Loading...
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_slab_state(node, WORKINGSET_NODES);
384 }
385 } else {
386 if (!list_empty(&node->private_list)) {
387 list_lru_del(&shadow_nodes, &node->private_list);
388 __dec_lruvec_slab_state(node, WORKINGSET_NODES);
389 }
390 }
391}
392
393static unsigned long count_shadow_nodes(struct shrinker *shrinker,
394 struct shrink_control *sc)
395{
396 unsigned long max_nodes;
397 unsigned long nodes;
398 unsigned long pages;
399
400 nodes = list_lru_shrink_count(&shadow_nodes, sc);
401
402 /*
403 * Approximate a reasonable limit for the nodes
404 * containing shadow entries. We don't need to keep more
405 * shadow entries than possible pages on the active list,
406 * since refault distances bigger than that are dismissed.
407 *
408 * The size of the active list converges toward 100% of
409 * overall page cache as memory grows, with only a tiny
410 * inactive list. Assume the total cache size for that.
411 *
412 * Nodes might be sparsely populated, with only one shadow
413 * entry in the extreme case. Obviously, we cannot keep one
414 * node for every eligible shadow entry, so compromise on a
415 * worst-case density of 1/8th. Below that, not all eligible
416 * refaults can be detected anymore.
417 *
418 * On 64-bit with 7 xa_nodes per page and 64 slots
419 * each, this will reclaim shadow entries when they consume
420 * ~1.8% of available memory:
421 *
422 * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE
423 */
424#ifdef CONFIG_MEMCG
425 if (sc->memcg) {
426 struct lruvec *lruvec;
427 int i;
428
429 lruvec = mem_cgroup_lruvec(NODE_DATA(sc->nid), sc->memcg);
430 for (pages = 0, i = 0; i < NR_LRU_LISTS; i++)
431 pages += lruvec_page_state_local(lruvec,
432 NR_LRU_BASE + i);
433 pages += lruvec_page_state_local(lruvec, NR_SLAB_RECLAIMABLE);
434 pages += lruvec_page_state_local(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_slab_state(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_slab_state(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);
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/mm_inline.h>
10#include <linux/writeback.h>
11#include <linux/shmem_fs.h>
12#include <linux/pagemap.h>
13#include <linux/atomic.h>
14#include <linux/module.h>
15#include <linux/swap.h>
16#include <linux/dax.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include "internal.h"
20
21/*
22 * Double CLOCK lists
23 *
24 * Per node, two clock lists are maintained for file pages: the
25 * inactive and the active list. Freshly faulted pages start out at
26 * the head of the inactive list and page reclaim scans pages from the
27 * tail. Pages that are accessed multiple times on the inactive list
28 * are promoted to the active list, to protect them from reclaim,
29 * whereas active pages are demoted to the inactive list when the
30 * active list grows too big.
31 *
32 * fault ------------------------+
33 * |
34 * +--------------+ | +-------------+
35 * reclaim <- | inactive | <-+-- demotion | active | <--+
36 * +--------------+ +-------------+ |
37 * | |
38 * +-------------- promotion ------------------+
39 *
40 *
41 * Access frequency and refault distance
42 *
43 * A workload is thrashing when its pages are frequently used but they
44 * are evicted from the inactive list every time before another access
45 * would have promoted them to the active list.
46 *
47 * In cases where the average access distance between thrashing pages
48 * is bigger than the size of memory there is nothing that can be
49 * done - the thrashing set could never fit into memory under any
50 * circumstance.
51 *
52 * However, the average access distance could be bigger than the
53 * inactive list, yet smaller than the size of memory. In this case,
54 * the set could fit into memory if it weren't for the currently
55 * active pages - which may be used more, hopefully less frequently:
56 *
57 * +-memory available to cache-+
58 * | |
59 * +-inactive------+-active----+
60 * a b | c d e f g h i | J K L M N |
61 * +---------------+-----------+
62 *
63 * It is prohibitively expensive to accurately track access frequency
64 * of pages. But a reasonable approximation can be made to measure
65 * thrashing on the inactive list, after which refaulting pages can be
66 * activated optimistically to compete with the existing active pages.
