mm/workingset.c

   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/shmem_fs.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/atomic.h>
  12 #include <linux/module.h>
  13 #include <linux/swap.h>
  14 #include <linux/dax.h>
  15 #include <linux/fs.h>
  16 #include <linux/mm.h>
  17
  18 /*
  19  *              Double CLOCK lists
  20  *
  21  * Per node, two clock lists are maintained for file pages: the
  22  * inactive and the active list.  Freshly faulted pages start out at
  23  * the head of the inactive list and page reclaim scans pages from the
  24  * tail.  Pages that are accessed multiple times on the inactive list
  25  * are promoted to the active list, to protect them from reclaim,
  26  * whereas active pages are demoted to the inactive list when the
  27  * active list grows too big.
  28  *
  29  *   fault ------------------------+
  30  *                                 |
  31  *              +--------------+   |            +-------------+
  32  *   reclaim <- |   inactive   | <-+-- demotion |    active   | <--+
  33  *              +--------------+                +-------------+    |
  34  *                     |                                           |
  35  *                     +-------------- promotion ------------------+
  36  *
  37  *
  38  *              Access frequency and refault distance
  39  *
  40  * A workload is thrashing when its pages are frequently used but they
  41  * are evicted from the inactive list every time before another access
  42  * would have promoted them to the active list.
  43  *
  44  * In cases where the average access distance between thrashing pages
  45  * is bigger than the size of memory there is nothing that can be
  46  * done - the thrashing set could never fit into memory under any
  47  * circumstance.
  48  *
  49  * However, the average access distance could be bigger than the
  50  * inactive list, yet smaller than the size of memory.  In this case,
  51  * the set could fit into memory if it weren't for the currently
  52  * active pages - which may be used more, hopefully less frequently:
  53  *
  54  *      +-memory available to cache-+
  55  *      |                           |
  56  *      +-inactive------+-active----+
  57  *  a b | c d e f g h i | J K L M N |
  58  *      +---------------+-----------+
  59  *
  60  * It is prohibitively expensive to accurately track access frequency
  61  * of pages.  But a reasonable approximation can be made to measure
  62  * thrashing on the inactive list, after which refaulting pages can be
  63  * activated optimistically to compete with the existing active pages.
  64  *
  65  * Approximating inactive page access frequency - Observations:
  66  *
  67  * 1. When a page is accessed for the first time, it is added to the
  68  *    head of the inactive list, slides every existing inactive page
  69  *    towards the tail by one slot, and pushes the current tail page
  70  *    out of memory.
  71  *
  72  * 2. When a page is accessed for the second time, it is promoted to
  73  *    the active list, shrinking the inactive list by one slot.  This
  74  *    also slides all inactive pages that were faulted into the cache
  75  *    more recently than the activated page towards the tail of the
  76  *    inactive list.
  77  *
  78  * Thus:
  79  *
  80  * 1. The sum of evictions and activations between any two points in
  81  *    time indicate the minimum number of inactive pages accessed in
  82  *    between.
  83  *
  84  * 2. Moving one inactive page N page slots towards the tail of the
  85  *    list requires at least N inactive page accesses.
  86  *
  87  * Combining these:
  88  *
  89  * 1. When a page is finally evicted from memory, the number of
  90  *    inactive pages accessed while the page was in cache is at least
  91  *    the number of page slots on the inactive list.
  92  *
  93  * 2. In addition, measuring the sum of evictions and activations (E)
  94  *    at the time of a page's eviction, and comparing it to another
  95  *    reading (R) at the time the page faults back into memory tells
  96  *    the minimum number of accesses while the page was not cached.
  97  *    This is called the refault distance.
  98  *
  99  * Because the first access of the page was the fault and the second
 100  * access the refault, we combine the in-cache distance with the
 101  * out-of-cache distance to get the complete minimum access distance
 102  * of this page:
 103  *
 104  *      NR_inactive + (R - E)
 105  *
 106  * And knowing the minimum access distance of a page, we can easily
 107  * tell if the page would be able to stay in cache assuming all page
 108  * slots in the cache were available:
 109  *
 110  *   NR_inactive + (R - E) <= NR_inactive + NR_active
 111  *
 112  * which can be further simplified to
 113  *
 114  *   (R - E) <= NR_active
 115  *
 116  * Put into words, the refault distance (out-of-cache) can be seen as
 117  * a deficit in inactive list space (in-cache).  If the inactive list
 118  * had (R - E) more page slots, the page would not have been evicted
 119  * in between accesses, but activated instead.  And on a full system,
 120  * the only thing eating into inactive list space is active pages.
