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_sem
25 * page->flags PG_locked (lock_page)
26 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
27 * mapping->i_mmap_rwsem
28 * anon_vma->rwsem
29 * mm->page_table_lock or pte_lock
30 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
34 * mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
35 * mapping->tree_lock (widely used)
36 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
37 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
38 * sb_lock (within inode_lock in fs/fs-writeback.c)
39 * mapping->tree_lock (widely used, in set_page_dirty,
40 * in arch-dependent flush_dcache_mmap_lock,
41 * within bdi.wb->list_lock in __sync_single_inode)
42 *
43 * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon)
44 * ->tasklist_lock
45 * pte map lock
46 */
47
48#include <linux/mm.h>
49#include <linux/pagemap.h>
50#include <linux/swap.h>
51#include <linux/swapops.h>
52#include <linux/slab.h>
53#include <linux/init.h>
54#include <linux/ksm.h>
55#include <linux/rmap.h>
56#include <linux/rcupdate.h>
57#include <linux/export.h>
58#include <linux/memcontrol.h>
59#include <linux/mmu_notifier.h>
60#include <linux/migrate.h>
61#include <linux/hugetlb.h>
62#include <linux/backing-dev.h>
63#include <linux/page_idle.h>
64
65#include <asm/tlbflush.h>
66
67#include <trace/events/tlb.h>
68
69#include "internal.h"
70
71static struct kmem_cache *anon_vma_cachep;
72static struct kmem_cache *anon_vma_chain_cachep;
73
74static inline struct anon_vma *anon_vma_alloc(void)
75{
76 struct anon_vma *anon_vma;
77
78 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
79 if (anon_vma) {
80 atomic_set(&anon_vma->refcount, 1);
81 anon_vma->degree = 1; /* Reference for first vma */
82 anon_vma->parent = anon_vma;
83 /*
84 * Initialise the anon_vma root to point to itself. If called
85 * from fork, the root will be reset to the parents anon_vma.
86 */
87 anon_vma->root = anon_vma;
88 }
89
90 return anon_vma;
91}
92
93static inline void anon_vma_free(struct anon_vma *anon_vma)
94{
95 VM_BUG_ON(atomic_read(&anon_vma->refcount));
96
97 /*
98 * Synchronize against page_lock_anon_vma_read() such that
99 * we can safely hold the lock without the anon_vma getting
100 * freed.
101 *
102 * Relies on the full mb implied by the atomic_dec_and_test() from
103 * put_anon_vma() against the acquire barrier implied by
104 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
105 *
106 * page_lock_anon_vma_read() VS put_anon_vma()
107 * down_read_trylock() atomic_dec_and_test()
108 * LOCK MB
109 * atomic_read() rwsem_is_locked()
110 *
111 * LOCK should suffice since the actual taking of the lock must
112 * happen _before_ what follows.
113 */
114 might_sleep();
115 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
116 anon_vma_lock_write(anon_vma);
117 anon_vma_unlock_write(anon_vma);
118 }
119
120 kmem_cache_free(anon_vma_cachep, anon_vma);
121}
122
123static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
124{
125 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
126}
127
128static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
129{
130 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
131}
132
133static void anon_vma_chain_link(struct vm_area_struct *vma,
134 struct anon_vma_chain *avc,
135 struct anon_vma *anon_vma)
136{
137 avc->vma = vma;
138 avc->anon_vma = anon_vma;
139 list_add(&avc->same_vma, &vma->anon_vma_chain);
140 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
141}
142
143/**
144 * anon_vma_prepare - attach an anon_vma to a memory region
145 * @vma: the memory region in question
146 *
147 * This makes sure the memory mapping described by 'vma' has
148 * an 'anon_vma' attached to it, so that we can associate the
149 * anonymous pages mapped into it with that anon_vma.
150 *
151 * The common case will be that we already have one, but if
152 * not we either need to find an adjacent mapping that we
153 * can re-use the anon_vma from (very common when the only
154 * reason for splitting a vma has been mprotect()), or we
155 * allocate a new one.
156 *
157 * Anon-vma allocations are very subtle, because we may have
158 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
159 * and that may actually touch the spinlock even in the newly
160 * allocated vma (it depends on RCU to make sure that the
161 * anon_vma isn't actually destroyed).
162 *
163 * As a result, we need to do proper anon_vma locking even
164 * for the new allocation. At the same time, we do not want
165 * to do any locking for the common case of already having
166 * an anon_vma.
167 *
168 * This must be called with the mmap_sem held for reading.
169 */
170int anon_vma_prepare(struct vm_area_struct *vma)
171{
172 struct anon_vma *anon_vma = vma->anon_vma;
173 struct anon_vma_chain *avc;
174
175 might_sleep();
176 if (unlikely(!anon_vma)) {
177 struct mm_struct *mm = vma->vm_mm;
178 struct anon_vma *allocated;
179
180 avc = anon_vma_chain_alloc(GFP_KERNEL);
181 if (!avc)
182 goto out_enomem;
183
184 anon_vma = find_mergeable_anon_vma(vma);
185 allocated = NULL;
186 if (!anon_vma) {
187 anon_vma = anon_vma_alloc();
188 if (unlikely(!anon_vma))
189 goto out_enomem_free_avc;
190 allocated = anon_vma;
191 }
192
193 anon_vma_lock_write(anon_vma);
194 /* page_table_lock to protect against threads */
195 spin_lock(&mm->page_table_lock);
196 if (likely(!vma->anon_vma)) {
197 vma->anon_vma = anon_vma;
198 anon_vma_chain_link(vma, avc, anon_vma);
199 /* vma reference or self-parent link for new root */
200 anon_vma->degree++;
201 allocated = NULL;
202 avc = NULL;
203 }
204 spin_unlock(&mm->page_table_lock);
205 anon_vma_unlock_write(anon_vma);
206
207 if (unlikely(allocated))
208 put_anon_vma(allocated);
209 if (unlikely(avc))
210 anon_vma_chain_free(avc);
211 }
212 return 0;
213
214 out_enomem_free_avc:
215 anon_vma_chain_free(avc);
216 out_enomem:
217 return -ENOMEM;
218}
219
220/*
221 * This is a useful helper function for locking the anon_vma root as
222 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
223 * have the same vma.
224 *
225 * Such anon_vma's should have the same root, so you'd expect to see
226 * just a single mutex_lock for the whole traversal.
227 */
228static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
229{
230 struct anon_vma *new_root = anon_vma->root;
231 if (new_root != root) {
232 if (WARN_ON_ONCE(root))
233 up_write(&root->rwsem);
234 root = new_root;
235 down_write(&root->rwsem);
236 }
237 return root;
238}
239
240static inline void unlock_anon_vma_root(struct anon_vma *root)
241{
242 if (root)
243 up_write(&root->rwsem);
244}
245
246/*
247 * Attach the anon_vmas from src to dst.
248 * Returns 0 on success, -ENOMEM on failure.
249 *
250 * If dst->anon_vma is NULL this function tries to find and reuse existing
251 * anon_vma which has no vmas and only one child anon_vma. This prevents
252 * degradation of anon_vma hierarchy to endless linear chain in case of
253 * constantly forking task. On the other hand, an anon_vma with more than one
254 * child isn't reused even if there was no alive vma, thus rmap walker has a
255 * good chance of avoiding scanning the whole hierarchy when it searches where
256 * page is mapped.
257 */
258int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
259{
260 struct anon_vma_chain *avc, *pavc;
261 struct anon_vma *root = NULL;
262
263 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
264 struct anon_vma *anon_vma;
265
266 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
267 if (unlikely(!avc)) {
268 unlock_anon_vma_root(root);
269 root = NULL;
270 avc = anon_vma_chain_alloc(GFP_KERNEL);
271 if (!avc)
272 goto enomem_failure;
273 }
274 anon_vma = pavc->anon_vma;
275 root = lock_anon_vma_root(root, anon_vma);
276 anon_vma_chain_link(dst, avc, anon_vma);
277
278 /*
279 * Reuse existing anon_vma if its degree lower than two,
280 * that means it has no vma and only one anon_vma child.
281 *
282 * Do not chose parent anon_vma, otherwise first child
283 * will always reuse it. Root anon_vma is never reused:
284 * it has self-parent reference and at least one child.
285 */
286 if (!dst->anon_vma && anon_vma != src->anon_vma &&
287 anon_vma->degree < 2)
288 dst->anon_vma = anon_vma;
289 }
290 if (dst->anon_vma)
291 dst->anon_vma->degree++;
292 unlock_anon_vma_root(root);
293 return 0;
294
295 enomem_failure:
296 /*
297 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
298 * decremented in unlink_anon_vmas().
299 * We can safely do this because callers of anon_vma_clone() don't care
300 * about dst->anon_vma if anon_vma_clone() failed.
301 */
302 dst->anon_vma = NULL;
303 unlink_anon_vmas(dst);
304 return -ENOMEM;
305}
306
307/*
308 * Attach vma to its own anon_vma, as well as to the anon_vmas that
309 * the corresponding VMA in the parent process is attached to.
310 * Returns 0 on success, non-zero on failure.
311 */
312int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
313{
314 struct anon_vma_chain *avc;
315 struct anon_vma *anon_vma;
316 int error;
317
318 /* Don't bother if the parent process has no anon_vma here. */
319 if (!pvma->anon_vma)
320 return 0;
321
322 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
323 vma->anon_vma = NULL;
324
325 /*
326 * First, attach the new VMA to the parent VMA's anon_vmas,
327 * so rmap can find non-COWed pages in child processes.
328 */
329 error = anon_vma_clone(vma, pvma);
330 if (error)
331 return error;
332
333 /* An existing anon_vma has been reused, all done then. */
334 if (vma->anon_vma)
335 return 0;
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 anon_vma->parent = pvma->anon_vma;
351 /*
352 * With refcounts, an anon_vma can stay around longer than the
353 * process it belongs to. The root anon_vma needs to be pinned until
354 * this anon_vma is freed, because the lock lives in the root.
355 */
356 get_anon_vma(anon_vma->root);
357 /* Mark this anon_vma as the one where our new (COWed) pages go. */
358 vma->anon_vma = anon_vma;
359 anon_vma_lock_write(anon_vma);
360 anon_vma_chain_link(vma, avc, anon_vma);
361 anon_vma->parent->degree++;
362 anon_vma_unlock_write(anon_vma);
363
364 return 0;
365
366 out_error_free_anon_vma:
367 put_anon_vma(anon_vma);
368 out_error:
369 unlink_anon_vmas(vma);
370 return -ENOMEM;
371}
372
373void unlink_anon_vmas(struct vm_area_struct *vma)
374{
375 struct anon_vma_chain *avc, *next;
376 struct anon_vma *root = NULL;
377
378 /*
379 * Unlink each anon_vma chained to the VMA. This list is ordered
380 * from newest to oldest, ensuring the root anon_vma gets freed last.
381 */
382 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
383 struct anon_vma *anon_vma = avc->anon_vma;
384
385 root = lock_anon_vma_root(root, anon_vma);
386 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
387
388 /*
389 * Leave empty anon_vmas on the list - we'll need
390 * to free them outside the lock.
391 */
392 if (RB_EMPTY_ROOT(&anon_vma->rb_root)) {
393 anon_vma->parent->degree--;
394 continue;
395 }
396
397 list_del(&avc->same_vma);
398 anon_vma_chain_free(avc);
399 }
400 if (vma->anon_vma)
401 vma->anon_vma->degree--;
402 unlock_anon_vma_root(root);
403
404 /*
405 * Iterate the list once more, it now only contains empty and unlinked
406 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
407 * needing to write-acquire the anon_vma->root->rwsem.
408 */
409 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
410 struct anon_vma *anon_vma = avc->anon_vma;
411
412 BUG_ON(anon_vma->degree);
413 put_anon_vma(anon_vma);
414
415 list_del(&avc->same_vma);
416 anon_vma_chain_free(avc);
417 }
418}
419
420static void anon_vma_ctor(void *data)
421{
422 struct anon_vma *anon_vma = data;
423
424 init_rwsem(&anon_vma->rwsem);
425 atomic_set(&anon_vma->refcount, 0);
426 anon_vma->rb_root = RB_ROOT;
427}
428
429void __init anon_vma_init(void)
430{
431 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
432 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
433 anon_vma_ctor);
434 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
435 SLAB_PANIC|SLAB_ACCOUNT);
436}
437
438/*
439 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
440 *
441 * Since there is no serialization what so ever against page_remove_rmap()
442 * the best this function can do is return a locked anon_vma that might
443 * have been relevant to this page.
444 *
445 * The page might have been remapped to a different anon_vma or the anon_vma
446 * returned may already be freed (and even reused).
447 *
448 * In case it was remapped to a different anon_vma, the new anon_vma will be a
449 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
450 * ensure that any anon_vma obtained from the page will still be valid for as
451 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
452 *
453 * All users of this function must be very careful when walking the anon_vma
454 * chain and verify that the page in question is indeed mapped in it
455 * [ something equivalent to page_mapped_in_vma() ].
456 *
457 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
458 * that the anon_vma pointer from page->mapping is valid if there is a
459 * mapcount, we can dereference the anon_vma after observing those.
460 */
461struct anon_vma *page_get_anon_vma(struct page *page)
462{
463 struct anon_vma *anon_vma = NULL;
464 unsigned long anon_mapping;
465
466 rcu_read_lock();
467 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
468 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
469 goto out;
470 if (!page_mapped(page))
471 goto out;
472
473 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
474 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
475 anon_vma = NULL;
476 goto out;
477 }
478
479 /*
480 * If this page is still mapped, then its anon_vma cannot have been
481 * freed. But if it has been unmapped, we have no security against the
482 * anon_vma structure being freed and reused (for another anon_vma:
483 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
484 * above cannot corrupt).
