2 * Copyright © 2008-2015 Intel Corporation
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:
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
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
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>
33 #include <drm/i915_drm.h>
36 #include "i915_trace.h"
38 static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock)
40 switch (mutex_trylock_recursive(&dev->struct_mutex)) {
41 case MUTEX_TRYLOCK_FAILED:
44 case MUTEX_TRYLOCK_SUCCESS:
48 case MUTEX_TRYLOCK_RECURSIVE:
56 static void i915_gem_shrinker_unlock(struct drm_device *dev, bool unlock)
61 mutex_unlock(&dev->struct_mutex);
64 static bool any_vma_pinned(struct drm_i915_gem_object *obj)
68 list_for_each_entry(vma, &obj->vma_list, obj_link)
69 if (i915_vma_is_pinned(vma))
75 static bool swap_available(void)
77 return get_nr_swap_pages() > 0;
80 static bool can_release_pages(struct drm_i915_gem_object *obj)
85 /* Consider only shrinkable ojects. */
86 if (!i915_gem_object_is_shrinkable(obj))
89 /* Only report true if by unbinding the object and putting its pages
90 * we can actually make forward progress towards freeing physical
93 * If the pages are pinned for any other reason than being bound
94 * to the GPU, simply unbinding from the GPU is not going to succeed
95 * in releasing our pin count on the pages themselves.
97 if (atomic_read(&obj->mm.pages_pin_count) > obj->bind_count)
100 if (any_vma_pinned(obj))
103 /* We can only return physical pages to the system if we can either
104 * discard the contents (because the user has marked them as being
105 * purgeable) or if we can move their contents out to swap.
107 return swap_available() || obj->mm.madv == I915_MADV_DONTNEED;
110 static bool unsafe_drop_pages(struct drm_i915_gem_object *obj)
112 if (i915_gem_object_unbind(obj) == 0)
113 __i915_gem_object_put_pages(obj, I915_MM_SHRINKER);
114 return !READ_ONCE(obj->mm.pages);
118 * i915_gem_shrink - Shrink buffer object caches
119 * @dev_priv: i915 device
120 * @target: amount of memory to make available, in pages
121 * @flags: control flags for selecting cache types
123 * This function is the main interface to the shrinker. It will try to release
124 * up to @target pages of main memory backing storage from buffer objects.
125 * Selection of the specific caches can be done with @flags. This is e.g. useful
126 * when purgeable objects should be removed from caches preferentially.
128 * Note that it's not guaranteed that released amount is actually available as
129 * free system memory - the pages might still be in-used to due to other reasons
130 * (like cpu mmaps) or the mm core has reused them before we could grab them.
131 * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
132 * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
134 * Also note that any kind of pinning (both per-vma address space pins and
135 * backing storage pins at the buffer object level) result in the shrinker code
136 * having to skip the object.
139 * The number of pages of backing storage actually released.
142 i915_gem_shrink(struct drm_i915_private *dev_priv,
143 unsigned long target, unsigned flags)
146 struct list_head *list;
149 { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
150 { &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
153 unsigned long count = 0;
156 if (!i915_gem_shrinker_lock(&dev_priv->drm, &unlock))
159 trace_i915_gem_shrink(dev_priv, target, flags);
160 i915_gem_retire_requests(dev_priv);
163 * Unbinding of objects will require HW access; Let us not wake the
164 * device just to recover a little memory. If absolutely necessary,
165 * we will force the wake during oom-notifier.
167 if ((flags & I915_SHRINK_BOUND) &&
168 !intel_runtime_pm_get_if_in_use(dev_priv))
169 flags &= ~I915_SHRINK_BOUND;
172 * As we may completely rewrite the (un)bound list whilst unbinding
173 * (due to retiring requests) we have to strictly process only
174 * one element of the list at the time, and recheck the list
175 * on every iteration.
177 * In particular, we must hold a reference whilst removing the
178 * object as we may end up waiting for and/or retiring the objects.
179 * This might release the final reference (held by the active list)
180 * and result in the object being freed from under us. This is
181 * similar to the precautions the eviction code must take whilst
184 * Also note that although these lists do not hold a reference to
185 * the object we can safely grab one here: The final object
186 * unreferencing and the bound_list are both protected by the
187 * dev->struct_mutex and so we won't ever be able to observe an
188 * object on the bound_list with a reference count equals 0.
