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 i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock)
  39 {
  40         switch (mutex_trylock_recursive(&dev->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 bool any_vma_pinned(struct drm_i915_gem_object *obj)
  57 {
  58         struct i915_vma *vma;
  59
  60         list_for_each_entry(vma, &obj->vma_list, obj_link)
  61                 if (i915_vma_is_pinned(vma))
  62                         return true;
  63
  64         return false;
  65 }
  66
  67 static bool swap_available(void)
  68 {
  69         return get_nr_swap_pages() > 0;
  70 }
  71
  72 static bool can_release_pages(struct drm_i915_gem_object *obj)
  73 {
  74         if (!obj->mm.pages)
  75                 return false;
  76
  77         /* Consider only shrinkable ojects. */
  78         if (!i915_gem_object_is_shrinkable(obj))
  79                 return false;
  80
  81         /* Only report true if by unbinding the object and putting its pages
  82          * we can actually make forward progress towards freeing physical
  83          * pages.
  84          *
  85          * If the pages are pinned for any other reason than being bound
  86          * to the GPU, simply unbinding from the GPU is not going to succeed
  87          * in releasing our pin count on the pages themselves.
  88          */
  89         if (atomic_read(&obj->mm.pages_pin_count) > obj->bind_count)
  90                 return false;
  91
  92         if (any_vma_pinned(obj))
  93                 return false;
  94
  95         /* We can only return physical pages to the system if we can either
  96          * discard the contents (because the user has marked them as being
  97          * purgeable) or if we can move their contents out to swap.
  98          */
  99         return swap_available() || obj->mm.madv == I915_MADV_DONTNEED;
 100 }
 101
 102 static bool unsafe_drop_pages(struct drm_i915_gem_object *obj)
 103 {
 104         if (i915_gem_object_unbind(obj) == 0)
 105                 __i915_gem_object_put_pages(obj, I915_MM_SHRINKER);
 106         return !READ_ONCE(obj->mm.pages);
 107 }
 108
 109 /**
 110  * i915_gem_shrink - Shrink buffer object caches
 111  * @dev_priv: i915 device
 112  * @target: amount of memory to make available, in pages
 113  * @flags: control flags for selecting cache types
 114  *
 115  * This function is the main interface to the shrinker. It will try to release
 116  * up to @target pages of main memory backing storage from buffer objects.
 117  * Selection of the specific caches can be done with @flags. This is e.g. useful
 118  * when purgeable objects should be removed from caches preferentially.
 119  *
 120  * Note that it's not guaranteed that released amount is actually available as
 121  * free system memory - the pages might still be in-used to due to other reasons
 122  * (like cpu mmaps) or the mm core has reused them before we could grab them.
 123  * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
 124  * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
 125  *
 126  * Also note that any kind of pinning (both per-vma address space pins and
 127  * backing storage pins at the buffer object level) result in the shrinker code
 128  * having to skip the object.
 129  *
 130  * Returns:
 131  * The number of pages of backing storage actually released.
 132  */
 133 unsigned long
 134 i915_gem_shrink(struct drm_i915_private *dev_priv,
 135                 unsigned long target, unsigned flags)
 136 {
 137         const struct {
 138                 struct list_head *list;
 139                 unsigned int bit;
 140         } phases[] = {
 141                 { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
 142                 { &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
 143                 { NULL, 0 },
 144         }, *phase;
 145         unsigned long count = 0;
 146         bool unlock;
 147
 148         if (!i915_gem_shrinker_lock(&dev_priv->drm, &unlock))
 149                 return 0;
 150
 151         trace_i915_gem_shrink(dev_priv, target, flags);
 152         i915_gem_retire_requests(dev_priv);
 153
 154         /*
 155          * Unbinding of objects will require HW access; Let us not wake the
 156          * device just to recover a little memory. If absolutely necessary,
 157          * we will force the wake during oom-notifier.
 158          */
 159         if ((flags & I915_SHRINK_BOUND) &&
 160             !intel_runtime_pm_get_if_in_use(dev_priv))
 161                 flags &= ~I915_SHRINK_BOUND;
 162
 163         /*
 164          * As we may completely rewrite the (un)bound list whilst unbinding
 165          * (due to retiring requests) we have to strictly process only
 166          * one element of the list at the time, and recheck the list
 167          * on every iteration.
