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[NR_LRU_LISTS], lru_add_pvecs);
  38 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
  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 __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  * @pages: list of pages threaded on page->lru
  82  *
  83  * Release a list of pages which are strung together on page.lru.  Currently
  84  * used by read_cache_pages() and related error recovery code.
  85  */
  86 void put_pages_list(struct list_head *pages)
  87 {
  88         while (!list_empty(pages)) {
  89                 struct page *victim;
  90
  91                 victim = list_entry(pages->prev, struct page, lru);
  92                 list_del(&victim->lru);
  93                 page_cache_release(victim);
  94         }
  95 }
  96 EXPORT_SYMBOL(put_pages_list);
  97
  98 /*
  99  * pagevec_move_tail() must be called with IRQ disabled.
 100  * Otherwise this may cause nasty races.
 101  */
 102 static void pagevec_move_tail(struct pagevec *pvec)
 103 {
 104         int i;
 105         int pgmoved = 0;
 106         struct zone *zone = NULL;
 107
 108         for (i = 0; i < pagevec_count(pvec); i++) {
 109                 struct page *page = pvec->pages[i];
 110                 struct zone *pagezone = page_zone(page);
 111
 112                 if (pagezone != zone) {
 113                         if (zone)
 114                                 spin_unlock(&zone->lru_lock);
 115                         zone = pagezone;
 116                         spin_lock(&zone->lru_lock);
 117                 }
 118                 if (PageLRU(page) && !PageActive(page)) {
 119                         list_move_tail(&page->lru, &zone->lru[LRU_INACTIVE].list);
 120                         pgmoved++;
 121                 }
 122         }
 123         if (zone)
 124                 spin_unlock(&zone->lru_lock);
 125         __count_vm_events(PGROTATED, pgmoved);
 126         release_pages(pvec->pages, pvec->nr, pvec->cold);
 127         pagevec_reinit(pvec);
 128 }
 129
 130 /*
 131  * Writeback is about to end against a page which has been marked for immediate
 132  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 133  * inactive list.
 134  */
 135 void  rotate_reclaimable_page(struct page *page)
 136 {
 137         if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
 138             PageLRU(page)) {
 139                 struct pagevec *pvec;
 140                 unsigned long flags;
 141
 142                 page_cache_get(page);
 143                 local_irq_save(flags);
 144                 pvec = &__get_cpu_var(lru_rotate_pvecs);
 145                 if (!pagevec_add(pvec, page))
 146                         pagevec_move_tail(pvec);
 147                 local_irq_restore(flags);
 148         }
 149 }
 150
 151 /*
 152  * FIXME: speed this up?
 153  */
 154 void activate_page(struct page *page)
 155 {
 156         struct zone *zone = page_zone(page);
 157
 158         spin_lock_irq(&zone->lru_lock);
 159         if (PageLRU(page) && !PageActive(page)) {
 160                 del_page_from_inactive_list(zone, page);
 161                 SetPageActive(page);
 162                 add_page_to_active_list(zone, page);
 163                 __count_vm_event(PGACTIVATE);
 164                 mem_cgroup_move_lists(page, true);
 165         }
 166         spin_unlock_irq(&zone->lru_lock);
 167 }
 168
 169 /*
 170  * Mark a page as having seen activity.
 171  *
 172  * inactive,unreferenced        ->      inactive,referenced
 173  * inactive,referenced          ->      active,unreferenced
 174  * active,unreferenced          ->      active,referenced
 175  */
 176 void mark_page_accessed(struct page *page)
 177 {
 178         if (!PageActive(page) && PageReferenced(page) && PageLRU(page)) {
 179                 activate_page(page);
 180                 ClearPageReferenced(page);
 181         } else if (!PageReferenced(page)) {
 182                 SetPageReferenced(page);
 183         }
 184 }
 185
 186 EXPORT_SYMBOL(mark_page_accessed);
 187
 188 void __lru_cache_add(struct page *page, enum lru_list lru)
 189 {
 190         struct pagevec *pvec = &get_cpu_var(lru_add_pvecs)[lru];
 191
 192         page_cache_get(page);
 193         if (!pagevec_add(pvec, page))
 194                 ____pagevec_lru_add(pvec, lru);
 195         put_cpu_var(lru_add_pvecs);
 196 }
 197
 198 /**
 199  * lru_cache_add_lru - add a page to a page list
 200  * @page: the page to be added to the LRU.
 201  * @lru: the LRU list to which the page is added.
