mm/swap.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  linux/mm/swap.c
   4  *
   5  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   6  */
   7
   8 /*
   9  * This file contains the default values for the operation of the
  10  * Linux VM subsystem. Fine-tuning documentation can be found in
  11  * Documentation/admin-guide/sysctl/vm.rst.
  12  * Started 18.12.91
  13  * Swap aging added 23.2.95, Stephen Tweedie.
  14  * Buffermem limits added 12.3.98, Rik van Riel.
  15  */
  16
  17 #include <linux/mm.h>
  18 #include <linux/sched.h>
  19 #include <linux/kernel_stat.h>
  20 #include <linux/swap.h>
  21 #include <linux/mman.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/pagevec.h>
  24 #include <linux/init.h>
  25 #include <linux/export.h>
  26 #include <linux/mm_inline.h>
  27 #include <linux/percpu_counter.h>
  28 #include <linux/memremap.h>
  29 #include <linux/percpu.h>
  30 #include <linux/cpu.h>
  31 #include <linux/notifier.h>
  32 #include <linux/backing-dev.h>
  33 #include <linux/memcontrol.h>
  34 #include <linux/gfp.h>
  35 #include <linux/uio.h>
  36 #include <linux/hugetlb.h>
  37 #include <linux/page_idle.h>
  38 #include <linux/local_lock.h>
  39
  40 #include "internal.h"
  41
  42 #define CREATE_TRACE_POINTS
  43 #include <trace/events/pagemap.h>
  44
  45 /* How many pages do we try to swap or page in/out together? */
  46 int page_cluster;
  47
  48 /* Protecting only lru_rotate.pvec which requires disabling interrupts */
  49 struct lru_rotate {
  50         local_lock_t lock;
  51         struct pagevec pvec;
  52 };
  53 static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
  54         .lock = INIT_LOCAL_LOCK(lock),
  55 };
  56
  57 /*
  58  * The following struct pagevec are grouped together because they are protected
  59  * by disabling preemption (and interrupts remain enabled).
  60  */
  61 struct lru_pvecs {
  62         local_lock_t lock;
  63         struct pagevec lru_add;
  64         struct pagevec lru_deactivate_file;
  65         struct pagevec lru_deactivate;
  66         struct pagevec lru_lazyfree;
  67 #ifdef CONFIG_SMP
  68         struct pagevec activate_page;
  69 #endif
  70 };
  71 static DEFINE_PER_CPU(struct lru_pvecs, lru_pvecs) = {
  72         .lock = INIT_LOCAL_LOCK(lock),
  73 };
  74
  75 /*
  76  * This path almost never happens for VM activity - pages are normally
  77  * freed via pagevecs.  But it gets used by networking.
  78  */
  79 static void __page_cache_release(struct page *page)
  80 {
  81         if (PageLRU(page)) {
  82                 pg_data_t *pgdat = page_pgdat(page);
  83                 struct lruvec *lruvec;
  84                 unsigned long flags;
  85
  86                 spin_lock_irqsave(&pgdat->lru_lock, flags);
  87                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
  88                 VM_BUG_ON_PAGE(!PageLRU(page), page);
  89                 __ClearPageLRU(page);
  90                 del_page_from_lru_list(page, lruvec, page_off_lru(page));
  91                 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
  92         }
  93         __ClearPageWaiters(page);
  94 }
  95
  96 static void __put_single_page(struct page *page)
  97 {
  98         __page_cache_release(page);
  99         mem_cgroup_uncharge(page);
 100         free_unref_page(page);
 101 }
 102
 103 static void __put_compound_page(struct page *page)
 104 {
 105         /*
 106          * __page_cache_release() is supposed to be called for thp, not for
 107          * hugetlb. This is because hugetlb page does never have PageLRU set
 108          * (it's never listed to any LRU lists) and no memcg routines should
 109          * be called for hugetlb (it has a separate hugetlb_cgroup.)
 110          */
 111         if (!PageHuge(page))
 112                 __page_cache_release(page);
 113         destroy_compound_page(page);
 114 }
 115
 116 void __put_page(struct page *page)
 117 {
 118         if (is_zone_device_page(page)) {
 119                 put_dev_pagemap(page->pgmap);
 120
 121                 /*
 122                  * The page belongs to the device that created pgmap. Do
 123                  * not return it to page allocator.
 124                  */
 125                 return;
 126         }
 127
 128         if (unlikely(PageCompound(page)))
 129                 __put_compound_page(page);
 130         else
 131                 __put_single_page(page);
 132 }
 133 EXPORT_SYMBOL(__put_page);
 134
 135 /**
 136  * put_pages_list() - release a list of pages
 137  * @pages: list of pages threaded on page->lru
 138  *
 139  * Release a list of pages which are strung together on page.lru.  Currently
 140  * used by read_cache_pages() and related error recovery code.
