vmscan.c revision 6fe6b7e35785e3232ffe7f81d3893f1316710a02
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_is_file_cache(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 /* In active use or really unfreeable? Activate it. */ 634 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && 635 referenced && page_mapping_inuse(page)) 636 goto activate_locked; 637 638 /* 639 * Anonymous process memory has backing store? 640 * Try to allocate it some swap space here. 641 */ 642 if (PageAnon(page) && !PageSwapCache(page)) { 643 if (!(sc->gfp_mask & __GFP_IO)) 644 goto keep_locked; 645 if (!add_to_swap(page)) 646 goto activate_locked; 647 may_enter_fs = 1; 648 } 649 650 mapping = page_mapping(page); 651 652 /* 653 * The page is mapped into the page tables of one or more 654 * processes. Try to unmap it here. 655 */ 656 if (page_mapped(page) && mapping) { 657 switch (try_to_unmap(page, 0)) { 658 case SWAP_FAIL: 659 goto activate_locked; 660 case SWAP_AGAIN: 661 goto keep_locked; 662 case SWAP_MLOCK: 663 goto cull_mlocked; 664 case SWAP_SUCCESS: 665 ; /* try to free the page below */ 666 } 667 } 668 669 if (PageDirty(page)) { 670 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced) 671 goto keep_locked; 672 if (!may_enter_fs) 673 goto keep_locked; 674 if (!sc->may_writepage) 675 goto keep_locked; 676 677 /* Page is dirty, try to write it out here */ 678 switch (pageout(page, mapping, sync_writeback)) { 679 case PAGE_KEEP: 680 goto keep_locked; 681 case PAGE_ACTIVATE: 682 goto activate_locked; 683 case PAGE_SUCCESS: 684 if (PageWriteback(page) || PageDirty(page)) 685 goto keep; 686 /* 687 * A synchronous write - probably a ramdisk. Go 688 * ahead and try to reclaim the page. 689 */ 690 if (!trylock_page(page)) 691 goto keep; 692 if (PageDirty(page) || PageWriteback(page)) 693 goto keep_locked; 694 mapping = page_mapping(page); 695 case PAGE_CLEAN: 696 ; /* try to free the page below */ 697 } 698 } 699 700 /* 701 * If the page has buffers, try to free the buffer mappings 702 * associated with this page. If we succeed we try to free 703 * the page as well. 704 * 705 * We do this even if the page is PageDirty(). 706 * try_to_release_page() does not perform I/O, but it is 707 * possible for a page to have PageDirty set, but it is actually 708 * clean (all its buffers are clean). This happens if the 709 * buffers were written out directly, with submit_bh(). ext3 710 * will do this, as well as the blockdev mapping. 711 * try_to_release_page() will discover that cleanness and will 712 * drop the buffers and mark the page clean - it can be freed. 713 * 714 * Rarely, pages can have buffers and no ->mapping. These are 715 * the pages which were not successfully invalidated in 716 * truncate_complete_page(). We try to drop those buffers here 717 * and if that worked, and the page is no longer mapped into 718 * process address space (page_count == 1) it can be freed. 719 * Otherwise, leave the page on the LRU so it is swappable. 720 */ 721 if (page_has_private(page)) { 722 if (!try_to_release_page(page, sc->gfp_mask)) 723 goto activate_locked; 724 if (!mapping && page_count(page) == 1) { 725 unlock_page(page); 726 if (put_page_testzero(page)) 727 goto free_it; 728 else { 729 /* 730 * rare race with speculative reference. 731 * the speculative reference will free 732 * this page shortly, so we may 733 * increment nr_reclaimed here (and 734 * leave it off the LRU). 735 */ 736 nr_reclaimed++; 737 continue; 738 } 739 } 740 } 741 742 if (!mapping || !__remove_mapping(mapping, page)) 743 goto keep_locked; 744 745 /* 746 * At this point, we have no other references and there is 747 * no way to pick any more up (removed from LRU, removed 748 * from pagecache). Can use non-atomic bitops now (and 749 * we obviously don't have to worry about waking up a process 750 * waiting on the page lock, because there are no references. 751 */ 752 __clear_page_locked(page); 753free_it: 754 nr_reclaimed++; 755 if (!pagevec_add(&freed_pvec, page)) { 756 __pagevec_free(&freed_pvec); 757 pagevec_reinit(&freed_pvec); 758 } 759 continue; 760 761cull_mlocked: 762 if (PageSwapCache(page)) 763 try_to_free_swap(page); 764 unlock_page(page); 765 putback_lru_page(page); 766 continue; 767 768activate_locked: 769 /* Not a candidate for swapping, so reclaim swap space. */ 770 if (PageSwapCache(page) && vm_swap_full()) 771 try_to_free_swap(page); 772 VM_BUG_ON(PageActive(page)); 773 SetPageActive(page); 774 pgactivate++; 775keep_locked: 776 unlock_page(page); 777keep: 778 list_add(&page->lru, &ret_pages); 779 VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); 780 } 781 list_splice(&ret_pages, page_list); 782 if (pagevec_count(&freed_pvec)) 783 __pagevec_free(&freed_pvec); 784 count_vm_events(PGACTIVATE, pgactivate); 785 return nr_reclaimed; 786} 787 788/* LRU Isolation modes. */ 789#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */ 790#define ISOLATE_ACTIVE 1 /* Isolate active pages. */ 791#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */ 792 793/* 794 * Attempt to remove the specified page from its LRU. Only take this page 795 * if it is of the appropriate PageActive status. Pages which are being 796 * freed elsewhere are also ignored. 797 * 798 * page: page to consider 799 * mode: one of the LRU isolation modes defined above 800 * 801 * returns 0 on success, -ve errno on failure. 802 */ 803int __isolate_lru_page(struct page *page, int mode, int file) 804{ 805 int ret = -EINVAL; 806 807 /* Only take pages on the LRU. */ 808 if (!PageLRU(page)) 809 return ret; 810 811 /* 812 * When checking the active state, we need to be sure we are 813 * dealing with comparible boolean values. Take the logical not 814 * of each. 815 */ 816 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode)) 817 return ret; 818 819 if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file)) 820 return ret; 821 822 /* 823 * When this function is being called for lumpy reclaim, we 824 * initially look into all LRU pages, active, inactive and 825 * unevictable; only give shrink_page_list evictable pages. 826 */ 827 if (PageUnevictable(page)) 828 return ret; 829 830 ret = -EBUSY; 831 832 if (likely(get_page_unless_zero(page))) { 833 /* 834 * Be careful not to clear PageLRU until after we're 835 * sure the page is not being freed elsewhere -- the 836 * page release code relies on it. 837 */ 838 ClearPageLRU(page); 839 ret = 0; 840 mem_cgroup_del_lru(page); 841 } 842 843 return ret; 844} 845 846/* 847 * zone->lru_lock is heavily contended. Some of the functions that 848 * shrink the lists perform better by taking out a batch of pages 849 * and working on them outside the LRU lock. 850 * 851 * For pagecache intensive workloads, this function is the hottest 852 * spot in the kernel (apart from copy_*_user functions). 853 * 854 * Appropriate locks must be held before calling this function. 855 * 856 * @nr_to_scan: The number of pages to look through on the list. 857 * @src: The LRU list to pull pages off. 858 * @dst: The temp list to put pages on to. 859 * @scanned: The number of pages that were scanned. 