67 *
68 * Approximating inactive page access frequency - Observations:
69 *
70 * 1. When a page is accessed for the first time, it is added to the
71 * head of the inactive list, slides every existing inactive page
72 * towards the tail by one slot, and pushes the current tail page
73 * out of memory.
74 *
75 * 2. When a page is accessed for the second time, it is promoted to
76 * the active list, shrinking the inactive list by one slot. This
77 * also slides all inactive pages that were faulted into the cache
78 * more recently than the activated page towards the tail of the
79 * inactive list.
80 *
81 * Thus:
82 *
83 * 1. The sum of evictions and activations between any two points in
84 * time indicate the minimum number of inactive pages accessed in
85 * between.
86 *
87 * 2. Moving one inactive page N page slots towards the tail of the
88 * list requires at least N inactive page accesses.
89 *
90 * Combining these:
91 *
92 * 1. When a page is finally evicted from memory, the number of
93 * inactive pages accessed while the page was in cache is at least
94 * the number of page slots on the inactive list.
95 *
96 * 2. In addition, measuring the sum of evictions and activations (E)
97 * at the time of a page's eviction, and comparing it to another
98 * reading (R) at the time the page faults back into memory tells
99 * the minimum number of accesses while the page was not cached.
100 * This is called the refault distance.
101 *
102 * Because the first access of the page was the fault and the second
103 * access the refault, we combine the in-cache distance with the
104 * out-of-cache distance to get the complete minimum access distance
105 * of this page:
106 *
107 * NR_inactive + (R - E)
108 *
109 * And knowing the minimum access distance of a page, we can easily
110 * tell if the page would be able to stay in cache assuming all page
111 * slots in the cache were available:
112 *
113 * NR_inactive + (R - E) <= NR_inactive + NR_active
114 *
115 * If we have swap we should consider about NR_inactive_anon and
116 * NR_active_anon, so for page cache and anonymous respectively:
117 *
118 * NR_inactive_file + (R - E) <= NR_inactive_file + NR_active_file
119 * + NR_inactive_anon + NR_active_anon
120 *
121 * NR_inactive_anon + (R - E) <= NR_inactive_anon + NR_active_anon
122 * + NR_inactive_file + NR_active_file
123 *
124 * Which can be further simplified to:
125 *
126 * (R - E) <= NR_active_file + NR_inactive_anon + NR_active_anon
127 *
128 * (R - E) <= NR_active_anon + NR_inactive_file + NR_active_file
129 *
130 * Put into words, the refault distance (out-of-cache) can be seen as
131 * a deficit in inactive list space (in-cache). If the inactive list
132 * had (R - E) more page slots, the page would not have been evicted
133 * in between accesses, but activated instead. And on a full system,
134 * the only thing eating into inactive list space is active pages.
135 *
136 *
137 * Refaulting inactive pages
138 *
139 * All that is known about the active list is that the pages have been
140 * accessed more than once in the past. This means that at any given
141 * time there is actually a good chance that pages on the active list
142 * are no longer in active use.
143 *
144 * So when a refault distance of (R - E) is observed and there are at
145 * least (R - E) pages in the userspace workingset, the refaulting page
146 * is activated optimistically in the hope that (R - E) pages are actually
147 * used less frequently than the refaulting page - or even not used at
148 * all anymore.
149 *
150 * That means if inactive cache is refaulting with a suitable refault
151 * distance, we assume the cache workingset is transitioning and put
152 * pressure on the current workingset.
153 *
154 * If this is wrong and demotion kicks in, the pages which are truly
155 * used more frequently will be reactivated while the less frequently
156 * used once will be evicted from memory.
157 *
158 * But if this is right, the stale pages will be pushed out of memory
159 * and the used pages get to stay in cache.
160 *
161 * Refaulting active pages
162 *
163 * If on the other hand the refaulting pages have recently been
164 * deactivated, it means that the active list is no longer protecting
165 * actively used cache from reclaim. The cache is NOT transitioning to
166 * a different workingset; the existing workingset is thrashing in the
167 * space allocated to the page cache.
