mm/workingset.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Workingset detection
   4  *
   5  * Copyright (C) 2013 Red Hat, Inc., Johannes Weiner
   6  */
   7
   8 #include <linux/memcontrol.h>
   9 #include <linux/writeback.h>
  10 #include <linux/shmem_fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/atomic.h>
  13 #include <linux/module.h>
  14 #include <linux/swap.h>
  15 #include <linux/dax.h>
  16 #include <linux/fs.h>
  17 #include <linux/mm.h>
  18
  19 /*
  20  *              Double CLOCK lists
  21  *
  22  * Per node, two clock lists are maintained for file pages: the
  23  * inactive and the active list.  Freshly faulted pages start out at
  24  * the head of the inactive list and page reclaim scans pages from the
  25  * tail.  Pages that are accessed multiple times on the inactive list
  26  * are promoted to the active list, to protect them from reclaim,
  27  * whereas active pages are demoted to the inactive list when the
  28  * active list grows too big.
  29  *
  30  *   fault ------------------------+
  31  *                                 |
  32  *              +--------------+   |            +-------------+
  33  *   reclaim <- |   inactive   | <-+-- demotion |    active   | <--+
  34  *              +--------------+                +-------------+    |
  35  *                     |                                           |
  36  *                     +-------------- promotion ------------------+
  37  *
  38  *
  39  *              Access frequency and refault distance
  40  *
  41  * A workload is thrashing when its pages are frequently used but they
  42  * are evicted from the inactive list every time before another access
  43  * would have promoted them to the active list.
  44  *
  45  * In cases where the average access distance between thrashing pages
  46  * is bigger than the size of memory there is nothing that can be
  47  * done - the thrashing set could never fit into memory under any
  48  * circumstance.
  49  *
  50  * However, the average access distance could be bigger than the
  51  * inactive list, yet smaller than the size of memory.  In this case,
  52  * the set could fit into memory if it weren't for the currently
  53  * active pages - which may be used more, hopefully less frequently:
  54  *
  55  *      +-memory available to cache-+
  56  *      |                           |
  57  *      +-inactive------+-active----+
  58  *  a b | c d e f g h i | J K L M N |
  59  *      +---------------+-----------+
  60  *
  61  * It is prohibitively expensive to accurately track access frequency
  62  * of pages.  But a reasonable approximation can be made to measure
  63  * thrashing on the inactive list, after which refaulting pages can be
  64  * activated optimistically to compete with the existing active pages.
  65  *
  66  * Approximating inactive page access frequency - Observations:
  67  *
  68  * 1. When a page is accessed for the first time, it is added to the
  69  *    head of the inactive list, slides every existing inactive page
  70  *    towards the tail by one slot, and pushes the current tail page
  71  *    out of memory.
  72  *
  73  * 2. When a page is accessed for the second time, it is promoted to
  74  *    the active list, shrinking the inactive list by one slot.  This
  75  *    also slides all inactive pages that were faulted into the cache
  76  *    more recently than the activated page towards the tail of the
  77  *    inactive list.
  78  *
  79  * Thus:
  80  *
  81  * 1. The sum of evictions and activations between any two points in
  82  *    time indicate the minimum number of inactive pages accessed in
  83  *    between.
  84  *
  85  * 2. Moving one inactive page N page slots towards the tail of the
  86  *    list requires at least N inactive page accesses.
  87  *
  88  * Combining these:
  89  *
  90  * 1. When a page is finally evicted from memory, the number of
  91  *    inactive pages accessed while the page was in cache is at least
  92  *    the number of page slots on the inactive list.
  93  *
  94  * 2. In addition, measuring the sum of evictions and activations (E)
  95  *    at the time of a page's eviction, and comparing it to another
  96  *    reading (R) at the time the page faults back into memory tells
  97  *    the minimum number of accesses while the page was not cached.
  98  *    This is called the refault distance.
  99  *
 100  * Because the first access of the page was the fault and the second
 101  * access the refault, we combine the in-cache distance with the
 102  * out-of-cache distance to get the complete minimum access distance
 103  * of this page:
 104  *
 105  *      NR_inactive + (R - E)
 106  *
 107  * And knowing the minimum access distance of a page, we can easily
 108  * tell if the page would be able to stay in cache assuming all page
 109  * slots in the cache were available:
 110  *
 111  *   NR_inactive + (R - E) <= NR_inactive + NR_active
 112  *
 113  * which can be further simplified to
 114  *
 115  *   (R - E) <= NR_active
 116  *
 117  * Put into words, the refault distance (out-of-cache) can be seen as
 118  * a deficit in inactive list space (in-cache).  If the inactive list
 119  * had (R - E) more page slots, the page would not have been evicted
 120  * in between accesses, but activated instead.  And on a full system,
 121  * the only thing eating into inactive list space is active pages.
