vmscan.c revision 66e1707bc34609f626e2e7b4fe7e454c9748bad5
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 42#include <asm/tlbflush.h> 43#include <asm/div64.h> 44 45#include <linux/swapops.h> 46 47#include "internal.h" 48 49struct scan_control { 50 /* Incremented by the number of inactive pages that were scanned */ 51 unsigned long nr_scanned; 52 53 /* This context's GFP mask */ 54 gfp_t gfp_mask; 55 56 int may_writepage; 57 58 /* Can pages be swapped as part of reclaim? */ 59 int may_swap; 60 61 /* This context's SWAP_CLUSTER_MAX. If freeing memory for 62 * suspend, we effectively ignore SWAP_CLUSTER_MAX. 63 * In this context, it doesn't matter that we scan the 64 * whole list at once. */ 65 int swap_cluster_max; 66 67 int swappiness; 68 69 int all_unreclaimable; 70 71 int order; 72 73 /* Which cgroup do we reclaim from */ 74 struct mem_cgroup *mem_cgroup; 75 76 /* Pluggable isolate pages callback */ 77 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst, 78 unsigned long *scanned, int order, int mode, 79 struct zone *z, struct mem_cgroup *mem_cont, 80 int active); 81}; 82 83#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) 84 85#ifdef ARCH_HAS_PREFETCH 86#define prefetch_prev_lru_page(_page, _base, _field) \ 87 do { \ 88 if ((_page)->lru.prev != _base) { \ 89 struct page *prev; \ 90 \ 91 prev = lru_to_page(&(_page->lru)); \ 92 prefetch(&prev->_field); \ 93 } \ 94 } while (0) 95#else 96#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) 97#endif 98 99#ifdef ARCH_HAS_PREFETCHW 100#define prefetchw_prev_lru_page(_page, _base, _field) \ 101 do { \ 102 if ((_page)->lru.prev != _base) { \ 103 struct page *prev; \ 104 \ 105 prev = lru_to_page(&(_page->lru)); \ 106 prefetchw(&prev->_field); \ 107 } \ 108 } while (0) 109#else 110#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) 111#endif 112 113/* 114 * From 0 .. 100. Higher means more swappy. 115 */ 116int vm_swappiness = 60; 117long vm_total_pages; /* The total number of pages which the VM controls */ 118 119static LIST_HEAD(shrinker_list); 120static DECLARE_RWSEM(shrinker_rwsem); 121 122/* 123 * Add a shrinker callback to be called from the vm 124 */ 125void register_shrinker(struct shrinker *shrinker) 126{ 127 shrinker->nr = 0; 128 down_write(&shrinker_rwsem); 129 list_add_tail(&shrinker->list, &shrinker_list); 130 up_write(&shrinker_rwsem); 131} 132EXPORT_SYMBOL(register_shrinker); 133 134/* 135 * Remove one 136 */ 137void unregister_shrinker(struct shrinker *shrinker) 138{ 139 down_write(&shrinker_rwsem); 140 list_del(&shrinker->list); 141 up_write(&shrinker_rwsem); 142} 143EXPORT_SYMBOL(unregister_shrinker); 144 145#define SHRINK_BATCH 128 146/* 147 * Call the shrink functions to age shrinkable caches 148 * 149 * Here we assume it costs one seek to replace a lru page and that it also 150 * takes a seek to recreate a cache object. With this in mind we age equal 151 * percentages of the lru and ageable caches. This should balance the seeks 152 * generated by these structures. 153 * 154 * If the vm encountered mapped pages on the LRU it increase the pressure on 155 * slab to avoid swapping. 156 * 157 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. 158 * 159 * `lru_pages' represents the number of on-LRU pages in all the zones which 160 * are eligible for the caller's allocation attempt. It is used for balancing 161 * slab reclaim versus page reclaim. 162 * 163 * Returns the number of slab objects which we shrunk. 164 */ 165unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask, 166 unsigned long lru_pages) 167{ 168 struct shrinker *shrinker; 169 unsigned long ret = 0; 170 171 if (scanned == 0) 172 scanned = SWAP_CLUSTER_MAX; 173 174 if (!down_read_trylock(&shrinker_rwsem)) 175 return 1; /* Assume we'll be able to shrink next time */ 176 177 list_for_each_entry(shrinker, &shrinker_list, list) { 178 unsigned long long delta; 179 unsigned long total_scan; 180 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask); 181 182 delta = (4 * scanned) / shrinker->seeks; 183 delta *= max_pass; 184 do_div(delta, lru_pages + 1); 185 shrinker->nr += delta; 186 if (shrinker->nr < 0) { 187 printk(KERN_ERR "%s: nr=%ld\n", 188 __FUNCTION__, shrinker->nr); 189 shrinker->nr = max_pass; 190 } 191 192 /* 193 * Avoid risking looping forever due to too large nr value: 194 * never try to free more than twice the estimate number of 195 * freeable entries. 196 */ 197 if (shrinker->nr > max_pass * 2) 198 shrinker->nr = max_pass * 2; 199 200 total_scan = shrinker->nr; 201 shrinker->nr = 0; 202 203 while (total_scan >= SHRINK_BATCH) { 204 long this_scan = SHRINK_BATCH; 205 int shrink_ret; 206 int nr_before; 207 208 nr_before = (*shrinker->shrink)(0, gfp_mask); 209 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask); 210 if (shrink_ret == -1) 211 break; 212 if (shrink_ret < nr_before) 213 ret += nr_before - shrink_ret; 214 count_vm_events(SLABS_SCANNED, this_scan); 215 total_scan -= this_scan; 216 217 cond_resched(); 218 } 219 220 shrinker->nr += total_scan; 221 } 222 up_read(&shrinker_rwsem); 223 return ret; 224} 225 226/* Called without lock on whether page is mapped, so answer is unstable */ 227static inline int page_mapping_inuse(struct page *page) 228{ 229 struct address_space *mapping; 230 231 /* Page is in somebody's page tables. */ 232 if (page_mapped(page)) 233 return 1; 234 235 /* Be more reluctant to reclaim swapcache than pagecache */ 236 if (PageSwapCache(page)) 237 return 1; 238 239 mapping = page_mapping(page); 240 if (!mapping) 241 return 0; 242 243 /* File is mmap'd by somebody? */ 244 return mapping_mapped(mapping); 245} 246 247static inline int is_page_cache_freeable(struct page *page) 248{ 249 return page_count(page) - !!PagePrivate(page) == 2; 250} 251 252static int may_write_to_queue(struct backing_dev_info *bdi) 253{ 254 if (current->flags & PF_SWAPWRITE) 255 return 1; 256 if (!bdi_write_congested(bdi)) 257 return 1; 258 if (bdi == current->backing_dev_info) 259 return 1; 260 return 0; 261} 262 263/* 264 * We detected a synchronous write error writing a page out. Probably 265 * -ENOSPC. We need to propagate that into the address_space for a subsequent 266 * fsync(), msync() or close(). 267 * 268 * The tricky part is that after writepage we cannot touch the mapping: nothing 269 * prevents it from being freed up. But we have a ref on the page and once 270 * that page is locked, the mapping is pinned. 271 * 272 * We're allowed to run sleeping lock_page() here because we know the caller has 273 * __GFP_FS. 274 */ 275static void handle_write_error(struct address_space *mapping, 276 struct page *page, int error) 277{ 278 lock_page(page); 279 if (page_mapping(page) == mapping) 280 mapping_set_error(mapping, error); 281 unlock_page(page); 282} 283 284/* Request for sync pageout. */ 285enum pageout_io { 286 PAGEOUT_IO_ASYNC, 287 PAGEOUT_IO_SYNC, 288}; 289 290/* possible outcome of pageout() */ 291typedef enum { 292 /* failed to write page out, page is locked */ 293 PAGE_KEEP, 294 /* move page to the active list, page is locked */ 295 PAGE_ACTIVATE, 296 /* page has been sent to the disk successfully, page is unlocked */ 297 PAGE_SUCCESS, 298 /* page is clean and locked */ 299 PAGE_CLEAN, 300} pageout_t; 301 302/* 303 * pageout is called by shrink_page_list() for each dirty page. 304 * Calls ->writepage(). 