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                 /*
  90                  * THP can not break up slab pages so avoid taking
  91                  * compound_lock(). Slab performs non-atomic bit ops
  92                  * on page->flags for better performance. In
  93                  * particular slab_unlock() in slub used to be a hot
  94                  * path. It is still hot on arches that do not support
  95                  * this_cpu_cmpxchg_double().
  96                  *
  97                  * If "page" is part of a slab or hugetlbfs page it
  98                  * cannot be splitted and the head page cannot change
  99                  * from under us. And if "page" is part of a THP page
 100                  * under splitting, if the head page pointed by the
 101                  * THP tail isn't a THP head anymore, we'll find
 102                  * PageTail clear after smp_rmb() and we'll treat it
 103                  * as a single page.
 104                  */
 105                 if (PageSlab(page_head) || PageHeadHuge(page_head)) {
 106                         /*
 107                          * If "page" is a THP tail, we must read the tail page
 108                          * flags after the head page flags. The
 109                          * split_huge_page side enforces write memory
 110                          * barriers between clearing PageTail and before the
 111                          * head page can be freed and reallocated.
 112                          */
 113                         smp_rmb();
 114                         if (likely(PageTail(page))) {
 115                                 /*
 116                                  * __split_huge_page_refcount
 117                                  * cannot race here.
 118                                  */
 119                                 VM_BUG_ON(!PageHead(page_head));
 120                                 VM_BUG_ON(page_mapcount(page) <= 0);
 121                                 atomic_dec(&page->_mapcount);
 122                                 if (put_page_testzero(page_head))
 123                                         __put_compound_page(page_head);
 124                                 return;
 125                         } else
 126                                 /*
 127                                  * __split_huge_page_refcount
 128                                  * run before us, "page" was a
 129                                  * THP tail. The split
 130                                  * page_head has been freed
 131                                  * and reallocated as slab or
 132                                  * hugetlbfs page of smaller
 133                                  * order (only possible if
 134                                  * reallocated as slab on
 135                                  * x86).
 136                                  */
 137                                 goto out_put_single;
 138                 }
 139
 140                 if (likely(page != page_head &&
 141                            get_page_unless_zero(page_head))) {
 142                         unsigned long flags;
 143
 144                         /*
 145                          * page_head wasn't a dangling pointer but it
 146                          * may not be a head page anymore by the time
 147                          * we obtain the lock. That is ok as long as it
 148                          * can't be freed from under us.
 149                          */
 150                         flags = compound_lock_irqsave(page_head);
 151                         if (unlikely(!PageTail(page))) {
 152                                 /* __split_huge_page_refcount run before us */
 153                                 compound_unlock_irqrestore(page_head, flags);
 154                                 if (put_page_testzero(page_head)) {
 155                                         /*
 156                                          * The head page may have been
 157                                          * freed and reallocated as a
 158                                          * compound page of smaller
 159                                          * order and then freed again.
 160                                          * All we know is that it
 161                                          * cannot have become: a THP
 162                                          * page, a compound page of
 163                                          * higher order, a tail page.
 164                                          * That is because we still
 165                                          * hold the refcount of the
 166                                          * split THP tail and
 167                                          * page_head was the THP head
 168                                          * before the split.
 169                                          */
 170                                         if (PageHead(page_head))
 171                                                 __put_compound_page(page_head);
 172                                         else
 173                                                 __put_single_page(page_head);
 174                                 }
 175 out_put_single:
 176                                 if (put_page_testzero(page))
 177                                         __put_single_page(page);
 178                                 return;
 179                         }
 180                         VM_BUG_ON(page_head != page->first_page);
 181                         /*
 182                          * We can release the refcount taken by
 183                          * get_page_unless_zero() now that
 184                          * __split_huge_page_refcount() is blocked on
 185                          * the compound_lock.
