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) += thp_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), thp_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 = thp_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 static 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 static 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                 workingset_activation(page);
 447         }
 448         if (page_is_idle(page))
 449                 clear_page_idle(page);
 450 }
 451 EXPORT_SYMBOL(mark_page_accessed);
 452
 453 /**
 454  * lru_cache_add - add a page to a page list
 455  * @page: the page to be added to the LRU.
 456  *
 457  * Queue the page for addition to the LRU via pagevec. The decision on whether
 458  * to add the page to the [in]active [file|anon] list is deferred until the
 459  * pagevec is drained. This gives a chance for the caller of lru_cache_add()
 460  * have the page added to the active list using mark_page_accessed().
 461  */
 462 void lru_cache_add(struct page *page)
 463 {
 464         struct pagevec *pvec;
 465
 466         VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
 467         VM_BUG_ON_PAGE(PageLRU(page), page);
 468
 469         get_page(page);
 470         local_lock(&lru_pvecs.lock);
 471         pvec = this_cpu_ptr(&lru_pvecs.lru_add);
 472         if (!pagevec_add(pvec, page) || PageCompound(page))
 473                 __pagevec_lru_add(pvec);
 474         local_unlock(&lru_pvecs.lock);
 475 }
 476 EXPORT_SYMBOL(lru_cache_add);
 477
 478 /**
 479  * lru_cache_add_inactive_or_unevictable
 480  * @page:  the page to be added to LRU
 481  * @vma:   vma in which page is mapped for determining reclaimability
 482  *
 483  * Place @page on the inactive or unevictable LRU list, depending on its
 484  * evictability.
 485  */
 486 void lru_cache_add_inactive_or_unevictable(struct page *page,
 487                                          struct vm_area_struct *vma)
 488 {
 489         bool unevictable;
 490
 491         VM_BUG_ON_PAGE(PageLRU(page), page);
 492
 493         unevictable = (vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED;
 494         if (unlikely(unevictable) && !TestSetPageMlocked(page)) {
 495                 int nr_pages = thp_nr_pages(page);
 496                 /*
 497                  * We use the irq-unsafe __mod_zone_page_stat because this
 498                  * counter is not modified from interrupt context, and the pte
 499                  * lock is held(spinlock), which implies preemption disabled.
 500                  */
 501                 __mod_zone_page_state(page_zone(page), NR_MLOCK, nr_pages);
 502                 count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
 503         }
 504         lru_cache_add(page);
 505 }
 506
 507 /*
 508  * If the page can not be invalidated, it is moved to the
 509  * inactive list to speed up its reclaim.  It is moved to the
 510  * head of the list, rather than the tail, to give the flusher
 511  * threads some time to write it out, as this is much more
 512  * effective than the single-page writeout from reclaim.
 513  *
 514  * If the page isn't page_mapped and dirty/writeback, the page
 515  * could reclaim asap using PG_reclaim.
 516  *
 517  * 1. active, mapped page -> none
 518  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 519  * 3. inactive, mapped page -> none
 520  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 521  * 5. inactive, clean -> inactive, tail
 522  * 6. Others -> none
 523  *
 524  * In 4, why it moves inactive's head, the VM expects the page would
 525  * be write it out by flusher threads as this is much more effective
 526  * than the single-page writeout from reclaim.
 527  */
 528 static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec,
 529                               void *arg)
 530 {
 531         int lru;
 532         bool active;
 533         int nr_pages = thp_nr_pages(page);
 534
 535         if (!PageLRU(page))
 536                 return;
 537
 538         if (PageUnevictable(page))
 539                 return;
 540
 541         /* Some processes are using the page */
 542         if (page_mapped(page))
 543                 return;
 544
 545         active = PageActive(page);
 546         lru = page_lru_base_type(page);
 547
 548         del_page_from_lru_list(page, lruvec, lru + active);
 549         ClearPageActive(page);
 550         ClearPageReferenced(page);
 551
 552         if (PageWriteback(page) || PageDirty(page)) {
 553                 /*
 554                  * PG_reclaim could be raced with end_page_writeback
 555                  * It can make readahead confusing.  But race window
 556                  * is _really_ small and  it's non-critical problem.
