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 static void put_compound_page(struct page *page)
  38 {
  39         page = (struct page *)page_private(page);
  40         if (put_page_testzero(page)) {
  41                 void (*dtor)(struct page *page);
  42
  43                 dtor = (void (*)(struct page *))page[1].lru.next;
  44                 (*dtor)(page);
  45         }
  46 }
  47
  48 void put_page(struct page *page)
  49 {
  50         if (unlikely(PageCompound(page)))
  51                 put_compound_page(page);
  52         else if (put_page_testzero(page))
  53                 __page_cache_release(page);
  54 }
  55 EXPORT_SYMBOL(put_page);
  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 static void __lru_add_drain(int cpu)
 163 {
 164         struct pagevec *pvec = &per_cpu(lru_add_pvecs, cpu);
 165
 166         /* CPU is dead, so no locking needed. */
 167         if (pagevec_count(pvec))
 168                 __pagevec_lru_add(pvec);
 169         pvec = &per_cpu(lru_add_active_pvecs, cpu);
 170         if (pagevec_count(pvec))
 171                 __pagevec_lru_add_active(pvec);
 172 }
 173
 174 void lru_add_drain(void)
 175 {
 176         __lru_add_drain(get_cpu());
 177         put_cpu();
 178 }
 179
 180 #ifdef CONFIG_NUMA
 181 static void lru_add_drain_per_cpu(void *dummy)
 182 {
 183         lru_add_drain();
 184 }
 185
 186 /*
 187  * Returns 0 for success
 188  */
 189 int lru_add_drain_all(void)
 190 {
 191         return schedule_on_each_cpu(lru_add_drain_per_cpu, NULL);
 192 }
 193
 194 #else
 195
 196 /*
 197  * Returns 0 for success
 198  */
 199 int lru_add_drain_all(void)
 200 {
 201         lru_add_drain();
 202         return 0;
 203 }
 204 #endif
 205
 206 /*
 207  * This path almost never happens for VM activity - pages are normally
 208  * freed via pagevecs.  But it gets used by networking.
 209  */
 210 void fastcall __page_cache_release(struct page *page)
 211 {
 212         if (PageLRU(page)) {
 213                 unsigned long flags;
 214                 struct zone *zone = page_zone(page);
 215
 216                 spin_lock_irqsave(&zone->lru_lock, flags);
 217                 BUG_ON(!PageLRU(page));
 218                 __ClearPageLRU(page);
 219                 del_page_from_lru(zone, page);
 220                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 221         }
 222         free_hot_page(page);
 223 }
 224 EXPORT_SYMBOL(__page_cache_release);
 225
 226 /*
 227  * Batched page_cache_release().  Decrement the reference count on all the
 228  * passed pages.  If it fell to zero then remove the page from the LRU and
 229  * free it.
 230  *
 231  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 232  * for the remainder of the operation.
 233  *
 234  * The locking in this function is against shrink_cache(): we recheck the
 235  * page count inside the lock to see whether shrink_cache grabbed the page
 236  * via the LRU.  If it did, give up: shrink_cache will free it.
 237  */
 238 void release_pages(struct page **pages, int nr, int cold)
 239 {
 240         int i;
 241         struct pagevec pages_to_free;
 242         struct zone *zone = NULL;
 243
 244         pagevec_init(&pages_to_free, cold);
 245         for (i = 0; i < nr; i++) {
 246                 struct page *page = pages[i];
 247
 248                 if (unlikely(PageCompound(page))) {
 249                         if (zone) {
 250                                 spin_unlock_irq(&zone->lru_lock);
 251                                 zone = NULL;
 252                         }
 253                         put_compound_page(page);
 254                         continue;
 255                 }
 256
 257                 if (!put_page_testzero(page))
 258                         continue;
 259
 260                 if (PageLRU(page)) {
 261                         struct zone *pagezone = page_zone(page);
 262                         if (pagezone != zone) {
 263                                 if (zone)
 264                                         spin_unlock_irq(&zone->lru_lock);
 265                                 zone = pagezone;
 266                                 spin_lock_irq(&zone->lru_lock);
 267                         }
 268                         BUG_ON(!PageLRU(page));
 269                         __ClearPageLRU(page);
 270                         del_page_from_lru(zone, page);
 271                 }
 272
 273                 if (!pagevec_add(&pages_to_free, page)) {
 274                         if (zone) {
 275                                 spin_unlock_irq(&zone->lru_lock);
 276                                 zone = NULL;
 277                         }
 278                         __pagevec_free(&pages_to_free);
 279                         pagevec_reinit(&pages_to_free);
 280                 }
 281         }
 282         if (zone)
 283                 spin_unlock_irq(&zone->lru_lock);
 284
 285         pagevec_free(&pages_to_free);
 286 }
 287
 288 /*
 289  * The pages which we're about to release may be in the deferred lru-addition
 290  * queues.  That would prevent them from really being freed right now.  That's
 291  * OK from a correctness point of view but is inefficient - those pages may be
 292  * cache-warm and we want to give them back to the page allocator ASAP.
