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>
 31#include <linux/vmalloc.h>
 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 shrinker_lock(struct drm_i915_private *i915, bool *unlock)
 39{
 40	switch (mutex_trylock_recursive(&i915->drm.struct_mutex)) {
 41	case MUTEX_TRYLOCK_RECURSIVE:
 42		*unlock = false;
 43		return true;
 44
 45	case MUTEX_TRYLOCK_FAILED:
 46		*unlock = false;
 47		preempt_disable();
 48		do {
 49			cpu_relax();
 50			if (mutex_trylock(&i915->drm.struct_mutex)) {
 51				*unlock = true;
 52				break;
 53			}
 54		} while (!need_resched());
 55		preempt_enable();
 56		return *unlock;
 57
 58	case MUTEX_TRYLOCK_SUCCESS:
 59		*unlock = true;
 60		return true;
 61	}
 62
 63	BUG();
 64}
 65
 66static void shrinker_unlock(struct drm_i915_private *i915, bool unlock)
 67{
 68	if (!unlock)
 69		return;
 70
 71	mutex_unlock(&i915->drm.struct_mutex);
 72}
 73
 74static bool swap_available(void)
 75{
 76	return get_nr_swap_pages() > 0;
 77}
 78
 79static bool can_release_pages(struct drm_i915_gem_object *obj)
 80{
 81	/* Consider only shrinkable ojects. */
 82	if (!i915_gem_object_is_shrinkable(obj))
 83		return false;
 84
 85	/* Only report true if by unbinding the object and putting its pages
 86	 * we can actually make forward progress towards freeing physical
 87	 * pages.
 88	 *
 89	 * If the pages are pinned for any other reason than being bound
 90	 * to the GPU, simply unbinding from the GPU is not going to succeed
 91	 * in releasing our pin count on the pages themselves.
 92	 */
 93	if (atomic_read(&obj->mm.pages_pin_count) > obj->bind_count)
 94		return false;
 95
 96	/* If any vma are "permanently" pinned, it will prevent us from
 97	 * reclaiming the obj->mm.pages. We only allow scanout objects to claim
 98	 * a permanent pin, along with a few others like the context objects.
 99	 * To simplify the scan, and to avoid walking the list of vma under the
100	 * object, we just check the count of its permanently pinned.
101	 */
102	if (READ_ONCE(obj->pin_global))
103		return false;
104
105	/* We can only return physical pages to the system if we can either
106	 * discard the contents (because the user has marked them as being
107	 * purgeable) or if we can move their contents out to swap.
108	 */
109	return swap_available() || obj->mm.madv == I915_MADV_DONTNEED;
110}
111
112static bool unsafe_drop_pages(struct drm_i915_gem_object *obj)
113{
114	if (i915_gem_object_unbind(obj) == 0)
115		__i915_gem_object_put_pages(obj, I915_MM_SHRINKER);
116	return !i915_gem_object_has_pages(obj);
117}
118
119/**
120 * i915_gem_shrink - Shrink buffer object caches
121 * @i915: i915 device
122 * @target: amount of memory to make available, in pages
123 * @nr_scanned: optional output for number of pages scanned (incremental)
124 * @flags: control flags for selecting cache types
125 *
126 * This function is the main interface to the shrinker. It will try to release
127 * up to @target pages of main memory backing storage from buffer objects.
128 * Selection of the specific caches can be done with @flags. This is e.g. useful
129 * when purgeable objects should be removed from caches preferentially.
130 *
131 * Note that it's not guaranteed that released amount is actually available as
132 * free system memory - the pages might still be in-used to due to other reasons
133 * (like cpu mmaps) or the mm core has reused them before we could grab them.
134 * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
135 * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
136 *
137 * Also note that any kind of pinning (both per-vma address space pins and
138 * backing storage pins at the buffer object level) result in the shrinker code
139 * having to skip the object.
140 *
141 * Returns:
142 * The number of pages of backing storage actually released.
143 */
144unsigned long
145i915_gem_shrink(struct drm_i915_private *i915,
146		unsigned long target,
147		unsigned long *nr_scanned,
148		unsigned flags)
149{
150	const struct {
151		struct list_head *list;
152		unsigned int bit;
153	} phases[] = {
154		{ &i915->mm.unbound_list, I915_SHRINK_UNBOUND },
155		{ &i915->mm.bound_list, I915_SHRINK_BOUND },
156		{ NULL, 0 },
157	}, *phase;
158	unsigned long count = 0;
159	unsigned long scanned = 0;
160	bool unlock;
161
162	if (!shrinker_lock(i915, &unlock))
163		return 0;
164
165	/*
166	 * When shrinking the active list, also consider active contexts.
