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         compound_page_dtor *dtor;
 106
 107         /*
 108          * __page_cache_release() is supposed to be called for thp, not for
 109          * hugetlb. This is because hugetlb page does never have PageLRU set
 110          * (it's never listed to any LRU lists) and no memcg routines should
 111          * be called for hugetlb (it has a separate hugetlb_cgroup.)
 112          */
 113         if (!PageHuge(page))
 114                 __page_cache_release(page);
 115         dtor = get_compound_page_dtor(page);
 116         (*dtor)(page);
 117 }
 118
 119 void __put_page(struct page *page)
 120 {
 121         if (is_zone_device_page(page)) {
 122                 put_dev_pagemap(page->pgmap);
 123
 124                 /*
 125                  * The page belongs to the device that created pgmap. Do
 126                  * not return it to page allocator.
 127                  */
 128                 return;
 129         }
 130
 131         if (unlikely(PageCompound(page)))
 132                 __put_compound_page(page);
 133         else
 134                 __put_single_page(page);
 135 }
 136 EXPORT_SYMBOL(__put_page);
 137
 138 /**
 139  * put_pages_list() - release a list of pages
 140  * @pages: list of pages threaded on page->lru
 141  *
 142  * Release a list of pages which are strung together on page.lru.  Currently
 143  * used by read_cache_pages() and related error recovery code.
 144  */
 145 void put_pages_list(struct list_head *pages)
 146 {
 147         while (!list_empty(pages)) {
 148                 struct page *victim;
 149
 150                 victim = lru_to_page(pages);
 151                 list_del(&victim->lru);
 152                 put_page(victim);
 153         }
 154 }
 155 EXPORT_SYMBOL(put_pages_list);
 156
 157 /*
 158  * get_kernel_pages() - pin kernel pages in memory
 159  * @kiov:       An array of struct kvec structures
 160  * @nr_segs:    number of segments to pin
 161  * @write:      pinning for read/write, currently ignored
 162  * @pages:      array that receives pointers to the pages pinned.
 163  *              Should be at least nr_segs long.
 164  *
 165  * Returns number of pages pinned. This may be fewer than the number
 166  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 167  * were pinned, returns -errno. Each page returned must be released
 168  * with a put_page() call when it is finished with.
 169  */
 170 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
 171                 struct page **pages)
 172 {
 173         int seg;
 174
 175         for (seg = 0; seg < nr_segs; seg++) {
 176                 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
 177                         return seg;
 178
 179                 pages[seg] = kmap_to_page(kiov[seg].iov_base);
 180                 get_page(pages[seg]);
 181         }
 182
 183         return seg;
 184 }
 185 EXPORT_SYMBOL_GPL(get_kernel_pages);
 186
 187 /*
 188  * get_kernel_page() - pin a kernel page in memory
 189  * @start:      starting kernel address
 190  * @write:      pinning for read/write, currently ignored
 191  * @pages:      array that receives pointer to the page pinned.
 192  *              Must be at least nr_segs long.
 193  *
 194  * Returns 1 if page is pinned. If the page was not pinned, returns
 195  * -errno. The page returned must be released with a put_page() call
 196  * when it is finished with.
 197  */
 198 int get_kernel_page(unsigned long start, int write, struct page **pages)
 199 {
 200         const struct kvec kiov = {
 201                 .iov_base = (void *)start,
 202                 .iov_len = PAGE_SIZE
 203         };
 204
 205         return get_kernel_pages(&kiov, 1, write, pages);
 206 }
 207 EXPORT_SYMBOL_GPL(get_kernel_page);
 208
 209 static void pagevec_lru_move_fn(struct pagevec *pvec,
 210         void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
 211         void *arg)
 212 {
 213         int i;
 214         struct pglist_data *pgdat = NULL;
 215         struct lruvec *lruvec;
 216         unsigned long flags = 0;
 217
 218         for (i = 0; i < pagevec_count(pvec); i++) {
 219                 struct page *page = pvec->pages[i];
 220                 struct pglist_data *pagepgdat = page_pgdat(page);
 221
 222                 if (pagepgdat != pgdat) {
 223                         if (pgdat)
 224                                 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
 225                         pgdat = pagepgdat;
 226                         spin_lock_irqsave(&pgdat->lru_lock, flags);
 227                 }
 228
 229                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
 230                 (*move_fn)(page, lruvec, arg);
 231         }
 232         if (pgdat)
 233                 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
 234         release_pages(pvec->pages, pvec->nr);
 235         pagevec_reinit(pvec);
 236 }
 237
 238 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
 239                                  void *arg)
 240 {
 241         int *pgmoved = arg;
 242
 243         if (PageLRU(page) && !PageUnevictable(page)) {
 244                 del_page_from_lru_list(page, lruvec, page_lru(page));
 245                 ClearPageActive(page);
 246                 add_page_to_lru_list_tail(page, lruvec, page_lru(page));
 247                 (*pgmoved)++;
 248         }
 249 }
 250
 251 /*
 252  * pagevec_move_tail() must be called with IRQ disabled.
