mm/swap.c

   1 /*
   2  *  linux/mm/swap.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  */
   6
   7 /*
   8  * This file contains the default values for the operation of the
   9  * Linux VM subsystem. Fine-tuning documentation can be found in
  10  * Documentation/sysctl/vm.txt.
  11  * Started 18.12.91
  12  * Swap aging added 23.2.95, Stephen Tweedie.
  13  * Buffermem limits added 12.3.98, Rik van Riel.
  14  */
  15
  16 #include <linux/mm.h>
  17 #include <linux/sched.h>
  18 #include <linux/kernel_stat.h>
  19 #include <linux/swap.h>
  20 #include <linux/mman.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/pagevec.h>
  23 #include <linux/init.h>
  24 #include <linux/export.h>
  25 #include <linux/mm_inline.h>
  26 #include <linux/percpu_counter.h>
  27 #include <linux/percpu.h>
  28 #include <linux/cpu.h>
  29 #include <linux/notifier.h>
  30 #include <linux/backing-dev.h>
  31 #include <linux/memcontrol.h>
  32 #include <linux/gfp.h>
  33 #include <linux/uio.h>
  34
  35 #include "internal.h"
  36
  37 #define CREATE_TRACE_POINTS
  38 #include <trace/events/pagemap.h>
  39
  40 /* How many pages do we try to swap or page in/out together? */
  41 int page_cluster;
  42
  43 static DEFINE_PER_CPU(struct pagevec, lru_add_pvec);
  44 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
  45 static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
  46
  47 /*
  48  * This path almost never happens for VM activity - pages are normally
  49  * freed via pagevecs.  But it gets used by networking.
  50  */
  51 static void __page_cache_release(struct page *page)
  52 {
  53         if (PageLRU(page)) {
  54                 struct zone *zone = page_zone(page);
  55                 struct lruvec *lruvec;
  56                 unsigned long flags;
  57
  58                 spin_lock_irqsave(&zone->lru_lock, flags);
  59                 lruvec = mem_cgroup_page_lruvec(page, zone);
  60                 VM_BUG_ON_PAGE(!PageLRU(page), page);
  61                 __ClearPageLRU(page);
  62                 del_page_from_lru_list(page, lruvec, page_off_lru(page));
  63                 spin_unlock_irqrestore(&zone->lru_lock, flags);
  64         }
  65 }
  66
  67 static void __put_single_page(struct page *page)
  68 {
  69         __page_cache_release(page);
  70         free_hot_cold_page(page, 0);
  71 }
  72
  73 static void __put_compound_page(struct page *page)
  74 {
  75         compound_page_dtor *dtor;
  76
  77         __page_cache_release(page);
  78         dtor = get_compound_page_dtor(page);
  79         (*dtor)(page);
  80 }
  81
  82 static void put_compound_page(struct page *page)
  83 {
  84         struct page *page_head;
  85
  86         if (likely(!PageTail(page))) {
  87                 if (put_page_testzero(page)) {
  88                         /*
  89                          * By the time all refcounts have been released
  90                          * split_huge_page cannot run anymore from under us.
  91                          */
  92                         if (PageHead(page))
  93                                 __put_compound_page(page);
  94                         else
  95                                 __put_single_page(page);
  96                 }
  97                 return;
  98         }
  99
 100         /* __split_huge_page_refcount can run under us */
 101         page_head = compound_head(page);
 102
 103         /*
 104          * THP can not break up slab pages so avoid taking
 105          * compound_lock() and skip the tail page refcounting (in
 106          * _mapcount) too. Slab performs non-atomic bit ops on
 107          * page->flags for better performance. In particular
 108          * slab_unlock() in slub used to be a hot path. It is still
 109          * hot on arches that do not support
 110          * this_cpu_cmpxchg_double().
 111          *
 112          * If "page" is part of a slab or hugetlbfs page it cannot be
 113          * splitted and the head page cannot change from under us. And
 114          * if "page" is part of a THP page under splitting, if the
 115          * head page pointed by the THP tail isn't a THP head anymore,
 116          * we'll find PageTail clear after smp_rmb() and we'll treat
 117          * it as a single page.
 118          */
 119         if (!__compound_tail_refcounted(page_head)) {
 120                 /*
 121                  * If "page" is a THP tail, we must read the tail page
 122                  * flags after the head page flags. The
 123                  * split_huge_page side enforces write memory barriers
 124                  * between clearing PageTail and before the head page
 125                  * can be freed and reallocated.
