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 /* 69 * Nodemask of nodes allowed by the caller. If NULL, all nodes 70 * are scanned. 71 */ 72 nodemask_t *nodemask; 73 74 /* 75 * The memory cgroup that hit its limit and as a result is the 76 * primary target of this reclaim invocation. 77 */ 78 struct mem_cgroup *target_mem_cgroup; 79 80 /* Writepage batching in laptop mode; RECLAIM_WRITE */ 81 unsigned int may_writepage:1; 82 83 /* Can mapped pages be reclaimed? */ 84 unsigned int may_unmap:1; 85 86 /* Can pages be swapped as part of reclaim? */ 87 unsigned int may_swap:1; 88 89 /* 90 * Cgroups are not reclaimed below their configured memory.low, 91 * unless we threaten to OOM. If any cgroups are skipped due to 92 * memory.low and nothing was reclaimed, go back for memory.low. 93 */ 94 unsigned int memcg_low_reclaim:1; 95 unsigned int memcg_low_skipped:1; 96 97 unsigned int hibernation_mode:1; 98 99 /* One of the zones is ready for compaction */ 100 unsigned int compaction_ready:1; 101 102 /* Allocation order */ 103 s8 order; 104 105 /* Scan (total_size >> priority) pages at once */ 106 s8 priority; 107 108 /* The highest zone to isolate pages for reclaim from */ 109 s8 reclaim_idx; 110 111 /* This context's GFP mask */ 112 gfp_t gfp_mask; 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 struct { 121 unsigned int dirty; 122 unsigned int unqueued_dirty; 123 unsigned int congested; 124 unsigned int writeback; 125 unsigned int immediate; 126 unsigned int file_taken; 127 unsigned int taken; 128 } nr; 129 }; 130 131 #ifdef ARCH_HAS_PREFETCH 132 #define prefetch_prev_lru_page(_page, _base, _field) \ 133 do { \ 134 if ((_page)->lru.prev != _base) { \ 135 struct page *prev; \ 136 \ 137 prev = lru_to_page(&(_page->lru)); \ 138 prefetch(&prev->_field); \ 139 } \ 140 } while (0) 141 #else 142 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) 143 #endif 144 145 #ifdef ARCH_HAS_PREFETCHW 146 #define prefetchw_prev_lru_page(_page, _base, _field) \ 147 do { \ 148 if ((_page)->lru.prev != _base) { \ 149 struct page *prev; \ 150 \ 151 prev = lru_to_page(&(_page->lru)); \ 152 prefetchw(&prev->_field); \ 153 } \ 154 } while (0) 155 #else 156 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) 157 #endif 158 159 /* 160 * From 0 .. 100. Higher means more swappy. 161 */ 162 int vm_swappiness = 60; 163 /* 164 * The total number of pages which are beyond the high watermark within all 165 * zones. 166 */ 167 unsigned long vm_total_pages; 168 169 static LIST_HEAD(shrinker_list); 170 static DECLARE_RWSEM(shrinker_rwsem); 171 172 #ifdef CONFIG_MEMCG_KMEM 173 174 /* 175 * We allow subsystems to populate their shrinker-related 176 * LRU lists before register_shrinker_prepared() is called 177 * for the shrinker, since we don't want to impose 178 * restrictions on their internal registration order. 179 * In this case shrink_slab_memcg() may find corresponding 180 * bit is set in the shrinkers map. 181 * 182 * This value is used by the function to detect registering 183 * shrinkers and to skip do_shrink_slab() calls for them. 184 */ 185 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL) 186 187 static DEFINE_IDR(shrinker_idr); 188 static int shrinker_nr_max; 189 190 static int prealloc_memcg_shrinker(struct shrinker *shrinker) 191 { 192 int id, ret = -ENOMEM; 193 194 down_write(&shrinker_rwsem); 195 /* This may call shrinker, so it must use down_read_trylock() */ 196 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL); 197 if (id < 0) 198 goto unlock; 199 200 if (id >= shrinker_nr_max) { 201 if (memcg_expand_shrinker_maps(id)) { 202 idr_remove(&shrinker_idr, id); 203 goto unlock; 204 } 205 206 shrinker_nr_max = id + 1; 207 } 208 shrinker->id = id; 209 ret = 0; 210 unlock: 211 up_write(&shrinker_rwsem); 212 return ret; 213 } 214 215 static void unregister_memcg_shrinker(struct shrinker *shrinker) 216 { 217 int id = shrinker->id; 218 219 BUG_ON(id < 0); 220 221 down_write(&shrinker_rwsem); 222 idr_remove(&shrinker_idr, id); 223 up_write(&shrinker_rwsem); 224 } 225 #else /* CONFIG_MEMCG_KMEM */ 226 static int prealloc_memcg_shrinker(struct shrinker *shrinker) 227 { 228 return 0; 229 } 230 231 static void unregister_memcg_shrinker(struct shrinker *shrinker) 232 { 233 } 234 #endif /* CONFIG_MEMCG_KMEM */ 235 236 #ifdef CONFIG_MEMCG 237 static bool global_reclaim(struct scan_control *sc) 238 { 239 return !sc->target_mem_cgroup; 240 } 241 242 /** 243 * sane_reclaim - is the usual dirty throttling mechanism operational? 244 * @sc: scan_control in question 245 * 246 * The normal page dirty throttling mechanism in balance_dirty_pages() is 247 * completely broken with the legacy memcg and direct stalling in 248 * shrink_page_list() is used for throttling instead, which lacks all the 249 * niceties such as fairness, adaptive pausing, bandwidth proportional 250 * allocation and configurability. 251 * 252 * This function tests whether the vmscan currently in progress can assume 253 * that the normal dirty throttling mechanism is operational. 254 */ 255 static bool sane_reclaim(struct scan_control *sc) 256 { 257 struct mem_cgroup *memcg = sc->target_mem_cgroup; 258 259 if (!memcg) 260 return true; 261 #ifdef CONFIG_CGROUP_WRITEBACK 262 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) 263 return true; 264 #endif 265 return false; 266 } 267 268 static void set_memcg_congestion(pg_data_t *pgdat, 269 struct mem_cgroup *memcg, 270 bool congested) 271 { 272 struct mem_cgroup_per_node *mn; 273 274 if (!memcg) 275 return; 276 277 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id); 278 WRITE_ONCE(mn->congested, congested); 279 } 280 281 static bool memcg_congested(pg_data_t *pgdat, 282 struct mem_cgroup *memcg) 283 { 284 struct mem_cgroup_per_node *mn; 285 286 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id); 287 return READ_ONCE(mn->congested); 288 289 } 290 #else 291 static bool global_reclaim(struct scan_control *sc) 292 { 293 return true; 294 } 295 296 static bool sane_reclaim(struct scan_control *sc) 297 { 298 return true; 299 } 300 301 static inline void set_memcg_congestion(struct pglist_data *pgdat, 302 struct mem_cgroup *memcg, bool congested) 303 { 304 } 305 306 static inline bool memcg_congested(struct pglist_data *pgdat, 307 struct mem_cgroup *memcg) 308 { 309 return false; 310 311 } 312 #endif 313 314 /* 315 * This misses isolated pages which are not accounted for to save counters. 316 * As the data only determines if reclaim or compaction continues, it is 317 * not expected that isolated pages will be a dominating factor. 318 */ 319 unsigned long zone_reclaimable_pages(struct zone *zone) 320 { 321 unsigned long nr; 322 323 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + 324 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); 325 if (get_nr_swap_pages() > 0) 326 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + 327 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); 328 329 return nr; 330 } 331 332 /** 333 * lruvec_lru_size - Returns the number of pages on the given LRU list. 334 * @lruvec: lru vector 335 * @lru: lru to use 336 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) 337 */ 338 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx) 339 { 340 unsigned long lru_size; 341 int zid; 342 343 if (!mem_cgroup_disabled()) 344 lru_size = mem_cgroup_get_lru_size(lruvec, lru); 345 else 346 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru); 347 348 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) { 349 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; 350 unsigned long size; 351 352 if (!managed_zone(zone)) 353 continue; 354 355 if (!mem_cgroup_disabled()) 356 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid); 357 else 358 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid], 359 NR_ZONE_LRU_BASE + lru); 360 lru_size -= min(size, lru_size); 361 } 362 363 return lru_size; 364 365 } 366 367 /* 368 * Add a shrinker callback to be called from the vm. 369 */ 370 int prealloc_shrinker(struct shrinker *shrinker) 371 { 372 size_t size = sizeof(*shrinker->nr_deferred); 373 374 if (shrinker->flags & SHRINKER_NUMA_AWARE) 375 size *= nr_node_ids; 376 377 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); 378 if (!shrinker->nr_deferred) 379 return -ENOMEM; 380 381 if (shrinker->flags & SHRINKER_MEMCG_AWARE) { 382 if (prealloc_memcg_shrinker(shrinker)) 383 goto free_deferred; 384 } 385 386 return 0; 387 388 free_deferred: 389 kfree(shrinker->nr_deferred); 390 shrinker->nr_deferred = NULL; 391 return -ENOMEM; 392 } 393 394 void free_prealloced_shrinker(struct shrinker *shrinker) 395 { 396 if (!shrinker->nr_deferred) 397 return; 398 399 if (shrinker->flags & SHRINKER_MEMCG_AWARE) 400 unregister_memcg_shrinker(shrinker); 401 402 kfree(shrinker->nr_deferred); 403 shrinker->nr_deferred = NULL; 404 } 405 406 void register_shrinker_prepared(struct shrinker *shrinker) 407 { 408 down_write(&shrinker_rwsem); 409 list_add_tail(&shrinker->list, &shrinker_list); 410 #ifdef CONFIG_MEMCG_KMEM 411 idr_replace(&shrinker_idr, shrinker, shrinker->id); 412 #endif 413 up_write(&shrinker_rwsem); 414 } 415 416 int register_shrinker(struct shrinker *shrinker) 417 { 418 int err = prealloc_shrinker(shrinker); 419 420 if (err) 421 return err; 422 register_shrinker_prepared(shrinker); 423 return 0; 424 } 425 EXPORT_SYMBOL(register_shrinker); 426 427 /* 428 * Remove one 429 */ 430 void unregister_shrinker(struct shrinker *shrinker) 431 { 432 if (!shrinker->nr_deferred) 433 return; 434 if (shrinker->flags & SHRINKER_MEMCG_AWARE) 435 unregister_memcg_shrinker(shrinker); 436 down_write(&shrinker_rwsem); 437 list_del(&shrinker->list); 438 up_write(&shrinker_rwsem); 439 kfree(shrinker->nr_deferred); 440 shrinker->nr_deferred = NULL; 441 } 442 EXPORT_SYMBOL(unregister_shrinker); 443 444 #define SHRINK_BATCH 128 445 446 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, 447 struct shrinker *shrinker, int priority) 448 { 449 unsigned long freed = 0; 450 unsigned long long delta; 451 long total_scan; 452 long freeable; 453 long nr; 454 long new_nr; 455 int nid = shrinkctl->nid; 456 long batch_size = shrinker->batch ? shrinker->batch 457 : SHRINK_BATCH; 458 long scanned = 0, next_deferred; 459 460 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) 461 nid = 0; 462 463 freeable = shrinker->count_objects(shrinker, shrinkctl); 464 if (freeable == 0 || freeable == SHRINK_EMPTY) 465 return freeable; 466 467 /* 468 * copy the current shrinker scan count into a local variable 469 * and zero it so that other concurrent shrinker invocations 470 * don't also do this scanning work. 471 */ 472 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); 473 474 total_scan = nr; 475 delta = freeable >> priority; 476 delta *= 4; 477 do_div(delta, shrinker->seeks); 478 total_scan += delta; 479 if (total_scan < 0) { 480 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", 481 shrinker->scan_objects, total_scan); 482 total_scan = freeable; 483 next_deferred = nr; 484 } else 485 next_deferred = total_scan; 486 487 /* 488 * We need to avoid excessive windup on filesystem shrinkers 489 * due to large numbers of GFP_NOFS allocations causing the 490 * shrinkers to return -1 all the time. This results in a large 491 * nr being built up so when a shrink that can do some work 492 * comes along it empties the entire cache due to nr >>> 493 * freeable. This is bad for sustaining a working set in 494 * memory. 495 * 496 * Hence only allow the shrinker to scan the entire cache when 497 * a large delta change is calculated directly. 498 */ 499 if (delta < freeable / 4) 500 total_scan = min(total_scan, freeable / 2); 501 502 /* 503 * Avoid risking looping forever due to too large nr value: 504 * never try to free more than twice the estimate number of 505 * freeable entries. 506 */ 507 if (total_scan > freeable * 2) 508 total_scan = freeable * 2; 509 510 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, 511 freeable, delta, total_scan, priority); 512 513 /* 514 * Normally, we should not scan less than batch_size objects in one 515 * pass to avoid too frequent shrinker calls, but if the slab has less 516 * than batch_size objects in total and we are really tight on memory, 517 * we will try to reclaim all available objects, otherwise we can end 518 * up failing allocations although there are plenty of reclaimable 519 * objects spread over several slabs with usage less than the 520 * batch_size. 521 * 522 * We detect the "tight on memory" situations by looking at the total 523 * number of objects we want to scan (total_scan). If it is greater 524 * than the total number of objects on slab (freeable), we must be 525 * scanning at high prio and therefore should try to reclaim as much as 526 * possible. 527 */ 528 while (total_scan >= batch_size || 529 total_scan >= freeable) { 530 unsigned long ret; 531 unsigned long nr_to_scan = min(batch_size, total_scan); 532 533 shrinkctl->nr_to_scan = nr_to_scan; 534 shrinkctl->nr_scanned = nr_to_scan; 535 ret = shrinker->scan_objects(shrinker, shrinkctl); 536 if (ret == SHRINK_STOP) 537 break; 538 freed += ret; 539 540 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned); 541 total_scan -= shrinkctl->nr_scanned; 542 scanned += shrinkctl->nr_scanned; 543 544 cond_resched(); 545 } 546 547 if (next_deferred >= scanned) 548 next_deferred -= scanned; 549 else 550 next_deferred = 0; 551 /* 552 * move the unused scan count back into the shrinker in a 553 * manner that handles concurrent updates. If we exhausted the 554 * scan, there is no need to do an update. 555 */ 556 if (next_deferred > 0) 557 new_nr = atomic_long_add_return(next_deferred, 558 &shrinker->nr_deferred[nid]); 559 else 560 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); 561 562 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); 563 return freed; 564 } 565 566 #ifdef CONFIG_MEMCG_KMEM 567 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, 568 struct mem_cgroup *memcg, int priority) 569 { 570 struct memcg_shrinker_map *map; 571 unsigned long freed = 0; 572 int ret, i; 573 574 if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)) 575 return 0; 576 577 if (!down_read_trylock(&shrinker_rwsem)) 578 return 0; 579 580 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map, 581 true); 582 if (unlikely(!map)) 583 goto unlock; 584 585 for_each_set_bit(i, map->map, shrinker_nr_max) { 586 struct shrink_control sc = { 587 .gfp_mask = gfp_mask, 588 .nid = nid, 589 .memcg = memcg, 590 }; 591 struct shrinker *shrinker; 592 593 shrinker = idr_find(&shrinker_idr, i); 594 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) { 595 if (!shrinker) 596 clear_bit(i, map->map); 597 continue; 598 } 599 600 ret = do_shrink_slab(&sc, shrinker, priority); 601 if (ret == SHRINK_EMPTY) { 602 clear_bit(i, map->map); 603 /* 604 * After the shrinker reported that it had no objects to 605 * free, but before we cleared the corresponding bit in 606 * the memcg shrinker map, a new object might have been 607 * added. To make sure, we have the bit set in this 608 * case, we invoke the shrinker one more time and reset 609 * the bit if it reports that it is not empty anymore. 610 * The memory barrier here pairs with the barrier in 611 * memcg_set_shrinker_bit(): 612 * 613 * list_lru_add() shrink_slab_memcg() 614 * list_add_tail() clear_bit() 615 * <MB> <MB> 616 * set_bit() do_shrink_slab() 617 */ 618 smp_mb__after_atomic(); 619 ret = do_shrink_slab(&sc, shrinker, priority); 620 if (ret == SHRINK_EMPTY) 621 ret = 0; 622 else 623 memcg_set_shrinker_bit(memcg, nid, i); 624 } 625 freed += ret; 626 627 if (rwsem_is_contended(&shrinker_rwsem)) { 628 freed = freed ? : 1; 629 break; 630 } 631 } 632 unlock: 633 up_read(&shrinker_rwsem); 634 return freed; 635 } 636 #else /* CONFIG_MEMCG_KMEM */ 637 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, 638 struct mem_cgroup *memcg, int priority) 639 { 640 return 0; 641 } 642 #endif /* CONFIG_MEMCG_KMEM */ 643 644 /** 645 * shrink_slab - shrink slab caches 646 * @gfp_mask: allocation context 647 * @nid: node whose slab caches to target 648 * @memcg: memory cgroup whose slab caches to target 649 * @priority: the reclaim priority 650 * 651 * Call the shrink functions to age shrinkable caches. 652 * 653 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, 654 * unaware shrinkers will receive a node id of 0 instead. 