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 /* How many pages do we try to swap or page in/out together? */
  38 int page_cluster;
  39
  40 static DEFINE_PER_CPU(struct pagevec[NR_LRU_LISTS], lru_add_pvecs);
  41 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
  42 static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
  43
  44 /*
  45  * This path almost never happens for VM activity - pages are normally
  46  * freed via pagevecs.  But it gets used by networking.
  47  */
  48 static void __page_cache_release(struct page *page)
  49 {
  50         if (PageLRU(page)) {
  51                 struct zone *zone = page_zone(page);
  52                 struct lruvec *lruvec;
  53                 unsigned long flags;
  54
  55                 spin_lock_irqsave(&zone->lru_lock, flags);
  56                 lruvec = mem_cgroup_page_lruvec(page, zone);
  57                 VM_BUG_ON(!PageLRU(page));
  58                 __ClearPageLRU(page);
  59                 del_page_from_lru_list(page, lruvec, page_off_lru(page));
  60                 spin_unlock_irqrestore(&zone->lru_lock, flags);
  61         }
  62 }
  63
  64 static void __put_single_page(struct page *page)
  65 {
  66         __page_cache_release(page);
  67         free_hot_cold_page(page, 0);
  68 }
  69
  70 static void __put_compound_page(struct page *page)
  71 {
  72         compound_page_dtor *dtor;
  73
  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                         /*
  90                          * THP can not break up slab pages so avoid taking
  91                          * compound_lock().  Slab performs non-atomic bit ops
  92                          * on page->flags for better performance.  In particular
  93                          * slab_unlock() in slub used to be a hot path.  It is
  94                          * still hot on arches that do not support
  95                          * this_cpu_cmpxchg_double().
  96                          */
  97                         if (PageSlab(page_head)) {
  98                                 if (PageTail(page)) {
  99                                         if (put_page_testzero(page_head))
 100                                                 VM_BUG_ON(1);
 101
 102                                         atomic_dec(&page->_mapcount);
 103                                         goto skip_lock_tail;
 104                                 } else
 105                                         goto skip_lock;
 106                         }
 107                         /*
 108                          * page_head wasn't a dangling pointer but it
 109                          * may not be a head page anymore by the time
 110                          * we obtain the lock. That is ok as long as it
 111                          * can't be freed from under us.
 112                          */
 113                         flags = compound_lock_irqsave(page_head);
 114                         if (unlikely(!PageTail(page))) {
 115                                 /* __split_huge_page_refcount run before us */
 116                                 compound_unlock_irqrestore(page_head, flags);
 117 skip_lock:
 118                                 if (put_page_testzero(page_head))
 119                                         __put_single_page(page_head);
 120 out_put_single:
 121                                 if (put_page_testzero(page))
 122                                         __put_single_page(page);
 123                                 return;
 124                         }
 125                         VM_BUG_ON(page_head != page->first_page);
 126                         /*
 127                          * We can release the refcount taken by
 128                          * get_page_unless_zero() now that
 129                          * __split_huge_page_refcount() is blocked on
 130                          * the compound_lock.
 131                          */
 132                         if (put_page_testzero(page_head))
 133                                 VM_BUG_ON(1);
 134                         /* __split_huge_page_refcount will wait now */
 135                         VM_BUG_ON(page_mapcount(page) <= 0);
 136                         atomic_dec(&page->_mapcount);
 137                         VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
 138                         VM_BUG_ON(atomic_read(&page->_count) != 0);
 139                         compound_unlock_irqrestore(page_head, flags);
 140
 141 skip_lock_tail:
 142                         if (put_page_testzero(page_head)) {
 143                                 if (PageHead(page_head))
 144                                         __put_compound_page(page_head);
 145                                 else
 146                                         __put_single_page(page_head);
 147                         }
 148                 } else {
 149                         /* page_head is a dangling pointer */
 150                         VM_BUG_ON(PageTail(page));
 151                         goto out_put_single;
 152                 }
 153         } else if (put_page_testzero(page)) {
 154                 if (PageHead(page))
 155                         __put_compound_page(page);
 156                 else
 157                         __put_single_page(page);
 158         }
 159 }
 160
 161 void put_page(struct page *page)
 162 {
 163         if (unlikely(PageCompound(page)))
 164                 put_compound_page(page);
 165         else if (put_page_testzero(page))
 166                 __put_single_page(page);
 167 }
 168 EXPORT_SYMBOL(put_page);
 169
 170 /*
 171  * This function is exported but must not be called by anything other
 172  * than get_page(). It implements the slow path of get_page().
