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
  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
  42 /*
  43  * This path almost never happens for VM activity - pages are normally
  44  * freed via pagevecs.  But it gets used by networking.
  45  */
  46 static void __page_cache_release(struct page *page)
  47 {
  48         if (PageLRU(page)) {
  49                 unsigned long flags;
  50                 struct zone *zone = page_zone(page);
  51
  52                 spin_lock_irqsave(&zone->lru_lock, flags);
  53                 VM_BUG_ON(!PageLRU(page));
  54                 __ClearPageLRU(page);
  55                 del_page_from_lru(zone, page);
  56                 spin_unlock_irqrestore(&zone->lru_lock, flags);
  57         }
  58         free_hot_page(page);
  59 }
  60
  61 static void put_compound_page(struct page *page)
  62 {
  63         page = compound_head(page);
  64         if (put_page_testzero(page)) {
  65                 compound_page_dtor *dtor;
  66
  67                 dtor = get_compound_page_dtor(page);
  68                 (*dtor)(page);
  69         }
  70 }
  71
  72 void put_page(struct page *page)
  73 {
  74         if (unlikely(PageCompound(page)))
  75                 put_compound_page(page);
  76         else if (put_page_testzero(page))
  77                 __page_cache_release(page);
  78 }
  79 EXPORT_SYMBOL(put_page);
  80
  81 /**
  82  * put_pages_list() - release a list of pages
  83  * @pages: list of pages threaded on page->lru
  84  *
  85  * Release a list of pages which are strung together on page.lru.  Currently
  86  * used by read_cache_pages() and related error recovery code.
  87  */
  88 void put_pages_list(struct list_head *pages)
  89 {
  90         while (!list_empty(pages)) {
  91                 struct page *victim;
  92
  93                 victim = list_entry(pages->prev, struct page, lru);
  94                 list_del(&victim->lru);
  95                 page_cache_release(victim);
  96         }
  97 }
  98 EXPORT_SYMBOL(put_pages_list);
  99
 100 /*
 101  * pagevec_move_tail() must be called with IRQ disabled.
 102  * Otherwise this may cause nasty races.
 103  */
 104 static void pagevec_move_tail(struct pagevec *pvec)
 105 {
 106         int i;
 107         int pgmoved = 0;
 108         struct zone *zone = NULL;
 109
 110         for (i = 0; i < pagevec_count(pvec); i++) {
 111                 struct page *page = pvec->pages[i];
 112                 struct zone *pagezone = page_zone(page);
 113
 114                 if (pagezone != zone) {
 115                         if (zone)
 116                                 spin_unlock(&zone->lru_lock);
 117                         zone = pagezone;
 118                         spin_lock(&zone->lru_lock);
 119                 }
 120                 if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 121                         int lru = page_is_file_cache(page);
 122                         list_move_tail(&page->lru, &zone->lru[lru].list);
 123                         pgmoved++;
 124                 }
 125         }
 126         if (zone)
 127                 spin_unlock(&zone->lru_lock);
 128         __count_vm_events(PGROTATED, pgmoved);
 129         release_pages(pvec->pages, pvec->nr, pvec->cold);
 130         pagevec_reinit(pvec);
 131 }
 132
 133 /*
 134  * Writeback is about to end against a page which has been marked for immediate
 135  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 136  * inactive list.
 137  */
 138 void  rotate_reclaimable_page(struct page *page)
 139 {
 140         if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
 141             !PageUnevictable(page) && PageLRU(page)) {
 142                 struct pagevec *pvec;
 143                 unsigned long flags;
 144
 145                 page_cache_get(page);
 146                 local_irq_save(flags);
 147                 pvec = &__get_cpu_var(lru_rotate_pvecs);
 148                 if (!pagevec_add(pvec, page))
 149                         pagevec_move_tail(pvec);
 150                 local_irq_restore(flags);
 151         }
 152 }
 153
 154 /*
 155  * FIXME: speed this up?