 253  * Otherwise this may cause nasty races.
 254  */
 255 static void pagevec_move_tail(struct pagevec *pvec)
 256 {
 257         int pgmoved = 0;
 258
 259         pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
 260         __count_vm_events(PGROTATED, pgmoved);
 261 }
 262
 263 /*
 264  * Writeback is about to end against a page which has been marked for immediate
 265  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 266  * inactive list.
 267  */
 268 void rotate_reclaimable_page(struct page *page)
 269 {
 270         if (!PageLocked(page) && !PageDirty(page) &&
 271             !PageUnevictable(page) && PageLRU(page)) {
 272                 struct pagevec *pvec;
 273                 unsigned long flags;
 274
 275                 get_page(page);
 276                 local_lock_irqsave(&lru_rotate.lock, flags);
 277                 pvec = this_cpu_ptr(&lru_rotate.pvec);
 278                 if (!pagevec_add(pvec, page) || PageCompound(page))
 279                         pagevec_move_tail(pvec);
 280                 local_unlock_irqrestore(&lru_rotate.lock, flags);
 281         }
 282 }
 283
 284 static void update_page_reclaim_stat(struct lruvec *lruvec,
 285                                      int file, int rotated)
 286 {
 287         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
 288
 289         reclaim_stat->recent_scanned[file]++;
 290         if (rotated)
 291                 reclaim_stat->recent_rotated[file]++;
 292 }
 293
 294 static void __activate_page(struct page *page, struct lruvec *lruvec,
 295                             void *arg)
 296 {
 297         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 298                 int file = page_is_file_lru(page);
 299                 int lru = page_lru_base_type(page);
 300
 301                 del_page_from_lru_list(page, lruvec, lru);
 302                 SetPageActive(page);
 303                 lru += LRU_ACTIVE;
 304                 add_page_to_lru_list(page, lruvec, lru);
 305                 trace_mm_lru_activate(page);
 306
 307                 __count_vm_event(PGACTIVATE);
 308                 update_page_reclaim_stat(lruvec, file, 1);
 309         }
 310 }
 311
 312 #ifdef CONFIG_SMP
 313 static void activate_page_drain(int cpu)
 314 {
 315         struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
 316
 317         if (pagevec_count(pvec))
 318                 pagevec_lru_move_fn(pvec, __activate_page, NULL);
 319 }
 320
 321 static bool need_activate_page_drain(int cpu)
 322 {
 323         return pagevec_count(&per_cpu(lru_pvecs.activate_page, cpu)) != 0;
 324 }
 325
 326 void activate_page(struct page *page)
 327 {
 328         page = compound_head(page);
 329         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 330                 struct pagevec *pvec;
 331
 332                 local_lock(&lru_pvecs.lock);
 333                 pvec = this_cpu_ptr(&lru_pvecs.activate_page);
 334                 get_page(page);
 335                 if (!pagevec_add(pvec, page) || PageCompound(page))
 336                         pagevec_lru_move_fn(pvec, __activate_page, NULL);
 337                 local_unlock(&lru_pvecs.lock);
 338         }
 339 }
 340
 341 #else
 342 static inline void activate_page_drain(int cpu)
 343 {
 344 }
 345
 346 void activate_page(struct page *page)
 347 {
 348         pg_data_t *pgdat = page_pgdat(page);
 349
 350         page = compound_head(page);
 351         spin_lock_irq(&pgdat->lru_lock);
 352         __activate_page(page, mem_cgroup_page_lruvec(page, pgdat), NULL);
 353         spin_unlock_irq(&pgdat->lru_lock);
 354 }
 355 #endif
 356
 357 static void __lru_cache_activate_page(struct page *page)
 358 {
 359         struct pagevec *pvec;
 360         int i;
 361
 362         local_lock(&lru_pvecs.lock);
 363         pvec = this_cpu_ptr(&lru_pvecs.lru_add);
 364
 365         /*
 366          * Search backwards on the optimistic assumption that the page being
 367          * activated has just been added to this pagevec. Note that only
 368          * the local pagevec is examined as a !PageLRU page could be in the
 369          * process of being released, reclaimed, migrated or on a remote
 370          * pagevec that is currently being drained. Furthermore, marking
 371          * a remote pagevec's page PageActive potentially hits a race where
 372          * a page is marked PageActive just after it is added to the inactive
 373          * list causing accounting errors and BUG_ON checks to trigger.
