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 <drm/drmP.h>
  32 #include <drm/i915_drm.h>
  33
  34 #include "i915_drv.h"
  35 #include "i915_trace.h"
  36
  37 static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
  38 {
  39         if (!mutex_is_locked(mutex))
  40                 return false;
  41
  42 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES)
  43         return mutex->owner == task;
  44 #else
  45         /* Since UP may be pre-empted, we cannot assume that we own the lock */
  46         return false;
  47 #endif
  48 }
  49
  50 /**
  51  * i915_gem_shrink - Shrink buffer object caches
  52  * @dev_priv: i915 device
  53  * @target: amount of memory to make available, in pages
  54  * @flags: control flags for selecting cache types
  55  *
  56  * This function is the main interface to the shrinker. It will try to release
  57  * up to @target pages of main memory backing storage from buffer objects.
  58  * Selection of the specific caches can be done with @flags. This is e.g. useful
  59  * when purgeable objects should be removed from caches preferentially.
  60  *
  61  * Note that it's not guaranteed that released amount is actually available as
  62  * free system memory - the pages might still be in-used to due to other reasons
  63  * (like cpu mmaps) or the mm core has reused them before we could grab them.
  64  * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
  65  * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
  66  *
  67  * Also note that any kind of pinning (both per-vma address space pins and
  68  * backing storage pins at the buffer object level) result in the shrinker code
  69  * having to skip the object.
  70  *
  71  * Returns:
  72  * The number of pages of backing storage actually released.
  73  */
  74 unsigned long
  75 i915_gem_shrink(struct drm_i915_private *dev_priv,
  76                 long target, unsigned flags)
  77 {
  78         const struct {
  79                 struct list_head *list;
  80                 unsigned int bit;
  81         } phases[] = {
  82                 { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
  83                 { &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
  84                 { NULL, 0 },
  85         }, *phase;
  86         unsigned long count = 0;
  87
  88         /*
  89          * As we may completely rewrite the (un)bound list whilst unbinding
  90          * (due to retiring requests) we have to strictly process only
  91          * one element of the list at the time, and recheck the list
  92          * on every iteration.
  93          *
  94          * In particular, we must hold a reference whilst removing the
  95          * object as we may end up waiting for and/or retiring the objects.
  96          * This might release the final reference (held by the active list)
  97          * and result in the object being freed from under us. This is
  98          * similar to the precautions the eviction code must take whilst
  99          * removing objects.
 100          *
 101          * Also note that although these lists do not hold a reference to
 102          * the object we can safely grab one here: The final object
 103          * unreferencing and the bound_list are both protected by the
 104          * dev->struct_mutex and so we won't ever be able to observe an
 105          * object on the bound_list with a reference count equals 0.
 106          */
 107         for (phase = phases; phase->list; phase++) {
 108                 struct list_head still_in_list;
 109
 110                 if ((flags & phase->bit) == 0)
 111                         continue;
 112
 113                 INIT_LIST_HEAD(&still_in_list);
 114                 while (count < target && !list_empty(phase->list)) {
 115                         struct drm_i915_gem_object *obj;
 116                         struct i915_vma *vma, *v;
 117
 118                         obj = list_first_entry(phase->list,
 119                                                typeof(*obj), global_list);
 120                         list_move_tail(&obj->global_list, &still_in_list);
 121
 122                         if (flags & I915_SHRINK_PURGEABLE &&
 123                             obj->madv != I915_MADV_DONTNEED)
 124                                 continue;
 125
 126                         drm_gem_object_reference(&obj->base);
 127
 128                         /* For the unbound phase, this should be a no-op! */
 129                         list_for_each_entry_safe(vma, v,
 130                                                  &obj->vma_list, vma_link)
 131                                 if (i915_vma_unbind(vma))
 132                                         break;
 133
 134                         if (i915_gem_object_put_pages(obj) == 0)
 135                                 count += obj->base.size >> PAGE_SHIFT;
 136
 137                         drm_gem_object_unreference(&obj->base);
 138                 }
 139                 list_splice(&still_in_list, phase->list);
 140         }
 141
 142         return count;
 143 }
 144
 145 /**
 146  * i915_gem_shrink_all - Shrink buffer object caches completely
 147  * @dev_priv: i915 device
 148  *
 149  * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 150  * caches completely. It also first waits for and retires all outstanding
 151  * requests to also be able to release backing storage for active objects.
 152  *
 153  * This should only be used in code to intentionally quiescent the gpu or as a
 154  * last-ditch effort when memory seems to have run out.
 155  *
 156  * Returns:
 157  * The number of pages of backing storage actually released.
