Loading...
1/*
2 * Workingset detection
3 *
4 * Copyright (C) 2013 Red Hat, Inc., Johannes Weiner
5 */
6
7#include <linux/memcontrol.h>
8#include <linux/writeback.h>
9#include <linux/pagemap.h>
10#include <linux/atomic.h>
11#include <linux/module.h>
12#include <linux/swap.h>
13#include <linux/fs.h>
14#include <linux/mm.h>
15
16/*
17 * Double CLOCK lists
18 *
19 * Per zone, two clock lists are maintained for file pages: the
20 * inactive and the active list. Freshly faulted pages start out at
21 * the head of the inactive list and page reclaim scans pages from the
22 * tail. Pages that are accessed multiple times on the inactive list
23 * are promoted to the active list, to protect them from reclaim,
24 * whereas active pages are demoted to the inactive list when the
25 * active list grows too big.
26 *
27 * fault ------------------------+
28 * |
29 * +--------------+ | +-------------+
30 * reclaim <- | inactive | <-+-- demotion | active | <--+
31 * +--------------+ +-------------+ |
32 * | |
33 * +-------------- promotion ------------------+
34 *
35 *
36 * Access frequency and refault distance
37 *
38 * A workload is thrashing when its pages are frequently used but they
39 * are evicted from the inactive list every time before another access
40 * would have promoted them to the active list.
41 *
42 * In cases where the average access distance between thrashing pages
43 * is bigger than the size of memory there is nothing that can be
44 * done - the thrashing set could never fit into memory under any
45 * circumstance.
46 *
47 * However, the average access distance could be bigger than the
48 * inactive list, yet smaller than the size of memory. In this case,
49 * the set could fit into memory if it weren't for the currently
50 * active pages - which may be used more, hopefully less frequently:
51 *
52 * +-memory available to cache-+
53 * | |
54 * +-inactive------+-active----+
55 * a b | c d e f g h i | J K L M N |
56 * +---------------+-----------+
57 *
58 * It is prohibitively expensive to accurately track access frequency
59 * of pages. But a reasonable approximation can be made to measure
60 * thrashing on the inactive list, after which refaulting pages can be
61 * activated optimistically to compete with the existing active pages.
62 *
63 * Approximating inactive page access frequency - Observations:
64 *
65 * 1. When a page is accessed for the first time, it is added to the
66 * head of the inactive list, slides every existing inactive page
67 * towards the tail by one slot, and pushes the current tail page
68 * out of memory.
69 *
70 * 2. When a page is accessed for the second time, it is promoted to
71 * the active list, shrinking the inactive list by one slot. This
72 * also slides all inactive pages that were faulted into the cache
73 * more recently than the activated page towards the tail of the
74 * inactive list.
75 *
76 * Thus:
77 *
78 * 1. The sum of evictions and activations between any two points in
79 * time indicate the minimum number of inactive pages accessed in
80 * between.
81 *
82 * 2. Moving one inactive page N page slots towards the tail of the
83 * list requires at least N inactive page accesses.
84 *
85 * Combining these:
86 *
87 * 1. When a page is finally evicted from memory, the number of
88 * inactive pages accessed while the page was in cache is at least
89 * the number of page slots on the inactive list.
90 *
91 * 2. In addition, measuring the sum of evictions and activations (E)
92 * at the time of a page's eviction, and comparing it to another
93 * reading (R) at the time the page faults back into memory tells
94 * the minimum number of accesses while the page was not cached.
95 * This is called the refault distance.
96 *
97 * Because the first access of the page was the fault and the second
98 * access the refault, we combine the in-cache distance with the
99 * out-of-cache distance to get the complete minimum access distance
100 * of this page:
101 *
102 * NR_inactive + (R - E)
103 *
104 * And knowing the minimum access distance of a page, we can easily
105 * tell if the page would be able to stay in cache assuming all page
106 * slots in the cache were available:
107 *
108 * NR_inactive + (R - E) <= NR_inactive + NR_active
109 *
110 * which can be further simplified to
111 *
112 * (R - E) <= NR_active
113 *
114 * Put into words, the refault distance (out-of-cache) can be seen as
115 * a deficit in inactive list space (in-cache). If the inactive list
116 * had (R - E) more page slots, the page would not have been evicted
117 * in between accesses, but activated instead. And on a full system,
118 * the only thing eating into inactive list space is active pages.
