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