168 *
169 *
170 * Implementation
171 *
172 * For each node's LRU lists, a counter for inactive evictions and
173 * activations is maintained (node->nonresident_age).
174 *
175 * On eviction, a snapshot of this counter (along with some bits to
176 * identify the node) is stored in the now empty page cache
177 * slot of the evicted page. This is called a shadow entry.
178 *
179 * On cache misses for which there are shadow entries, an eligible
180 * refault distance will immediately activate the refaulting page.
181 */
182
183#define WORKINGSET_SHIFT 1
184#define EVICTION_SHIFT ((BITS_PER_LONG - BITS_PER_XA_VALUE) + \
185 WORKINGSET_SHIFT + NODES_SHIFT + \
186 MEM_CGROUP_ID_SHIFT)
187#define EVICTION_MASK (~0UL >> EVICTION_SHIFT)
188
189/*
190 * Eviction timestamps need to be able to cover the full range of
191 * actionable refaults. However, bits are tight in the xarray
192 * entry, and after storing the identifier for the lruvec there might
193 * not be enough left to represent every single actionable refault. In
194 * that case, we have to sacrifice granularity for distance, and group
195 * evictions into coarser buckets by shaving off lower timestamp bits.
196 */
197static unsigned int bucket_order __read_mostly;
198
199static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction,
200 bool workingset)
201{
202 eviction &= EVICTION_MASK;
203 eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
204 eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
205 eviction = (eviction << WORKINGSET_SHIFT) | workingset;
206
207 return xa_mk_value(eviction);
208}
209
210static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
211 unsigned long *evictionp, bool *workingsetp)
212{
213 unsigned long entry = xa_to_value(shadow);
214 int memcgid, nid;
215 bool workingset;
216
217 workingset = entry & ((1UL << WORKINGSET_SHIFT) - 1);
218 entry >>= WORKINGSET_SHIFT;
219 nid = entry & ((1UL << NODES_SHIFT) - 1);
220 entry >>= NODES_SHIFT;
221 memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
222 entry >>= MEM_CGROUP_ID_SHIFT;
223
224 *memcgidp = memcgid;
225 *pgdat = NODE_DATA(nid);
226 *evictionp = entry;
227 *workingsetp = workingset;
228}
229
230#ifdef CONFIG_LRU_GEN
231
232static void *lru_gen_eviction(struct folio *folio)
233{
234 int hist;
235 unsigned long token;
236 unsigned long min_seq;
237 struct lruvec *lruvec;
238 struct lru_gen_folio *lrugen;
239 int type = folio_is_file_lru(folio);
240 int delta = folio_nr_pages(folio);
241 int refs = folio_lru_refs(folio);
242 int tier = lru_tier_from_refs(refs);
243 struct mem_cgroup *memcg = folio_memcg(folio);
244 struct pglist_data *pgdat = folio_pgdat(folio);
245
246 BUILD_BUG_ON(LRU_GEN_WIDTH + LRU_REFS_WIDTH > BITS_PER_LONG - EVICTION_SHIFT);
247
248 lruvec = mem_cgroup_lruvec(memcg, pgdat);
249 lrugen = &lruvec->lrugen;
250 min_seq = READ_ONCE(lrugen->min_seq[type]);
251 token = (min_seq << LRU_REFS_WIDTH) | max(refs - 1, 0);
252
253 hist = lru_hist_from_seq(min_seq);
254 atomic_long_add(delta, &lrugen->evicted[hist][type][tier]);
255
256 return pack_shadow(mem_cgroup_id(memcg), pgdat, token, refs);
257}
258
259/*
260 * Tests if the shadow entry is for a folio that was recently evicted.
261 * Fills in @lruvec, @token, @workingset with the values unpacked from shadow.
