vmscan.c revision e1dbeda60a7ea9e82a908d93c07308d104d50d79
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 40#include <asm/tlbflush.h> 41#include <asm/div64.h> 42 43#include <linux/swapops.h> 44 45#include "internal.h" 46 47struct scan_control { 48 /* Incremented by the number of inactive pages that were scanned */ 49 unsigned long nr_scanned; 50 51 /* This context's GFP mask */ 52 gfp_t gfp_mask; 53 54 int may_writepage; 55 56 /* Can pages be swapped as part of reclaim? */ 57 int may_swap; 58 59 /* This context's SWAP_CLUSTER_MAX. If freeing memory for 60 * suspend, we effectively ignore SWAP_CLUSTER_MAX. 61 * In this context, it doesn't matter that we scan the 62 * whole list at once. */ 63 int swap_cluster_max; 64 65 int swappiness; 66 67 int all_unreclaimable; 68}; 69 70/* 71 * The list of shrinker callbacks used by to apply pressure to 72 * ageable caches. 73 */ 74struct shrinker { 75 shrinker_t shrinker; 76 struct list_head list; 77 int seeks; /* seeks to recreate an obj */ 78 long nr; /* objs pending delete */ 79}; 80 81#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) 82 83#ifdef ARCH_HAS_PREFETCH 84#define prefetch_prev_lru_page(_page, _base, _field) \ 85 do { \ 86 if ((_page)->lru.prev != _base) { \ 87 struct page *prev; \ 88 \ 89 prev = lru_to_page(&(_page->lru)); \ 90 prefetch(&prev->_field); \ 91 } \ 92 } while (0) 93#else 94#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) 95#endif 96 97#ifdef ARCH_HAS_PREFETCHW 98#define prefetchw_prev_lru_page(_page, _base, _field) \ 99 do { \ 100 if ((_page)->lru.prev != _base) { \ 101 struct page *prev; \ 102 \ 103 prev = lru_to_page(&(_page->lru)); \ 104 prefetchw(&prev->_field); \ 105 } \ 106 } while (0) 107#else 108#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) 109#endif 110 111/* 112 * From 0 .. 100. Higher means more swappy. 113 */ 114int vm_swappiness = 60; 115long vm_total_pages; /* The total number of pages which the VM controls */ 116 117static LIST_HEAD(shrinker_list); 118static DECLARE_RWSEM(shrinker_rwsem); 119 120/* 121 * Add a shrinker callback to be called from the vm 122 */ 123struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker) 124{ 125 struct shrinker *shrinker; 126 127 shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL); 128 if (shrinker) { 129 shrinker->shrinker = theshrinker; 130 shrinker->seeks = seeks; 131 shrinker->nr = 0; 132 down_write(&shrinker_rwsem); 133 list_add_tail(&shrinker->list, &shrinker_list); 134 up_write(&shrinker_rwsem); 135 } 136 return shrinker; 137} 138EXPORT_SYMBOL(set_shrinker); 139 140/* 141 * Remove one 142 */ 143void remove_shrinker(struct shrinker *shrinker) 144{ 145 down_write(&shrinker_rwsem); 146 list_del(&shrinker->list); 147 up_write(&shrinker_rwsem); 148 kfree(shrinker); 149} 150EXPORT_SYMBOL(remove_shrinker); 151 152#define SHRINK_BATCH 128 153/* 154 * Call the shrink functions to age shrinkable caches 155 * 156 * Here we assume it costs one seek to replace a lru page and that it also 157 * takes a seek to recreate a cache object. With this in mind we age equal 158 * percentages of the lru and ageable caches. This should balance the seeks 159 * generated by these structures. 160 * 161 * If the vm encounted mapped pages on the LRU it increase the pressure on 162 * slab to avoid swapping. 163 * 164 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. 165 * 166 * `lru_pages' represents the number of on-LRU pages in all the zones which 167 * are eligible for the caller's allocation attempt. It is used for balancing 168 * slab reclaim versus page reclaim. 169 * 170 * Returns the number of slab objects which we shrunk. 171 */ 172unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask, 173 unsigned long lru_pages) 174{ 175 struct shrinker *shrinker; 176 unsigned long ret = 0; 177 178 if (scanned == 0) 179 scanned = SWAP_CLUSTER_MAX; 180 181 if (!down_read_trylock(&shrinker_rwsem)) 182 return 1; /* Assume we'll be able to shrink next time */ 183 184 list_for_each_entry(shrinker, &shrinker_list, list) { 185 unsigned long long delta; 186 unsigned long total_scan; 187 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask); 188 189 delta = (4 * scanned) / shrinker->seeks; 190 delta *= max_pass; 191 do_div(delta, lru_pages + 1); 192 shrinker->nr += delta; 193 if (shrinker->nr < 0) { 194 printk(KERN_ERR "%s: nr=%ld\n", 195 __FUNCTION__, shrinker->nr); 196 shrinker->nr = max_pass; 197 } 198 199 /* 200 * Avoid risking looping forever due to too large nr value: 201 * never try to free more than twice the estimate number of 202 * freeable entries. 203 */ 204 if (shrinker->nr > max_pass * 2) 205 shrinker->nr = max_pass * 2; 206 207 total_scan = shrinker->nr; 208 shrinker->nr = 0; 209 210 while (total_scan >= SHRINK_BATCH) { 211 long this_scan = SHRINK_BATCH; 212 int shrink_ret; 213 int nr_before; 214 215 nr_before = (*shrinker->shrinker)(0, gfp_mask); 216 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask); 217 if (shrink_ret == -1) 218 break; 219 if (shrink_ret < nr_before) 220 ret += nr_before - shrink_ret; 221 count_vm_events(SLABS_SCANNED, this_scan); 222 total_scan -= this_scan; 223 224 cond_resched(); 225 } 226 227 shrinker->nr += total_scan; 228 } 229 up_read(&shrinker_rwsem); 230 return ret; 231} 232 233/* Called without lock on whether page is mapped, so answer is unstable */ 234static inline int page_mapping_inuse(struct page *page) 235{ 236 struct address_space *mapping; 237 238 /* Page is in somebody's page tables. */ 239 if (page_mapped(page)) 240 return 1; 241 242 /* Be more reluctant to reclaim swapcache than pagecache */ 243 if (PageSwapCache(page)) 244 return 1; 245 246 mapping = page_mapping(page); 247 if (!mapping) 248 return 0; 249 250 /* File is mmap'd by somebody? */ 251 return mapping_mapped(mapping); 252} 253 254static inline int is_page_cache_freeable(struct page *page) 255{ 256 return page_count(page) - !!PagePrivate(page) == 2; 257} 258 259static int may_write_to_queue(struct backing_dev_info *bdi) 260{ 261 if (current->flags & PF_SWAPWRITE) 262 return 1; 263 if (!bdi_write_congested(bdi)) 264 return 1; 265 if (bdi == current->backing_dev_info) 266 return 1; 267 return 0; 268} 269 270/* 271 * We detected a synchronous write error writing a page out. Probably 272 * -ENOSPC. We need to propagate that into the address_space for a subsequent 273 * fsync(), msync() or close(). 