 156  */
 157 void activate_page(struct page *page)
 158 {
 159         struct zone *zone = page_zone(page);
 160
 161         spin_lock_irq(&zone->lru_lock);
 162         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 163                 int file = page_is_file_cache(page);
 164                 int lru = LRU_BASE + file;
 165                 del_page_from_lru_list(zone, page, lru);
 166
 167                 SetPageActive(page);
 168                 lru += LRU_ACTIVE;
 169                 add_page_to_lru_list(zone, page, lru);
 170                 __count_vm_event(PGACTIVATE);
 171                 mem_cgroup_move_lists(page, lru);
 172
 173                 zone->recent_rotated[!!file]++;
 174                 zone->recent_scanned[!!file]++;
 175         }
 176         spin_unlock_irq(&zone->lru_lock);
 177 }
 178
 179 /*
 180  * Mark a page as having seen activity.
 181  *
 182  * inactive,unreferenced        ->      inactive,referenced
 183  * inactive,referenced          ->      active,unreferenced
 184  * active,unreferenced          ->      active,referenced
 185  */
 186 void mark_page_accessed(struct page *page)
 187 {
 188         if (!PageActive(page) && !PageUnevictable(page) &&
 189                         PageReferenced(page) && PageLRU(page)) {
 190                 activate_page(page);
 191                 ClearPageReferenced(page);
 192         } else if (!PageReferenced(page)) {
 193                 SetPageReferenced(page);
 194         }
 195 }
 196
 197 EXPORT_SYMBOL(mark_page_accessed);
 198
 199 void __lru_cache_add(struct page *page, enum lru_list lru)
 200 {
 201         struct pagevec *pvec = &get_cpu_var(lru_add_pvecs)[lru];
 202
 203         page_cache_get(page);
 204         if (!pagevec_add(pvec, page))
 205                 ____pagevec_lru_add(pvec, lru);
 206         put_cpu_var(lru_add_pvecs);
 207 }
 208
 209 /**
 210  * lru_cache_add_lru - add a page to a page list
 211  * @page: the page to be added to the LRU.
 212  * @lru: the LRU list to which the page is added.
 213  */
 214 void lru_cache_add_lru(struct page *page, enum lru_list lru)
 215 {
 216         if (PageActive(page)) {
 217                 VM_BUG_ON(PageUnevictable(page));
 218                 ClearPageActive(page);
 219         } else if (PageUnevictable(page)) {
 220                 VM_BUG_ON(PageActive(page));
 221                 ClearPageUnevictable(page);
 222         }
 223
 224         VM_BUG_ON(PageLRU(page) || PageActive(page) || PageUnevictable(page));
 225         __lru_cache_add(page, lru);
 226 }
 227
 228 /**
 229  * add_page_to_unevictable_list - add a page to the unevictable list
 230  * @page:  the page to be added to the unevictable list
 231  *
 232  * Add page directly to its zone's unevictable list.  To avoid races with
 233  * tasks that might be making the page evictable, through eg. munlock,
 234  * munmap or exit, while it's not on the lru, we want to add the page
 235  * while it's locked or otherwise "invisible" to other tasks.  This is
 236  * difficult to do when using the pagevec cache, so bypass that.
 237  */
 238 void add_page_to_unevictable_list(struct page *page)
 239 {
 240         struct zone *zone = page_zone(page);
 241
 242         spin_lock_irq(&zone->lru_lock);
 243         SetPageUnevictable(page);
 244         SetPageLRU(page);
 245         add_page_to_lru_list(zone, page, LRU_UNEVICTABLE);
 246         spin_unlock_irq(&zone->lru_lock);
 247 }
 248
 249 /**
 250  * lru_cache_add_active_or_unevictable
 251  * @page:  the page to be added to LRU
 252  * @vma:   vma in which page is mapped for determining reclaimability
 253  *
 254  * place @page on active or unevictable LRU list, depending on
 255  * page_evictable().  Note that if the page is not evictable,
 256  * it goes directly back onto it's zone's unevictable list.  It does
 257  * NOT use a per cpu pagevec.
