mm/swap.c

   1 /*
   2  *  linux/mm/swap.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5  */
   6
   7 /*
   8  * This file contains the default values for the operation of the
   9  * Linux VM subsystem. Fine-tuning documentation can be found in
  10  * Documentation/sysctl/vm.txt.
  11  * Started 18.12.91
  12  * Swap aging added 23.2.95, Stephen Tweedie.
  13  * Buffermem limits added 12.3.98, Rik van Riel.
  14  */
  15
  16 #include <linux/mm.h>
  17 #include <linux/sched.h>
  18 #include <linux/kernel_stat.h>
  19 #include <linux/swap.h>
  20 #include <linux/mman.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/pagevec.h>
  23 #include <linux/init.h>
  24 #include <linux/export.h>
  25 #include <linux/mm_inline.h>
  26 #include <linux/percpu_counter.h>
  27 #include <linux/percpu.h>
  28 #include <linux/cpu.h>
  29 #include <linux/notifier.h>
  30 #include <linux/backing-dev.h>
  31 #include <linux/memcontrol.h>
  32 #include <linux/gfp.h>
  33 #include <linux/uio.h>
  34
  35 #include "internal.h"
  36
  37 #define CREATE_TRACE_POINTS
  38 #include <trace/events/pagemap.h>
  39
  40 /* How many pages do we try to swap or page in/out together? */
  41 int page_cluster;
  42
  43 static DEFINE_PER_CPU(struct pagevec, lru_add_pvec);
  44 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
  45 static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
  46
  47 /*
  48  * This path almost never happens for VM activity - pages are normally
  49  * freed via pagevecs.  But it gets used by networking.
  50  */
  51 static void __page_cache_release(struct page *page)
  52 {
  53         if (PageLRU(page)) {
  54                 struct zone *zone = page_zone(page);
  55                 struct lruvec *lruvec;
  56                 unsigned long flags;
  57
  58                 spin_lock_irqsave(&zone->lru_lock, flags);
  59                 lruvec = mem_cgroup_page_lruvec(page, zone);
  60                 VM_BUG_ON(!PageLRU(page));
  61                 __ClearPageLRU(page);
  62                 del_page_from_lru_list(page, lruvec, page_off_lru(page));
  63                 spin_unlock_irqrestore(&zone->lru_lock, flags);
  64         }
  65 }
  66
  67 static void __put_single_page(struct page *page)
  68 {
  69         __page_cache_release(page);
  70         free_hot_cold_page(page, 0);
  71 }
  72
  73 static void __put_compound_page(struct page *page)
  74 {
  75         compound_page_dtor *dtor;
  76
  77         __page_cache_release(page);
  78         dtor = get_compound_page_dtor(page);
  79         (*dtor)(page);
  80 }
  81
  82 static void put_compound_page(struct page *page)
  83 {
  84         if (unlikely(PageTail(page))) {
  85                 /* __split_huge_page_refcount can run under us */
  86                 struct page *page_head = compound_trans_head(page);
  87
  88                 if (likely(page != page_head &&
  89                            get_page_unless_zero(page_head))) {
  90                         unsigned long flags;
  91
  92                         /*
  93                          * THP can not break up slab pages so avoid taking
  94                          * compound_lock().  Slab performs non-atomic bit ops
  95                          * on page->flags for better performance.  In particular
  96                          * slab_unlock() in slub used to be a hot path.  It is
  97                          * still hot on arches that do not support
  98                          * this_cpu_cmpxchg_double().
  99                          */
 100                         if (PageSlab(page_head)) {
 101                                 if (PageTail(page)) {
 102                                         if (put_page_testzero(page_head))
 103                                                 VM_BUG_ON(1);
 104
 105                                         atomic_dec(&page->_mapcount);
 106                                         goto skip_lock_tail;
 107                                 } else
 108                                         goto skip_lock;
 109                         }
 110                         /*
 111                          * page_head wasn't a dangling pointer but it
 112                          * may not be a head page anymore by the time
 113                          * we obtain the lock. That is ok as long as it
 114                          * can't be freed from under us.