 141  */
 142 void put_pages_list(struct list_head *pages)
 143 {
 144         while (!list_empty(pages)) {
 145                 struct page *victim;
 146
 147                 victim = lru_to_page(pages);
 148                 list_del(&victim->lru);
 149                 put_page(victim);
 150         }
 151 }
 152 EXPORT_SYMBOL(put_pages_list);
 153
 154 /*
 155  * get_kernel_pages() - pin kernel pages in memory
 156  * @kiov:       An array of struct kvec structures
 157  * @nr_segs:    number of segments to pin
 158  * @write:      pinning for read/write, currently ignored
 159  * @pages:      array that receives pointers to the pages pinned.
 160  *              Should be at least nr_segs long.
 161  *
 162  * Returns number of pages pinned. This may be fewer than the number
 163  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 164  * were pinned, returns -errno. Each page returned must be released
 165  * with a put_page() call when it is finished with.
 166  */
 167 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
 168                 struct page **pages)
 169 {
 170         int seg;
 171
 172         for (seg = 0; seg < nr_segs; seg++) {
 173                 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
 174                         return seg;
 175
 176                 pages[seg] = kmap_to_page(kiov[seg].iov_base);
 177                 get_page(pages[seg]);
 178         }
 179
 180         return seg;
 181 }
 182 EXPORT_SYMBOL_GPL(get_kernel_pages);
 183
 184 /*
 185  * get_kernel_page() - pin a kernel page in memory
 186  * @start:      starting kernel address
 187  * @write:      pinning for read/write, currently ignored
 188  * @pages:      array that receives pointer to the page pinned.
 189  *              Must be at least nr_segs long.
 190  *
 191  * Returns 1 if page is pinned. If the page was not pinned, returns
 192  * -errno. The page returned must be released with a put_page() call
 193  * when it is finished with.
 194  */
 195 int get_kernel_page(unsigned long start, int write, struct page **pages)
 196 {
 197         const struct kvec kiov = {
 198                 .iov_base = (void *)start,
 199                 .iov_len = PAGE_SIZE
 200         };
 201
 202         return get_kernel_pages(&kiov, 1, write, pages);
 203 }
 204 EXPORT_SYMBOL_GPL(get_kernel_page);
 205
 206 static void pagevec_lru_move_fn(struct pagevec *pvec,
 207         void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
 208         void *arg)
 209 {
 210         int i;
 211         struct pglist_data *pgdat = NULL;
 212         struct lruvec *lruvec;
 213         unsigned long flags = 0;
 214
 215         for (i = 0; i < pagevec_count(pvec); i++) {
 216                 struct page *page = pvec->pages[i];
 217                 struct pglist_data *pagepgdat = page_pgdat(page);
 218
 219                 if (pagepgdat != pgdat) {
 220                         if (pgdat)
 221                                 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
 222                         pgdat = pagepgdat;
 223                         spin_lock_irqsave(&pgdat->lru_lock, flags);
 224                 }
 225
 226                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
 227                 (*move_fn)(page, lruvec, arg);
 228         }
 229         if (pgdat)
 230                 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
 231         release_pages(pvec->pages, pvec->nr);
 232         pagevec_reinit(pvec);
 233 }
 234
 235 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
 236                                  void *arg)
 237 {
 238         int *pgmoved = arg;
 239
 240         if (PageLRU(page) && !PageUnevictable(page)) {
 241                 del_page_from_lru_list(page, lruvec, page_lru(page));
 242                 ClearPageActive(page);
 243                 add_page_to_lru_list_tail(page, lruvec, page_lru(page));
 244                 (*pgmoved)++;
 245         }
 246 }
 247
 248 /*
 249  * pagevec_move_tail() must be called with IRQ disabled.
 250  * Otherwise this may cause nasty races.
 251  */
 252 static void pagevec_move_tail(struct pagevec *pvec)
 253 {
 254         int pgmoved = 0;
 255
 256         pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
 257         __count_vm_events(PGROTATED, pgmoved);
 258 }
 259
 260 /*
 261  * Writeback is about to end against a page which has been marked for immediate
 262  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 263  * inactive list.
