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() */ 43 typedef 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 54 struct scan_control { 55 /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */ 56 unsigned long nr_to_scan; 57 58 /* Incremented by the number of inactive pages that were scanned */ 59 unsigned long nr_scanned; 60 61 /* Incremented by the number of pages reclaimed */ 62 unsigned long nr_reclaimed; 63 64 unsigned long nr_mapped; /* From page_state */ 65 66 /* How many pages shrink_cache() should reclaim */ 67 int nr_to_reclaim; 68 69 /* Ask shrink_caches, or shrink_zone to scan at this priority */ 70 unsigned int priority; 71 72 /* This context's GFP mask */ 73 unsigned int gfp_mask; 74 75 int may_writepage; 76 77 /* Can pages be swapped as part of reclaim? */ 78 int may_swap; 79 80 /* This context's SWAP_CLUSTER_MAX. If freeing memory for 81 * suspend, we effectively ignore SWAP_CLUSTER_MAX. 82 * In this context, it doesn't matter that we scan the 83 * whole list at once. */ 84 int swap_cluster_max; 85 }; 86 87 /* 88 * The list of shrinker callbacks used by to apply pressure to 89 * ageable caches. 90 */ 91 struct shrinker { 92 shrinker_t shrinker; 93 struct list_head list; 94 int seeks; /* seeks to recreate an obj */ 95 long nr; /* objs pending delete */ 96 }; 97 98 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) 99 100 #ifdef ARCH_HAS_PREFETCH 101 #define prefetch_prev_lru_page(_page, _base, _field) \ 102 do { \ 103 if ((_page)->lru.prev != _base) { \ 104 struct page *prev; \ 105 \ 106 prev = lru_to_page(&(_page->lru)); \ 107 prefetch(&prev->_field); \ 108 } \ 109 } while (0) 110 #else 111 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) 112 #endif 113 114 #ifdef ARCH_HAS_PREFETCHW 115 #define prefetchw_prev_lru_page(_page, _base, _field) \ 116 do { \ 117 if ((_page)->lru.prev != _base) { \ 118 struct page *prev; \ 119 \ 120 prev = lru_to_page(&(_page->lru)); \ 121 prefetchw(&prev->_field); \ 122 } \ 123 } while (0) 124 #else 125 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) 126 #endif 127 128 /* 129 * From 0 .. 100. Higher means more swappy. 130 */ 131 int vm_swappiness = 60; 132 static long total_memory; 133 134 static LIST_HEAD(shrinker_list); 135 static DECLARE_RWSEM(shrinker_rwsem); 136 137 /* 138 * Add a shrinker callback to be called from the vm 139 */ 140 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker) 141 { 142 struct shrinker *shrinker; 143 144 shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL); 145 if (shrinker) { 146 shrinker->shrinker = theshrinker; 147 shrinker->seeks = seeks; 148 shrinker->nr = 0; 149 down_write(&shrinker_rwsem); 150 list_add_tail(&shrinker->list, &shrinker_list); 151 up_write(&shrinker_rwsem); 152 } 153 return shrinker; 154 } 155 EXPORT_SYMBOL(set_shrinker); 156 157 /* 158 * Remove one 159 */ 160 void remove_shrinker(struct shrinker *shrinker) 161 { 162 down_write(&shrinker_rwsem); 163 list_del(&shrinker->list); 164 up_write(&shrinker_rwsem); 165 kfree(shrinker); 166 } 167 EXPORT_SYMBOL(remove_shrinker); 168 169 #define SHRINK_BATCH 128 170 /* 171 * Call the shrink functions to age shrinkable caches 172 * 173 * Here we assume it costs one seek to replace a lru page and that it also 174 * takes a seek to recreate a cache object. With this in mind we age equal 175 * percentages of the lru and ageable caches. This should balance the seeks 176 * generated by these structures. 177 * 178 * If the vm encounted mapped pages on the LRU it increase the pressure on 179 * slab to avoid swapping. 180 * 181 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits. 182 * 183 * `lru_pages' represents the number of on-LRU pages in all the zones which 184 * are eligible for the caller's allocation attempt. It is used for balancing 185 * slab reclaim versus page reclaim. 186 * 187 * Returns the number of slab objects which we shrunk. 188 */ 189 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask, 190 unsigned long lru_pages) 191 { 192 struct shrinker *shrinker; 193 int ret = 0; 194 195 if (scanned == 0) 196 scanned = SWAP_CLUSTER_MAX; 197 198 if (!down_read_trylock(&shrinker_rwsem)) 199 return 1; /* Assume we'll be able to shrink next time */ 200 201 list_for_each_entry(shrinker, &shrinker_list, list) { 202 unsigned long long delta; 203 unsigned long total_scan; 204 205 delta = (4 * scanned) / shrinker->seeks; 206 delta *= (*shrinker->shrinker)(0, gfp_mask); 207 do_div(delta, lru_pages + 1); 208 shrinker->nr += delta; 209 if (shrinker->nr < 0) 210 shrinker->nr = LONG_MAX; /* It wrapped! */ 211 212 total_scan = shrinker->nr; 213 shrinker->nr = 