USB: misc: iowarrior: clean up urb->status usage
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / rmap.c
blobfede5c7910beebe57894af31a37db9cb8fc8dee4
1 /*
2 * mm/rmap.c - physical to virtual reverse mappings
4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5 * Released under the General Public License (GPL).
7 * Simple, low overhead reverse mapping scheme.
8 * Please try to keep this thing as modular as possible.
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.
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
21 * Lock ordering in mm:
23 * inode->i_mutex (while writing or truncating, not reading or faulting)
24 * inode->i_alloc_sem (vmtruncate_range)
25 * mm->mmap_sem
26 * page->flags PG_locked (lock_page)
27 * mapping->i_mmap_lock
28 * anon_vma->lock
29 * mm->page_table_lock or pte_lock
30 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
34 * inode_lock (in set_page_dirty's __mark_inode_dirty)
35 * sb_lock (within inode_lock in fs/fs-writeback.c)
36 * mapping->tree_lock (widely used, in set_page_dirty,
37 * in arch-dependent flush_dcache_mmap_lock,
38 * within inode_lock in __sync_single_inode)
41 #include <linux/mm.h>
42 #include <linux/pagemap.h>
43 #include <linux/swap.h>
44 #include <linux/swapops.h>
45 #include <linux/slab.h>
46 #include <linux/init.h>
47 #include <linux/rmap.h>
48 #include <linux/rcupdate.h>
49 #include <linux/module.h>
50 #include <linux/kallsyms.h>
52 #include <asm/tlbflush.h>
54 struct kmem_cache *anon_vma_cachep;
56 /* This must be called under the mmap_sem. */
57 int anon_vma_prepare(struct vm_area_struct *vma)
59 struct anon_vma *anon_vma = vma->anon_vma;
61 might_sleep();
62 if (unlikely(!anon_vma)) {
63 struct mm_struct *mm = vma->vm_mm;
64 struct anon_vma *allocated, *locked;
66 anon_vma = find_mergeable_anon_vma(vma);
67 if (anon_vma) {
68 allocated = NULL;
69 locked = anon_vma;
70 spin_lock(&locked->lock);
71 } else {
72 anon_vma = anon_vma_alloc();
73 if (unlikely(!anon_vma))
74 return -ENOMEM;
75 allocated = anon_vma;
76 locked = NULL;
79 /* page_table_lock to protect against threads */
80 spin_lock(&mm->page_table_lock);
81 if (likely(!vma->anon_vma)) {
82 vma->anon_vma = anon_vma;
83 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
84 allocated = NULL;
86 spin_unlock(&mm->page_table_lock);
88 if (locked)
89 spin_unlock(&locked->lock);
90 if (unlikely(allocated))
91 anon_vma_free(allocated);
93 return 0;
96 void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
98 BUG_ON(vma->anon_vma != next->anon_vma);
99 list_del(&next->anon_vma_node);
102 void __anon_vma_link(struct vm_area_struct *vma)
104 struct anon_vma *anon_vma = vma->anon_vma;
106 if (anon_vma)
107 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
110 void anon_vma_link(struct vm_area_struct *vma)
112 struct anon_vma *anon_vma = vma->anon_vma;
114 if (anon_vma) {
115 spin_lock(&anon_vma->lock);
116 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
117 spin_unlock(&anon_vma->lock);
121 void anon_vma_unlink(struct vm_area_struct *vma)
123 struct anon_vma *anon_vma = vma->anon_vma;
124 int empty;
126 if (!anon_vma)
127 return;
129 spin_lock(&anon_vma->lock);
130 list_del(&vma->anon_vma_node);
132 /* We must garbage collect the anon_vma if it's empty */
133 empty = list_empty(&anon_vma->head);
134 spin_unlock(&anon_vma->lock);
136 if (empty)
137 anon_vma_free(anon_vma);
140 static void anon_vma_ctor(void *data, struct kmem_cache *cachep,
141 unsigned long flags)
143 struct anon_vma *anon_vma = data;
145 spin_lock_init(&anon_vma->lock);
146 INIT_LIST_HEAD(&anon_vma->head);
149 void __init anon_vma_init(void)
151 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
152 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL);
156 * Getting a lock on a stable anon_vma from a page off the LRU is
157 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
159 static struct anon_vma *page_lock_anon_vma(struct page *page)
161 struct anon_vma *anon_vma;
162 unsigned long anon_mapping;
164 rcu_read_lock();
165 anon_mapping = (unsigned long) page->mapping;
166 if (!(anon_mapping & PAGE_MAPPING_ANON))
167 goto out;
168 if (!page_mapped(page))
169 goto out;
171 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
172 spin_lock(&anon_vma->lock);
173 return anon_vma;
174 out:
175 rcu_read_unlock();
176 return NULL;
179 static void page_unlock_anon_vma(struct anon_vma *anon_vma)
181 spin_unlock(&anon_vma->lock);
182 rcu_read_unlock();
186 * At what user virtual address is page expected in vma?
