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  *
  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  * @pages: list of pages threaded on page->lru
  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)) {
 121                         list_move_tail(&page->lru, &zone->inactive_list);
 122                         pgmoved++;
 123                 }
 124         }
 125         if (zone)
 126                 spin_unlock(&zone->lru_lock);
 127         __count_vm_events(PGROTATED, pgmoved);
 128         release_pages(pvec->pages, pvec->nr, pvec->cold);
 129         pagevec_reinit(pvec);
 130 }
 131
 132 /*
 133  * Writeback is about to end against a page which has been marked for immediate
 134  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 135  * inactive list.
 136  *
 137  * Returns zero if it cleared PG_writeback.
 138  */
 139 int rotate_reclaimable_page(struct page *page)
 140 {
 141         struct pagevec *pvec;
 142         unsigned long flags;
 143
 144         if (PageLocked(page))
 145                 return 1;
 146         if (PageDirty(page))
 147                 return 1;
 148         if (PageActive(page))
 149                 return 1;
 150         if (!PageLRU(page))
 151                 return 1;
 152
 153         page_cache_get(page);
 154         local_irq_save(flags);
 155         pvec = &__get_cpu_var(lru_rotate_pvecs);
 156         if (!pagevec_add(pvec, page))
 157                 pagevec_move_tail(pvec);
 158         local_irq_restore(flags);
 159
 160         if (!test_clear_page_writeback(page))
 161                 BUG();
 162
 163         return 0;
 164 }
 165
 166 /*
 167  * FIXME: speed this up?
 168  */
 169 void activate_page(struct page *page)
 170 {
 171         struct zone *zone = page_zone(page);
 172
 173         spin_lock_irq(&zone->lru_lock);
 174         if (PageLRU(page) && !PageActive(page)) {
 175                 del_page_from_inactive_list(zone, page);
 176                 SetPageActive(page);
 177                 add_page_to_active_list(zone, page);
 178                 __count_vm_event(PGACTIVATE);
 179                 mem_cgroup_move_lists(page, true);
 180         }
 181         spin_unlock_irq(&zone->lru_lock);
 182 }
 183
 184 /*
 185  * Mark a page as having seen activity.
 186  *
 187  * inactive,unreferenced        ->      inactive,referenced
 188  * inactive,referenced          ->      active,unreferenced
 189  * active,unreferenced          ->      active,referenced
 190  */
 191 void mark_page_accessed(struct page *page)
 192 {
 193         if (!PageActive(page) && PageReferenced(page) && PageLRU(page)) {
 194                 activate_page(page);
 195                 ClearPageReferenced(page);
 196         } else if (!PageReferenced(page)) {
 197                 SetPageReferenced(page);
 198         }
 199 }
 200
 201 EXPORT_SYMBOL(mark_page_accessed);
 202
 203 /**
 204  * lru_cache_add: add a page to the page lists
 205  * @page: the page to add
 206  */
 207 void lru_cache_add(struct page *page)
 208 {
 209         struct pagevec *pvec = &get_cpu_var(lru_add_pvecs);
 210
 211         page_cache_get(page);
 212         if (!pagevec_add(pvec, page))
 213                 __pagevec_lru_add(pvec);
 214         put_cpu_var(lru_add_pvecs);
 215 }
 216
 217 void lru_cache_add_active(struct page *page)
 218 {
 219         struct pagevec *pvec = &get_cpu_var(lru_add_active_pvecs);
 220
 221         page_cache_get(page);
 222         if (!pagevec_add(pvec, page))
 223                 __pagevec_lru_add_active(pvec);
 224         put_cpu_var(lru_add_active_pvecs);
 225 }
 226
 227 /*
 228  * Drain pages out of the cpu's pagevecs.
 229  * Either "cpu" is the current CPU, and preemption has already been
 230  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 231  */
 232 static void drain_cpu_pagevecs(int cpu)
 233 {
 234         struct pagevec *pvec;
 235
 236         pvec = &per_cpu(lru_add_pvecs, cpu);
 237         if (pagevec_count(pvec))
 238                 __pagevec_lru_add(pvec);
 239
 240         pvec = &per_cpu(lru_add_active_pvecs, cpu);
 241         if (pagevec_count(pvec))
 242                 __pagevec_lru_add_active(pvec);
 243
 244         pvec = &per_cpu(lru_rotate_pvecs, cpu);
 245         if (pagevec_count(pvec)) {
 246                 unsigned long flags;
 247
 248                 /* No harm done if a racing interrupt already did this */
 249                 local_irq_save(flags);
 250                 pagevec_move_tail(pvec);
 251                 local_irq_restore(flags);
 252         }
 253 }
 254
 255 void lru_add_drain(void)
 256 {
 257         drain_cpu_pagevecs(get_cpu());
 258         put_cpu();
 259 }
 260
 261 #ifdef CONFIG_NUMA
 262 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 263 {
 264         lru_add_drain();
 265 }
 266
 267 /*
 268  * Returns 0 for success
 269  */
 270 int lru_add_drain_all(void)
 271 {
 272         return schedule_on_each_cpu(lru_add_drain_per_cpu);
 273 }
 274
 275 #else
 276
 277 /*
 278  * Returns 0 for success
 279  */
 280 int lru_add_drain_all(void)
 281 {
 282         lru_add_drain();
 283         return 0;
 284 }
 285 #endif
 286
 287 /*
 288  * Batched page_cache_release().  Decrement the reference count on all the
 289  * passed pages.  If it fell to zero then remove the page from the LRU and
 290  * free it.
