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#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 15 16#include <linux/mm.h> 17#include <linux/module.h> 18#include <linux/gfp.h> 19#include <linux/kernel_stat.h> 20#include <linux/swap.h> 21#include <linux/pagemap.h> 22#include <linux/init.h> 23#include <linux/highmem.h> 24#include <linux/vmpressure.h> 25#include <linux/vmstat.h> 26#include <linux/file.h> 27#include <linux/writeback.h> 28#include <linux/blkdev.h> 29#include <linux/buffer_head.h> /* for try_to_release_page(), 30 buffer_heads_over_limit */ 31#include <linux/mm_inline.h> 32#include <linux/backing-dev.h> 33#include <linux/rmap.h> 34#include <linux/topology.h> 35#include <linux/cpu.h> 36#include <linux/cpuset.h> 37#include <linux/compaction.h> 38#include <linux/notifier.h> 39#include <linux/rwsem.h> 40#include <linux/delay.h> 41#include <linux/kthread.h> 42#include <linux/freezer.h> 43#include <linux/memcontrol.h> 44#include <linux/delayacct.h> 45#include <linux/sysctl.h> 46#include <linux/oom.h> 47#include <linux/prefetch.h> 48#include <linux/printk.h> 49 50#include <asm/tlbflush.h> 51#include <asm/div64.h> 52 53#include <linux/swapops.h> 54#include <linux/balloon_compaction.h> 55 56#include "internal.h" 57 58#define CREATE_TRACE_POINTS 59#include <trace/events/vmscan.h> 60 61struct scan_control { 62 /* How many pages shrink_list() should reclaim */ 63 unsigned long nr_to_reclaim; 64 65 /* This context's GFP mask */ 66 gfp_t gfp_mask; 67 68 /* Allocation order */ 69 int order; 70 71 /* 72 * Nodemask of nodes allowed by the caller. If NULL, all nodes 73 * are scanned. 74 */ 75 nodemask_t *nodemask; 76 77 /* 78 * The memory cgroup that hit its limit and as a result is the 79 * primary target of this reclaim invocation. 80 */ 81 struct mem_cgroup *target_mem_cgroup; 82 83 /* Scan (total_size >> priority) pages at once */ 84 int priority; 85 86 unsigned int may_writepage:1; 87 88 /* Can mapped pages be reclaimed? */ 89 unsigned int may_unmap:1; 90 91 /* Can pages be swapped as part of reclaim? */ 92 unsigned int may_swap:1; 93 94 /* Can cgroups be reclaimed below their normal consumption range? */ 95 unsigned int may_thrash:1; 96 97 unsigned int hibernation_mode:1; 98 99 /* One of the zones is ready for compaction */ 100 unsigned int compaction_ready:1; 101 102 /* Incremented by the number of inactive pages that were scanned */ 103 unsigned long nr_scanned; 104 105 /* Number of pages freed so far during a call to shrink_zones() */ 106 unsigned long nr_reclaimed; 107}; 108 109#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) 110 111#ifdef ARCH_HAS_PREFETCH 112#define prefetch_prev_lru_page(_page, _base, _field) \ 113 do { \ 114 if ((_page)->lru.prev != _base) { \ 115 struct page *prev; \ 116 \ 117 prev = lru_to_page(&(_page->lru)); \ 118 prefetch(&prev->_field); \ 119 } \ 120 } while (0) 121#else 122#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) 123#endif 124 125#ifdef ARCH_HAS_PREFETCHW 126#define prefetchw_prev_lru_page(_page, _base, _field) \ 127 do { \ 128 if ((_page)->lru.prev != _base) { \ 129 struct page *prev; \ 130 \ 131 prev = lru_to_page(&(_page->lru)); \ 132 prefetchw(&prev->_field); \ 133 } \ 134 } while (0) 135#else 136#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) 137#endif 138 139/* 140 * From 0 .. 100. Higher means more swappy. 141 */ 142int vm_swappiness = 60; 143/* 144 * The total number of pages which are beyond the high watermark within all 145 * zones. 146 */ 147unsigned long vm_total_pages; 148 149static LIST_HEAD(shrinker_list); 150static DECLARE_RWSEM(shrinker_rwsem); 151 152#ifdef CONFIG_MEMCG 153static bool global_reclaim(struct scan_control *sc) 154{ 155 return !sc->target_mem_cgroup; 156} 157#else 158static bool global_reclaim(struct scan_control *sc) 159{ 160 return true; 161} 162#endif 163 164static unsigned long zone_reclaimable_pages(struct zone *zone) 165{ 166 int nr; 167 168 nr = zone_page_state(zone, NR_ACTIVE_FILE) + 169 zone_page_state(zone, NR_INACTIVE_FILE); 170 171 if (get_nr_swap_pages() > 0) 172 nr += zone_page_state(zone, NR_ACTIVE_ANON) + 173 zone_page_state(zone, NR_INACTIVE_ANON); 174 175 return nr; 176} 177 178bool zone_reclaimable(struct zone *zone) 179{ 180 return zone_page_state(zone, NR_PAGES_SCANNED) < 181 zone_reclaimable_pages(zone) * 6; 182} 183 184static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru) 185{ 186 if (!mem_cgroup_disabled()) 187 return mem_cgroup_get_lru_size(lruvec, lru); 188 189 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru); 190} 191 192/* 193 * Add a shrinker callback to be called from the vm. 194 */ 195int register_shrinker(struct shrinker *shrinker) 196{ 197 size_t size = sizeof(*shrinker->nr_deferred); 198 199 /* 200 * If we only have one possible node in the system anyway, save 201 * ourselves the trouble and disable NUMA aware behavior. This way we 202 * will save memory and some small loop time later. 203 */ 204 if (nr_node_ids == 1) 205 shrinker->flags &= ~SHRINKER_NUMA_AWARE; 206 207 if (shrinker->flags & SHRINKER_NUMA_AWARE) 208 size *= nr_node_ids; 209 210 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); 211 if (!shrinker->nr_deferred) 212 return -ENOMEM; 213 214 down_write(&shrinker_rwsem); 215 list_add_tail(&shrinker->list, &shrinker_list); 216 up_write(&shrinker_rwsem); 217 return 0; 218} 219EXPORT_SYMBOL(register_shrinker); 220 221/* 222 * Remove one 223 */ 224void unregister_shrinker(struct shrinker *shrinker) 225{ 226 down_write(&shrinker_rwsem); 227 list_del(&shrinker->list); 228 up_write(&shrinker_rwsem); 229 kfree(shrinker->nr_deferred); 230} 231EXPORT_SYMBOL(unregister_shrinker); 232 233#define SHRINK_BATCH 128 234 235static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, 236 struct shrinker *shrinker, 237 unsigned long nr_scanned, 238 unsigned long nr_eligible) 239{ 240 unsigned long freed = 0; 241 unsigned long long delta; 242 long total_scan; 243 long freeable; 244 long nr; 245 long new_nr; 246 int nid = shrinkctl->nid; 247 long batch_size = shrinker->batch ? shrinker->batch 248 : SHRINK_BATCH; 249 250 freeable = shrinker->count_objects(shrinker, shrinkctl); 251 if (freeable == 0) 252 return 0; 253 254 /* 255 * copy the current shrinker scan count into a local variable 256 * and zero it so that other concurrent shrinker invocations 257 * don't also do this scanning work. 258 */ 259 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); 260 261 total_scan = nr; 262 delta = (4 * nr_scanned) / shrinker->seeks; 263 delta *= freeable; 264 do_div(delta, nr_eligible + 1); 265 total_scan += delta; 266 if (total_scan < 0) { 267 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", 268 shrinker->scan_objects, total_scan); 269 total_scan = freeable; 270 } 271 272 /* 273 * We need to avoid excessive windup on filesystem shrinkers 274 * due to large numbers of GFP_NOFS allocations causing the 275 * shrinkers to return -1 all the time. This results in a large 276 * nr being built up so when a shrink that can do some work 277 * comes along it empties the entire cache due to nr >>> 278 * freeable. This is bad for sustaining a working set in 279 * memory. 280 * 281 * Hence only allow the shrinker to scan the entire cache when 282 * a large delta change is calculated directly. 283 */ 284 if (delta < freeable / 4) 285 total_scan = min(total_scan, freeable / 2); 286 287 /* 288 * Avoid risking looping forever due to too large nr value: 289 * never try to free more than twice the estimate number of 290 * freeable entries. 291 */ 292 if (total_scan > freeable * 2) 293 total_scan = freeable * 2; 294 295 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, 296 nr_scanned, nr_eligible, 297 freeable, delta, total_scan); 298 299 /* 300 * Normally, we should not scan less than batch_size objects in one 301 * pass to avoid too frequent shrinker calls, but if the slab has less 302 * than batch_size objects in total and we are really tight on memory, 303 * we will try to reclaim all available objects, otherwise we can end 304 * up failing allocations although there are plenty of reclaimable 305 * objects spread over several slabs with usage less than the 306 * batch_size. 307 * 308 * We detect the "tight on memory" situations by looking at the total 309 * number of objects we want to scan (total_scan). If it is greater 310 * than the total number of objects on slab (freeable), we must be 311 * scanning at high prio and therefore should try to reclaim as much as 312 * possible. 313 */ 314 while (total_scan >= batch_size || 315 total_scan >= freeable) { 316 unsigned long ret; 317 unsigned long nr_to_scan = min(batch_size, total_scan); 318 319 shrinkctl->nr_to_scan = nr_to_scan; 320 ret = shrinker->scan_objects(shrinker, shrinkctl); 321 if (ret == SHRINK_STOP) 322 break; 323 freed += ret; 324 325 count_vm_events(SLABS_SCANNED, nr_to_scan); 326 total_scan -= nr_to_scan; 327 328 cond_resched(); 329 } 330 331 /* 332 * move the unused scan count back into the shrinker in a 333 * manner that handles concurrent updates. If we exhausted the 334 * scan, there is no need to do an update. 335 */ 336 if (total_scan > 0) 337 new_nr = atomic_long_add_return(total_scan, 338 &shrinker->nr_deferred[nid]); 339 else 340 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); 341 342 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); 343 return freed; 344} 345 346/** 347 * shrink_slab - shrink slab caches 348 * @gfp_mask: allocation context 349 * @nid: node whose slab caches to target 350 * @memcg: memory cgroup whose slab caches to target 351 * @nr_scanned: pressure numerator 352 * @nr_eligible: pressure denominator 353 * 354 * Call the shrink functions to age shrinkable caches. 355 * 356 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, 357 * unaware shrinkers will receive a node id of 0 instead. 358 * 359 * @memcg specifies the memory cgroup to target. If it is not NULL, 360 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan 361 * objects from the memory cgroup specified. Otherwise all shrinkers 362 * are called, and memcg aware shrinkers are supposed to scan the 363 * global list then. 364 * 365 * @nr_scanned and @nr_eligible form a ratio that indicate how much of 366 * the available objects should be scanned. Page reclaim for example 367 * passes the number of pages scanned and the number of pages on the 368 * LRU lists that it considered on @nid, plus a bias in @nr_scanned 369 * when it encountered mapped pages. The ratio is further biased by 370 * the ->seeks setting of the shrink function, which indicates the 371 * cost to recreate an object relative to that of an LRU page. 372 * 373 * Returns the number of reclaimed slab objects. 374 */ 375static unsigned long shrink_slab(gfp_t gfp_mask, int nid, 376 struct mem_cgroup *memcg, 377 unsigned long nr_scanned, 378 unsigned long nr_eligible) 379{ 380 struct shrinker *shrinker; 381 unsigned long freed = 0; 382 383 if (memcg && !memcg_kmem_is_active(memcg)) 384 return 0; 385 386 if (nr_scanned == 0) 387 nr_scanned = SWAP_CLUSTER_MAX; 388 389 if (!down_read_trylock(&shrinker_rwsem)) { 390 /* 391 * If we would return 0, our callers would understand that we 392 * have nothing else to shrink and give up trying. By returning 393 * 1 we keep it going and assume we'll be able to shrink next 394 * time. 395 */ 396 freed = 1; 397 goto out; 398 } 399 400 list_for_each_entry(shrinker, &shrinker_list, list) { 401 struct shrink_control sc = { 402 .gfp_mask = gfp_mask, 403 .nid = nid, 404 .memcg = memcg, 405 }; 406 407 if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE)) 408 continue; 409 410 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) 411 sc.nid = 0; 412 413 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible); 414 } 415 416 up_read(&shrinker_rwsem); 417out: 418 cond_resched(); 419 return freed; 420} 421 422void drop_slab_node(int nid) 423{ 424 unsigned long freed; 425 426 do { 427 struct mem_cgroup *memcg = NULL; 428 429 freed = 0; 430 do { 431 freed += shrink_slab(GFP_KERNEL, nid, memcg, 432 1000, 1000); 433 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); 434 } while (freed > 10); 435} 436 437void drop_slab(void) 438{ 439 int nid; 440 441 for_each_online_node(nid) 442 drop_slab_node(nid); 443} 444 445static inline int is_page_cache_freeable(struct page *page) 446{ 447 /* 448 * A freeable page cache page is referenced only by the caller 449 * that isolated the page, the page cache radix tree and 450 * optional buffer heads at page->private. 451 */ 452 return page_count(page) - page_has_private(page) == 2; 453} 454 455static int may_write_to_queue(struct backing_dev_info *bdi, 456 struct scan_control *sc) 457{ 458 if (current->flags & PF_SWAPWRITE) 459 return 1; 460 if (!bdi_write_congested(bdi)) 461 return 1; 462 if (bdi == current->backing_dev_info) 463 return 1; 464 return 0; 465} 466 467/* 468 * We detected a synchronous write error writing a page out. Probably 469 * -ENOSPC. We need to propagate that into the address_space for a subsequent 470 * fsync(), msync() or close(). 471 * 472 * The tricky part is that after writepage we cannot touch the mapping: nothing 473 * prevents it from being freed up. But we have a ref on the page and once 474 * that page is locked, the mapping is pinned. 475 * 476 * We're allowed to run sleeping lock_page() here because we know the caller has 477 * __GFP_FS. 478 */ 479static void handle_write_error(struct address_space *mapping, 480 struct page *page, int error) 481{ 482 lock_page(page); 483 if (page_mapping(page) == mapping) 484 mapping_set_error(mapping, error); 485 unlock_page(page); 486} 487 488/* possible outcome of pageout() */ 489typedef enum { 490 /* failed to write page out, page is locked */ 491 PAGE_KEEP, 492 /* move page to the active list, page is locked */ 493 PAGE_ACTIVATE, 494 /* page has been sent to the disk successfully, page is unlocked */ 495 PAGE_SUCCESS, 496 /* page is clean and locked */ 497 PAGE_CLEAN, 498} pageout_t; 499 500/* 501 * pageout is called by shrink_page_list() for each dirty page. 502 * Calls ->writepage(). 503 */ 504static pageout_t pageout(struct page *page, struct address_space *mapping, 505 struct scan_control *sc) 506{ 507 /* 508 * If the page is dirty, only perform writeback if that write 509 * will be non-blocking. To prevent this allocation from being 510 * stalled by pagecache activity. But note that there may be 511 * stalls if we need to run get_block(). We could test 512 * PagePrivate for that. 513 * 514 * If this process is currently in __generic_file_write_iter() against 515 * this page's queue, we can perform writeback even if that 516 * will block. 517 * 518 * If the page is swapcache, write it back even if that would 519 * block, for some throttling. This happens by accident, because 520 * swap_backing_dev_info is bust: it doesn't reflect the 521 * congestion state of the swapdevs. Easy to fix, if needed. 522 */ 523 if (!is_page_cache_freeable(page)) 524 return PAGE_KEEP; 525 if (!mapping) { 526 /* 527 * Some data journaling orphaned pages can have 528 * page->mapping == NULL while being dirty with clean buffers. 