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