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_mutex
28 * anon_vma->mutex
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->i_lock (in set_page_dirty's __mark_inode_dirty)
35 * inode_wb_list_lock (in set_page_dirty's __mark_inode_dirty)
36 * sb_lock (within inode_lock in fs/fs-writeback.c)
37 * mapping->tree_lock (widely used, in set_page_dirty,
38 * in arch-dependent flush_dcache_mmap_lock,
39 * within inode_wb_list_lock in __sync_single_inode)
40 *
41 * anon_vma->mutex,mapping->i_mutex (memory_failure, collect_procs_anon)
42 * ->tasklist_lock
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    struct anon_vma *anon_vma;
71
72    anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
73    if (anon_vma) {
74        atomic_set(&anon_vma->refcount, 1);
75        /*
76         * Initialise the anon_vma root to point to itself. If called
77         * from fork, the root will be reset to the parents anon_vma.
78         */
79        anon_vma->root = anon_vma;
80    }
81
82    return anon_vma;
83}
84
85static inline void anon_vma_free(struct anon_vma *anon_vma)
86{
87    VM_BUG_ON(atomic_read(&anon_vma->refcount));
88
89    /*
90     * Synchronize against page_lock_anon_vma() such that
91     * we can safely hold the lock without the anon_vma getting
92     * freed.
93     *
94     * Relies on the full mb implied by the atomic_dec_and_test() from
95     * put_anon_vma() against the acquire barrier implied by
96     * mutex_trylock() from page_lock_anon_vma(). This orders:
97     *
98     * page_lock_anon_vma() VS put_anon_vma()
99     * mutex_trylock() atomic_dec_and_test()
100     * LOCK MB
101     * atomic_read() mutex_is_locked()
102     *
103     * LOCK should suffice since the actual taking of the lock must
104     * happen _before_ what follows.
105     */
106    if (mutex_is_locked(&anon_vma->root->mutex)) {
107        anon_vma_lock(anon_vma);
108        anon_vma_unlock(anon_vma);
109    }
110
111    kmem_cache_free(anon_vma_cachep, anon_vma);
112}
113
114static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
115{
116    return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
117}
118
119static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
120{
121    kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
122}
123
124/**
125 * anon_vma_prepare - attach an anon_vma to a memory region
126 * @vma: the memory region in question
127 *
128 * This makes sure the memory mapping described by 'vma' has
129 * an 'anon_vma' attached to it, so that we can associate the
130 * anonymous pages mapped into it with that anon_vma.
131 *
132 * The common case will be that we already have one, but if
133 * not we either need to find an adjacent mapping that we
134 * can re-use the anon_vma from (very common when the only
135 * reason for splitting a vma has been mprotect()), or we
136 * allocate a new one.
137 *
138 * Anon-vma allocations are very subtle, because we may have
139 * optimistically looked up an anon_vma in page_lock_anon_vma()
140 * and that may actually touch the spinlock even in the newly
141 * allocated vma (it depends on RCU to make sure that the
142 * anon_vma isn't actually destroyed).
143 *
144 * As a result, we need to do proper anon_vma locking even
145 * for the new allocation. At the same time, we do not want
146 * to do any locking for the common case of already having
147 * an anon_vma.
148 *
149 * This must be called with the mmap_sem held for reading.
150 */
151int anon_vma_prepare(struct vm_area_struct *vma)
152{
153    struct anon_vma *anon_vma = vma->anon_vma;
154    struct anon_vma_chain *avc;
155
156    might_sleep();
157    if (unlikely(!anon_vma)) {
158        struct mm_struct *mm = vma->vm_mm;
159        struct anon_vma *allocated;
160
161        avc = anon_vma_chain_alloc(GFP_KERNEL);
162        if (!avc)
163            goto out_enomem;
164
165        anon_vma = find_mergeable_anon_vma(vma);
166        allocated = NULL;
167        if (!anon_vma) {
168            anon_vma = anon_vma_alloc();
169            if (unlikely(!anon_vma))
170                goto out_enomem_free_avc;
171            allocated = anon_vma;
172        }
173
174        anon_vma_lock(anon_vma);
175        /* page_table_lock to protect against threads */
176        spin_lock(&mm->page_table_lock);
177        if (likely(!vma->anon_vma)) {
178            vma->anon_vma = anon_vma;
179            avc->anon_vma = anon_vma;
180            avc->vma = vma;
181            list_add(&avc->same_vma, &vma->anon_vma_chain);
182            list_add_tail(&avc->same_anon_vma, &anon_vma->head);
183            allocated = NULL;
184            avc = NULL;
185        }
186        spin_unlock(&mm->page_table_lock);
187        anon_vma_unlock(anon_vma);
188
189        if (unlikely(allocated))
190            put_anon_vma(allocated);
191        if (unlikely(avc))
192            anon_vma_chain_free(avc);
193    }
194    return 0;
195
196 out_enomem_free_avc:
197    anon_vma_chain_free(avc);
198 out_enomem:
199    return -ENOMEM;
200}
201
202/*
203 * This is a useful helper function for locking the anon_vma root as
204 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
205 * have the same vma.
206 *
207 * Such anon_vma's should have the same root, so you'd expect to see
208 * just a single mutex_lock for the whole traversal.
209 */
210static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
211{
212    struct anon_vma *new_root = anon_vma->root;
213    if (new_root != root) {
214        if (WARN_ON_ONCE(root))
215            mutex_unlock(&root->mutex);
216        root = new_root;
217        mutex_lock(&root->mutex);
218    }
219    return root;
220}
221
222static inline void unlock_anon_vma_root(struct anon_vma *root)
223{
224    if (root)
225        mutex_unlock(&root->mutex);
226}
227
228static void anon_vma_chain_link(struct vm_area_struct *vma,
229                struct anon_vma_chain *avc,
230                struct anon_vma *anon_vma)
231{
232    avc->vma = vma;
233    avc->anon_vma = anon_vma;
234    list_add(&avc->same_vma, &vma->anon_vma_chain);
235
236    /*
237     * It's critical to add new vmas to the tail of the anon_vma,
238     * see comment in huge_memory.c:__split_huge_page().
239     */
240    list_add_tail(&avc->same_anon_vma, &anon_vma->head);
241}
242
243/*
244 * Attach the anon_vmas from src to dst.
245 * Returns 0 on success, -ENOMEM on failure.
246 */
247int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
248{
249    struct anon_vma_chain *avc, *pavc;
250    struct anon_vma *root = NULL;
251
252    list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
253        struct anon_vma *anon_vma;
254
255        avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
256        if (unlikely(!avc)) {
257            unlock_anon_vma_root(root);
258            root = NULL;
259            avc = anon_vma_chain_alloc(GFP_KERNEL);
260            if (!avc)
261                goto enomem_failure;
262        }
263        anon_vma = pavc->anon_vma;
264        root = lock_anon_vma_root(root, anon_vma);
265        anon_vma_chain_link(dst, avc, anon_vma);
266    }
267    unlock_anon_vma_root(root);
268    return 0;
269
270 enomem_failure:
271    unlink_anon_vmas(dst);
272    return -ENOMEM;
273}
274
275/*
276 * Attach vma to its own anon_vma, as well as to the anon_vmas that
277 * the corresponding VMA in the parent process is attached to.