655 * 656 * @memcg specifies the memory cgroup to target. Unaware shrinkers 657 * are called only if it is the root cgroup. 658 * 659 * @priority is sc->priority, we take the number of objects and >> by priority 660 * in order to get the scan target. 661 * 662 * Returns the number of reclaimed slab objects. 663 */ 664 static unsigned long shrink_slab(gfp_t gfp_mask, int nid, 665 struct mem_cgroup *memcg, 666 int priority) 667 { 668 struct shrinker *shrinker; 669 unsigned long freed = 0; 670 int ret; 671 672 if (!mem_cgroup_is_root(memcg)) 673 return shrink_slab_memcg(gfp_mask, nid, memcg, priority); 674 675 if (!down_read_trylock(&shrinker_rwsem)) 676 goto out; 677 678 list_for_each_entry(shrinker, &shrinker_list, list) { 679 struct shrink_control sc = { 680 .gfp_mask = gfp_mask, 681 .nid = nid, 682 .memcg = memcg, 683 }; 684 685 ret = do_shrink_slab(&sc, shrinker, priority); 686 if (ret == SHRINK_EMPTY) 687 ret = 0; 688 freed += ret; 689 /* 690 * Bail out if someone want to register a new shrinker to 691 * prevent the regsitration from being stalled for long periods 692 * by parallel ongoing shrinking. 693 */ 694 if (rwsem_is_contended(&shrinker_rwsem)) { 695 freed = freed ? : 1; 696 break; 697 } 698 } 699 700 up_read(&shrinker_rwsem); 701 out: 702 cond_resched(); 703 return freed; 704 } 705 706 void drop_slab_node(int nid) 707 { 708 unsigned long freed; 709 710 do { 711 struct mem_cgroup *memcg = NULL; 712 713 freed = 0; 714 memcg = mem_cgroup_iter(NULL, NULL, NULL); 715 do { 716 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0); 717 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); 718 } while (freed > 10); 719 } 720 721 void drop_slab(void) 722 { 723 int nid; 724 725 for_each_online_node(nid) 726 drop_slab_node(nid); 727 } 728 729 static inline int is_page_cache_freeable(struct page *page) 730 { 731 /* 732 * A freeable page cache page is referenced only by the caller 733 * that isolated the page, the page cache radix tree and 734 * optional buffer heads at page->private. 735 */ 736 int radix_pins = PageTransHuge(page) && PageSwapCache(page) ? 737 HPAGE_PMD_NR : 1; 738 return page_count(page) - page_has_private(page) == 1 + radix_pins; 739 } 740 741 static int may_write_to_inode(struct inode *inode, struct scan_control *sc) 742 { 743 if (current->flags & PF_SWAPWRITE) 744 return 1; 745 if (!inode_write_congested(inode)) 746 return 1; 747 if (inode_to_bdi(inode) == current->backing_dev_info) 748 return 1; 749 return 0; 750 } 751 752 /* 753 * We detected a synchronous write error writing a page out. Probably 754 * -ENOSPC. We need to propagate that into the address_space for a subsequent 755 * fsync(), msync() or close(). 756 * 757 * The tricky part is that after writepage we cannot touch the mapping: nothing 758 * prevents it from being freed up. But we have a ref on the page and once 759 * that page is locked, the mapping is pinned. 760 * 761 * We're allowed to run sleeping lock_page() here because we know the caller has 762 * __GFP_FS. 763 */ 764 static void handle_write_error(struct address_space *mapping, 765 struct page *page, int error) 766 { 767 lock_page(page); 768 if (page_mapping(page) == mapping) 769 mapping_set_error(mapping, error); 770 unlock_page(page); 771 } 772 773 /* possible outcome of pageout() */ 774 typedef enum { 775 /* failed to write page out, page is locked */ 776 PAGE_KEEP, 777 /* move page to the active list, page is locked */ 778 PAGE_ACTIVATE, 779 /* page has been sent to the disk successfully, page is unlocked */ 780 PAGE_SUCCESS, 781 /* page is clean and locked */ 782 PAGE_CLEAN, 783 } pageout_t; 784 785 /* 786 * pageout is called by shrink_page_list() for each dirty page. 787 * Calls ->writepage(). 788 */ 789 static pageout_t pageout(struct page *page, struct address_space *mapping, 790 struct scan_control *sc) 791 { 792 /* 793 * If the page is dirty, only perform writeback if that write 794 * will be non-blocking. To prevent this allocation from being 795 * stalled by pagecache activity. But note that there may be 796 * stalls if we need to run get_block(). We could test 797 * PagePrivate for that. 798 * 799 * If this process is currently in __generic_file_write_iter() against 800 * this page's queue, we can perform writeback even if that 801 * will block. 802 * 803 * If the page is swapcache, write it back even if that would 804 * block, for some throttling. This happens by accident, because 805 * swap_backing_dev_info is bust: it doesn't reflect the 806 * congestion state of the swapdevs. Easy to fix, if needed. 807 */ 808 if (!is_page_cache_freeable(page)) 809 return PAGE_KEEP; 810 if (!mapping) { 811 /* 812 * Some data journaling orphaned pages can have 813 * page->mapping == NULL while being dirty with clean buffers. 814 */ 815 if (page_has_private(page)) { 816 if (try_to_free_buffers(page)) { 817 ClearPageDirty(page); 818 pr_info("%s: orphaned page\n", __func__); 819 return PAGE_CLEAN; 820 } 821 } 822 return PAGE_KEEP; 823 } 824 if (mapping->a_ops->writepage == NULL) 825 return PAGE_ACTIVATE; 826 if (!may_write_to_inode(mapping->host, sc)) 827 return PAGE_KEEP; 828 829 if (clear_page_dirty_for_io(page)) { 830 int res; 831 struct writeback_control wbc = { 832 .sync_mode = WB_SYNC_NONE, 833 .nr_to_write = SWAP_CLUSTER_MAX, 834 .range_start = 0, 835 .range_end = LLONG_MAX, 836 .for_reclaim = 1, 837 }; 838 839 SetPageReclaim(page); 840 res = mapping->a_ops->writepage(page, &wbc); 841 if (res < 0) 842 handle_write_error(mapping, page, res); 843 if (res == AOP_WRITEPAGE_ACTIVATE) { 844 ClearPageReclaim(page); 845 return PAGE_ACTIVATE; 846 } 847 848 if (!PageWriteback(page)) { 849 /* synchronous write or broken a_ops? */ 850 ClearPageReclaim(page); 851 } 852 trace_mm_vmscan_writepage(page); 853 inc_node_page_state(page, NR_VMSCAN_WRITE); 854 return PAGE_SUCCESS; 855 } 856 857 return PAGE_CLEAN; 858 } 859 860 /* 861 * Same as remove_mapping, but if the page is removed from the mapping, it 862 * gets returned with a refcount of 0. 863 */ 864 static int __remove_mapping(struct address_space *mapping, struct page *page, 865 bool reclaimed) 866 { 867 unsigned long flags; 868 int refcount; 869 870 BUG_ON(!PageLocked(page)); 871 BUG_ON(mapping != page_mapping(page)); 872 873 xa_lock_irqsave(&mapping->i_pages, flags); 874 /* 875 * The non racy check for a busy page. 876 * 877 * Must be careful with the order of the tests. When someone has 878 * a ref to the page, it may be possible that they dirty it then 879 * drop the reference. So if PageDirty is tested before page_count 880 * here, then the following race may occur: 881 * 882 * get_user_pages(&page); 883 * [user mapping goes away] 884 * write_to(page); 885 * !PageDirty(page) [good] 886 * SetPageDirty(page); 887 * put_page(page); 888 * !page_count(page) [good, discard it] 889 * 890 * [oops, our write_to data is lost] 891 * 892 * Reversing the order of the tests ensures such a situation cannot 893 * escape unnoticed. The smp_rmb is needed to ensure the page->flags 894 * load is not satisfied before that of page->_refcount. 895 * 896 * Note that if SetPageDirty is always performed via set_page_dirty, 897 * and thus under the i_pages lock, then this ordering is not required. 898 */ 899 if (unlikely(PageTransHuge(page)) && PageSwapCache(page)) 900 refcount = 1 + HPAGE_PMD_NR; 901 else 902 refcount = 2; 903 if (!page_ref_freeze(page, refcount)) 904 goto cannot_free; 905 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ 906 if (unlikely(PageDirty(page))) { 907 page_ref_unfreeze(page, refcount); 908 goto cannot_free; 909 } 910 911 if (PageSwapCache(page)) { 912 swp_entry_t swap = { .val = page_private(page) }; 913 mem_cgroup_swapout(page, swap); 914 __delete_from_swap_cache(page); 915 xa_unlock_irqrestore(&mapping->i_pages, flags); 916 put_swap_page(page, swap); 917 } else { 918 void (*freepage)(struct page *); 919 void *shadow = NULL; 920 921 freepage = mapping->a_ops->freepage; 922 /* 923 * Remember a shadow entry for reclaimed file cache in 924 * order to detect refaults, thus thrashing, later on. 925 * 926 * But don't store shadows in an address space that is 927 * already exiting. This is not just an optizimation, 928 * inode reclaim needs to empty out the radix tree or 929 * the nodes are lost. Don't plant shadows behind its 930 * back. 931 * 932 * We also don't store shadows for DAX mappings because the 933 * only page cache pages found in these are zero pages 934 * covering holes, and because we don't want to mix DAX 935 * exceptional entries and shadow exceptional entries in the 936 * same address_space. 937 */ 938 if (reclaimed && page_is_file_cache(page) && 939 !mapping_exiting(mapping) && !dax_mapping(mapping)) 940 shadow = workingset_eviction(mapping, page); 941 __delete_from_page_cache(page, shadow); 942 xa_unlock_irqrestore(&mapping->i_pages, flags); 943 944 if (freepage != NULL) 945 freepage(page); 946 } 947 948 return 1; 949 950 cannot_free: 951 xa_unlock_irqrestore(&mapping->i_pages, flags); 952 return 0; 953 } 954 955 /* 956 * Attempt to detach a locked page from its ->mapping. If it is dirty or if 957 * someone else has a ref on the page, abort and return 0. If it was 958 * successfully detached, return 1. Assumes the caller has a single ref on 959 * this page. 960 */ 961 int remove_mapping(struct address_space *mapping, struct page *page) 962 { 963 if (__remove_mapping(mapping, page, false)) { 964 /* 965 * Unfreezing the refcount with 1 rather than 2 effectively 966 * drops the pagecache ref for us without requiring another 967 * atomic operation. 968 */ 969 page_ref_unfreeze(page, 1); 970 return 1; 971 } 972 return 0; 973 } 974 975 /** 976 * putback_lru_page - put previously isolated page onto appropriate LRU list 977 * @page: page to be put back to appropriate lru list 978 * 979 * Add previously isolated @page to appropriate LRU list. 980 * Page may still be unevictable for other reasons. 981 * 982 * lru_lock must not be held, interrupts must be enabled. 983 */ 984 void putback_lru_page(struct page *page) 985 { 986 lru_cache_add(page); 987 put_page(page); /* drop ref from isolate */ 988 } 989 990 enum page_references { 991 PAGEREF_RECLAIM, 992 PAGEREF_RECLAIM_CLEAN, 993 PAGEREF_KEEP, 994 PAGEREF_ACTIVATE, 995 }; 996 997 static enum page_references page_check_references(struct page *page, 998 struct scan_control *sc) 999 { 1000 int referenced_ptes, referenced_page; 1001 unsigned long vm_flags; 1002 1003 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, 1004 &vm_flags); 1005 referenced_page = TestClearPageReferenced(page); 1006 1007 /* 1008 * Mlock lost the isolation race with us. Let try_to_unmap() 1009 * move the page to the unevictable list. 1010 */ 1011 if (vm_flags & VM_LOCKED) 1012 return PAGEREF_RECLAIM; 1013 1014 if (referenced_ptes) { 1015 if (PageSwapBacked(page)) 1016 return PAGEREF_ACTIVATE; 1017 /* 1018 * All mapped pages start out with page table 1019 * references from the instantiating fault, so we need 1020 * to look twice if a mapped file page is used more 1021 * than once. 1022 * 1023 * Mark it and spare it for another trip around the 1024 * inactive list. Another page table reference will 1025 * lead to its activation. 1026 * 1027 * Note: the mark is set for activated pages as well 1028 * so that recently deactivated but used pages are 1029 * quickly recovered. 1030 */ 1031 SetPageReferenced(page); 1032 1033 if (referenced_page || referenced_ptes > 1) 1034 return PAGEREF_ACTIVATE; 1035 1036 /* 1037 * Activate file-backed executable pages after first usage. 1038 */ 1039 if (vm_flags & VM_EXEC) 1040 return PAGEREF_ACTIVATE; 1041 1042 return PAGEREF_KEEP; 1043 } 1044 1045 /* Reclaim if clean, defer dirty pages to writeback */ 1046 if (referenced_page && !PageSwapBacked(page)) 1047 return PAGEREF_RECLAIM_CLEAN; 1048 1049 return PAGEREF_RECLAIM; 1050 } 1051 1052 /* Check if a page is dirty or under writeback */ 1053 static void page_check_dirty_writeback(struct page *page, 1054 bool *dirty, bool *writeback) 1055 { 1056 struct address_space *mapping; 1057 1058 /* 1059 * Anonymous pages are not handled by flushers and must be written 1060 * from reclaim context. Do not stall reclaim based on them 1061 */ 1062 if (!page_is_file_cache(page) || 1063 (PageAnon(page) && !PageSwapBacked(page))) { 1064 *dirty = false; 1065 *writeback = false; 1066 return; 1067 } 1068 1069 /* By default assume that the page flags are accurate */ 1070 *dirty = PageDirty(page); 1071 *writeback = PageWriteback(page); 1072 1073 /* Verify dirty/writeback state if the filesystem supports it */ 1074 if (!page_has_private(page)) 1075 return; 1076 1077 mapping = page_mapping(page); 1078 if (mapping && mapping->a_ops->is_dirty_writeback) 1079 mapping->a_ops->is_dirty_writeback(page, dirty, writeback); 1080 } 1081 1082 /* 1083 * shrink_page_list() returns the number of reclaimed pages 1084 */ 1085 static unsigned long shrink_page_list(struct list_head *page_list, 1086 struct pglist_data *pgdat, 1087 struct scan_control *sc, 1088 enum ttu_flags ttu_flags, 1089 struct reclaim_stat *stat, 1090 bool force_reclaim) 1091 { 1092 LIST_HEAD(ret_pages); 1093 LIST_HEAD(free_pages); 1094 int pgactivate = 0; 1095 unsigned nr_unqueued_dirty = 0; 1096 unsigned nr_dirty = 0; 1097 unsigned nr_congested = 0; 1098 unsigned nr_reclaimed = 0; 1099 unsigned nr_writeback = 0; 1100 unsigned nr_immediate = 0; 1101 unsigned nr_ref_keep = 0; 1102 unsigned nr_unmap_fail = 0; 1103 1104 cond_resched(); 1105 1106 while (!list_empty(page_list)) { 1107 struct address_space *mapping; 1108 struct page *page; 1109 int may_enter_fs; 1110 enum page_references references = PAGEREF_RECLAIM_CLEAN; 1111 bool dirty, writeback; 1112 1113 cond_resched(); 1114 1115 page = lru_to_page(page_list); 1116 list_del(&page->lru); 1117 1118 if (!trylock_page(page)) 1119 goto keep; 1120 1121 VM_BUG_ON_PAGE(PageActive(page), page); 1122 1123 sc->nr_scanned++; 1124 1125 if (unlikely(!page_evictable(page))) 1126 goto activate_locked; 1127 1128 if (!sc->may_unmap && page_mapped(page)) 1129 goto keep_locked; 1130 1131 /* Double the slab pressure for mapped and swapcache pages */ 1132 if ((page_mapped(page) || PageSwapCache(page)) && 1133 !(PageAnon(page) && !PageSwapBacked(page))) 1134 sc->nr_scanned++; 1135 1136 may_enter_fs = (sc->gfp_mask & __GFP_FS) || 1137 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); 1138 1139 /* 1140 * The number of dirty pages determines if a node is marked 1141 * reclaim_congested which affects wait_iff_congested. kswapd 1142 * will stall and start writing pages if the tail of the LRU 1143 * is all dirty unqueued pages. 1144 */ 1145 page_check_dirty_writeback(page, &dirty, &writeback); 1146 if (dirty || writeback) 1147 nr_dirty++; 1148 1149 if (dirty && !writeback) 1150 nr_unqueued_dirty++; 1151 1152 /* 1153 * Treat this page as congested if the underlying BDI is or if 1154 * pages are cycling through the LRU so quickly that the 1155 * pages marked for immediate reclaim are making it to the 1156 * end of the LRU a second time. 1157 */ 1158 mapping = page_mapping(page); 1159 if (((dirty || writeback) && mapping && 1160 inode_write_congested(mapping->host)) || 1161 (writeback && PageReclaim(page))) 1162 nr_congested++; 1163 1164 /* 1165 * If a page at the tail of the LRU is under writeback, there 1166 * are three cases to consider. 1167 * 1168 * 1) If reclaim is encountering an excessive number of pages 1169 * under writeback and this page is both under writeback and 1170 * PageReclaim then it indicates that pages are being queued 1171 * for IO but are being recycled through the LRU before the 1172 * IO can complete. Waiting on the page itself risks an 1173 * indefinite stall if it is impossible to writeback the 1174 * page due to IO error or disconnected storage so instead 1175 * note that the LRU is being scanned too quickly and the 1176 * caller can stall after page list has been processed. 1177 * 1178 * 2) Global or new memcg reclaim encounters a page that is 1179 * not marked for immediate reclaim, or the caller does not 1180 * have __GFP_FS (or __GFP_IO if it's simply going to swap, 1181 * not to fs). In this case mark the page for immediate 1182 * reclaim and continue scanning. 1183 * 1184 * Require may_enter_fs because we would wait on fs, which 1185 * may not have submitted IO yet. And the loop driver might 1186 * enter reclaim, and deadlock if it waits on a page for 1187 * which it is needed to do the write (loop masks off 1188 * __GFP_IO|__GFP_FS for this reason); but more thought 1189 * would probably show more reasons. 