 173  */
 174 bool __get_page_tail(struct page *page)
 175 {
 176         /*
 177          * This takes care of get_page() if run on a tail page
 178          * returned by one of the get_user_pages/follow_page variants.
 179          * get_user_pages/follow_page itself doesn't need the compound
 180          * lock because it runs __get_page_tail_foll() under the
 181          * proper PT lock that already serializes against
 182          * split_huge_page().
 183          */
 184         unsigned long flags;
 185         bool got = false;
 186         struct page *page_head = compound_trans_head(page);
 187
 188         if (likely(page != page_head && get_page_unless_zero(page_head))) {
 189
 190                 /* Ref to put_compound_page() comment. */
 191                 if (PageSlab(page_head)) {
 192                         if (likely(PageTail(page))) {
 193                                 __get_page_tail_foll(page, false);
 194                                 return true;
 195                         } else {
 196                                 put_page(page_head);
 197                                 return false;
 198                         }
 199                 }
 200
 201                 /*
 202                  * page_head wasn't a dangling pointer but it
 203                  * may not be a head page anymore by the time
 204                  * we obtain the lock. That is ok as long as it
 205                  * can't be freed from under us.
 206                  */
 207                 flags = compound_lock_irqsave(page_head);
 208                 /* here __split_huge_page_refcount won't run anymore */
 209                 if (likely(PageTail(page))) {
 210                         __get_page_tail_foll(page, false);
 211                         got = true;
 212                 }
 213                 compound_unlock_irqrestore(page_head, flags);
 214                 if (unlikely(!got))
 215                         put_page(page_head);
 216         }
 217         return got;
 218 }
 219 EXPORT_SYMBOL(__get_page_tail);
 220
 221 /**
 222  * put_pages_list() - release a list of pages
 223  * @pages: list of pages threaded on page->lru
 224  *
 225  * Release a list of pages which are strung together on page.lru.  Currently
 226  * used by read_cache_pages() and related error recovery code.
 227  */
 228 void put_pages_list(struct list_head *pages)
 229 {
 230         while (!list_empty(pages)) {
 231                 struct page *victim;
 232
 233                 victim = list_entry(pages->prev, struct page, lru);
 234                 list_del(&victim->lru);
 235                 page_cache_release(victim);
 236         }
 237 }
 238 EXPORT_SYMBOL(put_pages_list);
 239
 240 /*
 241  * get_kernel_pages() - pin kernel pages in memory
 242  * @kiov:       An array of struct kvec structures
 243  * @nr_segs:    number of segments to pin
 244  * @write:      pinning for read/write, currently ignored
 245  * @pages:      array that receives pointers to the pages pinned.
 246  *              Should be at least nr_segs long.
 247  *
 248  * Returns number of pages pinned. This may be fewer than the number
 249  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 250  * were pinned, returns -errno. Each page returned must be released
 251  * with a put_page() call when it is finished with.
 252  */
 253 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
 254                 struct page **pages)
 255 {
 256         int seg;
 257
 258         for (seg = 0; seg < nr_segs; seg++) {
 259                 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
 260                         return seg;
 261
 262                 pages[seg] = kmap_to_page(kiov[seg].iov_base);
 263                 page_cache_get(pages[seg]);
 264         }
 265
 266         return seg;
 267 }
 268 EXPORT_SYMBOL_GPL(get_kernel_pages);
 269
 270 /*
 271  * get_kernel_page() - pin a kernel page in memory
 272  * @start:      starting kernel address
 273  * @write:      pinning for read/write, currently ignored
 274  * @pages:      array that receives pointer to the page pinned.
