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#include <linux/hugetlb.h>
60
61#include <asm/tlbflush.h>
62
63#include "internal.h"
64
65static struct kmem_cache *anon_vma_cachep;
66static struct kmem_cache *anon_vma_chain_cachep;
67
68static inline struct anon_vma *anon_vma_alloc(void)
69{
70    return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
71}
72
73void anon_vma_free(struct anon_vma *anon_vma)
74{
75    kmem_cache_free(anon_vma_cachep, anon_vma);
76}
77
78static inline struct anon_vma_chain *anon_vma_chain_alloc(void)
79{
80    return kmem_cache_alloc(anon_vma_chain_cachep, GFP_KERNEL);
81}
82
83void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
84{
85    kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
86}
87
88/**
89 * anon_vma_prepare - attach an anon_vma to a memory region
90 * @vma: the memory region in question
91 *
92 * This makes sure the memory mapping described by 'vma' has
93 * an 'anon_vma' attached to it, so that we can associate the
94 * anonymous pages mapped into it with that anon_vma.
95 *
96 * The common case will be that we already have one, but if
97 * if not we either need to find an adjacent mapping that we
98 * can re-use the anon_vma from (very common when the only
99 * reason for splitting a vma has been mprotect()), or we
100 * allocate a new one.
101 *
102 * Anon-vma allocations are very subtle, because we may have
103 * optimistically looked up an anon_vma in page_lock_anon_vma()
104 * and that may actually touch the spinlock even in the newly
105 * allocated vma (it depends on RCU to make sure that the
106 * anon_vma isn't actually destroyed).
107 *
108 * As a result, we need to do proper anon_vma locking even
109 * for the new allocation. At the same time, we do not want
110 * to do any locking for the common case of already having
111 * an anon_vma.
112 *
113 * This must be called with the mmap_sem held for reading.
114 */
115int anon_vma_prepare(struct vm_area_struct *vma)
116{
117    struct anon_vma *anon_vma = vma->anon_vma;
118    struct anon_vma_chain *avc;
119
120    might_sleep();
121    if (unlikely(!anon_vma)) {
122        struct mm_struct *mm = vma->vm_mm;
123        struct anon_vma *allocated;
124
125        avc = anon_vma_chain_alloc();
126        if (!avc)
127            goto out_enomem;
128
129        anon_vma = find_mergeable_anon_vma(vma);
130        allocated = NULL;
131        if (!anon_vma) {
132            anon_vma = anon_vma_alloc();
133            if (unlikely(!anon_vma))
134                goto out_enomem_free_avc;
135            allocated = anon_vma;
136            /*
137             * This VMA had no anon_vma yet. This anon_vma is
138             * the root of any anon_vma tree that might form.
139             */
140            anon_vma->root = anon_vma;
141        }
142
143        anon_vma_lock(anon_vma);
144        /* page_table_lock to protect against threads */
145        spin_lock(&mm->page_table_lock);
146        if (likely(!vma->anon_vma)) {
147            vma->anon_vma = anon_vma;
148            avc->anon_vma = anon_vma;
149            avc->vma = vma;
150            list_add(&avc->same_vma, &vma->anon_vma_chain);
151            list_add_tail(&avc->same_anon_vma, &anon_vma->head);
152            allocated = NULL;
153            avc = NULL;
154        }
155        spin_unlock(&mm->page_table_lock);
156        anon_vma_unlock(anon_vma);
157
158        if (unlikely(allocated))
159            anon_vma_free(allocated);
160        if (unlikely(avc))
161            anon_vma_chain_free(avc);
162    }
163    return 0;
164
165 out_enomem_free_avc:
166    anon_vma_chain_free(avc);
167 out_enomem:
168    return -ENOMEM;
169}
170
171static void anon_vma_chain_link(struct vm_area_struct *vma,
172                struct anon_vma_chain *avc,
173                struct anon_vma *anon_vma)
174{
175    avc->vma = vma;
176    avc->anon_vma = anon_vma;
177    list_add(&avc->same_vma, &vma->anon_vma_chain);
178
179    anon_vma_lock(anon_vma);
180    list_add_tail(&avc->same_anon_vma, &anon_vma->head);
181    anon_vma_unlock(anon_vma);
182}
183
184/*
185 * Attach the anon_vmas from src to dst.
186 * Returns 0 on success, -ENOMEM on failure.
187 */
188int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
189{
190    struct anon_vma_chain *avc, *pavc;
191
192    list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
193        avc = anon_vma_chain_alloc();
194        if (!avc)
195            goto enomem_failure;
196        anon_vma_chain_link(dst, avc, pavc->anon_vma);
197    }
198    return 0;
199
200 enomem_failure:
201    unlink_anon_vmas(dst);
202    return -ENOMEM;
203}
204
205/*
206 * Attach vma to its own anon_vma, as well as to the anon_vmas that
207 * the corresponding VMA in the parent process is attached to.
208 * Returns 0 on success, non-zero on failure.
209 */
210int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
211{
212    struct anon_vma_chain *avc;
213    struct anon_vma *anon_vma;
214
215    /* Don't bother if the parent process has no anon_vma here. */
216    if (!pvma->anon_vma)
217        return 0;
218
219    /*
220     * First, attach the new VMA to the parent VMA's anon_vmas,
221     * so rmap can find non-COWed pages in child processes.
222     */
223    if (anon_vma_clone(vma, pvma))
224        return -ENOMEM;
225
226    /* Then add our own anon_vma. */
227    anon_vma = anon_vma_alloc();
228    if (!anon_vma)
229        goto out_error;
230    avc = anon_vma_chain_alloc();
231    if (!avc)
232        goto out_error_free_anon_vma;
233
234    /*
235     * The root anon_vma's spinlock is the lock actually used when we
236     * lock any of the anon_vmas in this anon_vma tree.
237     */
238    anon_vma->root = pvma->anon_vma->root;
239    /*
240     * With KSM refcounts, an anon_vma can stay around longer than the
241     * process it belongs to. The root anon_vma needs to be pinned
242     * until this anon_vma is freed, because the lock lives in the root.
