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