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_FS (or __GFP_IO if it's simply going to swap, 978 * not to fs). In this case mark the page for immediate 979 * reclaim and continue scanning. 980 * 981 * Require may_enter_fs because we would wait on fs, which 982 * may not have submitted IO yet. And the loop driver might 983 * enter reclaim, and deadlock if it waits on a page for 984 * which it is needed to do the write (loop masks off 985 * __GFP_IO|__GFP_FS for this reason); but more thought 986 * would probably show more reasons. 987 * 988 * 3) Legacy memcg encounters a page that is not already marked 989 * PageReclaim. memcg does not have any dirty pages 990 * throttling so we could easily OOM just because too many 991 * pages are in writeback and there is nothing else to 992 * reclaim. Wait for the writeback to complete. 993 */ 994 if (PageWriteback(page)) { 995 /* Case 1 above */ 996 if (current_is_kswapd() && 997 PageReclaim(page) && 998 test_bit(ZONE_WRITEBACK, &zone->flags)) { 999 nr_immediate++; 1000 goto keep_locked; 1001 1002 /* Case 2 above */ 1003 } else if (sane_reclaim(sc) || 1004 !PageReclaim(page) || !may_enter_fs) { 1005 /* 1006 * This is slightly racy - end_page_writeback() 1007 * might have just cleared PageReclaim, then 1008 * setting PageReclaim here end up interpreted 1009 * as PageReadahead - but that does not matter 1010 * enough to care. What we do want is for this 1011 * page to have PageReclaim set next time memcg 1012 * reclaim reaches the tests above, so it will 1013 * then wait_on_page_writeback() to avoid OOM; 1014 * and it's also appropriate in global reclaim. 1015 */ 1016 SetPageReclaim(page); 1017 nr_writeback++; 1018 1019 goto keep_locked; 1020 1021 /* Case 3 above */ 1022 } else { 1023 wait_on_page_writeback(page); 1024 } 1025 } 1026 1027 if (!force_reclaim) 1028 references = page_check_references(page, sc); 1029 1030 switch (references) { 1031 case PAGEREF_ACTIVATE: 1032 goto activate_locked; 1033 case PAGEREF_KEEP: 1034 goto keep_locked; 1035 case PAGEREF_RECLAIM: 1036 case PAGEREF_RECLAIM_CLEAN: 1037 ; /* try to reclaim the page below */ 1038 } 1039 1040 /* 1041 * Anonymous process memory has backing store? 1042 * Try to allocate it some swap space here. 1043 */ 1044 if (PageAnon(page) && !PageSwapCache(page)) { 1045 if (!(sc->gfp_mask & __GFP_IO)) 1046 goto keep_locked; 1047 if (!add_to_swap(page, page_list)) 1048 goto activate_locked; 1049 may_enter_fs = 1; 1050 1051 /* Adding to swap updated mapping */ 1052 mapping = page_mapping(page); 1053 } 1054 1055 /* 1056 * The page is mapped into the page tables of one or more 1057 * processes. Try to unmap it here. 1058 */ 1059 if (page_mapped(page) && mapping) { 1060 switch (try_to_unmap(page, ttu_flags)) { 1061 case SWAP_FAIL: 1062 goto activate_locked; 1063 case SWAP_AGAIN: 1064 goto keep_locked; 1065 case SWAP_MLOCK: 1066 goto cull_mlocked; 1067 case SWAP_SUCCESS: 1068 ; /* try to free the page below */ 1069 } 1070 } 1071 1072 if (PageDirty(page)) { 1073 /* 1074 * Only kswapd can writeback filesystem pages to 1075 * avoid risk of stack overflow but only writeback 1076 * if many dirty pages have been encountered. 1077 */ 1078 if (page_is_file_cache(page) && 1079 (!current_is_kswapd() || 1080 !test_bit(ZONE_DIRTY, &zone->flags))) { 1081 /* 1082 * Immediately reclaim when written back. 1083 * Similar in principal to deactivate_page() 1084 * except we already have the page isolated 1085 * and know it's dirty 1086 */ 1087 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE); 1088 SetPageReclaim(page); 1089 1090 goto keep_locked; 1091 } 1092 1093 if (references == PAGEREF_RECLAIM_CLEAN) 1094 goto keep_locked; 1095 if (!may_enter_fs) 1096 goto keep_locked; 1097 if (!sc->may_writepage) 1098 goto keep_locked; 1099 1100 /* Page is dirty, try to write it out here */ 1101 switch (pageout(page, mapping, sc)) { 1102 case PAGE_KEEP: 1103 goto keep_locked; 1104 case PAGE_ACTIVATE: 1105 goto activate_locked; 1106 case PAGE_SUCCESS: 1107 if (PageWriteback(page)) 1108 goto keep; 1109 if (PageDirty(page)) 1110 goto keep; 1111 1112 /* 1113 * A synchronous write - probably a ramdisk. Go 1114 * ahead and try to reclaim the page. 1115 */ 1116 if (!trylock_page(page)) 1117 goto keep; 1118 if (PageDirty(page) || PageWriteback(page)) 1119 goto keep_locked; 1120 mapping = page_mapping(page); 1121 case PAGE_CLEAN: 1122 ; /* try to free the page below */ 1123 } 1124 } 1125 1126 /* 1127 * If the page has buffers, try to free the buffer mappings 1128 * associated with this page. If we succeed we try to free 1129 * the page as well. 1130 * 1131 * We do this even if the page is PageDirty(). 1132 * try_to_release_page() does not perform I/O, but it is 1133 * possible for a page to have PageDirty set, but it is actually 1134 * clean (all its buffers are clean). This happens if the 1135 * buffers were written out directly, with submit_bh(). ext3 1136 * will do this, as well as the blockdev mapping. 1137 * try_to_release_page() will discover that cleanness and will 1138 * drop the buffers and mark the page clean - it can be freed. 1139 * 1140 * Rarely, pages can have buffers and no ->mapping. These are 1141 * the pages which were not successfully invalidated in 1142 * truncate_complete_page(). We try to drop those buffers here 1143 * and if that worked, and the page is no longer mapped into 1144 * process address space (page_count == 1) it can be freed. 1145 * Otherwise, leave the page on the LRU so it is swappable. 1146 */ 1147 if (page_has_private(page)) { 1148 if (!try_to_release_page(page, sc->gfp_mask)) 1149 goto activate_locked; 1150 if (!mapping && page_count(page) == 1) { 1151 unlock_page(page); 1152 if (put_page_testzero(page)) 1153 goto free_it; 1154 else { 1155 /* 1156 * rare race with speculative reference. 1157 * the speculative reference will free 1158 * this page shortly, so we may 1159 * increment nr_reclaimed here (and 1160 * leave it off the LRU). 1161 */ 1162 nr_reclaimed++; 1163 continue; 1164 } 1165 } 1166 } 1167 1168 if (!mapping || !__remove_mapping(mapping, page, true)) 1169 goto keep_locked; 1170 1171 /* 1172 * At this point, we have no other references and there is 1173 * no way to pick any more up (removed from LRU, removed 1174 * from pagecache). Can use non-atomic bitops now (and 1175 * we obviously don't have to worry about waking up a process 1176 * waiting on the page lock, because there are no references. 1177 */ 1178 __clear_page_locked(page); 1179 free_it: 1180 nr_reclaimed++; 1181 1182 /* 1183 * Is there need to periodically free_page_list? It would 1184 * appear not as the counts should be low 1185 */ 1186 list_add(&page->lru, &free_pages); 1187 continue; 1188 1189 cull_mlocked: 1190 if (PageSwapCache(page)) 1191 try_to_free_swap(page); 1192 unlock_page(page); 1193 putback_lru_page(page); 1194 continue; 1195 1196 activate_locked: 1197 /* Not a candidate for swapping, so reclaim swap space. */ 1198 if (PageSwapCache(page) && vm_swap_full()) 1199 try_to_free_swap(page); 1200 VM_BUG_ON_PAGE(PageActive(page), page); 1201 SetPageActive(page); 1202 pgactivate++; 1203 keep_locked: 1204 unlock_page(page); 1205 keep: 1206 list_add(&page->lru, &ret_pages); 1207 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); 1208 } 1209 1210 mem_cgroup_uncharge_list(&free_pages); 1211 free_hot_cold_page_list(&free_pages, true); 1212 1213 list_splice(&ret_pages, page_list); 1214 count_vm_events(PGACTIVATE, pgactivate); 1215 1216 *ret_nr_dirty += nr_dirty; 1217 *ret_nr_congested += nr_congested; 1218 *ret_nr_unqueued_dirty += nr_unqueued_dirty; 1219 *ret_nr_writeback += nr_writeback; 1220 *ret_nr_immediate += nr_immediate; 1221 return nr_reclaimed; 1222 } 1223 1224 unsigned long reclaim_clean_pages_from_list(struct zone *zone, 1225 struct list_head *page_list) 1226 { 1227 struct scan_control sc = { 1228 .gfp_mask = GFP_KERNEL, 1229 .priority = DEF_PRIORITY, 1230 .may_unmap = 1, 1231 }; 1232 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5; 1233 struct page *page, *next; 1234 LIST_HEAD(clean_pages); 1235 1236 list_for_each_entry_safe(page, next, page_list, lru) { 1237 if (page_is_file_cache(page) && !PageDirty(page) && 1238 !isolated_balloon_page(page)) { 1239 ClearPageActive(page); 1240 list_move(&page->lru, &clean_pages); 1241 } 1242 } 1243 1244 ret = shrink_page_list(&clean_pages, zone, &sc, 1245 TTU_UNMAP|TTU_IGNORE_ACCESS, 1246 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true); 1247 list_splice(&clean_pages, page_list); 1248 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret); 1249 return ret; 1250 } 1251 1252 /* 1253 * Attempt to remove the specified page from its LRU. Only take this page 1254 * if it is of the appropriate PageActive status. Pages which are being 1255 * freed elsewhere are also ignored. 1256 * 1257 * page: page to consider 1258 * mode: one of the LRU isolation modes defined above 1259 * 1260 * returns 0 on success, -ve errno on failure. 1261 */ 1262 int __isolate_lru_page(struct page *page, isolate_mode_t mode) 1263 { 1264 int ret = -EINVAL; 1265 1266 /* Only take pages on the LRU. */ 1267 if (!PageLRU(page)) 1268 return ret; 1269 1270 /* Compaction should not handle unevictable pages but CMA can do so */ 1271 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) 1272 return ret; 1273 1274 ret = -EBUSY; 1275 1276 /* 1277 * To minimise LRU disruption, the caller can indicate that it only 1278 * wants to isolate pages it will be able to operate on without 1279 * blocking - clean pages for the most part. 1280 * 1281 * ISOLATE_CLEAN means that only clean pages should be isolated. This 1282 * is used by reclaim when it is cannot write to backing storage 1283 * 1284 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages 1285 * that it is possible to migrate without blocking 1286 */ 1287 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) { 1288 /* All the caller can do on PageWriteback is block */ 1289 if (PageWriteback(page)) 1290 return ret; 1291 1292 if (PageDirty(page)) { 1293 struct address_space *mapping; 1294 1295 /* ISOLATE_CLEAN means only clean pages */ 1296 if (mode & ISOLATE_CLEAN) 1297 return ret; 1298 1299 /* 1300 * Only pages without mappings or that have a 1301 * ->migratepage callback are possible to migrate 1302 * without blocking 1303 */ 1304 mapping = page_mapping(page); 1305 if (mapping && !mapping->a_ops->migratepage) 1306 return ret; 1307 } 1308 } 1309 1310 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) 1311 return ret; 1312 1313 if (likely(get_page_unless_zero(page))) { 1314 /* 1315 * Be careful not to clear PageLRU until after we're 1316 * sure the page is not being freed elsewhere -- the 1317 * page release code relies on it. 1318 */ 1319 ClearPageLRU(page); 1320 ret = 0; 1321 } 1322 1323 return ret; 1324 } 1325 1326 /* 1327 * zone->lru_lock is heavily contended. Some of the functions that 1328 * shrink the lists perform better by taking out a batch of pages 1329 * and working on them outside the LRU lock. 1330 * 1331 * For pagecache intensive workloads, this function is the hottest 1332 * spot in the kernel (apart from copy_*_user functions). 1333 * 1334 * Appropriate locks must be held before calling this function. 1335 * 1336 * @nr_to_scan: The number of pages to look through on the list. 1337 * @lruvec: The LRU vector to pull pages from. 1338 * @dst: The temp list to put pages on to. 1339 * @nr_scanned: The number of pages that were scanned. 