188 static inline unsigned long
189 vma_address(struct page *page, struct vm_area_struct *vma)
191 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
192 unsigned long address;
194 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
195 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
196 /* page should be within any vma from prio_tree_next */
197 BUG_ON(!PageAnon(page));
198 return -EFAULT;
200 return address;
204 * At what user virtual address is page expected in vma? checking that the
205 * page matches the vma: currently only used on anon pages, by unuse_vma;
207 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
209 if (PageAnon(page)) {
210 if ((void *)vma->anon_vma !=
211 (void *)page->mapping - PAGE_MAPPING_ANON)
212 return -EFAULT;
213 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
214 if (!vma->vm_file ||
215 vma->vm_file->f_mapping != page->mapping)
216 return -EFAULT;
217 } else
218 return -EFAULT;
219 return vma_address(page, vma);
223 * Check that @page is mapped at @address into @mm.
225 * On success returns with pte mapped and locked.
227 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
228 unsigned long address, spinlock_t **ptlp)
230 pgd_t *pgd;
231 pud_t *pud;
232 pmd_t *pmd;
233 pte_t *pte;
234 spinlock_t *ptl;
236 pgd = pgd_offset(mm, address);
237 if (!pgd_present(*pgd))
238 return NULL;
240 pud = pud_offset(pgd, address);
241 if (!pud_present(*pud))
242 return NULL;
244 pmd = pmd_offset(pud, address);
245 if (!pmd_present(*pmd))
246 return NULL;
248 pte = pte_offset_map(pmd, address);
249 /* Make a quick check before getting the lock */
250 if (!pte_present(*pte)) {
251 pte_unmap(pte);
252 return NULL;
255 ptl = pte_lockptr(mm, pmd);
256 spin_lock(ptl);
257 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
258 *ptlp = ptl;
259 return pte;
261 pte_unmap_unlock(pte, ptl);
262 return NULL;
266 * Subfunctions of page_referenced: page_referenced_one called
267 * repeatedly from either page_referenced_anon or page_referenced_file.
269 static int page_referenced_one(struct page *page,
270 struct vm_area_struct *vma, unsigned int *mapcount)
272 struct mm_struct *mm = vma->vm_mm;
273 unsigned long address;
274 pte_t *pte;
275 spinlock_t *ptl;
276 int referenced = 0;
278 address = vma_address(page, vma);
279 if (address == -EFAULT)
280 goto out;
282 pte = page_check_address(page, mm, address, &ptl);
283 if (!pte)
284 goto out;
286 if (ptep_clear_flush_young(vma, address, pte))
287 referenced++;
289 /* Pretend the page is referenced if the task has the
290 swap token and is in the middle of a page fault. */
291 if (mm != current->mm && has_swap_token(mm) &&
292 rwsem_is_locked(&mm->mmap_sem))
293 referenced++;
295 (*mapcount)--;
296 pte_unmap_unlock(pte, ptl);
297 out:
298 return referenced;
301 static int page_referenced_anon(struct page *page)
303 unsigned int mapcount;
304 struct anon_vma *anon_vma;
305 struct vm_area_struct *vma;
306 int referenced = 0;
308 anon_vma = page_lock_anon_vma(page);
309 if (!anon_vma)
310 return referenced;
312 mapcount = page_mapcount(page);
313 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
314 referenced += page_referenced_one(page, vma, &mapcount);
315 if (!mapcount)
316 break;
319 page_unlock_anon_vma(anon_vma);
320 return referenced;
324 * page_referenced_file - referenced check for object-based rmap
325 * @page: the page we're checking references on.
