workingset.c 17 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508
  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 node, 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 node's file LRU lists, a counter for inactive evictions
  143. * and activations is maintained (node->inactive_age).
  144. *
  145. * On eviction, a snapshot of this counter (along with some bits to
  146. * identify the node) 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. #define EVICTION_SHIFT (RADIX_TREE_EXCEPTIONAL_ENTRY + \
  153. NODES_SHIFT + \
  154. MEM_CGROUP_ID_SHIFT)
  155. #define EVICTION_MASK (~0UL >> EVICTION_SHIFT)
  156. /*
  157. * Eviction timestamps need to be able to cover the full range of
  158. * actionable refaults. However, bits are tight in the radix tree
  159. * entry, and after storing the identifier for the lruvec there might
  160. * not be enough left to represent every single actionable refault. In
  161. * that case, we have to sacrifice granularity for distance, and group
  162. * evictions into coarser buckets by shaving off lower timestamp bits.
  163. */
  164. static unsigned int bucket_order __read_mostly;
  165. static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction)
  166. {
  167. eviction >>= bucket_order;
  168. eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
  169. eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
  170. eviction = (eviction << RADIX_TREE_EXCEPTIONAL_SHIFT);
  171. return (void *)(eviction | RADIX_TREE_EXCEPTIONAL_ENTRY);
  172. }
  173. static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
  174. unsigned long *evictionp)
  175. {
  176. unsigned long entry = (unsigned long)shadow;
  177. int memcgid, nid;
  178. entry >>= RADIX_TREE_EXCEPTIONAL_SHIFT;
  179. nid = entry & ((1UL << NODES_SHIFT) - 1);
  180. entry >>= NODES_SHIFT;
  181. memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
  182. entry >>= MEM_CGROUP_ID_SHIFT;
  183. *memcgidp = memcgid;
  184. *pgdat = NODE_DATA(nid);
  185. *evictionp = entry << bucket_order;
  186. }
  187. /**
  188. * workingset_eviction - note the eviction of a page from memory
  189. * @mapping: address space the page was backing
  190. * @page: the page being evicted
  191. *
  192. * Returns a shadow entry to be stored in @mapping->page_tree in place
  193. * of the evicted @page so that a later refault can be detected.
  194. */
  195. void *workingset_eviction(struct address_space *mapping, struct page *page)
  196. {
  197. struct mem_cgroup *memcg = page_memcg(page);
  198. struct pglist_data *pgdat = page_pgdat(page);
  199. int memcgid = mem_cgroup_id(memcg);
  200. unsigned long eviction;
  201. struct lruvec *lruvec;
  202. /* Page is fully exclusive and pins page->mem_cgroup */
  203. VM_BUG_ON_PAGE(PageLRU(page), page);
  204. VM_BUG_ON_PAGE(page_count(page), page);
  205. VM_BUG_ON_PAGE(!PageLocked(page), page);
  206. lruvec = mem_cgroup_lruvec(pgdat, memcg);
  207. eviction = atomic_long_inc_return(&lruvec->inactive_age);
  208. return pack_shadow(memcgid, pgdat, eviction);
  209. }
  210. /**
  211. * workingset_refault - evaluate the refault of a previously evicted page
  212. * @shadow: shadow entry of the evicted page
  213. *
  214. * Calculates and evaluates the refault distance of the previously
  215. * evicted page in the context of the node it was allocated in.
  216. *
  217. * Returns %true if the page should be activated, %false otherwise.
  218. */
  219. bool workingset_refault(void *shadow)
  220. {
  221. unsigned long refault_distance;
  222. unsigned long active_file;
  223. struct mem_cgroup *memcg;
  224. unsigned long eviction;
  225. struct lruvec *lruvec;
  226. unsigned long refault;
  227. struct pglist_data *pgdat;
  228. int memcgid;
  229. unpack_shadow(shadow, &memcgid, &pgdat, &eviction);
  230. rcu_read_lock();
  231. /*
  232. * Look up the memcg associated with the stored ID. It might
  233. * have been deleted since the page's eviction.
  234. *
  235. * Note that in rare events the ID could have been recycled
  236. * for a new cgroup that refaults a shared page. This is
  237. * impossible to tell from the available data. However, this
  238. * should be a rare and limited disturbance, and activations
  239. * are always speculative anyway. Ultimately, it's the aging
  240. * algorithm's job to shake out the minimum access frequency
  241. * for the active cache.
  242. *
  243. * XXX: On !CONFIG_MEMCG, this will always return NULL; it
  244. * would be better if the root_mem_cgroup existed in all
  245. * configurations instead.
