i915_gem_shrinker.c 14.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
/*
 * Copyright © 2008-2015 Intel Corporation
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 *
 */

#include <linux/oom.h>
#include <linux/shmem_fs.h>
#include <linux/slab.h>
#include <linux/swap.h>
#include <linux/pci.h>
#include <linux/dma-buf.h>
31
#include <linux/vmalloc.h>
32 33 34 35 36 37
#include <drm/drmP.h>
#include <drm/i915_drm.h>

#include "i915_drv.h"
#include "i915_trace.h"

38
static bool shrinker_lock(struct drm_i915_private *dev_priv, bool *unlock)
39
{
40
	switch (mutex_trylock_recursive(&dev_priv->drm.struct_mutex)) {
41
	case MUTEX_TRYLOCK_RECURSIVE:
42
		*unlock = false;
43
		return true;
44 45

	case MUTEX_TRYLOCK_FAILED:
46 47
		*unlock = false;
		preempt_disable();
48 49 50 51
		do {
			cpu_relax();
			if (mutex_trylock(&dev_priv->drm.struct_mutex)) {
				*unlock = true;
52
				break;
53 54
			}
		} while (!need_resched());
55 56
		preempt_enable();
		return *unlock;
57

58 59 60
	case MUTEX_TRYLOCK_SUCCESS:
		*unlock = true;
		return true;
61 62
	}

63
	BUG();
64 65
}

66
static void shrinker_unlock(struct drm_i915_private *dev_priv, bool unlock)
67 68 69 70
{
	if (!unlock)
		return;

71
	mutex_unlock(&dev_priv->drm.struct_mutex);
72 73
}

74
static bool any_vma_pinned(struct drm_i915_gem_object *obj)
75 76 77
{
	struct i915_vma *vma;

78 79 80 81 82 83 84 85
	list_for_each_entry(vma, &obj->vma_list, obj_link) {
		/* Only GGTT vma may be permanently pinned, and are always
		 * at the start of the list. We can stop hunting as soon
		 * as we see a ppGTT vma.
		 */
		if (!i915_vma_is_ggtt(vma))
			break;

86
		if (i915_vma_is_pinned(vma))
87
			return true;
88
	}
89

90
	return false;
91 92 93 94 95 96 97 98 99
}

static bool swap_available(void)
{
	return get_nr_swap_pages() > 0;
}

static bool can_release_pages(struct drm_i915_gem_object *obj)
{
100 101 102
	if (!obj->mm.pages)
		return false;

103 104
	/* Consider only shrinkable ojects. */
	if (!i915_gem_object_is_shrinkable(obj))
105 106
		return false;

107 108 109 110 111 112 113 114
	/* Only report true if by unbinding the object and putting its pages
	 * we can actually make forward progress towards freeing physical
	 * pages.
	 *
	 * If the pages are pinned for any other reason than being bound
	 * to the GPU, simply unbinding from the GPU is not going to succeed
	 * in releasing our pin count on the pages themselves.
	 */
115
	if (atomic_read(&obj->mm.pages_pin_count) > obj->bind_count)
116 117 118
		return false;

	if (any_vma_pinned(obj))
119 120 121 122 123 124
		return false;

	/* We can only return physical pages to the system if we can either
	 * discard the contents (because the user has marked them as being
	 * purgeable) or if we can move their contents out to swap.
	 */
C
Chris Wilson 已提交
125
	return swap_available() || obj->mm.madv == I915_MADV_DONTNEED;
126 127
}

128 129 130
static bool unsafe_drop_pages(struct drm_i915_gem_object *obj)
{
	if (i915_gem_object_unbind(obj) == 0)
131
		__i915_gem_object_put_pages(obj, I915_MM_SHRINKER);
132
	return !READ_ONCE(obj->mm.pages);
133 134
}