 122  *
 123  *
 124  *              Refaulting inactive pages
 125  *
 126  * All that is known about the active list is that the pages have been
 127  * accessed more than once in the past.  This means that at any given
 128  * time there is actually a good chance that pages on the active list
 129  * are no longer in active use.
 130  *
 131  * So when a refault distance of (R - E) is observed and there are at
 132  * least (R - E) active pages, the refaulting page is activated
 133  * optimistically in the hope that (R - E) active pages are actually
 134  * used less frequently than the refaulting page - or even not used at
 135  * all anymore.
 136  *
 137  * That means if inactive cache is refaulting with a suitable refault
 138  * distance, we assume the cache workingset is transitioning and put
 139  * pressure on the current active list.
 140  *
 141  * If this is wrong and demotion kicks in, the pages which are truly
 142  * used more frequently will be reactivated while the less frequently
 143  * used once will be evicted from memory.
 144  *
 145  * But if this is right, the stale pages will be pushed out of memory
 146  * and the used pages get to stay in cache.
 147  *
 148  *              Refaulting active pages
 149  *
 150  * If on the other hand the refaulting pages have recently been
 151  * deactivated, it means that the active list is no longer protecting
 152  * actively used cache from reclaim. The cache is NOT transitioning to
 153  * a different workingset; the existing workingset is thrashing in the
 154  * space allocated to the page cache.
 155  *
 156  *
 157  *              Implementation
 158  *
 159  * For each node's file LRU lists, a counter for inactive evictions
 160  * and activations is maintained (node->inactive_age).
 161  *
 162  * On eviction, a snapshot of this counter (along with some bits to
 163  * identify the node) is stored in the now empty page cache radix tree
 164  * slot of the evicted page.  This is called a shadow entry.
 165  *
 166  * On cache misses for which there are shadow entries, an eligible
 167  * refault distance will immediately activate the refaulting page.
 168  */
 169
 170 #define EVICTION_SHIFT  (RADIX_TREE_EXCEPTIONAL_ENTRY + \
 171                          1 + NODES_SHIFT + MEM_CGROUP_ID_SHIFT)
 172 #define EVICTION_MASK   (~0UL >> EVICTION_SHIFT)
 173
 174 /*
 175  * Eviction timestamps need to be able to cover the full range of
 176  * actionable refaults. However, bits are tight in the radix tree
 177  * entry, and after storing the identifier for the lruvec there might
 178  * not be enough left to represent every single actionable refault. In
 179  * that case, we have to sacrifice granularity for distance, and group
 180  * evictions into coarser buckets by shaving off lower timestamp bits.
 181  */
 182 static unsigned int bucket_order __read_mostly;
 183
 184 static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction,
 185                          bool workingset)
 186 {
 187         eviction >>= bucket_order;
 188         eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
 189         eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
 190         eviction = (eviction << 1) | workingset;
 191         eviction = (eviction << RADIX_TREE_EXCEPTIONAL_SHIFT);
 192
 193         return (void *)(eviction | RADIX_TREE_EXCEPTIONAL_ENTRY);
 194 }
 195
 196 static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
 197                           unsigned long *evictionp, bool *workingsetp)
 198 {
 199         unsigned long entry = (unsigned long)shadow;
 200         int memcgid, nid;
 201         bool workingset;
 202
 203         entry >>= RADIX_TREE_EXCEPTIONAL_SHIFT;
 204         workingset = entry & 1;
 205         entry >>= 1;
 206         nid = entry & ((1UL << NODES_SHIFT) - 1);
 207         entry >>= NODES_SHIFT;
 208         memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
 209         entry >>= MEM_CGROUP_ID_SHIFT;
 210
 211         *memcgidp = memcgid;
 212         *pgdat = NODE_DATA(nid);
 213         *evictionp = entry << bucket_order;
 214         *workingsetp = workingset;
 215 }
 216
 217 /**
 218  * workingset_eviction - note the eviction of a page from memory
 219  * @mapping: address space the page was backing
 220  * @page: the page being evicted
 221  *
 222  * Returns a shadow entry to be stored in @mapping->page_tree in place
 223  * of the evicted @page so that a later refault can be detected.
