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 #define CREATE_TRACE_POINTS
  39 #include <trace/events/pagemap.h>
  40
  41 /* How many pages do we try to swap or page in/out together? */
  42 int page_cluster;
  43
  44 static DEFINE_PER_CPU(struct pagevec, lru_add_pvec);
  45 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
  46 static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
  47
  48 /*
  49  * This path almost never happens for VM activity - pages are normally
  50  * freed via pagevecs.  But it gets used by networking.
  51  */
  52 static void __page_cache_release(struct page *page)
  53 {
  54         if (PageLRU(page)) {
  55                 struct zone *zone = page_zone(page);
  56                 struct lruvec *lruvec;
  57                 unsigned long flags;
  58
  59                 spin_lock_irqsave(&zone->lru_lock, flags);
  60                 lruvec = mem_cgroup_page_lruvec(page, zone);
  61                 VM_BUG_ON(!PageLRU(page));
  62                 __ClearPageLRU(page);
  63                 del_page_from_lru_list(page, lruvec, page_off_lru(page));
  64                 spin_unlock_irqrestore(&zone->lru_lock, flags);
  65         }
  66 }
  67
  68 static void __put_single_page(struct page *page)
  69 {
  70         __page_cache_release(page);
  71         free_hot_cold_page(page, 0);
  72 }
  73
  74 static void __put_compound_page(struct page *page)
  75 {
  76         compound_page_dtor *dtor;
  77
  78         __page_cache_release(page);
  79         dtor = get_compound_page_dtor(page);
  80         (*dtor)(page);
  81 }
  82
  83 static void put_compound_page(struct page *page)
  84 {
  85         if (unlikely(PageTail(page))) {
  86                 /* __split_huge_page_refcount can run under us */
  87                 struct page *page_head = compound_trans_head(page);
  88
  89                 /*
  90                  * THP can not break up slab pages so avoid taking
  91                  * compound_lock() and skip the tail page refcounting
  92                  * (in _mapcount) too. Slab performs non-atomic bit
  93                  * ops on page->flags for better performance. In
  94                  * particular slab_unlock() in slub used to be a hot
  95                  * path. It is still hot on arches that do not support
  96                  * this_cpu_cmpxchg_double().
  97                  *
  98                  * If "page" is part of a slab or hugetlbfs page it
  99                  * cannot be splitted and the head page cannot change
 100                  * from under us. And if "page" is part of a THP page
 101                  * under splitting, if the head page pointed by the
 102                  * THP tail isn't a THP head anymore, we'll find
 103                  * PageTail clear after smp_rmb() and we'll treat it
 104                  * as a single page.
 105                  */
 106                 if (!__compound_tail_refcounted(page_head)) {
 107                         /*
 108                          * If "page" is a THP tail, we must read the tail page
 109                          * flags after the head page flags. The
 110                          * split_huge_page side enforces write memory
 111                          * barriers between clearing PageTail and before the
 112                          * head page can be freed and reallocated.
 113                          */
 114                         smp_rmb();
 115                         if (likely(PageTail(page))) {
 116                                 /*
 117                                  * __split_huge_page_refcount
 118                                  * cannot race here.
 119                                  */
 120                                 VM_BUG_ON(!PageHead(page_head));
 121                                 VM_BUG_ON(page_mapcount(page) != 0);
 122                                 if (put_page_testzero(page_head)) {
 123                                         /*
 124                                          * If this is the tail of a
 125                                          * slab compound page, the
 126                                          * tail pin must not be the
 127                                          * last reference held on the
 128                                          * page, because the PG_slab
 129                                          * cannot be cleared before
 130                                          * all tail pins (which skips
 131                                          * the _mapcount tail
 132                                          * refcounting) have been
 133                                          * released. For hugetlbfs the
 134                                          * tail pin may be the last
 135                                          * reference on the page
 136                                          * instead, because
 137                                          * PageHeadHuge will not go
 138                                          * away until the compound
 139                                          * page enters the buddy
 140                                          * allocator.
