blob: 40ee02c83978e6a4bd650b2da6d31e5b866c5b80 [file] [log] [blame]
Greg Kroah-Hartmanb2441312017-11-01 15:07:57 +01001// SPDX-License-Identifier: GPL-2.0
Johannes Weinera5289102014-04-03 14:47:51 -07002/*
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>
Hugh Dickins3a4f8a02017-02-24 14:59:36 -080010#include <linux/shmem_fs.h>
Johannes Weinera5289102014-04-03 14:47:51 -070011#include <linux/pagemap.h>
12#include <linux/atomic.h>
13#include <linux/module.h>
14#include <linux/swap.h>
Johannes Weiner14b46872016-12-12 16:43:52 -080015#include <linux/dax.h>
Johannes Weinera5289102014-04-03 14:47:51 -070016#include <linux/fs.h>
17#include <linux/mm.h>
18
19/*
20 * Double CLOCK lists
21 *
Mel Gorman1e6b10852016-07-28 15:46:08 -070022 * Per node, two clock lists are maintained for file pages: the
Johannes Weinera5289102014-04-03 14:47:51 -070023 * 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 *
Mel Gorman1e6b10852016-07-28 15:46:08 -0700147 * For each node's file LRU lists, a counter for inactive evictions
148 * and activations is maintained (node->inactive_age).
Johannes Weinera5289102014-04-03 14:47:51 -0700149 *
150 * On eviction, a snapshot of this counter (along with some bits to
Mel Gorman1e6b10852016-07-28 15:46:08 -0700151 * identify the node) is stored in the now empty page cache radix tree
Johannes Weinera5289102014-04-03 14:47:51 -0700152 * 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
Johannes Weiner689c94f2016-03-15 14:57:07 -0700158#define EVICTION_SHIFT (RADIX_TREE_EXCEPTIONAL_ENTRY + \
Mel Gorman1e6b10852016-07-28 15:46:08 -0700159 NODES_SHIFT + \
Johannes Weiner23047a92016-03-15 14:57:16 -0700160 MEM_CGROUP_ID_SHIFT)
Johannes Weiner689c94f2016-03-15 14:57:07 -0700161#define EVICTION_MASK (~0UL >> EVICTION_SHIFT)
162
Johannes Weiner612e4492016-03-15 14:57:13 -0700163/*
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
Mel Gorman1e6b10852016-07-28 15:46:08 -0700173static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction)
Johannes Weinera5289102014-04-03 14:47:51 -0700174{
Johannes Weiner612e4492016-03-15 14:57:13 -0700175 eviction >>= bucket_order;
Johannes Weiner23047a92016-03-15 14:57:16 -0700176 eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
Mel Gorman1e6b10852016-07-28 15:46:08 -0700177 eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
Johannes Weinera5289102014-04-03 14:47:51 -0700178 eviction = (eviction << RADIX_TREE_EXCEPTIONAL_SHIFT);
179
180 return (void *)(eviction | RADIX_TREE_EXCEPTIONAL_ENTRY);
181}
182
Mel Gorman1e6b10852016-07-28 15:46:08 -0700183static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
Johannes Weiner162453b2016-03-15 14:57:10 -0700184 unsigned long *evictionp)
Johannes Weinera5289102014-04-03 14:47:51 -0700185{
186 unsigned long entry = (unsigned long)shadow;
Mel Gorman1e6b10852016-07-28 15:46:08 -0700187 int memcgid, nid;
Johannes Weinera5289102014-04-03 14:47:51 -0700188
189 entry >>= RADIX_TREE_EXCEPTIONAL_SHIFT;
Johannes Weinera5289102014-04-03 14:47:51 -0700190 nid = entry & ((1UL << NODES_SHIFT) - 1);
191 entry >>= NODES_SHIFT;
Johannes Weiner23047a92016-03-15 14:57:16 -0700192 memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
193 entry >>= MEM_CGROUP_ID_SHIFT;
Johannes Weinera5289102014-04-03 14:47:51 -0700194
Johannes Weiner23047a92016-03-15 14:57:16 -0700195 *memcgidp = memcgid;
Mel Gorman1e6b10852016-07-28 15:46:08 -0700196 *pgdat = NODE_DATA(nid);
Johannes Weiner612e4492016-03-15 14:57:13 -0700197 *evictionp = entry << bucket_order;
Johannes Weinera5289102014-04-03 14:47:51 -0700198}
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 *
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700205 * Returns a shadow entry to be stored in @mapping->i_pages in place
Johannes Weinera5289102014-04-03 14:47:51 -0700206 * of the evicted @page so that a later refault can be detected.