 168          *
 169          * In particular, we must hold a reference whilst removing the
 170          * object as we may end up waiting for and/or retiring the objects.
 171          * This might release the final reference (held by the active list)
 172          * and result in the object being freed from under us. This is
 173          * similar to the precautions the eviction code must take whilst
 174          * removing objects.
 175          *
 176          * Also note that although these lists do not hold a reference to
 177          * the object we can safely grab one here: The final object
 178          * unreferencing and the bound_list are both protected by the
 179          * dev->struct_mutex and so we won't ever be able to observe an
 180          * object on the bound_list with a reference count equals 0.
 181          */
 182         for (phase = phases; phase->list; phase++) {
 183                 struct list_head still_in_list;
 184                 struct drm_i915_gem_object *obj;
 185
 186                 if ((flags & phase->bit) == 0)
 187                         continue;
 188
 189                 INIT_LIST_HEAD(&still_in_list);
 190                 while (count < target &&
 191                        (obj = list_first_entry_or_null(phase->list,
 192                                                        typeof(*obj),
 193                                                        global_link))) {
 194                         list_move_tail(&obj->global_link, &still_in_list);
 195                         if (!obj->mm.pages) {
 196                                 list_del_init(&obj->global_link);
 197                                 continue;
 198                         }
 199
 200                         if (flags & I915_SHRINK_PURGEABLE &&
 201                             obj->mm.madv != I915_MADV_DONTNEED)
 202                                 continue;
 203
 204                         if (flags & I915_SHRINK_VMAPS &&
 205                             !is_vmalloc_addr(obj->mm.mapping))
 206                                 continue;
 207
 208                         if (!(flags & I915_SHRINK_ACTIVE) &&
 209                             (i915_gem_object_is_active(obj) ||
 210                              i915_gem_object_is_framebuffer(obj)))
 211                                 continue;
 212
 213                         if (!can_release_pages(obj))
 214                                 continue;
 215
 216                         if (unsafe_drop_pages(obj)) {
 217                                 /* May arrive from get_pages on another bo */
 218                                 mutex_lock_nested(&obj->mm.lock,
 219                                                   I915_MM_SHRINKER);
 220                                 if (!obj->mm.pages) {
 221                                         __i915_gem_object_invalidate(obj);
 222                                         list_del_init(&obj->global_link);
 223                                         count += obj->base.size >> PAGE_SHIFT;
 224                                 }
 225                                 mutex_unlock(&obj->mm.lock);
 226                         }
 227                 }
 228                 list_splice_tail(&still_in_list, phase->list);
 229         }
 230
 231         if (flags & I915_SHRINK_BOUND)
 232                 intel_runtime_pm_put(dev_priv);
 233
 234         i915_gem_retire_requests(dev_priv);
 235         if (unlock)
 236                 mutex_unlock(&dev_priv->drm.struct_mutex);
 237
 238         /* expedite the RCU grace period to free some request slabs */
 239         synchronize_rcu_expedited();
 240
 241         return count;
 242 }
 243
 244 /**
 245  * i915_gem_shrink_all - Shrink buffer object caches completely
 246  * @dev_priv: i915 device
 247  *
 248  * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 249  * caches completely. It also first waits for and retires all outstanding
 250  * requests to also be able to release backing storage for active objects.
 251  *
 252  * This should only be used in code to intentionally quiescent the gpu or as a
 253  * last-ditch effort when memory seems to have run out.
 254  *
 255  * Returns:
 256  * The number of pages of backing storage actually released.
