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 void __lru_cache_add(struct page *page, enum lru_list lru)
 450 {
 451         struct pagevec *pvec = &get_cpu_var(lru_add_pvecs)[lru];
 452
 453         page_cache_get(page);
 454         if (!pagevec_add(pvec, page))
 455                 __pagevec_lru_add(pvec, lru);
 456         put_cpu_var(lru_add_pvecs);
 457 }
 458 EXPORT_SYMBOL(__lru_cache_add);
 459
 460 /**
 461  * lru_cache_add_lru - add a page to a page list
 462  * @page: the page to be added to the LRU.
 463  * @lru: the LRU list to which the page is added.
 464  */
 465 void lru_cache_add_lru(struct page *page, enum lru_list lru)
 466 {
 467         if (PageActive(page)) {
 468                 VM_BUG_ON(PageUnevictable(page));
 469                 ClearPageActive(page);
 470         } else if (PageUnevictable(page)) {
 471                 VM_BUG_ON(PageActive(page));
 472                 ClearPageUnevictable(page);
 473         }
 474
 475         VM_BUG_ON(PageLRU(page) || PageActive(page) || PageUnevictable(page));
 476         __lru_cache_add(page, lru);
 477 }
 478
 479 /**
 480  * add_page_to_unevictable_list - add a page to the unevictable list
 481  * @page:  the page to be added to the unevictable list
 482  *
 483  * Add page directly to its zone's unevictable list.  To avoid races with
 484  * tasks that might be making the page evictable, through eg. munlock,
 485  * munmap or exit, while it's not on the lru, we want to add the page
 486  * while it's locked or otherwise "invisible" to other tasks.  This is
 487  * difficult to do when using the pagevec cache, so bypass that.
 488  */
 489 void add_page_to_unevictable_list(struct page *page)
 490 {
 491         struct zone *zone = page_zone(page);
 492         struct lruvec *lruvec;
 493
 494         spin_lock_irq(&zone->lru_lock);
 495         lruvec = mem_cgroup_page_lruvec(page, zone);
 496         SetPageUnevictable(page);
 497         SetPageLRU(page);
 498         add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
 499         spin_unlock_irq(&zone->lru_lock);
 500 }
 501
 502 /*
 503  * If the page can not be invalidated, it is moved to the
 504  * inactive list to speed up its reclaim.  It is moved to the
 505  * head of the list, rather than the tail, to give the flusher
 506  * threads some time to write it out, as this is much more
 507  * effective than the single-page writeout from reclaim.
 508  *
 509  * If the page isn't page_mapped and dirty/writeback, the page
 510  * could reclaim asap using PG_reclaim.
 511  *
 512  * 1. active, mapped page -> none
 513  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 514  * 3. inactive, mapped page -> none
 515  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 516  * 5. inactive, clean -> inactive, tail
 517  * 6. Others -> none
 518  *
 519  * In 4, why it moves inactive's head, the VM expects the page would
 520  * be write it out by flusher threads as this is much more effective
 521  * than the single-page writeout from reclaim.
 522  */
 523 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 524                               void *arg)
 525 {
 526         int lru, file;
 527         bool active;
 528
 529         if (!PageLRU(page))
 530                 return;
 531
 532         if (PageUnevictable(page))
 533                 return;
 534
 535         /* Some processes are using the page */
 536         if (page_mapped(page))
 537                 return;
 538
 539         active = PageActive(page);
 540         file = page_is_file_cache(page);
 541         lru = page_lru_base_type(page);
 542
 543         del_page_from_lru_list(page, lruvec, lru + active);
 544         ClearPageActive(page);
 545         ClearPageReferenced(page);
 546         add_page_to_lru_list(page, lruvec, lru);
 547
 548         if (PageWriteback(page) || PageDirty(page)) {
 549                 /*
 550                  * PG_reclaim could be raced with end_page_writeback
 551                  * It can make readahead confusing.  But race window
 552                  * is _really_ small and  it's non-critical problem.
 553                  */
 554                 SetPageReclaim(page);
 555         } else {
 556                 /*
 557                  * The page's writeback ends up during pagevec
 558                  * We moves tha page into tail of inactive.
 559                  */
 560                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 561                 __count_vm_event(PGROTATED);
 562         }
 563
 564         if (active)
 565                 __count_vm_event(PGDEACTIVATE);
 566         update_page_reclaim_stat(lruvec, file, 0);
 567 }
 568
 569 /*
 570  * Drain pages out of the cpu's pagevecs.