1190 * 1191 * 3) Legacy memcg encounters a page that is already marked 1192 * PageReclaim. memcg does not have any dirty pages 1193 * throttling so we could easily OOM just because too many 1194 * pages are in writeback and there is nothing else to 1195 * reclaim. Wait for the writeback to complete. 1196 * 1197 * In cases 1) and 2) we activate the pages to get them out of 1198 * the way while we continue scanning for clean pages on the 1199 * inactive list and refilling from the active list. The 1200 * observation here is that waiting for disk writes is more 1201 * expensive than potentially causing reloads down the line. 1202 * Since they're marked for immediate reclaim, they won't put 1203 * memory pressure on the cache working set any longer than it 1204 * takes to write them to disk. 1205 */ 1206 if (PageWriteback(page)) { 1207 /* Case 1 above */ 1208 if (current_is_kswapd() && 1209 PageReclaim(page) && 1210 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { 1211 nr_immediate++; 1212 goto activate_locked; 1213 1214 /* Case 2 above */ 1215 } else if (sane_reclaim(sc) || 1216 !PageReclaim(page) || !may_enter_fs) { 1217 /* 1218 * This is slightly racy - end_page_writeback() 1219 * might have just cleared PageReclaim, then 1220 * setting PageReclaim here end up interpreted 1221 * as PageReadahead - but that does not matter 1222 * enough to care. What we do want is for this 1223 * page to have PageReclaim set next time memcg 1224 * reclaim reaches the tests above, so it will 1225 * then wait_on_page_writeback() to avoid OOM; 1226 * and it's also appropriate in global reclaim. 1227 */ 1228 SetPageReclaim(page); 1229 nr_writeback++; 1230 goto activate_locked; 1231 1232 /* Case 3 above */ 1233 } else { 1234 unlock_page(page); 1235 wait_on_page_writeback(page); 1236 /* then go back and try same page again */ 1237 list_add_tail(&page->lru, page_list); 1238 continue; 1239 } 1240 } 1241 1242 if (!force_reclaim) 1243 references = page_check_references(page, sc); 1244 1245 switch (references) { 1246 case PAGEREF_ACTIVATE: 1247 goto activate_locked; 1248 case PAGEREF_KEEP: 1249 nr_ref_keep++; 1250 goto keep_locked; 1251 case PAGEREF_RECLAIM: 1252 case PAGEREF_RECLAIM_CLEAN: 1253 ; /* try to reclaim the page below */ 1254 } 1255 1256 /* 1257 * Anonymous process memory has backing store? 1258 * Try to allocate it some swap space here. 1259 * Lazyfree page could be freed directly 1260 */ 1261 if (PageAnon(page) && PageSwapBacked(page)) { 1262 if (!PageSwapCache(page)) { 1263 if (!(sc->gfp_mask & __GFP_IO)) 1264 goto keep_locked; 1265 if (PageTransHuge(page)) { 1266 /* cannot split THP, skip it */ 1267 if (!can_split_huge_page(page, NULL)) 1268 goto activate_locked; 1269 /* 1270 * Split pages without a PMD map right 1271 * away. Chances are some or all of the 1272 * tail pages can be freed without IO. 1273 */ 1274 if (!compound_mapcount(page) && 1275 split_huge_page_to_list(page, 1276 page_list)) 1277 goto activate_locked; 1278 } 1279 if (!add_to_swap(page)) { 1280 if (!PageTransHuge(page)) 1281 goto activate_locked; 1282 /* Fallback to swap normal pages */ 1283 if (split_huge_page_to_list(page, 1284 page_list)) 1285 goto activate_locked; 1286 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 1287 count_vm_event(THP_SWPOUT_FALLBACK); 1288 #endif 1289 if (!add_to_swap(page)) 1290 goto activate_locked; 1291 } 1292 1293 may_enter_fs = 1; 1294 1295 /* Adding to swap updated mapping */ 1296 mapping = page_mapping(page); 1297 } 1298 } else if (unlikely(PageTransHuge(page))) { 1299 /* Split file THP */ 1300 if (split_huge_page_to_list(page, page_list)) 1301 goto keep_locked; 1302 } 1303 1304 /* 1305 * The page is mapped into the page tables of one or more 1306 * processes. Try to unmap it here. 1307 */ 1308 if (page_mapped(page)) { 1309 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH; 1310 1311 if (unlikely(PageTransHuge(page))) 1312 flags |= TTU_SPLIT_HUGE_PMD; 1313 if (!try_to_unmap(page, flags)) { 1314 nr_unmap_fail++; 1315 goto activate_locked; 1316 } 1317 } 1318 1319 if (PageDirty(page)) { 1320 /* 1321 * Only kswapd can writeback filesystem pages 1322 * to avoid risk of stack overflow. But avoid 1323 * injecting inefficient single-page IO into 1324 * flusher writeback as much as possible: only 1325 * write pages when we've encountered many 1326 * dirty pages, and when we've already scanned 1327 * the rest of the LRU for clean pages and see 1328 * the same dirty pages again (PageReclaim). 1329 */ 1330 if (page_is_file_cache(page) && 1331 (!current_is_kswapd() || !PageReclaim(page) || 1332 !test_bit(PGDAT_DIRTY, &pgdat->flags))) { 1333 /* 1334 * Immediately reclaim when written back. 1335 * Similar in principal to deactivate_page() 1336 * except we already have the page isolated 1337 * and know it's dirty 1338 */ 1339 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); 1340 SetPageReclaim(page); 1341 1342 goto activate_locked; 1343 } 1344 1345 if (references == PAGEREF_RECLAIM_CLEAN) 1346 goto keep_locked; 1347 if (!may_enter_fs) 1348 goto keep_locked; 1349 if (!sc->may_writepage) 1350 goto keep_locked; 1351 1352 /* 1353 * Page is dirty. Flush the TLB if a writable entry 1354 * potentially exists to avoid CPU writes after IO 1355 * starts and then write it out here. 1356 */ 1357 try_to_unmap_flush_dirty(); 1358 switch (pageout(page, mapping, sc)) { 1359 case PAGE_KEEP: 1360 goto keep_locked; 1361 case PAGE_ACTIVATE: 1362 goto activate_locked; 1363 case PAGE_SUCCESS: 1364 if (PageWriteback(page)) 1365 goto keep; 1366 if (PageDirty(page)) 1367 goto keep; 1368 1369 /* 1370 * A synchronous write - probably a ramdisk. Go 1371 * ahead and try to reclaim the page. 1372 */ 1373 if (!trylock_page(page)) 1374 goto keep; 1375 if (PageDirty(page) || PageWriteback(page)) 1376 goto keep_locked; 1377 mapping = page_mapping(page); 1378 case PAGE_CLEAN: 1379 ; /* try to free the page below */ 1380 } 1381 } 1382 1383 /* 1384 * If the page has buffers, try to free the buffer mappings 1385 * associated with this page. If we succeed we try to free 1386 * the page as well. 1387 * 1388 * We do this even if the page is PageDirty(). 1389 * try_to_release_page() does not perform I/O, but it is 1390 * possible for a page to have PageDirty set, but it is actually 1391 * clean (all its buffers are clean). This happens if the 1392 * buffers were written out directly, with submit_bh(). ext3 1393 * will do this, as well as the blockdev mapping. 1394 * try_to_release_page() will discover that cleanness and will 1395 * drop the buffers and mark the page clean - it can be freed. 1396 * 1397 * Rarely, pages can have buffers and no ->mapping. These are 1398 * the pages which were not successfully invalidated in 1399 * truncate_complete_page(). We try to drop those buffers here 1400 * and if that worked, and the page is no longer mapped into 1401 * process address space (page_count == 1) it can be freed. 1402 * Otherwise, leave the page on the LRU so it is swappable. 1403 */ 1404 if (page_has_private(page)) { 1405 if (!try_to_release_page(page, sc->gfp_mask)) 1406 goto activate_locked; 1407 if (!mapping && page_count(page) == 1) { 1408 unlock_page(page); 1409 if (put_page_testzero(page)) 1410 goto free_it; 1411 else { 1412 /* 1413 * rare race with speculative reference. 1414 * the speculative reference will free 1415 * this page shortly, so we may 1416 * increment nr_reclaimed here (and 1417 * leave it off the LRU). 1418 */ 1419 nr_reclaimed++; 1420 continue; 1421 } 1422 } 1423 } 1424 1425 if (PageAnon(page) && !PageSwapBacked(page)) { 1426 /* follow __remove_mapping for reference */ 1427 if (!page_ref_freeze(page, 1)) 1428 goto keep_locked; 1429 if (PageDirty(page)) { 1430 page_ref_unfreeze(page, 1); 1431 goto keep_locked; 1432 } 1433 1434 count_vm_event(PGLAZYFREED); 1435 count_memcg_page_event(page, PGLAZYFREED); 1436 } else if (!mapping || !__remove_mapping(mapping, page, true)) 1437 goto keep_locked; 1438 /* 1439 * At this point, we have no other references and there is 1440 * no way to pick any more up (removed from LRU, removed 1441 * from pagecache). Can use non-atomic bitops now (and 1442 * we obviously don't have to worry about waking up a process 1443 * waiting on the page lock, because there are no references. 1444 */ 1445 __ClearPageLocked(page); 1446 free_it: 1447 nr_reclaimed++; 1448 1449 /* 1450 * Is there need to periodically free_page_list? It would 1451 * appear not as the counts should be low 1452 */ 1453 if (unlikely(PageTransHuge(page))) { 1454 mem_cgroup_uncharge(page); 1455 (*get_compound_page_dtor(page))(page); 1456 } else 1457 list_add(&page->lru, &free_pages); 1458 continue; 1459 1460 activate_locked: 1461 /* Not a candidate for swapping, so reclaim swap space. */ 1462 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) || 1463 PageMlocked(page))) 1464 try_to_free_swap(page); 1465 VM_BUG_ON_PAGE(PageActive(page), page); 1466 if (!PageMlocked(page)) { 1467 SetPageActive(page); 1468 pgactivate++; 1469 count_memcg_page_event(page, PGACTIVATE); 1470 } 1471 keep_locked: 1472 unlock_page(page); 1473 keep: 1474 list_add(&page->lru, &ret_pages); 1475 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); 1476 } 1477 1478 mem_cgroup_uncharge_list(&free_pages); 1479 try_to_unmap_flush(); 1480 free_unref_page_list(&free_pages); 1481 1482 list_splice(&ret_pages, page_list); 1483 count_vm_events(PGACTIVATE, pgactivate); 1484 1485 if (stat) { 1486 stat->nr_dirty = nr_dirty; 1487 stat->nr_congested = nr_congested; 1488 stat->nr_unqueued_dirty = nr_unqueued_dirty; 1489 stat->nr_writeback = nr_writeback; 1490 stat->nr_immediate = nr_immediate; 1491 stat->nr_activate = pgactivate; 1492 stat->nr_ref_keep = nr_ref_keep; 1493 stat->nr_unmap_fail = nr_unmap_fail; 1494 } 1495 return nr_reclaimed; 1496 } 1497 1498 unsigned long reclaim_clean_pages_from_list(struct zone *zone, 1499 struct list_head *page_list) 1500 { 1501 struct scan_control sc = { 1502 .gfp_mask = GFP_KERNEL, 1503 .priority = DEF_PRIORITY, 1504 .may_unmap = 1, 1505 }; 1506 unsigned long ret; 1507 struct page *page, *next; 1508 LIST_HEAD(clean_pages); 1509 1510 list_for_each_entry_safe(page, next, page_list, lru) { 1511 if (page_is_file_cache(page) && !PageDirty(page) && 1512 !__PageMovable(page)) { 1513 ClearPageActive(page); 1514 list_move(&page->lru, &clean_pages); 1515 } 1516 } 1517 1518 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, 1519 TTU_IGNORE_ACCESS, NULL, true); 1520 list_splice(&clean_pages, page_list); 1521 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); 1522 return ret; 1523 } 1524 1525 /* 1526 * Attempt to remove the specified page from its LRU. Only take this page 1527 * if it is of the appropriate PageActive status. Pages which are being 1528 * freed elsewhere are also ignored. 1529 * 1530 * page: page to consider 1531 * mode: one of the LRU isolation modes defined above 1532 * 1533 * returns 0 on success, -ve errno on failure. 1534 */ 1535 int __isolate_lru_page(struct page *page, isolate_mode_t mode) 1536 { 1537 int ret = -EINVAL; 1538 1539 /* Only take pages on the LRU. */ 1540 if (!PageLRU(page)) 1541 return ret; 1542 1543 /* Compaction should not handle unevictable pages but CMA can do so */ 1544 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) 1545 return ret; 1546 1547 ret = -EBUSY; 1548 1549 /* 1550 * To minimise LRU disruption, the caller can indicate that it only 1551 * wants to isolate pages it will be able to operate on without 1552 * blocking - clean pages for the most part. 1553 * 1554 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages 1555 * that it is possible to migrate without blocking 1556 */ 1557 if (mode & ISOLATE_ASYNC_MIGRATE) { 1558 /* All the caller can do on PageWriteback is block */ 1559 if (PageWriteback(page)) 1560 return ret; 1561 1562 if (PageDirty(page)) { 1563 struct address_space *mapping; 1564 bool migrate_dirty; 1565 1566 /* 1567 * Only pages without mappings or that have a 1568 * ->migratepage callback are possible to migrate 1569 * without blocking. However, we can be racing with 1570 * truncation so it's necessary to lock the page 1571 * to stabilise the mapping as truncation holds 1572 * the page lock until after the page is removed 1573 * from the page cache. 1574 */ 1575 if (!trylock_page(page)) 1576 return ret; 1577 1578 mapping = page_mapping(page); 1579 migrate_dirty = !mapping || mapping->a_ops->migratepage; 1580 unlock_page(page); 1581 if (!migrate_dirty) 1582 return ret; 1583 } 1584 } 1585 1586 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) 1587 return ret; 1588 1589 if (likely(get_page_unless_zero(page))) { 1590 /* 1591 * Be careful not to clear PageLRU until after we're 1592 * sure the page is not being freed elsewhere -- the 1593 * page release code relies on it. 1594 */ 1595 ClearPageLRU(page); 1596 ret = 0; 1597 } 1598 1599 return ret; 1600 } 1601 1602 1603 /* 1604 * Update LRU sizes after isolating pages. The LRU size updates must 1605 * be complete before mem_cgroup_update_lru_size due to a santity check. 1606 */ 1607 static __always_inline void update_lru_sizes(struct lruvec *lruvec, 1608 enum lru_list lru, unsigned long *nr_zone_taken) 1609 { 1610 int zid; 1611 1612 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 1613 if (!nr_zone_taken[zid]) 1614 continue; 1615 1616 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); 1617 #ifdef CONFIG_MEMCG 1618 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); 1619 #endif 1620 } 1621 1622 } 1623 1624 /* 1625 * zone_lru_lock is heavily contended. Some of the functions that 1626 * shrink the lists perform better by taking out a batch of pages 1627 * and working on them outside the LRU lock. 1628 * 1629 * For pagecache intensive workloads, this function is the hottest 1630 * spot in the kernel (apart from copy_*_user functions). 1631 * 1632 * Appropriate locks must be held before calling this function. 1633 * 1634 * @nr_to_scan: The number of eligible pages to look through on the list. 1635 * @lruvec: The LRU vector to pull pages from. 1636 * @dst: The temp list to put pages on to. 1637 * @nr_scanned: The number of pages that were scanned. 1638 * @sc: The scan_control struct for this reclaim session 1639 * @mode: One of the LRU isolation modes 1640 * @lru: LRU list id for isolating 1641 * 1642 * returns how many pages were moved onto *@dst. 1643 */ 1644 static unsigned long isolate_lru_pages(unsigned long nr_to_scan, 1645 struct lruvec *lruvec, struct list_head *dst, 1646 unsigned long *nr_scanned, struct scan_control *sc, 1647 isolate_mode_t mode, enum lru_list lru) 1648 { 1649 struct list_head *src = &lruvec->lists[lru]; 1650 unsigned long nr_taken = 0; 1651 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; 1652 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; 1653 unsigned long skipped = 0; 1654 unsigned long scan, total_scan, nr_pages; 1655 LIST_HEAD(pages_skipped); 1656 1657 scan = 0; 1658 for (total_scan = 0; 1659 scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src); 1660 total_scan++) { 1661 struct page *page; 1662 1663 page = lru_to_page(src); 1664 prefetchw_prev_lru_page(page, src, flags); 1665 1666 VM_BUG_ON_PAGE(!PageLRU(page), page); 1667 1668 if (page_zonenum(page) > sc->reclaim_idx) { 1669 list_move(&page->lru, &pages_skipped); 1670 nr_skipped[page_zonenum(page)]++; 1671 continue; 1672 } 1673 1674 /* 1675 * Do not count skipped pages because that makes the function 1676 * return with no isolated pages if the LRU mostly contains 1677 * ineligible pages. This causes the VM to not reclaim any 1678 * pages, triggering a premature OOM. 1679 */ 1680 scan++; 1681 switch (__isolate_lru_page(page, mode)) { 1682 case 0: 1683 nr_pages = hpage_nr_pages(page); 1684 nr_taken += nr_pages; 1685 nr_zone_taken[page_zonenum(page)] += nr_pages; 1686 list_move(&page->lru, dst); 1687 break; 1688 1689 case -EBUSY: 1690 /* else it is being freed elsewhere */ 1691 list_move(&page->lru, src); 1692 continue; 1693 1694 default: 1695 BUG(); 1696 } 1697 } 1698 1699 /* 1700 * Splice any skipped pages to the start of the LRU list. Note that 1701 * this disrupts the LRU order when reclaiming for lower zones but 1702 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX 1703 * scanning would soon rescan the same pages to skip and put the 1704 * system at risk of premature OOM. 1705 */ 1706 if (!list_empty(&pages_skipped)) { 1707 int zid; 1708 1709 list_splice(&pages_skipped, src); 1710 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 1711 if (!nr_skipped[zid]) 1712 continue; 1713 1714 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); 1715 skipped += nr_skipped[zid]; 1716 } 1717 } 1718 *nr_scanned = total_scan; 1719 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, 1720 total_scan, skipped, nr_taken, mode, lru); 1721 update_lru_sizes(lruvec, lru, nr_zone_taken); 1722 return nr_taken; 1723 } 1724 1725 /** 1726 * isolate_lru_page - tries to isolate a page from its LRU list 1727 * @page: page to isolate from its LRU list 1728 * 1729 * Isolates a @page from an LRU list, clears PageLRU and adjusts the 1730 * vmstat statistic corresponding to whatever LRU list the page was on. 1731 * 1732 * Returns 0 if the page was removed from an LRU list. 1733 * Returns -EBUSY if the page was not on an LRU list. 1734 * 1735 * The returned page will have PageLRU() cleared. If it was found on 1736 * the active list, it will have PageActive set. If it was found on 1737 * the unevictable list, it will have the PageUnevictable bit set. That flag 1738 * may need to be cleared by the caller before letting the page go. 1739 * 1740 * The vmstat statistic corresponding to the list on which the page was 1741 * found will be decremented. 