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(!PageLRU(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         if (unlikely(PageTail(page))) {
  85                 /* __split_huge_page_refcount can run under us */
  86                 struct page *page_head = compound_trans_head(page);
  87
  88                 /*
  89                  * THP can not break up slab pages so avoid taking
  90                  * compound_lock() and skip the tail page refcounting
  91                  * (in _mapcount) too. Slab performs non-atomic bit
  92                  * ops on page->flags for better performance. In
  93                  * particular slab_unlock() in slub used to be a hot
  94                  * path. It is still hot on arches that do not support
  95                  * this_cpu_cmpxchg_double().
  96                  *
  97                  * If "page" is part of a slab or hugetlbfs page it
  98                  * cannot be splitted and the head page cannot change
  99                  * from under us. And if "page" is part of a THP page
 100                  * under splitting, if the head page pointed by the
 101                  * THP tail isn't a THP head anymore, we'll find
 102                  * PageTail clear after smp_rmb() and we'll treat it
 103                  * as a single page.
 104                  */
 105                 if (!__compound_tail_refcounted(page_head)) {
 106                         /*
 107                          * If "page" is a THP tail, we must read the tail page
 108                          * flags after the head page flags. The
 109                          * split_huge_page side enforces write memory
 110                          * barriers between clearing PageTail and before the
 111                          * head page can be freed and reallocated.
 112                          */
 113                         smp_rmb();
 114                         if (likely(PageTail(page))) {
 115                                 /*
 116                                  * __split_huge_page_refcount
 117                                  * cannot race here.
 118                                  */
 119                                 VM_BUG_ON(!PageHead(page_head));
 120                                 VM_BUG_ON(page_mapcount(page) != 0);
 121                                 if (put_page_testzero(page_head)) {
 122                                         /*
 123                                          * If this is the tail of a
 124                                          * slab compound page, the
 125                                          * tail pin must not be the
 126                                          * last reference held on the
 127                                          * page, because the PG_slab
 128                                          * cannot be cleared before
 129                                          * all tail pins (which skips
 130                                          * the _mapcount tail
 131                                          * refcounting) have been
 132                                          * released. For hugetlbfs the
 133                                          * tail pin may be the last
 134                                          * reference on the page
 135                                          * instead, because
 136                                          * PageHeadHuge will not go
 137                                          * away until the compound
 138                                          * page enters the buddy
 139                                          * allocator.
 140                                          */
 141                                         VM_BUG_ON(PageSlab(page_head));
 142                                         __put_compound_page(page_head);
 143                                 }
 144                                 return;
 145                         } else
 146                                 /*
 147                                  * __split_huge_page_refcount
 148                                  * run before us, "page" was a
 149                                  * THP tail. The split
 150                                  * page_head has been freed
 151                                  * and reallocated as slab or
 152                                  * hugetlbfs page of smaller
 153                                  * order (only possible if
 154                                  * reallocated as slab on
 155                                  * x86).
 156                                  */
 157                                 goto out_put_single;
 158                 }
 159
 160                 if (likely(page != page_head &&
 161                            get_page_unless_zero(page_head))) {
 162                         unsigned long flags;
 163
 164                         /*
 165                          * page_head wasn't a dangling pointer but it
 166                          * may not be a head page anymore by the time
 167                          * we obtain the lock. That is ok as long as it
 168                          * can't be freed from under us.
 169                          */
 170                         flags = compound_lock_irqsave(page_head);
 171                         if (unlikely(!PageTail(page))) {
 172                                 /* __split_huge_page_refcount run before us */
 173                                 compound_unlock_irqrestore(page_head, flags);
 174                                 if (put_page_testzero(page_head)) {
 175                                         /*
 176                                          * The head page may have been
 177                                          * freed and reallocated as a
 178                                          * compound page of smaller
 179                                          * order and then freed again.
 180                                          * All we know is that it
 181                                          * cannot have become: a THP
 182                                          * page, a compound page of
 183                                          * higher order, a tail page.