278 * Returns 0 on success, non-zero on failure.
279 */
280int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
281{
282    struct anon_vma_chain *avc;
283    struct anon_vma *anon_vma;
284
285    /* Don't bother if the parent process has no anon_vma here. */
286    if (!pvma->anon_vma)
287        return 0;
288
289    /*
290     * First, attach the new VMA to the parent VMA's anon_vmas,
291     * so rmap can find non-COWed pages in child processes.
292     */
293    if (anon_vma_clone(vma, pvma))
294        return -ENOMEM;
295
296    /* Then add our own anon_vma. */
297    anon_vma = anon_vma_alloc();
298    if (!anon_vma)
299        goto out_error;
300    avc = anon_vma_chain_alloc(GFP_KERNEL);
301    if (!avc)
302        goto out_error_free_anon_vma;
303
304    /*
305     * The root anon_vma's spinlock is the lock actually used when we
306     * lock any of the anon_vmas in this anon_vma tree.
307     */
308    anon_vma->root = pvma->anon_vma->root;
309    /*
310     * With refcounts, an anon_vma can stay around longer than the
311     * process it belongs to. The root anon_vma needs to be pinned until
312     * this anon_vma is freed, because the lock lives in the root.
313     */
314    get_anon_vma(anon_vma->root);
315    /* Mark this anon_vma as the one where our new (COWed) pages go. */
316    vma->anon_vma = anon_vma;
317    anon_vma_lock(anon_vma);
318    anon_vma_chain_link(vma, avc, anon_vma);
319    anon_vma_unlock(anon_vma);
320
321    return 0;
322
323 out_error_free_anon_vma:
324    put_anon_vma(anon_vma);
325 out_error:
326    unlink_anon_vmas(vma);
327    return -ENOMEM;
328}
329
330void unlink_anon_vmas(struct vm_area_struct *vma)
331{
332    struct anon_vma_chain *avc, *next;
333    struct anon_vma *root = NULL;
334
335    /*
336     * Unlink each anon_vma chained to the VMA. This list is ordered
337     * from newest to oldest, ensuring the root anon_vma gets freed last.
338     */
339    list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
340        struct anon_vma *anon_vma = avc->anon_vma;
341
342        root = lock_anon_vma_root(root, anon_vma);
343        list_del(&avc->same_anon_vma);
344
345        /*
346         * Leave empty anon_vmas on the list - we'll need
347         * to free them outside the lock.
348         */
349        if (list_empty(&anon_vma->head))
350            continue;
351
352        list_del(&avc->same_vma);
353        anon_vma_chain_free(avc);
354    }
355    unlock_anon_vma_root(root);
356
357    /*
358     * Iterate the list once more, it now only contains empty and unlinked
359     * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
360     * needing to acquire the anon_vma->root->mutex.
361     */
362    list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
363        struct anon_vma *anon_vma = avc->anon_vma;
364
365        put_anon_vma(anon_vma);
366
367        list_del(&avc->same_vma);
368        anon_vma_chain_free(avc);
369    }
370}
371
372static void anon_vma_ctor(void *data)
373{
374    struct anon_vma *anon_vma = data;
375
376    mutex_init(&anon_vma->mutex);
377    atomic_set(&anon_vma->refcount, 0);
378    INIT_LIST_HEAD(&anon_vma->head);
379}
380
381void __init anon_vma_init(void)
382{
383    anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
384            0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
385    anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
386}
387
388/*
389 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
390 *
391 * Since there is no serialization what so ever against page_remove_rmap()
392 * the best this function can do is return a locked anon_vma that might
393 * have been relevant to this page.
394 *
395 * The page might have been remapped to a different anon_vma or the anon_vma
396 * returned may already be freed (and even reused).
397 *
398 * In case it was remapped to a different anon_vma, the new anon_vma will be a
399 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
400 * ensure that any anon_vma obtained from the page will still be valid for as
401 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
402 *
403 * All users of this function must be very careful when walking the anon_vma
404 * chain and verify that the page in question is indeed mapped in it
405 * [ something equivalent to page_mapped_in_vma() ].
406 *
407 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
408 * that the anon_vma pointer from page->mapping is valid if there is a
409 * mapcount, we can dereference the anon_vma after observing those.
410 */
411struct anon_vma *page_get_anon_vma(struct page *page)
412{
413    struct anon_vma *anon_vma = NULL;
414    unsigned long anon_mapping;
415
416    rcu_read_lock();
417    anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
418    if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
419        goto out;
420    if (!page_mapped(page))
421        goto out;
422
423    anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
424    if (!atomic_inc_not_zero(&anon_vma->refcount)) {
425        anon_vma = NULL;
426        goto out;
427    }
428
429    /*
430     * If this page is still mapped, then its anon_vma cannot have been
431     * freed. But if it has been unmapped, we have no security against the
432     * anon_vma structure being freed and reused (for another anon_vma:
433     * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
434     * above cannot corrupt).
435     */
436    if (!page_mapped(page)) {
437        put_anon_vma(anon_vma);
438        anon_vma = NULL;
439    }
440out:
441    rcu_read_unlock();
442
443    return anon_vma;
444}
445
446/*
447 * Similar to page_get_anon_vma() except it locks the anon_vma.
448 *
449 * Its a little more complex as it tries to keep the fast path to a single
450 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
451 * reference like with page_get_anon_vma() and then block on the mutex.
452 */
453struct anon_vma *page_lock_anon_vma(struct page *page)
454{
455    struct anon_vma *anon_vma = NULL;
456    struct anon_vma *root_anon_vma;
457    unsigned long anon_mapping;
458
459    rcu_read_lock();
460    anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
461    if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
462        goto out;
463    if (!page_mapped(page))
464        goto out;
465
466    anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
467    root_anon_vma = ACCESS_ONCE(anon_vma->root);
468    if (mutex_trylock(&root_anon_vma->mutex)) {
469        /*
470         * If the page is still mapped, then this anon_vma is still
471         * its anon_vma, and holding the mutex ensures that it will
472         * not go away, see anon_vma_free().
473         */
474        if (!page_mapped(page)) {
475            mutex_unlock(&root_anon_vma->mutex);
476            anon_vma = NULL;
477        }
478        goto out;
479    }
480
481    /* trylock failed, we got to sleep */
482    if (!atomic_inc_not_zero(&anon_vma->refcount)) {
483        anon_vma = NULL;
484        goto out;
485    }
486
487    if (!page_mapped(page)) {
488        put_anon_vma(anon_vma);
489        anon_vma = NULL;
490        goto out;
491    }
492
493    /* we pinned the anon_vma, its safe to sleep */
494    rcu_read_unlock();
495    anon_vma_lock(anon_vma);
496
497    if (atomic_dec_and_test(&anon_vma->refcount)) {
498        /*
499         * Oops, we held the last refcount, release the lock
500         * and bail -- can't simply use put_anon_vma() because
501         * we'll deadlock on the anon_vma_lock() recursion.
