linux/mm/vmscan.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   4 *
   5 *  Swap reorganised 29.12.95, Stephen Tweedie.
   6 *  kswapd added: 7.1.96  sct
   7 *  Removed kswapd_ctl limits, and swap out as many pages as needed
   8 *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
   9 *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
  10 *  Multiqueue VM started 5.8.00, Rik van Riel.
  11 */
  12
  13#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  14
  15#include <linux/mm.h>
  16#include <linux/sched/mm.h>
  17#include <linux/module.h>
  18#include <linux/gfp.h>
  19#include <linux/kernel_stat.h>
  20#include <linux/swap.h>
  21#include <linux/pagemap.h>
  22#include <linux/init.h>
  23#include <linux/highmem.h>
  24#include <linux/vmpressure.h>
  25#include <linux/vmstat.h>
  26#include <linux/file.h>
  27#include <linux/writeback.h>
  28#include <linux/blkdev.h>
  29#include <linux/buffer_head.h>  /* for try_to_release_page(),
  30                                        buffer_heads_over_limit */
  31#include <linux/mm_inline.h>
  32#include <linux/backing-dev.h>
  33#include <linux/rmap.h>
  34#include <linux/topology.h>
  35#include <linux/cpu.h>
  36#include <linux/cpuset.h>
  37#include <linux/compaction.h>
  38#include <linux/notifier.h>
  39#include <linux/rwsem.h>
  40#include <linux/delay.h>
  41#include <linux/kthread.h>
  42#include <linux/freezer.h>
  43#include <linux/memcontrol.h>
  44#include <linux/migrate.h>
  45#include <linux/delayacct.h>
  46#include <linux/sysctl.h>
  47#include <linux/oom.h>
  48#include <linux/pagevec.h>
  49#include <linux/prefetch.h>
  50#include <linux/printk.h>
  51#include <linux/dax.h>
  52#include <linux/psi.h>
  53
  54#include <asm/tlbflush.h>
  55#include <asm/div64.h>
  56
  57#include <linux/swapops.h>
  58#include <linux/balloon_compaction.h>
  59
  60#include "internal.h"
  61
  62#define CREATE_TRACE_POINTS
  63#include <trace/events/vmscan.h>
  64
  65struct scan_control {
  66        /* How many pages shrink_list() should reclaim */
  67        unsigned long nr_to_reclaim;
  68
  69        /*
  70         * Nodemask of nodes allowed by the caller. If NULL, all nodes
  71         * are scanned.
  72         */
  73        nodemask_t      *nodemask;
  74
  75        /*
  76         * The memory cgroup that hit its limit and as a result is the
  77         * primary target of this reclaim invocation.
  78         */
  79        struct mem_cgroup *target_mem_cgroup;
  80
  81        /*
  82         * Scan pressure balancing between anon and file LRUs
  83         */
  84        unsigned long   anon_cost;
  85        unsigned long   file_cost;
  86
  87        /* Can active pages be deactivated as part of reclaim? */
  88#define DEACTIVATE_ANON 1
  89#define DEACTIVATE_FILE 2
  90        unsigned int may_deactivate:2;
  91        unsigned int force_deactivate:1;
  92        unsigned int skipped_deactivate:1;
  93
  94        /* Writepage batching in laptop mode; RECLAIM_WRITE */
  95        unsigned int may_writepage:1;
  96
  97        /* Can mapped pages be reclaimed? */
  98        unsigned int may_unmap:1;
  99
 100        /* Can pages be swapped as part of reclaim? */
 101        unsigned int may_swap:1;
 102
 103        /*
 104         * Cgroup memory below memory.low is protected as long as we
 105         * don't threaten to OOM. If any cgroup is reclaimed at
 106         * reduced force or passed over entirely due to its memory.low
 107         * setting (memcg_low_skipped), and nothing is reclaimed as a
 108         * result, then go back for one more cycle that reclaims the protected
 109         * memory (memcg_low_reclaim) to avert OOM.
 110         */
 111        unsigned int memcg_low_reclaim:1;
 112        unsigned int memcg_low_skipped:1;
 113
 114        unsigned int hibernation_mode:1;
 115
 116        /* One of the zones is ready for compaction */
 117        unsigned int compaction_ready:1;
 118
 119        /* There is easily reclaimable cold cache in the current node */
 120        unsigned int cache_trim_mode:1;
 121
 122        /* The file pages on the current node are dangerously low */
 123        unsigned int file_is_tiny:1;
 124
 125        /* Always discard instead of demoting to lower tier memory */
 126        unsigned int no_demotion:1;
 127
 128        /* Allocation order */
 129        s8 order;
 130
 131        /* Scan (total_size >> priority) pages at once */
 132        s8 priority;
 133
 134        /* The highest zone to isolate pages for reclaim from */
 135        s8 reclaim_idx;
 136
 137        /* This context's GFP mask */
 138        gfp_t gfp_mask;
 139
 140        /* Incremented by the number of inactive pages that were scanned */
 141        unsigned long nr_scanned;
 142
 143        /* Number of pages freed so far during a call to shrink_zones() */
 144        unsigned long nr_reclaimed;
 145
 146        struct {
 147                unsigned int dirty;
 148                unsigned int unqueued_dirty;
 149                unsigned int congested;
 150                unsigned int writeback;
 151                unsigned int immediate;
 152                unsigned int file_taken;
 153                unsigned int taken;
 154        } nr;
 155
 156        /* for recording the reclaimed slab by now */
 157        struct reclaim_state reclaim_state;
 158};
 159
 160#ifdef ARCH_HAS_PREFETCHW
 161#define prefetchw_prev_lru_page(_page, _base, _field)                   \
 162        do {                                                            \
 163                if ((_page)->lru.prev != _base) {                       \
 164                        struct page *prev;                              \
 165                                                                        \
 166                        prev = lru_to_page(&(_page->lru));              \
 167                        prefetchw(&prev->_field);                       \
 168                }                                                       \
 169        } while (0)
 170#else
 171#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 172#endif
 173
 174/*
 175 * From 0 .. 200.  Higher means more swappy.
 176 */
 177int vm_swappiness = 60;
 178
 179static void set_task_reclaim_state(struct task_struct *task,
 180                                   struct reclaim_state *rs)
 181{
 182        /* Check for an overwrite */
 183        WARN_ON_ONCE(rs && task->reclaim_state);
 184
 185        /* Check for the nulling of an already-nulled member */
 186        WARN_ON_ONCE(!rs && !task->reclaim_state);
 187
 188        task->reclaim_state = rs;
 189}
 190
 191static LIST_HEAD(shrinker_list);
 192static DECLARE_RWSEM(shrinker_rwsem);
 193
 194#ifdef CONFIG_MEMCG
 195static int shrinker_nr_max;
 196
 197/* The shrinker_info is expanded in a batch of BITS_PER_LONG */
 198static inline int shrinker_map_size(int nr_items)
 199{
 200        return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long));
 201}
 202
 203static inline int shrinker_defer_size(int nr_items)
 204{
 205        return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t));
 206}
 207
 208static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
 209                                                     int nid)
 210{
 211        return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info,
 212                                         lockdep_is_held(&shrinker_rwsem));
 213}
 214
 215static int expand_one_shrinker_info(struct mem_cgroup *memcg,
 216                                    int map_size, int defer_size,
 217                                    int old_map_size, int old_defer_size)
 218{
 219        struct shrinker_info *new, *old;
 220        struct mem_cgroup_per_node *pn;
 221        int nid;
 222        int size = map_size + defer_size;
 223
 224        for_each_node(nid) {
 225                pn = memcg->nodeinfo[nid];
 226                old = shrinker_info_protected(memcg, nid);
 227                /* Not yet online memcg */
 228                if (!old)
 229                        return 0;
 230
 231                new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
 232                if (!new)
 233                        return -ENOMEM;
 234
 235                new->nr_deferred = (atomic_long_t *)(new + 1);
 236                new->map = (void *)new->nr_deferred + defer_size;
 237
 238                /* map: set all old bits, clear all new bits */
 239                memset(new->map, (int)0xff, old_map_size);
 240                memset((void *)new->map + old_map_size, 0, map_size - old_map_size);
 241                /* nr_deferred: copy old values, clear all new values */
 242                memcpy(new->nr_deferred, old->nr_deferred, old_defer_size);
 243                memset((void *)new->nr_deferred + old_defer_size, 0,
 244                       defer_size - old_defer_size);
 245
 246                rcu_assign_pointer(pn->shrinker_info, new);
 247                kvfree_rcu(old, rcu);
 248        }
 249
 250        return 0;
 251}
 252
 253void free_shrinker_info(struct mem_cgroup *memcg)
 254{
 255        struct mem_cgroup_per_node *pn;
 256        struct shrinker_info *info;
 257        int nid;
 258
 259        for_each_node(nid) {
 260                pn = memcg->nodeinfo[nid];
 261                info = rcu_dereference_protected(pn->shrinker_info, true);
 262                kvfree(info);
 263                rcu_assign_pointer(pn->shrinker_info, NULL);
 264        }
 265}
 266
 267int alloc_shrinker_info(struct mem_cgroup *memcg)
 268{
 269        struct shrinker_info *info;
 270        int nid, size, ret = 0;
 271        int map_size, defer_size = 0;
 272
 273        down_write(&shrinker_rwsem);
 274        map_size = shrinker_map_size(shrinker_nr_max);
 275        defer_size = shrinker_defer_size(shrinker_nr_max);
 276        size = map_size + defer_size;
 277        for_each_node(nid) {
 278                info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid);
 279                if (!info) {
 280                        free_shrinker_info(memcg);
 281                        ret = -ENOMEM;
 282                        break;
 283                }
 284                info->nr_deferred = (atomic_long_t *)(info + 1);
 285                info->map = (void *)info->nr_deferred + defer_size;
 286                rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
 287        }
 288        up_write(&shrinker_rwsem);
 289
 290        return ret;
 291}
 292
 293static inline bool need_expand(int nr_max)
 294{
 295        return round_up(nr_max, BITS_PER_LONG) >
 296               round_up(shrinker_nr_max, BITS_PER_LONG);
 297}
 298
 299static int expand_shrinker_info(int new_id)
 300{
 301        int ret = 0;
 302        int new_nr_max = new_id + 1;
 303        int map_size, defer_size = 0;
 304        int old_map_size, old_defer_size = 0;
 305        struct mem_cgroup *memcg;
 306
 307        if (!need_expand(new_nr_max))
 308                goto out;
 309
 310        if (!root_mem_cgroup)
 311                goto out;
 312
 313        lockdep_assert_held(&shrinker_rwsem);
 314
 315        map_size = shrinker_map_size(new_nr_max);
 316        defer_size = shrinker_defer_size(new_nr_max);
 317        old_map_size = shrinker_map_size(shrinker_nr_max);
 318        old_defer_size = shrinker_defer_size(shrinker_nr_max);
 319
 320        memcg = mem_cgroup_iter(NULL, NULL, NULL);
 321        do {
 322                ret = expand_one_shrinker_info(memcg, map_size, defer_size,
 323                                               old_map_size, old_defer_size);
 324                if (ret) {
 325                        mem_cgroup_iter_break(NULL, memcg);
 326                        goto out;
 327                }
 328        } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
 329out:
 330        if (!ret)
 331                shrinker_nr_max = new_nr_max;
 332
 333        return ret;
 334}
 335
 336void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
 337{
 338        if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
 339                struct shrinker_info *info;
 340
 341                rcu_read_lock();
 342                info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
 343                /* Pairs with smp mb in shrink_slab() */
 344                smp_mb__before_atomic();
 345                set_bit(shrinker_id, info->map);
 346                rcu_read_unlock();
 347        }
 348}
 349
 350static DEFINE_IDR(shrinker_idr);
 351
 352static int prealloc_memcg_shrinker(struct shrinker *shrinker)
 353{
 354        int id, ret = -ENOMEM;
 355
 356        if (mem_cgroup_disabled())
 357                return -ENOSYS;
 358
 359        down_write(&shrinker_rwsem);
 360        /* This may call shrinker, so it must use down_read_trylock() */
 361        id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
 362        if (id < 0)
 363                goto unlock;
 364
 365        if (id >= shrinker_nr_max) {
 366                if (expand_shrinker_info(id)) {
 367                        idr_remove(&shrinker_idr, id);
 368                        goto unlock;
 369                }
 370        }
 371        shrinker->id = id;
 372        ret = 0;
 373unlock:
 374        up_write(&shrinker_rwsem);
 375        return ret;
 376}
 377
 378static void unregister_memcg_shrinker(struct shrinker *shrinker)
 379{
 380        int id = shrinker->id;
 381
 382        BUG_ON(id < 0);
 383
 384        lockdep_assert_held(&shrinker_rwsem);
 385
 386        idr_remove(&shrinker_idr, id);
 387}
 388
 389static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
 390                                   struct mem_cgroup *memcg)
 391{
 392        struct shrinker_info *info;
 393
 394        info = shrinker_info_protected(memcg, nid);
 395        return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0);
 396}
 397
 398static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
 399                                  struct mem_cgroup *memcg)
 400{
 401        struct shrinker_info *info;
 402
 403        info = shrinker_info_protected(memcg, nid);
 404        return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]);
 405}
 406
 407void reparent_shrinker_deferred(struct mem_cgroup *memcg)
 408{
 409        int i, nid;
 410        long nr;
 411        struct mem_cgroup *parent;
 412        struct shrinker_info *child_info, *parent_info;
 413
 414        parent = parent_mem_cgroup(memcg);
 415        if (!parent)
 416                parent = root_mem_cgroup;
 417
 418        /* Prevent from concurrent shrinker_info expand */
 419        down_read(&shrinker_rwsem);
 420        for_each_node(nid) {
 421                child_info = shrinker_info_protected(memcg, nid);
 422                parent_info = shrinker_info_protected(parent, nid);
 423                for (i = 0; i < shrinker_nr_max; i++) {
 424                        nr = atomic_long_read(&child_info->nr_deferred[i]);
 425                        atomic_long_add(nr, &parent_info->nr_deferred[i]);
 426                }
 427        }
 428        up_read(&shrinker_rwsem);
 429}
 430
 431static bool cgroup_reclaim(struct scan_control *sc)
 432{
 433        return sc->target_mem_cgroup;
 434}
 435
 436/**
 437 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
 438 * @sc: scan_control in question
 439 *
 440 * The normal page dirty throttling mechanism in balance_dirty_pages() is
 441 * completely broken with the legacy memcg and direct stalling in
 442 * shrink_page_list() is used for throttling instead, which lacks all the
 443 * niceties such as fairness, adaptive pausing, bandwidth proportional
 444 * allocation and configurability.
 445 *
 446 * This function tests whether the vmscan currently in progress can assume
 447 * that the normal dirty throttling mechanism is operational.
 448 */
 449static bool writeback_throttling_sane(struct scan_control *sc)
 450{
 451        if (!cgroup_reclaim(sc))
 452                return true;
 453#ifdef CONFIG_CGROUP_WRITEBACK
 454        if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
 455                return true;
 456#endif
 457        return false;
 458}
 459#else
 460static int prealloc_memcg_shrinker(struct shrinker *shrinker)
 461{
 462        return -ENOSYS;
 463}
 464
 465static void unregister_memcg_shrinker(struct shrinker *shrinker)
 466{
 467}
 468
 469static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
 470                                   struct mem_cgroup *memcg)
 471{
 472        return 0;
 473}
 474
 475static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
 476                                  struct mem_cgroup *memcg)
 477{
 478        return 0;
 479}
 480
 481static bool cgroup_reclaim(struct scan_control *sc)
 482{
 483        return false;
 484}
 485
 486static bool writeback_throttling_sane(struct scan_control *sc)
 487{
 488        return true;
 489}
 490#endif
 491
 492static long xchg_nr_deferred(struct shrinker *shrinker,
 493                             struct shrink_control *sc)
 494{
 495        int nid = sc->nid;
 496
 497        if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
 498                nid = 0;
 499
 500        if (sc->memcg &&
 501            (shrinker->flags & SHRINKER_MEMCG_AWARE))
 502                return xchg_nr_deferred_memcg(nid, shrinker,
 503                                              sc->memcg);
 504
 505        return atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
 506}
 507
 508
 509static long add_nr_deferred(long nr, struct shrinker *shrinker,
 510                            struct shrink_control *sc)
 511{
 512        int nid = sc->nid;
 513
 514        if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
 515                nid = 0;
 516
 517        if (sc->memcg &&
 518            (shrinker->flags & SHRINKER_MEMCG_AWARE))
 519                return add_nr_deferred_memcg(nr, nid, shrinker,
 520                                             sc->memcg);
 521
 522        return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]);
 523}
 524
 525static bool can_demote(int nid, struct scan_control *sc)
 526{
 527        if (!numa_demotion_enabled)
 528                return false;
 529        if (sc) {
 530                if (sc->no_demotion)
 531                        return false;
 532                /* It is pointless to do demotion in memcg reclaim */
 533                if (cgroup_reclaim(sc))
 534                        return false;
 535        }
 536        if (next_demotion_node(nid) == NUMA_NO_NODE)
 537                return false;
 538
 539        return true;
 540}
 541
 542static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
 543                                          int nid,
 544                                          struct scan_control *sc)
 545{
 546        if (memcg == NULL) {
 547                /*
 548                 * For non-memcg reclaim, is there
 549                 * space in any swap device?
 550                 */
 551                if (get_nr_swap_pages() > 0)
 552                        return true;
 553        } else {
 554                /* Is the memcg below its swap limit? */
 555                if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
 556                        return true;
 557        }
 558
 559        /*
 560         * The page can not be swapped.
 561         *
 562         * Can it be reclaimed from this node via demotion?
 563         */
 564        return can_demote(nid, sc);
 565}
 566
 567/*
 568 * This misses isolated pages which are not accounted for to save counters.
 569 * As the data only determines if reclaim or compaction continues, it is
 570 * not expected that isolated pages will be a dominating factor.
 571 */
 572unsigned long zone_reclaimable_pages(struct zone *zone)
 573{
 574        unsigned long nr;
 575
 576        nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
 577                zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
 578        if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
 579                nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
 580                        zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
 581
 582        return nr;
 583}
 584
 585/**
 586 * lruvec_lru_size -  Returns the number of pages on the given LRU list.
 587 * @lruvec: lru vector
 588 * @lru: lru to use
 589 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
 590 */
 591static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
 592                                     int zone_idx)
 593{
 594        unsigned long size = 0;
 595        int zid;
 596
 597        for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
 598                struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
 599
 600                if (!managed_zone(zone))
 601                        continue;
 602
 603                if (!mem_cgroup_disabled())
 604                        size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
 605                else
 606                        size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
 607        }
 608        return size;
 609}
 610
 611/*
 612 * Add a shrinker callback to be called from the vm.