 186                          */
 187                         if (put_page_testzero(page_head))
 188                                 VM_BUG_ON(1);
 189                         /* __split_huge_page_refcount will wait now */
 190                         VM_BUG_ON(page_mapcount(page) <= 0);
 191                         atomic_dec(&page->_mapcount);
 192                         VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
 193                         VM_BUG_ON(atomic_read(&page->_count) != 0);
 194                         compound_unlock_irqrestore(page_head, flags);
 195
 196                         if (put_page_testzero(page_head)) {
 197                                 if (PageHead(page_head))
 198                                         __put_compound_page(page_head);
 199                                 else
 200                                         __put_single_page(page_head);
 201                         }
 202                 } else {
 203                         /* page_head is a dangling pointer */
 204                         VM_BUG_ON(PageTail(page));
 205                         goto out_put_single;
 206                 }
 207         } else if (put_page_testzero(page)) {
 208                 if (PageHead(page))
 209                         __put_compound_page(page);
 210                 else
 211                         __put_single_page(page);
 212         }
 213 }
 214
 215 void put_page(struct page *page)
 216 {
 217         if (unlikely(PageCompound(page)))
 218                 put_compound_page(page);
 219         else if (put_page_testzero(page))
 220                 __put_single_page(page);
 221 }
 222 EXPORT_SYMBOL(put_page);
 223
 224 /*
 225  * This function is exported but must not be called by anything other
 226  * than get_page(). It implements the slow path of get_page().
 227  */
 228 bool __get_page_tail(struct page *page)
 229 {
 230         /*
 231          * This takes care of get_page() if run on a tail page
 232          * returned by one of the get_user_pages/follow_page variants.
 233          * get_user_pages/follow_page itself doesn't need the compound
 234          * lock because it runs __get_page_tail_foll() under the
 235          * proper PT lock that already serializes against
 236          * split_huge_page().
 237          */
 238         unsigned long flags;
 239         bool got;
 240         struct page *page_head = compound_trans_head(page);
 241
 242         /* Ref to put_compound_page() comment. */
 243         if (PageSlab(page_head) || PageHeadHuge(page_head)) {
 244                 smp_rmb();
 245                 if (likely(PageTail(page))) {
 246                         /*
 247                          * This is a hugetlbfs page or a slab
 248                          * page. __split_huge_page_refcount
 249                          * cannot race here.
 250                          */
 251                         VM_BUG_ON(!PageHead(page_head));
 252                         __get_page_tail_foll(page, true);
 253                         return true;
 254                 } else {
 255                         /*
 256                          * __split_huge_page_refcount run
 257                          * before us, "page" was a THP
 258                          * tail. The split page_head has been
 259                          * freed and reallocated as slab or
 260                          * hugetlbfs page of smaller order
 261                          * (only possible if reallocated as
 262                          * slab on x86).
 263                          */
 264                         return false;
 265                 }
 266         }
 267
 268         got = false;
 269         if (likely(page != page_head && get_page_unless_zero(page_head))) {
 270                 /*
 271                  * page_head wasn't a dangling pointer but it
 272                  * may not be a head page anymore by the time
 273                  * we obtain the lock. That is ok as long as it
 274                  * can't be freed from under us.
 275                  */
 276                 flags = compound_lock_irqsave(page_head);
 277                 /* here __split_huge_page_refcount won't run anymore */
 278                 if (likely(PageTail(page))) {
 279                         __get_page_tail_foll(page, false);
 280                         got = true;
 281                 }
 282                 compound_unlock_irqrestore(page_head, flags);
 283                 if (unlikely(!got))
 284                         put_page(page_head);
 285         }
 286         return got;
 287 }
 288 EXPORT_SYMBOL(__get_page_tail);
 289
 290 /**
 291  * put_pages_list() - release a list of pages
 292  * @pages: list of pages threaded on page->lru
 293  *
 294  * Release a list of pages which are strung together on page.lru.  Currently
 295  * used by read_cache_pages() and related error recovery code.