 202  */
 203 void lru_cache_add_lru(struct page *page, enum lru_list lru)
 204 {
 205         if (PageActive(page)) {
 206                 ClearPageActive(page);
 207         }
 208
 209         VM_BUG_ON(PageLRU(page) || PageActive(page));
 210         __lru_cache_add(page, lru);
 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 *pvecs = per_cpu(lru_add_pvecs, cpu);
 221         struct pagevec *pvec;
 222         int lru;
 223
 224         for_each_lru(lru) {
 225                 pvec = &pvecs[lru - LRU_BASE];
 226                 if (pagevec_count(pvec))
 227                         ____pagevec_lru_add(pvec, lru);
 228         }
 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_inactive_list(): we recheck
 282  * the page count inside the lock to see whether shrink_inactive_list()
 283  * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
 284  * will free it.
 285  */
 286 void release_pages(struct page **pages, int nr, int cold)
 287 {
 288         int i;
 289         struct pagevec pages_to_free;
 290         struct zone *zone = NULL;
 291         unsigned long uninitialized_var(flags);
 292
 293         pagevec_init(&pages_to_free, cold);
 294         for (i = 0; i < nr; i++) {
 295                 struct page *page = pages[i];
 296
 297                 if (unlikely(PageCompound(page))) {
 298                         if (zone) {
 299                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 300                                 zone = NULL;
 301                         }
 302                         put_compound_page(page);
 303                         continue;
 304                 }
 305
 306                 if (!put_page_testzero(page))
 307                         continue;
 308
 309                 if (PageLRU(page)) {
 310                         struct zone *pagezone = page_zone(page);
 311                         if (pagezone != zone) {
 312                                 if (zone)
 313                                         spin_unlock_irqrestore(&zone->lru_lock,
 314                                                                         flags);
 315                                 zone = pagezone;
 316                                 spin_lock_irqsave(&zone->lru_lock, flags);
 317                         }
 318                         VM_BUG_ON(!PageLRU(page));
 319                         __ClearPageLRU(page);
 320                         del_page_from_lru(zone, page);
 321                 }
 322
 323                 if (!pagevec_add(&pages_to_free, page)) {
 324                         if (zone) {
 325                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 326                                 zone = NULL;
 327                         }
 328                         __pagevec_free(&pages_to_free);
 329                         pagevec_reinit(&pages_to_free);
 330                 }
 331         }
 332         if (zone)
 333                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 334
 335         pagevec_free(&pages_to_free);
 336 }
 337
 338 /*
 339  * The pages which we're about to release may be in the deferred lru-addition
 340  * queues.  That would prevent them from really being freed right now.  That's
 341  * OK from a correctness point of view but is inefficient - those pages may be
 342  * cache-warm and we want to give them back to the page allocator ASAP.
 343  *
 344  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 345  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 346  * mutual recursion.
 347  */
 348 void __pagevec_release(struct pagevec *pvec)
 349 {
 350         lru_add_drain();
 351         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 352         pagevec_reinit(pvec);
 353 }
 354
 355 EXPORT_SYMBOL(__pagevec_release);
 356
 357 /*
 358  * pagevec_release() for pages which are known to not be on the LRU
 359  *
 360  * This function reinitialises the caller's pagevec.
 361  */
 362 void __pagevec_release_nonlru(struct pagevec *pvec)
 363 {
 364         int i;
 365         struct pagevec pages_to_free;
 366
 367         pagevec_init(&pages_to_free, pvec->cold);
 368         for (i = 0; i < pagevec_count(pvec); i++) {
 369                 struct page *page = pvec->pages[i];
 370
 371                 VM_BUG_ON(PageLRU(page));
 372                 if (put_page_testzero(page))
 373                         pagevec_add(&pages_to_free, page);
 374         }
 375         pagevec_free(&pages_to_free);
 376         pagevec_reinit(pvec);
 377 }
 378
 379 /*
 380  * Add the passed pages to the LRU, then drop the caller's refcount
 381  * on them.  Reinitialises the caller's pagevec.