 613 */
 614int prealloc_shrinker(struct shrinker *shrinker)
 615{
 616        unsigned int size;
 617        int err;
 618
 619        if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
 620                err = prealloc_memcg_shrinker(shrinker);
 621                if (err != -ENOSYS)
 622                        return err;
 623
 624                shrinker->flags &= ~SHRINKER_MEMCG_AWARE;
 625        }
 626
 627        size = sizeof(*shrinker->nr_deferred);
 628        if (shrinker->flags & SHRINKER_NUMA_AWARE)
 629                size *= nr_node_ids;
 630
 631        shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
 632        if (!shrinker->nr_deferred)
 633                return -ENOMEM;
 634
 635        return 0;
 636}
 637
 638void free_prealloced_shrinker(struct shrinker *shrinker)
 639{
 640        if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
 641                down_write(&shrinker_rwsem);
 642                unregister_memcg_shrinker(shrinker);
 643                up_write(&shrinker_rwsem);
 644                return;
 645        }
 646
 647        kfree(shrinker->nr_deferred);
 648        shrinker->nr_deferred = NULL;
 649}
 650
 651void register_shrinker_prepared(struct shrinker *shrinker)
 652{
 653        down_write(&shrinker_rwsem);
 654        list_add_tail(&shrinker->list, &shrinker_list);
 655        shrinker->flags |= SHRINKER_REGISTERED;
 656        up_write(&shrinker_rwsem);
 657}
 658
 659int register_shrinker(struct shrinker *shrinker)
 660{
 661        int err = prealloc_shrinker(shrinker);
 662
 663        if (err)
 664                return err;
 665        register_shrinker_prepared(shrinker);
 666        return 0;
 667}
 668EXPORT_SYMBOL(register_shrinker);
 669
 670/*
 671 * Remove one
 672 */
 673void unregister_shrinker(struct shrinker *shrinker)
 674{
 675        if (!(shrinker->flags & SHRINKER_REGISTERED))
 676                return;
 677
 678        down_write(&shrinker_rwsem);
 679        list_del(&shrinker->list);
 680        shrinker->flags &= ~SHRINKER_REGISTERED;
 681        if (shrinker->flags & SHRINKER_MEMCG_AWARE)
 682                unregister_memcg_shrinker(shrinker);
 683        up_write(&shrinker_rwsem);
 684
 685        kfree(shrinker->nr_deferred);
 686        shrinker->nr_deferred = NULL;
 687}
 688EXPORT_SYMBOL(unregister_shrinker);
 689
 690#define SHRINK_BATCH 128
 691
 692static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
 693                                    struct shrinker *shrinker, int priority)
 694{
 695        unsigned long freed = 0;
 696        unsigned long long delta;
 697        long total_scan;
 698        long freeable;
 699        long nr;
 700        long new_nr;
 701        long batch_size = shrinker->batch ? shrinker->batch
 702                                          : SHRINK_BATCH;
 703        long scanned = 0, next_deferred;
 704
 705        freeable = shrinker->count_objects(shrinker, shrinkctl);
 706        if (freeable == 0 || freeable == SHRINK_EMPTY)
 707                return freeable;
 708
 709        /*
 710         * copy the current shrinker scan count into a local variable
 711         * and zero it so that other concurrent shrinker invocations
 712         * don't also do this scanning work.
 713         */
 714        nr = xchg_nr_deferred(shrinker, shrinkctl);
 715
 716        if (shrinker->seeks) {
 717                delta = freeable >> priority;
 718                delta *= 4;
 719                do_div(delta, shrinker->seeks);
 720        } else {
 721                /*
 722                 * These objects don't require any IO to create. Trim
 723                 * them aggressively under memory pressure to keep
 724                 * them from causing refetches in the IO caches.
 725                 */
 726                delta = freeable / 2;
 727        }
 728
 729        total_scan = nr >> priority;
 730        total_scan += delta;
 731        total_scan = min(total_scan, (2 * freeable));
 732
 733        trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
 734                                   freeable, delta, total_scan, priority);
 735
 736        /*
 737         * Normally, we should not scan less than batch_size objects in one
 738         * pass to avoid too frequent shrinker calls, but if the slab has less
 739         * than batch_size objects in total and we are really tight on memory,
 740         * we will try to reclaim all available objects, otherwise we can end
 741         * up failing allocations although there are plenty of reclaimable
 742         * objects spread over several slabs with usage less than the
 743         * batch_size.
 744         *
 745         * We detect the "tight on memory" situations by looking at the total
 746         * number of objects we want to scan (total_scan). If it is greater
 747         * than the total number of objects on slab (freeable), we must be
 748         * scanning at high prio and therefore should try to reclaim as much as
 749         * possible.
 750         */
 751        while (total_scan >= batch_size ||
 752               total_scan >= freeable) {
 753                unsigned long ret;
 754                unsigned long nr_to_scan = min(batch_size, total_scan);
 755
 756                shrinkctl->nr_to_scan = nr_to_scan;
 757                shrinkctl->nr_scanned = nr_to_scan;
 758                ret = shrinker->scan_objects(shrinker, shrinkctl);
 759                if (ret == SHRINK_STOP)
 760                        break;
 761                freed += ret;
 762
 763                count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
 764                total_scan -= shrinkctl->nr_scanned;
 765                scanned += shrinkctl->nr_scanned;
 766
 767                cond_resched();
 768        }
 769
 770        /*
 771         * The deferred work is increased by any new work (delta) that wasn't
 772         * done, decreased by old deferred work that was done now.
 773         *
 774         * And it is capped to two times of the freeable items.
 775         */
 776        next_deferred = max_t(long, (nr + delta - scanned), 0);
 777        next_deferred = min(next_deferred, (2 * freeable));
 778
 779        /*
 780         * move the unused scan count back into the shrinker in a
 781         * manner that handles concurrent updates.
 782         */
 783        new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
 784
 785        trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
 786        return freed;
 787}
 788
 789#ifdef CONFIG_MEMCG
 790static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
 791                        struct mem_cgroup *memcg, int priority)
 792{
 793        struct shrinker_info *info;
 794        unsigned long ret, freed = 0;
 795        int i;
 796
 797        if (!mem_cgroup_online(memcg))
 798                return 0;
 799
 800        if (!down_read_trylock(&shrinker_rwsem))
 801                return 0;
 802
 803        info = shrinker_info_protected(memcg, nid);
 804        if (unlikely(!info))
 805                goto unlock;
 806
 807        for_each_set_bit(i, info->map, shrinker_nr_max) {
 808                struct shrink_control sc = {
 809                        .gfp_mask = gfp_mask,
 810                        .nid = nid,
 811                        .memcg = memcg,
 812                };
 813                struct shrinker *shrinker;
 814
 815                shrinker = idr_find(&shrinker_idr, i);
 816                if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
 817                        if (!shrinker)
 818                                clear_bit(i, info->map);
 819                        continue;
 820                }
 821
 822                /* Call non-slab shrinkers even though kmem is disabled */
 823                if (!memcg_kmem_enabled() &&
 824                    !(shrinker->flags & SHRINKER_NONSLAB))
 825                        continue;
 826
 827                ret = do_shrink_slab(&sc, shrinker, priority);
 828                if (ret == SHRINK_EMPTY) {
 829                        clear_bit(i, info->map);
 830                        /*
 831                         * After the shrinker reported that it had no objects to
 832                         * free, but before we cleared the corresponding bit in
 833                         * the memcg shrinker map, a new object might have been
 834                         * added. To make sure, we have the bit set in this
 835                         * case, we invoke the shrinker one more time and reset
 836                         * the bit if it reports that it is not empty anymore.
 837                         * The memory barrier here pairs with the barrier in
 838                         * set_shrinker_bit():
 839                         *
 840                         * list_lru_add()     shrink_slab_memcg()
 841                         *   list_add_tail()    clear_bit()
 842                         *   <MB>               <MB>
 843                         *   set_bit()          do_shrink_slab()
 844                         */
 845                        smp_mb__after_atomic();
 846                        ret = do_shrink_slab(&sc, shrinker, priority);
 847                        if (ret == SHRINK_EMPTY)
 848                                ret = 0;
 849                        else
 850                                set_shrinker_bit(memcg, nid, i);
 851                }
 852                freed += ret;
 853
 854                if (rwsem_is_contended(&shrinker_rwsem)) {
 855                        freed = freed ? : 1;
 856                        break;
 857                }
 858        }
 859unlock:
 860        up_read(&shrinker_rwsem);
 861        return freed;
 862}
 863#else /* CONFIG_MEMCG */
 864static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
 865                        struct mem_cgroup *memcg, int priority)
 866{
 867        return 0;
 868}
 869#endif /* CONFIG_MEMCG */
 870
 871/**
 872 * shrink_slab - shrink slab caches
 873 * @gfp_mask: allocation context
 874 * @nid: node whose slab caches to target
 875 * @memcg: memory cgroup whose slab caches to target
 876 * @priority: the reclaim priority
 877 *
 878 * Call the shrink functions to age shrinkable caches.
 879 *
 880 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
 881 * unaware shrinkers will receive a node id of 0 instead.
 882 *
 883 * @memcg specifies the memory cgroup to target. Unaware shrinkers
 884 * are called only if it is the root cgroup.
 885 *
 886 * @priority is sc->priority, we take the number of objects and >> by priority
 887 * in order to get the scan target.
 888 *
 889 * Returns the number of reclaimed slab objects.
 890 */
 891static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
 892                                 struct mem_cgroup *memcg,
 893                                 int priority)
 894{
 895        unsigned long ret, freed = 0;
 896        struct shrinker *shrinker;
 897
 898        /*
 899         * The root memcg might be allocated even though memcg is disabled
 900         * via "cgroup_disable=memory" boot parameter.  This could make
 901         * mem_cgroup_is_root() return false, then just run memcg slab
 902         * shrink, but skip global shrink.  This may result in premature
 903         * oom.
 904         */
 905        if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
 906                return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
 907
 908        if (!down_read_trylock(&shrinker_rwsem))
 909                goto out;
 910
 911        list_for_each_entry(shrinker, &shrinker_list, list) {
 912                struct shrink_control sc = {
 913                        .gfp_mask = gfp_mask,
 914                        .nid = nid,
 915                        .memcg = memcg,
 916                };
 917
 918                ret = do_shrink_slab(&sc, shrinker, priority);
 919                if (ret == SHRINK_EMPTY)
 920                        ret = 0;
 921                freed += ret;
 922                /*
 923                 * Bail out if someone want to register a new shrinker to
 924                 * prevent the registration from being stalled for long periods
 925                 * by parallel ongoing shrinking.
 926                 */
 927                if (rwsem_is_contended(&shrinker_rwsem)) {
 928                        freed = freed ? : 1;
 929                        break;
 930                }
 931        }
 932
 933        up_read(&shrinker_rwsem);
 934out:
 935        cond_resched();
 936        return freed;
 937}
 938
 939void drop_slab_node(int nid)
 940{
 941        unsigned long freed;
 942        int shift = 0;
 943
 944        do {
 945                struct mem_cgroup *memcg = NULL;
 946
 947                if (fatal_signal_pending(current))
 948                        return;
 949
 950                freed = 0;
 951                memcg = mem_cgroup_iter(NULL, NULL, NULL);
 952                do {
 953                        freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
 954                } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
 955        } while ((freed >> shift++) > 1);
 956}
 957
 958void drop_slab(void)
 959{
 960        int nid;
 961
 962        for_each_online_node(nid)
 963                drop_slab_node(nid);
 964}
 965
 966static inline int is_page_cache_freeable(struct page *page)
 967{
 968        /*
 969         * A freeable page cache page is referenced only by the caller
 970         * that isolated the page, the page cache and optional buffer
 971         * heads at page->private.
 972         */
 973        int page_cache_pins = thp_nr_pages(page);
 974        return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
 975}
 976
 977static int may_write_to_inode(struct inode *inode)
 978{
 979        if (current->flags & PF_SWAPWRITE)
 980                return 1;
 981        if (!inode_write_congested(inode))
 982                return 1;
 983        if (inode_to_bdi(inode) == current->backing_dev_info)
 984                return 1;
 985        return 0;
 986}
 987
 988/*
 989 * We detected a synchronous write error writing a page out.  Probably
 990 * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 991 * fsync(), msync() or close().
 992 *
 993 * The tricky part is that after writepage we cannot touch the mapping: nothing
 994 * prevents it from being freed up.  But we have a ref on the page and once
 995 * that page is locked, the mapping is pinned.
 996 *
 997 * We're allowed to run sleeping lock_page() here because we know the caller has
 998 * __GFP_FS.
 999 */
1000static void handle_write_error(struct address_space *mapping,
1001                                struct page *page, int error)
1002{
1003        lock_page(page);
1004        if (page_mapping(page) == mapping)
1005                mapping_set_error(mapping, error);
1006        unlock_page(page);
1007}
1008
1009/* possible outcome of pageout() */
1010typedef enum {
1011        /* failed to write page out, page is locked */
1012        PAGE_KEEP,
1013        /* move page to the active list, page is locked */
1014        PAGE_ACTIVATE,
1015        /* page has been sent to the disk successfully, page is unlocked */
1016        PAGE_SUCCESS,
1017        /* page is clean and locked */
1018        PAGE_CLEAN,
1019} pageout_t;
1020
1021/*
1022 * pageout is called by shrink_page_list() for each dirty page.
1023 * Calls ->writepage().
1024 */
1025static pageout_t pageout(struct page *page, struct address_space *mapping)
1026{
1027        /*
1028         * If the page is dirty, only perform writeback if that write
1029         * will be non-blocking.  To prevent this allocation from being
1030         * stalled by pagecache activity.  But note that there may be
1031         * stalls if we need to run get_block().  We could test
1032         * PagePrivate for that.
1033         *
1034         * If this process is currently in __generic_file_write_iter() against
1035         * this page's queue, we can perform writeback even if that
1036         * will block.
1037         *
1038         * If the page is swapcache, write it back even if that would
1039         * block, for some throttling. This happens by accident, because
1040         * swap_backing_dev_info is bust: it doesn't reflect the
1041         * congestion state of the swapdevs.  Easy to fix, if needed.
1042         */
1043        if (!is_page_cache_freeable(page))
1044                return PAGE_KEEP;
1045        if (!mapping) {
1046                /*
1047                 * Some data journaling orphaned pages can have
1048                 * page->mapping == NULL while being dirty with clean buffers.
1049                 */
1050                if (page_has_private(page)) {
1051                        if (try_to_free_buffers(page)) {
1052                                ClearPageDirty(page);
1053                                pr_info("%s: orphaned page\n", __func__);
1054                                return PAGE_CLEAN;
1055                        }
1056                }
1057                return PAGE_KEEP;
1058        }
1059        if (mapping->a_ops->writepage == NULL)
1060                return PAGE_ACTIVATE;
1061        if (!may_write_to_inode(mapping->host))
1062                return PAGE_KEEP;
1063
1064        if (clear_page_dirty_for_io(page)) {
1065                int res;
1066                struct writeback_control wbc = {
1067                        .sync_mode = WB_SYNC_NONE,
1068                        .nr_to_write = SWAP_CLUSTER_MAX,
1069                        .range_start = 0,
1070                        .range_end = LLONG_MAX,
1071                        .for_reclaim = 1,
1072                };
1073
1074                SetPageReclaim(page);
1075                res = mapping->a_ops->writepage(page, &wbc);
1076                if (res < 0)
1077                        handle_write_error(mapping, page, res);
1078                if (res == AOP_WRITEPAGE_ACTIVATE) {
1079                        ClearPageReclaim(page);
1080                        return PAGE_ACTIVATE;
1081                }
1082
1083                if (!PageWriteback(page)) {
1084                        /* synchronous write or broken a_ops? */
1085                        ClearPageReclaim(page);
1086                }
1087                trace_mm_vmscan_writepage(page);
1088                inc_node_page_state(page, NR_VMSCAN_WRITE);
1089                return PAGE_SUCCESS;
1090        }
1091
1092        return PAGE_CLEAN;
1093}
1094
1095/*
1096 * Same as remove_mapping, but if the page is removed from the mapping, it
1097 * gets returned with a refcount of 0.
1098 */
1099static int __remove_mapping(struct address_space *mapping, struct page *page,
1100                            bool reclaimed, struct mem_cgroup *target_memcg)
1101{
1102        int refcount;
1103        void *shadow = NULL;
1104
1105        BUG_ON(!PageLocked(page));
1106        BUG_ON(mapping != page_mapping(page));
1107
1108        xa_lock_irq(&mapping->i_pages);
1109        /*
1110         * The non racy check for a busy page.
1111         *
1112         * Must be careful with the order of the tests. When someone has
1113         * a ref to the page, it may be possible that they dirty it then
1114         * drop the reference. So if PageDirty is tested before page_count
1115         * here, then the following race may occur:
1116         *
1117         * get_user_pages(&page);
1118         * [user mapping goes away]
1119         * write_to(page);
1120         *                              !PageDirty(page)    [good]
1121         * SetPageDirty(page);
1122         * put_page(page);
1123         *                              !page_count(page)   [good, discard it]
1124         *
1125         * [oops, our write_to data is lost]
1126         *
1127         * Reversing the order of the tests ensures such a situation cannot
1128         * escape unnoticed. The smp_rmb is needed to ensure the page->flags
1129         * load is not satisfied before that of page->_refcount.
1130         *
1131         * Note that if SetPageDirty is always performed via set_page_dirty,
1132         * and thus under the i_pages lock, then this ordering is not required.
1133         */
1134        refcount = 1 + compound_nr(page);
1135        if (!page_ref_freeze(page, refcount))
1136                goto cannot_free;
1137        /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
1138        if (unlikely(PageDirty(page))) {
1139                page_ref_unfreeze(page, refcount);
1140                goto cannot_free;
1141        }
1142
1143        if (PageSwapCache(page)) {
1144                swp_entry_t swap = { .val = page_private(page) };
1145                mem_cgroup_swapout(page, swap);
1146                if (reclaimed && !mapping_exiting(mapping))
1147                        shadow = workingset_eviction(page, target_memcg);
1148                __delete_from_swap_cache(page, swap, shadow);
1149                xa_unlock_irq(&mapping->i_pages);
1150                put_swap_page(page, swap);
1151        } else {
1152                void (*freepage)(struct page *);
1153
1154                freepage = mapping->a_ops->freepage;
1155                /*
1156                 * Remember a shadow entry for reclaimed file cache in
1157                 * order to detect refaults, thus thrashing, later on.
1158                 *
1159                 * But don't store shadows in an address space that is
1160                 * already exiting.  This is not just an optimization,
1161                 * inode reclaim needs to empty out the radix tree or
1162                 * the nodes are lost.  Don't plant shadows behind its
1163                 * back.
1164                 *
1165                 * We also don't store shadows for DAX mappings because the
1166                 * only page cache pages found in these are zero pages
1167                 * covering holes, and because we don't want to mix DAX
1168                 * exceptional entries and shadow exceptional entries in the
1169                 * same address_space.
1170                 */
1171                if (reclaimed && page_is_file_lru(page) &&
1172                    !mapping_exiting(mapping) && !dax_mapping(mapping))
1173                        shadow = workingset_eviction(page, target_memcg);
1174                __delete_from_page_cache(page, shadow);
1175                xa_unlock_irq(&mapping->i_pages);
1176
1177                if (freepage != NULL)
1178                        freepage(page);
1179        }
1180
1181        return 1;
1182
1183cannot_free:
1184        xa_unlock_irq(&mapping->i_pages);
1185        return 0;
1186}
1187
1188/*
1189 * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
1190 * someone else has a ref on the page, abort and return 0.  If it was
1191 * successfully detached, return 1.  Assumes the caller has a single ref on
1192 * this page.
1193 */
1194int remove_mapping(struct address_space *mapping, struct page *page)
1195{
1196        if (__remove_mapping(mapping, page, false, NULL)) {
1197                /*
1198                 * Unfreezing the refcount with 1 rather than 2 effectively
1199                 * drops the pagecache ref for us without requiring another
1200                 * atomic operation.
1201                 */
1202                page_ref_unfreeze(page, 1);
1203                return 1;
1204        }
1205        return 0;
1206}
1207
1208/**
1209 * putback_lru_page - put previously isolated page onto appropriate LRU list
1210 * @page: page to be put back to appropriate lru list
1211 *
1212 * Add previously isolated @page to appropriate LRU list.
