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/module.h>
  25 #include <linux/mm_inline.h>
  26 #include <linux/buffer_head.h>  /* for try_to_release_page() */
  27 #include <linux/percpu_counter.h>
  28 #include <linux/percpu.h>
  29 #include <linux/cpu.h>
  30 #include <linux/notifier.h>
  31 #include <linux/backing-dev.h>
  32 #include <linux/memcontrol.h>
  33 #include <linux/gfp.h>
  34
  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
  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(zone, 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 = page->first_page;
  81                 smp_rmb();
  82                 /*
  83                  * If PageTail is still set after smp_rmb() we can be sure
  84                  * that the page->first_page we read wasn't a dangling pointer.
  85                  * See __split_huge_page_refcount() smp_wmb().
  86                  */
  87                 if (likely(PageTail(page) && get_page_unless_zero(page_head))) {
  88                         unsigned long flags;
  89                         /*
  90                          * Verify that our page_head wasn't converted
  91                          * to a a regular page before we got a
  92                          * reference on it.
  93                          */
  94                         if (unlikely(!PageHead(page_head))) {
  95                                 /* PageHead is cleared after PageTail */
  96                                 smp_rmb();
  97                                 VM_BUG_ON(PageTail(page));
  98                                 goto out_put_head;
  99                         }
 100                         /*
 101                          * Only run compound_lock on a valid PageHead,
 102                          * after having it pinned with
 103                          * get_page_unless_zero() above.
 104                          */
 105                         smp_mb();
 106                         /* page_head wasn't a dangling pointer */
 107                         flags = compound_lock_irqsave(page_head);
 108                         if (unlikely(!PageTail(page))) {
 109                                 /* __split_huge_page_refcount run before us */
 110                                 compound_unlock_irqrestore(page_head, flags);
 111                                 VM_BUG_ON(PageHead(page_head));
 112                         out_put_head:
 113                                 if (put_page_testzero(page_head))
 114                                         __put_single_page(page_head);
 115                         out_put_single:
 116                                 if (put_page_testzero(page))
 117                                         __put_single_page(page);
 118                                 return;
 119                         }
 120                         VM_BUG_ON(page_head != page->first_page);
 121                         /*
 122                          * We can release the refcount taken by
 123                          * get_page_unless_zero now that
 124                          * split_huge_page_refcount is blocked on the
 125                          * compound_lock.
 126                          */
 127                         if (put_page_testzero(page_head))
 128                                 VM_BUG_ON(1);
 129                         /* __split_huge_page_refcount will wait now */
 130                         VM_BUG_ON(atomic_read(&page->_count) <= 0);
 131                         atomic_dec(&page->_count);
 132                         VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
 133                         compound_unlock_irqrestore(page_head, flags);
 134                         if (put_page_testzero(page_head)) {
 135                                 if (PageHead(page_head))
 136                                         __put_compound_page(page_head);
 137                                 else
 138                                         __put_single_page(page_head);
 139                         }
 140                 } else {
 141                         /* page_head is a dangling pointer */
 142                         VM_BUG_ON(PageTail(page));
 143                         goto out_put_single;
 144                 }
 145         } else if (put_page_testzero(page)) {
 146                 if (PageHead(page))
 147                         __put_compound_page(page);
 148                 else
 149                         __put_single_page(page);
 150         }
 151 }
 152
 153 void put_page(struct page *page)
 154 {
 155         if (unlikely(PageCompound(page)))
 156                 put_compound_page(page);
 157         else if (put_page_testzero(page))
 158                 __put_single_page(page);
 159 }
 160 EXPORT_SYMBOL(put_page);
 161
 162 /**
 163  * put_pages_list() - release a list of pages
 164  * @pages: list of pages threaded on page->lru
 165  *
 166  * Release a list of pages which are strung together on page.lru.  Currently
 167  * used by read_cache_pages() and related error recovery code.
