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