vmscan.c revision 6c0b13519d1c755d874e82c8fb8a6dcef0ee402c
1/* 2 * linux/mm/vmscan.c 3 * 4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 5 * 6 * Swap reorganised 29.12.95, Stephen Tweedie. 7 * kswapd added: 7.1.96 sct 8 * Removed kswapd_ctl limits, and swap out as many pages as needed 9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel. 10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). 11 * Multiqueue VM started 5.8.00, Rik van Riel. 12 */ 13 14#include <linux/mm.h> 15#include <linux/module.h> 16#include <linux/slab.h> 17#include <linux/kernel_stat.h> 18#include <linux/swap.h> 19#include <linux/pagemap.h> 20#include <linux/init.h> 21#include <linux/highmem.h> 22#include <linux/vmstat.h> 23#include <linux/file.h> 24#include <linux/writeback.h> 25#include <linux/blkdev.h> 26#include <linux/buffer_head.h> /* for try_to_release_page(), 27 buffer_heads_over_limit */ 28#include <linux/mm_inline.h> 29#include <linux/pagevec.h> 30#include <linux/backing-dev.h> 31#include <linux/rmap.h> 32#include <linux/topology.h> 33#include <linux/cpu.h> 34#include <linux/cpuset.h> 35#include <linux/notifier.h> 36#include <linux/rwsem.h> 37#include <linux/delay.h> 38#include <linux/kthread.h> 39#include <linux/freezer.h> 40#include <linux/memcontrol.h> 41#include <linux/delayacct.h> 42#include <linux/sysctl.h> 43 44#include <asm/tlbflush.h> 45#include <asm/div64.h> 46 47#include <linux/swapops.h> 48 49#include "internal.h" 50 51struct scan_control { 52 /* Incremented by the number of inactive pages that were scanned */ 53 unsigned long nr_scanned; 54 55 /* Number of pages freed so far during a call to shrink_zones() */ 56 unsigned long nr_reclaimed; 57 58 /* This context's GFP mask */ 59 gfp_t gfp_mask; 60 61 int may_writepage; 62 63 /* Can mapped pages be reclaimed? */ 64 int may_unmap; 65 66 /* Can pages be swapped as part of reclaim? */ 67 int may_swap; 68 69 /* This context's SWAP_CLUSTER_MAX. If freeing memory for 70 * suspend, we effectively ignore SWAP_CLUSTER_MAX. 71 * In this context, it doesn't matter that we scan the 72 * whole list at once. */ 73 int swap_cluster_max; 74 75 int swappiness; 76 77 int all_unreclaimable; 78 79 int order; 80 81 /* Which cgroup do we reclaim from */ 82 struct mem_cgroup *mem_cgroup; 83 84 /* 85 * Nodemask of nodes allowed by the caller. If NULL, all nodes 86 * are scanned. 87 */ 88 nodemask_t *nodemask; 89 90 /* Pluggable isolate pages callback */ 91 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst, 92 unsigned long *scanned, int order, int mode, 93 struct zone *z, struct mem_cgroup *mem_cont, 94 int active, int file); 95}; 96 97#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) 98 99#ifdef ARCH_HAS_PREFETCH 100#define prefetch_prev_lru_page(_page, _base, _field) \ 101 do { \ 102 if ((_page)->lru.prev != _base) { \ 103 struct page *prev; \ 104 \ 105 prev = lru_to_page(&(_page->lru)); \ 106 prefetch(&prev->_field); \ 107 } \ 108 } while (0) 109#else 110#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) 111#endif 112 113#ifdef ARCH_HAS_PREFETCHW 114#define prefetchw_prev_lru_page(_page, _base, _field) \ 115 do { \ 116 if ((_page)->lru.prev != _base) { \ 117 struct page *prev; \ 118 \ 119 prev = lru_to_page(&(_page->lru)); \ 120 prefetchw(&prev->_field); \ 121 } \ 122 } while (0) 123#else 124#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) 125#endif 126 127/* 128 * From 0 .. 100. Higher means more swappy. 129 */ 130int vm_swappiness = 60; 131long vm_total_pages; /* The total number of pages which the VM controls */ 132 133static LIST_HEAD(shrinker_list); 134static DECLARE_RWSEM(shrinker_rwsem); 135 136#ifdef CONFIG_CGROUP_MEM_RES_CTLR 137#define scanning_global_lru(sc) (!(sc)->mem_cgroup) 138#else 139#define scanning_global_lru(sc) (1) 140#endif 141 142static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone, 143 struct scan_control *sc) 144{ 145 if (!scanning_global_lru(sc)) 146 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone); 147 148 return &zone->reclaim_stat; 149} 150 151static unsigned long zone_nr_pages(struct zone *zone, struct scan_control *sc, 152 enum lru_list lru) 153{ 154 if (!scanning_global_lru(sc)) 155 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru); 156 157 return zone_page_state(zone, NR_LRU_BASE + lru); 158} 159 160 161/* 162 * Add a shrinker callback to be called from the vm 163 */ 164void register_shrinker(struct shrinker *shrinker) 165{ 166 shrinker->nr = 0; 167 down_write(&shrinker_rwsem); 168 list_add_tail(&shrinker->list, &shrinker_list); 169 up_write(&shrinker_rwsem); 170} 171EXPORT_SYMBOL(register_shrinker); 172 173/* 174 * Remove one 175 */ 176void unregister_shrinker(struct shrinker *shrinker) 177{ 178 down_write(&shrinker_rwsem); 179 list_del(&shrinker->list); 180 up_write(&shrinker_rwsem); 181} 182EXPORT_SYMBOL(unregister_shrinker); 183 184#define SHRINK_BATCH 128 185/* 186 * Call the shrink functions to age shrinkable caches 187 * 188 * Here we assume it costs one seek to replace a lru page and that it also 189 * takes a seek to recreate a cache object. With this in mind we age equal 190 * percentages of the lru and ageable caches. This should balance the seeks 191 * generated by these structures. 192 * 193 * If the vm encountered mapped pages on the LRU it increase the pressure on 194 * slab to avoid swapping. 195 * 196 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. 197 * 198 * `lru_pages' represents the number of on-LRU pages in all the zones which 199 * are eligible for the caller's allocation attempt. It is used for balancing 200 * slab reclaim versus page reclaim. 201 * 202 * Returns the number of slab objects which we shrunk. 203 */ 204unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask, 205 unsigned long lru_pages) 206{ 207 struct shrinker *shrinker; 208 unsigned long ret = 0; 209 210 if (scanned == 0) 211 scanned = SWAP_CLUSTER_MAX; 212 213 if (!down_read_trylock(&shrinker_rwsem)) 214 return 1; /* Assume we'll be able to shrink next time */ 215 216 list_for_each_entry(shrinker, &shrinker_list, list) { 217 unsigned long long delta; 218 unsigned long total_scan; 219 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask); 220 221 delta = (4 * scanned) / shrinker->seeks; 222 delta *= max_pass; 223 do_div(delta, lru_pages + 1); 224 shrinker->nr += delta; 225 if (shrinker->nr < 0) { 226 printk(KERN_ERR "shrink_slab: %pF negative objects to " 227 "delete nr=%ld\n", 228 shrinker->shrink, shrinker->nr); 229 shrinker->nr = max_pass; 230 } 231 232 /* 233 * Avoid risking looping forever due to too large nr value: 234 * never try to free more than twice the estimate number of 235 * freeable entries. 236 */ 237 if (shrinker->nr > max_pass * 2) 238 shrinker->nr = max_pass * 2; 239 240 total_scan = shrinker->nr; 241 shrinker->nr = 0; 242 243 while (total_scan >= SHRINK_BATCH) { 244 long this_scan = SHRINK_BATCH; 245 int shrink_ret; 246 int nr_before; 247 248 nr_before = (*shrinker->shrink)(0, gfp_mask); 249 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask); 250 if (shrink_ret == -1) 251 break; 252 if (shrink_ret < nr_before) 253 ret += nr_before - shrink_ret; 254 count_vm_events(SLABS_SCANNED, this_scan); 255 total_scan -= this_scan; 256 257 cond_resched(); 258 } 259 260 shrinker->nr += total_scan; 261 } 262 up_read(&shrinker_rwsem); 263 return ret; 264} 265 266/* Called without lock on whether page is mapped, so answer is unstable */ 267static inline int page_mapping_inuse(struct page *page) 268{ 269 struct address_space *mapping; 270 271 /* Page is in somebody's page tables. */ 272 if (page_mapped(page)) 273 return 1; 274 275 /* Be more reluctant to reclaim swapcache than pagecache */ 276 if (PageSwapCache(page)) 277 return 1; 278 279 mapping = page_mapping(page); 280 if (!mapping) 281 return 0; 282 283 /* File is mmap'd by somebody? */ 284 return mapping_mapped(mapping); 285} 286 287static inline int is_page_cache_freeable(struct page *page) 288{ 289 return page_count(page) - !!page_has_private(page) == 2; 290} 291 292static int may_write_to_queue(struct backing_dev_info *bdi) 293{ 294 if (current->flags & PF_SWAPWRITE) 295 return 1; 296 if (!bdi_write_congested(bdi)) 297 return 1; 298 if (bdi == current->backing_dev_info) 299 return 1; 300 return 0; 301} 302 303/* 304 * We detected a synchronous write error writing a page out. Probably 305 * -ENOSPC. We need to propagate that into the address_space for a subsequent 306 * fsync(), msync() or close(). 307 * 308 * The tricky part is that after writepage we cannot touch the mapping: nothing 309 * prevents it from being freed up. But we have a ref on the page and once 310 * that page is locked, the mapping is pinned. 311 * 312 * We're allowed to run sleeping lock_page() here because we know the caller has 313 * __GFP_FS. 314 */ 315static void handle_write_error(struct address_space *mapping, 316 struct page *page, int error) 317{ 318 lock_page(page); 319 if (page_mapping(page) == mapping) 320 mapping_set_error(mapping, error); 321 unlock_page(page); 322} 323 324/* Request for sync pageout. */ 325enum pageout_io { 326 PAGEOUT_IO_ASYNC, 327 PAGEOUT_IO_SYNC, 328}; 329 330/* possible outcome of pageout() */ 331typedef enum { 332 /* failed to write page out, page is locked */ 333 PAGE_KEEP, 334 /* move page to the active list, page is locked */ 335 PAGE_ACTIVATE, 336 /* page has been sent to the disk successfully, page is unlocked */ 337 PAGE_SUCCESS, 338 /* page is clean and locked */ 339 PAGE_CLEAN, 340} pageout_t; 341 342/* 343 * pageout is called by shrink_page_list() for each dirty page. 344 * Calls ->writepage(). 345 */ 346static pageout_t pageout(struct page *page, struct address_space *mapping, 347 enum pageout_io sync_writeback) 348{ 349 /* 350 * If the page is dirty, only perform writeback if that write 351 * will be non-blocking. To prevent this allocation from being 352 * stalled by pagecache activity. But note that there may be 353 * stalls if we need to run get_block(). We could test 354 * PagePrivate for that. 355 * 356 * If this process is currently in generic_file_write() against 357 * this page's queue, we can perform writeback even if that 358 * will block. 359 * 360 * If the page is swapcache, write it back even if that would 361 * block, for some throttling. This happens by accident, because 362 * swap_backing_dev_info is bust: it doesn't reflect the 363 * congestion state of the swapdevs. Easy to fix, if needed. 364 * See swapfile.c:page_queue_congested(). 365 */ 366 if (!is_page_cache_freeable(page)) 367 return PAGE_KEEP; 368 if (!mapping) { 369 /* 370 * Some data journaling orphaned pages can have 371 * page->mapping == NULL while being dirty with clean buffers. 