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  *              Activating refaulting 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  * If this is wrong and demotion kicks in, the pages which are truly
 138  * used more frequently will be reactivated while the less frequently
 139  * used once will be evicted from memory.
 140  *
 141  * But if this is right, the stale pages will be pushed out of memory
 142  * and the used pages get to stay in cache.
 143  *
 144  *
 145  *              Implementation
 146  *
 147  * For each node's file LRU lists, a counter for inactive evictions
 148  * and activations is maintained (node->inactive_age).
 149  *
 150  * On eviction, a snapshot of this counter (along with some bits to
 151  * identify the node) is stored in the now empty page cache radix tree
 152  * slot of the evicted page.  This is called a shadow entry.
 153  *
 154  * On cache misses for which there are shadow entries, an eligible
 155  * refault distance will immediately activate the refaulting page.
 156  */
 157
 158 #define EVICTION_SHIFT  (RADIX_TREE_EXCEPTIONAL_ENTRY + \
 159                          NODES_SHIFT +  \
 160                          MEM_CGROUP_ID_SHIFT)
 161 #define EVICTION_MASK   (~0UL >> EVICTION_SHIFT)
 162
 163 /*
 164  * Eviction timestamps need to be able to cover the full range of
 165  * actionable refaults. However, bits are tight in the radix tree
 166  * entry, and after storing the identifier for the lruvec there might
 167  * not be enough left to represent every single actionable refault. In
 168  * that case, we have to sacrifice granularity for distance, and group
 169  * evictions into coarser buckets by shaving off lower timestamp bits.
 170  */
 171 static unsigned int bucket_order __read_mostly;
 172
 173 static void *pack_shadow(int memcgid, pg_data_t *pgdat, unsigned long eviction)
 174 {
 175         eviction >>= bucket_order;
 176         eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
 177         eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
 178         eviction = (eviction << RADIX_TREE_EXCEPTIONAL_SHIFT);
 179
 180         return (void *)(eviction | RADIX_TREE_EXCEPTIONAL_ENTRY);
 181 }
 182
 183 static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
 184                           unsigned long *evictionp)
 185 {
 186         unsigned long entry = (unsigned long)shadow;
 187         int memcgid, nid;
 188
 189         entry >>= RADIX_TREE_EXCEPTIONAL_SHIFT;
 190         nid = entry & ((1UL << NODES_SHIFT) - 1);
 191         entry >>= NODES_SHIFT;
 192         memcgid = entry & ((1UL << MEM_CGROUP_ID_SHIFT) - 1);
 193         entry >>= MEM_CGROUP_ID_SHIFT;
 194
 195         *memcgidp = memcgid;
 196         *pgdat = NODE_DATA(nid);
 197         *evictionp = entry << bucket_order;
 198 }
 199
 200 /**
 201  * workingset_eviction - note the eviction of a page from memory
 202  * @mapping: address space the page was backing
 203  * @page: the page being evicted
 204  *
 205  * Returns a shadow entry to be stored in @mapping->i_pages in place
 206  * of the evicted @page so that a later refault can be detected.
 207  */
 208 void *workingset_eviction(struct address_space *mapping, struct page *page)
 209 {
 210         struct mem_cgroup *memcg = page_memcg(page);
 211         struct pglist_data *pgdat = page_pgdat(page);
 212         int memcgid = mem_cgroup_id(memcg);
 213         unsigned long eviction;
 214         struct lruvec *lruvec;
 215
 216         /* Page is fully exclusive and pins page->mem_cgroup */
 217         VM_BUG_ON_PAGE(PageLRU(page), page);
 218         VM_BUG_ON_PAGE(page_count(page), page);
 219         VM_BUG_ON_PAGE(!PageLocked(page), page);
 220
 221         lruvec = mem_cgroup_lruvec(pgdat, memcg);
 222         eviction = atomic_long_inc_return(&lruvec->inactive_age);
 223         return pack_shadow(memcgid, pgdat, eviction);
 224 }
 225
 226 /**
 227  * workingset_refault - evaluate the refault of a previously evicted page
 228  * @shadow: shadow entry of the evicted page
 229  *
 230  * Calculates and evaluates the refault distance of the previously
 231  * evicted page in the context of the node it was allocated in.
