xref: /openbmc/linux/mm/workingset.c (revision f80d2f08)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Workingset detection
4  *
5  * Copyright (C) 2013 Red Hat, Inc., Johannes Weiner
6  */
7 
8 #include <linux/memcontrol.h>
9 #include <linux/writeback.h>
10 #include <linux/shmem_fs.h>
11 #include <linux/pagemap.h>
12 #include <linux/atomic.h>
13 #include <linux/module.h>
14 #include <linux/swap.h>
15 #include <linux/dax.h>
16 #include <linux/fs.h>
17 #include <linux/mm.h>
18 
19 /*
20  *		Double CLOCK lists
21  *
22  * Per node, two clock lists are maintained for file pages: the
23  * inactive and the active list.  Freshly faulted pages start out at
24  * the head of the inactive list and page reclaim scans pages from the
25  * tail.  Pages that are accessed multiple times on the inactive list
26  * are promoted to the active list, to protect them from reclaim,
27  * whereas active pages are demoted to the inactive list when the
28  * active list grows too big.
29  *
30  *   fault ------------------------+
31  *                                 |
32  *              +--------------+   |            +-------------+
33  *   reclaim <- |   inactive   | <-+-- demotion |    active   | <--+
34  *              +--------------+                +-------------+    |
35  *                     |                                           |
36  *                     +-------------- promotion ------------------+
37  *
38  *
39  *		Access frequency and refault distance
40  *
41  * A workload is thrashing when its pages are frequently used but they
42  * are evicted from the inactive list every time before another access
43  * would have promoted them to the active list.
44  *
45  * In cases where the average access distance between thrashing pages
46  * is bigger than the size of memory there is nothing that can be
47  * done - the thrashing set could never fit into memory under any
48  * circumstance.
49  *
50  * However, the average access distance could be bigger than the
51  * inactive list, yet smaller than the size of memory.  In this case,
52  * the set could fit into memory if it weren't for the currently
53  * active pages - which may be used more, hopefully less frequently:
54  *
55  *      +-memory available to cache-+
56  *      |                           |
57  *      +-inactive------+-active----+
58  *  a b | c d e f g h i | J K L M N |
59  *      +---------------+-----------+
60  *
61  * It is prohibitively expensive to accurately track access frequency
62  * of pages.  But a reasonable approximation can be made to measure
63  * thrashing on the inactive list, after which refaulting pages can be
64  * activated optimistically to compete with the existing active pages.
65  *
66  * Approximating inactive page access frequency - Observations:
67  *
68  * 1. When a page is accessed for the first time, it is added to the
69  *    head of the inactive list, slides every existing inactive page
70  *    towards the tail by one slot, and pushes the current tail page
71  *    out of memory.
72  *
73  * 2. When a page is accessed for the second time, it is promoted to
74  *    the active list, shrinking the inactive list by one slot.  This
75  *    also slides all inactive pages that were faulted into the cache
76  *    more recently than the activated page towards the tail of the
77  *    inactive list.
78  *
79  * Thus:
80  *
81  * 1. The sum of evictions and activations between any two points in
82  *    time indicate the minimum number of inactive pages accessed in
83  *    between.
84  *
85  * 2. Moving one inactive page N page slots towards the tail of the
86  *    list requires at least N inactive page accesses.
87  *
88  * Combining these:
89  *
90  * 1. When a page is finally evicted from memory, the number of
91  *    inactive pages accessed while the page was in cache is at least
92  *    the number of page slots on the inactive list.
93  *
94  * 2. In addition, measuring the sum of evictions and activations (E)
95  *    at the time of a page's eviction, and comparing it to another
96  *    reading (R) at the time the page faults back into memory tells
97  *    the minimum number of accesses while the page was not cached.
98  *    This is called the refault distance.
99  *
100  * Because the first access of the page was the fault and the second
101  * access the refault, we combine the in-cache distance with the
102  * out-of-cache distance to get the complete minimum access distance
103  * of this page:
104  *
105  *      NR_inactive + (R - E)
106  *
107  * And knowing the minimum access distance of a page, we can easily
108  * tell if the page would be able to stay in cache assuming all page
109  * slots in the cache were available:
110  *
111  *   NR_inactive + (R - E) <= NR_inactive + NR_active
112  *
113  * which can be further simplified to
114  *
115  *   (R - E) <= NR_active
116  *
117  * Put into words, the refault distance (out-of-cache) can be seen as
118  * a deficit in inactive list space (in-cache).  If the inactive list
119  * had (R - E) more page slots, the page would not have been evicted
120  * in between accesses, but activated instead.  And on a full system,
121  * the only thing eating into inactive list space is active pages.
