rmap.c revision 9617d95e6e9ffd883cf90a89724fe60d7ab22f9a
1/* 2 * mm/rmap.c - physical to virtual reverse mappings 3 * 4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br> 5 * Released under the General Public License (GPL). 6 * 7 * Simple, low overhead reverse mapping scheme. 8 * Please try to keep this thing as modular as possible. 9 * 10 * Provides methods for unmapping each kind of mapped page: 11 * the anon methods track anonymous pages, and 12 * the file methods track pages belonging to an inode. 13 * 14 * Original design by Rik van Riel <riel@conectiva.com.br> 2001 15 * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004 16 * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004 17 * Contributions by Hugh Dickins <hugh@veritas.com> 2003, 2004 18 */ 19 20/* 21 * Lock ordering in mm: 22 * 23 * inode->i_sem (while writing or truncating, not reading or faulting) 24 * inode->i_alloc_sem 25 * 26 * When a page fault occurs in writing from user to file, down_read 27 * of mmap_sem nests within i_sem; in sys_msync, i_sem nests within 28 * down_read of mmap_sem; i_sem and down_write of mmap_sem are never 29 * taken together; in truncation, i_sem is taken outermost. 30 * 31 * mm->mmap_sem 32 * page->flags PG_locked (lock_page) 33 * mapping->i_mmap_lock 34 * anon_vma->lock 35 * mm->page_table_lock or pte_lock 36 * zone->lru_lock (in mark_page_accessed) 37 * swap_lock (in swap_duplicate, swap_info_get) 38 * mmlist_lock (in mmput, drain_mmlist and others) 39 * mapping->private_lock (in __set_page_dirty_buffers) 40 * inode_lock (in set_page_dirty's __mark_inode_dirty) 41 * sb_lock (within inode_lock in fs/fs-writeback.c) 42 * mapping->tree_lock (widely used, in set_page_dirty, 43 * in arch-dependent flush_dcache_mmap_lock, 44 * within inode_lock in __sync_single_inode) 45 */ 46 47#include <linux/mm.h> 48#include <linux/pagemap.h> 49#include <linux/swap.h> 50#include <linux/swapops.h> 51#include <linux/slab.h> 52#include <linux/init.h> 53#include <linux/rmap.h> 54#include <linux/rcupdate.h> 55 56#include <asm/tlbflush.h> 57 58//#define RMAP_DEBUG /* can be enabled only for debugging */ 59 60kmem_cache_t *anon_vma_cachep; 61 62static inline void validate_anon_vma(struct vm_area_struct *find_vma) 63{ 64#ifdef RMAP_DEBUG 65 struct anon_vma *anon_vma = find_vma->anon_vma; 66 struct vm_area_struct *vma; 67 unsigned int mapcount = 0; 68 int found = 0; 69 70 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { 71 mapcount++; 72 BUG_ON(mapcount > 100000); 73 if (vma == find_vma) 74 found = 1; 75 } 76 BUG_ON(!found); 77#endif 78} 79 80/* This must be called under the mmap_sem. */ 81int anon_vma_prepare(struct vm_area_struct *vma) 82{ 83 struct anon_vma *anon_vma = vma->anon_vma; 84 85 might_sleep(); 86 if (unlikely(!anon_vma)) { 87 struct mm_struct *mm = vma->vm_mm; 88 struct anon_vma *allocated, *locked; 89 90 anon_vma = find_mergeable_anon_vma(vma); 91 if (anon_vma) { 92 allocated = NULL; 93 locked = anon_vma; 94 spin_lock(&locked->lock); 95 } else { 96 anon_vma = anon_vma_alloc(); 97 if (unlikely(!anon_vma)) 98 return -ENOMEM; 99 allocated = anon_vma; 100 locked = NULL; 101 } 102 103 /* page_table_lock to protect against threads */ 104 spin_lock(&mm->page_table_lock); 105 if (likely(!vma->anon_vma)) { 106 vma->anon_vma = anon_vma; 107 list_add(&vma->anon_vma_node, &anon_vma->head); 108 allocated = NULL; 109 } 110 spin_unlock(&mm->page_table_lock); 111 112 if (locked) 113 spin_unlock(&locked->lock); 114 if (unlikely(allocated)) 115 anon_vma_free(allocated); 116 } 117 return 0; 118} 119 120void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next) 121{ 122 BUG_ON(vma->anon_vma != next->anon_vma); 123 list_del(&next->anon_vma_node); 124} 125 126void __anon_vma_link(struct vm_area_struct *vma) 127{ 128 struct anon_vma *anon_vma = vma->anon_vma; 129 130 if (anon_vma) { 131 list_add(&vma->anon_vma_node, &anon_vma->head); 132 validate_anon_vma(vma); 133 } 134} 135 136void anon_vma_link(struct vm_area_struct *vma) 137{ 138 struct anon_vma *anon_vma = vma->anon_vma; 139 140 if (anon_vma) { 141 spin_lock(&anon_vma->lock); 142 list_add(&vma->anon_vma_node, &anon_vma->head); 143 validate_anon_vma(vma); 144 spin_unlock(&anon_vma->lock); 145 } 146} 147 148void anon_vma_unlink(struct vm_area_struct *vma) 149{ 150 struct anon_vma *anon_vma = vma->anon_vma; 151 int empty; 152 153 if (!anon_vma) 154 return; 155 156 spin_lock(&anon_vma->lock); 157 validate_anon_vma(vma); 158 list_del(&vma->anon_vma_node); 159 160 /* We must garbage collect the anon_vma if it's empty */ 161 empty = list_empty(&anon_vma->head); 162 spin_unlock(&anon_vma->lock); 163 164 if (empty) 165 anon_vma_free(anon_vma); 166} 167 168static void anon_vma_ctor(void *data, kmem_cache_t *cachep, unsigned long flags) 169{ 170 if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) == 171 SLAB_CTOR_CONSTRUCTOR) { 172 struct anon_vma *anon_vma = data; 173 174 spin_lock_init(&anon_vma->lock); 175 INIT_LIST_HEAD(&anon_vma->head); 176 } 177} 178 179void __init anon_vma_init(void) 180{ 181 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), 182 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL); 183} 184 185/* 186 * Getting a lock on a stable anon_vma from a page off the LRU is 187 * tricky: page_lock_anon_vma rely on RCU to guard against the races. 188 */ 189static struct anon_vma *page_lock_anon_vma(struct page *page) 190{ 191 struct anon_vma *anon_vma = NULL; 192 unsigned long anon_mapping; 193 194 rcu_read_lock(); 195 anon_mapping = (unsigned long) page->mapping; 196 if (!(anon_mapping & PAGE_MAPPING_ANON)) 197 goto out; 198 if (!page_mapped(page)) 199 goto out; 200 201 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); 202 spin_lock(&anon_vma->lock); 203out: 204 rcu_read_unlock(); 205 return anon_vma; 206} 207 208/* 209 * At what user virtual address is page expected in vma? 210 */ 211static inline unsigned long 212vma_address(struct page *page, struct vm_area_struct *vma) 213{ 214 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 215 unsigned long address; 216 217 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 218 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { 219 /* page should be within any vma from prio_tree_next */ 220 BUG_ON(!PageAnon(page)); 221 return -EFAULT; 222 } 223 return address; 224} 225 226/* 227 * At what user virtual address is page expected in vma? checking that the 228 * page matches the vma: currently only used on anon pages, by unuse_vma; 229 */ 230unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) 231{ 232 if (PageAnon(page)) { 233 if ((void *)vma->anon_vma != 234 (void *)page->mapping - PAGE_MAPPING_ANON) 235 return -EFAULT; 236 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { 237 if (!vma->vm_file || 238 vma->vm_file->f_mapping != page->mapping) 239 return -EFAULT; 240 } else 241 return -EFAULT; 242 return vma_address(page, vma); 243} 244 245/* 246 * Check that @page is mapped at @address into @mm. 247 * 248 * On success returns with pte mapped and locked. 