Root/mm/rmap.c

1/*
2 * mm/rmap.c - physical to virtual reverse mappings
3 *
4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5 * Released under the General Public License (GPL).
6 *
7 * Simple, low overhead reverse mapping scheme.
8 * Please try to keep this thing as modular as possible.
9 *
10 * Provides methods for unmapping each kind of mapped page:
11 * the anon methods track anonymous pages, and
12 * the file methods track pages belonging to an inode.
13 *
14 * Original design by Rik van Riel <riel@conectiva.com.br> 2001
15 * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
16 * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
17 * Contributions by Hugh Dickins 2003, 2004
18 */
19
20/*
21 * Lock ordering in mm:
22 *
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)
39 *
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
44 */
45
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/ksm.h>
53#include <linux/rmap.h>
54#include <linux/rcupdate.h>
55#include <linux/module.h>
56#include <linux/memcontrol.h>
57#include <linux/mmu_notifier.h>
58#include <linux/migrate.h>
59
60#include <asm/tlbflush.h>
61
62#include "internal.h"
63
64static struct kmem_cache *anon_vma_cachep;
65static struct kmem_cache *anon_vma_chain_cachep;
66
67static inline struct anon_vma *anon_vma_alloc(void)
68{
69    return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
70}
71
72void anon_vma_free(struct anon_vma *anon_vma)
73{
74    kmem_cache_free(anon_vma_cachep, anon_vma);
75}
76
77static inline struct anon_vma_chain *anon_vma_chain_alloc(void)
78{
79    return kmem_cache_alloc(anon_vma_chain_cachep, GFP_KERNEL);
80}
81
82void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
83{
84    kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
85}
86
87/**
88 * anon_vma_prepare - attach an anon_vma to a memory region
89 * @vma: the memory region in question
90 *
91 * This makes sure the memory mapping described by 'vma' has
92 * an 'anon_vma' attached to it, so that we can associate the
93 * anonymous pages mapped into it with that anon_vma.
94 *
95 * The common case will be that we already have one, but if
96 * if not we either need to find an adjacent mapping that we
97 * can re-use the anon_vma from (very common when the only
98 * reason for splitting a vma has been mprotect()), or we
99 * allocate a new one.
100 *
101 * Anon-vma allocations are very subtle, because we may have
102 * optimistically looked up an anon_vma in page_lock_anon_vma()
103 * and that may actually touch the spinlock even in the newly
104 * allocated vma (it depends on RCU to make sure that the
105 * anon_vma isn't actually destroyed).
106 *
107 * As a result, we need to do proper anon_vma locking even
108 * for the new allocation. At the same time, we do not want
109 * to do any locking for the common case of already having
110 * an anon_vma.
111 *
112 * This must be called with the mmap_sem held for reading.
113 */
114int anon_vma_prepare(struct vm_area_struct *vma)
115{
116    struct anon_vma *anon_vma = vma->anon_vma;
117    struct anon_vma_chain *avc;
118
119    might_sleep();
120    if (unlikely(!anon_vma)) {
121        struct mm_struct *mm = vma->vm_mm;
122        struct anon_vma *allocated;
123
124        avc = anon_vma_chain_alloc();
125        if (!avc)
126            goto out_enomem;
127
128        anon_vma = find_mergeable_anon_vma(vma);
129        allocated = NULL;
130        if (!anon_vma) {
131            anon_vma = anon_vma_alloc();
132            if (unlikely(!anon_vma))
133                goto out_enomem_free_avc;
134            allocated = anon_vma;
135        }
136        spin_lock(&anon_vma->lock);
137
138        /* page_table_lock to protect against threads */
139        spin_lock(&mm->page_table_lock);
140        if (likely(!vma->anon_vma)) {
141            vma->anon_vma = anon_vma;
142            avc->anon_vma = anon_vma;
143            avc->vma = vma;
144            list_add(&avc->same_vma, &vma->anon_vma_chain);
145            list_add(&avc->same_anon_vma, &anon_vma->head);
146            allocated = NULL;
147        }
148        spin_unlock(&mm->page_table_lock);
149
150        spin_unlock(&anon_vma->lock);
151        if (unlikely(allocated)) {
152            anon_vma_free(allocated);
153            anon_vma_chain_free(avc);
154        }
155    }
156    return 0;
157
158 out_enomem_free_avc:
159    anon_vma_chain_free(avc);
160 out_enomem:
161    return -ENOMEM;
162}
163
164static void anon_vma_chain_link(struct vm_area_struct *vma,
165                struct anon_vma_chain *avc,
166                struct anon_vma *anon_vma)
167{
168    avc->vma = vma;
169    avc->anon_vma = anon_vma;
170    list_add(&avc->same_vma, &vma->anon_vma_chain);
171
172    spin_lock(&anon_vma->lock);
173    list_add_tail(&avc->same_anon_vma, &anon_vma->head);
174    spin_unlock(&anon_vma->lock);
175}
176
177/*
178 * Attach the anon_vmas from src to dst.
179 * Returns 0 on success, -ENOMEM on failure.
180 */
181int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
182{
183    struct anon_vma_chain *avc, *pavc;
184
185    list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
186        avc = anon_vma_chain_alloc();
187        if (!avc)
188            goto enomem_failure;
189        anon_vma_chain_link(dst, avc, pavc->anon_vma);
190    }
191    return 0;
192
193 enomem_failure:
194    unlink_anon_vmas(dst);
195    return -ENOMEM;
196}
197
198/*
199 * Attach vma to its own anon_vma, as well as to the anon_vmas that
200 * the corresponding VMA in the parent process is attached to.
201 * Returns 0 on success, non-zero on failure.