119 *
120 *
121 * Activating refaulting pages
122 *
123 * All that is known about the active list is that the pages have been
124 * accessed more than once in the past. This means that at any given
125 * time there is actually a good chance that pages on the active list
126 * are no longer in active use.
127 *
128 * So when a refault distance of (R - E) is observed and there are at
129 * least (R - E) active pages, the refaulting page is activated
130 * optimistically in the hope that (R - E) active pages are actually
131 * used less frequently than the refaulting page - or even not used at
132 * all anymore.
133 *
134 * If this is wrong and demotion kicks in, the pages which are truly
135 * used more frequently will be reactivated while the less frequently
136 * used once will be evicted from memory.
137 *
138 * But if this is right, the stale pages will be pushed out of memory
139 * and the used pages get to stay in cache.
140 *
141 *
142 * Implementation
143 *
144 * For each zone's file LRU lists, a counter for inactive evictions
145 * and activations is maintained (zone->inactive_age).
146 *
147 * On eviction, a snapshot of this counter (along with some bits to
148 * identify the zone) is stored in the now empty page cache radix tree
149 * slot of the evicted page. This is called a shadow entry.
150 *
151 * On cache misses for which there are shadow entries, an eligible
152 * refault distance will immediately activate the refaulting page.
153 */
154
155#define EVICTION_SHIFT (RADIX_TREE_EXCEPTIONAL_ENTRY + \
156 ZONES_SHIFT + NODES_SHIFT + \
157 MEM_CGROUP_ID_SHIFT)
158#define EVICTION_MASK (~0UL >> EVICTION_SHIFT)
159
160/*
161 * Eviction timestamps need to be able to cover the full range of
162 * actionable refaults. However, bits are tight in the radix tree
163 * entry, and after storing the identifier for the lruvec there might
164 * not be enough left to represent every single actionable refault. In
165 * that case, we have to sacrifice granularity for distance, and group
166 * evictions into coarser buckets by shaving off lower timestamp bits.
167 */
168static unsigned int bucket_order __read_mostly;
169
170static void *pack_shadow(int memcgid, struct zone *zone, unsigned long eviction)
171{
172 eviction >>= bucket_order;
173 eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
174 eviction = (eviction << NODES_SHIFT) | zone_to_nid(zone);
175 eviction = (eviction << ZONES_SHIFT) | zone_idx(zone);
176 eviction = (eviction << RADIX_TREE_EXCEPTIONAL_SHIFT);
177
178 return (void *)(eviction | RADIX_TREE_EXCEPTIONAL_ENTRY);
179}
180
181static void unpack_shadow(void *shadow, int *memcgidp, struct zone **zonep,
182 unsigned long *evictionp)
183{
184 unsigned long entry = (unsigned long)shadow;
185 int memcgid, nid, zid;
186
187 entry >>= RADIX_TREE_EXCEPTIONAL_SHIFT;
188 zid = entry & ((1UL << ZONES_SHIFT) - 1);
189 entry >>= ZONES_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 *zonep = NODE_DATA(nid)->node_zones + zid;
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->page_tree 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 zone *zone = page_zone(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_zone_lruvec(zone, memcg);
222 eviction = atomic_long_inc_return(&lruvec->inactive_age);
223 return pack_shadow(memcgid, zone, 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 zone 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 zone *zone;
244 int memcgid;
245
246 unpack_shadow(shadow, &memcgid, &zone, &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_zone_lruvec(zone, memcg);
271 refault = atomic_long_read(&lruvec->inactive_age);
272 active_file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE);
273 rcu_read_unlock();
274
275 /*
276 * The unsigned subtraction here gives an accurate distance
277 * across inactive_age overflows in most cases.
278 *
279 * There is a special case: usually, shadow entries have a
280 * short lifetime and are either refaulted or reclaimed along
281 * with the inode before they get too old. But it is not
282 * impossible for the inactive_age to lap a shadow entry in
283 * the field, which can then can result in a false small
284 * refault distance, leading to a false activation should this
285 * old entry actually refault again. However, earlier kernels
286 * used to deactivate unconditionally with *every* reclaim
287 * invocation for the longest time, so the occasional
288 * inappropriate activation leading to pressure on the active
289 * list is not a problem.