190 for (phase = phases; phase->list; phase++) {
191 struct list_head still_in_list;
192 struct drm_i915_gem_object *obj;
194 if ((flags & phase->bit) == 0)
197 INIT_LIST_HEAD(&still_in_list);
198 while (count < target &&
199 (obj = list_first_entry_or_null(phase->list,
202 list_move_tail(&obj->global_link, &still_in_list);
203 if (!obj->mm.pages) {
204 list_del_init(&obj->global_link);
208 if (flags & I915_SHRINK_PURGEABLE &&
209 obj->mm.madv != I915_MADV_DONTNEED)
212 if (flags & I915_SHRINK_VMAPS &&
213 !is_vmalloc_addr(obj->mm.mapping))
216 if (!(flags & I915_SHRINK_ACTIVE) &&
217 (i915_gem_object_is_active(obj) ||
218 i915_gem_object_is_framebuffer(obj)))
221 if (!can_release_pages(obj))
224 if (unsafe_drop_pages(obj)) {
225 /* May arrive from get_pages on another bo */
226 mutex_lock_nested(&obj->mm.lock,
228 if (!obj->mm.pages) {
229 __i915_gem_object_invalidate(obj);
230 list_del_init(&obj->global_link);
231 count += obj->base.size >> PAGE_SHIFT;
233 mutex_unlock(&obj->mm.lock);
236 list_splice_tail(&still_in_list, phase->list);
239 if (flags & I915_SHRINK_BOUND)
240 intel_runtime_pm_put(dev_priv);
242 i915_gem_retire_requests(dev_priv);
244 i915_gem_shrinker_unlock(&dev_priv->drm, unlock);
250 * i915_gem_shrink_all - Shrink buffer object caches completely
251 * @dev_priv: i915 device
253 * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
254 * caches completely. It also first waits for and retires all outstanding
255 * requests to also be able to release backing storage for active objects.
257 * This should only be used in code to intentionally quiescent the gpu or as a
258 * last-ditch effort when memory seems to have run out.
261 * The number of pages of backing storage actually released.
263 unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv)
267 intel_runtime_pm_get(dev_priv);
268 freed = i915_gem_shrink(dev_priv, -1UL,
270 I915_SHRINK_UNBOUND |
272 intel_runtime_pm_put(dev_priv);
278 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
280 struct drm_i915_private *dev_priv =
281 container_of(shrinker, struct drm_i915_private, mm.shrinker);
282 struct drm_device *dev = &dev_priv->drm;
283 struct drm_i915_gem_object *obj;
287 if (!i915_gem_shrinker_lock(dev, &unlock))
290 i915_gem_retire_requests(dev_priv);
293 list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_link)
294 if (can_release_pages(obj))
295 count += obj->base.size >> PAGE_SHIFT;
297 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_link) {
298 if (!i915_gem_object_is_active(obj) && can_release_pages(obj))
299 count += obj->base.size >> PAGE_SHIFT;
302 i915_gem_shrinker_unlock(dev, unlock);
308 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
310 struct drm_i915_private *dev_priv =
311 container_of(shrinker, struct drm_i915_private, mm.shrinker);
312 struct drm_device *dev = &dev_priv->drm;
316 if (!i915_gem_shrinker_lock(dev, &unlock))
319 freed = i915_gem_shrink(dev_priv,
322 I915_SHRINK_UNBOUND |
323 I915_SHRINK_PURGEABLE);
324 if (freed < sc->nr_to_scan)
325 freed += i915_gem_shrink(dev_priv,
326 sc->nr_to_scan - freed,
328 I915_SHRINK_UNBOUND);
330 i915_gem_shrinker_unlock(dev, unlock);
335 struct shrinker_lock_uninterruptible {
336 bool was_interruptible;
341 i915_gem_shrinker_lock_uninterruptible(struct drm_i915_private *dev_priv,
342 struct shrinker_lock_uninterruptible *slu,
345 unsigned long timeout = jiffies + msecs_to_jiffies_timeout(timeout_ms);
348 if (i915_gem_wait_for_idle(dev_priv, 0) == 0 &&
349 i915_gem_shrinker_lock(&dev_priv->drm, &slu->unlock))
352 schedule_timeout_killable(1);
353 if (fatal_signal_pending(current))
356 if (time_after(jiffies, timeout)) {
357 pr_err("Unable to lock GPU to purge memory.\n");
362 slu->was_interruptible = dev_priv->mm.interruptible;
363 dev_priv->mm.interruptible = false;
