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 mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
  39 {
  40         if (!mutex_is_locked(mutex))
  41                 return false;
  42
  43 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES)
  44         return mutex->owner == task;
  45 #else
  46         /* Since UP may be pre-empted, we cannot assume that we own the lock */
  47         return false;
  48 #endif
  49 }
  50
  51 static int num_vma_bound(struct drm_i915_gem_object *obj)
  52 {
  53         struct i915_vma *vma;
  54         int count = 0;
  55
  56         list_for_each_entry(vma, &obj->vma_list, obj_link) {
  57                 if (drm_mm_node_allocated(&vma->node))
  58                         count++;
  59                 if (vma->pin_count)
  60                         count++;
  61         }
  62
  63         return count;
  64 }
  65
  66 static bool swap_available(void)
  67 {
  68         return get_nr_swap_pages() > 0;
  69 }
  70
  71 static bool can_release_pages(struct drm_i915_gem_object *obj)
  72 {
  73         /* Only shmemfs objects are backed by swap */
  74         if (!obj->base.filp)
  75                 return false;
  76
  77         /* Only report true if by unbinding the object and putting its pages
  78          * we can actually make forward progress towards freeing physical
  79          * pages.
  80          *
  81          * If the pages are pinned for any other reason than being bound
  82          * to the GPU, simply unbinding from the GPU is not going to succeed
  83          * in releasing our pin count on the pages themselves.
  84          */
  85         if (obj->pages_pin_count != num_vma_bound(obj))
  86                 return false;
  87
  88         /* We can only return physical pages to the system if we can either
  89          * discard the contents (because the user has marked them as being
  90          * purgeable) or if we can move their contents out to swap.
  91          */
  92         return swap_available() || obj->madv == I915_MADV_DONTNEED;
  93 }
  94
  95 /**
  96  * i915_gem_shrink - Shrink buffer object caches
  97  * @dev_priv: i915 device
  98  * @target: amount of memory to make available, in pages
  99  * @flags: control flags for selecting cache types
 100  *
 101  * This function is the main interface to the shrinker. It will try to release
 102  * up to @target pages of main memory backing storage from buffer objects.
 103  * Selection of the specific caches can be done with @flags. This is e.g. useful
 104  * when purgeable objects should be removed from caches preferentially.
 105  *
 106  * Note that it's not guaranteed that released amount is actually available as
 107  * free system memory - the pages might still be in-used to due to other reasons
 108  * (like cpu mmaps) or the mm core has reused them before we could grab them.
 109  * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
 110  * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
 111  *
 112  * Also note that any kind of pinning (both per-vma address space pins and
 113  * backing storage pins at the buffer object level) result in the shrinker code
 114  * having to skip the object.
 115  *
 116  * Returns:
 117  * The number of pages of backing storage actually released.
 118  */
 119 unsigned long
 120 i915_gem_shrink(struct drm_i915_private *dev_priv,
 121                 unsigned long target, unsigned flags)
 122 {
 123         const struct {
 124                 struct list_head *list;
 125                 unsigned int bit;
 126         } phases[] = {
 127                 { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
 128                 { &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
 129                 { NULL, 0 },
 130         }, *phase;
 131         unsigned long count = 0;
 132
 133         trace_i915_gem_shrink(dev_priv, target, flags);
 134         i915_gem_retire_requests(dev_priv);
 135
 136         /*
 137          * Unbinding of objects will require HW access; Let us not wake the
 138          * device just to recover a little memory. If absolutely necessary,
 139          * we will force the wake during oom-notifier.
 140          */
 141         if ((flags & I915_SHRINK_BOUND) &&
 142             !intel_runtime_pm_get_if_in_use(dev_priv))
 143                 flags &= ~I915_SHRINK_BOUND;
 144
 145         /*
 146          * As we may completely rewrite the (un)bound list whilst unbinding
 147          * (due to retiring requests) we have to strictly process only
 148          * one element of the list at the time, and recheck the list
 149          * on every iteration.