 291  *
 292  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 293  * for the remainder of the operation.
 294  *
 295  * The locking in this function is against shrink_cache(): we recheck the
 296  * page count inside the lock to see whether shrink_cache grabbed the page
 297  * via the LRU.  If it did, give up: shrink_cache will free it.
 298  */
 299 void release_pages(struct page **pages, int nr, int cold)
 300 {
 301         int i;
 302         struct pagevec pages_to_free;
 303         struct zone *zone = NULL;
 304         unsigned long uninitialized_var(flags);
 305
 306         pagevec_init(&pages_to_free, cold);
 307         for (i = 0; i < nr; i++) {
 308                 struct page *page = pages[i];
 309
 310                 if (unlikely(PageCompound(page))) {
 311                         if (zone) {
 312                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 313                                 zone = NULL;
 314                         }
 315                         put_compound_page(page);
 316                         continue;
 317                 }
 318
 319                 if (!put_page_testzero(page))
 320                         continue;
 321
 322                 if (PageLRU(page)) {
 323                         struct zone *pagezone = page_zone(page);
 324                         if (pagezone != zone) {
 325                                 if (zone)
 326                                         spin_unlock_irqrestore(&zone->lru_lock,
 327                                                                         flags);
 328                                 zone = pagezone;
 329                                 spin_lock_irqsave(&zone->lru_lock, flags);
 330                         }
 331                         VM_BUG_ON(!PageLRU(page));
 332                         __ClearPageLRU(page);
 333                         del_page_from_lru(zone, page);
 334                 }
 335
 336                 if (!pagevec_add(&pages_to_free, page)) {
 337                         if (zone) {
 338                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 339                                 zone = NULL;
 340                         }
 341                         __pagevec_free(&pages_to_free);
 342                         pagevec_reinit(&pages_to_free);
 343                 }
 344         }
 345         if (zone)
 346                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 347
 348         pagevec_free(&pages_to_free);
 349 }
 350
 351 /*
 352  * The pages which we're about to release may be in the deferred lru-addition
 353  * queues.  That would prevent them from really being freed right now.  That's
 354  * OK from a correctness point of view but is inefficient - those pages may be
 355  * cache-warm and we want to give them back to the page allocator ASAP.
 356  *
 357  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 358  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 359  * mutual recursion.
 360  */
 361 void __pagevec_release(struct pagevec *pvec)
 362 {
 363         lru_add_drain();
 364         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 365         pagevec_reinit(pvec);
 366 }
 367
 368 EXPORT_SYMBOL(__pagevec_release);
 369
 370 /*
 371  * pagevec_release() for pages which are known to not be on the LRU
 372  *
 373  * This function reinitialises the caller's pagevec.
 374  */
 375 void __pagevec_release_nonlru(struct pagevec *pvec)
 376 {
 377         int i;
 378         struct pagevec pages_to_free;
 379
 380         pagevec_init(&pages_to_free, pvec->cold);
 381         for (i = 0; i < pagevec_count(pvec); i++) {
 382                 struct page *page = pvec->pages[i];
 383
 384                 VM_BUG_ON(PageLRU(page));
 385                 if (put_page_testzero(page))
 386                         pagevec_add(&pages_to_free, page);
 387         }
 388         pagevec_free(&pages_to_free);
 389         pagevec_reinit(pvec);
 390 }
 391
 392 /*
 393  * Add the passed pages to the LRU, then drop the caller's refcount
 394  * on them.  Reinitialises the caller's pagevec.