1742 * 1743 * Restrictions: 1744 * 1745 * (1) Must be called with an elevated refcount on the page. This is a 1746 * fundamentnal difference from isolate_lru_pages (which is called 1747 * without a stable reference). 1748 * (2) the lru_lock must not be held. 1749 * (3) interrupts must be enabled. 1750 */ 1751 int isolate_lru_page(struct page *page) 1752 { 1753 int ret = -EBUSY; 1754 1755 VM_BUG_ON_PAGE(!page_count(page), page); 1756 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); 1757 1758 if (PageLRU(page)) { 1759 struct zone *zone = page_zone(page); 1760 struct lruvec *lruvec; 1761 1762 spin_lock_irq(zone_lru_lock(zone)); 1763 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); 1764 if (PageLRU(page)) { 1765 int lru = page_lru(page); 1766 get_page(page); 1767 ClearPageLRU(page); 1768 del_page_from_lru_list(page, lruvec, lru); 1769 ret = 0; 1770 } 1771 spin_unlock_irq(zone_lru_lock(zone)); 1772 } 1773 return ret; 1774 } 1775 1776 /* 1777 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and 1778 * then get resheduled. When there are massive number of tasks doing page 1779 * allocation, such sleeping direct reclaimers may keep piling up on each CPU, 1780 * the LRU list will go small and be scanned faster than necessary, leading to 1781 * unnecessary swapping, thrashing and OOM. 1782 */ 1783 static int too_many_isolated(struct pglist_data *pgdat, int file, 1784 struct scan_control *sc) 1785 { 1786 unsigned long inactive, isolated; 1787 1788 if (current_is_kswapd()) 1789 return 0; 1790 1791 if (!sane_reclaim(sc)) 1792 return 0; 1793 1794 if (file) { 1795 inactive = node_page_state(pgdat, NR_INACTIVE_FILE); 1796 isolated = node_page_state(pgdat, NR_ISOLATED_FILE); 1797 } else { 1798 inactive = node_page_state(pgdat, NR_INACTIVE_ANON); 1799 isolated = node_page_state(pgdat, NR_ISOLATED_ANON); 1800 } 1801 1802 /* 1803 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they 1804 * won't get blocked by normal direct-reclaimers, forming a circular 1805 * deadlock. 1806 */ 1807 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) 1808 inactive >>= 3; 1809 1810 return isolated > inactive; 1811 } 1812 1813 static noinline_for_stack void 1814 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) 1815 { 1816 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1817 struct pglist_data *pgdat = lruvec_pgdat(lruvec); 1818 LIST_HEAD(pages_to_free); 1819 1820 /* 1821 * Put back any unfreeable pages. 1822 */ 1823 while (!list_empty(page_list)) { 1824 struct page *page = lru_to_page(page_list); 1825 int lru; 1826 1827 VM_BUG_ON_PAGE(PageLRU(page), page); 1828 list_del(&page->lru); 1829 if (unlikely(!page_evictable(page))) { 1830 spin_unlock_irq(&pgdat->lru_lock); 1831 putback_lru_page(page); 1832 spin_lock_irq(&pgdat->lru_lock); 1833 continue; 1834 } 1835 1836 lruvec = mem_cgroup_page_lruvec(page, pgdat); 1837 1838 SetPageLRU(page); 1839 lru = page_lru(page); 1840 add_page_to_lru_list(page, lruvec, lru); 1841 1842 if (is_active_lru(lru)) { 1843 int file = is_file_lru(lru); 1844 int numpages = hpage_nr_pages(page); 1845 reclaim_stat->recent_rotated[file] += numpages; 1846 } 1847 if (put_page_testzero(page)) { 1848 __ClearPageLRU(page); 1849 __ClearPageActive(page); 1850 del_page_from_lru_list(page, lruvec, lru); 1851 1852 if (unlikely(PageCompound(page))) { 1853 spin_unlock_irq(&pgdat->lru_lock); 1854 mem_cgroup_uncharge(page); 1855 (*get_compound_page_dtor(page))(page); 1856 spin_lock_irq(&pgdat->lru_lock); 1857 } else 1858 list_add(&page->lru, &pages_to_free); 1859 } 1860 } 1861 1862 /* 1863 * To save our caller's stack, now use input list for pages to free. 1864 */ 1865 list_splice(&pages_to_free, page_list); 1866 } 1867 1868 /* 1869 * If a kernel thread (such as nfsd for loop-back mounts) services 1870 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. 1871 * In that case we should only throttle if the backing device it is 1872 * writing to is congested. In other cases it is safe to throttle. 1873 */ 1874 static int current_may_throttle(void) 1875 { 1876 return !(current->flags & PF_LESS_THROTTLE) || 1877 current->backing_dev_info == NULL || 1878 bdi_write_congested(current->backing_dev_info); 1879 } 1880 1881 /* 1882 * shrink_inactive_list() is a helper for shrink_node(). It returns the number 1883 * of reclaimed pages 1884 */ 1885 static noinline_for_stack unsigned long 1886 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, 1887 struct scan_control *sc, enum lru_list lru) 1888 { 1889 LIST_HEAD(page_list); 1890 unsigned long nr_scanned; 1891 unsigned long nr_reclaimed = 0; 1892 unsigned long nr_taken; 1893 struct reclaim_stat stat = {}; 1894 isolate_mode_t isolate_mode = 0; 1895 int file = is_file_lru(lru); 1896 struct pglist_data *pgdat = lruvec_pgdat(lruvec); 1897 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 1898 bool stalled = false; 1899 1900 while (unlikely(too_many_isolated(pgdat, file, sc))) { 1901 if (stalled) 1902 return 0; 1903 1904 /* wait a bit for the reclaimer. */ 1905 msleep(100); 1906 stalled = true; 1907 1908 /* We are about to die and free our memory. Return now. */ 1909 if (fatal_signal_pending(current)) 1910 return SWAP_CLUSTER_MAX; 1911 } 1912 1913 lru_add_drain(); 1914 1915 if (!sc->may_unmap) 1916 isolate_mode |= ISOLATE_UNMAPPED; 1917 1918 spin_lock_irq(&pgdat->lru_lock); 1919 1920 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, 1921 &nr_scanned, sc, isolate_mode, lru); 1922 1923 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); 1924 reclaim_stat->recent_scanned[file] += nr_taken; 1925 1926 if (current_is_kswapd()) { 1927 if (global_reclaim(sc)) 1928 __count_vm_events(PGSCAN_KSWAPD, nr_scanned); 1929 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD, 1930 nr_scanned); 1931 } else { 1932 if (global_reclaim(sc)) 1933 __count_vm_events(PGSCAN_DIRECT, nr_scanned); 1934 count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT, 1935 nr_scanned); 1936 } 1937 spin_unlock_irq(&pgdat->lru_lock); 1938 1939 if (nr_taken == 0) 1940 return 0; 1941 1942 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0, 1943 &stat, false); 1944 1945 spin_lock_irq(&pgdat->lru_lock); 1946 1947 if (current_is_kswapd()) { 1948 if (global_reclaim(sc)) 1949 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed); 1950 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD, 1951 nr_reclaimed); 1952 } else { 1953 if (global_reclaim(sc)) 1954 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed); 1955 count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT, 1956 nr_reclaimed); 1957 } 1958 1959 putback_inactive_pages(lruvec, &page_list); 1960 1961 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); 1962 1963 spin_unlock_irq(&pgdat->lru_lock); 1964 1965 mem_cgroup_uncharge_list(&page_list); 1966 free_unref_page_list(&page_list); 1967 1968 /* 1969 * If dirty pages are scanned that are not queued for IO, it 1970 * implies that flushers are not doing their job. This can 1971 * happen when memory pressure pushes dirty pages to the end of 1972 * the LRU before the dirty limits are breached and the dirty 1973 * data has expired. It can also happen when the proportion of 1974 * dirty pages grows not through writes but through memory 1975 * pressure reclaiming all the clean cache. And in some cases, 1976 * the flushers simply cannot keep up with the allocation 1977 * rate. Nudge the flusher threads in case they are asleep. 1978 */ 1979 if (stat.nr_unqueued_dirty == nr_taken) 1980 wakeup_flusher_threads(WB_REASON_VMSCAN); 1981 1982 sc->nr.dirty += stat.nr_dirty; 1983 sc->nr.congested += stat.nr_congested; 1984 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty; 1985 sc->nr.writeback += stat.nr_writeback; 1986 sc->nr.immediate += stat.nr_immediate; 1987 sc->nr.taken += nr_taken; 1988 if (file) 1989 sc->nr.file_taken += nr_taken; 1990 1991 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, 1992 nr_scanned, nr_reclaimed, &stat, sc->priority, file); 1993 return nr_reclaimed; 1994 } 1995 1996 /* 1997 * This moves pages from the active list to the inactive list. 1998 * 1999 * We move them the other way if the page is referenced by one or more 2000 * processes, from rmap. 2001 * 2002 * If the pages are mostly unmapped, the processing is fast and it is 2003 * appropriate to hold zone_lru_lock across the whole operation. But if 2004 * the pages are mapped, the processing is slow (page_referenced()) so we 2005 * should drop zone_lru_lock around each page. It's impossible to balance 2006 * this, so instead we remove the pages from the LRU while processing them. 2007 * It is safe to rely on PG_active against the non-LRU pages in here because 2008 * nobody will play with that bit on a non-LRU page. 2009 * 2010 * The downside is that we have to touch page->_refcount against each page. 2011 * But we had to alter page->flags anyway. 2012 * 2013 * Returns the number of pages moved to the given lru. 2014 */ 2015 2016 static unsigned move_active_pages_to_lru(struct lruvec *lruvec, 2017 struct list_head *list, 2018 struct list_head *pages_to_free, 2019 enum lru_list lru) 2020 { 2021 struct pglist_data *pgdat = lruvec_pgdat(lruvec); 2022 struct page *page; 2023 int nr_pages; 2024 int nr_moved = 0; 2025 2026 while (!list_empty(list)) { 2027 page = lru_to_page(list); 2028 lruvec = mem_cgroup_page_lruvec(page, pgdat); 2029 2030 VM_BUG_ON_PAGE(PageLRU(page), page); 2031 SetPageLRU(page); 2032 2033 nr_pages = hpage_nr_pages(page); 2034 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); 2035 list_move(&page->lru, &lruvec->lists[lru]); 2036 2037 if (put_page_testzero(page)) { 2038 __ClearPageLRU(page); 2039 __ClearPageActive(page); 2040 del_page_from_lru_list(page, lruvec, lru); 2041 2042 if (unlikely(PageCompound(page))) { 2043 spin_unlock_irq(&pgdat->lru_lock); 2044 mem_cgroup_uncharge(page); 2045 (*get_compound_page_dtor(page))(page); 2046 spin_lock_irq(&pgdat->lru_lock); 2047 } else 2048 list_add(&page->lru, pages_to_free); 2049 } else { 2050 nr_moved += nr_pages; 2051 } 2052 } 2053 2054 if (!is_active_lru(lru)) { 2055 __count_vm_events(PGDEACTIVATE, nr_moved); 2056 count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, 2057 nr_moved); 2058 } 2059 2060 return nr_moved; 2061 } 2062 2063 static void shrink_active_list(unsigned long nr_to_scan, 2064 struct lruvec *lruvec, 2065 struct scan_control *sc, 2066 enum lru_list lru) 2067 { 2068 unsigned long nr_taken; 2069 unsigned long nr_scanned; 2070 unsigned long vm_flags; 2071 LIST_HEAD(l_hold); /* The pages which were snipped off */ 2072 LIST_HEAD(l_active); 2073 LIST_HEAD(l_inactive); 2074 struct page *page; 2075 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 2076 unsigned nr_deactivate, nr_activate; 2077 unsigned nr_rotated = 0; 2078 isolate_mode_t isolate_mode = 0; 2079 int file = is_file_lru(lru); 2080 struct pglist_data *pgdat = lruvec_pgdat(lruvec); 2081 2082 lru_add_drain(); 2083 2084 if (!sc->may_unmap) 2085 isolate_mode |= ISOLATE_UNMAPPED; 2086 2087 spin_lock_irq(&pgdat->lru_lock); 2088 2089 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, 2090 &nr_scanned, sc, isolate_mode, lru); 2091 2092 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); 2093 reclaim_stat->recent_scanned[file] += nr_taken; 2094 2095 __count_vm_events(PGREFILL, nr_scanned); 2096 count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); 2097 2098 spin_unlock_irq(&pgdat->lru_lock); 2099 2100 while (!list_empty(&l_hold)) { 2101 cond_resched(); 2102 page = lru_to_page(&l_hold); 2103 list_del(&page->lru); 2104 2105 if (unlikely(!page_evictable(page))) { 2106 putback_lru_page(page); 2107 continue; 2108 } 2109 2110 if (unlikely(buffer_heads_over_limit)) { 2111 if (page_has_private(page) && trylock_page(page)) { 2112 if (page_has_private(page)) 2113 try_to_release_page(page, 0); 2114 unlock_page(page); 2115 } 2116 } 2117 2118 if (page_referenced(page, 0, sc->target_mem_cgroup, 2119 &vm_flags)) { 2120 nr_rotated += hpage_nr_pages(page); 2121 /* 2122 * Identify referenced, file-backed active pages and 2123 * give them one more trip around the active list. So 2124 * that executable code get better chances to stay in 2125 * memory under moderate memory pressure. Anon pages 2126 * are not likely to be evicted by use-once streaming 2127 * IO, plus JVM can create lots of anon VM_EXEC pages, 2128 * so we ignore them here. 2129 */ 2130 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { 2131 list_add(&page->lru, &l_active); 2132 continue; 2133 } 2134 } 2135 2136 ClearPageActive(page); /* we are de-activating */ 2137 list_add(&page->lru, &l_inactive); 2138 } 2139 2140 /* 2141 * Move pages back to the lru list. 2142 */ 2143 spin_lock_irq(&pgdat->lru_lock); 2144 /* 2145 * Count referenced pages from currently used mappings as rotated, 2146 * even though only some of them are actually re-activated. This 2147 * helps balance scan pressure between file and anonymous pages in 2148 * get_scan_count. 2149 */ 2150 reclaim_stat->recent_rotated[file] += nr_rotated; 2151 2152 nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); 2153 nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); 2154 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); 2155 spin_unlock_irq(&pgdat->lru_lock); 2156 2157 mem_cgroup_uncharge_list(&l_hold); 2158 free_unref_page_list(&l_hold); 2159 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, 2160 nr_deactivate, nr_rotated, sc->priority, file); 2161 } 2162 2163 /* 2164 * The inactive anon list should be small enough that the VM never has 2165 * to do too much work. 2166 * 2167 * The inactive file list should be small enough to leave most memory 2168 * to the established workingset on the scan-resistant active list, 2169 * but large enough to avoid thrashing the aggregate readahead window. 2170 * 2171 * Both inactive lists should also be large enough that each inactive 2172 * page has a chance to be referenced again before it is reclaimed. 2173 * 2174 * If that fails and refaulting is observed, the inactive list grows. 2175 * 2176 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages 2177 * on this LRU, maintained by the pageout code. An inactive_ratio 2178 * of 3 means 3:1 or 25% of the pages are kept on the inactive list. 2179 * 2180 * total target max 2181 * memory ratio inactive 2182 * ------------------------------------- 2183 * 10MB 1 5MB 2184 * 100MB 1 50MB 2185 * 1GB 3 250MB 2186 * 10GB 10 0.9GB 2187 * 100GB 31 3GB 2188 * 1TB 101 10GB 2189 * 10TB 320 32GB 2190 */ 2191 static bool inactive_list_is_low(struct lruvec *lruvec, bool file, 2192 struct mem_cgroup *memcg, 2193 struct scan_control *sc, bool actual_reclaim) 2194 { 2195 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE; 2196 struct pglist_data *pgdat = lruvec_pgdat(lruvec); 2197 enum lru_list inactive_lru = file * LRU_FILE; 2198 unsigned long inactive, active; 2199 unsigned long inactive_ratio; 2200 unsigned long refaults; 2201 unsigned long gb; 2202 2203 /* 2204 * If we don't have swap space, anonymous page deactivation 2205 * is pointless. 