372 */ 373 if (page_has_private(page)) { 374 if (try_to_free_buffers(page)) { 375 ClearPageDirty(page); 376 printk("%s: orphaned page\n", __func__); 377 return PAGE_CLEAN; 378 } 379 } 380 return PAGE_KEEP; 381 } 382 if (mapping->a_ops->writepage == NULL) 383 return PAGE_ACTIVATE; 384 if (!may_write_to_queue(mapping->backing_dev_info)) 385 return PAGE_KEEP; 386 387 if (clear_page_dirty_for_io(page)) { 388 int res; 389 struct writeback_control wbc = { 390 .sync_mode = WB_SYNC_NONE, 391 .nr_to_write = SWAP_CLUSTER_MAX, 392 .range_start = 0, 393 .range_end = LLONG_MAX, 394 .nonblocking = 1, 395 .for_reclaim = 1, 396 }; 397 398 SetPageReclaim(page); 399 res = mapping->a_ops->writepage(page, &wbc); 400 if (res < 0) 401 handle_write_error(mapping, page, res); 402 if (res == AOP_WRITEPAGE_ACTIVATE) { 403 ClearPageReclaim(page); 404 return PAGE_ACTIVATE; 405 } 406 407 /* 408 * Wait on writeback if requested to. This happens when 409 * direct reclaiming a large contiguous area and the 410 * first attempt to free a range of pages fails. 411 */ 412 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC) 413 wait_on_page_writeback(page); 414 415 if (!PageWriteback(page)) { 416 /* synchronous write or broken a_ops? */ 417 ClearPageReclaim(page); 418 } 419 inc_zone_page_state(page, NR_VMSCAN_WRITE); 420 return PAGE_SUCCESS; 421 } 422 423 return PAGE_CLEAN; 424} 425 426/* 427 * Same as remove_mapping, but if the page is removed from the mapping, it 428 * gets returned with a refcount of 0. 429 */ 430static int __remove_mapping(struct address_space *mapping, struct page *page) 431{ 432 BUG_ON(!PageLocked(page)); 433 BUG_ON(mapping != page_mapping(page)); 434 435 spin_lock_irq(&mapping->tree_lock); 436 /* 437 * The non racy check for a busy page. 438 * 439 * Must be careful with the order of the tests. When someone has 440 * a ref to the page, it may be possible that they dirty it then 441 * drop the reference. So if PageDirty is tested before page_count 442 * here, then the following race may occur: 443 * 444 * get_user_pages(&page); 445 * [user mapping goes away] 446 * write_to(page); 447 * !PageDirty(page) [good] 448 * SetPageDirty(page); 449 * put_page(page); 450 * !page_count(page) [good, discard it] 451 * 452 * [oops, our write_to data is lost] 453 * 454 * Reversing the order of the tests ensures such a situation cannot 455 * escape unnoticed. The smp_rmb is needed to ensure the page->flags 456 * load is not satisfied before that of page->_count. 457 * 458 * Note that if SetPageDirty is always performed via set_page_dirty, 459 * and thus under tree_lock, then this ordering is not required. 460 */ 461 if (!page_freeze_refs(page, 2)) 462 goto cannot_free; 463 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ 464 if (unlikely(PageDirty(page))) { 465 page_unfreeze_refs(page, 2); 466 goto cannot_free; 467 } 468 469 if (PageSwapCache(page)) { 470 swp_entry_t swap = { .val = page_private(page) }; 471 __delete_from_swap_cache(page); 472 spin_unlock_irq(&mapping->tree_lock); 473 swapcache_free(swap, page); 474 } else { 475 __remove_from_page_cache(page); 476 spin_unlock_irq(&mapping->tree_lock); 477 mem_cgroup_uncharge_cache_page(page); 478 } 479 480 return 1; 481 482cannot_free: 483 spin_unlock_irq(&mapping->tree_lock); 484 return 0; 485} 486 487/* 488 * Attempt to detach a locked page from its ->mapping. If it is dirty or if 489 * someone else has a ref on the page, abort and return 0. If it was 490 * successfully detached, return 1. Assumes the caller has a single ref on 491 * this page. 492 */ 493int remove_mapping(struct address_space *mapping, struct page *page) 494{ 495 if (__remove_mapping(mapping, page)) { 496 /* 497 * Unfreezing the refcount with 1 rather than 2 effectively 498 * drops the pagecache ref for us without requiring another 499 * atomic operation. 500 */ 501 page_unfreeze_refs(page, 1); 502 return 1; 503 } 504 return 0; 505} 506 507/** 508 * putback_lru_page - put previously isolated page onto appropriate LRU list 509 * @page: page to be put back to appropriate lru list 510 * 511 * Add previously isolated @page to appropriate LRU list. 512 * Page may still be unevictable for other reasons. 513 * 514 * lru_lock must not be held, interrupts must be enabled. 515 */ 516void putback_lru_page(struct page *page) 517{ 518 int lru; 519 int active = !!TestClearPageActive(page); 520 int was_unevictable = PageUnevictable(page); 521 522 VM_BUG_ON(PageLRU(page)); 523 524redo: 525 ClearPageUnevictable(page); 526 527 if (page_evictable(page, NULL)) { 528 /* 529 * For evictable pages, we can use the cache. 530 * In event of a race, worst case is we end up with an 531 * unevictable page on [in]active list. 532 * We know how to handle that. 533 */ 534 lru = active + page_lru_base_type(page); 535 lru_cache_add_lru(page, lru); 536 } else { 537 /* 538 * Put unevictable pages directly on zone's unevictable 539 * list. 540 */ 541 lru = LRU_UNEVICTABLE; 542 add_page_to_unevictable_list(page); 543 } 544 545 /* 546 * page's status can change while we move it among lru. If an evictable 547 * page is on unevictable list, it never be freed. To avoid that, 548 * check after we added it to the list, again. 549 */ 550 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) { 551 if (!isolate_lru_page(page)) { 552 put_page(page); 553 goto redo; 554 } 555 /* This means someone else dropped this page from LRU 556 * So, it will be freed or putback to LRU again. There is 557 * nothing to do here. 558 */ 559 } 560 561 if (was_unevictable && lru != LRU_UNEVICTABLE) 562 count_vm_event(UNEVICTABLE_PGRESCUED); 563 else if (!was_unevictable && lru == LRU_UNEVICTABLE) 564 count_vm_event(UNEVICTABLE_PGCULLED); 565 566 put_page(page); /* drop ref from isolate */ 567} 568 569/* 570 * shrink_page_list() returns the number of reclaimed pages 571 */ 572static unsigned long shrink_page_list(struct list_head *page_list, 573 struct scan_control *sc, 574 enum pageout_io sync_writeback) 575{ 576 LIST_HEAD(ret_pages); 577 struct pagevec freed_pvec; 578 int pgactivate = 0; 579 unsigned long nr_reclaimed = 0; 580 unsigned long vm_flags; 581 582 cond_resched(); 583 584 pagevec_init(&freed_pvec, 1); 585 while (!list_empty(page_list)) { 586 struct address_space *mapping; 587 struct page *page; 588 int may_enter_fs; 589 int referenced; 590 591 cond_resched(); 592 593 page = lru_to_page(page_list); 594 list_del(&page->lru); 595 596 if (!trylock_page(page)) 597 goto keep; 598 599 VM_BUG_ON(PageActive(page)); 600 601 sc->nr_scanned++; 602 603 if (unlikely(!page_evictable(page, NULL))) 604 goto cull_mlocked; 605 606 if (!sc->may_unmap && page_mapped(page)) 607 goto keep_locked; 608 609 /* Double the slab pressure for mapped and swapcache pages */ 610 if (page_mapped(page) || PageSwapCache(page)) 611 sc->nr_scanned++; 612 613 may_enter_fs = (sc->gfp_mask & __GFP_FS) || 614 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); 615 616 if (PageWriteback(page)) { 617 /* 618 * Synchronous reclaim is performed in two passes, 619 * first an asynchronous pass over the list to 620 * start parallel writeback, and a second synchronous 621 * pass to wait for the IO to complete. Wait here 622 * for any page for which writeback has already 623 * started. 624 */ 625 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs) 626 wait_on_page_writeback(page); 627 else 628 goto keep_locked; 629 } 630 631 referenced = page_referenced(page, 1, 632 sc->mem_cgroup, &vm_flags); 633 /* 634 * In active use or really unfreeable? Activate it. 635 * If page which have PG_mlocked lost isoltation race, 636 * try_to_unmap moves it to unevictable list 637 */ 638 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && 639 referenced && page_mapping_inuse(page) 640 && !(vm_flags & VM_LOCKED)) 641 goto activate_locked; 642 643 /* 644 * Anonymous process memory has backing store? 645 * Try to allocate it some swap space here. 646 */ 647 if (PageAnon(page) && !PageSwapCache(page)) { 648 if (!(sc->gfp_mask & __GFP_IO)) 649 goto keep_locked; 650 if (!add_to_swap(page)) 651 goto activate_locked; 652 may_enter_fs = 1; 653 } 654 655 mapping = page_mapping(page); 656 657 /* 658 * The page is mapped into the page tables of one or more 659 * processes. Try to unmap it here. 660 */ 661 if (page_mapped(page) && mapping) { 662 switch (try_to_unmap(page, 0)) { 663 case SWAP_FAIL: 664 goto activate_locked; 665 case SWAP_AGAIN: 666 goto keep_locked; 667 case SWAP_MLOCK: 668 goto cull_mlocked; 669 case SWAP_SUCCESS: 670 ; /* try to free the page below */ 671 } 672 } 673 674 if (PageDirty(page)) { 675 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced) 676 goto keep_locked; 677 if (!may_enter_fs) 678 goto keep_locked; 679 if (!sc->may_writepage) 680 goto keep_locked; 681 682 /* Page is dirty, try to write it out here */ 683 switch (pageout(page, mapping, sync_writeback)) { 684 case PAGE_KEEP: 685 goto keep_locked; 686 case PAGE_ACTIVATE: 687 goto activate_locked; 688 case PAGE_SUCCESS: 689 if (PageWriteback(page) || PageDirty(page)) 690 goto keep; 691 /* 692 * A synchronous write - probably a ramdisk. Go 693 * ahead and try to reclaim the page. 694 */ 695 if (!trylock_page(page)) 696 goto keep; 697 if (PageDirty(page) || PageWriteback(page)) 698 goto keep_locked; 699 mapping = page_mapping(page); 700 case PAGE_CLEAN: 701 ; /* try to free the page below */ 702 } 703 } 704 705 /* 706 * If the page has buffers, try to free the buffer mappings 707 * associated with this page. If we succeed we try to free 708 * the page as well. 709 * 710 * We do this even if the page is PageDirty(). 711 * try_to_release_page() does not perform I/O, but it is 712 * possible for a page to have PageDirty set, but it is actually 713 * clean (all its buffers are clean). This happens if the 714 * buffers were written out directly, with submit_bh(). ext3 715 * will do this, as well as the blockdev mapping. 716 * try_to_release_page() will discover that cleanness and will 717 * drop the buffers and mark the page clean - it can be freed. 718 * 719 * Rarely, pages can have buffers and no ->mapping. These are 720 * the pages which were not successfully invalidated in 721 * truncate_complete_page(). We try to drop those buffers here 722 * and if that worked, and the page is no longer mapped into 723 * process address space (page_count == 1) it can be freed. 724 * Otherwise, leave the page on the LRU so it is swappable. 725 */ 726 if (page_has_private(page)) { 727 if (!try_to_release_page(page, sc->gfp_mask)) 728 goto activate_locked; 729 if (!mapping && page_count(page) == 1) { 730 unlock_page(page); 731 if (put_page_testzero(page)) 732 goto free_it; 733 else { 734 /* 735 * rare race with speculative reference. 736 * the speculative reference will free 737 * this page shortly, so we may 738 * increment nr_reclaimed here (and 739 * leave it off the LRU). 740 */ 741 nr_reclaimed++; 742 continue; 743 } 744 } 745 } 746 747 if (!mapping || !__remove_mapping(mapping, page)) 748 goto keep_locked; 749 750 /* 751 * At this point, we have no other references and there is 752 * no way to pick any more up (removed from LRU, removed 753 * from pagecache). Can use non-atomic bitops now (and 754 * we obviously don't have to worry about waking up a process 755 * waiting on the page lock, because there are no references. 