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 opereation 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/module.h>
  28 #include <linux/percpu_counter.h>
  29 #include <linux/percpu.h>
  30 #include <linux/cpu.h>
  31 #include <linux/notifier.h>
  32 #include <linux/init.h>
  33
  34 /* How many pages do we try to swap or page in/out together? */
  35 int page_cluster;
  36
  37 #ifdef CONFIG_HUGETLB_PAGE
  38
  39 void put_page(struct page *page)
  40 {
  41         if (unlikely(PageCompound(page))) {
  42                 page = (struct page *)page_private(page);
  43                 if (put_page_testzero(page)) {
  44                         void (*dtor)(struct page *page);
  45
  46                         dtor = (void (*)(struct page *))page[1].mapping;
  47                         (*dtor)(page);
  48                 }
  49                 return;
  50         }
  51         if (put_page_testzero(page))
  52                 __page_cache_release(page);
  53 }
  54 EXPORT_SYMBOL(put_page);
  55 #endif
  56
  57 /*
  58  * Writeback is about to end against a page which has been marked for immediate
  59  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
  60  * inactive list.  The page still has PageWriteback set, which will pin it.
  61  *
  62  * We don't expect many pages to come through here, so don't bother batching
  63  * things up.
  64  *
  65  * To avoid placing the page at the tail of the LRU while PG_writeback is still
  66  * set, this function will clear PG_writeback before performing the page
  67  * motion.  Do that inside the lru lock because once PG_writeback is cleared
  68  * we may not touch the page.
  69  *
  70  * Returns zero if it cleared PG_writeback.
  71  */
  72 int rotate_reclaimable_page(struct page *page)
  73 {
  74         struct zone *zone;
  75         unsigned long flags;
  76
  77         if (PageLocked(page))
  78                 return 1;
  79         if (PageDirty(page))
  80                 return 1;
  81         if (PageActive(page))
  82                 return 1;
  83         if (!PageLRU(page))
  84                 return 1;
  85
  86         zone = page_zone(page);
  87         spin_lock_irqsave(&zone->lru_lock, flags);
  88         if (PageLRU(page) && !PageActive(page)) {
  89                 list_del(&page->lru);
  90                 list_add_tail(&page->lru, &zone->inactive_list);
  91                 inc_page_state(pgrotated);
  92         }
  93         if (!test_clear_page_writeback(page))
  94                 BUG();
  95         spin_unlock_irqrestore(&zone->lru_lock, flags);
  96         return 0;
  97 }
  98
  99 /*
 100  * FIXME: speed this up?
 101  */
 102 void fastcall activate_page(struct page *page)
 103 {
 104         struct zone *zone = page_zone(page);
 105
 106         spin_lock_irq(&zone->lru_lock);
 107         if (PageLRU(page) && !PageActive(page)) {
 108                 del_page_from_inactive_list(zone, page);
 109                 SetPageActive(page);
 110                 add_page_to_active_list(zone, page);
 111                 inc_page_state(pgactivate);
 112         }
 113         spin_unlock_irq(&zone->lru_lock);
 114 }
 115
 116 /*
 117  * Mark a page as having seen activity.
