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 mutex_is_locked_by(struct mutex
*mutex
, struct task_struct
*task
)
40 if (!mutex_is_locked(mutex
))
43 #if defined(CONFIG_DEBUG_MUTEXES) || defined(CONFIG_MUTEX_SPIN_ON_OWNER)
44 return mutex
->owner
== task
;
46 /* Since UP may be pre-empted, we cannot assume that we own the lock */
51 static bool any_vma_pinned(struct drm_i915_gem_object
*obj
)
55 list_for_each_entry(vma
, &obj
->vma_list
, obj_link
)
56 if (i915_vma_is_pinned(vma
))
62 static bool swap_available(void)
64 return get_nr_swap_pages() > 0;
67 static bool can_release_pages(struct drm_i915_gem_object
*obj
)
69 /* Only shmemfs objects are backed by swap */
73 /* Only report true if by unbinding the object and putting its pages
74 * we can actually make forward progress towards freeing physical
77 * If the pages are pinned for any other reason than being bound
78 * to the GPU, simply unbinding from the GPU is not going to succeed
79 * in releasing our pin count on the pages themselves.
81 if (obj
->pages_pin_count
> obj
->bind_count
)
84 if (any_vma_pinned(obj
))
87 /* We can only return physical pages to the system if we can either
88 * discard the contents (because the user has marked them as being
89 * purgeable) or if we can move their contents out to swap.
91 return swap_available() || obj
->madv
== I915_MADV_DONTNEED
;
95 * i915_gem_shrink - Shrink buffer object caches
96 * @dev_priv: i915 device
97 * @target: amount of memory to make available, in pages
98 * @flags: control flags for selecting cache types
100 * This function is the main interface to the shrinker. It will try to release
101 * up to @target pages of main memory backing storage from buffer objects.
102 * Selection of the specific caches can be done with @flags. This is e.g. useful
103 * when purgeable objects should be removed from caches preferentially.
105 * Note that it's not guaranteed that released amount is actually available as
106 * free system memory - the pages might still be in-used to due to other reasons
107 * (like cpu mmaps) or the mm core has reused them before we could grab them.
108 * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
109 * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
111 * Also note that any kind of pinning (both per-vma address space pins and
112 * backing storage pins at the buffer object level) result in the shrinker code
113 * having to skip the object.
116 * The number of pages of backing storage actually released.
119 i915_gem_shrink(struct drm_i915_private
*dev_priv
,
120 unsigned long target
, unsigned flags
)
123 struct list_head
*list
;
126 { &dev_priv
->mm
.unbound_list
, I915_SHRINK_UNBOUND
},
127 { &dev_priv
->mm
.bound_list
, I915_SHRINK_BOUND
},
130 unsigned long count
= 0;
132 trace_i915_gem_shrink(dev_priv
, target
, flags
);
133 i915_gem_retire_requests(dev_priv
);
136 * Unbinding of objects will require HW access; Let us not wake the
137 * device just to recover a little memory. If absolutely necessary,
138 * we will force the wake during oom-notifier.
140 if ((flags
& I915_SHRINK_BOUND
) &&
141 !intel_runtime_pm_get_if_in_use(dev_priv
))
142 flags
&= ~I915_SHRINK_BOUND
;
145 * As we may completely rewrite the (un)bound list whilst unbinding
146 * (due to retiring requests) we have to strictly process only
147 * one element of the list at the time, and recheck the list
148 * on every iteration.
150 * In particular, we must hold a reference whilst removing the
151 * object as we may end up waiting for and/or retiring the objects.
152 * This might release the final reference (held by the active list)
153 * and result in the object being freed from under us. This is
154 * similar to the precautions the eviction code must take whilst
157 * Also note that although these lists do not hold a reference to
158 * the object we can safely grab one here: The final object
159 * unreferencing and the bound_list are both protected by the
160 * dev->struct_mutex and so we won't ever be able to observe an
161 * object on the bound_list with a reference count equals 0.
163 for (phase
= phases
; phase
->list
; phase
++) {
164 struct list_head still_in_list
;
165 struct drm_i915_gem_object
*obj
;
167 if ((flags
& phase
->bit
) == 0)
170 INIT_LIST_HEAD(&still_in_list
);
171 while (count
< target
&&
172 (obj
= list_first_entry_or_null(phase
->list
,
175 list_move_tail(&obj
->global_list
, &still_in_list
);
177 if (flags
& I915_SHRINK_PURGEABLE
&&
178 obj
->madv
!= I915_MADV_DONTNEED
)
181 if (flags
& I915_SHRINK_VMAPS
&&
182 !is_vmalloc_addr(obj
->mapping
))
185 if ((flags
& I915_SHRINK_ACTIVE
) == 0 &&
186 i915_gem_object_is_active(obj
))
189 if (!can_release_pages(obj
))
192 i915_gem_object_get(obj
);
194 /* For the unbound phase, this should be a no-op! */
195 i915_gem_object_unbind(obj
);
196 if (i915_gem_object_put_pages(obj
) == 0)
197 count
+= obj
->base
.size
>> PAGE_SHIFT
;
199 i915_gem_object_put(obj
);
201 list_splice(&still_in_list
, phase
->list
);
204 if (flags
& I915_SHRINK_BOUND
)
205 intel_runtime_pm_put(dev_priv
);
207 i915_gem_retire_requests(dev_priv
);
208 /* expedite the RCU grace period to free some request slabs */
209 synchronize_rcu_expedited();
215 * i915_gem_shrink_all - Shrink buffer object caches completely
216 * @dev_priv: i915 device
218 * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
219 * caches completely. It also first waits for and retires all outstanding
220 * requests to also be able to release backing storage for active objects.
