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) += hpage_nr_pages(page);
 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 void lru_note_cost(struct lruvec *lruvec, bool file, unsigned int nr_pages)
 282 {
 283         do {
 284                 unsigned long lrusize;
 285
 286                 /* Record cost event */
 287                 if (file)
 288                         lruvec->file_cost += nr_pages;
 289                 else
 290                         lruvec->anon_cost += nr_pages;
 291
 292                 /*
 293                  * Decay previous events
 294                  *
 295                  * Because workloads change over time (and to avoid
 296                  * overflow) we keep these statistics as a floating
 297                  * average, which ends up weighing recent refaults
 298                  * more than old ones.
 299                  */
 300                 lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
 301                           lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
 302                           lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
 303                           lruvec_page_state(lruvec, NR_ACTIVE_FILE);
 304
 305                 if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
 306                         lruvec->file_cost /= 2;
 307                         lruvec->anon_cost /= 2;
 308                 }
 309         } while ((lruvec = parent_lruvec(lruvec)));
 310 }
 311
 312 void lru_note_cost_page(struct page *page)
 313 {
 314         lru_note_cost(mem_cgroup_page_lruvec(page, page_pgdat(page)),
 315                       page_is_file_lru(page), hpage_nr_pages(page));
 316 }
 317
 318 static void __activate_page(struct page *page, struct lruvec *lruvec,
 319                             void *arg)
 320 {
 321         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 322                 int lru = page_lru_base_type(page);
 323                 int nr_pages = hpage_nr_pages(page);
 324
 325                 del_page_from_lru_list(page, lruvec, lru);
 326                 SetPageActive(page);
 327                 lru += LRU_ACTIVE;
 328                 add_page_to_lru_list(page, lruvec, lru);
 329                 trace_mm_lru_activate(page);
 330
 331                 __count_vm_events(PGACTIVATE, nr_pages);
 332                 __count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
 333                                      nr_pages);
 334         }
 335 }
 336
 337 #ifdef CONFIG_SMP
 338 static void activate_page_drain(int cpu)
 339 {
 340         struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
 341
 342         if (pagevec_count(pvec))
 343                 pagevec_lru_move_fn(pvec, __activate_page, NULL);
 344 }
 345
 346 static bool need_activate_page_drain(int cpu)
 347 {
 348         return pagevec_count(&per_cpu(lru_pvecs.activate_page, cpu)) != 0;
 349 }
 350
 351 void activate_page(struct page *page)
 352 {
 353         page = compound_head(page);
 354         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 355                 struct pagevec *pvec;
 356
 357                 local_lock(&lru_pvecs.lock);
 358                 pvec = this_cpu_ptr(&lru_pvecs.activate_page);
 359                 get_page(page);
 360                 if (!pagevec_add(pvec, page) || PageCompound(page))
 361                         pagevec_lru_move_fn(pvec, __activate_page, NULL);
 362                 local_unlock(&lru_pvecs.lock);
 363         }
 364 }
 365
 366 #else
 367 static inline void activate_page_drain(int cpu)
 368 {
 369 }
 370
 371 void activate_page(struct page *page)
 372 {
 373         pg_data_t *pgdat = page_pgdat(page);
 374
 375         page = compound_head(page);
 376         spin_lock_irq(&pgdat->lru_lock);
 377         __activate_page(page, mem_cgroup_page_lruvec(page, pgdat), NULL);
 378         spin_unlock_irq(&pgdat->lru_lock);
 379 }
 380 #endif
 381
 382 static void __lru_cache_activate_page(struct page *page)
 383 {
 384         struct pagevec *pvec;
 385         int i;
 386
 387         local_lock(&lru_pvecs.lock);
 388         pvec = this_cpu_ptr(&lru_pvecs.lru_add);
 389
 390         /*
 391          * Search backwards on the optimistic assumption that the page being
 392          * activated has just been added to this pagevec. Note that only
 393          * the local pagevec is examined as a !PageLRU page could be in the
 394          * process of being released, reclaimed, migrated or on a remote
 395          * pagevec that is currently being drained. Furthermore, marking
 396          * a remote pagevec's page PageActive potentially hits a race where
 397          * a page is marked PageActive just after it is added to the inactive
 398          * list causing accounting errors and BUG_ON checks to trigger.
