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