305 */ 306static pageout_t pageout(struct page *page, struct address_space *mapping, 307 enum pageout_io sync_writeback) 308{ 309 /* 310 * If the page is dirty, only perform writeback if that write 311 * will be non-blocking. To prevent this allocation from being 312 * stalled by pagecache activity. But note that there may be 313 * stalls if we need to run get_block(). We could test 314 * PagePrivate for that. 315 * 316 * If this process is currently in generic_file_write() against 317 * this page's queue, we can perform writeback even if that 318 * will block. 319 * 320 * If the page is swapcache, write it back even if that would 321 * block, for some throttling. This happens by accident, because 322 * swap_backing_dev_info is bust: it doesn't reflect the 323 * congestion state of the swapdevs. Easy to fix, if needed. 324 * See swapfile.c:page_queue_congested(). 325 */ 326 if (!is_page_cache_freeable(page)) 327 return PAGE_KEEP; 328 if (!mapping) { 329 /* 330 * Some data journaling orphaned pages can have 331 * page->mapping == NULL while being dirty with clean buffers. 332 */ 333 if (PagePrivate(page)) { 334 if (try_to_free_buffers(page)) { 335 ClearPageDirty(page); 336 printk("%s: orphaned page\n", __FUNCTION__); 337 return PAGE_CLEAN; 338 } 339 } 340 return PAGE_KEEP; 341 } 342 if (mapping->a_ops->writepage == NULL) 343 return PAGE_ACTIVATE; 344 if (!may_write_to_queue(mapping->backing_dev_info)) 345 return PAGE_KEEP; 346 347 if (clear_page_dirty_for_io(page)) { 348 int res; 349 struct writeback_control wbc = { 350 .sync_mode = WB_SYNC_NONE, 351 .nr_to_write = SWAP_CLUSTER_MAX, 352 .range_start = 0, 353 .range_end = LLONG_MAX, 354 .nonblocking = 1, 355 .for_reclaim = 1, 356 }; 357 358 SetPageReclaim(page); 359 res = mapping->a_ops->writepage(page, &wbc); 360 if (res < 0) 361 handle_write_error(mapping, page, res); 362 if (res == AOP_WRITEPAGE_ACTIVATE) { 363 ClearPageReclaim(page); 364 return PAGE_ACTIVATE; 365 } 366 367 /* 368 * Wait on writeback if requested to. This happens when 369 * direct reclaiming a large contiguous area and the 370 * first attempt to free a range of pages fails. 371 */ 372 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC) 373 wait_on_page_writeback(page); 374 375 if (!PageWriteback(page)) { 376 /* synchronous write or broken a_ops? */ 377 ClearPageReclaim(page); 378 } 379 inc_zone_page_state(page, NR_VMSCAN_WRITE); 380 return PAGE_SUCCESS; 381 } 382 383 return PAGE_CLEAN; 384} 385 386/* 387 * Attempt to detach a locked page from its ->mapping. If it is dirty or if 388 * someone else has a ref on the page, abort and return 0. If it was 389 * successfully detached, return 1. Assumes the caller has a single ref on 390 * this page. 391 */ 392int remove_mapping(struct address_space *mapping, struct page *page) 393{ 394 BUG_ON(!PageLocked(page)); 395 BUG_ON(mapping != page_mapping(page)); 396 397 write_lock_irq(&mapping->tree_lock); 398 /* 399 * The non racy check for a busy page. 400 * 401 * Must be careful with the order of the tests. When someone has 402 * a ref to the page, it may be possible that they dirty it then 403 * drop the reference. So if PageDirty is tested before page_count 404 * here, then the following race may occur: 405 * 406 * get_user_pages(&page); 407 * [user mapping goes away] 408 * write_to(page); 409 * !PageDirty(page) [good] 410 * SetPageDirty(page); 411 * put_page(page); 412 * !page_count(page) [good, discard it] 413 * 414 * [oops, our write_to data is lost] 415 * 416 * Reversing the order of the tests ensures such a situation cannot 417 * escape unnoticed. The smp_rmb is needed to ensure the page->flags 418 * load is not satisfied before that of page->_count. 419 * 420 * Note that if SetPageDirty is always performed via set_page_dirty, 421 * and thus under tree_lock, then this ordering is not required. 422 */ 423 if (unlikely(page_count(page) != 2)) 424 goto cannot_free; 425 smp_rmb(); 426 if (unlikely(PageDirty(page))) 427 goto cannot_free; 428 429 if (PageSwapCache(page)) { 430 swp_entry_t swap = { .val = page_private(page) }; 431 __delete_from_swap_cache(page); 432 write_unlock_irq(&mapping->tree_lock); 433 swap_free(swap); 434 __put_page(page); /* The pagecache ref */ 435 return 1; 436 } 437 438 __remove_from_page_cache(page); 439 write_unlock_irq(&mapping->tree_lock); 440 __put_page(page); 441 return 1; 442 443cannot_free: 444 write_unlock_irq(&mapping->tree_lock); 445 return 0; 446} 447 448/* 449 * shrink_page_list() returns the number of reclaimed pages 450 */ 451static unsigned long shrink_page_list(struct list_head *page_list, 452 struct scan_control *sc, 453 enum pageout_io sync_writeback) 454{ 455 LIST_HEAD(ret_pages); 456 struct pagevec freed_pvec; 457 int pgactivate = 0; 458 unsigned long nr_reclaimed = 0; 459 460 cond_resched(); 461 462 pagevec_init(&freed_pvec, 1); 463 while (!list_empty(page_list)) { 464 struct address_space *mapping; 465 struct page *page; 466 int may_enter_fs; 467 int referenced; 468 469 cond_resched(); 470 471 page = lru_to_page(page_list); 472 list_del(&page->lru); 473 474 if (TestSetPageLocked(page)) 475 goto keep; 476 477 VM_BUG_ON(PageActive(page)); 478 479 sc->nr_scanned++; 480 481 if (!sc->may_swap && page_mapped(page)) 482 goto keep_locked; 483 484 /* Double the slab pressure for mapped and swapcache pages */ 485 if (page_mapped(page) || PageSwapCache(page)) 486 sc->nr_scanned++; 487 488 may_enter_fs = (sc->gfp_mask & __GFP_FS) || 489 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); 490 491 if (PageWriteback(page)) { 492 /* 493 * Synchronous reclaim is performed in two passes, 494 * first an asynchronous pass over the list to 495 * start parallel writeback, and a second synchronous 496 * pass to wait for the IO to complete. Wait here 497 * for any page for which writeback has already 498 * started. 499 */ 500 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs) 501 wait_on_page_writeback(page); 502 else 503 goto keep_locked; 504 } 505 506 referenced = page_referenced(page, 1); 507 /* In active use or really unfreeable? Activate it. */ 508 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && 509 referenced && page_mapping_inuse(page)) 510 goto activate_locked; 511 512#ifdef CONFIG_SWAP 513 /* 514 * Anonymous process memory has backing store? 515 * Try to allocate it some swap space here. 516 */ 517 if (PageAnon(page) && !PageSwapCache(page)) 518 if (!add_to_swap(page, GFP_ATOMIC)) 519 goto activate_locked; 520#endif /* CONFIG_SWAP */ 521 522 mapping = page_mapping(page); 523 524 /* 525 * The page is mapped into the page tables of one or more 526 * processes. Try to unmap it here. 527 */ 528 if (page_mapped(page) && mapping) { 529 switch (try_to_unmap(page, 0)) { 530 case SWAP_FAIL: 531 goto activate_locked; 532 case SWAP_AGAIN: 533 goto keep_locked; 534 case SWAP_SUCCESS: 535 ; /* try to free the page below */ 536 } 537 } 538 539 if (PageDirty(page)) { 540 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced) 541 goto keep_locked; 542 if (!may_enter_fs) 543 goto keep_locked; 544 if (!sc->may_writepage) 545 goto keep_locked; 546 547 /* Page is dirty, try to write it out here */ 548 switch (pageout(page, mapping, sync_writeback)) { 549 case PAGE_KEEP: 550 goto keep_locked; 551 case PAGE_ACTIVATE: 552 goto activate_locked; 553 case PAGE_SUCCESS: 554 if (PageWriteback(page) || PageDirty(page)) 555 goto keep; 556 /* 557 * A synchronous write - probably a ramdisk. Go 558 * ahead and try to reclaim the page. 559 */ 560 if (TestSetPageLocked(page)) 561 goto keep; 562 if (PageDirty(page) || PageWriteback(page)) 563 goto keep_locked; 564 mapping = page_mapping(page); 565 case PAGE_CLEAN: 566 ; /* try to free the page below */ 567 } 568 } 569 570 /* 571 * If the page has buffers, try to free the buffer mappings 572 * associated with this page. If we succeed we try to free 573 * the page as well. 574 * 575 * We do this even if the page is PageDirty(). 