 115                          */
 116                         flags = compound_lock_irqsave(page_head);
 117                         if (unlikely(!PageTail(page))) {
 118                                 /* __split_huge_page_refcount run before us */
 119                                 compound_unlock_irqrestore(page_head, flags);
 120 skip_lock:
 121                                 if (put_page_testzero(page_head))
 122                                         __put_single_page(page_head);
 123 out_put_single:
 124                                 if (put_page_testzero(page))
 125                                         __put_single_page(page);
 126                                 return;
 127                         }
 128                         VM_BUG_ON(page_head != page->first_page);
 129                         /*
 130                          * We can release the refcount taken by
 131                          * get_page_unless_zero() now that
 132                          * __split_huge_page_refcount() is blocked on
 133                          * the compound_lock.
 134                          */
 135                         if (put_page_testzero(page_head))
 136                                 VM_BUG_ON(1);
 137                         /* __split_huge_page_refcount will wait now */
 138                         VM_BUG_ON(page_mapcount(page) <= 0);
 139                         atomic_dec(&page->_mapcount);
 140                         VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
 141                         VM_BUG_ON(atomic_read(&page->_count) != 0);
 142                         compound_unlock_irqrestore(page_head, flags);
 143
 144 skip_lock_tail:
 145                         if (put_page_testzero(page_head)) {
 146                                 if (PageHead(page_head))
 147                                         __put_compound_page(page_head);
 148                                 else
 149                                         __put_single_page(page_head);
 150                         }
 151                 } else {
 152                         /* page_head is a dangling pointer */
 153                         VM_BUG_ON(PageTail(page));
 154                         goto out_put_single;
 155                 }
 156         } else if (put_page_testzero(page)) {
 157                 if (PageHead(page))
 158                         __put_compound_page(page);
 159                 else
 160                         __put_single_page(page);
 161         }
 162 }
 163
 164 void put_page(struct page *page)
 165 {
 166         if (unlikely(PageCompound(page)))
 167                 put_compound_page(page);
 168         else if (put_page_testzero(page))
 169                 __put_single_page(page);
 170 }
 171 EXPORT_SYMBOL(put_page);
 172
 173 /*
 174  * This function is exported but must not be called by anything other
 175  * than get_page(). It implements the slow path of get_page().
 176  */
 177 bool __get_page_tail(struct page *page)
 178 {
 179         /*
 180          * This takes care of get_page() if run on a tail page
 181          * returned by one of the get_user_pages/follow_page variants.
 182          * get_user_pages/follow_page itself doesn't need the compound
 183          * lock because it runs __get_page_tail_foll() under the
 184          * proper PT lock that already serializes against
 185          * split_huge_page().
 186          */
 187         unsigned long flags;
 188         bool got = false;
 189         struct page *page_head = compound_trans_head(page);
 190
 191         if (likely(page != page_head && get_page_unless_zero(page_head))) {
 192
 193                 /* Ref to put_compound_page() comment. */
 194                 if (PageSlab(page_head)) {
 195                         if (likely(PageTail(page))) {
 196                                 __get_page_tail_foll(page, false);
 197                                 return true;
 198                         } else {
 199                                 put_page(page_head);
 200                                 return false;
 201                         }
 202                 }
 203
 204                 /*
 205                  * page_head wasn't a dangling pointer but it
 206                  * may not be a head page anymore by the time
 207                  * we obtain the lock. That is ok as long as it
 208                  * can't be freed from under us.
 209                  */
 210                 flags = compound_lock_irqsave(page_head);
 211                 /* here __split_huge_page_refcount won't run anymore */
 212                 if (likely(PageTail(page))) {
 213                         __get_page_tail_foll(page, false);
 214                         got = true;
 215                 }
 216                 compound_unlock_irqrestore(page_head, flags);
 217                 if (unlikely(!got))
 218                         put_page(page_head);
 219         }
 220         return got;
 221 }
 222 EXPORT_SYMBOL(__get_page_tail);
 223
 224 /**
 225  * put_pages_list() - release a list of pages
 226  * @pages: list of pages threaded on page->lru
 227  *
 228  * Release a list of pages which are strung together on page.lru.  Currently
 229  * used by read_cache_pages() and related error recovery code.
 230  */
 231 void put_pages_list(struct list_head *pages)
 232 {
 233         while (!list_empty(pages)) {
 234                 struct page *victim;
 235
 236                 victim = list_entry(pages->prev, struct page, lru);
 237                 list_del(&victim->lru);
 238                 page_cache_release(victim);
 239         }
 240 }
 241 EXPORT_SYMBOL(put_pages_list);
 242
 243 /*
 244  * get_kernel_pages() - pin kernel pages in memory
 245  * @kiov:       An array of struct kvec structures
 246  * @nr_segs:    number of segments to pin
 247  * @write:      pinning for read/write, currently ignored
 248  * @pages:      array that receives pointers to the pages pinned.
