vmscan.c revision 8695949a1d7c99e039595db00af8e0fe4722307d
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/file.h> 23#include <linux/writeback.h> 24#include <linux/blkdev.h> 25#include <linux/buffer_head.h> /* for try_to_release_page(), 26 buffer_heads_over_limit */ 27#include <linux/mm_inline.h> 28#include <linux/pagevec.h> 29#include <linux/backing-dev.h> 30#include <linux/rmap.h> 31#include <linux/topology.h> 32#include <linux/cpu.h> 33#include <linux/cpuset.h> 34#include <linux/notifier.h> 35#include <linux/rwsem.h> 36 37#include <asm/tlbflush.h> 38#include <asm/div64.h> 39 40#include <linux/swapops.h> 41 42/* possible outcome of pageout() */ 43typedef enum { 44 /* failed to write page out, page is locked */ 45 PAGE_KEEP, 46 /* move page to the active list, page is locked */ 47 PAGE_ACTIVATE, 48 /* page has been sent to the disk successfully, page is unlocked */ 49 PAGE_SUCCESS, 50 /* page is clean and locked */ 51 PAGE_CLEAN, 52} pageout_t; 53 54struct scan_control { 55 /* Incremented by the number of inactive pages that were scanned */ 56 unsigned long nr_scanned; 57 58 /* Incremented by the number of pages reclaimed */ 59 unsigned long nr_reclaimed; 60 61 unsigned long nr_mapped; /* From page_state */ 62 63 /* This context's GFP mask */ 64 gfp_t gfp_mask; 65 66 int may_writepage; 67 68 /* Can pages be swapped as part of reclaim? */ 69 int may_swap; 70 71 /* This context's SWAP_CLUSTER_MAX. If freeing memory for 72 * suspend, we effectively ignore SWAP_CLUSTER_MAX. 73 * In this context, it doesn't matter that we scan the 74 * whole list at once. */ 75 int swap_cluster_max; 76}; 77 78/* 79 * The list of shrinker callbacks used by to apply pressure to 80 * ageable caches. 81 */ 82struct shrinker { 83 shrinker_t shrinker; 84 struct list_head list; 85 int seeks; /* seeks to recreate an obj */ 86 long nr; /* objs pending delete */ 87}; 88 89#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) 90 91#ifdef ARCH_HAS_PREFETCH 92#define prefetch_prev_lru_page(_page, _base, _field) \ 93 do { \ 94 if ((_page)->lru.prev != _base) { \ 95 struct page *prev; \ 96 \ 97 prev = lru_to_page(&(_page->lru)); \ 98 prefetch(&prev->_field); \ 99 } \ 100 } while (0) 101#else 102#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) 103#endif 104 105#ifdef ARCH_HAS_PREFETCHW 106#define prefetchw_prev_lru_page(_page, _base, _field) \ 107 do { \ 108 if ((_page)->lru.prev != _base) { \ 109 struct page *prev; \ 110 \ 111 prev = lru_to_page(&(_page->lru)); \ 112 prefetchw(&prev->_field); \ 113 } \ 114 } while (0) 115#else 116#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) 117#endif 118 119/* 120 * From 0 .. 100. Higher means more swappy. 121 */ 122int vm_swappiness = 60; 123static long total_memory; 124 125static LIST_HEAD(shrinker_list); 126static DECLARE_RWSEM(shrinker_rwsem); 127 128/* 129 * Add a shrinker callback to be called from the vm 130 */ 131struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker) 132{ 133 struct shrinker *shrinker; 134 135 shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL); 136 if (shrinker) { 137 shrinker->shrinker = theshrinker; 138 shrinker->seeks = seeks; 139 shrinker->nr = 0; 140 down_write(&shrinker_rwsem); 141 list_add_tail(&shrinker->list, &shrinker_list); 142 up_write(&shrinker_rwsem); 143 } 144 return shrinker; 145} 146EXPORT_SYMBOL(set_shrinker); 147 148/* 149 * Remove one 150 */ 151void remove_shrinker(struct shrinker *shrinker) 152{ 153 down_write(&shrinker_rwsem); 154 list_del(&shrinker->list); 155 up_write(&shrinker_rwsem); 156 kfree(shrinker); 157} 158EXPORT_SYMBOL(remove_shrinker); 159 160#define SHRINK_BATCH 128 161/* 162 * Call the shrink functions to age shrinkable caches 163 * 164 * Here we assume it costs one seek to replace a lru page and that it also 165 * takes a seek to recreate a cache object. With this in mind we age equal 166 * percentages of the lru and ageable caches. This should balance the seeks 167 * generated by these structures. 168 * 169 * If the vm encounted mapped pages on the LRU it increase the pressure on 170 * slab to avoid swapping. 171 * 172 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. 173 * 174 * `lru_pages' represents the number of on-LRU pages in all the zones which 175 * are eligible for the caller's allocation attempt. It is used for balancing 176 * slab reclaim versus page reclaim. 177 * 178 * Returns the number of slab objects which we shrunk. 179 */ 180int shrink_slab(unsigned long scanned, gfp_t gfp_mask, unsigned long lru_pages) 181{ 182 struct shrinker *shrinker; 183 int ret = 0; 184 185 if (scanned == 0) 186 scanned = SWAP_CLUSTER_MAX; 187 188 if (!down_read_trylock(&shrinker_rwsem)) 189 return 1; /* Assume we'll be able to shrink next time */ 190 191 list_for_each_entry(shrinker, &shrinker_list, list) { 192 unsigned long long delta; 193 unsigned long total_scan; 194 unsigned long max_pass = (*shrinker->shrinker)(0, gfp_mask); 195 196 delta = (4 * scanned) / shrinker->seeks; 197 delta *= max_pass; 198 do_div(delta, lru_pages + 1); 199 shrinker->nr += delta; 200 if (shrinker->nr < 0) { 201 printk(KERN_ERR "%s: nr=%ld\n", 202 __FUNCTION__, shrinker->nr); 203 shrinker->nr = max_pass; 204 } 205 206 /* 207 * Avoid risking looping forever due to too large nr value: 208 * never try to free more than twice the estimate number of 209 * freeable entries. 210 */ 211 if (shrinker->nr > max_pass * 2) 212 shrinker->nr = max_pass * 2; 213 214 total_scan = shrinker->nr; 215 shrinker->nr = 0; 216 217 while (total_scan >= SHRINK_BATCH) { 218 long this_scan = SHRINK_BATCH; 219 int shrink_ret; 220 int nr_before; 221 222 nr_before = (*shrinker->shrinker)(0, gfp_mask); 223 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask); 224 if (shrink_ret == -1) 225 break; 226 if (shrink_ret < nr_before) 227 ret += nr_before - shrink_ret; 228 mod_page_state(slabs_scanned, this_scan); 229 total_scan -= this_scan; 230 231 cond_resched(); 232 } 233 234 shrinker->nr += total_scan; 235 } 236 up_read(&shrinker_rwsem); 237 return ret; 238} 239 240/* Called without lock on whether page is mapped, so answer is unstable */ 241static inline int page_mapping_inuse(struct page *page) 242{ 243 struct address_space *mapping; 244 245 /* Page is in somebody's page tables. */ 246 if (page_mapped(page)) 247 return 1; 248 249 /* Be more reluctant to reclaim swapcache than pagecache */ 250 if (PageSwapCache(page)) 251 return 1; 252 253 mapping = page_mapping(page); 254 if (!mapping) 255 return 0; 256 257 /* File is mmap'd by somebody? */ 258 return mapping_mapped(mapping); 259} 260 261static inline int is_page_cache_freeable(struct page *page) 262{ 263 return page_count(page) - !!PagePrivate(page) == 2; 264} 265 266static int may_write_to_queue(struct backing_dev_info *bdi) 267{ 268 if (current->flags & PF_SWAPWRITE) 269 return 1; 270 if (!bdi_write_congested(bdi)) 271 return 1; 272 if (bdi == current->backing_dev_info) 273 return 1; 274 return 0; 275} 276 277/* 278 * We detected a synchronous write error writing a page out. Probably 279 * -ENOSPC. We need to propagate that into the address_space for a subsequent 280 * fsync(), msync() or close(). 