Loading...
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_lock
25 * page->flags PG_locked (lock_page) * (see huegtlbfs below)
26 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
27 * mapping->i_mmap_rwsem
28 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
29 * anon_vma->rwsem
30 * mm->page_table_lock or pte_lock
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
34 * lock_page_memcg move_lock (in __set_page_dirty_buffers)
35 * i_pages lock (widely used)
36 * lruvec->lru_lock (in lock_page_lruvec_irq)
37 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
38 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
39 * sb_lock (within inode_lock in fs/fs-writeback.c)
40 * i_pages lock (widely used, in set_page_dirty,
41 * in arch-dependent flush_dcache_mmap_lock,
42 * within bdi.wb->list_lock in __sync_single_inode)
43 *
44 * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon)
45 * ->tasklist_lock
46 * pte map lock
47 *
48 * * hugetlbfs PageHuge() pages take locks in this order:
49 * mapping->i_mmap_rwsem
50 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
51 * page->flags PG_locked (lock_page)
52 */
53
54#include <linux/mm.h>
55#include <linux/sched/mm.h>
56#include <linux/sched/task.h>
57#include <linux/pagemap.h>
58#include <linux/swap.h>
59#include <linux/swapops.h>
60#include <linux/slab.h>
61#include <linux/init.h>
62#include <linux/ksm.h>
63#include <linux/rmap.h>
64#include <linux/rcupdate.h>
65#include <linux/export.h>
66#include <linux/memcontrol.h>
67#include <linux/mmu_notifier.h>
68#include <linux/migrate.h>
69#include <linux/hugetlb.h>
70#include <linux/huge_mm.h>
71#include <linux/backing-dev.h>
72#include <linux/page_idle.h>
73#include <linux/memremap.h>
74#include <linux/userfaultfd_k.h>
75
76#include <asm/tlbflush.h>
77
78#include <trace/events/tlb.h>
79
80#include "internal.h"
81
82static struct kmem_cache *anon_vma_cachep;
83static struct kmem_cache *anon_vma_chain_cachep;
84
85static inline struct anon_vma *anon_vma_alloc(void)
86{
87 struct anon_vma *anon_vma;
88
89 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
90 if (anon_vma) {
91 atomic_set(&anon_vma->refcount, 1);
92 anon_vma->degree = 1; /* Reference for first vma */
93 anon_vma->parent = anon_vma;
94 /*
95 * Initialise the anon_vma root to point to itself. If called
96 * from fork, the root will be reset to the parents anon_vma.
97 */
98 anon_vma->root = anon_vma;
99 }
100
101 return anon_vma;
102}
103
104static inline void anon_vma_free(struct anon_vma *anon_vma)
105{
106 VM_BUG_ON(atomic_read(&anon_vma->refcount));
107
108 /*
109 * Synchronize against page_lock_anon_vma_read() such that
110 * we can safely hold the lock without the anon_vma getting
111 * freed.
112 *
113 * Relies on the full mb implied by the atomic_dec_and_test() from
114 * put_anon_vma() against the acquire barrier implied by
115 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
116 *
117 * page_lock_anon_vma_read() VS put_anon_vma()
118 * down_read_trylock() atomic_dec_and_test()
119 * LOCK MB
120 * atomic_read() rwsem_is_locked()
121 *
122 * LOCK should suffice since the actual taking of the lock must
123 * happen _before_ what follows.
124 */
125 might_sleep();
126 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
127 anon_vma_lock_write(anon_vma);
128 anon_vma_unlock_write(anon_vma);
129 }
130
131 kmem_cache_free(anon_vma_cachep, anon_vma);
132}
133
134static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
135{
136 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
137}
138
139static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
140{
141 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
142}
143
144static void anon_vma_chain_link(struct vm_area_struct *vma,
145 struct anon_vma_chain *avc,
146 struct anon_vma *anon_vma)
147{
148 avc->vma = vma;
149 avc->anon_vma = anon_vma;
150 list_add(&avc->same_vma, &vma->anon_vma_chain);
151 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
152}
153
154/**
155 * __anon_vma_prepare - attach an anon_vma to a memory region
156 * @vma: the memory region in question
157 *
158 * This makes sure the memory mapping described by 'vma' has
159 * an 'anon_vma' attached to it, so that we can associate the
160 * anonymous pages mapped into it with that anon_vma.
161 *
162 * The common case will be that we already have one, which
163 * is handled inline by anon_vma_prepare(). But if
164 * not we either need to find an adjacent mapping that we
165 * can re-use the anon_vma from (very common when the only
166 * reason for splitting a vma has been mprotect()), or we
167 * allocate a new one.
168 *
169 * Anon-vma allocations are very subtle, because we may have
170 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
171 * and that may actually touch the rwsem even in the newly
172 * allocated vma (it depends on RCU to make sure that the
173 * anon_vma isn't actually destroyed).
174 *
175 * As a result, we need to do proper anon_vma locking even
176 * for the new allocation. At the same time, we do not want
177 * to do any locking for the common case of already having
178 * an anon_vma.
179 *
180 * This must be called with the mmap_lock held for reading.
181 */
182int __anon_vma_prepare(struct vm_area_struct *vma)
183{
184 struct mm_struct *mm = vma->vm_mm;
185 struct anon_vma *anon_vma, *allocated;
186 struct anon_vma_chain *avc;
187
188 might_sleep();
189
190 avc = anon_vma_chain_alloc(GFP_KERNEL);
191 if (!avc)
192 goto out_enomem;
193
194 anon_vma = find_mergeable_anon_vma(vma);
195 allocated = NULL;
196 if (!anon_vma) {
197 anon_vma = anon_vma_alloc();
198 if (unlikely(!anon_vma))
199 goto out_enomem_free_avc;
200 allocated = anon_vma;
201 }
202
203 anon_vma_lock_write(anon_vma);
204 /* page_table_lock to protect against threads */
205 spin_lock(&mm->page_table_lock);
206 if (likely(!vma->anon_vma)) {
207 vma->anon_vma = anon_vma;
208 anon_vma_chain_link(vma, avc, anon_vma);
209 /* vma reference or self-parent link for new root */
210 anon_vma->degree++;
211 allocated = NULL;
212 avc = NULL;
213 }
214 spin_unlock(&mm->page_table_lock);
215 anon_vma_unlock_write(anon_vma);
216
217 if (unlikely(allocated))
218 put_anon_vma(allocated);
219 if (unlikely(avc))
220 anon_vma_chain_free(avc);
221
222 return 0;
223
224 out_enomem_free_avc:
225 anon_vma_chain_free(avc);
226 out_enomem:
227 return -ENOMEM;
228}
229
230/*
231 * This is a useful helper function for locking the anon_vma root as
232 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
233 * have the same vma.
234 *
235 * Such anon_vma's should have the same root, so you'd expect to see
236 * just a single mutex_lock for the whole traversal.
237 */
238static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
239{
240 struct anon_vma *new_root = anon_vma->root;
241 if (new_root != root) {
242 if (WARN_ON_ONCE(root))
243 up_write(&root->rwsem);
244 root = new_root;
245 down_write(&root->rwsem);
246 }
247 return root;
248}
249
250static inline void unlock_anon_vma_root(struct anon_vma *root)
251{
252 if (root)
253 up_write(&root->rwsem);
254}
255
256/*
257 * Attach the anon_vmas from src to dst.
258 * Returns 0 on success, -ENOMEM on failure.
259 *
260 * anon_vma_clone() is called by __vma_adjust(), __split_vma(), copy_vma() and
261 * anon_vma_fork(). The first three want an exact copy of src, while the last
262 * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
263 * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
264 * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
265 *
266 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
267 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
268 * This prevents degradation of anon_vma hierarchy to endless linear chain in
269 * case of constantly forking task. On the other hand, an anon_vma with more
270 * than one child isn't reused even if there was no alive vma, thus rmap
271 * walker has a good chance of avoiding scanning the whole hierarchy when it
272 * searches where page is mapped.
273 */
274int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
275{
276 struct anon_vma_chain *avc, *pavc;
277 struct anon_vma *root = NULL;
278
279 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
280 struct anon_vma *anon_vma;
281
282 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
283 if (unlikely(!avc)) {
284 unlock_anon_vma_root(root);
285 root = NULL;
286 avc = anon_vma_chain_alloc(GFP_KERNEL);
287 if (!avc)
288 goto enomem_failure;
289 }
290 anon_vma = pavc->anon_vma;
291 root = lock_anon_vma_root(root, anon_vma);
292 anon_vma_chain_link(dst, avc, anon_vma);
293
294 /*
295 * Reuse existing anon_vma if its degree lower than two,
296 * that means it has no vma and only one anon_vma child.
297 *
298 * Do not chose parent anon_vma, otherwise first child
299 * will always reuse it. Root anon_vma is never reused:
300 * it has self-parent reference and at least one child.
301 */
302 if (!dst->anon_vma && src->anon_vma &&
303 anon_vma != src->anon_vma && anon_vma->degree < 2)
304 dst->anon_vma = anon_vma;
305 }
306 if (dst->anon_vma)
307 dst->anon_vma->degree++;
308 unlock_anon_vma_root(root);
309 return 0;
310
311 enomem_failure:
312 /*
313 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
314 * decremented in unlink_anon_vmas().
315 * We can safely do this because callers of anon_vma_clone() don't care
316 * about dst->anon_vma if anon_vma_clone() failed.
317 */
318 dst->anon_vma = NULL;
319 unlink_anon_vmas(dst);
320 return -ENOMEM;
321}
322
323/*
324 * Attach vma to its own anon_vma, as well as to the anon_vmas that
325 * the corresponding VMA in the parent process is attached to.
326 * Returns 0 on success, non-zero on failure.
327 */
328int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
329{
330 struct anon_vma_chain *avc;
331 struct anon_vma *anon_vma;
332 int error;
333
334 /* Don't bother if the parent process has no anon_vma here. */
335 if (!pvma->anon_vma)
336 return 0;
337
338 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
339 vma->anon_vma = NULL;
340
341 /*
342 * First, attach the new VMA to the parent VMA's anon_vmas,
343 * so rmap can find non-COWed pages in child processes.
344 */
345 error = anon_vma_clone(vma, pvma);
346 if (error)
347 return error;
348
349 /* An existing anon_vma has been reused, all done then. */
350 if (vma->anon_vma)
351 return 0;
352
353 /* Then add our own anon_vma. */
354 anon_vma = anon_vma_alloc();
355 if (!anon_vma)
356 goto out_error;
357 avc = anon_vma_chain_alloc(GFP_KERNEL);
358 if (!avc)
359 goto out_error_free_anon_vma;
360
361 /*
362 * The root anon_vma's rwsem is the lock actually used when we
363 * lock any of the anon_vmas in this anon_vma tree.
364 */
365 anon_vma->root = pvma->anon_vma->root;
366 anon_vma->parent = pvma->anon_vma;
367 /*
368 * With refcounts, an anon_vma can stay around longer than the
369 * process it belongs to. The root anon_vma needs to be pinned until
370 * this anon_vma is freed, because the lock lives in the root.
371 */
372 get_anon_vma(anon_vma->root);
373 /* Mark this anon_vma as the one where our new (COWed) pages go. */
374 vma->anon_vma = anon_vma;
375 anon_vma_lock_write(anon_vma);
376 anon_vma_chain_link(vma, avc, anon_vma);
377 anon_vma->parent->degree++;
378 anon_vma_unlock_write(anon_vma);
379
380 return 0;
381
382 out_error_free_anon_vma:
383 put_anon_vma(anon_vma);
384 out_error:
385 unlink_anon_vmas(vma);
386 return -ENOMEM;
387}
388
389void unlink_anon_vmas(struct vm_area_struct *vma)
390{
391 struct anon_vma_chain *avc, *next;
392 struct anon_vma *root = NULL;
393
394 /*
395 * Unlink each anon_vma chained to the VMA. This list is ordered
396 * from newest to oldest, ensuring the root anon_vma gets freed last.
397 */
398 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
399 struct anon_vma *anon_vma = avc->anon_vma;
400
401 root = lock_anon_vma_root(root, anon_vma);
402 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
403
404 /*
405 * Leave empty anon_vmas on the list - we'll need
406 * to free them outside the lock.
407 */
408 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
409 anon_vma->parent->degree--;
410 continue;
411 }
412
413 list_del(&avc->same_vma);
414 anon_vma_chain_free(avc);
415 }
416 if (vma->anon_vma) {
417 vma->anon_vma->degree--;
418
419 /*
420 * vma would still be needed after unlink, and anon_vma will be prepared
421 * when handle fault.
422 */
423 vma->anon_vma = NULL;
424 }
425 unlock_anon_vma_root(root);
426
427 /*
428 * Iterate the list once more, it now only contains empty and unlinked
429 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
430 * needing to write-acquire the anon_vma->root->rwsem.
431 */
432 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
433 struct anon_vma *anon_vma = avc->anon_vma;
434
435 VM_WARN_ON(anon_vma->degree);
436 put_anon_vma(anon_vma);
437
438 list_del(&avc->same_vma);
439 anon_vma_chain_free(avc);
440 }
441}
442
443static void anon_vma_ctor(void *data)
444{
445 struct anon_vma *anon_vma = data;
446
447 init_rwsem(&anon_vma->rwsem);
448 atomic_set(&anon_vma->refcount, 0);
449 anon_vma->rb_root = RB_ROOT_CACHED;
450}
451
452void __init anon_vma_init(void)
453{
454 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
455 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
456 anon_vma_ctor);
457 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
458 SLAB_PANIC|SLAB_ACCOUNT);
459}
460
461/*
462 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
463 *
464 * Since there is no serialization what so ever against page_remove_rmap()
465 * the best this function can do is return a refcount increased anon_vma
466 * that might have been relevant to this page.
467 *
468 * The page might have been remapped to a different anon_vma or the anon_vma
469 * returned may already be freed (and even reused).
470 *
471 * In case it was remapped to a different anon_vma, the new anon_vma will be a
472 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
473 * ensure that any anon_vma obtained from the page will still be valid for as
474 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
475 *
476 * All users of this function must be very careful when walking the anon_vma
477 * chain and verify that the page in question is indeed mapped in it
478 * [ something equivalent to page_mapped_in_vma() ].
479 *
480 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
481 * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
482 * if there is a mapcount, we can dereference the anon_vma after observing
483 * those.
484 */
485struct anon_vma *page_get_anon_vma(struct page *page)
486{
487 struct anon_vma *anon_vma = NULL;
488 unsigned long anon_mapping;
489
490 rcu_read_lock();
491 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
492 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
493 goto out;
494 if (!page_mapped(page))
495 goto out;
496
497 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
498 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
499 anon_vma = NULL;
500 goto out;
501 }
502
503 /*
504 * If this page is still mapped, then its anon_vma cannot have been
505 * freed. But if it has been unmapped, we have no security against the
506 * anon_vma structure being freed and reused (for another anon_vma:
507 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
508 * above cannot corrupt).
509 */
510 if (!page_mapped(page)) {
511 rcu_read_unlock();
512 put_anon_vma(anon_vma);
513 return NULL;
514 }
515out:
516 rcu_read_unlock();
517
518 return anon_vma;
519}
520
521/*
522 * Similar to page_get_anon_vma() except it locks the anon_vma.
523 *
524 * Its a little more complex as it tries to keep the fast path to a single
525 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
526 * reference like with page_get_anon_vma() and then block on the mutex.
527 */
528struct anon_vma *page_lock_anon_vma_read(struct page *page)
529{
530 struct anon_vma *anon_vma = NULL;
531 struct anon_vma *root_anon_vma;
532 unsigned long anon_mapping;
533
534 rcu_read_lock();
535 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
536 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
537 goto out;
538 if (!page_mapped(page))
539 goto out;
540
541 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
542 root_anon_vma = READ_ONCE(anon_vma->root);
543 if (down_read_trylock(&root_anon_vma->rwsem)) {
544 /*
545 * If the page is still mapped, then this anon_vma is still
546 * its anon_vma, and holding the mutex ensures that it will
547 * not go away, see anon_vma_free().
