summaryrefslogtreecommitdiff
path: root/mm/rmap.c (plain)
blob: a7276d8c96f33f95ef7743e6e7247ecb83c07861
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 VM_WARN_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 = &current->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 = &current->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 = &current->tlb_ubc;
615
616 cpumask_or(&tlb_ubc->cpumask, &tlb_ubc->cpumask, mm_cpumask(mm));
617 tlb_ubc->flush_required = true;
618
619 /*
620 * Ensure compiler does not re-order the setting of tlb_flush_batched
621 * before the PTE is cleared.
622 */
623 barrier();
624 mm->tlb_flush_batched = true;
625
626 /*
627 * If the PTE was dirty then it's best to assume it's writable. The
628 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
629 * before the page is queued for IO.
630 */
631 if (writable)
632 tlb_ubc->writable = true;
633}
634
635/*
636 * Returns true if the TLB flush should be deferred to the end of a batch of
637 * unmap operations to reduce IPIs.
638 */
639static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
640{
641 bool should_defer = false;
642
643 if (!(flags & TTU_BATCH_FLUSH))
644 return false;
645
646 /* If remote CPUs need to be flushed then defer batch the flush */
647 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
648 should_defer = true;
649 put_cpu();
650
651 return should_defer;
652}
653
654/*
655 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
656 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
657 * operation such as mprotect or munmap to race between reclaim unmapping
658 * the page and flushing the page. If this race occurs, it potentially allows
659 * access to data via a stale TLB entry. Tracking all mm's that have TLB
660 * batching in flight would be expensive during reclaim so instead track
661 * whether TLB batching occurred in the past and if so then do a flush here
662 * if required. This will cost one additional flush per reclaim cycle paid
663 * by the first operation at risk such as mprotect and mumap.
664 *
665 * This must be called under the PTL so that an access to tlb_flush_batched
666 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
667 * via the PTL.
668 */
669void flush_tlb_batched_pending(struct mm_struct *mm)
670{
671 if (mm->tlb_flush_batched) {
672 flush_tlb_mm(mm);
673
674 /*
675 * Do not allow the compiler to re-order the clearing of
676 * tlb_flush_batched before the tlb is flushed.
677 */
678 barrier();
679 mm->tlb_flush_batched = false;
680 }
681}
682#else
683static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
684 struct page *page, bool writable)
685{
686}
687
688static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
689{
690 return false;
691}
692#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
693
694/*
695 * At what user virtual address is page expected in vma?
696 * Caller should check the page is actually part of the vma.
697 */
698unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
699{
700 unsigned long address;
701 if (PageAnon(page)) {
702 struct anon_vma *page__anon_vma = page_anon_vma(page);
703 /*
704 * Note: swapoff's unuse_vma() is more efficient with this
705 * check, and needs it to match anon_vma when KSM is active.
706 */
707 if (!vma->anon_vma || !page__anon_vma ||
708 vma->anon_vma->root != page__anon_vma->root)
709 return -EFAULT;
710 } else if (page->mapping) {
711 if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
712 return -EFAULT;
713 } else
714 return -EFAULT;
715 address = __vma_address(page, vma);
716 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
717 return -EFAULT;
718 return address;
719}
720
721pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
722{
723 pgd_t *pgd;
724 pud_t *pud;
725 pmd_t *pmd = NULL;
726 pmd_t pmde;
727
728 pgd = pgd_offset(mm, address);
729 if (!pgd_present(*pgd))
730 goto out;
731
732 pud = pud_offset(pgd, address);
733 if (!pud_present(*pud))
734 goto out;
735
736 pmd = pmd_offset(pud, address);
737 /*
738 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
739 * without holding anon_vma lock for write. So when looking for a
740 * genuine pmde (in which to find pte), test present and !THP together.
741 */
742 pmde = *pmd;
743 barrier();
744 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
745 pmd = NULL;
746out:
747 return pmd;
748}
749
750/*
751 * Check that @page is mapped at @address into @mm.
752 *
753 * If @sync is false, page_check_address may perform a racy check to avoid
754 * the page table lock when the pte is not present (helpful when reclaiming
755 * highly shared pages).
756 *
757 * On success returns with pte mapped and locked.
758 */
759pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
760 unsigned long address, spinlock_t **ptlp, int sync)
761{
762 pmd_t *pmd;
763 pte_t *pte;
764 spinlock_t *ptl;
765
766 if (unlikely(PageHuge(page))) {
767 /* when pud is not present, pte will be NULL */
768 pte = huge_pte_offset(mm, address);
769 if (!pte)
770 return NULL;
771
772 ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
773 goto check;
774 }
775
776 pmd = mm_find_pmd(mm, address);
777 if (!pmd)
778 return NULL;
779
780 pte = pte_offset_map(pmd, address);
781 /* Make a quick check before getting the lock */
782 if (!sync && !pte_present(*pte)) {
783 pte_unmap(pte);
784 return NULL;
785 }
786
787 ptl = pte_lockptr(mm, pmd);
788check:
789 spin_lock(ptl);
790 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
791 *ptlp = ptl;
792 return pte;
793 }
794 pte_unmap_unlock(pte, ptl);
795 return NULL;
796}
797
798/**
799 * page_mapped_in_vma - check whether a page is really mapped in a VMA
800 * @page: the page to test
801 * @vma: the VMA to test
802 *
803 * Returns 1 if the page is mapped into the page tables of the VMA, 0
804 * if the page is not mapped into the page tables of this VMA. Only
805 * valid for normal file or anonymous VMAs.