274 * 275 * The tricky part is that after writepage we cannot touch the mapping: nothing 276 * prevents it from being freed up. But we have a ref on the page and once 277 * that page is locked, the mapping is pinned. 278 * 279 * We're allowed to run sleeping lock_page() here because we know the caller has 280 * __GFP_FS. 281 */ 282static void handle_write_error(struct address_space *mapping, 283 struct page *page, int error) 284{ 285 lock_page(page); 286 if (page_mapping(page) == mapping) { 287 if (error == -ENOSPC) 288 set_bit(AS_ENOSPC, &mapping->flags); 289 else 290 set_bit(AS_EIO, &mapping->flags); 291 } 292 unlock_page(page); 293} 294 295/* possible outcome of pageout() */ 296typedef enum { 297 /* failed to write page out, page is locked */ 298 PAGE_KEEP, 299 /* move page to the active list, page is locked */ 300 PAGE_ACTIVATE, 301 /* page has been sent to the disk successfully, page is unlocked */ 302 PAGE_SUCCESS, 303 /* page is clean and locked */ 304 PAGE_CLEAN, 305} pageout_t; 306 307/* 308 * pageout is called by shrink_page_list() for each dirty page. 309 * Calls ->writepage(). 310 */ 311static pageout_t pageout(struct page *page, struct address_space *mapping) 312{ 313 /* 314 * If the page is dirty, only perform writeback if that write 315 * will be non-blocking. To prevent this allocation from being 316 * stalled by pagecache activity. But note that there may be 317 * stalls if we need to run get_block(). We could test 318 * PagePrivate for that. 319 * 320 * If this process is currently in generic_file_write() against 321 * this page's queue, we can perform writeback even if that 322 * will block. 323 * 324 * If the page is swapcache, write it back even if that would 325 * block, for some throttling. This happens by accident, because 326 * swap_backing_dev_info is bust: it doesn't reflect the 327 * congestion state of the swapdevs. Easy to fix, if needed. 328 * See swapfile.c:page_queue_congested(). 329 */ 330 if (!is_page_cache_freeable(page)) 331 return PAGE_KEEP; 332 if (!mapping) { 333 /* 334 * Some data journaling orphaned pages can have 335 * page->mapping == NULL while being dirty with clean buffers. 336 */ 337 if (PagePrivate(page)) { 338 if (try_to_free_buffers(page)) { 339 ClearPageDirty(page); 340 printk("%s: orphaned page\n", __FUNCTION__); 341 return PAGE_CLEAN; 342 } 343 } 344 return PAGE_KEEP; 345 } 346 if (mapping->a_ops->writepage == NULL) 347 return PAGE_ACTIVATE; 348 if (!may_write_to_queue(mapping->backing_dev_info)) 349 return PAGE_KEEP; 350 351 if (clear_page_dirty_for_io(page)) { 352 int res; 353 struct writeback_control wbc = { 354 .sync_mode = WB_SYNC_NONE, 355 .nr_to_write = SWAP_CLUSTER_MAX, 356 .range_start = 0, 357 .range_end = LLONG_MAX, 358 .nonblocking = 1, 359 .for_reclaim = 1, 360 }; 361 362 SetPageReclaim(page); 363 res = mapping->a_ops->writepage(page, &wbc); 364 if (res < 0) 365 handle_write_error(mapping, page, res); 366 if (res == AOP_WRITEPAGE_ACTIVATE) { 367 ClearPageReclaim(page); 368 return PAGE_ACTIVATE; 369 } 370 if (!PageWriteback(page)) { 371 /* synchronous write or broken a_ops? */ 372 ClearPageReclaim(page); 373 } 374 inc_zone_page_state(page, NR_VMSCAN_WRITE); 375 return PAGE_SUCCESS; 376 } 377 378 return PAGE_CLEAN; 379} 380 381/* 382 * Attempt to detach a locked page from its ->mapping. If it is dirty or if 383 * someone else has a ref on the page, abort and return 0. If it was 384 * successfully detached, return 1. Assumes the caller has a single ref on 385 * this page. 386 */ 387int remove_mapping(struct address_space *mapping, struct page *page) 388{ 389 BUG_ON(!PageLocked(page)); 390 BUG_ON(mapping != page_mapping(page)); 391 392 write_lock_irq(&mapping->tree_lock); 393 /* 394 * The non racy check for a busy page. 395 * 396 * Must be careful with the order of the tests. When someone has 397 * a ref to the page, it may be possible that they dirty it then 398 * drop the reference. So if PageDirty is tested before page_count 399 * here, then the following race may occur: 400 * 401 * get_user_pages(&page); 402 * [user mapping goes away] 403 * write_to(page); 404 * !PageDirty(page) [good] 405 * SetPageDirty(page); 406 * put_page(page); 407 * !page_count(page) [good, discard it] 408 * 409 * [oops, our write_to data is lost] 410 * 411 * Reversing the order of the tests ensures such a situation cannot 412 * escape unnoticed. The smp_rmb is needed to ensure the page->flags 413 * load is not satisfied before that of page->_count. 414 * 415 * Note that if SetPageDirty is always performed via set_page_dirty, 416 * and thus under tree_lock, then this ordering is not required. 417 */ 418 if (unlikely(page_count(page) != 2)) 419 goto cannot_free; 420 smp_rmb(); 421 if (unlikely(PageDirty(page))) 422 goto cannot_free; 423 424 if (PageSwapCache(page)) { 425 swp_entry_t swap = { .val = page_private(page) }; 426 __delete_from_swap_cache(page); 427 write_unlock_irq(&mapping->tree_lock); 428 swap_free(swap); 429 __put_page(page); /* The pagecache ref */ 430 return 1; 431 } 432 433 __remove_from_page_cache(page); 434 write_unlock_irq(&mapping->tree_lock); 435 __put_page(page); 436 return 1; 437 438cannot_free: 439 write_unlock_irq(&mapping->tree_lock); 440 return 0; 441} 442 443/* 444 * shrink_page_list() returns the number of reclaimed pages 445 */ 446static unsigned long shrink_page_list(struct list_head *page_list, 447 struct scan_control *sc) 448{ 449 LIST_HEAD(ret_pages); 450 struct pagevec freed_pvec; 451 int pgactivate = 0; 452 unsigned long nr_reclaimed = 0; 453 454 cond_resched(); 455 456 pagevec_init(&freed_pvec, 1); 457 while (!list_empty(page_list)) { 458 struct address_space *mapping; 459 struct page *page; 460 int may_enter_fs; 461 int referenced; 462 463 cond_resched(); 464 465 page = lru_to_page(page_list); 466 list_del(&page->lru); 467 468 if (TestSetPageLocked(page)) 469 goto keep; 470 471 VM_BUG_ON(PageActive(page)); 472 473 sc->nr_scanned++; 474 475 if (!sc->may_swap && page_mapped(page)) 476 goto keep_locked; 477 478 /* Double the slab pressure for mapped and swapcache pages */ 479 if (page_mapped(page) || PageSwapCache(page)) 480 sc->nr_scanned++; 481 482 if (PageWriteback(page)) 483 goto keep_locked; 484 485 referenced = page_referenced(page, 1); 486 /* In active use or really unfreeable? Activate it. */ 487 if (referenced && page_mapping_inuse(page)) 488 goto activate_locked; 489 490#ifdef CONFIG_SWAP 491 /* 492 * Anonymous process memory has backing store? 