202 */
203int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
204{
205    struct anon_vma_chain *avc;
206    struct anon_vma *anon_vma;
207
208    /* Don't bother if the parent process has no anon_vma here. */
209    if (!pvma->anon_vma)
210        return 0;
211
212    /*
213     * First, attach the new VMA to the parent VMA's anon_vmas,
214     * so rmap can find non-COWed pages in child processes.
215     */
216    if (anon_vma_clone(vma, pvma))
217        return -ENOMEM;
218
219    /* Then add our own anon_vma. */
220    anon_vma = anon_vma_alloc();
221    if (!anon_vma)
222        goto out_error;
223    avc = anon_vma_chain_alloc();
224    if (!avc)
225        goto out_error_free_anon_vma;
226    anon_vma_chain_link(vma, avc, anon_vma);
227    /* Mark this anon_vma as the one where our new (COWed) pages go. */
228    vma->anon_vma = anon_vma;
229
230    return 0;
231
232 out_error_free_anon_vma:
233    anon_vma_free(anon_vma);
234 out_error:
235    unlink_anon_vmas(vma);
236    return -ENOMEM;
237}
238
239static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain)
240{
241    struct anon_vma *anon_vma = anon_vma_chain->anon_vma;
242    int empty;
243
244    /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */
245    if (!anon_vma)
246        return;
247
248    spin_lock(&anon_vma->lock);
249    list_del(&anon_vma_chain->same_anon_vma);
250
251    /* We must garbage collect the anon_vma if it's empty */
252    empty = list_empty(&anon_vma->head) && !ksm_refcount(anon_vma);
253    spin_unlock(&anon_vma->lock);
254
255    if (empty)
256        anon_vma_free(anon_vma);
257}
258
259void unlink_anon_vmas(struct vm_area_struct *vma)
260{
261    struct anon_vma_chain *avc, *next;
262
263    /* Unlink each anon_vma chained to the VMA. */
264    list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
265        anon_vma_unlink(avc);
266        list_del(&avc->same_vma);
267        anon_vma_chain_free(avc);
268    }
269}
270
271static void anon_vma_ctor(void *data)
272{
273    struct anon_vma *anon_vma = data;
274
275    spin_lock_init(&anon_vma->lock);
276    ksm_refcount_init(anon_vma);
277    INIT_LIST_HEAD(&anon_vma->head);
278}
279
280void __init anon_vma_init(void)
281{
282    anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
283            0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
284    anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
285}
286
287/*
288 * Getting a lock on a stable anon_vma from a page off the LRU is
289 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
290 */
291struct anon_vma *page_lock_anon_vma(struct page *page)
292{
293    struct anon_vma *anon_vma;
294    unsigned long anon_mapping;
295
296    rcu_read_lock();
297    anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
298    if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
299        goto out;
300    if (!page_mapped(page))
301        goto out;
302
303    anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
304    spin_lock(&anon_vma->lock);
305    return anon_vma;
306out:
307    rcu_read_unlock();
308    return NULL;
309}
310
311void page_unlock_anon_vma(struct anon_vma *anon_vma)
312{
313    spin_unlock(&anon_vma->lock);
314    rcu_read_unlock();
315}
316
317/*
318 * At what user virtual address is page expected in @vma?
319 * Returns virtual address or -EFAULT if page's index/offset is not
320 * within the range mapped the @vma.
321 */
322static inline unsigned long
323vma_address(struct page *page, struct vm_area_struct *vma)
324{
325    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
326    unsigned long address;
327
328    address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
329    if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
330        /* page should be within @vma mapping range */
331        return -EFAULT;
332    }
333    return address;
334}
335
336/*
337 * At what user virtual address is page expected in vma?
338 * checking that the page matches the vma.
339 */
340unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
341{
342    if (PageAnon(page)) {
343        if (vma->anon_vma != page_anon_vma(page))
344            return -EFAULT;
345    } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
346        if (!vma->vm_file ||
347            vma->vm_file->f_mapping != page->mapping)
348            return -EFAULT;
349    } else
350        return -EFAULT;
351    return vma_address(page, vma);
352}
353
354/*
355 * Check that @page is mapped at @address into @mm.
356 *
357 * If @sync is false, page_check_address may perform a racy check to avoid
358 * the page table lock when the pte is not present (helpful when reclaiming
359 * highly shared pages).
360 *
361 * On success returns with pte mapped and locked.
362 */
363pte_t *page_check_address(struct page *page, struct mm_struct *mm,
364              unsigned long address, spinlock_t **ptlp, int sync)
365{
366    pgd_t *pgd;
367    pud_t *pud;
368    pmd_t *pmd;
369    pte_t *pte;
370    spinlock_t *ptl;
371
372    pgd = pgd_offset(mm, address);
373    if (!pgd_present(*pgd))
374        return NULL;
375
376    pud = pud_offset(pgd, address);
377    if (!pud_present(*pud))
378        return NULL;
379
380    pmd = pmd_offset(pud, address);
381    if (!pmd_present(*pmd))
382        return NULL;
383
384    pte = pte_offset_map(pmd, address);
385    /* Make a quick check before getting the lock */
386    if (!sync && !pte_present(*pte)) {
387        pte_unmap(pte);
388        return NULL;
389    }
390
391    ptl = pte_lockptr(mm, pmd);
392    spin_lock(ptl);
393    if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
394        *ptlp = ptl;
395        return pte;
396    }
397    pte_unmap_unlock(pte, ptl);
398    return NULL;
399}
400
401/**
402 * page_mapped_in_vma - check whether a page is really mapped in a VMA
403 * @page: the page to test
404 * @vma: the VMA to test
405 *
406 * Returns 1 if the page is mapped into the page tables of the VMA, 0
407 * if the page is not mapped into the page tables of this VMA. Only
408 * valid for normal file or anonymous VMAs.