1340 * @sc: The scan_control struct for this reclaim session 1341 * @mode: One of the LRU isolation modes 1342 * @lru: LRU list id for isolating 1343 * 1344 * returns how many pages were moved onto *@dst. 1345 */ 1346 static unsigned long isolate_lru_pages(unsigned long nr_to_scan, 1347 struct lruvec *lruvec, struct list_head *dst, 1348 unsigned long *nr_scanned, struct scan_control *sc, 1349 isolate_mode_t mode, enum lru_list lru) 1350 { 1351 struct list_head *src = &lruvec->lists[lru]; 1352 unsigned long nr_taken = 0; 1353 unsigned long scan; 1354 1355 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { 1356 struct page *page; 1357 int nr_pages; 1358 1359 page = lru_to_page(src); 1360 prefetchw_prev_lru_page(page, src, flags); 1361 1362 VM_BUG_ON_PAGE(!PageLRU(page), page); 1363 1364 switch (__isolate_lru_page(page, mode)) { 1365 case 0: 1366 nr_pages = hpage_nr_pages(page); 1367 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages); 1368 list_move(&page->lru, dst); 1369 nr_taken += nr_pages; 1370 break; 1371 1372 case -EBUSY: 1373 /* else it is being freed elsewhere */ 1374 list_move(&page->lru, src); 1375 continue; 1376 1377 default: 1378 BUG(); 1379 } 1380 } 1381 1382 *nr_scanned = scan; 1383 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan, 1384 nr_taken, mode, is_file_lru(lru)); 1385 return nr_taken; 1386 } 1387 1388 /** 1389 * isolate_lru_page - tries to isolate a page from its LRU list 1390 * @page: page to isolate from its LRU list 1391 * 1392 * Isolates a @page from an LRU list, clears PageLRU and adjusts the 1393 * vmstat statistic corresponding to whatever LRU list the page was on. 1394 * 1395 * Returns 0 if the page was removed from an LRU list. 1396 * Returns -EBUSY if the page was not on an LRU list. 1397 * 1398 * The returned page will have PageLRU() cleared. If it was found on 1399 * the active list, it will have PageActive set. If it was found on 1400 * the unevictable list, it will have the PageUnevictable bit set. That flag 1401 * may need to be cleared by the caller before letting the page go. 1402 * 1403 * The vmstat statistic corresponding to the list on which the page was 1404 * found will be decremented. 1405 * 1406 * Restrictions: 1407 * (1) Must be called with an elevated refcount on the page. This is a 1408 * fundamentnal difference from isolate_lru_pages (which is called 1409 * without a stable reference). 1410 * (2) the lru_lock must not be held. 1411 * (3) interrupts must be enabled. 1412 */ 1413 int isolate_lru_page(struct page *page) 1414 { 1415 int ret = -EBUSY; 1416 1417 VM_BUG_ON_PAGE(!page_count(page), page); 1418 1419 if (PageLRU(page)) { 1420 struct zone *zone = page_zone(page); 1421 struct lruvec *lruvec; 1422 1423 spin_lock_irq(&zone->lru_lock); 1424 lruvec = mem_cgroup_page_lruvec(page, zone); 1425 if (PageLRU(page)) { 1426 int lru = page_lru(page); 1427 get_page(page); 1428 ClearPageLRU(page); 1429 del_page_from_lru_list(page, lruvec, lru); 1430 ret = 0; 1431 } 1432 spin_unlock_irq(&zone->lru_lock); 1433 } 1434 return ret; 1435 } 1436 1437 /* 1438 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and 1439 * then get resheduled. When there are massive number of tasks doing page 1440 * allocation, such sleeping direct reclaimers may keep piling up on each CPU, 1441 * the LRU list will go small and be scanned faster than necessary, leading to 1442 * unnecessary swapping, thrashing and OOM. 1443 */ 1444 static int too_many_isolated(struct zone *zone, int file, 1445 struct scan_control *sc) 1446 { 1447 unsigned long inactive, isolated; 1448 1449 if (current_is_kswapd()) 1450 return 0; 1451 1452 if (!sane_reclaim(sc)) 1453 return 0; 1454 1455 if (file) { 1456 inactive = zone_page_state(zone, NR_INACTIVE_FILE); 1457 isolated = zone_page_state(zone, NR_ISOLATED_FILE); 1458 } else { 1459 inactive = zone_page_state(zone, NR_INACTIVE_ANON); 1460 isolated = zone_page_state(zone, NR_ISOLATED_ANON); 1461 } 1462 1463 /* 1464 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they 1465 * won't get blocked by normal direct-reclaimers, forming a circular 1466 * deadlock. 1467 */ 1468 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS) 1469 inactive >>= 3; 1470 1471 return isolated > inactive; 1472 } 1473 1474 static noinline_for_stack void 1475 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) 1476 { 1477 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1478 struct zone *zone = lruvec_zone(lruvec); 1479 LIST_HEAD(pages_to_free); 1480 1481 /* 1482 * Put back any unfreeable pages. 1483 */ 1484 while (!list_empty(page_list)) { 1485 struct page *page = lru_to_page(page_list); 1486 int lru; 1487 1488 VM_BUG_ON_PAGE(PageLRU(page), page); 1489 list_del(&page->lru); 1490 if (unlikely(!page_evictable(page))) { 1491 spin_unlock_irq(&zone->lru_lock); 1492 putback_lru_page(page); 1493 spin_lock_irq(&zone->lru_lock); 1494 continue; 1495 } 1496 1497 lruvec = mem_cgroup_page_lruvec(page, zone); 1498 1499 SetPageLRU(page); 1500 lru = page_lru(page); 1501 add_page_to_lru_list(page, lruvec, lru); 1502 1503 if (is_active_lru(lru)) { 1504 int file = is_file_lru(lru); 1505 int numpages = hpage_nr_pages(page); 1506 reclaim_stat->recent_rotated[file] += numpages; 1507 } 1508 if (put_page_testzero(page)) { 1509 __ClearPageLRU(page); 1510 __ClearPageActive(page); 1511 del_page_from_lru_list(page, lruvec, lru); 1512 1513 if (unlikely(PageCompound(page))) { 1514 spin_unlock_irq(&zone->lru_lock); 1515 mem_cgroup_uncharge(page); 1516 (*get_compound_page_dtor(page))(page); 1517 spin_lock_irq(&zone->lru_lock); 1518 } else 1519 list_add(&page->lru, &pages_to_free); 1520 } 1521 } 1522 1523 /* 1524 * To save our caller's stack, now use input list for pages to free. 1525 */ 1526 list_splice(&pages_to_free, page_list); 1527 } 1528 1529 /* 1530 * If a kernel thread (such as nfsd for loop-back mounts) services 1531 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. 1532 * In that case we should only throttle if the backing device it is 1533 * writing to is congested. In other cases it is safe to throttle. 1534 */ 1535 static int current_may_throttle(void) 1536 { 1537 return !(current->flags & PF_LESS_THROTTLE) || 1538 current->backing_dev_info == NULL || 1539 bdi_write_congested(current->backing_dev_info); 1540 } 1541 1542 /* 1543 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number 1544 * of reclaimed pages 1545 */ 1546 static noinline_for_stack unsigned long 1547 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, 1548 struct scan_control *sc, enum lru_list lru) 1549 { 1550 LIST_HEAD(page_list); 1551 unsigned long nr_scanned; 1552 unsigned long nr_reclaimed = 0; 1553 unsigned long nr_taken; 1554 unsigned long nr_dirty = 0; 1555 unsigned long nr_congested = 0; 1556 unsigned long nr_unqueued_dirty = 0; 1557 unsigned long nr_writeback = 0; 1558 unsigned long nr_immediate = 0; 1559 isolate_mode_t isolate_mode = 0; 1560 int file = is_file_lru(lru); 1561 struct zone *zone = lruvec_zone(lruvec); 1562 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1563 1564 while (unlikely(too_many_isolated(zone, file, sc))) { 1565 congestion_wait(BLK_RW_ASYNC, HZ/10); 1566 1567 /* We are about to die and free our memory. Return now. */ 1568 if (fatal_signal_pending(current)) 1569 return SWAP_CLUSTER_MAX; 1570 } 1571 1572 lru_add_drain(); 1573 1574 if (!sc->may_unmap) 1575 isolate_mode |= ISOLATE_UNMAPPED; 1576 if (!sc->may_writepage) 1577 isolate_mode |= ISOLATE_CLEAN; 1578 1579 spin_lock_irq(&zone->lru_lock); 1580 1581 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, 1582 &nr_scanned, sc, isolate_mode, lru); 1583 1584 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); 1585 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); 1586 1587 if (global_reclaim(sc)) { 1588 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned); 1589 if (current_is_kswapd()) 1590 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned); 1591 else 1592 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned); 1593 } 1594 spin_unlock_irq(&zone->lru_lock); 1595 1596 if (nr_taken == 0) 1597 return 0; 1598 1599 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP, 1600 &nr_dirty, &nr_unqueued_dirty, &nr_congested, 1601 &nr_writeback, &nr_immediate, 1602 false); 1603 1604 spin_lock_irq(&zone->lru_lock); 1605 1606 reclaim_stat->recent_scanned[file] += nr_taken; 1607 1608 if (global_reclaim(sc)) { 1609 if (current_is_kswapd()) 1610 __count_zone_vm_events(PGSTEAL_KSWAPD, zone, 1611 nr_reclaimed); 1612 else 1613 __count_zone_vm_events(PGSTEAL_DIRECT, zone, 1614 nr_reclaimed); 1615 } 1616 1617 putback_inactive_pages(lruvec, &page_list); 1618 1619 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); 1620 1621 spin_unlock_irq(&zone->lru_lock); 1622 1623 mem_cgroup_uncharge_list(&page_list); 1624 free_hot_cold_page_list(&page_list, true); 1625 1626 /* 1627 * If reclaim is isolating dirty pages under writeback, it implies 1628 * that the long-lived page allocation rate is exceeding the page 1629 * laundering rate. Either the global limits are not being effective 1630 * at throttling processes due to the page distribution throughout 1631 * zones or there is heavy usage of a slow backing device. The 1632 * only option is to throttle from reclaim context which is not ideal 1633 * as there is no guarantee the dirtying process is throttled in the 1634 * same way balance_dirty_pages() manages. 1635 * 1636 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number 1637 * of pages under pages flagged for immediate reclaim and stall if any 1638 * are encountered in the nr_immediate check below. 1639 */ 1640 if (nr_writeback && nr_writeback == nr_taken) 1641 set_bit(ZONE_WRITEBACK, &zone->flags); 1642 1643 /* 1644 * Legacy memcg will stall in page writeback so avoid forcibly 1645 * stalling here. 1646 */ 1647 if (sane_reclaim(sc)) { 1648 /* 1649 * Tag a zone as congested if all the dirty pages scanned were 1650 * backed by a congested BDI and wait_iff_congested will stall. 1651 */ 1652 if (nr_dirty && nr_dirty == nr_congested) 1653 set_bit(ZONE_CONGESTED, &zone->flags); 1654 1655 /* 1656 * If dirty pages are scanned that are not queued for IO, it 1657 * implies that flushers are not keeping up. In this case, flag 1658 * the zone ZONE_DIRTY and kswapd will start writing pages from 1659 * reclaim context. 1660 */ 1661 if (nr_unqueued_dirty == nr_taken) 1662 set_bit(ZONE_DIRTY, &zone->flags); 1663 1664 /* 1665 * If kswapd scans pages marked marked for immediate 1666 * reclaim and under writeback (nr_immediate), it implies 1667 * that pages are cycling through the LRU faster than 1668 * they are written so also forcibly stall. 1669 */ 1670 if (nr_immediate && current_may_throttle()) 1671 congestion_wait(BLK_RW_ASYNC, HZ/10); 1672 } 1673 1674 /* 1675 * Stall direct reclaim for IO completions if underlying BDIs or zone 1676 * is congested. Allow kswapd to continue until it starts encountering 1677 * unqueued dirty pages or cycling through the LRU too quickly. 1678 */ 1679 if (!sc->hibernation_mode && !current_is_kswapd() && 1680 current_may_throttle()) 1681 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10); 1682 1683 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id, 1684 zone_idx(zone), 1685 nr_scanned, nr_reclaimed, 1686 sc->priority, 1687 trace_shrink_flags(file)); 1688 return nr_reclaimed; 1689 } 1690 1691 /* 1692 * This moves pages from the active list to the inactive list. 1693 * 1694 * We move them the other way if the page is referenced by one or more 1695 * processes, from rmap. 1696 * 1697 * If the pages are mostly unmapped, the processing is fast and it is 1698 * appropriate to hold zone->lru_lock across the whole operation. But if 1699 * the pages are mapped, the processing is slow (page_referenced()) so we 1700 * should drop zone->lru_lock around each page. It's impossible to balance 1701 * this, so instead we remove the pages from the LRU while processing them. 1702 * It is safe to rely on PG_active against the non-LRU pages in here because 1703 * nobody will play with that bit on a non-LRU page. 1704 * 1705 * The downside is that we have to touch page->_count against each page. 1706 * But we had to alter page->flags anyway. 