290 */
291 refault_distance = (refault - eviction) & EVICTION_MASK;
292
293 inc_zone_state(zone, WORKINGSET_REFAULT);
294
295 if (refault_distance <= active_file) {
296 inc_zone_state(zone, WORKINGSET_ACTIVATE);
297 return true;
298 }
299 return false;
300}
301
302/**
303 * workingset_activation - note a page activation
304 * @page: page that is being activated
305 */
306void workingset_activation(struct page *page)
307{
308 struct lruvec *lruvec;
309
310 lock_page_memcg(page);
311 /*
312 * Filter non-memcg pages here, e.g. unmap can call
313 * mark_page_accessed() on VDSO pages.
314 *
315 * XXX: See workingset_refault() - this should return
316 * root_mem_cgroup even for !CONFIG_MEMCG.
317 */
318 if (!mem_cgroup_disabled() && !page_memcg(page))
319 goto out;
320 lruvec = mem_cgroup_zone_lruvec(page_zone(page), page_memcg(page));
321 atomic_long_inc(&lruvec->inactive_age);
322out:
323 unlock_page_memcg(page);
324}
325
326/*
327 * Shadow entries reflect the share of the working set that does not
328 * fit into memory, so their number depends on the access pattern of
329 * the workload. In most cases, they will refault or get reclaimed
330 * along with the inode, but a (malicious) workload that streams
331 * through files with a total size several times that of available
332 * memory, while preventing the inodes from being reclaimed, can
333 * create excessive amounts of shadow nodes. To keep a lid on this,
334 * track shadow nodes and reclaim them when they grow way past the
335 * point where they would still be useful.
336 */
337
338struct list_lru workingset_shadow_nodes;
339
340static unsigned long count_shadow_nodes(struct shrinker *shrinker,
341 struct shrink_control *sc)
342{
343 unsigned long shadow_nodes;
344 unsigned long max_nodes;
345 unsigned long pages;
346
347 /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
348 local_irq_disable();
349 shadow_nodes = list_lru_shrink_count(&workingset_shadow_nodes, sc);
350 local_irq_enable();
351
352 if (memcg_kmem_enabled())
353 pages = mem_cgroup_node_nr_lru_pages(sc->memcg, sc->nid,
354 LRU_ALL_FILE);
355 else
356 pages = node_page_state(sc->nid, NR_ACTIVE_FILE) +
357 node_page_state(sc->nid, NR_INACTIVE_FILE);
358
359 /*
360 * Active cache pages are limited to 50% of memory, and shadow
361 * entries that represent a refault distance bigger than that
362 * do not have any effect. Limit the number of shadow nodes
363 * such that shadow entries do not exceed the number of active
364 * cache pages, assuming a worst-case node population density
365 * of 1/8th on average.
366 *
367 * On 64-bit with 7 radix_tree_nodes per page and 64 slots
368 * each, this will reclaim shadow entries when they consume
369 * ~2% of available memory:
370 *
371 * PAGE_SIZE / radix_tree_nodes / node_entries / PAGE_SIZE
372 */
373 max_nodes = pages >> (1 + RADIX_TREE_MAP_SHIFT - 3);
374
375 if (shadow_nodes <= max_nodes)
376 return 0;
377
378 return shadow_nodes - max_nodes;
379}
380
381static enum lru_status shadow_lru_isolate(struct list_head *item,
382 struct list_lru_one *lru,
383 spinlock_t *lru_lock,
384 void *arg)
385{
386 struct address_space *mapping;
387 struct radix_tree_node *node;
388 unsigned int i;
389 int ret;
390
391 /*
392 * Page cache insertions and deletions synchroneously maintain
393 * the shadow node LRU under the mapping->tree_lock and the
394 * lru_lock. Because the page cache tree is emptied before
395 * the inode can be destroyed, holding the lru_lock pins any
396 * address_space that has radix tree nodes on the LRU.