 374          */
 375         for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
 376                 struct page *pagevec_page = pvec->pages[i];
 377
 378                 if (pagevec_page == page) {
 379                         SetPageActive(page);
 380                         break;
 381                 }
 382         }
 383
 384         local_unlock(&lru_pvecs.lock);
 385 }
 386
 387 /*
 388  * Mark a page as having seen activity.
 389  *
 390  * inactive,unreferenced        ->      inactive,referenced
 391  * inactive,referenced          ->      active,unreferenced
 392  * active,unreferenced          ->      active,referenced
 393  *
 394  * When a newly allocated page is not yet visible, so safe for non-atomic ops,
 395  * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
 396  */
 397 void mark_page_accessed(struct page *page)
 398 {
 399         page = compound_head(page);
 400
 401         if (!PageReferenced(page)) {
 402                 SetPageReferenced(page);
 403         } else if (PageUnevictable(page)) {
 404                 /*
 405                  * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
 406                  * this list is never rotated or maintained, so marking an
 407                  * evictable page accessed has no effect.
 408                  */
 409         } else if (!PageActive(page)) {
 410                 /*
 411                  * If the page is on the LRU, queue it for activation via
 412                  * lru_pvecs.activate_page. Otherwise, assume the page is on a
 413                  * pagevec, mark it active and it'll be moved to the active
 414                  * LRU on the next drain.
 415                  */
 416                 if (PageLRU(page))
 417                         activate_page(page);
 418                 else
 419                         __lru_cache_activate_page(page);
 420                 ClearPageReferenced(page);
 421                 if (page_is_file_lru(page))
 422                         workingset_activation(page);
 423         }
 424         if (page_is_idle(page))
 425                 clear_page_idle(page);
 426 }
 427 EXPORT_SYMBOL(mark_page_accessed);
 428
 429 static void __lru_cache_add(struct page *page)
 430 {
 431         struct pagevec *pvec;
 432
 433         local_lock(&lru_pvecs.lock);
 434         pvec = this_cpu_ptr(&lru_pvecs.lru_add);
 435         get_page(page);
 436         if (!pagevec_add(pvec, page) || PageCompound(page))
 437                 __pagevec_lru_add(pvec);
 438         local_unlock(&lru_pvecs.lock);
 439 }
 440
 441 /**
 442  * lru_cache_add_anon - add a page to the page lists
 443  * @page: the page to add
 444  */
 445 void lru_cache_add_anon(struct page *page)
 446 {
 447         if (PageActive(page))
 448                 ClearPageActive(page);
 449         __lru_cache_add(page);
 450 }
 451
 452 void lru_cache_add_file(struct page *page)
 453 {
 454         if (PageActive(page))
 455                 ClearPageActive(page);
 456         __lru_cache_add(page);
 457 }
 458 EXPORT_SYMBOL(lru_cache_add_file);
 459
 460 /**
 461  * lru_cache_add - add a page to a page list
 462  * @page: the page to be added to the LRU.
 463  *
 464  * Queue the page for addition to the LRU via pagevec. The decision on whether
 465  * to add the page to the [in]active [file|anon] list is deferred until the
 466  * pagevec is drained. This gives a chance for the caller of lru_cache_add()
 467  * have the page added to the active list using mark_page_accessed().
 468  */
 469 void lru_cache_add(struct page *page)
 470 {
 471         VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
 472         VM_BUG_ON_PAGE(PageLRU(page), page);
 473         __lru_cache_add(page);
 474 }
 475
 476 /**
 477  * lru_cache_add_active_or_unevictable
 478  * @page:  the page to be added to LRU
 479  * @vma:   vma in which page is mapped for determining reclaimability
 480  *
 481  * Place @page on the active or unevictable LRU list, depending on its
 482  * evictability.  Note that if the page is not evictable, it goes
 483  * directly back onto it's zone's unevictable list, it does NOT use a
 484  * per cpu pagevec.
 485  */
 486 void lru_cache_add_active_or_unevictable(struct page *page,
 487                                          struct vm_area_struct *vma)
 488 {
 489         VM_BUG_ON_PAGE(PageLRU(page), page);
 490
 491         if (likely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) != VM_LOCKED))
 492                 SetPageActive(page);
 493         else if (!TestSetPageMlocked(page)) {
 494                 /*
 495                  * We use the irq-unsafe __mod_zone_page_stat because this
 496                  * counter is not modified from interrupt context, and the pte
 497                  * lock is held(spinlock), which implies preemption disabled.
 498                  */
 499                 __mod_zone_page_state(page_zone(page), NR_MLOCK,
 500                                     hpage_nr_pages(page));
 501                 count_vm_event(UNEVICTABLE_PGMLOCKED);
 502         }
 503         lru_cache_add(page);
 504 }
 505
 506 /*
 507  * If the page can not be invalidated, it is moved to the
 508  * inactive list to speed up its reclaim.  It is moved to the
 509  * head of the list, rather than the tail, to give the flusher
 510  * threads some time to write it out, as this is much more
 511  * effective than the single-page writeout from reclaim.