2206 */ 2207 if (!file && !total_swap_pages) 2208 return false; 2209 2210 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx); 2211 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx); 2212 2213 if (memcg) 2214 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); 2215 else 2216 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); 2217 2218 /* 2219 * When refaults are being observed, it means a new workingset 2220 * is being established. Disable active list protection to get 2221 * rid of the stale workingset quickly. 2222 */ 2223 if (file && actual_reclaim && lruvec->refaults != refaults) { 2224 inactive_ratio = 0; 2225 } else { 2226 gb = (inactive + active) >> (30 - PAGE_SHIFT); 2227 if (gb) 2228 inactive_ratio = int_sqrt(10 * gb); 2229 else 2230 inactive_ratio = 1; 2231 } 2232 2233 if (actual_reclaim) 2234 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx, 2235 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive, 2236 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active, 2237 inactive_ratio, file); 2238 2239 return inactive * inactive_ratio < active; 2240 } 2241 2242 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, 2243 struct lruvec *lruvec, struct mem_cgroup *memcg, 2244 struct scan_control *sc) 2245 { 2246 if (is_active_lru(lru)) { 2247 if (inactive_list_is_low(lruvec, is_file_lru(lru), 2248 memcg, sc, true)) 2249 shrink_active_list(nr_to_scan, lruvec, sc, lru); 2250 return 0; 2251 } 2252 2253 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); 2254 } 2255 2256 enum scan_balance { 2257 SCAN_EQUAL, 2258 SCAN_FRACT, 2259 SCAN_ANON, 2260 SCAN_FILE, 2261 }; 2262 2263 /* 2264 * Determine how aggressively the anon and file LRU lists should be 2265 * scanned. The relative value of each set of LRU lists is determined 2266 * by looking at the fraction of the pages scanned we did rotate back 2267 * onto the active list instead of evict. 2268 * 2269 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan 2270 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan 2271 */ 2272 static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, 2273 struct scan_control *sc, unsigned long *nr, 2274 unsigned long *lru_pages) 2275 { 2276 int swappiness = mem_cgroup_swappiness(memcg); 2277 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; 2278 u64 fraction[2]; 2279 u64 denominator = 0; /* gcc */ 2280 struct pglist_data *pgdat = lruvec_pgdat(lruvec); 2281 unsigned long anon_prio, file_prio; 2282 enum scan_balance scan_balance; 2283 unsigned long anon, file; 2284 unsigned long ap, fp; 2285 enum lru_list lru; 2286 2287 /* If we have no swap space, do not bother scanning anon pages. */ 2288 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { 2289 scan_balance = SCAN_FILE; 2290 goto out; 2291 } 2292 2293 /* 2294 * Global reclaim will swap to prevent OOM even with no 2295 * swappiness, but memcg users want to use this knob to 2296 * disable swapping for individual groups completely when 2297 * using the memory controller's swap limit feature would be 2298 * too expensive. 2299 */ 2300 if (!global_reclaim(sc) && !swappiness) { 2301 scan_balance = SCAN_FILE; 2302 goto out; 2303 } 2304 2305 /* 2306 * Do not apply any pressure balancing cleverness when the 2307 * system is close to OOM, scan both anon and file equally 2308 * (unless the swappiness setting disagrees with swapping). 2309 */ 2310 if (!sc->priority && swappiness) { 2311 scan_balance = SCAN_EQUAL; 2312 goto out; 2313 } 2314 2315 /* 2316 * Prevent the reclaimer from falling into the cache trap: as 2317 * cache pages start out inactive, every cache fault will tip 2318 * the scan balance towards the file LRU. And as the file LRU 2319 * shrinks, so does the window for rotation from references. 2320 * This means we have a runaway feedback loop where a tiny 2321 * thrashing file LRU becomes infinitely more attractive than 2322 * anon pages. Try to detect this based on file LRU size. 2323 */ 2324 if (global_reclaim(sc)) { 2325 unsigned long pgdatfile; 2326 unsigned long pgdatfree; 2327 int z; 2328 unsigned long total_high_wmark = 0; 2329 2330 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); 2331 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + 2332 node_page_state(pgdat, NR_INACTIVE_FILE); 2333 2334 for (z = 0; z < MAX_NR_ZONES; z++) { 2335 struct zone *zone = &pgdat->node_zones[z]; 2336 if (!managed_zone(zone)) 2337 continue; 2338 2339 total_high_wmark += high_wmark_pages(zone); 2340 } 2341 2342 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { 2343 /* 2344 * Force SCAN_ANON if there are enough inactive 2345 * anonymous pages on the LRU in eligible zones. 2346 * Otherwise, the small LRU gets thrashed. 2347 */ 2348 if (!inactive_list_is_low(lruvec, false, memcg, sc, false) && 2349 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx) 2350 >> sc->priority) { 2351 scan_balance = SCAN_ANON; 2352 goto out; 2353 } 2354 } 2355 } 2356 2357 /* 2358 * If there is enough inactive page cache, i.e. if the size of the 2359 * inactive list is greater than that of the active list *and* the 2360 * inactive list actually has some pages to scan on this priority, we 2361 * do not reclaim anything from the anonymous working set right now. 2362 * Without the second condition we could end up never scanning an 2363 * lruvec even if it has plenty of old anonymous pages unless the 2364 * system is under heavy pressure. 2365 */ 2366 if (!inactive_list_is_low(lruvec, true, memcg, sc, false) && 2367 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) { 2368 scan_balance = SCAN_FILE; 2369 goto out; 2370 } 2371 2372 scan_balance = SCAN_FRACT; 2373 2374 /* 2375 * With swappiness at 100, anonymous and file have the same priority. 2376 * This scanning priority is essentially the inverse of IO cost. 2377 */ 2378 anon_prio = swappiness; 2379 file_prio = 200 - anon_prio; 2380 2381 /* 2382 * OK, so we have swap space and a fair amount of page cache 2383 * pages. We use the recently rotated / recently scanned 2384 * ratios to determine how valuable each cache is. 2385 * 2386 * Because workloads change over time (and to avoid overflow) 2387 * we keep these statistics as a floating average, which ends 2388 * up weighing recent references more than old ones. 2389 * 2390 * anon in [0], file in [1] 2391 */ 2392 2393 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + 2394 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); 2395 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + 2396 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); 2397 2398 spin_lock_irq(&pgdat->lru_lock); 2399 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { 2400 reclaim_stat->recent_scanned[0] /= 2; 2401 reclaim_stat->recent_rotated[0] /= 2; 2402 } 2403 2404 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { 2405 reclaim_stat->recent_scanned[1] /= 2; 2406 reclaim_stat->recent_rotated[1] /= 2; 2407 } 2408 2409 /* 2410 * The amount of pressure on anon vs file pages is inversely 2411 * proportional to the fraction of recently scanned pages on 2412 * each list that were recently referenced and in active use. 2413 */ 2414 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); 2415 ap /= reclaim_stat->recent_rotated[0] + 1; 2416 2417 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); 2418 fp /= reclaim_stat->recent_rotated[1] + 1; 2419 spin_unlock_irq(&pgdat->lru_lock); 2420 2421 fraction[0] = ap; 2422 fraction[1] = fp; 2423 denominator = ap + fp + 1; 2424 out: 2425 *lru_pages = 0; 2426 for_each_evictable_lru(lru) { 2427 int file = is_file_lru(lru); 2428 unsigned long size; 2429 unsigned long scan; 2430 2431 size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); 2432 scan = size >> sc->priority; 2433 /* 2434 * If the cgroup's already been deleted, make sure to 2435 * scrape out the remaining cache. 2436 */ 2437 if (!scan && !mem_cgroup_online(memcg)) 2438 scan = min(size, SWAP_CLUSTER_MAX); 2439 2440 switch (scan_balance) { 2441 case SCAN_EQUAL: 2442 /* Scan lists relative to size */ 2443 break; 2444 case SCAN_FRACT: 2445 /* 2446 * Scan types proportional to swappiness and 2447 * their relative recent reclaim efficiency. 2448 */ 2449 scan = div64_u64(scan * fraction[file], 2450 denominator); 2451 break; 2452 case SCAN_FILE: 2453 case SCAN_ANON: 2454 /* Scan one type exclusively */ 2455 if ((scan_balance == SCAN_FILE) != file) { 2456 size = 0; 2457 scan = 0; 2458 } 2459 break; 2460 default: 2461 /* Look ma, no brain */ 2462 BUG(); 2463 } 2464 2465 *lru_pages += size; 2466 nr[lru] = scan; 2467 } 2468 } 2469 2470 /* 2471 * This is a basic per-node page freer. Used by both kswapd and direct reclaim. 2472 */ 2473 static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg, 2474 struct scan_control *sc, unsigned long *lru_pages) 2475 { 2476 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); 2477 unsigned long nr[NR_LRU_LISTS]; 2478 unsigned long targets[NR_LRU_LISTS]; 2479 unsigned long nr_to_scan; 2480 enum lru_list lru; 2481 unsigned long nr_reclaimed = 0; 2482 unsigned long nr_to_reclaim = sc->nr_to_reclaim; 2483 struct blk_plug plug; 2484 bool scan_adjusted; 2485 2486 get_scan_count(lruvec, memcg, sc, nr, lru_pages); 2487 2488 /* Record the original scan target for proportional adjustments later */ 2489 memcpy(targets, nr, sizeof(nr)); 2490 2491 /* 2492 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal 2493 * event that can occur when there is little memory pressure e.g. 2494 * multiple streaming readers/writers. Hence, we do not abort scanning 2495 * when the requested number of pages are reclaimed when scanning at 2496 * DEF_PRIORITY on the assumption that the fact we are direct 2497 * reclaiming implies that kswapd is not keeping up and it is best to 2498 * do a batch of work at once. For memcg reclaim one check is made to 2499 * abort proportional reclaim if either the file or anon lru has already 2500 * dropped to zero at the first pass. 2501 */ 2502 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && 2503 sc->priority == DEF_PRIORITY); 2504 2505 blk_start_plug(&plug); 2506 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || 2507 nr[LRU_INACTIVE_FILE]) { 2508 unsigned long nr_anon, nr_file, percentage; 2509 unsigned long nr_scanned; 2510 2511 for_each_evictable_lru(lru) { 2512 if (nr[lru]) { 2513 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); 2514 nr[lru] -= nr_to_scan; 2515 2516 nr_reclaimed += shrink_list(lru, nr_to_scan, 2517 lruvec, memcg, sc); 2518 } 2519 } 2520 2521 cond_resched(); 2522 2523 if (nr_reclaimed < nr_to_reclaim || scan_adjusted) 2524 continue; 2525 2526 /* 2527 * For kswapd and memcg, reclaim at least the number of pages 2528 * requested. Ensure that the anon and file LRUs are scanned 2529 * proportionally what was requested by get_scan_count(). We 2530 * stop reclaiming one LRU and reduce the amount scanning 2531 * proportional to the original scan target. 2532 */ 2533 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; 2534 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; 2535 2536 /* 2537 * It's just vindictive to attack the larger once the smaller 2538 * has gone to zero. And given the way we stop scanning the 2539 * smaller below, this makes sure that we only make one nudge 2540 * towards proportionality once we've got nr_to_reclaim. 2541 */ 2542 if (!nr_file || !nr_anon) 2543 break; 2544 2545 if (nr_file > nr_anon) { 2546 unsigned long scan_target = targets[LRU_INACTIVE_ANON] + 2547 targets[LRU_ACTIVE_ANON] + 1; 2548 lru = LRU_BASE; 2549 percentage = nr_anon * 100 / scan_target; 2550 } else { 2551 unsigned long scan_target = targets[LRU_INACTIVE_FILE] + 2552 targets[LRU_ACTIVE_FILE] + 1; 2553 lru = LRU_FILE; 2554 percentage = nr_file * 100 / scan_target; 2555 } 2556 2557 /* Stop scanning the smaller of the LRU */ 2558 nr[lru] = 0; 2559 nr[lru + LRU_ACTIVE] = 0; 2560 2561 /* 2562 * Recalculate the other LRU scan count based on its original 2563 * scan target and the percentage scanning already complete 2564 */ 2565 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; 2566 nr_scanned = targets[lru] - nr[lru]; 2567 nr[lru] = targets[lru] * (100 - percentage) / 100; 2568 nr[lru] -= min(nr[lru], nr_scanned); 2569 2570 lru += LRU_ACTIVE; 2571 nr_scanned = targets[lru] - nr[lru]; 2572 nr[lru] = targets[lru] * (100 - percentage) / 100; 2573 nr[lru] -= min(nr[lru], nr_scanned); 2574 2575 scan_adjusted = true; 2576 } 2577 blk_finish_plug(&plug); 2578 sc->nr_reclaimed += nr_reclaimed; 2579 2580 /* 2581 * Even if we did not try to evict anon pages at all, we want to 2582 * rebalance the anon lru active/inactive ratio. 2583 */ 2584 if (inactive_list_is_low(lruvec, false, memcg, sc, true)) 2585 shrink_active_list(SWAP_CLUSTER_MAX, lruvec, 2586 sc, LRU_ACTIVE_ANON); 2587 } 2588 2589 /* Use reclaim/compaction for costly allocs or under memory pressure */ 2590 static bool in_reclaim_compaction(struct scan_control *sc) 2591 { 2592 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && 2593 (sc->order > PAGE_ALLOC_COSTLY_ORDER || 2594 sc->priority < DEF_PRIORITY - 2)) 2595 return true; 2596 2597 return false; 2598 } 2599 2600 /* 2601 * Reclaim/compaction is used for high-order allocation requests. It reclaims 2602 * order-0 pages before compacting the zone. should_continue_reclaim() returns 2603 * true if more pages should be reclaimed such that when the page allocator 2604 * calls try_to_compact_zone() that it will have enough free pages to succeed. 2605 * It will give up earlier than that if there is difficulty reclaiming pages. 2606 */ 2607 static inline bool should_continue_reclaim(struct pglist_data *pgdat, 2608 unsigned long nr_reclaimed, 2609 unsigned long nr_scanned, 2610 struct scan_control *sc) 2611 { 2612 unsigned long pages_for_compaction; 2613 unsigned long inactive_lru_pages; 2614 int z; 2615 2616 /* If not in reclaim/compaction mode, stop */ 2617 if (!in_reclaim_compaction(sc)) 2618 return false; 2619 2620 /* Consider stopping depending on scan and reclaim activity */ 2621 if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) { 2622 /* 2623 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the 2624 * full LRU list has been scanned and we are still failing 2625 * to reclaim pages. This full LRU scan is potentially 2626 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed 2627 */ 2628 if (!nr_reclaimed && !nr_scanned) 2629 return false; 2630 } else { 2631 /* 2632 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably 2633 * fail without consequence, stop if we failed to reclaim 2634 * any pages from the last SWAP_CLUSTER_MAX number of 2635 * pages that were scanned. This will return to the 2636 * caller faster at the risk reclaim/compaction and 2637 * the resulting allocation attempt fails 2638 */ 2639 if (!nr_reclaimed) 2640 return false; 2641 } 2642 2643 /* 2644 * If we have not reclaimed enough pages for compaction and the 2645 * inactive lists are large enough, continue reclaiming 2646 */ 2647 pages_for_compaction = compact_gap(sc->order); 2648 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); 2649 if (get_nr_swap_pages() > 0) 2650 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); 2651 if (sc->nr_reclaimed < pages_for_compaction && 2652 inactive_lru_pages > pages_for_compaction) 2653 return true; 2654 2655 /* If compaction would go ahead or the allocation would succeed, stop */ 2656 for (z = 0; z <= sc->reclaim_idx; z++) { 2657 struct zone *zone = &pgdat->node_zones[z]; 2658 if (!managed_zone(zone)) 2659 continue; 2660 2661 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { 2662 case COMPACT_SUCCESS: 2663 case COMPACT_CONTINUE: 2664 return false; 2665 default: 2666 /* check next zone */ 2667 ; 2668 } 2669 } 2670 return true; 2671 } 2672 2673 static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg) 2674 { 2675 return test_bit(PGDAT_CONGESTED, &pgdat->flags) || 2676 (memcg && memcg_congested(pgdat, memcg)); 2677 } 2678 2679 static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) 2680 { 2681 struct reclaim_state *reclaim_state = current->reclaim_state; 2682 unsigned long nr_reclaimed, nr_scanned; 2683 bool reclaimable = false; 2684 2685 do { 2686 struct mem_cgroup *root = sc->target_mem_cgroup; 2687 struct mem_cgroup_reclaim_cookie reclaim = { 2688 .pgdat = pgdat, 2689 .priority = sc->priority, 2690 }; 2691 unsigned long node_lru_pages = 0; 2692 struct mem_cgroup *memcg; 2693 2694 memset(&sc->nr, 0, sizeof(sc->nr)); 2695 2696 nr_reclaimed = sc->nr_reclaimed; 2697 nr_scanned = sc->nr_scanned; 2698 2699 memcg = mem_cgroup_iter(root, NULL, &reclaim); 2700 do { 2701 unsigned long lru_pages; 2702 unsigned long reclaimed; 2703 unsigned long scanned; 2704 2705 switch (mem_cgroup_protected(root, memcg)) { 2706 case MEMCG_PROT_MIN: 2707 /* 2708 * Hard protection. 