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 * mm->mmap_sem
25 * page->flags PG_locked (lock_page)
26 * mapping->i_mmap_mutex
27 * anon_vma->mutex
28 * mm->page_table_lock or pte_lock
29 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
30 * swap_lock (in swap_duplicate, swap_info_get)
31 * mmlist_lock (in mmput, drain_mmlist and others)
32 * mapping->private_lock (in __set_page_dirty_buffers)
33 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
34 * bdi.wb->list_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 bdi.wb->list_lock in __sync_single_inode)
39 *
40 * anon_vma->mutex,mapping->i_mutex (memory_failure, collect_procs_anon)
41 * ->tasklist_lock
42 * pte map lock
43 */
44
45#include <linux/mm.h>
46#include <linux/pagemap.h>
47#include <linux/swap.h>
48#include <linux/swapops.h>
49#include <linux/slab.h>
50#include <linux/init.h>
51#include <linux/ksm.h>
52#include <linux/rmap.h>
53#include <linux/rcupdate.h>
54#include <linux/export.h>
55#include <linux/memcontrol.h>
56#include <linux/mmu_notifier.h>
57#include <linux/migrate.h>
58#include <linux/hugetlb.h>
59
60#include <asm/tlbflush.h>
61
62#include "internal.h"
63
64static struct kmem_cache *anon_vma_cachep;
65static struct kmem_cache *anon_vma_chain_cachep;
66
67static inline struct anon_vma *anon_vma_alloc(void)
68{
69    struct anon_vma *anon_vma;
70
71    anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
72    if (anon_vma) {
73        atomic_set(&anon_vma->refcount, 1);
74        /*
75         * Initialise the anon_vma root to point to itself. If called
76         * from fork, the root will be reset to the parents anon_vma.
77         */
78        anon_vma->root = anon_vma;
79    }
80
81    return anon_vma;
82}
83
84static inline void anon_vma_free(struct anon_vma *anon_vma)
85{
86    VM_BUG_ON(atomic_read(&anon_vma->refcount));
87
88    /*
89     * Synchronize against page_lock_anon_vma() such that
90     * we can safely hold the lock without the anon_vma getting
91     * freed.
92     *
93     * Relies on the full mb implied by the atomic_dec_and_test() from
94     * put_anon_vma() against the acquire barrier implied by
95     * mutex_trylock() from page_lock_anon_vma(). This orders:
96     *
97     * page_lock_anon_vma() VS put_anon_vma()
98     * mutex_trylock() atomic_dec_and_test()
99     * LOCK MB
100     * atomic_read() mutex_is_locked()
101     *
102     * LOCK should suffice since the actual taking of the lock must
103     * happen _before_ what follows.
104     */
105    if (mutex_is_locked(&anon_vma->root->mutex)) {
106        anon_vma_lock(anon_vma);
107        anon_vma_unlock(anon_vma);
108    }
109
110    kmem_cache_free(anon_vma_cachep, anon_vma);
111}
112
113static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
114{
115    return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
116}
117
118static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
119{
120    kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
121}
122
123static void anon_vma_chain_link(struct vm_area_struct *vma,
124                struct anon_vma_chain *avc,
125                struct anon_vma *anon_vma)
126{
127    avc->vma = vma;
128    avc->anon_vma = anon_vma;
129    list_add(&avc->same_vma, &vma->anon_vma_chain);
130
131    /*
132     * It's critical to add new vmas to the tail of the anon_vma,
133     * see comment in huge_memory.c:__split_huge_page().
134     */
135    list_add_tail(&avc->same_anon_vma, &anon_vma->head);
136}
137
138/**
139 * anon_vma_prepare - attach an anon_vma to a memory region
140 * @vma: the memory region in question
141 *
142 * This makes sure the memory mapping described by 'vma' has
143 * an 'anon_vma' attached to it, so that we can associate the
144 * anonymous pages mapped into it with that anon_vma.
145 *
146 * The common case will be that we already have one, but if
147 * not we either need to find an adjacent mapping that we
148 * can re-use the anon_vma from (very common when the only
149 * reason for splitting a vma has been mprotect()), or we
150 * allocate a new one.
151 *
152 * Anon-vma allocations are very subtle, because we may have
153 * optimistically looked up an anon_vma in page_lock_anon_vma()
154 * and that may actually touch the spinlock even in the newly
155 * allocated vma (it depends on RCU to make sure that the
156 * anon_vma isn't actually destroyed).
157 *
158 * As a result, we need to do proper anon_vma locking even
159 * for the new allocation. At the same time, we do not want
160 * to do any locking for the common case of already having
161 * an anon_vma.
162 *
163 * This must be called with the mmap_sem held for reading.
164 */
165int anon_vma_prepare(struct vm_area_struct *vma)
166{
167    struct anon_vma *anon_vma = vma->anon_vma;
168    struct anon_vma_chain *avc;
169
170    might_sleep();
171    if (unlikely(!anon_vma)) {
172        struct mm_struct *mm = vma->vm_mm;
173        struct anon_vma *allocated;
174
175        avc = anon_vma_chain_alloc(GFP_KERNEL);
176        if (!avc)
177            goto out_enomem;
178
179        anon_vma = find_mergeable_anon_vma(vma);
180        allocated = NULL;
181        if (!anon_vma) {
182            anon_vma = anon_vma_alloc();
183            if (unlikely(!anon_vma))
184                goto out_enomem_free_avc;
185            allocated = anon_vma;
186        }
187
188        anon_vma_lock(anon_vma);
189        /* page_table_lock to protect against threads */
190        spin_lock(&mm->page_table_lock);
191        if (likely(!vma->anon_vma)) {
192            vma->anon_vma = anon_vma;
193            anon_vma_chain_link(vma, avc, anon_vma);
194            allocated = NULL;
195            avc = NULL;
196        }
197        spin_unlock(&mm->page_table_lock);
198        anon_vma_unlock(anon_vma);
199
200        if (unlikely(allocated))
201            put_anon_vma(allocated);
202        if (unlikely(avc))
203            anon_vma_chain_free(avc);
204    }
205    return 0;
206
207 out_enomem_free_avc:
208    anon_vma_chain_free(avc);
209 out_enomem:
210    return -ENOMEM;
211}
212
213/*
214 * This is a useful helper function for locking the anon_vma root as
215 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
216 * have the same vma.
217 *
218 * Such anon_vma's should have the same root, so you'd expect to see
219 * just a single mutex_lock for the whole traversal.
220 */
221static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
222{
223    struct anon_vma *new_root = anon_vma->root;
224    if (new_root != root) {
225        if (WARN_ON_ONCE(root))
226            mutex_unlock(&root->mutex);
227        root = new_root;
228        mutex_lock(&root->mutex);
229    }
230    return root;
231}
232
233static inline void unlock_anon_vma_root(struct anon_vma *root)
234{
235    if (root)
236        mutex_unlock(&root->mutex);
237}
238
239/*
240 * Attach the anon_vmas from src to dst.
241 * Returns 0 on success, -ENOMEM on failure.