122  *
123  *
124  *		Refaulting inactive pages
125  *
126  * All that is known about the active list is that the pages have been
127  * accessed more than once in the past.  This means that at any given
128  * time there is actually a good chance that pages on the active list
129  * are no longer in active use.
130  *
131  * So when a refault distance of (R - E) is observed and there are at
132  * least (R - E) active pages, the refaulting page is activated
133  * optimistically in the hope that (R - E) active pages are actually
134  * used less frequently than the refaulting page - or even not used at
135  * all anymore.
136  *
137  * That means if inactive cache is refaulting with a suitable refault
138  * distance, we assume the cache workingset is transitioning and put
139  * pressure on the current active list.
140  *
141  * If this is wrong and demotion kicks in, the pages which are truly
142  * used more frequently will be reactivated while the less frequently
143  * used once will be evicted from memory.
144  *
145  * But if this is right, the stale pages will be pushed out of memory
146  * and the used pages get to stay in cache.
147  *
148  *		Refaulting active pages
149  *
150  * If on the other hand the refaulting pages have recently been
151  * deactivated, it means that the active list is no longer protecting
152  * actively used cache from reclaim. The cache is NOT transitioning to
153  * a different workingset; the existing workingset is thrashing in the
154  * space allocated to the page cache.
155  *
156  *
157  *		Implementation
158  *
159  * For each node's file LRU lists, a counter for inactive evictions
160  * and activations is maintained (node->inactive_age).
161  *
162  * On eviction, a snapshot of this counter (along with some bits to
163  * identify the node) is stored in the now empty page cache
164  * slot of the evicted page.  This is called a shadow entry.
165  *
166  * On cache misses for which there are shadow entries, an eligible
167  * refault distance will immediately activate the refaulting page.
168  */
169 
170 #define EVICTION_SHIFT	((BITS_PER_LONG - BITS_PER_XA_VALUE) +	\
171 			 1 + NODES_SHIFT + MEM_CGROUP_ID_SHIFT)
172 #define EVICTION_MASK	(~0UL >> EVICTION_SHIFT)
173 
174 /*
175  * Eviction timestamps need to be able to cover the full range of
176  * actionable refaults. However, bits are tight in the xarray
177  * entry, and after storing the identifier for the lruvec there might
178  * not be enough left to represent every single actionable refault. In
179  * that case, we have to sacrifice granularity for distance, and group
180  * evictions into coarser buckets by shaving off lower timestamp bits.
181  */
182 static unsigned int bucket_order __read_mostly;
183 
184 static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction,
185 			 bool workingset)
186 {
187 	eviction >>= bucket_order;
188 	eviction &= EVICTION_MASK;
189 	eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
190 	eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
191 	eviction = (eviction << 1) | workingset;
192 
193 	return xa_mk_value(eviction);
194 }
195 
196 static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
197 			  unsigned long *evictionp, bool *workingsetp)
198 {
199 	unsigned long entry = xa_to_value(shadow);
200 	int memcgid, nid;
201 	bool workingset;
202 
203 	workingset = entry & 1;
204 	entry >>= 1;
205 	nid = entry & ((1UL << NODES_SHIFT) - 1);
206 	entry >>= NODES_SHIFT;
207 	memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
208 	entry >>= MEM_CGROUP_ID_SHIFT;
209 
210 	*memcgidp = memcgid;
211 	*pgdat = NODE_DATA(nid);
212 	*evictionp = entry << bucket_order;
213 	*workingsetp = workingset;
214 }
215 
216 /**
217  * workingset_eviction - note the eviction of a page from memory
218  * @mapping: address space the page was backing
219  * @page: the page being evicted
220  *
221  * Returns a shadow entry to be stored in @mapping->i_pages in place
222  * of the evicted @page so that a later refault can be detected.