 121  *
 122  *
 123  *              Activating refaulting pages
 124  *
 125  * All that is known about the active list is that the pages have been
 126  * accessed more than once in the past.  This means that at any given
 127  * time there is actually a good chance that pages on the active list
 128  * are no longer in active use.
 129  *
 130  * So when a refault distance of (R - E) is observed and there are at
 131  * least (R - E) active pages, the refaulting page is activated
 132  * optimistically in the hope that (R - E) active pages are actually
 133  * used less frequently than the refaulting page - or even not used at
 134  * all anymore.
 135  *
 136  * If this is wrong and demotion kicks in, the pages which are truly
 137  * used more frequently will be reactivated while the less frequently
 138  * used once will be evicted from memory.
 139  *
 140  * But if this is right, the stale pages will be pushed out of memory
 141  * and the used pages get to stay in cache.
 142  *
 143  *
 144  *              Implementation
 145  *
 146  * For each node's file LRU lists, a counter for inactive evictions
 147  * and activations is maintained (node->inactive_age).
 148  *
 149  * On eviction, a snapshot of this counter (along with some bits to
 150  * identify the node) is stored in the now empty page cache radix tree
 151  * slot of the evicted page.  This is called a shadow entry.
 152  *
 153  * On cache misses for which there are shadow entries, an eligible
 154  * refault distance will immediately activate the refaulting page.
 155  */
 156
 157 #define EVICTION_SHIFT  (RADIX_TREE_EXCEPTIONAL_ENTRY + \
 158                          NODES_SHIFT +  \
 159                          MEM_CGROUP_ID_SHIFT)
 160 #define EVICTION_MASK   (~0UL >> EVICTION_SHIFT)
 161
 162 /*
 163  * Eviction timestamps need to be able to cover the full range of
 164  * actionable refaults. However, bits are tight in the radix tree
 165  * entry, and after storing the identifier for the lruvec there might
 166  * not be enough left to represent every single actionable refault. In
 167  * that case, we have to sacrifice granularity for distance, and group
 168  * evictions into coarser buckets by shaving off lower timestamp bits.
 169  */
 170 static unsigned int bucket_order __read_mostly;
 171
 172 static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction)
 173 {
 174         eviction >>= bucket_order;
 175         eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
 176         eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
 177         eviction = (eviction << RADIX_TREE_EXCEPTIONAL_SHIFT);
 178
 179         return (void *)(eviction | RADIX_TREE_EXCEPTIONAL_ENTRY);
 180 }
 181
 182 static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
 183                           unsigned long *evictionp)
 184 {
 185         unsigned long entry = (unsigned long)shadow;
 186         int memcgid, nid;
 187
 188         entry >>= RADIX_TREE_EXCEPTIONAL_SHIFT;
 189         nid = entry & ((1UL << NODES_SHIFT) - 1);
 190         entry >>= NODES_SHIFT;
 191         memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
 192         entry >>= MEM_CGROUP_ID_SHIFT;
 193
 194         *memcgidp = memcgid;
 195         *pgdat = NODE_DATA(nid);
 196         *evictionp = entry << bucket_order;
 197 }
 198
 199 /**
 200  * workingset_eviction - note the eviction of a page from memory
 201  * @mapping: address space the page was backing
 202  * @page: the page being evicted
 203  *
 204  * Returns a shadow entry to be stored in @mapping->page_tree in place
 205  * of the evicted @page so that a later refault can be detected.
