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_RECURSIVE:
  42                 *unlock = false;
  43                 return true;
  44
  45         case MUTEX_TRYLOCK_FAILED:
  46                 *unlock = false;
  47                 preempt_disable();
  48                 do {
  49                         cpu_relax();
  50                         if (mutex_trylock(&dev_priv->drm.struct_mutex)) {
  51                                 *unlock = true;
  52                                 break;
  53                         }
  54                 } while (!need_resched());
  55                 preempt_enable();
  56                 return *unlock;
  57
  58         case MUTEX_TRYLOCK_SUCCESS:
  59                 *unlock = true;
  60                 return true;
  61         }
  62
  63         BUG();
  64 }
  65
  66 static void shrinker_unlock(struct drm_i915_private *dev_priv, bool unlock)
  67 {
  68         if (!unlock)
  69                 return;
  70
  71         mutex_unlock(&dev_priv->drm.struct_mutex);
  72 }
  73
  74 static bool any_vma_pinned(struct drm_i915_gem_object *obj)
  75 {
  76         struct i915_vma *vma;
  77
  78         list_for_each_entry(vma, &obj->vma_list, obj_link) {
  79                 /* Only GGTT vma may be permanently pinned, and are always
  80                  * at the start of the list. We can stop hunting as soon
  81                  * as we see a ppGTT vma.
  82                  */
  83                 if (!i915_vma_is_ggtt(vma))
  84                         break;
  85
  86                 if (i915_vma_is_pinned(vma))
  87                         return true;
  88         }
  89
  90         return false;
  91 }
  92
  93 static bool swap_available(void)
  94 {
  95         return get_nr_swap_pages() > 0;
  96 }
  97
  98 static bool can_release_pages(struct drm_i915_gem_object *obj)
  99 {
 100         if (!obj->mm.pages)
 101                 return false;
 102
 103         /* Consider only shrinkable ojects. */
 104         if (!i915_gem_object_is_shrinkable(obj))
 105                 return false;
 106
 107         /* Only report true if by unbinding the object and putting its pages
 108          * we can actually make forward progress towards freeing physical
 109          * pages.
 110          *
 111          * If the pages are pinned for any other reason than being bound
 112          * to the GPU, simply unbinding from the GPU is not going to succeed
 113          * in releasing our pin count on the pages themselves.
 114          */
 115         if (atomic_read(&obj->mm.pages_pin_count) > obj->bind_count)
 116                 return false;
 117
 118         if (any_vma_pinned(obj))
 119                 return false;
 120
 121         /* We can only return physical pages to the system if we can either
 122          * discard the contents (because the user has marked them as being
 123          * purgeable) or if we can move their contents out to swap.
 124          */
 125         return swap_available() || obj->mm.madv == I915_MADV_DONTNEED;
 126 }
 127
 128 static bool unsafe_drop_pages(struct drm_i915_gem_object *obj)
 129 {
 130         if (i915_gem_object_unbind(obj) == 0)
 131                 __i915_gem_object_put_pages(obj, I915_MM_SHRINKER);
 132         return !READ_ONCE(obj->mm.pages);
 133 }
 134
 135 /**
 136  * i915_gem_shrink - Shrink buffer object caches
 137  * @dev_priv: i915 device
 138  * @target: amount of memory to make available, in pages
 139  * @flags: control flags for selecting cache types
 140  *
 141  * This function is the main interface to the shrinker. It will try to release
 142  * up to @target pages of main memory backing storage from buffer objects.
 143  * Selection of the specific caches can be done with @flags. This is e.g. useful
 144  * when purgeable objects should be removed from caches preferentially.
 145  *
 146  * Note that it's not guaranteed that released amount is actually available as
 147  * free system memory - the pages might still be in-used to due to other reasons
 148  * (like cpu mmaps) or the mm core has reused them before we could grab them.
 149  * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
 150  * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
 151  *
 152  * Also note that any kind of pinning (both per-vma address space pins and
 153  * backing storage pins at the buffer object level) result in the shrinker code
 154  * having to skip the object.
 155  *
 156  * Returns:
 157  * The number of pages of backing storage actually released.