207 */
208void *workingset_eviction(struct address_space *mapping, struct page *page)
209{
Johannes Weiner23047a92016-03-15 14:57:16 -0700210 struct mem_cgroup *memcg = page_memcg(page);
Mel Gorman1e6b10852016-07-28 15:46:08 -0700211 struct pglist_data *pgdat = page_pgdat(page);
Johannes Weiner23047a92016-03-15 14:57:16 -0700212 int memcgid = mem_cgroup_id(memcg);
Johannes Weinera5289102014-04-03 14:47:51 -0700213 unsigned long eviction;
Johannes Weiner23047a92016-03-15 14:57:16 -0700214 struct lruvec *lruvec;
Johannes Weinera5289102014-04-03 14:47:51 -0700215
Johannes Weiner23047a92016-03-15 14:57:16 -0700216 /* 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
Mel Gorman1e6b10852016-07-28 15:46:08 -0700221 lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weiner23047a92016-03-15 14:57:16 -0700222 eviction = atomic_long_inc_return(&lruvec->inactive_age);
Mel Gorman1e6b10852016-07-28 15:46:08 -0700223 return pack_shadow(memcgid, pgdat, eviction);
Johannes Weinera5289102014-04-03 14:47:51 -0700224}
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
Mel Gorman1e6b10852016-07-28 15:46:08 -0700231 * evicted page in the context of the node it was allocated in.
Johannes Weinera5289102014-04-03 14:47:51 -0700232 *
233 * Returns %true if the page should be activated, %false otherwise.
234 */
235bool workingset_refault(void *shadow)
236{
237 unsigned long refault_distance;
Johannes Weiner23047a92016-03-15 14:57:16 -0700238 unsigned long active_file;
239 struct mem_cgroup *memcg;
Johannes Weiner162453b2016-03-15 14:57:10 -0700240 unsigned long eviction;
Johannes Weiner23047a92016-03-15 14:57:16 -0700241 struct lruvec *lruvec;
Johannes Weiner162453b2016-03-15 14:57:10 -0700242 unsigned long refault;
Mel Gorman1e6b10852016-07-28 15:46:08 -0700243 struct pglist_data *pgdat;
Johannes Weiner23047a92016-03-15 14:57:16 -0700244 int memcgid;
Johannes Weinera5289102014-04-03 14:47:51 -0700245
Mel Gorman1e6b10852016-07-28 15:46:08 -0700246 unpack_shadow(shadow, &memcgid, &pgdat, &eviction);
Johannes Weiner162453b2016-03-15 14:57:10 -0700247
Johannes Weiner23047a92016-03-15 14:57:16 -0700248 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 }
Mel Gorman1e6b10852016-07-28 15:46:08 -0700270 lruvec = mem_cgroup_lruvec(pgdat, memcg);
Johannes Weiner23047a92016-03-15 14:57:16 -0700271 refault = atomic_long_read(&lruvec->inactive_age);
Michal Hockofd538802017-02-22 15:45:58 -0800272 active_file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES);
Johannes Weiner162453b2016-03-15 14:57:10 -0700273
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
Johannes Weiner00f3ca22017-07-06 15:40:52 -0700292 inc_lruvec_state(lruvec, WORKINGSET_REFAULT);
Johannes Weinera5289102014-04-03 14:47:51 -0700293
Johannes Weiner23047a92016-03-15 14:57:16 -0700294 if (refault_distance <= active_file) {
Johannes Weiner00f3ca22017-07-06 15:40:52 -0700295 inc_lruvec_state(lruvec, WORKINGSET_ACTIVATE);
Johannes Weiner2a2e4882017-05-03 14:55:03 -0700296 rcu_read_unlock();
Johannes Weinera5289102014-04-03 14:47:51 -0700297 return true;
298 }
Johannes Weiner2a2e4882017-05-03 14:55:03 -0700299 rcu_read_unlock();
Johannes Weinera5289102014-04-03 14:47:51 -0700300 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{
Johannes Weiner55779ec2016-07-28 15:45:10 -0700309 struct mem_cgroup *memcg;
Johannes Weiner23047a92016-03-15 14:57:16 -0700310 struct lruvec *lruvec;
311
Johannes Weiner55779ec2016-07-28 15:45:10 -0700312 rcu_read_lock();
Johannes Weiner23047a92016-03-15 14:57:16 -0700313 /*
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 */
Johannes Weiner55779ec2016-07-28 15:45:10 -0700320 memcg = page_memcg_rcu(page);
321 if (!mem_cgroup_disabled() && !memcg)
Johannes Weiner23047a92016-03-15 14:57:16 -0700322 goto out;
Mel Gormanef8f2322016-07-28 15:46:05 -0700323 lruvec = mem_cgroup_lruvec(page_pgdat(page), memcg);
Johannes Weiner23047a92016-03-15 14:57:16 -0700324 atomic_long_inc(&lruvec->inactive_age);
325out:
Johannes Weiner55779ec2016-07-28 15:45:10 -0700326 rcu_read_unlock();
Johannes Weinera5289102014-04-03 14:47:51 -0700327}
Johannes Weiner449dd692014-04-03 14:47:56 -0700328
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
Johannes Weiner14b46872016-12-12 16:43:52 -0800341static struct list_lru shadow_nodes;
342
Mel Gormanc7df8ad2017-11-15 17:37:41 -0800343void workingset_update_node(struct radix_tree_node *node)
Johannes Weiner14b46872016-12-12 16:43:52 -0800344{
Johannes Weiner14b46872016-12-12 16:43:52 -0800345 /*
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
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700351 * as node->private_list is protected by the i_pages lock.
