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 61 struct 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 */ 142 int vm_swappiness = 60; 143 /* 144 * The total number of pages which are beyond the high watermark within all 145 * zones. 146 */ 147 unsigned long vm_total_pages; 148 149 static LIST_HEAD(shrinker_list); 150 static DECLARE_RWSEM(shrinker_rwsem); 151 152 #ifdef CONFIG_MEMCG 153 static bool global_reclaim(struct scan_control *sc) 154 { 155 return !sc->target_mem_cgroup; 156 } 157 158 /** 159 * sane_reclaim - is the usual dirty throttling mechanism operational? 160 * @sc: scan_control in question 161 * 162 * The normal page dirty throttling mechanism in balance_dirty_pages() is 163 * completely broken with the legacy memcg and direct stalling in 164 * shrink_page_list() is used for throttling instead, which lacks all the 165 * niceties such as fairness, adaptive pausing, bandwidth proportional 166 * allocation and configurability. 167 * 168 * This function tests whether the vmscan currently in progress can assume 169 * that the normal dirty throttling mechanism is operational. 170 */ 171 static bool sane_reclaim(struct scan_control *sc) 172 { 173 struct mem_cgroup *memcg = sc->target_mem_cgroup; 174 175 if (!memcg) 176 return true; 177 #ifdef CONFIG_CGROUP_WRITEBACK 178 if (cgroup_on_dfl(mem_cgroup_css(memcg)->cgroup)) 179 return true; 180 #endif 181 return false; 182 } 183 #else 184 static bool global_reclaim(struct scan_control *sc) 185 { 186 return true; 187 } 188 189 static bool sane_reclaim(struct scan_control *sc) 190 { 191 return true; 192 } 193 #endif 194 195 static unsigned long zone_reclaimable_pages(struct zone *zone) 196 { 197 int nr; 198 199 nr = zone_page_state(zone, NR_ACTIVE_FILE) + 200 zone_page_state(zone, NR_INACTIVE_FILE); 201 202 if (get_nr_swap_pages() > 0) 203 nr += zone_page_state(zone, NR_ACTIVE_ANON) + 204 zone_page_state(zone, NR_INACTIVE_ANON); 205 206 return nr; 207 } 208 209 bool zone_reclaimable(struct zone *zone) 210 { 211 return zone_page_state(zone, NR_PAGES_SCANNED) < 212 zone_reclaimable_pages(zone) * 6; 213 } 214 215 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru) 216 { 217 if (!mem_cgroup_disabled()) 218 return mem_cgroup_get_lru_size(lruvec, lru); 219 220 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru); 221 } 222 223 /* 224 * Add a shrinker callback to be called from the vm. 225 */ 226 int register_shrinker(struct shrinker *shrinker) 227 { 228 size_t size = sizeof(*shrinker->nr_deferred); 229 230 /* 231 * If we only have one possible node in the system anyway, save 232 * ourselves the trouble and disable NUMA aware behavior. This way we 233 * will save memory and some small loop time later. 234 */ 235 if (nr_node_ids == 1) 236 shrinker->flags &= ~SHRINKER_NUMA_AWARE; 237 238 if (shrinker->flags & SHRINKER_NUMA_AWARE) 239 size *= nr_node_ids; 240 241 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); 242 if (!shrinker->nr_deferred) 243 return -ENOMEM; 244 245 down_write(&shrinker_rwsem); 246 list_add_tail(&shrinker->list, &shrinker_list); 247 up_write(&shrinker_rwsem); 248 return 0; 249 } 250 EXPORT_SYMBOL(register_shrinker); 251 252 /* 253 * Remove one 254 */ 255 void unregister_shrinker(struct shrinker *shrinker) 256 { 257 down_write(&shrinker_rwsem); 258 list_del(&shrinker->list); 259 up_write(&shrinker_rwsem); 260 kfree(shrinker->nr_deferred); 261 } 262 EXPORT_SYMBOL(unregister_shrinker); 263 264 #define SHRINK_BATCH 128 265 266 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, 267 struct shrinker *shrinker, 268 unsigned long nr_scanned, 269 unsigned long nr_eligible) 270 { 271 unsigned long freed = 0; 272 unsigned long long delta; 273 long total_scan; 274 long freeable; 275 long nr; 276 long new_nr; 277 int nid = shrinkctl->nid; 278 long batch_size = shrinker->batch ? shrinker->batch 279 : SHRINK_BATCH; 280 281 freeable = shrinker->count_objects(shrinker, shrinkctl); 282 if (freeable == 0) 283 return 0; 284 285 /* 286 * copy the current shrinker scan count into a local variable 287 * and zero it so that other concurrent shrinker invocations 288 * don't also do this scanning work. 289 */ 290 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); 291 292 total_scan = nr; 293 delta = (4 * nr_scanned) / shrinker->seeks; 294 delta *= freeable; 295 do_div(delta, nr_eligible + 1); 296 total_scan += delta; 297 if (total_scan < 0) { 298 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", 299 shrinker->scan_objects, total_scan); 300 total_scan = freeable; 301 } 302 303 /* 304 * We need to avoid excessive windup on filesystem shrinkers 305 * due to large numbers of GFP_NOFS allocations causing the 306 * shrinkers to return -1 all the time. This results in a large 307 * nr being built up so when a shrink that can do some work 308 * comes along it empties the entire cache due to nr >>> 309 * freeable. This is bad for sustaining a working set in 310 * memory. 311 * 312 * Hence only allow the shrinker to scan the entire cache when 313 * a large delta change is calculated directly. 314 */ 315 if (delta < freeable / 4) 316 total_scan = min(total_scan, freeable / 2); 317 318 /* 319 * Avoid risking looping forever due to too large nr value: 320 * never try to free more than twice the estimate number of 321 * freeable entries. 322 */ 323 if (total_scan > freeable * 2) 324 total_scan = freeable * 2; 325 326 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, 327 nr_scanned, nr_eligible, 328 freeable, delta, total_scan); 329 330 /* 331 * Normally, we should not scan less than batch_size objects in one 332 * pass to avoid too frequent shrinker calls, but if the slab has less 333 * than batch_size objects in total and we are really tight on memory, 334 * we will try to reclaim all available objects, otherwise we can end 335 * up failing allocations although there are plenty of reclaimable 336 * objects spread over several slabs with usage less than the 337 * batch_size. 338 * 339 * We detect the "tight on memory" situations by looking at the total 340 * number of objects we want to scan (total_scan). If it is greater 341 * than the total number of objects on slab (freeable), we must be 342 * scanning at high prio and therefore should try to reclaim as much as 343 * possible. 344 */ 345 while (total_scan >= batch_size || 346 total_scan >= freeable) { 347 unsigned long ret; 348 unsigned long nr_to_scan = min(batch_size, total_scan); 349 350 shrinkctl->nr_to_scan = nr_to_scan; 351 ret = shrinker->scan_objects(shrinker, shrinkctl); 352 if (ret == SHRINK_STOP) 353 break; 354 freed += ret; 355 356 count_vm_events(SLABS_SCANNED, nr_to_scan); 357 total_scan -= nr_to_scan; 358 359 cond_resched(); 360 } 361 362 /* 363 * move the unused scan count back into the shrinker in a 364 * manner that handles concurrent updates. If we exhausted the 365 * scan, there is no need to do an update. 366 */ 367 if (total_scan > 0) 368 new_nr = atomic_long_add_return(total_scan, 369 &shrinker->nr_deferred[nid]); 370 else 371 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); 372 373 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); 374 return freed; 375 } 376 377 /** 378 * shrink_slab - shrink slab caches 379 * @gfp_mask: allocation context 380 * @nid: node whose slab caches to target 381 * @memcg: memory cgroup whose slab caches to target 382 * @nr_scanned: pressure numerator 383 * @nr_eligible: pressure denominator 384 * 385 * Call the shrink functions to age shrinkable caches. 386 * 387 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, 388 * unaware shrinkers will receive a node id of 0 instead. 389 * 390 * @memcg specifies the memory cgroup to target. If it is not NULL, 391 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan 392 * objects from the memory cgroup specified. Otherwise all shrinkers 393 * are called, and memcg aware shrinkers are supposed to scan the 394 * global list then. 395 * 396 * @nr_scanned and @nr_eligible form a ratio that indicate how much of 397 * the available objects should be scanned. Page reclaim for example 398 * passes the number of pages scanned and the number of pages on the 399 * LRU lists that it considered on @nid, plus a bias in @nr_scanned 400 * when it encountered mapped pages. The ratio is further biased by 401 * the ->seeks setting of the shrink function, which indicates the 402 * cost to recreate an object relative to that of an LRU page. 403 * 404 * Returns the number of reclaimed slab objects. 405 */ 406 static unsigned long shrink_slab(gfp_t gfp_mask, int nid, 407 struct mem_cgroup *memcg, 408 unsigned long nr_scanned, 409 unsigned long nr_eligible) 410 { 411 struct shrinker *shrinker; 412 unsigned long freed = 0; 413 414 if (memcg && !memcg_kmem_is_active(memcg)) 415 return 0; 416 417 if (nr_scanned == 0) 418 nr_scanned = SWAP_CLUSTER_MAX; 419 420 if (!down_read_trylock(&shrinker_rwsem)) { 421 /* 422 * If we would return 0, our callers would understand that we 423 * have nothing else to shrink and give up trying. By returning 424 * 1 we keep it going and assume we'll be able to shrink next 425 * time. 426 */ 427 freed = 1; 428 goto out; 429 } 430 431 list_for_each_entry(shrinker, &shrinker_list, list) { 432 struct shrink_control sc = { 433 .gfp_mask = gfp_mask, 434 .nid = nid, 435 .memcg = memcg, 436 }; 437 438 if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE)) 439 continue; 440 441 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) 442 sc.nid = 0; 443 444 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible); 445 } 446 447 up_read(&shrinker_rwsem); 448 out: 449 cond_resched(); 450 return freed; 451 } 452 453 void drop_slab_node(int nid) 454 { 455 unsigned long freed; 456 457 do { 458 struct mem_cgroup *memcg = NULL; 459 460 freed = 0; 461 do { 462 freed += shrink_slab(GFP_KERNEL, nid, memcg, 463 1000, 1000); 464 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); 465 } while (freed > 10); 466 } 467 468 void drop_slab(void) 469 { 470 int nid; 471 472 for_each_online_node(nid) 473 drop_slab_node(nid); 474 } 475 476 static inline int is_page_cache_freeable(struct page *page) 477 { 478 /* 479 * A freeable page cache page is referenced only by the caller 480 * that isolated the page, the page cache radix tree and 481 * optional buffer heads at page->private. 482 */ 483 return page_count(page) - page_has_private(page) == 2; 484 } 485 486 static int may_write_to_inode(struct inode *inode, struct scan_control *sc) 487 { 488 if (current->flags & PF_SWAPWRITE) 489 return 1; 490 if (!inode_write_congested(inode)) 491 return 1; 492 if (inode_to_bdi(inode) == current->backing_dev_info) 493 return 1; 494 return 0; 495 } 496 497 /* 498 * We detected a synchronous write error writing a page out. Probably 499 * -ENOSPC. We need to propagate that into the address_space for a subsequent 500 * fsync(), msync() or close(). 501 * 502 * The tricky part is that after writepage we cannot touch the mapping: nothing 503 * prevents it from being freed up. But we have a ref on the page and once 504 * that page is locked, the mapping is pinned. 505 * 506 * We're allowed to run sleeping lock_page() here because we know the caller has 507 * __GFP_FS. 508 */ 509 static void handle_write_error(struct address_space *mapping, 510 struct page *page, int error) 511 { 512 lock_page(page); 513 if (page_mapping(page) == mapping) 514 mapping_set_error(mapping, error); 515 unlock_page(page); 516 } 517 518 /* possible outcome of pageout() */ 519 typedef enum { 520 /* failed to write page out, page is locked */ 521 PAGE_KEEP, 522 /* move page to the active list, page is locked */ 523 PAGE_ACTIVATE, 524 /* page has been sent to the disk successfully, page is unlocked */ 525 PAGE_SUCCESS, 526 /* page is clean and locked */ 527 PAGE_CLEAN, 528 } pageout_t; 529 530 /* 531 * pageout is called by shrink_page_list() for each dirty page. 532 * Calls ->writepage(). 533 */ 534 static pageout_t pageout(struct page *page, struct address_space *mapping, 535 struct scan_control *sc) 536 { 537 /* 538 * If the page is dirty, only perform writeback if that write 539 * will be non-blocking. To prevent this allocation from being 540 * stalled by pagecache activity. But note that there may be 541 * stalls if we need to run get_block(). We could test 542 * PagePrivate for that. 543 * 544 * If this process is currently in __generic_file_write_iter() against 545 * this page's queue, we can perform writeback even if that 546 * will block. 547 * 548 * If the page is swapcache, write it back even if that would 549 * block, for some throttling. This happens by accident, because 550 * swap_backing_dev_info is bust: it doesn't reflect the 551 * congestion state of the swapdevs. Easy to fix, if needed. 552 */ 553 if (!is_page_cache_freeable(page)) 554 return PAGE_KEEP; 555 if (!mapping) { 556 /* 557 * Some data journaling orphaned pages can have 558 * page->mapping == NULL while being dirty with clean buffers. 559 */ 560 if (page_has_private(page)) { 561 if (try_to_free_buffers(page)) { 562 ClearPageDirty(page); 563 pr_info("%s: orphaned page\n", __func__); 564 return PAGE_CLEAN; 565 } 566 } 567 return PAGE_KEEP; 568 } 569 if (mapping->a_ops->writepage == NULL) 570 return PAGE_ACTIVATE; 571 if (!may_write_to_inode(mapping->host, sc)) 572 return PAGE_KEEP; 573 574 if (clear_page_dirty_for_io(page)) { 575 int res; 576 struct writeback_control wbc = { 577 .sync_mode = WB_SYNC_NONE, 578 .nr_to_write = SWAP_CLUSTER_MAX, 579 .range_start = 0, 580 .range_end = LLONG_MAX, 581 .for_reclaim = 1, 582 }; 583 584 SetPageReclaim(page); 585 res = mapping->a_ops->writepage(page, &wbc); 586 if (res < 0) 587 handle_write_error(mapping, page, res); 588 if (res == AOP_WRITEPAGE_ACTIVATE) { 589 ClearPageReclaim(page); 590 return PAGE_ACTIVATE; 591 } 592 593 if (!PageWriteback(page)) { 594 /* synchronous write or broken a_ops? */ 595 ClearPageReclaim(page); 596 } 597 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page)); 598 inc_zone_page_state(page, NR_VMSCAN_WRITE); 599 return PAGE_SUCCESS; 600 } 601 602 return PAGE_CLEAN; 603 } 604 605 /* 606 * Same as remove_mapping, but if the page is removed from the mapping, it 607 * gets returned with a refcount of 0. 608 */ 609 static int __remove_mapping(struct address_space *mapping, struct page *page, 610 bool reclaimed) 611 { 612 unsigned long flags; 613 struct mem_cgroup *memcg; 614 615 BUG_ON(!PageLocked(page)); 616 BUG_ON(mapping != page_mapping(page)); 617 618 memcg = mem_cgroup_begin_page_stat(page); 619 spin_lock_irqsave(&mapping->tree_lock, flags); 620 /* 621 * The non racy check for a busy page. 622 * 623 * Must be careful with the order of the tests. When someone has 624 * a ref to the page, it may be possible that they dirty it then 625 * drop the reference. So if PageDirty is tested before page_count 626 * here, then the following race may occur: 627 * 628 * get_user_pages(&page); 629 * [user mapping goes away] 630 * write_to(page); 631 * !PageDirty(page) [good] 632 * SetPageDirty(page); 633 * put_page(page); 634 * !page_count(page) [good, discard it] 635 * 636 * [oops, our write_to data is lost] 637 * 638 * Reversing the order of the tests ensures such a situation cannot 639 * escape unnoticed. The smp_rmb is needed to ensure the page->flags 640 * load is not satisfied before that of page->_count. 