397 *
398 * We can then safely transition to the mapping->tree_lock to
399 * pin only the address_space of the particular node we want
400 * to reclaim, take the node off-LRU, and drop the lru_lock.
401 */
402
403 node = container_of(item, struct radix_tree_node, private_list);
404 mapping = node->private_data;
405
406 /* Coming from the list, invert the lock order */
407 if (!spin_trylock(&mapping->tree_lock)) {
408 spin_unlock(lru_lock);
409 ret = LRU_RETRY;
410 goto out;
411 }
412
413 list_lru_isolate(lru, item);
414 spin_unlock(lru_lock);
415
416 /*
417 * The nodes should only contain one or more shadow entries,
418 * no pages, so we expect to be able to remove them all and
419 * delete and free the empty node afterwards.
420 */
421
422 BUG_ON(!node->count);
423 BUG_ON(node->count & RADIX_TREE_COUNT_MASK);
424
425 for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
426 if (node->slots[i]) {
427 BUG_ON(!radix_tree_exceptional_entry(node->slots[i]));
428 node->slots[i] = NULL;
429 BUG_ON(node->count < (1U << RADIX_TREE_COUNT_SHIFT));
430 node->count -= 1U << RADIX_TREE_COUNT_SHIFT;
431 BUG_ON(!mapping->nrexceptional);
432 mapping->nrexceptional--;
433 }
434 }
435 BUG_ON(node->count);
436 inc_zone_state(page_zone(virt_to_page(node)), WORKINGSET_NODERECLAIM);
437 if (!__radix_tree_delete_node(&mapping->page_tree, node))
438 BUG();
439
440 spin_unlock(&mapping->tree_lock);
441 ret = LRU_REMOVED_RETRY;
442out:
443 local_irq_enable();
444 cond_resched();
445 local_irq_disable();
446 spin_lock(lru_lock);
447 return ret;
448}
449
450static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
451 struct shrink_control *sc)
452{
453 unsigned long ret;
454
455 /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
456 local_irq_disable();
457 ret = list_lru_shrink_walk(&workingset_shadow_nodes, sc,
458 shadow_lru_isolate, NULL);
459 local_irq_enable();
460 return ret;
461}
462
463static struct shrinker workingset_shadow_shrinker = {
464 .count_objects = count_shadow_nodes,
465 .scan_objects = scan_shadow_nodes,
466 .seeks = DEFAULT_SEEKS,
467 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
468};
469
470/*
471 * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
472 * mapping->tree_lock.
473 */
474static struct lock_class_key shadow_nodes_key;
475
476static int __init workingset_init(void)
477{
478 unsigned int timestamp_bits;
479 unsigned int max_order;
480 int ret;
481
482 BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
483 /*
484 * Calculate the eviction bucket size to cover the longest
485 * actionable refault distance, which is currently half of
486 * memory (totalram_pages/2). However, memory hotplug may add
487 * some more pages at runtime, so keep working with up to
488 * double the initial memory by using totalram_pages as-is.
489 */
490 timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
491 max_order = fls_long(totalram_pages - 1);
492 if (max_order > timestamp_bits)
493 bucket_order = max_order - timestamp_bits;
494 printk("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
495 timestamp_bits, max_order, bucket_order);
496
497 ret = list_lru_init_key(&workingset_shadow_nodes, &shadow_nodes_key);
498 if (ret)
499 goto err;
500 ret = register_shrinker(&workingset_shadow_shrinker);
501 if (ret)
502 goto err_list_lru;
503 return 0;
504err_list_lru:
505 list_lru_destroy(&workingset_shadow_nodes);
506err:
507 return ret;
508}
509module_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 EVICTION_SHIFT ((BITS_PER_LONG - BITS_PER_XA_VALUE) + \
172 1 + NODES_SHIFT + MEM_CGROUP_ID_SHIFT)
173#define EVICTION_MASK (~0UL >> EVICTION_SHIFT)
174
175/*
176 * Eviction timestamps need to be able to cover the full range of
177 * actionable refaults. However, bits are tight in the xarray
178 * entry, and after storing the identifier for the lruvec there might
179 * not be enough left to represent every single actionable refault. In
180 * that case, we have to sacrifice granularity for distance, and group
181 * evictions into coarser buckets by shaving off lower timestamp bits.