1213 * Page may still be unevictable for other reasons.
1214 *
1215 * lru_lock must not be held, interrupts must be enabled.
1216 */
1217void putback_lru_page(struct page *page)
1218{
1219        lru_cache_add(page);
1220        put_page(page);         /* drop ref from isolate */
1221}
1222
1223enum page_references {
1224        PAGEREF_RECLAIM,
1225        PAGEREF_RECLAIM_CLEAN,
1226        PAGEREF_KEEP,
1227        PAGEREF_ACTIVATE,
1228};
1229
1230static enum page_references page_check_references(struct page *page,
1231                                                  struct scan_control *sc)
1232{
1233        int referenced_ptes, referenced_page;
1234        unsigned long vm_flags;
1235
1236        referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1237                                          &vm_flags);
1238        referenced_page = TestClearPageReferenced(page);
1239
1240        /*
1241         * Mlock lost the isolation race with us.  Let try_to_unmap()
1242         * move the page to the unevictable list.
1243         */
1244        if (vm_flags & VM_LOCKED)
1245                return PAGEREF_RECLAIM;
1246
1247        if (referenced_ptes) {
1248                /*
1249                 * All mapped pages start out with page table
1250                 * references from the instantiating fault, so we need
1251                 * to look twice if a mapped file page is used more
1252                 * than once.
1253                 *
1254                 * Mark it and spare it for another trip around the
1255                 * inactive list.  Another page table reference will
1256                 * lead to its activation.
1257                 *
1258                 * Note: the mark is set for activated pages as well
1259                 * so that recently deactivated but used pages are
1260                 * quickly recovered.
1261                 */
1262                SetPageReferenced(page);
1263
1264                if (referenced_page || referenced_ptes > 1)
1265                        return PAGEREF_ACTIVATE;
1266
1267                /*
1268                 * Activate file-backed executable pages after first usage.
1269                 */
1270                if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
1271                        return PAGEREF_ACTIVATE;
1272
1273                return PAGEREF_KEEP;
1274        }
1275
1276        /* Reclaim if clean, defer dirty pages to writeback */
1277        if (referenced_page && !PageSwapBacked(page))
1278                return PAGEREF_RECLAIM_CLEAN;
1279
1280        return PAGEREF_RECLAIM;
1281}
1282
1283/* Check if a page is dirty or under writeback */
1284static void page_check_dirty_writeback(struct page *page,
1285                                       bool *dirty, bool *writeback)
1286{
1287        struct address_space *mapping;
1288
1289        /*
1290         * Anonymous pages are not handled by flushers and must be written
1291         * from reclaim context. Do not stall reclaim based on them
1292         */
1293        if (!page_is_file_lru(page) ||
1294            (PageAnon(page) && !PageSwapBacked(page))) {
1295                *dirty = false;
1296                *writeback = false;
1297                return;
1298        }
1299
1300        /* By default assume that the page flags are accurate */
1301        *dirty = PageDirty(page);
1302        *writeback = PageWriteback(page);
1303
1304        /* Verify dirty/writeback state if the filesystem supports it */
1305        if (!page_has_private(page))
1306                return;
1307
1308        mapping = page_mapping(page);
1309        if (mapping && mapping->a_ops->is_dirty_writeback)
1310                mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1311}
1312
1313static struct page *alloc_demote_page(struct page *page, unsigned long node)
1314{
1315        struct migration_target_control mtc = {
1316                /*
1317                 * Allocate from 'node', or fail quickly and quietly.
1318                 * When this happens, 'page' will likely just be discarded
1319                 * instead of migrated.
1320                 */
1321                .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) |
1322                            __GFP_THISNODE  | __GFP_NOWARN |
1323                            __GFP_NOMEMALLOC | GFP_NOWAIT,
1324                .nid = node
1325        };
1326
1327        return alloc_migration_target(page, (unsigned long)&mtc);
1328}
1329
1330/*
1331 * Take pages on @demote_list and attempt to demote them to
1332 * another node.  Pages which are not demoted are left on
1333 * @demote_pages.
1334 */
1335static unsigned int demote_page_list(struct list_head *demote_pages,
1336                                     struct pglist_data *pgdat)
1337{
1338        int target_nid = next_demotion_node(pgdat->node_id);
1339        unsigned int nr_succeeded;
1340        int err;
1341
1342        if (list_empty(demote_pages))
1343                return 0;
1344
1345        if (target_nid == NUMA_NO_NODE)
1346                return 0;
1347
1348        /* Demotion ignores all cpuset and mempolicy settings */
1349        err = migrate_pages(demote_pages, alloc_demote_page, NULL,
1350                            target_nid, MIGRATE_ASYNC, MR_DEMOTION,
1351                            &nr_succeeded);
1352
1353        if (current_is_kswapd())
1354                __count_vm_events(PGDEMOTE_KSWAPD, nr_succeeded);
1355        else
1356                __count_vm_events(PGDEMOTE_DIRECT, nr_succeeded);
1357
1358        return nr_succeeded;
1359}
1360
1361/*
1362 * shrink_page_list() returns the number of reclaimed pages
1363 */
1364static unsigned int shrink_page_list(struct list_head *page_list,
1365                                     struct pglist_data *pgdat,
1366                                     struct scan_control *sc,
1367                                     struct reclaim_stat *stat,
1368                                     bool ignore_references)
1369{
1370        LIST_HEAD(ret_pages);
1371        LIST_HEAD(free_pages);
1372        LIST_HEAD(demote_pages);
1373        unsigned int nr_reclaimed = 0;
1374        unsigned int pgactivate = 0;
1375        bool do_demote_pass;
1376
1377        memset(stat, 0, sizeof(*stat));
1378        cond_resched();
1379        do_demote_pass = can_demote(pgdat->node_id, sc);
1380
1381retry:
1382        while (!list_empty(page_list)) {
1383                struct address_space *mapping;
1384                struct page *page;
1385                enum page_references references = PAGEREF_RECLAIM;
1386                bool dirty, writeback, may_enter_fs;
1387                unsigned int nr_pages;
1388
1389                cond_resched();
1390
1391                page = lru_to_page(page_list);
1392                list_del(&page->lru);
1393
1394                if (!trylock_page(page))
1395                        goto keep;
1396
1397                VM_BUG_ON_PAGE(PageActive(page), page);
1398
1399                nr_pages = compound_nr(page);
1400
1401                /* Account the number of base pages even though THP */
1402                sc->nr_scanned += nr_pages;
1403
1404                if (unlikely(!page_evictable(page)))
1405                        goto activate_locked;
1406
1407                if (!sc->may_unmap && page_mapped(page))
1408                        goto keep_locked;
1409
1410                may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1411                        (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1412
1413                /*
1414                 * The number of dirty pages determines if a node is marked
1415                 * reclaim_congested which affects wait_iff_congested. kswapd
1416                 * will stall and start writing pages if the tail of the LRU
1417                 * is all dirty unqueued pages.
1418                 */
1419                page_check_dirty_writeback(page, &dirty, &writeback);
1420                if (dirty || writeback)
1421                        stat->nr_dirty++;
1422
1423                if (dirty && !writeback)
1424                        stat->nr_unqueued_dirty++;
1425
1426                /*
1427                 * Treat this page as congested if the underlying BDI is or if
1428                 * pages are cycling through the LRU so quickly that the
1429                 * pages marked for immediate reclaim are making it to the
1430                 * end of the LRU a second time.
1431                 */
1432                mapping = page_mapping(page);
1433                if (((dirty || writeback) && mapping &&
1434                     inode_write_congested(mapping->host)) ||
1435                    (writeback && PageReclaim(page)))
1436                        stat->nr_congested++;
1437
1438                /*
1439                 * If a page at the tail of the LRU is under writeback, there
1440                 * are three cases to consider.
1441                 *
1442                 * 1) If reclaim is encountering an excessive number of pages
1443                 *    under writeback and this page is both under writeback and
1444                 *    PageReclaim then it indicates that pages are being queued
1445                 *    for IO but are being recycled through the LRU before the
1446                 *    IO can complete. Waiting on the page itself risks an
1447                 *    indefinite stall if it is impossible to writeback the
1448                 *    page due to IO error or disconnected storage so instead
1449                 *    note that the LRU is being scanned too quickly and the
1450                 *    caller can stall after page list has been processed.
1451                 *
1452                 * 2) Global or new memcg reclaim encounters a page that is
1453                 *    not marked for immediate reclaim, or the caller does not
1454                 *    have __GFP_FS (or __GFP_IO if it's simply going to swap,
1455                 *    not to fs). In this case mark the page for immediate
1456                 *    reclaim and continue scanning.
1457                 *
1458                 *    Require may_enter_fs because we would wait on fs, which
1459                 *    may not have submitted IO yet. And the loop driver might
1460                 *    enter reclaim, and deadlock if it waits on a page for
1461                 *    which it is needed to do the write (loop masks off
1462                 *    __GFP_IO|__GFP_FS for this reason); but more thought
1463                 *    would probably show more reasons.
1464                 *
1465                 * 3) Legacy memcg encounters a page that is already marked
1466                 *    PageReclaim. memcg does not have any dirty pages
1467                 *    throttling so we could easily OOM just because too many
1468                 *    pages are in writeback and there is nothing else to
1469                 *    reclaim. Wait for the writeback to complete.
1470                 *
1471                 * In cases 1) and 2) we activate the pages to get them out of
1472                 * the way while we continue scanning for clean pages on the
1473                 * inactive list and refilling from the active list. The
1474                 * observation here is that waiting for disk writes is more
1475                 * expensive than potentially causing reloads down the line.
1476                 * Since they're marked for immediate reclaim, they won't put
1477                 * memory pressure on the cache working set any longer than it
1478                 * takes to write them to disk.
1479                 */
1480                if (PageWriteback(page)) {
1481                        /* Case 1 above */
1482                        if (current_is_kswapd() &&
1483                            PageReclaim(page) &&
1484                            test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1485                                stat->nr_immediate++;
1486                                goto activate_locked;
1487
1488                        /* Case 2 above */
1489                        } else if (writeback_throttling_sane(sc) ||
1490                            !PageReclaim(page) || !may_enter_fs) {
1491                                /*
1492                                 * This is slightly racy - end_page_writeback()
1493                                 * might have just cleared PageReclaim, then
1494                                 * setting PageReclaim here end up interpreted
1495                                 * as PageReadahead - but that does not matter
1496                                 * enough to care.  What we do want is for this
1497                                 * page to have PageReclaim set next time memcg
1498                                 * reclaim reaches the tests above, so it will
1499                                 * then wait_on_page_writeback() to avoid OOM;
1500                                 * and it's also appropriate in global reclaim.
1501                                 */
1502                                SetPageReclaim(page);
1503                                stat->nr_writeback++;
1504                                goto activate_locked;
1505
1506                        /* Case 3 above */
1507                        } else {
1508                                unlock_page(page);
1509                                wait_on_page_writeback(page);
1510                                /* then go back and try same page again */
1511                                list_add_tail(&page->lru, page_list);
1512                                continue;
1513                        }
1514                }
1515
1516                if (!ignore_references)
1517                        references = page_check_references(page, sc);
1518
1519                switch (references) {
1520                case PAGEREF_ACTIVATE:
1521                        goto activate_locked;
1522                case PAGEREF_KEEP:
1523                        stat->nr_ref_keep += nr_pages;
1524                        goto keep_locked;
1525                case PAGEREF_RECLAIM:
1526                case PAGEREF_RECLAIM_CLEAN:
1527                        ; /* try to reclaim the page below */
1528                }
1529
1530                /*
1531                 * Before reclaiming the page, try to relocate
1532                 * its contents to another node.
1533                 */
1534                if (do_demote_pass &&
1535                    (thp_migration_supported() || !PageTransHuge(page))) {
1536                        list_add(&page->lru, &demote_pages);
1537                        unlock_page(page);
1538                        continue;
1539                }
1540
1541                /*
1542                 * Anonymous process memory has backing store?
1543                 * Try to allocate it some swap space here.
1544                 * Lazyfree page could be freed directly
1545                 */
1546                if (PageAnon(page) && PageSwapBacked(page)) {
1547                        if (!PageSwapCache(page)) {
1548                                if (!(sc->gfp_mask & __GFP_IO))
1549                                        goto keep_locked;
1550                                if (page_maybe_dma_pinned(page))
1551                                        goto keep_locked;
1552                                if (PageTransHuge(page)) {
1553                                        /* cannot split THP, skip it */
1554                                        if (!can_split_huge_page(page, NULL))
1555                                                goto activate_locked;
1556                                        /*
1557                                         * Split pages without a PMD map right
1558                                         * away. Chances are some or all of the
1559                                         * tail pages can be freed without IO.
1560                                         */
1561                                        if (!compound_mapcount(page) &&
1562                                            split_huge_page_to_list(page,
1563                                                                    page_list))
1564                                                goto activate_locked;
1565                                }
1566                                if (!add_to_swap(page)) {
1567                                        if (!PageTransHuge(page))
1568                                                goto activate_locked_split;
1569                                        /* Fallback to swap normal pages */
1570                                        if (split_huge_page_to_list(page,
1571                                                                    page_list))
1572                                                goto activate_locked;
1573#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1574                                        count_vm_event(THP_SWPOUT_FALLBACK);
1575#endif
1576                                        if (!add_to_swap(page))
1577                                                goto activate_locked_split;
1578                                }
1579
1580                                may_enter_fs = true;
1581
1582                                /* Adding to swap updated mapping */
1583                                mapping = page_mapping(page);
1584                        }
1585                } else if (unlikely(PageTransHuge(page))) {
1586                        /* Split file THP */
1587                        if (split_huge_page_to_list(page, page_list))
1588                                goto keep_locked;
1589                }
1590
1591                /*
1592                 * THP may get split above, need minus tail pages and update
1593                 * nr_pages to avoid accounting tail pages twice.
1594                 *
1595                 * The tail pages that are added into swap cache successfully
1596                 * reach here.
1597                 */
1598                if ((nr_pages > 1) && !PageTransHuge(page)) {
1599                        sc->nr_scanned -= (nr_pages - 1);
1600                        nr_pages = 1;
1601                }
1602
1603                /*
1604                 * The page is mapped into the page tables of one or more
1605                 * processes. Try to unmap it here.
1606                 */
1607                if (page_mapped(page)) {
1608                        enum ttu_flags flags = TTU_BATCH_FLUSH;
1609                        bool was_swapbacked = PageSwapBacked(page);
1610
1611                        if (unlikely(PageTransHuge(page)))
1612                                flags |= TTU_SPLIT_HUGE_PMD;
1613
1614                        try_to_unmap(page, flags);
1615                        if (page_mapped(page)) {
1616                                stat->nr_unmap_fail += nr_pages;
1617                                if (!was_swapbacked && PageSwapBacked(page))
1618                                        stat->nr_lazyfree_fail += nr_pages;
1619                                goto activate_locked;
1620                        }
1621                }
1622
1623                if (PageDirty(page)) {
1624                        /*
1625                         * Only kswapd can writeback filesystem pages
1626                         * to avoid risk of stack overflow. But avoid
1627                         * injecting inefficient single-page IO into
1628                         * flusher writeback as much as possible: only
1629                         * write pages when we've encountered many
1630                         * dirty pages, and when we've already scanned
1631                         * the rest of the LRU for clean pages and see
1632                         * the same dirty pages again (PageReclaim).
1633                         */
1634                        if (page_is_file_lru(page) &&
1635                            (!current_is_kswapd() || !PageReclaim(page) ||
1636                             !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1637                                /*
1638                                 * Immediately reclaim when written back.
1639                                 * Similar in principal to deactivate_page()
1640                                 * except we already have the page isolated
1641                                 * and know it's dirty
1642                                 */
1643                                inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1644                                SetPageReclaim(page);
1645
1646                                goto activate_locked;
1647                        }
1648
1649                        if (references == PAGEREF_RECLAIM_CLEAN)
1650                                goto keep_locked;
1651                        if (!may_enter_fs)
1652                                goto keep_locked;
1653                        if (!sc->may_writepage)
1654                                goto keep_locked;
1655
1656                        /*
1657                         * Page is dirty. Flush the TLB if a writable entry
1658                         * potentially exists to avoid CPU writes after IO
1659                         * starts and then write it out here.
1660                         */
1661                        try_to_unmap_flush_dirty();
1662                        switch (pageout(page, mapping)) {
1663                        case PAGE_KEEP:
1664                                goto keep_locked;
1665                        case PAGE_ACTIVATE:
1666                                goto activate_locked;
1667                        case PAGE_SUCCESS:
1668                                stat->nr_pageout += thp_nr_pages(page);
1669
1670                                if (PageWriteback(page))
1671                                        goto keep;
1672                                if (PageDirty(page))
1673                                        goto keep;
1674
1675                                /*
1676                                 * A synchronous write - probably a ramdisk.  Go
1677                                 * ahead and try to reclaim the page.
1678                                 */
1679                                if (!trylock_page(page))
1680                                        goto keep;
1681                                if (PageDirty(page) || PageWriteback(page))
1682                                        goto keep_locked;
1683                                mapping = page_mapping(page);
1684                                fallthrough;
1685                        case PAGE_CLEAN:
1686                                ; /* try to free the page below */
1687                        }
1688                }
1689
1690                /*
1691                 * If the page has buffers, try to free the buffer mappings
1692                 * associated with this page. If we succeed we try to free
1693                 * the page as well.
1694                 *
1695                 * We do this even if the page is PageDirty().
1696                 * try_to_release_page() does not perform I/O, but it is
1697                 * possible for a page to have PageDirty set, but it is actually
1698                 * clean (all its buffers are clean).  This happens if the
1699                 * buffers were written out directly, with submit_bh(). ext3
1700                 * will do this, as well as the blockdev mapping.
1701                 * try_to_release_page() will discover that cleanness and will
1702                 * drop the buffers and mark the page clean - it can be freed.
1703                 *
1704                 * Rarely, pages can have buffers and no ->mapping.  These are
1705                 * the pages which were not successfully invalidated in
1706                 * truncate_cleanup_page().  We try to drop those buffers here
1707                 * and if that worked, and the page is no longer mapped into
1708                 * process address space (page_count == 1) it can be freed.
1709                 * Otherwise, leave the page on the LRU so it is swappable.
1710                 */
1711                if (page_has_private(page)) {
1712                        if (!try_to_release_page(page, sc->gfp_mask))
1713                                goto activate_locked;
1714                        if (!mapping && page_count(page) == 1) {
1715                                unlock_page(page);
1716                                if (put_page_testzero(page))
1717                                        goto free_it;
1718                                else {
1719                                        /*
1720                                         * rare race with speculative reference.
1721                                         * the speculative reference will free
1722                                         * this page shortly, so we may
1723                                         * increment nr_reclaimed here (and
1724                                         * leave it off the LRU).
1725                                         */
1726                                        nr_reclaimed++;
1727                                        continue;
1728                                }
1729                        }
1730                }
1731
1732                if (PageAnon(page) && !PageSwapBacked(page)) {
1733                        /* follow __remove_mapping for reference */
1734                        if (!page_ref_freeze(page, 1))
1735                                goto keep_locked;
1736                        /*
1737                         * The page has only one reference left, which is
1738                         * from the isolation. After the caller puts the
1739                         * page back on lru and drops the reference, the
1740                         * page will be freed anyway. It doesn't matter
1741                         * which lru it goes. So we don't bother checking
1742                         * PageDirty here.