 296  */
 297 void put_pages_list(struct list_head *pages)
 298 {
 299         while (!list_empty(pages)) {
 300                 struct page *victim;
 301
 302                 victim = list_entry(pages->prev, struct page, lru);
 303                 list_del(&victim->lru);
 304                 page_cache_release(victim);
 305         }
 306 }
 307 EXPORT_SYMBOL(put_pages_list);
 308
 309 /*
 310  * get_kernel_pages() - pin kernel pages in memory
 311  * @kiov:       An array of struct kvec structures
 312  * @nr_segs:    number of segments to pin
 313  * @write:      pinning for read/write, currently ignored
 314  * @pages:      array that receives pointers to the pages pinned.
 315  *              Should be at least nr_segs long.
 316  *
 317  * Returns number of pages pinned. This may be fewer than the number
 318  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 319  * were pinned, returns -errno. Each page returned must be released
 320  * with a put_page() call when it is finished with.
 321  */
 322 int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
 323                 struct page **pages)
 324 {
 325         int seg;
 326
 327         for (seg = 0; seg < nr_segs; seg++) {
 328                 if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
 329                         return seg;
 330
 331                 pages[seg] = kmap_to_page(kiov[seg].iov_base);
 332                 page_cache_get(pages[seg]);
 333         }
 334
 335         return seg;
 336 }
 337 EXPORT_SYMBOL_GPL(get_kernel_pages);
 338
 339 /*
 340  * get_kernel_page() - pin a kernel page in memory
 341  * @start:      starting kernel address
 342  * @write:      pinning for read/write, currently ignored
 343  * @pages:      array that receives pointer to the page pinned.
 344  *              Must be at least nr_segs long.
 345  *
 346  * Returns 1 if page is pinned. If the page was not pinned, returns
 347  * -errno. The page returned must be released with a put_page() call
 348  * when it is finished with.
 349  */
 350 int get_kernel_page(unsigned long start, int write, struct page **pages)
 351 {
 352         const struct kvec kiov = {
 353                 .iov_base = (void *)start,
 354                 .iov_len = PAGE_SIZE
 355         };
 356
 357         return get_kernel_pages(&kiov, 1, write, pages);
 358 }
 359 EXPORT_SYMBOL_GPL(get_kernel_page);
 360
 361 static void pagevec_lru_move_fn(struct pagevec *pvec,
 362         void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
 363         void *arg)
 364 {
 365         int i;
 366         struct zone *zone = NULL;
 367         struct lruvec *lruvec;
 368         unsigned long flags = 0;
 369
 370         for (i = 0; i < pagevec_count(pvec); i++) {
 371                 struct page *page = pvec->pages[i];
 372                 struct zone *pagezone = page_zone(page);
 373
 374                 if (pagezone != zone) {
 375                         if (zone)
 376                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 377                         zone = pagezone;
 378                         spin_lock_irqsave(&zone->lru_lock, flags);
 379                 }
 380
 381                 lruvec = mem_cgroup_page_lruvec(page, zone);
 382                 (*move_fn)(page, lruvec, arg);
 383         }
 384         if (zone)
 385                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 386         release_pages(pvec->pages, pvec->nr, pvec->cold);
 387         pagevec_reinit(pvec);
 388 }
 389
 390 static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
 391                                  void *arg)
 392 {
 393         int *pgmoved = arg;
 394
 395         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 396                 enum lru_list lru = page_lru_base_type(page);
 397                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 398                 (*pgmoved)++;
 399         }
 400 }
 401
 402 /*
 403  * pagevec_move_tail() must be called with IRQ disabled.
 404  * Otherwise this may cause nasty races.
 405  */
 406 static void pagevec_move_tail(struct pagevec *pvec)
 407 {
 408         int pgmoved = 0;
 409
 410         pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
 411         __count_vm_events(PGROTATED, pgmoved);
 412 }
 413
 414 /*
 415  * Writeback is about to end against a page which has been marked for immediate
 416  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 417  * inactive list.