0; 214 215 while (total_scan >= SHRINK_BATCH) { 216 long this_scan = SHRINK_BATCH; 217 int shrink_ret; 218 int nr_before; 219 220 nr_before = (*shrinker->shrinker)(0, gfp_mask); 221 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask); 222 if (shrink_ret == -1) 223 break; 224 if (shrink_ret < nr_before) 225 ret += nr_before - shrink_ret; 226 mod_page_state(slabs_scanned, this_scan); 227 total_scan -= this_scan; 228 229 cond_resched(); 230 } 231 232 shrinker->nr += total_scan; 233 } 234 up_read(&shrinker_rwsem); 235 return ret; 236 } 237 238 /* Called without lock on whether page is mapped, so answer is unstable */ 239 static inline int page_mapping_inuse(struct page *page) 240 { 241 struct address_space *mapping; 242 243 /* Page is in somebody's page tables. */ 244 if (page_mapped(page)) 245 return 1; 246 247 /* Be more reluctant to reclaim swapcache than pagecache */ 248 if (PageSwapCache(page)) 249 return 1; 250 251 mapping = page_mapping(page); 252 if (!mapping) 253 return 0; 254 255 /* File is mmap'd by somebody? */ 256 return mapping_mapped(mapping); 257 } 258 259 static inline int is_page_cache_freeable(struct page *page) 260 { 261 return page_count(page) - !!PagePrivate(page) == 2; 262 } 263 264 static int may_write_to_queue(struct backing_dev_info *bdi) 265 { 266 if (current_is_kswapd()) 267 return 1; 268 if (current_is_pdflush()) /* This is unlikely, but why not... */ 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 */ 289 static 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 */ 305 static 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 == 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 } 372 373 /* 374 * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed 375 */ 376 static int shrink_list(struct list_head *page_list, struct scan_control *sc) 377 { 378 LIST_HEAD(ret_pages); 379 struct pagevec freed_pvec; 380 int pgactivate = 0; 381 int reclaimed = 0; 382 383 cond_resched(); 384 385 pagevec_init(&freed_pvec, 1); 386 while (!list_empty(page_list)) { 387 struct address_space *mapping; 388 struct page *page; 389 int may_enter_fs; 390 int referenced; 391 392 cond_resched(); 393 394 page = lru_to_page(page_list); 395 list_del(&page->lru); 396 397 if (TestSetPageLocked(page)) 398 goto keep; 399 400 BUG_ON(PageActive(page)); 401 402 sc->nr_scanned++; 403 /* Double the slab pressure for mapped and swapcache pages */ 404 if (page_mapped(page) || PageSwapCache(page)) 405 sc->nr_scanned++; 406 407 if (PageWriteback(page)) 408 goto keep_locked; 409 410 referenced = page_referenced(page, 1, sc->priority <= 0); 411 /* In active use or really unfreeable? Activate it. */ 412 if (referenced && page_mapping_inuse(page)) 413 goto activate_locked; 414 415 #ifdef CONFIG_SWAP 416 /* 417 * Anonymous process memory has backing store? 418 * Try to allocate it some swap space here. 419 */ 420 if (PageAnon(page) && !PageSwapCache(page) && sc->may_swap) { 421 if (!add_to_swap(page)) 422 goto activate_locked; 423 } 424 #endif /* CONFIG_SWAP */ 425 426 mapping = page_mapping(page); 427 may_enter_fs = (sc->gfp_mask & __GFP_FS) || 428 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); 429 430 /* 431 * The page is mapped into the page tables of one or more 432 * processes. Try to unmap it here. 433 */ 434 if (page_mapped(page) && mapping) { 435 switch (try_to_unmap(page)) { 436 case SWAP_FAIL: 437 goto activate_locked; 438 case SWAP_AGAIN: 439 goto keep_locked; 440 case SWAP_SUCCESS: 441 ; /* try to free the page below */ 442 } 443 } 444 445 if (PageDirty(page)) { 446 if (referenced) 447 goto keep_locked; 448 if (!may_enter_fs) 449 goto keep_locked; 450 if (laptop_mode && !sc->may_writepage) 451 goto keep_locked; 452 453 /* Page is dirty, try to write it out here */ 454 switch(pageout(page, mapping)) { 455 case PAGE_KEEP: 456 goto keep_locked; 457 case PAGE_ACTIVATE: 458 goto activate_locked; 459 case PAGE_SUCCESS: 460 if (PageWriteback(page) || PageDirty(page)) 461 goto keep; 462 /* 463 * A synchronous write - probably a ramdisk. Go 464 * ahead and try to reclaim the page. 465 */ 466 if (TestSetPageLocked(page)) 467 goto keep; 468 if (PageDirty(page) || PageWriteback(page)) 469 goto keep_locked; 470 mapping = page_mapping(page); 471 case PAGE_CLEAN: 472 ; /* try to free the page below */ 473 } 474 } 475 476 /* 477 * If the page has buffers, try to free the buffer mappings 478 * associated with this page. If we succeed we try to free 479 * the page as well. 480 * 481 * We do this even if the page is PageDirty(). 