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 void put_page(struct page *page)
  38 {
  39         if (unlikely(PageCompound(page))) {
  40                 page = (struct page *)page_private(page);
  41                 if (put_page_testzero(page)) {
  42                         void (*dtor)(struct page *page);
  43
  44                         dtor = (void (*)(struct page *))page[1].mapping;
  45                         (*dtor)(page);
  46                 }
  47                 return;
  48         }
  49         if (put_page_testzero(page))
  50                 __page_cache_release(page);
  51 }
  52 EXPORT_SYMBOL(put_page);
  53
  54 /*
  55  * Writeback is about to end against a page which has been marked for immediate
  56  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
  57  * inactive list.  The page still has PageWriteback set, which will pin it.
  58  *
  59  * We don't expect many pages to come through here, so don't bother batching
  60  * things up.
  61  *
  62  * To avoid placing the page at the tail of the LRU while PG_writeback is still
  63  * set, this function will clear PG_writeback before performing the page
  64  * motion.  Do that inside the lru lock because once PG_writeback is cleared
  65  * we may not touch the page.
  66  *
  67  * Returns zero if it cleared PG_writeback.
  68  */
  69 int rotate_reclaimable_page(struct page *page)
  70 {
  71         struct zone *zone;
  72         unsigned long flags;
  73
  74         if (PageLocked(page))
  75                 return 1;
  76         if (PageDirty(page))
  77                 return 1;
  78         if (PageActive(page))
  79                 return 1;
  80         if (!PageLRU(page))
  81                 return 1;
  82
  83         zone = page_zone(page);
  84         spin_lock_irqsave(&zone->lru_lock, flags);
  85         if (PageLRU(page) && !PageActive(page)) {
  86                 list_del(&page->lru);
  87                 list_add_tail(&page->lru, &zone->inactive_list);
  88                 inc_page_state(pgrotated);
  89         }
  90         if (!test_clear_page_writeback(page))
  91                 BUG();
  92         spin_unlock_irqrestore(&zone->lru_lock, flags);
  93         return 0;
  94 }
  95
  96 /*
  97  * FIXME: speed this up?
  98  */
  99 void fastcall activate_page(struct page *page)
 100 {
 101         struct zone *zone = page_zone(page);
 102
 103         spin_lock_irq(&zone->lru_lock);
 104         if (PageLRU(page) && !PageActive(page)) {
 105                 del_page_from_inactive_list(zone, page);
 106                 SetPageActive(page);
 107                 add_page_to_active_list(zone, page);
 108                 inc_page_state(pgactivate);
 109         }
 110         spin_unlock_irq(&zone->lru_lock);
 111 }
 112
 113 /*
 114  * Mark a page as having seen activity.
 115  *
 116  * inactive,unreferenced        ->      inactive,referenced
 117  * inactive,referenced          ->      active,unreferenced
 118  * active,unreferenced          ->      active,referenced
 119  */
 120 void fastcall mark_page_accessed(struct page *page)
 121 {
 122         if (!PageActive(page) && PageReferenced(page) && PageLRU(page)) {
 123                 activate_page(page);
 124                 ClearPageReferenced(page);
 125         } else if (!PageReferenced(page)) {
 126                 SetPageReferenced(page);
 127         }
 128 }
 129
 130 EXPORT_SYMBOL(mark_page_accessed);
 131
 132 /**
 133  * lru_cache_add: add a page to the page lists
 134  * @page: the page to add
 135  */
 136 static DEFINE_PER_CPU(struct pagevec, lru_add_pvecs) = { 0, };
 137 static DEFINE_PER_CPU(struct pagevec, lru_add_active_pvecs) = { 0, };
 138
 139 void fastcall lru_cache_add(struct page *page)
 140 {
 141         struct pagevec *pvec = &get_cpu_var(lru_add_pvecs);
 142
 143         page_cache_get(page);
 144         if (!pagevec_add(pvec, page))
 145                 __pagevec_lru_add(pvec);
 146         put_cpu_var(lru_add_pvecs);
 147 }
 148
 149 void fastcall lru_cache_add_active(struct page *page)
 150 {
 151         struct pagevec *pvec = &get_cpu_var(lru_add_active_pvecs);
 152
 153         page_cache_get(page);
 154         if (!pagevec_add(pvec, page))
