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/buffer_head.h>  /* for try_to_release_page() */
  27 #include <linux/percpu_counter.h>
  28 #include <linux/percpu.h>
  29 #include <linux/cpu.h>
  30 #include <linux/notifier.h>
  31 #include <linux/backing-dev.h>
  32 #include <linux/memcontrol.h>
  33 #include <linux/gfp.h>
  34 #include <linux/hugetlb.h>
  35
  36 #include "internal.h"
  37
  38 /* How many pages do we try to swap or page in/out together? */
  39 int page_cluster;
  40
  41 static DEFINE_PER_CPU(struct pagevec[NR_LRU_LISTS], lru_add_pvecs);
  42 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
  43 static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
  44
  45 /*
  46  * This path almost never happens for VM activity - pages are normally
  47  * freed via pagevecs.  But it gets used by networking.
  48  */
  49 static void __page_cache_release(struct page *page)
  50 {
  51         if (PageLRU(page)) {
  52                 unsigned long flags;
  53                 struct zone *zone = page_zone(page);
  54
  55                 spin_lock_irqsave(&zone->lru_lock, flags);
  56                 VM_BUG_ON(!PageLRU(page));
  57                 __ClearPageLRU(page);
  58                 del_page_from_lru(zone, 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         if (!PageHuge(page))
  74                 __page_cache_release(page);
  75         dtor = get_compound_page_dtor(page);
  76         (*dtor)(page);
  77 }
  78
  79 static void put_compound_page(struct page *page)
  80 {
  81         if (unlikely(PageTail(page))) {
  82                 /* __split_huge_page_refcount can run under us */
  83                 struct page *page_head = compound_trans_head(page);
  84
  85                 if (likely(page != page_head &&
  86                            get_page_unless_zero(page_head))) {
  87                         unsigned long flags;
  88
  89                          if (PageHeadHuge(page_head)) {
  90                                 if (likely(PageTail(page))) {
  91                                         /*
  92                                          * __split_huge_page_refcount
  93                                          * cannot race here.
  94                                          */
  95                                         VM_BUG_ON(!PageHead(page_head));
  96                                         atomic_dec(&page->_mapcount);
  97                                         if (put_page_testzero(page_head))
  98                                                 VM_BUG_ON(1);
  99                                         if (put_page_testzero(page_head))
 100                                                 __put_compound_page(page_head);
 101                                         return;
 102                                 } else {
 103                                         /*
 104                                          * __split_huge_page_refcount
 105                                          * run before us, "page" was a
 106                                          * THP tail. The split
 107                                          * page_head has been freed
 108                                          * and reallocated as slab or
 109                                          * hugetlbfs page of smaller
 110                                          * order (only possible if
 111                                          * reallocated as slab on
 112                                          * x86).
 113                                          */
 114                                         goto skip_lock;
 115                                 }
 116                         }
 117                         /*
 118                          * page_head wasn't a dangling pointer but it
 119                          * may not be a head page anymore by the time
 120                          * we obtain the lock. That is ok as long as it
 121                          * can't be freed from under us.
 122                          */
 123                         flags = compound_lock_irqsave(page_head);
 124                         if (unlikely(!PageTail(page))) {
 125                                 /* __split_huge_page_refcount run before us */
 126                                 compound_unlock_irqrestore(page_head, flags);
 127                                 VM_BUG_ON(PageHead(page_head));
 128 skip_lock:
 129                                 if (put_page_testzero(page_head)) {
 130                                         /*
 131                                          * The head page may have been
 132                                          * freed and reallocated as a
 133                                          * compound page of smaller
 134                                          * order and then freed again.
 135                                          * All we know is that it
 136                                          * cannot have become: a THP
 137                                          * page, a compound page of
 138                                          * higher order, a tail page.
 139                                          * That is because we still
 140                                          * hold the refcount of the
 141                                          * split THP tail and
 142                                          * page_head was the THP head
 143                                          * before the split.