242 */
243int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
244{
245    struct anon_vma_chain *avc, *pavc;
246    struct anon_vma *root = NULL;
247
248    list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
249        struct anon_vma *anon_vma;
250
251        avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
252        if (unlikely(!avc)) {
253            unlock_anon_vma_root(root);
254            root = NULL;
255            avc = anon_vma_chain_alloc(GFP_KERNEL);
256            if (!avc)
257                goto enomem_failure;
258        }
259        anon_vma = pavc->anon_vma;
260        root = lock_anon_vma_root(root, anon_vma);
261        anon_vma_chain_link(dst, avc, anon_vma);
262    }
263    unlock_anon_vma_root(root);
264    return 0;
265
266 enomem_failure:
267    unlink_anon_vmas(dst);
268    return -ENOMEM;
269}
270
271/*
272 * Some rmap walk that needs to find all ptes/hugepmds without false
273 * negatives (like migrate and split_huge_page) running concurrent
274 * with operations that copy or move pagetables (like mremap() and
275 * fork()) to be safe. They depend on the anon_vma "same_anon_vma"
276 * list to be in a certain order: the dst_vma must be placed after the
277 * src_vma in the list. This is always guaranteed by fork() but
278 * mremap() needs to call this function to enforce it in case the
279 * dst_vma isn't newly allocated and chained with the anon_vma_clone()
280 * function but just an extension of a pre-existing vma through
281 * vma_merge.
282 *
283 * NOTE: the same_anon_vma list can still be changed by other
284 * processes while mremap runs because mremap doesn't hold the
285 * anon_vma mutex to prevent modifications to the list while it
286 * runs. All we need to enforce is that the relative order of this
287 * process vmas isn't changing (we don't care about other vmas
288 * order). Each vma corresponds to an anon_vma_chain structure so
289 * there's no risk that other processes calling anon_vma_moveto_tail()
290 * and changing the same_anon_vma list under mremap() will screw with
291 * the relative order of this process vmas in the list, because we
292 * they can't alter the order of any vma that belongs to this
293 * process. And there can't be another anon_vma_moveto_tail() running
294 * concurrently with mremap() coming from this process because we hold
295 * the mmap_sem for the whole mremap(). fork() ordering dependency
296 * also shouldn't be affected because fork() only cares that the
297 * parent vmas are placed in the list before the child vmas and
298 * anon_vma_moveto_tail() won't reorder vmas from either the fork()
299 * parent or child.
300 */
301void anon_vma_moveto_tail(struct vm_area_struct *dst)
302{
303    struct anon_vma_chain *pavc;
304    struct anon_vma *root = NULL;
305
306    list_for_each_entry_reverse(pavc, &dst->anon_vma_chain, same_vma) {
307        struct anon_vma *anon_vma = pavc->anon_vma;
308        VM_BUG_ON(pavc->vma != dst);
309        root = lock_anon_vma_root(root, anon_vma);
310        list_del(&pavc->same_anon_vma);
311        list_add_tail(&pavc->same_anon_vma, &anon_vma->head);
312    }
313    unlock_anon_vma_root(root);
314}
315
316/*
317 * Attach vma to its own anon_vma, as well as to the anon_vmas that
318 * the corresponding VMA in the parent process is attached to.
319 * Returns 0 on success, non-zero on failure.
320 */
321int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
322{
323    struct anon_vma_chain *avc;
324    struct anon_vma *anon_vma;
325
326    /* Don't bother if the parent process has no anon_vma here. */
327    if (!pvma->anon_vma)
328        return 0;
329
330    /*
331     * First, attach the new VMA to the parent VMA's anon_vmas,
332     * so rmap can find non-COWed pages in child processes.
333     */
334    if (anon_vma_clone(vma, pvma))
335        return -ENOMEM;
336
337    /* Then add our own anon_vma. */
338    anon_vma = anon_vma_alloc();
339    if (!anon_vma)
340        goto out_error;
341    avc = anon_vma_chain_alloc(GFP_KERNEL);
342    if (!avc)
343        goto out_error_free_anon_vma;
344
345    /*
346     * The root anon_vma's spinlock is the lock actually used when we
347     * lock any of the anon_vmas in this anon_vma tree.
348     */
349    anon_vma->root = pvma->anon_vma->root;
350    /*
351     * With refcounts, an anon_vma can stay around longer than the
352     * process it belongs to. The root anon_vma needs to be pinned until
353     * this anon_vma is freed, because the lock lives in the root.
354     */
355    get_anon_vma(anon_vma->root);
356    /* Mark this anon_vma as the one where our new (COWed) pages go. */
357    vma->anon_vma = anon_vma;
358    anon_vma_lock(anon_vma);
359    anon_vma_chain_link(vma, avc, anon_vma);
360    anon_vma_unlock(anon_vma);
361
362    return 0;
363
364 out_error_free_anon_vma:
365    put_anon_vma(anon_vma);
366 out_error:
367    unlink_anon_vmas(vma);
368    return -ENOMEM;
369}
370
371void unlink_anon_vmas(struct vm_area_struct *vma)
372{
373    struct anon_vma_chain *avc, *next;
374    struct anon_vma *root = NULL;
375
376    /*
377     * Unlink each anon_vma chained to the VMA. This list is ordered
378     * from newest to oldest, ensuring the root anon_vma gets freed last.
379     */
380    list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
381        struct anon_vma *anon_vma = avc->anon_vma;
382
383        root = lock_anon_vma_root(root, anon_vma);
384        list_del(&avc->same_anon_vma);
385
386        /*
387         * Leave empty anon_vmas on the list - we'll need
388         * to free them outside the lock.
389         */
390        if (list_empty(&anon_vma->head))
391            continue;
392
393        list_del(&avc->same_vma);
394        anon_vma_chain_free(avc);
395    }
396    unlock_anon_vma_root(root);
397
398    /*
399     * Iterate the list once more, it now only contains empty and unlinked
400     * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
401     * needing to acquire the anon_vma->root->mutex.
402     */
403    list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
404        struct anon_vma *anon_vma = avc->anon_vma;
405
406        put_anon_vma(anon_vma);
407
408        list_del(&avc->same_vma);
409        anon_vma_chain_free(avc);
410    }
411}
412
413static void anon_vma_ctor(void *data)
414{
415    struct anon_vma *anon_vma = data;
416
417    mutex_init(&anon_vma->mutex);
418    atomic_set(&anon_vma->refcount, 0);
419    INIT_LIST_HEAD(&anon_vma->head);
420}
421
422void __init anon_vma_init(void)
423{
424    anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
425            0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
426    anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
427}
428
429/*
430 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
431 *
432 * Since there is no serialization what so ever against page_remove_rmap()
433 * the best this function can do is return a locked anon_vma that might
434 * have been relevant to this page.
435 *
436 * The page might have been remapped to a different anon_vma or the anon_vma
437 * returned may already be freed (and even reused).
438 *
439 * In case it was remapped to a different anon_vma, the new anon_vma will be a
440 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
441 * ensure that any anon_vma obtained from the page will still be valid for as
442 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
443 *
444 * All users of this function must be very careful when walking the anon_vma
445 * chain and verify that the page in question is indeed mapped in it
446 * [ something equivalent to page_mapped_in_vma() ].
447 *
448 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
449 * that the anon_vma pointer from page->mapping is valid if there is a
450 * mapcount, we can dereference the anon_vma after observing those.
451 */
452struct anon_vma *page_get_anon_vma(struct page *page)
453{
454    struct anon_vma *anon_vma = NULL;
455    unsigned long anon_mapping;
456
457    rcu_read_lock();
458    anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
459    if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
460        goto out;
461    if (!page_mapped(page))
462        goto out;
463
464    anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
465    if (!atomic_inc_not_zero(&anon_vma->refcount)) {
466        anon_vma = NULL;
467        goto out;
468    }
469
470    /*
471     * If this page is still mapped, then its anon_vma cannot have been
472     * freed. But if it has been unmapped, we have no security against the
473     * anon_vma structure being freed and reused (for another anon_vma:
474     * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
475     * above cannot corrupt).