 155                 __pagevec_lru_add_active(pvec);
 156         put_cpu_var(lru_add_active_pvecs);
 157 }
 158
 159 static void __lru_add_drain(int cpu)
 160 {
 161         struct pagevec *pvec = &per_cpu(lru_add_pvecs, cpu);
 162
 163         /* CPU is dead, so no locking needed. */
 164         if (pagevec_count(pvec))
 165                 __pagevec_lru_add(pvec);
 166         pvec = &per_cpu(lru_add_active_pvecs, cpu);
 167         if (pagevec_count(pvec))
 168                 __pagevec_lru_add_active(pvec);
 169 }
 170
 171 void lru_add_drain(void)
 172 {
 173         __lru_add_drain(get_cpu());
 174         put_cpu();
 175 }
 176
 177 #ifdef CONFIG_NUMA
 178 static void lru_add_drain_per_cpu(void *dummy)
 179 {
 180         lru_add_drain();
 181 }
 182
 183 /*
 184  * Returns 0 for success
 185  */
 186 int lru_add_drain_all(void)
 187 {
 188         return schedule_on_each_cpu(lru_add_drain_per_cpu, NULL);
 189 }
 190
 191 #else
 192
 193 /*
 194  * Returns 0 for success
 195  */
 196 int lru_add_drain_all(void)
 197 {
 198         lru_add_drain();
 199         return 0;
 200 }
 201 #endif
 202
 203 /*
 204  * This path almost never happens for VM activity - pages are normally
 205  * freed via pagevecs.  But it gets used by networking.
 206  */
 207 void fastcall __page_cache_release(struct page *page)
 208 {
 209         unsigned long flags;
 210         struct zone *zone = page_zone(page);
 211
 212         spin_lock_irqsave(&zone->lru_lock, flags);
 213         if (TestClearPageLRU(page))
 214                 del_page_from_lru(zone, page);
 215         if (page_count(page) != 0)
 216                 page = NULL;
 217         spin_unlock_irqrestore(&zone->lru_lock, flags);
 218         if (page)
 219                 free_hot_page(page);
 220 }
 221
 222 EXPORT_SYMBOL(__page_cache_release);
 223
 224 /*
 225  * Batched page_cache_release().  Decrement the reference count on all the
 226  * passed pages.  If it fell to zero then remove the page from the LRU and
 227  * free it.
 228  *
 229  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 230  * for the remainder of the operation.
 231  *
 232  * The locking in this function is against shrink_cache(): we recheck the
 233  * page count inside the lock to see whether shrink_cache grabbed the page
 234  * via the LRU.  If it did, give up: shrink_cache will free it.
 235  */
 236 void release_pages(struct page **pages, int nr, int cold)
 237 {
 238         int i;
 239         struct pagevec pages_to_free;
 240         struct zone *zone = NULL;
 241
 242         pagevec_init(&pages_to_free, cold);
 243         for (i = 0; i < nr; i++) {
 244                 struct page *page = pages[i];
 245                 struct zone *pagezone;
 246
 247                 if (!put_page_testzero(page))
 248                         continue;
 249
 250                 pagezone = page_zone(page);
 251                 if (pagezone != zone) {
 252                         if (zone)
 253                                 spin_unlock_irq(&zone->lru_lock);
 254                         zone = pagezone;
 255                         spin_lock_irq(&zone->lru_lock);
 256                 }
 257                 if (TestClearPageLRU(page))
 258                         del_page_from_lru(zone, page);
 259                 if (page_count(page) == 0) {
 260                         if (!pagevec_add(&pages_to_free, page)) {
 261                                 spin_unlock_irq(&zone->lru_lock);
 262                                 __pagevec_free(&pages_to_free);
 263                                 pagevec_reinit(&pages_to_free);
 264                                 zone = NULL;    /* No lock is held */
 265                         }
 266                 }
 267         }
 268         if (zone)
 269                 spin_unlock_irq(&zone->lru_lock);
 270
 271         pagevec_free(&pages_to_free);
 272 }
 273
 274 /*
 275  * The pages which we're about to release may be in the deferred lru-addition
 276  * queues.  That would prevent them from really being freed right now.  That's
 277  * OK from a correctness point of view but is inefficient - those pages may be
 278  * cache-warm and we want to give them back to the page allocator ASAP.