 275  *              Must be at least nr_segs long.
 276  *
 277  * Returns 1 if page is pinned. If the page was not pinned, returns
 278  * -errno. The page returned must be released with a put_page() call
 279  * when it is finished with.
 280  */
 281 int get_kernel_page(unsigned long start, int write, struct page **pages)
 282 {
 283         const struct kvec kiov = {
 284                 .iov_base = (void *)start,
 285                 .iov_len = PAGE_SIZE
 286         };
 287
 288         return get_kernel_pages(&kiov, 1, write, pages);
 289 }
 290 EXPORT_SYMBOL_GPL(get_kernel_page);
 291
 292 static void pagevec_lru_move_fn(struct pagevec *pvec,
 293         void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
 294         void *arg)
 295 {
 296         int i;
 297         struct zone *zone = NULL;
 298         struct lruvec *lruvec;
 299         unsigned long flags = 0;
 300
 301         for (i = 0; i < pagevec_count(pvec); i++) {
 302                 struct page *page = pvec->pages[i];
 303                 struct zone *pagezone = page_zone(page);
 304
 305                 if (pagezone != zone) {
 306                         if (zone)
 307                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 308                         zone = pagezone;
 309                         spin_lock_irqsave(&zone->lru_lock, flags);
 310                 }
 311
 312                 lruvec = mem_cgroup_page_lruvec(page, zone);
 313                 (*move_fn)(page, lruvec, arg);
 314         }
 315         if (zone)
 316                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 317         release_pages(pvec->pages, pvec->nr, pvec->cold);
 318         pagevec_reinit(pvec);
 319 }
 320
 321 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
 322                                  void *arg)
 323 {
 324         int *pgmoved = arg;
 325
 326         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 327                 enum lru_list lru = page_lru_base_type(page);
 328                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 329                 (*pgmoved)++;
 330         }
 331 }
 332
 333 /*
 334  * pagevec_move_tail() must be called with IRQ disabled.
 335  * Otherwise this may cause nasty races.
 336  */
 337 static void pagevec_move_tail(struct pagevec *pvec)
 338 {
 339         int pgmoved = 0;
 340
 341         pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
 342         __count_vm_events(PGROTATED, pgmoved);
 343 }
 344
 345 /*
 346  * Writeback is about to end against a page which has been marked for immediate
 347  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 348  * inactive list.
 349  */
 350 void rotate_reclaimable_page(struct page *page)
 351 {
 352         if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
 353             !PageUnevictable(page) && PageLRU(page)) {
 354                 struct pagevec *pvec;
 355                 unsigned long flags;
 356
 357                 page_cache_get(page);
 358                 local_irq_save(flags);
 359                 pvec = &__get_cpu_var(lru_rotate_pvecs);
 360                 if (!pagevec_add(pvec, page))
 361                         pagevec_move_tail(pvec);
 362                 local_irq_restore(flags);
 363         }
 364 }
 365
 366 static void update_page_reclaim_stat(struct lruvec *lruvec,
 367                                      int file, int rotated)
 368 {
 369         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
 370
 371         reclaim_stat->recent_scanned[file]++;
 372         if (rotated)
 373                 reclaim_stat->recent_rotated[file]++;
 374 }
 375
 376 static void __activate_page(struct page *page, struct lruvec *lruvec,
 377                             void *arg)
 378 {
 379         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 380                 int file = page_is_file_cache(page);
 381                 int lru = page_lru_base_type(page);
 382
 383                 del_page_from_lru_list(page, lruvec, lru);
 384                 SetPageActive(page);
 385                 lru += LRU_ACTIVE;