 126                  */
 127                 smp_rmb();
 128                 if (likely(PageTail(page))) {
 129                         /*
 130                          * __split_huge_page_refcount cannot race
 131                          * here.
 132                          */
 133                         VM_BUG_ON_PAGE(!PageHead(page_head), page_head);
 134                         VM_BUG_ON_PAGE(page_mapcount(page) != 0, page);
 135                         if (put_page_testzero(page_head)) {
 136                                 /*
 137                                  * If this is the tail of a slab
 138                                  * compound page, the tail pin must
 139                                  * not be the last reference held on
 140                                  * the page, because the PG_slab
 141                                  * cannot be cleared before all tail
 142                                  * pins (which skips the _mapcount
 143                                  * tail refcounting) have been
 144                                  * released. For hugetlbfs the tail
 145                                  * pin may be the last reference on
 146                                  * the page instead, because
 147                                  * PageHeadHuge will not go away until
 148                                  * the compound page enters the buddy
 149                                  * allocator.
 150                                  */
 151                                 VM_BUG_ON_PAGE(PageSlab(page_head), page_head);
 152                                 __put_compound_page(page_head);
 153                         }
 154                         return;
 155                 } else
 156                         /*
 157                          * __split_huge_page_refcount run before us,
 158                          * "page" was a THP tail. The split page_head
 159                          * has been freed and reallocated as slab or
 160                          * hugetlbfs page of smaller order (only
 161                          * possible if reallocated as slab on x86).
 162                          */
 163                         goto out_put_single;
 164         }
 165
 166         if (likely(page != page_head && get_page_unless_zero(page_head))) {
 167                 unsigned long flags;
 168
 169                 /*
 170                  * page_head wasn't a dangling pointer but it may not
 171                  * be a head page anymore by the time we obtain the
 172                  * lock. That is ok as long as it can't be freed from
 173                  * under us.
 174                  */
 175                 flags = compound_lock_irqsave(page_head);
 176                 if (unlikely(!PageTail(page))) {
 177                         /* __split_huge_page_refcount run before us */
 178                         compound_unlock_irqrestore(page_head, flags);
 179                         if (put_page_testzero(page_head)) {
 180                                 /*
 181                                  * The head page may have been freed
 182                                  * and reallocated as a compound page
 183                                  * of smaller order and then freed
 184                                  * again.  All we know is that it
 185                                  * cannot have become: a THP page, a
 186                                  * compound page of higher order, a
 187                                  * tail page.  That is because we
 188                                  * still hold the refcount of the
 189                                  * split THP tail and page_head was
 190                                  * the THP head before the split.
 191                                  */
 192                                 if (PageHead(page_head))
 193                                         __put_compound_page(page_head);
 194                                 else
 195                                         __put_single_page(page_head);
 196                         }
 197 out_put_single:
 198                         if (put_page_testzero(page))
 199                                 __put_single_page(page);
 200                         return;
 201                 }
 202                 VM_BUG_ON_PAGE(page_head != page->first_page, page);
 203                 /*
 204                  * We can release the refcount taken by
 205                  * get_page_unless_zero() now that
 206                  * __split_huge_page_refcount() is blocked on the
 207                  * compound_lock.
 208                  */
 209                 if (put_page_testzero(page_head))
 210                         VM_BUG_ON_PAGE(1, page_head);
 211                 /* __split_huge_page_refcount will wait now */
 212                 VM_BUG_ON_PAGE(page_mapcount(page) <= 0, page);
 213                 atomic_dec(&page->_mapcount);
 214                 VM_BUG_ON_PAGE(atomic_read(&page_head->_count) <= 0, page_head);
 215                 VM_BUG_ON_PAGE(atomic_read(&page->_count) != 0, page);
 216                 compound_unlock_irqrestore(page_head, flags);
 217
 218                 if (put_page_testzero(page_head)) {
 219                         if (PageHead(page_head))
 220                                 __put_compound_page(page_head);
 221                         else
 222                                 __put_single_page(page_head);
 223                 }
 224         } else {
 225                 /* page_head is a dangling pointer */
 226                 VM_BUG_ON_PAGE(PageTail(page), page);
 227                 goto out_put_single;
 228         }
 229 }
 230
 231 void put_page(struct page *page)
 232 {
 233         if (unlikely(PageCompound(page)))
 234                 put_compound_page(page);
 235         else if (put_page_testzero(page))
 236                 __put_single_page(page);
 237 }
 238 EXPORT_SYMBOL(put_page);
 239
 240 /*
 241  * This function is exported but must not be called by anything other
 242  * than get_page(). It implements the slow path of get_page().