1707 */ 1708 1709 static void move_active_pages_to_lru(struct lruvec *lruvec, 1710 struct list_head *list, 1711 struct list_head *pages_to_free, 1712 enum lru_list lru) 1713 { 1714 struct zone *zone = lruvec_zone(lruvec); 1715 unsigned long pgmoved = 0; 1716 struct page *page; 1717 int nr_pages; 1718 1719 while (!list_empty(list)) { 1720 page = lru_to_page(list); 1721 lruvec = mem_cgroup_page_lruvec(page, zone); 1722 1723 VM_BUG_ON_PAGE(PageLRU(page), page); 1724 SetPageLRU(page); 1725 1726 nr_pages = hpage_nr_pages(page); 1727 mem_cgroup_update_lru_size(lruvec, lru, nr_pages); 1728 list_move(&page->lru, &lruvec->lists[lru]); 1729 pgmoved += nr_pages; 1730 1731 if (put_page_testzero(page)) { 1732 __ClearPageLRU(page); 1733 __ClearPageActive(page); 1734 del_page_from_lru_list(page, lruvec, lru); 1735 1736 if (unlikely(PageCompound(page))) { 1737 spin_unlock_irq(&zone->lru_lock); 1738 mem_cgroup_uncharge(page); 1739 (*get_compound_page_dtor(page))(page); 1740 spin_lock_irq(&zone->lru_lock); 1741 } else 1742 list_add(&page->lru, pages_to_free); 1743 } 1744 } 1745 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); 1746 if (!is_active_lru(lru)) 1747 __count_vm_events(PGDEACTIVATE, pgmoved); 1748 } 1749 1750 static void shrink_active_list(unsigned long nr_to_scan, 1751 struct lruvec *lruvec, 1752 struct scan_control *sc, 1753 enum lru_list lru) 1754 { 1755 unsigned long nr_taken; 1756 unsigned long nr_scanned; 1757 unsigned long vm_flags; 1758 LIST_HEAD(l_hold); /* The pages which were snipped off */ 1759 LIST_HEAD(l_active); 1760 LIST_HEAD(l_inactive); 1761 struct page *page; 1762 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1763 unsigned long nr_rotated = 0; 1764 isolate_mode_t isolate_mode = 0; 1765 int file = is_file_lru(lru); 1766 struct zone *zone = lruvec_zone(lruvec); 1767 1768 lru_add_drain(); 1769 1770 if (!sc->may_unmap) 1771 isolate_mode |= ISOLATE_UNMAPPED; 1772 if (!sc->may_writepage) 1773 isolate_mode |= ISOLATE_CLEAN; 1774 1775 spin_lock_irq(&zone->lru_lock); 1776 1777 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, 1778 &nr_scanned, sc, isolate_mode, lru); 1779 if (global_reclaim(sc)) 1780 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned); 1781 1782 reclaim_stat->recent_scanned[file] += nr_taken; 1783 1784 __count_zone_vm_events(PGREFILL, zone, nr_scanned); 1785 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken); 1786 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken); 1787 spin_unlock_irq(&zone->lru_lock); 1788 1789 while (!list_empty(&l_hold)) { 1790 cond_resched(); 1791 page = lru_to_page(&l_hold); 1792 list_del(&page->lru); 1793 1794 if (unlikely(!page_evictable(page))) { 1795 putback_lru_page(page); 1796 continue; 1797 } 1798 1799 if (unlikely(buffer_heads_over_limit)) { 1800 if (page_has_private(page) && trylock_page(page)) { 1801 if (page_has_private(page)) 1802 try_to_release_page(page, 0); 1803 unlock_page(page); 1804 } 1805 } 1806 1807 if (page_referenced(page, 0, sc->target_mem_cgroup, 1808 &vm_flags)) { 1809 nr_rotated += hpage_nr_pages(page); 1810 /* 1811 * Identify referenced, file-backed active pages and 1812 * give them one more trip around the active list. So 1813 * that executable code get better chances to stay in 1814 * memory under moderate memory pressure. Anon pages 1815 * are not likely to be evicted by use-once streaming 1816 * IO, plus JVM can create lots of anon VM_EXEC pages, 1817 * so we ignore them here. 1818 */ 1819 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { 1820 list_add(&page->lru, &l_active); 1821 continue; 1822 } 1823 } 1824 1825 ClearPageActive(page); /* we are de-activating */ 1826 list_add(&page->lru, &l_inactive); 1827 } 1828 1829 /* 1830 * Move pages back to the lru list. 1831 */ 1832 spin_lock_irq(&zone->lru_lock); 1833 /* 1834 * Count referenced pages from currently used mappings as rotated, 1835 * even though only some of them are actually re-activated. This 1836 * helps balance scan pressure between file and anonymous pages in 1837 * get_scan_count. 1838 */ 1839 reclaim_stat->recent_rotated[file] += nr_rotated; 1840 1841 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); 1842 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); 1843 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken); 1844 spin_unlock_irq(&zone->lru_lock); 1845 1846 mem_cgroup_uncharge_list(&l_hold); 1847 free_hot_cold_page_list(&l_hold, true); 1848 } 1849 1850 #ifdef CONFIG_SWAP 1851 static int inactive_anon_is_low_global(struct zone *zone) 1852 { 1853 unsigned long active, inactive; 1854 1855 active = zone_page_state(zone, NR_ACTIVE_ANON); 1856 inactive = zone_page_state(zone, NR_INACTIVE_ANON); 1857 1858 if (inactive * zone->inactive_ratio < active) 1859 return 1; 1860 1861 return 0; 1862 } 1863 1864 /** 1865 * inactive_anon_is_low - check if anonymous pages need to be deactivated 1866 * @lruvec: LRU vector to check 1867 * 1868 * Returns true if the zone does not have enough inactive anon pages, 1869 * meaning some active anon pages need to be deactivated. 1870 */ 1871 static int inactive_anon_is_low(struct lruvec *lruvec) 1872 { 1873 /* 1874 * If we don't have swap space, anonymous page deactivation 1875 * is pointless. 1876 */ 1877 if (!total_swap_pages) 1878 return 0; 1879 1880 if (!mem_cgroup_disabled()) 1881 return mem_cgroup_inactive_anon_is_low(lruvec); 1882 1883 return inactive_anon_is_low_global(lruvec_zone(lruvec)); 1884 } 1885 #else 1886 static inline int inactive_anon_is_low(struct lruvec *lruvec) 1887 { 1888 return 0; 1889 } 1890 #endif 1891 1892 /** 1893 * inactive_file_is_low - check if file pages need to be deactivated 1894 * @lruvec: LRU vector to check 1895 * 1896 * When the system is doing streaming IO, memory pressure here 1897 * ensures that active file pages get deactivated, until more 1898 * than half of the file pages are on the inactive list. 1899 * 1900 * Once we get to that situation, protect the system's working 1901 * set from being evicted by disabling active file page aging. 1902 * 1903 * This uses a different ratio than the anonymous pages, because 1904 * the page cache uses a use-once replacement algorithm. 1905 */ 1906 static int inactive_file_is_low(struct lruvec *lruvec) 1907 { 1908 unsigned long inactive; 1909 unsigned long active; 1910 1911 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE); 1912 active = get_lru_size(lruvec, LRU_ACTIVE_FILE); 1913 1914 return active > inactive; 1915 } 1916 1917 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru) 1918 { 1919 if (is_file_lru(lru)) 1920 return inactive_file_is_low(lruvec); 1921 else 1922 return inactive_anon_is_low(lruvec); 1923 } 1924 1925 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, 1926 struct lruvec *lruvec, struct scan_control *sc) 1927 { 1928 if (is_active_lru(lru)) { 1929 if (inactive_list_is_low(lruvec, lru)) 1930 shrink_active_list(nr_to_scan, lruvec, sc, lru); 1931 return 0; 1932 } 1933 1934 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); 1935 } 1936 1937 enum scan_balance { 1938 SCAN_EQUAL, 1939 SCAN_FRACT, 1940 SCAN_ANON, 1941 SCAN_FILE, 1942 }; 1943 1944 /* 1945 * Determine how aggressively the anon and file LRU lists should be 1946 * scanned. The relative value of each set of LRU lists is determined 1947 * by looking at the fraction of the pages scanned we did rotate back 1948 * onto the active list instead of evict. 1949 * 1950 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan 1951 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan 1952 */ 1953 static void get_scan_count(struct lruvec *lruvec, int swappiness, 1954 struct scan_control *sc, unsigned long *nr, 1955 unsigned long *lru_pages) 1956 { 1957 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1958 u64 fraction[2]; 1959 u64 denominator = 0; /* gcc */ 1960 struct zone *zone = lruvec_zone(lruvec); 1961 unsigned long anon_prio, file_prio; 1962 enum scan_balance scan_balance; 1963 unsigned long anon, file; 1964 bool force_scan = false; 1965 unsigned long ap, fp; 1966 enum lru_list lru; 1967 bool some_scanned; 1968 int pass; 1969 1970 /* 1971 * If the zone or memcg is small, nr[l] can be 0. This 1972 * results in no scanning on this priority and a potential 1973 * priority drop. Global direct reclaim can go to the next 1974 * zone and tends to have no problems. Global kswapd is for 1975 * zone balancing and it needs to scan a minimum amount. When 1976 * reclaiming for a memcg, a priority drop can cause high 1977 * latencies, so it's better to scan a minimum amount there as 1978 * well. 1979 */ 1980 if (current_is_kswapd()) { 1981 if (!zone_reclaimable(zone)) 1982 force_scan = true; 1983 if (!mem_cgroup_lruvec_online(lruvec)) 1984 force_scan = true; 1985 } 1986 if (!global_reclaim(sc)) 1987 force_scan = true; 1988 1989 /* If we have no swap space, do not bother scanning anon pages. */ 1990 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) { 1991 scan_balance = SCAN_FILE; 1992 goto out; 1993 } 1994 1995 /* 1996 * Global reclaim will swap to prevent OOM even with no 1997 * swappiness, but memcg users want to use this knob to 1998 * disable swapping for individual groups completely when 1999 * using the memory controller's swap limit feature would be 2000 * too expensive. 2001 */ 2002 if (!global_reclaim(sc) && !swappiness) { 2003 scan_balance = SCAN_FILE; 2004 goto out; 2005 } 2006 2007 /* 2008 * Do not apply any pressure balancing cleverness when the 2009 * system is close to OOM, scan both anon and file equally 2010 * (unless the swappiness setting disagrees with swapping). 2011 */ 2012 if (!sc->priority && swappiness) { 2013 scan_balance = SCAN_EQUAL; 2014 goto out; 2015 } 2016 2017 /* 2018 * Prevent the reclaimer from falling into the cache trap: as 2019 * cache pages start out inactive, every cache fault will tip 2020 * the scan balance towards the file LRU. And as the file LRU 2021 * shrinks, so does the window for rotation from references. 2022 * This means we have a runaway feedback loop where a tiny 2023 * thrashing file LRU becomes infinitely more attractive than 2024 * anon pages. Try to detect this based on file LRU size. 2025 */ 2026 if (global_reclaim(sc)) { 2027 unsigned long zonefile; 2028 unsigned long zonefree; 2029 2030 zonefree = zone_page_state(zone, NR_FREE_PAGES); 2031 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) + 2032 zone_page_state(zone, NR_INACTIVE_FILE); 2033 2034 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) { 2035 scan_balance = SCAN_ANON; 2036 goto out; 2037 } 2038 } 2039 2040 /* 2041 * There is enough inactive page cache, do not reclaim 2042 * anything from the anonymous working set right now. 2043 */ 2044 if (!inactive_file_is_low(lruvec)) { 2045 scan_balance = SCAN_FILE; 2046 goto out; 2047 } 2048 2049 scan_balance = SCAN_FRACT; 2050 2051 /* 2052 * With swappiness at 100, anonymous and file have the same priority. 