222 * This should only be used in code to intentionally quiescent the gpu or as a
223 * last-ditch effort when memory seems to have run out.
226 * The number of pages of backing storage actually released.
228 unsigned long i915_gem_shrink_all(struct drm_i915_private
*dev_priv
)
232 freed
= i915_gem_shrink(dev_priv
, -1UL,
234 I915_SHRINK_UNBOUND
|
236 rcu_barrier(); /* wait until our RCU delayed slab frees are completed */
241 static bool i915_gem_shrinker_lock(struct drm_device
*dev
, bool *unlock
)
243 if (!mutex_trylock(&dev
->struct_mutex
)) {
244 if (!mutex_is_locked_by(&dev
->struct_mutex
, current
))
255 i915_gem_shrinker_count(struct shrinker
*shrinker
, struct shrink_control
*sc
)
257 struct drm_i915_private
*dev_priv
=
258 container_of(shrinker
, struct drm_i915_private
, mm
.shrinker
);
259 struct drm_device
*dev
= &dev_priv
->drm
;
260 struct drm_i915_gem_object
*obj
;
264 if (!i915_gem_shrinker_lock(dev
, &unlock
))
267 i915_gem_retire_requests(dev_priv
);
270 list_for_each_entry(obj
, &dev_priv
->mm
.unbound_list
, global_list
)
271 if (can_release_pages(obj
))
272 count
+= obj
->base
.size
>> PAGE_SHIFT
;
274 list_for_each_entry(obj
, &dev_priv
->mm
.bound_list
, global_list
) {
275 if (!i915_gem_object_is_active(obj
) && can_release_pages(obj
))
276 count
+= obj
->base
.size
>> PAGE_SHIFT
;
280 mutex_unlock(&dev
->struct_mutex
);
286 i915_gem_shrinker_scan(struct shrinker
*shrinker
, struct shrink_control
*sc
)
288 struct drm_i915_private
*dev_priv
=
289 container_of(shrinker
, struct drm_i915_private
, mm
.shrinker
);
290 struct drm_device
*dev
= &dev_priv
->drm
;
294 if (!i915_gem_shrinker_lock(dev
, &unlock
))
297 freed
= i915_gem_shrink(dev_priv
,
300 I915_SHRINK_UNBOUND
|
301 I915_SHRINK_PURGEABLE
);
302 if (freed
< sc
->nr_to_scan
)
303 freed
+= i915_gem_shrink(dev_priv
,
304 sc
->nr_to_scan
- freed
,
306 I915_SHRINK_UNBOUND
);
308 mutex_unlock(&dev
->struct_mutex
);
313 struct shrinker_lock_uninterruptible
{
314 bool was_interruptible
;
319 i915_gem_shrinker_lock_uninterruptible(struct drm_i915_private
*dev_priv
,
320 struct shrinker_lock_uninterruptible
*slu
,
323 unsigned long timeout
= jiffies
+ msecs_to_jiffies_timeout(timeout_ms
);
326 if (i915_gem_wait_for_idle(dev_priv
, false) == 0 &&
327 i915_gem_shrinker_lock(&dev_priv
->drm
, &slu
->unlock
))
330 schedule_timeout_killable(1);
331 if (fatal_signal_pending(current
))
334 if (time_after(jiffies
, timeout
)) {
335 pr_err("Unable to lock GPU to purge memory.\n");
340 slu
->was_interruptible
= dev_priv
->mm
.interruptible
;
341 dev_priv
->mm
.interruptible
= false;
346 i915_gem_shrinker_unlock_uninterruptible(struct drm_i915_private
*dev_priv
,
347 struct shrinker_lock_uninterruptible
*slu
)
349 dev_priv
->mm
.interruptible
= slu
->was_interruptible
;
351 mutex_unlock(&dev_priv
->drm
.struct_mutex
);
355 i915_gem_shrinker_oom(struct notifier_block
*nb
, unsigned long event
, void *ptr
)
357 struct drm_i915_private
*dev_priv
=
358 container_of(nb
, struct drm_i915_private
, mm
.oom_notifier
);
359 struct shrinker_lock_uninterruptible slu
;
360 struct drm_i915_gem_object
*obj
;
361 unsigned long unevictable
, bound
, unbound
, freed_pages
;
363 if (!i915_gem_shrinker_lock_uninterruptible(dev_priv
, &slu
, 5000))
366 intel_runtime_pm_get(dev_priv
);
367 freed_pages
= i915_gem_shrink_all(dev_priv
);
368 intel_runtime_pm_put(dev_priv
);
370 /* Because we may be allocating inside our own driver, we cannot
371 * assert that there are no objects with pinned pages that are not
372 * being pointed to by hardware.