 571  * Either "cpu" is the current CPU, and preemption has already been
 572  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 573  */
 574 void lru_add_drain_cpu(int cpu)
 575 {
 576         struct pagevec *pvecs = per_cpu(lru_add_pvecs, cpu);
 577         struct pagevec *pvec;
 578         int lru;
 579
 580         for_each_lru(lru) {
 581                 pvec = &pvecs[lru - LRU_BASE];
 582                 if (pagevec_count(pvec))
 583                         __pagevec_lru_add(pvec, lru);
 584         }
 585
 586         pvec = &per_cpu(lru_rotate_pvecs, cpu);
 587         if (pagevec_count(pvec)) {
 588                 unsigned long flags;
 589
 590                 /* No harm done if a racing interrupt already did this */
 591                 local_irq_save(flags);
 592                 pagevec_move_tail(pvec);
 593                 local_irq_restore(flags);
 594         }
 595
 596         pvec = &per_cpu(lru_deactivate_pvecs, cpu);
 597         if (pagevec_count(pvec))
 598                 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 599
 600         activate_page_drain(cpu);
 601 }
 602
 603 /**
 604  * deactivate_page - forcefully deactivate a page
 605  * @page: page to deactivate
 606  *
 607  * This function hints the VM that @page is a good reclaim candidate,
 608  * for example if its invalidation fails due to the page being dirty
 609  * or under writeback.
 610  */
 611 void deactivate_page(struct page *page)
 612 {
 613         /*
 614          * In a workload with many unevictable page such as mprotect, unevictable
 615          * page deactivation for accelerating reclaim is pointless.
 616          */
 617         if (PageUnevictable(page))
 618                 return;
 619
 620         if (likely(get_page_unless_zero(page))) {
 621                 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
 622
 623                 if (!pagevec_add(pvec, page))
 624                         pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 625                 put_cpu_var(lru_deactivate_pvecs);
 626         }
 627 }
 628
 629 void lru_add_drain(void)
 630 {
 631         lru_add_drain_cpu(get_cpu());
 632         put_cpu();
 633 }
 634
 635 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 636 {
 637         lru_add_drain();
 638 }
 639
 640 /*
 641  * Returns 0 for success
 642  */
 643 int lru_add_drain_all(void)
 644 {
 645         return schedule_on_each_cpu(lru_add_drain_per_cpu);
 646 }
 647
 648 /*
 649  * Batched page_cache_release().  Decrement the reference count on all the
 650  * passed pages.  If it fell to zero then remove the page from the LRU and
 651  * free it.
 652  *
 653  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 654  * for the remainder of the operation.
 655  *
 656  * The locking in this function is against shrink_inactive_list(): we recheck
 657  * the page count inside the lock to see whether shrink_inactive_list()
 658  * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
 659  * will free it.
 660  */
 661 void release_pages(struct page **pages, int nr, int cold)
 662 {
 663         int i;
 664         LIST_HEAD(pages_to_free);
 665         struct zone *zone = NULL;
 666         struct lruvec *lruvec;
 667         unsigned long uninitialized_var(flags);
 668
 669         for (i = 0; i < nr; i++) {
 670                 struct page *page = pages[i];
 671
 672                 if (unlikely(PageCompound(page))) {
 673                         if (zone) {
 674                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 675                                 zone = NULL;
 676                         }
 677                         put_compound_page(page);
 678                         continue;
 679                 }
 680
 681                 if (!put_page_testzero(page))
 682                         continue;
 683
 684                 if (PageLRU(page)) {
 685                         struct zone *pagezone = page_zone(page);
 686
 687                         if (pagezone != zone) {
 688                                 if (zone)
 689                                         spin_unlock_irqrestore(&zone->lru_lock,
 690                                                                         flags);
 691                                 zone = pagezone;
 692                                 spin_lock_irqsave(&zone->lru_lock, flags);
 693                         }
 694
 695                         lruvec = mem_cgroup_page_lruvec(page, zone);
 696                         VM_BUG_ON(!PageLRU(page));
 697                         __ClearPageLRU(page);
 698                         del_page_from_lru_list(page, lruvec, page_off_lru(page));
 699                 }
 700
 701                 list_add(&page->lru, &pages_to_free);
 702         }
 703         if (zone)
 704                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 705
 706         free_hot_cold_page_list(&pages_to_free, cold);
 707 }
 708 EXPORT_SYMBOL(release_pages);
 709
 710 /*
 711  * The pages which we're about to release may be in the deferred lru-addition
 712  * queues.  That would prevent them from really being freed right now.  That's
 713  * OK from a correctness point of view but is inefficient - those pages may be
 714  * cache-warm and we want to give them back to the page allocator ASAP.
 715  *
 716  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 717  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 718  * mutual recursion.
