Merge commit 'jwb/next' into next
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / rmap.c
blobdd43373a483fa764ce35a3bdc541ea4148653af4
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 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)
40 * (code doesn't rely on that order so it could be switched around)
41 * ->tasklist_lock
42 * anon_vma->lock (memory_failure, collect_procs_anon)
43 * pte map lock
46 #include <linux/mm.h>
47 #include <linux/pagemap.h>
48 #include <linux/swap.h>
49 #include <linux/swapops.h>
50 #include <linux/slab.h>
51 #include <linux/init.h>
52 #include <linux/rmap.h>
53 #include <linux/rcupdate.h>
54 #include <linux/module.h>
55 #include <linux/memcontrol.h>
56 #include <linux/mmu_notifier.h>
57 #include <linux/migrate.h>
59 #include <asm/tlbflush.h>
61 #include "internal.h"
63 static struct kmem_cache *anon_vma_cachep;
65 static inline struct anon_vma *anon_vma_alloc(void)
67 return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
70 static inline void anon_vma_free(struct anon_vma *anon_vma)
72 kmem_cache_free(anon_vma_cachep, anon_vma);
75 /**
76 * anon_vma_prepare - attach an anon_vma to a memory region
77 * @vma: the memory region in question
79 * This makes sure the memory mapping described by 'vma' has
80 * an 'anon_vma' attached to it, so that we can associate the
81 * anonymous pages mapped into it with that anon_vma.
83 * The common case will be that we already have one, but if
84 * if not we either need to find an adjacent mapping that we
85 * can re-use the anon_vma from (very common when the only
86 * reason for splitting a vma has been mprotect()), or we
87 * allocate a new one.
89 * Anon-vma allocations are very subtle, because we may have
90 * optimistically looked up an anon_vma in page_lock_anon_vma()
91 * and that may actually touch the spinlock even in the newly
92 * allocated vma (it depends on RCU to make sure that the
93 * anon_vma isn't actually destroyed).
95 * As a result, we need to do proper anon_vma locking even
96 * for the new allocation. At the same time, we do not want
97 * to do any locking for the common case of already having
98 * an anon_vma.
100 * This must be called with the mmap_sem held for reading.
102 int anon_vma_prepare(struct vm_area_struct *vma)
104 struct anon_vma *anon_vma = vma->anon_vma;
106 might_sleep();
107 if (unlikely(!anon_vma)) {
108 struct mm_struct *mm = vma->vm_mm;
109 struct anon_vma *allocated;
111 anon_vma = find_mergeable_anon_vma(vma);
112 allocated = NULL;
113 if (!anon_vma) {
114 anon_vma = anon_vma_alloc();
115 if (unlikely(!anon_vma))
116 return -ENOMEM;
117 allocated = anon_vma;
119 spin_lock(&anon_vma->lock);
121 /* page_table_lock to protect against threads */
122 spin_lock(&mm->page_table_lock);
123 if (likely(!vma->anon_vma)) {
124 vma->anon_vma = anon_vma;
125 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
126 allocated = NULL;
128 spin_unlock(&mm->page_table_lock);
130 spin_unlock(&anon_vma->lock);
131 if (unlikely(allocated))
132 anon_vma_free(allocated);
134 return 0;
137 void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
139 BUG_ON(vma->anon_vma != next->anon_vma);
140 list_del(&next->anon_vma_node);
143 void __anon_vma_link(struct vm_area_struct *vma)
145 struct anon_vma *anon_vma = vma->anon_vma;
147 if (anon_vma)
148 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
151 void anon_vma_link(struct vm_area_struct *vma)
153 struct anon_vma *anon_vma = vma->anon_vma;
155 if (anon_vma) {
156 spin_lock(&anon_vma->lock);
157 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
158 spin_unlock(&anon_vma->lock);
162 void anon_vma_unlink(struct vm_area_struct *vma)
164 struct anon_vma *anon_vma = vma->anon_vma;
165 int empty;
167 if (!anon_vma)
168 return;
170 spin_lock(&anon_vma->lock);
171 list_del(&vma->anon_vma_node);
173 /* We must garbage collect the anon_vma if it's empty */
174 empty = list_empty(&anon_vma->head);
175 spin_unlock(&anon_vma->lock);
177 if (empty)
178 anon_vma_free(anon_vma);
181 static void anon_vma_ctor(void *data)
183 struct anon_vma *anon_vma = data;
185 spin_lock_init(&anon_vma->lock);
186 INIT_LIST_HEAD(&anon_vma->head);
189 void __init anon_vma_init(void)
191 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
192 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
196 * Getting a lock on a stable anon_vma from a page off the LRU is
197 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
199 struct anon_vma *page_lock_anon_vma(struct page *page)
201 struct anon_vma *anon_vma;
202 unsigned long anon_mapping;
204 rcu_read_lock();
205 anon_mapping = (unsigned long) page->mapping;
206 if (!(anon_mapping & PAGE_MAPPING_ANON))
207 goto out;
208 if (!page_mapped(page))
209 goto out;
211 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
212 spin_lock(&anon_vma->lock);
213 return anon_vma;
214 out:
215 rcu_read_unlock();
216 return NULL;
219 void page_unlock_anon_vma(struct anon_vma *anon_vma)
221 spin_unlock(&anon_vma->lock);
222 rcu_read_unlock();
226 * At what user virtual address is page expected in @vma?
