rbtree_postorder_for_each_entry_safe(mmo, mn, &obj->mmo.offsets, offset)
 134                 drm_vma_node_revoke(&mmo->vma_node, file);
 135         spin_unlock(&obj->mmo.lock);
 136
 137         list_for_each_entry_safe(lut, ln, &close, obj_link) {
 138                 struct i915_gem_context *ctx = lut->ctx;
 139                 struct i915_vma *vma;
 140
 141                 /*
 142                  * We allow the process to have multiple handles to the same
 143                  * vma, in the same fd namespace, by virtue of flink/open.
 144                  */
 145
 146                 mutex_lock(&ctx->lut_mutex);
 147                 vma = radix_tree_delete(&ctx->handles_vma, lut->handle);
 148                 if (vma) {
 149                         GEM_BUG_ON(vma->obj != obj);
 150                         GEM_BUG_ON(!atomic_read(&vma->open_count));
 151                         i915_vma_close(vma);
 152                 }
 153                 mutex_unlock(&ctx->lut_mutex);
 154
 155                 i915_gem_context_put(lut->ctx);
 156                 i915_lut_handle_free(lut);
 157                 i915_gem_object_put(obj);
 158         }
 159 }
 160
 161 static void __i915_gem_free_object_rcu(struct rcu_head *head)
 162 {
 163         struct drm_i915_gem_object *obj =
 164                 container_of(head, typeof(*obj), rcu);
 165         struct drm_i915_private *i915 = to_i915(obj->base.dev);
 166
 167         dma_resv_fini(&obj->base._resv);
 168         i915_gem_object_free(obj);
 169
 170         GEM_BUG_ON(!atomic_read(&i915->mm.free_count));
 171         atomic_dec(&i915->mm.free_count);
 172 }
 173
 174 static void __i915_gem_object_free_mmaps(struct drm_i915_gem_object *obj)
 175 {
 176         /* Skip serialisation and waking the device if known to be not used. */
 177
 178         if (obj->userfault_count)
 179                 i915_gem_object_release_mmap_gtt(obj);
 180
 181         if (!RB_EMPTY_ROOT(&obj->mmo.offsets)) {
 182                 struct i915_mmap_offset *mmo, *mn;
 183
 184                 i915_gem_object_release_mmap_offset(obj);
 185
 186                 rbtree_postorder_for_each_entry_safe(mmo, mn,
 187                                                      &obj->mmo.offsets,
 188                                                      offset) {
 189                         drm_vma_offset_remove(obj->base.dev->vma_offset_manager,
 190                                               &mmo->vma_node);
 191                         kfree(mmo);
 192                 }
 193                 obj->mmo.offsets = RB_ROOT;
 194         }
 195 }
 196
 197 static void __i915_gem_free_objects(struct drm_i915_private *i915,
 198                                     struct llist_node *freed)
 199 {
 200         struct drm_i915_gem_object *obj, *on;
 201
 202         llist_for_each_entry_safe(obj, on, freed, freed) {
 203                 trace_i915_gem_object_destroy(obj);
 204
 205                 if (!list_empty(&obj->vma.list)) {
 206                         struct i915_vma *vma;
 207
 208                         /*
 209                          * Note that the vma keeps an object reference while
 210                          * it is active, so it *should* not sleep while we
 211                          * destroy it. Our debug code errs insits it *might*.
 212                          * For the moment, play along.
 213                          */
 214                         spin_lock(&obj->vma.lock);
 215                         while ((vma = list_first_entry_or_null(&obj->vma.list,
 216                                                                struct i915_vma,
 217                                                                obj_link))) {
 218                                 GEM_BUG_ON(vma->obj != obj);
 219                                 spin_unlock(&obj->vma.lock);
 220
 221                                 __i915_vma_put(vma);
 222
 223                                 spin_lock(&obj->vma.lock);
 224                         }
 225                         spin_unlock(&obj->vma.lock);
 226                 }
 227
 228                 __i915_gem_object_free_mmaps(obj);
 229
 230                 GEM_BUG_ON(!list_empty(&obj->lut_list));
 231
 232                 atomic_set(&obj->mm.pages_pin_count, 0);
 233                 __i915_gem_object_put_pages(obj);
 234                 GEM_BUG_ON(i915_gem_object_has_pages(obj));
 235                 bitmap_free(obj->bit_17);
 236
 237                 if (obj->base.import_attach)
 238                         drm_prime_gem_destroy(&obj->base, NULL);
 239
 240                 drm_gem_free_mmap_offset(&obj->base);
 241
 242                 if (obj->ops->release)
 243                         obj->ops->release(obj);
 244
 245                 /* But keep the pointer alive for RCU-protected lookups */
 246                 call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
 247                 cond_resched();
 248         }
 249 }
 250
 251 void i915_gem_flush_free_objects(struct drm_i915_private *i915)
 252 {
 253         struct llist_node *freed = llist_del_all(&i915->mm.free_list);
 254
 255         if (unlikely(freed))
 256                 __i915_gem_free_objects(i915, freed);
 257 }
 258
 259 static void __i915_gem_free_work(struct work_struct *work)
 260 {
 261         struct drm_i915_private *i915 =
 262                 container_of(work, struct drm_i915_private, mm.free_work);
 263
 264         i915_gem_flush_free_objects(i915);
 265 }
 266
 267 void i915_gem_free_object(struct drm_gem_object *gem_obj)
 268 {
 269         struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
 270         struct drm_i915_private *i915 = to_i915(obj->base.dev);
 271
 272         GEM_BUG_ON(i915_gem_object_is_framebuffer(obj));
 273
 274         /*
 275          * Before we free the object, make sure any pure RCU-only
 276          * read-side critical sections are complete, e.g.