641 * 642 * Note that if SetPageDirty is always performed via set_page_dirty, 643 * and thus under tree_lock, then this ordering is not required. 644 */ 645 if (!page_freeze_refs(page, 2)) 646 goto cannot_free; 647 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ 648 if (unlikely(PageDirty(page))) { 649 page_unfreeze_refs(page, 2); 650 goto cannot_free; 651 } 652 653 if (PageSwapCache(page)) { 654 swp_entry_t swap = { .val = page_private(page) }; 655 mem_cgroup_swapout(page, swap); 656 __delete_from_swap_cache(page); 657 spin_unlock_irqrestore(&mapping->tree_lock, flags); 658 mem_cgroup_end_page_stat(memcg); 659 swapcache_free(swap); 660 } else { 661 void (*freepage)(struct page *); 662 void *shadow = NULL; 663 664 freepage = mapping->a_ops->freepage; 665 /* 666 * Remember a shadow entry for reclaimed file cache in 667 * order to detect refaults, thus thrashing, later on. 668 * 669 * But don't store shadows in an address space that is 670 * already exiting. This is not just an optizimation, 671 * inode reclaim needs to empty out the radix tree or 672 * the nodes are lost. Don't plant shadows behind its 673 * back. 674 */ 675 if (reclaimed && page_is_file_cache(page) && 676 !mapping_exiting(mapping)) 677 shadow = workingset_eviction(mapping, page); 678 __delete_from_page_cache(page, shadow, memcg); 679 spin_unlock_irqrestore(&mapping->tree_lock, flags); 680 mem_cgroup_end_page_stat(memcg); 681 682 if (freepage != NULL) 683 freepage(page); 684 } 685 686 return 1; 687 688 cannot_free: 689 spin_unlock_irqrestore(&mapping->tree_lock, flags); 690 mem_cgroup_end_page_stat(memcg); 691 return 0; 692 } 693 694 /* 695 * Attempt to detach a locked page from its ->mapping. If it is dirty or if 696 * someone else has a ref on the page, abort and return 0. If it was 697 * successfully detached, return 1. Assumes the caller has a single ref on 698 * this page. 699 */ 700 int remove_mapping(struct address_space *mapping, struct page *page) 701 { 702 if (__remove_mapping(mapping, page, false)) { 703 /* 704 * Unfreezing the refcount with 1 rather than 2 effectively 705 * drops the pagecache ref for us without requiring another 706 * atomic operation. 707 */ 708 page_unfreeze_refs(page, 1); 709 return 1; 710 } 711 return 0; 712 } 713 714 /** 715 * putback_lru_page - put previously isolated page onto appropriate LRU list 716 * @page: page to be put back to appropriate lru list 717 * 718 * Add previously isolated @page to appropriate LRU list. 719 * Page may still be unevictable for other reasons. 720 * 721 * lru_lock must not be held, interrupts must be enabled. 722 */ 723 void putback_lru_page(struct page *page) 724 { 725 bool is_unevictable; 726 int was_unevictable = PageUnevictable(page); 727 728 VM_BUG_ON_PAGE(PageLRU(page), page); 729 730 redo: 731 ClearPageUnevictable(page); 732 733 if (page_evictable(page)) { 734 /* 735 * For evictable pages, we can use the cache. 736 * In event of a race, worst case is we end up with an 737 * unevictable page on [in]active list. 738 * We know how to handle that. 739 */ 740 is_unevictable = false; 741 lru_cache_add(page); 742 } else { 743 /* 744 * Put unevictable pages directly on zone's unevictable 745 * list. 746 */ 747 is_unevictable = true; 748 add_page_to_unevictable_list(page); 749 /* 750 * When racing with an mlock or AS_UNEVICTABLE clearing 751 * (page is unlocked) make sure that if the other thread 752 * does not observe our setting of PG_lru and fails 753 * isolation/check_move_unevictable_pages, 754 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move 755 * the page back to the evictable list. 756 * 757 * The other side is TestClearPageMlocked() or shmem_lock(). 758 */ 759 smp_mb(); 760 } 761 762 /* 763 * page's status can change while we move it among lru. If an evictable 764 * page is on unevictable list, it never be freed. To avoid that, 765 * check after we added it to the list, again. 766 */ 767 if (is_unevictable && page_evictable(page)) { 768 if (!isolate_lru_page(page)) { 769 put_page(page); 770 goto redo; 771 } 772 /* This means someone else dropped this page from LRU 773 * So, it will be freed or putback to LRU again. There is 774 * nothing to do here. 775 */ 776 } 777 778 if (was_unevictable && !is_unevictable) 779 count_vm_event(UNEVICTABLE_PGRESCUED); 780 else if (!was_unevictable && is_unevictable) 781 count_vm_event(UNEVICTABLE_PGCULLED); 782 783 put_page(page); /* drop ref from isolate */ 784 } 785 786 enum page_references { 787 PAGEREF_RECLAIM, 788 PAGEREF_RECLAIM_CLEAN, 789 PAGEREF_KEEP, 790 PAGEREF_ACTIVATE, 791 }; 792 793 static enum page_references page_check_references(struct page *page, 794 struct scan_control *sc) 795 { 796 int referenced_ptes, referenced_page; 797 unsigned long vm_flags; 798 799 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, 800 &vm_flags); 801 referenced_page = TestClearPageReferenced(page); 802 803 /* 804 * Mlock lost the isolation race with us. Let try_to_unmap() 805 * move the page to the unevictable list. 806 */ 807 if (vm_flags & VM_LOCKED) 808 return PAGEREF_RECLAIM; 809 810 if (referenced_ptes) { 811 if (PageSwapBacked(page)) 812 return PAGEREF_ACTIVATE; 813 /* 814 * All mapped pages start out with page table 815 * references from the instantiating fault, so we need 816 * to look twice if a mapped file page is used more 817 * than once. 818 * 819 * Mark it and spare it for another trip around the 820 * inactive list. Another page table reference will 821 * lead to its activation. 822 * 823 * Note: the mark is set for activated pages as well 824 * so that recently deactivated but used pages are 825 * quickly recovered. 826 */ 827 SetPageReferenced(page); 828 829 if (referenced_page || referenced_ptes > 1) 830 return PAGEREF_ACTIVATE; 831 832 /* 833 * Activate file-backed executable pages after first usage. 834 */ 835 if (vm_flags & VM_EXEC) 836 return PAGEREF_ACTIVATE; 837 838 return PAGEREF_KEEP; 839 } 840 841 /* Reclaim if clean, defer dirty pages to writeback */ 842 if (referenced_page && !PageSwapBacked(page)) 843 return PAGEREF_RECLAIM_CLEAN; 844 845 return PAGEREF_RECLAIM; 846 } 847 848 /* Check if a page is dirty or under writeback */ 849 static void page_check_dirty_writeback(struct page *page, 850 bool *dirty, bool *writeback) 851 { 852 struct address_space *mapping; 853 854 /* 855 * Anonymous pages are not handled by flushers and must be written 856 * from reclaim context. Do not stall reclaim based on them 857 */ 858 if (!page_is_file_cache(page)) { 859 *dirty = false; 860 *writeback = false; 861 return; 862 } 863 864 /* By default assume that the page flags are accurate */ 865 *dirty = PageDirty(page); 866 *writeback = PageWriteback(page); 867 868 /* Verify dirty/writeback state if the filesystem supports it */ 869 if (!page_has_private(page)) 870 return; 871 872 mapping = page_mapping(page); 873 if (mapping && mapping->a_ops->is_dirty_writeback) 874 mapping->a_ops->is_dirty_writeback(page, dirty, writeback); 875 } 876 877 /* 878 * shrink_page_list() returns the number of reclaimed pages 879 */ 880 static unsigned long shrink_page_list(struct list_head *page_list, 881 struct zone *zone, 882 struct scan_control *sc, 883 enum ttu_flags ttu_flags, 884 unsigned long *ret_nr_dirty, 885 unsigned long *ret_nr_unqueued_dirty, 886 unsigned long *ret_nr_congested, 887 unsigned long *ret_nr_writeback, 888 unsigned long *ret_nr_immediate, 889 bool force_reclaim) 890 { 891 LIST_HEAD(ret_pages); 892 LIST_HEAD(free_pages); 893 int pgactivate = 0; 894 unsigned long nr_unqueued_dirty = 0; 895 unsigned long nr_dirty = 0; 896 unsigned long nr_congested = 0; 897 unsigned long nr_reclaimed = 0; 898 unsigned long nr_writeback = 0; 899 unsigned long nr_immediate = 0; 900 901 cond_resched(); 902 903 while (!list_empty(page_list)) { 904 struct address_space *mapping; 905 struct page *page; 906 int may_enter_fs; 907 enum page_references references = PAGEREF_RECLAIM_CLEAN; 908 bool dirty, writeback; 909 910 cond_resched(); 911 912 page = lru_to_page(page_list); 913 list_del(&page->lru); 914 915 if (!trylock_page(page)) 916 goto keep; 917 918 VM_BUG_ON_PAGE(PageActive(page), page); 919 VM_BUG_ON_PAGE(page_zone(page) != zone, page); 920 921 sc->nr_scanned++; 922 923 if (unlikely(!page_evictable(page))) 924 goto cull_mlocked; 925 926 if (!sc->may_unmap && page_mapped(page)) 927 goto keep_locked; 928 929 /* Double the slab pressure for mapped and swapcache pages */ 930 if (page_mapped(page) || PageSwapCache(page)) 931 sc->nr_scanned++; 932 933 may_enter_fs = (sc->gfp_mask & __GFP_FS) || 934 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); 935 936 /* 937 * The number of dirty pages determines if a zone is marked 938 * reclaim_congested which affects wait_iff_congested. kswapd 939 * will stall and start writing pages if the tail of the LRU 940 * is all dirty unqueued pages. 941 */ 942 page_check_dirty_writeback(page, &dirty, &writeback); 943 if (dirty || writeback) 944 nr_dirty++; 945 946 if (dirty && !writeback) 947 nr_unqueued_dirty++; 948 949 /* 950 * Treat this page as congested if the underlying BDI is or if 951 * pages are cycling through the LRU so quickly that the 952 * pages marked for immediate reclaim are making it to the 953 * end of the LRU a second time. 954 */ 955 mapping = page_mapping(page); 956 if (((dirty || writeback) && mapping && 957 inode_write_congested(mapping->host)) || 958 (writeback && PageReclaim(page))) 959 nr_congested++; 960 961 /* 962 * If a page at the tail of the LRU is under writeback, there 963 * are three cases to consider. 964 * 965 * 1) If reclaim is encountering an excessive number of pages 966 * under writeback and this page is both under writeback and 967 * PageReclaim then it indicates that pages are being queued 968 * for IO but are being recycled through the LRU before the 969 * IO can complete. Waiting on the page itself risks an 970 * indefinite stall if it is impossible to writeback the 971 * page due to IO error or disconnected storage so instead 972 * note that the LRU is being scanned too quickly and the 973 * caller can stall after page list has been processed. 974 * 975 * 2) Global or new memcg reclaim encounters a page that is 976 * not marked for immediate reclaim or the caller does not 977 * have __GFP_IO. In this case mark the page for immediate 978 * reclaim and continue scanning. 979 * 980 * __GFP_IO is checked because a loop driver thread might 981 * enter reclaim, and deadlock if it waits on a page for 982 * which it is needed to do the write (loop masks off 983 * __GFP_IO|__GFP_FS for this reason); but more thought 984 * would probably show more reasons. 985 * 986 * Don't require __GFP_FS, since we're not going into the 987 * FS, just waiting on its writeback completion. Worryingly, 988 * ext4 gfs2 and xfs allocate pages with 989 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing 990 * may_enter_fs here is liable to OOM on them. 991 * 992 * 3) Legacy memcg encounters a page that is not already marked 993 * PageReclaim. memcg does not have any dirty pages 994 * throttling so we could easily OOM just because too many 995 * pages are in writeback and there is nothing else to 996 * reclaim. Wait for the writeback to complete. 997 */ 998 if (PageWriteback(page)) { 999 /* Case 1 above */ 1000 if (current_is_kswapd() && 1001 PageReclaim(page) && 1002 test_bit(ZONE_WRITEBACK, &zone->flags)) { 1003 nr_immediate++; 1004 goto keep_locked; 1005 1006 /* Case 2 above */ 1007 } else if (sane_reclaim(sc) || 1008 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) { 1009 /* 1010 * This is slightly racy - end_page_writeback() 1011 * might have just cleared PageReclaim, then 1012 * setting PageReclaim here end up interpreted 1013 * as PageReadahead - but that does not matter 1014 * enough to care. What we do want is for this 1015 * page to have PageReclaim set next time memcg 1016 * reclaim reaches the tests above, so it will 1017 * then wait_on_page_writeback() to avoid OOM; 1018 * and it's also appropriate in global reclaim. 1019 */ 1020 SetPageReclaim(page); 1021 nr_writeback++; 1022 1023 goto keep_locked; 1024 1025 /* Case 3 above */ 1026 } else { 1027 wait_on_page_writeback(page); 1028 } 1029 } 1030 1031 if (!force_reclaim) 1032 references = page_check_references(page, sc); 1033 1034 switch (references) { 1035 case PAGEREF_ACTIVATE: 1036 goto activate_locked; 1037 case PAGEREF_KEEP: 1038 goto keep_locked; 1039 case PAGEREF_RECLAIM: 1040 case PAGEREF_RECLAIM_CLEAN: 1041 ; /* try to reclaim the page below */ 1042 } 1043 1044 /* 1045 * Anonymous process memory has backing store? 1046 * Try to allocate it some swap space here. 1047 */ 1048 if (PageAnon(page) && !PageSwapCache(page)) { 1049 if (!(sc->gfp_mask & __GFP_IO)) 1050 goto keep_locked; 1051 if (!add_to_swap(page, page_list)) 1052 goto activate_locked; 1053 may_enter_fs = 1; 1054 1055 /* Adding to swap updated mapping */ 1056 mapping = page_mapping(page); 1057 } 1058 1059 /* 1060 * The page is mapped into the page tables of one or more 1061 * processes. Try to unmap it here. 1062 */ 1063 if (page_mapped(page) && mapping) { 1064 switch (try_to_unmap(page, ttu_flags)) { 1065 case SWAP_FAIL: 1066 goto activate_locked; 1067 case SWAP_AGAIN: 1068 goto keep_locked; 1069 case SWAP_MLOCK: 1070 goto cull_mlocked; 1071 case SWAP_SUCCESS: 1072 ; /* try to free the page below */ 1073 } 1074 } 1075 1076 if (PageDirty(page)) { 1077 /* 1078 * Only kswapd can writeback filesystem pages to 1079 * avoid risk of stack overflow but only writeback 1080 * if many dirty pages have been encountered. 1081 */ 1082 if (page_is_file_cache(page) && 1083 (!current_is_kswapd() || 1084 !test_bit(ZONE_DIRTY, &zone->flags))) { 1085 /* 1086 * Immediately reclaim when written back. 1087 * Similar in principal to deactivate_page() 1088 * except we already have the page isolated 1089 * and know it's dirty 1090 */ 1091 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); 1092 SetPageReclaim(page); 1093 1094 goto keep_locked; 1095 } 1096 1097 if (references == PAGEREF_RECLAIM_CLEAN) 1098 goto keep_locked; 1099 if (!may_enter_fs) 1100 goto keep_locked; 1101 if (!sc->may_writepage) 1102 goto keep_locked; 1103 1104 /* Page is dirty, try to write it out here */ 1105 switch (pageout(page, mapping, sc)) { 1106 case PAGE_KEEP: 1107 goto keep_locked; 1108 case PAGE_ACTIVATE: 1109 goto activate_locked; 1110 case PAGE_SUCCESS: 1111 if (PageWriteback(page)) 1112 goto keep; 1113 if (PageDirty(page)) 1114 goto keep; 1115 1116 /* 1117 * A synchronous write - probably a ramdisk. Go 1118 * ahead and try to reclaim the page. 1119 */ 1120 if (!trylock_page(page)) 1121 goto keep; 1122 if (PageDirty(page) || PageWriteback(page)) 1123 goto keep_locked; 1124 mapping = page_mapping(page); 1125 case PAGE_CLEAN: 1126 ; /* try to free the page below */ 1127 } 1128 } 1129 1130 /* 1131 * If the page has buffers, try to free the buffer mappings 1132 * associated with this page. If we succeed we try to free 1133 * the page as well. 1134 * 1135 * We do this even if the page is PageDirty(). 