 557                  */
 558                 add_page_to_lru_list(page, lruvec, lru);
 559                 SetPageReclaim(page);
 560         } else {
 561                 /*
 562                  * The page's writeback ends up during pagevec
 563                  * We moves tha page into tail of inactive.
 564                  */
 565                 add_page_to_lru_list_tail(page, lruvec, lru);
 566                 __count_vm_events(PGROTATED, nr_pages);
 567         }
 568
 569         if (active) {
 570                 __count_vm_events(PGDEACTIVATE, nr_pages);
 571                 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
 572                                      nr_pages);
 573         }
 574 }
 575
 576 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 577                             void *arg)
 578 {
 579         if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
 580                 int lru = page_lru_base_type(page);
 581                 int nr_pages = thp_nr_pages(page);
 582
 583                 del_page_from_lru_list(page, lruvec, lru + LRU_ACTIVE);
 584                 ClearPageActive(page);
 585                 ClearPageReferenced(page);
 586                 add_page_to_lru_list(page, lruvec, lru);
 587
 588                 __count_vm_events(PGDEACTIVATE, nr_pages);
 589                 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
 590                                      nr_pages);
 591         }
 592 }
 593
 594 static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec,
 595                             void *arg)
 596 {
 597         if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
 598             !PageSwapCache(page) && !PageUnevictable(page)) {
 599                 bool active = PageActive(page);
 600                 int nr_pages = thp_nr_pages(page);
 601
 602                 del_page_from_lru_list(page, lruvec,
 603                                        LRU_INACTIVE_ANON + active);
 604                 ClearPageActive(page);
 605                 ClearPageReferenced(page);
 606                 /*
 607                  * Lazyfree pages are clean anonymous pages.  They have
 608                  * PG_swapbacked flag cleared, to distinguish them from normal
 609                  * anonymous pages
 610                  */
 611                 ClearPageSwapBacked(page);
 612                 add_page_to_lru_list(page, lruvec, LRU_INACTIVE_FILE);
 613
 614                 __count_vm_events(PGLAZYFREE, nr_pages);
 615                 __count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
 616                                      nr_pages);
 617         }
 618 }
 619
 620 /*
 621  * Drain pages out of the cpu's pagevecs.
 622  * Either "cpu" is the current CPU, and preemption has already been
 623  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 624  */
 625 void lru_add_drain_cpu(int cpu)
 626 {
 627         struct pagevec *pvec = &per_cpu(lru_pvecs.lru_add, cpu);
 628
 629         if (pagevec_count(pvec))
 630                 __pagevec_lru_add(pvec);
 631
 632         pvec = &per_cpu(lru_rotate.pvec, cpu);
 633         /* Disabling interrupts below acts as a compiler barrier. */
 634         if (data_race(pagevec_count(pvec))) {
 635                 unsigned long flags;
 636
 637                 /* No harm done if a racing interrupt already did this */
 638                 local_lock_irqsave(&lru_rotate.lock, flags);
 639                 pagevec_move_tail(pvec);
 640                 local_unlock_irqrestore(&lru_rotate.lock, flags);
 641         }
 642
 643         pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
 644         if (pagevec_count(pvec))
 645                 pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
 646
 647         pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
 648         if (pagevec_count(pvec))
 649                 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 650
 651         pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
 652         if (pagevec_count(pvec))
 653                 pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
 654
 655         activate_page_drain(cpu);
 656 }
 657
 658 /**
 659  * deactivate_file_page - forcefully deactivate a file page
 660  * @page: page to deactivate
 661  *
 662  * This function hints the VM that @page is a good reclaim candidate,
 663  * for example if its invalidation fails due to the page being dirty
 664  * or under writeback.