281 * 282 * The tricky part is that after writepage we cannot touch the mapping: nothing 283 * prevents it from being freed up. But we have a ref on the page and once 284 * that page is locked, the mapping is pinned. 285 * 286 * We're allowed to run sleeping lock_page() here because we know the caller has 287 * __GFP_FS. 288 */ 289static void handle_write_error(struct address_space *mapping, 290 struct page *page, int error) 291{ 292 lock_page(page); 293 if (page_mapping(page) == mapping) { 294 if (error == -ENOSPC) 295 set_bit(AS_ENOSPC, &mapping->flags); 296 else 297 set_bit(AS_EIO, &mapping->flags); 298 } 299 unlock_page(page); 300} 301 302/* 303 * pageout is called by shrink_list() for each dirty page. Calls ->writepage(). 304 */ 305static pageout_t pageout(struct page *page, struct address_space *mapping) 306{ 307 /* 308 * If the page is dirty, only perform writeback if that write 309 * will be non-blocking. To prevent this allocation from being 310 * stalled by pagecache activity. But note that there may be 311 * stalls if we need to run get_block(). We could test 312 * PagePrivate for that. 313 * 314 * If this process is currently in generic_file_write() against 315 * this page's queue, we can perform writeback even if that 316 * will block. 317 * 318 * If the page is swapcache, write it back even if that would 319 * block, for some throttling. This happens by accident, because 320 * swap_backing_dev_info is bust: it doesn't reflect the 321 * congestion state of the swapdevs. Easy to fix, if needed. 322 * See swapfile.c:page_queue_congested(). 323 */ 324 if (!is_page_cache_freeable(page)) 325 return PAGE_KEEP; 326 if (!mapping) { 327 /* 328 * Some data journaling orphaned pages can have 329 * page->mapping == NULL while being dirty with clean buffers. 330 */ 331 if (PagePrivate(page)) { 332 if (try_to_free_buffers(page)) { 333 ClearPageDirty(page); 334 printk("%s: orphaned page\n", __FUNCTION__); 335 return PAGE_CLEAN; 336 } 337 } 338 return PAGE_KEEP; 339 } 340 if (mapping->a_ops->writepage == NULL) 341 return PAGE_ACTIVATE; 342 if (!may_write_to_queue(mapping->backing_dev_info)) 343 return PAGE_KEEP; 344 345 if (clear_page_dirty_for_io(page)) { 346 int res; 347 struct writeback_control wbc = { 348 .sync_mode = WB_SYNC_NONE, 349 .nr_to_write = SWAP_CLUSTER_MAX, 350 .nonblocking = 1, 351 .for_reclaim = 1, 352 }; 353 354 SetPageReclaim(page); 355 res = mapping->a_ops->writepage(page, &wbc); 356 if (res < 0) 357 handle_write_error(mapping, page, res); 358 if (res == AOP_WRITEPAGE_ACTIVATE) { 359 ClearPageReclaim(page); 360 return PAGE_ACTIVATE; 361 } 362 if (!PageWriteback(page)) { 363 /* synchronous write or broken a_ops? */ 364 ClearPageReclaim(page); 365 } 366 367 return PAGE_SUCCESS; 368 } 369 370 return PAGE_CLEAN; 371} 372 373static int remove_mapping(struct address_space *mapping, struct page *page) 374{ 375 if (!mapping) 376 return 0; /* truncate got there first */ 377 378 write_lock_irq(&mapping->tree_lock); 379 380 /* 381 * The non-racy check for busy page. It is critical to check 382 * PageDirty _after_ making sure that the page is freeable and 383 * not in use by anybody. (pagecache + us == 2) 384 */ 385 if (unlikely(page_count(page) != 2)) 386 goto cannot_free; 387 smp_rmb(); 388 if (unlikely(PageDirty(page))) 389 goto cannot_free; 390 391 if (PageSwapCache(page)) { 392 swp_entry_t swap = { .val = page_private(page) }; 393 __delete_from_swap_cache(page); 394 write_unlock_irq(&mapping->tree_lock); 395 swap_free(swap); 396 __put_page(page); /* The pagecache ref */ 397 return 1; 398 } 399 400 __remove_from_page_cache(page); 401 write_unlock_irq(&mapping->tree_lock); 402 __put_page(page); 403 return 1; 404 405cannot_free: 406 write_unlock_irq(&mapping->tree_lock); 407 return 0; 408} 409 410/* 411 * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed 412 */ 413static int shrink_list(struct list_head *page_list, struct scan_control *sc) 414{ 415 LIST_HEAD(ret_pages); 416 struct pagevec freed_pvec; 417 int pgactivate = 0; 418 int reclaimed = 0; 419 420 cond_resched(); 421 422 pagevec_init(&freed_pvec, 1); 423 while (!list_empty(page_list)) { 424 struct address_space *mapping; 425 struct page *page; 426 int may_enter_fs; 427 int referenced; 428 429 cond_resched(); 430 431 page = lru_to_page(page_list); 432 list_del(&page->lru); 433 434 if (TestSetPageLocked(page)) 435 goto keep; 436 437 BUG_ON(PageActive(page)); 438 439 sc->nr_scanned++; 440 441 if (!sc->may_swap && page_mapped(page)) 442 goto keep_locked; 443 444 /* Double the slab pressure for mapped and swapcache pages */ 445 if (page_mapped(page) || PageSwapCache(page)) 446 sc->nr_scanned++; 447 448 if (PageWriteback(page)) 449 goto keep_locked; 450 451 referenced = page_referenced(page, 1); 452 /* In active use or really unfreeable? Activate it. */ 453 if (referenced && page_mapping_inuse(page)) 454 goto activate_locked; 455 456#ifdef CONFIG_SWAP 457 /* 458 * Anonymous process memory has backing store? 459 * Try to allocate it some swap space here. 460 */ 461 if (PageAnon(page) && !PageSwapCache(page)) { 462 if (!sc->may_swap) 463 goto keep_locked; 464 if (!add_to_swap(page, GFP_ATOMIC)) 465 goto activate_locked; 466 } 467#endif /* CONFIG_SWAP */ 468 469 mapping = page_mapping(page); 470 may_enter_fs = (sc->gfp_mask & __GFP_FS) || 471 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); 472 473 /* 474 * The page is mapped into the page tables of one or more 475 * processes. Try to unmap it here. 476 */ 477 if (page_mapped(page) && mapping) { 478 /* 479 * No unmapping if we do not swap 480 */ 481 if (!sc->may_swap) 482 goto keep_locked; 483 484 switch (try_to_unmap(page, 0)) { 485 case SWAP_FAIL: 486 goto activate_locked; 487 case SWAP_AGAIN: 488 goto keep_locked; 489 case SWAP_SUCCESS: 490 ; /* try to free the page below */ 491 } 492 } 493 494 if (PageDirty(page)) { 495 if (referenced) 496 goto keep_locked; 497 if (!may_enter_fs) 498 goto keep_locked; 499 if (!sc->may_writepage) 500 goto keep_locked; 501 502 /* Page is dirty, try to write it out here */ 503 switch(pageout(page, mapping)) { 504 case PAGE_KEEP: 505 goto keep_locked; 506 case PAGE_ACTIVATE: 507 goto activate_locked; 508 case PAGE_SUCCESS: 509 if (PageWriteback(page) || PageDirty(page)) 510 goto keep; 511 /* 512 * A synchronous write - probably a ramdisk. Go 513 * ahead and try to reclaim the page. 514 */ 515 if (TestSetPageLocked(page)) 516 goto keep; 517 if (PageDirty(page) || PageWriteback(page)) 518 goto keep_locked; 519 mapping = page_mapping(page); 520 case PAGE_CLEAN: 521 ; /* try to free the page below */ 522 } 523 } 524 525 /* 526 * If the page has buffers, try to free the buffer mappings 527 * associated with this page. If we succeed we try to free 528 * the page as well. 529 * 530 * We do this even if the page is PageDirty(). 