502         */
503        anon_vma_unlock(anon_vma);
504        __put_anon_vma(anon_vma);
505        anon_vma = NULL;
506    }
507
508    return anon_vma;
509
510out:
511    rcu_read_unlock();
512    return anon_vma;
513}
514
515void page_unlock_anon_vma(struct anon_vma *anon_vma)
516{
517    anon_vma_unlock(anon_vma);
518}
519
520/*
521 * At what user virtual address is page expected in @vma?
522 * Returns virtual address or -EFAULT if page's index/offset is not
523 * within the range mapped the @vma.
524 */
525inline unsigned long
526vma_address(struct page *page, struct vm_area_struct *vma)
527{
528    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
529    unsigned long address;
530
531    if (unlikely(is_vm_hugetlb_page(vma)))
532        pgoff = page->index << huge_page_order(page_hstate(page));
533    address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
534    if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
535        /* page should be within @vma mapping range */
536        return -EFAULT;
537    }
538    return address;
539}
540
541/*
542 * At what user virtual address is page expected in vma?
543 * Caller should check the page is actually part of the vma.
544 */
545unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
546{
547    if (PageAnon(page)) {
548        struct anon_vma *page__anon_vma = page_anon_vma(page);
549        /*
550         * Note: swapoff's unuse_vma() is more efficient with this
551         * check, and needs it to match anon_vma when KSM is active.
552         */
553        if (!vma->anon_vma || !page__anon_vma ||
554            vma->anon_vma->root != page__anon_vma->root)
555            return -EFAULT;
556    } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
557        if (!vma->vm_file ||
558            vma->vm_file->f_mapping != page->mapping)
559            return -EFAULT;
560    } else
561        return -EFAULT;
562    return vma_address(page, vma);
563}
564
565/*
566 * Check that @page is mapped at @address into @mm.
567 *
568 * If @sync is false, page_check_address may perform a racy check to avoid
569 * the page table lock when the pte is not present (helpful when reclaiming
570 * highly shared pages).
571 *
572 * On success returns with pte mapped and locked.
573 */
574pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
575              unsigned long address, spinlock_t **ptlp, int sync)
576{
577    pgd_t *pgd;
578    pud_t *pud;
579    pmd_t *pmd;
580    pte_t *pte;
581    spinlock_t *ptl;
582
583    if (unlikely(PageHuge(page))) {
584        pte = huge_pte_offset(mm, address);
585        ptl = &mm->page_table_lock;
586        goto check;
587    }
588
589    pgd = pgd_offset(mm, address);
590    if (!pgd_present(*pgd))
591        return NULL;
592
593    pud = pud_offset(pgd, address);
594    if (!pud_present(*pud))
595        return NULL;
596
597    pmd = pmd_offset(pud, address);
598    if (!pmd_present(*pmd))
599        return NULL;
600    if (pmd_trans_huge(*pmd))
601        return NULL;
602
603    pte = pte_offset_map(pmd, address);
604    /* Make a quick check before getting the lock */
605    if (!sync && !pte_present(*pte)) {
606        pte_unmap(pte);
607        return NULL;
608    }
609
610    ptl = pte_lockptr(mm, pmd);
611check:
612    spin_lock(ptl);
613    if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
614        *ptlp = ptl;
615        return pte;
616    }
617    pte_unmap_unlock(pte, ptl);
618    return NULL;
619}
620
621/**
622 * page_mapped_in_vma - check whether a page is really mapped in a VMA
623 * @page: the page to test
624 * @vma: the VMA to test
625 *
626 * Returns 1 if the page is mapped into the page tables of the VMA, 0
627 * if the page is not mapped into the page tables of this VMA. Only
628 * valid for normal file or anonymous VMAs.
629 */
630int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
631{
632    unsigned long address;
633    pte_t *pte;
634    spinlock_t *ptl;
635
636    address = vma_address(page, vma);
637    if (address == -EFAULT) /* out of vma range */
638        return 0;
639    pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
640    if (!pte) /* the page is not in this mm */
641        return 0;
642    pte_unmap_unlock(pte, ptl);
643
644    return 1;
645}
646
647/*
648 * Subfunctions of page_referenced: page_referenced_one called
649 * repeatedly from either page_referenced_anon or page_referenced_file.
650 */
651int page_referenced_one(struct page *page, struct vm_area_struct *vma,
652            unsigned long address, unsigned int *mapcount,
653            unsigned long *vm_flags)
654{
655    struct mm_struct *mm = vma->vm_mm;
656    int referenced = 0;
657
658    if (unlikely(PageTransHuge(page))) {
659        pmd_t *pmd;
660
661        spin_lock(&mm->page_table_lock);
662        /*
663         * rmap might return false positives; we must filter
664         * these out using page_check_address_pmd().
665         */
666        pmd = page_check_address_pmd(page, mm, address,
667                         PAGE_CHECK_ADDRESS_PMD_FLAG);
668        if (!pmd) {
669            spin_unlock(&mm->page_table_lock);
670            goto out;
671        }
672
673        if (vma->vm_flags & VM_LOCKED) {
674            spin_unlock(&mm->page_table_lock);
675            *mapcount = 0; /* break early from loop */
676            *vm_flags |= VM_LOCKED;
677            goto out;
678        }
679
680        /* go ahead even if the pmd is pmd_trans_splitting() */
681        if (pmdp_clear_flush_young_notify(vma, address, pmd))
682            referenced++;
683        spin_unlock(&mm->page_table_lock);
684    } else {
685        pte_t *pte;
686        spinlock_t *ptl;
687
688        /*
689         * rmap might return false positives; we must filter
690         * these out using page_check_address().
691         */
692        pte = page_check_address(page, mm, address, &ptl, 0);
693        if (!pte)
694            goto out;
695
696        if (vma->vm_flags & VM_LOCKED) {
697            pte_unmap_unlock(pte, ptl);
698            *mapcount = 0; /* break early from loop */
699            *vm_flags |= VM_LOCKED;
700            goto out;
701        }
702
703        if (ptep_clear_flush_young_notify(vma, address, pte)) {
704            /*
705             * Don't treat a reference through a sequentially read
706             * mapping as such. If the page has been used in
707             * another mapping, we will catch it; if this other
708             * mapping is already gone, the unmap path will have
709             * set PG_referenced or activated the page.