2053 * This scanning priority is essentially the inverse of IO cost. 2054 */ 2055 anon_prio = swappiness; 2056 file_prio = 200 - anon_prio; 2057 2058 /* 2059 * OK, so we have swap space and a fair amount of page cache 2060 * pages. We use the recently rotated / recently scanned 2061 * ratios to determine how valuable each cache is. 2062 * 2063 * Because workloads change over time (and to avoid overflow) 2064 * we keep these statistics as a floating average, which ends 2065 * up weighing recent references more than old ones. 2066 * 2067 * anon in [0], file in [1] 2068 */ 2069 2070 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) + 2071 get_lru_size(lruvec, LRU_INACTIVE_ANON); 2072 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) + 2073 get_lru_size(lruvec, LRU_INACTIVE_FILE); 2074 2075 spin_lock_irq(&zone->lru_lock); 2076 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { 2077 reclaim_stat->recent_scanned[0] /= 2; 2078 reclaim_stat->recent_rotated[0] /= 2; 2079 } 2080 2081 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { 2082 reclaim_stat->recent_scanned[1] /= 2; 2083 reclaim_stat->recent_rotated[1] /= 2; 2084 } 2085 2086 /* 2087 * The amount of pressure on anon vs file pages is inversely 2088 * proportional to the fraction of recently scanned pages on 2089 * each list that were recently referenced and in active use. 2090 */ 2091 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); 2092 ap /= reclaim_stat->recent_rotated[0] + 1; 2093 2094 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); 2095 fp /= reclaim_stat->recent_rotated[1] + 1; 2096 spin_unlock_irq(&zone->lru_lock); 2097 2098 fraction[0] = ap; 2099 fraction[1] = fp; 2100 denominator = ap + fp + 1; 2101 out: 2102 some_scanned = false; 2103 /* Only use force_scan on second pass. */ 2104 for (pass = 0; !some_scanned && pass < 2; pass++) { 2105 *lru_pages = 0; 2106 for_each_evictable_lru(lru) { 2107 int file = is_file_lru(lru); 2108 unsigned long size; 2109 unsigned long scan; 2110 2111 size = get_lru_size(lruvec, lru); 2112 scan = size >> sc->priority; 2113 2114 if (!scan && pass && force_scan) 2115 scan = min(size, SWAP_CLUSTER_MAX); 2116 2117 switch (scan_balance) { 2118 case SCAN_EQUAL: 2119 /* Scan lists relative to size */ 2120 break; 2121 case SCAN_FRACT: 2122 /* 2123 * Scan types proportional to swappiness and 2124 * their relative recent reclaim efficiency. 2125 */ 2126 scan = div64_u64(scan * fraction[file], 2127 denominator); 2128 break; 2129 case SCAN_FILE: 2130 case SCAN_ANON: 2131 /* Scan one type exclusively */ 2132 if ((scan_balance == SCAN_FILE) != file) { 2133 size = 0; 2134 scan = 0; 2135 } 2136 break; 2137 default: 2138 /* Look ma, no brain */ 2139 BUG(); 2140 } 2141 2142 *lru_pages += size; 2143 nr[lru] = scan; 2144 2145 /* 2146 * Skip the second pass and don't force_scan, 2147 * if we found something to scan. 2148 */ 2149 some_scanned |= !!scan; 2150 } 2151 } 2152 } 2153 2154 /* 2155 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. 2156 */ 2157 static void shrink_lruvec(struct lruvec *lruvec, int swappiness, 2158 struct scan_control *sc, unsigned long *lru_pages) 2159 { 2160 unsigned long nr[NR_LRU_LISTS]; 2161 unsigned long targets[NR_LRU_LISTS]; 2162 unsigned long nr_to_scan; 2163 enum lru_list lru; 2164 unsigned long nr_reclaimed = 0; 2165 unsigned long nr_to_reclaim = sc->nr_to_reclaim; 2166 struct blk_plug plug; 2167 bool scan_adjusted; 2168 2169 get_scan_count(lruvec, swappiness, sc, nr, lru_pages); 2170 2171 /* Record the original scan target for proportional adjustments later */ 2172 memcpy(targets, nr, sizeof(nr)); 2173 2174 /* 2175 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal 2176 * event that can occur when there is little memory pressure e.g. 2177 * multiple streaming readers/writers. Hence, we do not abort scanning 2178 * when the requested number of pages are reclaimed when scanning at 2179 * DEF_PRIORITY on the assumption that the fact we are direct 2180 * reclaiming implies that kswapd is not keeping up and it is best to 2181 * do a batch of work at once. For memcg reclaim one check is made to 2182 * abort proportional reclaim if either the file or anon lru has already 2183 * dropped to zero at the first pass. 2184 */ 2185 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && 2186 sc->priority == DEF_PRIORITY); 2187 2188 blk_start_plug(&plug); 2189 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || 2190 nr[LRU_INACTIVE_FILE]) { 2191 unsigned long nr_anon, nr_file, percentage; 2192 unsigned long nr_scanned; 2193 2194 for_each_evictable_lru(lru) { 2195 if (nr[lru]) { 2196 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); 2197 nr[lru] -= nr_to_scan; 2198 2199 nr_reclaimed += shrink_list(lru, nr_to_scan, 2200 lruvec, sc); 2201 } 2202 } 2203 2204 if (nr_reclaimed < nr_to_reclaim || scan_adjusted) 2205 continue; 2206 2207 /* 2208 * For kswapd and memcg, reclaim at least the number of pages 2209 * requested. Ensure that the anon and file LRUs are scanned 2210 * proportionally what was requested by get_scan_count(). We 2211 * stop reclaiming one LRU and reduce the amount scanning 2212 * proportional to the original scan target. 2213 */ 2214 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; 2215 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; 2216 2217 /* 2218 * It's just vindictive to attack the larger once the smaller 2219 * has gone to zero. And given the way we stop scanning the 2220 * smaller below, this makes sure that we only make one nudge 2221 * towards proportionality once we've got nr_to_reclaim. 2222 */ 2223 if (!nr_file || !nr_anon) 2224 break; 2225 2226 if (nr_file > nr_anon) { 2227 unsigned long scan_target = targets[LRU_INACTIVE_ANON] + 2228 targets[LRU_ACTIVE_ANON] + 1; 2229 lru = LRU_BASE; 2230 percentage = nr_anon * 100 / scan_target; 2231 } else { 2232 unsigned long scan_target = targets[LRU_INACTIVE_FILE] + 2233 targets[LRU_ACTIVE_FILE] + 1; 2234 lru = LRU_FILE; 2235 percentage = nr_file * 100 / scan_target; 2236 } 2237 2238 /* Stop scanning the smaller of the LRU */ 2239 nr[lru] = 0; 2240 nr[lru + LRU_ACTIVE] = 0; 2241 2242 /* 2243 * Recalculate the other LRU scan count based on its original 2244 * scan target and the percentage scanning already complete 2245 */ 2246 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; 2247 nr_scanned = targets[lru] - nr[lru]; 2248 nr[lru] = targets[lru] * (100 - percentage) / 100; 2249 nr[lru] -= min(nr[lru], nr_scanned); 2250 2251 lru += LRU_ACTIVE; 2252 nr_scanned = targets[lru] - nr[lru]; 2253 nr[lru] = targets[lru] * (100 - percentage) / 100; 2254 nr[lru] -= min(nr[lru], nr_scanned); 2255 2256 scan_adjusted = true; 2257 } 2258 blk_finish_plug(&plug); 2259 sc->nr_reclaimed += nr_reclaimed; 2260 2261 /* 2262 * Even if we did not try to evict anon pages at all, we want to 2263 * rebalance the anon lru active/inactive ratio. 2264 */ 2265 if (inactive_anon_is_low(lruvec)) 2266 shrink_active_list(SWAP_CLUSTER_MAX, lruvec, 2267 sc, LRU_ACTIVE_ANON); 2268 2269 throttle_vm_writeout(sc->gfp_mask); 2270 } 2271 2272 /* Use reclaim/compaction for costly allocs or under memory pressure */ 2273 static bool in_reclaim_compaction(struct scan_control *sc) 2274 { 2275 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && 2276 (sc->order > PAGE_ALLOC_COSTLY_ORDER || 2277 sc->priority < DEF_PRIORITY - 2)) 2278 return true; 2279 2280 return false; 2281 } 2282 2283 /* 2284 * Reclaim/compaction is used for high-order allocation requests. It reclaims 2285 * order-0 pages before compacting the zone. should_continue_reclaim() returns 2286 * true if more pages should be reclaimed such that when the page allocator 2287 * calls try_to_compact_zone() that it will have enough free pages to succeed. 2288 * It will give up earlier than that if there is difficulty reclaiming pages. 2289 */ 2290 static inline bool should_continue_reclaim(struct zone *zone, 2291 unsigned long nr_reclaimed, 2292 unsigned long nr_scanned, 2293 struct scan_control *sc) 2294 { 2295 unsigned long pages_for_compaction; 2296 unsigned long inactive_lru_pages; 2297 2298 /* If not in reclaim/compaction mode, stop */ 2299 if (!in_reclaim_compaction(sc)) 2300 return false; 2301 2302 /* Consider stopping depending on scan and reclaim activity */ 2303 if (sc->gfp_mask & __GFP_REPEAT) { 2304 /* 2305 * For __GFP_REPEAT allocations, stop reclaiming if the 2306 * full LRU list has been scanned and we are still failing 2307 * to reclaim pages. This full LRU scan is potentially 2308 * expensive but a __GFP_REPEAT caller really wants to succeed 2309 */ 2310 if (!nr_reclaimed && !nr_scanned) 2311 return false; 2312 } else { 2313 /* 2314 * For non-__GFP_REPEAT allocations which can presumably 2315 * fail without consequence, stop if we failed to reclaim 2316 * any pages from the last SWAP_CLUSTER_MAX number of 2317 * pages that were scanned. This will return to the 2318 * caller faster at the risk reclaim/compaction and 2319 * the resulting allocation attempt fails 2320 */ 2321 if (!nr_reclaimed) 2322 return false; 2323 } 2324 2325 /* 2326 * If we have not reclaimed enough pages for compaction and the 2327 * inactive lists are large enough, continue reclaiming 2328 */ 2329 pages_for_compaction = (2UL << sc->order); 2330 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE); 2331 if (get_nr_swap_pages() > 0) 2332 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON); 2333 if (sc->nr_reclaimed < pages_for_compaction && 2334 inactive_lru_pages > pages_for_compaction) 2335 return true; 2336 2337 /* If compaction would go ahead or the allocation would succeed, stop */ 2338 switch (compaction_suitable(zone, sc->order, 0, 0)) { 2339 case COMPACT_PARTIAL: 2340 case COMPACT_CONTINUE: 2341 return false; 2342 default: 2343 return true; 2344 } 2345 } 2346 2347 static bool shrink_zone(struct zone *zone, struct scan_control *sc, 2348 bool is_classzone) 2349 { 2350 struct reclaim_state *reclaim_state = current->reclaim_state; 2351 unsigned long nr_reclaimed, nr_scanned; 2352 bool reclaimable = false; 2353 2354 do { 2355 struct mem_cgroup *root = sc->target_mem_cgroup; 2356 struct mem_cgroup_reclaim_cookie reclaim = { 2357 .zone = zone, 2358 .priority = sc->priority, 2359 }; 2360 unsigned long zone_lru_pages = 0; 2361 struct mem_cgroup *memcg; 2362 2363 nr_reclaimed = sc->nr_reclaimed; 2364 nr_scanned = sc->nr_scanned; 2365 2366 memcg = mem_cgroup_iter(root, NULL, &reclaim); 2367 do { 2368 unsigned long lru_pages; 2369 unsigned long scanned; 2370 struct lruvec *lruvec; 2371 int swappiness; 2372 2373 if (mem_cgroup_low(root, memcg)) { 2374 if (!sc->may_thrash) 2375 continue; 2376 mem_cgroup_events(memcg, MEMCG_LOW, 1); 2377 } 2378 2379 lruvec = mem_cgroup_zone_lruvec(zone, memcg); 2380 swappiness = mem_cgroup_swappiness(memcg); 2381 scanned = sc->nr_scanned; 2382 2383 shrink_lruvec(lruvec, swappiness, sc, &lru_pages); 2384 zone_lru_pages += lru_pages; 2385 2386 if (memcg && is_classzone) 2387 shrink_slab(sc->gfp_mask, zone_to_nid(zone), 2388 memcg, sc->nr_scanned - scanned, 2389 lru_pages); 2390 2391 /* 2392 * Direct reclaim and kswapd have to scan all memory 2393 * cgroups to fulfill the overall scan target for the 2394 * zone. 2395 * 2396 * Limit reclaim, on the other hand, only cares about 2397 * nr_to_reclaim pages to be reclaimed and it will 2398 * retry with decreasing priority if one round over the 2399 * whole hierarchy is not sufficient. 2400 */ 2401 if (!global_reclaim(sc) && 2402 sc->nr_reclaimed >= sc->nr_to_reclaim) { 2403 mem_cgroup_iter_break(root, memcg); 2404 break; 2405 } 2406 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); 2407 2408 /* 2409 * Shrink the slab caches in the same proportion that 2410 * the eligible LRU pages were scanned. 2411 */ 2412 if (global_reclaim(sc) && is_classzone) 2413 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL, 2414 sc->nr_scanned - nr_scanned, 2415 zone_lru_pages); 2416 2417 if (reclaim_state) { 2418 sc->nr_reclaimed += reclaim_state->reclaimed_slab; 2419 reclaim_state->reclaimed_slab = 0; 2420 } 2421 2422 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, 2423 sc->nr_scanned - nr_scanned, 2424 sc->nr_reclaimed - nr_reclaimed); 2425 2426 if (sc->nr_reclaimed - nr_reclaimed) 2427 reclaimable = true; 2428 2429 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed, 2430 sc->nr_scanned - nr_scanned, sc)); 2431 2432 return reclaimable; 2433 } 2434 2435 /* 2436 * Returns true if compaction should go ahead for a high-order request, or 2437 * the high-order allocation would succeed without compaction. 