sfc: Improve reliability of RX queue flushing
[linux-2.6/libata-dev.git] / mm / rmap.c
blob836c6c63e1f2dbee16a922f7195da2f3ac237d83
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)
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/memcontrol.h>
51 #include <linux/mmu_notifier.h>
52 #include <linux/migrate.h>
54 #include <asm/tlbflush.h>
56 #include "internal.h"
58 static struct kmem_cache *anon_vma_cachep;
60 static inline struct anon_vma *anon_vma_alloc(void)
62 return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
65 static inline void anon_vma_free(struct anon_vma *anon_vma)
67 kmem_cache_free(anon_vma_cachep, anon_vma);
70 /**
71 * anon_vma_prepare - attach an anon_vma to a memory region
72 * @vma: the memory region in question
74 * This makes sure the memory mapping described by 'vma' has
75 * an 'anon_vma' attached to it, so that we can associate the
76 * anonymous pages mapped into it with that anon_vma.
78 * The common case will be that we already have one, but if
79 * if not we either need to find an adjacent mapping that we
80 * can re-use the anon_vma from (very common when the only
81 * reason for splitting a vma has been mprotect()), or we
82 * allocate a new one.
84 * Anon-vma allocations are very subtle, because we may have
85 * optimistically looked up an anon_vma in page_lock_anon_vma()
86 * and that may actually touch the spinlock even in the newly
87 * allocated vma (it depends on RCU to make sure that the
88 * anon_vma isn't actually destroyed).
90 * As a result, we need to do proper anon_vma locking even
91 * for the new allocation. At the same time, we do not want
92 * to do any locking for the common case of already having
93 * an anon_vma.
95 * This must be called with the mmap_sem held for reading.
97 int anon_vma_prepare(struct vm_area_struct *vma)
99 struct anon_vma *anon_vma = vma->anon_vma;
101 might_sleep();
102 if (unlikely(!anon_vma)) {
103 struct mm_struct *mm = vma->vm_mm;
104 struct anon_vma *allocated;
106 anon_vma = find_mergeable_anon_vma(vma);
107 allocated = NULL;
108 if (!anon_vma) {
109 anon_vma = anon_vma_alloc();
110 if (unlikely(!anon_vma))
111 return -ENOMEM;
112 allocated = anon_vma;
114 spin_lock(&anon_vma->lock);
116 /* page_table_lock to protect against threads */
117 spin_lock(&mm->page_table_lock);
118 if (likely(!vma->anon_vma)) {
119 vma->anon_vma = anon_vma;
120 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
121 allocated = NULL;
123 spin_unlock(&mm->page_table_lock);
125 spin_unlock(&anon_vma->lock);
126 if (unlikely(allocated))
127 anon_vma_free(allocated);
129 return 0;
132 void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
134 BUG_ON(vma->anon_vma != next->anon_vma);
135 list_del(&next->anon_vma_node);
138 void __anon_vma_link(struct vm_area_struct *vma)
140 struct anon_vma *anon_vma = vma->anon_vma;
142 if (anon_vma)
143 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
146 void anon_vma_link(struct vm_area_struct *vma)
148 struct anon_vma *anon_vma = vma->anon_vma;
150 if (anon_vma) {
151 spin_lock(&anon_vma->lock);
152 list_add_tail(&vma->anon_vma_node, &anon_vma->head);
153 spin_unlock(&anon_vma->lock);
157 void anon_vma_unlink(struct vm_area_struct *vma)
159 struct anon_vma *anon_vma = vma->anon_vma;
160 int empty;
162 if (!anon_vma)
163 return;
165 spin_lock(&anon_vma->lock);
166 list_del(&vma->anon_vma_node);
168 /* We must garbage collect the anon_vma if it's empty */
169 empty = list_empty(&anon_vma->head);
170 spin_unlock(&anon_vma->lock);
172 if (empty)
173 anon_vma_free(anon_vma);
176 static void anon_vma_ctor(void *data)
178 struct anon_vma *anon_vma = data;
180 spin_lock_init(&anon_vma->lock);
181 INIT_LIST_HEAD(&anon_vma->head);
184 void __init anon_vma_init(void)
186 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
187 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
191 * Getting a lock on a stable anon_vma from a page off the LRU is
192 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
194 static struct anon_vma *page_lock_anon_vma(struct page *page)
196 struct anon_vma *anon_vma;
197 unsigned long anon_mapping;
199 rcu_read_lock();
200 anon_mapping = (unsigned long) page->mapping;
201 if (!(anon_mapping & PAGE_MAPPING_ANON))
202 goto out;
203 if (!page_mapped(page))
204 goto out;
206 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
207 spin_lock(&anon_vma->lock);
208 return anon_vma;
209 out:
210 rcu_read_unlock();
211 return NULL;
214 static void page_unlock_anon_vma(struct anon_vma *anon_vma)
216 spin_unlock(&anon_vma->lock);
217 rcu_read_unlock();
221 * At what user virtual address is page expected in @vma?
