4 * Copyright (C) 2013 Red Hat, Inc., Johannes Weiner
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>
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.
27 * fault ------------------------+
29 * +--------------+ | +-------------+
30 * reclaim <- | inactive | <-+-- demotion | active | <--+
31 * +--------------+ +-------------+ |
33 * +-------------- promotion ------------------+
36 * Access frequency and refault distance
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.
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
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:
52 * +-memory available to cache-+
54 * +-inactive------+-active----+
55 * a b | c d e f g h i | J K L M N |
56 * +---------------+-----------+
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.
63 * Approximating inactive page access frequency - Observations:
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
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
78 * 1. The sum of evictions and activations between any two points in
79 * time indicate the minimum number of inactive pages accessed in
82 * 2. Moving one inactive page N page slots towards the tail of the
83 * list requires at least N inactive page accesses.
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.
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.
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
102 * NR_inactive + (R - E)
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:
108 * NR_inactive + (R - E) <= NR_inactive + NR_active
110 * which can be further simplified to
112 * (R - E) <= NR_active
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.
121 * Activating refaulting pages
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.
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
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.
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.
144 * For each zone's file LRU lists, a counter for inactive evictions
145 * and activations is maintained (zone->inactive_age).
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.
151 * On cache misses for which there are shadow entries, an eligible
152 * refault distance will immediately activate the refaulting page.
155 static void *pack_shadow(unsigned long eviction, struct zone *zone)
157 eviction = (eviction << NODES_SHIFT) | zone_to_nid(zone);
158 eviction = (eviction << ZONES_SHIFT) | zone_idx(zone);
159 eviction = (eviction << RADIX_TREE_EXCEPTIONAL_SHIFT);
161 return (void *)(eviction | RADIX_TREE_EXCEPTIONAL_ENTRY);
164 static void unpack_shadow(void *shadow,
166 unsigned long *distance)
168 unsigned long entry = (unsigned long)shadow;
169 unsigned long eviction;
170 unsigned long refault;
174 entry >>= RADIX_TREE_EXCEPTIONAL_SHIFT;
175 zid = entry & ((1UL << ZONES_SHIFT) - 1);
176 entry >>= ZONES_SHIFT;
177 nid = entry & ((1UL << NODES_SHIFT) - 1);
178 entry >>= NODES_SHIFT;
181 *zone = NODE_DATA(nid)->node_zones + zid;
183 refault = atomic_long_read(&(*zone)->inactive_age);
184 mask = ~0UL >> (NODES_SHIFT + ZONES_SHIFT +
185 RADIX_TREE_EXCEPTIONAL_SHIFT);
187 * The unsigned subtraction here gives an accurate distance
188 * across inactive_age overflows in most cases.
190 * There is a special case: usually, shadow entries have a
191 * short lifetime and are either refaulted or reclaimed along
192 * with the inode before they get too old. But it is not
193 * impossible for the inactive_age to lap a shadow entry in
194 * the field, which can then can result in a false small
195 * refault distance, leading to a false activation should this
196 * old entry actually refault again. However, earlier kernels
197 * used to deactivate unconditionally with *every* reclaim
198 * invocation for the longest time, so the occasional
199 * inappropriate activation leading to pressure on the active
200 * list is not a problem.
202 *distance = (refault - eviction) & mask;
206 * workingset_eviction - note the eviction of a page from memory
207 * @mapping: address space the page was backing
208 * @page: the page being evicted
210 * Returns a shadow entry to be stored in @mapping->page_tree in place
211 * of the evicted @page so that a later refault can be detected.
213 void *workingset_eviction(struct address_space *mapping, struct page *page)
215 struct zone *zone = page_zone(page);
216 unsigned long eviction;
218 eviction = atomic_long_inc_return(&zone->inactive_age);
219 return pack_shadow(eviction, zone);
223 * workingset_refault - evaluate the refault of a previously evicted page
224 * @shadow: shadow entry of the evicted page
226 * Calculates and evaluates the refault distance of the previously
227 * evicted page in the context of the zone it was allocated in.
229 * Returns %true if the page should be activated, %false otherwise.
231 bool workingset_refault(void *shadow)
233 unsigned long refault_distance;
236 unpack_shadow(shadow, &zone, &refault_distance);
237 inc_zone_state(zone, WORKINGSET_REFAULT);
239 if (refault_distance <= zone_page_state(zone, NR_ACTIVE_FILE)) {
240 inc_zone_state(zone, WORKINGSET_ACTIVATE);
247 * workingset_activation - note a page activation
248 * @page: page that is being activated
250 void workingset_activation(struct page *page)
252 atomic_long_inc(&page_zone(page)->inactive_age);
256 * Shadow entries reflect the share of the working set that does not
257 * fit into memory, so their number depends on the access pattern of
258 * the workload. In most cases, they will refault or get reclaimed
259 * along with the inode, but a (malicious) workload that streams
260 * through files with a total size several times that of available
261 * memory, while preventing the inodes from being reclaimed, can
262 * create excessive amounts of shadow nodes. To keep a lid on this,
263 * track shadow nodes and reclaim them when they grow way past the
264 * point where they would still be useful.