2438 */ 2439 static inline bool compaction_ready(struct zone *zone, int order) 2440 { 2441 unsigned long balance_gap, watermark; 2442 bool watermark_ok; 2443 2444 /* 2445 * Compaction takes time to run and there are potentially other 2446 * callers using the pages just freed. Continue reclaiming until 2447 * there is a buffer of free pages available to give compaction 2448 * a reasonable chance of completing and allocating the page 2449 */ 2450 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( 2451 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); 2452 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order); 2453 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0); 2454 2455 /* 2456 * If compaction is deferred, reclaim up to a point where 2457 * compaction will have a chance of success when re-enabled 2458 */ 2459 if (compaction_deferred(zone, order)) 2460 return watermark_ok; 2461 2462 /* 2463 * If compaction is not ready to start and allocation is not likely 2464 * to succeed without it, then keep reclaiming. 2465 */ 2466 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED) 2467 return false; 2468 2469 return watermark_ok; 2470 } 2471 2472 /* 2473 * This is the direct reclaim path, for page-allocating processes. We only 2474 * try to reclaim pages from zones which will satisfy the caller's allocation 2475 * request. 2476 * 2477 * We reclaim from a zone even if that zone is over high_wmark_pages(zone). 2478 * Because: 2479 * a) The caller may be trying to free *extra* pages to satisfy a higher-order 2480 * allocation or 2481 * b) The target zone may be at high_wmark_pages(zone) but the lower zones 2482 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min' 2483 * zone defense algorithm. 2484 * 2485 * If a zone is deemed to be full of pinned pages then just give it a light 2486 * scan then give up on it. 2487 * 2488 * Returns true if a zone was reclaimable. 2489 */ 2490 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc) 2491 { 2492 struct zoneref *z; 2493 struct zone *zone; 2494 unsigned long nr_soft_reclaimed; 2495 unsigned long nr_soft_scanned; 2496 gfp_t orig_mask; 2497 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask); 2498 bool reclaimable = false; 2499 2500 /* 2501 * If the number of buffer_heads in the machine exceeds the maximum 2502 * allowed level, force direct reclaim to scan the highmem zone as 2503 * highmem pages could be pinning lowmem pages storing buffer_heads 2504 */ 2505 orig_mask = sc->gfp_mask; 2506 if (buffer_heads_over_limit) 2507 sc->gfp_mask |= __GFP_HIGHMEM; 2508 2509 for_each_zone_zonelist_nodemask(zone, z, zonelist, 2510 requested_highidx, sc->nodemask) { 2511 enum zone_type classzone_idx; 2512 2513 if (!populated_zone(zone)) 2514 continue; 2515 2516 classzone_idx = requested_highidx; 2517 while (!populated_zone(zone->zone_pgdat->node_zones + 2518 classzone_idx)) 2519 classzone_idx--; 2520 2521 /* 2522 * Take care memory controller reclaiming has small influence 2523 * to global LRU. 2524 */ 2525 if (global_reclaim(sc)) { 2526 if (!cpuset_zone_allowed(zone, 2527 GFP_KERNEL | __GFP_HARDWALL)) 2528 continue; 2529 2530 if (sc->priority != DEF_PRIORITY && 2531 !zone_reclaimable(zone)) 2532 continue; /* Let kswapd poll it */ 2533 2534 /* 2535 * If we already have plenty of memory free for 2536 * compaction in this zone, don't free any more. 2537 * Even though compaction is invoked for any 2538 * non-zero order, only frequent costly order 2539 * reclamation is disruptive enough to become a 2540 * noticeable problem, like transparent huge 2541 * page allocations. 2542 */ 2543 if (IS_ENABLED(CONFIG_COMPACTION) && 2544 sc->order > PAGE_ALLOC_COSTLY_ORDER && 2545 zonelist_zone_idx(z) <= requested_highidx && 2546 compaction_ready(zone, sc->order)) { 2547 sc->compaction_ready = true; 2548 continue; 2549 } 2550 2551 /* 2552 * This steals pages from memory cgroups over softlimit 2553 * and returns the number of reclaimed pages and 2554 * scanned pages. This works for global memory pressure 2555 * and balancing, not for a memcg's limit. 2556 */ 2557 nr_soft_scanned = 0; 2558 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, 2559 sc->order, sc->gfp_mask, 2560 &nr_soft_scanned); 2561 sc->nr_reclaimed += nr_soft_reclaimed; 2562 sc->nr_scanned += nr_soft_scanned; 2563 if (nr_soft_reclaimed) 2564 reclaimable = true; 2565 /* need some check for avoid more shrink_zone() */ 2566 } 2567 2568 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx)) 2569 reclaimable = true; 2570 2571 if (global_reclaim(sc) && 2572 !reclaimable && zone_reclaimable(zone)) 2573 reclaimable = true; 2574 } 2575 2576 /* 2577 * Restore to original mask to avoid the impact on the caller if we 2578 * promoted it to __GFP_HIGHMEM. 2579 */ 2580 sc->gfp_mask = orig_mask; 2581 2582 return reclaimable; 2583 } 2584 2585 /* 2586 * This is the main entry point to direct page reclaim. 2587 * 2588 * If a full scan of the inactive list fails to free enough memory then we 2589 * are "out of memory" and something needs to be killed. 2590 * 2591 * If the caller is !__GFP_FS then the probability of a failure is reasonably 2592 * high - the zone may be full of dirty or under-writeback pages, which this 2593 * caller can't do much about. We kick the writeback threads and take explicit 2594 * naps in the hope that some of these pages can be written. But if the 2595 * allocating task holds filesystem locks which prevent writeout this might not 2596 * work, and the allocation attempt will fail. 2597 * 2598 * returns: 0, if no pages reclaimed 2599 * else, the number of pages reclaimed 2600 */ 2601 static unsigned long do_try_to_free_pages(struct zonelist *zonelist, 2602 struct scan_control *sc) 2603 { 2604 int initial_priority = sc->priority; 2605 unsigned long total_scanned = 0; 2606 unsigned long writeback_threshold; 2607 bool zones_reclaimable; 2608 retry: 2609 delayacct_freepages_start(); 2610 2611 if (global_reclaim(sc)) 2612 count_vm_event(ALLOCSTALL); 2613 2614 do { 2615 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, 2616 sc->priority); 2617 sc->nr_scanned = 0; 2618 zones_reclaimable = shrink_zones(zonelist, sc); 2619 2620 total_scanned += sc->nr_scanned; 2621 if (sc->nr_reclaimed >= sc->nr_to_reclaim) 2622 break; 2623 2624 if (sc->compaction_ready) 2625 break; 2626 2627 /* 2628 * If we're getting trouble reclaiming, start doing 2629 * writepage even in laptop mode. 2630 */ 2631 if (sc->priority < DEF_PRIORITY - 2) 2632 sc->may_writepage = 1; 2633 2634 /* 2635 * Try to write back as many pages as we just scanned. This 2636 * tends to cause slow streaming writers to write data to the 2637 * disk smoothly, at the dirtying rate, which is nice. But 2638 * that's undesirable in laptop mode, where we *want* lumpy 2639 * writeout. So in laptop mode, write out the whole world. 2640 */ 2641 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; 2642 if (total_scanned > writeback_threshold) { 2643 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned, 2644 WB_REASON_TRY_TO_FREE_PAGES); 2645 sc->may_writepage = 1; 2646 } 2647 } while (--sc->priority >= 0); 2648 2649 delayacct_freepages_end(); 2650 2651 if (sc->nr_reclaimed) 2652 return sc->nr_reclaimed; 2653 2654 /* Aborted reclaim to try compaction? don't OOM, then */ 2655 if (sc->compaction_ready) 2656 return 1; 2657 2658 /* Untapped cgroup reserves? Don't OOM, retry. */ 2659 if (!sc->may_thrash) { 2660 sc->priority = initial_priority; 2661 sc->may_thrash = 1; 2662 goto retry; 2663 } 2664 2665 /* Any of the zones still reclaimable? Don't OOM. */ 2666 if (zones_reclaimable) 2667 return 1; 2668 2669 return 0; 2670 } 2671 2672 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat) 2673 { 2674 struct zone *zone; 2675 unsigned long pfmemalloc_reserve = 0; 2676 unsigned long free_pages = 0; 2677 int i; 2678 bool wmark_ok; 2679 2680 for (i = 0; i <= ZONE_NORMAL; i++) { 2681 zone = &pgdat->node_zones[i]; 2682 if (!populated_zone(zone) || 2683 zone_reclaimable_pages(zone) == 0) 2684 continue; 2685 2686 pfmemalloc_reserve += min_wmark_pages(zone); 2687 free_pages += zone_page_state(zone, NR_FREE_PAGES); 2688 } 2689 2690 /* If there are no reserves (unexpected config) then do not throttle */ 2691 if (!pfmemalloc_reserve) 2692 return true; 2693 2694 wmark_ok = free_pages > pfmemalloc_reserve / 2; 2695 2696 /* kswapd must be awake if processes are being throttled */ 2697 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { 2698 pgdat->classzone_idx = min(pgdat->classzone_idx, 2699 (enum zone_type)ZONE_NORMAL); 2700 wake_up_interruptible(&pgdat->kswapd_wait); 2701 } 2702 2703 return wmark_ok; 2704 } 2705 2706 /* 2707 * Throttle direct reclaimers if backing storage is backed by the network 2708 * and the PFMEMALLOC reserve for the preferred node is getting dangerously 2709 * depleted. kswapd will continue to make progress and wake the processes 2710 * when the low watermark is reached. 2711 * 2712 * Returns true if a fatal signal was delivered during throttling. If this 2713 * happens, the page allocator should not consider triggering the OOM killer. 2714 */ 2715 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, 2716 nodemask_t *nodemask) 2717 { 2718 struct zoneref *z; 2719 struct zone *zone; 2720 pg_data_t *pgdat = NULL; 2721 2722 /* 2723 * Kernel threads should not be throttled as they may be indirectly 2724 * responsible for cleaning pages necessary for reclaim to make forward 2725 * progress. kjournald for example may enter direct reclaim while 2726 * committing a transaction where throttling it could forcing other 2727 * processes to block on log_wait_commit(). 2728 */ 2729 if (current->flags & PF_KTHREAD) 2730 goto out; 2731 2732 /* 2733 * If a fatal signal is pending, this process should not throttle. 2734 * It should return quickly so it can exit and free its memory 2735 */ 2736 if (fatal_signal_pending(current)) 2737 goto out; 2738 2739 /* 2740 * Check if the pfmemalloc reserves are ok by finding the first node 2741 * with a usable ZONE_NORMAL or lower zone. The expectation is that 2742 * GFP_KERNEL will be required for allocating network buffers when 2743 * swapping over the network so ZONE_HIGHMEM is unusable. 2744 * 2745 * Throttling is based on the first usable node and throttled processes 2746 * wait on a queue until kswapd makes progress and wakes them. There 2747 * is an affinity then between processes waking up and where reclaim 2748 * progress has been made assuming the process wakes on the same node. 2749 * More importantly, processes running on remote nodes will not compete 2750 * for remote pfmemalloc reserves and processes on different nodes 2751 * should make reasonable progress. 2752 */ 2753 for_each_zone_zonelist_nodemask(zone, z, zonelist, 2754 gfp_zone(gfp_mask), nodemask) { 2755 if (zone_idx(zone) > ZONE_NORMAL) 2756 continue; 2757 2758 /* Throttle based on the first usable node */ 2759 pgdat = zone->zone_pgdat; 2760 if (pfmemalloc_watermark_ok(pgdat)) 2761 goto out; 2762 break; 2763 } 2764 2765 /* If no zone was usable by the allocation flags then do not throttle */ 2766 if (!pgdat) 2767 goto out; 2768 2769 /* Account for the throttling */ 2770 count_vm_event(PGSCAN_DIRECT_THROTTLE); 2771 2772 /* 2773 * If the caller cannot enter the filesystem, it's possible that it 2774 * is due to the caller holding an FS lock or performing a journal 2775 * transaction in the case of a filesystem like ext[3|4]. In this case, 2776 * it is not safe to block on pfmemalloc_wait as kswapd could be 2777 * blocked waiting on the same lock. Instead, throttle for up to a 2778 * second before continuing. 