860 * @order: The caller's attempted allocation order 861 * @mode: One of the LRU isolation modes 862 * @file: True [1] if isolating file [!anon] pages 863 * 864 * returns how many pages were moved onto *@dst. 865 */ 866static unsigned long isolate_lru_pages(unsigned long nr_to_scan, 867 struct list_head *src, struct list_head *dst, 868 unsigned long *scanned, int order, int mode, int file) 869{ 870 unsigned long nr_taken = 0; 871 unsigned long scan; 872 873 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { 874 struct page *page; 875 unsigned long pfn; 876 unsigned long end_pfn; 877 unsigned long page_pfn; 878 int zone_id; 879 880 page = lru_to_page(src); 881 prefetchw_prev_lru_page(page, src, flags); 882 883 VM_BUG_ON(!PageLRU(page)); 884 885 switch (__isolate_lru_page(page, mode, file)) { 886 case 0: 887 list_move(&page->lru, dst); 888 nr_taken++; 889 break; 890 891 case -EBUSY: 892 /* else it is being freed elsewhere */ 893 list_move(&page->lru, src); 894 continue; 895 896 default: 897 BUG(); 898 } 899 900 if (!order) 901 continue; 902 903 /* 904 * Attempt to take all pages in the order aligned region 905 * surrounding the tag page. Only take those pages of 906 * the same active state as that tag page. We may safely 907 * round the target page pfn down to the requested order 908 * as the mem_map is guarenteed valid out to MAX_ORDER, 909 * where that page is in a different zone we will detect 910 * it from its zone id and abort this block scan. 911 */ 912 zone_id = page_zone_id(page); 913 page_pfn = page_to_pfn(page); 914 pfn = page_pfn & ~((1 << order) - 1); 915 end_pfn = pfn + (1 << order); 916 for (; pfn < end_pfn; pfn++) { 917 struct page *cursor_page; 918 919 /* The target page is in the block, ignore it. */ 920 if (unlikely(pfn == page_pfn)) 921 continue; 922 923 /* Avoid holes within the zone. */ 924 if (unlikely(!pfn_valid_within(pfn))) 925 break; 926 927 cursor_page = pfn_to_page(pfn); 928 929 /* Check that we have not crossed a zone boundary. */ 930 if (unlikely(page_zone_id(cursor_page) != zone_id)) 931 continue; 932 switch (__isolate_lru_page(cursor_page, mode, file)) { 933 case 0: 934 list_move(&cursor_page->lru, dst); 935 nr_taken++; 936 scan++; 937 break; 938 939 case -EBUSY: 940 /* else it is being freed elsewhere */ 941 list_move(&cursor_page->lru, src); 942 default: 943 break; /* ! on LRU or wrong list */ 944 } 945 } 946 } 947 948 *scanned = scan; 949 return nr_taken; 950} 951 952static unsigned long isolate_pages_global(unsigned long nr, 953 struct list_head *dst, 954 unsigned long *scanned, int order, 955 int mode, struct zone *z, 956 struct mem_cgroup *mem_cont, 957 int active, int file) 958{ 959 int lru = LRU_BASE; 960 if (active) 961 lru += LRU_ACTIVE; 962 if (file) 963 lru += LRU_FILE; 964 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order, 965 mode, !!file); 966} 967 968/* 969 * clear_active_flags() is a helper for shrink_active_list(), clearing 970 * any active bits from the pages in the list. 971 */ 972static unsigned long clear_active_flags(struct list_head *page_list, 973 unsigned int *count) 974{ 975 int nr_active = 0; 976 int lru; 977 struct page *page; 978 979 list_for_each_entry(page, page_list, lru) { 980 lru = page_is_file_cache(page); 981 if (PageActive(page)) { 982 lru += LRU_ACTIVE; 983 ClearPageActive(page); 984 nr_active++; 985 } 986 count[lru]++; 987 } 988 989 return nr_active; 990} 991 992/** 993 * isolate_lru_page - tries to isolate a page from its LRU list 994 * @page: page to isolate from its LRU list 995 * 996 * Isolates a @page from an LRU list, clears PageLRU and adjusts the 997 * vmstat statistic corresponding to whatever LRU list the page was on. 998 * 999 * Returns 0 if the page was removed from an LRU list. 1000 * Returns -EBUSY if the page was not on an LRU list. 1001 * 1002 * The returned page will have PageLRU() cleared. If it was found on 1003 * the active list, it will have PageActive set. If it was found on 1004 * the unevictable list, it will have the PageUnevictable bit set. That flag 1005 * may need to be cleared by the caller before letting the page go. 1006 * 1007 * The vmstat statistic corresponding to the list on which the page was 1008 * found will be decremented. 1009 * 1010 * Restrictions: 1011 * (1) Must be called with an elevated refcount on the page. This is a 1012 * fundamentnal difference from isolate_lru_pages (which is called 1013 * without a stable reference). 1014 * (2) the lru_lock must not be held. 1015 * (3) interrupts must be enabled. 1016 */ 1017int isolate_lru_page(struct page *page) 1018{ 1019 int ret = -EBUSY; 1020 1021 if (PageLRU(page)) { 1022 struct zone *zone = page_zone(page); 1023 1024 spin_lock_irq(&zone->lru_lock); 1025 if (PageLRU(page) && get_page_unless_zero(page)) { 1026 int lru = page_lru(page); 1027 ret = 0; 1028 ClearPageLRU(page); 1029 1030 del_page_from_lru_list(zone, page, lru); 1031 } 1032 spin_unlock_irq(&zone->lru_lock); 1033 } 1034 return ret; 1035} 1036 1037/* 1038 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number 1039 * of reclaimed pages 1040 */ 1041static unsigned long shrink_inactive_list(unsigned long max_scan, 1042 struct zone *zone, struct scan_control *sc, 1043 int priority, int file) 1044{ 1045 LIST_HEAD(page_list); 1046 struct pagevec pvec; 1047 unsigned long nr_scanned = 0; 1048 unsigned long nr_reclaimed = 0; 1049 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); 1050 int lumpy_reclaim = 0; 1051 1052 /* 1053 * If we need a large contiguous chunk of memory, or have 1054 * trouble getting a small set of contiguous pages, we 1055 * will reclaim both active and inactive pages. 1056 * 1057 * We use the same threshold as pageout congestion_wait below. 1058 */ 1059 if (sc->order > PAGE_ALLOC_COSTLY_ORDER) 1060 lumpy_reclaim = 1; 1061 else if (sc->order && priority < DEF_PRIORITY - 2) 1062 lumpy_reclaim = 1; 1063 1064 pagevec_init(&pvec, 1); 1065 1066 lru_add_drain(); 1067 spin_lock_irq(&zone->lru_lock); 1068 do { 1069 struct page *page; 1070 unsigned long nr_taken; 1071 unsigned long nr_scan; 1072 unsigned long nr_freed; 1073 unsigned long nr_active; 1074 unsigned int count[NR_LRU_LISTS] = { 0, }; 1075 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE; 1076 1077 nr_taken = sc->isolate_pages(sc->swap_cluster_max, 1078 &page_list, &nr_scan, sc->order, mode, 1079 zone, sc->mem_cgroup, 0, file); 1080 nr_active = clear_active_flags(&page_list, count); 1081 __count_vm_events(PGDEACTIVATE, nr_active); 1082 1083 __mod_zone_page_state(zone, NR_ACTIVE_FILE, 1084 -count[LRU_ACTIVE_FILE]); 1085 __mod_zone_page_state(zone, NR_INACTIVE_FILE, 1086 -count[LRU_INACTIVE_FILE]); 1087 __mod_zone_page_state(zone, NR_ACTIVE_ANON, 1088 -count[LRU_ACTIVE_ANON]); 1089 __mod_zone_page_state(zone, NR_INACTIVE_ANON, 1090 -count[LRU_INACTIVE_ANON]); 1091 1092 if (scanning_global_lru(sc)) 1093 zone->pages_scanned += nr_scan; 1094 1095 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON]; 1096 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON]; 1097 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE]; 1098 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE]; 1099 1100 spin_unlock_irq(&zone->lru_lock); 1101 1102 nr_scanned += nr_scan; 1103 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC); 1104 1105 /* 1106 * If we are direct reclaiming for contiguous pages and we do 1107 * not reclaim everything in the list, try again and wait 1108 * for IO to complete. This will stall high-order allocations 1109 * but that should be acceptable to the caller 1110 */ 1111 if (nr_freed < nr_taken && !current_is_kswapd() && 1112 lumpy_reclaim) { 1113 congestion_wait(WRITE, HZ/10); 1114 1115 /* 1116 * The attempt at page out may have made some 1117 * of the pages active, mark them inactive again. 1118 */ 1119 nr_active = clear_active_flags(&page_list, count); 1120 count_vm_events(PGDEACTIVATE, nr_active); 1121 1122 nr_freed += shrink_page_list(&page_list, sc, 1123 PAGEOUT_IO_SYNC); 1124 } 1125 1126 nr_reclaimed += nr_freed; 1127 local_irq_disable(); 1128 if (current_is_kswapd()) { 1129 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan); 1130 __count_vm_events(KSWAPD_STEAL, nr_freed); 1131 } else if (scanning_global_lru(sc)) 1132 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan); 1133 1134 __count_zone_vm_events(PGSTEAL, zone, nr_freed); 1135 1136 if (nr_taken == 0) 1137 goto done; 1138 1139 spin_lock(&zone->lru_lock); 1140 /* 1141 * Put back any unfreeable pages. 1142 */ 1143 while (!list_empty(&page_list)) { 1144 int lru; 1145 page = lru_to_page(&page_list); 1146 VM_BUG_ON(PageLRU(page)); 1147 list_del(&page->lru); 1148 if (unlikely(!page_evictable(page, NULL))) { 1149 spin_unlock_irq(&zone->lru_lock); 1150 putback_lru_page(page); 1151 spin_lock_irq(&zone->lru_lock); 1152 continue; 1153 } 1154 SetPageLRU(page); 1155 lru = page_lru(page); 1156 add_page_to_lru_list(zone, page, lru); 1157 if (PageActive(page)) { 1158 int file = !!page_is_file_cache(page); 1159 reclaim_stat->recent_rotated[file]++; 1160 } 1161 if (!pagevec_add(&pvec, page)) { 1162 spin_unlock_irq(&zone->lru_lock); 1163 __pagevec_release(&pvec); 1164 spin_lock_irq(&zone->lru_lock); 1165 } 1166 } 1167 } while (nr_scanned < max_scan); 1168 spin_unlock(&zone->lru_lock); 1169done: 1170 local_irq_enable(); 1171 pagevec_release(&pvec); 1172 return nr_reclaimed; 1173} 1174 1175/* 1176 * We are about to scan this zone at a certain priority level. If that priority 1177 * level is smaller (ie: more urgent) than the previous priority, then note 1178 * that priority level within the zone. This is done so that when the next 1179 * process comes in to scan this zone, it will immediately start out at this 1180 * priority level rather than having to build up its own scanning priority. 