]>
git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - drivers/gpu/drm/i915/i915_gem_shrinker.c
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 shrinker_lock(struct drm_i915_private
*dev_priv
, bool *unlock
)
40 switch (mutex_trylock_recursive(&dev_priv
->drm
.struct_mutex
)) {
41 case MUTEX_TRYLOCK_RECURSIVE
:
45 case MUTEX_TRYLOCK_FAILED
:
50 if (mutex_trylock(&dev_priv
->drm
.struct_mutex
)) {
54 } while (!need_resched());
58 case MUTEX_TRYLOCK_SUCCESS
:
66 static void shrinker_unlock(struct drm_i915_private
*dev_priv
, bool unlock
)
71 mutex_unlock(&dev_priv
->drm
.struct_mutex
);
74 static bool any_vma_pinned(struct drm_i915_gem_object
*obj
)
78 list_for_each_entry(vma
, &obj
->vma_list
, obj_link
) {
79 /* Only GGTT vma may be permanently pinned, and are always
80 * at the start of the list. We can stop hunting as soon
81 * as we see a ppGTT vma.
83 if (!i915_vma_is_ggtt(vma
))
86 if (i915_vma_is_pinned(vma
))
93 static bool swap_available(void)
95 return get_nr_swap_pages() > 0;
98 static bool can_release_pages(struct drm_i915_gem_object
*obj
)
103 /* Consider only shrinkable ojects. */
104 if (!i915_gem_object_is_shrinkable(obj
))
107 /* Only report true if by unbinding the object and putting its pages
108 * we can actually make forward progress towards freeing physical
111 * If the pages are pinned for any other reason than being bound
112 * to the GPU, simply unbinding from the GPU is not going to succeed
113 * in releasing our pin count on the pages themselves.
115 if (atomic_read(&obj
->mm
.pages_pin_count
) > obj
->bind_count
)
118 if (any_vma_pinned(obj
))
121 /* We can only return physical pages to the system if we can either
122 * discard the contents (because the user has marked them as being
123 * purgeable) or if we can move their contents out to swap.
125 return swap_available() || obj
->mm
.madv
== I915_MADV_DONTNEED
;
128 static bool unsafe_drop_pages(struct drm_i915_gem_object
*obj
)
130 if (i915_gem_object_unbind(obj
) == 0)
131 __i915_gem_object_put_pages(obj
, I915_MM_SHRINKER
);
132 return !READ_ONCE(obj
->mm
.pages
);
136 * i915_gem_shrink - Shrink buffer object caches
137 * @dev_priv: i915 device
138 * @target: amount of memory to make available, in pages
139 * @nr_scanned: optional output for number of pages scanned (incremental)
140 * @flags: control flags for selecting cache types
142 * This function is the main interface to the shrinker. It will try to release
143 * up to @target pages of main memory backing storage from buffer objects.
144 * Selection of the specific caches can be done with @flags. This is e.g. useful
145 * when purgeable objects should be removed from caches preferentially.
147 * Note that it's not guaranteed that released amount is actually available as
148 * free system memory - the pages might still be in-used to due to other reasons
149 * (like cpu mmaps) or the mm core has reused them before we could grab them.
150 * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
151 * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
153 * Also note that any kind of pinning (both per-vma address space pins and
154 * backing storage pins at the buffer object level) result in the shrinker code
155 * having to skip the object.
158 * The number of pages of backing storage actually released.
161 i915_gem_shrink(struct drm_i915_private
*dev_priv
,
162 unsigned long target
,
163 unsigned long *nr_scanned
,
167 struct list_head
*list
;
170 { &dev_priv
->mm
.unbound_list
, I915_SHRINK_UNBOUND
},
171 { &dev_priv
->mm
.bound_list
, I915_SHRINK_BOUND
},
174 unsigned long count
= 0;
175 unsigned long scanned
= 0;
178 if (!shrinker_lock(dev_priv
, &unlock
))
181 trace_i915_gem_shrink(dev_priv
, target
, flags
);
182 i915_gem_retire_requests(dev_priv
);
185 * Unbinding of objects will require HW access; Let us not wake the
186 * device just to recover a little memory. If absolutely necessary,
187 * we will force the wake during oom-notifier.
189 if ((flags
& I915_SHRINK_BOUND
) &&
190 !intel_runtime_pm_get_if_in_use(dev_priv
))
191 flags
&= ~I915_SHRINK_BOUND
;
194 * As we may completely rewrite the (un)bound list whilst unbinding
195 * (due to retiring requests) we have to strictly process only
196 * one element of the list at the time, and recheck the list
197 * on every iteration.
199 * In particular, we must hold a reference whilst removing the
200 * object as we may end up waiting for and/or retiring the objects.
201 * This might release the final reference (held by the active list)
202 * and result in the object being freed from under us. This is
203 * similar to the precautions the eviction code must take whilst
206 * Also note that although these lists do not hold a reference to
207 * the object we can safely grab one here: The final object
208 * unreferencing and the bound_list are both protected by the
209 * dev->struct_mutex and so we won't ever be able to observe an
210 * object on the bound_list with a reference count equals 0.