267 struct list_lru __workingset_shadow_nodes;
268 DEFINE_LOCAL_IRQ_LOCK(workingset_shadow_lock);
270 static unsigned long count_shadow_nodes(struct shrinker *shrinker,
271 struct shrink_control *sc)
273 unsigned long shadow_nodes;
274 unsigned long max_nodes;
277 /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
278 local_lock_irq(workingset_shadow_lock);
279 shadow_nodes = list_lru_shrink_count(&__workingset_shadow_nodes, sc);
280 local_unlock_irq(workingset_shadow_lock);
282 pages = node_present_pages(sc->nid);
284 * Active cache pages are limited to 50% of memory, and shadow
285 * entries that represent a refault distance bigger than that
286 * do not have any effect. Limit the number of shadow nodes
287 * such that shadow entries do not exceed the number of active
288 * cache pages, assuming a worst-case node population density
289 * of 1/8th on average.
291 * On 64-bit with 7 radix_tree_nodes per page and 64 slots
292 * each, this will reclaim shadow entries when they consume
293 * ~2% of available memory:
295 * PAGE_SIZE / radix_tree_nodes / node_entries / PAGE_SIZE
297 max_nodes = pages >> (1 + RADIX_TREE_MAP_SHIFT - 3);
299 if (shadow_nodes <= max_nodes)
302 return shadow_nodes - max_nodes;
305 static enum lru_status shadow_lru_isolate(struct list_head *item,
306 struct list_lru_one *lru,
307 spinlock_t *lru_lock,
310 struct address_space *mapping;
311 struct radix_tree_node *node;
316 * Page cache insertions and deletions synchroneously maintain
317 * the shadow node LRU under the mapping->tree_lock and the
318 * lru_lock. Because the page cache tree is emptied before
319 * the inode can be destroyed, holding the lru_lock pins any
320 * address_space that has radix tree nodes on the LRU.
322 * We can then safely transition to the mapping->tree_lock to
323 * pin only the address_space of the particular node we want
324 * to reclaim, take the node off-LRU, and drop the lru_lock.
327 node = container_of(item, struct radix_tree_node, private_list);
328 mapping = node->private_data;
330 /* Coming from the list, invert the lock order */
331 if (!spin_trylock(&mapping->tree_lock)) {
332 spin_unlock(lru_lock);
337 list_lru_isolate(lru, item);
338 spin_unlock(lru_lock);
341 * The nodes should only contain one or more shadow entries,
342 * no pages, so we expect to be able to remove them all and
343 * delete and free the empty node afterwards.
346 BUG_ON(!node->count);
347 BUG_ON(node->count & RADIX_TREE_COUNT_MASK);
349 for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
350 if (node->slots[i]) {
351 BUG_ON(!radix_tree_exceptional_entry(node->slots[i]));
352 node->slots[i] = NULL;
353 BUG_ON(node->count < (1U << RADIX_TREE_COUNT_SHIFT));
354 node->count -= 1U << RADIX_TREE_COUNT_SHIFT;
355 BUG_ON(!mapping->nrshadows);
356 mapping->nrshadows--;
360 inc_zone_state(page_zone(virt_to_page(node)), WORKINGSET_NODERECLAIM);
361 if (!__radix_tree_delete_node(&mapping->page_tree, node))
364 spin_unlock(&mapping->tree_lock);
365 ret = LRU_REMOVED_RETRY;
367 local_unlock_irq(workingset_shadow_lock);
369 local_lock_irq(workingset_shadow_lock);
374 static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
375 struct shrink_control *sc)
379 /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
380 local_lock_irq(workingset_shadow_lock);
381 ret = list_lru_shrink_walk(&__workingset_shadow_nodes, sc,
382 shadow_lru_isolate, NULL);
383 local_unlock_irq(workingset_shadow_lock);
387 static struct shrinker workingset_shadow_shrinker = {
388 .count_objects = count_shadow_nodes,
389 .scan_objects = scan_shadow_nodes,
390 .seeks = DEFAULT_SEEKS,
391 .flags = SHRINKER_NUMA_AWARE,
395 * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
396 * mapping->tree_lock.
398 static struct lock_class_key shadow_nodes_key;
400 static int __init workingset_init(void)
404 ret = list_lru_init_key(&__workingset_shadow_nodes, &shadow_nodes_key);
407 ret = register_shrinker(&workingset_shadow_shrinker);
412 list_lru_destroy(&__workingset_shadow_nodes);
416 module_init(workingset_init);