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);
  38 static DEFINE_PER_CPU(struct pagevec, lru_add_active_pvecs);
  39 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
  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 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 248 {
 249         lru_add_drain();
 250 }
 251
 252 /*
 253  * Returns 0 for success
 254  */
 255 int lru_add_drain_all(void)
 256 {
 257         return schedule_on_each_cpu(lru_add_drain_per_cpu);
 258 }
 259
 260 /*
 261  * Batched page_cache_release().  Decrement the reference count on all the
 262  * passed pages.  If it fell to zero then remove the page from the LRU and
 263  * free it.
 264  *
 265  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 266  * for the remainder of the operation.
 267  *
 268  * The locking in this function is against shrink_inactive_list(): we recheck
 269  * the page count inside the lock to see whether shrink_inactive_list()
 270  * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
 271  * will free it.
 272  */
 273 void release_pages(struct page **pages, int nr, int cold)
 274 {
 275         int i;
 276         struct pagevec pages_to_free;
 277         struct zone *zone = NULL;
 278         unsigned long uninitialized_var(flags);
 279
 280         pagevec_init(&pages_to_free, cold);
 281         for (i = 0; i < nr; i++) {
 282                 struct page *page = pages[i];
 283
 284                 if (unlikely(PageCompound(page))) {
 285                         if (zone) {
 286                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 287                                 zone = NULL;
 288                         }
 289                         put_compound_page(page);
 290                         continue;
 291                 }
 292
 293                 if (!put_page_testzero(page))
 294                         continue;
 295
 296                 if (PageLRU(page)) {
 297                         struct zone *pagezone = page_zone(page);
 298                         if (pagezone != zone) {
 299                                 if (zone)
 300                                         spin_unlock_irqrestore(&zone->lru_lock,
 301                                                                         flags);
 302                                 zone = pagezone;
 303                                 spin_lock_irqsave(&zone->lru_lock, flags);
 304                         }
 305                         VM_BUG_ON(!PageLRU(page));
 306                         __ClearPageLRU(page);
 307                         del_page_from_lru(zone, page);
 308                 }
 309
 310                 if (!pagevec_add(&pages_to_free, page)) {
 311                         if (zone) {
 312                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 313                                 zone = NULL;
 314                         }
 315                         __pagevec_free(&pages_to_free);
 316                         pagevec_reinit(&pages_to_free);
 317                 }
 318         }
 319         if (zone)
 320                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 321
 322         pagevec_free(&pages_to_free);
 323 }
 324
 325 /*
 326  * The pages which we're about to release may be in the deferred lru-addition
 327  * queues.  That would prevent them from really being freed right now.  That's
 328  * OK from a correctness point of view but is inefficient - those pages may be
 329  * cache-warm and we want to give them back to the page allocator ASAP.
 330  *
 331  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 332  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 333  * mutual recursion.
 334  */
 335 void __pagevec_release(struct pagevec *pvec)
 336 {
 337         lru_add_drain();
 338         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 339         pagevec_reinit(pvec);
 340 }
 341
 342 EXPORT_SYMBOL(__pagevec_release);
 343
 344 /*
 345  * pagevec_release() for pages which are known to not be on the LRU
 346  *
 347  * This function reinitialises the caller's pagevec.
 348  */
 349 void __pagevec_release_nonlru(struct pagevec *pvec)
 350 {
 351         int i;
 352         struct pagevec pages_to_free;
 353
 354         pagevec_init(&pages_to_free, pvec->cold);
 355         for (i = 0; i < pagevec_count(pvec); i++) {
 356                 struct page *page = pvec->pages[i];
 357
 358                 VM_BUG_ON(PageLRU(page));
 359                 if (put_page_testzero(page))
 360                         pagevec_add(&pages_to_free, page);
 361         }
 362         pagevec_free(&pages_to_free);
 363         pagevec_reinit(pvec);
 364 }
 365
 366 /*
 367  * Add the passed pages to the LRU, then drop the caller's refcount
 368  * on them.  Reinitialises the caller's pagevec.