409 */
410int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
411{
412    unsigned long address;
413    pte_t *pte;
414    spinlock_t *ptl;
415
416    address = vma_address(page, vma);
417    if (address == -EFAULT) /* out of vma range */
418        return 0;
419    pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
420    if (!pte) /* the page is not in this mm */
421        return 0;
422    pte_unmap_unlock(pte, ptl);
423
424    return 1;
425}
426
427/*
428 * Subfunctions of page_referenced: page_referenced_one called
429 * repeatedly from either page_referenced_anon or page_referenced_file.
430 */
431int page_referenced_one(struct page *page, struct vm_area_struct *vma,
432            unsigned long address, unsigned int *mapcount,
433            unsigned long *vm_flags)
434{
435    struct mm_struct *mm = vma->vm_mm;
436    pte_t *pte;
437    spinlock_t *ptl;
438    int referenced = 0;
439
440    pte = page_check_address(page, mm, address, &ptl, 0);
441    if (!pte)
442        goto out;
443
444    /*
445     * Don't want to elevate referenced for mlocked page that gets this far,
446     * in order that it progresses to try_to_unmap and is moved to the
447     * unevictable list.
448     */
449    if (vma->vm_flags & VM_LOCKED) {
450        *mapcount = 1; /* break early from loop */
451        *vm_flags |= VM_LOCKED;
452        goto out_unmap;
453    }
454
455    if (ptep_clear_flush_young_notify(vma, address, pte)) {
456        /*
457         * Don't treat a reference through a sequentially read
458         * mapping as such. If the page has been used in
459         * another mapping, we will catch it; if this other
460         * mapping is already gone, the unmap path will have
461         * set PG_referenced or activated the page.
462         */
463        if (likely(!VM_SequentialReadHint(vma)))
464            referenced++;
465    }
466
467    /* Pretend the page is referenced if the task has the
468       swap token and is in the middle of a page fault. */
469    if (mm != current->mm && has_swap_token(mm) &&
470            rwsem_is_locked(&mm->mmap_sem))
471        referenced++;
472
473out_unmap:
474    (*mapcount)--;
475    pte_unmap_unlock(pte, ptl);
476
477    if (referenced)
478        *vm_flags |= vma->vm_flags;
479out:
480    return referenced;
481}
482
483static int page_referenced_anon(struct page *page,
484                struct mem_cgroup *mem_cont,
485                unsigned long *vm_flags)
486{
487    unsigned int mapcount;
488    struct anon_vma *anon_vma;
489    struct anon_vma_chain *avc;
490    int referenced = 0;
491
492    anon_vma = page_lock_anon_vma(page);
493    if (!anon_vma)
494        return referenced;
495
496    mapcount = page_mapcount(page);
497    list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
498        struct vm_area_struct *vma = avc->vma;
499        unsigned long address = vma_address(page, vma);
500        if (address == -EFAULT)
501            continue;
502        /*
503         * If we are reclaiming on behalf of a cgroup, skip
504         * counting on behalf of references from different
505         * cgroups
506         */
507        if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
508            continue;
509        referenced += page_referenced_one(page, vma, address,
510                          &mapcount, vm_flags);
511        if (!mapcount)
512            break;
513    }
514
515    page_unlock_anon_vma(anon_vma);
516    return referenced;
517}
518
519/**
520 * page_referenced_file - referenced check for object-based rmap
521 * @page: the page we're checking references on.
522 * @mem_cont: target memory controller
523 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
524 *
525 * For an object-based mapped page, find all the places it is mapped and
526 * check/clear the referenced flag. This is done by following the page->mapping
527 * pointer, then walking the chain of vmas it holds. It returns the number
528 * of references it found.
529 *
530 * This function is only called from page_referenced for object-based pages.
531 */
532static int page_referenced_file(struct page *page,
533                struct mem_cgroup *mem_cont,
534                unsigned long *vm_flags)
535{
536    unsigned int mapcount;
537    struct address_space *mapping = page->mapping;
538    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
539    struct vm_area_struct *vma;
540    struct prio_tree_iter iter;
541    int referenced = 0;
542
543    /*
544     * The caller's checks on page->mapping and !PageAnon have made
545     * sure that this is a file page: the check for page->mapping
546     * excludes the case just before it gets set on an anon page.
547     */
548    BUG_ON(PageAnon(page));
549
550    /*
551     * The page lock not only makes sure that page->mapping cannot
552     * suddenly be NULLified by truncation, it makes sure that the
553     * structure at mapping cannot be freed and reused yet,
554     * so we can safely take mapping->i_mmap_lock.
555     */
556    BUG_ON(!PageLocked(page));
557
558    spin_lock(&mapping->i_mmap_lock);
559
560    /*
561     * i_mmap_lock does not stabilize mapcount at all, but mapcount
562     * is more likely to be accurate if we note it after spinning.
563     */
564    mapcount = page_mapcount(page);
565
566    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
567        unsigned long address = vma_address(page, vma);
568        if (address == -EFAULT)
569            continue;
570        /*
571         * If we are reclaiming on behalf of a cgroup, skip
572         * counting on behalf of references from different
573         * cgroups
574         */
575        if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
576            continue;
577        referenced += page_referenced_one(page, vma, address,
578                          &mapcount, vm_flags);
579        if (!mapcount)
580            break;
581    }
582
583    spin_unlock(&mapping->i_mmap_lock);
584    return referenced;
585}
586
587/**
588 * page_referenced - test if the page was referenced
589 * @page: the page to test
590 * @is_locked: caller holds lock on the page
591 * @mem_cont: target memory controller
592 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
593 *
594 * Quick test_and_clear_referenced for all mappings to a page,
595 * returns the number of ptes which referenced the page.