222 * Returns virtual address or -EFAULT if page's index/offset is not
223 * within the range mapped the @vma.
225 static inline unsigned long
226 vma_address(struct page *page, struct vm_area_struct *vma)
228 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
229 unsigned long address;
231 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
232 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
233 /* page should be within @vma mapping range */
234 return -EFAULT;
236 return address;
240 * At what user virtual address is page expected in vma? checking that the
241 * page matches the vma: currently only used on anon pages, by unuse_vma;
243 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
245 if (PageAnon(page)) {
246 if ((void *)vma->anon_vma !=
247 (void *)page->mapping - PAGE_MAPPING_ANON)
248 return -EFAULT;
249 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
250 if (!vma->vm_file ||
251 vma->vm_file->f_mapping != page->mapping)
252 return -EFAULT;
253 } else
254 return -EFAULT;
255 return vma_address(page, vma);
259 * Check that @page is mapped at @address into @mm.
261 * If @sync is false, page_check_address may perform a racy check to avoid
262 * the page table lock when the pte is not present (helpful when reclaiming
263 * highly shared pages).
265 * On success returns with pte mapped and locked.
267 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
268 unsigned long address, spinlock_t **ptlp, int sync)
270 pgd_t *pgd;
271 pud_t *pud;
272 pmd_t *pmd;
273 pte_t *pte;
274 spinlock_t *ptl;
276 pgd = pgd_offset(mm, address);
277 if (!pgd_present(*pgd))
278 return NULL;
280 pud = pud_offset(pgd, address);
281 if (!pud_present(*pud))
282 return NULL;
284 pmd = pmd_offset(pud, address);
285 if (!pmd_present(*pmd))
286 return NULL;
288 pte = pte_offset_map(pmd, address);
289 /* Make a quick check before getting the lock */
290 if (!sync && !pte_present(*pte)) {
291 pte_unmap(pte);
292 return NULL;
295 ptl = pte_lockptr(mm, pmd);
296 spin_lock(ptl);
297 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
298 *ptlp = ptl;
299 return pte;
301 pte_unmap_unlock(pte, ptl);
302 return NULL;
306 * page_mapped_in_vma - check whether a page is really mapped in a VMA
307 * @page: the page to test
308 * @vma: the VMA to test
310 * Returns 1 if the page is mapped into the page tables of the VMA, 0
311 * if the page is not mapped into the page tables of this VMA. Only
312 * valid for normal file or anonymous VMAs.
314 static int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
316 unsigned long address;
317 pte_t *pte;
318 spinlock_t *ptl;
320 address = vma_address(page, vma);
321 if (address == -EFAULT) /* out of vma range */
322 return 0;
323 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
324 if (!pte) /* the page is not in this mm */
325 return 0;
326 pte_unmap_unlock(pte, ptl);
328 return 1;
332 * Subfunctions of page_referenced: page_referenced_one called
333 * repeatedly from either page_referenced_anon or page_referenced_file.
335 static int page_referenced_one(struct page *page,
336 struct vm_area_struct *vma,
337 unsigned int *mapcount,
338 unsigned long *vm_flags)
340 struct mm_struct *mm = vma->vm_mm;
341 unsigned long address;
342 pte_t *pte;
343 spinlock_t *ptl;
344 int referenced = 0;
346 address = vma_address(page, vma);
347 if (address == -EFAULT)
348 goto out;
350 pte = page_check_address(page, mm, address, &ptl, 0);
351 if (!pte)
352 goto out;
355 * Don't want to elevate referenced for mlocked page that gets this far,
356 * in order that it progresses to try_to_unmap and is moved to the
357 * unevictable list.
359 if (vma->vm_flags & VM_LOCKED) {
360 *mapcount = 1; /* break early from loop */
361 goto out_unmap;
364 if (ptep_clear_flush_young_notify(vma, address, pte)) {
366 * Don't treat a reference through a sequentially read
367 * mapping as such. If the page has been used in
368 * another mapping, we will catch it; if this other
369 * mapping is already gone, the unmap path will have
370 * set PG_referenced or activated the page.
