1/*
  2 * SPDX-License-Identifier: MIT
  3 *
  4 * Copyright © 2008-2015 Intel Corporation
  5 */
  6
  7#include <linux/oom.h>
  8#include <linux/sched/mm.h>
  9#include <linux/shmem_fs.h>
 10#include <linux/slab.h>
 11#include <linux/swap.h>
 12#include <linux/pci.h>
 13#include <linux/dma-buf.h>
 14#include <linux/vmalloc.h>
 15
 16#include "gt/intel_gt_requests.h"
 17
 18#include "i915_trace.h"
 19
 20static bool swap_available(void)
 21{
 22	return get_nr_swap_pages() > 0;
 23}
 24
 25static bool can_release_pages(struct drm_i915_gem_object *obj)
 26{
 27	/* Consider only shrinkable ojects. */
 28	if (!i915_gem_object_is_shrinkable(obj))
 29		return false;
 30
 31	/*
 32	 * We can only return physical pages to the system if we can either
 33	 * discard the contents (because the user has marked them as being
 34	 * purgeable) or if we can move their contents out to swap.
 35	 */
 36	return swap_available() || obj->mm.madv == I915_MADV_DONTNEED;
 37}
 38
 39static bool unsafe_drop_pages(struct drm_i915_gem_object *obj,
 40			      unsigned long shrink)
 41{
 42	unsigned long flags;
 43
 44	flags = 0;
 45	if (shrink & I915_SHRINK_ACTIVE)
 46		flags = I915_GEM_OBJECT_UNBIND_ACTIVE;
 47	if (!(shrink & I915_SHRINK_BOUND))
 48		flags = I915_GEM_OBJECT_UNBIND_TEST;
 49
 50	if (i915_gem_object_unbind(obj, flags) == 0)
 51		__i915_gem_object_put_pages(obj);
 52
 53	return !i915_gem_object_has_pages(obj);
 54}
 55
 56static void try_to_writeback(struct drm_i915_gem_object *obj,
 57			     unsigned int flags)
 58{
 59	switch (obj->mm.madv) {
 60	case I915_MADV_DONTNEED:
 61		i915_gem_object_truncate(obj);
 62	case __I915_MADV_PURGED:
 63		return;
 64	}
 65
 66	if (flags & I915_SHRINK_WRITEBACK)
 67		i915_gem_object_writeback(obj);
 68}
 69
 70/**
 71 * i915_gem_shrink - Shrink buffer object caches
 72 * @i915: i915 device
 73 * @target: amount of memory to make available, in pages
 74 * @nr_scanned: optional output for number of pages scanned (incremental)
 75 * @shrink: control flags for selecting cache types
 76 *
 77 * This function is the main interface to the shrinker. It will try to release
 78 * up to @target pages of main memory backing storage from buffer objects.
 79 * Selection of the specific caches can be done with @flags. This is e.g. useful
 80 * when purgeable objects should be removed from caches preferentially.
 81 *
 82 * Note that it's not guaranteed that released amount is actually available as
 83 * free system memory - the pages might still be in-used to due to other reasons
 84 * (like cpu mmaps) or the mm core has reused them before we could grab them.
 85 * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
 86 * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
 87 *
 88 * Also note that any kind of pinning (both per-vma address space pins and
 89 * backing storage pins at the buffer object level) result in the shrinker code
 90 * having to skip the object.
 91 *
 92 * Returns:
 93 * The number of pages of backing storage actually released.
 94 */
 95unsigned long
 96i915_gem_shrink(struct drm_i915_private *i915,
 97		unsigned long target,
 98		unsigned long *nr_scanned,
 99		unsigned int shrink)
100{
101	const struct {
102		struct list_head *list;
103		unsigned int bit;
104	} phases[] = {
105		{ &i915->mm.purge_list, ~0u },
106		{
107			&i915->mm.shrink_list,
108			I915_SHRINK_BOUND | I915_SHRINK_UNBOUND
109		},
110		{ NULL, 0 },
111	}, *phase;
112	intel_wakeref_t wakeref = 0;
113	unsigned long count = 0;
114	unsigned long scanned = 0;
115
116	trace_i915_gem_shrink(i915, target, shrink);
117
118	/*
119	 * Unbinding of objects will require HW access; Let us not wake the
120	 * device just to recover a little memory. If absolutely necessary,
121	 * we will force the wake during oom-notifier.