327 * For an object-based mapped page, find all the places it is mapped and
328 * check/clear the referenced flag. This is done by following the page->mapping
329 * pointer, then walking the chain of vmas it holds. It returns the number
330 * of references it found.
332 * This function is only called from page_referenced for object-based pages.
334 static int page_referenced_file(struct page *page)
336 unsigned int mapcount;
337 struct address_space *mapping = page->mapping;
338 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
339 struct vm_area_struct *vma;
340 struct prio_tree_iter iter;
341 int referenced = 0;
344 * The caller's checks on page->mapping and !PageAnon have made
345 * sure that this is a file page: the check for page->mapping
346 * excludes the case just before it gets set on an anon page.
348 BUG_ON(PageAnon(page));
351 * The page lock not only makes sure that page->mapping cannot
352 * suddenly be NULLified by truncation, it makes sure that the
353 * structure at mapping cannot be freed and reused yet,
354 * so we can safely take mapping->i_mmap_lock.
356 BUG_ON(!PageLocked(page));
358 spin_lock(&mapping->i_mmap_lock);
361 * i_mmap_lock does not stabilize mapcount at all, but mapcount
362 * is more likely to be accurate if we note it after spinning.
364 mapcount = page_mapcount(page);
366 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
367 if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
368 == (VM_LOCKED|VM_MAYSHARE)) {
369 referenced++;
370 break;
372 referenced += page_referenced_one(page, vma, &mapcount);
373 if (!mapcount)
374 break;
377 spin_unlock(&mapping->i_mmap_lock);
378 return referenced;
382 * page_referenced - test if the page was referenced
383 * @page: the page to test
384 * @is_locked: caller holds lock on the page
386 * Quick test_and_clear_referenced for all mappings to a page,
387 * returns the number of ptes which referenced the page.
389 int page_referenced(struct page *page, int is_locked)
391 int referenced = 0;
393 if (page_test_and_clear_young(page))
394 referenced++;
396 if (TestClearPageReferenced(page))
397 referenced++;
399 if (page_mapped(page) && page->mapping) {
400 if (PageAnon(page))
401 referenced += page_referenced_anon(page);
402 else if (is_locked)
403 referenced += page_referenced_file(page);
404 else if (TestSetPageLocked(page))
405 referenced++;
406 else {
407 if (page->mapping)
408 referenced += page_referenced_file(page);
409 unlock_page(page);
412 return referenced;
415 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma)
417 struct mm_struct *mm = vma->vm_mm;
418 unsigned long address;
419 pte_t *pte;
420 spinlock_t *ptl;
421 int ret = 0;
423 address = vma_address(page, vma);
424 if (address == -EFAULT)
425 goto out;
427 pte = page_check_address(page, mm, address, &ptl);
428 if (!pte)
429 goto out;
431 if (pte_dirty(*pte) || pte_write(*pte)) {
432 pte_t entry;
434 flush_cache_page(vma, address, pte_pfn(*pte));
435 entry = ptep_clear_flush(vma, address, pte);
436 entry = pte_wrprotect(entry);
437 entry = pte_mkclean(entry);
438 set_pte_at(mm, address, pte, entry);
439 lazy_mmu_prot_update(entry);
440 ret = 1;
443 pte_unmap_unlock(pte, ptl);
444 out:
445 return ret;
448 static int page_mkclean_file(struct address_space *mapping, struct page *page)
450 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
451 struct vm_area_struct *vma;
452 struct prio_tree_iter iter;
453 int ret = 0;
455 BUG_ON(PageAnon(page));
457 spin_lock(&mapping->i_mmap_lock);
458 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
459 if (vma->vm_flags & VM_SHARED)
460 ret += page_mkclean_one(page, vma);
462 spin_unlock(&mapping->i_mmap_lock);
463 return ret;
466 int page_mkclean(struct page *page)
468 int ret = 0;
470 BUG_ON(!PageLocked(page));
472 if (page_mapped(page)) {
473 struct address_space *mapping = page_mapping(page);
474 if (mapping)
475 ret = page_mkclean_file(mapping, page);
476 if (page_test_dirty(page)) {
477 page_clear_dirty(page);
478 ret = 1;
482 return ret;
484 EXPORT_SYMBOL_GPL(page_mkclean);
487 * page_set_anon_rmap - setup new anonymous rmap
488 * @page: the page to add the mapping to
489 * @vma: the vm area in which the mapping is added
490 * @address: the user virtual address mapped
492 static void __page_set_anon_rmap(struct page *page,
493 struct vm_area_struct *vma, unsigned long address)
495 struct anon_vma *anon_vma = vma->anon_vma;
497 BUG_ON(!anon_vma);
498 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
499 page->mapping = (struct address_space *) anon_vma;
501 page->index = linear_page_index(vma, address);
504 * nr_mapped state can be updated without turning off
505 * interrupts because it is not modified via interrupt.
