1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * linux/mm/vmscan.c 4 * 5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 6 * 7 * Swap reorganised 29.12.95, Stephen Tweedie. 8 * kswapd added: 7.1.96 sct 9 * Removed kswapd_ctl limits, and swap out as many pages as needed 10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel. 11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). 12 * Multiqueue VM started 5.8.00, Rik van Riel. 13 */ 14 15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 16 17 #include <linux/mm.h> 18 #include <linux/sched/mm.h> 19 #include <linux/module.h> 20 #include <linux/gfp.h> 21 #include <linux/kernel_stat.h> 22 #include <linux/swap.h> 23 #include <linux/pagemap.h> 24 #include <linux/init.h> 25 #include <linux/highmem.h> 26 #include <linux/vmpressure.h> 27 #include <linux/vmstat.h> 28 #include <linux/file.h> 29 #include <linux/writeback.h> 30 #include <linux/blkdev.h> 31 #include <linux/buffer_head.h> /* for try_to_release_page(), 32 buffer_heads_over_limit */ 33 #include <linux/mm_inline.h> 34 #include <linux/backing-dev.h> 35 #include <linux/rmap.h> 36 #include <linux/topology.h> 37 #include <linux/cpu.h> 38 #include <linux/cpuset.h> 39 #include <linux/compaction.h> 40 #include <linux/notifier.h> 41 #include <linux/rwsem.h> 42 #include <linux/delay.h> 43 #include <linux/kthread.h> 44 #include <linux/freezer.h> 45 #include <linux/memcontrol.h> 46 #include <linux/delayacct.h> 47 #include <linux/sysctl.h> 48 #include <linux/oom.h> 49 #include <linux/prefetch.h> 50 #include <linux/printk.h> 51 #include <linux/dax.h> 52 53 #include <asm/tlbflush.h> 54 #include <asm/div64.h> 55 56 #include <linux/swapops.h> 57 #include <linux/balloon_compaction.h> 58 59 #include "internal.h" 60 61 #define CREATE_TRACE_POINTS 62 #include <trace/events/vmscan.h> 63 64 struct scan_control { 65 /* How many pages shrink_list() should reclaim */ 66 unsigned long nr_to_reclaim; 67 68 /* This context's GFP mask */ 69 gfp_t gfp_mask; 70 71 /* Allocation order */ 72 int order; 73 74 /* 75 * Nodemask of nodes allowed by the caller. If NULL, all nodes 76 * are scanned. 77 */ 78 nodemask_t *nodemask; 79 80 /* 81 * The memory cgroup that hit its limit and as a result is the 82 * primary target of this reclaim invocation. 83 */ 84 struct mem_cgroup *target_mem_cgroup; 85 86 /* Scan (total_size >> priority) pages at once */ 87 int priority; 88 89 /* The highest zone to isolate pages for reclaim from */ 90 enum zone_type reclaim_idx; 91 92 /* Writepage batching in laptop mode; RECLAIM_WRITE */ 93 unsigned int may_writepage:1; 94 95 /* Can mapped pages be reclaimed? */ 96 unsigned int may_unmap:1; 97 98 /* Can pages be swapped as part of reclaim? */ 99 unsigned int may_swap:1; 100 101 /* 102 * Cgroups are not reclaimed below their configured memory.low, 103 * unless we threaten to OOM. If any cgroups are skipped due to 104 * memory.low and nothing was reclaimed, go back for memory.low. 105 */ 106 unsigned int memcg_low_reclaim:1; 107 unsigned int memcg_low_skipped:1; 108 109 unsigned int hibernation_mode:1; 110 111 /* One of the zones is ready for compaction */ 112 unsigned int compaction_ready:1; 113 114 /* Incremented by the number of inactive pages that were scanned */ 115 unsigned long nr_scanned; 116 117 /* Number of pages freed so far during a call to shrink_zones() */ 118 unsigned long nr_reclaimed; 119 }; 120 121 #ifdef ARCH_HAS_PREFETCH 122 #define prefetch_prev_lru_page(_page, _base, _field) \ 123 do { \ 124 if ((_page)->lru.prev != _base) { \ 125 struct page *prev; \ 126 \ 127 prev = lru_to_page(&(_page->lru)); \ 128 prefetch(&prev->_field); \ 129 } \ 130 } while (0) 131 #else 132 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) 133 #endif 134 135 #ifdef ARCH_HAS_PREFETCHW 136 #define prefetchw_prev_lru_page(_page, _base, _field) \ 137 do { \ 138 if ((_page)->lru.prev != _base) { \ 139 struct page *prev; \ 140 \ 141 prev = lru_to_page(&(_page->lru)); \ 142 prefetchw(&prev->_field); \ 143 } \ 144 } while (0) 145 #else 146 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) 147 #endif 148 149 /* 150 * From 0 .. 100. Higher means more swappy. 151 */ 152 int vm_swappiness = 60; 153 /* 154 * The total number of pages which are beyond the high watermark within all 155 * zones. 156 */ 157 unsigned long vm_total_pages; 158 159 static LIST_HEAD(shrinker_list); 160 static DECLARE_RWSEM(shrinker_rwsem); 161 162 #ifdef CONFIG_MEMCG 163 static bool global_reclaim(struct scan_control *sc) 164 { 165 return !sc->target_mem_cgroup; 166 } 167 168 /** 169 * sane_reclaim - is the usual dirty throttling mechanism operational? 170 * @sc: scan_control in question 171 * 172 * The normal page dirty throttling mechanism in balance_dirty_pages() is 173 * completely broken with the legacy memcg and direct stalling in 174 * shrink_page_list() is used for throttling instead, which lacks all the 175 * niceties such as fairness, adaptive pausing, bandwidth proportional 176 * allocation and configurability. 177 * 178 * This function tests whether the vmscan currently in progress can assume 179 * that the normal dirty throttling mechanism is operational. 180 */ 181 static bool sane_reclaim(struct scan_control *sc) 182 { 183 struct mem_cgroup *memcg = sc->target_mem_cgroup; 184 185 if (!memcg) 186 return true; 187 #ifdef CONFIG_CGROUP_WRITEBACK 188 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) 189 return true; 190 #endif 191 return false; 192 } 193 #else 194 static bool global_reclaim(struct scan_control *sc) 195 { 196 return true; 197 } 198 199 static bool sane_reclaim(struct scan_control *sc) 200 { 201 return true; 202 } 203 #endif 204 205 /* 206 * This misses isolated pages which are not accounted for to save counters. 207 * As the data only determines if reclaim or compaction continues, it is 208 * not expected that isolated pages will be a dominating factor. 209 */ 210 unsigned long zone_reclaimable_pages(struct zone *zone) 211 { 212 unsigned long nr; 213 214 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + 215 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); 216 if (get_nr_swap_pages() > 0) 217 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + 218 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); 219 220 return nr; 221 } 222 223 /** 224 * lruvec_lru_size - Returns the number of pages on the given LRU list. 225 * @lruvec: lru vector 226 * @lru: lru to use 227 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) 228 */ 229 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx) 230 { 231 unsigned long lru_size; 232 int zid; 233 234 if (!mem_cgroup_disabled()) 235 lru_size = mem_cgroup_get_lru_size(lruvec, lru); 236 else 237 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru); 238 239 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) { 240 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; 241 unsigned long size; 242 243 if (!managed_zone(zone)) 244 continue; 245 246 if (!mem_cgroup_disabled()) 247 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid); 248 else 249 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid], 250 NR_ZONE_LRU_BASE + lru); 251 lru_size -= min(size, lru_size); 252 } 253 254 return lru_size; 255 256 } 257 258 /* 259 * Add a shrinker callback to be called from the vm. 260 */ 261 int register_shrinker(struct shrinker *shrinker) 262 { 263 size_t size = sizeof(*shrinker->nr_deferred); 264 265 if (shrinker->flags & SHRINKER_NUMA_AWARE) 266 size *= nr_node_ids; 267 268 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); 269 if (!shrinker->nr_deferred) 270 return -ENOMEM; 271 272 down_write(&shrinker_rwsem); 273 list_add_tail(&shrinker->list, &shrinker_list); 274 up_write(&shrinker_rwsem); 275 return 0; 276 } 277 EXPORT_SYMBOL(register_shrinker); 278 279 /* 280 * Remove one 281 */ 282 void unregister_shrinker(struct shrinker *shrinker) 283 { 284 if (!shrinker->nr_deferred) 285 return; 286 down_write(&shrinker_rwsem); 287 list_del(&shrinker->list); 288 up_write(&shrinker_rwsem); 289 kfree(shrinker->nr_deferred); 290 shrinker->nr_deferred = NULL; 291 } 292 EXPORT_SYMBOL(unregister_shrinker); 293 294 #define SHRINK_BATCH 128 295 296 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, 297 struct shrinker *shrinker, int priority) 298 { 299 unsigned long freed = 0; 300 unsigned long long delta; 301 long total_scan; 302 long freeable; 303 long nr; 304 long new_nr; 305 int nid = shrinkctl->nid; 306 long batch_size = shrinker->batch ? shrinker->batch 307 : SHRINK_BATCH; 308 long scanned = 0, next_deferred; 309 310 freeable = shrinker->count_objects(shrinker, shrinkctl); 311 if (freeable == 0) 312 return 0; 313 314 /* 315 * copy the current shrinker scan count into a local variable 316 * and zero it so that other concurrent shrinker invocations 317 * don't also do this scanning work. 318 */ 319 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); 320 321 total_scan = nr; 322 delta = freeable >> priority; 323 delta *= 4; 324 do_div(delta, shrinker->seeks); 325 total_scan += delta; 326 if (total_scan < 0) { 327 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", 328 shrinker->scan_objects, total_scan); 329 total_scan = freeable; 330 next_deferred = nr; 331 } else 332 next_deferred = total_scan; 333 334 /* 335 * We need to avoid excessive windup on filesystem shrinkers 336 * due to large numbers of GFP_NOFS allocations causing the 337 * shrinkers to return -1 all the time. This results in a large 338 * nr being built up so when a shrink that can do some work 339 * comes along it empties the entire cache due to nr >>> 340 * freeable. This is bad for sustaining a working set in 341 * memory. 342 * 343 * Hence only allow the shrinker to scan the entire cache when 344 * a large delta change is calculated directly. 345 */ 346 if (delta < freeable / 4) 347 total_scan = min(total_scan, freeable / 2); 348 349 /* 350 * Avoid risking looping forever due to too large nr value: 351 * never try to free more than twice the estimate number of 352 * freeable entries. 353 */ 354 if (total_scan > freeable * 2) 355 total_scan = freeable * 2; 356 357 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, 358 freeable, delta, total_scan, priority); 359 360 /* 361 * Normally, we should not scan less than batch_size objects in one 362 * pass to avoid too frequent shrinker calls, but if the slab has less 363 * than batch_size objects in total and we are really tight on memory, 364 * we will try to reclaim all available objects, otherwise we can end 365 * up failing allocations although there are plenty of reclaimable 366 * objects spread over several slabs with usage less than the 367 * batch_size. 368 * 369 * We detect the "tight on memory" situations by looking at the total 370 * number of objects we want to scan (total_scan). If it is greater 371 * than the total number of objects on slab (freeable), we must be 372 * scanning at high prio and therefore should try to reclaim as much as 373 * possible. 374 */ 375 while (total_scan >= batch_size || 376 total_scan >= freeable) { 377 unsigned long ret; 378 unsigned long nr_to_scan = min(batch_size, total_scan); 379 380 shrinkctl->nr_to_scan = nr_to_scan; 381 shrinkctl->nr_scanned = nr_to_scan; 382 ret = shrinker->scan_objects(shrinker, shrinkctl); 383 if (ret == SHRINK_STOP) 384 break; 385 freed += ret; 386 387 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned); 388 total_scan -= shrinkctl->nr_scanned; 389 scanned += shrinkctl->nr_scanned; 390 391 cond_resched(); 392 } 393 394 if (next_deferred >= scanned) 395 next_deferred -= scanned; 396 else 397 next_deferred = 0; 398 /* 399 * move the unused scan count back into the shrinker in a 400 * manner that handles concurrent updates. If we exhausted the 401 * scan, there is no need to do an update. 402 */ 403 if (next_deferred > 0) 404 new_nr = atomic_long_add_return(next_deferred, 405 &shrinker->nr_deferred[nid]); 406 else 407 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); 408 409 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); 410 return freed; 411 } 412 413 /** 414 * shrink_slab - shrink slab caches 415 * @gfp_mask: allocation context 416 * @nid: node whose slab caches to target 417 * @memcg: memory cgroup whose slab caches to target 418 * @priority: the reclaim priority 419 * 420 * Call the shrink functions to age shrinkable caches. 421 * 422 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, 423 * unaware shrinkers will receive a node id of 0 instead. 424 * 425 * @memcg specifies the memory cgroup to target. If it is not NULL, 426 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan 427 * objects from the memory cgroup specified. Otherwise, only unaware 428 * shrinkers are called. 429 * 430 * @priority is sc->priority, we take the number of objects and >> by priority 431 * in order to get the scan target. 432 * 433 * Returns the number of reclaimed slab objects. 434 */ 435 static unsigned long shrink_slab(gfp_t gfp_mask, int nid, 436 struct mem_cgroup *memcg, 437 int priority) 438 { 439 struct shrinker *shrinker; 440 unsigned long freed = 0; 441 442 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg))) 443 return 0; 444 445 if (!down_read_trylock(&shrinker_rwsem)) { 446 /* 447 * If we would return 0, our callers would understand that we 448 * have nothing else to shrink and give up trying. By returning 449 * 1 we keep it going and assume we'll be able to shrink next 450 * time. 451 */ 452 freed = 1; 453 goto out; 454 } 455 456 list_for_each_entry(shrinker, &shrinker_list, list) { 457 struct shrink_control sc = { 458 .gfp_mask = gfp_mask, 459 .nid = nid, 460 .memcg = memcg, 461 }; 462 463 /* 464 * If kernel memory accounting is disabled, we ignore 465 * SHRINKER_MEMCG_AWARE flag and call all shrinkers 466 * passing NULL for memcg. 467 */ 468 if (memcg_kmem_enabled() && 469 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE)) 470 continue; 471 472 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) 473 sc.nid = 0; 474 475 freed += do_shrink_slab(&sc, shrinker, priority); 476 /* 477 * Bail out if someone want to register a new shrinker to 478 * prevent the regsitration from being stalled for long periods 479 * by parallel ongoing shrinking. 480 */ 481 if (rwsem_is_contended(&shrinker_rwsem)) { 482 freed = freed ? : 1; 483 break; 484 } 485 } 486 487 up_read(&shrinker_rwsem); 488 out: 489 cond_resched(); 490 return freed; 491 } 492 493 void drop_slab_node(int nid) 494 { 495 unsigned long freed; 496 497 do { 498 struct mem_cgroup *memcg = NULL; 499 500 freed = 0; 501 do { 502 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0); 503 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); 504 } while (freed > 10); 505 } 506 507 void drop_slab(void) 508 { 509 int nid; 510 511 for_each_online_node(nid) 512 drop_slab_node(nid); 513 } 514 515 static inline int is_page_cache_freeable(struct page *page) 516 { 517 /* 518 * A freeable page cache page is referenced only by the caller 519 * that isolated the page, the page cache radix tree and 520 * optional buffer heads at page->private. 521 */ 522 int radix_pins = PageTransHuge(page) && PageSwapCache(page) ? 