576 * try_to_release_page() does not perform I/O, but it is 577 * possible for a page to have PageDirty set, but it is actually 578 * clean (all its buffers are clean). This happens if the 579 * buffers were written out directly, with submit_bh(). ext3 580 * will do this, as well as the blockdev mapping. 581 * try_to_release_page() will discover that cleanness and will 582 * drop the buffers and mark the page clean - it can be freed. 583 * 584 * Rarely, pages can have buffers and no ->mapping. These are 585 * the pages which were not successfully invalidated in 586 * truncate_complete_page(). We try to drop those buffers here 587 * and if that worked, and the page is no longer mapped into 588 * process address space (page_count == 1) it can be freed. 589 * Otherwise, leave the page on the LRU so it is swappable. 590 */ 591 if (PagePrivate(page)) { 592 if (!try_to_release_page(page, sc->gfp_mask)) 593 goto activate_locked; 594 if (!mapping && page_count(page) == 1) 595 goto free_it; 596 } 597 598 if (!mapping || !remove_mapping(mapping, page)) 599 goto keep_locked; 600 601free_it: 602 unlock_page(page); 603 nr_reclaimed++; 604 if (!pagevec_add(&freed_pvec, page)) 605 __pagevec_release_nonlru(&freed_pvec); 606 continue; 607 608activate_locked: 609 SetPageActive(page); 610 pgactivate++; 611keep_locked: 612 unlock_page(page); 613keep: 614 list_add(&page->lru, &ret_pages); 615 VM_BUG_ON(PageLRU(page)); 616 } 617 list_splice(&ret_pages, page_list); 618 if (pagevec_count(&freed_pvec)) 619 __pagevec_release_nonlru(&freed_pvec); 620 count_vm_events(PGACTIVATE, pgactivate); 621 return nr_reclaimed; 622} 623 624/* LRU Isolation modes. */ 625#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */ 626#define ISOLATE_ACTIVE 1 /* Isolate active pages. */ 627#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */ 628 629/* 630 * Attempt to remove the specified page from its LRU. Only take this page 631 * if it is of the appropriate PageActive status. Pages which are being 632 * freed elsewhere are also ignored. 633 * 634 * page: page to consider 635 * mode: one of the LRU isolation modes defined above 636 * 637 * returns 0 on success, -ve errno on failure. 638 */ 639int __isolate_lru_page(struct page *page, int mode) 640{ 641 int ret = -EINVAL; 642 643 /* Only take pages on the LRU. */ 644 if (!PageLRU(page)) 645 return ret; 646 647 /* 648 * When checking the active state, we need to be sure we are 649 * dealing with comparible boolean values. Take the logical not 650 * of each. 651 */ 652 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode)) 653 return ret; 654 655 ret = -EBUSY; 656 if (likely(get_page_unless_zero(page))) { 657 /* 658 * Be careful not to clear PageLRU until after we're 659 * sure the page is not being freed elsewhere -- the 660 * page release code relies on it. 661 */ 662 ClearPageLRU(page); 663 ret = 0; 664 } 665 666 return ret; 667} 668 669/* 670 * zone->lru_lock is heavily contended. Some of the functions that 671 * shrink the lists perform better by taking out a batch of pages 672 * and working on them outside the LRU lock. 673 * 674 * For pagecache intensive workloads, this function is the hottest 675 * spot in the kernel (apart from copy_*_user functions). 676 * 677 * Appropriate locks must be held before calling this function. 678 * 679 * @nr_to_scan: The number of pages to look through on the list. 680 * @src: The LRU list to pull pages off. 681 * @dst: The temp list to put pages on to. 682 * @scanned: The number of pages that were scanned. 683 * @order: The caller's attempted allocation order 684 * @mode: One of the LRU isolation modes 685 * 686 * returns how many pages were moved onto *@dst. 687 */ 688static unsigned long isolate_lru_pages(unsigned long nr_to_scan, 689 struct list_head *src, struct list_head *dst, 690 unsigned long *scanned, int order, int mode) 691{ 692 unsigned long nr_taken = 0; 693 unsigned long scan; 694 695 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) { 696 struct page *page; 697 unsigned long pfn; 698 unsigned long end_pfn; 699 unsigned long page_pfn; 700 int zone_id; 701 702 page = lru_to_page(src); 703 prefetchw_prev_lru_page(page, src, flags); 704 705 VM_BUG_ON(!PageLRU(page)); 706 707 switch (__isolate_lru_page(page, mode)) { 708 case 0: 709 list_move(&page->lru, dst); 710 nr_taken++; 711 break; 712 713 case -EBUSY: 714 /* else it is being freed elsewhere */ 715 list_move(&page->lru, src); 716 continue; 717 718 default: 719 BUG(); 720 } 721 722 if (!order) 723 continue; 724 725 /* 726 * Attempt to take all pages in the order aligned region 727 * surrounding the tag page. Only take those pages of 728 * the same active state as that tag page. We may safely 729 * round the target page pfn down to the requested order 730 * as the mem_map is guarenteed valid out to MAX_ORDER, 731 * where that page is in a different zone we will detect 732 * it from its zone id and abort this block scan. 733 */ 734 zone_id = page_zone_id(page); 735 page_pfn = page_to_pfn(page); 736 pfn = page_pfn & ~((1 << order) - 1); 737 end_pfn = pfn + (1 << order); 738 for (; pfn < end_pfn; pfn++) { 739 struct page *cursor_page; 740 741 /* The target page is in the block, ignore it. */ 742 if (unlikely(pfn == page_pfn)) 743 continue; 744 745 /* Avoid holes within the zone. */ 746 if (unlikely(!pfn_valid_within(pfn))) 747 break; 748 749 cursor_page = pfn_to_page(pfn); 750 /* Check that we have not crossed a zone boundary. */ 751 if (unlikely(page_zone_id(cursor_page) != zone_id)) 752 continue; 753 switch (__isolate_lru_page(cursor_page, mode)) { 754 case 0: 755 list_move(&cursor_page->lru, dst); 756 nr_taken++; 757 scan++; 758 break; 759 760 case -EBUSY: 761 /* else it is being freed elsewhere */ 762 list_move(&cursor_page->lru, src); 763 default: 764 break; 765 } 766 } 767 } 768 769 *scanned = scan; 770 return nr_taken; 771} 772 773static unsigned long isolate_pages_global(unsigned long nr, 774 struct list_head *dst, 775 unsigned long *scanned, int order, 776 int mode, struct zone *z, 777 struct mem_cgroup *mem_cont, 778 int active) 779{ 780 if (active) 781 return isolate_lru_pages(nr, &z->active_list, dst, 782 scanned, order, mode); 783 else 784 return isolate_lru_pages(nr, &z->inactive_list, dst, 785 scanned, order, mode); 786} 787 788/* 789 * clear_active_flags() is a helper for shrink_active_list(), clearing 790 * any active bits from the pages in the list. 791 */ 792static unsigned long clear_active_flags(struct list_head *page_list) 793{ 794 int nr_active = 0; 795 struct page *page; 796 797 list_for_each_entry(page, page_list, lru) 798 if (PageActive(page)) { 799 ClearPageActive(page); 800 nr_active++; 801 } 802 803 return nr_active; 804} 805 806/* 807 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number 808 * of reclaimed pages 809 */ 810static unsigned long shrink_inactive_list(unsigned long max_scan, 811 struct zone *zone, struct scan_control *sc) 812{ 813 LIST_HEAD(page_list); 814 struct pagevec pvec; 815 unsigned long nr_scanned = 0; 816 unsigned long nr_reclaimed = 0; 817 818 pagevec_init(&pvec, 1); 819 820 lru_add_drain(); 821 spin_lock_irq(&zone->lru_lock); 822 do { 823 struct page *page; 824 unsigned long nr_taken; 825 unsigned long nr_scan; 826 unsigned long nr_freed; 827 unsigned long nr_active; 828 829 nr_taken = sc->isolate_pages(sc->swap_cluster_max, 830 &page_list, &nr_scan, sc->order, 831 (sc->order > PAGE_ALLOC_COSTLY_ORDER)? 