drm/i915: Use radixtree to jump start intel_partial_pages()
[linux-block.git] / drivers / gpu / drm / i915 / i915_gem_shrinker.c
CommitLineData
be6a0376
DV
1/*
2 * Copyright © 2008-2015 Intel Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 *
23 */
24
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>
e87666b5 31#include <linux/vmalloc.h>
be6a0376
DV
32#include <drm/drmP.h>
33#include <drm/i915_drm.h>
34
35#include "i915_drv.h"
36#include "i915_trace.h"
37
38static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
39{
40 if (!mutex_is_locked(mutex))
41 return false;
42
4f074a53 43#if defined(CONFIG_DEBUG_MUTEXES) || defined(CONFIG_MUTEX_SPIN_ON_OWNER)
be6a0376
DV
44 return mutex->owner == task;
45#else
46 /* Since UP may be pre-empted, we cannot assume that we own the lock */
47 return false;
48#endif
49}
50
15717de2 51static bool any_vma_pinned(struct drm_i915_gem_object *obj)
c1a415e2
CW
52{
53 struct i915_vma *vma;
c1a415e2 54
15717de2 55 list_for_each_entry(vma, &obj->vma_list, obj_link)
3272db53 56 if (i915_vma_is_pinned(vma))
15717de2 57 return true;
c1a415e2 58
15717de2 59 return false;
c1a415e2
CW
60}
61
62static bool swap_available(void)
63{
64 return get_nr_swap_pages() > 0;
65}
66
67static bool can_release_pages(struct drm_i915_gem_object *obj)
68{
1bec9b0b
CW
69 /* Only shmemfs objects are backed by swap */
70 if (!obj->base.filp)
71 return false;
72
c1a415e2
CW
73 /* Only report true if by unbinding the object and putting its pages
74 * we can actually make forward progress towards freeing physical
75 * pages.
76 *
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.
80 */
15717de2
CW
81 if (obj->pages_pin_count > obj->bind_count)
82 return false;
83
84 if (any_vma_pinned(obj))
c1a415e2
CW
85 return false;
86
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.
90 */
91 return swap_available() || obj->madv == I915_MADV_DONTNEED;
92}
93
eb0b44ad
DV
94/**
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
99 *
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.
104 *
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().
110 *
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.
114 *
115 * Returns:
116 * The number of pages of backing storage actually released.
117 */
be6a0376
DV
118unsigned long
119i915_gem_shrink(struct drm_i915_private *dev_priv,
14387540 120 unsigned long target, unsigned flags)
be6a0376
DV
121{
122 const struct {
123 struct list_head *list;
124 unsigned int bit;
125 } phases[] = {
126 { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
127 { &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
128 { NULL, 0 },
129 }, *phase;
130 unsigned long count = 0;
131
3abafa53 132 trace_i915_gem_shrink(dev_priv, target, flags);
c033666a 133 i915_gem_retire_requests(dev_priv);
3abafa53 134
178a30c9
PP
135 /*
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.
139 */
140 if ((flags & I915_SHRINK_BOUND) &&
141 !intel_runtime_pm_get_if_in_use(dev_priv))
142 flags &= ~I915_SHRINK_BOUND;
143
be6a0376
DV
144 /*
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.
149 *
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
155 * removing objects.
156 *
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.