531 * try_to_release_page() does not perform I/O, but it is 532 * possible for a page to have PageDirty set, but it is actually 533 * clean (all its buffers are clean). This happens if the 534 * buffers were written out directly, with submit_bh(). ext3 535 * will do this, as well as the blockdev mapping. 536 * try_to_release_page() will discover that cleanness and will 537 * drop the buffers and mark the page clean - it can be freed. 538 * 539 * Rarely, pages can have buffers and no ->mapping. These are 540 * the pages which were not successfully invalidated in 541 * truncate_complete_page(). We try to drop those buffers here 542 * and if that worked, and the page is no longer mapped into 543 * process address space (page_count == 1) it can be freed. 544 * Otherwise, leave the page on the LRU so it is swappable. 545 */ 546 if (PagePrivate(page)) { 547 if (!try_to_release_page(page, sc->gfp_mask)) 548 goto activate_locked; 549 if (!mapping && page_count(page) == 1) 550 goto free_it; 551 } 552 553 if (!remove_mapping(mapping, page)) 554 goto keep_locked; 555 556free_it: 557 unlock_page(page); 558 reclaimed++; 559 if (!pagevec_add(&freed_pvec, page)) 560 __pagevec_release_nonlru(&freed_pvec); 561 continue; 562 563activate_locked: 564 SetPageActive(page); 565 pgactivate++; 566keep_locked: 567 unlock_page(page); 568keep: 569 list_add(&page->lru, &ret_pages); 570 BUG_ON(PageLRU(page)); 571 } 572 list_splice(&ret_pages, page_list); 573 if (pagevec_count(&freed_pvec)) 574 __pagevec_release_nonlru(&freed_pvec); 575 mod_page_state(pgactivate, pgactivate); 576 sc->nr_reclaimed += reclaimed; 577 return reclaimed; 578} 579 580#ifdef CONFIG_MIGRATION 581static inline void move_to_lru(struct page *page) 582{ 583 list_del(&page->lru); 584 if (PageActive(page)) { 585 /* 586 * lru_cache_add_active checks that 587 * the PG_active bit is off. 588 */ 589 ClearPageActive(page); 590 lru_cache_add_active(page); 591 } else { 592 lru_cache_add(page); 593 } 594 put_page(page); 595} 596 597/* 598 * Add isolated pages on the list back to the LRU. 599 * 600 * returns the number of pages put back. 601 */ 602int putback_lru_pages(struct list_head *l) 603{ 604 struct page *page; 605 struct page *page2; 606 int count = 0; 607 608 list_for_each_entry_safe(page, page2, l, lru) { 609 move_to_lru(page); 610 count++; 611 } 612 return count; 613} 614 615/* 616 * Non migratable page 617 */ 618int fail_migrate_page(struct page *newpage, struct page *page) 619{ 620 return -EIO; 621} 622EXPORT_SYMBOL(fail_migrate_page); 623 624/* 625 * swapout a single page 626 * page is locked upon entry, unlocked on exit 627 */ 628static int swap_page(struct page *page) 629{ 630 struct address_space *mapping = page_mapping(page); 631 632 if (page_mapped(page) && mapping) 633 if (try_to_unmap(page, 1) != SWAP_SUCCESS) 634 goto unlock_retry; 635 636 if (PageDirty(page)) { 637 /* Page is dirty, try to write it out here */ 638 switch(pageout(page, mapping)) { 639 case PAGE_KEEP: 640 case PAGE_ACTIVATE: 641 goto unlock_retry; 642 643 case PAGE_SUCCESS: 644 goto retry; 645 646 case PAGE_CLEAN: 647 ; /* try to free the page below */ 648 } 649 } 650 651 if (PagePrivate(page)) { 652 if (!try_to_release_page(page, GFP_KERNEL) || 653 (!mapping && page_count(page) == 1)) 654 goto unlock_retry; 655 } 656 657 if (remove_mapping(mapping, page)) { 658 /* Success */ 659 unlock_page(page); 660 return 0; 661 } 662 663unlock_retry: 664 unlock_page(page); 665 666retry: 667 return -EAGAIN; 668} 669EXPORT_SYMBOL(swap_page); 670 671/* 672 * Page migration was first developed in the context of the memory hotplug 673 * project. The main authors of the migration code are: 674 * 675 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> 676 * Hirokazu Takahashi <taka@valinux.co.jp> 677 * Dave Hansen <haveblue@us.ibm.com> 678 * Christoph Lameter <clameter@sgi.com> 679 */ 680 681/* 682 * Remove references for a page and establish the new page with the correct 683 * basic settings to be able to stop accesses to the page. 684 */ 685int migrate_page_remove_references(struct page *newpage, 686 struct page *page, int nr_refs) 687{ 688 struct address_space *mapping = page_mapping(page); 689 struct page **radix_pointer; 690 691 /* 692 * Avoid doing any of the following work if the page count 693 * indicates that the page is in use or truncate has removed 694 * the page. 695 */ 696 if (!mapping || page_mapcount(page) + nr_refs != page_count(page)) 697 return -EAGAIN; 698 699 /* 700 * Establish swap ptes for anonymous pages or destroy pte 701 * maps for files. 702 * 703 * In order to reestablish file backed mappings the fault handlers 704 * will take the radix tree_lock which may then be used to stop 705 * processses from accessing this page until the new page is ready. 706 * 707 * A process accessing via a swap pte (an anonymous page) will take a 708 * page_lock on the old page which will block the process until the 709 * migration attempt is complete. At that time the PageSwapCache bit 710 * will be examined. If the page was migrated then the PageSwapCache 711 * bit will be clear and the operation to retrieve the page will be 712 * retried which will find the new page in the radix tree. Then a new 713 * direct mapping may be generated based on the radix tree contents. 714 * 715 * If the page was not migrated then the PageSwapCache bit 716 * is still set and the operation may continue. 717 */ 718 if (try_to_unmap(page, 1) == SWAP_FAIL) 719 /* A vma has VM_LOCKED set -> Permanent failure */ 720 return -EPERM; 721 722 /* 723 * Give up if we were unable to remove all mappings. 724 */ 725 if (page_mapcount(page)) 726 return -EAGAIN; 727 728 write_lock_irq(&mapping->tree_lock); 729 730 radix_pointer = (struct page **)radix_tree_lookup_slot( 731 &mapping->page_tree, 732 page_index(page)); 733 734 if (!page_mapping(page) || page_count(page) != nr_refs || 735 *radix_pointer != page) { 736 write_unlock_irq(&mapping->tree_lock); 737 return -EAGAIN; 738 } 739 740 /* 741 * Now we know that no one else is looking at the page. 742 * 743 * Certain minimal information about a page must be available 744 * in order for other subsystems to properly handle the page if they 745 * find it through the radix tree update before we are finished 746 * copying the page. 747 */ 748 get_page(newpage); 749 newpage->index = page->index; 750 newpage->mapping = page->mapping; 751 if (PageSwapCache(page)) { 752 SetPageSwapCache(newpage); 753 set_page_private(newpage, page_private(page)); 754 } 755 756 *radix_pointer = newpage; 757 __put_page(page); 758 write_unlock_irq(&mapping->tree_lock); 759 760 return 0; 761} 762EXPORT_SYMBOL(migrate_page_remove_references); 763 764/* 765 * Copy the page to its new location 766 */ 767void migrate_page_copy(struct page *newpage, struct page *page) 768{ 769 copy_highpage(newpage, page); 770 771 if (PageError(page)) 772 SetPageError(newpage); 773 if (PageReferenced(page)) 774 SetPageReferenced(newpage); 775 if (PageUptodate(page)) 776 SetPageUptodate(newpage); 777 if (PageActive(page)) 778 SetPageActive(newpage); 779 if (PageChecked(page)) 780 SetPageChecked(newpage); 781 if (PageMappedToDisk(page)) 782 SetPageMappedToDisk(newpage); 783 784 if (PageDirty(page)) { 785 clear_page_dirty_for_io(page); 786 set_page_dirty(newpage); 787 } 788 789 ClearPageSwapCache(page); 790 ClearPageActive(page); 791 ClearPagePrivate(page); 792 set_page_private(page, 0); 793 page->mapping = NULL; 794 795 /* 796 * If any waiters have accumulated on the new page then 797 * wake them up. 798 */ 799 if (PageWriteback(newpage)) 800 end_page_writeback(newpage); 801} 802EXPORT_SYMBOL(migrate_page_copy); 803 804/* 805 * Common logic to directly migrate a single page suitable for 806 * pages that do not use PagePrivate. 