243     */
244    get_anon_vma(anon_vma->root);
245    /* Mark this anon_vma as the one where our new (COWed) pages go. */
246    vma->anon_vma = anon_vma;
247    anon_vma_chain_link(vma, avc, anon_vma);
248
249    return 0;
250
251 out_error_free_anon_vma:
252    anon_vma_free(anon_vma);
253 out_error:
254    unlink_anon_vmas(vma);
255    return -ENOMEM;
256}
257
258static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain)
259{
260    struct anon_vma *anon_vma = anon_vma_chain->anon_vma;
261    int empty;
262
263    /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */
264    if (!anon_vma)
265        return;
266
267    anon_vma_lock(anon_vma);
268    list_del(&anon_vma_chain->same_anon_vma);
269
270    /* We must garbage collect the anon_vma if it's empty */
271    empty = list_empty(&anon_vma->head) && !anonvma_external_refcount(anon_vma);
272    anon_vma_unlock(anon_vma);
273
274    if (empty) {
275        /* We no longer need the root anon_vma */
276        if (anon_vma->root != anon_vma)
277            drop_anon_vma(anon_vma->root);
278        anon_vma_free(anon_vma);
279    }
280}
281
282void unlink_anon_vmas(struct vm_area_struct *vma)
283{
284    struct anon_vma_chain *avc, *next;
285
286    /*
287     * Unlink each anon_vma chained to the VMA. This list is ordered
288     * from newest to oldest, ensuring the root anon_vma gets freed last.
289     */
290    list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
291        anon_vma_unlink(avc);
292        list_del(&avc->same_vma);
293        anon_vma_chain_free(avc);
294    }
295}
296
297static void anon_vma_ctor(void *data)
298{
299    struct anon_vma *anon_vma = data;
300
301    spin_lock_init(&anon_vma->lock);
302    anonvma_external_refcount_init(anon_vma);
303    INIT_LIST_HEAD(&anon_vma->head);
304}
305
306void __init anon_vma_init(void)
307{
308    anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
309            0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
310    anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
311}
312
313/*
314 * Getting a lock on a stable anon_vma from a page off the LRU is
315 * tricky: page_lock_anon_vma rely on RCU to guard against the races.
316 */
317struct anon_vma *page_lock_anon_vma(struct page *page)
318{
319    struct anon_vma *anon_vma, *root_anon_vma;
320    unsigned long anon_mapping;
321
322    rcu_read_lock();
323    anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
324    if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
325        goto out;
326    if (!page_mapped(page))
327        goto out;
328
329    anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
330    root_anon_vma = ACCESS_ONCE(anon_vma->root);
331    spin_lock(&root_anon_vma->lock);
332
333    /*
334     * If this page is still mapped, then its anon_vma cannot have been
335     * freed. But if it has been unmapped, we have no security against
336     * the anon_vma structure being freed and reused (for another anon_vma:
337     * SLAB_DESTROY_BY_RCU guarantees that - so the spin_lock above cannot
338     * corrupt): with anon_vma_prepare() or anon_vma_fork() redirecting
339     * anon_vma->root before page_unlock_anon_vma() is called to unlock.
340     */
341    if (page_mapped(page))
342        return anon_vma;
343
344    spin_unlock(&root_anon_vma->lock);
345out:
346    rcu_read_unlock();
347    return NULL;
348}
349
350void page_unlock_anon_vma(struct anon_vma *anon_vma)
351{
352    anon_vma_unlock(anon_vma);
353    rcu_read_unlock();
354}
355
356/*
357 * At what user virtual address is page expected in @vma?
358 * Returns virtual address or -EFAULT if page's index/offset is not
359 * within the range mapped the @vma.
360 */
361static inline unsigned long
362vma_address(struct page *page, struct vm_area_struct *vma)
363{
364    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
365    unsigned long address;
366
367    if (unlikely(is_vm_hugetlb_page(vma)))
368        pgoff = page->index << huge_page_order(page_hstate(page));
369    address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
370    if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
371        /* page should be within @vma mapping range */
372        return -EFAULT;
373    }
374    return address;
375}
376
377/*
378 * At what user virtual address is page expected in vma?
379 * Caller should check the page is actually part of the vma.
380 */
381unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
382{
383    if (PageAnon(page)) {
384        struct anon_vma *page__anon_vma = page_anon_vma(page);
385        /*
386         * Note: swapoff's unuse_vma() is more efficient with this
387         * check, and needs it to match anon_vma when KSM is active.
388         */
389        if (!vma->anon_vma || !page__anon_vma ||
390            vma->anon_vma->root != page__anon_vma->root)
391            return -EFAULT;
392    } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
393        if (!vma->vm_file ||
394            vma->vm_file->f_mapping != page->mapping)
395            return -EFAULT;
396    } else
397        return -EFAULT;
398    return vma_address(page, vma);
399}
400
401/*
402 * Check that @page is mapped at @address into @mm.
403 *
404 * If @sync is false, page_check_address may perform a racy check to avoid
405 * the page table lock when the pte is not present (helpful when reclaiming
406 * highly shared pages).
407 *
408 * On success returns with pte mapped and locked.
409 */
410pte_t *page_check_address(struct page *page, struct mm_struct *mm,
411              unsigned long address, spinlock_t **ptlp, int sync)
412{
413    pgd_t *pgd;
414    pud_t *pud;
415    pmd_t *pmd;
416    pte_t *pte;
417    spinlock_t *ptl;
418
419    if (unlikely(PageHuge(page))) {
420        pte = huge_pte_offset(mm, address);
421        ptl = &mm->page_table_lock;
422        goto check;
423    }
424
425    pgd = pgd_offset(mm, address);
426    if (!pgd_present(*pgd))
427        return NULL;
428
429    pud = pud_offset(pgd, address);
430    if (!pud_present(*pud))
431        return NULL;
432
433    pmd = pmd_offset(pud, address);
434    if (!pmd_present(*pmd))
435        return NULL;
436
437    pte = pte_offset_map(pmd, address);
438    /* Make a quick check before getting the lock */
439    if (!sync && !pte_present(*pte)) {
440        pte_unmap(pte);
441        return NULL;
442    }
443
444    ptl = pte_lockptr(mm, pmd);
445check:
446    spin_lock(ptl);
447    if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
448        *ptlp = ptl;
449        return pte;
450    }
451    pte_unmap_unlock(pte, ptl);
452    return NULL;
453}
454
455/**
456 * page_mapped_in_vma - check whether a page is really mapped in a VMA
457 * @page: the page to test
458 * @vma: the VMA to test
459 *
460 * Returns 1 if the page is mapped into the page tables of the VMA, 0
461 * if the page is not mapped into the page tables of this VMA. Only
462 * valid for normal file or anonymous VMAs.
