]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - mm/rmap.c
Btrfs: fix device replace of a missing RAID 5/6 device
[mirror_ubuntu-artful-kernel.git] / 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_rwsem
27 * anon_vma->rwsem
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 * mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
34 * mapping->tree_lock (widely used)
35 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
36 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
37 * sb_lock (within inode_lock in fs/fs-writeback.c)
38 * mapping->tree_lock (widely used, in set_page_dirty,
39 * in arch-dependent flush_dcache_mmap_lock,
40 * within bdi.wb->list_lock in __sync_single_inode)
41 *
42 * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon)
43 * ->tasklist_lock
44 * pte map lock
45 */
46
47 #include <linux/mm.h>
48 #include <linux/pagemap.h>
49 #include <linux/swap.h>
50 #include <linux/swapops.h>
51 #include <linux/slab.h>
52 #include <linux/init.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/rcupdate.h>
56 #include <linux/export.h>
57 #include <linux/memcontrol.h>
58 #include <linux/mmu_notifier.h>
59 #include <linux/migrate.h>
60 #include <linux/hugetlb.h>
61 #include <linux/backing-dev.h>
62
63 #include <asm/tlbflush.h>
64
65 #include "internal.h"
66
67 static struct kmem_cache *anon_vma_cachep;
68 static struct kmem_cache *anon_vma_chain_cachep;
69
70 static inline struct anon_vma *anon_vma_alloc(void)
71 {
72 struct anon_vma *anon_vma;
73
74 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
75 if (anon_vma) {
76 atomic_set(&anon_vma->refcount, 1);
77 anon_vma->degree = 1; /* Reference for first vma */
78 anon_vma->parent = anon_vma;
79 /*
80 * Initialise the anon_vma root to point to itself. If called
81 * from fork, the root will be reset to the parents anon_vma.
82 */
83 anon_vma->root = anon_vma;
84 }
85
86 return anon_vma;
87 }
88
89 static inline void anon_vma_free(struct anon_vma *anon_vma)
90 {
91 VM_BUG_ON(atomic_read(&anon_vma->refcount));
92
93 /*
94 * Synchronize against page_lock_anon_vma_read() such that
95 * we can safely hold the lock without the anon_vma getting
96 * freed.
97 *
98 * Relies on the full mb implied by the atomic_dec_and_test() from
99 * put_anon_vma() against the acquire barrier implied by
100 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
101 *
102 * page_lock_anon_vma_read() VS put_anon_vma()
103 * down_read_trylock() atomic_dec_and_test()
104 * LOCK MB
105 * atomic_read() rwsem_is_locked()
106 *
107 * LOCK should suffice since the actual taking of the lock must
108 * happen _before_ what follows.
109 */
110 might_sleep();
111 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
112 anon_vma_lock_write(anon_vma);
113 anon_vma_unlock_write(anon_vma);
114 }
115
116 kmem_cache_free(anon_vma_cachep, anon_vma);
117 }
118
119 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
120 {
121 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
122 }
123
124 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
125 {
126 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
127 }
128
129 static void anon_vma_chain_link(struct vm_area_struct *vma,
130 struct anon_vma_chain *avc,
131 struct anon_vma *anon_vma)
132 {
133 avc->vma = vma;
134 avc->anon_vma = anon_vma;
135 list_add(&avc->same_vma, &vma->anon_vma_chain);
136 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
137 }
138
139 /**
140 * anon_vma_prepare - attach an anon_vma to a memory region
141 * @vma: the memory region in question
142 *
143 * This makes sure the memory mapping described by 'vma' has
144 * an 'anon_vma' attached to it, so that we can associate the
145 * anonymous pages mapped into it with that anon_vma.
146 *
147 * The common case will be that we already have one, but if
148 * not we either need to find an adjacent mapping that we
149 * can re-use the anon_vma from (very common when the only
150 * reason for splitting a vma has been mprotect()), or we
151 * allocate a new one.
152 *
153 * Anon-vma allocations are very subtle, because we may have
154 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
155 * and that may actually touch the spinlock even in the newly
156 * allocated vma (it depends on RCU to make sure that the
157 * anon_vma isn't actually destroyed).
158 *
159 * As a result, we need to do proper anon_vma locking even
160 * for the new allocation. At the same time, we do not want
161 * to do any locking for the common case of already having
162 * an anon_vma.
163 *
164 * This must be called with the mmap_sem held for reading.
165 */
166 int anon_vma_prepare(struct vm_area_struct *vma)
167 {
168 struct anon_vma *anon_vma = vma->anon_vma;
169 struct anon_vma_chain *avc;
170
171 might_sleep();
172 if (unlikely(!anon_vma)) {
173 struct mm_struct *mm = vma->vm_mm;
174 struct anon_vma *allocated;
175
176 avc = anon_vma_chain_alloc(GFP_KERNEL);
177 if (!avc)
178 goto out_enomem;
179
180 anon_vma = find_mergeable_anon_vma(vma);
181 allocated = NULL;
182 if (!anon_vma) {
183 anon_vma = anon_vma_alloc();
184 if (unlikely(!anon_vma))
185 goto out_enomem_free_avc;
186 allocated = anon_vma;
187 }
188
189 anon_vma_lock_write(anon_vma);
190 /* page_table_lock to protect against threads */
191 spin_lock(&mm->page_table_lock);
192 if (likely(!vma->anon_vma)) {
193 vma->anon_vma = anon_vma;
194 anon_vma_chain_link(vma, avc, anon_vma);
195 /* vma reference or self-parent link for new root */
196 anon_vma->degree++;
197 allocated = NULL;
198 avc = NULL;
199 }
200 spin_unlock(&mm->page_table_lock);
201 anon_vma_unlock_write(anon_vma);
202
203 if (unlikely(allocated))
204 put_anon_vma(allocated);
205 if (unlikely(avc))
206 anon_vma_chain_free(avc);
207 }
208 return 0;
209
210 out_enomem_free_avc:
211 anon_vma_chain_free(avc);
212 out_enomem:
213 return -ENOMEM;
214 }
215
216 /*
217 * This is a useful helper function for locking the anon_vma root as
218 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
219 * have the same vma.
220 *
221 * Such anon_vma's should have the same root, so you'd expect to see
222 * just a single mutex_lock for the whole traversal.
