vmscan.c revision 4356f21d09283dc6d39a6f7287a65ddab61e2808
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/gfp.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/compaction.h> 36#include <linux/notifier.h> 37#include <linux/rwsem.h> 38#include <linux/delay.h> 39#include <linux/kthread.h> 40#include <linux/freezer.h> 41#include <linux/memcontrol.h> 42#include <linux/delayacct.h> 43#include <linux/sysctl.h> 44#include <linux/oom.h> 45#include <linux/prefetch.h> 46 47#include <asm/tlbflush.h> 48#include <asm/div64.h> 49 50#include <linux/swapops.h> 51 52#include "internal.h" 53 54#define CREATE_TRACE_POINTS 55#include <trace/events/vmscan.h> 56 57/* 58 * reclaim_mode determines how the inactive list is shrunk 59 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages 60 * RECLAIM_MODE_ASYNC: Do not block 61 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback 62 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference 63 * page from the LRU and reclaim all pages within a 64 * naturally aligned range 65 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of 66 * order-0 pages and then compact the zone 67 */ 68typedef unsigned __bitwise__ reclaim_mode_t; 69#define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u) 70#define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u) 71#define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u) 72#define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u) 73#define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u) 74 75struct scan_control { 76 /* Incremented by the number of inactive pages that were scanned */ 77 unsigned long nr_scanned; 78 79 /* Number of pages freed so far during a call to shrink_zones() */ 80 unsigned long nr_reclaimed; 81 82 /* How many pages shrink_list() should reclaim */ 83 unsigned long nr_to_reclaim; 84 85 unsigned long hibernation_mode; 86 87 /* This context's GFP mask */ 88 gfp_t gfp_mask; 89 90 int may_writepage; 91 92 /* Can mapped pages be reclaimed? */ 93 int may_unmap; 94 95 /* Can pages be swapped as part of reclaim? */ 96 int may_swap; 97 98 int order; 99 100 /* 101 * Intend to reclaim enough continuous memory rather than reclaim 102 * enough amount of memory. i.e, mode for high order allocation. 103 */ 104 reclaim_mode_t reclaim_mode; 105 106 /* Which cgroup do we reclaim from */ 107 struct mem_cgroup *mem_cgroup; 108 109 /* 110 * Nodemask of nodes allowed by the caller. If NULL, all nodes 111 * are scanned. 112 */ 113 nodemask_t *nodemask; 114}; 115 116#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) 117 118#ifdef ARCH_HAS_PREFETCH 119#define prefetch_prev_lru_page(_page, _base, _field) \ 120 do { \ 121 if ((_page)->lru.prev != _base) { \ 122 struct page *prev; \ 123 \ 124 prev = lru_to_page(&(_page->lru)); \ 125 prefetch(&prev->_field); \ 126 } \ 127 } while (0) 128#else 129#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) 130#endif 131 132#ifdef ARCH_HAS_PREFETCHW 133#define prefetchw_prev_lru_page(_page, _base, _field) \ 134 do { \ 135 if ((_page)->lru.prev != _base) { \ 136 struct page *prev; \ 137 \ 138 prev = lru_to_page(&(_page->lru)); \ 139 prefetchw(&prev->_field); \ 140 } \ 141 } while (0) 142#else 143#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) 144#endif 145 146/* 147 * From 0 .. 100. Higher means more swappy. 148 */ 149int vm_swappiness = 60; 150long vm_total_pages; /* The total number of pages which the VM controls */ 151 152static LIST_HEAD(shrinker_list); 153static DECLARE_RWSEM(shrinker_rwsem); 154 155#ifdef CONFIG_CGROUP_MEM_RES_CTLR 156#define scanning_global_lru(sc) (!(sc)->mem_cgroup) 157#else 158#define scanning_global_lru(sc) (1) 159#endif 160 161static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone, 162 struct scan_control *sc) 163{ 164 if (!scanning_global_lru(sc)) 165 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone); 166 167 return &zone->reclaim_stat; 168} 169 170static unsigned long zone_nr_lru_pages(struct zone *zone, 171 struct scan_control *sc, enum lru_list lru) 172{ 173 if (!scanning_global_lru(sc)) 174 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup, 175 zone_to_nid(zone), zone_idx(zone), BIT(lru)); 176 177 return zone_page_state(zone, NR_LRU_BASE + lru); 178} 179 180 181/* 182 * Add a shrinker callback to be called from the vm 183 */ 184void register_shrinker(struct shrinker *shrinker) 185{ 186 shrinker->nr = 0; 187 down_write(&shrinker_rwsem); 188 list_add_tail(&shrinker->list, &shrinker_list); 189 up_write(&shrinker_rwsem); 190} 191EXPORT_SYMBOL(register_shrinker); 192 193/* 194 * Remove one 195 */ 196void unregister_shrinker(struct shrinker *shrinker) 197{ 198 down_write(&shrinker_rwsem); 199 list_del(&shrinker->list); 200 up_write(&shrinker_rwsem); 201} 202EXPORT_SYMBOL(unregister_shrinker); 203 204static inline int do_shrinker_shrink(struct shrinker *shrinker, 205 struct shrink_control *sc, 206 unsigned long nr_to_scan) 207{ 208 sc->nr_to_scan = nr_to_scan; 209 return (*shrinker->shrink)(shrinker, sc); 210} 211 212#define SHRINK_BATCH 128 213/* 214 * Call the shrink functions to age shrinkable caches 215 * 216 * Here we assume it costs one seek to replace a lru page and that it also 217 * takes a seek to recreate a cache object. With this in mind we age equal 218 * percentages of the lru and ageable caches. This should balance the seeks 219 * generated by these structures. 220 * 221 * If the vm encountered mapped pages on the LRU it increase the pressure on 222 * slab to avoid swapping. 223 * 224 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. 225 * 226 * `lru_pages' represents the number of on-LRU pages in all the zones which 227 * are eligible for the caller's allocation attempt. It is used for balancing 228 * slab reclaim versus page reclaim. 229 * 230 * Returns the number of slab objects which we shrunk. 231 */ 232unsigned long shrink_slab(struct shrink_control *shrink, 233 unsigned long nr_pages_scanned, 234 unsigned long lru_pages) 235{ 236 struct shrinker *shrinker; 237 unsigned long ret = 0; 238 239 if (nr_pages_scanned == 0) 240 nr_pages_scanned = SWAP_CLUSTER_MAX; 241 242 if (!down_read_trylock(&shrinker_rwsem)) { 243 /* Assume we'll be able to shrink next time */ 244 ret = 1; 245 goto out; 246 } 247 248 list_for_each_entry(shrinker, &shrinker_list, list) { 249 unsigned long long delta; 250 unsigned long total_scan; 251 unsigned long max_pass; 252 int shrink_ret = 0; 253 long nr; 254 long new_nr; 255 long batch_size = shrinker->batch ? shrinker->batch 256 : SHRINK_BATCH; 257 258 /* 259 * copy the current shrinker scan count into a local variable 260 * and zero it so that other concurrent shrinker invocations 261 * don't also do this scanning work. 262 */ 263 do { 264 nr = shrinker->nr; 265 } while (cmpxchg(&shrinker->nr, nr, 0) != nr); 266 267 total_scan = nr; 268 max_pass = do_shrinker_shrink(shrinker, shrink, 0); 269 delta = (4 * nr_pages_scanned) / shrinker->seeks; 270 delta *= max_pass; 271 do_div(delta, lru_pages + 1); 272 total_scan += delta; 273 if (total_scan < 0) { 274 printk(KERN_ERR "shrink_slab: %pF negative objects to " 275 "delete nr=%ld\n", 276 shrinker->shrink, total_scan); 277 total_scan = max_pass; 278 } 279 280 /* 281 * We need to avoid excessive windup on filesystem shrinkers 282 * due to large numbers of GFP_NOFS allocations causing the 283 * shrinkers to return -1 all the time. This results in a large 284 * nr being built up so when a shrink that can do some work 285 * comes along it empties the entire cache due to nr >>> 286 * max_pass. This is bad for sustaining a working set in 287 * memory. 288 * 289 * Hence only allow the shrinker to scan the entire cache when 290 * a large delta change is calculated directly. 291 */ 292 if (delta < max_pass / 4) 293 total_scan = min(total_scan, max_pass / 2); 294 295 /* 296 * Avoid risking looping forever due to too large nr value: 297 * never try to free more than twice the estimate number of 298 * freeable entries. 299 */ 300 if (total_scan > max_pass * 2) 301 total_scan = max_pass * 2; 302 303 trace_mm_shrink_slab_start(shrinker, shrink, nr, 304 nr_pages_scanned, lru_pages, 305 max_pass, delta, total_scan); 306 307 while (total_scan >= batch_size) { 308 int nr_before; 309 310 nr_before = do_shrinker_shrink(shrinker, shrink, 0); 311 shrink_ret = do_shrinker_shrink(shrinker, shrink, 312 batch_size); 313 if (shrink_ret == -1) 314 break; 315 if (shrink_ret < nr_before) 316 ret += nr_before - shrink_ret; 317 count_vm_events(SLABS_SCANNED, batch_size); 318 total_scan -= batch_size; 319 320 cond_resched(); 321 } 322 323 /* 324 * move the unused scan count back into the shrinker in a 325 * manner that handles concurrent updates. If we exhausted the 326 * scan, there is no need to do an update. 327 */ 328 do { 329 nr = shrinker->nr; 330 new_nr = total_scan + nr; 331 if (total_scan <= 0) 332 break; 333 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr); 334 335 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr); 336 } 337 up_read(&shrinker_rwsem); 338out: 339 cond_resched(); 340 return ret; 341} 342 343static void set_reclaim_mode(int priority, struct scan_control *sc, 344 bool sync) 345{ 346 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC; 347 348 /* 349 * Initially assume we are entering either lumpy reclaim or 350 * reclaim/compaction.Depending on the order, we will either set the 351 * sync mode or just reclaim order-0 pages later. 352 */ 353 if (COMPACTION_BUILD) 354 sc->reclaim_mode = RECLAIM_MODE_COMPACTION; 355 else 356 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM; 357 358 /* 359 * Avoid using lumpy reclaim or reclaim/compaction if possible by 360 * restricting when its set to either costly allocations or when 361 * under memory pressure 362 */ 363 if (sc->order > PAGE_ALLOC_COSTLY_ORDER) 364 sc->reclaim_mode |= syncmode; 365 else if (sc->order && priority < DEF_PRIORITY - 2) 366 sc->reclaim_mode |= syncmode; 367 else 368 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC; 369} 370 371static void reset_reclaim_mode(struct scan_control *sc) 372{ 373 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC; 374} 375 376static inline int is_page_cache_freeable(struct page *page) 377{ 378 /* 379 * A freeable page cache page is referenced only by the caller 380 * that isolated the page, the page cache radix tree and 381 * optional buffer heads at page->private. 382 */ 383 return page_count(page) - page_has_private(page) == 2; 384} 385 386static int may_write_to_queue(struct backing_dev_info *bdi, 387 struct scan_control *sc) 388{ 389 if (current->flags & PF_SWAPWRITE) 390 return 1; 391 if (!bdi_write_congested(bdi)) 392 return 1; 393 if (bdi == current->backing_dev_info) 394 return 1; 395 396 /* lumpy reclaim for hugepage often need a lot of write */ 397 if (sc->order > PAGE_ALLOC_COSTLY_ORDER) 398 return 1; 399 return 0; 400} 401 402/* 403 * We detected a synchronous write error writing a page out. Probably 404 * -ENOSPC. We need to propagate that into the address_space for a subsequent 405 * fsync(), msync() or close(). 406 * 407 * The tricky part is that after writepage we cannot touch the mapping: nothing 408 * prevents it from being freed up. But we have a ref on the page and once 409 * that page is locked, the mapping is pinned. 410 * 411 * We're allowed to run sleeping lock_page() here because we know the caller has 412 * __GFP_FS. 413 */ 414static void handle_write_error(struct address_space *mapping, 415 struct page *page, int error) 416{ 417 lock_page(page); 418 if (page_mapping(page) == mapping) 419 mapping_set_error(mapping, error); 420 unlock_page(page); 421} 422 423/* possible outcome of pageout() */ 424typedef enum { 425 /* failed to write page out, page is locked */ 426 PAGE_KEEP, 427 /* move page to the active list, page is locked */ 428 PAGE_ACTIVATE, 429 /* page has been sent to the disk successfully, page is unlocked */ 430 PAGE_SUCCESS, 431 /* page is clean and locked */ 432 PAGE_CLEAN, 433} pageout_t; 434 435/* 436 * pageout is called by shrink_page_list() for each dirty page. 437 * Calls ->writepage(). 438 */ 439static pageout_t pageout(struct page *page, struct address_space *mapping, 440 struct scan_control *sc) 441{ 442 /* 443 * If the page is dirty, only perform writeback if that write 444 * will be non-blocking. To prevent this allocation from being 445 * stalled by pagecache activity. But note that there may be 446 * stalls if we need to run get_block(). We could test 447 * PagePrivate for that. 448 * 449 * If this process is currently in __generic_file_aio_write() against 450 * this page's queue, we can perform writeback even if that 451 * will block. 452 * 453 * If the page is swapcache, write it back even if that would 454 * block, for some throttling. This happens by accident, because 455 * swap_backing_dev_info is bust: it doesn't reflect the 456 * congestion state of the swapdevs. Easy to fix, if needed. 457 */ 458 if (!is_page_cache_freeable(page)) 459 return PAGE_KEEP; 460 if (!mapping) { 461 /* 462 * Some data journaling orphaned pages can have 463 * page->mapping == NULL while being dirty with clean buffers. 