2779 */ 2780 if (!(gfp_mask & __GFP_FS)) { 2781 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, 2782 pfmemalloc_watermark_ok(pgdat), HZ); 2783 2784 goto check_pending; 2785 } 2786 2787 /* Throttle until kswapd wakes the process */ 2788 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, 2789 pfmemalloc_watermark_ok(pgdat)); 2790 2791 check_pending: 2792 if (fatal_signal_pending(current)) 2793 return true; 2794 2795 out: 2796 return false; 2797 } 2798 2799 unsigned long try_to_free_pages(struct zonelist *zonelist, int order, 2800 gfp_t gfp_mask, nodemask_t *nodemask) 2801 { 2802 unsigned long nr_reclaimed; 2803 struct scan_control sc = { 2804 .nr_to_reclaim = SWAP_CLUSTER_MAX, 2805 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), 2806 .order = order, 2807 .nodemask = nodemask, 2808 .priority = DEF_PRIORITY, 2809 .may_writepage = !laptop_mode, 2810 .may_unmap = 1, 2811 .may_swap = 1, 2812 }; 2813 2814 /* 2815 * Do not enter reclaim if fatal signal was delivered while throttled. 2816 * 1 is returned so that the page allocator does not OOM kill at this 2817 * point. 2818 */ 2819 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask)) 2820 return 1; 2821 2822 trace_mm_vmscan_direct_reclaim_begin(order, 2823 sc.may_writepage, 2824 gfp_mask); 2825 2826 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 2827 2828 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); 2829 2830 return nr_reclaimed; 2831 } 2832 2833 #ifdef CONFIG_MEMCG 2834 2835 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg, 2836 gfp_t gfp_mask, bool noswap, 2837 struct zone *zone, 2838 unsigned long *nr_scanned) 2839 { 2840 struct scan_control sc = { 2841 .nr_to_reclaim = SWAP_CLUSTER_MAX, 2842 .target_mem_cgroup = memcg, 2843 .may_writepage = !laptop_mode, 2844 .may_unmap = 1, 2845 .may_swap = !noswap, 2846 }; 2847 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); 2848 int swappiness = mem_cgroup_swappiness(memcg); 2849 unsigned long lru_pages; 2850 2851 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | 2852 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); 2853 2854 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, 2855 sc.may_writepage, 2856 sc.gfp_mask); 2857 2858 /* 2859 * NOTE: Although we can get the priority field, using it 2860 * here is not a good idea, since it limits the pages we can scan. 2861 * if we don't reclaim here, the shrink_zone from balance_pgdat 2862 * will pick up pages from other mem cgroup's as well. We hack 2863 * the priority and make it zero. 2864 */ 2865 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages); 2866 2867 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); 2868 2869 *nr_scanned = sc.nr_scanned; 2870 return sc.nr_reclaimed; 2871 } 2872 2873 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, 2874 unsigned long nr_pages, 2875 gfp_t gfp_mask, 2876 bool may_swap) 2877 { 2878 struct zonelist *zonelist; 2879 unsigned long nr_reclaimed; 2880 int nid; 2881 struct scan_control sc = { 2882 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), 2883 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | 2884 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), 2885 .target_mem_cgroup = memcg, 2886 .priority = DEF_PRIORITY, 2887 .may_writepage = !laptop_mode, 2888 .may_unmap = 1, 2889 .may_swap = may_swap, 2890 }; 2891 2892 /* 2893 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't 2894 * take care of from where we get pages. So the node where we start the 2895 * scan does not need to be the current node. 2896 */ 2897 nid = mem_cgroup_select_victim_node(memcg); 2898 2899 zonelist = NODE_DATA(nid)->node_zonelists; 2900 2901 trace_mm_vmscan_memcg_reclaim_begin(0, 2902 sc.may_writepage, 2903 sc.gfp_mask); 2904 2905 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 2906 2907 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); 2908 2909 return nr_reclaimed; 2910 } 2911 #endif 2912 2913 static void age_active_anon(struct zone *zone, struct scan_control *sc) 2914 { 2915 struct mem_cgroup *memcg; 2916 2917 if (!total_swap_pages) 2918 return; 2919 2920 memcg = mem_cgroup_iter(NULL, NULL, NULL); 2921 do { 2922 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg); 2923 2924 if (inactive_anon_is_low(lruvec)) 2925 shrink_active_list(SWAP_CLUSTER_MAX, lruvec, 2926 sc, LRU_ACTIVE_ANON); 2927 2928 memcg = mem_cgroup_iter(NULL, memcg, NULL); 2929 } while (memcg); 2930 } 2931 2932 static bool zone_balanced(struct zone *zone, int order, 2933 unsigned long balance_gap, int classzone_idx) 2934 { 2935 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) + 2936 balance_gap, classzone_idx, 0)) 2937 return false; 2938 2939 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone, 2940 order, 0, classzone_idx) == COMPACT_SKIPPED) 2941 return false; 2942 2943 return true; 2944 } 2945 2946 /* 2947 * pgdat_balanced() is used when checking if a node is balanced. 2948 * 2949 * For order-0, all zones must be balanced! 2950 * 2951 * For high-order allocations only zones that meet watermarks and are in a 2952 * zone allowed by the callers classzone_idx are added to balanced_pages. The 2953 * total of balanced pages must be at least 25% of the zones allowed by 2954 * classzone_idx for the node to be considered balanced. Forcing all zones to 2955 * be balanced for high orders can cause excessive reclaim when there are 2956 * imbalanced zones. 2957 * The choice of 25% is due to 2958 * o a 16M DMA zone that is balanced will not balance a zone on any 2959 * reasonable sized machine 2960 * o On all other machines, the top zone must be at least a reasonable 2961 * percentage of the middle zones. For example, on 32-bit x86, highmem 2962 * would need to be at least 256M for it to be balance a whole node. 2963 * Similarly, on x86-64 the Normal zone would need to be at least 1G 2964 * to balance a node on its own. These seemed like reasonable ratios. 2965 */ 2966 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) 2967 { 2968 unsigned long managed_pages = 0; 2969 unsigned long balanced_pages = 0; 2970 int i; 2971 2972 /* Check the watermark levels */ 2973 for (i = 0; i <= classzone_idx; i++) { 2974 struct zone *zone = pgdat->node_zones + i; 2975 2976 if (!populated_zone(zone)) 2977 continue; 2978 2979 managed_pages += zone->managed_pages; 2980 2981 /* 2982 * A special case here: 2983 * 2984 * balance_pgdat() skips over all_unreclaimable after 2985 * DEF_PRIORITY. Effectively, it considers them balanced so 2986 * they must be considered balanced here as well! 2987 */ 2988 if (!zone_reclaimable(zone)) { 2989 balanced_pages += zone->managed_pages; 2990 continue; 2991 } 2992 2993 if (zone_balanced(zone, order, 0, i)) 2994 balanced_pages += zone->managed_pages; 2995 else if (!order) 2996 return false; 2997 } 2998 2999 if (order) 3000 return balanced_pages >= (managed_pages >> 2); 3001 else 3002 return true; 3003 } 3004 3005 /* 3006 * Prepare kswapd for sleeping. This verifies that there are no processes 3007 * waiting in throttle_direct_reclaim() and that watermarks have been met. 3008 * 3009 * Returns true if kswapd is ready to sleep 3010 */ 3011 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining, 3012 int classzone_idx) 3013 { 3014 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */ 3015 if (remaining) 3016 return false; 3017 3018 /* 3019 * The throttled processes are normally woken up in balance_pgdat() as 3020 * soon as pfmemalloc_watermark_ok() is true. But there is a potential 3021 * race between when kswapd checks the watermarks and a process gets 3022 * throttled. There is also a potential race if processes get 3023 * throttled, kswapd wakes, a large process exits thereby balancing the 3024 * zones, which causes kswapd to exit balance_pgdat() before reaching 3025 * the wake up checks. If kswapd is going to sleep, no process should 3026 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If 3027 * the wake up is premature, processes will wake kswapd and get 3028 * throttled again. The difference from wake ups in balance_pgdat() is 3029 * that here we are under prepare_to_wait(). 3030 */ 3031 if (waitqueue_active(&pgdat->pfmemalloc_wait)) 3032 wake_up_all(&pgdat->pfmemalloc_wait); 3033 3034 return pgdat_balanced(pgdat, order, classzone_idx); 3035 } 3036 3037 /* 3038 * kswapd shrinks the zone by the number of pages required to reach 3039 * the high watermark. 3040 * 3041 * Returns true if kswapd scanned at least the requested number of pages to 3042 * reclaim or if the lack of progress was due to pages under writeback. 3043 * This is used to determine if the scanning priority needs to be raised. 3044 */ 3045 static bool kswapd_shrink_zone(struct zone *zone, 3046 int classzone_idx, 3047 struct scan_control *sc, 3048 unsigned long *nr_attempted) 3049 { 3050 int testorder = sc->order; 3051 unsigned long balance_gap; 3052 bool lowmem_pressure; 3053 3054 /* Reclaim above the high watermark. */ 3055 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone)); 3056 3057 /* 3058 * Kswapd reclaims only single pages with compaction enabled. Trying 3059 * too hard to reclaim until contiguous free pages have become 3060 * available can hurt performance by evicting too much useful data 3061 * from memory. Do not reclaim more than needed for compaction. 3062 */ 3063 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && 3064 compaction_suitable(zone, sc->order, 0, classzone_idx) 3065 != COMPACT_SKIPPED) 3066 testorder = 0; 3067 3068 /* 3069 * We put equal pressure on every zone, unless one zone has way too 3070 * many pages free already. The "too many pages" is defined as the 3071 * high wmark plus a "gap" where the gap is either the low 3072 * watermark or 1% of the zone, whichever is smaller. 3073 */ 3074 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP( 3075 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO)); 3076 3077 /* 3078 * If there is no low memory pressure or the zone is balanced then no 3079 * reclaim is necessary 3080 */ 3081 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone)); 3082 if (!lowmem_pressure && zone_balanced(zone, testorder, 3083 balance_gap, classzone_idx)) 3084 return true; 3085 3086 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx); 3087 3088 /* Account for the number of pages attempted to reclaim */ 3089 *nr_attempted += sc->nr_to_reclaim; 3090 3091 clear_bit(ZONE_WRITEBACK, &zone->flags); 3092 3093 /* 3094 * If a zone reaches its high watermark, consider it to be no longer 3095 * congested. It's possible there are dirty pages backed by congested 3096 * BDIs but as pressure is relieved, speculatively avoid congestion 3097 * waits. 3098 */ 3099 if (zone_reclaimable(zone) && 3100 zone_balanced(zone, testorder, 0, classzone_idx)) { 3101 clear_bit(ZONE_CONGESTED, &zone->flags); 3102 clear_bit(ZONE_DIRTY, &zone->flags); 3103 } 3104 3105 return sc->nr_scanned >= sc->nr_to_reclaim; 3106 } 3107 3108 /* 3109 * For kswapd, balance_pgdat() will work across all this node's zones until 3110 * they are all at high_wmark_pages(zone). 3111 * 3112 * Returns the final order kswapd was reclaiming at 3113 * 3114 * There is special handling here for zones which are full of pinned pages. 3115 * This can happen if the pages are all mlocked, or if they are all used by 3116 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. 