workingset.c 14 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413
  1. /*
  2. * Workingset detection
  3. *
  4. * Copyright (C) 2013 Red Hat, Inc., Johannes Weiner
  5. */
  6. #include <linux/memcontrol.h>
  7. #include <linux/writeback.h>
  8. #include <linux/pagemap.h>
  9. #include <linux/atomic.h>
  10. #include <linux/module.h>
  11. #include <linux/swap.h>
  12. #include <linux/fs.h>
  13. #include <linux/mm.h>
  14. /*
  15. * Double CLOCK lists
  16. *
  17. * Per zone, two clock lists are maintained for file pages: the
  18. * inactive and the active list. Freshly faulted pages start out at
  19. * the head of the inactive list and page reclaim scans pages from the
  20. * tail. Pages that are accessed multiple times on the inactive list
  21. * are promoted to the active list, to protect them from reclaim,
  22. * whereas active pages are demoted to the inactive list when the
  23. * active list grows too big.
  24. *
  25. * fault ------------------------+
  26. * |
  27. * +--------------+ | +-------------+
  28. * reclaim <- | inactive | <-+-- demotion | active | <--+
  29. * +--------------+ +-------------+ |
  30. * | |
  31. * +-------------- promotion ------------------+
  32. *
  33. *
  34. * Access frequency and refault distance
  35. *
  36. * A workload is thrashing when its pages are frequently used but they
  37. * are evicted from the inactive list every time before another access
  38. * would have promoted them to the active list.
  39. *
  40. * In cases where the average access distance between thrashing pages
  41. * is bigger than the size of memory there is nothing that can be
  42. * done - the thrashing set could never fit into memory under any
  43. * circumstance.
  44. *
  45. * However, the average access distance could be bigger than the
  46. * inactive list, yet smaller than the size of memory. In this case,
  47. * the set could fit into memory if it weren't for the currently
  48. * active pages - which may be used more, hopefully less frequently:
  49. *
  50. * +-memory available to cache-+
  51. * | |
  52. * +-inactive------+-active----+
  53. * a b | c d e f g h i | J K L M N |
  54. * +---------------+-----------+
  55. *
  56. * It is prohibitively expensive to accurately track access frequency
  57. * of pages. But a reasonable approximation can be made to measure
  58. * thrashing on the inactive list, after which refaulting pages can be
  59. * activated optimistically to compete with the existing active pages.
  60. *
  61. * Approximating inactive page access frequency - Observations:
  62. *
  63. * 1. When a page is accessed for the first time, it is added to the
  64. * head of the inactive list, slides every existing inactive page
  65. * towards the tail by one slot, and pushes the current tail page
  66. * out of memory.
  67. *
  68. * 2. When a page is accessed for the second time, it is promoted to
  69. * the active list, shrinking the inactive list by one slot. This
  70. * also slides all inactive pages that were faulted into the cache
  71. * more recently than the activated page towards the tail of the
  72. * inactive list.
  73. *
  74. * Thus:
  75. *
  76. * 1. The sum of evictions and activations between any two points in
  77. * time indicate the minimum number of inactive pages accessed in
  78. * between.
  79. *
  80. * 2. Moving one inactive page N page slots towards the tail of the
  81. * list requires at least N inactive page accesses.
  82. *
  83. * Combining these:
  84. *
  85. * 1. When a page is finally evicted from memory, the number of
  86. * inactive pages accessed while the page was in cache is at least
  87. * the number of page slots on the inactive list.
  88. *
  89. * 2. In addition, measuring the sum of evictions and activations (E)
  90. * at the time of a page's eviction, and comparing it to another
  91. * reading (R) at the time the page faults back into memory tells
  92. * the minimum number of accesses while the page was not cached.
  93. * This is called the refault distance.