223  */
224 void *workingset_eviction(struct address_space *mapping, struct page *page)
225 {
226 	struct pglist_data *pgdat = page_pgdat(page);
227 	struct mem_cgroup *memcg = page_memcg(page);
228 	int memcgid = mem_cgroup_id(memcg);
229 	unsigned long eviction;
230 	struct lruvec *lruvec;
231 
232 	/* Page is fully exclusive and pins page->mem_cgroup */
233 	VM_BUG_ON_PAGE(PageLRU(page), page);
234 	VM_BUG_ON_PAGE(page_count(page), page);
235 	VM_BUG_ON_PAGE(!PageLocked(page), page);
236 
237 	lruvec = mem_cgroup_lruvec(pgdat, memcg);
238 	eviction = atomic_long_inc_return(&lruvec->inactive_age);
239 	return pack_shadow(memcgid, pgdat, eviction, PageWorkingset(page));
240 }
241 
242 /**
243  * workingset_refault - evaluate the refault of a previously evicted page
244  * @page: the freshly allocated replacement page
245  * @shadow: shadow entry of the evicted page
246  *
247  * Calculates and evaluates the refault distance of the previously
248  * evicted page in the context of the node it was allocated in.
249  */
250 void workingset_refault(struct page *page, void *shadow)
251 {
252 	unsigned long refault_distance;
253 	struct pglist_data *pgdat;
254 	unsigned long active_file;
255 	struct mem_cgroup *memcg;
256 	unsigned long eviction;
257 	struct lruvec *lruvec;
258 	unsigned long refault;
259 	bool workingset;
260 	int memcgid;
261 
262 	unpack_shadow(shadow, &memcgid, &pgdat, &eviction, &workingset);
263 
264 	rcu_read_lock();
265 	/*
266 	 * Look up the memcg associated with the stored ID. It might
267 	 * have been deleted since the page's eviction.
268 	 *
269 	 * Note that in rare events the ID could have been recycled
270 	 * for a new cgroup that refaults a shared page. This is
271 	 * impossible to tell from the available data. However, this
272 	 * should be a rare and limited disturbance, and activations
273 	 * are always speculative anyway. Ultimately, it's the aging
274 	 * algorithm's job to shake out the minimum access frequency
275 	 * for the active cache.
276 	 *
277 	 * XXX: On !CONFIG_MEMCG, this will always return NULL; it
278 	 * would be better if the root_mem_cgroup existed in all
279 	 * configurations instead.
280 	 */
281 	memcg = mem_cgroup_from_id(memcgid);
282 	if (!mem_cgroup_disabled() && !memcg)
283 		goto out;
284 	lruvec = mem_cgroup_lruvec(pgdat, memcg);
285 	refault = atomic_long_read(&lruvec->inactive_age);
286 	active_file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES);
287 
288 	/*
289 	 * Calculate the refault distance
290 	 *
291 	 * The unsigned subtraction here gives an accurate distance
292 	 * across inactive_age overflows in most cases. There is a
293 	 * special case: usually, shadow entries have a short lifetime
294 	 * and are either refaulted or reclaimed along with the inode
295 	 * before they get too old.  But it is not impossible for the
296 	 * inactive_age to lap a shadow entry in the field, which can
297 	 * then result in a false small refault distance, leading to a
298 	 * false activation should this old entry actually refault
299 	 * again.  However, earlier kernels used to deactivate
300 	 * unconditionally with *every* reclaim invocation for the
301 	 * longest time, so the occasional inappropriate activation
302 	 * leading to pressure on the active list is not a problem.
303 	 */
304 	refault_distance = (refault - eviction) & EVICTION_MASK;
305 
306 	inc_lruvec_state(lruvec, WORKINGSET_REFAULT);
307 
308 	/*
309 	 * Compare the distance to the existing workingset size. We
310 	 * don't act on pages that couldn't stay resident even if all
311 	 * the memory was available to the page cache.
312 	 */
313 	if (refault_distance > active_file)
314 		goto out;
315 
316 	SetPageActive(page);
317 	atomic_long_inc(&lruvec->inactive_age);
318 	inc_lruvec_state(lruvec, WORKINGSET_ACTIVATE);
319 
320 	/* Page was active prior to eviction */
321 	if (workingset) {
322 		SetPageWorkingset(page);
323 		inc_lruvec_state(lruvec, WORKINGSET_RESTORE);
324 	}
325 out:
326 	rcu_read_unlock();
327 }
328 
329 /**
330  * workingset_activation - note a page activation
331  * @page: page that is being activated
332  */
333 void workingset_activation(struct page *page)
334 {
335 	struct mem_cgroup *memcg;
336 	struct lruvec *lruvec;
337 
338 	rcu_read_lock();
339 	/*
340 	 * Filter non-memcg pages here, e.g. unmap can call
341 	 * mark_page_accessed() on VDSO pages.
342 	 *
343 	 * XXX: See workingset_refault() - this should return
344 	 * root_mem_cgroup even for !CONFIG_MEMCG.