 665  */
 666 void deactivate_file_page(struct page *page)
 667 {
 668         /*
 669          * In a workload with many unevictable page such as mprotect,
 670          * unevictable page deactivation for accelerating reclaim is pointless.
 671          */
 672         if (PageUnevictable(page))
 673                 return;
 674
 675         if (likely(get_page_unless_zero(page))) {
 676                 struct pagevec *pvec;
 677
 678                 local_lock(&lru_pvecs.lock);
 679                 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
 680
 681                 if (!pagevec_add(pvec, page) || PageCompound(page))
 682                         pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
 683                 local_unlock(&lru_pvecs.lock);
 684         }
 685 }
 686
 687 /*
 688  * deactivate_page - deactivate a page
 689  * @page: page to deactivate
 690  *
 691  * deactivate_page() moves @page to the inactive list if @page was on the active
 692  * list and was not an unevictable page.  This is done to accelerate the reclaim
 693  * of @page.
 694  */
 695 void deactivate_page(struct page *page)
 696 {
 697         if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
 698                 struct pagevec *pvec;
 699
 700                 local_lock(&lru_pvecs.lock);
 701                 pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
 702                 get_page(page);
 703                 if (!pagevec_add(pvec, page) || PageCompound(page))
 704                         pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 705                 local_unlock(&lru_pvecs.lock);
 706         }
 707 }
 708
 709 /**
 710  * mark_page_lazyfree - make an anon page lazyfree
 711  * @page: page to deactivate
 712  *
 713  * mark_page_lazyfree() moves @page to the inactive file list.
 714  * This is done to accelerate the reclaim of @page.
 715  */
 716 void mark_page_lazyfree(struct page *page)
 717 {
 718         if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
 719             !PageSwapCache(page) && !PageUnevictable(page)) {
 720                 struct pagevec *pvec;
 721
 722                 local_lock(&lru_pvecs.lock);
 723                 pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
 724                 get_page(page);
 725                 if (!pagevec_add(pvec, page) || PageCompound(page))
 726                         pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
 727                 local_unlock(&lru_pvecs.lock);
 728         }
 729 }
 730
 731 void lru_add_drain(void)
 732 {
 733         local_lock(&lru_pvecs.lock);
 734         lru_add_drain_cpu(smp_processor_id());
 735         local_unlock(&lru_pvecs.lock);
 736 }
 737
 738 void lru_add_drain_cpu_zone(struct zone *zone)
 739 {
 740         local_lock(&lru_pvecs.lock);
 741         lru_add_drain_cpu(smp_processor_id());
 742         drain_local_pages(zone);
 743         local_unlock(&lru_pvecs.lock);
 744 }
 745
 746 #ifdef CONFIG_SMP
 747
 748 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 749
 750 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 751 {
 752         lru_add_drain();
 753 }
 754
 755 /*
 756  * Doesn't need any cpu hotplug locking because we do rely on per-cpu
 757  * kworkers being shut down before our page_alloc_cpu_dead callback is
 758  * executed on the offlined cpu.
 759  * Calling this function with cpu hotplug locks held can actually lead
 760  * to obscure indirect dependencies via WQ context.
 761  */
 762 void lru_add_drain_all(void)
 763 {
 764         /*
 765          * lru_drain_gen - Global pages generation number
 766          *
 767          * (A) Definition: global lru_drain_gen = x implies that all generations
 768          *     0 < n <= x are already *scheduled* for draining.
 769          *
 770          * This is an optimization for the highly-contended use case where a
 771          * user space workload keeps constantly generating a flow of pages for
 772          * each CPU.
 773          */
 774         static unsigned int lru_drain_gen;
 775         static struct cpumask has_work;
 776         static DEFINE_MUTEX(lock);
 777         unsigned cpu, this_gen;
 778
 779         /*
 780          * Make sure nobody triggers this path before mm_percpu_wq is fully
 781          * initialized.
 782          */
 783         if (WARN_ON(!mm_percpu_wq))
 784                 return;
 785
 786         /*
 787          * Guarantee pagevec counter stores visible by this CPU are visible to
 788          * other CPUs before loading the current drain generation.