596 */
597int page_referenced(struct page *page,
598            int is_locked,
599            struct mem_cgroup *mem_cont,
600            unsigned long *vm_flags)
601{
602    int referenced = 0;
603    int we_locked = 0;
604
605    *vm_flags = 0;
606    if (page_mapped(page) && page_rmapping(page)) {
607        if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
608            we_locked = trylock_page(page);
609            if (!we_locked) {
610                referenced++;
611                goto out;
612            }
613        }
614        if (unlikely(PageKsm(page)))
615            referenced += page_referenced_ksm(page, mem_cont,
616                                vm_flags);
617        else if (PageAnon(page))
618            referenced += page_referenced_anon(page, mem_cont,
619                                vm_flags);
620        else if (page->mapping)
621            referenced += page_referenced_file(page, mem_cont,
622                                vm_flags);
623        if (we_locked)
624            unlock_page(page);
625    }
626out:
627    if (page_test_and_clear_young(page))
628        referenced++;
629
630    return referenced;
631}
632
633static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
634                unsigned long address)
635{
636    struct mm_struct *mm = vma->vm_mm;
637    pte_t *pte;
638    spinlock_t *ptl;
639    int ret = 0;
640
641    pte = page_check_address(page, mm, address, &ptl, 1);
642    if (!pte)
643        goto out;
644
645    if (pte_dirty(*pte) || pte_write(*pte)) {
646        pte_t entry;
647
648        flush_cache_page(vma, address, pte_pfn(*pte));
649        entry = ptep_clear_flush_notify(vma, address, pte);
650        entry = pte_wrprotect(entry);
651        entry = pte_mkclean(entry);
652        set_pte_at(mm, address, pte, entry);
653        ret = 1;
654    }
655
656    pte_unmap_unlock(pte, ptl);
657out:
658    return ret;
659}
660
661static int page_mkclean_file(struct address_space *mapping, struct page *page)
662{
663    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
664    struct vm_area_struct *vma;
665    struct prio_tree_iter iter;
666    int ret = 0;
667
668    BUG_ON(PageAnon(page));
669
670    spin_lock(&mapping->i_mmap_lock);
671    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
672        if (vma->vm_flags & VM_SHARED) {
673            unsigned long address = vma_address(page, vma);
674            if (address == -EFAULT)
675                continue;
676            ret += page_mkclean_one(page, vma, address);
677        }
678    }
679    spin_unlock(&mapping->i_mmap_lock);
680    return ret;
681}
682
683int page_mkclean(struct page *page)
684{
685    int ret = 0;
686
687    BUG_ON(!PageLocked(page));
688
689    if (page_mapped(page)) {
690        struct address_space *mapping = page_mapping(page);
691        if (mapping) {
692            ret = page_mkclean_file(mapping, page);
693            if (page_test_dirty(page)) {
694                page_clear_dirty(page);
695                ret = 1;
696            }
697        }
698    }
699
700    return ret;
701}
702EXPORT_SYMBOL_GPL(page_mkclean);
703
704/**
705 * page_move_anon_rmap - move a page to our anon_vma
706 * @page: the page to move to our anon_vma
707 * @vma: the vma the page belongs to
708 * @address: the user virtual address mapped
709 *
710 * When a page belongs exclusively to one process after a COW event,
711 * that page can be moved into the anon_vma that belongs to just that
712 * process, so the rmap code will not search the parent or sibling
713 * processes.
714 */
715void page_move_anon_rmap(struct page *page,
716    struct vm_area_struct *vma, unsigned long address)
717{
718    struct anon_vma *anon_vma = vma->anon_vma;
719
720    VM_BUG_ON(!PageLocked(page));
721    VM_BUG_ON(!anon_vma);
722    VM_BUG_ON(page->index != linear_page_index(vma, address));
723
724    anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
725    page->mapping = (struct address_space *) anon_vma;
726}
727
728/**
729 * __page_set_anon_rmap - setup new anonymous rmap
730 * @page: the page to add the mapping to
731 * @vma: the vm area in which the mapping is added
732 * @address: the user virtual address mapped
733 * @exclusive: the page is exclusively owned by the current process
734 */
735static void __page_set_anon_rmap(struct page *page,
736    struct vm_area_struct *vma, unsigned long address, int exclusive)
737{
738    struct anon_vma *anon_vma = vma->anon_vma;
739
740    BUG_ON(!anon_vma);
741
742    /*
743     * If the page isn't exclusively mapped into this vma,
744     * we must use the _oldest_ possible anon_vma for the
745     * page mapping!
746     *
747     * So take the last AVC chain entry in the vma, which is
748     * the deepest ancestor, and use the anon_vma from that.
749     */
750    if (!exclusive) {
751        struct anon_vma_chain *avc;
752        avc = list_entry(vma->anon_vma_chain.prev, struct anon_vma_chain, same_vma);
753        anon_vma = avc->anon_vma;
754    }
755
756    anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
757    page->mapping = (struct address_space *) anon_vma;
758    page->index = linear_page_index(vma, address);
759}
760
761/**
762 * __page_check_anon_rmap - sanity check anonymous rmap addition
763 * @page: the page to add the mapping to
764 * @vma: the vm area in which the mapping is added
765 * @address: the user virtual address mapped
766 */
767static void __page_check_anon_rmap(struct page *page,
768    struct vm_area_struct *vma, unsigned long address)
769{
770#ifdef CONFIG_DEBUG_VM
771    /*
772     * The page's anon-rmap details (mapping and index) are guaranteed to
773     * be set up correctly at this point.