 257  */
 258 unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv)
 259 {
 260         unsigned long freed;
 261
 262         intel_runtime_pm_get(dev_priv);
 263         freed = i915_gem_shrink(dev_priv, -1UL,
 264                                 I915_SHRINK_BOUND |
 265                                 I915_SHRINK_UNBOUND |
 266                                 I915_SHRINK_ACTIVE);
 267         intel_runtime_pm_put(dev_priv);
 268
 269         rcu_barrier(); /* wait until our RCU delayed slab frees are completed */
 270
 271         return freed;
 272 }
 273
 274 static unsigned long
 275 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 276 {
 277         struct drm_i915_private *dev_priv =
 278                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 279         struct drm_device *dev = &dev_priv->drm;
 280         struct drm_i915_gem_object *obj;
 281         unsigned long count;
 282         bool unlock;
 283
 284         if (!i915_gem_shrinker_lock(dev, &unlock))
 285                 return 0;
 286
 287         i915_gem_retire_requests(dev_priv);
 288
 289         count = 0;
 290         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_link)
 291                 if (can_release_pages(obj))
 292                         count += obj->base.size >> PAGE_SHIFT;
 293
 294         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_link) {
 295                 if (!i915_gem_object_is_active(obj) && can_release_pages(obj))
 296                         count += obj->base.size >> PAGE_SHIFT;
 297         }
 298
 299         if (unlock)
 300                 mutex_unlock(&dev->struct_mutex);
 301
 302         return count;
 303 }
 304
 305 static unsigned long
 306 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 307 {
 308         struct drm_i915_private *dev_priv =
 309                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 310         struct drm_device *dev = &dev_priv->drm;
 311         unsigned long freed;
 312         bool unlock;
 313
 314         if (!i915_gem_shrinker_lock(dev, &unlock))
 315                 return SHRINK_STOP;
 316
 317         freed = i915_gem_shrink(dev_priv,
 318                                 sc->nr_to_scan,
 319                                 I915_SHRINK_BOUND |
 320                                 I915_SHRINK_UNBOUND |
 321                                 I915_SHRINK_PURGEABLE);
 322         if (freed < sc->nr_to_scan)
 323                 freed += i915_gem_shrink(dev_priv,
 324                                          sc->nr_to_scan - freed,
 325                                          I915_SHRINK_BOUND |
 326                                          I915_SHRINK_UNBOUND);
 327         if (unlock)
 328                 mutex_unlock(&dev->struct_mutex);
 329
 330         return freed;
 331 }
 332
 333 struct shrinker_lock_uninterruptible {
 334         bool was_interruptible;
 335         bool unlock;
 336 };
 337
 338 static bool
 339 i915_gem_shrinker_lock_uninterruptible(struct drm_i915_private *dev_priv,
 340                                        struct shrinker_lock_uninterruptible *slu,
 341                                        int timeout_ms)
 342 {
 343         unsigned long timeout = jiffies + msecs_to_jiffies_timeout(timeout_ms);
 344
 345         do {
 346                 if (i915_gem_wait_for_idle(dev_priv, 0) == 0 &&
 347                     i915_gem_shrinker_lock(&dev_priv->drm, &slu->unlock))
 348                         break;
 349
 350                 schedule_timeout_killable(1);
 351                 if (fatal_signal_pending(current))
 352                         return false;
 353
 354                 if (time_after(jiffies, timeout)) {
 355                         pr_err("Unable to lock GPU to purge memory.\n");
 356                         return false;
 357                 }
 358         } while (1);
 359
 360         slu->was_interruptible = dev_priv->mm.interruptible;
 361         dev_priv->mm.interruptible = false;
 362         return true;
 363 }
 364
 365 static void
 366 i915_gem_shrinker_unlock_uninterruptible(struct drm_i915_private *dev_priv,
 367                                          struct shrinker_lock_uninterruptible *slu)
 368 {
 369         dev_priv->mm.interruptible = slu->was_interruptible;
 370         if (slu->unlock)
 371                 mutex_unlock(&dev_priv->drm.struct_mutex);
 372 }
 373
 374 static int
 375 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 376 {
 377         struct drm_i915_private *dev_priv =
 378                 container_of(nb, struct drm_i915_private, mm.oom_notifier);
 379         struct shrinker_lock_uninterruptible slu;
 380         struct drm_i915_gem_object *obj;
 381         unsigned long unevictable, bound, unbound, freed_pages;
 382
 383         if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000))
 384                 return NOTIFY_DONE;
 385
 386         freed_pages = i915_gem_shrink_all(dev_priv);
 387
 388         /* Because we may be allocating inside our own driver, we cannot
 389          * assert that there are no objects with pinned pages that are not
 390          * being pointed to by hardware.