485 */
486 if (!page_mapped(page)) {
487 rcu_read_unlock();
488 put_anon_vma(anon_vma);
489 return NULL;
490 }
491out:
492 rcu_read_unlock();
493
494 return anon_vma;
495}
496
497/*
498 * Similar to page_get_anon_vma() except it locks the anon_vma.
499 *
500 * Its a little more complex as it tries to keep the fast path to a single
501 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
502 * reference like with page_get_anon_vma() and then block on the mutex.
503 */
504struct anon_vma *page_lock_anon_vma_read(struct page *page)
505{
506 struct anon_vma *anon_vma = NULL;
507 struct anon_vma *root_anon_vma;
508 unsigned long anon_mapping;
509
510 rcu_read_lock();
511 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
512 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
513 goto out;
514 if (!page_mapped(page))
515 goto out;
516
517 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
518 root_anon_vma = READ_ONCE(anon_vma->root);
519 if (down_read_trylock(&root_anon_vma->rwsem)) {
520 /*
521 * If the page is still mapped, then this anon_vma is still
522 * its anon_vma, and holding the mutex ensures that it will
523 * not go away, see anon_vma_free().
524 */
525 if (!page_mapped(page)) {
526 up_read(&root_anon_vma->rwsem);
527 anon_vma = NULL;
528 }
529 goto out;
530 }
531
532 /* trylock failed, we got to sleep */
533 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
534 anon_vma = NULL;
535 goto out;
536 }
537
538 if (!page_mapped(page)) {
539 rcu_read_unlock();
540 put_anon_vma(anon_vma);
541 return NULL;
542 }
543
544 /* we pinned the anon_vma, its safe to sleep */
545 rcu_read_unlock();
546 anon_vma_lock_read(anon_vma);
547
548 if (atomic_dec_and_test(&anon_vma->refcount)) {
549 /*
550 * Oops, we held the last refcount, release the lock
551 * and bail -- can't simply use put_anon_vma() because
552 * we'll deadlock on the anon_vma_lock_write() recursion.
553 */
554 anon_vma_unlock_read(anon_vma);
555 __put_anon_vma(anon_vma);
556 anon_vma = NULL;
557 }
558
559 return anon_vma;
560
561out:
562 rcu_read_unlock();
563 return anon_vma;
564}
565
566void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
567{
568 anon_vma_unlock_read(anon_vma);
569}
570
571#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
572/*
573 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
574 * important if a PTE was dirty when it was unmapped that it's flushed
575 * before any IO is initiated on the page to prevent lost writes. Similarly,
576 * it must be flushed before freeing to prevent data leakage.
577 */
578void try_to_unmap_flush(void)
579{
580 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
581 int cpu;
582
583 if (!tlb_ubc->flush_required)
584 return;
585
586 cpu = get_cpu();
587
588 if (cpumask_test_cpu(cpu, &tlb_ubc->cpumask)) {
589 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
590 local_flush_tlb();
591 trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
592 }
593
594 if (cpumask_any_but(&tlb_ubc->cpumask, cpu) < nr_cpu_ids)
595 flush_tlb_others(&tlb_ubc->cpumask, NULL, 0, TLB_FLUSH_ALL);
596 cpumask_clear(&tlb_ubc->cpumask);
597 tlb_ubc->flush_required = false;
598 tlb_ubc->writable = false;
599 put_cpu();
600}
601
602/* Flush iff there are potentially writable TLB entries that can race with IO */
603void try_to_unmap_flush_dirty(void)
604{
605 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
606
607 if (tlb_ubc->writable)
608 try_to_unmap_flush();
609}
610
611static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
612 struct page *page, bool writable)
613{
614 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
615
616 cpumask_or(&tlb_ubc->cpumask, &tlb_ubc->cpumask, mm_cpumask(mm));
617 tlb_ubc->flush_required = true;
618
619 /*
620 * If the PTE was dirty then it's best to assume it's writable. The
621 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
622 * before the page is queued for IO.
623 */
624 if (writable)
625 tlb_ubc->writable = true;
626}
627
628/*
629 * Returns true if the TLB flush should be deferred to the end of a batch of
630 * unmap operations to reduce IPIs.
631 */
632static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
633{
634 bool should_defer = false;
635
636 if (!(flags & TTU_BATCH_FLUSH))
637 return false;
638
639 /* If remote CPUs need to be flushed then defer batch the flush */
640 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
641 should_defer = true;
642 put_cpu();
643
644 return should_defer;
645}
646#else
647static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
648 struct page *page, bool writable)
649{
650}
651
652static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
653{
654 return false;
655}
656#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
657
658/*
659 * At what user virtual address is page expected in vma?
660 * Caller should check the page is actually part of the vma.
661 */
662unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
663{
664 unsigned long address;
665 if (PageAnon(page)) {
666 struct anon_vma *page__anon_vma = page_anon_vma(page);
667 /*
668 * Note: swapoff's unuse_vma() is more efficient with this
669 * check, and needs it to match anon_vma when KSM is active.
670 */
671 if (!vma->anon_vma || !page__anon_vma ||
672 vma->anon_vma->root != page__anon_vma->root)
673 return -EFAULT;
674 } else if (page->mapping) {
675 if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
676 return -EFAULT;
677 } else
678 return -EFAULT;
679 address = __vma_address(page, vma);
680 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
681 return -EFAULT;
682 return address;
683}
684
685pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
686{
687 pgd_t *pgd;
688 pud_t *pud;
689 pmd_t *pmd = NULL;
690 pmd_t pmde;
691
692 pgd = pgd_offset(mm, address);
693 if (!pgd_present(*pgd))
694 goto out;
695
696 pud = pud_offset(pgd, address);
697 if (!pud_present(*pud))
698 goto out;
699
700 pmd = pmd_offset(pud, address);
701 /*
702 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
703 * without holding anon_vma lock for write. So when looking for a
704 * genuine pmde (in which to find pte), test present and !THP together.
705 */
706 pmde = *pmd;
707 barrier();
708 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
709 pmd = NULL;
710out:
711 return pmd;
712}
713
714/*
715 * Check that @page is mapped at @address into @mm.
716 *
717 * If @sync is false, page_check_address may perform a racy check to avoid
718 * the page table lock when the pte is not present (helpful when reclaiming
719 * highly shared pages).
720 *
721 * On success returns with pte mapped and locked.
722 */
723pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
724 unsigned long address, spinlock_t **ptlp, int sync)
725{
726 pmd_t *pmd;
727 pte_t *pte;
728 spinlock_t *ptl;
729
730 if (unlikely(PageHuge(page))) {
731 /* when pud is not present, pte will be NULL */
732 pte = huge_pte_offset(mm, address);
733 if (!pte)
734 return NULL;
735
736 ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
737 goto check;
738 }
739
740 pmd = mm_find_pmd(mm, address);
741 if (!pmd)
742 return NULL;
743
744 pte = pte_offset_map(pmd, address);
745 /* Make a quick check before getting the lock */
746 if (!sync && !pte_present(*pte)) {
747 pte_unmap(pte);
748 return NULL;
749 }
750
751 ptl = pte_lockptr(mm, pmd);
752check:
753 spin_lock(ptl);
754 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
755 *ptlp = ptl;
756 return pte;
757 }
758 pte_unmap_unlock(pte, ptl);
759 return NULL;
760}
761
762/**
763 * page_mapped_in_vma - check whether a page is really mapped in a VMA
764 * @page: the page to test
765 * @vma: the VMA to test
766 *
767 * Returns 1 if the page is mapped into the page tables of the VMA, 0
768 * if the page is not mapped into the page tables of this VMA. Only
769 * valid for normal file or anonymous VMAs.
770 */
771int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
772{
773 unsigned long address;
774 pte_t *pte;
775 spinlock_t *ptl;
776
777 address = __vma_address(page, vma);
778 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
779 return 0;
780 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
781 if (!pte) /* the page is not in this mm */
782 return 0;
783 pte_unmap_unlock(pte, ptl);
784
785 return 1;
786}
787
788#ifdef CONFIG_TRANSPARENT_HUGEPAGE
789/*
790 * Check that @page is mapped at @address into @mm. In contrast to
791 * page_check_address(), this function can handle transparent huge pages.
792 *
793 * On success returns true with pte mapped and locked. For PMD-mapped
794 * transparent huge pages *@ptep is set to NULL.
795 */
796bool page_check_address_transhuge(struct page *page, struct mm_struct *mm,
797 unsigned long address, pmd_t **pmdp,
798 pte_t **ptep, spinlock_t **ptlp)
799{
800 pgd_t *pgd;
801 pud_t *pud;
802 pmd_t *pmd;
803 pte_t *pte;
804 spinlock_t *ptl;
805
806 if (unlikely(PageHuge(page))) {
807 /* when pud is not present, pte will be NULL */
808 pte = huge_pte_offset(mm, address);
809 if (!pte)
810 return false;
811
812 ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
813 pmd = NULL;
814 goto check_pte;
815 }
816
817 pgd = pgd_offset(mm, address);
818 if (!pgd_present(*pgd))
819 return false;
820 pud = pud_offset(pgd, address);
821 if (!pud_present(*pud))
822 return false;
823 pmd = pmd_offset(pud, address);
824
825 if (pmd_trans_huge(*pmd)) {
826 ptl = pmd_lock(mm, pmd);
827 if (!pmd_present(*pmd))
828 goto unlock_pmd;
829 if (unlikely(!pmd_trans_huge(*pmd))) {
830 spin_unlock(ptl);
831 goto map_pte;
832 }
833
834 if (pmd_page(*pmd) != page)
835 goto unlock_pmd;
836
837 pte = NULL;
838 goto found;
839unlock_pmd:
840 spin_unlock(ptl);
841 return false;
842 } else {
843 pmd_t pmde = *pmd;
844
845 barrier();
846 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
847 return false;
848 }
849map_pte:
850 pte = pte_offset_map(pmd, address);
851 if (!pte_present(*pte)) {
852 pte_unmap(pte);
853 return false;
854 }
855
856 ptl = pte_lockptr(mm, pmd);
857check_pte:
858 spin_lock(ptl);
859
860 if (!pte_present(*pte)) {
861 pte_unmap_unlock(pte, ptl);
862 return false;
863 }
864
865 /* THP can be referenced by any subpage */
866 if (pte_pfn(*pte) - page_to_pfn(page) >= hpage_nr_pages(page)) {
867 pte_unmap_unlock(pte, ptl);
868 return false;
869 }
870found:
871 *ptep = pte;
872 *pmdp = pmd;
873 *ptlp = ptl;
874 return true;
875}
876#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
877
878struct page_referenced_arg {
879 int mapcount;
880 int referenced;
881 unsigned long vm_flags;
882 struct mem_cgroup *memcg;
883};
884/*
885 * arg: page_referenced_arg will be passed
886 */
887static int page_referenced_one(struct page *page, struct vm_area_struct *vma,
888 unsigned long address, void *arg)
889{
890 struct mm_struct *mm = vma->vm_mm;
891 struct page_referenced_arg *pra = arg;
892 pmd_t *pmd;
893 pte_t *pte;
894 spinlock_t *ptl;
895 int referenced = 0;
896
897 if (!page_check_address_transhuge(page, mm, address, &pmd, &pte, &ptl))
898 return SWAP_AGAIN;
899
900 if (vma->vm_flags & VM_LOCKED) {
901 if (pte)
902 pte_unmap(pte);
903 spin_unlock(ptl);
904 pra->vm_flags |= VM_LOCKED;
905 return SWAP_FAIL; /* To break the loop */
906 }
907
908 if (pte) {
909 if (ptep_clear_flush_young_notify(vma, address, pte)) {
910 /*
911 * Don't treat a reference through a sequentially read
912 * mapping as such. If the page has been used in
913 * another mapping, we will catch it; if this other
914 * mapping is already gone, the unmap path will have
915 * set PG_referenced or activated the page.
916 */
917 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
918 referenced++;
919 }
920 pte_unmap(pte);
921 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
922 if (pmdp_clear_flush_young_notify(vma, address, pmd))
923 referenced++;
924 } else {
925 /* unexpected pmd-mapped page? */
926 WARN_ON_ONCE(1);
927 }
928 spin_unlock(ptl);
929
930 if (referenced)
931 clear_page_idle(page);
932 if (test_and_clear_page_young(page))
933 referenced++;
934
935 if (referenced) {
936 pra->referenced++;
937 pra->vm_flags |= vma->vm_flags;
938 }
939
940 pra->mapcount--;
941 if (!pra->mapcount)
942 return SWAP_SUCCESS; /* To break the loop */
943
944 return SWAP_AGAIN;
945}
946
947static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
948{
949 struct page_referenced_arg *pra = arg;
950 struct mem_cgroup *memcg = pra->memcg;
951
952 if (!mm_match_cgroup(vma->vm_mm, memcg))
953 return true;
954
955 return false;
956}
957
958/**
959 * page_referenced - test if the page was referenced
960 * @page: the page to test
961 * @is_locked: caller holds lock on the page
962 * @memcg: target memory cgroup
963 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
964 *
965 * Quick test_and_clear_referenced for all mappings to a page,
966 * returns the number of ptes which referenced the page.