807 * 808 * Pages are locked upon entry and exit. 809 */ 810int migrate_page(struct page *newpage, struct page *page) 811{ 812 int rc; 813 814 BUG_ON(PageWriteback(page)); /* Writeback must be complete */ 815 816 rc = migrate_page_remove_references(newpage, page, 2); 817 818 if (rc) 819 return rc; 820 821 migrate_page_copy(newpage, page); 822 823 /* 824 * Remove auxiliary swap entries and replace 825 * them with real ptes. 826 * 827 * Note that a real pte entry will allow processes that are not 828 * waiting on the page lock to use the new page via the page tables 829 * before the new page is unlocked. 830 */ 831 remove_from_swap(newpage); 832 return 0; 833} 834EXPORT_SYMBOL(migrate_page); 835 836/* 837 * migrate_pages 838 * 839 * Two lists are passed to this function. The first list 840 * contains the pages isolated from the LRU to be migrated. 841 * The second list contains new pages that the pages isolated 842 * can be moved to. If the second list is NULL then all 843 * pages are swapped out. 844 * 845 * The function returns after 10 attempts or if no pages 846 * are movable anymore because to has become empty 847 * or no retryable pages exist anymore. 848 * 849 * Return: Number of pages not migrated when "to" ran empty. 850 */ 851int migrate_pages(struct list_head *from, struct list_head *to, 852 struct list_head *moved, struct list_head *failed) 853{ 854 int retry; 855 int nr_failed = 0; 856 int pass = 0; 857 struct page *page; 858 struct page *page2; 859 int swapwrite = current->flags & PF_SWAPWRITE; 860 int rc; 861 862 if (!swapwrite) 863 current->flags |= PF_SWAPWRITE; 864 865redo: 866 retry = 0; 867 868 list_for_each_entry_safe(page, page2, from, lru) { 869 struct page *newpage = NULL; 870 struct address_space *mapping; 871 872 cond_resched(); 873 874 rc = 0; 875 if (page_count(page) == 1) 876 /* page was freed from under us. So we are done. */ 877 goto next; 878 879 if (to && list_empty(to)) 880 break; 881 882 /* 883 * Skip locked pages during the first two passes to give the 884 * functions holding the lock time to release the page. Later we 885 * use lock_page() to have a higher chance of acquiring the 886 * lock. 887 */ 888 rc = -EAGAIN; 889 if (pass > 2) 890 lock_page(page); 891 else 892 if (TestSetPageLocked(page)) 893 goto next; 894 895 /* 896 * Only wait on writeback if we have already done a pass where 897 * we we may have triggered writeouts for lots of pages. 898 */ 899 if (pass > 0) { 900 wait_on_page_writeback(page); 901 } else { 902 if (PageWriteback(page)) 903 goto unlock_page; 904 } 905 906 /* 907 * Anonymous pages must have swap cache references otherwise 908 * the information contained in the page maps cannot be 909 * preserved. 910 */ 911 if (PageAnon(page) && !PageSwapCache(page)) { 912 if (!add_to_swap(page, GFP_KERNEL)) { 913 rc = -ENOMEM; 914 goto unlock_page; 915 } 916 } 917 918 if (!to) { 919 rc = swap_page(page); 920 goto next; 921 } 922 923 newpage = lru_to_page(to); 924 lock_page(newpage); 925 926 /* 927 * Pages are properly locked and writeback is complete. 928 * Try to migrate the page. 929 */ 930 mapping = page_mapping(page); 931 if (!mapping) 932 goto unlock_both; 933 934 if (mapping->a_ops->migratepage) { 935 /* 936 * Most pages have a mapping and most filesystems 937 * should provide a migration function. Anonymous 938 * pages are part of swap space which also has its 939 * own migration function. This is the most common 940 * path for page migration. 941 */ 942 rc = mapping->a_ops->migratepage(newpage, page); 943 goto unlock_both; 944 } 945 946 /* 947 * Default handling if a filesystem does not provide 948 * a migration function. We can only migrate clean 949 * pages so try to write out any dirty pages first. 950 */ 951 if (PageDirty(page)) { 952 switch (pageout(page, mapping)) { 953 case PAGE_KEEP: 954 case PAGE_ACTIVATE: 955 goto unlock_both; 956 957 case PAGE_SUCCESS: 958 unlock_page(newpage); 959 goto next; 960 961 case PAGE_CLEAN: 962 ; /* try to migrate the page below */ 963 } 964 } 965 966 /* 967 * Buffers are managed in a filesystem specific way. 968 * We must have no buffers or drop them. 969 */ 970 if (!page_has_buffers(page) || 971 try_to_release_page(page, GFP_KERNEL)) { 972 rc = migrate_page(newpage, page); 973 goto unlock_both; 974 } 975 976 /* 977 * On early passes with mapped pages simply 978 * retry. There may be a lock held for some 979 * buffers that may go away. Later 980 * swap them out. 981 */ 982 if (pass > 4) { 983 /* 984 * Persistently unable to drop buffers..... As a 985 * measure of last resort we fall back to 986 * swap_page(). 987 */ 988 unlock_page(newpage); 989 newpage = NULL; 990 rc = swap_page(page); 991 goto next; 992 } 993 994unlock_both: 995 unlock_page(newpage); 996 997unlock_page: 998 unlock_page(page); 999 1000next: 1001 if (rc == -EAGAIN) { 1002 retry++; 1003 } else if (rc) { 1004 /* Permanent failure */ 1005 list_move(&page->lru, failed); 1006 nr_failed++; 1007 } else { 1008 if (newpage) { 1009 /* Successful migration. Return page to LRU */ 1010 move_to_lru(newpage); 1011 } 1012 list_move(&page->lru, moved); 1013 } 1014 } 1015 if (retry && pass++ < 10) 1016 goto redo; 1017 1018 if (!swapwrite) 1019 current->flags &= ~PF_SWAPWRITE; 1020 1021 return nr_failed + retry; 1022} 1023 1024/* 1025 * Isolate one page from the LRU lists and put it on the 1026 * indicated list with elevated refcount. 1027 * 1028 * Result: 1029 * 0 = page not on LRU list 1030 * 1 = page removed from LRU list and added to the specified list. 1031 */ 1032int isolate_lru_page(struct page *page) 1033{ 1034 int ret = 0; 1035 1036 if (PageLRU(page)) { 1037 struct zone *zone = page_zone(page); 1038 spin_lock_irq(&zone->lru_lock); 1039 if (PageLRU(page)) { 1040 ret = 1; 1041 get_page(page); 1042 ClearPageLRU(page); 1043 if (PageActive(page)) 1044 del_page_from_active_list(zone, page); 1045 else 1046 del_page_from_inactive_list(zone, page); 1047 } 1048 spin_unlock_irq(&zone->lru_lock); 1049 } 1050 1051 return ret; 1052} 1053#endif 1054 1055/* 1056 * zone->lru_lock is heavily contended. Some of the functions that 1057 * shrink the lists perform better by taking out a batch of pages 1058 * and working on them outside the LRU lock. 1059 * 1060 * For pagecache intensive workloads, this function is the hottest 1061 * spot in the kernel (apart from copy_*_user functions). 