3117 * What we do is to detect the case where all pages in the zone have been 3118 * scanned twice and there has been zero successful reclaim. Mark the zone as 3119 * dead and from now on, only perform a short scan. Basically we're polling 3120 * the zone for when the problem goes away. 3121 * 3122 * kswapd scans the zones in the highmem->normal->dma direction. It skips 3123 * zones which have free_pages > high_wmark_pages(zone), but once a zone is 3124 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the 3125 * lower zones regardless of the number of free pages in the lower zones. This 3126 * interoperates with the page allocator fallback scheme to ensure that aging 3127 * of pages is balanced across the zones. 3128 */ 3129 static unsigned long balance_pgdat(pg_data_t *pgdat, int order, 3130 int *classzone_idx) 3131 { 3132 int i; 3133 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ 3134 unsigned long nr_soft_reclaimed; 3135 unsigned long nr_soft_scanned; 3136 struct scan_control sc = { 3137 .gfp_mask = GFP_KERNEL, 3138 .order = order, 3139 .priority = DEF_PRIORITY, 3140 .may_writepage = !laptop_mode, 3141 .may_unmap = 1, 3142 .may_swap = 1, 3143 }; 3144 count_vm_event(PAGEOUTRUN); 3145 3146 do { 3147 unsigned long nr_attempted = 0; 3148 bool raise_priority = true; 3149 bool pgdat_needs_compaction = (order > 0); 3150 3151 sc.nr_reclaimed = 0; 3152 3153 /* 3154 * Scan in the highmem->dma direction for the highest 3155 * zone which needs scanning 3156 */ 3157 for (i = pgdat->nr_zones - 1; i >= 0; i--) { 3158 struct zone *zone = pgdat->node_zones + i; 3159 3160 if (!populated_zone(zone)) 3161 continue; 3162 3163 if (sc.priority != DEF_PRIORITY && 3164 !zone_reclaimable(zone)) 3165 continue; 3166 3167 /* 3168 * Do some background aging of the anon list, to give 3169 * pages a chance to be referenced before reclaiming. 3170 */ 3171 age_active_anon(zone, &sc); 3172 3173 /* 3174 * If the number of buffer_heads in the machine 3175 * exceeds the maximum allowed level and this node 3176 * has a highmem zone, force kswapd to reclaim from 3177 * it to relieve lowmem pressure. 3178 */ 3179 if (buffer_heads_over_limit && is_highmem_idx(i)) { 3180 end_zone = i; 3181 break; 3182 } 3183 3184 if (!zone_balanced(zone, order, 0, 0)) { 3185 end_zone = i; 3186 break; 3187 } else { 3188 /* 3189 * If balanced, clear the dirty and congested 3190 * flags 3191 */ 3192 clear_bit(ZONE_CONGESTED, &zone->flags); 3193 clear_bit(ZONE_DIRTY, &zone->flags); 3194 } 3195 } 3196 3197 if (i < 0) 3198 goto out; 3199 3200 for (i = 0; i <= end_zone; i++) { 3201 struct zone *zone = pgdat->node_zones + i; 3202 3203 if (!populated_zone(zone)) 3204 continue; 3205 3206 /* 3207 * If any zone is currently balanced then kswapd will 3208 * not call compaction as it is expected that the 3209 * necessary pages are already available. 3210 */ 3211 if (pgdat_needs_compaction && 3212 zone_watermark_ok(zone, order, 3213 low_wmark_pages(zone), 3214 *classzone_idx, 0)) 3215 pgdat_needs_compaction = false; 3216 } 3217 3218 /* 3219 * If we're getting trouble reclaiming, start doing writepage 3220 * even in laptop mode. 3221 */ 3222 if (sc.priority < DEF_PRIORITY - 2) 3223 sc.may_writepage = 1; 3224 3225 /* 3226 * Now scan the zone in the dma->highmem direction, stopping 3227 * at the last zone which needs scanning. 3228 * 3229 * We do this because the page allocator works in the opposite 3230 * direction. This prevents the page allocator from allocating 3231 * pages behind kswapd's direction of progress, which would 3232 * cause too much scanning of the lower zones. 3233 */ 3234 for (i = 0; i <= end_zone; i++) { 3235 struct zone *zone = pgdat->node_zones + i; 3236 3237 if (!populated_zone(zone)) 3238 continue; 3239 3240 if (sc.priority != DEF_PRIORITY && 3241 !zone_reclaimable(zone)) 3242 continue; 3243 3244 sc.nr_scanned = 0; 3245 3246 nr_soft_scanned = 0; 3247 /* 3248 * Call soft limit reclaim before calling shrink_zone. 3249 */ 3250 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone, 3251 order, sc.gfp_mask, 3252 &nr_soft_scanned); 3253 sc.nr_reclaimed += nr_soft_reclaimed; 3254 3255 /* 3256 * There should be no need to raise the scanning 3257 * priority if enough pages are already being scanned 3258 * that that high watermark would be met at 100% 3259 * efficiency. 3260 */ 3261 if (kswapd_shrink_zone(zone, end_zone, 3262 &sc, &nr_attempted)) 3263 raise_priority = false; 3264 } 3265 3266 /* 3267 * If the low watermark is met there is no need for processes 3268 * to be throttled on pfmemalloc_wait as they should not be 3269 * able to safely make forward progress. Wake them 3270 */ 3271 if (waitqueue_active(&pgdat->pfmemalloc_wait) && 3272 pfmemalloc_watermark_ok(pgdat)) 3273 wake_up_all(&pgdat->pfmemalloc_wait); 3274 3275 /* 3276 * Fragmentation may mean that the system cannot be rebalanced 3277 * for high-order allocations in all zones. If twice the 3278 * allocation size has been reclaimed and the zones are still 3279 * not balanced then recheck the watermarks at order-0 to 3280 * prevent kswapd reclaiming excessively. Assume that a 3281 * process requested a high-order can direct reclaim/compact. 3282 */ 3283 if (order && sc.nr_reclaimed >= 2UL << order) 3284 order = sc.order = 0; 3285 3286 /* Check if kswapd should be suspending */ 3287 if (try_to_freeze() || kthread_should_stop()) 3288 break; 3289 3290 /* 3291 * Compact if necessary and kswapd is reclaiming at least the 3292 * high watermark number of pages as requsted 3293 */ 3294 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted) 3295 compact_pgdat(pgdat, order); 3296 3297 /* 3298 * Raise priority if scanning rate is too low or there was no 3299 * progress in reclaiming pages 3300 */ 3301 if (raise_priority || !sc.nr_reclaimed) 3302 sc.priority--; 3303 } while (sc.priority >= 1 && 3304 !pgdat_balanced(pgdat, order, *classzone_idx)); 3305 3306 out: 3307 /* 3308 * Return the order we were reclaiming at so prepare_kswapd_sleep() 3309 * makes a decision on the order we were last reclaiming at. However, 3310 * if another caller entered the allocator slow path while kswapd 3311 * was awake, order will remain at the higher level 3312 */ 3313 *classzone_idx = end_zone; 3314 return order; 3315 } 3316 3317 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx) 3318 { 3319 long remaining = 0; 3320 DEFINE_WAIT(wait); 3321 3322 if (freezing(current) || kthread_should_stop()) 3323 return; 3324 3325 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 3326 3327 /* Try to sleep for a short interval */ 3328 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { 3329 remaining = schedule_timeout(HZ/10); 3330 finish_wait(&pgdat->kswapd_wait, &wait); 3331 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 3332 } 3333 3334 /* 3335 * After a short sleep, check if it was a premature sleep. If not, then 3336 * go fully to sleep until explicitly woken up. 3337 */ 3338 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) { 3339 trace_mm_vmscan_kswapd_sleep(pgdat->node_id); 3340 3341 /* 3342 * vmstat counters are not perfectly accurate and the estimated 3343 * value for counters such as NR_FREE_PAGES can deviate from the 3344 * true value by nr_online_cpus * threshold. To avoid the zone 3345 * watermarks being breached while under pressure, we reduce the 3346 * per-cpu vmstat threshold while kswapd is awake and restore 3347 * them before going back to sleep. 3348 */ 3349 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); 3350 3351 /* 3352 * Compaction records what page blocks it recently failed to 3353 * isolate pages from and skips them in the future scanning. 3354 * When kswapd is going to sleep, it is reasonable to assume 3355 * that pages and compaction may succeed so reset the cache. 3356 */ 3357 reset_isolation_suitable(pgdat); 3358 3359 if (!kthread_should_stop()) 3360 schedule(); 3361 3362 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); 3363 } else { 3364 if (remaining) 3365 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); 3366 else 3367 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); 3368 } 3369 finish_wait(&pgdat->kswapd_wait, &wait); 3370 } 3371 3372 /* 3373 * The background pageout daemon, started as a kernel thread 3374 * from the init process. 3375 * 3376 * This basically trickles out pages so that we have _some_ 3377 * free memory available even if there is no other activity 3378 * that frees anything up. This is needed for things like routing 3379 * etc, where we otherwise might have all activity going on in 3380 * asynchronous contexts that cannot page things out. 3381 * 3382 * If there are applications that are active memory-allocators 3383 * (most normal use), this basically shouldn't matter. 3384 */ 3385 static int kswapd(void *p) 3386 { 3387 unsigned long order, new_order; 3388 unsigned balanced_order; 3389 int classzone_idx, new_classzone_idx; 3390 int balanced_classzone_idx; 3391 pg_data_t *pgdat = (pg_data_t*)p; 3392 struct task_struct *tsk = current; 3393 3394 struct reclaim_state reclaim_state = { 3395 .reclaimed_slab = 0, 3396 }; 3397 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); 3398 3399 lockdep_set_current_reclaim_state(GFP_KERNEL); 3400 3401 if (!cpumask_empty(cpumask)) 3402 set_cpus_allowed_ptr(tsk, cpumask); 3403 current->reclaim_state = &reclaim_state; 3404 3405 /* 3406 * Tell the memory management that we're a "memory allocator", 3407 * and that if we need more memory we should get access to it 3408 * regardless (see "__alloc_pages()"). "kswapd" should 3409 * never get caught in the normal page freeing logic. 3410 * 3411 * (Kswapd normally doesn't need memory anyway, but sometimes 3412 * you need a small amount of memory in order to be able to 3413 * page out something else, and this flag essentially protects 3414 * us from recursively trying to free more memory as we're 3415 * trying to free the first piece of memory in the first place). 3416 */ 3417 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; 3418 set_freezable(); 3419 3420 order = new_order = 0; 3421 balanced_order = 0; 3422 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1; 3423 balanced_classzone_idx = classzone_idx; 3424 for ( ; ; ) { 3425 bool ret; 3426 3427 /* 3428 * If the last balance_pgdat was unsuccessful it's unlikely a 3429 * new request of a similar or harder type will succeed soon 3430 * so consider going to sleep on the basis we reclaimed at 3431 */ 3432 if (balanced_classzone_idx >= new_classzone_idx && 3433 balanced_order == new_order) { 3434 new_order = pgdat->kswapd_max_order; 3435 new_classzone_idx = pgdat->classzone_idx; 3436 pgdat->kswapd_max_order = 0; 3437 pgdat->classzone_idx = pgdat->nr_zones - 1; 3438 } 3439 3440 if (order < new_order || classzone_idx > new_classzone_idx) { 3441 /* 3442 * Don't sleep if someone wants a larger 'order' 3443 * allocation or has tigher zone constraints 3444 */ 3445 order = new_order; 3446 classzone_idx = new_classzone_idx; 3447 } else { 3448 kswapd_try_to_sleep(pgdat, balanced_order, 3449 balanced_classzone_idx); 3450 order = pgdat->kswapd_max_order; 3451 classzone_idx = pgdat->classzone_idx; 3452 new_order = order; 3453 new_classzone_idx = classzone_idx; 3454 pgdat->kswapd_max_order = 0; 3455 pgdat->classzone_idx = pgdat->nr_zones - 1; 3456 } 3457 3458 ret = try_to_freeze(); 3459 if (kthread_should_stop()) 3460 break; 3461 3462 /* 3463 * We can speed up thawing tasks if we don't call balance_pgdat 3464 * after returning from the refrigerator 3465 */ 3466 if (!ret) { 3467 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); 3468 balanced_classzone_idx = classzone_idx; 3469 balanced_order = balance_pgdat(pgdat, order, 3470 &balanced_classzone_idx); 3471 } 3472 } 3473 3474 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); 3475 current->reclaim_state = NULL; 3476 lockdep_clear_current_reclaim_state(); 3477 3478 return 0; 3479 } 3480 3481 /* 3482 * A zone is low on free memory, so wake its kswapd task to service it. 3483 */ 3484 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) 3485 { 3486 pg_data_t *pgdat; 3487 3488 if (!populated_zone(zone)) 3489 return; 3490 3491 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) 3492 return; 3493 pgdat = zone->zone_pgdat; 3494 if (pgdat->kswapd_max_order < order) { 3495 pgdat->kswapd_max_order = order; 3496 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx); 3497 } 3498 if (!waitqueue_active(&pgdat->kswapd_wait)) 3499 return; 3500 if (zone_balanced(zone, order, 0, 0)) 3501 return; 3502 3503 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order); 3504 wake_up_interruptible(&pgdat->kswapd_wait); 3505 } 3506 3507 #ifdef CONFIG_HIBERNATION 3508 /* 3509 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of 3510 * freed pages. 