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 bool any_vma_pinned(struct drm_i915_gem_object
*obj
)
60 list_for_each_entry(vma
, &obj
->vma_list
, obj_link
)
61 if (i915_vma_is_pinned(vma
))
67 static bool swap_available(void)
69 return get_nr_swap_pages() > 0;
72 static bool can_release_pages(struct drm_i915_gem_object
*obj
)
77 /* Consider only shrinkable ojects. */
78 if (!i915_gem_object_is_shrinkable(obj
))
81 /* Only report true if by unbinding the object and putting its pages
82 * we can actually make forward progress towards freeing physical
85 * If the pages are pinned for any other reason than being bound
86 * to the GPU, simply unbinding from the GPU is not going to succeed
87 * in releasing our pin count on the pages themselves.
89 if (atomic_read(&obj
->mm
.pages_pin_count
) > obj
->bind_count
)
92 if (any_vma_pinned(obj
))
95 /* We can only return physical pages to the system if we can either
96 * discard the contents (because the user has marked them as being
97 * purgeable) or if we can move their contents out to swap.
99 return swap_available() || obj
->mm
.madv
== I915_MADV_DONTNEED
;
102 static bool unsafe_drop_pages(struct drm_i915_gem_object
*obj
)
104 if (i915_gem_object_unbind(obj
) == 0)
105 __i915_gem_object_put_pages(obj
, I915_MM_SHRINKER
);
106 return !READ_ONCE(obj
->mm
.pages
);
110 * i915_gem_shrink - Shrink buffer object caches
111 * @dev_priv: i915 device
112 * @target: amount of memory to make available, in pages
113 * @flags: control flags for selecting cache types
115 * This function is the main interface to the shrinker. It will try to release
116 * up to @target pages of main memory backing storage from buffer objects.
117 * Selection of the specific caches can be done with @flags. This is e.g. useful
118 * when purgeable objects should be removed from caches preferentially.
120 * Note that it's not guaranteed that released amount is actually available as
121 * free system memory - the pages might still be in-used to due to other reasons
122 * (like cpu mmaps) or the mm core has reused them before we could grab them.
123 * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
124 * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
126 * Also note that any kind of pinning (both per-vma address space pins and
127 * backing storage pins at the buffer object level) result in the shrinker code
128 * having to skip the object.
131 * The number of pages of backing storage actually released.
134 i915_gem_shrink(struct drm_i915_private
*dev_priv
,
135 unsigned long target
, unsigned flags
)
138 struct list_head
*list
;
141 { &dev_priv
->mm
.unbound_list
, I915_SHRINK_UNBOUND
},
142 { &dev_priv
->mm
.bound_list
, I915_SHRINK_BOUND
},
145 unsigned long count
= 0;
148 if (!i915_gem_shrinker_lock(&dev_priv
->drm
, &unlock
))
151 trace_i915_gem_shrink(dev_priv
, target
, flags
);
152 i915_gem_retire_requests(dev_priv
);
155 * Unbinding of objects will require HW access; Let us not wake the
156 * device just to recover a little memory. If absolutely necessary,
157 * we will force the wake during oom-notifier.
159 if ((flags
& I915_SHRINK_BOUND
) &&
160 !intel_runtime_pm_get_if_in_use(dev_priv
))
161 flags
&= ~I915_SHRINK_BOUND
;
164 * As we may completely rewrite the (un)bound list whilst unbinding
165 * (due to retiring requests) we have to strictly process only
166 * one element of the list at the time, and recheck the list
167 * on every iteration.
169 * In particular, we must hold a reference whilst removing the
170 * object as we may end up waiting for and/or retiring the objects.
171 * This might release the final reference (held by the active list)
172 * and result in the object being freed from under us. This is
173 * similar to the precautions the eviction code must take whilst
176 * Also note that although these lists do not hold a reference to
177 * the object we can safely grab one here: The final object
178 * unreferencing and the bound_list are both protected by the
179 * dev->struct_mutex and so we won't ever be able to observe an
180 * object on the bound_list with a reference count equals 0.
