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