 418  */
 419 void rotate_reclaimable_page(struct page *page)
 420 {
 421         if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
 422             !PageUnevictable(page) && PageLRU(page)) {
 423                 struct pagevec *pvec;
 424                 unsigned long flags;
 425
 426                 page_cache_get(page);
 427                 local_irq_save(flags);
 428                 pvec = &__get_cpu_var(lru_rotate_pvecs);
 429                 if (!pagevec_add(pvec, page))
 430                         pagevec_move_tail(pvec);
 431                 local_irq_restore(flags);
 432         }
 433 }
 434
 435 static void update_page_reclaim_stat(struct lruvec *lruvec,
 436                                      int file, int rotated)
 437 {
 438         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
 439
 440         reclaim_stat->recent_scanned[file]++;
 441         if (rotated)
 442                 reclaim_stat->recent_rotated[file]++;
 443 }
 444
 445 static void __activate_page(struct page *page, struct lruvec *lruvec,
 446                             void *arg)
 447 {
 448         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 449                 int file = page_is_file_cache(page);
 450                 int lru = page_lru_base_type(page);
 451
 452                 del_page_from_lru_list(page, lruvec, lru);
 453                 SetPageActive(page);
 454                 lru += LRU_ACTIVE;
 455                 add_page_to_lru_list(page, lruvec, lru);
 456                 trace_mm_lru_activate(page, page_to_pfn(page));
 457
 458                 __count_vm_event(PGACTIVATE);
 459                 update_page_reclaim_stat(lruvec, file, 1);
 460         }
 461 }
 462
 463 #ifdef CONFIG_SMP
 464 static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
 465
 466 static void activate_page_drain(int cpu)
 467 {
 468         struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
 469
 470         if (pagevec_count(pvec))
 471                 pagevec_lru_move_fn(pvec, __activate_page, NULL);
 472 }
 473
 474 static bool need_activate_page_drain(int cpu)
 475 {
 476         return pagevec_count(&per_cpu(activate_page_pvecs, cpu)) != 0;
 477 }
 478
 479 void activate_page(struct page *page)
 480 {
 481         if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 482                 struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
 483
 484                 page_cache_get(page);
 485                 if (!pagevec_add(pvec, page))
 486                         pagevec_lru_move_fn(pvec, __activate_page, NULL);
 487                 put_cpu_var(activate_page_pvecs);
 488         }
 489 }
 490
 491 #else
 492 static inline void activate_page_drain(int cpu)
 493 {
 494 }
 495
 496 static bool need_activate_page_drain(int cpu)
 497 {
 498         return false;
 499 }
 500
 501 void activate_page(struct page *page)
 502 {
 503         struct zone *zone = page_zone(page);
 504
 505         spin_lock_irq(&zone->lru_lock);
 506         __activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
 507         spin_unlock_irq(&zone->lru_lock);
 508 }
 509 #endif
 510
 511 static void __lru_cache_activate_page(struct page *page)
 512 {
 513         struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
 514         int i;
 515
 516         /*
 517          * Search backwards on the optimistic assumption that the page being
 518          * activated has just been added to this pagevec. Note that only
 519          * the local pagevec is examined as a !PageLRU page could be in the
 520          * process of being released, reclaimed, migrated or on a remote
 521          * pagevec that is currently being drained. Furthermore, marking
 522          * a remote pagevec's page PageActive potentially hits a race where
 523          * a page is marked PageActive just after it is added to the inactive
 524          * list causing accounting errors and BUG_ON checks to trigger.
 525          */
 526         for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
 527                 struct page *pagevec_page = pvec->pages[i];
 528
 529                 if (pagevec_page == page) {
 530                         SetPageActive(page);
 531                         break;
 532                 }
 533         }
 534
 535         put_cpu_var(lru_add_pvec);
 536 }
 537
 538 /*
 539  * Mark a page as having seen activity.