967 */
968int page_referenced(struct page *page,
969 int is_locked,
970 struct mem_cgroup *memcg,
971 unsigned long *vm_flags)
972{
973 int ret;
974 int we_locked = 0;
975 struct page_referenced_arg pra = {
976 .mapcount = total_mapcount(page),
977 .memcg = memcg,
978 };
979 struct rmap_walk_control rwc = {
980 .rmap_one = page_referenced_one,
981 .arg = (void *)&pra,
982 .anon_lock = page_lock_anon_vma_read,
983 };
984
985 *vm_flags = 0;
986 if (!page_mapped(page))
987 return 0;
988
989 if (!page_rmapping(page))
990 return 0;
991
992 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
993 we_locked = trylock_page(page);
994 if (!we_locked)
995 return 1;
996 }
997
998 /*
999 * If we are reclaiming on behalf of a cgroup, skip
1000 * counting on behalf of references from different
1001 * cgroups
1002 */
1003 if (memcg) {
1004 rwc.invalid_vma = invalid_page_referenced_vma;
1005 }
1006
1007 ret = rmap_walk(page, &rwc);
1008 *vm_flags = pra.vm_flags;
1009
1010 if (we_locked)
1011 unlock_page(page);
1012
1013 return pra.referenced;
1014}
1015
1016static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
1017 unsigned long address, void *arg)
1018{
1019 struct mm_struct *mm = vma->vm_mm;
1020 pte_t *pte;
1021 spinlock_t *ptl;
1022 int ret = 0;
1023 int *cleaned = arg;
1024
1025 pte = page_check_address(page, mm, address, &ptl, 1);
1026 if (!pte)
1027 goto out;
1028
1029 if (pte_dirty(*pte) || pte_write(*pte)) {
1030 pte_t entry;
1031
1032 flush_cache_page(vma, address, pte_pfn(*pte));
1033 entry = ptep_clear_flush(vma, address, pte);
1034 entry = pte_wrprotect(entry);
1035 entry = pte_mkclean(entry);
1036 set_pte_at(mm, address, pte, entry);
1037 ret = 1;
1038 }
1039
1040 pte_unmap_unlock(pte, ptl);
1041
1042 if (ret) {
1043 mmu_notifier_invalidate_page(mm, address);
1044 (*cleaned)++;
1045 }
1046out:
1047 return SWAP_AGAIN;
1048}
1049
1050static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1051{
1052 if (vma->vm_flags & VM_SHARED)
1053 return false;
1054
1055 return true;
1056}
1057
1058int page_mkclean(struct page *page)
1059{
1060 int cleaned = 0;
1061 struct address_space *mapping;
1062 struct rmap_walk_control rwc = {
1063 .arg = (void *)&cleaned,
1064 .rmap_one = page_mkclean_one,
1065 .invalid_vma = invalid_mkclean_vma,
1066 };
1067
1068 BUG_ON(!PageLocked(page));
1069
1070 if (!page_mapped(page))
1071 return 0;
1072
1073 mapping = page_mapping(page);
1074 if (!mapping)
1075 return 0;
1076
1077 rmap_walk(page, &rwc);
1078
1079 return cleaned;
1080}
1081EXPORT_SYMBOL_GPL(page_mkclean);
1082
1083/**
1084 * page_move_anon_rmap - move a page to our anon_vma
1085 * @page: the page to move to our anon_vma
1086 * @vma: the vma the page belongs to
1087 * @address: the user virtual address mapped
1088 *
1089 * When a page belongs exclusively to one process after a COW event,
1090 * that page can be moved into the anon_vma that belongs to just that
1091 * process, so the rmap code will not search the parent or sibling
1092 * processes.
1093 */
1094void page_move_anon_rmap(struct page *page,
1095 struct vm_area_struct *vma, unsigned long address)
1096{
1097 struct anon_vma *anon_vma = vma->anon_vma;
1098
1099 VM_BUG_ON_PAGE(!PageLocked(page), page);
1100 VM_BUG_ON_VMA(!anon_vma, vma);
1101 VM_BUG_ON_PAGE(page->index != linear_page_index(vma, address), page);
1102
1103 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1104 /*
1105 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1106 * simultaneously, so a concurrent reader (eg page_referenced()'s
1107 * PageAnon()) will not see one without the other.
1108 */
1109 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1110}
1111
1112/**
1113 * __page_set_anon_rmap - set up new anonymous rmap
1114 * @page: Page to add to rmap
1115 * @vma: VM area to add page to.
1116 * @address: User virtual address of the mapping
1117 * @exclusive: the page is exclusively owned by the current process
1118 */
1119static void __page_set_anon_rmap(struct page *page,
1120 struct vm_area_struct *vma, unsigned long address, int exclusive)
1121{
1122 struct anon_vma *anon_vma = vma->anon_vma;
1123
1124 BUG_ON(!anon_vma);
1125
1126 if (PageAnon(page))
1127 return;
1128
1129 /*
1130 * If the page isn't exclusively mapped into this vma,
1131 * we must use the _oldest_ possible anon_vma for the
1132 * page mapping!
1133 */
1134 if (!exclusive)
1135 anon_vma = anon_vma->root;
1136
1137 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1138 page->mapping = (struct address_space *) anon_vma;
1139 page->index = linear_page_index(vma, address);
1140}
1141
1142/**
1143 * __page_check_anon_rmap - sanity check anonymous rmap addition
1144 * @page: the page to add the mapping to
1145 * @vma: the vm area in which the mapping is added
1146 * @address: the user virtual address mapped
1147 */
1148static void __page_check_anon_rmap(struct page *page,
1149 struct vm_area_struct *vma, unsigned long address)
1150{
1151#ifdef CONFIG_DEBUG_VM
1152 /*
1153 * The page's anon-rmap details (mapping and index) are guaranteed to
1154 * be set up correctly at this point.
1155 *
1156 * We have exclusion against page_add_anon_rmap because the caller
1157 * always holds the page locked, except if called from page_dup_rmap,
1158 * in which case the page is already known to be setup.
1159 *
1160 * We have exclusion against page_add_new_anon_rmap because those pages
1161 * are initially only visible via the pagetables, and the pte is locked
1162 * over the call to page_add_new_anon_rmap.
1163 */
1164 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1165 BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
1166#endif
1167}
1168
1169/**
1170 * page_add_anon_rmap - add pte mapping to an anonymous page
1171 * @page: the page to add the mapping to
1172 * @vma: the vm area in which the mapping is added
1173 * @address: the user virtual address mapped
1174 * @compound: charge the page as compound or small page
1175 *
1176 * The caller needs to hold the pte lock, and the page must be locked in
1177 * the anon_vma case: to serialize mapping,index checking after setting,
1178 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1179 * (but PageKsm is never downgraded to PageAnon).
1180 */
1181void page_add_anon_rmap(struct page *page,
1182 struct vm_area_struct *vma, unsigned long address, bool compound)
1183{
1184 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1185}
1186
1187/*
1188 * Special version of the above for do_swap_page, which often runs
1189 * into pages that are exclusively owned by the current process.
1190 * Everybody else should continue to use page_add_anon_rmap above.
1191 */
1192void do_page_add_anon_rmap(struct page *page,
1193 struct vm_area_struct *vma, unsigned long address, int flags)
1194{
1195 bool compound = flags & RMAP_COMPOUND;
1196 bool first;
1197
1198 if (compound) {
1199 atomic_t *mapcount;
1200 VM_BUG_ON_PAGE(!PageLocked(page), page);
1201 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1202 mapcount = compound_mapcount_ptr(page);
1203 first = atomic_inc_and_test(mapcount);
1204 } else {
1205 first = atomic_inc_and_test(&page->_mapcount);
1206 }
1207
1208 if (first) {
1209 int nr = compound ? hpage_nr_pages(page) : 1;
1210 /*
1211 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1212 * these counters are not modified in interrupt context, and
1213 * pte lock(a spinlock) is held, which implies preemption
1214 * disabled.
1215 */
1216 if (compound) {
1217 __inc_zone_page_state(page,
1218 NR_ANON_TRANSPARENT_HUGEPAGES);
1219 }
1220 __mod_zone_page_state(page_zone(page), NR_ANON_PAGES, nr);
1221 }
1222 if (unlikely(PageKsm(page)))
1223 return;
1224
1225 VM_BUG_ON_PAGE(!PageLocked(page), page);
1226
1227 /* address might be in next vma when migration races vma_adjust */
1228 if (first)
1229 __page_set_anon_rmap(page, vma, address,
1230 flags & RMAP_EXCLUSIVE);
1231 else
1232 __page_check_anon_rmap(page, vma, address);
1233}
1234
1235/**
1236 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1237 * @page: the page to add the mapping to
1238 * @vma: the vm area in which the mapping is added
1239 * @address: the user virtual address mapped
1240 * @compound: charge the page as compound or small page
1241 *
1242 * Same as page_add_anon_rmap but must only be called on *new* pages.
1243 * This means the inc-and-test can be bypassed.
1244 * Page does not have to be locked.
1245 */
1246void page_add_new_anon_rmap(struct page *page,
1247 struct vm_area_struct *vma, unsigned long address, bool compound)
1248{
1249 int nr = compound ? hpage_nr_pages(page) : 1;
1250
1251 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1252 SetPageSwapBacked(page);
1253 if (compound) {
1254 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1255 /* increment count (starts at -1) */
1256 atomic_set(compound_mapcount_ptr(page), 0);
1257 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1258 } else {
1259 /* Anon THP always mapped first with PMD */
1260 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1261 /* increment count (starts at -1) */
1262 atomic_set(&page->_mapcount, 0);
1263 }
1264 __mod_zone_page_state(page_zone(page), NR_ANON_PAGES, nr);
1265 __page_set_anon_rmap(page, vma, address, 1);
1266}
1267
1268/**
1269 * page_add_file_rmap - add pte mapping to a file page
1270 * @page: the page to add the mapping to
1271 *
1272 * The caller needs to hold the pte lock.
1273 */
1274void page_add_file_rmap(struct page *page)
1275{
1276 lock_page_memcg(page);
1277 if (atomic_inc_and_test(&page->_mapcount)) {
1278 __inc_zone_page_state(page, NR_FILE_MAPPED);
1279 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
1280 }
1281 unlock_page_memcg(page);
1282}
1283
1284static void page_remove_file_rmap(struct page *page)
1285{
1286 lock_page_memcg(page);
1287
1288 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1289 if (unlikely(PageHuge(page))) {
1290 /* hugetlb pages are always mapped with pmds */
1291 atomic_dec(compound_mapcount_ptr(page));
1292 goto out;
1293 }
1294
1295 /* page still mapped by someone else? */
1296 if (!atomic_add_negative(-1, &page->_mapcount))
1297 goto out;
1298
1299 /*
1300 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1301 * these counters are not modified in interrupt context, and
1302 * pte lock(a spinlock) is held, which implies preemption disabled.
1303 */
1304 __dec_zone_page_state(page, NR_FILE_MAPPED);
1305 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
1306
1307 if (unlikely(PageMlocked(page)))
1308 clear_page_mlock(page);
1309out:
1310 unlock_page_memcg(page);
1311}
1312
1313static void page_remove_anon_compound_rmap(struct page *page)
1314{
1315 int i, nr;
1316
1317 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1318 return;
1319
1320 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1321 if (unlikely(PageHuge(page)))
1322 return;
1323
1324 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1325 return;
1326
1327 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1328
1329 if (TestClearPageDoubleMap(page)) {
1330 /*
1331 * Subpages can be mapped with PTEs too. Check how many of
1332 * themi are still mapped.
1333 */
1334 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1335 if (atomic_add_negative(-1, &page[i]._mapcount))
1336 nr++;
1337 }
1338 } else {
1339 nr = HPAGE_PMD_NR;
1340 }
1341
1342 if (unlikely(PageMlocked(page)))
1343 clear_page_mlock(page);
1344
1345 if (nr) {
1346 __mod_zone_page_state(page_zone(page), NR_ANON_PAGES, -nr);
1347 deferred_split_huge_page(page);
1348 }
1349}
1350
1351/**
1352 * page_remove_rmap - take down pte mapping from a page
1353 * @page: page to remove mapping from
1354 * @compound: uncharge the page as compound or small page
1355 *
1356 * The caller needs to hold the pte lock.
1357 */
1358void page_remove_rmap(struct page *page, bool compound)
1359{
1360 if (!PageAnon(page)) {
1361 VM_BUG_ON_PAGE(compound && !PageHuge(page), page);
1362 page_remove_file_rmap(page);
1363 return;
1364 }
1365
1366 if (compound)
1367 return page_remove_anon_compound_rmap(page);
1368
1369 /* page still mapped by someone else? */
1370 if (!atomic_add_negative(-1, &page->_mapcount))
1371 return;
1372
1373 /*
1374 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1375 * these counters are not modified in interrupt context, and
1376 * pte lock(a spinlock) is held, which implies preemption disabled.
1377 */
1378 __dec_zone_page_state(page, NR_ANON_PAGES);
1379
1380 if (unlikely(PageMlocked(page)))
1381 clear_page_mlock(page);
1382
1383 if (PageTransCompound(page))
1384 deferred_split_huge_page(compound_head(page));
1385
1386 /*
1387 * It would be tidy to reset the PageAnon mapping here,
1388 * but that might overwrite a racing page_add_anon_rmap
1389 * which increments mapcount after us but sets mapping
1390 * before us: so leave the reset to free_hot_cold_page,
1391 * and remember that it's only reliable while mapped.
1392 * Leaving it set also helps swapoff to reinstate ptes
1393 * faster for those pages still in swapcache.