368 i915_gem_shrinker_unlock_uninterruptible(struct drm_i915_private *dev_priv,
369 struct shrinker_lock_uninterruptible *slu)
371 dev_priv->mm.interruptible = slu->was_interruptible;
372 i915_gem_shrinker_unlock(&dev_priv->drm, slu->unlock);
376 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
378 struct drm_i915_private *dev_priv =
379 container_of(nb, struct drm_i915_private, mm.oom_notifier);
380 struct shrinker_lock_uninterruptible slu;
381 struct drm_i915_gem_object *obj;
382 unsigned long unevictable, bound, unbound, freed_pages;
384 if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000))
387 freed_pages = i915_gem_shrink_all(dev_priv);
389 /* Because we may be allocating inside our own driver, we cannot
390 * assert that there are no objects with pinned pages that are not
391 * being pointed to by hardware.
393 unbound = bound = unevictable = 0;
394 list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_link) {
398 if (!can_release_pages(obj))
399 unevictable += obj->base.size >> PAGE_SHIFT;
401 unbound += obj->base.size >> PAGE_SHIFT;
403 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_link) {
407 if (!can_release_pages(obj))
408 unevictable += obj->base.size >> PAGE_SHIFT;
410 bound += obj->base.size >> PAGE_SHIFT;
413 i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
415 if (freed_pages || unbound || bound)
416 pr_info("Purging GPU memory, %lu pages freed, "
417 "%lu pages still pinned.\n",
418 freed_pages, unevictable);
419 if (unbound || bound)
420 pr_err("%lu and %lu pages still available in the "
421 "bound and unbound GPU page lists.\n",
424 *(unsigned long *)ptr += freed_pages;
429 i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
431 struct drm_i915_private *dev_priv =
432 container_of(nb, struct drm_i915_private, mm.vmap_notifier);
433 struct shrinker_lock_uninterruptible slu;
434 struct i915_vma *vma, *next;
435 unsigned long freed_pages = 0;
438 if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000))
441 /* Force everything onto the inactive lists */
442 ret = i915_gem_wait_for_idle(dev_priv, I915_WAIT_LOCKED);
446 intel_runtime_pm_get(dev_priv);
447 freed_pages += i915_gem_shrink(dev_priv, -1UL,
449 I915_SHRINK_UNBOUND |
452 intel_runtime_pm_put(dev_priv);
454 /* We also want to clear any cached iomaps as they wrap vmap */
455 list_for_each_entry_safe(vma, next,
456 &dev_priv->ggtt.base.inactive_list, vm_link) {
457 unsigned long count = vma->node.size >> PAGE_SHIFT;
458 if (vma->iomap && i915_vma_unbind(vma) == 0)
459 freed_pages += count;
463 i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
465 *(unsigned long *)ptr += freed_pages;
470 * i915_gem_shrinker_init - Initialize i915 shrinker
471 * @dev_priv: i915 device
473 * This function registers and sets up the i915 shrinker and OOM handler.
475 void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
477 dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
478 dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
479 dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
480 WARN_ON(register_shrinker(&dev_priv->mm.shrinker));
482 dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
483 WARN_ON(register_oom_notifier(&dev_priv->mm.oom_notifier));
485 dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
486 WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
490 * i915_gem_shrinker_cleanup - Clean up i915 shrinker
491 * @dev_priv: i915 device
493 * This function unregisters the i915 shrinker and OOM handler.
495 void i915_gem_shrinker_cleanup(struct drm_i915_private *dev_priv)
497 WARN_ON(unregister_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
498 WARN_ON(unregister_oom_notifier(&dev_priv->mm.oom_notifier));
499 unregister_shrinker(&dev_priv->mm.shrinker);