 719  */
 720 void __pagevec_release(struct pagevec *pvec)
 721 {
 722         lru_add_drain();
 723         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 724         pagevec_reinit(pvec);
 725 }
 726 EXPORT_SYMBOL(__pagevec_release);
 727
 728 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 729 /* used by __split_huge_page_refcount() */
 730 void lru_add_page_tail(struct page *page, struct page *page_tail,
 731                        struct lruvec *lruvec)
 732 {
 733         int uninitialized_var(active);
 734         enum lru_list lru;
 735         const int file = 0;
 736
 737         VM_BUG_ON(!PageHead(page));
 738         VM_BUG_ON(PageCompound(page_tail));
 739         VM_BUG_ON(PageLRU(page_tail));
 740         VM_BUG_ON(NR_CPUS != 1 &&
 741                   !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
 742
 743         SetPageLRU(page_tail);
 744
 745         if (page_evictable(page_tail, NULL)) {
 746                 if (PageActive(page)) {
 747                         SetPageActive(page_tail);
 748                         active = 1;
 749                         lru = LRU_ACTIVE_ANON;
 750                 } else {
 751                         active = 0;
 752                         lru = LRU_INACTIVE_ANON;
 753                 }
 754         } else {
 755                 SetPageUnevictable(page_tail);
 756                 lru = LRU_UNEVICTABLE;
 757         }
 758
 759         if (likely(PageLRU(page)))
 760                 list_add_tail(&page_tail->lru, &page->lru);
 761         else {
 762                 struct list_head *list_head;
 763                 /*
 764                  * Head page has not yet been counted, as an hpage,
 765                  * so we must account for each subpage individually.
 766                  *
 767                  * Use the standard add function to put page_tail on the list,
 768                  * but then correct its position so they all end up in order.
 769                  */
 770                 add_page_to_lru_list(page_tail, lruvec, lru);
 771                 list_head = page_tail->lru.prev;
 772                 list_move_tail(&page_tail->lru, list_head);
 773         }
 774
 775         if (!PageUnevictable(page))
 776                 update_page_reclaim_stat(lruvec, file, active);
 777 }
 778 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 779
 780 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
 781                                  void *arg)
 782 {
 783         enum lru_list lru = (enum lru_list)arg;
 784         int file = is_file_lru(lru);
 785         int active = is_active_lru(lru);
 786
 787         VM_BUG_ON(PageActive(page));
 788         VM_BUG_ON(PageUnevictable(page));
 789         VM_BUG_ON(PageLRU(page));
 790
 791         SetPageLRU(page);
 792         if (active)
 793                 SetPageActive(page);
 794         add_page_to_lru_list(page, lruvec, lru);
 795         update_page_reclaim_stat(lruvec, file, active);
 796 }
 797
 798 /*
 799  * Add the passed pages to the LRU, then drop the caller's refcount
 800  * on them.  Reinitialises the caller's pagevec.
 801  */
 802 void __pagevec_lru_add(struct pagevec *pvec, enum lru_list lru)
 803 {
 804         VM_BUG_ON(is_unevictable_lru(lru));
 805
 806         pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, (void *)lru);
 807 }
 808 EXPORT_SYMBOL(__pagevec_lru_add);
 809
 810 /**
 811  * pagevec_lookup - gang pagecache lookup
 812  * @pvec:       Where the resulting pages are placed
 813  * @mapping:    The address_space to search
 814  * @start:      The starting page index
 815  * @nr_pages:   The maximum number of pages
 816  *
 817  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 818  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 819  * reference against the pages in @pvec.
 820  *
 821  * The search returns a group of mapping-contiguous pages with ascending
 822  * indexes.  There may be holes in the indices due to not-present pages.
 823  *
 824  * pagevec_lookup() returns the number of pages which were found.
 825  */
 826 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 827                 pgoff_t start, unsigned nr_pages)
 828 {
 829         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 830         return pagevec_count(pvec);
 831 }
 832 EXPORT_SYMBOL(pagevec_lookup);
 833
 834 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 835                 pgoff_t *index, int tag, unsigned nr_pages)
 836 {
 837         pvec->nr = find_get_pages_tag(mapping, index, tag,
 838                                         nr_pages, pvec->pages);
 839         return pagevec_count(pvec);
 840 }
 841 EXPORT_SYMBOL(pagevec_lookup_tag);
 842
 843 /*
 844  * Perform any setup for the swap system
 845  */
 846 void __init swap_setup(void)
 847 {
 848         unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
 849
 850 #ifdef CONFIG_SWAP
 851         bdi_init(swapper_space.backing_dev_info);
 852 #endif
 853
 854         /* Use a smaller cluster for small-memory machines */
 855         if (megs < 16)
 856                 page_cluster = 2;
 857         else
 858                 page_cluster = 3;
 859         /*
 860          * Right now other parts of the system means that we
 861          * _really_ don't want to cluster much more
 862          */
 863 }