 243  */
 244 bool __get_page_tail(struct page *page)
 245 {
 246         /*
 247          * This takes care of get_page() if run on a tail page
 248          * returned by one of the get_user_pages/follow_page variants.
 249          * get_user_pages/follow_page itself doesn't need the compound
 250          * lock because it runs __get_page_tail_foll() under the
 251          * proper PT lock that already serializes against
 252          * split_huge_page().
 253          */
 254         unsigned long flags;
 255         bool got;
 256         struct page *page_head = compound_head(page);
 257
 258         /* Ref to put_compound_page() comment. */
 259         if (!__compound_tail_refcounted(page_head)) {
 260                 smp_rmb();
 261                 if (likely(PageTail(page))) {
 262                         /*
 263                          * This is a hugetlbfs page or a slab
 264                          * page. __split_huge_page_refcount
 265                          * cannot race here.
 266                          */
 267                         VM_BUG_ON_PAGE(!PageHead(page_head), page_head);
 268                         __get_page_tail_foll(page, true);
 269                         return true;
 270                 } else {
 271                         /*
 272                          * __split_huge_page_refcount run
 273                          * before us, "page" was a THP
 274                          * tail. The split page_head has been
 275                          * freed and reallocated as slab or
 276                          * hugetlbfs page of smaller order
 277                          * (only possible if reallocated as
 278                          * slab on x86).
 279                          */
 280                         return false;
 281                 }
 282         }
 283
 284         got = false;
 285         if (likely(page != page_head && get_page_unless_zero(page_head))) {
 286                 /*
 287                  * page_head wasn't a dangling pointer but it
 288                  * may not be a head page anymore by the time
 289                  * we obtain the lock. That is ok as long as it
 290                  * can't be freed from under us.
 291                  */
 292                 flags = compound_lock_irqsave(page_head);
 293                 /* here __split_huge_page_refcount won't run anymore */
 294                 if (likely(PageTail(page))) {
 295                         __get_page_tail_foll(page, false);
 296                         got = true;
 297                 }
 298                 compound_unlock_irqrestore(page_head, flags);
 299                 if (unlikely(!got))
 300                         put_page(page_head);
 301         }
 302         return got;
 303 }
 304 EXPORT_SYMBOL(__get_page_tail);
 305
 306 /**
 307  * put_pages_list() - release a list of pages
 308  * @pages: list of pages threaded on page->lru
 309  *
 310  * Release a list of pages which are strung together on page.lru.  Currently
 311  * used by read_cache_pages() and related error recovery code.
 312  */
 313 void put_pages_list(struct list_head *pages)
 314 {
 315         while (!list_empty(pages)) {
 316                 struct page *victim;
 317
 318                 victim = list_entry(pages->prev, struct page, lru);
 319                 list_del(&victim->lru);
 320                 page_cache_release(victim);
 321         }
 322 }
 323 EXPORT_SYMBOL(put_pages_list);
 324
 325 /*
 326  * get_kernel_pages() - pin kernel pages in memory
 327  * @kiov:       An array of struct kvec structures
 328  * @nr_segs:    number of segments to pin
 329  * @write:      pinning for read/write, currently ignored
 330  * @pages:      array that receives pointers to the pages pinned.
 331  *              Should be at least nr_segs long.
 332  *
 333  * Returns number of pages pinned. This may be fewer than the number
 334  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 335  * were pinned, returns -errno. Each page returned must be released
 336  * with a put_page() call when it is finished with.
 337  */
 338 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
 339                 struct page **pages)
 340 {
 341         int seg;
 342
 343         for (seg = 0; seg < nr_segs; seg++) {
 344                 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
 345                         return seg;
 346
 347                 pages[seg] = kmap_to_page(kiov[seg].iov_base);
 348                 page_cache_get(pages[seg]);
 349         }
 350
 351         return seg;
 352 }
 353 EXPORT_SYMBOL_GPL(get_kernel_pages);
 354
 355 /*
 356  * get_kernel_page() - pin a kernel page in memory
 357  * @start:      starting kernel address
 358  * @write:      pinning for read/write, currently ignored
 359  * @pages:      array that receives pointer to the page pinned.
