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 * pgdat->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 * i_pages 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 * i_pages 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/sched/mm.h>
50 #include <linux/sched/task.h>
51 #include <linux/pagemap.h>
52 #include <linux/swap.h>
53 #include <linux/swapops.h>
54 #include <linux/slab.h>
55 #include <linux/init.h>
56 #include <linux/ksm.h>
57 #include <linux/rmap.h>
58 #include <linux/rcupdate.h>
59 #include <linux/export.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/migrate.h>
63 #include <linux/hugetlb.h>
64 #include <linux/huge_mm.h>
65 #include <linux/backing-dev.h>
66 #include <linux/page_idle.h>
67 #include <linux/memremap.h>
68 #include <linux/userfaultfd_k.h>
69
70 #include <asm/tlbflush.h>
71
72 #include <trace/events/tlb.h>
73
74 #include "internal.h"
75
76 static struct kmem_cache *anon_vma_cachep;
77 static struct kmem_cache *anon_vma_chain_cachep;
78
79 static inline struct anon_vma *anon_vma_alloc(void)
80 {
81 struct anon_vma *anon_vma;
82
83 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
84 if (anon_vma) {
85 atomic_set(&anon_vma->refcount, 1);
86 anon_vma->degree = 1; /* Reference for first vma */
87 anon_vma->parent = anon_vma;
88 /*
89 * Initialise the anon_vma root to point to itself. If called
90 * from fork, the root will be reset to the parents anon_vma.
91 */
92 anon_vma->root = anon_vma;
93 }
94
95 return anon_vma;
96 }
97
98 static inline void anon_vma_free(struct anon_vma *anon_vma)
99 {
100 VM_BUG_ON(atomic_read(&anon_vma->refcount));
101
102 /*
103 * Synchronize against page_lock_anon_vma_read() such that
104 * we can safely hold the lock without the anon_vma getting
105 * freed.
106 *
107 * Relies on the full mb implied by the atomic_dec_and_test() from
108 * put_anon_vma() against the acquire barrier implied by
109 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
110 *
111 * page_lock_anon_vma_read() VS put_anon_vma()
112 * down_read_trylock() atomic_dec_and_test()
113 * LOCK MB
114 * atomic_read() rwsem_is_locked()
115 *
116 * LOCK should suffice since the actual taking of the lock must
117 * happen _before_ what follows.
118 */
119 might_sleep();
120 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
121 anon_vma_lock_write(anon_vma);
122 anon_vma_unlock_write(anon_vma);
123 }
124
125 kmem_cache_free(anon_vma_cachep, anon_vma);
126 }
127
128 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
129 {
130 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
131 }
132
133 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
134 {
135 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
136 }
137
138 static void anon_vma_chain_link(struct vm_area_struct *vma,
139 struct anon_vma_chain *avc,
140 struct anon_vma *anon_vma)
141 {
142 avc->vma = vma;
143 avc->anon_vma = anon_vma;
144 list_add(&avc->same_vma, &vma->anon_vma_chain);
145 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
146 }
147
148 /**
149 * __anon_vma_prepare - attach an anon_vma to a memory region
150 * @vma: the memory region in question
151 *
152 * This makes sure the memory mapping described by 'vma' has
153 * an 'anon_vma' attached to it, so that we can associate the
154 * anonymous pages mapped into it with that anon_vma.
155 *
156 * The common case will be that we already have one, which
157 * is handled inline by anon_vma_prepare(). But if
158 * not we either need to find an adjacent mapping that we
159 * can re-use the anon_vma from (very common when the only
160 * reason for splitting a vma has been mprotect()), or we
161 * allocate a new one.
162 *
163 * Anon-vma allocations are very subtle, because we may have
164 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
165 * and that may actually touch the spinlock even in the newly
166 * allocated vma (it depends on RCU to make sure that the
167 * anon_vma isn't actually destroyed).
168 *
169 * As a result, we need to do proper anon_vma locking even
170 * for the new allocation. At the same time, we do not want
171 * to do any locking for the common case of already having
172 * an anon_vma.
173 *
174 * This must be called with the mmap_sem held for reading.
