Commit | Line | Data |
---|---|---|
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 | ||
38 | static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task) | |
39 | { | |
40 | if (!mutex_is_locked(mutex)) | |
41 | return false; | |
42 | ||
43 | #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES) | |
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 | ||
c1a415e2 CW |
51 | static int num_vma_bound(struct drm_i915_gem_object *obj) |
52 | { | |
53 | struct i915_vma *vma; | |
54 | int count = 0; | |
55 | ||
1c7f4bca | 56 | list_for_each_entry(vma, &obj->vma_list, obj_link) { |
c1a415e2 CW |
57 | if (drm_mm_node_allocated(&vma->node)) |
58 | count++; | |
59 | if (vma->pin_count) | |
60 | count++; | |
61 | } | |
62 | ||
63 | return count; | |
64 | } | |
65 | ||
66 | static bool swap_available(void) | |
67 | { | |
68 | return get_nr_swap_pages() > 0; | |
69 | } | |
70 | ||
71 | static bool can_release_pages(struct drm_i915_gem_object *obj) | |
72 | { | |
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 | */ | |
81 | if (obj->pages_pin_count != num_vma_bound(obj)) | |
82 | return false; | |
83 | ||
84 | /* We can only return physical pages to the system if we can either | |
85 | * discard the contents (because the user has marked them as being | |
86 | * purgeable) or if we can move their contents out to swap. | |
87 | */ | |
88 | return swap_available() || obj->madv == I915_MADV_DONTNEED; | |
89 | } | |
90 | ||
eb0b44ad DV |
91 | /** |
92 | * i915_gem_shrink - Shrink buffer object caches | |
93 | * @dev_priv: i915 device | |
94 | * @target: amount of memory to make available, in pages | |
95 | * @flags: control flags for selecting cache types | |
96 | * | |
97 | * This function is the main interface to the shrinker. It will try to release | |
98 | * up to @target pages of main memory backing storage from buffer objects. | |
99 | * Selection of the specific caches can be done with @flags. This is e.g. useful | |
100 | * when purgeable objects should be removed from caches preferentially. | |
101 | * | |
102 | * Note that it's not guaranteed that released amount is actually available as | |
103 | * free system memory - the pages might still be in-used to due to other reasons | |
104 | * (like cpu mmaps) or the mm core has reused them before we could grab them. | |
105 | * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to | |
106 | * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all(). | |
107 | * | |
108 | * Also note that any kind of pinning (both per-vma address space pins and | |
109 | * backing storage pins at the buffer object level) result in the shrinker code | |
110 | * having to skip the object. | |
111 | * | |
112 | * Returns: | |
113 | * The number of pages of backing storage actually released. | |
114 | */ | |
be6a0376 DV |
115 | unsigned long |
116 | i915_gem_shrink(struct drm_i915_private *dev_priv, | |
14387540 | 117 | unsigned long target, unsigned flags) |
be6a0376 DV |
118 | { |
119 | const struct { | |
120 | struct list_head *list; | |
121 | unsigned int bit; | |
122 | } phases[] = { | |
123 | { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND }, | |
124 | { &dev_priv->mm.bound_list, I915_SHRINK_BOUND }, | |
125 | { NULL, 0 }, | |
126 | }, *phase; | |
127 | unsigned long count = 0; | |
128 | ||
3abafa53 | 129 | trace_i915_gem_shrink(dev_priv, target, flags); |
c9c0f5ea | 130 | i915_gem_retire_requests(dev_priv->dev); |
3abafa53 | 131 | |
be6a0376 DV |
132 | /* |
133 | * As we may completely rewrite the (un)bound list whilst unbinding | |
134 | * (due to retiring requests) we have to strictly process only | |
135 | * one element of the list at the time, and recheck the list | |
136 | * on every iteration. | |
137 | * | |
138 | * In particular, we must hold a reference whilst removing the | |
