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 /* How many pages do we try to swap or page in/out together? */
  35 int page_cluster;
  36
  37 static DEFINE_PER_CPU(struct pagevec, lru_add_pvecs) = { 0, };
  38 static DEFINE_PER_CPU(struct pagevec, lru_add_active_pvecs) = { 0, };
  39 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs) = { 0, };
  40
  41 /*
  42  * This path almost never happens for VM activity - pages are normally
  43  * freed via pagevecs.  But it gets used by networking.
  44  */
  45 static void __page_cache_release(struct page *page)
  46 {
  47         if (PageLRU(page)) {
  48                 unsigned long flags;
  49                 struct zone *zone = page_zone(page);
  50
  51                 spin_lock_irqsave(&zone->lru_lock, flags);
  52                 VM_BUG_ON(!PageLRU(page));
  53                 __ClearPageLRU(page);
  54                 del_page_from_lru(zone, page);
  55                 spin_unlock_irqrestore(&zone->lru_lock, flags);
  56         }
  57         free_hot_page(page);
  58 }
  59
  60 static void put_compound_page(struct page *page)
  61 {
  62         page = compound_head(page);
  63         if (put_page_testzero(page)) {
  64                 compound_page_dtor *dtor;
  65
  66                 dtor = get_compound_page_dtor(page);
  67                 (*dtor)(page);
  68         }
  69 }
  70
  71 void put_page(struct page *page)
  72 {
  73         if (unlikely(PageCompound(page)))
  74                 put_compound_page(page);
  75         else if (put_page_testzero(page))
  76                 __page_cache_release(page);
  77 }
  78 EXPORT_SYMBOL(put_page);
  79
  80 /**
  81  * put_pages_list() - release a list of pages
  82  * @pages: list of pages threaded on page->lru
  83  *
  84  * Release a list of pages which are strung together on page.lru.  Currently
  85  * used by read_cache_pages() and related error recovery code.
  86  */
  87 void put_pages_list(struct list_head *pages)
  88 {
  89         while (!list_empty(pages)) {
  90                 struct page *victim;
  91
  92                 victim = list_entry(pages->prev, struct page, lru);
  93                 list_del(&victim->lru);
  94                 page_cache_release(victim);
  95         }
  96 }
  97 EXPORT_SYMBOL(put_pages_list);
  98
  99 /*
 100  * pagevec_move_tail() must be called with IRQ disabled.
 101  * Otherwise this may cause nasty races.
 102  */
 103 static void pagevec_move_tail(struct pagevec *pvec)
 104 {
 105         int i;
 106         int pgmoved = 0;
 107         struct zone *zone = NULL;
 108
 109         for (i = 0; i < pagevec_count(pvec); i++) {
 110                 struct page *page = pvec->pages[i];
 111                 struct zone *pagezone = page_zone(page);
 112
 113                 if (pagezone != zone) {
 114                         if (zone)
 115                                 spin_unlock(&zone->lru_lock);
 116                         zone = pagezone;
 117                         spin_lock(&zone->lru_lock);
 118                 }
 119                 if (PageLRU(page) && !PageActive(page)) {
 120                         list_move_tail(&page->lru, &zone->inactive_list);
 121                         pgmoved++;
 122                 }
 123         }
 124         if (zone)
 125                 spin_unlock(&zone->lru_lock);
 126         __count_vm_events(PGROTATED, pgmoved);
 127         release_pages(pvec->pages, pvec->nr, pvec->cold);
 128         pagevec_reinit(pvec);
 129 }
 130
 131 /*
 132  * Writeback is about to end against a page which has been marked for immediate
 133  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 134  * inactive list.
 135  *
 136  * Returns zero if it cleared PG_writeback.
 137  */
 138 int rotate_reclaimable_page(struct page *page)
 139 {
 140         struct pagevec *pvec;
 141         unsigned long flags;
 142
 143         if (PageLocked(page))
 144                 return 1;
 145         if (PageDirty(page))
 146                 return 1;
 147         if (PageActive(page))
 148                 return 1;
 149         if (!PageLRU(page))
 150                 return 1;
 151
 152         page_cache_get(page);
 153         local_irq_save(flags);
 154         pvec = &__get_cpu_var(lru_rotate_pvecs);
 155         if (!pagevec_add(pvec, page))
 156                 pagevec_move_tail(pvec);
 157         local_irq_restore(flags);
 158
 159         if (!test_clear_page_writeback(page))
 160                 BUG();
 161
 162         return 0;
 163 }
 164
 165 /*
 166  * FIXME: speed this up?