 512  *
 513  * If the page isn't page_mapped and dirty/writeback, the page
 514  * could reclaim asap using PG_reclaim.
 515  *
 516  * 1. active, mapped page -> none
 517  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 518  * 3. inactive, mapped page -> none
 519  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 520  * 5. inactive, clean -> inactive, tail
 521  * 6. Others -> none
 522  *
 523  * In 4, why it moves inactive's head, the VM expects the page would
 524  * be write it out by flusher threads as this is much more effective
 525  * than the single-page writeout from reclaim.
 526  */
 527 static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec,
 528                               void *arg)
 529 {
 530         int lru, file;
 531         bool active;
 532
 533         if (!PageLRU(page))
 534                 return;
 535
 536         if (PageUnevictable(page))
 537                 return;
 538
 539         /* Some processes are using the page */
 540         if (page_mapped(page))
 541                 return;
 542
 543         active = PageActive(page);
 544         file = page_is_file_lru(page);
 545         lru = page_lru_base_type(page);
 546
 547         del_page_from_lru_list(page, lruvec, lru + active);
 548         ClearPageActive(page);
 549         ClearPageReferenced(page);
 550
 551         if (PageWriteback(page) || PageDirty(page)) {
 552                 /*
 553                  * PG_reclaim could be raced with end_page_writeback
 554                  * It can make readahead confusing.  But race window
 555                  * is _really_ small and  it's non-critical problem.
 556                  */
 557                 add_page_to_lru_list(page, lruvec, lru);
 558                 SetPageReclaim(page);
 559         } else {
 560                 /*
 561                  * The page's writeback ends up during pagevec
 562                  * We moves tha page into tail of inactive.
 563                  */
 564                 add_page_to_lru_list_tail(page, lruvec, lru);
 565                 __count_vm_event(PGROTATED);
 566         }
 567
 568         if (active)
 569                 __count_vm_event(PGDEACTIVATE);
 570         update_page_reclaim_stat(lruvec, file, 0);
 571 }
 572
 573 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 574                             void *arg)
 575 {
 576         if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
 577                 int file = page_is_file_lru(page);
 578                 int lru = page_lru_base_type(page);
 579
 580                 del_page_from_lru_list(page, lruvec, lru + LRU_ACTIVE);
 581                 ClearPageActive(page);
 582                 ClearPageReferenced(page);
 583                 add_page_to_lru_list(page, lruvec, lru);
 584
 585                 __count_vm_events(PGDEACTIVATE, hpage_nr_pages(page));
 586                 update_page_reclaim_stat(lruvec, file, 0);
 587         }
 588 }
 589
 590 static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec,
 591                             void *arg)
 592 {
 593         if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
 594             !PageSwapCache(page) && !PageUnevictable(page)) {
 595                 bool active = PageActive(page);
 596
 597                 del_page_from_lru_list(page, lruvec,
 598                                        LRU_INACTIVE_ANON + active);
 599                 ClearPageActive(page);
 600                 ClearPageReferenced(page);
 601                 /*
 602                  * Lazyfree pages are clean anonymous pages.  They have
 603                  * PG_swapbacked flag cleared, to distinguish them from normal
 604                  * anonymous pages
 605                  */
 606                 ClearPageSwapBacked(page);
 607                 add_page_to_lru_list(page, lruvec, LRU_INACTIVE_FILE);
 608
 609                 __count_vm_events(PGLAZYFREE, hpage_nr_pages(page));
 610                 count_memcg_page_event(page, PGLAZYFREE);
 611                 update_page_reclaim_stat(lruvec, 1, 0);
 612         }
 613 }
 614
 615 /*
 616  * Drain pages out of the cpu's pagevecs.
 617  * Either "cpu" is the current CPU, and preemption has already been
 618  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 619  */
 620 void lru_add_drain_cpu(int cpu)
 621 {
 622         struct pagevec *pvec = &per_cpu(lru_pvecs.lru_add, cpu);
 623
 624         if (pagevec_count(pvec))
 625                 __pagevec_lru_add(pvec);
 626
 627         pvec = &per_cpu(lru_rotate.pvec, cpu);
 628         if (pagevec_count(pvec)) {
 629                 unsigned long flags;
 630
 631                 /* No harm done if a racing interrupt already did this */
 632                 local_lock_irqsave(&lru_rotate.lock, flags);
 633                 pagevec_move_tail(pvec);
 634                 local_unlock_irqrestore(&lru_rotate.lock, flags);
 635         }
 636
 637         pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
 638         if (pagevec_count(pvec))
 639                 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
 640
 641         pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
 642         if (pagevec_count(pvec))
 643                 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 644
 645         pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
 646         if (pagevec_count(pvec))
 647                 pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
 648
 649         activate_page_drain(cpu);
 650 }
 651
 652 /**
 653  * deactivate_file_page - forcefully deactivate a file page
 654  * @page: page to deactivate
 655  *
 656  * This function hints the VM that @page is a good reclaim candidate,
 657  * for example if its invalidation fails due to the page being dirty
 658  * or under writeback.
