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
);
63 /* expedite the RCU grace period to free some request slabs */
64 synchronize_rcu_expedited();
67 static bool any_vma_pinned(struct drm_i915_gem_object
*obj
)
71 list_for_each_entry(vma
, &obj
->vma_list
, obj_link
)
72 if (i915_vma_is_pinned(vma
))
78 static bool swap_available(void)
80 return get_nr_swap_pages() > 0;
83 static bool can_release_pages(struct drm_i915_gem_object
*obj
)
88 /* Consider only shrinkable ojects. */
89 if (!i915_gem_object_is_shrinkable(obj
))
92 /* Only report true if by unbinding the object and putting its pages
93 * we can actually make forward progress towards freeing physical
96 * If the pages are pinned for any other reason than being bound
97 * to the GPU, simply unbinding from the GPU is not going to succeed
98 * in releasing our pin count on the pages themselves.
100 if (atomic_read(&obj
->mm
.pages_pin_count
) > obj
->bind_count
)
103 if (any_vma_pinned(obj
))
106 /* We can only return physical pages to the system if we can either
107 * discard the contents (because the user has marked them as being
108 * purgeable) or if we can move their contents out to swap.
110 return swap_available() || obj
->mm
.madv
== I915_MADV_DONTNEED
;
113 static bool unsafe_drop_pages(struct drm_i915_gem_object
*obj
)
115 if (i915_gem_object_unbind(obj
) == 0)
116 __i915_gem_object_put_pages(obj
, I915_MM_SHRINKER
);
117 return !READ_ONCE(obj
->mm
.pages
);
121 * i915_gem_shrink - Shrink buffer object caches
122 * @dev_priv: i915 device
123 * @target: amount of memory to make available, in pages
124 * @flags: control flags for selecting cache types
126 * This function is the main interface to the shrinker. It will try to release
127 * up to @target pages of main memory backing storage from buffer objects.
128 * Selection of the specific caches can be done with @flags. This is e.g. useful
129 * when purgeable objects should be removed from caches preferentially.
131 * Note that it's not guaranteed that released amount is actually available as
132 * free system memory - the pages might still be in-used to due to other reasons
133 * (like cpu mmaps) or the mm core has reused them before we could grab them.
134 * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
135 * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
137 * Also note that any kind of pinning (both per-vma address space pins and
138 * backing storage pins at the buffer object level) result in the shrinker code
139 * having to skip the object.
142 * The number of pages of backing storage actually released.
145 i915_gem_shrink(struct drm_i915_private
*dev_priv
,
146 unsigned long target
, unsigned flags
)
149 struct list_head
*list
;
152 { &dev_priv
->mm
.unbound_list
, I915_SHRINK_UNBOUND
},
153 { &dev_priv
->mm
.bound_list
, I915_SHRINK_BOUND
},
156 unsigned long count
= 0;
159 if (!i915_gem_shrinker_lock(&dev_priv
->drm
, &unlock
))
162 trace_i915_gem_shrink(dev_priv
, target
, flags
);
163 i915_gem_retire_requests(dev_priv
);
166 * Unbinding of objects will require HW access; Let us not wake the
167 * device just to recover a little memory. If absolutely necessary,
168 * we will force the wake during oom-notifier.
170 if ((flags
& I915_SHRINK_BOUND
) &&
171 !intel_runtime_pm_get_if_in_use(dev_priv
))
172 flags
&= ~I915_SHRINK_BOUND
;
175 * As we may completely rewrite the (un)bound list whilst unbinding
176 * (due to retiring requests) we have to strictly process only
177 * one element of the list at the time, and recheck the list
178 * on every iteration.
180 * In particular, we must hold a reference whilst removing the
181 * object as we may end up waiting for and/or retiring the objects.
182 * This might release the final reference (held by the active list)
183 * and result in the object being freed from under us. This is
184 * similar to the precautions the eviction code must take whilst
187 * Also note that although these lists do not hold a reference to
188 * the object we can safely grab one here: The final object
189 * unreferencing and the bound_list are both protected by the
190 * dev->struct_mutex and so we won't ever be able to observe an
191 * object on the bound_list with a reference count equals 0.