507 __inc_zone_page_state(page, NR_ANON_PAGES);
511 * page_set_anon_rmap - sanity check anonymous rmap addition
512 * @page: the page to add the mapping to
513 * @vma: the vm area in which the mapping is added
514 * @address: the user virtual address mapped
516 static void __page_check_anon_rmap(struct page *page,
517 struct vm_area_struct *vma, unsigned long address)
519 #ifdef CONFIG_DEBUG_VM
521 * The page's anon-rmap details (mapping and index) are guaranteed to
522 * be set up correctly at this point.
524 * We have exclusion against page_add_anon_rmap because the caller
525 * always holds the page locked, except if called from page_dup_rmap,
526 * in which case the page is already known to be setup.
528 * We have exclusion against page_add_new_anon_rmap because those pages
529 * are initially only visible via the pagetables, and the pte is locked
530 * over the call to page_add_new_anon_rmap.
532 struct anon_vma *anon_vma = vma->anon_vma;
533 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
534 BUG_ON(page->mapping != (struct address_space *)anon_vma);
535 BUG_ON(page->index != linear_page_index(vma, address));
536 #endif
540 * page_add_anon_rmap - add pte mapping to an anonymous page
541 * @page: the page to add the mapping to
542 * @vma: the vm area in which the mapping is added
543 * @address: the user virtual address mapped
545 * The caller needs to hold the pte lock and the page must be locked.
547 void page_add_anon_rmap(struct page *page,
548 struct vm_area_struct *vma, unsigned long address)
550 VM_BUG_ON(!PageLocked(page));
551 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
552 if (atomic_inc_and_test(&page->_mapcount))
553 __page_set_anon_rmap(page, vma, address);
554 else
555 __page_check_anon_rmap(page, vma, address);
559 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
560 * @page: the page to add the mapping to
561 * @vma: the vm area in which the mapping is added
562 * @address: the user virtual address mapped
564 * Same as page_add_anon_rmap but must only be called on *new* pages.
565 * This means the inc-and-test can be bypassed.
566 * Page does not have to be locked.
568 void page_add_new_anon_rmap(struct page *page,
569 struct vm_area_struct *vma, unsigned long address)
571 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
572 atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */
573 __page_set_anon_rmap(page, vma, address);
577 * page_add_file_rmap - add pte mapping to a file page
578 * @page: the page to add the mapping to
580 * The caller needs to hold the pte lock.
582 void page_add_file_rmap(struct page *page)
584 if (atomic_inc_and_test(&page->_mapcount))
585 __inc_zone_page_state(page, NR_FILE_MAPPED);
588 #ifdef CONFIG_DEBUG_VM
590 * page_dup_rmap - duplicate pte mapping to a page
591 * @page: the page to add the mapping to
593 * For copy_page_range only: minimal extract from page_add_file_rmap /
594 * page_add_anon_rmap, avoiding unnecessary tests (already checked) so it's
595 * quicker.
597 * The caller needs to hold the pte lock.
