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 | { | |
1bec9b0b CW |
73 | /* Only shmemfs objects are backed by swap */ |
74 | if (!obj->base.filp) | |
75 | return false; | |
76 | ||
c1a415e2 CW |
77 | /* Only report true if by unbinding the object and putting its pages |
78 | * we can actually make forward progress towards freeing physical | |
79 | * pages. | |
80 | * | |
81 | * If the pages are pinned for any other reason than being bound | |
82 | * to the GPU, simply unbinding from the GPU is not going to succeed | |
83 | * in releasing our pin count on the pages themselves. | |
84 | */ | |
85 | if (obj->pages_pin_count != num_vma_bound(obj)) | |
86 | return false; | |
87 | ||
88 | /* We can only return physical pages to the system if we can either | |
89 | * discard the contents (because the user has marked them as being | |
90 | * purgeable) or if we can move their contents out to swap. | |
91 | */ | |
92 | return swap_available() || obj->madv == I915_MADV_DONTNEED; | |
93 | } | |
94 | ||
eb0b44ad DV |
95 | /** |
96 | * i915_gem_shrink - Shrink buffer object caches | |
97 | * @dev_priv: i915 device | |
98 | * @target: amount of memory to make available, in pages | |
99 | * @flags: control flags for selecting cache types | |
100 | * | |
101 | * This function is the main interface to the shrinker. It will try to release | |
102 | * up to @target pages of main memory backing storage from buffer objects. | |
103 | * Selection of the specific caches can be done with @flags. This is e.g. useful | |
104 | * when purgeable objects should be removed from caches preferentially. | |
105 | * | |
106 | * Note that it's not guaranteed that released amount is actually available as | |
107 | * free system memory - the pages might still be in-used to due to other reasons | |
108 | * (like cpu mmaps) or the mm core has reused them before we could grab them. | |
109 | * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to | |
110 | * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all(). | |
111 | * | |
112 | * Also note that any kind of pinning (both per-vma address space pins and | |
113 | * backing storage pins at the buffer object level) result in the shrinker code | |
114 | * having to skip the object. | |
115 | * | |
116 | * Returns: | |
117 | * The number of pages of backing storage actually released. | |
118 | */ | |
be6a0376 DV |
119 | unsigned long |
120 | i915_gem_shrink(struct drm_i915_private *dev_priv, | |
14387540 | 121 | unsigned long target, unsigned flags) |
be6a0376 DV |
122 | { |
123 | const struct { | |
124 | struct list_head *list; | |
125 | unsigned int bit; | |
126 | } phases[] = { | |
127 | { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND }, | |
128 | { &dev_priv->mm.bound_list, I915_SHRINK_BOUND }, | |
129 | { NULL, 0 }, | |
130 | }, *phase; | |
131 | unsigned long count = 0; | |
132 | ||
3abafa53 | 133 | trace_i915_gem_shrink(dev_priv, target, flags); |
c9c0f5ea | 134 | i915_gem_retire_requests(dev_priv->dev); |
3abafa53 | 135 | |
be6a0376 DV |
136 | /* |
137 | * As we may completely rewrite the (un)bound list whilst unbinding | |
138 | * (due to retiring requests) we have to strictly process only | |
139 | * one element of the list at the time, and recheck the list | |
140 | * on every iteration. | |
141 | * | |
142 | * In particular, we must hold a reference whilst removing the | |
143 | * object as we may end up waiting for and/or retiring the objects. | |
144 | * This might release the final reference (held by the active list) | |
145 | * and result in the object being freed from under us. This is | |
146 | * similar to the precautions the eviction code must take whilst | |
147 | * removing objects. | |