523 HPAGE_PMD_NR : 1; 524 return page_count(page) - page_has_private(page) == 1 + radix_pins; 525 } 526 527 static int may_write_to_inode(struct inode *inode, struct scan_control *sc) 528 { 529 if (current->flags & PF_SWAPWRITE) 530 return 1; 531 if (!inode_write_congested(inode)) 532 return 1; 533 if (inode_to_bdi(inode) == current->backing_dev_info) 534 return 1; 535 return 0; 536 } 537 538 /* 539 * We detected a synchronous write error writing a page out. Probably 540 * -ENOSPC. We need to propagate that into the address_space for a subsequent 541 * fsync(), msync() or close(). 542 * 543 * The tricky part is that after writepage we cannot touch the mapping: nothing 544 * prevents it from being freed up. But we have a ref on the page and once 545 * that page is locked, the mapping is pinned. 546 * 547 * We're allowed to run sleeping lock_page() here because we know the caller has 548 * __GFP_FS. 549 */ 550 static void handle_write_error(struct address_space *mapping, 551 struct page *page, int error) 552 { 553 lock_page(page); 554 if (page_mapping(page) == mapping) 555 mapping_set_error(mapping, error); 556 unlock_page(page); 557 } 558 559 /* possible outcome of pageout() */ 560 typedef enum { 561 /* failed to write page out, page is locked */ 562 PAGE_KEEP, 563 /* move page to the active list, page is locked */ 564 PAGE_ACTIVATE, 565 /* page has been sent to the disk successfully, page is unlocked */ 566 PAGE_SUCCESS, 567 /* page is clean and locked */ 568 PAGE_CLEAN, 569 } pageout_t; 570 571 /* 572 * pageout is called by shrink_page_list() for each dirty page. 573 * Calls ->writepage(). 574 */ 575 static pageout_t pageout(struct page *page, struct address_space *mapping, 576 struct scan_control *sc) 577 { 578 /* 579 * If the page is dirty, only perform writeback if that write 580 * will be non-blocking. To prevent this allocation from being 581 * stalled by pagecache activity. But note that there may be 582 * stalls if we need to run get_block(). We could test 583 * PagePrivate for that. 584 * 585 * If this process is currently in __generic_file_write_iter() against 586 * this page's queue, we can perform writeback even if that 587 * will block. 588 * 589 * If the page is swapcache, write it back even if that would 590 * block, for some throttling. This happens by accident, because 591 * swap_backing_dev_info is bust: it doesn't reflect the 592 * congestion state of the swapdevs. Easy to fix, if needed. 593 */ 594 if (!is_page_cache_freeable(page)) 595 return PAGE_KEEP; 596 if (!mapping) { 597 /* 598 * Some data journaling orphaned pages can have 599 * page->mapping == NULL while being dirty with clean buffers. 600 */ 601 if (page_has_private(page)) { 602 if (try_to_free_buffers(page)) { 603 ClearPageDirty(page); 604 pr_info("%s: orphaned page\n", __func__); 605 return PAGE_CLEAN; 606 } 607 } 608 return PAGE_KEEP; 609 } 610 if (mapping->a_ops->writepage == NULL) 611 return PAGE_ACTIVATE; 612 if (!may_write_to_inode(mapping->host, sc)) 613 return PAGE_KEEP; 614 615 if (clear_page_dirty_for_io(page)) { 616 int res; 617 struct writeback_control wbc = { 618 .sync_mode = WB_SYNC_NONE, 619 .nr_to_write = SWAP_CLUSTER_MAX, 620 .range_start = 0, 621 .range_end = LLONG_MAX, 622 .for_reclaim = 1, 623 }; 624 625 SetPageReclaim(page); 626 res = mapping->a_ops->writepage(page, &wbc); 627 if (res < 0) 628 handle_write_error(mapping, page, res); 629 if (res == AOP_WRITEPAGE_ACTIVATE) { 630 ClearPageReclaim(page); 631 return PAGE_ACTIVATE; 632 } 633 634 if (!PageWriteback(page)) { 635 /* synchronous write or broken a_ops? */ 636 ClearPageReclaim(page); 637 } 638 trace_mm_vmscan_writepage(page); 639 inc_node_page_state(page, NR_VMSCAN_WRITE); 640 return PAGE_SUCCESS; 641 } 642 643 return PAGE_CLEAN; 644 } 645 646 /* 647 * Same as remove_mapping, but if the page is removed from the mapping, it 648 * gets returned with a refcount of 0. 649 */ 650 static int __remove_mapping(struct address_space *mapping, struct page *page, 651 bool reclaimed) 652 { 653 unsigned long flags; 654 int refcount; 655 656 BUG_ON(!PageLocked(page)); 657 BUG_ON(mapping != page_mapping(page)); 658 659 spin_lock_irqsave(&mapping->tree_lock, flags); 660 /* 661 * The non racy check for a busy page. 662 * 663 * Must be careful with the order of the tests. When someone has 664 * a ref to the page, it may be possible that they dirty it then 665 * drop the reference. So if PageDirty is tested before page_count 666 * here, then the following race may occur: 667 * 668 * get_user_pages(&page); 669 * [user mapping goes away] 670 * write_to(page); 671 * !PageDirty(page) [good] 672 * SetPageDirty(page); 673 * put_page(page); 674 * !page_count(page) [good, discard it] 675 * 676 * [oops, our write_to data is lost] 677 * 678 * Reversing the order of the tests ensures such a situation cannot 679 * escape unnoticed. The smp_rmb is needed to ensure the page->flags 680 * load is not satisfied before that of page->_refcount. 681 * 682 * Note that if SetPageDirty is always performed via set_page_dirty, 683 * and thus under tree_lock, then this ordering is not required. 684 */ 685 if (unlikely(PageTransHuge(page)) && PageSwapCache(page)) 686 refcount = 1 + HPAGE_PMD_NR; 687 else 688 refcount = 2; 689 if (!page_ref_freeze(page, refcount)) 690 goto cannot_free; 691 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ 692 if (unlikely(PageDirty(page))) { 693 page_ref_unfreeze(page, refcount); 694 goto cannot_free; 695 } 696 697 if (PageSwapCache(page)) { 698 swp_entry_t swap = { .val = page_private(page) }; 699 mem_cgroup_swapout(page, swap); 700 __delete_from_swap_cache(page); 701 spin_unlock_irqrestore(&mapping->tree_lock, flags); 702 put_swap_page(page, swap); 703 } else { 704 void (*freepage)(struct page *); 705 void *shadow = NULL; 706 707 freepage = mapping->a_ops->freepage; 708 /* 709 * Remember a shadow entry for reclaimed file cache in 710 * order to detect refaults, thus thrashing, later on. 711 * 712 * But don't store shadows in an address space that is 713 * already exiting. This is not just an optizimation, 714 * inode reclaim needs to empty out the radix tree or 715 * the nodes are lost. Don't plant shadows behind its 716 * back. 717 * 718 * We also don't store shadows for DAX mappings because the 719 * only page cache pages found in these are zero pages 720 * covering holes, and because we don't want to mix DAX 721 * exceptional entries and shadow exceptional entries in the 722 * same page_tree. 723 */ 724 if (reclaimed && page_is_file_cache(page) && 725 !mapping_exiting(mapping) && !dax_mapping(mapping)) 726 shadow = workingset_eviction(mapping, page); 727 __delete_from_page_cache(page, shadow); 728 spin_unlock_irqrestore(&mapping->tree_lock, flags); 729 730 if (freepage != NULL) 731 freepage(page); 732 } 733 734 return 1; 735 736 cannot_free: 737 spin_unlock_irqrestore(&mapping->tree_lock, flags); 738 return 0; 739 } 740 741 /* 742 * Attempt to detach a locked page from its ->mapping. If it is dirty or if 743 * someone else has a ref on the page, abort and return 0. If it was 744 * successfully detached, return 1. Assumes the caller has a single ref on 745 * this page. 746 */ 747 int remove_mapping(struct address_space *mapping, struct page *page) 748 { 749 if (__remove_mapping(mapping, page, false)) { 750 /* 751 * Unfreezing the refcount with 1 rather than 2 effectively 752 * drops the pagecache ref for us without requiring another 753 * atomic operation. 754 */ 755 page_ref_unfreeze(page, 1); 756 return 1; 757 } 758 return 0; 759 } 760 761 /** 762 * putback_lru_page - put previously isolated page onto appropriate LRU list 763 * @page: page to be put back to appropriate lru list 764 * 765 * Add previously isolated @page to appropriate LRU list. 766 * Page may still be unevictable for other reasons. 767 * 768 * lru_lock must not be held, interrupts must be enabled. 769 */ 770 void putback_lru_page(struct page *page) 771 { 772 lru_cache_add(page); 773 put_page(page); /* drop ref from isolate */ 774 } 775 776 enum page_references { 777 PAGEREF_RECLAIM, 778 PAGEREF_RECLAIM_CLEAN, 779 PAGEREF_KEEP, 780 PAGEREF_ACTIVATE, 781 }; 782 783 static enum page_references page_check_references(struct page *page, 784 struct scan_control *sc) 785 { 786 int referenced_ptes, referenced_page; 787 unsigned long vm_flags; 788 789 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, 790 &vm_flags); 791 referenced_page = TestClearPageReferenced(page); 792 793 /* 794 * Mlock lost the isolation race with us. Let try_to_unmap() 795 * move the page to the unevictable list. 796 */ 797 if (vm_flags & VM_LOCKED) 798 return PAGEREF_RECLAIM; 799 800 if (referenced_ptes) { 801 if (PageSwapBacked(page)) 802 return PAGEREF_ACTIVATE; 803 /* 804 * All mapped pages start out with page table 805 * references from the instantiating fault, so we need 806 * to look twice if a mapped file page is used more 807 * than once. 808 * 809 * Mark it and spare it for another trip around the 810 * inactive list. Another page table reference will 811 * lead to its activation. 812 * 813 * Note: the mark is set for activated pages as well 814 * so that recently deactivated but used pages are 815 * quickly recovered. 816 */ 817 SetPageReferenced(page); 818 819 if (referenced_page || referenced_ptes > 1) 820 return PAGEREF_ACTIVATE; 821 822 /* 823 * Activate file-backed executable pages after first usage. 824 */ 825 if (vm_flags & VM_EXEC) 826 return PAGEREF_ACTIVATE; 827 828 return PAGEREF_KEEP; 829 } 830 831 /* Reclaim if clean, defer dirty pages to writeback */ 832 if (referenced_page && !PageSwapBacked(page)) 833 return PAGEREF_RECLAIM_CLEAN; 834 835 return PAGEREF_RECLAIM; 836 } 837 838 /* Check if a page is dirty or under writeback */ 839 static void page_check_dirty_writeback(struct page *page, 840 bool *dirty, bool *writeback) 841 { 842 struct address_space *mapping; 843 844 /* 845 * Anonymous pages are not handled by flushers and must be written 846 * from reclaim context. Do not stall reclaim based on them 847 */ 848 if (!page_is_file_cache(page) || 849 (PageAnon(page) && !PageSwapBacked(page))) { 850 *dirty = false; 851 *writeback = false; 852 return; 853 } 854 855 /* By default assume that the page flags are accurate */ 856 *dirty = PageDirty(page); 857 *writeback = PageWriteback(page); 858 859 /* Verify dirty/writeback state if the filesystem supports it */ 860 if (!page_has_private(page)) 861 return; 862 863 mapping = page_mapping(page); 864 if (mapping && mapping->a_ops->is_dirty_writeback) 865 mapping->a_ops->is_dirty_writeback(page, dirty, writeback); 866 } 867 868 struct reclaim_stat { 869 unsigned nr_dirty; 870 unsigned nr_unqueued_dirty; 871 unsigned nr_congested; 872 unsigned nr_writeback; 873 unsigned nr_immediate; 874 unsigned nr_activate; 875 unsigned nr_ref_keep; 876 unsigned nr_unmap_fail; 877 }; 878 879 /* 880 * shrink_page_list() returns the number of reclaimed pages 881 */ 882 static unsigned long shrink_page_list(struct list_head *page_list, 883 struct pglist_data *pgdat, 884 struct scan_control *sc, 885 enum ttu_flags ttu_flags, 886 struct reclaim_stat *stat, 887 bool force_reclaim) 888 { 889 LIST_HEAD(ret_pages); 890 LIST_HEAD(free_pages); 891 int pgactivate = 0; 892 unsigned nr_unqueued_dirty = 0; 893 unsigned nr_dirty = 0; 894 unsigned nr_congested = 0; 895 unsigned nr_reclaimed = 0; 896 unsigned nr_writeback = 0; 897 unsigned nr_immediate = 0; 898 unsigned nr_ref_keep = 0; 899 unsigned nr_unmap_fail = 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); 919 920 sc->nr_scanned++; 921 922 if (unlikely(!page_evictable(page))) 923 goto activate_locked; 924 925 if (!sc->may_unmap && page_mapped(page)) 926 goto keep_locked; 927 928 /* Double the slab pressure for mapped and swapcache pages */ 929 if ((page_mapped(page) || PageSwapCache(page)) && 930 !(PageAnon(page) && !PageSwapBacked(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 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 * In cases 1) and 2) we activate the pages to get them out of 995 * the way while we continue scanning for clean pages on the 996 * inactive list and refilling from the active list. The 997 * observation here is that waiting for disk writes is more 998 * expensive than potentially causing reloads down the line. 999 * Since they're marked for immediate reclaim, they won't put 1000 * memory pressure on the cache working set any longer than it 1001 * takes to write them to disk. 1002 */ 1003 if (PageWriteback(page)) { 1004 /* Case 1 above */ 1005 if (current_is_kswapd() && 1006 PageReclaim(page) && 1007 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { 1008 nr_immediate++; 1009 goto activate_locked; 1010 1011 /* Case 2 above */ 1012 } else if (sane_reclaim(sc) || 1013 !PageReclaim(page) || !may_enter_fs) { 1014 /* 1015 * This is slightly racy - end_page_writeback() 1016 * might have just cleared PageReclaim, then 1017 * setting PageReclaim here end up interpreted 1018 * as PageReadahead - but that does not matter 1019 * enough to care. What we do want is for this 1020 * page to have PageReclaim set next time memcg 1021 * reclaim reaches the tests above, so it will 1022 * then wait_on_page_writeback() to avoid OOM; 1023 * and it's also appropriate in global reclaim. 1024 */ 1025 SetPageReclaim(page); 1026 nr_writeback++; 1027 goto activate_locked; 1028 1029 /* Case 3 above */ 1030 } else { 1031 unlock_page(page); 1032 wait_on_page_writeback(page); 1033 /* then go back and try same page again */ 1034 list_add_tail(&page->lru, page_list); 1035 continue; 1036 } 1037 } 1038 1039 if (!force_reclaim) 1040 references = page_check_references(page, sc); 1041 1042 switch (references) { 1043 case PAGEREF_ACTIVATE: 1044 goto activate_locked; 1045 case PAGEREF_KEEP: 1046 nr_ref_keep++; 1047 goto keep_locked; 1048 case PAGEREF_RECLAIM: 1049 case PAGEREF_RECLAIM_CLEAN: 1050 ; /* try to reclaim the page below */ 1051 } 1052 1053 /* 1054 * Anonymous process memory has backing store? 1055 * Try to allocate it some swap space here. 1056 * Lazyfree page could be freed directly 1057 */ 1058 if (PageAnon(page) && PageSwapBacked(page)) { 1059 if (!PageSwapCache(page)) { 1060 if (!(sc->gfp_mask & __GFP_IO)) 1061 goto keep_locked; 1062 if (PageTransHuge(page)) { 1063 /* cannot split THP, skip it */ 1064 if (!can_split_huge_page(page, NULL)) 1065 goto activate_locked; 1066 /* 1067 * Split pages without a PMD map right 1068 * away. Chances are some or all of the 1069 * tail pages can be freed without IO. 1070 */ 1071 if (!compound_mapcount(page) && 1072 split_huge_page_to_list(page, 1073 page_list)) 1074 goto activate_locked; 1075 } 1076 if (!add_to_swap(page)) { 1077 if (!PageTransHuge(page)) 1078 goto activate_locked; 1079 /* Fallback to swap normal pages */ 1080 if (split_huge_page_to_list(page, 1081 page_list)) 1082 goto activate_locked; 1083 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 1084 count_vm_event(THP_SWPOUT_FALLBACK); 1085 #endif 1086 if (!add_to_swap(page)) 1087 goto activate_locked; 1088 } 1089 1090 may_enter_fs = 1; 1091 1092 /* Adding to swap updated mapping */ 1093 mapping = page_mapping(page); 1094 } 1095 } else if (unlikely(PageTransHuge(page))) { 1096 /* Split file THP */ 1097 if (split_huge_page_to_list(page, page_list)) 1098 goto keep_locked; 1099 } 1100 1101 /* 1102 * The page is mapped into the page tables of one or more 1103 * processes. Try to unmap it here. 1104 */ 1105 if (page_mapped(page)) { 1106 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH; 1107 1108 if (unlikely(PageTransHuge(page))) 1109 flags |= TTU_SPLIT_HUGE_PMD; 1110 if (!try_to_unmap(page, flags)) { 1111 nr_unmap_fail++; 1112 goto activate_locked; 1113 } 1114 } 1115 1116 if (PageDirty(page)) { 1117 /* 1118 * Only kswapd can writeback filesystem pages 1119 * to avoid risk of stack overflow. But avoid 1120 * injecting inefficient single-page IO into 1121 * flusher writeback as much as possible: only 1122 * write pages when we've encountered many 1123 * dirty pages, and when we've already scanned 1124 * the rest of the LRU for clean pages and see 1125 * the same dirty pages again (PageReclaim). 