476     */
477    if (!page_mapped(page)) {
478        put_anon_vma(anon_vma);
479        anon_vma = NULL;
480    }
481out:
482    rcu_read_unlock();
483
484    return anon_vma;
485}
486
487/*
488 * Similar to page_get_anon_vma() except it locks the anon_vma.
489 *
490 * Its a little more complex as it tries to keep the fast path to a single
491 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
492 * reference like with page_get_anon_vma() and then block on the mutex.
493 */
494struct anon_vma *page_lock_anon_vma(struct page *page)
495{
496    struct anon_vma *anon_vma = NULL;
497    struct anon_vma *root_anon_vma;
498    unsigned long anon_mapping;
499
500    rcu_read_lock();
501    anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
502    if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
503        goto out;
504    if (!page_mapped(page))
505        goto out;
506
507    anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
508    root_anon_vma = ACCESS_ONCE(anon_vma->root);
509    if (mutex_trylock(&root_anon_vma->mutex)) {
510        /*
511         * If the page is still mapped, then this anon_vma is still
512         * its anon_vma, and holding the mutex ensures that it will
513         * not go away, see anon_vma_free().
514         */
515        if (!page_mapped(page)) {
516            mutex_unlock(&root_anon_vma->mutex);
517            anon_vma = NULL;
518        }
519        goto out;
520    }
521
522    /* trylock failed, we got to sleep */
523    if (!atomic_inc_not_zero(&anon_vma->refcount)) {
524        anon_vma = NULL;
525        goto out;
526    }
527
528    if (!page_mapped(page)) {
529        put_anon_vma(anon_vma);
530        anon_vma = NULL;
531        goto out;
532    }
533
534    /* we pinned the anon_vma, its safe to sleep */
535    rcu_read_unlock();
536    anon_vma_lock(anon_vma);
537
538    if (atomic_dec_and_test(&anon_vma->refcount)) {
539        /*
540         * Oops, we held the last refcount, release the lock
541         * and bail -- can't simply use put_anon_vma() because
542         * we'll deadlock on the anon_vma_lock() recursion.
543         */
544        anon_vma_unlock(anon_vma);
545        __put_anon_vma(anon_vma);
546        anon_vma = NULL;
547    }
548
549    return anon_vma;
550
551out:
552    rcu_read_unlock();
553    return anon_vma;
554}
555
556void page_unlock_anon_vma(struct anon_vma *anon_vma)
557{
558    anon_vma_unlock(anon_vma);
559}
560
561/*
562 * At what user virtual address is page expected in @vma?
563 * Returns virtual address or -EFAULT if page's index/offset is not
564 * within the range mapped the @vma.
565 */
566inline unsigned long
567vma_address(struct page *page, struct vm_area_struct *vma)
568{
569    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
570    unsigned long address;
571
572    if (unlikely(is_vm_hugetlb_page(vma)))
573        pgoff = page->index << huge_page_order(page_hstate(page));
574    address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
575    if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
576        /* page should be within @vma mapping range */
577        return -EFAULT;
578    }
579    return address;
580}
581
582/*
583 * At what user virtual address is page expected in vma?
584 * Caller should check the page is actually part of the vma.
585 */
586unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
587{
588    if (PageAnon(page)) {
589        struct anon_vma *page__anon_vma = page_anon_vma(page);
590        /*
591         * Note: swapoff's unuse_vma() is more efficient with this
592         * check, and needs it to match anon_vma when KSM is active.
593         */
594        if (!vma->anon_vma || !page__anon_vma ||
595            vma->anon_vma->root != page__anon_vma->root)
596            return -EFAULT;
597    } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
598        if (!vma->vm_file ||
599            vma->vm_file->f_mapping != page->mapping)
600            return -EFAULT;
601    } else
602        return -EFAULT;
603    return vma_address(page, vma);
604}
605
606/*
607 * Check that @page is mapped at @address into @mm.
608 *
609 * If @sync is false, page_check_address may perform a racy check to avoid
610 * the page table lock when the pte is not present (helpful when reclaiming
611 * highly shared pages).
612 *
613 * On success returns with pte mapped and locked.
614 */
615pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
616              unsigned long address, spinlock_t **ptlp, int sync)
617{
618    pgd_t *pgd;
619    pud_t *pud;
620    pmd_t *pmd;
621    pte_t *pte;
622    spinlock_t *ptl;
623
624    if (unlikely(PageHuge(page))) {
625        pte = huge_pte_offset(mm, address);
626        ptl = &mm->page_table_lock;
627        goto check;
628    }
629
630    pgd = pgd_offset(mm, address);
631    if (!pgd_present(*pgd))
632        return NULL;
633
634    pud = pud_offset(pgd, address);
635    if (!pud_present(*pud))
636        return NULL;
637
638    pmd = pmd_offset(pud, address);
639    if (!pmd_present(*pmd))
640        return NULL;
641    if (pmd_trans_huge(*pmd))
642        return NULL;
643
644    pte = pte_offset_map(pmd, address);
645    /* Make a quick check before getting the lock */
646    if (!sync && !pte_present(*pte)) {
647        pte_unmap(pte);
648        return NULL;
649    }
650
651    ptl = pte_lockptr(mm, pmd);
652check:
653    spin_lock(ptl);
654    if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
655        *ptlp = ptl;
656        return pte;
657    }
658    pte_unmap_unlock(pte, ptl);
659    return NULL;
660}
661
662/**
663 * page_mapped_in_vma - check whether a page is really mapped in a VMA
664 * @page: the page to test
665 * @vma: the VMA to test
666 *
667 * Returns 1 if the page is mapped into the page tables of the VMA, 0
668 * if the page is not mapped into the page tables of this VMA. Only
669 * valid for normal file or anonymous VMAs.
670 */
671int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
672{
673    unsigned long address;
674    pte_t *pte;
675    spinlock_t *ptl;
676
677    address = vma_address(page, vma);
678    if (address == -EFAULT) /* out of vma range */
679        return 0;
680    pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
681    if (!pte) /* the page is not in this mm */
682        return 0;
683    pte_unmap_unlock(pte, ptl);
684
685    return 1;
686}
687
688/*
689 * Subfunctions of page_referenced: page_referenced_one called
690 * repeatedly from either page_referenced_anon or page_referenced_file.
691 */
692int page_referenced_one(struct page *page, struct vm_area_struct *vma,
693            unsigned long address, unsigned int *mapcount,
694            unsigned long *vm_flags)
695{
696    struct mm_struct *mm = vma->vm_mm;
697    int referenced = 0;
698
699    if (unlikely(PageTransHuge(page))) {
700        pmd_t *pmd;
701
702        spin_lock(&mm->page_table_lock);
703        /*
704         * rmap might return false positives; we must filter
705         * these out using page_check_address_pmd().
706         */
707        pmd = page_check_address_pmd(page, mm, address,
708                         PAGE_CHECK_ADDRESS_PMD_FLAG);
709        if (!pmd) {
710            spin_unlock(&mm->page_table_lock);
711            goto out;
712        }
713
714        if (vma->vm_flags & VM_LOCKED) {
715            spin_unlock(&mm->page_table_lock);
716            *mapcount = 0; /* break early from loop */
717            *vm_flags |= VM_LOCKED;
718            goto out;
719        }
720
721        /* go ahead even if the pmd is pmd_trans_splitting() */
722        if (pmdp_clear_flush_young_notify(vma, address, pmd))
723            referenced++;
724        spin_unlock(&mm->page_table_lock);
725    } else {
726        pte_t *pte;
727        spinlock_t *ptl;
728
729        /*
730         * rmap might return false positives; we must filter
731         * these out using page_check_address().