548 */
549 if (!page_mapped(page)) {
550 up_read(&root_anon_vma->rwsem);
551 anon_vma = NULL;
552 }
553 goto out;
554 }
555
556 /* trylock failed, we got to sleep */
557 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
558 anon_vma = NULL;
559 goto out;
560 }
561
562 if (!page_mapped(page)) {
563 rcu_read_unlock();
564 put_anon_vma(anon_vma);
565 return NULL;
566 }
567
568 /* we pinned the anon_vma, its safe to sleep */
569 rcu_read_unlock();
570 anon_vma_lock_read(anon_vma);
571
572 if (atomic_dec_and_test(&anon_vma->refcount)) {
573 /*
574 * Oops, we held the last refcount, release the lock
575 * and bail -- can't simply use put_anon_vma() because
576 * we'll deadlock on the anon_vma_lock_write() recursion.
577 */
578 anon_vma_unlock_read(anon_vma);
579 __put_anon_vma(anon_vma);
580 anon_vma = NULL;
581 }
582
583 return anon_vma;
584
585out:
586 rcu_read_unlock();
587 return anon_vma;
588}
589
590void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
591{
592 anon_vma_unlock_read(anon_vma);
593}
594
595#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
596/*
597 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
598 * important if a PTE was dirty when it was unmapped that it's flushed
599 * before any IO is initiated on the page to prevent lost writes. Similarly,
600 * it must be flushed before freeing to prevent data leakage.
601 */
602void try_to_unmap_flush(void)
603{
604 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
605
606 if (!tlb_ubc->flush_required)
607 return;
608
609 arch_tlbbatch_flush(&tlb_ubc->arch);
610 tlb_ubc->flush_required = false;
611 tlb_ubc->writable = false;
612}
613
614/* Flush iff there are potentially writable TLB entries that can race with IO */
615void try_to_unmap_flush_dirty(void)
616{
617 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
618
619 if (tlb_ubc->writable)
620 try_to_unmap_flush();
621}
622
623static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
624{
625 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
626
627 arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
628 tlb_ubc->flush_required = true;
629
630 /*
631 * Ensure compiler does not re-order the setting of tlb_flush_batched
632 * before the PTE is cleared.
633 */
634 barrier();
635 mm->tlb_flush_batched = true;
636
637 /*
638 * If the PTE was dirty then it's best to assume it's writable. The
639 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
640 * before the page is queued for IO.
641 */
642 if (writable)
643 tlb_ubc->writable = true;
644}
645
646/*
647 * Returns true if the TLB flush should be deferred to the end of a batch of
648 * unmap operations to reduce IPIs.
649 */
650static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
651{
652 bool should_defer = false;
653
654 if (!(flags & TTU_BATCH_FLUSH))
655 return false;
656
657 /* If remote CPUs need to be flushed then defer batch the flush */
658 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
659 should_defer = true;
660 put_cpu();
661
662 return should_defer;
663}
664
665/*
666 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
667 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
668 * operation such as mprotect or munmap to race between reclaim unmapping
669 * the page and flushing the page. If this race occurs, it potentially allows
670 * access to data via a stale TLB entry. Tracking all mm's that have TLB
671 * batching in flight would be expensive during reclaim so instead track
672 * whether TLB batching occurred in the past and if so then do a flush here
673 * if required. This will cost one additional flush per reclaim cycle paid
674 * by the first operation at risk such as mprotect and mumap.
675 *
676 * This must be called under the PTL so that an access to tlb_flush_batched
677 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
678 * via the PTL.
679 */
680void flush_tlb_batched_pending(struct mm_struct *mm)
681{
682 if (data_race(mm->tlb_flush_batched)) {
683 flush_tlb_mm(mm);
684
685 /*
686 * Do not allow the compiler to re-order the clearing of
687 * tlb_flush_batched before the tlb is flushed.
688 */
689 barrier();
690 mm->tlb_flush_batched = false;
691 }
692}
693#else
694static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
695{
696}
697
698static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
699{
700 return false;
701}
702#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
703
704/*
705 * At what user virtual address is page expected in vma?
706 * Caller should check the page is actually part of the vma.
707 */
708unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
709{
710 if (PageAnon(page)) {
711 struct anon_vma *page__anon_vma = page_anon_vma(page);
712 /*
713 * Note: swapoff's unuse_vma() is more efficient with this
714 * check, and needs it to match anon_vma when KSM is active.
715 */
716 if (!vma->anon_vma || !page__anon_vma ||
717 vma->anon_vma->root != page__anon_vma->root)
718 return -EFAULT;
719 } else if (!vma->vm_file) {
720 return -EFAULT;
721 } else if (vma->vm_file->f_mapping != compound_head(page)->mapping) {
722 return -EFAULT;
723 }
724
725 return vma_address(page, vma);
726}
727
728pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
729{
730 pgd_t *pgd;
731 p4d_t *p4d;
732 pud_t *pud;
733 pmd_t *pmd = NULL;
734 pmd_t pmde;
735
736 pgd = pgd_offset(mm, address);
737 if (!pgd_present(*pgd))
738 goto out;
739
740 p4d = p4d_offset(pgd, address);
741 if (!p4d_present(*p4d))
742 goto out;
743
744 pud = pud_offset(p4d, address);
745 if (!pud_present(*pud))
746 goto out;
747
748 pmd = pmd_offset(pud, address);
749 /*
750 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
751 * without holding anon_vma lock for write. So when looking for a
752 * genuine pmde (in which to find pte), test present and !THP together.
753 */
754 pmde = *pmd;
755 barrier();
756 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
757 pmd = NULL;
758out:
759 return pmd;
760}
761
762struct page_referenced_arg {
763 int mapcount;
764 int referenced;
765 unsigned long vm_flags;
766 struct mem_cgroup *memcg;
767};
768/*
769 * arg: page_referenced_arg will be passed
770 */
771static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
772 unsigned long address, void *arg)
773{
774 struct page_referenced_arg *pra = arg;
775 struct page_vma_mapped_walk pvmw = {
776 .page = page,
777 .vma = vma,
778 .address = address,
779 };
780 int referenced = 0;
781
782 while (page_vma_mapped_walk(&pvmw)) {
783 address = pvmw.address;
784
785 if (vma->vm_flags & VM_LOCKED) {
786 page_vma_mapped_walk_done(&pvmw);
787 pra->vm_flags |= VM_LOCKED;
788 return false; /* To break the loop */
789 }
790
791 if (pvmw.pte) {
792 if (ptep_clear_flush_young_notify(vma, address,
793 pvmw.pte)) {
794 /*
795 * Don't treat a reference through
796 * a sequentially read mapping as such.
797 * If the page has been used in another mapping,
798 * we will catch it; if this other mapping is
799 * already gone, the unmap path will have set
800 * PG_referenced or activated the page.
801 */
802 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
803 referenced++;
804 }
805 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
806 if (pmdp_clear_flush_young_notify(vma, address,
807 pvmw.pmd))
808 referenced++;
809 } else {
810 /* unexpected pmd-mapped page? */
811 WARN_ON_ONCE(1);
812 }
813
814 pra->mapcount--;
815 }
816
817 if (referenced)
818 clear_page_idle(page);
819 if (test_and_clear_page_young(page))
820 referenced++;
821
822 if (referenced) {
823 pra->referenced++;
824 pra->vm_flags |= vma->vm_flags;
825 }
826
827 if (!pra->mapcount)
828 return false; /* To break the loop */
829
830 return true;
831}
832
833static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
834{
835 struct page_referenced_arg *pra = arg;
836 struct mem_cgroup *memcg = pra->memcg;
837
838 if (!mm_match_cgroup(vma->vm_mm, memcg))
839 return true;
840
841 return false;
842}
843
844/**
845 * page_referenced - test if the page was referenced
846 * @page: the page to test
847 * @is_locked: caller holds lock on the page
848 * @memcg: target memory cgroup
849 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
850 *
851 * Quick test_and_clear_referenced for all mappings to a page,
852 * returns the number of ptes which referenced the page.
853 */
854int page_referenced(struct page *page,
855 int is_locked,
856 struct mem_cgroup *memcg,
857 unsigned long *vm_flags)
858{
859 int we_locked = 0;
860 struct page_referenced_arg pra = {
861 .mapcount = total_mapcount(page),
862 .memcg = memcg,
863 };
864 struct rmap_walk_control rwc = {
865 .rmap_one = page_referenced_one,
866 .arg = (void *)&pra,
867 .anon_lock = page_lock_anon_vma_read,
868 };
869
870 *vm_flags = 0;
871 if (!pra.mapcount)
872 return 0;
873
874 if (!page_rmapping(page))
875 return 0;
876
877 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
878 we_locked = trylock_page(page);
879 if (!we_locked)
880 return 1;
881 }
882
883 /*
884 * If we are reclaiming on behalf of a cgroup, skip
885 * counting on behalf of references from different
886 * cgroups
887 */
888 if (memcg) {
889 rwc.invalid_vma = invalid_page_referenced_vma;
890 }
891
892 rmap_walk(page, &rwc);
893 *vm_flags = pra.vm_flags;
894
895 if (we_locked)
896 unlock_page(page);
897
898 return pra.referenced;
899}
900
901static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
902 unsigned long address, void *arg)
903{
904 struct page_vma_mapped_walk pvmw = {
905 .page = page,
906 .vma = vma,
907 .address = address,
908 .flags = PVMW_SYNC,
909 };
910 struct mmu_notifier_range range;
911 int *cleaned = arg;
912
913 /*
914 * We have to assume the worse case ie pmd for invalidation. Note that
915 * the page can not be free from this function.
916 */
917 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
918 0, vma, vma->vm_mm, address,
919 vma_address_end(page, vma));
920 mmu_notifier_invalidate_range_start(&range);
921
922 while (page_vma_mapped_walk(&pvmw)) {
923 int ret = 0;
924
925 address = pvmw.address;
926 if (pvmw.pte) {
927 pte_t entry;
928 pte_t *pte = pvmw.pte;
929
930 if (!pte_dirty(*pte) && !pte_write(*pte))
931 continue;
932
933 flush_cache_page(vma, address, pte_pfn(*pte));
934 entry = ptep_clear_flush(vma, address, pte);
935 entry = pte_wrprotect(entry);
936 entry = pte_mkclean(entry);
937 set_pte_at(vma->vm_mm, address, pte, entry);
938 ret = 1;
939 } else {
940#ifdef CONFIG_TRANSPARENT_HUGEPAGE
941 pmd_t *pmd = pvmw.pmd;
942 pmd_t entry;
943
944 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
945 continue;
946
947 flush_cache_page(vma, address, page_to_pfn(page));
948 entry = pmdp_invalidate(vma, address, pmd);
949 entry = pmd_wrprotect(entry);
950 entry = pmd_mkclean(entry);
951 set_pmd_at(vma->vm_mm, address, pmd, entry);
952 ret = 1;
953#else
954 /* unexpected pmd-mapped page? */
955 WARN_ON_ONCE(1);
956#endif
957 }
958
959 /*
960 * No need to call mmu_notifier_invalidate_range() as we are
961 * downgrading page table protection not changing it to point
962 * to a new page.
963 *
964 * See Documentation/vm/mmu_notifier.rst
965 */
966 if (ret)
967 (*cleaned)++;
968 }
969
970 mmu_notifier_invalidate_range_end(&range);
971
972 return true;
973}
974
975static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
976{
977 if (vma->vm_flags & VM_SHARED)
978 return false;
979
980 return true;
981}
982
983int page_mkclean(struct page *page)
984{
985 int cleaned = 0;
986 struct address_space *mapping;
987 struct rmap_walk_control rwc = {
988 .arg = (void *)&cleaned,
989 .rmap_one = page_mkclean_one,
990 .invalid_vma = invalid_mkclean_vma,
991 };
992
993 BUG_ON(!PageLocked(page));
994
995 if (!page_mapped(page))
996 return 0;
997
998 mapping = page_mapping(page);
999 if (!mapping)
1000 return 0;
1001
1002 rmap_walk(page, &rwc);
1003
1004 return cleaned;
1005}
1006EXPORT_SYMBOL_GPL(page_mkclean);
1007
1008/**
1009 * page_move_anon_rmap - move a page to our anon_vma
1010 * @page: the page to move to our anon_vma
1011 * @vma: the vma the page belongs to
1012 *
1013 * When a page belongs exclusively to one process after a COW event,
1014 * that page can be moved into the anon_vma that belongs to just that
1015 * process, so the rmap code will not search the parent or sibling
1016 * processes.
1017 */
1018void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1019{
1020 struct anon_vma *anon_vma = vma->anon_vma;
1021
1022 page = compound_head(page);
1023
1024 VM_BUG_ON_PAGE(!PageLocked(page), page);
1025 VM_BUG_ON_VMA(!anon_vma, vma);
1026
1027 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1028 /*
1029 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1030 * simultaneously, so a concurrent reader (eg page_referenced()'s
1031 * PageAnon()) will not see one without the other.
1032 */
1033 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1034}
1035
1036/**
1037 * __page_set_anon_rmap - set up new anonymous rmap
1038 * @page: Page or Hugepage to add to rmap
1039 * @vma: VM area to add page to.
1040 * @address: User virtual address of the mapping
1041 * @exclusive: the page is exclusively owned by the current process
1042 */
1043static void __page_set_anon_rmap(struct page *page,
1044 struct vm_area_struct *vma, unsigned long address, int exclusive)
1045{
1046 struct anon_vma *anon_vma = vma->anon_vma;
1047
1048 BUG_ON(!anon_vma);
1049
1050 if (PageAnon(page))
1051 return;
1052
1053 /*
1054 * If the page isn't exclusively mapped into this vma,
1055 * we must use the _oldest_ possible anon_vma for the
1056 * page mapping!
1057 */
1058 if (!exclusive)
1059 anon_vma = anon_vma->root;
1060
1061 /*
1062 * page_idle does a lockless/optimistic rmap scan on page->mapping.
1063 * Make sure the compiler doesn't split the stores of anon_vma and
1064 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1065 * could mistake the mapping for a struct address_space and crash.
1066 */
1067 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1068 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1069 page->index = linear_page_index(vma, address);
1070}
1071
1072/**
1073 * __page_check_anon_rmap - sanity check anonymous rmap addition
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 */
1078static void __page_check_anon_rmap(struct page *page,
1079 struct vm_area_struct *vma, unsigned long address)
1080{
1081 /*
1082 * The page's anon-rmap details (mapping and index) are guaranteed to
1083 * be set up correctly at this point.
1084 *
1085 * We have exclusion against page_add_anon_rmap because the caller
1086 * always holds the page locked.
1087 *
1088 * We have exclusion against page_add_new_anon_rmap because those pages
1089 * are initially only visible via the pagetables, and the pte is locked
1090 * over the call to page_add_new_anon_rmap.
1091 */
1092 VM_BUG_ON_PAGE(page_anon_vma(page)->root != vma->anon_vma->root, page);
1093 VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1094 page);
1095}
1096
1097/**
1098 * page_add_anon_rmap - add pte mapping to an anonymous page
1099 * @page: the page to add the mapping to
1100 * @vma: the vm area in which the mapping is added
1101 * @address: the user virtual address mapped
1102 * @compound: charge the page as compound or small page
1103 *
1104 * The caller needs to hold the pte lock, and the page must be locked in
1105 * the anon_vma case: to serialize mapping,index checking after setting,
1106 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1107 * (but PageKsm is never downgraded to PageAnon).
1108 */
1109void page_add_anon_rmap(struct page *page,
1110 struct vm_area_struct *vma, unsigned long address, bool compound)
1111{
1112 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1113}
1114
1115/*
1116 * Special version of the above for do_swap_page, which often runs
1117 * into pages that are exclusively owned by the current process.
1118 * Everybody else should continue to use page_add_anon_rmap above.