 141                                          */
 142                                         VM_BUG_ON(PageSlab(page_head));
 143                                         __put_compound_page(page_head);
 144                                 }
 145                                 return;
 146                         } else
 147                                 /*
 148                                  * __split_huge_page_refcount
 149                                  * run before us, "page" was a
 150                                  * THP tail. The split
 151                                  * page_head has been freed
 152                                  * and reallocated as slab or
 153                                  * hugetlbfs page of smaller
 154                                  * order (only possible if
 155                                  * reallocated as slab on
 156                                  * x86).
 157                                  */
 158                                 goto out_put_single;
 159                 }
 160
 161                 if (likely(page != page_head &&
 162                            get_page_unless_zero(page_head))) {
 163                         unsigned long flags;
 164
 165                         /*
 166                          * page_head wasn't a dangling pointer but it
 167                          * may not be a head page anymore by the time
 168                          * we obtain the lock. That is ok as long as it
 169                          * can't be freed from under us.
 170                          */
 171                         flags = compound_lock_irqsave(page_head);
 172                         if (unlikely(!PageTail(page))) {
 173                                 /* __split_huge_page_refcount run before us */
 174                                 compound_unlock_irqrestore(page_head, flags);
 175                                 if (put_page_testzero(page_head)) {
 176                                         /*
 177                                          * The head page may have been
 178                                          * freed and reallocated as a
 179                                          * compound page of smaller
 180                                          * order and then freed again.
 181                                          * All we know is that it
 182                                          * cannot have become: a THP
 183                                          * page, a compound page of
 184                                          * higher order, a tail page.
 185                                          * That is because we still
 186                                          * hold the refcount of the
 187                                          * split THP tail and
 188                                          * page_head was the THP head
 189                                          * before the split.
 190                                          */
 191                                         if (PageHead(page_head))
 192                                                 __put_compound_page(page_head);
 193                                         else
 194                                                 __put_single_page(page_head);
 195                                 }
 196 out_put_single:
 197                                 if (put_page_testzero(page))
 198                                         __put_single_page(page);
 199                                 return;
 200                         }
 201                         VM_BUG_ON(page_head != page->first_page);
 202                         /*
 203                          * We can release the refcount taken by
 204                          * get_page_unless_zero() now that
 205                          * __split_huge_page_refcount() is blocked on
 206                          * the compound_lock.
 207                          */
 208                         if (put_page_testzero(page_head))
 209                                 VM_BUG_ON(1);
 210                         /* __split_huge_page_refcount will wait now */
 211                         VM_BUG_ON(page_mapcount(page) <= 0);
 212                         atomic_dec(&page->_mapcount);
 213                         VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
 214                         VM_BUG_ON(atomic_read(&page->_count) != 0);
 215                         compound_unlock_irqrestore(page_head, flags);
 216
 217                         if (put_page_testzero(page_head)) {
 218                                 if (PageHead(page_head))
 219                                         __put_compound_page(page_head);
 220                                 else
 221                                         __put_single_page(page_head);
 222                         }
 223                 } else {
 224                         /* page_head is a dangling pointer */
 225                         VM_BUG_ON(PageTail(page));
 226                         goto out_put_single;
 227                 }
 228         } else if (put_page_testzero(page)) {
 229                 if (PageHead(page))
 230                         __put_compound_page(page);
 231                 else
 232                         __put_single_page(page);
 233         }
 234 }
 235
 236 void put_page(struct page *page)
 237 {
 238         if (unlikely(PageCompound(page)))
 239                 put_compound_page(page);
 240         else if (put_page_testzero(page))
 241                 __put_single_page(page);
 242 }
 243 EXPORT_SYMBOL(put_page);
 244
 245 /*
 246  * This function is exported but must not be called by anything other
 247  * than get_page(). It implements the slow path of get_page().
 248  */
 249 bool __get_page_tail(struct page *page)
 250 {
 251         /*
 252          * This takes care of get_page() if run on a tail page
 253          * returned by one of the get_user_pages/follow_page variants.
 254          * get_user_pages/follow_page itself doesn't need the compound
 255          * lock because it runs __get_page_tail_foll() under the
 256          * proper PT lock that already serializes against
 257          * split_huge_page().
 258          */
 259         unsigned long flags;
 260         bool got;
 261         struct page *page_head = compound_trans_head(page);
 262
 263         /* Ref to put_compound_page() comment. */
 264         if (PageSlab(page_head) || PageHeadHuge(page_head)) {
 265                 smp_rmb();
 266                 if (likely(PageTail(page))) {
 267                         /*
 268                          * This is a hugetlbfs page or a slab
 269                          * page. __split_huge_page_refcount
 270                          * cannot race here.