493 * Try to allocate it some swap space here. 494 */ 495 if (PageAnon(page) && !PageSwapCache(page)) 496 if (!add_to_swap(page, GFP_ATOMIC)) 497 goto activate_locked; 498#endif /* CONFIG_SWAP */ 499 500 mapping = page_mapping(page); 501 may_enter_fs = (sc->gfp_mask & __GFP_FS) || 502 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); 503 504 /* 505 * The page is mapped into the page tables of one or more 506 * processes. Try to unmap it here. 507 */ 508 if (page_mapped(page) && mapping) { 509 switch (try_to_unmap(page, 0)) { 510 case SWAP_FAIL: 511 goto activate_locked; 512 case SWAP_AGAIN: 513 goto keep_locked; 514 case SWAP_SUCCESS: 515 ; /* try to free the page below */ 516 } 517 } 518 519 if (PageDirty(page)) { 520 if (referenced) 521 goto keep_locked; 522 if (!may_enter_fs) 523 goto keep_locked; 524 if (!sc->may_writepage) 525 goto keep_locked; 526 527 /* Page is dirty, try to write it out here */ 528 switch(pageout(page, mapping)) { 529 case PAGE_KEEP: 530 goto keep_locked; 531 case PAGE_ACTIVATE: 532 goto activate_locked; 533 case PAGE_SUCCESS: 534 if (PageWriteback(page) || PageDirty(page)) 535 goto keep; 536 /* 537 * A synchronous write - probably a ramdisk. Go 538 * ahead and try to reclaim the page. 539 */ 540 if (TestSetPageLocked(page)) 541 goto keep; 542 if (PageDirty(page) || PageWriteback(page)) 543 goto keep_locked; 544 mapping = page_mapping(page); 545 case PAGE_CLEAN: 546 ; /* try to free the page below */ 547 } 548 } 549 550 /* 551 * If the page has buffers, try to free the buffer mappings 552 * associated with this page. If we succeed we try to free 553 * the page as well. 554 * 555 * We do this even if the page is PageDirty(). 556 * try_to_release_page() does not perform I/O, but it is 557 * possible for a page to have PageDirty set, but it is actually 558 * clean (all its buffers are clean). This happens if the 559 * buffers were written out directly, with submit_bh(). ext3 560 * will do this, as well as the blockdev mapping. 561 * try_to_release_page() will discover that cleanness and will 562 * drop the buffers and mark the page clean - it can be freed. 563 * 564 * Rarely, pages can have buffers and no ->mapping. These are 565 * the pages which were not successfully invalidated in 566 * truncate_complete_page(). We try to drop those buffers here 567 * and if that worked, and the page is no longer mapped into 568 * process address space (page_count == 1) it can be freed. 569 * Otherwise, leave the page on the LRU so it is swappable. 570 */ 571 if (PagePrivate(page)) { 572 if (!try_to_release_page(page, sc->gfp_mask)) 573 goto activate_locked; 574 if (!mapping && page_count(page) == 1) 575 goto free_it; 576 } 577 578 if (!mapping || !remove_mapping(mapping, page)) 579 goto keep_locked; 580 581free_it: 582 unlock_page(page); 583 nr_reclaimed++; 584 if (!pagevec_add(&freed_pvec, page)) 585 __pagevec_release_nonlru(&freed_pvec); 586 continue; 587 588activate_locked: 589 SetPageActive(page); 590 pgactivate++; 591keep_locked: 592 unlock_page(page); 593keep: 594 list_add(&page->lru, &ret_pages); 595 VM_BUG_ON(PageLRU(page)); 596 } 597 list_splice(&ret_pages, page_list); 598 if (pagevec_count(&freed_pvec)) 599 __pagevec_release_nonlru(&freed_pvec); 600 count_vm_events(PGACTIVATE, pgactivate); 601 return nr_reclaimed; 602} 603 604/* 605 * zone->lru_lock is heavily contended. Some of the functions that 606 * shrink the lists perform better by taking out a batch of pages 607 * and working on them outside the LRU lock. 608 * 609 * For pagecache intensive workloads, this function is the hottest 610 * spot in the kernel (apart from copy_*_user functions). 611 * 612 * Appropriate locks must be held before calling this function. 613 * 614 * @nr_to_scan: The number of pages to look through on the list. 615 * @src: The LRU list to pull pages off. 616 * @dst: The temp list to put pages on to. 617 * @scanned: The number of pages that were scanned. 618 * 619 * returns how many pages were moved onto *@dst. 620 */ 621static unsigned long isolate_lru_pages(unsigned long nr_to_scan, 622 struct list_head *src, struct list_head *dst, 623 unsigned long *scanned) 624{ 625 unsigned long nr_taken = 0; 626 struct page *page; 627 unsigned long scan; 628 629 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { 630 struct list_head *target; 631 page = lru_to_page(src); 632 prefetchw_prev_lru_page(page, src, flags); 633 634 VM_BUG_ON(!PageLRU(page)); 635 636 list_del(&page->lru); 637 target = src; 638 if (likely(get_page_unless_zero(page))) { 639 /* 640 * Be careful not to clear PageLRU until after we're 641 * sure the page is not being freed elsewhere -- the 642 * page release code relies on it. 643 */ 644 ClearPageLRU(page); 645 target = dst; 646 nr_taken++; 647 } /* else it is being freed elsewhere */ 648 649 list_add(&page->lru, target); 650 } 651 652 *scanned = scan; 653 return nr_taken; 654} 655 656/* 657 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number 658 * of reclaimed pages 659 */ 660static unsigned long shrink_inactive_list(unsigned long max_scan, 661 struct zone *zone, struct scan_control *sc) 662{ 663 LIST_HEAD(page_list); 664 struct pagevec pvec; 665 unsigned long nr_scanned = 0; 666 unsigned long nr_reclaimed = 0; 667 668 pagevec_init(&pvec, 1); 669 670 lru_add_drain(); 671 spin_lock_irq(&zone->lru_lock); 672 do { 673 struct page *page; 674 unsigned long nr_taken; 675 unsigned long nr_scan; 676 unsigned long nr_freed; 677 678 nr_taken = isolate_lru_pages(sc->swap_cluster_max, 679 &zone->inactive_list, 680 &page_list, &nr_scan); 681 zone->nr_inactive -= nr_taken; 682 zone->pages_scanned += nr_scan; 683 spin_unlock_irq(&zone->lru_lock); 684 685 nr_scanned += nr_scan; 686 nr_freed = shrink_page_list(&page_list, sc); 687 nr_reclaimed += nr_freed; 688 local_irq_disable(); 689 if (current_is_kswapd()) { 690 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan); 691 __count_vm_events(KSWAPD_STEAL, nr_freed); 692 } else 693 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan); 694 __count_vm_events(PGACTIVATE, nr_freed); 695 696 if (nr_taken == 0) 697 goto done; 698 699 spin_lock(&zone->lru_lock); 700 /* 701 * Put back any unfreeable pages. 