774     *
775     * We have exclusion against page_add_anon_rmap because the caller
776     * always holds the page locked, except if called from page_dup_rmap,
777     * in which case the page is already known to be setup.
778     *
779     * We have exclusion against page_add_new_anon_rmap because those pages
780     * are initially only visible via the pagetables, and the pte is locked
781     * over the call to page_add_new_anon_rmap.
782     */
783    BUG_ON(page->index != linear_page_index(vma, address));
784#endif
785}
786
787/**
788 * page_add_anon_rmap - add pte mapping to an anonymous page
789 * @page: the page to add the mapping to
790 * @vma: the vm area in which the mapping is added
791 * @address: the user virtual address mapped
792 *
793 * The caller needs to hold the pte lock, and the page must be locked in
794 * the anon_vma case: to serialize mapping,index checking after setting,
795 * and to ensure that PageAnon is not being upgraded racily to PageKsm
796 * (but PageKsm is never downgraded to PageAnon).
797 */
798void page_add_anon_rmap(struct page *page,
799    struct vm_area_struct *vma, unsigned long address)
800{
801    int first = atomic_inc_and_test(&page->_mapcount);
802    if (first)
803        __inc_zone_page_state(page, NR_ANON_PAGES);
804    if (unlikely(PageKsm(page)))
805        return;
806
807    VM_BUG_ON(!PageLocked(page));
808    VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
809    if (first)
810        __page_set_anon_rmap(page, vma, address, 0);
811    else
812        __page_check_anon_rmap(page, vma, address);
813}
814
815/**
816 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
817 * @page: the page to add the mapping to
818 * @vma: the vm area in which the mapping is added
819 * @address: the user virtual address mapped
820 *
821 * Same as page_add_anon_rmap but must only be called on *new* pages.
822 * This means the inc-and-test can be bypassed.
823 * Page does not have to be locked.
824 */
825void page_add_new_anon_rmap(struct page *page,
826    struct vm_area_struct *vma, unsigned long address)
827{
828    VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
829    SetPageSwapBacked(page);
830    atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
831    __inc_zone_page_state(page, NR_ANON_PAGES);
832    __page_set_anon_rmap(page, vma, address, 1);
833    if (page_evictable(page, vma))
834        lru_cache_add_lru(page, LRU_ACTIVE_ANON);
835    else
836        add_page_to_unevictable_list(page);
837}
838
839/**
840 * page_add_file_rmap - add pte mapping to a file page
841 * @page: the page to add the mapping to
842 *
843 * The caller needs to hold the pte lock.
844 */
845void page_add_file_rmap(struct page *page)
846{
847    if (atomic_inc_and_test(&page->_mapcount)) {
848        __inc_zone_page_state(page, NR_FILE_MAPPED);
849        mem_cgroup_update_file_mapped(page, 1);
850    }
851}
852
853/**
854 * page_remove_rmap - take down pte mapping from a page
855 * @page: page to remove mapping from
856 *
857 * The caller needs to hold the pte lock.
858 */
859void page_remove_rmap(struct page *page)
860{
861    /* page still mapped by someone else? */
862    if (!atomic_add_negative(-1, &page->_mapcount))
863        return;
864
865    /*
866     * Now that the last pte has gone, s390 must transfer dirty
867     * flag from storage key to struct page. We can usually skip
868     * this if the page is anon, so about to be freed; but perhaps
869     * not if it's in swapcache - there might be another pte slot
870     * containing the swap entry, but page not yet written to swap.
871     */
872    if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
873        page_clear_dirty(page);
874        set_page_dirty(page);
875    }
876    if (PageAnon(page)) {
877        mem_cgroup_uncharge_page(page);
878        __dec_zone_page_state(page, NR_ANON_PAGES);
879    } else {
880        __dec_zone_page_state(page, NR_FILE_MAPPED);
881        mem_cgroup_update_file_mapped(page, -1);
882    }
883    /*
884     * It would be tidy to reset the PageAnon mapping here,
885     * but that might overwrite a racing page_add_anon_rmap
886     * which increments mapcount after us but sets mapping
887     * before us: so leave the reset to free_hot_cold_page,
888     * and remember that it's only reliable while mapped.
889     * Leaving it set also helps swapoff to reinstate ptes
890     * faster for those pages still in swapcache.
891     */
892}
893
894/*
895 * Subfunctions of try_to_unmap: try_to_unmap_one called
896 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
897 */
898int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
899             unsigned long address, enum ttu_flags flags)
900{
901    struct mm_struct *mm = vma->vm_mm;
902    pte_t *pte;
903    pte_t pteval;
904    spinlock_t *ptl;
905    int ret = SWAP_AGAIN;
906
907    pte = page_check_address(page, mm, address, &ptl, 0);
908    if (!pte)
909        goto out;
910
911    /*
912     * If the page is mlock()d, we cannot swap it out.
913     * If it's recently referenced (perhaps page_referenced
914     * skipped over this mm) then we should reactivate it.