464 */ 465 if (page_has_private(page)) { 466 if (try_to_free_buffers(page)) { 467 ClearPageDirty(page); 468 printk("%s: orphaned page\n", __func__); 469 return PAGE_CLEAN; 470 } 471 } 472 return PAGE_KEEP; 473 } 474 if (mapping->a_ops->writepage == NULL) 475 return PAGE_ACTIVATE; 476 if (!may_write_to_queue(mapping->backing_dev_info, sc)) 477 return PAGE_KEEP; 478 479 if (clear_page_dirty_for_io(page)) { 480 int res; 481 struct writeback_control wbc = { 482 .sync_mode = WB_SYNC_NONE, 483 .nr_to_write = SWAP_CLUSTER_MAX, 484 .range_start = 0, 485 .range_end = LLONG_MAX, 486 .for_reclaim = 1, 487 }; 488 489 SetPageReclaim(page); 490 res = mapping->a_ops->writepage(page, &wbc); 491 if (res < 0) 492 handle_write_error(mapping, page, res); 493 if (res == AOP_WRITEPAGE_ACTIVATE) { 494 ClearPageReclaim(page); 495 return PAGE_ACTIVATE; 496 } 497 498 /* 499 * Wait on writeback if requested to. This happens when 500 * direct reclaiming a large contiguous area and the 501 * first attempt to free a range of pages fails. 502 */ 503 if (PageWriteback(page) && 504 (sc->reclaim_mode & RECLAIM_MODE_SYNC)) 505 wait_on_page_writeback(page); 506 507 if (!PageWriteback(page)) { 508 /* synchronous write or broken a_ops? */ 509 ClearPageReclaim(page); 510 } 511 trace_mm_vmscan_writepage(page, 512 trace_reclaim_flags(page, sc->reclaim_mode)); 513 inc_zone_page_state(page, NR_VMSCAN_WRITE); 514 return PAGE_SUCCESS; 515 } 516 517 return PAGE_CLEAN; 518} 519 520/* 521 * Same as remove_mapping, but if the page is removed from the mapping, it 522 * gets returned with a refcount of 0. 523 */ 524static int __remove_mapping(struct address_space *mapping, struct page *page) 525{ 526 BUG_ON(!PageLocked(page)); 527 BUG_ON(mapping != page_mapping(page)); 528 529 spin_lock_irq(&mapping->tree_lock); 530 /* 531 * The non racy check for a busy page. 532 * 533 * Must be careful with the order of the tests. When someone has 534 * a ref to the page, it may be possible that they dirty it then 535 * drop the reference. So if PageDirty is tested before page_count 536 * here, then the following race may occur: 537 * 538 * get_user_pages(&page); 539 * [user mapping goes away] 540 * write_to(page); 541 * !PageDirty(page) [good] 542 * SetPageDirty(page); 543 * put_page(page); 544 * !page_count(page) [good, discard it] 545 * 546 * [oops, our write_to data is lost] 547 * 548 * Reversing the order of the tests ensures such a situation cannot 549 * escape unnoticed. The smp_rmb is needed to ensure the page->flags 550 * load is not satisfied before that of page->_count. 551 * 552 * Note that if SetPageDirty is always performed via set_page_dirty, 553 * and thus under tree_lock, then this ordering is not required. 554 */ 555 if (!page_freeze_refs(page, 2)) 556 goto cannot_free; 557 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ 558 if (unlikely(PageDirty(page))) { 559 page_unfreeze_refs(page, 2); 560 goto cannot_free; 561 } 562 563 if (PageSwapCache(page)) { 564 swp_entry_t swap = { .val = page_private(page) }; 565 __delete_from_swap_cache(page); 566 spin_unlock_irq(&mapping->tree_lock); 567 swapcache_free(swap, page); 568 } else { 569 void (*freepage)(struct page *); 570 571 freepage = mapping->a_ops->freepage; 572 573 __delete_from_page_cache(page); 574 spin_unlock_irq(&mapping->tree_lock); 575 mem_cgroup_uncharge_cache_page(page); 576 577 if (freepage != NULL) 578 freepage(page); 579 } 580 581 return 1; 582 583cannot_free: 584 spin_unlock_irq(&mapping->tree_lock); 585 return 0; 586} 587 588/* 589 * Attempt to detach a locked page from its ->mapping. If it is dirty or if 590 * someone else has a ref on the page, abort and return 0. If it was 591 * successfully detached, return 1. Assumes the caller has a single ref on 592 * this page. 593 */ 594int remove_mapping(struct address_space *mapping, struct page *page) 595{ 596 if (__remove_mapping(mapping, page)) { 597 /* 598 * Unfreezing the refcount with 1 rather than 2 effectively 599 * drops the pagecache ref for us without requiring another 600 * atomic operation. 601 */ 602 page_unfreeze_refs(page, 1); 603 return 1; 604 } 605 return 0; 606} 607 608/** 609 * putback_lru_page - put previously isolated page onto appropriate LRU list 610 * @page: page to be put back to appropriate lru list 611 * 612 * Add previously isolated @page to appropriate LRU list. 613 * Page may still be unevictable for other reasons. 614 * 615 * lru_lock must not be held, interrupts must be enabled. 616 */ 617void putback_lru_page(struct page *page) 618{ 619 int lru; 620 int active = !!TestClearPageActive(page); 621 int was_unevictable = PageUnevictable(page); 622 623 VM_BUG_ON(PageLRU(page)); 624 625redo: 626 ClearPageUnevictable(page); 627 628 if (page_evictable(page, NULL)) { 629 /* 630 * For evictable pages, we can use the cache. 631 * In event of a race, worst case is we end up with an 632 * unevictable page on [in]active list. 633 * We know how to handle that. 634 */ 635 lru = active + page_lru_base_type(page); 636 lru_cache_add_lru(page, lru); 637 } else { 638 /* 639 * Put unevictable pages directly on zone's unevictable 640 * list. 641 */ 642 lru = LRU_UNEVICTABLE; 643 add_page_to_unevictable_list(page); 644 /* 645 * When racing with an mlock clearing (page is 646 * unlocked), make sure that if the other thread does 647 * not observe our setting of PG_lru and fails 648 * isolation, we see PG_mlocked cleared below and move 649 * the page back to the evictable list. 650 * 651 * The other side is TestClearPageMlocked(). 652 */ 653 smp_mb(); 654 } 655 656 /* 657 * page's status can change while we move it among lru. If an evictable 658 * page is on unevictable list, it never be freed. To avoid that, 659 * check after we added it to the list, again. 660 */ 661 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) { 662 if (!isolate_lru_page(page)) { 663 put_page(page); 664 goto redo; 665 } 666 /* This means someone else dropped this page from LRU 667 * So, it will be freed or putback to LRU again. There is 668 * nothing to do here. 669 */ 670 } 671 672 if (was_unevictable && lru != LRU_UNEVICTABLE) 673 count_vm_event(UNEVICTABLE_PGRESCUED); 674 else if (!was_unevictable && lru == LRU_UNEVICTABLE) 675 count_vm_event(UNEVICTABLE_PGCULLED); 676 677 put_page(page); /* drop ref from isolate */ 678} 679 680enum page_references { 681 PAGEREF_RECLAIM, 682 PAGEREF_RECLAIM_CLEAN, 683 PAGEREF_KEEP, 684 PAGEREF_ACTIVATE, 685}; 686 687static enum page_references page_check_references(struct page *page, 688 struct scan_control *sc) 689{ 690 int referenced_ptes, referenced_page; 691 unsigned long vm_flags; 692 693 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags); 694 referenced_page = TestClearPageReferenced(page); 695 696 /* Lumpy reclaim - ignore references */ 697 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM) 698 return PAGEREF_RECLAIM; 699 700 /* 701 * Mlock lost the isolation race with us. Let try_to_unmap() 702 * move the page to the unevictable list. 703 */ 704 if (vm_flags & VM_LOCKED) 705 return PAGEREF_RECLAIM; 706 707 if (referenced_ptes) { 708 if (PageAnon(page)) 709 return PAGEREF_ACTIVATE; 710 /* 711 * All mapped pages start out with page table 712 * references from the instantiating fault, so we need 713 * to look twice if a mapped file page is used more 714 * than once. 715 * 716 * Mark it and spare it for another trip around the 717 * inactive list. Another page table reference will 718 * lead to its activation. 719 * 720 * Note: the mark is set for activated pages as well 721 * so that recently deactivated but used pages are 722 * quickly recovered. 723 */ 724 SetPageReferenced(page); 725 726 if (referenced_page) 727 return PAGEREF_ACTIVATE; 728 729 return PAGEREF_KEEP; 730 } 731 732 /* Reclaim if clean, defer dirty pages to writeback */ 733 if (referenced_page && !PageSwapBacked(page)) 734 return PAGEREF_RECLAIM_CLEAN; 735 736 return PAGEREF_RECLAIM; 737} 738 739static noinline_for_stack void free_page_list(struct list_head *free_pages) 740{ 741 struct pagevec freed_pvec; 742 struct page *page, *tmp; 743 744 pagevec_init(&freed_pvec, 1); 745 746 list_for_each_entry_safe(page, tmp, free_pages, lru) { 747 list_del(&page->lru); 748 if (!pagevec_add(&freed_pvec, page)) { 749 __pagevec_free(&freed_pvec); 750 pagevec_reinit(&freed_pvec); 751 } 752 } 753 754 pagevec_free(&freed_pvec); 755} 756 757/* 758 * shrink_page_list() returns the number of reclaimed pages 759 */ 760static unsigned long shrink_page_list(struct list_head *page_list, 761 struct zone *zone, 762 struct scan_control *sc) 763{ 764 LIST_HEAD(ret_pages); 765 LIST_HEAD(free_pages); 766 int pgactivate = 0; 767 unsigned long nr_dirty = 0; 768 unsigned long nr_congested = 0; 769 unsigned long nr_reclaimed = 0; 770 771 cond_resched(); 772 773 while (!list_empty(page_list)) { 774 enum page_references references; 775 struct address_space *mapping; 776 struct page *page; 777 int may_enter_fs; 778 779 cond_resched(); 780 781 page = lru_to_page(page_list); 782 list_del(&page->lru); 783 784 if (!trylock_page(page)) 785 goto keep; 786 787 VM_BUG_ON(PageActive(page)); 788 VM_BUG_ON(page_zone(page) != zone); 789 790 sc->nr_scanned++; 791 792 if (unlikely(!page_evictable(page, NULL))) 793 goto cull_mlocked; 794 795 if (!sc->may_unmap && page_mapped(page)) 796 goto keep_locked; 797 798 /* Double the slab pressure for mapped and swapcache pages */ 799 if (page_mapped(page) || PageSwapCache(page)) 800 sc->nr_scanned++; 801 802 may_enter_fs = (sc->gfp_mask & __GFP_FS) || 803 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); 804 805 if (PageWriteback(page)) { 806 /* 807 * Synchronous reclaim is performed in two passes, 808 * first an asynchronous pass over the list to 809 * start parallel writeback, and a second synchronous 810 * pass to wait for the IO to complete. Wait here 811 * for any page for which writeback has already 812 * started. 813 */ 814 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) && 815 may_enter_fs) 816 wait_on_page_writeback(page); 817 else { 818 unlock_page(page); 819 goto keep_lumpy; 820 } 821 } 822 823 references = page_check_references(page, sc); 824 switch (references) { 825 case PAGEREF_ACTIVATE: 826 goto activate_locked; 827 case PAGEREF_KEEP: 828 goto keep_locked; 829 case PAGEREF_RECLAIM: 830 case PAGEREF_RECLAIM_CLEAN: 831 ; /* try to reclaim the page below */ 832 } 833 834 /* 835 * Anonymous process memory has backing store? 836 * Try to allocate it some swap space here. 837 */ 838 if (PageAnon(page) && !PageSwapCache(page)) { 839 if (!(sc->gfp_mask & __GFP_IO)) 840 goto keep_locked; 841 if (!add_to_swap(page)) 842 goto activate_locked; 843 may_enter_fs = 1; 844 } 845 846 mapping = page_mapping(page); 847 848 /* 849 * The page is mapped into the page tables of one or more 850 * processes. Try to unmap it here. 851 */ 852 if (page_mapped(page) && mapping) { 853 switch (try_to_unmap(page, TTU_UNMAP)) { 854 case SWAP_FAIL: 855 goto activate_locked; 856 case SWAP_AGAIN: 857 goto keep_locked; 858 case SWAP_MLOCK: 859 goto cull_mlocked; 860 case SWAP_SUCCESS: 861 ; /* try to free the page below */ 862 } 863 } 864 865 if (PageDirty(page)) { 866 nr_dirty++; 867 868 if (references == PAGEREF_RECLAIM_CLEAN) 869 goto keep_locked; 870 if (!may_enter_fs) 871 goto keep_locked; 872 if (!sc->may_writepage) 873 goto keep_locked; 874 875 /* Page is dirty, try to write it out here */ 876 switch (pageout(page, mapping, sc)) { 877 case PAGE_KEEP: 878 nr_congested++; 879 goto keep_locked; 880 case PAGE_ACTIVATE: 881 goto activate_locked; 882 case PAGE_SUCCESS: 883 if (PageWriteback(page)) 884 goto keep_lumpy; 885 if (PageDirty(page)) 886 goto keep; 887 888 /* 889 * A synchronous write - probably a ramdisk. Go 890 * ahead and try to reclaim the page. 891 */ 892 if (!trylock_page(page)) 893 goto keep; 894 if (PageDirty(page) || PageWriteback(page)) 895 goto keep_locked; 896 mapping = page_mapping(page); 897 case PAGE_CLEAN: 898 ; /* try to free the page below */ 899 } 900 } 901 902 /* 903 * If the page has buffers, try to free the buffer mappings 904 * associated with this page. If we succeed we try to free 905 * the page as well. 906 * 907 * We do this even if the page is PageDirty(). 908 * try_to_release_page() does not perform I/O, but it is 909 * possible for a page to have PageDirty set, but it is actually 910 * clean (all its buffers are clean). This happens if the 911 * buffers were written out directly, with submit_bh(). ext3 912 * will do this, as well as the blockdev mapping. 913 * try_to_release_page() will discover that cleanness and will 914 * drop the buffers and mark the page clean - it can be freed. 915 * 916 * Rarely, pages can have buffers and no ->mapping. These are 917 * the pages which were not successfully invalidated in 918 * truncate_complete_page(). We try to drop those buffers here 919 * and if that worked, and the page is no longer mapped into 920 * process address space (page_count == 1) it can be freed. 921 * Otherwise, leave the page on the LRU so it is swappable. 922 */ 923 if (page_has_private(page)) { 924 if (!try_to_release_page(page, sc->gfp_mask)) 925 goto activate_locked; 926 if (!mapping && page_count(page) == 1) { 927 unlock_page(page); 928 if (put_page_testzero(page)) 929 goto free_it; 930 else { 931 /* 932 * rare race with speculative reference. 933 * the speculative reference will free 934 * this page shortly, so we may 935 * increment nr_reclaimed here (and 936 * leave it off the LRU). 937 */ 938 nr_reclaimed++; 939 continue; 940 } 941 } 942 } 943 944 if (!mapping || !