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 */
182static unsigned int bucket_order __read_mostly;
183
184static 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
196static 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 * @page: the page being evicted
219 *
220 * Returns a shadow entry to be stored in @page->mapping->i_pages in place
221 * of the evicted @page so that a later refault can be detected.
222 */
223void *workingset_eviction(struct page *page)
224{
225 struct pglist_data *pgdat = page_pgdat(page);
226 struct mem_cgroup *memcg = page_memcg(page);
227 int memcgid = mem_cgroup_id(memcg);
228 unsigned long eviction;
229 struct lruvec *lruvec;
230
231 /* Page is fully exclusive and pins page->mem_cgroup */
232 VM_BUG_ON_PAGE(PageLRU(page), page);
233 VM_BUG_ON_PAGE(page_count(page), page);
234 VM_BUG_ON_PAGE(!PageLocked(page), page);
235
236 lruvec = mem_cgroup_lruvec(pgdat, memcg);
237 eviction = atomic_long_inc_return(&lruvec->inactive_age);
238 return pack_shadow(memcgid, pgdat, eviction, PageWorkingset(page));
239}
240
241/**
242 * workingset_refault - evaluate the refault of a previously evicted page
243 * @page: the freshly allocated replacement page
244 * @shadow: shadow entry of the evicted page
245 *
246 * Calculates and evaluates the refault distance of the previously
247 * evicted page in the context of the node it was allocated in.
248 */
249void workingset_refault(struct page *page, void *shadow)
250{
251 unsigned long refault_distance;
252 struct pglist_data *pgdat;
253 unsigned long active_file;
254 struct mem_cgroup *memcg;
255 unsigned long eviction;
256 struct lruvec *lruvec;
257 unsigned long refault;
258 bool workingset;
259 int memcgid;
260
261 unpack_shadow(shadow, &memcgid, &pgdat, &eviction, &workingset);
262
263 rcu_read_lock();
264 /*
265 * Look up the memcg associated with the stored ID. It might
266 * have been deleted since the page's eviction.
267 *
268 * Note that in rare events the ID could have been recycled
269 * for a new cgroup that refaults a shared page. This is
270 * impossible to tell from the available data. However, this
271 * should be a rare and limited disturbance, and activations
272 * are always speculative anyway. Ultimately, it's the aging
273 * algorithm's job to shake out the minimum access frequency
274 * for the active cache.
275 *
276 * XXX: On !CONFIG_MEMCG, this will always return NULL; it
277 * would be better if the root_mem_cgroup existed in all
278 * configurations instead.
279 */
280 memcg = mem_cgroup_from_id(memcgid);
281 if (!mem_cgroup_disabled() && !memcg)
282 goto out;
283 lruvec = mem_cgroup_lruvec(pgdat, memcg);
284 refault = atomic_long_read(&lruvec->inactive_age);
285 active_file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES);
286
287 /*
288 * Calculate the refault distance
289 *
290 * The unsigned subtraction here gives an accurate distance
291 * across inactive_age overflows in most cases. There is a
292 * special case: usually, shadow entries have a short lifetime
293 * and are either refaulted or reclaimed along with the inode
294 * before they get too old. But it is not impossible for the
295 * inactive_age to lap a shadow entry in the field, which can
296 * then result in a false small refault distance, leading to a
297 * false activation should this old entry actually refault
298 * again. However, earlier kernels used to deactivate
299 * unconditionally with *every* reclaim invocation for the
300 * longest time, so the occasional inappropriate activation
301 * leading to pressure on the active list is not a problem.
302 */
303 refault_distance = (refault - eviction) & EVICTION_MASK;
304
305 inc_lruvec_state(lruvec, WORKINGSET_REFAULT);
306
307 /*
308 * Compare the distance to the existing workingset size. We
309 * don't act on pages that couldn't stay resident even if all
310 * the memory was available to the page cache.
311 */
312 if (refault_distance > active_file)
313 goto out;
314
315 SetPageActive(page);
316 atomic_long_inc(&lruvec->inactive_age);
317 inc_lruvec_state(lruvec, WORKINGSET_ACTIVATE);
318
319 /* Page was active prior to eviction */
320 if (workingset) {
321 SetPageWorkingset(page);
322 inc_lruvec_state(lruvec, WORKINGSET_RESTORE);
323 }
324out:
325 rcu_read_unlock();
326}
327
328/**
329 * workingset_activation - note a page activation
330 * @page: page that is being activated
331 */
332void workingset_activation(struct page *page)
333{
334 struct mem_cgroup *memcg;
335 struct lruvec *lruvec;
336
337 rcu_read_lock();
338 /*
339 * Filter non-memcg pages here, e.g. unmap can call
340 * mark_page_accessed() on VDSO pages.