148 | * | |
149 | * Also note that although these lists do not hold a reference to | |
150 | * the object we can safely grab one here: The final object | |
151 | * unreferencing and the bound_list are both protected by the | |
152 | * dev->struct_mutex and so we won't ever be able to observe an | |
153 | * object on the bound_list with a reference count equals 0. | |
154 | */ | |
155 | for (phase = phases; phase->list; phase++) { | |
156 | struct list_head still_in_list; | |
157 | ||
158 | if ((flags & phase->bit) == 0) | |
159 | continue; | |
160 | ||
161 | INIT_LIST_HEAD(&still_in_list); | |
162 | while (count < target && !list_empty(phase->list)) { | |
163 | struct drm_i915_gem_object *obj; | |
164 | struct i915_vma *vma, *v; | |
165 | ||
166 | obj = list_first_entry(phase->list, | |
167 | typeof(*obj), global_list); | |
168 | list_move_tail(&obj->global_list, &still_in_list); | |
169 | ||
170 | if (flags & I915_SHRINK_PURGEABLE && | |
171 | obj->madv != I915_MADV_DONTNEED) | |
172 | continue; | |
173 | ||
eae2c43b CW |
174 | if (flags & I915_SHRINK_VMAPS && |
175 | !is_vmalloc_addr(obj->mapping)) | |
176 | continue; | |
177 | ||
5763ff04 CW |
178 | if ((flags & I915_SHRINK_ACTIVE) == 0 && obj->active) |
179 | continue; | |
180 | ||
c1a415e2 CW |
181 | if (!can_release_pages(obj)) |
182 | continue; | |
183 | ||
be6a0376 DV |
184 | drm_gem_object_reference(&obj->base); |
185 | ||
186 | /* For the unbound phase, this should be a no-op! */ | |
187 | list_for_each_entry_safe(vma, v, | |
1c7f4bca | 188 | &obj->vma_list, obj_link) |
be6a0376 DV |
189 | if (i915_vma_unbind(vma)) |
190 | break; | |
191 | ||
192 | if (i915_gem_object_put_pages(obj) == 0) | |
193 | count += obj->base.size >> PAGE_SHIFT; | |
194 | ||
195 | drm_gem_object_unreference(&obj->base); | |
196 | } | |
197 | list_splice(&still_in_list, phase->list); | |
198 | } | |
199 | ||
c9c0f5ea CW |
200 | i915_gem_retire_requests(dev_priv->dev); |
201 | ||
be6a0376 DV |
202 | return count; |
203 | } | |
204 | ||
eb0b44ad | 205 | /** |
1f2449cd | 206 | * i915_gem_shrink_all - Shrink buffer object caches completely |
eb0b44ad DV |
207 | * @dev_priv: i915 device |
208 | * | |
209 | * This is a simple wraper around i915_gem_shrink() to aggressively shrink all | |
210 | * caches completely. It also first waits for and retires all outstanding | |
211 | * requests to also be able to release backing storage for active objects. | |
212 | * | |
213 | * This should only be used in code to intentionally quiescent the gpu or as a | |
214 | * last-ditch effort when memory seems to have run out. | |
215 | * | |
216 | * Returns: | |
217 | * The number of pages of backing storage actually released. | |
218 | */ | |
be6a0376 DV |
219 | unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv) |
220 | { | |
14387540 | 221 | return i915_gem_shrink(dev_priv, -1UL, |
5763ff04 CW |
222 | I915_SHRINK_BOUND | |
223 | I915_SHRINK_UNBOUND | | |
224 | I915_SHRINK_ACTIVE); | |
be6a0376 DV |
225 | } |
226 | ||
227 | static bool i915_gem_shrinker_lock(struct drm_device *dev, bool *unlock) | |
228 | { | |
229 | if (!mutex_trylock(&dev->struct_mutex)) { | |
230 | if (!mutex_is_locked_by(&dev->struct_mutex, current)) | |
231 | return false; | |
232 | ||
233 | if (to_i915(dev)->mm.shrinker_no_lock_stealing) | |
234 | return false; | |
235 | ||
236 | *unlock = false; | |
237 | } else | |
238 | *unlock = true; | |
239 | ||
240 | return true; | |
241 | } | |
242 | ||
be6a0376 DV |
243 | static unsigned long |
244 | i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc) | |
245 | { | |
246 | struct drm_i915_private *dev_priv = | |
247 | container_of(shrinker, struct drm_i915_private, mm.shrinker); | |