732         */
733        pte = page_check_address(page, mm, address, &ptl, 0);
734        if (!pte)
735            goto out;
736
737        if (vma->vm_flags & VM_LOCKED) {
738            pte_unmap_unlock(pte, ptl);
739            *mapcount = 0; /* break early from loop */
740            *vm_flags |= VM_LOCKED;
741            goto out;
742        }
743
744        if (ptep_clear_flush_young_notify(vma, address, pte)) {
745            /*
746             * Don't treat a reference through a sequentially read
747             * mapping as such. If the page has been used in
748             * another mapping, we will catch it; if this other
749             * mapping is already gone, the unmap path will have
750             * set PG_referenced or activated the page.
751             */
752            if (likely(!VM_SequentialReadHint(vma)))
753                referenced++;
754        }
755        pte_unmap_unlock(pte, ptl);
756    }
757
758    /* Pretend the page is referenced if the task has the
759       swap token and is in the middle of a page fault. */
760    if (mm != current->mm && has_swap_token(mm) &&
761            rwsem_is_locked(&mm->mmap_sem))
762        referenced++;
763
764    (*mapcount)--;
765
766    if (referenced)
767        *vm_flags |= vma->vm_flags;
768out:
769    return referenced;
770}
771
772static int page_referenced_anon(struct page *page,
773                struct mem_cgroup *memcg,
774                unsigned long *vm_flags)
775{
776    unsigned int mapcount;
777    struct anon_vma *anon_vma;
778    struct anon_vma_chain *avc;
779    int referenced = 0;
780
781    anon_vma = page_lock_anon_vma(page);
782    if (!anon_vma)
783        return referenced;
784
785    mapcount = page_mapcount(page);
786    list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
787        struct vm_area_struct *vma = avc->vma;
788        unsigned long address = vma_address(page, vma);
789        if (address == -EFAULT)
790            continue;
791        /*
792         * If we are reclaiming on behalf of a cgroup, skip
793         * counting on behalf of references from different
794         * cgroups
795         */
796        if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
797            continue;
798        referenced += page_referenced_one(page, vma, address,
799                          &mapcount, vm_flags);
800        if (!mapcount)
801            break;
802    }
803
804    page_unlock_anon_vma(anon_vma);
805    return referenced;
806}
807
808/**
809 * page_referenced_file - referenced check for object-based rmap
810 * @page: the page we're checking references on.
811 * @memcg: target memory control group
812 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
813 *
814 * For an object-based mapped page, find all the places it is mapped and
815 * check/clear the referenced flag. This is done by following the page->mapping
816 * pointer, then walking the chain of vmas it holds. It returns the number
817 * of references it found.
818 *
819 * This function is only called from page_referenced for object-based pages.
820 */
821static int page_referenced_file(struct page *page,
822                struct mem_cgroup *memcg,
823                unsigned long *vm_flags)
824{
825    unsigned int mapcount;
826    struct address_space *mapping = page->mapping;
827    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
828    struct vm_area_struct *vma;
829    struct prio_tree_iter iter;
830    int referenced = 0;
831
832    /*
833     * The caller's checks on page->mapping and !PageAnon have made
834     * sure that this is a file page: the check for page->mapping
835     * excludes the case just before it gets set on an anon page.
836     */
837    BUG_ON(PageAnon(page));
838
839    /*
840     * The page lock not only makes sure that page->mapping cannot
841     * suddenly be NULLified by truncation, it makes sure that the
842     * structure at mapping cannot be freed and reused yet,
843     * so we can safely take mapping->i_mmap_mutex.
844     */
845    BUG_ON(!PageLocked(page));
846
847    mutex_lock(&mapping->i_mmap_mutex);
848
849    /*
850     * i_mmap_mutex does not stabilize mapcount at all, but mapcount
851     * is more likely to be accurate if we note it after spinning.
852     */
853    mapcount = page_mapcount(page);
854
855    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
856        unsigned long address = vma_address(page, vma);
857        if (address == -EFAULT)
858            continue;
859        /*
860         * If we are reclaiming on behalf of a cgroup, skip
861         * counting on behalf of references from different
862         * cgroups
863         */
864        if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
865            continue;
866        referenced += page_referenced_one(page, vma, address,
867                          &mapcount, vm_flags);
868        if (!mapcount)
869            break;
870    }
871
872    mutex_unlock(&mapping->i_mmap_mutex);
873    return referenced;
874}
875
876/**
877 * page_referenced - test if the page was referenced
878 * @page: the page to test
879 * @is_locked: caller holds lock on the page
880 * @memcg: target memory cgroup
881 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
882 *
883 * Quick test_and_clear_referenced for all mappings to a page,
884 * returns the number of ptes which referenced the page.
885 */
886int page_referenced(struct page *page,
887            int is_locked,
888            struct mem_cgroup *memcg,
889            unsigned long *vm_flags)
890{
891    int referenced = 0;
892    int we_locked = 0;
893
894    *vm_flags = 0;
895    if (page_mapped(page) && page_rmapping(page)) {
896        if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
897            we_locked = trylock_page(page);
898            if (!we_locked) {
899                referenced++;
900                goto out;
901            }
902        }
903        if (unlikely(PageKsm(page)))
904            referenced += page_referenced_ksm(page, memcg,
905                                vm_flags);
906        else if (PageAnon(page))
907            referenced += page_referenced_anon(page, memcg,
908                                vm_flags);
909        else if (page->mapping)
910            referenced += page_referenced_file(page, memcg,
911                                vm_flags);
912        if (we_locked)
913            unlock_page(page);
914
915        if (page_test_and_clear_young(page_to_pfn(page)))
916            referenced++;
917    }
918out:
919    return referenced;
920}
921
922static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
923                unsigned long address)
924{
925    struct mm_struct *mm = vma->vm_mm;
926    pte_t *pte;
927    spinlock_t *ptl;
928    int ret = 0;
929
930    pte = page_check_address(page, mm, address, &ptl, 1);
931    if (!pte)
932        goto out;
933
934    if (pte_dirty(*pte) || pte_write(*pte)) {
935        pte_t entry;
936
937        flush_cache_page(vma, address, pte_pfn(*pte));
938        entry = ptep_clear_flush_notify(vma, address, pte);
939        entry = pte_wrprotect(entry);
940        entry = pte_mkclean(entry);
941        set_pte_at(mm, address, pte, entry);
942        ret = 1;
943    }
944
945    pte_unmap_unlock(pte, ptl);
946out:
947    return ret;
948}
949
950static int page_mkclean_file(struct address_space *mapping, struct page *page)
951{
952    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
953    struct vm_area_struct *vma;
954    struct prio_tree_iter iter;
955    int ret = 0;
956
957    BUG_ON(PageAnon(page));
958
959    mutex_lock(&mapping->i_mmap_mutex);
960    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
961        if (vma->vm_flags & VM_SHARED) {
962            unsigned long address = vma_address(page, vma);
963            if (address == -EFAULT)
964                continue;
965            ret += page_mkclean_one(page, vma, address);
966        }
967    }
968    mutex_unlock(&mapping->i_mmap_mutex);
969    return ret;
970}
971
972int page_mkclean(struct page *page)
973{
974    int ret = 0;
975
976    BUG_ON(!PageLocked(page));
977
978    if (page_mapped(page)) {
979        struct address_space *mapping = page_mapping(page);
980        if (mapping) {
981            ret = page_mkclean_file(mapping, page);
982            if (page_test_and_clear_dirty(page_to_pfn(page), 1))
983                ret = 1;
984        }
985    }
986
987    return ret;
988}
989EXPORT_SYMBOL_GPL(page_mkclean);
990
991/**
992 * page_move_anon_rmap - move a page to our anon_vma
993 * @page: the page to move to our anon_vma
994 * @vma: the vma the page belongs to
995 * @address: the user virtual address mapped
996 *
997 * When a page belongs exclusively to one process after a COW event,
998 * that page can be moved into the anon_vma that belongs to just that
999 * process, so the rmap code will not search the parent or sibling
1000 * processes.