345 	 */
346 	memcg = page_memcg_rcu(page);
347 	if (!mem_cgroup_disabled() && !memcg)
348 		goto out;
349 	lruvec = mem_cgroup_lruvec(page_pgdat(page), memcg);
350 	atomic_long_inc(&lruvec->inactive_age);
351 out:
352 	rcu_read_unlock();
353 }
354 
355 /*
356  * Shadow entries reflect the share of the working set that does not
357  * fit into memory, so their number depends on the access pattern of
358  * the workload.  In most cases, they will refault or get reclaimed
359  * along with the inode, but a (malicious) workload that streams
360  * through files with a total size several times that of available
361  * memory, while preventing the inodes from being reclaimed, can
362  * create excessive amounts of shadow nodes.  To keep a lid on this,
363  * track shadow nodes and reclaim them when they grow way past the
364  * point where they would still be useful.
365  */
366 
367 static struct list_lru shadow_nodes;
368 
369 void workingset_update_node(struct xa_node *node)
370 {
371 	/*
372 	 * Track non-empty nodes that contain only shadow entries;
373 	 * unlink those that contain pages or are being freed.
374 	 *
375 	 * Avoid acquiring the list_lru lock when the nodes are
376 	 * already where they should be. The list_empty() test is safe
377 	 * as node->private_list is protected by the i_pages lock.
378 	 */
379 	VM_WARN_ON_ONCE(!irqs_disabled());  /* For __inc_lruvec_page_state */
380 
381 	if (node->count && node->count == node->nr_values) {
382 		if (list_empty(&node->private_list)) {
383 			list_lru_add(&shadow_nodes, &node->private_list);
384 			__inc_lruvec_page_state(virt_to_page(node),
385 						WORKINGSET_NODES);
386 		}
387 	} else {
388 		if (!list_empty(&node->private_list)) {
389 			list_lru_del(&shadow_nodes, &node->private_list);
390 			__dec_lruvec_page_state(virt_to_page(node),
391 						WORKINGSET_NODES);
392 		}
393 	}
394 }
395 
396 static unsigned long count_shadow_nodes(struct shrinker *shrinker,
397 					struct shrink_control *sc)
398 {
399 	unsigned long max_nodes;
400 	unsigned long nodes;
401 	unsigned long pages;
402 
403 	nodes = list_lru_shrink_count(&shadow_nodes, sc);
404 
405 	/*
406 	 * Approximate a reasonable limit for the nodes
407 	 * containing shadow entries. We don't need to keep more
408 	 * shadow entries than possible pages on the active list,
409 	 * since refault distances bigger than that are dismissed.
410 	 *
411 	 * The size of the active list converges toward 100% of
412 	 * overall page cache as memory grows, with only a tiny
413 	 * inactive list. Assume the total cache size for that.
414 	 *
415 	 * Nodes might be sparsely populated, with only one shadow
416 	 * entry in the extreme case. Obviously, we cannot keep one
417 	 * node for every eligible shadow entry, so compromise on a
418 	 * worst-case density of 1/8th. Below that, not all eligible
419 	 * refaults can be detected anymore.
420 	 *
421 	 * On 64-bit with 7 xa_nodes per page and 64 slots
422 	 * each, this will reclaim shadow entries when they consume
423 	 * ~1.8% of available memory:
424 	 *
425 	 * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE
426 	 */
427 #ifdef CONFIG_MEMCG
428 	if (sc->memcg) {
429 		struct lruvec *lruvec;
430 
431 		pages = mem_cgroup_node_nr_lru_pages(sc->memcg, sc->nid,
432 						     LRU_ALL);
433 		lruvec = mem_cgroup_lruvec(NODE_DATA(sc->nid), sc->memcg);
434 		pages += lruvec_page_state(lruvec, NR_SLAB_RECLAIMABLE);
435 		pages += lruvec_page_state(lruvec, NR_SLAB_UNRECLAIMABLE);
436 	} else
437 #endif
438 		pages = node_present_pages(sc->nid);
439 
440 	max_nodes = pages >> (XA_CHUNK_SHIFT - 3);
441 
442 	if (!nodes)
443 		return SHRINK_EMPTY;
444 
445 	if (nodes <= max_nodes)
446 		return 0;
447 	return nodes - max_nodes;
448 }
449 
450 static enum lru_status shadow_lru_isolate(struct list_head *item,
451 					  struct list_lru_one *lru,
452 					  spinlock_t *lru_lock,
453 					  void *arg) __must_hold(lru_lock)
454 {
455 	struct xa_node *node = container_of(item, struct xa_node, private_list);
456 	XA_STATE(xas, node->array, 0);
457 	struct address_space *mapping;
458 	int ret;
459 
460 	/*
461 	 * Page cache insertions and deletions synchroneously maintain
462 	 * the shadow node LRU under the i_pages lock and the
463 	 * lru_lock.  Because the page cache tree is emptied before
464 	 * the inode can be destroyed, holding the lru_lock pins any
465 	 * address_space that has nodes on the LRU.