262 */
263static bool lru_gen_test_recent(void *shadow, bool file, struct lruvec **lruvec,
264 unsigned long *token, bool *workingset)
265{
266 int memcg_id;
267 unsigned long min_seq;
268 struct mem_cgroup *memcg;
269 struct pglist_data *pgdat;
270
271 unpack_shadow(shadow, &memcg_id, &pgdat, token, workingset);
272
273 memcg = mem_cgroup_from_id(memcg_id);
274 *lruvec = mem_cgroup_lruvec(memcg, pgdat);
275
276 min_seq = READ_ONCE((*lruvec)->lrugen.min_seq[file]);
277 return (*token >> LRU_REFS_WIDTH) == (min_seq & (EVICTION_MASK >> LRU_REFS_WIDTH));
278}
279
280static void lru_gen_refault(struct folio *folio, void *shadow)
281{
282 bool recent;
283 int hist, tier, refs;
284 bool workingset;
285 unsigned long token;
286 struct lruvec *lruvec;
287 struct lru_gen_folio *lrugen;
288 int type = folio_is_file_lru(folio);
289 int delta = folio_nr_pages(folio);
290
291 rcu_read_lock();
292
293 recent = lru_gen_test_recent(shadow, type, &lruvec, &token, &workingset);
294 if (lruvec != folio_lruvec(folio))
295 goto unlock;
296
297 mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + type, delta);
298
299 if (!recent)
300 goto unlock;
301
302 lrugen = &lruvec->lrugen;
303
304 hist = lru_hist_from_seq(READ_ONCE(lrugen->min_seq[type]));
305 /* see the comment in folio_lru_refs() */
306 refs = (token & (BIT(LRU_REFS_WIDTH) - 1)) + workingset;
307 tier = lru_tier_from_refs(refs);
308
309 atomic_long_add(delta, &lrugen->refaulted[hist][type][tier]);
310 mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta);
311
312 /*
313 * Count the following two cases as stalls:
314 * 1. For pages accessed through page tables, hotter pages pushed out
315 * hot pages which refaulted immediately.
316 * 2. For pages accessed multiple times through file descriptors,
317 * they would have been protected by sort_folio().
318 */
319 if (lru_gen_in_fault() || refs >= BIT(LRU_REFS_WIDTH) - 1) {
320 set_mask_bits(&folio->flags, 0, LRU_REFS_MASK | BIT(PG_workingset));
321 mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + type, delta);
322 }
323unlock:
324 rcu_read_unlock();
325}
326
327#else /* !CONFIG_LRU_GEN */
328
329static void *lru_gen_eviction(struct folio *folio)
330{
331 return NULL;
332}
333
334static bool lru_gen_test_recent(void *shadow, bool file, struct lruvec **lruvec,
335 unsigned long *token, bool *workingset)
336{
337 return false;
338}
339
340static void lru_gen_refault(struct folio *folio, void *shadow)
341{
342}
343
344#endif /* CONFIG_LRU_GEN */
345
346/**
347 * workingset_age_nonresident - age non-resident entries as LRU ages
348 * @lruvec: the lruvec that was aged
349 * @nr_pages: the number of pages to count
350 *
351 * As in-memory pages are aged, non-resident pages need to be aged as
352 * well, in order for the refault distances later on to be comparable
353 * to the in-memory dimensions. This function allows reclaim and LRU
354 * operations to drive the non-resident aging along in parallel.
355 */
356void workingset_age_nonresident(struct lruvec *lruvec, unsigned long nr_pages)
357{
358 /*
359 * Reclaiming a cgroup means reclaiming all its children in a
360 * round-robin fashion. That means that each cgroup has an LRU
361 * order that is composed of the LRU orders of its child
362 * cgroups; and every page has an LRU position not just in the
363 * cgroup that owns it, but in all of that group's ancestors.
364 *
365 * So when the physical inactive list of a leaf cgroup ages,
366 * the virtual inactive lists of all its parents, including
367 * the root cgroup's, age as well.
368 */
369 do {
370 atomic_long_add(nr_pages, &lruvec->nonresident_age);
371 } while ((lruvec = parent_lruvec(lruvec)));
372}
373
374/**
375 * workingset_eviction - note the eviction of a folio from memory
376 * @target_memcg: the cgroup that is causing the reclaim
377 * @folio: the folio being evicted
378 *
379 * Return: a shadow entry to be stored in @folio->mapping->i_pages in place
380 * of the evicted @folio so that a later refault can be detected.