122	 */
123	if (shrink & I915_SHRINK_BOUND) {
124		wakeref = intel_runtime_pm_get_if_in_use(&i915->runtime_pm);
125		if (!wakeref)
126			shrink &= ~I915_SHRINK_BOUND;
127	}
128
129	/*
130	 * When shrinking the active list, we should also consider active
131	 * contexts. Active contexts are pinned until they are retired, and
132	 * so can not be simply unbound to retire and unpin their pages. To
133	 * shrink the contexts, we must wait until the gpu is idle and
134	 * completed its switch to the kernel context. In short, we do
135	 * not have a good mechanism for idling a specific context, but
136	 * what we can do is give them a kick so that we do not keep idle
137	 * contexts around longer than is necessary.
138	 */
139	if (shrink & I915_SHRINK_ACTIVE)
140		/* Retire requests to unpin all idle contexts */
141		intel_gt_retire_requests(&i915->gt);
142
143	/*
144	 * As we may completely rewrite the (un)bound list whilst unbinding
145	 * (due to retiring requests) we have to strictly process only
146	 * one element of the list at the time, and recheck the list
147	 * on every iteration.
148	 *
149	 * In particular, we must hold a reference whilst removing the
150	 * object as we may end up waiting for and/or retiring the objects.
151	 * This might release the final reference (held by the active list)
152	 * and result in the object being freed from under us. This is
153	 * similar to the precautions the eviction code must take whilst
154	 * removing objects.
155	 *
156	 * Also note that although these lists do not hold a reference to
157	 * the object we can safely grab one here: The final object
158	 * unreferencing and the bound_list are both protected by the
159	 * dev->struct_mutex and so we won't ever be able to observe an
160	 * object on the bound_list with a reference count equals 0.
161	 */
162	for (phase = phases; phase->list; phase++) {
163		struct list_head still_in_list;
164		struct drm_i915_gem_object *obj;
165		unsigned long flags;
166
167		if ((shrink & phase->bit) == 0)
168			continue;
169
170		INIT_LIST_HEAD(&still_in_list);
171
172		/*
173		 * We serialize our access to unreferenced objects through
174		 * the use of the struct_mutex. While the objects are not
175		 * yet freed (due to RCU then a workqueue) we still want
176		 * to be able to shrink their pages, so they remain on
177		 * the unbound/bound list until actually freed.
178		 */
179		spin_lock_irqsave(&i915->mm.obj_lock, flags);
180		while (count < target &&
181		       (obj = list_first_entry_or_null(phase->list,
182						       typeof(*obj),
183						       mm.link))) {
184			list_move_tail(&obj->mm.link, &still_in_list);
185
186			if (shrink & I915_SHRINK_VMAPS &&
187			    !is_vmalloc_addr(obj->mm.mapping))
188				continue;
189
190			if (!(shrink & I915_SHRINK_ACTIVE) &&
191			    i915_gem_object_is_framebuffer(obj))
192				continue;
193
194			if (!can_release_pages(obj))
195				continue;
196
197			if (!kref_get_unless_zero(&obj->base.refcount))
198				continue;
199
200			spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
201
202			if (unsafe_drop_pages(obj, shrink)) {
203				/* May arrive from get_pages on another bo */
204				mutex_lock(&obj->mm.lock);
205				if (!i915_gem_object_has_pages(obj)) {
206					try_to_writeback(obj, shrink);
207					count += obj->base.size >> PAGE_SHIFT;
208				}
209				mutex_unlock(&obj->mm.lock);
210			}
211
212			scanned += obj->base.size >> PAGE_SHIFT;
213			i915_gem_object_put(obj);
214
215			spin_lock_irqsave(&i915->mm.obj_lock, flags);
216		}
217		list_splice_tail(&still_in_list, phase->list);
218		spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
219	}
220
221	if (shrink & I915_SHRINK_BOUND)
222		intel_runtime_pm_put(&i915->runtime_pm, wakeref);
223
224	if (nr_scanned)
225		*nr_scanned += scanned;
226	return count;
227}
228
229/**
230 * i915_gem_shrink_all - Shrink buffer object caches completely
231 * @i915: i915 device
232 *
233 * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
234 * caches completely. It also first waits for and retires all outstanding
235 * requests to also be able to release backing storage for active objects.
236 *
237 * This should only be used in code to intentionally quiescent the gpu or as a
238 * last-ditch effort when memory seems to have run out.
239 *
240 * Returns:
241 * The number of pages of backing storage actually released.