 271                          */
 272                         VM_BUG_ON(!PageHead(page_head));
 273                         __get_page_tail_foll(page, true);
 274                         return true;
 275                 } else {
 276                         /*
 277                          * __split_huge_page_refcount run
 278                          * before us, "page" was a THP
 279                          * tail. The split page_head has been
 280                          * freed and reallocated as slab or
 281                          * hugetlbfs page of smaller order
 282                          * (only possible if reallocated as
 283                          * slab on x86).
 284                          */
 285                         return false;
 286                 }
 287         }
 288
 289         got = false;
 290         if (likely(page != page_head && get_page_unless_zero(page_head))) {
 291                 /*
 292                  * page_head wasn't a dangling pointer but it
 293                  * may not be a head page anymore by the time
 294                  * we obtain the lock. That is ok as long as it
 295                  * can't be freed from under us.
 296                  */
 297                 flags = compound_lock_irqsave(page_head);
 298                 /* here __split_huge_page_refcount won't run anymore */
 299                 if (likely(PageTail(page))) {
 300                         __get_page_tail_foll(page, false);
 301                         got = true;
 302                 }
 303                 compound_unlock_irqrestore(page_head, flags);
 304                 if (unlikely(!got))
 305                         put_page(page_head);
 306         }
 307         return got;
 308 }
 309 EXPORT_SYMBOL(__get_page_tail);
 310
 311 /**
 312  * put_pages_list() - release a list of pages
 313  * @pages: list of pages threaded on page->lru
 314  *
 315  * Release a list of pages which are strung together on page.lru.  Currently
 316  * used by read_cache_pages() and related error recovery code.
 317  */
 318 void put_pages_list(struct list_head *pages)
 319 {
 320         while (!list_empty(pages)) {
 321                 struct page *victim;
 322
 323                 victim = list_entry(pages->prev, struct page, lru);
 324                 list_del(&victim->lru);
 325                 page_cache_release(victim);
 326         }
 327 }
 328 EXPORT_SYMBOL(put_pages_list);
 329
 330 /*
 331  * get_kernel_pages() - pin kernel pages in memory
 332  * @kiov:       An array of struct kvec structures
 333  * @nr_segs:    number of segments to pin
 334  * @write:      pinning for read/write, currently ignored
 335  * @pages:      array that receives pointers to the pages pinned.
 336  *              Should be at least nr_segs long.
 337  *
 338  * Returns number of pages pinned. This may be fewer than the number
 339  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 340  * were pinned, returns -errno. Each page returned must be released
 341  * with a put_page() call when it is finished with.
 342  */
 343 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
 344                 struct page **pages)
 345 {
 346         int seg;
 347
 348         for (seg = 0; seg < nr_segs; seg++) {
 349                 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
 350                         return seg;
 351
 352                 pages[seg] = kmap_to_page(kiov[seg].iov_base);
 353                 page_cache_get(pages[seg]);
 354         }
 355
 356         return seg;
 357 }
 358 EXPORT_SYMBOL_GPL(get_kernel_pages);
 359
 360 /*
 361  * get_kernel_page() - pin a kernel page in memory
 362  * @start:      starting kernel address
 363  * @write:      pinning for read/write, currently ignored
 364  * @pages:      array that receives pointer to the page pinned.
 365  *              Must be at least nr_segs long.
 366  *
 367  * Returns 1 if page is pinned. If the page was not pinned, returns
 368  * -errno. The page returned must be released with a put_page() call
 369  * when it is finished with.