 118  *
 119  * inactive,unreferenced        ->      inactive,referenced
 120  * inactive,referenced          ->      active,unreferenced
 121  * active,unreferenced          ->      active,referenced
 122  */
 123 void fastcall mark_page_accessed(struct page *page)
 124 {
 125         if (!PageActive(page) && PageReferenced(page) && PageLRU(page)) {
 126                 activate_page(page);
 127                 ClearPageReferenced(page);
 128         } else if (!PageReferenced(page)) {
 129                 SetPageReferenced(page);
 130         }
 131 }
 132
 133 EXPORT_SYMBOL(mark_page_accessed);
 134
 135 /**
 136  * lru_cache_add: add a page to the page lists
 137  * @page: the page to add
 138  */
 139 static DEFINE_PER_CPU(struct pagevec, lru_add_pvecs) = { 0, };
 140 static DEFINE_PER_CPU(struct pagevec, lru_add_active_pvecs) = { 0, };
 141
 142 void fastcall lru_cache_add(struct page *page)
 143 {
 144         struct pagevec *pvec = &get_cpu_var(lru_add_pvecs);
 145
 146         page_cache_get(page);
 147         if (!pagevec_add(pvec, page))
 148                 __pagevec_lru_add(pvec);
 149         put_cpu_var(lru_add_pvecs);
 150 }
 151
 152 void fastcall lru_cache_add_active(struct page *page)
 153 {
 154         struct pagevec *pvec = &get_cpu_var(lru_add_active_pvecs);
 155
 156         page_cache_get(page);
 157         if (!pagevec_add(pvec, page))
 158                 __pagevec_lru_add_active(pvec);
 159         put_cpu_var(lru_add_active_pvecs);
 160 }
 161
 162 void lru_add_drain(void)
 163 {
 164         struct pagevec *pvec = &get_cpu_var(lru_add_pvecs);
 165
 166         if (pagevec_count(pvec))
 167                 __pagevec_lru_add(pvec);
 168         pvec = &__get_cpu_var(lru_add_active_pvecs);
 169         if (pagevec_count(pvec))
 170                 __pagevec_lru_add_active(pvec);
 171         put_cpu_var(lru_add_pvecs);
 172 }
 173
 174 /*
 175  * This path almost never happens for VM activity - pages are normally
 176  * freed via pagevecs.  But it gets used by networking.
 177  */
 178 void fastcall __page_cache_release(struct page *page)
 179 {
 180         unsigned long flags;
 181         struct zone *zone = page_zone(page);
 182
 183         spin_lock_irqsave(&zone->lru_lock, flags);
 184         if (TestClearPageLRU(page))
 185                 del_page_from_lru(zone, page);
 186         if (page_count(page) != 0)
 187                 page = NULL;
 188         spin_unlock_irqrestore(&zone->lru_lock, flags);
 189         if (page)
 190                 free_hot_page(page);
 191 }
 192
 193 EXPORT_SYMBOL(__page_cache_release);
 194
 195 /*
 196  * Batched page_cache_release().  Decrement the reference count on all the
 197  * passed pages.  If it fell to zero then remove the page from the LRU and
 198  * free it.
 199  *
 200  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 201  * for the remainder of the operation.
 202  *
 203  * The locking in this function is against shrink_cache(): we recheck the
 204  * page count inside the lock to see whether shrink_cache grabbed the page
 205  * via the LRU.  If it did, give up: shrink_cache will free it.
 206  */
 207 void release_pages(struct page **pages, int nr, int cold)
 208 {
 209         int i;
 210         struct pagevec pages_to_free;
 211         struct zone *zone = NULL;
 212
 213         pagevec_init(&pages_to_free, cold);
 214         for (i = 0; i < nr; i++) {
 215                 struct page *page = pages[i];
 216                 struct zone *pagezone;
 217
 218                 if (!put_page_testzero(page))
 219                         continue;
 220
 221                 pagezone = page_zone(page);
 222                 if (pagezone != zone) {
 223                         if (zone)
 224                                 spin_unlock_irq(&zone->lru_lock);
 225                         zone = pagezone;
 226                         spin_lock_irq(&zone->lru_lock);
 227                 }
 228                 if (TestClearPageLRU(page))
 229                         del_page_from_lru(zone, page);
 230                 if (page_count(page) == 0) {
 231                         if (!pagevec_add(&pages_to_free, page)) {
 232                                 spin_unlock_irq(&zone->lru_lock);
 233                                 __pagevec_free(&pages_to_free);
 234                                 pagevec_reinit(&pages_to_free);
 235                                 zone = NULL;    /* No lock is held */
 236                         }
 237                 }
 238         }
 239         if (zone)
 240                 spin_unlock_irq(&zone->lru_lock);
 241
 242         pagevec_free(&pages_to_free);
 243 }
 244
 245 /*
 246  * The pages which we're about to release may be in the deferred lru-addition
 247  * queues.  That would prevent them from really being freed right now.  That's
 248  * OK from a correctness point of view but is inefficient - those pages may be
 249  * cache-warm and we want to give them back to the page allocator ASAP.