1126 */ 1127 if (page_is_file_cache(page) && 1128 (!current_is_kswapd() || !PageReclaim(page) || 1129 !test_bit(PGDAT_DIRTY, &pgdat->flags))) { 1130 /* 1131 * Immediately reclaim when written back. 1132 * Similar in principal to deactivate_page() 1133 * except we already have the page isolated 1134 * and know it's dirty 1135 */ 1136 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); 1137 SetPageReclaim(page); 1138 1139 goto activate_locked; 1140 } 1141 1142 if (references == PAGEREF_RECLAIM_CLEAN) 1143 goto keep_locked; 1144 if (!may_enter_fs) 1145 goto keep_locked; 1146 if (!sc->may_writepage) 1147 goto keep_locked; 1148 1149 /* 1150 * Page is dirty. Flush the TLB if a writable entry 1151 * potentially exists to avoid CPU writes after IO 1152 * starts and then write it out here. 1153 */ 1154 try_to_unmap_flush_dirty(); 1155 switch (pageout(page, mapping, sc)) { 1156 case PAGE_KEEP: 1157 goto keep_locked; 1158 case PAGE_ACTIVATE: 1159 goto activate_locked; 1160 case PAGE_SUCCESS: 1161 if (PageWriteback(page)) 1162 goto keep; 1163 if (PageDirty(page)) 1164 goto keep; 1165 1166 /* 1167 * A synchronous write - probably a ramdisk. Go 1168 * ahead and try to reclaim the page. 1169 */ 1170 if (!trylock_page(page)) 1171 goto keep; 1172 if (PageDirty(page) || PageWriteback(page)) 1173 goto keep_locked; 1174 mapping = page_mapping(page); 1175 case PAGE_CLEAN: 1176 ; /* try to free the page below */ 1177 } 1178 } 1179 1180 /* 1181 * If the page has buffers, try to free the buffer mappings 1182 * associated with this page. If we succeed we try to free 1183 * the page as well. 1184 * 1185 * We do this even if the page is PageDirty(). 1186 * try_to_release_page() does not perform I/O, but it is 1187 * possible for a page to have PageDirty set, but it is actually 1188 * clean (all its buffers are clean). This happens if the 1189 * buffers were written out directly, with submit_bh(). ext3 1190 * will do this, as well as the blockdev mapping. 1191 * try_to_release_page() will discover that cleanness and will 1192 * drop the buffers and mark the page clean - it can be freed. 1193 * 1194 * Rarely, pages can have buffers and no ->mapping. These are 1195 * the pages which were not successfully invalidated in 1196 * truncate_complete_page(). We try to drop those buffers here 1197 * and if that worked, and the page is no longer mapped into 1198 * process address space (page_count == 1) it can be freed. 1199 * Otherwise, leave the page on the LRU so it is swappable. 1200 */ 1201 if (page_has_private(page)) { 1202 if (!try_to_release_page(page, sc->gfp_mask)) 1203 goto activate_locked; 1204 if (!mapping && page_count(page) == 1) { 1205 unlock_page(page); 1206 if (put_page_testzero(page)) 1207 goto free_it; 1208 else { 1209 /* 1210 * rare race with speculative reference. 1211 * the speculative reference will free 1212 * this page shortly, so we may 1213 * increment nr_reclaimed here (and 1214 * leave it off the LRU). 1215 */ 1216 nr_reclaimed++; 1217 continue; 1218 } 1219 } 1220 } 1221 1222 if (PageAnon(page) && !PageSwapBacked(page)) { 1223 /* follow __remove_mapping for reference */ 1224 if (!page_ref_freeze(page, 1)) 1225 goto keep_locked; 1226 if (PageDirty(page)) { 1227 page_ref_unfreeze(page, 1); 1228 goto keep_locked; 1229 } 1230 1231 count_vm_event(PGLAZYFREED); 1232 count_memcg_page_event(page, PGLAZYFREED); 1233 } else if (!mapping || !__remove_mapping(mapping, page, true)) 1234 goto keep_locked; 1235 /* 1236 * At this point, we have no other references and there is 1237 * no way to pick any more up (removed from LRU, removed 1238 * from pagecache). Can use non-atomic bitops now (and 1239 * we obviously don't have to worry about waking up a process 1240 * waiting on the page lock, because there are no references. 1241 */ 1242 __ClearPageLocked(page); 1243 free_it: 1244 nr_reclaimed++; 1245 1246 /* 1247 * Is there need to periodically free_page_list? It would 1248 * appear not as the counts should be low 1249 */ 1250 if (unlikely(PageTransHuge(page))) { 1251 mem_cgroup_uncharge(page); 1252 (*get_compound_page_dtor(page))(page); 1253 } else 1254 list_add(&page->lru, &free_pages); 1255 continue; 1256 1257 activate_locked: 1258 /* Not a candidate for swapping, so reclaim swap space. */ 1259 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) || 1260 PageMlocked(page))) 1261 try_to_free_swap(page); 1262 VM_BUG_ON_PAGE(PageActive(page), page); 1263 if (!PageMlocked(page)) { 1264 SetPageActive(page); 1265 pgactivate++; 1266 count_memcg_page_event(page, PGACTIVATE); 1267 } 1268 keep_locked: 1269 unlock_page(page); 1270 keep: 1271 list_add(&page->lru, &ret_pages); 1272 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); 1273 } 1274 1275 mem_cgroup_uncharge_list(&free_pages); 1276 try_to_unmap_flush(); 1277 free_unref_page_list(&free_pages); 1278 1279 list_splice(&ret_pages, page_list); 1280 count_vm_events(PGACTIVATE, pgactivate); 1281 1282 if (stat) { 1283 stat->nr_dirty = nr_dirty; 1284 stat->nr_congested = nr_congested; 1285 stat->nr_unqueued_dirty = nr_unqueued_dirty; 1286 stat->nr_writeback = nr_writeback; 1287 stat->nr_immediate = nr_immediate; 1288 stat->nr_activate = pgactivate; 1289 stat->nr_ref_keep = nr_ref_keep; 1290 stat->nr_unmap_fail = nr_unmap_fail; 1291 } 1292 return nr_reclaimed; 1293 } 1294 1295 unsigned long reclaim_clean_pages_from_list(struct zone *zone, 1296 struct list_head *page_list) 1297 { 1298 struct scan_control sc = { 1299 .gfp_mask = GFP_KERNEL, 1300 .priority = DEF_PRIORITY, 1301 .may_unmap = 1, 1302 }; 1303 unsigned long ret; 1304 struct page *page, *next; 1305 LIST_HEAD(clean_pages); 1306 1307 list_for_each_entry_safe(page, next, page_list, lru) { 1308 if (page_is_file_cache(page) && !PageDirty(page) && 1309 !__PageMovable(page)) { 1310 ClearPageActive(page); 1311 list_move(&page->lru, &clean_pages); 1312 } 1313 } 1314 1315 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, 1316 TTU_IGNORE_ACCESS, NULL, true); 1317 list_splice(&clean_pages, page_list); 1318 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); 1319 return ret; 1320 } 1321 1322 /* 1323 * Attempt to remove the specified page from its LRU. Only take this page 1324 * if it is of the appropriate PageActive status. Pages which are being 1325 * freed elsewhere are also ignored. 1326 * 1327 * page: page to consider 1328 * mode: one of the LRU isolation modes defined above 1329 * 1330 * returns 0 on success, -ve errno on failure. 1331 */ 1332 int __isolate_lru_page(struct page *page, isolate_mode_t mode) 1333 { 1334 int ret = -EINVAL; 1335 1336 /* Only take pages on the LRU. */ 1337 if (!PageLRU(page)) 1338 return ret; 1339 1340 /* Compaction should not handle unevictable pages but CMA can do so */ 1341 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) 1342 return ret; 1343 1344 ret = -EBUSY; 1345 1346 /* 1347 * To minimise LRU disruption, the caller can indicate that it only 1348 * wants to isolate pages it will be able to operate on without 1349 * blocking - clean pages for the most part. 1350 * 1351 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages 1352 * that it is possible to migrate without blocking 1353 */ 1354 if (mode & ISOLATE_ASYNC_MIGRATE) { 1355 /* All the caller can do on PageWriteback is block */ 1356 if (PageWriteback(page)) 1357 return ret; 1358 1359 if (PageDirty(page)) { 1360 struct address_space *mapping; 1361 bool migrate_dirty; 1362 1363 /* 1364 * Only pages without mappings or that have a 1365 * ->migratepage callback are possible to migrate 1366 * without blocking. However, we can be racing with 1367 * truncation so it's necessary to lock the page 1368 * to stabilise the mapping as truncation holds 1369 * the page lock until after the page is removed 1370 * from the page cache. 1371 */ 1372 if (!trylock_page(page)) 1373 return ret; 1374 1375 mapping = page_mapping(page); 1376 migrate_dirty = mapping && mapping->a_ops->migratepage; 1377 unlock_page(page); 1378 if (!migrate_dirty) 1379 return ret; 1380 } 1381 } 1382 1383 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) 1384 return ret; 1385 1386 if (likely(get_page_unless_zero(page))) { 1387 /* 1388 * Be careful not to clear PageLRU until after we're 1389 * sure the page is not being freed elsewhere -- the 1390 * page release code relies on it. 1391 */ 1392 ClearPageLRU(page); 1393 ret = 0; 1394 } 1395 1396 return ret; 1397 } 1398 1399 1400 /* 1401 * Update LRU sizes after isolating pages. The LRU size updates must 1402 * be complete before mem_cgroup_update_lru_size due to a santity check. 1403 */ 1404 static __always_inline void update_lru_sizes(struct lruvec *lruvec, 1405 enum lru_list lru, unsigned long *nr_zone_taken) 1406 { 1407 int zid; 1408 1409 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 1410 if (!nr_zone_taken[zid]) 1411 continue; 1412 1413 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); 1414 #ifdef CONFIG_MEMCG 1415 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); 1416 #endif 1417 } 1418 1419 } 1420 1421 /* 1422 * zone_lru_lock is heavily contended. Some of the functions that 1423 * shrink the lists perform better by taking out a batch of pages 1424 * and working on them outside the LRU lock. 1425 * 1426 * For pagecache intensive workloads, this function is the hottest 1427 * spot in the kernel (apart from copy_*_user functions). 1428 * 1429 * Appropriate locks must be held before calling this function. 1430 * 1431 * @nr_to_scan: The number of eligible pages to look through on the list. 1432 * @lruvec: The LRU vector to pull pages from. 1433 * @dst: The temp list to put pages on to. 1434 * @nr_scanned: The number of pages that were scanned. 1435 * @sc: The scan_control struct for this reclaim session 1436 * @mode: One of the LRU isolation modes 1437 * @lru: LRU list id for isolating 1438 * 1439 * returns how many pages were moved onto *@dst. 1440 */ 1441 static unsigned long isolate_lru_pages(unsigned long nr_to_scan, 1442 struct lruvec *lruvec, struct list_head *dst, 1443 unsigned long *nr_scanned, struct scan_control *sc, 1444 isolate_mode_t mode, enum lru_list lru) 1445 { 1446 struct list_head *src = &lruvec->lists[lru]; 1447 unsigned long nr_taken = 0; 1448 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; 1449 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; 1450 unsigned long skipped = 0; 1451 unsigned long scan, total_scan, nr_pages; 1452 LIST_HEAD(pages_skipped); 1453 1454 scan = 0; 1455 for (total_scan = 0; 1456 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src); 1457 total_scan++) { 1458 struct page *page; 1459 1460 page = lru_to_page(src); 1461 prefetchw_prev_lru_page(page, src, flags); 1462 1463 VM_BUG_ON_PAGE(!PageLRU(page), page); 1464 1465 if (page_zonenum(page) > sc->reclaim_idx) { 1466 list_move(&page->lru, &pages_skipped); 1467 nr_skipped[page_zonenum(page)]++; 1468 continue; 1469 } 1470 1471 /* 1472 * Do not count skipped pages because that makes the function 1473 * return with no isolated pages if the LRU mostly contains 1474 * ineligible pages. This causes the VM to not reclaim any 1475 * pages, triggering a premature OOM. 1476 */ 1477 scan++; 1478 switch (__isolate_lru_page(page, mode)) { 1479 case 0: 1480 nr_pages = hpage_nr_pages(page); 1481 nr_taken += nr_pages; 1482 nr_zone_taken[page_zonenum(page)] += nr_pages; 1483 list_move(&page->lru, dst); 1484 break; 1485 1486 case -EBUSY: 1487 /* else it is being freed elsewhere */ 1488 list_move(&page->lru, src); 1489 continue; 1490 1491 default: 1492 BUG(); 1493 } 1494 } 1495 1496 /* 1497 * Splice any skipped pages to the start of the LRU list. Note that 1498 * this disrupts the LRU order when reclaiming for lower zones but 1499 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX 1500 * scanning would soon rescan the same pages to skip and put the 1501 * system at risk of premature OOM. 1502 */ 1503 if (!list_empty(&pages_skipped)) { 1504 int zid; 1505 1506 list_splice(&pages_skipped, src); 1507 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 1508 if (!nr_skipped[zid]) 1509 continue; 1510 1511 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); 1512 skipped += nr_skipped[zid]; 1513 } 1514 } 1515 *nr_scanned = total_scan; 1516 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, 1517 total_scan, skipped, nr_taken, mode, lru); 1518 update_lru_sizes(lruvec, lru, nr_zone_taken); 1519 return nr_taken; 1520 } 1521 1522 /** 1523 * isolate_lru_page - tries to isolate a page from its LRU list 1524 * @page: page to isolate from its LRU list 1525 * 1526 * Isolates a @page from an LRU list, clears PageLRU and adjusts the 1527 * vmstat statistic corresponding to whatever LRU list the page was on. 1528 * 1529 * Returns 0 if the page was removed from an LRU list. 1530 * Returns -EBUSY if the page was not on an LRU list. 1531 * 1532 * The returned page will have PageLRU() cleared. If it was found on 1533 * the active list, it will have PageActive set. If it was found on 1534 * the unevictable list, it will have the PageUnevictable bit set. That flag 1535 * may need to be cleared by the caller before letting the page go. 1536 * 1537 * The vmstat statistic corresponding to the list on which the page was 1538 * found will be decremented. 1539 * 1540 * Restrictions: 1541 * 1542 * (1) Must be called with an elevated refcount on the page. This is a 1543 * fundamentnal difference from isolate_lru_pages (which is called 1544 * without a stable reference). 1545 * (2) the lru_lock must not be held. 1546 * (3) interrupts must be enabled. 1547 */ 1548 int isolate_lru_page(struct page *page) 1549 { 1550 int ret = -EBUSY; 1551 1552 VM_BUG_ON_PAGE(!page_count(page), page); 1553 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); 1554 1555 if (PageLRU(page)) { 1556 struct zone *zone = page_zone(page); 1557 struct lruvec *lruvec; 1558 1559 spin_lock_irq(zone_lru_lock(zone)); 1560 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); 1561 if (PageLRU(page)) { 1562 int lru = page_lru(page); 1563 get_page(page); 1564 ClearPageLRU(page); 1565 del_page_from_lru_list(page, lruvec, lru); 1566 ret = 0; 1567 } 1568 spin_unlock_irq(zone_lru_lock(zone)); 1569 } 1570 return ret; 1571 } 1572 1573 /* 1574 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and 1575 * then get resheduled. When there are massive number of tasks doing page 1576 * allocation, such sleeping direct reclaimers may keep piling up on each CPU, 1577 * the LRU list will go small and be scanned faster than necessary, leading to 1578 * unnecessary swapping, thrashing and OOM. 1579 */ 1580 static int too_many_isolated(struct pglist_data *pgdat, int file, 1581 struct scan_control *sc) 1582 { 1583 unsigned long inactive, isolated; 1584 1585 if (current_is_kswapd()) 1586 return 0; 1587 1588 if (!sane_reclaim(sc)) 1589 return 0; 1590 1591 if (file) { 1592 inactive = node_page_state(pgdat, NR_INACTIVE_FILE); 1593 isolated = node_page_state(pgdat, NR_ISOLATED_FILE); 1594 } else { 1595 inactive = node_page_state(pgdat, NR_INACTIVE_ANON); 1596 isolated = node_page_state(pgdat, NR_ISOLATED_ANON); 1597 } 1598 1599 /* 1600 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they 1601 * won't get blocked by normal direct-reclaimers, forming a circular 1602 * deadlock. 