710             */
711            if (likely(!VM_SequentialReadHint(vma)))
712                referenced++;
713        }
714        pte_unmap_unlock(pte, ptl);
715    }
716
717    /* Pretend the page is referenced if the task has the
718       swap token and is in the middle of a page fault. */
719    if (mm != current->mm && has_swap_token(mm) &&
720            rwsem_is_locked(&mm->mmap_sem))
721        referenced++;
722
723    (*mapcount)--;
724
725    if (referenced)
726        *vm_flags |= vma->vm_flags;
727out:
728    return referenced;
729}
730
731static int page_referenced_anon(struct page *page,
732                struct mem_cgroup *mem_cont,
733                unsigned long *vm_flags)
734{
735    unsigned int mapcount;
736    struct anon_vma *anon_vma;
737    struct anon_vma_chain *avc;
738    int referenced = 0;
739
740    anon_vma = page_lock_anon_vma(page);
741    if (!anon_vma)
742        return referenced;
743
744    mapcount = page_mapcount(page);
745    list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
746        struct vm_area_struct *vma = avc->vma;
747        unsigned long address = vma_address(page, vma);
748        if (address == -EFAULT)
749            continue;
750        /*
751         * If we are reclaiming on behalf of a cgroup, skip
752         * counting on behalf of references from different
753         * cgroups
754         */
755        if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
756            continue;
757        referenced += page_referenced_one(page, vma, address,
758                          &mapcount, vm_flags);
759        if (!mapcount)
760            break;
761    }
762
763    page_unlock_anon_vma(anon_vma);
764    return referenced;
765}
766
767/**
768 * page_referenced_file - referenced check for object-based rmap
769 * @page: the page we're checking references on.
770 * @mem_cont: target memory controller
771 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
772 *
773 * For an object-based mapped page, find all the places it is mapped and
774 * check/clear the referenced flag. This is done by following the page->mapping
775 * pointer, then walking the chain of vmas it holds. It returns the number
776 * of references it found.
777 *
778 * This function is only called from page_referenced for object-based pages.
779 */
780static int page_referenced_file(struct page *page,
781                struct mem_cgroup *mem_cont,
782                unsigned long *vm_flags)
783{
784    unsigned int mapcount;
785    struct address_space *mapping = page->mapping;
786    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
787    struct vm_area_struct *vma;
788    struct prio_tree_iter iter;
789    int referenced = 0;
790
791    /*
792     * The caller's checks on page->mapping and !PageAnon have made
793     * sure that this is a file page: the check for page->mapping
794     * excludes the case just before it gets set on an anon page.
795     */
796    BUG_ON(PageAnon(page));
797
798    /*
799     * The page lock not only makes sure that page->mapping cannot
800     * suddenly be NULLified by truncation, it makes sure that the
801     * structure at mapping cannot be freed and reused yet,
802     * so we can safely take mapping->i_mmap_mutex.
803     */
804    BUG_ON(!PageLocked(page));
805
806    mutex_lock(&mapping->i_mmap_mutex);
807
808    /*
809     * i_mmap_mutex does not stabilize mapcount at all, but mapcount
810     * is more likely to be accurate if we note it after spinning.
811     */
812    mapcount = page_mapcount(page);
813
814    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
815        unsigned long address = vma_address(page, vma);
816        if (address == -EFAULT)
817            continue;
818        /*
819         * If we are reclaiming on behalf of a cgroup, skip
820         * counting on behalf of references from different
821         * cgroups
822         */
823        if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
824            continue;
825        referenced += page_referenced_one(page, vma, address,
826                          &mapcount, vm_flags);
827        if (!mapcount)
828            break;
829    }
830
831    mutex_unlock(&mapping->i_mmap_mutex);
832    return referenced;
833}
834
835/**
836 * page_referenced - test if the page was referenced
837 * @page: the page to test
838 * @is_locked: caller holds lock on the page
839 * @mem_cont: target memory controller
840 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
841 *
842 * Quick test_and_clear_referenced for all mappings to a page,
843 * returns the number of ptes which referenced the page.
844 */
845int page_referenced(struct page *page,
846            int is_locked,
847            struct mem_cgroup *mem_cont,
848            unsigned long *vm_flags)
849{
850    int referenced = 0;
851    int we_locked = 0;
852
853    *vm_flags = 0;
854    if (page_mapped(page) && page_rmapping(page)) {
855        if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
856            we_locked = trylock_page(page);
857            if (!we_locked) {
858                referenced++;
859                goto out;
860            }
861        }
862        if (unlikely(PageKsm(page)))
863            referenced += page_referenced_ksm(page, mem_cont,
864                                vm_flags);
865        else if (PageAnon(page))
866            referenced += page_referenced_anon(page, mem_cont,
867                                vm_flags);
868        else if (page->mapping)
869            referenced += page_referenced_file(page, mem_cont,
870                                vm_flags);
871        if (we_locked)
872            unlock_page(page);
873    }
874out:
875    if (page_test_and_clear_young(page_to_pfn(page)))
876        referenced++;
877
878    return referenced;
879}
880
881static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
882                unsigned long address)
883{
884    struct mm_struct *mm = vma->vm_mm;
885    pte_t *pte;
886    spinlock_t *ptl;
887    int ret = 0;
888
889    pte = page_check_address(page, mm, address, &ptl, 1);
890    if (!pte)
891        goto out;
892
893    if (pte_dirty(*pte) || pte_write(*pte)) {
894        pte_t entry;
895
896        flush_cache_page(vma, address, pte_pfn(*pte));
897        entry = ptep_clear_flush_notify(vma, address, pte);
898        entry = pte_wrprotect(entry);
899        entry = pte_mkclean(entry);
900        set_pte_at(mm, address, pte, entry);
901        ret = 1;
902    }
903
904    pte_unmap_unlock(pte, ptl);
905out:
906    return ret;
907}
908
909static int page_mkclean_file(struct address_space *mapping, struct page *page)
910{
911    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
912    struct vm_area_struct *vma;
913    struct prio_tree_iter iter;
914    int ret = 0;
915
916    BUG_ON(PageAnon(page));
917
918    mutex_lock(&mapping->i_mmap_mutex);
919    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
920        if (vma->vm_flags & VM_SHARED) {
921            unsigned long address = vma_address(page, vma);
922            if (address == -EFAULT)
923                continue;
924            ret += page_mkclean_one(page, vma, address);
925        }
926    }
927    mutex_unlock(&mapping->i_mmap_mutex);
928    return ret;
929}
930
931int page_mkclean(struct page *page)
932{
933    int ret = 0;
934
935    BUG_ON(!PageLocked(page));
936
937    if (page_mapped(page)) {
938        struct address_space *mapping = page_mapping(page);
939        if (mapping) {
940            ret = page_mkclean_file(mapping, page);
941            if (page_test_and_clear_dirty(page_to_pfn(page), 1))
942                ret = 1;
943        }
944    }
945
946    return ret;
947}
948EXPORT_SYMBOL_GPL(page_mkclean);
949
950/**
951 * page_move_anon_rmap - move a page to our anon_vma
952 * @page: the page to move to our anon_vma
953 * @vma: the vma the page belongs to
954 * @address: the user virtual address mapped
955 *
956 * When a page belongs exclusively to one process after a COW event,
957 * that page can be moved into the anon_vma that belongs to just that
958 * process, so the rmap code will not search the parent or sibling
959 * processes.
