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 shrinker_lock(struct drm_i915_private *dev_priv, bool *unlock)
  39 {
  40         switch (mutex_trylock_recursive(&dev_priv->drm.struct_mutex)) {
  41         case MUTEX_TRYLOCK_FAILED:
  42                 return false;
  43
  44         case MUTEX_TRYLOCK_SUCCESS:
  45                 *unlock = true;
  46                 return true;
  47
  48         case MUTEX_TRYLOCK_RECURSIVE:
  49                 *unlock = false;
  50                 return true;
  51         }
  52
  53         BUG();
  54 }
  55
  56 static void shrinker_unlock(struct drm_i915_private *dev_priv, bool unlock)
  57 {
  58         if (!unlock)
  59                 return;
  60
  61         mutex_unlock(&dev_priv->drm.struct_mutex);
  62
  63         /* expedite the RCU grace period to free some request slabs */
  64         synchronize_rcu_expedited();
  65 }
  66
  67 static bool any_vma_pinned(struct drm_i915_gem_object *obj)
  68 {
  69         struct i915_vma *vma;
  70
  71         list_for_each_entry(vma, &obj->vma_list, obj_link)
  72                 if (i915_vma_is_pinned(vma))
  73                         return true;
  74
  75         return false;
  76 }
  77
  78 static bool swap_available(void)
  79 {
  80         return get_nr_swap_pages() > 0;
  81 }
  82
  83 static bool can_release_pages(struct drm_i915_gem_object *obj)
  84 {
  85         if (!obj->mm.pages)
  86                 return false;
  87
  88         /* Consider only shrinkable ojects. */
  89         if (!i915_gem_object_is_shrinkable(obj))
  90                 return false;
  91
  92         /* Only report true if by unbinding the object and putting its pages
  93          * we can actually make forward progress towards freeing physical
  94          * pages.
  95          *
  96          * If the pages are pinned for any other reason than being bound
  97          * to the GPU, simply unbinding from the GPU is not going to succeed
  98          * in releasing our pin count on the pages themselves.
  99          */
 100         if (atomic_read(&obj->mm.pages_pin_count) > obj->bind_count)
 101                 return false;
 102
 103         if (any_vma_pinned(obj))
 104                 return false;
 105
 106         /* We can only return physical pages to the system if we can either
 107          * discard the contents (because the user has marked them as being
 108          * purgeable) or if we can move their contents out to swap.
 109          */
 110         return swap_available() || obj->mm.madv == I915_MADV_DONTNEED;
 111 }
 112
 113 static bool unsafe_drop_pages(struct drm_i915_gem_object *obj)
 114 {
 115         if (i915_gem_object_unbind(obj) == 0)
 116                 __i915_gem_object_put_pages(obj, I915_MM_SHRINKER);
 117         return !READ_ONCE(obj->mm.pages);
 118 }
 119
 120 /**
 121  * i915_gem_shrink - Shrink buffer object caches
 122  * @dev_priv: i915 device
 123  * @target: amount of memory to make available, in pages
 124  * @flags: control flags for selecting cache types
 125  *
 126  * This function is the main interface to the shrinker. It will try to release
 127  * up to @target pages of main memory backing storage from buffer objects.
 128  * Selection of the specific caches can be done with @flags. This is e.g. useful
 129  * when purgeable objects should be removed from caches preferentially.
 130  *
 131  * Note that it's not guaranteed that released amount is actually available as
 132  * free system memory - the pages might still be in-used to due to other reasons
 133  * (like cpu mmaps) or the mm core has reused them before we could grab them.
 134  * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
 135  * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
 136  *
 137  * Also note that any kind of pinning (both per-vma address space pins and
 138  * backing storage pins at the buffer object level) result in the shrinker code
 139  * having to skip the object.
 140  *
 141  * Returns:
 142  * The number of pages of backing storage actually released.