1603 */ 1604 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) 1605 inactive >>= 3; 1606 1607 return isolated > inactive; 1608 } 1609 1610 static noinline_for_stack void 1611 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) 1612 { 1613 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1614 struct pglist_data *pgdat = lruvec_pgdat(lruvec); 1615 LIST_HEAD(pages_to_free); 1616 1617 /* 1618 * Put back any unfreeable pages. 1619 */ 1620 while (!list_empty(page_list)) { 1621 struct page *page = lru_to_page(page_list); 1622 int lru; 1623 1624 VM_BUG_ON_PAGE(PageLRU(page), page); 1625 list_del(&page->lru); 1626 if (unlikely(!page_evictable(page))) { 1627 spin_unlock_irq(&pgdat->lru_lock); 1628 putback_lru_page(page); 1629 spin_lock_irq(&pgdat->lru_lock); 1630 continue; 1631 } 1632 1633 lruvec = mem_cgroup_page_lruvec(page, pgdat); 1634 1635 SetPageLRU(page); 1636 lru = page_lru(page); 1637 add_page_to_lru_list(page, lruvec, lru); 1638 1639 if (is_active_lru(lru)) { 1640 int file = is_file_lru(lru); 1641 int numpages = hpage_nr_pages(page); 1642 reclaim_stat->recent_rotated[file] += numpages; 1643 } 1644 if (put_page_testzero(page)) { 1645 __ClearPageLRU(page); 1646 __ClearPageActive(page); 1647 del_page_from_lru_list(page, lruvec, lru); 1648 1649 if (unlikely(PageCompound(page))) { 1650 spin_unlock_irq(&pgdat->lru_lock); 1651 mem_cgroup_uncharge(page); 1652 (*get_compound_page_dtor(page))(page); 1653 spin_lock_irq(&pgdat->lru_lock); 1654 } else 1655 list_add(&page->lru, &pages_to_free); 1656 } 1657 } 1658 1659 /* 1660 * To save our caller's stack, now use input list for pages to free. 1661 */ 1662 list_splice(&pages_to_free, page_list); 1663 } 1664 1665 /* 1666 * If a kernel thread (such as nfsd for loop-back mounts) services 1667 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. 1668 * In that case we should only throttle if the backing device it is 1669 * writing to is congested. In other cases it is safe to throttle. 1670 */ 1671 static int current_may_throttle(void) 1672 { 1673 return !(current->flags & PF_LESS_THROTTLE) || 1674 current->backing_dev_info == NULL || 1675 bdi_write_congested(current->backing_dev_info); 1676 } 1677 1678 /* 1679 * shrink_inactive_list() is a helper for shrink_node(). It returns the number 1680 * of reclaimed pages 1681 */ 1682 static noinline_for_stack unsigned long 1683 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, 1684 struct scan_control *sc, enum lru_list lru) 1685 { 1686 LIST_HEAD(page_list); 1687 unsigned long nr_scanned; 1688 unsigned long nr_reclaimed = 0; 1689 unsigned long nr_taken; 1690 struct reclaim_stat stat = {}; 1691 isolate_mode_t isolate_mode = 0; 1692 int file = is_file_lru(lru); 1693 struct pglist_data *pgdat = lruvec_pgdat(lruvec); 1694 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1695 bool stalled = false; 1696 1697 while (unlikely(too_many_isolated(pgdat, file, sc))) { 1698 if (stalled) 1699 return 0; 1700 1701 /* wait a bit for the reclaimer. */ 1702 msleep(100); 1703 stalled = true; 1704 1705 /* We are about to die and free our memory. Return now. */ 1706 if (fatal_signal_pending(current)) 1707 return SWAP_CLUSTER_MAX; 1708 } 1709 1710 lru_add_drain(); 1711 1712 if (!sc->may_unmap) 1713 isolate_mode |= ISOLATE_UNMAPPED; 1714 1715 spin_lock_irq(&pgdat->lru_lock); 1716 1717 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, 1718 &nr_scanned, sc, isolate_mode, lru); 1719 1720 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); 1721 reclaim_stat->recent_scanned[file] += nr_taken; 1722 1723 if (current_is_kswapd()) { 1724 if (global_reclaim(sc)) 1725 __count_vm_events(PGSCAN_KSWAPD, nr_scanned); 1726 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD, 1727 nr_scanned); 1728 } else { 1729 if (global_reclaim(sc)) 1730 __count_vm_events(PGSCAN_DIRECT, nr_scanned); 1731 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT, 1732 nr_scanned); 1733 } 1734 spin_unlock_irq(&pgdat->lru_lock); 1735 1736 if (nr_taken == 0) 1737 return 0; 1738 1739 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0, 1740 &stat, false); 1741 1742 spin_lock_irq(&pgdat->lru_lock); 1743 1744 if (current_is_kswapd()) { 1745 if (global_reclaim(sc)) 1746 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed); 1747 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD, 1748 nr_reclaimed); 1749 } else { 1750 if (global_reclaim(sc)) 1751 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed); 1752 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT, 1753 nr_reclaimed); 1754 } 1755 1756 putback_inactive_pages(lruvec, &page_list); 1757 1758 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); 1759 1760 spin_unlock_irq(&pgdat->lru_lock); 1761 1762 mem_cgroup_uncharge_list(&page_list); 1763 free_unref_page_list(&page_list); 1764 1765 /* 1766 * If reclaim is isolating dirty pages under writeback, it implies 1767 * that the long-lived page allocation rate is exceeding the page 1768 * laundering rate. Either the global limits are not being effective 1769 * at throttling processes due to the page distribution throughout 1770 * zones or there is heavy usage of a slow backing device. The 1771 * only option is to throttle from reclaim context which is not ideal 1772 * as there is no guarantee the dirtying process is throttled in the 1773 * same way balance_dirty_pages() manages. 1774 * 1775 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number 1776 * of pages under pages flagged for immediate reclaim and stall if any 1777 * are encountered in the nr_immediate check below. 1778 */ 1779 if (stat.nr_writeback && stat.nr_writeback == nr_taken) 1780 set_bit(PGDAT_WRITEBACK, &pgdat->flags); 1781 1782 /* 1783 * Legacy memcg will stall in page writeback so avoid forcibly 1784 * stalling here. 1785 */ 1786 if (sane_reclaim(sc)) { 1787 /* 1788 * Tag a zone as congested if all the dirty pages scanned were 1789 * backed by a congested BDI and wait_iff_congested will stall. 1790 */ 1791 if (stat.nr_dirty && stat.nr_dirty == stat.nr_congested) 1792 set_bit(PGDAT_CONGESTED, &pgdat->flags); 1793 1794 /* 1795 * If dirty pages are scanned that are not queued for IO, it 1796 * implies that flushers are not doing their job. This can 1797 * happen when memory pressure pushes dirty pages to the end of 1798 * the LRU before the dirty limits are breached and the dirty 1799 * data has expired. It can also happen when the proportion of 1800 * dirty pages grows not through writes but through memory 1801 * pressure reclaiming all the clean cache. And in some cases, 1802 * the flushers simply cannot keep up with the allocation 1803 * rate. Nudge the flusher threads in case they are asleep, but 1804 * also allow kswapd to start writing pages during reclaim. 1805 */ 1806 if (stat.nr_unqueued_dirty == nr_taken) { 1807 wakeup_flusher_threads(WB_REASON_VMSCAN); 1808 set_bit(PGDAT_DIRTY, &pgdat->flags); 1809 } 1810 1811 /* 1812 * If kswapd scans pages marked marked for immediate 1813 * reclaim and under writeback (nr_immediate), it implies 1814 * that pages are cycling through the LRU faster than 1815 * they are written so also forcibly stall. 1816 */ 1817 if (stat.nr_immediate && current_may_throttle()) 1818 congestion_wait(BLK_RW_ASYNC, HZ/10); 1819 } 1820 1821 /* 1822 * Stall direct reclaim for IO completions if underlying BDIs or zone 1823 * is congested. Allow kswapd to continue until it starts encountering 1824 * unqueued dirty pages or cycling through the LRU too quickly. 1825 */ 1826 if (!sc->hibernation_mode && !current_is_kswapd() && 1827 current_may_throttle()) 1828 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10); 1829 1830 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, 1831 nr_scanned, nr_reclaimed, 1832 stat.nr_dirty, stat.nr_writeback, 1833 stat.nr_congested, stat.nr_immediate, 1834 stat.nr_activate, stat.nr_ref_keep, 1835 stat.nr_unmap_fail, 1836 sc->priority, file); 1837 return nr_reclaimed; 1838 } 1839 1840 /* 1841 * This moves pages from the active list to the inactive list. 1842 * 1843 * We move them the other way if the page is referenced by one or more 1844 * processes, from rmap. 1845 * 1846 * If the pages are mostly unmapped, the processing is fast and it is 1847 * appropriate to hold zone_lru_lock across the whole operation. But if 1848 * the pages are mapped, the processing is slow (page_referenced()) so we 1849 * should drop zone_lru_lock around each page. It's impossible to balance 1850 * this, so instead we remove the pages from the LRU while processing them. 1851 * It is safe to rely on PG_active against the non-LRU pages in here because 1852 * nobody will play with that bit on a non-LRU page. 1853 * 1854 * The downside is that we have to touch page->_refcount against each page. 1855 * But we had to alter page->flags anyway. 1856 * 1857 * Returns the number of pages moved to the given lru. 1858 */ 1859 1860 static unsigned move_active_pages_to_lru(struct lruvec *lruvec, 1861 struct list_head *list, 1862 struct list_head *pages_to_free, 1863 enum lru_list lru) 1864 { 1865 struct pglist_data *pgdat = lruvec_pgdat(lruvec); 1866 struct page *page; 1867 int nr_pages; 1868 int nr_moved = 0; 1869 1870 while (!list_empty(list)) { 1871 page = lru_to_page(list); 1872 lruvec = mem_cgroup_page_lruvec(page, pgdat); 1873 1874 VM_BUG_ON_PAGE(PageLRU(page), page); 1875 SetPageLRU(page); 1876 1877 nr_pages = hpage_nr_pages(page); 1878 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); 1879 list_move(&page->lru, &lruvec->lists[lru]); 1880 1881 if (put_page_testzero(page)) { 1882 __ClearPageLRU(page); 1883 __ClearPageActive(page); 1884 del_page_from_lru_list(page, lruvec, lru); 1885 1886 if (unlikely(PageCompound(page))) { 1887 spin_unlock_irq(&pgdat->lru_lock); 1888 mem_cgroup_uncharge(page); 1889 (*get_compound_page_dtor(page))(page); 1890 spin_lock_irq(&pgdat->lru_lock); 1891 } else 1892 list_add(&page->lru, pages_to_free); 1893 } else { 1894 nr_moved += nr_pages; 1895 } 1896 } 1897 1898 if (!is_active_lru(lru)) { 1899 __count_vm_events(PGDEACTIVATE, nr_moved); 1900 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, 1901 nr_moved); 1902 } 1903 1904 return nr_moved; 1905 } 1906 1907 static void shrink_active_list(unsigned long nr_to_scan, 1908 struct lruvec *lruvec, 1909 struct scan_control *sc, 1910 enum lru_list lru) 1911 { 1912 unsigned long nr_taken; 1913 unsigned long nr_scanned; 1914 unsigned long vm_flags; 1915 LIST_HEAD(l_hold); /* The pages which were snipped off */ 1916 LIST_HEAD(l_active); 1917 LIST_HEAD(l_inactive); 1918 struct page *page; 1919 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1920 unsigned nr_deactivate, nr_activate; 1921 unsigned nr_rotated = 0; 1922 isolate_mode_t isolate_mode = 0; 1923 int file = is_file_lru(lru); 1924 struct pglist_data *pgdat = lruvec_pgdat(lruvec); 1925 1926 lru_add_drain(); 1927 1928 if (!sc->may_unmap) 1929 isolate_mode |= ISOLATE_UNMAPPED; 1930 1931 spin_lock_irq(&pgdat->lru_lock); 1932 1933 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, 1934 &nr_scanned, sc, isolate_mode, lru); 1935 1936 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); 1937 reclaim_stat->recent_scanned[file] += nr_taken; 1938 1939 __count_vm_events(PGREFILL, nr_scanned); 1940 count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); 1941 1942 spin_unlock_irq(&pgdat->lru_lock); 1943 1944 while (!list_empty(&l_hold)) { 1945 cond_resched(); 1946 page = lru_to_page(&l_hold); 1947 list_del(&page->lru); 1948 1949 if (unlikely(!page_evictable(page))) { 1950 putback_lru_page(page); 1951 continue; 1952 } 1953 1954 if (unlikely(buffer_heads_over_limit)) { 1955 if (page_has_private(page) && trylock_page(page)) { 1956 if (page_has_private(page)) 1957 try_to_release_page(page, 0); 1958 unlock_page(page); 1959 } 1960 } 1961 1962 if (page_referenced(page, 0, sc->target_mem_cgroup, 1963 &vm_flags)) { 1964 nr_rotated += hpage_nr_pages(page); 1965 /* 1966 * Identify referenced, file-backed active pages and 1967 * give them one more trip around the active list. So 1968 * that executable code get better chances to stay in 1969 * memory under moderate memory pressure. Anon pages 1970 * are not likely to be evicted by use-once streaming 1971 * IO, plus JVM can create lots of anon VM_EXEC pages, 1972 * so we ignore them here. 1973 */ 1974 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { 1975 list_add(&page->lru, &l_active); 1976 continue; 1977 } 1978 } 1979 1980 ClearPageActive(page); /* we are de-activating */ 1981 list_add(&page->lru, &l_inactive); 1982 } 1983 1984 /* 1985 * Move pages back to the lru list. 1986 */ 1987 spin_lock_irq(&pgdat->lru_lock); 1988 /* 1989 * Count referenced pages from currently used mappings as rotated, 1990 * even though only some of them are actually re-activated. This 1991 * helps balance scan pressure between file and anonymous pages in 1992 * get_scan_count. 1993 */ 1994 reclaim_stat->recent_rotated[file] += nr_rotated; 1995 1996 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); 1997 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); 1998 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); 1999 spin_unlock_irq(&pgdat->lru_lock); 2000 2001 mem_cgroup_uncharge_list(&l_hold); 2002 free_unref_page_list(&l_hold); 2003 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, 2004 nr_deactivate, nr_rotated, sc->priority, file); 2005 } 2006 2007 /* 2008 * The inactive anon list should be small enough that the VM never has 2009 * to do too much work. 2010 * 2011 * The inactive file list should be small enough to leave most memory 2012 * to the established workingset on the scan-resistant active list, 2013 * but large enough to avoid thrashing the aggregate readahead window. 2014 * 2015 * Both inactive lists should also be large enough that each inactive 2016 * page has a chance to be referenced again before it is reclaimed. 2017 * 2018 * If that fails and refaulting is observed, the inactive list grows. 2019 * 2020 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages 2021 * on this LRU, maintained by the pageout code. An inactive_ratio 2022 * of 3 means 3:1 or 25% of the pages are kept on the inactive list. 2023 * 2024 * total target max 2025 * memory ratio inactive 2026 * ------------------------------------- 2027 * 10MB 1 5MB 2028 * 100MB 1 50MB 2029 * 1GB 3 250MB 2030 * 10GB 10 0.9GB 2031 * 100GB 31 3GB 2032 * 1TB 101 10GB 2033 * 10TB 320 32GB 2034 */ 2035 static bool inactive_list_is_low(struct lruvec *lruvec, bool file, 2036 struct mem_cgroup *memcg, 2037 struct scan_control *sc, bool actual_reclaim) 2038 { 2039 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE; 2040 struct pglist_data *pgdat = lruvec_pgdat(lruvec); 2041 enum lru_list inactive_lru = file * LRU_FILE; 2042 unsigned long inactive, active; 2043 unsigned long inactive_ratio; 2044 unsigned long refaults; 2045 unsigned long gb; 2046 2047 /* 2048 * If we don't have swap space, anonymous page deactivation 2049 * is pointless. 