 158  */
 159 unsigned long
 160 i915_gem_shrink(struct drm_i915_private *dev_priv,
 161                 unsigned long target, unsigned flags)
 162 {
 163         const struct {
 164                 struct list_head *list;
 165                 unsigned int bit;
 166         } phases[] = {
 167                 { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
 168                 { &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
 169                 { NULL, 0 },
 170         }, *phase;
 171         unsigned long count = 0;
 172         bool unlock;
 173
 174         if (!shrinker_lock(dev_priv, &unlock))
 175                 return 0;
 176
 177         trace_i915_gem_shrink(dev_priv, target, flags);
 178         i915_gem_retire_requests(dev_priv);
 179
 180         /*
 181          * Unbinding of objects will require HW access; Let us not wake the
 182          * device just to recover a little memory. If absolutely necessary,
 183          * we will force the wake during oom-notifier.
 184          */
 185         if ((flags & I915_SHRINK_BOUND) &&
 186             !intel_runtime_pm_get_if_in_use(dev_priv))
 187                 flags &= ~I915_SHRINK_BOUND;
 188
 189         /*
 190          * As we may completely rewrite the (un)bound list whilst unbinding
 191          * (due to retiring requests) we have to strictly process only
 192          * one element of the list at the time, and recheck the list
 193          * on every iteration.
 194          *
 195          * In particular, we must hold a reference whilst removing the
 196          * object as we may end up waiting for and/or retiring the objects.
 197          * This might release the final reference (held by the active list)
 198          * and result in the object being freed from under us. This is
 199          * similar to the precautions the eviction code must take whilst
 200          * removing objects.
 201          *
 202          * Also note that although these lists do not hold a reference to
 203          * the object we can safely grab one here: The final object
 204          * unreferencing and the bound_list are both protected by the
 205          * dev->struct_mutex and so we won't ever be able to observe an
 206          * object on the bound_list with a reference count equals 0.
 207          */
 208         for (phase = phases; phase->list; phase++) {
 209                 struct list_head still_in_list;
 210                 struct drm_i915_gem_object *obj;
 211
 212                 if ((flags & phase->bit) == 0)
 213                         continue;
 214
 215                 INIT_LIST_HEAD(&still_in_list);
 216                 while (count < target &&
 217                        (obj = list_first_entry_or_null(phase->list,
 218                                                        typeof(*obj),
 219                                                        global_link))) {
 220                         list_move_tail(&obj->global_link, &still_in_list);
 221                         if (!obj->mm.pages) {
 222                                 list_del_init(&obj->global_link);
 223                                 continue;
 224                         }
 225
 226                         if (flags & I915_SHRINK_PURGEABLE &&
 227                             obj->mm.madv != I915_MADV_DONTNEED)
 228                                 continue;
 229
 230                         if (flags & I915_SHRINK_VMAPS &&
 231                             !is_vmalloc_addr(obj->mm.mapping))
 232                                 continue;
 233
 234                         if (!(flags & I915_SHRINK_ACTIVE) &&
 235                             (i915_gem_object_is_active(obj) ||
 236                              i915_gem_object_is_framebuffer(obj)))
 237                                 continue;
 238
 239                         if (!can_release_pages(obj))
 240                                 continue;
 241
 242                         if (unsafe_drop_pages(obj)) {
 243                                 /* May arrive from get_pages on another bo */
 244                                 mutex_lock_nested(&obj->mm.lock,
 245                                                   I915_MM_SHRINKER);
 246                                 if (!obj->mm.pages) {
 247                                         __i915_gem_object_invalidate(obj);
 248                                         list_del_init(&obj->global_link);
 249                                         count += obj->base.size >> PAGE_SHIFT;
 250                                 }
 251                                 mutex_unlock(&obj->mm.lock);
 252                         }
 253                 }
 254                 list_splice_tail(&still_in_list, phase->list);
 255         }
 256
 257         if (flags & I915_SHRINK_BOUND)
 258                 intel_runtime_pm_put(dev_priv);
 259
 260         i915_gem_retire_requests(dev_priv);
 261
 262         shrinker_unlock(dev_priv, unlock);
 263
 264         return count;
 265 }
 266
 267 /**
 268  * i915_gem_shrink_all - Shrink buffer object caches completely
 269  * @dev_priv: i915 device
 270  *
 271  * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 272  * caches completely. It also first waits for and retires all outstanding
 273  * requests to also be able to release backing storage for active objects.