__remove_mapping(mapping, page)) 945 goto keep_locked; 946 947 /* 948 * At this point, we have no other references and there is 949 * no way to pick any more up (removed from LRU, removed 950 * from pagecache). Can use non-atomic bitops now (and 951 * we obviously don't have to worry about waking up a process 952 * waiting on the page lock, because there are no references. 953 */ 954 __clear_page_locked(page); 955free_it: 956 nr_reclaimed++; 957 958 /* 959 * Is there need to periodically free_page_list? It would 960 * appear not as the counts should be low 961 */ 962 list_add(&page->lru, &free_pages); 963 continue; 964 965cull_mlocked: 966 if (PageSwapCache(page)) 967 try_to_free_swap(page); 968 unlock_page(page); 969 putback_lru_page(page); 970 reset_reclaim_mode(sc); 971 continue; 972 973activate_locked: 974 /* Not a candidate for swapping, so reclaim swap space. */ 975 if (PageSwapCache(page) && vm_swap_full()) 976 try_to_free_swap(page); 977 VM_BUG_ON(PageActive(page)); 978 SetPageActive(page); 979 pgactivate++; 980keep_locked: 981 unlock_page(page); 982keep: 983 reset_reclaim_mode(sc); 984keep_lumpy: 985 list_add(&page->lru, &ret_pages); 986 VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); 987 } 988 989 /* 990 * Tag a zone as congested if all the dirty pages encountered were 991 * backed by a congested BDI. In this case, reclaimers should just 992 * back off and wait for congestion to clear because further reclaim 993 * will encounter the same problem 994 */ 995 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc)) 996 zone_set_flag(zone, ZONE_CONGESTED); 997 998 free_page_list(&free_pages); 999 1000 list_splice(&ret_pages, page_list); 1001 count_vm_events(PGACTIVATE, pgactivate); 1002 return nr_reclaimed; 1003} 1004 1005/* 1006 * Attempt to remove the specified page from its LRU. Only take this page 1007 * if it is of the appropriate PageActive status. Pages which are being 1008 * freed elsewhere are also ignored. 1009 * 1010 * page: page to consider 1011 * mode: one of the LRU isolation modes defined above 1012 * 1013 * returns 0 on success, -ve errno on failure. 1014 */ 1015int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file) 1016{ 1017 bool all_lru_mode; 1018 int ret = -EINVAL; 1019 1020 /* Only take pages on the LRU. */ 1021 if (!PageLRU(page)) 1022 return ret; 1023 1024 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) == 1025 (ISOLATE_ACTIVE|ISOLATE_INACTIVE); 1026 1027 /* 1028 * When checking the active state, we need to be sure we are 1029 * dealing with comparible boolean values. Take the logical not 1030 * of each. 1031 */ 1032 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE)) 1033 return ret; 1034 1035 if (!all_lru_mode && !!page_is_file_cache(page) != file) 1036 return ret; 1037 1038 /* 1039 * When this function is being called for lumpy reclaim, we 1040 * initially look into all LRU pages, active, inactive and 1041 * unevictable; only give shrink_page_list evictable pages. 1042 */ 1043 if (PageUnevictable(page)) 1044 return ret; 1045 1046 ret = -EBUSY; 1047 1048 if (likely(get_page_unless_zero(page))) { 1049 /* 1050 * Be careful not to clear PageLRU until after we're 1051 * sure the page is not being freed elsewhere -- the 1052 * page release code relies on it. 1053 */ 1054 ClearPageLRU(page); 1055 ret = 0; 1056 } 1057 1058 return ret; 1059} 1060 1061/* 1062 * zone->lru_lock is heavily contended. Some of the functions that 1063 * shrink the lists perform better by taking out a batch of pages 1064 * and working on them outside the LRU lock. 1065 * 1066 * For pagecache intensive workloads, this function is the hottest 1067 * spot in the kernel (apart from copy_*_user functions). 1068 * 1069 * Appropriate locks must be held before calling this function. 1070 * 1071 * @nr_to_scan: The number of pages to look through on the list. 1072 * @src: The LRU list to pull pages off. 1073 * @dst: The temp list to put pages on to. 1074 * @scanned: The number of pages that were scanned. 1075 * @order: The caller's attempted allocation order 1076 * @mode: One of the LRU isolation modes 1077 * @file: True [1] if isolating file [!anon] pages 1078 * 1079 * returns how many pages were moved onto *@dst. 1080 */ 1081static unsigned long isolate_lru_pages(unsigned long nr_to_scan, 1082 struct list_head *src, struct list_head *dst, 1083 unsigned long *scanned, int order, isolate_mode_t mode, 1084 int file) 1085{ 1086 unsigned long nr_taken = 0; 1087 unsigned long nr_lumpy_taken = 0; 1088 unsigned long nr_lumpy_dirty = 0; 1089 unsigned long nr_lumpy_failed = 0; 1090 unsigned long scan; 1091 1092 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { 1093 struct page *page; 1094 unsigned long pfn; 1095 unsigned long end_pfn; 1096 unsigned long page_pfn; 1097 int zone_id; 1098 1099 page = lru_to_page(src); 1100 prefetchw_prev_lru_page(page, src, flags); 1101 1102 VM_BUG_ON(!PageLRU(page)); 1103 1104 switch (__isolate_lru_page(page, mode, file)) { 1105 case 0: 1106 list_move(&page->lru, dst); 1107 mem_cgroup_del_lru(page); 1108 nr_taken += hpage_nr_pages(page); 1109 break; 1110 1111 case -EBUSY: 1112 /* else it is being freed elsewhere */ 1113 list_move(&page->lru, src); 1114 mem_cgroup_rotate_lru_list(page, page_lru(page)); 1115 continue; 1116 1117 default: 1118 BUG(); 1119 } 1120 1121 if (!order) 1122 continue; 1123 1124 /* 1125 * Attempt to take all pages in the order aligned region 1126 * surrounding the tag page. Only take those pages of 1127 * the same active state as that tag page. We may safely 1128 * round the target page pfn down to the requested order 1129 * as the mem_map is guaranteed valid out to MAX_ORDER, 1130 * where that page is in a different zone we will detect 1131 * it from its zone id and abort this block scan. 1132 */ 1133 zone_id = page_zone_id(page); 1134 page_pfn = page_to_pfn(page); 1135 pfn = page_pfn & ~((1 << order) - 1); 1136 end_pfn = pfn + (1 << order); 1137 for (; pfn < end_pfn; pfn++) { 1138 struct page *cursor_page; 1139 1140 /* The target page is in the block, ignore it. */ 1141 if (unlikely(pfn == page_pfn)) 1142 continue; 1143 1144 /* Avoid holes within the zone. */ 1145 if (unlikely(!pfn_valid_within(pfn))) 1146 break; 1147 1148 cursor_page = pfn_to_page(pfn); 1149 1150 /* Check that we have not crossed a zone boundary. */ 1151 if (unlikely(page_zone_id(cursor_page) != zone_id)) 1152 break; 1153 1154 /* 1155 * If we don't have enough swap space, reclaiming of 1156 * anon page which don't already have a swap slot is 1157 * pointless. 1158 */ 1159 if (nr_swap_pages <= 0 && PageAnon(cursor_page) && 1160 !PageSwapCache(cursor_page)) 1161 break; 1162 1163 if (__isolate_lru_page(cursor_page, mode, file) == 0) { 1164 list_move(&cursor_page->lru, dst); 1165 mem_cgroup_del_lru(cursor_page); 1166 nr_taken += hpage_nr_pages(page); 1167 nr_lumpy_taken++; 1168 if (PageDirty(cursor_page)) 1169 nr_lumpy_dirty++; 1170 scan++; 1171 } else { 1172 /* 1173 * Check if the page is freed already. 1174 * 1175 * We can't use page_count() as that 1176 * requires compound_head and we don't 1177 * have a pin on the page here. If a 1178 * page is tail, we may or may not 1179 * have isolated the head, so assume 1180 * it's not free, it'd be tricky to 1181 * track the head status without a 1182 * page pin. 1183 */ 1184 if (!PageTail(cursor_page) && 1185 !atomic_read(&cursor_page->_count)) 1186 continue; 1187 break; 1188 } 1189 } 1190 1191 /* If we break out of the loop above, lumpy reclaim failed */ 1192 if (pfn < end_pfn) 1193 nr_lumpy_failed++; 1194 } 1195 1196 *scanned = scan; 1197 1198 trace_mm_vmscan_lru_isolate(order, 1199 nr_to_scan, scan, 1200 nr_taken, 1201 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed, 1202 mode); 1203 return nr_taken; 1204} 1205 1206static unsigned long isolate_pages_global(unsigned long nr, 1207 struct list_head *dst, 1208 unsigned long *scanned, int order, 1209 isolate_mode_t mode, 1210 struct zone *z, int active, int file) 1211{ 1212 int lru = LRU_BASE; 1213 if (active) 1214 lru += LRU_ACTIVE; 1215 if (file) 1216 lru += LRU_FILE; 1217 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order, 1218 mode, file); 1219} 1220 1221/* 1222 * clear_active_flags() is a helper for shrink_active_list(), clearing 1223 * any active bits from the pages in the list. 1224 */ 1225static unsigned long clear_active_flags(struct list_head *page_list, 1226 unsigned int *count) 1227{ 1228 int nr_active = 0; 1229 int lru; 1230 struct page *page; 1231 1232 list_for_each_entry(page, page_list, lru) { 1233 int numpages = hpage_nr_pages(page); 1234 lru = page_lru_base_type(page); 1235 if (PageActive(page)) { 1236 lru += LRU_ACTIVE; 1237 ClearPageActive(page); 1238 nr_active += numpages; 1239 } 1240 if (count) 1241 count[lru] += numpages; 1242 } 1243 1244 return nr_active; 1245} 1246 1247/** 1248 * isolate_lru_page - tries to isolate a page from its LRU list 1249 * @page: page to isolate from its LRU list 1250 * 1251 * Isolates a @page from an LRU list, clears PageLRU and adjusts the 1252 * vmstat statistic corresponding to whatever LRU list the page was on. 1253 * 1254 * Returns 0 if the page was removed from an LRU list. 1255 * Returns -EBUSY if the page was not on an LRU list. 1256 * 1257 * The returned page will have PageLRU() cleared. If it was found on 1258 * the active list, it will have PageActive set. If it was found on 1259 * the unevictable list, it will have the PageUnevictable bit set. That flag 1260 * may need to be cleared by the caller before letting the page go. 1261 * 1262 * The vmstat statistic corresponding to the list on which the page was 1263 * found will be decremented. 1264 * 1265 * Restrictions: 1266 * (1) Must be called with an elevated refcount on the page. This is a 1267 * fundamentnal difference from isolate_lru_pages (which is called 1268 * without a stable reference). 1269 * (2) the lru_lock must not be held. 1270 * (3) interrupts must be enabled. 1271 */ 1272int isolate_lru_page(struct page *page) 1273{ 1274 int ret = -EBUSY; 1275 1276 VM_BUG_ON(!page_count(page)); 1277 1278 if (PageLRU(page)) { 1279 struct zone *zone = page_zone(page); 1280 1281 spin_lock_irq(&zone->lru_lock); 1282 if (PageLRU(page)) { 1283 int lru = page_lru(page); 1284 ret = 0; 1285 get_page(page); 1286 ClearPageLRU(page); 1287 1288 del_page_from_lru_list(zone, page, lru); 1289 } 1290 spin_unlock_irq(&zone->lru_lock); 1291 } 1292 return ret; 1293} 1294 1295/* 1296 * Are there way too many processes in the direct reclaim path already? 1297 */ 1298static int too_many_isolated(struct zone *zone, int file, 1299 struct scan_control *sc) 1300{ 1301 unsigned long inactive, isolated; 1302 1303 if (current_is_kswapd()) 1304 return 0; 1305 1306 if (!scanning_global_lru(sc)) 1307 return 0; 1308 1309 if (file) { 1310 inactive = zone_page_state(zone, NR_INACTIVE_FILE); 1311 isolated = zone_page_state(zone, NR_ISOLATED_FILE); 1312 } else { 1313 inactive = zone_page_state(zone, NR_INACTIVE_ANON); 1314 isolated = zone_page_state(zone, NR_ISOLATED_ANON); 1315 } 1316 1317 return isolated > inactive; 1318} 1319 1320/* 1321 * TODO: Try merging with migrations version of putback_lru_pages 1322 */ 1323static noinline_for_stack void 1324putback_lru_pages(struct zone *zone, struct scan_control *sc, 1325 unsigned long nr_anon, unsigned long nr_file, 1326 struct list_head *page_list) 1327{ 1328 struct page *page; 1329 struct pagevec pvec; 1330 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); 1331 1332 pagevec_init(&pvec, 1); 1333 1334 /* 1335 * Put back any unfreeable pages. 1336 */ 1337 spin_lock(&zone->lru_lock); 1338 while (!list_empty(page_list)) { 1339 int lru; 1340 page = lru_to_page(page_list); 1341 VM_BUG_ON(PageLRU(page)); 1342 list_del(&page->lru); 1343 if (unlikely(!page_evictable(page, NULL))) { 1344 spin_unlock_irq(&zone->lru_lock); 1345 putback_lru_page(page); 1346 spin_lock_irq(&zone->lru_lock); 1347 continue; 1348 } 1349 SetPageLRU(page); 1350 lru = page_lru(page); 1351 add_page_to_lru_list(zone, page, lru); 1352 if (is_active_lru(lru)) { 1353 int file = is_file_lru(lru); 1354 int numpages = hpage_nr_pages(page); 1355 reclaim_stat->recent_rotated[file] += numpages; 1356 } 1357 if (!pagevec_add(&pvec, page)) { 1358 spin_unlock_irq(&zone->lru_lock); 1359 __pagevec_release(&pvec); 1360 spin_lock_irq(&zone->lru_lock); 1361 } 1362 } 1363 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon); 1364 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file); 1365 1366 spin_unlock_irq(&zone->lru_lock); 1367 pagevec_release(&pvec); 1368} 1369 1370static noinline_for_stack void update_isolated_counts(struct zone *zone, 1371 struct scan_control *sc, 1372 unsigned long *nr_anon, 1373 unsigned long *nr_file, 1374 struct list_head *isolated_list) 1375{ 1376 unsigned long nr_active; 1377 unsigned int count[NR_LRU_LISTS] = { 0, }; 1378 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); 1379 1380 nr_active = clear_active_flags(isolated_list, count); 1381 __count_vm_events(PGDEACTIVATE, nr_active); 1382 1383 __mod_zone_page_state(zone, NR_ACTIVE_FILE, 1384 -count[LRU_ACTIVE_FILE]); 1385 __mod_zone_page_state(zone, NR_INACTIVE_FILE, 1386 -count[LRU_INACTIVE_FILE]); 1387 __mod_zone_page_state(zone, NR_ACTIVE_ANON, 1388 -count[LRU_ACTIVE_ANON]); 1389 __mod_zone_page_state(zone, NR_INACTIVE_ANON, 1390 -count[LRU_INACTIVE_ANON]); 1391 1392 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON]; 1393 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE]; 1394 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon); 1395 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file); 1396 1397 reclaim_stat->recent_scanned[0] += *nr_anon; 1398 reclaim_stat->recent_scanned[1] += *nr_file; 1399} 1400 1401/* 1402 * Returns true if the caller should wait to clean dirty/writeback pages. 