1136 * try_to_release_page() does not perform I/O, but it is 1137 * possible for a page to have PageDirty set, but it is actually 1138 * clean (all its buffers are clean). This happens if the 1139 * buffers were written out directly, with submit_bh(). ext3 1140 * will do this, as well as the blockdev mapping. 1141 * try_to_release_page() will discover that cleanness and will 1142 * drop the buffers and mark the page clean - it can be freed. 1143 * 1144 * Rarely, pages can have buffers and no ->mapping. These are 1145 * the pages which were not successfully invalidated in 1146 * truncate_complete_page(). We try to drop those buffers here 1147 * and if that worked, and the page is no longer mapped into 1148 * process address space (page_count == 1) it can be freed. 1149 * Otherwise, leave the page on the LRU so it is swappable. 1150 */ 1151 if (page_has_private(page)) { 1152 if (!try_to_release_page(page, sc->gfp_mask)) 1153 goto activate_locked; 1154 if (!mapping && page_count(page) == 1) { 1155 unlock_page(page); 1156 if (put_page_testzero(page)) 1157 goto free_it; 1158 else { 1159 /* 1160 * rare race with speculative reference. 1161 * the speculative reference will free 1162 * this page shortly, so we may 1163 * increment nr_reclaimed here (and 1164 * leave it off the LRU). 1165 */ 1166 nr_reclaimed++; 1167 continue; 1168 } 1169 } 1170 } 1171 1172 if (!mapping || !__remove_mapping(mapping, page, true)) 1173 goto keep_locked; 1174 1175 /* 1176 * At this point, we have no other references and there is 1177 * no way to pick any more up (removed from LRU, removed 1178 * from pagecache). Can use non-atomic bitops now (and 1179 * we obviously don't have to worry about waking up a process 1180 * waiting on the page lock, because there are no references. 1181 */ 1182 __clear_page_locked(page); 1183 free_it: 1184 nr_reclaimed++; 1185 1186 /* 1187 * Is there need to periodically free_page_list? It would 1188 * appear not as the counts should be low 1189 */ 1190 list_add(&page->lru, &free_pages); 1191 continue; 1192 1193 cull_mlocked: 1194 if (PageSwapCache(page)) 1195 try_to_free_swap(page); 1196 unlock_page(page); 1197 putback_lru_page(page); 1198 continue; 1199 1200 activate_locked: 1201 /* Not a candidate for swapping, so reclaim swap space. */ 1202 if (PageSwapCache(page) && vm_swap_full()) 1203 try_to_free_swap(page); 1204 VM_BUG_ON_PAGE(PageActive(page), page); 1205 SetPageActive(page); 1206 pgactivate++; 1207 keep_locked: 1208 unlock_page(page); 1209 keep: 1210 list_add(&page->lru, &ret_pages); 1211 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); 1212 } 1213 1214 mem_cgroup_uncharge_list(&free_pages); 1215 free_hot_cold_page_list(&free_pages, true); 1216 1217 list_splice(&ret_pages, page_list); 1218 count_vm_events(PGACTIVATE, pgactivate); 1219 1220 *ret_nr_dirty += nr_dirty; 1221 *ret_nr_congested += nr_congested; 1222 *ret_nr_unqueued_dirty += nr_unqueued_dirty; 1223 *ret_nr_writeback += nr_writeback; 1224 *ret_nr_immediate += nr_immediate; 1225 return nr_reclaimed; 1226 } 1227 1228 unsigned long reclaim_clean_pages_from_list(struct zone *zone, 1229 struct list_head *page_list) 1230 { 1231 struct scan_control sc = { 1232 .gfp_mask = GFP_KERNEL, 1233 .priority = DEF_PRIORITY, 1234 .may_unmap = 1, 1235 }; 1236 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5; 1237 struct page *page, *next; 1238 LIST_HEAD(clean_pages); 1239 1240 list_for_each_entry_safe(page, next, page_list, lru) { 1241 if (page_is_file_cache(page) && !PageDirty(page) && 1242 !isolated_balloon_page(page)) { 1243 ClearPageActive(page); 1244 list_move(&page->lru, &clean_pages); 1245 } 1246 } 1247 1248 ret = shrink_page_list(&clean_pages, zone, &sc, 1249 TTU_UNMAP|TTU_IGNORE_ACCESS, 1250 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true); 1251 list_splice(&clean_pages, page_list); 1252 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret); 1253 return ret; 1254 } 1255 1256 /* 1257 * Attempt to remove the specified page from its LRU. Only take this page 1258 * if it is of the appropriate PageActive status. Pages which are being 1259 * freed elsewhere are also ignored. 1260 * 1261 * page: page to consider 1262 * mode: one of the LRU isolation modes defined above 1263 * 1264 * returns 0 on success, -ve errno on failure. 1265 */ 1266 int __isolate_lru_page(struct page *page, isolate_mode_t mode) 1267 { 1268 int ret = -EINVAL; 1269 1270 /* Only take pages on the LRU. */ 1271 if (!PageLRU(page)) 1272 return ret; 1273 1274 /* Compaction should not handle unevictable pages but CMA can do so */ 1275 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) 1276 return ret; 1277 1278 ret = -EBUSY; 1279 1280 /* 1281 * To minimise LRU disruption, the caller can indicate that it only 1282 * wants to isolate pages it will be able to operate on without 1283 * blocking - clean pages for the most part. 1284 * 1285 * ISOLATE_CLEAN means that only clean pages should be isolated. This 1286 * is used by reclaim when it is cannot write to backing storage 1287 * 1288 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages 1289 * that it is possible to migrate without blocking 1290 */ 1291 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { 1292 /* All the caller can do on PageWriteback is block */ 1293 if (PageWriteback(page)) 1294 return ret; 1295 1296 if (PageDirty(page)) { 1297 struct address_space *mapping; 1298 1299 /* ISOLATE_CLEAN means only clean pages */ 1300 if (mode & ISOLATE_CLEAN) 1301 return ret; 1302 1303 /* 1304 * Only pages without mappings or that have a 1305 * ->migratepage callback are possible to migrate 1306 * without blocking 1307 */ 1308 mapping = page_mapping(page); 1309 if (mapping && !mapping->a_ops->migratepage) 1310 return ret; 1311 } 1312 } 1313 1314 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) 1315 return ret; 1316 1317 if (likely(get_page_unless_zero(page))) { 1318 /* 1319 * Be careful not to clear PageLRU until after we're 1320 * sure the page is not being freed elsewhere -- the 1321 * page release code relies on it. 1322 */ 1323 ClearPageLRU(page); 1324 ret = 0; 1325 } 1326 1327 return ret; 1328 } 1329 1330 /* 1331 * zone->lru_lock is heavily contended. Some of the functions that 1332 * shrink the lists perform better by taking out a batch of pages 1333 * and working on them outside the LRU lock. 1334 * 1335 * For pagecache intensive workloads, this function is the hottest 1336 * spot in the kernel (apart from copy_*_user functions). 1337 * 1338 * Appropriate locks must be held before calling this function. 1339 * 1340 * @nr_to_scan: The number of pages to look through on the list. 1341 * @lruvec: The LRU vector to pull pages from. 1342 * @dst: The temp list to put pages on to. 1343 * @nr_scanned: The number of pages that were scanned. 1344 * @sc: The scan_control struct for this reclaim session 1345 * @mode: One of the LRU isolation modes 1346 * @lru: LRU list id for isolating 1347 * 1348 * returns how many pages were moved onto *@dst. 1349 */ 1350 static unsigned long isolate_lru_pages(unsigned long nr_to_scan, 1351 struct lruvec *lruvec, struct list_head *dst, 1352 unsigned long *nr_scanned, struct scan_control *sc, 1353 isolate_mode_t mode, enum lru_list lru) 1354 { 1355 struct list_head *src = &lruvec->lists[lru]; 1356 unsigned long nr_taken = 0; 1357 unsigned long scan; 1358 1359 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { 1360 struct page *page; 1361 int nr_pages; 1362 1363 page = lru_to_page(src); 1364 prefetchw_prev_lru_page(page, src, flags); 1365 1366 VM_BUG_ON_PAGE(!PageLRU(page), page); 1367 1368 switch (__isolate_lru_page(page, mode)) { 1369 case 0: 1370 nr_pages = hpage_nr_pages(page); 1371 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages); 1372 list_move(&page->lru, dst); 1373 nr_taken += nr_pages; 1374 break; 1375 1376 case -EBUSY: 1377 /* else it is being freed elsewhere */ 1378 list_move(&page->lru, src); 1379 continue; 1380 1381 default: 1382 BUG(); 1383 } 1384 } 1385 1386 *nr_scanned = scan; 1387 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan, 1388 nr_taken, mode, is_file_lru(lru)); 1389 return nr_taken; 1390 } 1391 1392 /** 1393 * isolate_lru_page - tries to isolate a page from its LRU list 1394 * @page: page to isolate from its LRU list 1395 * 1396 * Isolates a @page from an LRU list, clears PageLRU and adjusts the 1397 * vmstat statistic corresponding to whatever LRU list the page was on. 1398 * 1399 * Returns 0 if the page was removed from an LRU list. 1400 * Returns -EBUSY if the page was not on an LRU list. 1401 * 1402 * The returned page will have PageLRU() cleared. If it was found on 1403 * the active list, it will have PageActive set. If it was found on 1404 * the unevictable list, it will have the PageUnevictable bit set. That flag 1405 * may need to be cleared by the caller before letting the page go. 1406 * 1407 * The vmstat statistic corresponding to the list on which the page was 1408 * found will be decremented. 1409 * 1410 * Restrictions: 1411 * (1) Must be called with an elevated refcount on the page. This is a 1412 * fundamentnal difference from isolate_lru_pages (which is called 1413 * without a stable reference). 1414 * (2) the lru_lock must not be held. 1415 * (3) interrupts must be enabled. 1416 */ 1417 int isolate_lru_page(struct page *page) 1418 { 1419 int ret = -EBUSY; 1420 1421 VM_BUG_ON_PAGE(!page_count(page), page); 1422 1423 if (PageLRU(page)) { 1424 struct zone *zone = page_zone(page); 1425 struct lruvec *lruvec; 1426 1427 spin_lock_irq(&zone->lru_lock); 1428 lruvec = mem_cgroup_page_lruvec(page, zone); 1429 if (PageLRU(page)) { 1430 int lru = page_lru(page); 1431 get_page(page); 1432 ClearPageLRU(page); 1433 del_page_from_lru_list(page, lruvec, lru); 1434 ret = 0; 1435 } 1436 spin_unlock_irq(&zone->lru_lock); 1437 } 1438 return ret; 1439 } 1440 1441 /* 1442 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and 1443 * then get resheduled. When there are massive number of tasks doing page 1444 * allocation, such sleeping direct reclaimers may keep piling up on each CPU, 1445 * the LRU list will go small and be scanned faster than necessary, leading to 1446 * unnecessary swapping, thrashing and OOM. 1447 */ 1448 static int too_many_isolated(struct zone *zone, int file, 1449 struct scan_control *sc) 1450 { 1451 unsigned long inactive, isolated; 1452 1453 if (current_is_kswapd()) 1454 return 0; 1455 1456 if (!sane_reclaim(sc)) 1457 return 0; 1458 1459 if (file) { 1460 inactive = zone_page_state(zone, NR_INACTIVE_FILE); 1461 isolated = zone_page_state(zone, NR_ISOLATED_FILE); 1462 } else { 1463 inactive = zone_page_state(zone, NR_INACTIVE_ANON); 1464 isolated = zone_page_state(zone, NR_ISOLATED_ANON); 1465 } 1466 1467 /* 1468 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they 1469 * won't get blocked by normal direct-reclaimers, forming a circular 1470 * deadlock. 1471 */ 1472 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS) 1473 inactive >>= 3; 1474 1475 return isolated > inactive; 1476 } 1477 1478 static noinline_for_stack void 1479 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) 1480 { 1481 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1482 struct zone *zone = lruvec_zone(lruvec); 1483 LIST_HEAD(pages_to_free); 1484 1485 /* 1486 * Put back any unfreeable pages. 1487 */ 1488 while (!list_empty(page_list)) { 1489 struct page *page = lru_to_page(page_list); 1490 int lru; 1491 1492 VM_BUG_ON_PAGE(PageLRU(page), page); 1493 list_del(&page->lru); 1494 if (unlikely(!page_evictable(page))) { 1495 spin_unlock_irq(&zone->lru_lock); 1496 putback_lru_page(page); 1497 spin_lock_irq(&zone->lru_lock); 1498 continue; 1499 } 1500 1501 lruvec = mem_cgroup_page_lruvec(page, zone); 1502 1503 SetPageLRU(page); 1504 lru = page_lru(page); 1505 add_page_to_lru_list(page, lruvec, lru); 1506 1507 if (is_active_lru(lru)) { 1508 int file = is_file_lru(lru); 1509 int numpages = hpage_nr_pages(page); 1510 reclaim_stat->recent_rotated[file] += numpages; 1511 } 1512 if (put_page_testzero(page)) { 1513 __ClearPageLRU(page); 1514 __ClearPageActive(page); 1515 del_page_from_lru_list(page, lruvec, lru); 1516 1517 if (unlikely(PageCompound(page))) { 1518 spin_unlock_irq(&zone->lru_lock); 1519 mem_cgroup_uncharge(page); 1520 (*get_compound_page_dtor(page))(page); 1521 spin_lock_irq(&zone->lru_lock); 1522 } else 1523 list_add(&page->lru, &pages_to_free); 1524 } 1525 } 1526 1527 /* 1528 * To save our caller's stack, now use input list for pages to free. 1529 */ 1530 list_splice(&pages_to_free, page_list); 1531 } 1532 1533 /* 1534 * If a kernel thread (such as nfsd for loop-back mounts) services 1535 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. 1536 * In that case we should only throttle if the backing device it is 1537 * writing to is congested. In other cases it is safe to throttle. 1538 */ 1539 static int current_may_throttle(void) 1540 { 1541 return !