 258  */
 259 void lru_cache_add_active_or_unevictable(struct page *page,
 260                                         struct vm_area_struct *vma)
 261 {
 262         if (page_evictable(page, vma))
 263                 lru_cache_add_lru(page, LRU_ACTIVE + page_is_file_cache(page));
 264         else
 265                 add_page_to_unevictable_list(page);
 266 }
 267
 268 /*
 269  * Drain pages out of the cpu's pagevecs.
 270  * Either "cpu" is the current CPU, and preemption has already been
 271  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 272  */
 273 static void drain_cpu_pagevecs(int cpu)
 274 {
 275         struct pagevec *pvecs = per_cpu(lru_add_pvecs, cpu);
 276         struct pagevec *pvec;
 277         int lru;
 278
 279         for_each_lru(lru) {
 280                 pvec = &pvecs[lru - LRU_BASE];
 281                 if (pagevec_count(pvec))
 282                         ____pagevec_lru_add(pvec, lru);
 283         }
 284
 285         pvec = &per_cpu(lru_rotate_pvecs, cpu);
 286         if (pagevec_count(pvec)) {
 287                 unsigned long flags;
 288
 289                 /* No harm done if a racing interrupt already did this */
 290                 local_irq_save(flags);
 291                 pagevec_move_tail(pvec);
 292                 local_irq_restore(flags);
 293         }
 294 }
 295
 296 void lru_add_drain(void)
 297 {
 298         drain_cpu_pagevecs(get_cpu());
 299         put_cpu();
 300 }
 301
 302 #if defined(CONFIG_NUMA) || defined(CONFIG_UNEVICTABLE_LRU)
 303 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 304 {
 305         lru_add_drain();
 306 }
 307
 308 /*
 309  * Returns 0 for success
 310  */
 311 int lru_add_drain_all(void)
 312 {
 313         return schedule_on_each_cpu(lru_add_drain_per_cpu);
 314 }
 315
 316 #else
 317
 318 /*
 319  * Returns 0 for success
 320  */
 321 int lru_add_drain_all(void)
 322 {
 323         lru_add_drain();
 324         return 0;
 325 }
 326 #endif
 327
 328 /*
 329  * Batched page_cache_release().  Decrement the reference count on all the
 330  * passed pages.  If it fell to zero then remove the page from the LRU and
 331  * free it.
 332  *
 333  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 334  * for the remainder of the operation.
 335  *
 336  * The locking in this function is against shrink_inactive_list(): we recheck
 337  * the page count inside the lock to see whether shrink_inactive_list()
 338  * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
 339  * will free it.
 340  */
 341 void release_pages(struct page **pages, int nr, int cold)
 342 {
 343         int i;
 344         struct pagevec pages_to_free;
 345         struct zone *zone = NULL;
 346         unsigned long uninitialized_var(flags);
 347
 348         pagevec_init(&pages_to_free, cold);
 349         for (i = 0; i < nr; i++) {
 350                 struct page *page = pages[i];
 351
 352                 if (unlikely(PageCompound(page))) {
 353                         if (zone) {
 354                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 355                                 zone = NULL;
 356                         }
 357                         put_compound_page(page);
 358                         continue;
 359                 }
 360
 361                 if (!put_page_testzero(page))
 362                         continue;
 363
 364                 if (PageLRU(page)) {
 365                         struct zone *pagezone = page_zone(page);
 366
 367                         if (pagezone != zone) {
 368                                 if (zone)
 369                                         spin_unlock_irqrestore(&zone->lru_lock,
 370                                                                         flags);
 371                                 zone = pagezone;
 372                                 spin_lock_irqsave(&zone->lru_lock, flags);
 373                         }
 374                         VM_BUG_ON(!PageLRU(page));
 375                         __ClearPageLRU(page);
 376                         del_page_from_lru(zone, page);
 377                 }
 378
 379                 if (!pagevec_add(&pages_to_free, page)) {
 380                         if (zone) {
 381                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 382                                 zone = NULL;
 383                         }
 384                         __pagevec_free(&pages_to_free);
 385                         pagevec_reinit(&pages_to_free);
 386                 }
 387         }
 388         if (zone)
 389                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 390
 391         pagevec_free(&pages_to_free);
 392 }
 393
 394 /*
 395  * The pages which we're about to release may be in the deferred lru-addition
 396  * queues.  That would prevent them from really being freed right now.  That's
 397  * OK from a correctness point of view but is inefficient - those pages may be
 398  * cache-warm and we want to give them back to the page allocator ASAP.