1743                         */
1744                        count_vm_event(PGLAZYFREED);
1745                        count_memcg_page_event(page, PGLAZYFREED);
1746                } else if (!mapping || !__remove_mapping(mapping, page, true,
1747                                                         sc->target_mem_cgroup))
1748                        goto keep_locked;
1749
1750                unlock_page(page);
1751free_it:
1752                /*
1753                 * THP may get swapped out in a whole, need account
1754                 * all base pages.
1755                 */
1756                nr_reclaimed += nr_pages;
1757
1758                /*
1759                 * Is there need to periodically free_page_list? It would
1760                 * appear not as the counts should be low
1761                 */
1762                if (unlikely(PageTransHuge(page)))
1763                        destroy_compound_page(page);
1764                else
1765                        list_add(&page->lru, &free_pages);
1766                continue;
1767
1768activate_locked_split:
1769                /*
1770                 * The tail pages that are failed to add into swap cache
1771                 * reach here.  Fixup nr_scanned and nr_pages.
1772                 */
1773                if (nr_pages > 1) {
1774                        sc->nr_scanned -= (nr_pages - 1);
1775                        nr_pages = 1;
1776                }
1777activate_locked:
1778                /* Not a candidate for swapping, so reclaim swap space. */
1779                if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1780                                                PageMlocked(page)))
1781                        try_to_free_swap(page);
1782                VM_BUG_ON_PAGE(PageActive(page), page);
1783                if (!PageMlocked(page)) {
1784                        int type = page_is_file_lru(page);
1785                        SetPageActive(page);
1786                        stat->nr_activate[type] += nr_pages;
1787                        count_memcg_page_event(page, PGACTIVATE);
1788                }
1789keep_locked:
1790                unlock_page(page);
1791keep:
1792                list_add(&page->lru, &ret_pages);
1793                VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1794        }
1795        /* 'page_list' is always empty here */
1796
1797        /* Migrate pages selected for demotion */
1798        nr_reclaimed += demote_page_list(&demote_pages, pgdat);
1799        /* Pages that could not be demoted are still in @demote_pages */
1800        if (!list_empty(&demote_pages)) {
1801                /* Pages which failed to demoted go back on @page_list for retry: */
1802                list_splice_init(&demote_pages, page_list);
1803                do_demote_pass = false;
1804                goto retry;
1805        }
1806
1807        pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1808
1809        mem_cgroup_uncharge_list(&free_pages);
1810        try_to_unmap_flush();
1811        free_unref_page_list(&free_pages);
1812
1813        list_splice(&ret_pages, page_list);
1814        count_vm_events(PGACTIVATE, pgactivate);
1815
1816        return nr_reclaimed;
1817}
1818
1819unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1820                                            struct list_head *page_list)
1821{
1822        struct scan_control sc = {
1823                .gfp_mask = GFP_KERNEL,
1824                .may_unmap = 1,
1825        };
1826        struct reclaim_stat stat;
1827        unsigned int nr_reclaimed;
1828        struct page *page, *next;
1829        LIST_HEAD(clean_pages);
1830        unsigned int noreclaim_flag;
1831
1832        list_for_each_entry_safe(page, next, page_list, lru) {
1833                if (!PageHuge(page) && page_is_file_lru(page) &&
1834                    !PageDirty(page) && !__PageMovable(page) &&
1835                    !PageUnevictable(page)) {
1836                        ClearPageActive(page);
1837                        list_move(&page->lru, &clean_pages);
1838                }
1839        }
1840
1841        /*
1842         * We should be safe here since we are only dealing with file pages and
1843         * we are not kswapd and therefore cannot write dirty file pages. But
1844         * call memalloc_noreclaim_save() anyway, just in case these conditions
1845         * change in the future.
1846         */
1847        noreclaim_flag = memalloc_noreclaim_save();
1848        nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1849                                        &stat, true);
1850        memalloc_noreclaim_restore(noreclaim_flag);
1851
1852        list_splice(&clean_pages, page_list);
1853        mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1854                            -(long)nr_reclaimed);
1855        /*
1856         * Since lazyfree pages are isolated from file LRU from the beginning,
1857         * they will rotate back to anonymous LRU in the end if it failed to
1858         * discard so isolated count will be mismatched.
1859         * Compensate the isolated count for both LRU lists.
1860         */
1861        mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1862                            stat.nr_lazyfree_fail);
1863        mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1864                            -(long)stat.nr_lazyfree_fail);
1865        return nr_reclaimed;
1866}
1867
1868/*
1869 * Attempt to remove the specified page from its LRU.  Only take this page
1870 * if it is of the appropriate PageActive status.  Pages which are being
1871 * freed elsewhere are also ignored.
1872 *
1873 * page:        page to consider
1874 * mode:        one of the LRU isolation modes defined above
1875 *
1876 * returns true on success, false on failure.
1877 */
1878bool __isolate_lru_page_prepare(struct page *page, isolate_mode_t mode)
1879{
1880        /* Only take pages on the LRU. */
1881        if (!PageLRU(page))
1882                return false;
1883
1884        /* Compaction should not handle unevictable pages but CMA can do so */
1885        if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1886                return false;
1887
1888        /*
1889         * To minimise LRU disruption, the caller can indicate that it only
1890         * wants to isolate pages it will be able to operate on without
1891         * blocking - clean pages for the most part.
1892         *
1893         * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1894         * that it is possible to migrate without blocking
1895         */
1896        if (mode & ISOLATE_ASYNC_MIGRATE) {
1897                /* All the caller can do on PageWriteback is block */
1898                if (PageWriteback(page))
1899                        return false;
1900
1901                if (PageDirty(page)) {
1902                        struct address_space *mapping;
1903                        bool migrate_dirty;
1904
1905                        /*
1906                         * Only pages without mappings or that have a
1907                         * ->migratepage callback are possible to migrate
1908                         * without blocking. However, we can be racing with
1909                         * truncation so it's necessary to lock the page
1910                         * to stabilise the mapping as truncation holds
1911                         * the page lock until after the page is removed
1912                         * from the page cache.
1913                         */
1914                        if (!trylock_page(page))
1915                                return false;
1916
1917                        mapping = page_mapping(page);
1918                        migrate_dirty = !mapping || mapping->a_ops->migratepage;
1919                        unlock_page(page);
1920                        if (!migrate_dirty)
1921                                return false;
1922                }
1923        }
1924
1925        if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1926                return false;
1927
1928        return true;
1929}
1930
1931/*
1932 * Update LRU sizes after isolating pages. The LRU size updates must
1933 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1934 */
1935static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1936                        enum lru_list lru, unsigned long *nr_zone_taken)
1937{
1938        int zid;
1939
1940        for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1941                if (!nr_zone_taken[zid])
1942                        continue;
1943
1944                update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1945        }
1946
1947}
1948
1949/*
1950 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1951 *
1952 * lruvec->lru_lock is heavily contended.  Some of the functions that
1953 * shrink the lists perform better by taking out a batch of pages
1954 * and working on them outside the LRU lock.
1955 *
1956 * For pagecache intensive workloads, this function is the hottest
1957 * spot in the kernel (apart from copy_*_user functions).
1958 *
1959 * Lru_lock must be held before calling this function.
1960 *
1961 * @nr_to_scan: The number of eligible pages to look through on the list.
1962 * @lruvec:     The LRU vector to pull pages from.
1963 * @dst:        The temp list to put pages on to.
1964 * @nr_scanned: The number of pages that were scanned.
1965 * @sc:         The scan_control struct for this reclaim session
1966 * @lru:        LRU list id for isolating
1967 *
1968 * returns how many pages were moved onto *@dst.
1969 */
1970static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1971                struct lruvec *lruvec, struct list_head *dst,
1972                unsigned long *nr_scanned, struct scan_control *sc,
1973                enum lru_list lru)
1974{
1975        struct list_head *src = &lruvec->lists[lru];
1976        unsigned long nr_taken = 0;
1977        unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1978        unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1979        unsigned long skipped = 0;
1980        unsigned long scan, total_scan, nr_pages;
1981        LIST_HEAD(pages_skipped);
1982        isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1983
1984        total_scan = 0;
1985        scan = 0;
1986        while (scan < nr_to_scan && !list_empty(src)) {
1987                struct page *page;
1988
1989                page = lru_to_page(src);
1990                prefetchw_prev_lru_page(page, src, flags);
1991
1992                nr_pages = compound_nr(page);
1993                total_scan += nr_pages;
1994
1995                if (page_zonenum(page) > sc->reclaim_idx) {
1996                        list_move(&page->lru, &pages_skipped);
1997                        nr_skipped[page_zonenum(page)] += nr_pages;
1998                        continue;
1999                }
2000
2001                /*
2002                 * Do not count skipped pages because that makes the function
2003                 * return with no isolated pages if the LRU mostly contains
2004                 * ineligible pages.  This causes the VM to not reclaim any
2005                 * pages, triggering a premature OOM.
2006                 *
2007                 * Account all tail pages of THP.  This would not cause
2008                 * premature OOM since __isolate_lru_page() returns -EBUSY
2009                 * only when the page is being freed somewhere else.
2010                 */
2011                scan += nr_pages;
2012                if (!__isolate_lru_page_prepare(page, mode)) {
2013                        /* It is being freed elsewhere */
2014                        list_move(&page->lru, src);
2015                        continue;
2016                }
2017                /*
2018                 * Be careful not to clear PageLRU until after we're
2019                 * sure the page is not being freed elsewhere -- the
2020                 * page release code relies on it.
2021                 */
2022                if (unlikely(!get_page_unless_zero(page))) {
2023                        list_move(&page->lru, src);
2024                        continue;
2025                }
2026
2027                if (!TestClearPageLRU(page)) {
2028                        /* Another thread is already isolating this page */
2029                        put_page(page);
2030                        list_move(&page->lru, src);
2031                        continue;
2032                }
2033
2034                nr_taken += nr_pages;
2035                nr_zone_taken[page_zonenum(page)] += nr_pages;
2036                list_move(&page->lru, dst);
2037        }
2038
2039        /*
2040         * Splice any skipped pages to the start of the LRU list. Note that
2041         * this disrupts the LRU order when reclaiming for lower zones but
2042         * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
2043         * scanning would soon rescan the same pages to skip and put the
2044         * system at risk of premature OOM.
2045         */
2046        if (!list_empty(&pages_skipped)) {
2047                int zid;
2048
2049                list_splice(&pages_skipped, src);
2050                for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2051                        if (!nr_skipped[zid])
2052                                continue;
2053
2054                        __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
2055                        skipped += nr_skipped[zid];
2056                }
2057        }
2058        *nr_scanned = total_scan;
2059        trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
2060                                    total_scan, skipped, nr_taken, mode, lru);
2061        update_lru_sizes(lruvec, lru, nr_zone_taken);
2062        return nr_taken;
2063}
2064
2065/**
2066 * isolate_lru_page - tries to isolate a page from its LRU list
2067 * @page: page to isolate from its LRU list
2068 *
2069 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
2070 * vmstat statistic corresponding to whatever LRU list the page was on.
2071 *
2072 * Returns 0 if the page was removed from an LRU list.
2073 * Returns -EBUSY if the page was not on an LRU list.
2074 *
2075 * The returned page will have PageLRU() cleared.  If it was found on
2076 * the active list, it will have PageActive set.  If it was found on
2077 * the unevictable list, it will have the PageUnevictable bit set. That flag
2078 * may need to be cleared by the caller before letting the page go.
2079 *
2080 * The vmstat statistic corresponding to the list on which the page was
2081 * found will be decremented.
2082 *
2083 * Restrictions:
2084 *
2085 * (1) Must be called with an elevated refcount on the page. This is a
2086 *     fundamental difference from isolate_lru_pages (which is called
2087 *     without a stable reference).
2088 * (2) the lru_lock must not be held.
2089 * (3) interrupts must be enabled.
2090 */
2091int isolate_lru_page(struct page *page)
2092{
2093        int ret = -EBUSY;
2094
2095        VM_BUG_ON_PAGE(!page_count(page), page);
2096        WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
2097
2098        if (TestClearPageLRU(page)) {
2099                struct lruvec *lruvec;
2100
2101                get_page(page);
2102                lruvec = lock_page_lruvec_irq(page);
2103                del_page_from_lru_list(page, lruvec);
2104                unlock_page_lruvec_irq(lruvec);
2105                ret = 0;
2106        }
2107
2108        return ret;
2109}
2110
2111/*
2112 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
2113 * then get rescheduled. When there are massive number of tasks doing page
2114 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
2115 * the LRU list will go small and be scanned faster than necessary, leading to
2116 * unnecessary swapping, thrashing and OOM.
2117 */
2118static int too_many_isolated(struct pglist_data *pgdat, int file,
2119                struct scan_control *sc)
2120{
2121        unsigned long inactive, isolated;
2122
2123        if (current_is_kswapd())
2124                return 0;
2125
2126        if (!writeback_throttling_sane(sc))
2127                return 0;
2128
2129        if (file) {
2130                inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
2131                isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
2132        } else {
2133                inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
2134                isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
2135        }
2136
2137        /*
2138         * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2139         * won't get blocked by normal direct-reclaimers, forming a circular
2140         * deadlock.
2141         */
2142        if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
2143                inactive >>= 3;
2144
2145        return isolated > inactive;
2146}
2147
2148/*
2149 * move_pages_to_lru() moves pages from private @list to appropriate LRU list.
2150 * On return, @list is reused as a list of pages to be freed by the caller.
2151 *
2152 * Returns the number of pages moved to the given lruvec.
2153 */
2154static unsigned int move_pages_to_lru(struct lruvec *lruvec,
2155                                      struct list_head *list)
2156{
2157        int nr_pages, nr_moved = 0;
2158        LIST_HEAD(pages_to_free);
2159        struct page *page;
2160
2161        while (!list_empty(list)) {
2162                page = lru_to_page(list);
2163                VM_BUG_ON_PAGE(PageLRU(page), page);
2164                list_del(&page->lru);
2165                if (unlikely(!page_evictable(page))) {
2166                        spin_unlock_irq(&lruvec->lru_lock);
2167                        putback_lru_page(page);
2168                        spin_lock_irq(&lruvec->lru_lock);
2169                        continue;
2170                }
2171
2172                /*
2173                 * The SetPageLRU needs to be kept here for list integrity.
2174                 * Otherwise:
2175                 *   #0 move_pages_to_lru             #1 release_pages
2176                 *   if !put_page_testzero
2177                 *                                    if (put_page_testzero())
2178                 *                                      !PageLRU //skip lru_lock
2179                 *     SetPageLRU()
2180                 *     list_add(&page->lru,)
2181                 *                                        list_add(&page->lru,)
2182                 */
2183                SetPageLRU(page);
2184
2185                if (unlikely(put_page_testzero(page))) {
2186                        __clear_page_lru_flags(page);
2187
2188                        if (unlikely(PageCompound(page))) {
2189                                spin_unlock_irq(&lruvec->lru_lock);
2190                                destroy_compound_page(page);
2191                                spin_lock_irq(&lruvec->lru_lock);
2192                        } else
2193                                list_add(&page->lru, &pages_to_free);
2194
2195                        continue;
2196                }
2197
2198                /*
2199                 * All pages were isolated from the same lruvec (and isolation
2200                 * inhibits memcg migration).
2201                 */
2202                VM_BUG_ON_PAGE(!page_matches_lruvec(page, lruvec), page);
2203                add_page_to_lru_list(page, lruvec);
2204                nr_pages = thp_nr_pages(page);
2205                nr_moved += nr_pages;
2206                if (PageActive(page))
2207                        workingset_age_nonresident(lruvec, nr_pages);
2208        }
2209
2210        /*
2211         * To save our caller's stack, now use input list for pages to free.
2212         */
2213        list_splice(&pages_to_free, list);
2214
2215        return nr_moved;
2216}
2217
2218/*
2219 * If a kernel thread (such as nfsd for loop-back mounts) services
2220 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
2221 * In that case we should only throttle if the backing device it is
2222 * writing to is congested.  In other cases it is safe to throttle.
2223 */
2224static int current_may_throttle(void)
2225{
2226        return !(current->flags & PF_LOCAL_THROTTLE) ||
2227                current->backing_dev_info == NULL ||
2228                bdi_write_congested(current->backing_dev_info);
2229}
2230
2231/*
2232 * shrink_inactive_list() is a helper for shrink_node().  It returns the number
2233 * of reclaimed pages
2234 */
2235static unsigned long
2236shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
2237                     struct scan_control *sc, enum lru_list lru)
2238{
2239        LIST_HEAD(page_list);
2240        unsigned long nr_scanned;
2241        unsigned int nr_reclaimed = 0;
2242        unsigned long nr_taken;
2243        struct reclaim_stat stat;
2244        bool file = is_file_lru(lru);
2245        enum vm_event_item item;
2246        struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2247        bool stalled = false;
2248
2249        while (unlikely(too_many_isolated(pgdat, file, sc))) {
2250                if (stalled)
2251                        return 0;
2252
2253                /* wait a bit for the reclaimer. */
2254                msleep(100);
2255                stalled = true;
2256
2257                /* We are about to die and free our memory. Return now. */
2258                if (fatal_signal_pending(current))
2259                        return SWAP_CLUSTER_MAX;
2260        }
2261
2262        lru_add_drain();
2263
2264        spin_lock_irq(&lruvec->lru_lock);
2265
2266        nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
2267                                     &nr_scanned, sc, lru);
2268
2269        __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2270        item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
2271        if (!cgroup_reclaim(sc))
2272                __count_vm_events(item, nr_scanned);
2273        __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
2274        __count_vm_events(PGSCAN_ANON + file, nr_scanned);
2275
2276        spin_unlock_irq(&lruvec->lru_lock);
2277
2278        if (nr_taken == 0)
2279                return 0;
2280
2281        nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
2282
2283        spin_lock_irq(&lruvec->lru_lock);
2284        move_pages_to_lru(lruvec, &page_list);
2285
2286        __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2287        item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2288        if (!cgroup_reclaim(sc))
2289                __count_vm_events(item, nr_reclaimed);
2290        __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2291        __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
2292        spin_unlock_irq(&lruvec->lru_lock);
2293
2294        lru_note_cost(lruvec, file, stat.nr_pageout);
2295        mem_cgroup_uncharge_list(&page_list);
2296        free_unref_page_list(&page_list);
2297
2298        /*
2299         * If dirty pages are scanned that are not queued for IO, it
2300         * implies that flushers are not doing their job. This can
2301         * happen when memory pressure pushes dirty pages to the end of
2302         * the LRU before the dirty limits are breached and the dirty
2303         * data has expired. It can also happen when the proportion of
2304         * dirty pages grows not through writes but through memory
2305         * pressure reclaiming all the clean cache. And in some cases,
2306         * the flushers simply cannot keep up with the allocation
2307         * rate. Nudge the flusher threads in case they are asleep.
2308         */
2309        if (stat.nr_unqueued_dirty == nr_taken)
2310                wakeup_flusher_threads(WB_REASON_VMSCAN);
2311
2312        sc->nr.dirty += stat.nr_dirty;
2313        sc->nr.congested += stat.nr_congested;
2314        sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2315        sc->nr.writeback += stat.nr_writeback;
2316        sc->nr.immediate += stat.nr_immediate;
2317        sc->nr.taken += nr_taken;
2318        if (file)
2319                sc->nr.file_taken += nr_taken;
2320
2321        trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2322                        nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2323        return nr_reclaimed;
2324}
2325
2326/*
2327 * shrink_active_list() moves pages from the active LRU to the inactive LRU.