 224  */
 225 void *workingset_eviction(struct address_space *mapping, struct page *page)
 226 {
 227         struct pglist_data *pgdat = page_pgdat(page);
 228         struct mem_cgroup *memcg = page_memcg(page);
 229         int memcgid = mem_cgroup_id(memcg);
 230         unsigned long eviction;
 231         struct lruvec *lruvec;
 232
 233         /* Page is fully exclusive and pins page->mem_cgroup */
 234         VM_BUG_ON_PAGE(PageLRU(page), page);
 235         VM_BUG_ON_PAGE(page_count(page), page);
 236         VM_BUG_ON_PAGE(!PageLocked(page), page);
 237
 238         lruvec = mem_cgroup_lruvec(pgdat, memcg);
 239         eviction = atomic_long_inc_return(&lruvec->inactive_age);
 240         return pack_shadow(memcgid, pgdat, eviction, PageWorkingset(page));
 241 }
 242
 243 /**
 244  * workingset_refault - evaluate the refault of a previously evicted page
 245  * @page: the freshly allocated replacement page
 246  * @shadow: shadow entry of the evicted page
 247  *
 248  * Calculates and evaluates the refault distance of the previously
 249  * evicted page in the context of the node it was allocated in.
 250  */
 251 void workingset_refault(struct page *page, void *shadow)
 252 {
 253         unsigned long refault_distance;
 254         struct pglist_data *pgdat;
 255         unsigned long active_file;
 256         struct mem_cgroup *memcg;
 257         unsigned long eviction;
 258         struct lruvec *lruvec;
 259         unsigned long refault;
 260         bool workingset;
 261         int memcgid;
 262
 263         unpack_shadow(shadow, &memcgid, &pgdat, &eviction, &workingset);
 264
 265         rcu_read_lock();
 266         /*
 267          * Look up the memcg associated with the stored ID. It might
 268          * have been deleted since the page's eviction.
 269          *
 270          * Note that in rare events the ID could have been recycled
 271          * for a new cgroup that refaults a shared page. This is
 272          * impossible to tell from the available data. However, this
 273          * should be a rare and limited disturbance, and activations
 274          * are always speculative anyway. Ultimately, it's the aging
 275          * algorithm's job to shake out the minimum access frequency
 276          * for the active cache.
 277          *
 278          * XXX: On !CONFIG_MEMCG, this will always return NULL; it
 279          * would be better if the root_mem_cgroup existed in all
 280          * configurations instead.
 281          */
 282         memcg = mem_cgroup_from_id(memcgid);
 283         if (!mem_cgroup_disabled() && !memcg)
 284                 goto out;
 285         lruvec = mem_cgroup_lruvec(pgdat, memcg);
 286         refault = atomic_long_read(&lruvec->inactive_age);
 287         active_file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES);
 288
 289         /*
 290          * Calculate the refault distance
 291          *
 292          * The unsigned subtraction here gives an accurate distance
 293          * across inactive_age overflows in most cases. There is a
 294          * special case: usually, shadow entries have a short lifetime
 295          * and are either refaulted or reclaimed along with the inode
 296          * before they get too old.  But it is not impossible for the
 297          * inactive_age to lap a shadow entry in the field, which can
 298          * then result in a false small refault distance, leading to a
 299          * false activation should this old entry actually refault
 300          * again.  However, earlier kernels used to deactivate
 301          * unconditionally with *every* reclaim invocation for the
 302          * longest time, so the occasional inappropriate activation
 303          * leading to pressure on the active list is not a problem.
 304          */
 305         refault_distance = (refault - eviction) & EVICTION_MASK;
 306
 307         inc_lruvec_state(lruvec, WORKINGSET_REFAULT);
 308
 309         /*
 310          * Compare the distance to the existing workingset size. We
 311          * don't act on pages that couldn't stay resident even if all
 312          * the memory was available to the page cache.
 313          */
 314         if (refault_distance > active_file)
 315                 goto out;
 316
 317         SetPageActive(page);
 318         atomic_long_inc(&lruvec->inactive_age);
 319         inc_lruvec_state(lruvec, WORKINGSET_ACTIVATE);
 320
 321         /* Page was active prior to eviction */
 322         if (workingset) {
 323                 SetPageWorkingset(page);
 324                 inc_lruvec_state(lruvec, WORKINGSET_RESTORE);
 325         }
 326 out:
 327         rcu_read_unlock();
 328 }
 329
 330 /**
 331  * workingset_activation - note a page activation
 332  * @page: page that is being activated
 333  */
 334 void workingset_activation(struct page *page)
 335 {
 336         struct mem_cgroup *memcg;
 337         struct lruvec *lruvec;
 338
 339         rcu_read_lock();
 340         /*
 341          * Filter non-memcg pages here, e.g. unmap can call
 342          * mark_page_accessed() on VDSO pages.