 540  *
 541  * inactive,unreferenced        ->      inactive,referenced
 542  * inactive,referenced          ->      active,unreferenced
 543  * active,unreferenced          ->      active,referenced
 544  */
 545 void mark_page_accessed(struct page *page)
 546 {
 547         if (!PageActive(page) && !PageUnevictable(page) &&
 548                         PageReferenced(page)) {
 549
 550                 /*
 551                  * If the page is on the LRU, queue it for activation via
 552                  * activate_page_pvecs. Otherwise, assume the page is on a
 553                  * pagevec, mark it active and it'll be moved to the active
 554                  * LRU on the next drain.
 555                  */
 556                 if (PageLRU(page))
 557                         activate_page(page);
 558                 else
 559                         __lru_cache_activate_page(page);
 560                 ClearPageReferenced(page);
 561         } else if (!PageReferenced(page)) {
 562                 SetPageReferenced(page);
 563         }
 564 }
 565 EXPORT_SYMBOL(mark_page_accessed);
 566
 567 /*
 568  * Queue the page for addition to the LRU via pagevec. The decision on whether
 569  * to add the page to the [in]active [file|anon] list is deferred until the
 570  * pagevec is drained. This gives a chance for the caller of __lru_cache_add()
 571  * have the page added to the active list using mark_page_accessed().
 572  */
 573 void __lru_cache_add(struct page *page)
 574 {
 575         struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
 576
 577         page_cache_get(page);
 578         if (!pagevec_space(pvec))
 579                 __pagevec_lru_add(pvec);
 580         pagevec_add(pvec, page);
 581         put_cpu_var(lru_add_pvec);
 582 }
 583 EXPORT_SYMBOL(__lru_cache_add);
 584
 585 /**
 586  * lru_cache_add - add a page to a page list
 587  * @page: the page to be added to the LRU.
 588  */
 589 void lru_cache_add(struct page *page)
 590 {
 591         VM_BUG_ON(PageActive(page) && PageUnevictable(page));
 592         VM_BUG_ON(PageLRU(page));
 593         __lru_cache_add(page);
 594 }
 595
 596 /**
 597  * add_page_to_unevictable_list - add a page to the unevictable list
 598  * @page:  the page to be added to the unevictable list
 599  *
 600  * Add page directly to its zone's unevictable list.  To avoid races with
 601  * tasks that might be making the page evictable, through eg. munlock,
 602  * munmap or exit, while it's not on the lru, we want to add the page
 603  * while it's locked or otherwise "invisible" to other tasks.  This is
 604  * difficult to do when using the pagevec cache, so bypass that.
 605  */
 606 void add_page_to_unevictable_list(struct page *page)
 607 {
 608         struct zone *zone = page_zone(page);
 609         struct lruvec *lruvec;
 610
 611         spin_lock_irq(&zone->lru_lock);
 612         lruvec = mem_cgroup_page_lruvec(page, zone);
 613         ClearPageActive(page);
 614         SetPageUnevictable(page);
 615         SetPageLRU(page);
 616         add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
 617         spin_unlock_irq(&zone->lru_lock);
 618 }
 619
 620 /*
 621  * If the page can not be invalidated, it is moved to the
 622  * inactive list to speed up its reclaim.  It is moved to the
 623  * head of the list, rather than the tail, to give the flusher
 624  * threads some time to write it out, as this is much more
 625  * effective than the single-page writeout from reclaim.
 626  *
 627  * If the page isn't page_mapped and dirty/writeback, the page
 628  * could reclaim asap using PG_reclaim.
 629  *
 630  * 1. active, mapped page -> none
 631  * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 632  * 3. inactive, mapped page -> none
 633  * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 634  * 5. inactive, clean -> inactive, tail
 635  * 6. Others -> none
 636  *
 637  * In 4, why it moves inactive's head, the VM expects the page would
 638  * be write it out by flusher threads as this is much more effective
 639  * than the single-page writeout from reclaim.