832 ISOLATE_BOTH : ISOLATE_INACTIVE, 833 zone, sc->mem_cgroup, 0); 834 nr_active = clear_active_flags(&page_list); 835 __count_vm_events(PGDEACTIVATE, nr_active); 836 837 __mod_zone_page_state(zone, NR_ACTIVE, -nr_active); 838 __mod_zone_page_state(zone, NR_INACTIVE, 839 -(nr_taken - nr_active)); 840 zone->pages_scanned += nr_scan; 841 spin_unlock_irq(&zone->lru_lock); 842 843 nr_scanned += nr_scan; 844 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC); 845 846 /* 847 * If we are direct reclaiming for contiguous pages and we do 848 * not reclaim everything in the list, try again and wait 849 * for IO to complete. This will stall high-order allocations 850 * but that should be acceptable to the caller 851 */ 852 if (nr_freed < nr_taken && !current_is_kswapd() && 853 sc->order > PAGE_ALLOC_COSTLY_ORDER) { 854 congestion_wait(WRITE, HZ/10); 855 856 /* 857 * The attempt at page out may have made some 858 * of the pages active, mark them inactive again. 859 */ 860 nr_active = clear_active_flags(&page_list); 861 count_vm_events(PGDEACTIVATE, nr_active); 862 863 nr_freed += shrink_page_list(&page_list, sc, 864 PAGEOUT_IO_SYNC); 865 } 866 867 nr_reclaimed += nr_freed; 868 local_irq_disable(); 869 if (current_is_kswapd()) { 870 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan); 871 __count_vm_events(KSWAPD_STEAL, nr_freed); 872 } else 873 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan); 874 __count_zone_vm_events(PGSTEAL, zone, nr_freed); 875 876 if (nr_taken == 0) 877 goto done; 878 879 spin_lock(&zone->lru_lock); 880 /* 881 * Put back any unfreeable pages. 882 */ 883 while (!list_empty(&page_list)) { 884 page = lru_to_page(&page_list); 885 VM_BUG_ON(PageLRU(page)); 886 SetPageLRU(page); 887 list_del(&page->lru); 888 if (PageActive(page)) 889 add_page_to_active_list(zone, page); 890 else 891 add_page_to_inactive_list(zone, page); 892 if (!pagevec_add(&pvec, page)) { 893 spin_unlock_irq(&zone->lru_lock); 894 __pagevec_release(&pvec); 895 spin_lock_irq(&zone->lru_lock); 896 } 897 } 898 } while (nr_scanned < max_scan); 899 spin_unlock(&zone->lru_lock); 900done: 901 local_irq_enable(); 902 pagevec_release(&pvec); 903 return nr_reclaimed; 904} 905 906/* 907 * We are about to scan this zone at a certain priority level. If that priority 908 * level is smaller (ie: more urgent) than the previous priority, then note 909 * that priority level within the zone. This is done so that when the next 910 * process comes in to scan this zone, it will immediately start out at this 911 * priority level rather than having to build up its own scanning priority. 912 * Here, this priority affects only the reclaim-mapped threshold. 913 */ 914static inline void note_zone_scanning_priority(struct zone *zone, int priority) 915{ 916 if (priority < zone->prev_priority) 917 zone->prev_priority = priority; 918} 919 920static inline int zone_is_near_oom(struct zone *zone) 921{ 922 return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE) 923 + zone_page_state(zone, NR_INACTIVE))*3; 924} 925 926/* 927 * This moves pages from the active list to the inactive list. 928 * 929 * We move them the other way if the page is referenced by one or more 930 * processes, from rmap. 931 * 932 * If the pages are mostly unmapped, the processing is fast and it is 933 * appropriate to hold zone->lru_lock across the whole operation. But if 934 * the pages are mapped, the processing is slow (page_referenced()) so we 935 * should drop zone->lru_lock around each page. It's impossible to balance 936 * this, so instead we remove the pages from the LRU while processing them. 937 * It is safe to rely on PG_active against the non-LRU pages in here because 938 * nobody will play with that bit on a non-LRU page. 939 * 940 * The downside is that we have to touch page->_count against each page. 941 * But we had to alter page->flags anyway. 942 */ 943static void shrink_active_list(unsigned long nr_pages, struct zone *zone, 944 struct scan_control *sc, int priority) 945{ 946 unsigned long pgmoved; 947 int pgdeactivate = 0; 948 unsigned long pgscanned; 949 LIST_HEAD(l_hold); /* The pages which were snipped off */ 950 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */ 951 LIST_HEAD(l_active); /* Pages to go onto the active_list */ 952 struct page *page; 953 struct pagevec pvec; 954 int reclaim_mapped = 0; 955 956 if (sc->may_swap) { 957 long mapped_ratio; 958 long distress; 959 long swap_tendency; 960 long imbalance; 961 962 if (zone_is_near_oom(zone)) 963 goto force_reclaim_mapped; 964 965 /* 966 * `distress' is a measure of how much trouble we're having 967 * reclaiming pages. 0 -> no problems. 100 -> great trouble. 968 */ 969 distress = 100 >> min(zone->prev_priority, priority); 970 971 /* 972 * The point of this algorithm is to decide when to start 973 * reclaiming mapped memory instead of just pagecache. Work out 974 * how much memory 975 * is mapped. 976 */ 977 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) + 978 global_page_state(NR_ANON_PAGES)) * 100) / 979 vm_total_pages; 980 981 /* 982 * Now decide how much we really want to unmap some pages. The 983 * mapped ratio is downgraded - just because there's a lot of 984 * mapped memory doesn't necessarily mean that page reclaim 985 * isn't succeeding. 986 * 987 * The distress ratio is important - we don't want to start 988 * going oom. 989 * 990 * A 100% value of vm_swappiness overrides this algorithm 991 * altogether. 992 */ 993 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness; 994 995 /* 996 * If there's huge imbalance between active and inactive 997 * (think active 100 times larger than inactive) we should 998 * become more permissive, or the system will take too much 999 * cpu before it start swapping during memory pressure. 1000 * Distress is about avoiding early-oom, this is about 1001 * making swappiness graceful despite setting it to low 1002 * values. 1003 * 1004 * Avoid div by zero with nr_inactive+1, and max resulting 1005 * value is vm_total_pages. 1006 */ 1007 imbalance = zone_page_state(zone, NR_ACTIVE); 1008 imbalance /= zone_page_state(zone, NR_INACTIVE) + 1; 1009 1010 /* 1011 * Reduce the effect of imbalance if swappiness is low, 1012 * this means for a swappiness very low, the imbalance 1013 * must be much higher than 100 for this logic to make 1014 * the difference. 1015 * 1016 * Max temporary value is vm_total_pages*100. 1017 */ 1018 imbalance *= (vm_swappiness + 1); 1019 imbalance /= 100; 1020 1021 /* 1022 * If not much of the ram is mapped, makes the imbalance 1023 * less relevant, it's high priority we refill the inactive 1024 * list with mapped pages only in presence of high ratio of 1025 * mapped pages. 1026 * 1027 * Max temporary value is vm_total_pages*100. 