1001 */
1002void page_move_anon_rmap(struct page *page,
1003    struct vm_area_struct *vma, unsigned long address)
1004{
1005    struct anon_vma *anon_vma = vma->anon_vma;
1006
1007    VM_BUG_ON(!PageLocked(page));
1008    VM_BUG_ON(!anon_vma);
1009    VM_BUG_ON(page->index != linear_page_index(vma, address));
1010
1011    anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1012    page->mapping = (struct address_space *) anon_vma;
1013}
1014
1015/**
1016 * __page_set_anon_rmap - set up new anonymous rmap
1017 * @page: Page to add to rmap
1018 * @vma: VM area to add page to.
1019 * @address: User virtual address of the mapping
1020 * @exclusive: the page is exclusively owned by the current process
1021 */
1022static void __page_set_anon_rmap(struct page *page,
1023    struct vm_area_struct *vma, unsigned long address, int exclusive)
1024{
1025    struct anon_vma *anon_vma = vma->anon_vma;
1026
1027    BUG_ON(!anon_vma);
1028
1029    if (PageAnon(page))
1030        return;
1031
1032    /*
1033     * If the page isn't exclusively mapped into this vma,
1034     * we must use the _oldest_ possible anon_vma for the
1035     * page mapping!
1036     */
1037    if (!exclusive)
1038        anon_vma = anon_vma->root;
1039
1040    anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1041    page->mapping = (struct address_space *) anon_vma;
1042    page->index = linear_page_index(vma, address);
1043}
1044
1045/**
1046 * __page_check_anon_rmap - sanity check anonymous rmap addition
1047 * @page: the page to add the mapping to
1048 * @vma: the vm area in which the mapping is added
1049 * @address: the user virtual address mapped
1050 */
1051static void __page_check_anon_rmap(struct page *page,
1052    struct vm_area_struct *vma, unsigned long address)
1053{
1054#ifdef CONFIG_DEBUG_VM
1055    /*
1056     * The page's anon-rmap details (mapping and index) are guaranteed to
1057     * be set up correctly at this point.
1058     *
1059     * We have exclusion against page_add_anon_rmap because the caller
1060     * always holds the page locked, except if called from page_dup_rmap,
1061     * in which case the page is already known to be setup.
1062     *
1063     * We have exclusion against page_add_new_anon_rmap because those pages
1064     * are initially only visible via the pagetables, and the pte is locked
1065     * over the call to page_add_new_anon_rmap.
1066     */
1067    BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1068    BUG_ON(page->index != linear_page_index(vma, address));
1069#endif
1070}
1071
1072/**
1073 * page_add_anon_rmap - add pte mapping to an anonymous page
1074 * @page: the page to add the mapping to
1075 * @vma: the vm area in which the mapping is added
1076 * @address: the user virtual address mapped
1077 *
1078 * The caller needs to hold the pte lock, and the page must be locked in
1079 * the anon_vma case: to serialize mapping,index checking after setting,
1080 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1081 * (but PageKsm is never downgraded to PageAnon).
1082 */
1083void page_add_anon_rmap(struct page *page,
1084    struct vm_area_struct *vma, unsigned long address)
1085{
1086    do_page_add_anon_rmap(page, vma, address, 0);
1087}
1088
1089/*
1090 * Special version of the above for do_swap_page, which often runs
1091 * into pages that are exclusively owned by the current process.
1092 * Everybody else should continue to use page_add_anon_rmap above.
1093 */
1094void do_page_add_anon_rmap(struct page *page,
1095    struct vm_area_struct *vma, unsigned long address, int exclusive)
1096{
1097    int first = atomic_inc_and_test(&page->_mapcount);
1098    if (first) {
1099        if (!PageTransHuge(page))
1100            __inc_zone_page_state(page, NR_ANON_PAGES);
1101        else
1102            __inc_zone_page_state(page,
1103                          NR_ANON_TRANSPARENT_HUGEPAGES);
1104    }
1105    if (unlikely(PageKsm(page)))
1106        return;
1107
1108    VM_BUG_ON(!PageLocked(page));
1109    /* address might be in next vma when migration races vma_adjust */
1110    if (first)
1111        __page_set_anon_rmap(page, vma, address, exclusive);
1112    else
1113        __page_check_anon_rmap(page, vma, address);
1114}
1115
1116/**
1117 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1118 * @page: the page to add the mapping to
1119 * @vma: the vm area in which the mapping is added
1120 * @address: the user virtual address mapped
1121 *
1122 * Same as page_add_anon_rmap but must only be called on *new* pages.
1123 * This means the inc-and-test can be bypassed.
1124 * Page does not have to be locked.
1125 */
1126void page_add_new_anon_rmap(struct page *page,
1127    struct vm_area_struct *vma, unsigned long address)
1128{
1129    VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1130    SetPageSwapBacked(page);
1131    atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1132    if (!PageTransHuge(page))
1133        __inc_zone_page_state(page, NR_ANON_PAGES);
1134    else
1135        __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1136    __page_set_anon_rmap(page, vma, address, 1);
1137    if (page_evictable(page, vma))
1138        lru_cache_add_lru(page, LRU_ACTIVE_ANON);
1139    else
1140        add_page_to_unevictable_list(page);
1141}
1142
1143/**
1144 * page_add_file_rmap - add pte mapping to a file page
1145 * @page: the page to add the mapping to
1146 *
1147 * The caller needs to hold the pte lock.
1148 */
1149void page_add_file_rmap(struct page *page)
1150{
1151    bool locked;
1152    unsigned long flags;
1153
1154    mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1155    if (atomic_inc_and_test(&page->_mapcount)) {
1156        __inc_zone_page_state(page, NR_FILE_MAPPED);
1157        mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
1158    }
1159    mem_cgroup_end_update_page_stat(page, &locked, &flags);
1160}
1161
1162/**
1163 * page_remove_rmap - take down pte mapping from a page
1164 * @page: page to remove mapping from
1165 *
1166 * The caller needs to hold the pte lock.
1167 */
1168void page_remove_rmap(struct page *page)
1169{
1170    bool anon = PageAnon(page);
1171    bool locked;
1172    unsigned long flags;
1173
1174    /*
1175     * The anon case has no mem_cgroup page_stat to update; but may
1176     * uncharge_page() below, where the lock ordering can deadlock if
1177     * we hold the lock against page_stat move: so avoid it on anon.
1178     */
1179    if (!anon)
1180        mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1181
1182    /* page still mapped by someone else? */
1183    if (!atomic_add_negative(-1, &page->_mapcount))
1184        goto out;
1185
1186    /*
1187     * Now that the last pte has gone, s390 must transfer dirty
1188     * flag from storage key to struct page. We can usually skip
1189     * this if the page is anon, so about to be freed; but perhaps
1190     * not if it's in swapcache - there might be another pte slot
1191     * containing the swap entry, but page not yet written to swap.