 395  */
 396 void __pagevec_lru_add(struct pagevec *pvec)
 397 {
 398         int i;
 399         struct zone *zone = NULL;
 400
 401         for (i = 0; i < pagevec_count(pvec); i++) {
 402                 struct page *page = pvec->pages[i];
 403                 struct zone *pagezone = page_zone(page);
 404
 405                 if (pagezone != zone) {
 406                         if (zone)
 407                                 spin_unlock_irq(&zone->lru_lock);
 408                         zone = pagezone;
 409                         spin_lock_irq(&zone->lru_lock);
 410                 }
 411                 VM_BUG_ON(PageLRU(page));
 412                 SetPageLRU(page);
 413                 add_page_to_inactive_list(zone, page);
 414         }
 415         if (zone)
 416                 spin_unlock_irq(&zone->lru_lock);
 417         release_pages(pvec->pages, pvec->nr, pvec->cold);
 418         pagevec_reinit(pvec);
 419 }
 420
 421 EXPORT_SYMBOL(__pagevec_lru_add);
 422
 423 void __pagevec_lru_add_active(struct pagevec *pvec)
 424 {
 425         int i;
 426         struct zone *zone = NULL;
 427
 428         for (i = 0; i < pagevec_count(pvec); i++) {
 429                 struct page *page = pvec->pages[i];
 430                 struct zone *pagezone = page_zone(page);
 431
 432                 if (pagezone != zone) {
 433                         if (zone)
 434                                 spin_unlock_irq(&zone->lru_lock);
 435                         zone = pagezone;
 436                         spin_lock_irq(&zone->lru_lock);
 437                 }
 438                 VM_BUG_ON(PageLRU(page));
 439                 SetPageLRU(page);
 440                 VM_BUG_ON(PageActive(page));
 441                 SetPageActive(page);
 442                 add_page_to_active_list(zone, page);
 443         }
 444         if (zone)
 445                 spin_unlock_irq(&zone->lru_lock);
 446         release_pages(pvec->pages, pvec->nr, pvec->cold);
 447         pagevec_reinit(pvec);
 448 }
 449
 450 /*
 451  * Try to drop buffers from the pages in a pagevec
 452  */
 453 void pagevec_strip(struct pagevec *pvec)
 454 {
 455         int i;
 456
 457         for (i = 0; i < pagevec_count(pvec); i++) {
 458                 struct page *page = pvec->pages[i];
 459
 460                 if (PagePrivate(page) && !TestSetPageLocked(page)) {
 461                         if (PagePrivate(page))
 462                                 try_to_release_page(page, 0);
 463                         unlock_page(page);
 464                 }
 465         }
 466 }
 467
 468 /**
 469  * pagevec_lookup - gang pagecache lookup
 470  * @pvec:       Where the resulting pages are placed
 471  * @mapping:    The address_space to search
 472  * @start:      The starting page index
 473  * @nr_pages:   The maximum number of pages
 474  *
 475  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 476  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 477  * reference against the pages in @pvec.
 478  *
 479  * The search returns a group of mapping-contiguous pages with ascending
 480  * indexes.  There may be holes in the indices due to not-present pages.
 481  *
 482  * pagevec_lookup() returns the number of pages which were found.
 483  */
 484 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 485                 pgoff_t start, unsigned nr_pages)
 486 {
 487         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 488         return pagevec_count(pvec);
 489 }
 490
 491 EXPORT_SYMBOL(pagevec_lookup);
 492
 493 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 494                 pgoff_t *index, int tag, unsigned nr_pages)
 495 {
 496         pvec->nr = find_get_pages_tag(mapping, index, tag,
 497                                         nr_pages, pvec->pages);
 498         return pagevec_count(pvec);
 499 }
 500
 501 EXPORT_SYMBOL(pagevec_lookup_tag);
 502
 503 #ifdef CONFIG_SMP
 504 /*
 505  * We tolerate a little inaccuracy to avoid ping-ponging the counter between
 506  * CPUs
 507  */
 508 #define ACCT_THRESHOLD  max(16, NR_CPUS * 2)
 509
 510 static DEFINE_PER_CPU(long, committed_space) = 0;
 511
 512 void vm_acct_memory(long pages)
 513 {
 514         long *local;
 515
 516         preempt_disable();
 517         local = &__get_cpu_var(committed_space);
 518         *local += pages;
 519         if (*local > ACCT_THRESHOLD || *local < -ACCT_THRESHOLD) {
 520                 atomic_add(*local, &vm_committed_space);
 521                 *local = 0;
 522         }
 523         preempt_enable();
 524 }
 525
 526 #ifdef CONFIG_HOTPLUG_CPU
 527
 528 /* Drop the CPU's cached committed space back into the central pool. */
 529 static int cpu_swap_callback(struct notifier_block *nfb,
 530                              unsigned long action,
 531                              void *hcpu)
 532 {
 533         long *committed;
 534
 535         committed = &per_cpu(committed_space, (long)hcpu);
 536         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
 537                 atomic_add(*committed, &vm_committed_space);
 538                 *committed = 0;
 539                 drain_cpu_pagevecs((long)hcpu);
 540         }
 541         return NOTIFY_OK;
 542 }
 543 #endif /* CONFIG_HOTPLUG_CPU */
 544 #endif /* CONFIG_SMP */
 545
 546 /*
 547  * Perform any setup for the swap system
 548  */
 549 void __init swap_setup(void)
 550 {
 551         unsigned long megs = num_physpages >> (20 - PAGE_SHIFT);
 552
 553 #ifdef CONFIG_SWAP
 554         bdi_init(swapper_space.backing_dev_info);
 555 #endif
 556
 557         /* Use a smaller cluster for small-memory machines */
 558         if (megs < 16)
 559                 page_cluster = 2;
 560         else
 561                 page_cluster = 3;
 562         /*
 563          * Right now other parts of the system means that we
 564          * _really_ don't want to cluster much more
 565          */
 566 #ifdef CONFIG_HOTPLUG_CPU
 567         hotcpu_notifier(cpu_swap_callback, 0);
 568 #endif
 569 }