482 * try_to_release_page() does not perform I/O, but it is 483 * possible for a page to have PageDirty set, but it is actually 484 * clean (all its buffers are clean). This happens if the 485 * buffers were written out directly, with submit_bh(). ext3 486 * will do this, as well as the blockdev mapping. 487 * try_to_release_page() will discover that cleanness and will 488 * drop the buffers and mark the page clean - it can be freed. 489 * 490 * Rarely, pages can have buffers and no ->mapping. These are 491 * the pages which were not successfully invalidated in 492 * truncate_complete_page(). We try to drop those buffers here 493 * and if that worked, and the page is no longer mapped into 494 * process address space (page_count == 1) it can be freed. 495 * Otherwise, leave the page on the LRU so it is swappable. 496 */ 497 if (PagePrivate(page)) { 498 if (!try_to_release_page(page, sc->gfp_mask)) 499 goto activate_locked; 500 if (!mapping && page_count(page) == 1) 501 goto free_it; 502 } 503 504 if (!mapping) 505 goto keep_locked; /* truncate got there first */ 506 507 write_lock_irq(&mapping->tree_lock); 508 509 /* 510 * The non-racy check for busy page. It is critical to check 511 * PageDirty _after_ making sure that the page is freeable and 512 * not in use by anybody. (pagecache + us == 2) 513 */ 514 if (page_count(page) != 2 || PageDirty(page)) { 515 write_unlock_irq(&mapping->tree_lock); 516 goto keep_locked; 517 } 518 519 #ifdef CONFIG_SWAP 520 if (PageSwapCache(page)) { 521 swp_entry_t swap = { .val = page->private }; 522 __delete_from_swap_cache(page); 523 write_unlock_irq(&mapping->tree_lock); 524 swap_free(swap); 525 __put_page(page); /* The pagecache ref */ 526 goto free_it; 527 } 528 #endif /* CONFIG_SWAP */ 529 530 __remove_from_page_cache(page); 531 write_unlock_irq(&mapping->tree_lock); 532 __put_page(page); 533 534 free_it: 535 unlock_page(page); 536 reclaimed++; 537 if (!pagevec_add(&freed_pvec, page)) 538 __pagevec_release_nonlru(&freed_pvec); 539 continue; 540 541 activate_locked: 542 SetPageActive(page); 543 pgactivate++; 544 keep_locked: 545 unlock_page(page); 546 keep: 547 list_add(&page->lru, &ret_pages); 548 BUG_ON(PageLRU(page)); 549 } 550 list_splice(&ret_pages, page_list); 551 if (pagevec_count(&freed_pvec)) 552 __pagevec_release_nonlru(&freed_pvec); 553 mod_page_state(pgactivate, pgactivate); 554 sc->nr_reclaimed += reclaimed; 555 return reclaimed; 556 } 557 558 /* 559 * zone->lru_lock is heavily contended. Some of the functions that 560 * shrink the lists perform better by taking out a batch of pages 561 * and working on them outside the LRU lock. 562 * 563 * For pagecache intensive workloads, this function is the hottest 564 * spot in the kernel (apart from copy_*_user functions). 565 * 566 * Appropriate locks must be held before calling this function. 567 * 568 * @nr_to_scan: The number of pages to look through on the list. 569 * @src: The LRU list to pull pages off. 570 * @dst: The temp list to put pages on to. 571 * @scanned: The number of pages that were scanned. 572 * 573 * returns how many pages were moved onto *@dst. 574 */ 575 static int isolate_lru_pages(int nr_to_scan, struct list_head *src, 576 struct list_head *dst, int *scanned) 577 { 578 int nr_taken = 0; 579 struct page *page; 580 int scan = 0; 581 582 while (scan++ < nr_to_scan && !list_empty(src)) { 583 page = lru_to_page(src); 584 prefetchw_prev_lru_page(page, src, flags); 585 586 if (!TestClearPageLRU(page)) 587 BUG(); 588 list_del(&page->lru); 589 if (get_page_testone(page)) { 590 /* 591 * It is being freed elsewhere 592 */ 593 __put_page(page); 594 SetPageLRU(page); 595 list_add(&page->lru, src); 596 continue; 597 } else { 598 list_add(&page->lru, dst); 599 nr_taken++; 600 } 601 } 602 603 *scanned = scan; 604 return nr_taken; 605 } 606 607 /* 608 * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed 609 */ 610 static void shrink_cache(struct zone *zone, struct scan_control *sc) 611 { 612 LIST_HEAD(page_list); 613 struct pagevec pvec; 614 int max_scan = sc->nr_to_scan; 615 616 pagevec_init(&pvec, 1); 617 618 lru_add_drain(); 619 spin_lock_irq(&zone->lru_lock); 620 while (max_scan > 0) { 621 struct page *page; 622 int nr_taken; 623 int nr_scan; 624 int nr_freed; 625 626 nr_taken = isolate_lru_pages(sc->swap_cluster_max, 627 &zone->inactive_list, 628 &page_list, &nr_scan); 629 zone->nr_inactive -= nr_taken; 630 zone->pages_scanned += nr_scan; 631 spin_unlock_irq(&zone->lru_lock); 632 633 if (nr_taken == 0) 634 goto done; 635 636 max_scan -= nr_scan; 637 if (current_is_kswapd()) 638 mod_page_state_zone(zone, pgscan_kswapd, nr_scan); 639 else 640 mod_page_state_zone(zone, pgscan_direct, nr_scan); 641 nr_freed = shrink_list(&page_list, sc); 642 if (current_is_kswapd()) 643 mod_page_state(kswapd_steal, nr_freed); 644 mod_page_state_zone(zone, pgsteal, nr_freed); 645 sc->nr_to_reclaim -= nr_freed; 646 647 spin_lock_irq(&zone->lru_lock); 648 /* 649 * Put back any unfreeable pages. 