372 if (likely(!VM_SequentialReadHint(vma)))
373 referenced++;
376 /* Pretend the page is referenced if the task has the
377 swap token and is in the middle of a page fault. */
378 if (mm != current->mm && has_swap_token(mm) &&
379 rwsem_is_locked(&mm->mmap_sem))
380 referenced++;
382 out_unmap:
383 (*mapcount)--;
384 pte_unmap_unlock(pte, ptl);
385 out:
386 if (referenced)
387 *vm_flags |= vma->vm_flags;
388 return referenced;
391 static int page_referenced_anon(struct page *page,
392 struct mem_cgroup *mem_cont,
393 unsigned long *vm_flags)
395 unsigned int mapcount;
396 struct anon_vma *anon_vma;
397 struct vm_area_struct *vma;
398 int referenced = 0;
400 anon_vma = page_lock_anon_vma(page);
401 if (!anon_vma)
402 return referenced;
404 mapcount = page_mapcount(page);
405 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
407 * If we are reclaiming on behalf of a cgroup, skip
408 * counting on behalf of references from different
409 * cgroups
411 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
412 continue;
413 referenced += page_referenced_one(page, vma,
414 &mapcount, vm_flags);
415 if (!mapcount)
416 break;
419 page_unlock_anon_vma(anon_vma);
420 return referenced;
424 * page_referenced_file - referenced check for object-based rmap
425 * @page: the page we're checking references on.
426 * @mem_cont: target memory controller
427 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
429 * For an object-based mapped page, find all the places it is mapped and
430 * check/clear the referenced flag. This is done by following the page->mapping
431 * pointer, then walking the chain of vmas it holds. It returns the number
432 * of references it found.
434 * This function is only called from page_referenced for object-based pages.
436 static int page_referenced_file(struct page *page,
437 struct mem_cgroup *mem_cont,
438 unsigned long *vm_flags)
440 unsigned int mapcount;
441 struct address_space *mapping = page->mapping;
442 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
443 struct vm_area_struct *vma;
444 struct prio_tree_iter iter;
445 int referenced = 0;
448 * The caller's checks on page->mapping and !PageAnon have made
449 * sure that this is a file page: the check for page->mapping
450 * excludes the case just before it gets set on an anon page.
452 BUG_ON(PageAnon(page));
455 * The page lock not only makes sure that page->mapping cannot
456 * suddenly be NULLified by truncation, it makes sure that the
457 * structure at mapping cannot be freed and reused yet,
458 * so we can safely take mapping->i_mmap_lock.
460 BUG_ON(!PageLocked(page));
462 spin_lock(&mapping->i_mmap_lock);
465 * i_mmap_lock does not stabilize mapcount at all, but mapcount
466 * is more likely to be accurate if we note it after spinning.
468 mapcount = page_mapcount(page);
470 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
472 * If we are reclaiming on behalf of a cgroup, skip
473 * counting on behalf of references from different
474 * cgroups
476 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
477 continue;
478 referenced += page_referenced_one(page, vma,
479 &mapcount, vm_flags);
480 if (!mapcount)
481 break;
484 spin_unlock(&mapping->i_mmap_lock);
485 return referenced;
489 * page_referenced - test if the page was referenced
490 * @page: the page to test
491 * @is_locked: caller holds lock on the page
492 * @mem_cont: target memory controller
493 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
495 * Quick test_and_clear_referenced for all mappings to a page,
496 * returns the number of ptes which referenced the page.