 370  */
 371 int get_kernel_page(unsigned long start, int write, struct page **pages)
 372 {
 373         const struct kvec kiov = {
 374                 .iov_base = (void *)start,
 375                 .iov_len = PAGE_SIZE
 376         };
 377
 378         return get_kernel_pages(&kiov, 1, write, pages);
 379 }
 380 EXPORT_SYMBOL_GPL(get_kernel_page);
 381
 382 static void pagevec_lru_move_fn(struct pagevec *pvec,
 383         void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
 384         void *arg)
 385 {
 386         int i;
 387         struct zone *zone = NULL;
 388         struct lruvec *lruvec;
 389         unsigned long flags = 0;
 390
 391         for (i = 0; i < pagevec_count(pvec); i++) {
 392                 struct page *page = pvec->pages[i];
 393                 struct zone *pagezone = page_zone(page);
 394
 395                 if (pagezone != zone) {
 396                         if (zone)
 397                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 398                         zone = pagezone;
 399                         spin_lock_irqsave(&zone->lru_lock, flags);
 400                 }
 401
 402                 lruvec = mem_cgroup_page_lruvec(page, zone);
 403                 (*move_fn)(page, lruvec, arg);
 404         }
 405         if (zone)
 406                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 407         release_pages(pvec->pages, pvec->nr, pvec->cold);
 408         pagevec_reinit(pvec);
 409 }
 410
 411 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
 412                                  void *arg)
 413 {
 414         int *pgmoved = arg;
 415
 416         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 417                 enum lru_list lru = page_lru_base_type(page);
 418                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 419                 (*pgmoved)++;
 420         }
 421 }
 422
 423 /*
 424  * pagevec_move_tail() must be called with IRQ disabled.
 425  * Otherwise this may cause nasty races.
 426  */
 427 static void pagevec_move_tail(struct pagevec *pvec)
 428 {
 429         int pgmoved = 0;
 430
 431         pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
 432         __count_vm_events(PGROTATED, pgmoved);
 433 }
 434
 435 /*
 436  * Writeback is about to end against a page which has been marked for immediate
 437  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 438  * inactive list.
 439  */
 440 void rotate_reclaimable_page(struct page *page)
 441 {
 442         if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
 443             !PageUnevictable(page) && PageLRU(page)) {
 444                 struct pagevec *pvec;
 445                 unsigned long flags;
 446
 447                 page_cache_get(page);
 448                 local_irq_save(flags);
 449                 pvec = &__get_cpu_var(lru_rotate_pvecs);
 450                 if (!pagevec_add(pvec, page))
 451                         pagevec_move_tail(pvec);
 452                 local_irq_restore(flags);
 453         }
 454 }
 455
 456 static void update_page_reclaim_stat(struct lruvec *lruvec,
 457                                      int file, int rotated)
 458 {
 459         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
 460
 461         reclaim_stat->recent_scanned[file]++;
 462         if (rotated)
 463                 reclaim_stat->recent_rotated[file]++;
 464 }
 465
 466 static void __activate_page(struct page *page, struct lruvec *lruvec,
 467                             void *arg)
 468 {
 469         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 470                 int file = page_is_file_cache(page);
 471                 int lru = page_lru_base_type(page);
 472
 473                 del_page_from_lru_list(page, lruvec, lru);
 474                 SetPageActive(page);
 475                 lru += LRU_ACTIVE;
 476                 add_page_to_lru_list(page, lruvec, lru);
 477                 trace_mm_lru_activate(page, page_to_pfn(page));
 478
 479                 __count_vm_event(PGACTIVATE);
 480                 update_page_reclaim_stat(lruvec, file, 1);
 481         }
 482 }
 483
 484 #ifdef CONFIG_SMP
 485 static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
 486
 487 static void activate_page_drain(int cpu)
 488 {
 489         struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
 490
 491         if (pagevec_count(pvec))
 492                 pagevec_lru_move_fn(pvec, __activate_page, NULL);
 493 }
 494
 495 static bool need_activate_page_drain(int cpu)
 496 {
 497         return pagevec_count(&per_cpu(activate_page_pvecs, cpu)) != 0;
 498 }
 499
 500 void activate_page(struct page *page)
 501 {
 502         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 503                 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
 504
 505                 page_cache_get(page);
 506                 if (!pagevec_add(pvec, page))
 507                         pagevec_lru_move_fn(pvec, __activate_page, NULL);
 508                 put_cpu_var(activate_page_pvecs);
 509         }
 510 }
 511
 512 #else
 513 static inline void activate_page_drain(int cpu)
 514 {
 515 }
 516
 517 static bool need_activate_page_drain(int cpu)
 518 {
 519         return false;
 520 }
 521
 522 void activate_page(struct page *page)
 523 {
 524         struct zone *zone = page_zone(page);
 525
 526         spin_lock_irq(&zone->lru_lock);
 527         __activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
 528         spin_unlock_irq(&zone->lru_lock);
 529 }
 530 #endif
 531
 532 static void __lru_cache_activate_page(struct page *page)
 533 {
 534         struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
 535         int i;
 536
 537         /*
 538          * Search backwards on the optimistic assumption that the page being
 539          * activated has just been added to this pagevec. Note that only
 540          * the local pagevec is examined as a !PageLRU page could be in the
 541          * process of being released, reclaimed, migrated or on a remote
 542          * pagevec that is currently being drained. Furthermore, marking
 543          * a remote pagevec's page PageActive potentially hits a race where
 544          * a page is marked PageActive just after it is added to the inactive
 545          * list causing accounting errors and BUG_ON checks to trigger.