463 */
464int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
465{
466    unsigned long address;
467    pte_t *pte;
468    spinlock_t *ptl;
469
470    address = vma_address(page, vma);
471    if (address == -EFAULT) /* out of vma range */
472        return 0;
473    pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
474    if (!pte) /* the page is not in this mm */
475        return 0;
476    pte_unmap_unlock(pte, ptl);
477
478    return 1;
479}
480
481/*
482 * Subfunctions of page_referenced: page_referenced_one called
483 * repeatedly from either page_referenced_anon or page_referenced_file.
484 */
485int page_referenced_one(struct page *page, struct vm_area_struct *vma,
486            unsigned long address, unsigned int *mapcount,
487            unsigned long *vm_flags)
488{
489    struct mm_struct *mm = vma->vm_mm;
490    pte_t *pte;
491    spinlock_t *ptl;
492    int referenced = 0;
493
494    pte = page_check_address(page, mm, address, &ptl, 0);
495    if (!pte)
496        goto out;
497
498    /*
499     * Don't want to elevate referenced for mlocked page that gets this far,
500     * in order that it progresses to try_to_unmap and is moved to the
501     * unevictable list.
502     */
503    if (vma->vm_flags & VM_LOCKED) {
504        *mapcount = 1; /* break early from loop */
505        *vm_flags |= VM_LOCKED;
506        goto out_unmap;
507    }
508
509    if (ptep_clear_flush_young_notify(vma, address, pte)) {
510        /*
511         * Don't treat a reference through a sequentially read
512         * mapping as such. If the page has been used in
513         * another mapping, we will catch it; if this other
514         * mapping is already gone, the unmap path will have
515         * set PG_referenced or activated the page.
516         */
517        if (likely(!VM_SequentialReadHint(vma)))
518            referenced++;
519    }
520
521    /* Pretend the page is referenced if the task has the
522       swap token and is in the middle of a page fault. */
523    if (mm != current->mm && has_swap_token(mm) &&
524            rwsem_is_locked(&mm->mmap_sem))
525        referenced++;
526
527out_unmap:
528    (*mapcount)--;
529    pte_unmap_unlock(pte, ptl);
530
531    if (referenced)
532        *vm_flags |= vma->vm_flags;
533out:
534    return referenced;
535}
536
537static int page_referenced_anon(struct page *page,
538                struct mem_cgroup *mem_cont,
539                unsigned long *vm_flags)
540{
541    unsigned int mapcount;
542    struct anon_vma *anon_vma;
543    struct anon_vma_chain *avc;
544    int referenced = 0;
545
546    anon_vma = page_lock_anon_vma(page);
547    if (!anon_vma)
548        return referenced;
549
550    mapcount = page_mapcount(page);
551    list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
552        struct vm_area_struct *vma = avc->vma;
553        unsigned long address = vma_address(page, vma);
554        if (address == -EFAULT)
555            continue;
556        /*
557         * If we are reclaiming on behalf of a cgroup, skip
558         * counting on behalf of references from different
559         * cgroups
560         */
561        if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
562            continue;
563        referenced += page_referenced_one(page, vma, address,
564                          &mapcount, vm_flags);
565        if (!mapcount)
566            break;
567    }
568
569    page_unlock_anon_vma(anon_vma);
570    return referenced;
571}
572
573/**
574 * page_referenced_file - referenced check for object-based rmap
575 * @page: the page we're checking references on.
576 * @mem_cont: target memory controller
577 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
578 *
579 * For an object-based mapped page, find all the places it is mapped and
580 * check/clear the referenced flag. This is done by following the page->mapping
581 * pointer, then walking the chain of vmas it holds. It returns the number
582 * of references it found.
583 *
584 * This function is only called from page_referenced for object-based pages.
585 */
586static int page_referenced_file(struct page *page,
587                struct mem_cgroup *mem_cont,
588                unsigned long *vm_flags)
589{
590    unsigned int mapcount;
591    struct address_space *mapping = page->mapping;
592    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
593    struct vm_area_struct *vma;
594    struct prio_tree_iter iter;
595    int referenced = 0;
596
597    /*
598     * The caller's checks on page->mapping and !PageAnon have made
599     * sure that this is a file page: the check for page->mapping
600     * excludes the case just before it gets set on an anon page.
601     */
602    BUG_ON(PageAnon(page));
603
604    /*
605     * The page lock not only makes sure that page->mapping cannot
606     * suddenly be NULLified by truncation, it makes sure that the
607     * structure at mapping cannot be freed and reused yet,
608     * so we can safely take mapping->i_mmap_lock.
609     */
610    BUG_ON(!PageLocked(page));
611
612    spin_lock(&mapping->i_mmap_lock);
613
614    /*
615     * i_mmap_lock does not stabilize mapcount at all, but mapcount
616     * is more likely to be accurate if we note it after spinning.
617     */
618    mapcount = page_mapcount(page);
619
620    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
621        unsigned long address = vma_address(page, vma);
622        if (address == -EFAULT)
623            continue;
624        /*
625         * If we are reclaiming on behalf of a cgroup, skip
626         * counting on behalf of references from different
627         * cgroups
628         */
629        if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
630            continue;
631        referenced += page_referenced_one(page, vma, address,
632                          &mapcount, vm_flags);
633        if (!mapcount)
634            break;
635    }
636
637    spin_unlock(&mapping->i_mmap_lock);
638    return referenced;
639}
640
641/**
642 * page_referenced - test if the page was referenced
643 * @page: the page to test
644 * @is_locked: caller holds lock on the page
645 * @mem_cont: target memory controller
646 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
647 *
648 * Quick test_and_clear_referenced for all mappings to a page,
649 * returns the number of ptes which referenced the page.