1403 * 1404 * If we are direct reclaiming for contiguous pages and we do not reclaim 1405 * everything in the list, try again and wait for writeback IO to complete. 1406 * This will stall high-order allocations noticeably. Only do that when really 1407 * need to free the pages under high memory pressure. 1408 */ 1409static inline bool should_reclaim_stall(unsigned long nr_taken, 1410 unsigned long nr_freed, 1411 int priority, 1412 struct scan_control *sc) 1413{ 1414 int lumpy_stall_priority; 1415 1416 /* kswapd should not stall on sync IO */ 1417 if (current_is_kswapd()) 1418 return false; 1419 1420 /* Only stall on lumpy reclaim */ 1421 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE) 1422 return false; 1423 1424 /* If we have reclaimed everything on the isolated list, no stall */ 1425 if (nr_freed == nr_taken) 1426 return false; 1427 1428 /* 1429 * For high-order allocations, there are two stall thresholds. 1430 * High-cost allocations stall immediately where as lower 1431 * order allocations such as stacks require the scanning 1432 * priority to be much higher before stalling. 1433 */ 1434 if (sc->order > PAGE_ALLOC_COSTLY_ORDER) 1435 lumpy_stall_priority = DEF_PRIORITY; 1436 else 1437 lumpy_stall_priority = DEF_PRIORITY / 3; 1438 1439 return priority <= lumpy_stall_priority; 1440} 1441 1442/* 1443 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number 1444 * of reclaimed pages 1445 */ 1446static noinline_for_stack unsigned long 1447shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone, 1448 struct scan_control *sc, int priority, int file) 1449{ 1450 LIST_HEAD(page_list); 1451 unsigned long nr_scanned; 1452 unsigned long nr_reclaimed = 0; 1453 unsigned long nr_taken; 1454 unsigned long nr_anon; 1455 unsigned long nr_file; 1456 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE; 1457 1458 while (unlikely(too_many_isolated(zone, file, sc))) { 1459 congestion_wait(BLK_RW_ASYNC, HZ/10); 1460 1461 /* We are about to die and free our memory. Return now. */ 1462 if (fatal_signal_pending(current)) 1463 return SWAP_CLUSTER_MAX; 1464 } 1465 1466 set_reclaim_mode(priority, sc, false); 1467 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM) 1468 reclaim_mode |= ISOLATE_ACTIVE; 1469 1470 lru_add_drain(); 1471 spin_lock_irq(&zone->lru_lock); 1472 1473 if (scanning_global_lru(sc)) { 1474 nr_taken = isolate_pages_global(nr_to_scan, &page_list, 1475 &nr_scanned, sc->order, reclaim_mode, zone, 0, file); 1476 zone->pages_scanned += nr_scanned; 1477 if (current_is_kswapd()) 1478 __count_zone_vm_events(PGSCAN_KSWAPD, zone, 1479 nr_scanned); 1480 else 1481 __count_zone_vm_events(PGSCAN_DIRECT, zone, 1482 nr_scanned); 1483 } else { 1484 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list, 1485 &nr_scanned, sc->order, reclaim_mode, zone, 1486 sc->mem_cgroup, 0, file); 1487 /* 1488 * mem_cgroup_isolate_pages() keeps track of 1489 * scanned pages on its own. 1490 */ 1491 } 1492 1493 if (nr_taken == 0) { 1494 spin_unlock_irq(&zone->lru_lock); 1495 return 0; 1496 } 1497 1498 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list); 1499 1500 spin_unlock_irq(&zone->lru_lock); 1501 1502 nr_reclaimed = shrink_page_list(&page_list, zone, sc); 1503 1504 /* Check if we should syncronously wait for writeback */ 1505 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) { 1506 set_reclaim_mode(priority, sc, true); 1507 nr_reclaimed += shrink_page_list(&page_list, zone, sc); 1508 } 1509 1510 local_irq_disable(); 1511 if (current_is_kswapd()) 1512 __count_vm_events(KSWAPD_STEAL, nr_reclaimed); 1513 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed); 1514 1515 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list); 1516 1517 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, 1518 zone_idx(zone), 1519 nr_scanned, nr_reclaimed, 1520 priority, 1521 trace_shrink_flags(file, sc->reclaim_mode)); 1522 return nr_reclaimed; 1523} 1524 1525/* 1526 * This moves pages from the active list to the inactive list. 1527 * 1528 * We move them the other way if the page is referenced by one or more 1529 * processes, from rmap. 1530 * 1531 * If the pages are mostly unmapped, the processing is fast and it is 1532 * appropriate to hold zone->lru_lock across the whole operation. But if 1533 * the pages are mapped, the processing is slow (page_referenced()) so we 1534 * should drop zone->lru_lock around each page. It's impossible to balance 1535 * this, so instead we remove the pages from the LRU while processing them. 1536 * It is safe to rely on PG_active against the non-LRU pages in here because 1537 * nobody will play with that bit on a non-LRU page. 1538 * 1539 * The downside is that we have to touch page->_count against each page. 1540 * But we had to alter page->flags anyway. 1541 */ 1542 1543static void move_active_pages_to_lru(struct zone *zone, 1544 struct list_head *list, 1545 enum lru_list lru) 1546{ 1547 unsigned long pgmoved = 0; 1548 struct pagevec pvec; 1549 struct page *page; 1550 1551 pagevec_init(&pvec, 1); 1552 1553 while (!list_empty(list)) { 1554 page = lru_to_page(list); 1555 1556 VM_BUG_ON(PageLRU(page)); 1557 SetPageLRU(page); 1558 1559 list_move(&page->lru, &zone->lru[lru].list); 1560 mem_cgroup_add_lru_list(page, lru); 1561 pgmoved += hpage_nr_pages(page); 1562 1563 if (!pagevec_add(&pvec, page) || list_empty(list)) { 1564 spin_unlock_irq(&zone->lru_lock); 1565 if (buffer_heads_over_limit) 1566 pagevec_strip(&pvec); 1567 __pagevec_release(&pvec); 1568 spin_lock_irq(&zone->lru_lock); 1569 } 1570 } 1571 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); 1572 if (!is_active_lru(lru)) 1573 __count_vm_events(PGDEACTIVATE, pgmoved); 1574} 1575 1576static void shrink_active_list(unsigned long nr_pages, struct zone *zone, 1577 struct scan_control *sc, int priority, int file) 1578{ 1579 unsigned long nr_taken; 1580 unsigned long pgscanned; 1581 unsigned long vm_flags; 1582 LIST_HEAD(l_hold); /* The pages which were snipped off */ 1583 LIST_HEAD(l_active); 1584 LIST_HEAD(l_inactive); 1585 struct page *page; 1586 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); 1587 unsigned long nr_rotated = 0; 1588 1589 lru_add_drain(); 1590 spin_lock_irq(&zone->lru_lock); 1591 if (scanning_global_lru(sc)) { 1592 nr_taken = isolate_pages_global(nr_pages, &l_hold, 1593 &pgscanned, sc->order, 1594 ISOLATE_ACTIVE, zone, 1595 1, file); 1596 zone->pages_scanned += pgscanned; 1597 } else { 1598 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold, 1599 &pgscanned, sc->order, 1600 ISOLATE_ACTIVE, zone, 1601 sc->mem_cgroup, 1, file); 1602 /* 1603 * mem_cgroup_isolate_pages() keeps track of 1604 * scanned pages on its own. 1605 */ 1606 } 1607 1608 reclaim_stat->recent_scanned[file] += nr_taken; 1609 1610 __count_zone_vm_events(PGREFILL, zone, pgscanned); 1611 if (file) 1612 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken); 1613 else 1614 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken); 1615 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); 1616 spin_unlock_irq(&zone->lru_lock); 1617 1618 while (!list_empty(&l_hold)) { 1619 cond_resched(); 1620 page = lru_to_page(&l_hold); 1621 list_del(&page->lru); 1622 1623 if (unlikely(!page_evictable(page, NULL))) { 1624 putback_lru_page(page); 1625 continue; 1626 } 1627 1628 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) { 1629 nr_rotated += hpage_nr_pages(page); 1630 /* 1631 * Identify referenced, file-backed active pages and 1632 * give them one more trip around the active list. So 1633 * that executable code get better chances to stay in 1634 * memory under moderate memory pressure. Anon pages 1635 * are not likely to be evicted by use-once streaming 1636 * IO, plus JVM can create lots of anon VM_EXEC pages, 1637 * so we ignore them here. 1638 */ 1639 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { 1640 list_add(&page->lru, &l_active); 1641 continue; 1642 } 1643 } 1644 1645 ClearPageActive(page); /* we are de-activating */ 1646 list_add(&page->lru, &l_inactive); 1647 } 1648 1649 /* 1650 * Move pages back to the lru list. 1651 */ 1652 spin_lock_irq(&zone->lru_lock); 1653 /* 1654 * Count referenced pages from currently used mappings as rotated, 1655 * even though only some of them are actually re-activated. This 1656 * helps balance scan pressure between file and anonymous pages in 1657 * get_scan_ratio. 1658 */ 1659 reclaim_stat->recent_rotated[file] += nr_rotated; 1660 1661 move_active_pages_to_lru(zone, &l_active, 1662 LRU_ACTIVE + file * LRU_FILE); 1663 move_active_pages_to_lru(zone, &l_inactive, 1664 LRU_BASE + file * LRU_FILE); 1665 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); 1666 spin_unlock_irq(&zone->lru_lock); 1667} 1668 1669#ifdef CONFIG_SWAP 1670static int inactive_anon_is_low_global(struct zone *zone) 1671{ 1672 unsigned long active, inactive; 1673 1674 active = zone_page_state(zone, NR_ACTIVE_ANON); 1675 inactive = zone_page_state(zone, NR_INACTIVE_ANON); 1676 1677 if (inactive * zone->inactive_ratio < active) 1678 return 1; 1679 1680 return 0; 1681} 1682 1683/** 1684 * inactive_anon_is_low - check if anonymous pages need to be deactivated 1685 * @zone: zone to check 1686 * @sc: scan control of this context 1687 * 1688 * Returns true if the zone does not have enough inactive anon pages, 1689 * meaning some active anon pages need to be deactivated. 1690 */ 1691static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc) 1692{ 1693 int low; 1694 1695 /* 1696 * If we don't have swap space, anonymous page deactivation 1697 * is pointless. 1698 */ 1699 if (!total_swap_pages) 1700 return 0; 1701 1702 if (scanning_global_lru(sc)) 1703 low = inactive_anon_is_low_global(zone); 1704 else 1705 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup); 1706 return low; 1707} 1708#else 1709static inline int inactive_anon_is_low(struct zone *zone, 1710 struct scan_control *sc) 1711{ 1712 return 0; 1713} 1714#endif 1715 1716static int inactive_file_is_low_global(struct zone *zone) 1717{ 1718 unsigned long active, inactive; 1719 1720 active = zone_page_state(zone, NR_ACTIVE_FILE); 1721 inactive = zone_page_state(zone, NR_INACTIVE_FILE); 1722 1723 return (active > inactive); 1724} 1725 1726/** 1727 * inactive_file_is_low - check if file pages need to be deactivated 1728 * @zone: zone to check 1729 * @sc: scan control of this context 1730 * 1731 * When the system is doing streaming IO, memory pressure here 1732 * ensures that active file pages get deactivated, until more 1733 * than half of the file pages are on the inactive list. 1734 * 1735 * Once we get to that situation, protect the system's working 1736 * set from being evicted by disabling active file page aging. 1737 * 1738 * This uses a different ratio than the anonymous pages, because 1739 * the page cache uses a use-once replacement algorithm. 1740 */ 1741static int inactive_file_is_low(struct zone *zone, struct scan_control *sc) 1742{ 1743 int low; 1744 1745 if (scanning_global_lru(sc)) 1746 low = inactive_file_is_low_global(zone); 1747 else 1748 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup); 1749 return low; 1750} 1751 1752static int inactive_list_is_low(struct zone *zone, struct scan_control *sc, 1753 int file) 1754{ 1755 if (file) 1756 return inactive_file_is_low(zone, sc); 1757 else 1758 return inactive_anon_is_low(zone, sc); 1759} 1760 1761static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, 1762 struct zone *zone, struct scan_control *sc, int priority) 1763{ 1764 int file = is_file_lru(lru); 1765 1766 if (is_active_lru(lru)) { 1767 if (inactive_list_is_low(zone, sc, file)) 1768 shrink_active_list(nr_to_scan, zone, sc, priority, file); 1769 return 0; 1770 } 1771 1772 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file); 1773} 1774 1775static int vmscan_swappiness(struct scan_control *sc) 1776{ 1777 if (scanning_global_lru(sc)) 1778 return vm_swappiness; 1779 return mem_cgroup_swappiness(sc->mem_cgroup); 1780} 1781 1782/* 1783 * Determine how aggressively the anon and file LRU lists should be 1784 * scanned. The relative value of each set of LRU lists is determined 1785 * by looking at the fraction of the pages scanned we did rotate back 1786 * onto the active list instead of evict. 1787 * 1788 * nr[0] = anon pages to scan; nr[1] = file pages to scan 1789 */ 1790static void get_scan_count(struct zone *zone, struct scan_control *sc, 1791 unsigned long *nr, int priority) 1792{ 1793 unsigned long anon, file, free; 1794 unsigned long anon_prio, file_prio; 1795 unsigned long ap, fp; 1796 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); 1797 u64 fraction[2], denominator; 1798 enum lru_list l; 1799 int noswap = 0; 1800 bool force_scan = false; 1801 unsigned long nr_force_scan[2]; 1802 1803 /* kswapd does zone balancing and needs to scan this zone */ 1804 if (scanning_global_lru(sc) && current_is_kswapd()) 1805 force_scan = true; 1806 /* memcg may have small limit and need to avoid priority drop */ 1807 if (!scanning_global_lru(sc)) 1808 force_scan = true; 1809 1810 /* If we have no swap space, do not bother scanning anon pages. */ 1811 if (!sc->may_swap || (nr_swap_pages <= 0)) { 1812 noswap = 1; 1813 fraction[0] = 0; 1814 fraction[1] = 1; 1815 denominator = 1; 1816 nr_force_scan[0] = 0; 1817 nr_force_scan[1] = SWAP_CLUSTER_MAX; 1818 goto out; 1819 } 1820 1821 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) + 1822 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON); 1823 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) + 1824 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE); 1825 1826 if (scanning_global_lru(sc)) { 1827 free = zone_page_state(zone, NR_FREE_PAGES); 1828 /* If we have very few page cache pages, 1829 force-scan anon pages. */ 1830 if (unlikely(file + free <= high_wmark_pages(zone))) { 1831 fraction[0] = 1; 1832 fraction[1] = 0; 1833 denominator = 1; 1834 nr_force_scan[0] = SWAP_CLUSTER_MAX; 1835 nr_force_scan[1] = 0; 1836 goto out; 1837 } 1838 } 1839 1840 /* 1841 * With swappiness at 100, anonymous and file have the same priority. 