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
  64 static bool any_vma_pinned(struct drm_i915_gem_object *obj)
  65 {
  66         struct i915_vma *vma;
  67
  68         list_for_each_entry(vma, &obj->vma_list, obj_link) {
  69                 /* Only GGTT vma may be permanently pinned, and are always
  70                  * at the start of the list. We can stop hunting as soon
  71                  * as we see a ppGTT vma.
  72                  */
  73                 if (!i915_vma_is_ggtt(vma))
  74                         break;
  75
  76                 if (i915_vma_is_pinned(vma))
  77                         return true;
  78         }
  79
  80         return false;
  81 }
  82
  83 static bool swap_available(void)
  84 {
  85         return get_nr_swap_pages() > 0;
  86 }
  87
  88 static bool can_release_pages(struct drm_i915_gem_object *obj)
  89 {
  90         if (!obj->mm.pages)
  91                 return false;
  92
  93         /* Consider only shrinkable ojects. */
  94         if (!i915_gem_object_is_shrinkable(obj))
  95                 return false;
  96
  97         /* Only report true if by unbinding the object and putting its pages
  98          * we can actually make forward progress towards freeing physical
  99          * pages.
 100          *
 101          * If the pages are pinned for any other reason than being bound
 102          * to the GPU, simply unbinding from the GPU is not going to succeed
 103          * in releasing our pin count on the pages themselves.
 104          */
 105         if (atomic_read(&obj->mm.pages_pin_count) > obj->bind_count)
 106                 return false;
 107
 108         if (any_vma_pinned(obj))
 109                 return false;
 110
 111         /* We can only return physical pages to the system if we can either
 112          * discard the contents (because the user has marked them as being
 113          * purgeable) or if we can move their contents out to swap.
 114          */
 115         return swap_available() || obj->mm.madv == I915_MADV_DONTNEED;
 116 }
 117
 118 static bool unsafe_drop_pages(struct drm_i915_gem_object *obj)
 119 {
 120         if (i915_gem_object_unbind(obj) == 0)
 121                 __i915_gem_object_put_pages(obj, I915_MM_SHRINKER);
 122         return !READ_ONCE(obj->mm.pages);
 123 }
 124
 125 /**
 126  * i915_gem_shrink - Shrink buffer object caches
 127  * @dev_priv: i915 device
 128  * @target: amount of memory to make available, in pages
 129  * @flags: control flags for selecting cache types
 130  *
 131  * This function is the main interface to the shrinker. It will try to release
 132  * up to @target pages of main memory backing storage from buffer objects.
 133  * Selection of the specific caches can be done with @flags. This is e.g. useful
 134  * when purgeable objects should be removed from caches preferentially.
 135  *
 136  * Note that it's not guaranteed that released amount is actually available as
 137  * free system memory - the pages might still be in-used to due to other reasons
 138  * (like cpu mmaps) or the mm core has reused them before we could grab them.
 139  * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
 140  * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
 141  *
 142  * Also note that any kind of pinning (both per-vma address space pins and
 143  * backing storage pins at the buffer object level) result in the shrinker code
 144  * having to skip the object.
 145  *
 146  * Returns:
 147  * The number of pages of backing storage actually released.
 148  */
 149 unsigned long
 150 i915_gem_shrink(struct drm_i915_private *dev_priv,
 151                 unsigned long target, unsigned flags)
 152 {
 153         const struct {
 154                 struct list_head *list;
 155                 unsigned int bit;
 156         } phases[] = {
 157                 { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
 158                 { &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
 159                 { NULL, 0 },
 160         }, *phase;
 161         unsigned long count = 0;
 162         bool unlock;
 163
 164         if (!shrinker_lock(dev_priv, &unlock))
 165                 return 0;
 166
 167         trace_i915_gem_shrink(dev_priv, target, flags);
 168         i915_gem_retire_requests(dev_priv);
 169
 170         /*
 171          * Unbinding of objects will require HW access; Let us not wake the
 172          * device just to recover a little memory. If absolutely necessary,
 173          * we will force the wake during oom-notifier.