1119 */
1120void do_page_add_anon_rmap(struct page *page,
1121 struct vm_area_struct *vma, unsigned long address, int flags)
1122{
1123 bool compound = flags & RMAP_COMPOUND;
1124 bool first;
1125
1126 if (unlikely(PageKsm(page)))
1127 lock_page_memcg(page);
1128 else
1129 VM_BUG_ON_PAGE(!PageLocked(page), page);
1130
1131 if (compound) {
1132 atomic_t *mapcount;
1133 VM_BUG_ON_PAGE(!PageLocked(page), page);
1134 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1135 mapcount = compound_mapcount_ptr(page);
1136 first = atomic_inc_and_test(mapcount);
1137 } else {
1138 first = atomic_inc_and_test(&page->_mapcount);
1139 }
1140
1141 if (first) {
1142 int nr = compound ? thp_nr_pages(page) : 1;
1143 /*
1144 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1145 * these counters are not modified in interrupt context, and
1146 * pte lock(a spinlock) is held, which implies preemption
1147 * disabled.
1148 */
1149 if (compound)
1150 __mod_lruvec_page_state(page, NR_ANON_THPS, nr);
1151 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1152 }
1153
1154 if (unlikely(PageKsm(page))) {
1155 unlock_page_memcg(page);
1156 return;
1157 }
1158
1159 /* address might be in next vma when migration races vma_adjust */
1160 if (first)
1161 __page_set_anon_rmap(page, vma, address,
1162 flags & RMAP_EXCLUSIVE);
1163 else
1164 __page_check_anon_rmap(page, vma, address);
1165}
1166
1167/**
1168 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1169 * @page: the page to add the mapping to
1170 * @vma: the vm area in which the mapping is added
1171 * @address: the user virtual address mapped
1172 * @compound: charge the page as compound or small page
1173 *
1174 * Same as page_add_anon_rmap but must only be called on *new* pages.
1175 * This means the inc-and-test can be bypassed.
1176 * Page does not have to be locked.
1177 */
1178void page_add_new_anon_rmap(struct page *page,
1179 struct vm_area_struct *vma, unsigned long address, bool compound)
1180{
1181 int nr = compound ? thp_nr_pages(page) : 1;
1182
1183 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1184 __SetPageSwapBacked(page);
1185 if (compound) {
1186 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1187 /* increment count (starts at -1) */
1188 atomic_set(compound_mapcount_ptr(page), 0);
1189 if (hpage_pincount_available(page))
1190 atomic_set(compound_pincount_ptr(page), 0);
1191
1192 __mod_lruvec_page_state(page, NR_ANON_THPS, nr);
1193 } else {
1194 /* Anon THP always mapped first with PMD */
1195 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1196 /* increment count (starts at -1) */
1197 atomic_set(&page->_mapcount, 0);
1198 }
1199 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1200 __page_set_anon_rmap(page, vma, address, 1);
1201}
1202
1203/**
1204 * page_add_file_rmap - add pte mapping to a file page
1205 * @page: the page to add the mapping to
1206 * @compound: charge the page as compound or small page
1207 *
1208 * The caller needs to hold the pte lock.
1209 */
1210void page_add_file_rmap(struct page *page, bool compound)
1211{
1212 int i, nr = 1;
1213
1214 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1215 lock_page_memcg(page);
1216 if (compound && PageTransHuge(page)) {
1217 int nr_pages = thp_nr_pages(page);
1218
1219 for (i = 0, nr = 0; i < nr_pages; i++) {
1220 if (atomic_inc_and_test(&page[i]._mapcount))
1221 nr++;
1222 }
1223 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1224 goto out;
1225 if (PageSwapBacked(page))
1226 __mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED,
1227 nr_pages);
1228 else
1229 __mod_lruvec_page_state(page, NR_FILE_PMDMAPPED,
1230 nr_pages);
1231 } else {
1232 if (PageTransCompound(page) && page_mapping(page)) {
1233 VM_WARN_ON_ONCE(!PageLocked(page));
1234
1235 SetPageDoubleMap(compound_head(page));
1236 if (PageMlocked(page))
1237 clear_page_mlock(compound_head(page));
1238 }
1239 if (!atomic_inc_and_test(&page->_mapcount))
1240 goto out;
1241 }
1242 __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1243out:
1244 unlock_page_memcg(page);
1245}
1246
1247static void page_remove_file_rmap(struct page *page, bool compound)
1248{
1249 int i, nr = 1;
1250
1251 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1252
1253 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1254 if (unlikely(PageHuge(page))) {
1255 /* hugetlb pages are always mapped with pmds */
1256 atomic_dec(compound_mapcount_ptr(page));
1257 return;
1258 }
1259
1260 /* page still mapped by someone else? */
1261 if (compound && PageTransHuge(page)) {
1262 int nr_pages = thp_nr_pages(page);
1263
1264 for (i = 0, nr = 0; i < nr_pages; i++) {
1265 if (atomic_add_negative(-1, &page[i]._mapcount))
1266 nr++;
1267 }
1268 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1269 return;
1270 if (PageSwapBacked(page))
1271 __mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED,
1272 -nr_pages);
1273 else
1274 __mod_lruvec_page_state(page, NR_FILE_PMDMAPPED,
1275 -nr_pages);
1276 } else {
1277 if (!atomic_add_negative(-1, &page->_mapcount))
1278 return;
1279 }
1280
1281 /*
1282 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1283 * these counters are not modified in interrupt context, and
1284 * pte lock(a spinlock) is held, which implies preemption disabled.
1285 */
1286 __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1287
1288 if (unlikely(PageMlocked(page)))
1289 clear_page_mlock(page);
1290}
1291
1292static void page_remove_anon_compound_rmap(struct page *page)
1293{
1294 int i, nr;
1295
1296 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1297 return;
1298
1299 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1300 if (unlikely(PageHuge(page)))
1301 return;
1302
1303 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1304 return;
1305
1306 __mod_lruvec_page_state(page, NR_ANON_THPS, -thp_nr_pages(page));
1307
1308 if (TestClearPageDoubleMap(page)) {
1309 /*
1310 * Subpages can be mapped with PTEs too. Check how many of
1311 * them are still mapped.
1312 */
1313 for (i = 0, nr = 0; i < thp_nr_pages(page); i++) {
1314 if (atomic_add_negative(-1, &page[i]._mapcount))
1315 nr++;
1316 }
1317
1318 /*
1319 * Queue the page for deferred split if at least one small
1320 * page of the compound page is unmapped, but at least one
1321 * small page is still mapped.
1322 */
1323 if (nr && nr < thp_nr_pages(page))
1324 deferred_split_huge_page(page);
1325 } else {
1326 nr = thp_nr_pages(page);
1327 }
1328
1329 if (unlikely(PageMlocked(page)))
1330 clear_page_mlock(page);
1331
1332 if (nr)
1333 __mod_lruvec_page_state(page, NR_ANON_MAPPED, -nr);
1334}
1335
1336/**
1337 * page_remove_rmap - take down pte mapping from a page
1338 * @page: page to remove mapping from
1339 * @compound: uncharge the page as compound or small page
1340 *
1341 * The caller needs to hold the pte lock.
1342 */
1343void page_remove_rmap(struct page *page, bool compound)
1344{
1345 lock_page_memcg(page);
1346
1347 if (!PageAnon(page)) {
1348 page_remove_file_rmap(page, compound);
1349 goto out;
1350 }
1351
1352 if (compound) {
1353 page_remove_anon_compound_rmap(page);
1354 goto out;
1355 }
1356
1357 /* page still mapped by someone else? */
1358 if (!atomic_add_negative(-1, &page->_mapcount))
1359 goto out;
1360
1361 /*
1362 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1363 * these counters are not modified in interrupt context, and
1364 * pte lock(a spinlock) is held, which implies preemption disabled.
1365 */
1366 __dec_lruvec_page_state(page, NR_ANON_MAPPED);
1367
1368 if (unlikely(PageMlocked(page)))
1369 clear_page_mlock(page);
1370
1371 if (PageTransCompound(page))
1372 deferred_split_huge_page(compound_head(page));
1373
1374 /*
1375 * It would be tidy to reset the PageAnon mapping here,
1376 * but that might overwrite a racing page_add_anon_rmap
1377 * which increments mapcount after us but sets mapping
1378 * before us: so leave the reset to free_unref_page,
1379 * and remember that it's only reliable while mapped.
1380 * Leaving it set also helps swapoff to reinstate ptes
1381 * faster for those pages still in swapcache.
1382 */
1383out:
1384 unlock_page_memcg(page);
1385}
1386
1387/*
1388 * @arg: enum ttu_flags will be passed to this argument
1389 */
1390static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1391 unsigned long address, void *arg)
1392{
1393 struct mm_struct *mm = vma->vm_mm;
1394 struct page_vma_mapped_walk pvmw = {
1395 .page = page,
1396 .vma = vma,
1397 .address = address,
1398 };
1399 pte_t pteval;
1400 struct page *subpage;
1401 bool ret = true;
1402 struct mmu_notifier_range range;
1403 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1404
1405 /*
1406 * When racing against e.g. zap_pte_range() on another cpu,
1407 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1408 * try_to_unmap() may return before page_mapped() has become false,
1409 * if page table locking is skipped: use TTU_SYNC to wait for that.
1410 */
1411 if (flags & TTU_SYNC)
1412 pvmw.flags = PVMW_SYNC;
1413
1414 if (flags & TTU_SPLIT_HUGE_PMD)
1415 split_huge_pmd_address(vma, address, false, page);
1416
1417 /*
1418 * For THP, we have to assume the worse case ie pmd for invalidation.
1419 * For hugetlb, it could be much worse if we need to do pud
1420 * invalidation in the case of pmd sharing.
1421 *
1422 * Note that the page can not be free in this function as call of
1423 * try_to_unmap() must hold a reference on the page.
1424 */
1425 range.end = PageKsm(page) ?
1426 address + PAGE_SIZE : vma_address_end(page, vma);
1427 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1428 address, range.end);
1429 if (PageHuge(page)) {
1430 /*
1431 * If sharing is possible, start and end will be adjusted
1432 * accordingly.
1433 */
1434 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1435 &range.end);
1436 }
1437 mmu_notifier_invalidate_range_start(&range);
1438
1439 while (page_vma_mapped_walk(&pvmw)) {
1440 /*
1441 * If the page is mlock()d, we cannot swap it out.
1442 */
1443 if (!(flags & TTU_IGNORE_MLOCK) &&
1444 (vma->vm_flags & VM_LOCKED)) {
1445 /*
1446 * PTE-mapped THP are never marked as mlocked: so do
1447 * not set it on a DoubleMap THP, nor on an Anon THP
1448 * (which may still be PTE-mapped after DoubleMap was
1449 * cleared). But stop unmapping even in those cases.
1450 */
1451 if (!PageTransCompound(page) || (PageHead(page) &&
1452 !PageDoubleMap(page) && !PageAnon(page)))
1453 mlock_vma_page(page);
1454 page_vma_mapped_walk_done(&pvmw);
1455 ret = false;
1456 break;
1457 }
1458
1459 /* Unexpected PMD-mapped THP? */
1460 VM_BUG_ON_PAGE(!pvmw.pte, page);
1461
1462 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1463 address = pvmw.address;
1464
1465 if (PageHuge(page) && !PageAnon(page)) {
1466 /*
1467 * To call huge_pmd_unshare, i_mmap_rwsem must be
1468 * held in write mode. Caller needs to explicitly
1469 * do this outside rmap routines.
1470 */
1471 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1472 if (huge_pmd_unshare(mm, vma, &address, pvmw.pte)) {
1473 /*
1474 * huge_pmd_unshare unmapped an entire PMD
1475 * page. There is no way of knowing exactly
1476 * which PMDs may be cached for this mm, so
1477 * we must flush them all. start/end were
1478 * already adjusted above to cover this range.
1479 */
1480 flush_cache_range(vma, range.start, range.end);
1481 flush_tlb_range(vma, range.start, range.end);
1482 mmu_notifier_invalidate_range(mm, range.start,
1483 range.end);
1484
1485 /*
1486 * The ref count of the PMD page was dropped
1487 * which is part of the way map counting
1488 * is done for shared PMDs. Return 'true'
1489 * here. When there is no other sharing,
1490 * huge_pmd_unshare returns false and we will
1491 * unmap the actual page and drop map count
1492 * to zero.
1493 */
1494 page_vma_mapped_walk_done(&pvmw);
1495 break;
1496 }
1497 }
1498
1499 /* Nuke the page table entry. */
1500 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1501 if (should_defer_flush(mm, flags)) {
1502 /*
1503 * We clear the PTE but do not flush so potentially
1504 * a remote CPU could still be writing to the page.
1505 * If the entry was previously clean then the
1506 * architecture must guarantee that a clear->dirty
1507 * transition on a cached TLB entry is written through
1508 * and traps if the PTE is unmapped.
1509 */
1510 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1511
1512 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1513 } else {
1514 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1515 }
1516
1517 /* Move the dirty bit to the page. Now the pte is gone. */
1518 if (pte_dirty(pteval))
1519 set_page_dirty(page);
1520
1521 /* Update high watermark before we lower rss */
1522 update_hiwater_rss(mm);
1523
1524 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1525 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1526 if (PageHuge(page)) {
1527 hugetlb_count_sub(compound_nr(page), mm);
1528 set_huge_swap_pte_at(mm, address,
1529 pvmw.pte, pteval,
1530 vma_mmu_pagesize(vma));
1531 } else {
1532 dec_mm_counter(mm, mm_counter(page));
1533 set_pte_at(mm, address, pvmw.pte, pteval);
1534 }
1535
1536 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1537 /*
1538 * The guest indicated that the page content is of no
1539 * interest anymore. Simply discard the pte, vmscan
1540 * will take care of the rest.
1541 * A future reference will then fault in a new zero
1542 * page. When userfaultfd is active, we must not drop
1543 * this page though, as its main user (postcopy
1544 * migration) will not expect userfaults on already
1545 * copied pages.
1546 */
1547 dec_mm_counter(mm, mm_counter(page));
1548 /* We have to invalidate as we cleared the pte */
1549 mmu_notifier_invalidate_range(mm, address,
1550 address + PAGE_SIZE);
1551 } else if (PageAnon(page)) {
1552 swp_entry_t entry = { .val = page_private(subpage) };
1553 pte_t swp_pte;
1554 /*
1555 * Store the swap location in the pte.
1556 * See handle_pte_fault() ...
1557 */
1558 if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1559 WARN_ON_ONCE(1);
1560 ret = false;
1561 /* We have to invalidate as we cleared the pte */
1562 mmu_notifier_invalidate_range(mm, address,
1563 address + PAGE_SIZE);
1564 page_vma_mapped_walk_done(&pvmw);
1565 break;
1566 }
1567
1568 /* MADV_FREE page check */
1569 if (!PageSwapBacked(page)) {
1570 if (!PageDirty(page)) {
1571 /* Invalidate as we cleared the pte */
1572 mmu_notifier_invalidate_range(mm,
1573 address, address + PAGE_SIZE);
1574 dec_mm_counter(mm, MM_ANONPAGES);
1575 goto discard;
1576 }
1577
1578 /*
1579 * If the page was redirtied, it cannot be
1580 * discarded. Remap the page to page table.