182 */
183static unsigned int bucket_order __read_mostly;
184
185static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction,
186 bool workingset)
187{
188 eviction >>= bucket_order;
189 eviction &= EVICTION_MASK;
190 eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
191 eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
192 eviction = (eviction << 1) | workingset;
193
194 return xa_mk_value(eviction);
195}
196
197static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
198 unsigned long *evictionp, bool *workingsetp)
199{
200 unsigned long entry = xa_to_value(shadow);
201 int memcgid, nid;
202 bool workingset;
203
204 workingset = entry & 1;
205 entry >>= 1;
206 nid = entry & ((1UL << NODES_SHIFT) - 1);
207 entry >>= NODES_SHIFT;
208 memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
209 entry >>= MEM_CGROUP_ID_SHIFT;
210
211 *memcgidp = memcgid;
212 *pgdat = NODE_DATA(nid);
213 *evictionp = entry << bucket_order;
214 *workingsetp = workingset;
215}
216
217/**
218 * workingset_age_nonresident - age non-resident entries as LRU ages
219 * @memcg: the lruvec that was aged
220 * @nr_pages: the number of pages to count
221 *
222 * As in-memory pages are aged, non-resident pages need to be aged as
223 * well, in order for the refault distances later on to be comparable
224 * to the in-memory dimensions. This function allows reclaim and LRU
225 * operations to drive the non-resident aging along in parallel.
226 */
227void workingset_age_nonresident(struct lruvec *lruvec, unsigned long nr_pages)
228{
229 /*
230 * Reclaiming a cgroup means reclaiming all its children in a
231 * round-robin fashion. That means that each cgroup has an LRU
232 * order that is composed of the LRU orders of its child
233 * cgroups; and every page has an LRU position not just in the
234 * cgroup that owns it, but in all of that group's ancestors.
235 *
236 * So when the physical inactive list of a leaf cgroup ages,
237 * the virtual inactive lists of all its parents, including
238 * the root cgroup's, age as well.
239 */
240 do {
241 atomic_long_add(nr_pages, &lruvec->nonresident_age);
242 } while ((lruvec = parent_lruvec(lruvec)));
243}
244
245/**
246 * workingset_eviction - note the eviction of a page from memory
247 * @target_memcg: the cgroup that is causing the reclaim
248 * @page: the page being evicted
249 *
250 * Returns a shadow entry to be stored in @page->mapping->i_pages in place
251 * of the evicted @page so that a later refault can be detected.
252 */
253void *workingset_eviction(struct page *page, struct mem_cgroup *target_memcg)
254{
255 struct pglist_data *pgdat = page_pgdat(page);
256 unsigned long eviction;
257 struct lruvec *lruvec;
258 int memcgid;
259
260 /* Page is fully exclusive and pins page->mem_cgroup */
261 VM_BUG_ON_PAGE(PageLRU(page), page);
262 VM_BUG_ON_PAGE(page_count(page), page);
263 VM_BUG_ON_PAGE(!PageLocked(page), page);
264
265 lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
266 workingset_age_nonresident(lruvec, thp_nr_pages(page));
267 /* XXX: target_memcg can be NULL, go through lruvec */
268 memcgid = mem_cgroup_id(lruvec_memcg(lruvec));
269 eviction = atomic_long_read(&lruvec->nonresident_age);
270 return pack_shadow(memcgid, pgdat, eviction, PageWorkingset(page));
271}
272
273/**
274 * workingset_refault - evaluate the refault of a previously evicted page
275 * @page: the freshly allocated replacement page
276 * @shadow: shadow entry of the evicted page
277 *
278 * Calculates and evaluates the refault distance of the previously
279 * evicted page in the context of the node and the memcg whose memory
280 * pressure caused the eviction.