 174          */
 175         if ((flags & I915_SHRINK_BOUND) &&
 176             !intel_runtime_pm_get_if_in_use(dev_priv))
 177                 flags &= ~I915_SHRINK_BOUND;
 178
 179         /*
 180          * As we may completely rewrite the (un)bound list whilst unbinding
 181          * (due to retiring requests) we have to strictly process only
 182          * one element of the list at the time, and recheck the list
 183          * on every iteration.
 184          *
 185          * In particular, we must hold a reference whilst removing the
 186          * object as we may end up waiting for and/or retiring the objects.
 187          * This might release the final reference (held by the active list)
 188          * and result in the object being freed from under us. This is
 189          * similar to the precautions the eviction code must take whilst
 190          * removing objects.
 191          *
 192          * Also note that although these lists do not hold a reference to
 193          * the object we can safely grab one here: The final object
 194          * unreferencing and the bound_list are both protected by the
 195          * dev->struct_mutex and so we won't ever be able to observe an
 196          * object on the bound_list with a reference count equals 0.
 197          */
 198         for (phase = phases; phase->list; phase++) {
 199                 struct list_head still_in_list;
 200                 struct drm_i915_gem_object *obj;
 201
 202                 if ((flags & phase->bit) == 0)
 203                         continue;
 204
 205                 INIT_LIST_HEAD(&still_in_list);
 206                 while (count < target &&
 207                        (obj = list_first_entry_or_null(phase->list,
 208                                                        typeof(*obj),
 209                                                        global_link))) {
 210                         list_move_tail(&obj->global_link, &still_in_list);
 211                         if (!obj->mm.pages) {
 212                                 list_del_init(&obj->global_link);
 213                                 continue;
 214                         }
 215
 216                         if (flags & I915_SHRINK_PURGEABLE &&
 217                             obj->mm.madv != I915_MADV_DONTNEED)
 218                                 continue;
 219
 220                         if (flags & I915_SHRINK_VMAPS &&
 221                             !is_vmalloc_addr(obj->mm.mapping))
 222                                 continue;
 223
 224                         if (!(flags & I915_SHRINK_ACTIVE) &&
 225                             (i915_gem_object_is_active(obj) ||
 226                              i915_gem_object_is_framebuffer(obj)))
 227                                 continue;
 228
 229                         if (!can_release_pages(obj))
 230                                 continue;
 231
 232                         if (unsafe_drop_pages(obj)) {
 233                                 /* May arrive from get_pages on another bo */
 234                                 mutex_lock_nested(&obj->mm.lock,
 235                                                   I915_MM_SHRINKER);
 236                                 if (!obj->mm.pages) {
 237                                         __i915_gem_object_invalidate(obj);
 238                                         list_del_init(&obj->global_link);
 239                                         count += obj->base.size >> PAGE_SHIFT;
 240                                 }
 241                                 mutex_unlock(&obj->mm.lock);
 242                         }
 243                 }
 244                 list_splice_tail(&still_in_list, phase->list);
 245         }
 246
 247         if (flags & I915_SHRINK_BOUND)
 248                 intel_runtime_pm_put(dev_priv);
 249
 250         i915_gem_retire_requests(dev_priv);
 251
 252         shrinker_unlock(dev_priv, unlock);
 253
 254         return count;
 255 }
 256
 257 /**
 258  * i915_gem_shrink_all - Shrink buffer object caches completely
 259  * @dev_priv: i915 device
 260  *
 261  * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 262  * caches completely. It also first waits for and retires all outstanding
 263  * requests to also be able to release backing storage for active objects.
 264  *
 265  * This should only be used in code to intentionally quiescent the gpu or as a
 266  * last-ditch effort when memory seems to have run out.
 267  *
 268  * Returns:
 269  * The number of pages of backing storage actually released.