 250  *
 251  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 252  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 253  * mutual recursion.
 254  */
 255 void __pagevec_release(struct pagevec *pvec)
 256 {
 257         lru_add_drain();
 258         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 259         pagevec_reinit(pvec);
 260 }
 261
 262 EXPORT_SYMBOL(__pagevec_release);
 263
 264 /*
 265  * pagevec_release() for pages which are known to not be on the LRU
 266  *
 267  * This function reinitialises the caller's pagevec.
 268  */
 269 void __pagevec_release_nonlru(struct pagevec *pvec)
 270 {
 271         int i;
 272         struct pagevec pages_to_free;
 273
 274         pagevec_init(&pages_to_free, pvec->cold);
 275         for (i = 0; i < pagevec_count(pvec); i++) {
 276                 struct page *page = pvec->pages[i];
 277
 278                 BUG_ON(PageLRU(page));
 279                 if (put_page_testzero(page))
 280                         pagevec_add(&pages_to_free, page);
 281         }
 282         pagevec_free(&pages_to_free);
 283         pagevec_reinit(pvec);
 284 }
 285
 286 /*
 287  * Add the passed pages to the LRU, then drop the caller's refcount
 288  * on them.  Reinitialises the caller's pagevec.
 289  */
 290 void __pagevec_lru_add(struct pagevec *pvec)
 291 {
 292         int i;
 293         struct zone *zone = NULL;
 294
 295         for (i = 0; i < pagevec_count(pvec); i++) {
 296                 struct page *page = pvec->pages[i];
 297                 struct zone *pagezone = page_zone(page);
 298
 299                 if (pagezone != zone) {
 300                         if (zone)
 301                                 spin_unlock_irq(&zone->lru_lock);
 302                         zone = pagezone;
 303                         spin_lock_irq(&zone->lru_lock);
 304                 }
 305                 if (TestSetPageLRU(page))
 306                         BUG();
 307                 add_page_to_inactive_list(zone, page);
 308         }
 309         if (zone)
 310                 spin_unlock_irq(&zone->lru_lock);
 311         release_pages(pvec->pages, pvec->nr, pvec->cold);
 312         pagevec_reinit(pvec);
 313 }
 314
 315 EXPORT_SYMBOL(__pagevec_lru_add);
 316
 317 void __pagevec_lru_add_active(struct pagevec *pvec)
 318 {
 319         int i;
 320         struct zone *zone = NULL;
 321
 322         for (i = 0; i < pagevec_count(pvec); i++) {
 323                 struct page *page = pvec->pages[i];
 324                 struct zone *pagezone = page_zone(page);
 325
 326                 if (pagezone != zone) {
 327                         if (zone)
 328                                 spin_unlock_irq(&zone->lru_lock);
 329                         zone = pagezone;
 330                         spin_lock_irq(&zone->lru_lock);
 331                 }
 332                 if (TestSetPageLRU(page))
 333                         BUG();
 334                 if (TestSetPageActive(page))
 335                         BUG();
 336                 add_page_to_active_list(zone, page);
 337         }
 338         if (zone)
 339                 spin_unlock_irq(&zone->lru_lock);
 340         release_pages(pvec->pages, pvec->nr, pvec->cold);
 341         pagevec_reinit(pvec);
 342 }
 343
 344 /*
 345  * Try to drop buffers from the pages in a pagevec
 346  */
 347 void pagevec_strip(struct pagevec *pvec)
 348 {
 349         int i;
 350
 351         for (i = 0; i < pagevec_count(pvec); i++) {
 352                 struct page *page = pvec->pages[i];
 353
 354                 if (PagePrivate(page) && !TestSetPageLocked(page)) {
 355                         try_to_release_page(page, 0);
 356                         unlock_page(page);
 357                 }
 358         }
 359 }
 360
 361 /**
 362  * pagevec_lookup - gang pagecache lookup
 363  * @pvec:       Where the resulting pages are placed
 364  * @mapping:    The address_space to search
 365  * @start:      The starting page index
 366  * @nr_pages:   The maximum number of pages
 367  *
 368  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 369  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 370  * reference against the pages in @pvec.