1394 */
1395}
1396
1397struct rmap_private {
1398 enum ttu_flags flags;
1399 int lazyfreed;
1400};
1401
1402/*
1403 * @arg: enum ttu_flags will be passed to this argument
1404 */
1405static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1406 unsigned long address, void *arg)
1407{
1408 struct mm_struct *mm = vma->vm_mm;
1409 pte_t *pte;
1410 pte_t pteval;
1411 spinlock_t *ptl;
1412 int ret = SWAP_AGAIN;
1413 struct rmap_private *rp = arg;
1414 enum ttu_flags flags = rp->flags;
1415
1416 /* munlock has nothing to gain from examining un-locked vmas */
1417 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1418 goto out;
1419
1420 if (flags & TTU_SPLIT_HUGE_PMD) {
1421 split_huge_pmd_address(vma, address,
1422 flags & TTU_MIGRATION, page);
1423 /* check if we have anything to do after split */
1424 if (page_mapcount(page) == 0)
1425 goto out;
1426 }
1427
1428 pte = page_check_address(page, mm, address, &ptl, 0);
1429 if (!pte)
1430 goto out;
1431
1432 /*
1433 * If the page is mlock()d, we cannot swap it out.
1434 * If it's recently referenced (perhaps page_referenced
1435 * skipped over this mm) then we should reactivate it.
1436 */
1437 if (!(flags & TTU_IGNORE_MLOCK)) {
1438 if (vma->vm_flags & VM_LOCKED) {
1439 /* Holding pte lock, we do *not* need mmap_sem here */
1440 mlock_vma_page(page);
1441 ret = SWAP_MLOCK;
1442 goto out_unmap;
1443 }
1444 if (flags & TTU_MUNLOCK)
1445 goto out_unmap;
1446 }
1447 if (!(flags & TTU_IGNORE_ACCESS)) {
1448 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1449 ret = SWAP_FAIL;
1450 goto out_unmap;
1451 }
1452 }
1453
1454 /* Nuke the page table entry. */
1455 flush_cache_page(vma, address, page_to_pfn(page));
1456 if (should_defer_flush(mm, flags)) {
1457 /*
1458 * We clear the PTE but do not flush so potentially a remote
1459 * CPU could still be writing to the page. If the entry was
1460 * previously clean then the architecture must guarantee that
1461 * a clear->dirty transition on a cached TLB entry is written
1462 * through and traps if the PTE is unmapped.
1463 */
1464 pteval = ptep_get_and_clear(mm, address, pte);
1465
1466 set_tlb_ubc_flush_pending(mm, page, pte_dirty(pteval));
1467 } else {
1468 pteval = ptep_clear_flush(vma, address, pte);
1469 }
1470
1471 /* Move the dirty bit to the physical page now the pte is gone. */
1472 if (pte_dirty(pteval))
1473 set_page_dirty(page);
1474
1475 /* Update high watermark before we lower rss */
1476 update_hiwater_rss(mm);
1477
1478 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1479 if (PageHuge(page)) {
1480 hugetlb_count_sub(1 << compound_order(page), mm);
1481 } else {
1482 dec_mm_counter(mm, mm_counter(page));
1483 }
1484 set_pte_at(mm, address, pte,
1485 swp_entry_to_pte(make_hwpoison_entry(page)));
1486 } else if (pte_unused(pteval)) {
1487 /*
1488 * The guest indicated that the page content is of no
1489 * interest anymore. Simply discard the pte, vmscan
1490 * will take care of the rest.
1491 */
1492 dec_mm_counter(mm, mm_counter(page));
1493 } else if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION)) {
1494 swp_entry_t entry;
1495 pte_t swp_pte;
1496 /*
1497 * Store the pfn of the page in a special migration
1498 * pte. do_swap_page() will wait until the migration
1499 * pte is removed and then restart fault handling.
1500 */
1501 entry = make_migration_entry(page, pte_write(pteval));
1502 swp_pte = swp_entry_to_pte(entry);
1503 if (pte_soft_dirty(pteval))
1504 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1505 set_pte_at(mm, address, pte, swp_pte);
1506 } else if (PageAnon(page)) {
1507 swp_entry_t entry = { .val = page_private(page) };
1508 pte_t swp_pte;
1509 /*
1510 * Store the swap location in the pte.
1511 * See handle_pte_fault() ...
1512 */
1513 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
1514
1515 if (!PageDirty(page) && (flags & TTU_LZFREE)) {
1516 /* It's a freeable page by MADV_FREE */
1517 dec_mm_counter(mm, MM_ANONPAGES);
1518 rp->lazyfreed++;
1519 goto discard;
1520 }
1521
1522 if (swap_duplicate(entry) < 0) {
1523 set_pte_at(mm, address, pte, pteval);
1524 ret = SWAP_FAIL;
1525 goto out_unmap;
1526 }
1527 if (list_empty(&mm->mmlist)) {
1528 spin_lock(&mmlist_lock);
1529 if (list_empty(&mm->mmlist))
1530 list_add(&mm->mmlist, &init_mm.mmlist);
1531 spin_unlock(&mmlist_lock);
1532 }
1533 dec_mm_counter(mm, MM_ANONPAGES);
1534 inc_mm_counter(mm, MM_SWAPENTS);
1535 swp_pte = swp_entry_to_pte(entry);
1536 if (pte_soft_dirty(pteval))
1537 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1538 set_pte_at(mm, address, pte, swp_pte);
1539 } else
1540 dec_mm_counter(mm, mm_counter_file(page));
1541
1542discard:
1543 page_remove_rmap(page, PageHuge(page));
1544 put_page(page);
1545
1546out_unmap:
1547 pte_unmap_unlock(pte, ptl);
1548 if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK))
1549 mmu_notifier_invalidate_page(mm, address);
1550out:
1551 return ret;
1552}
1553
1554bool is_vma_temporary_stack(struct vm_area_struct *vma)
1555{
1556 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1557
1558 if (!maybe_stack)
1559 return false;
1560
1561 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1562 VM_STACK_INCOMPLETE_SETUP)
1563 return true;
1564
1565 return false;
1566}
1567
1568static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1569{
1570 return is_vma_temporary_stack(vma);
1571}
1572
1573static int page_mapcount_is_zero(struct page *page)
1574{
1575 return !page_mapcount(page);
1576}
1577
1578/**
1579 * try_to_unmap - try to remove all page table mappings to a page
1580 * @page: the page to get unmapped
1581 * @flags: action and flags
1582 *
1583 * Tries to remove all the page table entries which are mapping this
1584 * page, used in the pageout path. Caller must hold the page lock.
1585 * Return values are:
1586 *
1587 * SWAP_SUCCESS - we succeeded in removing all mappings
1588 * SWAP_AGAIN - we missed a mapping, try again later
1589 * SWAP_FAIL - the page is unswappable
1590 * SWAP_MLOCK - page is mlocked.
1591 */
1592int try_to_unmap(struct page *page, enum ttu_flags flags)
1593{
1594 int ret;
1595 struct rmap_private rp = {
1596 .flags = flags,
1597 .lazyfreed = 0,
1598 };
1599
1600 struct rmap_walk_control rwc = {
1601 .rmap_one = try_to_unmap_one,
1602 .arg = &rp,
1603 .done = page_mapcount_is_zero,
1604 .anon_lock = page_lock_anon_vma_read,
1605 };
1606
1607 /*
1608 * During exec, a temporary VMA is setup and later moved.
1609 * The VMA is moved under the anon_vma lock but not the
1610 * page tables leading to a race where migration cannot
1611 * find the migration ptes. Rather than increasing the
1612 * locking requirements of exec(), migration skips
1613 * temporary VMAs until after exec() completes.
1614 */
1615 if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page))
1616 rwc.invalid_vma = invalid_migration_vma;
1617
1618 if (flags & TTU_RMAP_LOCKED)
1619 ret = rmap_walk_locked(page, &rwc);
1620 else
1621 ret = rmap_walk(page, &rwc);
1622
1623 if (ret != SWAP_MLOCK && !page_mapcount(page)) {
1624 ret = SWAP_SUCCESS;
1625 if (rp.lazyfreed && !PageDirty(page))
1626 ret = SWAP_LZFREE;
1627 }
1628 return ret;
1629}
1630
1631static int page_not_mapped(struct page *page)
1632{
1633 return !page_mapped(page);
1634};
1635
1636/**
1637 * try_to_munlock - try to munlock a page
1638 * @page: the page to be munlocked
1639 *
1640 * Called from munlock code. Checks all of the VMAs mapping the page
1641 * to make sure nobody else has this page mlocked. The page will be
1642 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1643 *
1644 * Return values are:
1645 *
1646 * SWAP_AGAIN - no vma is holding page mlocked, or,
1647 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1648 * SWAP_FAIL - page cannot be located at present
1649 * SWAP_MLOCK - page is now mlocked.
1650 */
1651int try_to_munlock(struct page *page)
1652{
1653 int ret;
1654 struct rmap_private rp = {
1655 .flags = TTU_MUNLOCK,
1656 .lazyfreed = 0,
1657 };
1658
1659 struct rmap_walk_control rwc = {
1660 .rmap_one = try_to_unmap_one,
1661 .arg = &rp,
1662 .done = page_not_mapped,
1663 .anon_lock = page_lock_anon_vma_read,
1664
1665 };
1666
1667 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1668
1669 ret = rmap_walk(page, &rwc);
1670 return ret;
1671}
1672
1673void __put_anon_vma(struct anon_vma *anon_vma)
1674{
1675 struct anon_vma *root = anon_vma->root;
1676
1677 anon_vma_free(anon_vma);
1678 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1679 anon_vma_free(root);
1680}
1681
1682static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1683 struct rmap_walk_control *rwc)
1684{
1685 struct anon_vma *anon_vma;
1686
1687 if (rwc->anon_lock)
1688 return rwc->anon_lock(page);
1689
1690 /*
1691 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1692 * because that depends on page_mapped(); but not all its usages
1693 * are holding mmap_sem. Users without mmap_sem are required to
1694 * take a reference count to prevent the anon_vma disappearing
1695 */
1696 anon_vma = page_anon_vma(page);
1697 if (!anon_vma)
1698 return NULL;
1699
1700 anon_vma_lock_read(anon_vma);
1701 return anon_vma;
1702}
1703
1704/*
1705 * rmap_walk_anon - do something to anonymous page using the object-based
1706 * rmap method
1707 * @page: the page to be handled
1708 * @rwc: control variable according to each walk type
1709 *
1710 * Find all the mappings of a page using the mapping pointer and the vma chains
1711 * contained in the anon_vma struct it points to.
1712 *
1713 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1714 * where the page was found will be held for write. So, we won't recheck
1715 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1716 * LOCKED.
1717 */
1718static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1719 bool locked)
1720{
1721 struct anon_vma *anon_vma;
1722 pgoff_t pgoff;
1723 struct anon_vma_chain *avc;
1724 int ret = SWAP_AGAIN;
1725
1726 if (locked) {
1727 anon_vma = page_anon_vma(page);
1728 /* anon_vma disappear under us? */
1729 VM_BUG_ON_PAGE(!anon_vma, page);
1730 } else {
1731 anon_vma = rmap_walk_anon_lock(page, rwc);
1732 }
1733 if (!anon_vma)
1734 return ret;
1735
1736 pgoff = page_to_pgoff(page);
1737 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1738 struct vm_area_struct *vma = avc->vma;
1739 unsigned long address = vma_address(page, vma);
1740
1741 cond_resched();
1742
1743 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1744 continue;
1745
1746 ret = rwc->rmap_one(page, vma, address, rwc->arg);
1747 if (ret != SWAP_AGAIN)
1748 break;
1749 if (rwc->done && rwc->done(page))
1750 break;
1751 }
1752
1753 if (!locked)
1754 anon_vma_unlock_read(anon_vma);
1755 return ret;
1756}
1757
1758/*
1759 * rmap_walk_file - do something to file page using the object-based rmap method
1760 * @page: the page to be handled
1761 * @rwc: control variable according to each walk type
1762 *
1763 * Find all the mappings of a page using the mapping pointer and the vma chains
1764 * contained in the address_space struct it points to.
1765 *
1766 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1767 * where the page was found will be held for write. So, we won't recheck
1768 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1769 * LOCKED.
1770 */
1771static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1772 bool locked)
1773{
1774 struct address_space *mapping = page_mapping(page);
1775 pgoff_t pgoff;
1776 struct vm_area_struct *vma;
1777 int ret = SWAP_AGAIN;
1778
1779 /*
1780 * The page lock not only makes sure that page->mapping cannot
1781 * suddenly be NULLified by truncation, it makes sure that the
1782 * structure at mapping cannot be freed and reused yet,
1783 * so we can safely take mapping->i_mmap_rwsem.
1784 */
1785 VM_BUG_ON_PAGE(!PageLocked(page), page);
1786
1787 if (!mapping)
1788 return ret;
1789
1790 pgoff = page_to_pgoff(page);
1791 if (!locked)
1792 i_mmap_lock_read(mapping);
1793 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1794 unsigned long address = vma_address(page, vma);
1795
1796 cond_resched();
1797
1798 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1799 continue;
1800
1801 ret = rwc->rmap_one(page, vma, address, rwc->arg);
1802 if (ret != SWAP_AGAIN)
1803 goto done;
1804 if (rwc->done && rwc->done(page))
1805 goto done;
1806 }
1807
1808done:
1809 if (!locked)
1810 i_mmap_unlock_read(mapping);
1811 return ret;
1812}
1813
1814int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1815{
1816 if (unlikely(PageKsm(page)))
1817 return rmap_walk_ksm(page, rwc);
1818 else if (PageAnon(page))
1819 return rmap_walk_anon(page, rwc, false);
1820 else
1821 return rmap_walk_file(page, rwc, false);
1822}
1823
1824/* Like rmap_walk, but caller holds relevant rmap lock */
1825int rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1826{
1827 /* no ksm support for now */
1828 VM_BUG_ON_PAGE(PageKsm(page), page);
1829 if (PageAnon(page))
1830 return rmap_walk_anon(page, rwc, true);
1831 else
1832 return rmap_walk_file(page, rwc, true);
1833}
1834
1835#ifdef CONFIG_HUGETLB_PAGE
1836/*
1837 * The following three functions are for anonymous (private mapped) hugepages.