 168  */
 169 void put_pages_list(struct list_head *pages)
 170 {
 171         while (!list_empty(pages)) {
 172                 struct page *victim;
 173
 174                 victim = list_entry(pages->prev, struct page, lru);
 175                 list_del(&victim->lru);
 176                 page_cache_release(victim);
 177         }
 178 }
 179 EXPORT_SYMBOL(put_pages_list);
 180
 181 static void pagevec_lru_move_fn(struct pagevec *pvec,
 182                                 void (*move_fn)(struct page *page, void *arg),
 183                                 void *arg)
 184 {
 185         int i;
 186         struct zone *zone = NULL;
 187         unsigned long flags = 0;
 188
 189         for (i = 0; i < pagevec_count(pvec); i++) {
 190                 struct page *page = pvec->pages[i];
 191                 struct zone *pagezone = page_zone(page);
 192
 193                 if (pagezone != zone) {
 194                         if (zone)
 195                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 196                         zone = pagezone;
 197                         spin_lock_irqsave(&zone->lru_lock, flags);
 198                 }
 199
 200                 (*move_fn)(page, arg);
 201         }
 202         if (zone)
 203                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 204         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 205         pagevec_reinit(pvec);
 206 }
 207
 208 static void pagevec_move_tail_fn(struct page *page, void *arg)
 209 {
 210         int *pgmoved = arg;
 211         struct zone *zone = page_zone(page);
 212
 213         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 214                 int lru = page_lru_base_type(page);
 215                 list_move_tail(&page->lru, &zone->lru[lru].list);
 216                 (*pgmoved)++;
 217         }
 218 }
 219
 220 /*
 221  * pagevec_move_tail() must be called with IRQ disabled.
 222  * Otherwise this may cause nasty races.
 223  */
 224 static void pagevec_move_tail(struct pagevec *pvec)
 225 {
 226         int pgmoved = 0;
 227
 228         pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
 229         __count_vm_events(PGROTATED, pgmoved);
 230 }
 231
 232 /*
 233  * Writeback is about to end against a page which has been marked for immediate
 234  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 235  * inactive list.
 236  */
 237 void rotate_reclaimable_page(struct page *page)
 238 {
 239         if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
 240             !PageUnevictable(page) && PageLRU(page)) {
 241                 struct pagevec *pvec;
 242                 unsigned long flags;
 243
 244                 page_cache_get(page);
 245                 local_irq_save(flags);
 246                 pvec = &__get_cpu_var(lru_rotate_pvecs);
 247                 if (!pagevec_add(pvec, page))
 248                         pagevec_move_tail(pvec);
 249                 local_irq_restore(flags);
 250         }
 251 }
 252
 253 static void update_page_reclaim_stat(struct zone *zone, struct page *page,
 254                                      int file, int rotated)
 255 {
 256         struct zone_reclaim_stat *reclaim_stat = &zone->reclaim_stat;
 257         struct zone_reclaim_stat *memcg_reclaim_stat;
 258
 259         memcg_reclaim_stat = mem_cgroup_get_reclaim_stat_from_page(page);
 260
 261         reclaim_stat->recent_scanned[file]++;
 262         if (rotated)
 263                 reclaim_stat->recent_rotated[file]++;
 264
 265         if (!memcg_reclaim_stat)
 266                 return;
 267
 268         memcg_reclaim_stat->recent_scanned[file]++;
 269         if (rotated)
 270                 memcg_reclaim_stat->recent_rotated[file]++;
 271 }
 272
 273 /*
 274  * FIXME: speed this up?
 275  */
 276 void activate_page(struct page *page)
 277 {
 278         struct zone *zone = page_zone(page);
 279
 280         spin_lock_irq(&zone->lru_lock);
 281         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 282                 int file = page_is_file_cache(page);
 283                 int lru = page_lru_base_type(page);
 284                 del_page_from_lru_list(zone, page, lru);
 285
 286                 SetPageActive(page);
 287                 lru += LRU_ACTIVE;
 288                 add_page_to_lru_list(zone, page, lru);
 289                 __count_vm_event(PGACTIVATE);
 290
 291                 update_page_reclaim_stat(zone, page, file, 1);
 292         }
 293         spin_unlock_irq(&zone->lru_lock);
 294 }
 295
 296 /*
 297  * Mark a page as having seen activity.