1028 */ 1029 imbalance *= mapped_ratio; 1030 imbalance /= 100; 1031 1032 /* apply imbalance feedback to swap_tendency */ 1033 swap_tendency += imbalance; 1034 1035 /* 1036 * Now use this metric to decide whether to start moving mapped 1037 * memory onto the inactive list. 1038 */ 1039 if (swap_tendency >= 100) 1040force_reclaim_mapped: 1041 reclaim_mapped = 1; 1042 } 1043 1044 lru_add_drain(); 1045 spin_lock_irq(&zone->lru_lock); 1046 pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order, 1047 ISOLATE_ACTIVE, zone, 1048 sc->mem_cgroup, 1); 1049 zone->pages_scanned += pgscanned; 1050 __mod_zone_page_state(zone, NR_ACTIVE, -pgmoved); 1051 spin_unlock_irq(&zone->lru_lock); 1052 1053 while (!list_empty(&l_hold)) { 1054 cond_resched(); 1055 page = lru_to_page(&l_hold); 1056 list_del(&page->lru); 1057 if (page_mapped(page)) { 1058 if (!reclaim_mapped || 1059 (total_swap_pages == 0 && PageAnon(page)) || 1060 page_referenced(page, 0)) { 1061 list_add(&page->lru, &l_active); 1062 continue; 1063 } 1064 } 1065 list_add(&page->lru, &l_inactive); 1066 } 1067 1068 pagevec_init(&pvec, 1); 1069 pgmoved = 0; 1070 spin_lock_irq(&zone->lru_lock); 1071 while (!list_empty(&l_inactive)) { 1072 page = lru_to_page(&l_inactive); 1073 prefetchw_prev_lru_page(page, &l_inactive, flags); 1074 VM_BUG_ON(PageLRU(page)); 1075 SetPageLRU(page); 1076 VM_BUG_ON(!PageActive(page)); 1077 ClearPageActive(page); 1078 1079 list_move(&page->lru, &zone->inactive_list); 1080 mem_cgroup_move_lists(page_get_page_cgroup(page), false); 1081 pgmoved++; 1082 if (!pagevec_add(&pvec, page)) { 1083 __mod_zone_page_state(zone, NR_INACTIVE, pgmoved); 1084 spin_unlock_irq(&zone->lru_lock); 1085 pgdeactivate += pgmoved; 1086 pgmoved = 0; 1087 if (buffer_heads_over_limit) 1088 pagevec_strip(&pvec); 1089 __pagevec_release(&pvec); 1090 spin_lock_irq(&zone->lru_lock); 1091 } 1092 } 1093 __mod_zone_page_state(zone, NR_INACTIVE, pgmoved); 1094 pgdeactivate += pgmoved; 1095 if (buffer_heads_over_limit) { 1096 spin_unlock_irq(&zone->lru_lock); 1097 pagevec_strip(&pvec); 1098 spin_lock_irq(&zone->lru_lock); 1099 } 1100 1101 pgmoved = 0; 1102 while (!list_empty(&l_active)) { 1103 page = lru_to_page(&l_active); 1104 prefetchw_prev_lru_page(page, &l_active, flags); 1105 VM_BUG_ON(PageLRU(page)); 1106 SetPageLRU(page); 1107 VM_BUG_ON(!PageActive(page)); 1108 list_move(&page->lru, &zone->active_list); 1109 mem_cgroup_move_lists(page_get_page_cgroup(page), true); 1110 pgmoved++; 1111 if (!pagevec_add(&pvec, page)) { 1112 __mod_zone_page_state(zone, NR_ACTIVE, pgmoved); 1113 pgmoved = 0; 1114 spin_unlock_irq(&zone->lru_lock); 1115 __pagevec_release(&pvec); 1116 spin_lock_irq(&zone->lru_lock); 1117 } 1118 } 1119 __mod_zone_page_state(zone, NR_ACTIVE, pgmoved); 1120 1121 __count_zone_vm_events(PGREFILL, zone, pgscanned); 1122 __count_vm_events(PGDEACTIVATE, pgdeactivate); 1123 spin_unlock_irq(&zone->lru_lock); 1124 1125 pagevec_release(&pvec); 1126} 1127 1128/* 1129 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. 1130 */ 1131static unsigned long shrink_zone(int priority, struct zone *zone, 1132 struct scan_control *sc) 1133{ 1134 unsigned long nr_active; 1135 unsigned long nr_inactive; 1136 unsigned long nr_to_scan; 1137 unsigned long nr_reclaimed = 0; 1138 1139 /* 1140 * Add one to `nr_to_scan' just to make sure that the kernel will 1141 * slowly sift through the active list. 1142 */ 1143 zone->nr_scan_active += 1144 (zone_page_state(zone, NR_ACTIVE) >> priority) + 1; 1145 nr_active = zone->nr_scan_active; 1146 if (nr_active >= sc->swap_cluster_max) 1147 zone->nr_scan_active = 0; 1148 else 1149 nr_active = 0; 1150 1151 zone->nr_scan_inactive += 1152 (zone_page_state(zone, NR_INACTIVE) >> priority) + 1; 1153 nr_inactive = zone->nr_scan_inactive; 1154 if (nr_inactive >= sc->swap_cluster_max) 1155 zone->nr_scan_inactive = 0; 1156 else 1157 nr_inactive = 0; 1158 1159 while (nr_active || nr_inactive) { 1160 if (nr_active) { 1161 nr_to_scan = min(nr_active, 1162 (unsigned long)sc->swap_cluster_max); 1163 nr_active -= nr_to_scan; 1164 shrink_active_list(nr_to_scan, zone, sc, priority); 1165 } 1166 1167 if (nr_inactive) { 1168 nr_to_scan = min(nr_inactive, 1169 (unsigned long)sc->swap_cluster_max); 1170 nr_inactive -= nr_to_scan; 1171 nr_reclaimed += shrink_inactive_list(nr_to_scan, zone, 1172 sc); 1173 } 1174 } 1175 1176 throttle_vm_writeout(sc->gfp_mask); 1177 return nr_reclaimed; 1178} 1179 1180/* 1181 * This is the direct reclaim path, for page-allocating processes. We only 1182 * try to reclaim pages from zones which will satisfy the caller's allocation 1183 * request. 1184 * 1185 * We reclaim from a zone even if that zone is over pages_high. Because: 1186 * a) The caller may be trying to free *extra* pages to satisfy a higher-order 1187 * allocation or 1188 * b) The zones may be over pages_high but they must go *over* pages_high to 1189 * satisfy the `incremental min' zone defense algorithm. 1190 * 1191 * Returns the number of reclaimed pages. 1192 * 1193 * If a zone is deemed to be full of pinned pages then just give it a light 1194 * scan then give up on it. 1195 */ 1196static unsigned long shrink_zones(int priority, struct zone **zones, 1197 struct scan_control *sc) 1198{ 1199 unsigned long nr_reclaimed = 0; 1200 int i; 1201 1202 sc->all_unreclaimable = 1; 1203 for (i = 0; zones[i] != NULL; i++) { 1204 struct zone *zone = zones[i]; 1205 1206 if (!populated_zone(zone)) 1207 continue; 1208 1209 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 1210 continue; 1211 1212 note_zone_scanning_priority(zone, priority); 1213 1214 if (zone_is_all_unreclaimable(zone) && priority != DEF_PRIORITY) 1215 continue; /* Let kswapd poll it */ 1216 1217 sc->all_unreclaimable = 0; 1218 1219 nr_reclaimed += shrink_zone(priority, zone, sc); 1220 } 1221 return nr_reclaimed; 1222} 1223 1224/* 1225 * This is the main entry point to direct page reclaim. 1226 * 1227 * If a full scan of the inactive list fails to free enough memory then we 1228 * are "out of memory" and something needs to be killed. 1229 * 1230 * If the caller is !__GFP_FS then the probability of a failure is reasonably 1231 * high - the zone may be full of dirty or under-writeback pages, which this 1232 * caller can't do much about. We kick pdflush and take explicit naps in the 1233 * hope that some of these pages can be written. But if the allocating task 1234 * holds filesystem locks which prevent writeout this might not work, and the 1235 * allocation attempt will fail. 1236 */ 1237static unsigned long do_try_to_free_pages(struct zone **zones, gfp_t gfp_mask, 1238 struct scan_control *sc) 1239{ 1240 int priority; 1241 int ret = 0; 1242 unsigned long total_scanned = 0; 1243 unsigned long nr_reclaimed = 0; 1244 struct reclaim_state *reclaim_state = current->reclaim_state; 1245 unsigned long lru_pages = 0; 1246 int i; 1247 1248 count_vm_event(ALLOCSTALL); 1249 1250 for (i = 0; zones[i] != NULL; i++) { 1251 struct zone *zone = zones[i]; 1252 1253 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 1254 continue; 1255 1256 lru_pages += zone_page_state(zone, NR_ACTIVE) 1257 + zone_page_state(zone, NR_INACTIVE); 1258 } 1259 1260 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 1261 sc->nr_scanned = 0; 1262 if (!priority) 1263 disable_swap_token(); 1264 nr_reclaimed += shrink_zones(priority, zones, sc); 1265 /* 1266 * Don't shrink slabs when reclaiming memory from 1267 * over limit cgroups 1268 */ 1269 if (sc->mem_cgroup == NULL) 1270 shrink_slab(sc->nr_scanned, gfp_mask, lru_pages); 1271 if (reclaim_state) { 1272 nr_reclaimed += reclaim_state->reclaimed_slab; 1273 reclaim_state->reclaimed_slab = 0; 1274 } 1275 total_scanned += sc->nr_scanned; 1276 if (nr_reclaimed >= sc->swap_cluster_max) { 1277 ret = 1; 1278 goto out; 1279 } 1280 1281 /* 1282 * Try to write back as many pages as we just scanned. This 1283 * tends to cause slow streaming writers to write data to the 1284 * disk smoothly, at the dirtying rate, which is nice. But 1285 * that's undesirable in laptop mode, where we *want* lumpy 1286 * writeout. So in laptop mode, write out the whole world. 1287 */ 1288 if (total_scanned > sc->swap_cluster_max + 1289 sc->swap_cluster_max / 2) { 1290 wakeup_pdflush(laptop_mode ? 