227 * Returns virtual address or -EFAULT if page's index/offset is not
228 * within the range mapped the @vma.
230 static inline unsigned long
231 vma_address(struct page *page, struct vm_area_struct *vma)
233 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
234 unsigned long address;
236 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
237 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
238 /* page should be within @vma mapping range */
239 return -EFAULT;
241 return address;
245 * At what user virtual address is page expected in vma?
246 * checking that the page matches the vma.
248 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
250 if (PageAnon(page)) {
251 if ((void *)vma->anon_vma !=
252 (void *)page->mapping - PAGE_MAPPING_ANON)
253 return -EFAULT;
254 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
255 if (!vma->vm_file ||
256 vma->vm_file->f_mapping != page->mapping)
257 return -EFAULT;
258 } else
259 return -EFAULT;
260 return vma_address(page, vma);
264 * Check that @page is mapped at @address into @mm.
266 * If @sync is false, page_check_address may perform a racy check to avoid
267 * the page table lock when the pte is not present (helpful when reclaiming
268 * highly shared pages).
270 * On success returns with pte mapped and locked.
272 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
273 unsigned long address, spinlock_t **ptlp, int sync)
275 pgd_t *pgd;
276 pud_t *pud;
277 pmd_t *pmd;
278 pte_t *pte;
279 spinlock_t *ptl;
281 pgd = pgd_offset(mm, address);
282 if (!pgd_present(*pgd))
283 return NULL;
285 pud = pud_offset(pgd, address);
286 if (!pud_present(*pud))
287 return NULL;
289 pmd = pmd_offset(pud, address);
290 if (!pmd_present(*pmd))
291 return NULL;
293 pte = pte_offset_map(pmd, address);
294 /* Make a quick check before getting the lock */
295 if (!sync && !pte_present(*pte)) {
296 pte_unmap(pte);
297 return NULL;
300 ptl = pte_lockptr(mm, pmd);
301 spin_lock(ptl);
302 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
303 *ptlp = ptl;
304 return pte;
306 pte_unmap_unlock(pte, ptl);
307 return NULL;
311 * page_mapped_in_vma - check whether a page is really mapped in a VMA
312 * @page: the page to test
313 * @vma: the VMA to test
315 * Returns 1 if the page is mapped into the page tables of the VMA, 0
316 * if the page is not mapped into the page tables of this VMA. Only
317 * valid for normal file or anonymous VMAs.
319 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
321 unsigned long address;
322 pte_t *pte;
323 spinlock_t *ptl;
325 address = vma_address(page, vma);
326 if (address == -EFAULT) /* out of vma range */
327 return 0;
328 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
329 if (!pte) /* the page is not in this mm */
330 return 0;
331 pte_unmap_unlock(pte, ptl);
333 return 1;
337 * Subfunctions of page_referenced: page_referenced_one called
338 * repeatedly from either page_referenced_anon or page_referenced_file.
340 static int page_referenced_one(struct page *page,
341 struct vm_area_struct *vma,
342 unsigned int *mapcount,
343 unsigned long *vm_flags)
345 struct mm_struct *mm = vma->vm_mm;
346 unsigned long address;
347 pte_t *pte;
348 spinlock_t *ptl;
349 int referenced = 0;
351 address = vma_address(page, vma);
352 if (address == -EFAULT)
353 goto out;
355 pte = page_check_address(page, mm, address, &ptl, 0);
356 if (!pte)
357 goto out;
360 * Don't want to elevate referenced for mlocked page that gets this far,
361 * in order that it progresses to try_to_unmap and is moved to the
362 * unevictable list.