1581 */
1582 set_pte_at(mm, address, pvmw.pte, pteval);
1583 SetPageSwapBacked(page);
1584 ret = false;
1585 page_vma_mapped_walk_done(&pvmw);
1586 break;
1587 }
1588
1589 if (swap_duplicate(entry) < 0) {
1590 set_pte_at(mm, address, pvmw.pte, pteval);
1591 ret = false;
1592 page_vma_mapped_walk_done(&pvmw);
1593 break;
1594 }
1595 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1596 set_pte_at(mm, address, pvmw.pte, pteval);
1597 ret = false;
1598 page_vma_mapped_walk_done(&pvmw);
1599 break;
1600 }
1601 if (list_empty(&mm->mmlist)) {
1602 spin_lock(&mmlist_lock);
1603 if (list_empty(&mm->mmlist))
1604 list_add(&mm->mmlist, &init_mm.mmlist);
1605 spin_unlock(&mmlist_lock);
1606 }
1607 dec_mm_counter(mm, MM_ANONPAGES);
1608 inc_mm_counter(mm, MM_SWAPENTS);
1609 swp_pte = swp_entry_to_pte(entry);
1610 if (pte_soft_dirty(pteval))
1611 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1612 if (pte_uffd_wp(pteval))
1613 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1614 set_pte_at(mm, address, pvmw.pte, swp_pte);
1615 /* Invalidate as we cleared the pte */
1616 mmu_notifier_invalidate_range(mm, address,
1617 address + PAGE_SIZE);
1618 } else {
1619 /*
1620 * This is a locked file-backed page, thus it cannot
1621 * be removed from the page cache and replaced by a new
1622 * page before mmu_notifier_invalidate_range_end, so no
1623 * concurrent thread might update its page table to
1624 * point at new page while a device still is using this
1625 * page.
1626 *
1627 * See Documentation/vm/mmu_notifier.rst
1628 */
1629 dec_mm_counter(mm, mm_counter_file(page));
1630 }
1631discard:
1632 /*
1633 * No need to call mmu_notifier_invalidate_range() it has be
1634 * done above for all cases requiring it to happen under page
1635 * table lock before mmu_notifier_invalidate_range_end()
1636 *
1637 * See Documentation/vm/mmu_notifier.rst
1638 */
1639 page_remove_rmap(subpage, PageHuge(page));
1640 put_page(page);
1641 }
1642
1643 mmu_notifier_invalidate_range_end(&range);
1644
1645 return ret;
1646}
1647
1648static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1649{
1650 return vma_is_temporary_stack(vma);
1651}
1652
1653static int page_not_mapped(struct page *page)
1654{
1655 return !page_mapped(page);
1656}
1657
1658/**
1659 * try_to_unmap - try to remove all page table mappings to a page
1660 * @page: the page to get unmapped
1661 * @flags: action and flags
1662 *
1663 * Tries to remove all the page table entries which are mapping this
1664 * page, used in the pageout path. Caller must hold the page lock.
1665 *
1666 * It is the caller's responsibility to check if the page is still
1667 * mapped when needed (use TTU_SYNC to prevent accounting races).
1668 */
1669void try_to_unmap(struct page *page, enum ttu_flags flags)
1670{
1671 struct rmap_walk_control rwc = {
1672 .rmap_one = try_to_unmap_one,
1673 .arg = (void *)flags,
1674 .done = page_not_mapped,
1675 .anon_lock = page_lock_anon_vma_read,
1676 };
1677
1678 if (flags & TTU_RMAP_LOCKED)
1679 rmap_walk_locked(page, &rwc);
1680 else
1681 rmap_walk(page, &rwc);
1682}
1683
1684/*
1685 * @arg: enum ttu_flags will be passed to this argument.
1686 *
1687 * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
1688 * containing migration entries.
1689 */
1690static bool try_to_migrate_one(struct page *page, struct vm_area_struct *vma,
1691 unsigned long address, void *arg)
1692{
1693 struct mm_struct *mm = vma->vm_mm;
1694 struct page_vma_mapped_walk pvmw = {
1695 .page = page,
1696 .vma = vma,
1697 .address = address,
1698 };
1699 pte_t pteval;
1700 struct page *subpage;
1701 bool ret = true;
1702 struct mmu_notifier_range range;
1703 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1704
1705 /*
1706 * When racing against e.g. zap_pte_range() on another cpu,
1707 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1708 * try_to_migrate() may return before page_mapped() has become false,
1709 * if page table locking is skipped: use TTU_SYNC to wait for that.
1710 */
1711 if (flags & TTU_SYNC)
1712 pvmw.flags = PVMW_SYNC;
1713
1714 /*
1715 * unmap_page() in mm/huge_memory.c is the only user of migration with
1716 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
1717 */
1718 if (flags & TTU_SPLIT_HUGE_PMD)
1719 split_huge_pmd_address(vma, address, true, page);
1720
1721 /*
1722 * For THP, we have to assume the worse case ie pmd for invalidation.
1723 * For hugetlb, it could be much worse if we need to do pud
1724 * invalidation in the case of pmd sharing.
1725 *
1726 * Note that the page can not be free in this function as call of
1727 * try_to_unmap() must hold a reference on the page.
1728 */
1729 range.end = PageKsm(page) ?
1730 address + PAGE_SIZE : vma_address_end(page, vma);
1731 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1732 address, range.end);
1733 if (PageHuge(page)) {
1734 /*
1735 * If sharing is possible, start and end will be adjusted
1736 * accordingly.
1737 */
1738 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1739 &range.end);
1740 }
1741 mmu_notifier_invalidate_range_start(&range);
1742
1743 while (page_vma_mapped_walk(&pvmw)) {
1744#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1745 /* PMD-mapped THP migration entry */
1746 if (!pvmw.pte) {
1747 VM_BUG_ON_PAGE(PageHuge(page) ||
1748 !PageTransCompound(page), page);
1749
1750 set_pmd_migration_entry(&pvmw, page);
1751 continue;
1752 }
1753#endif
1754
1755 /* Unexpected PMD-mapped THP? */
1756 VM_BUG_ON_PAGE(!pvmw.pte, page);
1757
1758 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1759 address = pvmw.address;
1760
1761 if (PageHuge(page) && !PageAnon(page)) {
1762 /*
1763 * To call huge_pmd_unshare, i_mmap_rwsem must be
1764 * held in write mode. Caller needs to explicitly
1765 * do this outside rmap routines.
1766 */
1767 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1768 if (huge_pmd_unshare(mm, vma, &address, pvmw.pte)) {
1769 /*
1770 * huge_pmd_unshare unmapped an entire PMD
1771 * page. There is no way of knowing exactly
1772 * which PMDs may be cached for this mm, so
1773 * we must flush them all. start/end were
1774 * already adjusted above to cover this range.
1775 */
1776 flush_cache_range(vma, range.start, range.end);
1777 flush_tlb_range(vma, range.start, range.end);
1778 mmu_notifier_invalidate_range(mm, range.start,
1779 range.end);
1780
1781 /*
1782 * The ref count of the PMD page was dropped
1783 * which is part of the way map counting
1784 * is done for shared PMDs. Return 'true'
1785 * here. When there is no other sharing,
1786 * huge_pmd_unshare returns false and we will
1787 * unmap the actual page and drop map count
1788 * to zero.
1789 */
1790 page_vma_mapped_walk_done(&pvmw);
1791 break;
1792 }
1793 }
1794
1795 /* Nuke the page table entry. */
1796 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1797 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1798
1799 /* Move the dirty bit to the page. Now the pte is gone. */
1800 if (pte_dirty(pteval))
1801 set_page_dirty(page);
1802
1803 /* Update high watermark before we lower rss */
1804 update_hiwater_rss(mm);
1805
1806 if (is_zone_device_page(page)) {
1807 swp_entry_t entry;
1808 pte_t swp_pte;
1809
1810 /*
1811 * Store the pfn of the page in a special migration
1812 * pte. do_swap_page() will wait until the migration
1813 * pte is removed and then restart fault handling.
1814 */
1815 entry = make_readable_migration_entry(
1816 page_to_pfn(page));
1817 swp_pte = swp_entry_to_pte(entry);
1818
1819 /*
1820 * pteval maps a zone device page and is therefore
1821 * a swap pte.
1822 */
1823 if (pte_swp_soft_dirty(pteval))
1824 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1825 if (pte_swp_uffd_wp(pteval))
1826 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1827 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1828 /*
1829 * No need to invalidate here it will synchronize on
1830 * against the special swap migration pte.
1831 *
1832 * The assignment to subpage above was computed from a
1833 * swap PTE which results in an invalid pointer.
1834 * Since only PAGE_SIZE pages can currently be
1835 * migrated, just set it to page. This will need to be
1836 * changed when hugepage migrations to device private
1837 * memory are supported.
1838 */
1839 subpage = page;
1840 } else if (PageHWPoison(page)) {
1841 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1842 if (PageHuge(page)) {
1843 hugetlb_count_sub(compound_nr(page), mm);
1844 set_huge_swap_pte_at(mm, address,
1845 pvmw.pte, pteval,
1846 vma_mmu_pagesize(vma));
1847 } else {
1848 dec_mm_counter(mm, mm_counter(page));
1849 set_pte_at(mm, address, pvmw.pte, pteval);
1850 }
1851
1852 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1853 /*
1854 * The guest indicated that the page content is of no
1855 * interest anymore. Simply discard the pte, vmscan
1856 * will take care of the rest.
1857 * A future reference will then fault in a new zero
1858 * page. When userfaultfd is active, we must not drop
1859 * this page though, as its main user (postcopy
1860 * migration) will not expect userfaults on already
1861 * copied pages.
1862 */
1863 dec_mm_counter(mm, mm_counter(page));
1864 /* We have to invalidate as we cleared the pte */
1865 mmu_notifier_invalidate_range(mm, address,
1866 address + PAGE_SIZE);
1867 } else {
1868 swp_entry_t entry;
1869 pte_t swp_pte;
1870
1871 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1872 set_pte_at(mm, address, pvmw.pte, pteval);
1873 ret = false;
1874 page_vma_mapped_walk_done(&pvmw);
1875 break;
1876 }
1877
1878 /*
1879 * Store the pfn of the page in a special migration
1880 * pte. do_swap_page() will wait until the migration
1881 * pte is removed and then restart fault handling.
1882 */
1883 if (pte_write(pteval))
1884 entry = make_writable_migration_entry(
1885 page_to_pfn(subpage));
1886 else
1887 entry = make_readable_migration_entry(
1888 page_to_pfn(subpage));
1889
1890 swp_pte = swp_entry_to_pte(entry);
1891 if (pte_soft_dirty(pteval))
1892 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1893 if (pte_uffd_wp(pteval))
1894 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1895 set_pte_at(mm, address, pvmw.pte, swp_pte);
1896 /*
1897 * No need to invalidate here it will synchronize on
1898 * against the special swap migration pte.
1899 */
1900 }
1901
1902 /*
1903 * No need to call mmu_notifier_invalidate_range() it has be
1904 * done above for all cases requiring it to happen under page
1905 * table lock before mmu_notifier_invalidate_range_end()
1906 *
1907 * See Documentation/vm/mmu_notifier.rst
1908 */
1909 page_remove_rmap(subpage, PageHuge(page));
1910 put_page(page);
1911 }
1912
1913 mmu_notifier_invalidate_range_end(&range);
1914
1915 return ret;
1916}
1917
1918/**
1919 * try_to_migrate - try to replace all page table mappings with swap entries
1920 * @page: the page to replace page table entries for
1921 * @flags: action and flags
1922 *
1923 * Tries to remove all the page table entries which are mapping this page and
1924 * replace them with special swap entries. Caller must hold the page lock.
1925 */
1926void try_to_migrate(struct page *page, enum ttu_flags flags)
1927{
1928 struct rmap_walk_control rwc = {
1929 .rmap_one = try_to_migrate_one,
1930 .arg = (void *)flags,
1931 .done = page_not_mapped,
1932 .anon_lock = page_lock_anon_vma_read,
1933 };
1934
1935 /*
1936 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
1937 * TTU_SPLIT_HUGE_PMD and TTU_SYNC flags.
1938 */
1939 if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
1940 TTU_SYNC)))
1941 return;
1942
1943 if (is_zone_device_page(page) && !is_device_private_page(page))
1944 return;
1945
1946 /*
1947 * During exec, a temporary VMA is setup and later moved.
1948 * The VMA is moved under the anon_vma lock but not the
1949 * page tables leading to a race where migration cannot
1950 * find the migration ptes. Rather than increasing the
1951 * locking requirements of exec(), migration skips
1952 * temporary VMAs until after exec() completes.
1953 */
1954 if (!PageKsm(page) && PageAnon(page))
1955 rwc.invalid_vma = invalid_migration_vma;
1956
1957 if (flags & TTU_RMAP_LOCKED)
1958 rmap_walk_locked(page, &rwc);
1959 else
1960 rmap_walk(page, &rwc);
1961}
1962
1963/*
1964 * Walks the vma's mapping a page and mlocks the page if any locked vma's are
1965 * found. Once one is found the page is locked and the scan can be terminated.
1966 */
1967static bool page_mlock_one(struct page *page, struct vm_area_struct *vma,
1968 unsigned long address, void *unused)
1969{
1970 struct page_vma_mapped_walk pvmw = {
1971 .page = page,
1972 .vma = vma,
1973 .address = address,
1974 };
1975
1976 /* An un-locked vma doesn't have any pages to lock, continue the scan */
1977 if (!(vma->vm_flags & VM_LOCKED))
1978 return true;
1979
1980 while (page_vma_mapped_walk(&pvmw)) {
1981 /*
1982 * Need to recheck under the ptl to serialise with
1983 * __munlock_pagevec_fill() after VM_LOCKED is cleared in
1984 * munlock_vma_pages_range().
1985 */
1986 if (vma->vm_flags & VM_LOCKED) {
1987 /*
1988 * PTE-mapped THP are never marked as mlocked; but
1989 * this function is never called on a DoubleMap THP,
1990 * nor on an Anon THP (which may still be PTE-mapped
1991 * after DoubleMap was cleared).
1992 */
1993 mlock_vma_page(page);
1994 /*
1995 * No need to scan further once the page is marked
1996 * as mlocked.
1997 */
1998 page_vma_mapped_walk_done(&pvmw);
1999 return false;
2000 }
2001 }
2002
2003 return true;
2004}
2005
2006/**
2007 * page_mlock - try to mlock a page
2008 * @page: the page to be mlocked
2009 *
2010 * Called from munlock code. Checks all of the VMAs mapping the page and mlocks
2011 * the page if any are found. The page will be returned with PG_mlocked cleared
2012 * if it is not mapped by any locked vmas.
2013 */
2014void page_mlock(struct page *page)
2015{
2016 struct rmap_walk_control rwc = {
2017 .rmap_one = page_mlock_one,
2018 .done = page_not_mapped,
2019 .anon_lock = page_lock_anon_vma_read,
2020
2021 };
2022
2023 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
2024 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
2025
2026 /* Anon THP are only marked as mlocked when singly mapped */
2027 if (PageTransCompound(page) && PageAnon(page))
2028 return;
2029
2030 rmap_walk(page, &rwc);
2031}
2032
2033#ifdef CONFIG_DEVICE_PRIVATE
2034struct make_exclusive_args {
2035 struct mm_struct *mm;
2036 unsigned long address;
2037 void *owner;
2038 bool valid;
2039};
2040
2041static bool page_make_device_exclusive_one(struct page *page,
2042 struct vm_area_struct *vma, unsigned long address, void *priv)
2043{
2044 struct mm_struct *mm = vma->vm_mm;
2045 struct page_vma_mapped_walk pvmw = {
2046 .page = page,
2047 .vma = vma,
2048 .address = address,
2049 };
2050 struct make_exclusive_args *args = priv;
2051 pte_t pteval;
2052 struct page *subpage;
2053 bool ret = true;
2054 struct mmu_notifier_range range;
2055 swp_entry_t entry;
2056 pte_t swp_pte;
2057
2058 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
2059 vma->vm_mm, address, min(vma->vm_end,
2060 address + page_size(page)), args->owner);
2061 mmu_notifier_invalidate_range_start(&range);
2062
2063 while (page_vma_mapped_walk(&pvmw)) {
2064 /* Unexpected PMD-mapped THP? */
2065 VM_BUG_ON_PAGE(!pvmw.pte, page);
2066
2067 if (!pte_present(*pvmw.pte)) {
2068 ret = false;
2069 page_vma_mapped_walk_done(&pvmw);
2070 break;
2071 }
2072
2073 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
2074 address = pvmw.address;
2075
2076 /* Nuke the page table entry. */
2077 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
2078 pteval = ptep_clear_flush(vma, address, pvmw.pte);
2079
2080 /* Move the dirty bit to the page. Now the pte is gone. */
2081 if (pte_dirty(pteval))
2082 set_page_dirty(page);
2083
2084 /*
2085 * Check that our target page is still mapped at the expected
2086 * address.