 640  */
 641 static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 642                               void *arg)
 643 {
 644         int lru, file;
 645         bool active;
 646
 647         if (!PageLRU(page))
 648                 return;
 649
 650         if (PageUnevictable(page))
 651                 return;
 652
 653         /* Some processes are using the page */
 654         if (page_mapped(page))
 655                 return;
 656
 657         active = PageActive(page);
 658         file = page_is_file_cache(page);
 659         lru = page_lru_base_type(page);
 660
 661         del_page_from_lru_list(page, lruvec, lru + active);
 662         ClearPageActive(page);
 663         ClearPageReferenced(page);
 664         add_page_to_lru_list(page, lruvec, lru);
 665
 666         if (PageWriteback(page) || PageDirty(page)) {
 667                 /*
 668                  * PG_reclaim could be raced with end_page_writeback
 669                  * It can make readahead confusing.  But race window
 670                  * is _really_ small and  it's non-critical problem.
 671                  */
 672                 SetPageReclaim(page);
 673         } else {
 674                 /*
 675                  * The page's writeback ends up during pagevec
 676                  * We moves tha page into tail of inactive.
 677                  */
 678                 list_move_tail(&page->lru, &lruvec->lists[lru]);
 679                 __count_vm_event(PGROTATED);
 680         }
 681
 682         if (active)
 683                 __count_vm_event(PGDEACTIVATE);
 684         update_page_reclaim_stat(lruvec, file, 0);
 685 }
 686
 687 /*
 688  * Drain pages out of the cpu's pagevecs.
 689  * Either "cpu" is the current CPU, and preemption has already been
 690  * disabled; or "cpu" is being hot-unplugged, and is already dead.
 691  */
 692 void lru_add_drain_cpu(int cpu)
 693 {
 694         struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
 695
 696         if (pagevec_count(pvec))
 697                 __pagevec_lru_add(pvec);
 698
 699         pvec = &per_cpu(lru_rotate_pvecs, cpu);
 700         if (pagevec_count(pvec)) {
 701                 unsigned long flags;
 702
 703                 /* No harm done if a racing interrupt already did this */
 704                 local_irq_save(flags);
 705                 pagevec_move_tail(pvec);
 706                 local_irq_restore(flags);
 707         }
 708
 709         pvec = &per_cpu(lru_deactivate_pvecs, cpu);
 710         if (pagevec_count(pvec))
 711                 pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 712
 713         activate_page_drain(cpu);
 714 }
 715
 716 /**
 717  * deactivate_page - forcefully deactivate a page
 718  * @page: page to deactivate
 719  *
 720  * This function hints the VM that @page is a good reclaim candidate,
 721  * for example if its invalidation fails due to the page being dirty
 722  * or under writeback.
 723  */
 724 void deactivate_page(struct page *page)
 725 {
 726         /*
 727          * In a workload with many unevictable page such as mprotect, unevictable
 728          * page deactivation for accelerating reclaim is pointless.
 729          */
 730         if (PageUnevictable(page))
 731                 return;
 732
 733         if (likely(get_page_unless_zero(page))) {
 734                 struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
 735
 736                 if (!pagevec_add(pvec, page))
 737                         pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 738                 put_cpu_var(lru_deactivate_pvecs);
 739         }
 740 }
 741
 742 void lru_add_drain(void)
 743 {
 744         lru_add_drain_cpu(get_cpu());
 745         put_cpu();
 746 }
 747
 748 static void lru_add_drain_per_cpu(struct work_struct *dummy)
 749 {
 750         lru_add_drain();
 751 }
 752
 753 static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 754
 755 void lru_add_drain_all(void)
 756 {
 757         static DEFINE_MUTEX(lock);
 758         static struct cpumask has_work;
 759         int cpu;
 760
 761         mutex_lock(&lock);
 762         get_online_cpus();
 763         cpumask_clear(&has_work);
 764
 765         for_each_online_cpu(cpu) {
 766                 struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 767
 768                 if (pagevec_count(&per_cpu(lru_add_pvec, cpu)) ||
 769                     pagevec_count(&per_cpu(lru_rotate_pvecs, cpu)) ||
 770                     pagevec_count(&per_cpu(lru_deactivate_pvecs, cpu)) ||
 771                     need_activate_page_drain(cpu)) {
 772                         INIT_WORK(work, lru_add_drain_per_cpu);
 773                         schedule_work_on(cpu, work);
 774                         cpumask_set_cpu(cpu, &has_work);
 775                 }
 776         }
 777
 778         for_each_cpu(cpu, &has_work)
 779                 flush_work(&per_cpu(lru_add_drain_work, cpu));
 780
 781         put_online_cpus();
 782         mutex_unlock(&lock);
 783 }
 784
 785 /*
 786  * Batched page_cache_release().  Decrement the reference count on all the
 787  * passed pages.  If it fell to zero then remove the page from the LRU and
 788  * free it.
