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