1181 * Here, this priority affects only the reclaim-mapped threshold. 1182 */ 1183static inline void note_zone_scanning_priority(struct zone *zone, int priority) 1184{ 1185 if (priority < zone->prev_priority) 1186 zone->prev_priority = priority; 1187} 1188 1189/* 1190 * This moves pages from the active list to the inactive list. 1191 * 1192 * We move them the other way if the page is referenced by one or more 1193 * processes, from rmap. 1194 * 1195 * If the pages are mostly unmapped, the processing is fast and it is 1196 * appropriate to hold zone->lru_lock across the whole operation. But if 1197 * the pages are mapped, the processing is slow (page_referenced()) so we 1198 * should drop zone->lru_lock around each page. It's impossible to balance 1199 * this, so instead we remove the pages from the LRU while processing them. 1200 * It is safe to rely on PG_active against the non-LRU pages in here because 1201 * nobody will play with that bit on a non-LRU page. 1202 * 1203 * The downside is that we have to touch page->_count against each page. 1204 * But we had to alter page->flags anyway. 1205 */ 1206 1207 1208static void shrink_active_list(unsigned long nr_pages, struct zone *zone, 1209 struct scan_control *sc, int priority, int file) 1210{ 1211 unsigned long pgmoved; 1212 unsigned long pgscanned; 1213 unsigned long vm_flags; 1214 LIST_HEAD(l_hold); /* The pages which were snipped off */ 1215 LIST_HEAD(l_inactive); 1216 struct page *page; 1217 struct pagevec pvec; 1218 enum lru_list lru; 1219 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); 1220 1221 lru_add_drain(); 1222 spin_lock_irq(&zone->lru_lock); 1223 pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order, 1224 ISOLATE_ACTIVE, zone, 1225 sc->mem_cgroup, 1, file); 1226 /* 1227 * zone->pages_scanned is used for detect zone's oom 1228 * mem_cgroup remembers nr_scan by itself. 1229 */ 1230 if (scanning_global_lru(sc)) { 1231 zone->pages_scanned += pgscanned; 1232 } 1233 reclaim_stat->recent_scanned[!!file] += pgmoved; 1234 1235 if (file) 1236 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved); 1237 else 1238 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved); 1239 spin_unlock_irq(&zone->lru_lock); 1240 1241 pgmoved = 0; /* count referenced (mapping) mapped pages */ 1242 while (!list_empty(&l_hold)) { 1243 cond_resched(); 1244 page = lru_to_page(&l_hold); 1245 list_del(&page->lru); 1246 1247 if (unlikely(!page_evictable(page, NULL))) { 1248 putback_lru_page(page); 1249 continue; 1250 } 1251 1252 /* page_referenced clears PageReferenced */ 1253 if (page_mapping_inuse(page) && 1254 page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) 1255 pgmoved++; 1256 1257 list_add(&page->lru, &l_inactive); 1258 } 1259 1260 /* 1261 * Move the pages to the [file or anon] inactive list. 1262 */ 1263 pagevec_init(&pvec, 1); 1264 lru = LRU_BASE + file * LRU_FILE; 1265 1266 spin_lock_irq(&zone->lru_lock); 1267 /* 1268 * Count referenced pages from currently used mappings as 1269 * rotated, even though they are moved to the inactive list. 1270 * This helps balance scan pressure between file and anonymous 1271 * pages in get_scan_ratio. 1272 */ 1273 reclaim_stat->recent_rotated[!!file] += pgmoved; 1274 1275 pgmoved = 0; /* count pages moved to inactive list */ 1276 while (!list_empty(&l_inactive)) { 1277 page = lru_to_page(&l_inactive); 1278 prefetchw_prev_lru_page(page, &l_inactive, flags); 1279 VM_BUG_ON(PageLRU(page)); 1280 SetPageLRU(page); 1281 VM_BUG_ON(!PageActive(page)); 1282 ClearPageActive(page); 1283 1284 list_move(&page->lru, &zone->lru[lru].list); 1285 mem_cgroup_add_lru_list(page, lru); 1286 pgmoved++; 1287 if (!pagevec_add(&pvec, page)) { 1288 spin_unlock_irq(&zone->lru_lock); 1289 if (buffer_heads_over_limit) 1290 pagevec_strip(&pvec); 1291 __pagevec_release(&pvec); 1292 spin_lock_irq(&zone->lru_lock); 1293 } 1294 } 1295 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); 1296 __count_zone_vm_events(PGREFILL, zone, pgscanned); 1297 __count_vm_events(PGDEACTIVATE, pgmoved); 1298 spin_unlock_irq(&zone->lru_lock); 1299 if (buffer_heads_over_limit) 1300 pagevec_strip(&pvec); 1301 pagevec_release(&pvec); 1302} 1303 1304static int inactive_anon_is_low_global(struct zone *zone) 1305{ 1306 unsigned long active, inactive; 1307 1308 active = zone_page_state(zone, NR_ACTIVE_ANON); 1309 inactive = zone_page_state(zone, NR_INACTIVE_ANON); 1310 1311 if (inactive * zone->inactive_ratio < active) 1312 return 1; 1313 1314 return 0; 1315} 1316 1317/** 1318 * inactive_anon_is_low - check if anonymous pages need to be deactivated 1319 * @zone: zone to check 1320 * @sc: scan control of this context 1321 * 1322 * Returns true if the zone does not have enough inactive anon pages, 1323 * meaning some active anon pages need to be deactivated. 1324 */ 1325static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc) 1326{ 1327 int low; 1328 1329 if (scanning_global_lru(sc)) 1330 low = inactive_anon_is_low_global(zone); 1331 else 1332 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup); 1333 return low; 1334} 1335 1336static int inactive_file_is_low_global(struct zone *zone) 1337{ 1338 unsigned long active, inactive; 1339 1340 active = zone_page_state(zone, NR_ACTIVE_FILE); 1341 inactive = zone_page_state(zone, NR_INACTIVE_FILE); 1342 1343 return (active > inactive); 1344} 1345 1346/** 1347 * inactive_file_is_low - check if file pages need to be deactivated 1348 * @zone: zone to check 1349 * @sc: scan control of this context 1350 * 1351 * When the system is doing streaming IO, memory pressure here 1352 * ensures that active file pages get deactivated, until more 1353 * than half of the file pages are on the inactive list. 1354 * 1355 * Once we get to that situation, protect the system's working 1356 * set from being evicted by disabling active file page aging. 1357 * 1358 * This uses a different ratio than the anonymous pages, because 1359 * the page cache uses a use-once replacement algorithm. 1360 */ 1361static int inactive_file_is_low(struct zone *zone, struct scan_control *sc) 1362{ 1363 int low; 1364 1365 if (scanning_global_lru(sc)) 1366 low = inactive_file_is_low_global(zone); 1367 else 1368 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup); 1369 return low; 1370} 1371 1372static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, 1373 struct zone *zone, struct scan_control *sc, int priority) 1374{ 1375 int file = is_file_lru(lru); 1376 1377 if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) { 1378 shrink_active_list(nr_to_scan, zone, sc, priority, file); 1379 return 0; 1380 } 1381 1382 if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) { 1383 shrink_active_list(nr_to_scan, zone, sc, priority, file); 1384 return 0; 1385 } 1386 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file); 1387} 1388 1389/* 1390 * Determine how aggressively the anon and file LRU lists should be 1391 * scanned. The relative value of each set of LRU lists is determined 1392 * by looking at the fraction of the pages scanned we did rotate back 1393 * onto the active list instead of evict. 