960 */
961void page_move_anon_rmap(struct page *page,
962    struct vm_area_struct *vma, unsigned long address)
963{
964    struct anon_vma *anon_vma = vma->anon_vma;
965
966    VM_BUG_ON(!PageLocked(page));
967    VM_BUG_ON(!anon_vma);
968    VM_BUG_ON(page->index != linear_page_index(vma, address));
969
970    anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
971    page->mapping = (struct address_space *) anon_vma;
972}
973
974/**
975 * __page_set_anon_rmap - set up new anonymous rmap
976 * @page: Page to add to rmap
977 * @vma: VM area to add page to.
978 * @address: User virtual address of the mapping
979 * @exclusive: the page is exclusively owned by the current process
980 */
981static void __page_set_anon_rmap(struct page *page,
982    struct vm_area_struct *vma, unsigned long address, int exclusive)
983{
984    struct anon_vma *anon_vma = vma->anon_vma;
985
986    BUG_ON(!anon_vma);
987
988    if (PageAnon(page))
989        return;
990
991    /*
992     * If the page isn't exclusively mapped into this vma,
993     * we must use the _oldest_ possible anon_vma for the
994     * page mapping!
995     */
996    if (!exclusive)
997        anon_vma = anon_vma->root;
998
999    anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1000    page->mapping = (struct address_space *) anon_vma;
1001    page->index = linear_page_index(vma, address);
1002}
1003
1004/**
1005 * __page_check_anon_rmap - sanity check anonymous rmap addition
1006 * @page: the page to add the mapping to
1007 * @vma: the vm area in which the mapping is added
1008 * @address: the user virtual address mapped
1009 */
1010static void __page_check_anon_rmap(struct page *page,
1011    struct vm_area_struct *vma, unsigned long address)
1012{
1013#ifdef CONFIG_DEBUG_VM
1014    /*
1015     * The page's anon-rmap details (mapping and index) are guaranteed to
1016     * be set up correctly at this point.
1017     *
1018     * We have exclusion against page_add_anon_rmap because the caller
1019     * always holds the page locked, except if called from page_dup_rmap,
1020     * in which case the page is already known to be setup.
1021     *
1022     * We have exclusion against page_add_new_anon_rmap because those pages
1023     * are initially only visible via the pagetables, and the pte is locked
1024     * over the call to page_add_new_anon_rmap.
1025     */
1026    BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1027    BUG_ON(page->index != linear_page_index(vma, address));
1028#endif
1029}
1030
1031/**
1032 * page_add_anon_rmap - add pte mapping to an anonymous page
1033 * @page: the page to add the mapping to
1034 * @vma: the vm area in which the mapping is added
1035 * @address: the user virtual address mapped
1036 *
1037 * The caller needs to hold the pte lock, and the page must be locked in
1038 * the anon_vma case: to serialize mapping,index checking after setting,
1039 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1040 * (but PageKsm is never downgraded to PageAnon).
1041 */
1042void page_add_anon_rmap(struct page *page,
1043    struct vm_area_struct *vma, unsigned long address)
1044{
1045    do_page_add_anon_rmap(page, vma, address, 0);
1046}
1047
1048/*
1049 * Special version of the above for do_swap_page, which often runs
1050 * into pages that are exclusively owned by the current process.
1051 * Everybody else should continue to use page_add_anon_rmap above.
1052 */
1053void do_page_add_anon_rmap(struct page *page,
1054    struct vm_area_struct *vma, unsigned long address, int exclusive)
1055{
1056    int first = atomic_inc_and_test(&page->_mapcount);
1057    if (first) {
1058        if (!PageTransHuge(page))
1059            __inc_zone_page_state(page, NR_ANON_PAGES);
1060        else
1061            __inc_zone_page_state(page,
1062                          NR_ANON_TRANSPARENT_HUGEPAGES);
1063    }
1064    if (unlikely(PageKsm(page)))
1065        return;
1066
1067    VM_BUG_ON(!PageLocked(page));
1068    /* address might be in next vma when migration races vma_adjust */
1069    if (first)
1070        __page_set_anon_rmap(page, vma, address, exclusive);
1071    else
1072        __page_check_anon_rmap(page, vma, address);
1073}
1074
1075/**
1076 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1077 * @page: the page to add the mapping to
1078 * @vma: the vm area in which the mapping is added
1079 * @address: the user virtual address mapped
1080 *
1081 * Same as page_add_anon_rmap but must only be called on *new* pages.
1082 * This means the inc-and-test can be bypassed.
1083 * Page does not have to be locked.
1084 */
1085void page_add_new_anon_rmap(struct page *page,
1086    struct vm_area_struct *vma, unsigned long address)
1087{
1088    VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1089    SetPageSwapBacked(page);
1090    atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1091    if (!PageTransHuge(page))
1092        __inc_zone_page_state(page, NR_ANON_PAGES);
1093    else
1094        __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1095    __page_set_anon_rmap(page, vma, address, 1);
1096    if (page_evictable(page, vma))
1097        lru_cache_add_lru(page, LRU_ACTIVE_ANON);
1098    else
1099        add_page_to_unevictable_list(page);
1100}
1101
1102/**
1103 * page_add_file_rmap - add pte mapping to a file page
1104 * @page: the page to add the mapping to
1105 *
1106 * The caller needs to hold the pte lock.
1107 */
1108void page_add_file_rmap(struct page *page)
1109{
1110    if (atomic_inc_and_test(&page->_mapcount)) {
1111        __inc_zone_page_state(page, NR_FILE_MAPPED);
1112        mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
1113    }
1114}
1115
1116/**
1117 * page_remove_rmap - take down pte mapping from a page
1118 * @page: page to remove mapping from
1119 *
1120 * The caller needs to hold the pte lock.
1121 */
1122void page_remove_rmap(struct page *page)
1123{
1124    /* page still mapped by someone else? */
1125    if (!atomic_add_negative(-1, &page->_mapcount))
1126        return;
1127
1128    /*
1129     * Now that the last pte has gone, s390 must transfer dirty
1130     * flag from storage key to struct page. We can usually skip
1131     * this if the page is anon, so about to be freed; but perhaps
1132     * not if it's in swapcache - there might be another pte slot
1133     * containing the swap entry, but page not yet written to swap.
1134     */
1135    if ((!PageAnon(page) || PageSwapCache(page)) &&
1136        page_test_and_clear_dirty(page_to_pfn(page), 1))
1137        set_page_dirty(page);
1138    /*
1139     * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1140     * and not charged by memcg for now.
1141     */
1142    if (unlikely(PageHuge(page)))
1143        return;
1144    if (PageAnon(page)) {
1145        mem_cgroup_uncharge_page(page);
1146        if (!PageTransHuge(page))
1147            __dec_zone_page_state(page, NR_ANON_PAGES);
1148        else
1149            __dec_zone_page_state(page,
1150                          NR_ANON_TRANSPARENT_HUGEPAGES);
1151    } else {
1152        __dec_zone_page_state(page, NR_FILE_MAPPED);
1153        mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
1154    }
1155    /*
1156     * It would be tidy to reset the PageAnon mapping here,
1157     * but that might overwrite a racing page_add_anon_rmap
1158     * which increments mapcount after us but sets mapping
1159     * before us: so leave the reset to free_hot_cold_page,
1160     * and remember that it's only reliable while mapped.