364 if (vma->vm_flags & VM_LOCKED) {
365 *mapcount = 1; /* break early from loop */
366 *vm_flags |= VM_LOCKED;
367 goto out_unmap;
370 if (ptep_clear_flush_young_notify(vma, address, pte)) {
372 * Don't treat a reference through a sequentially read
373 * mapping as such. If the page has been used in
374 * another mapping, we will catch it; if this other
375 * mapping is already gone, the unmap path will have
376 * set PG_referenced or activated the page.
378 if (likely(!VM_SequentialReadHint(vma)))
379 referenced++;
382 /* Pretend the page is referenced if the task has the
383 swap token and is in the middle of a page fault. */
384 if (mm != current->mm && has_swap_token(mm) &&
385 rwsem_is_locked(&mm->mmap_sem))
386 referenced++;
388 out_unmap:
389 (*mapcount)--;
390 pte_unmap_unlock(pte, ptl);
391 out:
392 if (referenced)
393 *vm_flags |= vma->vm_flags;
394 return referenced;
397 static int page_referenced_anon(struct page *page,
398 struct mem_cgroup *mem_cont,
399 unsigned long *vm_flags)
401 unsigned int mapcount;
402 struct anon_vma *anon_vma;
403 struct vm_area_struct *vma;
404 int referenced = 0;
406 anon_vma = page_lock_anon_vma(page);
407 if (!anon_vma)
408 return referenced;
410 mapcount = page_mapcount(page);
411 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
413 * If we are reclaiming on behalf of a cgroup, skip
414 * counting on behalf of references from different
415 * cgroups
417 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
418 continue;
419 referenced += page_referenced_one(page, vma,
420 &mapcount, vm_flags);
421 if (!mapcount)
422 break;
425 page_unlock_anon_vma(anon_vma);
426 return referenced;
430 * page_referenced_file - referenced check for object-based rmap
431 * @page: the page we're checking references on.
432 * @mem_cont: target memory controller
433 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
435 * For an object-based mapped page, find all the places it is mapped and
436 * check/clear the referenced flag. This is done by following the page->mapping
437 * pointer, then walking the chain of vmas it holds. It returns the number
438 * of references it found.
440 * This function is only called from page_referenced for object-based pages.
442 static int page_referenced_file(struct page *page,
443 struct mem_cgroup *mem_cont,
444 unsigned long *vm_flags)
446 unsigned int mapcount;
447 struct address_space *mapping = page->mapping;
448 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
449 struct vm_area_struct *vma;
450 struct prio_tree_iter iter;
451 int referenced = 0;
454 * The caller's checks on page->mapping and !PageAnon have made
455 * sure that this is a file page: the check for page->mapping
456 * excludes the case just before it gets set on an anon page.
458 BUG_ON(PageAnon(page));
461 * The page lock not only makes sure that page->mapping cannot
462 * suddenly be NULLified by truncation, it makes sure that the
463 * structure at mapping cannot be freed and reused yet,
464 * so we can safely take mapping->i_mmap_lock.
466 BUG_ON(!PageLocked(page));
468 spin_lock(&mapping->i_mmap_lock);
471 * i_mmap_lock does not stabilize mapcount at all, but mapcount
472 * is more likely to be accurate if we note it after spinning.
474 mapcount = page_mapcount(page);
476 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
478 * If we are reclaiming on behalf of a cgroup, skip
479 * counting on behalf of references from different
480 * cgroups
482 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
483 continue;
484 referenced += page_referenced_one(page, vma,
485 &mapcount, vm_flags);
486 if (!mapcount)
487 break;
490 spin_unlock(&mapping->i_mmap_lock);
491 return referenced;
495 * page_referenced - test if the page was referenced
496 * @page: the page to test
497 * @is_locked: caller holds lock on the page
498 * @mem_cont: target memory controller
499 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
501 * Quick test_and_clear_referenced for all mappings to a page,
502 * returns the number of ptes which referenced the page.