1838 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1839 * and no lru code, because we handle hugepages differently from common pages.
1840 */
1841static void __hugepage_set_anon_rmap(struct page *page,
1842 struct vm_area_struct *vma, unsigned long address, int exclusive)
1843{
1844 struct anon_vma *anon_vma = vma->anon_vma;
1845
1846 BUG_ON(!anon_vma);
1847
1848 if (PageAnon(page))
1849 return;
1850 if (!exclusive)
1851 anon_vma = anon_vma->root;
1852
1853 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1854 page->mapping = (struct address_space *) anon_vma;
1855 page->index = linear_page_index(vma, address);
1856}
1857
1858void hugepage_add_anon_rmap(struct page *page,
1859 struct vm_area_struct *vma, unsigned long address)
1860{
1861 struct anon_vma *anon_vma = vma->anon_vma;
1862 int first;
1863
1864 BUG_ON(!PageLocked(page));
1865 BUG_ON(!anon_vma);
1866 /* address might be in next vma when migration races vma_adjust */
1867 first = atomic_inc_and_test(compound_mapcount_ptr(page));
1868 if (first)
1869 __hugepage_set_anon_rmap(page, vma, address, 0);
1870}
1871
1872void hugepage_add_new_anon_rmap(struct page *page,
1873 struct vm_area_struct *vma, unsigned long address)
1874{
1875 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1876 atomic_set(compound_mapcount_ptr(page), 0);
1877 __hugepage_set_anon_rmap(page, vma, address, 1);
1878}
1879#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->rwsem
28 * mm->page_table_lock or pte_lock
29 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
30 * swap_lock (in swap_duplicate, swap_info_get)
31 * mmlist_lock (in mmput, drain_mmlist and others)
32 * mapping->private_lock (in __set_page_dirty_buffers)
33 * 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->rwsem,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#include <linux/backing-dev.h>
60
61#include <asm/tlbflush.h>
62
63#include "internal.h"
64
65static struct kmem_cache *anon_vma_cachep;
66static struct kmem_cache *anon_vma_chain_cachep;
67
68static inline struct anon_vma *anon_vma_alloc(void)
69{
70 struct anon_vma *anon_vma;
71
72 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
73 if (anon_vma) {
74 atomic_set(&anon_vma->refcount, 1);
75 /*
76 * Initialise the anon_vma root to point to itself. If called
77 * from fork, the root will be reset to the parents anon_vma.
78 */
79 anon_vma->root = anon_vma;
80 }
81
82 return anon_vma;
83}
84
85static inline void anon_vma_free(struct anon_vma *anon_vma)
86{
87 VM_BUG_ON(atomic_read(&anon_vma->refcount));
88
89 /*
90 * Synchronize against page_lock_anon_vma_read() such that
91 * we can safely hold the lock without the anon_vma getting
92 * freed.
93 *
94 * Relies on the full mb implied by the atomic_dec_and_test() from
95 * put_anon_vma() against the acquire barrier implied by
96 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
97 *
98 * page_lock_anon_vma_read() VS put_anon_vma()
99 * down_read_trylock() atomic_dec_and_test()
100 * LOCK MB
101 * atomic_read() rwsem_is_locked()
102 *
103 * LOCK should suffice since the actual taking of the lock must
104 * happen _before_ what follows.
105 */
106 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
107 anon_vma_lock_write(anon_vma);
108 anon_vma_unlock_write(anon_vma);
109 }
110
111 kmem_cache_free(anon_vma_cachep, anon_vma);
112}
113
114static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
115{
116 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
117}
118
119static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
120{
121 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
122}
123
124static void anon_vma_chain_link(struct vm_area_struct *vma,
125 struct anon_vma_chain *avc,
126 struct anon_vma *anon_vma)
127{
128 avc->vma = vma;
129 avc->anon_vma = anon_vma;
130 list_add(&avc->same_vma, &vma->anon_vma_chain);
131 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
132}
133
134/**
135 * anon_vma_prepare - attach an anon_vma to a memory region
136 * @vma: the memory region in question
137 *
138 * This makes sure the memory mapping described by 'vma' has
139 * an 'anon_vma' attached to it, so that we can associate the
140 * anonymous pages mapped into it with that anon_vma.
141 *
142 * The common case will be that we already have one, but if
143 * not we either need to find an adjacent mapping that we
144 * can re-use the anon_vma from (very common when the only
145 * reason for splitting a vma has been mprotect()), or we
146 * allocate a new one.
147 *
148 * Anon-vma allocations are very subtle, because we may have
149 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
150 * and that may actually touch the spinlock even in the newly
151 * allocated vma (it depends on RCU to make sure that the
152 * anon_vma isn't actually destroyed).
153 *
154 * As a result, we need to do proper anon_vma locking even
155 * for the new allocation. At the same time, we do not want
156 * to do any locking for the common case of already having
157 * an anon_vma.
158 *
159 * This must be called with the mmap_sem held for reading.
160 */
161int anon_vma_prepare(struct vm_area_struct *vma)
162{
163 struct anon_vma *anon_vma = vma->anon_vma;
164 struct anon_vma_chain *avc;
165
166 might_sleep();
167 if (unlikely(!anon_vma)) {
168 struct mm_struct *mm = vma->vm_mm;
169 struct anon_vma *allocated;
170
171 avc = anon_vma_chain_alloc(GFP_KERNEL);
172 if (!avc)
173 goto out_enomem;
174
175 anon_vma = find_mergeable_anon_vma(vma);
176 allocated = NULL;
177 if (!anon_vma) {
178 anon_vma = anon_vma_alloc();
179 if (unlikely(!anon_vma))
180 goto out_enomem_free_avc;
181 allocated = anon_vma;
182 }
183
184 anon_vma_lock_write(anon_vma);
185 /* page_table_lock to protect against threads */
186 spin_lock(&mm->page_table_lock);
187 if (likely(!vma->anon_vma)) {
188 vma->anon_vma = anon_vma;
189 anon_vma_chain_link(vma, avc, anon_vma);
190 allocated = NULL;
191 avc = NULL;
192 }
193 spin_unlock(&mm->page_table_lock);
194 anon_vma_unlock_write(anon_vma);
195
196 if (unlikely(allocated))
197 put_anon_vma(allocated);
198 if (unlikely(avc))
199 anon_vma_chain_free(avc);
200 }
201 return 0;
202
203 out_enomem_free_avc:
204 anon_vma_chain_free(avc);
205 out_enomem:
206 return -ENOMEM;
207}
208
209/*
210 * This is a useful helper function for locking the anon_vma root as
211 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
212 * have the same vma.
213 *
214 * Such anon_vma's should have the same root, so you'd expect to see
215 * just a single mutex_lock for the whole traversal.
216 */
217static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
218{
219 struct anon_vma *new_root = anon_vma->root;
220 if (new_root != root) {
221 if (WARN_ON_ONCE(root))
222 up_write(&root->rwsem);
223 root = new_root;
224 down_write(&root->rwsem);
225 }
226 return root;
227}
228
229static inline void unlock_anon_vma_root(struct anon_vma *root)
230{
231 if (root)
232 up_write(&root->rwsem);
233}
234
235/*
236 * Attach the anon_vmas from src to dst.
237 * Returns 0 on success, -ENOMEM on failure.
238 */
239int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
240{
241 struct anon_vma_chain *avc, *pavc;
242 struct anon_vma *root = NULL;
243
244 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
245 struct anon_vma *anon_vma;
246
247 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
248 if (unlikely(!avc)) {
249 unlock_anon_vma_root(root);
250 root = NULL;
251 avc = anon_vma_chain_alloc(GFP_KERNEL);
252 if (!avc)
253 goto enomem_failure;
254 }
255 anon_vma = pavc->anon_vma;
256 root = lock_anon_vma_root(root, anon_vma);
257 anon_vma_chain_link(dst, avc, anon_vma);
258 }
259 unlock_anon_vma_root(root);
260 return 0;
261
262 enomem_failure:
263 unlink_anon_vmas(dst);
264 return -ENOMEM;
265}
266
267/*
268 * Attach vma to its own anon_vma, as well as to the anon_vmas that
269 * the corresponding VMA in the parent process is attached to.
270 * Returns 0 on success, non-zero on failure.
271 */
272int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
273{
274 struct anon_vma_chain *avc;
275 struct anon_vma *anon_vma;
276
277 /* Don't bother if the parent process has no anon_vma here. */
278 if (!pvma->anon_vma)
279 return 0;
280
281 /*
282 * First, attach the new VMA to the parent VMA's anon_vmas,
283 * so rmap can find non-COWed pages in child processes.
284 */
285 if (anon_vma_clone(vma, pvma))
286 return -ENOMEM;
287
288 /* Then add our own anon_vma. */
289 anon_vma = anon_vma_alloc();
290 if (!anon_vma)
291 goto out_error;
292 avc = anon_vma_chain_alloc(GFP_KERNEL);
293 if (!avc)
294 goto out_error_free_anon_vma;
295
296 /*
297 * The root anon_vma's spinlock is the lock actually used when we
298 * lock any of the anon_vmas in this anon_vma tree.
299 */
300 anon_vma->root = pvma->anon_vma->root;
301 /*
302 * With refcounts, an anon_vma can stay around longer than the
303 * process it belongs to. The root anon_vma needs to be pinned until
304 * this anon_vma is freed, because the lock lives in the root.
305 */
306 get_anon_vma(anon_vma->root);
307 /* Mark this anon_vma as the one where our new (COWed) pages go. */
308 vma->anon_vma = anon_vma;
309 anon_vma_lock_write(anon_vma);
310 anon_vma_chain_link(vma, avc, anon_vma);
311 anon_vma_unlock_write(anon_vma);
312
313 return 0;
314
315 out_error_free_anon_vma:
316 put_anon_vma(anon_vma);
317 out_error:
318 unlink_anon_vmas(vma);
319 return -ENOMEM;
320}
321
322void unlink_anon_vmas(struct vm_area_struct *vma)
323{
324 struct anon_vma_chain *avc, *next;
325 struct anon_vma *root = NULL;
326
327 /*
328 * Unlink each anon_vma chained to the VMA. This list is ordered
329 * from newest to oldest, ensuring the root anon_vma gets freed last.
330 */
331 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
332 struct anon_vma *anon_vma = avc->anon_vma;
333
334 root = lock_anon_vma_root(root, anon_vma);
335 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
336
337 /*
338 * Leave empty anon_vmas on the list - we'll need
339 * to free them outside the lock.
340 */
341 if (RB_EMPTY_ROOT(&anon_vma->rb_root))
342 continue;
343
344 list_del(&avc->same_vma);
345 anon_vma_chain_free(avc);
346 }
347 unlock_anon_vma_root(root);
348
349 /*
350 * Iterate the list once more, it now only contains empty and unlinked
351 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
352 * needing to write-acquire the anon_vma->root->rwsem.
353 */
354 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
355 struct anon_vma *anon_vma = avc->anon_vma;
356
357 put_anon_vma(anon_vma);
358
359 list_del(&avc->same_vma);
360 anon_vma_chain_free(avc);
361 }
362}
363
364static void anon_vma_ctor(void *data)
365{
366 struct anon_vma *anon_vma = data;
367
368 init_rwsem(&anon_vma->rwsem);
369 atomic_set(&anon_vma->refcount, 0);
370 anon_vma->rb_root = RB_ROOT;
371}
372
373void __init anon_vma_init(void)
374{
375 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
376 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
377 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
378}
379
380/*
381 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
382 *
383 * Since there is no serialization what so ever against page_remove_rmap()
384 * the best this function can do is return a locked anon_vma that might
385 * have been relevant to this page.
386 *
387 * The page might have been remapped to a different anon_vma or the anon_vma
388 * returned may already be freed (and even reused).
389 *
390 * In case it was remapped to a different anon_vma, the new anon_vma will be a
391 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
392 * ensure that any anon_vma obtained from the page will still be valid for as
393 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
394 *
395 * All users of this function must be very careful when walking the anon_vma
396 * chain and verify that the page in question is indeed mapped in it
397 * [ something equivalent to page_mapped_in_vma() ].
398 *
399 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
400 * that the anon_vma pointer from page->mapping is valid if there is a
401 * mapcount, we can dereference the anon_vma after observing those.
402 */
403struct anon_vma *page_get_anon_vma(struct page *page)
404{
405 struct anon_vma *anon_vma = NULL;
406 unsigned long anon_mapping;
407
408 rcu_read_lock();
409 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
410 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
411 goto out;
412 if (!page_mapped(page))
413 goto out;
414
415 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
416 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
417 anon_vma = NULL;
418 goto out;
419 }
420
421 /*
422 * If this page is still mapped, then its anon_vma cannot have been
423 * freed. But if it has been unmapped, we have no security against the
424 * anon_vma structure being freed and reused (for another anon_vma:
425 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
426 * above cannot corrupt).
427 */
428 if (!page_mapped(page)) {
429 put_anon_vma(anon_vma);
430 anon_vma = NULL;
431 }
432out:
433 rcu_read_unlock();
434
435 return anon_vma;
436}
437
438/*
439 * Similar to page_get_anon_vma() except it locks the anon_vma.
440 *
441 * Its a little more complex as it tries to keep the fast path to a single
442 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
443 * reference like with page_get_anon_vma() and then block on the mutex.
