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 operation 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/export.h>
  25 #include <linux/mm_inline.h>
  26 #include <linux/percpu_counter.h>
  27 #include <linux/percpu.h>
  28 #include <linux/cpu.h>
  29 #include <linux/notifier.h>
  30 #include <linux/backing-dev.h>
  31 #include <linux/memcontrol.h>
  32 #include <linux/gfp.h>
  33 #include <linux/uio.h>
  34 #include <linux/hugetlb.h>
  35
  36 #include "internal.h"
  37
  38 #define CREATE_TRACE_POINTS
  39 #include <trace/events/pagemap.h>
  40
  41 /* How many pages do we try to swap or page in/out together? */
  42 int page_cluster;
  43
  44 static DEFINE_PER_CPU(struct pagevec, lru_add_pvec);
  45 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
  46 static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
  47
  48 /*
  49  * This path almost never happens for VM activity - pages are normally
  50  * freed via pagevecs.  But it gets used by networking.
  51  */
  52 static void __page_cache_release(struct page *page)
  53 {
  54         if (PageLRU(page)) {
  55                 struct zone *zone = page_zone(page);
  56                 struct lruvec *lruvec;
  57                 unsigned long flags;
  58
  59                 spin_lock_irqsave(&zone->lru_lock, flags);
  60                 lruvec = mem_cgroup_page_lruvec(page, zone);
  61                 VM_BUG_ON(!PageLRU(page));
  62                 __ClearPageLRU(page);
  63                 del_page_from_lru_list(page, lruvec, page_off_lru(page));
  64                 spin_unlock_irqrestore(&zone->lru_lock, flags);
  65         }
  66 }
  67
  68 static void __put_single_page(struct page *page)
  69 {
  70         __page_cache_release(page);
  71         free_hot_cold_page(page, 0);
  72 }
  73
  74 static void __put_compound_page(struct page *page)
  75 {
  76         compound_page_dtor *dtor;
  77
  78         __page_cache_release(page);
  79         dtor = get_compound_page_dtor(page);
  80         (*dtor)(page);
  81 }
  82
  83 static void put_compound_page(struct page *page)
  84 {
  85         if (unlikely(PageTail(page))) {
  86                 /* __split_huge_page_refcount can run under us */
  87                 struct page *page_head = compound_trans_head(page);
  88
  89                 if (likely(page != page_head &&
  90                            get_page_unless_zero(page_head))) {
  91                         unsigned long flags;
  92
  93                         /*
  94                          * THP can not break up slab pages so avoid taking
  95                          * compound_lock().  Slab performs non-atomic bit ops
  96                          * on page->flags for better performance.  In particular
  97                          * slab_unlock() in slub used to be a hot path.  It is
  98                          * still hot on arches that do not support
  99                          * this_cpu_cmpxchg_double().
 100                          */
 101                         if (PageSlab(page_head) || PageHeadHuge(page_head)) {
 102                                 if (likely(PageTail(page))) {
 103                                         /*
 104                                          * __split_huge_page_refcount
 105                                          * cannot race here.
 106                                          */
 107                                         VM_BUG_ON(!PageHead(page_head));
 108                                         atomic_dec(&page->_mapcount);
 109                                         if (put_page_testzero(page_head))
 110                                                 VM_BUG_ON(1);
 111                                         if (put_page_testzero(page_head))
 112                                                 __put_compound_page(page_head);
 113                                         return;
 114                                 } else
 115                                         /*
 116                                          * __split_huge_page_refcount
 117                                          * run before us, "page" was a
 118                                          * THP tail. The split
 119                                          * page_head has been freed
 120                                          * and reallocated as slab or
 121                                          * hugetlbfs page of smaller
 122                                          * order (only possible if
 123                                          * reallocated as slab on
 124                                          * x86).
 125                                          */
 126                                         goto skip_lock;
 127                         }
 128                         /*
 129                          * page_head wasn't a dangling pointer but it
 130                          * may not be a head page anymore by the time
 131                          * we obtain the lock. That is ok as long as it
 132                          * can't be freed from under us.