381 */
382void *workingset_eviction(struct folio *folio, struct mem_cgroup *target_memcg)
383{
384 struct pglist_data *pgdat = folio_pgdat(folio);
385 unsigned long eviction;
386 struct lruvec *lruvec;
387 int memcgid;
388
389 /* Folio is fully exclusive and pins folio's memory cgroup pointer */
390 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
391 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
392 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
393
394 if (lru_gen_enabled())
395 return lru_gen_eviction(folio);
396
397 lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
398 /* XXX: target_memcg can be NULL, go through lruvec */
399 memcgid = mem_cgroup_id(lruvec_memcg(lruvec));
400 eviction = atomic_long_read(&lruvec->nonresident_age);
401 eviction >>= bucket_order;
402 workingset_age_nonresident(lruvec, folio_nr_pages(folio));
403 return pack_shadow(memcgid, pgdat, eviction,
404 folio_test_workingset(folio));
405}
406
407/**
408 * workingset_test_recent - tests if the shadow entry is for a folio that was
409 * recently evicted. Also fills in @workingset with the value unpacked from
410 * shadow.
411 * @shadow: the shadow entry to be tested.
412 * @file: whether the corresponding folio is from the file lru.
413 * @workingset: where the workingset value unpacked from shadow should
414 * be stored.
415 * @flush: whether to flush cgroup rstat.
416 *
417 * Return: true if the shadow is for a recently evicted folio; false otherwise.
418 */
419bool workingset_test_recent(void *shadow, bool file, bool *workingset,
420 bool flush)
421{
422 struct mem_cgroup *eviction_memcg;
423 struct lruvec *eviction_lruvec;
424 unsigned long refault_distance;
425 unsigned long workingset_size;
426 unsigned long refault;
427 int memcgid;
428 struct pglist_data *pgdat;
429 unsigned long eviction;
430
431 rcu_read_lock();
432
433 if (lru_gen_enabled()) {
434 bool recent = lru_gen_test_recent(shadow, file,
435 &eviction_lruvec, &eviction, workingset);
436
437 rcu_read_unlock();
438 return recent;
439 }
440
441
442 unpack_shadow(shadow, &memcgid, &pgdat, &eviction, workingset);
443 eviction <<= bucket_order;
444
445 /*
446 * Look up the memcg associated with the stored ID. It might
447 * have been deleted since the folio's eviction.
448 *
449 * Note that in rare events the ID could have been recycled
450 * for a new cgroup that refaults a shared folio. This is
451 * impossible to tell from the available data. However, this
452 * should be a rare and limited disturbance, and activations
453 * are always speculative anyway. Ultimately, it's the aging
454 * algorithm's job to shake out the minimum access frequency
455 * for the active cache.
456 *
457 * XXX: On !CONFIG_MEMCG, this will always return NULL; it
458 * would be better if the root_mem_cgroup existed in all
459 * configurations instead.
460 */
461 eviction_memcg = mem_cgroup_from_id(memcgid);
462 if (!mem_cgroup_disabled() &&
463 (!eviction_memcg || !mem_cgroup_tryget(eviction_memcg))) {
464 rcu_read_unlock();
465 return false;
466 }
467
468 rcu_read_unlock();
469
470 /*
471 * Flush stats (and potentially sleep) outside the RCU read section.
472 *
473 * Note that workingset_test_recent() itself might be called in RCU read
474 * section (for e.g, in cachestat) - these callers need to skip flushing
475 * stats (via the flush argument).
476 *
477 * XXX: With per-memcg flushing and thresholding, is ratelimiting
478 * still needed here?