498 int page_referenced(struct page *page,
499 int is_locked,
500 struct mem_cgroup *mem_cont,
501 unsigned long *vm_flags)
503 int referenced = 0;
505 if (TestClearPageReferenced(page))
506 referenced++;
508 *vm_flags = 0;
509 if (page_mapped(page) && page->mapping) {
510 if (PageAnon(page))
511 referenced += page_referenced_anon(page, mem_cont,
512 vm_flags);
513 else if (is_locked)
514 referenced += page_referenced_file(page, mem_cont,
515 vm_flags);
516 else if (!trylock_page(page))
517 referenced++;
518 else {
519 if (page->mapping)
520 referenced += page_referenced_file(page,
521 mem_cont, vm_flags);
522 unlock_page(page);
526 if (page_test_and_clear_young(page))
527 referenced++;
529 return referenced;
532 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma)
534 struct mm_struct *mm = vma->vm_mm;
535 unsigned long address;
536 pte_t *pte;
537 spinlock_t *ptl;
538 int ret = 0;
540 address = vma_address(page, vma);
541 if (address == -EFAULT)
542 goto out;
544 pte = page_check_address(page, mm, address, &ptl, 1);
545 if (!pte)
546 goto out;
548 if (pte_dirty(*pte) || pte_write(*pte)) {
549 pte_t entry;
551 flush_cache_page(vma, address, pte_pfn(*pte));
552 entry = ptep_clear_flush_notify(vma, address, pte);
553 entry = pte_wrprotect(entry);
554 entry = pte_mkclean(entry);
555 set_pte_at(mm, address, pte, entry);
556 ret = 1;
559 pte_unmap_unlock(pte, ptl);
560 out:
561 return ret;
564 static int page_mkclean_file(struct address_space *mapping, struct page *page)
566 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
567 struct vm_area_struct *vma;
568 struct prio_tree_iter iter;
569 int ret = 0;
571 BUG_ON(PageAnon(page));
573 spin_lock(&mapping->i_mmap_lock);
574 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
575 if (vma->vm_flags & VM_SHARED)
576 ret += page_mkclean_one(page, vma);
578 spin_unlock(&mapping->i_mmap_lock);
579 return ret;
582 int page_mkclean(struct page *page)
584 int ret = 0;
586 BUG_ON(!PageLocked(page));
588 if (page_mapped(page)) {
589 struct address_space *mapping = page_mapping(page);
590 if (mapping) {
591 ret = page_mkclean_file(mapping, page);
592 if (page_test_dirty(page)) {
593 page_clear_dirty(page);
594 ret = 1;
599 return ret;
601 EXPORT_SYMBOL_GPL(page_mkclean);
604 * __page_set_anon_rmap - setup new anonymous rmap
605 * @page: the page to add the mapping to
606 * @vma: the vm area in which the mapping is added
607 * @address: the user virtual address mapped
609 static void __page_set_anon_rmap(struct page *page,
610 struct vm_area_struct *vma, unsigned long address)
612 struct anon_vma *anon_vma = vma->anon_vma;
614 BUG_ON(!anon_vma);
615 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
616 page->mapping = (struct address_space *) anon_vma;
618 page->index = linear_page_index(vma, address);
621 * nr_mapped state can be updated without turning off
622 * interrupts because it is not modified via interrupt.
624 __inc_zone_page_state(page, NR_ANON_PAGES);
628 * __page_check_anon_rmap - sanity check anonymous rmap addition
629 * @page: the page to add the mapping to
630 * @vma: the vm area in which the mapping is added
631 * @address: the user virtual address mapped
633 static void __page_check_anon_rmap(struct page *page,
634 struct vm_area_struct *vma, unsigned long address)
636 #ifdef CONFIG_DEBUG_VM
638 * The page's anon-rmap details (mapping and index) are guaranteed to
639 * be set up correctly at this point.
641 * We have exclusion against page_add_anon_rmap because the caller
642 * always holds the page locked, except if called from page_dup_rmap,
643 * in which case the page is already known to be setup.
645 * We have exclusion against page_add_new_anon_rmap because those pages
646 * are initially only visible via the pagetables, and the pte is locked
647 * over the call to page_add_new_anon_rmap.
649 struct anon_vma *anon_vma = vma->anon_vma;
650 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
651 BUG_ON(page->mapping != (struct address_space *)anon_vma);
652 BUG_ON(page->index != linear_page_index(vma, address));
653 #endif
657 * page_add_anon_rmap - add pte mapping to an anonymous page
658 * @page: the page to add the mapping to
659 * @vma: the vm area in which the mapping is added
660 * @address: the user virtual address mapped
662 * The caller needs to hold the pte lock and the page must be locked.
664 void page_add_anon_rmap(struct page *page,
665 struct vm_area_struct *vma, unsigned long address)
667 VM_BUG_ON(!PageLocked(page));
668 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
669 if (atomic_inc_and_test(&page->_mapcount))
670 __page_set_anon_rmap(page, vma, address);
671 else
672 __page_check_anon_rmap(page, vma, address);
676 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
677 * @page: the page to add the mapping to
678 * @vma: the vm area in which the mapping is added
679 * @address: the user virtual address mapped
681 * Same as page_add_anon_rmap but must only be called on *new* pages.
682 * This means the inc-and-test can be bypassed.
683 * Page does not have to be locked.