 206  */
 207 void *workingset_eviction(struct address_space *mapping, struct page *page)
 208 {
 209         struct mem_cgroup *memcg = page_memcg(page);
 210         struct pglist_data *pgdat = page_pgdat(page);
 211         int memcgid = mem_cgroup_id(memcg);
 212         unsigned long eviction;
 213         struct lruvec *lruvec;
 214
 215         /* Page is fully exclusive and pins page->mem_cgroup */
 216         VM_BUG_ON_PAGE(PageLRU(page), page);
 217         VM_BUG_ON_PAGE(page_count(page), page);
 218         VM_BUG_ON_PAGE(!PageLocked(page), page);
 219
 220         lruvec = mem_cgroup_lruvec(pgdat, memcg);
 221         eviction = atomic_long_inc_return(&lruvec->inactive_age);
 222         return pack_shadow(memcgid, pgdat, eviction);
 223 }
 224
 225 /**
 226  * workingset_refault - evaluate the refault of a previously evicted page
 227  * @shadow: shadow entry of the evicted page
 228  *
 229  * Calculates and evaluates the refault distance of the previously
 230  * evicted page in the context of the node it was allocated in.
 231  *
 232  * Returns %true if the page should be activated, %false otherwise.
 233  */
 234 bool workingset_refault(void *shadow)
 235 {
 236         unsigned long refault_distance;
 237         unsigned long active_file;
 238         struct mem_cgroup *memcg;
 239         unsigned long eviction;
 240         struct lruvec *lruvec;
 241         unsigned long refault;
 242         struct pglist_data *pgdat;
 243         int memcgid;
 244
 245         unpack_shadow(shadow, &memcgid, &pgdat, &eviction);
 246
 247         rcu_read_lock();
 248         /*
 249          * Look up the memcg associated with the stored ID. It might
 250          * have been deleted since the page's eviction.
 251          *
 252          * Note that in rare events the ID could have been recycled
 253          * for a new cgroup that refaults a shared page. This is
 254          * impossible to tell from the available data. However, this
 255          * should be a rare and limited disturbance, and activations
 256          * are always speculative anyway. Ultimately, it's the aging
 257          * algorithm's job to shake out the minimum access frequency
 258          * for the active cache.
 259          *
 260          * XXX: On !CONFIG_MEMCG, this will always return NULL; it
 261          * would be better if the root_mem_cgroup existed in all
 262          * configurations instead.
 263          */
 264         memcg = mem_cgroup_from_id(memcgid);
 265         if (!mem_cgroup_disabled() && !memcg) {
 266                 rcu_read_unlock();
 267                 return false;
 268         }
 269         lruvec = mem_cgroup_lruvec(pgdat, memcg);
 270         refault = atomic_long_read(&lruvec->inactive_age);
 271         active_file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES);
 272         rcu_read_unlock();
 273
 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
 292         inc_node_state(pgdat, WORKINGSET_REFAULT);
 293
 294         if (refault_distance <= active_file) {
 295                 inc_node_state(pgdat, WORKINGSET_ACTIVATE);
 296                 return true;
 297         }
 298         return false;
 299 }
 300
 301 /**
 302  * workingset_activation - note a page activation
 303  * @page: page that is being activated
 304  */
 305 void workingset_activation(struct page *page)
 306 {
 307         struct mem_cgroup *memcg;
 308         struct lruvec *lruvec;
 309
 310         rcu_read_lock();
 311         /*
 312          * Filter non-memcg pages here, e.g. unmap can call
 313          * mark_page_accessed() on VDSO pages.
 314          *
 315          * XXX: See workingset_refault() - this should return
 316          * root_mem_cgroup even for !CONFIG_MEMCG.
 317          */
 318         memcg = page_memcg_rcu(page);
 319         if (!mem_cgroup_disabled() && !memcg)
 320                 goto out;
 321         lruvec = mem_cgroup_lruvec(page_pgdat(page), memcg);
 322         atomic_long_inc(&lruvec->inactive_age);
 323 out:
 324         rcu_read_unlock();
 325 }
 326
 327 /*
 328  * Shadow entries reflect the share of the working set that does not
 329  * fit into memory, so their number depends on the access pattern of
 330  * the workload.  In most cases, they will refault or get reclaimed
 331  * along with the inode, but a (malicious) workload that streams
 332  * through files with a total size several times that of available
 333  * memory, while preventing the inodes from being reclaimed, can
 334  * create excessive amounts of shadow nodes.  To keep a lid on this,
 335  * track shadow nodes and reclaim them when they grow way past the
 336  * point where they would still be useful.