1062 * 1063 * Appropriate locks must be held before calling this function. 1064 * 1065 * @nr_to_scan: The number of pages to look through on the list. 1066 * @src: The LRU list to pull pages off. 1067 * @dst: The temp list to put pages on to. 1068 * @scanned: The number of pages that were scanned. 1069 * 1070 * returns how many pages were moved onto *@dst. 1071 */ 1072static int isolate_lru_pages(int nr_to_scan, struct list_head *src, 1073 struct list_head *dst, int *scanned) 1074{ 1075 int nr_taken = 0; 1076 struct page *page; 1077 int scan = 0; 1078 1079 while (scan++ < nr_to_scan && !list_empty(src)) { 1080 struct list_head *target; 1081 page = lru_to_page(src); 1082 prefetchw_prev_lru_page(page, src, flags); 1083 1084 BUG_ON(!PageLRU(page)); 1085 1086 list_del(&page->lru); 1087 target = src; 1088 if (likely(get_page_unless_zero(page))) { 1089 /* 1090 * Be careful not to clear PageLRU until after we're 1091 * sure the page is not being freed elsewhere -- the 1092 * page release code relies on it. 1093 */ 1094 ClearPageLRU(page); 1095 target = dst; 1096 nr_taken++; 1097 } /* else it is being freed elsewhere */ 1098 1099 list_add(&page->lru, target); 1100 } 1101 1102 *scanned = scan; 1103 return nr_taken; 1104} 1105 1106/* 1107 * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed 1108 */ 1109static void shrink_cache(int max_scan, struct zone *zone, struct scan_control *sc) 1110{ 1111 LIST_HEAD(page_list); 1112 struct pagevec pvec; 1113 1114 pagevec_init(&pvec, 1); 1115 1116 lru_add_drain(); 1117 spin_lock_irq(&zone->lru_lock); 1118 while (max_scan > 0) { 1119 struct page *page; 1120 int nr_taken; 1121 int nr_scan; 1122 int nr_freed; 1123 1124 nr_taken = isolate_lru_pages(sc->swap_cluster_max, 1125 &zone->inactive_list, 1126 &page_list, &nr_scan); 1127 zone->nr_inactive -= nr_taken; 1128 zone->pages_scanned += nr_scan; 1129 spin_unlock_irq(&zone->lru_lock); 1130 1131 if (nr_taken == 0) 1132 goto done; 1133 1134 max_scan -= nr_scan; 1135 nr_freed = shrink_list(&page_list, sc); 1136 1137 local_irq_disable(); 1138 if (current_is_kswapd()) { 1139 __mod_page_state_zone(zone, pgscan_kswapd, nr_scan); 1140 __mod_page_state(kswapd_steal, nr_freed); 1141 } else 1142 __mod_page_state_zone(zone, pgscan_direct, nr_scan); 1143 __mod_page_state_zone(zone, pgsteal, nr_freed); 1144 1145 spin_lock(&zone->lru_lock); 1146 /* 1147 * Put back any unfreeable pages. 1148 */ 1149 while (!list_empty(&page_list)) { 1150 page = lru_to_page(&page_list); 1151 BUG_ON(PageLRU(page)); 1152 SetPageLRU(page); 1153 list_del(&page->lru); 1154 if (PageActive(page)) 1155 add_page_to_active_list(zone, page); 1156 else 1157 add_page_to_inactive_list(zone, page); 1158 if (!pagevec_add(&pvec, page)) { 1159 spin_unlock_irq(&zone->lru_lock); 1160 __pagevec_release(&pvec); 1161 spin_lock_irq(&zone->lru_lock); 1162 } 1163 } 1164 } 1165 spin_unlock_irq(&zone->lru_lock); 1166done: 1167 pagevec_release(&pvec); 1168} 1169 1170/* 1171 * This moves pages from the active list to the inactive list. 1172 * 1173 * We move them the other way if the page is referenced by one or more 1174 * processes, from rmap. 1175 * 1176 * If the pages are mostly unmapped, the processing is fast and it is 1177 * appropriate to hold zone->lru_lock across the whole operation. But if 1178 * the pages are mapped, the processing is slow (page_referenced()) so we 1179 * should drop zone->lru_lock around each page. It's impossible to balance 1180 * this, so instead we remove the pages from the LRU while processing them. 1181 * It is safe to rely on PG_active against the non-LRU pages in here because 1182 * nobody will play with that bit on a non-LRU page. 1183 * 1184 * The downside is that we have to touch page->_count against each page. 1185 * But we had to alter page->flags anyway. 1186 */ 1187static void 1188refill_inactive_zone(int nr_pages, struct zone *zone, struct scan_control *sc) 1189{ 1190 int pgmoved; 1191 int pgdeactivate = 0; 1192 int pgscanned; 1193 LIST_HEAD(l_hold); /* The pages which were snipped off */ 1194 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */ 1195 LIST_HEAD(l_active); /* Pages to go onto the active_list */ 1196 struct page *page; 1197 struct pagevec pvec; 1198 int reclaim_mapped = 0; 1199 1200 if (unlikely(sc->may_swap)) { 1201 long mapped_ratio; 1202 long distress; 1203 long swap_tendency; 1204 1205 /* 1206 * `distress' is a measure of how much trouble we're having 1207 * reclaiming pages. 0 -> no problems. 100 -> great trouble. 1208 */ 1209 distress = 100 >> zone->prev_priority; 1210 1211 /* 1212 * The point of this algorithm is to decide when to start 1213 * reclaiming mapped memory instead of just pagecache. Work out 1214 * how much memory 1215 * is mapped. 1216 */ 1217 mapped_ratio = (sc->nr_mapped * 100) / total_memory; 1218 1219 /* 1220 * Now decide how much we really want to unmap some pages. The 1221 * mapped ratio is downgraded - just because there's a lot of 1222 * mapped memory doesn't necessarily mean that page reclaim 1223 * isn't succeeding. 1224 * 1225 * The distress ratio is important - we don't want to start 1226 * going oom. 1227 * 1228 * A 100% value of vm_swappiness overrides this algorithm 1229 * altogether. 1230 */ 1231 swap_tendency = mapped_ratio / 2 + distress + vm_swappiness; 1232 1233 /* 1234 * Now use this metric to decide whether to start moving mapped 1235 * memory onto the inactive list. 1236 */ 1237 if (swap_tendency >= 100) 1238 reclaim_mapped = 1; 1239 } 1240 1241 lru_add_drain(); 1242 spin_lock_irq(&zone->lru_lock); 1243 pgmoved = isolate_lru_pages(nr_pages, &zone->active_list, 1244 &l_hold, &pgscanned); 1245 zone->pages_scanned += pgscanned; 1246 zone->nr_active -= pgmoved; 1247 spin_unlock_irq(&zone->lru_lock); 1248 1249 while (!list_empty(&l_hold)) { 1250 cond_resched(); 1251 page = lru_to_page(&l_hold); 1252 list_del(&page->lru); 1253 if (page_mapped(page)) { 1254 if (!reclaim_mapped || 1255 (total_swap_pages == 0 && PageAnon(page)) || 1256 page_referenced(page, 0)) { 1257 list_add(&page->lru, &l_active); 1258 continue; 1259 } 1260 } 1261 list_add(&page->lru, &l_inactive); 1262 } 1263 1264 pagevec_init(&pvec, 1); 1265 pgmoved = 0; 1266 spin_lock_irq(&zone->lru_lock); 1267 while (!list_empty(&l_inactive)) { 1268 page = lru_to_page(&l_inactive); 1269 prefetchw_prev_lru_page(page, &l_inactive, flags); 1270 BUG_ON(PageLRU(page)); 1271 SetPageLRU(page); 1272 BUG_ON(!PageActive(page)); 1273 ClearPageActive(page); 1274 1275 list_move(&page->lru, &zone->inactive_list); 1276 pgmoved++; 1277 if (!pagevec_add(&pvec, page)) { 1278 zone->nr_inactive += pgmoved; 1279 spin_unlock_irq(&zone->lru_lock); 1280 pgdeactivate += pgmoved; 1281 pgmoved = 0; 1282 if (buffer_heads_over_limit) 