504 int page_referenced(struct page *page,
505 int is_locked,
506 struct mem_cgroup *mem_cont,
507 unsigned long *vm_flags)
509 int referenced = 0;
511 if (TestClearPageReferenced(page))
512 referenced++;
514 *vm_flags = 0;
515 if (page_mapped(page) && page->mapping) {
516 if (PageAnon(page))
517 referenced += page_referenced_anon(page, mem_cont,
518 vm_flags);
519 else if (is_locked)
520 referenced += page_referenced_file(page, mem_cont,
521 vm_flags);
522 else if (!trylock_page(page))
523 referenced++;
524 else {
525 if (page->mapping)
526 referenced += page_referenced_file(page,
527 mem_cont, vm_flags);
528 unlock_page(page);
532 if (page_test_and_clear_young(page))
533 referenced++;
535 return referenced;
538 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma)
540 struct mm_struct *mm = vma->vm_mm;
541 unsigned long address;
542 pte_t *pte;
543 spinlock_t *ptl;
544 int ret = 0;
546 address = vma_address(page, vma);
547 if (address == -EFAULT)
548 goto out;
550 pte = page_check_address(page, mm, address, &ptl, 1);
551 if (!pte)
552 goto out;
554 if (pte_dirty(*pte) || pte_write(*pte)) {
555 pte_t entry;
557 flush_cache_page(vma, address, pte_pfn(*pte));
558 entry = ptep_clear_flush_notify(vma, address, pte);
559 entry = pte_wrprotect(entry);
560 entry = pte_mkclean(entry);
561 set_pte_at(mm, address, pte, entry);
562 ret = 1;
565 pte_unmap_unlock(pte, ptl);
566 out:
567 return ret;
570 static int page_mkclean_file(struct address_space *mapping, struct page *page)
572 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
573 struct vm_area_struct *vma;
574 struct prio_tree_iter iter;
575 int ret = 0;
577 BUG_ON(PageAnon(page));
579 spin_lock(&mapping->i_mmap_lock);
580 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
581 if (vma->vm_flags & VM_SHARED)
582 ret += page_mkclean_one(page, vma);
584 spin_unlock(&mapping->i_mmap_lock);
585 return ret;
588 int page_mkclean(struct page *page)
590 int ret = 0;
592 BUG_ON(!PageLocked(page));
594 if (page_mapped(page)) {
595 struct address_space *mapping = page_mapping(page);
596 if (mapping) {
597 ret = page_mkclean_file(mapping, page);
598 if (page_test_dirty(page)) {
599 page_clear_dirty(page);
600 ret = 1;
605 return ret;
607 EXPORT_SYMBOL_GPL(page_mkclean);
610 * __page_set_anon_rmap - setup new anonymous rmap
611 * @page: the page to add the mapping to
612 * @vma: the vm area in which the mapping is added
613 * @address: the user virtual address mapped
615 static void __page_set_anon_rmap(struct page *page,
616 struct vm_area_struct *vma, unsigned long address)
618 struct anon_vma *anon_vma = vma->anon_vma;
620 BUG_ON(!anon_vma);
621 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
622 page->mapping = (struct address_space *) anon_vma;
624 page->index = linear_page_index(vma, address);
627 * nr_mapped state can be updated without turning off
628 * interrupts because it is not modified via interrupt.
630 __inc_zone_page_state(page, NR_ANON_PAGES);
634 * __page_check_anon_rmap - sanity check anonymous rmap addition
635 * @page: the page to add the mapping to
636 * @vma: the vm area in which the mapping is added
637 * @address: the user virtual address mapped
639 static void __page_check_anon_rmap(struct page *page,
640 struct vm_area_struct *vma, unsigned long address)
642 #ifdef CONFIG_DEBUG_VM
644 * The page's anon-rmap details (mapping and index) are guaranteed to
645 * be set up correctly at this point.
647 * We have exclusion against page_add_anon_rmap because the caller
648 * always holds the page locked, except if called from page_dup_rmap,
649 * in which case the page is already known to be setup.
651 * We have exclusion against page_add_new_anon_rmap because those pages
652 * are initially only visible via the pagetables, and the pte is locked
653 * over the call to page_add_new_anon_rmap.
655 struct anon_vma *anon_vma = vma->anon_vma;
656 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
657 BUG_ON(page->mapping != (struct address_space *)anon_vma);
658 BUG_ON(page->index != linear_page_index(vma, address));
659 #endif
663 * page_add_anon_rmap - add pte mapping to an anonymous page
664 * @page: the page to add the mapping to
665 * @vma: the vm area in which the mapping is added
666 * @address: the user virtual address mapped
668 * The caller needs to hold the pte lock and the page must be locked.
670 void page_add_anon_rmap(struct page *page,
671 struct vm_area_struct *vma, unsigned long address)
673 VM_BUG_ON(!PageLocked(page));
674 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
675 if (atomic_inc_and_test(&page->_mapcount))
676 __page_set_anon_rmap(page, vma, address);
677 else
678 __page_check_anon_rmap(page, vma, address);
682 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
683 * @page: the page to add the mapping to
684 * @vma: the vm area in which the mapping is added
685 * @address: the user virtual address mapped
687 * Same as page_add_anon_rmap but must only be called on *new* pages.