 274  *
 275  * This should only be used in code to intentionally quiescent the gpu or as a
 276  * last-ditch effort when memory seems to have run out.
 277  *
 278  * Returns:
 279  * The number of pages of backing storage actually released.
 280  */
 281 unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv)
 282 {
 283         unsigned long freed;
 284
 285         intel_runtime_pm_get(dev_priv);
 286         freed = i915_gem_shrink(dev_priv, -1UL,
 287                                 I915_SHRINK_BOUND |
 288                                 I915_SHRINK_UNBOUND |
 289                                 I915_SHRINK_ACTIVE);
 290         intel_runtime_pm_put(dev_priv);
 291
 292         return freed;
 293 }
 294
 295 static unsigned long
 296 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 297 {
 298         struct drm_i915_private *dev_priv =
 299                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 300         struct drm_i915_gem_object *obj;
 301         unsigned long count;
 302         bool unlock;
 303
 304         if (!shrinker_lock(dev_priv, &unlock))
 305                 return 0;
 306
 307         i915_gem_retire_requests(dev_priv);
 308
 309         count = 0;
 310         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_link)
 311                 if (can_release_pages(obj))
 312                         count += obj->base.size >> PAGE_SHIFT;
 313
 314         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_link) {
 315                 if (!i915_gem_object_is_active(obj) && can_release_pages(obj))
 316                         count += obj->base.size >> PAGE_SHIFT;
 317         }
 318
 319         shrinker_unlock(dev_priv, unlock);
 320
 321         return count;
 322 }
 323
 324 static unsigned long
 325 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 326 {
 327         struct drm_i915_private *dev_priv =
 328                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 329         unsigned long freed;
 330         bool unlock;
 331
 332         if (!shrinker_lock(dev_priv, &unlock))
 333                 return SHRINK_STOP;
 334
 335         freed = i915_gem_shrink(dev_priv,
 336                                 sc->nr_to_scan,
 337                                 I915_SHRINK_BOUND |
 338                                 I915_SHRINK_UNBOUND |
 339                                 I915_SHRINK_PURGEABLE);
 340         if (freed < sc->nr_to_scan)
 341                 freed += i915_gem_shrink(dev_priv,
 342                                          sc->nr_to_scan - freed,
 343                                          I915_SHRINK_BOUND |
 344                                          I915_SHRINK_UNBOUND);
 345         if (freed < sc->nr_to_scan && current_is_kswapd()) {
 346                 intel_runtime_pm_get(dev_priv);
 347                 freed += i915_gem_shrink(dev_priv,
 348                                          sc->nr_to_scan - freed,
 349                                          I915_SHRINK_ACTIVE |
 350                                          I915_SHRINK_BOUND |
 351                                          I915_SHRINK_UNBOUND);
 352                 intel_runtime_pm_put(dev_priv);
 353         }
 354
 355         shrinker_unlock(dev_priv, unlock);
 356
 357         return freed;
 358 }
 359
 360 static bool
 361 shrinker_lock_uninterruptible(struct drm_i915_private *dev_priv, bool *unlock,
 362                               int timeout_ms)
 363 {
 364         unsigned long timeout = jiffies + msecs_to_jiffies_timeout(timeout_ms);
 365
 366         do {
 367                 if (i915_gem_wait_for_idle(dev_priv, 0) == 0 &&
 368                     shrinker_lock(dev_priv, unlock))
 369                         break;
 370
 371                 schedule_timeout_killable(1);
 372                 if (fatal_signal_pending(current))
 373                         return false;
 374
 375                 if (time_after(jiffies, timeout)) {
 376                         pr_err("Unable to lock GPU to purge memory.\n");
 377                         return false;
 378                 }
 379         } while (1);
 380
 381         return true;
 382 }
 383
 384 static int
 385 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 386 {
 387         struct drm_i915_private *dev_priv =
 388                 container_of(nb, struct drm_i915_private, mm.oom_notifier);
 389         struct drm_i915_gem_object *obj;
 390         unsigned long unevictable, bound, unbound, freed_pages;
 391         bool unlock;
 392
 393         if (!shrinker_lock_uninterruptible(dev_priv, &unlock, 5000))
 394                 return NOTIFY_DONE;
 395
 396         freed_pages = i915_gem_shrink_all(dev_priv);
 397
 398         /* Because we may be allocating inside our own driver, we cannot
 399          * assert that there are no objects with pinned pages that are not
 400          * being pointed to by hardware.