529 */ 530 if (page_has_private(page)) { 531 if (try_to_free_buffers(page)) { 532 ClearPageDirty(page); 533 pr_info("%s: orphaned page\n", __func__); 534 return PAGE_CLEAN; 535 } 536 } 537 return PAGE_KEEP; 538 } 539 if (mapping->a_ops->writepage == NULL) 540 return PAGE_ACTIVATE; 541 if (!may_write_to_queue(inode_to_bdi(mapping->host), sc)) 542 return PAGE_KEEP; 543 544 if (clear_page_dirty_for_io(page)) { 545 int res; 546 struct writeback_control wbc = { 547 .sync_mode = WB_SYNC_NONE, 548 .nr_to_write = SWAP_CLUSTER_MAX, 549 .range_start = 0, 550 .range_end = LLONG_MAX, 551 .for_reclaim = 1, 552 }; 553 554 SetPageReclaim(page); 555 res = mapping->a_ops->writepage(page, &wbc); 556 if (res < 0) 557 handle_write_error(mapping, page, res); 558 if (res == AOP_WRITEPAGE_ACTIVATE) { 559 ClearPageReclaim(page); 560 return PAGE_ACTIVATE; 561 } 562 563 if (!PageWriteback(page)) { 564 /* synchronous write or broken a_ops? */ 565 ClearPageReclaim(page); 566 } 567 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page)); 568 inc_zone_page_state(page, NR_VMSCAN_WRITE); 569 return PAGE_SUCCESS; 570 } 571 572 return PAGE_CLEAN; 573} 574 575/* 576 * Same as remove_mapping, but if the page is removed from the mapping, it 577 * gets returned with a refcount of 0. 578 */ 579static int __remove_mapping(struct address_space *mapping, struct page *page, 580 bool reclaimed) 581{ 582 BUG_ON(!PageLocked(page)); 583 BUG_ON(mapping != page_mapping(page)); 584 585 spin_lock_irq(&mapping->tree_lock); 586 /* 587 * The non racy check for a busy page. 588 * 589 * Must be careful with the order of the tests. When someone has 590 * a ref to the page, it may be possible that they dirty it then 591 * drop the reference. So if PageDirty is tested before page_count 592 * here, then the following race may occur: 593 * 594 * get_user_pages(&page); 595 * [user mapping goes away] 596 * write_to(page); 597 * !PageDirty(page) [good] 598 * SetPageDirty(page); 599 * put_page(page); 600 * !page_count(page) [good, discard it] 601 * 602 * [oops, our write_to data is lost] 603 * 604 * Reversing the order of the tests ensures such a situation cannot 605 * escape unnoticed. The smp_rmb is needed to ensure the page->flags 606 * load is not satisfied before that of page->_count. 607 * 608 * Note that if SetPageDirty is always performed via set_page_dirty, 609 * and thus under tree_lock, then this ordering is not required. 610 */ 611 if (!page_freeze_refs(page, 2)) 612 goto cannot_free; 613 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ 614 if (unlikely(PageDirty(page))) { 615 page_unfreeze_refs(page, 2); 616 goto cannot_free; 617 } 618 619 if (PageSwapCache(page)) { 620 swp_entry_t swap = { .val = page_private(page) }; 621 mem_cgroup_swapout(page, swap); 622 __delete_from_swap_cache(page); 623 spin_unlock_irq(&mapping->tree_lock); 624 swapcache_free(swap); 625 } else { 626 void (*freepage)(struct page *); 627 void *shadow = NULL; 628 629 freepage = mapping->a_ops->freepage; 630 /* 631 * Remember a shadow entry for reclaimed file cache in 632 * order to detect refaults, thus thrashing, later on. 633 * 634 * But don't store shadows in an address space that is 635 * already exiting. This is not just an optizimation, 636 * inode reclaim needs to empty out the radix tree or 637 * the nodes are lost. Don't plant shadows behind its 638 * back. 639 */ 640 if (reclaimed && page_is_file_cache(page) && 641 !mapping_exiting(mapping)) 642 shadow = workingset_eviction(mapping, page); 643 __delete_from_page_cache(page, shadow); 644 spin_unlock_irq(&mapping->tree_lock); 645 646 if (freepage != NULL) 647 freepage(page); 648 } 649 650 return 1; 651 652cannot_free: 653 spin_unlock_irq(&mapping->tree_lock); 654 return 0; 655} 656 657/* 658 * Attempt to detach a locked page from its ->mapping. If it is dirty or if 659 * someone else has a ref on the page, abort and return 0. If it was 660 * successfully detached, return 1. Assumes the caller has a single ref on 661 * this page. 662 */ 663int remove_mapping(struct address_space *mapping, struct page *page) 664{ 665 if (__remove_mapping(mapping, page, false)) { 666 /* 667 * Unfreezing the refcount with 1 rather than 2 effectively 668 * drops the pagecache ref for us without requiring another 669 * atomic operation. 670 */ 671 page_unfreeze_refs(page, 1); 672 return 1; 673 } 674 return 0; 675} 676 677/** 678 * putback_lru_page - put previously isolated page onto appropriate LRU list 679 * @page: page to be put back to appropriate lru list 680 * 681 * Add previously isolated @page to appropriate LRU list. 682 * Page may still be unevictable for other reasons. 683 * 684 * lru_lock must not be held, interrupts must be enabled. 685 */ 686void putback_lru_page(struct page *page) 687{ 688 bool is_unevictable; 689 int was_unevictable = PageUnevictable(page); 690 691 VM_BUG_ON_PAGE(PageLRU(page), page); 692 693redo: 694 ClearPageUnevictable(page); 695 696 if (page_evictable(page)) { 697 /* 698 * For evictable pages, we can use the cache. 699 * In event of a race, worst case is we end up with an 700 * unevictable page on [in]active list. 701 * We know how to handle that. 702 */ 703 is_unevictable = false; 704 lru_cache_add(page); 705 } else { 706 /* 707 * Put unevictable pages directly on zone's unevictable 708 * list. 709 */ 710 is_unevictable = true; 711 add_page_to_unevictable_list(page); 712 /* 713 * When racing with an mlock or AS_UNEVICTABLE clearing 714 * (page is unlocked) make sure that if the other thread 715 * does not observe our setting of PG_lru and fails 716 * isolation/check_move_unevictable_pages, 717 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move 718 * the page back to the evictable list. 719 * 720 * The other side is TestClearPageMlocked() or shmem_lock(). 721 */ 722 smp_mb(); 723 } 724 725 /* 726 * page's status can change while we move it among lru. If an evictable 727 * page is on unevictable list, it never be freed. To avoid that, 728 * check after we added it to the list, again. 729 */ 730 if (is_unevictable && page_evictable(page)) { 731 if (!isolate_lru_page(page)) { 732 put_page(page); 733 goto redo; 734 } 735 /* This means someone else dropped this page from LRU 736 * So, it will be freed or putback to LRU again. There is 737 * nothing to do here. 738 */ 739 } 740 741 if (was_unevictable && !is_unevictable) 742 count_vm_event(UNEVICTABLE_PGRESCUED); 743 else if (!was_unevictable && is_unevictable) 744 count_vm_event(UNEVICTABLE_PGCULLED); 745 746 put_page(page); /* drop ref from isolate */ 747} 748 749enum page_references { 750 PAGEREF_RECLAIM, 751 PAGEREF_RECLAIM_CLEAN, 752 PAGEREF_KEEP, 753 PAGEREF_ACTIVATE, 754}; 755 756static enum page_references page_check_references(struct page *page, 757 struct scan_control *sc) 758{ 759 int referenced_ptes, referenced_page; 760 unsigned long vm_flags; 761 762 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, 763 &vm_flags); 764 referenced_page = TestClearPageReferenced(page); 765 766 /* 767 * Mlock lost the isolation race with us. Let try_to_unmap() 768 * move the page to the unevictable list. 769 */ 770 if (vm_flags & VM_LOCKED) 771 return PAGEREF_RECLAIM; 772 773 if (referenced_ptes) { 774 if (PageSwapBacked(page)) 775 return PAGEREF_ACTIVATE; 776 /* 777 * All mapped pages start out with page table 778 * references from the instantiating fault, so we need 779 * to look twice if a mapped file page is used more 780 * than once. 781 * 782 * Mark it and spare it for another trip around the 783 * inactive list. Another page table reference will 784 * lead to its activation. 785 * 786 * Note: the mark is set for activated pages as well 787 * so that recently deactivated but used pages are 788 * quickly recovered. 789 */ 790 SetPageReferenced(page); 791 792 if (referenced_page || referenced_ptes > 1) 793 return PAGEREF_ACTIVATE; 794 795 /* 796 * Activate file-backed executable pages after first usage. 797 */ 798 if (vm_flags & VM_EXEC) 799 return PAGEREF_ACTIVATE; 800 801 return PAGEREF_KEEP; 802 } 803 804 /* Reclaim if clean, defer dirty pages to writeback */ 805 if (referenced_page && !PageSwapBacked(page)) 806 return PAGEREF_RECLAIM_CLEAN; 807 808 return PAGEREF_RECLAIM; 809} 810 811/* Check if a page is dirty or under writeback */ 812static void page_check_dirty_writeback(struct page *page, 813 bool *dirty, bool *writeback) 814{ 815 struct address_space *mapping; 816 817 /* 818 * Anonymous pages are not handled by flushers and must be written 819 * from reclaim context. Do not stall reclaim based on them 820 */ 821 if (!page_is_file_cache(page)) { 822 *dirty = false; 823 *writeback = false; 824 return; 825 } 826 827 /* By default assume that the page flags are accurate */ 828 *dirty = PageDirty(page); 829 *writeback = PageWriteback(page); 830 831 /* Verify dirty/writeback state if the filesystem supports it */ 832 if (!page_has_private(page)) 833 return; 834 835 mapping = page_mapping(page); 836 if (mapping && mapping->a_ops->is_dirty_writeback) 837 mapping->a_ops->is_dirty_writeback(page, dirty, writeback); 838} 839 840/* 841 * shrink_page_list() returns the number of reclaimed pages 842 */ 843static unsigned long shrink_page_list(struct list_head *page_list, 844 struct zone *zone, 845 struct scan_control *sc, 846 enum ttu_flags ttu_flags, 847 unsigned long *ret_nr_dirty, 848 unsigned long *ret_nr_unqueued_dirty, 849 unsigned long *ret_nr_congested, 850 unsigned long *ret_nr_writeback, 851 unsigned long *ret_nr_immediate, 852 bool force_reclaim) 853{ 854 LIST_HEAD(ret_pages); 855 LIST_HEAD(free_pages); 856 int pgactivate = 0; 857 unsigned long nr_unqueued_dirty = 0; 858 unsigned long nr_dirty = 0; 859 unsigned long nr_congested = 0; 860 unsigned long nr_reclaimed = 0; 861 unsigned long nr_writeback = 0; 862 unsigned long nr_immediate = 0; 863 864 cond_resched(); 865 866 while (!list_empty(page_list)) { 867 struct address_space *mapping; 868 struct page *page; 869 int may_enter_fs; 870 enum page_references references = PAGEREF_RECLAIM_CLEAN; 871 bool dirty, writeback; 872 873 cond_resched(); 874 875 page = lru_to_page(page_list); 876 list_del(&page->lru); 877 878 if (!trylock_page(page)) 879 goto keep; 880 881 VM_BUG_ON_PAGE(PageActive(page), page); 882 VM_BUG_ON_PAGE(page_zone(page) != zone, page); 883 884 sc->nr_scanned++; 885 886 if (unlikely(!page_evictable(page))) 887 goto cull_mlocked; 888 889 if (!sc->may_unmap && page_mapped(page)) 890 goto keep_locked; 891 892 /* Double the slab pressure for mapped and swapcache pages */ 893 if (page_mapped(page) || PageSwapCache(page)) 894 sc->nr_scanned++; 895 896 may_enter_fs = (sc->gfp_mask & __GFP_FS) || 897 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); 898 899 /* 900 * The number of dirty pages determines if a zone is marked 901 * reclaim_congested which affects wait_iff_congested. kswapd 902 * will stall and start writing pages if the tail of the LRU 903 * is all dirty unqueued pages. 904 */ 905 page_check_dirty_writeback(page, &dirty, &writeback); 906 if (dirty || writeback) 907 nr_dirty++; 908 909 if (dirty && !writeback) 910 nr_unqueued_dirty++; 911 912 /* 913 * Treat this page as congested if the underlying BDI is or if 914 * pages are cycling through the LRU so quickly that the 915 * pages marked for immediate reclaim are making it to the 916 * end of the LRU a second time. 917 */ 918 mapping = page_mapping(page); 919 if (((dirty || writeback) && mapping && 920 bdi_write_congested(inode_to_bdi(mapping->host))) || 921 (writeback && PageReclaim(page))) 922 nr_congested++; 923 924 /* 925 * If a page at the tail of the LRU is under writeback, there 926 * are three cases to consider. 927 * 928 * 1) If reclaim is encountering an excessive number of pages 929 * under writeback and this page is both under writeback and 930 * PageReclaim then it indicates that pages are being queued 931 * for IO but are being recycled through the LRU before the 932 * IO can complete. Waiting on the page itself risks an 933 * indefinite stall if it is impossible to writeback the 934 * page due to IO error or disconnected storage so instead 935 * note that the LRU is being scanned too quickly and the 936 * caller can stall after page list has been processed. 937 * 938 * 2) Global reclaim encounters a page, memcg encounters a 939 * page that is not marked for immediate reclaim or 940 * the caller does not have __GFP_FS (or __GFP_IO if it's 941 * simply going to swap, not to fs). In this case mark 942 * the page for immediate reclaim and continue scanning. 943 * 944 * Require may_enter_fs because we would wait on fs, which 945 * may not have submitted IO yet. And the loop driver might 946 * enter reclaim, and deadlock if it waits on a page for 947 * which it is needed to do the write (loop masks off 948 * __GFP_IO|__GFP_FS for this reason); but more thought 949 * would probably show more reasons. 950 * 951 * 3) memcg encounters a page that is not already marked 952 * PageReclaim. memcg does not have any dirty pages 953 * throttling so we could easily OOM just because too many 954 * pages are in writeback and there is nothing else to 955 * reclaim. Wait for the writeback to complete. 956 */ 957 if (PageWriteback(page)) { 958 /* Case 1 above */ 959 if (current_is_kswapd() && 960 PageReclaim(page) && 961 test_bit(ZONE_WRITEBACK, &zone->flags)) { 962 nr_immediate++; 963 goto keep_locked; 964 965 /* Case 2 above */ 966 } else if (global_reclaim(sc) || 967 !PageReclaim(page) || !may_enter_fs) { 968 /* 969 * This is slightly racy - end_page_writeback() 970 * might have just cleared PageReclaim, then 971 * setting PageReclaim here end up interpreted 972 * as PageReadahead - but that does not matter 973 * enough to care. What we do want is for this 974 * page to have PageReclaim set next time memcg 975 * reclaim reaches the tests above, so it will 976 * then wait_on_page_writeback() to avoid OOM; 977 * and it's also appropriate in global reclaim. 978 */ 979 SetPageReclaim(page); 980 nr_writeback++; 981 982 goto keep_locked; 983 984 /* Case 3 above */ 985 } else { 986 wait_on_page_writeback(page); 987 } 988 } 989 990 if (!force_reclaim) 991 references = page_check_references(page, sc); 992 993 switch (references) { 994 case PAGEREF_ACTIVATE: 995 goto activate_locked; 996 case PAGEREF_KEEP: 997 goto keep_locked; 998 case PAGEREF_RECLAIM: 999 case PAGEREF_RECLAIM_CLEAN: 1000 ; /* try to reclaim the page below */ 1001 } 1002 1003 /* 1004 * Anonymous process memory has backing store? 1005 * Try to allocate it some swap space here. 1006 */ 1007 if (PageAnon(page) && !PageSwapCache(page)) { 1008 if (!(sc->gfp_mask & __GFP_IO)) 1009 goto keep_locked; 1010 if (!add_to_swap(page, page_list)) 1011 goto activate_locked; 1012 may_enter_fs = 1; 1013 1014 /* Adding to swap updated mapping */ 1015 mapping = page_mapping(page); 1016 } 1017 1018 /* 1019 * The page is mapped into the page tables of one or more 1020 * processes. Try to unmap it here. 1021 */ 1022 if (page_mapped(page) && mapping) { 1023 switch (try_to_unmap(page, ttu_flags)) { 1024 case SWAP_FAIL: 1025 goto activate_locked; 1026 case SWAP_AGAIN: 1027 goto keep_locked; 1028 case SWAP_MLOCK: 1029 goto cull_mlocked; 1030 case SWAP_SUCCESS: 1031 ; /* try to free the page below */ 1032 } 1033 } 1034 1035 if (PageDirty(page)) { 1036 /* 1037 * Only kswapd can writeback filesystem pages to 1038 * avoid risk of stack overflow but only writeback 1039 * if many dirty pages have been encountered. 