248 | struct drm_device *dev = dev_priv->dev; | |
249 | struct drm_i915_gem_object *obj; | |
250 | unsigned long count; | |
251 | bool unlock; | |
252 | ||
253 | if (!i915_gem_shrinker_lock(dev, &unlock)) | |
254 | return 0; | |
255 | ||
256 | count = 0; | |
257 | list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) | |
6f0ac204 | 258 | if (can_release_pages(obj)) |
be6a0376 DV |
259 | count += obj->base.size >> PAGE_SHIFT; |
260 | ||
261 | list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) { | |
c1a415e2 | 262 | if (!obj->active && can_release_pages(obj)) |
be6a0376 DV |
263 | count += obj->base.size >> PAGE_SHIFT; |
264 | } | |
265 | ||
266 | if (unlock) | |
267 | mutex_unlock(&dev->struct_mutex); | |
268 | ||
269 | return count; | |
270 | } | |
271 | ||
272 | static unsigned long | |
273 | i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc) | |
274 | { | |
275 | struct drm_i915_private *dev_priv = | |
276 | container_of(shrinker, struct drm_i915_private, mm.shrinker); | |
277 | struct drm_device *dev = dev_priv->dev; | |
278 | unsigned long freed; | |
279 | bool unlock; | |
280 | ||
281 | if (!i915_gem_shrinker_lock(dev, &unlock)) | |
282 | return SHRINK_STOP; | |
283 | ||
284 | freed = i915_gem_shrink(dev_priv, | |
285 | sc->nr_to_scan, | |
286 | I915_SHRINK_BOUND | | |
287 | I915_SHRINK_UNBOUND | | |
288 | I915_SHRINK_PURGEABLE); | |
289 | if (freed < sc->nr_to_scan) | |
290 | freed += i915_gem_shrink(dev_priv, | |
291 | sc->nr_to_scan - freed, | |
292 | I915_SHRINK_BOUND | | |
293 | I915_SHRINK_UNBOUND); | |
294 | if (unlock) | |
295 | mutex_unlock(&dev->struct_mutex); | |
296 | ||
297 | return freed; | |
298 | } | |
299 | ||
168cf367 CW |
300 | struct shrinker_lock_uninterruptible { |
301 | bool was_interruptible; | |
302 | bool unlock; | |
303 | }; | |
304 | ||
305 | static bool | |
306 | i915_gem_shrinker_lock_uninterruptible(struct drm_i915_private *dev_priv, | |
307 | struct shrinker_lock_uninterruptible *slu, | |
308 | int timeout_ms) | |
309 | { | |
310 | unsigned long timeout = msecs_to_jiffies(timeout_ms) + 1; | |
311 | ||
312 | while (!i915_gem_shrinker_lock(dev_priv->dev, &slu->unlock)) { | |
313 | schedule_timeout_killable(1); | |
314 | if (fatal_signal_pending(current)) | |
315 | return false; | |
316 | if (--timeout == 0) { | |
317 | pr_err("Unable to lock GPU to purge memory.\n"); | |
318 | return false; | |
319 | } | |
320 | } | |
321 | ||
322 | slu->was_interruptible = dev_priv->mm.interruptible; | |
323 | dev_priv->mm.interruptible = false; | |
324 | return true; | |
325 | } | |
326 | ||
327 | static void | |
328 | i915_gem_shrinker_unlock_uninterruptible(struct drm_i915_private *dev_priv, | |
329 | struct shrinker_lock_uninterruptible *slu) | |
330 | { | |
331 | dev_priv->mm.interruptible = slu->was_interruptible; | |
332 | if (slu->unlock) | |
333 | mutex_unlock(&dev_priv->dev->struct_mutex); | |
334 | } | |
335 | ||
be6a0376 DV |
336 | static int |
337 | i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr) | |
338 | { | |
339 | struct drm_i915_private *dev_priv = | |
340 | container_of(nb, struct drm_i915_private, mm.oom_notifier); | |
168cf367 | 341 | struct shrinker_lock_uninterruptible slu; |
be6a0376 | 342 | struct drm_i915_gem_object *obj; |
1768d455 | 343 | unsigned long unevictable, bound, unbound, freed_pages; |
be6a0376 | 344 | |
168cf367 | 345 | if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000)) |
be6a0376 | 346 | return NOTIFY_DONE; |
be6a0376 DV |
347 | |
348 | freed_pages = i915_gem_shrink_all(dev_priv); | |
349 | ||
be6a0376 DV |
350 | /* Because we may be allocating inside our own driver, we cannot |
351 | * assert that there are no objects with pinned pages that are not | |