 133                          */
 134                         flags = compound_lock_irqsave(page_head);
 135                         if (unlikely(!PageTail(page))) {
 136                                 /* __split_huge_page_refcount run before us */
 137                                 compound_unlock_irqrestore(page_head, flags);
 138 skip_lock:
 139                                 if (put_page_testzero(page_head)) {
 140                                         /*
 141                                          * The head page may have been
 142                                          * freed and reallocated as a
 143                                          * compound page of smaller
 144                                          * order and then freed again.
 145                                          * All we know is that it
 146                                          * cannot have become: a THP
 147                                          * page, a compound page of
 148                                          * higher order, a tail page.
 149                                          * That is because we still
 150                                          * hold the refcount of the
 151                                          * split THP tail and
 152                                          * page_head was the THP head
 153                                          * before the split.
 154                                          */
 155                                         if (PageHead(page_head))
 156                                                 __put_compound_page(page_head);
 157                                         else
 158                                                 __put_single_page(page_head);
 159                                 }
 160 out_put_single:
 161                                 if (put_page_testzero(page))
 162                                         __put_single_page(page);
 163                                 return;
 164                         }
 165                         VM_BUG_ON(page_head != page->first_page);
 166                         /*
 167                          * We can release the refcount taken by
 168                          * get_page_unless_zero() now that
 169                          * __split_huge_page_refcount() is blocked on
 170                          * the compound_lock.
 171                          */
 172                         if (put_page_testzero(page_head))
 173                                 VM_BUG_ON(1);
 174                         /* __split_huge_page_refcount will wait now */
 175                         VM_BUG_ON(page_mapcount(page) <= 0);
 176                         atomic_dec(&page->_mapcount);
 177                         VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
 178                         VM_BUG_ON(atomic_read(&page->_count) != 0);
 179                         compound_unlock_irqrestore(page_head, flags);
 180
 181                         if (put_page_testzero(page_head)) {
 182                                 if (PageHead(page_head))
 183                                         __put_compound_page(page_head);
 184                                 else
 185                                         __put_single_page(page_head);
 186                         }
 187                 } else {
 188                         /* page_head is a dangling pointer */
 189                         VM_BUG_ON(PageTail(page));
 190                         goto out_put_single;
 191                 }
 192         } else if (put_page_testzero(page)) {
 193                 if (PageHead(page))
 194                         __put_compound_page(page);
 195                 else
 196                         __put_single_page(page);
 197         }
 198 }
 199
 200 void put_page(struct page *page)
 201 {
 202         if (unlikely(PageCompound(page)))
 203                 put_compound_page(page);
 204         else if (put_page_testzero(page))
 205                 __put_single_page(page);
 206 }
 207 EXPORT_SYMBOL(put_page);
 208
 209 /*
 210  * This function is exported but must not be called by anything other
 211  * than get_page(). It implements the slow path of get_page().
 212  */
 213 bool __get_page_tail(struct page *page)
 214 {
 215         /*
 216          * This takes care of get_page() if run on a tail page
 217          * returned by one of the get_user_pages/follow_page variants.
 218          * get_user_pages/follow_page itself doesn't need the compound
 219          * lock because it runs __get_page_tail_foll() under the
 220          * proper PT lock that already serializes against
 221          * split_huge_page().
 222          */
 223         unsigned long flags;
 224         bool got = false;
 225         struct page *page_head = compound_trans_head(page);
 226
 227         if (likely(page != page_head && get_page_unless_zero(page_head))) {
 228                 /* Ref to put_compound_page() comment. */
 229                 if (PageSlab(page_head) || PageHeadHuge(page_head)) {
 230                         if (likely(PageTail(page))) {
 231                                 /*
 232                                  * This is a hugetlbfs page or a slab
 233                                  * page. __split_huge_page_refcount
 234                                  * cannot race here.
 235                                  */
 236                                 VM_BUG_ON(!PageHead(page_head));
 237                                 __get_page_tail_foll(page, false);
 238                                 return true;
 239                         } else {
 240                                 /*
 241                                  * __split_huge_page_refcount run
 242                                  * before us, "page" was a THP
 243                                  * tail. The split page_head has been
 244                                  * freed and reallocated as slab or
 245                                  * hugetlbfs page of smaller order
 246                                  * (only possible if reallocated as
 247                                  * slab on x86).