 167  */
 168 void activate_page(struct page *page)
 169 {
 170         struct zone *zone = page_zone(page);
 171
 172         spin_lock_irq(&zone->lru_lock);
 173         if (PageLRU(page) && !PageActive(page)) {
 174                 del_page_from_inactive_list(zone, page);
 175                 SetPageActive(page);
 176                 add_page_to_active_list(zone, page);
 177                 __count_vm_event(PGACTIVATE);
 178                 mem_cgroup_move_lists(page, true);
 179         }
 180         spin_unlock_irq(&zone->lru_lock);
 181 }
 182
 183 /*
 184  * Mark a page as having seen activity.
 185  *
 186  * inactive,unreferenced        ->      inactive,referenced
 187  * inactive,referenced          ->      active,unreferenced
 188  * active,unreferenced          ->      active,referenced
 189  */
 190 void mark_page_accessed(struct page *page)
 191 {
 192         if (!PageActive(page) && PageReferenced(page) && PageLRU(page)) {
 193                 activate_page(page);
 194                 ClearPageReferenced(page);
 195         } else if (!PageReferenced(page)) {
 196                 SetPageReferenced(page);
 197         }
 198 }
 199
 200 EXPORT_SYMBOL(mark_page_accessed);
 201
 202 /**
 203  * lru_cache_add: add a page to the page lists
 204  * @page: the page to add
 205  */
 206 void lru_cache_add(struct page *page)
 207 {
 208         struct pagevec *pvec = &get_cpu_var(lru_add_pvecs);
 209
 210         page_cache_get(page);
 211         if (!pagevec_add(pvec, page))
 212                 __pagevec_lru_add(pvec);
 213         put_cpu_var(lru_add_pvecs);
 214 }
 215
 216 void lru_cache_add_active(struct page *page)
 217 {
 218         struct pagevec *pvec = &get_cpu_var(lru_add_active_pvecs);
 219
 220         page_cache_get(page);
 221         if (!pagevec_add(pvec, page))
 222                 __pagevec_lru_add_active(pvec);
 223         put_cpu_var(lru_add_active_pvecs);
 224 }
 225
 226 /*
 227  * Drain pages out of the cpu's pagevecs.
 228  * Either "cpu" is the current CPU, and preemption has already been
 229  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 230  */
 231 static void drain_cpu_pagevecs(int cpu)
 232 {
 233         struct pagevec *pvec;
 234
 235         pvec = &per_cpu(lru_add_pvecs, cpu);
 236         if (pagevec_count(pvec))
 237                 __pagevec_lru_add(pvec);
 238
 239         pvec = &per_cpu(lru_add_active_pvecs, cpu);
 240         if (pagevec_count(pvec))
 241                 __pagevec_lru_add_active(pvec);
 242
 243         pvec = &per_cpu(lru_rotate_pvecs, cpu);
 244         if (pagevec_count(pvec)) {
 245                 unsigned long flags;
 246
 247                 /* No harm done if a racing interrupt already did this */
 248                 local_irq_save(flags);
 249                 pagevec_move_tail(pvec);
 250                 local_irq_restore(flags);
 251         }
 252 }
 253
 254 void lru_add_drain(void)
 255 {
 256         drain_cpu_pagevecs(get_cpu());
 257         put_cpu();
 258 }
 259
 260 #ifdef CONFIG_NUMA
 261 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 262 {
 263         lru_add_drain();
 264 }
 265
 266 /*
 267  * Returns 0 for success
 268  */
 269 int lru_add_drain_all(void)
 270 {
 271         return schedule_on_each_cpu(lru_add_drain_per_cpu);
 272 }
 273
 274 #else
 275
 276 /*
 277  * Returns 0 for success
 278  */
 279 int lru_add_drain_all(void)
 280 {
 281         lru_add_drain();
 282         return 0;
 283 }
 284 #endif
 285
 286 /*
 287  * Batched page_cache_release().  Decrement the reference count on all the
 288  * passed pages.  If it fell to zero then remove the page from the LRU and
 289  * free it.
 290  *
 291  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 292  * for the remainder of the operation.
 293  *
 294  * The locking in this function is against shrink_cache(): we recheck the
 295  * page count inside the lock to see whether shrink_cache grabbed the page
 296  * via the LRU.  If it did, give up: shrink_cache will free it.