688 * This means the inc-and-test can be bypassed.
689 * Page does not have to be locked.
691 void page_add_new_anon_rmap(struct page *page,
692 struct vm_area_struct *vma, unsigned long address)
694 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
695 SetPageSwapBacked(page);
696 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
697 __page_set_anon_rmap(page, vma, address);
698 if (page_evictable(page, vma))
699 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
700 else
701 add_page_to_unevictable_list(page);
705 * page_add_file_rmap - add pte mapping to a file page
706 * @page: the page to add the mapping to
708 * The caller needs to hold the pte lock.
710 void page_add_file_rmap(struct page *page)
712 if (atomic_inc_and_test(&page->_mapcount)) {
713 __inc_zone_page_state(page, NR_FILE_MAPPED);
714 mem_cgroup_update_mapped_file_stat(page, 1);
719 * page_remove_rmap - take down pte mapping from a page
720 * @page: page to remove mapping from
722 * The caller needs to hold the pte lock.
724 void page_remove_rmap(struct page *page)
726 /* page still mapped by someone else? */
727 if (!atomic_add_negative(-1, &page->_mapcount))
728 return;
731 * Now that the last pte has gone, s390 must transfer dirty
732 * flag from storage key to struct page. We can usually skip
733 * this if the page is anon, so about to be freed; but perhaps
734 * not if it's in swapcache - there might be another pte slot
735 * containing the swap entry, but page not yet written to swap.
737 if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
738 page_clear_dirty(page);
739 set_page_dirty(page);
741 if (PageAnon(page)) {
742 mem_cgroup_uncharge_page(page);
743 __dec_zone_page_state(page, NR_ANON_PAGES);
744 } else {
745 __dec_zone_page_state(page, NR_FILE_MAPPED);
747 mem_cgroup_update_mapped_file_stat(page, -1);
749 * It would be tidy to reset the PageAnon mapping here,
750 * but that might overwrite a racing page_add_anon_rmap
751 * which increments mapcount after us but sets mapping
752 * before us: so leave the reset to free_hot_cold_page,
753 * and remember that it's only reliable while mapped.
754 * Leaving it set also helps swapoff to reinstate ptes
755 * faster for those pages still in swapcache.
760 * Subfunctions of try_to_unmap: try_to_unmap_one called
761 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
763 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
764 enum ttu_flags flags)
766 struct mm_struct *mm = vma->vm_mm;
767 unsigned long address;
768 pte_t *pte;
769 pte_t pteval;
770 spinlock_t *ptl;
771 int ret = SWAP_AGAIN;
773 address = vma_address(page, vma);
774 if (address == -EFAULT)
775 goto out;
777 pte = page_check_address(page, mm, address, &ptl, 0);
778 if (!pte)
779 goto out;
782 * If the page is mlock()d, we cannot swap it out.
783 * If it's recently referenced (perhaps page_referenced
784 * skipped over this mm) then we should reactivate it.
786 if (!(flags & TTU_IGNORE_MLOCK)) {
787 if (vma->vm_flags & VM_LOCKED) {
788 ret = SWAP_MLOCK;
789 goto out_unmap;
792 if (!(flags & TTU_IGNORE_ACCESS)) {
793 if (ptep_clear_flush_young_notify(vma, address, pte)) {
794 ret = SWAP_FAIL;
795 goto out_unmap;
799 /* Nuke the page table entry. */
800 flush_cache_page(vma, address, page_to_pfn(page));
801 pteval = ptep_clear_flush_notify(vma, address, pte);
803 /* Move the dirty bit to the physical page now the pte is gone. */
804 if (pte_dirty(pteval))
805 set_page_dirty(page);
807 /* Update high watermark before we lower rss */
808 update_hiwater_rss(mm);
810 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
811 if (PageAnon(page))
812 dec_mm_counter(mm, anon_rss);
813 else
814 dec_mm_counter(mm, file_rss);
815 set_pte_at(mm, address, pte,
816 swp_entry_to_pte(make_hwpoison_entry(page)));
817 } else if (PageAnon(page)) {
818 swp_entry_t entry = { .val = page_private(page) };
820 if (PageSwapCache(page)) {
822 * Store the swap location in the pte.
823 * See handle_pte_fault() ...