212 for (phase
= phases
; phase
->list
; phase
++) {
213 struct list_head still_in_list
;
214 struct drm_i915_gem_object
*obj
;
216 if ((flags
& phase
->bit
) == 0)
219 INIT_LIST_HEAD(&still_in_list
);
220 while (count
< target
&&
221 (obj
= list_first_entry_or_null(phase
->list
,
224 list_move_tail(&obj
->global_link
, &still_in_list
);
225 if (!obj
->mm
.pages
) {
226 list_del_init(&obj
->global_link
);
230 if (flags
& I915_SHRINK_PURGEABLE
&&
231 obj
->mm
.madv
!= I915_MADV_DONTNEED
)
234 if (flags
& I915_SHRINK_VMAPS
&&
235 !is_vmalloc_addr(obj
->mm
.mapping
))
238 if (!(flags
& I915_SHRINK_ACTIVE
) &&
239 (i915_gem_object_is_active(obj
) ||
240 i915_gem_object_is_framebuffer(obj
)))
243 if (!can_release_pages(obj
))
246 if (unsafe_drop_pages(obj
)) {
247 /* May arrive from get_pages on another bo */
248 mutex_lock_nested(&obj
->mm
.lock
,
250 if (!obj
->mm
.pages
) {
251 __i915_gem_object_invalidate(obj
);
252 list_del_init(&obj
->global_link
);
253 count
+= obj
->base
.size
>> PAGE_SHIFT
;
255 mutex_unlock(&obj
->mm
.lock
);
256 scanned
+= obj
->base
.size
>> PAGE_SHIFT
;
259 list_splice_tail(&still_in_list
, phase
->list
);
262 if (flags
& I915_SHRINK_BOUND
)
263 intel_runtime_pm_put(dev_priv
);
265 i915_gem_retire_requests(dev_priv
);
267 shrinker_unlock(dev_priv
, unlock
);
270 *nr_scanned
+= scanned
;
275 * i915_gem_shrink_all - Shrink buffer object caches completely
276 * @dev_priv: i915 device
278 * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
279 * caches completely. It also first waits for and retires all outstanding
280 * requests to also be able to release backing storage for active objects.
282 * This should only be used in code to intentionally quiescent the gpu or as a
283 * last-ditch effort when memory seems to have run out.
286 * The number of pages of backing storage actually released.
288 unsigned long i915_gem_shrink_all(struct drm_i915_private
*dev_priv
)
292 intel_runtime_pm_get(dev_priv
);
293 freed
= i915_gem_shrink(dev_priv
, -1UL, NULL
,
295 I915_SHRINK_UNBOUND
|
297 intel_runtime_pm_put(dev_priv
);
303 i915_gem_shrinker_count(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_i915_gem_object
*obj
;
311 if (!shrinker_lock(dev_priv
, &unlock
))
314 i915_gem_retire_requests(dev_priv
);
317 list_for_each_entry(obj
, &dev_priv
->mm
.unbound_list
, global_link
)
318 if (can_release_pages(obj
))
319 count
+= obj
->base
.size
>> PAGE_SHIFT
;
321 list_for_each_entry(obj
, &dev_priv
->mm
.bound_list
, global_link
) {
322 if (!i915_gem_object_is_active(obj
) && can_release_pages(obj
))
323 count
+= obj
->base
.size
>> PAGE_SHIFT
;
326 shrinker_unlock(dev_priv
, unlock
);
332 i915_gem_shrinker_scan(struct shrinker
*shrinker
, struct shrink_control
*sc
)
334 struct drm_i915_private
*dev_priv
=
335 container_of(shrinker
, struct drm_i915_private
, mm
.shrinker
);
341 if (!shrinker_lock(dev_priv
, &unlock
))
344 freed
= i915_gem_shrink(dev_priv
,
348 I915_SHRINK_UNBOUND
|
349 I915_SHRINK_PURGEABLE
);
350 if (freed
< sc
->nr_to_scan
)
351 freed
+= i915_gem_shrink(dev_priv
,
352 sc
->nr_to_scan
- sc
->nr_scanned
,
355 I915_SHRINK_UNBOUND
);
356 if (freed
< sc
->nr_to_scan
&& current_is_kswapd()) {
357 intel_runtime_pm_get(dev_priv
);
358 freed
+= i915_gem_shrink(dev_priv
,
359 sc
->nr_to_scan
- sc
->nr_scanned
,
363 I915_SHRINK_UNBOUND
);
364 intel_runtime_pm_put(dev_priv
);
367 shrinker_unlock(dev_priv
, unlock
);
369 return sc
->nr_scanned
? freed
: SHRINK_STOP
;
373 shrinker_lock_uninterruptible(struct drm_i915_private
*dev_priv
, bool *unlock
,
376 unsigned long timeout
= jiffies
+ msecs_to_jiffies_timeout(timeout_ms
);
379 if (i915_gem_wait_for_idle(dev_priv
, 0) == 0 &&
380 shrinker_lock(dev_priv
, unlock
))
383 schedule_timeout_killable(1);
384 if (fatal_signal_pending(current
))
387 if (time_after(jiffies
, timeout
)) {
388 pr_err("Unable to lock GPU to purge memory.\n");
397 i915_gem_shrinker_oom(struct notifier_block
*nb
, unsigned long event
, void *ptr
)
399 struct drm_i915_private
*dev_priv
=
400 container_of(nb
, struct drm_i915_private
, mm
.oom_notifier
);
401 struct drm_i915_gem_object
*obj
;
402 unsigned long unevictable
, bound
, unbound
, freed_pages
;
405 if (!shrinker_lock_uninterruptible(dev_priv
, &unlock
, 5000))
408 freed_pages
= i915_gem_shrink_all(dev_priv
);
410 /* Because we may be allocating inside our own driver, we cannot
411 * assert that there are no objects with pinned pages that are not
412 * being pointed to by hardware.