1394 * 1395 * percent[0] specifies how much pressure to put on ram/swap backed 1396 * memory, while percent[1] determines pressure on the file LRUs. 1397 */ 1398static void get_scan_ratio(struct zone *zone, struct scan_control *sc, 1399 unsigned long *percent) 1400{ 1401 unsigned long anon, file, free; 1402 unsigned long anon_prio, file_prio; 1403 unsigned long ap, fp; 1404 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); 1405 1406 /* If we have no swap space, do not bother scanning anon pages. */ 1407 if (!sc->may_swap || (nr_swap_pages <= 0)) { 1408 percent[0] = 0; 1409 percent[1] = 100; 1410 return; 1411 } 1412 1413 anon = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) + 1414 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON); 1415 file = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) + 1416 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE); 1417 1418 if (scanning_global_lru(sc)) { 1419 free = zone_page_state(zone, NR_FREE_PAGES); 1420 /* If we have very few page cache pages, 1421 force-scan anon pages. */ 1422 if (unlikely(file + free <= high_wmark_pages(zone))) { 1423 percent[0] = 100; 1424 percent[1] = 0; 1425 return; 1426 } 1427 } 1428 1429 /* 1430 * OK, so we have swap space and a fair amount of page cache 1431 * pages. We use the recently rotated / recently scanned 1432 * ratios to determine how valuable each cache is. 1433 * 1434 * Because workloads change over time (and to avoid overflow) 1435 * we keep these statistics as a floating average, which ends 1436 * up weighing recent references more than old ones. 1437 * 1438 * anon in [0], file in [1] 1439 */ 1440 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { 1441 spin_lock_irq(&zone->lru_lock); 1442 reclaim_stat->recent_scanned[0] /= 2; 1443 reclaim_stat->recent_rotated[0] /= 2; 1444 spin_unlock_irq(&zone->lru_lock); 1445 } 1446 1447 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { 1448 spin_lock_irq(&zone->lru_lock); 1449 reclaim_stat->recent_scanned[1] /= 2; 1450 reclaim_stat->recent_rotated[1] /= 2; 1451 spin_unlock_irq(&zone->lru_lock); 1452 } 1453 1454 /* 1455 * With swappiness at 100, anonymous and file have the same priority. 1456 * This scanning priority is essentially the inverse of IO cost. 1457 */ 1458 anon_prio = sc->swappiness; 1459 file_prio = 200 - sc->swappiness; 1460 1461 /* 1462 * The amount of pressure on anon vs file pages is inversely 1463 * proportional to the fraction of recently scanned pages on 1464 * each list that were recently referenced and in active use. 1465 */ 1466 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1); 1467 ap /= reclaim_stat->recent_rotated[0] + 1; 1468 1469 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1); 1470 fp /= reclaim_stat->recent_rotated[1] + 1; 1471 1472 /* Normalize to percentages */ 1473 percent[0] = 100 * ap / (ap + fp + 1); 1474 percent[1] = 100 - percent[0]; 1475} 1476 1477/* 1478 * Smallish @nr_to_scan's are deposited in @nr_saved_scan, 1479 * until we collected @swap_cluster_max pages to scan. 1480 */ 1481static unsigned long nr_scan_try_batch(unsigned long nr_to_scan, 1482 unsigned long *nr_saved_scan, 1483 unsigned long swap_cluster_max) 1484{ 1485 unsigned long nr; 1486 1487 *nr_saved_scan += nr_to_scan; 1488 nr = *nr_saved_scan; 1489 1490 if (nr >= swap_cluster_max) 1491 *nr_saved_scan = 0; 1492 else 1493 nr = 0; 1494 1495 return nr; 1496} 1497 1498/* 1499 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. 1500 */ 1501static void shrink_zone(int priority, struct zone *zone, 1502 struct scan_control *sc) 1503{ 1504 unsigned long nr[NR_LRU_LISTS]; 1505 unsigned long nr_to_scan; 1506 unsigned long percent[2]; /* anon @ 0; file @ 1 */ 1507 enum lru_list l; 1508 unsigned long nr_reclaimed = sc->nr_reclaimed; 1509 unsigned long swap_cluster_max = sc->swap_cluster_max; 1510 1511 get_scan_ratio(zone, sc, percent); 1512 1513 for_each_evictable_lru(l) { 1514 int file = is_file_lru(l); 1515 unsigned long scan; 1516 1517 scan = zone_nr_pages(zone, sc, l); 1518 if (priority) { 1519 scan >>= priority; 1520 scan = (scan * percent[file]) / 100; 1521 } 1522 if (scanning_global_lru(sc)) 1523 nr[l] = nr_scan_try_batch(scan, 1524 &zone->lru[l].nr_saved_scan, 1525 swap_cluster_max); 1526 else 1527 nr[l] = scan; 1528 } 1529 1530 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || 1531 nr[LRU_INACTIVE_FILE]) { 1532 for_each_evictable_lru(l) { 1533 if (nr[l]) { 1534 nr_to_scan = min(nr[l], swap_cluster_max); 1535 nr[l] -= nr_to_scan; 1536 1537 nr_reclaimed += shrink_list(l, nr_to_scan, 1538 zone, sc, priority); 1539 } 1540 } 1541 /* 1542 * On large memory systems, scan >> priority can become 1543 * really large. This is fine for the starting priority; 1544 * we want to put equal scanning pressure on each zone. 1545 * However, if the VM has a harder time of freeing pages, 1546 * with multiple processes reclaiming pages, the total 1547 * freeing target can get unreasonably large. 1548 */ 1549 if (nr_reclaimed > swap_cluster_max && 1550 priority < DEF_PRIORITY && !current_is_kswapd()) 1551 break; 1552 } 1553 1554 sc->nr_reclaimed = nr_reclaimed; 1555 1556 /* 1557 * Even if we did not try to evict anon pages at all, we want to 1558 * rebalance the anon lru active/inactive ratio. 1559 */ 1560 if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0) 1561 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0); 1562 1563 throttle_vm_writeout(sc->gfp_mask); 1564} 1565 1566/* 1567 * This is the direct reclaim path, for page-allocating processes. We only 1568 * try to reclaim pages from zones which will satisfy the caller's allocation 1569 * request. 1570 * 1571 * We reclaim from a zone even if that zone is over high_wmark_pages(zone). 1572 * Because: 1573 * a) The caller may be trying to free *extra* pages to satisfy a higher-order 1574 * allocation or 1575 * b) The target zone may be at high_wmark_pages(zone) but the lower zones 1576 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' 1577 * zone defense algorithm. 1578 * 1579 * If a zone is deemed to be full of pinned pages then just give it a light 1580 * scan then give up on it. 1581 */ 1582static void shrink_zones(int priority, struct zonelist *zonelist, 1583 struct scan_control *sc) 1584{ 1585 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask); 1586 struct zoneref *z; 1587 struct zone *zone; 1588 1589 sc->all_unreclaimable = 1; 1590 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx, 1591 sc->nodemask) { 1592 if (!populated_zone(zone)) 1593 continue; 1594 /* 1595 * Take care memory controller reclaiming has small influence 1596 * to global LRU. 1597 */ 1598 if (scanning_global_lru(sc)) { 1599 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 1600 continue; 1601 note_zone_scanning_priority(zone, priority); 1602 1603 if (zone_is_all_unreclaimable(zone) && 1604 priority != DEF_PRIORITY) 1605 continue; /* Let kswapd poll it */ 1606 sc->all_unreclaimable = 0; 1607 } else { 1608 /* 1609 * Ignore cpuset limitation here. We just want to reduce 1610 * # of used pages by us regardless of memory shortage. 1611 */ 1612 sc->all_unreclaimable = 0; 1613 mem_cgroup_note_reclaim_priority(sc->mem_cgroup, 1614 priority); 1615 } 1616 1617 shrink_zone(priority, zone, sc); 1618 } 1619} 1620 1621/* 1622 * This is the main entry point to direct page reclaim. 1623 * 1624 * If a full scan of the inactive list fails to free enough memory then we 1625 * are "out of memory" and something needs to be killed. 1626 * 1627 * If the caller is !__GFP_FS then the probability of a failure is reasonably 1628 * high - the zone may be full of dirty or under-writeback pages, which this 1629 * caller can't do much about. We kick pdflush and take explicit naps in the 1630 * hope that some of these pages can be written. But if the allocating task 1631 * holds filesystem locks which prevent writeout this might not work, and the 1632 * allocation attempt will fail. 1633 * 1634 * returns: 0, if no pages reclaimed 1635 * else, the number of pages reclaimed 1636 */ 1637static unsigned long do_try_to_free_pages(struct zonelist *zonelist, 1638 struct scan_control *sc) 1639{ 1640 int priority; 1641 unsigned long ret = 0; 1642 unsigned long total_scanned = 0; 1643 struct reclaim_state *reclaim_state = current->reclaim_state; 1644 unsigned long lru_pages = 0; 1645 struct zoneref *z; 1646 struct zone *zone; 1647 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask); 1648 1649 delayacct_freepages_start(); 1650 1651 if (scanning_global_lru(sc)) 1652 count_vm_event(ALLOCSTALL); 1653 /* 1654 * mem_cgroup will not do shrink_slab. 1655 */ 1656 if (scanning_global_lru(sc)) { 1657 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1658 1659 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 1660 continue; 1661 1662 lru_pages += zone_lru_pages(zone); 1663 } 1664 } 1665 1666 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 1667 sc->nr_scanned = 0; 1668 if (!priority) 1669 disable_swap_token(); 1670 shrink_zones(priority, zonelist, sc); 1671 /* 1672 * Don't shrink slabs when reclaiming memory from 1673 * over limit cgroups 1674 */ 1675 if (scanning_global_lru(sc)) { 1676 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages); 1677 if (reclaim_state) { 1678 sc->nr_reclaimed += reclaim_state->reclaimed_slab; 1679 reclaim_state->reclaimed_slab = 0; 1680 } 1681 } 1682 total_scanned += sc->nr_scanned; 1683 if (sc->nr_reclaimed >= sc->swap_cluster_max) { 1684 ret = sc->nr_reclaimed; 1685 goto out; 1686 } 1687 1688 /* 1689 * Try to write back as many pages as we just scanned. This 1690 * tends to cause slow streaming writers to write data to the 1691 * disk smoothly, at the dirtying rate, which is nice. But 1692 * that's undesirable in laptop mode, where we *want* lumpy 1693 * writeout. So in laptop mode, write out the whole world. 