 298  *
 299  * inactive,unreferenced        ->      inactive,referenced
 300  * inactive,referenced          ->      active,unreferenced
 301  * active,unreferenced          ->      active,referenced
 302  */
 303 void mark_page_accessed(struct page *page)
 304 {
 305         if (!PageActive(page) && !PageUnevictable(page) &&
 306                         PageReferenced(page) && PageLRU(page)) {
 307                 activate_page(page);
 308                 ClearPageReferenced(page);
 309         } else if (!PageReferenced(page)) {
 310                 SetPageReferenced(page);
 311         }
 312 }
 313
 314 EXPORT_SYMBOL(mark_page_accessed);
 315
 316 void __lru_cache_add(struct page *page, enum lru_list lru)
 317 {
 318         struct pagevec *pvec = &get_cpu_var(lru_add_pvecs)[lru];
 319
 320         page_cache_get(page);
 321         if (!pagevec_add(pvec, page))
 322                 ____pagevec_lru_add(pvec, lru);
 323         put_cpu_var(lru_add_pvecs);
 324 }
 325 EXPORT_SYMBOL(__lru_cache_add);
 326
 327 /**
 328  * lru_cache_add_lru - add a page to a page list
 329  * @page: the page to be added to the LRU.
 330  * @lru: the LRU list to which the page is added.
 331  */
 332 void lru_cache_add_lru(struct page *page, enum lru_list lru)
 333 {
 334         if (PageActive(page)) {
 335                 VM_BUG_ON(PageUnevictable(page));
 336                 ClearPageActive(page);
 337         } else if (PageUnevictable(page)) {
 338                 VM_BUG_ON(PageActive(page));
 339                 ClearPageUnevictable(page);
 340         }
 341
 342         VM_BUG_ON(PageLRU(page) || PageActive(page) || PageUnevictable(page));
 343         __lru_cache_add(page, lru);
 344 }
 345
 346 /**
 347  * add_page_to_unevictable_list - add a page to the unevictable list
 348  * @page:  the page to be added to the unevictable list
 349  *
 350  * Add page directly to its zone's unevictable list.  To avoid races with
 351  * tasks that might be making the page evictable, through eg. munlock,
 352  * munmap or exit, while it's not on the lru, we want to add the page
 353  * while it's locked or otherwise "invisible" to other tasks.  This is
 354  * difficult to do when using the pagevec cache, so bypass that.
 355  */
 356 void add_page_to_unevictable_list(struct page *page)
 357 {
 358         struct zone *zone = page_zone(page);
 359
 360         spin_lock_irq(&zone->lru_lock);
 361         SetPageUnevictable(page);
 362         SetPageLRU(page);
 363         add_page_to_lru_list(zone, page, LRU_UNEVICTABLE);
 364         spin_unlock_irq(&zone->lru_lock);
 365 }
 366
 367 /*
 368  * Drain pages out of the cpu's pagevecs.
 369  * Either "cpu" is the current CPU, and preemption has already been
 370  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 371  */
 372 static void drain_cpu_pagevecs(int cpu)
 373 {
 374         struct pagevec *pvecs = per_cpu(lru_add_pvecs, cpu);
 375         struct pagevec *pvec;
 376         int lru;
 377
 378         for_each_lru(lru) {
 379                 pvec = &pvecs[lru - LRU_BASE];
 380                 if (pagevec_count(pvec))
 381                         ____pagevec_lru_add(pvec, lru);
 382         }
 383
 384         pvec = &per_cpu(lru_rotate_pvecs, cpu);
 385         if (pagevec_count(pvec)) {
 386                 unsigned long flags;
 387
 388                 /* No harm done if a racing interrupt already did this */
 389                 local_irq_save(flags);
 390                 pagevec_move_tail(pvec);
 391                 local_irq_restore(flags);
 392         }
 393 }
 394
 395 void lru_add_drain(void)
 396 {
 397         drain_cpu_pagevecs(get_cpu());
 398         put_cpu();
 399 }
 400
 401 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 402 {
 403         lru_add_drain();
 404 }
 405
 406 /*
 407  * Returns 0 for success
 408  */
 409 int lru_add_drain_all(void)
 410 {
 411         return schedule_on_each_cpu(lru_add_drain_per_cpu);
 412 }
 413
 414 /*
 415  * Batched page_cache_release().  Decrement the reference count on all the
 416  * passed pages.  If it fell to zero then remove the page from the LRU and
 417  * free it.