806 */
807int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
808{
809 unsigned long address;
810 pte_t *pte;
811 spinlock_t *ptl;
812
813 address = __vma_address(page, vma);
814 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
815 return 0;
816 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
817 if (!pte) /* the page is not in this mm */
818 return 0;
819 pte_unmap_unlock(pte, ptl);
820
821 return 1;
822}
823
824#ifdef CONFIG_TRANSPARENT_HUGEPAGE
825/*
826 * Check that @page is mapped at @address into @mm. In contrast to
827 * page_check_address(), this function can handle transparent huge pages.
828 *
829 * On success returns true with pte mapped and locked. For PMD-mapped
830 * transparent huge pages *@ptep is set to NULL.
831 */
832bool page_check_address_transhuge(struct page *page, struct mm_struct *mm,
833 unsigned long address, pmd_t **pmdp,
834 pte_t **ptep, spinlock_t **ptlp)
835{
836 pgd_t *pgd;
837 pud_t *pud;
838 pmd_t *pmd;
839 pte_t *pte;
840 spinlock_t *ptl;
841
842 if (unlikely(PageHuge(page))) {
843 /* when pud is not present, pte will be NULL */
844 pte = huge_pte_offset(mm, address);
845 if (!pte)
846 return false;
847
848 ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
849 pmd = NULL;
850 goto check_pte;
851 }
852
853 pgd = pgd_offset(mm, address);
854 if (!pgd_present(*pgd))
855 return false;
856 pud = pud_offset(pgd, address);
857 if (!pud_present(*pud))
858 return false;
859 pmd = pmd_offset(pud, address);
860
861 if (pmd_trans_huge(*pmd)) {
862 ptl = pmd_lock(mm, pmd);
863 if (!pmd_present(*pmd))
864 goto unlock_pmd;
865 if (unlikely(!pmd_trans_huge(*pmd))) {
866 spin_unlock(ptl);
867 goto map_pte;
868 }
869
870 if (pmd_page(*pmd) != page)
871 goto unlock_pmd;
872
873 pte = NULL;
874 goto found;
875unlock_pmd:
876 spin_unlock(ptl);
877 return false;
878 } else {
879 pmd_t pmde = *pmd;
880
881 barrier();
882 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
883 return false;
884 }
885map_pte:
886 pte = pte_offset_map(pmd, address);
887 if (!pte_present(*pte)) {
888 pte_unmap(pte);
889 return false;
890 }
891
892 ptl = pte_lockptr(mm, pmd);
893check_pte:
894 spin_lock(ptl);
895
896 if (!pte_present(*pte)) {
897 pte_unmap_unlock(pte, ptl);
898 return false;
899 }
900
901 /* THP can be referenced by any subpage */
902 if (pte_pfn(*pte) - page_to_pfn(page) >= hpage_nr_pages(page)) {
903 pte_unmap_unlock(pte, ptl);
904 return false;
905 }
906found:
907 *ptep = pte;
908 *pmdp = pmd;
909 *ptlp = ptl;
910 return true;
911}
912#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
913
914struct page_referenced_arg {
915 int mapcount;
916 int referenced;
917 unsigned long vm_flags;
918 struct mem_cgroup *memcg;
919};
920/*
921 * arg: page_referenced_arg will be passed
922 */
923static int page_referenced_one(struct page *page, struct vm_area_struct *vma,
924 unsigned long address, void *arg)
925{
926 struct mm_struct *mm = vma->vm_mm;
927 struct page_referenced_arg *pra = arg;
928 pmd_t *pmd;
929 pte_t *pte;
930 spinlock_t *ptl;
931 int referenced = 0;
932
933 if (!page_check_address_transhuge(page, mm, address, &pmd, &pte, &ptl))
934 return SWAP_AGAIN;
935
936 if (vma->vm_flags & VM_LOCKED) {
937 if (pte)
938 pte_unmap(pte);
939 spin_unlock(ptl);
940 pra->vm_flags |= VM_LOCKED;
941 return SWAP_FAIL; /* To break the loop */
942 }
943
944 if (pte) {
945 if (ptep_clear_flush_young_notify(vma, address, pte)) {
946 /*
947 * Don't treat a reference through a sequentially read
948 * mapping as such. If the page has been used in
949 * another mapping, we will catch it; if this other
950 * mapping is already gone, the unmap path will have
951 * set PG_referenced or activated the page.
952 */
953 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
954 referenced++;
955 }
956 pte_unmap(pte);
957 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
958 if (pmdp_clear_flush_young_notify(vma, address, pmd))
959 referenced++;
960 } else {
961 /* unexpected pmd-mapped page? */
962 WARN_ON_ONCE(1);
963 }
964 spin_unlock(ptl);
965
966 if (referenced)
967 clear_page_idle(page);
968 if (test_and_clear_page_young(page))
969 referenced++;
970
971 if (referenced) {
972 pra->referenced++;
973 pra->vm_flags |= vma->vm_flags;
974 }
975
976 pra->mapcount--;
977 if (!pra->mapcount)
978 return SWAP_SUCCESS; /* To break the loop */
979
980 return SWAP_AGAIN;
981}
982
983static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
984{
985 struct page_referenced_arg *pra = arg;
986 struct mem_cgroup *memcg = pra->memcg;
987
988 if (!mm_match_cgroup(vma->vm_mm, memcg))
989 return true;
990
991 return false;
992}
993
994/**
995 * page_referenced - test if the page was referenced
996 * @page: the page to test
997 * @is_locked: caller holds lock on the page
998 * @memcg: target memory cgroup
999 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
1000 *
1001 * Quick test_and_clear_referenced for all mappings to a page,
1002 * returns the number of ptes which referenced the page.