242 */
243unsigned long i915_gem_shrink_all(struct drm_i915_private *i915)
244{
245	intel_wakeref_t wakeref;
246	unsigned long freed = 0;
247
248	with_intel_runtime_pm(&i915->runtime_pm, wakeref) {
249		freed = i915_gem_shrink(i915, -1UL, NULL,
250					I915_SHRINK_BOUND |
251					I915_SHRINK_UNBOUND);
252	}
253
254	return freed;
255}
256
257static unsigned long
258i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
259{
260	struct drm_i915_private *i915 =
261		container_of(shrinker, struct drm_i915_private, mm.shrinker);
262	unsigned long num_objects;
263	unsigned long count;
264
265	count = READ_ONCE(i915->mm.shrink_memory) >> PAGE_SHIFT;
266	num_objects = READ_ONCE(i915->mm.shrink_count);
267
268	/*
269	 * Update our preferred vmscan batch size for the next pass.
270	 * Our rough guess for an effective batch size is roughly 2
271	 * available GEM objects worth of pages. That is we don't want
272	 * the shrinker to fire, until it is worth the cost of freeing an
273	 * entire GEM object.
274	 */
275	if (num_objects) {
276		unsigned long avg = 2 * count / num_objects;
277
278		i915->mm.shrinker.batch =
279			max((i915->mm.shrinker.batch + avg) >> 1,
280			    128ul /* default SHRINK_BATCH */);
281	}
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 *i915 =
290		container_of(shrinker, struct drm_i915_private, mm.shrinker);
291	unsigned long freed;
292
293	sc->nr_scanned = 0;
294
295	freed = i915_gem_shrink(i915,
296				sc->nr_to_scan,
297				&sc->nr_scanned,
298				I915_SHRINK_BOUND |
299				I915_SHRINK_UNBOUND);
300	if (sc->nr_scanned < sc->nr_to_scan && current_is_kswapd()) {
301		intel_wakeref_t wakeref;
302
303		with_intel_runtime_pm(&i915->runtime_pm, wakeref) {
304			freed += i915_gem_shrink(i915,
305						 sc->nr_to_scan - sc->nr_scanned,
306						 &sc->nr_scanned,
307						 I915_SHRINK_ACTIVE |
308						 I915_SHRINK_BOUND |
309						 I915_SHRINK_UNBOUND |
310						 I915_SHRINK_WRITEBACK);
311		}
312	}
313
314	return sc->nr_scanned ? freed : SHRINK_STOP;
315}
316
317static int
318i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
319{
320	struct drm_i915_private *i915 =
321		container_of(nb, struct drm_i915_private, mm.oom_notifier);
322	struct drm_i915_gem_object *obj;
323	unsigned long unevictable, available, freed_pages;
324	intel_wakeref_t wakeref;
325	unsigned long flags;
326
327	freed_pages = 0;
328	with_intel_runtime_pm(&i915->runtime_pm, wakeref)
329		freed_pages += i915_gem_shrink(i915, -1UL, NULL,
330					       I915_SHRINK_BOUND |
331					       I915_SHRINK_UNBOUND |
332					       I915_SHRINK_WRITEBACK);
333
334	/* Because we may be allocating inside our own driver, we cannot
335	 * assert that there are no objects with pinned pages that are not
336	 * being pointed to by hardware.
337	 */
338	available = unevictable = 0;
339	spin_lock_irqsave(&i915->mm.obj_lock, flags);
340	list_for_each_entry(obj, &i915->mm.shrink_list, mm.link) {
341		if (!can_release_pages(obj))
342			unevictable += obj->base.size >> PAGE_SHIFT;
343		else
344			available += obj->base.size >> PAGE_SHIFT;
345	}
346	spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
347
348	if (freed_pages || available)
349		pr_info("Purging GPU memory, %lu pages freed, "
350			"%lu pages still pinned, %lu pages left available.\n",
351			freed_pages, unevictable, available);
352
353	*(unsigned long *)ptr += freed_pages;
354	return NOTIFY_DONE;
355}
356
357static int
358i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
359{
360	struct drm_i915_private *i915 =
361		container_of(nb, struct drm_i915_private, mm.vmap_notifier);
362	struct i915_vma *vma, *next;
363	unsigned long freed_pages = 0;
364	intel_wakeref_t wakeref;
365
366	with_intel_runtime_pm(&i915->runtime_pm, wakeref)
367		freed_pages += i915_gem_shrink(i915, -1UL, NULL,
368					       I915_SHRINK_BOUND |
369					       I915_SHRINK_UNBOUND |
370					       I915_SHRINK_VMAPS);