 232  *
 233  * Returns %true if the page should be activated, %false otherwise.
 234  */
 235 bool workingset_refault(void *shadow)
 236 {
 237         unsigned long refault_distance;
 238         unsigned long active_file;
 239         struct mem_cgroup *memcg;
 240         unsigned long eviction;
 241         struct lruvec *lruvec;
 242         unsigned long refault;
 243         struct pglist_data *pgdat;
 244         int memcgid;
 245
 246         unpack_shadow(shadow, &memcgid, &pgdat, &eviction);
 247
 248         rcu_read_lock();
 249         /*
 250          * Look up the memcg associated with the stored ID. It might
 251          * have been deleted since the page's eviction.
 252          *
 253          * Note that in rare events the ID could have been recycled
 254          * for a new cgroup that refaults a shared page. This is
 255          * impossible to tell from the available data. However, this
 256          * should be a rare and limited disturbance, and activations
 257          * are always speculative anyway. Ultimately, it's the aging
 258          * algorithm's job to shake out the minimum access frequency
 259          * for the active cache.
 260          *
 261          * XXX: On !CONFIG_MEMCG, this will always return NULL; it
 262          * would be better if the root_mem_cgroup existed in all
 263          * configurations instead.
 264          */
 265         memcg = mem_cgroup_from_id(memcgid);
 266         if (!mem_cgroup_disabled() && !memcg) {
 267                 rcu_read_unlock();
 268                 return false;
 269         }
 270         lruvec = mem_cgroup_lruvec(pgdat, memcg);
 271         refault = atomic_long_read(&lruvec->inactive_age);
 272         active_file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES);
 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_lruvec_state(lruvec, WORKINGSET_REFAULT);
 293
 294         if (refault_distance <= active_file) {
 295                 inc_lruvec_state(lruvec, WORKINGSET_ACTIVATE);
 296                 rcu_read_unlock();
 297                 return true;
 298         }
 299         rcu_read_unlock();
 300         return false;
 301 }
 302
 303 /**
 304  * workingset_activation - note a page activation
 305  * @page: page that is being activated
 306  */
 307 void workingset_activation(struct page *page)
 308 {
 309         struct mem_cgroup *memcg;
 310         struct lruvec *lruvec;
 311
 312         rcu_read_lock();
 313         /*
 314          * Filter non-memcg pages here, e.g. unmap can call
 315          * mark_page_accessed() on VDSO pages.
 316          *
 317          * XXX: See workingset_refault() - this should return
 318          * root_mem_cgroup even for !CONFIG_MEMCG.
 319          */
 320         memcg = page_memcg_rcu(page);
 321         if (!mem_cgroup_disabled() && !memcg)
 322                 goto out;
 323         lruvec = mem_cgroup_lruvec(page_pgdat(page), memcg);
 324         atomic_long_inc(&lruvec->inactive_age);
 325 out:
 326         rcu_read_unlock();
 327 }
 328
 329 /*
 330  * Shadow entries reflect the share of the working set that does not
 331  * fit into memory, so their number depends on the access pattern of
 332  * the workload.  In most cases, they will refault or get reclaimed
 333  * along with the inode, but a (malicious) workload that streams
 334  * through files with a total size several times that of available
 335  * memory, while preventing the inodes from being reclaimed, can
 336  * create excessive amounts of shadow nodes.  To keep a lid on this,
 337  * track shadow nodes and reclaim them when they grow way past the
 338  * point where they would still be useful.
 339  */
 340
 341 static struct list_lru shadow_nodes;
 342
 343 void workingset_update_node(struct radix_tree_node *node)
 344 {
 345         /*
 346          * Track non-empty nodes that contain only shadow entries;
 347          * unlink those that contain pages or are being freed.