650 */
651int page_referenced(struct page *page,
652            int is_locked,
653            struct mem_cgroup *mem_cont,
654            unsigned long *vm_flags)
655{
656    int referenced = 0;
657    int we_locked = 0;
658
659    *vm_flags = 0;
660    if (page_mapped(page) && page_rmapping(page)) {
661        if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
662            we_locked = trylock_page(page);
663            if (!we_locked) {
664                referenced++;
665                goto out;
666            }
667        }
668        if (unlikely(PageKsm(page)))
669            referenced += page_referenced_ksm(page, mem_cont,
670                                vm_flags);
671        else if (PageAnon(page))
672            referenced += page_referenced_anon(page, mem_cont,
673                                vm_flags);
674        else if (page->mapping)
675            referenced += page_referenced_file(page, mem_cont,
676                                vm_flags);
677        if (we_locked)
678            unlock_page(page);
679    }
680out:
681    if (page_test_and_clear_young(page))
682        referenced++;
683
684    return referenced;
685}
686
687static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
688                unsigned long address)
689{
690    struct mm_struct *mm = vma->vm_mm;
691    pte_t *pte;
692    spinlock_t *ptl;
693    int ret = 0;
694
695    pte = page_check_address(page, mm, address, &ptl, 1);
696    if (!pte)
697        goto out;
698
699    if (pte_dirty(*pte) || pte_write(*pte)) {
700        pte_t entry;
701
702        flush_cache_page(vma, address, pte_pfn(*pte));
703        entry = ptep_clear_flush_notify(vma, address, pte);
704        entry = pte_wrprotect(entry);
705        entry = pte_mkclean(entry);
706        set_pte_at(mm, address, pte, entry);
707        ret = 1;
708    }
709
710    pte_unmap_unlock(pte, ptl);
711out:
712    return ret;
713}
714
715static int page_mkclean_file(struct address_space *mapping, struct page *page)
716{
717    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
718    struct vm_area_struct *vma;
719    struct prio_tree_iter iter;
720    int ret = 0;
721
722    BUG_ON(PageAnon(page));
723
724    spin_lock(&mapping->i_mmap_lock);
725    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
726        if (vma->vm_flags & VM_SHARED) {
727            unsigned long address = vma_address(page, vma);
728            if (address == -EFAULT)
729                continue;
730            ret += page_mkclean_one(page, vma, address);
731        }
732    }
733    spin_unlock(&mapping->i_mmap_lock);
734    return ret;
735}
736
737int page_mkclean(struct page *page)
738{
739    int ret = 0;
740
741    BUG_ON(!PageLocked(page));
742
743    if (page_mapped(page)) {
744        struct address_space *mapping = page_mapping(page);
745        if (mapping) {
746            ret = page_mkclean_file(mapping, page);
747            if (page_test_dirty(page)) {
748                page_clear_dirty(page);
749                ret = 1;
750            }
751        }
752    }
753
754    return ret;
755}
756EXPORT_SYMBOL_GPL(page_mkclean);
757
758/**
759 * page_move_anon_rmap - move a page to our anon_vma
760 * @page: the page to move to our anon_vma
761 * @vma: the vma the page belongs to
762 * @address: the user virtual address mapped
763 *
764 * When a page belongs exclusively to one process after a COW event,
765 * that page can be moved into the anon_vma that belongs to just that
766 * process, so the rmap code will not search the parent or sibling
767 * processes.
768 */
769void page_move_anon_rmap(struct page *page,
770    struct vm_area_struct *vma, unsigned long address)
771{
772    struct anon_vma *anon_vma = vma->anon_vma;
773
774    VM_BUG_ON(!PageLocked(page));
775    VM_BUG_ON(!anon_vma);
776    VM_BUG_ON(page->index != linear_page_index(vma, address));
777
778    anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
779    page->mapping = (struct address_space *) anon_vma;
780}
781
782/**
783 * __page_set_anon_rmap - setup new anonymous rmap
784 * @page: the page to add the mapping to
785 * @vma: the vm area in which the mapping is added
786 * @address: the user virtual address mapped
787 * @exclusive: the page is exclusively owned by the current process
788 */
789static void __page_set_anon_rmap(struct page *page,
790    struct vm_area_struct *vma, unsigned long address, int exclusive)
791{
792    struct anon_vma *anon_vma = vma->anon_vma;
793
794    BUG_ON(!anon_vma);
795
796    /*
797     * If the page isn't exclusively mapped into this vma,
798     * we must use the _oldest_ possible anon_vma for the
799     * page mapping!
800     */
801    if (!exclusive) {
802        if (PageAnon(page))
803            return;
804        anon_vma = anon_vma->root;
805    } else {
806        /*
807         * In this case, swapped-out-but-not-discarded swap-cache
808         * is remapped. So, no need to update page->mapping here.
809         * We convice anon_vma poitned by page->mapping is not obsolete
810         * because vma->anon_vma is necessary to be a family of it.
811         */
812        if (PageAnon(page))
813            return;
814    }
815
816    anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
817    page->mapping = (struct address_space *) anon_vma;
818    page->index = linear_page_index(vma, address);
819}
820
821/**
822 * __page_check_anon_rmap - sanity check anonymous rmap addition
823 * @page: the page to add the mapping to
824 * @vma: the vm area in which the mapping is added
825 * @address: the user virtual address mapped
826 */
827static void __page_check_anon_rmap(struct page *page,
828    struct vm_area_struct *vma, unsigned long address)
829{
830#ifdef CONFIG_DEBUG_VM
831    /*
832     * The page's anon-rmap details (mapping and index) are guaranteed to
833     * be set up correctly at this point.
834     *
835     * We have exclusion against page_add_anon_rmap because the caller
836     * always holds the page locked, except if called from page_dup_rmap,
837     * in which case the page is already known to be setup.
838     *
839     * We have exclusion against page_add_new_anon_rmap because those pages
840     * are initially only visible via the pagetables, and the pte is locked
841     * over the call to page_add_new_anon_rmap.
842     */
843    BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
844    BUG_ON(page->index != linear_page_index(vma, address));
845#endif
846}
847
848/**
849 * page_add_anon_rmap - add pte mapping to an anonymous page
850 * @page: the page to add the mapping to
851 * @vma: the vm area in which the mapping is added
852 * @address: the user virtual address mapped
853 *
854 * The caller needs to hold the pte lock, and the page must be locked in
855 * the anon_vma case: to serialize mapping,index checking after setting,
856 * and to ensure that PageAnon is not being upgraded racily to PageKsm
857 * (but PageKsm is never downgraded to PageAnon).
858 */
859void page_add_anon_rmap(struct page *page,
860    struct vm_area_struct *vma, unsigned long address)
861{
862    do_page_add_anon_rmap(page, vma, address, 0);
863}
864
865/*
866 * Special version of the above for do_swap_page, which often runs
867 * into pages that are exclusively owned by the current process.
868 * Everybody else should continue to use page_add_anon_rmap above.