 248                                  */
 249                                 put_page(page_head);
 250                                 return false;
 251                         }
 252                 }
 253
 254                 /*
 255                  * page_head wasn't a dangling pointer but it
 256                  * may not be a head page anymore by the time
 257                  * we obtain the lock. That is ok as long as it
 258                  * can't be freed from under us.
 259                  */
 260                 flags = compound_lock_irqsave(page_head);
 261                 /* here __split_huge_page_refcount won't run anymore */
 262                 if (likely(PageTail(page))) {
 263                         __get_page_tail_foll(page, false);
 264                         got = true;
 265                 }
 266                 compound_unlock_irqrestore(page_head, flags);
 267                 if (unlikely(!got))
 268                         put_page(page_head);
 269         }
 270         return got;
 271 }
 272 EXPORT_SYMBOL(__get_page_tail);
 273
 274 /**
 275  * put_pages_list() - release a list of pages
 276  * @pages: list of pages threaded on page->lru
 277  *
 278  * Release a list of pages which are strung together on page.lru.  Currently
 279  * used by read_cache_pages() and related error recovery code.
 280  */
 281 void put_pages_list(struct list_head *pages)
 282 {
 283         while (!list_empty(pages)) {
 284                 struct page *victim;
 285
 286                 victim = list_entry(pages->prev, struct page, lru);
 287                 list_del(&victim->lru);
 288                 page_cache_release(victim);
 289         }
 290 }
 291 EXPORT_SYMBOL(put_pages_list);
 292
 293 /*
 294  * get_kernel_pages() - pin kernel pages in memory
 295  * @kiov:       An array of struct kvec structures
 296  * @nr_segs:    number of segments to pin
 297  * @write:      pinning for read/write, currently ignored
 298  * @pages:      array that receives pointers to the pages pinned.
 299  *              Should be at least nr_segs long.
 300  *
 301  * Returns number of pages pinned. This may be fewer than the number
 302  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 303  * were pinned, returns -errno. Each page returned must be released
 304  * with a put_page() call when it is finished with.
 305  */
 306 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
 307                 struct page **pages)
 308 {
 309         int seg;
 310
 311         for (seg = 0; seg < nr_segs; seg++) {
 312                 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
 313                         return seg;
 314
 315                 pages[seg] = kmap_to_page(kiov[seg].iov_base);
 316                 page_cache_get(pages[seg]);
 317         }
 318
 319         return seg;
 320 }
 321 EXPORT_SYMBOL_GPL(get_kernel_pages);
 322
 323 /*
 324  * get_kernel_page() - pin a kernel page in memory
 325  * @start:      starting kernel address
 326  * @write:      pinning for read/write, currently ignored
 327  * @pages:      array that receives pointer to the page pinned.
 328  *              Must be at least nr_segs long.
 329  *
 330  * Returns 1 if page is pinned. If the page was not pinned, returns
 331  * -errno. The page returned must be released with a put_page() call
 332  * when it is finished with.
 333  */
 334 int get_kernel_page(unsigned long start, int write, struct page **pages)
 335 {
 336         const struct kvec kiov = {
 337                 .iov_base = (void *)start,
 338                 .iov_len = PAGE_SIZE
 339         };
 340
 341         return get_kernel_pages(&kiov, 1, write, pages);
 342 }
 343 EXPORT_SYMBOL_GPL(get_kernel_page);
 344
 345 static void pagevec_lru_move_fn(struct pagevec *pvec,
 346         void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
 347         void *arg)
 348 {
 349         int i;
 350         struct zone *zone = NULL;
 351         struct lruvec *lruvec;
 352         unsigned long flags = 0;
 353
 354         for (i = 0; i < pagevec_count(pvec); i++) {
 355                 struct page *page = pvec->pages[i];
 356                 struct zone *pagezone = page_zone(page);
 357
 358                 if (pagezone != zone) {
 359                         if (zone)
 360                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 361                         zone = pagezone;
 362                         spin_lock_irqsave(&zone->lru_lock, flags);
 363                 }
 364
 365                 lruvec = mem_cgroup_page_lruvec(page, zone);
 366                 (*move_fn)(page, lruvec, arg);
 367         }
 368         if (zone)
 369                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 370         release_pages(pvec->pages, pvec->nr, pvec->cold);
 371         pagevec_reinit(pvec);
 372 }
 373
 374 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
 375                                  void *arg)
 376 {
 377         int *pgmoved = arg;
 378
 379         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 380                 enum lru_list lru = page_lru_base_type(page);
 381                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 382                 (*pgmoved)++;
 383         }
 384 }
 385
 386 /*
 387  * pagevec_move_tail() must be called with IRQ disabled.