 348          *
 349          * Avoid acquiring the list_lru lock when the nodes are
 350          * already where they should be. The list_empty() test is safe
 351          * as node->private_list is protected by the i_pages lock.
 352          */
 353         if (node->count && node->count == node->exceptional) {
 354                 if (list_empty(&node->private_list))
 355                         list_lru_add(&shadow_nodes, &node->private_list);
 356         } else {
 357                 if (!list_empty(&node->private_list))
 358                         list_lru_del(&shadow_nodes, &node->private_list);
 359         }
 360 }
 361
 362 static unsigned long count_shadow_nodes(struct shrinker *shrinker,
 363                                         struct shrink_control *sc)
 364 {
 365         unsigned long max_nodes;
 366         unsigned long nodes;
 367         unsigned long cache;
 368
 369         nodes = list_lru_shrink_count(&shadow_nodes, sc);
 370
 371         /*
 372          * Approximate a reasonable limit for the radix tree nodes
 373          * containing shadow entries. We don't need to keep more
 374          * shadow entries than possible pages on the active list,
 375          * since refault distances bigger than that are dismissed.
 376          *
 377          * The size of the active list converges toward 100% of
 378          * overall page cache as memory grows, with only a tiny
 379          * inactive list. Assume the total cache size for that.
 380          *
 381          * Nodes might be sparsely populated, with only one shadow
 382          * entry in the extreme case. Obviously, we cannot keep one
 383          * node for every eligible shadow entry, so compromise on a
 384          * worst-case density of 1/8th. Below that, not all eligible
 385          * refaults can be detected anymore.
 386          *
 387          * On 64-bit with 7 radix_tree_nodes per page and 64 slots
 388          * each, this will reclaim shadow entries when they consume
 389          * ~1.8% of available memory:
 390          *
 391          * PAGE_SIZE / radix_tree_nodes / node_entries * 8 / PAGE_SIZE
 392          */
 393         if (sc->memcg) {
 394                 cache = mem_cgroup_node_nr_lru_pages(sc->memcg, sc->nid,
 395                                                      LRU_ALL_FILE);
 396         } else {
 397                 cache = node_page_state(NODE_DATA(sc->nid), NR_ACTIVE_FILE) +
 398                         node_page_state(NODE_DATA(sc->nid), NR_INACTIVE_FILE);
 399         }
 400         max_nodes = cache >> (RADIX_TREE_MAP_SHIFT - 3);
 401
 402         if (!nodes)
 403                 return SHRINK_EMPTY;
 404
 405         if (nodes <= max_nodes)
 406                 return 0;
 407         return nodes - max_nodes;
 408 }
 409
 410 static enum lru_status shadow_lru_isolate(struct list_head *item,
 411                                           struct list_lru_one *lru,
 412                                           spinlock_t *lru_lock,
 413                                           void *arg)
 414 {
 415         struct address_space *mapping;
 416         struct radix_tree_node *node;
 417         unsigned int i;
 418         int ret;
 419
 420         /*
 421          * Page cache insertions and deletions synchroneously maintain
 422          * the shadow node LRU under the i_pages lock and the
 423          * lru_lock.  Because the page cache tree is emptied before
 424          * the inode can be destroyed, holding the lru_lock pins any
 425          * address_space that has radix tree nodes on the LRU.
 426          *
 427          * We can then safely transition to the i_pages lock to
 428          * pin only the address_space of the particular node we want
 429          * to reclaim, take the node off-LRU, and drop the lru_lock.
 430          */
 431
 432         node = container_of(item, struct radix_tree_node, private_list);
 433         mapping = container_of(node->root, struct address_space, i_pages);
 434
 435         /* Coming from the list, invert the lock order */
 436         if (!xa_trylock(&mapping->i_pages)) {
 437                 spin_unlock_irq(lru_lock);
 438                 ret = LRU_RETRY;
 439                 goto out;
 440         }
 441
 442         list_lru_isolate(lru, item);
 443         spin_unlock(lru_lock);
 444
 445         /*
 446          * The nodes should only contain one or more shadow entries,
 447          * no pages, so we expect to be able to remove them all and
 448          * delete and free the empty node afterwards.