 360  *              Must be at least nr_segs long.
 361  *
 362  * Returns 1 if page is pinned. If the page was not pinned, returns
 363  * -errno. The page returned must be released with a put_page() call
 364  * when it is finished with.
 365  */
 366 int get_kernel_page(unsigned long start, int write, struct page **pages)
 367 {
 368         const struct kvec kiov = {
 369                 .iov_base = (void *)start,
 370                 .iov_len = PAGE_SIZE
 371         };
 372
 373         return get_kernel_pages(&kiov, 1, write, pages);
 374 }
 375 EXPORT_SYMBOL_GPL(get_kernel_page);
 376
 377 static void pagevec_lru_move_fn(struct pagevec *pvec,
 378         void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
 379         void *arg)
 380 {
 381         int i;
 382         struct zone *zone = NULL;
 383         struct lruvec *lruvec;
 384         unsigned long flags = 0;
 385
 386         for (i = 0; i < pagevec_count(pvec); i++) {
 387                 struct page *page = pvec->pages[i];
 388                 struct zone *pagezone = page_zone(page);
 389
 390                 if (pagezone != zone) {
 391                         if (zone)
 392                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 393                         zone = pagezone;
 394                         spin_lock_irqsave(&zone->lru_lock, flags);
 395                 }
 396
 397                 lruvec = mem_cgroup_page_lruvec(page, zone);
 398                 (*move_fn)(page, lruvec, arg);
 399         }
 400         if (zone)
 401                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 402         release_pages(pvec->pages, pvec->nr, pvec->cold);
 403         pagevec_reinit(pvec);
 404 }
 405
 406 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
 407                                  void *arg)
 408 {
 409         int *pgmoved = arg;
 410
 411         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 412                 enum lru_list lru = page_lru_base_type(page);
 413                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 414                 (*pgmoved)++;
 415         }
 416 }
 417
 418 /*
 419  * pagevec_move_tail() must be called with IRQ disabled.
 420  * Otherwise this may cause nasty races.
 421  */
 422 static void pagevec_move_tail(struct pagevec *pvec)
 423 {
 424         int pgmoved = 0;
 425
 426         pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
 427         __count_vm_events(PGROTATED, pgmoved);
 428 }
 429
 430 /*
 431  * Writeback is about to end against a page which has been marked for immediate
 432  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 433  * inactive list.
 434  */
 435 void rotate_reclaimable_page(struct page *page)
 436 {
 437         if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
 438             !PageUnevictable(page) && PageLRU(page)) {
 439                 struct pagevec *pvec;
 440                 unsigned long flags;
 441
 442                 page_cache_get(page);
 443                 local_irq_save(flags);
 444                 pvec = &__get_cpu_var(lru_rotate_pvecs);
 445                 if (!pagevec_add(pvec, page))
 446                         pagevec_move_tail(pvec);
 447                 local_irq_restore(flags);
 448         }
 449 }
 450
 451 static void update_page_reclaim_stat(struct lruvec *lruvec,
 452                                      int file, int rotated)
 453 {
 454         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
 455
 456         reclaim_stat->recent_scanned[file]++;
 457         if (rotated)
 458                 reclaim_stat->recent_rotated[file]++;
 459 }
 460
 461 static void __activate_page(struct page *page, struct lruvec *lruvec,
 462                             void *arg)
 463 {
 464         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 465                 int file = page_is_file_cache(page);
 466                 int lru = page_lru_base_type(page);
 467
 468                 del_page_from_lru_list(page, lruvec, lru);
 469                 SetPageActive(page);
 470                 lru += LRU_ACTIVE;
 471                 add_page_to_lru_list(page, lruvec, lru);
 472                 trace_mm_lru_activate(page, page_to_pfn(page));
 473
 474                 __count_vm_event(PGACTIVATE);
 475                 update_page_reclaim_stat(lruvec, file, 1);
 476         }
 477 }
 478
 479 #ifdef CONFIG_SMP
 480 static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
 481
 482 static void activate_page_drain(int cpu)
 483 {
 484         struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
 485
 486         if (pagevec_count(pvec))
 487                 pagevec_lru_move_fn(pvec, __activate_page, NULL);
 488 }
 489
 490 static bool need_activate_page_drain(int cpu)
 491 {
 492         return pagevec_count(&per_cpu(activate_page_pvecs, cpu)) != 0;
 493 }
 494