 337  */
 338
 339 static struct list_lru shadow_nodes;
 340
 341 void workingset_update_node(struct radix_tree_node *node, void *private)
 342 {
 343         struct address_space *mapping = private;
 344
 345         /* Only regular page cache has shadow entries */
 346         if (dax_mapping(mapping) || shmem_mapping(mapping))
 347                 return;
 348
 349         /*
 350          * Track non-empty nodes that contain only shadow entries;
 351          * unlink those that contain pages or are being freed.
 352          *
 353          * Avoid acquiring the list_lru lock when the nodes are
 354          * already where they should be. The list_empty() test is safe
 355          * as node->private_list is protected by &mapping->tree_lock.
 356          */
 357         if (node->count && node->count == node->exceptional) {
 358                 if (list_empty(&node->private_list))
 359                         list_lru_add(&shadow_nodes, &node->private_list);
 360         } else {
 361                 if (!list_empty(&node->private_list))
 362                         list_lru_del(&shadow_nodes, &node->private_list);
 363         }
 364 }
 365
 366 static unsigned long count_shadow_nodes(struct shrinker *shrinker,
 367                                         struct shrink_control *sc)
 368 {
 369         unsigned long max_nodes;
 370         unsigned long nodes;
 371         unsigned long cache;
 372
 373         /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
 374         local_irq_disable();
 375         nodes = list_lru_shrink_count(&shadow_nodes, sc);
 376         local_irq_enable();
 377
 378         /*
 379          * Approximate a reasonable limit for the radix tree nodes
 380          * containing shadow entries. We don't need to keep more
 381          * shadow entries than possible pages on the active list,
 382          * since refault distances bigger than that are dismissed.
 383          *
 384          * The size of the active list converges toward 100% of
 385          * overall page cache as memory grows, with only a tiny
 386          * inactive list. Assume the total cache size for that.
 387          *
 388          * Nodes might be sparsely populated, with only one shadow
 389          * entry in the extreme case. Obviously, we cannot keep one
 390          * node for every eligible shadow entry, so compromise on a
 391          * worst-case density of 1/8th. Below that, not all eligible
 392          * refaults can be detected anymore.
 393          *
 394          * On 64-bit with 7 radix_tree_nodes per page and 64 slots
 395          * each, this will reclaim shadow entries when they consume
 396          * ~1.8% of available memory:
 397          *
 398          * PAGE_SIZE / radix_tree_nodes / node_entries * 8 / PAGE_SIZE
 399          */
 400         if (sc->memcg) {
 401                 cache = mem_cgroup_node_nr_lru_pages(sc->memcg, sc->nid,
 402                                                      LRU_ALL_FILE);
 403         } else {
 404                 cache = node_page_state(NODE_DATA(sc->nid), NR_ACTIVE_FILE) +
 405                         node_page_state(NODE_DATA(sc->nid), NR_INACTIVE_FILE);
 406         }
 407         max_nodes = cache >> (RADIX_TREE_MAP_SHIFT - 3);
 408
 409         if (nodes <= max_nodes)
 410                 return 0;
 411         return nodes - max_nodes;
 412 }
 413
 414 static enum lru_status shadow_lru_isolate(struct list_head *item,
 415                                           struct list_lru_one *lru,
 416                                           spinlock_t *lru_lock,
 417                                           void *arg)
 418 {
 419         struct address_space *mapping;
 420         struct radix_tree_node *node;
 421         unsigned int i;
 422         int ret;
 423
 424         /*
 425          * Page cache insertions and deletions synchroneously maintain
 426          * the shadow node LRU under the mapping->tree_lock and the
 427          * lru_lock.  Because the page cache tree is emptied before
 428          * the inode can be destroyed, holding the lru_lock pins any
 429          * address_space that has radix tree nodes on the LRU.
 430          *
 431          * We can then safely transition to the mapping->tree_lock to
 432          * pin only the address_space of the particular node we want
 433          * to reclaim, take the node off-LRU, and drop the lru_lock.