1694 */ 1695 if (total_scanned > sc->swap_cluster_max + 1696 sc->swap_cluster_max / 2) { 1697 wakeup_pdflush(laptop_mode ? 0 : total_scanned); 1698 sc->may_writepage = 1; 1699 } 1700 1701 /* Take a nap, wait for some writeback to complete */ 1702 if (sc->nr_scanned && priority < DEF_PRIORITY - 2) 1703 congestion_wait(WRITE, HZ/10); 1704 } 1705 /* top priority shrink_zones still had more to do? don't OOM, then */ 1706 if (!sc->all_unreclaimable && scanning_global_lru(sc)) 1707 ret = sc->nr_reclaimed; 1708out: 1709 /* 1710 * Now that we've scanned all the zones at this priority level, note 1711 * that level within the zone so that the next thread which performs 1712 * scanning of this zone will immediately start out at this priority 1713 * level. This affects only the decision whether or not to bring 1714 * mapped pages onto the inactive list. 1715 */ 1716 if (priority < 0) 1717 priority = 0; 1718 1719 if (scanning_global_lru(sc)) { 1720 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1721 1722 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 1723 continue; 1724 1725 zone->prev_priority = priority; 1726 } 1727 } else 1728 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority); 1729 1730 delayacct_freepages_end(); 1731 1732 return ret; 1733} 1734 1735unsigned long try_to_free_pages(struct zonelist *zonelist, int order, 1736 gfp_t gfp_mask, nodemask_t *nodemask) 1737{ 1738 struct scan_control sc = { 1739 .gfp_mask = gfp_mask, 1740 .may_writepage = !laptop_mode, 1741 .swap_cluster_max = SWAP_CLUSTER_MAX, 1742 .may_unmap = 1, 1743 .may_swap = 1, 1744 .swappiness = vm_swappiness, 1745 .order = order, 1746 .mem_cgroup = NULL, 1747 .isolate_pages = isolate_pages_global, 1748 .nodemask = nodemask, 1749 }; 1750 1751 return do_try_to_free_pages(zonelist, &sc); 1752} 1753 1754#ifdef CONFIG_CGROUP_MEM_RES_CTLR 1755 1756unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont, 1757 gfp_t gfp_mask, 1758 bool noswap, 1759 unsigned int swappiness) 1760{ 1761 struct scan_control sc = { 1762 .may_writepage = !laptop_mode, 1763 .may_unmap = 1, 1764 .may_swap = !noswap, 1765 .swap_cluster_max = SWAP_CLUSTER_MAX, 1766 .swappiness = swappiness, 1767 .order = 0, 1768 .mem_cgroup = mem_cont, 1769 .isolate_pages = mem_cgroup_isolate_pages, 1770 .nodemask = NULL, /* we don't care the placement */ 1771 }; 1772 struct zonelist *zonelist; 1773 1774 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | 1775 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); 1776 zonelist = NODE_DATA(numa_node_id())->node_zonelists; 1777 return do_try_to_free_pages(zonelist, &sc); 1778} 1779#endif 1780 1781/* 1782 * For kswapd, balance_pgdat() will work across all this node's zones until 1783 * they are all at high_wmark_pages(zone). 1784 * 1785 * Returns the number of pages which were actually freed. 1786 * 1787 * There is special handling here for zones which are full of pinned pages. 1788 * This can happen if the pages are all mlocked, or if they are all used by 1789 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. 1790 * What we do is to detect the case where all pages in the zone have been 1791 * scanned twice and there has been zero successful reclaim. Mark the zone as 1792 * dead and from now on, only perform a short scan. Basically we're polling 1793 * the zone for when the problem goes away. 1794 * 1795 * kswapd scans the zones in the highmem->normal->dma direction. It skips 1796 * zones which have free_pages > high_wmark_pages(zone), but once a zone is 1797 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the 1798 * lower zones regardless of the number of free pages in the lower zones. This 1799 * interoperates with the page allocator fallback scheme to ensure that aging 1800 * of pages is balanced across the zones. 1801 */ 1802static unsigned long balance_pgdat(pg_data_t *pgdat, int order) 1803{ 1804 int all_zones_ok; 1805 int priority; 1806 int i; 1807 unsigned long total_scanned; 1808 struct reclaim_state *reclaim_state = current->reclaim_state; 1809 struct scan_control sc = { 1810 .gfp_mask = GFP_KERNEL, 1811 .may_unmap = 1, 1812 .may_swap = 1, 1813 .swap_cluster_max = SWAP_CLUSTER_MAX, 1814 .swappiness = vm_swappiness, 1815 .order = order, 1816 .mem_cgroup = NULL, 1817 .isolate_pages = isolate_pages_global, 1818 }; 1819 /* 1820 * temp_priority is used to remember the scanning priority at which 1821 * this zone was successfully refilled to 1822 * free_pages == high_wmark_pages(zone). 1823 */ 1824 int temp_priority[MAX_NR_ZONES]; 1825 1826loop_again: 1827 total_scanned = 0; 1828 sc.nr_reclaimed = 0; 1829 sc.may_writepage = !laptop_mode; 1830 count_vm_event(PAGEOUTRUN); 1831 1832 for (i = 0; i < pgdat->nr_zones; i++) 1833 temp_priority[i] = DEF_PRIORITY; 1834 1835 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 1836 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ 1837 unsigned long lru_pages = 0; 1838 1839 /* The swap token gets in the way of swapout... */ 1840 if (!priority) 1841 disable_swap_token(); 1842 1843 all_zones_ok = 1; 1844 1845 /* 1846 * Scan in the highmem->dma direction for the highest 1847 * zone which needs scanning 1848 */ 1849 for (i = pgdat->nr_zones - 1; i >= 0; i--) { 1850 struct zone *zone = pgdat->node_zones + i; 1851 1852 if (!populated_zone(zone)) 1853 continue; 1854 1855 if (zone_is_all_unreclaimable(zone) && 1856 priority != DEF_PRIORITY) 1857 continue; 1858 1859 /* 1860 * Do some background aging of the anon list, to give 1861 * pages a chance to be referenced before reclaiming. 1862 */ 1863 if (inactive_anon_is_low(zone, &sc)) 1864 shrink_active_list(SWAP_CLUSTER_MAX, zone, 1865 &sc, priority, 0); 1866 1867 if (!zone_watermark_ok(zone, order, 1868 high_wmark_pages(zone), 0, 0)) { 1869 end_zone = i; 1870 break; 1871 } 1872 } 1873 if (i < 0) 1874 goto out; 1875 1876 for (i = 0; i <= end_zone; i++) { 1877 struct zone *zone = pgdat->node_zones + i; 1878 1879 lru_pages += zone_lru_pages(zone); 1880 } 1881 1882 /* 1883 * Now scan the zone in the dma->highmem direction, stopping 1884 * at the last zone which needs scanning. 1885 * 1886 * We do this because the page allocator works in the opposite 1887 * direction. This prevents the page allocator from allocating 1888 * pages behind kswapd's direction of progress, which would 1889 * cause too much scanning of the lower zones. 1890 */ 1891 for (i = 0; i <= end_zone; i++) { 1892 struct zone *zone = pgdat->node_zones + i; 1893 int nr_slab; 1894 1895 if (!populated_zone(zone)) 1896 continue; 1897 1898 if (zone_is_all_unreclaimable(zone) && 1899 priority != DEF_PRIORITY) 1900 continue; 1901 1902 if (!zone_watermark_ok(zone, order, 1903 high_wmark_pages(zone), end_zone, 0)) 1904 all_zones_ok = 0; 1905 temp_priority[i] = priority; 1906 sc.nr_scanned = 0; 1907 note_zone_scanning_priority(zone, priority); 1908 /* 1909 * We put equal pressure on every zone, unless one 1910 * zone has way too many pages free already. 