 399          */
 400         for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
 401                 struct page *pagevec_page = pvec->pages[i];
 402
 403                 if (pagevec_page == page) {
 404                         SetPageActive(page);
 405                         break;
 406                 }
 407         }
 408
 409         local_unlock(&lru_pvecs.lock);
 410 }
 411
 412 /*
 413  * Mark a page as having seen activity.
 414  *
 415  * inactive,unreferenced        ->      inactive,referenced
 416  * inactive,referenced          ->      active,unreferenced
 417  * active,unreferenced          ->      active,referenced
 418  *
 419  * When a newly allocated page is not yet visible, so safe for non-atomic ops,
 420  * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
 421  */
 422 void mark_page_accessed(struct page *page)
 423 {
 424         page = compound_head(page);
 425
 426         if (!PageReferenced(page)) {
 427                 SetPageReferenced(page);
 428         } else if (PageUnevictable(page)) {
 429                 /*
 430                  * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
 431                  * this list is never rotated or maintained, so marking an
 432                  * evictable page accessed has no effect.
 433                  */
 434         } else if (!PageActive(page)) {
 435                 /*
 436                  * If the page is on the LRU, queue it for activation via
 437                  * lru_pvecs.activate_page. Otherwise, assume the page is on a
 438                  * pagevec, mark it active and it'll be moved to the active
 439                  * LRU on the next drain.
 440                  */
 441                 if (PageLRU(page))
 442                         activate_page(page);
 443                 else
 444                         __lru_cache_activate_page(page);
 445                 ClearPageReferenced(page);
 446                 if (page_is_file_lru(page))
 447                         workingset_activation(page);
 448         }
 449         if (page_is_idle(page))
 450                 clear_page_idle(page);
 451 }
 452 EXPORT_SYMBOL(mark_page_accessed);
 453
 454 /**
 455  * lru_cache_add - add a page to a page list
 456  * @page: the page to be added to the LRU.
 457  *
 458  * Queue the page for addition to the LRU via pagevec. The decision on whether
 459  * to add the page to the [in]active [file|anon] list is deferred until the
 460  * pagevec is drained. This gives a chance for the caller of lru_cache_add()
 461  * have the page added to the active list using mark_page_accessed().
 462  */
 463 void lru_cache_add(struct page *page)
 464 {
 465         struct pagevec *pvec;
 466
 467         VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
 468         VM_BUG_ON_PAGE(PageLRU(page), page);
 469
 470         get_page(page);
 471         local_lock(&lru_pvecs.lock);
 472         pvec = this_cpu_ptr(&lru_pvecs.lru_add);
 473         if (!pagevec_add(pvec, page) || PageCompound(page))
 474                 __pagevec_lru_add(pvec);
 475         local_unlock(&lru_pvecs.lock);
 476 }
 477 EXPORT_SYMBOL(lru_cache_add);
 478
 479 /**
 480  * lru_cache_add_active_or_unevictable
 481  * @page:  the page to be added to LRU
 482  * @vma:   vma in which page is mapped for determining reclaimability
 483  *
 484  * Place @page on the active or unevictable LRU list, depending on its
 485  * evictability.  Note that if the page is not evictable, it goes
 486  * directly back onto it's zone's unevictable list, it does NOT use a
 487  * per cpu pagevec.
 488  */
 489 void lru_cache_add_active_or_unevictable(struct page *page,
 490                                          struct vm_area_struct *vma)
 491 {
 492         VM_BUG_ON_PAGE(PageLRU(page), page);
 493
 494         if (likely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) != VM_LOCKED))
 495                 SetPageActive(page);
 496         else if (!TestSetPageMlocked(page)) {
 497                 /*
 498                  * We use the irq-unsafe __mod_zone_page_stat because this
 499                  * counter is not modified from interrupt context, and the pte
 500                  * lock is held(spinlock), which implies preemption disabled.