223 */
224 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
225 {
226 struct anon_vma *new_root = anon_vma->root;
227 if (new_root != root) {
228 if (WARN_ON_ONCE(root))
229 up_write(&root->rwsem);
230 root = new_root;
231 down_write(&root->rwsem);
232 }
233 return root;
234 }
235
236 static inline void unlock_anon_vma_root(struct anon_vma *root)
237 {
238 if (root)
239 up_write(&root->rwsem);
240 }
241
242 /*
243 * Attach the anon_vmas from src to dst.
244 * Returns 0 on success, -ENOMEM on failure.
245 *
246 * If dst->anon_vma is NULL this function tries to find and reuse existing
247 * anon_vma which has no vmas and only one child anon_vma. This prevents
248 * degradation of anon_vma hierarchy to endless linear chain in case of
249 * constantly forking task. On the other hand, an anon_vma with more than one
250 * child isn't reused even if there was no alive vma, thus rmap walker has a
251 * good chance of avoiding scanning the whole hierarchy when it searches where
252 * page is mapped.
253 */
254 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
255 {
256 struct anon_vma_chain *avc, *pavc;
257 struct anon_vma *root = NULL;
258
259 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
260 struct anon_vma *anon_vma;
261
262 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
263 if (unlikely(!avc)) {
264 unlock_anon_vma_root(root);
265 root = NULL;
266 avc = anon_vma_chain_alloc(GFP_KERNEL);
267 if (!avc)
268 goto enomem_failure;
269 }
270 anon_vma = pavc->anon_vma;
271 root = lock_anon_vma_root(root, anon_vma);
272 anon_vma_chain_link(dst, avc, anon_vma);
273
274 /*
275 * Reuse existing anon_vma if its degree lower than two,
276 * that means it has no vma and only one anon_vma child.
277 *
278 * Do not chose parent anon_vma, otherwise first child
279 * will always reuse it. Root anon_vma is never reused:
280 * it has self-parent reference and at least one child.
281 */
282 if (!dst->anon_vma && anon_vma != src->anon_vma &&
283 anon_vma->degree < 2)
284 dst->anon_vma = anon_vma;
285 }
286 if (dst->anon_vma)
287 dst->anon_vma->degree++;
288 unlock_anon_vma_root(root);
289 return 0;
290
291 enomem_failure:
292 /*
293 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
294 * decremented in unlink_anon_vmas().
295 * We can safely do this because callers of anon_vma_clone() don't care
296 * about dst->anon_vma if anon_vma_clone() failed.
297 */
298 dst->anon_vma = NULL;
299 unlink_anon_vmas(dst);
300 return -ENOMEM;
301 }
302
303 /*
304 * Attach vma to its own anon_vma, as well as to the anon_vmas that
305 * the corresponding VMA in the parent process is attached to.
306 * Returns 0 on success, non-zero on failure.
307 */
308 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
309 {
310 struct anon_vma_chain *avc;
311 struct anon_vma *anon_vma;
312 int error;
313
314 /* Don't bother if the parent process has no anon_vma here. */
315 if (!pvma->anon_vma)
316 return 0;
317
318 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
319 vma->anon_vma = NULL;
320
321 /*
322 * First, attach the new VMA to the parent VMA's anon_vmas,
323 * so rmap can find non-COWed pages in child processes.
324 */
325 error = anon_vma_clone(vma, pvma);
326 if (error)
327 return error;
328
329 /* An existing anon_vma has been reused, all done then. */
330 if (vma->anon_vma)
331 return 0;
332
333 /* Then add our own anon_vma. */
334 anon_vma = anon_vma_alloc();
335 if (!anon_vma)
336 goto out_error;
337 avc = anon_vma_chain_alloc(GFP_KERNEL);
338 if (!avc)
339 goto out_error_free_anon_vma;
340
341 /*
342 * The root anon_vma's spinlock is the lock actually used when we
343 * lock any of the anon_vmas in this anon_vma tree.
344 */
345 anon_vma->root = pvma->anon_vma->root;
346 anon_vma->parent = pvma->anon_vma;
347 /*
348 * With refcounts, an anon_vma can stay around longer than the
349 * process it belongs to. The root anon_vma needs to be pinned until
350 * this anon_vma is freed, because the lock lives in the root.
351 */
352 get_anon_vma(anon_vma->root);
353 /* Mark this anon_vma as the one where our new (COWed) pages go. */
354 vma->anon_vma = anon_vma;
355 anon_vma_lock_write(anon_vma);
356 anon_vma_chain_link(vma, avc, anon_vma);
357 anon_vma->parent->degree++;
358 anon_vma_unlock_write(anon_vma);
359
360 return 0;
361
362 out_error_free_anon_vma:
363 put_anon_vma(anon_vma);
364 out_error:
365 unlink_anon_vmas(vma);
366 return -ENOMEM;
367 }
368
369 void unlink_anon_vmas(struct vm_area_struct *vma)
370 {
371 struct anon_vma_chain *avc, *next;
372 struct anon_vma *root = NULL;
373
374 /*
375 * Unlink each anon_vma chained to the VMA. This list is ordered
376 * from newest to oldest, ensuring the root anon_vma gets freed last.
377 */
378 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
379 struct anon_vma *anon_vma = avc->anon_vma;
380
381 root = lock_anon_vma_root(root, anon_vma);
382 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
383
384 /*
385 * Leave empty anon_vmas on the list - we'll need
386 * to free them outside the lock.
387 */
388 if (RB_EMPTY_ROOT(&anon_vma->rb_root)) {
389 anon_vma->parent->degree--;
390 continue;
391 }
392
393 list_del(&avc->same_vma);
394 anon_vma_chain_free(avc);
395 }
396 if (vma->anon_vma)
397 vma->anon_vma->degree--;
398 unlock_anon_vma_root(root);
399
400 /*
401 * Iterate the list once more, it now only contains empty and unlinked
402 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
403 * needing to write-acquire the anon_vma->root->rwsem.