869 */
870void do_page_add_anon_rmap(struct page *page,
871    struct vm_area_struct *vma, unsigned long address, int exclusive)
872{
873    int first = atomic_inc_and_test(&page->_mapcount);
874    if (first)
875        __inc_zone_page_state(page, NR_ANON_PAGES);
876    if (unlikely(PageKsm(page)))
877        return;
878
879    VM_BUG_ON(!PageLocked(page));
880    VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
881    if (first)
882        __page_set_anon_rmap(page, vma, address, exclusive);
883    else
884        __page_check_anon_rmap(page, vma, address);
885}
886
887/**
888 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
889 * @page: the page to add the mapping to
890 * @vma: the vm area in which the mapping is added
891 * @address: the user virtual address mapped
892 *
893 * Same as page_add_anon_rmap but must only be called on *new* pages.
894 * This means the inc-and-test can be bypassed.
895 * Page does not have to be locked.
896 */
897void page_add_new_anon_rmap(struct page *page,
898    struct vm_area_struct *vma, unsigned long address)
899{
900    VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
901    SetPageSwapBacked(page);
902    atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
903    __inc_zone_page_state(page, NR_ANON_PAGES);
904    __page_set_anon_rmap(page, vma, address, 1);
905    if (page_evictable(page, vma))
906        lru_cache_add_lru(page, LRU_ACTIVE_ANON);
907    else
908        add_page_to_unevictable_list(page);
909}
910
911/**
912 * page_add_file_rmap - add pte mapping to a file page
913 * @page: the page to add the mapping to
914 *
915 * The caller needs to hold the pte lock.
916 */
917void page_add_file_rmap(struct page *page)
918{
919    if (atomic_inc_and_test(&page->_mapcount)) {
920        __inc_zone_page_state(page, NR_FILE_MAPPED);
921        mem_cgroup_update_file_mapped(page, 1);
922    }
923}
924
925/**
926 * page_remove_rmap - take down pte mapping from a page
927 * @page: page to remove mapping from
928 *
929 * The caller needs to hold the pte lock.
930 */
931void page_remove_rmap(struct page *page)
932{
933    /* page still mapped by someone else? */
934    if (!atomic_add_negative(-1, &page->_mapcount))
935        return;
936
937    /*
938     * Now that the last pte has gone, s390 must transfer dirty
939     * flag from storage key to struct page. We can usually skip
940     * this if the page is anon, so about to be freed; but perhaps
941     * not if it's in swapcache - there might be another pte slot
942     * containing the swap entry, but page not yet written to swap.
943     */
944    if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) {
945        page_clear_dirty(page);
946        set_page_dirty(page);
947    }
948    /*
949     * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
950     * and not charged by memcg for now.
951     */
952    if (unlikely(PageHuge(page)))
953        return;
954    if (PageAnon(page)) {
955        mem_cgroup_uncharge_page(page);
956        __dec_zone_page_state(page, NR_ANON_PAGES);
957    } else {
958        __dec_zone_page_state(page, NR_FILE_MAPPED);
959        mem_cgroup_update_file_mapped(page, -1);
960    }
961    /*
962     * It would be tidy to reset the PageAnon mapping here,
963     * but that might overwrite a racing page_add_anon_rmap
964     * which increments mapcount after us but sets mapping
965     * before us: so leave the reset to free_hot_cold_page,
966     * and remember that it's only reliable while mapped.
967     * Leaving it set also helps swapoff to reinstate ptes
968     * faster for those pages still in swapcache.
969     */
970}
971
972/*
973 * Subfunctions of try_to_unmap: try_to_unmap_one called
974 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
975 */
976int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
977             unsigned long address, enum ttu_flags flags)
978{
979    struct mm_struct *mm = vma->vm_mm;
980    pte_t *pte;
981    pte_t pteval;
982    spinlock_t *ptl;
983    int ret = SWAP_AGAIN;
984
985    pte = page_check_address(page, mm, address, &ptl, 0);
986    if (!pte)
987        goto out;
988
989    /*
990     * If the page is mlock()d, we cannot swap it out.
991     * If it's recently referenced (perhaps page_referenced
992     * skipped over this mm) then we should reactivate it.
993     */
994    if (!(flags & TTU_IGNORE_MLOCK)) {
995        if (vma->vm_flags & VM_LOCKED)
996            goto out_mlock;
997
998        if (TTU_ACTION(flags) == TTU_MUNLOCK)
999            goto out_unmap;
1000    }
1001    if (!(flags & TTU_IGNORE_ACCESS)) {
1002        if (ptep_clear_flush_young_notify(vma, address, pte)) {
1003            ret = SWAP_FAIL;
1004            goto out_unmap;
1005        }
1006      }
1007
1008    /* Nuke the page table entry. */
1009    flush_cache_page(vma, address, page_to_pfn(page));
1010    pteval = ptep_clear_flush_notify(vma, address, pte);
1011
1012    /* Move the dirty bit to the physical page now the pte is gone. */
1013    if (pte_dirty(pteval))
1014        set_page_dirty(page);
1015
1016    /* Update high watermark before we lower rss */
1017    update_hiwater_rss(mm);
1018
1019    if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1020        if (PageAnon(page))
1021            dec_mm_counter(mm, MM_ANONPAGES);
1022        else
1023            dec_mm_counter(mm, MM_FILEPAGES);
1024        set_pte_at(mm, address, pte,
1025                swp_entry_to_pte(make_hwpoison_entry(page)));
1026    } else if (PageAnon(page)) {
1027        swp_entry_t entry = { .val = page_private(page) };
1028
1029        if (PageSwapCache(page)) {
1030            /*
1031             * Store the swap location in the pte.
1032             * See handle_pte_fault() ...
1033             */
1034            if (swap_duplicate(entry) < 0) {
1035                set_pte_at(mm, address, pte, pteval);
1036                ret = SWAP_FAIL;
1037                goto out_unmap;
1038            }
1039            if (list_empty(&mm->mmlist)) {
1040                spin_lock(&mmlist_lock);
1041                if (list_empty(&mm->mmlist))
1042                    list_add(&mm->mmlist, &init_mm.mmlist);
1043                spin_unlock(&mmlist_lock);
1044            }
1045            dec_mm_counter(mm, MM_ANONPAGES);
1046            inc_mm_counter(mm, MM_SWAPENTS);
1047        } else if (PAGE_MIGRATION) {
1048            /*
1049             * Store the pfn of the page in a special migration
1050             * pte. do_swap_page() will wait until the migration
1051             * pte is removed and then restart fault handling.