175 */
176 int __anon_vma_prepare(struct vm_area_struct *vma)
177 {
178 struct mm_struct *mm = vma->vm_mm;
179 struct anon_vma *anon_vma, *allocated;
180 struct anon_vma_chain *avc;
181
182 might_sleep();
183
184 avc = anon_vma_chain_alloc(GFP_KERNEL);
185 if (!avc)
186 goto out_enomem;
187
188 anon_vma = find_mergeable_anon_vma(vma);
189 allocated = NULL;
190 if (!anon_vma) {
191 anon_vma = anon_vma_alloc();
192 if (unlikely(!anon_vma))
193 goto out_enomem_free_avc;
194 allocated = anon_vma;
195 }
196
197 anon_vma_lock_write(anon_vma);
198 /* page_table_lock to protect against threads */
199 spin_lock(&mm->page_table_lock);
200 if (likely(!vma->anon_vma)) {
201 vma->anon_vma = anon_vma;
202 anon_vma_chain_link(vma, avc, anon_vma);
203 /* vma reference or self-parent link for new root */
204 anon_vma->degree++;
205 allocated = NULL;
206 avc = NULL;
207 }
208 spin_unlock(&mm->page_table_lock);
209 anon_vma_unlock_write(anon_vma);
210
211 if (unlikely(allocated))
212 put_anon_vma(allocated);
213 if (unlikely(avc))
214 anon_vma_chain_free(avc);
215
216 return 0;
217
218 out_enomem_free_avc:
219 anon_vma_chain_free(avc);
220 out_enomem:
221 return -ENOMEM;
222 }
223
224 /*
225 * This is a useful helper function for locking the anon_vma root as
226 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
227 * have the same vma.
228 *
229 * Such anon_vma's should have the same root, so you'd expect to see
230 * just a single mutex_lock for the whole traversal.
231 */
232 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
233 {
234 struct anon_vma *new_root = anon_vma->root;
235 if (new_root != root) {
236 if (WARN_ON_ONCE(root))
237 up_write(&root->rwsem);
238 root = new_root;
239 down_write(&root->rwsem);
240 }
241 return root;
242 }
243
244 static inline void unlock_anon_vma_root(struct anon_vma *root)
245 {
246 if (root)
247 up_write(&root->rwsem);
248 }
249
250 /*
251 * Attach the anon_vmas from src to dst.
252 * Returns 0 on success, -ENOMEM on failure.
253 *
254 * If dst->anon_vma is NULL this function tries to find and reuse existing
255 * anon_vma which has no vmas and only one child anon_vma. This prevents
256 * degradation of anon_vma hierarchy to endless linear chain in case of
257 * constantly forking task. On the other hand, an anon_vma with more than one
258 * child isn't reused even if there was no alive vma, thus rmap walker has a
259 * good chance of avoiding scanning the whole hierarchy when it searches where
260 * page is mapped.
261 */
262 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
263 {
264 struct anon_vma_chain *avc, *pavc;
265 struct anon_vma *root = NULL;
266
267 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
268 struct anon_vma *anon_vma;
269
270 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
271 if (unlikely(!avc)) {
272 unlock_anon_vma_root(root);
273 root = NULL;
274 avc = anon_vma_chain_alloc(GFP_KERNEL);
275 if (!avc)
276 goto enomem_failure;
277 }
278 anon_vma = pavc->anon_vma;
279 root = lock_anon_vma_root(root, anon_vma);
280 anon_vma_chain_link(dst, avc, anon_vma);
281
282 /*
283 * Reuse existing anon_vma if its degree lower than two,
284 * that means it has no vma and only one anon_vma child.
285 *
286 * Do not chose parent anon_vma, otherwise first child
287 * will always reuse it. Root anon_vma is never reused:
288 * it has self-parent reference and at least one child.
289 */
290 if (!dst->anon_vma && anon_vma != src->anon_vma &&
291 anon_vma->degree < 2)
292 dst->anon_vma = anon_vma;
293 }
294 if (dst->anon_vma)
295 dst->anon_vma->degree++;
296 unlock_anon_vma_root(root);
297 return 0;
298
299 enomem_failure:
300 /*
301 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
302 * decremented in unlink_anon_vmas().