2328 *
2329 * We move them the other way if the page is referenced by one or more
2330 * processes.
2331 *
2332 * If the pages are mostly unmapped, the processing is fast and it is
2333 * appropriate to hold lru_lock across the whole operation.  But if
2334 * the pages are mapped, the processing is slow (page_referenced()), so
2335 * we should drop lru_lock around each page.  It's impossible to balance
2336 * this, so instead we remove the pages from the LRU while processing them.
2337 * It is safe to rely on PG_active against the non-LRU pages in here because
2338 * nobody will play with that bit on a non-LRU page.
2339 *
2340 * The downside is that we have to touch page->_refcount against each page.
2341 * But we had to alter page->flags anyway.
2342 */
2343static void shrink_active_list(unsigned long nr_to_scan,
2344                               struct lruvec *lruvec,
2345                               struct scan_control *sc,
2346                               enum lru_list lru)
2347{
2348        unsigned long nr_taken;
2349        unsigned long nr_scanned;
2350        unsigned long vm_flags;
2351        LIST_HEAD(l_hold);      /* The pages which were snipped off */
2352        LIST_HEAD(l_active);
2353        LIST_HEAD(l_inactive);
2354        struct page *page;
2355        unsigned nr_deactivate, nr_activate;
2356        unsigned nr_rotated = 0;
2357        int file = is_file_lru(lru);
2358        struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2359
2360        lru_add_drain();
2361
2362        spin_lock_irq(&lruvec->lru_lock);
2363
2364        nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2365                                     &nr_scanned, sc, lru);
2366
2367        __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2368
2369        if (!cgroup_reclaim(sc))
2370                __count_vm_events(PGREFILL, nr_scanned);
2371        __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2372
2373        spin_unlock_irq(&lruvec->lru_lock);
2374
2375        while (!list_empty(&l_hold)) {
2376                cond_resched();
2377                page = lru_to_page(&l_hold);
2378                list_del(&page->lru);
2379
2380                if (unlikely(!page_evictable(page))) {
2381                        putback_lru_page(page);
2382                        continue;
2383                }
2384
2385                if (unlikely(buffer_heads_over_limit)) {
2386                        if (page_has_private(page) && trylock_page(page)) {
2387                                if (page_has_private(page))
2388                                        try_to_release_page(page, 0);
2389                                unlock_page(page);
2390                        }
2391                }
2392
2393                if (page_referenced(page, 0, sc->target_mem_cgroup,
2394                                    &vm_flags)) {
2395                        /*
2396                         * Identify referenced, file-backed active pages and
2397                         * give them one more trip around the active list. So
2398                         * that executable code get better chances to stay in
2399                         * memory under moderate memory pressure.  Anon pages
2400                         * are not likely to be evicted by use-once streaming
2401                         * IO, plus JVM can create lots of anon VM_EXEC pages,
2402                         * so we ignore them here.
2403                         */
2404                        if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
2405                                nr_rotated += thp_nr_pages(page);
2406                                list_add(&page->lru, &l_active);
2407                                continue;
2408                        }
2409                }
2410
2411                ClearPageActive(page);  /* we are de-activating */
2412                SetPageWorkingset(page);
2413                list_add(&page->lru, &l_inactive);
2414        }
2415
2416        /*
2417         * Move pages back to the lru list.
2418         */
2419        spin_lock_irq(&lruvec->lru_lock);
2420
2421        nr_activate = move_pages_to_lru(lruvec, &l_active);
2422        nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2423        /* Keep all free pages in l_active list */
2424        list_splice(&l_inactive, &l_active);
2425
2426        __count_vm_events(PGDEACTIVATE, nr_deactivate);
2427        __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2428
2429        __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2430        spin_unlock_irq(&lruvec->lru_lock);
2431
2432        mem_cgroup_uncharge_list(&l_active);
2433        free_unref_page_list(&l_active);
2434        trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2435                        nr_deactivate, nr_rotated, sc->priority, file);
2436}
2437
2438unsigned long reclaim_pages(struct list_head *page_list)
2439{
2440        int nid = NUMA_NO_NODE;
2441        unsigned int nr_reclaimed = 0;
2442        LIST_HEAD(node_page_list);
2443        struct reclaim_stat dummy_stat;
2444        struct page *page;
2445        unsigned int noreclaim_flag;
2446        struct scan_control sc = {
2447                .gfp_mask = GFP_KERNEL,
2448                .may_writepage = 1,
2449                .may_unmap = 1,
2450                .may_swap = 1,
2451                .no_demotion = 1,
2452        };
2453
2454        noreclaim_flag = memalloc_noreclaim_save();
2455
2456        while (!list_empty(page_list)) {
2457                page = lru_to_page(page_list);
2458                if (nid == NUMA_NO_NODE) {
2459                        nid = page_to_nid(page);
2460                        INIT_LIST_HEAD(&node_page_list);
2461                }
2462
2463                if (nid == page_to_nid(page)) {
2464                        ClearPageActive(page);
2465                        list_move(&page->lru, &node_page_list);
2466                        continue;
2467                }
2468
2469                nr_reclaimed += shrink_page_list(&node_page_list,
2470                                                NODE_DATA(nid),
2471                                                &sc, &dummy_stat, false);
2472                while (!list_empty(&node_page_list)) {
2473                        page = lru_to_page(&node_page_list);
2474                        list_del(&page->lru);
2475                        putback_lru_page(page);
2476                }
2477
2478                nid = NUMA_NO_NODE;
2479        }
2480
2481        if (!list_empty(&node_page_list)) {
2482                nr_reclaimed += shrink_page_list(&node_page_list,
2483                                                NODE_DATA(nid),
2484                                                &sc, &dummy_stat, false);
2485                while (!list_empty(&node_page_list)) {
2486                        page = lru_to_page(&node_page_list);
2487                        list_del(&page->lru);
2488                        putback_lru_page(page);
2489                }
2490        }
2491
2492        memalloc_noreclaim_restore(noreclaim_flag);
2493
2494        return nr_reclaimed;
2495}
2496
2497static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2498                                 struct lruvec *lruvec, struct scan_control *sc)
2499{
2500        if (is_active_lru(lru)) {
2501                if (sc->may_deactivate & (1 << is_file_lru(lru)))
2502                        shrink_active_list(nr_to_scan, lruvec, sc, lru);
2503                else
2504                        sc->skipped_deactivate = 1;
2505                return 0;
2506        }
2507
2508        return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2509}
2510
2511/*
2512 * The inactive anon list should be small enough that the VM never has
2513 * to do too much work.
2514 *
2515 * The inactive file list should be small enough to leave most memory
2516 * to the established workingset on the scan-resistant active list,
2517 * but large enough to avoid thrashing the aggregate readahead window.
2518 *
2519 * Both inactive lists should also be large enough that each inactive
2520 * page has a chance to be referenced again before it is reclaimed.
2521 *
2522 * If that fails and refaulting is observed, the inactive list grows.
2523 *
2524 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2525 * on this LRU, maintained by the pageout code. An inactive_ratio
2526 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2527 *
2528 * total     target    max
2529 * memory    ratio     inactive
2530 * -------------------------------------
2531 *   10MB       1         5MB
2532 *  100MB       1        50MB
2533 *    1GB       3       250MB
2534 *   10GB      10       0.9GB
2535 *  100GB      31         3GB
2536 *    1TB     101        10GB
2537 *   10TB     320        32GB
2538 */
2539static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2540{
2541        enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2542        unsigned long inactive, active;
2543        unsigned long inactive_ratio;
2544        unsigned long gb;
2545
2546        inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2547        active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2548
2549        gb = (inactive + active) >> (30 - PAGE_SHIFT);
2550        if (gb)
2551                inactive_ratio = int_sqrt(10 * gb);
2552        else
2553                inactive_ratio = 1;
2554
2555        return inactive * inactive_ratio < active;
2556}
2557
2558enum scan_balance {
2559        SCAN_EQUAL,
2560        SCAN_FRACT,
2561        SCAN_ANON,
2562        SCAN_FILE,
2563};
2564
2565/*
2566 * Determine how aggressively the anon and file LRU lists should be
2567 * scanned.  The relative value of each set of LRU lists is determined
2568 * by looking at the fraction of the pages scanned we did rotate back
2569 * onto the active list instead of evict.
2570 *
2571 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2572 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2573 */
2574static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2575                           unsigned long *nr)
2576{
2577        struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2578        struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2579        unsigned long anon_cost, file_cost, total_cost;
2580        int swappiness = mem_cgroup_swappiness(memcg);
2581        u64 fraction[ANON_AND_FILE];
2582        u64 denominator = 0;    /* gcc */
2583        enum scan_balance scan_balance;
2584        unsigned long ap, fp;
2585        enum lru_list lru;
2586
2587        /* If we have no swap space, do not bother scanning anon pages. */
2588        if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2589                scan_balance = SCAN_FILE;
2590                goto out;
2591        }
2592
2593        /*
2594         * Global reclaim will swap to prevent OOM even with no
2595         * swappiness, but memcg users want to use this knob to
2596         * disable swapping for individual groups completely when
2597         * using the memory controller's swap limit feature would be
2598         * too expensive.
2599         */
2600        if (cgroup_reclaim(sc) && !swappiness) {
2601                scan_balance = SCAN_FILE;
2602                goto out;
2603        }
2604
2605        /*
2606         * Do not apply any pressure balancing cleverness when the
2607         * system is close to OOM, scan both anon and file equally
2608         * (unless the swappiness setting disagrees with swapping).
2609         */
2610        if (!sc->priority && swappiness) {
2611                scan_balance = SCAN_EQUAL;
2612                goto out;
2613        }
2614
2615        /*
2616         * If the system is almost out of file pages, force-scan anon.
2617         */
2618        if (sc->file_is_tiny) {
2619                scan_balance = SCAN_ANON;
2620                goto out;
2621        }
2622
2623        /*
2624         * If there is enough inactive page cache, we do not reclaim
2625         * anything from the anonymous working right now.
2626         */
2627        if (sc->cache_trim_mode) {
2628                scan_balance = SCAN_FILE;
2629                goto out;
2630        }
2631
2632        scan_balance = SCAN_FRACT;
2633        /*
2634         * Calculate the pressure balance between anon and file pages.
2635         *
2636         * The amount of pressure we put on each LRU is inversely
2637         * proportional to the cost of reclaiming each list, as
2638         * determined by the share of pages that are refaulting, times
2639         * the relative IO cost of bringing back a swapped out
2640         * anonymous page vs reloading a filesystem page (swappiness).
2641         *
2642         * Although we limit that influence to ensure no list gets
2643         * left behind completely: at least a third of the pressure is
2644         * applied, before swappiness.
2645         *
2646         * With swappiness at 100, anon and file have equal IO cost.
2647         */
2648        total_cost = sc->anon_cost + sc->file_cost;
2649        anon_cost = total_cost + sc->anon_cost;
2650        file_cost = total_cost + sc->file_cost;
2651        total_cost = anon_cost + file_cost;
2652
2653        ap = swappiness * (total_cost + 1);
2654        ap /= anon_cost + 1;
2655
2656        fp = (200 - swappiness) * (total_cost + 1);
2657        fp /= file_cost + 1;
2658
2659        fraction[0] = ap;
2660        fraction[1] = fp;
2661        denominator = ap + fp;
2662out:
2663        for_each_evictable_lru(lru) {
2664                int file = is_file_lru(lru);
2665                unsigned long lruvec_size;
2666                unsigned long low, min;
2667                unsigned long scan;
2668
2669                lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2670                mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2671                                      &min, &low);
2672
2673                if (min || low) {
2674                        /*
2675                         * Scale a cgroup's reclaim pressure by proportioning
2676                         * its current usage to its memory.low or memory.min
2677                         * setting.
2678                         *
2679                         * This is important, as otherwise scanning aggression
2680                         * becomes extremely binary -- from nothing as we
2681                         * approach the memory protection threshold, to totally
2682                         * nominal as we exceed it.  This results in requiring
2683                         * setting extremely liberal protection thresholds. It
2684                         * also means we simply get no protection at all if we
2685                         * set it too low, which is not ideal.
2686                         *
2687                         * If there is any protection in place, we reduce scan
2688                         * pressure by how much of the total memory used is
2689                         * within protection thresholds.
2690                         *
2691                         * There is one special case: in the first reclaim pass,
2692                         * we skip over all groups that are within their low
2693                         * protection. If that fails to reclaim enough pages to
2694                         * satisfy the reclaim goal, we come back and override
2695                         * the best-effort low protection. However, we still
2696                         * ideally want to honor how well-behaved groups are in
2697                         * that case instead of simply punishing them all
2698                         * equally. As such, we reclaim them based on how much
2699                         * memory they are using, reducing the scan pressure
2700                         * again by how much of the total memory used is under
2701                         * hard protection.
2702                         */
2703                        unsigned long cgroup_size = mem_cgroup_size(memcg);
2704                        unsigned long protection;
2705
2706                        /* memory.low scaling, make sure we retry before OOM */
2707                        if (!sc->memcg_low_reclaim && low > min) {
2708                                protection = low;
2709                                sc->memcg_low_skipped = 1;
2710                        } else {
2711                                protection = min;
2712                        }
2713
2714                        /* Avoid TOCTOU with earlier protection check */
2715                        cgroup_size = max(cgroup_size, protection);
2716
2717                        scan = lruvec_size - lruvec_size * protection /
2718                                (cgroup_size + 1);
2719
2720                        /*
2721                         * Minimally target SWAP_CLUSTER_MAX pages to keep
2722                         * reclaim moving forwards, avoiding decrementing
2723                         * sc->priority further than desirable.
2724                         */
2725                        scan = max(scan, SWAP_CLUSTER_MAX);
2726                } else {
2727                        scan = lruvec_size;
2728                }
2729
2730                scan >>= sc->priority;
2731
2732                /*
2733                 * If the cgroup's already been deleted, make sure to
2734                 * scrape out the remaining cache.
2735                 */
2736                if (!scan && !mem_cgroup_online(memcg))
2737                        scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2738
2739                switch (scan_balance) {
2740                case SCAN_EQUAL:
2741                        /* Scan lists relative to size */
2742                        break;
2743                case SCAN_FRACT:
2744                        /*
2745                         * Scan types proportional to swappiness and
2746                         * their relative recent reclaim efficiency.
2747                         * Make sure we don't miss the last page on
2748                         * the offlined memory cgroups because of a
2749                         * round-off error.
2750                         */
2751                        scan = mem_cgroup_online(memcg) ?
2752                               div64_u64(scan * fraction[file], denominator) :
2753                               DIV64_U64_ROUND_UP(scan * fraction[file],
2754                                                  denominator);
2755                        break;
2756                case SCAN_FILE:
2757                case SCAN_ANON:
2758                        /* Scan one type exclusively */
2759                        if ((scan_balance == SCAN_FILE) != file)
2760                                scan = 0;
2761                        break;
2762                default:
2763                        /* Look ma, no brain */
2764                        BUG();
2765                }
2766
2767                nr[lru] = scan;
2768        }
2769}
2770
2771/*
2772 * Anonymous LRU management is a waste if there is
2773 * ultimately no way to reclaim the memory.
2774 */
2775static bool can_age_anon_pages(struct pglist_data *pgdat,
2776                               struct scan_control *sc)
2777{
2778        /* Aging the anon LRU is valuable if swap is present: */
2779        if (total_swap_pages > 0)
2780                return true;
2781
2782        /* Also valuable if anon pages can be demoted: */
2783        return can_demote(pgdat->node_id, sc);
2784}
2785
2786static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2787{
2788        unsigned long nr[NR_LRU_LISTS];
2789        unsigned long targets[NR_LRU_LISTS];
2790        unsigned long nr_to_scan;
2791        enum lru_list lru;
2792        unsigned long nr_reclaimed = 0;
2793        unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2794        struct blk_plug plug;
2795        bool scan_adjusted;
2796
2797        get_scan_count(lruvec, sc, nr);
2798
2799        /* Record the original scan target for proportional adjustments later */
2800        memcpy(targets, nr, sizeof(nr));
2801
2802        /*
2803         * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2804         * event that can occur when there is little memory pressure e.g.
2805         * multiple streaming readers/writers. Hence, we do not abort scanning
2806         * when the requested number of pages are reclaimed when scanning at
2807         * DEF_PRIORITY on the assumption that the fact we are direct
2808         * reclaiming implies that kswapd is not keeping up and it is best to
2809         * do a batch of work at once. For memcg reclaim one check is made to
2810         * abort proportional reclaim if either the file or anon lru has already
2811         * dropped to zero at the first pass.
2812         */
2813        scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2814                         sc->priority == DEF_PRIORITY);
2815
2816        blk_start_plug(&plug);
2817        while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2818                                        nr[LRU_INACTIVE_FILE]) {
2819                unsigned long nr_anon, nr_file, percentage;
2820                unsigned long nr_scanned;
2821
2822                for_each_evictable_lru(lru) {
2823                        if (nr[lru]) {
2824                                nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2825                                nr[lru] -= nr_to_scan;
2826
2827                                nr_reclaimed += shrink_list(lru, nr_to_scan,
2828                                                            lruvec, sc);
2829                        }
2830                }
2831
2832                cond_resched();
2833
2834                if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2835                        continue;
2836
2837                /*
2838                 * For kswapd and memcg, reclaim at least the number of pages
2839                 * requested. Ensure that the anon and file LRUs are scanned
2840                 * proportionally what was requested by get_scan_count(). We
2841                 * stop reclaiming one LRU and reduce the amount scanning
2842                 * proportional to the original scan target.
2843                 */
2844                nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2845                nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2846
2847                /*
2848                 * It's just vindictive to attack the larger once the smaller
2849                 * has gone to zero.  And given the way we stop scanning the
2850                 * smaller below, this makes sure that we only make one nudge
2851                 * towards proportionality once we've got nr_to_reclaim.
2852                 */
2853                if (!nr_file || !nr_anon)
2854                        break;
2855
2856                if (nr_file > nr_anon) {
2857                        unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2858                                                targets[LRU_ACTIVE_ANON] + 1;
2859                        lru = LRU_BASE;
2860                        percentage = nr_anon * 100 / scan_target;
2861                } else {
2862                        unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2863                                                targets[LRU_ACTIVE_FILE] + 1;
2864                        lru = LRU_FILE;
2865                        percentage = nr_file * 100 / scan_target;
2866                }
2867
2868                /* Stop scanning the smaller of the LRU */
2869                nr[lru] = 0;
2870                nr[lru + LRU_ACTIVE] = 0;
2871
2872                /*
2873                 * Recalculate the other LRU scan count based on its original
2874                 * scan target and the percentage scanning already complete
2875                 */
2876                lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2877                nr_scanned = targets[lru] - nr[lru];
2878                nr[lru] = targets[lru] * (100 - percentage) / 100;
2879                nr[lru] -= min(nr[lru], nr_scanned);
2880
2881                lru += LRU_ACTIVE;
2882                nr_scanned = targets[lru] - nr[lru];
2883                nr[lru] = targets[lru] * (100 - percentage) / 100;
2884                nr[lru] -= min(nr[lru], nr_scanned);
2885
2886                scan_adjusted = true;
2887        }
2888        blk_finish_plug(&plug);
2889        sc->nr_reclaimed += nr_reclaimed;
2890
2891        /*
2892         * Even if we did not try to evict anon pages at all, we want to
2893         * rebalance the anon lru active/inactive ratio.
