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_DEBUG_MUTEXES) || defined(CONFIG_MUTEX_SPIN_ON_OWNER)
  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                 unsigned 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         trace_i915_gem_shrink(dev_priv, target, flags);
  89         i915_gem_retire_requests(dev_priv->dev);
  90
  91         /*
  92          * As we may completely rewrite the (un)bound list whilst unbinding
  93          * (due to retiring requests) we have to strictly process only
  94          * one element of the list at the time, and recheck the list
  95          * on every iteration.
  96          *
  97          * In particular, we must hold a reference whilst removing the
  98          * object as we may end up waiting for and/or retiring the objects.
  99          * This might release the final reference (held by the active list)
 100          * and result in the object being freed from under us. This is
 101          * similar to the precautions the eviction code must take whilst
 102          * removing objects.
 103          *
 104          * Also note that although these lists do not hold a reference to
 105          * the object we can safely grab one here: The final object
 106          * unreferencing and the bound_list are both protected by the
 107          * dev->struct_mutex and so we won't ever be able to observe an
 108          * object on the bound_list with a reference count equals 0.
 109          */
 110         for (phase = phases; phase->list; phase++) {
 111                 struct list_head still_in_list;
 112
 113                 if ((flags & phase->bit) == 0)
 114                         continue;
 115
 116                 INIT_LIST_HEAD(&still_in_list);
 117                 while (count < target && !list_empty(phase->list)) {
 118                         struct drm_i915_gem_object *obj;
 119                         struct i915_vma *vma, *v;
 120
 121                         obj = list_first_entry(phase->list,
 122                                                typeof(*obj), global_list);
 123                         list_move_tail(&obj->global_list, &still_in_list);
 124
 125                         if (flags & I915_SHRINK_PURGEABLE &&
 126                             obj->madv != I915_MADV_DONTNEED)
 127                                 continue;
 128
 129                         if ((flags & I915_SHRINK_ACTIVE) == 0 && obj->active)
 130                                 continue;
 131
 132                         drm_gem_object_reference(&obj->base);
 133
 134                         /* For the unbound phase, this should be a no-op! */
 135                         list_for_each_entry_safe(vma, v,
 136                                                  &obj->vma_list, vma_link)
 137                                 if (i915_vma_unbind(vma))
 138                                         break;
 139
 140                         if (i915_gem_object_put_pages(obj) == 0)
 141                                 count += obj->base.size >> PAGE_SHIFT;
 142
 143                         drm_gem_object_unreference(&obj->base);
 144                 }
 145                 list_splice(&still_in_list, phase->list);
 146         }
 147
 148         i915_gem_retire_requests(dev_priv->dev);
 149
 150         return count;
 151 }
 152
 153 /**
 154  * i915_gem_shrink_all - Shrink buffer object caches completely
 155  * @dev_priv: i915 device
 156  *
 157  * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 158  * caches completely. It also first waits for and retires all outstanding
 159  * requests to also be able to release backing storage for active objects.
 160  *
 161  * This should only be used in code to intentionally quiescent the gpu or as a
 162  * last-ditch effort when memory seems to have run out.
 163  *
 164  * Returns:
 165  * The number of pages of backing storage actually released.
 166  */
 167 unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv)
 168 {
 169         return i915_gem_shrink(dev_priv, -1UL,
 170                                I915_SHRINK_BOUND |
 171                                I915_SHRINK_UNBOUND |
 172                                I915_SHRINK_ACTIVE);
 173 }
 174
 175 static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock)
 176 {
 177         if (!mutex_trylock(&dev->struct_mutex)) {
 178                 if (!mutex_is_locked_by(&dev->struct_mutex, current))
 179                         return false;
 180
 181                 if (to_i915(dev)->mm.shrinker_no_lock_stealing)
 182                         return false;
 183
 184                 *unlock = false;
 185         } else
 186                 *unlock = true;
 187
 188         return true;
 189 }
 190
 191 static int num_vma_bound(struct drm_i915_gem_object *obj)
 192 {
 193         struct i915_vma *vma;
 194         int count = 0;
 195
 196         list_for_each_entry(vma, &obj->vma_list, vma_link) {
 197                 if (drm_mm_node_allocated(&vma->node))
 198                         count++;
 199                 if (vma->pin_count)
 200                         count++;
 201         }
 202
 203         return count;
 204 }
 205
 206 static unsigned long
 207 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 208 {
 209         struct drm_i915_private *dev_priv =
 210                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 211         struct drm_device *dev = dev_priv->dev;