404 */
405 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
406 struct anon_vma *anon_vma = avc->anon_vma;
407
408 BUG_ON(anon_vma->degree);
409 put_anon_vma(anon_vma);
410
411 list_del(&avc->same_vma);
412 anon_vma_chain_free(avc);
413 }
414 }
415
416 static void anon_vma_ctor(void *data)
417 {
418 struct anon_vma *anon_vma = data;
419
420 init_rwsem(&anon_vma->rwsem);
421 atomic_set(&anon_vma->refcount, 0);
422 anon_vma->rb_root = RB_ROOT;
423 }
424
425 void __init anon_vma_init(void)
426 {
427 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
428 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
429 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
430 }
431
432 /*
433 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
434 *
435 * Since there is no serialization what so ever against page_remove_rmap()
436 * the best this function can do is return a locked anon_vma that might
437 * have been relevant to this page.
438 *
439 * The page might have been remapped to a different anon_vma or the anon_vma
440 * returned may already be freed (and even reused).
441 *
442 * In case it was remapped to a different anon_vma, the new anon_vma will be a
443 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
444 * ensure that any anon_vma obtained from the page will still be valid for as
445 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
446 *
447 * All users of this function must be very careful when walking the anon_vma
448 * chain and verify that the page in question is indeed mapped in it
449 * [ something equivalent to page_mapped_in_vma() ].
450 *
451 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
452 * that the anon_vma pointer from page->mapping is valid if there is a
453 * mapcount, we can dereference the anon_vma after observing those.
454 */
455 struct anon_vma *page_get_anon_vma(struct page *page)
456 {
457 struct anon_vma *anon_vma = NULL;
458 unsigned long anon_mapping;
459
460 rcu_read_lock();
461 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
462 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
463 goto out;
464 if (!page_mapped(page))
465 goto out;
466
467 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
468 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
469 anon_vma = NULL;
470 goto out;
471 }
472
473 /*
474 * If this page is still mapped, then its anon_vma cannot have been
475 * freed. But if it has been unmapped, we have no security against the
476 * anon_vma structure being freed and reused (for another anon_vma:
477 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
478 * above cannot corrupt).
479 */
480 if (!page_mapped(page)) {
481 rcu_read_unlock();
482 put_anon_vma(anon_vma);
483 return NULL;
484 }
485 out:
486 rcu_read_unlock();
487
488 return anon_vma;
489 }
490
491 /*
492 * Similar to page_get_anon_vma() except it locks the anon_vma.
493 *
494 * Its a little more complex as it tries to keep the fast path to a single
495 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
496 * reference like with page_get_anon_vma() and then block on the mutex.
497 */
498 struct anon_vma *page_lock_anon_vma_read(struct page *page)
499 {
500 struct anon_vma *anon_vma = NULL;
501 struct anon_vma *root_anon_vma;
502 unsigned long anon_mapping;
503
504 rcu_read_lock();
505 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
506 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
507 goto out;
508 if (!page_mapped(page))
509 goto out;
510
511 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
512 root_anon_vma = READ_ONCE(anon_vma->root);
513 if (down_read_trylock(&root_anon_vma->rwsem)) {
514 /*
515 * If the page is still mapped, then this anon_vma is still
516 * its anon_vma, and holding the mutex ensures that it will
517 * not go away, see anon_vma_free().
518 */
519 if (!page_mapped(page)) {
520 up_read(&root_anon_vma->rwsem);
521 anon_vma = NULL;
522 }
523 goto out;
524 }
525
526 /* trylock failed, we got to sleep */
527 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
528 anon_vma = NULL;
529 goto out;
530 }
531
532 if (!page_mapped(page)) {
533 rcu_read_unlock();
534 put_anon_vma(anon_vma);
535 return NULL;
536 }
537
538 /* we pinned the anon_vma, its safe to sleep */
539 rcu_read_unlock();
540 anon_vma_lock_read(anon_vma);
541
542 if (atomic_dec_and_test(&anon_vma->refcount)) {
543 /*
544 * Oops, we held the last refcount, release the lock
545 * and bail -- can't simply use put_anon_vma() because
546 * we'll deadlock on the anon_vma_lock_write() recursion.
547 */
548 anon_vma_unlock_read(anon_vma);
549 __put_anon_vma(anon_vma);
550 anon_vma = NULL;
551 }
552
553 return anon_vma;
554
555 out:
556 rcu_read_unlock();
557 return anon_vma;
558 }
559
560 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
561 {
562 anon_vma_unlock_read(anon_vma);
563 }
564
565 /*
566 * At what user virtual address is page expected in @vma?
567 */
568 static inline unsigned long
569 __vma_address(struct page *page, struct vm_area_struct *vma)
570 {
571 pgoff_t pgoff = page_to_pgoff(page);
572 return vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
573 }
574
575 inline unsigned long
576 vma_address(struct page *page, struct vm_area_struct *vma)
577 {
578 unsigned long address = __vma_address(page, vma);
579
580 /* page should be within @vma mapping range */
581 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
582
583 return address;
584 }
585
586 /*
587 * At what user virtual address is page expected in vma?
588 * Caller should check the page is actually part of the vma.
589 */
590 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
591 {
592 unsigned long address;
593 if (PageAnon(page)) {
594 struct anon_vma *page__anon_vma = page_anon_vma(page);
595 /*
596 * Note: swapoff's unuse_vma() is more efficient with this
597 * check, and needs it to match anon_vma when KSM is active.
598 */
599 if (!vma->anon_vma || !page__anon_vma ||
600 vma->anon_vma->root != page__anon_vma->root)
601 return -EFAULT;
602 } else if (page->mapping) {
603 if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
604 return -EFAULT;
605 } else
606 return -EFAULT;
607 address = __vma_address(page, vma);
608 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
609 return -EFAULT;
610 return address;
611 }
612
613 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
614 {
615 pgd_t *pgd;
616 pud_t *pud;
617 pmd_t *pmd = NULL;
618 pmd_t pmde;
619
620 pgd = pgd_offset(mm, address);
621 if (!pgd_present(*pgd))
622 goto out;
623
624 pud = pud_offset(pgd, address);
625 if (!pud_present(*pud))
626 goto out;
627
628 pmd = pmd_offset(pud, address);
629 /*
630 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
631 * without holding anon_vma lock for write. So when looking for a
632 * genuine pmde (in which to find pte), test present and !THP together.