 293  *
 294  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 295  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 296  * mutual recursion.
 297  */
 298 void __pagevec_release(struct pagevec *pvec)
 299 {
 300         lru_add_drain();
 301         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 302         pagevec_reinit(pvec);
 303 }
 304
 305 EXPORT_SYMBOL(__pagevec_release);
 306
 307 /*
 308  * pagevec_release() for pages which are known to not be on the LRU
 309  *
 310  * This function reinitialises the caller's pagevec.
 311  */
 312 void __pagevec_release_nonlru(struct pagevec *pvec)
 313 {
 314         int i;
 315         struct pagevec pages_to_free;
 316
 317         pagevec_init(&pages_to_free, pvec->cold);
 318         for (i = 0; i < pagevec_count(pvec); i++) {
 319                 struct page *page = pvec->pages[i];
 320
 321                 BUG_ON(PageLRU(page));
 322                 if (put_page_testzero(page))
 323                         pagevec_add(&pages_to_free, page);
 324         }
 325         pagevec_free(&pages_to_free);
 326         pagevec_reinit(pvec);
 327 }
 328
 329 /*
 330  * Add the passed pages to the LRU, then drop the caller's refcount
 331  * on them.  Reinitialises the caller's pagevec.
 332  */
 333 void __pagevec_lru_add(struct pagevec *pvec)
 334 {
 335         int i;
 336         struct zone *zone = NULL;
 337
 338         for (i = 0; i < pagevec_count(pvec); i++) {
 339                 struct page *page = pvec->pages[i];
 340                 struct zone *pagezone = page_zone(page);
 341
 342                 if (pagezone != zone) {
 343                         if (zone)
 344                                 spin_unlock_irq(&zone->lru_lock);
 345                         zone = pagezone;
 346                         spin_lock_irq(&zone->lru_lock);
 347                 }
 348                 BUG_ON(PageLRU(page));
 349                 SetPageLRU(page);
 350                 add_page_to_inactive_list(zone, page);
 351         }
 352         if (zone)
 353                 spin_unlock_irq(&zone->lru_lock);
 354         release_pages(pvec->pages, pvec->nr, pvec->cold);
 355         pagevec_reinit(pvec);
 356 }
 357
 358 EXPORT_SYMBOL(__pagevec_lru_add);
 359
 360 void __pagevec_lru_add_active(struct pagevec *pvec)
 361 {
 362         int i;
 363         struct zone *zone = NULL;
 364
 365         for (i = 0; i < pagevec_count(pvec); i++) {
 366                 struct page *page = pvec->pages[i];
 367                 struct zone *pagezone = page_zone(page);
 368
 369                 if (pagezone != zone) {
 370                         if (zone)
 371                                 spin_unlock_irq(&zone->lru_lock);
 372                         zone = pagezone;
 373                         spin_lock_irq(&zone->lru_lock);
 374                 }
 375                 BUG_ON(PageLRU(page));
 376                 SetPageLRU(page);
 377                 BUG_ON(PageActive(page));
 378                 SetPageActive(page);
 379                 add_page_to_active_list(zone, page);
 380         }
 381         if (zone)
 382                 spin_unlock_irq(&zone->lru_lock);
 383         release_pages(pvec->pages, pvec->nr, pvec->cold);
 384         pagevec_reinit(pvec);
 385 }
 386
 387 /*
 388  * Try to drop buffers from the pages in a pagevec
 389  */
 390 void pagevec_strip(struct pagevec *pvec)
 391 {
 392         int i;
 393
 394         for (i = 0; i < pagevec_count(pvec); i++) {
 395                 struct page *page = pvec->pages[i];
 396
 397                 if (PagePrivate(page) && !TestSetPageLocked(page)) {
 398                         if (PagePrivate(page))
 399                                 try_to_release_page(page, 0);
 400                         unlock_page(page);
 401                 }
 402         }
 403 }
 404
 405 /**
 406  * pagevec_lookup - gang pagecache lookup
 407  * @pvec:       Where the resulting pages are placed
 408  * @mapping:    The address_space to search
 409  * @start:      The starting page index
 410  * @nr_pages:   The maximum number of pages
 411  *
 412  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 413  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 414  * reference against the pages in @pvec.