1040 */ 1041 if (page_is_file_cache(page) && 1042 (!current_is_kswapd() || 1043 !test_bit(ZONE_DIRTY, &zone->flags))) { 1044 /* 1045 * Immediately reclaim when written back. 1046 * Similar in principal to deactivate_page() 1047 * except we already have the page isolated 1048 * and know it's dirty 1049 */ 1050 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); 1051 SetPageReclaim(page); 1052 1053 goto keep_locked; 1054 } 1055 1056 if (references == PAGEREF_RECLAIM_CLEAN) 1057 goto keep_locked; 1058 if (!may_enter_fs) 1059 goto keep_locked; 1060 if (!sc->may_writepage) 1061 goto keep_locked; 1062 1063 /* Page is dirty, try to write it out here */ 1064 switch (pageout(page, mapping, sc)) { 1065 case PAGE_KEEP: 1066 goto keep_locked; 1067 case PAGE_ACTIVATE: 1068 goto activate_locked; 1069 case PAGE_SUCCESS: 1070 if (PageWriteback(page)) 1071 goto keep; 1072 if (PageDirty(page)) 1073 goto keep; 1074 1075 /* 1076 * A synchronous write - probably a ramdisk. Go 1077 * ahead and try to reclaim the page. 1078 */ 1079 if (!trylock_page(page)) 1080 goto keep; 1081 if (PageDirty(page) || PageWriteback(page)) 1082 goto keep_locked; 1083 mapping = page_mapping(page); 1084 case PAGE_CLEAN: 1085 ; /* try to free the page below */ 1086 } 1087 } 1088 1089 /* 1090 * If the page has buffers, try to free the buffer mappings 1091 * associated with this page. If we succeed we try to free 1092 * the page as well. 1093 * 1094 * We do this even if the page is PageDirty(). 1095 * try_to_release_page() does not perform I/O, but it is 1096 * possible for a page to have PageDirty set, but it is actually 1097 * clean (all its buffers are clean). This happens if the 1098 * buffers were written out directly, with submit_bh(). ext3 1099 * will do this, as well as the blockdev mapping. 1100 * try_to_release_page() will discover that cleanness and will 1101 * drop the buffers and mark the page clean - it can be freed. 1102 * 1103 * Rarely, pages can have buffers and no ->mapping. These are 1104 * the pages which were not successfully invalidated in 1105 * truncate_complete_page(). We try to drop those buffers here 1106 * and if that worked, and the page is no longer mapped into 1107 * process address space (page_count == 1) it can be freed. 1108 * Otherwise, leave the page on the LRU so it is swappable. 1109 */ 1110 if (page_has_private(page)) { 1111 if (!try_to_release_page(page, sc->gfp_mask)) 1112 goto activate_locked; 1113 if (!mapping && page_count(page) == 1) { 1114 unlock_page(page); 1115 if (put_page_testzero(page)) 1116 goto free_it; 1117 else { 1118 /* 1119 * rare race with speculative reference. 1120 * the speculative reference will free 1121 * this page shortly, so we may 1122 * increment nr_reclaimed here (and 1123 * leave it off the LRU). 1124 */ 1125 nr_reclaimed++; 1126 continue; 1127 } 1128 } 1129 } 1130 1131 if (!mapping || !__remove_mapping(mapping, page, true)) 1132 goto keep_locked; 1133 1134 /* 1135 * At this point, we have no other references and there is 1136 * no way to pick any more up (removed from LRU, removed 1137 * from pagecache). Can use non-atomic bitops now (and 1138 * we obviously don't have to worry about waking up a process 1139 * waiting on the page lock, because there are no references. 1140 */ 1141 __clear_page_locked(page); 1142free_it: 1143 nr_reclaimed++; 1144 1145 /* 1146 * Is there need to periodically free_page_list? It would 1147 * appear not as the counts should be low 1148 */ 1149 list_add(&page->lru, &free_pages); 1150 continue; 1151 1152cull_mlocked: 1153 if (PageSwapCache(page)) 1154 try_to_free_swap(page); 1155 unlock_page(page); 1156 list_add(&page->lru, &ret_pages); 1157 continue; 1158 1159activate_locked: 1160 /* Not a candidate for swapping, so reclaim swap space. */ 1161 if (PageSwapCache(page) && vm_swap_full()) 1162 try_to_free_swap(page); 1163 VM_BUG_ON_PAGE(PageActive(page), page); 1164 SetPageActive(page); 1165 pgactivate++; 1166keep_locked: 1167 unlock_page(page); 1168keep: 1169 list_add(&page->lru, &ret_pages); 1170 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); 1171 } 1172 1173 mem_cgroup_uncharge_list(&free_pages); 1174 free_hot_cold_page_list(&free_pages, true); 1175 1176 list_splice(&ret_pages, page_list); 1177 count_vm_events(PGACTIVATE, pgactivate); 1178 1179 *ret_nr_dirty += nr_dirty; 1180 *ret_nr_congested += nr_congested; 1181 *ret_nr_unqueued_dirty += nr_unqueued_dirty; 1182 *ret_nr_writeback += nr_writeback; 1183 *ret_nr_immediate += nr_immediate; 1184 return nr_reclaimed; 1185} 1186 1187unsigned long reclaim_clean_pages_from_list(struct zone *zone, 1188 struct list_head *page_list) 1189{ 1190 struct scan_control sc = { 1191 .gfp_mask = GFP_KERNEL, 1192 .priority = DEF_PRIORITY, 1193 .may_unmap = 1, 1194 }; 1195 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5; 1196 struct page *page, *next; 1197 LIST_HEAD(clean_pages); 1198 1199 list_for_each_entry_safe(page, next, page_list, lru) { 1200 if (page_is_file_cache(page) && !PageDirty(page) && 1201 !isolated_balloon_page(page)) { 1202 ClearPageActive(page); 1203 list_move(&page->lru, &clean_pages); 1204 } 1205 } 1206 1207 ret = shrink_page_list(&clean_pages, zone, &sc, 1208 TTU_UNMAP|TTU_IGNORE_ACCESS, 1209 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true); 1210 list_splice(&clean_pages, page_list); 1211 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret); 1212 return ret; 1213} 1214 1215/* 1216 * Attempt to remove the specified page from its LRU. Only take this page 1217 * if it is of the appropriate PageActive status. Pages which are being 1218 * freed elsewhere are also ignored. 1219 * 1220 * page: page to consider 1221 * mode: one of the LRU isolation modes defined above 1222 * 1223 * returns 0 on success, -ve errno on failure. 1224 */ 1225int __isolate_lru_page(struct page *page, isolate_mode_t mode) 1226{ 1227 int ret = -EINVAL; 1228 1229 /* Only take pages on the LRU. */ 1230 if (!PageLRU(page)) 1231 return ret; 1232 1233 /* Compaction should not handle unevictable pages but CMA can do so */ 1234 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) 1235 return ret; 1236 1237 ret = -EBUSY; 1238 1239 /* 1240 * To minimise LRU disruption, the caller can indicate that it only 1241 * wants to isolate pages it will be able to operate on without 1242 * blocking - clean pages for the most part. 1243 * 1244 * ISOLATE_CLEAN means that only clean pages should be isolated. This 1245 * is used by reclaim when it is cannot write to backing storage 1246 * 1247 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages 1248 * that it is possible to migrate without blocking 1249 */ 1250 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { 1251 /* All the caller can do on PageWriteback is block */ 1252 if (PageWriteback(page)) 1253 return ret; 1254 1255 if (PageDirty(page)) { 1256 struct address_space *mapping; 1257 1258 /* ISOLATE_CLEAN means only clean pages */ 1259 if (mode & ISOLATE_CLEAN) 1260 return ret; 1261 1262 /* 1263 * Only pages without mappings or that have a 1264 * ->migratepage callback are possible to migrate 1265 * without blocking 1266 */ 1267 mapping = page_mapping(page); 1268 if (mapping && !mapping->a_ops->migratepage) 1269 return ret; 1270 } 1271 } 1272 1273 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) 1274 return ret; 1275 1276 if (likely(get_page_unless_zero(page))) { 1277 /* 1278 * Be careful not to clear PageLRU until after we're 1279 * sure the page is not being freed elsewhere -- the 1280 * page release code relies on it. 1281 */ 1282 ClearPageLRU(page); 1283 ret = 0; 1284 } 1285 1286 return ret; 1287} 1288 1289/* 1290 * zone->lru_lock is heavily contended. Some of the functions that 1291 * shrink the lists perform better by taking out a batch of pages 1292 * and working on them outside the LRU lock. 1293 * 1294 * For pagecache intensive workloads, this function is the hottest 1295 * spot in the kernel (apart from copy_*_user functions). 1296 * 1297 * Appropriate locks must be held before calling this function. 1298 * 1299 * @nr_to_scan: The number of pages to look through on the list. 1300 * @lruvec: The LRU vector to pull pages from. 1301 * @dst: The temp list to put pages on to. 1302 * @nr_scanned: The number of pages that were scanned. 1303 * @sc: The scan_control struct for this reclaim session 1304 * @mode: One of the LRU isolation modes 1305 * @lru: LRU list id for isolating 1306 * 1307 * returns how many pages were moved onto *@dst. 1308 */ 1309static unsigned long isolate_lru_pages(unsigned long nr_to_scan, 1310 struct lruvec *lruvec, struct list_head *dst, 1311 unsigned long *nr_scanned, struct scan_control *sc, 1312 isolate_mode_t mode, enum lru_list lru) 1313{ 1314 struct list_head *src = &lruvec->lists[lru]; 1315 unsigned long nr_taken = 0; 1316 unsigned long scan; 1317 1318 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { 1319 struct page *page; 1320 int nr_pages; 1321 1322 page = lru_to_page(src); 1323 prefetchw_prev_lru_page(page, src, flags); 1324 1325 VM_BUG_ON_PAGE(!PageLRU(page), page); 1326 1327 switch (__isolate_lru_page(page, mode)) { 1328 case 0: 1329 nr_pages = hpage_nr_pages(page); 1330 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages); 1331 list_move(&page->lru, dst); 1332 nr_taken += nr_pages; 1333 break; 1334 1335 case -EBUSY: 1336 /* else it is being freed elsewhere */ 1337 list_move(&page->lru, src); 1338 continue; 1339 1340 default: 1341 BUG(); 1342 } 1343 } 1344 1345 *nr_scanned = scan; 1346 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan, 1347 nr_taken, mode, is_file_lru(lru)); 1348 return nr_taken; 1349} 1350 1351/** 1352 * isolate_lru_page - tries to isolate a page from its LRU list 1353 * @page: page to isolate from its LRU list 1354 * 1355 * Isolates a @page from an LRU list, clears PageLRU and adjusts the 1356 * vmstat statistic corresponding to whatever LRU list the page was on. 1357 * 1358 * Returns 0 if the page was removed from an LRU list. 1359 * Returns -EBUSY if the page was not on an LRU list. 1360 * 1361 * The returned page will have PageLRU() cleared. If it was found on 1362 * the active list, it will have PageActive set. If it was found on 1363 * the unevictable list, it will have the PageUnevictable bit set. That flag 1364 * may need to be cleared by the caller before letting the page go. 1365 * 1366 * The vmstat statistic corresponding to the list on which the page was 1367 * found will be decremented. 1368 * 1369 * Restrictions: 1370 * (1) Must be called with an elevated refcount on the page. This is a 1371 * fundamentnal difference from isolate_lru_pages (which is called 1372 * without a stable reference). 1373 * (2) the lru_lock must not be held. 1374 * (3) interrupts must be enabled. 1375 */ 1376int isolate_lru_page(struct page *page) 1377{ 1378 int ret = -EBUSY; 1379 1380 VM_BUG_ON_PAGE(!page_count(page), page); 1381 1382 if (PageLRU(page)) { 1383 struct zone *zone = page_zone(page); 1384 struct lruvec *lruvec; 1385 1386 spin_lock_irq(&zone->lru_lock); 1387 lruvec = mem_cgroup_page_lruvec(page, zone); 1388 if (PageLRU(page)) { 1389 int lru = page_lru(page); 1390 get_page(page); 1391 ClearPageLRU(page); 1392 del_page_from_lru_list(page, lruvec, lru); 1393 ret = 0; 1394 } 1395 spin_unlock_irq(&zone->lru_lock); 1396 } 1397 return ret; 1398} 1399 1400/* 1401 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and 1402 * then get resheduled. When there are massive number of tasks doing page 1403 * allocation, such sleeping direct reclaimers may keep piling up on each CPU, 1404 * the LRU list will go small and be scanned faster than necessary, leading to 1405 * unnecessary swapping, thrashing and OOM. 1406 */ 1407static int too_many_isolated(struct zone *zone, int file, 1408 struct scan_control *sc) 1409{ 1410 unsigned long inactive, isolated; 1411 1412 if (current_is_kswapd()) 1413 return 0; 1414 1415 if (!global_reclaim(sc)) 1416 return 0; 1417 1418 if (file) { 1419 inactive = zone_page_state(zone, NR_INACTIVE_FILE); 1420 isolated = zone_page_state(zone, NR_ISOLATED_FILE); 1421 } else { 1422 inactive = zone_page_state(zone, NR_INACTIVE_ANON); 1423 isolated = zone_page_state(zone, NR_ISOLATED_ANON); 1424 } 1425 1426 /* 1427 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they 1428 * won't get blocked by normal direct-reclaimers, forming a circular 1429 * deadlock. 1430 */ 1431 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS) 1432 inactive >>= 3; 1433 1434 return isolated > inactive; 1435} 1436 1437static noinline_for_stack void 1438putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) 1439{ 1440 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1441 struct zone *zone = lruvec_zone(lruvec); 1442 LIST_HEAD(pages_to_free); 1443 1444 /* 1445 * Put back any unfreeable pages. 1446 */ 1447 while (!list_empty(page_list)) { 1448 struct page *page = lru_to_page(page_list); 1449 int lru; 1450 1451 VM_BUG_ON_PAGE(PageLRU(page), page); 1452 list_del(&page->lru); 1453 if (unlikely(!page_evictable(page))) { 1454 spin_unlock_irq(&zone->lru_lock); 1455 putback_lru_page(page); 1456 spin_lock_irq(&zone->lru_lock); 1457 continue; 1458 } 1459 1460 lruvec = mem_cgroup_page_lruvec(page, zone); 1461 1462 SetPageLRU(page); 1463 lru = page_lru(page); 1464 add_page_to_lru_list(page, lruvec, lru); 1465 1466 if (is_active_lru(lru)) { 1467 int file = is_file_lru(lru); 1468 int numpages = hpage_nr_pages(page); 1469 reclaim_stat->recent_rotated[file] += numpages; 1470 } 1471 if (put_page_testzero(page)) { 1472 __ClearPageLRU(page); 1473 __ClearPageActive(page); 1474 del_page_from_lru_list(page, lruvec, lru); 1475 1476 if (unlikely(PageCompound(page))) { 1477 spin_unlock_irq(&zone->lru_lock); 1478 mem_cgroup_uncharge(page); 1479 (*get_compound_page_dtor(page))(page); 1480 spin_lock_irq(&zone->lru_lock); 1481 } else 1482 list_add(&page->lru, &pages_to_free); 1483 } 1484 } 1485 1486 /* 1487 * To save our caller's stack, now use input list for pages to free. 1488 */ 1489 list_splice(&pages_to_free, page_list); 1490} 1491 1492/* 1493 * If a kernel thread (such as nfsd for loop-back mounts) services 1494 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. 1495 * In that case we should only throttle if the backing device it is 1496 * writing to is congested. In other cases it is safe to throttle. 1497 */ 1498static int current_may_throttle(void) 1499{ 1500 return !