  94. *
  95. * Because the first access of the page was the fault and the second
  96. * access the refault, we combine the in-cache distance with the
  97. * out-of-cache distance to get the complete minimum access distance
  98. * of this page:
  99. *
  100. * NR_inactive + (R - E)
  101. *
  102. * And knowing the minimum access distance of a page, we can easily
  103. * tell if the page would be able to stay in cache assuming all page
  104. * slots in the cache were available:
  105. *
  106. * NR_inactive + (R - E) <= NR_inactive + NR_active
  107. *
  108. * which can be further simplified to
  109. *
  110. * (R - E) <= NR_active
  111. *
  112. * Put into words, the refault distance (out-of-cache) can be seen as
  113. * a deficit in inactive list space (in-cache). If the inactive list
  114. * had (R - E) more page slots, the page would not have been evicted
  115. * in between accesses, but activated instead. And on a full system,
  116. * the only thing eating into inactive list space is active pages.
  117. *
  118. *
  119. * Activating refaulting pages
  120. *
  121. * All that is known about the active list is that the pages have been
  122. * accessed more than once in the past. This means that at any given
  123. * time there is actually a good chance that pages on the active list
  124. * are no longer in active use.
  125. *
  126. * So when a refault distance of (R - E) is observed and there are at
  127. * least (R - E) active pages, the refaulting page is activated
  128. * optimistically in the hope that (R - E) active pages are actually
  129. * used less frequently than the refaulting page - or even not used at
  130. * all anymore.
  131. *
  132. * If this is wrong and demotion kicks in, the pages which are truly
  133. * used more frequently will be reactivated while the less frequently
  134. * used once will be evicted from memory.
  135. *
  136. * But if this is right, the stale pages will be pushed out of memory
  137. * and the used pages get to stay in cache.
  138. *
  139. *
  140. * Implementation
  141. *
  142. * For each zone's file LRU lists, a counter for inactive evictions
  143. * and activations is maintained (zone->inactive_age).
  144. *
  145. * On eviction, a snapshot of this counter (along with some bits to
  146. * identify the zone) is stored in the now empty page cache radix tree
  147. * slot of the evicted page. This is called a shadow entry.
  148. *
  149. * On cache misses for which there are shadow entries, an eligible
  150. * refault distance will immediately activate the refaulting page.
  151. */
  152. static void *pack_shadow(unsigned long eviction, struct zone *zone)
  153. {
  154. eviction = (eviction << NODES_SHIFT) | zone_to_nid(zone);
  155. eviction = (eviction << ZONES_SHIFT) | zone_idx(zone);
  156. eviction = (eviction << RADIX_TREE_EXCEPTIONAL_SHIFT);
  157. return (void *)(eviction | RADIX_TREE_EXCEPTIONAL_ENTRY);
  158. }
  159. static void unpack_shadow(void *shadow,
  160. struct zone **zone,
  161. unsigned long *distance)
  162. {
  163. unsigned long entry = (unsigned long)shadow;
  164. unsigned long eviction;
  165. unsigned long refault;
  166. unsigned long mask;
  167. int zid, nid;
  168. entry >>= RADIX_TREE_EXCEPTIONAL_SHIFT;
  169. zid = entry & ((1UL << ZONES_SHIFT) - 1);
  170. entry >>= ZONES_SHIFT;
  171. nid = entry & ((1UL << NODES_SHIFT) - 1);
  172. entry >>= NODES_SHIFT;
  173. eviction = entry;
  174. *zone = NODE_DATA(nid)->node_zones + zid;
  175. refault = atomic_long_read(&(*zone)->inactive_age);
  176. mask = ~0UL >> (NODES_SHIFT + ZONES_SHIFT +
  177. RADIX_TREE_EXCEPTIONAL_SHIFT);
  178. /*
  179. * The unsigned subtraction here gives an accurate distance
  180. * across inactive_age overflows in most cases.