825 swap_duplicate(entry);
826 if (list_empty(&mm->mmlist)) {
827 spin_lock(&mmlist_lock);
828 if (list_empty(&mm->mmlist))
829 list_add(&mm->mmlist, &init_mm.mmlist);
830 spin_unlock(&mmlist_lock);
832 dec_mm_counter(mm, anon_rss);
833 } else if (PAGE_MIGRATION) {
835 * Store the pfn of the page in a special migration
836 * pte. do_swap_page() will wait until the migration
837 * pte is removed and then restart fault handling.
839 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
840 entry = make_migration_entry(page, pte_write(pteval));
842 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
843 BUG_ON(pte_file(*pte));
844 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
845 /* Establish migration entry for a file page */
846 swp_entry_t entry;
847 entry = make_migration_entry(page, pte_write(pteval));
848 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
849 } else
850 dec_mm_counter(mm, file_rss);
853 page_remove_rmap(page);
854 page_cache_release(page);
856 out_unmap:
857 pte_unmap_unlock(pte, ptl);
858 out:
859 return ret;
863 * objrmap doesn't work for nonlinear VMAs because the assumption that
864 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
865 * Consequently, given a particular page and its ->index, we cannot locate the
866 * ptes which are mapping that page without an exhaustive linear search.
868 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
869 * maps the file to which the target page belongs. The ->vm_private_data field
870 * holds the current cursor into that scan. Successive searches will circulate
871 * around the vma's virtual address space.
873 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
874 * more scanning pressure is placed against them as well. Eventually pages
875 * will become fully unmapped and are eligible for eviction.
877 * For very sparsely populated VMAs this is a little inefficient - chances are
878 * there there won't be many ptes located within the scan cluster. In this case
879 * maybe we could scan further - to the end of the pte page, perhaps.
881 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
882 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
883 * rather than unmapping them. If we encounter the "check_page" that vmscan is
884 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
886 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
887 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
889 static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
890 struct vm_area_struct *vma, struct page *check_page)
892 struct mm_struct *mm = vma->vm_mm;
893 pgd_t *pgd;
894 pud_t *pud;
895 pmd_t *pmd;
896 pte_t *pte;
897 pte_t pteval;
898 spinlock_t *ptl;
899 struct page *page;
900 unsigned long address;
901 unsigned long end;
902 int ret = SWAP_AGAIN;
903 int locked_vma = 0;
905 address = (vma->vm_start + cursor) & CLUSTER_MASK;
906 end = address + CLUSTER_SIZE;
907 if (address < vma->vm_start)
908 address = vma->vm_start;
909 if (end > vma->vm_end)
910 end = vma->vm_end;
912 pgd = pgd_offset(mm, address);
913 if (!pgd_present(*pgd))
914 return ret;
916 pud = pud_offset(pgd, address);
917 if (!pud_present(*pud))
918 return ret;
920 pmd = pmd_offset(pud, address);
921 if (!pmd_present(*pmd))
922 return ret;
925 * MLOCK_PAGES => feature is configured.
926 * if we can acquire the mmap_sem for read, and vma is VM_LOCKED,
927 * keep the sem while scanning the cluster for mlocking pages.
929 if (MLOCK_PAGES && down_read_trylock(&vma->vm_mm->mmap_sem)) {
930 locked_vma = (vma->vm_flags & VM_LOCKED);
931 if (!locked_vma)
932 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
935 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
937 /* Update high watermark before we lower rss */
938 update_hiwater_rss(mm);
940 for (; address < end; pte++, address += PAGE_SIZE) {
941 if (!pte_present(*pte))
942 continue;
943 page = vm_normal_page(vma, address, *pte);
944 BUG_ON(!page || PageAnon(page));
946 if (locked_vma) {
947 mlock_vma_page(page); /* no-op if already mlocked */
948 if (page == check_page)
949 ret = SWAP_MLOCK;
950 continue; /* don't unmap */
953 if (ptep_clear_flush_young_notify(vma, address, pte))
954 continue;
956 /* Nuke the page table entry. */
957 flush_cache_page(vma, address, pte_pfn(*pte));
958 pteval = ptep_clear_flush_notify(vma, address, pte);
960 /* If nonlinear, store the file page offset in the pte. */
961 if (page->index != linear_page_index(vma, address))
962 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
964 /* Move the dirty bit to the physical page now the pte is gone. */
965 if (pte_dirty(pteval))
966 set_page_dirty(page);
968 page_remove_rmap(page);
969 page_cache_release(page);
970 dec_mm_counter(mm, file_rss);
971 (*mapcount)--;
973 pte_unmap_unlock(pte - 1, ptl);
974 if (locked_vma)
975 up_read(&vma->vm_mm->mmap_sem);
976 return ret;
980 * common handling for pages mapped in VM_LOCKED vmas
982 static int try_to_mlock_page(struct page *page, struct vm_area_struct *vma)
984 int mlocked = 0;
986 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
987 if (vma->vm_flags & VM_LOCKED) {
988 mlock_vma_page(page);
989 mlocked++; /* really mlocked the page */
991 up_read(&vma->vm_mm->mmap_sem);
993 return mlocked;
997 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
998 * rmap method
999 * @page: the page to unmap/unlock
1000 * @unlock: request for unlock rather than unmap [unlikely]
1001 * @migration: unmapping for migration - ignored if @unlock
1003 * Find all the mappings of a page using the mapping pointer and the vma chains
1004 * contained in the anon_vma struct it points to.