756 */ 757 __clear_page_locked(page); 758free_it: 759 nr_reclaimed++; 760 if (!pagevec_add(&freed_pvec, page)) { 761 __pagevec_free(&freed_pvec); 762 pagevec_reinit(&freed_pvec); 763 } 764 continue; 765 766cull_mlocked: 767 if (PageSwapCache(page)) 768 try_to_free_swap(page); 769 unlock_page(page); 770 putback_lru_page(page); 771 continue; 772 773activate_locked: 774 /* Not a candidate for swapping, so reclaim swap space. */ 775 if (PageSwapCache(page) && vm_swap_full()) 776 try_to_free_swap(page); 777 VM_BUG_ON(PageActive(page)); 778 SetPageActive(page); 779 pgactivate++; 780keep_locked: 781 unlock_page(page); 782keep: 783 list_add(&page->lru, &ret_pages); 784 VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); 785 } 786 list_splice(&ret_pages, page_list); 787 if (pagevec_count(&freed_pvec)) 788 __pagevec_free(&freed_pvec); 789 count_vm_events(PGACTIVATE, pgactivate); 790 return nr_reclaimed; 791} 792 793/* LRU Isolation modes. */ 794#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */ 795#define ISOLATE_ACTIVE 1 /* Isolate active pages. */ 796#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */ 797 798/* 799 * Attempt to remove the specified page from its LRU. Only take this page 800 * if it is of the appropriate PageActive status. Pages which are being 801 * freed elsewhere are also ignored. 802 * 803 * page: page to consider 804 * mode: one of the LRU isolation modes defined above 805 * 806 * returns 0 on success, -ve errno on failure. 807 */ 808int __isolate_lru_page(struct page *page, int mode, int file) 809{ 810 int ret = -EINVAL; 811 812 /* Only take pages on the LRU. */ 813 if (!PageLRU(page)) 814 return ret; 815 816 /* 817 * When checking the active state, we need to be sure we are 818 * dealing with comparible boolean values. Take the logical not 819 * of each. 820 */ 821 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode)) 822 return ret; 823 824 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file) 825 return ret; 826 827 /* 828 * When this function is being called for lumpy reclaim, we 829 * initially look into all LRU pages, active, inactive and 830 * unevictable; only give shrink_page_list evictable pages. 831 */ 832 if (PageUnevictable(page)) 833 return ret; 834 835 ret = -EBUSY; 836 837 if (likely(get_page_unless_zero(page))) { 838 /* 839 * Be careful not to clear PageLRU until after we're 840 * sure the page is not being freed elsewhere -- the 841 * page release code relies on it. 842 */ 843 ClearPageLRU(page); 844 ret = 0; 845 } 846 847 return ret; 848} 849 850/* 851 * zone->lru_lock is heavily contended. Some of the functions that 852 * shrink the lists perform better by taking out a batch of pages 853 * and working on them outside the LRU lock. 854 * 855 * For pagecache intensive workloads, this function is the hottest 856 * spot in the kernel (apart from copy_*_user functions). 857 * 858 * Appropriate locks must be held before calling this function. 859 * 860 * @nr_to_scan: The number of pages to look through on the list. 861 * @src: The LRU list to pull pages off. 862 * @dst: The temp list to put pages on to. 863 * @scanned: The number of pages that were scanned. 864 * @order: The caller's attempted allocation order 865 * @mode: One of the LRU isolation modes 866 * @file: True [1] if isolating file [!anon] pages 867 * 868 * returns how many pages were moved onto *@dst. 869 */ 870static unsigned long isolate_lru_pages(unsigned long nr_to_scan, 871 struct list_head *src, struct list_head *dst, 872 unsigned long *scanned, int order, int mode, int file) 873{ 874 unsigned long nr_taken = 0; 875 unsigned long scan; 876 877 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { 878 struct page *page; 879 unsigned long pfn; 880 unsigned long end_pfn; 881 unsigned long page_pfn; 882 int zone_id; 883 884 page = lru_to_page(src); 885 prefetchw_prev_lru_page(page, src, flags); 886 887 VM_BUG_ON(!PageLRU(page)); 888 889 switch (__isolate_lru_page(page, mode, file)) { 890 case 0: 891 list_move(&page->lru, dst); 892 mem_cgroup_del_lru(page); 893 nr_taken++; 894 break; 895 896 case -EBUSY: 897 /* else it is being freed elsewhere */ 898 list_move(&page->lru, src); 899 mem_cgroup_rotate_lru_list(page, page_lru(page)); 900 continue; 901 902 default: 903 BUG(); 904 } 905 906 if (!order) 907 continue; 908 909 /* 910 * Attempt to take all pages in the order aligned region 911 * surrounding the tag page. Only take those pages of 912 * the same active state as that tag page. We may safely 913 * round the target page pfn down to the requested order 914 * as the mem_map is guarenteed valid out to MAX_ORDER, 915 * where that page is in a different zone we will detect 916 * it from its zone id and abort this block scan. 917 */ 918 zone_id = page_zone_id(page); 919 page_pfn = page_to_pfn(page); 920 pfn = page_pfn & ~((1 << order) - 1); 921 end_pfn = pfn + (1 << order); 922 for (; pfn < end_pfn; pfn++) { 923 struct page *cursor_page; 924 925 /* The target page is in the block, ignore it. */ 926 if (unlikely(pfn == page_pfn)) 927 continue; 928 929 /* Avoid holes within the zone. */ 930 if (unlikely(!pfn_valid_within(pfn))) 931 break; 932 933 cursor_page = pfn_to_page(pfn); 934 935 /* Check that we have not crossed a zone boundary. */ 936 if (unlikely(page_zone_id(cursor_page) != zone_id)) 937 continue; 938 939 /* 940 * If we don't have enough swap space, reclaiming of 941 * anon page which don't already have a swap slot is 942 * pointless. 943 */ 944 if (nr_swap_pages <= 0 && PageAnon(cursor_page) && 945 !PageSwapCache(cursor_page)) 946 continue; 947 948 if (__isolate_lru_page(cursor_page, mode, file) == 0) { 949 list_move(&cursor_page->lru, dst); 950 mem_cgroup_del_lru(cursor_page); 951 nr_taken++; 952 scan++; 953 } 954 } 955 } 956 957 *scanned = scan; 958 return nr_taken; 959} 960 961static unsigned long isolate_pages_global(unsigned long nr, 962 struct list_head *dst, 963 unsigned long *scanned, int order, 964 int mode, struct zone *z, 965 struct mem_cgroup *mem_cont, 966 int active, int file) 967{ 968 int lru = LRU_BASE; 969 if (active) 970 lru += LRU_ACTIVE; 971 if (file) 972 lru += LRU_FILE; 973 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order, 974 mode, file); 975} 976 977/* 978 * clear_active_flags() is a helper for shrink_active_list(), clearing 979 * any active bits from the pages in the list. 980 */ 981static unsigned long clear_active_flags(struct list_head *page_list, 982 unsigned int *count) 983{ 984 int nr_active = 0; 985 int lru; 986 struct page *page; 987 988 list_for_each_entry(page, page_list, lru) { 989 lru = page_lru_base_type(page); 990 if (PageActive(page)) { 991 lru += LRU_ACTIVE; 992 ClearPageActive(page); 993 nr_active++; 994 } 995 count[lru]++; 996 } 997 998 return nr_active; 999} 1000 1001/** 1002 * isolate_lru_page - tries to isolate a page from its LRU list 1003 * @page: page to isolate from its LRU list 1004 * 1005 * Isolates a @page from an LRU list, clears PageLRU and adjusts the 1006 * vmstat statistic corresponding to whatever LRU list the page was on. 1007 * 1008 * Returns 0 if the page was removed from an LRU list. 1009 * Returns -EBUSY if the page was not on an LRU list. 1010 * 1011 * The returned page will have PageLRU() cleared. If it was found on 1012 * the active list, it will have PageActive set. If it was found on 1013 * the unevictable list, it will have the PageUnevictable bit set. That flag 1014 * may need to be cleared by the caller before letting the page go. 1015 * 1016 * The vmstat statistic corresponding to the list on which the page was 1017 * found will be decremented. 1018 * 1019 * Restrictions: 1020 * (1) Must be called with an elevated refcount on the page. This is a 1021 * fundamentnal difference from isolate_lru_pages (which is called 1022 * without a stable reference). 1023 * (2) the lru_lock must not be held. 1024 * (3) interrupts must be enabled. 1025 */ 1026int isolate_lru_page(struct page *page) 1027{ 1028 int ret = -EBUSY; 1029 1030 if (PageLRU(page)) { 1031 struct zone *zone = page_zone(page); 1032 1033 spin_lock_irq(&zone->lru_lock); 1034 if (PageLRU(page) && get_page_unless_zero(page)) { 1035 int lru = page_lru(page); 1036 ret = 0; 1037 ClearPageLRU(page); 1038 1039 del_page_from_lru_list(zone, page, lru); 1040 } 1041 spin_unlock_irq(&zone->lru_lock); 1042 } 1043 return ret; 1044} 1045 1046/* 1047 * Are there way too many processes in the direct reclaim path already? 1048 */ 1049static int too_many_isolated(struct zone *zone, int file, 1050 struct scan_control *sc) 1051{ 1052 unsigned long inactive, isolated; 1053 1054 if (current_is_kswapd()) 1055 return 0; 1056 1057 if (!scanning_global_lru(sc)) 1058 return 0; 1059 1060 if (file) { 1061 inactive = zone_page_state(zone, NR_INACTIVE_FILE); 1062 isolated = zone_page_state(zone, NR_ISOLATED_FILE); 1063 } else { 1064 inactive = zone_page_state(zone, NR_INACTIVE_ANON); 1065 isolated = zone_page_state(zone, NR_ISOLATED_ANON); 1066 } 1067 1068 return isolated > inactive; 1069} 1070 1071/* 1072 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number 1073 * of reclaimed pages 1074 */ 1075static unsigned long shrink_inactive_list(unsigned long max_scan, 1076 struct zone *zone, struct scan_control *sc, 1077 int priority, int file) 1078{ 1079 LIST_HEAD(page_list); 1080 struct pagevec pvec; 1081 unsigned long nr_scanned = 0; 1082 unsigned long nr_reclaimed = 0; 1083 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); 1084 int lumpy_reclaim = 0; 1085 1086 while (unlikely(too_many_isolated(zone, file, sc))) { 1087 congestion_wait(WRITE, HZ/10); 1088 1089 /* We are about to die and free our memory. Return now. */ 1090 if (fatal_signal_pending(current)) 1091 return SWAP_CLUSTER_MAX; 1092 } 1093 1094 /* 1095 * If we need a large contiguous chunk of memory, or have 1096 * trouble getting a small set of contiguous pages, we 1097 * will reclaim both active and inactive pages. 1098 * 1099 * We use the same threshold as pageout congestion_wait below. 