685 void page_add_new_anon_rmap(struct page *page,
686 struct vm_area_struct *vma, unsigned long address)
688 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
689 SetPageSwapBacked(page);
690 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
691 __page_set_anon_rmap(page, vma, address);
692 if (page_evictable(page, vma))
693 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
694 else
695 add_page_to_unevictable_list(page);
699 * page_add_file_rmap - add pte mapping to a file page
700 * @page: the page to add the mapping to
702 * The caller needs to hold the pte lock.
704 void page_add_file_rmap(struct page *page)
706 if (atomic_inc_and_test(&page->_mapcount)) {
707 __inc_zone_page_state(page, NR_FILE_MAPPED);
708 mem_cgroup_update_mapped_file_stat(page, 1);
712 #ifdef CONFIG_DEBUG_VM
714 * page_dup_rmap - duplicate pte mapping to a page
715 * @page: the page to add the mapping to
716 * @vma: the vm area being duplicated
717 * @address: the user virtual address mapped
719 * For copy_page_range only: minimal extract from page_add_file_rmap /
720 * page_add_anon_rmap, avoiding unnecessary tests (already checked) so it's
721 * quicker.
723 * The caller needs to hold the pte lock.
725 void page_dup_rmap(struct page *page, struct vm_area_struct *vma, unsigned long address)
727 if (PageAnon(page))
728 __page_check_anon_rmap(page, vma, address);
729 atomic_inc(&page->_mapcount);
731 #endif
734 * page_remove_rmap - take down pte mapping from a page
735 * @page: page to remove mapping from
737 * The caller needs to hold the pte lock.
739 void page_remove_rmap(struct page *page)
741 if (atomic_add_negative(-1, &page->_mapcount)) {
743 * Now that the last pte has gone, s390 must transfer dirty
744 * flag from storage key to struct page. We can usually skip
745 * this if the page is anon, so about to be freed; but perhaps
746 * not if it's in swapcache - there might be another pte slot
747 * containing the swap entry, but page not yet written to swap.
749 if ((!PageAnon(page) || PageSwapCache(page)) &&
750 page_test_dirty(page)) {
751 page_clear_dirty(page);
752 set_page_dirty(page);
754 if (PageAnon(page))
755 mem_cgroup_uncharge_page(page);
756 __dec_zone_page_state(page,
757 PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED);
758 mem_cgroup_update_mapped_file_stat(page, -1);
760 * It would be tidy to reset the PageAnon mapping here,
761 * but that might overwrite a racing page_add_anon_rmap
762 * which increments mapcount after us but sets mapping
763 * before us: so leave the reset to free_hot_cold_page,
764 * and remember that it's only reliable while mapped.
765 * Leaving it set also helps swapoff to reinstate ptes
766 * faster for those pages still in swapcache.
772 * Subfunctions of try_to_unmap: try_to_unmap_one called
773 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
775 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
776 int migration)
778 struct mm_struct *mm = vma->vm_mm;
779 unsigned long address;
780 pte_t *pte;
781 pte_t pteval;
782 spinlock_t *ptl;
783 int ret = SWAP_AGAIN;
785 address = vma_address(page, vma);
786 if (address == -EFAULT)
787 goto out;
789 pte = page_check_address(page, mm, address, &ptl, 0);
790 if (!pte)
791 goto out;
794 * If the page is mlock()d, we cannot swap it out.
795 * If it's recently referenced (perhaps page_referenced
796 * skipped over this mm) then we should reactivate it.
798 if (!migration) {
799 if (vma->vm_flags & VM_LOCKED) {
800 ret = SWAP_MLOCK;
801 goto out_unmap;
803 if (ptep_clear_flush_young_notify(vma, address, pte)) {
804 ret = SWAP_FAIL;
805 goto out_unmap;
809 /* Nuke the page table entry. */
810 flush_cache_page(vma, address, page_to_pfn(page));
811 pteval = ptep_clear_flush_notify(vma, address, pte);
813 /* Move the dirty bit to the physical page now the pte is gone. */
814 if (pte_dirty(pteval))
815 set_page_dirty(page);
817 /* Update high watermark before we lower rss */
818 update_hiwater_rss(mm);
820 if (PageAnon(page)) {
821 swp_entry_t entry = { .val = page_private(page) };
823 if (PageSwapCache(page)) {
825 * Store the swap location in the pte.
826 * See handle_pte_fault() ...
828 swap_duplicate(entry);
829 if (list_empty(&mm->mmlist)) {
830 spin_lock(&mmlist_lock);
831 if (list_empty(&mm->mmlist))
832 list_add(&mm->mmlist, &init_mm.mmlist);
833 spin_unlock(&mmlist_lock);
835 dec_mm_counter(mm, anon_rss);
836 } else if (PAGE_MIGRATION) {
838 * Store the pfn of the page in a special migration
839 * pte. do_swap_page() will wait until the migration
840 * pte is removed and then restart fault handling.