915     */
916    if (!(flags & TTU_IGNORE_MLOCK)) {
917        if (vma->vm_flags & VM_LOCKED)
918            goto out_mlock;
919
920        if (TTU_ACTION(flags) == TTU_MUNLOCK)
921            goto out_unmap;
922    }
923    if (!(flags & TTU_IGNORE_ACCESS)) {
924        if (ptep_clear_flush_young_notify(vma, address, pte)) {
925            ret = SWAP_FAIL;
926            goto out_unmap;
927        }
928      }
929
930    /* Nuke the page table entry. */
931    flush_cache_page(vma, address, page_to_pfn(page));
932    pteval = ptep_clear_flush_notify(vma, address, pte);
933
934    /* Move the dirty bit to the physical page now the pte is gone. */
935    if (pte_dirty(pteval))
936        set_page_dirty(page);
937
938    /* Update high watermark before we lower rss */
939    update_hiwater_rss(mm);
940
941    if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
942        if (PageAnon(page))
943            dec_mm_counter(mm, MM_ANONPAGES);
944        else
945            dec_mm_counter(mm, MM_FILEPAGES);
946        set_pte_at(mm, address, pte,
947                swp_entry_to_pte(make_hwpoison_entry(page)));
948    } else if (PageAnon(page)) {
949        swp_entry_t entry = { .val = page_private(page) };
950
951        if (PageSwapCache(page)) {
952            /*
953             * Store the swap location in the pte.
954             * See handle_pte_fault() ...
955             */
956            if (swap_duplicate(entry) < 0) {
957                set_pte_at(mm, address, pte, pteval);
958                ret = SWAP_FAIL;
959                goto out_unmap;
960            }
961            if (list_empty(&mm->mmlist)) {
962                spin_lock(&mmlist_lock);
963                if (list_empty(&mm->mmlist))
964                    list_add(&mm->mmlist, &init_mm.mmlist);
965                spin_unlock(&mmlist_lock);
966            }
967            dec_mm_counter(mm, MM_ANONPAGES);
968            inc_mm_counter(mm, MM_SWAPENTS);
969        } else if (PAGE_MIGRATION) {
970            /*
971             * Store the pfn of the page in a special migration
972             * pte. do_swap_page() will wait until the migration
973             * pte is removed and then restart fault handling.
974             */
975            BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
976            entry = make_migration_entry(page, pte_write(pteval));
977        }
978        set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
979        BUG_ON(pte_file(*pte));
980    } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
981        /* Establish migration entry for a file page */
982        swp_entry_t entry;
983        entry = make_migration_entry(page, pte_write(pteval));
984        set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
985    } else
986        dec_mm_counter(mm, MM_FILEPAGES);
987
988    page_remove_rmap(page);
989    page_cache_release(page);
990
991out_unmap:
992    pte_unmap_unlock(pte, ptl);
993out:
994    return ret;
995
996out_mlock:
997    pte_unmap_unlock(pte, ptl);
998
999
1000    /*
1001     * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1002     * unstable result and race. Plus, We can't wait here because
1003     * we now hold anon_vma->lock or mapping->i_mmap_lock.
1004     * if trylock failed, the page remain in evictable lru and later
1005     * vmscan could retry to move the page to unevictable lru if the
1006     * page is actually mlocked.
1007     */
1008    if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1009        if (vma->vm_flags & VM_LOCKED) {
1010            mlock_vma_page(page);
1011            ret = SWAP_MLOCK;
1012        }
1013        up_read(&vma->vm_mm->mmap_sem);
1014    }
1015    return ret;
1016}
1017
1018/*
1019 * objrmap doesn't work for nonlinear VMAs because the assumption that
1020 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1021 * Consequently, given a particular page and its ->index, we cannot locate the
1022 * ptes which are mapping that page without an exhaustive linear search.
1023 *
1024 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1025 * maps the file to which the target page belongs. The ->vm_private_data field
1026 * holds the current cursor into that scan. Successive searches will circulate
1027 * around the vma's virtual address space.
1028 *
1029 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1030 * more scanning pressure is placed against them as well. Eventually pages
1031 * will become fully unmapped and are eligible for eviction.
1032 *
1033 * For very sparsely populated VMAs this is a little inefficient - chances are
1034 * there there won't be many ptes located within the scan cluster. In this case
1035 * maybe we could scan further - to the end of the pte page, perhaps.
1036 *
1037 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1038 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1039 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1040 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1041 */
1042#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1043#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1044
1045static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1046        struct vm_area_struct *vma, struct page *check_page)
1047{
1048    struct mm_struct *mm = vma->vm_mm;
1049    pgd_t *pgd;
1050    pud_t *pud;
1051    pmd_t *pmd;
1052    pte_t *pte;
1053    pte_t pteval;
1054    spinlock_t *ptl;
1055    struct page *page;
1056    unsigned long address;
1057    unsigned long end;
1058    int ret = SWAP_AGAIN;
1059    int locked_vma = 0;
1060
1061    address = (vma->vm_start + cursor) & CLUSTER_MASK;
1062    end = address + CLUSTER_SIZE;
1063    if (address < vma->vm_start)
1064        address = vma->vm_start;
1065    if (end > vma->vm_end)
1066        end = vma->vm_end;
1067
1068    pgd = pgd_offset(mm, address);
1069    if (!pgd_present(*pgd))
1070        return ret;
1071
1072    pud = pud_offset(pgd, address);
1073    if (!pud_present(*pud))
1074        return ret;
1075
1076    pmd = pmd_offset(pud, address);
1077    if (!pmd_present(*pmd))
1078        return ret;
1079
1080    /*
1081     * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1082     * keep the sem while scanning the cluster for mlocking pages.