650 */ 651 while (!list_empty(&page_list)) { 652 page = lru_to_page(&page_list); 653 if (TestSetPageLRU(page)) 654 BUG(); 655 list_del(&page->lru); 656 if (PageActive(page)) 657 add_page_to_active_list(zone, page); 658 else 659 add_page_to_inactive_list(zone, page); 660 if (!pagevec_add(&pvec, page)) { 661 spin_unlock_irq(&zone->lru_lock); 662 __pagevec_release(&pvec); 663 spin_lock_irq(&zone->lru_lock); 664 } 665 } 666 } 667 spin_unlock_irq(&zone->lru_lock); 668 done: 669 pagevec_release(&pvec); 670 } 671 672 /* 673 * This moves pages from the active list to the inactive list. 674 * 675 * We move them the other way if the page is referenced by one or more 676 * processes, from rmap. 677 * 678 * If the pages are mostly unmapped, the processing is fast and it is 679 * appropriate to hold zone->lru_lock across the whole operation. But if 680 * the pages are mapped, the processing is slow (page_referenced()) so we 681 * should drop zone->lru_lock around each page. It's impossible to balance 682 * this, so instead we remove the pages from the LRU while processing them. 683 * It is safe to rely on PG_active against the non-LRU pages in here because 684 * nobody will play with that bit on a non-LRU page. 685 * 686 * The downside is that we have to touch page->_count against each page. 687 * But we had to alter page->flags anyway. 688 */ 689 static void 690 refill_inactive_zone(struct zone *zone, struct scan_control *sc) 691 { 692 int pgmoved; 693 int pgdeactivate = 0; 694 int pgscanned; 695 int nr_pages = sc->nr_to_scan; 696 LIST_HEAD(l_hold); /* The pages which were snipped off */ 697 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */ 698 LIST_HEAD(l_active); /* Pages to go onto the active_list */ 699 struct page *page; 700 struct pagevec pvec; 701 int reclaim_mapped = 0; 702 long mapped_ratio; 703 long distress; 704 long swap_tendency; 705 706 lru_add_drain(); 707 spin_lock_irq(&zone->lru_lock); 708 pgmoved = isolate_lru_pages(nr_pages, &zone->active_list, 709 &l_hold, &pgscanned); 710 zone->pages_scanned += pgscanned; 711 zone->nr_active -= pgmoved; 712 spin_unlock_irq(&zone->lru_lock); 713 714 /* 715 * `distress' is a measure of how much trouble we're having reclaiming 716 * pages. 0 -> no problems. 100 -> great trouble. 717 */ 718 distress = 100 >> zone->prev_priority; 719 720 /* 721 * The point of this algorithm is to decide when to start reclaiming 722 * mapped memory instead of just pagecache. Work out how much memory 723 * is mapped. 724 */ 725 mapped_ratio = (sc->nr_mapped * 100) / total_memory; 726 727 /* 728 * Now decide how much we really want to unmap some pages. The mapped 729 * ratio is downgraded - just because there's a lot of mapped memory 730 * doesn't necessarily mean that page reclaim isn't succeeding. 731 * 732 * The distress ratio is important - we don't want to start going oom. 733 * 734 * A 100% value of vm_swappiness overrides this algorithm altogether. 735 */ 736 swap_tendency = mapped_ratio / 2 + distress + vm_swappiness; 737 738 /* 739 * Now use this metric to decide whether to start moving mapped memory 740 * onto the inactive list. 741 */ 742 if (swap_tendency >= 100) 743 reclaim_mapped = 1; 744 745 while (!list_empty(&l_hold)) { 746 cond_resched(); 747 page = lru_to_page(&l_hold); 748 list_del(&page->lru); 749 if (page_mapped(page)) { 750 if (!reclaim_mapped || 751 (total_swap_pages == 0 && PageAnon(page)) || 752 page_referenced(page, 0, sc->priority <= 0)) { 753 list_add(&page->lru, &l_active); 754 continue; 755 } 756 } 757 list_add(&page->lru, &l_inactive); 758 } 759 760 pagevec_init(&pvec, 1); 761 pgmoved = 0; 762 spin_lock_irq(&zone->lru_lock); 763 while (!list_empty(&l_inactive)) { 764 page = lru_to_page(&l_inactive); 765 prefetchw_prev_lru_page(page, &l_inactive, flags); 766 if (TestSetPageLRU(page)) 767 BUG(); 768 if (!TestClearPageActive(page)) 769 BUG(); 770 list_move(&page->lru, &zone->inactive_list); 771 pgmoved++; 772 if (!pagevec_add(&pvec, page)) { 773 zone->nr_inactive += pgmoved; 774 spin_unlock_irq(&zone->lru_lock); 775 pgdeactivate += pgmoved; 776 pgmoved = 0; 777 if (buffer_heads_over_limit) 778 pagevec_strip(&pvec); 779 __pagevec_release(&pvec); 780 spin_lock_irq(&zone->lru_lock); 781 } 782 } 783 zone->nr_inactive += pgmoved; 784 pgdeactivate += pgmoved; 785 if (buffer_heads_over_limit) { 786 spin_unlock_irq(&zone->lru_lock); 787 