vmscan.c revision cf40bd16fdad42c053040bcd3988f5fdedbb6c57
1/* 2 * linux/mm/vmscan.c 3 * 4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 5 * 6 * Swap reorganised 29.12.95, Stephen Tweedie. 7 * kswapd added: 7.1.96 sct 8 * Removed kswapd_ctl limits, and swap out as many pages as needed 9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel. 10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). 11 * Multiqueue VM started 5.8.00, Rik van Riel. 12 */ 13 14#include <linux/mm.h> 15#include <linux/module.h> 16#include <linux/slab.h> 17#include <linux/kernel_stat.h> 18#include <linux/swap.h> 19#include <linux/pagemap.h> 20#include <linux/init.h> 21#include <linux/highmem.h> 22#include <linux/vmstat.h> 23#include <linux/file.h> 24#include <linux/writeback.h> 25#include <linux/blkdev.h> 26#include <linux/buffer_head.h> /* for try_to_release_page(), 27 buffer_heads_over_limit */ 28#include <linux/mm_inline.h> 29#include <linux/pagevec.h> 30#include <linux/backing-dev.h> 31#include <linux/rmap.h> 32#include <linux/topology.h> 33#include <linux/cpu.h> 34#include <linux/cpuset.h> 35#include <linux/notifier.h> 36#include <linux/rwsem.h> 37#include <linux/delay.h> 38#include <linux/kthread.h> 39#include <linux/freezer.h> 40#include <linux/memcontrol.h> 41#include <linux/delayacct.h> 42#include <linux/sysctl.h> 43 44#include <asm/tlbflush.h> 45#include <asm/div64.h> 46 47#include <linux/swapops.h> 48 49#include "internal.h" 50 51struct scan_control { 52 /* Incremented by the number of inactive pages that were scanned */ 53 unsigned long nr_scanned; 54 55 /* Number of pages freed so far during a call to shrink_zones() */ 56 unsigned long nr_reclaimed; 57 58 /* This context's GFP mask */ 59 gfp_t gfp_mask; 60 61 int may_writepage; 62 63 /* Can pages be swapped as part of reclaim? */ 64 int may_swap; 65 66 /* This context's SWAP_CLUSTER_MAX. If freeing memory for 67 * suspend, we effectively ignore SWAP_CLUSTER_MAX. 68 * In this context, it doesn't matter that we scan the 69 * whole list at once. */ 70 int swap_cluster_max; 71 72 int swappiness; 73 74 int all_unreclaimable; 75 76 int order; 77 78 /* Which cgroup do we reclaim from */ 79 struct mem_cgroup *mem_cgroup; 80 81 /* Pluggable isolate pages callback */ 82 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst, 83 unsigned long *scanned, int order, int mode, 84 struct zone *z, struct mem_cgroup *mem_cont, 85 int active, int file); 86}; 87 88#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) 89 90#ifdef ARCH_HAS_PREFETCH 91#define prefetch_prev_lru_page(_page, _base, _field) \ 92 do { \ 93 if ((_page)->lru.prev != _base) { \ 94 struct page *prev; \ 95 \ 96 prev = lru_to_page(&(_page->lru)); \ 97 prefetch(&prev->_field); \ 98 } \ 99 } while (0) 100#else 101#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) 102#endif 103 104#ifdef ARCH_HAS_PREFETCHW 105#define prefetchw_prev_lru_page(_page, _base, _field) \ 106 do { \ 107 if ((_page)->lru.prev != _base) { \ 108 struct page *prev; \ 109 \ 110 prev = lru_to_page(&(_page->lru)); \ 111 prefetchw(&prev->_field); \ 112 } \ 113 } while (0) 114#else 115#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) 116#endif 117 118/* 119 * From 0 .. 100. Higher means more swappy. 120 */ 121int vm_swappiness = 60; 122long vm_total_pages; /* The total number of pages which the VM controls */ 123 124static LIST_HEAD(shrinker_list); 125static DECLARE_RWSEM(shrinker_rwsem); 126 127#ifdef CONFIG_CGROUP_MEM_RES_CTLR 128#define scanning_global_lru(sc) (!(sc)->mem_cgroup) 129#else 130#define scanning_global_lru(sc) (1) 131#endif 132 133static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone, 134 struct scan_control *sc) 135{ 136 if (!scanning_global_lru(sc)) 137 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone); 138 139 return &zone->reclaim_stat; 140} 141 142static unsigned long zone_nr_pages(struct zone *zone, struct scan_control *sc, 143 enum lru_list lru) 144{ 145 if (!scanning_global_lru(sc)) 146 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru); 147 148 return zone_page_state(zone, NR_LRU_BASE + lru); 149} 150 151 152/* 153 * Add a shrinker callback to be called from the vm 154 */ 155void register_shrinker(struct shrinker *shrinker) 156{ 157 shrinker->nr = 0; 158 down_write(&shrinker_rwsem); 159 list_add_tail(&shrinker->list, &shrinker_list); 160 up_write(&shrinker_rwsem); 161} 162EXPORT_SYMBOL(register_shrinker); 163 164/* 165 * Remove one 166 */ 167void unregister_shrinker(struct shrinker *shrinker) 168{ 169 down_write(&shrinker_rwsem); 170 list_del(&shrinker->list); 171 up_write(&shrinker_rwsem); 172} 173EXPORT_SYMBOL(unregister_shrinker); 174 175#define SHRINK_BATCH 128 176/* 177 * Call the shrink functions to age shrinkable caches 178 * 179 * Here we assume it costs one seek to replace a lru page and that it also 180 * takes a seek to recreate a cache object. With this in mind we age equal 181 * percentages of the lru and ageable caches. This should balance the seeks 182 * generated by these structures. 183 * 184 * If the vm encountered mapped pages on the LRU it increase the pressure on 185 * slab to avoid swapping. 186 * 187 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. 188 * 189 * `lru_pages' represents the number of on-LRU pages in all the zones which 190 * are eligible for the caller's allocation attempt. It is used for balancing 191 * slab reclaim versus page reclaim. 192 * 193 * Returns the number of slab objects which we shrunk. 194 */ 195unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask, 196 unsigned long lru_pages) 197{ 198 struct shrinker *shrinker; 199 unsigned long ret = 0; 200 201 if (scanned == 0) 202 scanned = SWAP_CLUSTER_MAX; 203 204 if (!down_read_trylock(&shrinker_rwsem)) 205 return 1; /* Assume we'll be able to shrink next time */ 206 207 list_for_each_entry(shrinker, &shrinker_list, list) { 208 unsigned long long delta; 209 unsigned long total_scan; 210 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask); 211 212 delta = (4 * scanned) / shrinker->seeks; 213 delta *= max_pass; 214 do_div(delta, lru_pages + 1); 215 shrinker->nr += delta; 216 if (shrinker->nr < 0) { 217 printk(KERN_ERR "%s: nr=%ld\n", 218 __func__, shrinker->nr); 219 shrinker->nr = max_pass; 220 } 221 222 /* 223 * Avoid risking looping forever due to too large nr value: 224 * never try to free more than twice the estimate number of 225 * freeable entries. 226 */ 227 if (shrinker->nr > max_pass * 2) 228 shrinker->nr = max_pass * 2; 229 230 total_scan = shrinker->nr; 231 shrinker->nr = 0; 232 233 while (total_scan >= SHRINK_BATCH) { 234 long this_scan = SHRINK_BATCH; 235 int shrink_ret; 236 int nr_before; 237 238 nr_before = (*shrinker->shrink)(0, gfp_mask); 239 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask); 240 if (shrink_ret == -1) 241 break; 242 if (shrink_ret < nr_before) 243 ret += nr_before - shrink_ret; 244 count_vm_events(SLABS_SCANNED, this_scan); 245 total_scan -= this_scan; 246 247 cond_resched(); 248 } 249 250 shrinker->nr += total_scan; 251 } 252 up_read(&shrinker_rwsem); 253 return ret; 254} 255 256/* Called without lock on whether page is mapped, so answer is unstable */ 257static inline int page_mapping_inuse(struct page *page) 258{ 259 struct address_space *mapping; 260 261 /* Page is in somebody's page tables. */ 262 if (page_mapped(page)) 263 return 1; 264 265 /* Be more reluctant to reclaim swapcache than pagecache */ 266 if (PageSwapCache(page)) 267 return 1; 268 269 mapping = page_mapping(page); 270 if (!mapping) 271 return 0; 272 273 /* File is mmap'd by somebody? */ 274 return mapping_mapped(mapping); 275} 276 277static inline int is_page_cache_freeable(struct page *page) 278{ 279 return page_count(page) - !!PagePrivate(page) == 2; 280} 281 282static int may_write_to_queue(struct backing_dev_info *bdi) 283{ 284 if (current->flags & PF_SWAPWRITE) 285 return 1; 286 if (!bdi_write_congested(bdi)) 287 return 1; 288 if (bdi == current->backing_dev_info) 289 return 1; 290 return 0; 291} 292 293/* 294 * We detected a synchronous write error writing a page out. Probably 295 * -ENOSPC. We need to propagate that into the address_space for a subsequent 296 * fsync(), msync() or close(). 297 * 298 * The tricky part is that after writepage we cannot touch the mapping: nothing 299 * prevents it from being freed up. But we have a ref on the page and once 300 * that page is locked, the mapping is pinned. 301 * 302 * We're allowed to run sleeping lock_page() here because we know the caller has 303 * __GFP_FS. 304 */ 305static void handle_write_error(struct address_space *mapping, 306 struct page *page, int error) 307{ 308 lock_page(page); 309 if (page_mapping(page) == mapping) 310 mapping_set_error(mapping, error); 311 unlock_page(page); 312} 313 314/* Request for sync pageout. */ 315enum pageout_io { 316 PAGEOUT_IO_ASYNC, 317 PAGEOUT_IO_SYNC, 318}; 319 320/* possible outcome of pageout() */ 321typedef enum { 322 /* failed to write page out, page is locked */ 323 PAGE_KEEP, 324 /* move page to the active list, page is locked */ 325 PAGE_ACTIVATE, 326 /* page has been sent to the disk successfully, page is unlocked */ 327 PAGE_SUCCESS, 328 /* page is clean and locked */ 329 PAGE_CLEAN, 330} pageout_t; 331 332/* 333 * pageout is called by shrink_page_list() for each dirty page. 334 * Calls ->writepage(). 335 */ 336static pageout_t pageout(struct page *page, struct address_space *mapping, 337 enum pageout_io sync_writeback) 338{ 339 /* 340 * If the page is dirty, only perform writeback if that write 341 * will be non-blocking. To prevent this allocation from being 342 * stalled by pagecache activity. But note that there may be 343 * stalls if we need to run get_block(). We could test 344 * PagePrivate for that. 345 * 346 * If this process is currently in generic_file_write() against 347 * this page's queue, we can perform writeback even if that 348 * will block. 349 * 350 * If the page is swapcache, write it back even if that would 351 * block, for some throttling. This happens by accident, because 352 * swap_backing_dev_info is bust: it doesn't reflect the 353 * congestion state of the swapdevs. Easy to fix, if needed. 