479 */
480 if (flush)
481 mem_cgroup_flush_stats_ratelimited(eviction_memcg);
482
483 eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
484 refault = atomic_long_read(&eviction_lruvec->nonresident_age);
485
486 /*
487 * Calculate the refault distance
488 *
489 * The unsigned subtraction here gives an accurate distance
490 * across nonresident_age overflows in most cases. There is a
491 * special case: usually, shadow entries have a short lifetime
492 * and are either refaulted or reclaimed along with the inode
493 * before they get too old. But it is not impossible for the
494 * nonresident_age to lap a shadow entry in the field, which
495 * can then result in a false small refault distance, leading
496 * to a false activation should this old entry actually
497 * refault again. However, earlier kernels used to deactivate
498 * unconditionally with *every* reclaim invocation for the
499 * longest time, so the occasional inappropriate activation
500 * leading to pressure on the active list is not a problem.
501 */
502 refault_distance = (refault - eviction) & EVICTION_MASK;
503
504 /*
505 * Compare the distance to the existing workingset size. We
506 * don't activate pages that couldn't stay resident even if
507 * all the memory was available to the workingset. Whether
508 * workingset competition needs to consider anon or not depends
509 * on having free swap space.
510 */
511 workingset_size = lruvec_page_state(eviction_lruvec, NR_ACTIVE_FILE);
512 if (!file) {
513 workingset_size += lruvec_page_state(eviction_lruvec,
514 NR_INACTIVE_FILE);
515 }
516 if (mem_cgroup_get_nr_swap_pages(eviction_memcg) > 0) {
517 workingset_size += lruvec_page_state(eviction_lruvec,
518 NR_ACTIVE_ANON);
519 if (file) {
520 workingset_size += lruvec_page_state(eviction_lruvec,
521 NR_INACTIVE_ANON);
522 }
523 }
524
525 mem_cgroup_put(eviction_memcg);
526 return refault_distance <= workingset_size;
527}
528
529/**
530 * workingset_refault - Evaluate the refault of a previously evicted folio.
531 * @folio: The freshly allocated replacement folio.
532 * @shadow: Shadow entry of the evicted folio.
533 *
534 * Calculates and evaluates the refault distance of the previously
535 * evicted folio in the context of the node and the memcg whose memory
536 * pressure caused the eviction.
537 */
538void workingset_refault(struct folio *folio, void *shadow)
539{
540 bool file = folio_is_file_lru(folio);
541 struct pglist_data *pgdat;
542 struct mem_cgroup *memcg;
543 struct lruvec *lruvec;
544 bool workingset;
545 long nr;
546
547 if (lru_gen_enabled()) {
548 lru_gen_refault(folio, shadow);
549 return;
550 }
551
552 /*
553 * The activation decision for this folio is made at the level
554 * where the eviction occurred, as that is where the LRU order
555 * during folio reclaim is being determined.
556 *
557 * However, the cgroup that will own the folio is the one that
558 * is actually experiencing the refault event. Make sure the folio is
559 * locked to guarantee folio_memcg() stability throughout.
560 */
561 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
562 nr = folio_nr_pages(folio);
563 memcg = folio_memcg(folio);
564 pgdat = folio_pgdat(folio);
565 lruvec = mem_cgroup_lruvec(memcg, pgdat);
566
567 mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr);
568
569 if (!workingset_test_recent(shadow, file, &workingset, true))
570 return;
571
572 folio_set_active(folio);
573 workingset_age_nonresident(lruvec, nr);
574 mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file, nr);
575
576 /* Folio was active prior to eviction */
577 if (workingset) {
578 folio_set_workingset(folio);
579 /*
580 * XXX: Move to folio_add_lru() when it supports new vs
581 * putback
582 */
583 lru_note_cost_refault(folio);
584 mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file, nr);
585 }
586}
587
588/**
589 * workingset_activation - note a page activation
590 * @folio: Folio that is being activated.
591 */
592void workingset_activation(struct folio *folio)
593{
594 /*
595 * Filter non-memcg pages here, e.g. unmap can call
596 * mark_page_accessed() on VDSO pages.