842 BUG_ON(!migration);
843 entry = make_migration_entry(page, pte_write(pteval));
845 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
846 BUG_ON(pte_file(*pte));
847 } else if (PAGE_MIGRATION && migration) {
848 /* Establish migration entry for a file page */
849 swp_entry_t entry;
850 entry = make_migration_entry(page, pte_write(pteval));
851 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
852 } else
853 dec_mm_counter(mm, file_rss);
856 page_remove_rmap(page);
857 page_cache_release(page);
859 out_unmap:
860 pte_unmap_unlock(pte, ptl);
861 out:
862 return ret;
866 * objrmap doesn't work for nonlinear VMAs because the assumption that
867 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
868 * Consequently, given a particular page and its ->index, we cannot locate the
869 * ptes which are mapping that page without an exhaustive linear search.
871 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
872 * maps the file to which the target page belongs. The ->vm_private_data field
873 * holds the current cursor into that scan. Successive searches will circulate
874 * around the vma's virtual address space.
876 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
877 * more scanning pressure is placed against them as well. Eventually pages
878 * will become fully unmapped and are eligible for eviction.
880 * For very sparsely populated VMAs this is a little inefficient - chances are
881 * there there won't be many ptes located within the scan cluster. In this case
882 * maybe we could scan further - to the end of the pte page, perhaps.
884 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
885 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
886 * rather than unmapping them. If we encounter the "check_page" that vmscan is
887 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
889 #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
890 #define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
892 static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
893 struct vm_area_struct *vma, struct page *check_page)
895 struct mm_struct *mm = vma->vm_mm;
896 pgd_t *pgd;
897 pud_t *pud;
898 pmd_t *pmd;
899 pte_t *pte;
900 pte_t pteval;
901 spinlock_t *ptl;
902 struct page *page;
903 unsigned long address;
904 unsigned long end;
905 int ret = SWAP_AGAIN;
906 int locked_vma = 0;
908 address = (vma->vm_start + cursor) & CLUSTER_MASK;
909 end = address + CLUSTER_SIZE;
910 if (address < vma->vm_start)
911 address = vma->vm_start;
912 if (end > vma->vm_end)
913 end = vma->vm_end;
915 pgd = pgd_offset(mm, address);
916 if (!pgd_present(*pgd))
917 return ret;
919 pud = pud_offset(pgd, address);
920 if (!pud_present(*pud))
921 return ret;
923 pmd = pmd_offset(pud, address);
924 if (!pmd_present(*pmd))
925 return ret;
928 * MLOCK_PAGES => feature is configured.
929 * if we can acquire the mmap_sem for read, and vma is VM_LOCKED,
930 * keep the sem while scanning the cluster for mlocking pages.
932 if (MLOCK_PAGES && down_read_trylock(&vma->vm_mm->mmap_sem)) {
933 locked_vma = (vma->vm_flags & VM_LOCKED);
934 if (!locked_vma)
935 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
938 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
940 /* Update high watermark before we lower rss */
941 update_hiwater_rss(mm);
943 for (; address < end; pte++, address += PAGE_SIZE) {
944 if (!pte_present(*pte))
945 continue;
946 page = vm_normal_page(vma, address, *pte);
947 BUG_ON(!page || PageAnon(page));
949 if (locked_vma) {
950 mlock_vma_page(page); /* no-op if already mlocked */
951 if (page == check_page)
952 ret = SWAP_MLOCK;
953 continue; /* don't unmap */
956 if (ptep_clear_flush_young_notify(vma, address, pte))
957 continue;
959 /* Nuke the page table entry. */
960 flush_cache_page(vma, address, pte_pfn(*pte));
961 pteval = ptep_clear_flush_notify(vma, address, pte);
963 /* If nonlinear, store the file page offset in the pte. */
964 if (page->index != linear_page_index(vma, address))
965 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
967 /* Move the dirty bit to the physical page now the pte is gone. */
968 if (pte_dirty(pteval))
969 set_page_dirty(page);
971 page_remove_rmap(page);
972 page_cache_release(page);
973 dec_mm_counter(mm, file_rss);
974 (*mapcount)--;
976 pte_unmap_unlock(pte - 1, ptl);
977 if (locked_vma)
978 up_read(&vma->vm_mm->mmap_sem);
979 return ret;
983 * common handling for pages mapped in VM_LOCKED vmas
985 static int try_to_mlock_page(struct page *page, struct vm_area_struct *vma)
987 int mlocked = 0;
989 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
990 if (vma->vm_flags & VM_LOCKED) {
991 mlock_vma_page(page);
992 mlocked++; /* really mlocked the page */
994 up_read(&vma->vm_mm->mmap_sem);
996 return mlocked;
1000 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1001 * rmap method
1002 * @page: the page to unmap/unlock
1003 * @unlock: request for unlock rather than unmap [unlikely]
1004 * @migration: unmapping for migration - ignored if @unlock
1006 * Find all the mappings of a page using the mapping pointer and the vma chains
1007 * contained in the anon_vma struct it points to.