1100 */ 1101 if (sc->order > PAGE_ALLOC_COSTLY_ORDER) 1102 lumpy_reclaim = 1; 1103 else if (sc->order && priority < DEF_PRIORITY - 2) 1104 lumpy_reclaim = 1; 1105 1106 pagevec_init(&pvec, 1); 1107 1108 lru_add_drain(); 1109 spin_lock_irq(&zone->lru_lock); 1110 do { 1111 struct page *page; 1112 unsigned long nr_taken; 1113 unsigned long nr_scan; 1114 unsigned long nr_freed; 1115 unsigned long nr_active; 1116 unsigned int count[NR_LRU_LISTS] = { 0, }; 1117 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE; 1118 unsigned long nr_anon; 1119 unsigned long nr_file; 1120 1121 nr_taken = sc->isolate_pages(sc->swap_cluster_max, 1122 &page_list, &nr_scan, sc->order, mode, 1123 zone, sc->mem_cgroup, 0, file); 1124 1125 if (scanning_global_lru(sc)) { 1126 zone->pages_scanned += nr_scan; 1127 if (current_is_kswapd()) 1128 __count_zone_vm_events(PGSCAN_KSWAPD, zone, 1129 nr_scan); 1130 else 1131 __count_zone_vm_events(PGSCAN_DIRECT, zone, 1132 nr_scan); 1133 } 1134 1135 if (nr_taken == 0) 1136 goto done; 1137 1138 nr_active = clear_active_flags(&page_list, count); 1139 __count_vm_events(PGDEACTIVATE, nr_active); 1140 1141 __mod_zone_page_state(zone, NR_ACTIVE_FILE, 1142 -count[LRU_ACTIVE_FILE]); 1143 __mod_zone_page_state(zone, NR_INACTIVE_FILE, 1144 -count[LRU_INACTIVE_FILE]); 1145 __mod_zone_page_state(zone, NR_ACTIVE_ANON, 1146 -count[LRU_ACTIVE_ANON]); 1147 __mod_zone_page_state(zone, NR_INACTIVE_ANON, 1148 -count[LRU_INACTIVE_ANON]); 1149 1150 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON]; 1151 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE]; 1152 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon); 1153 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file); 1154 1155 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON]; 1156 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON]; 1157 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE]; 1158 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE]; 1159 1160 spin_unlock_irq(&zone->lru_lock); 1161 1162 nr_scanned += nr_scan; 1163 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC); 1164 1165 /* 1166 * If we are direct reclaiming for contiguous pages and we do 1167 * not reclaim everything in the list, try again and wait 1168 * for IO to complete. This will stall high-order allocations 1169 * but that should be acceptable to the caller 1170 */ 1171 if (nr_freed < nr_taken && !current_is_kswapd() && 1172 lumpy_reclaim) { 1173 congestion_wait(BLK_RW_ASYNC, HZ/10); 1174 1175 /* 1176 * The attempt at page out may have made some 1177 * of the pages active, mark them inactive again. 1178 */ 1179 nr_active = clear_active_flags(&page_list, count); 1180 count_vm_events(PGDEACTIVATE, nr_active); 1181 1182 nr_freed += shrink_page_list(&page_list, sc, 1183 PAGEOUT_IO_SYNC); 1184 } 1185 1186 nr_reclaimed += nr_freed; 1187 1188 local_irq_disable(); 1189 if (current_is_kswapd()) 1190 __count_vm_events(KSWAPD_STEAL, nr_freed); 1191 __count_zone_vm_events(PGSTEAL, zone, nr_freed); 1192 1193 spin_lock(&zone->lru_lock); 1194 /* 1195 * Put back any unfreeable pages. 1196 */ 1197 while (!list_empty(&page_list)) { 1198 int lru; 1199 page = lru_to_page(&page_list); 1200 VM_BUG_ON(PageLRU(page)); 1201 list_del(&page->lru); 1202 if (unlikely(!page_evictable(page, NULL))) { 1203 spin_unlock_irq(&zone->lru_lock); 1204 putback_lru_page(page); 1205 spin_lock_irq(&zone->lru_lock); 1206 continue; 1207 } 1208 SetPageLRU(page); 1209 lru = page_lru(page); 1210 add_page_to_lru_list(zone, page, lru); 1211 if (is_active_lru(lru)) { 1212 int file = is_file_lru(lru); 1213 reclaim_stat->recent_rotated[file]++; 1214 } 1215 if (!pagevec_add(&pvec, page)) { 1216 spin_unlock_irq(&zone->lru_lock); 1217 __pagevec_release(&pvec); 1218 spin_lock_irq(&zone->lru_lock); 1219 } 1220 } 1221 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon); 1222 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file); 1223 1224 } while (nr_scanned < max_scan); 1225 1226done: 1227 spin_unlock_irq(&zone->lru_lock); 1228 pagevec_release(&pvec); 1229 return nr_reclaimed; 1230} 1231 1232/* 1233 * We are about to scan this zone at a certain priority level. If that priority 1234 * level is smaller (ie: more urgent) than the previous priority, then note 1235 * that priority level within the zone. This is done so that when the next 1236 * process comes in to scan this zone, it will immediately start out at this 1237 * priority level rather than having to build up its own scanning priority. 1238 * Here, this priority affects only the reclaim-mapped threshold. 1239 */ 1240static inline void note_zone_scanning_priority(struct zone *zone, int priority) 1241{ 1242 if (priority < zone->prev_priority) 1243 zone->prev_priority = priority; 1244} 1245 1246/* 1247 * This moves pages from the active list to the inactive list. 1248 * 1249 * We move them the other way if the page is referenced by one or more 1250 * processes, from rmap. 1251 * 1252 * If the pages are mostly unmapped, the processing is fast and it is 1253 * appropriate to hold zone->lru_lock across the whole operation. But if 1254 * the pages are mapped, the processing is slow (page_referenced()) so we 1255 * should drop zone->lru_lock around each page. It's impossible to balance 1256 * this, so instead we remove the pages from the LRU while processing them. 1257 * It is safe to rely on PG_active against the non-LRU pages in here because 1258 * nobody will play with that bit on a non-LRU page. 1259 * 1260 * The downside is that we have to touch page->_count against each page. 1261 * But we had to alter page->flags anyway. 1262 */ 1263 1264static void move_active_pages_to_lru(struct zone *zone, 1265 struct list_head *list, 1266 enum lru_list lru) 1267{ 1268 unsigned long pgmoved = 0; 1269 struct pagevec pvec; 1270 struct page *page; 1271 1272 pagevec_init(&pvec, 1); 1273 1274 while (!list_empty(list)) { 1275 page = lru_to_page(list); 1276 1277 VM_BUG_ON(PageLRU(page)); 1278 SetPageLRU(page); 1279 1280 list_move(&page->lru, &zone->lru[lru].list); 1281 mem_cgroup_add_lru_list(page, lru); 1282 pgmoved++; 1283 1284 if (!pagevec_add(&pvec, page) || list_empty(list)) { 1285 spin_unlock_irq(&zone->lru_lock); 1286 if (buffer_heads_over_limit) 1287 pagevec_strip(&pvec); 1288 __pagevec_release(&pvec); 1289 spin_lock_irq(&zone->lru_lock); 1290 } 1291 } 1292 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); 1293 if (!is_active_lru(lru)) 1294 __count_vm_events(PGDEACTIVATE, pgmoved); 1295} 1296 1297static void shrink_active_list(unsigned long nr_pages, struct zone *zone, 1298 struct scan_control *sc, int priority, int file) 1299{ 1300 unsigned long nr_taken; 1301 unsigned long pgscanned; 1302 unsigned long vm_flags; 1303 LIST_HEAD(l_hold); /* The pages which were snipped off */ 1304 LIST_HEAD(l_active); 1305 LIST_HEAD(l_inactive); 1306 struct page *page; 1307 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); 1308 unsigned long nr_rotated = 0; 1309 1310 lru_add_drain(); 1311 spin_lock_irq(&zone->lru_lock); 1312 nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order, 1313 ISOLATE_ACTIVE, zone, 1314 sc->mem_cgroup, 1, file); 1315 /* 1316 * zone->pages_scanned is used for detect zone's oom 1317 * mem_cgroup remembers nr_scan by itself. 1318 */ 1319 if (scanning_global_lru(sc)) { 1320 zone->pages_scanned += pgscanned; 1321 } 1322 reclaim_stat->recent_scanned[file] += nr_taken; 1323 1324 __count_zone_vm_events(PGREFILL, zone, pgscanned); 1325 if (file) 1326 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken); 1327 else 1328 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken); 1329 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); 1330 spin_unlock_irq(&zone->lru_lock); 1331 1332 while (!list_empty(&l_hold)) { 1333 cond_resched(); 1334 page = lru_to_page(&l_hold); 1335 list_del(&page->lru); 1336 1337 if (unlikely(!page_evictable(page, NULL))) { 1338 putback_lru_page(page); 1339 continue; 1340 } 1341 1342 /* page_referenced clears PageReferenced */ 1343 if (page_mapping_inuse(page) && 1344 page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) { 1345 nr_rotated++; 1346 /* 1347 * Identify referenced, file-backed active pages and 1348 * give them one more trip around the active list. So 1349 * that executable code get better chances to stay in 1350 * memory under moderate memory pressure. Anon pages 1351 * are not likely to be evicted by use-once streaming 1352 * IO, plus JVM can create lots of anon VM_EXEC pages, 1353 * so we ignore them here. 1354 */ 1355 if ((vm_flags & VM_EXEC) && !PageAnon(page)) { 1356 list_add(&page->lru, &l_active); 1357 continue; 1358 } 1359 } 1360 1361 ClearPageActive(page); /* we are de-activating */ 1362 list_add(&page->lru, &l_inactive); 1363 } 1364 1365 /* 1366 * Move pages back to the lru list. 1367 */ 1368 spin_lock_irq(&zone->lru_lock); 1369 /* 1370 * Count referenced pages from currently used mappings as rotated, 1371 * even though only some of them are actually re-activated. This 1372 * helps balance scan pressure between file and anonymous pages in 1373 * get_scan_ratio. 1374 */ 1375 reclaim_stat->recent_rotated[file] += nr_rotated; 1376 1377 move_active_pages_to_lru(zone, &l_active, 1378 LRU_ACTIVE + file * LRU_FILE); 1379 move_active_pages_to_lru(zone, &l_inactive, 1380 LRU_BASE + file * LRU_FILE); 1381 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); 1382 spin_unlock_irq(&zone->lru_lock); 1383} 1384 1385static int inactive_anon_is_low_global(struct zone *zone) 1386{ 1387 unsigned long active, inactive; 1388 1389 active = zone_page_state(zone, NR_ACTIVE_ANON); 1390 inactive = zone_page_state(zone, NR_INACTIVE_ANON); 1391 1392 if (inactive * zone->inactive_ratio < active) 1393 return 1; 1394 1395 return 0; 1396} 1397 1398/** 1399 * inactive_anon_is_low - check if anonymous pages need to be deactivated 1400 * @zone: zone to check 1401 * @sc: scan control of this context 1402 * 1403 * Returns true if the zone does not have enough inactive anon pages, 1404 * meaning some active anon pages need to be deactivated. 1405 */ 1406static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc) 1407{ 1408 int low; 1409 1410 if (scanning_global_lru(sc)) 1411 low = inactive_anon_is_low_global(zone); 1412 else 1413 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup); 1414 return low; 1415} 1416 1417static int inactive_file_is_low_global(struct zone *zone) 1418{ 1419 unsigned long active, inactive; 1420 1421 active = zone_page_state(zone, NR_ACTIVE_FILE); 1422 inactive = zone_page_state(zone, NR_INACTIVE_FILE); 1423 1424 return (active > inactive); 1425} 1426 1427/** 1428 * inactive_file_is_low - check if file pages need to be deactivated 1429 * @zone: zone to check 1430 * @sc: scan control of this context 1431 * 1432 * When the system is doing streaming IO, memory pressure here 1433 * ensures that active file pages get deactivated, until more 1434 * than half of the file pages are on the inactive list. 1435 * 1436 * Once we get to that situation, protect the system's working 1437 * set from being evicted by disabling active file page aging. 1438 * 1439 * This uses a different ratio than the anonymous pages, because 1440 * the page cache uses a use-once replacement algorithm. 1441 */ 1442static int inactive_file_is_low(struct zone *zone, struct scan_control *sc) 1443{ 1444 int low; 1445 1446 if (scanning_global_lru(sc)) 1447 low = inactive_file_is_low_global(zone); 1448 else 1449 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup); 1450 return low; 1451} 1452 1453static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, 1454 struct zone *zone, struct scan_control *sc, int priority) 1455{ 1456 int file = is_file_lru(lru); 1457 1458 if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) { 1459 shrink_active_list(nr_to_scan, zone, sc, priority, file); 1460 return 0; 1461 } 1462 1463 if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) { 1464 shrink_active_list(nr_to_scan, zone, sc, priority, file); 1465 return 0; 1466 } 1467 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file); 1468} 1469 1470/* 1471 * Determine how aggressively the anon and file LRU lists should be 1472 * scanned. The relative value of each set of LRU lists is determined 1473 * by looking at the fraction of the pages scanned we did rotate back 1474 * onto the active list instead of evict. 