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