 789          */
 790         smp_mb();
 791
 792         /*
 793          * (B) Locally cache global LRU draining generation number
 794          *
 795          * The read barrier ensures that the counter is loaded before the mutex
 796          * is taken. It pairs with smp_mb() inside the mutex critical section
 797          * at (D).
 798          */
 799         this_gen = smp_load_acquire(&lru_drain_gen);
 800
 801         mutex_lock(&lock);
 802
 803         /*
 804          * (C) Exit the draining operation if a newer generation, from another
 805          * lru_add_drain_all(), was already scheduled for draining. Check (A).
 806          */
 807         if (unlikely(this_gen != lru_drain_gen))
 808                 goto done;
 809
 810         /*
 811          * (D) Increment global generation number
 812          *
 813          * Pairs with smp_load_acquire() at (B), outside of the critical
 814          * section. Use a full memory barrier to guarantee that the new global
 815          * drain generation number is stored before loading pagevec counters.
 816          *
 817          * This pairing must be done here, before the for_each_online_cpu loop
 818          * below which drains the page vectors.
 819          *
 820          * Let x, y, and z represent some system CPU numbers, where x < y < z.
 821          * Assume CPU #z is is in the middle of the for_each_online_cpu loop
 822          * below and has already reached CPU #y's per-cpu data. CPU #x comes
 823          * along, adds some pages to its per-cpu vectors, then calls
 824          * lru_add_drain_all().
 825          *
 826          * If the paired barrier is done at any later step, e.g. after the
 827          * loop, CPU #x will just exit at (C) and miss flushing out all of its
 828          * added pages.
 829          */
 830         WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
 831         smp_mb();
 832
 833         cpumask_clear(&has_work);
 834         for_each_online_cpu(cpu) {
 835                 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 836
 837                 if (pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
 838                     data_race(pagevec_count(&per_cpu(lru_rotate.pvec, cpu))) ||
 839                     pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
 840                     pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
 841                     pagevec_count(&per_cpu(lru_pvecs.lru_lazyfree, cpu)) ||
 842                     need_activate_page_drain(cpu)) {
 843                         INIT_WORK(work, lru_add_drain_per_cpu);
 844                         queue_work_on(cpu, mm_percpu_wq, work);
 845                         __cpumask_set_cpu(cpu, &has_work);
 846                 }
 847         }
 848
 849         for_each_cpu(cpu, &has_work)
 850                 flush_work(&per_cpu(lru_add_drain_work, cpu));
 851
 852 done:
 853         mutex_unlock(&lock);
 854 }
 855 #else
 856 void lru_add_drain_all(void)
 857 {
 858         lru_add_drain();
 859 }
 860 #endif /* CONFIG_SMP */
 861
 862 /**
 863  * release_pages - batched put_page()
 864  * @pages: array of pages to release
 865  * @nr: number of pages
 866  *
 867  * Decrement the reference count on all the pages in @pages.  If it
 868  * fell to zero, remove the page from the LRU and free it.
 869  */
 870 void release_pages(struct page **pages, int nr)
 871 {
 872         int i;
 873         LIST_HEAD(pages_to_free);
 874         struct pglist_data *locked_pgdat = NULL;
 875         struct lruvec *lruvec;
 876         unsigned long flags;
 877         unsigned int lock_batch;
 878
 879         for (i = 0; i < nr; i++) {
 880                 struct page *page = pages[i];
 881
 882                 /*
 883                  * Make sure the IRQ-safe lock-holding time does not get
 884                  * excessive with a continuous string of pages from the
 885                  * same pgdat. The lock is held only if pgdat != NULL.