1083     */
1084    if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1085        locked_vma = (vma->vm_flags & VM_LOCKED);
1086        if (!locked_vma)
1087            up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1088    }
1089
1090    pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1091
1092    /* Update high watermark before we lower rss */
1093    update_hiwater_rss(mm);
1094
1095    for (; address < end; pte++, address += PAGE_SIZE) {
1096        if (!pte_present(*pte))
1097            continue;
1098        page = vm_normal_page(vma, address, *pte);
1099        BUG_ON(!page || PageAnon(page));
1100
1101        if (locked_vma) {
1102            mlock_vma_page(page); /* no-op if already mlocked */
1103            if (page == check_page)
1104                ret = SWAP_MLOCK;
1105            continue; /* don't unmap */
1106        }
1107
1108        if (ptep_clear_flush_young_notify(vma, address, pte))
1109            continue;
1110
1111        /* Nuke the page table entry. */
1112        flush_cache_page(vma, address, pte_pfn(*pte));
1113        pteval = ptep_clear_flush_notify(vma, address, pte);
1114
1115        /* If nonlinear, store the file page offset in the pte. */
1116        if (page->index != linear_page_index(vma, address))
1117            set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1118
1119        /* Move the dirty bit to the physical page now the pte is gone. */
1120        if (pte_dirty(pteval))
1121            set_page_dirty(page);
1122
1123        page_remove_rmap(page);
1124        page_cache_release(page);
1125        dec_mm_counter(mm, MM_FILEPAGES);
1126        (*mapcount)--;
1127    }
1128    pte_unmap_unlock(pte - 1, ptl);
1129    if (locked_vma)
1130        up_read(&vma->vm_mm->mmap_sem);
1131    return ret;
1132}
1133
1134/**
1135 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1136 * rmap method
1137 * @page: the page to unmap/unlock
1138 * @flags: action and flags
1139 *
1140 * Find all the mappings of a page using the mapping pointer and the vma chains
1141 * contained in the anon_vma struct it points to.
1142 *
1143 * This function is only called from try_to_unmap/try_to_munlock for
1144 * anonymous pages.
1145 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1146 * where the page was found will be held for write. So, we won't recheck
1147 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1148 * 'LOCKED.
1149 */
1150static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1151{
1152    struct anon_vma *anon_vma;
1153    struct anon_vma_chain *avc;
1154    int ret = SWAP_AGAIN;
1155
1156    anon_vma = page_lock_anon_vma(page);
1157    if (!anon_vma)
1158        return ret;
1159
1160    list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1161        struct vm_area_struct *vma = avc->vma;
1162        unsigned long address = vma_address(page, vma);
1163        if (address == -EFAULT)
1164            continue;
1165        ret = try_to_unmap_one(page, vma, address, flags);
1166        if (ret != SWAP_AGAIN || !page_mapped(page))
1167            break;
1168    }
1169
1170    page_unlock_anon_vma(anon_vma);
1171    return ret;
1172}
1173
1174/**
1175 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1176 * @page: the page to unmap/unlock
1177 * @flags: action and flags
1178 *
1179 * Find all the mappings of a page using the mapping pointer and the vma chains
1180 * contained in the address_space struct it points to.
1181 *
1182 * This function is only called from try_to_unmap/try_to_munlock for
1183 * object-based pages.
1184 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1185 * where the page was found will be held for write. So, we won't recheck
1186 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1187 * 'LOCKED.
1188 */
1189static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1190{
1191    struct address_space *mapping = page->mapping;
1192    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1193    struct vm_area_struct *vma;
1194    struct prio_tree_iter iter;
1195    int ret = SWAP_AGAIN;
1196    unsigned long cursor;
1197    unsigned long max_nl_cursor = 0;
1198    unsigned long max_nl_size = 0;
1199    unsigned int mapcount;
1200
1201    spin_lock(&mapping->i_mmap_lock);
1202    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1203        unsigned long address = vma_address(page, vma);
1204        if (address == -EFAULT)
1205            continue;
1206        ret = try_to_unmap_one(page, vma, address, flags);
1207        if (ret != SWAP_AGAIN || !page_mapped(page))
1208            goto out;
1209    }
1210
1211    if (list_empty(&mapping->i_mmap_nonlinear))
1212        goto out;
1213
1214    /*
1215     * We don't bother to try to find the munlocked page in nonlinears.
1216     * It's costly. Instead, later, page reclaim logic may call
1217     * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1218     */
1219    if (TTU_ACTION(flags) == TTU_MUNLOCK)
1220        goto out;
1221
1222    list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1223                        shared.vm_set.list) {
1224        cursor = (unsigned long) vma->vm_private_data;
1225        if (cursor > max_nl_cursor)
1226            max_nl_cursor = cursor;
1227        cursor = vma->vm_end - vma->vm_start;
1228        if (cursor > max_nl_size)
1229            max_nl_size = cursor;
1230    }
1231
1232    if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1233        ret = SWAP_FAIL;
1234        goto out;
1235    }
1236
1237    /*
1238     * We don't try to search for this page in the nonlinear vmas,
1239     * and page_referenced wouldn't have found it anyway. Instead
1240     * just walk the nonlinear vmas trying to age and unmap some.
1241     * The mapcount of the page we came in with is irrelevant,
1242     * but even so use it as a guide to how hard we should try?
1243     */
1244    mapcount = page_mapcount(page);
1245    if (!mapcount)
1246        goto out;
1247    cond_resched_lock(&mapping->i_mmap_lock);
1248
1249    max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1250    if (max_nl_cursor == 0)
1251        max_nl_cursor = CLUSTER_SIZE;
1252
1253    do {
1254        list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1255                        shared.vm_set.list) {
1256            cursor = (unsigned long) vma->vm_private_data;
1257            while ( cursor < max_nl_cursor &&
1258                cursor < vma->vm_end - vma->vm_start) {
1259                if (try_to_unmap_cluster(cursor, &mapcount,
1260                        vma, page) == SWAP_MLOCK)
1261                    ret = SWAP_MLOCK;
1262                cursor += CLUSTER_SIZE;
1263                vma->vm_private_data = (void *) cursor;
1264                if ((int)mapcount <= 0)
1265                    goto out;
1266            }
1267            vma->vm_private_data = (void *) max_nl_cursor;
1268        }
1269        cond_resched_lock(&mapping->i_mmap_lock);
1270        max_nl_cursor += CLUSTER_SIZE;
1271    } while (max_nl_cursor <= max_nl_size);
1272
1273    /*
1274     * Don't loop forever (perhaps all the remaining pages are
1275     * in locked vmas). Reset cursor on all unreserved nonlinear
1276     * vmas, now forgetting on which ones it had fallen behind.