(current->flags & PF_LESS_THROTTLE) || 1501 current->backing_dev_info == NULL || 1502 bdi_write_congested(current->backing_dev_info); 1503} 1504 1505/* 1506 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number 1507 * of reclaimed pages 1508 */ 1509static noinline_for_stack unsigned long 1510shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, 1511 struct scan_control *sc, enum lru_list lru) 1512{ 1513 LIST_HEAD(page_list); 1514 unsigned long nr_scanned; 1515 unsigned long nr_reclaimed = 0; 1516 unsigned long nr_taken; 1517 unsigned long nr_dirty = 0; 1518 unsigned long nr_congested = 0; 1519 unsigned long nr_unqueued_dirty = 0; 1520 unsigned long nr_writeback = 0; 1521 unsigned long nr_immediate = 0; 1522 isolate_mode_t isolate_mode = 0; 1523 int file = is_file_lru(lru); 1524 struct zone *zone = lruvec_zone(lruvec); 1525 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1526 1527 while (unlikely(too_many_isolated(zone, file, sc))) { 1528 congestion_wait(BLK_RW_ASYNC, HZ/10); 1529 1530 /* We are about to die and free our memory. Return now. */ 1531 if (fatal_signal_pending(current)) 1532 return SWAP_CLUSTER_MAX; 1533 } 1534 1535 lru_add_drain(); 1536 1537 if (!sc->may_unmap) 1538 isolate_mode |= ISOLATE_UNMAPPED; 1539 if (!sc->may_writepage) 1540 isolate_mode |= ISOLATE_CLEAN; 1541 1542 spin_lock_irq(&zone->lru_lock); 1543 1544 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, 1545 &nr_scanned, sc, isolate_mode, lru); 1546 1547 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); 1548 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); 1549 1550 if (global_reclaim(sc)) { 1551 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned); 1552 if (current_is_kswapd()) 1553 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned); 1554 else 1555 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned); 1556 } 1557 spin_unlock_irq(&zone->lru_lock); 1558 1559 if (nr_taken == 0) 1560 return 0; 1561 1562 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP, 1563 &nr_dirty, &nr_unqueued_dirty, &nr_congested, 1564 &nr_writeback, &nr_immediate, 1565 false); 1566 1567 spin_lock_irq(&zone->lru_lock); 1568 1569 reclaim_stat->recent_scanned[file] += nr_taken; 1570 1571 if (global_reclaim(sc)) { 1572 if (current_is_kswapd()) 1573 __count_zone_vm_events(PGSTEAL_KSWAPD, zone, 1574 nr_reclaimed); 1575 else 1576 __count_zone_vm_events(PGSTEAL_DIRECT, zone, 1577 nr_reclaimed); 1578 } 1579 1580 putback_inactive_pages(lruvec, &page_list); 1581 1582 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); 1583 1584 spin_unlock_irq(&zone->lru_lock); 1585 1586 mem_cgroup_uncharge_list(&page_list); 1587 free_hot_cold_page_list(&page_list, true); 1588 1589 /* 1590 * If reclaim is isolating dirty pages under writeback, it implies 1591 * that the long-lived page allocation rate is exceeding the page 1592 * laundering rate. Either the global limits are not being effective 1593 * at throttling processes due to the page distribution throughout 1594 * zones or there is heavy usage of a slow backing device. The 1595 * only option is to throttle from reclaim context which is not ideal 1596 * as there is no guarantee the dirtying process is throttled in the 1597 * same way balance_dirty_pages() manages. 1598 * 1599 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number 1600 * of pages under pages flagged for immediate reclaim and stall if any 1601 * are encountered in the nr_immediate check below. 1602 */ 1603 if (nr_writeback && nr_writeback == nr_taken) 1604 set_bit(ZONE_WRITEBACK, &zone->flags); 1605 1606 /* 1607 * memcg will stall in page writeback so only consider forcibly 1608 * stalling for global reclaim 1609 */ 1610 if (global_reclaim(sc)) { 1611 /* 1612 * Tag a zone as congested if all the dirty pages scanned were 1613 * backed by a congested BDI and wait_iff_congested will stall. 1614 */ 1615 if (nr_dirty && nr_dirty == nr_congested) 1616 set_bit(ZONE_CONGESTED, &zone->flags); 1617 1618 /* 1619 * If dirty pages are scanned that are not queued for IO, it 1620 * implies that flushers are not keeping up. In this case, flag 1621 * the zone ZONE_DIRTY and kswapd will start writing pages from 1622 * reclaim context. 1623 */ 1624 if (nr_unqueued_dirty == nr_taken) 1625 set_bit(ZONE_DIRTY, &zone->flags); 1626 1627 /* 1628 * If kswapd scans pages marked marked for immediate 1629 * reclaim and under writeback (nr_immediate), it implies 1630 * that pages are cycling through the LRU faster than 1631 * they are written so also forcibly stall. 1632 */ 1633 if (nr_immediate && current_may_throttle()) 1634 congestion_wait(BLK_RW_ASYNC, HZ/10); 1635 } 1636 1637 /* 1638 * Stall direct reclaim for IO completions if underlying BDIs or zone 1639 * is congested. Allow kswapd to continue until it starts encountering 1640 * unqueued dirty pages or cycling through the LRU too quickly. 1641 */ 1642 if (!sc->hibernation_mode && !current_is_kswapd() && 1643 current_may_throttle()) 1644 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); 1645 1646 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, 1647 zone_idx(zone), 1648 nr_scanned, nr_reclaimed, 1649 sc->priority, 1650 trace_shrink_flags(file)); 1651 return nr_reclaimed; 1652} 1653 1654/* 1655 * This moves pages from the active list to the inactive list. 1656 * 1657 * We move them the other way if the page is referenced by one or more 1658 * processes, from rmap. 1659 * 1660 * If the pages are mostly unmapped, the processing is fast and it is 1661 * appropriate to hold zone->lru_lock across the whole operation. But if 1662 * the pages are mapped, the processing is slow (page_referenced()) so we 1663 * should drop zone->lru_lock around each page. It's impossible to balance 1664 * this, so instead we remove the pages from the LRU while processing them. 1665 * It is safe to rely on PG_active against the non-LRU pages in here because 1666 * nobody will play with that bit on a non-LRU page. 1667 * 1668 * The downside is that we have to touch page->_count against each page. 1669 * But we had to alter page->flags anyway. 1670 */ 1671 1672static void move_active_pages_to_lru(struct lruvec *lruvec, 1673 struct list_head *list, 1674 struct list_head *pages_to_free, 1675 enum lru_list lru) 1676{ 1677 struct zone *zone = lruvec_zone(lruvec); 1678 unsigned long pgmoved = 0; 1679 struct page *page; 1680 int nr_pages; 1681 1682 while (!list_empty(list)) { 1683 page = lru_to_page(list); 1684 lruvec = mem_cgroup_page_lruvec(page, zone); 1685 1686 VM_BUG_ON_PAGE(PageLRU(page), page); 1687 SetPageLRU(page); 1688 1689 nr_pages = hpage_nr_pages(page); 1690 mem_cgroup_update_lru_size(lruvec, lru, nr_pages); 1691 list_move(&page->lru, &lruvec->lists[lru]); 1692 pgmoved += nr_pages; 1693 1694 if (put_page_testzero(page)) { 1695 __ClearPageLRU(page); 1696 __ClearPageActive(page); 1697 del_page_from_lru_list(page, lruvec, lru); 1698 1699 if (unlikely(PageCompound(page))) { 1700 spin_unlock_irq(&zone->lru_lock); 1701 mem_cgroup_uncharge(page); 1702 (*get_compound_page_dtor(page))(page); 1703 spin_lock_irq(&zone->lru_lock); 1704 } else 1705 list_add(&page->lru, pages_to_free); 1706 } 1707 } 1708 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); 1709 if (!is_active_lru(lru)) 1710 __count_vm_events(PGDEACTIVATE, pgmoved); 1711} 1712 1713static void shrink_active_list(unsigned long nr_to_scan, 1714 struct lruvec *lruvec, 1715 struct scan_control *sc, 1716 enum lru_list lru) 1717{ 1718 unsigned long nr_taken; 1719 unsigned long nr_scanned; 1720 unsigned long vm_flags; 1721 LIST_HEAD(l_hold); /* The pages which were snipped off */ 1722 LIST_HEAD(l_active); 1723 LIST_HEAD(l_inactive); 1724 struct page *page; 1725 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1726 unsigned long nr_rotated = 0; 1727 isolate_mode_t isolate_mode = 0; 1728 int file = is_file_lru(lru); 1729 struct zone *zone = lruvec_zone(lruvec); 1730 1731 lru_add_drain(); 1732 1733 if (!sc->may_unmap) 1734 isolate_mode |= ISOLATE_UNMAPPED; 1735 if (!sc->may_writepage) 1736 isolate_mode |= ISOLATE_CLEAN; 1737 1738 spin_lock_irq(&zone->lru_lock); 1739 1740 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, 1741 &nr_scanned, sc, isolate_mode, lru); 1742 if (global_reclaim(sc)) 1743 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned); 1744 1745 reclaim_stat->recent_scanned[file] += nr_taken; 1746 1747 __count_zone_vm_events(PGREFILL, zone, nr_scanned); 1748 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); 1749 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); 1750 spin_unlock_irq(&zone->lru_lock); 1751 1752 while (!list_empty(&l_hold)) { 1753 cond_resched(); 1754 page = lru_to_page(&l_hold); 1755 list_del(&page->lru); 1756 1757 if (unlikely(!page_evictable(page))) { 1758 putback_lru_page(page); 1759 continue; 1760 } 1761 1762 if (unlikely(buffer_heads_over_limit)) { 1763 if (page_has_private(page) && trylock_page(page)) { 1764 if (page_has_private(page)) 1765 try_to_release_page(page, 0); 1766 unlock_page(page); 1767 } 1768 } 1769 1770 if (page_referenced(page, 0, sc->target_mem_cgroup, 1771 &vm_flags)) { 1772 nr_rotated += hpage_nr_pages(page); 1773 /* 1774 * Identify referenced, file-backed active pages and 1775 * give them one more trip around the active list. So 1776 * that executable code get better chances to stay in 1777 * memory under moderate memory pressure. Anon pages 1778 * are not likely to be evicted by use-once streaming 1779 * IO, plus JVM can create lots of anon VM_EXEC pages, 1780 * so we ignore them here. 1781 */ 1782 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { 1783 list_add(&page->lru, &l_active); 1784 continue; 1785 } 1786 } 1787 1788 ClearPageActive(page); /* we are de-activating */ 1789 list_add(&page->lru, &l_inactive); 1790 } 1791 1792 /* 1793 * Move pages back to the lru list. 1794 */ 1795 spin_lock_irq(&zone->lru_lock); 1796 /* 1797 * Count referenced pages from currently used mappings as rotated, 1798 * even though only some of them are actually re-activated. This 1799 * helps balance scan pressure between file and anonymous pages in 1800 * get_scan_count. 1801 */ 1802 reclaim_stat->recent_rotated[file] += nr_rotated; 1803 1804 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); 1805 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); 1806 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); 1807 spin_unlock_irq(&zone->lru_lock); 1808 1809 mem_cgroup_uncharge_list(&l_hold); 1810 free_hot_cold_page_list(&l_hold, true); 1811} 1812 1813#ifdef CONFIG_SWAP 1814static int inactive_anon_is_low_global(struct zone *zone) 1815{ 1816 unsigned long active, inactive; 1817 1818 active = zone_page_state(zone, NR_ACTIVE_ANON); 1819 inactive = zone_page_state(zone, NR_INACTIVE_ANON); 1820 1821 if (inactive * zone->inactive_ratio < active) 1822 return 1; 1823 1824 return 0; 1825} 1826 1827/** 1828 * inactive_anon_is_low - check if anonymous pages need to be deactivated 1829 * @lruvec: LRU vector to check 1830 * 1831 * Returns true if the zone does not have enough inactive anon pages, 1832 * meaning some active anon pages need to be deactivated. 1833 */ 1834static int inactive_anon_is_low(struct lruvec *lruvec) 1835{ 1836 /* 1837 * If we don't have swap space, anonymous page deactivation 1838 * is pointless. 1839 */ 1840 if (!total_swap_pages) 1841 return 0; 1842 1843 if (!mem_cgroup_disabled()) 1844 return mem_cgroup_inactive_anon_is_low(lruvec); 1845 1846 return inactive_anon_is_low_global(lruvec_zone(lruvec)); 1847} 1848#else 1849static inline int inactive_anon_is_low(struct lruvec *lruvec) 1850{ 1851 return 0; 1852} 1853#endif 1854 1855/** 1856 * inactive_file_is_low - check if file pages need to be deactivated 1857 * @lruvec: LRU vector to check 1858 * 1859 * When the system is doing streaming IO, memory pressure here 1860 * ensures that active file pages get deactivated, until more 1861 * than half of the file pages are on the inactive list. 1862 * 1863 * Once we get to that situation, protect the system's working 1864 * set from being evicted by disabling active file page aging. 1865 * 1866 * This uses a different ratio than the anonymous pages, because 1867 * the page cache uses a use-once replacement algorithm. 1868 */ 1869static int inactive_file_is_low(struct lruvec *lruvec) 1870{ 1871 unsigned long inactive; 1872 unsigned long active; 1873 1874 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE); 1875 active = get_lru_size(lruvec, LRU_ACTIVE_FILE); 1876 1877 return active > inactive; 1878} 1879 1880static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru) 1881{ 1882 if (is_file_lru(lru)) 1883 return inactive_file_is_low(lruvec); 1884 else 1885 return inactive_anon_is_low(lruvec); 1886} 1887 1888static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, 1889 struct lruvec *lruvec, struct scan_control *sc) 1890{ 1891 if (is_active_lru(lru)) { 1892 if (inactive_list_is_low(lruvec, lru)) 1893 shrink_active_list(nr_to_scan, lruvec, sc, lru); 1894 return 0; 1895 } 1896 1897 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); 1898} 1899 1900enum scan_balance { 1901 SCAN_EQUAL, 1902 SCAN_FRACT, 1903 SCAN_ANON, 1904 SCAN_FILE, 1905}; 1906 1907/* 1908 * Determine how aggressively the anon and file LRU lists should be 1909 * scanned. The relative value of each set of LRU lists is determined 1910 * by looking at the fraction of the pages scanned we did rotate back 1911 * onto the active list instead of evict. 1912 * 1913 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan 1914 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan 1915 */ 1916static void get_scan_count(struct lruvec *lruvec, int swappiness, 1917 struct scan_control *sc, unsigned long *nr, 1918 unsigned long *lru_pages) 1919{ 1920 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1921 u64 fraction[2]; 1922 u64 denominator = 0; /* gcc */ 1923 struct zone *zone = lruvec_zone(lruvec); 1924 unsigned long anon_prio, file_prio; 1925 enum scan_balance scan_balance; 1926 unsigned long anon, file; 1927 bool force_scan = false; 1928 unsigned long ap, fp; 1929 enum lru_list lru; 1930 bool some_scanned; 1931 int pass; 1932 1933 /* 1934 * If the zone or memcg is small, nr[l] can be 0. This 1935 * results in no scanning on this priority and a potential 1936 * priority drop. Global direct reclaim can go to the next 1937 * zone and tends to have no problems. Global kswapd is for 1938 * zone balancing and it needs to scan a minimum amount. When 1939 * reclaiming for a memcg, a priority drop can cause high 1940 * latencies, so it's better to scan a minimum amount there as 1941 * well. 1942 */ 1943 if (current_is_kswapd()) { 1944 if (!zone_reclaimable(zone)) 1945 force_scan = true; 1946 if (!mem_cgroup_lruvec_online(lruvec)) 1947 force_scan = true; 1948 } 1949 if (!global_reclaim(sc)) 1950 force_scan = true; 1951 1952 /* If we have no swap space, do not bother scanning anon pages. */ 1953 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) { 1954 scan_balance = SCAN_FILE; 1955 goto out; 1956 } 1957 1958 /* 1959 * Global reclaim will swap to prevent OOM even with no 1960 * swappiness, but memcg users want to use this knob to 1961 * disable swapping for individual groups completely when 1962 * using the memory controller's swap limit feature would be 1963 * too expensive. 1964 */ 1965 if (!global_reclaim(sc) && !swappiness) { 1966 scan_balance = SCAN_FILE; 1967 goto out; 1968 } 1969 1970 /* 1971 * Do not apply any pressure balancing cleverness when the 1972 * system is close to OOM, scan both anon and file equally 1973 * (unless the swappiness setting disagrees with swapping). 