249 */ 250pte_t *page_check_address(struct page *page, struct mm_struct *mm, 251 unsigned long address, spinlock_t **ptlp) 252{ 253 pgd_t *pgd; 254 pud_t *pud; 255 pmd_t *pmd; 256 pte_t *pte; 257 spinlock_t *ptl; 258 259 pgd = pgd_offset(mm, address); 260 if (!pgd_present(*pgd)) 261 return NULL; 262 263 pud = pud_offset(pgd, address); 264 if (!pud_present(*pud)) 265 return NULL; 266 267 pmd = pmd_offset(pud, address); 268 if (!pmd_present(*pmd)) 269 return NULL; 270 271 pte = pte_offset_map(pmd, address); 272 /* Make a quick check before getting the lock */ 273 if (!pte_present(*pte)) { 274 pte_unmap(pte); 275 return NULL; 276 } 277 278 ptl = pte_lockptr(mm, pmd); 279 spin_lock(ptl); 280 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { 281 *ptlp = ptl; 282 return pte; 283 } 284 pte_unmap_unlock(pte, ptl); 285 return NULL; 286} 287 288/* 289 * Subfunctions of page_referenced: page_referenced_one called 290 * repeatedly from either page_referenced_anon or page_referenced_file. 291 */ 292static int page_referenced_one(struct page *page, 293 struct vm_area_struct *vma, unsigned int *mapcount) 294{ 295 struct mm_struct *mm = vma->vm_mm; 296 unsigned long address; 297 pte_t *pte; 298 spinlock_t *ptl; 299 int referenced = 0; 300 301 address = vma_address(page, vma); 302 if (address == -EFAULT) 303 goto out; 304 305 pte = page_check_address(page, mm, address, &ptl); 306 if (!pte) 307 goto out; 308 309 if (ptep_clear_flush_young(vma, address, pte)) 310 referenced++; 311 312 /* Pretend the page is referenced if the task has the 313 swap token and is in the middle of a page fault. */ 314 if (mm != current->mm && has_swap_token(mm) && 315 rwsem_is_locked(&mm->mmap_sem)) 316 referenced++; 317 318 (*mapcount)--; 319 pte_unmap_unlock(pte, ptl); 320out: 321 return referenced; 322} 323 324static int page_referenced_anon(struct page *page) 325{ 326 unsigned int mapcount; 327 struct anon_vma *anon_vma; 328 struct vm_area_struct *vma; 329 int referenced = 0; 330 331 anon_vma = page_lock_anon_vma(page); 332 if (!anon_vma) 333 return referenced; 334 335 mapcount = page_mapcount(page); 336 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { 337 referenced += page_referenced_one(page, vma, &mapcount); 338 if (!mapcount) 339 break; 340 } 341 spin_unlock(&anon_vma->lock); 342 return referenced; 343} 344 345/** 346 * page_referenced_file - referenced check for object-based rmap 347 * @page: the page we're checking references on. 348 * 349 * For an object-based mapped page, find all the places it is mapped and 350 * check/clear the referenced flag. This is done by following the page->mapping 351 * pointer, then walking the chain of vmas it holds. It returns the number 352 * of references it found. 353 * 354 * This function is only called from page_referenced for object-based pages. 355 */ 356static int page_referenced_file(struct page *page) 357{ 358 unsigned int mapcount; 359 struct address_space *mapping = page->mapping; 360 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 361 struct vm_area_struct *vma; 362 struct prio_tree_iter iter; 363 int referenced = 0; 364 365 /* 366 * The caller's checks on page->mapping and !PageAnon have made 367 * sure that this is a file page: the check for page->mapping 368 * excludes the case just before it gets set on an anon page. 369 */ 370 BUG_ON(PageAnon(page)); 371 372 /* 373 * The page lock not only makes sure that page->mapping cannot 374 * suddenly be NULLified by truncation, it makes sure that the 375 * structure at mapping cannot be freed and reused yet, 376 * so we can safely take mapping->i_mmap_lock. 377 */ 378 BUG_ON(!PageLocked(page)); 379 380 spin_lock(&mapping->i_mmap_lock); 381 382 /* 383 * i_mmap_lock does not stabilize mapcount at all, but mapcount 384 * is more likely to be accurate if we note it after spinning. 