 343          *
 344          * XXX: See workingset_refault() - this should return
 345          * root_mem_cgroup even for !CONFIG_MEMCG.
 346          */
 347         memcg = page_memcg_rcu(page);
 348         if (!mem_cgroup_disabled() && !memcg)
 349                 goto out;
 350         lruvec = mem_cgroup_lruvec(page_pgdat(page), memcg);
 351         atomic_long_inc(&lruvec->inactive_age);
 352 out:
 353         rcu_read_unlock();
 354 }
 355
 356 /*
 357  * Shadow entries reflect the share of the working set that does not
 358  * fit into memory, so their number depends on the access pattern of
 359  * the workload.  In most cases, they will refault or get reclaimed
 360  * along with the inode, but a (malicious) workload that streams
 361  * through files with a total size several times that of available
 362  * memory, while preventing the inodes from being reclaimed, can
 363  * create excessive amounts of shadow nodes.  To keep a lid on this,
 364  * track shadow nodes and reclaim them when they grow way past the
 365  * point where they would still be useful.
 366  */
 367
 368 static struct list_lru shadow_nodes;
 369
 370 void workingset_update_node(struct radix_tree_node *node, void *private)
 371 {
 372         struct address_space *mapping = private;
 373
 374         /* Only regular page cache has shadow entries */
 375         if (dax_mapping(mapping) || shmem_mapping(mapping))
 376                 return;
 377
 378         /*
 379          * Track non-empty nodes that contain only shadow entries;
 380          * unlink those that contain pages or are being freed.
 381          *
 382          * Avoid acquiring the list_lru lock when the nodes are
 383          * already where they should be. The list_empty() test is safe
 384          * as node->private_list is protected by &mapping->tree_lock.
 385          */
 386         if (node->count && node->count == node->exceptional) {
 387                 if (list_empty(&node->private_list))
 388                         list_lru_add(&shadow_nodes, &node->private_list);
 389         } else {
 390                 if (!list_empty(&node->private_list))
 391                         list_lru_del(&shadow_nodes, &node->private_list);
 392         }
 393 }
 394
 395 static unsigned long count_shadow_nodes(struct shrinker *shrinker,
 396                                         struct shrink_control *sc)
 397 {
 398         unsigned long max_nodes;
 399         unsigned long nodes;
 400         unsigned long cache;
 401
 402         /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
 403         local_irq_disable();
 404         nodes = list_lru_shrink_count(&shadow_nodes, sc);
 405         local_irq_enable();
 406
 407         /*
 408          * Approximate a reasonable limit for the radix tree nodes
 409          * containing shadow entries. We don't need to keep more
 410          * shadow entries than possible pages on the active list,
 411          * since refault distances bigger than that are dismissed.
 412          *
 413          * The size of the active list converges toward 100% of
 414          * overall page cache as memory grows, with only a tiny
 415          * inactive list. Assume the total cache size for that.
 416          *
 417          * Nodes might be sparsely populated, with only one shadow
 418          * entry in the extreme case. Obviously, we cannot keep one
 419          * node for every eligible shadow entry, so compromise on a
 420          * worst-case density of 1/8th. Below that, not all eligible
 421          * refaults can be detected anymore.
 422          *
 423          * On 64-bit with 7 radix_tree_nodes per page and 64 slots
 424          * each, this will reclaim shadow entries when they consume
 425          * ~1.8% of available memory:
 426          *
 427          * PAGE_SIZE / radix_tree_nodes / node_entries * 8 / PAGE_SIZE
 428          */
 429         if (sc->memcg) {
 430                 cache = mem_cgroup_node_nr_lru_pages(sc->memcg, sc->nid,
 431                                                      LRU_ALL_FILE);
 432         } else {
 433                 cache = node_page_state(NODE_DATA(sc->nid), NR_ACTIVE_FILE) +
 434                         node_page_state(NODE_DATA(sc->nid), NR_INACTIVE_FILE);
 435         }
 436         max_nodes = cache >> (RADIX_TREE_MAP_SHIFT - 3);
 437
 438         if (nodes <= max_nodes)
 439                 return 0;
 440         return nodes - max_nodes;
 441 }
 442
 443 static enum lru_status shadow_lru_isolate(struct list_head *item,
 444                                           struct list_lru_one *lru,
 445                                           spinlock_t *lru_lock,
 446                                           void *arg)
 447 {
 448         struct address_space *mapping;
 449         struct radix_tree_node *node;
 450         unsigned int i;
 451         int ret;
 452
 453         /*
 454          * Page cache insertions and deletions synchroneously maintain
 455          * the shadow node LRU under the mapping->tree_lock and the
 456          * lru_lock.  Because the page cache tree is emptied before
 457          * the inode can be destroyed, holding the lru_lock pins any
 458          * address_space that has radix tree nodes on the LRU.