 789  *
 790  * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 791  * for the remainder of the operation.
 792  *
 793  * The locking in this function is against shrink_inactive_list(): we recheck
 794  * the page count inside the lock to see whether shrink_inactive_list()
 795  * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
 796  * will free it.
 797  */
 798 void release_pages(struct page **pages, int nr, int cold)
 799 {
 800         int i;
 801         LIST_HEAD(pages_to_free);
 802         struct zone *zone = NULL;
 803         struct lruvec *lruvec;
 804         unsigned long uninitialized_var(flags);
 805
 806         for (i = 0; i < nr; i++) {
 807                 struct page *page = pages[i];
 808
 809                 if (unlikely(PageCompound(page))) {
 810                         if (zone) {
 811                                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 812                                 zone = NULL;
 813                         }
 814                         put_compound_page(page);
 815                         continue;
 816                 }
 817
 818                 if (!put_page_testzero(page))
 819                         continue;
 820
 821                 if (PageLRU(page)) {
 822                         struct zone *pagezone = page_zone(page);
 823
 824                         if (pagezone != zone) {
 825                                 if (zone)
 826                                         spin_unlock_irqrestore(&zone->lru_lock,
 827                                                                         flags);
 828                                 zone = pagezone;
 829                                 spin_lock_irqsave(&zone->lru_lock, flags);
 830                         }
 831
 832                         lruvec = mem_cgroup_page_lruvec(page, zone);
 833                         VM_BUG_ON(!PageLRU(page));
 834                         __ClearPageLRU(page);
 835                         del_page_from_lru_list(page, lruvec, page_off_lru(page));
 836                 }
 837
 838                 /* Clear Active bit in case of parallel mark_page_accessed */
 839                 ClearPageActive(page);
 840
 841                 list_add(&page->lru, &pages_to_free);
 842         }
 843         if (zone)
 844                 spin_unlock_irqrestore(&zone->lru_lock, flags);
 845
 846         free_hot_cold_page_list(&pages_to_free, cold);
 847 }
 848 EXPORT_SYMBOL(release_pages);
 849
 850 /*
 851  * The pages which we're about to release may be in the deferred lru-addition
 852  * queues.  That would prevent them from really being freed right now.  That's
 853  * OK from a correctness point of view but is inefficient - those pages may be
 854  * cache-warm and we want to give them back to the page allocator ASAP.
 855  *
 856  * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 857  * and __pagevec_lru_add_active() call release_pages() directly to avoid
 858  * mutual recursion.
 859  */
 860 void __pagevec_release(struct pagevec *pvec)
 861 {
 862         lru_add_drain();
 863         release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 864         pagevec_reinit(pvec);
 865 }
 866 EXPORT_SYMBOL(__pagevec_release);
 867
 868 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 869 /* used by __split_huge_page_refcount() */
 870 void lru_add_page_tail(struct page *page, struct page *page_tail,
 871                        struct lruvec *lruvec, struct list_head *list)
 872 {
 873         const int file = 0;
 874
 875         VM_BUG_ON(!PageHead(page));
 876         VM_BUG_ON(PageCompound(page_tail));
 877         VM_BUG_ON(PageLRU(page_tail));
 878         VM_BUG_ON(NR_CPUS != 1 &&
 879                   !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
 880
 881         if (!list)
 882                 SetPageLRU(page_tail);
 883
 884         if (likely(PageLRU(page)))
 885                 list_add_tail(&page_tail->lru, &page->lru);
 886         else if (list) {
 887                 /* page reclaim is reclaiming a huge page */
 888                 get_page(page_tail);
 889                 list_add_tail(&page_tail->lru, list);
 890         } else {
 891                 struct list_head *list_head;
 892                 /*
 893                  * Head page has not yet been counted, as an hpage,
 894                  * so we must account for each subpage individually.