385 */ 386 mapcount = page_mapcount(page); 387 388 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 389 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE)) 390 == (VM_LOCKED|VM_MAYSHARE)) { 391 referenced++; 392 break; 393 } 394 referenced += page_referenced_one(page, vma, &mapcount); 395 if (!mapcount) 396 break; 397 } 398 399 spin_unlock(&mapping->i_mmap_lock); 400 return referenced; 401} 402 403/** 404 * page_referenced - test if the page was referenced 405 * @page: the page to test 406 * @is_locked: caller holds lock on the page 407 * 408 * Quick test_and_clear_referenced for all mappings to a page, 409 * returns the number of ptes which referenced the page. 410 */ 411int page_referenced(struct page *page, int is_locked) 412{ 413 int referenced = 0; 414 415 if (page_test_and_clear_young(page)) 416 referenced++; 417 418 if (TestClearPageReferenced(page)) 419 referenced++; 420 421 if (page_mapped(page) && page->mapping) { 422 if (PageAnon(page)) 423 referenced += page_referenced_anon(page); 424 else if (is_locked) 425 referenced += page_referenced_file(page); 426 else if (TestSetPageLocked(page)) 427 referenced++; 428 else { 429 if (page->mapping) 430 referenced += page_referenced_file(page); 431 unlock_page(page); 432 } 433 } 434 return referenced; 435} 436 437/** 438 * page_set_anon_rmap - setup new anonymous rmap 439 * @page: the page to add the mapping to 440 * @vma: the vm area in which the mapping is added 441 * @address: the user virtual address mapped 442 */ 443static void __page_set_anon_rmap(struct page *page, 444 struct vm_area_struct *vma, unsigned long address) 445{ 446 struct anon_vma *anon_vma = vma->anon_vma; 447 448 BUG_ON(!anon_vma); 449 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 450 page->mapping = (struct address_space *) anon_vma; 451 452 page->index = linear_page_index(vma, address); 453 454 inc_page_state(nr_mapped); 455} 456 457/** 458 * page_add_anon_rmap - add pte mapping to an anonymous page 459 * @page: the page to add the mapping to 460 * @vma: the vm area in which the mapping is added 461 * @address: the user virtual address mapped 462 * 463 * The caller needs to hold the pte lock. 464 */ 465void page_add_anon_rmap(struct page *page, 466 struct vm_area_struct *vma, unsigned long address) 467{ 468 if (atomic_inc_and_test(&page->_mapcount)) 469 __page_set_anon_rmap(page, vma, address); 470 /* else checking page index and mapping is racy */ 471} 472 473/* 474 * page_add_new_anon_rmap - add pte mapping to a new anonymous page 475 * @page: the page to add the mapping to 476 * @vma: the vm area in which the mapping is added 477 * @address: the user virtual address mapped 478 * 479 * Same as page_add_anon_rmap but must only be called on *new* pages. 480 * This means the inc-and-test can be bypassed. 