0 : total_scanned); 1291 sc->may_writepage = 1; 1292 } 1293 1294 /* Take a nap, wait for some writeback to complete */ 1295 if (sc->nr_scanned && priority < DEF_PRIORITY - 2) 1296 congestion_wait(WRITE, HZ/10); 1297 } 1298 /* top priority shrink_caches still had more to do? don't OOM, then */ 1299 if (!sc->all_unreclaimable && sc->mem_cgroup == NULL) 1300 ret = 1; 1301out: 1302 /* 1303 * Now that we've scanned all the zones at this priority level, note 1304 * that level within the zone so that the next thread which performs 1305 * scanning of this zone will immediately start out at this priority 1306 * level. This affects only the decision whether or not to bring 1307 * mapped pages onto the inactive list. 1308 */ 1309 if (priority < 0) 1310 priority = 0; 1311 for (i = 0; zones[i] != NULL; i++) { 1312 struct zone *zone = zones[i]; 1313 1314 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 1315 continue; 1316 1317 zone->prev_priority = priority; 1318 } 1319 return ret; 1320} 1321 1322unsigned long try_to_free_pages(struct zone **zones, int order, gfp_t gfp_mask) 1323{ 1324 struct scan_control sc = { 1325 .gfp_mask = gfp_mask, 1326 .may_writepage = !laptop_mode, 1327 .swap_cluster_max = SWAP_CLUSTER_MAX, 1328 .may_swap = 1, 1329 .swappiness = vm_swappiness, 1330 .order = order, 1331 .mem_cgroup = NULL, 1332 .isolate_pages = isolate_pages_global, 1333 }; 1334 1335 return do_try_to_free_pages(zones, gfp_mask, &sc); 1336} 1337 1338#ifdef CONFIG_CGROUP_MEM_CONT 1339 1340#ifdef CONFIG_HIGHMEM 1341#define ZONE_USERPAGES ZONE_HIGHMEM 1342#else 1343#define ZONE_USERPAGES ZONE_NORMAL 1344#endif 1345 1346unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont) 1347{ 1348 struct scan_control sc = { 1349 .gfp_mask = GFP_KERNEL, 1350 .may_writepage = !laptop_mode, 1351 .may_swap = 1, 1352 .swap_cluster_max = SWAP_CLUSTER_MAX, 1353 .swappiness = vm_swappiness, 1354 .order = 0, 1355 .mem_cgroup = mem_cont, 1356 .isolate_pages = mem_cgroup_isolate_pages, 1357 }; 1358 int node; 1359 struct zone **zones; 1360 1361 for_each_online_node(node) { 1362 zones = NODE_DATA(node)->node_zonelists[ZONE_USERPAGES].zones; 1363 if (do_try_to_free_pages(zones, sc.gfp_mask, &sc)) 1364 return 1; 1365 } 1366 return 0; 1367} 1368#endif 1369 1370/* 1371 * For kswapd, balance_pgdat() will work across all this node's zones until 1372 * they are all at pages_high. 1373 * 1374 * Returns the number of pages which were actually freed. 1375 * 1376 * There is special handling here for zones which are full of pinned pages. 1377 * This can happen if the pages are all mlocked, or if they are all used by 1378 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. 1379 * What we do is to detect the case where all pages in the zone have been 1380 * scanned twice and there has been zero successful reclaim. Mark the zone as 1381 * dead and from now on, only perform a short scan. Basically we're polling 1382 * the zone for when the problem goes away. 1383 * 1384 * kswapd scans the zones in the highmem->normal->dma direction. It skips 1385 * zones which have free_pages > pages_high, but once a zone is found to have 1386 * free_pages <= pages_high, we scan that zone and the lower zones regardless 1387 * of the number of free pages in the lower zones. This interoperates with 1388 * the page allocator fallback scheme to ensure that aging of pages is balanced 1389 * across the zones. 1390 */ 1391static unsigned long balance_pgdat(pg_data_t *pgdat, int order) 1392{ 1393 int all_zones_ok; 1394 int priority; 1395 int i; 1396 unsigned long total_scanned; 1397 unsigned long nr_reclaimed; 1398 struct reclaim_state *reclaim_state = current->reclaim_state; 1399 struct scan_control sc = { 1400 .gfp_mask = GFP_KERNEL, 1401 .may_swap = 1, 1402 .swap_cluster_max = SWAP_CLUSTER_MAX, 1403 .swappiness = vm_swappiness, 1404 .order = order, 1405 .mem_cgroup = NULL, 1406 .isolate_pages = isolate_pages_global, 1407 }; 1408 /* 1409 * temp_priority is used to remember the scanning priority at which 1410 * this zone was successfully refilled to free_pages == pages_high. 1411 */ 1412 int temp_priority[MAX_NR_ZONES]; 1413 1414loop_again: 1415 total_scanned = 0; 1416 nr_reclaimed = 0; 1417 sc.may_writepage = !laptop_mode; 1418 count_vm_event(PAGEOUTRUN); 1419 1420 for (i = 0; i < pgdat->nr_zones; i++) 1421 temp_priority[i] = DEF_PRIORITY; 1422 1423 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 1424 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ 1425 unsigned long lru_pages = 0; 1426 1427 /* The swap token gets in the way of swapout... */ 1428 if (!priority) 1429 disable_swap_token(); 1430 1431 all_zones_ok = 1; 1432 1433 /* 1434 * Scan in the highmem->dma direction for the highest 1435 * zone which needs scanning 1436 */ 1437 for (i = pgdat->nr_zones - 1; i >= 0; i--) { 1438 struct zone *zone = pgdat->node_zones + i; 1439 1440 if (!populated_zone(zone)) 1441 continue; 1442 1443 if (zone_is_all_unreclaimable(zone) && 1444 priority != DEF_PRIORITY) 1445 continue; 1446 1447 if (!zone_watermark_ok(zone, order, zone->pages_high, 1448 0, 0)) { 1449 end_zone = i; 1450 break; 1451 } 1452 } 1453 if (i < 0) 1454 goto out; 1455 1456 for (i = 0; i <= end_zone; i++) { 1457 struct zone *zone = pgdat->node_zones + i; 1458 1459 lru_pages += zone_page_state(zone, NR_ACTIVE) 1460 + zone_page_state(zone, NR_INACTIVE); 1461 } 1462 1463 /* 1464 * Now scan the zone in the dma->highmem direction, stopping 1465 * at the last zone which needs scanning. 1466 * 1467 * We do this because the page allocator works in the opposite 1468 * direction. This prevents the page allocator from allocating 1469 * pages behind kswapd's direction of progress, which would 1470 * cause too much scanning of the lower zones. 1471 */ 1472 for (i = 0; i <= end_zone; i++) { 1473 struct zone *zone = pgdat->node_zones + i; 1474 int nr_slab; 1475 1476 if (!populated_zone(zone)) 1477 continue; 1478 1479 if (zone_is_all_unreclaimable(zone) && 1480 priority != DEF_PRIORITY) 1481 continue; 1482 1483 if (!zone_watermark_ok(zone, order, zone->pages_high, 1484 end_zone, 0)) 1485 all_zones_ok = 0; 1486 temp_priority[i] = priority; 1487 sc.nr_scanned = 0; 1488 note_zone_scanning_priority(zone, priority); 1489 /* 1490 * We put equal pressure on every zone, unless one 1491 * zone has way too many pages free already. 1492 */ 1493 if (!zone_watermark_ok(zone, order, 8*zone->pages_high, 1494 end_zone, 0)) 1495 nr_reclaimed += shrink_zone(priority, zone, &sc); 1496 reclaim_state->reclaimed_slab = 0; 1497 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL, 1498 lru_pages); 1499 nr_reclaimed += reclaim_state->reclaimed_slab; 1500 total_scanned += sc.nr_scanned; 1501 if (zone_is_all_unreclaimable(zone)) 1502 continue; 1503 if (nr_slab == 0 && zone->pages_scanned >= 1504 (zone_page_state(zone, NR_ACTIVE) 1505 + zone_page_state(zone, NR_INACTIVE)) * 6) 1506 zone_set_flag(zone, 1507 ZONE_ALL_UNRECLAIMABLE); 1508 /* 1509 * If we've done a decent amount of scanning and 1510 * the reclaim ratio is low, start doing writepage 1511 * even in laptop mode 1512 */ 1513 if (total_scanned > SWAP_CLUSTER_MAX * 2 && 1514 total_scanned > nr_reclaimed + nr_reclaimed / 2) 1515 sc.may_writepage = 1; 1516 } 1517 if (all_zones_ok) 1518 break; /* kswapd: all done */ 1519 /* 1520 * OK, kswapd is getting into trouble. Take a nap, then take 1521 * another pass across the zones. 