(current->flags & PF_LESS_THROTTLE) || 1542 current->backing_dev_info == NULL || 1543 bdi_write_congested(current->backing_dev_info); 1544 } 1545 1546 /* 1547 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number 1548 * of reclaimed pages 1549 */ 1550 static noinline_for_stack unsigned long 1551 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, 1552 struct scan_control *sc, enum lru_list lru) 1553 { 1554 LIST_HEAD(page_list); 1555 unsigned long nr_scanned; 1556 unsigned long nr_reclaimed = 0; 1557 unsigned long nr_taken; 1558 unsigned long nr_dirty = 0; 1559 unsigned long nr_congested = 0; 1560 unsigned long nr_unqueued_dirty = 0; 1561 unsigned long nr_writeback = 0; 1562 unsigned long nr_immediate = 0; 1563 isolate_mode_t isolate_mode = 0; 1564 int file = is_file_lru(lru); 1565 struct zone *zone = lruvec_zone(lruvec); 1566 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1567 1568 while (unlikely(too_many_isolated(zone, file, sc))) { 1569 congestion_wait(BLK_RW_ASYNC, HZ/10); 1570 1571 /* We are about to die and free our memory. Return now. */ 1572 if (fatal_signal_pending(current)) 1573 return SWAP_CLUSTER_MAX; 1574 } 1575 1576 lru_add_drain(); 1577 1578 if (!sc->may_unmap) 1579 isolate_mode |= ISOLATE_UNMAPPED; 1580 if (!sc->may_writepage) 1581 isolate_mode |= ISOLATE_CLEAN; 1582 1583 spin_lock_irq(&zone->lru_lock); 1584 1585 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, 1586 &nr_scanned, sc, isolate_mode, lru); 1587 1588 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); 1589 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); 1590 1591 if (global_reclaim(sc)) { 1592 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned); 1593 if (current_is_kswapd()) 1594 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned); 1595 else 1596 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned); 1597 } 1598 spin_unlock_irq(&zone->lru_lock); 1599 1600 if (nr_taken == 0) 1601 return 0; 1602 1603 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP, 1604 &nr_dirty, &nr_unqueued_dirty, &nr_congested, 1605 &nr_writeback, &nr_immediate, 1606 false); 1607 1608 spin_lock_irq(&zone->lru_lock); 1609 1610 reclaim_stat->recent_scanned[file] += nr_taken; 1611 1612 if (global_reclaim(sc)) { 1613 if (current_is_kswapd()) 1614 __count_zone_vm_events(PGSTEAL_KSWAPD, zone, 1615 nr_reclaimed); 1616 else 1617 __count_zone_vm_events(PGSTEAL_DIRECT, zone, 1618 nr_reclaimed); 1619 } 1620 1621 putback_inactive_pages(lruvec, &page_list); 1622 1623 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); 1624 1625 spin_unlock_irq(&zone->lru_lock); 1626 1627 mem_cgroup_uncharge_list(&page_list); 1628 free_hot_cold_page_list(&page_list, true); 1629 1630 /* 1631 * If reclaim is isolating dirty pages under writeback, it implies 1632 * that the long-lived page allocation rate is exceeding the page 1633 * laundering rate. Either the global limits are not being effective 1634 * at throttling processes due to the page distribution throughout 1635 * zones or there is heavy usage of a slow backing device. The 1636 * only option is to throttle from reclaim context which is not ideal 1637 * as there is no guarantee the dirtying process is throttled in the 1638 * same way balance_dirty_pages() manages. 1639 * 1640 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number 1641 * of pages under pages flagged for immediate reclaim and stall if any 1642 * are encountered in the nr_immediate check below. 1643 */ 1644 if (nr_writeback && nr_writeback == nr_taken) 1645 set_bit(ZONE_WRITEBACK, &zone->flags); 1646 1647 /* 1648 * Legacy memcg will stall in page writeback so avoid forcibly 1649 * stalling here. 1650 */ 1651 if (sane_reclaim(sc)) { 1652 /* 1653 * Tag a zone as congested if all the dirty pages scanned were 1654 * backed by a congested BDI and wait_iff_congested will stall. 1655 */ 1656 if (nr_dirty && nr_dirty == nr_congested) 1657 set_bit(ZONE_CONGESTED, &zone->flags); 1658 1659 /* 1660 * If dirty pages are scanned that are not queued for IO, it 1661 * implies that flushers are not keeping up. In this case, flag 1662 * the zone ZONE_DIRTY and kswapd will start writing pages from 1663 * reclaim context. 1664 */ 1665 if (nr_unqueued_dirty == nr_taken) 1666 set_bit(ZONE_DIRTY, &zone->flags); 1667 1668 /* 1669 * If kswapd scans pages marked marked for immediate 1670 * reclaim and under writeback (nr_immediate), it implies 1671 * that pages are cycling through the LRU faster than 1672 * they are written so also forcibly stall. 1673 */ 1674 if (nr_immediate && current_may_throttle()) 1675 congestion_wait(BLK_RW_ASYNC, HZ/10); 1676 } 1677 1678 /* 1679 * Stall direct reclaim for IO completions if underlying BDIs or zone 1680 * is congested. Allow kswapd to continue until it starts encountering 1681 * unqueued dirty pages or cycling through the LRU too quickly. 1682 */ 1683 if (!sc->hibernation_mode && !current_is_kswapd() && 1684 current_may_throttle()) 1685 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); 1686 1687 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, 1688 zone_idx(zone), 1689 nr_scanned, nr_reclaimed, 1690 sc->priority, 1691 trace_shrink_flags(file)); 1692 return nr_reclaimed; 1693 } 1694 1695 /* 1696 * This moves pages from the active list to the inactive list. 1697 * 1698 * We move them the other way if the page is referenced by one or more 1699 * processes, from rmap. 1700 * 1701 * If the pages are mostly unmapped, the processing is fast and it is 1702 * appropriate to hold zone->lru_lock across the whole operation. But if 1703 * the pages are mapped, the processing is slow (page_referenced()) so we 1704 * should drop zone->lru_lock around each page. It's impossible to balance 1705 * this, so instead we remove the pages from the LRU while processing them. 1706 * It is safe to rely on PG_active against the non-LRU pages in here because 1707 * nobody will play with that bit on a non-LRU page. 1708 * 1709 * The downside is that we have to touch page->_count against each page. 1710 * But we had to alter page->flags anyway. 1711 */ 1712 1713 static void move_active_pages_to_lru(struct lruvec *lruvec, 1714 struct list_head *list, 1715 struct list_head *pages_to_free, 1716 enum lru_list lru) 1717 { 1718 struct zone *zone = lruvec_zone(lruvec); 1719 unsigned long pgmoved = 0; 1720 struct page *page; 1721 int nr_pages; 1722 1723 while (!list_empty(list)) { 1724 page = lru_to_page(list); 1725 lruvec = mem_cgroup_page_lruvec(page, zone); 1726 1727 VM_BUG_ON_PAGE(PageLRU(page), page); 1728 SetPageLRU(page); 1729 1730 nr_pages = hpage_nr_pages(page); 1731 mem_cgroup_update_lru_size(lruvec, lru, nr_pages); 1732 list_move(&page->lru, &lruvec->lists[lru]); 1733 pgmoved += nr_pages; 1734 1735 if (put_page_testzero(page)) { 1736 __ClearPageLRU(page); 1737 __ClearPageActive(page); 1738 del_page_from_lru_list(page, lruvec, lru); 1739 1740 if (unlikely(PageCompound(page))) { 1741 spin_unlock_irq(&zone->lru_lock); 1742 mem_cgroup_uncharge(page); 1743 (*get_compound_page_dtor(page))(page); 1744 spin_lock_irq(&zone->lru_lock); 1745 } else 1746 list_add(&page->lru, pages_to_free); 1747 } 1748 } 1749 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); 1750 if (!is_active_lru(lru)) 1751 __count_vm_events(PGDEACTIVATE, pgmoved); 1752 } 1753 1754 static void shrink_active_list(unsigned long nr_to_scan, 1755 struct lruvec *lruvec, 1756 struct scan_control *sc, 1757 enum lru_list lru) 1758 { 1759 unsigned long nr_taken; 1760 unsigned long nr_scanned; 1761 unsigned long vm_flags; 1762 LIST_HEAD(l_hold); /* The pages which were snipped off */ 1763 LIST_HEAD(l_active); 1764 LIST_HEAD(l_inactive); 1765 struct page *page; 1766 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1767 unsigned long nr_rotated = 0; 1768 isolate_mode_t isolate_mode = 0; 1769 int file = is_file_lru(lru); 1770 struct zone *zone = lruvec_zone(lruvec); 1771 1772 lru_add_drain(); 1773 1774 if (!sc->may_unmap) 1775 isolate_mode |= ISOLATE_UNMAPPED; 1776 if (!sc->may_writepage) 1777 isolate_mode |= ISOLATE_CLEAN; 1778 1779 spin_lock_irq(&zone->lru_lock); 1780 1781 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, 1782 &nr_scanned, sc, isolate_mode, lru); 1783 if (global_reclaim(sc)) 1784 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned); 1785 1786 reclaim_stat->recent_scanned[file] += nr_taken; 1787 1788 __count_zone_vm_events(PGREFILL, zone, nr_scanned); 1789 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); 1790 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); 1791 spin_unlock_irq(&zone->lru_lock); 1792 1793 while (!list_empty(&l_hold)) { 1794 cond_resched(); 1795 page = lru_to_page(&l_hold); 1796 list_del(&page->lru); 1797 1798 if (unlikely(!page_evictable(page))) { 1799 putback_lru_page(page); 1800 continue; 1801 } 1802 1803 if (unlikely(buffer_heads_over_limit)) { 1804 if (page_has_private(page) && trylock_page(page)) { 1805 if (page_has_private(page)) 1806 try_to_release_page(page, 0); 1807 unlock_page(page); 1808 } 1809 } 1810 1811 if (page_referenced(page, 0, sc->target_mem_cgroup, 1812 &vm_flags)) { 1813 nr_rotated += hpage_nr_pages(page); 1814 /* 1815 * Identify referenced, file-backed active pages and 1816 * give them one more trip around the active list. So 1817 * that executable code get better chances to stay in 1818 * memory under moderate memory pressure. Anon pages 1819 * are not likely to be evicted by use-once streaming 1820 * IO, plus JVM can create lots of anon VM_EXEC pages, 1821 * so we ignore them here. 1822 */ 1823 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { 1824 list_add(&page->lru, &l_active); 1825 continue; 1826 } 1827 } 1828 1829 ClearPageActive(page); /* we are de-activating */ 1830 list_add(&page->lru, &l_inactive); 1831 } 1832 1833 /* 1834 * Move pages back to the lru list. 1835 */ 1836 spin_lock_irq(&zone->lru_lock); 1837 /* 1838 * Count referenced pages from currently used mappings as rotated, 1839 * even though only some of them are actually re-activated. This 1840 * helps balance scan pressure between file and anonymous pages in 1841 * get_scan_count. 1842 */ 1843 reclaim_stat->recent_rotated[file] += nr_rotated; 1844 1845 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); 1846 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); 1847 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); 1848 spin_unlock_irq(&zone->lru_lock); 1849 1850 mem_cgroup_uncharge_list(&l_hold); 1851 free_hot_cold_page_list(&l_hold, true); 1852 } 1853 1854 #ifdef CONFIG_SWAP 1855 static int inactive_anon_is_low_global(struct zone *zone) 1856 { 1857 unsigned long active, inactive; 1858 1859 active = zone_page_state(zone, NR_ACTIVE_ANON); 1860 inactive = zone_page_state(zone, NR_INACTIVE_ANON); 1861 1862 if (inactive * zone->inactive_ratio < active) 1863 return 1; 1864 1865 return 0; 1866 } 1867 1868 /** 1869 * inactive_anon_is_low - check if anonymous pages need to be deactivated 1870 * @lruvec: LRU vector to check 1871 * 1872 * Returns true if the zone does not have enough inactive anon pages, 1873 * meaning some active anon pages need to be deactivated. 1874 */ 1875 static int inactive_anon_is_low(struct lruvec *lruvec) 1876 { 1877 /* 1878 * If we don't have swap space, anonymous page deactivation 1879 * is pointless. 1880 */ 1881 if (!total_swap_pages) 1882 return 0; 1883 1884 if (!mem_cgroup_disabled()) 1885 return mem_cgroup_inactive_anon_is_low(lruvec); 1886 1887 return inactive_anon_is_low_global(lruvec_zone(lruvec)); 1888 } 1889 #else 1890 static inline int inactive_anon_is_low(struct lruvec *lruvec) 1891 { 1892 return 0; 1893 } 1894 #endif 1895 1896 /** 1897 * inactive_file_is_low - check if file pages need to be deactivated 1898 * @lruvec: LRU vector to check 1899 * 1900 * When the system is doing streaming IO, memory pressure here 1901 * ensures that active file pages get deactivated, until more 1902 * than half of the file pages are on the inactive list. 1903 * 1904 * Once we get to that situation, protect the system's working 1905 * set from being evicted by disabling active file page aging. 1906 * 1907 * This uses a different ratio than the anonymous pages, because 1908 * the page cache uses a use-once replacement algorithm. 1909 */ 1910 static int inactive_file_is_low(struct lruvec *lruvec) 1911 { 1912 unsigned long inactive; 1913 unsigned long active; 1914 1915 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE); 1916 active = get_lru_size(lruvec, LRU_ACTIVE_FILE); 1917 1918 return active > inactive; 1919 } 1920 1921 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru) 1922 { 1923 if (is_file_lru(lru)) 1924 return inactive_file_is_low(lruvec); 1925 else 1926 return inactive_anon_is_low(lruvec); 1927 } 1928 1929 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, 1930 struct lruvec *lruvec, struct scan_control *sc) 1931 { 1932 if (is_active_lru(lru)) { 1933 if (inactive_list_is_low(lruvec, lru)) 1934 shrink_active_list(nr_to_scan, lruvec, sc, lru); 1935 return 0; 1936 } 1937 1938 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); 1939 } 1940 1941 enum scan_balance { 1942 SCAN_EQUAL, 1943 SCAN_FRACT, 1944 SCAN_ANON, 1945 SCAN_FILE, 1946 }; 1947 1948 /* 1949 * Determine how aggressively the anon and file LRU lists should be 1950 * scanned. The relative value of each set of LRU lists is determined 1951 * by looking at the fraction of the pages scanned we did rotate back 1952 * onto the active list instead of evict. 1953 * 1954 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan 1955 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan 1956 */ 1957 static void get_scan_count(struct lruvec *lruvec, int swappiness, 1958 struct scan_control *sc, unsigned long *nr, 1959 unsigned long *lru_pages) 1960 { 1961 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1962 u64 fraction[2]; 1963 u64 denominator = 0; /* gcc */ 1964 struct zone *zone = lruvec_zone(lruvec); 1965 unsigned long anon_prio, file_prio; 1966 enum scan_balance scan_balance; 1967 unsigned long anon, file; 1968 bool force_scan = false; 1969 unsigned long ap, fp; 1970 enum lru_list lru; 1971 bool some_scanned; 1972 int pass; 1973 1974 /* 1975 * If the zone or memcg is small, nr[l] can be 0. This 1976 * results in no scanning on this priority and a potential 1977 * priority drop. Global direct reclaim can go to the next 1978 * zone and tends to have no problems. Global kswapd is for 1979 * zone balancing and it needs to scan a minimum amount. When 1980 * reclaiming for a memcg, a priority drop can cause high 1981 * latencies, so it's better to scan a minimum amount there as 1982 * well. 1983 */ 1984 if (current_is_kswapd()) { 1985 if (!zone_reclaimable(zone)) 1986 force_scan = true; 1987 if (!mem_cgroup_lruvec_online(lruvec)) 1988 force_scan = true; 1989 } 1990 if (!global_reclaim(sc)) 1991 force_scan = true; 1992 1993 /* If we have no swap space, do not bother scanning anon pages. */ 1994 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) { 1995 scan_balance = SCAN_FILE; 1996 goto out; 1997 } 1998 1999 /* 2000 * Global reclaim will swap to prevent OOM even with no 2001 * swappiness, but memcg users want to use this knob to 2002 * disable swapping for individual groups completely when 2003 * using the memory controller's swap limit feature would be 2004 * too expensive. 2005 */ 2006 if (!global_reclaim(sc) && !swappiness) { 2007 scan_balance = SCAN_FILE; 2008 goto out; 2009 } 2010 2011 /* 2012 * Do not apply any pressure balancing cleverness when the 2013 * system is close to OOM, scan both anon and file equally 2014 * (unless the swappiness setting disagrees with swapping). 2015 */ 2016 if (!sc->priority && swappiness) { 2017 scan_balance = SCAN_EQUAL; 2018 goto out; 2019 } 2020 2021 /* 2022 * Prevent the reclaimer from falling into the cache trap: as 2023 * cache pages start out inactive, every cache fault will tip 2024 * the scan balance towards the file LRU. And as the file LRU 2025 * shrinks, so does the window for rotation from references. 