1003 */
1004int page_referenced(struct page *page,
1005 int is_locked,
1006 struct mem_cgroup *memcg,
1007 unsigned long *vm_flags)
1008{
1009 int ret;
1010 int we_locked = 0;
1011 struct page_referenced_arg pra = {
1012 .mapcount = total_mapcount(page),
1013 .memcg = memcg,
1014 };
1015 struct rmap_walk_control rwc = {
1016 .rmap_one = page_referenced_one,
1017 .arg = (void *)&pra,
1018 .anon_lock = page_lock_anon_vma_read,
1019 };
1020
1021 *vm_flags = 0;
1022 if (!page_mapped(page))
1023 return 0;
1024
1025 if (!page_rmapping(page))
1026 return 0;
1027
1028 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
1029 we_locked = trylock_page(page);
1030 if (!we_locked)
1031 return 1;
1032 }
1033
1034 /*
1035 * If we are reclaiming on behalf of a cgroup, skip
1036 * counting on behalf of references from different
1037 * cgroups
1038 */
1039 if (memcg) {
1040 rwc.invalid_vma = invalid_page_referenced_vma;
1041 }
1042
1043 ret = rmap_walk(page, &rwc);
1044 *vm_flags = pra.vm_flags;
1045
1046 if (we_locked)
1047 unlock_page(page);
1048
1049 return pra.referenced;
1050}
1051
1052static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
1053 unsigned long address, void *arg)
1054{
1055 struct mm_struct *mm = vma->vm_mm;
1056 pte_t *pte;
1057 spinlock_t *ptl;
1058 int ret = 0;
1059 int *cleaned = arg;
1060
1061 pte = page_check_address(page, mm, address, &ptl, 1);
1062 if (!pte)
1063 goto out;
1064
1065 if (pte_dirty(*pte) || pte_write(*pte)) {
1066 pte_t entry;
1067
1068 flush_cache_page(vma, address, pte_pfn(*pte));
1069 entry = ptep_clear_flush(vma, address, pte);
1070 entry = pte_wrprotect(entry);
1071 entry = pte_mkclean(entry);
1072 set_pte_at(mm, address, pte, entry);
1073 ret = 1;
1074 }
1075
1076 pte_unmap_unlock(pte, ptl);
1077
1078 if (ret) {
1079 mmu_notifier_invalidate_page(mm, address);
1080 (*cleaned)++;
1081 }
1082out:
1083 return SWAP_AGAIN;
1084}
1085
1086static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1087{
1088 if (vma->vm_flags & VM_SHARED)
1089 return false;
1090
1091 return true;
1092}
1093
1094int page_mkclean(struct page *page)
1095{
1096 int cleaned = 0;
1097 struct address_space *mapping;
1098 struct rmap_walk_control rwc = {
1099 .arg = (void *)&cleaned,
1100 .rmap_one = page_mkclean_one,
1101 .invalid_vma = invalid_mkclean_vma,
1102 };
1103
1104 BUG_ON(!PageLocked(page));
1105
1106 if (!page_mapped(page))
1107 return 0;
1108
1109 mapping = page_mapping(page);
1110 if (!mapping)
1111 return 0;
1112
1113 rmap_walk(page, &rwc);
1114
1115 return cleaned;
1116}
1117EXPORT_SYMBOL_GPL(page_mkclean);
1118
1119/**
1120 * page_move_anon_rmap - move a page to our anon_vma
1121 * @page: the page to move to our anon_vma
1122 * @vma: the vma the page belongs to
1123 *
1124 * When a page belongs exclusively to one process after a COW event,
1125 * that page can be moved into the anon_vma that belongs to just that
1126 * process, so the rmap code will not search the parent or sibling
1127 * processes.
1128 */
1129void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1130{
1131 struct anon_vma *anon_vma = vma->anon_vma;
1132
1133 page = compound_head(page);
1134
1135 VM_BUG_ON_PAGE(!PageLocked(page), page);
1136 VM_BUG_ON_VMA(!anon_vma, vma);
1137
1138 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1139 /*
1140 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1141 * simultaneously, so a concurrent reader (eg page_referenced()'s
1142 * PageAnon()) will not see one without the other.
1143 */
1144 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1145}
1146
1147/**
1148 * __page_set_anon_rmap - set up new anonymous rmap
1149 * @page: Page to add to rmap
1150 * @vma: VM area to add page to.
1151 * @address: User virtual address of the mapping
1152 * @exclusive: the page is exclusively owned by the current process
1153 */
1154static void __page_set_anon_rmap(struct page *page,
1155 struct vm_area_struct *vma, unsigned long address, int exclusive)
1156{
1157 struct anon_vma *anon_vma = vma->anon_vma;
1158
1159 BUG_ON(!anon_vma);
1160
1161 if (PageAnon(page))
1162 return;
1163
1164 /*
1165 * If the page isn't exclusively mapped into this vma,
1166 * we must use the _oldest_ possible anon_vma for the
1167 * page mapping!
1168 */
1169 if (!exclusive)
1170 anon_vma = anon_vma->root;
1171
1172 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1173 page->mapping = (struct address_space *) anon_vma;
1174 page->index = linear_page_index(vma, address);
1175}
1176
1177/**
1178 * __page_check_anon_rmap - sanity check anonymous rmap addition
1179 * @page: the page to add the mapping to
1180 * @vma: the vm area in which the mapping is added
1181 * @address: the user virtual address mapped
1182 */
1183static void __page_check_anon_rmap(struct page *page,
1184 struct vm_area_struct *vma, unsigned long address)
1185{
1186#ifdef CONFIG_DEBUG_VM
1187 /*
1188 * The page's anon-rmap details (mapping and index) are guaranteed to
1189 * be set up correctly at this point.