 279  *
 280  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 281  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 282  * mutual recursion.
 283  */
 284 void __pagevec_release(struct pagevec *pvec)
 285 {
 286         lru_add_drain();
 287         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 288         pagevec_reinit(pvec);
 289 }
 290
 291 EXPORT_SYMBOL(__pagevec_release);
 292
 293 /*
 294  * pagevec_release() for pages which are known to not be on the LRU
 295  *
 296  * This function reinitialises the caller's pagevec.
 297  */
 298 void __pagevec_release_nonlru(struct pagevec *pvec)
 299 {
 300         int i;
 301         struct pagevec pages_to_free;
 302
 303         pagevec_init(&pages_to_free, pvec->cold);
 304         for (i = 0; i < pagevec_count(pvec); i++) {
 305                 struct page *page = pvec->pages[i];
 306
 307                 BUG_ON(PageLRU(page));
 308                 if (put_page_testzero(page))
 309                         pagevec_add(&pages_to_free, page);
 310         }
 311         pagevec_free(&pages_to_free);
 312         pagevec_reinit(pvec);
 313 }
 314
 315 /*
 316  * Add the passed pages to the LRU, then drop the caller's refcount
 317  * on them.  Reinitialises the caller's pagevec.
 318  */
 319 void __pagevec_lru_add(struct pagevec *pvec)
 320 {
 321         int i;
 322         struct zone *zone = NULL;
 323
 324         for (i = 0; i < pagevec_count(pvec); i++) {
 325                 struct page *page = pvec->pages[i];
 326                 struct zone *pagezone = page_zone(page);
 327
 328                 if (pagezone != zone) {
 329                         if (zone)
 330                                 spin_unlock_irq(&zone->lru_lock);
 331                         zone = pagezone;
 332                         spin_lock_irq(&zone->lru_lock);
 333                 }
 334                 if (TestSetPageLRU(page))
 335                         BUG();
 336                 add_page_to_inactive_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 EXPORT_SYMBOL(__pagevec_lru_add);
 345
 346 void __pagevec_lru_add_active(struct pagevec *pvec)
 347 {
 348         int i;
 349         struct zone *zone = NULL;
 350
 351         for (i = 0; i < pagevec_count(pvec); i++) {
 352                 struct page *page = pvec->pages[i];
 353                 struct zone *pagezone = page_zone(page);
 354
 355                 if (pagezone != zone) {
 356                         if (zone)
 357                                 spin_unlock_irq(&zone->lru_lock);
 358                         zone = pagezone;
 359                         spin_lock_irq(&zone->lru_lock);
 360                 }
 361                 if (TestSetPageLRU(page))
 362                         BUG();
 363                 if (TestSetPageActive(page))
 364                         BUG();
 365                 add_page_to_active_list(zone, page);
 366         }
 367         if (zone)
 368                 spin_unlock_irq(&zone->lru_lock);
 369         release_pages(pvec->pages, pvec->nr, pvec->cold);
 370         pagevec_reinit(pvec);
 371 }
 372
 373 /*
 374  * Try to drop buffers from the pages in a pagevec
 375  */
 376 void pagevec_strip(struct pagevec *pvec)
 377 {
 378         int i;
 379
 380         for (i = 0; i < pagevec_count(pvec); i++) {
 381                 struct page *page = pvec->pages[i];
 382
 383                 if (PagePrivate(page) && !TestSetPageLocked(page)) {
 384                         try_to_release_page(page, 0);
 385                         unlock_page(page);
 386                 }
 387         }
 388 }
 389
 390 /**
 391  * pagevec_lookup - gang pagecache lookup
 392  * @pvec:       Where the resulting pages are placed
 393  * @mapping:    The address_space to search
 394  * @start:      The starting page index
 395  * @nr_pages:   The maximum number of pages
 396  *
 397  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 398  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 399  * reference against the pages in @pvec.