 369  */
 370 void __pagevec_lru_add(struct pagevec *pvec)
 371 {
 372         int i;
 373         struct zone *zone = NULL;
 374
 375         for (i = 0; i < pagevec_count(pvec); i++) {
 376                 struct page *page = pvec->pages[i];
 377                 struct zone *pagezone = page_zone(page);
 378
 379                 if (pagezone != zone) {
 380                         if (zone)
 381                                 spin_unlock_irq(&zone->lru_lock);
 382                         zone = pagezone;
 383                         spin_lock_irq(&zone->lru_lock);
 384                 }
 385                 VM_BUG_ON(PageLRU(page));
 386                 SetPageLRU(page);
 387                 add_page_to_inactive_list(zone, page);
 388         }
 389         if (zone)
 390                 spin_unlock_irq(&zone->lru_lock);
 391         release_pages(pvec->pages, pvec->nr, pvec->cold);
 392         pagevec_reinit(pvec);
 393 }
 394
 395 EXPORT_SYMBOL(__pagevec_lru_add);
 396
 397 void __pagevec_lru_add_active(struct pagevec *pvec)
 398 {
 399         int i;
 400         struct zone *zone = NULL;
 401
 402         for (i = 0; i < pagevec_count(pvec); i++) {
 403                 struct page *page = pvec->pages[i];
 404                 struct zone *pagezone = page_zone(page);
 405
 406                 if (pagezone != zone) {
 407                         if (zone)
 408                                 spin_unlock_irq(&zone->lru_lock);
 409                         zone = pagezone;
 410                         spin_lock_irq(&zone->lru_lock);
 411                 }
 412                 VM_BUG_ON(PageLRU(page));
 413                 SetPageLRU(page);
 414                 VM_BUG_ON(PageActive(page));
 415                 SetPageActive(page);
 416                 add_page_to_active_list(zone, page);
 417         }
 418         if (zone)
 419                 spin_unlock_irq(&zone->lru_lock);
 420         release_pages(pvec->pages, pvec->nr, pvec->cold);
 421         pagevec_reinit(pvec);
 422 }
 423
 424 /*
 425  * Try to drop buffers from the pages in a pagevec
 426  */
 427 void pagevec_strip(struct pagevec *pvec)
 428 {
 429         int i;
 430
 431         for (i = 0; i < pagevec_count(pvec); i++) {
 432                 struct page *page = pvec->pages[i];
 433
 434                 if (PagePrivate(page) && trylock_page(page)) {
 435                         if (PagePrivate(page))
 436                                 try_to_release_page(page, 0);
 437                         unlock_page(page);
 438                 }
 439         }
 440 }
 441
 442 /**
 443  * pagevec_lookup - gang pagecache lookup
 444  * @pvec:       Where the resulting pages are placed
 445  * @mapping:    The address_space to search
 446  * @start:      The starting page index
 447  * @nr_pages:   The maximum number of pages
 448  *
 449  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 450  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 451  * reference against the pages in @pvec.
 452  *
 453  * The search returns a group of mapping-contiguous pages with ascending
 454  * indexes.  There may be holes in the indices due to not-present pages.
 455  *
 456  * pagevec_lookup() returns the number of pages which were found.
 457  */
 458 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 459                 pgoff_t start, unsigned nr_pages)
 460 {
 461         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 462         return pagevec_count(pvec);
 463 }
 464
 465 EXPORT_SYMBOL(pagevec_lookup);
 466
 467 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 468                 pgoff_t *index, int tag, unsigned nr_pages)
 469 {
 470         pvec->nr = find_get_pages_tag(mapping, index, tag,
 471                                         nr_pages, pvec->pages);
 472         return pagevec_count(pvec);
 473 }
 474
 475 EXPORT_SYMBOL(pagevec_lookup_tag);
 476
 477 #ifdef CONFIG_SMP
 478 /*
 479  * We tolerate a little inaccuracy to avoid ping-ponging the counter between
 480  * CPUs
 481  */
 482 #define ACCT_THRESHOLD  max(16, NR_CPUS * 2)
 483
 484 static DEFINE_PER_CPU(long, committed_space);
 485
 486 void vm_acct_memory(long pages)
 487 {
 488         long *local;
 489
 490         preempt_disable();
 491         local = &__get_cpu_var(committed_space);
 492         *local += pages;
 493         if (*local > ACCT_THRESHOLD || *local < -ACCT_THRESHOLD) {
 494                 atomic_long_add(*local, &vm_committed_space);
 495                 *local = 0;
 496         }
 497         preempt_enable();
 498 }
 499
 500 #ifdef CONFIG_HOTPLUG_CPU
 501
 502 /* Drop the CPU's cached committed space back into the central pool. */
 503 static int cpu_swap_callback(struct notifier_block *nfb,
 504                              unsigned long action,
 505                              void *hcpu)
 506 {
 507         long *committed;
 508
 509         committed = &per_cpu(committed_space, (long)hcpu);
 510         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
 511                 atomic_long_add(*committed, &vm_committed_space);
 512                 *committed = 0;
 513                 drain_cpu_pagevecs((long)hcpu);
 514         }
 515         return NOTIFY_OK;
 516 }
 517 #endif /* CONFIG_HOTPLUG_CPU */
 518 #endif /* CONFIG_SMP */
 519
 520 /*
 521  * Perform any setup for the swap system
 522  */
 523 void __init swap_setup(void)
 524 {
 525         unsigned long megs = num_physpages >> (20 - PAGE_SHIFT);
 526
 527 #ifdef CONFIG_SWAP
 528         bdi_init(swapper_space.backing_dev_info);
 529 #endif
 530
 531         /* Use a smaller cluster for small-memory machines */
 532         if (megs < 16)
 533                 page_cluster = 2;
 534         else
 535                 page_cluster = 3;
 536         /*
 537          * Right now other parts of the system means that we
 538          * _really_ don't want to cluster much more
 539          */
 540 #ifdef CONFIG_HOTPLUG_CPU
 541         hotcpu_notifier(cpu_swap_callback, 0);
 542 #endif
 543 }