 501                  */
 502                 __mod_zone_page_state(page_zone(page), NR_MLOCK,
 503                                     hpage_nr_pages(page));
 504                 count_vm_event(UNEVICTABLE_PGMLOCKED);
 505         }
 506         lru_cache_add(page);
 507 }
 508
 509 /*
 510  * If the page can not be invalidated, it is moved to the
 511  * inactive list to speed up its reclaim.  It is moved to the
 512  * head of the list, rather than the tail, to give the flusher
 513  * threads some time to write it out, as this is much more
 514  * effective than the single-page writeout from reclaim.
 515  *
 516  * If the page isn't page_mapped and dirty/writeback, the page
 517  * could reclaim asap using PG_reclaim.
 518  *
 519  * 1. active, mapped page -> none
 520  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 521  * 3. inactive, mapped page -> none
 522  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 523  * 5. inactive, clean -> inactive, tail
 524  * 6. Others -> none
 525  *
 526  * In 4, why it moves inactive's head, the VM expects the page would
 527  * be write it out by flusher threads as this is much more effective
 528  * than the single-page writeout from reclaim.
 529  */
 530 static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec,
 531                               void *arg)
 532 {
 533         int lru;
 534         bool active;
 535         int nr_pages = hpage_nr_pages(page);
 536
 537         if (!PageLRU(page))
 538                 return;
 539
 540         if (PageUnevictable(page))
 541                 return;
 542
 543         /* Some processes are using the page */
 544         if (page_mapped(page))
 545                 return;
 546
 547         active = PageActive(page);
 548         lru = page_lru_base_type(page);
 549
 550         del_page_from_lru_list(page, lruvec, lru + active);
 551         ClearPageActive(page);
 552         ClearPageReferenced(page);
 553
 554         if (PageWriteback(page) || PageDirty(page)) {
 555                 /*
 556                  * PG_reclaim could be raced with end_page_writeback
 557                  * It can make readahead confusing.  But race window
 558                  * is _really_ small and  it's non-critical problem.
 559                  */
 560                 add_page_to_lru_list(page, lruvec, lru);
 561                 SetPageReclaim(page);
 562         } else {
 563                 /*
 564                  * The page's writeback ends up during pagevec
 565                  * We moves tha page into tail of inactive.
 566                  */
 567                 add_page_to_lru_list_tail(page, lruvec, lru);
 568                 __count_vm_events(PGROTATED, nr_pages);
 569         }
 570
 571         if (active) {
 572                 __count_vm_events(PGDEACTIVATE, nr_pages);
 573                 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
 574                                      nr_pages);
 575         }
 576 }
 577
 578 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 579                             void *arg)
 580 {
 581         if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
 582                 int lru = page_lru_base_type(page);
 583                 int nr_pages = hpage_nr_pages(page);
 584
 585                 del_page_from_lru_list(page, lruvec, lru + LRU_ACTIVE);
 586                 ClearPageActive(page);
 587                 ClearPageReferenced(page);
 588                 add_page_to_lru_list(page, lruvec, lru);
 589
 590                 __count_vm_events(PGDEACTIVATE, nr_pages);
 591                 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
 592                                      nr_pages);
 593         }
 594 }
 595
 596 static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec,
 597                             void *arg)
 598 {
 599         if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
 600             !PageSwapCache(page) && !PageUnevictable(page)) {
 601                 bool active = PageActive(page);
 602                 int nr_pages = hpage_nr_pages(page);
 603
 604                 del_page_from_lru_list(page, lruvec,
 605                                        LRU_INACTIVE_ANON + active);
 606                 ClearPageActive(page);
 607                 ClearPageReferenced(page);
 608                 /*
 609                  * Lazyfree pages are clean anonymous pages.  They have
 610                  * PG_swapbacked flag cleared, to distinguish them from normal
 611                  * anonymous pages
 612                  */
 613                 ClearPageSwapBacked(page);
 614                 add_page_to_lru_list(page, lruvec, LRU_INACTIVE_FILE);
 615
 616                 __count_vm_events(PGLAZYFREE, nr_pages);
 617                 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
 618                                      nr_pages);
 619         }
 620 }
 621
 622 /*
 623  * Drain pages out of the cpu's pagevecs.