1190 *
1191 * We have exclusion against page_add_anon_rmap because the caller
1192 * always holds the page locked, except if called from page_dup_rmap,
1193 * in which case the page is already known to be setup.
1194 *
1195 * We have exclusion against page_add_new_anon_rmap because those pages
1196 * are initially only visible via the pagetables, and the pte is locked
1197 * over the call to page_add_new_anon_rmap.
1198 */
1199 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1200 BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
1201#endif
1202}
1203
1204/**
1205 * page_add_anon_rmap - add pte mapping to an anonymous page
1206 * @page: the page to add the mapping to
1207 * @vma: the vm area in which the mapping is added
1208 * @address: the user virtual address mapped
1209 * @compound: charge the page as compound or small page
1210 *
1211 * The caller needs to hold the pte lock, and the page must be locked in
1212 * the anon_vma case: to serialize mapping,index checking after setting,
1213 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1214 * (but PageKsm is never downgraded to PageAnon).
1215 */
1216void page_add_anon_rmap(struct page *page,
1217 struct vm_area_struct *vma, unsigned long address, bool compound)
1218{
1219 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1220}
1221
1222/*
1223 * Special version of the above for do_swap_page, which often runs
1224 * into pages that are exclusively owned by the current process.
1225 * Everybody else should continue to use page_add_anon_rmap above.
1226 */
1227void do_page_add_anon_rmap(struct page *page,
1228 struct vm_area_struct *vma, unsigned long address, int flags)
1229{
1230 bool compound = flags & RMAP_COMPOUND;
1231 bool first;
1232
1233 if (compound) {
1234 atomic_t *mapcount;
1235 VM_BUG_ON_PAGE(!PageLocked(page), page);
1236 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1237 mapcount = compound_mapcount_ptr(page);
1238 first = atomic_inc_and_test(mapcount);
1239 } else {
1240 first = atomic_inc_and_test(&page->_mapcount);
1241 }
1242
1243 if (first) {
1244 int nr = compound ? hpage_nr_pages(page) : 1;
1245 /*
1246 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1247 * these counters are not modified in interrupt context, and
1248 * pte lock(a spinlock) is held, which implies preemption
1249 * disabled.
1250 */
1251 if (compound)
1252 __inc_node_page_state(page, NR_ANON_THPS);
1253 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1254 }
1255 if (unlikely(PageKsm(page)))
1256 return;
1257
1258 VM_BUG_ON_PAGE(!PageLocked(page), page);
1259
1260 /* address might be in next vma when migration races vma_adjust */
1261 if (first)
1262 __page_set_anon_rmap(page, vma, address,
1263 flags & RMAP_EXCLUSIVE);
1264 else
1265 __page_check_anon_rmap(page, vma, address);
1266}
1267
1268/**
1269 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1270 * @page: the page to add the mapping to
1271 * @vma: the vm area in which the mapping is added
1272 * @address: the user virtual address mapped
1273 * @compound: charge the page as compound or small page
1274 *
1275 * Same as page_add_anon_rmap but must only be called on *new* pages.
1276 * This means the inc-and-test can be bypassed.
1277 * Page does not have to be locked.
1278 */
1279void page_add_new_anon_rmap(struct page *page,
1280 struct vm_area_struct *vma, unsigned long address, bool compound)
1281{
1282 int nr = compound ? hpage_nr_pages(page) : 1;
1283
1284 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1285 __SetPageSwapBacked(page);
1286 if (compound) {
1287 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1288 /* increment count (starts at -1) */
1289 atomic_set(compound_mapcount_ptr(page), 0);
1290 __inc_node_page_state(page, NR_ANON_THPS);
1291 } else {
1292 /* Anon THP always mapped first with PMD */
1293 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1294 /* increment count (starts at -1) */
1295 atomic_set(&page->_mapcount, 0);
1296 }
1297 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1298 __page_set_anon_rmap(page, vma, address, 1);
1299}
1300
1301/**
1302 * page_add_file_rmap - add pte mapping to a file page
1303 * @page: the page to add the mapping to
1304 *
1305 * The caller needs to hold the pte lock.
1306 */
1307void page_add_file_rmap(struct page *page, bool compound)
1308{
1309 int i, nr = 1;
1310
1311 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1312 lock_page_memcg(page);
1313 if (compound && PageTransHuge(page)) {
1314 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1315 if (atomic_inc_and_test(&page[i]._mapcount))
1316 nr++;
1317 }
1318 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1319 goto out;
1320 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1321 __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1322 } else {
1323 if (PageTransCompound(page) && page_mapping(page)) {
1324 VM_WARN_ON_ONCE(!PageLocked(page));
1325
1326 SetPageDoubleMap(compound_head(page));
1327 if (PageMlocked(page))
1328 clear_page_mlock(compound_head(page));
1329 }
1330 if (!atomic_inc_and_test(&page->_mapcount))
1331 goto out;
1332 }
1333 __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, nr);
1334 mem_cgroup_update_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED, nr);
1335out:
1336 unlock_page_memcg(page);
1337}
1338
1339static void page_remove_file_rmap(struct page *page, bool compound)
1340{
1341 int i, nr = 1;
1342
1343 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1344 lock_page_memcg(page);
1345
1346 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1347 if (unlikely(PageHuge(page))) {
1348 /* hugetlb pages are always mapped with pmds */
1349 atomic_dec(compound_mapcount_ptr(page));
1350 goto out;
1351 }
1352
1353 /* page still mapped by someone else? */
1354 if (compound && PageTransHuge(page)) {
1355 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1356 if (atomic_add_negative(-1, &page[i]._mapcount))
1357 nr++;
1358 }
1359 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1360 goto out;
1361 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1362 __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1363 } else {
1364 if (!atomic_add_negative(-1, &page->_mapcount))
1365 goto out;
1366 }
1367
1368 /*
1369 * We use the irq-unsafe __{inc|mod}_zone_page_state because
1370 * these counters are not modified in interrupt context, and
1371 * pte lock(a spinlock) is held, which implies preemption disabled.