 388  * Otherwise this may cause nasty races.
 389  */
 390 static void pagevec_move_tail(struct pagevec *pvec)
 391 {
 392         int pgmoved = 0;
 393
 394         pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
 395         __count_vm_events(PGROTATED, pgmoved);
 396 }
 397
 398 /*
 399  * Writeback is about to end against a page which has been marked for immediate
 400  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 401  * inactive list.
 402  */
 403 void rotate_reclaimable_page(struct page *page)
 404 {
 405         if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
 406             !PageUnevictable(page) && PageLRU(page)) {
 407                 struct pagevec *pvec;
 408                 unsigned long flags;
 409
 410                 page_cache_get(page);
 411                 local_irq_save(flags);
 412                 pvec = &__get_cpu_var(lru_rotate_pvecs);
 413                 if (!pagevec_add(pvec, page))
 414                         pagevec_move_tail(pvec);
 415                 local_irq_restore(flags);
 416         }
 417 }
 418
 419 static void update_page_reclaim_stat(struct lruvec *lruvec,
 420                                      int file, int rotated)
 421 {
 422         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
 423
 424         reclaim_stat->recent_scanned[file]++;
 425         if (rotated)
 426                 reclaim_stat->recent_rotated[file]++;
 427 }
 428
 429 static void __activate_page(struct page *page, struct lruvec *lruvec,
 430                             void *arg)
 431 {
 432         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 433                 int file = page_is_file_cache(page);
 434                 int lru = page_lru_base_type(page);
 435
 436                 del_page_from_lru_list(page, lruvec, lru);
 437                 SetPageActive(page);
 438                 lru += LRU_ACTIVE;
 439                 add_page_to_lru_list(page, lruvec, lru);
 440                 trace_mm_lru_activate(page, page_to_pfn(page));
 441
 442                 __count_vm_event(PGACTIVATE);
 443                 update_page_reclaim_stat(lruvec, file, 1);
 444         }
 445 }
 446
 447 #ifdef CONFIG_SMP
 448 static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
 449
 450 static void activate_page_drain(int cpu)
 451 {
 452         struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
 453
 454         if (pagevec_count(pvec))
 455                 pagevec_lru_move_fn(pvec, __activate_page, NULL);
 456 }
 457
 458 static bool need_activate_page_drain(int cpu)
 459 {
 460         return pagevec_count(&per_cpu(activate_page_pvecs, cpu)) != 0;
 461 }
 462
 463 void activate_page(struct page *page)
 464 {
 465         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 466                 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
 467
 468                 page_cache_get(page);
 469                 if (!pagevec_add(pvec, page))
 470                         pagevec_lru_move_fn(pvec, __activate_page, NULL);
 471                 put_cpu_var(activate_page_pvecs);
 472         }
 473 }
 474
 475 #else
 476 static inline void activate_page_drain(int cpu)
 477 {
 478 }
 479
 480 static bool need_activate_page_drain(int cpu)
 481 {
 482         return false;
 483 }
 484
 485 void activate_page(struct page *page)
 486 {
 487         struct zone *zone = page_zone(page);
 488
 489         spin_lock_irq(&zone->lru_lock);
 490         __activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
 491         spin_unlock_irq(&zone->lru_lock);
 492 }
 493 #endif
 494
 495 static void __lru_cache_activate_page(struct page *page)
 496 {
 497         struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
 498         int i;
 499
 500         /*
 501          * Search backwards on the optimistic assumption that the page being
 502          * activated has just been added to this pagevec. Note that only
 503          * the local pagevec is examined as a !PageLRU page could be in the
 504          * process of being released, reclaimed, migrated or on a remote
 505          * pagevec that is currently being drained. Furthermore, marking
 506          * a remote pagevec's page PageActive potentially hits a race where
 507          * a page is marked PageActive just after it is added to the inactive
 508          * list causing accounting errors and BUG_ON checks to trigger.