 624  * Either "cpu" is the current CPU, and preemption has already been
 625  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 626  */
 627 void lru_add_drain_cpu(int cpu)
 628 {
 629         struct pagevec *pvec = &per_cpu(lru_pvecs.lru_add, cpu);
 630
 631         if (pagevec_count(pvec))
 632                 __pagevec_lru_add(pvec);
 633
 634         pvec = &per_cpu(lru_rotate.pvec, cpu);
 635         if (pagevec_count(pvec)) {
 636                 unsigned long flags;
 637
 638                 /* No harm done if a racing interrupt already did this */
 639                 local_lock_irqsave(&lru_rotate.lock, flags);
 640                 pagevec_move_tail(pvec);
 641                 local_unlock_irqrestore(&lru_rotate.lock, flags);
 642         }
 643
 644         pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
 645         if (pagevec_count(pvec))
 646                 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
 647
 648         pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
 649         if (pagevec_count(pvec))
 650                 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 651
 652         pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
 653         if (pagevec_count(pvec))
 654                 pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
 655
 656         activate_page_drain(cpu);
 657 }
 658
 659 /**
 660  * deactivate_file_page - forcefully deactivate a file page
 661  * @page: page to deactivate
 662  *
 663  * This function hints the VM that @page is a good reclaim candidate,
 664  * for example if its invalidation fails due to the page being dirty
 665  * or under writeback.
 666  */
 667 void deactivate_file_page(struct page *page)
 668 {
 669         /*
 670          * In a workload with many unevictable page such as mprotect,
 671          * unevictable page deactivation for accelerating reclaim is pointless.
 672          */
 673         if (PageUnevictable(page))
 674                 return;
 675
 676         if (likely(get_page_unless_zero(page))) {
 677                 struct pagevec *pvec;
 678
 679                 local_lock(&lru_pvecs.lock);
 680                 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
 681
 682                 if (!pagevec_add(pvec, page) || PageCompound(page))
 683                         pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
 684                 local_unlock(&lru_pvecs.lock);
 685         }
 686 }
 687
 688 /*
 689  * deactivate_page - deactivate a page
 690  * @page: page to deactivate
 691  *
 692  * deactivate_page() moves @page to the inactive list if @page was on the active
 693  * list and was not an unevictable page.  This is done to accelerate the reclaim
 694  * of @page.
 695  */
 696 void deactivate_page(struct page *page)
 697 {
 698         if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
 699                 struct pagevec *pvec;
 700
 701                 local_lock(&lru_pvecs.lock);
 702                 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
 703                 get_page(page);
 704                 if (!pagevec_add(pvec, page) || PageCompound(page))
 705                         pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 706                 local_unlock(&lru_pvecs.lock);
 707         }
 708 }
 709
 710 /**
 711  * mark_page_lazyfree - make an anon page lazyfree
 712  * @page: page to deactivate
 713  *
 714  * mark_page_lazyfree() moves @page to the inactive file list.
 715  * This is done to accelerate the reclaim of @page.