2050 */ 2051 if (!file && !total_swap_pages) 2052 return false; 2053 2054 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx); 2055 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx); 2056 2057 if (memcg) 2058 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); 2059 else 2060 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); 2061 2062 /* 2063 * When refaults are being observed, it means a new workingset 2064 * is being established. Disable active list protection to get 2065 * rid of the stale workingset quickly. 2066 */ 2067 if (file && actual_reclaim && lruvec->refaults != refaults) { 2068 inactive_ratio = 0; 2069 } else { 2070 gb = (inactive + active) >> (30 - PAGE_SHIFT); 2071 if (gb) 2072 inactive_ratio = int_sqrt(10 * gb); 2073 else 2074 inactive_ratio = 1; 2075 } 2076 2077 if (actual_reclaim) 2078 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx, 2079 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive, 2080 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active, 2081 inactive_ratio, file); 2082 2083 return inactive * inactive_ratio < active; 2084 } 2085 2086 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, 2087 struct lruvec *lruvec, struct mem_cgroup *memcg, 2088 struct scan_control *sc) 2089 { 2090 if (is_active_lru(lru)) { 2091 if (inactive_list_is_low(lruvec, is_file_lru(lru), 2092 memcg, sc, true)) 2093 shrink_active_list(nr_to_scan, lruvec, sc, lru); 2094 return 0; 2095 } 2096 2097 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); 2098 } 2099 2100 enum scan_balance { 2101 SCAN_EQUAL, 2102 SCAN_FRACT, 2103 SCAN_ANON, 2104 SCAN_FILE, 2105 }; 2106 2107 /* 2108 * Determine how aggressively the anon and file LRU lists should be 2109 * scanned. The relative value of each set of LRU lists is determined 2110 * by looking at the fraction of the pages scanned we did rotate back 2111 * onto the active list instead of evict. 2112 * 2113 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan 2114 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan 2115 */ 2116 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, 2117 struct scan_control *sc, unsigned long *nr, 2118 unsigned long *lru_pages) 2119 { 2120 int swappiness = mem_cgroup_swappiness(memcg); 2121 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 2122 u64 fraction[2]; 2123 u64 denominator = 0; /* gcc */ 2124 struct pglist_data *pgdat = lruvec_pgdat(lruvec); 2125 unsigned long anon_prio, file_prio; 2126 enum scan_balance scan_balance; 2127 unsigned long anon, file; 2128 unsigned long ap, fp; 2129 enum lru_list lru; 2130 2131 /* If we have no swap space, do not bother scanning anon pages. */ 2132 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { 2133 scan_balance = SCAN_FILE; 2134 goto out; 2135 } 2136 2137 /* 2138 * Global reclaim will swap to prevent OOM even with no 2139 * swappiness, but memcg users want to use this knob to 2140 * disable swapping for individual groups completely when 2141 * using the memory controller's swap limit feature would be 2142 * too expensive. 2143 */ 2144 if (!global_reclaim(sc) && !swappiness) { 2145 scan_balance = SCAN_FILE; 2146 goto out; 2147 } 2148 2149 /* 2150 * Do not apply any pressure balancing cleverness when the 2151 * system is close to OOM, scan both anon and file equally 2152 * (unless the swappiness setting disagrees with swapping). 2153 */ 2154 if (!sc->priority && swappiness) { 2155 scan_balance = SCAN_EQUAL; 2156 goto out; 2157 } 2158 2159 /* 2160 * Prevent the reclaimer from falling into the cache trap: as 2161 * cache pages start out inactive, every cache fault will tip 2162 * the scan balance towards the file LRU. And as the file LRU 2163 * shrinks, so does the window for rotation from references. 2164 * This means we have a runaway feedback loop where a tiny 2165 * thrashing file LRU becomes infinitely more attractive than 2166 * anon pages. Try to detect this based on file LRU size. 2167 */ 2168 if (global_reclaim(sc)) { 2169 unsigned long pgdatfile; 2170 unsigned long pgdatfree; 2171 int z; 2172 unsigned long total_high_wmark = 0; 2173 2174 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); 2175 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + 2176 node_page_state(pgdat, NR_INACTIVE_FILE); 2177 2178 for (z = 0; z < MAX_NR_ZONES; z++) { 2179 struct zone *zone = &pgdat->node_zones[z]; 2180 if (!managed_zone(zone)) 2181 continue; 2182 2183 total_high_wmark += high_wmark_pages(zone); 2184 } 2185 2186 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { 2187 /* 2188 * Force SCAN_ANON if there are enough inactive 2189 * anonymous pages on the LRU in eligible zones. 2190 * Otherwise, the small LRU gets thrashed. 2191 */ 2192 if (!inactive_list_is_low(lruvec, false, memcg, sc, false) && 2193 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx) 2194 >> sc->priority) { 2195 scan_balance = SCAN_ANON; 2196 goto out; 2197 } 2198 } 2199 } 2200 2201 /* 2202 * If there is enough inactive page cache, i.e. if the size of the 2203 * inactive list is greater than that of the active list *and* the 2204 * inactive list actually has some pages to scan on this priority, we 2205 * do not reclaim anything from the anonymous working set right now. 2206 * Without the second condition we could end up never scanning an 2207 * lruvec even if it has plenty of old anonymous pages unless the 2208 * system is under heavy pressure. 2209 */ 2210 if (!inactive_list_is_low(lruvec, true, memcg, sc, false) && 2211 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) { 2212 scan_balance = SCAN_FILE; 2213 goto out; 2214 } 2215 2216 scan_balance = SCAN_FRACT; 2217 2218 /* 2219 * With swappiness at 100, anonymous and file have the same priority. 2220 * This scanning priority is essentially the inverse of IO cost. 2221 */ 2222 anon_prio = swappiness; 2223 file_prio = 200 - anon_prio; 2224 2225 /* 2226 * OK, so we have swap space and a fair amount of page cache 2227 * pages. We use the recently rotated / recently scanned 2228 * ratios to determine how valuable each cache is. 2229 * 2230 * Because workloads change over time (and to avoid overflow) 2231 * we keep these statistics as a floating average, which ends 2232 * up weighing recent references more than old ones. 2233 * 2234 * anon in [0], file in [1] 2235 */ 2236 2237 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + 2238 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); 2239 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + 2240 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); 2241 2242 spin_lock_irq(&pgdat->lru_lock); 2243 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { 2244 reclaim_stat->recent_scanned[0] /= 2; 2245 reclaim_stat->recent_rotated[0] /= 2; 2246 } 2247 2248 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { 2249 reclaim_stat->recent_scanned[1] /= 2; 2250 reclaim_stat->recent_rotated[1] /= 2; 2251 } 2252 2253 /* 2254 * The amount of pressure on anon vs file pages is inversely 2255 * proportional to the fraction of recently scanned pages on 2256 * each list that were recently referenced and in active use. 2257 */ 2258 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); 2259 ap /= reclaim_stat->recent_rotated[0] + 1; 2260 2261 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); 2262 fp /= reclaim_stat->recent_rotated[1] + 1; 2263 spin_unlock_irq(&pgdat->lru_lock); 2264 2265 fraction[0] = ap; 2266 fraction[1] = fp; 2267 denominator = ap + fp + 1; 2268 out: 2269 *lru_pages = 0; 2270 for_each_evictable_lru(lru) { 2271 int file = is_file_lru(lru); 2272 unsigned long size; 2273 unsigned long scan; 2274 2275 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); 2276 scan = size >> sc->priority; 2277 /* 2278 * If the cgroup's already been deleted, make sure to 2279 * scrape out the remaining cache. 2280 */ 2281 if (!scan && !mem_cgroup_online(memcg)) 2282 scan = min(size, SWAP_CLUSTER_MAX); 2283 2284 switch (scan_balance) { 2285 case SCAN_EQUAL: 2286 /* Scan lists relative to size */ 2287 break; 2288 case SCAN_FRACT: 2289 /* 2290 * Scan types proportional to swappiness and 2291 * their relative recent reclaim efficiency. 2292 */ 2293 scan = div64_u64(scan * fraction[file], 2294 denominator); 2295 break; 2296 case SCAN_FILE: 2297 case SCAN_ANON: 2298 /* Scan one type exclusively */ 2299 if ((scan_balance == SCAN_FILE) != file) { 2300 size = 0; 2301 scan = 0; 2302 } 2303 break; 2304 default: 2305 /* Look ma, no brain */ 2306 BUG(); 2307 } 2308 2309 *lru_pages += size; 2310 nr[lru] = scan; 2311 } 2312 } 2313 2314 /* 2315 * This is a basic per-node page freer. Used by both kswapd and direct reclaim. 2316 */ 2317 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg, 2318 struct scan_control *sc, unsigned long *lru_pages) 2319 { 2320 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); 2321 unsigned long nr[NR_LRU_LISTS]; 2322 unsigned long targets[NR_LRU_LISTS]; 2323 unsigned long nr_to_scan; 2324 enum lru_list lru; 2325 unsigned long nr_reclaimed = 0; 2326 unsigned long nr_to_reclaim = sc->nr_to_reclaim; 2327 struct blk_plug plug; 2328 bool scan_adjusted; 2329 2330 get_scan_count(lruvec, memcg, sc, nr, lru_pages); 2331 2332 /* Record the original scan target for proportional adjustments later */ 2333 memcpy(targets, nr, sizeof(nr)); 2334 2335 /* 2336 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal 2337 * event that can occur when there is little memory pressure e.g. 2338 * multiple streaming readers/writers. Hence, we do not abort scanning 2339 * when the requested number of pages are reclaimed when scanning at 2340 * DEF_PRIORITY on the assumption that the fact we are direct 2341 * reclaiming implies that kswapd is not keeping up and it is best to 2342 * do a batch of work at once. For memcg reclaim one check is made to 2343 * abort proportional reclaim if either the file or anon lru has already 2344 * dropped to zero at the first pass. 2345 */ 2346 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && 2347 sc->priority == DEF_PRIORITY); 2348 2349 blk_start_plug(&plug); 2350 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || 2351 nr[LRU_INACTIVE_FILE]) { 2352 unsigned long nr_anon, nr_file, percentage; 2353 unsigned long nr_scanned; 2354 2355 for_each_evictable_lru(lru) { 2356 if (nr[lru]) { 2357 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); 2358 nr[lru] -= nr_to_scan; 2359 2360 nr_reclaimed += shrink_list(lru, nr_to_scan, 2361 lruvec, memcg, sc); 2362 } 2363 } 2364 2365 cond_resched(); 2366 2367 if (nr_reclaimed < nr_to_reclaim || scan_adjusted) 2368 continue; 2369 2370 /* 2371 * For kswapd and memcg, reclaim at least the number of pages 2372 * requested. Ensure that the anon and file LRUs are scanned 2373 * proportionally what was requested by get_scan_count(). We 2374 * stop reclaiming one LRU and reduce the amount scanning 2375 * proportional to the original scan target. 2376 */ 2377 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; 2378 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; 2379 2380 /* 2381 * It's just vindictive to attack the larger once the smaller 2382 * has gone to zero. And given the way we stop scanning the 2383 * smaller below, this makes sure that we only make one nudge 2384 * towards proportionality once we've got nr_to_reclaim. 2385 */ 2386 if (!nr_file || !nr_anon) 2387 break; 2388 2389 if (nr_file > nr_anon) { 2390 unsigned long scan_target = targets[LRU_INACTIVE_ANON] + 2391 targets[LRU_ACTIVE_ANON] + 1; 2392 lru = LRU_BASE; 2393 percentage = nr_anon * 100 / scan_target; 2394 } else { 2395 unsigned long scan_target = targets[LRU_INACTIVE_FILE] + 2396 targets[LRU_ACTIVE_FILE] + 1; 2397 lru = LRU_FILE; 2398 percentage = nr_file * 100 / scan_target; 2399 } 2400 2401 /* Stop scanning the smaller of the LRU */ 2402 nr[lru] = 0; 2403 nr[lru + LRU_ACTIVE] = 0; 2404 2405 /* 2406 * Recalculate the other LRU scan count based on its original 2407 * scan target and the percentage scanning already complete 2408 */ 2409 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; 2410 nr_scanned = targets[lru] - nr[lru]; 2411 nr[lru] = targets[lru] * (100 - percentage) / 100; 2412 nr[lru] -= min(nr[lru], nr_scanned); 2413 2414 lru += LRU_ACTIVE; 2415 nr_scanned = targets[lru] - nr[lru]; 2416 nr[lru] = targets[lru] * (100 - percentage) / 100; 2417 nr[lru] -= min(nr[lru], nr_scanned); 2418 2419 scan_adjusted = true; 2420 } 2421 blk_finish_plug(&plug); 2422 sc->nr_reclaimed += nr_reclaimed; 2423 2424 /* 2425 * Even if we did not try to evict anon pages at all, we want to 2426 * rebalance the anon lru active/inactive ratio. 2427 */ 2428 if (inactive_list_is_low(lruvec, false, memcg, sc, true)) 2429 shrink_active_list(SWAP_CLUSTER_MAX, lruvec, 2430 sc, LRU_ACTIVE_ANON); 2431 } 2432 2433 /* Use reclaim/compaction for costly allocs or under memory pressure */ 2434 static bool in_reclaim_compaction(struct scan_control *sc) 2435 { 2436 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && 2437 (sc->order > PAGE_ALLOC_COSTLY_ORDER || 2438 sc->priority < DEF_PRIORITY - 2)) 2439 return true; 2440 2441 return false; 2442 } 2443 2444 /* 2445 * Reclaim/compaction is used for high-order allocation requests. It reclaims 2446 * order-0 pages before compacting the zone. should_continue_reclaim() returns 2447 * true if more pages should be reclaimed such that when the page allocator 2448 * calls try_to_compact_zone() that it will have enough free pages to succeed. 2449 * It will give up earlier than that if there is difficulty reclaiming pages. 2450 */ 2451 static inline bool should_continue_reclaim(struct pglist_data *pgdat, 2452 unsigned long nr_reclaimed, 2453 unsigned long nr_scanned, 2454 struct scan_control *sc) 2455 { 2456 unsigned long pages_for_compaction; 2457 unsigned long inactive_lru_pages; 2458 int z; 2459 2460 /* If not in reclaim/compaction mode, stop */ 2461 if (!in_reclaim_compaction(sc)) 2462 return false; 2463 2464 /* Consider stopping depending on scan and reclaim activity */ 2465 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) { 2466 /* 2467 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the 2468 * full LRU list has been scanned and we are still failing 2469 * to reclaim pages. This full LRU scan is potentially 2470 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed 2471 */ 2472 if (!nr_reclaimed && !nr_scanned) 2473 return false; 2474 } else { 2475 /* 2476 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably 2477 * fail without consequence, stop if we failed to reclaim 2478 * any pages from the last SWAP_CLUSTER_MAX number of 2479 * pages that were scanned. This will return to the 2480 * caller faster at the risk reclaim/compaction and 2481 * the resulting allocation attempt fails 2482 */ 2483 if (!nr_reclaimed) 2484 return false; 2485 } 2486 2487 /* 2488 * If we have not reclaimed enough pages for compaction and the 2489 * inactive lists are large enough, continue reclaiming 2490 */ 2491 pages_for_compaction = compact_gap(sc->order); 2492 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); 2493 if (get_nr_swap_pages() > 0) 2494 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); 2495 if (sc->nr_reclaimed < pages_for_compaction && 2496 inactive_lru_pages > pages_for_compaction) 2497 return true; 2498 2499 /* If compaction would go ahead or the allocation would succeed, stop */ 2500 for (z = 0; z <= sc->reclaim_idx; z++) { 2501 struct zone *zone = &pgdat->node_zones[z]; 2502 if (!managed_zone(zone)) 2503 continue; 2504 2505 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { 2506 case COMPACT_SUCCESS: 2507 case COMPACT_CONTINUE: 2508 return false; 2509 default: 2510 /* check next zone */ 2511 ; 2512 } 2513 } 2514 return true; 2515 } 2516 2517 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) 2518 { 2519 struct reclaim_state *reclaim_state = current->reclaim_state; 2520 unsigned long nr_reclaimed, nr_scanned; 2521 bool reclaimable = false; 2522 2523 do { 2524 struct mem_cgroup *root = sc->target_mem_cgroup; 2525 struct mem_cgroup_reclaim_cookie reclaim = { 2526 .pgdat = pgdat, 2527 .priority = sc->priority, 2528 }; 2529 unsigned long node_lru_pages = 0; 2530 struct mem_cgroup *memcg; 2531 2532 nr_reclaimed = sc->nr_reclaimed; 2533 nr_scanned = sc->nr_scanned; 2534 2535 memcg = mem_cgroup_iter(root, NULL, &reclaim); 2536 do { 2537 unsigned long lru_pages; 2538 unsigned long reclaimed; 2539 unsigned long scanned; 2540 2541 if (mem_cgroup_low(root, memcg)) { 2542 if (!sc->memcg_low_reclaim) { 2543 sc->memcg_low_skipped = 1; 2544 continue; 2545 } 2546 mem_cgroup_event(memcg, MEMCG_LOW); 2547 } 2548 2549 reclaimed = sc->nr_reclaimed; 2550 scanned = sc->nr_scanned; 2551 shrink_node_memcg(pgdat, memcg, sc, &lru_pages); 2552 node_lru_pages += lru_pages; 2553 2554 if (memcg) 2555 shrink_slab(sc->gfp_mask, pgdat->node_id, 2556 memcg, sc->priority); 2557 2558 /* Record the group's reclaim efficiency */ 2559 vmpressure(sc->gfp_mask, memcg, false, 2560 sc->nr_scanned - scanned, 2561 sc->nr_reclaimed - reclaimed); 2562 2563 /* 2564 * Direct reclaim and kswapd have to scan all memory 2565 * cgroups to fulfill the overall scan target for the 2566 * node. 2567 * 2568 * Limit reclaim, on the other hand, only cares about 2569 * nr_to_reclaim pages to be reclaimed and it will 2570 * retry with decreasing priority if one round over the 2571 * whole hierarchy is not sufficient. 