1277     */
1278    list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1279        vma->vm_private_data = NULL;
1280out:
1281    spin_unlock(&mapping->i_mmap_lock);
1282    return ret;
1283}
1284
1285/**
1286 * try_to_unmap - try to remove all page table mappings to a page
1287 * @page: the page to get unmapped
1288 * @flags: action and flags
1289 *
1290 * Tries to remove all the page table entries which are mapping this
1291 * page, used in the pageout path. Caller must hold the page lock.
1292 * Return values are:
1293 *
1294 * SWAP_SUCCESS - we succeeded in removing all mappings
1295 * SWAP_AGAIN - we missed a mapping, try again later
1296 * SWAP_FAIL - the page is unswappable
1297 * SWAP_MLOCK - page is mlocked.
1298 */
1299int try_to_unmap(struct page *page, enum ttu_flags flags)
1300{
1301    int ret;
1302
1303    BUG_ON(!PageLocked(page));
1304
1305    if (unlikely(PageKsm(page)))
1306        ret = try_to_unmap_ksm(page, flags);
1307    else if (PageAnon(page))
1308        ret = try_to_unmap_anon(page, flags);
1309    else
1310        ret = try_to_unmap_file(page, flags);
1311    if (ret != SWAP_MLOCK && !page_mapped(page))
1312        ret = SWAP_SUCCESS;
1313    return ret;
1314}
1315
1316/**
1317 * try_to_munlock - try to munlock a page
1318 * @page: the page to be munlocked
1319 *
1320 * Called from munlock code. Checks all of the VMAs mapping the page
1321 * to make sure nobody else has this page mlocked. The page will be
1322 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1323 *
1324 * Return values are:
1325 *
1326 * SWAP_AGAIN - no vma is holding page mlocked, or,
1327 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1328 * SWAP_FAIL - page cannot be located at present
1329 * SWAP_MLOCK - page is now mlocked.
1330 */
1331int try_to_munlock(struct page *page)
1332{
1333    VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1334
1335    if (unlikely(PageKsm(page)))
1336        return try_to_unmap_ksm(page, TTU_MUNLOCK);
1337    else if (PageAnon(page))
1338        return try_to_unmap_anon(page, TTU_MUNLOCK);
1339    else
1340        return try_to_unmap_file(page, TTU_MUNLOCK);
1341}
1342
1343#ifdef CONFIG_MIGRATION
1344/*
1345 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1346 * Called by migrate.c to remove migration ptes, but might be used more later.
1347 */
1348static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1349        struct vm_area_struct *, unsigned long, void *), void *arg)
1350{
1351    struct anon_vma *anon_vma;
1352    struct anon_vma_chain *avc;
1353    int ret = SWAP_AGAIN;
1354
1355    /*
1356     * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1357     * because that depends on page_mapped(); but not all its usages
1358     * are holding mmap_sem, which also gave the necessary guarantee
1359     * (that this anon_vma's slab has not already been destroyed).
1360     * This needs to be reviewed later: avoiding page_lock_anon_vma()
1361     * is risky, and currently limits the usefulness of rmap_walk().
1362     */
1363    anon_vma = page_anon_vma(page);
1364    if (!anon_vma)
1365        return ret;
1366    spin_lock(&anon_vma->lock);
1367    list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1368        struct vm_area_struct *vma = avc->vma;
1369        unsigned long address = vma_address(page, vma);
1370        if (address == -EFAULT)
1371            continue;
1372        ret = rmap_one(page, vma, address, arg);
1373        if (ret != SWAP_AGAIN)
1374            break;
1375    }
1376    spin_unlock(&anon_vma->lock);
1377    return ret;
1378}
1379
1380static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1381        struct vm_area_struct *, unsigned long, void *), void *arg)
1382{
1383    struct address_space *mapping = page->mapping;
1384    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1385    struct vm_area_struct *vma;
1386    struct prio_tree_iter iter;
1387    int ret = SWAP_AGAIN;
1388
1389    if (!mapping)
1390        return ret;
1391    spin_lock(&mapping->i_mmap_lock);
1392    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1393        unsigned long address = vma_address(page, vma);
1394        if (address == -EFAULT)
1395            continue;
1396        ret = rmap_one(page, vma, address, arg);
1397        if (ret != SWAP_AGAIN)
1398            break;
1399    }
1400    /*
1401     * No nonlinear handling: being always shared, nonlinear vmas
1402     * never contain migration ptes. Decide what to do about this
1403     * limitation to linear when we need rmap_walk() on nonlinear.
1404     */
1405    spin_unlock(&mapping->i_mmap_lock);
1406    return ret;
1407}
1408
1409int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1410        struct vm_area_struct *, unsigned long, void *), void *arg)
1411{
1412    VM_BUG_ON(!PageLocked(page));
1413
1414    if (unlikely(PageKsm(page)))
1415        return rmap_walk_ksm(page, rmap_one, arg);
1416    else if (PageAnon(page))
1417        return rmap_walk_anon(page, rmap_one, arg);
1418    else
1419        return rmap_walk_file(page, rmap_one, arg);
1420}
1421#endif /* CONFIG_MIGRATION */
1422

Archive Download this file



interactive