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