1052             */
1053            BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1054            entry = make_migration_entry(page, pte_write(pteval));
1055        }
1056        set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1057        BUG_ON(pte_file(*pte));
1058    } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
1059        /* Establish migration entry for a file page */
1060        swp_entry_t entry;
1061        entry = make_migration_entry(page, pte_write(pteval));
1062        set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1063    } else
1064        dec_mm_counter(mm, MM_FILEPAGES);
1065
1066    page_remove_rmap(page);
1067    page_cache_release(page);
1068
1069out_unmap:
1070    pte_unmap_unlock(pte, ptl);
1071out:
1072    return ret;
1073
1074out_mlock:
1075    pte_unmap_unlock(pte, ptl);
1076
1077
1078    /*
1079     * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1080     * unstable result and race. Plus, We can't wait here because
1081     * we now hold anon_vma->lock or mapping->i_mmap_lock.
1082     * if trylock failed, the page remain in evictable lru and later
1083     * vmscan could retry to move the page to unevictable lru if the
1084     * page is actually mlocked.
1085     */
1086    if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1087        if (vma->vm_flags & VM_LOCKED) {
1088            mlock_vma_page(page);
1089            ret = SWAP_MLOCK;
1090        }
1091        up_read(&vma->vm_mm->mmap_sem);
1092    }
1093    return ret;
1094}
1095
1096/*
1097 * objrmap doesn't work for nonlinear VMAs because the assumption that
1098 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1099 * Consequently, given a particular page and its ->index, we cannot locate the
1100 * ptes which are mapping that page without an exhaustive linear search.
1101 *
1102 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1103 * maps the file to which the target page belongs. The ->vm_private_data field
1104 * holds the current cursor into that scan. Successive searches will circulate
1105 * around the vma's virtual address space.
1106 *
1107 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1108 * more scanning pressure is placed against them as well. Eventually pages
1109 * will become fully unmapped and are eligible for eviction.
1110 *
1111 * For very sparsely populated VMAs this is a little inefficient - chances are
1112 * there there won't be many ptes located within the scan cluster. In this case
1113 * maybe we could scan further - to the end of the pte page, perhaps.
1114 *
1115 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1116 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1117 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1118 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1119 */
1120#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1121#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1122
1123static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1124        struct vm_area_struct *vma, struct page *check_page)
1125{
1126    struct mm_struct *mm = vma->vm_mm;
1127    pgd_t *pgd;
1128    pud_t *pud;
1129    pmd_t *pmd;
1130    pte_t *pte;
1131    pte_t pteval;
1132    spinlock_t *ptl;
1133    struct page *page;
1134    unsigned long address;
1135    unsigned long end;
1136    int ret = SWAP_AGAIN;
1137    int locked_vma = 0;
1138
1139    address = (vma->vm_start + cursor) & CLUSTER_MASK;
1140    end = address + CLUSTER_SIZE;
1141    if (address < vma->vm_start)
1142        address = vma->vm_start;
1143    if (end > vma->vm_end)
1144        end = vma->vm_end;
1145
1146    pgd = pgd_offset(mm, address);
1147    if (!pgd_present(*pgd))
1148        return ret;
1149
1150    pud = pud_offset(pgd, address);
1151    if (!pud_present(*pud))
1152        return ret;
1153
1154    pmd = pmd_offset(pud, address);
1155    if (!pmd_present(*pmd))
1156        return ret;
1157
1158    /*
1159     * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1160     * keep the sem while scanning the cluster for mlocking pages.
1161     */
1162    if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1163        locked_vma = (vma->vm_flags & VM_LOCKED);
1164        if (!locked_vma)
1165            up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1166    }
1167
1168    pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1169
1170    /* Update high watermark before we lower rss */
1171    update_hiwater_rss(mm);
1172
1173    for (; address < end; pte++, address += PAGE_SIZE) {
1174        if (!pte_present(*pte))
1175            continue;
1176        page = vm_normal_page(vma, address, *pte);
1177        BUG_ON(!page || PageAnon(page));
1178
1179        if (locked_vma) {
1180            mlock_vma_page(page); /* no-op if already mlocked */
1181            if (page == check_page)
1182                ret = SWAP_MLOCK;
1183            continue; /* don't unmap */
1184        }
1185
1186        if (ptep_clear_flush_young_notify(vma, address, pte))
1187            continue;
1188
1189        /* Nuke the page table entry. */
1190        flush_cache_page(vma, address, pte_pfn(*pte));
1191        pteval = ptep_clear_flush_notify(vma, address, pte);
1192
1193        /* If nonlinear, store the file page offset in the pte. */
1194        if (page->index != linear_page_index(vma, address))
1195            set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1196
1197        /* Move the dirty bit to the physical page now the pte is gone. */
1198        if (pte_dirty(pteval))
1199            set_page_dirty(page);
1200
1201        page_remove_rmap(page);
1202        page_cache_release(page);
1203        dec_mm_counter(mm, MM_FILEPAGES);
1204        (*mapcount)--;
1205    }
1206    pte_unmap_unlock(pte - 1, ptl);
1207    if (locked_vma)
1208        up_read(&vma->vm_mm->mmap_sem);
1209    return ret;
1210}
1211
1212static bool is_vma_temporary_stack(struct vm_area_struct *vma)
1213{
1214    int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1215
1216    if (!maybe_stack)
1217        return false;
1218
1219    if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1220                        VM_STACK_INCOMPLETE_SETUP)
1221        return true;
1222
1223    return false;
1224}
1225
1226/**
1227 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1228 * rmap method
1229 * @page: the page to unmap/unlock
1230 * @flags: action and flags
1231 *
1232 * Find all the mappings of a page using the mapping pointer and the vma chains
1233 * contained in the anon_vma struct it points to.
1234 *
1235 * This function is only called from try_to_unmap/try_to_munlock for
1236 * anonymous pages.
1237 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1238 * where the page was found will be held for write. So, we won't recheck
1239 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1240 * 'LOCKED.
1241 */
1242static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1243{
1244    struct anon_vma *anon_vma;
1245    struct anon_vma_chain *avc;
1246    int ret = SWAP_AGAIN;
1247
1248    anon_vma = page_lock_anon_vma(page);
1249    if (!anon_vma)
1250        return ret;
1251
1252    list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1253        struct vm_area_struct *vma = avc->vma;
1254        unsigned long address;
1255
1256        /*
1257         * During exec, a temporary VMA is setup and later moved.