 158  */
 159 unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv)
 160 {
 161         i915_gem_evict_everything(dev_priv->dev);
 162         return i915_gem_shrink(dev_priv, LONG_MAX,
 163                                I915_SHRINK_BOUND | I915_SHRINK_UNBOUND);
 164 }
 165
 166 static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock)
 167 {
 168         if (!mutex_trylock(&dev->struct_mutex)) {
 169                 if (!mutex_is_locked_by(&dev->struct_mutex, current))
 170                         return false;
 171
 172                 if (to_i915(dev)->mm.shrinker_no_lock_stealing)
 173                         return false;
 174
 175                 *unlock = false;
 176         } else
 177                 *unlock = true;
 178
 179         return true;
 180 }
 181
 182 static int num_vma_bound(struct drm_i915_gem_object *obj)
 183 {
 184         struct i915_vma *vma;
 185         int count = 0;
 186
 187         list_for_each_entry(vma, &obj->vma_list, vma_link) {
 188                 if (drm_mm_node_allocated(&vma->node))
 189                         count++;
 190                 if (vma->pin_count)
 191                         count++;
 192         }
 193
 194         return count;
 195 }
 196
 197 static unsigned long
 198 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 199 {
 200         struct drm_i915_private *dev_priv =
 201                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 202         struct drm_device *dev = dev_priv->dev;
 203         struct drm_i915_gem_object *obj;
 204         unsigned long count;
 205         bool unlock;
 206
 207         if (!i915_gem_shrinker_lock(dev, &unlock))
 208                 return 0;
 209
 210         count = 0;
 211         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list)
 212                 if (obj->pages_pin_count == 0)
 213                         count += obj->base.size >> PAGE_SHIFT;
 214
 215         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
 216                 if (obj->pages_pin_count == num_vma_bound(obj))
 217                         count += obj->base.size >> PAGE_SHIFT;
 218         }
 219
 220         if (unlock)
 221                 mutex_unlock(&dev->struct_mutex);
 222
 223         return count;
 224 }
 225
 226 static unsigned long
 227 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 228 {
 229         struct drm_i915_private *dev_priv =
 230                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 231         struct drm_device *dev = dev_priv->dev;
 232         unsigned long freed;
 233         bool unlock;
 234
 235         if (!i915_gem_shrinker_lock(dev, &unlock))
 236                 return SHRINK_STOP;
 237
 238         freed = i915_gem_shrink(dev_priv,
 239                                 sc->nr_to_scan,
 240                                 I915_SHRINK_BOUND |
 241                                 I915_SHRINK_UNBOUND |
 242                                 I915_SHRINK_PURGEABLE);
 243         if (freed < sc->nr_to_scan)
 244                 freed += i915_gem_shrink(dev_priv,
 245                                          sc->nr_to_scan - freed,
 246                                          I915_SHRINK_BOUND |
 247                                          I915_SHRINK_UNBOUND);
 248         if (unlock)
 249                 mutex_unlock(&dev->struct_mutex);
 250
 251         return freed;
 252 }
 253
 254 static int
 255 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 256 {
 257         struct drm_i915_private *dev_priv =
 258                 container_of(nb, struct drm_i915_private, mm.oom_notifier);
 259         struct drm_device *dev = dev_priv->dev;
 260         struct drm_i915_gem_object *obj;
 261         unsigned long timeout = msecs_to_jiffies(5000) + 1;
 262         unsigned long pinned, bound, unbound, freed_pages;
 263         bool was_interruptible;
 264         bool unlock;
 265
 266         while (!i915_gem_shrinker_lock(dev, &unlock) && --timeout) {
 267                 schedule_timeout_killable(1);
 268                 if (fatal_signal_pending(current))
 269                         return NOTIFY_DONE;
 270         }
 271         if (timeout == 0) {
 272                 pr_err("Unable to purge GPU memory due lock contention.\n");
 273                 return NOTIFY_DONE;
 274         }
 275
 276         was_interruptible = dev_priv->mm.interruptible;
 277         dev_priv->mm.interruptible = false;
 278
 279         freed_pages = i915_gem_shrink_all(dev_priv);
 280
 281         dev_priv->mm.interruptible = was_interruptible;
 282
 283         /* Because we may be allocating inside our own driver, we cannot
 284          * assert that there are no objects with pinned pages that are not
 285          * being pointed to by hardware.
 286          */
 287         unbound = bound = pinned = 0;
 288         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) {
 289                 if (!obj->base.filp) /* not backed by a freeable object */
 290                         continue;
 291
 292                 if (obj->pages_pin_count)
 293                         pinned += obj->base.size;
 294                 else
 295                         unbound += obj->base.size;
 296         }
 297         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
 298                 if (!obj->base.filp)
 299                         continue;
 300
 301                 if (obj->pages_pin_count)
 302                         pinned += obj->base.size;
 303                 else
 304                         bound += obj->base.size;
 305         }
 306
 307         if (unlock)
 308                 mutex_unlock(&dev->struct_mutex);
 309
 310         if (freed_pages || unbound || bound)
 311                 pr_info("Purging GPU memory, %lu bytes freed, %lu bytes still pinned.\n",
 312                         freed_pages << PAGE_SHIFT, pinned);
 313         if (unbound || bound)
 314                 pr_err("%lu and %lu bytes still available in the "
 315                        "bound and unbound GPU page lists.\n",
 316                        bound, unbound);
 317
 318         *(unsigned long *)ptr += freed_pages;
 319         return NOTIFY_DONE;
 320 }
 321
 322 /**
 323  * i915_gem_shrinker_init - Initialize i915 shrinker
 324  * @dev_priv: i915 device
 325  *
 326  * This function registers and sets up the i915 shrinker and OOM handler.
 327  */
 328 void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
 329 {
 330         dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
 331         dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
 332         dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
 333         register_shrinker(&dev_priv->mm.shrinker);
 334
 335         dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 336         register_oom_notifier(&dev_priv->mm.oom_notifier);
 337 }