 249  *              Should be at least nr_segs long.
 250  *
 251  * Returns number of pages pinned. This may be fewer than the number
 252  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 253  * were pinned, returns -errno. Each page returned must be released
 254  * with a put_page() call when it is finished with.
 255  */
 256 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
 257                 struct page **pages)
 258 {
 259         int seg;
 260
 261         for (seg = 0; seg < nr_segs; seg++) {
 262                 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
 263                         return seg;
 264
 265                 pages[seg] = kmap_to_page(kiov[seg].iov_base);
 266                 page_cache_get(pages[seg]);
 267         }
 268
 269         return seg;
 270 }
 271 EXPORT_SYMBOL_GPL(get_kernel_pages);
 272
 273 /*
 274  * get_kernel_page() - pin a kernel page in memory
 275  * @start:      starting kernel address
 276  * @write:      pinning for read/write, currently ignored
 277  * @pages:      array that receives pointer to the page pinned.
 278  *              Must be at least nr_segs long.
 279  *
 280  * Returns 1 if page is pinned. If the page was not pinned, returns
 281  * -errno. The page returned must be released with a put_page() call
 282  * when it is finished with.
 283  */
 284 int get_kernel_page(unsigned long start, int write, struct page **pages)
 285 {
 286         const struct kvec kiov = {
 287                 .iov_base = (void *)start,
 288                 .iov_len = PAGE_SIZE
 289         };
 290
 291         return get_kernel_pages(&kiov, 1, write, pages);
 292 }
 293 EXPORT_SYMBOL_GPL(get_kernel_page);
 294
 295 static void pagevec_lru_move_fn(struct pagevec *pvec,
 296         void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
 297         void *arg)
 298 {
 299         int i;
 300         struct zone *zone = NULL;
 301         struct lruvec *lruvec;
 302         unsigned long flags = 0;
 303
 304         for (i = 0; i < pagevec_count(pvec); i++) {
 305                 struct page *page = pvec->pages[i];
 306                 struct zone *pagezone = page_zone(page);
 307
 308                 if (pagezone != zone) {
 309                         if (zone)
 310                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 311                         zone = pagezone;
 312                         spin_lock_irqsave(&zone->lru_lock, flags);
 313                 }
 314
 315                 lruvec = mem_cgroup_page_lruvec(page, zone);
 316                 (*move_fn)(page, lruvec, arg);
 317         }
 318         if (zone)
 319                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 320         release_pages(pvec->pages, pvec->nr, pvec->cold);
 321         pagevec_reinit(pvec);
 322 }
 323
 324 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
 325                                  void *arg)
 326 {
 327         int *pgmoved = arg;
 328
 329         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 330                 enum lru_list lru = page_lru_base_type(page);
 331                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 332                 (*pgmoved)++;
 333         }
 334 }
 335
 336 /*
 337  * pagevec_move_tail() must be called with IRQ disabled.
 338  * Otherwise this may cause nasty races.
 339  */
 340 static void pagevec_move_tail(struct pagevec *pvec)
 341 {
 342         int pgmoved = 0;
 343
 344         pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
 345         __count_vm_events(PGROTATED, pgmoved);
 346 }
 347
 348 /*
 349  * Writeback is about to end against a page which has been marked for immediate
 350  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 351  * inactive list.
