]>
Commit | Line | Data |
---|---|---|
a528910e JW |
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 | ||
689c94f0 JW |
155 | #define EVICTION_SHIFT (RADIX_TREE_EXCEPTIONAL_ENTRY + \ |
156 | ZONES_SHIFT + NODES_SHIFT) | |
157 | #define EVICTION_MASK (~0UL >> EVICTION_SHIFT) | |
158 | ||
612e4493 JW |
159 | /* |
160 | * Eviction timestamps need to be able to cover the full range of | |
161 | * actionable refaults. However, bits are tight in the radix tree | |
162 | * entry, and after storing the identifier for the lruvec there might | |
163 | * not be enough left to represent every single actionable refault. In | |
164 | * that case, we have to sacrifice granularity for distance, and group | |
165 | * evictions into coarser buckets by shaving off lower timestamp bits. | |
166 | */ | |
167 | static unsigned int bucket_order __read_mostly; | |
168 | ||
a528910e JW |
169 | static void *pack_shadow(unsigned long eviction, struct zone *zone) |
170 | { | |
612e4493 | 171 | eviction >>= bucket_order; |
a528910e JW |
172 | eviction = (eviction << NODES_SHIFT) | zone_to_nid(zone); |
173 | eviction = (eviction << ZONES_SHIFT) | zone_idx(zone); | |
174 | eviction = (eviction << RADIX_TREE_EXCEPTIONAL_SHIFT); | |
175 | ||
176 | return (void *)(eviction | RADIX_TREE_EXCEPTIONAL_ENTRY); | |
177 | } | |
178 | ||
162453bf JW |
179 | static void unpack_shadow(void *shadow, struct zone **zonep, |
180 | unsigned long *evictionp) | |
a528910e JW |
181 | { |
182 | unsigned long entry = (unsigned long)shadow; | |
a528910e JW |
183 | int zid, nid; |
184 | ||
185 | entry >>= RADIX_TREE_EXCEPTIONAL_SHIFT; | |
186 | zid = entry & ((1UL << ZONES_SHIFT) - 1); | |
187 | entry >>= ZONES_SHIFT; | |
188 | nid = entry & ((1UL << NODES_SHIFT) - 1); | |
189 | entry >>= NODES_SHIFT; | |
a528910e | 190 | |
162453bf | 191 | *zonep = NODE_DATA(nid)->node_zones + zid; |
612e4493 | 192 | *evictionp = entry << bucket_order; |
a528910e JW |
193 | } |
194 | ||
195 | /** | |
196 | * workingset_eviction - note the eviction of a page from memory | |
197 | * @mapping: address space the page was backing | |
198 | * @page: the page being evicted | |
199 | * | |
200 | * Returns a shadow entry to be stored in @mapping->page_tree in place | |
201 | * of the evicted @page so that a later refault can be detected. | |
202 | */ | |
203 | void *workingset_eviction(struct address_space *mapping, struct page *page) | |
204 | { | |
205 | struct zone *zone = page_zone(page); | |
206 | unsigned long eviction; | |
207 | ||
208 | eviction = atomic_long_inc_return(&zone->inactive_age); | |
209 | return pack_shadow(eviction, zone); | |
210 | } | |
211 | ||
212 | /** | |
213 | * workingset_refault - evaluate the refault of a previously evicted page | |
214 | * @shadow: shadow entry of the evicted page | |
215 | * | |
216 | * Calculates and evaluates the refault distance of the previously | |
217 | * evicted page in the context of the zone it was allocated in. | |
218 | * | |
219 | * Returns %true if the page should be activated, %false otherwise. | |
220 | */ | |
221 | bool workingset_refault(void *shadow) | |
222 | { | |
223 | unsigned long refault_distance; | |
162453bf JW |
224 | unsigned long eviction; |
225 | unsigned long refault; | |
a528910e JW |
226 | struct zone *zone; |
227 | ||
162453bf JW |
228 | unpack_shadow(shadow, &zone, &eviction); |
229 | ||
230 | refault = atomic_long_read(&zone->inactive_age); | |
231 | ||
232 | /* | |
233 | * The unsigned subtraction here gives an accurate distance | |
234 | * across inactive_age overflows in most cases. | |
235 | * | |
236 | * There is a special case: usually, shadow entries have a | |
237 | * short lifetime and are either refaulted or reclaimed along | |
238 | * with the inode before they get too old. But it is not | |
239 | * impossible for the inactive_age to lap a shadow entry in | |
240 | * the field, which can then can result in a false small | |
241 | * refault distance, leading to a false activation should this | |