182 for (phase
= phases
; phase
->list
; phase
++) {
183 struct list_head still_in_list
;
184 struct drm_i915_gem_object
*obj
;
186 if ((flags
& phase
->bit
) == 0)
189 INIT_LIST_HEAD(&still_in_list
);
190 while (count
< target
&&
191 (obj
= list_first_entry_or_null(phase
->list
,
194 list_move_tail(&obj
->global_link
, &still_in_list
);
195 if (!obj
->mm
.pages
) {
196 list_del_init(&obj
->global_link
);
200 if (flags
& I915_SHRINK_PURGEABLE
&&
201 obj
->mm
.madv
!= I915_MADV_DONTNEED
)
204 if (flags
& I915_SHRINK_VMAPS
&&
205 !is_vmalloc_addr(obj
->mm
.mapping
))
208 if (!(flags
& I915_SHRINK_ACTIVE
) &&
209 (i915_gem_object_is_active(obj
) ||
210 obj
->framebuffer_references
))
213 if (!can_release_pages(obj
))
216 if (unsafe_drop_pages(obj
)) {
217 /* May arrive from get_pages on another bo */
218 mutex_lock_nested(&obj
->mm
.lock
,
220 if (!obj
->mm
.pages
) {
221 __i915_gem_object_invalidate(obj
);
222 list_del_init(&obj
->global_link
);
223 count
+= obj
->base
.size
>> PAGE_SHIFT
;
225 mutex_unlock(&obj
->mm
.lock
);
228 list_splice_tail(&still_in_list
, phase
->list
);
231 if (flags
& I915_SHRINK_BOUND
)
232 intel_runtime_pm_put(dev_priv
);
234 i915_gem_retire_requests(dev_priv
);
236 mutex_unlock(&dev_priv
->drm
.struct_mutex
);
238 /* expedite the RCU grace period to free some request slabs */
239 synchronize_rcu_expedited();
245 * i915_gem_shrink_all - Shrink buffer object caches completely
246 * @dev_priv: i915 device
248 * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
249 * caches completely. It also first waits for and retires all outstanding
250 * requests to also be able to release backing storage for active objects.
252 * This should only be used in code to intentionally quiescent the gpu or as a
253 * last-ditch effort when memory seems to have run out.
256 * The number of pages of backing storage actually released.
258 unsigned long i915_gem_shrink_all(struct drm_i915_private
*dev_priv
)
262 freed
= i915_gem_shrink(dev_priv
, -1UL,
264 I915_SHRINK_UNBOUND
|
266 rcu_barrier(); /* wait until our RCU delayed slab frees are completed */
272 i915_gem_shrinker_count(struct shrinker
*shrinker
, struct shrink_control
*sc
)
274 struct drm_i915_private
*dev_priv
=
275 container_of(shrinker
, struct drm_i915_private
, mm
.shrinker
);
276 struct drm_device
*dev
= &dev_priv
->drm
;
277 struct drm_i915_gem_object
*obj
;
281 if (!i915_gem_shrinker_lock(dev
, &unlock
))
284 i915_gem_retire_requests(dev_priv
);
287 list_for_each_entry(obj
, &dev_priv
->mm
.unbound_list
, global_link
)
288 if (can_release_pages(obj
))
289 count
+= obj
->base
.size
>> PAGE_SHIFT
;
291 list_for_each_entry(obj
, &dev_priv
->mm
.bound_list
, global_link
) {
292 if (!i915_gem_object_is_active(obj
) && can_release_pages(obj
))
293 count
+= obj
->base
.size
>> PAGE_SHIFT
;
297 mutex_unlock(&dev
->struct_mutex
);
303 i915_gem_shrinker_scan(struct shrinker
*shrinker
, struct shrink_control
*sc
)
305 struct drm_i915_private
*dev_priv
=
306 container_of(shrinker
, struct drm_i915_private
, mm
.shrinker
);
307 struct drm_device
*dev
= &dev_priv
->drm
;
311 if (!i915_gem_shrinker_lock(dev
, &unlock
))
314 freed
= i915_gem_shrink(dev_priv
,
317 I915_SHRINK_UNBOUND
|
318 I915_SHRINK_PURGEABLE
);
319 if (freed
< sc
->nr_to_scan
)
320 freed
+= i915_gem_shrink(dev_priv
,
321 sc
->nr_to_scan
- freed
,
323 I915_SHRINK_UNBOUND
);
325 mutex_unlock(&dev
->struct_mutex
);
330 struct shrinker_lock_uninterruptible
{
331 bool was_interruptible
;
336 i915_gem_shrinker_lock_uninterruptible(struct drm_i915_private
*dev_priv
,
337 struct shrinker_lock_uninterruptible
*slu
,
340 unsigned long timeout
= jiffies
+ msecs_to_jiffies_timeout(timeout_ms
);
343 if (i915_gem_wait_for_idle(dev_priv
, 0) == 0 &&
344 i915_gem_shrinker_lock(&dev_priv
->drm
, &slu
->unlock
))
347 schedule_timeout_killable(1);
348 if (fatal_signal_pending(current
))
351 if (time_after(jiffies
, timeout
)) {
352 pr_err("Unable to lock GPU to purge memory.\n");
357 slu
->was_interruptible
= dev_priv
->mm
.interruptible
;
358 dev_priv
->mm
.interruptible
= false;
363 i915_gem_shrinker_unlock_uninterruptible(struct drm_i915_private
*dev_priv
,
364 struct shrinker_lock_uninterruptible
*slu
)
366 dev_priv
->mm
.interruptible
= slu
->was_interruptible
;
368 mutex_unlock(&dev_priv
->drm
.struct_mutex
);
372 i915_gem_shrinker_oom(struct notifier_block
*nb
, unsigned long event
, void *ptr
)
374 struct drm_i915_private
*dev_priv
=
375 container_of(nb
, struct drm_i915_private
, mm
.oom_notifier
);
376 struct shrinker_lock_uninterruptible slu
;
377 struct drm_i915_gem_object
*obj
;
378 unsigned long unevictable
, bound
, unbound
, freed_pages
;
380 if (!i915_gem_shrinker_lock_uninterruptible(dev_priv
, &slu
, 5000))
383 intel_runtime_pm_get(dev_priv
);
384 freed_pages
= i915_gem_shrink_all(dev_priv
);
385 intel_runtime_pm_put(dev_priv
);
387 /* Because we may be allocating inside our own driver, we cannot
388 * assert that there are no objects with pinned pages that are not
389 * being pointed to by hardware.