 886                  */
 887                 if (locked_pgdat && ++lock_batch == SWAP_CLUSTER_MAX) {
 888                         spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 889                         locked_pgdat = NULL;
 890                 }
 891
 892                 page = compound_head(page);
 893                 if (is_huge_zero_page(page))
 894                         continue;
 895
 896                 if (is_zone_device_page(page)) {
 897                         if (locked_pgdat) {
 898                                 spin_unlock_irqrestore(&locked_pgdat->lru_lock,
 899                                                        flags);
 900                                 locked_pgdat = NULL;
 901                         }
 902                         /*
 903                          * ZONE_DEVICE pages that return 'false' from
 904                          * page_is_devmap_managed() do not require special
 905                          * processing, and instead, expect a call to
 906                          * put_page_testzero().
 907                          */
 908                         if (page_is_devmap_managed(page)) {
 909                                 put_devmap_managed_page(page);
 910                                 continue;
 911                         }
 912                 }
 913
 914                 if (!put_page_testzero(page))
 915                         continue;
 916
 917                 if (PageCompound(page)) {
 918                         if (locked_pgdat) {
 919                                 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 920                                 locked_pgdat = NULL;
 921                         }
 922                         __put_compound_page(page);
 923                         continue;
 924                 }
 925
 926                 if (PageLRU(page)) {
 927                         struct pglist_data *pgdat = page_pgdat(page);
 928
 929                         if (pgdat != locked_pgdat) {
 930                                 if (locked_pgdat)
 931                                         spin_unlock_irqrestore(&locked_pgdat->lru_lock,
 932                                                                         flags);
 933                                 lock_batch = 0;
 934                                 locked_pgdat = pgdat;
 935                                 spin_lock_irqsave(&locked_pgdat->lru_lock, flags);
 936                         }
 937
 938                         lruvec = mem_cgroup_page_lruvec(page, locked_pgdat);
 939                         VM_BUG_ON_PAGE(!PageLRU(page), page);
 940                         __ClearPageLRU(page);
 941                         del_page_from_lru_list(page, lruvec, page_off_lru(page));
 942                 }
 943
 944                 __ClearPageWaiters(page);
 945
 946                 list_add(&page->lru, &pages_to_free);
 947         }
 948         if (locked_pgdat)
 949                 spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 950
 951         mem_cgroup_uncharge_list(&pages_to_free);
 952         free_unref_page_list(&pages_to_free);
 953 }
 954 EXPORT_SYMBOL(release_pages);
 955
 956 /*
 957  * The pages which we're about to release may be in the deferred lru-addition
 958  * queues.  That would prevent them from really being freed right now.  That's
 959  * OK from a correctness point of view but is inefficient - those pages may be
 960  * cache-warm and we want to give them back to the page allocator ASAP.
 961  *
 962  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 963  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 964  * mutual recursion.
 965  */
 966 void __pagevec_release(struct pagevec *pvec)
 967 {
 968         if (!pvec->percpu_pvec_drained) {
 969                 lru_add_drain();
 970                 pvec->percpu_pvec_drained = true;
 971         }
 972         release_pages(pvec->pages, pagevec_count(pvec));
 973         pagevec_reinit(pvec);
 974 }
 975 EXPORT_SYMBOL(__pagevec_release);
 976
 977 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 978 /* used by __split_huge_page_refcount() */
 979 void lru_add_page_tail(struct page *page, struct page *page_tail,
 980                        struct lruvec *lruvec, struct list_head *list)
 981 {
 982         VM_BUG_ON_PAGE(!PageHead(page), page);
 983         VM_BUG_ON_PAGE(PageCompound(page_tail), page);
 984         VM_BUG_ON_PAGE(PageLRU(page_tail), page);
 985         lockdep_assert_held(&lruvec_pgdat(lruvec)->lru_lock);
 986
 987         if (!list)
 988                 SetPageLRU(page_tail);
 989
 990         if (likely(PageLRU(page)))
 991                 list_add_tail(&page_tail->lru, &page->lru);
 992         else if (list) {
 993                 /* page reclaim is reclaiming a huge page */
 994                 get_page(page_tail);
 995                 list_add_tail(&page_tail->lru, list);
 996         } else {
 997                 /*
 998                  * Head page has not yet been counted, as an hpage,
 999                  * so we must account for each subpage individually.