139 | * object as we may end up waiting for and/or retiring the objects. | |
140 | * This might release the final reference (held by the active list) | |
141 | * and result in the object being freed from under us. This is | |
142 | * similar to the precautions the eviction code must take whilst | |
143 | * removing objects. | |
144 | * | |
145 | * Also note that although these lists do not hold a reference to | |
146 | * the object we can safely grab one here: The final object | |
147 | * unreferencing and the bound_list are both protected by the | |
148 | * dev->struct_mutex and so we won't ever be able to observe an | |
149 | * object on the bound_list with a reference count equals 0. | |
150 | */ | |
151 | for (phase = phases; phase->list; phase++) { | |
152 | struct list_head still_in_list; | |
153 | ||
154 | if ((flags & phase->bit) == 0) | |
155 | continue; | |
156 | ||
157 | INIT_LIST_HEAD(&still_in_list); | |
158 | while (count < target && !list_empty(phase->list)) { | |
159 | struct drm_i915_gem_object *obj; | |
160 | struct i915_vma *vma, *v; | |
161 | ||
162 | obj = list_first_entry(phase->list, | |
163 | typeof(*obj), global_list); | |
164 | list_move_tail(&obj->global_list, &still_in_list); | |
165 | ||
166 | if (flags & I915_SHRINK_PURGEABLE && | |
167 | obj->madv != I915_MADV_DONTNEED) | |
168 | continue; | |
169 | ||
eae2c43b CW |
170 | if (flags & I915_SHRINK_VMAPS && |
171 | !is_vmalloc_addr(obj->mapping)) | |
172 | continue; | |
173 | ||
5763ff04 CW |
174 | if ((flags & I915_SHRINK_ACTIVE) == 0 && obj->active) |
175 | continue; | |
176 | ||
c1a415e2 CW |
177 | if (!can_release_pages(obj)) |
178 | continue; | |
179 | ||
be6a0376 DV |
180 | drm_gem_object_reference(&obj->base); |
181 | ||
182 | /* For the unbound phase, this should be a no-op! */ | |
183 | list_for_each_entry_safe(vma, v, | |
1c7f4bca | 184 | &obj->vma_list, obj_link) |
be6a0376 DV |
185 | if (i915_vma_unbind(vma)) |
186 | break; | |
187 | ||
188 | if (i915_gem_object_put_pages(obj) == 0) | |
189 | count += obj->base.size >> PAGE_SHIFT; | |
190 | ||
191 | drm_gem_object_unreference(&obj->base); | |
192 | } | |
193 | list_splice(&still_in_list, phase->list); | |
194 | } | |
195 | ||
c9c0f5ea CW |
196 | i915_gem_retire_requests(dev_priv->dev); |
197 | ||
be6a0376 DV |
198 | return count; |
199 | } | |
200 | ||
eb0b44ad | 201 | /** |
1f2449cd | 202 | * i915_gem_shrink_all - Shrink buffer object caches completely |
eb0b44ad DV |
203 | * @dev_priv: i915 device |
204 | * | |
205 | * This is a simple wraper around i915_gem_shrink() to aggressively shrink all | |
206 | * caches completely. It also first waits for and retires all outstanding | |
207 | * requests to also be able to release backing storage for active objects. | |
208 | * | |
209 | * This should only be used in code to intentionally quiescent the gpu or as a | |
210 | * last-ditch effort when memory seems to have run out. | |
211 | * | |
212 | * Returns: | |
213 | * The number of pages of backing storage actually released. | |
214 | */ | |
be6a0376 DV |
215 | unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv) |
216 | { | |
14387540 | 217 | return i915_gem_shrink(dev_priv, -1UL, |
5763ff04 CW |
218 | I915_SHRINK_BOUND | |
219 | I915_SHRINK_UNBOUND | | |
220 | I915_SHRINK_ACTIVE); | |
be6a0376 DV |
221 | } |
222 | ||
223 | static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock) | |
224 | { | |
225 | if (!mutex_trylock(&dev->struct_mutex)) { | |
226 | if (!mutex_is_locked_by(&dev->struct_mutex, current)) | |
227 | return false; | |
228 | ||
229 | if (to_i915(dev)->mm.shrinker_no_lock_stealing) | |
230 | return false; | |
231 | ||
232 | *unlock = false; | |
233 | } else | |
234 | *unlock = true; | |
235 | ||
236 | return true; | |
237 | } | |
238 | ||