1911 */ 1912 if (!zone_watermark_ok(zone, order, 1913 8*high_wmark_pages(zone), end_zone, 0)) 1914 shrink_zone(priority, zone, &sc); 1915 reclaim_state->reclaimed_slab = 0; 1916 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL, 1917 lru_pages); 1918 sc.nr_reclaimed += reclaim_state->reclaimed_slab; 1919 total_scanned += sc.nr_scanned; 1920 if (zone_is_all_unreclaimable(zone)) 1921 continue; 1922 if (nr_slab == 0 && zone->pages_scanned >= 1923 (zone_lru_pages(zone) * 6)) 1924 zone_set_flag(zone, 1925 ZONE_ALL_UNRECLAIMABLE); 1926 /* 1927 * If we've done a decent amount of scanning and 1928 * the reclaim ratio is low, start doing writepage 1929 * even in laptop mode 1930 */ 1931 if (total_scanned > SWAP_CLUSTER_MAX * 2 && 1932 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2) 1933 sc.may_writepage = 1; 1934 } 1935 if (all_zones_ok) 1936 break; /* kswapd: all done */ 1937 /* 1938 * OK, kswapd is getting into trouble. Take a nap, then take 1939 * another pass across the zones. 1940 */ 1941 if (total_scanned && priority < DEF_PRIORITY - 2) 1942 congestion_wait(WRITE, HZ/10); 1943 1944 /* 1945 * We do this so kswapd doesn't build up large priorities for 1946 * example when it is freeing in parallel with allocators. It 1947 * matches the direct reclaim path behaviour in terms of impact 1948 * on zone->*_priority. 1949 */ 1950 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) 1951 break; 1952 } 1953out: 1954 /* 1955 * Note within each zone the priority level at which this zone was 1956 * brought into a happy state. So that the next thread which scans this 1957 * zone will start out at that priority level. 1958 */ 1959 for (i = 0; i < pgdat->nr_zones; i++) { 1960 struct zone *zone = pgdat->node_zones + i; 1961 1962 zone->prev_priority = temp_priority[i]; 1963 } 1964 if (!all_zones_ok) { 1965 cond_resched(); 1966 1967 try_to_freeze(); 1968 1969 /* 1970 * Fragmentation may mean that the system cannot be 1971 * rebalanced for high-order allocations in all zones. 1972 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX, 1973 * it means the zones have been fully scanned and are still 1974 * not balanced. For high-order allocations, there is 1975 * little point trying all over again as kswapd may 1976 * infinite loop. 1977 * 1978 * Instead, recheck all watermarks at order-0 as they 1979 * are the most important. If watermarks are ok, kswapd will go 1980 * back to sleep. High-order users can still perform direct 1981 * reclaim if they wish. 1982 */ 1983 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX) 1984 order = sc.order = 0; 1985 1986 goto loop_again; 1987 } 1988 1989 return sc.nr_reclaimed; 1990} 1991 1992/* 1993 * The background pageout daemon, started as a kernel thread 1994 * from the init process. 1995 * 1996 * This basically trickles out pages so that we have _some_ 1997 * free memory available even if there is no other activity 1998 * that frees anything up. This is needed for things like routing 1999 * etc, where we otherwise might have all activity going on in 2000 * asynchronous contexts that cannot page things out. 2001 * 2002 * If there are applications that are active memory-allocators 2003 * (most normal use), this basically shouldn't matter. 2004 */ 2005static int kswapd(void *p) 2006{ 2007 unsigned long order; 2008 pg_data_t *pgdat = (pg_data_t*)p; 2009 struct task_struct *tsk = current; 2010 DEFINE_WAIT(wait); 2011 struct reclaim_state reclaim_state = { 2012 .reclaimed_slab = 0, 2013 }; 2014 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); 2015 2016 lockdep_set_current_reclaim_state(GFP_KERNEL); 2017 2018 if (!cpumask_empty(cpumask)) 2019 set_cpus_allowed_ptr(tsk, cpumask); 2020 current->reclaim_state = &reclaim_state; 2021 2022 /* 2023 * Tell the memory management that we're a "memory allocator", 2024 * and that if we need more memory we should get access to it 2025 * regardless (see "__alloc_pages()"). "kswapd" should 2026 * never get caught in the normal page freeing logic. 2027 * 2028 * (Kswapd normally doesn't need memory anyway, but sometimes 2029 * you need a small amount of memory in order to be able to 2030 * page out something else, and this flag essentially protects 2031 * us from recursively trying to free more memory as we're 2032 * trying to free the first piece of memory in the first place). 2033 */ 2034 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; 2035 set_freezable(); 2036 2037 order = 0; 2038 for ( ; ; ) { 2039 unsigned long new_order; 2040 2041 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 2042 new_order = pgdat->kswapd_max_order; 2043 pgdat->kswapd_max_order = 0; 2044 if (order < new_order) { 2045 /* 2046 * Don't sleep if someone wants a larger 'order' 2047 * allocation 2048 */ 2049 order = new_order; 2050 } else { 2051 if (!freezing(current)) 2052 schedule(); 2053 2054 order = pgdat->kswapd_max_order; 2055 } 2056 finish_wait(&pgdat->kswapd_wait, &wait); 2057 2058 if (!try_to_freeze()) { 2059 /* We can speed up thawing tasks if we don't call 2060 * balance_pgdat after returning from the refrigerator 2061 */ 2062 balance_pgdat(pgdat, order); 2063 } 2064 } 2065 return 0; 2066} 2067 2068/* 2069 * A zone is low on free memory, so wake its kswapd task to service it. 2070 */ 2071void wakeup_kswapd(struct zone *zone, int order) 2072{ 2073 pg_data_t *pgdat; 2074 2075 if (!populated_zone(zone)) 2076 return; 2077 2078 pgdat = zone->zone_pgdat; 2079 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0)) 2080 return; 2081 if (pgdat->kswapd_max_order < order) 2082 pgdat->kswapd_max_order = order; 2083 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 2084 return; 2085 if (!waitqueue_active(&pgdat->kswapd_wait)) 2086 return; 2087 wake_up_interruptible(&pgdat->kswapd_wait); 2088} 2089 2090unsigned long global_lru_pages(void) 2091{ 2092 return global_page_state(NR_ACTIVE_ANON) 2093 + global_page_state(NR_ACTIVE_FILE) 2094 + global_page_state(NR_INACTIVE_ANON) 2095 + global_page_state(NR_INACTIVE_FILE); 2096} 2097 2098#ifdef CONFIG_HIBERNATION 2099/* 2100 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages 2101 * from LRU lists system-wide, for given pass and priority. 2102 * 2103 * For pass > 3 we also try to shrink the LRU lists that contain a few pages 2104 */ 2105static void shrink_all_zones(unsigned long nr_pages, int prio, 2106 int pass, struct scan_control *sc) 2107{ 2108 struct zone *zone; 2109 unsigned long nr_reclaimed = 0; 2110 2111 for_each_populated_zone(zone) { 2112 enum lru_list l; 2113 2114 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY) 2115 continue; 2116 2117 for_each_evictable_lru(l) { 2118 enum zone_stat_item ls = NR_LRU_BASE + l; 2119 unsigned long lru_pages = zone_page_state(zone, ls); 2120 2121 /* For pass = 0, we don't shrink the active list */ 2122 if (pass == 0 && (l == LRU_ACTIVE_ANON || 2123 l == LRU_ACTIVE_FILE)) 2124 continue; 2125 2126 zone->lru[l].nr_saved_scan += (lru_pages >> prio) + 1; 2127 if (zone->lru[l].nr_saved_scan >= nr_pages || pass > 3) { 2128 unsigned long nr_to_scan; 2129 2130 zone->lru[l].nr_saved_scan = 0; 2131 nr_to_scan = min(nr_pages, lru_pages); 2132 nr_reclaimed += shrink_list(l, nr_to_scan, zone, 2133 sc, prio); 2134 if (nr_reclaimed >= nr_pages) { 2135 sc->nr_reclaimed += nr_reclaimed; 2136 return; 2137 } 2138 } 2139 } 2140 } 2141 sc->nr_reclaimed += nr_reclaimed; 2142} 2143 2144/* 2145 * Try to free `nr_pages' of memory, system-wide, and return the number of 2146 * freed pages. 