 716  */
 717 void mark_page_lazyfree(struct page *page)
 718 {
 719         if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
 720             !PageSwapCache(page) && !PageUnevictable(page)) {
 721                 struct pagevec *pvec;
 722
 723                 local_lock(&lru_pvecs.lock);
 724                 pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
 725                 get_page(page);
 726                 if (!pagevec_add(pvec, page) || PageCompound(page))
 727                         pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
 728                 local_unlock(&lru_pvecs.lock);
 729         }
 730 }
 731
 732 void lru_add_drain(void)
 733 {
 734         local_lock(&lru_pvecs.lock);
 735         lru_add_drain_cpu(smp_processor_id());
 736         local_unlock(&lru_pvecs.lock);
 737 }
 738
 739 void lru_add_drain_cpu_zone(struct zone *zone)
 740 {
 741         local_lock(&lru_pvecs.lock);
 742         lru_add_drain_cpu(smp_processor_id());
 743         drain_local_pages(zone);
 744         local_unlock(&lru_pvecs.lock);
 745 }
 746
 747 #ifdef CONFIG_SMP
 748
 749 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 750
 751 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 752 {
 753         lru_add_drain();
 754 }
 755
 756 /*
 757  * Doesn't need any cpu hotplug locking because we do rely on per-cpu
 758  * kworkers being shut down before our page_alloc_cpu_dead callback is
 759  * executed on the offlined cpu.
 760  * Calling this function with cpu hotplug locks held can actually lead
 761  * to obscure indirect dependencies via WQ context.
 762  */
 763 void lru_add_drain_all(void)
 764 {
 765         static seqcount_t seqcount = SEQCNT_ZERO(seqcount);
 766         static DEFINE_MUTEX(lock);
 767         static struct cpumask has_work;
 768         int cpu, seq;
 769
 770         /*
 771          * Make sure nobody triggers this path before mm_percpu_wq is fully
 772          * initialized.
 773          */
 774         if (WARN_ON(!mm_percpu_wq))
 775                 return;
 776
 777         seq = raw_read_seqcount_latch(&seqcount);
 778
 779         mutex_lock(&lock);
 780
 781         /*
 782          * Piggyback on drain started and finished while we waited for lock:
 783          * all pages pended at the time of our enter were drained from vectors.
 784          */
 785         if (__read_seqcount_retry(&seqcount, seq))
 786                 goto done;
 787
 788         raw_write_seqcount_latch(&seqcount);
 789
 790         cpumask_clear(&has_work);
 791
 792         for_each_online_cpu(cpu) {
 793                 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 794
 795                 if (pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
 796                     pagevec_count(&per_cpu(lru_rotate.pvec, cpu)) ||
 797                     pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
 798                     pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
 799                     pagevec_count(&per_cpu(lru_pvecs.lru_lazyfree, cpu)) ||
 800                     need_activate_page_drain(cpu)) {
 801                         INIT_WORK(work, lru_add_drain_per_cpu);
 802                         queue_work_on(cpu, mm_percpu_wq, work);
 803                         cpumask_set_cpu(cpu, &has_work);
 804                 }
 805         }
 806
 807         for_each_cpu(cpu, &has_work)
 808                 flush_work(&per_cpu(lru_add_drain_work, cpu));
 809
 810 done:
 811         mutex_unlock(&lock);
 812 }
 813 #else
 814 void lru_add_drain_all(void)
 815 {
 816         lru_add_drain();
 817 }
 818 #endif
 819
 820 /**
 821  * release_pages - batched put_page()
 822  * @pages: array of pages to release
 823  * @nr: number of pages
 824  *
 825  * Decrement the reference count on all the pages in @pages.  If it
 826  * fell to zero, remove the page from the LRU and free it.
 827  */
 828 void release_pages(struct page **pages, int nr)
 829 {
 830         int i;
 831         LIST_HEAD(pages_to_free);
 832         struct pglist_data *locked_pgdat = NULL;
 833         struct lruvec *lruvec;
 834         unsigned long uninitialized_var(flags);
 835         unsigned int uninitialized_var(lock_batch);
 836
 837         for (i = 0; i < nr; i++) {
 838                 struct page *page = pages[i];
 839
 840                 /*
 841                  * Make sure the IRQ-safe lock-holding time does not get
 842                  * excessive with a continuous string of pages from the
 843                  * same pgdat. The lock is held only if pgdat != NULL.