be6a0376 DV |
239 | static unsigned long |
240 | i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc) | |
241 | { | |
242 | struct drm_i915_private *dev_priv = | |
243 | container_of(shrinker, struct drm_i915_private, mm.shrinker); | |
244 | struct drm_device *dev = dev_priv->dev; | |
245 | struct drm_i915_gem_object *obj; | |
246 | unsigned long count; | |
247 | bool unlock; | |
248 | ||
249 | if (!i915_gem_shrinker_lock(dev, &unlock)) | |
250 | return 0; | |
251 | ||
252 | count = 0; | |
253 | list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) | |
6f0ac204 | 254 | if (can_release_pages(obj)) |
be6a0376 DV |
255 | count += obj->base.size >> PAGE_SHIFT; |
256 | ||
257 | list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) { | |
c1a415e2 | 258 | if (!obj->active && can_release_pages(obj)) |
be6a0376 DV |
259 | count += obj->base.size >> PAGE_SHIFT; |
260 | } | |
261 | ||
262 | if (unlock) | |
263 | mutex_unlock(&dev->struct_mutex); | |
264 | ||
265 | return count; | |
266 | } | |
267 | ||
268 | static unsigned long | |
269 | i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc) | |
270 | { | |
271 | struct drm_i915_private *dev_priv = | |
272 | container_of(shrinker, struct drm_i915_private, mm.shrinker); | |
273 | struct drm_device *dev = dev_priv->dev; | |
274 | unsigned long freed; | |
275 | bool unlock; | |
276 | ||
277 | if (!i915_gem_shrinker_lock(dev, &unlock)) | |
278 | return SHRINK_STOP; | |
279 | ||
280 | freed = i915_gem_shrink(dev_priv, | |
281 | sc->nr_to_scan, | |
282 | I915_SHRINK_BOUND | | |
283 | I915_SHRINK_UNBOUND | | |
284 | I915_SHRINK_PURGEABLE); | |
285 | if (freed < sc->nr_to_scan) | |
286 | freed += i915_gem_shrink(dev_priv, | |
287 | sc->nr_to_scan - freed, | |
288 | I915_SHRINK_BOUND | | |
289 | I915_SHRINK_UNBOUND); | |
290 | if (unlock) | |
291 | mutex_unlock(&dev->struct_mutex); | |
292 | ||
293 | return freed; | |
294 | } | |
295 | ||
168cf367 CW |
296 | struct shrinker_lock_uninterruptible { |
297 | bool was_interruptible; | |
298 | bool unlock; | |
299 | }; | |
300 | ||
301 | static bool | |
302 | i915_gem_shrinker_lock_uninterruptible(struct drm_i915_private *dev_priv, | |
303 | struct shrinker_lock_uninterruptible *slu, | |
304 | int timeout_ms) | |
305 | { | |
306 | unsigned long timeout = msecs_to_jiffies(timeout_ms) + 1; | |
307 | ||
308 | while (!i915_gem_shrinker_lock(dev_priv->dev, &slu->unlock)) { | |
309 | schedule_timeout_killable(1); | |
310 | if (fatal_signal_pending(current)) | |
311 | return false; | |
312 | if (--timeout == 0) { | |
313 | pr_err("Unable to lock GPU to purge memory.\n"); | |
314 | return false; | |
315 | } | |
316 | } | |
317 | ||
318 | slu->was_interruptible = dev_priv->mm.interruptible; | |
319 | dev_priv->mm.interruptible = false; | |
320 | return true; | |
321 | } | |
322 | ||
323 | static void | |
324 | i915_gem_shrinker_unlock_uninterruptible(struct drm_i915_private *dev_priv, | |
325 | struct shrinker_lock_uninterruptible *slu) | |
326 | { | |
327 | dev_priv->mm.interruptible = slu->was_interruptible; | |
328 | if (slu->unlock) | |
329 | mutex_unlock(&dev_priv->dev->struct_mutex); | |
330 | } | |
331 | ||
be6a0376 DV |
332 | static int |
333 | i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr) | |
334 | { | |
335 | struct drm_i915_private *dev_priv = | |
336 | container_of(nb, struct drm_i915_private, mm.oom_notifier); | |
168cf367 | 337 | struct shrinker_lock_uninterruptible slu; |
be6a0376 | 338 | struct drm_i915_gem_object *obj; |
be6a0376 | 339 | unsigned long pinned, bound, unbound, freed_pages; |
be6a0376 | 340 | |
168cf367 | 341 | if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000)) |
be6a0376 | 342 | return NOTIFY_DONE; |
be6a0376 DV |
343 | |
344 | freed_pages = i915_gem_shrink_all(dev_priv); | |
345 | ||
be6a0376 DV |