 495 void activate_page(struct page *page)
 496 {
 497         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 498                 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
 499
 500                 page_cache_get(page);
 501                 if (!pagevec_add(pvec, page))
 502                         pagevec_lru_move_fn(pvec, __activate_page, NULL);
 503                 put_cpu_var(activate_page_pvecs);
 504         }
 505 }
 506
 507 #else
 508 static inline void activate_page_drain(int cpu)
 509 {
 510 }
 511
 512 static bool need_activate_page_drain(int cpu)
 513 {
 514         return false;
 515 }
 516
 517 void activate_page(struct page *page)
 518 {
 519         struct zone *zone = page_zone(page);
 520
 521         spin_lock_irq(&zone->lru_lock);
 522         __activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
 523         spin_unlock_irq(&zone->lru_lock);
 524 }
 525 #endif
 526
 527 static void __lru_cache_activate_page(struct page *page)
 528 {
 529         struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
 530         int i;
 531
 532         /*
 533          * Search backwards on the optimistic assumption that the page being
 534          * activated has just been added to this pagevec. Note that only
 535          * the local pagevec is examined as a !PageLRU page could be in the
 536          * process of being released, reclaimed, migrated or on a remote
 537          * pagevec that is currently being drained. Furthermore, marking
 538          * a remote pagevec's page PageActive potentially hits a race where
 539          * a page is marked PageActive just after it is added to the inactive
 540          * list causing accounting errors and BUG_ON checks to trigger.
 541          */
 542         for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
 543                 struct page *pagevec_page = pvec->pages[i];
 544
 545                 if (pagevec_page == page) {
 546                         SetPageActive(page);
 547                         break;
 548                 }
 549         }
 550
 551         put_cpu_var(lru_add_pvec);
 552 }
 553
 554 /*
 555  * Mark a page as having seen activity.
 556  *
 557  * inactive,unreferenced        ->      inactive,referenced
 558  * inactive,referenced          ->      active,unreferenced
 559  * active,unreferenced          ->      active,referenced
 560  */
 561 void mark_page_accessed(struct page *page)
 562 {
 563         if (!PageActive(page) && !PageUnevictable(page) &&
 564                         PageReferenced(page)) {
 565
 566                 /*
 567                  * If the page is on the LRU, queue it for activation via
 568                  * activate_page_pvecs. Otherwise, assume the page is on a
 569                  * pagevec, mark it active and it'll be moved to the active
 570                  * LRU on the next drain.
 571                  */
 572                 if (PageLRU(page))
 573                         activate_page(page);
 574                 else
 575                         __lru_cache_activate_page(page);
 576                 ClearPageReferenced(page);
 577         } else if (!PageReferenced(page)) {
 578                 SetPageReferenced(page);
 579         }
 580 }
 581 EXPORT_SYMBOL(mark_page_accessed);
 582
 583 /*
 584  * Queue the page for addition to the LRU via pagevec. The decision on whether
 585  * to add the page to the [in]active [file|anon] list is deferred until the
 586  * pagevec is drained. This gives a chance for the caller of __lru_cache_add()
 587  * have the page added to the active list using mark_page_accessed().
 588  */
 589 void __lru_cache_add(struct page *page)
 590 {
 591         struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
 592
 593         page_cache_get(page);
 594         if (!pagevec_space(pvec))
 595                 __pagevec_lru_add(pvec);
 596         pagevec_add(pvec, page);
 597         put_cpu_var(lru_add_pvec);
 598 }
 599 EXPORT_SYMBOL(__lru_cache_add);
 600
 601 /**
 602  * lru_cache_add - add a page to a page list
 603  * @page: the page to be added to the LRU.
 604  */
 605 void lru_cache_add(struct page *page)
 606 {
 607         VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
 608         VM_BUG_ON_PAGE(PageLRU(page), page);
 609         __lru_cache_add(page);
 610 }
 611
 612 /**
 613  * add_page_to_unevictable_list - add a page to the unevictable list
 614  * @page:  the page to be added to the unevictable list
 615  *
 616  * Add page directly to its zone's unevictable list.  To avoid races with
 617  * tasks that might be making the page evictable, through eg. munlock,
 618  * munmap or exit, while it's not on the lru, we want to add the page
 619  * while it's locked or otherwise "invisible" to other tasks.  This is
 620  * difficult to do when using the pagevec cache, so bypass that.