1842 * This scanning priority is essentially the inverse of IO cost. 1843 */ 1844 anon_prio = vmscan_swappiness(sc); 1845 file_prio = 200 - vmscan_swappiness(sc); 1846 1847 /* 1848 * OK, so we have swap space and a fair amount of page cache 1849 * pages. We use the recently rotated / recently scanned 1850 * ratios to determine how valuable each cache is. 1851 * 1852 * Because workloads change over time (and to avoid overflow) 1853 * we keep these statistics as a floating average, which ends 1854 * up weighing recent references more than old ones. 1855 * 1856 * anon in [0], file in [1] 1857 */ 1858 spin_lock_irq(&zone->lru_lock); 1859 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { 1860 reclaim_stat->recent_scanned[0] /= 2; 1861 reclaim_stat->recent_rotated[0] /= 2; 1862 } 1863 1864 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { 1865 reclaim_stat->recent_scanned[1] /= 2; 1866 reclaim_stat->recent_rotated[1] /= 2; 1867 } 1868 1869 /* 1870 * The amount of pressure on anon vs file pages is inversely 1871 * proportional to the fraction of recently scanned pages on 1872 * each list that were recently referenced and in active use. 1873 */ 1874 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1); 1875 ap /= reclaim_stat->recent_rotated[0] + 1; 1876 1877 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1); 1878 fp /= reclaim_stat->recent_rotated[1] + 1; 1879 spin_unlock_irq(&zone->lru_lock); 1880 1881 fraction[0] = ap; 1882 fraction[1] = fp; 1883 denominator = ap + fp + 1; 1884 if (force_scan) { 1885 unsigned long scan = SWAP_CLUSTER_MAX; 1886 nr_force_scan[0] = div64_u64(scan * ap, denominator); 1887 nr_force_scan[1] = div64_u64(scan * fp, denominator); 1888 } 1889out: 1890 for_each_evictable_lru(l) { 1891 int file = is_file_lru(l); 1892 unsigned long scan; 1893 1894 scan = zone_nr_lru_pages(zone, sc, l); 1895 if (priority || noswap) { 1896 scan >>= priority; 1897 scan = div64_u64(scan * fraction[file], denominator); 1898 } 1899 1900 /* 1901 * If zone is small or memcg is small, nr[l] can be 0. 1902 * This results no-scan on this priority and priority drop down. 1903 * For global direct reclaim, it can visit next zone and tend 1904 * not to have problems. For global kswapd, it's for zone 1905 * balancing and it need to scan a small amounts. When using 1906 * memcg, priority drop can cause big latency. So, it's better 1907 * to scan small amount. See may_noscan above. 1908 */ 1909 if (!scan && force_scan) 1910 scan = nr_force_scan[file]; 1911 nr[l] = scan; 1912 } 1913} 1914 1915/* 1916 * Reclaim/compaction depends on a number of pages being freed. To avoid 1917 * disruption to the system, a small number of order-0 pages continue to be 1918 * rotated and reclaimed in the normal fashion. However, by the time we get 1919 * back to the allocator and call try_to_compact_zone(), we ensure that 1920 * there are enough free pages for it to be likely successful 1921 */ 1922static inline bool should_continue_reclaim(struct zone *zone, 1923 unsigned long nr_reclaimed, 1924 unsigned long nr_scanned, 1925 struct scan_control *sc) 1926{ 1927 unsigned long pages_for_compaction; 1928 unsigned long inactive_lru_pages; 1929 1930 /* If not in reclaim/compaction mode, stop */ 1931 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION)) 1932 return false; 1933 1934 /* Consider stopping depending on scan and reclaim activity */ 1935 if (sc->gfp_mask & __GFP_REPEAT) { 1936 /* 1937 * For __GFP_REPEAT allocations, stop reclaiming if the 1938 * full LRU list has been scanned and we are still failing 1939 * to reclaim pages. This full LRU scan is potentially 1940 * expensive but a __GFP_REPEAT caller really wants to succeed 1941 */ 1942 if (!nr_reclaimed && !nr_scanned) 1943 return false; 1944 } else { 1945 /* 1946 * For non-__GFP_REPEAT allocations which can presumably 1947 * fail without consequence, stop if we failed to reclaim 1948 * any pages from the last SWAP_CLUSTER_MAX number of 1949 * pages that were scanned. This will return to the 1950 * caller faster at the risk reclaim/compaction and 1951 * the resulting allocation attempt fails 1952 */ 1953 if (!nr_reclaimed) 1954 return false; 1955 } 1956 1957 /* 1958 * If we have not reclaimed enough pages for compaction and the 1959 * inactive lists are large enough, continue reclaiming 1960 */ 1961 pages_for_compaction = (2UL << sc->order); 1962 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) + 1963 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE); 1964 if (sc->nr_reclaimed < pages_for_compaction && 1965 inactive_lru_pages > pages_for_compaction) 1966 return true; 1967 1968 /* If compaction would go ahead or the allocation would succeed, stop */ 1969 switch (compaction_suitable(zone, sc->order)) { 1970 case COMPACT_PARTIAL: 1971 case COMPACT_CONTINUE: 1972 return false; 1973 default: 1974 return true; 1975 } 1976} 1977 1978/* 1979 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. 1980 */ 1981static void shrink_zone(int priority, struct zone *zone, 1982 struct scan_control *sc) 1983{ 1984 unsigned long nr[NR_LRU_LISTS]; 1985 unsigned long nr_to_scan; 1986 enum lru_list l; 1987 unsigned long nr_reclaimed, nr_scanned; 1988 unsigned long nr_to_reclaim = sc->nr_to_reclaim; 1989 1990restart: 1991 nr_reclaimed = 0; 1992 nr_scanned = sc->nr_scanned; 1993 get_scan_count(zone, sc, nr, priority); 1994 1995 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || 1996 nr[LRU_INACTIVE_FILE]) { 1997 for_each_evictable_lru(l) { 1998 if (nr[l]) { 1999 nr_to_scan = min_t(unsigned long, 2000 nr[l], SWAP_CLUSTER_MAX); 2001 nr[l] -= nr_to_scan; 2002 2003 nr_reclaimed += shrink_list(l, nr_to_scan, 2004 zone, sc, priority); 2005 } 2006 } 2007 /* 2008 * On large memory systems, scan >> priority can become 2009 * really large. This is fine for the starting priority; 2010 * we want to put equal scanning pressure on each zone. 2011 * However, if the VM has a harder time of freeing pages, 2012 * with multiple processes reclaiming pages, the total 2013 * freeing target can get unreasonably large. 2014 */ 2015 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY) 2016 break; 2017 } 2018 sc->nr_reclaimed += nr_reclaimed; 2019 2020 /* 2021 * Even if we did not try to evict anon pages at all, we want to 2022 * rebalance the anon lru active/inactive ratio. 2023 */ 2024 if (inactive_anon_is_low(zone, sc)) 2025 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0); 2026 2027 /* reclaim/compaction might need reclaim to continue */ 2028 if (should_continue_reclaim(zone, nr_reclaimed, 2029 sc->nr_scanned - nr_scanned, sc)) 2030 goto restart; 2031 2032 throttle_vm_writeout(sc->gfp_mask); 2033} 2034 2035/* 2036 * This is the direct reclaim path, for page-allocating processes. We only 2037 * try to reclaim pages from zones which will satisfy the caller's allocation 2038 * request. 2039 * 2040 * We reclaim from a zone even if that zone is over high_wmark_pages(zone). 2041 * Because: 2042 * a) The caller may be trying to free *extra* pages to satisfy a higher-order 2043 * allocation or 2044 * b) The target zone may be at high_wmark_pages(zone) but the lower zones 2045 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' 2046 * zone defense algorithm. 2047 * 2048 * If a zone is deemed to be full of pinned pages then just give it a light 2049 * scan then give up on it. 2050 */ 2051static void shrink_zones(int priority, struct zonelist *zonelist, 2052 struct scan_control *sc) 2053{ 2054 struct zoneref *z; 2055 struct zone *zone; 2056 unsigned long nr_soft_reclaimed; 2057 unsigned long nr_soft_scanned; 2058 2059 for_each_zone_zonelist_nodemask(zone, z, zonelist, 2060 gfp_zone(sc->gfp_mask), sc->nodemask) { 2061 if (!populated_zone(zone)) 2062 continue; 2063 /* 2064 * Take care memory controller reclaiming has small influence 2065 * to global LRU. 2066 */ 2067 if (scanning_global_lru(sc)) { 2068 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 2069 continue; 2070 if (zone->all_unreclaimable && priority != DEF_PRIORITY) 2071 continue; /* Let kswapd poll it */ 2072 /* 2073 * This steals pages from memory cgroups over softlimit 2074 * and returns the number of reclaimed pages and 2075 * scanned pages. This works for global memory pressure 2076 * and balancing, not for a memcg's limit. 2077 */ 2078 nr_soft_scanned = 0; 2079 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, 2080 sc->order, sc->gfp_mask, 2081 &nr_soft_scanned); 2082 sc->nr_reclaimed += nr_soft_reclaimed; 2083 sc->nr_scanned += nr_soft_scanned; 2084 /* need some check for avoid more shrink_zone() */ 2085 } 2086 2087 shrink_zone(priority, zone, sc); 2088 } 2089} 2090 2091static bool zone_reclaimable(struct zone *zone) 2092{ 2093 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6; 2094} 2095 2096/* All zones in zonelist are unreclaimable? */ 2097static bool all_unreclaimable(struct zonelist *zonelist, 2098 struct scan_control *sc) 2099{ 2100 struct zoneref *z; 2101 struct zone *zone; 2102 2103 for_each_zone_zonelist_nodemask(zone, z, zonelist, 2104 gfp_zone(sc->gfp_mask), sc->nodemask) { 2105 if (!populated_zone(zone)) 2106 continue; 2107 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 2108 continue; 2109 if (!zone->all_unreclaimable) 2110 return false; 2111 } 2112 2113 return true; 2114} 2115 2116/* 2117 * This is the main entry point to direct page reclaim. 2118 * 2119 * If a full scan of the inactive list fails to free enough memory then we 2120 * are "out of memory" and something needs to be killed. 2121 * 2122 * If the caller is !__GFP_FS then the probability of a failure is reasonably 2123 * high - the zone may be full of dirty or under-writeback pages, which this 2124 * caller can't do much about. We kick the writeback threads and take explicit 2125 * naps in the hope that some of these pages can be written. But if the 2126 * allocating task holds filesystem locks which prevent writeout this might not 2127 * work, and the allocation attempt will fail. 2128 * 2129 * returns: 0, if no pages reclaimed 2130 * else, the number of pages reclaimed 2131 */ 2132static unsigned long do_try_to_free_pages(struct zonelist *zonelist, 2133 struct scan_control *sc, 2134 struct shrink_control *shrink) 2135{ 2136 int priority; 2137 unsigned long total_scanned = 0; 2138 struct reclaim_state *reclaim_state = current->reclaim_state; 2139 struct zoneref *z; 2140 struct zone *zone; 2141 unsigned long writeback_threshold; 2142 2143 get_mems_allowed(); 2144 delayacct_freepages_start(); 2145 2146 if (scanning_global_lru(sc)) 2147 count_vm_event(ALLOCSTALL); 2148 2149 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 2150 sc->nr_scanned = 0; 2151 if (!priority) 2152 disable_swap_token(sc->mem_cgroup); 2153 shrink_zones(priority, zonelist, sc); 2154 /* 2155 * Don't shrink slabs when reclaiming memory from 2156 * over limit cgroups 2157 */ 2158 if (scanning_global_lru(sc)) { 2159 unsigned long lru_pages = 0; 2160 for_each_zone_zonelist(zone, z, zonelist, 2161 gfp_zone(sc->gfp_mask)) { 2162 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 2163 continue; 2164 2165 lru_pages += zone_reclaimable_pages(zone); 2166 } 2167 2168 shrink_slab(shrink, sc->nr_scanned, lru_pages); 2169 if (reclaim_state) { 2170 sc->nr_reclaimed += reclaim_state->reclaimed_slab; 2171 reclaim_state->reclaimed_slab = 0; 2172 } 2173 } 2174 total_scanned += sc->nr_scanned; 2175 if (sc->nr_reclaimed >= sc->nr_to_reclaim) 2176 goto out; 2177 2178 /* 2179 * Try to write back as many pages as we just scanned. This 2180 * tends to cause slow streaming writers to write data to the 2181 * disk smoothly, at the dirtying rate, which is nice. But 2182 * that's undesirable in laptop mode, where we *want* lumpy 2183 * writeout. So in laptop mode, write out the whole world. 2184 */ 2185 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; 2186 if (total_scanned > writeback_threshold) { 2187 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned); 2188 sc->may_writepage = 1; 2189 } 2190 2191 /* Take a nap, wait for some writeback to complete */ 2192 if (!sc->hibernation_mode && sc->nr_scanned && 2193 priority < DEF_PRIORITY - 2) { 2194 struct zone *preferred_zone; 2195 2196 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask), 2197 &cpuset_current_mems_allowed, 2198 &preferred_zone); 2199 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10); 2200 } 2201 } 2202 2203out: 2204 delayacct_freepages_end(); 2205 put_mems_allowed(); 2206 2207 if (sc->nr_reclaimed) 2208 return sc->nr_reclaimed; 2209 2210 /* 2211 * As hibernation is going on, kswapd is freezed so that it can't mark 2212 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable 2213 * check. 