  181. *
  182. * There is a special case: usually, shadow entries have a
  183. * short lifetime and are either refaulted or reclaimed along
  184. * with the inode before they get too old. But it is not
  185. * impossible for the inactive_age to lap a shadow entry in
  186. * the field, which can then can result in a false small
  187. * refault distance, leading to a false activation should this
  188. * old entry actually refault again. However, earlier kernels
  189. * used to deactivate unconditionally with *every* reclaim
  190. * invocation for the longest time, so the occasional
  191. * inappropriate activation leading to pressure on the active
  192. * list is not a problem.
  193. */
  194. *distance = (refault - eviction) & mask;
  195. }
  196. /**
  197. * workingset_eviction - note the eviction of a page from memory
  198. * @mapping: address space the page was backing
  199. * @page: the page being evicted
  200. *
  201. * Returns a shadow entry to be stored in @mapping->page_tree in place
  202. * of the evicted @page so that a later refault can be detected.
  203. */
  204. void *workingset_eviction(struct address_space *mapping, struct page *page)
  205. {
  206. struct zone *zone = page_zone(page);
  207. unsigned long eviction;
  208. eviction = atomic_long_inc_return(&zone->inactive_age);
  209. return pack_shadow(eviction, zone);
  210. }
  211. /**
  212. * workingset_refault - evaluate the refault of a previously evicted page
  213. * @shadow: shadow entry of the evicted page
  214. *
  215. * Calculates and evaluates the refault distance of the previously
  216. * evicted page in the context of the zone it was allocated in.
  217. *
  218. * Returns %true if the page should be activated, %false otherwise.
  219. */
  220. bool workingset_refault(void *shadow)
  221. {
  222. unsigned long refault_distance;
  223. struct zone *zone;
  224. unpack_shadow(shadow, &zone, &refault_distance);
  225. inc_zone_state(zone, WORKINGSET_REFAULT);
  226. if (refault_distance <= zone_page_state(zone, NR_ACTIVE_FILE)) {
  227. inc_zone_state(zone, WORKINGSET_ACTIVATE);
  228. return true;
  229. }
  230. return false;
  231. }
  232. /**
  233. * workingset_activation - note a page activation
  234. * @page: page that is being activated
  235. */
  236. void workingset_activation(struct page *page)
  237. {
  238. atomic_long_inc(&page_zone(page)->inactive_age);
  239. }
  240. /*
  241. * Shadow entries reflect the share of the working set that does not
  242. * fit into memory, so their number depends on the access pattern of
  243. * the workload. In most cases, they will refault or get reclaimed
  244. * along with the inode, but a (malicious) workload that streams
  245. * through files with a total size several times that of available
  246. * memory, while preventing the inodes from being reclaimed, can
  247. * create excessive amounts of shadow nodes. To keep a lid on this,
  248. * track shadow nodes and reclaim them when they grow way past the
  249. * point where they would still be useful.
  250. */
  251. struct list_lru workingset_shadow_nodes;
  252. static unsigned long count_shadow_nodes(struct shrinker *shrinker,
  253. struct shrink_control *sc)
  254. {
  255. unsigned long shadow_nodes;
  256. unsigned long max_nodes;
  257. unsigned long pages;
  258. /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
  259. local_irq_disable();
  260. shadow_nodes = list_lru_shrink_count(&workingset_shadow_nodes, sc);
  261. local_irq_enable();
  262. pages = node_present_pages(sc->nid);
  263. /*
  264. * Active cache pages are limited to 50% of memory, and shadow
  265. * entries that represent a refault distance bigger than that
  266. * do not have any effect. Limit the number of shadow nodes
  267. * such that shadow entries do not exceed the number of active
  268. * cache pages, assuming a worst-case node population density
  269. * of 1/8th on average.