 352  */
 353 void rotate_reclaimable_page(struct page *page)
 354 {
 355         if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
 356             !PageUnevictable(page) && PageLRU(page)) {
 357                 struct pagevec *pvec;
 358                 unsigned long flags;
 359
 360                 page_cache_get(page);
 361                 local_irq_save(flags);
 362                 pvec = &__get_cpu_var(lru_rotate_pvecs);
 363                 if (!pagevec_add(pvec, page))
 364                         pagevec_move_tail(pvec);
 365                 local_irq_restore(flags);
 366         }
 367 }
 368
 369 static void update_page_reclaim_stat(struct lruvec *lruvec,
 370                                      int file, int rotated)
 371 {
 372         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
 373
 374         reclaim_stat->recent_scanned[file]++;
 375         if (rotated)
 376                 reclaim_stat->recent_rotated[file]++;
 377 }
 378
 379 static void __activate_page(struct page *page, struct lruvec *lruvec,
 380                             void *arg)
 381 {
 382         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 383                 int file = page_is_file_cache(page);
 384                 int lru = page_lru_base_type(page);
 385
 386                 del_page_from_lru_list(page, lruvec, lru);
 387                 SetPageActive(page);
 388                 lru += LRU_ACTIVE;
 389                 add_page_to_lru_list(page, lruvec, lru);
 390                 trace_mm_lru_activate(page, page_to_pfn(page));
 391
 392                 __count_vm_event(PGACTIVATE);
 393                 update_page_reclaim_stat(lruvec, file, 1);
 394         }
 395 }
 396
 397 #ifdef CONFIG_SMP
 398 static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
 399
 400 static void activate_page_drain(int cpu)
 401 {
 402         struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
 403
 404         if (pagevec_count(pvec))
 405                 pagevec_lru_move_fn(pvec, __activate_page, NULL);
 406 }
 407
 408 void activate_page(struct page *page)
 409 {
 410         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 411                 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
 412
 413                 page_cache_get(page);
 414                 if (!pagevec_add(pvec, page))
 415                         pagevec_lru_move_fn(pvec, __activate_page, NULL);
 416                 put_cpu_var(activate_page_pvecs);
 417         }
 418 }
 419
 420 #else
 421 static inline void activate_page_drain(int cpu)
 422 {
 423 }
 424
 425 void activate_page(struct page *page)
 426 {
 427         struct zone *zone = page_zone(page);
 428
 429         spin_lock_irq(&zone->lru_lock);
 430         __activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
 431         spin_unlock_irq(&zone->lru_lock);
 432 }
 433 #endif
 434
 435 /*
 436  * Mark a page as having seen activity.
 437  *
 438  * inactive,unreferenced        ->      inactive,referenced
 439  * inactive,referenced          ->      active,unreferenced
 440  * active,unreferenced          ->      active,referenced
 441  */
 442 void mark_page_accessed(struct page *page)
 443 {
 444         if (!PageActive(page) && !PageUnevictable(page) &&
 445                         PageReferenced(page) && PageLRU(page)) {
 446                 activate_page(page);
 447                 ClearPageReferenced(page);
 448         } else if (!PageReferenced(page)) {
 449                 SetPageReferenced(page);
 450         }
 451 }
 452 EXPORT_SYMBOL(mark_page_accessed);
 453
 454 /*
 455  * Queue the page for addition to the LRU via pagevec. The decision on whether
 456  * to add the page to the [in]active [file|anon] list is deferred until the
 457  * pagevec is drained. This gives a chance for the caller of __lru_cache_add()
 458  * have the page added to the active list using mark_page_accessed().
 459  */
 460 void __lru_cache_add(struct page *page, enum lru_list lru)
 461 {
 462         struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
 463
 464         if (is_active_lru(lru))
 465                 SetPageActive(page);
 466         else
 467                 ClearPageActive(page);
 468
 469         page_cache_get(page);
 470         if (!pagevec_space(pvec))
 471                 __pagevec_lru_add(pvec, lru);
 472         pagevec_add(pvec, page);
 473         put_cpu_var(lru_add_pvec);
 474 }
 475 EXPORT_SYMBOL(__lru_cache_add);
 476
 477 /**
 478  * lru_cache_add_lru - add a page to a page list
 479  * @page: the page to be added to the LRU.
 480  * @lru: the LRU list to which the page is added.
 481  */
 482 void lru_cache_add_lru(struct page *page, enum lru_list lru)
 483 {
 484         if (PageActive(page)) {
 485                 VM_BUG_ON(PageUnevictable(page));
 486         } else if (PageUnevictable(page)) {
 487                 VM_BUG_ON(PageActive(page));
 488         }
 489
 490         VM_BUG_ON(PageLRU(page));
 491         __lru_cache_add(page, lru);
 492 }
 493
 494 /**
 495  * add_page_to_unevictable_list - add a page to the unevictable list
 496  * @page:  the page to be added to the unevictable list
 497  *
 498  * Add page directly to its zone's unevictable list.  To avoid races with
 499  * tasks that might be making the page evictable, through eg. munlock,
 500  * munmap or exit, while it's not on the lru, we want to add the page
 501  * while it's locked or otherwise "invisible" to other tasks.  This is
 502  * difficult to do when using the pagevec cache, so bypass that.