702 */ 703 while (!list_empty(&page_list)) { 704 page = lru_to_page(&page_list); 705 VM_BUG_ON(PageLRU(page)); 706 SetPageLRU(page); 707 list_del(&page->lru); 708 if (PageActive(page)) 709 add_page_to_active_list(zone, page); 710 else 711 add_page_to_inactive_list(zone, page); 712 if (!pagevec_add(&pvec, page)) { 713 spin_unlock_irq(&zone->lru_lock); 714 __pagevec_release(&pvec); 715 spin_lock_irq(&zone->lru_lock); 716 } 717 } 718 } while (nr_scanned < max_scan); 719 spin_unlock(&zone->lru_lock); 720done: 721 local_irq_enable(); 722 pagevec_release(&pvec); 723 return nr_reclaimed; 724} 725 726/* 727 * We are about to scan this zone at a certain priority level. If that priority 728 * level is smaller (ie: more urgent) than the previous priority, then note 729 * that priority level within the zone. This is done so that when the next 730 * process comes in to scan this zone, it will immediately start out at this 731 * priority level rather than having to build up its own scanning priority. 732 * Here, this priority affects only the reclaim-mapped threshold. 733 */ 734static inline void note_zone_scanning_priority(struct zone *zone, int priority) 735{ 736 if (priority < zone->prev_priority) 737 zone->prev_priority = priority; 738} 739 740static inline int zone_is_near_oom(struct zone *zone) 741{ 742 return zone->pages_scanned >= (zone->nr_active + zone->nr_inactive)*3; 743} 744 745/* 746 * This moves pages from the active list to the inactive list. 747 * 748 * We move them the other way if the page is referenced by one or more 749 * processes, from rmap. 750 * 751 * If the pages are mostly unmapped, the processing is fast and it is 752 * appropriate to hold zone->lru_lock across the whole operation. But if 753 * the pages are mapped, the processing is slow (page_referenced()) so we 754 * should drop zone->lru_lock around each page. It's impossible to balance 755 * this, so instead we remove the pages from the LRU while processing them. 756 * It is safe to rely on PG_active against the non-LRU pages in here because 757 * nobody will play with that bit on a non-LRU page. 758 * 759 * The downside is that we have to touch page->_count against each page. 760 * But we had to alter page->flags anyway. 761 */ 762static void shrink_active_list(unsigned long nr_pages, struct zone *zone, 763 struct scan_control *sc, int priority) 764{ 765 unsigned long pgmoved; 766 int pgdeactivate = 0; 767 unsigned long pgscanned; 768 LIST_HEAD(l_hold); /* The pages which were snipped off */ 769 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */ 770 LIST_HEAD(l_active); /* Pages to go onto the active_list */ 771 struct page *page; 772 struct pagevec pvec; 773 int reclaim_mapped = 0; 774 775 if (sc->may_swap) { 776 long mapped_ratio; 777 long distress; 778 long swap_tendency; 779 780 if (zone_is_near_oom(zone)) 781 goto force_reclaim_mapped; 782 783 /* 784 * `distress' is a measure of how much trouble we're having 785 * reclaiming pages. 0 -> no problems. 100 -> great trouble. 786 */ 787 distress = 100 >> min(zone->prev_priority, priority); 788 789 /* 790 * The point of this algorithm is to decide when to start 791 * reclaiming mapped memory instead of just pagecache. Work out 792 * how much memory 793 * is mapped. 794 */ 795 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) + 796 global_page_state(NR_ANON_PAGES)) * 100) / 797 vm_total_pages; 798 799 /* 800 * Now decide how much we really want to unmap some pages. The 801 * mapped ratio is downgraded - just because there's a lot of 802 * mapped memory doesn't necessarily mean that page reclaim 803 * isn't succeeding. 804 * 805 * The distress ratio is important - we don't want to start 806 * going oom. 807 * 808 * A 100% value of vm_swappiness overrides this algorithm 809 * altogether. 810 */ 811 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness; 812 813 /* 814 * Now use this metric to decide whether to start moving mapped 815 * memory onto the inactive list. 816 */ 817 if (swap_tendency >= 100) 818force_reclaim_mapped: 819 reclaim_mapped = 1; 820 } 821 822 lru_add_drain(); 823 spin_lock_irq(&zone->lru_lock); 824 pgmoved = isolate_lru_pages(nr_pages, &zone->active_list, 825 &l_hold, &pgscanned); 826 zone->pages_scanned += pgscanned; 827 zone->nr_active -= pgmoved; 828 spin_unlock_irq(&zone->lru_lock); 829 830 while (!list_empty(&l_hold)) { 831 cond_resched(); 832 page = lru_to_page(&l_hold); 833 list_del(&page->lru); 834 if (page_mapped(page)) { 835 if (!reclaim_mapped || 836 (total_swap_pages == 0 && PageAnon(page)) || 837 page_referenced(page, 0)) { 838 list_add(&page->lru, &l_active); 839 continue; 840 } 841 } 842 list_add(&page->lru, &l_inactive); 843 } 844 845 pagevec_init(&pvec, 1); 846 pgmoved = 0; 847 spin_lock_irq(&zone->lru_lock); 848 while (!list_empty(&l_inactive)) { 849 page = lru_to_page(&l_inactive); 850 prefetchw_prev_lru_page(page, &l_inactive, flags); 851 VM_BUG_ON(PageLRU(page)); 852 SetPageLRU(page); 853 VM_BUG_ON(!PageActive(page)); 854 ClearPageActive(page); 855 856 list_move(&page->lru, &zone->inactive_list); 857 pgmoved++; 858 if (!pagevec_add(&pvec, page)) { 859 zone->nr_inactive += pgmoved; 860 spin_unlock_irq(&zone->lru_lock); 861 pgdeactivate += pgmoved; 862 pgmoved = 0; 863 if (buffer_heads_over_limit) 864 pagevec_strip(&pvec); 865 __pagevec_release(&pvec); 866 spin_lock_irq(&zone->lru_lock); 867 } 868 } 869 zone->nr_inactive += pgmoved; 870 pgdeactivate += pgmoved; 871 if (buffer_heads_over_limit) { 872 spin_unlock_irq(&zone->lru_lock); 873 pagevec_strip(&pvec); 874 spin_lock_irq(&zone->lru_lock); 875 } 876 877 pgmoved = 0; 878 while (!list_empty(&l_active)) { 879 page = lru_to_page(&l_active); 880 prefetchw_prev_lru_page(page, &l_active, flags); 881 VM_BUG_ON(PageLRU(page)); 882 SetPageLRU(page); 883 VM_BUG_ON(!PageActive(page)); 884 list_move(&page->lru, &zone->active_list); 885 pgmoved++; 886 if (!pagevec_add(&pvec, page)) { 887 zone->nr_active += pgmoved; 888 pgmoved = 0; 889 spin_unlock_irq(&zone->lru_lock); 890 __pagevec_release(&pvec); 891 spin_lock_irq(&zone->lru_lock); 892 } 893 } 894 zone->nr_active += pgmoved; 895 896 __count_zone_vm_events(PGREFILL, zone, pgscanned); 897 __count_vm_events(PGDEACTIVATE, pgdeactivate); 898 spin_unlock_irq(&zone->lru_lock); 899 900 pagevec_release(&pvec); 901} 902 903/* 904 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. 