2709 * If there is no reclaimable memory, OOM. 2710 */ 2711 continue; 2712 case MEMCG_PROT_LOW: 2713 /* 2714 * Soft protection. 2715 * Respect the protection only as long as 2716 * there is an unprotected supply 2717 * of reclaimable memory from other cgroups. 2718 */ 2719 if (!sc->memcg_low_reclaim) { 2720 sc->memcg_low_skipped = 1; 2721 continue; 2722 } 2723 memcg_memory_event(memcg, MEMCG_LOW); 2724 break; 2725 case MEMCG_PROT_NONE: 2726 break; 2727 } 2728 2729 reclaimed = sc->nr_reclaimed; 2730 scanned = sc->nr_scanned; 2731 shrink_node_memcg(pgdat, memcg, sc, &lru_pages); 2732 node_lru_pages += lru_pages; 2733 2734 shrink_slab(sc->gfp_mask, pgdat->node_id, 2735 memcg, sc->priority); 2736 2737 /* Record the group's reclaim efficiency */ 2738 vmpressure(sc->gfp_mask, memcg, false, 2739 sc->nr_scanned - scanned, 2740 sc->nr_reclaimed - reclaimed); 2741 2742 /* 2743 * Direct reclaim and kswapd have to scan all memory 2744 * cgroups to fulfill the overall scan target for the 2745 * node. 2746 * 2747 * Limit reclaim, on the other hand, only cares about 2748 * nr_to_reclaim pages to be reclaimed and it will 2749 * retry with decreasing priority if one round over the 2750 * whole hierarchy is not sufficient. 2751 */ 2752 if (!global_reclaim(sc) && 2753 sc->nr_reclaimed >= sc->nr_to_reclaim) { 2754 mem_cgroup_iter_break(root, memcg); 2755 break; 2756 } 2757 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); 2758 2759 if (reclaim_state) { 2760 sc->nr_reclaimed += reclaim_state->reclaimed_slab; 2761 reclaim_state->reclaimed_slab = 0; 2762 } 2763 2764 /* Record the subtree's reclaim efficiency */ 2765 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, 2766 sc->nr_scanned - nr_scanned, 2767 sc->nr_reclaimed - nr_reclaimed); 2768 2769 if (sc->nr_reclaimed - nr_reclaimed) 2770 reclaimable = true; 2771 2772 if (current_is_kswapd()) { 2773 /* 2774 * If reclaim is isolating dirty pages under writeback, 2775 * it implies that the long-lived page allocation rate 2776 * is exceeding the page laundering rate. Either the 2777 * global limits are not being effective at throttling 2778 * processes due to the page distribution throughout 2779 * zones or there is heavy usage of a slow backing 2780 * device. The only option is to throttle from reclaim 2781 * context which is not ideal as there is no guarantee 2782 * the dirtying process is throttled in the same way 2783 * balance_dirty_pages() manages. 2784 * 2785 * Once a node is flagged PGDAT_WRITEBACK, kswapd will 2786 * count the number of pages under pages flagged for 2787 * immediate reclaim and stall if any are encountered 2788 * in the nr_immediate check below. 2789 */ 2790 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken) 2791 set_bit(PGDAT_WRITEBACK, &pgdat->flags); 2792 2793 /* 2794 * Tag a node as congested if all the dirty pages 2795 * scanned were backed by a congested BDI and 2796 * wait_iff_congested will stall. 2797 */ 2798 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) 2799 set_bit(PGDAT_CONGESTED, &pgdat->flags); 2800 2801 /* Allow kswapd to start writing pages during reclaim.*/ 2802 if (sc->nr.unqueued_dirty == sc->nr.file_taken) 2803 set_bit(PGDAT_DIRTY, &pgdat->flags); 2804 2805 /* 2806 * If kswapd scans pages marked marked for immediate 2807 * reclaim and under writeback (nr_immediate), it 2808 * implies that pages are cycling through the LRU 2809 * faster than they are written so also forcibly stall. 2810 */ 2811 if (sc->nr.immediate) 2812 congestion_wait(BLK_RW_ASYNC, HZ/10); 2813 } 2814 2815 /* 2816 * Legacy memcg will stall in page writeback so avoid forcibly 2817 * stalling in wait_iff_congested(). 2818 */ 2819 if (!global_reclaim(sc) && sane_reclaim(sc) && 2820 sc->nr.dirty && sc->nr.dirty == sc->nr.congested) 2821 set_memcg_congestion(pgdat, root, true); 2822 2823 /* 2824 * Stall direct reclaim for IO completions if underlying BDIs 2825 * and node is congested. Allow kswapd to continue until it 2826 * starts encountering unqueued dirty pages or cycling through 2827 * the LRU too quickly. 2828 */ 2829 if (!sc->hibernation_mode && !current_is_kswapd() && 2830 current_may_throttle() && pgdat_memcg_congested(pgdat, root)) 2831 wait_iff_congested(BLK_RW_ASYNC, HZ/10); 2832 2833 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, 2834 sc->nr_scanned - nr_scanned, sc)); 2835 2836 /* 2837 * Kswapd gives up on balancing particular nodes after too 2838 * many failures to reclaim anything from them and goes to 2839 * sleep. On reclaim progress, reset the failure counter. A 2840 * successful direct reclaim run will revive a dormant kswapd. 2841 */ 2842 if (reclaimable) 2843 pgdat->kswapd_failures = 0; 2844 2845 return reclaimable; 2846 } 2847 2848 /* 2849 * Returns true if compaction should go ahead for a costly-order request, or 2850 * the allocation would already succeed without compaction. Return false if we 2851 * should reclaim first. 2852 */ 2853 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) 2854 { 2855 unsigned long watermark; 2856 enum compact_result suitable; 2857 2858 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); 2859 if (suitable == COMPACT_SUCCESS) 2860 /* Allocation should succeed already. Don't reclaim. */ 2861 return true; 2862 if (suitable == COMPACT_SKIPPED) 2863 /* Compaction cannot yet proceed. Do reclaim. */ 2864 return false; 2865 2866 /* 2867 * Compaction is already possible, but it takes time to run and there 2868 * are potentially other callers using the pages just freed. So proceed 2869 * with reclaim to make a buffer of free pages available to give 2870 * compaction a reasonable chance of completing and allocating the page. 2871 * Note that we won't actually reclaim the whole buffer in one attempt 2872 * as the target watermark in should_continue_reclaim() is lower. But if 2873 * we are already above the high+gap watermark, don't reclaim at all. 2874 */ 2875 watermark = high_wmark_pages(zone) + compact_gap(sc->order); 2876 2877 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); 2878 } 2879 2880 /* 2881 * This is the direct reclaim path, for page-allocating processes. We only 2882 * try to reclaim pages from zones which will satisfy the caller's allocation 2883 * request. 2884 * 2885 * If a zone is deemed to be full of pinned pages then just give it a light 2886 * scan then give up on it. 2887 */ 2888 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) 2889 { 2890 struct zoneref *z; 2891 struct zone *zone; 2892 unsigned long nr_soft_reclaimed; 2893 unsigned long nr_soft_scanned; 2894 gfp_t orig_mask; 2895 pg_data_t *last_pgdat = NULL; 2896 2897 /* 2898 * If the number of buffer_heads in the machine exceeds the maximum 2899 * allowed level, force direct reclaim to scan the highmem zone as 2900 * highmem pages could be pinning lowmem pages storing buffer_heads 2901 */ 2902 orig_mask = sc->gfp_mask; 2903 if (buffer_heads_over_limit) { 2904 sc->gfp_mask |= __GFP_HIGHMEM; 2905 sc->reclaim_idx = gfp_zone(sc->gfp_mask); 2906 } 2907 2908 for_each_zone_zonelist_nodemask(zone, z, zonelist, 2909 sc->reclaim_idx, sc->nodemask) { 2910 /* 2911 * Take care memory controller reclaiming has small influence 2912 * to global LRU. 2913 */ 2914 if (global_reclaim(sc)) { 2915 if (!cpuset_zone_allowed(zone, 2916 GFP_KERNEL | __GFP_HARDWALL)) 2917 continue; 2918 2919 /* 2920 * If we already have plenty of memory free for 2921 * compaction in this zone, don't free any more. 2922 * Even though compaction is invoked for any 2923 * non-zero order, only frequent costly order 2924 * reclamation is disruptive enough to become a 2925 * noticeable problem, like transparent huge 2926 * page allocations. 2927 */ 2928 if (IS_ENABLED(CONFIG_COMPACTION) && 2929 sc->order > PAGE_ALLOC_COSTLY_ORDER && 2930 compaction_ready(zone, sc)) { 2931 sc->compaction_ready = true; 2932 continue; 2933 } 2934 2935 /* 2936 * Shrink each node in the zonelist once. If the 2937 * zonelist is ordered by zone (not the default) then a 2938 * node may be shrunk multiple times but in that case 2939 * the user prefers lower zones being preserved. 2940 */ 2941 if (zone->zone_pgdat == last_pgdat) 2942 continue; 2943 2944 /* 2945 * This steals pages from memory cgroups over softlimit 2946 * and returns the number of reclaimed pages and 2947 * scanned pages. This works for global memory pressure 2948 * and balancing, not for a memcg's limit. 2949 */ 2950 nr_soft_scanned = 0; 2951 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, 2952 sc->order, sc->gfp_mask, 2953 &nr_soft_scanned); 2954 sc->nr_reclaimed += nr_soft_reclaimed; 2955 sc->nr_scanned += nr_soft_scanned; 2956 /* need some check for avoid more shrink_zone() */ 2957 } 2958 2959 /* See comment about same check for global reclaim above */ 2960 if (zone->zone_pgdat == last_pgdat) 2961 continue; 2962 last_pgdat = zone->zone_pgdat; 2963 shrink_node(zone->zone_pgdat, sc); 2964 } 2965 2966 /* 2967 * Restore to original mask to avoid the impact on the caller if we 2968 * promoted it to __GFP_HIGHMEM. 2969 */ 2970 sc->gfp_mask = orig_mask; 2971 } 2972 2973 static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat) 2974 { 2975 struct mem_cgroup *memcg; 2976 2977 memcg = mem_cgroup_iter(root_memcg, NULL, NULL); 2978 do { 2979 unsigned long refaults; 2980 struct lruvec *lruvec; 2981 2982 if (memcg) 2983 refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); 2984 else 2985 refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); 2986 2987 lruvec = mem_cgroup_lruvec(pgdat, memcg); 2988 lruvec->refaults = refaults; 2989 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL))); 2990 } 2991 2992 /* 2993 * This is the main entry point to direct page reclaim. 2994 * 2995 * If a full scan of the inactive list fails to free enough memory then we 2996 * are "out of memory" and something needs to be killed. 2997 * 2998 * If the caller is !__GFP_FS then the probability of a failure is reasonably 2999 * high - the zone may be full of dirty or under-writeback pages, which this 3000 * caller can't do much about. We kick the writeback threads and take explicit 3001 * naps in the hope that some of these pages can be written. But if the 3002 * allocating task holds filesystem locks which prevent writeout this might not 3003 * work, and the allocation attempt will fail. 3004 * 3005 * returns: 0, if no pages reclaimed 3006 * else, the number of pages reclaimed 3007 */ 3008 static unsigned long do_try_to_free_pages(struct zonelist *zonelist, 3009 struct scan_control *sc) 3010 { 3011 int initial_priority = sc->priority; 3012 pg_data_t *last_pgdat; 3013 struct zoneref *z; 3014 struct zone *zone; 3015 retry: 3016 delayacct_freepages_start(); 3017 3018 if (global_reclaim(sc)) 3019 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); 3020 3021 do { 3022 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, 3023 sc->priority); 3024 sc->nr_scanned = 0; 3025 shrink_zones(zonelist, sc); 3026 3027 if (sc->nr_reclaimed >= sc->nr_to_reclaim) 3028 break; 3029 3030 if (sc->compaction_ready) 3031 break; 3032 3033 /* 3034 * If we're getting trouble reclaiming, start doing 3035 * writepage even in laptop mode. 3036 */ 3037 if (sc->priority < DEF_PRIORITY - 2) 3038 sc->may_writepage = 1; 3039 } while (--sc->priority >= 0); 3040 3041 last_pgdat = NULL; 3042 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, 3043 sc->nodemask) { 3044 if (zone->zone_pgdat == last_pgdat) 3045 continue; 3046 last_pgdat = zone->zone_pgdat; 3047 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); 3048 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false); 3049 } 3050 3051 delayacct_freepages_end(); 3052 3053 if (sc->nr_reclaimed) 3054 return sc->nr_reclaimed; 3055 3056 /* Aborted reclaim to try compaction? don't OOM, then */ 3057 if (sc->compaction_ready) 3058 return 1; 3059 3060 /* Untapped cgroup reserves? Don't OOM, retry. */ 3061 if (sc->memcg_low_skipped) { 3062 sc->priority = initial_priority; 3063 sc->memcg_low_reclaim = 1; 3064 sc->memcg_low_skipped = 0; 3065 goto retry; 3066 } 3067 3068 return 0; 3069 } 3070 3071 static bool allow_direct_reclaim(pg_data_t *pgdat) 3072 { 3073 struct zone *zone; 3074 unsigned long pfmemalloc_reserve = 0; 3075 unsigned long free_pages = 0; 3076 int i; 3077 bool wmark_ok; 3078 3079 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) 3080 return true; 3081 3082 for (i = 0; i <= ZONE_NORMAL; i++) { 3083 zone = &pgdat->node_zones[i]; 3084 if (!managed_zone(zone)) 3085 continue; 3086 3087 if (!zone_reclaimable_pages(zone)) 3088 continue; 3089 3090 pfmemalloc_reserve += min_wmark_pages(zone); 3091 free_pages += zone_page_state(zone, NR_FREE_PAGES); 3092 } 3093 3094 /* If there are no reserves (unexpected config) then do not throttle */ 3095 if (!pfmemalloc_reserve) 3096 return true; 3097 3098 wmark_ok = free_pages > pfmemalloc_reserve / 2; 3099 3100 /* kswapd must be awake if processes are being throttled */ 3101 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { 3102 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, 3103 (enum zone_type)ZONE_NORMAL); 3104 wake_up_interruptible(&pgdat->kswapd_wait); 3105 } 3106 3107 return wmark_ok; 3108 } 3109 3110 /* 3111 * Throttle direct reclaimers if backing storage is backed by the network 3112 * and the PFMEMALLOC reserve for the preferred node is getting dangerously 3113 * depleted. kswapd will continue to make progress and wake the processes 3114 * when the low watermark is reached. 3115 * 3116 * Returns true if a fatal signal was delivered during throttling. If this 3117 * happens, the page allocator should not consider triggering the OOM killer. 3118 */ 3119 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, 3120 nodemask_t *nodemask) 3121 { 3122 struct zoneref *z; 3123 struct zone *zone; 3124 pg_data_t *pgdat = NULL; 3125 3126 /* 3127 * Kernel threads should not be throttled as they may be indirectly 3128 * responsible for cleaning pages necessary for reclaim to make forward 3129 * progress. kjournald for example may enter direct reclaim while 3130 * committing a transaction where throttling it could forcing other 3131 * processes to block on log_wait_commit(). 3132 */ 3133 if (current->flags & PF_KTHREAD) 3134 goto out; 3135 3136 /* 3137 * If a fatal signal is pending, this process should not throttle. 3138 * It should return quickly so it can exit and free its memory 3139 */ 3140 if (fatal_signal_pending(current)) 3141 goto out; 3142 3143 /* 3144 * Check if the pfmemalloc reserves are ok by finding the first node 3145 * with a usable ZONE_NORMAL or lower zone. The expectation is that 3146 * GFP_KERNEL will be required for allocating network buffers when 3147 * swapping over the network so ZONE_HIGHMEM is unusable. 3148 * 3149 * Throttling is based on the first usable node and throttled processes 3150 * wait on a queue until kswapd makes progress and wakes them. There 3151 * is an affinity then between processes waking up and where reclaim 3152 * progress has been made assuming the process wakes on the same node. 3153 * More importantly, processes running on remote nodes will not compete 3154 * for remote pfmemalloc reserves and processes on different nodes 3155 * should make reasonable progress. 3156 */ 3157 for_each_zone_zonelist_nodemask(zone, z, zonelist, 3158 gfp_zone(gfp_mask), nodemask) { 3159 if (zone_idx(zone) > ZONE_NORMAL) 3160 continue; 3161 3162 /* Throttle based on the first usable node */ 3163 pgdat = zone->zone_pgdat; 3164 if (allow_direct_reclaim(pgdat)) 3165 goto out; 3166 break; 3167 } 3168 3169 /* If no zone was usable by the allocation flags then do not throttle */ 3170 if (!pgdat) 3171 goto out; 3172 3173 /* Account for the throttling */ 3174 count_vm_event(PGSCAN_DIRECT_THROTTLE); 3175 3176 /* 3177 * If the caller cannot enter the filesystem, it's possible that it 3178 * is due to the caller holding an FS lock or performing a journal 3179 * transaction in the case of a filesystem like ext[3|4]. In this case, 3180 * it is not safe to block on pfmemalloc_wait as kswapd could be 3181 * blocked waiting on the same lock. Instead, throttle for up to a 3182 * second before continuing. 3183 */ 3184 if (!