135 136 137 138
/**
 * i915_gem_shrink - Shrink buffer object caches
 * @dev_priv: i915 device
 * @target: amount of memory to make available, in pages
139
 * @nr_scanned: optional output for number of pages scanned (incremental)
140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159
 * @flags: control flags for selecting cache types
 *
 * This function is the main interface to the shrinker. It will try to release
 * up to @target pages of main memory backing storage from buffer objects.
 * Selection of the specific caches can be done with @flags. This is e.g. useful
 * when purgeable objects should be removed from caches preferentially.
 *
 * Note that it's not guaranteed that released amount is actually available as
 * free system memory - the pages might still be in-used to due to other reasons
 * (like cpu mmaps) or the mm core has reused them before we could grab them.
 * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
 * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
 *
 * Also note that any kind of pinning (both per-vma address space pins and
 * backing storage pins at the buffer object level) result in the shrinker code
 * having to skip the object.
 *
 * Returns:
 * The number of pages of backing storage actually released.
 */
160 161
unsigned long
i915_gem_shrink(struct drm_i915_private *dev_priv,
162 163 164
		unsigned long target,
		unsigned long *nr_scanned,
		unsigned flags)
165 166 167 168 169 170 171 172 173 174
{
	const struct {
		struct list_head *list;
		unsigned int bit;
	} phases[] = {
		{ &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
		{ &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
		{ NULL, 0 },
	}, *phase;
	unsigned long count = 0;
175
	unsigned long scanned = 0;
176 177
	bool unlock;

178
	if (!shrinker_lock(dev_priv, &unlock))
179
		return 0;
180

181
	trace_i915_gem_shrink(dev_priv, target, flags);
182
	i915_gem_retire_requests(dev_priv);
183

184 185 186 187 188 189 190 191 192
	/*
	 * Unbinding of objects will require HW access; Let us not wake the
	 * device just to recover a little memory. If absolutely necessary,
	 * we will force the wake during oom-notifier.
	 */
	if ((flags & I915_SHRINK_BOUND) &&
	    !intel_runtime_pm_get_if_in_use(dev_priv))
		flags &= ~I915_SHRINK_BOUND;

193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213
	/*
	 * As we may completely rewrite the (un)bound list whilst unbinding
	 * (due to retiring requests) we have to strictly process only
	 * one element of the list at the time, and recheck the list
	 * on every iteration.
	 *
	 * In particular, we must hold a reference whilst removing the
	 * object as we may end up waiting for and/or retiring the objects.
	 * This might release the final reference (held by the active list)
	 * and result in the object being freed from under us. This is
	 * similar to the precautions the eviction code must take whilst
	 * removing objects.
	 *
	 * Also note that although these lists do not hold a reference to
	 * the object we can safely grab one here: The final object
	 * unreferencing and the bound_list are both protected by the
	 * dev->struct_mutex and so we won't ever be able to observe an
	 * object on the bound_list with a reference count equals 0.
	 */
	for (phase = phases; phase->list; phase++) {
		struct list_head still_in_list;
214
		struct drm_i915_gem_object *obj;
215 216 217 218 219

		if ((flags & phase->bit) == 0)
			continue;

		INIT_LIST_HEAD(&still_in_list);
220 221 222
		while (count < target &&
		       (obj = list_first_entry_or_null(phase->list,
						       typeof(*obj),
223 224
						       global_link))) {
			list_move_tail(&obj->global_link, &still_in_list);
225
			if (!obj->mm.pages) {
226
				list_del_init(&obj->global_link);
227 228
				continue;
			}
229 230

			if (flags & I915_SHRINK_PURGEABLE &&
C
Chris Wilson 已提交
231
			    obj->mm.madv != I915_MADV_DONTNEED)
232 233
				continue;

234
			if (flags & I915_SHRINK_VMAPS &&
C
Chris Wilson 已提交
235
			    !is_vmalloc_addr(obj->mm.mapping))
236 237
				continue;

238 239
			if (!(flags & I915_SHRINK_ACTIVE) &&
			    (i915_gem_object_is_active(obj) ||
240
			     i915_gem_object_is_framebuffer(obj)))
241 242
				continue;

243 244 245
			if (!can_release_pages(obj))
				continue;

246
			if (unsafe_drop_pages(obj)) {
247 248
				/* May arrive from get_pages on another bo */
				mutex_lock_nested(&obj->mm.lock,
249
						  I915_MM_SHRINKER);
250 251
				if (!obj->mm.pages) {
					__i915_gem_object_invalidate(obj);
252
					list_del_init(&obj->global_link);
253 254 255
					count += obj->base.size >> PAGE_SHIFT;
				}
				mutex_unlock(&obj->mm.lock);
256
				scanned += obj->base.size >> PAGE_SHIFT;
257
			}
258
		}
259
		list_splice_tail(&still_in_list, phase->list);
260 261
	}

262 263 264
	if (flags & I915_SHRINK_BOUND)
		intel_runtime_pm_put(dev_priv);