444 */
445struct anon_vma *page_lock_anon_vma_read(struct page *page)
446{
447 struct anon_vma *anon_vma = NULL;
448 struct anon_vma *root_anon_vma;
449 unsigned long anon_mapping;
450
451 rcu_read_lock();
452 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
453 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
454 goto out;
455 if (!page_mapped(page))
456 goto out;
457
458 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
459 root_anon_vma = ACCESS_ONCE(anon_vma->root);
460 if (down_read_trylock(&root_anon_vma->rwsem)) {
461 /*
462 * If the page is still mapped, then this anon_vma is still
463 * its anon_vma, and holding the mutex ensures that it will
464 * not go away, see anon_vma_free().
465 */
466 if (!page_mapped(page)) {
467 up_read(&root_anon_vma->rwsem);
468 anon_vma = NULL;
469 }
470 goto out;
471 }
472
473 /* trylock failed, we got to sleep */
474 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
475 anon_vma = NULL;
476 goto out;
477 }
478
479 if (!page_mapped(page)) {
480 put_anon_vma(anon_vma);
481 anon_vma = NULL;
482 goto out;
483 }
484
485 /* we pinned the anon_vma, its safe to sleep */
486 rcu_read_unlock();
487 anon_vma_lock_read(anon_vma);
488
489 if (atomic_dec_and_test(&anon_vma->refcount)) {
490 /*
491 * Oops, we held the last refcount, release the lock
492 * and bail -- can't simply use put_anon_vma() because
493 * we'll deadlock on the anon_vma_lock_write() recursion.
494 */
495 anon_vma_unlock_read(anon_vma);
496 __put_anon_vma(anon_vma);
497 anon_vma = NULL;
498 }
499
500 return anon_vma;
501
502out:
503 rcu_read_unlock();
504 return anon_vma;
505}
506
507void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
508{
509 anon_vma_unlock_read(anon_vma);
510}
511
512/*
513 * At what user virtual address is page expected in @vma?
514 */
515static inline unsigned long
516__vma_address(struct page *page, struct vm_area_struct *vma)
517{
518 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
519
520 if (unlikely(is_vm_hugetlb_page(vma)))
521 pgoff = page->index << huge_page_order(page_hstate(page));
522
523 return vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
524}
525
526inline unsigned long
527vma_address(struct page *page, struct vm_area_struct *vma)
528{
529 unsigned long address = __vma_address(page, vma);
530
531 /* page should be within @vma mapping range */
532 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
533
534 return address;
535}
536
537/*
538 * At what user virtual address is page expected in vma?
539 * Caller should check the page is actually part of the vma.
540 */
541unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
542{
543 unsigned long address;
544 if (PageAnon(page)) {
545 struct anon_vma *page__anon_vma = page_anon_vma(page);
546 /*
547 * Note: swapoff's unuse_vma() is more efficient with this
548 * check, and needs it to match anon_vma when KSM is active.
549 */
550 if (!vma->anon_vma || !page__anon_vma ||
551 vma->anon_vma->root != page__anon_vma->root)
552 return -EFAULT;
553 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
554 if (!vma->vm_file ||
555 vma->vm_file->f_mapping != page->mapping)
556 return -EFAULT;
557 } else
558 return -EFAULT;
559 address = __vma_address(page, vma);
560 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
561 return -EFAULT;
562 return address;
563}
564
565pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
566{
567 pgd_t *pgd;
568 pud_t *pud;
569 pmd_t *pmd = NULL;
570
571 pgd = pgd_offset(mm, address);
572 if (!pgd_present(*pgd))
573 goto out;
574
575 pud = pud_offset(pgd, address);
576 if (!pud_present(*pud))
577 goto out;
578
579 pmd = pmd_offset(pud, address);
580 if (!pmd_present(*pmd))
581 pmd = NULL;
582out:
583 return pmd;
584}
585
586/*
587 * Check that @page is mapped at @address into @mm.
588 *
589 * If @sync is false, page_check_address may perform a racy check to avoid
590 * the page table lock when the pte is not present (helpful when reclaiming
591 * highly shared pages).
592 *
593 * On success returns with pte mapped and locked.
594 */
595pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
596 unsigned long address, spinlock_t **ptlp, int sync)
597{
598 pmd_t *pmd;
599 pte_t *pte;
600 spinlock_t *ptl;
601
602 if (unlikely(PageHuge(page))) {
603 /* when pud is not present, pte will be NULL */
604 pte = huge_pte_offset(mm, address);
605 if (!pte)
606 return NULL;
607
608 ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
609 goto check;
610 }
611
612 pmd = mm_find_pmd(mm, address);
613 if (!pmd)
614 return NULL;
615
616 if (pmd_trans_huge(*pmd))
617 return NULL;
618
619 pte = pte_offset_map(pmd, address);
620 /* Make a quick check before getting the lock */
621 if (!sync && !pte_present(*pte)) {
622 pte_unmap(pte);
623 return NULL;
624 }
625
626 ptl = pte_lockptr(mm, pmd);
627check:
628 spin_lock(ptl);
629 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
630 *ptlp = ptl;
631 return pte;
632 }
633 pte_unmap_unlock(pte, ptl);
634 return NULL;
635}
636
637/**
638 * page_mapped_in_vma - check whether a page is really mapped in a VMA
639 * @page: the page to test
640 * @vma: the VMA to test
641 *
642 * Returns 1 if the page is mapped into the page tables of the VMA, 0
643 * if the page is not mapped into the page tables of this VMA. Only
644 * valid for normal file or anonymous VMAs.
645 */
646int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
647{
648 unsigned long address;
649 pte_t *pte;
650 spinlock_t *ptl;
651
652 address = __vma_address(page, vma);
653 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
654 return 0;
655 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
656 if (!pte) /* the page is not in this mm */
657 return 0;
658 pte_unmap_unlock(pte, ptl);
659
660 return 1;
661}
662
663struct page_referenced_arg {
664 int mapcount;
665 int referenced;
666 unsigned long vm_flags;
667 struct mem_cgroup *memcg;
668};
669/*
670 * arg: page_referenced_arg will be passed
671 */
672int page_referenced_one(struct page *page, struct vm_area_struct *vma,
673 unsigned long address, void *arg)
674{
675 struct mm_struct *mm = vma->vm_mm;
676 spinlock_t *ptl;
677 int referenced = 0;
678 struct page_referenced_arg *pra = arg;
679
680 if (unlikely(PageTransHuge(page))) {
681 pmd_t *pmd;
682
683 /*
684 * rmap might return false positives; we must filter
685 * these out using page_check_address_pmd().
686 */
687 pmd = page_check_address_pmd(page, mm, address,
688 PAGE_CHECK_ADDRESS_PMD_FLAG, &ptl);
689 if (!pmd)
690 return SWAP_AGAIN;
691
692 if (vma->vm_flags & VM_LOCKED) {
693 spin_unlock(ptl);
694 pra->vm_flags |= VM_LOCKED;
695 return SWAP_FAIL; /* To break the loop */
696 }
697
698 /* go ahead even if the pmd is pmd_trans_splitting() */
699 if (pmdp_clear_flush_young_notify(vma, address, pmd))
700 referenced++;
701 spin_unlock(ptl);
702 } else {
703 pte_t *pte;
704
705 /*
706 * rmap might return false positives; we must filter
707 * these out using page_check_address().
708 */
709 pte = page_check_address(page, mm, address, &ptl, 0);
710 if (!pte)
711 return SWAP_AGAIN;
712
713 if (vma->vm_flags & VM_LOCKED) {
714 pte_unmap_unlock(pte, ptl);
715 pra->vm_flags |= VM_LOCKED;
716 return SWAP_FAIL; /* To break the loop */
717 }
718
719 if (ptep_clear_flush_young_notify(vma, address, pte)) {
720 /*
721 * Don't treat a reference through a sequentially read
722 * mapping as such. If the page has been used in
723 * another mapping, we will catch it; if this other
724 * mapping is already gone, the unmap path will have
725 * set PG_referenced or activated the page.
726 */
727 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
728 referenced++;
729 }
730 pte_unmap_unlock(pte, ptl);
731 }
732
733 if (referenced) {
734 pra->referenced++;
735 pra->vm_flags |= vma->vm_flags;
736 }
737
738 pra->mapcount--;
739 if (!pra->mapcount)
740 return SWAP_SUCCESS; /* To break the loop */
741
742 return SWAP_AGAIN;
743}
744
745static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
746{
747 struct page_referenced_arg *pra = arg;
748 struct mem_cgroup *memcg = pra->memcg;
749
750 if (!mm_match_cgroup(vma->vm_mm, memcg))
751 return true;
752
753 return false;
754}
755
756/**
757 * page_referenced - test if the page was referenced
758 * @page: the page to test
759 * @is_locked: caller holds lock on the page
760 * @memcg: target memory cgroup
761 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
762 *
763 * Quick test_and_clear_referenced for all mappings to a page,
764 * returns the number of ptes which referenced the page.
765 */
766int page_referenced(struct page *page,
767 int is_locked,
768 struct mem_cgroup *memcg,
769 unsigned long *vm_flags)
770{
771 int ret;
772 int we_locked = 0;
773 struct page_referenced_arg pra = {
774 .mapcount = page_mapcount(page),
775 .memcg = memcg,
776 };
777 struct rmap_walk_control rwc = {
778 .rmap_one = page_referenced_one,
779 .arg = (void *)&pra,
780 .anon_lock = page_lock_anon_vma_read,
781 };
782
783 *vm_flags = 0;
784 if (!page_mapped(page))
785 return 0;
786
787 if (!page_rmapping(page))
788 return 0;
789
790 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
791 we_locked = trylock_page(page);
792 if (!we_locked)
793 return 1;
794 }
795
796 /*
797 * If we are reclaiming on behalf of a cgroup, skip
798 * counting on behalf of references from different
799 * cgroups
800 */
801 if (memcg) {
802 rwc.invalid_vma = invalid_page_referenced_vma;
803 }
804
805 ret = rmap_walk(page, &rwc);
806 *vm_flags = pra.vm_flags;
807
808 if (we_locked)
809 unlock_page(page);
810
811 return pra.referenced;
812}
813
814static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
815 unsigned long address, void *arg)
816{
817 struct mm_struct *mm = vma->vm_mm;
818 pte_t *pte;
819 spinlock_t *ptl;
820 int ret = 0;
821 int *cleaned = arg;
822
823 pte = page_check_address(page, mm, address, &ptl, 1);
824 if (!pte)
825 goto out;
826
827 if (pte_dirty(*pte) || pte_write(*pte)) {
828 pte_t entry;
829
830 flush_cache_page(vma, address, pte_pfn(*pte));
831 entry = ptep_clear_flush(vma, address, pte);
832 entry = pte_wrprotect(entry);
833 entry = pte_mkclean(entry);
834 set_pte_at(mm, address, pte, entry);
835 ret = 1;
836 }
837
838 pte_unmap_unlock(pte, ptl);
839
840 if (ret) {
841 mmu_notifier_invalidate_page(mm, address);
842 (*cleaned)++;
843 }
844out:
845 return SWAP_AGAIN;
846}
847
848static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
849{
850 if (vma->vm_flags & VM_SHARED)
851 return false;
852
853 return true;
854}
855
856int page_mkclean(struct page *page)
857{
858 int cleaned = 0;
859 struct address_space *mapping;
860 struct rmap_walk_control rwc = {
861 .arg = (void *)&cleaned,
862 .rmap_one = page_mkclean_one,
863 .invalid_vma = invalid_mkclean_vma,
864 };
865
866 BUG_ON(!PageLocked(page));
867
868 if (!page_mapped(page))
869 return 0;
870
871 mapping = page_mapping(page);
872 if (!mapping)
873 return 0;
874
875 rmap_walk(page, &rwc);
876
877 return cleaned;
878}
879EXPORT_SYMBOL_GPL(page_mkclean);
880
881/**
882 * page_move_anon_rmap - move a page to our anon_vma
883 * @page: the page to move to our anon_vma
884 * @vma: the vma the page belongs to
885 * @address: the user virtual address mapped
886 *
887 * When a page belongs exclusively to one process after a COW event,
888 * that page can be moved into the anon_vma that belongs to just that
889 * process, so the rmap code will not search the parent or sibling
890 * processes.
891 */
892void page_move_anon_rmap(struct page *page,
893 struct vm_area_struct *vma, unsigned long address)
894{
895 struct anon_vma *anon_vma = vma->anon_vma;
896
897 VM_BUG_ON_PAGE(!PageLocked(page), page);
898 VM_BUG_ON(!anon_vma);
899 VM_BUG_ON_PAGE(page->index != linear_page_index(vma, address), page);
900
901 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
902 page->mapping = (struct address_space *) anon_vma;
903}
904
905/**
906 * __page_set_anon_rmap - set up new anonymous rmap
907 * @page: Page to add to rmap
908 * @vma: VM area to add page to.
909 * @address: User virtual address of the mapping
910 * @exclusive: the page is exclusively owned by the current process
911 */
912static void __page_set_anon_rmap(struct page *page,
913 struct vm_area_struct *vma, unsigned long address, int exclusive)
914{
915 struct anon_vma *anon_vma = vma->anon_vma;
916
917 BUG_ON(!anon_vma);
918
919 if (PageAnon(page))
920 return;
921
922 /*
923 * If the page isn't exclusively mapped into this vma,
924 * we must use the _oldest_ possible anon_vma for the
925 * page mapping!