 184                                          * That is because we still
 185                                          * hold the refcount of the
 186                                          * split THP tail and
 187                                          * page_head was the THP head
 188                                          * before the split.
 189                                          */
 190                                         if (PageHead(page_head))
 191                                                 __put_compound_page(page_head);
 192                                         else
 193                                                 __put_single_page(page_head);
 194                                 }
 195 out_put_single:
 196                                 if (put_page_testzero(page))
 197                                         __put_single_page(page);
 198                                 return;
 199                         }
 200                         VM_BUG_ON(page_head != page->first_page);
 201                         /*
 202                          * We can release the refcount taken by
 203                          * get_page_unless_zero() now that
 204                          * __split_huge_page_refcount() is blocked on
 205                          * the compound_lock.
 206                          */
 207                         if (put_page_testzero(page_head))
 208                                 VM_BUG_ON(1);
 209                         /* __split_huge_page_refcount will wait now */
 210                         VM_BUG_ON(page_mapcount(page) <= 0);
 211                         atomic_dec(&page->_mapcount);
 212                         VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
 213                         VM_BUG_ON(atomic_read(&page->_count) != 0);
 214                         compound_unlock_irqrestore(page_head, flags);
 215
 216                         if (put_page_testzero(page_head)) {
 217                                 if (PageHead(page_head))
 218                                         __put_compound_page(page_head);
 219                                 else
 220                                         __put_single_page(page_head);
 221                         }
 222                 } else {
 223                         /* page_head is a dangling pointer */
 224                         VM_BUG_ON(PageTail(page));
 225                         goto out_put_single;
 226                 }
 227         } else if (put_page_testzero(page)) {
 228                 if (PageHead(page))
 229                         __put_compound_page(page);
 230                 else
 231                         __put_single_page(page);
 232         }
 233 }
 234
 235 void put_page(struct page *page)
 236 {
 237         if (unlikely(PageCompound(page)))
 238                 put_compound_page(page);
 239         else if (put_page_testzero(page))
 240                 __put_single_page(page);
 241 }
 242 EXPORT_SYMBOL(put_page);
 243
 244 /*
 245  * This function is exported but must not be called by anything other
 246  * than get_page(). It implements the slow path of get_page().
 247  */
 248 bool __get_page_tail(struct page *page)
 249 {
 250         /*
 251          * This takes care of get_page() if run on a tail page
 252          * returned by one of the get_user_pages/follow_page variants.
 253          * get_user_pages/follow_page itself doesn't need the compound
 254          * lock because it runs __get_page_tail_foll() under the
 255          * proper PT lock that already serializes against
 256          * split_huge_page().
 257          */
 258         unsigned long flags;
 259         bool got;
 260         struct page *page_head = compound_trans_head(page);
 261
 262         /* Ref to put_compound_page() comment. */
 263         if (!__compound_tail_refcounted(page_head)) {
 264                 smp_rmb();
 265                 if (likely(PageTail(page))) {
 266                         /*
 267                          * This is a hugetlbfs page or a slab
 268                          * page. __split_huge_page_refcount
 269                          * cannot race here.
 270                          */
 271                         VM_BUG_ON(!PageHead(page_head));
 272                         __get_page_tail_foll(page, true);
 273                         return true;
 274                 } else {
 275                         /*
 276                          * __split_huge_page_refcount run
 277                          * before us, "page" was a THP
 278                          * tail. The split page_head has been
 279                          * freed and reallocated as slab or
 280                          * hugetlbfs page of smaller order
 281                          * (only possible if reallocated as
 282                          * slab on x86).
 283                          */
 284                         return false;
 285                 }
 286         }
 287
 288         got = false;
 289         if (likely(page != page_head && get_page_unless_zero(page_head))) {
 290                 /*
 291                  * page_head wasn't a dangling pointer but it
 292                  * may not be a head page anymore by the time
 293                  * we obtain the lock. That is ok as long as it
 294                  * can't be freed from under us.