265
	i915_gem_retire_requests(dev_priv);
266

267
	shrinker_unlock(dev_priv, unlock);
268

269 270
	if (nr_scanned)
		*nr_scanned += scanned;
271 272 273
	return count;
}

274
/**
275
 * i915_gem_shrink_all - Shrink buffer object caches completely
276 277 278 279 280 281 282 283 284 285 286 287
 * @dev_priv: i915 device
 *
 * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 * caches completely. It also first waits for and retires all outstanding
 * requests to also be able to release backing storage for active objects.
 *
 * This should only be used in code to intentionally quiescent the gpu or as a
 * last-ditch effort when memory seems to have run out.
 *
 * Returns:
 * The number of pages of backing storage actually released.
 */
288 289
unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv)
{
290 291
	unsigned long freed;

292
	intel_runtime_pm_get(dev_priv);
293
	freed = i915_gem_shrink(dev_priv, -1UL, NULL,
294 295 296
				I915_SHRINK_BOUND |
				I915_SHRINK_UNBOUND |
				I915_SHRINK_ACTIVE);
297 298
	intel_runtime_pm_put(dev_priv);

299
	return freed;
300 301 302 303 304 305 306 307 308 309 310
}

static unsigned long
i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
{
	struct drm_i915_private *dev_priv =
		container_of(shrinker, struct drm_i915_private, mm.shrinker);
	struct drm_i915_gem_object *obj;
	unsigned long count;
	bool unlock;

311
	if (!shrinker_lock(dev_priv, &unlock))
312 313
		return 0;

314 315
	i915_gem_retire_requests(dev_priv);

316
	count = 0;
317
	list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_link)
318
		if (can_release_pages(obj))
319 320
			count += obj->base.size >> PAGE_SHIFT;

321
	list_for_each_entry(obj, &dev_priv->mm.bound_list, global_link) {
322
		if (!i915_gem_object_is_active(obj) && can_release_pages(obj))
323 324 325
			count += obj->base.size >> PAGE_SHIFT;
	}

326
	shrinker_unlock(dev_priv, unlock);
327 328 329 330 331 332 333 334 335 336 337 338

	return count;
}

static unsigned long
i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
{
	struct drm_i915_private *dev_priv =
		container_of(shrinker, struct drm_i915_private, mm.shrinker);
	unsigned long freed;
	bool unlock;

339 340
	sc->nr_scanned = 0;

341
	if (!shrinker_lock(dev_priv, &unlock))
342 343 344 345
		return SHRINK_STOP;

	freed = i915_gem_shrink(dev_priv,
				sc->nr_to_scan,
346
				&sc->nr_scanned,
347 348 349 350 351
				I915_SHRINK_BOUND |
				I915_SHRINK_UNBOUND |
				I915_SHRINK_PURGEABLE);
	if (freed < sc->nr_to_scan)
		freed += i915_gem_shrink(dev_priv,
352 353
					 sc->nr_to_scan - sc->nr_scanned,
					 &sc->nr_scanned,
354 355
					 I915_SHRINK_BOUND |
					 I915_SHRINK_UNBOUND);
356 357 358
	if (freed < sc->nr_to_scan && current_is_kswapd()) {
		intel_runtime_pm_get(dev_priv);
		freed += i915_gem_shrink(dev_priv,
359 360
					 sc->nr_to_scan - sc->nr_scanned,
					 &sc->nr_scanned,
361 362 363 364 365
					 I915_SHRINK_ACTIVE |
					 I915_SHRINK_BOUND |
					 I915_SHRINK_UNBOUND);
		intel_runtime_pm_put(dev_priv);
	}
366

367
	shrinker_unlock(dev_priv, unlock);
368

369
	return sc->nr_scanned ? freed : SHRINK_STOP;
370 371
}

372
static bool
373 374
shrinker_lock_uninterruptible(struct drm_i915_private *dev_priv, bool *unlock,
			      int timeout_ms)
375
{
376 377 378
	unsigned long timeout = jiffies + msecs_to_jiffies_timeout(timeout_ms);

	do {
379
		if (i915_gem_wait_for_idle(dev_priv, 0) == 0 &&
380
		    shrinker_lock(dev_priv, unlock))
381
			break;
382 383 384 385

		schedule_timeout_killable(1);
		if (fatal_signal_pending(current))
			return false;
386 387

		if (time_after(jiffies, timeout)) {
388 389 390
			pr_err("Unable to lock GPU to purge memory.\n");
			return false;
		}
391
	} while (1);
392 393 394 395

	return true;
}

396 397 398 399 400 401
static int
i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
{
	struct drm_i915_private *dev_priv =
		container_of(nb, struct drm_i915_private, mm.oom_notifier);
	struct drm_i915_gem_object *obj;
402
	unsigned long unevictable, bound, unbound, freed_pages;
403
	bool unlock;
404