 546          */
 547         for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
 548                 struct page *pagevec_page = pvec->pages[i];
 549
 550                 if (pagevec_page == page) {
 551                         SetPageActive(page);
 552                         break;
 553                 }
 554         }
 555
 556         put_cpu_var(lru_add_pvec);
 557 }
 558
 559 /*
 560  * Mark a page as having seen activity.
 561  *
 562  * inactive,unreferenced        ->      inactive,referenced
 563  * inactive,referenced          ->      active,unreferenced
 564  * active,unreferenced          ->      active,referenced
 565  */
 566 void mark_page_accessed(struct page *page)
 567 {
 568         if (!PageActive(page) && !PageUnevictable(page) &&
 569                         PageReferenced(page)) {
 570
 571                 /*
 572                  * If the page is on the LRU, queue it for activation via
 573                  * activate_page_pvecs. Otherwise, assume the page is on a
 574                  * pagevec, mark it active and it'll be moved to the active
 575                  * LRU on the next drain.
 576                  */
 577                 if (PageLRU(page))
 578                         activate_page(page);
 579                 else
 580                         __lru_cache_activate_page(page);
 581                 ClearPageReferenced(page);
 582         } else if (!PageReferenced(page)) {
 583                 SetPageReferenced(page);
 584         }
 585 }
 586 EXPORT_SYMBOL(mark_page_accessed);
 587
 588 /*
 589  * Queue the page for addition to the LRU via pagevec. The decision on whether
 590  * to add the page to the [in]active [file|anon] list is deferred until the
 591  * pagevec is drained. This gives a chance for the caller of __lru_cache_add()
 592  * have the page added to the active list using mark_page_accessed().
 593  */
 594 void __lru_cache_add(struct page *page)
 595 {
 596         struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
 597
 598         page_cache_get(page);
 599         if (!pagevec_space(pvec))
 600                 __pagevec_lru_add(pvec);
 601         pagevec_add(pvec, page);
 602         put_cpu_var(lru_add_pvec);
 603 }
 604 EXPORT_SYMBOL(__lru_cache_add);
 605
 606 /**
 607  * lru_cache_add - add a page to a page list
 608  * @page: the page to be added to the LRU.
 609  */
 610 void lru_cache_add(struct page *page)
 611 {
 612         VM_BUG_ON(PageActive(page) && PageUnevictable(page));
 613         VM_BUG_ON(PageLRU(page));
 614         __lru_cache_add(page);
 615 }
 616
 617 /**
 618  * add_page_to_unevictable_list - add a page to the unevictable list
 619  * @page:  the page to be added to the unevictable list
 620  *
 621  * Add page directly to its zone's unevictable list.  To avoid races with
 622  * tasks that might be making the page evictable, through eg. munlock,
 623  * munmap or exit, while it's not on the lru, we want to add the page
 624  * while it's locked or otherwise "invisible" to other tasks.  This is
 625  * difficult to do when using the pagevec cache, so bypass that.
 626  */
 627 void add_page_to_unevictable_list(struct page *page)
 628 {
 629         struct zone *zone = page_zone(page);
 630         struct lruvec *lruvec;
 631
 632         spin_lock_irq(&zone->lru_lock);
 633         lruvec = mem_cgroup_page_lruvec(page, zone);
 634         ClearPageActive(page);
 635         SetPageUnevictable(page);
 636         SetPageLRU(page);
 637         add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
 638         spin_unlock_irq(&zone->lru_lock);
 639 }
 640
 641 /*
 642  * If the page can not be invalidated, it is moved to the
 643  * inactive list to speed up its reclaim.  It is moved to the
 644  * head of the list, rather than the tail, to give the flusher
 645  * threads some time to write it out, as this is much more
 646  * effective than the single-page writeout from reclaim.