 295                  */
 296                 flags = compound_lock_irqsave(page_head);
 297                 /* here __split_huge_page_refcount won't run anymore */
 298                 if (likely(PageTail(page))) {
 299                         __get_page_tail_foll(page, false);
 300                         got = true;
 301                 }
 302                 compound_unlock_irqrestore(page_head, flags);
 303                 if (unlikely(!got))
 304                         put_page(page_head);
 305         }
 306         return got;
 307 }
 308 EXPORT_SYMBOL(__get_page_tail);
 309
 310 /**
 311  * put_pages_list() - release a list of pages
 312  * @pages: list of pages threaded on page->lru
 313  *
 314  * Release a list of pages which are strung together on page.lru.  Currently
 315  * used by read_cache_pages() and related error recovery code.
 316  */
 317 void put_pages_list(struct list_head *pages)
 318 {
 319         while (!list_empty(pages)) {
 320                 struct page *victim;
 321
 322                 victim = list_entry(pages->prev, struct page, lru);
 323                 list_del(&victim->lru);
 324                 page_cache_release(victim);
 325         }
 326 }
 327 EXPORT_SYMBOL(put_pages_list);
 328
 329 /*
 330  * get_kernel_pages() - pin kernel pages in memory
 331  * @kiov:       An array of struct kvec structures
 332  * @nr_segs:    number of segments to pin
 333  * @write:      pinning for read/write, currently ignored
 334  * @pages:      array that receives pointers to the pages pinned.
 335  *              Should be at least nr_segs long.
 336  *
 337  * Returns number of pages pinned. This may be fewer than the number
 338  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 339  * were pinned, returns -errno. Each page returned must be released
 340  * with a put_page() call when it is finished with.
 341  */
 342 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
 343                 struct page **pages)
 344 {
 345         int seg;
 346
 347         for (seg = 0; seg < nr_segs; seg++) {
 348                 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
 349                         return seg;
 350
 351                 pages[seg] = kmap_to_page(kiov[seg].iov_base);
 352                 page_cache_get(pages[seg]);
 353         }
 354
 355         return seg;
 356 }
 357 EXPORT_SYMBOL_GPL(get_kernel_pages);
 358
 359 /*
 360  * get_kernel_page() - pin a kernel page in memory
 361  * @start:      starting kernel address
 362  * @write:      pinning for read/write, currently ignored
 363  * @pages:      array that receives pointer to the page pinned.
 364  *              Must be at least nr_segs long.
 365  *
 366  * Returns 1 if page is pinned. If the page was not pinned, returns
 367  * -errno. The page returned must be released with a put_page() call
 368  * when it is finished with.
 369  */
 370 int get_kernel_page(unsigned long start, int write, struct page **pages)
 371 {
 372         const struct kvec kiov = {
 373                 .iov_base = (void *)start,
 374                 .iov_len = PAGE_SIZE
 375         };
 376
 377         return get_kernel_pages(&kiov, 1, write, pages);
 378 }
 379 EXPORT_SYMBOL_GPL(get_kernel_page);
 380
 381 static void pagevec_lru_move_fn(struct pagevec *pvec,
 382         void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
 383         void *arg)
 384 {
 385         int i;
 386         struct zone *zone = NULL;
 387         struct lruvec *lruvec;
 388         unsigned long flags = 0;
 389
 390         for (i = 0; i < pagevec_count(pvec); i++) {
 391                 struct page *page = pvec->pages[i];
 392                 struct zone *pagezone = page_zone(page);
 393
 394                 if (pagezone != zone) {
 395                         if (zone)
 396                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 397                         zone = pagezone;
 398                         spin_lock_irqsave(&zone->lru_lock, flags);
 399                 }
 400
 401                 lruvec = mem_cgroup_page_lruvec(page, zone);
 402                 (*move_fn)(page, lruvec, arg);
 403         }
 404         if (zone)
 405                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 406         release_pages(pvec->pages, pvec->nr, pvec->cold);
 407         pagevec_reinit(pvec);
 408 }
 409
 410 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
 411                                  void *arg)
 412 {
 413         int *pgmoved = arg;
 414
 415         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 416                 enum lru_list lru = page_lru_base_type(page);
 417                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 418                 (*pgmoved)++;
 419         }
 420 }
 421
 422 /*
 423  * pagevec_move_tail() must be called with IRQ disabled.