905 */ 906static unsigned long shrink_zone(int priority, struct zone *zone, 907 struct scan_control *sc) 908{ 909 unsigned long nr_active; 910 unsigned long nr_inactive; 911 unsigned long nr_to_scan; 912 unsigned long nr_reclaimed = 0; 913 914 atomic_inc(&zone->reclaim_in_progress); 915 916 /* 917 * Add one to `nr_to_scan' just to make sure that the kernel will 918 * slowly sift through the active list. 919 */ 920 zone->nr_scan_active += (zone->nr_active >> priority) + 1; 921 nr_active = zone->nr_scan_active; 922 if (nr_active >= sc->swap_cluster_max) 923 zone->nr_scan_active = 0; 924 else 925 nr_active = 0; 926 927 zone->nr_scan_inactive += (zone->nr_inactive >> priority) + 1; 928 nr_inactive = zone->nr_scan_inactive; 929 if (nr_inactive >= sc->swap_cluster_max) 930 zone->nr_scan_inactive = 0; 931 else 932 nr_inactive = 0; 933 934 while (nr_active || nr_inactive) { 935 if (nr_active) { 936 nr_to_scan = min(nr_active, 937 (unsigned long)sc->swap_cluster_max); 938 nr_active -= nr_to_scan; 939 shrink_active_list(nr_to_scan, zone, sc, priority); 940 } 941 942 if (nr_inactive) { 943 nr_to_scan = min(nr_inactive, 944 (unsigned long)sc->swap_cluster_max); 945 nr_inactive -= nr_to_scan; 946 nr_reclaimed += shrink_inactive_list(nr_to_scan, zone, 947 sc); 948 } 949 } 950 951 throttle_vm_writeout(); 952 953 atomic_dec(&zone->reclaim_in_progress); 954 return nr_reclaimed; 955} 956 957/* 958 * This is the direct reclaim path, for page-allocating processes. We only 959 * try to reclaim pages from zones which will satisfy the caller's allocation 960 * request. 961 * 962 * We reclaim from a zone even if that zone is over pages_high. Because: 963 * a) The caller may be trying to free *extra* pages to satisfy a higher-order 964 * allocation or 965 * b) The zones may be over pages_high but they must go *over* pages_high to 966 * satisfy the `incremental min' zone defense algorithm. 967 * 968 * Returns the number of reclaimed pages. 969 * 970 * If a zone is deemed to be full of pinned pages then just give it a light 971 * scan then give up on it. 972 */ 973static unsigned long shrink_zones(int priority, struct zone **zones, 974 struct scan_control *sc) 975{ 976 unsigned long nr_reclaimed = 0; 977 int i; 978 979 sc->all_unreclaimable = 1; 980 for (i = 0; zones[i] != NULL; i++) { 981 struct zone *zone = zones[i]; 982 983 if (!populated_zone(zone)) 984 continue; 985 986 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL)) 987 continue; 988 989 note_zone_scanning_priority(zone, priority); 990 991 if (zone->all_unreclaimable && priority != DEF_PRIORITY) 992 continue; /* Let kswapd poll it */ 993 994 sc->all_unreclaimable = 0; 995 996 nr_reclaimed += shrink_zone(priority, zone, sc); 997 } 998 return nr_reclaimed; 999} 1000 1001/* 1002 * This is the main entry point to direct page reclaim. 1003 * 1004 * If a full scan of the inactive list fails to free enough memory then we 1005 * are "out of memory" and something needs to be killed. 1006 * 1007 * If the caller is !__GFP_FS then the probability of a failure is reasonably 1008 * high - the zone may be full of dirty or under-writeback pages, which this 1009 * caller can't do much about. We kick pdflush and take explicit naps in the 1010 * hope that some of these pages can be written. But if the allocating task 1011 * holds filesystem locks which prevent writeout this might not work, and the 1012 * allocation attempt will fail. 1013 */ 1014unsigned long try_to_free_pages(struct zone **zones, gfp_t gfp_mask) 1015{ 1016 int priority; 1017 int ret = 0; 1018 unsigned long total_scanned = 0; 1019 unsigned long nr_reclaimed = 0; 1020 struct reclaim_state *reclaim_state = current->reclaim_state; 1021 unsigned long lru_pages = 0; 1022 int i; 1023 struct scan_control sc = { 1024 .gfp_mask = gfp_mask, 1025 .may_writepage = !laptop_mode, 1026 .swap_cluster_max = SWAP_CLUSTER_MAX, 1027 .may_swap = 1, 1028 .swappiness = vm_swappiness, 1029 }; 1030 1031 count_vm_event(ALLOCSTALL); 1032 1033 for (i = 0; zones[i] != NULL; i++) { 1034 struct zone *zone = zones[i]; 1035 1036 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL)) 1037 continue; 1038 1039 lru_pages += zone->nr_active + zone->nr_inactive; 1040 } 1041 1042 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 1043 sc.nr_scanned = 0; 1044 if (!priority) 1045 disable_swap_token(); 1046 nr_reclaimed += shrink_zones(priority, zones, &sc); 1047 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages); 1048 if (reclaim_state) { 1049 nr_reclaimed += reclaim_state->reclaimed_slab; 1050 reclaim_state->reclaimed_slab = 0; 1051 } 1052 total_scanned += sc.nr_scanned; 1053 if (nr_reclaimed >= sc.swap_cluster_max) { 1054 ret = 1; 1055 goto out; 1056 } 1057 1058 /* 1059 * Try to write back as many pages as we just scanned. This 1060 * tends to cause slow streaming writers to write data to the 1061 * disk smoothly, at the dirtying rate, which is nice. But 1062 * that's undesirable in laptop mode, where we *want* lumpy 1063 * writeout. So in laptop mode, write out the whole world. 1064 */ 1065 if (total_scanned > sc.swap_cluster_max + 1066 sc.swap_cluster_max / 2) { 1067 wakeup_pdflush(laptop_mode ? 0 : total_scanned); 1068 sc.may_writepage = 1; 1069 } 1070 1071 /* Take a nap, wait for some writeback to complete */ 1072 if (sc.nr_scanned && priority < DEF_PRIORITY - 2) 1073 congestion_wait(WRITE, HZ/10); 1074 } 1075 /* top priority shrink_caches still had more to do? don't OOM, then */ 1076 if (!sc.all_unreclaimable) 1077 ret = 1; 1078out: 1079 /* 1080 * Now that we've scanned all the zones at this priority level, note 1081 * that level within the zone so that the next thread which performs 1082 * scanning of this zone will immediately start out at this priority 1083 * level. This affects only the decision whether or not to bring 1084 * mapped pages onto the inactive list. 1085 */ 1086 if (priority < 0) 1087 priority = 0; 1088 for (i = 0; zones[i] != 0; i++) { 1089 struct zone *zone = zones[i]; 1090 1091 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL)) 1092 continue; 1093 1094 zone->prev_priority = priority; 1095 } 1096 return ret; 1097} 1098 1099/* 1100 * For kswapd, balance_pgdat() will work across all this node's zones until 1101 * they are all at pages_high. 1102 * 1103 * Returns the number of pages which were actually freed. 1104 * 1105 * There is special handling here for zones which are full of pinned pages. 1106 * This can happen if the pages are all mlocked, or if they are all used by 1107 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. 