2087 */
2088 if (args->mm == mm && args->address == address &&
2089 pte_write(pteval))
2090 args->valid = true;
2091
2092 /*
2093 * Store the pfn of the page in a special migration
2094 * pte. do_swap_page() will wait until the migration
2095 * pte is removed and then restart fault handling.
2096 */
2097 if (pte_write(pteval))
2098 entry = make_writable_device_exclusive_entry(
2099 page_to_pfn(subpage));
2100 else
2101 entry = make_readable_device_exclusive_entry(
2102 page_to_pfn(subpage));
2103 swp_pte = swp_entry_to_pte(entry);
2104 if (pte_soft_dirty(pteval))
2105 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2106 if (pte_uffd_wp(pteval))
2107 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2108
2109 set_pte_at(mm, address, pvmw.pte, swp_pte);
2110
2111 /*
2112 * There is a reference on the page for the swap entry which has
2113 * been removed, so shouldn't take another.
2114 */
2115 page_remove_rmap(subpage, false);
2116 }
2117
2118 mmu_notifier_invalidate_range_end(&range);
2119
2120 return ret;
2121}
2122
2123/**
2124 * page_make_device_exclusive - mark the page exclusively owned by a device
2125 * @page: the page to replace page table entries for
2126 * @mm: the mm_struct where the page is expected to be mapped
2127 * @address: address where the page is expected to be mapped
2128 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2129 *
2130 * Tries to remove all the page table entries which are mapping this page and
2131 * replace them with special device exclusive swap entries to grant a device
2132 * exclusive access to the page. Caller must hold the page lock.
2133 *
2134 * Returns false if the page is still mapped, or if it could not be unmapped
2135 * from the expected address. Otherwise returns true (success).
2136 */
2137static bool page_make_device_exclusive(struct page *page, struct mm_struct *mm,
2138 unsigned long address, void *owner)
2139{
2140 struct make_exclusive_args args = {
2141 .mm = mm,
2142 .address = address,
2143 .owner = owner,
2144 .valid = false,
2145 };
2146 struct rmap_walk_control rwc = {
2147 .rmap_one = page_make_device_exclusive_one,
2148 .done = page_not_mapped,
2149 .anon_lock = page_lock_anon_vma_read,
2150 .arg = &args,
2151 };
2152
2153 /*
2154 * Restrict to anonymous pages for now to avoid potential writeback
2155 * issues. Also tail pages shouldn't be passed to rmap_walk so skip
2156 * those.
2157 */
2158 if (!PageAnon(page) || PageTail(page))
2159 return false;
2160
2161 rmap_walk(page, &rwc);
2162
2163 return args.valid && !page_mapcount(page);
2164}
2165
2166/**
2167 * make_device_exclusive_range() - Mark a range for exclusive use by a device
2168 * @mm: mm_struct of assoicated target process
2169 * @start: start of the region to mark for exclusive device access
2170 * @end: end address of region
2171 * @pages: returns the pages which were successfully marked for exclusive access
2172 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2173 *
2174 * Returns: number of pages found in the range by GUP. A page is marked for
2175 * exclusive access only if the page pointer is non-NULL.
2176 *
2177 * This function finds ptes mapping page(s) to the given address range, locks
2178 * them and replaces mappings with special swap entries preventing userspace CPU
2179 * access. On fault these entries are replaced with the original mapping after
2180 * calling MMU notifiers.
2181 *
2182 * A driver using this to program access from a device must use a mmu notifier
2183 * critical section to hold a device specific lock during programming. Once
2184 * programming is complete it should drop the page lock and reference after
2185 * which point CPU access to the page will revoke the exclusive access.
2186 */
2187int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
2188 unsigned long end, struct page **pages,
2189 void *owner)
2190{
2191 long npages = (end - start) >> PAGE_SHIFT;
2192 long i;
2193
2194 npages = get_user_pages_remote(mm, start, npages,
2195 FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
2196 pages, NULL, NULL);
2197 if (npages < 0)
2198 return npages;
2199
2200 for (i = 0; i < npages; i++, start += PAGE_SIZE) {
2201 if (!trylock_page(pages[i])) {
2202 put_page(pages[i]);
2203 pages[i] = NULL;
2204 continue;
2205 }
2206
2207 if (!page_make_device_exclusive(pages[i], mm, start, owner)) {
2208 unlock_page(pages[i]);
2209 put_page(pages[i]);
2210 pages[i] = NULL;
2211 }
2212 }
2213
2214 return npages;
2215}
2216EXPORT_SYMBOL_GPL(make_device_exclusive_range);
2217#endif
2218
2219void __put_anon_vma(struct anon_vma *anon_vma)
2220{
2221 struct anon_vma *root = anon_vma->root;
2222
2223 anon_vma_free(anon_vma);
2224 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
2225 anon_vma_free(root);
2226}
2227
2228static struct anon_vma *rmap_walk_anon_lock(struct page *page,
2229 struct rmap_walk_control *rwc)
2230{
2231 struct anon_vma *anon_vma;
2232
2233 if (rwc->anon_lock)
2234 return rwc->anon_lock(page);
2235
2236 /*
2237 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
2238 * because that depends on page_mapped(); but not all its usages
2239 * are holding mmap_lock. Users without mmap_lock are required to
2240 * take a reference count to prevent the anon_vma disappearing
2241 */
2242 anon_vma = page_anon_vma(page);
2243 if (!anon_vma)
2244 return NULL;
2245
2246 anon_vma_lock_read(anon_vma);
2247 return anon_vma;
2248}
2249
2250/*
2251 * rmap_walk_anon - do something to anonymous page using the object-based
2252 * rmap method
2253 * @page: the page to be handled
2254 * @rwc: control variable according to each walk type
2255 *
2256 * Find all the mappings of a page using the mapping pointer and the vma chains
2257 * contained in the anon_vma struct it points to.
2258 *
2259 * When called from page_mlock(), the mmap_lock of the mm containing the vma
2260 * where the page was found will be held for write. So, we won't recheck
2261 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
2262 * LOCKED.
2263 */
2264static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
2265 bool locked)
2266{
2267 struct anon_vma *anon_vma;
2268 pgoff_t pgoff_start, pgoff_end;
2269 struct anon_vma_chain *avc;
2270
2271 if (locked) {
2272 anon_vma = page_anon_vma(page);
2273 /* anon_vma disappear under us? */
2274 VM_BUG_ON_PAGE(!anon_vma, page);
2275 } else {
2276 anon_vma = rmap_walk_anon_lock(page, rwc);
2277 }
2278 if (!anon_vma)
2279 return;
2280
2281 pgoff_start = page_to_pgoff(page);
2282 pgoff_end = pgoff_start + thp_nr_pages(page) - 1;
2283 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
2284 pgoff_start, pgoff_end) {
2285 struct vm_area_struct *vma = avc->vma;
2286 unsigned long address = vma_address(page, vma);
2287
2288 VM_BUG_ON_VMA(address == -EFAULT, vma);
2289 cond_resched();
2290
2291 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2292 continue;
2293
2294 if (!rwc->rmap_one(page, vma, address, rwc->arg))
2295 break;
2296 if (rwc->done && rwc->done(page))
2297 break;
2298 }
2299
2300 if (!locked)
2301 anon_vma_unlock_read(anon_vma);
2302}
2303
2304/*
2305 * rmap_walk_file - do something to file page using the object-based rmap method
2306 * @page: the page to be handled
2307 * @rwc: control variable according to each walk type
2308 *
2309 * Find all the mappings of a page using the mapping pointer and the vma chains
2310 * contained in the address_space struct it points to.
2311 *
2312 * When called from page_mlock(), the mmap_lock of the mm containing the vma
2313 * where the page was found will be held for write. So, we won't recheck
2314 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
2315 * LOCKED.
2316 */
2317static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
2318 bool locked)
2319{
2320 struct address_space *mapping = page_mapping(page);
2321 pgoff_t pgoff_start, pgoff_end;
2322 struct vm_area_struct *vma;
2323
2324 /*
2325 * The page lock not only makes sure that page->mapping cannot
2326 * suddenly be NULLified by truncation, it makes sure that the
2327 * structure at mapping cannot be freed and reused yet,
2328 * so we can safely take mapping->i_mmap_rwsem.
2329 */
2330 VM_BUG_ON_PAGE(!PageLocked(page), page);
2331
2332 if (!mapping)
2333 return;
2334
2335 pgoff_start = page_to_pgoff(page);
2336 pgoff_end = pgoff_start + thp_nr_pages(page) - 1;
2337 if (!locked)
2338 i_mmap_lock_read(mapping);
2339 vma_interval_tree_foreach(vma, &mapping->i_mmap,
2340 pgoff_start, pgoff_end) {
2341 unsigned long address = vma_address(page, vma);
2342
2343 VM_BUG_ON_VMA(address == -EFAULT, vma);
2344 cond_resched();
2345
2346 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2347 continue;
2348
2349 if (!rwc->rmap_one(page, vma, address, rwc->arg))
2350 goto done;
2351 if (rwc->done && rwc->done(page))
2352 goto done;
2353 }
2354
2355done:
2356 if (!locked)
2357 i_mmap_unlock_read(mapping);
2358}
2359
2360void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
2361{
2362 if (unlikely(PageKsm(page)))
2363 rmap_walk_ksm(page, rwc);
2364 else if (PageAnon(page))
2365 rmap_walk_anon(page, rwc, false);
2366 else
2367 rmap_walk_file(page, rwc, false);
2368}
2369
2370/* Like rmap_walk, but caller holds relevant rmap lock */
2371void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
2372{
2373 /* no ksm support for now */
2374 VM_BUG_ON_PAGE(PageKsm(page), page);
2375 if (PageAnon(page))
2376 rmap_walk_anon(page, rwc, true);
2377 else
2378 rmap_walk_file(page, rwc, true);
2379}
2380
2381#ifdef CONFIG_HUGETLB_PAGE
2382/*
2383 * The following two functions are for anonymous (private mapped) hugepages.
2384 * Unlike common anonymous pages, anonymous hugepages have no accounting code
2385 * and no lru code, because we handle hugepages differently from common pages.
2386 */
2387void hugepage_add_anon_rmap(struct page *page,
2388 struct vm_area_struct *vma, unsigned long address)
2389{
2390 struct anon_vma *anon_vma = vma->anon_vma;
2391 int first;
2392
2393 BUG_ON(!PageLocked(page));
2394 BUG_ON(!anon_vma);
2395 /* address might be in next vma when migration races vma_adjust */
2396 first = atomic_inc_and_test(compound_mapcount_ptr(page));
2397 if (first)
2398 __page_set_anon_rmap(page, vma, address, 0);
2399}
2400
2401void hugepage_add_new_anon_rmap(struct page *page,
2402 struct vm_area_struct *vma, unsigned long address)
2403{
2404 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2405 atomic_set(compound_mapcount_ptr(page), 0);
2406 if (hpage_pincount_available(page))
2407 atomic_set(compound_pincount_ptr(page), 0);
2408
2409 __page_set_anon_rmap(page, vma, address, 1);
2410}
2411#endif /* CONFIG_HUGETLB_PAGE */
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/module.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
123/**
124 * anon_vma_prepare - attach an anon_vma to a memory region
125 * @vma: the memory region in question
126 *
127 * This makes sure the memory mapping described by 'vma' has
128 * an 'anon_vma' attached to it, so that we can associate the
129 * anonymous pages mapped into it with that anon_vma.
130 *
131 * The common case will be that we already have one, but if
132 * not we either need to find an adjacent mapping that we
133 * can re-use the anon_vma from (very common when the only
134 * reason for splitting a vma has been mprotect()), or we
135 * allocate a new one.
136 *
137 * Anon-vma allocations are very subtle, because we may have
138 * optimistically looked up an anon_vma in page_lock_anon_vma()
139 * and that may actually touch the spinlock even in the newly
140 * allocated vma (it depends on RCU to make sure that the
141 * anon_vma isn't actually destroyed).
142 *
143 * As a result, we need to do proper anon_vma locking even
144 * for the new allocation. At the same time, we do not want
145 * to do any locking for the common case of already having
146 * an anon_vma.
147 *
148 * This must be called with the mmap_sem held for reading.
149 */
150int anon_vma_prepare(struct vm_area_struct *vma)
151{
152 struct anon_vma *anon_vma = vma->anon_vma;
153 struct anon_vma_chain *avc;
154
155 might_sleep();
156 if (unlikely(!anon_vma)) {
157 struct mm_struct *mm = vma->vm_mm;
158 struct anon_vma *allocated;
159
160 avc = anon_vma_chain_alloc(GFP_KERNEL);
161 if (!avc)
162 goto out_enomem;
163
164 anon_vma = find_mergeable_anon_vma(vma);
165 allocated = NULL;
166 if (!anon_vma) {
167 anon_vma = anon_vma_alloc();
168 if (unlikely(!anon_vma))
169 goto out_enomem_free_avc;
170 allocated = anon_vma;
171 }
172
173 anon_vma_lock(anon_vma);
174 /* page_table_lock to protect against threads */
175 spin_lock(&mm->page_table_lock);
176 if (likely(!vma->anon_vma)) {
177 vma->anon_vma = anon_vma;
178 avc->anon_vma = anon_vma;
179 avc->vma = vma;
180 list_add(&avc->same_vma, &vma->anon_vma_chain);
181 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
182 allocated = NULL;
183 avc = NULL;
184 }
185 spin_unlock(&mm->page_table_lock);
186 anon_vma_unlock(anon_vma);
187
188 if (unlikely(allocated))
189 put_anon_vma(allocated);
190 if (unlikely(avc))
191 anon_vma_chain_free(avc);
192 }
193 return 0;
194
195 out_enomem_free_avc:
196 anon_vma_chain_free(avc);
197 out_enomem:
198 return -ENOMEM;
199}
200
201/*
202 * This is a useful helper function for locking the anon_vma root as
203 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
204 * have the same vma.
205 *
206 * Such anon_vma's should have the same root, so you'd expect to see
207 * just a single mutex_lock for the whole traversal.
208 */
209static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
210{
211 struct anon_vma *new_root = anon_vma->root;
212 if (new_root != root) {
213 if (WARN_ON_ONCE(root))
214 mutex_unlock(&root->mutex);
215 root = new_root;
216 mutex_lock(&root->mutex);
217 }
218 return root;
219}
220
221static inline void unlock_anon_vma_root(struct anon_vma *root)
222{
223 if (root)
224 mutex_unlock(&root->mutex);
225}
226
227static void anon_vma_chain_link(struct vm_area_struct *vma,
228 struct anon_vma_chain *avc,
229 struct anon_vma *anon_vma)
230{
231 avc->vma = vma;
232 avc->anon_vma = anon_vma;
233 list_add(&avc->same_vma, &vma->anon_vma_chain);
234
235 /*
236 * It's critical to add new vmas to the tail of the anon_vma,
237 * see comment in huge_memory.c:__split_huge_page().
238 */
239 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
240}
241
242/*
243 * Attach the anon_vmas from src to dst.
244 * Returns 0 on success, -ENOMEM on failure.