2572 */ 2573 if (!global_reclaim(sc) && 2574 sc->nr_reclaimed >= sc->nr_to_reclaim) { 2575 mem_cgroup_iter_break(root, memcg); 2576 break; 2577 } 2578 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); 2579 2580 if (global_reclaim(sc)) 2581 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL, 2582 sc->priority); 2583 2584 if (reclaim_state) { 2585 sc->nr_reclaimed += reclaim_state->reclaimed_slab; 2586 reclaim_state->reclaimed_slab = 0; 2587 } 2588 2589 /* Record the subtree's reclaim efficiency */ 2590 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, 2591 sc->nr_scanned - nr_scanned, 2592 sc->nr_reclaimed - nr_reclaimed); 2593 2594 if (sc->nr_reclaimed - nr_reclaimed) 2595 reclaimable = true; 2596 2597 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, 2598 sc->nr_scanned - nr_scanned, sc)); 2599 2600 /* 2601 * Kswapd gives up on balancing particular nodes after too 2602 * many failures to reclaim anything from them and goes to 2603 * sleep. On reclaim progress, reset the failure counter. A 2604 * successful direct reclaim run will revive a dormant kswapd. 2605 */ 2606 if (reclaimable) 2607 pgdat->kswapd_failures = 0; 2608 2609 return reclaimable; 2610 } 2611 2612 /* 2613 * Returns true if compaction should go ahead for a costly-order request, or 2614 * the allocation would already succeed without compaction. Return false if we 2615 * should reclaim first. 2616 */ 2617 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) 2618 { 2619 unsigned long watermark; 2620 enum compact_result suitable; 2621 2622 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); 2623 if (suitable == COMPACT_SUCCESS) 2624 /* Allocation should succeed already. Don't reclaim. */ 2625 return true; 2626 if (suitable == COMPACT_SKIPPED) 2627 /* Compaction cannot yet proceed. Do reclaim. */ 2628 return false; 2629 2630 /* 2631 * Compaction is already possible, but it takes time to run and there 2632 * are potentially other callers using the pages just freed. So proceed 2633 * with reclaim to make a buffer of free pages available to give 2634 * compaction a reasonable chance of completing and allocating the page. 2635 * Note that we won't actually reclaim the whole buffer in one attempt 2636 * as the target watermark in should_continue_reclaim() is lower. But if 2637 * we are already above the high+gap watermark, don't reclaim at all. 2638 */ 2639 watermark = high_wmark_pages(zone) + compact_gap(sc->order); 2640 2641 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); 2642 } 2643 2644 /* 2645 * This is the direct reclaim path, for page-allocating processes. We only 2646 * try to reclaim pages from zones which will satisfy the caller's allocation 2647 * request. 2648 * 2649 * If a zone is deemed to be full of pinned pages then just give it a light 2650 * scan then give up on it. 2651 */ 2652 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) 2653 { 2654 struct zoneref *z; 2655 struct zone *zone; 2656 unsigned long nr_soft_reclaimed; 2657 unsigned long nr_soft_scanned; 2658 gfp_t orig_mask; 2659 pg_data_t *last_pgdat = NULL; 2660 2661 /* 2662 * If the number of buffer_heads in the machine exceeds the maximum 2663 * allowed level, force direct reclaim to scan the highmem zone as 2664 * highmem pages could be pinning lowmem pages storing buffer_heads 2665 */ 2666 orig_mask = sc->gfp_mask; 2667 if (buffer_heads_over_limit) { 2668 sc->gfp_mask |= __GFP_HIGHMEM; 2669 sc->reclaim_idx = gfp_zone(sc->gfp_mask); 2670 } 2671 2672 for_each_zone_zonelist_nodemask(zone, z, zonelist, 2673 sc->reclaim_idx, sc->nodemask) { 2674 /* 2675 * Take care memory controller reclaiming has small influence 2676 * to global LRU. 2677 */ 2678 if (global_reclaim(sc)) { 2679 if (!cpuset_zone_allowed(zone, 2680 GFP_KERNEL | __GFP_HARDWALL)) 2681 continue; 2682 2683 /* 2684 * If we already have plenty of memory free for 2685 * compaction in this zone, don't free any more. 2686 * Even though compaction is invoked for any 2687 * non-zero order, only frequent costly order 2688 * reclamation is disruptive enough to become a 2689 * noticeable problem, like transparent huge 2690 * page allocations. 2691 */ 2692 if (IS_ENABLED(CONFIG_COMPACTION) && 2693 sc->order > PAGE_ALLOC_COSTLY_ORDER && 2694 compaction_ready(zone, sc)) { 2695 sc->compaction_ready = true; 2696 continue; 2697 } 2698 2699 /* 2700 * Shrink each node in the zonelist once. If the 2701 * zonelist is ordered by zone (not the default) then a 2702 * node may be shrunk multiple times but in that case 2703 * the user prefers lower zones being preserved. 2704 */ 2705 if (zone->zone_pgdat == last_pgdat) 2706 continue; 2707 2708 /* 2709 * This steals pages from memory cgroups over softlimit 2710 * and returns the number of reclaimed pages and 2711 * scanned pages. This works for global memory pressure 2712 * and balancing, not for a memcg's limit. 2713 */ 2714 nr_soft_scanned = 0; 2715 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, 2716 sc->order, sc->gfp_mask, 2717 &nr_soft_scanned); 2718 sc->nr_reclaimed += nr_soft_reclaimed; 2719 sc->nr_scanned += nr_soft_scanned; 2720 /* need some check for avoid more shrink_zone() */ 2721 } 2722 2723 /* See comment about same check for global reclaim above */ 2724 if (zone->zone_pgdat == last_pgdat) 2725 continue; 2726 last_pgdat = zone->zone_pgdat; 2727 shrink_node(zone->zone_pgdat, sc); 2728 } 2729 2730 /* 2731 * Restore to original mask to avoid the impact on the caller if we 2732 * promoted it to __GFP_HIGHMEM. 2733 */ 2734 sc->gfp_mask = orig_mask; 2735 } 2736 2737 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat) 2738 { 2739 struct mem_cgroup *memcg; 2740 2741 memcg = mem_cgroup_iter(root_memcg, NULL, NULL); 2742 do { 2743 unsigned long refaults; 2744 struct lruvec *lruvec; 2745 2746 if (memcg) 2747 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); 2748 else 2749 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); 2750 2751 lruvec = mem_cgroup_lruvec(pgdat, memcg); 2752 lruvec->refaults = refaults; 2753 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL))); 2754 } 2755 2756 /* 2757 * This is the main entry point to direct page reclaim. 2758 * 2759 * If a full scan of the inactive list fails to free enough memory then we 2760 * are "out of memory" and something needs to be killed. 2761 * 2762 * If the caller is !__GFP_FS then the probability of a failure is reasonably 2763 * high - the zone may be full of dirty or under-writeback pages, which this 2764 * caller can't do much about. We kick the writeback threads and take explicit 2765 * naps in the hope that some of these pages can be written. But if the 2766 * allocating task holds filesystem locks which prevent writeout this might not 2767 * work, and the allocation attempt will fail. 2768 * 2769 * returns: 0, if no pages reclaimed 2770 * else, the number of pages reclaimed 2771 */ 2772 static unsigned long do_try_to_free_pages(struct zonelist *zonelist, 2773 struct scan_control *sc) 2774 { 2775 int initial_priority = sc->priority; 2776 pg_data_t *last_pgdat; 2777 struct zoneref *z; 2778 struct zone *zone; 2779 retry: 2780 delayacct_freepages_start(); 2781 2782 if (global_reclaim(sc)) 2783 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); 2784 2785 do { 2786 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, 2787 sc->priority); 2788 sc->nr_scanned = 0; 2789 shrink_zones(zonelist, sc); 2790 2791 if (sc->nr_reclaimed >= sc->nr_to_reclaim) 2792 break; 2793 2794 if (sc->compaction_ready) 2795 break; 2796 2797 /* 2798 * If we're getting trouble reclaiming, start doing 2799 * writepage even in laptop mode. 2800 */ 2801 if (sc->priority < DEF_PRIORITY - 2) 2802 sc->may_writepage = 1; 2803 } while (--sc->priority >= 0); 2804 2805 last_pgdat = NULL; 2806 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, 2807 sc->nodemask) { 2808 if (zone->zone_pgdat == last_pgdat) 2809 continue; 2810 last_pgdat = zone->zone_pgdat; 2811 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); 2812 } 2813 2814 delayacct_freepages_end(); 2815 2816 if (sc->nr_reclaimed) 2817 return sc->nr_reclaimed; 2818 2819 /* Aborted reclaim to try compaction? don't OOM, then */ 2820 if (sc->compaction_ready) 2821 return 1; 2822 2823 /* Untapped cgroup reserves? Don't OOM, retry. */ 2824 if (sc->memcg_low_skipped) { 2825 sc->priority = initial_priority; 2826 sc->memcg_low_reclaim = 1; 2827 sc->memcg_low_skipped = 0; 2828 goto retry; 2829 } 2830 2831 return 0; 2832 } 2833 2834 static bool allow_direct_reclaim(pg_data_t *pgdat) 2835 { 2836 struct zone *zone; 2837 unsigned long pfmemalloc_reserve = 0; 2838 unsigned long free_pages = 0; 2839 int i; 2840 bool wmark_ok; 2841 2842 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) 2843 return true; 2844 2845 for (i = 0; i <= ZONE_NORMAL; i++) { 2846 zone = &pgdat->node_zones[i]; 2847 if (!managed_zone(zone)) 2848 continue; 2849 2850 if (!zone_reclaimable_pages(zone)) 2851 continue; 2852 2853 pfmemalloc_reserve += min_wmark_pages(zone); 2854 free_pages += zone_page_state(zone, NR_FREE_PAGES); 2855 } 2856 2857 /* If there are no reserves (unexpected config) then do not throttle */ 2858 if (!pfmemalloc_reserve) 2859 return true; 2860 2861 wmark_ok = free_pages > pfmemalloc_reserve / 2; 2862 2863 /* kswapd must be awake if processes are being throttled */ 2864 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { 2865 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, 2866 (enum zone_type)ZONE_NORMAL); 2867 wake_up_interruptible(&pgdat->kswapd_wait); 2868 } 2869 2870 return wmark_ok; 2871 } 2872 2873 /* 2874 * Throttle direct reclaimers if backing storage is backed by the network 2875 * and the PFMEMALLOC reserve for the preferred node is getting dangerously 2876 * depleted. kswapd will continue to make progress and wake the processes 2877 * when the low watermark is reached. 2878 * 2879 * Returns true if a fatal signal was delivered during throttling. If this 2880 * happens, the page allocator should not consider triggering the OOM killer. 2881 */ 2882 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, 2883 nodemask_t *nodemask) 2884 { 2885 struct zoneref *z; 2886 struct zone *zone; 2887 pg_data_t *pgdat = NULL; 2888 2889 /* 2890 * Kernel threads should not be throttled as they may be indirectly 2891 * responsible for cleaning pages necessary for reclaim to make forward 2892 * progress. kjournald for example may enter direct reclaim while 2893 * committing a transaction where throttling it could forcing other 2894 * processes to block on log_wait_commit(). 2895 */ 2896 if (current->flags & PF_KTHREAD) 2897 goto out; 2898 2899 /* 2900 * If a fatal signal is pending, this process should not throttle. 2901 * It should return quickly so it can exit and free its memory 2902 */ 2903 if (fatal_signal_pending(current)) 2904 goto out; 2905 2906 /* 2907 * Check if the pfmemalloc reserves are ok by finding the first node 2908 * with a usable ZONE_NORMAL or lower zone. The expectation is that 2909 * GFP_KERNEL will be required for allocating network buffers when 2910 * swapping over the network so ZONE_HIGHMEM is unusable. 2911 * 2912 * Throttling is based on the first usable node and throttled processes 2913 * wait on a queue until kswapd makes progress and wakes them. There 2914 * is an affinity then between processes waking up and where reclaim 2915 * progress has been made assuming the process wakes on the same node. 2916 * More importantly, processes running on remote nodes will not compete 2917 * for remote pfmemalloc reserves and processes on different nodes 2918 * should make reasonable progress. 2919 */ 2920 for_each_zone_zonelist_nodemask(zone, z, zonelist, 2921 gfp_zone(gfp_mask), nodemask) { 2922 if (zone_idx(zone) > ZONE_NORMAL) 2923 continue; 2924 2925 /* Throttle based on the first usable node */ 2926 pgdat = zone->zone_pgdat; 2927 if (allow_direct_reclaim(pgdat)) 2928 goto out; 2929 break; 2930 } 2931 2932 /* If no zone was usable by the allocation flags then do not throttle */ 2933 if (!pgdat) 2934 goto out; 2935 2936 /* Account for the throttling */ 2937 count_vm_event(PGSCAN_DIRECT_THROTTLE); 2938 2939 /* 2940 * If the caller cannot enter the filesystem, it's possible that it 2941 * is due to the caller holding an FS lock or performing a journal 2942 * transaction in the case of a filesystem like ext[3|4]. In this case, 2943 * it is not safe to block on pfmemalloc_wait as kswapd could be 2944 * blocked waiting on the same lock. Instead, throttle for up to a 2945 * second before continuing. 2946 */ 2947 if (!(gfp_mask & __GFP_FS)) { 2948 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, 2949 allow_direct_reclaim(pgdat), HZ); 2950 2951 goto check_pending; 2952 } 2953 2954 /* Throttle until kswapd wakes the process */ 2955 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, 2956 allow_direct_reclaim(pgdat)); 2957 2958 check_pending: 2959 if (fatal_signal_pending(current)) 2960 return true; 2961 2962 out: 2963 return false; 2964 } 2965 2966 unsigned long try_to_free_pages(struct zonelist *zonelist, int order, 2967 gfp_t gfp_mask, nodemask_t *nodemask) 2968 { 2969 unsigned long nr_reclaimed; 2970 struct scan_control sc = { 2971 .nr_to_reclaim = SWAP_CLUSTER_MAX, 2972 .gfp_mask = current_gfp_context(gfp_mask), 2973 .reclaim_idx = gfp_zone(gfp_mask), 2974 .order = order, 2975 .nodemask = nodemask, 2976 .priority = DEF_PRIORITY, 2977 .may_writepage = !laptop_mode, 2978 .may_unmap = 1, 2979 .may_swap = 1, 2980 }; 2981 2982 /* 2983 * Do not enter reclaim if fatal signal was delivered while throttled. 