481 */ 482void page_add_new_anon_rmap(struct page *page, 483 struct vm_area_struct *vma, unsigned long address) 484{ 485 atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */ 486 __page_set_anon_rmap(page, vma, address); 487} 488 489/** 490 * page_add_file_rmap - add pte mapping to a file page 491 * @page: the page to add the mapping to 492 * 493 * The caller needs to hold the pte lock. 494 */ 495void page_add_file_rmap(struct page *page) 496{ 497 BUG_ON(PageAnon(page)); 498 BUG_ON(!pfn_valid(page_to_pfn(page))); 499 500 if (atomic_inc_and_test(&page->_mapcount)) 501 inc_page_state(nr_mapped); 502} 503 504/** 505 * page_remove_rmap - take down pte mapping from a page 506 * @page: page to remove mapping from 507 * 508 * The caller needs to hold the pte lock. 509 */ 510void page_remove_rmap(struct page *page) 511{ 512 if (atomic_add_negative(-1, &page->_mapcount)) { 513 BUG_ON(page_mapcount(page) < 0); 514 /* 515 * It would be tidy to reset the PageAnon mapping here, 516 * but that might overwrite a racing page_add_anon_rmap 517 * which increments mapcount after us but sets mapping 518 * before us: so leave the reset to free_hot_cold_page, 519 * and remember that it's only reliable while mapped. 520 * Leaving it set also helps swapoff to reinstate ptes 521 * faster for those pages still in swapcache. 522 */ 523 if (page_test_and_clear_dirty(page)) 524 set_page_dirty(page); 525 dec_page_state(nr_mapped); 526 } 527} 528 529/* 530 * Subfunctions of try_to_unmap: try_to_unmap_one called 531 * repeatedly from either try_to_unmap_anon or try_to_unmap_file. 532 */ 533static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma) 534{ 535 struct mm_struct *mm = vma->vm_mm; 536 unsigned long address; 537 pte_t *pte; 538 pte_t pteval; 539 spinlock_t *ptl; 540 int ret = SWAP_AGAIN; 541 542 address = vma_address(page, vma); 543 if (address == -EFAULT) 544 goto out; 545 546 pte = page_check_address(page, mm, address, &ptl); 547 if (!pte) 548 goto out; 549 550 /* 551 * If the page is mlock()d, we cannot swap it out. 552 * If it's recently referenced (perhaps page_referenced 553 * skipped over this mm) then we should reactivate it. 554 */ 555 if ((vma->vm_flags & VM_LOCKED) || 556 ptep_clear_flush_young(vma, address, pte)) { 557 ret = SWAP_FAIL; 558 goto out_unmap; 559 } 560 561 /* Nuke the page table entry. */ 562 flush_cache_page(vma, address, page_to_pfn(page)); 563 pteval = ptep_clear_flush(vma, address, pte); 564 565 /* Move the dirty bit to the physical page now the pte is gone. */ 566 if (pte_dirty(pteval)) 567 set_page_dirty(page); 568 569 /* Update high watermark before we lower rss */ 570 update_hiwater_rss(mm); 571 572 if (PageAnon(page)) { 573 swp_entry_t entry = { .val = page_private(page) }; 574 /* 575 * Store the swap location in the pte. 576 * See handle_pte_fault() ... 577 */ 578 BUG_ON(!PageSwapCache(page)); 579 swap_duplicate(entry); 580 if (list_empty(&mm->mmlist)) { 581 spin_lock(&mmlist_lock); 582 if (list_empty(&mm->mmlist)) 583 list_add(&mm->mmlist, &init_mm.mmlist); 584 spin_unlock(&mmlist_lock); 585 } 586 set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); 587 BUG_ON(pte_file(*pte)); 588 dec_mm_counter(mm, anon_rss); 589 } else 590 dec_mm_counter(mm, file_rss); 591 592 page_remove_rmap(page); 593 page_cache_release(page); 594 595out_unmap: 596 pte_unmap_unlock(pte, ptl); 597out: 598 return ret; 599} 600 601/* 602 * objrmap doesn't work for nonlinear VMAs because the assumption that 603 * offset-into-file correlates with offset-into-virtual-addresses does not hold. 604 * Consequently, given a particular page and its ->index, we cannot locate the 605 * ptes which are mapping that page without an exhaustive linear search. 