371
372	/* We also want to clear any cached iomaps as they wrap vmap */
373	mutex_lock(&i915->ggtt.vm.mutex);
374	list_for_each_entry_safe(vma, next,
375				 &i915->ggtt.vm.bound_list, vm_link) {
376		unsigned long count = vma->node.size >> PAGE_SHIFT;
377
378		if (!vma->iomap || i915_vma_is_active(vma))
379			continue;
380
381		if (__i915_vma_unbind(vma) == 0)
382			freed_pages += count;
383	}
384	mutex_unlock(&i915->ggtt.vm.mutex);
385
386	*(unsigned long *)ptr += freed_pages;
387	return NOTIFY_DONE;
388}
389
390void i915_gem_driver_register__shrinker(struct drm_i915_private *i915)
391{
392	i915->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
393	i915->mm.shrinker.count_objects = i915_gem_shrinker_count;
394	i915->mm.shrinker.seeks = DEFAULT_SEEKS;
395	i915->mm.shrinker.batch = 4096;
396	drm_WARN_ON(&i915->drm, register_shrinker(&i915->mm.shrinker));
397
398	i915->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
399	drm_WARN_ON(&i915->drm, register_oom_notifier(&i915->mm.oom_notifier));
400
401	i915->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
402	drm_WARN_ON(&i915->drm,
403		    register_vmap_purge_notifier(&i915->mm.vmap_notifier));
404}
405
406void i915_gem_driver_unregister__shrinker(struct drm_i915_private *i915)
407{
408	drm_WARN_ON(&i915->drm,
409		    unregister_vmap_purge_notifier(&i915->mm.vmap_notifier));
410	drm_WARN_ON(&i915->drm,
411		    unregister_oom_notifier(&i915->mm.oom_notifier));
412	unregister_shrinker(&i915->mm.shrinker);
413}
414
415void i915_gem_shrinker_taints_mutex(struct drm_i915_private *i915,
416				    struct mutex *mutex)
417{
418	if (!IS_ENABLED(CONFIG_LOCKDEP))
419		return;
420
421	fs_reclaim_acquire(GFP_KERNEL);
422
423	mutex_acquire(&mutex->dep_map, 0, 0, _RET_IP_);
424	mutex_release(&mutex->dep_map, _RET_IP_);
425
426	fs_reclaim_release(GFP_KERNEL);
427}
428
429#define obj_to_i915(obj__) to_i915((obj__)->base.dev)
430
431void i915_gem_object_make_unshrinkable(struct drm_i915_gem_object *obj)
432{
433	struct drm_i915_private *i915 = obj_to_i915(obj);
434	unsigned long flags;
435
436	/*
437	 * We can only be called while the pages are pinned or when
438	 * the pages are released. If pinned, we should only be called
439	 * from a single caller under controlled conditions; and on release
440	 * only one caller may release us. Neither the two may cross.
441	 */
442	if (atomic_add_unless(&obj->mm.shrink_pin, 1, 0))
443		return;
444
445	spin_lock_irqsave(&i915->mm.obj_lock, flags);
446	if (!atomic_fetch_inc(&obj->mm.shrink_pin) &&
447	    !list_empty(&obj->mm.link)) {
448		list_del_init(&obj->mm.link);
449		i915->mm.shrink_count--;
450		i915->mm.shrink_memory -= obj->base.size;
451	}
452	spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
453}
454
455static void __i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj,
456					      struct list_head *head)
457{
458	struct drm_i915_private *i915 = obj_to_i915(obj);
459	unsigned long flags;
460
461	GEM_BUG_ON(!i915_gem_object_has_pages(obj));
462	if (!i915_gem_object_is_shrinkable(obj))
463		return;
464
465	if (atomic_add_unless(&obj->mm.shrink_pin, -1, 1))
466		return;
467
468	spin_lock_irqsave(&i915->mm.obj_lock, flags);
469	GEM_BUG_ON(!kref_read(&obj->base.refcount));
470	if (atomic_dec_and_test(&obj->mm.shrink_pin)) {
471		GEM_BUG_ON(!list_empty(&obj->mm.link));
472
473		list_add_tail(&obj->mm.link, head);
474		i915->mm.shrink_count++;
475		i915->mm.shrink_memory += obj->base.size;
476
477	}
478	spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
479}
480
481void i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj)
482{
483	__i915_gem_object_make_shrinkable(obj,
484					  &obj_to_i915(obj)->mm.shrink_list);
485}
486
487void i915_gem_object_make_purgeable(struct drm_i915_gem_object *obj)
488{
489	__i915_gem_object_make_shrinkable(obj,
490					  &obj_to_i915(obj)->mm.purge_list);
491}