 399  *
 400  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 401  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 402  * mutual recursion.
 403  */
 404 void __pagevec_release(struct pagevec *pvec)
 405 {
 406         lru_add_drain();
 407         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 408         pagevec_reinit(pvec);
 409 }
 410
 411 EXPORT_SYMBOL(__pagevec_release);
 412
 413 /*
 414  * pagevec_release() for pages which are known to not be on the LRU
 415  *
 416  * This function reinitialises the caller's pagevec.
 417  */
 418 void __pagevec_release_nonlru(struct pagevec *pvec)
 419 {
 420         int i;
 421         struct pagevec pages_to_free;
 422
 423         pagevec_init(&pages_to_free, pvec->cold);
 424         for (i = 0; i < pagevec_count(pvec); i++) {
 425                 struct page *page = pvec->pages[i];
 426
 427                 VM_BUG_ON(PageLRU(page));
 428                 if (put_page_testzero(page))
 429                         pagevec_add(&pages_to_free, page);
 430         }
 431         pagevec_free(&pages_to_free);
 432         pagevec_reinit(pvec);
 433 }
 434
 435 /*
 436  * Add the passed pages to the LRU, then drop the caller's refcount
 437  * on them.  Reinitialises the caller's pagevec.
 438  */
 439 void ____pagevec_lru_add(struct pagevec *pvec, enum lru_list lru)
 440 {
 441         int i;
 442         struct zone *zone = NULL;
 443         VM_BUG_ON(is_unevictable_lru(lru));
 444
 445         for (i = 0; i < pagevec_count(pvec); i++) {
 446                 struct page *page = pvec->pages[i];
 447                 struct zone *pagezone = page_zone(page);
 448
 449                 if (pagezone != zone) {
 450                         if (zone)
 451                                 spin_unlock_irq(&zone->lru_lock);
 452                         zone = pagezone;
 453                         spin_lock_irq(&zone->lru_lock);
 454                 }
 455                 VM_BUG_ON(PageActive(page));
 456                 VM_BUG_ON(PageUnevictable(page));
 457                 VM_BUG_ON(PageLRU(page));
 458                 SetPageLRU(page);
 459                 if (is_active_lru(lru))
 460                         SetPageActive(page);
 461                 add_page_to_lru_list(zone, page, lru);
 462         }
 463         if (zone)
 464                 spin_unlock_irq(&zone->lru_lock);
 465         release_pages(pvec->pages, pvec->nr, pvec->cold);
 466         pagevec_reinit(pvec);
 467 }
 468
 469 EXPORT_SYMBOL(____pagevec_lru_add);
 470
 471 /*
 472  * Try to drop buffers from the pages in a pagevec
 473  */
 474 void pagevec_strip(struct pagevec *pvec)
 475 {
 476         int i;
 477
 478         for (i = 0; i < pagevec_count(pvec); i++) {
 479                 struct page *page = pvec->pages[i];
 480
 481                 if (PagePrivate(page) && trylock_page(page)) {
 482                         if (PagePrivate(page))
 483                                 try_to_release_page(page, 0);
 484                         unlock_page(page);
 485                 }
 486         }
 487 }
 488
 489 /**
 490  * pagevec_swap_free - try to free swap space from the pages in a pagevec
 491  * @pvec: pagevec with swapcache pages to free the swap space of
 492  *
 493  * The caller needs to hold an extra reference to each page and
 494  * not hold the page lock on the pages.  This function uses a
 495  * trylock on the page lock so it may not always free the swap
 496  * space associated with a page.