633 */
634 pmde = *pmd;
635 barrier();
636 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
637 pmd = NULL;
638 out:
639 return pmd;
640 }
641
642 /*
643 * Check that @page is mapped at @address into @mm.
644 *
645 * If @sync is false, page_check_address may perform a racy check to avoid
646 * the page table lock when the pte is not present (helpful when reclaiming
647 * highly shared pages).
648 *
649 * On success returns with pte mapped and locked.
650 */
651 pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
652 unsigned long address, spinlock_t **ptlp, int sync)
653 {
654 pmd_t *pmd;
655 pte_t *pte;
656 spinlock_t *ptl;
657
658 if (unlikely(PageHuge(page))) {
659 /* when pud is not present, pte will be NULL */
660 pte = huge_pte_offset(mm, address);
661 if (!pte)
662 return NULL;
663
664 ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
665 goto check;
666 }
667
668 pmd = mm_find_pmd(mm, address);
669 if (!pmd)
670 return NULL;
671
672 pte = pte_offset_map(pmd, address);
673 /* Make a quick check before getting the lock */
674 if (!sync && !pte_present(*pte)) {
675 pte_unmap(pte);
676 return NULL;
677 }
678
679 ptl = pte_lockptr(mm, pmd);
680 check:
681 spin_lock(ptl);
682 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
683 *ptlp = ptl;
684 return pte;
685 }
686 pte_unmap_unlock(pte, ptl);
687 return NULL;
688 }
689
690 /**
691 * page_mapped_in_vma - check whether a page is really mapped in a VMA
692 * @page: the page to test
693 * @vma: the VMA to test
694 *
695 * Returns 1 if the page is mapped into the page tables of the VMA, 0
696 * if the page is not mapped into the page tables of this VMA. Only
697 * valid for normal file or anonymous VMAs.
698 */
699 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
700 {
701 unsigned long address;
702 pte_t *pte;
703 spinlock_t *ptl;
704
705 address = __vma_address(page, vma);
706 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
707 return 0;
708 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
709 if (!pte) /* the page is not in this mm */
710 return 0;
711 pte_unmap_unlock(pte, ptl);
712
713 return 1;
714 }
715
716 struct page_referenced_arg {
717 int mapcount;
718 int referenced;
719 unsigned long vm_flags;
720 struct mem_cgroup *memcg;
721 };
722 /*
723 * arg: page_referenced_arg will be passed
724 */
725 static int page_referenced_one(struct page *page, struct vm_area_struct *vma,
726 unsigned long address, void *arg)
727 {
728 struct mm_struct *mm = vma->vm_mm;
729 spinlock_t *ptl;
730 int referenced = 0;
731 struct page_referenced_arg *pra = arg;
732
733 if (unlikely(PageTransHuge(page))) {
734 pmd_t *pmd;
735
736 /*
737 * rmap might return false positives; we must filter
738 * these out using page_check_address_pmd().
739 */
740 pmd = page_check_address_pmd(page, mm, address,
741 PAGE_CHECK_ADDRESS_PMD_FLAG, &ptl);
742 if (!pmd)
743 return SWAP_AGAIN;
744
745 if (vma->vm_flags & VM_LOCKED) {
746 spin_unlock(ptl);
747 pra->vm_flags |= VM_LOCKED;
748 return SWAP_FAIL; /* To break the loop */
749 }
750
751 /* go ahead even if the pmd is pmd_trans_splitting() */
752 if (pmdp_clear_flush_young_notify(vma, address, pmd))
753 referenced++;
754 spin_unlock(ptl);
755 } else {
756 pte_t *pte;
757
758 /*
759 * rmap might return false positives; we must filter
760 * these out using page_check_address().
761 */
762 pte = page_check_address(page, mm, address, &ptl, 0);
763 if (!pte)
764 return SWAP_AGAIN;
765
766 if (vma->vm_flags & VM_LOCKED) {
767 pte_unmap_unlock(pte, ptl);
768 pra->vm_flags |= VM_LOCKED;
769 return SWAP_FAIL; /* To break the loop */
770 }
771
772 if (ptep_clear_flush_young_notify(vma, address, pte)) {
773 /*
774 * Don't treat a reference through a sequentially read
775 * mapping as such. If the page has been used in
776 * another mapping, we will catch it; if this other
777 * mapping is already gone, the unmap path will have
778 * set PG_referenced or activated the page.
779 */
780 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
781 referenced++;
782 }
783 pte_unmap_unlock(pte, ptl);
784 }
785
786 if (referenced) {
787 pra->referenced++;
788 pra->vm_flags |= vma->vm_flags;
789 }
790
791 pra->mapcount--;
792 if (!pra->mapcount)
793 return SWAP_SUCCESS; /* To break the loop */
794
795 return SWAP_AGAIN;
796 }
797
798 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
799 {
800 struct page_referenced_arg *pra = arg;
801 struct mem_cgroup *memcg = pra->memcg;
802
803 if (!mm_match_cgroup(vma->vm_mm, memcg))
804 return true;
805
806 return false;
807 }
808
809 /**
810 * page_referenced - test if the page was referenced
811 * @page: the page to test
812 * @is_locked: caller holds lock on the page
813 * @memcg: target memory cgroup
814 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
815 *
816 * Quick test_and_clear_referenced for all mappings to a page,
817 * returns the number of ptes which referenced the page.