 895                  *
 896                  * Use the standard add function to put page_tail on the list,
 897                  * but then correct its position so they all end up in order.
 898                  */
 899                 add_page_to_lru_list(page_tail, lruvec, page_lru(page_tail));
 900                 list_head = page_tail->lru.prev;
 901                 list_move_tail(&page_tail->lru, list_head);
 902         }
 903
 904         if (!PageUnevictable(page))
 905                 update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
 906 }
 907 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 908
 909 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
 910                                  void *arg)
 911 {
 912         int file = page_is_file_cache(page);
 913         int active = PageActive(page);
 914         enum lru_list lru = page_lru(page);
 915
 916         VM_BUG_ON(PageLRU(page));
 917
 918         SetPageLRU(page);
 919         add_page_to_lru_list(page, lruvec, lru);
 920         update_page_reclaim_stat(lruvec, file, active);
 921         trace_mm_lru_insertion(page, page_to_pfn(page), lru, trace_pagemap_flags(page));
 922 }
 923
 924 /*
 925  * Add the passed pages to the LRU, then drop the caller's refcount
 926  * on them.  Reinitialises the caller's pagevec.
 927  */
 928 void __pagevec_lru_add(struct pagevec *pvec)
 929 {
 930         pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
 931 }
 932 EXPORT_SYMBOL(__pagevec_lru_add);
 933
 934 /**
 935  * pagevec_lookup - gang pagecache lookup
 936  * @pvec:       Where the resulting pages are placed
 937  * @mapping:    The address_space to search
 938  * @start:      The starting page index
 939  * @nr_pages:   The maximum number of pages
 940  *
 941  * pagevec_lookup() will search for and return a group of up to @nr_pages pages
 942  * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 943  * reference against the pages in @pvec.
 944  *
 945  * The search returns a group of mapping-contiguous pages with ascending
 946  * indexes.  There may be holes in the indices due to not-present pages.
 947  *
 948  * pagevec_lookup() returns the number of pages which were found.
 949  */
 950 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 951                 pgoff_t start, unsigned nr_pages)
 952 {
 953         pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 954         return pagevec_count(pvec);
 955 }
 956 EXPORT_SYMBOL(pagevec_lookup);
 957
 958 unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
 959                 pgoff_t *index, int tag, unsigned nr_pages)
 960 {
 961         pvec->nr = find_get_pages_tag(mapping, index, tag,
 962                                         nr_pages, pvec->pages);
 963         return pagevec_count(pvec);
 964 }
 965 EXPORT_SYMBOL(pagevec_lookup_tag);
 966
 967 /*
 968  * Perform any setup for the swap system
 969  */
 970 void __init swap_setup(void)
 971 {
 972         unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
 973 #ifdef CONFIG_SWAP
 974         int i;
 975
 976         if (bdi_init(swapper_spaces[0].backing_dev_info))
 977                 panic("Failed to init swap bdi");
 978         for (i = 0; i < MAX_SWAPFILES; i++) {
 979                 spin_lock_init(&swapper_spaces[i].tree_lock);
 980                 INIT_LIST_HEAD(&swapper_spaces[i].i_mmap_nonlinear);
 981         }
 982 #endif
 983
 984         /* Use a smaller cluster for small-memory machines */
 985         if (megs < 16)
 986                 page_cluster = 2;
 987         else
 988                 page_cluster = 3;
 989         /*
 990          * Right now other parts of the system means that we
 991          * _really_ don't want to cluster much more
 992          */
 993 }