Johannes Weiner14b46872016-12-12 16:43:52 -0800352 */
353 if (node->count && node->count == node->exceptional) {
Matthew Wilcoxd58275b2017-01-16 17:10:21 -0500354 if (list_empty(&node->private_list))
Johannes Weiner14b46872016-12-12 16:43:52 -0800355 list_lru_add(&shadow_nodes, &node->private_list);
Johannes Weiner14b46872016-12-12 16:43:52 -0800356 } else {
357 if (!list_empty(&node->private_list))
358 list_lru_del(&shadow_nodes, &node->private_list);
359 }
360}
Johannes Weiner449dd692014-04-03 14:47:56 -0700361
362static unsigned long count_shadow_nodes(struct shrinker *shrinker,
363 struct shrink_control *sc)
364{
Johannes Weiner449dd692014-04-03 14:47:56 -0700365 unsigned long max_nodes;
Johannes Weiner14b46872016-12-12 16:43:52 -0800366 unsigned long nodes;
Johannes Weinerb5388992016-12-12 16:43:58 -0800367 unsigned long cache;
Johannes Weiner449dd692014-04-03 14:47:56 -0700368
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700369 /* list_lru lock nests inside the IRQ-safe i_pages lock */
Johannes Weiner449dd692014-04-03 14:47:56 -0700370 local_irq_disable();
Johannes Weiner14b46872016-12-12 16:43:52 -0800371 nodes = list_lru_shrink_count(&shadow_nodes, sc);
Johannes Weiner449dd692014-04-03 14:47:56 -0700372 local_irq_enable();
373
Johannes Weiner449dd692014-04-03 14:47:56 -0700374 /*
Johannes Weinerb5388992016-12-12 16:43:58 -0800375 * 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.
Johannes Weiner449dd692014-04-03 14:47:56 -0700389 *
390 * On 64-bit with 7 radix_tree_nodes per page and 64 slots
391 * each, this will reclaim shadow entries when they consume
Johannes Weinerb5388992016-12-12 16:43:58 -0800392 * ~1.8% of available memory:
Johannes Weiner449dd692014-04-03 14:47:56 -0700393 *
Johannes Weinerb5388992016-12-12 16:43:58 -0800394 * PAGE_SIZE / radix_tree_nodes / node_entries * 8 / PAGE_SIZE
Johannes Weiner449dd692014-04-03 14:47:56 -0700395 */
Johannes Weinerb5388992016-12-12 16:43:58 -0800396 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);
Johannes Weiner449dd692014-04-03 14:47:56 -0700404
Johannes Weiner14b46872016-12-12 16:43:52 -0800405 if (nodes <= max_nodes)
Johannes Weiner449dd692014-04-03 14:47:56 -0700406 return 0;
Johannes Weiner14b46872016-12-12 16:43:52 -0800407 return nodes - max_nodes;
Johannes Weiner449dd692014-04-03 14:47:56 -0700408}
409
410static enum lru_status shadow_lru_isolate(struct list_head *item,
Vladimir Davydov3f97b162015-02-12 14:59:35 -0800411 struct list_lru_one *lru,
Johannes Weiner449dd692014-04-03 14:47:56 -0700412 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
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700422 * the shadow node LRU under the i_pages lock and the
Johannes Weiner449dd692014-04-03 14:47:56 -0700423 * 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 *
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700427 * We can then safely transition to the i_pages lock to
Johannes Weiner449dd692014-04-03 14:47:56 -0700428 * 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);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700433 mapping = container_of(node->root, struct address_space, i_pages);
Johannes Weiner449dd692014-04-03 14:47:56 -0700434
435 /* Coming from the list, invert the lock order */
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700436 if (!xa_trylock(&mapping->i_pages)) {
Johannes Weiner449dd692014-04-03 14:47:56 -0700437 spin_unlock(lru_lock);
438 ret = LRU_RETRY;
439 goto out;
440 }
441
Vladimir Davydov3f97b162015-02-12 14:59:35 -0800442 list_lru_isolate(lru, item);
Johannes Weiner449dd692014-04-03 14:47:56 -0700443 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 */