 647  *
 648  * If the page isn't page_mapped and dirty/writeback, the page
 649  * could reclaim asap using PG_reclaim.
 650  *
 651  * 1. active, mapped page -> none
 652  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 653  * 3. inactive, mapped page -> none
 654  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 655  * 5. inactive, clean -> inactive, tail
 656  * 6. Others -> none
 657  *
 658  * In 4, why it moves inactive's head, the VM expects the page would
 659  * be write it out by flusher threads as this is much more effective
 660  * than the single-page writeout from reclaim.
 661  */
 662 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 663                               void *arg)
 664 {
 665         int lru, file;
 666         bool active;
 667
 668         if (!PageLRU(page))
 669                 return;
 670
 671         if (PageUnevictable(page))
 672                 return;
 673
 674         /* Some processes are using the page */
 675         if (page_mapped(page))
 676                 return;
 677
 678         active = PageActive(page);
 679         file = page_is_file_cache(page);
 680         lru = page_lru_base_type(page);
 681
 682         del_page_from_lru_list(page, lruvec, lru + active);
 683         ClearPageActive(page);
 684         ClearPageReferenced(page);
 685         add_page_to_lru_list(page, lruvec, lru);
 686
 687         if (PageWriteback(page) || PageDirty(page)) {
 688                 /*
 689                  * PG_reclaim could be raced with end_page_writeback
 690                  * It can make readahead confusing.  But race window
 691                  * is _really_ small and  it's non-critical problem.
 692                  */
 693                 SetPageReclaim(page);
 694         } else {
 695                 /*
 696                  * The page's writeback ends up during pagevec
 697                  * We moves tha page into tail of inactive.
 698                  */
 699                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 700                 __count_vm_event(PGROTATED);
 701         }
 702
 703         if (active)
 704                 __count_vm_event(PGDEACTIVATE);
 705         update_page_reclaim_stat(lruvec, file, 0);
 706 }
 707
 708 /*
 709  * Drain pages out of the cpu's pagevecs.
 710  * Either "cpu" is the current CPU, and preemption has already been
 711  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 712  */
 713 void lru_add_drain_cpu(int cpu)
 714 {
 715         struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
 716
 717         if (pagevec_count(pvec))
 718                 __pagevec_lru_add(pvec);
 719
 720         pvec = &per_cpu(lru_rotate_pvecs, cpu);
 721         if (pagevec_count(pvec)) {
 722                 unsigned long flags;
 723
 724                 /* No harm done if a racing interrupt already did this */
 725                 local_irq_save(flags);
 726                 pagevec_move_tail(pvec);
 727                 local_irq_restore(flags);
 728         }
 729
 730         pvec = &per_cpu(lru_deactivate_pvecs, cpu);
 731         if (pagevec_count(pvec))
 732                 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 733
 734         activate_page_drain(cpu);
 735 }
 736
 737 /**
 738  * deactivate_page - forcefully deactivate a page
 739  * @page: page to deactivate
 740  *
 741  * This function hints the VM that @page is a good reclaim candidate,
 742  * for example if its invalidation fails due to the page being dirty
 743  * or under writeback.
 744  */
 745 void deactivate_page(struct page *page)
 746 {
 747         /*
 748          * In a workload with many unevictable page such as mprotect, unevictable
 749          * page deactivation for accelerating reclaim is pointless.