(gfp_mask & __GFP_FS)) { 3185 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, 3186 allow_direct_reclaim(pgdat), HZ); 3187 3188 goto check_pending; 3189 } 3190 3191 /* Throttle until kswapd wakes the process */ 3192 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, 3193 allow_direct_reclaim(pgdat)); 3194 3195 check_pending: 3196 if (fatal_signal_pending(current)) 3197 return true; 3198 3199 out: 3200 return false; 3201 } 3202 3203 unsigned long try_to_free_pages(struct zonelist *zonelist, int order, 3204 gfp_t gfp_mask, nodemask_t *nodemask) 3205 { 3206 unsigned long nr_reclaimed; 3207 struct scan_control sc = { 3208 .nr_to_reclaim = SWAP_CLUSTER_MAX, 3209 .gfp_mask = current_gfp_context(gfp_mask), 3210 .reclaim_idx = gfp_zone(gfp_mask), 3211 .order = order, 3212 .nodemask = nodemask, 3213 .priority = DEF_PRIORITY, 3214 .may_writepage = !laptop_mode, 3215 .may_unmap = 1, 3216 .may_swap = 1, 3217 }; 3218 3219 /* 3220 * scan_control uses s8 fields for order, priority, and reclaim_idx. 3221 * Confirm they are large enough for max values. 3222 */ 3223 BUILD_BUG_ON(MAX_ORDER > S8_MAX); 3224 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX); 3225 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX); 3226 3227 /* 3228 * Do not enter reclaim if fatal signal was delivered while throttled. 3229 * 1 is returned so that the page allocator does not OOM kill at this 3230 * point. 3231 */ 3232 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) 3233 return 1; 3234 3235 trace_mm_vmscan_direct_reclaim_begin(order, 3236 sc.may_writepage, 3237 sc.gfp_mask, 3238 sc.reclaim_idx); 3239 3240 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 3241 3242 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); 3243 3244 return nr_reclaimed; 3245 } 3246 3247 #ifdef CONFIG_MEMCG 3248 3249 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, 3250 gfp_t gfp_mask, bool noswap, 3251 pg_data_t *pgdat, 3252 unsigned long *nr_scanned) 3253 { 3254 struct scan_control sc = { 3255 .nr_to_reclaim = SWAP_CLUSTER_MAX, 3256 .target_mem_cgroup = memcg, 3257 .may_writepage = !laptop_mode, 3258 .may_unmap = 1, 3259 .reclaim_idx = MAX_NR_ZONES - 1, 3260 .may_swap = !noswap, 3261 }; 3262 unsigned long lru_pages; 3263 3264 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | 3265 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); 3266 3267 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, 3268 sc.may_writepage, 3269 sc.gfp_mask, 3270 sc.reclaim_idx); 3271 3272 /* 3273 * NOTE: Although we can get the priority field, using it 3274 * here is not a good idea, since it limits the pages we can scan. 3275 * if we don't reclaim here, the shrink_node from balance_pgdat 3276 * will pick up pages from other mem cgroup's as well. We hack 3277 * the priority and make it zero. 3278 */ 3279 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages); 3280 3281 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); 3282 3283 *nr_scanned = sc.nr_scanned; 3284 return sc.nr_reclaimed; 3285 } 3286 3287 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, 3288 unsigned long nr_pages, 3289 gfp_t gfp_mask, 3290 bool may_swap) 3291 { 3292 struct zonelist *zonelist; 3293 unsigned long nr_reclaimed; 3294 int nid; 3295 unsigned int noreclaim_flag; 3296 struct scan_control sc = { 3297 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), 3298 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | 3299 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), 3300 .reclaim_idx = MAX_NR_ZONES - 1, 3301 .target_mem_cgroup = memcg, 3302 .priority = DEF_PRIORITY, 3303 .may_writepage = !laptop_mode, 3304 .may_unmap = 1, 3305 .may_swap = may_swap, 3306 }; 3307 3308 /* 3309 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't 3310 * take care of from where we get pages. So the node where we start the 3311 * scan does not need to be the current node. 3312 */ 3313 nid = mem_cgroup_select_victim_node(memcg); 3314 3315 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK]; 3316 3317 trace_mm_vmscan_memcg_reclaim_begin(0, 3318 sc.may_writepage, 3319 sc.gfp_mask, 3320 sc.reclaim_idx); 3321 3322 noreclaim_flag = memalloc_noreclaim_save(); 3323 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 3324 memalloc_noreclaim_restore(noreclaim_flag); 3325 3326 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); 3327 3328 return nr_reclaimed; 3329 } 3330 #endif 3331 3332 static void age_active_anon(struct pglist_data *pgdat, 3333 struct scan_control *sc) 3334 { 3335 struct mem_cgroup *memcg; 3336 3337 if (!total_swap_pages) 3338 return; 3339 3340 memcg = mem_cgroup_iter(NULL, NULL, NULL); 3341 do { 3342 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); 3343 3344 if (inactive_list_is_low(lruvec, false, memcg, sc, true)) 3345 shrink_active_list(SWAP_CLUSTER_MAX, lruvec, 3346 sc, LRU_ACTIVE_ANON); 3347 3348 memcg = mem_cgroup_iter(NULL, memcg, NULL); 3349 } while (memcg); 3350 } 3351 3352 /* 3353 * Returns true if there is an eligible zone balanced for the request order 3354 * and classzone_idx 3355 */ 3356 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) 3357 { 3358 int i; 3359 unsigned long mark = -1; 3360 struct zone *zone; 3361 3362 for (i = 0; i <= classzone_idx; i++) { 3363 zone = pgdat->node_zones + i; 3364 3365 if (!managed_zone(zone)) 3366 continue; 3367 3368 mark = high_wmark_pages(zone); 3369 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx)) 3370 return true; 3371 } 3372 3373 /* 3374 * If a node has no populated zone within classzone_idx, it does not 3375 * need balancing by definition. This can happen if a zone-restricted 3376 * allocation tries to wake a remote kswapd. 3377 */ 3378 if (mark == -1) 3379 return true; 3380 3381 return false; 3382 } 3383 3384 /* Clear pgdat state for congested, dirty or under writeback. */ 3385 static void clear_pgdat_congested(pg_data_t *pgdat) 3386 { 3387 clear_bit(PGDAT_CONGESTED, &pgdat->flags); 3388 clear_bit(PGDAT_DIRTY, &pgdat->flags); 3389 clear_bit(PGDAT_WRITEBACK, &pgdat->flags); 3390 } 3391 3392 /* 3393 * Prepare kswapd for sleeping. This verifies that there are no processes 3394 * waiting in throttle_direct_reclaim() and that watermarks have been met. 3395 * 3396 * Returns true if kswapd is ready to sleep 3397 */ 3398 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) 3399 { 3400 /* 3401 * The throttled processes are normally woken up in balance_pgdat() as 3402 * soon as allow_direct_reclaim() is true. But there is a potential 3403 * race between when kswapd checks the watermarks and a process gets 3404 * throttled. There is also a potential race if processes get 3405 * throttled, kswapd wakes, a large process exits thereby balancing the 3406 * zones, which causes kswapd to exit balance_pgdat() before reaching 3407 * the wake up checks. If kswapd is going to sleep, no process should 3408 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If 3409 * the wake up is premature, processes will wake kswapd and get 3410 * throttled again. The difference from wake ups in balance_pgdat() is 3411 * that here we are under prepare_to_wait(). 3412 */ 3413 if (waitqueue_active(&pgdat->pfmemalloc_wait)) 3414 wake_up_all(&pgdat->pfmemalloc_wait); 3415 3416 /* Hopeless node, leave it to direct reclaim */ 3417 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) 3418 return true; 3419 3420 if (pgdat_balanced(pgdat, order, classzone_idx)) { 3421 clear_pgdat_congested(pgdat); 3422 return true; 3423 } 3424 3425 return false; 3426 } 3427 3428 /* 3429 * kswapd shrinks a node of pages that are at or below the highest usable 3430 * zone that is currently unbalanced. 3431 * 3432 * Returns true if kswapd scanned at least the requested number of pages to 3433 * reclaim or if the lack of progress was due to pages under writeback. 3434 * This is used to determine if the scanning priority needs to be raised. 3435 */ 3436 static bool kswapd_shrink_node(pg_data_t *pgdat, 3437 struct scan_control *sc) 3438 { 3439 struct zone *zone; 3440 int z; 3441 3442 /* Reclaim a number of pages proportional to the number of zones */ 3443 sc->nr_to_reclaim = 0; 3444 for (z = 0; z <= sc->reclaim_idx; z++) { 3445 zone = pgdat->node_zones + z; 3446 if (!managed_zone(zone)) 3447 continue; 3448 3449 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); 3450 } 3451 3452 /* 3453 * Historically care was taken to put equal pressure on all zones but 3454 * now pressure is applied based on node LRU order. 3455 */ 3456 shrink_node(pgdat, sc); 3457 3458 /* 3459 * Fragmentation may mean that the system cannot be rebalanced for 3460 * high-order allocations. If twice the allocation size has been 3461 * reclaimed then recheck watermarks only at order-0 to prevent 3462 * excessive reclaim. Assume that a process requested a high-order 3463 * can direct reclaim/compact. 3464 */ 3465 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) 3466 sc->order = 0; 3467 3468 return sc->nr_scanned >= sc->nr_to_reclaim; 3469 } 3470 3471 /* 3472 * For kswapd, balance_pgdat() will reclaim pages across a node from zones 3473 * that are eligible for use by the caller until at least one zone is 3474 * balanced. 3475 * 3476 * Returns the order kswapd finished reclaiming at. 3477 * 3478 * kswapd scans the zones in the highmem->normal->dma direction. It skips 3479 * zones which have free_pages > high_wmark_pages(zone), but once a zone is 3480 * found to have free_pages <= high_wmark_pages(zone), any page is that zone 3481 * or lower is eligible for reclaim until at least one usable zone is 3482 * balanced. 3483 */ 3484 static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) 3485 { 3486 int i; 3487 unsigned long nr_soft_reclaimed; 3488 unsigned long nr_soft_scanned; 3489 struct zone *zone; 3490 struct scan_control sc = { 3491 .gfp_mask = GFP_KERNEL, 3492 .order = order, 3493 .priority = DEF_PRIORITY, 3494 .may_writepage = !laptop_mode, 3495 .may_unmap = 1, 3496 .may_swap = 1, 3497 }; 3498 3499 __fs_reclaim_acquire(); 3500 3501 count_vm_event(PAGEOUTRUN); 3502 3503 do { 3504 unsigned long nr_reclaimed = sc.nr_reclaimed; 3505 bool raise_priority = true; 3506 bool ret; 3507 3508 sc.reclaim_idx = classzone_idx; 3509 3510 /* 3511 * If the number of buffer_heads exceeds the maximum allowed 3512 * then consider reclaiming from all zones. This has a dual 3513 * purpose -- on 64-bit systems it is expected that 3514 * buffer_heads are stripped during active rotation. On 32-bit 3515 * systems, highmem pages can pin lowmem memory and shrinking 3516 * buffers can relieve lowmem pressure. Reclaim may still not 3517 * go ahead if all eligible zones for the original allocation 3518 * request are balanced to avoid excessive reclaim from kswapd. 3519 */ 3520 if (buffer_heads_over_limit) { 3521 for (i = MAX_NR_ZONES - 1; i >= 0; i--) { 3522 zone = pgdat->node_zones + i; 3523 if (!managed_zone(zone)) 3524 continue; 3525 3526 sc.reclaim_idx = i; 3527 break; 3528 } 3529 } 3530 3531 /* 3532 * Only reclaim if there are no eligible zones. Note that 3533 * sc.reclaim_idx is not used as buffer_heads_over_limit may 3534 * have adjusted it. 3535 */ 3536 if (pgdat_balanced(pgdat, sc.order, classzone_idx)) 3537 goto out; 3538 3539 /* 3540 * Do some background aging of the anon list, to give 3541 * pages a chance to be referenced before reclaiming. All 3542 * pages are rotated regardless of classzone as this is 3543 * about consistent aging. 3544 */ 3545 age_active_anon(pgdat, &sc); 3546 3547 /* 3548 * If we're getting trouble reclaiming, start doing writepage 3549 * even in laptop mode. 3550 */ 3551 if (sc.priority < DEF_PRIORITY - 2) 3552 sc.may_writepage = 1; 3553 3554 /* Call soft limit reclaim before calling shrink_node. */ 3555 sc.nr_scanned = 0; 3556 nr_soft_scanned = 0; 3557 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, 3558 sc.gfp_mask, &nr_soft_scanned); 3559 sc.nr_reclaimed += nr_soft_reclaimed; 3560 3561 /* 3562 * There should be no need to raise the scanning priority if 3563 * enough pages are already being scanned that that high 3564 * watermark would be met at 100% efficiency. 3565 */ 3566 if (kswapd_shrink_node(pgdat, &sc)) 3567 raise_priority = false; 3568 3569 /* 3570 * If the low watermark is met there is no need for processes 3571 * to be throttled on pfmemalloc_wait as they should not be 3572 * able to safely make forward progress. Wake them 3573 */ 3574 if (waitqueue_active(&pgdat->pfmemalloc_wait) && 3575 allow_direct_reclaim(pgdat)) 3576 wake_up_all(&pgdat->pfmemalloc_wait); 3577 3578 /* Check if kswapd should be suspending */ 3579 __fs_reclaim_release(); 3580 ret = try_to_freeze(); 3581 __fs_reclaim_acquire(); 3582 if (ret || kthread_should_stop()) 3583 break; 3584 3585 /* 3586 * Raise priority if scanning rate is too low or there was no 3587 * progress in reclaiming pages 3588 */ 3589 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; 3590 if (raise_priority || !nr_reclaimed) 3591 sc.priority--; 3592 } while (sc.priority >= 1); 3593 3594 if (!sc.nr_reclaimed) 3595 pgdat->kswapd_failures++; 3596 3597 out: 3598 snapshot_refaults(NULL, pgdat); 3599 __fs_reclaim_release(); 3600 /* 3601 * Return the order kswapd stopped reclaiming at as 3602 * prepare_kswapd_sleep() takes it into account. If another caller 3603 * entered the allocator slow path while kswapd was awake, order will 3604 * remain at the higher level. 3605 */ 3606 return sc.order; 3607 } 3608 3609 /* 3610 * pgdat->kswapd_classzone_idx is the highest zone index that a recent 3611 * allocation request woke kswapd for. When kswapd has not woken recently, 3612 * the value is MAX_NR_ZONES which is not a valid index. This compares a 3613 * given classzone and returns it or the highest classzone index kswapd 3614 * was recently woke for. 3615 */ 3616 static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat, 3617 enum zone_type classzone_idx) 3618 { 3619 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) 3620 return classzone_idx; 3621 3622 return max(pgdat->kswapd_classzone_idx, classzone_idx); 3623 } 3624 3625 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, 3626 unsigned int classzone_idx) 3627 { 3628 long remaining = 0; 3629 DEFINE_WAIT(wait); 3630 3631 if (freezing(current) || kthread_should_stop()) 3632 return; 3633 3634 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 3635 3636 /* 3637 * Try to sleep for a short interval. Note that kcompactd will only be 3638 * woken if it is possible to sleep for a short interval. This is 3639 * deliberate on the assumption that if reclaim cannot keep an 3640 * eligible zone balanced that it's also unlikely that compaction will 3641 * succeed. 3642 */ 3643 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { 3644 /* 3645 * Compaction records what page blocks it recently failed to 3646 * isolate pages from and skips them in the future scanning. 3647 * When kswapd is going to sleep, it is reasonable to assume 3648 * that pages and compaction may succeed so reset the cache. 3649 */ 3650 reset_isolation_suitable(pgdat); 3651 3652 /* 3653 * We have freed the memory, now we should compact it to make 3654 * allocation of the requested order possible. 3655 */ 3656 wakeup_kcompactd(pgdat, alloc_order, classzone_idx); 3657 3658 remaining = schedule_timeout(HZ/10); 3659 3660 /* 3661 * If woken prematurely then reset kswapd_classzone_idx and 3662 * order. The values will either be from a wakeup request or 3663 * the previous request that slept prematurely. 3664 */ 3665 if (remaining) { 3666 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); 3667 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); 3668 } 3669 3670 finish_wait(&pgdat->kswapd_wait, &wait); 3671 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 3672 } 3673 3674 /* 3675 * After a short sleep, check if it was a premature sleep. If not, then 3676 * go fully to sleep until explicitly woken up. 3677 */ 3678 if (!remaining && 3679 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { 3680 trace_mm_vmscan_kswapd_sleep(pgdat->node_id); 3681 3682 /* 3683 * vmstat counters are not perfectly accurate and the estimated 3684 * value for counters such as NR_FREE_PAGES can deviate from the 3685 * true value by nr_online_cpus * threshold. To avoid the zone 3686 * watermarks being breached while under pressure, we reduce the 3687 * per-cpu vmstat threshold while kswapd is awake and restore 3688 * them before going back to sleep. 