303 * We can safely do this because callers of anon_vma_clone() don't care
304 * about dst->anon_vma if anon_vma_clone() failed.
305 */
306 dst->anon_vma = NULL;
307 unlink_anon_vmas(dst);
308 return -ENOMEM;
309 }
310
311 /*
312 * Attach vma to its own anon_vma, as well as to the anon_vmas that
313 * the corresponding VMA in the parent process is attached to.
314 * Returns 0 on success, non-zero on failure.
315 */
316 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
317 {
318 struct anon_vma_chain *avc;
319 struct anon_vma *anon_vma;
320 int error;
321
322 /* Don't bother if the parent process has no anon_vma here. */
323 if (!pvma->anon_vma)
324 return 0;
325
326 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
327 vma->anon_vma = NULL;
328
329 /*
330 * First, attach the new VMA to the parent VMA's anon_vmas,
331 * so rmap can find non-COWed pages in child processes.
332 */
333 error = anon_vma_clone(vma, pvma);
334 if (error)
335 return error;
336
337 /* An existing anon_vma has been reused, all done then. */
338 if (vma->anon_vma)
339 return 0;
340
341 /* Then add our own anon_vma. */
342 anon_vma = anon_vma_alloc();
343 if (!anon_vma)
344 goto out_error;
345 avc = anon_vma_chain_alloc(GFP_KERNEL);
346 if (!avc)
347 goto out_error_free_anon_vma;
348
349 /*
350 * The root anon_vma's spinlock is the lock actually used when we
351 * lock any of the anon_vmas in this anon_vma tree.
352 */
353 anon_vma->root = pvma->anon_vma->root;
354 anon_vma->parent = pvma->anon_vma;
355 /*
356 * With refcounts, an anon_vma can stay around longer than the
357 * process it belongs to. The root anon_vma needs to be pinned until
358 * this anon_vma is freed, because the lock lives in the root.
359 */
360 get_anon_vma(anon_vma->root);
361 /* Mark this anon_vma as the one where our new (COWed) pages go. */
362 vma->anon_vma = anon_vma;
363 anon_vma_lock_write(anon_vma);
364 anon_vma_chain_link(vma, avc, anon_vma);
365 anon_vma->parent->degree++;
366 anon_vma_unlock_write(anon_vma);
367
368 return 0;
369
370 out_error_free_anon_vma:
371 put_anon_vma(anon_vma);
372 out_error:
373 unlink_anon_vmas(vma);
374 return -ENOMEM;
375 }
376
377 void unlink_anon_vmas(struct vm_area_struct *vma)
378 {
379 struct anon_vma_chain *avc, *next;
380 struct anon_vma *root = NULL;
381
382 /*
383 * Unlink each anon_vma chained to the VMA. This list is ordered
384 * from newest to oldest, ensuring the root anon_vma gets freed last.
385 */
386 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
387 struct anon_vma *anon_vma = avc->anon_vma;
388
389 root = lock_anon_vma_root(root, anon_vma);
390 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
391
392 /*
393 * Leave empty anon_vmas on the list - we'll need
394 * to free them outside the lock.
395 */
396 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
397 anon_vma->parent->degree--;
398 continue;
399 }
400
401 list_del(&avc->same_vma);
402 anon_vma_chain_free(avc);
403 }
404 if (vma->anon_vma)
405 vma->anon_vma->degree--;
406 unlock_anon_vma_root(root);
407
408 /*
409 * Iterate the list once more, it now only contains empty and unlinked
410 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
411 * needing to write-acquire the anon_vma->root->rwsem.