 415  *
 416  * The search returns a group of mapping-contiguous pages with ascending
 417  * indexes.  There may be holes in the indices due to not-present pages.
 418  *
 419  * pagevec_lookup() returns the number of pages which were found.
 420  */
 421 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 422                 pgoff_t start, unsigned nr_pages)
 423 {
 424         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 425         return pagevec_count(pvec);
 426 }
 427
 428 EXPORT_SYMBOL(pagevec_lookup);
 429
 430 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 431                 pgoff_t *index, int tag, unsigned nr_pages)
 432 {
 433         pvec->nr = find_get_pages_tag(mapping, index, tag,
 434                                         nr_pages, pvec->pages);
 435         return pagevec_count(pvec);
 436 }
 437
 438 EXPORT_SYMBOL(pagevec_lookup_tag);
 439
 440 #ifdef CONFIG_SMP
 441 /*
 442  * We tolerate a little inaccuracy to avoid ping-ponging the counter between
 443  * CPUs
 444  */
 445 #define ACCT_THRESHOLD  max(16, NR_CPUS * 2)
 446
 447 static DEFINE_PER_CPU(long, committed_space) = 0;
 448
 449 void vm_acct_memory(long pages)
 450 {
 451         long *local;
 452
 453         preempt_disable();
 454         local = &__get_cpu_var(committed_space);
 455         *local += pages;
 456         if (*local > ACCT_THRESHOLD || *local < -ACCT_THRESHOLD) {
 457                 atomic_add(*local, &vm_committed_space);
 458                 *local = 0;
 459         }
 460         preempt_enable();
 461 }
 462
 463 #ifdef CONFIG_HOTPLUG_CPU
 464
 465 /* Drop the CPU's cached committed space back into the central pool. */
 466 static int cpu_swap_callback(struct notifier_block *nfb,
 467                              unsigned long action,
 468                              void *hcpu)
 469 {
 470         long *committed;
 471
 472         committed = &per_cpu(committed_space, (long)hcpu);
 473         if (action == CPU_DEAD) {
 474                 atomic_add(*committed, &vm_committed_space);
 475                 *committed = 0;
 476                 __lru_add_drain((long)hcpu);
 477         }
 478         return NOTIFY_OK;
 479 }
 480 #endif /* CONFIG_HOTPLUG_CPU */
 481 #endif /* CONFIG_SMP */
 482
 483 #ifdef CONFIG_SMP
 484 void percpu_counter_mod(struct percpu_counter *fbc, long amount)
 485 {
 486         long count;
 487         long *pcount;
 488         int cpu = get_cpu();
 489
 490         pcount = per_cpu_ptr(fbc->counters, cpu);
 491         count = *pcount + amount;
 492         if (count >= FBC_BATCH || count <= -FBC_BATCH) {
 493                 spin_lock(&fbc->lock);
 494                 fbc->count += count;
 495                 *pcount = 0;
 496                 spin_unlock(&fbc->lock);
 497         } else {
 498                 *pcount = count;
 499         }
 500         put_cpu();
 501 }
 502 EXPORT_SYMBOL(percpu_counter_mod);
 503
 504 /*
 505  * Add up all the per-cpu counts, return the result.  This is a more accurate
 506  * but much slower version of percpu_counter_read_positive()
 507  */
 508 long percpu_counter_sum(struct percpu_counter *fbc)
 509 {
 510         long ret;
 511         int cpu;
 512
 513         spin_lock(&fbc->lock);
 514         ret = fbc->count;
 515         for_each_possible_cpu(cpu) {
 516                 long *pcount = per_cpu_ptr(fbc->counters, cpu);
 517                 ret += *pcount;
 518         }
 519         spin_unlock(&fbc->lock);
 520         return ret < 0 ? 0 : ret;
 521 }
 522 EXPORT_SYMBOL(percpu_counter_sum);
 523 #endif
 524
 525 /*
 526  * Perform any setup for the swap system
 527  */
 528 void __init swap_setup(void)
 529 {
 530         unsigned long megs = num_physpages >> (20 - PAGE_SHIFT);
 531
 532         /* Use a smaller cluster for small-memory machines */
 533         if (megs < 16)
 534                 page_cluster = 2;
 535         else
 536                 page_cluster = 3;
 537         /*
 538          * Right now other parts of the system means that we
 539          * _really_ don't want to cluster much more
 540          */
 541         hotcpu_notifier(cpu_swap_callback, 0);
 542 }