1192     */
1193    if ((!anon || PageSwapCache(page)) &&
1194        page_test_and_clear_dirty(page_to_pfn(page), 1))
1195        set_page_dirty(page);
1196    /*
1197     * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1198     * and not charged by memcg for now.
1199     */
1200    if (unlikely(PageHuge(page)))
1201        goto out;
1202    if (anon) {
1203        mem_cgroup_uncharge_page(page);
1204        if (!PageTransHuge(page))
1205            __dec_zone_page_state(page, NR_ANON_PAGES);
1206        else
1207            __dec_zone_page_state(page,
1208                          NR_ANON_TRANSPARENT_HUGEPAGES);
1209    } else {
1210        __dec_zone_page_state(page, NR_FILE_MAPPED);
1211        mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
1212    }
1213    /*
1214     * It would be tidy to reset the PageAnon mapping here,
1215     * but that might overwrite a racing page_add_anon_rmap
1216     * which increments mapcount after us but sets mapping
1217     * before us: so leave the reset to free_hot_cold_page,
1218     * and remember that it's only reliable while mapped.
1219     * Leaving it set also helps swapoff to reinstate ptes
1220     * faster for those pages still in swapcache.
1221     */
1222out:
1223    if (!anon)
1224        mem_cgroup_end_update_page_stat(page, &locked, &flags);
1225}
1226
1227/*
1228 * Subfunctions of try_to_unmap: try_to_unmap_one called
1229 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
1230 */
1231int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1232             unsigned long address, enum ttu_flags flags)
1233{
1234    struct mm_struct *mm = vma->vm_mm;
1235    pte_t *pte;
1236    pte_t pteval;
1237    spinlock_t *ptl;
1238    int ret = SWAP_AGAIN;
1239
1240    pte = page_check_address(page, mm, address, &ptl, 0);
1241    if (!pte)
1242        goto out;
1243
1244    /*
1245     * If the page is mlock()d, we cannot swap it out.
1246     * If it's recently referenced (perhaps page_referenced
1247     * skipped over this mm) then we should reactivate it.
1248     */
1249    if (!(flags & TTU_IGNORE_MLOCK)) {
1250        if (vma->vm_flags & VM_LOCKED)
1251            goto out_mlock;
1252
1253        if (TTU_ACTION(flags) == TTU_MUNLOCK)
1254            goto out_unmap;
1255    }
1256    if (!(flags & TTU_IGNORE_ACCESS)) {
1257        if (ptep_clear_flush_young_notify(vma, address, pte)) {
1258            ret = SWAP_FAIL;
1259            goto out_unmap;
1260        }
1261      }
1262
1263    /* Nuke the page table entry. */
1264    flush_cache_page(vma, address, page_to_pfn(page));
1265    pteval = ptep_clear_flush_notify(vma, address, pte);
1266
1267    /* Move the dirty bit to the physical page now the pte is gone. */
1268    if (pte_dirty(pteval))
1269        set_page_dirty(page);
1270
1271    /* Update high watermark before we lower rss */
1272    update_hiwater_rss(mm);
1273
1274    if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1275        if (PageAnon(page))
1276            dec_mm_counter(mm, MM_ANONPAGES);
1277        else
1278            dec_mm_counter(mm, MM_FILEPAGES);
1279        set_pte_at(mm, address, pte,
1280                swp_entry_to_pte(make_hwpoison_entry(page)));
1281    } else if (PageAnon(page)) {
1282        swp_entry_t entry = { .val = page_private(page) };
1283
1284        if (PageSwapCache(page)) {
1285            /*
1286             * Store the swap location in the pte.
1287             * See handle_pte_fault() ...
1288             */
1289            if (swap_duplicate(entry) < 0) {
1290                set_pte_at(mm, address, pte, pteval);
1291                ret = SWAP_FAIL;
1292                goto out_unmap;
1293            }
1294            if (list_empty(&mm->mmlist)) {
1295                spin_lock(&mmlist_lock);
1296                if (list_empty(&mm->mmlist))
1297                    list_add(&mm->mmlist, &init_mm.mmlist);
1298                spin_unlock(&mmlist_lock);
1299            }
1300            dec_mm_counter(mm, MM_ANONPAGES);
1301            inc_mm_counter(mm, MM_SWAPENTS);
1302        } else if (IS_ENABLED(CONFIG_MIGRATION)) {
1303            /*
1304             * Store the pfn of the page in a special migration
1305             * pte. do_swap_page() will wait until the migration
1306             * pte is removed and then restart fault handling.
1307             */
1308            BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1309            entry = make_migration_entry(page, pte_write(pteval));
1310        }
1311        set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1312        BUG_ON(pte_file(*pte));
1313    } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1314           (TTU_ACTION(flags) == TTU_MIGRATION)) {
1315        /* Establish migration entry for a file page */
1316        swp_entry_t entry;
1317        entry = make_migration_entry(page, pte_write(pteval));
1318        set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1319    } else
1320        dec_mm_counter(mm, MM_FILEPAGES);
1321
1322    page_remove_rmap(page);
1323    page_cache_release(page);
1324
1325out_unmap:
1326    pte_unmap_unlock(pte, ptl);
1327out:
1328    return ret;
1329
1330out_mlock:
1331    pte_unmap_unlock(pte, ptl);
1332
1333
1334    /*
1335     * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1336     * unstable result and race. Plus, We can't wait here because
1337     * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1338     * if trylock failed, the page remain in evictable lru and later
1339     * vmscan could retry to move the page to unevictable lru if the
1340     * page is actually mlocked.
1341     */
1342    if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1343        if (vma->vm_flags & VM_LOCKED) {
1344            mlock_vma_page(page);
1345            ret = SWAP_MLOCK;
1346        }
1347        up_read(&vma->vm_mm->mmap_sem);
1348    }
1349    return ret;
1350}
1351
1352/*
1353 * objrmap doesn't work for nonlinear VMAs because the assumption that
1354 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1355 * Consequently, given a particular page and its ->index, we cannot locate the
1356 * ptes which are mapping that page without an exhaustive linear search.
1357 *
1358 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1359 * maps the file to which the target page belongs. The ->vm_private_data field
1360 * holds the current cursor into that scan. Successive searches will circulate
1361 * around the vma's virtual address space.
1362 *
1363 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1364 * more scanning pressure is placed against them as well. Eventually pages
1365 * will become fully unmapped and are eligible for eviction.
1366 *
1367 * For very sparsely populated VMAs this is a little inefficient - chances are
1368 * there there won't be many ptes located within the scan cluster. In this case
1369 * maybe we could scan further - to the end of the pte page, perhaps.
1370 *
1371 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1372 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1373 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1374 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1375 */
1376#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1377#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1378
1379static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1380        struct vm_area_struct *vma, struct page *check_page)
1381{
1382    struct mm_struct *mm = vma->vm_mm;
1383    pgd_t *pgd;
1384    pud_t *pud;
1385    pmd_t *pmd;
1386    pte_t *pte;
1387    pte_t pteval;
1388    spinlock_t *ptl;
1389    struct page *page;
1390    unsigned long address;
1391    unsigned long end;
1392    int ret = SWAP_AGAIN;
1393    int locked_vma = 0;
1394
1395    address = (vma->vm_start + cursor) & CLUSTER_MASK;
1396    end = address + CLUSTER_SIZE;
1397    if (address < vma->vm_start)
1398        address = vma->vm_start;
1399    if (end > vma->vm_end)
1400        end = vma->vm_end;
1401
1402    pgd = pgd_offset(mm, address);
1403    if (!pgd_present(*pgd))
1404        return ret;
1405
1406    pud = pud_offset(pgd, address);
1407    if (!pud_present(*pud))
1408        return ret;
1409
1410    pmd = pmd_offset(pud, address);
1411    if (!pmd_present(*pmd))
1412        return ret;
1413
1414    /*
1415     * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1416     * keep the sem while scanning the cluster for mlocking pages.