  246. */
  247. memcg = mem_cgroup_from_id(memcgid);
  248. if (!mem_cgroup_disabled() && !memcg) {
  249. rcu_read_unlock();
  250. return false;
  251. }
  252. lruvec = mem_cgroup_lruvec(pgdat, memcg);
  253. refault = atomic_long_read(&lruvec->inactive_age);
  254. active_file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES);
  255. rcu_read_unlock();
  256. /*
  257. * The unsigned subtraction here gives an accurate distance
  258. * across inactive_age overflows in most cases.
  259. *
  260. * There is a special case: usually, shadow entries have a
  261. * short lifetime and are either refaulted or reclaimed along
  262. * with the inode before they get too old. But it is not
  263. * impossible for the inactive_age to lap a shadow entry in
  264. * the field, which can then can result in a false small
  265. * refault distance, leading to a false activation should this
  266. * old entry actually refault again. However, earlier kernels
  267. * used to deactivate unconditionally with *every* reclaim
  268. * invocation for the longest time, so the occasional
  269. * inappropriate activation leading to pressure on the active
  270. * list is not a problem.
  271. */
  272. refault_distance = (refault - eviction) & EVICTION_MASK;
  273. inc_node_state(pgdat, WORKINGSET_REFAULT);
  274. if (refault_distance <= active_file) {
  275. inc_node_state(pgdat, WORKINGSET_ACTIVATE);
  276. return true;
  277. }
  278. return false;
  279. }
  280. /**
  281. * workingset_activation - note a page activation
  282. * @page: page that is being activated
  283. */
  284. void workingset_activation(struct page *page)
  285. {
  286. struct mem_cgroup *memcg;
  287. struct lruvec *lruvec;
  288. rcu_read_lock();
  289. /*
  290. * Filter non-memcg pages here, e.g. unmap can call
  291. * mark_page_accessed() on VDSO pages.
  292. *
  293. * XXX: See workingset_refault() - this should return
  294. * root_mem_cgroup even for !CONFIG_MEMCG.
  295. */
  296. memcg = page_memcg_rcu(page);
  297. if (!mem_cgroup_disabled() && !memcg)
  298. goto out;
  299. lruvec = mem_cgroup_lruvec(page_pgdat(page), memcg);
  300. atomic_long_inc(&lruvec->inactive_age);
  301. out:
  302. rcu_read_unlock();
  303. }
  304. /*
  305. * Shadow entries reflect the share of the working set that does not
  306. * fit into memory, so their number depends on the access pattern of
  307. * the workload. In most cases, they will refault or get reclaimed
  308. * along with the inode, but a (malicious) workload that streams
  309. * through files with a total size several times that of available
  310. * memory, while preventing the inodes from being reclaimed, can
  311. * create excessive amounts of shadow nodes. To keep a lid on this,
  312. * track shadow nodes and reclaim them when they grow way past the
  313. * point where they would still be useful.
  314. */
  315. struct list_lru workingset_shadow_nodes;
  316. static unsigned long count_shadow_nodes(struct shrinker *shrinker,
  317. struct shrink_control *sc)
  318. {
  319. unsigned long shadow_nodes;
  320. unsigned long max_nodes;
  321. unsigned long pages;
  322. /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
  323. local_irq_disable();
  324. shadow_nodes = list_lru_shrink_count(&workingset_shadow_nodes, sc);
  325. local_irq_enable();
  326. if (sc->memcg) {
  327. pages = mem_cgroup_node_nr_lru_pages(sc->memcg, sc->nid,
  328. LRU_ALL_FILE);
  329. } else {
  330. pages = node_page_state(NODE_DATA(sc->nid), NR_ACTIVE_FILE) +
  331. node_page_state(NODE_DATA(sc->nid), NR_INACTIVE_FILE);
  332. }
  333. /*
  334. * Active cache pages are limited to 50% of memory, and shadow
  335. * entries that represent a refault distance bigger than that
  336. * do not have any effect. Limit the number of shadow nodes
  337. * such that shadow entries do not exceed the number of active
  338. * cache pages, assuming a worst-case node population density
  339. * of 1/8th on average.