926 */
927 if (!exclusive)
928 anon_vma = anon_vma->root;
929
930 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
931 page->mapping = (struct address_space *) anon_vma;
932 page->index = linear_page_index(vma, address);
933}
934
935/**
936 * __page_check_anon_rmap - sanity check anonymous rmap addition
937 * @page: the page to add the mapping to
938 * @vma: the vm area in which the mapping is added
939 * @address: the user virtual address mapped
940 */
941static void __page_check_anon_rmap(struct page *page,
942 struct vm_area_struct *vma, unsigned long address)
943{
944#ifdef CONFIG_DEBUG_VM
945 /*
946 * The page's anon-rmap details (mapping and index) are guaranteed to
947 * be set up correctly at this point.
948 *
949 * We have exclusion against page_add_anon_rmap because the caller
950 * always holds the page locked, except if called from page_dup_rmap,
951 * in which case the page is already known to be setup.
952 *
953 * We have exclusion against page_add_new_anon_rmap because those pages
954 * are initially only visible via the pagetables, and the pte is locked
955 * over the call to page_add_new_anon_rmap.
956 */
957 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
958 BUG_ON(page->index != linear_page_index(vma, address));
959#endif
960}
961
962/**
963 * page_add_anon_rmap - add pte mapping to an anonymous page
964 * @page: the page to add the mapping to
965 * @vma: the vm area in which the mapping is added
966 * @address: the user virtual address mapped
967 *
968 * The caller needs to hold the pte lock, and the page must be locked in
969 * the anon_vma case: to serialize mapping,index checking after setting,
970 * and to ensure that PageAnon is not being upgraded racily to PageKsm
971 * (but PageKsm is never downgraded to PageAnon).
972 */
973void page_add_anon_rmap(struct page *page,
974 struct vm_area_struct *vma, unsigned long address)
975{
976 do_page_add_anon_rmap(page, vma, address, 0);
977}
978
979/*
980 * Special version of the above for do_swap_page, which often runs
981 * into pages that are exclusively owned by the current process.
982 * Everybody else should continue to use page_add_anon_rmap above.
983 */
984void do_page_add_anon_rmap(struct page *page,
985 struct vm_area_struct *vma, unsigned long address, int exclusive)
986{
987 int first = atomic_inc_and_test(&page->_mapcount);
988 if (first) {
989 if (PageTransHuge(page))
990 __inc_zone_page_state(page,
991 NR_ANON_TRANSPARENT_HUGEPAGES);
992 __mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
993 hpage_nr_pages(page));
994 }
995 if (unlikely(PageKsm(page)))
996 return;
997
998 VM_BUG_ON_PAGE(!PageLocked(page), page);
999 /* address might be in next vma when migration races vma_adjust */
1000 if (first)
1001 __page_set_anon_rmap(page, vma, address, exclusive);
1002 else
1003 __page_check_anon_rmap(page, vma, address);
1004}
1005
1006/**
1007 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1008 * @page: the page to add the mapping to
1009 * @vma: the vm area in which the mapping is added
1010 * @address: the user virtual address mapped
1011 *
1012 * Same as page_add_anon_rmap but must only be called on *new* pages.
1013 * This means the inc-and-test can be bypassed.
1014 * Page does not have to be locked.
1015 */
1016void page_add_new_anon_rmap(struct page *page,
1017 struct vm_area_struct *vma, unsigned long address)
1018{
1019 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1020 SetPageSwapBacked(page);
1021 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1022 if (PageTransHuge(page))
1023 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1024 __mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1025 hpage_nr_pages(page));
1026 __page_set_anon_rmap(page, vma, address, 1);
1027 if (!mlocked_vma_newpage(vma, page)) {
1028 SetPageActive(page);
1029 lru_cache_add(page);
1030 } else
1031 add_page_to_unevictable_list(page);
1032}
1033
1034/**
1035 * page_add_file_rmap - add pte mapping to a file page
1036 * @page: the page to add the mapping to
1037 *
1038 * The caller needs to hold the pte lock.
1039 */
1040void page_add_file_rmap(struct page *page)
1041{
1042 bool locked;
1043 unsigned long flags;
1044
1045 mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1046 if (atomic_inc_and_test(&page->_mapcount)) {
1047 __inc_zone_page_state(page, NR_FILE_MAPPED);
1048 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
1049 }
1050 mem_cgroup_end_update_page_stat(page, &locked, &flags);
1051}
1052
1053/**
1054 * page_remove_rmap - take down pte mapping from a page
1055 * @page: page to remove mapping from
1056 *
1057 * The caller needs to hold the pte lock.
1058 */
1059void page_remove_rmap(struct page *page)
1060{
1061 bool anon = PageAnon(page);
1062 bool locked;
1063 unsigned long flags;
1064
1065 /*
1066 * The anon case has no mem_cgroup page_stat to update; but may
1067 * uncharge_page() below, where the lock ordering can deadlock if
1068 * we hold the lock against page_stat move: so avoid it on anon.
1069 */
1070 if (!anon)
1071 mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1072
1073 /* page still mapped by someone else? */
1074 if (!atomic_add_negative(-1, &page->_mapcount))
1075 goto out;
1076
1077 /*
1078 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1079 * and not charged by memcg for now.
1080 */
1081 if (unlikely(PageHuge(page)))
1082 goto out;
1083 if (anon) {
1084 mem_cgroup_uncharge_page(page);
1085 if (PageTransHuge(page))
1086 __dec_zone_page_state(page,
1087 NR_ANON_TRANSPARENT_HUGEPAGES);
1088 __mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1089 -hpage_nr_pages(page));
1090 } else {
1091 __dec_zone_page_state(page, NR_FILE_MAPPED);
1092 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
1093 mem_cgroup_end_update_page_stat(page, &locked, &flags);
1094 }
1095 if (unlikely(PageMlocked(page)))
1096 clear_page_mlock(page);
1097 /*
1098 * It would be tidy to reset the PageAnon mapping here,
1099 * but that might overwrite a racing page_add_anon_rmap
1100 * which increments mapcount after us but sets mapping
1101 * before us: so leave the reset to free_hot_cold_page,
1102 * and remember that it's only reliable while mapped.
1103 * Leaving it set also helps swapoff to reinstate ptes
1104 * faster for those pages still in swapcache.
1105 */
1106 return;
1107out:
1108 if (!anon)
1109 mem_cgroup_end_update_page_stat(page, &locked, &flags);
1110}
1111
1112/*
1113 * @arg: enum ttu_flags will be passed to this argument
1114 */
1115int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1116 unsigned long address, void *arg)
1117{
1118 struct mm_struct *mm = vma->vm_mm;
1119 pte_t *pte;
1120 pte_t pteval;
1121 spinlock_t *ptl;
1122 int ret = SWAP_AGAIN;
1123 enum ttu_flags flags = (enum ttu_flags)arg;
1124
1125 pte = page_check_address(page, mm, address, &ptl, 0);
1126 if (!pte)
1127 goto out;
1128
1129 /*
1130 * If the page is mlock()d, we cannot swap it out.
1131 * If it's recently referenced (perhaps page_referenced
1132 * skipped over this mm) then we should reactivate it.
1133 */
1134 if (!(flags & TTU_IGNORE_MLOCK)) {
1135 if (vma->vm_flags & VM_LOCKED)
1136 goto out_mlock;
1137
1138 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1139 goto out_unmap;
1140 }
1141 if (!(flags & TTU_IGNORE_ACCESS)) {
1142 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1143 ret = SWAP_FAIL;
1144 goto out_unmap;
1145 }
1146 }
1147
1148 /* Nuke the page table entry. */
1149 flush_cache_page(vma, address, page_to_pfn(page));
1150 pteval = ptep_clear_flush(vma, address, pte);
1151
1152 /* Move the dirty bit to the physical page now the pte is gone. */
1153 if (pte_dirty(pteval))
1154 set_page_dirty(page);
1155
1156 /* Update high watermark before we lower rss */
1157 update_hiwater_rss(mm);
1158
1159 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1160 if (!PageHuge(page)) {
1161 if (PageAnon(page))
1162 dec_mm_counter(mm, MM_ANONPAGES);
1163 else
1164 dec_mm_counter(mm, MM_FILEPAGES);
1165 }
1166 set_pte_at(mm, address, pte,
1167 swp_entry_to_pte(make_hwpoison_entry(page)));
1168 } else if (pte_unused(pteval)) {
1169 /*
1170 * The guest indicated that the page content is of no
1171 * interest anymore. Simply discard the pte, vmscan
1172 * will take care of the rest.
1173 */
1174 if (PageAnon(page))
1175 dec_mm_counter(mm, MM_ANONPAGES);
1176 else
1177 dec_mm_counter(mm, MM_FILEPAGES);
1178 } else if (PageAnon(page)) {
1179 swp_entry_t entry = { .val = page_private(page) };
1180 pte_t swp_pte;
1181
1182 if (PageSwapCache(page)) {
1183 /*
1184 * Store the swap location in the pte.
1185 * See handle_pte_fault() ...
1186 */
1187 if (swap_duplicate(entry) < 0) {
1188 set_pte_at(mm, address, pte, pteval);
1189 ret = SWAP_FAIL;
1190 goto out_unmap;
1191 }
1192 if (list_empty(&mm->mmlist)) {
1193 spin_lock(&mmlist_lock);
1194 if (list_empty(&mm->mmlist))
1195 list_add(&mm->mmlist, &init_mm.mmlist);
1196 spin_unlock(&mmlist_lock);
1197 }
1198 dec_mm_counter(mm, MM_ANONPAGES);
1199 inc_mm_counter(mm, MM_SWAPENTS);
1200 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
1201 /*
1202 * Store the pfn of the page in a special migration
1203 * pte. do_swap_page() will wait until the migration
1204 * pte is removed and then restart fault handling.
1205 */
1206 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1207 entry = make_migration_entry(page, pte_write(pteval));
1208 }
1209 swp_pte = swp_entry_to_pte(entry);
1210 if (pte_soft_dirty(pteval))
1211 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1212 set_pte_at(mm, address, pte, swp_pte);
1213 BUG_ON(pte_file(*pte));
1214 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1215 (TTU_ACTION(flags) == TTU_MIGRATION)) {
1216 /* Establish migration entry for a file page */
1217 swp_entry_t entry;
1218 entry = make_migration_entry(page, pte_write(pteval));
1219 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1220 } else
1221 dec_mm_counter(mm, MM_FILEPAGES);
1222
1223 page_remove_rmap(page);
1224 page_cache_release(page);
1225
1226out_unmap:
1227 pte_unmap_unlock(pte, ptl);
1228 if (ret != SWAP_FAIL)
1229 mmu_notifier_invalidate_page(mm, address);
1230out:
1231 return ret;
1232
1233out_mlock:
1234 pte_unmap_unlock(pte, ptl);
1235
1236
1237 /*
1238 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1239 * unstable result and race. Plus, We can't wait here because
1240 * we now hold anon_vma->rwsem or mapping->i_mmap_mutex.
1241 * if trylock failed, the page remain in evictable lru and later
1242 * vmscan could retry to move the page to unevictable lru if the
1243 * page is actually mlocked.
1244 */
1245 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1246 if (vma->vm_flags & VM_LOCKED) {
1247 mlock_vma_page(page);
1248 ret = SWAP_MLOCK;
1249 }
1250 up_read(&vma->vm_mm->mmap_sem);
1251 }
1252 return ret;
1253}
1254
1255/*
1256 * objrmap doesn't work for nonlinear VMAs because the assumption that
1257 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1258 * Consequently, given a particular page and its ->index, we cannot locate the
1259 * ptes which are mapping that page without an exhaustive linear search.
1260 *
1261 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1262 * maps the file to which the target page belongs. The ->vm_private_data field
1263 * holds the current cursor into that scan. Successive searches will circulate
1264 * around the vma's virtual address space.
1265 *
1266 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1267 * more scanning pressure is placed against them as well. Eventually pages
1268 * will become fully unmapped and are eligible for eviction.
1269 *
1270 * For very sparsely populated VMAs this is a little inefficient - chances are
1271 * there there won't be many ptes located within the scan cluster. In this case
1272 * maybe we could scan further - to the end of the pte page, perhaps.
1273 *
1274 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1275 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1276 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1277 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1278 */
1279#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1280#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1281
1282static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1283 struct vm_area_struct *vma, struct page *check_page)
1284{
1285 struct mm_struct *mm = vma->vm_mm;
1286 pmd_t *pmd;
1287 pte_t *pte;
1288 pte_t pteval;
1289 spinlock_t *ptl;
1290 struct page *page;
1291 unsigned long address;
1292 unsigned long mmun_start; /* For mmu_notifiers */
1293 unsigned long mmun_end; /* For mmu_notifiers */
1294 unsigned long end;
1295 int ret = SWAP_AGAIN;
1296 int locked_vma = 0;
1297
1298 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1299 end = address + CLUSTER_SIZE;
1300 if (address < vma->vm_start)
1301 address = vma->vm_start;
1302 if (end > vma->vm_end)
1303 end = vma->vm_end;
1304
1305 pmd = mm_find_pmd(mm, address);
1306 if (!pmd)
1307 return ret;
1308
1309 mmun_start = address;
1310 mmun_end = end;
1311 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1312
1313 /*
1314 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1315 * keep the sem while scanning the cluster for mlocking pages.
1316 */
1317 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1318 locked_vma = (vma->vm_flags & VM_LOCKED);
1319 if (!locked_vma)
1320 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1321 }
1322
1323 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1324
1325 /* Update high watermark before we lower rss */
1326 update_hiwater_rss(mm);
1327
1328 for (; address < end; pte++, address += PAGE_SIZE) {
1329 if (!pte_present(*pte))
1330 continue;
1331 page = vm_normal_page(vma, address, *pte);
1332 BUG_ON(!page || PageAnon(page));
1333
1334 if (locked_vma) {
1335 if (page == check_page) {
1336 /* we know we have check_page locked */
1337 mlock_vma_page(page);
1338 ret = SWAP_MLOCK;
1339 } else if (trylock_page(page)) {
1340 /*
1341 * If we can lock the page, perform mlock.