1000                  *
1001                  * Put page_tail on the list at the correct position
1002                  * so they all end up in order.
1003                  */
1004                 add_page_to_lru_list_tail(page_tail, lruvec,
1005                                           page_lru(page_tail));
1006         }
1007 }
1008 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1009
1010 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
1011                                  void *arg)
1012 {
1013         enum lru_list lru;
1014         int was_unevictable = TestClearPageUnevictable(page);
1015         int nr_pages = thp_nr_pages(page);
1016
1017         VM_BUG_ON_PAGE(PageLRU(page), page);
1018
1019         /*
1020          * Page becomes evictable in two ways:
1021          * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
1022          * 2) Before acquiring LRU lock to put the page to correct LRU and then
1023          *   a) do PageLRU check with lock [check_move_unevictable_pages]
1024          *   b) do PageLRU check before lock [clear_page_mlock]
1025          *
1026          * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
1027          * following strict ordering:
1028          *
1029          * #0: __pagevec_lru_add_fn             #1: clear_page_mlock
1030          *
1031          * SetPageLRU()                         TestClearPageMlocked()
1032          * smp_mb() // explicit ordering        // above provides strict
1033          *                                      // ordering
1034          * PageMlocked()                        PageLRU()
1035          *
1036          *
1037          * if '#1' does not observe setting of PG_lru by '#0' and fails
1038          * isolation, the explicit barrier will make sure that page_evictable
1039          * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
1040          * can be reordered after PageMlocked check and can make '#1' to fail
1041          * the isolation of the page whose Mlocked bit is cleared (#0 is also
1042          * looking at the same page) and the evictable page will be stranded
1043          * in an unevictable LRU.
1044          */
1045         SetPageLRU(page);
1046         smp_mb__after_atomic();
1047
1048         if (page_evictable(page)) {
1049                 lru = page_lru(page);
1050                 if (was_unevictable)
1051                         __count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
1052         } else {
1053                 lru = LRU_UNEVICTABLE;
1054                 ClearPageActive(page);
1055                 SetPageUnevictable(page);
1056                 if (!was_unevictable)
1057                         __count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
1058         }
1059
1060         add_page_to_lru_list(page, lruvec, lru);
1061         trace_mm_lru_insertion(page, lru);
1062 }
1063
1064 /*
1065  * Add the passed pages to the LRU, then drop the caller's refcount
1066  * on them.  Reinitialises the caller's pagevec.
1067  */
1068 void __pagevec_lru_add(struct pagevec *pvec)
1069 {
1070         pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
1071 }
1072
1073 /**
1074  * pagevec_lookup_entries - gang pagecache lookup
1075  * @pvec:       Where the resulting entries are placed
1076  * @mapping:    The address_space to search
1077  * @start:      The starting entry index
1078  * @nr_entries: The maximum number of pages
1079  * @indices:    The cache indices corresponding to the entries in @pvec
1080  *
1081  * pagevec_lookup_entries() will search for and return a group of up
1082  * to @nr_pages pages and shadow entries in the mapping.  All
1083  * entries are placed in @pvec.  pagevec_lookup_entries() takes a
1084  * reference against actual pages in @pvec.
1085  *
1086  * The search returns a group of mapping-contiguous entries with
1087  * ascending indexes.  There may be holes in the indices due to
1088  * not-present entries.
1089  *
1090  * Only one subpage of a Transparent Huge Page is returned in one call:
1091  * allowing truncate_inode_pages_range() to evict the whole THP without
1092  * cycling through a pagevec of extra references.
1093  *
1094  * pagevec_lookup_entries() returns the number of entries which were
1095  * found.