 509          */
 510         for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
 511                 struct page *pagevec_page = pvec->pages[i];
 512
 513                 if (pagevec_page == page) {
 514                         SetPageActive(page);
 515                         break;
 516                 }
 517         }
 518
 519         put_cpu_var(lru_add_pvec);
 520 }
 521
 522 /*
 523  * Mark a page as having seen activity.
 524  *
 525  * inactive,unreferenced        ->      inactive,referenced
 526  * inactive,referenced          ->      active,unreferenced
 527  * active,unreferenced          ->      active,referenced
 528  */
 529 void mark_page_accessed(struct page *page)
 530 {
 531         if (!PageActive(page) && !PageUnevictable(page) &&
 532                         PageReferenced(page)) {
 533
 534                 /*
 535                  * If the page is on the LRU, queue it for activation via
 536                  * activate_page_pvecs. Otherwise, assume the page is on a
 537                  * pagevec, mark it active and it'll be moved to the active
 538                  * LRU on the next drain.
 539                  */
 540                 if (PageLRU(page))
 541                         activate_page(page);
 542                 else
 543                         __lru_cache_activate_page(page);
 544                 ClearPageReferenced(page);
 545         } else if (!PageReferenced(page)) {
 546                 SetPageReferenced(page);
 547         }
 548 }
 549 EXPORT_SYMBOL(mark_page_accessed);
 550
 551 /*
 552  * Queue the page for addition to the LRU via pagevec. The decision on whether
 553  * to add the page to the [in]active [file|anon] list is deferred until the
 554  * pagevec is drained. This gives a chance for the caller of __lru_cache_add()
 555  * have the page added to the active list using mark_page_accessed().
 556  */
 557 void __lru_cache_add(struct page *page)
 558 {
 559         struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
 560
 561         page_cache_get(page);
 562         if (!pagevec_space(pvec))
 563                 __pagevec_lru_add(pvec);
 564         pagevec_add(pvec, page);
 565         put_cpu_var(lru_add_pvec);
 566 }
 567 EXPORT_SYMBOL(__lru_cache_add);
 568
 569 /**
 570  * lru_cache_add - add a page to a page list
 571  * @page: the page to be added to the LRU.
 572  */
 573 void lru_cache_add(struct page *page)
 574 {
 575         VM_BUG_ON(PageActive(page) && PageUnevictable(page));
 576         VM_BUG_ON(PageLRU(page));
 577         __lru_cache_add(page);
 578 }
 579
 580 /**
 581  * add_page_to_unevictable_list - add a page to the unevictable list
 582  * @page:  the page to be added to the unevictable list
 583  *
 584  * Add page directly to its zone's unevictable list.  To avoid races with
 585  * tasks that might be making the page evictable, through eg. munlock,
 586  * munmap or exit, while it's not on the lru, we want to add the page
 587  * while it's locked or otherwise "invisible" to other tasks.  This is
 588  * difficult to do when using the pagevec cache, so bypass that.
 589  */
 590 void add_page_to_unevictable_list(struct page *page)
 591 {
 592         struct zone *zone = page_zone(page);
 593         struct lruvec *lruvec;
 594
 595         spin_lock_irq(&zone->lru_lock);
 596         lruvec = mem_cgroup_page_lruvec(page, zone);
 597         ClearPageActive(page);
 598         SetPageUnevictable(page);
 599         SetPageLRU(page);
 600         add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
 601         spin_unlock_irq(&zone->lru_lock);
 602 }
 603
 604 /*
 605  * If the page can not be invalidated, it is moved to the
 606  * inactive list to speed up its reclaim.  It is moved to the
 607  * head of the list, rather than the tail, to give the flusher
 608  * threads some time to write it out, as this is much more
 609  * effective than the single-page writeout from reclaim.