1108 * What we do is to detect the case where all pages in the zone have been 1109 * scanned twice and there has been zero successful reclaim. Mark the zone as 1110 * dead and from now on, only perform a short scan. Basically we're polling 1111 * the zone for when the problem goes away. 1112 * 1113 * kswapd scans the zones in the highmem->normal->dma direction. It skips 1114 * zones which have free_pages > pages_high, but once a zone is found to have 1115 * free_pages <= pages_high, we scan that zone and the lower zones regardless 1116 * of the number of free pages in the lower zones. This interoperates with 1117 * the page allocator fallback scheme to ensure that aging of pages is balanced 1118 * across the zones. 1119 */ 1120static unsigned long balance_pgdat(pg_data_t *pgdat, int order) 1121{ 1122 int all_zones_ok; 1123 int priority; 1124 int i; 1125 unsigned long total_scanned; 1126 unsigned long nr_reclaimed; 1127 struct reclaim_state *reclaim_state = current->reclaim_state; 1128 struct scan_control sc = { 1129 .gfp_mask = GFP_KERNEL, 1130 .may_swap = 1, 1131 .swap_cluster_max = SWAP_CLUSTER_MAX, 1132 .swappiness = vm_swappiness, 1133 }; 1134 /* 1135 * temp_priority is used to remember the scanning priority at which 1136 * this zone was successfully refilled to free_pages == pages_high. 1137 */ 1138 int temp_priority[MAX_NR_ZONES]; 1139 1140loop_again: 1141 total_scanned = 0; 1142 nr_reclaimed = 0; 1143 sc.may_writepage = !laptop_mode; 1144 count_vm_event(PAGEOUTRUN); 1145 1146 for (i = 0; i < pgdat->nr_zones; i++) 1147 temp_priority[i] = DEF_PRIORITY; 1148 1149 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 1150 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ 1151 unsigned long lru_pages = 0; 1152 1153 /* The swap token gets in the way of swapout... */ 1154 if (!priority) 1155 disable_swap_token(); 1156 1157 all_zones_ok = 1; 1158 1159 /* 1160 * Scan in the highmem->dma direction for the highest 1161 * zone which needs scanning 1162 */ 1163 for (i = pgdat->nr_zones - 1; i >= 0; i--) { 1164 struct zone *zone = pgdat->node_zones + i; 1165 1166 if (!populated_zone(zone)) 1167 continue; 1168 1169 if (zone->all_unreclaimable && priority != DEF_PRIORITY) 1170 continue; 1171 1172 if (!zone_watermark_ok(zone, order, zone->pages_high, 1173 0, 0)) { 1174 end_zone = i; 1175 break; 1176 } 1177 } 1178 if (i < 0) 1179 goto out; 1180 1181 for (i = 0; i <= end_zone; i++) { 1182 struct zone *zone = pgdat->node_zones + i; 1183 1184 lru_pages += zone->nr_active + zone->nr_inactive; 1185 } 1186 1187 /* 1188 * Now scan the zone in the dma->highmem direction, stopping 1189 * at the last zone which needs scanning. 1190 * 1191 * We do this because the page allocator works in the opposite 1192 * direction. This prevents the page allocator from allocating 1193 * pages behind kswapd's direction of progress, which would 1194 * cause too much scanning of the lower zones. 1195 */ 1196 for (i = 0; i <= end_zone; i++) { 1197 struct zone *zone = pgdat->node_zones + i; 1198 int nr_slab; 1199 1200 if (!populated_zone(zone)) 1201 continue; 1202 1203 if (zone->all_unreclaimable && priority != DEF_PRIORITY) 1204 continue; 1205 1206 if (!zone_watermark_ok(zone, order, zone->pages_high, 1207 end_zone, 0)) 1208 all_zones_ok = 0; 1209 temp_priority[i] = priority; 1210 sc.nr_scanned = 0; 1211 note_zone_scanning_priority(zone, priority); 1212 nr_reclaimed += shrink_zone(priority, zone, &sc); 1213 reclaim_state->reclaimed_slab = 0; 1214 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL, 1215 lru_pages); 1216 nr_reclaimed += reclaim_state->reclaimed_slab; 1217 total_scanned += sc.nr_scanned; 1218 if (zone->all_unreclaimable) 1219 continue; 1220 if (nr_slab == 0 && zone->pages_scanned >= 1221 (zone->nr_active + zone->nr_inactive) * 6) 1222 zone->all_unreclaimable = 1; 1223 /* 1224 * If we've done a decent amount of scanning and 1225 * the reclaim ratio is low, start doing writepage 1226 * even in laptop mode 1227 */ 1228 if (total_scanned > SWAP_CLUSTER_MAX * 2 && 1229 total_scanned > nr_reclaimed + nr_reclaimed / 2) 1230 sc.may_writepage = 1; 1231 } 1232 if (all_zones_ok) 1233 break; /* kswapd: all done */ 1234 /* 1235 * OK, kswapd is getting into trouble. Take a nap, then take 1236 * another pass across the zones. 1237 */ 1238 if (total_scanned && priority < DEF_PRIORITY - 2) 1239 congestion_wait(WRITE, HZ/10); 1240 1241 /* 1242 * We do this so kswapd doesn't build up large priorities for 1243 * example when it is freeing in parallel with allocators. It 1244 * matches the direct reclaim path behaviour in terms of impact 1245 * on zone->*_priority. 1246 */ 1247 if (nr_reclaimed >= SWAP_CLUSTER_MAX) 1248 break; 1249 } 1250out: 1251 /* 1252 * Note within each zone the priority level at which this zone was 1253 * brought into a happy state. So that the next thread which scans this 1254 * zone will start out at that priority level. 1255 */ 1256 for (i = 0; i < pgdat->nr_zones; i++) { 1257 struct zone *zone = pgdat->node_zones + i; 1258 1259 zone->prev_priority = temp_priority[i]; 1260 } 1261 if (!all_zones_ok) { 1262 cond_resched(); 1263 goto loop_again; 1264 } 1265 1266 return nr_reclaimed; 1267} 1268 1269/* 1270 * The background pageout daemon, started as a kernel thread 1271 * from the init process. 1272 * 1273 * This basically trickles out pages so that we have _some_ 1274 * free memory available even if there is no other activity 1275 * that frees anything up. This is needed for things like routing 1276 * etc, where we otherwise might have all activity going on in 1277 * asynchronous contexts that cannot page things out. 1278 * 1279 * If there are applications that are active memory-allocators 1280 * (most normal use), this basically shouldn't matter. 1281 */ 1282static int kswapd(void *p) 1283{ 1284 unsigned long order; 1285 pg_data_t *pgdat = (pg_data_t*)p; 1286 struct task_struct *tsk = current; 1287 DEFINE_WAIT(wait); 1288 struct reclaim_state reclaim_state = { 1289 .reclaimed_slab = 0, 1290 }; 1291 cpumask_t cpumask; 1292 1293 cpumask = node_to_cpumask(pgdat->node_id); 1294 if (!cpus_empty(cpumask)) 1295 set_cpus_allowed(tsk, cpumask); 1296 current->reclaim_state = &reclaim_state; 1297 1298 /* 1299 * Tell the memory management that we're a "memory allocator", 1300 * and that if we need more memory we should get access to it 1301 * regardless (see "__alloc_pages()"). "kswapd" should 1302 * never get caught in the normal page freeing logic. 