 297  */
 298 void release_pages(struct page **pages, int nr, int cold)
 299 {
 300         int i;
 301         struct pagevec pages_to_free;
 302         struct zone *zone = NULL;
 303         unsigned long uninitialized_var(flags);
 304
 305         pagevec_init(&pages_to_free, cold);
 306         for (i = 0; i < nr; i++) {
 307                 struct page *page = pages[i];
 308
 309                 if (unlikely(PageCompound(page))) {
 310                         if (zone) {
 311                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 312                                 zone = NULL;
 313                         }
 314                         put_compound_page(page);
 315                         continue;
 316                 }
 317
 318                 if (!put_page_testzero(page))
 319                         continue;
 320
 321                 if (PageLRU(page)) {
 322                         struct zone *pagezone = page_zone(page);
 323                         if (pagezone != zone) {
 324                                 if (zone)
 325                                         spin_unlock_irqrestore(&zone->lru_lock,
 326                                                                         flags);
 327                                 zone = pagezone;
 328                                 spin_lock_irqsave(&zone->lru_lock, flags);
 329                         }
 330                         VM_BUG_ON(!PageLRU(page));
 331                         __ClearPageLRU(page);
 332                         del_page_from_lru(zone, page);
 333                 }
 334
 335                 if (!pagevec_add(&pages_to_free, page)) {
 336                         if (zone) {
 337                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 338                                 zone = NULL;
 339                         }
 340                         __pagevec_free(&pages_to_free);
 341                         pagevec_reinit(&pages_to_free);
 342                 }
 343         }
 344         if (zone)
 345                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 346
 347         pagevec_free(&pages_to_free);
 348 }
 349
 350 /*
 351  * The pages which we're about to release may be in the deferred lru-addition
 352  * queues.  That would prevent them from really being freed right now.  That's
 353  * OK from a correctness point of view but is inefficient - those pages may be
 354  * cache-warm and we want to give them back to the page allocator ASAP.
 355  *
 356  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 357  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 358  * mutual recursion.
 359  */
 360 void __pagevec_release(struct pagevec *pvec)
 361 {
 362         lru_add_drain();
 363         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 364         pagevec_reinit(pvec);
 365 }
 366
 367 EXPORT_SYMBOL(__pagevec_release);
 368
 369 /*
 370  * pagevec_release() for pages which are known to not be on the LRU
 371  *
 372  * This function reinitialises the caller's pagevec.
 373  */
 374 void __pagevec_release_nonlru(struct pagevec *pvec)
 375 {
 376         int i;
 377         struct pagevec pages_to_free;
 378
 379         pagevec_init(&pages_to_free, pvec->cold);
 380         for (i = 0; i < pagevec_count(pvec); i++) {
 381                 struct page *page = pvec->pages[i];
 382
 383                 VM_BUG_ON(PageLRU(page));
 384                 if (put_page_testzero(page))
 385                         pagevec_add(&pages_to_free, page);
 386         }
 387         pagevec_free(&pages_to_free);
 388         pagevec_reinit(pvec);
 389 }
 390
 391 /*
 392  * Add the passed pages to the LRU, then drop the caller's refcount
 393  * on them.  Reinitialises the caller's pagevec.
 394  */
 395 void __pagevec_lru_add(struct pagevec *pvec)
 396 {
 397         int i;
 398         struct zone *zone = NULL;
 399
 400         for (i = 0; i < pagevec_count(pvec); i++) {
 401                 struct page *page = pvec->pages[i];
 402                 struct zone *pagezone = page_zone(page);
 403
 404                 if (pagezone != zone) {
 405                         if (zone)
 406                                 spin_unlock_irq(&zone->lru_lock);
 407                         zone = pagezone;
 408                         spin_lock_irq(&zone->lru_lock);
 409                 }
 410                 VM_BUG_ON(PageLRU(page));
 411                 SetPageLRU(page);
 412                 add_page_to_inactive_list(zone, page);
 413         }
 414         if (zone)
 415                 spin_unlock_irq(&zone->lru_lock);
 416         release_pages(pvec->pages, pvec->nr, pvec->cold);
 417         pagevec_reinit(pvec);
 418 }
 419
 420 EXPORT_SYMBOL(__pagevec_lru_add);
 421
 422 void __pagevec_lru_add_active(struct pagevec *pvec)
 423 {
 424         int i;
 425         struct zone *zone = NULL;
 426
