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
 273         /*
 274          * The unsigned subtraction here gives an accurate distance
 275          * across inactive_age overflows in most cases.
 276          *
 277          * There is a special case: usually, shadow entries have a
 278          * short lifetime and are either refaulted or reclaimed along
 279          * with the inode before they get too old.  But it is not
 280          * impossible for the inactive_age to lap a shadow entry in
 281          * the field, which can then can result in a false small
 282          * refault distance, leading to a false activation should this
 283          * old entry actually refault again.  However, earlier kernels
 284          * used to deactivate unconditionally with *every* reclaim
 285          * invocation for the longest time, so the occasional
 286          * inappropriate activation leading to pressure on the active
 287          * list is not a problem.
 288          */
 289         refault_distance = (refault - eviction) & EVICTION_MASK;
 290
 291         inc_node_state(pgdat, WORKINGSET_REFAULT);
 292         inc_memcg_state(memcg, WORKINGSET_REFAULT);
 293
 294         if (refault_distance <= active_file) {
 295                 inc_node_state(pgdat, WORKINGSET_ACTIVATE);
 296                 inc_memcg_state(memcg, WORKINGSET_ACTIVATE);
 297                 rcu_read_unlock();
 298                 return true;
 299         }
 300         rcu_read_unlock();
 301         return false;
 302 }
 303
 304 /**
 305  * workingset_activation - note a page activation
 306  * @page: page that is being activated
 307  */
 308 void workingset_activation(struct page *page)
 309 {
 310         struct mem_cgroup *memcg;
 311         struct lruvec *lruvec;
 312
 313         rcu_read_lock();
 314         /*
 315          * Filter non-memcg pages here, e.g. unmap can call
 316          * mark_page_accessed() on VDSO pages.
 317          *
 318          * XXX: See workingset_refault() - this should return
 319          * root_mem_cgroup even for !CONFIG_MEMCG.
 320          */
 321         memcg = page_memcg_rcu(page);
 322         if (!mem_cgroup_disabled() && !memcg)
 323                 goto out;
 324         lruvec = mem_cgroup_lruvec(page_pgdat(page), memcg);
 325         atomic_long_inc(&lruvec->inactive_age);
 326 out:
 327         rcu_read_unlock();
 328 }
 329
 330 /*
 331  * Shadow entries reflect the share of the working set that does not
 332  * fit into memory, so their number depends on the access pattern of
 333  * the workload.  In most cases, they will refault or get reclaimed
 334  * along with the inode, but a (malicious) workload that streams
 335  * through files with a total size several times that of available
 336  * memory, while preventing the inodes from being reclaimed, can
 337  * create excessive amounts of shadow nodes.  To keep a lid on this,
 338  * track shadow nodes and reclaim them when they grow way past the
 339  * point where they would still be useful.
 340  */
 341
 342 static struct list_lru shadow_nodes;
 343
 344 void workingset_update_node(struct radix_tree_node *node, void *private)
 345 {
 346         struct address_space *mapping = private;
 347
 348         /* Only regular page cache has shadow entries */
 349         if (dax_mapping(mapping) || shmem_mapping(mapping))
 350                 return;
 351
 352         /*
 353          * Track non-empty nodes that contain only shadow entries;
 354          * unlink those that contain pages or are being freed.
 355          *
 356          * Avoid acquiring the list_lru lock when the nodes are
 357          * already where they should be. The list_empty() test is safe
 358          * as node->private_list is protected by &mapping->tree_lock.
 359          */
 360         if (node->count && node->count == node->exceptional) {
 361                 if (list_empty(&node->private_list))
 362                         list_lru_add(&shadow_nodes, &node->private_list);
 363         } else {
 364                 if (!list_empty(&node->private_list))
 365                         list_lru_del(&shadow_nodes, &node->private_list);
 366         }
 367 }
 368
 369 static unsigned long count_shadow_nodes(struct shrinker *shrinker,
 370                                         struct shrink_control *sc)
 371 {
 372         unsigned long max_nodes;
 373         unsigned long nodes;
 374         unsigned long cache;
 375
 376         /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
 377         local_irq_disable();
 378         nodes = list_lru_shrink_count(&shadow_nodes, sc);
 379         local_irq_enable();
 380
 381         /*
 382          * Approximate a reasonable limit for the radix tree nodes
 383          * containing shadow entries. We don't need to keep more
 384          * shadow entries than possible pages on the active list,
 385          * since refault distances bigger than that are dismissed.
 386          *
 387          * The size of the active list converges toward 100% of
 388          * overall page cache as memory grows, with only a tiny
 389          * inactive list. Assume the total cache size for that.
 390          *
 391          * Nodes might be sparsely populated, with only one shadow
 392          * entry in the extreme case. Obviously, we cannot keep one
 393          * node for every eligible shadow entry, so compromise on a
 394          * worst-case density of 1/8th. Below that, not all eligible
 395          * refaults can be detected anymore.