2026 * This means we have a runaway feedback loop where a tiny 2027 * thrashing file LRU becomes infinitely more attractive than 2028 * anon pages. Try to detect this based on file LRU size. 2029 */ 2030 if (global_reclaim(sc)) { 2031 unsigned long zonefile; 2032 unsigned long zonefree; 2033 2034 zonefree = zone_page_state(zone, NR_FREE_PAGES); 2035 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) + 2036 zone_page_state(zone, NR_INACTIVE_FILE); 2037 2038 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) { 2039 scan_balance = SCAN_ANON; 2040 goto out; 2041 } 2042 } 2043 2044 /* 2045 * There is enough inactive page cache, do not reclaim 2046 * anything from the anonymous working set right now. 2047 */ 2048 if (!inactive_file_is_low(lruvec)) { 2049 scan_balance = SCAN_FILE; 2050 goto out; 2051 } 2052 2053 scan_balance = SCAN_FRACT; 2054 2055 /* 2056 * With swappiness at 100, anonymous and file have the same priority. 2057 * This scanning priority is essentially the inverse of IO cost. 2058 */ 2059 anon_prio = swappiness; 2060 file_prio = 200 - anon_prio; 2061 2062 /* 2063 * OK, so we have swap space and a fair amount of page cache 2064 * pages. We use the recently rotated / recently scanned 2065 * ratios to determine how valuable each cache is. 2066 * 2067 * Because workloads change over time (and to avoid overflow) 2068 * we keep these statistics as a floating average, which ends 2069 * up weighing recent references more than old ones. 2070 * 2071 * anon in [0], file in [1] 2072 */ 2073 2074 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) + 2075 get_lru_size(lruvec, LRU_INACTIVE_ANON); 2076 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) + 2077 get_lru_size(lruvec, LRU_INACTIVE_FILE); 2078 2079 spin_lock_irq(&zone->lru_lock); 2080 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { 2081 reclaim_stat->recent_scanned[0] /= 2; 2082 reclaim_stat->recent_rotated[0] /= 2; 2083 } 2084 2085 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { 2086 reclaim_stat->recent_scanned[1] /= 2; 2087 reclaim_stat->recent_rotated[1] /= 2; 2088 } 2089 2090 /* 2091 * The amount of pressure on anon vs file pages is inversely 2092 * proportional to the fraction of recently scanned pages on 2093 * each list that were recently referenced and in active use. 2094 */ 2095 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); 2096 ap /= reclaim_stat->recent_rotated[0] + 1; 2097 2098 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); 2099 fp /= reclaim_stat->recent_rotated[1] + 1; 2100 spin_unlock_irq(&zone->lru_lock); 2101 2102 fraction[0] = ap; 2103 fraction[1] = fp; 2104 denominator = ap + fp + 1; 2105 out: 2106 some_scanned = false; 2107 /* Only use force_scan on second pass. */ 2108 for (pass = 0; !some_scanned && pass < 2; pass++) { 2109 *lru_pages = 0; 2110 for_each_evictable_lru(lru) { 2111 int file = is_file_lru(lru); 2112 unsigned long size; 2113 unsigned long scan; 2114 2115 size = get_lru_size(lruvec, lru); 2116 scan = size >> sc->priority; 2117 2118 if (!scan && pass && force_scan) 2119 scan = min(size, SWAP_CLUSTER_MAX); 2120 2121 switch (scan_balance) { 2122 case SCAN_EQUAL: 2123 /* Scan lists relative to size */ 2124 break; 2125 case SCAN_FRACT: 2126 /* 2127 * Scan types proportional to swappiness and 2128 * their relative recent reclaim efficiency. 2129 */ 2130 scan = div64_u64(scan * fraction[file], 2131 denominator); 2132 break; 2133 case SCAN_FILE: 2134 case SCAN_ANON: 2135 /* Scan one type exclusively */ 2136 if ((scan_balance == SCAN_FILE) != file) { 2137 size = 0; 2138 scan = 0; 2139 } 2140 break; 2141 default: 2142 /* Look ma, no brain */ 2143 BUG(); 2144 } 2145 2146 *lru_pages += size; 2147 nr[lru] = scan; 2148 2149 /* 2150 * Skip the second pass and don't force_scan, 2151 * if we found something to scan. 2152 */ 2153 some_scanned |= !!scan; 2154 } 2155 } 2156 } 2157 2158 /* 2159 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. 2160 */ 2161 static void shrink_lruvec(struct lruvec *lruvec, int swappiness, 2162 struct scan_control *sc, unsigned long *lru_pages) 2163 { 2164 unsigned long nr[NR_LRU_LISTS]; 2165 unsigned long targets[NR_LRU_LISTS]; 2166 unsigned long nr_to_scan; 2167 enum lru_list lru; 2168 unsigned long nr_reclaimed = 0; 2169 unsigned long nr_to_reclaim = sc->nr_to_reclaim; 2170 struct blk_plug plug; 2171 bool scan_adjusted; 2172 2173 get_scan_count(lruvec, swappiness, sc, nr, lru_pages); 2174 2175 /* Record the original scan target for proportional adjustments later */ 2176 memcpy(targets, nr, sizeof(nr)); 2177 2178 /* 2179 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal 2180 * event that can occur when there is little memory pressure e.g. 2181 * multiple streaming readers/writers. Hence, we do not abort scanning 2182 * when the requested number of pages are reclaimed when scanning at 2183 * DEF_PRIORITY on the assumption that the fact we are direct 2184 * reclaiming implies that kswapd is not keeping up and it is best to 2185 * do a batch of work at once. For memcg reclaim one check is made to 2186 * abort proportional reclaim if either the file or anon lru has already 2187 * dropped to zero at the first pass. 2188 */ 2189 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && 2190 sc->priority == DEF_PRIORITY); 2191 2192 blk_start_plug(&plug); 2193 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || 2194 nr[LRU_INACTIVE_FILE]) { 2195 unsigned long nr_anon, nr_file, percentage; 2196 unsigned long nr_scanned; 2197 2198 for_each_evictable_lru(lru) { 2199 if (nr[lru]) { 2200 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); 2201 nr[lru] -= nr_to_scan; 2202 2203 nr_reclaimed += shrink_list(lru, nr_to_scan, 2204 lruvec, sc); 2205 } 2206 } 2207 2208 if (nr_reclaimed < nr_to_reclaim || scan_adjusted) 2209 continue; 2210 2211 /* 2212 * For kswapd and memcg, reclaim at least the number of pages 2213 * requested. Ensure that the anon and file LRUs are scanned 2214 * proportionally what was requested by get_scan_count(). We 2215 * stop reclaiming one LRU and reduce the amount scanning 2216 * proportional to the original scan target. 2217 */ 2218 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; 2219 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; 2220 2221 /* 2222 * It's just vindictive to attack the larger once the smaller 2223 * has gone to zero. And given the way we stop scanning the 2224 * smaller below, this makes sure that we only make one nudge 2225 * towards proportionality once we've got nr_to_reclaim. 2226 */ 2227 if (!nr_file || !nr_anon) 2228 break; 2229 2230 if (nr_file > nr_anon) { 2231 unsigned long scan_target = targets[LRU_INACTIVE_ANON] + 2232 targets[LRU_ACTIVE_ANON] + 1; 2233 lru = LRU_BASE; 2234 percentage = nr_anon * 100 / scan_target; 2235 } else { 2236 unsigned long scan_target = targets[LRU_INACTIVE_FILE] + 2237 targets[LRU_ACTIVE_FILE] + 1; 2238 lru = LRU_FILE; 2239 percentage = nr_file * 100 / scan_target; 2240 } 2241 2242 /* Stop scanning the smaller of the LRU */ 2243 nr[lru] = 0; 2244 nr[lru + LRU_ACTIVE] = 0; 2245 2246 /* 2247 * Recalculate the other LRU scan count based on its original 2248 * scan target and the percentage scanning already complete 2249 */ 2250 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; 2251 nr_scanned = targets[lru] - nr[lru]; 2252 nr[lru] = targets[lru] * (100 - percentage) / 100; 2253 nr[lru] -= min(nr[lru], nr_scanned); 2254 2255 lru += LRU_ACTIVE; 2256 nr_scanned = targets[lru] - nr[lru]; 2257 nr[lru] = targets[lru] * (100 - percentage) / 100; 2258 nr[lru] -= min(nr[lru], nr_scanned); 2259 2260 scan_adjusted = true; 2261 } 2262 blk_finish_plug(&plug); 2263 sc->nr_reclaimed += nr_reclaimed; 2264 2265 /* 2266 * Even if we did not try to evict anon pages at all, we want to 2267 * rebalance the anon lru active/inactive ratio. 2268 */ 2269 if (inactive_anon_is_low(lruvec)) 2270 shrink_active_list(SWAP_CLUSTER_MAX, lruvec, 2271 sc, LRU_ACTIVE_ANON); 2272 2273 throttle_vm_writeout(sc->gfp_mask); 2274 } 2275 2276 /* Use reclaim/compaction for costly allocs or under memory pressure */ 2277 static bool in_reclaim_compaction(struct scan_control *sc) 2278 { 2279 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && 2280 (sc->order > PAGE_ALLOC_COSTLY_ORDER || 2281 sc->priority < DEF_PRIORITY - 2)) 2282 return true; 2283 2284 return false; 2285 } 2286 2287 /* 2288 * Reclaim/compaction is used for high-order allocation requests. It reclaims 2289 * order-0 pages before compacting the zone. should_continue_reclaim() returns 2290 * true if more pages should be reclaimed such that when the page allocator 2291 * calls try_to_compact_zone() that it will have enough free pages to succeed. 2292 * It will give up earlier than that if there is difficulty reclaiming pages. 2293 */ 2294 static inline bool should_continue_reclaim(struct zone *zone, 2295 unsigned long nr_reclaimed, 2296 unsigned long nr_scanned, 2297 struct scan_control *sc) 2298 { 2299 unsigned long pages_for_compaction; 2300 unsigned long inactive_lru_pages; 2301 2302 /* If not in reclaim/compaction mode, stop */ 2303 if (!in_reclaim_compaction(sc)) 2304 return false; 2305 2306 /* Consider stopping depending on scan and reclaim activity */ 2307 if (sc->gfp_mask & __GFP_REPEAT) { 2308 /* 2309 * For __GFP_REPEAT allocations, stop reclaiming if the 2310 * full LRU list has been scanned and we are still failing 2311 * to reclaim pages. This full LRU scan is potentially 2312 * expensive but a __GFP_REPEAT caller really wants to succeed 2313 */ 2314 if (!nr_reclaimed && !nr_scanned) 2315 return false; 2316 } else { 2317 /* 2318 * For non-__GFP_REPEAT allocations which can presumably 2319 * fail without consequence, stop if we failed to reclaim 2320 * any pages from the last SWAP_CLUSTER_MAX number of 2321 * pages that were scanned. This will return to the 2322 * caller faster at the risk reclaim/compaction and 2323 * the resulting allocation attempt fails 2324 */ 2325 if (!nr_reclaimed) 2326 return false; 2327 } 2328 2329 /* 2330 * If we have not reclaimed enough pages for compaction and the 2331 * inactive lists are large enough, continue reclaiming 2332 */ 2333 pages_for_compaction = (2UL << sc->order); 2334 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE); 2335 if (get_nr_swap_pages() > 0) 2336 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON); 2337 if (sc->nr_reclaimed < pages_for_compaction && 2338 inactive_lru_pages > pages_for_compaction) 2339 return true; 2340 2341 /* If compaction would go ahead or the allocation would succeed, stop */ 2342 switch (compaction_suitable(zone, sc->order, 0, 0)) { 2343 case COMPACT_PARTIAL: 2344 case COMPACT_CONTINUE: 2345 return false; 2346 default: 2347 return true; 2348 } 2349 } 2350 2351 static bool shrink_zone(struct zone *zone, struct scan_control *sc, 2352 bool is_classzone) 2353 { 2354 struct reclaim_state *reclaim_state = current->reclaim_state; 2355 unsigned long nr_reclaimed, nr_scanned; 2356 bool reclaimable = false; 2357 2358 do { 2359 struct mem_cgroup *root = sc->target_mem_cgroup; 2360 struct mem_cgroup_reclaim_cookie reclaim = { 2361 .zone = zone, 2362 .priority = sc->priority, 2363 }; 2364 unsigned long zone_lru_pages = 0; 2365 struct mem_cgroup *memcg; 2366 2367 nr_reclaimed = sc->nr_reclaimed; 2368 nr_scanned = sc->nr_scanned; 2369 2370 memcg = mem_cgroup_iter(root, NULL, &reclaim); 2371 do { 2372 unsigned long lru_pages; 2373 unsigned long scanned; 2374 struct lruvec *lruvec; 2375 int swappiness; 2376 2377 if (mem_cgroup_low(root, memcg)) { 2378 if (!sc->may_thrash) 2379 continue; 2380 mem_cgroup_events(memcg, MEMCG_LOW, 1); 2381 } 2382 2383 lruvec = mem_cgroup_zone_lruvec(zone, memcg); 2384 swappiness = mem_cgroup_swappiness(memcg); 2385 scanned = sc->nr_scanned; 2386 2387 shrink_lruvec(lruvec, swappiness, sc, &lru_pages); 2388 zone_lru_pages += lru_pages; 2389 2390 if (memcg && is_classzone) 2391 shrink_slab(sc->gfp_mask, zone_to_nid(zone), 2392 memcg, sc->nr_scanned - scanned, 2393 lru_pages); 2394 2395 /* 2396 * Direct reclaim and kswapd have to scan all memory 2397 * cgroups to fulfill the overall scan target for the 2398 * zone. 2399 * 2400 * Limit reclaim, on the other hand, only cares about 2401 * nr_to_reclaim pages to be reclaimed and it will 2402 * retry with decreasing priority if one round over the 2403 * whole hierarchy is not sufficient. 2404 */ 2405 if (!global_reclaim(sc) && 2406 sc->nr_reclaimed >= sc->nr_to_reclaim) { 2407 mem_cgroup_iter_break(root, memcg); 2408 break; 2409 } 2410 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); 2411 2412 /* 2413 * Shrink the slab caches in the same proportion that 2414 * the eligible LRU pages were scanned. 2415 */ 2416 if (global_reclaim(sc) && is_classzone) 2417 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL, 2418 sc->nr_scanned - nr_scanned, 2419 zone_lru_pages); 2420 2421 if (reclaim_state) { 2422 sc->nr_reclaimed += reclaim_state->reclaimed_slab; 2423 reclaim_state->reclaimed_slab = 0; 2424 } 2425 2426 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, 2427 sc->nr_scanned - nr_scanned, 2428 sc->nr_reclaimed - nr_reclaimed); 2429 2430 if (sc->nr_reclaimed - nr_reclaimed) 2431 reclaimable = true; 2432 2433 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed, 2434 sc->nr_scanned - nr_scanned, sc)); 2435 2436 return reclaimable; 2437 } 2438 2439 /* 2440 * Returns true if compaction should go ahead for a high-order request, or 2441 * the high-order allocation would succeed without compaction. 2442 */ 2443 static inline bool compaction_ready(struct zone *zone, int order) 2444 { 2445 unsigned long balance_gap, watermark; 2446 bool watermark_ok; 2447 2448 /* 2449 * Compaction takes time to run and there are potentially other 2450 * callers using the pages just freed. Continue reclaiming until 2451 * there is a buffer of free pages available to give compaction 2452 * a reasonable chance of completing and allocating the page 2453 */ 2454 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( 2455 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); 2456 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order); 2457 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0); 2458 2459 /* 2460 * If compaction is deferred, reclaim up to a point where 2461 * compaction will have a chance of success when re-enabled 2462 */ 2463 if (compaction_deferred(zone, order)) 2464 return watermark_ok; 2465 2466 /* 2467 * If compaction is not ready to start and allocation is not likely 2468 * to succeed without it, then keep reclaiming. 2469 */ 2470 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED) 2471 return false; 2472 2473 return watermark_ok; 2474 } 2475 2476 /* 2477 * This is the direct reclaim path, for page-allocating processes. We only 2478 * try to reclaim pages from zones which will satisfy the caller's allocation 2479 * request. 2480 * 2481 * We reclaim from a zone even if that zone is over high_wmark_pages(zone). 2482 * Because: 2483 * a) The caller may be trying to free *extra* pages to satisfy a higher-order 2484 * allocation or 2485 * b) The target zone may be at high_wmark_pages(zone) but the lower zones 2486 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' 2487 * zone defense algorithm. 2488 * 2489 * If a zone is deemed to be full of pinned pages then just give it a light 2490 * scan then give up on it. 2491 * 2492 * Returns true if a zone was reclaimable. 