 264  */
 265 void rotate_reclaimable_page(struct page *page)
 266 {
 267         if (!PageLocked(page) && !PageDirty(page) &&
 268             !PageUnevictable(page) && PageLRU(page)) {
 269                 struct pagevec *pvec;
 270                 unsigned long flags;
 271
 272                 get_page(page);
 273                 local_lock_irqsave(&lru_rotate.lock, flags);
 274                 pvec = this_cpu_ptr(&lru_rotate.pvec);
 275                 if (!pagevec_add(pvec, page) || PageCompound(page))
 276                         pagevec_move_tail(pvec);
 277                 local_unlock_irqrestore(&lru_rotate.lock, flags);
 278         }
 279 }
 280
 281 static void update_page_reclaim_stat(struct lruvec *lruvec,
 282                                      int file, int rotated,
 283                                      unsigned int nr_pages)
 284 {
 285         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
 286
 287         reclaim_stat->recent_scanned[file] += nr_pages;
 288         if (rotated)
 289                 reclaim_stat->recent_rotated[file] += nr_pages;
 290 }
 291
 292 static void __activate_page(struct page *page, struct lruvec *lruvec,
 293                             void *arg)
 294 {
 295         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 296                 int lru = page_lru_base_type(page);
 297
 298                 del_page_from_lru_list(page, lruvec, lru);
 299                 SetPageActive(page);
 300                 lru += LRU_ACTIVE;
 301                 add_page_to_lru_list(page, lruvec, lru);
 302                 trace_mm_lru_activate(page);
 303
 304                 __count_vm_event(PGACTIVATE);
 305         }
 306 }
 307
 308 #ifdef CONFIG_SMP
 309 static void activate_page_drain(int cpu)
 310 {
 311         struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
 312
 313         if (pagevec_count(pvec))
 314                 pagevec_lru_move_fn(pvec, __activate_page, NULL);
 315 }
 316
 317 static bool need_activate_page_drain(int cpu)
 318 {
 319         return pagevec_count(&per_cpu(lru_pvecs.activate_page, cpu)) != 0;
 320 }
 321
 322 void activate_page(struct page *page)
 323 {
 324         page = compound_head(page);
 325         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 326                 struct pagevec *pvec;
 327
 328                 local_lock(&lru_pvecs.lock);
 329                 pvec = this_cpu_ptr(&lru_pvecs.activate_page);
 330                 get_page(page);
 331                 if (!pagevec_add(pvec, page) || PageCompound(page))
 332                         pagevec_lru_move_fn(pvec, __activate_page, NULL);
 333                 local_unlock(&lru_pvecs.lock);
 334         }
 335 }
 336
 337 #else
 338 static inline void activate_page_drain(int cpu)
 339 {
 340 }
 341
 342 void activate_page(struct page *page)
 343 {
 344         pg_data_t *pgdat = page_pgdat(page);
 345
 346         page = compound_head(page);
 347         spin_lock_irq(&pgdat->lru_lock);
 348         __activate_page(page, mem_cgroup_page_lruvec(page, pgdat), NULL);
 349         spin_unlock_irq(&pgdat->lru_lock);
 350 }
 351 #endif
 352
 353 static void __lru_cache_activate_page(struct page *page)
 354 {
 355         struct pagevec *pvec;
 356         int i;
 357
 358         local_lock(&lru_pvecs.lock);
 359         pvec = this_cpu_ptr(&lru_pvecs.lru_add);
 360
 361         /*
 362          * Search backwards on the optimistic assumption that the page being
 363          * activated has just been added to this pagevec. Note that only
 364          * the local pagevec is examined as a !PageLRU page could be in the
 365          * process of being released, reclaimed, migrated or on a remote
 366          * pagevec that is currently being drained. Furthermore, marking
 367          * a remote pagevec's page PageActive potentially hits a race where
 368          * a page is marked PageActive just after it is added to the inactive
 369          * list causing accounting errors and BUG_ON checks to trigger.
 370          */
 371         for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
 372                 struct page *pagevec_page = pvec->pages[i];
 373
 374                 if (pagevec_page == page) {
 375                         SetPageActive(page);
 376                         break;
 377                 }
 378         }
 379
 380         local_unlock(&lru_pvecs.lock);
 381 }
 382
 383 /*
 384  * Mark a page as having seen activity.
 385  *
 386  * inactive,unreferenced        ->      inactive,referenced
 387  * inactive,referenced          ->      active,unreferenced
 388  * active,unreferenced          ->      active,referenced
 389  *
 390  * When a newly allocated page is not yet visible, so safe for non-atomic ops,
 391  * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
 392  */
 393 void mark_page_accessed(struct page *page)
 394 {
 395         page = compound_head(page);
 396
 397         if (!PageReferenced(page)) {
 398                 SetPageReferenced(page);
 399         } else if (PageUnevictable(page)) {
 400                 /*
 401                  * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
 402                  * this list is never rotated or maintained, so marking an
 403                  * evictable page accessed has no effect.
 404                  */
 405         } else if (!PageActive(page)) {
 406                 /*
 407                  * If the page is on the LRU, queue it for activation via
 408                  * lru_pvecs.activate_page. Otherwise, assume the page is on a
 409                  * pagevec, mark it active and it'll be moved to the active
 410                  * LRU on the next drain.
 411                  */
 412                 if (PageLRU(page))
 413                         activate_page(page);
 414                 else
 415                         __lru_cache_activate_page(page);
 416                 ClearPageReferenced(page);
 417                 if (page_is_file_lru(page))
 418                         workingset_activation(page);
 419         }
 420         if (page_is_idle(page))
 421                 clear_page_idle(page);
 422 }
 423 EXPORT_SYMBOL(mark_page_accessed);
 424
 425 /**
 426  * lru_cache_add - add a page to a page list
 427  * @page: the page to be added to the LRU.
 428  *
 429  * Queue the page for addition to the LRU via pagevec. The decision on whether
 430  * to add the page to the [in]active [file|anon] list is deferred until the
 431  * pagevec is drained. This gives a chance for the caller of lru_cache_add()
 432  * have the page added to the active list using mark_page_accessed().