354 * See swapfile.c:page_queue_congested(). 355 */ 356 if (!is_page_cache_freeable(page)) 357 return PAGE_KEEP; 358 if (!mapping) { 359 /* 360 * Some data journaling orphaned pages can have 361 * page->mapping == NULL while being dirty with clean buffers. 362 */ 363 if (PagePrivate(page)) { 364 if (try_to_free_buffers(page)) { 365 ClearPageDirty(page); 366 printk("%s: orphaned page\n", __func__); 367 return PAGE_CLEAN; 368 } 369 } 370 return PAGE_KEEP; 371 } 372 if (mapping->a_ops->writepage == NULL) 373 return PAGE_ACTIVATE; 374 if (!may_write_to_queue(mapping->backing_dev_info)) 375 return PAGE_KEEP; 376 377 if (clear_page_dirty_for_io(page)) { 378 int res; 379 struct writeback_control wbc = { 380 .sync_mode = WB_SYNC_NONE, 381 .nr_to_write = SWAP_CLUSTER_MAX, 382 .range_start = 0, 383 .range_end = LLONG_MAX, 384 .nonblocking = 1, 385 .for_reclaim = 1, 386 }; 387 388 SetPageReclaim(page); 389 res = mapping->a_ops->writepage(page, &wbc); 390 if (res < 0) 391 handle_write_error(mapping, page, res); 392 if (res == AOP_WRITEPAGE_ACTIVATE) { 393 ClearPageReclaim(page); 394 return PAGE_ACTIVATE; 395 } 396 397 /* 398 * Wait on writeback if requested to. This happens when 399 * direct reclaiming a large contiguous area and the 400 * first attempt to free a range of pages fails. 401 */ 402 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC) 403 wait_on_page_writeback(page); 404 405 if (!PageWriteback(page)) { 406 /* synchronous write or broken a_ops? */ 407 ClearPageReclaim(page); 408 } 409 inc_zone_page_state(page, NR_VMSCAN_WRITE); 410 return PAGE_SUCCESS; 411 } 412 413 return PAGE_CLEAN; 414} 415 416/* 417 * Same as remove_mapping, but if the page is removed from the mapping, it 418 * gets returned with a refcount of 0. 419 */ 420static int __remove_mapping(struct address_space *mapping, struct page *page) 421{ 422 BUG_ON(!PageLocked(page)); 423 BUG_ON(mapping != page_mapping(page)); 424 425 spin_lock_irq(&mapping->tree_lock); 426 /* 427 * The non racy check for a busy page. 428 * 429 * Must be careful with the order of the tests. When someone has 430 * a ref to the page, it may be possible that they dirty it then 431 * drop the reference. So if PageDirty is tested before page_count 432 * here, then the following race may occur: 433 * 434 * get_user_pages(&page); 435 * [user mapping goes away] 436 * write_to(page); 437 * !PageDirty(page) [good] 438 * SetPageDirty(page); 439 * put_page(page); 440 * !page_count(page) [good, discard it] 441 * 442 * [oops, our write_to data is lost] 443 * 444 * Reversing the order of the tests ensures such a situation cannot 445 * escape unnoticed. The smp_rmb is needed to ensure the page->flags 446 * load is not satisfied before that of page->_count. 447 * 448 * Note that if SetPageDirty is always performed via set_page_dirty, 449 * and thus under tree_lock, then this ordering is not required. 450 */ 451 if (!page_freeze_refs(page, 2)) 452 goto cannot_free; 453 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ 454 if (unlikely(PageDirty(page))) { 455 page_unfreeze_refs(page, 2); 456 goto cannot_free; 457 } 458 459 if (PageSwapCache(page)) { 460 swp_entry_t swap = { .val = page_private(page) }; 461 __delete_from_swap_cache(page); 462 spin_unlock_irq(&mapping->tree_lock); 463 swap_free(swap); 464 } else { 465 __remove_from_page_cache(page); 466 spin_unlock_irq(&mapping->tree_lock); 467 } 468 469 return 1; 470 471cannot_free: 472 spin_unlock_irq(&mapping->tree_lock); 473 return 0; 474} 475 476/* 477 * Attempt to detach a locked page from its ->mapping. If it is dirty or if 478 * someone else has a ref on the page, abort and return 0. If it was 479 * successfully detached, return 1. Assumes the caller has a single ref on 480 * this page. 481 */ 482int remove_mapping(struct address_space *mapping, struct page *page) 483{ 484 if (__remove_mapping(mapping, page)) { 485 /* 486 * Unfreezing the refcount with 1 rather than 2 effectively 487 * drops the pagecache ref for us without requiring another 488 * atomic operation. 489 */ 490 page_unfreeze_refs(page, 1); 491 return 1; 492 } 493 return 0; 494} 495 496/** 497 * putback_lru_page - put previously isolated page onto appropriate LRU list 498 * @page: page to be put back to appropriate lru list 499 * 500 * Add previously isolated @page to appropriate LRU list. 501 * Page may still be unevictable for other reasons. 502 * 503 * lru_lock must not be held, interrupts must be enabled. 504 */ 505#ifdef CONFIG_UNEVICTABLE_LRU 506void putback_lru_page(struct page *page) 507{ 508 int lru; 509 int active = !!TestClearPageActive(page); 510 int was_unevictable = PageUnevictable(page); 511 512 VM_BUG_ON(PageLRU(page)); 513 514redo: 515 ClearPageUnevictable(page); 516 517 if (page_evictable(page, NULL)) { 518 /* 519 * For evictable pages, we can use the cache. 520 * In event of a race, worst case is we end up with an 521 * unevictable page on [in]active list. 522 * We know how to handle that. 523 */ 524 lru = active + page_is_file_cache(page); 525 lru_cache_add_lru(page, lru); 526 } else { 527 /* 528 * Put unevictable pages directly on zone's unevictable 529 * list. 530 */ 531 lru = LRU_UNEVICTABLE; 532 add_page_to_unevictable_list(page); 533 } 534 535 /* 536 * page's status can change while we move it among lru. If an evictable 537 * page is on unevictable list, it never be freed. To avoid that, 538 * check after we added it to the list, again. 539 */ 540 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) { 541 if (!isolate_lru_page(page)) { 542 put_page(page); 543 goto redo; 544 } 545 /* This means someone else dropped this page from LRU 546 * So, it will be freed or putback to LRU again. There is 547 * nothing to do here. 548 */ 549 } 550 551 if (was_unevictable && lru != LRU_UNEVICTABLE) 552 count_vm_event(UNEVICTABLE_PGRESCUED); 553 else if (!was_unevictable && lru == LRU_UNEVICTABLE) 554 count_vm_event(UNEVICTABLE_PGCULLED); 555 556 put_page(page); /* drop ref from isolate */ 557} 558 559#else /* CONFIG_UNEVICTABLE_LRU */ 560 561void putback_lru_page(struct page *page) 562{ 563 int lru; 564 VM_BUG_ON(PageLRU(page)); 565 566 lru = !!TestClearPageActive(page) + page_is_file_cache(page); 567 lru_cache_add_lru(page, lru); 568 put_page(page); 569} 570#endif /* CONFIG_UNEVICTABLE_LRU */ 571 572 573/* 574 * shrink_page_list() returns the number of reclaimed pages 575 */ 576static unsigned long shrink_page_list(struct list_head *page_list, 577 struct scan_control *sc, 578 enum pageout_io sync_writeback) 579{ 580 LIST_HEAD(ret_pages); 581 struct pagevec freed_pvec; 582 int pgactivate = 0; 583 unsigned long nr_reclaimed = 0; 584 585 cond_resched(); 586 587 pagevec_init(&freed_pvec, 1); 588 while (!list_empty(page_list)) { 589 struct address_space *mapping; 590 struct page *page; 591 int may_enter_fs; 592 int referenced; 593 594 cond_resched(); 595 596 page = lru_to_page(page_list); 597 list_del(&page->lru); 598 599 if (!trylock_page(page)) 600 goto keep; 601 602 VM_BUG_ON(PageActive(page)); 603 604 sc->nr_scanned++; 605 606 if (unlikely(!page_evictable(page, NULL))) 607 goto cull_mlocked; 608 609 if (!sc->may_swap && page_mapped(page)) 610 goto keep_locked; 611 612 /* Double the slab pressure for mapped and swapcache pages */ 613 if (page_mapped(page) || PageSwapCache(page)) 614 sc->nr_scanned++; 615 616 may_enter_fs = (sc->gfp_mask & __GFP_FS) || 617 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); 618 619 if (PageWriteback(page)) { 620 /* 621 * Synchronous reclaim is performed in two passes, 622 * first an asynchronous pass over the list to 623 * start parallel writeback, and a second synchronous 624 * pass to wait for the IO to complete. Wait here 625 * for any page for which writeback has already 626 * started. 627 */ 628 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs) 629 wait_on_page_writeback(page); 630 else 631 goto keep_locked; 632 } 633 634 referenced = page_referenced(page, 1, sc->mem_cgroup); 635 /* In active use or really unfreeable? Activate it. */ 636 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && 637 referenced && page_mapping_inuse(page)) 638 goto activate_locked; 639 640 /* 641 * Anonymous process memory has backing store? 642 * Try to allocate it some swap space here. 643 */ 644 if (PageAnon(page) && !PageSwapCache(page)) { 645 if (!(sc->gfp_mask & __GFP_IO)) 646 goto keep_locked; 647 if (!add_to_swap(page)) 648 goto activate_locked; 649 may_enter_fs = 1; 650 } 651 652 mapping = page_mapping(page); 653 654 /* 655 * The page is mapped into the page tables of one or more 656 * processes. Try to unmap it here. 657 */ 658 if (page_mapped(page) && mapping) { 659 switch (try_to_unmap(page, 0)) { 660 case SWAP_FAIL: 661 goto activate_locked; 662 case SWAP_AGAIN: 663 goto keep_locked; 664 case SWAP_MLOCK: 665 goto cull_mlocked; 666 case SWAP_SUCCESS: 667 ; /* try to free the page below */ 668 } 669 } 670 671 if (PageDirty(page)) { 672 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced) 673 goto keep_locked; 674 if (!may_enter_fs) 675 goto keep_locked; 676 if (!sc->may_writepage) 677 goto keep_locked; 678 679 /* Page is dirty, try to write it out here */ 680 switch (pageout(page, mapping, sync_writeback)) { 681 case PAGE_KEEP: 682 goto keep_locked; 683 case PAGE_ACTIVATE: 684 goto activate_locked; 685 case PAGE_SUCCESS: 686 if (PageWriteback(page) || PageDirty(page)) 687 goto keep; 688 /* 689 * A synchronous write - probably a ramdisk. Go 690 * ahead and try to reclaim the page. 691 */ 692 if (!trylock_page(page)) 693 goto keep; 694 if (PageDirty(page) || PageWriteback(page)) 695 goto keep_locked; 696 mapping = page_mapping(page); 697 case PAGE_CLEAN: 698 ; /* try to free the page below */ 699 } 700 } 701 702 /* 703 * If the page has buffers, try to free the buffer mappings 704 * associated with this page. If we succeed we try to free 705 * the page as well. 706 * 707 * We do this even if the page is PageDirty(). 708 * try_to_release_page() does not perform I/O, but it is 709 * possible for a page to have PageDirty set, but it is actually 710 * clean (all its buffers are clean). This happens if the 711 * buffers were written out directly, with submit_bh(). ext3 712 * will do this, as well as the blockdev mapping. 