 424  * Otherwise this may cause nasty races.
 425  */
 426 static void pagevec_move_tail(struct pagevec *pvec)
 427 {
 428         int pgmoved = 0;
 429
 430         pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
 431         __count_vm_events(PGROTATED, pgmoved);
 432 }
 433
 434 /*
 435  * Writeback is about to end against a page which has been marked for immediate
 436  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 437  * inactive list.
 438  */
 439 void rotate_reclaimable_page(struct page *page)
 440 {
 441         if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
 442             !PageUnevictable(page) && PageLRU(page)) {
 443                 struct pagevec *pvec;
 444                 unsigned long flags;
 445
 446                 page_cache_get(page);
 447                 local_irq_save(flags);
 448                 pvec = &__get_cpu_var(lru_rotate_pvecs);
 449                 if (!pagevec_add(pvec, page))
 450                         pagevec_move_tail(pvec);
 451                 local_irq_restore(flags);
 452         }
 453 }
 454
 455 static void update_page_reclaim_stat(struct lruvec *lruvec,
 456                                      int file, int rotated)
 457 {
 458         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
 459
 460         reclaim_stat->recent_scanned[file]++;
 461         if (rotated)
 462                 reclaim_stat->recent_rotated[file]++;
 463 }
 464
 465 static void __activate_page(struct page *page, struct lruvec *lruvec,
 466                             void *arg)
 467 {
 468         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 469                 int file = page_is_file_cache(page);
 470                 int lru = page_lru_base_type(page);
 471
 472                 del_page_from_lru_list(page, lruvec, lru);
 473                 SetPageActive(page);
 474                 lru += LRU_ACTIVE;
 475                 add_page_to_lru_list(page, lruvec, lru);
 476                 trace_mm_lru_activate(page, page_to_pfn(page));
 477
 478                 __count_vm_event(PGACTIVATE);
 479                 update_page_reclaim_stat(lruvec, file, 1);
 480         }
 481 }
 482
 483 #ifdef CONFIG_SMP
 484 static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
 485
 486 static void activate_page_drain(int cpu)
 487 {
 488         struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
 489
 490         if (pagevec_count(pvec))
 491                 pagevec_lru_move_fn(pvec, __activate_page, NULL);
 492 }
 493
 494 static bool need_activate_page_drain(int cpu)
 495 {
 496         return pagevec_count(&per_cpu(activate_page_pvecs, cpu)) != 0;
 497 }
 498
 499 void activate_page(struct page *page)
 500 {
 501         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 502                 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
 503
 504                 page_cache_get(page);
 505                 if (!pagevec_add(pvec, page))
 506                         pagevec_lru_move_fn(pvec, __activate_page, NULL);
 507                 put_cpu_var(activate_page_pvecs);
 508         }
 509 }
 510
 511 #else
 512 static inline void activate_page_drain(int cpu)
 513 {
 514 }
 515
 516 static bool need_activate_page_drain(int cpu)
 517 {
 518         return false;
 519 }
 520
 521 void activate_page(struct page *page)
 522 {
 523         struct zone *zone = page_zone(page);
 524
 525         spin_lock_irq(&zone->lru_lock);
 526         __activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
 527         spin_unlock_irq(&zone->lru_lock);
 528 }
 529 #endif
 530
 531 static void __lru_cache_activate_page(struct page *page)
 532 {
 533         struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
 534         int i;
 535
 536         /*
 537          * Search backwards on the optimistic assumption that the page being
 538          * activated has just been added to this pagevec. Note that only
 539          * the local pagevec is examined as a !PageLRU page could be in the
 540          * process of being released, reclaimed, migrated or on a remote
 541          * pagevec that is currently being drained. Furthermore, marking
 542          * a remote pagevec's page PageActive potentially hits a race where
 543          * a page is marked PageActive just after it is added to the inactive
 544          * list causing accounting errors and BUG_ON checks to trigger.