 449          */
 450         if (WARN_ON_ONCE(!node->exceptional))
 451                 goto out_invalid;
 452         if (WARN_ON_ONCE(node->count != node->exceptional))
 453                 goto out_invalid;
 454         for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
 455                 if (node->slots[i]) {
 456                         if (WARN_ON_ONCE(!radix_tree_exceptional_entry(node->slots[i])))
 457                                 goto out_invalid;
 458                         if (WARN_ON_ONCE(!node->exceptional))
 459                                 goto out_invalid;
 460                         if (WARN_ON_ONCE(!mapping->nrexceptional))
 461                                 goto out_invalid;
 462                         node->slots[i] = NULL;
 463                         node->exceptional--;
 464                         node->count--;
 465                         mapping->nrexceptional--;
 466                 }
 467         }
 468         if (WARN_ON_ONCE(node->exceptional))
 469                 goto out_invalid;
 470         inc_lruvec_page_state(virt_to_page(node), WORKINGSET_NODERECLAIM);
 471         __radix_tree_delete_node(&mapping->i_pages, node,
 472                                  workingset_lookup_update(mapping));
 473
 474 out_invalid:
 475         xa_unlock_irq(&mapping->i_pages);
 476         ret = LRU_REMOVED_RETRY;
 477 out:
 478         cond_resched();
 479         spin_lock_irq(lru_lock);
 480         return ret;
 481 }
 482
 483 static unsigned long scan_shadow_nodes(struct shrinker *shrinker,
 484                                        struct shrink_control *sc)
 485 {
 486         /* list_lru lock nests inside the IRQ-safe i_pages lock */
 487         return list_lru_shrink_walk_irq(&shadow_nodes, sc, shadow_lru_isolate,
 488                                         NULL);
 489 }
 490
 491 static struct shrinker workingset_shadow_shrinker = {
 492         .count_objects = count_shadow_nodes,
 493         .scan_objects = scan_shadow_nodes,
 494         .seeks = DEFAULT_SEEKS,
 495         .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
 496 };
 497
 498 /*
 499  * Our list_lru->lock is IRQ-safe as it nests inside the IRQ-safe
 500  * i_pages lock.
 501  */
 502 static struct lock_class_key shadow_nodes_key;
 503
 504 static int __init workingset_init(void)
 505 {
 506         unsigned int timestamp_bits;
 507         unsigned int max_order;
 508         int ret;
 509
 510         BUILD_BUG_ON(BITS_PER_LONG < EVICTION_SHIFT);
 511         /*
 512          * Calculate the eviction bucket size to cover the longest
 513          * actionable refault distance, which is currently half of
 514          * memory (totalram_pages/2). However, memory hotplug may add
 515          * some more pages at runtime, so keep working with up to
 516          * double the initial memory by using totalram_pages as-is.
 517          */
 518         timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
 519         max_order = fls_long(totalram_pages - 1);
 520         if (max_order > timestamp_bits)
 521                 bucket_order = max_order - timestamp_bits;
 522         pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",
 523                timestamp_bits, max_order, bucket_order);
 524
 525         ret = prealloc_shrinker(&workingset_shadow_shrinker);
 526         if (ret)
 527                 goto err;
 528         ret = __list_lru_init(&shadow_nodes, true, &shadow_nodes_key,
 529                               &workingset_shadow_shrinker);
 530         if (ret)
 531                 goto err_list_lru;
 532         register_shrinker_prepared(&workingset_shadow_shrinker);
 533         return 0;
 534 err_list_lru:
 535         free_prealloced_shrinker(&workingset_shadow_shrinker);
 536 err:
 537         return ret;
 538 }
 539 module_init(workingset_init);