599 void page_dup_rmap(struct page *page, struct vm_area_struct *vma, unsigned long address)
601 BUG_ON(page_mapcount(page) == 0);
602 if (PageAnon(page))
603 __page_check_anon_rmap(page, vma, address);
604 atomic_inc(&page->_mapcount);
606 #endif
609 * page_remove_rmap - take down pte mapping from a page
610 * @page: page to remove mapping from
612 * The caller needs to hold the pte lock.
614 void page_remove_rmap(struct page *page, struct vm_area_struct *vma)
616 if (atomic_add_negative(-1, &page->_mapcount)) {
617 if (unlikely(page_mapcount(page) < 0)) {
618 printk (KERN_EMERG "Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page));
619 printk (KERN_EMERG " page pfn = %lx\n", page_to_pfn(page));
620 printk (KERN_EMERG " page->flags = %lx\n", page->flags);
621 printk (KERN_EMERG " page->count = %x\n", page_count(page));
622 printk (KERN_EMERG " page->mapping = %p\n", page->mapping);
623 print_symbol (KERN_EMERG " vma->vm_ops = %s\n", (unsigned long)vma->vm_ops);
624 if (vma->vm_ops) {
625 print_symbol (KERN_EMERG " vma->vm_ops->nopage = %s\n", (unsigned long)vma->vm_ops->nopage);
626 print_symbol (KERN_EMERG " vma->vm_ops->fault = %s\n", (unsigned long)vma->vm_ops->fault);
628 if (vma->vm_file && vma->vm_file->f_op)
629 print_symbol (KERN_EMERG " vma->vm_file->f_op->mmap = %s\n", (unsigned long)vma->vm_file->f_op->mmap);
630 BUG();
634 * It would be tidy to reset the PageAnon mapping here,
635 * but that might overwrite a racing page_add_anon_rmap
636 * which increments mapcount after us but sets mapping
637 * before us: so leave the reset to free_hot_cold_page,
638 * and remember that it's only reliable while mapped.
639 * Leaving it set also helps swapoff to reinstate ptes
640 * faster for those pages still in swapcache.
642 if (page_test_dirty(page)) {
643 page_clear_dirty(page);
644 set_page_dirty(page);
646 __dec_zone_page_state(page,
647 PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED);
652 * Subfunctions of try_to_unmap: try_to_unmap_one called
653 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
655 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
656 int migration)
658 struct mm_struct *mm = vma->vm_mm;
659 unsigned long address;
660 pte_t *pte;
661 pte_t pteval;
662 spinlock_t *ptl;
663 int ret = SWAP_AGAIN;
665 address = vma_address(page, vma);
666 if (address == -EFAULT)
667 goto out;
669 pte = page_check_address(page, mm, address, &ptl);
670 if (!pte)
671 goto out;
674 * If the page is mlock()d, we cannot swap it out.
675 * If it's recently referenced (perhaps page_referenced
676 * skipped over this mm) then we should reactivate it.
678 if (!migration && ((vma->vm_flags & VM_LOCKED) ||
679 (ptep_clear_flush_young(vma, address, pte)))) {
680 ret = SWAP_FAIL;
681 goto out_unmap;
684 /* Nuke the page table entry. */
685 flush_cache_page(vma, address, page_to_pfn(page));
686 pteval = ptep_clear_flush(vma, address, pte);
688 /* Move the dirty bit to the physical page now the pte is gone. */
689 if (pte_dirty(pteval))
690 set_page_dirty(page);
692 /* Update high watermark before we lower rss */
693 update_hiwater_rss(mm);
695 if (PageAnon(page)) {
696 swp_entry_t entry = { .val = page_private(page) };
698 if (PageSwapCache(page)) {
700 * Store the swap location in the pte.
701 * See handle_pte_fault() ...
703 swap_duplicate(entry);
704 if (list_empty(&mm->mmlist)) {
705 spin_lock(&mmlist_lock);
706 if (list_empty(&mm->mmlist))
707 list_add(&mm->mmlist, &init_mm.mmlist);
708 spin_unlock(&mmlist_lock);
710 dec_mm_counter(mm, anon_rss);
711 #ifdef CONFIG_MIGRATION
712 } else {
714 * Store the pfn of the page in a special migration
715 * pte. do_swap_page() will wait until the migration
716 * pte is removed and then restart fault handling.