 844                  */
 845                 if (locked_pgdat && ++lock_batch == SWAP_CLUSTER_MAX) {
 846                         spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 847                         locked_pgdat = NULL;
 848                 }
 849
 850                 if (is_huge_zero_page(page))
 851                         continue;
 852
 853                 if (is_zone_device_page(page)) {
 854                         if (locked_pgdat) {
 855                                 spin_unlock_irqrestore(&locked_pgdat->lru_lock,
 856                                                        flags);
 857                                 locked_pgdat = NULL;
 858                         }
 859                         /*
 860                          * ZONE_DEVICE pages that return 'false' from
 861                          * put_devmap_managed_page() do not require special
 862                          * processing, and instead, expect a call to
 863                          * put_page_testzero().
 864                          */
 865                         if (page_is_devmap_managed(page)) {
 866                                 put_devmap_managed_page(page);
 867                                 continue;
 868                         }
 869                 }
 870
 871                 page = compound_head(page);
 872                 if (!put_page_testzero(page))
 873                         continue;
 874
 875                 if (PageCompound(page)) {
 876                         if (locked_pgdat) {
 877                                 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 878                                 locked_pgdat = NULL;
 879                         }
 880                         __put_compound_page(page);
 881                         continue;
 882                 }
 883
 884                 if (PageLRU(page)) {
 885                         struct pglist_data *pgdat = page_pgdat(page);
 886
 887                         if (pgdat != locked_pgdat) {
 888                                 if (locked_pgdat)
 889                                         spin_unlock_irqrestore(&locked_pgdat->lru_lock,
 890                                                                         flags);
 891                                 lock_batch = 0;
 892                                 locked_pgdat = pgdat;
 893                                 spin_lock_irqsave(&locked_pgdat->lru_lock, flags);
 894                         }
 895
 896                         lruvec = mem_cgroup_page_lruvec(page, locked_pgdat);
 897                         VM_BUG_ON_PAGE(!PageLRU(page), page);
 898                         __ClearPageLRU(page);
 899                         del_page_from_lru_list(page, lruvec, page_off_lru(page));
 900                 }
 901
 902                 /* Clear Active bit in case of parallel mark_page_accessed */
 903                 __ClearPageActive(page);
 904                 __ClearPageWaiters(page);
 905
 906                 list_add(&page->lru, &pages_to_free);
 907         }
 908         if (locked_pgdat)
 909                 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 910
 911         mem_cgroup_uncharge_list(&pages_to_free);
 912         free_unref_page_list(&pages_to_free);
 913 }
 914 EXPORT_SYMBOL(release_pages);
 915
 916 /*
 917  * The pages which we're about to release may be in the deferred lru-addition
 918  * queues.  That would prevent them from really being freed right now.  That's
 919  * OK from a correctness point of view but is inefficient - those pages may be
 920  * cache-warm and we want to give them back to the page allocator ASAP.
 921  *
 922  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 923  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 924  * mutual recursion.
 925  */
 926 void __pagevec_release(struct pagevec *pvec)
 927 {
 928         if (!pvec->percpu_pvec_drained) {
 929                 lru_add_drain();
 930                 pvec->percpu_pvec_drained = true;
 931         }
 932         release_pages(pvec->pages, pagevec_count(pvec));
 933         pagevec_reinit(pvec);
 934 }
 935 EXPORT_SYMBOL(__pagevec_release);
 936
 937 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 938 /* used by __split_huge_page_refcount() */
 939 void lru_add_page_tail(struct page *page, struct page *page_tail,
 940                        struct lruvec *lruvec, struct list_head *list)
 941 {
 942         VM_BUG_ON_PAGE(!PageHead(page), page);
 943         VM_BUG_ON_PAGE(PageCompound(page_tail), page);
 944         VM_BUG_ON_PAGE(PageLRU(page_tail), page);
 945         lockdep_assert_held(&lruvec_pgdat(lruvec)->lru_lock);
 946
 947         if (!list)
 948                 SetPageLRU(page_tail);
 949
 950         if (likely(PageLRU(page)))
 951                 list_add_tail(&page_tail->lru, &page->lru);
 952         else if (list) {
 953                 /* page reclaim is reclaiming a huge page */
 954                 get_page(page_tail);
 955                 list_add_tail(&page_tail->lru, list);
 956         } else {
 957                 /*
 958                  * Head page has not yet been counted, as an hpage,
 959                  * so we must account for each subpage individually.