374 unbound
= bound
= unevictable
= 0;
375 list_for_each_entry(obj
, &dev_priv
->mm
.unbound_list
, global_list
) {
376 if (!can_release_pages(obj
))
377 unevictable
+= obj
->base
.size
>> PAGE_SHIFT
;
379 unbound
+= obj
->base
.size
>> PAGE_SHIFT
;
381 list_for_each_entry(obj
, &dev_priv
->mm
.bound_list
, global_list
) {
382 if (!can_release_pages(obj
))
383 unevictable
+= obj
->base
.size
>> PAGE_SHIFT
;
385 bound
+= obj
->base
.size
>> PAGE_SHIFT
;
388 i915_gem_shrinker_unlock_uninterruptible(dev_priv
, &slu
);
390 if (freed_pages
|| unbound
|| bound
)
391 pr_info("Purging GPU memory, %lu pages freed, "
392 "%lu pages still pinned.\n",
393 freed_pages
, unevictable
);
394 if (unbound
|| bound
)
395 pr_err("%lu and %lu pages still available in the "
396 "bound and unbound GPU page lists.\n",
399 *(unsigned long *)ptr
+= freed_pages
;
404 i915_gem_shrinker_vmap(struct notifier_block
*nb
, unsigned long event
, void *ptr
)
406 struct drm_i915_private
*dev_priv
=
407 container_of(nb
, struct drm_i915_private
, mm
.vmap_notifier
);
408 struct shrinker_lock_uninterruptible slu
;
409 struct i915_vma
*vma
, *next
;
410 unsigned long freed_pages
= 0;
413 if (!i915_gem_shrinker_lock_uninterruptible(dev_priv
, &slu
, 5000))
416 /* Force everything onto the inactive lists */
417 ret
= i915_gem_wait_for_idle(dev_priv
, false);
421 intel_runtime_pm_get(dev_priv
);
422 freed_pages
+= i915_gem_shrink(dev_priv
, -1UL,
424 I915_SHRINK_UNBOUND
|
427 intel_runtime_pm_put(dev_priv
);
429 /* We also want to clear any cached iomaps as they wrap vmap */
430 list_for_each_entry_safe(vma
, next
,
431 &dev_priv
->ggtt
.base
.inactive_list
, vm_link
) {
432 unsigned long count
= vma
->node
.size
>> PAGE_SHIFT
;
433 if (vma
->iomap
&& i915_vma_unbind(vma
) == 0)
434 freed_pages
+= count
;
438 i915_gem_shrinker_unlock_uninterruptible(dev_priv
, &slu
);
440 *(unsigned long *)ptr
+= freed_pages
;
445 * i915_gem_shrinker_init - Initialize i915 shrinker
446 * @dev_priv: i915 device
448 * This function registers and sets up the i915 shrinker and OOM handler.
450 void i915_gem_shrinker_init(struct drm_i915_private
*dev_priv
)
452 dev_priv
->mm
.shrinker
.scan_objects
= i915_gem_shrinker_scan
;
453 dev_priv
->mm
.shrinker
.count_objects
= i915_gem_shrinker_count
;
454 dev_priv
->mm
.shrinker
.seeks
= DEFAULT_SEEKS
;
455 WARN_ON(register_shrinker(&dev_priv
->mm
.shrinker
));
457 dev_priv
->mm
.oom_notifier
.notifier_call
= i915_gem_shrinker_oom
;
458 WARN_ON(register_oom_notifier(&dev_priv
->mm
.oom_notifier
));
460 dev_priv
->mm
.vmap_notifier
.notifier_call
= i915_gem_shrinker_vmap
;
461 WARN_ON(register_vmap_purge_notifier(&dev_priv
->mm
.vmap_notifier
));
465 * i915_gem_shrinker_cleanup - Clean up i915 shrinker
466 * @dev_priv: i915 device
468 * This function unregisters the i915 shrinker and OOM handler.
470 void i915_gem_shrinker_cleanup(struct drm_i915_private
*dev_priv
)
472 WARN_ON(unregister_vmap_purge_notifier(&dev_priv
->mm
.vmap_notifier
));
473 WARN_ON(unregister_oom_notifier(&dev_priv
->mm
.oom_notifier
));
474 unregister_shrinker(&dev_priv
->mm
.shrinker
);