 434          */
 435
 436         node = container_of(item, struct radix_tree_node, private_list);
 437         mapping = container_of(node->root, struct address_space, page_tree);
 438
 439         /* Coming from the list, invert the lock order */
 440         if (!spin_trylock(&mapping->tree_lock)) {
 441                 spin_unlock(lru_lock);
 442                 ret = LRU_RETRY;
 443                 goto out;
 444         }
 445
 446         list_lru_isolate(lru, item);
 447         spin_unlock(lru_lock);
 448
 449         /*
 450          * The nodes should only contain one or more shadow entries,
 451          * no pages, so we expect to be able to remove them all and
 452          * delete and free the empty node afterwards.
 453          */
 454         if (WARN_ON_ONCE(!node->exceptional))
 455                 goto out_invalid;
 456         if (WARN_ON_ONCE(node->count != node->exceptional))
 457                 goto out_invalid;
 458         for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
 459                 if (node->slots[i]) {
 460                         if (WARN_ON_ONCE(!radix_tree_exceptional_entry(node->slots[i])))
 461                                 goto out_invalid;
 462                         if (WARN_ON_ONCE(!node->exceptional))
 463                                 goto out_invalid;
 464                         if (WARN_ON_ONCE(!mapping->nrexceptional))
 465                                 goto out_invalid;
 466                         node->slots[i] = NULL;
 467                         node->exceptional--;
 468                         node->count--;
 469                         mapping->nrexceptional--;
 470                 }
 471         }
 472         if (WARN_ON_ONCE(node->exceptional))
 473                 goto out_invalid;
 474         inc_node_state(page_pgdat(virt_to_page(node)), WORKINGSET_NODERECLAIM);
 475         __radix_tree_delete_node(&mapping->page_tree, node,
 476                                  workingset_update_node, mapping);
 477
 478 out_invalid:
 479         spin_unlock(&mapping->tree_lock);
 480         ret = LRU_REMOVED_RETRY;
 481 out:
 482         local_irq_enable();
 483         cond_resched();
 484         local_irq_disable();
 485         spin_lock(lru_lock);
 486         return ret;
 487 }
 488
 489 static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
 490                                        struct shrink_control *sc)
 491 {
 492         unsigned long ret;
 493
 494         /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
 495         local_irq_disable();
 496         ret = list_lru_shrink_walk(&shadow_nodes, sc, shadow_lru_isolate, NULL);
 497         local_irq_enable();
 498         return ret;
 499 }
 500
 501 static struct shrinker workingset_shadow_shrinker = {
 502         .count_objects = count_shadow_nodes,
 503         .scan_objects = scan_shadow_nodes,
 504         .seeks = DEFAULT_SEEKS,
 505         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
 506 };
 507
 508 /*
 509  * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
 510  * mapping->tree_lock.
 511  */
 512 static struct lock_class_key shadow_nodes_key;
 513
 514 static int __init workingset_init(void)
 515 {
 516         unsigned int timestamp_bits;
 517         unsigned int max_order;
 518         int ret;
 519
 520         BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
 521         /*
 522          * Calculate the eviction bucket size to cover the longest
 523          * actionable refault distance, which is currently half of
 524          * memory (totalram_pages/2). However, memory hotplug may add
 525          * some more pages at runtime, so keep working with up to
 526          * double the initial memory by using totalram_pages as-is.
 527          */
 528         timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
 529         max_order = fls_long(totalram_pages - 1);
 530         if (max_order > timestamp_bits)
 531                 bucket_order = max_order - timestamp_bits;
 532         pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
 533                timestamp_bits, max_order, bucket_order);
 534
 535         ret = list_lru_init_key(&shadow_nodes, &shadow_nodes_key);
 536         if (ret)
 537                 goto err;
 538         ret = register_shrinker(&workingset_shadow_shrinker);
 539         if (ret)
 540                 goto err_list_lru;
 541         return 0;
 542 err_list_lru:
 543         list_lru_destroy(&shadow_nodes);
 544 err:
 545         return ret;
 546 }
 547 module_init(workingset_init);