 433  */
 434 void lru_cache_add(struct page *page)
 435 {
 436         struct pagevec *pvec;
 437
 438         VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
 439         VM_BUG_ON_PAGE(PageLRU(page), page);
 440
 441         get_page(page);
 442         local_lock(&lru_pvecs.lock);
 443         pvec = this_cpu_ptr(&lru_pvecs.lru_add);
 444         if (!pagevec_add(pvec, page) || PageCompound(page))
 445                 __pagevec_lru_add(pvec);
 446         local_unlock(&lru_pvecs.lock);
 447 }
 448 EXPORT_SYMBOL(lru_cache_add);
 449
 450 /**
 451  * lru_cache_add_active_or_unevictable
 452  * @page:  the page to be added to LRU
 453  * @vma:   vma in which page is mapped for determining reclaimability
 454  *
 455  * Place @page on the active or unevictable LRU list, depending on its
 456  * evictability.  Note that if the page is not evictable, it goes
 457  * directly back onto it's zone's unevictable list, it does NOT use a
 458  * per cpu pagevec.
 459  */
 460 void lru_cache_add_active_or_unevictable(struct page *page,
 461                                          struct vm_area_struct *vma)
 462 {
 463         VM_BUG_ON_PAGE(PageLRU(page), page);
 464
 465         if (likely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) != VM_LOCKED))
 466                 SetPageActive(page);
 467         else if (!TestSetPageMlocked(page)) {
 468                 /*
 469                  * We use the irq-unsafe __mod_zone_page_stat because this
 470                  * counter is not modified from interrupt context, and the pte
 471                  * lock is held(spinlock), which implies preemption disabled.
 472                  */
 473                 __mod_zone_page_state(page_zone(page), NR_MLOCK,
 474                                     hpage_nr_pages(page));
 475                 count_vm_event(UNEVICTABLE_PGMLOCKED);
 476         }
 477         lru_cache_add(page);
 478 }
 479
 480 /*
 481  * If the page can not be invalidated, it is moved to the
 482  * inactive list to speed up its reclaim.  It is moved to the
 483  * head of the list, rather than the tail, to give the flusher
 484  * threads some time to write it out, as this is much more
 485  * effective than the single-page writeout from reclaim.
 486  *
 487  * If the page isn't page_mapped and dirty/writeback, the page
 488  * could reclaim asap using PG_reclaim.
 489  *
 490  * 1. active, mapped page -> none
 491  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 492  * 3. inactive, mapped page -> none
 493  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 494  * 5. inactive, clean -> inactive, tail
 495  * 6. Others -> none
 496  *
 497  * In 4, why it moves inactive's head, the VM expects the page would
 498  * be write it out by flusher threads as this is much more effective
 499  * than the single-page writeout from reclaim.
 500  */
 501 static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec,
 502                               void *arg)
 503 {
 504         int lru, file;
 505         bool active;
 506
 507         if (!PageLRU(page))
 508                 return;
 509
 510         if (PageUnevictable(page))
 511                 return;
 512
 513         /* Some processes are using the page */
 514         if (page_mapped(page))
 515                 return;
 516
 517         active = PageActive(page);
 518         file = page_is_file_lru(page);
 519         lru = page_lru_base_type(page);
 520
 521         del_page_from_lru_list(page, lruvec, lru + active);
 522         ClearPageActive(page);
 523         ClearPageReferenced(page);
 524
 525         if (PageWriteback(page) || PageDirty(page)) {
 526                 /*
 527                  * PG_reclaim could be raced with end_page_writeback
 528                  * It can make readahead confusing.  But race window
 529                  * is _really_ small and  it's non-critical problem.
 530                  */
 531                 add_page_to_lru_list(page, lruvec, lru);
 532                 SetPageReclaim(page);
 533         } else {
 534                 /*
 535                  * The page's writeback ends up during pagevec
 536                  * We moves tha page into tail of inactive.
 537                  */
 538                 add_page_to_lru_list_tail(page, lruvec, lru);
 539                 __count_vm_event(PGROTATED);
 540         }
 541
 542         if (active)
 543                 __count_vm_event(PGDEACTIVATE);
 544         update_page_reclaim_stat(lruvec, file, 0, hpage_nr_pages(page));
 545 }
 546
 547 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 548                             void *arg)
 549 {
 550         if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
 551                 int file = page_is_file_lru(page);
 552                 int lru = page_lru_base_type(page);
 553
 554                 del_page_from_lru_list(page, lruvec, lru + LRU_ACTIVE);
 555                 ClearPageActive(page);
 556                 ClearPageReferenced(page);
 557                 add_page_to_lru_list(page, lruvec, lru);
 558
 559                 __count_vm_events(PGDEACTIVATE, hpage_nr_pages(page));
 560                 update_page_reclaim_stat(lruvec, file, 0, hpage_nr_pages(page));
 561         }
 562 }
 563
 564 static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec,
 565                             void *arg)
 566 {
 567         if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
 568             !PageSwapCache(page) && !PageUnevictable(page)) {
 569                 bool active = PageActive(page);
 570
 571                 del_page_from_lru_list(page, lruvec,
 572                                        LRU_INACTIVE_ANON + active);
 573                 ClearPageActive(page);
 574                 ClearPageReferenced(page);
 575                 /*
 576                  * Lazyfree pages are clean anonymous pages.  They have
 577                  * PG_swapbacked flag cleared, to distinguish them from normal
 578                  * anonymous pages
 579                  */
 580                 ClearPageSwapBacked(page);
 581                 add_page_to_lru_list(page, lruvec, LRU_INACTIVE_FILE);
 582
 583                 __count_vm_events(PGLAZYFREE, hpage_nr_pages(page));
 584                 count_memcg_page_event(page, PGLAZYFREE);
 585                 update_page_reclaim_stat(lruvec, 1, 0, hpage_nr_pages(page));
 586         }
 587 }
 588
 589 /*
 590  * Drain pages out of the cpu's pagevecs.