193 for (phase
= phases
; phase
->list
; phase
++) {
194 struct list_head still_in_list
;
195 struct drm_i915_gem_object
*obj
;
197 if ((flags
& phase
->bit
) == 0)
200 INIT_LIST_HEAD(&still_in_list
);
201 while (count
< target
&&
202 (obj
= list_first_entry_or_null(phase
->list
,
205 list_move_tail(&obj
->global_link
, &still_in_list
);
206 if (!obj
->mm
.pages
) {
207 list_del_init(&obj
->global_link
);
211 if (flags
& I915_SHRINK_PURGEABLE
&&
212 obj
->mm
.madv
!= I915_MADV_DONTNEED
)
215 if (flags
& I915_SHRINK_VMAPS
&&
216 !is_vmalloc_addr(obj
->mm
.mapping
))
219 if (!(flags
& I915_SHRINK_ACTIVE
) &&
220 (i915_gem_object_is_active(obj
) ||
221 obj
->framebuffer_references
))
224 if (!can_release_pages(obj
))
227 if (unsafe_drop_pages(obj
)) {
228 /* May arrive from get_pages on another bo */
229 mutex_lock_nested(&obj
->mm
.lock
,
231 if (!obj
->mm
.pages
) {
232 __i915_gem_object_invalidate(obj
);
233 list_del_init(&obj
->global_link
);
234 count
+= obj
->base
.size
>> PAGE_SHIFT
;
236 mutex_unlock(&obj
->mm
.lock
);
239 list_splice_tail(&still_in_list
, phase
->list
);
242 if (flags
& I915_SHRINK_BOUND
)
243 intel_runtime_pm_put(dev_priv
);
245 i915_gem_retire_requests(dev_priv
);
247 i915_gem_shrinker_unlock(&dev_priv
->drm
, unlock
);
253 * i915_gem_shrink_all - Shrink buffer object caches completely
254 * @dev_priv: i915 device
256 * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
257 * caches completely. It also first waits for and retires all outstanding
258 * requests to also be able to release backing storage for active objects.
260 * This should only be used in code to intentionally quiescent the gpu or as a
261 * last-ditch effort when memory seems to have run out.
264 * The number of pages of backing storage actually released.
266 unsigned long i915_gem_shrink_all(struct drm_i915_private
*dev_priv
)
270 freed
= i915_gem_shrink(dev_priv
, -1UL,
272 I915_SHRINK_UNBOUND
|
274 synchronize_rcu(); /* wait for our earlier RCU delayed slab frees */
280 i915_gem_shrinker_count(struct shrinker
*shrinker
, struct shrink_control
*sc
)
282 struct drm_i915_private
*dev_priv
=
283 container_of(shrinker
, struct drm_i915_private
, mm
.shrinker
);
284 struct drm_device
*dev
= &dev_priv
->drm
;
285 struct drm_i915_gem_object
*obj
;
289 if (!i915_gem_shrinker_lock(dev
, &unlock
))
292 i915_gem_retire_requests(dev_priv
);
295 list_for_each_entry(obj
, &dev_priv
->mm
.unbound_list
, global_link
)
296 if (can_release_pages(obj
))
297 count
+= obj
->base
.size
>> PAGE_SHIFT
;
299 list_for_each_entry(obj
, &dev_priv
->mm
.bound_list
, global_link
) {
300 if (!i915_gem_object_is_active(obj
) && can_release_pages(obj
))
301 count
+= obj
->base
.size
>> PAGE_SHIFT
;
304 i915_gem_shrinker_unlock(dev
, unlock
);
310 i915_gem_shrinker_scan(struct shrinker
*shrinker
, struct shrink_control
*sc
)
312 struct drm_i915_private
*dev_priv
=
313 container_of(shrinker
, struct drm_i915_private
, mm
.shrinker
);
314 struct drm_device
*dev
= &dev_priv
->drm
;
318 if (!i915_gem_shrinker_lock(dev
, &unlock
))
321 freed
= i915_gem_shrink(dev_priv
,
324 I915_SHRINK_UNBOUND
|
325 I915_SHRINK_PURGEABLE
);
326 if (freed
< sc
->nr_to_scan
)
327 freed
+= i915_gem_shrink(dev_priv
,
328 sc
->nr_to_scan
- freed
,
330 I915_SHRINK_UNBOUND
);
332 i915_gem_shrinker_unlock(dev
, unlock
);
337 struct shrinker_lock_uninterruptible
{
338 bool was_interruptible
;
343 i915_gem_shrinker_lock_uninterruptible(struct drm_i915_private
*dev_priv
,
344 struct shrinker_lock_uninterruptible
*slu
,
347 unsigned long timeout
= jiffies
+ msecs_to_jiffies_timeout(timeout_ms
);
350 if (i915_gem_wait_for_idle(dev_priv
, 0) == 0 &&
351 i915_gem_shrinker_lock(&dev_priv
->drm
, &slu
->unlock
))
354 schedule_timeout_killable(1);
355 if (fatal_signal_pending(current
))
358 if (time_after(jiffies
, timeout
)) {
359 pr_err("Unable to lock GPU to purge memory.\n");
364 slu
->was_interruptible
= dev_priv
->mm
.interruptible
;
365 dev_priv
->mm
.interruptible
= false;
370 i915_gem_shrinker_unlock_uninterruptible(struct drm_i915_private
*dev_priv
,
371 struct shrinker_lock_uninterruptible
*slu
)
373 dev_priv
->mm
.interruptible
= slu
->was_interruptible
;
374 i915_gem_shrinker_unlock(&dev_priv
->drm
, slu
->unlock
);
378 i915_gem_shrinker_oom(struct notifier_block
*nb
, unsigned long event
, void *ptr
)
380 struct drm_i915_private
*dev_priv
=
381 container_of(nb
, struct drm_i915_private
, mm
.oom_notifier
);
382 struct shrinker_lock_uninterruptible slu
;
383 struct drm_i915_gem_object
*obj
;
384 unsigned long unevictable
, bound
, unbound
, freed_pages
;
386 if (!i915_gem_shrinker_lock_uninterruptible(dev_priv
, &slu
, 5000))
389 intel_runtime_pm_get(dev_priv
);
390 freed_pages
= i915_gem_shrink_all(dev_priv
);
391 intel_runtime_pm_put(dev_priv
);
393 /* Because we may be allocating inside our own driver, we cannot
394 * assert that there are no objects with pinned pages that are not
395 * being pointed to by hardware.