 960                  *
 961                  * Put page_tail on the list at the correct position
 962                  * so they all end up in order.
 963                  */
 964                 add_page_to_lru_list_tail(page_tail, lruvec,
 965                                           page_lru(page_tail));
 966         }
 967 }
 968 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 969
 970 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
 971                                  void *arg)
 972 {
 973         enum lru_list lru;
 974         int was_unevictable = TestClearPageUnevictable(page);
 975         int nr_pages = hpage_nr_pages(page);
 976
 977         VM_BUG_ON_PAGE(PageLRU(page), page);
 978
 979         /*
 980          * Page becomes evictable in two ways:
 981          * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
 982          * 2) Before acquiring LRU lock to put the page to correct LRU and then
 983          *   a) do PageLRU check with lock [check_move_unevictable_pages]
 984          *   b) do PageLRU check before lock [clear_page_mlock]
 985          *
 986          * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
 987          * following strict ordering:
 988          *
 989          * #0: __pagevec_lru_add_fn             #1: clear_page_mlock
 990          *
 991          * SetPageLRU()                         TestClearPageMlocked()
 992          * smp_mb() // explicit ordering        // above provides strict
 993          *                                      // ordering
 994          * PageMlocked()                        PageLRU()
 995          *
 996          *
 997          * if '#1' does not observe setting of PG_lru by '#0' and fails
 998          * isolation, the explicit barrier will make sure that page_evictable
 999          * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
1000          * can be reordered after PageMlocked check and can make '#1' to fail
1001          * the isolation of the page whose Mlocked bit is cleared (#0 is also
1002          * looking at the same page) and the evictable page will be stranded
1003          * in an unevictable LRU.
1004          */
1005         SetPageLRU(page);
1006         smp_mb__after_atomic();
1007
1008         if (page_evictable(page)) {
1009                 lru = page_lru(page);
1010                 if (was_unevictable)
1011                         __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
1012         } else {
1013                 lru = LRU_UNEVICTABLE;
1014                 ClearPageActive(page);
1015                 SetPageUnevictable(page);
1016                 if (!was_unevictable)
1017                         __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
1018         }
1019
1020         add_page_to_lru_list(page, lruvec, lru);
1021         trace_mm_lru_insertion(page, lru);
1022 }
1023
1024 /*
1025  * Add the passed pages to the LRU, then drop the caller's refcount
1026  * on them.  Reinitialises the caller's pagevec.
1027  */
1028 void __pagevec_lru_add(struct pagevec *pvec)
1029 {
1030         pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
1031 }
1032
1033 /**
1034  * pagevec_lookup_entries - gang pagecache lookup
1035  * @pvec:       Where the resulting entries are placed
1036  * @mapping:    The address_space to search
1037  * @start:      The starting entry index
1038  * @nr_entries: The maximum number of pages
1039  * @indices:    The cache indices corresponding to the entries in @pvec
1040  *
1041  * pagevec_lookup_entries() will search for and return a group of up
1042  * to @nr_pages pages and shadow entries in the mapping.  All
1043  * entries are placed in @pvec.  pagevec_lookup_entries() takes a
1044  * reference against actual pages in @pvec.
1045  *
1046  * The search returns a group of mapping-contiguous entries with
1047  * ascending indexes.  There may be holes in the indices due to
1048  * not-present entries.
1049  *
1050  * Only one subpage of a Transparent Huge Page is returned in one call:
1051  * allowing truncate_inode_pages_range() to evict the whole THP without
1052  * cycling through a pagevec of extra references.
1053  *
1054  * pagevec_lookup_entries() returns the number of entries which were
1055  * found.