 591  * Either "cpu" is the current CPU, and preemption has already been
 592  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 593  */
 594 void lru_add_drain_cpu(int cpu)
 595 {
 596         struct pagevec *pvec = &per_cpu(lru_pvecs.lru_add, cpu);
 597
 598         if (pagevec_count(pvec))
 599                 __pagevec_lru_add(pvec);
 600
 601         pvec = &per_cpu(lru_rotate.pvec, cpu);
 602         if (pagevec_count(pvec)) {
 603                 unsigned long flags;
 604
 605                 /* No harm done if a racing interrupt already did this */
 606                 local_lock_irqsave(&lru_rotate.lock, flags);
 607                 pagevec_move_tail(pvec);
 608                 local_unlock_irqrestore(&lru_rotate.lock, flags);
 609         }
 610
 611         pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
 612         if (pagevec_count(pvec))
 613                 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
 614
 615         pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
 616         if (pagevec_count(pvec))
 617                 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 618
 619         pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
 620         if (pagevec_count(pvec))
 621                 pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
 622
 623         activate_page_drain(cpu);
 624 }
 625
 626 /**
 627  * deactivate_file_page - forcefully deactivate a file page
 628  * @page: page to deactivate
 629  *
 630  * This function hints the VM that @page is a good reclaim candidate,
 631  * for example if its invalidation fails due to the page being dirty
 632  * or under writeback.
 633  */
 634 void deactivate_file_page(struct page *page)
 635 {
 636         /*
 637          * In a workload with many unevictable page such as mprotect,
 638          * unevictable page deactivation for accelerating reclaim is pointless.
 639          */
 640         if (PageUnevictable(page))
 641                 return;
 642
 643         if (likely(get_page_unless_zero(page))) {
 644                 struct pagevec *pvec;
 645
 646                 local_lock(&lru_pvecs.lock);
 647                 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
 648
 649                 if (!pagevec_add(pvec, page) || PageCompound(page))
 650                         pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
 651                 local_unlock(&lru_pvecs.lock);
 652         }
 653 }
 654
 655 /*
 656  * deactivate_page - deactivate a page
 657  * @page: page to deactivate
 658  *
 659  * deactivate_page() moves @page to the inactive list if @page was on the active
 660  * list and was not an unevictable page.  This is done to accelerate the reclaim
 661  * of @page.
 662  */
 663 void deactivate_page(struct page *page)
 664 {
 665         if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
 666                 struct pagevec *pvec;
 667
 668                 local_lock(&lru_pvecs.lock);
 669                 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
 670                 get_page(page);
 671                 if (!pagevec_add(pvec, page) || PageCompound(page))
 672                         pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 673                 local_unlock(&lru_pvecs.lock);
 674         }
 675 }
 676
 677 /**
 678  * mark_page_lazyfree - make an anon page lazyfree
 679  * @page: page to deactivate
 680  *
 681  * mark_page_lazyfree() moves @page to the inactive file list.
 682  * This is done to accelerate the reclaim of @page.
 683  */
 684 void mark_page_lazyfree(struct page *page)
 685 {
 686         if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
 687             !PageSwapCache(page) && !PageUnevictable(page)) {
 688                 struct pagevec *pvec;
 689
 690                 local_lock(&lru_pvecs.lock);
 691                 pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
 692                 get_page(page);
 693                 if (!pagevec_add(pvec, page) || PageCompound(page))
 694                         pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
 695                 local_unlock(&lru_pvecs.lock);
 696         }
 697 }
 698
 699 void lru_add_drain(void)
 700 {
 701         local_lock(&lru_pvecs.lock);
 702         lru_add_drain_cpu(smp_processor_id());
 703         local_unlock(&lru_pvecs.lock);
 704 }
 705
 706 void lru_add_drain_cpu_zone(struct zone *zone)
 707 {
 708         local_lock(&lru_pvecs.lock);
 709         lru_add_drain_cpu(smp_processor_id());
 710         drain_local_pages(zone);
 711         local_unlock(&lru_pvecs.lock);
 712 }
 713
 714 #ifdef CONFIG_SMP
 715
 716 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 717
 718 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 719 {
 720         lru_add_drain();
 721 }
 722
 723 /*
 724  * Doesn't need any cpu hotplug locking because we do rely on per-cpu
 725  * kworkers being shut down before our page_alloc_cpu_dead callback is
 726  * executed on the offlined cpu.