713 * try_to_release_page() will discover that cleanness and will 714 * drop the buffers and mark the page clean - it can be freed. 715 * 716 * Rarely, pages can have buffers and no ->mapping. These are 717 * the pages which were not successfully invalidated in 718 * truncate_complete_page(). We try to drop those buffers here 719 * and if that worked, and the page is no longer mapped into 720 * process address space (page_count == 1) it can be freed. 721 * Otherwise, leave the page on the LRU so it is swappable. 722 */ 723 if (PagePrivate(page)) { 724 if (!try_to_release_page(page, sc->gfp_mask)) 725 goto activate_locked; 726 if (!mapping && page_count(page) == 1) { 727 unlock_page(page); 728 if (put_page_testzero(page)) 729 goto free_it; 730 else { 731 /* 732 * rare race with speculative reference. 733 * the speculative reference will free 734 * this page shortly, so we may 735 * increment nr_reclaimed here (and 736 * leave it off the LRU). 737 */ 738 nr_reclaimed++; 739 continue; 740 } 741 } 742 } 743 744 if (!mapping || !__remove_mapping(mapping, page)) 745 goto keep_locked; 746 747 /* 748 * At this point, we have no other references and there is 749 * no way to pick any more up (removed from LRU, removed 750 * from pagecache). Can use non-atomic bitops now (and 751 * we obviously don't have to worry about waking up a process 752 * waiting on the page lock, because there are no references. 753 */ 754 __clear_page_locked(page); 755free_it: 756 nr_reclaimed++; 757 if (!pagevec_add(&freed_pvec, page)) { 758 __pagevec_free(&freed_pvec); 759 pagevec_reinit(&freed_pvec); 760 } 761 continue; 762 763cull_mlocked: 764 if (PageSwapCache(page)) 765 try_to_free_swap(page); 766 unlock_page(page); 767 putback_lru_page(page); 768 continue; 769 770activate_locked: 771 /* Not a candidate for swapping, so reclaim swap space. */ 772 if (PageSwapCache(page) && vm_swap_full()) 773 try_to_free_swap(page); 774 VM_BUG_ON(PageActive(page)); 775 SetPageActive(page); 776 pgactivate++; 777keep_locked: 778 unlock_page(page); 779keep: 780 list_add(&page->lru, &ret_pages); 781 VM_BUG_ON(PageLRU(page) || PageUnevictable(page)); 782 } 783 list_splice(&ret_pages, page_list); 784 if (pagevec_count(&freed_pvec)) 785 __pagevec_free(&freed_pvec); 786 count_vm_events(PGACTIVATE, pgactivate); 787 return nr_reclaimed; 788} 789 790/* LRU Isolation modes. */ 791#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */ 792#define ISOLATE_ACTIVE 1 /* Isolate active pages. */ 793#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */ 794 795/* 796 * Attempt to remove the specified page from its LRU. Only take this page 797 * if it is of the appropriate PageActive status. Pages which are being 798 * freed elsewhere are also ignored. 799 * 800 * page: page to consider 801 * mode: one of the LRU isolation modes defined above 802 * 803 * returns 0 on success, -ve errno on failure. 804 */ 805int __isolate_lru_page(struct page *page, int mode, int file) 806{ 807 int ret = -EINVAL; 808 809 /* Only take pages on the LRU. */ 810 if (!PageLRU(page)) 811 return ret; 812 813 /* 814 * When checking the active state, we need to be sure we are 815 * dealing with comparible boolean values. Take the logical not 816 * of each. 817 */ 818 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode)) 819 return ret; 820 821 if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file)) 822 return ret; 823 824 /* 825 * When this function is being called for lumpy reclaim, we 826 * initially look into all LRU pages, active, inactive and 827 * unevictable; only give shrink_page_list evictable pages. 828 */ 829 if (PageUnevictable(page)) 830 return ret; 831 832 ret = -EBUSY; 833 834 if (likely(get_page_unless_zero(page))) { 835 /* 836 * Be careful not to clear PageLRU until after we're 837 * sure the page is not being freed elsewhere -- the 838 * page release code relies on it. 839 */ 840 ClearPageLRU(page); 841 ret = 0; 842 mem_cgroup_del_lru(page); 843 } 844 845 return ret; 846} 847 848/* 849 * zone->lru_lock is heavily contended. Some of the functions that 850 * shrink the lists perform better by taking out a batch of pages 851 * and working on them outside the LRU lock. 852 * 853 * For pagecache intensive workloads, this function is the hottest 854 * spot in the kernel (apart from copy_*_user functions). 855 * 856 * Appropriate locks must be held before calling this function. 857 * 858 * @nr_to_scan: The number of pages to look through on the list. 859 * @src: The LRU list to pull pages off. 860 * @dst: The temp list to put pages on to. 861 * @scanned: The number of pages that were scanned. 862 * @order: The caller's attempted allocation order 863 * @mode: One of the LRU isolation modes 864 * @file: True [1] if isolating file [!anon] pages 865 * 866 * returns how many pages were moved onto *@dst. 867 */ 868static unsigned long isolate_lru_pages(unsigned long nr_to_scan, 869 struct list_head *src, struct list_head *dst, 870 unsigned long *scanned, int order, int mode, int file) 871{ 872 unsigned long nr_taken = 0; 873 unsigned long scan; 874 875 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { 876 struct page *page; 877 unsigned long pfn; 878 unsigned long end_pfn; 879 unsigned long page_pfn; 880 int zone_id; 881 882 page = lru_to_page(src); 883 prefetchw_prev_lru_page(page, src, flags); 884 885 VM_BUG_ON(!PageLRU(page)); 886 887 switch (__isolate_lru_page(page, mode, file)) { 888 case 0: 889 list_move(&page->lru, dst); 890 nr_taken++; 891 break; 892 893 case -EBUSY: 894 /* else it is being freed elsewhere */ 895 list_move(&page->lru, src); 896 continue; 897 898 default: 899 BUG(); 900 } 901 902 if (!order) 903 continue; 904 905 /* 906 * Attempt to take all pages in the order aligned region 907 * surrounding the tag page. Only take those pages of 908 * the same active state as that tag page. We may safely 909 * round the target page pfn down to the requested order 910 * as the mem_map is guarenteed valid out to MAX_ORDER, 911 * where that page is in a different zone we will detect 912 * it from its zone id and abort this block scan. 913 */ 914 zone_id = page_zone_id(page); 915 page_pfn = page_to_pfn(page); 916 pfn = page_pfn & ~((1 << order) - 1); 917 end_pfn = pfn + (1 << order); 918 for (; pfn < end_pfn; pfn++) { 919 struct page *cursor_page; 920 921 /* The target page is in the block, ignore it. */ 922 if (unlikely(pfn == page_pfn)) 923 continue; 924 925 /* Avoid holes within the zone. */ 926 if (unlikely(!pfn_valid_within(pfn))) 927 break; 928 929 cursor_page = pfn_to_page(pfn); 930 931 /* Check that we have not crossed a zone boundary. */ 932 if (unlikely(page_zone_id(cursor_page) != zone_id)) 933 continue; 934 switch (__isolate_lru_page(cursor_page, mode, file)) { 935 case 0: 936 list_move(&cursor_page->lru, dst); 937 nr_taken++; 938 scan++; 939 break; 940 941 case -EBUSY: 942 /* else it is being freed elsewhere */ 943 list_move(&cursor_page->lru, src); 944 default: 945 break; /* ! on LRU or wrong list */ 946 } 947 } 948 } 949 950 *scanned = scan; 951 return nr_taken; 952} 953 954static unsigned long isolate_pages_global(unsigned long nr, 955 struct list_head *dst, 956 unsigned long *scanned, int order, 957 int mode, struct zone *z, 958 struct mem_cgroup *mem_cont, 959 int active, int file) 960{ 961 int lru = LRU_BASE; 962 if (active) 963 lru += LRU_ACTIVE; 964 if (file) 965 lru += LRU_FILE; 966 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order, 967 mode, !!file); 968} 969 970/* 971 * clear_active_flags() is a helper for shrink_active_list(), clearing 972 * any active bits from the pages in the list. 973 */ 974static unsigned long clear_active_flags(struct list_head *page_list, 975 unsigned int *count) 976{ 977 int nr_active = 0; 978 int lru; 979 struct page *page; 980 981 list_for_each_entry(page, page_list, lru) { 982 lru = page_is_file_cache(page); 983 if (PageActive(page)) { 984 lru += LRU_ACTIVE; 985 ClearPageActive(page); 986 nr_active++; 987 } 988 count[lru]++; 989 } 990 991 return nr_active; 992} 993 994/** 995 * isolate_lru_page - tries to isolate a page from its LRU list 996 * @page: page to isolate from its LRU list 997 * 998 * Isolates a @page from an LRU list, clears PageLRU and adjusts the 999 * vmstat statistic corresponding to whatever LRU list the page was on. 1000 * 1001 * Returns 0 if the page was removed from an LRU list. 1002 * Returns -EBUSY if the page was not on an LRU list. 1003 * 1004 * The returned page will have PageLRU() cleared. If it was found on 1005 * the active list, it will have PageActive set. If it was found on 1006 * the unevictable list, it will have the PageUnevictable bit set. That flag 1007 * may need to be cleared by the caller before letting the page go. 1008 * 1009 * The vmstat statistic corresponding to the list on which the page was 1010 * found will be decremented. 1011 * 1012 * Restrictions: 1013 * (1) Must be called with an elevated refcount on the page. This is a 1014 * fundamentnal difference from isolate_lru_pages (which is called 1015 * without a stable reference). 1016 * (2) the lru_lock must not be held. 1017 * (3) interrupts must be enabled. 1018 */ 1019int isolate_lru_page(struct page *page) 1020{ 1021 int ret = -EBUSY; 1022 1023 if (PageLRU(page)) { 1024 struct zone *zone = page_zone(page); 1025 1026 spin_lock_irq(&zone->lru_lock); 1027 if (PageLRU(page) && get_page_unless_zero(page)) { 1028 int lru = page_lru(page); 1029 ret = 0; 1030 ClearPageLRU(page); 1031 1032 del_page_from_lru_list(zone, page, lru); 1033 } 1034 spin_unlock_irq(&zone->lru_lock); 1035 } 1036 return ret; 1037} 1038 1039/* 1040 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number 1041 * of reclaimed pages 1042 */ 1043static unsigned long shrink_inactive_list(unsigned long max_scan, 1044 struct zone *zone, struct scan_control *sc, 1045 int priority, int file) 1046{ 1047 LIST_HEAD(page_list); 1048 struct pagevec pvec; 1049 unsigned long nr_scanned = 0; 1050 unsigned long nr_reclaimed = 0; 1051 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); 1052 1053 pagevec_init(&pvec, 1); 1054 1055 lru_add_drain(); 1056 spin_lock_irq(&zone->lru_lock); 1057 do { 1058 struct page *page; 1059 unsigned long nr_taken; 1060 unsigned long nr_scan; 1061 unsigned long nr_freed; 1062 unsigned long nr_active; 1063 unsigned int count[NR_LRU_LISTS] = { 0, }; 1064 int mode = ISOLATE_INACTIVE; 1065 1066 /* 1067 * If we need a large contiguous chunk of memory, or have 1068 * trouble getting a small set of contiguous pages, we 1069 * will reclaim both active and inactive pages. 