 503  */
 504 void add_page_to_unevictable_list(struct page *page)
 505 {
 506         struct zone *zone = page_zone(page);
 507         struct lruvec *lruvec;
 508
 509         spin_lock_irq(&zone->lru_lock);
 510         lruvec = mem_cgroup_page_lruvec(page, zone);
 511         SetPageUnevictable(page);
 512         SetPageLRU(page);
 513         add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
 514         spin_unlock_irq(&zone->lru_lock);
 515 }
 516
 517 /*
 518  * If the page can not be invalidated, it is moved to the
 519  * inactive list to speed up its reclaim.  It is moved to the
 520  * head of the list, rather than the tail, to give the flusher
 521  * threads some time to write it out, as this is much more
 522  * effective than the single-page writeout from reclaim.
 523  *
 524  * If the page isn't page_mapped and dirty/writeback, the page
 525  * could reclaim asap using PG_reclaim.
 526  *
 527  * 1. active, mapped page -> none
 528  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 529  * 3. inactive, mapped page -> none
 530  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 531  * 5. inactive, clean -> inactive, tail
 532  * 6. Others -> none
 533  *
 534  * In 4, why it moves inactive's head, the VM expects the page would
 535  * be write it out by flusher threads as this is much more effective
 536  * than the single-page writeout from reclaim.
 537  */
 538 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 539                               void *arg)
 540 {
 541         int lru, file;
 542         bool active;
 543
 544         if (!PageLRU(page))
 545                 return;
 546
 547         if (PageUnevictable(page))
 548                 return;
 549
 550         /* Some processes are using the page */
 551         if (page_mapped(page))
 552                 return;
 553
 554         active = PageActive(page);
 555         file = page_is_file_cache(page);
 556         lru = page_lru_base_type(page);
 557
 558         del_page_from_lru_list(page, lruvec, lru + active);
 559         ClearPageActive(page);
 560         ClearPageReferenced(page);
 561         add_page_to_lru_list(page, lruvec, lru);
 562
 563         if (PageWriteback(page) || PageDirty(page)) {
 564                 /*
 565                  * PG_reclaim could be raced with end_page_writeback
 566                  * It can make readahead confusing.  But race window
 567                  * is _really_ small and  it's non-critical problem.
 568                  */
 569                 SetPageReclaim(page);
 570         } else {
 571                 /*
 572                  * The page's writeback ends up during pagevec
 573                  * We moves tha page into tail of inactive.
 574                  */
 575                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 576                 __count_vm_event(PGROTATED);
 577         }
 578
 579         if (active)
 580                 __count_vm_event(PGDEACTIVATE);
 581         update_page_reclaim_stat(lruvec, file, 0);
 582 }
 583
 584 /*
 585  * Drain pages out of the cpu's pagevecs.
 586  * Either "cpu" is the current CPU, and preemption has already been
 587  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 588  */
 589 void lru_add_drain_cpu(int cpu)
 590 {
 591         struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
 592
 593         if (pagevec_count(pvec))
 594                 __pagevec_lru_add(pvec, NR_LRU_LISTS);
 595
 596         pvec = &per_cpu(lru_rotate_pvecs, cpu);
 597         if (pagevec_count(pvec)) {
 598                 unsigned long flags;
 599
 600                 /* No harm done if a racing interrupt already did this */
 601                 local_irq_save(flags);
 602                 pagevec_move_tail(pvec);
 603                 local_irq_restore(flags);
 604         }
 605
 606         pvec = &per_cpu(lru_deactivate_pvecs, cpu);
 607         if (pagevec_count(pvec))
 608                 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 609
 610         activate_page_drain(cpu);
 611 }
 612
 613 /**
 614  * deactivate_page - forcefully deactivate a page
 615  * @page: page to deactivate
 616  *
 617  * This function hints the VM that @page is a good reclaim candidate,
 618  * for example if its invalidation fails due to the page being dirty
 619  * or under writeback.
 620  */
 621 void deactivate_page(struct page *page)
 622 {
 623         /*
 624          * In a workload with many unevictable page such as mprotect, unevictable
 625          * page deactivation for accelerating reclaim is pointless.
