drivers/gpu/drm/i915/i915_gem_shrinker.c

   1 /*
   2  * Copyright © 2008-2015 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/oom.h>
  26 #include <linux/shmem_fs.h>
  27 #include <linux/slab.h>
  28 #include <linux/swap.h>
  29 #include <linux/pci.h>
  30 #include <linux/dma-buf.h>
  31 #include <linux/vmalloc.h>
  32 #include <drm/drmP.h>
  33 #include <drm/i915_drm.h>
  34
  35 #include "i915_drv.h"
  36 #include "i915_trace.h"
  37
  38 static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
  39 {
  40         if (!mutex_is_locked(mutex))
  41                 return false;
  42
  43 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES)
  44         return mutex->owner == task;
  45 #else
  46         /* Since UP may be pre-empted, we cannot assume that we own the lock */
  47         return false;
  48 #endif
  49 }
  50
  51 static int num_vma_bound(struct drm_i915_gem_object *obj)
  52 {
  53         struct i915_vma *vma;
  54         int count = 0;
  55
  56         list_for_each_entry(vma, &obj->vma_list, obj_link) {
  57                 if (drm_mm_node_allocated(&vma->node))
  58                         count++;
  59                 if (vma->pin_count)
  60                         count++;
  61         }
  62
  63         return count;
  64 }
  65
  66 static bool swap_available(void)
  67 {
  68         return get_nr_swap_pages() > 0;
  69 }
  70
  71 static bool can_release_pages(struct drm_i915_gem_object *obj)
  72 {
  73         /* Only report true if by unbinding the object and putting its pages
  74          * we can actually make forward progress towards freeing physical
  75          * pages.
  76          *
  77          * If the pages are pinned for any other reason than being bound
  78          * to the GPU, simply unbinding from the GPU is not going to succeed
  79          * in releasing our pin count on the pages themselves.
  80          */
  81         if (obj->pages_pin_count != num_vma_bound(obj))
  82                 return false;
  83
  84         /* We can only return physical pages to the system if we can either
  85          * discard the contents (because the user has marked them as being
  86          * purgeable) or if we can move their contents out to swap.
  87          */
  88         return swap_available() || obj->madv == I915_MADV_DONTNEED;
  89 }
  90
  91 /**
  92  * i915_gem_shrink - Shrink buffer object caches
  93  * @dev_priv: i915 device
  94  * @target: amount of memory to make available, in pages
  95  * @flags: control flags for selecting cache types
  96  *
  97  * This function is the main interface to the shrinker. It will try to release
  98  * up to @target pages of main memory backing storage from buffer objects.
  99  * Selection of the specific caches can be done with @flags. This is e.g. useful
 100  * when purgeable objects should be removed from caches preferentially.
 101  *
 102  * Note that it's not guaranteed that released amount is actually available as
 103  * free system memory - the pages might still be in-used to due to other reasons
 104  * (like cpu mmaps) or the mm core has reused them before we could grab them.
 105  * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
 106  * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
 107  *
 108  * Also note that any kind of pinning (both per-vma address space pins and
 109  * backing storage pins at the buffer object level) result in the shrinker code
 110  * having to skip the object.
 111  *
 112  * Returns:
 113  * The number of pages of backing storage actually released.
 114  */
 115 unsigned long
 116 i915_gem_shrink(struct drm_i915_private *dev_priv,
 117                 unsigned long target, unsigned flags)
 118 {
 119         const struct {
 120                 struct list_head *list;
 121                 unsigned int bit;
 122         } phases[] = {
 123                 { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
 124                 { &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
 125                 { NULL, 0 },
 126         }, *phase;
 127         unsigned long count = 0;
 128
 129         trace_i915_gem_shrink(dev_priv, target, flags);
 130         i915_gem_retire_requests(dev_priv->dev);
 131
 132         /*
 133          * As we may completely rewrite the (un)bound list whilst unbinding
 134          * (due to retiring requests) we have to strictly process only
 135          * one element of the list at the time, and recheck the list
 136          * on every iteration.
 137          *
 138          * In particular, we must hold a reference whilst removing the
 139          * object as we may end up waiting for and/or retiring the objects.
 140          * This might release the final reference (held by the active list)
 141          * and result in the object being freed from under us. This is
 142          * similar to the precautions the eviction code must take whilst
 143          * removing objects.
 144          *
 145          * Also note that although these lists do not hold a reference to
 146          * the object we can safely grab one here: The final object
 147          * unreferencing and the bound_list are both protected by the
 148          * dev->struct_mutex and so we won't ever be able to observe an
 149          * object on the bound_list with a reference count equals 0.