1283 pagevec_strip(&pvec); 1284 __pagevec_release(&pvec); 1285 spin_lock_irq(&zone->lru_lock); 1286 } 1287 } 1288 zone->nr_inactive += pgmoved; 1289 pgdeactivate += pgmoved; 1290 if (buffer_heads_over_limit) { 1291 spin_unlock_irq(&zone->lru_lock); 1292 pagevec_strip(&pvec); 1293 spin_lock_irq(&zone->lru_lock); 1294 } 1295 1296 pgmoved = 0; 1297 while (!list_empty(&l_active)) { 1298 page = lru_to_page(&l_active); 1299 prefetchw_prev_lru_page(page, &l_active, flags); 1300 BUG_ON(PageLRU(page)); 1301 SetPageLRU(page); 1302 BUG_ON(!PageActive(page)); 1303 list_move(&page->lru, &zone->active_list); 1304 pgmoved++; 1305 if (!pagevec_add(&pvec, page)) { 1306 zone->nr_active += pgmoved; 1307 pgmoved = 0; 1308 spin_unlock_irq(&zone->lru_lock); 1309 __pagevec_release(&pvec); 1310 spin_lock_irq(&zone->lru_lock); 1311 } 1312 } 1313 zone->nr_active += pgmoved; 1314 spin_unlock(&zone->lru_lock); 1315 1316 __mod_page_state_zone(zone, pgrefill, pgscanned); 1317 __mod_page_state(pgdeactivate, pgdeactivate); 1318 local_irq_enable(); 1319 1320 pagevec_release(&pvec); 1321} 1322 1323/* 1324 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. 1325 */ 1326static void 1327shrink_zone(int priority, struct zone *zone, struct scan_control *sc) 1328{ 1329 unsigned long nr_active; 1330 unsigned long nr_inactive; 1331 unsigned long nr_to_scan; 1332 1333 atomic_inc(&zone->reclaim_in_progress); 1334 1335 /* 1336 * Add one to `nr_to_scan' just to make sure that the kernel will 1337 * slowly sift through the active list. 1338 */ 1339 zone->nr_scan_active += (zone->nr_active >> priority) + 1; 1340 nr_active = zone->nr_scan_active; 1341 if (nr_active >= sc->swap_cluster_max) 1342 zone->nr_scan_active = 0; 1343 else 1344 nr_active = 0; 1345 1346 zone->nr_scan_inactive += (zone->nr_inactive >> priority) + 1; 1347 nr_inactive = zone->nr_scan_inactive; 1348 if (nr_inactive >= sc->swap_cluster_max) 1349 zone->nr_scan_inactive = 0; 1350 else 1351 nr_inactive = 0; 1352 1353 while (nr_active || nr_inactive) { 1354 if (nr_active) { 1355 nr_to_scan = min(nr_active, 1356 (unsigned long)sc->swap_cluster_max); 1357 nr_active -= nr_to_scan; 1358 refill_inactive_zone(nr_to_scan, zone, sc); 1359 } 1360 1361 if (nr_inactive) { 1362 nr_to_scan = min(nr_inactive, 1363 (unsigned long)sc->swap_cluster_max); 1364 nr_inactive -= nr_to_scan; 1365 shrink_cache(nr_to_scan, zone, sc); 1366 } 1367 } 1368 1369 throttle_vm_writeout(); 1370 1371 atomic_dec(&zone->reclaim_in_progress); 1372} 1373 1374/* 1375 * This is the direct reclaim path, for page-allocating processes. We only 1376 * try to reclaim pages from zones which will satisfy the caller's allocation 1377 * request. 1378 * 1379 * We reclaim from a zone even if that zone is over pages_high. Because: 1380 * a) The caller may be trying to free *extra* pages to satisfy a higher-order 1381 * allocation or 1382 * b) The zones may be over pages_high but they must go *over* pages_high to 1383 * satisfy the `incremental min' zone defense algorithm. 1384 * 1385 * Returns the number of reclaimed pages. 1386 * 1387 * If a zone is deemed to be full of pinned pages then just give it a light 1388 * scan then give up on it. 1389 */ 1390static void 1391shrink_caches(int priority, struct zone **zones, struct scan_control *sc) 1392{ 1393 int i; 1394 1395 for (i = 0; zones[i] != NULL; i++) { 1396 struct zone *zone = zones[i]; 1397 1398 if (!populated_zone(zone)) 1399 continue; 1400 1401 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL)) 1402 continue; 1403 1404 zone->temp_priority = priority; 1405 if (zone->prev_priority > priority) 1406 zone->prev_priority = priority; 1407 1408 if (zone->all_unreclaimable && priority != DEF_PRIORITY) 1409 continue; /* Let kswapd poll it */ 1410 1411 shrink_zone(priority, zone, sc); 1412 } 1413} 1414 1415/* 1416 * This is the main entry point to direct page reclaim. 1417 * 1418 * If a full scan of the inactive list fails to free enough memory then we 1419 * are "out of memory" and something needs to be killed. 1420 * 1421 * If the caller is !__GFP_FS then the probability of a failure is reasonably 1422 * high - the zone may be full of dirty or under-writeback pages, which this 1423 * caller can't do much about. We kick pdflush and take explicit naps in the 1424 * hope that some of these pages can be written. But if the allocating task 1425 * holds filesystem locks which prevent writeout this might not work, and the 1426 * allocation attempt will fail. 1427 */ 1428int try_to_free_pages(struct zone **zones, gfp_t gfp_mask) 1429{ 1430 int priority; 1431 int ret = 0; 1432 int total_scanned = 0, total_reclaimed = 0; 1433 struct reclaim_state *reclaim_state = current->reclaim_state; 1434 struct scan_control sc; 1435 unsigned long lru_pages = 0; 1436 int i; 1437 1438 sc.gfp_mask = gfp_mask; 1439 sc.may_writepage = !laptop_mode; 1440 sc.may_swap = 1; 1441 1442 inc_page_state(allocstall); 1443 1444 for (i = 0; zones[i] != NULL; i++) { 1445 struct zone *zone = zones[i]; 1446 1447 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL)) 1448 continue; 1449 1450 zone->temp_priority = DEF_PRIORITY; 1451 lru_pages += zone->nr_active + zone->nr_inactive; 1452 } 1453 1454 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 1455 sc.nr_mapped = read_page_state(nr_mapped); 1456 sc.nr_scanned = 0; 1457 sc.nr_reclaimed = 0; 1458 sc.swap_cluster_max = SWAP_CLUSTER_MAX; 1459 if (!priority) 1460 disable_swap_token(); 1461 shrink_caches(priority, zones, &sc); 1462 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages); 1463 if (reclaim_state) { 1464 sc.nr_reclaimed += reclaim_state->reclaimed_slab; 1465 reclaim_state->reclaimed_slab = 0; 1466 } 1467 total_scanned += sc.nr_scanned; 1468 total_reclaimed += sc.nr_reclaimed; 1469 if (total_reclaimed >= sc.swap_cluster_max) { 1470 ret = 1; 1471 goto out; 1472 } 1473 1474 /* 1475 * Try to write back as many pages as we just scanned. This 1476 * tends to cause slow streaming writers to write data to the 1477 * disk smoothly, at the dirtying rate, which is nice. But 1478 * that's undesirable in laptop mode, where we *want* lumpy 1479 * writeout. So in laptop mode, write out the whole world. 1480 */ 1481 if (total_scanned > sc.swap_cluster_max + sc.swap_cluster_max/2) { 1482 wakeup_pdflush(laptop_mode ? 0 : total_scanned); 1483 sc.may_writepage = 1; 1484 } 1485 1486 /* Take a nap, wait for some writeback to complete */ 1487 if (sc.nr_scanned && priority < DEF_PRIORITY - 2) 1488 blk_congestion_wait(WRITE, HZ/10); 1489 } 1490out: 1491 for (i = 0; zones[i] != 0; i++) { 1492 struct zone *zone = zones[i]; 1493 1494 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL)) 1495 continue; 1496 1497 zone->prev_priority = zone->temp_priority; 1498 } 1499 return ret; 1500} 1501 1502/* 1503 * For kswapd, balance_pgdat() will work across all this node's zones until 1504 * they are all at pages_high. 1505 * 1506 * If `nr_pages' is non-zero then it is the number of pages which are to be 1507 * reclaimed, regardless of the zone occupancies. This is a software suspend 1508 * special. 