 497  */
 498 void pagevec_swap_free(struct pagevec *pvec)
 499 {
 500         int i;
 501
 502         for (i = 0; i < pagevec_count(pvec); i++) {
 503                 struct page *page = pvec->pages[i];
 504
 505                 if (PageSwapCache(page) && trylock_page(page)) {
 506                         if (PageSwapCache(page))
 507                                 remove_exclusive_swap_page_ref(page);
 508                         unlock_page(page);
 509                 }
 510         }
 511 }
 512
 513 /**
 514  * pagevec_lookup - gang pagecache lookup
 515  * @pvec:       Where the resulting pages are placed
 516  * @mapping:    The address_space to search
 517  * @start:      The starting page index
 518  * @nr_pages:   The maximum number of pages
 519  *
 520  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 521  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 522  * reference against the pages in @pvec.
 523  *
 524  * The search returns a group of mapping-contiguous pages with ascending
 525  * indexes.  There may be holes in the indices due to not-present pages.
 526  *
 527  * pagevec_lookup() returns the number of pages which were found.
 528  */
 529 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 530                 pgoff_t start, unsigned nr_pages)
 531 {
 532         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 533         return pagevec_count(pvec);
 534 }
 535
 536 EXPORT_SYMBOL(pagevec_lookup);
 537
 538 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 539                 pgoff_t *index, int tag, unsigned nr_pages)
 540 {
 541         pvec->nr = find_get_pages_tag(mapping, index, tag,
 542                                         nr_pages, pvec->pages);
 543         return pagevec_count(pvec);
 544 }
 545
 546 EXPORT_SYMBOL(pagevec_lookup_tag);
 547
 548 #ifdef CONFIG_SMP
 549 /*
 550  * We tolerate a little inaccuracy to avoid ping-ponging the counter between
 551  * CPUs
 552  */
 553 #define ACCT_THRESHOLD  max(16, NR_CPUS * 2)
 554
 555 static DEFINE_PER_CPU(long, committed_space);
 556
 557 void vm_acct_memory(long pages)
 558 {
 559         long *local;
 560
 561         preempt_disable();
 562         local = &__get_cpu_var(committed_space);
 563         *local += pages;
 564         if (*local > ACCT_THRESHOLD || *local < -ACCT_THRESHOLD) {
 565                 atomic_long_add(*local, &vm_committed_space);
 566                 *local = 0;
 567         }
 568         preempt_enable();
 569 }
 570
 571 #ifdef CONFIG_HOTPLUG_CPU
 572
 573 /* Drop the CPU's cached committed space back into the central pool. */
 574 static int cpu_swap_callback(struct notifier_block *nfb,
 575                              unsigned long action,
 576                              void *hcpu)
 577 {
 578         long *committed;
 579
 580         committed = &per_cpu(committed_space, (long)hcpu);
 581         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
 582                 atomic_long_add(*committed, &vm_committed_space);
 583                 *committed = 0;
 584                 drain_cpu_pagevecs((long)hcpu);
 585         }
 586         return NOTIFY_OK;
 587 }
 588 #endif /* CONFIG_HOTPLUG_CPU */
 589 #endif /* CONFIG_SMP */
 590
 591 /*
 592  * Perform any setup for the swap system
 593  */
 594 void __init swap_setup(void)
 595 {
 596         unsigned long megs = num_physpages >> (20 - PAGE_SHIFT);
 597
 598 #ifdef CONFIG_SWAP
 599         bdi_init(swapper_space.backing_dev_info);
 600 #endif
 601
 602         /* Use a smaller cluster for small-memory machines */
 603         if (megs < 16)
 604                 page_cluster = 2;
 605         else
 606                 page_cluster = 3;
 607         /*
 608          * Right now other parts of the system means that we
 609          * _really_ don't want to cluster much more
 610          */
 611 #ifdef CONFIG_HOTPLUG_CPU
 612         hotcpu_notifier(cpu_swap_callback, 0);
 613 #endif
 614 }