2493 */ 2494 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc) 2495 { 2496 struct zoneref *z; 2497 struct zone *zone; 2498 unsigned long nr_soft_reclaimed; 2499 unsigned long nr_soft_scanned; 2500 gfp_t orig_mask; 2501 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask); 2502 bool reclaimable = false; 2503 2504 /* 2505 * If the number of buffer_heads in the machine exceeds the maximum 2506 * allowed level, force direct reclaim to scan the highmem zone as 2507 * highmem pages could be pinning lowmem pages storing buffer_heads 2508 */ 2509 orig_mask = sc->gfp_mask; 2510 if (buffer_heads_over_limit) 2511 sc->gfp_mask |= __GFP_HIGHMEM; 2512 2513 for_each_zone_zonelist_nodemask(zone, z, zonelist, 2514 requested_highidx, sc->nodemask) { 2515 enum zone_type classzone_idx; 2516 2517 if (!populated_zone(zone)) 2518 continue; 2519 2520 classzone_idx = requested_highidx; 2521 while (!populated_zone(zone->zone_pgdat->node_zones + 2522 classzone_idx)) 2523 classzone_idx--; 2524 2525 /* 2526 * Take care memory controller reclaiming has small influence 2527 * to global LRU. 2528 */ 2529 if (global_reclaim(sc)) { 2530 if (!cpuset_zone_allowed(zone, 2531 GFP_KERNEL | __GFP_HARDWALL)) 2532 continue; 2533 2534 if (sc->priority != DEF_PRIORITY && 2535 !zone_reclaimable(zone)) 2536 continue; /* Let kswapd poll it */ 2537 2538 /* 2539 * If we already have plenty of memory free for 2540 * compaction in this zone, don't free any more. 2541 * Even though compaction is invoked for any 2542 * non-zero order, only frequent costly order 2543 * reclamation is disruptive enough to become a 2544 * noticeable problem, like transparent huge 2545 * page allocations. 2546 */ 2547 if (IS_ENABLED(CONFIG_COMPACTION) && 2548 sc->order > PAGE_ALLOC_COSTLY_ORDER && 2549 zonelist_zone_idx(z) <= requested_highidx && 2550 compaction_ready(zone, sc->order)) { 2551 sc->compaction_ready = true; 2552 continue; 2553 } 2554 2555 /* 2556 * This steals pages from memory cgroups over softlimit 2557 * and returns the number of reclaimed pages and 2558 * scanned pages. This works for global memory pressure 2559 * and balancing, not for a memcg's limit. 2560 */ 2561 nr_soft_scanned = 0; 2562 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, 2563 sc->order, sc->gfp_mask, 2564 &nr_soft_scanned); 2565 sc->nr_reclaimed += nr_soft_reclaimed; 2566 sc->nr_scanned += nr_soft_scanned; 2567 if (nr_soft_reclaimed) 2568 reclaimable = true; 2569 /* need some check for avoid more shrink_zone() */ 2570 } 2571 2572 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx)) 2573 reclaimable = true; 2574 2575 if (global_reclaim(sc) && 2576 !reclaimable && zone_reclaimable(zone)) 2577 reclaimable = true; 2578 } 2579 2580 /* 2581 * Restore to original mask to avoid the impact on the caller if we 2582 * promoted it to __GFP_HIGHMEM. 2583 */ 2584 sc->gfp_mask = orig_mask; 2585 2586 return reclaimable; 2587 } 2588 2589 /* 2590 * This is the main entry point to direct page reclaim. 2591 * 2592 * If a full scan of the inactive list fails to free enough memory then we 2593 * are "out of memory" and something needs to be killed. 2594 * 2595 * If the caller is !__GFP_FS then the probability of a failure is reasonably 2596 * high - the zone may be full of dirty or under-writeback pages, which this 2597 * caller can't do much about. We kick the writeback threads and take explicit 2598 * naps in the hope that some of these pages can be written. But if the 2599 * allocating task holds filesystem locks which prevent writeout this might not 2600 * work, and the allocation attempt will fail. 2601 * 2602 * returns: 0, if no pages reclaimed 2603 * else, the number of pages reclaimed 2604 */ 2605 static unsigned long do_try_to_free_pages(struct zonelist *zonelist, 2606 struct scan_control *sc) 2607 { 2608 int initial_priority = sc->priority; 2609 unsigned long total_scanned = 0; 2610 unsigned long writeback_threshold; 2611 bool zones_reclaimable; 2612 retry: 2613 delayacct_freepages_start(); 2614 2615 if (global_reclaim(sc)) 2616 count_vm_event(ALLOCSTALL); 2617 2618 do { 2619 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, 2620 sc->priority); 2621 sc->nr_scanned = 0; 2622 zones_reclaimable = shrink_zones(zonelist, sc); 2623 2624 total_scanned += sc->nr_scanned; 2625 if (sc->nr_reclaimed >= sc->nr_to_reclaim) 2626 break; 2627 2628 if (sc->compaction_ready) 2629 break; 2630 2631 /* 2632 * If we're getting trouble reclaiming, start doing 2633 * writepage even in laptop mode. 2634 */ 2635 if (sc->priority < DEF_PRIORITY - 2) 2636 sc->may_writepage = 1; 2637 2638 /* 2639 * Try to write back as many pages as we just scanned. This 2640 * tends to cause slow streaming writers to write data to the 2641 * disk smoothly, at the dirtying rate, which is nice. But 2642 * that's undesirable in laptop mode, where we *want* lumpy 2643 * writeout. So in laptop mode, write out the whole world. 2644 */ 2645 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; 2646 if (total_scanned > writeback_threshold) { 2647 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, 2648 WB_REASON_TRY_TO_FREE_PAGES); 2649 sc->may_writepage = 1; 2650 } 2651 } while (--sc->priority >= 0); 2652 2653 delayacct_freepages_end(); 2654 2655 if (sc->nr_reclaimed) 2656 return sc->nr_reclaimed; 2657 2658 /* Aborted reclaim to try compaction? don't OOM, then */ 2659 if (sc->compaction_ready) 2660 return 1; 2661 2662 /* Untapped cgroup reserves? Don't OOM, retry. */ 2663 if (!sc->may_thrash) { 2664 sc->priority = initial_priority; 2665 sc->may_thrash = 1; 2666 goto retry; 2667 } 2668 2669 /* Any of the zones still reclaimable? Don't OOM. */ 2670 if (zones_reclaimable) 2671 return 1; 2672 2673 return 0; 2674 } 2675 2676 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) 2677 { 2678 struct zone *zone; 2679 unsigned long pfmemalloc_reserve = 0; 2680 unsigned long free_pages = 0; 2681 int i; 2682 bool wmark_ok; 2683 2684 for (i = 0; i <= ZONE_NORMAL; i++) { 2685 zone = &pgdat->node_zones[i]; 2686 if (!populated_zone(zone) || 2687 zone_reclaimable_pages(zone) == 0) 2688 continue; 2689 2690 pfmemalloc_reserve += min_wmark_pages(zone); 2691 free_pages += zone_page_state(zone, NR_FREE_PAGES); 2692 } 2693 2694 /* If there are no reserves (unexpected config) then do not throttle */ 2695 if (!pfmemalloc_reserve) 2696 return true; 2697 2698 wmark_ok = free_pages > pfmemalloc_reserve / 2; 2699 2700 /* kswapd must be awake if processes are being throttled */ 2701 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { 2702 pgdat->classzone_idx = min(pgdat->classzone_idx, 2703 (enum zone_type)ZONE_NORMAL); 2704 wake_up_interruptible(&pgdat->kswapd_wait); 2705 } 2706 2707 return wmark_ok; 2708 } 2709 2710 /* 2711 * Throttle direct reclaimers if backing storage is backed by the network 2712 * and the PFMEMALLOC reserve for the preferred node is getting dangerously 2713 * depleted. kswapd will continue to make progress and wake the processes 2714 * when the low watermark is reached. 2715 * 2716 * Returns true if a fatal signal was delivered during throttling. If this 2717 * happens, the page allocator should not consider triggering the OOM killer. 2718 */ 2719 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, 2720 nodemask_t *nodemask) 2721 { 2722 struct zoneref *z; 2723 struct zone *zone; 2724 pg_data_t *pgdat = NULL; 2725 2726 /* 2727 * Kernel threads should not be throttled as they may be indirectly 2728 * responsible for cleaning pages necessary for reclaim to make forward 2729 * progress. kjournald for example may enter direct reclaim while 2730 * committing a transaction where throttling it could forcing other 2731 * processes to block on log_wait_commit(). 2732 */ 2733 if (current->flags & PF_KTHREAD) 2734 goto out; 2735 2736 /* 2737 * If a fatal signal is pending, this process should not throttle. 2738 * It should return quickly so it can exit and free its memory 2739 */ 2740 if (fatal_signal_pending(current)) 2741 goto out; 2742 2743 /* 2744 * Check if the pfmemalloc reserves are ok by finding the first node 2745 * with a usable ZONE_NORMAL or lower zone. The expectation is that 2746 * GFP_KERNEL will be required for allocating network buffers when 2747 * swapping over the network so ZONE_HIGHMEM is unusable. 2748 * 2749 * Throttling is based on the first usable node and throttled processes 2750 * wait on a queue until kswapd makes progress and wakes them. There 2751 * is an affinity then between processes waking up and where reclaim 2752 * progress has been made assuming the process wakes on the same node. 2753 * More importantly, processes running on remote nodes will not compete 2754 * for remote pfmemalloc reserves and processes on different nodes 2755 * should make reasonable progress. 2756 */ 2757 for_each_zone_zonelist_nodemask(zone, z, zonelist, 2758 gfp_zone(gfp_mask), nodemask) { 2759 if (zone_idx(zone) > ZONE_NORMAL) 2760 continue; 2761 2762 /* Throttle based on the first usable node */ 2763 pgdat = zone->zone_pgdat; 2764 if (pfmemalloc_watermark_ok(pgdat)) 2765 goto out; 2766 break; 2767 } 2768 2769 /* If no zone was usable by the allocation flags then do not throttle */ 2770 if (!pgdat) 2771 goto out; 2772 2773 /* Account for the throttling */ 2774 count_vm_event(PGSCAN_DIRECT_THROTTLE); 2775 2776 /* 2777 * If the caller cannot enter the filesystem, it's possible that it 2778 * is due to the caller holding an FS lock or performing a journal 2779 * transaction in the case of a filesystem like ext[3|4]. In this case, 2780 * it is not safe to block on pfmemalloc_wait as kswapd could be 2781 * blocked waiting on the same lock. Instead, throttle for up to a 2782 * second before continuing. 2783 */ 2784 if (!(gfp_mask & __GFP_FS)) { 2785 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, 2786 pfmemalloc_watermark_ok(pgdat), HZ); 2787 2788 goto check_pending; 2789 } 2790 2791 /* Throttle until kswapd wakes the process */ 2792 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, 2793 pfmemalloc_watermark_ok(pgdat)); 2794 2795 check_pending: 2796 if (fatal_signal_pending(current)) 2797 return true; 2798 2799 out: 2800 return false; 2801 } 2802 2803 unsigned long try_to_free_pages(struct zonelist *zonelist, int order, 2804 gfp_t gfp_mask, nodemask_t *nodemask) 2805 { 2806 unsigned long nr_reclaimed; 2807 struct scan_control sc = { 2808 .nr_to_reclaim = SWAP_CLUSTER_MAX, 2809 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), 2810 .order = order, 2811 .nodemask = nodemask, 2812 .priority = DEF_PRIORITY, 2813 .may_writepage = !laptop_mode, 2814 .may_unmap = 1, 2815 .may_swap = 1, 2816 }; 2817 2818 /* 2819 * Do not enter reclaim if fatal signal was delivered while throttled. 2820 * 1 is returned so that the page allocator does not OOM kill at this 2821 * point. 2822 */ 2823 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) 2824 return 1; 2825 2826 trace_mm_vmscan_direct_reclaim_begin(order, 2827 sc.may_writepage, 2828 gfp_mask); 2829 2830 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 2831 2832 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); 2833 2834 return nr_reclaimed; 2835 } 2836 2837 #ifdef CONFIG_MEMCG 2838 2839 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, 2840 gfp_t gfp_mask, bool noswap, 2841 struct zone *zone, 2842 unsigned long *nr_scanned) 2843 { 2844 struct scan_control sc = { 2845 .nr_to_reclaim = SWAP_CLUSTER_MAX, 2846 .target_mem_cgroup = memcg, 2847 .may_writepage = !laptop_mode, 2848 .may_unmap = 1, 2849 .may_swap = !noswap, 2850 }; 2851 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); 2852 int swappiness = mem_cgroup_swappiness(memcg); 2853 unsigned long lru_pages; 2854 2855 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | 2856 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); 2857 2858 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, 2859 sc.may_writepage, 2860 sc.gfp_mask); 2861 2862 /* 2863 * NOTE: Although we can get the priority field, using it 2864 * here is not a good idea, since it limits the pages we can scan. 2865 * if we don't reclaim here, the shrink_zone from balance_pgdat 2866 * will pick up pages from other mem cgroup's as well. We hack 2867 * the priority and make it zero. 2868 */ 2869 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages); 2870 2871 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); 2872 2873 *nr_scanned = sc.nr_scanned; 2874 return sc.nr_reclaimed; 2875 } 2876 2877 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, 2878 unsigned long nr_pages, 2879 gfp_t gfp_mask, 2880 bool may_swap) 2881 { 2882 struct zonelist *zonelist; 2883 unsigned long nr_reclaimed; 2884 int nid; 2885 struct scan_control sc = { 2886 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), 2887 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | 2888 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), 2889 .target_mem_cgroup = memcg, 2890 .priority = DEF_PRIORITY, 2891 .may_writepage = !laptop_mode, 2892 .may_unmap = 1, 2893 .may_swap = may_swap, 2894 }; 2895 2896 /* 2897 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't 2898 * take care of from where we get pages. So the node where we start the 2899 * scan does not need to be the current node. 2900 */ 2901 nid = mem_cgroup_select_victim_node(memcg); 2902 2903 zonelist = NODE_DATA(nid)->node_zonelists; 2904 2905 trace_mm_vmscan_memcg_reclaim_begin(0, 2906 sc.may_writepage, 2907 sc.gfp_mask); 2908 2909 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 2910 2911 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); 2912 2913 return nr_reclaimed; 2914 } 2915 #endif 2916 2917 static void age_active_anon(struct zone *zone, struct scan_control *sc) 2918 { 2919 struct mem_cgroup *memcg; 2920 2921 if (!total_swap_pages) 2922 return; 2923 2924 memcg = mem_cgroup_iter(NULL, NULL, NULL); 2925 do { 2926 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); 2927 2928 if (inactive_anon_is_low(lruvec)) 2929 shrink_active_list(SWAP_CLUSTER_MAX, lruvec, 2930 sc, LRU_ACTIVE_ANON); 2931 2932 memcg = mem_cgroup_iter(NULL, memcg, NULL); 2933 } while (memcg); 2934 } 2935 2936 static bool zone_balanced(struct zone *zone, int order, 2937 unsigned long balance_gap, int classzone_idx) 2938 { 2939 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) + 2940 balance_gap, classzone_idx, 0)) 2941 return false; 2942 2943 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone, 2944 order, 0, classzone_idx) == COMPACT_SKIPPED) 2945 return false; 2946 2947 return true; 2948 } 2949 2950 /* 2951 * pgdat_balanced() is used when checking if a node is balanced. 2952 * 2953 * For order-0, all zones must be balanced! 