1006 * This function is only called from try_to_unmap/try_to_munlock for
1007 * anonymous pages.
1008 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1009 * where the page was found will be held for write. So, we won't recheck
1010 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1011 * 'LOCKED.
1013 static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1015 struct anon_vma *anon_vma;
1016 struct vm_area_struct *vma;
1017 unsigned int mlocked = 0;
1018 int ret = SWAP_AGAIN;
1019 int unlock = TTU_ACTION(flags) == TTU_MUNLOCK;
1021 if (MLOCK_PAGES && unlikely(unlock))
1022 ret = SWAP_SUCCESS; /* default for try_to_munlock() */
1024 anon_vma = page_lock_anon_vma(page);
1025 if (!anon_vma)
1026 return ret;
1028 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
1029 if (MLOCK_PAGES && unlikely(unlock)) {
1030 if (!((vma->vm_flags & VM_LOCKED) &&
1031 page_mapped_in_vma(page, vma)))
1032 continue; /* must visit all unlocked vmas */
1033 ret = SWAP_MLOCK; /* saw at least one mlocked vma */
1034 } else {
1035 ret = try_to_unmap_one(page, vma, flags);
1036 if (ret == SWAP_FAIL || !page_mapped(page))
1037 break;
1039 if (ret == SWAP_MLOCK) {
1040 mlocked = try_to_mlock_page(page, vma);
1041 if (mlocked)
1042 break; /* stop if actually mlocked page */
1046 page_unlock_anon_vma(anon_vma);
1048 if (mlocked)
1049 ret = SWAP_MLOCK; /* actually mlocked the page */
1050 else if (ret == SWAP_MLOCK)
1051 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */
1053 return ret;
1057 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1058 * @page: the page to unmap/unlock
1059 * @flags: action and flags
1061 * Find all the mappings of a page using the mapping pointer and the vma chains
1062 * contained in the address_space struct it points to.
1064 * This function is only called from try_to_unmap/try_to_munlock for
1065 * object-based pages.
1066 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1067 * where the page was found will be held for write. So, we won't recheck
1068 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1069 * 'LOCKED.
1071 static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1073 struct address_space *mapping = page->mapping;
1074 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1075 struct vm_area_struct *vma;
1076 struct prio_tree_iter iter;
1077 int ret = SWAP_AGAIN;
1078 unsigned long cursor;
1079 unsigned long max_nl_cursor = 0;
1080 unsigned long max_nl_size = 0;
1081 unsigned int mapcount;
1082 unsigned int mlocked = 0;
1083 int unlock = TTU_ACTION(flags) == TTU_MUNLOCK;
1085 if (MLOCK_PAGES && unlikely(unlock))
1086 ret = SWAP_SUCCESS; /* default for try_to_munlock() */
1088 spin_lock(&mapping->i_mmap_lock);
1089 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1090 if (MLOCK_PAGES && unlikely(unlock)) {
1091 if (!((vma->vm_flags & VM_LOCKED) &&
1092 page_mapped_in_vma(page, vma)))
1093 continue; /* must visit all vmas */
1094 ret = SWAP_MLOCK;
1095 } else {
1096 ret = try_to_unmap_one(page, vma, flags);
1097 if (ret == SWAP_FAIL || !page_mapped(page))
1098 goto out;
1100 if (ret == SWAP_MLOCK) {
1101 mlocked = try_to_mlock_page(page, vma);
1102 if (mlocked)
1103 break; /* stop if actually mlocked page */
1107 if (mlocked)
1108 goto out;
1110 if (list_empty(&mapping->i_mmap_nonlinear))
1111 goto out;
1113 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1114 shared.vm_set.list) {
1115 if (MLOCK_PAGES && unlikely(unlock)) {
1116 if (!(vma->vm_flags & VM_LOCKED))
1117 continue; /* must visit all vmas */
1118 ret = SWAP_MLOCK; /* leave mlocked == 0 */
1119 goto out; /* no need to look further */
1121 if (!MLOCK_PAGES && !(flags & TTU_IGNORE_MLOCK) &&
1122 (vma->vm_flags & VM_LOCKED))
1123 continue;
1124 cursor = (unsigned long) vma->vm_private_data;
1125 if (cursor > max_nl_cursor)
1126 max_nl_cursor = cursor;
1127 cursor = vma->vm_end - vma->vm_start;
1128 if (cursor > max_nl_size)
1129 max_nl_size = cursor;
1132 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1133 ret = SWAP_FAIL;
1134 goto out;
1138 * We don't try to search for this page in the nonlinear vmas,
1139 * and page_referenced wouldn't have found it anyway. Instead
1140 * just walk the nonlinear vmas trying to age and unmap some.