1522 */ 1523 if (total_scanned && priority < DEF_PRIORITY - 2) 1524 congestion_wait(WRITE, HZ/10); 1525 1526 /* 1527 * We do this so kswapd doesn't build up large priorities for 1528 * example when it is freeing in parallel with allocators. It 1529 * matches the direct reclaim path behaviour in terms of impact 1530 * on zone->*_priority. 1531 */ 1532 if (nr_reclaimed >= SWAP_CLUSTER_MAX) 1533 break; 1534 } 1535out: 1536 /* 1537 * Note within each zone the priority level at which this zone was 1538 * brought into a happy state. So that the next thread which scans this 1539 * zone will start out at that priority level. 1540 */ 1541 for (i = 0; i < pgdat->nr_zones; i++) { 1542 struct zone *zone = pgdat->node_zones + i; 1543 1544 zone->prev_priority = temp_priority[i]; 1545 } 1546 if (!all_zones_ok) { 1547 cond_resched(); 1548 1549 try_to_freeze(); 1550 1551 goto loop_again; 1552 } 1553 1554 return nr_reclaimed; 1555} 1556 1557/* 1558 * The background pageout daemon, started as a kernel thread 1559 * from the init process. 1560 * 1561 * This basically trickles out pages so that we have _some_ 1562 * free memory available even if there is no other activity 1563 * that frees anything up. This is needed for things like routing 1564 * etc, where we otherwise might have all activity going on in 1565 * asynchronous contexts that cannot page things out. 1566 * 1567 * If there are applications that are active memory-allocators 1568 * (most normal use), this basically shouldn't matter. 1569 */ 1570static int kswapd(void *p) 1571{ 1572 unsigned long order; 1573 pg_data_t *pgdat = (pg_data_t*)p; 1574 struct task_struct *tsk = current; 1575 DEFINE_WAIT(wait); 1576 struct reclaim_state reclaim_state = { 1577 .reclaimed_slab = 0, 1578 }; 1579 cpumask_t cpumask; 1580 1581 cpumask = node_to_cpumask(pgdat->node_id); 1582 if (!cpus_empty(cpumask)) 1583 set_cpus_allowed(tsk, cpumask); 1584 current->reclaim_state = &reclaim_state; 1585 1586 /* 1587 * Tell the memory management that we're a "memory allocator", 1588 * and that if we need more memory we should get access to it 1589 * regardless (see "__alloc_pages()"). "kswapd" should 1590 * never get caught in the normal page freeing logic. 1591 * 1592 * (Kswapd normally doesn't need memory anyway, but sometimes 1593 * you need a small amount of memory in order to be able to 1594 * page out something else, and this flag essentially protects 1595 * us from recursively trying to free more memory as we're 1596 * trying to free the first piece of memory in the first place). 1597 */ 1598 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; 1599 set_freezable(); 1600 1601 order = 0; 1602 for ( ; ; ) { 1603 unsigned long new_order; 1604 1605 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 1606 new_order = pgdat->kswapd_max_order; 1607 pgdat->kswapd_max_order = 0; 1608 if (order < new_order) { 1609 /* 1610 * Don't sleep if someone wants a larger 'order' 1611 * allocation 1612 */ 1613 order = new_order; 1614 } else { 1615 if (!freezing(current)) 1616 schedule(); 1617 1618 order = pgdat->kswapd_max_order; 1619 } 1620 finish_wait(&pgdat->kswapd_wait, &wait); 1621 1622 if (!try_to_freeze()) { 1623 /* We can speed up thawing tasks if we don't call 1624 * balance_pgdat after returning from the refrigerator 1625 */ 1626 balance_pgdat(pgdat, order); 1627 } 1628 } 1629 return 0; 1630} 1631 1632/* 1633 * A zone is low on free memory, so wake its kswapd task to service it. 1634 */ 1635void wakeup_kswapd(struct zone *zone, int order) 1636{ 1637 pg_data_t *pgdat; 1638 1639 if (!populated_zone(zone)) 1640 return; 1641 1642 pgdat = zone->zone_pgdat; 1643 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0)) 1644 return; 1645 if (pgdat->kswapd_max_order < order) 1646 pgdat->kswapd_max_order = order; 1647 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL)) 1648 return; 1649 if (!waitqueue_active(&pgdat->kswapd_wait)) 1650 return; 1651 wake_up_interruptible(&pgdat->kswapd_wait); 1652} 1653 1654#ifdef CONFIG_PM 1655/* 1656 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages 1657 * from LRU lists system-wide, for given pass and priority, and returns the 1658 * number of reclaimed pages 1659 * 1660 * For pass > 3 we also try to shrink the LRU lists that contain a few pages 1661 */ 1662static unsigned long shrink_all_zones(unsigned long nr_pages, int prio, 1663 int pass, struct scan_control *sc) 1664{ 1665 struct zone *zone; 1666 unsigned long nr_to_scan, ret = 0; 1667 1668 for_each_zone(zone) { 1669 1670 if (!populated_zone(zone)) 1671 continue; 1672 1673 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY) 1674 continue; 1675 1676 /* For pass = 0 we don't shrink the active list */ 1677 if (pass > 0) { 1678 zone->nr_scan_active += 1679 (zone_page_state(zone, NR_ACTIVE) >> prio) + 1; 1680 if (zone->nr_scan_active >= nr_pages || pass > 3) { 1681 zone->nr_scan_active = 0; 1682 nr_to_scan = min(nr_pages, 1683 zone_page_state(zone, NR_ACTIVE)); 1684 shrink_active_list(nr_to_scan, zone, sc, prio); 1685 } 1686 } 1687 1688 zone->nr_scan_inactive += 1689 (zone_page_state(zone, NR_INACTIVE) >> prio) + 1; 1690 if (zone->nr_scan_inactive >= nr_pages || pass > 3) { 1691 zone->nr_scan_inactive = 0; 1692 nr_to_scan = min(nr_pages, 1693 zone_page_state(zone, NR_INACTIVE)); 1694 ret += shrink_inactive_list(nr_to_scan, zone, sc); 1695 if (ret >= nr_pages) 1696 return ret; 1697 } 1698 } 1699 1700 return ret; 1701} 1702 1703static unsigned long count_lru_pages(void) 1704{ 1705 return global_page_state(NR_ACTIVE) + global_page_state(NR_INACTIVE); 1706} 1707 1708/* 1709 * Try to free `nr_pages' of memory, system-wide, and return the number of 1710 * freed pages. 1711 * 1712 * Rather than trying to age LRUs the aim is to preserve the overall 1713 * LRU order by reclaiming preferentially 1714 * inactive > active > active referenced > active mapped 1715 */ 1716unsigned long shrink_all_memory(unsigned long nr_pages) 1717{ 1718 unsigned long lru_pages, nr_slab; 1719 unsigned long ret = 0; 1720 int pass; 1721 struct reclaim_state reclaim_state; 1722 struct scan_control sc = { 1723 .gfp_mask = GFP_KERNEL, 1724 .may_swap = 0, 1725 .swap_cluster_max = nr_pages, 1726 .may_writepage = 1, 1727 .swappiness = vm_swappiness, 1728 .isolate_pages = isolate_pages_global, 1729 }; 1730 1731 current->reclaim_state = &reclaim_state; 1732 1733 lru_pages = count_lru_pages(); 1734 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE); 1735 /* If slab caches are huge, it's better to hit them first */ 1736 while (nr_slab >= lru_pages) { 1737 reclaim_state.reclaimed_slab = 0; 1738 shrink_slab(nr_pages, sc.gfp_mask, lru_pages); 1739 if (!reclaim_state.reclaimed_slab) 1740 break; 1741 1742 ret += reclaim_state.reclaimed_slab; 1743 if (ret >= nr_pages) 1744 goto out; 1745 1746 nr_slab -= reclaim_state.reclaimed_slab; 1747 } 1748 1749 /* 1750 * We try to shrink LRUs in 5 passes: 1751 * 0 = Reclaim from inactive_list only 1752 * 1 = Reclaim from active list but don't reclaim mapped 1753 * 2 = 2nd pass of type 1 1754 * 3 = Reclaim mapped (normal reclaim) 1755 * 4 = 2nd pass of type 3 1756 */ 1757 for (pass = 0; pass < 5; pass++) { 1758 int prio; 1759 1760 /* Force reclaiming mapped pages in the passes #3 and #4 */ 1761 if (pass > 2) { 1762 sc.may_swap = 1; 1763 sc.swappiness = 100; 1764 } 1765 1766 for (prio = DEF_PRIORITY; prio >= 0; prio--) { 1767 unsigned long nr_to_scan = nr_pages - ret; 1768 1769 sc.nr_scanned = 0; 1770 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc); 1771 if (ret >= nr_pages) 1772 goto out; 1773 1774 reclaim_state.reclaimed_slab = 0; 1775 