1070 * 1071 * We use the same threshold as pageout congestion_wait below. 1072 */ 1073 if (sc->order > PAGE_ALLOC_COSTLY_ORDER) 1074 mode = ISOLATE_BOTH; 1075 else if (sc->order && priority < DEF_PRIORITY - 2) 1076 mode = ISOLATE_BOTH; 1077 1078 nr_taken = sc->isolate_pages(sc->swap_cluster_max, 1079 &page_list, &nr_scan, sc->order, mode, 1080 zone, sc->mem_cgroup, 0, file); 1081 nr_active = clear_active_flags(&page_list, count); 1082 __count_vm_events(PGDEACTIVATE, nr_active); 1083 1084 __mod_zone_page_state(zone, NR_ACTIVE_FILE, 1085 -count[LRU_ACTIVE_FILE]); 1086 __mod_zone_page_state(zone, NR_INACTIVE_FILE, 1087 -count[LRU_INACTIVE_FILE]); 1088 __mod_zone_page_state(zone, NR_ACTIVE_ANON, 1089 -count[LRU_ACTIVE_ANON]); 1090 __mod_zone_page_state(zone, NR_INACTIVE_ANON, 1091 -count[LRU_INACTIVE_ANON]); 1092 1093 if (scanning_global_lru(sc)) 1094 zone->pages_scanned += nr_scan; 1095 1096 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON]; 1097 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON]; 1098 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE]; 1099 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE]; 1100 1101 spin_unlock_irq(&zone->lru_lock); 1102 1103 nr_scanned += nr_scan; 1104 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC); 1105 1106 /* 1107 * If we are direct reclaiming for contiguous pages and we do 1108 * not reclaim everything in the list, try again and wait 1109 * for IO to complete. This will stall high-order allocations 1110 * but that should be acceptable to the caller 1111 */ 1112 if (nr_freed < nr_taken && !current_is_kswapd() && 1113 sc->order > PAGE_ALLOC_COSTLY_ORDER) { 1114 congestion_wait(WRITE, HZ/10); 1115 1116 /* 1117 * The attempt at page out may have made some 1118 * of the pages active, mark them inactive again. 1119 */ 1120 nr_active = clear_active_flags(&page_list, count); 1121 count_vm_events(PGDEACTIVATE, nr_active); 1122 1123 nr_freed += shrink_page_list(&page_list, sc, 1124 PAGEOUT_IO_SYNC); 1125 } 1126 1127 nr_reclaimed += nr_freed; 1128 local_irq_disable(); 1129 if (current_is_kswapd()) { 1130 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan); 1131 __count_vm_events(KSWAPD_STEAL, nr_freed); 1132 } else if (scanning_global_lru(sc)) 1133 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan); 1134 1135 __count_zone_vm_events(PGSTEAL, zone, nr_freed); 1136 1137 if (nr_taken == 0) 1138 goto done; 1139 1140 spin_lock(&zone->lru_lock); 1141 /* 1142 * Put back any unfreeable pages. 1143 */ 1144 while (!list_empty(&page_list)) { 1145 int lru; 1146 page = lru_to_page(&page_list); 1147 VM_BUG_ON(PageLRU(page)); 1148 list_del(&page->lru); 1149 if (unlikely(!page_evictable(page, NULL))) { 1150 spin_unlock_irq(&zone->lru_lock); 1151 putback_lru_page(page); 1152 spin_lock_irq(&zone->lru_lock); 1153 continue; 1154 } 1155 SetPageLRU(page); 1156 lru = page_lru(page); 1157 add_page_to_lru_list(zone, page, lru); 1158 if (PageActive(page)) { 1159 int file = !!page_is_file_cache(page); 1160 reclaim_stat->recent_rotated[file]++; 1161 } 1162 if (!pagevec_add(&pvec, page)) { 1163 spin_unlock_irq(&zone->lru_lock); 1164 __pagevec_release(&pvec); 1165 spin_lock_irq(&zone->lru_lock); 1166 } 1167 } 1168 } while (nr_scanned < max_scan); 1169 spin_unlock(&zone->lru_lock); 1170done: 1171 local_irq_enable(); 1172 pagevec_release(&pvec); 1173 return nr_reclaimed; 1174} 1175 1176/* 1177 * We are about to scan this zone at a certain priority level. If that priority 1178 * level is smaller (ie: more urgent) than the previous priority, then note 1179 * that priority level within the zone. This is done so that when the next 1180 * process comes in to scan this zone, it will immediately start out at this 1181 * priority level rather than having to build up its own scanning priority. 1182 * Here, this priority affects only the reclaim-mapped threshold. 1183 */ 1184static inline void note_zone_scanning_priority(struct zone *zone, int priority) 1185{ 1186 if (priority < zone->prev_priority) 1187 zone->prev_priority = priority; 1188} 1189 1190/* 1191 * This moves pages from the active list to the inactive list. 1192 * 1193 * We move them the other way if the page is referenced by one or more 1194 * processes, from rmap. 1195 * 1196 * If the pages are mostly unmapped, the processing is fast and it is 1197 * appropriate to hold zone->lru_lock across the whole operation. But if 1198 * the pages are mapped, the processing is slow (page_referenced()) so we 1199 * should drop zone->lru_lock around each page. It's impossible to balance 1200 * this, so instead we remove the pages from the LRU while processing them. 1201 * It is safe to rely on PG_active against the non-LRU pages in here because 1202 * nobody will play with that bit on a non-LRU page. 1203 * 1204 * The downside is that we have to touch page->_count against each page. 1205 * But we had to alter page->flags anyway. 1206 */ 1207 1208 1209static void shrink_active_list(unsigned long nr_pages, struct zone *zone, 1210 struct scan_control *sc, int priority, int file) 1211{ 1212 unsigned long pgmoved; 1213 int pgdeactivate = 0; 1214 unsigned long pgscanned; 1215 LIST_HEAD(l_hold); /* The pages which were snipped off */ 1216 LIST_HEAD(l_inactive); 1217 struct page *page; 1218 struct pagevec pvec; 1219 enum lru_list lru; 1220 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); 1221 1222 lru_add_drain(); 1223 spin_lock_irq(&zone->lru_lock); 1224 pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order, 1225 ISOLATE_ACTIVE, zone, 1226 sc->mem_cgroup, 1, file); 1227 /* 1228 * zone->pages_scanned is used for detect zone's oom 1229 * mem_cgroup remembers nr_scan by itself. 1230 */ 1231 if (scanning_global_lru(sc)) { 1232 zone->pages_scanned += pgscanned; 1233 } 1234 reclaim_stat->recent_scanned[!!file] += pgmoved; 1235 1236 if (file) 1237 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved); 1238 else 1239 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved); 1240 spin_unlock_irq(&zone->lru_lock); 1241 1242 pgmoved = 0; 1243 while (!list_empty(&l_hold)) { 1244 cond_resched(); 1245 page = lru_to_page(&l_hold); 1246 list_del(&page->lru); 1247 1248 if (unlikely(!page_evictable(page, NULL))) { 1249 putback_lru_page(page); 1250 continue; 1251 } 1252 1253 /* page_referenced clears PageReferenced */ 1254 if (page_mapping_inuse(page) && 1255 page_referenced(page, 0, sc->mem_cgroup)) 1256 pgmoved++; 1257 1258 list_add(&page->lru, &l_inactive); 1259 } 1260 1261 /* 1262 * Move the pages to the [file or anon] inactive list. 1263 */ 1264 pagevec_init(&pvec, 1); 1265 pgmoved = 0; 1266 lru = LRU_BASE + file * LRU_FILE; 1267 1268 spin_lock_irq(&zone->lru_lock); 1269 /* 1270 * Count referenced pages from currently used mappings as 1271 * rotated, even though they are moved to the inactive list. 1272 * This helps balance scan pressure between file and anonymous 1273 * pages in get_scan_ratio. 1274 */ 1275 reclaim_stat->recent_rotated[!!file] += pgmoved; 1276 1277 while (!list_empty(&l_inactive)) { 1278 page = lru_to_page(&l_inactive); 1279 prefetchw_prev_lru_page(page, &l_inactive, flags); 1280 VM_BUG_ON(PageLRU(page)); 1281 SetPageLRU(page); 1282 VM_BUG_ON(!PageActive(page)); 1283 ClearPageActive(page); 1284 1285 list_move(&page->lru, &zone->lru[lru].list); 1286 mem_cgroup_add_lru_list(page, lru); 1287 pgmoved++; 1288 if (!pagevec_add(&pvec, page)) { 1289 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); 1290 spin_unlock_irq(&zone->lru_lock); 1291 pgdeactivate += pgmoved; 1292 pgmoved = 0; 1293 if (buffer_heads_over_limit) 1294 pagevec_strip(&pvec); 1295 __pagevec_release(&pvec); 1296 spin_lock_irq(&zone->lru_lock); 1297 } 1298 } 1299 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved); 1300 pgdeactivate += pgmoved; 1301 if (buffer_heads_over_limit) { 1302 spin_unlock_irq(&zone->lru_lock); 1303 pagevec_strip(&pvec); 1304 spin_lock_irq(&zone->lru_lock); 1305 } 1306 __count_zone_vm_events(PGREFILL, zone, pgscanned); 1307 __count_vm_events(PGDEACTIVATE, pgdeactivate); 1308 spin_unlock_irq(&zone->lru_lock); 1309 if (vm_swap_full()) 1310 pagevec_swap_free(&pvec); 1311 1312 pagevec_release(&pvec); 1313} 1314 1315static int inactive_anon_is_low_global(struct zone *zone) 1316{ 1317 unsigned long active, inactive; 1318 1319 active = zone_page_state(zone, NR_ACTIVE_ANON); 1320 inactive = zone_page_state(zone, NR_INACTIVE_ANON); 1321 1322 if (inactive * zone->inactive_ratio < active) 1323 return 1; 1324 1325 return 0; 1326} 1327 1328/** 1329 * inactive_anon_is_low - check if anonymous pages need to be deactivated 1330 * @zone: zone to check 1331 * @sc: scan control of this context 1332 * 1333 * Returns true if the zone does not have enough inactive anon pages, 1334 * meaning some active anon pages need to be deactivated. 1335 */ 1336static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc) 1337{ 1338 int low; 1339 1340 if (scanning_global_lru(sc)) 1341 low = inactive_anon_is_low_global(zone); 1342 else 1343 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup); 1344 return low; 1345} 1346 1347static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, 1348 struct zone *zone, struct scan_control *sc, int priority) 1349{ 1350 int file = is_file_lru(lru); 1351 1352 if (lru == LRU_ACTIVE_FILE) { 1353 shrink_active_list(nr_to_scan, zone, sc, priority, file); 1354 return 0; 1355 } 1356 1357 if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) { 1358 shrink_active_list(nr_to_scan, zone, sc, priority, file); 1359 return 0; 1360 } 1361 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file); 1362} 1363 1364/* 1365 * Determine how aggressively the anon and file LRU lists should be 1366 * scanned. The relative value of each set of LRU lists is determined 1367 * by looking at the fraction of the pages scanned we did rotate back 1368 * onto the active list instead of evict. 