1141 * The mapcount of the page we came in with is irrelevant,
1142 * but even so use it as a guide to how hard we should try?
1144 mapcount = page_mapcount(page);
1145 if (!mapcount)
1146 goto out;
1147 cond_resched_lock(&mapping->i_mmap_lock);
1149 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1150 if (max_nl_cursor == 0)
1151 max_nl_cursor = CLUSTER_SIZE;
1153 do {
1154 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1155 shared.vm_set.list) {
1156 if (!MLOCK_PAGES && !(flags & TTU_IGNORE_MLOCK) &&
1157 (vma->vm_flags & VM_LOCKED))
1158 continue;
1159 cursor = (unsigned long) vma->vm_private_data;
1160 while ( cursor < max_nl_cursor &&
1161 cursor < vma->vm_end - vma->vm_start) {
1162 ret = try_to_unmap_cluster(cursor, &mapcount,
1163 vma, page);
1164 if (ret == SWAP_MLOCK)
1165 mlocked = 2; /* to return below */
1166 cursor += CLUSTER_SIZE;
1167 vma->vm_private_data = (void *) cursor;
1168 if ((int)mapcount <= 0)
1169 goto out;
1171 vma->vm_private_data = (void *) max_nl_cursor;
1173 cond_resched_lock(&mapping->i_mmap_lock);
1174 max_nl_cursor += CLUSTER_SIZE;
1175 } while (max_nl_cursor <= max_nl_size);
1178 * Don't loop forever (perhaps all the remaining pages are
1179 * in locked vmas). Reset cursor on all unreserved nonlinear
1180 * vmas, now forgetting on which ones it had fallen behind.
1182 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1183 vma->vm_private_data = NULL;
1184 out:
1185 spin_unlock(&mapping->i_mmap_lock);
1186 if (mlocked)
1187 ret = SWAP_MLOCK; /* actually mlocked the page */
1188 else if (ret == SWAP_MLOCK)
1189 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */
1190 return ret;
1194 * try_to_unmap - try to remove all page table mappings to a page
1195 * @page: the page to get unmapped
1196 * @flags: action and flags
1198 * Tries to remove all the page table entries which are mapping this
1199 * page, used in the pageout path. Caller must hold the page lock.
1200 * Return values are:
1202 * SWAP_SUCCESS - we succeeded in removing all mappings
1203 * SWAP_AGAIN - we missed a mapping, try again later
1204 * SWAP_FAIL - the page is unswappable
1205 * SWAP_MLOCK - page is mlocked.
1207 int try_to_unmap(struct page *page, enum ttu_flags flags)
1209 int ret;
1211 BUG_ON(!PageLocked(page));
1213 if (PageAnon(page))
1214 ret = try_to_unmap_anon(page, flags);
1215 else
1216 ret = try_to_unmap_file(page, flags);
1217 if (ret != SWAP_MLOCK && !page_mapped(page))
1218 ret = SWAP_SUCCESS;
1219 return ret;
1223 * try_to_munlock - try to munlock a page
1224 * @page: the page to be munlocked
1226 * Called from munlock code. Checks all of the VMAs mapping the page
1227 * to make sure nobody else has this page mlocked. The page will be
1228 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1230 * Return values are:
1232 * SWAP_SUCCESS - no vma's holding page mlocked.
1233 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1234 * SWAP_MLOCK - page is now mlocked.
1236 int try_to_munlock(struct page *page)
1238 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1240 if (PageAnon(page))
1241 return try_to_unmap_anon(page, TTU_MUNLOCK);
1242 else
1243 return try_to_unmap_file(page, TTU_MUNLOCK);