1009 * This function is only called from try_to_unmap/try_to_munlock for
1010 * anonymous pages.
1011 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1012 * where the page was found will be held for write. So, we won't recheck
1013 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1014 * 'LOCKED.
1016 static int try_to_unmap_anon(struct page *page, int unlock, int migration)
1018 struct anon_vma *anon_vma;
1019 struct vm_area_struct *vma;
1020 unsigned int mlocked = 0;
1021 int ret = SWAP_AGAIN;
1023 if (MLOCK_PAGES && unlikely(unlock))
1024 ret = SWAP_SUCCESS; /* default for try_to_munlock() */
1026 anon_vma = page_lock_anon_vma(page);
1027 if (!anon_vma)
1028 return ret;
1030 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
1031 if (MLOCK_PAGES && unlikely(unlock)) {
1032 if (!((vma->vm_flags & VM_LOCKED) &&
1033 page_mapped_in_vma(page, vma)))
1034 continue; /* must visit all unlocked vmas */
1035 ret = SWAP_MLOCK; /* saw at least one mlocked vma */
1036 } else {
1037 ret = try_to_unmap_one(page, vma, migration);
1038 if (ret == SWAP_FAIL || !page_mapped(page))
1039 break;
1041 if (ret == SWAP_MLOCK) {
1042 mlocked = try_to_mlock_page(page, vma);
1043 if (mlocked)
1044 break; /* stop if actually mlocked page */
1048 page_unlock_anon_vma(anon_vma);
1050 if (mlocked)
1051 ret = SWAP_MLOCK; /* actually mlocked the page */
1052 else if (ret == SWAP_MLOCK)
1053 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */
1055 return ret;
1059 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1060 * @page: the page to unmap/unlock
1061 * @unlock: request for unlock rather than unmap [unlikely]
1062 * @migration: unmapping for migration - ignored if @unlock
1064 * Find all the mappings of a page using the mapping pointer and the vma chains
1065 * contained in the address_space struct it points to.
1067 * This function is only called from try_to_unmap/try_to_munlock for
1068 * object-based pages.
1069 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1070 * where the page was found will be held for write. So, we won't recheck
1071 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1072 * 'LOCKED.
1074 static int try_to_unmap_file(struct page *page, int unlock, int migration)
1076 struct address_space *mapping = page->mapping;
1077 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1078 struct vm_area_struct *vma;
1079 struct prio_tree_iter iter;
1080 int ret = SWAP_AGAIN;
1081 unsigned long cursor;
1082 unsigned long max_nl_cursor = 0;
1083 unsigned long max_nl_size = 0;
1084 unsigned int mapcount;
1085 unsigned int mlocked = 0;
1087 if (MLOCK_PAGES && unlikely(unlock))
1088 ret = SWAP_SUCCESS; /* default for try_to_munlock() */
1090 spin_lock(&mapping->i_mmap_lock);
1091 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1092 if (MLOCK_PAGES && unlikely(unlock)) {
1093 if (!((vma->vm_flags & VM_LOCKED) &&
1094 page_mapped_in_vma(page, vma)))
1095 continue; /* must visit all vmas */
1096 ret = SWAP_MLOCK;
1097 } else {
1098 ret = try_to_unmap_one(page, vma, migration);
1099 if (ret == SWAP_FAIL || !page_mapped(page))
1100 goto out;
1102 if (ret == SWAP_MLOCK) {
1103 mlocked = try_to_mlock_page(page, vma);
1104 if (mlocked)
1105 break; /* stop if actually mlocked page */
1109 if (mlocked)
1110 goto out;
1112 if (list_empty(&mapping->i_mmap_nonlinear))
1113 goto out;
1115 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1116 shared.vm_set.list) {
1117 if (MLOCK_PAGES && unlikely(unlock)) {
1118 if (!(vma->vm_flags & VM_LOCKED))
1119 continue; /* must visit all vmas */
1120 ret = SWAP_MLOCK; /* leave mlocked == 0 */
1121 goto out; /* no need to look further */
1123 if (!MLOCK_PAGES && !migration && (vma->vm_flags & VM_LOCKED))
1124 continue;
1125 cursor = (unsigned long) vma->vm_private_data;
1126 if (cursor > max_nl_cursor)
1127 max_nl_cursor = cursor;
1128 cursor = vma->vm_end - vma->vm_start;
1129 if (cursor > max_nl_size)
1130 max_nl_size = cursor;
1133 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1134 ret = SWAP_FAIL;
1135 goto out;
1139 * We don't try to search for this page in the nonlinear vmas,
1140 * and page_referenced wouldn't have found it anyway. Instead
1141 * just walk the nonlinear vmas trying to age and unmap some.