242 | * old entry actually refault again. However, earlier kernels | |
243 | * used to deactivate unconditionally with *every* reclaim | |
244 | * invocation for the longest time, so the occasional | |
245 | * inappropriate activation leading to pressure on the active | |
246 | * list is not a problem. | |
247 | */ | |
248 | refault_distance = (refault - eviction) & EVICTION_MASK; | |
249 | ||
a528910e JW |
250 | inc_zone_state(zone, WORKINGSET_REFAULT); |
251 | ||
252 | if (refault_distance <= zone_page_state(zone, NR_ACTIVE_FILE)) { | |
253 | inc_zone_state(zone, WORKINGSET_ACTIVATE); | |
254 | return true; | |
255 | } | |
256 | return false; | |
257 | } | |
258 | ||
259 | /** | |
260 | * workingset_activation - note a page activation | |
261 | * @page: page that is being activated | |
262 | */ | |
263 | void workingset_activation(struct page *page) | |
264 | { | |
265 | atomic_long_inc(&page_zone(page)->inactive_age); | |
266 | } | |
449dd698 JW |
267 | |
268 | /* | |
269 | * Shadow entries reflect the share of the working set that does not | |
270 | * fit into memory, so their number depends on the access pattern of | |
271 | * the workload. In most cases, they will refault or get reclaimed | |
272 | * along with the inode, but a (malicious) workload that streams | |
273 | * through files with a total size several times that of available | |
274 | * memory, while preventing the inodes from being reclaimed, can | |
275 | * create excessive amounts of shadow nodes. To keep a lid on this, | |
276 | * track shadow nodes and reclaim them when they grow way past the | |
277 | * point where they would still be useful. | |
278 | */ | |
279 | ||
280 | struct list_lru workingset_shadow_nodes; | |
281 | ||
282 | static unsigned long count_shadow_nodes(struct shrinker *shrinker, | |
283 | struct shrink_control *sc) | |
284 | { | |
285 | unsigned long shadow_nodes; | |
286 | unsigned long max_nodes; | |
287 | unsigned long pages; | |
288 | ||
289 | /* list_lru lock nests inside IRQ-safe mapping->tree_lock */ | |
290 | local_irq_disable(); | |
503c358c | 291 | shadow_nodes = list_lru_shrink_count(&workingset_shadow_nodes, sc); |
449dd698 JW |
292 | local_irq_enable(); |
293 | ||
294 | pages = node_present_pages(sc->nid); | |
295 | /* | |
296 | * Active cache pages are limited to 50% of memory, and shadow | |
297 | * entries that represent a refault distance bigger than that | |
298 | * do not have any effect. Limit the number of shadow nodes | |
299 | * such that shadow entries do not exceed the number of active | |
300 | * cache pages, assuming a worst-case node population density | |
301 | * of 1/8th on average. | |
302 | * | |
303 | * On 64-bit with 7 radix_tree_nodes per page and 64 slots | |
304 | * each, this will reclaim shadow entries when they consume | |
305 | * ~2% of available memory: | |
306 | * | |
307 | * PAGE_SIZE / radix_tree_nodes / node_entries / PAGE_SIZE | |
308 | */ | |
309 | max_nodes = pages >> (1 + RADIX_TREE_MAP_SHIFT - 3); | |
310 | ||
311 | if (shadow_nodes <= max_nodes) | |
312 | return 0; | |
313 | ||
314 | return shadow_nodes - max_nodes; | |
315 | } | |
316 | ||
317 | static enum lru_status shadow_lru_isolate(struct list_head *item, | |
3f97b163 | 318 | struct list_lru_one *lru, |
449dd698 JW |
319 | spinlock_t *lru_lock, |
320 | void *arg) | |
321 | { | |
322 | struct address_space *mapping; | |
323 | struct radix_tree_node *node; | |
324 | unsigned int i; | |
325 | int ret; | |
326 | ||
327 | /* | |
328 | * Page cache insertions and deletions synchroneously maintain | |
329 | * the shadow node LRU under the mapping->tree_lock and the | |
330 | * lru_lock. Because the page cache tree is emptied before | |
331 | * the inode can be destroyed, holding the lru_lock pins any | |
332 | * address_space that has radix tree nodes on the LRU. | |
333 | * | |
334 | * We can then safely transition to the mapping->tree_lock to | |
335 | * pin only the address_space of the particular node we want | |
336 | * to reclaim, take the node off-LRU, and drop the lru_lock. | |