1096  */
1097 unsigned pagevec_lookup_entries(struct pagevec *pvec,
1098                                 struct address_space *mapping,
1099                                 pgoff_t start, unsigned nr_entries,
1100                                 pgoff_t *indices)
1101 {
1102         pvec->nr = find_get_entries(mapping, start, nr_entries,
1103                                     pvec->pages, indices);
1104         return pagevec_count(pvec);
1105 }
1106
1107 /**
1108  * pagevec_remove_exceptionals - pagevec exceptionals pruning
1109  * @pvec:       The pagevec to prune
1110  *
1111  * pagevec_lookup_entries() fills both pages and exceptional radix
1112  * tree entries into the pagevec.  This function prunes all
1113  * exceptionals from @pvec without leaving holes, so that it can be
1114  * passed on to page-only pagevec operations.
1115  */
1116 void pagevec_remove_exceptionals(struct pagevec *pvec)
1117 {
1118         int i, j;
1119
1120         for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
1121                 struct page *page = pvec->pages[i];
1122                 if (!xa_is_value(page))
1123                         pvec->pages[j++] = page;
1124         }
1125         pvec->nr = j;
1126 }
1127
1128 /**
1129  * pagevec_lookup_range - gang pagecache lookup
1130  * @pvec:       Where the resulting pages are placed
1131  * @mapping:    The address_space to search
1132  * @start:      The starting page index
1133  * @end:        The final page index
1134  *
1135  * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1136  * pages in the mapping starting from index @start and upto index @end
1137  * (inclusive).  The pages are placed in @pvec.  pagevec_lookup() takes a
1138  * reference against the pages in @pvec.
1139  *
1140  * The search returns a group of mapping-contiguous pages with ascending
1141  * indexes.  There may be holes in the indices due to not-present pages. We
1142  * also update @start to index the next page for the traversal.
1143  *
1144  * pagevec_lookup_range() returns the number of pages which were found. If this
1145  * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1146  * reached.
1147  */
1148 unsigned pagevec_lookup_range(struct pagevec *pvec,
1149                 struct address_space *mapping, pgoff_t *start, pgoff_t end)
1150 {
1151         pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1152                                         pvec->pages);
1153         return pagevec_count(pvec);
1154 }
1155 EXPORT_SYMBOL(pagevec_lookup_range);
1156
1157 unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1158                 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1159                 xa_mark_t tag)
1160 {
1161         pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1162                                         PAGEVEC_SIZE, pvec->pages);
1163         return pagevec_count(pvec);
1164 }
1165 EXPORT_SYMBOL(pagevec_lookup_range_tag);
1166
1167 unsigned pagevec_lookup_range_nr_tag(struct pagevec *pvec,
1168                 struct address_space *mapping, pgoff_t *index, pgoff_t end,
1169                 xa_mark_t tag, unsigned max_pages)
1170 {
1171         pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1172                 min_t(unsigned int, max_pages, PAGEVEC_SIZE), pvec->pages);
1173         return pagevec_count(pvec);
1174 }
1175 EXPORT_SYMBOL(pagevec_lookup_range_nr_tag);
1176 /*
1177  * Perform any setup for the swap system
1178  */
1179 void __init swap_setup(void)
1180 {
1181         unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1182
1183         /* Use a smaller cluster for small-memory machines */
1184         if (megs < 16)
1185                 page_cluster = 2;
1186         else
1187                 page_cluster = 3;
1188         /*
1189          * Right now other parts of the system means that we
1190          * _really_ don't want to cluster much more
1191          */
1192 }
1193
1194 #ifdef CONFIG_DEV_PAGEMAP_OPS
1195 void put_devmap_managed_page(struct page *page)
1196 {
1197         int count;
1198
1199         if (WARN_ON_ONCE(!page_is_devmap_managed(page)))
1200                 return;
1201
1202         count = page_ref_dec_return(page);
1203
1204         /*
1205          * devmap page refcounts are 1-based, rather than 0-based: if
1206          * refcount is 1, then the page is free and the refcount is
1207          * stable because nobody holds a reference on the page.
1208          */
1209         if (count == 1)
1210                 free_devmap_managed_page(page);
1211         else if (!count)
1212                 __put_page(page);
1213 }
1214 EXPORT_SYMBOL(put_devmap_managed_page);
1215 #endif