1369 * 1370 * percent[0] specifies how much pressure to put on ram/swap backed 1371 * memory, while percent[1] determines pressure on the file LRUs. 1372 */ 1373static void get_scan_ratio(struct zone *zone, struct scan_control *sc, 1374 unsigned long *percent) 1375{ 1376 unsigned long anon, file, free; 1377 unsigned long anon_prio, file_prio; 1378 unsigned long ap, fp; 1379 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc); 1380 1381 /* If we have no swap space, do not bother scanning anon pages. */ 1382 if (nr_swap_pages <= 0) { 1383 percent[0] = 0; 1384 percent[1] = 100; 1385 return; 1386 } 1387 1388 anon = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) + 1389 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON); 1390 file = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) + 1391 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE); 1392 1393 if (scanning_global_lru(sc)) { 1394 free = zone_page_state(zone, NR_FREE_PAGES); 1395 /* If we have very few page cache pages, 1396 force-scan anon pages. */ 1397 if (unlikely(file + free <= zone->pages_high)) { 1398 percent[0] = 100; 1399 percent[1] = 0; 1400 return; 1401 } 1402 } 1403 1404 /* 1405 * OK, so we have swap space and a fair amount of page cache 1406 * pages. We use the recently rotated / recently scanned 1407 * ratios to determine how valuable each cache is. 1408 * 1409 * Because workloads change over time (and to avoid overflow) 1410 * we keep these statistics as a floating average, which ends 1411 * up weighing recent references more than old ones. 1412 * 1413 * anon in [0], file in [1] 1414 */ 1415 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { 1416 spin_lock_irq(&zone->lru_lock); 1417 reclaim_stat->recent_scanned[0] /= 2; 1418 reclaim_stat->recent_rotated[0] /= 2; 1419 spin_unlock_irq(&zone->lru_lock); 1420 } 1421 1422 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { 1423 spin_lock_irq(&zone->lru_lock); 1424 reclaim_stat->recent_scanned[1] /= 2; 1425 reclaim_stat->recent_rotated[1] /= 2; 1426 spin_unlock_irq(&zone->lru_lock); 1427 } 1428 1429 /* 1430 * With swappiness at 100, anonymous and file have the same priority. 1431 * This scanning priority is essentially the inverse of IO cost. 1432 */ 1433 anon_prio = sc->swappiness; 1434 file_prio = 200 - sc->swappiness; 1435 1436 /* 1437 * The amount of pressure on anon vs file pages is inversely 1438 * proportional to the fraction of recently scanned pages on 1439 * each list that were recently referenced and in active use. 1440 */ 1441 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1); 1442 ap /= reclaim_stat->recent_rotated[0] + 1; 1443 1444 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1); 1445 fp /= reclaim_stat->recent_rotated[1] + 1; 1446 1447 /* Normalize to percentages */ 1448 percent[0] = 100 * ap / (ap + fp + 1); 1449 percent[1] = 100 - percent[0]; 1450} 1451 1452 1453/* 1454 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. 1455 */ 1456static void shrink_zone(int priority, struct zone *zone, 1457 struct scan_control *sc) 1458{ 1459 unsigned long nr[NR_LRU_LISTS]; 1460 unsigned long nr_to_scan; 1461 unsigned long percent[2]; /* anon @ 0; file @ 1 */ 1462 enum lru_list l; 1463 unsigned long nr_reclaimed = sc->nr_reclaimed; 1464 unsigned long swap_cluster_max = sc->swap_cluster_max; 1465 1466 get_scan_ratio(zone, sc, percent); 1467 1468 for_each_evictable_lru(l) { 1469 int file = is_file_lru(l); 1470 int scan; 1471 1472 scan = zone_page_state(zone, NR_LRU_BASE + l); 1473 if (priority) { 1474 scan >>= priority; 1475 scan = (scan * percent[file]) / 100; 1476 } 1477 if (scanning_global_lru(sc)) { 1478 zone->lru[l].nr_scan += scan; 1479 nr[l] = zone->lru[l].nr_scan; 1480 if (nr[l] >= swap_cluster_max) 1481 zone->lru[l].nr_scan = 0; 1482 else 1483 nr[l] = 0; 1484 } else 1485 nr[l] = scan; 1486 } 1487 1488 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || 1489 nr[LRU_INACTIVE_FILE]) { 1490 for_each_evictable_lru(l) { 1491 if (nr[l]) { 1492 nr_to_scan = min(nr[l], swap_cluster_max); 1493 nr[l] -= nr_to_scan; 1494 1495 nr_reclaimed += shrink_list(l, nr_to_scan, 1496 zone, sc, priority); 1497 } 1498 } 1499 /* 1500 * On large memory systems, scan >> priority can become 1501 * really large. This is fine for the starting priority; 1502 * we want to put equal scanning pressure on each zone. 1503 * However, if the VM has a harder time of freeing pages, 1504 * with multiple processes reclaiming pages, the total 1505 * freeing target can get unreasonably large. 1506 */ 1507 if (nr_reclaimed > swap_cluster_max && 1508 priority < DEF_PRIORITY && !current_is_kswapd()) 1509 break; 1510 } 1511 1512 sc->nr_reclaimed = nr_reclaimed; 1513 1514 /* 1515 * Even if we did not try to evict anon pages at all, we want to 1516 * rebalance the anon lru active/inactive ratio. 1517 */ 1518 if (inactive_anon_is_low(zone, sc)) 1519 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0); 1520 1521 throttle_vm_writeout(sc->gfp_mask); 1522} 1523 1524/* 1525 * This is the direct reclaim path, for page-allocating processes. We only 1526 * try to reclaim pages from zones which will satisfy the caller's allocation 1527 * request. 1528 * 1529 * We reclaim from a zone even if that zone is over pages_high. Because: 1530 * a) The caller may be trying to free *extra* pages to satisfy a higher-order 1531 * allocation or 1532 * b) The zones may be over pages_high but they must go *over* pages_high to 1533 * satisfy the `incremental min' zone defense algorithm. 1534 * 1535 * If a zone is deemed to be full of pinned pages then just give it a light 1536 * scan then give up on it. 1537 */ 1538static void shrink_zones(int priority, struct zonelist *zonelist, 1539 struct scan_control *sc) 1540{ 1541 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask); 1542 struct zoneref *z; 1543 struct zone *zone; 1544 1545 sc->all_unreclaimable = 1; 1546 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1547 if (!populated_zone(zone)) 1548 continue; 1549 /* 1550 * Take care memory controller reclaiming has small influence 1551 * to global LRU. 1552 */ 1553 if (scanning_global_lru(sc)) { 1554 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 1555 continue; 1556 note_zone_scanning_priority(zone, priority); 1557 1558 if (zone_is_all_unreclaimable(zone) && 1559 priority != DEF_PRIORITY) 1560 continue; /* Let kswapd poll it */ 1561 sc->all_unreclaimable = 0; 1562 } else { 1563 /* 1564 * Ignore cpuset limitation here. We just want to reduce 1565 * # of used pages by us regardless of memory shortage. 1566 */ 1567 sc->all_unreclaimable = 0; 1568 mem_cgroup_note_reclaim_priority(sc->mem_cgroup, 1569 priority); 1570 } 1571 1572 shrink_zone(priority, zone, sc); 1573 } 1574} 1575 1576/* 1577 * This is the main entry point to direct page reclaim. 1578 * 1579 * If a full scan of the inactive list fails to free enough memory then we 1580 * are "out of memory" and something needs to be killed. 1581 * 1582 * If the caller is !__GFP_FS then the probability of a failure is reasonably 1583 * high - the zone may be full of dirty or under-writeback pages, which this 1584 * caller can't do much about. We kick pdflush and take explicit naps in the 1585 * hope that some of these pages can be written. But if the allocating task 1586 * holds filesystem locks which prevent writeout this might not work, and the 1587 * allocation attempt will fail. 1588 * 1589 * returns: 0, if no pages reclaimed 1590 * else, the number of pages reclaimed 1591 */ 1592static unsigned long do_try_to_free_pages(struct zonelist *zonelist, 1593 struct scan_control *sc) 1594{ 1595 int priority; 1596 unsigned long ret = 0; 1597 unsigned long total_scanned = 0; 1598 struct reclaim_state *reclaim_state = current->reclaim_state; 1599 unsigned long lru_pages = 0; 1600 struct zoneref *z; 1601 struct zone *zone; 1602 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask); 1603 1604 delayacct_freepages_start(); 1605 1606 if (scanning_global_lru(sc)) 1607 count_vm_event(ALLOCSTALL); 1608 /* 1609 * mem_cgroup will not do shrink_slab. 1610 */ 1611 if (scanning_global_lru(sc)) { 1612 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1613 1614 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 1615 continue; 1616 1617 lru_pages += zone_lru_pages(zone); 1618 } 1619 } 1620 1621 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 1622 sc->nr_scanned = 0; 1623 if (!priority) 1624 disable_swap_token(); 1625 shrink_zones(priority, zonelist, sc); 1626 /* 1627 * Don't shrink slabs when reclaiming memory from 1628 * over limit cgroups 1629 */ 1630 if (scanning_global_lru(sc)) { 1631 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages); 1632 if (reclaim_state) { 1633 sc->nr_reclaimed += reclaim_state->reclaimed_slab; 1634 reclaim_state->reclaimed_slab = 0; 1635 } 1636 } 1637 total_scanned += sc->nr_scanned; 1638 if (sc->nr_reclaimed >= sc->swap_cluster_max) { 1639 ret = sc->nr_reclaimed; 1640 goto out; 1641 } 1642 1643 /* 1644 * Try to write back as many pages as we just scanned. This 1645 * tends to cause slow streaming writers to write data to the 1646 * disk smoothly, at the dirtying rate, which is nice. But 1647 * that's undesirable in laptop mode, where we *want* lumpy 1648 * writeout. So in laptop mode, write out the whole world. 1649 */ 1650 if (total_scanned > sc->swap_cluster_max + 1651 sc->swap_cluster_max / 2) { 1652 wakeup_pdflush(laptop_mode ? 