167	 * Active contexts are pinned until they are retired, and so can
168	 * not be simply unbound to retire and unpin their pages. To shrink
169	 * the contexts, we must wait until the gpu is idle.
170	 *
171	 * We don't care about errors here; if we cannot wait upon the GPU,
172	 * we will free as much as we can and hope to get a second chance.
173	 */
174	if (flags & I915_SHRINK_ACTIVE)
175		i915_gem_wait_for_idle(i915, I915_WAIT_LOCKED);
176
177	trace_i915_gem_shrink(i915, target, flags);
178	i915_retire_requests(i915);
179
180	/*
181	 * Unbinding of objects will require HW access; Let us not wake the
182	 * device just to recover a little memory. If absolutely necessary,
183	 * we will force the wake during oom-notifier.
184	 */
185	if ((flags & I915_SHRINK_BOUND) &&
186	    !intel_runtime_pm_get_if_in_use(i915))
187		flags &= ~I915_SHRINK_BOUND;
188
189	/*
190	 * As we may completely rewrite the (un)bound list whilst unbinding
191	 * (due to retiring requests) we have to strictly process only
192	 * one element of the list at the time, and recheck the list
193	 * on every iteration.
194	 *
195	 * In particular, we must hold a reference whilst removing the
196	 * object as we may end up waiting for and/or retiring the objects.
197	 * This might release the final reference (held by the active list)
198	 * and result in the object being freed from under us. This is
199	 * similar to the precautions the eviction code must take whilst
200	 * removing objects.
201	 *
202	 * Also note that although these lists do not hold a reference to
203	 * the object we can safely grab one here: The final object
204	 * unreferencing and the bound_list are both protected by the
205	 * dev->struct_mutex and so we won't ever be able to observe an
206	 * object on the bound_list with a reference count equals 0.
207	 */
208	for (phase = phases; phase->list; phase++) {
209		struct list_head still_in_list;
210		struct drm_i915_gem_object *obj;
211
212		if ((flags & phase->bit) == 0)
213			continue;
214
215		INIT_LIST_HEAD(&still_in_list);
216
217		/*
218		 * We serialize our access to unreferenced objects through
219		 * the use of the struct_mutex. While the objects are not
220		 * yet freed (due to RCU then a workqueue) we still want
221		 * to be able to shrink their pages, so they remain on
222		 * the unbound/bound list until actually freed.
223		 */
224		spin_lock(&i915->mm.obj_lock);
225		while (count < target &&
226		       (obj = list_first_entry_or_null(phase->list,
227						       typeof(*obj),
228						       mm.link))) {
229			list_move_tail(&obj->mm.link, &still_in_list);
230
231			if (flags & I915_SHRINK_PURGEABLE &&
232			    obj->mm.madv != I915_MADV_DONTNEED)
233				continue;
234
235			if (flags & I915_SHRINK_VMAPS &&
236			    !is_vmalloc_addr(obj->mm.mapping))
237				continue;
238
239			if (!(flags & I915_SHRINK_ACTIVE) &&
240			    (i915_gem_object_is_active(obj) ||
241			     i915_gem_object_is_framebuffer(obj)))
242				continue;
243
244			if (!can_release_pages(obj))
245				continue;
246
247			spin_unlock(&i915->mm.obj_lock);
248
249			if (unsafe_drop_pages(obj)) {
250				/* May arrive from get_pages on another bo */
251				mutex_lock_nested(&obj->mm.lock,
252						  I915_MM_SHRINKER);
253				if (!i915_gem_object_has_pages(obj)) {
254					__i915_gem_object_invalidate(obj);
255					count += obj->base.size >> PAGE_SHIFT;
256				}
257				mutex_unlock(&obj->mm.lock);
258			}
259			scanned += obj->base.size >> PAGE_SHIFT;
260
261			spin_lock(&i915->mm.obj_lock);
262		}
263		list_splice_tail(&still_in_list, phase->list);
264		spin_unlock(&i915->mm.obj_lock);
265	}
266
267	if (flags & I915_SHRINK_BOUND)
268		intel_runtime_pm_put(i915);
269
270	i915_retire_requests(i915);
271
272	shrinker_unlock(i915, unlock);
273
274	if (nr_scanned)
275		*nr_scanned += scanned;
276	return count;
277}
278
279/**
280 * i915_gem_shrink_all - Shrink buffer object caches completely
281 * @i915: i915 device
282 *
283 * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
284 * caches completely. It also first waits for and retires all outstanding
285 * requests to also be able to release backing storage for active objects.