245 */
246int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
247{
248 struct anon_vma_chain *avc, *pavc;
249 struct anon_vma *root = NULL;
250
251 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
252 struct anon_vma *anon_vma;
253
254 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
255 if (unlikely(!avc)) {
256 unlock_anon_vma_root(root);
257 root = NULL;
258 avc = anon_vma_chain_alloc(GFP_KERNEL);
259 if (!avc)
260 goto enomem_failure;
261 }
262 anon_vma = pavc->anon_vma;
263 root = lock_anon_vma_root(root, anon_vma);
264 anon_vma_chain_link(dst, avc, anon_vma);
265 }
266 unlock_anon_vma_root(root);
267 return 0;
268
269 enomem_failure:
270 unlink_anon_vmas(dst);
271 return -ENOMEM;
272}
273
274/*
275 * Attach vma to its own anon_vma, as well as to the anon_vmas that
276 * the corresponding VMA in the parent process is attached to.
277 * Returns 0 on success, non-zero on failure.
278 */
279int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
280{
281 struct anon_vma_chain *avc;
282 struct anon_vma *anon_vma;
283
284 /* Don't bother if the parent process has no anon_vma here. */
285 if (!pvma->anon_vma)
286 return 0;
287
288 /*
289 * First, attach the new VMA to the parent VMA's anon_vmas,
290 * so rmap can find non-COWed pages in child processes.
291 */
292 if (anon_vma_clone(vma, pvma))
293 return -ENOMEM;
294
295 /* Then add our own anon_vma. */
296 anon_vma = anon_vma_alloc();
297 if (!anon_vma)
298 goto out_error;
299 avc = anon_vma_chain_alloc(GFP_KERNEL);
300 if (!avc)
301 goto out_error_free_anon_vma;
302
303 /*
304 * The root anon_vma's spinlock is the lock actually used when we
305 * lock any of the anon_vmas in this anon_vma tree.
306 */
307 anon_vma->root = pvma->anon_vma->root;
308 /*
309 * With refcounts, an anon_vma can stay around longer than the
310 * process it belongs to. The root anon_vma needs to be pinned until
311 * this anon_vma is freed, because the lock lives in the root.
312 */
313 get_anon_vma(anon_vma->root);
314 /* Mark this anon_vma as the one where our new (COWed) pages go. */
315 vma->anon_vma = anon_vma;
316 anon_vma_lock(anon_vma);
317 anon_vma_chain_link(vma, avc, anon_vma);
318 anon_vma_unlock(anon_vma);
319
320 return 0;
321
322 out_error_free_anon_vma:
323 put_anon_vma(anon_vma);
324 out_error:
325 unlink_anon_vmas(vma);
326 return -ENOMEM;
327}
328
329void unlink_anon_vmas(struct vm_area_struct *vma)
330{
331 struct anon_vma_chain *avc, *next;
332 struct anon_vma *root = NULL;
333
334 /*
335 * Unlink each anon_vma chained to the VMA. This list is ordered
336 * from newest to oldest, ensuring the root anon_vma gets freed last.
337 */
338 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
339 struct anon_vma *anon_vma = avc->anon_vma;
340
341 root = lock_anon_vma_root(root, anon_vma);
342 list_del(&avc->same_anon_vma);
343
344 /*
345 * Leave empty anon_vmas on the list - we'll need
346 * to free them outside the lock.
347 */
348 if (list_empty(&anon_vma->head))
349 continue;
350
351 list_del(&avc->same_vma);
352 anon_vma_chain_free(avc);
353 }
354 unlock_anon_vma_root(root);
355
356 /*
357 * Iterate the list once more, it now only contains empty and unlinked
358 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
359 * needing to acquire the anon_vma->root->mutex.
360 */
361 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
362 struct anon_vma *anon_vma = avc->anon_vma;
363
364 put_anon_vma(anon_vma);
365
366 list_del(&avc->same_vma);
367 anon_vma_chain_free(avc);
368 }
369}
370
371static void anon_vma_ctor(void *data)
372{
373 struct anon_vma *anon_vma = data;
374
375 mutex_init(&anon_vma->mutex);
376 atomic_set(&anon_vma->refcount, 0);
377 INIT_LIST_HEAD(&anon_vma->head);
378}
379
380void __init anon_vma_init(void)
381{
382 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
383 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
384 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
385}
386
387/*
388 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
389 *
390 * Since there is no serialization what so ever against page_remove_rmap()
391 * the best this function can do is return a locked anon_vma that might
392 * have been relevant to this page.
393 *
394 * The page might have been remapped to a different anon_vma or the anon_vma
395 * returned may already be freed (and even reused).
396 *
397 * In case it was remapped to a different anon_vma, the new anon_vma will be a
398 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
399 * ensure that any anon_vma obtained from the page will still be valid for as
400 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
401 *
402 * All users of this function must be very careful when walking the anon_vma
403 * chain and verify that the page in question is indeed mapped in it
404 * [ something equivalent to page_mapped_in_vma() ].
405 *
406 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
407 * that the anon_vma pointer from page->mapping is valid if there is a
408 * mapcount, we can dereference the anon_vma after observing those.
409 */
410struct anon_vma *page_get_anon_vma(struct page *page)
411{
412 struct anon_vma *anon_vma = NULL;
413 unsigned long anon_mapping;
414
415 rcu_read_lock();
416 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
417 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
418 goto out;
419 if (!page_mapped(page))
420 goto out;
421
422 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
423 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
424 anon_vma = NULL;
425 goto out;
426 }
427
428 /*
429 * If this page is still mapped, then its anon_vma cannot have been
430 * freed. But if it has been unmapped, we have no security against the
431 * anon_vma structure being freed and reused (for another anon_vma:
432 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
433 * above cannot corrupt).
434 */
435 if (!page_mapped(page)) {
436 put_anon_vma(anon_vma);
437 anon_vma = NULL;
438 }
439out:
440 rcu_read_unlock();
441
442 return anon_vma;
443}
444
445/*
446 * Similar to page_get_anon_vma() except it locks the anon_vma.
447 *
448 * Its a little more complex as it tries to keep the fast path to a single
449 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
450 * reference like with page_get_anon_vma() and then block on the mutex.
451 */
452struct anon_vma *page_lock_anon_vma(struct page *page)
453{
454 struct anon_vma *anon_vma = NULL;
455 struct anon_vma *root_anon_vma;
456 unsigned long anon_mapping;
457
458 rcu_read_lock();
459 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
460 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
461 goto out;
462 if (!page_mapped(page))
463 goto out;
464
465 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
466 root_anon_vma = ACCESS_ONCE(anon_vma->root);
467 if (mutex_trylock(&root_anon_vma->mutex)) {
468 /*
469 * If the page is still mapped, then this anon_vma is still
470 * its anon_vma, and holding the mutex ensures that it will
471 * not go away, see anon_vma_free().
472 */
473 if (!page_mapped(page)) {
474 mutex_unlock(&root_anon_vma->mutex);
475 anon_vma = NULL;
476 }
477 goto out;
478 }
479
480 /* trylock failed, we got to sleep */
481 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
482 anon_vma = NULL;
483 goto out;
484 }
485
486 if (!page_mapped(page)) {
487 put_anon_vma(anon_vma);
488 anon_vma = NULL;
489 goto out;
490 }
491
492 /* we pinned the anon_vma, its safe to sleep */
493 rcu_read_unlock();
494 anon_vma_lock(anon_vma);
495
496 if (atomic_dec_and_test(&anon_vma->refcount)) {
497 /*
498 * Oops, we held the last refcount, release the lock
499 * and bail -- can't simply use put_anon_vma() because
500 * we'll deadlock on the anon_vma_lock() recursion.
501 */
502 anon_vma_unlock(anon_vma);
503 __put_anon_vma(anon_vma);
504 anon_vma = NULL;
505 }
506
507 return anon_vma;
508
509out:
510 rcu_read_unlock();
511 return anon_vma;
512}
513
514void page_unlock_anon_vma(struct anon_vma *anon_vma)
515{
516 anon_vma_unlock(anon_vma);
517}
518
519/*
520 * At what user virtual address is page expected in @vma?
521 * Returns virtual address or -EFAULT if page's index/offset is not
522 * within the range mapped the @vma.
523 */
524inline unsigned long
525vma_address(struct page *page, struct vm_area_struct *vma)
526{
527 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
528 unsigned long address;
529
530 if (unlikely(is_vm_hugetlb_page(vma)))
531 pgoff = page->index << huge_page_order(page_hstate(page));
532 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
533 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
534 /* page should be within @vma mapping range */
535 return -EFAULT;
536 }
537 return address;
538}
539
540/*
541 * At what user virtual address is page expected in vma?
542 * Caller should check the page is actually part of the vma.
543 */
544unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
545{
546 if (PageAnon(page)) {
547 struct anon_vma *page__anon_vma = page_anon_vma(page);
548 /*
549 * Note: swapoff's unuse_vma() is more efficient with this
550 * check, and needs it to match anon_vma when KSM is active.
551 */
552 if (!vma->anon_vma || !page__anon_vma ||
553 vma->anon_vma->root != page__anon_vma->root)
554 return -EFAULT;
555 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
556 if (!vma->vm_file ||
557 vma->vm_file->f_mapping != page->mapping)
558 return -EFAULT;
559 } else
560 return -EFAULT;
561 return vma_address(page, vma);
562}
563
564/*
565 * Check that @page is mapped at @address into @mm.
566 *
567 * If @sync is false, page_check_address may perform a racy check to avoid
568 * the page table lock when the pte is not present (helpful when reclaiming
569 * highly shared pages).
570 *
571 * On success returns with pte mapped and locked.
572 */
573pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
574 unsigned long address, spinlock_t **ptlp, int sync)
575{
576 pgd_t *pgd;
577 pud_t *pud;
578 pmd_t *pmd;
579 pte_t *pte;
580 spinlock_t *ptl;
581
582 if (unlikely(PageHuge(page))) {
583 pte = huge_pte_offset(mm, address);
584 ptl = &mm->page_table_lock;
585 goto check;
586 }
587
588 pgd = pgd_offset(mm, address);
589 if (!pgd_present(*pgd))
590 return NULL;
591
592 pud = pud_offset(pgd, address);
593 if (!pud_present(*pud))
594 return NULL;
595
596 pmd = pmd_offset(pud, address);
597 if (!pmd_present(*pmd))
598 return NULL;
599 if (pmd_trans_huge(*pmd))
600 return NULL;
601
602 pte = pte_offset_map(pmd, address);
603 /* Make a quick check before getting the lock */
604 if (!sync && !pte_present(*pte)) {
605 pte_unmap(pte);
606 return NULL;
607 }
608
609 ptl = pte_lockptr(mm, pmd);
610check:
611 spin_lock(ptl);
612 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
613 *ptlp = ptl;
614 return pte;
615 }
616 pte_unmap_unlock(pte, ptl);
617 return NULL;
618}
619
620/**
621 * page_mapped_in_vma - check whether a page is really mapped in a VMA
622 * @page: the page to test
623 * @vma: the VMA to test
624 *
625 * Returns 1 if the page is mapped into the page tables of the VMA, 0
626 * if the page is not mapped into the page tables of this VMA. Only
627 * valid for normal file or anonymous VMAs.
628 */
629int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
630{
631 unsigned long address;
632 pte_t *pte;
633 spinlock_t *ptl;
634
635 address = vma_address(page, vma);
636 if (address == -EFAULT) /* out of vma range */
637 return 0;
638 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
639 if (!pte) /* the page is not in this mm */
640 return 0;
641 pte_unmap_unlock(pte, ptl);
642
643 return 1;
644}
645
646/*
647 * Subfunctions of page_referenced: page_referenced_one called
648 * repeatedly from either page_referenced_anon or page_referenced_file.
649 */
650int page_referenced_one(struct page *page, struct vm_area_struct *vma,
651 unsigned long address, unsigned int *mapcount,
652 unsigned long *vm_flags)
653{
654 struct mm_struct *mm = vma->vm_mm;
655 int referenced = 0;
656
657 if (unlikely(PageTransHuge(page))) {
658 pmd_t *pmd;
659
660 spin_lock(&mm->page_table_lock);
661 /*
662 * rmap might return false positives; we must filter
663 * these out using page_check_address_pmd().
664 */
665 pmd = page_check_address_pmd(page, mm, address,
666 PAGE_CHECK_ADDRESS_PMD_FLAG);
667 if (!pmd) {
668 spin_unlock(&mm->page_table_lock);
669 goto out;
670 }
671
672 if (vma->vm_flags & VM_LOCKED) {
673 spin_unlock(&mm->page_table_lock);
674 *mapcount = 0; /* break early from loop */
675 *vm_flags |= VM_LOCKED;
676 goto out;
677 }
678
679 /* go ahead even if the pmd is pmd_trans_splitting() */
680 if (pmdp_clear_flush_young_notify(vma, address, pmd))
681 referenced++;
682 spin_unlock(&mm->page_table_lock);
683 } else {
684 pte_t *pte;
685 spinlock_t *ptl;
686
687 /*
688 * rmap might return false positives; we must filter
689 * these out using page_check_address().
690 */
691 pte = page_check_address(page, mm, address, &ptl, 0);
692 if (!pte)
693 goto out;
694
695 if (vma->vm_flags & VM_LOCKED) {
696 pte_unmap_unlock(pte, ptl);
697 *mapcount = 0; /* break early from loop */
698 *vm_flags |= VM_LOCKED;
699 goto out;
700 }
701
702 if (ptep_clear_flush_young_notify(vma, address, pte)) {
703 /*
704 * Don't treat a reference through a sequentially read
705 * mapping as such. If the page has been used in
706 * another mapping, we will catch it; if this other
707 * mapping is already gone, the unmap path will have
708 * set PG_referenced or activated the page.
709 */
710 if (likely(!VM_SequentialReadHint(vma)))
711 referenced++;
712 }
713 pte_unmap_unlock(pte, ptl);
714 }
715
716 /* Pretend the page is referenced if the task has the
717 swap token and is in the middle of a page fault. */
718 if (mm != current->mm && has_swap_token(mm) &&
719 rwsem_is_locked(&mm->mmap_sem))
720 referenced++;
721
722 (*mapcount)--;
723
724 if (referenced)
725 *vm_flags |= vma->vm_flags;
726out:
727 return referenced;
728}
729
730static int page_referenced_anon(struct page *page,
731 struct mem_cgroup *mem_cont,
732 unsigned long *vm_flags)
733{
734 unsigned int mapcount;
735 struct anon_vma *anon_vma;
736 struct anon_vma_chain *avc;
737 int referenced = 0;
738
739 anon_vma = page_lock_anon_vma(page);
740 if (!anon_vma)
741 return referenced;
742
743 mapcount = page_mapcount(page);
744 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
745 struct vm_area_struct *vma = avc->vma;
746 unsigned long address = vma_address(page, vma);
747 if (address == -EFAULT)
748 continue;
749 /*
750 * If we are reclaiming on behalf of a cgroup, skip
751 * counting on behalf of references from different
752 * cgroups
753 */
754 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
755 continue;
756 referenced += page_referenced_one(page, vma, address,
757 &mapcount, vm_flags);
758 if (!mapcount)
759 break;
760 }
761
762 page_unlock_anon_vma(anon_vma);
763 return referenced;
764}
765
766/**
767 * page_referenced_file - referenced check for object-based rmap
768 * @page: the page we're checking references on.
769 * @mem_cont: target memory controller
770 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
771 *
772 * For an object-based mapped page, find all the places it is mapped and
773 * check/clear the referenced flag. This is done by following the page->mapping
774 * pointer, then walking the chain of vmas it holds. It returns the number
775 * of references it found.
776 *
777 * This function is only called from page_referenced for object-based pages.
778 */
779static int page_referenced_file(struct page *page,
780 struct mem_cgroup *mem_cont,
781 unsigned long *vm_flags)
782{
783 unsigned int mapcount;
784 struct address_space *mapping = page->mapping;
785 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
786 struct vm_area_struct *vma;
787 struct prio_tree_iter iter;
788 int referenced = 0;
789
790 /*
791 * The caller's checks on page->mapping and !PageAnon have made
792 * sure that this is a file page: the check for page->mapping
793 * excludes the case just before it gets set on an anon page.