 150          *
 151          * In particular, we must hold a reference whilst removing the
 152          * object as we may end up waiting for and/or retiring the objects.
 153          * This might release the final reference (held by the active list)
 154          * and result in the object being freed from under us. This is
 155          * similar to the precautions the eviction code must take whilst
 156          * removing objects.
 157          *
 158          * Also note that although these lists do not hold a reference to
 159          * the object we can safely grab one here: The final object
 160          * unreferencing and the bound_list are both protected by the
 161          * dev->struct_mutex and so we won't ever be able to observe an
 162          * object on the bound_list with a reference count equals 0.
 163          */
 164         for (phase = phases; phase->list; phase++) {
 165                 struct list_head still_in_list;
 166
 167                 if ((flags & phase->bit) == 0)
 168                         continue;
 169
 170                 INIT_LIST_HEAD(&still_in_list);
 171                 while (count < target && !list_empty(phase->list)) {
 172                         struct drm_i915_gem_object *obj;
 173                         struct i915_vma *vma, *v;
 174
 175                         obj = list_first_entry(phase->list,
 176                                                typeof(*obj), global_list);
 177                         list_move_tail(&obj->global_list, &still_in_list);
 178
 179                         if (flags & I915_SHRINK_PURGEABLE &&
 180                             obj->madv != I915_MADV_DONTNEED)
 181                                 continue;
 182
 183                         if (flags & I915_SHRINK_VMAPS &&
 184                             !is_vmalloc_addr(obj->mapping))
 185                                 continue;
 186
 187                         if ((flags & I915_SHRINK_ACTIVE) == 0 && obj->active)
 188                                 continue;
 189
 190                         if (!can_release_pages(obj))
 191                                 continue;
 192
 193                         drm_gem_object_reference(&obj->base);
 194
 195                         /* For the unbound phase, this should be a no-op! */
 196                         list_for_each_entry_safe(vma, v,
 197                                                  &obj->vma_list, obj_link)
 198                                 if (i915_vma_unbind(vma))
 199                                         break;
 200
 201                         if (i915_gem_object_put_pages(obj) == 0)
 202                                 count += obj->base.size >> PAGE_SHIFT;
 203
 204                         drm_gem_object_unreference(&obj->base);
 205                 }
 206                 list_splice(&still_in_list, phase->list);
 207         }
 208
 209         if (flags & I915_SHRINK_BOUND)
 210                 intel_runtime_pm_put(dev_priv);
 211
 212         i915_gem_retire_requests(dev_priv);
 213
 214         return count;
 215 }
 216
 217 /**
 218  * i915_gem_shrink_all - Shrink buffer object caches completely
 219  * @dev_priv: i915 device
 220  *
 221  * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 222  * caches completely. It also first waits for and retires all outstanding
 223  * requests to also be able to release backing storage for active objects.
 224  *
 225  * This should only be used in code to intentionally quiescent the gpu or as a
 226  * last-ditch effort when memory seems to have run out.
 227  *
 228  * Returns:
 229  * The number of pages of backing storage actually released.