3689 */ 3690 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); 3691 3692 if (!kthread_should_stop()) 3693 schedule(); 3694 3695 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); 3696 } else { 3697 if (remaining) 3698 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); 3699 else 3700 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); 3701 } 3702 finish_wait(&pgdat->kswapd_wait, &wait); 3703 } 3704 3705 /* 3706 * The background pageout daemon, started as a kernel thread 3707 * from the init process. 3708 * 3709 * This basically trickles out pages so that we have _some_ 3710 * free memory available even if there is no other activity 3711 * that frees anything up. This is needed for things like routing 3712 * etc, where we otherwise might have all activity going on in 3713 * asynchronous contexts that cannot page things out. 3714 * 3715 * If there are applications that are active memory-allocators 3716 * (most normal use), this basically shouldn't matter. 3717 */ 3718 static int kswapd(void *p) 3719 { 3720 unsigned int alloc_order, reclaim_order; 3721 unsigned int classzone_idx = MAX_NR_ZONES - 1; 3722 pg_data_t *pgdat = (pg_data_t*)p; 3723 struct task_struct *tsk = current; 3724 3725 struct reclaim_state reclaim_state = { 3726 .reclaimed_slab = 0, 3727 }; 3728 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); 3729 3730 if (!cpumask_empty(cpumask)) 3731 set_cpus_allowed_ptr(tsk, cpumask); 3732 current->reclaim_state = &reclaim_state; 3733 3734 /* 3735 * Tell the memory management that we're a "memory allocator", 3736 * and that if we need more memory we should get access to it 3737 * regardless (see "__alloc_pages()"). "kswapd" should 3738 * never get caught in the normal page freeing logic. 3739 * 3740 * (Kswapd normally doesn't need memory anyway, but sometimes 3741 * you need a small amount of memory in order to be able to 3742 * page out something else, and this flag essentially protects 3743 * us from recursively trying to free more memory as we're 3744 * trying to free the first piece of memory in the first place). 3745 */ 3746 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; 3747 set_freezable(); 3748 3749 pgdat->kswapd_order = 0; 3750 pgdat->kswapd_classzone_idx = MAX_NR_ZONES; 3751 for ( ; ; ) { 3752 bool ret; 3753 3754 alloc_order = reclaim_order = pgdat->kswapd_order; 3755 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); 3756 3757 kswapd_try_sleep: 3758 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, 3759 classzone_idx); 3760 3761 /* Read the new order and classzone_idx */ 3762 alloc_order = reclaim_order = pgdat->kswapd_order; 3763 classzone_idx = kswapd_classzone_idx(pgdat, 0); 3764 pgdat->kswapd_order = 0; 3765 pgdat->kswapd_classzone_idx = MAX_NR_ZONES; 3766 3767 ret = try_to_freeze(); 3768 if (kthread_should_stop()) 3769 break; 3770 3771 /* 3772 * We can speed up thawing tasks if we don't call balance_pgdat 3773 * after returning from the refrigerator 3774 */ 3775 if (ret) 3776 continue; 3777 3778 /* 3779 * Reclaim begins at the requested order but if a high-order 3780 * reclaim fails then kswapd falls back to reclaiming for 3781 * order-0. If that happens, kswapd will consider sleeping 3782 * for the order it finished reclaiming at (reclaim_order) 3783 * but kcompactd is woken to compact for the original 3784 * request (alloc_order). 3785 */ 3786 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, 3787 alloc_order); 3788 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); 3789 if (reclaim_order < alloc_order) 3790 goto kswapd_try_sleep; 3791 } 3792 3793 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); 3794 current->reclaim_state = NULL; 3795 3796 return 0; 3797 } 3798 3799 /* 3800 * A zone is low on free memory or too fragmented for high-order memory. If 3801 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's 3802 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim 3803 * has failed or is not needed, still wake up kcompactd if only compaction is 3804 * needed. 3805 */ 3806 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order, 3807 enum zone_type classzone_idx) 3808 { 3809 pg_data_t *pgdat; 3810 3811 if (!managed_zone(zone)) 3812 return; 3813 3814 if (!cpuset_zone_allowed(zone, gfp_flags)) 3815 return; 3816 pgdat = zone->zone_pgdat; 3817 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, 3818 classzone_idx); 3819 pgdat->kswapd_order = max(pgdat->kswapd_order, order); 3820 if (!waitqueue_active(&pgdat->kswapd_wait)) 3821 return; 3822 3823 /* Hopeless node, leave it to direct reclaim if possible */ 3824 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES || 3825 pgdat_balanced(pgdat, order, classzone_idx)) { 3826 /* 3827 * There may be plenty of free memory available, but it's too 3828 * fragmented for high-order allocations. Wake up kcompactd 3829 * and rely on compaction_suitable() to determine if it's 3830 * needed. If it fails, it will defer subsequent attempts to 3831 * ratelimit its work. 3832 */ 3833 if (!(gfp_flags & __GFP_DIRECT_RECLAIM)) 3834 wakeup_kcompactd(pgdat, order, classzone_idx); 3835 return; 3836 } 3837 3838 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order, 3839 gfp_flags); 3840 wake_up_interruptible(&pgdat->kswapd_wait); 3841 } 3842 3843 #ifdef CONFIG_HIBERNATION 3844 /* 3845 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of 3846 * freed pages. 3847 * 3848 * Rather than trying to age LRUs the aim is to preserve the overall 3849 * LRU order by reclaiming preferentially 3850 * inactive > active > active referenced > active mapped 3851 */ 3852 unsigned long shrink_all_memory(unsigned long nr_to_reclaim) 3853 { 3854 struct reclaim_state reclaim_state; 3855 struct scan_control sc = { 3856 .nr_to_reclaim = nr_to_reclaim, 3857 .gfp_mask = GFP_HIGHUSER_MOVABLE, 3858 .reclaim_idx = MAX_NR_ZONES - 1, 3859 .priority = DEF_PRIORITY, 3860 .may_writepage = 1, 3861 .may_unmap = 1, 3862 .may_swap = 1, 3863 .hibernation_mode = 1, 3864 }; 3865 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); 3866 struct task_struct *p = current; 3867 unsigned long nr_reclaimed; 3868 unsigned int noreclaim_flag; 3869 3870 fs_reclaim_acquire(sc.gfp_mask); 3871 noreclaim_flag = memalloc_noreclaim_save(); 3872 reclaim_state.reclaimed_slab = 0; 3873 p->reclaim_state = &reclaim_state; 3874 3875 nr_reclaimed = do_try_to_free_pages(zonelist, &sc); 3876 3877 p->reclaim_state = NULL; 3878 memalloc_noreclaim_restore(noreclaim_flag); 3879 fs_reclaim_release(sc.gfp_mask); 3880 3881 return nr_reclaimed; 3882 } 3883 #endif /* CONFIG_HIBERNATION */ 3884 3885 /* It's optimal to keep kswapds on the same CPUs as their memory, but 3886 not required for correctness. So if the last cpu in a node goes 3887 away, we get changed to run anywhere: as the first one comes back, 3888 restore their cpu bindings. */ 3889 static int kswapd_cpu_online(unsigned int cpu) 3890 { 3891 int nid; 3892 3893 for_each_node_state(nid, N_MEMORY) { 3894 pg_data_t *pgdat = NODE_DATA(nid); 3895 const struct cpumask *mask; 3896 3897 mask = cpumask_of_node(pgdat->node_id); 3898 3899 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) 3900 /* One of our CPUs online: restore mask */ 3901 set_cpus_allowed_ptr(pgdat->kswapd, mask); 3902 } 3903 return 0; 3904 } 3905 3906 /* 3907 * This kswapd start function will be called by init and node-hot-add. 3908 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. 3909 */ 3910 int kswapd_run(int nid) 3911 { 3912 pg_data_t *pgdat = NODE_DATA(nid); 3913 int ret = 0; 3914 3915 if (pgdat->kswapd) 3916 return 0; 3917 3918 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); 3919 if (IS_ERR(pgdat->kswapd)) { 3920 /* failure at boot is fatal */ 3921 BUG_ON(system_state < SYSTEM_RUNNING); 3922 pr_err("Failed to start kswapd on node %d\n", nid); 3923 ret = PTR_ERR(pgdat->kswapd); 3924 pgdat->kswapd = NULL; 3925 } 3926 return ret; 3927 } 3928 3929 /* 3930 * Called by memory hotplug when all memory in a node is offlined. Caller must 3931 * hold mem_hotplug_begin/end(). 3932 */ 3933 void kswapd_stop(int nid) 3934 { 3935 struct task_struct *kswapd = NODE_DATA(nid)->kswapd; 3936 3937 if (kswapd) { 3938 kthread_stop(kswapd); 3939 NODE_DATA(nid)->kswapd = NULL; 3940 } 3941 } 3942 3943 static int __init kswapd_init(void) 3944 { 3945 int nid, ret; 3946 3947 swap_setup(); 3948 for_each_node_state(nid, N_MEMORY) 3949 kswapd_run(nid); 3950 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, 3951 "mm/vmscan:online", kswapd_cpu_online, 3952 NULL); 3953 WARN_ON(ret < 0); 3954 return 0; 3955 } 3956 3957 module_init(kswapd_init) 3958 3959 #ifdef CONFIG_NUMA 3960 /* 3961 * Node reclaim mode 3962 * 3963 * If non-zero call node_reclaim when the number of free pages falls below 3964 * the watermarks. 3965 */ 3966 int node_reclaim_mode __read_mostly; 3967 3968 #define RECLAIM_OFF 0 3969 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ 3970 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ 3971 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ 3972 3973 /* 3974 * Priority for NODE_RECLAIM. This determines the fraction of pages 3975 * of a node considered for each zone_reclaim. 4 scans 1/16th of 3976 * a zone. 3977 */ 3978 #define NODE_RECLAIM_PRIORITY 4 3979 3980 /* 3981 * Percentage of pages in a zone that must be unmapped for node_reclaim to 3982 * occur. 3983 */ 3984 int sysctl_min_unmapped_ratio = 1; 3985 3986 /* 3987 * If the number of slab pages in a zone grows beyond this percentage then 3988 * slab reclaim needs to occur. 3989 */ 3990 int sysctl_min_slab_ratio = 5; 3991 3992 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) 3993 { 3994 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); 3995 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + 3996 node_page_state(pgdat, NR_ACTIVE_FILE); 3997 3998 /* 3999 * It's possible for there to be more file mapped pages than 4000 * accounted for by the pages on the file LRU lists because 4001 * tmpfs pages accounted for as ANON can also be FILE_MAPPED 4002 */ 4003 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; 4004 } 4005 4006 /* Work out how many page cache pages we can reclaim in this reclaim_mode */ 4007 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) 4008 { 4009 unsigned long nr_pagecache_reclaimable; 4010 unsigned long delta = 0; 4011 4012 /* 4013 * If RECLAIM_UNMAP is set, then all file pages are considered 4014 * potentially reclaimable. Otherwise, we have to worry about 4015 * pages like swapcache and node_unmapped_file_pages() provides 4016 * a better estimate 4017 */ 4018 if (node_reclaim_mode & RECLAIM_UNMAP) 4019 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); 4020 else 4021 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); 4022 4023 /* If we can't clean pages, remove dirty pages from consideration */ 4024 if (!(node_reclaim_mode & RECLAIM_WRITE)) 4025 delta += node_page_state(pgdat, NR_FILE_DIRTY); 4026 4027 /* Watch for any possible underflows due to delta */ 4028 if (unlikely(delta > nr_pagecache_reclaimable)) 4029 delta = nr_pagecache_reclaimable; 4030 4031 return nr_pagecache_reclaimable - delta; 4032 } 4033 4034 /* 4035 * Try to free up some pages from this node through reclaim. 4036 */ 4037 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) 4038 { 4039 /* Minimum pages needed in order to stay on node */ 4040 const unsigned long nr_pages = 1 << order; 4041 struct task_struct *p = current; 4042 struct reclaim_state reclaim_state; 4043 unsigned int noreclaim_flag; 4044 struct scan_control sc = { 4045 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), 4046 .gfp_mask = current_gfp_context(gfp_mask), 4047 .order = order, 4048 .priority = NODE_RECLAIM_PRIORITY, 4049 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), 4050 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), 4051 .may_swap = 1, 4052 .reclaim_idx = gfp_zone(gfp_mask), 4053 }; 4054 4055 cond_resched(); 4056 fs_reclaim_acquire(sc.gfp_mask); 4057 /* 4058 * We need to be able to allocate from the reserves for RECLAIM_UNMAP 4059 * and we also need to be able to write out pages for RECLAIM_WRITE 4060 * and RECLAIM_UNMAP. 4061 */ 4062 noreclaim_flag = memalloc_noreclaim_save(); 4063 p->flags |= PF_SWAPWRITE; 4064 reclaim_state.reclaimed_slab = 0; 4065 p->reclaim_state = &reclaim_state; 4066 4067 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { 4068 /* 4069 * Free memory by calling shrink node with increasing 4070 * priorities until we have enough memory freed. 4071 */ 4072 do { 4073 shrink_node(pgdat, &sc); 4074 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); 4075 } 4076 4077 p->reclaim_state = NULL; 4078 current->flags &= ~PF_SWAPWRITE; 4079 memalloc_noreclaim_restore(noreclaim_flag); 4080 fs_reclaim_release(sc.gfp_mask); 4081 return sc.nr_reclaimed >= nr_pages; 4082 } 4083 4084 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) 4085 { 4086 int ret; 4087 4088 /* 4089 * Node reclaim reclaims unmapped file backed pages and 4090 * slab pages if we are over the defined limits. 4091 * 4092 * A small portion of unmapped file backed pages is needed for 4093 * file I/O otherwise pages read by file I/O will be immediately 4094 * thrown out if the node is overallocated. So we do not reclaim 4095 * if less than a specified percentage of the node is used by 4096 * unmapped file backed pages. 4097 */ 4098 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && 4099 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) 4100 return NODE_RECLAIM_FULL; 4101 4102 /* 4103 * Do not scan if the allocation should not be delayed. 4104 */ 4105 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) 4106 return NODE_RECLAIM_NOSCAN; 4107 4108 /* 4109 * Only run node reclaim on the local node or on nodes that do not 4110 * have associated processors. This will favor the local processor 4111 * over remote processors and spread off node memory allocations 4112 * as wide as possible. 4113 */ 4114 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) 4115 return NODE_RECLAIM_NOSCAN; 4116 4117 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) 4118 return NODE_RECLAIM_NOSCAN; 4119 4120 ret = __node_reclaim(pgdat, gfp_mask, order); 4121 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); 4122 4123 if (!ret) 4124 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); 4125 4126 return ret; 4127 } 4128 #endif 4129 4130 /* 4131 * page_evictable - test whether a page is evictable 4132 * @page: the page to test 4133 * 4134 * Test whether page is evictable--i.e., should be placed on active/inactive 4135 * lists vs unevictable list. 4136 * 4137 * Reasons page might not be evictable: 4138 * (1) page's mapping marked unevictable 4139 * (2) page is part of an mlocked VMA 4140 * 4141 */ 4142 int page_evictable(struct page *page) 4143 { 4144 int ret; 4145 4146 /* Prevent address_space of inode and swap cache from being freed */ 4147 rcu_read_lock(); 4148 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); 4149 rcu_read_unlock(); 4150 return ret; 4151 } 4152 4153 #ifdef CONFIG_SHMEM 4154 /** 4155 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list 4156 * @pages: array of pages to check 4157 * @nr_pages: number of pages to check 4158 * 4159 * Checks pages for evictability and moves them to the appropriate lru list. 4160 * 4161 * This function is only used for SysV IPC SHM_UNLOCK. 4162 */ 4163 void check_move_unevictable_pages(struct page **pages, int nr_pages) 4164 { 4165 struct lruvec *lruvec; 4166 struct pglist_data *pgdat = NULL; 4167 int pgscanned = 0; 4168 int pgrescued = 0; 4169 int i; 4170 4171 for (i = 0; i < nr_pages; i++) { 4172 struct page *page = pages[i]; 4173 struct pglist_data *pagepgdat = page_pgdat(page); 4174 4175 pgscanned++; 4176 if (pagepgdat != pgdat) { 4177 if (pgdat) 4178 spin_unlock_irq(&pgdat->lru_lock); 4179 pgdat = pagepgdat; 4180 spin_lock_irq(&pgdat->lru_lock); 4181 } 4182 lruvec = mem_cgroup_page_lruvec(page, pgdat); 4183 4184 if (!PageLRU(page) || !PageUnevictable(page)) 4185 continue; 4186 4187 if (page_evictable(page)) { 4188 enum lru_list lru = page_lru_base_type(page); 4189 4190 VM_BUG_ON_PAGE(PageActive(page), page); 4191 ClearPageUnevictable(page); 4192 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); 4193 add_page_to_lru_list(page, lruvec, lru); 4194 pgrescued++; 4195 } 4196 } 4197 4198 if (pgdat) { 4199 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); 4200 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); 4201 spin_unlock_irq(&pgdat->lru_lock); 4202 } 4203 } 4204 #endif /* CONFIG_SHMEM */ 4205