 212         struct drm_i915_gem_object *obj;
 213         unsigned long count;
 214         bool unlock;
 215
 216         if (!i915_gem_shrinker_lock(dev, &unlock))
 217                 return 0;
 218
 219         count = 0;
 220         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list)
 221                 if (obj->pages_pin_count == 0)
 222                         count += obj->base.size >> PAGE_SHIFT;
 223
 224         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
 225                 if (!obj->active && obj->pages_pin_count == num_vma_bound(obj))
 226                         count += obj->base.size >> PAGE_SHIFT;
 227         }
 228
 229         if (unlock)
 230                 mutex_unlock(&dev->struct_mutex);
 231
 232         return count;
 233 }
 234
 235 static unsigned long
 236 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 237 {
 238         struct drm_i915_private *dev_priv =
 239                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 240         struct drm_device *dev = dev_priv->dev;
 241         unsigned long freed;
 242         bool unlock;
 243
 244         if (!i915_gem_shrinker_lock(dev, &unlock))
 245                 return SHRINK_STOP;
 246
 247         freed = i915_gem_shrink(dev_priv,
 248                                 sc->nr_to_scan,
 249                                 I915_SHRINK_BOUND |
 250                                 I915_SHRINK_UNBOUND |
 251                                 I915_SHRINK_PURGEABLE);
 252         if (freed < sc->nr_to_scan)
 253                 freed += i915_gem_shrink(dev_priv,
 254                                          sc->nr_to_scan - freed,
 255                                          I915_SHRINK_BOUND |
 256                                          I915_SHRINK_UNBOUND);
 257         if (unlock)
 258                 mutex_unlock(&dev->struct_mutex);
 259
 260         return freed;
 261 }
 262
 263 static int
 264 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 265 {
 266         struct drm_i915_private *dev_priv =
 267                 container_of(nb, struct drm_i915_private, mm.oom_notifier);
 268         struct drm_device *dev = dev_priv->dev;
 269         struct drm_i915_gem_object *obj;
 270         unsigned long timeout = msecs_to_jiffies(5000) + 1;
 271         unsigned long pinned, bound, unbound, freed_pages;
 272         bool was_interruptible;
 273         bool unlock;
 274
 275         while (!i915_gem_shrinker_lock(dev, &unlock) && --timeout) {
 276                 schedule_timeout_killable(1);
 277                 if (fatal_signal_pending(current))
 278                         return NOTIFY_DONE;
 279         }
 280         if (timeout == 0) {
 281                 pr_err("Unable to purge GPU memory due lock contention.\n");
 282                 return NOTIFY_DONE;
 283         }
 284
 285         was_interruptible = dev_priv->mm.interruptible;
 286         dev_priv->mm.interruptible = false;
 287
 288         freed_pages = i915_gem_shrink_all(dev_priv);
 289
 290         dev_priv->mm.interruptible = was_interruptible;
 291
 292         /* Because we may be allocating inside our own driver, we cannot
 293          * assert that there are no objects with pinned pages that are not
 294          * being pointed to by hardware.
 295          */
 296         unbound = bound = pinned = 0;
 297         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) {
 298                 if (!obj->base.filp) /* not backed by a freeable object */
 299                         continue;
 300
 301                 if (obj->pages_pin_count)
 302                         pinned += obj->base.size;
 303                 else
 304                         unbound += obj->base.size;
 305         }
 306         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
 307                 if (!obj->base.filp)
 308                         continue;
 309
 310                 if (obj->pages_pin_count)
 311                         pinned += obj->base.size;
 312                 else
 313                         bound += obj->base.size;
 314         }
 315
 316         if (unlock)
 317                 mutex_unlock(&dev->struct_mutex);
 318
 319         if (freed_pages || unbound || bound)
 320                 pr_info("Purging GPU memory, %lu bytes freed, %lu bytes still pinned.\n",
 321                         freed_pages << PAGE_SHIFT, pinned);
 322         if (unbound || bound)
 323                 pr_err("%lu and %lu bytes still available in the "
 324                        "bound and unbound GPU page lists.\n",
 325                        bound, unbound);
 326
 327         *(unsigned long *)ptr += freed_pages;
 328         return NOTIFY_DONE;
 329 }
 330
 331 /**
 332  * i915_gem_shrinker_init - Initialize i915 shrinker
 333  * @dev_priv: i915 device
 334  *
 335  * This function registers and sets up the i915 shrinker and OOM handler.
 336  */
 337 void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
 338 {
 339         dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
 340         dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
 341         dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
 342         register_shrinker(&dev_priv->mm.shrinker);
 343
 344         dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 345         register_oom_notifier(&dev_priv->mm.oom_notifier);
 346 }