281 */
282void workingset_refault(struct page *page, void *shadow)
283{
284 bool file = page_is_file_lru(page);
285 struct mem_cgroup *eviction_memcg;
286 struct lruvec *eviction_lruvec;
287 unsigned long refault_distance;
288 unsigned long workingset_size;
289 struct pglist_data *pgdat;
290 struct mem_cgroup *memcg;
291 unsigned long eviction;
292 struct lruvec *lruvec;
293 unsigned long refault;
294 bool workingset;
295 int memcgid;
296
297 unpack_shadow(shadow, &memcgid, &pgdat, &eviction, &workingset);
298
299 rcu_read_lock();
300 /*
301 * Look up the memcg associated with the stored ID. It might
302 * have been deleted since the page's eviction.
303 *
304 * Note that in rare events the ID could have been recycled
305 * for a new cgroup that refaults a shared page. This is
306 * impossible to tell from the available data. However, this
307 * should be a rare and limited disturbance, and activations
308 * are always speculative anyway. Ultimately, it's the aging
309 * algorithm's job to shake out the minimum access frequency
310 * for the active cache.
311 *
312 * XXX: On !CONFIG_MEMCG, this will always return NULL; it
313 * would be better if the root_mem_cgroup existed in all
314 * configurations instead.
315 */
316 eviction_memcg = mem_cgroup_from_id(memcgid);
317 if (!mem_cgroup_disabled() && !eviction_memcg)
318 goto out;
319 eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
320 refault = atomic_long_read(&eviction_lruvec->nonresident_age);
321
322 /*
323 * Calculate the refault distance
324 *
325 * The unsigned subtraction here gives an accurate distance
326 * across nonresident_age overflows in most cases. There is a
327 * special case: usually, shadow entries have a short lifetime
328 * and are either refaulted or reclaimed along with the inode
329 * before they get too old. But it is not impossible for the
330 * nonresident_age to lap a shadow entry in the field, which
331 * can then result in a false small refault distance, leading
332 * to a false activation should this old entry actually
333 * refault again. However, earlier kernels used to deactivate
334 * unconditionally with *every* reclaim invocation for the
335 * longest time, so the occasional inappropriate activation
336 * leading to pressure on the active list is not a problem.
337 */
338 refault_distance = (refault - eviction) & EVICTION_MASK;
339
340 /*
341 * The activation decision for this page is made at the level
342 * where the eviction occurred, as that is where the LRU order
343 * during page reclaim is being determined.
344 *
345 * However, the cgroup that will own the page is the one that
346 * is actually experiencing the refault event.
347 */
348 memcg = page_memcg(page);
349 lruvec = mem_cgroup_lruvec(memcg, pgdat);
350
351 inc_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file);
352
353 /*
354 * Compare the distance to the existing workingset size. We
355 * don't activate pages that couldn't stay resident even if
356 * all the memory was available to the workingset. Whether
357 * workingset competition needs to consider anon or not depends
358 * on having swap.
359 */
360 workingset_size = lruvec_page_state(eviction_lruvec, NR_ACTIVE_FILE);
361 if (!file) {
362 workingset_size += lruvec_page_state(eviction_lruvec,
363 NR_INACTIVE_FILE);
364 }
365 if (mem_cgroup_get_nr_swap_pages(memcg) > 0) {
366 workingset_size += lruvec_page_state(eviction_lruvec,
367 NR_ACTIVE_ANON);
368 if (file) {
369 workingset_size += lruvec_page_state(eviction_lruvec,
370 NR_INACTIVE_ANON);
371 }
372 }
373 if (refault_distance > workingset_size)
374 goto out;
375
376 SetPageActive(page);
377 workingset_age_nonresident(lruvec, thp_nr_pages(page));
378 inc_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file);
379
380 /* Page was active prior to eviction */
381 if (workingset) {
382 SetPageWorkingset(page);
383 /* XXX: Move to lru_cache_add() when it supports new vs putback */
384 spin_lock_irq(&page_pgdat(page)->lru_lock);
385 lru_note_cost_page(page);
386 spin_unlock_irq(&page_pgdat(page)->lru_lock);
387 inc_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file);
388 }
389out:
390 rcu_read_unlock();
391}
392
393/**
394 * workingset_activation - note a page activation
395 * @page: page that is being activated
396 */
397void workingset_activation(struct page *page)
398{
399 struct mem_cgroup *memcg;
400 struct lruvec *lruvec;
401
402 rcu_read_lock();
403 /*
404 * Filter non-memcg pages here, e.g. unmap can call
405 * mark_page_accessed() on VDSO pages.