162 */
163 for (phase = phases; phase->list; phase++) {
164 struct list_head still_in_list;
2a1d7752 165 struct drm_i915_gem_object *obj;
be6a0376
DV
166
167 if ((flags & phase->bit) == 0)
168 continue;
169
170 INIT_LIST_HEAD(&still_in_list);
2a1d7752
CW
171 while (count < target &&
172 (obj = list_first_entry_or_null(phase->list,
173 typeof(*obj),
174 global_list))) {
be6a0376
DV
175 list_move_tail(&obj->global_list, &still_in_list);
176
177 if (flags & I915_SHRINK_PURGEABLE &&
178 obj->madv != I915_MADV_DONTNEED)
179 continue;
180
eae2c43b
CW
181 if (flags & I915_SHRINK_VMAPS &&
182 !is_vmalloc_addr(obj->mapping))
183 continue;
184
45353ce5
CW
185 if (!(flags & I915_SHRINK_ACTIVE) &&
186 (i915_gem_object_is_active(obj) ||
187 obj->framebuffer_references))
5763ff04
CW
188 continue;
189
c1a415e2
CW
190 if (!can_release_pages(obj))
191 continue;
192
25dc556a 193 i915_gem_object_get(obj);
be6a0376
DV
194
195 /* For the unbound phase, this should be a no-op! */
aa653a68 196 i915_gem_object_unbind(obj);
be6a0376
DV
197 if (i915_gem_object_put_pages(obj) == 0)
198 count += obj->base.size >> PAGE_SHIFT;
199
f8c417cd 200 i915_gem_object_put(obj);
be6a0376
DV
201 }
202 list_splice(&still_in_list, phase->list);
203 }
204
178a30c9
PP
205 if (flags & I915_SHRINK_BOUND)
206 intel_runtime_pm_put(dev_priv);
207
c033666a 208 i915_gem_retire_requests(dev_priv);
0eafec6d
CW
209 /* expedite the RCU grace period to free some request slabs */
210 synchronize_rcu_expedited();
c9c0f5ea 211
be6a0376
DV
212 return count;
213}
214
eb0b44ad 215/**
1f2449cd 216 * i915_gem_shrink_all - Shrink buffer object caches completely
eb0b44ad
DV
217 * @dev_priv: i915 device
218 *
219 * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
220 * caches completely. It also first waits for and retires all outstanding
221 * requests to also be able to release backing storage for active objects.
222 *
223 * This should only be used in code to intentionally quiescent the gpu or as a
224 * last-ditch effort when memory seems to have run out.
225 *
226 * Returns:
227 * The number of pages of backing storage actually released.
228 */
be6a0376
DV
229unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv)
230{
0eafec6d
CW
231 unsigned long freed;
232
233 freed = i915_gem_shrink(dev_priv, -1UL,
234 I915_SHRINK_BOUND |
235 I915_SHRINK_UNBOUND |
236 I915_SHRINK_ACTIVE);
237 rcu_barrier(); /* wait until our RCU delayed slab frees are completed */
238
239 return freed;
be6a0376
DV
240}
241
242static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock)
243{
244 if (!mutex_trylock(&dev->struct_mutex)) {
245 if (!mutex_is_locked_by(&dev->struct_mutex, current))
246 return false;
247
be6a0376
DV
248 *unlock = false;
249 } else
250 *unlock = true;
251
252 return true;
253}
254
be6a0376
DV
255static unsigned long
256i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
257{
258 struct drm_i915_private *dev_priv =
259 container_of(shrinker, struct drm_i915_private, mm.shrinker);
91c8a326 260 struct drm_device *dev = &dev_priv->drm;
be6a0376
DV
261 struct drm_i915_gem_object *obj;
262 unsigned long count;
263 bool unlock;
264
265 if (!i915_gem_shrinker_lock(dev, &unlock))
266 return 0;
267
bed50aea
CW
268 i915_gem_retire_requests(dev_priv);
269
be6a0376
DV
270 count = 0;