0 : total_scanned); 1653 sc->may_writepage = 1; 1654 } 1655 1656 /* Take a nap, wait for some writeback to complete */ 1657 if (sc->nr_scanned && priority < DEF_PRIORITY - 2) 1658 congestion_wait(WRITE, HZ/10); 1659 } 1660 /* top priority shrink_zones still had more to do? don't OOM, then */ 1661 if (!sc->all_unreclaimable && scanning_global_lru(sc)) 1662 ret = sc->nr_reclaimed; 1663out: 1664 /* 1665 * Now that we've scanned all the zones at this priority level, note 1666 * that level within the zone so that the next thread which performs 1667 * scanning of this zone will immediately start out at this priority 1668 * level. This affects only the decision whether or not to bring 1669 * mapped pages onto the inactive list. 1670 */ 1671 if (priority < 0) 1672 priority = 0; 1673 1674 if (scanning_global_lru(sc)) { 1675 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1676 1677 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 1678 continue; 1679 1680 zone->prev_priority = priority; 1681 } 1682 } else 1683 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority); 1684 1685 delayacct_freepages_end(); 1686 1687 return ret; 1688} 1689 1690unsigned long try_to_free_pages(struct zonelist *zonelist, int order, 1691 gfp_t gfp_mask) 1692{ 1693 struct scan_control sc = { 1694 .gfp_mask = gfp_mask, 1695 .may_writepage = !laptop_mode, 1696 .swap_cluster_max = SWAP_CLUSTER_MAX, 1697 .may_swap = 1, 1698 .swappiness = vm_swappiness, 1699 .order = order, 1700 .mem_cgroup = NULL, 1701 .isolate_pages = isolate_pages_global, 1702 }; 1703 1704 return do_try_to_free_pages(zonelist, &sc); 1705} 1706 1707#ifdef CONFIG_CGROUP_MEM_RES_CTLR 1708 1709unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont, 1710 gfp_t gfp_mask, 1711 bool noswap, 1712 unsigned int swappiness) 1713{ 1714 struct scan_control sc = { 1715 .may_writepage = !laptop_mode, 1716 .may_swap = 1, 1717 .swap_cluster_max = SWAP_CLUSTER_MAX, 1718 .swappiness = swappiness, 1719 .order = 0, 1720 .mem_cgroup = mem_cont, 1721 .isolate_pages = mem_cgroup_isolate_pages, 1722 }; 1723 struct zonelist *zonelist; 1724 1725 if (noswap) 1726 sc.may_swap = 0; 1727 1728 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | 1729 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); 1730 zonelist = NODE_DATA(numa_node_id())->node_zonelists; 1731 return do_try_to_free_pages(zonelist, &sc); 1732} 1733#endif 1734 1735/* 1736 * For kswapd, balance_pgdat() will work across all this node's zones until 1737 * they are all at pages_high. 1738 * 1739 * Returns the number of pages which were actually freed. 1740 * 1741 * There is special handling here for zones which are full of pinned pages. 1742 * This can happen if the pages are all mlocked, or if they are all used by 1743 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. 1744 * What we do is to detect the case where all pages in the zone have been 1745 * scanned twice and there has been zero successful reclaim. Mark the zone as 1746 * dead and from now on, only perform a short scan. Basically we're polling 1747 * the zone for when the problem goes away. 1748 * 1749 * kswapd scans the zones in the highmem->normal->dma direction. It skips 1750 * zones which have free_pages > pages_high, but once a zone is found to have 1751 * free_pages <= pages_high, we scan that zone and the lower zones regardless 1752 * of the number of free pages in the lower zones. This interoperates with 1753 * the page allocator fallback scheme to ensure that aging of pages is balanced 1754 * across the zones. 1755 */ 1756static unsigned long balance_pgdat(pg_data_t *pgdat, int order) 1757{ 1758 int all_zones_ok; 1759 int priority; 1760 int i; 1761 unsigned long total_scanned; 1762 struct reclaim_state *reclaim_state = current->reclaim_state; 1763 struct scan_control sc = { 1764 .gfp_mask = GFP_KERNEL, 1765 .may_swap = 1, 1766 .swap_cluster_max = SWAP_CLUSTER_MAX, 1767 .swappiness = vm_swappiness, 1768 .order = order, 1769 .mem_cgroup = NULL, 1770 .isolate_pages = isolate_pages_global, 1771 }; 1772 /* 1773 * temp_priority is used to remember the scanning priority at which 1774 * this zone was successfully refilled to free_pages == pages_high. 1775 */ 1776 int temp_priority[MAX_NR_ZONES]; 1777 1778loop_again: 1779 total_scanned = 0; 1780 sc.nr_reclaimed = 0; 1781 sc.may_writepage = !laptop_mode; 1782 count_vm_event(PAGEOUTRUN); 1783 1784 for (i = 0; i < pgdat->nr_zones; i++) 1785 temp_priority[i] = DEF_PRIORITY; 1786 1787 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 1788 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ 1789 unsigned long lru_pages = 0; 1790 1791 /* The swap token gets in the way of swapout... */ 1792 if (!priority) 1793 disable_swap_token(); 1794 1795 all_zones_ok = 1; 1796 1797 /* 1798 * Scan in the highmem->dma direction for the highest 1799 * zone which needs scanning 1800 */ 1801 for (i = pgdat->nr_zones - 1; i >= 0; i--) { 1802 struct zone *zone = pgdat->node_zones + i; 1803 1804 if (!populated_zone(zone)) 1805 continue; 1806 1807 if (zone_is_all_unreclaimable(zone) && 1808 priority != DEF_PRIORITY) 1809 continue; 1810 1811 /* 1812 * Do some background aging of the anon list, to give 1813 * pages a chance to be referenced before reclaiming. 1814 */ 1815 if (inactive_anon_is_low(zone, &sc)) 1816 shrink_active_list(SWAP_CLUSTER_MAX, zone, 1817 &sc, priority, 0); 1818 1819 if (!zone_watermark_ok(zone, order, zone->pages_high, 1820 0, 0)) { 1821 end_zone = i; 1822 break; 1823 } 1824 } 1825 if (i < 0) 1826 goto out; 1827 1828 for (i = 0; i <= end_zone; i++) { 1829 struct zone *zone = pgdat->node_zones + i; 1830 1831 lru_pages += zone_lru_pages(zone); 1832 } 1833 1834 /* 1835 * Now scan the zone in the dma->highmem direction, stopping 1836 * at the last zone which needs scanning. 1837 * 1838 * We do this because the page allocator works in the opposite 1839 * direction. This prevents the page allocator from allocating 1840 * pages behind kswapd's direction of progress, which would 1841 * cause too much scanning of the lower zones. 1842 */ 1843 for (i = 0; i <= end_zone; i++) { 1844 struct zone *zone = pgdat->node_zones + i; 1845 int nr_slab; 1846 1847 if (!populated_zone(zone)) 1848 continue; 1849 1850 if (zone_is_all_unreclaimable(zone) && 1851 priority != DEF_PRIORITY) 1852 continue; 1853 1854 if (!zone_watermark_ok(zone, order, zone->pages_high, 1855 end_zone, 0)) 1856 all_zones_ok = 0; 1857 temp_priority[i] = priority; 1858 sc.nr_scanned = 0; 1859 note_zone_scanning_priority(zone, priority); 1860 /* 1861 * We put equal pressure on every zone, unless one 1862 * zone has way too many pages free already. 1863 */ 1864 if (!zone_watermark_ok(zone, order, 8*zone->pages_high, 1865 end_zone, 0)) 1866 shrink_zone(priority, zone, &sc); 1867 reclaim_state->reclaimed_slab = 0; 1868 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL, 1869 lru_pages); 1870 sc.nr_reclaimed += reclaim_state->reclaimed_slab; 1871 total_scanned += sc.nr_scanned; 1872 if (zone_is_all_unreclaimable(zone)) 1873 continue; 1874 if (nr_slab == 0 && zone->pages_scanned >= 1875 (zone_lru_pages(zone) * 6)) 1876 zone_set_flag(zone, 1877 ZONE_ALL_UNRECLAIMABLE); 1878 /* 1879 * If we've done a decent amount of scanning and 1880 * the reclaim ratio is low, start doing writepage 1881 * even in laptop mode 1882 */ 1883 if (total_scanned > SWAP_CLUSTER_MAX * 2 && 1884 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2) 1885 sc.may_writepage = 1; 1886 } 1887 if (all_zones_ok) 1888 break; /* kswapd: all done */ 1889 /* 1890 * OK, kswapd is getting into trouble. Take a nap, then take 1891 * another pass across the zones. 1892 */ 1893 if (total_scanned && priority < DEF_PRIORITY - 2) 1894 congestion_wait(WRITE, HZ/10); 1895 1896 /* 1897 * We do this so kswapd doesn't build up large priorities for 1898 * example when it is freeing in parallel with allocators. It 1899 * matches the direct reclaim path behaviour in terms of impact 1900 * on zone->*_priority. 1901 */ 1902 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) 1903 break; 1904 } 1905out: 1906 /* 1907 * Note within each zone the priority level at which this zone was 1908 * brought into a happy state. So that the next thread which scans this 1909 * zone will start out at that priority level. 1910 */ 1911 for (i = 0; i < pgdat->nr_zones; i++) { 1912 struct zone *zone = pgdat->node_zones + i; 1913 1914 zone->prev_priority = temp_priority[i]; 1915 } 1916 if (!all_zones_ok) { 1917 cond_resched(); 1918 1919 try_to_freeze(); 1920 1921 /* 1922 * Fragmentation may mean that the system cannot be 1923 * rebalanced for high-order allocations in all zones. 1924 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX, 1925 * it means the zones have been fully scanned and are still 1926 * not balanced. For high-order allocations, there is 1927 * little point trying all over again as kswapd may 1928 * infinite loop. 1929 * 1930 * Instead, recheck all watermarks at order-0 as they 1931 * are the most important. If watermarks are ok, kswapd will go 1932 * back to sleep. High-order users can still perform direct 1933 * reclaim if they wish. 1934 */ 1935 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX) 1936 order = sc.order = 0; 1937 1938 goto loop_again; 1939 } 1940 1941 return sc.nr_reclaimed; 1942} 1943 1944/* 1945 * The background pageout daemon, started as a kernel thread 1946 * from the init process. 1947 * 1948 * This basically trickles out pages so that we have _some_ 1949 * free memory available even if there is no other activity 1950 * that frees anything up. This is needed for things like routing 1951 * etc, where we otherwise might have all activity going on in 1952 * asynchronous contexts that cannot page things out. 1953 * 1954 * If there are applications that are active memory-allocators 1955 * (most normal use), this basically shouldn't matter. 1956 */ 1957static int kswapd(void *p) 1958{ 1959 unsigned long order; 1960 pg_data_t *pgdat = (pg_data_t*)p; 1961 struct task_struct *tsk = current; 1962 DEFINE_WAIT(wait); 1963 struct reclaim_state reclaim_state = { 1964 .reclaimed_slab = 0, 1965 }; 1966 1967 lockdep_set_current_reclaim_state(GFP_KERNEL); 1968 1969 node_to_cpumask_ptr(cpumask, pgdat->node_id); 1970 1971 if (!cpumask_empty(cpumask)) 1972 set_cpus_allowed_ptr(tsk, cpumask); 1973 current->reclaim_state = &reclaim_state; 1974 1975 /* 1976 * Tell the memory management that we're a "memory allocator", 1977 * and that if we need more memory we should get access to it 1978 * regardless (see "__alloc_pages()"). "kswapd" should 1979 * never get caught in the normal page freeing logic. 