 144                                          */
 145                                         if (PageHead(page_head))
 146                                                 __put_compound_page(page_head);
 147                                         else
 148                                                 __put_single_page(page_head);
 149                                 }
 150 out_put_single:
 151                                 if (put_page_testzero(page))
 152                                         __put_single_page(page);
 153                                 return;
 154                         }
 155                         VM_BUG_ON(page_head != page->first_page);
 156                         /*
 157                          * We can release the refcount taken by
 158                          * get_page_unless_zero() now that
 159                          * __split_huge_page_refcount() is blocked on
 160                          * the compound_lock.
 161                          */
 162                         if (put_page_testzero(page_head))
 163                                 VM_BUG_ON(1);
 164                         /* __split_huge_page_refcount will wait now */
 165                         VM_BUG_ON(page_mapcount(page) <= 0);
 166                         atomic_dec(&page->_mapcount);
 167                         VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
 168                         VM_BUG_ON(atomic_read(&page->_count) != 0);
 169                         compound_unlock_irqrestore(page_head, flags);
 170                         if (put_page_testzero(page_head)) {
 171                                 if (PageHead(page_head))
 172                                         __put_compound_page(page_head);
 173                                 else
 174                                         __put_single_page(page_head);
 175                         }
 176                 } else {
 177                         /* page_head is a dangling pointer */
 178                         VM_BUG_ON(PageTail(page));
 179                         goto out_put_single;
 180                 }
 181         } else if (put_page_testzero(page)) {
 182                 if (PageHead(page))
 183                         __put_compound_page(page);
 184                 else
 185                         __put_single_page(page);
 186         }
 187 }
 188
 189 void put_page(struct page *page)
 190 {
 191         if (unlikely(PageCompound(page)))
 192                 put_compound_page(page);
 193         else if (put_page_testzero(page))
 194                 __put_single_page(page);
 195 }
 196 EXPORT_SYMBOL(put_page);
 197
 198 /*
 199  * This function is exported but must not be called by anything other
 200  * than get_page(). It implements the slow path of get_page().
 201  */
 202 bool __get_page_tail(struct page *page)
 203 {
 204         /*
 205          * This takes care of get_page() if run on a tail page
 206          * returned by one of the get_user_pages/follow_page variants.
 207          * get_user_pages/follow_page itself doesn't need the compound
 208          * lock because it runs __get_page_tail_foll() under the
 209          * proper PT lock that already serializes against
 210          * split_huge_page().
 211          */
 212         unsigned long flags;
 213         bool got = false;
 214         struct page *page_head = compound_trans_head(page);
 215
 216         if (likely(page != page_head && get_page_unless_zero(page_head))) {
 217                 /* Ref to put_compound_page() comment. */
 218                 if (PageHeadHuge(page_head)) {
 219                         if (likely(PageTail(page))) {
 220                                 /*
 221                                  * This is a hugetlbfs
 222                                  * page. __split_huge_page_refcount
 223                                  * cannot race here.
 224                                  */
 225                                 VM_BUG_ON(!PageHead(page_head));
 226                                 __get_page_tail_foll(page, false);
 227                                 return true;
 228                         } else {
 229                                 /*
 230                                  * __split_huge_page_refcount run
 231                                  * before us, "page" was a THP
 232                                  * tail. The split page_head has been
 233                                  * freed and reallocated as slab or
 234                                  * hugetlbfs page of smaller order
 235                                  * (only possible if reallocated as
 236                                  * slab on x86).
 237                                  */
 238                                 put_page(page_head);
 239                                 return false;
 240                         }
 241                 }
 242                 /*
 243                  * page_head wasn't a dangling pointer but it
 244                  * may not be a head page anymore by the time
 245                  * we obtain the lock. That is ok as long as it
 246                  * can't be freed from under us.