818 */
819 int page_referenced(struct page *page,
820 int is_locked,
821 struct mem_cgroup *memcg,
822 unsigned long *vm_flags)
823 {
824 int ret;
825 int we_locked = 0;
826 struct page_referenced_arg pra = {
827 .mapcount = page_mapcount(page),
828 .memcg = memcg,
829 };
830 struct rmap_walk_control rwc = {
831 .rmap_one = page_referenced_one,
832 .arg = (void *)&pra,
833 .anon_lock = page_lock_anon_vma_read,
834 };
835
836 *vm_flags = 0;
837 if (!page_mapped(page))
838 return 0;
839
840 if (!page_rmapping(page))
841 return 0;
842
843 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
844 we_locked = trylock_page(page);
845 if (!we_locked)
846 return 1;
847 }
848
849 /*
850 * If we are reclaiming on behalf of a cgroup, skip
851 * counting on behalf of references from different
852 * cgroups
853 */
854 if (memcg) {
855 rwc.invalid_vma = invalid_page_referenced_vma;
856 }
857
858 ret = rmap_walk(page, &rwc);
859 *vm_flags = pra.vm_flags;
860
861 if (we_locked)
862 unlock_page(page);
863
864 return pra.referenced;
865 }
866
867 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
868 unsigned long address, void *arg)
869 {
870 struct mm_struct *mm = vma->vm_mm;
871 pte_t *pte;
872 spinlock_t *ptl;
873 int ret = 0;
874 int *cleaned = arg;
875
876 pte = page_check_address(page, mm, address, &ptl, 1);
877 if (!pte)
878 goto out;
879
880 if (pte_dirty(*pte) || pte_write(*pte)) {
881 pte_t entry;
882
883 flush_cache_page(vma, address, pte_pfn(*pte));
884 entry = ptep_clear_flush(vma, address, pte);
885 entry = pte_wrprotect(entry);
886 entry = pte_mkclean(entry);
887 set_pte_at(mm, address, pte, entry);
888 ret = 1;
889 }
890
891 pte_unmap_unlock(pte, ptl);
892
893 if (ret) {
894 mmu_notifier_invalidate_page(mm, address);
895 (*cleaned)++;
896 }
897 out:
898 return SWAP_AGAIN;
899 }
900
901 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
902 {
903 if (vma->vm_flags & VM_SHARED)
904 return false;
905
906 return true;
907 }
908
909 int page_mkclean(struct page *page)
910 {
911 int cleaned = 0;
912 struct address_space *mapping;
913 struct rmap_walk_control rwc = {
914 .arg = (void *)&cleaned,
915 .rmap_one = page_mkclean_one,
916 .invalid_vma = invalid_mkclean_vma,
917 };
918
919 BUG_ON(!PageLocked(page));
920
921 if (!page_mapped(page))
922 return 0;
923
924 mapping = page_mapping(page);
925 if (!mapping)
926 return 0;
927
928 rmap_walk(page, &rwc);
929
930 return cleaned;
931 }
932 EXPORT_SYMBOL_GPL(page_mkclean);
933
934 /**
935 * page_move_anon_rmap - move a page to our anon_vma
936 * @page: the page to move to our anon_vma
937 * @vma: the vma the page belongs to
938 * @address: the user virtual address mapped
939 *
940 * When a page belongs exclusively to one process after a COW event,
941 * that page can be moved into the anon_vma that belongs to just that
942 * process, so the rmap code will not search the parent or sibling
943 * processes.
944 */
945 void page_move_anon_rmap(struct page *page,
946 struct vm_area_struct *vma, unsigned long address)
947 {
948 struct anon_vma *anon_vma = vma->anon_vma;
949
950 VM_BUG_ON_PAGE(!PageLocked(page), page);
951 VM_BUG_ON_VMA(!anon_vma, vma);
952 VM_BUG_ON_PAGE(page->index != linear_page_index(vma, address), page);
953
954 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
955 /*
956 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
957 * simultaneously, so a concurrent reader (eg page_referenced()'s
958 * PageAnon()) will not see one without the other.
959 */
960 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
961 }
962
963 /**
964 * __page_set_anon_rmap - set up new anonymous rmap
965 * @page: Page to add to rmap
966 * @vma: VM area to add page to.
967 * @address: User virtual address of the mapping
968 * @exclusive: the page is exclusively owned by the current process
969 */
970 static void __page_set_anon_rmap(struct page *page,
971 struct vm_area_struct *vma, unsigned long address, int exclusive)
972 {
973 struct anon_vma *anon_vma = vma->anon_vma;
974
975 BUG_ON(!anon_vma);
976
977 if (PageAnon(page))
978 return;
979
980 /*
981 * If the page isn't exclusively mapped into this vma,
982 * we must use the _oldest_ possible anon_vma for the
983 * page mapping!
984 */
985 if (!exclusive)
986 anon_vma = anon_vma->root;
987
988 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
989 page->mapping = (struct address_space *) anon_vma;
990 page->index = linear_page_index(vma, address);
991 }
992
993 /**
994 * __page_check_anon_rmap - sanity check anonymous rmap addition
995 * @page: the page to add the mapping to
996 * @vma: the vm area in which the mapping is added
997 * @address: the user virtual address mapped
998 */
999 static void __page_check_anon_rmap(struct page *page,
1000 struct vm_area_struct *vma, unsigned long address)
1001 {
1002 #ifdef CONFIG_DEBUG_VM
1003 /*
1004 * The page's anon-rmap details (mapping and index) are guaranteed to
1005 * be set up correctly at this point.
1006 *
1007 * We have exclusion against page_add_anon_rmap because the caller
1008 * always holds the page locked, except if called from page_dup_rmap,
1009 * in which case the page is already known to be setup.
1010 *
1011 * We have exclusion against page_add_new_anon_rmap because those pages
1012 * are initially only visible via the pagetables, and the pte is locked
1013 * over the call to page_add_new_anon_rmap.
1014 */
1015 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1016 BUG_ON(page->index != linear_page_index(vma, address));
1017 #endif
1018 }
1019
1020 /**
1021 * page_add_anon_rmap - add pte mapping to an anonymous page
1022 * @page: the page to add the mapping to
1023 * @vma: the vm area in which the mapping is added
1024 * @address: the user virtual address mapped
1025 *
1026 * The caller needs to hold the pte lock, and the page must be locked in
1027 * the anon_vma case: to serialize mapping,index checking after setting,
1028 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1029 * (but PageKsm is never downgraded to PageAnon).
1030 */
1031 void page_add_anon_rmap(struct page *page,
1032 struct vm_area_struct *vma, unsigned long address)
1033 {
1034 do_page_add_anon_rmap(page, vma, address, 0);
1035 }
1036
1037 /*
1038 * Special version of the above for do_swap_page, which often runs
1039 * into pages that are exclusively owned by the current process.
1040 * Everybody else should continue to use page_add_anon_rmap above.
1041 */
1042 void do_page_add_anon_rmap(struct page *page,
1043 struct vm_area_struct *vma, unsigned long address, int exclusive)
1044 {
1045 int first = atomic_inc_and_test(&page->_mapcount);
1046 if (first) {
1047 /*
1048 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1049 * these counters are not modified in interrupt context, and
1050 * pte lock(a spinlock) is held, which implies preemption
1051 * disabled.