286 *
287 * This should only be used in code to intentionally quiescent the gpu or as a
288 * last-ditch effort when memory seems to have run out.
289 *
290 * Returns:
291 * The number of pages of backing storage actually released.
292 */
293unsigned long i915_gem_shrink_all(struct drm_i915_private *i915)
294{
295	unsigned long freed;
296
297	intel_runtime_pm_get(i915);
298	freed = i915_gem_shrink(i915, -1UL, NULL,
299				I915_SHRINK_BOUND |
300				I915_SHRINK_UNBOUND |
301				I915_SHRINK_ACTIVE);
302	intel_runtime_pm_put(i915);
303
304	return freed;
305}
306
307static unsigned long
308i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
309{
310	struct drm_i915_private *i915 =
311		container_of(shrinker, struct drm_i915_private, mm.shrinker);
312	struct drm_i915_gem_object *obj;
313	unsigned long num_objects = 0;
314	unsigned long count = 0;
315
316	spin_lock(&i915->mm.obj_lock);
317	list_for_each_entry(obj, &i915->mm.unbound_list, mm.link)
318		if (can_release_pages(obj)) {
319			count += obj->base.size >> PAGE_SHIFT;
320			num_objects++;
321		}
322
323	list_for_each_entry(obj, &i915->mm.bound_list, mm.link)
324		if (!i915_gem_object_is_active(obj) && can_release_pages(obj)) {
325			count += obj->base.size >> PAGE_SHIFT;
326			num_objects++;
327		}
328	spin_unlock(&i915->mm.obj_lock);
329
330	/* Update our preferred vmscan batch size for the next pass.
331	 * Our rough guess for an effective batch size is roughly 2
332	 * available GEM objects worth of pages. That is we don't want
333	 * the shrinker to fire, until it is worth the cost of freeing an
334	 * entire GEM object.
335	 */
336	if (num_objects) {
337		unsigned long avg = 2 * count / num_objects;
338
339		i915->mm.shrinker.batch =
340			max((i915->mm.shrinker.batch + avg) >> 1,
341			    128ul /* default SHRINK_BATCH */);
342	}
343
344	return count;
345}
346
347static unsigned long
348i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
349{
350	struct drm_i915_private *i915 =
351		container_of(shrinker, struct drm_i915_private, mm.shrinker);
352	unsigned long freed;
353	bool unlock;
354
355	sc->nr_scanned = 0;
356
357	if (!shrinker_lock(i915, &unlock))
358		return SHRINK_STOP;
359
360	freed = i915_gem_shrink(i915,
361				sc->nr_to_scan,
362				&sc->nr_scanned,
363				I915_SHRINK_BOUND |
364				I915_SHRINK_UNBOUND |
365				I915_SHRINK_PURGEABLE);
366	if (sc->nr_scanned < sc->nr_to_scan)
367		freed += i915_gem_shrink(i915,
368					 sc->nr_to_scan - sc->nr_scanned,
369					 &sc->nr_scanned,
370					 I915_SHRINK_BOUND |
371					 I915_SHRINK_UNBOUND);
372	if (sc->nr_scanned < sc->nr_to_scan && current_is_kswapd()) {
373		intel_runtime_pm_get(i915);
374		freed += i915_gem_shrink(i915,
375					 sc->nr_to_scan - sc->nr_scanned,
376					 &sc->nr_scanned,
377					 I915_SHRINK_ACTIVE |
378					 I915_SHRINK_BOUND |
379					 I915_SHRINK_UNBOUND);
380		intel_runtime_pm_put(i915);
381	}
382
383	shrinker_unlock(i915, unlock);
384
385	return sc->nr_scanned ? freed : SHRINK_STOP;
386}
387
388static bool
389shrinker_lock_uninterruptible(struct drm_i915_private *i915, bool *unlock,
390			      int timeout_ms)
391{
392	unsigned long timeout = jiffies + msecs_to_jiffies_timeout(timeout_ms);
393
394	do {
395		if (i915_gem_wait_for_idle(i915, 0) == 0 &&
396		    shrinker_lock(i915, unlock))
397			break;
398
399		schedule_timeout_killable(1);
400		if (fatal_signal_pending(current))
401			return false;
402
403		if (time_after(jiffies, timeout)) {
404			pr_err("Unable to lock GPU to purge memory.\n");
405			return false;
406		}
407	} while (1);
408
409	return true;
410}
411
412static int
413i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
414{
415	struct drm_i915_private *i915 =
416		container_of(nb, struct drm_i915_private, mm.oom_notifier);
417	struct drm_i915_gem_object *obj;
418	unsigned long unevictable, bound, unbound, freed_pages;
419
420	freed_pages = i915_gem_shrink_all(i915);
421
422	/* Because we may be allocating inside our own driver, we cannot
423	 * assert that there are no objects with pinned pages that are not
424	 * being pointed to by hardware.