 459          *
 460          * We can then safely transition to the mapping->tree_lock to
 461          * pin only the address_space of the particular node we want
 462          * to reclaim, take the node off-LRU, and drop the lru_lock.
 463          */
 464
 465         node = container_of(item, struct radix_tree_node, private_list);
 466         mapping = container_of(node->root, struct address_space, page_tree);
 467
 468         /* Coming from the list, invert the lock order */
 469         if (!spin_trylock(&mapping->tree_lock)) {
 470                 spin_unlock(lru_lock);
 471                 ret = LRU_RETRY;
 472                 goto out;
 473         }
 474
 475         list_lru_isolate(lru, item);
 476         spin_unlock(lru_lock);
 477
 478         /*
 479          * The nodes should only contain one or more shadow entries,
 480          * no pages, so we expect to be able to remove them all and
 481          * delete and free the empty node afterwards.
 482          */
 483         if (WARN_ON_ONCE(!node->exceptional))
 484                 goto out_invalid;
 485         if (WARN_ON_ONCE(node->count != node->exceptional))
 486                 goto out_invalid;
 487         for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
 488                 if (node->slots[i]) {
 489                         if (WARN_ON_ONCE(!radix_tree_exceptional_entry(node->slots[i])))
 490                                 goto out_invalid;
 491                         if (WARN_ON_ONCE(!node->exceptional))
 492                                 goto out_invalid;
 493                         if (WARN_ON_ONCE(!mapping->nrexceptional))
 494                                 goto out_invalid;
 495                         node->slots[i] = NULL;
 496                         node->exceptional--;
 497                         node->count--;
 498                         mapping->nrexceptional--;
 499                 }
 500         }
 501         if (WARN_ON_ONCE(node->exceptional))
 502                 goto out_invalid;
 503         inc_lruvec_page_state(virt_to_page(node), WORKINGSET_NODERECLAIM);
 504         __radix_tree_delete_node(&mapping->page_tree, node,
 505                                  workingset_update_node, mapping);
 506
 507 out_invalid:
 508         spin_unlock(&mapping->tree_lock);
 509         ret = LRU_REMOVED_RETRY;
 510 out:
 511         local_irq_enable();
 512         cond_resched();
 513         local_irq_disable();
 514         spin_lock(lru_lock);
 515         return ret;
 516 }
 517
 518 static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
 519                                        struct shrink_control *sc)
 520 {
 521         unsigned long ret;
 522
 523         /* list_lru lock nests inside IRQ-safe mapping->tree_lock */
 524         local_irq_disable();
 525         ret = list_lru_shrink_walk(&shadow_nodes, sc, shadow_lru_isolate, NULL);
 526         local_irq_enable();
 527         return ret;
 528 }
 529
 530 static struct shrinker workingset_shadow_shrinker = {
 531         .count_objects = count_shadow_nodes,
 532         .scan_objects = scan_shadow_nodes,
 533         .seeks = DEFAULT_SEEKS,
 534         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
 535 };
 536
 537 /*
 538  * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
 539  * mapping->tree_lock.
 540  */
 541 static struct lock_class_key shadow_nodes_key;
 542
 543 static int __init workingset_init(void)
 544 {
 545         unsigned int timestamp_bits;
 546         unsigned int max_order;
 547         int ret;
 548
 549         BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
 550         /*
 551          * Calculate the eviction bucket size to cover the longest
 552          * actionable refault distance, which is currently half of
 553          * memory (totalram_pages/2). However, memory hotplug may add
 554          * some more pages at runtime, so keep working with up to
 555          * double the initial memory by using totalram_pages as-is.
 556          */
 557         timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
 558         max_order = fls_long(totalram_pages - 1);
 559         if (max_order > timestamp_bits)
 560                 bucket_order = max_order - timestamp_bits;
 561         pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
 562                timestamp_bits, max_order, bucket_order);
 563
 564         ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key);
 565         if (ret)
 566                 goto err;
 567         ret = register_shrinker(&workingset_shadow_shrinker);
 568         if (ret)
 569                 goto err_list_lru;
 570         return 0;
 571 err_list_lru:
 572         list_lru_destroy(&shadow_nodes);
 573 err:
 574         return ret;
 575 }
 576 module_init(workingset_init);