 621  */
 622 void add_page_to_unevictable_list(struct page *page)
 623 {
 624         struct zone *zone = page_zone(page);
 625         struct lruvec *lruvec;
 626
 627         spin_lock_irq(&zone->lru_lock);
 628         lruvec = mem_cgroup_page_lruvec(page, zone);
 629         ClearPageActive(page);
 630         SetPageUnevictable(page);
 631         SetPageLRU(page);
 632         add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
 633         spin_unlock_irq(&zone->lru_lock);
 634 }
 635
 636 /*
 637  * If the page can not be invalidated, it is moved to the
 638  * inactive list to speed up its reclaim.  It is moved to the
 639  * head of the list, rather than the tail, to give the flusher
 640  * threads some time to write it out, as this is much more
 641  * effective than the single-page writeout from reclaim.
 642  *
 643  * If the page isn't page_mapped and dirty/writeback, the page
 644  * could reclaim asap using PG_reclaim.
 645  *
 646  * 1. active, mapped page -> none
 647  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 648  * 3. inactive, mapped page -> none
 649  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 650  * 5. inactive, clean -> inactive, tail
 651  * 6. Others -> none
 652  *
 653  * In 4, why it moves inactive's head, the VM expects the page would
 654  * be write it out by flusher threads as this is much more effective
 655  * than the single-page writeout from reclaim.
 656  */
 657 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 658                               void *arg)
 659 {
 660         int lru, file;
 661         bool active;
 662
 663         if (!PageLRU(page))
 664                 return;
 665
 666         if (PageUnevictable(page))
 667                 return;
 668
 669         /* Some processes are using the page */
 670         if (page_mapped(page))
 671                 return;
 672
 673         active = PageActive(page);
 674         file = page_is_file_cache(page);
 675         lru = page_lru_base_type(page);
 676
 677         del_page_from_lru_list(page, lruvec, lru + active);
 678         ClearPageActive(page);
 679         ClearPageReferenced(page);
 680         add_page_to_lru_list(page, lruvec, lru);
 681
 682         if (PageWriteback(page) || PageDirty(page)) {
 683                 /*
 684                  * PG_reclaim could be raced with end_page_writeback
 685                  * It can make readahead confusing.  But race window
 686                  * is _really_ small and  it's non-critical problem.
 687                  */
 688                 SetPageReclaim(page);
 689         } else {
 690                 /*
 691                  * The page's writeback ends up during pagevec
 692                  * We moves tha page into tail of inactive.
 693                  */
 694                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 695                 __count_vm_event(PGROTATED);
 696         }
 697
 698         if (active)
 699                 __count_vm_event(PGDEACTIVATE);
 700         update_page_reclaim_stat(lruvec, file, 0);
 701 }
 702
 703 /*
 704  * Drain pages out of the cpu's pagevecs.
 705  * Either "cpu" is the current CPU, and preemption has already been
 706  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 707  */
 708 void lru_add_drain_cpu(int cpu)
 709 {
 710         struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
 711
 712         if (pagevec_count(pvec))
 713                 __pagevec_lru_add(pvec);
 714
 715         pvec = &per_cpu(lru_rotate_pvecs, cpu);
 716         if (pagevec_count(pvec)) {
 717                 unsigned long flags;
 718
 719                 /* No harm done if a racing interrupt already did this */
 720                 local_irq_save(flags);
 721                 pagevec_move_tail(pvec);
 722                 local_irq_restore(flags);
 723         }
 724
 725         pvec = &per_cpu(lru_deactivate_pvecs, cpu);
 726         if (pagevec_count(pvec))
 727                 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 728
 729         activate_page_drain(cpu);
 730 }
 731
 732 /**
 733  * deactivate_page - forcefully deactivate a page
 734  * @page: page to deactivate
 735  *
 736  * This function hints the VM that @page is a good reclaim candidate,
 737  * for example if its invalidation fails due to the page being dirty
 738  * or under writeback.
 739  */
 740 void deactivate_page(struct page *page)
 741 {
 742         /*
 743          * In a workload with many unevictable page such as mprotect, unevictable
 744          * page deactivation for accelerating reclaim is pointless.