 659  */
 660 void deactivate_file_page(struct page *page)
 661 {
 662         /*
 663          * In a workload with many unevictable page such as mprotect,
 664          * unevictable page deactivation for accelerating reclaim is pointless.
 665          */
 666         if (PageUnevictable(page))
 667                 return;
 668
 669         if (likely(get_page_unless_zero(page))) {
 670                 struct pagevec *pvec;
 671
 672                 local_lock(&lru_pvecs.lock);
 673                 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
 674
 675                 if (!pagevec_add(pvec, page) || PageCompound(page))
 676                         pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
 677                 local_unlock(&lru_pvecs.lock);
 678         }
 679 }
 680
 681 /*
 682  * deactivate_page - deactivate a page
 683  * @page: page to deactivate
 684  *
 685  * deactivate_page() moves @page to the inactive list if @page was on the active
 686  * list and was not an unevictable page.  This is done to accelerate the reclaim
 687  * of @page.
 688  */
 689 void deactivate_page(struct page *page)
 690 {
 691         if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
 692                 struct pagevec *pvec;
 693
 694                 local_lock(&lru_pvecs.lock);
 695                 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
 696                 get_page(page);
 697                 if (!pagevec_add(pvec, page) || PageCompound(page))
 698                         pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 699                 local_unlock(&lru_pvecs.lock);
 700         }
 701 }
 702
 703 /**
 704  * mark_page_lazyfree - make an anon page lazyfree
 705  * @page: page to deactivate
 706  *
 707  * mark_page_lazyfree() moves @page to the inactive file list.
 708  * This is done to accelerate the reclaim of @page.
 709  */
 710 void mark_page_lazyfree(struct page *page)
 711 {
 712         if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
 713             !PageSwapCache(page) && !PageUnevictable(page)) {
 714                 struct pagevec *pvec;
 715
 716                 local_lock(&lru_pvecs.lock);
 717                 pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
 718                 get_page(page);
 719                 if (!pagevec_add(pvec, page) || PageCompound(page))
 720                         pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
 721                 local_unlock(&lru_pvecs.lock);
 722         }
 723 }
 724
 725 void lru_add_drain(void)
 726 {
 727         local_lock(&lru_pvecs.lock);
 728         lru_add_drain_cpu(smp_processor_id());
 729         local_unlock(&lru_pvecs.lock);
 730 }
 731
 732 void lru_add_drain_cpu_zone(struct zone *zone)
 733 {
 734         local_lock(&lru_pvecs.lock);
 735         lru_add_drain_cpu(smp_processor_id());
 736         drain_local_pages(zone);
 737         local_unlock(&lru_pvecs.lock);
 738 }
 739
 740 #ifdef CONFIG_SMP
 741
 742 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 743
 744 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 745 {
 746         lru_add_drain();
 747 }
 748
 749 /*
 750  * Doesn't need any cpu hotplug locking because we do rely on per-cpu
 751  * kworkers being shut down before our page_alloc_cpu_dead callback is
 752  * executed on the offlined cpu.
 753  * Calling this function with cpu hotplug locks held can actually lead
 754  * to obscure indirect dependencies via WQ context.
 755  */
 756 void lru_add_drain_all(void)
 757 {
 758         static seqcount_t seqcount = SEQCNT_ZERO(seqcount);
 759         static DEFINE_MUTEX(lock);
 760         static struct cpumask has_work;
 761         int cpu, seq;
 762
 763         /*
 764          * Make sure nobody triggers this path before mm_percpu_wq is fully
 765          * initialized.
 766          */
 767         if (WARN_ON(!mm_percpu_wq))
 768                 return;
 769
 770         seq = raw_read_seqcount_latch(&seqcount);
 771
 772         mutex_lock(&lock);
 773
 774         /*
 775          * Piggyback on drain started and finished while we waited for lock:
 776          * all pages pended at the time of our enter were drained from vectors.