 610  *
 611  * If the page isn't page_mapped and dirty/writeback, the page
 612  * could reclaim asap using PG_reclaim.
 613  *
 614  * 1. active, mapped page -> none
 615  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 616  * 3. inactive, mapped page -> none
 617  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 618  * 5. inactive, clean -> inactive, tail
 619  * 6. Others -> none
 620  *
 621  * In 4, why it moves inactive's head, the VM expects the page would
 622  * be write it out by flusher threads as this is much more effective
 623  * than the single-page writeout from reclaim.
 624  */
 625 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 626                               void *arg)
 627 {
 628         int lru, file;
 629         bool active;
 630
 631         if (!PageLRU(page))
 632                 return;
 633
 634         if (PageUnevictable(page))
 635                 return;
 636
 637         /* Some processes are using the page */
 638         if (page_mapped(page))
 639                 return;
 640
 641         active = PageActive(page);
 642         file = page_is_file_cache(page);
 643         lru = page_lru_base_type(page);
 644
 645         del_page_from_lru_list(page, lruvec, lru + active);
 646         ClearPageActive(page);
 647         ClearPageReferenced(page);
 648         add_page_to_lru_list(page, lruvec, lru);
 649
 650         if (PageWriteback(page) || PageDirty(page)) {
 651                 /*
 652                  * PG_reclaim could be raced with end_page_writeback
 653                  * It can make readahead confusing.  But race window
 654                  * is _really_ small and  it's non-critical problem.
 655                  */
 656                 SetPageReclaim(page);
 657         } else {
 658                 /*
 659                  * The page's writeback ends up during pagevec
 660                  * We moves tha page into tail of inactive.
 661                  */
 662                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 663                 __count_vm_event(PGROTATED);
 664         }
 665
 666         if (active)
 667                 __count_vm_event(PGDEACTIVATE);
 668         update_page_reclaim_stat(lruvec, file, 0);
 669 }
 670
 671 /*
 672  * Drain pages out of the cpu's pagevecs.
 673  * Either "cpu" is the current CPU, and preemption has already been
 674  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 675  */
 676 void lru_add_drain_cpu(int cpu)
 677 {
 678         struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
 679
 680         if (pagevec_count(pvec))
 681                 __pagevec_lru_add(pvec);
 682
 683         pvec = &per_cpu(lru_rotate_pvecs, cpu);
 684         if (pagevec_count(pvec)) {
 685                 unsigned long flags;
 686
 687                 /* No harm done if a racing interrupt already did this */
 688                 local_irq_save(flags);
 689                 pagevec_move_tail(pvec);
 690                 local_irq_restore(flags);
 691         }
 692
 693         pvec = &per_cpu(lru_deactivate_pvecs, cpu);
 694         if (pagevec_count(pvec))
 695                 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 696
 697         activate_page_drain(cpu);
 698 }
 699
 700 /**
 701  * deactivate_page - forcefully deactivate a page
 702  * @page: page to deactivate
 703  *
 704  * This function hints the VM that @page is a good reclaim candidate,
 705  * for example if its invalidation fails due to the page being dirty
 706  * or under writeback.
 707  */
 708 void deactivate_page(struct page *page)
 709 {
 710         /*
 711          * In a workload with many unevictable page such as mprotect, unevictable
 712          * page deactivation for accelerating reclaim is pointless.