 745          */
 746         if (PageUnevictable(page))
 747                 return;
 748
 749         if (likely(get_page_unless_zero(page))) {
 750                 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
 751
 752                 if (!pagevec_add(pvec, page))
 753                         pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 754                 put_cpu_var(lru_deactivate_pvecs);
 755         }
 756 }
 757
 758 void lru_add_drain(void)
 759 {
 760         lru_add_drain_cpu(get_cpu());
 761         put_cpu();
 762 }
 763
 764 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 765 {
 766         lru_add_drain();
 767 }
 768
 769 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 770
 771 void lru_add_drain_all(void)
 772 {
 773         static DEFINE_MUTEX(lock);
 774         static struct cpumask has_work;
 775         int cpu;
 776
 777         mutex_lock(&lock);
 778         get_online_cpus();
 779         cpumask_clear(&has_work);
 780
 781         for_each_online_cpu(cpu) {
 782                 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 783
 784                 if (pagevec_count(&per_cpu(lru_add_pvec, cpu)) ||
 785                     pagevec_count(&per_cpu(lru_rotate_pvecs, cpu)) ||
 786                     pagevec_count(&per_cpu(lru_deactivate_pvecs, cpu)) ||
 787                     need_activate_page_drain(cpu)) {
 788                         INIT_WORK(work, lru_add_drain_per_cpu);
 789                         schedule_work_on(cpu, work);
 790                         cpumask_set_cpu(cpu, &has_work);
 791                 }
 792         }
 793
 794         for_each_cpu(cpu, &has_work)
 795                 flush_work(&per_cpu(lru_add_drain_work, cpu));
 796
 797         put_online_cpus();
 798         mutex_unlock(&lock);
 799 }
 800
 801 /*
 802  * Batched page_cache_release().  Decrement the reference count on all the
 803  * passed pages.  If it fell to zero then remove the page from the LRU and
 804  * free it.
 805  *
 806  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 807  * for the remainder of the operation.
 808  *
 809  * The locking in this function is against shrink_inactive_list(): we recheck
 810  * the page count inside the lock to see whether shrink_inactive_list()
 811  * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
 812  * will free it.
 813  */
 814 void release_pages(struct page **pages, int nr, int cold)
 815 {
 816         int i;
 817         LIST_HEAD(pages_to_free);
 818         struct zone *zone = NULL;
 819         struct lruvec *lruvec;
 820         unsigned long uninitialized_var(flags);
 821
 822         for (i = 0; i < nr; i++) {
 823                 struct page *page = pages[i];
 824
 825                 if (unlikely(PageCompound(page))) {
 826                         if (zone) {
 827                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 828                                 zone = NULL;
 829                         }
 830                         put_compound_page(page);
 831                         continue;
 832                 }
 833
 834                 if (!put_page_testzero(page))
 835                         continue;
 836
 837                 if (PageLRU(page)) {
 838                         struct zone *pagezone = page_zone(page);
 839
 840                         if (pagezone != zone) {
 841                                 if (zone)
 842                                         spin_unlock_irqrestore(&zone->lru_lock,
 843                                                                         flags);
 844                                 zone = pagezone;
 845                                 spin_lock_irqsave(&zone->lru_lock, flags);
 846                         }
 847
 848                         lruvec = mem_cgroup_page_lruvec(page, zone);
 849                         VM_BUG_ON_PAGE(!PageLRU(page), page);
 850                         __ClearPageLRU(page);
 851                         del_page_from_lru_list(page, lruvec, page_off_lru(page));
 852                 }
 853
 854                 /* Clear Active bit in case of parallel mark_page_accessed */
 855                 ClearPageActive(page);
 856
 857                 list_add(&page->lru, &pages_to_free);
 858         }
 859         if (zone)
 860                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 861
 862         free_hot_cold_page_list(&pages_to_free, cold);
 863 }
 864 EXPORT_SYMBOL(release_pages);
 865
 866 /*
 867  * The pages which we're about to release may be in the deferred lru-addition
 868  * queues.  That would prevent them from really being freed right now.  That's
 869  * OK from a correctness point of view but is inefficient - those pages may be
 870  * cache-warm and we want to give them back to the page allocator ASAP.
 871  *
 872  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 873  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 874  * mutual recursion.