 401          */
 402         unbound = bound = unevictable = 0;
 403         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_link) {
 404                 if (!obj->mm.pages)
 405                         continue;
 406
 407                 if (!can_release_pages(obj))
 408                         unevictable += obj->base.size >> PAGE_SHIFT;
 409                 else
 410                         unbound += obj->base.size >> PAGE_SHIFT;
 411         }
 412         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_link) {
 413                 if (!obj->mm.pages)
 414                         continue;
 415
 416                 if (!can_release_pages(obj))
 417                         unevictable += obj->base.size >> PAGE_SHIFT;
 418                 else
 419                         bound += obj->base.size >> PAGE_SHIFT;
 420         }
 421
 422         shrinker_unlock(dev_priv, unlock);
 423
 424         if (freed_pages || unbound || bound)
 425                 pr_info("Purging GPU memory, %lu pages freed, "
 426                         "%lu pages still pinned.\n",
 427                         freed_pages, unevictable);
 428         if (unbound || bound)
 429                 pr_err("%lu and %lu pages still available in the "
 430                        "bound and unbound GPU page lists.\n",
 431                        bound, unbound);
 432
 433         *(unsigned long *)ptr += freed_pages;
 434         return NOTIFY_DONE;
 435 }
 436
 437 static int
 438 i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
 439 {
 440         struct drm_i915_private *dev_priv =
 441                 container_of(nb, struct drm_i915_private, mm.vmap_notifier);
 442         struct i915_vma *vma, *next;
 443         unsigned long freed_pages = 0;
 444         bool unlock;
 445         int ret;
 446
 447         if (!shrinker_lock_uninterruptible(dev_priv, &unlock, 5000))
 448                 return NOTIFY_DONE;
 449
 450         /* Force everything onto the inactive lists */
 451         ret = i915_gem_wait_for_idle(dev_priv, I915_WAIT_LOCKED);
 452         if (ret)
 453                 goto out;
 454
 455         intel_runtime_pm_get(dev_priv);
 456         freed_pages += i915_gem_shrink(dev_priv, -1UL,
 457                                        I915_SHRINK_BOUND |
 458                                        I915_SHRINK_UNBOUND |
 459                                        I915_SHRINK_ACTIVE |
 460                                        I915_SHRINK_VMAPS);
 461         intel_runtime_pm_put(dev_priv);
 462
 463         /* We also want to clear any cached iomaps as they wrap vmap */
 464         list_for_each_entry_safe(vma, next,
 465                                  &dev_priv->ggtt.base.inactive_list, vm_link) {
 466                 unsigned long count = vma->node.size >> PAGE_SHIFT;
 467                 if (vma->iomap && i915_vma_unbind(vma) == 0)
 468                         freed_pages += count;
 469         }
 470
 471 out:
 472         shrinker_unlock(dev_priv, unlock);
 473
 474         *(unsigned long *)ptr += freed_pages;
 475         return NOTIFY_DONE;
 476 }
 477
 478 /**
 479  * i915_gem_shrinker_init - Initialize i915 shrinker
 480  * @dev_priv: i915 device
 481  *
 482  * This function registers and sets up the i915 shrinker and OOM handler.
 483  */
 484 void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
 485 {
 486         dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
 487         dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
 488         dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
 489         WARN_ON(register_shrinker(&dev_priv->mm.shrinker));
 490
 491         dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 492         WARN_ON(register_oom_notifier(&dev_priv->mm.oom_notifier));
 493
 494         dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
 495         WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
 496 }
 497
 498 /**
 499  * i915_gem_shrinker_cleanup - Clean up i915 shrinker
 500  * @dev_priv: i915 device
 501  *
 502  * This function unregisters the i915 shrinker and OOM handler.
 503  */
 504 void i915_gem_shrinker_cleanup(struct drm_i915_private *dev_priv)
 505 {
 506         WARN_ON(unregister_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
 507         WARN_ON(unregister_oom_notifier(&dev_priv->mm.oom_notifier));
 508         unregister_shrinker(&dev_priv->mm.shrinker);
 509 }