1056  */
1057 unsigned pagevec_lookup_entries(struct pagevec *pvec,
1058                                 struct address_space *mapping,
1059                                 pgoff_t start, unsigned nr_entries,
1060                                 pgoff_t *indices)
1061 {
1062         pvec->nr = find_get_entries(mapping, start, nr_entries,
1063                                     pvec->pages, indices);
1064         return pagevec_count(pvec);
1065 }
1066
1067 /**
1068  * pagevec_remove_exceptionals - pagevec exceptionals pruning
1069  * @pvec:       The pagevec to prune
1070  *
1071  * pagevec_lookup_entries() fills both pages and exceptional radix
1072  * tree entries into the pagevec.  This function prunes all
1073  * exceptionals from @pvec without leaving holes, so that it can be
1074  * passed on to page-only pagevec operations.
1075  */
1076 void pagevec_remove_exceptionals(struct pagevec *pvec)
1077 {
1078         int i, j;
1079
1080         for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
1081                 struct page *page = pvec->pages[i];
1082                 if (!xa_is_value(page))
1083                         pvec->pages[j++] = page;
1084         }
1085         pvec->nr = j;
1086 }
1087
1088 /**
1089  * pagevec_lookup_range - gang pagecache lookup
1090  * @pvec:       Where the resulting pages are placed
1091  * @mapping:    The address_space to search
1092  * @start:      The starting page index
1093  * @end:        The final page index
1094  *
1095  * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1096  * pages in the mapping starting from index @start and upto index @end
1097  * (inclusive).  The pages are placed in @pvec.  pagevec_lookup() takes a
1098  * reference against the pages in @pvec.
1099  *
1100  * The search returns a group of mapping-contiguous pages with ascending
1101  * indexes.  There may be holes in the indices due to not-present pages. We
1102  * also update @start to index the next page for the traversal.
1103  *
1104  * pagevec_lookup_range() returns the number of pages which were found. If this
1105  * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1106  * reached.
1107  */
1108 unsigned pagevec_lookup_range(struct pagevec *pvec,
1109                 struct address_space *mapping, pgoff_t *start, pgoff_t end)
1110 {
1111         pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1112                                         pvec->pages);
1113         return pagevec_count(pvec);
1114 }
1115 EXPORT_SYMBOL(pagevec_lookup_range);
1116
1117 unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1118                 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1119                 xa_mark_t tag)
1120 {
1121         pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1122                                         PAGEVEC_SIZE, pvec->pages);
1123         return pagevec_count(pvec);
1124 }
1125 EXPORT_SYMBOL(pagevec_lookup_range_tag);
1126
1127 unsigned pagevec_lookup_range_nr_tag(struct pagevec *pvec,
1128                 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1129                 xa_mark_t tag, unsigned max_pages)
1130 {
1131         pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1132                 min_t(unsigned int, max_pages, PAGEVEC_SIZE), pvec->pages);
1133         return pagevec_count(pvec);
1134 }
1135 EXPORT_SYMBOL(pagevec_lookup_range_nr_tag);
1136 /*
1137  * Perform any setup for the swap system
1138  */
1139 void __init swap_setup(void)
1140 {
1141         unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1142
1143         /* Use a smaller cluster for small-memory machines */
1144         if (megs < 16)
1145                 page_cluster = 2;
1146         else
1147                 page_cluster = 3;
1148         /*
1149          * Right now other parts of the system means that we
1150          * _really_ don't want to cluster much more
1151          */
1152 }
1153
1154 #ifdef CONFIG_DEV_PAGEMAP_OPS
1155 void put_devmap_managed_page(struct page *page)
1156 {
1157         int count;
1158
1159         if (WARN_ON_ONCE(!page_is_devmap_managed(page)))
1160                 return;
1161
1162         count = page_ref_dec_return(page);
1163
1164         /*
1165          * devmap page refcounts are 1-based, rather than 0-based: if
1166          * refcount is 1, then the page is free and the refcount is
1167          * stable because nobody holds a reference on the page.
1168          */
1169         if (count == 1)
1170                 free_devmap_managed_page(page);
1171         else if (!count)
1172                 __put_page(page);
1173 }
1174 EXPORT_SYMBOL(put_devmap_managed_page);
1175 #endif