1417     */
1418    if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1419        locked_vma = (vma->vm_flags & VM_LOCKED);
1420        if (!locked_vma)
1421            up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1422    }
1423
1424    pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1425
1426    /* Update high watermark before we lower rss */
1427    update_hiwater_rss(mm);
1428
1429    for (; address < end; pte++, address += PAGE_SIZE) {
1430        if (!pte_present(*pte))
1431            continue;
1432        page = vm_normal_page(vma, address, *pte);
1433        BUG_ON(!page || PageAnon(page));
1434
1435        if (locked_vma) {
1436            mlock_vma_page(page); /* no-op if already mlocked */
1437            if (page == check_page)
1438                ret = SWAP_MLOCK;
1439            continue; /* don't unmap */
1440        }
1441
1442        if (ptep_clear_flush_young_notify(vma, address, pte))
1443            continue;
1444
1445        /* Nuke the page table entry. */
1446        flush_cache_page(vma, address, pte_pfn(*pte));
1447        pteval = ptep_clear_flush_notify(vma, address, pte);
1448
1449        /* If nonlinear, store the file page offset in the pte. */
1450        if (page->index != linear_page_index(vma, address))
1451            set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1452
1453        /* Move the dirty bit to the physical page now the pte is gone. */
1454        if (pte_dirty(pteval))
1455            set_page_dirty(page);
1456
1457        page_remove_rmap(page);
1458        page_cache_release(page);
1459        dec_mm_counter(mm, MM_FILEPAGES);
1460        (*mapcount)--;
1461    }
1462    pte_unmap_unlock(pte - 1, ptl);
1463    if (locked_vma)
1464        up_read(&vma->vm_mm->mmap_sem);
1465    return ret;
1466}
1467
1468bool is_vma_temporary_stack(struct vm_area_struct *vma)
1469{
1470    int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1471
1472    if (!maybe_stack)
1473        return false;
1474
1475    if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1476                        VM_STACK_INCOMPLETE_SETUP)
1477        return true;
1478
1479    return false;
1480}
1481
1482/**
1483 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1484 * rmap method
1485 * @page: the page to unmap/unlock
1486 * @flags: action and flags
1487 *
1488 * Find all the mappings of a page using the mapping pointer and the vma chains
1489 * contained in the anon_vma struct it points to.
1490 *
1491 * This function is only called from try_to_unmap/try_to_munlock for
1492 * anonymous pages.
1493 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1494 * where the page was found will be held for write. So, we won't recheck
1495 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1496 * 'LOCKED.
1497 */
1498static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1499{
1500    struct anon_vma *anon_vma;
1501    struct anon_vma_chain *avc;
1502    int ret = SWAP_AGAIN;
1503
1504    anon_vma = page_lock_anon_vma(page);
1505    if (!anon_vma)
1506        return ret;
1507
1508    list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1509        struct vm_area_struct *vma = avc->vma;
1510        unsigned long address;
1511
1512        /*
1513         * During exec, a temporary VMA is setup and later moved.
1514         * The VMA is moved under the anon_vma lock but not the
1515         * page tables leading to a race where migration cannot
1516         * find the migration ptes. Rather than increasing the
1517         * locking requirements of exec(), migration skips
1518         * temporary VMAs until after exec() completes.
1519         */
1520        if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1521                is_vma_temporary_stack(vma))
1522            continue;
1523
1524        address = vma_address(page, vma);
1525        if (address == -EFAULT)
1526            continue;
1527        ret = try_to_unmap_one(page, vma, address, flags);
1528        if (ret != SWAP_AGAIN || !page_mapped(page))
1529            break;
1530    }
1531
1532    page_unlock_anon_vma(anon_vma);
1533    return ret;
1534}
1535
1536/**
1537 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1538 * @page: the page to unmap/unlock
1539 * @flags: action and flags
1540 *
1541 * Find all the mappings of a page using the mapping pointer and the vma chains
1542 * contained in the address_space struct it points to.
1543 *
1544 * This function is only called from try_to_unmap/try_to_munlock for
1545 * object-based pages.
1546 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1547 * where the page was found will be held for write. So, we won't recheck
1548 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1549 * 'LOCKED.
1550 */
1551static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1552{
1553    struct address_space *mapping = page->mapping;
1554    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1555    struct vm_area_struct *vma;
1556    struct prio_tree_iter iter;
1557    int ret = SWAP_AGAIN;
1558    unsigned long cursor;
1559    unsigned long max_nl_cursor = 0;
1560    unsigned long max_nl_size = 0;
1561    unsigned int mapcount;
1562
1563    mutex_lock(&mapping->i_mmap_mutex);
1564    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1565        unsigned long address = vma_address(page, vma);
1566        if (address == -EFAULT)
1567            continue;
1568        ret = try_to_unmap_one(page, vma, address, flags);
1569        if (ret != SWAP_AGAIN || !page_mapped(page))
1570            goto out;
1571    }
1572
1573    if (list_empty(&mapping->i_mmap_nonlinear))
1574        goto out;
1575
1576    /*
1577     * We don't bother to try to find the munlocked page in nonlinears.
1578     * It's costly. Instead, later, page reclaim logic may call
1579     * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1580     */
1581    if (TTU_ACTION(flags) == TTU_MUNLOCK)
1582        goto out;
1583
1584    list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1585                        shared.vm_set.list) {
1586        cursor = (unsigned long) vma->vm_private_data;
1587        if (cursor > max_nl_cursor)
1588            max_nl_cursor = cursor;
1589        cursor = vma->vm_end - vma->vm_start;
1590        if (cursor > max_nl_size)
1591            max_nl_size = cursor;
1592    }
1593
1594    if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1595        ret = SWAP_FAIL;
1596        goto out;
1597    }
1598
1599    /*
1600     * We don't try to search for this page in the nonlinear vmas,
1601     * and page_referenced wouldn't have found it anyway. Instead
1602     * just walk the nonlinear vmas trying to age and unmap some.
1603     * The mapcount of the page we came in with is irrelevant,
1604     * but even so use it as a guide to how hard we should try?
1605     */
1606    mapcount = page_mapcount(page);
1607    if (!mapcount)
1608        goto out;
1609    cond_resched();
1610
1611    max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1612    if (max_nl_cursor == 0)
1613        max_nl_cursor = CLUSTER_SIZE;
1614
1615    do {
1616        list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1617                        shared.vm_set.list) {
1618            cursor = (unsigned long) vma->vm_private_data;
1619            while ( cursor < max_nl_cursor &&
1620                cursor < vma->vm_end - vma->vm_start) {
1621                if (try_to_unmap_cluster(cursor, &mapcount,
1622                        vma, page) == SWAP_MLOCK)
1623                    ret = SWAP_MLOCK;
1624                cursor += CLUSTER_SIZE;
1625                vma->vm_private_data = (void *) cursor;
1626                if ((int)mapcount <= 0)
1627                    goto out;
1628            }
1629            vma->vm_private_data = (void *) max_nl_cursor;
1630        }
1631        cond_resched();
1632        max_nl_cursor += CLUSTER_SIZE;
1633    } while (max_nl_cursor <= max_nl_size);
1634
1635    /*
1636     * Don't loop forever (perhaps all the remaining pages are
1637     * in locked vmas). Reset cursor on all unreserved nonlinear
1638     * vmas, now forgetting on which ones it had fallen behind.