 727  * Calling this function with cpu hotplug locks held can actually lead
 728  * to obscure indirect dependencies via WQ context.
 729  */
 730 void lru_add_drain_all(void)
 731 {
 732         static seqcount_t seqcount = SEQCNT_ZERO(seqcount);
 733         static DEFINE_MUTEX(lock);
 734         static struct cpumask has_work;
 735         int cpu, seq;
 736
 737         /*
 738          * Make sure nobody triggers this path before mm_percpu_wq is fully
 739          * initialized.
 740          */
 741         if (WARN_ON(!mm_percpu_wq))
 742                 return;
 743
 744         seq = raw_read_seqcount_latch(&seqcount);
 745
 746         mutex_lock(&lock);
 747
 748         /*
 749          * Piggyback on drain started and finished while we waited for lock:
 750          * all pages pended at the time of our enter were drained from vectors.
 751          */
 752         if (__read_seqcount_retry(&seqcount, seq))
 753                 goto done;
 754
 755         raw_write_seqcount_latch(&seqcount);
 756
 757         cpumask_clear(&has_work);
 758
 759         for_each_online_cpu(cpu) {
 760                 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 761
 762                 if (pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
 763                     pagevec_count(&per_cpu(lru_rotate.pvec, cpu)) ||
 764                     pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
 765                     pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
 766                     pagevec_count(&per_cpu(lru_pvecs.lru_lazyfree, cpu)) ||
 767                     need_activate_page_drain(cpu)) {
 768                         INIT_WORK(work, lru_add_drain_per_cpu);
 769                         queue_work_on(cpu, mm_percpu_wq, work);
 770                         cpumask_set_cpu(cpu, &has_work);
 771                 }
 772         }
 773
 774         for_each_cpu(cpu, &has_work)
 775                 flush_work(&per_cpu(lru_add_drain_work, cpu));
 776
 777 done:
 778         mutex_unlock(&lock);
 779 }
 780 #else
 781 void lru_add_drain_all(void)
 782 {
 783         lru_add_drain();
 784 }
 785 #endif
 786
 787 /**
 788  * release_pages - batched put_page()
 789  * @pages: array of pages to release
 790  * @nr: number of pages
 791  *
 792  * Decrement the reference count on all the pages in @pages.  If it
 793  * fell to zero, remove the page from the LRU and free it.
 794  */
 795 void release_pages(struct page **pages, int nr)
 796 {
 797         int i;
 798         LIST_HEAD(pages_to_free);
 799         struct pglist_data *locked_pgdat = NULL;
 800         struct lruvec *lruvec;
 801         unsigned long uninitialized_var(flags);
 802         unsigned int uninitialized_var(lock_batch);
 803
 804         for (i = 0; i < nr; i++) {
 805                 struct page *page = pages[i];
 806
 807                 /*
 808                  * Make sure the IRQ-safe lock-holding time does not get
 809                  * excessive with a continuous string of pages from the
 810                  * same pgdat. The lock is held only if pgdat != NULL.
 811                  */
 812                 if (locked_pgdat && ++lock_batch == SWAP_CLUSTER_MAX) {
 813                         spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 814                         locked_pgdat = NULL;
 815                 }
 816
 817                 if (is_huge_zero_page(page))
 818                         continue;
 819
 820                 if (is_zone_device_page(page)) {
 821                         if (locked_pgdat) {
 822                                 spin_unlock_irqrestore(&locked_pgdat->lru_lock,
 823                                                        flags);
 824                                 locked_pgdat = NULL;
 825                         }
 826                         /*
 827                          * ZONE_DEVICE pages that return 'false' from
 828                          * put_devmap_managed_page() do not require special
 829                          * processing, and instead, expect a call to
 830                          * put_page_testzero().
 831                          */
 832                         if (page_is_devmap_managed(page)) {
 833                                 put_devmap_managed_page(page);
 834                                 continue;
 835                         }
 836                 }
 837
 838                 page = compound_head(page);
 839                 if (!put_page_testzero(page))
 840                         continue;
 841
 842                 if (PageCompound(page)) {
 843                         if (locked_pgdat) {
 844                                 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 845                                 locked_pgdat = NULL;
 846                         }
 847                         __put_compound_page(page);
 848                         continue;
 849                 }
 850
 851                 if (PageLRU(page)) {
 852                         struct pglist_data *pgdat = page_pgdat(page);
 853
 854                         if (pgdat != locked_pgdat) {
 855                                 if (locked_pgdat)
 856                                         spin_unlock_irqrestore(&locked_pgdat->lru_lock,
 857                                                                         flags);
 858                                 lock_batch = 0;
 859                                 locked_pgdat = pgdat;
 860                                 spin_lock_irqsave(&locked_pgdat->lru_lock, flags);
 861                         }
 862
 863                         lruvec = mem_cgroup_page_lruvec(page, locked_pgdat);
 864                         VM_BUG_ON_PAGE(!PageLRU(page), page);
 865                         __ClearPageLRU(page);
 866                         del_page_from_lru_list(page, lruvec, page_off_lru(page));
 867                 }
 868
 869                 /* Clear Active bit in case of parallel mark_page_accessed */
 870                 __ClearPageActive(page);
 871                 __ClearPageWaiters(page);
 872
 873                 list_add(&page->lru, &pages_to_free);
 874         }
 875         if (locked_pgdat)
 876                 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 877
 878         mem_cgroup_uncharge_list(&pages_to_free);
 879         free_unref_page_list(&pages_to_free);
 880 }
 881 EXPORT_SYMBOL(release_pages);
 882
 883 /*
 884  * The pages which we're about to release may be in the deferred lru-addition
 885  * queues.  That would prevent them from really being freed right now.  That's
 886  * OK from a correctness point of view but is inefficient - those pages may be
 887  * cache-warm and we want to give them back to the page allocator ASAP.