 626          */
 627         if (PageUnevictable(page))
 628                 return;
 629
 630         if (likely(get_page_unless_zero(page))) {
 631                 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
 632
 633                 if (!pagevec_add(pvec, page))
 634                         pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 635                 put_cpu_var(lru_deactivate_pvecs);
 636         }
 637 }
 638
 639 void lru_add_drain(void)
 640 {
 641         lru_add_drain_cpu(get_cpu());
 642         put_cpu();
 643 }
 644
 645 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 646 {
 647         lru_add_drain();
 648 }
 649
 650 /*
 651  * Returns 0 for success
 652  */
 653 int lru_add_drain_all(void)
 654 {
 655         return schedule_on_each_cpu(lru_add_drain_per_cpu);
 656 }
 657
 658 /*
 659  * Batched page_cache_release().  Decrement the reference count on all the
 660  * passed pages.  If it fell to zero then remove the page from the LRU and
 661  * free it.
 662  *
 663  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 664  * for the remainder of the operation.
 665  *
 666  * The locking in this function is against shrink_inactive_list(): we recheck
 667  * the page count inside the lock to see whether shrink_inactive_list()
 668  * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
 669  * will free it.
 670  */
 671 void release_pages(struct page **pages, int nr, int cold)
 672 {
 673         int i;
 674         LIST_HEAD(pages_to_free);
 675         struct zone *zone = NULL;
 676         struct lruvec *lruvec;
 677         unsigned long uninitialized_var(flags);
 678
 679         for (i = 0; i < nr; i++) {
 680                 struct page *page = pages[i];
 681
 682                 if (unlikely(PageCompound(page))) {
 683                         if (zone) {
 684                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 685                                 zone = NULL;
 686                         }
 687                         put_compound_page(page);
 688                         continue;
 689                 }
 690
 691                 if (!put_page_testzero(page))
 692                         continue;
 693
 694                 if (PageLRU(page)) {
 695                         struct zone *pagezone = page_zone(page);
 696
 697                         if (pagezone != zone) {
 698                                 if (zone)
 699                                         spin_unlock_irqrestore(&zone->lru_lock,
 700                                                                         flags);
 701                                 zone = pagezone;
 702                                 spin_lock_irqsave(&zone->lru_lock, flags);
 703                         }
 704
 705                         lruvec = mem_cgroup_page_lruvec(page, zone);
 706                         VM_BUG_ON(!PageLRU(page));
 707                         __ClearPageLRU(page);
 708                         del_page_from_lru_list(page, lruvec, page_off_lru(page));
 709                 }
 710
 711                 list_add(&page->lru, &pages_to_free);
 712         }
 713         if (zone)
 714                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 715
 716         free_hot_cold_page_list(&pages_to_free, cold);
 717 }
 718 EXPORT_SYMBOL(release_pages);
 719
 720 /*
 721  * The pages which we're about to release may be in the deferred lru-addition
 722  * queues.  That would prevent them from really being freed right now.  That's
 723  * OK from a correctness point of view but is inefficient - those pages may be
 724  * cache-warm and we want to give them back to the page allocator ASAP.
 725  *
 726  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 727  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 728  * mutual recursion.
 729  */
 730 void __pagevec_release(struct pagevec *pvec)
 731 {
 732         lru_add_drain();
 733         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 734         pagevec_reinit(pvec);
 735 }
 736 EXPORT_SYMBOL(__pagevec_release);
 737
 738 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 739 /* used by __split_huge_page_refcount() */
 740 void lru_add_page_tail(struct page *page, struct page *page_tail,
 741                        struct lruvec *lruvec, struct list_head *list)
 742 {
 743         int uninitialized_var(active);
 744         enum lru_list lru;
 745         const int file = 0;
 746
 747         VM_BUG_ON(!PageHead(page));
 748         VM_BUG_ON(PageCompound(page_tail));
 749         VM_BUG_ON(PageLRU(page_tail));
 750         VM_BUG_ON(NR_CPUS != 1 &&
 751                   !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
 752
 753         if (!list)
 754                 SetPageLRU(page_tail);
 755
 756         if (page_evictable(page_tail)) {
 757                 if (PageActive(page)) {
 758                         SetPageActive(page_tail);
 759                         active = 1;
 760                         lru = LRU_ACTIVE_ANON;
 761                 } else {
 762                         active = 0;
 763                         lru = LRU_INACTIVE_ANON;
 764                 }
 765         } else {
 766                 SetPageUnevictable(page_tail);
 767                 lru = LRU_UNEVICTABLE;
 768         }
 769
 770         if (likely(PageLRU(page)))
 771                 list_add_tail(&page_tail->lru, &page->lru);
 772         else if (list) {
 773                 /* page reclaim is reclaiming a huge page */
 774                 get_page(page_tail);
 775                 list_add_tail(&page_tail->lru, list);
 776         } else {
 777                 struct list_head *list_head;
 778                 /*
 779                  * Head page has not yet been counted, as an hpage,
 780                  * so we must account for each subpage individually.