1509 * 1510 * Returns the number of pages which were actually freed. 1511 * 1512 * There is special handling here for zones which are full of pinned pages. 1513 * This can happen if the pages are all mlocked, or if they are all used by 1514 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. 1515 * What we do is to detect the case where all pages in the zone have been 1516 * scanned twice and there has been zero successful reclaim. Mark the zone as 1517 * dead and from now on, only perform a short scan. Basically we're polling 1518 * the zone for when the problem goes away. 1519 * 1520 * kswapd scans the zones in the highmem->normal->dma direction. It skips 1521 * zones which have free_pages > pages_high, but once a zone is found to have 1522 * free_pages <= pages_high, we scan that zone and the lower zones regardless 1523 * of the number of free pages in the lower zones. This interoperates with 1524 * the page allocator fallback scheme to ensure that aging of pages is balanced 1525 * across the zones. 1526 */ 1527static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order) 1528{ 1529 int to_free = nr_pages; 1530 int all_zones_ok; 1531 int priority; 1532 int i; 1533 int total_scanned, total_reclaimed; 1534 struct reclaim_state *reclaim_state = current->reclaim_state; 1535 struct scan_control sc; 1536 1537loop_again: 1538 total_scanned = 0; 1539 total_reclaimed = 0; 1540 sc.gfp_mask = GFP_KERNEL; 1541 sc.may_writepage = !laptop_mode; 1542 sc.may_swap = 1; 1543 sc.nr_mapped = read_page_state(nr_mapped); 1544 1545 inc_page_state(pageoutrun); 1546 1547 for (i = 0; i < pgdat->nr_zones; i++) { 1548 struct zone *zone = pgdat->node_zones + i; 1549 1550 zone->temp_priority = DEF_PRIORITY; 1551 } 1552 1553 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 1554 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ 1555 unsigned long lru_pages = 0; 1556 1557 /* The swap token gets in the way of swapout... */ 1558 if (!priority) 1559 disable_swap_token(); 1560 1561 all_zones_ok = 1; 1562 1563 if (nr_pages == 0) { 1564 /* 1565 * Scan in the highmem->dma direction for the highest 1566 * zone which needs scanning 1567 */ 1568 for (i = pgdat->nr_zones - 1; i >= 0; i--) { 1569 struct zone *zone = pgdat->node_zones + i; 1570 1571 if (!populated_zone(zone)) 1572 continue; 1573 1574 if (zone->all_unreclaimable && 1575 priority != DEF_PRIORITY) 1576 continue; 1577 1578 if (!zone_watermark_ok(zone, order, 1579 zone->pages_high, 0, 0)) { 1580 end_zone = i; 1581 goto scan; 1582 } 1583 } 1584 goto out; 1585 } else { 1586 end_zone = pgdat->nr_zones - 1; 1587 } 1588scan: 1589 for (i = 0; i <= end_zone; i++) { 1590 struct zone *zone = pgdat->node_zones + i; 1591 1592 lru_pages += zone->nr_active + zone->nr_inactive; 1593 } 1594 1595 /* 1596 * Now scan the zone in the dma->highmem direction, stopping 1597 * at the last zone which needs scanning. 1598 * 1599 * We do this because the page allocator works in the opposite 1600 * direction. This prevents the page allocator from allocating 1601 * pages behind kswapd's direction of progress, which would 1602 * cause too much scanning of the lower zones. 1603 */ 1604 for (i = 0; i <= end_zone; i++) { 1605 struct zone *zone = pgdat->node_zones + i; 1606 int nr_slab; 1607 1608 if (!populated_zone(zone)) 1609 continue; 1610 1611 if (zone->all_unreclaimable && priority != DEF_PRIORITY) 1612 continue; 1613 1614 if (nr_pages == 0) { /* Not software suspend */ 1615 if (!zone_watermark_ok(zone, order, 1616 zone->pages_high, end_zone, 0)) 1617 all_zones_ok = 0; 1618 } 1619 zone->temp_priority = priority; 1620 if (zone->prev_priority > priority) 1621 zone->prev_priority = priority; 1622 sc.nr_scanned = 0; 1623 sc.nr_reclaimed = 0; 1624 sc.swap_cluster_max = nr_pages? nr_pages : SWAP_CLUSTER_MAX; 1625 shrink_zone(priority, zone, &sc); 1626 reclaim_state->reclaimed_slab = 0; 1627 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL, 1628 lru_pages); 1629 sc.nr_reclaimed += reclaim_state->reclaimed_slab; 1630 total_reclaimed += sc.nr_reclaimed; 1631 total_scanned += sc.nr_scanned; 1632 if (zone->all_unreclaimable) 1633 continue; 1634 if (nr_slab == 0 && zone->pages_scanned >= 1635 (zone->nr_active + zone->nr_inactive) * 4) 1636 zone->all_unreclaimable = 1; 1637 /* 1638 * If we've done a decent amount of scanning and 1639 * the reclaim ratio is low, start doing writepage 1640 * even in laptop mode 1641 */ 1642 if (total_scanned > SWAP_CLUSTER_MAX * 2 && 1643 total_scanned > total_reclaimed+total_reclaimed/2) 1644 sc.may_writepage = 1; 1645 } 1646 if (nr_pages && to_free > total_reclaimed) 1647 continue; /* swsusp: need to do more work */ 1648 if (all_zones_ok) 1649 break; /* kswapd: all done */ 1650 /* 1651 * OK, kswapd is getting into trouble. Take a nap, then take 1652 * another pass across the zones. 1653 */ 1654 if (total_scanned && priority < DEF_PRIORITY - 2) 1655 blk_congestion_wait(WRITE, HZ/10); 1656 1657 /* 1658 * We do this so kswapd doesn't build up large priorities for 1659 * example when it is freeing in parallel with allocators. It 1660 * matches the direct reclaim path behaviour in terms of impact 1661 * on zone->*_priority. 1662 */ 1663 if ((total_reclaimed >= SWAP_CLUSTER_MAX) && (!nr_pages)) 1664 break; 1665 } 1666out: 1667 for (i = 0; i < pgdat->nr_zones; i++) { 1668 struct zone *zone = pgdat->node_zones + i; 1669 1670 zone->prev_priority = zone->temp_priority; 1671 } 1672 if (!all_zones_ok) { 1673 cond_resched(); 1674 goto loop_again; 1675 } 1676 1677 return total_reclaimed; 1678} 1679 1680/* 1681 * The background pageout daemon, started as a kernel thread 1682 * from the init process. 1683 * 1684 * This basically trickles out pages so that we have _some_ 1685 * free memory available even if there is no other activity 1686 * that frees anything up. This is needed for things like routing 1687 * etc, where we otherwise might have all activity going on in 1688 * asynchronous contexts that cannot page things out. 1689 * 1690 * If there are applications that are active memory-allocators 1691 * (most normal use), this basically shouldn't matter. 1692 */ 1693static int kswapd(void *p) 1694{ 1695 unsigned long order; 1696 pg_data_t *pgdat = (pg_data_t*)p; 1697 struct task_struct *tsk = current; 1698 DEFINE_WAIT(wait); 1699 struct reclaim_state reclaim_state = { 1700 .reclaimed_slab = 0, 1701 }; 1702 cpumask_t cpumask; 1703 1704 daemonize("kswapd%d", pgdat->node_id); 1705 cpumask = node_to_cpumask(pgdat->node_id); 1706 if (!cpus_empty(cpumask)) 1707 set_cpus_allowed(tsk, cpumask); 1708 current->reclaim_state = &reclaim_state; 1709 1710 /* 1711 * Tell the memory management that we're a "memory allocator", 1712 * and that if we need more memory we should get access to it 1713 * regardless (see "__alloc_pages()"). "kswapd" should 1714 * never get caught in the normal page freeing logic. 