2214 */ 2215 if (oom_killer_disabled) 2216 return 0; 2217 2218 /* top priority shrink_zones still had more to do? don't OOM, then */ 2219 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc)) 2220 return 1; 2221 2222 return 0; 2223} 2224 2225unsigned long try_to_free_pages(struct zonelist *zonelist, int order, 2226 gfp_t gfp_mask, nodemask_t *nodemask) 2227{ 2228 unsigned long nr_reclaimed; 2229 struct scan_control sc = { 2230 .gfp_mask = gfp_mask, 2231 .may_writepage = !laptop_mode, 2232 .nr_to_reclaim = SWAP_CLUSTER_MAX, 2233 .may_unmap = 1, 2234 .may_swap = 1, 2235 .order = order, 2236 .mem_cgroup = NULL, 2237 .nodemask = nodemask, 2238 }; 2239 struct shrink_control shrink = { 2240 .gfp_mask = sc.gfp_mask, 2241 }; 2242 2243 trace_mm_vmscan_direct_reclaim_begin(order, 2244 sc.may_writepage, 2245 gfp_mask); 2246 2247 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); 2248 2249 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); 2250 2251 return nr_reclaimed; 2252} 2253 2254#ifdef CONFIG_CGROUP_MEM_RES_CTLR 2255 2256unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem, 2257 gfp_t gfp_mask, bool noswap, 2258 struct zone *zone, 2259 unsigned long *nr_scanned) 2260{ 2261 struct scan_control sc = { 2262 .nr_scanned = 0, 2263 .nr_to_reclaim = SWAP_CLUSTER_MAX, 2264 .may_writepage = !laptop_mode, 2265 .may_unmap = 1, 2266 .may_swap = !noswap, 2267 .order = 0, 2268 .mem_cgroup = mem, 2269 }; 2270 2271 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | 2272 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); 2273 2274 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0, 2275 sc.may_writepage, 2276 sc.gfp_mask); 2277 2278 /* 2279 * NOTE: Although we can get the priority field, using it 2280 * here is not a good idea, since it limits the pages we can scan. 2281 * if we don't reclaim here, the shrink_zone from balance_pgdat 2282 * will pick up pages from other mem cgroup's as well. We hack 2283 * the priority and make it zero. 2284 */ 2285 shrink_zone(0, zone, &sc); 2286 2287 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); 2288 2289 *nr_scanned = sc.nr_scanned; 2290 return sc.nr_reclaimed; 2291} 2292 2293unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont, 2294 gfp_t gfp_mask, 2295 bool noswap) 2296{ 2297 struct zonelist *zonelist; 2298 unsigned long nr_reclaimed; 2299 int nid; 2300 struct scan_control sc = { 2301 .may_writepage = !laptop_mode, 2302 .may_unmap = 1, 2303 .may_swap = !noswap, 2304 .nr_to_reclaim = SWAP_CLUSTER_MAX, 2305 .order = 0, 2306 .mem_cgroup = mem_cont, 2307 .nodemask = NULL, /* we don't care the placement */ 2308 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | 2309 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), 2310 }; 2311 struct shrink_control shrink = { 2312 .gfp_mask = sc.gfp_mask, 2313 }; 2314 2315 /* 2316 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't 2317 * take care of from where we get pages. So the node where we start the 2318 * scan does not need to be the current node. 2319 */ 2320 nid = mem_cgroup_select_victim_node(mem_cont); 2321 2322 zonelist = NODE_DATA(nid)->node_zonelists; 2323 2324 trace_mm_vmscan_memcg_reclaim_begin(0, 2325 sc.may_writepage, 2326 sc.gfp_mask); 2327 2328 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); 2329 2330 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); 2331 2332 return nr_reclaimed; 2333} 2334#endif 2335 2336/* 2337 * pgdat_balanced is used when checking if a node is balanced for high-order 2338 * allocations. Only zones that meet watermarks and are in a zone allowed 2339 * by the callers classzone_idx are added to balanced_pages. The total of 2340 * balanced pages must be at least 25% of the zones allowed by classzone_idx 2341 * for the node to be considered balanced. Forcing all zones to be balanced 2342 * for high orders can cause excessive reclaim when there are imbalanced zones. 2343 * The choice of 25% is due to 2344 * o a 16M DMA zone that is balanced will not balance a zone on any 2345 * reasonable sized machine 2346 * o On all other machines, the top zone must be at least a reasonable 2347 * percentage of the middle zones. For example, on 32-bit x86, highmem 2348 * would need to be at least 256M for it to be balance a whole node. 2349 * Similarly, on x86-64 the Normal zone would need to be at least 1G 2350 * to balance a node on its own. These seemed like reasonable ratios. 2351 */ 2352static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages, 2353 int classzone_idx) 2354{ 2355 unsigned long present_pages = 0; 2356 int i; 2357 2358 for (i = 0; i <= classzone_idx; i++) 2359 present_pages += pgdat->node_zones[i].present_pages; 2360 2361 /* A special case here: if zone has no page, we think it's balanced */ 2362 return balanced_pages >= (present_pages >> 2); 2363} 2364 2365/* is kswapd sleeping prematurely? */ 2366static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining, 2367 int classzone_idx) 2368{ 2369 int i; 2370 unsigned long balanced = 0; 2371 bool all_zones_ok = true; 2372 2373 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ 2374 if (remaining) 2375 return true; 2376 2377 /* Check the watermark levels */ 2378 for (i = 0; i <= classzone_idx; i++) { 2379 struct zone *zone = pgdat->node_zones + i; 2380 2381 if (!populated_zone(zone)) 2382 continue; 2383 2384 /* 2385 * balance_pgdat() skips over all_unreclaimable after 2386 * DEF_PRIORITY. Effectively, it considers them balanced so 2387 * they must be considered balanced here as well if kswapd 2388 * is to sleep 2389 */ 2390 if (zone->all_unreclaimable) { 2391 balanced += zone->present_pages; 2392 continue; 2393 } 2394 2395 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone), 2396 i, 0)) 2397 all_zones_ok = false; 2398 else 2399 balanced += zone->present_pages; 2400 } 2401 2402 /* 2403 * For high-order requests, the balanced zones must contain at least 2404 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones 2405 * must be balanced 2406 */ 2407 if (order) 2408 return !pgdat_balanced(pgdat, balanced, classzone_idx); 2409 else 2410 return !all_zones_ok; 2411} 2412 2413/* 2414 * For kswapd, balance_pgdat() will work across all this node's zones until 2415 * they are all at high_wmark_pages(zone). 2416 * 2417 * Returns the final order kswapd was reclaiming at 2418 * 2419 * There is special handling here for zones which are full of pinned pages. 2420 * This can happen if the pages are all mlocked, or if they are all used by 2421 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. 2422 * What we do is to detect the case where all pages in the zone have been 2423 * scanned twice and there has been zero successful reclaim. Mark the zone as 2424 * dead and from now on, only perform a short scan. Basically we're polling 2425 * the zone for when the problem goes away. 2426 * 2427 * kswapd scans the zones in the highmem->normal->dma direction. It skips 2428 * zones which have free_pages > high_wmark_pages(zone), but once a zone is 2429 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the 2430 * lower zones regardless of the number of free pages in the lower zones. This 2431 * interoperates with the page allocator fallback scheme to ensure that aging 2432 * of pages is balanced across the zones. 2433 */ 2434static unsigned long balance_pgdat(pg_data_t *pgdat, int order, 2435 int *classzone_idx) 2436{ 2437 int all_zones_ok; 2438 unsigned long balanced; 2439 int priority; 2440 int i; 2441 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ 2442 unsigned long total_scanned; 2443 struct reclaim_state *reclaim_state = current->reclaim_state; 2444 unsigned long nr_soft_reclaimed; 2445 unsigned long nr_soft_scanned; 2446 struct scan_control sc = { 2447 .gfp_mask = GFP_KERNEL, 2448 .may_unmap = 1, 2449 .may_swap = 1, 2450 /* 2451 * kswapd doesn't want to be bailed out while reclaim. because 2452 * we want to put equal scanning pressure on each zone. 2453 */ 2454 .nr_to_reclaim = ULONG_MAX, 2455 .order = order, 2456 .mem_cgroup = NULL, 2457 }; 2458 struct shrink_control shrink = { 2459 .gfp_mask = sc.gfp_mask, 2460 }; 2461loop_again: 2462 total_scanned = 0; 2463 sc.nr_reclaimed = 0; 2464 sc.may_writepage = !laptop_mode; 2465 count_vm_event(PAGEOUTRUN); 2466 2467 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 2468 unsigned long lru_pages = 0; 2469 int has_under_min_watermark_zone = 0; 2470 2471 /* The swap token gets in the way of swapout... */ 2472 if (!priority) 2473 disable_swap_token(NULL); 2474 2475 all_zones_ok = 1; 2476 balanced = 0; 2477 2478 /* 2479 * Scan in the highmem->dma direction for the highest 2480 * zone which needs scanning 2481 */ 2482 for (i = pgdat->nr_zones - 1; i >= 0; i--) { 2483 struct zone *zone = pgdat->node_zones + i; 2484 2485 if (!populated_zone(zone)) 2486 continue; 2487 2488 if (zone->all_unreclaimable && priority != DEF_PRIORITY) 2489 continue; 2490 2491 /* 2492 * Do some background aging of the anon list, to give 2493 * pages a chance to be referenced before reclaiming. 2494 */ 2495 if (inactive_anon_is_low(zone, &sc)) 2496 shrink_active_list(SWAP_CLUSTER_MAX, zone, 2497 &sc, priority, 0); 2498 2499 if (!zone_watermark_ok_safe(zone, order, 2500 high_wmark_pages(zone), 0, 0)) { 2501 end_zone = i; 2502 break; 2503 } else { 2504 /* If balanced, clear the congested flag */ 2505 zone_clear_flag(zone, ZONE_CONGESTED); 2506 } 2507 } 2508 if (i < 0) 2509 goto out; 2510 2511 for (i = 0; i <= end_zone; i++) { 2512 struct zone *zone = pgdat->node_zones + i; 2513 2514 lru_pages += zone_reclaimable_pages(zone); 2515 } 2516 2517 /* 2518 * Now scan the zone in the dma->highmem direction, stopping 2519 * at the last zone which needs scanning. 2520 * 2521 * We do this because the page allocator works in the opposite 2522 * direction. This prevents the page allocator from allocating 2523 * pages behind kswapd's direction of progress, which would 2524 * cause too much scanning of the lower zones. 2525 */ 2526 for (i = 0; i <= end_zone; i++) { 2527 struct zone *zone = pgdat->node_zones + i; 2528 int nr_slab; 2529 unsigned long balance_gap; 2530 2531 if (!populated_zone(zone)) 2532 continue; 2533 2534 if (zone->all_unreclaimable && priority != DEF_PRIORITY) 2535 continue; 2536 2537 sc.nr_scanned = 0; 2538 2539 nr_soft_scanned = 0; 2540 /* 2541 * Call soft limit reclaim before calling shrink_zone. 2542 */ 2543 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, 2544 order, sc.gfp_mask, 2545 &nr_soft_scanned); 2546 sc.nr_reclaimed += nr_soft_reclaimed; 2547 total_scanned += nr_soft_scanned; 2548 2549 /* 2550 * We put equal pressure on every zone, unless 2551 * one zone has way too many pages free 2552 * already. The "too many pages" is defined 2553 * as the high wmark plus a "gap" where the 2554 * gap is either the low watermark or 1% 2555 * of the zone, whichever is smaller. 2556 */ 2557 balance_gap = min(low_wmark_pages(zone), 2558 (zone->present_pages + 2559 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) / 2560 KSWAPD_ZONE_BALANCE_GAP_RATIO); 2561 if (!zone_watermark_ok_safe(zone, order, 2562 high_wmark_pages(zone) + balance_gap, 2563 end_zone, 0)) { 2564 shrink_zone(priority, zone, &sc); 2565 2566 reclaim_state->reclaimed_slab = 0; 2567 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages); 2568 sc.nr_reclaimed += reclaim_state->reclaimed_slab; 2569 total_scanned += sc.nr_scanned; 2570 2571 if (nr_slab == 0 && !zone_reclaimable(zone)) 2572 zone->all_unreclaimable = 1; 2573 } 2574 2575 /* 2576 * If we've done a decent amount of scanning and 2577 * the reclaim ratio is low, start doing writepage 2578 * even in laptop mode 2579 */ 2580 if (total_scanned > SWAP_CLUSTER_MAX * 2 && 2581 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2) 2582 sc.may_writepage = 1; 2583 2584 if (zone->all_unreclaimable) { 2585 if (end_zone && end_zone == i) 2586 end_zone--; 2587 continue; 2588 } 2589 2590 if (!zone_watermark_ok_safe(zone, order, 2591 high_wmark_pages(zone), end_zone, 0)) { 2592 all_zones_ok = 0; 2593 /* 2594 * We are still under min water mark. This 2595 * means that we have a GFP_ATOMIC allocation 2596 * failure risk. Hurry up! 2597 */ 2598 if (!zone_watermark_ok_safe(zone, order, 2599 min_wmark_pages(zone), end_zone, 0)) 2600 has_under_min_watermark_zone = 1; 2601 } else { 2602 /* 2603 * If a zone reaches its high watermark, 2604 * consider it to be no longer congested. It's 2605 * possible there are dirty pages backed by 2606 * congested BDIs but as pressure is relieved, 2607 * spectulatively avoid congestion waits 2608 */ 2609 zone_clear_flag(zone, ZONE_CONGESTED); 2610 if (i <= *classzone_idx) 2611 balanced += zone->present_pages; 2612 } 2613 2614 } 2615 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx))) 2616 break; /* kswapd: all done */ 2617 /* 2618 * OK, kswapd is getting into trouble. Take a nap, then take 2619 * another pass across the zones. 2620 */ 2621 if (total_scanned && (priority < DEF_PRIORITY - 2)) { 2622 if (has_under_min_watermark_zone) 2623 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT); 2624 else 2625 congestion_wait(BLK_RW_ASYNC, HZ/10); 2626 } 2627 2628 /* 2629 * We do this so kswapd doesn't build up large priorities for 2630 * example when it is freeing in parallel with allocators. It 2631 * matches the direct reclaim path behaviour in terms of impact 2632 * on zone->*_priority. 