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