 247                  */
 248                 flags = compound_lock_irqsave(page_head);
 249                 /* here __split_huge_page_refcount won't run anymore */
 250                 if (likely(PageTail(page))) {
 251                         __get_page_tail_foll(page, false);
 252                         got = true;
 253                 }
 254                 compound_unlock_irqrestore(page_head, flags);
 255                 if (unlikely(!got))
 256                         put_page(page_head);
 257         }
 258         return got;
 259 }
 260 EXPORT_SYMBOL(__get_page_tail);
 261
 262 /**
 263  * put_pages_list() - release a list of pages
 264  * @pages: list of pages threaded on page->lru
 265  *
 266  * Release a list of pages which are strung together on page.lru.  Currently
 267  * used by read_cache_pages() and related error recovery code.
 268  */
 269 void put_pages_list(struct list_head *pages)
 270 {
 271         while (!list_empty(pages)) {
 272                 struct page *victim;
 273
 274                 victim = list_entry(pages->prev, struct page, lru);
 275                 list_del(&victim->lru);
 276                 page_cache_release(victim);
 277         }
 278 }
 279 EXPORT_SYMBOL(put_pages_list);
 280
 281 static void pagevec_lru_move_fn(struct pagevec *pvec,
 282                                 void (*move_fn)(struct page *page, void *arg),
 283                                 void *arg)
 284 {
 285         int i;
 286         struct zone *zone = NULL;
 287         unsigned long flags = 0;
 288
 289         for (i = 0; i < pagevec_count(pvec); i++) {
 290                 struct page *page = pvec->pages[i];
 291                 struct zone *pagezone = page_zone(page);
 292
 293                 if (pagezone != zone) {
 294                         if (zone)
 295                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 296                         zone = pagezone;
 297                         spin_lock_irqsave(&zone->lru_lock, flags);
 298                 }
 299
 300                 (*move_fn)(page, arg);
 301         }
 302         if (zone)
 303                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 304         release_pages(pvec->pages, pvec->nr, pvec->cold);
 305         pagevec_reinit(pvec);
 306 }
 307
 308 static void pagevec_move_tail_fn(struct page *page, void *arg)
 309 {
 310         int *pgmoved = arg;
 311         struct zone *zone = page_zone(page);
 312
 313         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 314                 enum lru_list lru = page_lru_base_type(page);
 315                 list_move_tail(&page->lru, &zone->lru[lru].list);
 316                 mem_cgroup_rotate_reclaimable_page(page);
 317                 (*pgmoved)++;
 318         }
 319 }
 320
 321 /*
 322  * pagevec_move_tail() must be called with IRQ disabled.
 323  * Otherwise this may cause nasty races.
 324  */
 325 static void pagevec_move_tail(struct pagevec *pvec)
 326 {
 327         int pgmoved = 0;
 328
 329         pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
 330         __count_vm_events(PGROTATED, pgmoved);
 331 }
 332
 333 /*
 334  * Writeback is about to end against a page which has been marked for immediate
 335  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 336  * inactive list.
 337  */
 338 void rotate_reclaimable_page(struct page *page)
 339 {
 340         if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
 341             !PageUnevictable(page) && PageLRU(page)) {
 342                 struct pagevec *pvec;
 343                 unsigned long flags;
 344
 345                 page_cache_get(page);
 346                 local_irq_save(flags);
 347                 pvec = &__get_cpu_var(lru_rotate_pvecs);
 348                 if (!pagevec_add(pvec, page))
 349                         pagevec_move_tail(pvec);
 350                 local_irq_restore(flags);
 351         }
 352 }
 353
 354 static void update_page_reclaim_stat(struct zone *zone, struct page *page,
 355                                      int file, int rotated)
 356 {
 357         struct zone_reclaim_stat *reclaim_stat = &zone->reclaim_stat;
 358         struct zone_reclaim_stat *memcg_reclaim_stat;
 359
 360         memcg_reclaim_stat = mem_cgroup_get_reclaim_stat_from_page(page);
 361
 362         reclaim_stat->recent_scanned[file]++;
 363         if (rotated)
 364                 reclaim_stat->recent_rotated[file]++;
 365
 366         if (!memcg_reclaim_stat)
 367                 return;
 368
 369         memcg_reclaim_stat->recent_scanned[file]++;
 370         if (rotated)
 371                 memcg_reclaim_stat->recent_rotated[file]++;
 372 }
 373
 374 static void __activate_page(struct page *page, void *arg)
 375 {
 376         struct zone *zone = page_zone(page);
 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                 del_page_from_lru_list(zone, page, lru);
 382
 383                 SetPageActive(page);
 384                 lru += LRU_ACTIVE;
 385                 add_page_to_lru_list(zone, page, lru);
 386                 __count_vm_event(PGACTIVATE);
 387
 388                 update_page_reclaim_stat(zone, page, 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, 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
 448 EXPORT_SYMBOL(mark_page_accessed);
 449
 450 void __lru_cache_add(struct page *page, enum lru_list lru)
 451 {
 452         struct pagevec *pvec = &get_cpu_var(lru_add_pvecs)[lru];
 453
 454         page_cache_get(page);
 455         if (!pagevec_add(pvec, page))
 456                 ____pagevec_lru_add(pvec, lru);
 457         put_cpu_var(lru_add_pvecs);
 458 }
 459 EXPORT_SYMBOL(__lru_cache_add);
 460
 461 /**
 462  * lru_cache_add_lru - add a page to a page list
 463  * @page: the page to be added to the LRU.