pagevec_strip(&pvec); 788 spin_lock_irq(&zone->lru_lock); 789 } 790 791 pgmoved = 0; 792 while (!list_empty(&l_active)) { 793 page = lru_to_page(&l_active); 794 prefetchw_prev_lru_page(page, &l_active, flags); 795 if (TestSetPageLRU(page)) 796 BUG(); 797 BUG_ON(!PageActive(page)); 798 list_move(&page->lru, &zone->active_list); 799 pgmoved++; 800 if (!pagevec_add(&pvec, page)) { 801 zone->nr_active += pgmoved; 802 pgmoved = 0; 803 spin_unlock_irq(&zone->lru_lock); 804 __pagevec_release(&pvec); 805 spin_lock_irq(&zone->lru_lock); 806 } 807 } 808 zone->nr_active += pgmoved; 809 spin_unlock_irq(&zone->lru_lock); 810 pagevec_release(&pvec); 811 812 mod_page_state_zone(zone, pgrefill, pgscanned); 813 mod_page_state(pgdeactivate, pgdeactivate); 814 } 815 816 /* 817 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim. 818 */ 819 static void 820 shrink_zone(struct zone *zone, struct scan_control *sc) 821 { 822 unsigned long nr_active; 823 unsigned long nr_inactive; 824 825 /* 826 * Add one to `nr_to_scan' just to make sure that the kernel will 827 * slowly sift through the active list. 828 */ 829 zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1; 830 nr_active = zone->nr_scan_active; 831 if (nr_active >= sc->swap_cluster_max) 832 zone->nr_scan_active = 0; 833 else 834 nr_active = 0; 835 836 zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1; 837 nr_inactive = zone->nr_scan_inactive; 838 if (nr_inactive >= sc->swap_cluster_max) 839 zone->nr_scan_inactive = 0; 840 else 841 nr_inactive = 0; 842 843 sc->nr_to_reclaim = sc->swap_cluster_max; 844 845 while (nr_active || nr_inactive) { 846 if (nr_active) { 847 sc->nr_to_scan = min(nr_active, 848 (unsigned long)sc->swap_cluster_max); 849 nr_active -= sc->nr_to_scan; 850 refill_inactive_zone(zone, sc); 851 } 852 853 if (nr_inactive) { 854 sc->nr_to_scan = min(nr_inactive, 855 (unsigned long)sc->swap_cluster_max); 856 nr_inactive -= sc->nr_to_scan; 857 shrink_cache(zone, sc); 858 if (sc->nr_to_reclaim <= 0) 859 break; 860 } 861 } 862 863 throttle_vm_writeout(); 864 } 865 866 /* 867 * This is the direct reclaim path, for page-allocating processes. We only 868 * try to reclaim pages from zones which will satisfy the caller's allocation 869 * request. 870 * 871 * We reclaim from a zone even if that zone is over pages_high. Because: 872 * a) The caller may be trying to free *extra* pages to satisfy a higher-order 873 * allocation or 874 * b) The zones may be over pages_high but they must go *over* pages_high to 875 * satisfy the `incremental min' zone defense algorithm. 876 * 877 * Returns the number of reclaimed pages. 878 * 879 * If a zone is deemed to be full of pinned pages then just give it a light 880 * scan then give up on it. 881 */ 882 static void 883 shrink_caches(struct zone **zones, struct scan_control *sc) 884 { 885 int i; 886 887 for (i = 0; zones[i] != NULL; i++) { 888 struct zone *zone = zones[i]; 889 890 if (zone->present_pages == 0) 891 continue; 892 893 if (!cpuset_zone_allowed(zone)) 894 continue; 895 896 zone->temp_priority = sc->priority; 897 if (zone->prev_priority > sc->priority) 898 zone->prev_priority = sc->priority; 899 900 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY) 901 continue; /* Let kswapd poll it */ 902 903 atomic_inc(&zone->reclaim_in_progress); 904 shrink_zone(zone, sc); 905 atomic_dec(&zone->reclaim_in_progress); 906 } 907 } 908 909 /* 910 * This is the main entry point to direct page reclaim. 911 * 912 * If a full scan of the inactive list fails to free enough memory then we 913 * are "out of memory" and something needs to be killed. 914 * 915 * If the caller is !__GFP_FS then the probability of a failure is reasonably 916 * high - the zone may be full of dirty or under-writeback pages, which this 917 * caller can't do much about. We kick pdflush and take explicit naps in the 918 * hope that some of these pages can be written. But if the allocating task 919 * holds filesystem locks which prevent writeout this might not work, and the 920 * allocation attempt will fail. 921 */ 922 int try_to_free_pages(struct zone **zones, unsigned int gfp_mask) 923 { 924 int priority; 925 int ret = 0; 926 int total_scanned = 0, total_reclaimed = 0; 927 struct reclaim_state *reclaim_state = current->reclaim_state; 928 struct scan_control sc; 929 unsigned long lru_pages = 0; 930 int i; 931 932 sc.gfp_mask = gfp_mask; 933 sc.may_writepage = 0; 934 sc.may_swap = 1; 935 936 inc_page_state(allocstall); 937 938 for (i = 0; zones[i] != NULL; i++) { 939 struct zone *zone = zones[i]; 940 941 if (!cpuset_zone_allowed(zone)) 942 continue; 943 944 zone->temp_priority = DEF_PRIORITY; 945 lru_pages += zone->nr_active + zone->nr_inactive; 946 } 947 948 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 949 sc.nr_mapped = read_page_state(nr_mapped); 950 sc.nr_scanned = 0; 951 sc.nr_reclaimed = 0; 952 sc.priority = priority; 953 sc.swap_cluster_max = SWAP_CLUSTER_MAX; 954 shrink_caches(zones, &sc); 955 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages); 956 if (reclaim_state) { 957 sc.nr_reclaimed += reclaim_state->reclaimed_slab; 958 reclaim_state->reclaimed_slab = 0; 959 } 960 total_scanned += sc.nr_scanned; 961 total_reclaimed += sc.nr_reclaimed; 962 if (total_reclaimed >= sc.swap_cluster_max) { 963 ret = 1; 964 goto out; 965 } 966 967 /* 968 * Try to write back as many pages as we just scanned. This 969 * tends to cause slow streaming writers to write data to the 970 * disk smoothly, at the dirtying rate, which is nice. But 971 * that's undesirable in laptop mode, where we *want* lumpy 972 * writeout. So in laptop mode, write out the whole world. 973 */ 974 if (total_scanned > sc.swap_cluster_max + sc.swap_cluster_max/2) { 975 wakeup_bdflush(laptop_mode ? 0 : total_scanned); 976 sc.may_writepage = 1; 977 } 978 979 /* Take a nap, wait for some writeback to complete */ 980 if (sc.nr_scanned && priority < DEF_PRIORITY - 2) 981 blk_congestion_wait(WRITE, HZ/10); 982 } 983 out: 984 for (i = 0; zones[i] != 0; i++) { 985 struct zone *zone = zones[i]; 986 987 if (!cpuset_zone_allowed(zone)) 988 continue; 989 990 zone->prev_priority = zone->temp_priority; 991 } 992 return ret; 993 } 994 995 /* 996 * For kswapd, balance_pgdat() will work across all this node's zones until 997 * they are all at pages_high. 998 * 999 * If `nr_pages' is non-zero then it is the number of pages which are to be 1000 * reclaimed, regardless of the zone occupancies. This is a software suspend 1001 * special. 1002 * 1003 * Returns the number of pages which were actually freed. 1004 * 1005 * There is special handling here for zones which are full of pinned pages. 1006 * This can happen if the pages are all mlocked, or if they are all used by 1007 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb. 1008 * What we do is to detect the case where all pages in the zone have been 1009 * scanned twice and there has been zero successful reclaim. Mark the zone as 1010 * dead and from now on, only perform a short scan. Basically we're polling 1011 * the zone for when the problem goes away. 1012 * 1013 * kswapd scans the zones in the highmem->normal->dma direction. It skips 1014 * zones which have free_pages > pages_high, but once a zone is found to have 1015 * free_pages <= pages_high, we scan that zone and the lower zones regardless 1016 * of the number of free pages in the lower zones. This interoperates with 1017 * the page allocator fallback scheme to ensure that aging of pages is balanced 1018 * across the zones. 1019 */ 1020 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order) 1021 { 1022 int to_free = nr_pages; 1023 int all_zones_ok; 1024 int priority; 1025 int i; 1026 int total_scanned, total_reclaimed; 1027 struct reclaim_state *reclaim_state = current->reclaim_state; 1028 struct scan_control sc; 1029 1030 loop_again: 1031 total_scanned = 0; 1032 total_reclaimed = 0; 1033 sc.gfp_mask = GFP_KERNEL; 1034 sc.may_writepage = 0; 1035 sc.may_swap = 1; 1036 sc.nr_mapped = read_page_state(nr_mapped); 1037 1038 inc_page_state(pageoutrun); 1039 1040 for (i = 0; i < pgdat->nr_zones; i++) { 1041 struct zone *zone = pgdat->node_zones + i; 1042 1043 zone->temp_priority = DEF_PRIORITY; 1044 } 1045 1046 for (priority = DEF_PRIORITY; priority >= 0; priority--) { 1047 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */ 1048 unsigned long lru_pages = 0; 1049 1050 all_zones_ok = 1; 1051 1052 if (nr_pages == 0) { 1053 /* 1054 * Scan in the highmem->dma direction for the highest 1055 * zone which needs scanning 1056 */ 1057 for (i = pgdat->nr_zones - 1; i >= 0; i--) { 1058 struct zone *zone = pgdat->node_zones + i; 1059 1060 if (zone->present_pages == 0) 1061 continue; 1062 1063 if (zone->all_unreclaimable && 1064 priority != DEF_PRIORITY) 1065 continue; 1066 1067 if (!zone_watermark_ok(zone, order, 1068 zone->pages_high, 0, 0, 0)) { 1069 end_zone = i; 1070 goto scan; 1071 } 1072 } 1073 goto out; 1074 } else { 1075 end_zone = pgdat->nr_zones - 1; 1076 } 1077 scan: 1078 for (i = 0; i <= end_zone; i++) { 1079 struct zone *zone = pgdat->node_zones + i; 1080 1081 lru_pages += zone->nr_active + zone->nr_inactive; 1082 } 1083 1084 /* 1085 * Now scan the zone in the dma->highmem direction, stopping 1086 * at the last zone which needs scanning. 