 750          */
 751         if (PageUnevictable(page))
 752                 return;
 753
 754         if (likely(get_page_unless_zero(page))) {
 755                 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
 756
 757                 if (!pagevec_add(pvec, page))
 758                         pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 759                 put_cpu_var(lru_deactivate_pvecs);
 760         }
 761 }
 762
 763 void lru_add_drain(void)
 764 {
 765         lru_add_drain_cpu(get_cpu());
 766         put_cpu();
 767 }
 768
 769 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 770 {
 771         lru_add_drain();
 772 }
 773
 774 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 775
 776 void lru_add_drain_all(void)
 777 {
 778         static DEFINE_MUTEX(lock);
 779         static struct cpumask has_work;
 780         int cpu;
 781
 782         mutex_lock(&lock);
 783         get_online_cpus();
 784         cpumask_clear(&has_work);
 785
 786         for_each_online_cpu(cpu) {
 787                 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 788
 789                 if (pagevec_count(&per_cpu(lru_add_pvec, cpu)) ||
 790                     pagevec_count(&per_cpu(lru_rotate_pvecs, cpu)) ||
 791                     pagevec_count(&per_cpu(lru_deactivate_pvecs, cpu)) ||
 792                     need_activate_page_drain(cpu)) {
 793                         INIT_WORK(work, lru_add_drain_per_cpu);
 794                         schedule_work_on(cpu, work);
 795                         cpumask_set_cpu(cpu, &has_work);
 796                 }
 797         }
 798
 799         for_each_cpu(cpu, &has_work)
 800                 flush_work(&per_cpu(lru_add_drain_work, cpu));
 801
 802         put_online_cpus();
 803         mutex_unlock(&lock);
 804 }
 805
 806 /*
 807  * Batched page_cache_release().  Decrement the reference count on all the
 808  * passed pages.  If it fell to zero then remove the page from the LRU and
 809  * free it.
 810  *
 811  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 812  * for the remainder of the operation.
 813  *
 814  * The locking in this function is against shrink_inactive_list(): we recheck
 815  * the page count inside the lock to see whether shrink_inactive_list()
 816  * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
 817  * will free it.
 818  */
 819 void release_pages(struct page **pages, int nr, int cold)
 820 {
 821         int i;
 822         LIST_HEAD(pages_to_free);
 823         struct zone *zone = NULL;
 824         struct lruvec *lruvec;
 825         unsigned long uninitialized_var(flags);
 826
 827         for (i = 0; i < nr; i++) {
 828                 struct page *page = pages[i];
 829
 830                 if (unlikely(PageCompound(page))) {
 831                         if (zone) {
 832                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 833                                 zone = NULL;
 834                         }
 835                         put_compound_page(page);
 836                         continue;
 837                 }
 838
 839                 if (!put_page_testzero(page))
 840                         continue;
 841
 842                 if (PageLRU(page)) {
 843                         struct zone *pagezone = page_zone(page);
 844
 845                         if (pagezone != zone) {
 846                                 if (zone)
 847                                         spin_unlock_irqrestore(&zone->lru_lock,
 848                                                                         flags);
 849                                 zone = pagezone;
 850                                 spin_lock_irqsave(&zone->lru_lock, flags);
 851                         }
 852
 853                         lruvec = mem_cgroup_page_lruvec(page, zone);
 854                         VM_BUG_ON(!PageLRU(page));
 855                         __ClearPageLRU(page);
 856                         del_page_from_lru_list(page, lruvec, page_off_lru(page));
 857                 }
 858
 859                 /* Clear Active bit in case of parallel mark_page_accessed */
 860                 ClearPageActive(page);
 861
 862                 list_add(&page->lru, &pages_to_free);
 863         }
 864         if (zone)
 865                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 866
 867         free_hot_cold_page_list(&pages_to_free, cold);
 868 }
 869 EXPORT_SYMBOL(release_pages);
 870
 871 /*
 872  * The pages which we're about to release may be in the deferred lru-addition
 873  * queues.  That would prevent them from really being freed right now.  That's
 874  * OK from a correctness point of view but is inefficient - those pages may be
 875  * cache-warm and we want to give them back to the page allocator ASAP.
 876  *
 877  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 878  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 879  * mutual recursion.