 545          */
 546         for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
 547                 struct page *pagevec_page = pvec->pages[i];
 548
 549                 if (pagevec_page == page) {
 550                         SetPageActive(page);
 551                         break;
 552                 }
 553         }
 554
 555         put_cpu_var(lru_add_pvec);
 556 }
 557
 558 /*
 559  * Mark a page as having seen activity.
 560  *
 561  * inactive,unreferenced        ->      inactive,referenced
 562  * inactive,referenced          ->      active,unreferenced
 563  * active,unreferenced          ->      active,referenced
 564  */
 565 void mark_page_accessed(struct page *page)
 566 {
 567         if (!PageActive(page) && !PageUnevictable(page) &&
 568                         PageReferenced(page)) {
 569
 570                 /*
 571                  * If the page is on the LRU, queue it for activation via
 572                  * activate_page_pvecs. Otherwise, assume the page is on a
 573                  * pagevec, mark it active and it'll be moved to the active
 574                  * LRU on the next drain.
 575                  */
 576                 if (PageLRU(page))
 577                         activate_page(page);
 578                 else
 579                         __lru_cache_activate_page(page);
 580                 ClearPageReferenced(page);
 581         } else if (!PageReferenced(page)) {
 582                 SetPageReferenced(page);
 583         }
 584 }
 585 EXPORT_SYMBOL(mark_page_accessed);
 586
 587 /*
 588  * Queue the page for addition to the LRU via pagevec. The decision on whether
 589  * to add the page to the [in]active [file|anon] list is deferred until the
 590  * pagevec is drained. This gives a chance for the caller of __lru_cache_add()
 591  * have the page added to the active list using mark_page_accessed().
 592  */
 593 void __lru_cache_add(struct page *page)
 594 {
 595         struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
 596
 597         page_cache_get(page);
 598         if (!pagevec_space(pvec))
 599                 __pagevec_lru_add(pvec);
 600         pagevec_add(pvec, page);
 601         put_cpu_var(lru_add_pvec);
 602 }
 603 EXPORT_SYMBOL(__lru_cache_add);
 604
 605 /**
 606  * lru_cache_add - add a page to a page list
 607  * @page: the page to be added to the LRU.
 608  */
 609 void lru_cache_add(struct page *page)
 610 {
 611         VM_BUG_ON(PageActive(page) && PageUnevictable(page));
 612         VM_BUG_ON(PageLRU(page));
 613         __lru_cache_add(page);
 614 }
 615
 616 /**
 617  * add_page_to_unevictable_list - add a page to the unevictable list
 618  * @page:  the page to be added to the unevictable list
 619  *
 620  * Add page directly to its zone's unevictable list.  To avoid races with
 621  * tasks that might be making the page evictable, through eg. munlock,
 622  * munmap or exit, while it's not on the lru, we want to add the page
 623  * while it's locked or otherwise "invisible" to other tasks.  This is
 624  * difficult to do when using the pagevec cache, so bypass that.
 625  */
 626 void add_page_to_unevictable_list(struct page *page)
 627 {
 628         struct zone *zone = page_zone(page);
 629         struct lruvec *lruvec;
 630
 631         spin_lock_irq(&zone->lru_lock);
 632         lruvec = mem_cgroup_page_lruvec(page, zone);
 633         ClearPageActive(page);
 634         SetPageUnevictable(page);
 635         SetPageLRU(page);
 636         add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
 637         spin_unlock_irq(&zone->lru_lock);
 638 }
 639
 640 /*
 641  * If the page can not be invalidated, it is moved to the
 642  * inactive list to speed up its reclaim.  It is moved to the
 643  * head of the list, rather than the tail, to give the flusher
 644  * threads some time to write it out, as this is much more
 645  * effective than the single-page writeout from reclaim.
 646  *
 647  * If the page isn't page_mapped and dirty/writeback, the page
 648  * could reclaim asap using PG_reclaim.