1980 * 1981 * (Kswapd normally doesn't need memory anyway, but sometimes 1982 * you need a small amount of memory in order to be able to 1983 * page out something else, and this flag essentially protects 1984 * us from recursively trying to free more memory as we're 1985 * trying to free the first piece of memory in the first place). 1986 */ 1987 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; 1988 set_freezable(); 1989 1990 order = 0; 1991 for ( ; ; ) { 1992 unsigned long new_order; 1993 1994 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 1995 new_order = pgdat->kswapd_max_order; 1996 pgdat->kswapd_max_order = 0; 1997 if (order < new_order) { 1998 /* 1999 * Don't sleep if someone wants a larger 'order' 2000 * allocation 2001 */ 2002 order = new_order; 2003 } else { 2004 if (!freezing(current)) 2005 schedule(); 2006 2007 order = pgdat->kswapd_max_order; 2008 } 2009 finish_wait(&pgdat->kswapd_wait, &wait); 2010 2011 if (!try_to_freeze()) { 2012 /* We can speed up thawing tasks if we don't call 2013 * balance_pgdat after returning from the refrigerator 2014 */ 2015 balance_pgdat(pgdat, order); 2016 } 2017 } 2018 return 0; 2019} 2020 2021/* 2022 * A zone is low on free memory, so wake its kswapd task to service it. 2023 */ 2024void wakeup_kswapd(struct zone *zone, int order) 2025{ 2026 pg_data_t *pgdat; 2027 2028 if (!populated_zone(zone)) 2029 return; 2030 2031 pgdat = zone->zone_pgdat; 2032 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0)) 2033 return; 2034 if (pgdat->kswapd_max_order < order) 2035 pgdat->kswapd_max_order = order; 2036 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 2037 return; 2038 if (!waitqueue_active(&pgdat->kswapd_wait)) 2039 return; 2040 wake_up_interruptible(&pgdat->kswapd_wait); 2041} 2042 2043unsigned long global_lru_pages(void) 2044{ 2045 return global_page_state(NR_ACTIVE_ANON) 2046 + global_page_state(NR_ACTIVE_FILE) 2047 + global_page_state(NR_INACTIVE_ANON) 2048 + global_page_state(NR_INACTIVE_FILE); 2049} 2050 2051#ifdef CONFIG_PM 2052/* 2053 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages 2054 * from LRU lists system-wide, for given pass and priority, and returns the 2055 * number of reclaimed pages 2056 * 2057 * For pass > 3 we also try to shrink the LRU lists that contain a few pages 2058 */ 2059static unsigned long shrink_all_zones(unsigned long nr_pages, int prio, 2060 int pass, struct scan_control *sc) 2061{ 2062 struct zone *zone; 2063 unsigned long nr_to_scan, ret = 0; 2064 enum lru_list l; 2065 2066 for_each_zone(zone) { 2067 2068 if (!populated_zone(zone)) 2069 continue; 2070 2071 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY) 2072 continue; 2073 2074 for_each_evictable_lru(l) { 2075 /* For pass = 0, we don't shrink the active list */ 2076 if (pass == 0 && 2077 (l == LRU_ACTIVE || l == LRU_ACTIVE_FILE)) 2078 continue; 2079 2080 zone->lru[l].nr_scan += 2081 (zone_page_state(zone, NR_LRU_BASE + l) 2082 >> prio) + 1; 2083 if (zone->lru[l].nr_scan >= nr_pages || pass > 3) { 2084 zone->lru[l].nr_scan = 0; 2085 nr_to_scan = min(nr_pages, 2086 zone_page_state(zone, 2087 NR_LRU_BASE + l)); 2088 ret += shrink_list(l, nr_to_scan, zone, 2089 sc, prio); 2090 if (ret >= nr_pages) 2091 return ret; 2092 } 2093 } 2094 } 2095 2096 return ret; 2097} 2098 2099/* 2100 * Try to free `nr_pages' of memory, system-wide, and return the number of 2101 * freed pages. 2102 * 2103 * Rather than trying to age LRUs the aim is to preserve the overall 2104 * LRU order by reclaiming preferentially 2105 * inactive > active > active referenced > active mapped 2106 */ 2107unsigned long shrink_all_memory(unsigned long nr_pages) 2108{ 2109 unsigned long lru_pages, nr_slab; 2110 unsigned long ret = 0; 2111 int pass; 2112 struct reclaim_state reclaim_state; 2113 struct scan_control sc = { 2114 .gfp_mask = GFP_KERNEL, 2115 .may_swap = 0, 2116 .swap_cluster_max = nr_pages, 2117 .may_writepage = 1, 2118 .swappiness = vm_swappiness, 2119 .isolate_pages = isolate_pages_global, 2120 }; 2121 2122 current->reclaim_state = &reclaim_state; 2123 2124 lru_pages = global_lru_pages(); 2125 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE); 2126 /* If slab caches are huge, it's better to hit them first */ 2127 while (nr_slab >= lru_pages) { 2128 reclaim_state.reclaimed_slab = 0; 2129 shrink_slab(nr_pages, sc.gfp_mask, lru_pages); 2130 if (!reclaim_state.reclaimed_slab) 2131 break; 2132 2133 ret += reclaim_state.reclaimed_slab; 2134 if (ret >= nr_pages) 2135 goto out; 2136 2137 nr_slab -= reclaim_state.reclaimed_slab; 2138 } 2139 2140 /* 2141 * We try to shrink LRUs in 5 passes: 2142 * 0 = Reclaim from inactive_list only 2143 * 1 = Reclaim from active list but don't reclaim mapped 2144 * 2 = 2nd pass of type 1 2145 * 3 = Reclaim mapped (normal reclaim) 2146 * 4 = 2nd pass of type 3 2147 */ 2148 for (pass = 0; pass < 5; pass++) { 2149 int prio; 2150 2151 /* Force reclaiming mapped pages in the passes #3 and #4 */ 2152 if (pass > 2) { 2153 sc.may_swap = 1; 2154 sc.swappiness = 100; 2155 } 2156 2157 for (prio = DEF_PRIORITY; prio >= 0; prio--) { 2158 unsigned long nr_to_scan = nr_pages - ret; 2159 2160 sc.nr_scanned = 0; 2161 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc); 2162 if (ret >= nr_pages) 2163 goto out; 2164 2165 reclaim_state.reclaimed_slab = 0; 2166 shrink_slab(sc.nr_scanned, sc.gfp_mask, 2167 global_lru_pages()); 2168 ret += reclaim_state.reclaimed_slab; 2169 if (ret >= nr_pages) 2170 goto out; 2171 2172 if (sc.nr_scanned && prio < DEF_PRIORITY - 2) 2173 congestion_wait(WRITE, HZ / 10); 2174 } 2175 } 2176 2177 /* 2178 * If ret = 0, we could not shrink LRUs, but there may be something 2179 * in slab caches 2180 */ 2181 if (!ret) { 2182 do { 2183 reclaim_state.reclaimed_slab = 0; 2184 shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages()); 2185 ret += reclaim_state.reclaimed_slab; 2186 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0); 2187 } 2188 2189out: 2190 current->reclaim_state = NULL; 2191 2192 return ret; 2193} 2194#endif 2195 2196/* It's optimal to keep kswapds on the same CPUs as their memory, but 2197 not required for correctness. So if the last cpu in a node goes 2198 away, we get changed to run anywhere: as the first one comes back, 2199 restore their cpu bindings. */ 2200static int __devinit cpu_callback(struct notifier_block *nfb, 2201 unsigned long action, void *hcpu) 2202{ 2203 int nid; 2204 2205 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { 2206 for_each_node_state(nid, N_HIGH_MEMORY) { 2207 pg_data_t *pgdat = NODE_DATA(nid); 2208 node_to_cpumask_ptr(mask, pgdat->node_id); 2209 2210 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) 2211 /* One of our CPUs online: restore mask */ 2212 set_cpus_allowed_ptr(pgdat->kswapd, mask); 2213 } 2214 } 2215 return NOTIFY_OK; 2216} 2217 2218/* 2219 * This kswapd start function will be called by init and node-hot-add. 2220 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. 2221 */ 2222int kswapd_run(int nid) 2223{ 2224 pg_data_t *pgdat = NODE_DATA(nid); 2225 int ret = 0; 2226 2227 if (pgdat->kswapd) 2228 return 0; 2229 2230 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); 2231 if (IS_ERR(pgdat->kswapd)) { 2232 /* failure at boot is fatal */ 2233 BUG_ON(system_state == SYSTEM_BOOTING); 2234 printk("Failed to start kswapd on node %d\n",nid); 2235 ret = -1; 2236 } 2237 return ret; 2238} 2239 2240static int __init kswapd_init(void) 2241{ 2242 int nid; 2243 2244 swap_setup(); 2245 for_each_node_state(nid, N_HIGH_MEMORY) 2246 kswapd_run(nid); 2247 hotcpu_notifier(cpu_callback, 0); 2248 return 0; 2249} 2250 2251module_init(kswapd_init) 2252 2253#ifdef CONFIG_NUMA 2254/* 2255 * Zone reclaim mode 2256 * 2257 * If non-zero call zone_reclaim when the number of free pages falls below 2258 * the watermarks. 2259 */ 2260int zone_reclaim_mode __read_mostly; 2261 2262#define RECLAIM_OFF 0 2263#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ 2264#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ 2265#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ 2266 2267/* 2268 * Priority for ZONE_RECLAIM. This determines the fraction of pages 2269 * of a node considered for each zone_reclaim. 4 scans 1/16th of 2270 * a zone. 2271 */ 2272#define ZONE_RECLAIM_PRIORITY 4 2273 2274/* 2275 * Percentage of pages in a zone that must be unmapped for zone_reclaim to 2276 * occur. 2277 */ 2278int sysctl_min_unmapped_ratio = 1; 2279 2280/* 2281 * If the number of slab pages in a zone grows beyond this percentage then 2282 * slab reclaim needs to occur. 2283 */ 2284int sysctl_min_slab_ratio = 5; 2285 2286/* 2287 * Try to free up some pages from this zone through reclaim. 2288 */ 2289static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 2290{ 2291 /* Minimum pages needed in order to stay on node */ 2292 const unsigned long nr_pages = 1 << order; 2293 struct task_struct *p = current; 2294 struct reclaim_state reclaim_state; 2295 int priority; 2296 struct scan_control sc = { 2297 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), 2298 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP), 2299 .swap_cluster_max = max_t(unsigned long, nr_pages, 2300 SWAP_CLUSTER_MAX), 2301 .gfp_mask = gfp_mask, 2302 .swappiness = vm_swappiness, 2303 .isolate_pages = isolate_pages_global, 2304 }; 2305 unsigned long slab_reclaimable; 2306 2307 disable_swap_token(); 2308 cond_resched(); 2309 /* 2310 * We need to be able to allocate from the reserves for RECLAIM_SWAP 2311 * and we also need to be able to write out pages for RECLAIM_WRITE 2312 * and RECLAIM_SWAP. 2313 */ 2314 p->flags |= PF_MEMALLOC | PF_SWAPWRITE; 2315 reclaim_state.reclaimed_slab = 0; 2316 p->reclaim_state = &reclaim_state; 2317 2318 if (zone_page_state(zone, NR_FILE_PAGES) - 2319 zone_page_state(zone, NR_FILE_MAPPED) > 2320 zone->min_unmapped_pages) { 2321 /* 2322 * Free memory by calling shrink zone with increasing 2323 * priorities until we have enough memory freed. 