 464  * @lru: the LRU list to which the page is added.
 465  */
 466 void lru_cache_add_lru(struct page *page, enum lru_list lru)
 467 {
 468         if (PageActive(page)) {
 469                 VM_BUG_ON(PageUnevictable(page));
 470                 ClearPageActive(page);
 471         } else if (PageUnevictable(page)) {
 472                 VM_BUG_ON(PageActive(page));
 473                 ClearPageUnevictable(page);
 474         }
 475
 476         VM_BUG_ON(PageLRU(page) || PageActive(page) || PageUnevictable(page));
 477         __lru_cache_add(page, lru);
 478 }
 479
 480 /**
 481  * add_page_to_unevictable_list - add a page to the unevictable list
 482  * @page:  the page to be added to the unevictable list
 483  *
 484  * Add page directly to its zone's unevictable list.  To avoid races with
 485  * tasks that might be making the page evictable, through eg. munlock,
 486  * munmap or exit, while it's not on the lru, we want to add the page
 487  * while it's locked or otherwise "invisible" to other tasks.  This is
 488  * difficult to do when using the pagevec cache, so bypass that.
 489  */
 490 void add_page_to_unevictable_list(struct page *page)
 491 {
 492         struct zone *zone = page_zone(page);
 493
 494         spin_lock_irq(&zone->lru_lock);
 495         SetPageUnevictable(page);
 496         SetPageLRU(page);
 497         add_page_to_lru_list(zone, page, LRU_UNEVICTABLE);
 498         spin_unlock_irq(&zone->lru_lock);
 499 }
 500
 501 /*
 502  * If the page can not be invalidated, it is moved to the
 503  * inactive list to speed up its reclaim.  It is moved to the
 504  * head of the list, rather than the tail, to give the flusher
 505  * threads some time to write it out, as this is much more
 506  * effective than the single-page writeout from reclaim.
 507  *
 508  * If the page isn't page_mapped and dirty/writeback, the page
 509  * could reclaim asap using PG_reclaim.
 510  *
 511  * 1. active, mapped page -> none
 512  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 513  * 3. inactive, mapped page -> none
 514  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 515  * 5. inactive, clean -> inactive, tail
 516  * 6. Others -> none
 517  *
 518  * In 4, why it moves inactive's head, the VM expects the page would
 519  * be write it out by flusher threads as this is much more effective
 520  * than the single-page writeout from reclaim.