 371  *
 372  * The search returns a group of mapping-contiguous pages with ascending
 373  * indexes.  There may be holes in the indices due to not-present pages.
 374  *
 375  * pagevec_lookup() returns the number of pages which were found.
 376  */
 377 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 378                 pgoff_t start, unsigned nr_pages)
 379 {
 380         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 381         return pagevec_count(pvec);
 382 }
 383
 384 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 385                 pgoff_t *index, int tag, unsigned nr_pages)
 386 {
 387         pvec->nr = find_get_pages_tag(mapping, index, tag,
 388                                         nr_pages, pvec->pages);
 389         return pagevec_count(pvec);
 390 }
 391
 392 EXPORT_SYMBOL(pagevec_lookup_tag);
 393
 394 #ifdef CONFIG_SMP
 395 /*
 396  * We tolerate a little inaccuracy to avoid ping-ponging the counter between
 397  * CPUs
 398  */
 399 #define ACCT_THRESHOLD  max(16, NR_CPUS * 2)
 400
 401 static DEFINE_PER_CPU(long, committed_space) = 0;
 402
 403 void vm_acct_memory(long pages)
 404 {
 405         long *local;
 406
 407         preempt_disable();
 408         local = &__get_cpu_var(committed_space);
 409         *local += pages;
 410         if (*local > ACCT_THRESHOLD || *local < -ACCT_THRESHOLD) {
 411                 atomic_add(*local, &vm_committed_space);
 412                 *local = 0;
 413         }
 414         preempt_enable();
 415 }
 416
 417 #ifdef CONFIG_HOTPLUG_CPU
 418 static void lru_drain_cache(unsigned int cpu)
 419 {
 420         struct pagevec *pvec = &per_cpu(lru_add_pvecs, cpu);
 421
 422         /* CPU is dead, so no locking needed. */
 423         if (pagevec_count(pvec))
 424                 __pagevec_lru_add(pvec);
 425         pvec = &per_cpu(lru_add_active_pvecs, cpu);
 426         if (pagevec_count(pvec))
 427                 __pagevec_lru_add_active(pvec);
 428 }
 429
 430 /* Drop the CPU's cached committed space back into the central pool. */
 431 static int cpu_swap_callback(struct notifier_block *nfb,
 432                              unsigned long action,
 433                              void *hcpu)
 434 {
 435         long *committed;
 436
 437         committed = &per_cpu(committed_space, (long)hcpu);
 438         if (action == CPU_DEAD) {
 439                 atomic_add(*committed, &vm_committed_space);
 440                 *committed = 0;
 441                 lru_drain_cache((long)hcpu);
 442         }
 443         return NOTIFY_OK;
 444 }
 445 #endif /* CONFIG_HOTPLUG_CPU */
 446 #endif /* CONFIG_SMP */
 447
 448 #ifdef CONFIG_SMP
 449 void percpu_counter_mod(struct percpu_counter *fbc, long amount)
 450 {
 451         long count;
 452         long *pcount;
 453         int cpu = get_cpu();
 454
 455         pcount = per_cpu_ptr(fbc->counters, cpu);
 456         count = *pcount + amount;
 457         if (count >= FBC_BATCH || count <= -FBC_BATCH) {
 458                 spin_lock(&fbc->lock);
 459                 fbc->count += count;
 460                 spin_unlock(&fbc->lock);
 461                 count = 0;
 462         }
 463         *pcount = count;
 464         put_cpu();
 465 }
 466 EXPORT_SYMBOL(percpu_counter_mod);
 467 #endif
 468
 469 /*
 470  * Perform any setup for the swap system
 471  */
 472 void __init swap_setup(void)
 473 {
 474         unsigned long megs = num_physpages >> (20 - PAGE_SHIFT);
 475
 476         /* Use a smaller cluster for small-memory machines */
 477         if (megs < 16)
 478                 page_cluster = 2;
 479         else
 480                 page_cluster = 3;
 481         /*
 482          * Right now other parts of the system means that we
 483          * _really_ don't want to cluster much more
 484          */
 485         hotcpu_notifier(cpu_swap_callback, 0);
 486 }