1258         * The VMA is moved under the anon_vma lock but not the
1259         * page tables leading to a race where migration cannot
1260         * find the migration ptes. Rather than increasing the
1261         * locking requirements of exec(), migration skips
1262         * temporary VMAs until after exec() completes.
1263         */
1264        if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1265                is_vma_temporary_stack(vma))
1266            continue;
1267
1268        address = vma_address(page, vma);
1269        if (address == -EFAULT)
1270            continue;
1271        ret = try_to_unmap_one(page, vma, address, flags);
1272        if (ret != SWAP_AGAIN || !page_mapped(page))
1273            break;
1274    }
1275
1276    page_unlock_anon_vma(anon_vma);
1277    return ret;
1278}
1279
1280/**
1281 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1282 * @page: the page to unmap/unlock
1283 * @flags: action and flags
1284 *
1285 * Find all the mappings of a page using the mapping pointer and the vma chains
1286 * contained in the address_space struct it points to.
1287 *
1288 * This function is only called from try_to_unmap/try_to_munlock for
1289 * object-based pages.
1290 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1291 * where the page was found will be held for write. So, we won't recheck
1292 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1293 * 'LOCKED.
1294 */
1295static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1296{
1297    struct address_space *mapping = page->mapping;
1298    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1299    struct vm_area_struct *vma;
1300    struct prio_tree_iter iter;
1301    int ret = SWAP_AGAIN;
1302    unsigned long cursor;
1303    unsigned long max_nl_cursor = 0;
1304    unsigned long max_nl_size = 0;
1305    unsigned int mapcount;
1306
1307    spin_lock(&mapping->i_mmap_lock);
1308    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1309        unsigned long address = vma_address(page, vma);
1310        if (address == -EFAULT)
1311            continue;
1312        ret = try_to_unmap_one(page, vma, address, flags);
1313        if (ret != SWAP_AGAIN || !page_mapped(page))
1314            goto out;
1315    }
1316
1317    if (list_empty(&mapping->i_mmap_nonlinear))
1318        goto out;
1319
1320    /*
1321     * We don't bother to try to find the munlocked page in nonlinears.
1322     * It's costly. Instead, later, page reclaim logic may call
1323     * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1324     */
1325    if (TTU_ACTION(flags) == TTU_MUNLOCK)
1326        goto out;
1327
1328    list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1329                        shared.vm_set.list) {
1330        cursor = (unsigned long) vma->vm_private_data;
1331        if (cursor > max_nl_cursor)
1332            max_nl_cursor = cursor;
1333        cursor = vma->vm_end - vma->vm_start;
1334        if (cursor > max_nl_size)
1335            max_nl_size = cursor;
1336    }
1337
1338    if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1339        ret = SWAP_FAIL;
1340        goto out;
1341    }
1342
1343    /*
1344     * We don't try to search for this page in the nonlinear vmas,
1345     * and page_referenced wouldn't have found it anyway. Instead
1346     * just walk the nonlinear vmas trying to age and unmap some.
1347     * The mapcount of the page we came in with is irrelevant,
1348     * but even so use it as a guide to how hard we should try?
1349     */
1350    mapcount = page_mapcount(page);
1351    if (!mapcount)
1352        goto out;
1353    cond_resched_lock(&mapping->i_mmap_lock);
1354
1355    max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1356    if (max_nl_cursor == 0)
1357        max_nl_cursor = CLUSTER_SIZE;
1358
1359    do {
1360        list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1361                        shared.vm_set.list) {
1362            cursor = (unsigned long) vma->vm_private_data;
1363            while ( cursor < max_nl_cursor &&
1364                cursor < vma->vm_end - vma->vm_start) {
1365                if (try_to_unmap_cluster(cursor, &mapcount,
1366                        vma, page) == SWAP_MLOCK)
1367                    ret = SWAP_MLOCK;
1368                cursor += CLUSTER_SIZE;
1369                vma->vm_private_data = (void *) cursor;
1370                if ((int)mapcount <= 0)
1371                    goto out;
1372            }
1373            vma->vm_private_data = (void *) max_nl_cursor;
1374        }
1375        cond_resched_lock(&mapping->i_mmap_lock);
1376        max_nl_cursor += CLUSTER_SIZE;
1377    } while (max_nl_cursor <= max_nl_size);
1378
1379    /*
1380     * Don't loop forever (perhaps all the remaining pages are
1381     * in locked vmas). Reset cursor on all unreserved nonlinear
1382     * vmas, now forgetting on which ones it had fallen behind.
1383     */
1384    list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1385        vma->vm_private_data = NULL;
1386out:
1387    spin_unlock(&mapping->i_mmap_lock);
1388    return ret;
1389}
1390
1391/**
1392 * try_to_unmap - try to remove all page table mappings to a page
1393 * @page: the page to get unmapped
1394 * @flags: action and flags
1395 *
1396 * Tries to remove all the page table entries which are mapping this
1397 * page, used in the pageout path. Caller must hold the page lock.
1398 * Return values are:
1399 *
1400 * SWAP_SUCCESS - we succeeded in removing all mappings
1401 * SWAP_AGAIN - we missed a mapping, try again later
1402 * SWAP_FAIL - the page is unswappable
1403 * SWAP_MLOCK - page is mlocked.
1404 */
1405int try_to_unmap(struct page *page, enum ttu_flags flags)
1406{
1407    int ret;
1408
1409    BUG_ON(!PageLocked(page));
1410
1411    if (unlikely(PageKsm(page)))
1412        ret = try_to_unmap_ksm(page, flags);
1413    else if (PageAnon(page))
1414        ret = try_to_unmap_anon(page, flags);
1415    else
1416        ret = try_to_unmap_file(page, flags);
1417    if (ret != SWAP_MLOCK && !page_mapped(page))
1418        ret = SWAP_SUCCESS;
1419    return ret;
1420}
1421
1422/**
1423 * try_to_munlock - try to munlock a page
1424 * @page: the page to be munlocked
1425 *
1426 * Called from munlock code. Checks all of the VMAs mapping the page
1427 * to make sure nobody else has this page mlocked. The page will be
1428 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1429 *
1430 * Return values are:
1431 *
1432 * SWAP_AGAIN - no vma is holding page mlocked, or,
1433 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1434 * SWAP_FAIL - page cannot be located at present
1435 * SWAP_MLOCK - page is now mlocked.