2147 * 2148 * Rather than trying to age LRUs the aim is to preserve the overall 2149 * LRU order by reclaiming preferentially 2150 * inactive > active > active referenced > active mapped 2151 */ 2152unsigned long shrink_all_memory(unsigned long nr_pages) 2153{ 2154 unsigned long lru_pages, nr_slab; 2155 int pass; 2156 struct reclaim_state reclaim_state; 2157 struct scan_control sc = { 2158 .gfp_mask = GFP_KERNEL, 2159 .may_unmap = 0, 2160 .may_writepage = 1, 2161 .isolate_pages = isolate_pages_global, 2162 .nr_reclaimed = 0, 2163 }; 2164 2165 current->reclaim_state = &reclaim_state; 2166 2167 lru_pages = global_lru_pages(); 2168 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE); 2169 /* If slab caches are huge, it's better to hit them first */ 2170 while (nr_slab >= lru_pages) { 2171 reclaim_state.reclaimed_slab = 0; 2172 shrink_slab(nr_pages, sc.gfp_mask, lru_pages); 2173 if (!reclaim_state.reclaimed_slab) 2174 break; 2175 2176 sc.nr_reclaimed += reclaim_state.reclaimed_slab; 2177 if (sc.nr_reclaimed >= nr_pages) 2178 goto out; 2179 2180 nr_slab -= reclaim_state.reclaimed_slab; 2181 } 2182 2183 /* 2184 * We try to shrink LRUs in 5 passes: 2185 * 0 = Reclaim from inactive_list only 2186 * 1 = Reclaim from active list but don't reclaim mapped 2187 * 2 = 2nd pass of type 1 2188 * 3 = Reclaim mapped (normal reclaim) 2189 * 4 = 2nd pass of type 3 2190 */ 2191 for (pass = 0; pass < 5; pass++) { 2192 int prio; 2193 2194 /* Force reclaiming mapped pages in the passes #3 and #4 */ 2195 if (pass > 2) 2196 sc.may_unmap = 1; 2197 2198 for (prio = DEF_PRIORITY; prio >= 0; prio--) { 2199 unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed; 2200 2201 sc.nr_scanned = 0; 2202 sc.swap_cluster_max = nr_to_scan; 2203 shrink_all_zones(nr_to_scan, prio, pass, &sc); 2204 if (sc.nr_reclaimed >= nr_pages) 2205 goto out; 2206 2207 reclaim_state.reclaimed_slab = 0; 2208 shrink_slab(sc.nr_scanned, sc.gfp_mask, 2209 global_lru_pages()); 2210 sc.nr_reclaimed += reclaim_state.reclaimed_slab; 2211 if (sc.nr_reclaimed >= nr_pages) 2212 goto out; 2213 2214 if (sc.nr_scanned && prio < DEF_PRIORITY - 2) 2215 congestion_wait(WRITE, HZ / 10); 2216 } 2217 } 2218 2219 /* 2220 * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be 2221 * something in slab caches 2222 */ 2223 if (!sc.nr_reclaimed) { 2224 do { 2225 reclaim_state.reclaimed_slab = 0; 2226 shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages()); 2227 sc.nr_reclaimed += reclaim_state.reclaimed_slab; 2228 } while (sc.nr_reclaimed < nr_pages && 2229 reclaim_state.reclaimed_slab > 0); 2230 } 2231 2232 2233out: 2234 current->reclaim_state = NULL; 2235 2236 return sc.nr_reclaimed; 2237} 2238#endif /* CONFIG_HIBERNATION */ 2239 2240/* It's optimal to keep kswapds on the same CPUs as their memory, but 2241 not required for correctness. So if the last cpu in a node goes 2242 away, we get changed to run anywhere: as the first one comes back, 2243 restore their cpu bindings. */ 2244static int __devinit cpu_callback(struct notifier_block *nfb, 2245 unsigned long action, void *hcpu) 2246{ 2247 int nid; 2248 2249 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { 2250 for_each_node_state(nid, N_HIGH_MEMORY) { 2251 pg_data_t *pgdat = NODE_DATA(nid); 2252 const struct cpumask *mask; 2253 2254 mask = cpumask_of_node(pgdat->node_id); 2255 2256 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) 2257 /* One of our CPUs online: restore mask */ 2258 set_cpus_allowed_ptr(pgdat->kswapd, mask); 2259 } 2260 } 2261 return NOTIFY_OK; 2262} 2263 2264/* 2265 * This kswapd start function will be called by init and node-hot-add. 2266 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. 2267 */ 2268int kswapd_run(int nid) 2269{ 2270 pg_data_t *pgdat = NODE_DATA(nid); 2271 int ret = 0; 2272 2273 if (pgdat->kswapd) 2274 return 0; 2275 2276 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); 2277 if (IS_ERR(pgdat->kswapd)) { 2278 /* failure at boot is fatal */ 2279 BUG_ON(system_state == SYSTEM_BOOTING); 2280 printk("Failed to start kswapd on node %d\n",nid); 2281 ret = -1; 2282 } 2283 return ret; 2284} 2285 2286static int __init kswapd_init(void) 2287{ 2288 int nid; 2289 2290 swap_setup(); 2291 for_each_node_state(nid, N_HIGH_MEMORY) 2292 kswapd_run(nid); 2293 hotcpu_notifier(cpu_callback, 0); 2294 return 0; 2295} 2296 2297module_init(kswapd_init) 2298 2299#ifdef CONFIG_NUMA 2300/* 2301 * Zone reclaim mode 2302 * 2303 * If non-zero call zone_reclaim when the number of free pages falls below 2304 * the watermarks. 2305 */ 2306int zone_reclaim_mode __read_mostly; 2307 2308#define RECLAIM_OFF 0 2309#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ 2310#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ 2311#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ 2312 2313/* 2314 * Priority for ZONE_RECLAIM. This determines the fraction of pages 2315 * of a node considered for each zone_reclaim. 4 scans 1/16th of 2316 * a zone. 2317 */ 2318#define ZONE_RECLAIM_PRIORITY 4 2319 2320/* 2321 * Percentage of pages in a zone that must be unmapped for zone_reclaim to 2322 * occur. 2323 */ 2324int sysctl_min_unmapped_ratio = 1; 2325 2326/* 2327 * If the number of slab pages in a zone grows beyond this percentage then 2328 * slab reclaim needs to occur. 2329 */ 2330int sysctl_min_slab_ratio = 5; 2331 2332/* 2333 * Try to free up some pages from this zone through reclaim. 2334 */ 2335static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 2336{ 2337 /* Minimum pages needed in order to stay on node */ 2338 const unsigned long nr_pages = 1 << order; 2339 struct task_struct *p = current; 2340 struct reclaim_state reclaim_state; 2341 int priority; 2342 struct scan_control sc = { 2343 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), 2344 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), 2345 .may_swap = 1, 2346 .swap_cluster_max = max_t(unsigned long, nr_pages, 2347 SWAP_CLUSTER_MAX), 2348 .gfp_mask = gfp_mask, 2349 .swappiness = vm_swappiness, 2350 .order = order, 2351 .isolate_pages = isolate_pages_global, 2352 }; 2353 unsigned long slab_reclaimable; 2354 2355 disable_swap_token(); 2356 cond_resched(); 2357 /* 2358 * We need to be able to allocate from the reserves for RECLAIM_SWAP 2359 * and we also need to be able to write out pages for RECLAIM_WRITE 2360 * and RECLAIM_SWAP. 2361 */ 2362 p->flags |= PF_MEMALLOC | PF_SWAPWRITE; 2363 reclaim_state.reclaimed_slab = 0; 2364 p->reclaim_state = &reclaim_state; 2365 2366 if (zone_page_state(zone, NR_FILE_PAGES) - 2367 zone_page_state(zone, NR_FILE_MAPPED) > 2368 zone->min_unmapped_pages) { 2369 /* 2370 * Free memory by calling shrink zone with increasing 2371 * priorities until we have enough memory freed. 2372 */ 2373 priority = ZONE_RECLAIM_PRIORITY; 2374 do { 2375 note_zone_scanning_priority(zone, priority); 2376 shrink_zone(priority, zone, &sc); 2377 priority--; 2378 } while (priority >= 0 && sc.nr_reclaimed < nr_pages); 2379 } 2380 2381 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE); 2382 if (slab_reclaimable > zone->min_slab_pages) { 2383 /* 2384 * shrink_slab() does not currently allow us to determine how 2385 * many pages were freed in this zone. So we take the current 2386 * number of slab pages and shake the slab until it is reduced 2387 * by the same nr_pages that we used for reclaiming unmapped 2388 * pages. 2389 * 2390 * Note that shrink_slab will free memory on all zones and may 2391 * take a long time. 2392 */ 2393 while (shrink_slab(sc.nr_scanned, gfp_mask, order) && 2394 zone_page_state(zone, NR_SLAB_RECLAIMABLE) > 2395 slab_reclaimable - nr_pages) 2396 ; 2397 2398 /* 2399 * Update nr_reclaimed by the number of slab pages we 2400 * reclaimed from this zone. 2401 */ 2402 sc.nr_reclaimed += slab_reclaimable - 2403 zone_page_state(zone, NR_SLAB_RECLAIMABLE); 2404 } 2405 2406 p->reclaim_state = NULL; 2407 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); 2408 return sc.nr_reclaimed >= nr_pages; 2409} 2410 2411int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 2412{ 2413 int node_id; 2414 int ret; 2415 2416 /* 2417 * Zone reclaim reclaims unmapped file backed pages and 2418 * slab pages if we are over the defined limits. 2419 * 2420 * A small portion of unmapped file backed pages is needed for 2421 * file I/O otherwise pages read by file I/O will be immediately 2422 * thrown out if the zone is overallocated. So we do not reclaim 2423 * if less than a specified percentage of the zone is used by 2424 * unmapped file backed pages. 2425 */ 2426 if (zone_page_state(zone, NR_FILE_PAGES) - 2427 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages 2428 && zone_page_state(zone, NR_SLAB_RECLAIMABLE) 2429 <= zone->min_slab_pages) 2430 return 0; 2431 2432 if (zone_is_all_unreclaimable(zone)) 2433 return 0; 2434 2435 /* 2436 * Do not scan if the allocation should not be delayed. 