 713          */
 714         if (PageUnevictable(page))
 715                 return;
 716
 717         if (likely(get_page_unless_zero(page))) {
 718                 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
 719
 720                 if (!pagevec_add(pvec, page))
 721                         pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 722                 put_cpu_var(lru_deactivate_pvecs);
 723         }
 724 }
 725
 726 void lru_add_drain(void)
 727 {
 728         lru_add_drain_cpu(get_cpu());
 729         put_cpu();
 730 }
 731
 732 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 733 {
 734         lru_add_drain();
 735 }
 736
 737 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 738
 739 void lru_add_drain_all(void)
 740 {
 741         static DEFINE_MUTEX(lock);
 742         static struct cpumask has_work;
 743         int cpu;
 744
 745         mutex_lock(&lock);
 746         get_online_cpus();
 747         cpumask_clear(&has_work);
 748
 749         for_each_online_cpu(cpu) {
 750                 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 751
 752                 if (pagevec_count(&per_cpu(lru_add_pvec, cpu)) ||
 753                     pagevec_count(&per_cpu(lru_rotate_pvecs, cpu)) ||
 754                     pagevec_count(&per_cpu(lru_deactivate_pvecs, cpu)) ||
 755                     need_activate_page_drain(cpu)) {
 756                         INIT_WORK(work, lru_add_drain_per_cpu);
 757                         schedule_work_on(cpu, work);
 758                         cpumask_set_cpu(cpu, &has_work);
 759                 }
 760         }
 761
 762         for_each_cpu(cpu, &has_work)
 763                 flush_work(&per_cpu(lru_add_drain_work, cpu));
 764
 765         put_online_cpus();
 766         mutex_unlock(&lock);
 767 }
 768
 769 /*
 770  * Batched page_cache_release().  Decrement the reference count on all the
 771  * passed pages.  If it fell to zero then remove the page from the LRU and
 772  * free it.
 773  *
 774  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 775  * for the remainder of the operation.
 776  *
 777  * The locking in this function is against shrink_inactive_list(): we recheck
 778  * the page count inside the lock to see whether shrink_inactive_list()
 779  * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
 780  * will free it.
 781  */
 782 void release_pages(struct page **pages, int nr, int cold)
 783 {
 784         int i;
 785         LIST_HEAD(pages_to_free);
 786         struct zone *zone = NULL;
 787         struct lruvec *lruvec;
 788         unsigned long uninitialized_var(flags);
 789
 790         for (i = 0; i < nr; i++) {
 791                 struct page *page = pages[i];
 792
 793                 if (unlikely(PageCompound(page))) {
 794                         if (zone) {
 795                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 796                                 zone = NULL;
 797                         }
 798                         put_compound_page(page);
 799                         continue;
 800                 }
 801
 802                 if (!put_page_testzero(page))
 803                         continue;
 804
 805                 if (PageLRU(page)) {
 806                         struct zone *pagezone = page_zone(page);
 807
 808                         if (pagezone != zone) {
 809                                 if (zone)
 810                                         spin_unlock_irqrestore(&zone->lru_lock,
 811                                                                         flags);
 812                                 zone = pagezone;
 813                                 spin_lock_irqsave(&zone->lru_lock, flags);
 814                         }
 815
 816                         lruvec = mem_cgroup_page_lruvec(page, zone);
 817                         VM_BUG_ON(!PageLRU(page));
 818                         __ClearPageLRU(page);
 819                         del_page_from_lru_list(page, lruvec, page_off_lru(page));
 820                 }
 821
 822                 /* Clear Active bit in case of parallel mark_page_accessed */
 823                 ClearPageActive(page);
 824
 825                 list_add(&page->lru, &pages_to_free);
 826         }
 827         if (zone)
 828                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 829
 830         free_hot_cold_page_list(&pages_to_free, cold);
 831 }
 832 EXPORT_SYMBOL(release_pages);
 833
 834 /*
 835  * The pages which we're about to release may be in the deferred lru-addition
 836  * queues.  That would prevent them from really being freed right now.  That's
 837  * OK from a correctness point of view but is inefficient - those pages may be
 838  * cache-warm and we want to give them back to the page allocator ASAP.
 839  *
 840  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 841  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 842  * mutual recursion.