 521  */
 522 static void lru_deactivate_fn(struct page *page, void *arg)
 523 {
 524         int lru, file;
 525         bool active;
 526         struct zone *zone = page_zone(page);
 527
 528         if (!PageLRU(page))
 529                 return;
 530
 531         if (PageUnevictable(page))
 532                 return;
 533
 534         /* Some processes are using the page */
 535         if (page_mapped(page))
 536                 return;
 537
 538         active = PageActive(page);
 539
 540         file = page_is_file_cache(page);
 541         lru = page_lru_base_type(page);
 542         del_page_from_lru_list(zone, page, lru + active);
 543         ClearPageActive(page);
 544         ClearPageReferenced(page);
 545         add_page_to_lru_list(zone, page, lru);
 546
 547         if (PageWriteback(page) || PageDirty(page)) {
 548                 /*
 549                  * PG_reclaim could be raced with end_page_writeback
 550                  * It can make readahead confusing.  But race window
 551                  * is _really_ small and  it's non-critical problem.
 552                  */
 553                 SetPageReclaim(page);
 554         } else {
 555                 /*
 556                  * The page's writeback ends up during pagevec
 557                  * We moves tha page into tail of inactive.
 558                  */
 559                 list_move_tail(&page->lru, &zone->lru[lru].list);
 560                 mem_cgroup_rotate_reclaimable_page(page);
 561                 __count_vm_event(PGROTATED);
 562         }
 563
 564         if (active)
 565                 __count_vm_event(PGDEACTIVATE);
 566         update_page_reclaim_stat(zone, page, 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 static void drain_cpu_pagevecs(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         drain_cpu_pagevecs(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         unsigned long uninitialized_var(flags);
 667
 668         for (i = 0; i < nr; i++) {
 669                 struct page *page = pages[i];
 670
 671                 if (unlikely(PageCompound(page))) {
 672                         if (zone) {
 673                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 674                                 zone = NULL;
 675                         }
 676                         put_compound_page(page);
 677                         continue;
 678                 }
 679
 680                 if (!put_page_testzero(page))
 681                         continue;
 682
 683                 if (PageLRU(page)) {
 684                         struct zone *pagezone = page_zone(page);
 685
 686                         if (pagezone != zone) {
 687                                 if (zone)
 688                                         spin_unlock_irqrestore(&zone->lru_lock,
 689                                                                         flags);
 690                                 zone = pagezone;
 691                                 spin_lock_irqsave(&zone->lru_lock, flags);
 692                         }
 693                         VM_BUG_ON(!PageLRU(page));
 694                         __ClearPageLRU(page);
 695                         del_page_from_lru(zone, page);
 696                 }
 697
 698                 list_add(&page->lru, &pages_to_free);
 699         }
 700         if (zone)
 701                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 702
 703         free_hot_cold_page_list(&pages_to_free, cold);
 704 }
 705 EXPORT_SYMBOL(release_pages);
 706
 707 /*
 708  * The pages which we're about to release may be in the deferred lru-addition
 709  * queues.  That would prevent them from really being freed right now.  That's
 710  * OK from a correctness point of view but is inefficient - those pages may be
 711  * cache-warm and we want to give them back to the page allocator ASAP.
 712  *
 713  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 714  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 715  * mutual recursion.