718 BUG_ON(!migration);
719 entry = make_migration_entry(page, pte_write(pteval));
720 #endif
722 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
723 BUG_ON(pte_file(*pte));
724 } else
725 #ifdef CONFIG_MIGRATION
726 if (migration) {
727 /* Establish migration entry for a file page */
728 swp_entry_t entry;
729 entry = make_migration_entry(page, pte_write(pteval));
730 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
731 } else
732 #endif
733 dec_mm_counter(mm, file_rss);
736 page_remove_rmap(page, vma);
737 page_cache_release(page);
739 out_unmap:
740 pte_unmap_unlock(pte, ptl);
741 out:
742 return ret;
746 * objrmap doesn't work for nonlinear VMAs because the assumption that
747 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
748 * Consequently, given a particular page and its ->index, we cannot locate the
749 * ptes which are mapping that page without an exhaustive linear search.
751 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
752 * maps the file to which the target page belongs. The ->vm_private_data field
753 * holds the current cursor into that scan. Successive searches will circulate
754 * around the vma's virtual address space.
756 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
757 * more scanning pressure is placed against them as well. Eventually pages
758 * will become fully unmapped and are eligible for eviction.
760 * For very sparsely populated VMAs this is a little inefficient - chances are
761 * there there won't be many ptes located within the scan cluster. In this case
762 * maybe we could scan further - to the end of the pte page, perhaps.
764 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
765 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
767 static void try_to_unmap_cluster(unsigned long cursor,
768 unsigned int *mapcount, struct vm_area_struct *vma)
770 struct mm_struct *mm = vma->vm_mm;
771 pgd_t *pgd;
772 pud_t *pud;
773 pmd_t *pmd;
774 pte_t *pte;
775 pte_t pteval;
776 spinlock_t *ptl;
777 struct page *page;
778 unsigned long address;
779 unsigned long end;
781 address = (vma->vm_start + cursor) & CLUSTER_MASK;
782 end = address + CLUSTER_SIZE;
783 if (address < vma->vm_start)
784 address = vma->vm_start;
785 if (end > vma->vm_end)
786 end = vma->vm_end;
788 pgd = pgd_offset(mm, address);
789 if (!pgd_present(*pgd))
790 return;
792 pud = pud_offset(pgd, address);
793 if (!pud_present(*pud))
794 return;
796 pmd = pmd_offset(pud, address);
797 if (!pmd_present(*pmd))
798 return;
800 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
802 /* Update high watermark before we lower rss */
803 update_hiwater_rss(mm);
805 for (; address < end; pte++, address += PAGE_SIZE) {
806 if (!pte_present(*pte))
807 continue;
808 page = vm_normal_page(vma, address, *pte);
809 BUG_ON(!page || PageAnon(page));
811 if (ptep_clear_flush_young(vma, address, pte))
812 continue;
814 /* Nuke the page table entry. */
815 flush_cache_page(vma, address, pte_pfn(*pte));
816 pteval = ptep_clear_flush(vma, address, pte);
818 /* If nonlinear, store the file page offset in the pte. */
819 if (page->index != linear_page_index(vma, address))
820 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
822 /* Move the dirty bit to the physical page now the pte is gone. */
823 if (pte_dirty(pteval))
824 set_page_dirty(page);
826 page_remove_rmap(page, vma);
827 page_cache_release(page);
828 dec_mm_counter(mm, file_rss);
829 (*mapcount)--;
831 pte_unmap_unlock(pte - 1, ptl);
834 static int try_to_unmap_anon(struct page *page, int migration)
836 struct anon_vma *anon_vma;
837 struct vm_area_struct *vma;
838 int ret = SWAP_AGAIN;
840 anon_vma = page_lock_anon_vma(page);
841 if (!anon_vma)
842 return ret;
844 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
845 ret = try_to_unmap_one(page, vma, migration);
846 if (ret == SWAP_FAIL || !page_mapped(page))
847 break;
850 page_unlock_anon_vma(anon_vma);
851 return ret;
855 * try_to_unmap_file - unmap file page using the object-based rmap method
856 * @page: the page to unmap
858 * Find all the mappings of a page using the mapping pointer and the vma chains
859 * contained in the address_space struct it points to.