 649  *
 650  * 1. active, mapped page -> none
 651  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 652  * 3. inactive, mapped page -> none
 653  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 654  * 5. inactive, clean -> inactive, tail
 655  * 6. Others -> none
 656  *
 657  * In 4, why it moves inactive's head, the VM expects the page would
 658  * be write it out by flusher threads as this is much more effective
 659  * than the single-page writeout from reclaim.
 660  */
 661 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 662                               void *arg)
 663 {
 664         int lru, file;
 665         bool active;
 666
 667         if (!PageLRU(page))
 668                 return;
 669
 670         if (PageUnevictable(page))
 671                 return;
 672
 673         /* Some processes are using the page */
 674         if (page_mapped(page))
 675                 return;
 676
 677         active = PageActive(page);
 678         file = page_is_file_cache(page);
 679         lru = page_lru_base_type(page);
 680
 681         del_page_from_lru_list(page, lruvec, lru + active);
 682         ClearPageActive(page);
 683         ClearPageReferenced(page);
 684         add_page_to_lru_list(page, lruvec, lru);
 685
 686         if (PageWriteback(page) || PageDirty(page)) {
 687                 /*
 688                  * PG_reclaim could be raced with end_page_writeback
 689                  * It can make readahead confusing.  But race window
 690                  * is _really_ small and  it's non-critical problem.
 691                  */
 692                 SetPageReclaim(page);
 693         } else {
 694                 /*
 695                  * The page's writeback ends up during pagevec
 696                  * We moves tha page into tail of inactive.
 697                  */
 698                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 699                 __count_vm_event(PGROTATED);
 700         }
 701
 702         if (active)
 703                 __count_vm_event(PGDEACTIVATE);
 704         update_page_reclaim_stat(lruvec, file, 0);
 705 }
 706
 707 /*
 708  * Drain pages out of the cpu's pagevecs.
 709  * Either "cpu" is the current CPU, and preemption has already been
 710  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 711  */
 712 void lru_add_drain_cpu(int cpu)
 713 {
 714         struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
 715
 716         if (pagevec_count(pvec))
 717                 __pagevec_lru_add(pvec);
 718
 719         pvec = &per_cpu(lru_rotate_pvecs, cpu);
 720         if (pagevec_count(pvec)) {
 721                 unsigned long flags;
 722
 723                 /* No harm done if a racing interrupt already did this */
 724                 local_irq_save(flags);
 725                 pagevec_move_tail(pvec);
 726                 local_irq_restore(flags);
 727         }
 728
 729         pvec = &per_cpu(lru_deactivate_pvecs, cpu);
 730         if (pagevec_count(pvec))
 731                 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 732
 733         activate_page_drain(cpu);
 734 }
 735
 736 /**
 737  * deactivate_page - forcefully deactivate a page
 738  * @page: page to deactivate
 739  *
 740  * This function hints the VM that @page is a good reclaim candidate,
 741  * for example if its invalidation fails due to the page being dirty
 742  * or under writeback.
 743  */
 744 void deactivate_page(struct page *page)
 745 {
 746         /*
 747          * In a workload with many unevictable page such as mprotect, unevictable
 748          * page deactivation for accelerating reclaim is pointless.