 888  *
 889  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 890  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 891  * mutual recursion.
 892  */
 893 void __pagevec_release(struct pagevec *pvec)
 894 {
 895         if (!pvec->percpu_pvec_drained) {
 896                 lru_add_drain();
 897                 pvec->percpu_pvec_drained = true;
 898         }
 899         release_pages(pvec->pages, pagevec_count(pvec));
 900         pagevec_reinit(pvec);
 901 }
 902 EXPORT_SYMBOL(__pagevec_release);
 903
 904 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 905 /* used by __split_huge_page_refcount() */
 906 void lru_add_page_tail(struct page *page, struct page *page_tail,
 907                        struct lruvec *lruvec, struct list_head *list)
 908 {
 909         VM_BUG_ON_PAGE(!PageHead(page), page);
 910         VM_BUG_ON_PAGE(PageCompound(page_tail), page);
 911         VM_BUG_ON_PAGE(PageLRU(page_tail), page);
 912         lockdep_assert_held(&lruvec_pgdat(lruvec)->lru_lock);
 913
 914         if (!list)
 915                 SetPageLRU(page_tail);
 916
 917         if (likely(PageLRU(page)))
 918                 list_add_tail(&page_tail->lru, &page->lru);
 919         else if (list) {
 920                 /* page reclaim is reclaiming a huge page */
 921                 get_page(page_tail);
 922                 list_add_tail(&page_tail->lru, list);
 923         } else {
 924                 /*
 925                  * Head page has not yet been counted, as an hpage,
 926                  * so we must account for each subpage individually.
 927                  *
 928                  * Put page_tail on the list at the correct position
 929                  * so they all end up in order.
 930                  */
 931                 add_page_to_lru_list_tail(page_tail, lruvec,
 932                                           page_lru(page_tail));
 933         }
 934 }
 935 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 936
 937 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
 938                                  void *arg)
 939 {
 940         enum lru_list lru;
 941         int was_unevictable = TestClearPageUnevictable(page);
 942
 943         VM_BUG_ON_PAGE(PageLRU(page), page);
 944
 945         /*
 946          * Page becomes evictable in two ways:
 947          * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
 948          * 2) Before acquiring LRU lock to put the page to correct LRU and then
 949          *   a) do PageLRU check with lock [check_move_unevictable_pages]
 950          *   b) do PageLRU check before lock [clear_page_mlock]
 951          *
 952          * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
 953          * following strict ordering:
 954          *
 955          * #0: __pagevec_lru_add_fn             #1: clear_page_mlock
 956          *
 957          * SetPageLRU()                         TestClearPageMlocked()
 958          * smp_mb() // explicit ordering        // above provides strict
 959          *                                      // ordering
 960          * PageMlocked()                        PageLRU()
 961          *
 962          *
 963          * if '#1' does not observe setting of PG_lru by '#0' and fails
 964          * isolation, the explicit barrier will make sure that page_evictable
 965          * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
 966          * can be reordered after PageMlocked check and can make '#1' to fail
 967          * the isolation of the page whose Mlocked bit is cleared (#0 is also
 968          * looking at the same page) and the evictable page will be stranded
 969          * in an unevictable LRU.
 970          */
 971         SetPageLRU(page);
 972         smp_mb__after_atomic();
 973
 974         if (page_evictable(page)) {
 975                 lru = page_lru(page);
 976                 if (was_unevictable)
 977                         count_vm_event(UNEVICTABLE_PGRESCUED);
 978         } else {
 979                 lru = LRU_UNEVICTABLE;
 980                 ClearPageActive(page);
 981                 SetPageUnevictable(page);
 982                 if (!was_unevictable)
 983                         count_vm_event(UNEVICTABLE_PGCULLED);
 984         }
 985
 986         add_page_to_lru_list(page, lruvec, lru);
 987         trace_mm_lru_insertion(page, lru);
 988 }
 989
 990 /*
 991  * Add the passed pages to the LRU, then drop the caller's refcount
 992  * on them.  Reinitialises the caller's pagevec.