1974 */ 1975 if (!sc->priority && swappiness) { 1976 scan_balance = SCAN_EQUAL; 1977 goto out; 1978 } 1979 1980 /* 1981 * Prevent the reclaimer from falling into the cache trap: as 1982 * cache pages start out inactive, every cache fault will tip 1983 * the scan balance towards the file LRU. And as the file LRU 1984 * shrinks, so does the window for rotation from references. 1985 * This means we have a runaway feedback loop where a tiny 1986 * thrashing file LRU becomes infinitely more attractive than 1987 * anon pages. Try to detect this based on file LRU size. 1988 */ 1989 if (global_reclaim(sc)) { 1990 unsigned long zonefile; 1991 unsigned long zonefree; 1992 1993 zonefree = zone_page_state(zone, NR_FREE_PAGES); 1994 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) + 1995 zone_page_state(zone, NR_INACTIVE_FILE); 1996 1997 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) { 1998 scan_balance = SCAN_ANON; 1999 goto out; 2000 } 2001 } 2002 2003 /* 2004 * There is enough inactive page cache, do not reclaim 2005 * anything from the anonymous working set right now. 2006 */ 2007 if (!inactive_file_is_low(lruvec)) { 2008 scan_balance = SCAN_FILE; 2009 goto out; 2010 } 2011 2012 scan_balance = SCAN_FRACT; 2013 2014 /* 2015 * With swappiness at 100, anonymous and file have the same priority. 2016 * This scanning priority is essentially the inverse of IO cost. 2017 */ 2018 anon_prio = swappiness; 2019 file_prio = 200 - anon_prio; 2020 2021 /* 2022 * OK, so we have swap space and a fair amount of page cache 2023 * pages. We use the recently rotated / recently scanned 2024 * ratios to determine how valuable each cache is. 2025 * 2026 * Because workloads change over time (and to avoid overflow) 2027 * we keep these statistics as a floating average, which ends 2028 * up weighing recent references more than old ones. 2029 * 2030 * anon in [0], file in [1] 2031 */ 2032 2033 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) + 2034 get_lru_size(lruvec, LRU_INACTIVE_ANON); 2035 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) + 2036 get_lru_size(lruvec, LRU_INACTIVE_FILE); 2037 2038 spin_lock_irq(&zone->lru_lock); 2039 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { 2040 reclaim_stat->recent_scanned[0] /= 2; 2041 reclaim_stat->recent_rotated[0] /= 2; 2042 } 2043 2044 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { 2045 reclaim_stat->recent_scanned[1] /= 2; 2046 reclaim_stat->recent_rotated[1] /= 2; 2047 } 2048 2049 /* 2050 * The amount of pressure on anon vs file pages is inversely 2051 * proportional to the fraction of recently scanned pages on 2052 * each list that were recently referenced and in active use. 2053 */ 2054 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); 2055 ap /= reclaim_stat->recent_rotated[0] + 1; 2056 2057 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); 2058 fp /= reclaim_stat->recent_rotated[1] + 1; 2059 spin_unlock_irq(&zone->lru_lock); 2060 2061 fraction[0] = ap; 2062 fraction[1] = fp; 2063 denominator = ap + fp + 1; 2064out: 2065 some_scanned = false; 2066 /* Only use force_scan on second pass. */ 2067 for (pass = 0; !some_scanned && pass < 2; pass++) { 2068 *lru_pages = 0; 2069 for_each_evictable_lru(lru) { 2070 int file = is_file_lru(lru); 2071 unsigned long size; 2072 unsigned long scan; 2073 2074 size = get_lru_size(lruvec, lru); 2075 scan = size >> sc->priority; 2076 2077 if (!scan && pass && force_scan) 2078 scan = min(size, SWAP_CLUSTER_MAX); 2079 2080 switch (scan_balance) { 2081 case SCAN_EQUAL: 2082 /* Scan lists relative to size */ 2083 break; 2084 case SCAN_FRACT: 2085 /* 2086 * Scan types proportional to swappiness and 2087 * their relative recent reclaim efficiency. 2088 */ 2089 scan = div64_u64(scan * fraction[file], 2090 denominator); 2091 break; 2092 case SCAN_FILE: 2093 case SCAN_ANON: 2094 /* Scan one type exclusively */ 2095 if ((scan_balance == SCAN_FILE) != file) { 2096 size = 0; 2097 scan = 0; 2098 } 2099 break; 2100 default: 2101 /* Look ma, no brain */ 2102 BUG(); 2103 } 2104 2105 *lru_pages += size; 2106 nr[lru] = scan; 2107 2108 /* 2109 * Skip the second pass and don't force_scan, 2110 * if we found something to scan. 2111 */ 2112 some_scanned |= !!scan; 2113 } 2114 } 2115} 2116 2117/* 2118 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. 2119 */ 2120static void shrink_lruvec(struct lruvec *lruvec, int swappiness, 2121 struct scan_control *sc, unsigned long *lru_pages) 2122{ 2123 unsigned long nr[NR_LRU_LISTS]; 2124 unsigned long targets[NR_LRU_LISTS]; 2125 unsigned long nr_to_scan; 2126 enum lru_list lru; 2127 unsigned long nr_reclaimed = 0; 2128 unsigned long nr_to_reclaim = sc->nr_to_reclaim; 2129 struct blk_plug plug; 2130 bool scan_adjusted; 2131 2132 get_scan_count(lruvec, swappiness, sc, nr, lru_pages); 2133 2134 /* Record the original scan target for proportional adjustments later */ 2135 memcpy(targets, nr, sizeof(nr)); 2136 2137 /* 2138 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal 2139 * event that can occur when there is little memory pressure e.g. 2140 * multiple streaming readers/writers. Hence, we do not abort scanning 2141 * when the requested number of pages are reclaimed when scanning at 2142 * DEF_PRIORITY on the assumption that the fact we are direct 2143 * reclaiming implies that kswapd is not keeping up and it is best to 2144 * do a batch of work at once. For memcg reclaim one check is made to 2145 * abort proportional reclaim if either the file or anon lru has already 2146 * dropped to zero at the first pass. 2147 */ 2148 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && 2149 sc->priority == DEF_PRIORITY); 2150 2151 blk_start_plug(&plug); 2152 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || 2153 nr[LRU_INACTIVE_FILE]) { 2154 unsigned long nr_anon, nr_file, percentage; 2155 unsigned long nr_scanned; 2156 2157 for_each_evictable_lru(lru) { 2158 if (nr[lru]) { 2159 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); 2160 nr[lru] -= nr_to_scan; 2161 2162 nr_reclaimed += shrink_list(lru, nr_to_scan, 2163 lruvec, sc); 2164 } 2165 } 2166 2167 if (nr_reclaimed < nr_to_reclaim || scan_adjusted) 2168 continue; 2169 2170 /* 2171 * For kswapd and memcg, reclaim at least the number of pages 2172 * requested. Ensure that the anon and file LRUs are scanned 2173 * proportionally what was requested by get_scan_count(). We 2174 * stop reclaiming one LRU and reduce the amount scanning 2175 * proportional to the original scan target. 2176 */ 2177 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; 2178 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; 2179 2180 /* 2181 * It's just vindictive to attack the larger once the smaller 2182 * has gone to zero. And given the way we stop scanning the 2183 * smaller below, this makes sure that we only make one nudge 2184 * towards proportionality once we've got nr_to_reclaim. 2185 */ 2186 if (!nr_file || !nr_anon) 2187 break; 2188 2189 if (nr_file > nr_anon) { 2190 unsigned long scan_target = targets[LRU_INACTIVE_ANON] + 2191 targets[LRU_ACTIVE_ANON] + 1; 2192 lru = LRU_BASE; 2193 percentage = nr_anon * 100 / scan_target; 2194 } else { 2195 unsigned long scan_target = targets[LRU_INACTIVE_FILE] + 2196 targets[LRU_ACTIVE_FILE] + 1; 2197 lru = LRU_FILE; 2198 percentage = nr_file * 100 / scan_target; 2199 } 2200 2201 /* Stop scanning the smaller of the LRU */ 2202 nr[lru] = 0; 2203 nr[lru + LRU_ACTIVE] = 0; 2204 2205 /* 2206 * Recalculate the other LRU scan count based on its original 2207 * scan target and the percentage scanning already complete 2208 */ 2209 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; 2210 nr_scanned = targets[lru] - nr[lru]; 2211 nr[lru] = targets[lru] * (100 - percentage) / 100; 2212 nr[lru] -= min(nr[lru], nr_scanned); 2213 2214 lru += LRU_ACTIVE; 2215 nr_scanned = targets[lru] - nr[lru]; 2216 nr[lru] = targets[lru] * (100 - percentage) / 100; 2217 nr[lru] -= min(nr[lru], nr_scanned); 2218 2219 scan_adjusted = true; 2220 } 2221 blk_finish_plug(&plug); 2222 sc->nr_reclaimed += nr_reclaimed; 2223 2224 /* 2225 * Even if we did not try to evict anon pages at all, we want to 2226 * rebalance the anon lru active/inactive ratio. 2227 */ 2228 if (inactive_anon_is_low(lruvec)) 2229 shrink_active_list(SWAP_CLUSTER_MAX, lruvec, 2230 sc, LRU_ACTIVE_ANON); 2231 2232 throttle_vm_writeout(sc->gfp_mask); 2233} 2234 2235/* Use reclaim/compaction for costly allocs or under memory pressure */ 2236static bool in_reclaim_compaction(struct scan_control *sc) 2237{ 2238 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && 2239 (sc->order > PAGE_ALLOC_COSTLY_ORDER || 2240 sc->priority < DEF_PRIORITY - 2)) 2241 return true; 2242 2243 return false; 2244} 2245 2246/* 2247 * Reclaim/compaction is used for high-order allocation requests. It reclaims 2248 * order-0 pages before compacting the zone. should_continue_reclaim() returns 2249 * true if more pages should be reclaimed such that when the page allocator 2250 * calls try_to_compact_zone() that it will have enough free pages to succeed. 2251 * It will give up earlier than that if there is difficulty reclaiming pages. 2252 */ 2253static inline bool should_continue_reclaim(struct zone *zone, 2254 unsigned long nr_reclaimed, 2255 unsigned long nr_scanned, 2256 struct scan_control *sc) 2257{ 2258 unsigned long pages_for_compaction; 2259 unsigned long inactive_lru_pages; 2260 2261 /* If not in reclaim/compaction mode, stop */ 2262 if (!in_reclaim_compaction(sc)) 2263 return false; 2264 2265 /* Consider stopping depending on scan and reclaim activity */ 2266 if (sc->gfp_mask & __GFP_REPEAT) { 2267 /* 2268 * For __GFP_REPEAT allocations, stop reclaiming if the 2269 * full LRU list has been scanned and we are still failing 2270 * to reclaim pages. This full LRU scan is potentially 2271 * expensive but a __GFP_REPEAT caller really wants to succeed 2272 */ 2273 if (!nr_reclaimed && !nr_scanned) 2274 return false; 2275 } else { 2276 /* 2277 * For non-__GFP_REPEAT allocations which can presumably 2278 * fail without consequence, stop if we failed to reclaim 2279 * any pages from the last SWAP_CLUSTER_MAX number of 2280 * pages that were scanned. This will return to the 2281 * caller faster at the risk reclaim/compaction and 2282 * the resulting allocation attempt fails 2283 */ 2284 if (!nr_reclaimed) 2285 return false; 2286 } 2287 2288 /* 2289 * If we have not reclaimed enough pages for compaction and the 2290 * inactive lists are large enough, continue reclaiming 2291 */ 2292 pages_for_compaction = (2UL << sc->order); 2293 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE); 2294 if (get_nr_swap_pages() > 0) 2295 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON); 2296 if (sc->nr_reclaimed < pages_for_compaction && 2297 inactive_lru_pages > pages_for_compaction) 2298 return true; 2299 2300 /* If compaction would go ahead or the allocation would succeed, stop */ 2301 switch (compaction_suitable(zone, sc->order, 0, 0)) { 2302 case COMPACT_PARTIAL: 2303 case COMPACT_CONTINUE: 2304 return false; 2305 default: 2306 return true; 2307 } 2308} 2309 2310static bool shrink_zone(struct zone *zone, struct scan_control *sc, 2311 bool is_classzone) 2312{ 2313 struct reclaim_state *reclaim_state = current->reclaim_state; 2314 unsigned long nr_reclaimed, nr_scanned; 2315 bool reclaimable = false; 2316 2317 do { 2318 struct mem_cgroup *root = sc->target_mem_cgroup; 2319 struct mem_cgroup_reclaim_cookie reclaim = { 2320 .zone = zone, 2321 .priority = sc->priority, 2322 }; 2323 unsigned long zone_lru_pages = 0; 2324 struct mem_cgroup *memcg; 2325 2326 nr_reclaimed = sc->nr_reclaimed; 2327 nr_scanned = sc->nr_scanned; 2328 2329 memcg = mem_cgroup_iter(root, NULL, &reclaim); 2330 do { 2331 unsigned long lru_pages; 2332 unsigned long scanned; 2333 struct lruvec *lruvec; 2334 int swappiness; 2335 2336 if (mem_cgroup_low(root, memcg)) { 2337 if (!sc->may_thrash) 2338 continue; 2339 mem_cgroup_events(memcg, MEMCG_LOW, 1); 2340 } 2341 2342 lruvec = mem_cgroup_zone_lruvec(zone, memcg); 2343 swappiness = mem_cgroup_swappiness(memcg); 2344 scanned = sc->nr_scanned; 2345 2346 shrink_lruvec(lruvec, swappiness, sc, &lru_pages); 2347 zone_lru_pages += lru_pages; 2348 2349 if (memcg && is_classzone) 2350 shrink_slab(sc->gfp_mask, zone_to_nid(zone), 2351 memcg, sc->nr_scanned - scanned, 2352 lru_pages); 2353 2354 /* 2355 * Direct reclaim and kswapd have to scan all memory 2356 * cgroups to fulfill the overall scan target for the 2357 * zone. 2358 * 2359 * Limit reclaim, on the other hand, only cares about 2360 * nr_to_reclaim pages to be reclaimed and it will 2361 * retry with decreasing priority if one round over the 2362 * whole hierarchy is not sufficient. 2363 */ 2364 if (!global_reclaim(sc) && 2365 sc->nr_reclaimed >= sc->nr_to_reclaim) { 2366 mem_cgroup_iter_break(root, memcg); 2367 break; 2368 } 2369 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); 2370 2371 /* 2372 * Shrink the slab caches in the same proportion that 2373 * the eligible LRU pages were scanned. 2374 */ 2375 if (global_reclaim(sc) && is_classzone) 2376 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL, 2377 sc->nr_scanned - nr_scanned, 2378 zone_lru_pages); 2379 2380 if (reclaim_state) { 2381 sc->nr_reclaimed += reclaim_state->reclaimed_slab; 2382 reclaim_state->reclaimed_slab = 0; 2383 } 2384 2385 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, 2386 sc->nr_scanned - nr_scanned, 2387 sc->nr_reclaimed - nr_reclaimed); 2388 2389 if (sc->nr_reclaimed - nr_reclaimed) 2390 reclaimable = true; 2391 2392 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed, 2393 sc->nr_scanned - nr_scanned, sc)); 2394 2395 return reclaimable; 2396} 2397 2398/* 2399 * Returns true if compaction should go ahead for a high-order request, or 2400 * the high-order allocation would succeed without compaction. 2401 */ 2402static inline bool compaction_ready(struct zone *zone, int order) 2403{ 2404 unsigned long balance_gap, watermark; 2405 bool watermark_ok; 2406 2407 /* 2408 * Compaction takes time to run and there are potentially other 2409 * callers using the pages just freed. Continue reclaiming until 2410 * there is a buffer of free pages available to give compaction 2411 * a reasonable chance of completing and allocating the page 2412 */ 2413 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( 2414 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); 2415 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order); 2416 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0); 2417 2418 /* 2419 * If compaction is deferred, reclaim up to a point where 2420 * compaction will have a chance of success when re-enabled 2421 */ 2422 if (compaction_deferred(zone, order)) 2423 return watermark_ok; 2424 2425 /* 2426 * If compaction is not ready to start and allocation is not likely 2427 * to succeed without it, then keep reclaiming. 