337 | */ | |
338 | ||
339 | node = container_of(item, struct radix_tree_node, private_list); | |
340 | mapping = node->private_data; | |
341 | ||
342 | /* Coming from the list, invert the lock order */ | |
343 | if (!spin_trylock(&mapping->tree_lock)) { | |
344 | spin_unlock(lru_lock); | |
345 | ret = LRU_RETRY; | |
346 | goto out; | |
347 | } | |
348 | ||
3f97b163 | 349 | list_lru_isolate(lru, item); |
449dd698 JW |
350 | spin_unlock(lru_lock); |
351 | ||
352 | /* | |
353 | * The nodes should only contain one or more shadow entries, | |
354 | * no pages, so we expect to be able to remove them all and | |
355 | * delete and free the empty node afterwards. | |
356 | */ | |
357 | ||
358 | BUG_ON(!node->count); | |
359 | BUG_ON(node->count & RADIX_TREE_COUNT_MASK); | |
360 | ||
361 | for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) { | |
362 | if (node->slots[i]) { | |
363 | BUG_ON(!radix_tree_exceptional_entry(node->slots[i])); | |
364 | node->slots[i] = NULL; | |
365 | BUG_ON(node->count < (1U << RADIX_TREE_COUNT_SHIFT)); | |
366 | node->count -= 1U << RADIX_TREE_COUNT_SHIFT; | |
f9fe48be RZ |
367 | BUG_ON(!mapping->nrexceptional); |
368 | mapping->nrexceptional--; | |
449dd698 JW |
369 | } |
370 | } | |
371 | BUG_ON(node->count); | |
372 | inc_zone_state(page_zone(virt_to_page(node)), WORKINGSET_NODERECLAIM); | |
373 | if (!__radix_tree_delete_node(&mapping->page_tree, node)) | |
374 | BUG(); | |
375 | ||
376 | spin_unlock(&mapping->tree_lock); | |
377 | ret = LRU_REMOVED_RETRY; | |
378 | out: | |
379 | local_irq_enable(); | |
380 | cond_resched(); | |
381 | local_irq_disable(); | |
382 | spin_lock(lru_lock); | |
383 | return ret; | |
384 | } | |
385 | ||
386 | static unsigned long scan_shadow_nodes(struct shrinker *shrinker, | |
387 | struct shrink_control *sc) | |
388 | { | |
389 | unsigned long ret; | |
390 | ||
391 | /* list_lru lock nests inside IRQ-safe mapping->tree_lock */ | |
392 | local_irq_disable(); | |
503c358c VD |
393 | ret = list_lru_shrink_walk(&workingset_shadow_nodes, sc, |
394 | shadow_lru_isolate, NULL); | |
449dd698 JW |
395 | local_irq_enable(); |
396 | return ret; | |
397 | } | |
398 | ||
399 | static struct shrinker workingset_shadow_shrinker = { | |
400 | .count_objects = count_shadow_nodes, | |
401 | .scan_objects = scan_shadow_nodes, | |
402 | .seeks = DEFAULT_SEEKS, | |
403 | .flags = SHRINKER_NUMA_AWARE, | |
404 | }; | |
405 | ||
406 | /* | |
407 | * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe | |
408 | * mapping->tree_lock. | |
409 | */ | |
410 | static struct lock_class_key shadow_nodes_key; | |
411 | ||
412 | static int __init workingset_init(void) | |
413 | { | |
612e4493 JW |
414 | unsigned int timestamp_bits; |
415 | unsigned int max_order; | |
449dd698 JW |
416 | int ret; |
417 | ||
612e4493 JW |
418 | BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT); |
419 | /* | |
420 | * Calculate the eviction bucket size to cover the longest | |
421 | * actionable refault distance, which is currently half of | |
422 | * memory (totalram_pages/2). However, memory hotplug may add | |
423 | * some more pages at runtime, so keep working with up to | |
424 | * double the initial memory by using totalram_pages as-is. | |
425 | */ | |
426 | timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT; | |
427 | max_order = fls_long(totalram_pages - 1); | |
428 | if (max_order > timestamp_bits) | |
429 | bucket_order = max_order - timestamp_bits; | |
430 | printk("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n", | |
431 | timestamp_bits, max_order, bucket_order); | |
432 | ||
449dd698 JW |
433 | ret = list_lru_init_key(&workingset_shadow_nodes, &shadow_nodes_key); |
434 | if (ret) | |
435 | goto err; | |
436 | ret = register_shrinker(&workingset_shadow_shrinker); | |
437 | if (ret) | |
438 | goto err_list_lru; | |
439 | return 0; | |
440 | err_list_lru: | |
441 | list_lru_destroy(&workingset_shadow_nodes); | |
442 | err: | |
443 | return ret; | |
444 | } | |
445 | module_init(workingset_init); |