1161     * Leaving it set also helps swapoff to reinstate ptes
1162     * faster for those pages still in swapcache.
1163     */
1164}
1165
1166/*
1167 * Subfunctions of try_to_unmap: try_to_unmap_one called
1168 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
1169 */
1170int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1171             unsigned long address, enum ttu_flags flags)
1172{
1173    struct mm_struct *mm = vma->vm_mm;
1174    pte_t *pte;
1175    pte_t pteval;
1176    spinlock_t *ptl;
1177    int ret = SWAP_AGAIN;
1178
1179    pte = page_check_address(page, mm, address, &ptl, 0);
1180    if (!pte)
1181        goto out;
1182
1183    /*
1184     * If the page is mlock()d, we cannot swap it out.
1185     * If it's recently referenced (perhaps page_referenced
1186     * skipped over this mm) then we should reactivate it.
1187     */
1188    if (!(flags & TTU_IGNORE_MLOCK)) {
1189        if (vma->vm_flags & VM_LOCKED)
1190            goto out_mlock;
1191
1192        if (TTU_ACTION(flags) == TTU_MUNLOCK)
1193            goto out_unmap;
1194    }
1195    if (!(flags & TTU_IGNORE_ACCESS)) {
1196        if (ptep_clear_flush_young_notify(vma, address, pte)) {
1197            ret = SWAP_FAIL;
1198            goto out_unmap;
1199        }
1200      }
1201
1202    /* Nuke the page table entry. */
1203    flush_cache_page(vma, address, page_to_pfn(page));
1204    pteval = ptep_clear_flush_notify(vma, address, pte);
1205
1206    /* Move the dirty bit to the physical page now the pte is gone. */
1207    if (pte_dirty(pteval))
1208        set_page_dirty(page);
1209
1210    /* Update high watermark before we lower rss */
1211    update_hiwater_rss(mm);
1212
1213    if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1214        if (PageAnon(page))
1215            dec_mm_counter(mm, MM_ANONPAGES);
1216        else
1217            dec_mm_counter(mm, MM_FILEPAGES);
1218        set_pte_at(mm, address, pte,
1219                swp_entry_to_pte(make_hwpoison_entry(page)));
1220    } else if (PageAnon(page)) {
1221        swp_entry_t entry = { .val = page_private(page) };
1222
1223        if (PageSwapCache(page)) {
1224            /*
1225             * Store the swap location in the pte.
1226             * See handle_pte_fault() ...
1227             */
1228            if (swap_duplicate(entry) < 0) {
1229                set_pte_at(mm, address, pte, pteval);
1230                ret = SWAP_FAIL;
1231                goto out_unmap;
1232            }
1233            if (list_empty(&mm->mmlist)) {
1234                spin_lock(&mmlist_lock);
1235                if (list_empty(&mm->mmlist))
1236                    list_add(&mm->mmlist, &init_mm.mmlist);
1237                spin_unlock(&mmlist_lock);
1238            }
1239            dec_mm_counter(mm, MM_ANONPAGES);
1240            inc_mm_counter(mm, MM_SWAPENTS);
1241        } else if (PAGE_MIGRATION) {
1242            /*
1243             * Store the pfn of the page in a special migration
1244             * pte. do_swap_page() will wait until the migration
1245             * pte is removed and then restart fault handling.
1246             */
1247            BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1248            entry = make_migration_entry(page, pte_write(pteval));
1249        }
1250        set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1251        BUG_ON(pte_file(*pte));
1252    } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
1253        /* Establish migration entry for a file page */
1254        swp_entry_t entry;
1255        entry = make_migration_entry(page, pte_write(pteval));
1256        set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1257    } else
1258        dec_mm_counter(mm, MM_FILEPAGES);
1259
1260    page_remove_rmap(page);
1261    page_cache_release(page);
1262
1263out_unmap:
1264    pte_unmap_unlock(pte, ptl);
1265out:
1266    return ret;
1267
1268out_mlock:
1269    pte_unmap_unlock(pte, ptl);
1270
1271
1272    /*
1273     * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1274     * unstable result and race. Plus, We can't wait here because
1275     * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1276     * if trylock failed, the page remain in evictable lru and later
1277     * vmscan could retry to move the page to unevictable lru if the
1278     * page is actually mlocked.
1279     */
1280    if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1281        if (vma->vm_flags & VM_LOCKED) {
1282            mlock_vma_page(page);
1283            ret = SWAP_MLOCK;
1284        }
1285        up_read(&vma->vm_mm->mmap_sem);
1286    }
1287    return ret;
1288}
1289
1290/*
1291 * objrmap doesn't work for nonlinear VMAs because the assumption that
1292 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1293 * Consequently, given a particular page and its ->index, we cannot locate the
1294 * ptes which are mapping that page without an exhaustive linear search.
1295 *
1296 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1297 * maps the file to which the target page belongs. The ->vm_private_data field
1298 * holds the current cursor into that scan. Successive searches will circulate
1299 * around the vma's virtual address space.
1300 *
1301 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1302 * more scanning pressure is placed against them as well. Eventually pages
1303 * will become fully unmapped and are eligible for eviction.
1304 *
1305 * For very sparsely populated VMAs this is a little inefficient - chances are
1306 * there there won't be many ptes located within the scan cluster. In this case
1307 * maybe we could scan further - to the end of the pte page, perhaps.
1308 *
1309 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1310 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1311 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1312 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1313 */
1314#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1315#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1316
1317static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1318        struct vm_area_struct *vma, struct page *check_page)
1319{
1320    struct mm_struct *mm = vma->vm_mm;
1321    pgd_t *pgd;
1322    pud_t *pud;
1323    pmd_t *pmd;
1324    pte_t *pte;
1325    pte_t pteval;
1326    spinlock_t *ptl;
1327    struct page *page;
1328    unsigned long address;
1329    unsigned long end;
1330    int ret = SWAP_AGAIN;
1331    int locked_vma = 0;
1332
1333    address = (vma->vm_start + cursor) & CLUSTER_MASK;
1334    end = address + CLUSTER_SIZE;
1335    if (address < vma->vm_start)
1336        address = vma->vm_start;
1337    if (end > vma->vm_end)
1338        end = vma->vm_end;
1339
1340    pgd = pgd_offset(mm, address);
1341    if (!pgd_present(*pgd))
1342        return ret;
1343
1344    pud = pud_offset(pgd, address);
1345    if (!pud_present(*pud))
1346        return ret;
1347
1348    pmd = pmd_offset(pud, address);
1349    if (!pmd_present(*pmd))
1350        return ret;
1351
1352    /*
1353     * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1354     * keep the sem while scanning the cluster for mlocking pages.