2428 */ 2429 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED) 2430 return false; 2431 2432 return watermark_ok; 2433} 2434 2435/* 2436 * This is the direct reclaim path, for page-allocating processes. We only 2437 * try to reclaim pages from zones which will satisfy the caller's allocation 2438 * request. 2439 * 2440 * We reclaim from a zone even if that zone is over high_wmark_pages(zone). 2441 * Because: 2442 * a) The caller may be trying to free *extra* pages to satisfy a higher-order 2443 * allocation or 2444 * b) The target zone may be at high_wmark_pages(zone) but the lower zones 2445 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' 2446 * zone defense algorithm. 2447 * 2448 * If a zone is deemed to be full of pinned pages then just give it a light 2449 * scan then give up on it. 2450 * 2451 * Returns true if a zone was reclaimable. 2452 */ 2453static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc) 2454{ 2455 struct zoneref *z; 2456 struct zone *zone; 2457 unsigned long nr_soft_reclaimed; 2458 unsigned long nr_soft_scanned; 2459 gfp_t orig_mask; 2460 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask); 2461 bool reclaimable = false; 2462 2463 /* 2464 * If the number of buffer_heads in the machine exceeds the maximum 2465 * allowed level, force direct reclaim to scan the highmem zone as 2466 * highmem pages could be pinning lowmem pages storing buffer_heads 2467 */ 2468 orig_mask = sc->gfp_mask; 2469 if (buffer_heads_over_limit) 2470 sc->gfp_mask |= __GFP_HIGHMEM; 2471 2472 for_each_zone_zonelist_nodemask(zone, z, zonelist, 2473 gfp_zone(sc->gfp_mask), sc->nodemask) { 2474 enum zone_type classzone_idx; 2475 2476 if (!populated_zone(zone)) 2477 continue; 2478 2479 classzone_idx = requested_highidx; 2480 while (!populated_zone(zone->zone_pgdat->node_zones + 2481 classzone_idx)) 2482 classzone_idx--; 2483 2484 /* 2485 * Take care memory controller reclaiming has small influence 2486 * to global LRU. 2487 */ 2488 if (global_reclaim(sc)) { 2489 if (!cpuset_zone_allowed(zone, 2490 GFP_KERNEL | __GFP_HARDWALL)) 2491 continue; 2492 2493 if (sc->priority != DEF_PRIORITY && 2494 !zone_reclaimable(zone)) 2495 continue; /* Let kswapd poll it */ 2496 2497 /* 2498 * If we already have plenty of memory free for 2499 * compaction in this zone, don't free any more. 2500 * Even though compaction is invoked for any 2501 * non-zero order, only frequent costly order 2502 * reclamation is disruptive enough to become a 2503 * noticeable problem, like transparent huge 2504 * page allocations. 2505 */ 2506 if (IS_ENABLED(CONFIG_COMPACTION) && 2507 sc->order > PAGE_ALLOC_COSTLY_ORDER && 2508 zonelist_zone_idx(z) <= requested_highidx && 2509 compaction_ready(zone, sc->order)) { 2510 sc->compaction_ready = true; 2511 continue; 2512 } 2513 2514 /* 2515 * This steals pages from memory cgroups over softlimit 2516 * and returns the number of reclaimed pages and 2517 * scanned pages. This works for global memory pressure 2518 * and balancing, not for a memcg's limit. 2519 */ 2520 nr_soft_scanned = 0; 2521 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, 2522 sc->order, sc->gfp_mask, 2523 &nr_soft_scanned); 2524 sc->nr_reclaimed += nr_soft_reclaimed; 2525 sc->nr_scanned += nr_soft_scanned; 2526 if (nr_soft_reclaimed) 2527 reclaimable = true; 2528 /* need some check for avoid more shrink_zone() */ 2529 } 2530 2531 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx)) 2532 reclaimable = true; 2533 2534 if (global_reclaim(sc) && 2535 !reclaimable && zone_reclaimable(zone)) 2536 reclaimable = true; 2537 } 2538 2539 /* 2540 * Restore to original mask to avoid the impact on the caller if we 2541 * promoted it to __GFP_HIGHMEM. 2542 */ 2543 sc->gfp_mask = orig_mask; 2544 2545 return reclaimable; 2546} 2547 2548/* 2549 * This is the main entry point to direct page reclaim. 2550 * 2551 * If a full scan of the inactive list fails to free enough memory then we 2552 * are "out of memory" and something needs to be killed. 2553 * 2554 * If the caller is !__GFP_FS then the probability of a failure is reasonably 2555 * high - the zone may be full of dirty or under-writeback pages, which this 2556 * caller can't do much about. We kick the writeback threads and take explicit 2557 * naps in the hope that some of these pages can be written. But if the 2558 * allocating task holds filesystem locks which prevent writeout this might not 2559 * work, and the allocation attempt will fail. 2560 * 2561 * returns: 0, if no pages reclaimed 2562 * else, the number of pages reclaimed 2563 */ 2564static unsigned long do_try_to_free_pages(struct zonelist *zonelist, 2565 struct scan_control *sc) 2566{ 2567 int initial_priority = sc->priority; 2568 unsigned long total_scanned = 0; 2569 unsigned long writeback_threshold; 2570 bool zones_reclaimable; 2571retry: 2572 delayacct_freepages_start(); 2573 2574 if (global_reclaim(sc)) 2575 count_vm_event(ALLOCSTALL); 2576 2577 do { 2578 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, 2579 sc->priority); 2580 sc->nr_scanned = 0; 2581 zones_reclaimable = shrink_zones(zonelist, sc); 2582 2583 total_scanned += sc->nr_scanned; 2584 if (sc->nr_reclaimed >= sc->nr_to_reclaim) 2585 break; 2586 2587 if (sc->compaction_ready) 2588 break; 2589 2590 /* 2591 * If we're getting trouble reclaiming, start doing 2592 * writepage even in laptop mode. 2593 */ 2594 if (sc->priority < DEF_PRIORITY - 2) 2595 sc->may_writepage = 1; 2596 2597 /* 2598 * Try to write back as many pages as we just scanned. This 2599 * tends to cause slow streaming writers to write data to the 2600 * disk smoothly, at the dirtying rate, which is nice. But 2601 * that's undesirable in laptop mode, where we *want* lumpy 2602 * writeout. So in laptop mode, write out the whole world. 2603 */ 2604 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; 2605 if (total_scanned > writeback_threshold) { 2606 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, 2607 WB_REASON_TRY_TO_FREE_PAGES); 2608 sc->may_writepage = 1; 2609 } 2610 } while (--sc->priority >= 0); 2611 2612 delayacct_freepages_end(); 2613 2614 if (sc->nr_reclaimed) 2615 return sc->nr_reclaimed; 2616 2617 /* Aborted reclaim to try compaction? don't OOM, then */ 2618 if (sc->compaction_ready) 2619 return 1; 2620 2621 /* Untapped cgroup reserves? Don't OOM, retry. */ 2622 if (!sc->may_thrash) { 2623 sc->priority = initial_priority; 2624 sc->may_thrash = 1; 2625 goto retry; 2626 } 2627 2628 /* Any of the zones still reclaimable? Don't OOM. */ 2629 if (zones_reclaimable) 2630 return 1; 2631 2632 return 0; 2633} 2634 2635static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) 2636{ 2637 struct zone *zone; 2638 unsigned long pfmemalloc_reserve = 0; 2639 unsigned long free_pages = 0; 2640 int i; 2641 bool wmark_ok; 2642 2643 for (i = 0; i <= ZONE_NORMAL; i++) { 2644 zone = &pgdat->node_zones[i]; 2645 if (!populated_zone(zone)) 2646 continue; 2647 2648 pfmemalloc_reserve += min_wmark_pages(zone); 2649 free_pages += zone_page_state(zone, NR_FREE_PAGES); 2650 } 2651 2652 /* If there are no reserves (unexpected config) then do not throttle */ 2653 if (!pfmemalloc_reserve) 2654 return true; 2655 2656 wmark_ok = free_pages > pfmemalloc_reserve / 2; 2657 2658 /* kswapd must be awake if processes are being throttled */ 2659 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { 2660 pgdat->classzone_idx = min(pgdat->classzone_idx, 2661 (enum zone_type)ZONE_NORMAL); 2662 wake_up_interruptible(&pgdat->kswapd_wait); 2663 } 2664 2665 return wmark_ok; 2666} 2667 2668/* 2669 * Throttle direct reclaimers if backing storage is backed by the network 2670 * and the PFMEMALLOC reserve for the preferred node is getting dangerously 2671 * depleted. kswapd will continue to make progress and wake the processes 2672 * when the low watermark is reached. 2673 * 2674 * Returns true if a fatal signal was delivered during throttling. If this 2675 * happens, the page allocator should not consider triggering the OOM killer. 2676 */ 2677static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, 2678 nodemask_t *nodemask) 2679{ 2680 struct zoneref *z; 2681 struct zone *zone; 2682 pg_data_t *pgdat = NULL; 2683 2684 /* 2685 * Kernel threads should not be throttled as they may be indirectly 2686 * responsible for cleaning pages necessary for reclaim to make forward 2687 * progress. kjournald for example may enter direct reclaim while 2688 * committing a transaction where throttling it could forcing other 2689 * processes to block on log_wait_commit(). 2690 */ 2691 if (current->flags & PF_KTHREAD) 2692 goto out; 2693 2694 /* 2695 * If a fatal signal is pending, this process should not throttle. 2696 * It should return quickly so it can exit and free its memory 2697 */ 2698 if (fatal_signal_pending(current)) 2699 goto out; 2700 2701 /* 2702 * Check if the pfmemalloc reserves are ok by finding the first node 2703 * with a usable ZONE_NORMAL or lower zone. The expectation is that 2704 * GFP_KERNEL will be required for allocating network buffers when 2705 * swapping over the network so ZONE_HIGHMEM is unusable. 2706 * 2707 * Throttling is based on the first usable node and throttled processes 2708 * wait on a queue until kswapd makes progress and wakes them. There 2709 * is an affinity then between processes waking up and where reclaim 2710 * progress has been made assuming the process wakes on the same node. 2711 * More importantly, processes running on remote nodes will not compete 2712 * for remote pfmemalloc reserves and processes on different nodes 2713 * should make reasonable progress. 2714 */ 2715 for_each_zone_zonelist_nodemask(zone, z, zonelist, 2716 gfp_zone(gfp_mask), nodemask) { 2717 if (zone_idx(zone) > ZONE_NORMAL) 2718 continue; 2719 2720 /* Throttle based on the first usable node */ 2721 pgdat = zone->zone_pgdat; 2722 if (pfmemalloc_watermark_ok(pgdat)) 2723 goto out; 2724 break; 2725 } 2726 2727 /* If no zone was usable by the allocation flags then do not throttle */ 2728 if (!pgdat) 2729 goto out; 2730 2731 /* Account for the throttling */ 2732 count_vm_event(PGSCAN_DIRECT_THROTTLE); 2733 2734 /* 2735 * If the caller cannot enter the filesystem, it's possible that it 2736 * is due to the caller holding an FS lock or performing a journal 2737 * transaction in the case of a filesystem like ext[3|4]. In this case, 2738 * it is not safe to block on pfmemalloc_wait as kswapd could be 2739 * blocked waiting on the same lock. Instead, throttle for up to a 2740 * second before continuing. 2741 */ 2742 if (!(gfp_mask & __GFP_FS)) { 2743 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, 2744 pfmemalloc_watermark_ok(pgdat), HZ); 2745 2746 goto check_pending; 2747 } 2748 2749 /* Throttle until kswapd wakes the process */ 2750 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, 2751 pfmemalloc_watermark_ok(pgdat)); 2752 2753check_pending: 2754 if (fatal_signal_pending(current)) 2755 return true; 2756 2757out: 2758 return false; 2759} 2760 2761unsigned long try_to_free_pages(struct zonelist *zonelist, int order, 2762 gfp_t gfp_mask, nodemask_t *nodemask) 2763{ 2764 unsigned long nr_reclaimed; 2765 struct scan_control sc = { 2766 .nr_to_reclaim = SWAP_CLUSTER_MAX, 2767 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), 2768 .order = order, 2769 .nodemask = nodemask, 2770 .priority = DEF_PRIORITY, 2771 .may_writepage = !laptop_mode, 2772 .may_unmap = 1, 2773 .may_swap = 1, 2774 }; 2775 2776 /* 2777 * Do not enter reclaim if fatal signal was delivered while throttled. 2778 * 1 is returned so that the page allocator does not OOM kill at this 2779 * point. 2780 */ 2781 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) 2782 return 1; 2783 2784 trace_mm_vmscan_direct_reclaim_begin(order, 2785 sc.may_writepage, 2786 gfp_mask); 2787 2788 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 2789 2790 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); 2791 2792 return nr_reclaimed; 2793} 2794 2795#ifdef CONFIG_MEMCG 2796 2797unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, 2798 gfp_t gfp_mask, bool noswap, 2799 struct zone *zone, 2800 unsigned long *nr_scanned) 2801{ 2802 struct scan_control sc = { 2803 .nr_to_reclaim = SWAP_CLUSTER_MAX, 2804 .target_mem_cgroup = memcg, 2805 .may_writepage = !laptop_mode, 2806 .may_unmap = 1, 2807 .may_swap = !noswap, 2808 }; 2809 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); 2810 int swappiness = mem_cgroup_swappiness(memcg); 2811 unsigned long lru_pages; 2812 2813 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | 2814 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); 2815 2816 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, 2817 sc.may_writepage, 2818 sc.gfp_mask); 2819 2820 /* 2821 * NOTE: Although we can get the priority field, using it 2822 * here is not a good idea, since it limits the pages we can scan. 2823 * if we don't reclaim here, the shrink_zone from balance_pgdat 2824 * will pick up pages from other mem cgroup's as well. We hack 2825 * the priority and make it zero. 2826 */ 2827 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages); 2828 2829 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); 2830 2831 *nr_scanned = sc.nr_scanned; 2832 return sc.nr_reclaimed; 2833} 2834 2835unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, 2836 unsigned long nr_pages, 2837 gfp_t gfp_mask, 2838 bool may_swap) 2839{ 2840 struct zonelist *zonelist; 2841 unsigned long nr_reclaimed; 2842 int nid; 2843 struct scan_control sc = { 2844 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), 2845 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | 2846 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), 2847 .target_mem_cgroup = memcg, 2848 .priority = DEF_PRIORITY, 2849 .may_writepage = !laptop_mode, 2850 .may_unmap = 1, 2851 .may_swap = may_swap, 2852 }; 2853 2854 /* 2855 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't 2856 * take care of from where we get pages. So the node where we start the 2857 * scan does not need to be the current node. 2858 */ 2859 nid = mem_cgroup_select_victim_node(memcg); 2860 2861 zonelist = NODE_DATA(nid)->node_zonelists; 2862 2863 trace_mm_vmscan_memcg_reclaim_begin(0, 2864 sc.may_writepage, 2865 sc.gfp_mask); 2866 2867 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 2868 2869 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); 2870 2871 return nr_reclaimed; 2872} 2873#endif 2874 2875static void age_active_anon(struct zone *zone, struct scan_control *sc) 2876{ 2877 struct mem_cgroup *memcg; 2878 2879 if (!total_swap_pages) 2880 return; 2881 2882 memcg = mem_cgroup_iter(NULL, NULL, NULL); 2883 do { 2884 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); 2885 2886 if (inactive_anon_is_low(lruvec)) 2887 shrink_active_list(SWAP_CLUSTER_MAX, lruvec, 2888 sc, LRU_ACTIVE_ANON); 2889 2890 memcg = mem_cgroup_iter(NULL, memcg, NULL); 2891 } while (memcg); 2892} 2893 2894static bool zone_balanced(struct zone *zone, int order, 2895 unsigned long balance_gap, int classzone_idx) 2896{ 2897 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) + 2898 balance_gap, classzone_idx, 0)) 2899 return false; 2900 2901 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone, 2902 order, 0, classzone_idx) == COMPACT_SKIPPED) 2903 return false; 2904 2905 return true; 2906} 2907 2908/* 2909 * pgdat_balanced() is used when checking if a node is balanced. 