2954 * 2955 * For high-order allocations only zones that meet watermarks and are in a 2956 * zone allowed by the callers classzone_idx are added to balanced_pages. The 2957 * total of balanced pages must be at least 25% of the zones allowed by 2958 * classzone_idx for the node to be considered balanced. Forcing all zones to 2959 * be balanced for high orders can cause excessive reclaim when there are 2960 * imbalanced zones. 2961 * The choice of 25% is due to 2962 * o a 16M DMA zone that is balanced will not balance a zone on any 2963 * reasonable sized machine 2964 * o On all other machines, the top zone must be at least a reasonable 2965 * percentage of the middle zones. For example, on 32-bit x86, highmem 2966 * would need to be at least 256M for it to be balance a whole node. 2967 * Similarly, on x86-64 the Normal zone would need to be at least 1G 2968 * to balance a node on its own. These seemed like reasonable ratios. 2969 */ 2970 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) 2971 { 2972 unsigned long managed_pages = 0; 2973 unsigned long balanced_pages = 0; 2974 int i; 2975 2976 /* Check the watermark levels */ 2977 for (i = 0; i <= classzone_idx; i++) { 2978 struct zone *zone = pgdat->node_zones + i; 2979 2980 if (!populated_zone(zone)) 2981 continue; 2982 2983 managed_pages += zone->managed_pages; 2984 2985 /* 2986 * A special case here: 2987 * 2988 * balance_pgdat() skips over all_unreclaimable after 2989 * DEF_PRIORITY. Effectively, it considers them balanced so 2990 * they must be considered balanced here as well! 2991 */ 2992 if (!zone_reclaimable(zone)) { 2993 balanced_pages += zone->managed_pages; 2994 continue; 2995 } 2996 2997 if (zone_balanced(zone, order, 0, i)) 2998 balanced_pages += zone->managed_pages; 2999 else if (!order) 3000 return false; 3001 } 3002 3003 if (order) 3004 return balanced_pages >= (managed_pages >> 2); 3005 else 3006 return true; 3007 } 3008 3009 /* 3010 * Prepare kswapd for sleeping. This verifies that there are no processes 3011 * waiting in throttle_direct_reclaim() and that watermarks have been met. 3012 * 3013 * Returns true if kswapd is ready to sleep 3014 */ 3015 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining, 3016 int classzone_idx) 3017 { 3018 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ 3019 if (remaining) 3020 return false; 3021 3022 /* 3023 * The throttled processes are normally woken up in balance_pgdat() as 3024 * soon as pfmemalloc_watermark_ok() is true. But there is a potential 3025 * race between when kswapd checks the watermarks and a process gets 3026 * throttled. There is also a potential race if processes get 3027 * throttled, kswapd wakes, a large process exits thereby balancing the 3028 * zones, which causes kswapd to exit balance_pgdat() before reaching 3029 * the wake up checks. If kswapd is going to sleep, no process should 3030 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If 3031 * the wake up is premature, processes will wake kswapd and get 3032 * throttled again. The difference from wake ups in balance_pgdat() is 3033 * that here we are under prepare_to_wait(). 3034 */ 3035 if (waitqueue_active(&pgdat->pfmemalloc_wait)) 3036 wake_up_all(&pgdat->pfmemalloc_wait); 3037 3038 return pgdat_balanced(pgdat, order, classzone_idx); 3039 } 3040 3041 /* 3042 * kswapd shrinks the zone by the number of pages required to reach 3043 * the high watermark. 3044 * 3045 * Returns true if kswapd scanned at least the requested number of pages to 3046 * reclaim or if the lack of progress was due to pages under writeback. 3047 * This is used to determine if the scanning priority needs to be raised. 3048 */ 3049 static bool kswapd_shrink_zone(struct zone *zone, 3050 int classzone_idx, 3051 struct scan_control *sc, 3052 unsigned long *nr_attempted) 3053 { 3054 int testorder = sc->order; 3055 unsigned long balance_gap; 3056 bool lowmem_pressure; 3057 3058 /* Reclaim above the high watermark. */ 3059 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone)); 3060 3061 /* 3062 * Kswapd reclaims only single pages with compaction enabled. Trying 3063 * too hard to reclaim until contiguous free pages have become 3064 * available can hurt performance by evicting too much useful data 3065 * from memory. Do not reclaim more than needed for compaction. 3066 */ 3067 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && 3068 compaction_suitable(zone, sc->order, 0, classzone_idx) 3069 != COMPACT_SKIPPED) 3070 testorder = 0; 3071 3072 /* 3073 * We put equal pressure on every zone, unless one zone has way too 3074 * many pages free already. The "too many pages" is defined as the 3075 * high wmark plus a "gap" where the gap is either the low 3076 * watermark or 1% of the zone, whichever is smaller. 3077 */ 3078 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( 3079 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); 3080 3081 /* 3082 * If there is no low memory pressure or the zone is balanced then no 3083 * reclaim is necessary 3084 */ 3085 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone)); 3086 if (!lowmem_pressure && zone_balanced(zone, testorder, 3087 balance_gap, classzone_idx)) 3088 return true; 3089 3090 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx); 3091 3092 /* Account for the number of pages attempted to reclaim */ 3093 *nr_attempted += sc->nr_to_reclaim; 3094 3095 clear_bit(ZONE_WRITEBACK, &zone->flags); 3096 3097 /* 3098 * If a zone reaches its high watermark, consider it to be no longer 3099 * congested. It's possible there are dirty pages backed by congested 3100 * BDIs but as pressure is relieved, speculatively avoid congestion 3101 * waits. 3102 */ 3103 if (zone_reclaimable(zone) && 3104 zone_balanced(zone, testorder, 0, classzone_idx)) { 3105 clear_bit(ZONE_CONGESTED, &zone->flags); 3106 clear_bit(ZONE_DIRTY, &zone->flags); 3107 } 3108 3109 return sc->nr_scanned >= sc->nr_to_reclaim; 3110 } 3111 3112 /* 3113 * For kswapd, balance_pgdat() will work across all this node's zones until 3114 * they are all at high_wmark_pages(zone). 3115 * 3116 * Returns the final order kswapd was reclaiming at 3117 * 3118 * There is special handling here for zones which are full of pinned pages. 3119 * This can happen if the pages are all mlocked, or if they are all used by 3120 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. 3121 * What we do is to detect the case where all pages in the zone have been 3122 * scanned twice and there has been zero successful reclaim. Mark the zone as 3123 * dead and from now on, only perform a short scan. Basically we're polling 3124 * the zone for when the problem goes away. 3125 * 3126 * kswapd scans the zones in the highmem->normal->dma direction. It skips 3127 * zones which have free_pages > high_wmark_pages(zone), but once a zone is 3128 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the 3129 * lower zones regardless of the number of free pages in the lower zones. This 3130 * interoperates with the page allocator fallback scheme to ensure that aging 3131 * of pages is balanced across the zones. 3132 */ 3133 static unsigned long balance_pgdat(pg_data_t *pgdat, int order, 3134 int *classzone_idx) 3135 { 3136 int i; 3137 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ 3138 unsigned long nr_soft_reclaimed; 3139 unsigned long nr_soft_scanned; 3140 struct scan_control sc = { 3141 .gfp_mask = GFP_KERNEL, 3142 .order = order, 3143 .priority = DEF_PRIORITY, 3144 .may_writepage = !laptop_mode, 3145 .may_unmap = 1, 3146 .may_swap = 1, 3147 }; 3148 count_vm_event(PAGEOUTRUN); 3149 3150 do { 3151 unsigned long nr_attempted = 0; 3152 bool raise_priority = true; 3153 bool pgdat_needs_compaction = (order > 0); 3154 3155 sc.nr_reclaimed = 0; 3156 3157 /* 3158 * Scan in the highmem->dma direction for the highest 3159 * zone which needs scanning 3160 */ 3161 for (i = pgdat->nr_zones - 1; i >= 0; i--) { 3162 struct zone *zone = pgdat->node_zones + i; 3163 3164 if (!populated_zone(zone)) 3165 continue; 3166 3167 if (sc.priority != DEF_PRIORITY && 3168 !zone_reclaimable(zone)) 3169 continue; 3170 3171 /* 3172 * Do some background aging of the anon list, to give 3173 * pages a chance to be referenced before reclaiming. 3174 */ 3175 age_active_anon(zone, &sc); 3176 3177 /* 3178 * If the number of buffer_heads in the machine 3179 * exceeds the maximum allowed level and this node 3180 * has a highmem zone, force kswapd to reclaim from 3181 * it to relieve lowmem pressure. 3182 */ 3183 if (buffer_heads_over_limit && is_highmem_idx(i)) { 3184 end_zone = i; 3185 break; 3186 } 3187 3188 if (!zone_balanced(zone, order, 0, 0)) { 3189 end_zone = i; 3190 break; 3191 } else { 3192 /* 3193 * If balanced, clear the dirty and congested 3194 * flags 3195 */ 3196 clear_bit(ZONE_CONGESTED, &zone->flags); 3197 clear_bit(ZONE_DIRTY, &zone->flags); 3198 } 3199 } 3200 3201 if (i < 0) 3202 goto out; 3203 3204 for (i = 0; i <= end_zone; i++) { 3205 struct zone *zone = pgdat->node_zones + i; 3206 3207 if (!populated_zone(zone)) 3208 continue; 3209 3210 /* 3211 * If any zone is currently balanced then kswapd will 3212 * not call compaction as it is expected that the 3213 * necessary pages are already available. 3214 */ 3215 if (pgdat_needs_compaction && 3216 zone_watermark_ok(zone, order, 3217 low_wmark_pages(zone), 3218 *classzone_idx, 0)) 3219 pgdat_needs_compaction = false; 3220 } 3221 3222 /* 3223 * If we're getting trouble reclaiming, start doing writepage 3224 * even in laptop mode. 3225 */ 3226 if (sc.priority < DEF_PRIORITY - 2) 3227 sc.may_writepage = 1; 3228 3229 /* 3230 * Now scan the zone in the dma->highmem direction, stopping 3231 * at the last zone which needs scanning. 3232 * 3233 * We do this because the page allocator works in the opposite 3234 * direction. This prevents the page allocator from allocating 3235 * pages behind kswapd's direction of progress, which would 3236 * cause too much scanning of the lower zones. 3237 */ 3238 for (i = 0; i <= end_zone; i++) { 3239 struct zone *zone = pgdat->node_zones + i; 3240 3241 if (!populated_zone(zone)) 3242 continue; 3243 3244 if (sc.priority != DEF_PRIORITY && 3245 !zone_reclaimable(zone)) 3246 continue; 3247 3248 sc.nr_scanned = 0; 3249 3250 nr_soft_scanned = 0; 3251 /* 3252 * Call soft limit reclaim before calling shrink_zone. 3253 */ 3254 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, 3255 order, sc.gfp_mask, 3256 &nr_soft_scanned); 3257 sc.nr_reclaimed += nr_soft_reclaimed; 3258 3259 /* 3260 * There should be no need to raise the scanning 3261 * priority if enough pages are already being scanned 3262 * that that high watermark would be met at 100% 3263 * efficiency. 3264 */ 3265 if (kswapd_shrink_zone(zone, end_zone, 3266 &sc, &nr_attempted)) 3267 raise_priority = false; 3268 } 3269 3270 /* 3271 * If the low watermark is met there is no need for processes 3272 * to be throttled on pfmemalloc_wait as they should not be 3273 * able to safely make forward progress. Wake them 3274 */ 3275 if (waitqueue_active(&pgdat->pfmemalloc_wait) && 3276 pfmemalloc_watermark_ok(pgdat)) 3277 wake_up_all(&pgdat->pfmemalloc_wait); 3278 3279 /* 3280 * Fragmentation may mean that the system cannot be rebalanced 3281 * for high-order allocations in all zones. If twice the 3282 * allocation size has been reclaimed and the zones are still 3283 * not balanced then recheck the watermarks at order-0 to 3284 * prevent kswapd reclaiming excessively. Assume that a 3285 * process requested a high-order can direct reclaim/compact. 3286 */ 3287 if (order && sc.nr_reclaimed >= 2UL << order) 3288 order = sc.order = 0; 3289 3290 /* Check if kswapd should be suspending */ 3291 if (try_to_freeze() || kthread_should_stop()) 3292 break; 3293 3294 /* 3295 * Compact if necessary and kswapd is reclaiming at least the 3296 * high watermark number of pages as requsted 3297 */ 3298 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted) 3299 compact_pgdat(pgdat, order); 3300 3301 /* 3302 * Raise priority if scanning rate is too low or there was no 3303 * progress in reclaiming pages 3304 */ 3305 if (raise_priority || !sc.nr_reclaimed) 3306 sc.priority--; 3307 } while (sc.priority >= 1 && 3308 !pgdat_balanced(pgdat, order, *classzone_idx)); 3309 3310 out: 3311 /* 3312 * Return the order we were reclaiming at so prepare_kswapd_sleep() 3313 * makes a decision on the order we were last reclaiming at. However, 3314 * if another caller entered the allocator slow path while kswapd 3315 * was awake, order will remain at the higher level 3316 */ 3317 *classzone_idx = end_zone; 3318 return order; 3319 } 3320 3321 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) 3322 { 3323 long remaining = 0; 3324 DEFINE_WAIT(wait); 3325 3326 if (freezing(current) || kthread_should_stop()) 3327 return; 3328 3329 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 3330 3331 /* Try to sleep for a short interval */ 3332 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { 3333 remaining = schedule_timeout(HZ/10); 3334 finish_wait(&pgdat->kswapd_wait, &wait); 3335 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 3336 } 3337 3338 /* 3339 * After a short sleep, check if it was a premature sleep. If not, then 3340 * go fully to sleep until explicitly woken up. 3341 */ 3342 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { 3343 trace_mm_vmscan_kswapd_sleep(pgdat->node_id); 3344 3345 /* 3346 * vmstat counters are not perfectly accurate and the estimated 3347 * value for counters such as NR_FREE_PAGES can deviate from the 3348 * true value by nr_online_cpus * threshold. To avoid the zone 3349 * watermarks being breached while under pressure, we reduce the 3350 * per-cpu vmstat threshold while kswapd is awake and restore 3351 * them before going back to sleep. 3352 */ 3353 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); 3354 3355 /* 3356 * Compaction records what page blocks it recently failed to 3357 * isolate pages from and skips them in the future scanning. 3358 * When kswapd is going to sleep, it is reasonable to assume 3359 * that pages and compaction may succeed so reset the cache. 3360 */ 3361 reset_isolation_suitable(pgdat); 3362 3363 if (!kthread_should_stop()) 3364 schedule(); 3365 3366 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); 3367 } else { 3368 if (remaining) 3369 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); 3370 else 3371 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); 3372 } 3373 finish_wait(&pgdat->kswapd_wait, &wait); 3374 } 3375 3376 /* 3377 * The background pageout daemon, started as a kernel thread 3378 * from the init process. 3379 * 3380 * This basically trickles out pages so that we have _some_ 3381 * free memory available even if there is no other activity 3382 * that frees anything up. This is needed for things like routing 3383 * etc, where we otherwise might have all activity going on in 3384 * asynchronous contexts that cannot page things out. 3385 * 3386 * If there are applications that are active memory-allocators 3387 * (most normal use), this basically shouldn't matter. 