2894         */
2895        if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
2896            inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2897                shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2898                                   sc, LRU_ACTIVE_ANON);
2899}
2900
2901/* Use reclaim/compaction for costly allocs or under memory pressure */
2902static bool in_reclaim_compaction(struct scan_control *sc)
2903{
2904        if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2905                        (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2906                         sc->priority < DEF_PRIORITY - 2))
2907                return true;
2908
2909        return false;
2910}
2911
2912/*
2913 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2914 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2915 * true if more pages should be reclaimed such that when the page allocator
2916 * calls try_to_compact_pages() that it will have enough free pages to succeed.
2917 * It will give up earlier than that if there is difficulty reclaiming pages.
2918 */
2919static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2920                                        unsigned long nr_reclaimed,
2921                                        struct scan_control *sc)
2922{
2923        unsigned long pages_for_compaction;
2924        unsigned long inactive_lru_pages;
2925        int z;
2926
2927        /* If not in reclaim/compaction mode, stop */
2928        if (!in_reclaim_compaction(sc))
2929                return false;
2930
2931        /*
2932         * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2933         * number of pages that were scanned. This will return to the caller
2934         * with the risk reclaim/compaction and the resulting allocation attempt
2935         * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2936         * allocations through requiring that the full LRU list has been scanned
2937         * first, by assuming that zero delta of sc->nr_scanned means full LRU
2938         * scan, but that approximation was wrong, and there were corner cases
2939         * where always a non-zero amount of pages were scanned.
2940         */
2941        if (!nr_reclaimed)
2942                return false;
2943
2944        /* If compaction would go ahead or the allocation would succeed, stop */
2945        for (z = 0; z <= sc->reclaim_idx; z++) {
2946                struct zone *zone = &pgdat->node_zones[z];
2947                if (!managed_zone(zone))
2948                        continue;
2949
2950                switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2951                case COMPACT_SUCCESS:
2952                case COMPACT_CONTINUE:
2953                        return false;
2954                default:
2955                        /* check next zone */
2956                        ;
2957                }
2958        }
2959
2960        /*
2961         * If we have not reclaimed enough pages for compaction and the
2962         * inactive lists are large enough, continue reclaiming
2963         */
2964        pages_for_compaction = compact_gap(sc->order);
2965        inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2966        if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
2967                inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2968
2969        return inactive_lru_pages > pages_for_compaction;
2970}
2971
2972static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2973{
2974        struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2975        struct mem_cgroup *memcg;
2976
2977        memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2978        do {
2979                struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2980                unsigned long reclaimed;
2981                unsigned long scanned;
2982
2983                /*
2984                 * This loop can become CPU-bound when target memcgs
2985                 * aren't eligible for reclaim - either because they
2986                 * don't have any reclaimable pages, or because their
2987                 * memory is explicitly protected. Avoid soft lockups.
2988                 */
2989                cond_resched();
2990
2991                mem_cgroup_calculate_protection(target_memcg, memcg);
2992
2993                if (mem_cgroup_below_min(memcg)) {
2994                        /*
2995                         * Hard protection.
2996                         * If there is no reclaimable memory, OOM.
2997                         */
2998                        continue;
2999                } else if (mem_cgroup_below_low(memcg)) {
3000                        /*
3001                         * Soft protection.
3002                         * Respect the protection only as long as
3003                         * there is an unprotected supply
3004                         * of reclaimable memory from other cgroups.
3005                         */
3006                        if (!sc->memcg_low_reclaim) {
3007                                sc->memcg_low_skipped = 1;
3008                                continue;
3009                        }
3010                        memcg_memory_event(memcg, MEMCG_LOW);
3011                }
3012
3013                reclaimed = sc->nr_reclaimed;
3014                scanned = sc->nr_scanned;
3015
3016                shrink_lruvec(lruvec, sc);
3017
3018                shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
3019                            sc->priority);
3020
3021                /* Record the group's reclaim efficiency */
3022                vmpressure(sc->gfp_mask, memcg, false,
3023                           sc->nr_scanned - scanned,
3024                           sc->nr_reclaimed - reclaimed);
3025
3026        } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
3027}
3028
3029static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
3030{
3031        struct reclaim_state *reclaim_state = current->reclaim_state;
3032        unsigned long nr_reclaimed, nr_scanned;
3033        struct lruvec *target_lruvec;
3034        bool reclaimable = false;
3035        unsigned long file;
3036
3037        target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
3038
3039again:
3040        /*
3041         * Flush the memory cgroup stats, so that we read accurate per-memcg
3042         * lruvec stats for heuristics.
3043         */
3044        mem_cgroup_flush_stats();
3045
3046        memset(&sc->nr, 0, sizeof(sc->nr));
3047
3048        nr_reclaimed = sc->nr_reclaimed;
3049        nr_scanned = sc->nr_scanned;
3050
3051        /*
3052         * Determine the scan balance between anon and file LRUs.
3053         */
3054        spin_lock_irq(&target_lruvec->lru_lock);
3055        sc->anon_cost = target_lruvec->anon_cost;
3056        sc->file_cost = target_lruvec->file_cost;
3057        spin_unlock_irq(&target_lruvec->lru_lock);
3058
3059        /*
3060         * Target desirable inactive:active list ratios for the anon
3061         * and file LRU lists.
3062         */
3063        if (!sc->force_deactivate) {
3064                unsigned long refaults;
3065
3066                refaults = lruvec_page_state(target_lruvec,
3067                                WORKINGSET_ACTIVATE_ANON);
3068                if (refaults != target_lruvec->refaults[0] ||
3069                        inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
3070                        sc->may_deactivate |= DEACTIVATE_ANON;
3071                else
3072                        sc->may_deactivate &= ~DEACTIVATE_ANON;
3073
3074                /*
3075                 * When refaults are being observed, it means a new
3076                 * workingset is being established. Deactivate to get
3077                 * rid of any stale active pages quickly.
3078                 */
3079                refaults = lruvec_page_state(target_lruvec,
3080                                WORKINGSET_ACTIVATE_FILE);
3081                if (refaults != target_lruvec->refaults[1] ||
3082                    inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
3083                        sc->may_deactivate |= DEACTIVATE_FILE;
3084                else
3085                        sc->may_deactivate &= ~DEACTIVATE_FILE;
3086        } else
3087                sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
3088
3089        /*
3090         * If we have plenty of inactive file pages that aren't
3091         * thrashing, try to reclaim those first before touching
3092         * anonymous pages.
3093         */
3094        file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
3095        if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
3096                sc->cache_trim_mode = 1;
3097        else
3098                sc->cache_trim_mode = 0;
3099
3100        /*
3101         * Prevent the reclaimer from falling into the cache trap: as
3102         * cache pages start out inactive, every cache fault will tip
3103         * the scan balance towards the file LRU.  And as the file LRU
3104         * shrinks, so does the window for rotation from references.
3105         * This means we have a runaway feedback loop where a tiny
3106         * thrashing file LRU becomes infinitely more attractive than
3107         * anon pages.  Try to detect this based on file LRU size.
3108         */
3109        if (!cgroup_reclaim(sc)) {
3110                unsigned long total_high_wmark = 0;
3111                unsigned long free, anon;
3112                int z;
3113
3114                free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
3115                file = node_page_state(pgdat, NR_ACTIVE_FILE) +
3116                           node_page_state(pgdat, NR_INACTIVE_FILE);
3117
3118                for (z = 0; z < MAX_NR_ZONES; z++) {
3119                        struct zone *zone = &pgdat->node_zones[z];
3120                        if (!managed_zone(zone))
3121                                continue;
3122
3123                        total_high_wmark += high_wmark_pages(zone);
3124                }
3125
3126                /*
3127                 * Consider anon: if that's low too, this isn't a
3128                 * runaway file reclaim problem, but rather just
3129                 * extreme pressure. Reclaim as per usual then.
3130                 */
3131                anon = node_page_state(pgdat, NR_INACTIVE_ANON);
3132
3133                sc->file_is_tiny =
3134                        file + free <= total_high_wmark &&
3135                        !(sc->may_deactivate & DEACTIVATE_ANON) &&
3136                        anon >> sc->priority;
3137        }
3138
3139        shrink_node_memcgs(pgdat, sc);
3140
3141        if (reclaim_state) {
3142                sc->nr_reclaimed += reclaim_state->reclaimed_slab;
3143                reclaim_state->reclaimed_slab = 0;
3144        }
3145
3146        /* Record the subtree's reclaim efficiency */
3147        vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
3148                   sc->nr_scanned - nr_scanned,
3149                   sc->nr_reclaimed - nr_reclaimed);
3150
3151        if (sc->nr_reclaimed - nr_reclaimed)
3152                reclaimable = true;
3153
3154        if (current_is_kswapd()) {
3155                /*
3156                 * If reclaim is isolating dirty pages under writeback,
3157                 * it implies that the long-lived page allocation rate
3158                 * is exceeding the page laundering rate. Either the
3159                 * global limits are not being effective at throttling
3160                 * processes due to the page distribution throughout
3161                 * zones or there is heavy usage of a slow backing
3162                 * device. The only option is to throttle from reclaim
3163                 * context which is not ideal as there is no guarantee
3164                 * the dirtying process is throttled in the same way
3165                 * balance_dirty_pages() manages.
3166                 *
3167                 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
3168                 * count the number of pages under pages flagged for
3169                 * immediate reclaim and stall if any are encountered
3170                 * in the nr_immediate check below.
3171                 */
3172                if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
3173                        set_bit(PGDAT_WRITEBACK, &pgdat->flags);
3174
3175                /* Allow kswapd to start writing pages during reclaim.*/
3176                if (sc->nr.unqueued_dirty == sc->nr.file_taken)
3177                        set_bit(PGDAT_DIRTY, &pgdat->flags);
3178
3179                /*
3180                 * If kswapd scans pages marked for immediate
3181                 * reclaim and under writeback (nr_immediate), it
3182                 * implies that pages are cycling through the LRU
3183                 * faster than they are written so also forcibly stall.
3184                 */
3185                if (sc->nr.immediate)
3186                        congestion_wait(BLK_RW_ASYNC, HZ/10);
3187        }
3188
3189        /*
3190         * Tag a node/memcg as congested if all the dirty pages
3191         * scanned were backed by a congested BDI and
3192         * wait_iff_congested will stall.
3193         *
3194         * Legacy memcg will stall in page writeback so avoid forcibly
3195         * stalling in wait_iff_congested().
3196         */
3197        if ((current_is_kswapd() ||
3198             (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
3199            sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
3200                set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
3201
3202        /*
3203         * Stall direct reclaim for IO completions if underlying BDIs
3204         * and node is congested. Allow kswapd to continue until it
3205         * starts encountering unqueued dirty pages or cycling through
3206         * the LRU too quickly.
3207         */
3208        if (!current_is_kswapd() && current_may_throttle() &&
3209            !sc->hibernation_mode &&
3210            test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
3211                wait_iff_congested(BLK_RW_ASYNC, HZ/10);
3212
3213        if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
3214                                    sc))
3215                goto again;
3216
3217        /*
3218         * Kswapd gives up on balancing particular nodes after too
3219         * many failures to reclaim anything from them and goes to
3220         * sleep. On reclaim progress, reset the failure counter. A
3221         * successful direct reclaim run will revive a dormant kswapd.
3222         */
3223        if (reclaimable)
3224                pgdat->kswapd_failures = 0;
3225}
3226
3227/*
3228 * Returns true if compaction should go ahead for a costly-order request, or
3229 * the allocation would already succeed without compaction. Return false if we
3230 * should reclaim first.
3231 */
3232static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
3233{
3234        unsigned long watermark;
3235        enum compact_result suitable;
3236
3237        suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
3238        if (suitable == COMPACT_SUCCESS)
3239                /* Allocation should succeed already. Don't reclaim. */
3240                return true;
3241        if (suitable == COMPACT_SKIPPED)
3242                /* Compaction cannot yet proceed. Do reclaim. */
3243                return false;
3244
3245        /*
3246         * Compaction is already possible, but it takes time to run and there
3247         * are potentially other callers using the pages just freed. So proceed
3248         * with reclaim to make a buffer of free pages available to give
3249         * compaction a reasonable chance of completing and allocating the page.
3250         * Note that we won't actually reclaim the whole buffer in one attempt
3251         * as the target watermark in should_continue_reclaim() is lower. But if
3252         * we are already above the high+gap watermark, don't reclaim at all.
3253         */
3254        watermark = high_wmark_pages(zone) + compact_gap(sc->order);
3255
3256        return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
3257}
3258
3259/*
3260 * This is the direct reclaim path, for page-allocating processes.  We only
3261 * try to reclaim pages from zones which will satisfy the caller's allocation
3262 * request.
3263 *
3264 * If a zone is deemed to be full of pinned pages then just give it a light
3265 * scan then give up on it.
3266 */
3267static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
3268{
3269        struct zoneref *z;
3270        struct zone *zone;
3271        unsigned long nr_soft_reclaimed;
3272        unsigned long nr_soft_scanned;
3273        gfp_t orig_mask;
3274        pg_data_t *last_pgdat = NULL;
3275
3276        /*
3277         * If the number of buffer_heads in the machine exceeds the maximum
3278         * allowed level, force direct reclaim to scan the highmem zone as
3279         * highmem pages could be pinning lowmem pages storing buffer_heads
3280         */
3281        orig_mask = sc->gfp_mask;
3282        if (buffer_heads_over_limit) {
3283                sc->gfp_mask |= __GFP_HIGHMEM;
3284                sc->reclaim_idx = gfp_zone(sc->gfp_mask);
3285        }
3286
3287        for_each_zone_zonelist_nodemask(zone, z, zonelist,
3288                                        sc->reclaim_idx, sc->nodemask) {
3289                /*
3290                 * Take care memory controller reclaiming has small influence
3291                 * to global LRU.
3292                 */
3293                if (!cgroup_reclaim(sc)) {
3294                        if (!cpuset_zone_allowed(zone,
3295                                                 GFP_KERNEL | __GFP_HARDWALL))
3296                                continue;
3297
3298                        /*
3299                         * If we already have plenty of memory free for
3300                         * compaction in this zone, don't free any more.
3301                         * Even though compaction is invoked for any
3302                         * non-zero order, only frequent costly order
3303                         * reclamation is disruptive enough to become a
3304                         * noticeable problem, like transparent huge
3305                         * page allocations.
3306                         */
3307                        if (IS_ENABLED(CONFIG_COMPACTION) &&
3308                            sc->order > PAGE_ALLOC_COSTLY_ORDER &&
3309                            compaction_ready(zone, sc)) {
3310                                sc->compaction_ready = true;
3311                                continue;
3312                        }
3313
3314                        /*
3315                         * Shrink each node in the zonelist once. If the
3316                         * zonelist is ordered by zone (not the default) then a
3317                         * node may be shrunk multiple times but in that case
3318                         * the user prefers lower zones being preserved.
3319                         */
3320                        if (zone->zone_pgdat == last_pgdat)
3321                                continue;
3322
3323                        /*
3324                         * This steals pages from memory cgroups over softlimit
3325                         * and returns the number of reclaimed pages and
3326                         * scanned pages. This works for global memory pressure
3327                         * and balancing, not for a memcg's limit.
3328                         */
3329                        nr_soft_scanned = 0;
3330                        nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
3331                                                sc->order, sc->gfp_mask,
3332                                                &nr_soft_scanned);
3333                        sc->nr_reclaimed += nr_soft_reclaimed;
3334                        sc->nr_scanned += nr_soft_scanned;
3335                        /* need some check for avoid more shrink_zone() */
3336                }
3337
3338                /* See comment about same check for global reclaim above */
3339                if (zone->zone_pgdat == last_pgdat)
3340                        continue;
3341                last_pgdat = zone->zone_pgdat;
3342                shrink_node(zone->zone_pgdat, sc);
3343        }
3344
3345        /*
3346         * Restore to original mask to avoid the impact on the caller if we
3347         * promoted it to __GFP_HIGHMEM.
3348         */
3349        sc->gfp_mask = orig_mask;
3350}
3351
3352static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
3353{
3354        struct lruvec *target_lruvec;
3355        unsigned long refaults;
3356
3357        target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
3358        refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
3359        target_lruvec->refaults[0] = refaults;
3360        refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
3361        target_lruvec->refaults[1] = refaults;
3362}
3363
3364/*
3365 * This is the main entry point to direct page reclaim.
3366 *
3367 * If a full scan of the inactive list fails to free enough memory then we
3368 * are "out of memory" and something needs to be killed.
3369 *
3370 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3371 * high - the zone may be full of dirty or under-writeback pages, which this
3372 * caller can't do much about.  We kick the writeback threads and take explicit
3373 * naps in the hope that some of these pages can be written.  But if the
3374 * allocating task holds filesystem locks which prevent writeout this might not
3375 * work, and the allocation attempt will fail.
3376 *
3377 * returns:     0, if no pages reclaimed
3378 *              else, the number of pages reclaimed
3379 */
3380static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3381                                          struct scan_control *sc)
3382{
3383        int initial_priority = sc->priority;
3384        pg_data_t *last_pgdat;
3385        struct zoneref *z;
3386        struct zone *zone;
3387retry:
3388        delayacct_freepages_start();
3389
3390        if (!cgroup_reclaim(sc))
3391                __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3392
3393        do {
3394                vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3395                                sc->priority);
3396                sc->nr_scanned = 0;
3397                shrink_zones(zonelist, sc);
3398
3399                if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3400                        break;
3401
3402                if (sc->compaction_ready)
3403                        break;
3404
3405                /*
3406                 * If we're getting trouble reclaiming, start doing
3407                 * writepage even in laptop mode.
3408                 */
3409                if (sc->priority < DEF_PRIORITY - 2)
3410                        sc->may_writepage = 1;
3411        } while (--sc->priority >= 0);
3412
3413        last_pgdat = NULL;
3414        for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3415                                        sc->nodemask) {
3416                if (zone->zone_pgdat == last_pgdat)
3417                        continue;
3418                last_pgdat = zone->zone_pgdat;
3419
3420                snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3421
3422                if (cgroup_reclaim(sc)) {
3423                        struct lruvec *lruvec;
3424
3425                        lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3426                                                   zone->zone_pgdat);
3427                        clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3428                }
3429        }
3430
3431        delayacct_freepages_end();
3432
3433        if (sc->nr_reclaimed)
3434                return sc->nr_reclaimed;
3435
3436        /* Aborted reclaim to try compaction? don't OOM, then */
3437        if (sc->compaction_ready)
3438                return 1;
3439
3440        /*
3441         * We make inactive:active ratio decisions based on the node's
3442         * composition of memory, but a restrictive reclaim_idx or a
3443         * memory.low cgroup setting can exempt large amounts of
3444         * memory from reclaim. Neither of which are very common, so
3445         * instead of doing costly eligibility calculations of the
3446         * entire cgroup subtree up front, we assume the estimates are
3447         * good, and retry with forcible deactivation if that fails.