794 */
795 BUG_ON(PageAnon(page));
796
797 /*
798 * The page lock not only makes sure that page->mapping cannot
799 * suddenly be NULLified by truncation, it makes sure that the
800 * structure at mapping cannot be freed and reused yet,
801 * so we can safely take mapping->i_mmap_mutex.
802 */
803 BUG_ON(!PageLocked(page));
804
805 mutex_lock(&mapping->i_mmap_mutex);
806
807 /*
808 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
809 * is more likely to be accurate if we note it after spinning.
810 */
811 mapcount = page_mapcount(page);
812
813 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
814 unsigned long address = vma_address(page, vma);
815 if (address == -EFAULT)
816 continue;
817 /*
818 * If we are reclaiming on behalf of a cgroup, skip
819 * counting on behalf of references from different
820 * cgroups
821 */
822 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
823 continue;
824 referenced += page_referenced_one(page, vma, address,
825 &mapcount, vm_flags);
826 if (!mapcount)
827 break;
828 }
829
830 mutex_unlock(&mapping->i_mmap_mutex);
831 return referenced;
832}
833
834/**
835 * page_referenced - test if the page was referenced
836 * @page: the page to test
837 * @is_locked: caller holds lock on the page
838 * @mem_cont: target memory controller
839 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
840 *
841 * Quick test_and_clear_referenced for all mappings to a page,
842 * returns the number of ptes which referenced the page.
843 */
844int page_referenced(struct page *page,
845 int is_locked,
846 struct mem_cgroup *mem_cont,
847 unsigned long *vm_flags)
848{
849 int referenced = 0;
850 int we_locked = 0;
851
852 *vm_flags = 0;
853 if (page_mapped(page) && page_rmapping(page)) {
854 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
855 we_locked = trylock_page(page);
856 if (!we_locked) {
857 referenced++;
858 goto out;
859 }
860 }
861 if (unlikely(PageKsm(page)))
862 referenced += page_referenced_ksm(page, mem_cont,
863 vm_flags);
864 else if (PageAnon(page))
865 referenced += page_referenced_anon(page, mem_cont,
866 vm_flags);
867 else if (page->mapping)
868 referenced += page_referenced_file(page, mem_cont,
869 vm_flags);
870 if (we_locked)
871 unlock_page(page);
872
873 if (page_test_and_clear_young(page_to_pfn(page)))
874 referenced++;
875 }
876out:
877 return referenced;
878}
879
880static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
881 unsigned long address)
882{
883 struct mm_struct *mm = vma->vm_mm;
884 pte_t *pte;
885 spinlock_t *ptl;
886 int ret = 0;
887
888 pte = page_check_address(page, mm, address, &ptl, 1);
889 if (!pte)
890 goto out;
891
892 if (pte_dirty(*pte) || pte_write(*pte)) {
893 pte_t entry;
894
895 flush_cache_page(vma, address, pte_pfn(*pte));
896 entry = ptep_clear_flush_notify(vma, address, pte);
897 entry = pte_wrprotect(entry);
898 entry = pte_mkclean(entry);
899 set_pte_at(mm, address, pte, entry);
900 ret = 1;
901 }
902
903 pte_unmap_unlock(pte, ptl);
904out:
905 return ret;
906}
907
908static int page_mkclean_file(struct address_space *mapping, struct page *page)
909{
910 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
911 struct vm_area_struct *vma;
912 struct prio_tree_iter iter;
913 int ret = 0;
914
915 BUG_ON(PageAnon(page));
916
917 mutex_lock(&mapping->i_mmap_mutex);
918 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
919 if (vma->vm_flags & VM_SHARED) {
920 unsigned long address = vma_address(page, vma);
921 if (address == -EFAULT)
922 continue;
923 ret += page_mkclean_one(page, vma, address);
924 }
925 }
926 mutex_unlock(&mapping->i_mmap_mutex);
927 return ret;
928}
929
930int page_mkclean(struct page *page)
931{
932 int ret = 0;
933
934 BUG_ON(!PageLocked(page));
935
936 if (page_mapped(page)) {
937 struct address_space *mapping = page_mapping(page);
938 if (mapping) {
939 ret = page_mkclean_file(mapping, page);
940 if (page_test_and_clear_dirty(page_to_pfn(page), 1))
941 ret = 1;
942 }
943 }
944
945 return ret;
946}
947EXPORT_SYMBOL_GPL(page_mkclean);
948
949/**
950 * page_move_anon_rmap - move a page to our anon_vma
951 * @page: the page to move to our anon_vma
952 * @vma: the vma the page belongs to
953 * @address: the user virtual address mapped
954 *
955 * When a page belongs exclusively to one process after a COW event,
956 * that page can be moved into the anon_vma that belongs to just that
957 * process, so the rmap code will not search the parent or sibling
958 * processes.
959 */
960void page_move_anon_rmap(struct page *page,
961 struct vm_area_struct *vma, unsigned long address)
962{
963 struct anon_vma *anon_vma = vma->anon_vma;
964
965 VM_BUG_ON(!PageLocked(page));
966 VM_BUG_ON(!anon_vma);
967 VM_BUG_ON(page->index != linear_page_index(vma, address));
968
969 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
970 page->mapping = (struct address_space *) anon_vma;
971}
972
973/**
974 * __page_set_anon_rmap - set up new anonymous rmap
975 * @page: Page to add to rmap
976 * @vma: VM area to add page to.
977 * @address: User virtual address of the mapping
978 * @exclusive: the page is exclusively owned by the current process
979 */
980static void __page_set_anon_rmap(struct page *page,
981 struct vm_area_struct *vma, unsigned long address, int exclusive)
982{
983 struct anon_vma *anon_vma = vma->anon_vma;
984
985 BUG_ON(!anon_vma);
986
987 if (PageAnon(page))
988 return;
989
990 /*
991 * If the page isn't exclusively mapped into this vma,
992 * we must use the _oldest_ possible anon_vma for the
993 * page mapping!
994 */
995 if (!exclusive)
996 anon_vma = anon_vma->root;
997
998 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
999 page->mapping = (struct address_space *) anon_vma;
1000 page->index = linear_page_index(vma, address);
1001}
1002
1003/**
1004 * __page_check_anon_rmap - sanity check anonymous rmap addition
1005 * @page: the page to add the mapping to
1006 * @vma: the vm area in which the mapping is added
1007 * @address: the user virtual address mapped
1008 */
1009static void __page_check_anon_rmap(struct page *page,
1010 struct vm_area_struct *vma, unsigned long address)
1011{
1012#ifdef CONFIG_DEBUG_VM
1013 /*
1014 * The page's anon-rmap details (mapping and index) are guaranteed to
1015 * be set up correctly at this point.
1016 *
1017 * We have exclusion against page_add_anon_rmap because the caller
1018 * always holds the page locked, except if called from page_dup_rmap,
1019 * in which case the page is already known to be setup.
1020 *
1021 * We have exclusion against page_add_new_anon_rmap because those pages
1022 * are initially only visible via the pagetables, and the pte is locked
1023 * over the call to page_add_new_anon_rmap.
1024 */
1025 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1026 BUG_ON(page->index != linear_page_index(vma, address));
1027#endif
1028}
1029
1030/**
1031 * page_add_anon_rmap - add pte mapping to an anonymous page
1032 * @page: the page to add the mapping to
1033 * @vma: the vm area in which the mapping is added
1034 * @address: the user virtual address mapped
1035 *
1036 * The caller needs to hold the pte lock, and the page must be locked in
1037 * the anon_vma case: to serialize mapping,index checking after setting,
1038 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1039 * (but PageKsm is never downgraded to PageAnon).
1040 */
1041void page_add_anon_rmap(struct page *page,
1042 struct vm_area_struct *vma, unsigned long address)
1043{
1044 do_page_add_anon_rmap(page, vma, address, 0);
1045}
1046
1047/*
1048 * Special version of the above for do_swap_page, which often runs
1049 * into pages that are exclusively owned by the current process.
1050 * Everybody else should continue to use page_add_anon_rmap above.
1051 */
1052void do_page_add_anon_rmap(struct page *page,
1053 struct vm_area_struct *vma, unsigned long address, int exclusive)
1054{
1055 int first = atomic_inc_and_test(&page->_mapcount);
1056 if (first) {
1057 if (!PageTransHuge(page))
1058 __inc_zone_page_state(page, NR_ANON_PAGES);
1059 else
1060 __inc_zone_page_state(page,
1061 NR_ANON_TRANSPARENT_HUGEPAGES);
1062 }
1063 if (unlikely(PageKsm(page)))
1064 return;
1065
1066 VM_BUG_ON(!PageLocked(page));
1067 /* address might be in next vma when migration races vma_adjust */
1068 if (first)
1069 __page_set_anon_rmap(page, vma, address, exclusive);
1070 else
1071 __page_check_anon_rmap(page, vma, address);
1072}
1073
1074/**
1075 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1076 * @page: the page to add the mapping to
1077 * @vma: the vm area in which the mapping is added
1078 * @address: the user virtual address mapped
1079 *
1080 * Same as page_add_anon_rmap but must only be called on *new* pages.
1081 * This means the inc-and-test can be bypassed.
1082 * Page does not have to be locked.
1083 */
1084void page_add_new_anon_rmap(struct page *page,
1085 struct vm_area_struct *vma, unsigned long address)
1086{
1087 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1088 SetPageSwapBacked(page);
1089 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1090 if (!PageTransHuge(page))
1091 __inc_zone_page_state(page, NR_ANON_PAGES);
1092 else
1093 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1094 __page_set_anon_rmap(page, vma, address, 1);
1095 if (page_evictable(page, vma))
1096 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
1097 else
1098 add_page_to_unevictable_list(page);
1099}
1100
1101/**
1102 * page_add_file_rmap - add pte mapping to a file page
1103 * @page: the page to add the mapping to
1104 *
1105 * The caller needs to hold the pte lock.
1106 */
1107void page_add_file_rmap(struct page *page)
1108{
1109 if (atomic_inc_and_test(&page->_mapcount)) {
1110 __inc_zone_page_state(page, NR_FILE_MAPPED);
1111 mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
1112 }
1113}
1114
1115/**
1116 * page_remove_rmap - take down pte mapping from a page
1117 * @page: page to remove mapping from
1118 *
1119 * The caller needs to hold the pte lock.
1120 */
1121void page_remove_rmap(struct page *page)
1122{
1123 /* page still mapped by someone else? */
1124 if (!atomic_add_negative(-1, &page->_mapcount))
1125 return;
1126
1127 /*
1128 * Now that the last pte has gone, s390 must transfer dirty
1129 * flag from storage key to struct page. We can usually skip
1130 * this if the page is anon, so about to be freed; but perhaps
1131 * not if it's in swapcache - there might be another pte slot
1132 * containing the swap entry, but page not yet written to swap.
1133 */
1134 if ((!PageAnon(page) || PageSwapCache(page)) &&
1135 page_test_and_clear_dirty(page_to_pfn(page), 1))
1136 set_page_dirty(page);
1137 /*
1138 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1139 * and not charged by memcg for now.
1140 */
1141 if (unlikely(PageHuge(page)))
1142 return;
1143 if (PageAnon(page)) {
1144 mem_cgroup_uncharge_page(page);
1145 if (!PageTransHuge(page))
1146 __dec_zone_page_state(page, NR_ANON_PAGES);
1147 else
1148 __dec_zone_page_state(page,
1149 NR_ANON_TRANSPARENT_HUGEPAGES);
1150 } else {
1151 __dec_zone_page_state(page, NR_FILE_MAPPED);
1152 mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
1153 }
1154 /*
1155 * It would be tidy to reset the PageAnon mapping here,
1156 * but that might overwrite a racing page_add_anon_rmap
1157 * which increments mapcount after us but sets mapping
1158 * before us: so leave the reset to free_hot_cold_page,
1159 * and remember that it's only reliable while mapped.
1160 * Leaving it set also helps swapoff to reinstate ptes
1161 * faster for those pages still in swapcache.
1162 */
1163}
1164
1165/*
1166 * Subfunctions of try_to_unmap: try_to_unmap_one called
1167 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
1168 */
1169int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1170 unsigned long address, enum ttu_flags flags)
1171{
1172 struct mm_struct *mm = vma->vm_mm;
1173 pte_t *pte;
1174 pte_t pteval;
1175 spinlock_t *ptl;
1176 int ret = SWAP_AGAIN;
1177
1178 pte = page_check_address(page, mm, address, &ptl, 0);
1179 if (!pte)
1180 goto out;
1181
1182 /*
1183 * If the page is mlock()d, we cannot swap it out.
1184 * If it's recently referenced (perhaps page_referenced
1185 * skipped over this mm) then we should reactivate it.
1186 */
1187 if (!(flags & TTU_IGNORE_MLOCK)) {
1188 if (vma->vm_flags & VM_LOCKED)
1189 goto out_mlock;
1190
1191 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1192 goto out_unmap;
1193 }
1194 if (!(flags & TTU_IGNORE_ACCESS)) {
1195 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1196 ret = SWAP_FAIL;
1197 goto out_unmap;
1198 }
1199 }
1200
1201 /* Nuke the page table entry. */
1202 flush_cache_page(vma, address, page_to_pfn(page));
1203 pteval = ptep_clear_flush_notify(vma, address, pte);
1204
1205 /* Move the dirty bit to the physical page now the pte is gone. */
1206 if (pte_dirty(pteval))
1207 set_page_dirty(page);
1208
1209 /* Update high watermark before we lower rss */
1210 update_hiwater_rss(mm);
1211
1212 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1213 if (PageAnon(page))
1214 dec_mm_counter(mm, MM_ANONPAGES);
1215 else
1216 dec_mm_counter(mm, MM_FILEPAGES);
1217 set_pte_at(mm, address, pte,
1218 swp_entry_to_pte(make_hwpoison_entry(page)));
1219 } else if (PageAnon(page)) {
1220 swp_entry_t entry = { .val = page_private(page) };
1221
1222 if (PageSwapCache(page)) {
1223 /*
1224 * Store the swap location in the pte.
1225 * See handle_pte_fault() ...
1226 */
1227 if (swap_duplicate(entry) < 0) {
1228 set_pte_at(mm, address, pte, pteval);
1229 ret = SWAP_FAIL;
1230 goto out_unmap;
1231 }
1232 if (list_empty(&mm->mmlist)) {
1233 spin_lock(&mmlist_lock);
1234 if (list_empty(&mm->mmlist))
1235 list_add(&mm->mmlist, &init_mm.mmlist);
1236 spin_unlock(&mmlist_lock);
1237 }
1238 dec_mm_counter(mm, MM_ANONPAGES);
1239 inc_mm_counter(mm, MM_SWAPENTS);
1240 } else if (PAGE_MIGRATION) {
1241 /*
1242 * Store the pfn of the page in a special migration
1243 * pte. do_swap_page() will wait until the migration
1244 * pte is removed and then restart fault handling.
1245 */
1246 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1247 entry = make_migration_entry(page, pte_write(pteval));
1248 }
1249 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1250 BUG_ON(pte_file(*pte));
1251 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
1252 /* Establish migration entry for a file page */
1253 swp_entry_t entry;
1254 entry = make_migration_entry(page, pte_write(pteval));
1255 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1256 } else
1257 dec_mm_counter(mm, MM_FILEPAGES);
1258
1259 page_remove_rmap(page);
1260 page_cache_release(page);
1261
1262out_unmap:
1263 pte_unmap_unlock(pte, ptl);
1264out:
1265 return ret;
1266
1267out_mlock:
1268 pte_unmap_unlock(pte, ptl);
1269
1270
1271 /*
1272 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1273 * unstable result and race. Plus, We can't wait here because
1274 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1275 * if trylock failed, the page remain in evictable lru and later
1276 * vmscan could retry to move the page to unevictable lru if the
1277 * page is actually mlocked.