 143  */
 144 unsigned long
 145 i915_gem_shrink(struct drm_i915_private *dev_priv,
 146                 unsigned long target, unsigned flags)
 147 {
 148         const struct {
 149                 struct list_head *list;
 150                 unsigned int bit;
 151         } phases[] = {
 152                 { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
 153                 { &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
 154                 { NULL, 0 },
 155         }, *phase;
 156         unsigned long count = 0;
 157         bool unlock;
 158
 159         if (!shrinker_lock(dev_priv, &unlock))
 160                 return 0;
 161
 162         trace_i915_gem_shrink(dev_priv, target, flags);
 163         i915_gem_retire_requests(dev_priv);
 164
 165         /*
 166          * Unbinding of objects will require HW access; Let us not wake the
 167          * device just to recover a little memory. If absolutely necessary,
 168          * we will force the wake during oom-notifier.
 169          */
 170         if ((flags & I915_SHRINK_BOUND) &&
 171             !intel_runtime_pm_get_if_in_use(dev_priv))
 172                 flags &= ~I915_SHRINK_BOUND;
 173
 174         /*
 175          * As we may completely rewrite the (un)bound list whilst unbinding
 176          * (due to retiring requests) we have to strictly process only
 177          * one element of the list at the time, and recheck the list
 178          * on every iteration.
 179          *
 180          * In particular, we must hold a reference whilst removing the
 181          * object as we may end up waiting for and/or retiring the objects.
 182          * This might release the final reference (held by the active list)
 183          * and result in the object being freed from under us. This is
 184          * similar to the precautions the eviction code must take whilst
 185          * removing objects.
 186          *
 187          * Also note that although these lists do not hold a reference to
 188          * the object we can safely grab one here: The final object
 189          * unreferencing and the bound_list are both protected by the
 190          * dev->struct_mutex and so we won't ever be able to observe an
 191          * object on the bound_list with a reference count equals 0.
 192          */
 193         for (phase = phases; phase->list; phase++) {
 194                 struct list_head still_in_list;
 195                 struct drm_i915_gem_object *obj;
 196
 197                 if ((flags & phase->bit) == 0)
 198                         continue;
 199
 200                 INIT_LIST_HEAD(&still_in_list);
 201                 while (count < target &&
 202                        (obj = list_first_entry_or_null(phase->list,
 203                                                        typeof(*obj),
 204                                                        global_link))) {
 205                         list_move_tail(&obj->global_link, &still_in_list);
 206                         if (!obj->mm.pages) {
 207                                 list_del_init(&obj->global_link);
 208                                 continue;
 209                         }
 210
 211                         if (flags & I915_SHRINK_PURGEABLE &&
 212                             obj->mm.madv != I915_MADV_DONTNEED)
 213                                 continue;
 214
 215                         if (flags & I915_SHRINK_VMAPS &&
 216                             !is_vmalloc_addr(obj->mm.mapping))
 217                                 continue;
 218
 219                         if (!(flags & I915_SHRINK_ACTIVE) &&
 220                             (i915_gem_object_is_active(obj) ||
 221                              i915_gem_object_is_framebuffer(obj)))
 222                                 continue;
 223
 224                         if (!can_release_pages(obj))
 225                                 continue;
 226
 227                         if (unsafe_drop_pages(obj)) {
 228                                 /* May arrive from get_pages on another bo */
 229                                 mutex_lock_nested(&obj->mm.lock,
 230                                                   I915_MM_SHRINKER);
 231                                 if (!obj->mm.pages) {
 232                                         __i915_gem_object_invalidate(obj);
 233                                         list_del_init(&obj->global_link);
 234                                         count += obj->base.size >> PAGE_SHIFT;
 235                                 }
 236                                 mutex_unlock(&obj->mm.lock);
 237                         }
 238                 }
 239                 list_splice_tail(&still_in_list, phase->list);
 240         }
 241
 242         if (flags & I915_SHRINK_BOUND)
 243                 intel_runtime_pm_put(dev_priv);
 244
 245         i915_gem_retire_requests(dev_priv);
 246
 247         shrinker_unlock(dev_priv, unlock);
 248
 249         return count;
 250 }
 251
 252 /**
 253  * i915_gem_shrink_all - Shrink buffer object caches completely
 254  * @dev_priv: i915 device
 255  *
 256  * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 257  * caches completely. It also first waits for and retires all outstanding
 258  * requests to also be able to release backing storage for active objects.