 150          */
 151         for (phase = phases; phase->list; phase++) {
 152                 struct list_head still_in_list;
 153
 154                 if ((flags & phase->bit) == 0)
 155                         continue;
 156
 157                 INIT_LIST_HEAD(&still_in_list);
 158                 while (count < target && !list_empty(phase->list)) {
 159                         struct drm_i915_gem_object *obj;
 160                         struct i915_vma *vma, *v;
 161
 162                         obj = list_first_entry(phase->list,
 163                                                typeof(*obj), global_list);
 164                         list_move_tail(&obj->global_list, &still_in_list);
 165
 166                         if (flags & I915_SHRINK_PURGEABLE &&
 167                             obj->madv != I915_MADV_DONTNEED)
 168                                 continue;
 169
 170                         if ((flags & I915_SHRINK_ACTIVE) == 0 && obj->active)
 171                                 continue;
 172
 173                         if (!can_release_pages(obj))
 174                                 continue;
 175
 176                         drm_gem_object_reference(&obj->base);
 177
 178                         /* For the unbound phase, this should be a no-op! */
 179                         list_for_each_entry_safe(vma, v,
 180                                                  &obj->vma_list, obj_link)
 181                                 if (i915_vma_unbind(vma))
 182                                         break;
 183
 184                         if (i915_gem_object_put_pages(obj) == 0)
 185                                 count += obj->base.size >> PAGE_SHIFT;
 186
 187                         drm_gem_object_unreference(&obj->base);
 188                 }
 189                 list_splice(&still_in_list, phase->list);
 190         }
 191
 192         i915_gem_retire_requests(dev_priv->dev);
 193
 194         return count;
 195 }
 196
 197 /**
 198  * i915_gem_shrink_all - Shrink buffer object caches completely
 199  * @dev_priv: i915 device
 200  *
 201  * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 202  * caches completely. It also first waits for and retires all outstanding
 203  * requests to also be able to release backing storage for active objects.
 204  *
 205  * This should only be used in code to intentionally quiescent the gpu or as a
 206  * last-ditch effort when memory seems to have run out.
 207  *
 208  * Returns:
 209  * The number of pages of backing storage actually released.
 210  */
 211 unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv)
 212 {
 213         return i915_gem_shrink(dev_priv, -1UL,
 214                                I915_SHRINK_BOUND |
 215                                I915_SHRINK_UNBOUND |
 216                                I915_SHRINK_ACTIVE);
 217 }
 218
 219 static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock)
 220 {
 221         if (!mutex_trylock(&dev->struct_mutex)) {
 222                 if (!mutex_is_locked_by(&dev->struct_mutex, current))
 223                         return false;
 224
 225                 if (to_i915(dev)->mm.shrinker_no_lock_stealing)
 226                         return false;
 227
 228                 *unlock = false;
 229         } else
 230                 *unlock = true;
 231
 232         return true;
 233 }
 234
 235 static unsigned long
 236 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 237 {
 238         struct drm_i915_private *dev_priv =
 239                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 240         struct drm_device *dev = dev_priv->dev;
 241         struct drm_i915_gem_object *obj;
 242         unsigned long count;
 243         bool unlock;
 244
 245         if (!i915_gem_shrinker_lock(dev, &unlock))
 246                 return 0;
 247
 248         count = 0;
 249         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list)
 250                 if (can_release_pages(obj))
 251                         count += obj->base.size >> PAGE_SHIFT;
 252
 253         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
 254                 if (!obj->active && can_release_pages(obj))
 255                         count += obj->base.size >> PAGE_SHIFT;
 256         }
 257
 258         if (unlock)
 259                 mutex_unlock(&dev->struct_mutex);
 260
 261         return count;
 262 }
 263
 264 static unsigned long
 265 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 266 {
 267         struct drm_i915_private *dev_priv =
 268                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 269         struct drm_device *dev = dev_priv->dev;
 270         unsigned long freed;
 271         bool unlock;
 272
 273         if (!i915_gem_shrinker_lock(dev, &unlock))
 274                 return SHRINK_STOP;
 275
 276         freed = i915_gem_shrink(dev_priv,
 277                                 sc->nr_to_scan,
 278                                 I915_SHRINK_BOUND |
 279                                 I915_SHRINK_UNBOUND |
 280                                 I915_SHRINK_PURGEABLE);
 281         if (freed < sc->nr_to_scan)
 282                 freed += i915_gem_shrink(dev_priv,
 283                                          sc->nr_to_scan - freed,
 284                                          I915_SHRINK_BOUND |
 285                                          I915_SHRINK_UNBOUND);
 286         if (unlock)
 287                 mutex_unlock(&dev->struct_mutex);
 288
 289         return freed;
 290 }
 291