1278 */
1279 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1280 if (vma->vm_flags & VM_LOCKED) {
1281 mlock_vma_page(page);
1282 ret = SWAP_MLOCK;
1283 }
1284 up_read(&vma->vm_mm->mmap_sem);
1285 }
1286 return ret;
1287}
1288
1289/*
1290 * objrmap doesn't work for nonlinear VMAs because the assumption that
1291 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1292 * Consequently, given a particular page and its ->index, we cannot locate the
1293 * ptes which are mapping that page without an exhaustive linear search.
1294 *
1295 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1296 * maps the file to which the target page belongs. The ->vm_private_data field
1297 * holds the current cursor into that scan. Successive searches will circulate
1298 * around the vma's virtual address space.
1299 *
1300 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1301 * more scanning pressure is placed against them as well. Eventually pages
1302 * will become fully unmapped and are eligible for eviction.
1303 *
1304 * For very sparsely populated VMAs this is a little inefficient - chances are
1305 * there there won't be many ptes located within the scan cluster. In this case
1306 * maybe we could scan further - to the end of the pte page, perhaps.
1307 *
1308 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1309 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1310 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1311 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1312 */
1313#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1314#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1315
1316static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1317 struct vm_area_struct *vma, struct page *check_page)
1318{
1319 struct mm_struct *mm = vma->vm_mm;
1320 pgd_t *pgd;
1321 pud_t *pud;
1322 pmd_t *pmd;
1323 pte_t *pte;
1324 pte_t pteval;
1325 spinlock_t *ptl;
1326 struct page *page;
1327 unsigned long address;
1328 unsigned long end;
1329 int ret = SWAP_AGAIN;
1330 int locked_vma = 0;
1331
1332 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1333 end = address + CLUSTER_SIZE;
1334 if (address < vma->vm_start)
1335 address = vma->vm_start;
1336 if (end > vma->vm_end)
1337 end = vma->vm_end;
1338
1339 pgd = pgd_offset(mm, address);
1340 if (!pgd_present(*pgd))
1341 return ret;
1342
1343 pud = pud_offset(pgd, address);
1344 if (!pud_present(*pud))
1345 return ret;
1346
1347 pmd = pmd_offset(pud, address);
1348 if (!pmd_present(*pmd))
1349 return ret;
1350
1351 /*
1352 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1353 * keep the sem while scanning the cluster for mlocking pages.
1354 */
1355 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1356 locked_vma = (vma->vm_flags & VM_LOCKED);
1357 if (!locked_vma)
1358 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1359 }
1360
1361 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1362
1363 /* Update high watermark before we lower rss */
1364 update_hiwater_rss(mm);
1365
1366 for (; address < end; pte++, address += PAGE_SIZE) {
1367 if (!pte_present(*pte))
1368 continue;
1369 page = vm_normal_page(vma, address, *pte);
1370 BUG_ON(!page || PageAnon(page));
1371
1372 if (locked_vma) {
1373 mlock_vma_page(page); /* no-op if already mlocked */
1374 if (page == check_page)
1375 ret = SWAP_MLOCK;
1376 continue; /* don't unmap */
1377 }
1378
1379 if (ptep_clear_flush_young_notify(vma, address, pte))
1380 continue;
1381
1382 /* Nuke the page table entry. */
1383 flush_cache_page(vma, address, pte_pfn(*pte));
1384 pteval = ptep_clear_flush_notify(vma, address, pte);
1385
1386 /* If nonlinear, store the file page offset in the pte. */
1387 if (page->index != linear_page_index(vma, address))
1388 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1389
1390 /* Move the dirty bit to the physical page now the pte is gone. */
1391 if (pte_dirty(pteval))
1392 set_page_dirty(page);
1393
1394 page_remove_rmap(page);
1395 page_cache_release(page);
1396 dec_mm_counter(mm, MM_FILEPAGES);
1397 (*mapcount)--;
1398 }
1399 pte_unmap_unlock(pte - 1, ptl);
1400 if (locked_vma)
1401 up_read(&vma->vm_mm->mmap_sem);
1402 return ret;
1403}
1404
1405bool is_vma_temporary_stack(struct vm_area_struct *vma)
1406{
1407 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1408
1409 if (!maybe_stack)
1410 return false;
1411
1412 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1413 VM_STACK_INCOMPLETE_SETUP)
1414 return true;
1415
1416 return false;
1417}
1418
1419/**
1420 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1421 * rmap method
1422 * @page: the page to unmap/unlock
1423 * @flags: action and flags
1424 *
1425 * Find all the mappings of a page using the mapping pointer and the vma chains
1426 * contained in the anon_vma struct it points to.
1427 *
1428 * This function is only called from try_to_unmap/try_to_munlock for
1429 * anonymous pages.
1430 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1431 * where the page was found will be held for write. So, we won't recheck
1432 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1433 * 'LOCKED.
1434 */
1435static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1436{
1437 struct anon_vma *anon_vma;
1438 struct anon_vma_chain *avc;
1439 int ret = SWAP_AGAIN;
1440
1441 anon_vma = page_lock_anon_vma(page);
1442 if (!anon_vma)
1443 return ret;
1444
1445 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1446 struct vm_area_struct *vma = avc->vma;
1447 unsigned long address;
1448
1449 /*
1450 * During exec, a temporary VMA is setup and later moved.
1451 * The VMA is moved under the anon_vma lock but not the
1452 * page tables leading to a race where migration cannot
1453 * find the migration ptes. Rather than increasing the
1454 * locking requirements of exec(), migration skips
1455 * temporary VMAs until after exec() completes.
1456 */
1457 if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1458 is_vma_temporary_stack(vma))
1459 continue;
1460
1461 address = vma_address(page, vma);
1462 if (address == -EFAULT)
1463 continue;
1464 ret = try_to_unmap_one(page, vma, address, flags);
1465 if (ret != SWAP_AGAIN || !page_mapped(page))
1466 break;
1467 }
1468
1469 page_unlock_anon_vma(anon_vma);
1470 return ret;
1471}
1472
1473/**
1474 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1475 * @page: the page to unmap/unlock
1476 * @flags: action and flags
1477 *
1478 * Find all the mappings of a page using the mapping pointer and the vma chains
1479 * contained in the address_space struct it points to.
1480 *
1481 * This function is only called from try_to_unmap/try_to_munlock for
1482 * object-based pages.
1483 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1484 * where the page was found will be held for write. So, we won't recheck
1485 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1486 * 'LOCKED.
1487 */
1488static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1489{
1490 struct address_space *mapping = page->mapping;
1491 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1492 struct vm_area_struct *vma;
1493 struct prio_tree_iter iter;
1494 int ret = SWAP_AGAIN;
1495 unsigned long cursor;
1496 unsigned long max_nl_cursor = 0;
1497 unsigned long max_nl_size = 0;
1498 unsigned int mapcount;
1499
1500 mutex_lock(&mapping->i_mmap_mutex);
1501 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1502 unsigned long address = vma_address(page, vma);
1503 if (address == -EFAULT)
1504 continue;
1505 ret = try_to_unmap_one(page, vma, address, flags);
1506 if (ret != SWAP_AGAIN || !page_mapped(page))
1507 goto out;
1508 }
1509
1510 if (list_empty(&mapping->i_mmap_nonlinear))
1511 goto out;
1512
1513 /*
1514 * We don't bother to try to find the munlocked page in nonlinears.
1515 * It's costly. Instead, later, page reclaim logic may call
1516 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1517 */
1518 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1519 goto out;
1520
1521 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1522 shared.vm_set.list) {
1523 cursor = (unsigned long) vma->vm_private_data;
1524 if (cursor > max_nl_cursor)
1525 max_nl_cursor = cursor;
1526 cursor = vma->vm_end - vma->vm_start;
1527 if (cursor > max_nl_size)
1528 max_nl_size = cursor;
1529 }
1530
1531 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1532 ret = SWAP_FAIL;
1533 goto out;
1534 }
1535
1536 /*
1537 * We don't try to search for this page in the nonlinear vmas,
1538 * and page_referenced wouldn't have found it anyway. Instead
1539 * just walk the nonlinear vmas trying to age and unmap some.
1540 * The mapcount of the page we came in with is irrelevant,
1541 * but even so use it as a guide to how hard we should try?
1542 */
1543 mapcount = page_mapcount(page);
1544 if (!mapcount)
1545 goto out;
1546 cond_resched();
1547
1548 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1549 if (max_nl_cursor == 0)
1550 max_nl_cursor = CLUSTER_SIZE;
1551
1552 do {
1553 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1554 shared.vm_set.list) {
1555 cursor = (unsigned long) vma->vm_private_data;
1556 while ( cursor < max_nl_cursor &&
1557 cursor < vma->vm_end - vma->vm_start) {
1558 if (try_to_unmap_cluster(cursor, &mapcount,
1559 vma, page) == SWAP_MLOCK)
1560 ret = SWAP_MLOCK;
1561 cursor += CLUSTER_SIZE;
1562 vma->vm_private_data = (void *) cursor;
1563 if ((int)mapcount <= 0)
1564 goto out;
1565 }
1566 vma->vm_private_data = (void *) max_nl_cursor;
1567 }
1568 cond_resched();
1569 max_nl_cursor += CLUSTER_SIZE;
1570 } while (max_nl_cursor <= max_nl_size);
1571
1572 /*
1573 * Don't loop forever (perhaps all the remaining pages are
1574 * in locked vmas). Reset cursor on all unreserved nonlinear
1575 * vmas, now forgetting on which ones it had fallen behind.
1576 */
1577 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1578 vma->vm_private_data = NULL;
1579out:
1580 mutex_unlock(&mapping->i_mmap_mutex);
1581 return ret;
1582}
1583
1584/**
1585 * try_to_unmap - try to remove all page table mappings to a page
1586 * @page: the page to get unmapped
1587 * @flags: action and flags
1588 *
1589 * Tries to remove all the page table entries which are mapping this
1590 * page, used in the pageout path. Caller must hold the page lock.
1591 * Return values are:
1592 *
1593 * SWAP_SUCCESS - we succeeded in removing all mappings
1594 * SWAP_AGAIN - we missed a mapping, try again later
1595 * SWAP_FAIL - the page is unswappable
1596 * SWAP_MLOCK - page is mlocked.
1597 */
1598int try_to_unmap(struct page *page, enum ttu_flags flags)
1599{
1600 int ret;
1601
1602 BUG_ON(!PageLocked(page));
1603 VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
1604
1605 if (unlikely(PageKsm(page)))
1606 ret = try_to_unmap_ksm(page, flags);
1607 else if (PageAnon(page))
1608 ret = try_to_unmap_anon(page, flags);
1609 else
1610 ret = try_to_unmap_file(page, flags);
1611 if (ret != SWAP_MLOCK && !page_mapped(page))
1612 ret = SWAP_SUCCESS;
1613 return ret;
1614}
1615
1616/**
1617 * try_to_munlock - try to munlock a page
1618 * @page: the page to be munlocked
1619 *
1620 * Called from munlock code. Checks all of the VMAs mapping the page
1621 * to make sure nobody else has this page mlocked. The page will be
1622 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1623 *
1624 * Return values are:
1625 *
1626 * SWAP_AGAIN - no vma is holding page mlocked, or,
1627 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1628 * SWAP_FAIL - page cannot be located at present
1629 * SWAP_MLOCK - page is now mlocked.
1630 */
1631int try_to_munlock(struct page *page)
1632{
1633 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1634
1635 if (unlikely(PageKsm(page)))
1636 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1637 else if (PageAnon(page))
1638 return try_to_unmap_anon(page, TTU_MUNLOCK);
1639 else
1640 return try_to_unmap_file(page, TTU_MUNLOCK);
1641}
1642
1643void __put_anon_vma(struct anon_vma *anon_vma)
1644{
1645 struct anon_vma *root = anon_vma->root;
1646
1647 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1648 anon_vma_free(root);
1649
1650 anon_vma_free(anon_vma);
1651}
1652
1653#ifdef CONFIG_MIGRATION
1654/*
1655 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1656 * Called by migrate.c to remove migration ptes, but might be used more later.
1657 */
1658static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1659 struct vm_area_struct *, unsigned long, void *), void *arg)
1660{
1661 struct anon_vma *anon_vma;
1662 struct anon_vma_chain *avc;
1663 int ret = SWAP_AGAIN;
1664
1665 /*
1666 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1667 * because that depends on page_mapped(); but not all its usages
1668 * are holding mmap_sem. Users without mmap_sem are required to
1669 * take a reference count to prevent the anon_vma disappearing
1670 */
1671 anon_vma = page_anon_vma(page);
1672 if (!anon_vma)
1673 return ret;
1674 anon_vma_lock(anon_vma);
1675 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1676 struct vm_area_struct *vma = avc->vma;
1677 unsigned long address = vma_address(page, vma);
1678 if (address == -EFAULT)
1679 continue;
1680 ret = rmap_one(page, vma, address, arg);
1681 if (ret != SWAP_AGAIN)
1682 break;
1683 }
1684 anon_vma_unlock(anon_vma);
1685 return ret;
1686}
1687
1688static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1689 struct vm_area_struct *, unsigned long, void *), void *arg)
1690{
1691 struct address_space *mapping = page->mapping;
1692 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1693 struct vm_area_struct *vma;
1694 struct prio_tree_iter iter;
1695 int ret = SWAP_AGAIN;
1696
1697 if (!mapping)
1698 return ret;
1699 mutex_lock(&mapping->i_mmap_mutex);
1700 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1701 unsigned long address = vma_address(page, vma);
1702 if (address == -EFAULT)
1703 continue;
1704 ret = rmap_one(page, vma, address, arg);
1705 if (ret != SWAP_AGAIN)
1706 break;
1707 }
1708 /*
1709 * No nonlinear handling: being always shared, nonlinear vmas
1710 * never contain migration ptes. Decide what to do about this
1711 * limitation to linear when we need rmap_walk() on nonlinear.
1712 */
1713 mutex_unlock(&mapping->i_mmap_mutex);
1714 return ret;
1715}
1716
1717int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1718 struct vm_area_struct *, unsigned long, void *), void *arg)
1719{
1720 VM_BUG_ON(!PageLocked(page));
1721
1722 if (unlikely(PageKsm(page)))
1723 return rmap_walk_ksm(page, rmap_one, arg);
1724 else if (PageAnon(page))
1725 return rmap_walk_anon(page, rmap_one, arg);
1726 else
1727 return rmap_walk_file(page, rmap_one, arg);
1728}
1729#endif /* CONFIG_MIGRATION */
1730
1731#ifdef CONFIG_HUGETLB_PAGE
1732/*
1733 * The following three functions are for anonymous (private mapped) hugepages.
1734 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1735 * and no lru code, because we handle hugepages differently from common pages.
1736 */
1737static void __hugepage_set_anon_rmap(struct page *page,
1738 struct vm_area_struct *vma, unsigned long address, int exclusive)
1739{
1740 struct anon_vma *anon_vma = vma->anon_vma;
1741
1742 BUG_ON(!anon_vma);
1743
1744 if (PageAnon(page))
1745 return;
1746 if (!exclusive)
1747 anon_vma = anon_vma->root;
1748
1749 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1750 page->mapping = (struct address_space *) anon_vma;
1751 page->index = linear_page_index(vma, address);
1752}
1753
1754void hugepage_add_anon_rmap(struct page *page,
1755 struct vm_area_struct *vma, unsigned long address)
1756{
1757 struct anon_vma *anon_vma = vma->anon_vma;
1758 int first;
1759
1760 BUG_ON(!PageLocked(page));
1761 BUG_ON(!anon_vma);
1762 /* address might be in next vma when migration races vma_adjust */
1763 first = atomic_inc_and_test(&page->_mapcount);
1764 if (first)
1765 __hugepage_set_anon_rmap(page, vma, address, 0);
1766}
1767
1768void hugepage_add_new_anon_rmap(struct page *page,
1769 struct vm_area_struct *vma, unsigned long address)
1770{
1771 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1772 atomic_set(&page->_mapcount, 0);
1773 __hugepage_set_anon_rmap(page, vma, address, 1);
1774}
1775#endif /* CONFIG_HUGETLB_PAGE */