606 * 607 * So what this code does is a mini "virtual scan" of each nonlinear VMA which 608 * maps the file to which the target page belongs. The ->vm_private_data field 609 * holds the current cursor into that scan. Successive searches will circulate 610 * around the vma's virtual address space. 611 * 612 * So as more replacement pressure is applied to the pages in a nonlinear VMA, 613 * more scanning pressure is placed against them as well. Eventually pages 614 * will become fully unmapped and are eligible for eviction. 615 * 616 * For very sparsely populated VMAs this is a little inefficient - chances are 617 * there there won't be many ptes located within the scan cluster. In this case 618 * maybe we could scan further - to the end of the pte page, perhaps. 619 */ 620#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) 621#define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) 622 623static void try_to_unmap_cluster(unsigned long cursor, 624 unsigned int *mapcount, struct vm_area_struct *vma) 625{ 626 struct mm_struct *mm = vma->vm_mm; 627 pgd_t *pgd; 628 pud_t *pud; 629 pmd_t *pmd; 630 pte_t *pte; 631 pte_t pteval; 632 spinlock_t *ptl; 633 struct page *page; 634 unsigned long address; 635 unsigned long end; 636 637 address = (vma->vm_start + cursor) & CLUSTER_MASK; 638 end = address + CLUSTER_SIZE; 639 if (address < vma->vm_start) 640 address = vma->vm_start; 641 if (end > vma->vm_end) 642 end = vma->vm_end; 643 644 pgd = pgd_offset(mm, address); 645 if (!pgd_present(*pgd)) 646 return; 647 648 pud = pud_offset(pgd, address); 649 if (!pud_present(*pud)) 650 return; 651 652 pmd = pmd_offset(pud, address); 653 if (!pmd_present(*pmd)) 654 return; 655 656 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 657 658 /* Update high watermark before we lower rss */ 659 update_hiwater_rss(mm); 660 661 for (; address < end; pte++, address += PAGE_SIZE) { 662 if (!pte_present(*pte)) 663 continue; 664 page = vm_normal_page(vma, address, *pte); 665 BUG_ON(!page || PageAnon(page)); 666 667 if (ptep_clear_flush_young(vma, address, pte)) 668 continue; 669 670 /* Nuke the page table entry. */ 671 flush_cache_page(vma, address, pte_pfn(*pte)); 672 pteval = ptep_clear_flush(vma, address, pte); 673 674 /* If nonlinear, store the file page offset in the pte. */ 675 if (page->index != linear_page_index(vma, address)) 676 set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); 677 678 /* Move the dirty bit to the physical page now the pte is gone. */ 679 if (pte_dirty(pteval)) 680 set_page_dirty(page); 681 682 page_remove_rmap(page); 683 page_cache_release(page); 684 dec_mm_counter(mm, file_rss); 685 (*mapcount)--; 686 } 687 pte_unmap_unlock(pte - 1, ptl); 688} 689 690static int try_to_unmap_anon(struct page *page) 691{ 692 struct anon_vma *anon_vma; 693 struct vm_area_struct *vma; 694 int ret = SWAP_AGAIN; 695 696 anon_vma = page_lock_anon_vma(page); 697 if (!anon_vma) 698 return ret; 699 700 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { 701 ret = try_to_unmap_one(page, vma); 702 if (ret == SWAP_FAIL || !page_mapped(page)) 703 break; 704 } 705 spin_unlock(&anon_vma->lock); 706 return ret; 707} 708 709/** 710 * try_to_unmap_file - unmap file page using the object-based rmap method 711 * @page: the page to unmap 712 * 713 * Find all the mappings of a page using the mapping pointer and the vma chains 714 * contained in the address_space struct it points to. 715 * 716 * This function is only called from try_to_unmap for object-based pages. 