 880  */
 881 void __pagevec_release(struct pagevec *pvec)
 882 {
 883         lru_add_drain();
 884         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 885         pagevec_reinit(pvec);
 886 }
 887 EXPORT_SYMBOL(__pagevec_release);
 888
 889 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 890 /* used by __split_huge_page_refcount() */
 891 void lru_add_page_tail(struct page *page, struct page *page_tail,
 892                        struct lruvec *lruvec, struct list_head *list)
 893 {
 894         const int file = 0;
 895
 896         VM_BUG_ON(!PageHead(page));
 897         VM_BUG_ON(PageCompound(page_tail));
 898         VM_BUG_ON(PageLRU(page_tail));
 899         VM_BUG_ON(NR_CPUS != 1 &&
 900                   !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
 901
 902         if (!list)
 903                 SetPageLRU(page_tail);
 904
 905         if (likely(PageLRU(page)))
 906                 list_add_tail(&page_tail->lru, &page->lru);
 907         else if (list) {
 908                 /* page reclaim is reclaiming a huge page */
 909                 get_page(page_tail);
 910                 list_add_tail(&page_tail->lru, list);
 911         } else {
 912                 struct list_head *list_head;
 913                 /*
 914                  * Head page has not yet been counted, as an hpage,
 915                  * so we must account for each subpage individually.
 916                  *
 917                  * Use the standard add function to put page_tail on the list,
 918                  * but then correct its position so they all end up in order.
 919                  */
 920                 add_page_to_lru_list(page_tail, lruvec, page_lru(page_tail));
 921                 list_head = page_tail->lru.prev;
 922                 list_move_tail(&page_tail->lru, list_head);
 923         }
 924
 925         if (!PageUnevictable(page))
 926                 update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
 927 }
 928 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 929
 930 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
 931                                  void *arg)
 932 {
 933         int file = page_is_file_cache(page);
 934         int active = PageActive(page);
 935         enum lru_list lru = page_lru(page);
 936
 937         VM_BUG_ON(PageLRU(page));
 938
 939         SetPageLRU(page);
 940         add_page_to_lru_list(page, lruvec, lru);
 941         update_page_reclaim_stat(lruvec, file, active);
 942         trace_mm_lru_insertion(page, page_to_pfn(page), lru, trace_pagemap_flags(page));
 943 }
 944
 945 /*
 946  * Add the passed pages to the LRU, then drop the caller's refcount
 947  * on them.  Reinitialises the caller's pagevec.
 948  */
 949 void __pagevec_lru_add(struct pagevec *pvec)
 950 {
 951         pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
 952 }
 953 EXPORT_SYMBOL(__pagevec_lru_add);
 954
 955 /**
 956  * pagevec_lookup - gang pagecache lookup
 957  * @pvec:       Where the resulting pages are placed
 958  * @mapping:    The address_space to search
 959  * @start:      The starting page index
 960  * @nr_pages:   The maximum number of pages
 961  *
 962  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 963  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 964  * reference against the pages in @pvec.
 965  *
 966  * The search returns a group of mapping-contiguous pages with ascending
 967  * indexes.  There may be holes in the indices due to not-present pages.
 968  *
 969  * pagevec_lookup() returns the number of pages which were found.
 970  */
 971 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 972                 pgoff_t start, unsigned nr_pages)
 973 {
 974         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 975         return pagevec_count(pvec);
 976 }
 977 EXPORT_SYMBOL(pagevec_lookup);
 978
 979 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 980                 pgoff_t *index, int tag, unsigned nr_pages)
 981 {
 982         pvec->nr = find_get_pages_tag(mapping, index, tag,
 983                                         nr_pages, pvec->pages);
 984         return pagevec_count(pvec);
 985 }
 986 EXPORT_SYMBOL(pagevec_lookup_tag);
 987
 988 /*
 989  * Perform any setup for the swap system
 990  */
 991 void __init swap_setup(void)
 992 {
 993         unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
 994 #ifdef CONFIG_SWAP
 995         int i;
 996
 997         if (bdi_init(swapper_spaces[0].backing_dev_info))
 998                 panic("Failed to init swap bdi");
 999         for (i = 0; i < MAX_SWAPFILES; i++) {
1000                 spin_lock_init(&swapper_spaces[i].tree_lock);
1001                 INIT_LIST_HEAD(&swapper_spaces[i].i_mmap_nonlinear);
1002         }
1003 #endif
1004
1005         /* Use a smaller cluster for small-memory machines */
1006         if (megs < 16)
1007                 page_cluster = 2;
1008         else
1009                 page_cluster = 3;
1010         /*
1011          * Right now other parts of the system means that we
1012          * _really_ don't want to cluster much more
1013          */
1014 }