  270. *
  271. * On 64-bit with 7 radix_tree_nodes per page and 64 slots
  272. * each, this will reclaim shadow entries when they consume
  273. * ~2% of available memory:
  274. *
  275. * PAGE_SIZE / radix_tree_nodes / node_entries / PAGE_SIZE
  276. */
  277. max_nodes = pages >> (1 + RADIX_TREE_MAP_SHIFT - 3);
  278. if (shadow_nodes <= max_nodes)
  279. return 0;
  280. return shadow_nodes - max_nodes;
  281. }
  282. static enum lru_status shadow_lru_isolate(struct list_head *item,
  283. struct list_lru_one *lru,
  284. spinlock_t *lru_lock,
  285. void *arg)
  286. {
  287. struct address_space *mapping;
  288. struct radix_tree_node *node;
  289. unsigned int i;
  290. int ret;
  291. /*
  292. * Page cache insertions and deletions synchroneously maintain
  293. * the shadow node LRU under the mapping->tree_lock and the
  294. * lru_lock. Because the page cache tree is emptied before
  295. * the inode can be destroyed, holding the lru_lock pins any
  296. * address_space that has radix tree nodes on the LRU.
  297. *
  298. * We can then safely transition to the mapping->tree_lock to
  299. * pin only the address_space of the particular node we want
  300. * to reclaim, take the node off-LRU, and drop the lru_lock.
  301. */
  302. node = container_of(item, struct radix_tree_node, private_list);
  303. mapping = node->private_data;
  304. /* Coming from the list, invert the lock order */
  305. if (!spin_trylock(&mapping->tree_lock)) {
  306. spin_unlock(lru_lock);
  307. ret = LRU_RETRY;
  308. goto out;
  309. }
  310. list_lru_isolate(lru, item);
  311. spin_unlock(lru_lock);
  312. /*
  313. * The nodes should only contain one or more shadow entries,
  314. * no pages, so we expect to be able to remove them all and
  315. * delete and free the empty node afterwards.
  316. */
  317. BUG_ON(!workingset_node_shadows(node));
  318. BUG_ON(workingset_node_pages(node));
  319. for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
  320. if (node->slots[i]) {
  321. BUG_ON(!radix_tree_exceptional_entry(node->slots[i]));
  322. node->slots[i] = NULL;
  323. workingset_node_shadows_dec(node);
  324. BUG_ON(!mapping->nrshadows);
  325. mapping->nrshadows--;
  326. }
  327. }
  328. BUG_ON(workingset_node_shadows(node));
  329. inc_zone_state(page_zone(virt_to_page(node)), WORKINGSET_NODERECLAIM);
  330. if (!__radix_tree_delete_node(&mapping->page_tree, node))
  331. BUG();
  332. spin_unlock(&mapping->tree_lock);
  333. ret = LRU_REMOVED_RETRY;
  334. out:
  335. local_irq_enable();
  336. cond_resched();
  337. local_irq_disable();
  338. spin_lock(lru_lock);
  339. return ret;
  340. }
  341. static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
  342. struct shrink_control *sc)
  343. {
  344. unsigned long ret;
  345. /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
  346. local_irq_disable();
  347. ret = list_lru_shrink_walk(&workingset_shadow_nodes, sc,
  348. shadow_lru_isolate, NULL);
  349. local_irq_enable();
  350. return ret;
  351. }
  352. static struct shrinker workingset_shadow_shrinker = {
  353. .count_objects = count_shadow_nodes,
  354. .scan_objects = scan_shadow_nodes,
  355. .seeks = DEFAULT_SEEKS,
  356. .flags = SHRINKER_NUMA_AWARE,
  357. };
  358. /*
  359. * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
  360. * mapping->tree_lock.
  361. */
  362. static struct lock_class_key shadow_nodes_key;
  363. static int __init workingset_init(void)
  364. {
  365. int ret;
  366. ret = list_lru_init_key(&workingset_shadow_nodes, &shadow_nodes_key);
  367. if (ret)
  368. goto err;
  369. ret = register_shrinker(&workingset_shadow_shrinker);
  370. if (ret)
  371. goto err_list_lru;
  372. return 0;
  373. err_list_lru:
  374. list_lru_destroy(&workingset_shadow_nodes);
  375. err:
  376. return ret;
  377. }
  378. module_init(workingset_init);