1475 * 1476 * percent[0] specifies how much pressure to put on ram/swap backed 1477 * memory, while percent[1] determines pressure on the file LRUs. 1478 */ 1479static void get_scan_ratio(struct zone *zone, struct scan_control *sc, 1480 unsigned long *percent) 1481{ 1482 unsigned long anon, file, free; 1483 unsigned long anon_prio, file_prio; 1484 unsigned long ap, fp; 1485 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); 1486 1487 anon = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) + 1488 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON); 1489 file = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) + 1490 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE); 1491 1492 if (scanning_global_lru(sc)) { 1493 free = zone_page_state(zone, NR_FREE_PAGES); 1494 /* If we have very few page cache pages, 1495 force-scan anon pages. */ 1496 if (unlikely(file + free <= high_wmark_pages(zone))) { 1497 percent[0] = 100; 1498 percent[1] = 0; 1499 return; 1500 } 1501 } 1502 1503 /* 1504 * OK, so we have swap space and a fair amount of page cache 1505 * pages. We use the recently rotated / recently scanned 1506 * ratios to determine how valuable each cache is. 1507 * 1508 * Because workloads change over time (and to avoid overflow) 1509 * we keep these statistics as a floating average, which ends 1510 * up weighing recent references more than old ones. 1511 * 1512 * anon in [0], file in [1] 1513 */ 1514 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { 1515 spin_lock_irq(&zone->lru_lock); 1516 reclaim_stat->recent_scanned[0] /= 2; 1517 reclaim_stat->recent_rotated[0] /= 2; 1518 spin_unlock_irq(&zone->lru_lock); 1519 } 1520 1521 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { 1522 spin_lock_irq(&zone->lru_lock); 1523 reclaim_stat->recent_scanned[1] /= 2; 1524 reclaim_stat->recent_rotated[1] /= 2; 1525 spin_unlock_irq(&zone->lru_lock); 1526 } 1527 1528 /* 1529 * With swappiness at 100, anonymous and file have the same priority. 1530 * This scanning priority is essentially the inverse of IO cost. 1531 */ 1532 anon_prio = sc->swappiness; 1533 file_prio = 200 - sc->swappiness; 1534 1535 /* 1536 * The amount of pressure on anon vs file pages is inversely 1537 * proportional to the fraction of recently scanned pages on 1538 * each list that were recently referenced and in active use. 1539 */ 1540 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1); 1541 ap /= reclaim_stat->recent_rotated[0] + 1; 1542 1543 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1); 1544 fp /= reclaim_stat->recent_rotated[1] + 1; 1545 1546 /* Normalize to percentages */ 1547 percent[0] = 100 * ap / (ap + fp + 1); 1548 percent[1] = 100 - percent[0]; 1549} 1550 1551/* 1552 * Smallish @nr_to_scan's are deposited in @nr_saved_scan, 1553 * until we collected @swap_cluster_max pages to scan. 1554 */ 1555static unsigned long nr_scan_try_batch(unsigned long nr_to_scan, 1556 unsigned long *nr_saved_scan, 1557 unsigned long swap_cluster_max) 1558{ 1559 unsigned long nr; 1560 1561 *nr_saved_scan += nr_to_scan; 1562 nr = *nr_saved_scan; 1563 1564 if (nr >= swap_cluster_max) 1565 *nr_saved_scan = 0; 1566 else 1567 nr = 0; 1568 1569 return nr; 1570} 1571 1572/* 1573 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. 1574 */ 1575static void shrink_zone(int priority, struct zone *zone, 1576 struct scan_control *sc) 1577{ 1578 unsigned long nr[NR_LRU_LISTS]; 1579 unsigned long nr_to_scan; 1580 unsigned long percent[2]; /* anon @ 0; file @ 1 */ 1581 enum lru_list l; 1582 unsigned long nr_reclaimed = sc->nr_reclaimed; 1583 unsigned long swap_cluster_max = sc->swap_cluster_max; 1584 int noswap = 0; 1585 1586 /* If we have no swap space, do not bother scanning anon pages. */ 1587 if (!sc->may_swap || (nr_swap_pages <= 0)) { 1588 noswap = 1; 1589 percent[0] = 0; 1590 percent[1] = 100; 1591 } else 1592 get_scan_ratio(zone, sc, percent); 1593 1594 for_each_evictable_lru(l) { 1595 int file = is_file_lru(l); 1596 unsigned long scan; 1597 1598 scan = zone_nr_pages(zone, sc, l); 1599 if (priority || noswap) { 1600 scan >>= priority; 1601 scan = (scan * percent[file]) / 100; 1602 } 1603 if (scanning_global_lru(sc)) 1604 nr[l] = nr_scan_try_batch(scan, 1605 &zone->lru[l].nr_saved_scan, 1606 swap_cluster_max); 1607 else 1608 nr[l] = scan; 1609 } 1610 1611 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || 1612 nr[LRU_INACTIVE_FILE]) { 1613 for_each_evictable_lru(l) { 1614 if (nr[l]) { 1615 nr_to_scan = min(nr[l], swap_cluster_max); 1616 nr[l] -= nr_to_scan; 1617 1618 nr_reclaimed += shrink_list(l, nr_to_scan, 1619 zone, sc, priority); 1620 } 1621 } 1622 /* 1623 * On large memory systems, scan >> priority can become 1624 * really large. This is fine for the starting priority; 1625 * we want to put equal scanning pressure on each zone. 1626 * However, if the VM has a harder time of freeing pages, 1627 * with multiple processes reclaiming pages, the total 1628 * freeing target can get unreasonably large. 1629 */ 1630 if (nr_reclaimed > swap_cluster_max && 1631 priority < DEF_PRIORITY && !current_is_kswapd()) 1632 break; 1633 } 1634 1635 sc->nr_reclaimed = nr_reclaimed; 1636 1637 /* 1638 * Even if we did not try to evict anon pages at all, we want to 1639 * rebalance the anon lru active/inactive ratio. 1640 */ 1641 if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0) 1642 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0); 1643 1644 throttle_vm_writeout(sc->gfp_mask); 1645} 1646 1647/* 1648 * This is the direct reclaim path, for page-allocating processes. We only 1649 * try to reclaim pages from zones which will satisfy the caller's allocation 1650 * request. 1651 * 1652 * We reclaim from a zone even if that zone is over high_wmark_pages(zone). 1653 * Because: 1654 * a) The caller may be trying to free *extra* pages to satisfy a higher-order 1655 * allocation or 1656 * b) The target zone may be at high_wmark_pages(zone) but the lower zones 1657 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' 1658 * zone defense algorithm. 1659 * 1660 * If a zone is deemed to be full of pinned pages then just give it a light 1661 * scan then give up on it. 1662 */ 1663static void shrink_zones(int priority, struct zonelist *zonelist, 1664 struct scan_control *sc) 1665{ 1666 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask); 1667 struct zoneref *z; 1668 struct zone *zone; 1669 1670 sc->all_unreclaimable = 1; 1671 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx, 1672 sc->nodemask) { 1673 if (!populated_zone(zone)) 1674 continue; 1675 /* 1676 * Take care memory controller reclaiming has small influence 1677 * to global LRU. 1678 */ 1679 if (scanning_global_lru(sc)) { 1680 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 1681 continue; 1682 note_zone_scanning_priority(zone, priority); 1683 1684 if (zone_is_all_unreclaimable(zone) && 1685 priority != DEF_PRIORITY) 1686 continue; /* Let kswapd poll it */ 1687 sc->all_unreclaimable = 0; 1688 } else { 1689 /* 1690 * Ignore cpuset limitation here. We just want to reduce 1691 * # of used pages by us regardless of memory shortage. 1692 */ 1693 sc->all_unreclaimable = 0; 1694 mem_cgroup_note_reclaim_priority(sc->mem_cgroup, 1695 priority); 1696 } 1697 1698 shrink_zone(priority, zone, sc); 1699 } 1700} 1701 1702/* 1703 * This is the main entry point to direct page reclaim. 1704 * 1705 * If a full scan of the inactive list fails to free enough memory then we 1706 * are "out of memory" and something needs to be killed. 1707 * 1708 * If the caller is !__GFP_FS then the probability of a failure is reasonably 1709 * high - the zone may be full of dirty or under-writeback pages, which this 1710 * caller can't do much about. We kick pdflush and take explicit naps in the 1711 * hope that some of these pages can be written. But if the allocating task 1712 * holds filesystem locks which prevent writeout this might not work, and the 1713 * allocation attempt will fail. 1714 * 1715 * returns: 0, if no pages reclaimed 1716 * else, the number of pages reclaimed 1717 */ 1718static unsigned long do_try_to_free_pages(struct zonelist *zonelist, 1719 struct scan_control *sc) 1720{ 1721 int priority; 1722 unsigned long ret = 0; 1723 unsigned long total_scanned = 0; 1724 struct reclaim_state *reclaim_state = current->reclaim_state; 1725 unsigned long lru_pages = 0; 1726 struct zoneref *z; 1727 struct zone *zone; 1728 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask); 1729 1730 delayacct_freepages_start(); 1731 1732 if (scanning_global_lru(sc)) 1733 count_vm_event(ALLOCSTALL); 1734 /* 1735 * mem_cgroup will not do shrink_slab. 1736 */ 1737 if (scanning_global_lru(sc)) { 1738 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1739 1740 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 1741 continue; 1742 1743 lru_pages += zone_reclaimable_pages(zone); 1744 } 1745 } 1746 1747 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 1748 sc->nr_scanned = 0; 1749 if (!priority) 1750 disable_swap_token(); 1751 shrink_zones(priority, zonelist, sc); 1752 /* 1753 * Don't shrink slabs when reclaiming memory from 1754 * over limit cgroups 1755 */ 1756 if (scanning_global_lru(sc)) { 1757 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages); 1758 if (reclaim_state) { 1759 sc->nr_reclaimed += reclaim_state->reclaimed_slab; 1760 reclaim_state->reclaimed_slab = 0; 1761 } 1762 } 1763 total_scanned += sc->nr_scanned; 1764 if (sc->nr_reclaimed >= sc->swap_cluster_max) { 1765 ret = sc->nr_reclaimed; 1766 goto out; 1767 } 1768 1769 /* 1770 * Try to write back as many pages as we just scanned. This 1771 * tends to cause slow streaming writers to write data to the 1772 * disk smoothly, at the dirtying rate, which is nice. But 1773 * that's undesirable in laptop mode, where we *want* lumpy 1774 * writeout. So in laptop mode, write out the whole world. 