3511 * 3512 * Rather than trying to age LRUs the aim is to preserve the overall 3513 * LRU order by reclaiming preferentially 3514 * inactive > active > active referenced > active mapped 3515 */ 3516 unsigned long shrink_all_memory(unsigned long nr_to_reclaim) 3517 { 3518 struct reclaim_state reclaim_state; 3519 struct scan_control sc = { 3520 .nr_to_reclaim = nr_to_reclaim, 3521 .gfp_mask = GFP_HIGHUSER_MOVABLE, 3522 .priority = DEF_PRIORITY, 3523 .may_writepage = 1, 3524 .may_unmap = 1, 3525 .may_swap = 1, 3526 .hibernation_mode = 1, 3527 }; 3528 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); 3529 struct task_struct *p = current; 3530 unsigned long nr_reclaimed; 3531 3532 p->flags |= PF_MEMALLOC; 3533 lockdep_set_current_reclaim_state(sc.gfp_mask); 3534 reclaim_state.reclaimed_slab = 0; 3535 p->reclaim_state = &reclaim_state; 3536 3537 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 3538 3539 p->reclaim_state = NULL; 3540 lockdep_clear_current_reclaim_state(); 3541 p->flags &= ~PF_MEMALLOC; 3542 3543 return nr_reclaimed; 3544 } 3545 #endif /* CONFIG_HIBERNATION */ 3546 3547 /* It's optimal to keep kswapds on the same CPUs as their memory, but 3548 not required for correctness. So if the last cpu in a node goes 3549 away, we get changed to run anywhere: as the first one comes back, 3550 restore their cpu bindings. */ 3551 static int cpu_callback(struct notifier_block *nfb, unsigned long action, 3552 void *hcpu) 3553 { 3554 int nid; 3555 3556 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { 3557 for_each_node_state(nid, N_MEMORY) { 3558 pg_data_t *pgdat = NODE_DATA(nid); 3559 const struct cpumask *mask; 3560 3561 mask = cpumask_of_node(pgdat->node_id); 3562 3563 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) 3564 /* One of our CPUs online: restore mask */ 3565 set_cpus_allowed_ptr(pgdat->kswapd, mask); 3566 } 3567 } 3568 return NOTIFY_OK; 3569 } 3570 3571 /* 3572 * This kswapd start function will be called by init and node-hot-add. 3573 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. 3574 */ 3575 int kswapd_run(int nid) 3576 { 3577 pg_data_t *pgdat = NODE_DATA(nid); 3578 int ret = 0; 3579 3580 if (pgdat->kswapd) 3581 return 0; 3582 3583 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); 3584 if (IS_ERR(pgdat->kswapd)) { 3585 /* failure at boot is fatal */ 3586 BUG_ON(system_state == SYSTEM_BOOTING); 3587 pr_err("Failed to start kswapd on node %d\n", nid); 3588 ret = PTR_ERR(pgdat->kswapd); 3589 pgdat->kswapd = NULL; 3590 } 3591 return ret; 3592 } 3593 3594 /* 3595 * Called by memory hotplug when all memory in a node is offlined. Caller must 3596 * hold mem_hotplug_begin/end(). 3597 */ 3598 void kswapd_stop(int nid) 3599 { 3600 struct task_struct *kswapd = NODE_DATA(nid)->kswapd; 3601 3602 if (kswapd) { 3603 kthread_stop(kswapd); 3604 NODE_DATA(nid)->kswapd = NULL; 3605 } 3606 } 3607 3608 static int __init kswapd_init(void) 3609 { 3610 int nid; 3611 3612 swap_setup(); 3613 for_each_node_state(nid, N_MEMORY) 3614 kswapd_run(nid); 3615 hotcpu_notifier(cpu_callback, 0); 3616 return 0; 3617 } 3618 3619 module_init(kswapd_init) 3620 3621 #ifdef CONFIG_NUMA 3622 /* 3623 * Zone reclaim mode 3624 * 3625 * If non-zero call zone_reclaim when the number of free pages falls below 3626 * the watermarks. 3627 */ 3628 int zone_reclaim_mode __read_mostly; 3629 3630 #define RECLAIM_OFF 0 3631 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ 3632 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ 3633 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ 3634 3635 /* 3636 * Priority for ZONE_RECLAIM. This determines the fraction of pages 3637 * of a node considered for each zone_reclaim. 4 scans 1/16th of 3638 * a zone. 3639 */ 3640 #define ZONE_RECLAIM_PRIORITY 4 3641 3642 /* 3643 * Percentage of pages in a zone that must be unmapped for zone_reclaim to 3644 * occur. 3645 */ 3646 int sysctl_min_unmapped_ratio = 1; 3647 3648 /* 3649 * If the number of slab pages in a zone grows beyond this percentage then 3650 * slab reclaim needs to occur. 3651 */ 3652 int sysctl_min_slab_ratio = 5; 3653 3654 static inline unsigned long zone_unmapped_file_pages(struct zone *zone) 3655 { 3656 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED); 3657 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) + 3658 zone_page_state(zone, NR_ACTIVE_FILE); 3659 3660 /* 3661 * It's possible for there to be more file mapped pages than 3662 * accounted for by the pages on the file LRU lists because 3663 * tmpfs pages accounted for as ANON can also be FILE_MAPPED 3664 */ 3665 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; 3666 } 3667 3668 /* Work out how many page cache pages we can reclaim in this reclaim_mode */ 3669 static long zone_pagecache_reclaimable(struct zone *zone) 3670 { 3671 long nr_pagecache_reclaimable; 3672 long delta = 0; 3673 3674 /* 3675 * If RECLAIM_UNMAP is set, then all file pages are considered 3676 * potentially reclaimable. Otherwise, we have to worry about 3677 * pages like swapcache and zone_unmapped_file_pages() provides 3678 * a better estimate 3679 */ 3680 if (zone_reclaim_mode & RECLAIM_UNMAP) 3681 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES); 3682 else 3683 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone); 3684 3685 /* If we can't clean pages, remove dirty pages from consideration */ 3686 if (!(zone_reclaim_mode & RECLAIM_WRITE)) 3687 delta += zone_page_state(zone, NR_FILE_DIRTY); 3688 3689 /* Watch for any possible underflows due to delta */ 3690 if (unlikely(delta > nr_pagecache_reclaimable)) 3691 delta = nr_pagecache_reclaimable; 3692 3693 return nr_pagecache_reclaimable - delta; 3694 } 3695 3696 /* 3697 * Try to free up some pages from this zone through reclaim. 3698 */ 3699 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 3700 { 3701 /* Minimum pages needed in order to stay on node */ 3702 const unsigned long nr_pages = 1 << order; 3703 struct task_struct *p = current; 3704 struct reclaim_state reclaim_state; 3705 struct scan_control sc = { 3706 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), 3707 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)), 3708 .order = order, 3709 .priority = ZONE_RECLAIM_PRIORITY, 3710 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), 3711 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP), 3712 .may_swap = 1, 3713 }; 3714 3715 cond_resched(); 3716 /* 3717 * We need to be able to allocate from the reserves for RECLAIM_UNMAP 3718 * and we also need to be able to write out pages for RECLAIM_WRITE 3719 * and RECLAIM_UNMAP. 3720 */ 3721 p->flags |= PF_MEMALLOC | PF_SWAPWRITE; 3722 lockdep_set_current_reclaim_state(gfp_mask); 3723 reclaim_state.reclaimed_slab = 0; 3724 p->reclaim_state = &reclaim_state; 3725 3726 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) { 3727 /* 3728 * Free memory by calling shrink zone with increasing 3729 * priorities until we have enough memory freed. 3730 */ 3731 do { 3732 shrink_zone(zone, &sc, true); 3733 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); 3734 } 3735 3736 p->reclaim_state = NULL; 3737 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); 3738 lockdep_clear_current_reclaim_state(); 3739 return sc.nr_reclaimed >= nr_pages; 3740 } 3741 3742 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 3743 { 3744 int node_id; 3745 int ret; 3746 3747 /* 3748 * Zone reclaim reclaims unmapped file backed pages and 3749 * slab pages if we are over the defined limits. 3750 * 3751 * A small portion of unmapped file backed pages is needed for 3752 * file I/O otherwise pages read by file I/O will be immediately 3753 * thrown out if the zone is overallocated. So we do not reclaim 3754 * if less than a specified percentage of the zone is used by 3755 * unmapped file backed pages. 3756 */ 3757 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages && 3758 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages) 3759 return ZONE_RECLAIM_FULL; 3760 3761 if (!zone_reclaimable(zone)) 3762 return ZONE_RECLAIM_FULL; 3763 3764 /* 3765 * Do not scan if the allocation should not be delayed. 3766 */ 3767 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) 3768 return ZONE_RECLAIM_NOSCAN; 3769 3770 /* 3771 * Only run zone reclaim on the local zone or on zones that do not 3772 * have associated processors. This will favor the local processor 3773 * over remote processors and spread off node memory allocations 3774 * as wide as possible. 3775 */ 3776 node_id = zone_to_nid(zone); 3777 if (node_state(node_id, N_CPU) && node_id != numa_node_id()) 3778 return ZONE_RECLAIM_NOSCAN; 3779 3780 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags)) 3781 return ZONE_RECLAIM_NOSCAN; 3782 3783 ret = __zone_reclaim(zone, gfp_mask, order); 3784 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags); 3785 3786 if (!ret) 3787 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); 3788 3789 return ret; 3790 } 3791 #endif 3792 3793 /* 3794 * page_evictable - test whether a page is evictable 3795 * @page: the page to test 3796 * 3797 * Test whether page is evictable--i.e., should be placed on active/inactive 3798 * lists vs unevictable list. 3799 * 3800 * Reasons page might not be evictable: 3801 * (1) page's mapping marked unevictable 3802 * (2) page is part of an mlocked VMA 3803 * 3804 */ 3805 int page_evictable(struct page *page) 3806 { 3807 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); 3808 } 3809 3810 #ifdef CONFIG_SHMEM 3811 /** 3812 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list 3813 * @pages: array of pages to check 3814 * @nr_pages: number of pages to check 3815 * 3816 * Checks pages for evictability and moves them to the appropriate lru list. 3817 * 3818 * This function is only used for SysV IPC SHM_UNLOCK. 3819 */ 3820 void check_move_unevictable_pages(struct page **pages, int nr_pages) 3821 { 3822 struct lruvec *lruvec; 3823 struct zone *zone = NULL; 3824 int pgscanned = 0; 3825 int pgrescued = 0; 3826 int i; 3827 3828 for (i = 0; i < nr_pages; i++) { 3829 struct page *page = pages[i]; 3830 struct zone *pagezone; 3831 3832 pgscanned++; 3833 pagezone = page_zone(page); 3834 if (pagezone != zone) { 3835 if (zone) 3836 spin_unlock_irq(&zone->lru_lock); 3837 zone = pagezone; 3838 spin_lock_irq(&zone->lru_lock); 3839 } 3840 lruvec = mem_cgroup_page_lruvec(page, zone); 3841 3842 if (!PageLRU(page) || !PageUnevictable(page)) 3843 continue; 3844 3845 if (page_evictable(page)) { 3846 enum lru_list lru = page_lru_base_type(page); 3847 3848 VM_BUG_ON_PAGE(PageActive(page), page); 3849 ClearPageUnevictable(page); 3850 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); 3851 add_page_to_lru_list(page, lruvec, lru); 3852 pgrescued++; 3853 } 3854 } 3855 3856 if (zone) { 3857 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); 3858 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); 3859 spin_unlock_irq(&zone->lru_lock); 3860 } 3861 } 3862 #endif /* CONFIG_SHMEM */ 3863