414 unbound
= bound
= unevictable
= 0;
415 list_for_each_entry(obj
, &dev_priv
->mm
.unbound_list
, global_link
) {
419 if (!can_release_pages(obj
))
420 unevictable
+= obj
->base
.size
>> PAGE_SHIFT
;
422 unbound
+= obj
->base
.size
>> PAGE_SHIFT
;
424 list_for_each_entry(obj
, &dev_priv
->mm
.bound_list
, global_link
) {
428 if (!can_release_pages(obj
))
429 unevictable
+= obj
->base
.size
>> PAGE_SHIFT
;
431 bound
+= obj
->base
.size
>> PAGE_SHIFT
;
434 shrinker_unlock(dev_priv
, unlock
);
436 if (freed_pages
|| unbound
|| bound
)
437 pr_info("Purging GPU memory, %lu pages freed, "
438 "%lu pages still pinned.\n",
439 freed_pages
, unevictable
);
440 if (unbound
|| bound
)
441 pr_err("%lu and %lu pages still available in the "
442 "bound and unbound GPU page lists.\n",
445 *(unsigned long *)ptr
+= freed_pages
;
450 i915_gem_shrinker_vmap(struct notifier_block
*nb
, unsigned long event
, void *ptr
)
452 struct drm_i915_private
*dev_priv
=
453 container_of(nb
, struct drm_i915_private
, mm
.vmap_notifier
);
454 struct i915_vma
*vma
, *next
;
455 unsigned long freed_pages
= 0;
459 if (!shrinker_lock_uninterruptible(dev_priv
, &unlock
, 5000))
462 /* Force everything onto the inactive lists */
463 ret
= i915_gem_wait_for_idle(dev_priv
, I915_WAIT_LOCKED
);
467 intel_runtime_pm_get(dev_priv
);
468 freed_pages
+= i915_gem_shrink(dev_priv
, -1UL, NULL
,
470 I915_SHRINK_UNBOUND
|
473 intel_runtime_pm_put(dev_priv
);
475 /* We also want to clear any cached iomaps as they wrap vmap */
476 list_for_each_entry_safe(vma
, next
,
477 &dev_priv
->ggtt
.base
.inactive_list
, vm_link
) {
478 unsigned long count
= vma
->node
.size
>> PAGE_SHIFT
;
479 if (vma
->iomap
&& i915_vma_unbind(vma
) == 0)
480 freed_pages
+= count
;
484 shrinker_unlock(dev_priv
, unlock
);
486 *(unsigned long *)ptr
+= freed_pages
;
491 * i915_gem_shrinker_init - Initialize i915 shrinker
492 * @dev_priv: i915 device
494 * This function registers and sets up the i915 shrinker and OOM handler.
496 void i915_gem_shrinker_init(struct drm_i915_private
*dev_priv
)
498 dev_priv
->mm
.shrinker
.scan_objects
= i915_gem_shrinker_scan
;
499 dev_priv
->mm
.shrinker
.count_objects
= i915_gem_shrinker_count
;
500 dev_priv
->mm
.shrinker
.seeks
= DEFAULT_SEEKS
;
501 WARN_ON(register_shrinker(&dev_priv
->mm
.shrinker
));
503 dev_priv
->mm
.oom_notifier
.notifier_call
= i915_gem_shrinker_oom
;
504 WARN_ON(register_oom_notifier(&dev_priv
->mm
.oom_notifier
));
506 dev_priv
->mm
.vmap_notifier
.notifier_call
= i915_gem_shrinker_vmap
;
507 WARN_ON(register_vmap_purge_notifier(&dev_priv
->mm
.vmap_notifier
));
511 * i915_gem_shrinker_cleanup - Clean up i915 shrinker
512 * @dev_priv: i915 device
514 * This function unregisters the i915 shrinker and OOM handler.
516 void i915_gem_shrinker_cleanup(struct drm_i915_private
*dev_priv
)
518 WARN_ON(unregister_vmap_purge_notifier(&dev_priv
->mm
.vmap_notifier
));
519 WARN_ON(unregister_oom_notifier(&dev_priv
->mm
.oom_notifier
));
520 unregister_shrinker(&dev_priv
->mm
.shrinker
);