1342 * Otherwise leave the page alone, it will be
1343 * eventually encountered again later.
1344 */
1345 mlock_vma_page(page);
1346 unlock_page(page);
1347 }
1348 continue; /* don't unmap */
1349 }
1350
1351 if (ptep_clear_flush_young_notify(vma, address, pte))
1352 continue;
1353
1354 /* Nuke the page table entry. */
1355 flush_cache_page(vma, address, pte_pfn(*pte));
1356 pteval = ptep_clear_flush(vma, address, pte);
1357
1358 /* If nonlinear, store the file page offset in the pte. */
1359 if (page->index != linear_page_index(vma, address)) {
1360 pte_t ptfile = pgoff_to_pte(page->index);
1361 if (pte_soft_dirty(pteval))
1362 pte_file_mksoft_dirty(ptfile);
1363 set_pte_at(mm, address, pte, ptfile);
1364 }
1365
1366 /* Move the dirty bit to the physical page now the pte is gone. */
1367 if (pte_dirty(pteval))
1368 set_page_dirty(page);
1369
1370 page_remove_rmap(page);
1371 page_cache_release(page);
1372 dec_mm_counter(mm, MM_FILEPAGES);
1373 (*mapcount)--;
1374 }
1375 pte_unmap_unlock(pte - 1, ptl);
1376 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1377 if (locked_vma)
1378 up_read(&vma->vm_mm->mmap_sem);
1379 return ret;
1380}
1381
1382static int try_to_unmap_nonlinear(struct page *page,
1383 struct address_space *mapping, void *arg)
1384{
1385 struct vm_area_struct *vma;
1386 int ret = SWAP_AGAIN;
1387 unsigned long cursor;
1388 unsigned long max_nl_cursor = 0;
1389 unsigned long max_nl_size = 0;
1390 unsigned int mapcount;
1391
1392 list_for_each_entry(vma,
1393 &mapping->i_mmap_nonlinear, shared.nonlinear) {
1394
1395 cursor = (unsigned long) vma->vm_private_data;
1396 if (cursor > max_nl_cursor)
1397 max_nl_cursor = cursor;
1398 cursor = vma->vm_end - vma->vm_start;
1399 if (cursor > max_nl_size)
1400 max_nl_size = cursor;
1401 }
1402
1403 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1404 return SWAP_FAIL;
1405 }
1406
1407 /*
1408 * We don't try to search for this page in the nonlinear vmas,
1409 * and page_referenced wouldn't have found it anyway. Instead
1410 * just walk the nonlinear vmas trying to age and unmap some.
1411 * The mapcount of the page we came in with is irrelevant,
1412 * but even so use it as a guide to how hard we should try?
1413 */
1414 mapcount = page_mapcount(page);
1415 if (!mapcount)
1416 return ret;
1417
1418 cond_resched();
1419
1420 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1421 if (max_nl_cursor == 0)
1422 max_nl_cursor = CLUSTER_SIZE;
1423
1424 do {
1425 list_for_each_entry(vma,
1426 &mapping->i_mmap_nonlinear, shared.nonlinear) {
1427
1428 cursor = (unsigned long) vma->vm_private_data;
1429 while (cursor < max_nl_cursor &&
1430 cursor < vma->vm_end - vma->vm_start) {
1431 if (try_to_unmap_cluster(cursor, &mapcount,
1432 vma, page) == SWAP_MLOCK)
1433 ret = SWAP_MLOCK;
1434 cursor += CLUSTER_SIZE;
1435 vma->vm_private_data = (void *) cursor;
1436 if ((int)mapcount <= 0)
1437 return ret;
1438 }
1439 vma->vm_private_data = (void *) max_nl_cursor;
1440 }
1441 cond_resched();
1442 max_nl_cursor += CLUSTER_SIZE;
1443 } while (max_nl_cursor <= max_nl_size);
1444
1445 /*
1446 * Don't loop forever (perhaps all the remaining pages are
1447 * in locked vmas). Reset cursor on all unreserved nonlinear
1448 * vmas, now forgetting on which ones it had fallen behind.
1449 */
1450 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.nonlinear)
1451 vma->vm_private_data = NULL;
1452
1453 return ret;
1454}
1455
1456bool is_vma_temporary_stack(struct vm_area_struct *vma)
1457{
1458 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1459
1460 if (!maybe_stack)
1461 return false;
1462
1463 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1464 VM_STACK_INCOMPLETE_SETUP)
1465 return true;
1466
1467 return false;
1468}
1469
1470static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1471{
1472 return is_vma_temporary_stack(vma);
1473}
1474
1475static int page_not_mapped(struct page *page)
1476{
1477 return !page_mapped(page);
1478};
1479
1480/**
1481 * try_to_unmap - try to remove all page table mappings to a page
1482 * @page: the page to get unmapped
1483 * @flags: action and flags
1484 *
1485 * Tries to remove all the page table entries which are mapping this
1486 * page, used in the pageout path. Caller must hold the page lock.
1487 * Return values are:
1488 *
1489 * SWAP_SUCCESS - we succeeded in removing all mappings
1490 * SWAP_AGAIN - we missed a mapping, try again later
1491 * SWAP_FAIL - the page is unswappable
1492 * SWAP_MLOCK - page is mlocked.
1493 */
1494int try_to_unmap(struct page *page, enum ttu_flags flags)
1495{
1496 int ret;
1497 struct rmap_walk_control rwc = {
1498 .rmap_one = try_to_unmap_one,
1499 .arg = (void *)flags,
1500 .done = page_not_mapped,
1501 .file_nonlinear = try_to_unmap_nonlinear,
1502 .anon_lock = page_lock_anon_vma_read,
1503 };
1504
1505 VM_BUG_ON_PAGE(!PageHuge(page) && PageTransHuge(page), page);
1506
1507 /*
1508 * During exec, a temporary VMA is setup and later moved.
1509 * The VMA is moved under the anon_vma lock but not the
1510 * page tables leading to a race where migration cannot
1511 * find the migration ptes. Rather than increasing the
1512 * locking requirements of exec(), migration skips
1513 * temporary VMAs until after exec() completes.
1514 */
1515 if (flags & TTU_MIGRATION && !PageKsm(page) && PageAnon(page))
1516 rwc.invalid_vma = invalid_migration_vma;
1517
1518 ret = rmap_walk(page, &rwc);
1519
1520 if (ret != SWAP_MLOCK && !page_mapped(page))
1521 ret = SWAP_SUCCESS;
1522 return ret;
1523}
1524
1525/**
1526 * try_to_munlock - try to munlock a page
1527 * @page: the page to be munlocked
1528 *
1529 * Called from munlock code. Checks all of the VMAs mapping the page
1530 * to make sure nobody else has this page mlocked. The page will be
1531 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1532 *
1533 * Return values are:
1534 *
1535 * SWAP_AGAIN - no vma is holding page mlocked, or,
1536 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1537 * SWAP_FAIL - page cannot be located at present
1538 * SWAP_MLOCK - page is now mlocked.
1539 */
1540int try_to_munlock(struct page *page)
1541{
1542 int ret;
1543 struct rmap_walk_control rwc = {
1544 .rmap_one = try_to_unmap_one,
1545 .arg = (void *)TTU_MUNLOCK,
1546 .done = page_not_mapped,
1547 /*
1548 * We don't bother to try to find the munlocked page in
1549 * nonlinears. It's costly. Instead, later, page reclaim logic
1550 * may call try_to_unmap() and recover PG_mlocked lazily.
1551 */
1552 .file_nonlinear = NULL,
1553 .anon_lock = page_lock_anon_vma_read,
1554
1555 };
1556
1557 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1558
1559 ret = rmap_walk(page, &rwc);
1560 return ret;
1561}
1562
1563void __put_anon_vma(struct anon_vma *anon_vma)
1564{
1565 struct anon_vma *root = anon_vma->root;
1566
1567 anon_vma_free(anon_vma);
1568 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1569 anon_vma_free(root);
1570}
1571
1572static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1573 struct rmap_walk_control *rwc)
1574{
1575 struct anon_vma *anon_vma;
1576
1577 if (rwc->anon_lock)
1578 return rwc->anon_lock(page);
1579
1580 /*
1581 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1582 * because that depends on page_mapped(); but not all its usages
1583 * are holding mmap_sem. Users without mmap_sem are required to
1584 * take a reference count to prevent the anon_vma disappearing
1585 */
1586 anon_vma = page_anon_vma(page);
1587 if (!anon_vma)
1588 return NULL;
1589
1590 anon_vma_lock_read(anon_vma);
1591 return anon_vma;
1592}
1593
1594/*
1595 * rmap_walk_anon - do something to anonymous page using the object-based
1596 * rmap method
1597 * @page: the page to be handled
1598 * @rwc: control variable according to each walk type
1599 *
1600 * Find all the mappings of a page using the mapping pointer and the vma chains
1601 * contained in the anon_vma struct it points to.
1602 *
1603 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1604 * where the page was found will be held for write. So, we won't recheck
1605 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1606 * LOCKED.
1607 */
1608static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc)
1609{
1610 struct anon_vma *anon_vma;
1611 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1612 struct anon_vma_chain *avc;
1613 int ret = SWAP_AGAIN;
1614
1615 anon_vma = rmap_walk_anon_lock(page, rwc);
1616 if (!anon_vma)
1617 return ret;
1618
1619 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1620 struct vm_area_struct *vma = avc->vma;
1621 unsigned long address = vma_address(page, vma);
1622
1623 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1624 continue;
1625
1626 ret = rwc->rmap_one(page, vma, address, rwc->arg);
1627 if (ret != SWAP_AGAIN)
1628 break;
1629 if (rwc->done && rwc->done(page))
1630 break;
1631 }
1632 anon_vma_unlock_read(anon_vma);
1633 return ret;
1634}
1635
1636/*
1637 * rmap_walk_file - do something to file page using the object-based rmap method
1638 * @page: the page to be handled
1639 * @rwc: control variable according to each walk type
1640 *
1641 * Find all the mappings of a page using the mapping pointer and the vma chains
1642 * contained in the address_space struct it points to.
1643 *
1644 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1645 * where the page was found will be held for write. So, we won't recheck
1646 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1647 * LOCKED.
1648 */
1649static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc)
1650{
1651 struct address_space *mapping = page->mapping;
1652 pgoff_t pgoff = page->index << compound_order(page);
1653 struct vm_area_struct *vma;
1654 int ret = SWAP_AGAIN;
1655
1656 /*
1657 * The page lock not only makes sure that page->mapping cannot
1658 * suddenly be NULLified by truncation, it makes sure that the
1659 * structure at mapping cannot be freed and reused yet,
1660 * so we can safely take mapping->i_mmap_mutex.
1661 */
1662 VM_BUG_ON(!PageLocked(page));
1663
1664 if (!mapping)
1665 return ret;
1666 mutex_lock(&mapping->i_mmap_mutex);
1667 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1668 unsigned long address = vma_address(page, vma);
1669
1670 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1671 continue;
1672
1673 ret = rwc->rmap_one(page, vma, address, rwc->arg);
1674 if (ret != SWAP_AGAIN)
1675 goto done;
1676 if (rwc->done && rwc->done(page))
1677 goto done;
1678 }
1679
1680 if (!rwc->file_nonlinear)
1681 goto done;
1682
1683 if (list_empty(&mapping->i_mmap_nonlinear))
1684 goto done;
1685
1686 ret = rwc->file_nonlinear(page, mapping, rwc->arg);
1687
1688done:
1689 mutex_unlock(&mapping->i_mmap_mutex);
1690 return ret;
1691}
1692
1693int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1694{
1695 if (unlikely(PageKsm(page)))
1696 return rmap_walk_ksm(page, rwc);
1697 else if (PageAnon(page))
1698 return rmap_walk_anon(page, rwc);
1699 else
1700 return rmap_walk_file(page, rwc);
1701}
1702
1703#ifdef CONFIG_HUGETLB_PAGE
1704/*
1705 * The following three functions are for anonymous (private mapped) hugepages.
1706 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1707 * and no lru code, because we handle hugepages differently from common pages.
1708 */
1709static void __hugepage_set_anon_rmap(struct page *page,
1710 struct vm_area_struct *vma, unsigned long address, int exclusive)
1711{
1712 struct anon_vma *anon_vma = vma->anon_vma;
1713
1714 BUG_ON(!anon_vma);
1715
1716 if (PageAnon(page))
1717 return;
1718 if (!exclusive)
1719 anon_vma = anon_vma->root;
1720
1721 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1722 page->mapping = (struct address_space *) anon_vma;
1723 page->index = linear_page_index(vma, address);
1724}
1725
1726void hugepage_add_anon_rmap(struct page *page,
1727 struct vm_area_struct *vma, unsigned long address)
1728{
1729 struct anon_vma *anon_vma = vma->anon_vma;
1730 int first;
1731
1732 BUG_ON(!PageLocked(page));
1733 BUG_ON(!anon_vma);
1734 /* address might be in next vma when migration races vma_adjust */
1735 first = atomic_inc_and_test(&page->_mapcount);
1736 if (first)
1737 __hugepage_set_anon_rmap(page, vma, address, 0);
1738}
1739
1740void hugepage_add_new_anon_rmap(struct page *page,
1741 struct vm_area_struct *vma, unsigned long address)
1742{
1743 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1744 atomic_set(&page->_mapcount, 0);
1745 __hugepage_set_anon_rmap(page, vma, address, 1);
1746}
1747#endif /* CONFIG_HUGETLB_PAGE */