3448         */
3449        if (sc->skipped_deactivate) {
3450                sc->priority = initial_priority;
3451                sc->force_deactivate = 1;
3452                sc->skipped_deactivate = 0;
3453                goto retry;
3454        }
3455
3456        /* Untapped cgroup reserves?  Don't OOM, retry. */
3457        if (sc->memcg_low_skipped) {
3458                sc->priority = initial_priority;
3459                sc->force_deactivate = 0;
3460                sc->memcg_low_reclaim = 1;
3461                sc->memcg_low_skipped = 0;
3462                goto retry;
3463        }
3464
3465        return 0;
3466}
3467
3468static bool allow_direct_reclaim(pg_data_t *pgdat)
3469{
3470        struct zone *zone;
3471        unsigned long pfmemalloc_reserve = 0;
3472        unsigned long free_pages = 0;
3473        int i;
3474        bool wmark_ok;
3475
3476        if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3477                return true;
3478
3479        for (i = 0; i <= ZONE_NORMAL; i++) {
3480                zone = &pgdat->node_zones[i];
3481                if (!managed_zone(zone))
3482                        continue;
3483
3484                if (!zone_reclaimable_pages(zone))
3485                        continue;
3486
3487                pfmemalloc_reserve += min_wmark_pages(zone);
3488                free_pages += zone_page_state(zone, NR_FREE_PAGES);
3489        }
3490
3491        /* If there are no reserves (unexpected config) then do not throttle */
3492        if (!pfmemalloc_reserve)
3493                return true;
3494
3495        wmark_ok = free_pages > pfmemalloc_reserve / 2;
3496
3497        /* kswapd must be awake if processes are being throttled */
3498        if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3499                if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3500                        WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3501
3502                wake_up_interruptible(&pgdat->kswapd_wait);
3503        }
3504
3505        return wmark_ok;
3506}
3507
3508/*
3509 * Throttle direct reclaimers if backing storage is backed by the network
3510 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3511 * depleted. kswapd will continue to make progress and wake the processes
3512 * when the low watermark is reached.
3513 *
3514 * Returns true if a fatal signal was delivered during throttling. If this
3515 * happens, the page allocator should not consider triggering the OOM killer.
3516 */
3517static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3518                                        nodemask_t *nodemask)
3519{
3520        struct zoneref *z;
3521        struct zone *zone;
3522        pg_data_t *pgdat = NULL;
3523
3524        /*
3525         * Kernel threads should not be throttled as they may be indirectly
3526         * responsible for cleaning pages necessary for reclaim to make forward
3527         * progress. kjournald for example may enter direct reclaim while
3528         * committing a transaction where throttling it could forcing other
3529         * processes to block on log_wait_commit().
3530         */
3531        if (current->flags & PF_KTHREAD)
3532                goto out;
3533
3534        /*
3535         * If a fatal signal is pending, this process should not throttle.
3536         * It should return quickly so it can exit and free its memory
3537         */
3538        if (fatal_signal_pending(current))
3539                goto out;
3540
3541        /*
3542         * Check if the pfmemalloc reserves are ok by finding the first node
3543         * with a usable ZONE_NORMAL or lower zone. The expectation is that
3544         * GFP_KERNEL will be required for allocating network buffers when
3545         * swapping over the network so ZONE_HIGHMEM is unusable.
3546         *
3547         * Throttling is based on the first usable node and throttled processes
3548         * wait on a queue until kswapd makes progress and wakes them. There
3549         * is an affinity then between processes waking up and where reclaim
3550         * progress has been made assuming the process wakes on the same node.
3551         * More importantly, processes running on remote nodes will not compete
3552         * for remote pfmemalloc reserves and processes on different nodes
3553         * should make reasonable progress.
3554         */
3555        for_each_zone_zonelist_nodemask(zone, z, zonelist,
3556                                        gfp_zone(gfp_mask), nodemask) {
3557                if (zone_idx(zone) > ZONE_NORMAL)
3558                        continue;
3559
3560                /* Throttle based on the first usable node */
3561                pgdat = zone->zone_pgdat;
3562                if (allow_direct_reclaim(pgdat))
3563                        goto out;
3564                break;
3565        }
3566
3567        /* If no zone was usable by the allocation flags then do not throttle */
3568        if (!pgdat)
3569                goto out;
3570
3571        /* Account for the throttling */
3572        count_vm_event(PGSCAN_DIRECT_THROTTLE);
3573
3574        /*
3575         * If the caller cannot enter the filesystem, it's possible that it
3576         * is due to the caller holding an FS lock or performing a journal
3577         * transaction in the case of a filesystem like ext[3|4]. In this case,
3578         * it is not safe to block on pfmemalloc_wait as kswapd could be
3579         * blocked waiting on the same lock. Instead, throttle for up to a
3580         * second before continuing.
3581         */
3582        if (!(gfp_mask & __GFP_FS))
3583                wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3584                        allow_direct_reclaim(pgdat), HZ);
3585        else
3586                /* Throttle until kswapd wakes the process */
3587                wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3588                        allow_direct_reclaim(pgdat));
3589
3590        if (fatal_signal_pending(current))
3591                return true;
3592
3593out:
3594        return false;
3595}
3596
3597unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3598                                gfp_t gfp_mask, nodemask_t *nodemask)
3599{
3600        unsigned long nr_reclaimed;
3601        struct scan_control sc = {
3602                .nr_to_reclaim = SWAP_CLUSTER_MAX,
3603                .gfp_mask = current_gfp_context(gfp_mask),
3604                .reclaim_idx = gfp_zone(gfp_mask),
3605                .order = order,
3606                .nodemask = nodemask,
3607                .priority = DEF_PRIORITY,
3608                .may_writepage = !laptop_mode,
3609                .may_unmap = 1,
3610                .may_swap = 1,
3611        };
3612
3613        /*
3614         * scan_control uses s8 fields for order, priority, and reclaim_idx.
3615         * Confirm they are large enough for max values.
3616         */
3617        BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3618        BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3619        BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3620
3621        /*
3622         * Do not enter reclaim if fatal signal was delivered while throttled.
3623         * 1 is returned so that the page allocator does not OOM kill at this
3624         * point.
3625         */
3626        if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3627                return 1;
3628
3629        set_task_reclaim_state(current, &sc.reclaim_state);
3630        trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3631
3632        nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3633
3634        trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3635        set_task_reclaim_state(current, NULL);
3636
3637        return nr_reclaimed;
3638}
3639
3640#ifdef CONFIG_MEMCG
3641
3642/* Only used by soft limit reclaim. Do not reuse for anything else. */
3643unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3644                                                gfp_t gfp_mask, bool noswap,
3645                                                pg_data_t *pgdat,
3646                                                unsigned long *nr_scanned)
3647{
3648        struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3649        struct scan_control sc = {
3650                .nr_to_reclaim = SWAP_CLUSTER_MAX,
3651                .target_mem_cgroup = memcg,
3652                .may_writepage = !laptop_mode,
3653                .may_unmap = 1,
3654                .reclaim_idx = MAX_NR_ZONES - 1,
3655                .may_swap = !noswap,
3656        };
3657
3658        WARN_ON_ONCE(!current->reclaim_state);
3659
3660        sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3661                        (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3662
3663        trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3664                                                      sc.gfp_mask);
3665
3666        /*
3667         * NOTE: Although we can get the priority field, using it
3668         * here is not a good idea, since it limits the pages we can scan.
3669         * if we don't reclaim here, the shrink_node from balance_pgdat
3670         * will pick up pages from other mem cgroup's as well. We hack
3671         * the priority and make it zero.
3672         */
3673        shrink_lruvec(lruvec, &sc);
3674
3675        trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3676
3677        *nr_scanned = sc.nr_scanned;
3678
3679        return sc.nr_reclaimed;
3680}
3681
3682unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3683                                           unsigned long nr_pages,
3684                                           gfp_t gfp_mask,
3685                                           bool may_swap)
3686{
3687        unsigned long nr_reclaimed;
3688        unsigned int noreclaim_flag;
3689        struct scan_control sc = {
3690                .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3691                .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3692                                (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3693                .reclaim_idx = MAX_NR_ZONES - 1,
3694                .target_mem_cgroup = memcg,
3695                .priority = DEF_PRIORITY,
3696                .may_writepage = !laptop_mode,
3697                .may_unmap = 1,
3698                .may_swap = may_swap,
3699        };
3700        /*
3701         * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3702         * equal pressure on all the nodes. This is based on the assumption that
3703         * the reclaim does not bail out early.
3704         */
3705        struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3706
3707        set_task_reclaim_state(current, &sc.reclaim_state);
3708        trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3709        noreclaim_flag = memalloc_noreclaim_save();
3710
3711        nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3712
3713        memalloc_noreclaim_restore(noreclaim_flag);
3714        trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3715        set_task_reclaim_state(current, NULL);
3716
3717        return nr_reclaimed;
3718}
3719#endif
3720
3721static void age_active_anon(struct pglist_data *pgdat,
3722                                struct scan_control *sc)
3723{
3724        struct mem_cgroup *memcg;
3725        struct lruvec *lruvec;
3726
3727        if (!can_age_anon_pages(pgdat, sc))
3728                return;
3729
3730        lruvec = mem_cgroup_lruvec(NULL, pgdat);
3731        if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3732                return;
3733
3734        memcg = mem_cgroup_iter(NULL, NULL, NULL);
3735        do {
3736                lruvec = mem_cgroup_lruvec(memcg, pgdat);
3737                shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3738                                   sc, LRU_ACTIVE_ANON);
3739                memcg = mem_cgroup_iter(NULL, memcg, NULL);
3740        } while (memcg);
3741}
3742
3743static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3744{
3745        int i;
3746        struct zone *zone;
3747
3748        /*
3749         * Check for watermark boosts top-down as the higher zones
3750         * are more likely to be boosted. Both watermarks and boosts
3751         * should not be checked at the same time as reclaim would
3752         * start prematurely when there is no boosting and a lower
3753         * zone is balanced.
3754         */
3755        for (i = highest_zoneidx; i >= 0; i--) {
3756                zone = pgdat->node_zones + i;
3757                if (!managed_zone(zone))
3758                        continue;
3759
3760                if (zone->watermark_boost)
3761                        return true;
3762        }
3763
3764        return false;
3765}
3766
3767/*
3768 * Returns true if there is an eligible zone balanced for the request order
3769 * and highest_zoneidx
3770 */
3771static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3772{
3773        int i;
3774        unsigned long mark = -1;
3775        struct zone *zone;
3776
3777        /*
3778         * Check watermarks bottom-up as lower zones are more likely to
3779         * meet watermarks.
3780         */
3781        for (i = 0; i <= highest_zoneidx; i++) {
3782                zone = pgdat->node_zones + i;
3783
3784                if (!managed_zone(zone))
3785                        continue;
3786
3787                mark = high_wmark_pages(zone);
3788                if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3789                        return true;
3790        }
3791
3792        /*
3793         * If a node has no populated zone within highest_zoneidx, it does not
3794         * need balancing by definition. This can happen if a zone-restricted
3795         * allocation tries to wake a remote kswapd.
3796         */
3797        if (mark == -1)
3798                return true;
3799
3800        return false;
3801}
3802
3803/* Clear pgdat state for congested, dirty or under writeback. */
3804static void clear_pgdat_congested(pg_data_t *pgdat)
3805{
3806        struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3807
3808        clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3809        clear_bit(PGDAT_DIRTY, &pgdat->flags);
3810        clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3811}
3812
3813/*
3814 * Prepare kswapd for sleeping. This verifies that there are no processes
3815 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3816 *
3817 * Returns true if kswapd is ready to sleep
3818 */
3819static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3820                                int highest_zoneidx)
3821{
3822        /*
3823         * The throttled processes are normally woken up in balance_pgdat() as
3824         * soon as allow_direct_reclaim() is true. But there is a potential
3825         * race between when kswapd checks the watermarks and a process gets
3826         * throttled. There is also a potential race if processes get
3827         * throttled, kswapd wakes, a large process exits thereby balancing the
3828         * zones, which causes kswapd to exit balance_pgdat() before reaching
3829         * the wake up checks. If kswapd is going to sleep, no process should
3830         * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3831         * the wake up is premature, processes will wake kswapd and get
3832         * throttled again. The difference from wake ups in balance_pgdat() is
3833         * that here we are under prepare_to_wait().
3834         */
3835        if (waitqueue_active(&pgdat->pfmemalloc_wait))
3836                wake_up_all(&pgdat->pfmemalloc_wait);
3837
3838        /* Hopeless node, leave it to direct reclaim */
3839        if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3840                return true;
3841
3842        if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
3843                clear_pgdat_congested(pgdat);
3844                return true;
3845        }
3846
3847        return false;
3848}
3849
3850/*
3851 * kswapd shrinks a node of pages that are at or below the highest usable
3852 * zone that is currently unbalanced.
3853 *
3854 * Returns true if kswapd scanned at least the requested number of pages to
3855 * reclaim or if the lack of progress was due to pages under writeback.
3856 * This is used to determine if the scanning priority needs to be raised.
3857 */
3858static bool kswapd_shrink_node(pg_data_t *pgdat,
3859                               struct scan_control *sc)
3860{
3861        struct zone *zone;
3862        int z;
3863
3864        /* Reclaim a number of pages proportional to the number of zones */
3865        sc->nr_to_reclaim = 0;
3866        for (z = 0; z <= sc->reclaim_idx; z++) {
3867                zone = pgdat->node_zones + z;
3868                if (!managed_zone(zone))
3869                        continue;
3870
3871                sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3872        }
3873
3874        /*
3875         * Historically care was taken to put equal pressure on all zones but
3876         * now pressure is applied based on node LRU order.
3877         */
3878        shrink_node(pgdat, sc);
3879
3880        /*
3881         * Fragmentation may mean that the system cannot be rebalanced for
3882         * high-order allocations. If twice the allocation size has been
3883         * reclaimed then recheck watermarks only at order-0 to prevent
3884         * excessive reclaim. Assume that a process requested a high-order
3885         * can direct reclaim/compact.
3886         */
3887        if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3888                sc->order = 0;
3889
3890        return sc->nr_scanned >= sc->nr_to_reclaim;
3891}
3892
3893/* Page allocator PCP high watermark is lowered if reclaim is active. */
3894static inline void
3895update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
3896{
3897        int i;
3898        struct zone *zone;
3899
3900        for (i = 0; i <= highest_zoneidx; i++) {
3901                zone = pgdat->node_zones + i;
3902
3903                if (!managed_zone(zone))
3904                        continue;
3905
3906                if (active)
3907                        set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
3908                else
3909                        clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
3910        }
3911}
3912
3913static inline void
3914set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
3915{
3916        update_reclaim_active(pgdat, highest_zoneidx, true);
3917}
3918
3919static inline void
3920clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
3921{
3922        update_reclaim_active(pgdat, highest_zoneidx, false);
3923}
3924
3925/*
3926 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3927 * that are eligible for use by the caller until at least one zone is
3928 * balanced.
3929 *
3930 * Returns the order kswapd finished reclaiming at.
3931 *
3932 * kswapd scans the zones in the highmem->normal->dma direction.  It skips
3933 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3934 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3935 * or lower is eligible for reclaim until at least one usable zone is
3936 * balanced.
3937 */
3938static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
3939{
3940        int i;
3941        unsigned long nr_soft_reclaimed;
3942        unsigned long nr_soft_scanned;
3943        unsigned long pflags;
3944        unsigned long nr_boost_reclaim;
3945        unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3946        bool boosted;
3947        struct zone *zone;
3948        struct scan_control sc = {
3949                .gfp_mask = GFP_KERNEL,
3950                .order = order,
3951                .may_unmap = 1,
3952        };
3953
3954        set_task_reclaim_state(current, &sc.reclaim_state);
3955        psi_memstall_enter(&pflags);
3956        __fs_reclaim_acquire(_THIS_IP_);
3957
3958        count_vm_event(PAGEOUTRUN);
3959
3960        /*
3961         * Account for the reclaim boost. Note that the zone boost is left in
3962         * place so that parallel allocations that are near the watermark will
3963         * stall or direct reclaim until kswapd is finished.
3964         */
3965        nr_boost_reclaim = 0;
3966        for (i = 0; i <= highest_zoneidx; i++) {
3967                zone = pgdat->node_zones + i;
3968                if (!managed_zone(zone))
3969                        continue;
3970
3971                nr_boost_reclaim += zone->watermark_boost;
3972                zone_boosts[i] = zone->watermark_boost;
3973        }
3974        boosted = nr_boost_reclaim;
3975
3976restart:
3977        set_reclaim_active(pgdat, highest_zoneidx);
3978        sc.priority = DEF_PRIORITY;
3979        do {
3980                unsigned long nr_reclaimed = sc.nr_reclaimed;
3981                bool raise_priority = true;
3982                bool balanced;
3983                bool ret;
3984
3985                sc.reclaim_idx = highest_zoneidx;
3986
3987                /*
3988                 * If the number of buffer_heads exceeds the maximum allowed
3989                 * then consider reclaiming from all zones. This has a dual
3990                 * purpose -- on 64-bit systems it is expected that
3991                 * buffer_heads are stripped during active rotation. On 32-bit
3992                 * systems, highmem pages can pin lowmem memory and shrinking
3993                 * buffers can relieve lowmem pressure. Reclaim may still not
3994                 * go ahead if all eligible zones for the original allocation
3995                 * request are balanced to avoid excessive reclaim from kswapd.
3996                 */
3997                if (buffer_heads_over_limit) {
3998                        for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3999                                zone = pgdat->node_zones + i;
4000                                if (!managed_zone(zone))
4001                                        continue;
4002
4003                                sc.reclaim_idx = i;
4004                                break;
4005                        }
4006                }
4007
4008                /*
4009                 * If the pgdat is imbalanced then ignore boosting and preserve
4010                 * the watermarks for a later time and restart. Note that the
4011                 * zone watermarks will be still reset at the end of balancing
4012                 * on the grounds that the normal reclaim should be enough to
4013                 * re-evaluate if boosting is required when kswapd next wakes.
4014                 */
4015                balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
4016                if (!balanced && nr_boost_reclaim) {
4017                        nr_boost_reclaim = 0;
4018                        goto restart;
4019                }
4020
4021                /*
4022                 * If boosting is not active then only reclaim if there are no
4023                 * eligible zones. Note that sc.reclaim_idx is not used as
4024                 * buffer_heads_over_limit may have adjusted it.
4025                 */
4026                if (!nr_boost_reclaim && balanced)
4027                        goto out;
4028
4029                /* Limit the priority of boosting to avoid reclaim writeback */
4030                if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
4031                        raise_priority = false;
4032
4033                /*
4034                 * Do not writeback or swap pages for boosted reclaim. The
4035                 * intent is to relieve pressure not issue sub-optimal IO
4036                 * from reclaim context. If no pages are reclaimed, the
4037                 * reclaim will be aborted.
4038                 */
4039                sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
4040                sc.may_swap = !nr_boost_reclaim;
4041
4042                /*
4043                 * Do some background aging of the anon list, to give
4044                 * pages a chance to be referenced before reclaiming. All
4045                 * pages are rotated regardless of classzone as this is
4046                 * about consistent aging.
4047                 */
4048                age_active_anon(pgdat, &sc);
4049
4050                /*
4051                 * If we're getting trouble reclaiming, start doing writepage
4052                 * even in laptop mode.
4053                 */
4054                if (sc.priority < DEF_PRIORITY - 2)
4055                        sc.may_writepage = 1;
4056
4057                /* Call soft limit reclaim before calling shrink_node. */
4058                sc.nr_scanned = 0;
4059                nr_soft_scanned = 0;
4060                nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
4061                                                sc.gfp_mask, &nr_soft_scanned);
4062                sc.nr_reclaimed += nr_soft_reclaimed;
4063
4064                /*
4065                 * There should be no need to raise the scanning priority if
4066                 * enough pages are already being scanned that that high
4067                 * watermark would be met at 100% efficiency.