341 *
342 * XXX: See workingset_refault() - this should return
343 * root_mem_cgroup even for !CONFIG_MEMCG.
344 */
345 memcg = page_memcg_rcu(page);
346 if (!mem_cgroup_disabled() && !memcg)
347 goto out;
348 lruvec = mem_cgroup_lruvec(page_pgdat(page), memcg);
349 atomic_long_inc(&lruvec->inactive_age);
350out:
351 rcu_read_unlock();
352}
353
354/*
355 * Shadow entries reflect the share of the working set that does not
356 * fit into memory, so their number depends on the access pattern of
357 * the workload. In most cases, they will refault or get reclaimed
358 * along with the inode, but a (malicious) workload that streams
359 * through files with a total size several times that of available
360 * memory, while preventing the inodes from being reclaimed, can
361 * create excessive amounts of shadow nodes. To keep a lid on this,
362 * track shadow nodes and reclaim them when they grow way past the
363 * point where they would still be useful.
364 */
365
366static struct list_lru shadow_nodes;
367
368void workingset_update_node(struct xa_node *node)
369{
370 /*
371 * Track non-empty nodes that contain only shadow entries;
372 * unlink those that contain pages or are being freed.
373 *
374 * Avoid acquiring the list_lru lock when the nodes are
375 * already where they should be. The list_empty() test is safe
376 * as node->private_list is protected by the i_pages lock.
377 */
378 VM_WARN_ON_ONCE(!irqs_disabled()); /* For __inc_lruvec_page_state */
379
380 if (node->count && node->count == node->nr_values) {
381 if (list_empty(&node->private_list)) {
382 list_lru_add(&shadow_nodes, &node->private_list);
383 __inc_lruvec_page_state(virt_to_page(node),
384 WORKINGSET_NODES);
385 }
386 } else {
387 if (!list_empty(&node->private_list)) {
388 list_lru_del(&shadow_nodes, &node->private_list);
389 __dec_lruvec_page_state(virt_to_page(node),
390 WORKINGSET_NODES);
391 }
392 }
393}
394
395static unsigned long count_shadow_nodes(struct shrinker *shrinker,
396 struct shrink_control *sc)
397{
398 unsigned long max_nodes;
399 unsigned long nodes;
400 unsigned long pages;
401
402 nodes = list_lru_shrink_count(&shadow_nodes, sc);
403
404 /*
405 * Approximate a reasonable limit for the nodes
406 * containing shadow entries. We don't need to keep more
407 * shadow entries than possible pages on the active list,
408 * since refault distances bigger than that are dismissed.
409 *
410 * The size of the active list converges toward 100% of
411 * overall page cache as memory grows, with only a tiny
412 * inactive list. Assume the total cache size for that.
413 *
414 * Nodes might be sparsely populated, with only one shadow
415 * entry in the extreme case. Obviously, we cannot keep one
416 * node for every eligible shadow entry, so compromise on a
417 * worst-case density of 1/8th. Below that, not all eligible
418 * refaults can be detected anymore.
419 *
420 * On 64-bit with 7 xa_nodes per page and 64 slots
421 * each, this will reclaim shadow entries when they consume
422 * ~1.8% of available memory:
423 *
424 * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE
425 */
426#ifdef CONFIG_MEMCG
427 if (sc->memcg) {
428 struct lruvec *lruvec;
429
430 pages = mem_cgroup_node_nr_lru_pages(sc->memcg, sc->nid,
431 LRU_ALL);
432 lruvec = mem_cgroup_lruvec(NODE_DATA(sc->nid), sc->memcg);
433 pages += lruvec_page_state(lruvec, NR_SLAB_RECLAIMABLE);
434 pages += lruvec_page_state(lruvec, NR_SLAB_UNRECLAIMABLE);
435 } else
436#endif
437 pages = node_present_pages(sc->nid);
438
439 max_nodes = pages >> (XA_CHUNK_SHIFT - 3);
440
441 if (!nodes)
442 return SHRINK_EMPTY;
443
444 if (nodes <= max_nodes)
445 return 0;
446 return nodes - max_nodes;
447}
448
449static enum lru_status shadow_lru_isolate(struct list_head *item,
450 struct list_lru_one *lru,
451 spinlock_t *lru_lock,
452 void *arg) __must_hold(lru_lock)
453{
454 struct xa_node *node = container_of(item, struct xa_node, private_list);
455 XA_STATE(xas, node->array, 0);
456 struct address_space *mapping;
457 int ret;
458
459 /*
460 * Page cache insertions and deletions synchroneously maintain
461 * the shadow node LRU under the i_pages lock and the
462 * lru_lock. Because the page cache tree is emptied before
463 * the inode can be destroyed, holding the lru_lock pins any
464 * address_space that has nodes on the LRU.