 427         for (i = 0; i < pagevec_count(pvec); i++) {
 428                 struct page *page = pvec->pages[i];
 429                 struct zone *pagezone = page_zone(page);
 430
 431                 if (pagezone != zone) {
 432                         if (zone)
 433                                 spin_unlock_irq(&zone->lru_lock);
 434                         zone = pagezone;
 435                         spin_lock_irq(&zone->lru_lock);
 436                 }
 437                 VM_BUG_ON(PageLRU(page));
 438                 SetPageLRU(page);
 439                 VM_BUG_ON(PageActive(page));
 440                 SetPageActive(page);
 441                 add_page_to_active_list(zone, page);
 442         }
 443         if (zone)
 444                 spin_unlock_irq(&zone->lru_lock);
 445         release_pages(pvec->pages, pvec->nr, pvec->cold);
 446         pagevec_reinit(pvec);
 447 }
 448
 449 /*
 450  * Try to drop buffers from the pages in a pagevec
 451  */
 452 void pagevec_strip(struct pagevec *pvec)
 453 {
 454         int i;
 455
 456         for (i = 0; i < pagevec_count(pvec); i++) {
 457                 struct page *page = pvec->pages[i];
 458
 459                 if (PagePrivate(page) && !TestSetPageLocked(page)) {
 460                         if (PagePrivate(page))
 461                                 try_to_release_page(page, 0);
 462                         unlock_page(page);
 463                 }
 464         }
 465 }
 466
 467 /**
 468  * pagevec_lookup - gang pagecache lookup
 469  * @pvec:       Where the resulting pages are placed
 470  * @mapping:    The address_space to search
 471  * @start:      The starting page index
 472  * @nr_pages:   The maximum number of pages
 473  *
 474  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 475  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 476  * reference against the pages in @pvec.
 477  *
 478  * The search returns a group of mapping-contiguous pages with ascending
 479  * indexes.  There may be holes in the indices due to not-present pages.
 480  *
 481  * pagevec_lookup() returns the number of pages which were found.
 482  */
 483 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 484                 pgoff_t start, unsigned nr_pages)
 485 {
 486         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 487         return pagevec_count(pvec);
 488 }
 489
 490 EXPORT_SYMBOL(pagevec_lookup);
 491
 492 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 493                 pgoff_t *index, int tag, unsigned nr_pages)
 494 {
 495         pvec->nr = find_get_pages_tag(mapping, index, tag,
 496                                         nr_pages, pvec->pages);
 497         return pagevec_count(pvec);
 498 }
 499
 500 EXPORT_SYMBOL(pagevec_lookup_tag);
 501
 502 #ifdef CONFIG_SMP
 503 /*
 504  * We tolerate a little inaccuracy to avoid ping-ponging the counter between
 505  * CPUs
 506  */
 507 #define ACCT_THRESHOLD  max(16, NR_CPUS * 2)
 508
 509 static DEFINE_PER_CPU(long, committed_space) = 0;
 510
 511 void vm_acct_memory(long pages)
 512 {
 513         long *local;
 514
 515         preempt_disable();
 516         local = &__get_cpu_var(committed_space);
 517         *local += pages;
 518         if (*local > ACCT_THRESHOLD || *local < -ACCT_THRESHOLD) {
 519                 atomic_add(*local, &vm_committed_space);
 520                 *local = 0;
 521         }
 522         preempt_enable();
 523 }
 524
 525 #ifdef CONFIG_HOTPLUG_CPU
 526
 527 /* Drop the CPU's cached committed space back into the central pool. */
 528 static int cpu_swap_callback(struct notifier_block *nfb,
 529                              unsigned long action,
 530                              void *hcpu)
 531 {
 532         long *committed;
 533
 534         committed = &per_cpu(committed_space, (long)hcpu);
 535         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
 536                 atomic_add(*committed, &vm_committed_space);
 537                 *committed = 0;
 538                 drain_cpu_pagevecs((long)hcpu);
 539         }
 540         return NOTIFY_OK;
 541 }
 542 #endif /* CONFIG_HOTPLUG_CPU */
 543 #endif /* CONFIG_SMP */
 544
 545 /*
 546  * Perform any setup for the swap system
 547  */
 548 void __init swap_setup(void)
 549 {
 550         unsigned long megs = num_physpages >> (20 - PAGE_SHIFT);
 551
 552 #ifdef CONFIG_SWAP
 553         bdi_init(swapper_space.backing_dev_info);
 554 #endif
 555
 556         /* Use a smaller cluster for small-memory machines */
 557         if (megs < 16)
 558                 page_cluster = 2;
 559         else
 560                 page_cluster = 3;
 561         /*
 562          * Right now other parts of the system means that we
 563          * _really_ don't want to cluster much more
 564          */
 565 #ifdef CONFIG_HOTPLUG_CPU
 566         hotcpu_notifier(cpu_swap_callback, 0);
 567 #endif
 568 }