391 unbound
= bound
= unevictable
= 0;
392 list_for_each_entry(obj
, &dev_priv
->mm
.unbound_list
, global_link
) {
396 if (!can_release_pages(obj
))
397 unevictable
+= obj
->base
.size
>> PAGE_SHIFT
;
399 unbound
+= obj
->base
.size
>> PAGE_SHIFT
;
401 list_for_each_entry(obj
, &dev_priv
->mm
.bound_list
, global_link
) {
405 if (!can_release_pages(obj
))
406 unevictable
+= obj
->base
.size
>> PAGE_SHIFT
;
408 bound
+= obj
->base
.size
>> PAGE_SHIFT
;
411 i915_gem_shrinker_unlock_uninterruptible(dev_priv
, &slu
);
413 if (freed_pages
|| unbound
|| bound
)
414 pr_info("Purging GPU memory, %lu pages freed, "
415 "%lu pages still pinned.\n",
416 freed_pages
, unevictable
);
417 if (unbound
|| bound
)
418 pr_err("%lu and %lu pages still available in the "
419 "bound and unbound GPU page lists.\n",
422 *(unsigned long *)ptr
+= freed_pages
;
427 i915_gem_shrinker_vmap(struct notifier_block
*nb
, unsigned long event
, void *ptr
)
429 struct drm_i915_private
*dev_priv
=
430 container_of(nb
, struct drm_i915_private
, mm
.vmap_notifier
);
431 struct shrinker_lock_uninterruptible slu
;
432 struct i915_vma
*vma
, *next
;
433 unsigned long freed_pages
= 0;
436 if (!i915_gem_shrinker_lock_uninterruptible(dev_priv
, &slu
, 5000))
439 /* Force everything onto the inactive lists */
440 ret
= i915_gem_wait_for_idle(dev_priv
, I915_WAIT_LOCKED
);
444 intel_runtime_pm_get(dev_priv
);
445 freed_pages
+= i915_gem_shrink(dev_priv
, -1UL,
447 I915_SHRINK_UNBOUND
|
450 intel_runtime_pm_put(dev_priv
);
452 /* We also want to clear any cached iomaps as they wrap vmap */
453 list_for_each_entry_safe(vma
, next
,
454 &dev_priv
->ggtt
.base
.inactive_list
, vm_link
) {
455 unsigned long count
= vma
->node
.size
>> PAGE_SHIFT
;
456 if (vma
->iomap
&& i915_vma_unbind(vma
) == 0)
457 freed_pages
+= count
;
461 i915_gem_shrinker_unlock_uninterruptible(dev_priv
, &slu
);
463 *(unsigned long *)ptr
+= freed_pages
;
468 * i915_gem_shrinker_init - Initialize i915 shrinker
469 * @dev_priv: i915 device
471 * This function registers and sets up the i915 shrinker and OOM handler.
473 void i915_gem_shrinker_init(struct drm_i915_private
*dev_priv
)
475 dev_priv
->mm
.shrinker
.scan_objects
= i915_gem_shrinker_scan
;
476 dev_priv
->mm
.shrinker
.count_objects
= i915_gem_shrinker_count
;
477 dev_priv
->mm
.shrinker
.seeks
= DEFAULT_SEEKS
;
478 WARN_ON(register_shrinker(&dev_priv
->mm
.shrinker
));
480 dev_priv
->mm
.oom_notifier
.notifier_call
= i915_gem_shrinker_oom
;
481 WARN_ON(register_oom_notifier(&dev_priv
->mm
.oom_notifier
));
483 dev_priv
->mm
.vmap_notifier
.notifier_call
= i915_gem_shrinker_vmap
;
484 WARN_ON(register_vmap_purge_notifier(&dev_priv
->mm
.vmap_notifier
));
488 * i915_gem_shrinker_cleanup - Clean up i915 shrinker
489 * @dev_priv: i915 device
491 * This function unregisters the i915 shrinker and OOM handler.
493 void i915_gem_shrinker_cleanup(struct drm_i915_private
*dev_priv
)
495 WARN_ON(unregister_vmap_purge_notifier(&dev_priv
->mm
.vmap_notifier
));
496 WARN_ON(unregister_oom_notifier(&dev_priv
->mm
.oom_notifier
));
497 unregister_shrinker(&dev_priv
->mm
.shrinker
);