  340. *
  341. * On 64-bit with 7 radix_tree_nodes per page and 64 slots
  342. * each, this will reclaim shadow entries when they consume
  343. * ~2% of available memory:
  344. *
  345. * PAGE_SIZE / radix_tree_nodes / node_entries / PAGE_SIZE
  346. */
  347. max_nodes = pages >> (1 + RADIX_TREE_MAP_SHIFT - 3);
  348. if (shadow_nodes <= max_nodes)
  349. return 0;
  350. return shadow_nodes - max_nodes;
  351. }
  352. static enum lru_status shadow_lru_isolate(struct list_head *item,
  353. struct list_lru_one *lru,
  354. spinlock_t *lru_lock,
  355. void *arg)
  356. {
  357. struct address_space *mapping;
  358. struct radix_tree_node *node;
  359. unsigned int i;
  360. int ret;
  361. /*
  362. * Page cache insertions and deletions synchroneously maintain
  363. * the shadow node LRU under the mapping->tree_lock and the
  364. * lru_lock. Because the page cache tree is emptied before
  365. * the inode can be destroyed, holding the lru_lock pins any
  366. * address_space that has radix tree nodes on the LRU.
  367. *
  368. * We can then safely transition to the mapping->tree_lock to
  369. * pin only the address_space of the particular node we want
  370. * to reclaim, take the node off-LRU, and drop the lru_lock.
  371. */
  372. node = container_of(item, struct radix_tree_node, private_list);
  373. mapping = node->private_data;
  374. /* Coming from the list, invert the lock order */
  375. if (!spin_trylock(&mapping->tree_lock)) {
  376. spin_unlock(lru_lock);
  377. ret = LRU_RETRY;
  378. goto out;
  379. }
  380. list_lru_isolate(lru, item);
  381. spin_unlock(lru_lock);
  382. /*
  383. * The nodes should only contain one or more shadow entries,
  384. * no pages, so we expect to be able to remove them all and
  385. * delete and free the empty node afterwards.
  386. */
  387. BUG_ON(!workingset_node_shadows(node));
  388. BUG_ON(workingset_node_pages(node));
  389. for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
  390. if (node->slots[i]) {
  391. BUG_ON(!radix_tree_exceptional_entry(node->slots[i]));
  392. node->slots[i] = NULL;
  393. workingset_node_shadows_dec(node);
  394. BUG_ON(!mapping->nrexceptional);
  395. mapping->nrexceptional--;
  396. }
  397. }
  398. BUG_ON(workingset_node_shadows(node));
  399. inc_node_state(page_pgdat(virt_to_page(node)), WORKINGSET_NODERECLAIM);
  400. if (!__radix_tree_delete_node(&mapping->page_tree, node))
  401. BUG();
  402. spin_unlock(&mapping->tree_lock);
  403. ret = LRU_REMOVED_RETRY;
  404. out:
  405. local_irq_enable();
  406. cond_resched();
  407. local_irq_disable();
  408. spin_lock(lru_lock);
  409. return ret;
  410. }
  411. static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
  412. struct shrink_control *sc)
  413. {
  414. unsigned long ret;
  415. /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
  416. local_irq_disable();
  417. ret = list_lru_shrink_walk(&workingset_shadow_nodes, sc,
  418. shadow_lru_isolate, NULL);
  419. local_irq_enable();
  420. return ret;
  421. }
  422. static struct shrinker workingset_shadow_shrinker = {
  423. .count_objects = count_shadow_nodes,
  424. .scan_objects = scan_shadow_nodes,
  425. .seeks = DEFAULT_SEEKS,
  426. .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
  427. };
  428. /*
  429. * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
  430. * mapping->tree_lock.
  431. */
  432. static struct lock_class_key shadow_nodes_key;
  433. static int __init workingset_init(void)
  434. {
  435. unsigned int timestamp_bits;
  436. unsigned int max_order;
  437. int ret;
  438. BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
  439. /*
  440. * Calculate the eviction bucket size to cover the longest
  441. * actionable refault distance, which is currently half of
  442. * memory (totalram_pages/2). However, memory hotplug may add
  443. * some more pages at runtime, so keep working with up to
  444. * double the initial memory by using totalram_pages as-is.
  445. */
  446. timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
  447. max_order = fls_long(totalram_pages - 1);
  448. if (max_order > timestamp_bits)
  449. bucket_order = max_order - timestamp_bits;
  450. pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
  451. timestamp_bits, max_order, bucket_order);
  452. ret = __list_lru_init(&workingset_shadow_nodes, true, &shadow_nodes_key);
  453. if (ret)
  454. goto err;
  455. ret = register_shrinker(&workingset_shadow_shrinker);
  456. if (ret)
  457. goto err_list_lru;
  458. return 0;
  459. err_list_lru:
  460. list_lru_destroy(&workingset_shadow_nodes);
  461. err:
  462. return ret;
  463. }
  464. module_init(workingset_init);