 875  */
 876 void __pagevec_release(struct pagevec *pvec)
 877 {
 878         lru_add_drain();
 879         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 880         pagevec_reinit(pvec);
 881 }
 882 EXPORT_SYMBOL(__pagevec_release);
 883
 884 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 885 /* used by __split_huge_page_refcount() */
 886 void lru_add_page_tail(struct page *page, struct page *page_tail,
 887                        struct lruvec *lruvec, struct list_head *list)
 888 {
 889         const int file = 0;
 890
 891         VM_BUG_ON_PAGE(!PageHead(page), page);
 892         VM_BUG_ON_PAGE(PageCompound(page_tail), page);
 893         VM_BUG_ON_PAGE(PageLRU(page_tail), page);
 894         VM_BUG_ON(NR_CPUS != 1 &&
 895                   !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
 896
 897         if (!list)
 898                 SetPageLRU(page_tail);
 899
 900         if (likely(PageLRU(page)))
 901                 list_add_tail(&page_tail->lru, &page->lru);
 902         else if (list) {
 903                 /* page reclaim is reclaiming a huge page */
 904                 get_page(page_tail);
 905                 list_add_tail(&page_tail->lru, list);
 906         } else {
 907                 struct list_head *list_head;
 908                 /*
 909                  * Head page has not yet been counted, as an hpage,
 910                  * so we must account for each subpage individually.
 911                  *
 912                  * Use the standard add function to put page_tail on the list,
 913                  * but then correct its position so they all end up in order.
 914                  */
 915                 add_page_to_lru_list(page_tail, lruvec, page_lru(page_tail));
 916                 list_head = page_tail->lru.prev;
 917                 list_move_tail(&page_tail->lru, list_head);
 918         }
 919
 920         if (!PageUnevictable(page))
 921                 update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
 922 }
 923 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 924
 925 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
 926                                  void *arg)
 927 {
 928         int file = page_is_file_cache(page);
 929         int active = PageActive(page);
 930         enum lru_list lru = page_lru(page);
 931
 932         VM_BUG_ON_PAGE(PageLRU(page), page);
 933
 934         SetPageLRU(page);
 935         add_page_to_lru_list(page, lruvec, lru);
 936         update_page_reclaim_stat(lruvec, file, active);
 937         trace_mm_lru_insertion(page, page_to_pfn(page), lru, trace_pagemap_flags(page));
 938 }
 939
 940 /*
 941  * Add the passed pages to the LRU, then drop the caller's refcount
 942  * on them.  Reinitialises the caller's pagevec.
 943  */
 944 void __pagevec_lru_add(struct pagevec *pvec)
 945 {
 946         pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
 947 }
 948 EXPORT_SYMBOL(__pagevec_lru_add);
 949
 950 /**
 951  * pagevec_lookup - gang pagecache lookup
 952  * @pvec:       Where the resulting pages are placed
 953  * @mapping:    The address_space to search
 954  * @start:      The starting page index
 955  * @nr_pages:   The maximum number of pages
 956  *
 957  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 958  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 959  * reference against the pages in @pvec.
 960  *
 961  * The search returns a group of mapping-contiguous pages with ascending
 962  * indexes.  There may be holes in the indices due to not-present pages.
 963  *
 964  * pagevec_lookup() returns the number of pages which were found.
 965  */
 966 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 967                 pgoff_t start, unsigned nr_pages)
 968 {
 969         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 970         return pagevec_count(pvec);
 971 }
 972 EXPORT_SYMBOL(pagevec_lookup);
 973
 974 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 975                 pgoff_t *index, int tag, unsigned nr_pages)
 976 {
 977         pvec->nr = find_get_pages_tag(mapping, index, tag,
 978                                         nr_pages, pvec->pages);
 979         return pagevec_count(pvec);
 980 }
 981 EXPORT_SYMBOL(pagevec_lookup_tag);
 982
 983 /*
 984  * Perform any setup for the swap system
 985  */
 986 void __init swap_setup(void)
 987 {
 988         unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
 989 #ifdef CONFIG_SWAP
 990         int i;
 991
 992         if (bdi_init(swapper_spaces[0].backing_dev_info))
 993                 panic("Failed to init swap bdi");
 994         for (i = 0; i < MAX_SWAPFILES; i++) {
 995                 spin_lock_init(&swapper_spaces[i].tree_lock);
 996                 INIT_LIST_HEAD(&swapper_spaces[i].i_mmap_nonlinear);
 997         }
 998 #endif
 999
1000         /* Use a smaller cluster for small-memory machines */
1001         if (megs < 16)
1002                 page_cluster = 2;
1003         else
1004                 page_cluster = 3;
1005         /*
1006          * Right now other parts of the system means that we
1007          * _really_ don't want to cluster much more
1008          */
1009 }