1372 */
1373 __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, -nr);
1374 mem_cgroup_update_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED, -nr);
1375
1376 if (unlikely(PageMlocked(page)))
1377 clear_page_mlock(page);
1378out:
1379 unlock_page_memcg(page);
1380}
1381
1382static void page_remove_anon_compound_rmap(struct page *page)
1383{
1384 int i, nr;
1385
1386 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1387 return;
1388
1389 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1390 if (unlikely(PageHuge(page)))
1391 return;
1392
1393 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1394 return;
1395
1396 __dec_node_page_state(page, NR_ANON_THPS);
1397
1398 if (TestClearPageDoubleMap(page)) {
1399 /*
1400 * Subpages can be mapped with PTEs too. Check how many of
1401 * themi are still mapped.
1402 */
1403 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1404 if (atomic_add_negative(-1, &page[i]._mapcount))
1405 nr++;
1406 }
1407 } else {
1408 nr = HPAGE_PMD_NR;
1409 }
1410
1411 if (unlikely(PageMlocked(page)))
1412 clear_page_mlock(page);
1413
1414 if (nr) {
1415 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
1416 deferred_split_huge_page(page);
1417 }
1418}
1419
1420/**
1421 * page_remove_rmap - take down pte mapping from a page
1422 * @page: page to remove mapping from
1423 * @compound: uncharge the page as compound or small page
1424 *
1425 * The caller needs to hold the pte lock.
1426 */
1427void page_remove_rmap(struct page *page, bool compound)
1428{
1429 if (!PageAnon(page))
1430 return page_remove_file_rmap(page, compound);
1431
1432 if (compound)
1433 return page_remove_anon_compound_rmap(page);
1434
1435 /* page still mapped by someone else? */
1436 if (!atomic_add_negative(-1, &page->_mapcount))
1437 return;
1438
1439 /*
1440 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1441 * these counters are not modified in interrupt context, and
1442 * pte lock(a spinlock) is held, which implies preemption disabled.
1443 */
1444 __dec_node_page_state(page, NR_ANON_MAPPED);
1445
1446 if (unlikely(PageMlocked(page)))
1447 clear_page_mlock(page);
1448
1449 if (PageTransCompound(page))
1450 deferred_split_huge_page(compound_head(page));
1451
1452 /*
1453 * It would be tidy to reset the PageAnon mapping here,
1454 * but that might overwrite a racing page_add_anon_rmap
1455 * which increments mapcount after us but sets mapping
1456 * before us: so leave the reset to free_hot_cold_page,
1457 * and remember that it's only reliable while mapped.
1458 * Leaving it set also helps swapoff to reinstate ptes
1459 * faster for those pages still in swapcache.
1460 */
1461}
1462
1463struct rmap_private {
1464 enum ttu_flags flags;
1465 int lazyfreed;
1466};
1467
1468/*
1469 * @arg: enum ttu_flags will be passed to this argument
1470 */
1471static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1472 unsigned long address, void *arg)
1473{
1474 struct mm_struct *mm = vma->vm_mm;
1475 pte_t *pte;
1476 pte_t pteval;
1477 spinlock_t *ptl;
1478 int ret = SWAP_AGAIN;
1479 unsigned long sh_address;
1480 bool pmd_sharing_possible = false;
1481 unsigned long spmd_start, spmd_end;
1482 struct rmap_private *rp = arg;
1483 enum ttu_flags flags = rp->flags;
1484
1485 /* munlock has nothing to gain from examining un-locked vmas */
1486 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1487 goto out;
1488
1489 if (flags & TTU_SPLIT_HUGE_PMD) {
1490 split_huge_pmd_address(vma, address,
1491 flags & TTU_MIGRATION, page);
1492 /* check if we have anything to do after split */
1493 if (page_mapcount(page) == 0)
1494 goto out;
1495 }
1496
1497 /*
1498 * Only use the range_start/end mmu notifiers if huge pmd sharing
1499 * is possible. In the normal case, mmu_notifier_invalidate_page
1500 * is sufficient as we only unmap a page. However, if we unshare
1501 * a pmd, we will unmap a PUD_SIZE range.
1502 */
1503 if (PageHuge(page)) {
1504 spmd_start = address;
1505 spmd_end = spmd_start + vma_mmu_pagesize(vma);
1506
1507 /*
1508 * Check if pmd sharing is possible. If possible, we could
1509 * unmap a PUD_SIZE range. spmd_start/spmd_end will be
1510 * modified if sharing is possible.