466 	 *
467 	 * We can then safely transition to the i_pages lock to
468 	 * pin only the address_space of the particular node we want
469 	 * to reclaim, take the node off-LRU, and drop the lru_lock.
470 	 */
471 
472 	mapping = container_of(node->array, struct address_space, i_pages);
473 
474 	/* Coming from the list, invert the lock order */
475 	if (!xa_trylock(&mapping->i_pages)) {
476 		spin_unlock_irq(lru_lock);
477 		ret = LRU_RETRY;
478 		goto out;
479 	}
480 
481 	list_lru_isolate(lru, item);
482 	__dec_lruvec_page_state(virt_to_page(node), WORKINGSET_NODES);
483 
484 	spin_unlock(lru_lock);
485 
486 	/*
487 	 * The nodes should only contain one or more shadow entries,
488 	 * no pages, so we expect to be able to remove them all and
489 	 * delete and free the empty node afterwards.
490 	 */
491 	if (WARN_ON_ONCE(!node->nr_values))
492 		goto out_invalid;
493 	if (WARN_ON_ONCE(node->count != node->nr_values))
494 		goto out_invalid;
495 	mapping->nrexceptional -= node->nr_values;
496 	xas.xa_node = xa_parent_locked(&mapping->i_pages, node);
497 	xas.xa_offset = node->offset;
498 	xas.xa_shift = node->shift + XA_CHUNK_SHIFT;
499 	xas_set_update(&xas, workingset_update_node);
500 	/*
501 	 * We could store a shadow entry here which was the minimum of the
502 	 * shadow entries we were tracking ...
503 	 */
504 	xas_store(&xas, NULL);
505 	__inc_lruvec_page_state(virt_to_page(node), WORKINGSET_NODERECLAIM);
506 
507 out_invalid:
508 	xa_unlock_irq(&mapping->i_pages);
509 	ret = LRU_REMOVED_RETRY;
510 out:
511 	cond_resched();
512 	spin_lock_irq(lru_lock);
513 	return ret;
514 }
515 
516 static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
517 				       struct shrink_control *sc)
518 {
519 	/* list_lru lock nests inside the IRQ-safe i_pages lock */
520 	return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate,
521 					NULL);
522 }
523 
524 static struct shrinker workingset_shadow_shrinker = {
525 	.count_objects = count_shadow_nodes,
526 	.scan_objects = scan_shadow_nodes,
527 	.seeks = 0, /* ->count reports only fully expendable nodes */
528 	.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
529 };
530 
531 /*
532  * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
533  * i_pages lock.
534  */
535 static struct lock_class_key shadow_nodes_key;
536 
537 static int __init workingset_init(void)
538 {
539 	unsigned int timestamp_bits;
540 	unsigned int max_order;
541 	int ret;
542 
543 	BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
544 	/*
545 	 * Calculate the eviction bucket size to cover the longest
546 	 * actionable refault distance, which is currently half of
547 	 * memory (totalram_pages/2). However, memory hotplug may add
548 	 * some more pages at runtime, so keep working with up to
549 	 * double the initial memory by using totalram_pages as-is.
550 	 */
551 	timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
552 	max_order = fls_long(totalram_pages() - 1);
553 	if (max_order > timestamp_bits)
554 		bucket_order = max_order - timestamp_bits;
555 	pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
556 	       timestamp_bits, max_order, bucket_order);
557 
558 	ret = prealloc_shrinker(&workingset_shadow_shrinker);
559 	if (ret)
560 		goto err;
561 	ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key,
562 			      &workingset_shadow_shrinker);
563 	if (ret)
564 		goto err_list_lru;
565 	register_shrinker_prepared(&workingset_shadow_shrinker);
566 	return 0;
567 err_list_lru:
568 	free_prealloced_shrinker(&workingset_shadow_shrinker);
569 err:
570 	return ret;
571 }
572 module_init(workingset_init);
573