406 *
407 * XXX: See workingset_refault() - this should return
408 * root_mem_cgroup even for !CONFIG_MEMCG.
409 */
410 memcg = page_memcg_rcu(page);
411 if (!mem_cgroup_disabled() && !memcg)
412 goto out;
413 lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
414 workingset_age_nonresident(lruvec, thp_nr_pages(page));
415out:
416 rcu_read_unlock();
417}
418
419/*
420 * Shadow entries reflect the share of the working set that does not
421 * fit into memory, so their number depends on the access pattern of
422 * the workload. In most cases, they will refault or get reclaimed
423 * along with the inode, but a (malicious) workload that streams
424 * through files with a total size several times that of available
425 * memory, while preventing the inodes from being reclaimed, can
426 * create excessive amounts of shadow nodes. To keep a lid on this,
427 * track shadow nodes and reclaim them when they grow way past the
428 * point where they would still be useful.
429 */
430
431static struct list_lru shadow_nodes;
432
433void workingset_update_node(struct xa_node *node)
434{
435 /*
436 * Track non-empty nodes that contain only shadow entries;
437 * unlink those that contain pages or are being freed.
438 *
439 * Avoid acquiring the list_lru lock when the nodes are
440 * already where they should be. The list_empty() test is safe
441 * as node->private_list is protected by the i_pages lock.
442 */
443 VM_WARN_ON_ONCE(!irqs_disabled()); /* For __inc_lruvec_page_state */
444
445 if (node->count && node->count == node->nr_values) {
446 if (list_empty(&node->private_list)) {
447 list_lru_add(&shadow_nodes, &node->private_list);
448 __inc_lruvec_slab_state(node, WORKINGSET_NODES);
449 }
450 } else {
451 if (!list_empty(&node->private_list)) {
452 list_lru_del(&shadow_nodes, &node->private_list);
453 __dec_lruvec_slab_state(node, WORKINGSET_NODES);
454 }
455 }
456}
457
458static unsigned long count_shadow_nodes(struct shrinker *shrinker,
459 struct shrink_control *sc)
460{
461 unsigned long max_nodes;
462 unsigned long nodes;
463 unsigned long pages;
464
465 nodes = list_lru_shrink_count(&shadow_nodes, sc);
466
467 /*
468 * Approximate a reasonable limit for the nodes
469 * containing shadow entries. We don't need to keep more
470 * shadow entries than possible pages on the active list,
471 * since refault distances bigger than that are dismissed.
472 *
473 * The size of the active list converges toward 100% of
474 * overall page cache as memory grows, with only a tiny
475 * inactive list. Assume the total cache size for that.
476 *
477 * Nodes might be sparsely populated, with only one shadow
478 * entry in the extreme case. Obviously, we cannot keep one
479 * node for every eligible shadow entry, so compromise on a
480 * worst-case density of 1/8th. Below that, not all eligible
481 * refaults can be detected anymore.
482 *
483 * On 64-bit with 7 xa_nodes per page and 64 slots
484 * each, this will reclaim shadow entries when they consume
485 * ~1.8% of available memory:
486 *
487 * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE
488 */
489#ifdef CONFIG_MEMCG
490 if (sc->memcg) {
491 struct lruvec *lruvec;
492 int i;
493
494 lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
495 for (pages = 0, i = 0; i < NR_LRU_LISTS; i++)
496 pages += lruvec_page_state_local(lruvec,
497 NR_LRU_BASE + i);
498 pages += lruvec_page_state_local(
499 lruvec, NR_SLAB_RECLAIMABLE_B) >> PAGE_SHIFT;
500 pages += lruvec_page_state_local(
501 lruvec, NR_SLAB_UNRECLAIMABLE_B) >> PAGE_SHIFT;
502 } else
503#endif
504 pages = node_present_pages(sc->nid);
505
506 max_nodes = pages >> (XA_CHUNK_SHIFT - 3);
507
508 if (!nodes)
509 return SHRINK_EMPTY;
510
511 if (nodes <= max_nodes)
512 return 0;
513 return nodes - max_nodes;
514}
515
516static enum lru_status shadow_lru_isolate(struct list_head *item,
517 struct list_lru_one *lru,
518 spinlock_t *lru_lock,
519 void *arg) __must_hold(lru_lock)
520{
521 struct xa_node *node = container_of(item, struct xa_node, private_list);
522 XA_STATE(xas, node->array, 0);
523 struct address_space *mapping;
524 int ret;
525
526 /*
527 * Page cache insertions and deletions synchroneously maintain
528 * the shadow node LRU under the i_pages lock and the
529 * lru_lock. Because the page cache tree is emptied before
530 * the inode can be destroyed, holding the lru_lock pins any
531 * address_space that has nodes on the LRU.