717 */ 718static int try_to_unmap_file(struct page *page) 719{ 720 struct address_space *mapping = page->mapping; 721 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 722 struct vm_area_struct *vma; 723 struct prio_tree_iter iter; 724 int ret = SWAP_AGAIN; 725 unsigned long cursor; 726 unsigned long max_nl_cursor = 0; 727 unsigned long max_nl_size = 0; 728 unsigned int mapcount; 729 730 spin_lock(&mapping->i_mmap_lock); 731 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 732 ret = try_to_unmap_one(page, vma); 733 if (ret == SWAP_FAIL || !page_mapped(page)) 734 goto out; 735 } 736 737 if (list_empty(&mapping->i_mmap_nonlinear)) 738 goto out; 739 740 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, 741 shared.vm_set.list) { 742 if (vma->vm_flags & VM_LOCKED) 743 continue; 744 cursor = (unsigned long) vma->vm_private_data; 745 if (cursor > max_nl_cursor) 746 max_nl_cursor = cursor; 747 cursor = vma->vm_end - vma->vm_start; 748 if (cursor > max_nl_size) 749 max_nl_size = cursor; 750 } 751 752 if (max_nl_size == 0) { /* any nonlinears locked or reserved */ 753 ret = SWAP_FAIL; 754 goto out; 755 } 756 757 /* 758 * We don't try to search for this page in the nonlinear vmas, 759 * and page_referenced wouldn't have found it anyway. Instead 760 * just walk the nonlinear vmas trying to age and unmap some. 761 * The mapcount of the page we came in with is irrelevant, 762 * but even so use it as a guide to how hard we should try? 763 */ 764 mapcount = page_mapcount(page); 765 if (!mapcount) 766 goto out; 767 cond_resched_lock(&mapping->i_mmap_lock); 768 769 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; 770 if (max_nl_cursor == 0) 771 max_nl_cursor = CLUSTER_SIZE; 772 773 do { 774 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, 775 shared.vm_set.list) { 776 if (vma->vm_flags & VM_LOCKED) 777 continue; 778 cursor = (unsigned long) vma->vm_private_data; 779 while ( cursor < max_nl_cursor && 780 cursor < vma->vm_end - vma->vm_start) { 781 try_to_unmap_cluster(cursor, &mapcount, vma); 782 cursor += CLUSTER_SIZE; 783 vma->vm_private_data = (void *) cursor; 784 if ((int)mapcount <= 0) 785 goto out; 786 } 787 vma->vm_private_data = (void *) max_nl_cursor; 788 } 789 cond_resched_lock(&mapping->i_mmap_lock); 790 max_nl_cursor += CLUSTER_SIZE; 791 } while (max_nl_cursor <= max_nl_size); 792 793 /* 794 * Don't loop forever (perhaps all the remaining pages are 795 * in locked vmas). Reset cursor on all unreserved nonlinear 796 * vmas, now forgetting on which ones it had fallen behind. 797 */ 798 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) 799 vma->vm_private_data = NULL; 800out: 801 spin_unlock(&mapping->i_mmap_lock); 802 return ret; 803} 804 805/** 806 * try_to_unmap - try to remove all page table mappings to a page 807 * @page: the page to get unmapped 808 * 809 * Tries to remove all the page table entries which are mapping this 810 * page, used in the pageout path. Caller must hold the page lock. 811 * Return values are: 812 * 813 * SWAP_SUCCESS - we succeeded in removing all mappings 814 * SWAP_AGAIN - we missed a mapping, try again later 815 * SWAP_FAIL - the page is unswappable 816 */ 817int try_to_unmap(struct page *page) 818{ 819 int ret; 820 821 BUG_ON(!PageLocked(page)); 822 823 if (PageAnon(page)) 824 ret = try_to_unmap_anon(page); 825 else 826 ret = try_to_unmap_file(page); 827 828 if (!page_mapped(page)) 829 ret = SWAP_SUCCESS; 830 return ret; 831} 832 833