shrink_slab(sc.nr_scanned, sc.gfp_mask, 1776 count_lru_pages()); 1777 ret += reclaim_state.reclaimed_slab; 1778 if (ret >= nr_pages) 1779 goto out; 1780 1781 if (sc.nr_scanned && prio < DEF_PRIORITY - 2) 1782 congestion_wait(WRITE, HZ / 10); 1783 } 1784 } 1785 1786 /* 1787 * If ret = 0, we could not shrink LRUs, but there may be something 1788 * in slab caches 1789 */ 1790 if (!ret) { 1791 do { 1792 reclaim_state.reclaimed_slab = 0; 1793 shrink_slab(nr_pages, sc.gfp_mask, count_lru_pages()); 1794 ret += reclaim_state.reclaimed_slab; 1795 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0); 1796 } 1797 1798out: 1799 current->reclaim_state = NULL; 1800 1801 return ret; 1802} 1803#endif 1804 1805/* It's optimal to keep kswapds on the same CPUs as their memory, but 1806 not required for correctness. So if the last cpu in a node goes 1807 away, we get changed to run anywhere: as the first one comes back, 1808 restore their cpu bindings. */ 1809static int __devinit cpu_callback(struct notifier_block *nfb, 1810 unsigned long action, void *hcpu) 1811{ 1812 pg_data_t *pgdat; 1813 cpumask_t mask; 1814 int nid; 1815 1816 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) { 1817 for_each_node_state(nid, N_HIGH_MEMORY) { 1818 pgdat = NODE_DATA(nid); 1819 mask = node_to_cpumask(pgdat->node_id); 1820 if (any_online_cpu(mask) != NR_CPUS) 1821 /* One of our CPUs online: restore mask */ 1822 set_cpus_allowed(pgdat->kswapd, mask); 1823 } 1824 } 1825 return NOTIFY_OK; 1826} 1827 1828/* 1829 * This kswapd start function will be called by init and node-hot-add. 1830 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. 1831 */ 1832int kswapd_run(int nid) 1833{ 1834 pg_data_t *pgdat = NODE_DATA(nid); 1835 int ret = 0; 1836 1837 if (pgdat->kswapd) 1838 return 0; 1839 1840 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); 1841 if (IS_ERR(pgdat->kswapd)) { 1842 /* failure at boot is fatal */ 1843 BUG_ON(system_state == SYSTEM_BOOTING); 1844 printk("Failed to start kswapd on node %d\n",nid); 1845 ret = -1; 1846 } 1847 return ret; 1848} 1849 1850static int __init kswapd_init(void) 1851{ 1852 int nid; 1853 1854 swap_setup(); 1855 for_each_node_state(nid, N_HIGH_MEMORY) 1856 kswapd_run(nid); 1857 hotcpu_notifier(cpu_callback, 0); 1858 return 0; 1859} 1860 1861module_init(kswapd_init) 1862 1863#ifdef CONFIG_NUMA 1864/* 1865 * Zone reclaim mode 1866 * 1867 * If non-zero call zone_reclaim when the number of free pages falls below 1868 * the watermarks. 1869 */ 1870int zone_reclaim_mode __read_mostly; 1871 1872#define RECLAIM_OFF 0 1873#define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */ 1874#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ 1875#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ 1876 1877/* 1878 * Priority for ZONE_RECLAIM. This determines the fraction of pages 1879 * of a node considered for each zone_reclaim. 4 scans 1/16th of 1880 * a zone. 1881 */ 1882#define ZONE_RECLAIM_PRIORITY 4 1883 1884/* 1885 * Percentage of pages in a zone that must be unmapped for zone_reclaim to 1886 * occur. 1887 */ 1888int sysctl_min_unmapped_ratio = 1; 1889 1890/* 1891 * If the number of slab pages in a zone grows beyond this percentage then 1892 * slab reclaim needs to occur. 1893 */ 1894int sysctl_min_slab_ratio = 5; 1895 1896/* 1897 * Try to free up some pages from this zone through reclaim. 1898 */ 1899static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 1900{ 1901 /* Minimum pages needed in order to stay on node */ 1902 const unsigned long nr_pages = 1 << order; 1903 struct task_struct *p = current; 1904 struct reclaim_state reclaim_state; 1905 int priority; 1906 unsigned long nr_reclaimed = 0; 1907 struct scan_control sc = { 1908 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE), 1909 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP), 1910 .swap_cluster_max = max_t(unsigned long, nr_pages, 1911 SWAP_CLUSTER_MAX), 1912 .gfp_mask = gfp_mask, 1913 .swappiness = vm_swappiness, 1914 .isolate_pages = isolate_pages_global, 1915 }; 1916 unsigned long slab_reclaimable; 1917 1918 disable_swap_token(); 1919 cond_resched(); 1920 /* 1921 * We need to be able to allocate from the reserves for RECLAIM_SWAP 1922 * and we also need to be able to write out pages for RECLAIM_WRITE 1923 * and RECLAIM_SWAP. 1924 */ 1925 p->flags |= PF_MEMALLOC | PF_SWAPWRITE; 1926 reclaim_state.reclaimed_slab = 0; 1927 p->reclaim_state = &reclaim_state; 1928 1929 if (zone_page_state(zone, NR_FILE_PAGES) - 1930 zone_page_state(zone, NR_FILE_MAPPED) > 1931 zone->min_unmapped_pages) { 1932 /* 1933 * Free memory by calling shrink zone with increasing 1934 * priorities until we have enough memory freed. 1935 */ 1936 priority = ZONE_RECLAIM_PRIORITY; 1937 do { 1938 note_zone_scanning_priority(zone, priority); 1939 nr_reclaimed += shrink_zone(priority, zone, &sc); 1940 priority--; 1941 } while (priority >= 0 && nr_reclaimed < nr_pages); 1942 } 1943 1944 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE); 1945 if (slab_reclaimable > zone->min_slab_pages) { 1946 /* 1947 * shrink_slab() does not currently allow us to determine how 1948 * many pages were freed in this zone. So we take the current 1949 * number of slab pages and shake the slab until it is reduced 1950 * by the same nr_pages that we used for reclaiming unmapped 1951 * pages. 1952 * 1953 * Note that shrink_slab will free memory on all zones and may 1954 * take a long time. 1955 */ 1956 while (shrink_slab(sc.nr_scanned, gfp_mask, order) && 1957 zone_page_state(zone, NR_SLAB_RECLAIMABLE) > 1958 slab_reclaimable - nr_pages) 1959 ; 1960 1961 /* 1962 * Update nr_reclaimed by the number of slab pages we 1963 * reclaimed from this zone. 1964 */ 1965 nr_reclaimed += slab_reclaimable - 1966 zone_page_state(zone, NR_SLAB_RECLAIMABLE); 1967 } 1968 1969 p->reclaim_state = NULL; 1970 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); 1971 return nr_reclaimed >= nr_pages; 1972} 1973 1974int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 1975{ 1976 int node_id; 1977 int ret; 1978 1979 /* 1980 * Zone reclaim reclaims unmapped file backed pages and 1981 * slab pages if we are over the defined limits. 1982 * 1983 * A small portion of unmapped file backed pages is needed for 1984 * file I/O otherwise pages read by file I/O will be immediately 1985 * thrown out if the zone is overallocated. So we do not reclaim 1986 * if less than a specified percentage of the zone is used by 1987 * unmapped file backed pages. 1988 */ 1989 if (zone_page_state(zone, NR_FILE_PAGES) - 1990 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages 1991 && zone_page_state(zone, NR_SLAB_RECLAIMABLE) 1992 <= zone->min_slab_pages) 1993 return 0; 1994 1995 if (zone_is_all_unreclaimable(zone)) 1996 return 0; 1997 1998 /* 1999 * Do not scan if the allocation should not be delayed. 2000 */ 2001 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC)) 2002 return 0; 2003 2004 /* 2005 * Only run zone reclaim on the local zone or on zones that do not 2006 * have associated processors. This will favor the local processor 2007 * over remote processors and spread off node memory allocations 2008 * as wide as possible. 2009 */ 2010 node_id = zone_to_nid(zone); 2011 if (node_state(node_id, N_CPU) && node_id != numa_node_id()) 2012 return 0; 2013 2014 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED)) 2015 return 0; 2016 ret = __zone_reclaim(zone, gfp_mask, order); 2017 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED); 2018 2019 return ret; 2020} 2021#endif 2022