 386                 add_page_to_lru_list(page, lruvec, lru);
 387
 388                 __count_vm_event(PGACTIVATE);
 389                 update_page_reclaim_stat(lruvec, file, 1);
 390         }
 391 }
 392
 393 #ifdef CONFIG_SMP
 394 static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
 395
 396 static void activate_page_drain(int cpu)
 397 {
 398         struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
 399
 400         if (pagevec_count(pvec))
 401                 pagevec_lru_move_fn(pvec, __activate_page, NULL);
 402 }
 403
 404 void activate_page(struct page *page)
 405 {
 406         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 407                 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
 408
 409                 page_cache_get(page);
 410                 if (!pagevec_add(pvec, page))
 411                         pagevec_lru_move_fn(pvec, __activate_page, NULL);
 412                 put_cpu_var(activate_page_pvecs);
 413         }
 414 }
 415
 416 #else
 417 static inline void activate_page_drain(int cpu)
 418 {
 419 }
 420
 421 void activate_page(struct page *page)
 422 {
 423         struct zone *zone = page_zone(page);
 424
 425         spin_lock_irq(&zone->lru_lock);
 426         __activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
 427         spin_unlock_irq(&zone->lru_lock);
 428 }
 429 #endif
 430
 431 /*
 432  * Mark a page as having seen activity.
 433  *
 434  * inactive,unreferenced        ->      inactive,referenced
 435  * inactive,referenced          ->      active,unreferenced
 436  * active,unreferenced          ->      active,referenced
 437  */
 438 void mark_page_accessed(struct page *page)
 439 {
 440         if (!PageActive(page) && !PageUnevictable(page) &&
 441                         PageReferenced(page) && PageLRU(page)) {
 442                 activate_page(page);
 443                 ClearPageReferenced(page);
 444         } else if (!PageReferenced(page)) {
 445                 SetPageReferenced(page);
 446         }
 447 }
 448 EXPORT_SYMBOL(mark_page_accessed);
 449
 450 /*
 451  * Order of operations is important: flush the pagevec when it's already
 452  * full, not when adding the last page, to make sure that last page is
 453  * not added to the LRU directly when passed to this function. Because
 454  * mark_page_accessed() (called after this when writing) only activates
 455  * pages that are on the LRU, linear writes in subpage chunks would see
 456  * every PAGEVEC_SIZE page activated, which is unexpected.
 457  */
 458 void __lru_cache_add(struct page *page, enum lru_list lru)
 459 {
 460         struct pagevec *pvec = &get_cpu_var(lru_add_pvecs)[lru];
 461
 462         page_cache_get(page);
 463         if (!pagevec_space(pvec))
 464                 __pagevec_lru_add(pvec, lru);
 465         pagevec_add(pvec, page);
 466         put_cpu_var(lru_add_pvecs);
 467 }
 468 EXPORT_SYMBOL(__lru_cache_add);
 469
 470 /**
 471  * lru_cache_add_lru - add a page to a page list
 472  * @page: the page to be added to the LRU.
 473  * @lru: the LRU list to which the page is added.
 474  */
 475 void lru_cache_add_lru(struct page *page, enum lru_list lru)
 476 {
 477         if (PageActive(page)) {
 478                 VM_BUG_ON(PageUnevictable(page));
 479                 ClearPageActive(page);
 480         } else if (PageUnevictable(page)) {
 481                 VM_BUG_ON(PageActive(page));
 482                 ClearPageUnevictable(page);
 483         }
 484
 485         VM_BUG_ON(PageLRU(page) || PageActive(page) || PageUnevictable(page));
 486         __lru_cache_add(page, lru);
 487 }
 488
 489 /**
 490  * add_page_to_unevictable_list - add a page to the unevictable list
 491  * @page:  the page to be added to the unevictable list
 492  *
 493  * Add page directly to its zone's unevictable list.  To avoid races with
 494  * tasks that might be making the page evictable, through eg. munlock,
 495  * munmap or exit, while it's not on the lru, we want to add the page
 496  * while it's locked or otherwise "invisible" to other tasks.  This is
 497  * difficult to do when using the pagevec cache, so bypass that.