425	 */
426	unbound = bound = unevictable = 0;
427	spin_lock(&i915->mm.obj_lock);
428	list_for_each_entry(obj, &i915->mm.unbound_list, mm.link) {
429		if (!can_release_pages(obj))
430			unevictable += obj->base.size >> PAGE_SHIFT;
431		else
432			unbound += obj->base.size >> PAGE_SHIFT;
433	}
434	list_for_each_entry(obj, &i915->mm.bound_list, mm.link) {
435		if (!can_release_pages(obj))
436			unevictable += obj->base.size >> PAGE_SHIFT;
437		else
438			bound += obj->base.size >> PAGE_SHIFT;
439	}
440	spin_unlock(&i915->mm.obj_lock);
441
442	if (freed_pages || unbound || bound)
443		pr_info("Purging GPU memory, %lu pages freed, "
444			"%lu pages still pinned.\n",
445			freed_pages, unevictable);
446	if (unbound || bound)
447		pr_err("%lu and %lu pages still available in the "
448		       "bound and unbound GPU page lists.\n",
449		       bound, unbound);
450
451	*(unsigned long *)ptr += freed_pages;
452	return NOTIFY_DONE;
453}
454
455static int
456i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
457{
458	struct drm_i915_private *i915 =
459		container_of(nb, struct drm_i915_private, mm.vmap_notifier);
460	struct i915_vma *vma, *next;
461	unsigned long freed_pages = 0;
462	bool unlock;
463	int ret;
464
465	if (!shrinker_lock_uninterruptible(i915, &unlock, 5000))
466		return NOTIFY_DONE;
467
468	/* Force everything onto the inactive lists */
469	ret = i915_gem_wait_for_idle(i915, I915_WAIT_LOCKED);
470	if (ret)
471		goto out;
472
473	intel_runtime_pm_get(i915);
474	freed_pages += i915_gem_shrink(i915, -1UL, NULL,
475				       I915_SHRINK_BOUND |
476				       I915_SHRINK_UNBOUND |
477				       I915_SHRINK_ACTIVE |
478				       I915_SHRINK_VMAPS);
479	intel_runtime_pm_put(i915);
480
481	/* We also want to clear any cached iomaps as they wrap vmap */
482	list_for_each_entry_safe(vma, next,
483				 &i915->ggtt.base.inactive_list, vm_link) {
484		unsigned long count = vma->node.size >> PAGE_SHIFT;
485		if (vma->iomap && i915_vma_unbind(vma) == 0)
486			freed_pages += count;
487	}
488
489out:
490	shrinker_unlock(i915, unlock);
491
492	*(unsigned long *)ptr += freed_pages;
493	return NOTIFY_DONE;
494}
495
496/**
497 * i915_gem_shrinker_register - Register the i915 shrinker
498 * @i915: i915 device
499 *
500 * This function registers and sets up the i915 shrinker and OOM handler.
501 */
502void i915_gem_shrinker_register(struct drm_i915_private *i915)
503{
504	i915->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
505	i915->mm.shrinker.count_objects = i915_gem_shrinker_count;
506	i915->mm.shrinker.seeks = DEFAULT_SEEKS;
507	i915->mm.shrinker.batch = 4096;
508	WARN_ON(register_shrinker(&i915->mm.shrinker));
509
510	i915->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
511	WARN_ON(register_oom_notifier(&i915->mm.oom_notifier));
512
513	i915->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
514	WARN_ON(register_vmap_purge_notifier(&i915->mm.vmap_notifier));
515}
516
517/**
518 * i915_gem_shrinker_unregister - Unregisters the i915 shrinker
519 * @i915: i915 device
520 *
521 * This function unregisters the i915 shrinker and OOM handler.
522 */
523void i915_gem_shrinker_unregister(struct drm_i915_private *i915)
524{
525	WARN_ON(unregister_vmap_purge_notifier(&i915->mm.vmap_notifier));
526	WARN_ON(unregister_oom_notifier(&i915->mm.oom_notifier));
527	unregister_shrinker(&i915->mm.shrinker);
528}