 259  *
 260  * This should only be used in code to intentionally quiescent the gpu or as a
 261  * last-ditch effort when memory seems to have run out.
 262  *
 263  * Returns:
 264  * The number of pages of backing storage actually released.
 265  */
 266 unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv)
 267 {
 268         unsigned long freed;
 269
 270         intel_runtime_pm_get(dev_priv);
 271         freed = i915_gem_shrink(dev_priv, -1UL,
 272                                 I915_SHRINK_BOUND |
 273                                 I915_SHRINK_UNBOUND |
 274                                 I915_SHRINK_ACTIVE);
 275         intel_runtime_pm_put(dev_priv);
 276
 277         synchronize_rcu(); /* wait for our earlier RCU delayed slab frees */
 278
 279         return freed;
 280 }
 281
 282 static unsigned long
 283 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 284 {
 285         struct drm_i915_private *dev_priv =
 286                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 287         struct drm_i915_gem_object *obj;
 288         unsigned long count;
 289         bool unlock;
 290
 291         if (!shrinker_lock(dev_priv, &unlock))
 292                 return 0;
 293
 294         i915_gem_retire_requests(dev_priv);
 295
 296         count = 0;
 297         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_link)
 298                 if (can_release_pages(obj))
 299                         count += obj->base.size >> PAGE_SHIFT;
 300
 301         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_link) {
 302                 if (!i915_gem_object_is_active(obj) && can_release_pages(obj))
 303                         count += obj->base.size >> PAGE_SHIFT;
 304         }
 305
 306         shrinker_unlock(dev_priv, unlock);
 307
 308         return count;
 309 }
 310
 311 static unsigned long
 312 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 313 {
 314         struct drm_i915_private *dev_priv =
 315                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 316         unsigned long freed;
 317         bool unlock;
 318
 319         if (!shrinker_lock(dev_priv, &unlock))
 320                 return SHRINK_STOP;
 321
 322         freed = i915_gem_shrink(dev_priv,
 323                                 sc->nr_to_scan,
 324                                 I915_SHRINK_BOUND |
 325                                 I915_SHRINK_UNBOUND |
 326                                 I915_SHRINK_PURGEABLE);
 327         if (freed < sc->nr_to_scan)
 328                 freed += i915_gem_shrink(dev_priv,
 329                                          sc->nr_to_scan - freed,
 330                                          I915_SHRINK_BOUND |
 331                                          I915_SHRINK_UNBOUND);
 332
 333         shrinker_unlock(dev_priv, unlock);
 334
 335         return freed;
 336 }
 337
 338 static bool
 339 shrinker_lock_uninterruptible(struct drm_i915_private *dev_priv, bool *unlock,
 340                               int timeout_ms)
 341 {
 342         unsigned long timeout = jiffies + msecs_to_jiffies_timeout(timeout_ms);
 343
 344         do {
 345                 if (i915_gem_wait_for_idle(dev_priv, 0) == 0 &&
 346                     shrinker_lock(dev_priv, unlock))
 347                         break;
 348
 349                 schedule_timeout_killable(1);
 350                 if (fatal_signal_pending(current))
 351                         return false;
 352
 353                 if (time_after(jiffies, timeout)) {
 354                         pr_err("Unable to lock GPU to purge memory.\n");
 355                         return false;
 356                 }
 357         } while (1);
 358
 359         return true;
 360 }
 361
 362 static int
 363 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 364 {
 365         struct drm_i915_private *dev_priv =
 366                 container_of(nb, struct drm_i915_private, mm.oom_notifier);
 367         struct drm_i915_gem_object *obj;
 368         unsigned long unevictable, bound, unbound, freed_pages;
 369         bool unlock;
 370
 371         if (!shrinker_lock_uninterruptible(dev_priv, &unlock, 5000))
 372                 return NOTIFY_DONE;
 373
 374         freed_pages = i915_gem_shrink_all(dev_priv);
 375
 376         /* Because we may be allocating inside our own driver, we cannot
 377          * assert that there are no objects with pinned pages that are not
 378          * being pointed to by hardware.