1052 */
1053 if (PageTransHuge(page))
1054 __inc_zone_page_state(page,
1055 NR_ANON_TRANSPARENT_HUGEPAGES);
1056 __mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1057 hpage_nr_pages(page));
1058 }
1059 if (unlikely(PageKsm(page)))
1060 return;
1061
1062 VM_BUG_ON_PAGE(!PageLocked(page), page);
1063 /* address might be in next vma when migration races vma_adjust */
1064 if (first)
1065 __page_set_anon_rmap(page, vma, address, exclusive);
1066 else
1067 __page_check_anon_rmap(page, vma, address);
1068 }
1069
1070 /**
1071 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1072 * @page: the page to add the mapping to
1073 * @vma: the vm area in which the mapping is added
1074 * @address: the user virtual address mapped
1075 *
1076 * Same as page_add_anon_rmap but must only be called on *new* pages.
1077 * This means the inc-and-test can be bypassed.
1078 * Page does not have to be locked.
1079 */
1080 void page_add_new_anon_rmap(struct page *page,
1081 struct vm_area_struct *vma, unsigned long address)
1082 {
1083 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1084 SetPageSwapBacked(page);
1085 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1086 if (PageTransHuge(page))
1087 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1088 __mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1089 hpage_nr_pages(page));
1090 __page_set_anon_rmap(page, vma, address, 1);
1091 }
1092
1093 /**
1094 * page_add_file_rmap - add pte mapping to a file page
1095 * @page: the page to add the mapping to
1096 *
1097 * The caller needs to hold the pte lock.
1098 */
1099 void page_add_file_rmap(struct page *page)
1100 {
1101 struct mem_cgroup *memcg;
1102
1103 memcg = mem_cgroup_begin_page_stat(page);
1104 if (atomic_inc_and_test(&page->_mapcount)) {
1105 __inc_zone_page_state(page, NR_FILE_MAPPED);
1106 mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
1107 }
1108 mem_cgroup_end_page_stat(memcg);
1109 }
1110
1111 static void page_remove_file_rmap(struct page *page)
1112 {
1113 struct mem_cgroup *memcg;
1114
1115 memcg = mem_cgroup_begin_page_stat(page);
1116
1117 /* page still mapped by someone else? */
1118 if (!atomic_add_negative(-1, &page->_mapcount))
1119 goto out;
1120
1121 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1122 if (unlikely(PageHuge(page)))
1123 goto out;
1124
1125 /*
1126 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1127 * these counters are not modified in interrupt context, and
1128 * pte lock(a spinlock) is held, which implies preemption disabled.
1129 */
1130 __dec_zone_page_state(page, NR_FILE_MAPPED);
1131 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
1132
1133 if (unlikely(PageMlocked(page)))
1134 clear_page_mlock(page);
1135 out:
1136 mem_cgroup_end_page_stat(memcg);
1137 }
1138
1139 /**
1140 * page_remove_rmap - take down pte mapping from a page
1141 * @page: page to remove mapping from
1142 *
1143 * The caller needs to hold the pte lock.
1144 */
1145 void page_remove_rmap(struct page *page)
1146 {
1147 if (!PageAnon(page)) {
1148 page_remove_file_rmap(page);
1149 return;
1150 }
1151
1152 /* page still mapped by someone else? */
1153 if (!atomic_add_negative(-1, &page->_mapcount))
1154 return;
1155
1156 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1157 if (unlikely(PageHuge(page)))
1158 return;
1159
1160 /*
1161 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1162 * these counters are not modified in interrupt context, and
1163 * pte lock(a spinlock) is held, which implies preemption disabled.
1164 */
1165 if (PageTransHuge(page))
1166 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1167
1168 __mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1169 -hpage_nr_pages(page));
1170
1171 if (unlikely(PageMlocked(page)))
1172 clear_page_mlock(page);
1173
1174 /*
1175 * It would be tidy to reset the PageAnon mapping here,
1176 * but that might overwrite a racing page_add_anon_rmap
1177 * which increments mapcount after us but sets mapping
1178 * before us: so leave the reset to free_hot_cold_page,
1179 * and remember that it's only reliable while mapped.
1180 * Leaving it set also helps swapoff to reinstate ptes
1181 * faster for those pages still in swapcache.
1182 */
1183 }
1184
1185 /*
1186 * @arg: enum ttu_flags will be passed to this argument
1187 */
1188 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1189 unsigned long address, void *arg)
1190 {
1191 struct mm_struct *mm = vma->vm_mm;
1192 pte_t *pte;
1193 pte_t pteval;
1194 spinlock_t *ptl;
1195 int ret = SWAP_AGAIN;
1196 enum ttu_flags flags = (enum ttu_flags)arg;
1197
1198 pte = page_check_address(page, mm, address, &ptl, 0);
1199 if (!pte)
1200 goto out;
1201
1202 /*
1203 * If the page is mlock()d, we cannot swap it out.
1204 * If it's recently referenced (perhaps page_referenced
1205 * skipped over this mm) then we should reactivate it.
1206 */
1207 if (!(flags & TTU_IGNORE_MLOCK)) {
1208 if (vma->vm_flags & VM_LOCKED)
1209 goto out_mlock;
1210
1211 if (flags & TTU_MUNLOCK)
1212 goto out_unmap;
1213 }
1214 if (!(flags & TTU_IGNORE_ACCESS)) {
1215 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1216 ret = SWAP_FAIL;
1217 goto out_unmap;
1218 }
1219 }
1220
1221 /* Nuke the page table entry. */
1222 flush_cache_page(vma, address, page_to_pfn(page));
1223 pteval = ptep_clear_flush(vma, address, pte);
1224
1225 /* Move the dirty bit to the physical page now the pte is gone. */
1226 if (pte_dirty(pteval))
1227 set_page_dirty(page);
1228
1229 /* Update high watermark before we lower rss */
1230 update_hiwater_rss(mm);
1231
1232 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1233 if (!PageHuge(page)) {
1234 if (PageAnon(page))
1235 dec_mm_counter(mm, MM_ANONPAGES);
1236 else
1237 dec_mm_counter(mm, MM_FILEPAGES);
1238 }
1239 set_pte_at(mm, address, pte,
1240 swp_entry_to_pte(make_hwpoison_entry(page)));
1241 } else if (pte_unused(pteval)) {
1242 /*
1243 * The guest indicated that the page content is of no
1244 * interest anymore. Simply discard the pte, vmscan
1245 * will take care of the rest.