271 list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list)
6f0ac204 272 if (can_release_pages(obj))
be6a0376
DV
273 count += obj->base.size >> PAGE_SHIFT;
274
275 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
573adb39 276 if (!i915_gem_object_is_active(obj) && can_release_pages(obj))
be6a0376
DV
277 count += obj->base.size >> PAGE_SHIFT;
278 }
279
280 if (unlock)
281 mutex_unlock(&dev->struct_mutex);
282
283 return count;
284}
285
286static unsigned long
287i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
288{
289 struct drm_i915_private *dev_priv =
290 container_of(shrinker, struct drm_i915_private, mm.shrinker);
91c8a326 291 struct drm_device *dev = &dev_priv->drm;
be6a0376
DV
292 unsigned long freed;
293 bool unlock;
294
295 if (!i915_gem_shrinker_lock(dev, &unlock))
296 return SHRINK_STOP;
297
298 freed = i915_gem_shrink(dev_priv,
299 sc->nr_to_scan,
300 I915_SHRINK_BOUND |
301 I915_SHRINK_UNBOUND |
302 I915_SHRINK_PURGEABLE);
303 if (freed < sc->nr_to_scan)
304 freed += i915_gem_shrink(dev_priv,
305 sc->nr_to_scan - freed,
306 I915_SHRINK_BOUND |
307 I915_SHRINK_UNBOUND);
308 if (unlock)
309 mutex_unlock(&dev->struct_mutex);
310
311 return freed;
312}
313
168cf367
CW
314struct shrinker_lock_uninterruptible {
315 bool was_interruptible;
316 bool unlock;
317};
318
319static bool
320i915_gem_shrinker_lock_uninterruptible(struct drm_i915_private *dev_priv,
321 struct shrinker_lock_uninterruptible *slu,
322 int timeout_ms)
323{
5cba5be6
CW
324 unsigned long timeout = jiffies + msecs_to_jiffies_timeout(timeout_ms);
325
326 do {
ea746f36 327 if (i915_gem_wait_for_idle(dev_priv, 0) == 0 &&
5cba5be6
CW
328 i915_gem_shrinker_lock(&dev_priv->drm, &slu->unlock))
329 break;
168cf367 330
168cf367
CW
331 schedule_timeout_killable(1);
332 if (fatal_signal_pending(current))
333 return false;
5cba5be6
CW
334
335 if (time_after(jiffies, timeout)) {
168cf367
CW
336 pr_err("Unable to lock GPU to purge memory.\n");
337 return false;
338 }
5cba5be6 339 } while (1);
168cf367
CW
340
341 slu->was_interruptible = dev_priv->mm.interruptible;
342 dev_priv->mm.interruptible = false;
343 return true;
344}
345
346static void
347i915_gem_shrinker_unlock_uninterruptible(struct drm_i915_private *dev_priv,
348 struct shrinker_lock_uninterruptible *slu)
349{
350 dev_priv->mm.interruptible = slu->was_interruptible;
351 if (slu->unlock)
91c8a326 352 mutex_unlock(&dev_priv->drm.struct_mutex);
168cf367
CW
353}
354
be6a0376
DV
355static int
356i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
357{
358 struct drm_i915_private *dev_priv =
359 container_of(nb, struct drm_i915_private, mm.oom_notifier);
168cf367 360 struct shrinker_lock_uninterruptible slu;
be6a0376 361 struct drm_i915_gem_object *obj;
1768d455 362 unsigned long unevictable, bound, unbound, freed_pages;
be6a0376 363
168cf367 364 if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000))
be6a0376 365 return NOTIFY_DONE;
be6a0376 366
ea9d9768 367 intel_runtime_pm_get(dev_priv);
be6a0376 368 freed_pages = i915_gem_shrink_all(dev_priv);
ea9d9768 369 intel_runtime_pm_put(dev_priv);
be6a0376 370
be6a0376
DV
371 /* Because we may be allocating inside our own driver, we cannot
372 * assert that there are no objects with pinned pages that are not
373 * being pointed to by hardware.