 230  */
 231 unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv)
 232 {
 233         return i915_gem_shrink(dev_priv, -1UL,
 234                                I915_SHRINK_BOUND |
 235                                I915_SHRINK_UNBOUND |
 236                                I915_SHRINK_ACTIVE);
 237 }
 238
 239 static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock)
 240 {
 241         if (!mutex_trylock(&dev->struct_mutex)) {
 242                 if (!mutex_is_locked_by(&dev->struct_mutex, current))
 243                         return false;
 244
 245                 if (to_i915(dev)->mm.shrinker_no_lock_stealing)
 246                         return false;
 247
 248                 *unlock = false;
 249         } else
 250                 *unlock = true;
 251
 252         return true;
 253 }
 254
 255 static unsigned long
 256 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 257 {
 258         struct drm_i915_private *dev_priv =
 259                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 260         struct drm_device *dev = dev_priv->dev;
 261         struct drm_i915_gem_object *obj;
 262         unsigned long count;
 263         bool unlock;
 264
 265         if (!i915_gem_shrinker_lock(dev, &unlock))
 266                 return 0;
 267
 268         count = 0;
 269         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list)
 270                 if (can_release_pages(obj))
 271                         count += obj->base.size >> PAGE_SHIFT;
 272
 273         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
 274                 if (!obj->active && can_release_pages(obj))
 275                         count += obj->base.size >> PAGE_SHIFT;
 276         }
 277
 278         if (unlock)
 279                 mutex_unlock(&dev->struct_mutex);
 280
 281         return count;
 282 }
 283
 284 static unsigned long
 285 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 286 {
 287         struct drm_i915_private *dev_priv =
 288                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 289         struct drm_device *dev = dev_priv->dev;
 290         unsigned long freed;
 291         bool unlock;
 292
 293         if (!i915_gem_shrinker_lock(dev, &unlock))
 294                 return SHRINK_STOP;
 295
 296         freed = i915_gem_shrink(dev_priv,
 297                                 sc->nr_to_scan,
 298                                 I915_SHRINK_BOUND |
 299                                 I915_SHRINK_UNBOUND |
 300                                 I915_SHRINK_PURGEABLE);
 301         if (freed < sc->nr_to_scan)
 302                 freed += i915_gem_shrink(dev_priv,
 303                                          sc->nr_to_scan - freed,
 304                                          I915_SHRINK_BOUND |
 305                                          I915_SHRINK_UNBOUND);
 306         if (unlock)
 307                 mutex_unlock(&dev->struct_mutex);
 308
 309         return freed;
 310 }
 311
 312 struct shrinker_lock_uninterruptible {
 313         bool was_interruptible;
 314         bool unlock;
 315 };
 316
 317 static bool
 318 i915_gem_shrinker_lock_uninterruptible(struct drm_i915_private *dev_priv,
 319                                        struct shrinker_lock_uninterruptible *slu,
 320                                        int timeout_ms)
 321 {
 322         unsigned long timeout = msecs_to_jiffies(timeout_ms) + 1;
 323
 324         while (!i915_gem_shrinker_lock(dev_priv->dev, &slu->unlock)) {
 325                 schedule_timeout_killable(1);
 326                 if (fatal_signal_pending(current))
 327                         return false;
 328                 if (--timeout == 0) {
 329                         pr_err("Unable to lock GPU to purge memory.\n");
 330                         return false;
 331                 }
 332         }
 333
 334         slu->was_interruptible = dev_priv->mm.interruptible;
 335         dev_priv->mm.interruptible = false;
 336         return true;
 337 }
 338
 339 static void
 340 i915_gem_shrinker_unlock_uninterruptible(struct drm_i915_private *dev_priv,
 341                                          struct shrinker_lock_uninterruptible *slu)
 342 {
 343         dev_priv->mm.interruptible = slu->was_interruptible;
 344         if (slu->unlock)
 345                 mutex_unlock(&dev_priv->dev->struct_mutex);
 346 }
 347
 348 static int
 349 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 350 {
 351         struct drm_i915_private *dev_priv =
 352                 container_of(nb, struct drm_i915_private, mm.oom_notifier);
 353         struct shrinker_lock_uninterruptible slu;
 354         struct drm_i915_gem_object *obj;
 355         unsigned long unevictable, bound, unbound, freed_pages;
 356
 357         if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000))
 358                 return NOTIFY_DONE;
 359
 360         intel_runtime_pm_get(dev_priv);
 361         freed_pages = i915_gem_shrink_all(dev_priv);
 362         intel_runtime_pm_put(dev_priv);
 363
 364         /* Because we may be allocating inside our own driver, we cannot
 365          * assert that there are no objects with pinned pages that are not
 366          * being pointed to by hardware.