397 unbound
= bound
= unevictable
= 0;
398 list_for_each_entry(obj
, &dev_priv
->mm
.unbound_list
, global_link
) {
402 if (!can_release_pages(obj
))
403 unevictable
+= obj
->base
.size
>> PAGE_SHIFT
;
405 unbound
+= obj
->base
.size
>> PAGE_SHIFT
;
407 list_for_each_entry(obj
, &dev_priv
->mm
.bound_list
, global_link
) {
411 if (!can_release_pages(obj
))
412 unevictable
+= obj
->base
.size
>> PAGE_SHIFT
;
414 bound
+= obj
->base
.size
>> PAGE_SHIFT
;
417 i915_gem_shrinker_unlock_uninterruptible(dev_priv
, &slu
);
419 if (freed_pages
|| unbound
|| bound
)
420 pr_info("Purging GPU memory, %lu pages freed, "
421 "%lu pages still pinned.\n",
422 freed_pages
, unevictable
);
423 if (unbound
|| bound
)
424 pr_err("%lu and %lu pages still available in the "
425 "bound and unbound GPU page lists.\n",
428 *(unsigned long *)ptr
+= freed_pages
;
433 i915_gem_shrinker_vmap(struct notifier_block
*nb
, unsigned long event
, void *ptr
)
435 struct drm_i915_private
*dev_priv
=
436 container_of(nb
, struct drm_i915_private
, mm
.vmap_notifier
);
437 struct shrinker_lock_uninterruptible slu
;
438 struct i915_vma
*vma
, *next
;
439 unsigned long freed_pages
= 0;
442 if (!i915_gem_shrinker_lock_uninterruptible(dev_priv
, &slu
, 5000))
445 /* Force everything onto the inactive lists */
446 ret
= i915_gem_wait_for_idle(dev_priv
, I915_WAIT_LOCKED
);
450 intel_runtime_pm_get(dev_priv
);
451 freed_pages
+= i915_gem_shrink(dev_priv
, -1UL,
453 I915_SHRINK_UNBOUND
|
456 intel_runtime_pm_put(dev_priv
);
458 /* We also want to clear any cached iomaps as they wrap vmap */
459 list_for_each_entry_safe(vma
, next
,
460 &dev_priv
->ggtt
.base
.inactive_list
, vm_link
) {
461 unsigned long count
= vma
->node
.size
>> PAGE_SHIFT
;
462 if (vma
->iomap
&& i915_vma_unbind(vma
) == 0)
463 freed_pages
+= count
;
467 i915_gem_shrinker_unlock_uninterruptible(dev_priv
, &slu
);
469 *(unsigned long *)ptr
+= freed_pages
;
474 * i915_gem_shrinker_init - Initialize i915 shrinker
475 * @dev_priv: i915 device
477 * This function registers and sets up the i915 shrinker and OOM handler.
479 void i915_gem_shrinker_init(struct drm_i915_private
*dev_priv
)
481 dev_priv
->mm
.shrinker
.scan_objects
= i915_gem_shrinker_scan
;
482 dev_priv
->mm
.shrinker
.count_objects
= i915_gem_shrinker_count
;
483 dev_priv
->mm
.shrinker
.seeks
= DEFAULT_SEEKS
;
484 WARN_ON(register_shrinker(&dev_priv
->mm
.shrinker
));
486 dev_priv
->mm
.oom_notifier
.notifier_call
= i915_gem_shrinker_oom
;
487 WARN_ON(register_oom_notifier(&dev_priv
->mm
.oom_notifier
));
489 dev_priv
->mm
.vmap_notifier
.notifier_call
= i915_gem_shrinker_vmap
;
490 WARN_ON(register_vmap_purge_notifier(&dev_priv
->mm
.vmap_notifier
));
494 * i915_gem_shrinker_cleanup - Clean up i915 shrinker
495 * @dev_priv: i915 device
497 * This function unregisters the i915 shrinker and OOM handler.
499 void i915_gem_shrinker_cleanup(struct drm_i915_private
*dev_priv
)
501 WARN_ON(unregister_vmap_purge_notifier(&dev_priv
->mm
.vmap_notifier
));
502 WARN_ON(unregister_oom_notifier(&dev_priv
->mm
.oom_notifier
));
503 unregister_shrinker(&dev_priv
->mm
.shrinker
);