2910 * 2911 * For order-0, all zones must be balanced! 2912 * 2913 * For high-order allocations only zones that meet watermarks and are in a 2914 * zone allowed by the callers classzone_idx are added to balanced_pages. The 2915 * total of balanced pages must be at least 25% of the zones allowed by 2916 * classzone_idx for the node to be considered balanced. Forcing all zones to 2917 * be balanced for high orders can cause excessive reclaim when there are 2918 * imbalanced zones. 2919 * The choice of 25% is due to 2920 * o a 16M DMA zone that is balanced will not balance a zone on any 2921 * reasonable sized machine 2922 * o On all other machines, the top zone must be at least a reasonable 2923 * percentage of the middle zones. For example, on 32-bit x86, highmem 2924 * would need to be at least 256M for it to be balance a whole node. 2925 * Similarly, on x86-64 the Normal zone would need to be at least 1G 2926 * to balance a node on its own. These seemed like reasonable ratios. 2927 */ 2928static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) 2929{ 2930 unsigned long managed_pages = 0; 2931 unsigned long balanced_pages = 0; 2932 int i; 2933 2934 /* Check the watermark levels */ 2935 for (i = 0; i <= classzone_idx; i++) { 2936 struct zone *zone = pgdat->node_zones + i; 2937 2938 if (!populated_zone(zone)) 2939 continue; 2940 2941 managed_pages += zone->managed_pages; 2942 2943 /* 2944 * A special case here: 2945 * 2946 * balance_pgdat() skips over all_unreclaimable after 2947 * DEF_PRIORITY. Effectively, it considers them balanced so 2948 * they must be considered balanced here as well! 2949 */ 2950 if (!zone_reclaimable(zone)) { 2951 balanced_pages += zone->managed_pages; 2952 continue; 2953 } 2954 2955 if (zone_balanced(zone, order, 0, i)) 2956 balanced_pages += zone->managed_pages; 2957 else if (!order) 2958 return false; 2959 } 2960 2961 if (order) 2962 return balanced_pages >= (managed_pages >> 2); 2963 else 2964 return true; 2965} 2966 2967/* 2968 * Prepare kswapd for sleeping. This verifies that there are no processes 2969 * waiting in throttle_direct_reclaim() and that watermarks have been met. 2970 * 2971 * Returns true if kswapd is ready to sleep 2972 */ 2973static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining, 2974 int classzone_idx) 2975{ 2976 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ 2977 if (remaining) 2978 return false; 2979 2980 /* 2981 * The throttled processes are normally woken up in balance_pgdat() as 2982 * soon as pfmemalloc_watermark_ok() is true. But there is a potential 2983 * race between when kswapd checks the watermarks and a process gets 2984 * throttled. There is also a potential race if processes get 2985 * throttled, kswapd wakes, a large process exits thereby balancing the 2986 * zones, which causes kswapd to exit balance_pgdat() before reaching 2987 * the wake up checks. If kswapd is going to sleep, no process should 2988 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If 2989 * the wake up is premature, processes will wake kswapd and get 2990 * throttled again. The difference from wake ups in balance_pgdat() is 2991 * that here we are under prepare_to_wait(). 2992 */ 2993 if (waitqueue_active(&pgdat->pfmemalloc_wait)) 2994 wake_up_all(&pgdat->pfmemalloc_wait); 2995 2996 return pgdat_balanced(pgdat, order, classzone_idx); 2997} 2998 2999/* 3000 * kswapd shrinks the zone by the number of pages required to reach 3001 * the high watermark. 3002 * 3003 * Returns true if kswapd scanned at least the requested number of pages to 3004 * reclaim or if the lack of progress was due to pages under writeback. 3005 * This is used to determine if the scanning priority needs to be raised. 3006 */ 3007static bool kswapd_shrink_zone(struct zone *zone, 3008 int classzone_idx, 3009 struct scan_control *sc, 3010 unsigned long *nr_attempted) 3011{ 3012 int testorder = sc->order; 3013 unsigned long balance_gap; 3014 bool lowmem_pressure; 3015 3016 /* Reclaim above the high watermark. */ 3017 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone)); 3018 3019 /* 3020 * Kswapd reclaims only single pages with compaction enabled. Trying 3021 * too hard to reclaim until contiguous free pages have become 3022 * available can hurt performance by evicting too much useful data 3023 * from memory. Do not reclaim more than needed for compaction. 3024 */ 3025 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && 3026 compaction_suitable(zone, sc->order, 0, classzone_idx) 3027 != COMPACT_SKIPPED) 3028 testorder = 0; 3029 3030 /* 3031 * We put equal pressure on every zone, unless one zone has way too 3032 * many pages free already. The "too many pages" is defined as the 3033 * high wmark plus a "gap" where the gap is either the low 3034 * watermark or 1% of the zone, whichever is smaller. 3035 */ 3036 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( 3037 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); 3038 3039 /* 3040 * If there is no low memory pressure or the zone is balanced then no 3041 * reclaim is necessary 3042 */ 3043 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone)); 3044 if (!lowmem_pressure && zone_balanced(zone, testorder, 3045 balance_gap, classzone_idx)) 3046 return true; 3047 3048 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx); 3049 3050 /* Account for the number of pages attempted to reclaim */ 3051 *nr_attempted += sc->nr_to_reclaim; 3052 3053 clear_bit(ZONE_WRITEBACK, &zone->flags); 3054 3055 /* 3056 * If a zone reaches its high watermark, consider it to be no longer 3057 * congested. It's possible there are dirty pages backed by congested 3058 * BDIs but as pressure is relieved, speculatively avoid congestion 3059 * waits. 3060 */ 3061 if (zone_reclaimable(zone) && 3062 zone_balanced(zone, testorder, 0, classzone_idx)) { 3063 clear_bit(ZONE_CONGESTED, &zone->flags); 3064 clear_bit(ZONE_DIRTY, &zone->flags); 3065 } 3066 3067 return sc->nr_scanned >= sc->nr_to_reclaim; 3068} 3069 3070/* 3071 * For kswapd, balance_pgdat() will work across all this node's zones until 3072 * they are all at high_wmark_pages(zone). 3073 * 3074 * Returns the final order kswapd was reclaiming at 3075 * 3076 * There is special handling here for zones which are full of pinned pages. 3077 * This can happen if the pages are all mlocked, or if they are all used by 3078 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. 3079 * What we do is to detect the case where all pages in the zone have been 3080 * scanned twice and there has been zero successful reclaim. Mark the zone as 3081 * dead and from now on, only perform a short scan. Basically we're polling 3082 * the zone for when the problem goes away. 3083 * 3084 * kswapd scans the zones in the highmem->normal->dma direction. It skips 3085 * zones which have free_pages > high_wmark_pages(zone), but once a zone is 3086 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the 3087 * lower zones regardless of the number of free pages in the lower zones. This 3088 * interoperates with the page allocator fallback scheme to ensure that aging 3089 * of pages is balanced across the zones. 3090 */ 3091static unsigned long balance_pgdat(pg_data_t *pgdat, int order, 3092 int *classzone_idx) 3093{ 3094 int i; 3095 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ 3096 unsigned long nr_soft_reclaimed; 3097 unsigned long nr_soft_scanned; 3098 struct scan_control sc = { 3099 .gfp_mask = GFP_KERNEL, 3100 .order = order, 3101 .priority = DEF_PRIORITY, 3102 .may_writepage = !laptop_mode, 3103 .may_unmap = 1, 3104 .may_swap = 1, 3105 }; 3106 count_vm_event(PAGEOUTRUN); 3107 3108 do { 3109 unsigned long nr_attempted = 0; 3110 bool raise_priority = true; 3111 bool pgdat_needs_compaction = (order > 0); 3112 3113 sc.nr_reclaimed = 0; 3114 3115 /* 3116 * Scan in the highmem->dma direction for the highest 3117 * zone which needs scanning 3118 */ 3119 for (i = pgdat->nr_zones - 1; i >= 0; i--) { 3120 struct zone *zone = pgdat->node_zones + i; 3121 3122 if (!populated_zone(zone)) 3123 continue; 3124 3125 if (sc.priority != DEF_PRIORITY && 3126 !zone_reclaimable(zone)) 3127 continue; 3128 3129 /* 3130 * Do some background aging of the anon list, to give 3131 * pages a chance to be referenced before reclaiming. 3132 */ 3133 age_active_anon(zone, &sc); 3134 3135 /* 3136 * If the number of buffer_heads in the machine 3137 * exceeds the maximum allowed level and this node 3138 * has a highmem zone, force kswapd to reclaim from 3139 * it to relieve lowmem pressure. 3140 */ 3141 if (buffer_heads_over_limit && is_highmem_idx(i)) { 3142 end_zone = i; 3143 break; 3144 } 3145 3146 if (!zone_balanced(zone, order, 0, 0)) { 3147 end_zone = i; 3148 break; 3149 } else { 3150 /* 3151 * If balanced, clear the dirty and congested 3152 * flags 3153 */ 3154 clear_bit(ZONE_CONGESTED, &zone->flags); 3155 clear_bit(ZONE_DIRTY, &zone->flags); 3156 } 3157 } 3158 3159 if (i < 0) 3160 goto out; 3161 3162 for (i = 0; i <= end_zone; i++) { 3163 struct zone *zone = pgdat->node_zones + i; 3164 3165 if (!populated_zone(zone)) 3166 continue; 3167 3168 /* 3169 * If any zone is currently balanced then kswapd will 3170 * not call compaction as it is expected that the 3171 * necessary pages are already available. 3172 */ 3173 if (pgdat_needs_compaction && 3174 zone_watermark_ok(zone, order, 3175 low_wmark_pages(zone), 3176 *classzone_idx, 0)) 3177 pgdat_needs_compaction = false; 3178 } 3179 3180 /* 3181 * If we're getting trouble reclaiming, start doing writepage 3182 * even in laptop mode. 3183 */ 3184 if (sc.priority < DEF_PRIORITY - 2) 3185 sc.may_writepage = 1; 3186 3187 /* 3188 * Now scan the zone in the dma->highmem direction, stopping 3189 * at the last zone which needs scanning. 3190 * 3191 * We do this because the page allocator works in the opposite 3192 * direction. This prevents the page allocator from allocating 3193 * pages behind kswapd's direction of progress, which would 3194 * cause too much scanning of the lower zones. 3195 */ 3196 for (i = 0; i <= end_zone; i++) { 3197 struct zone *zone = pgdat->node_zones + i; 3198 3199 if (!populated_zone(zone)) 3200 continue; 3201 3202 if (sc.priority != DEF_PRIORITY && 3203 !zone_reclaimable(zone)) 3204 continue; 3205 3206 sc.nr_scanned = 0; 3207 3208 nr_soft_scanned = 0; 3209 /* 3210 * Call soft limit reclaim before calling shrink_zone. 3211 */ 3212 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, 3213 order, sc.gfp_mask, 3214 &nr_soft_scanned); 3215 sc.nr_reclaimed += nr_soft_reclaimed; 3216 3217 /* 3218 * There should be no need to raise the scanning 3219 * priority if enough pages are already being scanned 3220 * that that high watermark would be met at 100% 3221 * efficiency. 3222 */ 3223 if (kswapd_shrink_zone(zone, end_zone, 3224 &sc, &nr_attempted)) 3225 raise_priority = false; 3226 } 3227 3228 /* 3229 * If the low watermark is met there is no need for processes 3230 * to be throttled on pfmemalloc_wait as they should not be 3231 * able to safely make forward progress. Wake them 3232 */ 3233 if (waitqueue_active(&pgdat->pfmemalloc_wait) && 3234 pfmemalloc_watermark_ok(pgdat)) 3235 wake_up_all(&pgdat->pfmemalloc_wait); 3236 3237 /* 3238 * Fragmentation may mean that the system cannot be rebalanced 3239 * for high-order allocations in all zones. If twice the 3240 * allocation size has been reclaimed and the zones are still 3241 * not balanced then recheck the watermarks at order-0 to 3242 * prevent kswapd reclaiming excessively. Assume that a 3243 * process requested a high-order can direct reclaim/compact. 3244 */ 3245 if (order && sc.nr_reclaimed >= 2UL << order) 3246 order = sc.order = 0; 3247 3248 /* Check if kswapd should be suspending */ 3249 if (try_to_freeze() || kthread_should_stop()) 3250 break; 3251 3252 /* 3253 * Compact if necessary and kswapd is reclaiming at least the 3254 * high watermark number of pages as requsted 3255 */ 3256 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted) 3257 compact_pgdat(pgdat, order); 3258 3259 /* 3260 * Raise priority if scanning rate is too low or there was no 3261 * progress in reclaiming pages 3262 */ 3263 if (raise_priority || !sc.nr_reclaimed) 3264 sc.priority--; 3265 } while (sc.priority >= 1 && 3266 !pgdat_balanced(pgdat, order, *classzone_idx)); 3267 3268out: 3269 /* 3270 * Return the order we were reclaiming at so prepare_kswapd_sleep() 3271 * makes a decision on the order we were last reclaiming at. However, 3272 * if another caller entered the allocator slow path while kswapd 3273 * was awake, order will remain at the higher level 3274 */ 3275 *classzone_idx = end_zone; 3276 return order; 3277} 3278 3279static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) 3280{ 3281 long remaining = 0; 3282 DEFINE_WAIT(wait); 3283 3284 if (freezing(current) || kthread_should_stop()) 3285 return; 3286 3287 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 3288 3289 /* Try to sleep for a short interval */ 3290 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { 3291 remaining = schedule_timeout(HZ/10); 3292 finish_wait(&pgdat->kswapd_wait, &wait); 3293 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 3294 } 3295 3296 /* 3297 * After a short sleep, check if it was a premature sleep. If not, then 3298 * go fully to sleep until explicitly woken up. 3299 */ 3300 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { 3301 trace_mm_vmscan_kswapd_sleep(pgdat->node_id); 3302 3303 /* 3304 * vmstat counters are not perfectly accurate and the estimated 3305 * value for counters such as NR_FREE_PAGES can deviate from the 3306 * true value by nr_online_cpus * threshold. To avoid the zone 3307 * watermarks being breached while under pressure, we reduce the 3308 * per-cpu vmstat threshold while kswapd is awake and restore 3309 * them before going back to sleep. 3310 */ 3311 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); 3312 3313 /* 3314 * Compaction records what page blocks it recently failed to 3315 * isolate pages from and skips them in the future scanning. 3316 * When kswapd is going to sleep, it is reasonable to assume 3317 * that pages and compaction may succeed so reset the cache. 3318 */ 3319 reset_isolation_suitable(pgdat); 3320 3321 if (!kthread_should_stop()) 3322 schedule(); 3323 3324 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); 3325 } else { 3326 if (remaining) 3327 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); 3328 else 3329 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); 3330 } 3331 finish_wait(&pgdat->kswapd_wait, &wait); 3332} 3333 3334/* 3335 * The background pageout daemon, started as a kernel thread 3336 * from the init process. 3337 * 3338 * This basically trickles out pages so that we have _some_ 3339 * free memory available even if there is no other activity 3340 * that frees anything up. This is needed for things like routing 3341 * etc, where we otherwise might have all activity going on in 3342 * asynchronous contexts that cannot page things out. 3343 * 3344 * If there are applications that are active memory-allocators 3345 * (most normal use), this basically shouldn't matter. 