1639     */
1640    list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1641        vma->vm_private_data = NULL;
1642out:
1643    mutex_unlock(&mapping->i_mmap_mutex);
1644    return ret;
1645}
1646
1647/**
1648 * try_to_unmap - try to remove all page table mappings to a page
1649 * @page: the page to get unmapped
1650 * @flags: action and flags
1651 *
1652 * Tries to remove all the page table entries which are mapping this
1653 * page, used in the pageout path. Caller must hold the page lock.
1654 * Return values are:
1655 *
1656 * SWAP_SUCCESS - we succeeded in removing all mappings
1657 * SWAP_AGAIN - we missed a mapping, try again later
1658 * SWAP_FAIL - the page is unswappable
1659 * SWAP_MLOCK - page is mlocked.
1660 */
1661int try_to_unmap(struct page *page, enum ttu_flags flags)
1662{
1663    int ret;
1664
1665    BUG_ON(!PageLocked(page));
1666    VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
1667
1668    if (unlikely(PageKsm(page)))
1669        ret = try_to_unmap_ksm(page, flags);
1670    else if (PageAnon(page))
1671        ret = try_to_unmap_anon(page, flags);
1672    else
1673        ret = try_to_unmap_file(page, flags);
1674    if (ret != SWAP_MLOCK && !page_mapped(page))
1675        ret = SWAP_SUCCESS;
1676    return ret;
1677}
1678
1679/**
1680 * try_to_munlock - try to munlock a page
1681 * @page: the page to be munlocked
1682 *
1683 * Called from munlock code. Checks all of the VMAs mapping the page
1684 * to make sure nobody else has this page mlocked. The page will be
1685 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1686 *
1687 * Return values are:
1688 *
1689 * SWAP_AGAIN - no vma is holding page mlocked, or,
1690 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1691 * SWAP_FAIL - page cannot be located at present
1692 * SWAP_MLOCK - page is now mlocked.
1693 */
1694int try_to_munlock(struct page *page)
1695{
1696    VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1697
1698    if (unlikely(PageKsm(page)))
1699        return try_to_unmap_ksm(page, TTU_MUNLOCK);
1700    else if (PageAnon(page))
1701        return try_to_unmap_anon(page, TTU_MUNLOCK);
1702    else
1703        return try_to_unmap_file(page, TTU_MUNLOCK);
1704}
1705
1706void __put_anon_vma(struct anon_vma *anon_vma)
1707{
1708    struct anon_vma *root = anon_vma->root;
1709
1710    if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1711        anon_vma_free(root);
1712
1713    anon_vma_free(anon_vma);
1714}
1715
1716#ifdef CONFIG_MIGRATION
1717/*
1718 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1719 * Called by migrate.c to remove migration ptes, but might be used more later.
1720 */
1721static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1722        struct vm_area_struct *, unsigned long, void *), void *arg)
1723{
1724    struct anon_vma *anon_vma;
1725    struct anon_vma_chain *avc;
1726    int ret = SWAP_AGAIN;
1727
1728    /*
1729     * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1730     * because that depends on page_mapped(); but not all its usages
1731     * are holding mmap_sem. Users without mmap_sem are required to
1732     * take a reference count to prevent the anon_vma disappearing
1733     */
1734    anon_vma = page_anon_vma(page);
1735    if (!anon_vma)
1736        return ret;
1737    anon_vma_lock(anon_vma);
1738    list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1739        struct vm_area_struct *vma = avc->vma;
1740        unsigned long address = vma_address(page, vma);
1741        if (address == -EFAULT)
1742            continue;
1743        ret = rmap_one(page, vma, address, arg);
1744        if (ret != SWAP_AGAIN)
1745            break;
1746    }
1747    anon_vma_unlock(anon_vma);
1748    return ret;
1749}
1750
1751static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1752        struct vm_area_struct *, unsigned long, void *), void *arg)
1753{
1754    struct address_space *mapping = page->mapping;
1755    pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1756    struct vm_area_struct *vma;
1757    struct prio_tree_iter iter;
1758    int ret = SWAP_AGAIN;
1759
1760    if (!mapping)
1761        return ret;
1762    mutex_lock(&mapping->i_mmap_mutex);
1763    vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1764        unsigned long address = vma_address(page, vma);
1765        if (address == -EFAULT)
1766            continue;
1767        ret = rmap_one(page, vma, address, arg);
1768        if (ret != SWAP_AGAIN)
1769            break;
1770    }
1771    /*
1772     * No nonlinear handling: being always shared, nonlinear vmas
1773     * never contain migration ptes. Decide what to do about this
1774     * limitation to linear when we need rmap_walk() on nonlinear.
1775     */
1776    mutex_unlock(&mapping->i_mmap_mutex);
1777    return ret;
1778}
1779
1780int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1781        struct vm_area_struct *, unsigned long, void *), void *arg)
1782{
1783    VM_BUG_ON(!PageLocked(page));
1784
1785    if (unlikely(PageKsm(page)))
1786        return rmap_walk_ksm(page, rmap_one, arg);
1787    else if (PageAnon(page))
1788        return rmap_walk_anon(page, rmap_one, arg);
1789    else
1790        return rmap_walk_file(page, rmap_one, arg);
1791}
1792#endif /* CONFIG_MIGRATION */
1793
1794#ifdef CONFIG_HUGETLB_PAGE
1795/*
1796 * The following three functions are for anonymous (private mapped) hugepages.
1797 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1798 * and no lru code, because we handle hugepages differently from common pages.
1799 */
1800static void __hugepage_set_anon_rmap(struct page *page,
1801    struct vm_area_struct *vma, unsigned long address, int exclusive)
1802{
1803    struct anon_vma *anon_vma = vma->anon_vma;
1804
1805    BUG_ON(!anon_vma);
1806
1807    if (PageAnon(page))
1808        return;
1809    if (!exclusive)
1810        anon_vma = anon_vma->root;
1811
1812    anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1813    page->mapping = (struct address_space *) anon_vma;
1814    page->index = linear_page_index(vma, address);
1815}
1816
1817void hugepage_add_anon_rmap(struct page *page,
1818                struct vm_area_struct *vma, unsigned long address)
1819{
1820    struct anon_vma *anon_vma = vma->anon_vma;
1821    int first;
1822
1823    BUG_ON(!PageLocked(page));
1824    BUG_ON(!anon_vma);
1825    /* address might be in next vma when migration races vma_adjust */
1826    first = atomic_inc_and_test(&page->_mapcount);
1827    if (first)
1828        __hugepage_set_anon_rmap(page, vma, address, 0);
1829}
1830
1831void hugepage_add_new_anon_rmap(struct page *page,
1832            struct vm_area_struct *vma, unsigned long address)
1833{
1834    BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1835    atomic_set(&page->_mapcount, 0);
1836    __hugepage_set_anon_rmap(page, vma, address, 1);
1837}
1838#endif /* CONFIG_HUGETLB_PAGE */
1839

Archive Download this file



interactive