1775 */ 1776 if (total_scanned > sc->swap_cluster_max + 1777 sc->swap_cluster_max / 2) { 1778 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned); 1779 sc->may_writepage = 1; 1780 } 1781 1782 /* Take a nap, wait for some writeback to complete */ 1783 if (sc->nr_scanned && priority < DEF_PRIORITY - 2) 1784 congestion_wait(BLK_RW_ASYNC, HZ/10); 1785 } 1786 /* top priority shrink_zones still had more to do? don't OOM, then */ 1787 if (!sc->all_unreclaimable && scanning_global_lru(sc)) 1788 ret = sc->nr_reclaimed; 1789out: 1790 /* 1791 * Now that we've scanned all the zones at this priority level, note 1792 * that level within the zone so that the next thread which performs 1793 * scanning of this zone will immediately start out at this priority 1794 * level. This affects only the decision whether or not to bring 1795 * mapped pages onto the inactive list. 1796 */ 1797 if (priority < 0) 1798 priority = 0; 1799 1800 if (scanning_global_lru(sc)) { 1801 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1802 1803 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 1804 continue; 1805 1806 zone->prev_priority = priority; 1807 } 1808 } else 1809 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority); 1810 1811 delayacct_freepages_end(); 1812 1813 return ret; 1814} 1815 1816unsigned long try_to_free_pages(struct zonelist *zonelist, int order, 1817 gfp_t gfp_mask, nodemask_t *nodemask) 1818{ 1819 struct scan_control sc = { 1820 .gfp_mask = gfp_mask, 1821 .may_writepage = !laptop_mode, 1822 .swap_cluster_max = SWAP_CLUSTER_MAX, 1823 .may_unmap = 1, 1824 .may_swap = 1, 1825 .swappiness = vm_swappiness, 1826 .order = order, 1827 .mem_cgroup = NULL, 1828 .isolate_pages = isolate_pages_global, 1829 .nodemask = nodemask, 1830 }; 1831 1832 return do_try_to_free_pages(zonelist, &sc); 1833} 1834 1835#ifdef CONFIG_CGROUP_MEM_RES_CTLR 1836 1837unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont, 1838 gfp_t gfp_mask, 1839 bool noswap, 1840 unsigned int swappiness) 1841{ 1842 struct scan_control sc = { 1843 .may_writepage = !laptop_mode, 1844 .may_unmap = 1, 1845 .may_swap = !noswap, 1846 .swap_cluster_max = SWAP_CLUSTER_MAX, 1847 .swappiness = swappiness, 1848 .order = 0, 1849 .mem_cgroup = mem_cont, 1850 .isolate_pages = mem_cgroup_isolate_pages, 1851 .nodemask = NULL, /* we don't care the placement */ 1852 }; 1853 struct zonelist *zonelist; 1854 1855 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | 1856 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); 1857 zonelist = NODE_DATA(numa_node_id())->node_zonelists; 1858 return do_try_to_free_pages(zonelist, &sc); 1859} 1860#endif 1861 1862/* 1863 * For kswapd, balance_pgdat() will work across all this node's zones until 1864 * they are all at high_wmark_pages(zone). 1865 * 1866 * Returns the number of pages which were actually freed. 1867 * 1868 * There is special handling here for zones which are full of pinned pages. 1869 * This can happen if the pages are all mlocked, or if they are all used by 1870 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. 1871 * What we do is to detect the case where all pages in the zone have been 1872 * scanned twice and there has been zero successful reclaim. Mark the zone as 1873 * dead and from now on, only perform a short scan. Basically we're polling 1874 * the zone for when the problem goes away. 1875 * 1876 * kswapd scans the zones in the highmem->normal->dma direction. It skips 1877 * zones which have free_pages > high_wmark_pages(zone), but once a zone is 1878 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the 1879 * lower zones regardless of the number of free pages in the lower zones. This 1880 * interoperates with the page allocator fallback scheme to ensure that aging 1881 * of pages is balanced across the zones. 1882 */ 1883static unsigned long balance_pgdat(pg_data_t *pgdat, int order) 1884{ 1885 int all_zones_ok; 1886 int priority; 1887 int i; 1888 unsigned long total_scanned; 1889 struct reclaim_state *reclaim_state = current->reclaim_state; 1890 struct scan_control sc = { 1891 .gfp_mask = GFP_KERNEL, 1892 .may_unmap = 1, 1893 .may_swap = 1, 1894 .swap_cluster_max = SWAP_CLUSTER_MAX, 1895 .swappiness = vm_swappiness, 1896 .order = order, 1897 .mem_cgroup = NULL, 1898 .isolate_pages = isolate_pages_global, 1899 }; 1900 /* 1901 * temp_priority is used to remember the scanning priority at which 1902 * this zone was successfully refilled to 1903 * free_pages == high_wmark_pages(zone). 1904 */ 1905 int temp_priority[MAX_NR_ZONES]; 1906 1907loop_again: 1908 total_scanned = 0; 1909 sc.nr_reclaimed = 0; 1910 sc.may_writepage = !laptop_mode; 1911 count_vm_event(PAGEOUTRUN); 1912 1913 for (i = 0; i < pgdat->nr_zones; i++) 1914 temp_priority[i] = DEF_PRIORITY; 1915 1916 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 1917 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ 1918 unsigned long lru_pages = 0; 1919 1920 /* The swap token gets in the way of swapout... */ 1921 if (!priority) 1922 disable_swap_token(); 1923 1924 all_zones_ok = 1; 1925 1926 /* 1927 * Scan in the highmem->dma direction for the highest 1928 * zone which needs scanning 1929 */ 1930 for (i = pgdat->nr_zones - 1; i >= 0; i--) { 1931 struct zone *zone = pgdat->node_zones + i; 1932 1933 if (!populated_zone(zone)) 1934 continue; 1935 1936 if (zone_is_all_unreclaimable(zone) && 1937 priority != DEF_PRIORITY) 1938 continue; 1939 1940 /* 1941 * Do some background aging of the anon list, to give 1942 * pages a chance to be referenced before reclaiming. 1943 */ 1944 if (inactive_anon_is_low(zone, &sc)) 1945 shrink_active_list(SWAP_CLUSTER_MAX, zone, 1946 &sc, priority, 0); 1947 1948 if (!zone_watermark_ok(zone, order, 1949 high_wmark_pages(zone), 0, 0)) { 1950 end_zone = i; 1951 break; 1952 } 1953 } 1954 if (i < 0) 1955 goto out; 1956 1957 for (i = 0; i <= end_zone; i++) { 1958 struct zone *zone = pgdat->node_zones + i; 1959 1960 lru_pages += zone_reclaimable_pages(zone); 1961 } 1962 1963 /* 1964 * Now scan the zone in the dma->highmem direction, stopping 1965 * at the last zone which needs scanning. 1966 * 1967 * We do this because the page allocator works in the opposite 1968 * direction. This prevents the page allocator from allocating 1969 * pages behind kswapd's direction of progress, which would 1970 * cause too much scanning of the lower zones. 1971 */ 1972 for (i = 0; i <= end_zone; i++) { 1973 struct zone *zone = pgdat->node_zones + i; 1974 int nr_slab; 1975 1976 if (!populated_zone(zone)) 1977 continue; 1978 1979 if (zone_is_all_unreclaimable(zone) && 1980 priority != DEF_PRIORITY) 1981 continue; 1982 1983 if (!zone_watermark_ok(zone, order, 1984 high_wmark_pages(zone), end_zone, 0)) 1985 all_zones_ok = 0; 1986 temp_priority[i] = priority; 1987 sc.nr_scanned = 0; 1988 note_zone_scanning_priority(zone, priority); 1989 /* 1990 * We put equal pressure on every zone, unless one 1991 * zone has way too many pages free already. 1992 */ 1993 if (!zone_watermark_ok(zone, order, 1994 8*high_wmark_pages(zone), end_zone, 0)) 1995 shrink_zone(priority, zone, &sc); 1996 reclaim_state->reclaimed_slab = 0; 1997 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL, 1998 lru_pages); 1999 sc.nr_reclaimed += reclaim_state->reclaimed_slab; 2000 total_scanned += sc.nr_scanned; 2001 if (zone_is_all_unreclaimable(zone)) 2002 continue; 2003 if (nr_slab == 0 && zone->pages_scanned >= 2004 (zone_reclaimable_pages(zone) * 6)) 2005 zone_set_flag(zone, 2006 ZONE_ALL_UNRECLAIMABLE); 2007 /* 2008 * If we've done a decent amount of scanning and 2009 * the reclaim ratio is low, start doing writepage 2010 * even in laptop mode 2011 */ 2012 if (total_scanned > SWAP_CLUSTER_MAX * 2 && 2013 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2) 2014 sc.may_writepage = 1; 2015 } 2016 if (all_zones_ok) 2017 break; /* kswapd: all done */ 2018 /* 2019 * OK, kswapd is getting into trouble. Take a nap, then take 2020 * another pass across the zones. 2021 */ 2022 if (total_scanned && priority < DEF_PRIORITY - 2) 2023 congestion_wait(BLK_RW_ASYNC, HZ/10); 2024 2025 /* 2026 * We do this so kswapd doesn't build up large priorities for 2027 * example when it is freeing in parallel with allocators. It 2028 * matches the direct reclaim path behaviour in terms of impact 2029 * on zone->*_priority. 2030 */ 2031 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) 2032 break; 2033 } 2034out: 2035 /* 2036 * Note within each zone the priority level at which this zone was 2037 * brought into a happy state. So that the next thread which scans this 2038 * zone will start out at that priority level. 2039 */ 2040 for (i = 0; i < pgdat->nr_zones; i++) { 2041 struct zone *zone = pgdat->node_zones + i; 2042 2043 zone->prev_priority = temp_priority[i]; 2044 } 2045 if (!all_zones_ok) { 2046 cond_resched(); 2047 2048 try_to_freeze(); 2049 2050 /* 2051 * Fragmentation may mean that the system cannot be 2052 * rebalanced for high-order allocations in all zones. 2053 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX, 2054 * it means the zones have been fully scanned and are still 2055 * not balanced. For high-order allocations, there is 2056 * little point trying all over again as kswapd may 2057 * infinite loop. 2058 * 2059 * Instead, recheck all watermarks at order-0 as they 2060 * are the most important. If watermarks are ok, kswapd will go 2061 * back to sleep. High-order users can still perform direct 2062 * reclaim if they wish. 2063 */ 2064 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX) 2065 order = sc.order = 0; 2066 2067 goto loop_again; 2068 } 2069 2070 return sc.nr_reclaimed; 2071} 2072 2073/* 2074 * The background pageout daemon, started as a kernel thread 2075 * from the init process. 2076 * 2077 * This basically trickles out pages so that we have _some_ 2078 * free memory available even if there is no other activity 2079 * that frees anything up. This is needed for things like routing 2080 * etc, where we otherwise might have all activity going on in 2081 * asynchronous contexts that cannot page things out. 2082 * 2083 * If there are applications that are active memory-allocators 2084 * (most normal use), this basically shouldn't matter. 2085 */ 2086static int kswapd(void *p) 2087{ 2088 unsigned long order; 2089 pg_data_t *pgdat = (pg_data_t*)p; 2090 struct task_struct *tsk = current; 2091 DEFINE_WAIT(wait); 2092 struct reclaim_state reclaim_state = { 2093 .reclaimed_slab = 0, 2094 }; 2095 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); 2096 2097 lockdep_set_current_reclaim_state(GFP_KERNEL); 2098 2099 if (!cpumask_empty(cpumask)) 2100 set_cpus_allowed_ptr(tsk, cpumask); 2101 current->reclaim_state = &reclaim_state; 2102 2103 /* 2104 * Tell the memory management that we're a "memory allocator", 2105 * and that if we need more memory we should get access to it 2106 * regardless (see "__alloc_pages()"). "kswapd" should 2107 * never get caught in the normal page freeing logic. 2108 * 2109 * (Kswapd normally doesn't need memory anyway, but sometimes 2110 * you need a small amount of memory in order to be able to 2111 * page out something else, and this flag essentially protects 2112 * us from recursively trying to free more memory as we're 2113 * trying to free the first piece of memory in the first place). 