 418  *
 419  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 420  * for the remainder of the operation.
 421  *
 422  * The locking in this function is against shrink_inactive_list(): we recheck
 423  * the page count inside the lock to see whether shrink_inactive_list()
 424  * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
 425  * will free it.
 426  */
 427 void release_pages(struct page **pages, int nr, int cold)
 428 {
 429         int i;
 430         struct pagevec pages_to_free;
 431         struct zone *zone = NULL;
 432         unsigned long uninitialized_var(flags);
 433
 434         pagevec_init(&pages_to_free, cold);
 435         for (i = 0; i < nr; i++) {
 436                 struct page *page = pages[i];
 437
 438                 if (unlikely(PageCompound(page))) {
 439                         if (zone) {
 440                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 441                                 zone = NULL;
 442                         }
 443                         put_compound_page(page);
 444                         continue;
 445                 }
 446
 447                 if (!put_page_testzero(page))
 448                         continue;
 449
 450                 if (PageLRU(page)) {
 451                         struct zone *pagezone = page_zone(page);
 452
 453                         if (pagezone != zone) {
 454                                 if (zone)
 455                                         spin_unlock_irqrestore(&zone->lru_lock,
 456                                                                         flags);
 457                                 zone = pagezone;
 458                                 spin_lock_irqsave(&zone->lru_lock, flags);
 459                         }
 460                         VM_BUG_ON(!PageLRU(page));
 461                         __ClearPageLRU(page);
 462                         del_page_from_lru(zone, page);
 463                 }
 464
 465                 if (!pagevec_add(&pages_to_free, page)) {
 466                         if (zone) {
 467                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 468                                 zone = NULL;
 469                         }
 470                         __pagevec_free(&pages_to_free);
 471                         pagevec_reinit(&pages_to_free);
 472                 }
 473         }
 474         if (zone)
 475                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 476
 477         pagevec_free(&pages_to_free);
 478 }
 479 EXPORT_SYMBOL(release_pages);
 480
 481 /*
 482  * The pages which we're about to release may be in the deferred lru-addition
 483  * queues.  That would prevent them from really being freed right now.  That's
 484  * OK from a correctness point of view but is inefficient - those pages may be
 485  * cache-warm and we want to give them back to the page allocator ASAP.
 486  *
 487  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 488  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 489  * mutual recursion.
