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