 993  */
 994 void __pagevec_lru_add(struct pagevec *pvec)
 995 {
 996         pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
 997 }
 998
 999 /**
1000  * pagevec_lookup_entries - gang pagecache lookup
1001  * @pvec:       Where the resulting entries are placed
1002  * @mapping:    The address_space to search
1003  * @start:      The starting entry index
1004  * @nr_entries: The maximum number of pages
1005  * @indices:    The cache indices corresponding to the entries in @pvec
1006  *
1007  * pagevec_lookup_entries() will search for and return a group of up
1008  * to @nr_pages pages and shadow entries in the mapping.  All
1009  * entries are placed in @pvec.  pagevec_lookup_entries() takes a
1010  * reference against actual pages in @pvec.
1011  *
1012  * The search returns a group of mapping-contiguous entries with
1013  * ascending indexes.  There may be holes in the indices due to
1014  * not-present entries.
1015  *
1016  * Only one subpage of a Transparent Huge Page is returned in one call:
1017  * allowing truncate_inode_pages_range() to evict the whole THP without
1018  * cycling through a pagevec of extra references.
1019  *
1020  * pagevec_lookup_entries() returns the number of entries which were
1021  * found.
1022  */
1023 unsigned pagevec_lookup_entries(struct pagevec *pvec,
1024                                 struct address_space *mapping,
1025                                 pgoff_t start, unsigned nr_entries,
1026                                 pgoff_t *indices)
1027 {
1028         pvec->nr = find_get_entries(mapping, start, nr_entries,
1029                                     pvec->pages, indices);
1030         return pagevec_count(pvec);
1031 }
1032
1033 /**
1034  * pagevec_remove_exceptionals - pagevec exceptionals pruning
1035  * @pvec:       The pagevec to prune
1036  *
1037  * pagevec_lookup_entries() fills both pages and exceptional radix
1038  * tree entries into the pagevec.  This function prunes all
1039  * exceptionals from @pvec without leaving holes, so that it can be
1040  * passed on to page-only pagevec operations.
1041  */
1042 void pagevec_remove_exceptionals(struct pagevec *pvec)
1043 {
1044         int i, j;
1045
1046         for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
1047                 struct page *page = pvec->pages[i];
1048                 if (!xa_is_value(page))
1049                         pvec->pages[j++] = page;
1050         }
1051         pvec->nr = j;
1052 }
1053
1054 /**
1055  * pagevec_lookup_range - gang pagecache lookup
1056  * @pvec:       Where the resulting pages are placed
1057  * @mapping:    The address_space to search
1058  * @start:      The starting page index
1059  * @end:        The final page index
1060  *
1061  * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1062  * pages in the mapping starting from index @start and upto index @end
1063  * (inclusive).  The pages are placed in @pvec.  pagevec_lookup() takes a
1064  * reference against the pages in @pvec.
1065  *
1066  * The search returns a group of mapping-contiguous pages with ascending
1067  * indexes.  There may be holes in the indices due to not-present pages. We
1068  * also update @start to index the next page for the traversal.
1069  *
1070  * pagevec_lookup_range() returns the number of pages which were found. If this
1071  * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1072  * reached.
1073  */
1074 unsigned pagevec_lookup_range(struct pagevec *pvec,
1075                 struct address_space *mapping, pgoff_t *start, pgoff_t end)
1076 {
1077         pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1078                                         pvec->pages);
1079         return pagevec_count(pvec);
1080 }
1081 EXPORT_SYMBOL(pagevec_lookup_range);
1082
1083 unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1084                 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1085                 xa_mark_t tag)
1086 {
1087         pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1088                                         PAGEVEC_SIZE, pvec->pages);
1089         return pagevec_count(pvec);
1090 }
1091 EXPORT_SYMBOL(pagevec_lookup_range_tag);
1092
1093 unsigned pagevec_lookup_range_nr_tag(struct pagevec *pvec,
1094                 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1095                 xa_mark_t tag, unsigned max_pages)
1096 {
1097         pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1098                 min_t(unsigned int, max_pages, PAGEVEC_SIZE), pvec->pages);
1099         return pagevec_count(pvec);
1100 }
1101 EXPORT_SYMBOL(pagevec_lookup_range_nr_tag);
1102 /*
1103  * Perform any setup for the swap system
1104  */
1105 void __init swap_setup(void)
1106 {
1107         unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1108
1109         /* Use a smaller cluster for small-memory machines */
1110         if (megs < 16)
1111                 page_cluster = 2;
1112         else
1113                 page_cluster = 3;
1114         /*
1115          * Right now other parts of the system means that we
1116          * _really_ don't want to cluster much more
1117          */
1118 }
1119
1120 #ifdef CONFIG_DEV_PAGEMAP_OPS
1121 void put_devmap_managed_page(struct page *page)
1122 {
1123         int count;
1124
1125         if (WARN_ON_ONCE(!page_is_devmap_managed(page)))
1126                 return;
1127
1128         count = page_ref_dec_return(page);
1129
1130         /*
1131          * devmap page refcounts are 1-based, rather than 0-based: if
1132          * refcount is 1, then the page is free and the refcount is
1133          * stable because nobody holds a reference on the page.
1134          */
1135         if (count == 1)
1136                 free_devmap_managed_page(page);
1137         else if (!count)
1138                 __put_page(page);
1139 }
1140 EXPORT_SYMBOL(put_devmap_managed_page);
1141 #endif