1511 */
1512 adjust_range_if_pmd_sharing_possible(vma, &spmd_start,
1513 &spmd_end);
1514 if (spmd_end - spmd_start != vma_mmu_pagesize(vma)) {
1515 sh_address = address;
1516
1517 pmd_sharing_possible = true;
1518 mmu_notifier_invalidate_range_start(vma->vm_mm,
1519 spmd_start, spmd_end);
1520 }
1521 }
1522
1523 pte = page_check_address(page, mm, address, &ptl,
1524 PageTransCompound(page));
1525 if (!pte)
1526 goto out;
1527
1528 /*
1529 * If the page is mlock()d, we cannot swap it out.
1530 * If it's recently referenced (perhaps page_referenced
1531 * skipped over this mm) then we should reactivate it.
1532 */
1533 if (!(flags & TTU_IGNORE_MLOCK)) {
1534 if (vma->vm_flags & VM_LOCKED) {
1535 /* PTE-mapped THP are never mlocked */
1536 if (!PageTransCompound(page)) {
1537 /*
1538 * Holding pte lock, we do *not* need
1539 * mmap_sem here
1540 */
1541 mlock_vma_page(page);
1542 }
1543 ret = SWAP_MLOCK;
1544 goto out_unmap;
1545 }
1546 if (flags & TTU_MUNLOCK)
1547 goto out_unmap;
1548 }
1549 if (!(flags & TTU_IGNORE_ACCESS)) {
1550 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1551 ret = SWAP_FAIL;
1552 goto out_unmap;
1553 }
1554 }
1555
1556 /*
1557 * Call huge_pmd_unshare to potentially unshare a huge pmd. Pass
1558 * sh_address as it will be modified if unsharing is successful.
1559 */
1560 if (PageHuge(page) && huge_pmd_unshare(mm, &sh_address, pte)) {
1561 /*
1562 * huge_pmd_unshare unmapped an entire PMD page. There is
1563 * no way of knowing exactly which PMDs may be cached for
1564 * this mm, so flush them all. spmd_start/spmd_end cover
1565 * this PUD_SIZE range.
1566 */
1567 flush_cache_range(vma, spmd_start, spmd_end);
1568 flush_tlb_range(vma, spmd_start, spmd_end);
1569
1570 /*
1571 * The ref count of the PMD page was dropped which is part
1572 * of the way map counting is done for shared PMDs. When
1573 * there is no other sharing, huge_pmd_unshare returns false
1574 * and we will unmap the actual page and drop map count
1575 * to zero.
1576 */
1577 goto out_unmap;
1578 }
1579
1580 /* Nuke the page table entry. */
1581 flush_cache_page(vma, address, page_to_pfn(page));
1582 if (should_defer_flush(mm, flags)) {
1583 /*
1584 * We clear the PTE but do not flush so potentially a remote
1585 * CPU could still be writing to the page. If the entry was
1586 * previously clean then the architecture must guarantee that
1587 * a clear->dirty transition on a cached TLB entry is written
1588 * through and traps if the PTE is unmapped.
1589 */
1590 pteval = ptep_get_and_clear(mm, address, pte);
1591
1592 set_tlb_ubc_flush_pending(mm, page, pte_dirty(pteval));
1593 } else {
1594 pteval = ptep_clear_flush(vma, address, pte);
1595 }
1596
1597 /* Move the dirty bit to the physical page now the pte is gone. */
1598 if (pte_dirty(pteval))
1599 set_page_dirty(page);
1600
1601 /* Update high watermark before we lower rss */
1602 update_hiwater_rss(mm);
1603
1604 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1605 if (PageHuge(page)) {
1606 hugetlb_count_sub(1 << compound_order(page), mm);
1607 } else {
1608 dec_mm_counter(mm, mm_counter(page));
1609 }
1610 set_pte_at(mm, address, pte,
1611 swp_entry_to_pte(make_hwpoison_entry(page)));
1612 } else if (pte_unused(pteval)) {
1613 /*
1614 * The guest indicated that the page content is of no
1615 * interest anymore. Simply discard the pte, vmscan
1616 * will take care of the rest.
1617 */
1618 dec_mm_counter(mm, mm_counter(page));
1619 } else if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION)) {
1620 swp_entry_t entry;
1621 pte_t swp_pte;
1622 /*
1623 * Store the pfn of the page in a special migration
1624 * pte. do_swap_page() will wait until the migration
1625 * pte is removed and then restart fault handling.
1626 */
1627 entry = make_migration_entry(page, pte_write(pteval));
1628 swp_pte = swp_entry_to_pte(entry);
1629 if (pte_soft_dirty(pteval))
1630 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1631 set_pte_at(mm, address, pte, swp_pte);
1632 } else if (PageAnon(page)) {
1633 swp_entry_t entry = { .val = page_private(page) };
1634 pte_t swp_pte;
1635 /*
1636 * Store the swap location in the pte.
1637 * See handle_pte_fault() ...