 379          */
 380         unbound = bound = unevictable = 0;
 381         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_link) {
 382                 if (!obj->mm.pages)
 383                         continue;
 384
 385                 if (!can_release_pages(obj))
 386                         unevictable += obj->base.size >> PAGE_SHIFT;
 387                 else
 388                         unbound += obj->base.size >> PAGE_SHIFT;
 389         }
 390         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_link) {
 391                 if (!obj->mm.pages)
 392                         continue;
 393
 394                 if (!can_release_pages(obj))
 395                         unevictable += obj->base.size >> PAGE_SHIFT;
 396                 else
 397                         bound += obj->base.size >> PAGE_SHIFT;
 398         }
 399
 400         shrinker_unlock(dev_priv, unlock);
 401
 402         if (freed_pages || unbound || bound)
 403                 pr_info("Purging GPU memory, %lu pages freed, "
 404                         "%lu pages still pinned.\n",
 405                         freed_pages, unevictable);
 406         if (unbound || bound)
 407                 pr_err("%lu and %lu pages still available in the "
 408                        "bound and unbound GPU page lists.\n",
 409                        bound, unbound);
 410
 411         *(unsigned long *)ptr += freed_pages;
 412         return NOTIFY_DONE;
 413 }
 414
 415 static int
 416 i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
 417 {
 418         struct drm_i915_private *dev_priv =
 419                 container_of(nb, struct drm_i915_private, mm.vmap_notifier);
 420         struct i915_vma *vma, *next;
 421         unsigned long freed_pages = 0;
 422         bool unlock;
 423         int ret;
 424
 425         if (!shrinker_lock_uninterruptible(dev_priv, &unlock, 5000))
 426                 return NOTIFY_DONE;
 427
 428         /* Force everything onto the inactive lists */
 429         ret = i915_gem_wait_for_idle(dev_priv, I915_WAIT_LOCKED);
 430         if (ret)
 431                 goto out;
 432
 433         intel_runtime_pm_get(dev_priv);
 434         freed_pages += i915_gem_shrink(dev_priv, -1UL,
 435                                        I915_SHRINK_BOUND |
 436                                        I915_SHRINK_UNBOUND |
 437                                        I915_SHRINK_ACTIVE |
 438                                        I915_SHRINK_VMAPS);
 439         intel_runtime_pm_put(dev_priv);
 440
 441         /* We also want to clear any cached iomaps as they wrap vmap */
 442         list_for_each_entry_safe(vma, next,
 443                                  &dev_priv->ggtt.base.inactive_list, vm_link) {
 444                 unsigned long count = vma->node.size >> PAGE_SHIFT;
 445                 if (vma->iomap && i915_vma_unbind(vma) == 0)
 446                         freed_pages += count;
 447         }
 448
 449 out:
 450         shrinker_unlock(dev_priv, unlock);
 451
 452         *(unsigned long *)ptr += freed_pages;
 453         return NOTIFY_DONE;
 454 }
 455
 456 /**
 457  * i915_gem_shrinker_init - Initialize i915 shrinker
 458  * @dev_priv: i915 device
 459  *
 460  * This function registers and sets up the i915 shrinker and OOM handler.
 461  */
 462 void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
 463 {
 464         dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
 465         dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
 466         dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
 467         WARN_ON(register_shrinker(&dev_priv->mm.shrinker));
 468
 469         dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 470         WARN_ON(register_oom_notifier(&dev_priv->mm.oom_notifier));
 471
 472         dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
 473         WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
 474 }
 475
 476 /**
 477  * i915_gem_shrinker_cleanup - Clean up i915 shrinker
 478  * @dev_priv: i915 device
 479  *
 480  * This function unregisters the i915 shrinker and OOM handler.
 481  */
 482 void i915_gem_shrinker_cleanup(struct drm_i915_private *dev_priv)
 483 {
 484         WARN_ON(unregister_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
 485         WARN_ON(unregister_oom_notifier(&dev_priv->mm.oom_notifier));
 486         unregister_shrinker(&dev_priv->mm.shrinker);
 487 }