1142 * The mapcount of the page we came in with is irrelevant,
1143 * but even so use it as a guide to how hard we should try?
1145 mapcount = page_mapcount(page);
1146 if (!mapcount)
1147 goto out;
1148 cond_resched_lock(&mapping->i_mmap_lock);
1150 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1151 if (max_nl_cursor == 0)
1152 max_nl_cursor = CLUSTER_SIZE;
1154 do {
1155 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1156 shared.vm_set.list) {
1157 if (!MLOCK_PAGES && !migration &&
1158 (vma->vm_flags & VM_LOCKED))
1159 continue;
1160 cursor = (unsigned long) vma->vm_private_data;
1161 while ( cursor < max_nl_cursor &&
1162 cursor < vma->vm_end - vma->vm_start) {
1163 ret = try_to_unmap_cluster(cursor, &mapcount,
1164 vma, page);
1165 if (ret == SWAP_MLOCK)
1166 mlocked = 2; /* to return below */
1167 cursor += CLUSTER_SIZE;
1168 vma->vm_private_data = (void *) cursor;
1169 if ((int)mapcount <= 0)
1170 goto out;
1172 vma->vm_private_data = (void *) max_nl_cursor;
1174 cond_resched_lock(&mapping->i_mmap_lock);
1175 max_nl_cursor += CLUSTER_SIZE;
1176 } while (max_nl_cursor <= max_nl_size);
1179 * Don't loop forever (perhaps all the remaining pages are
1180 * in locked vmas). Reset cursor on all unreserved nonlinear
1181 * vmas, now forgetting on which ones it had fallen behind.
1183 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1184 vma->vm_private_data = NULL;
1185 out:
1186 spin_unlock(&mapping->i_mmap_lock);
1187 if (mlocked)
1188 ret = SWAP_MLOCK; /* actually mlocked the page */
1189 else if (ret == SWAP_MLOCK)
1190 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */
1191 return ret;
1195 * try_to_unmap - try to remove all page table mappings to a page
1196 * @page: the page to get unmapped
1197 * @migration: migration flag
1199 * Tries to remove all the page table entries which are mapping this
1200 * page, used in the pageout path. Caller must hold the page lock.
1201 * Return values are:
1203 * SWAP_SUCCESS - we succeeded in removing all mappings
1204 * SWAP_AGAIN - we missed a mapping, try again later
1205 * SWAP_FAIL - the page is unswappable
1206 * SWAP_MLOCK - page is mlocked.
1208 int try_to_unmap(struct page *page, int migration)
1210 int ret;
1212 BUG_ON(!PageLocked(page));
1214 if (PageAnon(page))
1215 ret = try_to_unmap_anon(page, 0, migration);
1216 else
1217 ret = try_to_unmap_file(page, 0, migration);
1218 if (ret != SWAP_MLOCK && !page_mapped(page))
1219 ret = SWAP_SUCCESS;
1220 return ret;
1224 * try_to_munlock - try to munlock a page
1225 * @page: the page to be munlocked
1227 * Called from munlock code. Checks all of the VMAs mapping the page
1228 * to make sure nobody else has this page mlocked. The page will be
1229 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1231 * Return values are:
1233 * SWAP_SUCCESS - no vma's holding page mlocked.
1234 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1235 * SWAP_MLOCK - page is now mlocked.
1237 int try_to_munlock(struct page *page)
1239 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1241 if (PageAnon(page))
1242 return try_to_unmap_anon(page, 1, 0);
1243 else
1244 return try_to_unmap_file(page, 1, 0);