 777          */
 778         if (__read_seqcount_retry(&seqcount, seq))
 779                 goto done;
 780
 781         raw_write_seqcount_latch(&seqcount);
 782
 783         cpumask_clear(&has_work);
 784
 785         for_each_online_cpu(cpu) {
 786                 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 787
 788                 if (pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
 789                     pagevec_count(&per_cpu(lru_rotate.pvec, cpu)) ||
 790                     pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
 791                     pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
 792                     pagevec_count(&per_cpu(lru_pvecs.lru_lazyfree, cpu)) ||
 793                     need_activate_page_drain(cpu)) {
 794                         INIT_WORK(work, lru_add_drain_per_cpu);
 795                         queue_work_on(cpu, mm_percpu_wq, work);
 796                         cpumask_set_cpu(cpu, &has_work);
 797                 }
 798         }
 799
 800         for_each_cpu(cpu, &has_work)
 801                 flush_work(&per_cpu(lru_add_drain_work, cpu));
 802
 803 done:
 804         mutex_unlock(&lock);
 805 }
 806 #else
 807 void lru_add_drain_all(void)
 808 {
 809         lru_add_drain();
 810 }
 811 #endif
 812
 813 /**
 814  * release_pages - batched put_page()
 815  * @pages: array of pages to release
 816  * @nr: number of pages
 817  *
 818  * Decrement the reference count on all the pages in @pages.  If it
 819  * fell to zero, remove the page from the LRU and free it.
 820  */
 821 void release_pages(struct page **pages, int nr)
 822 {
 823         int i;
 824         LIST_HEAD(pages_to_free);
 825         struct pglist_data *locked_pgdat = NULL;
 826         struct lruvec *lruvec;
 827         unsigned long uninitialized_var(flags);
 828         unsigned int uninitialized_var(lock_batch);
 829
 830         for (i = 0; i < nr; i++) {
 831                 struct page *page = pages[i];
 832
 833                 /*
 834                  * Make sure the IRQ-safe lock-holding time does not get
 835                  * excessive with a continuous string of pages from the
 836                  * same pgdat. The lock is held only if pgdat != NULL.
 837                  */
 838                 if (locked_pgdat && ++lock_batch == SWAP_CLUSTER_MAX) {
 839                         spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 840                         locked_pgdat = NULL;
 841                 }
 842
 843                 if (is_huge_zero_page(page))
 844                         continue;
 845
 846                 if (is_zone_device_page(page)) {
 847                         if (locked_pgdat) {
 848                                 spin_unlock_irqrestore(&locked_pgdat->lru_lock,
 849                                                        flags);
 850                                 locked_pgdat = NULL;
 851                         }
 852                         /*
 853                          * ZONE_DEVICE pages that return 'false' from
 854                          * put_devmap_managed_page() do not require special
 855                          * processing, and instead, expect a call to
 856                          * put_page_testzero().
 857                          */
 858                         if (page_is_devmap_managed(page)) {
 859                                 put_devmap_managed_page(page);
 860                                 continue;
 861                         }
 862                 }
 863
 864                 page = compound_head(page);
 865                 if (!put_page_testzero(page))
 866                         continue;
 867
 868                 if (PageCompound(page)) {
 869                         if (locked_pgdat) {
 870                                 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 871                                 locked_pgdat = NULL;
 872                         }
 873                         __put_compound_page(page);
 874                         continue;
 875                 }
 876
 877                 if (PageLRU(page)) {
 878                         struct pglist_data *pgdat = page_pgdat(page);
 879
 880                         if (pgdat != locked_pgdat) {
 881                                 if (locked_pgdat)
 882                                         spin_unlock_irqrestore(&locked_pgdat->lru_lock,
 883                                                                         flags);
 884                                 lock_batch = 0;
 885                                 locked_pgdat = pgdat;
 886                                 spin_lock_irqsave(&locked_pgdat->lru_lock, flags);
 887                         }
 888
 889                         lruvec = mem_cgroup_page_lruvec(page, locked_pgdat);
 890                         VM_BUG_ON_PAGE(!PageLRU(page), page);
 891                         __ClearPageLRU(page);
 892                         del_page_from_lru_list(page, lruvec, page_off_lru(page));
 893                 }
 894
 895                 /* Clear Active bit in case of parallel mark_page_accessed */
 896                 __ClearPageActive(page);
 897                 __ClearPageWaiters(page);
 898
 899                 list_add(&page->lru, &pages_to_free);
 900         }
 901         if (locked_pgdat)
 902                 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 903
 904         mem_cgroup_uncharge_list(&pages_to_free);
 905         free_unref_page_list(&pages_to_free);
 906 }
 907 EXPORT_SYMBOL(release_pages);
 908
 909 /*
 910  * The pages which we're about to release may be in the deferred lru-addition
 911  * queues.  That would prevent them from really being freed right now.  That's
 912  * OK from a correctness point of view but is inefficient - those pages may be
 913  * cache-warm and we want to give them back to the page allocator ASAP.
 914  *
 915  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 916  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 917  * mutual recursion.