1087 * 1088 * We do this because the page allocator works in the opposite 1089 * direction. This prevents the page allocator from allocating 1090 * pages behind kswapd's direction of progress, which would 1091 * cause too much scanning of the lower zones. 1092 */ 1093 for (i = 0; i <= end_zone; i++) { 1094 struct zone *zone = pgdat->node_zones + i; 1095 int nr_slab; 1096 1097 if (zone->present_pages == 0) 1098 continue; 1099 1100 if (zone->all_unreclaimable && priority != DEF_PRIORITY) 1101 continue; 1102 1103 if (nr_pages == 0) { /* Not software suspend */ 1104 if (!zone_watermark_ok(zone, order, 1105 zone->pages_high, end_zone, 0, 0)) 1106 all_zones_ok = 0; 1107 } 1108 zone->temp_priority = priority; 1109 if (zone->prev_priority > priority) 1110 zone->prev_priority = priority; 1111 sc.nr_scanned = 0; 1112 sc.nr_reclaimed = 0; 1113 sc.priority = priority; 1114 sc.swap_cluster_max = nr_pages? nr_pages : SWAP_CLUSTER_MAX; 1115 atomic_inc(&zone->reclaim_in_progress); 1116 shrink_zone(zone, &sc); 1117 atomic_dec(&zone->reclaim_in_progress); 1118 reclaim_state->reclaimed_slab = 0; 1119 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL, 1120 lru_pages); 1121 sc.nr_reclaimed += reclaim_state->reclaimed_slab; 1122 total_reclaimed += sc.nr_reclaimed; 1123 total_scanned += sc.nr_scanned; 1124 if (zone->all_unreclaimable) 1125 continue; 1126 if (nr_slab == 0 && zone->pages_scanned >= 1127 (zone->nr_active + zone->nr_inactive) * 4) 1128 zone->all_unreclaimable = 1; 1129 /* 1130 * If we've done a decent amount of scanning and 1131 * the reclaim ratio is low, start doing writepage 1132 * even in laptop mode 1133 */ 1134 if (total_scanned > SWAP_CLUSTER_MAX * 2 && 1135 total_scanned > total_reclaimed+total_reclaimed/2) 1136 sc.may_writepage = 1; 1137 } 1138 if (nr_pages && to_free > total_reclaimed) 1139 continue; /* swsusp: need to do more work */ 1140 if (all_zones_ok) 1141 break; /* kswapd: all done */ 1142 /* 1143 * OK, kswapd is getting into trouble. Take a nap, then take 1144 * another pass across the zones. 1145 */ 1146 if (total_scanned && priority < DEF_PRIORITY - 2) 1147 blk_congestion_wait(WRITE, HZ/10); 1148 1149 /* 1150 * We do this so kswapd doesn't build up large priorities for 1151 * example when it is freeing in parallel with allocators. It 1152 * matches the direct reclaim path behaviour in terms of impact 1153 * on zone->*_priority. 1154 */ 1155 if ((total_reclaimed >= SWAP_CLUSTER_MAX) && (!nr_pages)) 1156 break; 1157 } 1158 out: 1159 for (i = 0; i < pgdat->nr_zones; i++) { 1160 struct zone *zone = pgdat->node_zones + i; 1161 1162 zone->prev_priority = zone->temp_priority; 1163 } 1164 if (!all_zones_ok) { 1165 cond_resched(); 1166 goto loop_again; 1167 } 1168 1169 return total_reclaimed; 1170 } 1171 1172 /* 1173 * The background pageout daemon, started as a kernel thread 1174 * from the init process. 1175 * 1176 * This basically trickles out pages so that we have _some_ 1177 * free memory available even if there is no other activity 1178 * that frees anything up. This is needed for things like routing 1179 * etc, where we otherwise might have all activity going on in 1180 * asynchronous contexts that cannot page things out. 1181 * 1182 * If there are applications that are active memory-allocators 1183 * (most normal use), this basically shouldn't matter. 1184 */ 1185 static int kswapd(void *p) 1186 { 1187 unsigned long order; 1188 pg_data_t *pgdat = (pg_data_t*)p; 1189 struct task_struct *tsk = current; 1190 DEFINE_WAIT(wait); 1191 struct reclaim_state reclaim_state = { 1192 .reclaimed_slab = 0, 1193 }; 1194 cpumask_t cpumask; 1195 1196 daemonize("kswapd%d", pgdat->node_id); 1197 cpumask = node_to_cpumask(pgdat->node_id); 1198 if (!cpus_empty(cpumask)) 1199 set_cpus_allowed(tsk, cpumask); 1200 current->reclaim_state = &reclaim_state; 1201 1202 /* 1203 * Tell the memory management that we're a "memory allocator", 1204 * and that if we need more memory we should get access to it 1205 * regardless (see "__alloc_pages()"). "kswapd" should 1206 * never get caught in the normal page freeing logic. 1207 * 1208 * (Kswapd normally doesn't need memory anyway, but sometimes 1209 * you need a small amount of memory in order to be able to 1210 * page out something else, and this flag essentially protects 1211 * us from recursively trying to free more memory as we're 1212 * trying to free the first piece of memory in the first place). 