3388 */ 3389 static int kswapd(void *p) 3390 { 3391 unsigned long order, new_order; 3392 unsigned balanced_order; 3393 int classzone_idx, new_classzone_idx; 3394 int balanced_classzone_idx; 3395 pg_data_t *pgdat = (pg_data_t*)p; 3396 struct task_struct *tsk = current; 3397 3398 struct reclaim_state reclaim_state = { 3399 .reclaimed_slab = 0, 3400 }; 3401 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); 3402 3403 lockdep_set_current_reclaim_state(GFP_KERNEL); 3404 3405 if (!cpumask_empty(cpumask)) 3406 set_cpus_allowed_ptr(tsk, cpumask); 3407 current->reclaim_state = &reclaim_state; 3408 3409 /* 3410 * Tell the memory management that we're a "memory allocator", 3411 * and that if we need more memory we should get access to it 3412 * regardless (see "__alloc_pages()"). "kswapd" should 3413 * never get caught in the normal page freeing logic. 3414 * 3415 * (Kswapd normally doesn't need memory anyway, but sometimes 3416 * you need a small amount of memory in order to be able to 3417 * page out something else, and this flag essentially protects 3418 * us from recursively trying to free more memory as we're 3419 * trying to free the first piece of memory in the first place). 3420 */ 3421 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; 3422 set_freezable(); 3423 3424 order = new_order = 0; 3425 balanced_order = 0; 3426 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; 3427 balanced_classzone_idx = classzone_idx; 3428 for ( ; ; ) { 3429 bool ret; 3430 3431 /* 3432 * If the last balance_pgdat was unsuccessful it's unlikely a 3433 * new request of a similar or harder type will succeed soon 3434 * so consider going to sleep on the basis we reclaimed at 3435 */ 3436 if (balanced_classzone_idx >= new_classzone_idx && 3437 balanced_order == new_order) { 3438 new_order = pgdat->kswapd_max_order; 3439 new_classzone_idx = pgdat->classzone_idx; 3440 pgdat->kswapd_max_order = 0; 3441 pgdat->classzone_idx = pgdat->nr_zones - 1; 3442 } 3443 3444 if (order < new_order || classzone_idx > new_classzone_idx) { 3445 /* 3446 * Don't sleep if someone wants a larger 'order' 3447 * allocation or has tigher zone constraints 3448 */ 3449 order = new_order; 3450 classzone_idx = new_classzone_idx; 3451 } else { 3452 kswapd_try_to_sleep(pgdat, balanced_order, 3453 balanced_classzone_idx); 3454 order = pgdat->kswapd_max_order; 3455 classzone_idx = pgdat->classzone_idx; 3456 new_order = order; 3457 new_classzone_idx = classzone_idx; 3458 pgdat->kswapd_max_order = 0; 3459 pgdat->classzone_idx = pgdat->nr_zones - 1; 3460 } 3461 3462 ret = try_to_freeze(); 3463 if (kthread_should_stop()) 3464 break; 3465 3466 /* 3467 * We can speed up thawing tasks if we don't call balance_pgdat 3468 * after returning from the refrigerator 3469 */ 3470 if (!ret) { 3471 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); 3472 balanced_classzone_idx = classzone_idx; 3473 balanced_order = balance_pgdat(pgdat, order, 3474 &balanced_classzone_idx); 3475 } 3476 } 3477 3478 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); 3479 current->reclaim_state = NULL; 3480 lockdep_clear_current_reclaim_state(); 3481 3482 return 0; 3483 } 3484 3485 /* 3486 * A zone is low on free memory, so wake its kswapd task to service it. 3487 */ 3488 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) 3489 { 3490 pg_data_t *pgdat; 3491 3492 if (!populated_zone(zone)) 3493 return; 3494 3495 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) 3496 return; 3497 pgdat = zone->zone_pgdat; 3498 if (pgdat->kswapd_max_order < order) { 3499 pgdat->kswapd_max_order = order; 3500 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); 3501 } 3502 if (!waitqueue_active(&pgdat->kswapd_wait)) 3503 return; 3504 if (zone_balanced(zone, order, 0, 0)) 3505 return; 3506 3507 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); 3508 wake_up_interruptible(&pgdat->kswapd_wait); 3509 } 3510 3511 #ifdef CONFIG_HIBERNATION 3512 /* 3513 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of 3514 * freed pages. 3515 * 3516 * Rather than trying to age LRUs the aim is to preserve the overall 3517 * LRU order by reclaiming preferentially 3518 * inactive > active > active referenced > active mapped 3519 */ 3520 unsigned long shrink_all_memory(unsigned long nr_to_reclaim) 3521 { 3522 struct reclaim_state reclaim_state; 3523 struct scan_control sc = { 3524 .nr_to_reclaim = nr_to_reclaim, 3525 .gfp_mask = GFP_HIGHUSER_MOVABLE, 3526 .priority = DEF_PRIORITY, 3527 .may_writepage = 1, 3528 .may_unmap = 1, 3529 .may_swap = 1, 3530 .hibernation_mode = 1, 3531 }; 3532 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); 3533 struct task_struct *p = current; 3534 unsigned long nr_reclaimed; 3535 3536 p->flags |= PF_MEMALLOC; 3537 lockdep_set_current_reclaim_state(sc.gfp_mask); 3538 reclaim_state.reclaimed_slab = 0; 3539 p->reclaim_state = &reclaim_state; 3540 3541 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 3542 3543 p->reclaim_state = NULL; 3544 lockdep_clear_current_reclaim_state(); 3545 p->flags &= ~PF_MEMALLOC; 3546 3547 return nr_reclaimed; 3548 } 3549 #endif /* CONFIG_HIBERNATION */ 3550 3551 /* It's optimal to keep kswapds on the same CPUs as their memory, but 3552 not required for correctness. So if the last cpu in a node goes 3553 away, we get changed to run anywhere: as the first one comes back, 3554 restore their cpu bindings. */ 3555 static int cpu_callback(struct notifier_block *nfb, unsigned long action, 3556 void *hcpu) 3557 { 3558 int nid; 3559 3560 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { 3561 for_each_node_state(nid, N_MEMORY) { 3562 pg_data_t *pgdat = NODE_DATA(nid); 3563 const struct cpumask *mask; 3564 3565 mask = cpumask_of_node(pgdat->node_id); 3566 3567 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) 3568 /* One of our CPUs online: restore mask */ 3569 set_cpus_allowed_ptr(pgdat->kswapd, mask); 3570 } 3571 } 3572 return NOTIFY_OK; 3573 } 3574 3575 /* 3576 * This kswapd start function will be called by init and node-hot-add. 3577 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. 3578 */ 3579 int kswapd_run(int nid) 3580 { 3581 pg_data_t *pgdat = NODE_DATA(nid); 3582 int ret = 0; 3583 3584 if (pgdat->kswapd) 3585 return 0; 3586 3587 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); 3588 if (IS_ERR(pgdat->kswapd)) { 3589 /* failure at boot is fatal */ 3590 BUG_ON(system_state == SYSTEM_BOOTING); 3591 pr_err("Failed to start kswapd on node %d\n", nid); 3592 ret = PTR_ERR(pgdat->kswapd); 3593 pgdat->kswapd = NULL; 3594 } 3595 return ret; 3596 } 3597 3598 /* 3599 * Called by memory hotplug when all memory in a node is offlined. Caller must 3600 * hold mem_hotplug_begin/end(). 3601 */ 3602 void kswapd_stop(int nid) 3603 { 3604 struct task_struct *kswapd = NODE_DATA(nid)->kswapd; 3605 3606 if (kswapd) { 3607 kthread_stop(kswapd); 3608 NODE_DATA(nid)->kswapd = NULL; 3609 } 3610 } 3611 3612 static int __init kswapd_init(void) 3613 { 3614 int nid; 3615 3616 swap_setup(); 3617 for_each_node_state(nid, N_MEMORY) 3618 kswapd_run(nid); 3619 hotcpu_notifier(cpu_callback, 0); 3620 return 0; 3621 } 3622 3623 module_init(kswapd_init) 3624 3625 #ifdef CONFIG_NUMA 3626 /* 3627 * Zone reclaim mode 3628 * 3629 * If non-zero call zone_reclaim when the number of free pages falls below 3630 * the watermarks. 3631 */ 3632 int zone_reclaim_mode __read_mostly; 3633 3634 #define RECLAIM_OFF 0 3635 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ 3636 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ 3637 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ 3638 3639 /* 3640 * Priority for ZONE_RECLAIM. This determines the fraction of pages 3641 * of a node considered for each zone_reclaim. 4 scans 1/16th of 3642 * a zone. 3643 */ 3644 #define ZONE_RECLAIM_PRIORITY 4 3645 3646 /* 3647 * Percentage of pages in a zone that must be unmapped for zone_reclaim to 3648 * occur. 3649 */ 3650 int sysctl_min_unmapped_ratio = 1; 3651 3652 /* 3653 * If the number of slab pages in a zone grows beyond this percentage then 3654 * slab reclaim needs to occur. 3655 */ 3656 int sysctl_min_slab_ratio = 5; 3657 3658 static inline unsigned long zone_unmapped_file_pages(struct zone *zone) 3659 { 3660 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); 3661 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + 3662 zone_page_state(zone, NR_ACTIVE_FILE); 3663 3664 /* 3665 * It's possible for there to be more file mapped pages than 3666 * accounted for by the pages on the file LRU lists because 3667 * tmpfs pages accounted for as ANON can also be FILE_MAPPED 3668 */ 3669 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; 3670 } 3671 3672 /* Work out how many page cache pages we can reclaim in this reclaim_mode */ 3673 static long zone_pagecache_reclaimable(struct zone *zone) 3674 { 3675 long nr_pagecache_reclaimable; 3676 long delta = 0; 3677 3678 /* 3679 * If RECLAIM_UNMAP is set, then all file pages are considered 3680 * potentially reclaimable. Otherwise, we have to worry about 3681 * pages like swapcache and zone_unmapped_file_pages() provides 3682 * a better estimate 3683 */ 3684 if (zone_reclaim_mode & RECLAIM_UNMAP) 3685 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); 3686 else 3687 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); 3688 3689 /* If we can't clean pages, remove dirty pages from consideration */ 3690 if (!(zone_reclaim_mode & RECLAIM_WRITE)) 3691 delta += zone_page_state(zone, NR_FILE_DIRTY); 3692 3693 /* Watch for any possible underflows due to delta */ 3694 if (unlikely(delta > nr_pagecache_reclaimable)) 3695 delta = nr_pagecache_reclaimable; 3696 3697 return nr_pagecache_reclaimable - delta; 3698 } 3699 3700 /* 3701 * Try to free up some pages from this zone through reclaim. 3702 */ 3703 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 3704 { 3705 /* Minimum pages needed in order to stay on node */ 3706 const unsigned long nr_pages = 1 << order; 3707 struct task_struct *p = current; 3708 struct reclaim_state reclaim_state; 3709 struct scan_control sc = { 3710 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), 3711 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), 3712 .order = order, 3713 .priority = ZONE_RECLAIM_PRIORITY, 3714 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), 3715 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP), 3716 .may_swap = 1, 3717 }; 3718 3719 cond_resched(); 3720 /* 3721 * We need to be able to allocate from the reserves for RECLAIM_UNMAP 3722 * and we also need to be able to write out pages for RECLAIM_WRITE 3723 * and RECLAIM_UNMAP. 3724 */ 3725 p->flags |= PF_MEMALLOC | PF_SWAPWRITE; 3726 lockdep_set_current_reclaim_state(gfp_mask); 3727 reclaim_state.reclaimed_slab = 0; 3728 p->reclaim_state = &reclaim_state; 3729 3730 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { 3731 /* 3732 * Free memory by calling shrink zone with increasing 3733 * priorities until we have enough memory freed. 3734 */ 3735 do { 3736 shrink_zone(zone, &sc, true); 3737 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); 3738 } 3739 3740 p->reclaim_state = NULL; 3741 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); 3742 lockdep_clear_current_reclaim_state(); 3743 return sc.nr_reclaimed >= nr_pages; 3744 } 3745 3746 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 3747 { 3748 int node_id; 3749 int ret; 3750 3751 /* 3752 * Zone reclaim reclaims unmapped file backed pages and 3753 * slab pages if we are over the defined limits. 3754 * 3755 * A small portion of unmapped file backed pages is needed for 3756 * file I/O otherwise pages read by file I/O will be immediately 3757 * thrown out if the zone is overallocated. So we do not reclaim 3758 * if less than a specified percentage of the zone is used by 3759 * unmapped file backed pages. 3760 */ 3761 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && 3762 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) 3763 return ZONE_RECLAIM_FULL; 3764 3765 if (!zone_reclaimable(zone)) 3766 return ZONE_RECLAIM_FULL; 3767 3768 /* 3769 * Do not scan if the allocation should not be delayed. 3770 */ 3771 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) 3772 return ZONE_RECLAIM_NOSCAN; 3773 3774 /* 3775 * Only run zone reclaim on the local zone or on zones that do not 3776 * have associated processors. This will favor the local processor 3777 * over remote processors and spread off node memory allocations 3778 * as wide as possible. 3779 */ 3780 node_id = zone_to_nid(zone); 3781 if (node_state(node_id, N_CPU) && node_id != numa_node_id()) 3782 return ZONE_RECLAIM_NOSCAN; 3783 3784 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags)) 3785 return ZONE_RECLAIM_NOSCAN; 3786 3787 ret = __zone_reclaim(zone, gfp_mask, order); 3788 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags); 3789 3790 if (!ret) 3791 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); 3792 3793 return ret; 3794 } 3795 #endif 3796 3797 /* 3798 * page_evictable - test whether a page is evictable 3799 * @page: the page to test 3800 * 3801 * Test whether page is evictable--i.e., should be placed on active/inactive 3802 * lists vs unevictable list. 3803 * 3804 * Reasons page might not be evictable: 3805 * (1) page's mapping marked unevictable 3806 * (2) page is part of an mlocked VMA 3807 * 3808 */ 3809 int page_evictable(struct page *page) 3810 { 3811 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); 3812 } 3813 3814 #ifdef CONFIG_SHMEM 3815 /** 3816 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list 3817 * @pages: array of pages to check 3818 * @nr_pages: number of pages to check 3819 * 3820 * Checks pages for evictability and moves them to the appropriate lru list. 3821 * 3822 * This function is only used for SysV IPC SHM_UNLOCK. 3823 */ 3824 void check_move_unevictable_pages(struct page **pages, int nr_pages) 3825 { 3826 struct lruvec *lruvec; 3827 struct zone *zone = NULL; 3828 int pgscanned = 0; 3829 int pgrescued = 0; 3830 int i; 3831 3832 for (i = 0; i < nr_pages; i++) { 3833 struct page *page = pages[i]; 3834 struct zone *pagezone; 3835 3836 pgscanned++; 3837 pagezone = page_zone(page); 3838 if (pagezone != zone) { 3839 if (zone) 3840 spin_unlock_irq(&zone->lru_lock); 3841 zone = pagezone; 3842 spin_lock_irq(&zone->lru_lock); 3843 } 3844 lruvec = mem_cgroup_page_lruvec(page, zone); 3845 3846 if (!PageLRU(page) || !PageUnevictable(page)) 3847 continue; 3848 3849 if (page_evictable(page)) { 3850 enum lru_list lru = page_lru_base_type(page); 3851 3852 VM_BUG_ON_PAGE(PageActive(page), page); 3853 ClearPageUnevictable(page); 3854 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); 3855 add_page_to_lru_list(page, lruvec, lru); 3856 pgrescued++; 3857 } 3858 } 3859 3860 if (zone) { 3861 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); 3862 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); 3863 spin_unlock_irq(&zone->lru_lock); 3864 } 3865 } 3866 #endif /* CONFIG_SHMEM */ 3867