 918  */
 919 void __pagevec_release(struct pagevec *pvec)
 920 {
 921         if (!pvec->percpu_pvec_drained) {
 922                 lru_add_drain();
 923                 pvec->percpu_pvec_drained = true;
 924         }
 925         release_pages(pvec->pages, pagevec_count(pvec));
 926         pagevec_reinit(pvec);
 927 }
 928 EXPORT_SYMBOL(__pagevec_release);
 929
 930 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 931 /* used by __split_huge_page_refcount() */
 932 void lru_add_page_tail(struct page *page, struct page *page_tail,
 933                        struct lruvec *lruvec, struct list_head *list)
 934 {
 935         const int file = 0;
 936
 937         VM_BUG_ON_PAGE(!PageHead(page), page);
 938         VM_BUG_ON_PAGE(PageCompound(page_tail), page);
 939         VM_BUG_ON_PAGE(PageLRU(page_tail), page);
 940         lockdep_assert_held(&lruvec_pgdat(lruvec)->lru_lock);
 941
 942         if (!list)
 943                 SetPageLRU(page_tail);
 944
 945         if (likely(PageLRU(page)))
 946                 list_add_tail(&page_tail->lru, &page->lru);
 947         else if (list) {
 948                 /* page reclaim is reclaiming a huge page */
 949                 get_page(page_tail);
 950                 list_add_tail(&page_tail->lru, list);
 951         } else {
 952                 /*
 953                  * Head page has not yet been counted, as an hpage,
 954                  * so we must account for each subpage individually.
 955                  *
 956                  * Put page_tail on the list at the correct position
 957                  * so they all end up in order.
 958                  */
 959                 add_page_to_lru_list_tail(page_tail, lruvec,
 960                                           page_lru(page_tail));
 961         }
 962
 963         if (!PageUnevictable(page))
 964                 update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
 965 }
 966 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 967
 968 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
 969                                  void *arg)
 970 {
 971         enum lru_list lru;
 972         int was_unevictable = TestClearPageUnevictable(page);
 973
 974         VM_BUG_ON_PAGE(PageLRU(page), page);
 975
 976         /*
 977          * Page becomes evictable in two ways:
 978          * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
 979          * 2) Before acquiring LRU lock to put the page to correct LRU and then
 980          *   a) do PageLRU check with lock [check_move_unevictable_pages]
 981          *   b) do PageLRU check before lock [clear_page_mlock]
 982          *
 983          * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
 984          * following strict ordering:
 985          *
 986          * #0: __pagevec_lru_add_fn             #1: clear_page_mlock
 987          *
 988          * SetPageLRU()                         TestClearPageMlocked()
 989          * smp_mb() // explicit ordering        // above provides strict
 990          *                                      // ordering
 991          * PageMlocked()                        PageLRU()
 992          *
 993          *
 994          * if '#1' does not observe setting of PG_lru by '#0' and fails
 995          * isolation, the explicit barrier will make sure that page_evictable
 996          * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
 997          * can be reordered after PageMlocked check and can make '#1' to fail
 998          * the isolation of the page whose Mlocked bit is cleared (#0 is also
 999          * looking at the same page) and the evictable page will be stranded
1000          * in an unevictable LRU.
1001          */
1002         SetPageLRU(page);
1003         smp_mb__after_atomic();
1004
1005         if (page_evictable(page)) {
1006                 lru = page_lru(page);
1007                 update_page_reclaim_stat(lruvec, page_is_file_lru(page),
1008                                          PageActive(page));
1009                 if (was_unevictable)
1010                         count_vm_event(UNEVICTABLE_PGRESCUED);
1011         } else {
1012                 lru = LRU_UNEVICTABLE;
1013                 ClearPageActive(page);
1014                 SetPageUnevictable(page);
1015                 if (!was_unevictable)
1016                         count_vm_event(UNEVICTABLE_PGCULLED);
1017         }
1018
1019         add_page_to_lru_list(page, lruvec, lru);
1020         trace_mm_lru_insertion(page, lru);
1021 }
1022
1023 /*
1024  * Add the passed pages to the LRU, then drop the caller's refcount
1025  * on them.  Reinitialises the caller's pagevec.
1026  */
1027 void __pagevec_lru_add(struct pagevec *pvec)
1028 {
1029         pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
1030 }
1031
1032 /**
1033  * pagevec_lookup_entries - gang pagecache lookup
1034  * @pvec:       Where the resulting entries are placed
1035  * @mapping:    The address_space to search
1036  * @start:      The starting entry index
1037  * @nr_entries: The maximum number of pages
1038  * @indices:    The cache indices corresponding to the entries in @pvec
1039  *
1040  * pagevec_lookup_entries() will search for and return a group of up
1041  * to @nr_pages pages and shadow entries in the mapping.  All
1042  * entries are placed in @pvec.  pagevec_lookup_entries() takes a
1043  * reference against actual pages in @pvec.