412 */
413 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
414 struct anon_vma *anon_vma = avc->anon_vma;
415
416 VM_WARN_ON(anon_vma->degree);
417 put_anon_vma(anon_vma);
418
419 list_del(&avc->same_vma);
420 anon_vma_chain_free(avc);
421 }
422 }
423
424 static void anon_vma_ctor(void *data)
425 {
426 struct anon_vma *anon_vma = data;
427
428 init_rwsem(&anon_vma->rwsem);
429 atomic_set(&anon_vma->refcount, 0);
430 anon_vma->rb_root = RB_ROOT_CACHED;
431 }
432
433 void __init anon_vma_init(void)
434 {
435 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
436 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
437 anon_vma_ctor);
438 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
439 SLAB_PANIC|SLAB_ACCOUNT);
440 }
441
442 /*
443 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
444 *
445 * Since there is no serialization what so ever against page_remove_rmap()
446 * the best this function can do is return a locked anon_vma that might
447 * have been relevant to this page.
448 *
449 * The page might have been remapped to a different anon_vma or the anon_vma
450 * returned may already be freed (and even reused).
451 *
452 * In case it was remapped to a different anon_vma, the new anon_vma will be a
453 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
454 * ensure that any anon_vma obtained from the page will still be valid for as
455 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
456 *
457 * All users of this function must be very careful when walking the anon_vma
458 * chain and verify that the page in question is indeed mapped in it
459 * [ something equivalent to page_mapped_in_vma() ].
460 *
461 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
462 * that the anon_vma pointer from page->mapping is valid if there is a
463 * mapcount, we can dereference the anon_vma after observing those.
464 */
465 struct anon_vma *page_get_anon_vma(struct page *page)
466 {
467 struct anon_vma *anon_vma = NULL;
468 unsigned long anon_mapping;
469
470 rcu_read_lock();
471 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
472 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
473 goto out;
474 if (!page_mapped(page))
475 goto out;
476
477 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
478 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
479 anon_vma = NULL;
480 goto out;
481 }
482
483 /*
484 * If this page is still mapped, then its anon_vma cannot have been
485 * freed. But if it has been unmapped, we have no security against the
486 * anon_vma structure being freed and reused (for another anon_vma:
487 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
488 * above cannot corrupt).
489 */
490 if (!page_mapped(page)) {
491 rcu_read_unlock();
492 put_anon_vma(anon_vma);
493 return NULL;
494 }
495 out:
496 rcu_read_unlock();
497
498 return anon_vma;
499 }
500
501 /*
502 * Similar to page_get_anon_vma() except it locks the anon_vma.
503 *
504 * Its a little more complex as it tries to keep the fast path to a single
505 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
506 * reference like with page_get_anon_vma() and then block on the mutex.
507 */
508 struct anon_vma *page_lock_anon_vma_read(struct page *page)
509 {
510 struct anon_vma *anon_vma = NULL;
511 struct anon_vma *root_anon_vma;
512 unsigned long anon_mapping;
513
514 rcu_read_lock();
515 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
516 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
517 goto out;
518 if (!page_mapped(page))
519 goto out;
520
521 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
522 root_anon_vma = READ_ONCE(anon_vma->root);
523 if (down_read_trylock(&root_anon_vma->rwsem)) {
524 /*
525 * If the page is still mapped, then this anon_vma is still
526 * its anon_vma, and holding the mutex ensures that it will
527 * not go away, see anon_vma_free().
528 */
529 if (!page_mapped(page)) {
530 up_read(&root_anon_vma->rwsem);
531 anon_vma = NULL;
532 }
533 goto out;
534 }
535
536 /* trylock failed, we got to sleep */
537 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
538 anon_vma = NULL;
539 goto out;
540 }
541
542 if (!page_mapped(page)) {
543 rcu_read_unlock();
544 put_anon_vma(anon_vma);
545 return NULL;
546 }
547
548 /* we pinned the anon_vma, its safe to sleep */
549 rcu_read_unlock();
550 anon_vma_lock_read(anon_vma);
551
552 if (atomic_dec_and_test(&anon_vma->refcount)) {
553 /*
554 * Oops, we held the last refcount, release the lock
555 * and bail -- can't simply use put_anon_vma() because
556 * we'll deadlock on the anon_vma_lock_write() recursion.