2633 */ 2634 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) 2635 break; 2636 } 2637out: 2638 2639 /* 2640 * order-0: All zones must meet high watermark for a balanced node 2641 * high-order: Balanced zones must make up at least 25% of the node 2642 * for the node to be balanced 2643 */ 2644 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) { 2645 cond_resched(); 2646 2647 try_to_freeze(); 2648 2649 /* 2650 * Fragmentation may mean that the system cannot be 2651 * rebalanced for high-order allocations in all zones. 2652 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX, 2653 * it means the zones have been fully scanned and are still 2654 * not balanced. For high-order allocations, there is 2655 * little point trying all over again as kswapd may 2656 * infinite loop. 2657 * 2658 * Instead, recheck all watermarks at order-0 as they 2659 * are the most important. If watermarks are ok, kswapd will go 2660 * back to sleep. High-order users can still perform direct 2661 * reclaim if they wish. 2662 */ 2663 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX) 2664 order = sc.order = 0; 2665 2666 goto loop_again; 2667 } 2668 2669 /* 2670 * If kswapd was reclaiming at a higher order, it has the option of 2671 * sleeping without all zones being balanced. Before it does, it must 2672 * ensure that the watermarks for order-0 on *all* zones are met and 2673 * that the congestion flags are cleared. The congestion flag must 2674 * be cleared as kswapd is the only mechanism that clears the flag 2675 * and it is potentially going to sleep here. 2676 */ 2677 if (order) { 2678 for (i = 0; i <= end_zone; i++) { 2679 struct zone *zone = pgdat->node_zones + i; 2680 2681 if (!populated_zone(zone)) 2682 continue; 2683 2684 if (zone->all_unreclaimable && priority != DEF_PRIORITY) 2685 continue; 2686 2687 /* Confirm the zone is balanced for order-0 */ 2688 if (!zone_watermark_ok(zone, 0, 2689 high_wmark_pages(zone), 0, 0)) { 2690 order = sc.order = 0; 2691 goto loop_again; 2692 } 2693 2694 /* If balanced, clear the congested flag */ 2695 zone_clear_flag(zone, ZONE_CONGESTED); 2696 } 2697 } 2698 2699 /* 2700 * Return the order we were reclaiming at so sleeping_prematurely() 2701 * makes a decision on the order we were last reclaiming at. However, 2702 * if another caller entered the allocator slow path while kswapd 2703 * was awake, order will remain at the higher level 2704 */ 2705 *classzone_idx = end_zone; 2706 return order; 2707} 2708 2709static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) 2710{ 2711 long remaining = 0; 2712 DEFINE_WAIT(wait); 2713 2714 if (freezing(current) || kthread_should_stop()) 2715 return; 2716 2717 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 2718 2719 /* Try to sleep for a short interval */ 2720 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) { 2721 remaining = schedule_timeout(HZ/10); 2722 finish_wait(&pgdat->kswapd_wait, &wait); 2723 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 2724 } 2725 2726 /* 2727 * After a short sleep, check if it was a premature sleep. If not, then 2728 * go fully to sleep until explicitly woken up. 2729 */ 2730 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) { 2731 trace_mm_vmscan_kswapd_sleep(pgdat->node_id); 2732 2733 /* 2734 * vmstat counters are not perfectly accurate and the estimated 2735 * value for counters such as NR_FREE_PAGES can deviate from the 2736 * true value by nr_online_cpus * threshold. To avoid the zone 2737 * watermarks being breached while under pressure, we reduce the 2738 * per-cpu vmstat threshold while kswapd is awake and restore 2739 * them before going back to sleep. 2740 */ 2741 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); 2742 schedule(); 2743 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); 2744 } else { 2745 if (remaining) 2746 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); 2747 else 2748 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); 2749 } 2750 finish_wait(&pgdat->kswapd_wait, &wait); 2751} 2752 2753/* 2754 * The background pageout daemon, started as a kernel thread 2755 * from the init process. 2756 * 2757 * This basically trickles out pages so that we have _some_ 2758 * free memory available even if there is no other activity 2759 * that frees anything up. This is needed for things like routing 2760 * etc, where we otherwise might have all activity going on in 2761 * asynchronous contexts that cannot page things out. 2762 * 2763 * If there are applications that are active memory-allocators 2764 * (most normal use), this basically shouldn't matter. 2765 */ 2766static int kswapd(void *p) 2767{ 2768 unsigned long order, new_order; 2769 int classzone_idx, new_classzone_idx; 2770 pg_data_t *pgdat = (pg_data_t*)p; 2771 struct task_struct *tsk = current; 2772 2773 struct reclaim_state reclaim_state = { 2774 .reclaimed_slab = 0, 2775 }; 2776 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); 2777 2778 lockdep_set_current_reclaim_state(GFP_KERNEL); 2779 2780 if (!cpumask_empty(cpumask)) 2781 set_cpus_allowed_ptr(tsk, cpumask); 2782 current->reclaim_state = &reclaim_state; 2783 2784 /* 2785 * Tell the memory management that we're a "memory allocator", 2786 * and that if we need more memory we should get access to it 2787 * regardless (see "__alloc_pages()"). "kswapd" should 2788 * never get caught in the normal page freeing logic. 2789 * 2790 * (Kswapd normally doesn't need memory anyway, but sometimes 2791 * you need a small amount of memory in order to be able to 2792 * page out something else, and this flag essentially protects 2793 * us from recursively trying to free more memory as we're 2794 * trying to free the first piece of memory in the first place). 2795 */ 2796 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; 2797 set_freezable(); 2798 2799 order = new_order = 0; 2800 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; 2801 for ( ; ; ) { 2802 int ret; 2803 2804 /* 2805 * If the last balance_pgdat was unsuccessful it's unlikely a 2806 * new request of a similar or harder type will succeed soon 2807 * so consider going to sleep on the basis we reclaimed at 2808 */ 2809 if (classzone_idx >= new_classzone_idx && order == new_order) { 2810 new_order = pgdat->kswapd_max_order; 2811 new_classzone_idx = pgdat->classzone_idx; 2812 pgdat->kswapd_max_order = 0; 2813 pgdat->classzone_idx = pgdat->nr_zones - 1; 2814 } 2815 2816 if (order < new_order || classzone_idx > new_classzone_idx) { 2817 /* 2818 * Don't sleep if someone wants a larger 'order' 2819 * allocation or has tigher zone constraints 2820 */ 2821 order = new_order; 2822 classzone_idx = new_classzone_idx; 2823 } else { 2824 kswapd_try_to_sleep(pgdat, order, classzone_idx); 2825 order = pgdat->kswapd_max_order; 2826 classzone_idx = pgdat->classzone_idx; 2827 pgdat->kswapd_max_order = 0; 2828 pgdat->classzone_idx = pgdat->nr_zones - 1; 2829 } 2830 2831 ret = try_to_freeze(); 2832 if (kthread_should_stop()) 2833 break; 2834 2835 /* 2836 * We can speed up thawing tasks if we don't call balance_pgdat 2837 * after returning from the refrigerator 2838 */ 2839 if (!ret) { 2840 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); 2841 order = balance_pgdat(pgdat, order, &classzone_idx); 2842 } 2843 } 2844 return 0; 2845} 2846 2847/* 2848 * A zone is low on free memory, so wake its kswapd task to service it. 2849 */ 2850void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) 2851{ 2852 pg_data_t *pgdat; 2853 2854 if (!populated_zone(zone)) 2855 return; 2856 2857 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 2858 return; 2859 pgdat = zone->zone_pgdat; 2860 if (pgdat->kswapd_max_order < order) { 2861 pgdat->kswapd_max_order = order; 2862 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); 2863 } 2864 if (!waitqueue_active(&pgdat->kswapd_wait)) 2865 return; 2866 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0)) 2867 return; 2868 2869 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); 2870 wake_up_interruptible(&pgdat->kswapd_wait); 2871} 2872 2873/* 2874 * The reclaimable count would be mostly accurate. 2875 * The less reclaimable pages may be 2876 * - mlocked pages, which will be moved to unevictable list when encountered 2877 * - mapped pages, which may require several travels to be reclaimed 2878 * - dirty pages, which is not "instantly" reclaimable 2879 */ 2880unsigned long global_reclaimable_pages(void) 2881{ 2882 int nr; 2883 2884 nr = global_page_state(NR_ACTIVE_FILE) + 2885 global_page_state(NR_INACTIVE_FILE); 2886 2887 if (nr_swap_pages > 0) 2888 nr += global_page_state(NR_ACTIVE_ANON) + 2889 global_page_state(NR_INACTIVE_ANON); 2890 2891 return nr; 2892} 2893 2894unsigned long zone_reclaimable_pages(struct zone *zone) 2895{ 2896 int nr; 2897 2898 nr = zone_page_state(zone, NR_ACTIVE_FILE) + 2899 zone_page_state(zone, NR_INACTIVE_FILE); 2900 2901 if (nr_swap_pages > 0) 2902 nr += zone_page_state(zone, NR_ACTIVE_ANON) + 2903 zone_page_state(zone, NR_INACTIVE_ANON); 2904 2905 return nr; 2906} 2907 2908#ifdef CONFIG_HIBERNATION 2909/* 2910 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of 2911 * freed pages. 2912 * 2913 * Rather than trying to age LRUs the aim is to preserve the overall 2914 * LRU order by reclaiming preferentially 2915 * inactive > active > active referenced > active mapped 2916 */ 2917unsigned long shrink_all_memory(unsigned long nr_to_reclaim) 2918{ 2919 struct reclaim_state reclaim_state; 2920 struct scan_control sc = { 2921 .gfp_mask = GFP_HIGHUSER_MOVABLE, 2922 .may_swap = 1, 2923 .may_unmap = 1, 2924 .may_writepage = 1, 2925 .nr_to_reclaim = nr_to_reclaim, 2926 .hibernation_mode = 1, 2927 .order = 0, 2928 }; 2929 struct shrink_control shrink = { 2930 .gfp_mask = sc.gfp_mask, 2931 }; 2932 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); 2933 struct task_struct *p = current; 2934 unsigned long nr_reclaimed; 2935 2936 p->flags |= PF_MEMALLOC; 2937 lockdep_set_current_reclaim_state(sc.gfp_mask); 2938 reclaim_state.reclaimed_slab = 0; 2939 p->reclaim_state = &reclaim_state; 2940 2941 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink); 2942 2943 p->reclaim_state = NULL; 2944 lockdep_clear_current_reclaim_state(); 2945 p->flags &= ~PF_MEMALLOC; 2946 2947 return nr_reclaimed; 2948} 2949#endif /* CONFIG_HIBERNATION */ 2950 2951/* It's optimal to keep kswapds on the same CPUs as their memory, but 2952 not required for correctness. So if the last cpu in a node goes 2953 away, we get changed to run anywhere: as the first one comes back, 2954 restore their cpu bindings. */ 2955static int __devinit cpu_callback(struct notifier_block *nfb, 2956 unsigned long action, void *hcpu) 2957{ 2958 int nid; 2959 2960 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { 2961 for_each_node_state(nid, N_HIGH_MEMORY) { 2962 pg_data_t *pgdat = NODE_DATA(nid); 2963 const struct cpumask *mask; 2964 2965 mask = cpumask_of_node(pgdat->node_id); 2966 2967 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) 2968 /* One of our CPUs online: restore mask */ 2969 set_cpus_allowed_ptr(pgdat->kswapd, mask); 2970 } 2971 } 2972 return NOTIFY_OK; 2973} 2974 2975/* 2976 * This kswapd start function will be called by init and node-hot-add. 2977 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. 2978 */ 2979int kswapd_run(int nid) 2980{ 2981 pg_data_t *pgdat = NODE_DATA(nid); 2982 int ret = 0; 2983 2984 if (pgdat->kswapd) 2985 return 0; 2986 2987 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); 2988 if (IS_ERR(pgdat->kswapd)) { 2989 /* failure at boot is fatal */ 2990 BUG_ON(system_state == SYSTEM_BOOTING); 2991 printk("Failed to start kswapd on node %d\n",nid); 2992 ret = -1; 2993 } 2994 return ret; 2995} 2996 2997/* 2998 * Called by memory hotplug when all memory in a node is offlined. 2999 */ 3000void kswapd_stop(int nid) 3001{ 3002 struct task_struct *kswapd = NODE_DATA(nid)->kswapd; 3003 3004 if (kswapd) 3005 kthread_stop(kswapd); 3006} 3007 3008static int __init kswapd_init(void) 3009{ 3010 int nid; 3011 3012 swap_setup(); 3013 for_each_node_state(nid, N_HIGH_MEMORY) 3014 kswapd_run(nid); 3015 hotcpu_notifier(cpu_callback, 0); 3016 return 0; 3017} 3018 3019module_init(kswapd_init) 3020 3021#ifdef CONFIG_NUMA 3022/* 3023 * Zone reclaim mode 3024 * 3025 * If non-zero call zone_reclaim when the number of free pages falls below 3026 * the watermarks. 3027 */ 3028int zone_reclaim_mode __read_mostly; 3029 3030#define RECLAIM_OFF 0 3031#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ 3032#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ 3033#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ 3034 3035/* 3036 * Priority for ZONE_RECLAIM. This determines the fraction of pages 3037 * of a node considered for each zone_reclaim. 4 scans 1/16th of 3038 * a zone. 3039 */ 3040#define ZONE_RECLAIM_PRIORITY 4 3041 3042/* 3043 * Percentage of pages in a zone that must be unmapped for zone_reclaim to 3044 * occur. 3045 */ 3046int sysctl_min_unmapped_ratio = 1; 3047 3048/* 3049 * If the number of slab pages in a zone grows beyond this percentage then 3050 * slab reclaim needs to occur. 