1246 */
1247 if (PageAnon(page))
1248 dec_mm_counter(mm, MM_ANONPAGES);
1249 else
1250 dec_mm_counter(mm, MM_FILEPAGES);
1251 } else if (PageAnon(page)) {
1252 swp_entry_t entry = { .val = page_private(page) };
1253 pte_t swp_pte;
1254
1255 if (PageSwapCache(page)) {
1256 /*
1257 * Store the swap location in the pte.
1258 * See handle_pte_fault() ...
1259 */
1260 if (swap_duplicate(entry) < 0) {
1261 set_pte_at(mm, address, pte, pteval);
1262 ret = SWAP_FAIL;
1263 goto out_unmap;
1264 }
1265 if (list_empty(&mm->mmlist)) {
1266 spin_lock(&mmlist_lock);
1267 if (list_empty(&mm->mmlist))
1268 list_add(&mm->mmlist, &init_mm.mmlist);
1269 spin_unlock(&mmlist_lock);
1270 }
1271 dec_mm_counter(mm, MM_ANONPAGES);
1272 inc_mm_counter(mm, MM_SWAPENTS);
1273 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
1274 /*
1275 * Store the pfn of the page in a special migration
1276 * pte. do_swap_page() will wait until the migration
1277 * pte is removed and then restart fault handling.
1278 */
1279 BUG_ON(!(flags & TTU_MIGRATION));
1280 entry = make_migration_entry(page, pte_write(pteval));
1281 }
1282 swp_pte = swp_entry_to_pte(entry);
1283 if (pte_soft_dirty(pteval))
1284 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1285 set_pte_at(mm, address, pte, swp_pte);
1286 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1287 (flags & TTU_MIGRATION)) {
1288 /* Establish migration entry for a file page */
1289 swp_entry_t entry;
1290 entry = make_migration_entry(page, pte_write(pteval));
1291 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1292 } else
1293 dec_mm_counter(mm, MM_FILEPAGES);
1294
1295 page_remove_rmap(page);
1296 page_cache_release(page);
1297
1298 out_unmap:
1299 pte_unmap_unlock(pte, ptl);
1300 if (ret != SWAP_FAIL && !(flags & TTU_MUNLOCK))
1301 mmu_notifier_invalidate_page(mm, address);
1302 out:
1303 return ret;
1304
1305 out_mlock:
1306 pte_unmap_unlock(pte, ptl);
1307
1308
1309 /*
1310 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1311 * unstable result and race. Plus, We can't wait here because
1312 * we now hold anon_vma->rwsem or mapping->i_mmap_rwsem.
1313 * if trylock failed, the page remain in evictable lru and later
1314 * vmscan could retry to move the page to unevictable lru if the
1315 * page is actually mlocked.
1316 */
1317 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1318 if (vma->vm_flags & VM_LOCKED) {
1319 mlock_vma_page(page);
1320 ret = SWAP_MLOCK;
1321 }
1322 up_read(&vma->vm_mm->mmap_sem);
1323 }
1324 return ret;
1325 }
1326
1327 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1328 {
1329 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1330
1331 if (!maybe_stack)
1332 return false;
1333
1334 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1335 VM_STACK_INCOMPLETE_SETUP)
1336 return true;
1337
1338 return false;
1339 }
1340
1341 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1342 {
1343 return is_vma_temporary_stack(vma);
1344 }
1345
1346 static int page_not_mapped(struct page *page)
1347 {
1348 return !page_mapped(page);
1349 };
1350
1351 /**
1352 * try_to_unmap - try to remove all page table mappings to a page
1353 * @page: the page to get unmapped
1354 * @flags: action and flags
1355 *
1356 * Tries to remove all the page table entries which are mapping this
1357 * page, used in the pageout path. Caller must hold the page lock.
1358 * Return values are:
1359 *
1360 * SWAP_SUCCESS - we succeeded in removing all mappings
1361 * SWAP_AGAIN - we missed a mapping, try again later
1362 * SWAP_FAIL - the page is unswappable
1363 * SWAP_MLOCK - page is mlocked.
1364 */
1365 int try_to_unmap(struct page *page, enum ttu_flags flags)
1366 {
1367 int ret;
1368 struct rmap_walk_control rwc = {
1369 .rmap_one = try_to_unmap_one,
1370 .arg = (void *)flags,
1371 .done = page_not_mapped,
1372 .anon_lock = page_lock_anon_vma_read,
1373 };
1374
1375 VM_BUG_ON_PAGE(!PageHuge(page) && PageTransHuge(page), page);
1376
1377 /*
1378 * During exec, a temporary VMA is setup and later moved.
1379 * The VMA is moved under the anon_vma lock but not the
1380 * page tables leading to a race where migration cannot
1381 * find the migration ptes. Rather than increasing the
1382 * locking requirements of exec(), migration skips
1383 * temporary VMAs until after exec() completes.
1384 */
1385 if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page))
1386 rwc.invalid_vma = invalid_migration_vma;
1387
1388 ret = rmap_walk(page, &rwc);
1389
1390 if (ret != SWAP_MLOCK && !page_mapped(page))
1391 ret = SWAP_SUCCESS;
1392 return ret;
1393 }
1394
1395 /**
1396 * try_to_munlock - try to munlock a page
1397 * @page: the page to be munlocked
1398 *
1399 * Called from munlock code. Checks all of the VMAs mapping the page
1400 * to make sure nobody else has this page mlocked. The page will be
1401 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1402 *
1403 * Return values are:
1404 *
1405 * SWAP_AGAIN - no vma is holding page mlocked, or,
1406 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1407 * SWAP_FAIL - page cannot be located at present
1408 * SWAP_MLOCK - page is now mlocked.