2984 * 1 is returned so that the page allocator does not OOM kill at this 2985 * point. 2986 */ 2987 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) 2988 return 1; 2989 2990 trace_mm_vmscan_direct_reclaim_begin(order, 2991 sc.may_writepage, 2992 sc.gfp_mask, 2993 sc.reclaim_idx); 2994 2995 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 2996 2997 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); 2998 2999 return nr_reclaimed; 3000 } 3001 3002 #ifdef CONFIG_MEMCG 3003 3004 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, 3005 gfp_t gfp_mask, bool noswap, 3006 pg_data_t *pgdat, 3007 unsigned long *nr_scanned) 3008 { 3009 struct scan_control sc = { 3010 .nr_to_reclaim = SWAP_CLUSTER_MAX, 3011 .target_mem_cgroup = memcg, 3012 .may_writepage = !laptop_mode, 3013 .may_unmap = 1, 3014 .reclaim_idx = MAX_NR_ZONES - 1, 3015 .may_swap = !noswap, 3016 }; 3017 unsigned long lru_pages; 3018 3019 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | 3020 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); 3021 3022 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, 3023 sc.may_writepage, 3024 sc.gfp_mask, 3025 sc.reclaim_idx); 3026 3027 /* 3028 * NOTE: Although we can get the priority field, using it 3029 * here is not a good idea, since it limits the pages we can scan. 3030 * if we don't reclaim here, the shrink_node from balance_pgdat 3031 * will pick up pages from other mem cgroup's as well. We hack 3032 * the priority and make it zero. 3033 */ 3034 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages); 3035 3036 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); 3037 3038 *nr_scanned = sc.nr_scanned; 3039 return sc.nr_reclaimed; 3040 } 3041 3042 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, 3043 unsigned long nr_pages, 3044 gfp_t gfp_mask, 3045 bool may_swap) 3046 { 3047 struct zonelist *zonelist; 3048 unsigned long nr_reclaimed; 3049 int nid; 3050 unsigned int noreclaim_flag; 3051 struct scan_control sc = { 3052 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), 3053 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | 3054 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), 3055 .reclaim_idx = MAX_NR_ZONES - 1, 3056 .target_mem_cgroup = memcg, 3057 .priority = DEF_PRIORITY, 3058 .may_writepage = !laptop_mode, 3059 .may_unmap = 1, 3060 .may_swap = may_swap, 3061 }; 3062 3063 /* 3064 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't 3065 * take care of from where we get pages. So the node where we start the 3066 * scan does not need to be the current node. 3067 */ 3068 nid = mem_cgroup_select_victim_node(memcg); 3069 3070 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK]; 3071 3072 trace_mm_vmscan_memcg_reclaim_begin(0, 3073 sc.may_writepage, 3074 sc.gfp_mask, 3075 sc.reclaim_idx); 3076 3077 noreclaim_flag = memalloc_noreclaim_save(); 3078 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 3079 memalloc_noreclaim_restore(noreclaim_flag); 3080 3081 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); 3082 3083 return nr_reclaimed; 3084 } 3085 #endif 3086 3087 static void age_active_anon(struct pglist_data *pgdat, 3088 struct scan_control *sc) 3089 { 3090 struct mem_cgroup *memcg; 3091 3092 if (!total_swap_pages) 3093 return; 3094 3095 memcg = mem_cgroup_iter(NULL, NULL, NULL); 3096 do { 3097 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); 3098 3099 if (inactive_list_is_low(lruvec, false, memcg, sc, true)) 3100 shrink_active_list(SWAP_CLUSTER_MAX, lruvec, 3101 sc, LRU_ACTIVE_ANON); 3102 3103 memcg = mem_cgroup_iter(NULL, memcg, NULL); 3104 } while (memcg); 3105 } 3106 3107 /* 3108 * Returns true if there is an eligible zone balanced for the request order 3109 * and classzone_idx 3110 */ 3111 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) 3112 { 3113 int i; 3114 unsigned long mark = -1; 3115 struct zone *zone; 3116 3117 for (i = 0; i <= classzone_idx; i++) { 3118 zone = pgdat->node_zones + i; 3119 3120 if (!managed_zone(zone)) 3121 continue; 3122 3123 mark = high_wmark_pages(zone); 3124 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx)) 3125 return true; 3126 } 3127 3128 /* 3129 * If a node has no populated zone within classzone_idx, it does not 3130 * need balancing by definition. This can happen if a zone-restricted 3131 * allocation tries to wake a remote kswapd. 3132 */ 3133 if (mark == -1) 3134 return true; 3135 3136 return false; 3137 } 3138 3139 /* Clear pgdat state for congested, dirty or under writeback. */ 3140 static void clear_pgdat_congested(pg_data_t *pgdat) 3141 { 3142 clear_bit(PGDAT_CONGESTED, &pgdat->flags); 3143 clear_bit(PGDAT_DIRTY, &pgdat->flags); 3144 clear_bit(PGDAT_WRITEBACK, &pgdat->flags); 3145 } 3146 3147 /* 3148 * Prepare kswapd for sleeping. This verifies that there are no processes 3149 * waiting in throttle_direct_reclaim() and that watermarks have been met. 3150 * 3151 * Returns true if kswapd is ready to sleep 3152 */ 3153 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) 3154 { 3155 /* 3156 * The throttled processes are normally woken up in balance_pgdat() as 3157 * soon as allow_direct_reclaim() is true. But there is a potential 3158 * race between when kswapd checks the watermarks and a process gets 3159 * throttled. There is also a potential race if processes get 3160 * throttled, kswapd wakes, a large process exits thereby balancing the 3161 * zones, which causes kswapd to exit balance_pgdat() before reaching 3162 * the wake up checks. If kswapd is going to sleep, no process should 3163 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If 3164 * the wake up is premature, processes will wake kswapd and get 3165 * throttled again. The difference from wake ups in balance_pgdat() is 3166 * that here we are under prepare_to_wait(). 3167 */ 3168 if (waitqueue_active(&pgdat->pfmemalloc_wait)) 3169 wake_up_all(&pgdat->pfmemalloc_wait); 3170 3171 /* Hopeless node, leave it to direct reclaim */ 3172 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) 3173 return true; 3174 3175 if (pgdat_balanced(pgdat, order, classzone_idx)) { 3176 clear_pgdat_congested(pgdat); 3177 return true; 3178 } 3179 3180 return false; 3181 } 3182 3183 /* 3184 * kswapd shrinks a node of pages that are at or below the highest usable 3185 * zone that is currently unbalanced. 3186 * 3187 * Returns true if kswapd scanned at least the requested number of pages to 3188 * reclaim or if the lack of progress was due to pages under writeback. 3189 * This is used to determine if the scanning priority needs to be raised. 3190 */ 3191 static bool kswapd_shrink_node(pg_data_t *pgdat, 3192 struct scan_control *sc) 3193 { 3194 struct zone *zone; 3195 int z; 3196 3197 /* Reclaim a number of pages proportional to the number of zones */ 3198 sc->nr_to_reclaim = 0; 3199 for (z = 0; z <= sc->reclaim_idx; z++) { 3200 zone = pgdat->node_zones + z; 3201 if (!managed_zone(zone)) 3202 continue; 3203 3204 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); 3205 } 3206 3207 /* 3208 * Historically care was taken to put equal pressure on all zones but 3209 * now pressure is applied based on node LRU order. 3210 */ 3211 shrink_node(pgdat, sc); 3212 3213 /* 3214 * Fragmentation may mean that the system cannot be rebalanced for 3215 * high-order allocations. If twice the allocation size has been 3216 * reclaimed then recheck watermarks only at order-0 to prevent 3217 * excessive reclaim. Assume that a process requested a high-order 3218 * can direct reclaim/compact. 3219 */ 3220 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) 3221 sc->order = 0; 3222 3223 return sc->nr_scanned >= sc->nr_to_reclaim; 3224 } 3225 3226 /* 3227 * For kswapd, balance_pgdat() will reclaim pages across a node from zones 3228 * that are eligible for use by the caller until at least one zone is 3229 * balanced. 3230 * 3231 * Returns the order kswapd finished reclaiming at. 3232 * 3233 * kswapd scans the zones in the highmem->normal->dma direction. It skips 3234 * zones which have free_pages > high_wmark_pages(zone), but once a zone is 3235 * found to have free_pages <= high_wmark_pages(zone), any page is that zone 3236 * or lower is eligible for reclaim until at least one usable zone is 3237 * balanced. 3238 */ 3239 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) 3240 { 3241 int i; 3242 unsigned long nr_soft_reclaimed; 3243 unsigned long nr_soft_scanned; 3244 struct zone *zone; 3245 struct scan_control sc = { 3246 .gfp_mask = GFP_KERNEL, 3247 .order = order, 3248 .priority = DEF_PRIORITY, 3249 .may_writepage = !laptop_mode, 3250 .may_unmap = 1, 3251 .may_swap = 1, 3252 }; 3253 count_vm_event(PAGEOUTRUN); 3254 3255 do { 3256 unsigned long nr_reclaimed = sc.nr_reclaimed; 3257 bool raise_priority = true; 3258 3259 sc.reclaim_idx = classzone_idx; 3260 3261 /* 3262 * If the number of buffer_heads exceeds the maximum allowed 3263 * then consider reclaiming from all zones. This has a dual 3264 * purpose -- on 64-bit systems it is expected that 3265 * buffer_heads are stripped during active rotation. On 32-bit 3266 * systems, highmem pages can pin lowmem memory and shrinking 3267 * buffers can relieve lowmem pressure. Reclaim may still not 3268 * go ahead if all eligible zones for the original allocation 3269 * request are balanced to avoid excessive reclaim from kswapd. 3270 */ 3271 if (buffer_heads_over_limit) { 3272 for (i = MAX_NR_ZONES - 1; i >= 0; i--) { 3273 zone = pgdat->node_zones + i; 3274 if (!managed_zone(zone)) 3275 continue; 3276 3277 sc.reclaim_idx = i; 3278 break; 3279 } 3280 } 3281 3282 /* 3283 * Only reclaim if there are no eligible zones. Note that 3284 * sc.reclaim_idx is not used as buffer_heads_over_limit may 3285 * have adjusted it. 3286 */ 3287 if (pgdat_balanced(pgdat, sc.order, classzone_idx)) 3288 goto out; 3289 3290 /* 3291 * Do some background aging of the anon list, to give 3292 * pages a chance to be referenced before reclaiming. All 3293 * pages are rotated regardless of classzone as this is 3294 * about consistent aging. 3295 */ 3296 age_active_anon(pgdat, &sc); 3297 3298 /* 3299 * If we're getting trouble reclaiming, start doing writepage 3300 * even in laptop mode. 3301 */ 3302 if (sc.priority < DEF_PRIORITY - 2) 3303 sc.may_writepage = 1; 3304 3305 /* Call soft limit reclaim before calling shrink_node. */ 3306 sc.nr_scanned = 0; 3307 nr_soft_scanned = 0; 3308 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, 3309 sc.gfp_mask, &nr_soft_scanned); 3310 sc.nr_reclaimed += nr_soft_reclaimed; 3311 3312 /* 3313 * There should be no need to raise the scanning priority if 3314 * enough pages are already being scanned that that high 3315 * watermark would be met at 100% efficiency. 3316 */ 3317 if (kswapd_shrink_node(pgdat, &sc)) 3318 raise_priority = false; 3319 3320 /* 3321 * If the low watermark is met there is no need for processes 3322 * to be throttled on pfmemalloc_wait as they should not be 3323 * able to safely make forward progress. Wake them 3324 */ 3325 if (waitqueue_active(&pgdat->pfmemalloc_wait) && 3326 allow_direct_reclaim(pgdat)) 3327 wake_up_all(&pgdat->pfmemalloc_wait); 3328 3329 /* Check if kswapd should be suspending */ 3330 if (try_to_freeze() || kthread_should_stop()) 3331 break; 3332 3333 /* 3334 * Raise priority if scanning rate is too low or there was no 3335 * progress in reclaiming pages 3336 */ 3337 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; 3338 if (raise_priority || !nr_reclaimed) 3339 sc.priority--; 3340 } while (sc.priority >= 1); 3341 3342 if (!sc.nr_reclaimed) 3343 pgdat->kswapd_failures++; 3344 3345 out: 3346 snapshot_refaults(NULL, pgdat); 3347 /* 3348 * Return the order kswapd stopped reclaiming at as 3349 * prepare_kswapd_sleep() takes it into account. If another caller 3350 * entered the allocator slow path while kswapd was awake, order will 3351 * remain at the higher level. 3352 */ 3353 return sc.order; 3354 } 3355 3356 /* 3357 * pgdat->kswapd_classzone_idx is the highest zone index that a recent 3358 * allocation request woke kswapd for. When kswapd has not woken recently, 3359 * the value is MAX_NR_ZONES which is not a valid index. This compares a 3360 * given classzone and returns it or the highest classzone index kswapd 3361 * was recently woke for. 3362 */ 3363 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat, 3364 enum zone_type classzone_idx) 3365 { 3366 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) 3367 return classzone_idx; 3368 3369 return max(pgdat->kswapd_classzone_idx, classzone_idx); 3370 } 3371 3372 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, 3373 unsigned int classzone_idx) 3374 { 3375 long remaining = 0; 3376 DEFINE_WAIT(wait); 3377 3378 if (freezing(current) || kthread_should_stop()) 3379 return; 3380 3381 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 3382 3383 /* 3384 * Try to sleep for a short interval. Note that kcompactd will only be 3385 * woken if it is possible to sleep for a short interval. This is 3386 * deliberate on the assumption that if reclaim cannot keep an 3387 * eligible zone balanced that it's also unlikely that compaction will 3388 * succeed. 3389 */ 3390 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { 3391 /* 3392 * Compaction records what page blocks it recently failed to 3393 * isolate pages from and skips them in the future scanning. 3394 * When kswapd is going to sleep, it is reasonable to assume 3395 * that pages and compaction may succeed so reset the cache. 3396 */ 3397 reset_isolation_suitable(pgdat); 3398 3399 /* 3400 * We have freed the memory, now we should compact it to make 3401 * allocation of the requested order possible. 3402 */ 3403 wakeup_kcompactd(pgdat, alloc_order, classzone_idx); 3404 3405 remaining = schedule_timeout(HZ/10); 3406 3407 /* 3408 * If woken prematurely then reset kswapd_classzone_idx and 3409 * order. The values will either be from a wakeup request or 3410 * the previous request that slept prematurely. 3411 */ 3412 if (remaining) { 3413 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); 3414 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); 3415 } 3416 3417 finish_wait(&pgdat->kswapd_wait, &wait); 3418 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 3419 } 3420 3421 /* 3422 * After a short sleep, check if it was a premature sleep. If not, then 3423 * go fully to sleep until explicitly woken up. 3424 */ 3425 if (!remaining && 3426 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { 3427 trace_mm_vmscan_kswapd_sleep(pgdat->node_id); 3428 3429 /* 3430 * vmstat counters are not perfectly accurate and the estimated 3431 * value for counters such as NR_FREE_PAGES can deviate from the 3432 * true value by nr_online_cpus * threshold. To avoid the zone 3433 * watermarks being breached while under pressure, we reduce the 3434 * per-cpu vmstat threshold while kswapd is awake and restore 3435 * them before going back to sleep. 3436 */ 3437 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); 3438 3439 if (!kthread_should_stop()) 3440 schedule(); 3441 3442 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); 3443 } else { 3444 if (remaining) 3445 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); 3446 else 3447 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); 3448 } 3449 finish_wait(&pgdat->kswapd_wait, &wait); 3450 } 3451 3452 /* 3453 * The background pageout daemon, started as a kernel thread 3454 * from the init process. 