2114 */ 2115 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; 2116 set_freezable(); 2117 2118 order = 0; 2119 for ( ; ; ) { 2120 unsigned long new_order; 2121 2122 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 2123 new_order = pgdat->kswapd_max_order; 2124 pgdat->kswapd_max_order = 0; 2125 if (order < new_order) { 2126 /* 2127 * Don't sleep if someone wants a larger 'order' 2128 * allocation 2129 */ 2130 order = new_order; 2131 } else { 2132 if (!freezing(current)) 2133 schedule(); 2134 2135 order = pgdat->kswapd_max_order; 2136 } 2137 finish_wait(&pgdat->kswapd_wait, &wait); 2138 2139 if (!try_to_freeze()) { 2140 /* We can speed up thawing tasks if we don't call 2141 * balance_pgdat after returning from the refrigerator 2142 */ 2143 balance_pgdat(pgdat, order); 2144 } 2145 } 2146 return 0; 2147} 2148 2149/* 2150 * A zone is low on free memory, so wake its kswapd task to service it. 2151 */ 2152void wakeup_kswapd(struct zone *zone, int order) 2153{ 2154 pg_data_t *pgdat; 2155 2156 if (!populated_zone(zone)) 2157 return; 2158 2159 pgdat = zone->zone_pgdat; 2160 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0)) 2161 return; 2162 if (pgdat->kswapd_max_order < order) 2163 pgdat->kswapd_max_order = order; 2164 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 2165 return; 2166 if (!waitqueue_active(&pgdat->kswapd_wait)) 2167 return; 2168 wake_up_interruptible(&pgdat->kswapd_wait); 2169} 2170 2171/* 2172 * The reclaimable count would be mostly accurate. 2173 * The less reclaimable pages may be 2174 * - mlocked pages, which will be moved to unevictable list when encountered 2175 * - mapped pages, which may require several travels to be reclaimed 2176 * - dirty pages, which is not "instantly" reclaimable 2177 */ 2178unsigned long global_reclaimable_pages(void) 2179{ 2180 int nr; 2181 2182 nr = global_page_state(NR_ACTIVE_FILE) + 2183 global_page_state(NR_INACTIVE_FILE); 2184 2185 if (nr_swap_pages > 0) 2186 nr += global_page_state(NR_ACTIVE_ANON) + 2187 global_page_state(NR_INACTIVE_ANON); 2188 2189 return nr; 2190} 2191 2192unsigned long zone_reclaimable_pages(struct zone *zone) 2193{ 2194 int nr; 2195 2196 nr = zone_page_state(zone, NR_ACTIVE_FILE) + 2197 zone_page_state(zone, NR_INACTIVE_FILE); 2198 2199 if (nr_swap_pages > 0) 2200 nr += zone_page_state(zone, NR_ACTIVE_ANON) + 2201 zone_page_state(zone, NR_INACTIVE_ANON); 2202 2203 return nr; 2204} 2205 2206#ifdef CONFIG_HIBERNATION 2207/* 2208 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages 2209 * from LRU lists system-wide, for given pass and priority. 2210 * 2211 * For pass > 3 we also try to shrink the LRU lists that contain a few pages 2212 */ 2213static void shrink_all_zones(unsigned long nr_pages, int prio, 2214 int pass, struct scan_control *sc) 2215{ 2216 struct zone *zone; 2217 unsigned long nr_reclaimed = 0; 2218 2219 for_each_populated_zone(zone) { 2220 enum lru_list l; 2221 2222 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY) 2223 continue; 2224 2225 for_each_evictable_lru(l) { 2226 enum zone_stat_item ls = NR_LRU_BASE + l; 2227 unsigned long lru_pages = zone_page_state(zone, ls); 2228 2229 /* For pass = 0, we don't shrink the active list */ 2230 if (pass == 0 && (l == LRU_ACTIVE_ANON || 2231 l == LRU_ACTIVE_FILE)) 2232 continue; 2233 2234 zone->lru[l].nr_saved_scan += (lru_pages >> prio) + 1; 2235 if (zone->lru[l].nr_saved_scan >= nr_pages || pass > 3) { 2236 unsigned long nr_to_scan; 2237 2238 zone->lru[l].nr_saved_scan = 0; 2239 nr_to_scan = min(nr_pages, lru_pages); 2240 nr_reclaimed += shrink_list(l, nr_to_scan, zone, 2241 sc, prio); 2242 if (nr_reclaimed >= nr_pages) { 2243 sc->nr_reclaimed += nr_reclaimed; 2244 return; 2245 } 2246 } 2247 } 2248 } 2249 sc->nr_reclaimed += nr_reclaimed; 2250} 2251 2252/* 2253 * Try to free `nr_pages' of memory, system-wide, and return the number of 2254 * freed pages. 2255 * 2256 * Rather than trying to age LRUs the aim is to preserve the overall 2257 * LRU order by reclaiming preferentially 2258 * inactive > active > active referenced > active mapped 2259 */ 2260unsigned long shrink_all_memory(unsigned long nr_pages) 2261{ 2262 unsigned long lru_pages, nr_slab; 2263 int pass; 2264 struct reclaim_state reclaim_state; 2265 struct scan_control sc = { 2266 .gfp_mask = GFP_KERNEL, 2267 .may_unmap = 0, 2268 .may_writepage = 1, 2269 .isolate_pages = isolate_pages_global, 2270 .nr_reclaimed = 0, 2271 }; 2272 2273 current->reclaim_state = &reclaim_state; 2274 2275 lru_pages = global_reclaimable_pages(); 2276 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE); 2277 /* If slab caches are huge, it's better to hit them first */ 2278 while (nr_slab >= lru_pages) { 2279 reclaim_state.reclaimed_slab = 0; 2280 shrink_slab(nr_pages, sc.gfp_mask, lru_pages); 2281 if (!reclaim_state.reclaimed_slab) 2282 break; 2283 2284 sc.nr_reclaimed += reclaim_state.reclaimed_slab; 2285 if (sc.nr_reclaimed >= nr_pages) 2286 goto out; 2287 2288 nr_slab -= reclaim_state.reclaimed_slab; 2289 } 2290 2291 /* 2292 * We try to shrink LRUs in 5 passes: 2293 * 0 = Reclaim from inactive_list only 2294 * 1 = Reclaim from active list but don't reclaim mapped 2295 * 2 = 2nd pass of type 1 2296 * 3 = Reclaim mapped (normal reclaim) 2297 * 4 = 2nd pass of type 3 2298 */ 2299 for (pass = 0; pass < 5; pass++) { 2300 int prio; 2301 2302 /* Force reclaiming mapped pages in the passes #3 and #4 */ 2303 if (pass > 2) 2304 sc.may_unmap = 1; 2305 2306 for (prio = DEF_PRIORITY; prio >= 0; prio--) { 2307 unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed; 2308 2309 sc.nr_scanned = 0; 2310 sc.swap_cluster_max = nr_to_scan; 2311 shrink_all_zones(nr_to_scan, prio, pass, &sc); 2312 if (sc.nr_reclaimed >= nr_pages) 2313 goto out; 2314 2315 reclaim_state.reclaimed_slab = 0; 2316 shrink_slab(sc.nr_scanned, sc.gfp_mask, 2317 global_reclaimable_pages()); 2318 sc.nr_reclaimed += reclaim_state.reclaimed_slab; 2319 if (sc.nr_reclaimed >= nr_pages) 2320 goto out; 2321 2322 if (sc.nr_scanned && prio < DEF_PRIORITY - 2) 2323 congestion_wait(BLK_RW_ASYNC, HZ / 10); 2324 } 2325 } 2326 2327 /* 2328 * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be 2329 * something in slab caches 2330 */ 2331 if (!sc.nr_reclaimed) { 2332 do { 2333 reclaim_state.reclaimed_slab = 0; 2334 shrink_slab(nr_pages, sc.gfp_mask, 2335 global_reclaimable_pages()); 2336 sc.nr_reclaimed += reclaim_state.reclaimed_slab; 2337 } while (sc.nr_reclaimed < nr_pages && 2338 reclaim_state.reclaimed_slab > 0); 2339 } 2340 2341 2342out: 2343 current->reclaim_state = NULL; 2344 2345 return sc.nr_reclaimed; 2346} 2347#endif /* CONFIG_HIBERNATION */ 2348 2349/* It's optimal to keep kswapds on the same CPUs as their memory, but 2350 not required for correctness. So if the last cpu in a node goes 2351 away, we get changed to run anywhere: as the first one comes back, 2352 restore their cpu bindings. */ 2353static int __devinit cpu_callback(struct notifier_block *nfb, 2354 unsigned long action, void *hcpu) 2355{ 2356 int nid; 2357 2358 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { 2359 for_each_node_state(nid, N_HIGH_MEMORY) { 2360 pg_data_t *pgdat = NODE_DATA(nid); 2361 const struct cpumask *mask; 2362 2363 mask = cpumask_of_node(pgdat->node_id); 2364 2365 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) 2366 /* One of our CPUs online: restore mask */ 2367 set_cpus_allowed_ptr(pgdat->kswapd, mask); 2368 } 2369 } 2370 return NOTIFY_OK; 2371} 2372 2373/* 2374 * This kswapd start function will be called by init and node-hot-add. 2375 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. 2376 */ 2377int kswapd_run(int nid) 2378{ 2379 pg_data_t *pgdat = NODE_DATA(nid); 2380 int ret = 0; 2381 2382 if (pgdat->kswapd) 2383 return 0; 2384 2385 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); 2386 if (IS_ERR(pgdat->kswapd)) { 2387 /* failure at boot is fatal */ 2388 BUG_ON(system_state == SYSTEM_BOOTING); 2389 printk("Failed to start kswapd on node %d\n",nid); 2390 ret = -1; 2391 } 2392 return ret; 2393} 2394 2395static int __init kswapd_init(void) 2396{ 2397 int nid; 2398 2399 swap_setup(); 2400 for_each_node_state(nid, N_HIGH_MEMORY) 2401 kswapd_run(nid); 2402 hotcpu_notifier(cpu_callback, 0); 2403 return 0; 2404} 2405 2406module_init(kswapd_init) 2407 2408#ifdef CONFIG_NUMA 2409/* 2410 * Zone reclaim mode 2411 * 2412 * If non-zero call zone_reclaim when the number of free pages falls below 2413 * the watermarks. 2414 */ 2415int zone_reclaim_mode __read_mostly; 2416 2417#define RECLAIM_OFF 0 2418#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ 2419#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ 2420#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ 2421 2422/* 2423 * Priority for ZONE_RECLAIM. This determines the fraction of pages 2424 * of a node considered for each zone_reclaim. 4 scans 1/16th of 2425 * a zone. 2426 */ 2427#define ZONE_RECLAIM_PRIORITY 4 2428 2429/* 2430 * Percentage of pages in a zone that must be unmapped for zone_reclaim to 2431 * occur. 2432 */ 2433int sysctl_min_unmapped_ratio = 1; 2434 2435/* 2436 * If the number of slab pages in a zone grows beyond this percentage then 2437 * slab reclaim needs to occur. 2438 */ 2439int sysctl_min_slab_ratio = 5; 2440 2441static inline unsigned long zone_unmapped_file_pages(struct zone *zone) 2442{ 2443 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); 2444 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + 2445 zone_page_state(zone, NR_ACTIVE_FILE); 2446 2447 /* 2448 * It's possible for there to be more file mapped pages than 2449 * accounted for by the pages on the file LRU lists because 2450 * tmpfs pages accounted for as ANON can also be FILE_MAPPED 2451 */ 2452 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; 2453} 2454 2455/* Work out how many page cache pages we can reclaim in this reclaim_mode */ 2456static long zone_pagecache_reclaimable(struct zone *zone) 2457{ 2458 long nr_pagecache_reclaimable; 2459 long delta = 0; 2460 2461 /* 2462 * If RECLAIM_SWAP is set, then all file pages are considered 2463 * potentially reclaimable. Otherwise, we have to worry about 2464 * pages like swapcache and zone_unmapped_file_pages() provides 2465 * a better estimate 2466 */ 2467 if (zone_reclaim_mode & RECLAIM_SWAP) 2468 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); 2469 else 2470 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); 2471 2472 /* If we can't clean pages, remove dirty pages from consideration */ 2473 if (!(zone_reclaim_mode & RECLAIM_WRITE)) 2474 delta += zone_page_state(zone, NR_FILE_DIRTY); 2475 2476 /* Watch for any possible underflows due to delta */ 2477 if (unlikely(delta > nr_pagecache_reclaimable)) 2478 delta = nr_pagecache_reclaimable; 2479 2480 return nr_pagecache_reclaimable - delta; 2481} 2482 2483/* 2484 * Try to free up some pages from this zone through reclaim. 2485 */ 2486static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 2487{ 2488 /* Minimum pages needed in order to stay on node */ 2489 const unsigned long nr_pages = 1 << order; 2490 struct task_struct *p = current; 2491 struct reclaim_state reclaim_state; 2492 int priority; 2493 struct scan_control sc = { 2494 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), 2495 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), 2496 .may_swap = 1, 2497 .swap_cluster_max = max_t(unsigned long, nr_pages, 2498 SWAP_CLUSTER_MAX), 2499 .gfp_mask = gfp_mask, 2500 .swappiness = vm_swappiness, 2501 .order = order, 2502 .isolate_pages = isolate_pages_global, 2503 }; 2504 unsigned long slab_reclaimable; 2505 2506 disable_swap_token(); 2507 cond_resched(); 2508 /* 2509 * We need to be able to allocate from the reserves for RECLAIM_SWAP 2510 * and we also need to be able to write out pages for RECLAIM_WRITE 2511 * and RECLAIM_SWAP. 2512 */ 2513 p->flags |= PF_MEMALLOC | PF_SWAPWRITE; 2514 reclaim_state.reclaimed_slab = 0; 2515 p->reclaim_state = &reclaim_state; 2516 2517 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { 2518 /* 2519 * Free memory by calling shrink zone with increasing 2520 * priorities until we have enough memory freed. 