 270  */
 271 unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv)
 272 {
 273         unsigned long freed;
 274
 275         intel_runtime_pm_get(dev_priv);
 276         freed = i915_gem_shrink(dev_priv, -1UL,
 277                                 I915_SHRINK_BOUND |
 278                                 I915_SHRINK_UNBOUND |
 279                                 I915_SHRINK_ACTIVE);
 280         intel_runtime_pm_put(dev_priv);
 281
 282         return freed;
 283 }
 284
 285 static unsigned long
 286 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 287 {
 288         struct drm_i915_private *dev_priv =
 289                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 290         struct drm_i915_gem_object *obj;
 291         unsigned long count;
 292         bool unlock;
 293
 294         if (!shrinker_lock(dev_priv, &unlock))
 295                 return 0;
 296
 297         i915_gem_retire_requests(dev_priv);
 298
 299         count = 0;
 300         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_link)
 301                 if (can_release_pages(obj))
 302                         count += obj->base.size >> PAGE_SHIFT;
 303
 304         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_link) {
 305                 if (!i915_gem_object_is_active(obj) && can_release_pages(obj))
 306                         count += obj->base.size >> PAGE_SHIFT;
 307         }
 308
 309         shrinker_unlock(dev_priv, unlock);
 310
 311         return count;
 312 }
 313
 314 static unsigned long
 315 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 316 {
 317         struct drm_i915_private *dev_priv =
 318                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 319         unsigned long freed;
 320         bool unlock;
 321
 322         if (!shrinker_lock(dev_priv, &unlock))
 323                 return SHRINK_STOP;
 324
 325         freed = i915_gem_shrink(dev_priv,
 326                                 sc->nr_to_scan,
 327                                 I915_SHRINK_BOUND |
 328                                 I915_SHRINK_UNBOUND |
 329                                 I915_SHRINK_PURGEABLE);
 330         if (freed < sc->nr_to_scan)
 331                 freed += i915_gem_shrink(dev_priv,
 332                                          sc->nr_to_scan - freed,
 333                                          I915_SHRINK_BOUND |
 334                                          I915_SHRINK_UNBOUND);
 335         if (freed < sc->nr_to_scan && current_is_kswapd()) {
 336                 intel_runtime_pm_get(dev_priv);
 337                 freed += i915_gem_shrink(dev_priv,
 338                                          sc->nr_to_scan - freed,
 339                                          I915_SHRINK_ACTIVE |
 340                                          I915_SHRINK_BOUND |
 341                                          I915_SHRINK_UNBOUND);
 342                 intel_runtime_pm_put(dev_priv);
 343         }
 344
 345         shrinker_unlock(dev_priv, unlock);
 346
 347         return freed;
 348 }
 349
 350 static bool
 351 shrinker_lock_uninterruptible(struct drm_i915_private *dev_priv, bool *unlock,
 352                               int timeout_ms)
 353 {
 354         unsigned long timeout = jiffies + msecs_to_jiffies_timeout(timeout_ms);
 355
 356         do {
 357                 if (i915_gem_wait_for_idle(dev_priv, 0) == 0 &&
 358                     shrinker_lock(dev_priv, unlock))
 359                         break;
 360
 361                 schedule_timeout_killable(1);
 362                 if (fatal_signal_pending(current))
 363                         return false;
 364
 365                 if (time_after(jiffies, timeout)) {
 366                         pr_err("Unable to lock GPU to purge memory.\n");
 367                         return false;
 368                 }
 369         } while (1);
 370
 371         return true;
 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 drm_i915_gem_object *obj;
 380         unsigned long unevictable, bound, unbound, freed_pages;
 381         bool unlock;
 382
 383         if (!shrinker_lock_uninterruptible(dev_priv, &unlock, 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         shrinker_unlock(dev_priv, unlock);
 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 i915_vma *vma, *next;
 433         unsigned long freed_pages = 0;
 434         bool unlock;
 435         int ret;
 436
 437         if (!shrinker_lock_uninterruptible(dev_priv, &unlock, 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         shrinker_unlock(dev_priv, unlock);
 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 }