1409 */
1410 int try_to_munlock(struct page *page)
1411 {
1412 int ret;
1413 struct rmap_walk_control rwc = {
1414 .rmap_one = try_to_unmap_one,
1415 .arg = (void *)TTU_MUNLOCK,
1416 .done = page_not_mapped,
1417 .anon_lock = page_lock_anon_vma_read,
1418
1419 };
1420
1421 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1422
1423 ret = rmap_walk(page, &rwc);
1424 return ret;
1425 }
1426
1427 void __put_anon_vma(struct anon_vma *anon_vma)
1428 {
1429 struct anon_vma *root = anon_vma->root;
1430
1431 anon_vma_free(anon_vma);
1432 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1433 anon_vma_free(root);
1434 }
1435
1436 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1437 struct rmap_walk_control *rwc)
1438 {
1439 struct anon_vma *anon_vma;
1440
1441 if (rwc->anon_lock)
1442 return rwc->anon_lock(page);
1443
1444 /*
1445 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1446 * because that depends on page_mapped(); but not all its usages
1447 * are holding mmap_sem. Users without mmap_sem are required to
1448 * take a reference count to prevent the anon_vma disappearing
1449 */
1450 anon_vma = page_anon_vma(page);
1451 if (!anon_vma)
1452 return NULL;
1453
1454 anon_vma_lock_read(anon_vma);
1455 return anon_vma;
1456 }
1457
1458 /*
1459 * rmap_walk_anon - do something to anonymous page using the object-based
1460 * rmap method
1461 * @page: the page to be handled
1462 * @rwc: control variable according to each walk type
1463 *
1464 * Find all the mappings of a page using the mapping pointer and the vma chains
1465 * contained in the anon_vma struct it points to.
1466 *
1467 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1468 * where the page was found will be held for write. So, we won't recheck
1469 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1470 * LOCKED.
1471 */
1472 static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc)
1473 {
1474 struct anon_vma *anon_vma;
1475 pgoff_t pgoff;
1476 struct anon_vma_chain *avc;
1477 int ret = SWAP_AGAIN;
1478
1479 anon_vma = rmap_walk_anon_lock(page, rwc);
1480 if (!anon_vma)
1481 return ret;
1482
1483 pgoff = page_to_pgoff(page);
1484 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1485 struct vm_area_struct *vma = avc->vma;
1486 unsigned long address = vma_address(page, vma);
1487
1488 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1489 continue;
1490
1491 ret = rwc->rmap_one(page, vma, address, rwc->arg);
1492 if (ret != SWAP_AGAIN)
1493 break;
1494 if (rwc->done && rwc->done(page))
1495 break;
1496 }
1497 anon_vma_unlock_read(anon_vma);
1498 return ret;
1499 }
1500
1501 /*
1502 * rmap_walk_file - do something to file page using the object-based rmap method
1503 * @page: the page to be handled
1504 * @rwc: control variable according to each walk type
1505 *
1506 * Find all the mappings of a page using the mapping pointer and the vma chains
1507 * contained in the address_space struct it points to.
1508 *
1509 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1510 * where the page was found will be held for write. So, we won't recheck
1511 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1512 * LOCKED.
1513 */
1514 static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc)
1515 {
1516 struct address_space *mapping = page->mapping;
1517 pgoff_t pgoff;
1518 struct vm_area_struct *vma;
1519 int ret = SWAP_AGAIN;
1520
1521 /*
1522 * The page lock not only makes sure that page->mapping cannot
1523 * suddenly be NULLified by truncation, it makes sure that the
1524 * structure at mapping cannot be freed and reused yet,
1525 * so we can safely take mapping->i_mmap_rwsem.
1526 */
1527 VM_BUG_ON_PAGE(!PageLocked(page), page);
1528
1529 if (!mapping)
1530 return ret;
1531
1532 pgoff = page_to_pgoff(page);
1533 i_mmap_lock_read(mapping);
1534 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1535 unsigned long address = vma_address(page, vma);
1536
1537 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1538 continue;
1539
1540 ret = rwc->rmap_one(page, vma, address, rwc->arg);
1541 if (ret != SWAP_AGAIN)
1542 goto done;
1543 if (rwc->done && rwc->done(page))
1544 goto done;
1545 }
1546
1547 done:
1548 i_mmap_unlock_read(mapping);
1549 return ret;
1550 }
1551
1552 int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1553 {
1554 if (unlikely(PageKsm(page)))
1555 return rmap_walk_ksm(page, rwc);
1556 else if (PageAnon(page))
1557 return rmap_walk_anon(page, rwc);
1558 else
1559 return rmap_walk_file(page, rwc);
1560 }
1561
1562 #ifdef CONFIG_HUGETLB_PAGE
1563 /*
1564 * The following three functions are for anonymous (private mapped) hugepages.
1565 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1566 * and no lru code, because we handle hugepages differently from common pages.
1567 */
1568 static void __hugepage_set_anon_rmap(struct page *page,
1569 struct vm_area_struct *vma, unsigned long address, int exclusive)
1570 {
1571 struct anon_vma *anon_vma = vma->anon_vma;
1572
1573 BUG_ON(!anon_vma);
1574
1575 if (PageAnon(page))
1576 return;
1577 if (!exclusive)
1578 anon_vma = anon_vma->root;
1579
1580 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1581 page->mapping = (struct address_space *) anon_vma;
1582 page->index = linear_page_index(vma, address);
1583 }
1584
1585 void hugepage_add_anon_rmap(struct page *page,
1586 struct vm_area_struct *vma, unsigned long address)
1587 {
1588 struct anon_vma *anon_vma = vma->anon_vma;
1589 int first;
1590
1591 BUG_ON(!PageLocked(page));
1592 BUG_ON(!anon_vma);
1593 /* address might be in next vma when migration races vma_adjust */
1594 first = atomic_inc_and_test(&page->_mapcount);
1595 if (first)
1596 __hugepage_set_anon_rmap(page, vma, address, 0);
1597 }
1598
1599 void hugepage_add_new_anon_rmap(struct page *page,
1600 struct vm_area_struct *vma, unsigned long address)
1601 {
1602 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1603 atomic_set(&page->_mapcount, 0);
1604 __hugepage_set_anon_rmap(page, vma, address, 1);
1605 }
1606 #endif /* CONFIG_HUGETLB_PAGE */