 400  *
 401  * The search returns a group of mapping-contiguous pages with ascending
 402  * indexes.  There may be holes in the indices due to not-present pages.
 403  *
 404  * pagevec_lookup() returns the number of pages which were found.
 405  */
 406 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 407                 pgoff_t start, unsigned nr_pages)
 408 {
 409         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 410         return pagevec_count(pvec);
 411 }
 412
 413 EXPORT_SYMBOL(pagevec_lookup);
 414
 415 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 416                 pgoff_t *index, int tag, unsigned nr_pages)
 417 {
 418         pvec->nr = find_get_pages_tag(mapping, index, tag,
 419                                         nr_pages, pvec->pages);
 420         return pagevec_count(pvec);
 421 }
 422
 423 EXPORT_SYMBOL(pagevec_lookup_tag);
 424
 425 #ifdef CONFIG_SMP
 426 /*
 427  * We tolerate a little inaccuracy to avoid ping-ponging the counter between
 428  * CPUs
 429  */
 430 #define ACCT_THRESHOLD  max(16, NR_CPUS * 2)
 431
 432 static DEFINE_PER_CPU(long, committed_space) = 0;
 433
 434 void vm_acct_memory(long pages)
 435 {
 436         long *local;
 437
 438         preempt_disable();
 439         local = &__get_cpu_var(committed_space);
 440         *local += pages;
 441         if (*local > ACCT_THRESHOLD || *local < -ACCT_THRESHOLD) {
 442                 atomic_add(*local, &vm_committed_space);
 443                 *local = 0;
 444         }
 445         preempt_enable();
 446 }
 447
 448 #ifdef CONFIG_HOTPLUG_CPU
 449
 450 /* Drop the CPU's cached committed space back into the central pool. */
 451 static int cpu_swap_callback(struct notifier_block *nfb,
 452                              unsigned long action,
 453                              void *hcpu)
 454 {
 455         long *committed;
 456
 457         committed = &per_cpu(committed_space, (long)hcpu);
 458         if (action == CPU_DEAD) {
 459                 atomic_add(*committed, &vm_committed_space);
 460                 *committed = 0;
 461                 __lru_add_drain((long)hcpu);
 462         }
 463         return NOTIFY_OK;
 464 }
 465 #endif /* CONFIG_HOTPLUG_CPU */
 466 #endif /* CONFIG_SMP */
 467
 468 #ifdef CONFIG_SMP
 469 void percpu_counter_mod(struct percpu_counter *fbc, long amount)
 470 {
 471         long count;
 472         long *pcount;
 473         int cpu = get_cpu();
 474
 475         pcount = per_cpu_ptr(fbc->counters, cpu);
 476         count = *pcount + amount;
 477         if (count >= FBC_BATCH || count <= -FBC_BATCH) {
 478                 spin_lock(&fbc->lock);
 479                 fbc->count += count;
 480                 spin_unlock(&fbc->lock);
 481                 count = 0;
 482         }
 483         *pcount = count;
 484         put_cpu();
 485 }
 486 EXPORT_SYMBOL(percpu_counter_mod);
 487 #endif
 488
 489 /*
 490  * Perform any setup for the swap system
 491  */
 492 void __init swap_setup(void)
 493 {
 494         unsigned long megs = num_physpages >> (20 - PAGE_SHIFT);
 495
 496         /* Use a smaller cluster for small-memory machines */
 497         if (megs < 16)
 498                 page_cluster = 2;
 499         else
 500                 page_cluster = 3;
 501         /*
 502          * Right now other parts of the system means that we
 503          * _really_ don't want to cluster much more
 504          */
 505         hotcpu_notifier(cpu_swap_callback, 0);
 506 }