 396          *
 397          * On 64-bit with 7 radix_tree_nodes per page and 64 slots
 398          * each, this will reclaim shadow entries when they consume
 399          * ~1.8% of available memory:
 400          *
 401          * PAGE_SIZE / radix_tree_nodes / node_entries * 8 / PAGE_SIZE
 402          */
 403         if (sc->memcg) {
 404                 cache = mem_cgroup_node_nr_lru_pages(sc->memcg, sc->nid,
 405                                                      LRU_ALL_FILE);
 406         } else {
 407                 cache = node_page_state(NODE_DATA(sc->nid), NR_ACTIVE_FILE) +
 408                         node_page_state(NODE_DATA(sc->nid), NR_INACTIVE_FILE);
 409         }
 410         max_nodes = cache >> (RADIX_TREE_MAP_SHIFT - 3);
 411
 412         if (nodes <= max_nodes)
 413                 return 0;
 414         return nodes - max_nodes;
 415 }
 416
 417 static enum lru_status shadow_lru_isolate(struct list_head *item,
 418                                           struct list_lru_one *lru,
 419                                           spinlock_t *lru_lock,
 420                                           void *arg)
 421 {
 422         struct address_space *mapping;
 423         struct radix_tree_node *node;
 424         unsigned int i;
 425         int ret;
 426
 427         /*
 428          * Page cache insertions and deletions synchroneously maintain
 429          * the shadow node LRU under the mapping->tree_lock and the
 430          * lru_lock.  Because the page cache tree is emptied before
 431          * the inode can be destroyed, holding the lru_lock pins any
 432          * address_space that has radix tree nodes on the LRU.
 433          *
 434          * We can then safely transition to the mapping->tree_lock to
 435          * pin only the address_space of the particular node we want
 436          * to reclaim, take the node off-LRU, and drop the lru_lock.
 437          */
 438
 439         node = container_of(item, struct radix_tree_node, private_list);
 440         mapping = container_of(node->root, struct address_space, page_tree);
 441
 442         /* Coming from the list, invert the lock order */
 443         if (!spin_trylock(&mapping->tree_lock)) {
 444                 spin_unlock(lru_lock);
 445                 ret = LRU_RETRY;
 446                 goto out;
 447         }
 448
 449         list_lru_isolate(lru, item);
 450         spin_unlock(lru_lock);
 451
 452         /*
 453          * The nodes should only contain one or more shadow entries,
 454          * no pages, so we expect to be able to remove them all and
 455          * delete and free the empty node afterwards.
 456          */
 457         if (WARN_ON_ONCE(!node->exceptional))
 458                 goto out_invalid;
 459         if (WARN_ON_ONCE(node->count != node->exceptional))
 460                 goto out_invalid;
 461         for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
 462                 if (node->slots[i]) {
 463                         if (WARN_ON_ONCE(!radix_tree_exceptional_entry(node->slots[i])))
 464                                 goto out_invalid;
 465                         if (WARN_ON_ONCE(!node->exceptional))
 466                                 goto out_invalid;
 467                         if (WARN_ON_ONCE(!mapping->nrexceptional))
 468                                 goto out_invalid;
 469                         node->slots[i] = NULL;
 470                         node->exceptional--;
 471                         node->count--;
 472                         mapping->nrexceptional--;
 473                 }
 474         }
 475         if (WARN_ON_ONCE(node->exceptional))
 476                 goto out_invalid;
 477         inc_node_state(page_pgdat(virt_to_page(node)), WORKINGSET_NODERECLAIM);
 478         inc_memcg_page_state(virt_to_page(node), WORKINGSET_NODERECLAIM);
 479         __radix_tree_delete_node(&mapping->page_tree, node,
 480                                  workingset_update_node, mapping);
 481
 482 out_invalid:
 483         spin_unlock(&mapping->tree_lock);
 484         ret = LRU_REMOVED_RETRY;
 485 out:
 486         local_irq_enable();
 487         cond_resched();
 488         local_irq_disable();
 489         spin_lock(lru_lock);
 490         return ret;
 491 }
 492
 493 static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
 494                                        struct shrink_control *sc)
 495 {
 496         unsigned long ret;
 497
 498         /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
 499         local_irq_disable();
 500         ret = list_lru_shrink_walk(&shadow_nodes, sc, shadow_lru_isolate, NULL);
 501         local_irq_enable();
 502         return ret;
 503 }
 504
 505 static struct shrinker workingset_shadow_shrinker = {
 506         .count_objects = count_shadow_nodes,
 507         .scan_objects = scan_shadow_nodes,
 508         .seeks = DEFAULT_SEEKS,
 509         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
 510 };
 511
 512 /*
 513  * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
 514  * mapping->tree_lock.
 515  */
 516 static struct lock_class_key shadow_nodes_key;
 517
 518 static int __init workingset_init(void)
 519 {
 520         unsigned int timestamp_bits;
 521         unsigned int max_order;
 522         int ret;
 523
 524         BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
 525         /*
 526          * Calculate the eviction bucket size to cover the longest
 527          * actionable refault distance, which is currently half of
 528          * memory (totalram_pages/2). However, memory hotplug may add
 529          * some more pages at runtime, so keep working with up to
 530          * double the initial memory by using totalram_pages as-is.
 531          */
 532         timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
 533         max_order = fls_long(totalram_pages - 1);
 534         if (max_order > timestamp_bits)
 535                 bucket_order = max_order - timestamp_bits;
 536         pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
 537                timestamp_bits, max_order, bucket_order);
 538
 539         ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key);
 540         if (ret)
 541                 goto err;
 542         ret = register_shrinker(&workingset_shadow_shrinker);
 543         if (ret)
 544                 goto err_list_lru;
 545         return 0;
 546 err_list_lru:
 547         list_lru_destroy(&shadow_nodes);
 548 err:
 549         return ret;
 550 }
 551 module_init(workingset_init);