 490  */
 491 void __pagevec_release(struct pagevec *pvec)
 492 {
 493         lru_add_drain();
 494         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 495         pagevec_reinit(pvec);
 496 }
 497
 498 EXPORT_SYMBOL(__pagevec_release);
 499
 500 /* used by __split_huge_page_refcount() */
 501 void lru_add_page_tail(struct zone* zone,
 502                        struct page *page, struct page *page_tail)
 503 {
 504         int active;
 505         enum lru_list lru;
 506         const int file = 0;
 507         struct list_head *head;
 508
 509         VM_BUG_ON(!PageHead(page));
 510         VM_BUG_ON(PageCompound(page_tail));
 511         VM_BUG_ON(PageLRU(page_tail));
 512         VM_BUG_ON(!spin_is_locked(&zone->lru_lock));
 513
 514         SetPageLRU(page_tail);
 515
 516         if (page_evictable(page_tail, NULL)) {
 517                 if (PageActive(page)) {
 518                         SetPageActive(page_tail);
 519                         active = 1;
 520                         lru = LRU_ACTIVE_ANON;
 521                 } else {
 522                         active = 0;
 523                         lru = LRU_INACTIVE_ANON;
 524                 }
 525                 update_page_reclaim_stat(zone, page_tail, file, active);
 526                 if (likely(PageLRU(page)))
 527                         head = page->lru.prev;
 528                 else
 529                         head = &zone->lru[lru].list;
 530                 __add_page_to_lru_list(zone, page_tail, lru, head);
 531         } else {
 532                 SetPageUnevictable(page_tail);
 533                 add_page_to_lru_list(zone, page_tail, LRU_UNEVICTABLE);
 534         }
 535 }
 536
 537 static void ____pagevec_lru_add_fn(struct page *page, void *arg)
 538 {
 539         enum lru_list lru = (enum lru_list)arg;
 540         struct zone *zone = page_zone(page);
 541         int file = is_file_lru(lru);
 542         int active = is_active_lru(lru);
 543
 544         VM_BUG_ON(PageActive(page));
 545         VM_BUG_ON(PageUnevictable(page));
 546         VM_BUG_ON(PageLRU(page));
 547
 548         SetPageLRU(page);
 549         if (active)
 550                 SetPageActive(page);
 551         update_page_reclaim_stat(zone, page, file, active);
 552         add_page_to_lru_list(zone, page, lru);
 553 }
 554
 555 /*
 556  * Add the passed pages to the LRU, then drop the caller's refcount
 557  * on them.  Reinitialises the caller's pagevec.
 558  */
 559 void ____pagevec_lru_add(struct pagevec *pvec, enum lru_list lru)
 560 {
 561         VM_BUG_ON(is_unevictable_lru(lru));
 562
 563         pagevec_lru_move_fn(pvec, ____pagevec_lru_add_fn, (void *)lru);
 564 }
 565
 566 EXPORT_SYMBOL(____pagevec_lru_add);
 567
 568 /*
 569  * Try to drop buffers from the pages in a pagevec
 570  */
 571 void pagevec_strip(struct pagevec *pvec)
 572 {
 573         int i;
 574
 575         for (i = 0; i < pagevec_count(pvec); i++) {
 576                 struct page *page = pvec->pages[i];
 577
 578                 if (page_has_private(page) && trylock_page(page)) {
 579                         if (page_has_private(page))
 580                                 try_to_release_page(page, 0);
 581                         unlock_page(page);
 582                 }
 583         }
 584 }
 585
 586 /**
 587  * pagevec_lookup - gang pagecache lookup
 588  * @pvec:       Where the resulting pages are placed
 589  * @mapping:    The address_space to search
 590  * @start:      The starting page index
 591  * @nr_pages:   The maximum number of pages
 592  *
 593  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 594  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 595  * reference against the pages in @pvec.
 596  *
 597  * The search returns a group of mapping-contiguous pages with ascending
 598  * indexes.  There may be holes in the indices due to not-present pages.
 599  *
 600  * pagevec_lookup() returns the number of pages which were found.
 601  */
 602 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 603                 pgoff_t start, unsigned nr_pages)
 604 {
 605         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 606         return pagevec_count(pvec);
 607 }
 608
 609 EXPORT_SYMBOL(pagevec_lookup);
 610
 611 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 612                 pgoff_t *index, int tag, unsigned nr_pages)
 613 {
 614         pvec->nr = find_get_pages_tag(mapping, index, tag,
 615                                         nr_pages, pvec->pages);
 616         return pagevec_count(pvec);
 617 }
 618
 619 EXPORT_SYMBOL(pagevec_lookup_tag);
 620
 621 /*
 622  * Perform any setup for the swap system
 623  */
 624 void __init swap_setup(void)
 625 {
 626         unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
 627
 628 #ifdef CONFIG_SWAP
 629         bdi_init(swapper_space.backing_dev_info);
 630 #endif
 631
 632         /* Use a smaller cluster for small-memory machines */
 633         if (megs < 16)
 634                 page_cluster = 2;
 635         else
 636                 page_cluster = 3;
 637         /*
 638          * Right now other parts of the system means that we
 639          * _really_ don't want to cluster much more
 640          */
 641 }