2521 */ 2522 priority = ZONE_RECLAIM_PRIORITY; 2523 do { 2524 note_zone_scanning_priority(zone, priority); 2525 shrink_zone(priority, zone, &sc); 2526 priority--; 2527 } while (priority >= 0 && sc.nr_reclaimed < nr_pages); 2528 } 2529 2530 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE); 2531 if (slab_reclaimable > zone->min_slab_pages) { 2532 /* 2533 * shrink_slab() does not currently allow us to determine how 2534 * many pages were freed in this zone. So we take the current 2535 * number of slab pages and shake the slab until it is reduced 2536 * by the same nr_pages that we used for reclaiming unmapped 2537 * pages. 2538 * 2539 * Note that shrink_slab will free memory on all zones and may 2540 * take a long time. 2541 */ 2542 while (shrink_slab(sc.nr_scanned, gfp_mask, order) && 2543 zone_page_state(zone, NR_SLAB_RECLAIMABLE) > 2544 slab_reclaimable - nr_pages) 2545 ; 2546 2547 /* 2548 * Update nr_reclaimed by the number of slab pages we 2549 * reclaimed from this zone. 2550 */ 2551 sc.nr_reclaimed += slab_reclaimable - 2552 zone_page_state(zone, NR_SLAB_RECLAIMABLE); 2553 } 2554 2555 p->reclaim_state = NULL; 2556 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); 2557 return sc.nr_reclaimed >= nr_pages; 2558} 2559 2560int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 2561{ 2562 int node_id; 2563 int ret; 2564 2565 /* 2566 * Zone reclaim reclaims unmapped file backed pages and 2567 * slab pages if we are over the defined limits. 2568 * 2569 * A small portion of unmapped file backed pages is needed for 2570 * file I/O otherwise pages read by file I/O will be immediately 2571 * thrown out if the zone is overallocated. So we do not reclaim 2572 * if less than a specified percentage of the zone is used by 2573 * unmapped file backed pages. 2574 */ 2575 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && 2576 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) 2577 return ZONE_RECLAIM_FULL; 2578 2579 if (zone_is_all_unreclaimable(zone)) 2580 return ZONE_RECLAIM_FULL; 2581 2582 /* 2583 * Do not scan if the allocation should not be delayed. 2584 */ 2585 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) 2586 return ZONE_RECLAIM_NOSCAN; 2587 2588 /* 2589 * Only run zone reclaim on the local zone or on zones that do not 2590 * have associated processors. This will favor the local processor 2591 * over remote processors and spread off node memory allocations 2592 * as wide as possible. 2593 */ 2594 node_id = zone_to_nid(zone); 2595 if (node_state(node_id, N_CPU) && node_id != numa_node_id()) 2596 return ZONE_RECLAIM_NOSCAN; 2597 2598 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) 2599 return ZONE_RECLAIM_NOSCAN; 2600 2601 ret = __zone_reclaim(zone, gfp_mask, order); 2602 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); 2603 2604 if (!ret) 2605 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); 2606 2607 return ret; 2608} 2609#endif 2610 2611/* 2612 * page_evictable - test whether a page is evictable 2613 * @page: the page to test 2614 * @vma: the VMA in which the page is or will be mapped, may be NULL 2615 * 2616 * Test whether page is evictable--i.e., should be placed on active/inactive 2617 * lists vs unevictable list. The vma argument is !NULL when called from the 2618 * fault path to determine how to instantate a new page. 2619 * 2620 * Reasons page might not be evictable: 2621 * (1) page's mapping marked unevictable 2622 * (2) page is part of an mlocked VMA 2623 * 2624 */ 2625int page_evictable(struct page *page, struct vm_area_struct *vma) 2626{ 2627 2628 if (mapping_unevictable(page_mapping(page))) 2629 return 0; 2630 2631 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page))) 2632 return 0; 2633 2634 return 1; 2635} 2636 2637/** 2638 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list 2639 * @page: page to check evictability and move to appropriate lru list 2640 * @zone: zone page is in 2641 * 2642 * Checks a page for evictability and moves the page to the appropriate 2643 * zone lru list. 2644 * 2645 * Restrictions: zone->lru_lock must be held, page must be on LRU and must 2646 * have PageUnevictable set. 2647 */ 2648static void check_move_unevictable_page(struct page *page, struct zone *zone) 2649{ 2650 VM_BUG_ON(PageActive(page)); 2651 2652retry: 2653 ClearPageUnevictable(page); 2654 if (page_evictable(page, NULL)) { 2655 enum lru_list l = page_lru_base_type(page); 2656 2657 __dec_zone_state(zone, NR_UNEVICTABLE); 2658 list_move(&page->lru, &zone->lru[l].list); 2659 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l); 2660 __inc_zone_state(zone, NR_INACTIVE_ANON + l); 2661 __count_vm_event(UNEVICTABLE_PGRESCUED); 2662 } else { 2663 /* 2664 * rotate unevictable list 2665 */ 2666 SetPageUnevictable(page); 2667 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list); 2668 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE); 2669 if (page_evictable(page, NULL)) 2670 goto retry; 2671 } 2672} 2673 2674/** 2675 * scan_mapping_unevictable_pages - scan an address space for evictable pages 2676 * @mapping: struct address_space to scan for evictable pages 2677 * 2678 * Scan all pages in mapping. Check unevictable pages for 2679 * evictability and move them to the appropriate zone lru list. 2680 */ 2681void scan_mapping_unevictable_pages(struct address_space *mapping) 2682{ 2683 pgoff_t next = 0; 2684 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >> 2685 PAGE_CACHE_SHIFT; 2686 struct zone *zone; 2687 struct pagevec pvec; 2688 2689 if (mapping->nrpages == 0) 2690 return; 2691 2692 pagevec_init(&pvec, 0); 2693 while (next < end && 2694 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) { 2695 int i; 2696 int pg_scanned = 0; 2697 2698 zone = NULL; 2699 2700 for (i = 0; i < pagevec_count(&pvec); i++) { 2701 struct page *page = pvec.pages[i]; 2702 pgoff_t page_index = page->index; 2703 struct zone *pagezone = page_zone(page); 2704 2705 pg_scanned++; 2706 if (page_index > next) 2707 next = page_index; 2708 next++; 2709 2710 if (pagezone != zone) { 2711 if (zone) 2712 spin_unlock_irq(&zone->lru_lock); 2713 zone = pagezone; 2714 spin_lock_irq(&zone->lru_lock); 2715 } 2716 2717 if (PageLRU(page) && PageUnevictable(page)) 2718 check_move_unevictable_page(page, zone); 2719 } 2720 if (zone) 2721 spin_unlock_irq(&zone->lru_lock); 2722 pagevec_release(&pvec); 2723 2724 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned); 2725 } 2726 2727} 2728 2729/** 2730 * scan_zone_unevictable_pages - check unevictable list for evictable pages 2731 * @zone - zone of which to scan the unevictable list 2732 * 2733 * Scan @zone's unevictable LRU lists to check for pages that have become 2734 * evictable. Move those that have to @zone's inactive list where they 2735 * become candidates for reclaim, unless shrink_inactive_zone() decides 2736 * to reactivate them. Pages that are still unevictable are rotated 2737 * back onto @zone's unevictable list. 2738 */ 2739#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */ 2740static void scan_zone_unevictable_pages(struct zone *zone) 2741{ 2742 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list; 2743 unsigned long scan; 2744 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE); 2745 2746 while (nr_to_scan > 0) { 2747 unsigned long batch_size = min(nr_to_scan, 2748 SCAN_UNEVICTABLE_BATCH_SIZE); 2749 2750 spin_lock_irq(&zone->lru_lock); 2751 for (scan = 0; scan < batch_size; scan++) { 2752 struct page *page = lru_to_page(l_unevictable); 2753 2754 if (!trylock_page(page)) 2755 continue; 2756 2757 prefetchw_prev_lru_page(page, l_unevictable, flags); 2758 2759 if (likely(PageLRU(page) && PageUnevictable(page))) 2760 check_move_unevictable_page(page, zone); 2761 2762 unlock_page(page); 2763 } 2764 spin_unlock_irq(&zone->lru_lock); 2765 2766 nr_to_scan -= batch_size; 2767 } 2768} 2769 2770 2771/** 2772 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages 2773 * 2774 * A really big hammer: scan all zones' unevictable LRU lists to check for 2775 * pages that have become evictable. Move those back to the zones' 2776 * inactive list where they become candidates for reclaim. 2777 * This occurs when, e.g., we have unswappable pages on the unevictable lists, 2778 * and we add swap to the system. As such, it runs in the context of a task 2779 * that has possibly/probably made some previously unevictable pages 2780 * evictable. 2781 */ 2782static void scan_all_zones_unevictable_pages(void) 2783{ 2784 struct zone *zone; 2785 2786 for_each_zone(zone) { 2787 scan_zone_unevictable_pages(zone); 2788 } 2789} 2790 2791/* 2792 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of 2793 * all nodes' unevictable lists for evictable pages 2794 */ 2795unsigned long scan_unevictable_pages; 2796 2797int scan_unevictable_handler(struct ctl_table *table, int write, 2798 struct file *file, void __user *buffer, 2799 size_t *length, loff_t *ppos) 2800{ 2801 proc_doulongvec_minmax(table, write, file, buffer, length, ppos); 2802 2803 if (write && *(unsigned long *)table->data) 2804 scan_all_zones_unevictable_pages(); 2805 2806 scan_unevictable_pages = 0; 2807 return 0; 2808} 2809 2810/* 2811 * per node 'scan_unevictable_pages' attribute. On demand re-scan of 2812 * a specified node's per zone unevictable lists for evictable pages. 2813 */ 2814 2815static ssize_t read_scan_unevictable_node(struct sys_device *dev, 2816 struct sysdev_attribute *attr, 2817 char *buf) 2818{ 2819 return sprintf(buf, "0\n"); /* always zero; should fit... */ 2820} 2821 2822static ssize_t write_scan_unevictable_node(struct sys_device *dev, 2823 struct sysdev_attribute *attr, 2824 const char *buf, size_t count) 2825{ 2826 struct zone *node_zones = NODE_DATA(dev->id)->node_zones; 2827 struct zone *zone; 2828 unsigned long res; 2829 unsigned long req = strict_strtoul(buf, 10, &res); 2830 2831 if (!req) 2832 return 1; /* zero is no-op */ 2833 2834 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) { 2835 if (!populated_zone(zone)) 2836 continue; 2837 scan_zone_unevictable_pages(zone); 2838 } 2839 return 1; 2840} 2841 2842 2843static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, 2844 read_scan_unevictable_node, 2845 write_scan_unevictable_node); 2846 2847int scan_unevictable_register_node(struct node *node) 2848{ 2849 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages); 2850} 2851 2852void scan_unevictable_unregister_node(struct node *node) 2853{ 2854 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages); 2855} 2856 2857