405
	if (!shrinker_lock_uninterruptible(dev_priv, &unlock, 5000))
406 407 408 409 410 411 412 413
		return NOTIFY_DONE;

	freed_pages = i915_gem_shrink_all(dev_priv);

	/* Because we may be allocating inside our own driver, we cannot
	 * assert that there are no objects with pinned pages that are not
	 * being pointed to by hardware.
	 */
414
	unbound = bound = unevictable = 0;
415
	list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_link) {
416 417 418
		if (!obj->mm.pages)
			continue;

419 420
		if (!can_release_pages(obj))
			unevictable += obj->base.size >> PAGE_SHIFT;
421
		else
422
			unbound += obj->base.size >> PAGE_SHIFT;
423
	}
424
	list_for_each_entry(obj, &dev_priv->mm.bound_list, global_link) {
425 426 427
		if (!obj->mm.pages)
			continue;

428 429
		if (!can_release_pages(obj))
			unevictable += obj->base.size >> PAGE_SHIFT;
430
		else
431
			bound += obj->base.size >> PAGE_SHIFT;
432 433
	}

434
	shrinker_unlock(dev_priv, unlock);
435 436

	if (freed_pages || unbound || bound)
437 438 439
		pr_info("Purging GPU memory, %lu pages freed, "
			"%lu pages still pinned.\n",
			freed_pages, unevictable);
440
	if (unbound || bound)
441
		pr_err("%lu and %lu pages still available in the "
442 443 444 445 446 447 448
		       "bound and unbound GPU page lists.\n",
		       bound, unbound);

	*(unsigned long *)ptr += freed_pages;
	return NOTIFY_DONE;
}

449 450 451 452 453
static int
i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
{
	struct drm_i915_private *dev_priv =
		container_of(nb, struct drm_i915_private, mm.vmap_notifier);
454 455
	struct i915_vma *vma, *next;
	unsigned long freed_pages = 0;
456
	bool unlock;
457
	int ret;
458

459
	if (!shrinker_lock_uninterruptible(dev_priv, &unlock, 5000))
460 461
		return NOTIFY_DONE;

462
	/* Force everything onto the inactive lists */
463
	ret = i915_gem_wait_for_idle(dev_priv, I915_WAIT_LOCKED);
464 465 466
	if (ret)
		goto out;

467
	intel_runtime_pm_get(dev_priv);
468
	freed_pages += i915_gem_shrink(dev_priv, -1UL, NULL,
469 470 471 472
				       I915_SHRINK_BOUND |
				       I915_SHRINK_UNBOUND |
				       I915_SHRINK_ACTIVE |
				       I915_SHRINK_VMAPS);
473
	intel_runtime_pm_put(dev_priv);
474 475 476 477 478 479 480 481

	/* We also want to clear any cached iomaps as they wrap vmap */
	list_for_each_entry_safe(vma, next,
				 &dev_priv->ggtt.base.inactive_list, vm_link) {
		unsigned long count = vma->node.size >> PAGE_SHIFT;
		if (vma->iomap && i915_vma_unbind(vma) == 0)
			freed_pages += count;
	}
482

483
out:
484
	shrinker_unlock(dev_priv, unlock);
485 486 487 488 489

	*(unsigned long *)ptr += freed_pages;
	return NOTIFY_DONE;
}

490 491 492 493 494 495
/**
 * i915_gem_shrinker_init - Initialize i915 shrinker
 * @dev_priv: i915 device
 *
 * This function registers and sets up the i915 shrinker and OOM handler.
 */
496 497 498 499 500
void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
{
	dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
	dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
	dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
501
	WARN_ON(register_shrinker(&dev_priv->mm.shrinker));
502 503

	dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
504
	WARN_ON(register_oom_notifier(&dev_priv->mm.oom_notifier));
505 506 507

	dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
	WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
508 509 510 511 512 513 514 515 516 517
}

/**
 * i915_gem_shrinker_cleanup - Clean up i915 shrinker
 * @dev_priv: i915 device
 *
 * This function unregisters the i915 shrinker and OOM handler.
 */
void i915_gem_shrinker_cleanup(struct drm_i915_private *dev_priv)
{
518
	WARN_ON(unregister_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
519 520
	WARN_ON(unregister_oom_notifier(&dev_priv->mm.oom_notifier));
	unregister_shrinker(&dev_priv->mm.shrinker);
521
}