597 */
598 if (mem_cgroup_disabled() || folio_memcg_charged(folio))
599 workingset_age_nonresident(folio_lruvec(folio), folio_nr_pages(folio));
600}
601
602/*
603 * Shadow entries reflect the share of the working set that does not
604 * fit into memory, so their number depends on the access pattern of
605 * the workload. In most cases, they will refault or get reclaimed
606 * along with the inode, but a (malicious) workload that streams
607 * through files with a total size several times that of available
608 * memory, while preventing the inodes from being reclaimed, can
609 * create excessive amounts of shadow nodes. To keep a lid on this,
610 * track shadow nodes and reclaim them when they grow way past the
611 * point where they would still be useful.
612 */
613
614struct list_lru shadow_nodes;
615
616void workingset_update_node(struct xa_node *node)
617{
618 struct address_space *mapping;
619 struct page *page = virt_to_page(node);
620
621 /*
622 * Track non-empty nodes that contain only shadow entries;
623 * unlink those that contain pages or are being freed.
624 *
625 * Avoid acquiring the list_lru lock when the nodes are
626 * already where they should be. The list_empty() test is safe
627 * as node->private_list is protected by the i_pages lock.
628 */
629 mapping = container_of(node->array, struct address_space, i_pages);
630 lockdep_assert_held(&mapping->i_pages.xa_lock);
631
632 if (node->count && node->count == node->nr_values) {
633 if (list_empty(&node->private_list)) {
634 list_lru_add_obj(&shadow_nodes, &node->private_list);
635 __inc_node_page_state(page, WORKINGSET_NODES);
636 }
637 } else {
638 if (!list_empty(&node->private_list)) {
639 list_lru_del_obj(&shadow_nodes, &node->private_list);
640 __dec_node_page_state(page, WORKINGSET_NODES);
641 }
642 }
643}
644
645static unsigned long count_shadow_nodes(struct shrinker *shrinker,
646 struct shrink_control *sc)
647{
648 unsigned long max_nodes;
649 unsigned long nodes;
650 unsigned long pages;
651
652 nodes = list_lru_shrink_count(&shadow_nodes, sc);
653 if (!nodes)
654 return SHRINK_EMPTY;
655
656 /*
657 * Approximate a reasonable limit for the nodes
658 * containing shadow entries. We don't need to keep more
659 * shadow entries than possible pages on the active list,
660 * since refault distances bigger than that are dismissed.
661 *
662 * The size of the active list converges toward 100% of
663 * overall page cache as memory grows, with only a tiny
664 * inactive list. Assume the total cache size for that.
665 *
666 * Nodes might be sparsely populated, with only one shadow
667 * entry in the extreme case. Obviously, we cannot keep one
668 * node for every eligible shadow entry, so compromise on a
669 * worst-case density of 1/8th. Below that, not all eligible
670 * refaults can be detected anymore.
671 *
672 * On 64-bit with 7 xa_nodes per page and 64 slots
673 * each, this will reclaim shadow entries when they consume
674 * ~1.8% of available memory:
675 *
676 * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE
677 */
678#ifdef CONFIG_MEMCG
679 if (sc->memcg) {
680 struct lruvec *lruvec;
681 int i;
682
683 mem_cgroup_flush_stats_ratelimited(sc->memcg);
684 lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
685 for (pages = 0, i = 0; i < NR_LRU_LISTS; i++)
686 pages += lruvec_page_state_local(lruvec,
687 NR_LRU_BASE + i);
688 pages += lruvec_page_state_local(
689 lruvec, NR_SLAB_RECLAIMABLE_B) >> PAGE_SHIFT;
690 pages += lruvec_page_state_local(
691 lruvec, NR_SLAB_UNRECLAIMABLE_B) >> PAGE_SHIFT;
692 } else
693#endif
694 pages = node_present_pages(sc->nid);
695
696 max_nodes = pages >> (XA_CHUNK_SHIFT - 3);
697
698 if (nodes <= max_nodes)
699 return 0;
700 return nodes - max_nodes;
701}
702
703static enum lru_status shadow_lru_isolate(struct list_head *item,
704 struct list_lru_one *lru,
705 void *arg) __must_hold(lru->lock)
706{
707 struct xa_node *node = container_of(item, struct xa_node, private_list);
708 struct address_space *mapping;
709 int ret;
710
711 /*
712 * Page cache insertions and deletions synchronously maintain
713 * the shadow node LRU under the i_pages lock and the
714 * &lru->lock. Because the page cache tree is emptied before
715 * the inode can be destroyed, holding the &lru->lock pins any
716 * address_space that has nodes on the LRU.