 292 struct shrinker_lock_uninterruptible {
 293         bool was_interruptible;
 294         bool unlock;
 295 };
 296
 297 static bool
 298 i915_gem_shrinker_lock_uninterruptible(struct drm_i915_private *dev_priv,
 299                                        struct shrinker_lock_uninterruptible *slu,
 300                                        int timeout_ms)
 301 {
 302         unsigned long timeout = msecs_to_jiffies(timeout_ms) + 1;
 303
 304         while (!i915_gem_shrinker_lock(dev_priv->dev, &slu->unlock)) {
 305                 schedule_timeout_killable(1);
 306                 if (fatal_signal_pending(current))
 307                         return false;
 308                 if (--timeout == 0) {
 309                         pr_err("Unable to lock GPU to purge memory.\n");
 310                         return false;
 311                 }
 312         }
 313
 314         slu->was_interruptible = dev_priv->mm.interruptible;
 315         dev_priv->mm.interruptible = false;
 316         return true;
 317 }
 318
 319 static void
 320 i915_gem_shrinker_unlock_uninterruptible(struct drm_i915_private *dev_priv,
 321                                          struct shrinker_lock_uninterruptible *slu)
 322 {
 323         dev_priv->mm.interruptible = slu->was_interruptible;
 324         if (slu->unlock)
 325                 mutex_unlock(&dev_priv->dev->struct_mutex);
 326 }
 327
 328 static int
 329 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 330 {
 331         struct drm_i915_private *dev_priv =
 332                 container_of(nb, struct drm_i915_private, mm.oom_notifier);
 333         struct shrinker_lock_uninterruptible slu;
 334         struct drm_i915_gem_object *obj;
 335         unsigned long pinned, bound, unbound, freed_pages;
 336
 337         if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000))
 338                 return NOTIFY_DONE;
 339
 340         freed_pages = i915_gem_shrink_all(dev_priv);
 341
 342         /* Because we may be allocating inside our own driver, we cannot
 343          * assert that there are no objects with pinned pages that are not
 344          * being pointed to by hardware.
 345          */
 346         unbound = bound = pinned = 0;
 347         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) {
 348                 if (!obj->base.filp) /* not backed by a freeable object */
 349                         continue;
 350
 351                 if (obj->pages_pin_count)
 352                         pinned += obj->base.size;
 353                 else
 354                         unbound += obj->base.size;
 355         }
 356         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
 357                 if (!obj->base.filp)
 358                         continue;
 359
 360                 if (obj->pages_pin_count)
 361                         pinned += obj->base.size;
 362                 else
 363                         bound += obj->base.size;
 364         }
 365
 366         i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
 367
 368         if (freed_pages || unbound || bound)
 369                 pr_info("Purging GPU memory, %lu bytes freed, %lu bytes still pinned.\n",
 370                         freed_pages << PAGE_SHIFT, pinned);
 371         if (unbound || bound)
 372                 pr_err("%lu and %lu bytes still available in the "
 373                        "bound and unbound GPU page lists.\n",
 374                        bound, unbound);
 375
 376         *(unsigned long *)ptr += freed_pages;
 377         return NOTIFY_DONE;
 378 }
 379
 380 static int
 381 i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
 382 {
 383         struct drm_i915_private *dev_priv =
 384                 container_of(nb, struct drm_i915_private, mm.vmap_notifier);
 385         struct shrinker_lock_uninterruptible slu;
 386         unsigned long freed_pages;
 387
 388         if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000))
 389                 return NOTIFY_DONE;
 390
 391         freed_pages = i915_gem_shrink_all(dev_priv);
 392
 393         i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
 394
 395         *(unsigned long *)ptr += freed_pages;
 396         return NOTIFY_DONE;
 397 }
 398
 399 /**
 400  * i915_gem_shrinker_init - Initialize i915 shrinker
 401  * @dev_priv: i915 device
 402  *
 403  * This function registers and sets up the i915 shrinker and OOM handler.
 404  */
 405 void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
 406 {
 407         dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
 408         dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
 409         dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
 410         WARN_ON(register_shrinker(&dev_priv->mm.shrinker));
 411
 412         dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 413         WARN_ON(register_oom_notifier(&dev_priv->mm.oom_notifier));
 414
 415         dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
 416         WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
 417 }
 418
 419 /**
 420  * i915_gem_shrinker_cleanup - Clean up i915 shrinker
 421  * @dev_priv: i915 device
 422  *
 423  * This function unregisters the i915 shrinker and OOM handler.
 424  */
 425 void i915_gem_shrinker_cleanup(struct drm_i915_private *dev_priv)
 426 {
 427         WARN_ON(unregister_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
 428         WARN_ON(unregister_oom_notifier(&dev_priv->mm.oom_notifier));
 429         unregister_shrinker(&dev_priv->mm.shrinker);
 430 }