352 | * being pointed to by hardware. | |
353 | */ | |
1768d455 | 354 | unbound = bound = unevictable = 0; |
be6a0376 | 355 | list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) { |
1768d455 CW |
356 | if (!can_release_pages(obj)) |
357 | unevictable += obj->base.size >> PAGE_SHIFT; | |
be6a0376 | 358 | else |
1768d455 | 359 | unbound += obj->base.size >> PAGE_SHIFT; |
be6a0376 DV |
360 | } |
361 | list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) { | |
1768d455 CW |
362 | if (!can_release_pages(obj)) |
363 | unevictable += obj->base.size >> PAGE_SHIFT; | |
be6a0376 | 364 | else |
1768d455 | 365 | bound += obj->base.size >> PAGE_SHIFT; |
be6a0376 DV |
366 | } |
367 | ||
168cf367 | 368 | i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu); |
be6a0376 DV |
369 | |
370 | if (freed_pages || unbound || bound) | |
1768d455 CW |
371 | pr_info("Purging GPU memory, %lu pages freed, " |
372 | "%lu pages still pinned.\n", | |
373 | freed_pages, unevictable); | |
be6a0376 | 374 | if (unbound || bound) |
1768d455 | 375 | pr_err("%lu and %lu pages still available in the " |
be6a0376 DV |
376 | "bound and unbound GPU page lists.\n", |
377 | bound, unbound); | |
378 | ||
379 | *(unsigned long *)ptr += freed_pages; | |
380 | return NOTIFY_DONE; | |
381 | } | |
382 | ||
e87666b5 CW |
383 | static int |
384 | i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr) | |
385 | { | |
386 | struct drm_i915_private *dev_priv = | |
387 | container_of(nb, struct drm_i915_private, mm.vmap_notifier); | |
168cf367 | 388 | struct shrinker_lock_uninterruptible slu; |
e87666b5 | 389 | unsigned long freed_pages; |
e87666b5 | 390 | |
168cf367 | 391 | if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 5000)) |
e87666b5 | 392 | return NOTIFY_DONE; |
e87666b5 | 393 | |
eae2c43b CW |
394 | freed_pages = i915_gem_shrink(dev_priv, -1UL, |
395 | I915_SHRINK_BOUND | | |
396 | I915_SHRINK_UNBOUND | | |
397 | I915_SHRINK_ACTIVE | | |
398 | I915_SHRINK_VMAPS); | |
e87666b5 | 399 | |
168cf367 | 400 | i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu); |
e87666b5 CW |
401 | |
402 | *(unsigned long *)ptr += freed_pages; | |
403 | return NOTIFY_DONE; | |
404 | } | |
405 | ||
eb0b44ad DV |
406 | /** |
407 | * i915_gem_shrinker_init - Initialize i915 shrinker | |
408 | * @dev_priv: i915 device | |
409 | * | |
410 | * This function registers and sets up the i915 shrinker and OOM handler. | |
411 | */ | |
be6a0376 DV |
412 | void i915_gem_shrinker_init(struct drm_i915_private *dev_priv) |
413 | { | |
414 | dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan; | |
415 | dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count; | |
416 | dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS; | |
a8a40589 | 417 | WARN_ON(register_shrinker(&dev_priv->mm.shrinker)); |
be6a0376 DV |
418 | |
419 | dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom; | |
a8a40589 | 420 | WARN_ON(register_oom_notifier(&dev_priv->mm.oom_notifier)); |
e87666b5 CW |
421 | |
422 | dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap; | |
423 | WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier)); | |
a8a40589 ID |
424 | } |
425 | ||
426 | /** | |
427 | * i915_gem_shrinker_cleanup - Clean up i915 shrinker | |
428 | * @dev_priv: i915 device | |
429 | * | |
430 | * This function unregisters the i915 shrinker and OOM handler. | |
431 | */ | |
432 | void i915_gem_shrinker_cleanup(struct drm_i915_private *dev_priv) | |
433 | { | |
e87666b5 | 434 | WARN_ON(unregister_vmap_purge_notifier(&dev_priv->mm.vmap_notifier)); |
a8a40589 ID |
435 | WARN_ON(unregister_oom_notifier(&dev_priv->mm.oom_notifier)); |
436 | unregister_shrinker(&dev_priv->mm.shrinker); | |
be6a0376 | 437 | } |