Johannes Weiner14b46872016-12-12 16:43:52 -0800450 if (WARN_ON_ONCE(!node->exceptional))
Johannes Weinerb9368872016-12-12 16:43:38 -0800451 goto out_invalid;
Johannes Weiner14b46872016-12-12 16:43:52 -0800452 if (WARN_ON_ONCE(node->count != node->exceptional))
Johannes Weinerb9368872016-12-12 16:43:38 -0800453 goto out_invalid;
Johannes Weiner449dd692014-04-03 14:47:56 -0700454 for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
455 if (node->slots[i]) {
Johannes Weinerb9368872016-12-12 16:43:38 -0800456 if (WARN_ON_ONCE(!radix_tree_exceptional_entry(node->slots[i])))
457 goto out_invalid;
Johannes Weiner14b46872016-12-12 16:43:52 -0800458 if (WARN_ON_ONCE(!node->exceptional))
459 goto out_invalid;
Johannes Weinerb9368872016-12-12 16:43:38 -0800460 if (WARN_ON_ONCE(!mapping->nrexceptional))
461 goto out_invalid;
Johannes Weiner449dd692014-04-03 14:47:56 -0700462 node->slots[i] = NULL;
Johannes Weiner14b46872016-12-12 16:43:52 -0800463 node->exceptional--;
464 node->count--;
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800465 mapping->nrexceptional--;
Johannes Weiner449dd692014-04-03 14:47:56 -0700466 }
467 }
Johannes Weiner14b46872016-12-12 16:43:52 -0800468 if (WARN_ON_ONCE(node->exceptional))
Johannes Weinerb9368872016-12-12 16:43:38 -0800469 goto out_invalid;
Johannes Weiner00f3ca22017-07-06 15:40:52 -0700470 inc_lruvec_page_state(virt_to_page(node), WORKINGSET_NODERECLAIM);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700471 __radix_tree_delete_node(&mapping->i_pages, node,
Mel Gormanc7df8ad2017-11-15 17:37:41 -0800472 workingset_lookup_update(mapping));
Johannes Weiner449dd692014-04-03 14:47:56 -0700473
Johannes Weinerb9368872016-12-12 16:43:38 -0800474out_invalid:
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700475 xa_unlock(&mapping->i_pages);
Johannes Weiner449dd692014-04-03 14:47:56 -0700476 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
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700490 /* list_lru lock nests inside the IRQ-safe i_pages lock */
Johannes Weiner449dd692014-04-03 14:47:56 -0700491 local_irq_disable();
Johannes Weiner14b46872016-12-12 16:43:52 -0800492 ret = list_lru_shrink_walk(&shadow_nodes, sc, shadow_lru_isolate, NULL);
Johannes Weiner449dd692014-04-03 14:47:56 -0700493 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,
Vladimir Davydov0a6b76d2016-03-17 14:18:42 -0700501 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
Johannes Weiner449dd692014-04-03 14:47:56 -0700502};
503
504/*
505 * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700506 * i_pages lock.
Johannes Weiner449dd692014-04-03 14:47:56 -0700507 */
508static struct lock_class_key shadow_nodes_key;
509
510static int __init workingset_init(void)
511{
Johannes Weiner612e4492016-03-15 14:57:13 -0700512 unsigned int timestamp_bits;
513 unsigned int max_order;
Johannes Weiner449dd692014-04-03 14:47:56 -0700514 int ret;
515
Johannes Weiner612e4492016-03-15 14:57:13 -0700516 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;
Anton Blanchardd3d36c42016-07-14 12:07:41 -0700528 pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
Johannes Weiner612e4492016-03-15 14:57:13 -0700529 timestamp_bits, max_order, bucket_order);
530
Johannes Weiner0cefabd2017-03-31 15:11:52 -0700531 ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key);
Johannes Weiner449dd692014-04-03 14:47:56 -0700532 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:
Johannes Weiner14b46872016-12-12 16:43:52 -0800539 list_lru_destroy(&shadow_nodes);
Johannes Weiner449dd692014-04-03 14:47:56 -0700540err:
541 return ret;
542}
543module_init(workingset_init);