1638 */
1639 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
1640
1641 if (!PageDirty(page) && (flags & TTU_LZFREE)) {
1642 /* It's a freeable page by MADV_FREE */
1643 dec_mm_counter(mm, MM_ANONPAGES);
1644 rp->lazyfreed++;
1645 goto discard;
1646 }
1647
1648 if (swap_duplicate(entry) < 0) {
1649 set_pte_at(mm, address, pte, pteval);
1650 ret = SWAP_FAIL;
1651 goto out_unmap;
1652 }
1653 if (list_empty(&mm->mmlist)) {
1654 spin_lock(&mmlist_lock);
1655 if (list_empty(&mm->mmlist))
1656 list_add(&mm->mmlist, &init_mm.mmlist);
1657 spin_unlock(&mmlist_lock);
1658 }
1659 dec_mm_counter(mm, MM_ANONPAGES);
1660 inc_mm_counter(mm, MM_SWAPENTS);
1661 swp_pte = swp_entry_to_pte(entry);
1662 if (pte_soft_dirty(pteval))
1663 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1664 set_pte_at(mm, address, pte, swp_pte);
1665 } else
1666 dec_mm_counter(mm, mm_counter_file(page));
1667
1668discard:
1669 page_remove_rmap(page, PageHuge(page));
1670 put_page(page);
1671
1672out_unmap:
1673 pte_unmap_unlock(pte, ptl);
1674 if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK))
1675 mmu_notifier_invalidate_page(mm, address);
1676out:
1677 if (pmd_sharing_possible)
1678 mmu_notifier_invalidate_range_end(vma->vm_mm,
1679 spmd_start, spmd_end);
1680 return ret;
1681}
1682
1683bool is_vma_temporary_stack(struct vm_area_struct *vma)
1684{
1685 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1686
1687 if (!maybe_stack)
1688 return false;
1689
1690 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1691 VM_STACK_INCOMPLETE_SETUP)
1692 return true;
1693
1694 return false;
1695}
1696
1697static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1698{
1699 return is_vma_temporary_stack(vma);
1700}
1701
1702static int page_mapcount_is_zero(struct page *page)
1703{
1704 return !page_mapcount(page);
1705}
1706
1707/**
1708 * try_to_unmap - try to remove all page table mappings to a page
1709 * @page: the page to get unmapped
1710 * @flags: action and flags
1711 *
1712 * Tries to remove all the page table entries which are mapping this
1713 * page, used in the pageout path. Caller must hold the page lock.
1714 * Return values are:
1715 *
1716 * SWAP_SUCCESS - we succeeded in removing all mappings
1717 * SWAP_AGAIN - we missed a mapping, try again later
1718 * SWAP_FAIL - the page is unswappable
1719 * SWAP_MLOCK - page is mlocked.
1720 */
1721int try_to_unmap(struct page *page, enum ttu_flags flags)
1722{
1723 int ret;
1724 struct rmap_private rp = {
1725 .flags = flags,
1726 .lazyfreed = 0,
1727 };
1728
1729 struct rmap_walk_control rwc = {
1730 .rmap_one = try_to_unmap_one,
1731 .arg = &rp,
1732 .done = page_mapcount_is_zero,
1733 .anon_lock = page_lock_anon_vma_read,
1734 };
1735
1736 /*
1737 * During exec, a temporary VMA is setup and later moved.
1738 * The VMA is moved under the anon_vma lock but not the
1739 * page tables leading to a race where migration cannot
1740 * find the migration ptes. Rather than increasing the
1741 * locking requirements of exec(), migration skips
1742 * temporary VMAs until after exec() completes.
1743 */
1744 if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page))
1745 rwc.invalid_vma = invalid_migration_vma;
1746
1747 if (flags & TTU_RMAP_LOCKED)
1748 ret = rmap_walk_locked(page, &rwc);
1749 else
1750 ret = rmap_walk(page, &rwc);
1751
1752 if (ret != SWAP_MLOCK && !page_mapcount(page)) {
1753 ret = SWAP_SUCCESS;
1754 if (rp.lazyfreed && !PageDirty(page))
1755 ret = SWAP_LZFREE;
1756 }
1757 return ret;
1758}
1759
1760static int page_not_mapped(struct page *page)
1761{
1762 return !page_mapped(page);
1763};
1764
1765/**
1766 * try_to_munlock - try to munlock a page
1767 * @page: the page to be munlocked
1768 *
1769 * Called from munlock code. Checks all of the VMAs mapping the page
1770 * to make sure nobody else has this page mlocked. The page will be
1771 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1772 *
1773 * Return values are:
1774 *
1775 * SWAP_AGAIN - no vma is holding page mlocked, or,
1776 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
1777 * SWAP_FAIL - page cannot be located at present
1778 * SWAP_MLOCK - page is now mlocked.
1779 */
1780int try_to_munlock(struct page *page)
1781{
1782 int ret;
1783 struct rmap_private rp = {
1784 .flags = TTU_MUNLOCK,
1785 .lazyfreed = 0,
1786 };
1787
1788 struct rmap_walk_control rwc = {
1789 .rmap_one = try_to_unmap_one,
1790 .arg = &rp,
1791 .done = page_not_mapped,
1792 .anon_lock = page_lock_anon_vma_read,
1793
1794 };
1795
1796 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1797
1798 ret = rmap_walk(page, &rwc);
1799 return ret;
1800}
1801
1802void __put_anon_vma(struct anon_vma *anon_vma)
1803{
1804 struct anon_vma *root = anon_vma->root;
1805
1806 anon_vma_free(anon_vma);
1807 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1808 anon_vma_free(root);
1809}
1810
1811static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1812 struct rmap_walk_control *rwc)
1813{
1814 struct anon_vma *anon_vma;
1815
1816 if (rwc->anon_lock)
1817 return rwc->anon_lock(page);
1818
1819 /*
1820 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1821 * because that depends on page_mapped(); but not all its usages
1822 * are holding mmap_sem. Users without mmap_sem are required to
1823 * take a reference count to prevent the anon_vma disappearing
1824 */
1825 anon_vma = page_anon_vma(page);
1826 if (!anon_vma)
1827 return NULL;
1828
1829 anon_vma_lock_read(anon_vma);
1830 return anon_vma;
1831}
1832
1833/*
1834 * rmap_walk_anon - do something to anonymous page using the object-based
1835 * rmap method
1836 * @page: the page to be handled
1837 * @rwc: control variable according to each walk type
1838 *
1839 * Find all the mappings of a page using the mapping pointer and the vma chains
1840 * contained in the anon_vma struct it points to.