1436 */
1437int try_to_munlock(struct page *page)
1438{
1439    VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1440
1441    if (unlikely(PageKsm(page)))
1442        return try_to_unmap_ksm(page, TTU_MUNLOCK);
1443    else if (PageAnon(page))
1444        return try_to_unmap_anon(page, TTU_MUNLOCK);
1445    else
1446        return try_to_unmap_file(page, TTU_MUNLOCK);
1447}
1448
1449#if defined(CONFIG_KSM) || defined(CONFIG_MIGRATION)
1450/*
1451 * Drop an anon_vma refcount, freeing the anon_vma and anon_vma->root
1452 * if necessary. Be careful to do all the tests under the lock. Once
1453 * we know we are the last user, nobody else can get a reference and we
1454 * can do the freeing without the lock.
1455 */
1456void drop_anon_vma(struct anon_vma *anon_vma)
1457{
1458    BUG_ON(atomic_read(&anon_vma->external_refcount) <= 0);
1459    if (atomic_dec_and_lock(&anon_vma->external_refcount, &anon_vma->root->lock)) {
1460        struct anon_vma *root = anon_vma->root;
1461        int empty = list_empty(&anon_vma->head);
1462        int last_root_user = 0;
1463        int root_empty = 0;
1464
1465        /*
1466         * The refcount on a non-root anon_vma got dropped. Drop
1467         * the refcount on the root and check if we need to free it.
1468         */
1469        if (empty && anon_vma != root) {
1470            BUG_ON(atomic_read(&root->external_refcount) <= 0);
1471            last_root_user = atomic_dec_and_test(&root->external_refcount);
1472            root_empty = list_empty(&root->head);
1473        }
1474        anon_vma_unlock(anon_vma);
1475
1476        if (empty) {
1477            anon_vma_free(anon_vma);
1478            if (root_empty && last_root_user)
1479                anon_vma_free(root);
1480        }
1481    }
1482}
1483#endif
1484
1485#ifdef CONFIG_MIGRATION
1486/*
1487 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1488 * Called by migrate.c to remove migration ptes, but might be used more later.
1489 */
1490static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1491        struct vm_area_struct *, unsigned long, void *), void *arg)
1492{
1493    struct anon_vma *anon_vma;
1494    struct anon_vma_chain *avc;
1495    int ret = SWAP_AGAIN;
1496
1497    /*
1498     * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1499     * because that depends on page_mapped(); but not all its usages
1500     * are holding mmap_sem. Users without mmap_sem are required to
1501     * take a reference count to prevent the anon_vma disappearing
1502     */
1503    anon_vma = page_anon_vma(page);
1504    if (!anon_vma)
1505        return ret;
1506    anon_vma_lock(anon_vma);
1507    list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1508        struct vm_area_struct *vma = avc->vma;
1509        unsigned long address = vma_address(page, vma);
1510        if (address == -EFAULT)
1511            continue;
1512        ret = rmap_one(page, vma, address, arg);
1513        if (ret != SWAP_AGAIN)
1514            break;
1515    }
1516    anon_vma_unlock(anon_vma);
1517    return ret;
1518}
1519
1520static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1521        struct vm_area_struct *, unsigned long, void *), void *arg)
1522{
1523    struct address_space *mapping = page->mapping;
1524    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1525    struct vm_area_struct *vma;
1526    struct prio_tree_iter iter;
1527    int ret = SWAP_AGAIN;
1528
1529    if (!mapping)
1530        return ret;
1531    spin_lock(&mapping->i_mmap_lock);
1532    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1533        unsigned long address = vma_address(page, vma);
1534        if (address == -EFAULT)
1535            continue;
1536        ret = rmap_one(page, vma, address, arg);
1537        if (ret != SWAP_AGAIN)
1538            break;
1539    }
1540    /*
1541     * No nonlinear handling: being always shared, nonlinear vmas
1542     * never contain migration ptes. Decide what to do about this
1543     * limitation to linear when we need rmap_walk() on nonlinear.
1544     */
1545    spin_unlock(&mapping->i_mmap_lock);
1546    return ret;
1547}
1548
1549int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1550        struct vm_area_struct *, unsigned long, void *), void *arg)
1551{
1552    VM_BUG_ON(!PageLocked(page));
1553
1554    if (unlikely(PageKsm(page)))
1555        return rmap_walk_ksm(page, rmap_one, arg);
1556    else if (PageAnon(page))
1557        return rmap_walk_anon(page, rmap_one, arg);
1558    else
1559        return rmap_walk_file(page, rmap_one, arg);
1560}
1561#endif /* CONFIG_MIGRATION */
1562
1563#ifdef CONFIG_HUGETLB_PAGE
1564/*
1565 * The following three functions are for anonymous (private mapped) hugepages.
1566 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1567 * and no lru code, because we handle hugepages differently from common pages.
1568 */
1569static void __hugepage_set_anon_rmap(struct page *page,
1570    struct vm_area_struct *vma, unsigned long address, int exclusive)
1571{
1572    struct anon_vma *anon_vma = vma->anon_vma;
1573
1574    BUG_ON(!anon_vma);
1575
1576    if (PageAnon(page))
1577        return;
1578    if (!exclusive)
1579        anon_vma = anon_vma->root;
1580
1581    anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1582    page->mapping = (struct address_space *) anon_vma;
1583    page->index = linear_page_index(vma, address);
1584}
1585
1586void hugepage_add_anon_rmap(struct page *page,
1587                struct vm_area_struct *vma, unsigned long address)
1588{
1589    struct anon_vma *anon_vma = vma->anon_vma;
1590    int first;
1591
1592    BUG_ON(!PageLocked(page));
1593    BUG_ON(!anon_vma);
1594    BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1595    first = atomic_inc_and_test(&page->_mapcount);
1596    if (first)
1597        __hugepage_set_anon_rmap(page, vma, address, 0);
1598}
1599
1600void hugepage_add_new_anon_rmap(struct page *page,
1601            struct vm_area_struct *vma, unsigned long address)
1602{
1603    BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1604    atomic_set(&page->_mapcount, 0);
1605    __hugepage_set_anon_rmap(page, vma, address, 1);
1606}
1607#endif /* CONFIG_HUGETLB_PAGE */
1608

Archive Download this file



interactive