2437 */ 2438 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) 2439 return 0; 2440 2441 /* 2442 * Only run zone reclaim on the local zone or on zones that do not 2443 * have associated processors. This will favor the local processor 2444 * over remote processors and spread off node memory allocations 2445 * as wide as possible. 2446 */ 2447 node_id = zone_to_nid(zone); 2448 if (node_state(node_id, N_CPU) && node_id != numa_node_id()) 2449 return 0; 2450 2451 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) 2452 return 0; 2453 ret = __zone_reclaim(zone, gfp_mask, order); 2454 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); 2455 2456 return ret; 2457} 2458#endif 2459 2460/* 2461 * page_evictable - test whether a page is evictable 2462 * @page: the page to test 2463 * @vma: the VMA in which the page is or will be mapped, may be NULL 2464 * 2465 * Test whether page is evictable--i.e., should be placed on active/inactive 2466 * lists vs unevictable list. The vma argument is !NULL when called from the 2467 * fault path to determine how to instantate a new page. 2468 * 2469 * Reasons page might not be evictable: 2470 * (1) page's mapping marked unevictable 2471 * (2) page is part of an mlocked VMA 2472 * 2473 */ 2474int page_evictable(struct page *page, struct vm_area_struct *vma) 2475{ 2476 2477 if (mapping_unevictable(page_mapping(page))) 2478 return 0; 2479 2480 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page))) 2481 return 0; 2482 2483 return 1; 2484} 2485 2486/** 2487 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list 2488 * @page: page to check evictability and move to appropriate lru list 2489 * @zone: zone page is in 2490 * 2491 * Checks a page for evictability and moves the page to the appropriate 2492 * zone lru list. 2493 * 2494 * Restrictions: zone->lru_lock must be held, page must be on LRU and must 2495 * have PageUnevictable set. 2496 */ 2497static void check_move_unevictable_page(struct page *page, struct zone *zone) 2498{ 2499 VM_BUG_ON(PageActive(page)); 2500 2501retry: 2502 ClearPageUnevictable(page); 2503 if (page_evictable(page, NULL)) { 2504 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page); 2505 2506 __dec_zone_state(zone, NR_UNEVICTABLE); 2507 list_move(&page->lru, &zone->lru[l].list); 2508 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l); 2509 __inc_zone_state(zone, NR_INACTIVE_ANON + l); 2510 __count_vm_event(UNEVICTABLE_PGRESCUED); 2511 } else { 2512 /* 2513 * rotate unevictable list 2514 */ 2515 SetPageUnevictable(page); 2516 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list); 2517 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE); 2518 if (page_evictable(page, NULL)) 2519 goto retry; 2520 } 2521} 2522 2523/** 2524 * scan_mapping_unevictable_pages - scan an address space for evictable pages 2525 * @mapping: struct address_space to scan for evictable pages 2526 * 2527 * Scan all pages in mapping. Check unevictable pages for 2528 * evictability and move them to the appropriate zone lru list. 2529 */ 2530void scan_mapping_unevictable_pages(struct address_space *mapping) 2531{ 2532 pgoff_t next = 0; 2533 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >> 2534 PAGE_CACHE_SHIFT; 2535 struct zone *zone; 2536 struct pagevec pvec; 2537 2538 if (mapping->nrpages == 0) 2539 return; 2540 2541 pagevec_init(&pvec, 0); 2542 while (next < end && 2543 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) { 2544 int i; 2545 int pg_scanned = 0; 2546 2547 zone = NULL; 2548 2549 for (i = 0; i < pagevec_count(&pvec); i++) { 2550 struct page *page = pvec.pages[i]; 2551 pgoff_t page_index = page->index; 2552 struct zone *pagezone = page_zone(page); 2553 2554 pg_scanned++; 2555 if (page_index > next) 2556 next = page_index; 2557 next++; 2558 2559 if (pagezone != zone) { 2560 if (zone) 2561 spin_unlock_irq(&zone->lru_lock); 2562 zone = pagezone; 2563 spin_lock_irq(&zone->lru_lock); 2564 } 2565 2566 if (PageLRU(page) && PageUnevictable(page)) 2567 check_move_unevictable_page(page, zone); 2568 } 2569 if (zone) 2570 spin_unlock_irq(&zone->lru_lock); 2571 pagevec_release(&pvec); 2572 2573 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned); 2574 } 2575 2576} 2577 2578/** 2579 * scan_zone_unevictable_pages - check unevictable list for evictable pages 2580 * @zone - zone of which to scan the unevictable list 2581 * 2582 * Scan @zone's unevictable LRU lists to check for pages that have become 2583 * evictable. Move those that have to @zone's inactive list where they 2584 * become candidates for reclaim, unless shrink_inactive_zone() decides 2585 * to reactivate them. Pages that are still unevictable are rotated 2586 * back onto @zone's unevictable list. 2587 */ 2588#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */ 2589static void scan_zone_unevictable_pages(struct zone *zone) 2590{ 2591 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list; 2592 unsigned long scan; 2593 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE); 2594 2595 while (nr_to_scan > 0) { 2596 unsigned long batch_size = min(nr_to_scan, 2597 SCAN_UNEVICTABLE_BATCH_SIZE); 2598 2599 spin_lock_irq(&zone->lru_lock); 2600 for (scan = 0; scan < batch_size; scan++) { 2601 struct page *page = lru_to_page(l_unevictable); 2602 2603 if (!trylock_page(page)) 2604 continue; 2605 2606 prefetchw_prev_lru_page(page, l_unevictable, flags); 2607 2608 if (likely(PageLRU(page) && PageUnevictable(page))) 2609 check_move_unevictable_page(page, zone); 2610 2611 unlock_page(page); 2612 } 2613 spin_unlock_irq(&zone->lru_lock); 2614 2615 nr_to_scan -= batch_size; 2616 } 2617} 2618 2619 2620/** 2621 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages 2622 * 2623 * A really big hammer: scan all zones' unevictable LRU lists to check for 2624 * pages that have become evictable. Move those back to the zones' 2625 * inactive list where they become candidates for reclaim. 2626 * This occurs when, e.g., we have unswappable pages on the unevictable lists, 2627 * and we add swap to the system. As such, it runs in the context of a task 2628 * that has possibly/probably made some previously unevictable pages 2629 * evictable. 2630 */ 2631static void scan_all_zones_unevictable_pages(void) 2632{ 2633 struct zone *zone; 2634 2635 for_each_zone(zone) { 2636 scan_zone_unevictable_pages(zone); 2637 } 2638} 2639 2640/* 2641 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of 2642 * all nodes' unevictable lists for evictable pages 2643 */ 2644unsigned long scan_unevictable_pages; 2645 2646int scan_unevictable_handler(struct ctl_table *table, int write, 2647 struct file *file, void __user *buffer, 2648 size_t *length, loff_t *ppos) 2649{ 2650 proc_doulongvec_minmax(table, write, file, buffer, length, ppos); 2651 2652 if (write && *(unsigned long *)table->data) 2653 scan_all_zones_unevictable_pages(); 2654 2655 scan_unevictable_pages = 0; 2656 return 0; 2657} 2658 2659/* 2660 * per node 'scan_unevictable_pages' attribute. On demand re-scan of 2661 * a specified node's per zone unevictable lists for evictable pages. 2662 */ 2663 2664static ssize_t read_scan_unevictable_node(struct sys_device *dev, 2665 struct sysdev_attribute *attr, 2666 char *buf) 2667{ 2668 return sprintf(buf, "0\n"); /* always zero; should fit... */ 2669} 2670 2671static ssize_t write_scan_unevictable_node(struct sys_device *dev, 2672 struct sysdev_attribute *attr, 2673 const char *buf, size_t count) 2674{ 2675 struct zone *node_zones = NODE_DATA(dev->id)->node_zones; 2676 struct zone *zone; 2677 unsigned long res; 2678 unsigned long req = strict_strtoul(buf, 10, &res); 2679 2680 if (!req) 2681 return 1; /* zero is no-op */ 2682 2683 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) { 2684 if (!populated_zone(zone)) 2685 continue; 2686 scan_zone_unevictable_pages(zone); 2687 } 2688 return 1; 2689} 2690 2691 2692static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, 2693 read_scan_unevictable_node, 2694 write_scan_unevictable_node); 2695 2696int scan_unevictable_register_node(struct node *node) 2697{ 2698 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages); 2699} 2700 2701void scan_unevictable_unregister_node(struct node *node) 2702{ 2703 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages); 2704} 2705 2706