4068                 */
4069                if (kswapd_shrink_node(pgdat, &sc))
4070                        raise_priority = false;
4071
4072                /*
4073                 * If the low watermark is met there is no need for processes
4074                 * to be throttled on pfmemalloc_wait as they should not be
4075                 * able to safely make forward progress. Wake them
4076                 */
4077                if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
4078                                allow_direct_reclaim(pgdat))
4079                        wake_up_all(&pgdat->pfmemalloc_wait);
4080
4081                /* Check if kswapd should be suspending */
4082                __fs_reclaim_release(_THIS_IP_);
4083                ret = try_to_freeze();
4084                __fs_reclaim_acquire(_THIS_IP_);
4085                if (ret || kthread_should_stop())
4086                        break;
4087
4088                /*
4089                 * Raise priority if scanning rate is too low or there was no
4090                 * progress in reclaiming pages
4091                 */
4092                nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
4093                nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
4094
4095                /*
4096                 * If reclaim made no progress for a boost, stop reclaim as
4097                 * IO cannot be queued and it could be an infinite loop in
4098                 * extreme circumstances.
4099                 */
4100                if (nr_boost_reclaim && !nr_reclaimed)
4101                        break;
4102
4103                if (raise_priority || !nr_reclaimed)
4104                        sc.priority--;
4105        } while (sc.priority >= 1);
4106
4107        if (!sc.nr_reclaimed)
4108                pgdat->kswapd_failures++;
4109
4110out:
4111        clear_reclaim_active(pgdat, highest_zoneidx);
4112
4113        /* If reclaim was boosted, account for the reclaim done in this pass */
4114        if (boosted) {
4115                unsigned long flags;
4116
4117                for (i = 0; i <= highest_zoneidx; i++) {
4118                        if (!zone_boosts[i])
4119                                continue;
4120
4121                        /* Increments are under the zone lock */
4122                        zone = pgdat->node_zones + i;
4123                        spin_lock_irqsave(&zone->lock, flags);
4124                        zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
4125                        spin_unlock_irqrestore(&zone->lock, flags);
4126                }
4127
4128                /*
4129                 * As there is now likely space, wakeup kcompact to defragment
4130                 * pageblocks.
4131                 */
4132                wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
4133        }
4134
4135        snapshot_refaults(NULL, pgdat);
4136        __fs_reclaim_release(_THIS_IP_);
4137        psi_memstall_leave(&pflags);
4138        set_task_reclaim_state(current, NULL);
4139
4140        /*
4141         * Return the order kswapd stopped reclaiming at as
4142         * prepare_kswapd_sleep() takes it into account. If another caller
4143         * entered the allocator slow path while kswapd was awake, order will
4144         * remain at the higher level.
4145         */
4146        return sc.order;
4147}
4148
4149/*
4150 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
4151 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
4152 * not a valid index then either kswapd runs for first time or kswapd couldn't
4153 * sleep after previous reclaim attempt (node is still unbalanced). In that
4154 * case return the zone index of the previous kswapd reclaim cycle.
4155 */
4156static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
4157                                           enum zone_type prev_highest_zoneidx)
4158{
4159        enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
4160
4161        return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
4162}
4163
4164static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
4165                                unsigned int highest_zoneidx)
4166{
4167        long remaining = 0;
4168        DEFINE_WAIT(wait);
4169
4170        if (freezing(current) || kthread_should_stop())
4171                return;
4172
4173        prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
4174
4175        /*
4176         * Try to sleep for a short interval. Note that kcompactd will only be
4177         * woken if it is possible to sleep for a short interval. This is
4178         * deliberate on the assumption that if reclaim cannot keep an
4179         * eligible zone balanced that it's also unlikely that compaction will
4180         * succeed.
4181         */
4182        if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
4183                /*
4184                 * Compaction records what page blocks it recently failed to
4185                 * isolate pages from and skips them in the future scanning.
4186                 * When kswapd is going to sleep, it is reasonable to assume
4187                 * that pages and compaction may succeed so reset the cache.
4188                 */
4189                reset_isolation_suitable(pgdat);
4190
4191                /*
4192                 * We have freed the memory, now we should compact it to make
4193                 * allocation of the requested order possible.
4194                 */
4195                wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
4196
4197                remaining = schedule_timeout(HZ/10);
4198
4199                /*
4200                 * If woken prematurely then reset kswapd_highest_zoneidx and
4201                 * order. The values will either be from a wakeup request or
4202                 * the previous request that slept prematurely.
4203                 */
4204                if (remaining) {
4205                        WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
4206                                        kswapd_highest_zoneidx(pgdat,
4207                                                        highest_zoneidx));
4208
4209                        if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
4210                                WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
4211                }
4212
4213                finish_wait(&pgdat->kswapd_wait, &wait);
4214                prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
4215        }
4216
4217        /*
4218         * After a short sleep, check if it was a premature sleep. If not, then
4219         * go fully to sleep until explicitly woken up.
4220         */
4221        if (!remaining &&
4222            prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
4223                trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
4224
4225                /*
4226                 * vmstat counters are not perfectly accurate and the estimated
4227                 * value for counters such as NR_FREE_PAGES can deviate from the
4228                 * true value by nr_online_cpus * threshold. To avoid the zone
4229                 * watermarks being breached while under pressure, we reduce the
4230                 * per-cpu vmstat threshold while kswapd is awake and restore
4231                 * them before going back to sleep.
4232                 */
4233                set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
4234
4235                if (!kthread_should_stop())
4236                        schedule();
4237
4238                set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
4239        } else {
4240                if (remaining)
4241                        count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
4242                else
4243                        count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
4244        }
4245        finish_wait(&pgdat->kswapd_wait, &wait);
4246}
4247
4248/*
4249 * The background pageout daemon, started as a kernel thread
4250 * from the init process.
4251 *
4252 * This basically trickles out pages so that we have _some_
4253 * free memory available even if there is no other activity
4254 * that frees anything up. This is needed for things like routing
4255 * etc, where we otherwise might have all activity going on in
4256 * asynchronous contexts that cannot page things out.
4257 *
4258 * If there are applications that are active memory-allocators
4259 * (most normal use), this basically shouldn't matter.
4260 */
4261static int kswapd(void *p)
4262{
4263        unsigned int alloc_order, reclaim_order;
4264        unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
4265        pg_data_t *pgdat = (pg_data_t *)p;
4266        struct task_struct *tsk = current;
4267        const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
4268
4269        if (!cpumask_empty(cpumask))
4270                set_cpus_allowed_ptr(tsk, cpumask);
4271
4272        /*
4273         * Tell the memory management that we're a "memory allocator",
4274         * and that if we need more memory we should get access to it
4275         * regardless (see "__alloc_pages()"). "kswapd" should
4276         * never get caught in the normal page freeing logic.
4277         *
4278         * (Kswapd normally doesn't need memory anyway, but sometimes
4279         * you need a small amount of memory in order to be able to
4280         * page out something else, and this flag essentially protects
4281         * us from recursively trying to free more memory as we're
4282         * trying to free the first piece of memory in the first place).
4283         */
4284        tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
4285        set_freezable();
4286
4287        WRITE_ONCE(pgdat->kswapd_order, 0);
4288        WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4289        for ( ; ; ) {
4290                bool ret;
4291
4292                alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
4293                highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4294                                                        highest_zoneidx);
4295
4296kswapd_try_sleep:
4297                kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
4298                                        highest_zoneidx);
4299
4300                /* Read the new order and highest_zoneidx */
4301                alloc_order = READ_ONCE(pgdat->kswapd_order);
4302                highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4303                                                        highest_zoneidx);
4304                WRITE_ONCE(pgdat->kswapd_order, 0);
4305                WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4306
4307                ret = try_to_freeze();
4308                if (kthread_should_stop())
4309                        break;
4310
4311                /*
4312                 * We can speed up thawing tasks if we don't call balance_pgdat
4313                 * after returning from the refrigerator
4314                 */
4315                if (ret)
4316                        continue;
4317
4318                /*
4319                 * Reclaim begins at the requested order but if a high-order
4320                 * reclaim fails then kswapd falls back to reclaiming for
4321                 * order-0. If that happens, kswapd will consider sleeping
4322                 * for the order it finished reclaiming at (reclaim_order)
4323                 * but kcompactd is woken to compact for the original
4324                 * request (alloc_order).
4325                 */
4326                trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
4327                                                alloc_order);
4328                reclaim_order = balance_pgdat(pgdat, alloc_order,
4329                                                highest_zoneidx);
4330                if (reclaim_order < alloc_order)
4331                        goto kswapd_try_sleep;
4332        }
4333
4334        tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
4335
4336        return 0;
4337}
4338
4339/*
4340 * A zone is low on free memory or too fragmented for high-order memory.  If
4341 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
4342 * pgdat.  It will wake up kcompactd after reclaiming memory.  If kswapd reclaim
4343 * has failed or is not needed, still wake up kcompactd if only compaction is
4344 * needed.
4345 */
4346void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
4347                   enum zone_type highest_zoneidx)
4348{
4349        pg_data_t *pgdat;
4350        enum zone_type curr_idx;
4351
4352        if (!managed_zone(zone))
4353                return;
4354
4355        if (!cpuset_zone_allowed(zone, gfp_flags))
4356                return;
4357
4358        pgdat = zone->zone_pgdat;
4359        curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
4360
4361        if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
4362                WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
4363
4364        if (READ_ONCE(pgdat->kswapd_order) < order)
4365                WRITE_ONCE(pgdat->kswapd_order, order);
4366
4367        if (!waitqueue_active(&pgdat->kswapd_wait))
4368                return;
4369
4370        /* Hopeless node, leave it to direct reclaim if possible */
4371        if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
4372            (pgdat_balanced(pgdat, order, highest_zoneidx) &&
4373             !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
4374                /*
4375                 * There may be plenty of free memory available, but it's too
4376                 * fragmented for high-order allocations.  Wake up kcompactd
4377                 * and rely on compaction_suitable() to determine if it's
4378                 * needed.  If it fails, it will defer subsequent attempts to
4379                 * ratelimit its work.
4380                 */
4381                if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
4382                        wakeup_kcompactd(pgdat, order, highest_zoneidx);
4383                return;
4384        }
4385
4386        trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
4387                                      gfp_flags);
4388        wake_up_interruptible(&pgdat->kswapd_wait);
4389}
4390
4391#ifdef CONFIG_HIBERNATION
4392/*
4393 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4394 * freed pages.
4395 *
4396 * Rather than trying to age LRUs the aim is to preserve the overall
4397 * LRU order by reclaiming preferentially
4398 * inactive > active > active referenced > active mapped
4399 */
4400unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4401{
4402        struct scan_control sc = {
4403                .nr_to_reclaim = nr_to_reclaim,
4404                .gfp_mask = GFP_HIGHUSER_MOVABLE,
4405                .reclaim_idx = MAX_NR_ZONES - 1,
4406                .priority = DEF_PRIORITY,
4407                .may_writepage = 1,
4408                .may_unmap = 1,
4409                .may_swap = 1,
4410                .hibernation_mode = 1,
4411        };
4412        struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4413        unsigned long nr_reclaimed;
4414        unsigned int noreclaim_flag;
4415
4416        fs_reclaim_acquire(sc.gfp_mask);
4417        noreclaim_flag = memalloc_noreclaim_save();
4418        set_task_reclaim_state(current, &sc.reclaim_state);
4419
4420        nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4421
4422        set_task_reclaim_state(current, NULL);
4423        memalloc_noreclaim_restore(noreclaim_flag);
4424        fs_reclaim_release(sc.gfp_mask);
4425
4426        return nr_reclaimed;
4427}
4428#endif /* CONFIG_HIBERNATION */
4429
4430/*
4431 * This kswapd start function will be called by init and node-hot-add.
4432 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4433 */
4434void kswapd_run(int nid)
4435{
4436        pg_data_t *pgdat = NODE_DATA(nid);
4437
4438        if (pgdat->kswapd)
4439                return;
4440
4441        pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4442        if (IS_ERR(pgdat->kswapd)) {
4443                /* failure at boot is fatal */
4444                BUG_ON(system_state < SYSTEM_RUNNING);
4445                pr_err("Failed to start kswapd on node %d\n", nid);
4446                pgdat->kswapd = NULL;
4447        }
4448}
4449
4450/*
4451 * Called by memory hotplug when all memory in a node is offlined.  Caller must
4452 * hold mem_hotplug_begin/end().
4453 */
4454void kswapd_stop(int nid)
4455{
4456        struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4457
4458        if (kswapd) {
4459                kthread_stop(kswapd);
4460                NODE_DATA(nid)->kswapd = NULL;
4461        }
4462}
4463
4464static int __init kswapd_init(void)
4465{
4466        int nid;
4467
4468        swap_setup();
4469        for_each_node_state(nid, N_MEMORY)
4470                kswapd_run(nid);
4471        return 0;
4472}
4473
4474module_init(kswapd_init)
4475
4476#ifdef CONFIG_NUMA
4477/*
4478 * Node reclaim mode
4479 *
4480 * If non-zero call node_reclaim when the number of free pages falls below
4481 * the watermarks.
4482 */
4483int node_reclaim_mode __read_mostly;
4484
4485/*
4486 * Priority for NODE_RECLAIM. This determines the fraction of pages
4487 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4488 * a zone.
4489 */
4490#define NODE_RECLAIM_PRIORITY 4
4491
4492/*
4493 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4494 * occur.
4495 */
4496int sysctl_min_unmapped_ratio = 1;
4497
4498/*
4499 * If the number of slab pages in a zone grows beyond this percentage then
4500 * slab reclaim needs to occur.
4501 */
4502int sysctl_min_slab_ratio = 5;
4503
4504static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4505{
4506        unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4507        unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4508                node_page_state(pgdat, NR_ACTIVE_FILE);
4509
4510        /*
4511         * It's possible for there to be more file mapped pages than
4512         * accounted for by the pages on the file LRU lists because
4513         * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4514         */
4515        return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4516}
4517
4518/* Work out how many page cache pages we can reclaim in this reclaim_mode */
4519static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4520{
4521        unsigned long nr_pagecache_reclaimable;
4522        unsigned long delta = 0;
4523
4524        /*
4525         * If RECLAIM_UNMAP is set, then all file pages are considered
4526         * potentially reclaimable. Otherwise, we have to worry about
4527         * pages like swapcache and node_unmapped_file_pages() provides
4528         * a better estimate
4529         */
4530        if (node_reclaim_mode & RECLAIM_UNMAP)
4531                nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4532        else
4533                nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4534
4535        /* If we can't clean pages, remove dirty pages from consideration */
4536        if (!(node_reclaim_mode & RECLAIM_WRITE))
4537                delta += node_page_state(pgdat, NR_FILE_DIRTY);
4538
4539        /* Watch for any possible underflows due to delta */
4540        if (unlikely(delta > nr_pagecache_reclaimable))
4541                delta = nr_pagecache_reclaimable;
4542
4543        return nr_pagecache_reclaimable - delta;
4544}
4545
4546/*
4547 * Try to free up some pages from this node through reclaim.
4548 */
4549static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4550{
4551        /* Minimum pages needed in order to stay on node */
4552        const unsigned long nr_pages = 1 << order;
4553        struct task_struct *p = current;
4554        unsigned int noreclaim_flag;
4555        struct scan_control sc = {
4556                .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4557                .gfp_mask = current_gfp_context(gfp_mask),
4558                .order = order,
4559                .priority = NODE_RECLAIM_PRIORITY,
4560                .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4561                .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4562                .may_swap = 1,
4563                .reclaim_idx = gfp_zone(gfp_mask),
4564        };
4565        unsigned long pflags;
4566
4567        trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4568                                           sc.gfp_mask);
4569
4570        cond_resched();
4571        psi_memstall_enter(&pflags);
4572        fs_reclaim_acquire(sc.gfp_mask);
4573        /*
4574         * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4575         * and we also need to be able to write out pages for RECLAIM_WRITE
4576         * and RECLAIM_UNMAP.
4577         */
4578        noreclaim_flag = memalloc_noreclaim_save();
4579        p->flags |= PF_SWAPWRITE;
4580        set_task_reclaim_state(p, &sc.reclaim_state);
4581
4582        if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4583                /*
4584                 * Free memory by calling shrink node with increasing
4585                 * priorities until we have enough memory freed.
4586                 */
4587                do {
4588                        shrink_node(pgdat, &sc);
4589                } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4590        }
4591
4592        set_task_reclaim_state(p, NULL);
4593        current->flags &= ~PF_SWAPWRITE;
4594        memalloc_noreclaim_restore(noreclaim_flag);
4595        fs_reclaim_release(sc.gfp_mask);
4596        psi_memstall_leave(&pflags);
4597
4598        trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4599
4600        return sc.nr_reclaimed >= nr_pages;
4601}
4602
4603int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4604{
4605        int ret;
4606
4607        /*
4608         * Node reclaim reclaims unmapped file backed pages and
4609         * slab pages if we are over the defined limits.
4610         *
4611         * A small portion of unmapped file backed pages is needed for
4612         * file I/O otherwise pages read by file I/O will be immediately
4613         * thrown out if the node is overallocated. So we do not reclaim
4614         * if less than a specified percentage of the node is used by
4615         * unmapped file backed pages.
4616         */
4617        if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4618            node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4619            pgdat->min_slab_pages)
4620                return NODE_RECLAIM_FULL;
4621
4622        /*
4623         * Do not scan if the allocation should not be delayed.
4624         */
4625        if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4626                return NODE_RECLAIM_NOSCAN;
4627
4628        /*
4629         * Only run node reclaim on the local node or on nodes that do not
4630         * have associated processors. This will favor the local processor
4631         * over remote processors and spread off node memory allocations
4632         * as wide as possible.
4633         */
4634        if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4635                return NODE_RECLAIM_NOSCAN;
4636
4637        if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4638                return NODE_RECLAIM_NOSCAN;
4639
4640        ret = __node_reclaim(pgdat, gfp_mask, order);
4641        clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4642
4643        if (!ret)
4644                count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4645
4646        return ret;
4647}
4648#endif
4649
4650/**
4651 * check_move_unevictable_pages - check pages for evictability and move to
4652 * appropriate zone lru list
4653 * @pvec: pagevec with lru pages to check
4654 *
4655 * Checks pages for evictability, if an evictable page is in the unevictable
4656 * lru list, moves it to the appropriate evictable lru list. This function
4657 * should be only used for lru pages.
4658 */
4659void check_move_unevictable_pages(struct pagevec *pvec)
4660{
4661        struct lruvec *lruvec = NULL;
4662        int pgscanned = 0;
4663        int pgrescued = 0;
4664        int i;
4665
4666        for (i = 0; i < pvec->nr; i++) {
4667                struct page *page = pvec->pages[i];
4668                int nr_pages;
4669
4670                if (PageTransTail(page))
4671                        continue;
4672
4673                nr_pages = thp_nr_pages(page);
4674                pgscanned += nr_pages;
4675
4676                /* block memcg migration during page moving between lru */
4677                if (!TestClearPageLRU(page))
4678                        continue;
4679
4680                lruvec = relock_page_lruvec_irq(page, lruvec);
4681                if (page_evictable(page) && PageUnevictable(page)) {
4682                        del_page_from_lru_list(page, lruvec);
4683                        ClearPageUnevictable(page);
4684                        add_page_to_lru_list(page, lruvec);
4685                        pgrescued += nr_pages;
4686                }
4687                SetPageLRU(page);
4688        }
4689
4690        if (lruvec) {
4691                __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4692                __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4693                unlock_page_lruvec_irq(lruvec);
4694        } else if (pgscanned) {
4695                count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4696        }
4697}
4698EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
4699