3051 */ 3052int sysctl_min_slab_ratio = 5; 3053 3054static inline unsigned long zone_unmapped_file_pages(struct zone *zone) 3055{ 3056 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); 3057 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + 3058 zone_page_state(zone, NR_ACTIVE_FILE); 3059 3060 /* 3061 * It's possible for there to be more file mapped pages than 3062 * accounted for by the pages on the file LRU lists because 3063 * tmpfs pages accounted for as ANON can also be FILE_MAPPED 3064 */ 3065 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; 3066} 3067 3068/* Work out how many page cache pages we can reclaim in this reclaim_mode */ 3069static long zone_pagecache_reclaimable(struct zone *zone) 3070{ 3071 long nr_pagecache_reclaimable; 3072 long delta = 0; 3073 3074 /* 3075 * If RECLAIM_SWAP is set, then all file pages are considered 3076 * potentially reclaimable. Otherwise, we have to worry about 3077 * pages like swapcache and zone_unmapped_file_pages() provides 3078 * a better estimate 3079 */ 3080 if (zone_reclaim_mode & RECLAIM_SWAP) 3081 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); 3082 else 3083 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); 3084 3085 /* If we can't clean pages, remove dirty pages from consideration */ 3086 if (!(zone_reclaim_mode & RECLAIM_WRITE)) 3087 delta += zone_page_state(zone, NR_FILE_DIRTY); 3088 3089 /* Watch for any possible underflows due to delta */ 3090 if (unlikely(delta > nr_pagecache_reclaimable)) 3091 delta = nr_pagecache_reclaimable; 3092 3093 return nr_pagecache_reclaimable - delta; 3094} 3095 3096/* 3097 * Try to free up some pages from this zone through reclaim. 3098 */ 3099static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 3100{ 3101 /* Minimum pages needed in order to stay on node */ 3102 const unsigned long nr_pages = 1 << order; 3103 struct task_struct *p = current; 3104 struct reclaim_state reclaim_state; 3105 int priority; 3106 struct scan_control sc = { 3107 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), 3108 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), 3109 .may_swap = 1, 3110 .nr_to_reclaim = max_t(unsigned long, nr_pages, 3111 SWAP_CLUSTER_MAX), 3112 .gfp_mask = gfp_mask, 3113 .order = order, 3114 }; 3115 struct shrink_control shrink = { 3116 .gfp_mask = sc.gfp_mask, 3117 }; 3118 unsigned long nr_slab_pages0, nr_slab_pages1; 3119 3120 cond_resched(); 3121 /* 3122 * We need to be able to allocate from the reserves for RECLAIM_SWAP 3123 * and we also need to be able to write out pages for RECLAIM_WRITE 3124 * and RECLAIM_SWAP. 3125 */ 3126 p->flags |= PF_MEMALLOC | PF_SWAPWRITE; 3127 lockdep_set_current_reclaim_state(gfp_mask); 3128 reclaim_state.reclaimed_slab = 0; 3129 p->reclaim_state = &reclaim_state; 3130 3131 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { 3132 /* 3133 * Free memory by calling shrink zone with increasing 3134 * priorities until we have enough memory freed. 3135 */ 3136 priority = ZONE_RECLAIM_PRIORITY; 3137 do { 3138 shrink_zone(priority, zone, &sc); 3139 priority--; 3140 } while (priority >= 0 && sc.nr_reclaimed < nr_pages); 3141 } 3142 3143 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); 3144 if (nr_slab_pages0 > zone->min_slab_pages) { 3145 /* 3146 * shrink_slab() does not currently allow us to determine how 3147 * many pages were freed in this zone. So we take the current 3148 * number of slab pages and shake the slab until it is reduced 3149 * by the same nr_pages that we used for reclaiming unmapped 3150 * pages. 3151 * 3152 * Note that shrink_slab will free memory on all zones and may 3153 * take a long time. 3154 */ 3155 for (;;) { 3156 unsigned long lru_pages = zone_reclaimable_pages(zone); 3157 3158 /* No reclaimable slab or very low memory pressure */ 3159 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages)) 3160 break; 3161 3162 /* Freed enough memory */ 3163 nr_slab_pages1 = zone_page_state(zone, 3164 NR_SLAB_RECLAIMABLE); 3165 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0) 3166 break; 3167 } 3168 3169 /* 3170 * Update nr_reclaimed by the number of slab pages we 3171 * reclaimed from this zone. 3172 */ 3173 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE); 3174 if (nr_slab_pages1 < nr_slab_pages0) 3175 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1; 3176 } 3177 3178 p->reclaim_state = NULL; 3179 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); 3180 lockdep_clear_current_reclaim_state(); 3181 return sc.nr_reclaimed >= nr_pages; 3182} 3183 3184int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 3185{ 3186 int node_id; 3187 int ret; 3188 3189 /* 3190 * Zone reclaim reclaims unmapped file backed pages and 3191 * slab pages if we are over the defined limits. 3192 * 3193 * A small portion of unmapped file backed pages is needed for 3194 * file I/O otherwise pages read by file I/O will be immediately 3195 * thrown out if the zone is overallocated. So we do not reclaim 3196 * if less than a specified percentage of the zone is used by 3197 * unmapped file backed pages. 3198 */ 3199 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && 3200 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) 3201 return ZONE_RECLAIM_FULL; 3202 3203 if (zone->all_unreclaimable) 3204 return ZONE_RECLAIM_FULL; 3205 3206 /* 3207 * Do not scan if the allocation should not be delayed. 3208 */ 3209 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) 3210 return ZONE_RECLAIM_NOSCAN; 3211 3212 /* 3213 * Only run zone reclaim on the local zone or on zones that do not 3214 * have associated processors. This will favor the local processor 3215 * over remote processors and spread off node memory allocations 3216 * as wide as possible. 3217 */ 3218 node_id = zone_to_nid(zone); 3219 if (node_state(node_id, N_CPU) && node_id != numa_node_id()) 3220 return ZONE_RECLAIM_NOSCAN; 3221 3222 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) 3223 return ZONE_RECLAIM_NOSCAN; 3224 3225 ret = __zone_reclaim(zone, gfp_mask, order); 3226 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); 3227 3228 if (!ret) 3229 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); 3230 3231 return ret; 3232} 3233#endif 3234 3235/* 3236 * page_evictable - test whether a page is evictable 3237 * @page: the page to test 3238 * @vma: the VMA in which the page is or will be mapped, may be NULL 3239 * 3240 * Test whether page is evictable--i.e., should be placed on active/inactive 3241 * lists vs unevictable list. The vma argument is !NULL when called from the 3242 * fault path to determine how to instantate a new page. 3243 * 3244 * Reasons page might not be evictable: 3245 * (1) page's mapping marked unevictable 3246 * (2) page is part of an mlocked VMA 3247 * 3248 */ 3249int page_evictable(struct page *page, struct vm_area_struct *vma) 3250{ 3251 3252 if (mapping_unevictable(page_mapping(page))) 3253 return 0; 3254 3255 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page))) 3256 return 0; 3257 3258 return 1; 3259} 3260 3261/** 3262 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list 3263 * @page: page to check evictability and move to appropriate lru list 3264 * @zone: zone page is in 3265 * 3266 * Checks a page for evictability and moves the page to the appropriate 3267 * zone lru list. 3268 * 3269 * Restrictions: zone->lru_lock must be held, page must be on LRU and must 3270 * have PageUnevictable set. 3271 */ 3272static void check_move_unevictable_page(struct page *page, struct zone *zone) 3273{ 3274 VM_BUG_ON(PageActive(page)); 3275 3276retry: 3277 ClearPageUnevictable(page); 3278 if (page_evictable(page, NULL)) { 3279 enum lru_list l = page_lru_base_type(page); 3280 3281 __dec_zone_state(zone, NR_UNEVICTABLE); 3282 list_move(&page->lru, &zone->lru[l].list); 3283 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l); 3284 __inc_zone_state(zone, NR_INACTIVE_ANON + l); 3285 __count_vm_event(UNEVICTABLE_PGRESCUED); 3286 } else { 3287 /* 3288 * rotate unevictable list 3289 */ 3290 SetPageUnevictable(page); 3291 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list); 3292 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE); 3293 if (page_evictable(page, NULL)) 3294 goto retry; 3295 } 3296} 3297 3298/** 3299 * scan_mapping_unevictable_pages - scan an address space for evictable pages 3300 * @mapping: struct address_space to scan for evictable pages 3301 * 3302 * Scan all pages in mapping. Check unevictable pages for 3303 * evictability and move them to the appropriate zone lru list. 3304 */ 3305void scan_mapping_unevictable_pages(struct address_space *mapping) 3306{ 3307 pgoff_t next = 0; 3308 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >> 3309 PAGE_CACHE_SHIFT; 3310 struct zone *zone; 3311 struct pagevec pvec; 3312 3313 if (mapping->nrpages == 0) 3314 return; 3315 3316 pagevec_init(&pvec, 0); 3317 while (next < end && 3318 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) { 3319 int i; 3320 int pg_scanned = 0; 3321 3322 zone = NULL; 3323 3324 for (i = 0; i < pagevec_count(&pvec); i++) { 3325 struct page *page = pvec.pages[i]; 3326 pgoff_t page_index = page->index; 3327 struct zone *pagezone = page_zone(page); 3328 3329 pg_scanned++; 3330 if (page_index > next) 3331 next = page_index; 3332 next++; 3333 3334 if (pagezone != zone) { 3335 if (zone) 3336 spin_unlock_irq(&zone->lru_lock); 3337 zone = pagezone; 3338 spin_lock_irq(&zone->lru_lock); 3339 } 3340 3341 if (PageLRU(page) && PageUnevictable(page)) 3342 check_move_unevictable_page(page, zone); 3343 } 3344 if (zone) 3345 spin_unlock_irq(&zone->lru_lock); 3346 pagevec_release(&pvec); 3347 3348 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned); 3349 } 3350 3351} 3352 3353/** 3354 * scan_zone_unevictable_pages - check unevictable list for evictable pages 3355 * @zone - zone of which to scan the unevictable list 3356 * 3357 * Scan @zone's unevictable LRU lists to check for pages that have become 3358 * evictable. Move those that have to @zone's inactive list where they 3359 * become candidates for reclaim, unless shrink_inactive_zone() decides 3360 * to reactivate them. Pages that are still unevictable are rotated 3361 * back onto @zone's unevictable list. 3362 */ 3363#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */ 3364static void scan_zone_unevictable_pages(struct zone *zone) 3365{ 3366 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list; 3367 unsigned long scan; 3368 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE); 3369 3370 while (nr_to_scan > 0) { 3371 unsigned long batch_size = min(nr_to_scan, 3372 SCAN_UNEVICTABLE_BATCH_SIZE); 3373 3374 spin_lock_irq(&zone->lru_lock); 3375 for (scan = 0; scan < batch_size; scan++) { 3376 struct page *page = lru_to_page(l_unevictable); 3377 3378 if (!trylock_page(page)) 3379 continue; 3380 3381 prefetchw_prev_lru_page(page, l_unevictable, flags); 3382 3383 if (likely(PageLRU(page) && PageUnevictable(page))) 3384 check_move_unevictable_page(page, zone); 3385 3386 unlock_page(page); 3387 } 3388 spin_unlock_irq(&zone->lru_lock); 3389 3390 nr_to_scan -= batch_size; 3391 } 3392} 3393 3394 3395/** 3396 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages 3397 * 3398 * A really big hammer: scan all zones' unevictable LRU lists to check for 3399 * pages that have become evictable. Move those back to the zones' 3400 * inactive list where they become candidates for reclaim. 3401 * This occurs when, e.g., we have unswappable pages on the unevictable lists, 3402 * and we add swap to the system. As such, it runs in the context of a task 3403 * that has possibly/probably made some previously unevictable pages 3404 * evictable. 3405 */ 3406static void scan_all_zones_unevictable_pages(void) 3407{ 3408 struct zone *zone; 3409 3410 for_each_zone(zone) { 3411 scan_zone_unevictable_pages(zone); 3412 } 3413} 3414 3415/* 3416 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of 3417 * all nodes' unevictable lists for evictable pages 3418 */ 3419unsigned long scan_unevictable_pages; 3420 3421int scan_unevictable_handler(struct ctl_table *table, int write, 3422 void __user *buffer, 3423 size_t *length, loff_t *ppos) 3424{ 3425 proc_doulongvec_minmax(table, write, buffer, length, ppos); 3426 3427 if (write && *(unsigned long *)table->data) 3428 scan_all_zones_unevictable_pages(); 3429 3430 scan_unevictable_pages = 0; 3431 return 0; 3432} 3433 3434#ifdef CONFIG_NUMA 3435/* 3436 * per node 'scan_unevictable_pages' attribute. On demand re-scan of 3437 * a specified node's per zone unevictable lists for evictable pages. 3438 */ 3439 3440static ssize_t read_scan_unevictable_node(struct sys_device *dev, 3441 struct sysdev_attribute *attr, 3442 char *buf) 3443{ 3444 return sprintf(buf, "0\n"); /* always zero; should fit... */ 3445} 3446 3447static ssize_t write_scan_unevictable_node(struct sys_device *dev, 3448 struct sysdev_attribute *attr, 3449 const char *buf, size_t count) 3450{ 3451 struct zone *node_zones = NODE_DATA(dev->id)->node_zones; 3452 struct zone *zone; 3453 unsigned long res; 3454 unsigned long req = strict_strtoul(buf, 10, &res); 3455 3456 if (!req) 3457 return 1; /* zero is no-op */ 3458 3459 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) { 3460 if (!populated_zone(zone)) 3461 continue; 3462 scan_zone_unevictable_pages(zone); 3463 } 3464 return 1; 3465} 3466 3467 3468static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, 3469 read_scan_unevictable_node, 3470 write_scan_unevictable_node); 3471 3472int scan_unevictable_register_node(struct node *node) 3473{ 3474 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages); 3475} 3476 3477void scan_unevictable_unregister_node(struct node *node) 3478{ 3479 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages); 3480} 3481#endif 3482