 382  */
 383 void ____pagevec_lru_add(struct pagevec *pvec, enum lru_list lru)
 384 {
 385         int i;
 386         struct zone *zone = NULL;
 387
 388         for (i = 0; i < pagevec_count(pvec); i++) {
 389                 struct page *page = pvec->pages[i];
 390                 struct zone *pagezone = page_zone(page);
 391
 392                 if (pagezone != zone) {
 393                         if (zone)
 394                                 spin_unlock_irq(&zone->lru_lock);
 395                         zone = pagezone;
 396                         spin_lock_irq(&zone->lru_lock);
 397                 }
 398                 VM_BUG_ON(PageLRU(page));
 399                 SetPageLRU(page);
 400                 if (is_active_lru(lru))
 401                         SetPageActive(page);
 402                 add_page_to_lru_list(zone, page, lru);
 403         }
 404         if (zone)
 405                 spin_unlock_irq(&zone->lru_lock);
 406         release_pages(pvec->pages, pvec->nr, pvec->cold);
 407         pagevec_reinit(pvec);
 408 }
 409
 410 EXPORT_SYMBOL(____pagevec_lru_add);
 411
 412 /*
 413  * Try to drop buffers from the pages in a pagevec
 414  */
 415 void pagevec_strip(struct pagevec *pvec)
 416 {
 417         int i;
 418
 419         for (i = 0; i < pagevec_count(pvec); i++) {
 420                 struct page *page = pvec->pages[i];
 421
 422                 if (PagePrivate(page) && trylock_page(page)) {
 423                         if (PagePrivate(page))
 424                                 try_to_release_page(page, 0);
 425                         unlock_page(page);
 426                 }
 427         }
 428 }
 429
 430 /**
 431  * pagevec_lookup - gang pagecache lookup
 432  * @pvec:       Where the resulting pages are placed
 433  * @mapping:    The address_space to search
 434  * @start:      The starting page index
 435  * @nr_pages:   The maximum number of pages
 436  *
 437  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 438  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 439  * reference against the pages in @pvec.
 440  *
 441  * The search returns a group of mapping-contiguous pages with ascending
 442  * indexes.  There may be holes in the indices due to not-present pages.
 443  *
 444  * pagevec_lookup() returns the number of pages which were found.
 445  */
 446 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 447                 pgoff_t start, unsigned nr_pages)
 448 {
 449         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 450         return pagevec_count(pvec);
 451 }
 452
 453 EXPORT_SYMBOL(pagevec_lookup);
 454
 455 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 456                 pgoff_t *index, int tag, unsigned nr_pages)
 457 {
 458         pvec->nr = find_get_pages_tag(mapping, index, tag,
 459                                         nr_pages, pvec->pages);
 460         return pagevec_count(pvec);
 461 }
 462
 463 EXPORT_SYMBOL(pagevec_lookup_tag);
 464
 465 #ifdef CONFIG_SMP
 466 /*
 467  * We tolerate a little inaccuracy to avoid ping-ponging the counter between
 468  * CPUs
 469  */
 470 #define ACCT_THRESHOLD  max(16, NR_CPUS * 2)
 471
 472 static DEFINE_PER_CPU(long, committed_space);
 473
 474 void vm_acct_memory(long pages)
 475 {
 476         long *local;
 477
 478         preempt_disable();
 479         local = &__get_cpu_var(committed_space);
 480         *local += pages;
 481         if (*local > ACCT_THRESHOLD || *local < -ACCT_THRESHOLD) {
 482                 atomic_long_add(*local, &vm_committed_space);
 483                 *local = 0;
 484         }
 485         preempt_enable();
 486 }
 487
 488 #ifdef CONFIG_HOTPLUG_CPU
 489
 490 /* Drop the CPU's cached committed space back into the central pool. */
 491 static int cpu_swap_callback(struct notifier_block *nfb,
 492                              unsigned long action,
 493                              void *hcpu)
 494 {
 495         long *committed;
 496
 497         committed = &per_cpu(committed_space, (long)hcpu);
 498         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
 499                 atomic_long_add(*committed, &vm_committed_space);
 500                 *committed = 0;
 501                 drain_cpu_pagevecs((long)hcpu);
 502         }
 503         return NOTIFY_OK;
 504 }
 505 #endif /* CONFIG_HOTPLUG_CPU */
 506 #endif /* CONFIG_SMP */
 507
 508 /*
 509  * Perform any setup for the swap system
 510  */
 511 void __init swap_setup(void)
 512 {
 513         unsigned long megs = num_physpages >> (20 - PAGE_SHIFT);
 514
 515 #ifdef CONFIG_SWAP
 516         bdi_init(swapper_space.backing_dev_info);
 517 #endif
 518
 519         /* Use a smaller cluster for small-memory machines */
 520         if (megs < 16)
 521                 page_cluster = 2;
 522         else
 523                 page_cluster = 3;
 524         /*
 525          * Right now other parts of the system means that we
 526          * _really_ don't want to cluster much more
 527          */
 528 #ifdef CONFIG_HOTPLUG_CPU
 529         hotcpu_notifier(cpu_swap_callback, 0);
 530 #endif
 531 }