861 * This function is only called from try_to_unmap for object-based pages.
863 static int try_to_unmap_file(struct page *page, int migration)
865 struct address_space *mapping = page->mapping;
866 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
867 struct vm_area_struct *vma;
868 struct prio_tree_iter iter;
869 int ret = SWAP_AGAIN;
870 unsigned long cursor;
871 unsigned long max_nl_cursor = 0;
872 unsigned long max_nl_size = 0;
873 unsigned int mapcount;
875 spin_lock(&mapping->i_mmap_lock);
876 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
877 ret = try_to_unmap_one(page, vma, migration);
878 if (ret == SWAP_FAIL || !page_mapped(page))
879 goto out;
882 if (list_empty(&mapping->i_mmap_nonlinear))
883 goto out;
885 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
886 shared.vm_set.list) {
887 if ((vma->vm_flags & VM_LOCKED) && !migration)
888 continue;
889 cursor = (unsigned long) vma->vm_private_data;
890 if (cursor > max_nl_cursor)
891 max_nl_cursor = cursor;
892 cursor = vma->vm_end - vma->vm_start;
893 if (cursor > max_nl_size)
894 max_nl_size = cursor;
897 if (max_nl_size == 0) { /* any nonlinears locked or reserved */
898 ret = SWAP_FAIL;
899 goto out;
903 * We don't try to search for this page in the nonlinear vmas,
904 * and page_referenced wouldn't have found it anyway. Instead
905 * just walk the nonlinear vmas trying to age and unmap some.
906 * The mapcount of the page we came in with is irrelevant,
907 * but even so use it as a guide to how hard we should try?
909 mapcount = page_mapcount(page);
910 if (!mapcount)
911 goto out;
912 cond_resched_lock(&mapping->i_mmap_lock);
914 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
915 if (max_nl_cursor == 0)
916 max_nl_cursor = CLUSTER_SIZE;
918 do {
919 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
920 shared.vm_set.list) {
921 if ((vma->vm_flags & VM_LOCKED) && !migration)
922 continue;
923 cursor = (unsigned long) vma->vm_private_data;
924 while ( cursor < max_nl_cursor &&
925 cursor < vma->vm_end - vma->vm_start) {
926 try_to_unmap_cluster(cursor, &mapcount, vma);
927 cursor += CLUSTER_SIZE;
928 vma->vm_private_data = (void *) cursor;
929 if ((int)mapcount <= 0)
930 goto out;
932 vma->vm_private_data = (void *) max_nl_cursor;
934 cond_resched_lock(&mapping->i_mmap_lock);
935 max_nl_cursor += CLUSTER_SIZE;
936 } while (max_nl_cursor <= max_nl_size);
939 * Don't loop forever (perhaps all the remaining pages are
940 * in locked vmas). Reset cursor on all unreserved nonlinear
941 * vmas, now forgetting on which ones it had fallen behind.
943 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
944 vma->vm_private_data = NULL;
945 out:
946 spin_unlock(&mapping->i_mmap_lock);
947 return ret;
951 * try_to_unmap - try to remove all page table mappings to a page
952 * @page: the page to get unmapped
954 * Tries to remove all the page table entries which are mapping this
955 * page, used in the pageout path. Caller must hold the page lock.
956 * Return values are:
958 * SWAP_SUCCESS - we succeeded in removing all mappings
959 * SWAP_AGAIN - we missed a mapping, try again later
960 * SWAP_FAIL - the page is unswappable
962 int try_to_unmap(struct page *page, int migration)
964 int ret;
966 BUG_ON(!PageLocked(page));
968 if (PageAnon(page))
969 ret = try_to_unmap_anon(page, migration);
970 else
971 ret = try_to_unmap_file(page, migration);
973 if (!page_mapped(page))
974 ret = SWAP_SUCCESS;
975 return ret;