465 *
466 * We can then safely transition to the i_pages lock to
467 * pin only the address_space of the particular node we want
468 * to reclaim, take the node off-LRU, and drop the lru_lock.
469 */
470
471 mapping = container_of(node->array, struct address_space, i_pages);
472
473 /* Coming from the list, invert the lock order */
474 if (!xa_trylock(&mapping->i_pages)) {
475 spin_unlock_irq(lru_lock);
476 ret = LRU_RETRY;
477 goto out;
478 }
479
480 list_lru_isolate(lru, item);
481 __dec_lruvec_page_state(virt_to_page(node), WORKINGSET_NODES);
482
483 spin_unlock(lru_lock);
484
485 /*
486 * The nodes should only contain one or more shadow entries,
487 * no pages, so we expect to be able to remove them all and
488 * delete and free the empty node afterwards.
489 */
490 if (WARN_ON_ONCE(!node->nr_values))
491 goto out_invalid;
492 if (WARN_ON_ONCE(node->count != node->nr_values))
493 goto out_invalid;
494 mapping->nrexceptional -= node->nr_values;
495 xas.xa_node = xa_parent_locked(&mapping->i_pages, node);
496 xas.xa_offset = node->offset;
497 xas.xa_shift = node->shift + XA_CHUNK_SHIFT;
498 xas_set_update(&xas, workingset_update_node);
499 /*
500 * We could store a shadow entry here which was the minimum of the
501 * shadow entries we were tracking ...
502 */
503 xas_store(&xas, NULL);
504 __inc_lruvec_page_state(virt_to_page(node), WORKINGSET_NODERECLAIM);
505
506out_invalid:
507 xa_unlock_irq(&mapping->i_pages);
508 ret = LRU_REMOVED_RETRY;
509out:
510 cond_resched();
511 spin_lock_irq(lru_lock);
512 return ret;
513}
514
515static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
516 struct shrink_control *sc)
517{
518 /* list_lru lock nests inside the IRQ-safe i_pages lock */
519 return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate,
520 NULL);
521}
522
523static struct shrinker workingset_shadow_shrinker = {
524 .count_objects = count_shadow_nodes,
525 .scan_objects = scan_shadow_nodes,
526 .seeks = 0, /* ->count reports only fully expendable nodes */
527 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
528};
529
530/*
531 * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
532 * i_pages lock.
533 */
534static struct lock_class_key shadow_nodes_key;
535
536static int __init workingset_init(void)
537{
538 unsigned int timestamp_bits;
539 unsigned int max_order;
540 int ret;
541
542 BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
543 /*
544 * Calculate the eviction bucket size to cover the longest
545 * actionable refault distance, which is currently half of
546 * memory (totalram_pages/2). However, memory hotplug may add
547 * some more pages at runtime, so keep working with up to
548 * double the initial memory by using totalram_pages as-is.
549 */
550 timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
551 max_order = fls_long(totalram_pages() - 1);
552 if (max_order > timestamp_bits)
553 bucket_order = max_order - timestamp_bits;
554 pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
555 timestamp_bits, max_order, bucket_order);
556
557 ret = prealloc_shrinker(&workingset_shadow_shrinker);
558 if (ret)
559 goto err;
560 ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key,
561 &workingset_shadow_shrinker);
562 if (ret)
563 goto err_list_lru;
564 register_shrinker_prepared(&workingset_shadow_shrinker);
565 return 0;
566err_list_lru:
567 free_prealloced_shrinker(&workingset_shadow_shrinker);
568err:
569 return ret;
570}
571module_init(workingset_init);
572