1715 * 1716 * (Kswapd normally doesn't need memory anyway, but sometimes 1717 * you need a small amount of memory in order to be able to 1718 * page out something else, and this flag essentially protects 1719 * us from recursively trying to free more memory as we're 1720 * trying to free the first piece of memory in the first place). 1721 */ 1722 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; 1723 1724 order = 0; 1725 for ( ; ; ) { 1726 unsigned long new_order; 1727 1728 try_to_freeze(); 1729 1730 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 1731 new_order = pgdat->kswapd_max_order; 1732 pgdat->kswapd_max_order = 0; 1733 if (order < new_order) { 1734 /* 1735 * Don't sleep if someone wants a larger 'order' 1736 * allocation 1737 */ 1738 order = new_order; 1739 } else { 1740 schedule(); 1741 order = pgdat->kswapd_max_order; 1742 } 1743 finish_wait(&pgdat->kswapd_wait, &wait); 1744 1745 balance_pgdat(pgdat, 0, order); 1746 } 1747 return 0; 1748} 1749 1750/* 1751 * A zone is low on free memory, so wake its kswapd task to service it. 1752 */ 1753void wakeup_kswapd(struct zone *zone, int order) 1754{ 1755 pg_data_t *pgdat; 1756 1757 if (!populated_zone(zone)) 1758 return; 1759 1760 pgdat = zone->zone_pgdat; 1761 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0)) 1762 return; 1763 if (pgdat->kswapd_max_order < order) 1764 pgdat->kswapd_max_order = order; 1765 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL)) 1766 return; 1767 if (!waitqueue_active(&pgdat->kswapd_wait)) 1768 return; 1769 wake_up_interruptible(&pgdat->kswapd_wait); 1770} 1771 1772#ifdef CONFIG_PM 1773/* 1774 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed 1775 * pages. 1776 */ 1777int shrink_all_memory(int nr_pages) 1778{ 1779 pg_data_t *pgdat; 1780 int nr_to_free = nr_pages; 1781 int ret = 0; 1782 struct reclaim_state reclaim_state = { 1783 .reclaimed_slab = 0, 1784 }; 1785 1786 current->reclaim_state = &reclaim_state; 1787 for_each_pgdat(pgdat) { 1788 int freed; 1789 freed = balance_pgdat(pgdat, nr_to_free, 0); 1790 ret += freed; 1791 nr_to_free -= freed; 1792 if (nr_to_free <= 0) 1793 break; 1794 } 1795 current->reclaim_state = NULL; 1796 return ret; 1797} 1798#endif 1799 1800#ifdef CONFIG_HOTPLUG_CPU 1801/* It's optimal to keep kswapds on the same CPUs as their memory, but 1802 not required for correctness. So if the last cpu in a node goes 1803 away, we get changed to run anywhere: as the first one comes back, 1804 restore their cpu bindings. */ 1805static int __devinit cpu_callback(struct notifier_block *nfb, 1806 unsigned long action, 1807 void *hcpu) 1808{ 1809 pg_data_t *pgdat; 1810 cpumask_t mask; 1811 1812 if (action == CPU_ONLINE) { 1813 for_each_pgdat(pgdat) { 1814 mask = node_to_cpumask(pgdat->node_id); 1815 if (any_online_cpu(mask) != NR_CPUS) 1816 /* One of our CPUs online: restore mask */ 1817 set_cpus_allowed(pgdat->kswapd, mask); 1818 } 1819 } 1820 return NOTIFY_OK; 1821} 1822#endif /* CONFIG_HOTPLUG_CPU */ 1823 1824static int __init kswapd_init(void) 1825{ 1826 pg_data_t *pgdat; 1827 swap_setup(); 1828 for_each_pgdat(pgdat) 1829 pgdat->kswapd 1830 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL)); 1831 total_memory = nr_free_pagecache_pages(); 1832 hotcpu_notifier(cpu_callback, 0); 1833 return 0; 1834} 1835 1836module_init(kswapd_init) 1837 1838#ifdef CONFIG_NUMA 1839/* 1840 * Zone reclaim mode 1841 * 1842 * If non-zero call zone_reclaim when the number of free pages falls below 1843 * the watermarks. 1844 * 1845 * In the future we may add flags to the mode. However, the page allocator 1846 * should only have to check that zone_reclaim_mode != 0 before calling 1847 * zone_reclaim(). 1848 */ 1849int zone_reclaim_mode __read_mostly; 1850 1851#define RECLAIM_OFF 0 1852#define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */ 1853#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ 1854#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */ 1855#define RECLAIM_SLAB (1<<3) /* Do a global slab shrink if the zone is out of memory */ 1856 1857/* 1858 * Mininum time between zone reclaim scans 1859 */ 1860int zone_reclaim_interval __read_mostly = 30*HZ; 1861 1862/* 1863 * Priority for ZONE_RECLAIM. This determines the fraction of pages 1864 * of a node considered for each zone_reclaim. 4 scans 1/16th of 1865 * a zone. 1866 */ 1867#define ZONE_RECLAIM_PRIORITY 4 1868 1869/* 1870 * Try to free up some pages from this zone through reclaim. 1871 */ 1872int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order) 1873{ 1874 int nr_pages; 1875 struct task_struct *p = current; 1876 struct reclaim_state reclaim_state; 1877 struct scan_control sc; 1878 cpumask_t mask; 1879 int node_id; 1880 int priority; 1881 1882 if (time_before(jiffies, 1883 zone->last_unsuccessful_zone_reclaim + zone_reclaim_interval)) 1884 return 0; 1885 1886 if (!(gfp_mask & __GFP_WAIT) || 1887 zone->all_unreclaimable || 1888 atomic_read(&zone->reclaim_in_progress) > 0 || 1889 (p->flags & PF_MEMALLOC)) 1890 return 0; 1891 1892 node_id = zone->zone_pgdat->node_id; 1893 mask = node_to_cpumask(node_id); 1894 if (!cpus_empty(mask) && node_id != numa_node_id()) 1895 return 0; 1896 1897 sc.may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE); 1898 sc.may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP); 1899 sc.nr_scanned = 0; 1900 sc.nr_reclaimed = 0; 1901 sc.nr_mapped = read_page_state(nr_mapped); 1902 sc.gfp_mask = gfp_mask; 1903 1904 disable_swap_token(); 1905 1906 nr_pages = 1 << order; 1907 if (nr_pages > SWAP_CLUSTER_MAX) 1908 sc.swap_cluster_max = nr_pages; 1909 else 1910 sc.swap_cluster_max = SWAP_CLUSTER_MAX; 1911 1912 cond_resched(); 1913 /* 1914 * We need to be able to allocate from the reserves for RECLAIM_SWAP 1915 * and we also need to be able to write out pages for RECLAIM_WRITE 1916 * and RECLAIM_SWAP. 1917 */ 1918 p->flags |= PF_MEMALLOC | PF_SWAPWRITE; 1919 reclaim_state.reclaimed_slab = 0; 1920 p->reclaim_state = &reclaim_state; 1921 1922 /* 1923 * Free memory by calling shrink zone with increasing priorities 1924 * until we have enough memory freed. 1925 */ 1926 priority = ZONE_RECLAIM_PRIORITY; 1927 do { 1928 shrink_zone(priority, zone, &sc); 1929 priority--; 1930 } while (priority >= 0 && sc.nr_reclaimed < nr_pages); 1931 1932 if (sc.nr_reclaimed < nr_pages && (zone_reclaim_mode & RECLAIM_SLAB)) { 1933 /* 1934 * shrink_slab does not currently allow us to determine 1935 * how many pages were freed in the zone. So we just 1936 * shake the slab and then go offnode for a single allocation. 1937 * 1938 * shrink_slab will free memory on all zones and may take 1939 * a long time. 1940 */ 1941 shrink_slab(sc.nr_scanned, gfp_mask, order); 1942 } 1943 1944 p->reclaim_state = NULL; 1945 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE); 1946 1947 if (sc.nr_reclaimed == 0) 1948 zone->last_unsuccessful_zone_reclaim = jiffies; 1949 1950 return sc.nr_reclaimed >= nr_pages; 1951} 1952#endif 1953 1954