532 *
533 * We can then safely transition to the i_pages lock to
534 * pin only the address_space of the particular node we want
535 * to reclaim, take the node off-LRU, and drop the lru_lock.
536 */
537
538 mapping = container_of(node->array, struct address_space, i_pages);
539
540 /* Coming from the list, invert the lock order */
541 if (!xa_trylock(&mapping->i_pages)) {
542 spin_unlock_irq(lru_lock);
543 ret = LRU_RETRY;
544 goto out;
545 }
546
547 list_lru_isolate(lru, item);
548 __dec_lruvec_slab_state(node, WORKINGSET_NODES);
549
550 spin_unlock(lru_lock);
551
552 /*
553 * The nodes should only contain one or more shadow entries,
554 * no pages, so we expect to be able to remove them all and
555 * delete and free the empty node afterwards.
556 */
557 if (WARN_ON_ONCE(!node->nr_values))
558 goto out_invalid;
559 if (WARN_ON_ONCE(node->count != node->nr_values))
560 goto out_invalid;
561 mapping->nrexceptional -= node->nr_values;
562 xas.xa_node = xa_parent_locked(&mapping->i_pages, node);
563 xas.xa_offset = node->offset;
564 xas.xa_shift = node->shift + XA_CHUNK_SHIFT;
565 xas_set_update(&xas, workingset_update_node);
566 /*
567 * We could store a shadow entry here which was the minimum of the
568 * shadow entries we were tracking ...
569 */
570 xas_store(&xas, NULL);
571 __inc_lruvec_slab_state(node, WORKINGSET_NODERECLAIM);
572
573out_invalid:
574 xa_unlock_irq(&mapping->i_pages);
575 ret = LRU_REMOVED_RETRY;
576out:
577 cond_resched();
578 spin_lock_irq(lru_lock);
579 return ret;
580}
581
582static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
583 struct shrink_control *sc)
584{
585 /* list_lru lock nests inside the IRQ-safe i_pages lock */
586 return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate,
587 NULL);
588}
589
590static struct shrinker workingset_shadow_shrinker = {
591 .count_objects = count_shadow_nodes,
592 .scan_objects = scan_shadow_nodes,
593 .seeks = 0, /* ->count reports only fully expendable nodes */
594 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
595};
596
597/*
598 * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
599 * i_pages lock.
600 */
601static struct lock_class_key shadow_nodes_key;
602
603static int __init workingset_init(void)
604{
605 unsigned int timestamp_bits;
606 unsigned int max_order;
607 int ret;
608
609 BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
610 /*
611 * Calculate the eviction bucket size to cover the longest
612 * actionable refault distance, which is currently half of
613 * memory (totalram_pages/2). However, memory hotplug may add
614 * some more pages at runtime, so keep working with up to
615 * double the initial memory by using totalram_pages as-is.
616 */
617 timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
618 max_order = fls_long(totalram_pages() - 1);
619 if (max_order > timestamp_bits)
620 bucket_order = max_order - timestamp_bits;
621 pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
622 timestamp_bits, max_order, bucket_order);
623
624 ret = prealloc_shrinker(&workingset_shadow_shrinker);
625 if (ret)
626 goto err;
627 ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key,
628 &workingset_shadow_shrinker);
629 if (ret)
630 goto err_list_lru;
631 register_shrinker_prepared(&workingset_shadow_shrinker);
632 return 0;
633err_list_lru:
634 free_prealloced_shrinker(&workingset_shadow_shrinker);
635err:
636 return ret;
637}
638module_init(workingset_init);