 843  */
 844 void __pagevec_release(struct pagevec *pvec)
 845 {
 846         lru_add_drain();
 847         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 848         pagevec_reinit(pvec);
 849 }
 850 EXPORT_SYMBOL(__pagevec_release);
 851
 852 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 853 /* used by __split_huge_page_refcount() */
 854 void lru_add_page_tail(struct page *page, struct page *page_tail,
 855                        struct lruvec *lruvec, struct list_head *list)
 856 {
 857         const int file = 0;
 858
 859         VM_BUG_ON(!PageHead(page));
 860         VM_BUG_ON(PageCompound(page_tail));
 861         VM_BUG_ON(PageLRU(page_tail));
 862         VM_BUG_ON(NR_CPUS != 1 &&
 863                   !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
 864
 865         if (!list)
 866                 SetPageLRU(page_tail);
 867
 868         if (likely(PageLRU(page)))
 869                 list_add_tail(&page_tail->lru, &page->lru);
 870         else if (list) {
 871                 /* page reclaim is reclaiming a huge page */
 872                 get_page(page_tail);
 873                 list_add_tail(&page_tail->lru, list);
 874         } else {
 875                 struct list_head *list_head;
 876                 /*
 877                  * Head page has not yet been counted, as an hpage,
 878                  * so we must account for each subpage individually.
 879                  *
 880                  * Use the standard add function to put page_tail on the list,
 881                  * but then correct its position so they all end up in order.
 882                  */
 883                 add_page_to_lru_list(page_tail, lruvec, page_lru(page_tail));
 884                 list_head = page_tail->lru.prev;
 885                 list_move_tail(&page_tail->lru, list_head);
 886         }
 887
 888         if (!PageUnevictable(page))
 889                 update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
 890 }
 891 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 892
 893 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
 894                                  void *arg)
 895 {
 896         int file = page_is_file_cache(page);
 897         int active = PageActive(page);
 898         enum lru_list lru = page_lru(page);
 899
 900         VM_BUG_ON(PageLRU(page));
 901
 902         SetPageLRU(page);
 903         add_page_to_lru_list(page, lruvec, lru);
 904         update_page_reclaim_stat(lruvec, file, active);
 905         trace_mm_lru_insertion(page, page_to_pfn(page), lru, trace_pagemap_flags(page));
 906 }
 907
 908 /*
 909  * Add the passed pages to the LRU, then drop the caller's refcount
 910  * on them.  Reinitialises the caller's pagevec.
 911  */
 912 void __pagevec_lru_add(struct pagevec *pvec)
 913 {
 914         pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
 915 }
 916 EXPORT_SYMBOL(__pagevec_lru_add);
 917
 918 /**
 919  * pagevec_lookup - gang pagecache lookup
 920  * @pvec:       Where the resulting pages are placed
 921  * @mapping:    The address_space to search
 922  * @start:      The starting page index
 923  * @nr_pages:   The maximum number of pages
 924  *
 925  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 926  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 927  * reference against the pages in @pvec.
 928  *
 929  * The search returns a group of mapping-contiguous pages with ascending
 930  * indexes.  There may be holes in the indices due to not-present pages.
 931  *
 932  * pagevec_lookup() returns the number of pages which were found.
 933  */
 934 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 935                 pgoff_t start, unsigned nr_pages)
 936 {
 937         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 938         return pagevec_count(pvec);
 939 }
 940 EXPORT_SYMBOL(pagevec_lookup);
 941
 942 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 943                 pgoff_t *index, int tag, unsigned nr_pages)
 944 {
 945         pvec->nr = find_get_pages_tag(mapping, index, tag,
 946                                         nr_pages, pvec->pages);
 947         return pagevec_count(pvec);
 948 }
 949 EXPORT_SYMBOL(pagevec_lookup_tag);
 950
 951 /*
 952  * Perform any setup for the swap system
 953  */
 954 void __init swap_setup(void)
 955 {
 956         unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
 957 #ifdef CONFIG_SWAP
 958         int i;
 959
 960         if (bdi_init(swapper_spaces[0].backing_dev_info))
 961                 panic("Failed to init swap bdi");
 962         for (i = 0; i < MAX_SWAPFILES; i++) {
 963                 spin_lock_init(&swapper_spaces[i].tree_lock);
 964                 INIT_LIST_HEAD(&swapper_spaces[i].i_mmap_nonlinear);
 965         }
 966 #endif
 967
 968         /* Use a smaller cluster for small-memory machines */
 969         if (megs < 16)
 970                 page_cluster = 2;
 971         else
 972                 page_cluster = 3;
 973         /*
 974          * Right now other parts of the system means that we
 975          * _really_ don't want to cluster much more
 976          */
 977 }