 498  */
 499 void add_page_to_unevictable_list(struct page *page)
 500 {
 501         struct zone *zone = page_zone(page);
 502         struct lruvec *lruvec;
 503
 504         spin_lock_irq(&zone->lru_lock);
 505         lruvec = mem_cgroup_page_lruvec(page, zone);
 506         SetPageUnevictable(page);
 507         SetPageLRU(page);
 508         add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
 509         spin_unlock_irq(&zone->lru_lock);
 510 }
 511
 512 /*
 513  * If the page can not be invalidated, it is moved to the
 514  * inactive list to speed up its reclaim.  It is moved to the
 515  * head of the list, rather than the tail, to give the flusher
 516  * threads some time to write it out, as this is much more
 517  * effective than the single-page writeout from reclaim.
 518  *
 519  * If the page isn't page_mapped and dirty/writeback, the page
 520  * could reclaim asap using PG_reclaim.
 521  *
 522  * 1. active, mapped page -> none
 523  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 524  * 3. inactive, mapped page -> none
 525  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 526  * 5. inactive, clean -> inactive, tail
 527  * 6. Others -> none
 528  *
 529  * In 4, why it moves inactive's head, the VM expects the page would
 530  * be write it out by flusher threads as this is much more effective
 531  * than the single-page writeout from reclaim.
 532  */
 533 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 534                               void *arg)
 535 {
 536         int lru, file;
 537         bool active;
 538
 539         if (!PageLRU(page))
 540                 return;
 541
 542         if (PageUnevictable(page))
 543                 return;
 544
 545         /* Some processes are using the page */
 546         if (page_mapped(page))
 547                 return;
 548
 549         active = PageActive(page);
 550         file = page_is_file_cache(page);
 551         lru = page_lru_base_type(page);
 552
 553         del_page_from_lru_list(page, lruvec, lru + active);
 554         ClearPageActive(page);
 555         ClearPageReferenced(page);
 556         add_page_to_lru_list(page, lruvec, lru);
 557
 558         if (PageWriteback(page) || PageDirty(page)) {
 559                 /*
 560                  * PG_reclaim could be raced with end_page_writeback
 561                  * It can make readahead confusing.  But race window
 562                  * is _really_ small and  it's non-critical problem.
 563                  */
 564                 SetPageReclaim(page);
 565         } else {
 566                 /*
 567                  * The page's writeback ends up during pagevec
 568                  * We moves tha page into tail of inactive.
 569                  */
 570                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 571                 __count_vm_event(PGROTATED);
 572         }
 573
 574         if (active)
 575                 __count_vm_event(PGDEACTIVATE);
 576         update_page_reclaim_stat(lruvec, file, 0);
 577 }
 578
 579 /*
 580  * Drain pages out of the cpu's pagevecs.
 581  * Either "cpu" is the current CPU, and preemption has already been
 582  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 583  */
 584 void lru_add_drain_cpu(int cpu)
 585 {
 586         struct pagevec *pvecs = per_cpu(lru_add_pvecs, cpu);
 587         struct pagevec *pvec;
 588         int lru;
 589
 590         for_each_lru(lru) {
 591                 pvec = &pvecs[lru - LRU_BASE];
 592                 if (pagevec_count(pvec))
 593                         __pagevec_lru_add(pvec, lru);
 594         }
 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 lru = (enum lru_list)arg;
 799         int file = is_file_lru(lru);
 800         int active = is_active_lru(lru);
 801
 802         VM_BUG_ON(PageActive(page));
 803         VM_BUG_ON(PageUnevictable(page));
 804         VM_BUG_ON(PageLRU(page));
 805
 806         SetPageLRU(page);
 807         if (active)
 808                 SetPageActive(page);
 809         add_page_to_lru_list(page, lruvec, lru);
 810         update_page_reclaim_stat(lruvec, file, active);
 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 }