 749          */
 750         if (PageUnevictable(page))
 751                 return;
 752
 753         if (likely(get_page_unless_zero(page))) {
 754                 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
 755
 756                 if (!pagevec_add(pvec, page))
 757                         pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 758                 put_cpu_var(lru_deactivate_pvecs);
 759         }
 760 }
 761
 762 void lru_add_drain(void)
 763 {
 764         lru_add_drain_cpu(get_cpu());
 765         put_cpu();
 766 }
 767
 768 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 769 {
 770         lru_add_drain();
 771 }
 772
 773 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 774
 775 void lru_add_drain_all(void)
 776 {
 777         static DEFINE_MUTEX(lock);
 778         static struct cpumask has_work;
 779         int cpu;
 780
 781         mutex_lock(&lock);
 782         get_online_cpus();
 783         cpumask_clear(&has_work);
 784
 785         for_each_online_cpu(cpu) {
 786                 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 787
 788                 if (pagevec_count(&per_cpu(lru_add_pvec, cpu)) ||
 789                     pagevec_count(&per_cpu(lru_rotate_pvecs, cpu)) ||
 790                     pagevec_count(&per_cpu(lru_deactivate_pvecs, cpu)) ||
 791                     need_activate_page_drain(cpu)) {
 792                         INIT_WORK(work, lru_add_drain_per_cpu);
 793                         schedule_work_on(cpu, work);
 794                         cpumask_set_cpu(cpu, &has_work);
 795                 }
 796         }
 797
 798         for_each_cpu(cpu, &has_work)
 799                 flush_work(&per_cpu(lru_add_drain_work, cpu));
 800
 801         put_online_cpus();
 802         mutex_unlock(&lock);
 803 }
 804
 805 /*
 806  * Batched page_cache_release().  Decrement the reference count on all the
 807  * passed pages.  If it fell to zero then remove the page from the LRU and
 808  * free it.
 809  *
 810  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 811  * for the remainder of the operation.
 812  *
 813  * The locking in this function is against shrink_inactive_list(): we recheck
 814  * the page count inside the lock to see whether shrink_inactive_list()
 815  * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
 816  * will free it.
 817  */
 818 void release_pages(struct page **pages, int nr, int cold)
 819 {
 820         int i;
 821         LIST_HEAD(pages_to_free);
 822         struct zone *zone = NULL;
 823         struct lruvec *lruvec;
 824         unsigned long uninitialized_var(flags);
 825
 826         for (i = 0; i < nr; i++) {
 827                 struct page *page = pages[i];
 828
 829                 if (unlikely(PageCompound(page))) {
 830                         if (zone) {
 831                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 832                                 zone = NULL;
 833                         }
 834                         put_compound_page(page);
 835                         continue;
 836                 }
 837
 838                 if (!put_page_testzero(page))
 839                         continue;
 840
 841                 if (PageLRU(page)) {
 842                         struct zone *pagezone = page_zone(page);
 843
 844                         if (pagezone != zone) {
 845                                 if (zone)
 846                                         spin_unlock_irqrestore(&zone->lru_lock,
 847                                                                         flags);
 848                                 zone = pagezone;
 849                                 spin_lock_irqsave(&zone->lru_lock, flags);
 850                         }
 851
 852                         lruvec = mem_cgroup_page_lruvec(page, zone);
 853                         VM_BUG_ON(!PageLRU(page));
 854                         __ClearPageLRU(page);
 855                         del_page_from_lru_list(page, lruvec, page_off_lru(page));
 856                 }
 857
 858                 /* Clear Active bit in case of parallel mark_page_accessed */
 859                 ClearPageActive(page);
 860
 861                 list_add(&page->lru, &pages_to_free);
 862         }
 863         if (zone)
 864                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 865
 866         free_hot_cold_page_list(&pages_to_free, cold);
 867 }
 868 EXPORT_SYMBOL(release_pages);
 869
 870 /*
 871  * The pages which we're about to release may be in the deferred lru-addition
 872  * queues.  That would prevent them from really being freed right now.  That's
 873  * OK from a correctness point of view but is inefficient - those pages may be
 874  * cache-warm and we want to give them back to the page allocator ASAP.
 875  *
 876  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 877  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 878  * mutual recursion.
 879  */
 880 void __pagevec_release(struct pagevec *pvec)
 881 {
 882         lru_add_drain();
 883         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 884         pagevec_reinit(pvec);
 885 }
 886 EXPORT_SYMBOL(__pagevec_release);
 887
 888 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 889 /* used by __split_huge_page_refcount() */
 890 void lru_add_page_tail(struct page *page, struct page *page_tail,
 891                        struct lruvec *lruvec, struct list_head *list)
 892 {
 893         const int file = 0;
 894
 895         VM_BUG_ON(!PageHead(page));
 896         VM_BUG_ON(PageCompound(page_tail));
 897         VM_BUG_ON(PageLRU(page_tail));
 898         VM_BUG_ON(NR_CPUS != 1 &&
 899                   !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
 900
 901         if (!list)
 902                 SetPageLRU(page_tail);
 903
 904         if (likely(PageLRU(page)))
 905                 list_add_tail(&page_tail->lru, &page->lru);
 906         else if (list) {
 907                 /* page reclaim is reclaiming a huge page */
 908                 get_page(page_tail);
 909                 list_add_tail(&page_tail->lru, list);
 910         } else {
 911                 struct list_head *list_head;
 912                 /*
 913                  * Head page has not yet been counted, as an hpage,
 914                  * so we must account for each subpage individually.