1213 */ 1214 tsk->flags |= PF_MEMALLOC|PF_KSWAPD; 1215 1216 order = 0; 1217 for ( ; ; ) { 1218 unsigned long new_order; 1219 if (current->flags & PF_FREEZE) 1220 refrigerator(PF_FREEZE); 1221 1222 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); 1223 new_order = pgdat->kswapd_max_order; 1224 pgdat->kswapd_max_order = 0; 1225 if (order < new_order) { 1226 /* 1227 * Don't sleep if someone wants a larger 'order' 1228 * allocation 1229 */ 1230 order = new_order; 1231 } else { 1232 schedule(); 1233 order = pgdat->kswapd_max_order; 1234 } 1235 finish_wait(&pgdat->kswapd_wait, &wait); 1236 1237 balance_pgdat(pgdat, 0, order); 1238 } 1239 return 0; 1240 } 1241 1242 /* 1243 * A zone is low on free memory, so wake its kswapd task to service it. 1244 */ 1245 void wakeup_kswapd(struct zone *zone, int order) 1246 { 1247 pg_data_t *pgdat; 1248 1249 if (zone->present_pages == 0) 1250 return; 1251 1252 pgdat = zone->zone_pgdat; 1253 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0, 0)) 1254 return; 1255 if (pgdat->kswapd_max_order < order) 1256 pgdat->kswapd_max_order = order; 1257 if (!cpuset_zone_allowed(zone)) 1258 return; 1259 if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait)) 1260 return; 1261 wake_up_interruptible(&zone->zone_pgdat->kswapd_wait); 1262 } 1263 1264 #ifdef CONFIG_PM 1265 /* 1266 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed 1267 * pages. 1268 */ 1269 int shrink_all_memory(int nr_pages) 1270 { 1271 pg_data_t *pgdat; 1272 int nr_to_free = nr_pages; 1273 int ret = 0; 1274 struct reclaim_state reclaim_state = { 1275 .reclaimed_slab = 0, 1276 }; 1277 1278 current->reclaim_state = &reclaim_state; 1279 for_each_pgdat(pgdat) { 1280 int freed; 1281 freed = balance_pgdat(pgdat, nr_to_free, 0); 1282 ret += freed; 1283 nr_to_free -= freed; 1284 if (nr_to_free <= 0) 1285 break; 1286 } 1287 current->reclaim_state = NULL; 1288 return ret; 1289 } 1290 #endif 1291 1292 #ifdef CONFIG_HOTPLUG_CPU 1293 /* It's optimal to keep kswapds on the same CPUs as their memory, but 1294 not required for correctness. So if the last cpu in a node goes 1295 away, we get changed to run anywhere: as the first one comes back, 1296 restore their cpu bindings. */ 1297 static int __devinit cpu_callback(struct notifier_block *nfb, 1298 unsigned long action, 1299 void *hcpu) 1300 { 1301 pg_data_t *pgdat; 1302 cpumask_t mask; 1303 1304 if (action == CPU_ONLINE) { 1305 for_each_pgdat(pgdat) { 1306 mask = node_to_cpumask(pgdat->node_id); 1307 if (any_online_cpu(mask) != NR_CPUS) 1308 /* One of our CPUs online: restore mask */ 1309 set_cpus_allowed(pgdat->kswapd, mask); 1310 } 1311 } 1312 return NOTIFY_OK; 1313 } 1314 #endif /* CONFIG_HOTPLUG_CPU */ 1315 1316 static int __init kswapd_init(void) 1317 { 1318 pg_data_t *pgdat; 1319 swap_setup(); 1320 for_each_pgdat(pgdat) 1321 pgdat->kswapd 1322 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL)); 1323 total_memory = nr_free_pagecache_pages(); 1324 hotcpu_notifier(cpu_callback, 0); 1325 return 0; 1326 } 1327 1328 module_init(kswapd_init) 1329 1330 1331 /* 1332 * Try to free up some pages from this zone through reclaim. 1333 */ 1334 int zone_reclaim(struct zone *zone, unsigned int gfp_mask, unsigned int order) 1335 { 1336 struct scan_control sc; 1337 int nr_pages = 1 << order; 1338 int total_reclaimed = 0; 1339 1340 /* The reclaim may sleep, so don't do it if sleep isn't allowed */ 1341 if (!(gfp_mask & __GFP_WAIT)) 1342 return 0; 1343 if (zone->all_unreclaimable) 1344 return 0; 1345 1346 sc.gfp_mask = gfp_mask; 1347 sc.may_writepage = 0; 1348 sc.may_swap = 0; 1349 sc.nr_mapped = read_page_state(nr_mapped); 1350 sc.nr_scanned = 0; 1351 sc.nr_reclaimed = 0; 1352 /* scan at the highest priority */ 1353 sc.priority = 0; 1354 1355 if (nr_pages > SWAP_CLUSTER_MAX) 1356 sc.swap_cluster_max = nr_pages; 1357 else 1358 sc.swap_cluster_max = SWAP_CLUSTER_MAX; 1359 1360 /* Don't reclaim the zone if there are other reclaimers active */ 1361 if (!atomic_inc_and_test(&zone->reclaim_in_progress)) 1362 goto out; 1363 1364 shrink_zone(zone, &sc); 1365 total_reclaimed = sc.nr_reclaimed; 1366 1367 out: 1368 atomic_dec(&zone->reclaim_in_progress); 1369 return total_reclaimed; 1370 } 1371 1372 asmlinkage long sys_set_zone_reclaim(unsigned int node, unsigned int zone, 1373 unsigned int state) 1374 { 1375 struct zone *z; 1376 int i; 1377 1378 if (node >= MAX_NUMNODES || !node_online(node)) 1379 return -EINVAL; 1380 1381 /* This will break if we ever add more zones */ 1382 if (!(zone & (1<<ZONE_DMA|1<<ZONE_NORMAL|1<<ZONE_HIGHMEM))) 1383 return -EINVAL; 1384 1385 for (i = 0; i < MAX_NR_ZONES; i++) { 1386 if (!(zone & 1<<i)) 1387 continue; 1388 1389 z = &NODE_DATA(node)->node_zones[i]; 1390 1391 if (state) 1392 z->reclaim_pages = 1; 1393 else 1394 z->reclaim_pages = 0; 1395 } 1396 1397 return 0; 1398 } 1399