3346 */ 3347static int kswapd(void *p) 3348{ 3349 unsigned long order, new_order; 3350 unsigned balanced_order; 3351 int classzone_idx, new_classzone_idx; 3352 int balanced_classzone_idx; 3353 pg_data_t *pgdat = (pg_data_t*)p; 3354 struct task_struct *tsk = current; 3355 3356 struct reclaim_state reclaim_state = { 3357 .reclaimed_slab = 0, 3358 }; 3359 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); 3360 3361 lockdep_set_current_reclaim_state(GFP_KERNEL); 3362 3363 if (!cpumask_empty(cpumask)) 3364 set_cpus_allowed_ptr(tsk, cpumask); 3365 current->reclaim_state = &reclaim_state; 3366 3367 /* 3368 * Tell the memory management that we're a "memory allocator", 3369 * and that if we need more memory we should get access to it 3370 * regardless (see "__alloc_pages()"). "kswapd" should 3371 * never get caught in the normal page freeing logic. 3372 * 3373 * (Kswapd normally doesn't need memory anyway, but sometimes 3374 * you need a small amount of memory in order to be able to 3375 * page out something else, and this flag essentially protects 3376 * us from recursively trying to free more memory as we're 3377 * trying to free the first piece of memory in the first place). 3378 */ 3379 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; 3380 set_freezable(); 3381 3382 order = new_order = 0; 3383 balanced_order = 0; 3384 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; 3385 balanced_classzone_idx = classzone_idx; 3386 for ( ; ; ) { 3387 bool ret; 3388 3389 /* 3390 * If the last balance_pgdat was unsuccessful it's unlikely a 3391 * new request of a similar or harder type will succeed soon 3392 * so consider going to sleep on the basis we reclaimed at 3393 */ 3394 if (balanced_classzone_idx >= new_classzone_idx && 3395 balanced_order == new_order) { 3396 new_order = pgdat->kswapd_max_order; 3397 new_classzone_idx = pgdat->classzone_idx; 3398 pgdat->kswapd_max_order = 0; 3399 pgdat->classzone_idx = pgdat->nr_zones - 1; 3400 } 3401 3402 if (order < new_order || classzone_idx > new_classzone_idx) { 3403 /* 3404 * Don't sleep if someone wants a larger 'order' 3405 * allocation or has tigher zone constraints 3406 */ 3407 order = new_order; 3408 classzone_idx = new_classzone_idx; 3409 } else { 3410 kswapd_try_to_sleep(pgdat, balanced_order, 3411 balanced_classzone_idx); 3412 order = pgdat->kswapd_max_order; 3413 classzone_idx = pgdat->classzone_idx; 3414 new_order = order; 3415 new_classzone_idx = classzone_idx; 3416 pgdat->kswapd_max_order = 0; 3417 pgdat->classzone_idx = pgdat->nr_zones - 1; 3418 } 3419 3420 ret = try_to_freeze(); 3421 if (kthread_should_stop()) 3422 break; 3423 3424 /* 3425 * We can speed up thawing tasks if we don't call balance_pgdat 3426 * after returning from the refrigerator 3427 */ 3428 if (!ret) { 3429 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); 3430 balanced_classzone_idx = classzone_idx; 3431 balanced_order = balance_pgdat(pgdat, order, 3432 &balanced_classzone_idx); 3433 } 3434 } 3435 3436 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); 3437 current->reclaim_state = NULL; 3438 lockdep_clear_current_reclaim_state(); 3439 3440 return 0; 3441} 3442 3443/* 3444 * A zone is low on free memory, so wake its kswapd task to service it. 3445 */ 3446void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) 3447{ 3448 pg_data_t *pgdat; 3449 3450 if (!populated_zone(zone)) 3451 return; 3452 3453 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) 3454 return; 3455 pgdat = zone->zone_pgdat; 3456 if (pgdat->kswapd_max_order < order) { 3457 pgdat->kswapd_max_order = order; 3458 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); 3459 } 3460 if (!waitqueue_active(&pgdat->kswapd_wait)) 3461 return; 3462 if (zone_balanced(zone, order, 0, 0)) 3463 return; 3464 3465 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); 3466 wake_up_interruptible(&pgdat->kswapd_wait); 3467} 3468 3469#ifdef CONFIG_HIBERNATION 3470/* 3471 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of 3472 * freed pages. 3473 * 3474 * Rather than trying to age LRUs the aim is to preserve the overall 3475 * LRU order by reclaiming preferentially 3476 * inactive > active > active referenced > active mapped 3477 */ 3478unsigned long shrink_all_memory(unsigned long nr_to_reclaim) 3479{ 3480 struct reclaim_state reclaim_state; 3481 struct scan_control sc = { 3482 .nr_to_reclaim = nr_to_reclaim, 3483 .gfp_mask = GFP_HIGHUSER_MOVABLE, 3484 .priority = DEF_PRIORITY, 3485 .may_writepage = 1, 3486 .may_unmap = 1, 3487 .may_swap = 1, 3488 .hibernation_mode = 1, 3489 }; 3490 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); 3491 struct task_struct *p = current; 3492 unsigned long nr_reclaimed; 3493 3494 p->flags |= PF_MEMALLOC; 3495 lockdep_set_current_reclaim_state(sc.gfp_mask); 3496 reclaim_state.reclaimed_slab = 0; 3497 p->reclaim_state = &reclaim_state; 3498 3499 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 3500 3501 p->reclaim_state = NULL; 3502 lockdep_clear_current_reclaim_state(); 3503 p->flags &= ~PF_MEMALLOC; 3504 3505 return nr_reclaimed; 3506} 3507#endif /* CONFIG_HIBERNATION */ 3508 3509/* It's optimal to keep kswapds on the same CPUs as their memory, but 3510 not required for correctness. So if the last cpu in a node goes 3511 away, we get changed to run anywhere: as the first one comes back, 3512 restore their cpu bindings. */ 3513static int cpu_callback(struct notifier_block *nfb, unsigned long action, 3514 void *hcpu) 3515{ 3516 int nid; 3517 3518 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { 3519 for_each_node_state(nid, N_MEMORY) { 3520 pg_data_t *pgdat = NODE_DATA(nid); 3521 const struct cpumask *mask; 3522 3523 mask = cpumask_of_node(pgdat->node_id); 3524 3525 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) 3526 /* One of our CPUs online: restore mask */ 3527 set_cpus_allowed_ptr(pgdat->kswapd, mask); 3528 } 3529 } 3530 return NOTIFY_OK; 3531} 3532 3533/* 3534 * This kswapd start function will be called by init and node-hot-add. 3535 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. 3536 */ 3537int kswapd_run(int nid) 3538{ 3539 pg_data_t *pgdat = NODE_DATA(nid); 3540 int ret = 0; 3541 3542 if (pgdat->kswapd) 3543 return 0; 3544 3545 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); 3546 if (IS_ERR(pgdat->kswapd)) { 3547 /* failure at boot is fatal */ 3548 BUG_ON(system_state == SYSTEM_BOOTING); 3549 pr_err("Failed to start kswapd on node %d\n", nid); 3550 ret = PTR_ERR(pgdat->kswapd); 3551 pgdat->kswapd = NULL; 3552 } 3553 return ret; 3554} 3555 3556/* 3557 * Called by memory hotplug when all memory in a node is offlined. Caller must 3558 * hold mem_hotplug_begin/end(). 3559 */ 3560void kswapd_stop(int nid) 3561{ 3562 struct task_struct *kswapd = NODE_DATA(nid)->kswapd; 3563 3564 if (kswapd) { 3565 kthread_stop(kswapd); 3566 NODE_DATA(nid)->kswapd = NULL; 3567 } 3568} 3569 3570static int __init kswapd_init(void) 3571{ 3572 int nid; 3573 3574 swap_setup(); 3575 for_each_node_state(nid, N_MEMORY) 3576 kswapd_run(nid); 3577 hotcpu_notifier(cpu_callback, 0); 3578 return 0; 3579} 3580 3581module_init(kswapd_init) 3582 3583#ifdef CONFIG_NUMA 3584/* 3585 * Zone reclaim mode 3586 * 3587 * If non-zero call zone_reclaim when the number of free pages falls below 3588 * the watermarks. 3589 */ 3590int zone_reclaim_mode __read_mostly; 3591 3592#define RECLAIM_OFF 0 3593#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ 3594#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ 3595#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ 3596 3597/* 3598 * Priority for ZONE_RECLAIM. This determines the fraction of pages 3599 * of a node considered for each zone_reclaim. 4 scans 1/16th of 3600 * a zone. 3601 */ 3602#define ZONE_RECLAIM_PRIORITY 4 3603 3604/* 3605 * Percentage of pages in a zone that must be unmapped for zone_reclaim to 3606 * occur. 3607 */ 3608int sysctl_min_unmapped_ratio = 1; 3609 3610/* 3611 * If the number of slab pages in a zone grows beyond this percentage then 3612 * slab reclaim needs to occur. 3613 */ 3614int sysctl_min_slab_ratio = 5; 3615 3616static inline unsigned long zone_unmapped_file_pages(struct zone *zone) 3617{ 3618 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); 3619 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + 3620 zone_page_state(zone, NR_ACTIVE_FILE); 3621 3622 /* 3623 * It's possible for there to be more file mapped pages than 3624 * accounted for by the pages on the file LRU lists because 3625 * tmpfs pages accounted for as ANON can also be FILE_MAPPED 3626 */ 3627 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; 3628} 3629 3630/* Work out how many page cache pages we can reclaim in this reclaim_mode */ 3631static long zone_pagecache_reclaimable(struct zone *zone) 3632{ 3633 long nr_pagecache_reclaimable; 3634 long delta = 0; 3635 3636 /* 3637 * If RECLAIM_SWAP is set, then all file pages are considered 3638 * potentially reclaimable. Otherwise, we have to worry about 3639 * pages like swapcache and zone_unmapped_file_pages() provides 3640 * a better estimate 3641 */ 3642 if (zone_reclaim_mode & RECLAIM_SWAP) 3643 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); 3644 else 3645 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); 3646 3647 /* If we can't clean pages, remove dirty pages from consideration */ 3648 if (!(zone_reclaim_mode & RECLAIM_WRITE)) 3649 delta += zone_page_state(zone, NR_FILE_DIRTY); 3650 3651 /* Watch for any possible underflows due to delta */ 3652 if (unlikely(delta > nr_pagecache_reclaimable)) 3653 delta = nr_pagecache_reclaimable; 3654 3655 return nr_pagecache_reclaimable - delta; 3656} 3657 3658/* 3659 * Try to free up some pages from this zone through reclaim. 3660 */ 3661static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 3662{ 3663 /* Minimum pages needed in order to stay on node */ 3664 const unsigned long nr_pages = 1 << order; 3665 struct task_struct *p = current; 3666 struct reclaim_state reclaim_state; 3667 struct scan_control sc = { 3668 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), 3669 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), 3670 .order = order, 3671 .priority = ZONE_RECLAIM_PRIORITY, 3672 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), 3673 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP), 3674 .may_swap = 1, 3675 }; 3676 3677 cond_resched(); 3678 /* 3679 * We need to be able to allocate from the reserves for RECLAIM_SWAP 3680 * and we also need to be able to write out pages for RECLAIM_WRITE 3681 * and RECLAIM_SWAP. 3682 */ 3683 p->flags |= PF_MEMALLOC | PF_SWAPWRITE; 3684 lockdep_set_current_reclaim_state(gfp_mask); 3685 reclaim_state.reclaimed_slab = 0; 3686 p->reclaim_state = &reclaim_state; 3687 3688 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { 3689 /* 3690 * Free memory by calling shrink zone with increasing 3691 * priorities until we have enough memory freed. 3692 */ 3693 do { 3694 shrink_zone(zone, &sc, true); 3695 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); 3696 } 3697 3698 p->reclaim_state = NULL; 3699 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); 3700 lockdep_clear_current_reclaim_state(); 3701 return sc.nr_reclaimed >= nr_pages; 3702} 3703 3704int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 3705{ 3706 int node_id; 3707 int ret; 3708 3709 /* 3710 * Zone reclaim reclaims unmapped file backed pages and 3711 * slab pages if we are over the defined limits. 3712 * 3713 * A small portion of unmapped file backed pages is needed for 3714 * file I/O otherwise pages read by file I/O will be immediately 3715 * thrown out if the zone is overallocated. So we do not reclaim 3716 * if less than a specified percentage of the zone is used by 3717 * unmapped file backed pages. 3718 */ 3719 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && 3720 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) 3721 return ZONE_RECLAIM_FULL; 3722 3723 if (!zone_reclaimable(zone)) 3724 return ZONE_RECLAIM_FULL; 3725 3726 /* 3727 * Do not scan if the allocation should not be delayed. 3728 */ 3729 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) 3730 return ZONE_RECLAIM_NOSCAN; 3731 3732 /* 3733 * Only run zone reclaim on the local zone or on zones that do not 3734 * have associated processors. This will favor the local processor 3735 * over remote processors and spread off node memory allocations 3736 * as wide as possible. 3737 */ 3738 node_id = zone_to_nid(zone); 3739 if (node_state(node_id, N_CPU) && node_id != numa_node_id()) 3740 return ZONE_RECLAIM_NOSCAN; 3741 3742 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags)) 3743 return ZONE_RECLAIM_NOSCAN; 3744 3745 ret = __zone_reclaim(zone, gfp_mask, order); 3746 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags); 3747 3748 if (!ret) 3749 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); 3750 3751 return ret; 3752} 3753#endif 3754 3755/* 3756 * page_evictable - test whether a page is evictable 3757 * @page: the page to test 3758 * 3759 * Test whether page is evictable--i.e., should be placed on active/inactive 3760 * lists vs unevictable list. 3761 * 3762 * Reasons page might not be evictable: 3763 * (1) page's mapping marked unevictable 3764 * (2) page is part of an mlocked VMA 3765 * 3766 */ 3767int page_evictable(struct page *page) 3768{ 3769 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); 3770} 3771 3772#ifdef CONFIG_SHMEM 3773/** 3774 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list 3775 * @pages: array of pages to check 3776 * @nr_pages: number of pages to check 3777 * 3778 * Checks pages for evictability and moves them to the appropriate lru list. 3779 * 3780 * This function is only used for SysV IPC SHM_UNLOCK. 3781 */ 3782void check_move_unevictable_pages(struct page **pages, int nr_pages) 3783{ 3784 struct lruvec *lruvec; 3785 struct zone *zone = NULL; 3786 int pgscanned = 0; 3787 int pgrescued = 0; 3788 int i; 3789 3790 for (i = 0; i < nr_pages; i++) { 3791 struct page *page = pages[i]; 3792 struct zone *pagezone; 3793 3794 pgscanned++; 3795 pagezone = page_zone(page); 3796 if (pagezone != zone) { 3797 if (zone) 3798 spin_unlock_irq(&zone->lru_lock); 3799 zone = pagezone; 3800 spin_lock_irq(&zone->lru_lock); 3801 } 3802 lruvec = mem_cgroup_page_lruvec(page, zone); 3803 3804 if (!PageLRU(page) || !PageUnevictable(page)) 3805 continue; 3806 3807 if (page_evictable(page)) { 3808 enum lru_list lru = page_lru_base_type(page); 3809 3810 VM_BUG_ON_PAGE(PageActive(page), page); 3811 ClearPageUnevictable(page); 3812 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); 3813 add_page_to_lru_list(page, lruvec, lru); 3814 pgrescued++; 3815 } 3816 } 3817 3818 if (zone) { 3819 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); 3820 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); 3821 spin_unlock_irq(&zone->lru_lock); 3822 } 3823} 3824#endif /* CONFIG_SHMEM */ 3825