 915                  *
 916                  * Use the standard add function to put page_tail on the list,
 917                  * but then correct its position so they all end up in order.
 918                  */
 919                 add_page_to_lru_list(page_tail, lruvec, page_lru(page_tail));
 920                 list_head = page_tail->lru.prev;
 921                 list_move_tail(&page_tail->lru, list_head);
 922         }
 923
 924         if (!PageUnevictable(page))
 925                 update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
 926 }
 927 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 928
 929 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
 930                                  void *arg)
 931 {
 932         int file = page_is_file_cache(page);
 933         int active = PageActive(page);
 934         enum lru_list lru = page_lru(page);
 935
 936         VM_BUG_ON(PageLRU(page));
 937
 938         SetPageLRU(page);
 939         add_page_to_lru_list(page, lruvec, lru);
 940         update_page_reclaim_stat(lruvec, file, active);
 941         trace_mm_lru_insertion(page, page_to_pfn(page), lru, trace_pagemap_flags(page));
 942 }
 943
 944 /*
 945  * Add the passed pages to the LRU, then drop the caller's refcount
 946  * on them.  Reinitialises the caller's pagevec.
 947  */
 948 void __pagevec_lru_add(struct pagevec *pvec)
 949 {
 950         pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
 951 }
 952 EXPORT_SYMBOL(__pagevec_lru_add);
 953
 954 /**
 955  * pagevec_lookup - gang pagecache lookup
 956  * @pvec:       Where the resulting pages are placed
 957  * @mapping:    The address_space to search
 958  * @start:      The starting page index
 959  * @nr_pages:   The maximum number of pages
 960  *
 961  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 962  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 963  * reference against the pages in @pvec.
 964  *
 965  * The search returns a group of mapping-contiguous pages with ascending
 966  * indexes.  There may be holes in the indices due to not-present pages.
 967  *
 968  * pagevec_lookup() returns the number of pages which were found.
 969  */
 970 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 971                 pgoff_t start, unsigned nr_pages)
 972 {
 973         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 974         return pagevec_count(pvec);
 975 }
 976 EXPORT_SYMBOL(pagevec_lookup);
 977
 978 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 979                 pgoff_t *index, int tag, unsigned nr_pages)
 980 {
 981         pvec->nr = find_get_pages_tag(mapping, index, tag,
 982                                         nr_pages, pvec->pages);
 983         return pagevec_count(pvec);
 984 }
 985 EXPORT_SYMBOL(pagevec_lookup_tag);
 986
 987 /*
 988  * Perform any setup for the swap system
 989  */
 990 void __init swap_setup(void)
 991 {
 992         unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
 993 #ifdef CONFIG_SWAP
 994         int i;
 995
 996         if (bdi_init(swapper_spaces[0].backing_dev_info))
 997                 panic("Failed to init swap bdi");
 998         for (i = 0; i < MAX_SWAPFILES; i++) {
 999                 spin_lock_init(&swapper_spaces[i].tree_lock);
1000                 INIT_LIST_HEAD(&swapper_spaces[i].i_mmap_nonlinear);
1001         }
1002 #endif
1003
1004         /* Use a smaller cluster for small-memory machines */
1005         if (megs < 16)
1006                 page_cluster = 2;
1007         else
1008                 page_cluster = 3;
1009         /*
1010          * Right now other parts of the system means that we
1011          * _really_ don't want to cluster much more
1012          */
1013 }