1303 * 1304 * (Kswapd normally doesn't need memory anyway, but sometimes 1305 * you need a small amount of memory in order to be able to 1306 * page out something else, and this flag essentially protects 1307 * us from recursively trying to free more memory as we're 1308 * trying to free the first piece of memory in the first place). 1309 */ 1310 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; 1311 1312 order = 0; 1313 for ( ; ; ) { 1314 unsigned long new_order; 1315 1316 try_to_freeze(); 1317 1318 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 1319 new_order = pgdat->kswapd_max_order; 1320 pgdat->kswapd_max_order = 0; 1321 if (order < new_order) { 1322 /* 1323 * Don't sleep if someone wants a larger 'order' 1324 * allocation 1325 */ 1326 order = new_order; 1327 } else { 1328 schedule(); 1329 order = pgdat->kswapd_max_order; 1330 } 1331 finish_wait(&pgdat->kswapd_wait, &wait); 1332 1333 balance_pgdat(pgdat, order); 1334 } 1335 return 0; 1336} 1337 1338/* 1339 * A zone is low on free memory, so wake its kswapd task to service it. 1340 */ 1341void wakeup_kswapd(struct zone *zone, int order) 1342{ 1343 pg_data_t *pgdat; 1344 1345 if (!populated_zone(zone)) 1346 return; 1347 1348 pgdat = zone->zone_pgdat; 1349 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0)) 1350 return; 1351 if (pgdat->kswapd_max_order < order) 1352 pgdat->kswapd_max_order = order; 1353 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL)) 1354 return; 1355 if (!waitqueue_active(&pgdat->kswapd_wait)) 1356 return; 1357 wake_up_interruptible(&pgdat->kswapd_wait); 1358} 1359 1360#ifdef CONFIG_PM 1361/* 1362 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages 1363 * from LRU lists system-wide, for given pass and priority, and returns the 1364 * number of reclaimed pages 1365 * 1366 * For pass > 3 we also try to shrink the LRU lists that contain a few pages 1367 */ 1368static unsigned long shrink_all_zones(unsigned long nr_pages, int pass, 1369 int prio, struct scan_control *sc) 1370{ 1371 struct zone *zone; 1372 unsigned long nr_to_scan, ret = 0; 1373 1374 for_each_zone(zone) { 1375 1376 if (!populated_zone(zone)) 1377 continue; 1378 1379 if (zone->all_unreclaimable && prio != DEF_PRIORITY) 1380 continue; 1381 1382 /* For pass = 0 we don't shrink the active list */ 1383 if (pass > 0) { 1384 zone->nr_scan_active += (zone->nr_active >> prio) + 1; 1385 if (zone->nr_scan_active >= nr_pages || pass > 3) { 1386 zone->nr_scan_active = 0; 1387 nr_to_scan = min(nr_pages, zone->nr_active); 1388 shrink_active_list(nr_to_scan, zone, sc, prio); 1389 } 1390 } 1391 1392 zone->nr_scan_inactive += (zone->nr_inactive >> prio) + 1; 1393 if (zone->nr_scan_inactive >= nr_pages || pass > 3) { 1394 zone->nr_scan_inactive = 0; 1395 nr_to_scan = min(nr_pages, zone->nr_inactive); 1396 ret += shrink_inactive_list(nr_to_scan, zone, sc); 1397 if (ret >= nr_pages) 1398 return ret; 1399 } 1400 } 1401 1402 return ret; 1403} 1404 1405/* 1406 * Try to free `nr_pages' of memory, system-wide, and return the number of 1407 * freed pages. 1408 * 1409 * Rather than trying to age LRUs the aim is to preserve the overall 1410 * LRU order by reclaiming preferentially 1411 * inactive > active > active referenced > active mapped 1412 */ 1413unsigned long shrink_all_memory(unsigned long nr_pages) 1414{ 1415 unsigned long lru_pages, nr_slab; 1416 unsigned long ret = 0; 1417 int pass; 1418 struct reclaim_state reclaim_state; 1419 struct zone *zone; 1420 struct scan_control sc = { 1421 .gfp_mask = GFP_KERNEL, 1422 .may_swap = 0, 1423 .swap_cluster_max = nr_pages, 1424 .may_writepage = 1, 1425 .swappiness = vm_swappiness, 1426 }; 1427 1428 current->reclaim_state = &reclaim_state; 1429 1430 lru_pages = 0; 1431 for_each_zone(zone) 1432 lru_pages += zone->nr_active + zone->nr_inactive; 1433 1434 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE); 1435 /* If slab caches are huge, it's better to hit them first */ 1436 while (nr_slab >= lru_pages) { 1437 reclaim_state.reclaimed_slab = 0; 1438 shrink_slab(nr_pages, sc.gfp_mask, lru_pages); 1439 if (!reclaim_state.reclaimed_slab) 1440 break; 1441 1442 ret += reclaim_state.reclaimed_slab; 1443 if (ret >= nr_pages) 1444 goto out; 1445 1446 nr_slab -= reclaim_state.reclaimed_slab; 1447 } 1448 1449 /* 1450 * We try to shrink LRUs in 5 passes: 1451 * 0 = Reclaim from inactive_list only 1452 * 1 = Reclaim from active list but don't reclaim mapped 1453 * 2 = 2nd pass of type 1 1454 * 3 = Reclaim mapped (normal reclaim) 1455 * 4 = 2nd pass of type 3 1456 */ 1457 for (pass = 0; pass < 5; pass++) { 1458 int prio; 1459 1460 /* Needed for shrinking slab caches later on */ 1461 if (!lru_pages) 1462 for_each_zone(zone) { 1463 lru_pages += zone->nr_active; 1464 lru_pages += zone->nr_inactive; 1465 } 1466 1467 /* Force reclaiming mapped pages in the passes #3 and #4 */ 1468 if (pass > 2) { 1469 sc.may_swap = 1; 1470 sc.swappiness = 100; 1471 } 1472 1473 for (prio = DEF_PRIORITY; prio >= 0; prio--) { 1474 unsigned long nr_to_scan = nr_pages - ret; 1475 1476 sc.nr_scanned = 0; 1477 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc); 1478 if (ret >= nr_pages) 1479 goto out; 1480 1481 reclaim_state.reclaimed_slab = 0; 1482 shrink_slab(sc.nr_scanned, sc.gfp_mask, lru_pages); 1483 ret += reclaim_state.reclaimed_slab; 1484 if (ret >= nr_pages) 1485 goto out; 1486 1487 if (sc.nr_scanned && prio < DEF_PRIORITY - 2) 1488 congestion_wait(WRITE, HZ / 10); 1489 } 1490 1491 lru_pages = 0; 1492 } 1493 1494 /* 1495 * If ret = 0, we could not shrink LRUs, but there may be something 1496 * in slab caches 1497 */ 1498 if (!ret) 1499 do { 1500 reclaim_state.reclaimed_slab = 0; 1501 shrink_slab(nr_pages, sc.gfp_mask, lru_pages); 1502 ret += reclaim_state.reclaimed_slab; 1503 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0); 1504 1505out: 1506 current->reclaim_state = NULL; 1507 1508 return ret; 1509} 1510#endif 1511 1512#ifdef CONFIG_HOTPLUG_CPU 1513/* It's optimal to keep kswapds on the same CPUs as their memory, but 1514 not required for correctness. So if the last cpu in a node goes 1515 away, we get changed to run anywhere: as the first one comes back, 1516 restore their cpu bindings. */ 1517static int __devinit cpu_callback(struct notifier_block *nfb, 1518 unsigned long action, void *hcpu) 1519{ 1520 pg_data_t *pgdat; 1521 cpumask_t mask; 1522 1523 if (action == CPU_ONLINE) { 1524 for_each_online_pgdat(pgdat) { 1525 mask = node_to_cpumask(pgdat->node_id); 1526 if (any_online_cpu(mask) != NR_CPUS) 1527 /* One of our CPUs online: restore mask */ 1528 set_cpus_allowed(pgdat->kswapd, mask); 1529 } 1530 } 1531 return NOTIFY_OK; 1532} 1533#endif /* CONFIG_HOTPLUG_CPU */ 1534 1535/* 1536 * This kswapd start function will be called by init and node-hot-add. 1537 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. 