2324 */ 2325 priority = ZONE_RECLAIM_PRIORITY; 2326 do { 2327 note_zone_scanning_priority(zone, priority); 2328 shrink_zone(priority, zone, &sc); 2329 priority--; 2330 } while (priority >= 0 && sc.nr_reclaimed < nr_pages); 2331 } 2332 2333 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE); 2334 if (slab_reclaimable > zone->min_slab_pages) { 2335 /* 2336 * shrink_slab() does not currently allow us to determine how 2337 * many pages were freed in this zone. So we take the current 2338 * number of slab pages and shake the slab until it is reduced 2339 * by the same nr_pages that we used for reclaiming unmapped 2340 * pages. 2341 * 2342 * Note that shrink_slab will free memory on all zones and may 2343 * take a long time. 2344 */ 2345 while (shrink_slab(sc.nr_scanned, gfp_mask, order) && 2346 zone_page_state(zone, NR_SLAB_RECLAIMABLE) > 2347 slab_reclaimable - nr_pages) 2348 ; 2349 2350 /* 2351 * Update nr_reclaimed by the number of slab pages we 2352 * reclaimed from this zone. 2353 */ 2354 sc.nr_reclaimed += slab_reclaimable - 2355 zone_page_state(zone, NR_SLAB_RECLAIMABLE); 2356 } 2357 2358 p->reclaim_state = NULL; 2359 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); 2360 return sc.nr_reclaimed >= nr_pages; 2361} 2362 2363int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 2364{ 2365 int node_id; 2366 int ret; 2367 2368 /* 2369 * Zone reclaim reclaims unmapped file backed pages and 2370 * slab pages if we are over the defined limits. 2371 * 2372 * A small portion of unmapped file backed pages is needed for 2373 * file I/O otherwise pages read by file I/O will be immediately 2374 * thrown out if the zone is overallocated. So we do not reclaim 2375 * if less than a specified percentage of the zone is used by 2376 * unmapped file backed pages. 2377 */ 2378 if (zone_page_state(zone, NR_FILE_PAGES) - 2379 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages 2380 && zone_page_state(zone, NR_SLAB_RECLAIMABLE) 2381 <= zone->min_slab_pages) 2382 return 0; 2383 2384 if (zone_is_all_unreclaimable(zone)) 2385 return 0; 2386 2387 /* 2388 * Do not scan if the allocation should not be delayed. 2389 */ 2390 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) 2391 return 0; 2392 2393 /* 2394 * Only run zone reclaim on the local zone or on zones that do not 2395 * have associated processors. This will favor the local processor 2396 * over remote processors and spread off node memory allocations 2397 * as wide as possible. 2398 */ 2399 node_id = zone_to_nid(zone); 2400 if (node_state(node_id, N_CPU) && node_id != numa_node_id()) 2401 return 0; 2402 2403 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) 2404 return 0; 2405 ret = __zone_reclaim(zone, gfp_mask, order); 2406 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); 2407 2408 return ret; 2409} 2410#endif 2411 2412#ifdef CONFIG_UNEVICTABLE_LRU 2413/* 2414 * page_evictable - test whether a page is evictable 2415 * @page: the page to test 2416 * @vma: the VMA in which the page is or will be mapped, may be NULL 2417 * 2418 * Test whether page is evictable--i.e., should be placed on active/inactive 2419 * lists vs unevictable list. The vma argument is !NULL when called from the 2420 * fault path to determine how to instantate a new page. 2421 * 2422 * Reasons page might not be evictable: 2423 * (1) page's mapping marked unevictable 2424 * (2) page is part of an mlocked VMA 2425 * 2426 */ 2427int page_evictable(struct page *page, struct vm_area_struct *vma) 2428{ 2429 2430 if (mapping_unevictable(page_mapping(page))) 2431 return 0; 2432 2433 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page))) 2434 return 0; 2435 2436 return 1; 2437} 2438 2439/** 2440 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list 2441 * @page: page to check evictability and move to appropriate lru list 2442 * @zone: zone page is in 2443 * 2444 * Checks a page for evictability and moves the page to the appropriate 2445 * zone lru list. 2446 * 2447 * Restrictions: zone->lru_lock must be held, page must be on LRU and must 2448 * have PageUnevictable set. 2449 */ 2450static void check_move_unevictable_page(struct page *page, struct zone *zone) 2451{ 2452 VM_BUG_ON(PageActive(page)); 2453 2454retry: 2455 ClearPageUnevictable(page); 2456 if (page_evictable(page, NULL)) { 2457 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page); 2458 2459 __dec_zone_state(zone, NR_UNEVICTABLE); 2460 list_move(&page->lru, &zone->lru[l].list); 2461 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l); 2462 __inc_zone_state(zone, NR_INACTIVE_ANON + l); 2463 __count_vm_event(UNEVICTABLE_PGRESCUED); 2464 } else { 2465 /* 2466 * rotate unevictable list 2467 */ 2468 SetPageUnevictable(page); 2469 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list); 2470 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE); 2471 if (page_evictable(page, NULL)) 2472 goto retry; 2473 } 2474} 2475 2476/** 2477 * scan_mapping_unevictable_pages - scan an address space for evictable pages 2478 * @mapping: struct address_space to scan for evictable pages 2479 * 2480 * Scan all pages in mapping. Check unevictable pages for 2481 * evictability and move them to the appropriate zone lru list. 2482 */ 2483void scan_mapping_unevictable_pages(struct address_space *mapping) 2484{ 2485 pgoff_t next = 0; 2486 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >> 2487 PAGE_CACHE_SHIFT; 2488 struct zone *zone; 2489 struct pagevec pvec; 2490 2491 if (mapping->nrpages == 0) 2492 return; 2493 2494 pagevec_init(&pvec, 0); 2495 while (next < end && 2496 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) { 2497 int i; 2498 int pg_scanned = 0; 2499 2500 zone = NULL; 2501 2502 for (i = 0; i < pagevec_count(&pvec); i++) { 2503 struct page *page = pvec.pages[i]; 2504 pgoff_t page_index = page->index; 2505 struct zone *pagezone = page_zone(page); 2506 2507 pg_scanned++; 2508 if (page_index > next) 2509 next = page_index; 2510 next++; 2511 2512 if (pagezone != zone) { 2513 if (zone) 2514 spin_unlock_irq(&zone->lru_lock); 2515 zone = pagezone; 2516 spin_lock_irq(&zone->lru_lock); 2517 } 2518 2519 if (PageLRU(page) && PageUnevictable(page)) 2520 check_move_unevictable_page(page, zone); 2521 } 2522 if (zone) 2523 spin_unlock_irq(&zone->lru_lock); 2524 pagevec_release(&pvec); 2525 2526 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned); 2527 } 2528 2529} 2530 2531/** 2532 * scan_zone_unevictable_pages - check unevictable list for evictable pages 2533 * @zone - zone of which to scan the unevictable list 2534 * 2535 * Scan @zone's unevictable LRU lists to check for pages that have become 2536 * evictable. Move those that have to @zone's inactive list where they 2537 * become candidates for reclaim, unless shrink_inactive_zone() decides 2538 * to reactivate them. Pages that are still unevictable are rotated 2539 * back onto @zone's unevictable list. 2540 */ 2541#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */ 2542static void scan_zone_unevictable_pages(struct zone *zone) 2543{ 2544 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list; 2545 unsigned long scan; 2546 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE); 2547 2548 while (nr_to_scan > 0) { 2549 unsigned long batch_size = min(nr_to_scan, 2550 SCAN_UNEVICTABLE_BATCH_SIZE); 2551 2552 spin_lock_irq(&zone->lru_lock); 2553 for (scan = 0; scan < batch_size; scan++) { 2554 struct page *page = lru_to_page(l_unevictable); 2555 2556 if (!trylock_page(page)) 2557 continue; 2558 2559 prefetchw_prev_lru_page(page, l_unevictable, flags); 2560 2561 if (likely(PageLRU(page) && PageUnevictable(page))) 2562 check_move_unevictable_page(page, zone); 2563 2564 unlock_page(page); 2565 } 2566 spin_unlock_irq(&zone->lru_lock); 2567 2568 nr_to_scan -= batch_size; 2569 } 2570} 2571 2572 2573/** 2574 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages 2575 * 2576 * A really big hammer: scan all zones' unevictable LRU lists to check for 2577 * pages that have become evictable. Move those back to the zones' 2578 * inactive list where they become candidates for reclaim. 2579 * This occurs when, e.g., we have unswappable pages on the unevictable lists, 2580 * and we add swap to the system. As such, it runs in the context of a task 2581 * that has possibly/probably made some previously unevictable pages 2582 * evictable. 2583 */ 2584static void scan_all_zones_unevictable_pages(void) 2585{ 2586 struct zone *zone; 2587 2588 for_each_zone(zone) { 2589 scan_zone_unevictable_pages(zone); 2590 } 2591} 2592 2593/* 2594 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of 2595 * all nodes' unevictable lists for evictable pages 2596 */ 2597unsigned long scan_unevictable_pages; 2598 2599int scan_unevictable_handler(struct ctl_table *table, int write, 2600 struct file *file, void __user *buffer, 2601 size_t *length, loff_t *ppos) 2602{ 2603 proc_doulongvec_minmax(table, write, file, buffer, length, ppos); 2604 2605 if (write && *(unsigned long *)table->data) 2606 scan_all_zones_unevictable_pages(); 2607 2608 scan_unevictable_pages = 0; 2609 return 0; 2610} 2611 2612/* 2613 * per node 'scan_unevictable_pages' attribute. On demand re-scan of 2614 * a specified node's per zone unevictable lists for evictable pages. 2615 */ 2616 2617static ssize_t read_scan_unevictable_node(struct sys_device *dev, 2618 struct sysdev_attribute *attr, 2619 char *buf) 2620{ 2621 return sprintf(buf, "0\n"); /* always zero; should fit... */ 2622} 2623 2624static ssize_t write_scan_unevictable_node(struct sys_device *dev, 2625 struct sysdev_attribute *attr, 2626 const char *buf, size_t count) 2627{ 2628 struct zone *node_zones = NODE_DATA(dev->id)->node_zones; 2629 struct zone *zone; 2630 unsigned long res; 2631 unsigned long req = strict_strtoul(buf, 10, &res); 2632 2633 if (!req) 2634 return 1; /* zero is no-op */ 2635 2636 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) { 2637 if (!populated_zone(zone)) 2638 continue; 2639 scan_zone_unevictable_pages(zone); 2640 } 2641 return 1; 2642} 2643 2644 2645static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR, 2646 read_scan_unevictable_node, 2647 write_scan_unevictable_node); 2648 2649int scan_unevictable_register_node(struct node *node) 2650{ 2651 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages); 2652} 2653 2654void scan_unevictable_unregister_node(struct node *node) 2655{ 2656 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages); 2657} 2658 2659#endif 2660