374 */
1768d455 375 unbound = bound = unevictable = 0;
be6a0376 376 list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) {
1768d455
CW
377 if (!can_release_pages(obj))
378 unevictable += obj->base.size >> PAGE_SHIFT;
be6a0376 379 else
1768d455 380 unbound += obj->base.size >> PAGE_SHIFT;
be6a0376
DV
381 }
382 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
1768d455
CW
383 if (!can_release_pages(obj))
384 unevictable += obj->base.size >> PAGE_SHIFT;
be6a0376 385 else
1768d455 386 bound += obj->base.size >> PAGE_SHIFT;
be6a0376
DV
387 }
388
168cf367 389 i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
be6a0376
DV
390
391 if (freed_pages || unbound || bound)
1768d455
CW
392 pr_info("Purging GPU memory, %lu pages freed, "
393 "%lu pages still pinned.\n",
394 freed_pages, unevictable);
be6a0376 395 if (unbound || bound)
1768d455 396 pr_err("%lu and %lu pages still available in the "
be6a0376
DV
397 "bound and unbound GPU page lists.\n",
398 bound, unbound);
399
400 *(unsigned long *)ptr += freed_pages;
401 return NOTIFY_DONE;
402}
403
e87666b5
CW
404static int
405i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
406{
407 struct drm_i915_private *dev_priv =
408 container_of(nb, struct drm_i915_private, mm.vmap_notifier);
168cf367 409 struct shrinker_lock_uninterruptible slu;
8ef8561f
CW
410 struct i915_vma *vma, *next;
411 unsigned long freed_pages = 0;
412 int ret;
e87666b5 413
168cf367 414 if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000))
e87666b5 415 return NOTIFY_DONE;
e87666b5 416
8ef8561f 417 /* Force everything onto the inactive lists */
22dd3bb9 418 ret = i915_gem_wait_for_idle(dev_priv, I915_WAIT_LOCKED);
8ef8561f
CW
419 if (ret)
420 goto out;
421
ea9d9768 422 intel_runtime_pm_get(dev_priv);
8ef8561f
CW
423 freed_pages += i915_gem_shrink(dev_priv, -1UL,
424 I915_SHRINK_BOUND |
425 I915_SHRINK_UNBOUND |
426 I915_SHRINK_ACTIVE |
427 I915_SHRINK_VMAPS);
ea9d9768 428 intel_runtime_pm_put(dev_priv);
8ef8561f
CW
429
430 /* We also want to clear any cached iomaps as they wrap vmap */
431 list_for_each_entry_safe(vma, next,
432 &dev_priv->ggtt.base.inactive_list, vm_link) {
433 unsigned long count = vma->node.size >> PAGE_SHIFT;
434 if (vma->iomap && i915_vma_unbind(vma) == 0)
435 freed_pages += count;
436 }
e87666b5 437
8ef8561f 438out:
168cf367 439 i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
e87666b5
CW
440
441 *(unsigned long *)ptr += freed_pages;
442 return NOTIFY_DONE;
443}
444
eb0b44ad
DV
445/**
446 * i915_gem_shrinker_init - Initialize i915 shrinker
447 * @dev_priv: i915 device
448 *
449 * This function registers and sets up the i915 shrinker and OOM handler.
450 */
be6a0376
DV
451void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
452{
453 dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
454 dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
455 dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
a8a40589 456 WARN_ON(register_shrinker(&dev_priv->mm.shrinker));
be6a0376
DV
457
458 dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
a8a40589 459 WARN_ON(register_oom_notifier(&dev_priv->mm.oom_notifier));
e87666b5
CW
460
461 dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
462 WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
a8a40589
ID
463}
464
465/**
466 * i915_gem_shrinker_cleanup - Clean up i915 shrinker
467 * @dev_priv: i915 device
468 *
469 * This function unregisters the i915 shrinker and OOM handler.
470 */
471void i915_gem_shrinker_cleanup(struct drm_i915_private *dev_priv)
472{
e87666b5 473 WARN_ON(unregister_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
a8a40589
ID
474 WARN_ON(unregister_oom_notifier(&dev_priv->mm.oom_notifier));
475 unregister_shrinker(&dev_priv->mm.shrinker);
be6a0376 476}