 391          */
 392         unbound = bound = unevictable = 0;
 393         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_link) {
 394                 if (!obj->mm.pages)
 395                         continue;
 396
 397                 if (!can_release_pages(obj))
 398                         unevictable += obj->base.size >> PAGE_SHIFT;
 399                 else
 400                         unbound += obj->base.size >> PAGE_SHIFT;
 401         }
 402         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_link) {
 403                 if (!obj->mm.pages)
 404                         continue;
 405
 406                 if (!can_release_pages(obj))
 407                         unevictable += obj->base.size >> PAGE_SHIFT;
 408                 else
 409                         bound += obj->base.size >> PAGE_SHIFT;
 410         }
 411
 412         i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
 413
 414         if (freed_pages || unbound || bound)
 415                 pr_info("Purging GPU memory, %lu pages freed, "
 416                         "%lu pages still pinned.\n",
 417                         freed_pages, unevictable);
 418         if (unbound || bound)
 419                 pr_err("%lu and %lu pages still available in the "
 420                        "bound and unbound GPU page lists.\n",
 421                        bound, unbound);
 422
 423         *(unsigned long *)ptr += freed_pages;
 424         return NOTIFY_DONE;
 425 }
 426
 427 static int
 428 i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
 429 {
 430         struct drm_i915_private *dev_priv =
 431                 container_of(nb, struct drm_i915_private, mm.vmap_notifier);
 432         struct shrinker_lock_uninterruptible slu;
 433         struct i915_vma *vma, *next;
 434         unsigned long freed_pages = 0;
 435         int ret;
 436
 437         if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000))
 438                 return NOTIFY_DONE;
 439
 440         /* Force everything onto the inactive lists */
 441         ret = i915_gem_wait_for_idle(dev_priv, I915_WAIT_LOCKED);
 442         if (ret)
 443                 goto out;
 444
 445         intel_runtime_pm_get(dev_priv);
 446         freed_pages += i915_gem_shrink(dev_priv, -1UL,
 447                                        I915_SHRINK_BOUND |
 448                                        I915_SHRINK_UNBOUND |
 449                                        I915_SHRINK_ACTIVE |
 450                                        I915_SHRINK_VMAPS);
 451         intel_runtime_pm_put(dev_priv);
 452
 453         /* We also want to clear any cached iomaps as they wrap vmap */
 454         list_for_each_entry_safe(vma, next,
 455                                  &dev_priv->ggtt.base.inactive_list, vm_link) {
 456                 unsigned long count = vma->node.size >> PAGE_SHIFT;
 457                 if (vma->iomap && i915_vma_unbind(vma) == 0)
 458                         freed_pages += count;
 459         }
 460
 461 out:
 462         i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
 463
 464         *(unsigned long *)ptr += freed_pages;
 465         return NOTIFY_DONE;
 466 }
 467
 468 /**
 469  * i915_gem_shrinker_init - Initialize i915 shrinker
 470  * @dev_priv: i915 device
 471  *
 472  * This function registers and sets up the i915 shrinker and OOM handler.
 473  */
 474 void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
 475 {
 476         dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
 477         dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
 478         dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
 479         WARN_ON(register_shrinker(&dev_priv->mm.shrinker));
 480
 481         dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 482         WARN_ON(register_oom_notifier(&dev_priv->mm.oom_notifier));
 483
 484         dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
 485         WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
 486 }
 487
 488 /**
 489  * i915_gem_shrinker_cleanup - Clean up i915 shrinker
 490  * @dev_priv: i915 device
 491  *
 492  * This function unregisters the i915 shrinker and OOM handler.
 493  */
 494 void i915_gem_shrinker_cleanup(struct drm_i915_private *dev_priv)
 495 {
 496         WARN_ON(unregister_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
 497         WARN_ON(unregister_oom_notifier(&dev_priv->mm.oom_notifier));
 498         unregister_shrinker(&dev_priv->mm.shrinker);
 499 }