 716  */
 717 void __pagevec_release(struct pagevec *pvec)
 718 {
 719         lru_add_drain();
 720         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 721         pagevec_reinit(pvec);
 722 }
 723
 724 EXPORT_SYMBOL(__pagevec_release);
 725
 726 /* used by __split_huge_page_refcount() */
 727 void lru_add_page_tail(struct zone* zone,
 728                        struct page *page, struct page *page_tail)
 729 {
 730         int active;
 731         enum lru_list lru;
 732         const int file = 0;
 733         struct list_head *head;
 734
 735         VM_BUG_ON(!PageHead(page));
 736         VM_BUG_ON(PageCompound(page_tail));
 737         VM_BUG_ON(PageLRU(page_tail));
 738         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&zone->lru_lock));
 739
 740         SetPageLRU(page_tail);
 741
 742         if (page_evictable(page_tail, NULL)) {
 743                 if (PageActive(page)) {
 744                         SetPageActive(page_tail);
 745                         active = 1;
 746                         lru = LRU_ACTIVE_ANON;
 747                 } else {
 748                         active = 0;
 749                         lru = LRU_INACTIVE_ANON;
 750                 }
 751                 update_page_reclaim_stat(zone, page_tail, file, active);
 752                 if (likely(PageLRU(page)))
 753                         head = page->lru.prev;
 754                 else
 755                         head = &zone->lru[lru].list;
 756                 __add_page_to_lru_list(zone, page_tail, lru, head);
 757         } else {
 758                 SetPageUnevictable(page_tail);
 759                 add_page_to_lru_list(zone, page_tail, LRU_UNEVICTABLE);
 760         }
 761 }
 762
 763 static void ____pagevec_lru_add_fn(struct page *page, void *arg)
 764 {
 765         enum lru_list lru = (enum lru_list)arg;
 766         struct zone *zone = page_zone(page);
 767         int file = is_file_lru(lru);
 768         int active = is_active_lru(lru);
 769
 770         VM_BUG_ON(PageActive(page));
 771         VM_BUG_ON(PageUnevictable(page));
 772         VM_BUG_ON(PageLRU(page));
 773
 774         SetPageLRU(page);
 775         if (active)
 776                 SetPageActive(page);
 777         update_page_reclaim_stat(zone, page, file, active);
 778         add_page_to_lru_list(zone, page, lru);
 779 }
 780
 781 /*
 782  * Add the passed pages to the LRU, then drop the caller's refcount
 783  * on them.  Reinitialises the caller's pagevec.
 784  */
 785 void ____pagevec_lru_add(struct pagevec *pvec, enum lru_list lru)
 786 {
 787         VM_BUG_ON(is_unevictable_lru(lru));
 788
 789         pagevec_lru_move_fn(pvec, ____pagevec_lru_add_fn, (void *)lru);
 790 }
 791
 792 EXPORT_SYMBOL(____pagevec_lru_add);
 793
 794 /*
 795  * Try to drop buffers from the pages in a pagevec
 796  */
 797 void pagevec_strip(struct pagevec *pvec)
 798 {
 799         int i;
 800
 801         for (i = 0; i < pagevec_count(pvec); i++) {
 802                 struct page *page = pvec->pages[i];
 803
 804                 if (page_has_private(page) && trylock_page(page)) {
 805                         if (page_has_private(page))
 806                                 try_to_release_page(page, 0);
 807                         unlock_page(page);
 808                 }
 809         }
 810 }
 811
 812 /**
 813  * pagevec_lookup - gang pagecache lookup
 814  * @pvec:       Where the resulting pages are placed
 815  * @mapping:    The address_space to search
 816  * @start:      The starting page index
 817  * @nr_pages:   The maximum number of pages
 818  *
 819  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 820  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 821  * reference against the pages in @pvec.
 822  *
 823  * The search returns a group of mapping-contiguous pages with ascending
 824  * indexes.  There may be holes in the indices due to not-present pages.
 825  *
 826  * pagevec_lookup() returns the number of pages which were found.
 827  */
 828 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 829                 pgoff_t start, unsigned nr_pages)
 830 {
 831         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 832         return pagevec_count(pvec);
 833 }
 834
 835 EXPORT_SYMBOL(pagevec_lookup);
 836
 837 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 838                 pgoff_t *index, int tag, unsigned nr_pages)
 839 {
 840         pvec->nr = find_get_pages_tag(mapping, index, tag,
 841                                         nr_pages, pvec->pages);
 842         return pagevec_count(pvec);
 843 }
 844
 845 EXPORT_SYMBOL(pagevec_lookup_tag);
 846
 847 /*
 848  * Perform any setup for the swap system
 849  */
 850 void __init swap_setup(void)
 851 {
 852         unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
 853
 854 #ifdef CONFIG_SWAP
 855         bdi_init(swapper_space.backing_dev_info);
 856 #endif
 857
 858         /* Use a smaller cluster for small-memory machines */
 859         if (megs < 16)
 860                 page_cluster = 2;
 861         else
 862                 page_cluster = 3;
 863         /*
 864          * Right now other parts of the system means that we
 865          * _really_ don't want to cluster much more
 866          */
 867 }