3455 * 3456 * This basically trickles out pages so that we have _some_ 3457 * free memory available even if there is no other activity 3458 * that frees anything up. This is needed for things like routing 3459 * etc, where we otherwise might have all activity going on in 3460 * asynchronous contexts that cannot page things out. 3461 * 3462 * If there are applications that are active memory-allocators 3463 * (most normal use), this basically shouldn't matter. 3464 */ 3465 static int kswapd(void *p) 3466 { 3467 unsigned int alloc_order, reclaim_order; 3468 unsigned int classzone_idx = MAX_NR_ZONES - 1; 3469 pg_data_t *pgdat = (pg_data_t*)p; 3470 struct task_struct *tsk = current; 3471 3472 struct reclaim_state reclaim_state = { 3473 .reclaimed_slab = 0, 3474 }; 3475 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); 3476 3477 if (!cpumask_empty(cpumask)) 3478 set_cpus_allowed_ptr(tsk, cpumask); 3479 current->reclaim_state = &reclaim_state; 3480 3481 /* 3482 * Tell the memory management that we're a "memory allocator", 3483 * and that if we need more memory we should get access to it 3484 * regardless (see "__alloc_pages()"). "kswapd" should 3485 * never get caught in the normal page freeing logic. 3486 * 3487 * (Kswapd normally doesn't need memory anyway, but sometimes 3488 * you need a small amount of memory in order to be able to 3489 * page out something else, and this flag essentially protects 3490 * us from recursively trying to free more memory as we're 3491 * trying to free the first piece of memory in the first place). 3492 */ 3493 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; 3494 set_freezable(); 3495 3496 pgdat->kswapd_order = 0; 3497 pgdat->kswapd_classzone_idx = MAX_NR_ZONES; 3498 for ( ; ; ) { 3499 bool ret; 3500 3501 alloc_order = reclaim_order = pgdat->kswapd_order; 3502 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); 3503 3504 kswapd_try_sleep: 3505 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, 3506 classzone_idx); 3507 3508 /* Read the new order and classzone_idx */ 3509 alloc_order = reclaim_order = pgdat->kswapd_order; 3510 classzone_idx = kswapd_classzone_idx(pgdat, 0); 3511 pgdat->kswapd_order = 0; 3512 pgdat->kswapd_classzone_idx = MAX_NR_ZONES; 3513 3514 ret = try_to_freeze(); 3515 if (kthread_should_stop()) 3516 break; 3517 3518 /* 3519 * We can speed up thawing tasks if we don't call balance_pgdat 3520 * after returning from the refrigerator 3521 */ 3522 if (ret) 3523 continue; 3524 3525 /* 3526 * Reclaim begins at the requested order but if a high-order 3527 * reclaim fails then kswapd falls back to reclaiming for 3528 * order-0. If that happens, kswapd will consider sleeping 3529 * for the order it finished reclaiming at (reclaim_order) 3530 * but kcompactd is woken to compact for the original 3531 * request (alloc_order). 3532 */ 3533 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, 3534 alloc_order); 3535 fs_reclaim_acquire(GFP_KERNEL); 3536 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); 3537 fs_reclaim_release(GFP_KERNEL); 3538 if (reclaim_order < alloc_order) 3539 goto kswapd_try_sleep; 3540 } 3541 3542 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); 3543 current->reclaim_state = NULL; 3544 3545 return 0; 3546 } 3547 3548 /* 3549 * A zone is low on free memory, so wake its kswapd task to service it. 3550 */ 3551 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx) 3552 { 3553 pg_data_t *pgdat; 3554 3555 if (!managed_zone(zone)) 3556 return; 3557 3558 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL)) 3559 return; 3560 pgdat = zone->zone_pgdat; 3561 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, 3562 classzone_idx); 3563 pgdat->kswapd_order = max(pgdat->kswapd_order, order); 3564 if (!waitqueue_active(&pgdat->kswapd_wait)) 3565 return; 3566 3567 /* Hopeless node, leave it to direct reclaim */ 3568 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) 3569 return; 3570 3571 if (pgdat_balanced(pgdat, order, classzone_idx)) 3572 return; 3573 3574 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order); 3575 wake_up_interruptible(&pgdat->kswapd_wait); 3576 } 3577 3578 #ifdef CONFIG_HIBERNATION 3579 /* 3580 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of 3581 * freed pages. 3582 * 3583 * Rather than trying to age LRUs the aim is to preserve the overall 3584 * LRU order by reclaiming preferentially 3585 * inactive > active > active referenced > active mapped 3586 */ 3587 unsigned long shrink_all_memory(unsigned long nr_to_reclaim) 3588 { 3589 struct reclaim_state reclaim_state; 3590 struct scan_control sc = { 3591 .nr_to_reclaim = nr_to_reclaim, 3592 .gfp_mask = GFP_HIGHUSER_MOVABLE, 3593 .reclaim_idx = MAX_NR_ZONES - 1, 3594 .priority = DEF_PRIORITY, 3595 .may_writepage = 1, 3596 .may_unmap = 1, 3597 .may_swap = 1, 3598 .hibernation_mode = 1, 3599 }; 3600 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); 3601 struct task_struct *p = current; 3602 unsigned long nr_reclaimed; 3603 unsigned int noreclaim_flag; 3604 3605 noreclaim_flag = memalloc_noreclaim_save(); 3606 fs_reclaim_acquire(sc.gfp_mask); 3607 reclaim_state.reclaimed_slab = 0; 3608 p->reclaim_state = &reclaim_state; 3609 3610 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 3611 3612 p->reclaim_state = NULL; 3613 fs_reclaim_release(sc.gfp_mask); 3614 memalloc_noreclaim_restore(noreclaim_flag); 3615 3616 return nr_reclaimed; 3617 } 3618 #endif /* CONFIG_HIBERNATION */ 3619 3620 /* It's optimal to keep kswapds on the same CPUs as their memory, but 3621 not required for correctness. So if the last cpu in a node goes 3622 away, we get changed to run anywhere: as the first one comes back, 3623 restore their cpu bindings. */ 3624 static int kswapd_cpu_online(unsigned int cpu) 3625 { 3626 int nid; 3627 3628 for_each_node_state(nid, N_MEMORY) { 3629 pg_data_t *pgdat = NODE_DATA(nid); 3630 const struct cpumask *mask; 3631 3632 mask = cpumask_of_node(pgdat->node_id); 3633 3634 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) 3635 /* One of our CPUs online: restore mask */ 3636 set_cpus_allowed_ptr(pgdat->kswapd, mask); 3637 } 3638 return 0; 3639 } 3640 3641 /* 3642 * This kswapd start function will be called by init and node-hot-add. 3643 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. 3644 */ 3645 int kswapd_run(int nid) 3646 { 3647 pg_data_t *pgdat = NODE_DATA(nid); 3648 int ret = 0; 3649 3650 if (pgdat->kswapd) 3651 return 0; 3652 3653 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); 3654 if (IS_ERR(pgdat->kswapd)) { 3655 /* failure at boot is fatal */ 3656 BUG_ON(system_state < SYSTEM_RUNNING); 3657 pr_err("Failed to start kswapd on node %d\n", nid); 3658 ret = PTR_ERR(pgdat->kswapd); 3659 pgdat->kswapd = NULL; 3660 } 3661 return ret; 3662 } 3663 3664 /* 3665 * Called by memory hotplug when all memory in a node is offlined. Caller must 3666 * hold mem_hotplug_begin/end(). 3667 */ 3668 void kswapd_stop(int nid) 3669 { 3670 struct task_struct *kswapd = NODE_DATA(nid)->kswapd; 3671 3672 if (kswapd) { 3673 kthread_stop(kswapd); 3674 NODE_DATA(nid)->kswapd = NULL; 3675 } 3676 } 3677 3678 static int __init kswapd_init(void) 3679 { 3680 int nid, ret; 3681 3682 swap_setup(); 3683 for_each_node_state(nid, N_MEMORY) 3684 kswapd_run(nid); 3685 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, 3686 "mm/vmscan:online", kswapd_cpu_online, 3687 NULL); 3688 WARN_ON(ret < 0); 3689 return 0; 3690 } 3691 3692 module_init(kswapd_init) 3693 3694 #ifdef CONFIG_NUMA 3695 /* 3696 * Node reclaim mode 3697 * 3698 * If non-zero call node_reclaim when the number of free pages falls below 3699 * the watermarks. 3700 */ 3701 int node_reclaim_mode __read_mostly; 3702 3703 #define RECLAIM_OFF 0 3704 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ 3705 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ 3706 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ 3707 3708 /* 3709 * Priority for NODE_RECLAIM. This determines the fraction of pages 3710 * of a node considered for each zone_reclaim. 4 scans 1/16th of 3711 * a zone. 3712 */ 3713 #define NODE_RECLAIM_PRIORITY 4 3714 3715 /* 3716 * Percentage of pages in a zone that must be unmapped for node_reclaim to 3717 * occur. 3718 */ 3719 int sysctl_min_unmapped_ratio = 1; 3720 3721 /* 3722 * If the number of slab pages in a zone grows beyond this percentage then 3723 * slab reclaim needs to occur. 3724 */ 3725 int sysctl_min_slab_ratio = 5; 3726 3727 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) 3728 { 3729 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); 3730 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + 3731 node_page_state(pgdat, NR_ACTIVE_FILE); 3732 3733 /* 3734 * It's possible for there to be more file mapped pages than 3735 * accounted for by the pages on the file LRU lists because 3736 * tmpfs pages accounted for as ANON can also be FILE_MAPPED 3737 */ 3738 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; 3739 } 3740 3741 /* Work out how many page cache pages we can reclaim in this reclaim_mode */ 3742 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) 3743 { 3744 unsigned long nr_pagecache_reclaimable; 3745 unsigned long delta = 0; 3746 3747 /* 3748 * If RECLAIM_UNMAP is set, then all file pages are considered 3749 * potentially reclaimable. Otherwise, we have to worry about 3750 * pages like swapcache and node_unmapped_file_pages() provides 3751 * a better estimate 3752 */ 3753 if (node_reclaim_mode & RECLAIM_UNMAP) 3754 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); 3755 else 3756 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); 3757 3758 /* If we can't clean pages, remove dirty pages from consideration */ 3759 if (!(node_reclaim_mode & RECLAIM_WRITE)) 3760 delta += node_page_state(pgdat, NR_FILE_DIRTY); 3761 3762 /* Watch for any possible underflows due to delta */ 3763 if (unlikely(delta > nr_pagecache_reclaimable)) 3764 delta = nr_pagecache_reclaimable; 3765 3766 return nr_pagecache_reclaimable - delta; 3767 } 3768 3769 /* 3770 * Try to free up some pages from this node through reclaim. 3771 */ 3772 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) 3773 { 3774 /* Minimum pages needed in order to stay on node */ 3775 const unsigned long nr_pages = 1 << order; 3776 struct task_struct *p = current; 3777 struct reclaim_state reclaim_state; 3778 unsigned int noreclaim_flag; 3779 struct scan_control sc = { 3780 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), 3781 .gfp_mask = current_gfp_context(gfp_mask), 3782 .order = order, 3783 .priority = NODE_RECLAIM_PRIORITY, 3784 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), 3785 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), 3786 .may_swap = 1, 3787 .reclaim_idx = gfp_zone(gfp_mask), 3788 }; 3789 3790 cond_resched(); 3791 /* 3792 * We need to be able to allocate from the reserves for RECLAIM_UNMAP 3793 * and we also need to be able to write out pages for RECLAIM_WRITE 3794 * and RECLAIM_UNMAP. 3795 */ 3796 noreclaim_flag = memalloc_noreclaim_save(); 3797 p->flags |= PF_SWAPWRITE; 3798 fs_reclaim_acquire(sc.gfp_mask); 3799 reclaim_state.reclaimed_slab = 0; 3800 p->reclaim_state = &reclaim_state; 3801 3802 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { 3803 /* 3804 * Free memory by calling shrink zone with increasing 3805 * priorities until we have enough memory freed. 3806 */ 3807 do { 3808 shrink_node(pgdat, &sc); 3809 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); 3810 } 3811 3812 p->reclaim_state = NULL; 3813 fs_reclaim_release(gfp_mask); 3814 current->flags &= ~PF_SWAPWRITE; 3815 memalloc_noreclaim_restore(noreclaim_flag); 3816 return sc.nr_reclaimed >= nr_pages; 3817 } 3818 3819 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) 3820 { 3821 int ret; 3822 3823 /* 3824 * Node reclaim reclaims unmapped file backed pages and 3825 * slab pages if we are over the defined limits. 3826 * 3827 * A small portion of unmapped file backed pages is needed for 3828 * file I/O otherwise pages read by file I/O will be immediately 3829 * thrown out if the node is overallocated. So we do not reclaim 3830 * if less than a specified percentage of the node is used by 3831 * unmapped file backed pages. 3832 */ 3833 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && 3834 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) 3835 return NODE_RECLAIM_FULL; 3836 3837 /* 3838 * Do not scan if the allocation should not be delayed. 3839 */ 3840 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) 3841 return NODE_RECLAIM_NOSCAN; 3842 3843 /* 3844 * Only run node reclaim on the local node or on nodes that do not 3845 * have associated processors. This will favor the local processor 3846 * over remote processors and spread off node memory allocations 3847 * as wide as possible. 3848 */ 3849 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) 3850 return NODE_RECLAIM_NOSCAN; 3851 3852 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) 3853 return NODE_RECLAIM_NOSCAN; 3854 3855 ret = __node_reclaim(pgdat, gfp_mask, order); 3856 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); 3857 3858 if (!ret) 3859 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); 3860 3861 return ret; 3862 } 3863 #endif 3864 3865 /* 3866 * page_evictable - test whether a page is evictable 3867 * @page: the page to test 3868 * 3869 * Test whether page is evictable--i.e., should be placed on active/inactive 3870 * lists vs unevictable list. 3871 * 3872 * Reasons page might not be evictable: 3873 * (1) page's mapping marked unevictable 3874 * (2) page is part of an mlocked VMA 3875 * 3876 */ 3877 int page_evictable(struct page *page) 3878 { 3879 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); 3880 } 3881 3882 #ifdef CONFIG_SHMEM 3883 /** 3884 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list 3885 * @pages: array of pages to check 3886 * @nr_pages: number of pages to check 3887 * 3888 * Checks pages for evictability and moves them to the appropriate lru list. 3889 * 3890 * This function is only used for SysV IPC SHM_UNLOCK. 3891 */ 3892 void check_move_unevictable_pages(struct page **pages, int nr_pages) 3893 { 3894 struct lruvec *lruvec; 3895 struct pglist_data *pgdat = NULL; 3896 int pgscanned = 0; 3897 int pgrescued = 0; 3898 int i; 3899 3900 for (i = 0; i < nr_pages; i++) { 3901 struct page *page = pages[i]; 3902 struct pglist_data *pagepgdat = page_pgdat(page); 3903 3904 pgscanned++; 3905 if (pagepgdat != pgdat) { 3906 if (pgdat) 3907 spin_unlock_irq(&pgdat->lru_lock); 3908 pgdat = pagepgdat; 3909 spin_lock_irq(&pgdat->lru_lock); 3910 } 3911 lruvec = mem_cgroup_page_lruvec(page, pgdat); 3912 3913 if (!PageLRU(page) || !PageUnevictable(page)) 3914 continue; 3915 3916 if (page_evictable(page)) { 3917 enum lru_list lru = page_lru_base_type(page); 3918 3919 VM_BUG_ON_PAGE(PageActive(page), page); 3920 ClearPageUnevictable(page); 3921 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); 3922 add_page_to_lru_list(page, lruvec, lru); 3923 pgrescued++; 3924 } 3925 } 3926 3927 if (pgdat) { 3928 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); 3929 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); 3930 spin_unlock_irq(&pgdat->lru_lock); 3931 } 3932 } 3933 #endif /* CONFIG_SHMEM */ 3934