 781                  *
 782                  * Use the standard add function to put page_tail on the list,
 783                  * but then correct its position so they all end up in order.
 784                  */
 785                 add_page_to_lru_list(page_tail, lruvec, lru);
 786                 list_head = page_tail->lru.prev;
 787                 list_move_tail(&page_tail->lru, list_head);
 788         }
 789
 790         if (!PageUnevictable(page))
 791                 update_page_reclaim_stat(lruvec, file, active);
 792 }
 793 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 794
 795 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
 796                                  void *arg)
 797 {
 798         enum lru_list requested_lru = (enum lru_list)arg;
 799         int file = page_is_file_cache(page);
 800         int active = PageActive(page);
 801         enum lru_list lru = page_lru(page);
 802
 803         WARN_ON_ONCE(requested_lru < NR_LRU_LISTS && requested_lru != lru);
 804         VM_BUG_ON(PageUnevictable(page));
 805         VM_BUG_ON(PageLRU(page));
 806
 807         SetPageLRU(page);
 808         add_page_to_lru_list(page, lruvec, lru);
 809         update_page_reclaim_stat(lruvec, file, active);
 810         trace_mm_lru_insertion(page, page_to_pfn(page), lru, trace_pagemap_flags(page));
 811 }
 812
 813 /*
 814  * Add the passed pages to the LRU, then drop the caller's refcount
 815  * on them.  Reinitialises the caller's pagevec.
 816  */
 817 void __pagevec_lru_add(struct pagevec *pvec, enum lru_list lru)
 818 {
 819         VM_BUG_ON(is_unevictable_lru(lru));
 820
 821         pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, (void *)lru);
 822 }
 823 EXPORT_SYMBOL(__pagevec_lru_add);
 824
 825 /**
 826  * pagevec_lookup - gang pagecache lookup
 827  * @pvec:       Where the resulting pages are placed
 828  * @mapping:    The address_space to search
 829  * @start:      The starting page index
 830  * @nr_pages:   The maximum number of pages
 831  *
 832  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 833  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 834  * reference against the pages in @pvec.
 835  *
 836  * The search returns a group of mapping-contiguous pages with ascending
 837  * indexes.  There may be holes in the indices due to not-present pages.
 838  *
 839  * pagevec_lookup() returns the number of pages which were found.
 840  */
 841 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 842                 pgoff_t start, unsigned nr_pages)
 843 {
 844         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 845         return pagevec_count(pvec);
 846 }
 847 EXPORT_SYMBOL(pagevec_lookup);
 848
 849 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 850                 pgoff_t *index, int tag, unsigned nr_pages)
 851 {
 852         pvec->nr = find_get_pages_tag(mapping, index, tag,
 853                                         nr_pages, pvec->pages);
 854         return pagevec_count(pvec);
 855 }
 856 EXPORT_SYMBOL(pagevec_lookup_tag);
 857
 858 /*
 859  * Perform any setup for the swap system
 860  */
 861 void __init swap_setup(void)
 862 {
 863         unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
 864 #ifdef CONFIG_SWAP
 865         int i;
 866
 867         bdi_init(swapper_spaces[0].backing_dev_info);
 868         for (i = 0; i < MAX_SWAPFILES; i++) {
 869                 spin_lock_init(&swapper_spaces[i].tree_lock);
 870                 INIT_LIST_HEAD(&swapper_spaces[i].i_mmap_nonlinear);
 871         }
 872 #endif
 873
 874         /* Use a smaller cluster for small-memory machines */
 875         if (megs < 16)
 876                 page_cluster = 2;
 877         else
 878                 page_cluster = 3;
 879         /*
 880          * Right now other parts of the system means that we
 881          * _really_ don't want to cluster much more
 882          */
 883 }