717 *
718 * We can then safely transition to the i_pages lock to
719 * pin only the address_space of the particular node we want
720 * to reclaim, take the node off-LRU, and drop the &lru->lock.
721 */
722
723 mapping = container_of(node->array, struct address_space, i_pages);
724
725 /* Coming from the list, invert the lock order */
726 if (!xa_trylock(&mapping->i_pages)) {
727 spin_unlock_irq(&lru->lock);
728 ret = LRU_RETRY;
729 goto out;
730 }
731
732 /* For page cache we need to hold i_lock */
733 if (mapping->host != NULL) {
734 if (!spin_trylock(&mapping->host->i_lock)) {
735 xa_unlock(&mapping->i_pages);
736 spin_unlock_irq(&lru->lock);
737 ret = LRU_RETRY;
738 goto out;
739 }
740 }
741
742 list_lru_isolate(lru, item);
743 __dec_node_page_state(virt_to_page(node), WORKINGSET_NODES);
744
745 spin_unlock(&lru->lock);
746
747 /*
748 * The nodes should only contain one or more shadow entries,
749 * no pages, so we expect to be able to remove them all and
750 * delete and free the empty node afterwards.
751 */
752 if (WARN_ON_ONCE(!node->nr_values))
753 goto out_invalid;
754 if (WARN_ON_ONCE(node->count != node->nr_values))
755 goto out_invalid;
756 xa_delete_node(node, workingset_update_node);
757 __inc_lruvec_kmem_state(node, WORKINGSET_NODERECLAIM);
758
759out_invalid:
760 xa_unlock_irq(&mapping->i_pages);
761 if (mapping->host != NULL) {
762 if (mapping_shrinkable(mapping))
763 inode_add_lru(mapping->host);
764 spin_unlock(&mapping->host->i_lock);
765 }
766 ret = LRU_REMOVED_RETRY;
767out:
768 cond_resched();
769 return ret;
770}
771
772static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
773 struct shrink_control *sc)
774{
775 /* list_lru lock nests inside the IRQ-safe i_pages lock */
776 return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate,
777 NULL);
778}
779
780/*
781 * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
782 * i_pages lock.
783 */
784static struct lock_class_key shadow_nodes_key;
785
786static int __init workingset_init(void)
787{
788 struct shrinker *workingset_shadow_shrinker;
789 unsigned int timestamp_bits;
790 unsigned int max_order;
791 int ret = -ENOMEM;
792
793 BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
794 /*
795 * Calculate the eviction bucket size to cover the longest
796 * actionable refault distance, which is currently half of
797 * memory (totalram_pages/2). However, memory hotplug may add
798 * some more pages at runtime, so keep working with up to
799 * double the initial memory by using totalram_pages as-is.
800 */
801 timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
802 max_order = fls_long(totalram_pages() - 1);
803 if (max_order > timestamp_bits)
804 bucket_order = max_order - timestamp_bits;
805 pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
806 timestamp_bits, max_order, bucket_order);
807
808 workingset_shadow_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE |
809 SHRINKER_MEMCG_AWARE,
810 "mm-shadow");
811 if (!workingset_shadow_shrinker)
812 goto err;
813
814 ret = list_lru_init_memcg_key(&shadow_nodes, workingset_shadow_shrinker,
815 &shadow_nodes_key);
816 if (ret)
817 goto err_list_lru;
818
819 workingset_shadow_shrinker->count_objects = count_shadow_nodes;
820 workingset_shadow_shrinker->scan_objects = scan_shadow_nodes;
821 /* ->count reports only fully expendable nodes */
822 workingset_shadow_shrinker->seeks = 0;
823
824 shrinker_register(workingset_shadow_shrinker);
825 return 0;
826err_list_lru:
827 shrinker_free(workingset_shadow_shrinker);
828err:
829 return ret;
830}
831module_init(workingset_init);