1538 */ 1539int kswapd_run(int nid) 1540{ 1541 pg_data_t *pgdat = NODE_DATA(nid); 1542 int ret = 0; 1543 1544 if (pgdat->kswapd) 1545 return 0; 1546 1547 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); 1548 if (IS_ERR(pgdat->kswapd)) { 1549 /* failure at boot is fatal */ 1550 BUG_ON(system_state == SYSTEM_BOOTING); 1551 printk("Failed to start kswapd on node %d\n",nid); 1552 ret = -1; 1553 } 1554 return ret; 1555} 1556 1557static int __init kswapd_init(void) 1558{ 1559 int nid; 1560 1561 swap_setup(); 1562 for_each_online_node(nid) 1563 kswapd_run(nid); 1564 hotcpu_notifier(cpu_callback, 0); 1565 return 0; 1566} 1567 1568module_init(kswapd_init) 1569 1570#ifdef CONFIG_NUMA 1571/* 1572 * Zone reclaim mode 1573 * 1574 * If non-zero call zone_reclaim when the number of free pages falls below 1575 * the watermarks. 1576 */ 1577int zone_reclaim_mode __read_mostly; 1578 1579#define RECLAIM_OFF 0 1580#define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */ 1581#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ 1582#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ 1583 1584/* 1585 * Priority for ZONE_RECLAIM. This determines the fraction of pages 1586 * of a node considered for each zone_reclaim. 4 scans 1/16th of 1587 * a zone. 1588 */ 1589#define ZONE_RECLAIM_PRIORITY 4 1590 1591/* 1592 * Percentage of pages in a zone that must be unmapped for zone_reclaim to 1593 * occur. 1594 */ 1595int sysctl_min_unmapped_ratio = 1; 1596 1597/* 1598 * If the number of slab pages in a zone grows beyond this percentage then 1599 * slab reclaim needs to occur. 1600 */ 1601int sysctl_min_slab_ratio = 5; 1602 1603/* 1604 * Try to free up some pages from this zone through reclaim. 1605 */ 1606static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 1607{ 1608 /* Minimum pages needed in order to stay on node */ 1609 const unsigned long nr_pages = 1 << order; 1610 struct task_struct *p = current; 1611 struct reclaim_state reclaim_state; 1612 int priority; 1613 unsigned long nr_reclaimed = 0; 1614 struct scan_control sc = { 1615 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), 1616 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP), 1617 .swap_cluster_max = max_t(unsigned long, nr_pages, 1618 SWAP_CLUSTER_MAX), 1619 .gfp_mask = gfp_mask, 1620 .swappiness = vm_swappiness, 1621 }; 1622 unsigned long slab_reclaimable; 1623 1624 disable_swap_token(); 1625 cond_resched(); 1626 /* 1627 * We need to be able to allocate from the reserves for RECLAIM_SWAP 1628 * and we also need to be able to write out pages for RECLAIM_WRITE 1629 * and RECLAIM_SWAP. 1630 */ 1631 p->flags |= PF_MEMALLOC | PF_SWAPWRITE; 1632 reclaim_state.reclaimed_slab = 0; 1633 p->reclaim_state = &reclaim_state; 1634 1635 if (zone_page_state(zone, NR_FILE_PAGES) - 1636 zone_page_state(zone, NR_FILE_MAPPED) > 1637 zone->min_unmapped_pages) { 1638 /* 1639 * Free memory by calling shrink zone with increasing 1640 * priorities until we have enough memory freed. 1641 */ 1642 priority = ZONE_RECLAIM_PRIORITY; 1643 do { 1644 note_zone_scanning_priority(zone, priority); 1645 nr_reclaimed += shrink_zone(priority, zone, &sc); 1646 priority--; 1647 } while (priority >= 0 && nr_reclaimed < nr_pages); 1648 } 1649 1650 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE); 1651 if (slab_reclaimable > zone->min_slab_pages) { 1652 /* 1653 * shrink_slab() does not currently allow us to determine how 1654 * many pages were freed in this zone. So we take the current 1655 * number of slab pages and shake the slab until it is reduced 1656 * by the same nr_pages that we used for reclaiming unmapped 1657 * pages. 1658 * 1659 * Note that shrink_slab will free memory on all zones and may 1660 * take a long time. 1661 */ 1662 while (shrink_slab(sc.nr_scanned, gfp_mask, order) && 1663 zone_page_state(zone, NR_SLAB_RECLAIMABLE) > 1664 slab_reclaimable - nr_pages) 1665 ; 1666 1667 /* 1668 * Update nr_reclaimed by the number of slab pages we 1669 * reclaimed from this zone. 1670 */ 1671 nr_reclaimed += slab_reclaimable - 1672 zone_page_state(zone, NR_SLAB_RECLAIMABLE); 1673 } 1674 1675 p->reclaim_state = NULL; 1676 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); 1677 return nr_reclaimed >= nr_pages; 1678} 1679 1680int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 1681{ 1682 cpumask_t mask; 1683 int node_id; 1684 1685 /* 1686 * Zone reclaim reclaims unmapped file backed pages and 1687 * slab pages if we are over the defined limits. 1688 * 1689 * A small portion of unmapped file backed pages is needed for 1690 * file I/O otherwise pages read by file I/O will be immediately 1691 * thrown out if the zone is overallocated. So we do not reclaim 1692 * if less than a specified percentage of the zone is used by 1693 * unmapped file backed pages. 1694 */ 1695 if (zone_page_state(zone, NR_FILE_PAGES) - 1696 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages 1697 && zone_page_state(zone, NR_SLAB_RECLAIMABLE) 1698 <= zone->min_slab_pages) 1699 return 0; 1700 1701 /* 1702 * Avoid concurrent zone reclaims, do not reclaim in a zone that does 1703 * not have reclaimable pages and if we should not delay the allocation 1704 * then do not scan. 1705 */ 1706 if (!(gfp_mask & __GFP_WAIT) || 1707 zone->all_unreclaimable || 1708 atomic_read(&zone->reclaim_in_progress) > 0 || 1709 (current->flags & PF_MEMALLOC)) 1710 return 0; 1711 1712 /* 1713 * Only run zone reclaim on the local zone or on zones that do not 1714 * have associated processors. This will favor the local processor 1715 * over remote processors and spread off node memory allocations 1716 * as wide as possible. 1717 */ 1718 node_id = zone_to_nid(zone); 1719 mask = node_to_cpumask(node_id); 1720 if (!cpus_empty(mask) && node_id != numa_node_id()) 1721 return 0; 1722 return __zone_reclaim(zone, gfp_mask, order); 1723} 1724#endif 1725