1355     */
1356    if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1357        locked_vma = (vma->vm_flags & VM_LOCKED);
1358        if (!locked_vma)
1359            up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1360    }
1361
1362    pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1363
1364    /* Update high watermark before we lower rss */
1365    update_hiwater_rss(mm);
1366
1367    for (; address < end; pte++, address += PAGE_SIZE) {
1368        if (!pte_present(*pte))
1369            continue;
1370        page = vm_normal_page(vma, address, *pte);
1371        BUG_ON(!page || PageAnon(page));
1372
1373        if (locked_vma) {
1374            mlock_vma_page(page); /* no-op if already mlocked */
1375            if (page == check_page)
1376                ret = SWAP_MLOCK;
1377            continue; /* don't unmap */
1378        }
1379
1380        if (ptep_clear_flush_young_notify(vma, address, pte))
1381            continue;
1382
1383        /* Nuke the page table entry. */
1384        flush_cache_page(vma, address, pte_pfn(*pte));
1385        pteval = ptep_clear_flush_notify(vma, address, pte);
1386
1387        /* If nonlinear, store the file page offset in the pte. */
1388        if (page->index != linear_page_index(vma, address))
1389            set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1390
1391        /* Move the dirty bit to the physical page now the pte is gone. */
1392        if (pte_dirty(pteval))
1393            set_page_dirty(page);
1394
1395        page_remove_rmap(page);
1396        page_cache_release(page);
1397        dec_mm_counter(mm, MM_FILEPAGES);
1398        (*mapcount)--;
1399    }
1400    pte_unmap_unlock(pte - 1, ptl);
1401    if (locked_vma)
1402        up_read(&vma->vm_mm->mmap_sem);
1403    return ret;
1404}
1405
1406bool is_vma_temporary_stack(struct vm_area_struct *vma)
1407{
1408    int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1409
1410    if (!maybe_stack)
1411        return false;
1412
1413    if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1414                        VM_STACK_INCOMPLETE_SETUP)
1415        return true;
1416
1417    return false;
1418}
1419
1420/**
1421 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1422 * rmap method
1423 * @page: the page to unmap/unlock
1424 * @flags: action and flags
1425 *
1426 * Find all the mappings of a page using the mapping pointer and the vma chains
1427 * contained in the anon_vma struct it points to.
1428 *
1429 * This function is only called from try_to_unmap/try_to_munlock for
1430 * anonymous pages.
1431 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1432 * where the page was found will be held for write. So, we won't recheck
1433 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1434 * 'LOCKED.
1435 */
1436static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1437{
1438    struct anon_vma *anon_vma;
1439    struct anon_vma_chain *avc;
1440    int ret = SWAP_AGAIN;
1441
1442    anon_vma = page_lock_anon_vma(page);
1443    if (!anon_vma)
1444        return ret;
1445
1446    list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1447        struct vm_area_struct *vma = avc->vma;
1448        unsigned long address;
1449
1450        /*
1451         * During exec, a temporary VMA is setup and later moved.
1452         * The VMA is moved under the anon_vma lock but not the
1453         * page tables leading to a race where migration cannot
1454         * find the migration ptes. Rather than increasing the
1455         * locking requirements of exec(), migration skips
1456         * temporary VMAs until after exec() completes.
1457         */
1458        if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1459                is_vma_temporary_stack(vma))
1460            continue;
1461
1462        address = vma_address(page, vma);
1463        if (address == -EFAULT)
1464            continue;
1465        ret = try_to_unmap_one(page, vma, address, flags);
1466        if (ret != SWAP_AGAIN || !page_mapped(page))
1467            break;
1468    }
1469
1470    page_unlock_anon_vma(anon_vma);
1471    return ret;
1472}
1473
1474/**
1475 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1476 * @page: the page to unmap/unlock
1477 * @flags: action and flags
1478 *
1479 * Find all the mappings of a page using the mapping pointer and the vma chains
1480 * contained in the address_space struct it points to.
1481 *
1482 * This function is only called from try_to_unmap/try_to_munlock for
1483 * object-based pages.
1484 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1485 * where the page was found will be held for write. So, we won't recheck
1486 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1487 * 'LOCKED.
1488 */
1489static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1490{
1491    struct address_space *mapping = page->mapping;
1492    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1493    struct vm_area_struct *vma;
1494    struct prio_tree_iter iter;
1495    int ret = SWAP_AGAIN;
1496    unsigned long cursor;
1497    unsigned long max_nl_cursor = 0;
1498    unsigned long max_nl_size = 0;
1499    unsigned int mapcount;
1500
1501    mutex_lock(&mapping->i_mmap_mutex);
1502    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1503        unsigned long address = vma_address(page, vma);
1504        if (address == -EFAULT)
1505            continue;
1506        ret = try_to_unmap_one(page, vma, address, flags);
1507        if (ret != SWAP_AGAIN || !page_mapped(page))
1508            goto out;
1509    }
1510
1511    if (list_empty(&mapping->i_mmap_nonlinear))
1512        goto out;
1513
1514    /*
1515     * We don't bother to try to find the munlocked page in nonlinears.
1516     * It's costly. Instead, later, page reclaim logic may call
1517     * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1518     */
1519    if (TTU_ACTION(flags) == TTU_MUNLOCK)
1520        goto out;
1521
1522    list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1523                        shared.vm_set.list) {
1524        cursor = (unsigned long) vma->vm_private_data;
1525        if (cursor > max_nl_cursor)
1526            max_nl_cursor = cursor;
1527        cursor = vma->vm_end - vma->vm_start;
1528        if (cursor > max_nl_size)
1529            max_nl_size = cursor;
1530    }
1531
1532    if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1533        ret = SWAP_FAIL;
1534        goto out;
1535    }
1536
1537    /*
1538     * We don't try to search for this page in the nonlinear vmas,
1539     * and page_referenced wouldn't have found it anyway. Instead
1540     * just walk the nonlinear vmas trying to age and unmap some.
1541     * The mapcount of the page we came in with is irrelevant,
1542     * but even so use it as a guide to how hard we should try?
1543     */
1544    mapcount = page_mapcount(page);
1545    if (!mapcount)
1546        goto out;
1547    cond_resched();
1548
1549    max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1550    if (max_nl_cursor == 0)
1551        max_nl_cursor = CLUSTER_SIZE;
1552
1553    do {
1554        list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1555                        shared.vm_set.list) {
1556            cursor = (unsigned long) vma->vm_private_data;
1557            while ( cursor < max_nl_cursor &&
1558                cursor < vma->vm_end - vma->vm_start) {
1559                if (try_to_unmap_cluster(cursor, &mapcount,
1560                        vma, page) == SWAP_MLOCK)
1561                    ret = SWAP_MLOCK;
1562                cursor += CLUSTER_SIZE;
1563                vma->vm_private_data = (void *) cursor;
1564                if ((int)mapcount <= 0)
1565                    goto out;
1566            }
1567            vma->vm_private_data = (void *) max_nl_cursor;
1568        }
1569        cond_resched();
1570        max_nl_cursor += CLUSTER_SIZE;
1571    } while (max_nl_cursor <= max_nl_size);
1572
1573    /*
1574     * Don't loop forever (perhaps all the remaining pages are
1575     * in locked vmas). Reset cursor on all unreserved nonlinear
1576     * vmas, now forgetting on which ones it had fallen behind.