346 | /* Because we may be allocating inside our own driver, we cannot |
347 | * assert that there are no objects with pinned pages that are not | |
348 | * being pointed to by hardware. | |
349 | */ | |
350 | unbound = bound = pinned = 0; | |
351 | list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) { | |
352 | if (!obj->base.filp) /* not backed by a freeable object */ | |
353 | continue; | |
354 | ||
355 | if (obj->pages_pin_count) | |
356 | pinned += obj->base.size; | |
357 | else | |
358 | unbound += obj->base.size; | |
359 | } | |
360 | list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) { | |
361 | if (!obj->base.filp) | |
362 | continue; | |
363 | ||
364 | if (obj->pages_pin_count) | |
365 | pinned += obj->base.size; | |
366 | else | |
367 | bound += obj->base.size; | |
368 | } | |
369 | ||
168cf367 | 370 | i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu); |
be6a0376 DV |
371 | |
372 | if (freed_pages || unbound || bound) | |
373 | pr_info("Purging GPU memory, %lu bytes freed, %lu bytes still pinned.\n", | |
374 | freed_pages << PAGE_SHIFT, pinned); | |
375 | if (unbound || bound) | |
376 | pr_err("%lu and %lu bytes still available in the " | |
377 | "bound and unbound GPU page lists.\n", | |
378 | bound, unbound); | |
379 | ||
380 | *(unsigned long *)ptr += freed_pages; | |
381 | return NOTIFY_DONE; | |
382 | } | |
383 | ||
e87666b5 CW |
384 | static int |
385 | i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr) | |
386 | { | |
387 | struct drm_i915_private *dev_priv = | |
388 | container_of(nb, struct drm_i915_private, mm.vmap_notifier); | |
168cf367 | 389 | struct shrinker_lock_uninterruptible slu; |
e87666b5 | 390 | unsigned long freed_pages; |
e87666b5 | 391 | |
168cf367 | 392 | if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000)) |
e87666b5 | 393 | return NOTIFY_DONE; |
e87666b5 | 394 | |
eae2c43b CW |
395 | freed_pages = i915_gem_shrink(dev_priv, -1UL, |
396 | I915_SHRINK_BOUND | | |
397 | I915_SHRINK_UNBOUND | | |
398 | I915_SHRINK_ACTIVE | | |
399 | I915_SHRINK_VMAPS); | |
e87666b5 | 400 | |
168cf367 | 401 | i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu); |
e87666b5 CW |
402 | |
403 | *(unsigned long *)ptr += freed_pages; | |
404 | return NOTIFY_DONE; | |
405 | } | |
406 | ||
eb0b44ad DV |
407 | /** |
408 | * i915_gem_shrinker_init - Initialize i915 shrinker | |
409 | * @dev_priv: i915 device | |
410 | * | |
411 | * This function registers and sets up the i915 shrinker and OOM handler. | |
412 | */ | |
be6a0376 DV |
413 | void i915_gem_shrinker_init(struct drm_i915_private *dev_priv) |
414 | { | |
415 | dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan; | |
416 | dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count; | |
417 | dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS; | |
a8a40589 | 418 | WARN_ON(register_shrinker(&dev_priv->mm.shrinker)); |
be6a0376 DV |
419 | |
420 | dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom; | |
a8a40589 | 421 | WARN_ON(register_oom_notifier(&dev_priv->mm.oom_notifier)); |
e87666b5 CW |
422 | |
423 | dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap; | |
424 | WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier)); | |
a8a40589 ID |
425 | } |
426 | ||
427 | /** | |
428 | * i915_gem_shrinker_cleanup - Clean up i915 shrinker | |
429 | * @dev_priv: i915 device | |
430 | * | |
431 | * This function unregisters the i915 shrinker and OOM handler. | |
432 | */ | |
433 | void i915_gem_shrinker_cleanup(struct drm_i915_private *dev_priv) | |
434 | { | |
e87666b5 | 435 | WARN_ON(unregister_vmap_purge_notifier(&dev_priv->mm.vmap_notifier)); |
a8a40589 ID |
436 | WARN_ON(unregister_oom_notifier(&dev_priv->mm.oom_notifier)); |
437 | unregister_shrinker(&dev_priv->mm.shrinker); | |
be6a0376 | 438 | } |