1841 *
1842 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1843 * where the page was found will be held for write. So, we won't recheck
1844 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1845 * LOCKED.
1846 */
1847static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1848 bool locked)
1849{
1850 struct anon_vma *anon_vma;
1851 pgoff_t pgoff;
1852 struct anon_vma_chain *avc;
1853 int ret = SWAP_AGAIN;
1854
1855 if (locked) {
1856 anon_vma = page_anon_vma(page);
1857 /* anon_vma disappear under us? */
1858 VM_BUG_ON_PAGE(!anon_vma, page);
1859 } else {
1860 anon_vma = rmap_walk_anon_lock(page, rwc);
1861 }
1862 if (!anon_vma)
1863 return ret;
1864
1865 pgoff = page_to_pgoff(page);
1866 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1867 struct vm_area_struct *vma = avc->vma;
1868 unsigned long address = vma_address(page, vma);
1869
1870 cond_resched();
1871
1872 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1873 continue;
1874
1875 ret = rwc->rmap_one(page, vma, address, rwc->arg);
1876 if (ret != SWAP_AGAIN)
1877 break;
1878 if (rwc->done && rwc->done(page))
1879 break;
1880 }
1881
1882 if (!locked)
1883 anon_vma_unlock_read(anon_vma);
1884 return ret;
1885}
1886
1887/*
1888 * rmap_walk_file - do something to file page using the object-based rmap method
1889 * @page: the page to be handled
1890 * @rwc: control variable according to each walk type
1891 *
1892 * Find all the mappings of a page using the mapping pointer and the vma chains
1893 * contained in the address_space struct it points to.
1894 *
1895 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1896 * where the page was found will be held for write. So, we won't recheck
1897 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1898 * LOCKED.
1899 */
1900static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1901 bool locked)
1902{
1903 struct address_space *mapping = page_mapping(page);
1904 pgoff_t pgoff;
1905 struct vm_area_struct *vma;
1906 int ret = SWAP_AGAIN;
1907
1908 /*
1909 * The page lock not only makes sure that page->mapping cannot
1910 * suddenly be NULLified by truncation, it makes sure that the
1911 * structure at mapping cannot be freed and reused yet,
1912 * so we can safely take mapping->i_mmap_rwsem.
1913 */
1914 VM_BUG_ON_PAGE(!PageLocked(page), page);
1915
1916 if (!mapping)
1917 return ret;
1918
1919 pgoff = page_to_pgoff(page);
1920 if (!locked)
1921 i_mmap_lock_read(mapping);
1922 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1923 unsigned long address = vma_address(page, vma);
1924
1925 cond_resched();
1926
1927 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1928 continue;
1929
1930 ret = rwc->rmap_one(page, vma, address, rwc->arg);
1931 if (ret != SWAP_AGAIN)
1932 goto done;
1933 if (rwc->done && rwc->done(page))
1934 goto done;
1935 }
1936
1937done:
1938 if (!locked)
1939 i_mmap_unlock_read(mapping);
1940 return ret;
1941}
1942
1943int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1944{
1945 if (unlikely(PageKsm(page)))
1946 return rmap_walk_ksm(page, rwc);
1947 else if (PageAnon(page))
1948 return rmap_walk_anon(page, rwc, false);
1949 else
1950 return rmap_walk_file(page, rwc, false);
1951}
1952
1953/* Like rmap_walk, but caller holds relevant rmap lock */
1954int rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1955{
1956 /* no ksm support for now */
1957 VM_BUG_ON_PAGE(PageKsm(page), page);
1958 if (PageAnon(page))
1959 return rmap_walk_anon(page, rwc, true);
1960 else
1961 return rmap_walk_file(page, rwc, true);
1962}
1963
1964#ifdef CONFIG_HUGETLB_PAGE
1965/*
1966 * The following three functions are for anonymous (private mapped) hugepages.
1967 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1968 * and no lru code, because we handle hugepages differently from common pages.
1969 */
1970static void __hugepage_set_anon_rmap(struct page *page,
1971 struct vm_area_struct *vma, unsigned long address, int exclusive)
1972{
1973 struct anon_vma *anon_vma = vma->anon_vma;
1974
1975 BUG_ON(!anon_vma);
1976
1977 if (PageAnon(page))
1978 return;
1979 if (!exclusive)
1980 anon_vma = anon_vma->root;
1981
1982 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1983 page->mapping = (struct address_space *) anon_vma;
1984 page->index = linear_page_index(vma, address);
1985}
1986
1987void hugepage_add_anon_rmap(struct page *page,
1988 struct vm_area_struct *vma, unsigned long address)
1989{
1990 struct anon_vma *anon_vma = vma->anon_vma;
1991 int first;
1992
1993 BUG_ON(!PageLocked(page));
1994 BUG_ON(!anon_vma);
1995 /* address might be in next vma when migration races vma_adjust */
1996 first = atomic_inc_and_test(compound_mapcount_ptr(page));
1997 if (first)
1998 __hugepage_set_anon_rmap(page, vma, address, 0);
1999}
2000
2001void hugepage_add_new_anon_rmap(struct page *page,
2002 struct vm_area_struct *vma, unsigned long address)
2003{
2004 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2005 atomic_set(compound_mapcount_ptr(page), 0);
2006 __hugepage_set_anon_rmap(page, vma, address, 1);
2007}
2008#endif /* CONFIG_HUGETLB_PAGE */
2009