 277          * i915_gem_busy_ioctl(). For the corresponding synchronized
 278          * lookup see i915_gem_object_lookup_rcu().
 279          */
 280         atomic_inc(&i915->mm.free_count);
 281
 282         /*
 283          * This serializes freeing with the shrinker. Since the free
 284          * is delayed, first by RCU then by the workqueue, we want the
 285          * shrinker to be able to free pages of unreferenced objects,
 286          * or else we may oom whilst there are plenty of deferred
 287          * freed objects.
 288          */
 289         i915_gem_object_make_unshrinkable(obj);
 290
 291         /*
 292          * Since we require blocking on struct_mutex to unbind the freed
 293          * object from the GPU before releasing resources back to the
 294          * system, we can not do that directly from the RCU callback (which may
 295          * be a softirq context), but must instead then defer that work onto a
 296          * kthread. We use the RCU callback rather than move the freed object
 297          * directly onto the work queue so that we can mix between using the
 298          * worker and performing frees directly from subsequent allocations for
 299          * crude but effective memory throttling.
 300          */
 301         if (llist_add(&obj->freed, &i915->mm.free_list))
 302                 queue_work(i915->wq, &i915->mm.free_work);
 303 }
 304
 305 static bool gpu_write_needs_clflush(struct drm_i915_gem_object *obj)
 306 {
 307         return !(obj->cache_level == I915_CACHE_NONE ||
 308                  obj->cache_level == I915_CACHE_WT);
 309 }
 310
 311 void
 312 i915_gem_object_flush_write_domain(struct drm_i915_gem_object *obj,
 313                                    unsigned int flush_domains)
 314 {
 315         struct i915_vma *vma;
 316
 317         assert_object_held(obj);
 318
 319         if (!(obj->write_domain & flush_domains))
 320                 return;
 321
 322         switch (obj->write_domain) {
 323         case I915_GEM_DOMAIN_GTT:
 324                 spin_lock(&obj->vma.lock);
 325                 for_each_ggtt_vma(vma, obj) {
 326                         if (i915_vma_unset_ggtt_write(vma))
 327                                 intel_gt_flush_ggtt_writes(vma->vm->gt);
 328                 }
 329                 spin_unlock(&obj->vma.lock);
 330
 331                 i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU);
 332                 break;
 333
 334         case I915_GEM_DOMAIN_WC:
 335                 wmb();
 336                 break;
 337
 338         case I915_GEM_DOMAIN_CPU:
 339                 i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC);
 340                 break;
 341
 342         case I915_GEM_DOMAIN_RENDER:
 343                 if (gpu_write_needs_clflush(obj))
 344                         obj->cache_dirty = true;
 345                 break;
 346         }
 347
 348         obj->write_domain = 0;
 349 }
 350
 351 void __i915_gem_object_flush_frontbuffer(struct drm_i915_gem_object *obj,
 352                                          enum fb_op_origin origin)
 353 {
 354         struct intel_frontbuffer *front;
 355
 356         front = __intel_frontbuffer_get(obj);
 357         if (front) {
 358                 intel_frontbuffer_flush(front, origin);
 359                 intel_frontbuffer_put(front);
 360         }
 361 }
 362
 363 void __i915_gem_object_invalidate_frontbuffer(struct drm_i915_gem_object *obj,
 364                                               enum fb_op_origin origin)
 365 {
 366         struct intel_frontbuffer *front;
 367
 368         front = __intel_frontbuffer_get(obj);
 369         if (front) {
 370                 intel_frontbuffer_invalidate(front, origin);
 371                 intel_frontbuffer_put(front);
 372         }
 373 }
 374
 375 void i915_gem_init__objects(struct drm_i915_private *i915)
 376 {
 377         INIT_WORK(&i915->mm.free_work, __i915_gem_free_work);
 378 }
 379
 380 static void i915_global_objects_shrink(void)
 381 {
 382         kmem_cache_shrink(global.slab_objects);
 383 }
 384
 385 static void i915_global_objects_exit(void)
 386 {
 387         kmem_cache_destroy(global.slab_objects);
 388 }
 389
 390 static struct i915_global_object global = { {
 391         .shrink = i915_global_objects_shrink,
 392         .exit = i915_global_objects_exit,
 393 } };
 394
 395 int __init i915_global_objects_init(void)
 396 {
 397         global.slab_objects =
 398                 KMEM_CACHE(drm_i915_gem_object, SLAB_HWCACHE_ALIGN);
 399         if (!global.slab_objects)
 400                 return -ENOMEM;
 401
 402         i915_global_register(&global.base);
 403         return 0;
 404 }
 405
 406 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
 407 #include "selftests/huge_gem_object.c"
 408 #include "selftests/huge_pages.c"
 409 #include "selftests/i915_gem_object.c"
 410 #include "selftests/i915_gem_coherency.c"
 411 #endif