 367          */
 368         unbound = bound = unevictable = 0;
 369         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) {
 370                 if (!can_release_pages(obj))
 371                         unevictable += obj->base.size >> PAGE_SHIFT;
 372                 else
 373                         unbound += obj->base.size >> PAGE_SHIFT;
 374         }
 375         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
 376                 if (!can_release_pages(obj))
 377                         unevictable += obj->base.size >> PAGE_SHIFT;
 378                 else
 379                         bound += obj->base.size >> PAGE_SHIFT;
 380         }
 381
 382         i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
 383
 384         if (freed_pages || unbound || bound)
 385                 pr_info("Purging GPU memory, %lu pages freed, "
 386                         "%lu pages still pinned.\n",
 387                         freed_pages, unevictable);
 388         if (unbound || bound)
 389                 pr_err("%lu and %lu pages still available in the "
 390                        "bound and unbound GPU page lists.\n",
 391                        bound, unbound);
 392
 393         *(unsigned long *)ptr += freed_pages;
 394         return NOTIFY_DONE;
 395 }
 396
 397 static int
 398 i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
 399 {
 400         struct drm_i915_private *dev_priv =
 401                 container_of(nb, struct drm_i915_private, mm.vmap_notifier);
 402         struct shrinker_lock_uninterruptible slu;
 403         struct i915_vma *vma, *next;
 404         unsigned long freed_pages = 0;
 405         int ret;
 406
 407         if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000))
 408                 return NOTIFY_DONE;
 409
 410         /* Force everything onto the inactive lists */
 411         ret = i915_gpu_idle(dev_priv->dev);
 412         if (ret)
 413                 goto out;
 414
 415         intel_runtime_pm_get(dev_priv);
 416         freed_pages += i915_gem_shrink(dev_priv, -1UL,
 417                                        I915_SHRINK_BOUND |
 418                                        I915_SHRINK_UNBOUND |
 419                                        I915_SHRINK_ACTIVE |
 420                                        I915_SHRINK_VMAPS);
 421         intel_runtime_pm_put(dev_priv);
 422
 423         /* We also want to clear any cached iomaps as they wrap vmap */
 424         list_for_each_entry_safe(vma, next,
 425                                  &dev_priv->ggtt.base.inactive_list, vm_link) {
 426                 unsigned long count = vma->node.size >> PAGE_SHIFT;
 427                 if (vma->iomap && i915_vma_unbind(vma) == 0)
 428                         freed_pages += count;
 429         }
 430
 431 out:
 432         i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
 433
 434         *(unsigned long *)ptr += freed_pages;
 435         return NOTIFY_DONE;
 436 }
 437
 438 /**
 439  * i915_gem_shrinker_init - Initialize i915 shrinker
 440  * @dev_priv: i915 device
 441  *
 442  * This function registers and sets up the i915 shrinker and OOM handler.
 443  */
 444 void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
 445 {
 446         dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
 447         dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
 448         dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
 449         WARN_ON(register_shrinker(&dev_priv->mm.shrinker));
 450
 451         dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 452         WARN_ON(register_oom_notifier(&dev_priv->mm.oom_notifier));
 453
 454         dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
 455         WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
 456 }
 457
 458 /**
 459  * i915_gem_shrinker_cleanup - Clean up i915 shrinker
 460  * @dev_priv: i915 device
 461  *
 462  * This function unregisters the i915 shrinker and OOM handler.
 463  */
 464 void i915_gem_shrinker_cleanup(struct drm_i915_private *dev_priv)
 465 {
 466         WARN_ON(unregister_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
 467         WARN_ON(unregister_oom_notifier(&dev_priv->mm.oom_notifier));
 468         unregister_shrinker(&dev_priv->mm.shrinker);
 469 }