1044  *
1045  * The search returns a group of mapping-contiguous entries with
1046  * ascending indexes.  There may be holes in the indices due to
1047  * not-present entries.
1048  *
1049  * Only one subpage of a Transparent Huge Page is returned in one call:
1050  * allowing truncate_inode_pages_range() to evict the whole THP without
1051  * cycling through a pagevec of extra references.
1052  *
1053  * pagevec_lookup_entries() returns the number of entries which were
1054  * found.
1055  */
1056 unsigned pagevec_lookup_entries(struct pagevec *pvec,
1057                                 struct address_space *mapping,
1058                                 pgoff_t start, unsigned nr_entries,
1059                                 pgoff_t *indices)
1060 {
1061         pvec->nr = find_get_entries(mapping, start, nr_entries,
1062                                     pvec->pages, indices);
1063         return pagevec_count(pvec);
1064 }
1065
1066 /**
1067  * pagevec_remove_exceptionals - pagevec exceptionals pruning
1068  * @pvec:       The pagevec to prune
1069  *
1070  * pagevec_lookup_entries() fills both pages and exceptional radix
1071  * tree entries into the pagevec.  This function prunes all
1072  * exceptionals from @pvec without leaving holes, so that it can be
1073  * passed on to page-only pagevec operations.
1074  */
1075 void pagevec_remove_exceptionals(struct pagevec *pvec)
1076 {
1077         int i, j;
1078
1079         for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
1080                 struct page *page = pvec->pages[i];
1081                 if (!xa_is_value(page))
1082                         pvec->pages[j++] = page;
1083         }
1084         pvec->nr = j;
1085 }
1086
1087 /**
1088  * pagevec_lookup_range - gang pagecache lookup
1089  * @pvec:       Where the resulting pages are placed
1090  * @mapping:    The address_space to search
1091  * @start:      The starting page index
1092  * @end:        The final page index
1093  *
1094  * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1095  * pages in the mapping starting from index @start and upto index @end
1096  * (inclusive).  The pages are placed in @pvec.  pagevec_lookup() takes a
1097  * reference against the pages in @pvec.
1098  *
1099  * The search returns a group of mapping-contiguous pages with ascending
1100  * indexes.  There may be holes in the indices due to not-present pages. We
1101  * also update @start to index the next page for the traversal.
1102  *
1103  * pagevec_lookup_range() returns the number of pages which were found. If this
1104  * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1105  * reached.
1106  */
1107 unsigned pagevec_lookup_range(struct pagevec *pvec,
1108                 struct address_space *mapping, pgoff_t *start, pgoff_t end)
1109 {
1110         pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1111                                         pvec->pages);
1112         return pagevec_count(pvec);
1113 }
1114 EXPORT_SYMBOL(pagevec_lookup_range);
1115
1116 unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1117                 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1118                 xa_mark_t tag)
1119 {
1120         pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1121                                         PAGEVEC_SIZE, pvec->pages);
1122         return pagevec_count(pvec);
1123 }
1124 EXPORT_SYMBOL(pagevec_lookup_range_tag);
1125
1126 unsigned pagevec_lookup_range_nr_tag(struct pagevec *pvec,
1127                 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1128                 xa_mark_t tag, unsigned max_pages)
1129 {
1130         pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1131                 min_t(unsigned int, max_pages, PAGEVEC_SIZE), pvec->pages);
1132         return pagevec_count(pvec);
1133 }
1134 EXPORT_SYMBOL(pagevec_lookup_range_nr_tag);
1135 /*
1136  * Perform any setup for the swap system
1137  */
1138 void __init swap_setup(void)
1139 {
1140         unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1141
1142         /* Use a smaller cluster for small-memory machines */
1143         if (megs < 16)
1144                 page_cluster = 2;
1145         else
1146                 page_cluster = 3;
1147         /*
1148          * Right now other parts of the system means that we
1149          * _really_ don't want to cluster much more
1150          */
1151 }
1152
1153 #ifdef CONFIG_DEV_PAGEMAP_OPS
1154 void put_devmap_managed_page(struct page *page)
1155 {
1156         int count;
1157
1158         if (WARN_ON_ONCE(!page_is_devmap_managed(page)))
1159                 return;
1160
1161         count = page_ref_dec_return(page);
1162
1163         /*
1164          * devmap page refcounts are 1-based, rather than 0-based: if
1165          * refcount is 1, then the page is free and the refcount is
1166          * stable because nobody holds a reference on the page.
1167          */
1168         if (count == 1)
1169                 free_devmap_managed_page(page);
1170         else if (!count)
1171                 __put_page(page);
1172 }
1173 EXPORT_SYMBOL(put_devmap_managed_page);
1174 #endif