1577     */
1578    list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1579        vma->vm_private_data = NULL;
1580out:
1581    mutex_unlock(&mapping->i_mmap_mutex);
1582    return ret;
1583}
1584
1585/**
1586 * try_to_unmap - try to remove all page table mappings to a page
1587 * @page: the page to get unmapped
1588 * @flags: action and flags
1589 *
1590 * Tries to remove all the page table entries which are mapping this
1591 * page, used in the pageout path. Caller must hold the page lock.
1592 * Return values are:
1593 *
1594 * SWAP_SUCCESS - we succeeded in removing all mappings
1595 * SWAP_AGAIN - we missed a mapping, try again later
1596 * SWAP_FAIL - the page is unswappable
1597 * SWAP_MLOCK - page is mlocked.
1598 */
1599int try_to_unmap(struct page *page, enum ttu_flags flags)
1600{
1601    int ret;
1602
1603    BUG_ON(!PageLocked(page));
1604    VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
1605
1606    if (unlikely(PageKsm(page)))
1607        ret = try_to_unmap_ksm(page, flags);
1608    else if (PageAnon(page))
1609        ret = try_to_unmap_anon(page, flags);
1610    else
1611        ret = try_to_unmap_file(page, flags);
1612    if (ret != SWAP_MLOCK && !page_mapped(page))
1613        ret = SWAP_SUCCESS;
1614    return ret;
1615}
1616
1617/**
1618 * try_to_munlock - try to munlock a page
1619 * @page: the page to be munlocked
1620 *
1621 * Called from munlock code. Checks all of the VMAs mapping the page
1622 * to make sure nobody else has this page mlocked. The page will be
1623 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1624 *
1625 * Return values are:
1626 *
1627 * SWAP_AGAIN - no vma is holding page mlocked, or,
1628 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1629 * SWAP_FAIL - page cannot be located at present
1630 * SWAP_MLOCK - page is now mlocked.
1631 */
1632int try_to_munlock(struct page *page)
1633{
1634    VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1635
1636    if (unlikely(PageKsm(page)))
1637        return try_to_unmap_ksm(page, TTU_MUNLOCK);
1638    else if (PageAnon(page))
1639        return try_to_unmap_anon(page, TTU_MUNLOCK);
1640    else
1641        return try_to_unmap_file(page, TTU_MUNLOCK);
1642}
1643
1644void __put_anon_vma(struct anon_vma *anon_vma)
1645{
1646    struct anon_vma *root = anon_vma->root;
1647
1648    if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1649        anon_vma_free(root);
1650
1651    anon_vma_free(anon_vma);
1652}
1653
1654#ifdef CONFIG_MIGRATION
1655/*
1656 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1657 * Called by migrate.c to remove migration ptes, but might be used more later.
1658 */
1659static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1660        struct vm_area_struct *, unsigned long, void *), void *arg)
1661{
1662    struct anon_vma *anon_vma;
1663    struct anon_vma_chain *avc;
1664    int ret = SWAP_AGAIN;
1665
1666    /*
1667     * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1668     * because that depends on page_mapped(); but not all its usages
1669     * are holding mmap_sem. Users without mmap_sem are required to
1670     * take a reference count to prevent the anon_vma disappearing
1671     */
1672    anon_vma = page_anon_vma(page);
1673    if (!anon_vma)
1674        return ret;
1675    anon_vma_lock(anon_vma);
1676    list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1677        struct vm_area_struct *vma = avc->vma;
1678        unsigned long address = vma_address(page, vma);
1679        if (address == -EFAULT)
1680            continue;
1681        ret = rmap_one(page, vma, address, arg);
1682        if (ret != SWAP_AGAIN)
1683            break;
1684    }
1685    anon_vma_unlock(anon_vma);
1686    return ret;
1687}
1688
1689static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1690        struct vm_area_struct *, unsigned long, void *), void *arg)
1691{
1692    struct address_space *mapping = page->mapping;
1693    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1694    struct vm_area_struct *vma;
1695    struct prio_tree_iter iter;
1696    int ret = SWAP_AGAIN;
1697
1698    if (!mapping)
1699        return ret;
1700    mutex_lock(&mapping->i_mmap_mutex);
1701    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1702        unsigned long address = vma_address(page, vma);
1703        if (address == -EFAULT)
1704            continue;
1705        ret = rmap_one(page, vma, address, arg);
1706        if (ret != SWAP_AGAIN)
1707            break;
1708    }
1709    /*
1710     * No nonlinear handling: being always shared, nonlinear vmas
1711     * never contain migration ptes. Decide what to do about this
1712     * limitation to linear when we need rmap_walk() on nonlinear.
1713     */
1714    mutex_unlock(&mapping->i_mmap_mutex);
1715    return ret;
1716}
1717
1718int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1719        struct vm_area_struct *, unsigned long, void *), void *arg)
1720{
1721    VM_BUG_ON(!PageLocked(page));
1722
1723    if (unlikely(PageKsm(page)))
1724        return rmap_walk_ksm(page, rmap_one, arg);
1725    else if (PageAnon(page))
1726        return rmap_walk_anon(page, rmap_one, arg);
1727    else
1728        return rmap_walk_file(page, rmap_one, arg);
1729}
1730#endif /* CONFIG_MIGRATION */
1731
1732#ifdef CONFIG_HUGETLB_PAGE
1733/*
1734 * The following three functions are for anonymous (private mapped) hugepages.
1735 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1736 * and no lru code, because we handle hugepages differently from common pages.
1737 */
1738static void __hugepage_set_anon_rmap(struct page *page,
1739    struct vm_area_struct *vma, unsigned long address, int exclusive)
1740{
1741    struct anon_vma *anon_vma = vma->anon_vma;
1742
1743    BUG_ON(!anon_vma);
1744
1745    if (PageAnon(page))
1746        return;
1747    if (!exclusive)
1748        anon_vma = anon_vma->root;
1749
1750    anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1751    page->mapping = (struct address_space *) anon_vma;
1752    page->index = linear_page_index(vma, address);
1753}
1754
1755void hugepage_add_anon_rmap(struct page *page,
1756                struct vm_area_struct *vma, unsigned long address)
1757{
1758    struct anon_vma *anon_vma = vma->anon_vma;
1759    int first;
1760
1761    BUG_ON(!PageLocked(page));
1762    BUG_ON(!anon_vma);
1763    /* address might be in next vma when migration races vma_adjust */
1764    first = atomic_inc_and_test(&page->_mapcount);
1765    if (first)
1766        __hugepage_set_anon_rmap(page, vma, address, 0);
1767}
1768
1769void hugepage_add_new_anon_rmap(struct page *page,
1770            struct vm_area_struct *vma, unsigned long address)
1771{
1772    BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1773    atomic_set(&page->_mapcount, 0);
1774    __hugepage_set_anon_rmap(page, vma, address, 1);
1775}
1776#endif /* CONFIG_HUGETLB_PAGE */
1777

Archive Download this file



interactive