mm/swap.c

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