557 */
558 anon_vma_unlock_read(anon_vma);
559 __put_anon_vma(anon_vma);
560 anon_vma = NULL;
561 }
562
563 return anon_vma;
564
565 out:
566 rcu_read_unlock();
567 return anon_vma;
568 }
569
570 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
571 {
572 anon_vma_unlock_read(anon_vma);
573 }
574
575 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
576 /*
577 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
578 * important if a PTE was dirty when it was unmapped that it's flushed
579 * before any IO is initiated on the page to prevent lost writes. Similarly,
580 * it must be flushed before freeing to prevent data leakage.
581 */
582 void try_to_unmap_flush(void)
583 {
584 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
585
586 if (!tlb_ubc->flush_required)
587 return;
588
589 arch_tlbbatch_flush(&tlb_ubc->arch);
590 tlb_ubc->flush_required = false;
591 tlb_ubc->writable = false;
592 }
593
594 /* Flush iff there are potentially writable TLB entries that can race with IO */
595 void try_to_unmap_flush_dirty(void)
596 {
597 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
598
599 if (tlb_ubc->writable)
600 try_to_unmap_flush();
601 }
602
603 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
604 {
605 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
606
607 arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
608 tlb_ubc->flush_required = true;
609
610 /*
611 * Ensure compiler does not re-order the setting of tlb_flush_batched
612 * before the PTE is cleared.
613 */
614 barrier();
615 mm->tlb_flush_batched = true;
616
617 /*
618 * If the PTE was dirty then it's best to assume it's writable. The
619 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
620 * before the page is queued for IO.
621 */
622 if (writable)
623 tlb_ubc->writable = true;
624 }
625
626 /*
627 * Returns true if the TLB flush should be deferred to the end of a batch of
628 * unmap operations to reduce IPIs.
629 */
630 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
631 {
632 bool should_defer = false;
633
634 if (!(flags & TTU_BATCH_FLUSH))
635 return false;
636
637 /* If remote CPUs need to be flushed then defer batch the flush */
638 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
639 should_defer = true;
640 put_cpu();
641
642 return should_defer;
643 }
644
645 /*
646 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
647 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
648 * operation such as mprotect or munmap to race between reclaim unmapping
649 * the page and flushing the page. If this race occurs, it potentially allows
650 * access to data via a stale TLB entry. Tracking all mm's that have TLB
651 * batching in flight would be expensive during reclaim so instead track
652 * whether TLB batching occurred in the past and if so then do a flush here
653 * if required. This will cost one additional flush per reclaim cycle paid
654 * by the first operation at risk such as mprotect and mumap.
655 *
656 * This must be called under the PTL so that an access to tlb_flush_batched
657 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
658 * via the PTL.
659 */
660 void flush_tlb_batched_pending(struct mm_struct *mm)
661 {
662 if (mm->tlb_flush_batched) {
663 flush_tlb_mm(mm);
664
665 /*
666 * Do not allow the compiler to re-order the clearing of
667 * tlb_flush_batched before the tlb is flushed.
668 */
669 barrier();
670 mm->tlb_flush_batched = false;
671 }
672 }
673 #else
674 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
675 {
676 }
677
678 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
679 {
680 return false;
681 }
682 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
683
684 /*
685 * At what user virtual address is page expected in vma?
686 * Caller should check the page is actually part of the vma.
687 */
688 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
689 {
690 unsigned long address;
691 if (PageAnon(page)) {
692 struct anon_vma *page__anon_vma = page_anon_vma(page);
693 /*
694 * Note: swapoff's unuse_vma() is more efficient with this
695 * check, and needs it to match anon_vma when KSM is active.
696 */
697 if (!vma->anon_vma || !page__anon_vma ||
698 vma->anon_vma->root != page__anon_vma->root)
699 return -EFAULT;
700 } else if (page->mapping) {
701 if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
702 return -EFAULT;
703 } else
704 return -EFAULT;
705 address = __vma_address(page, vma);
706 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
707 return -EFAULT;
708 return address;
709 }
710
711 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
712 {
713 pgd_t *pgd;
714 p4d_t *p4d;
715 pud_t *pud;
716 pmd_t *pmd = NULL;
717 pmd_t pmde;
718
719 pgd = pgd_offset(mm, address);
720 if (!pgd_present(*pgd))
721 goto out;
722
723 p4d = p4d_offset(pgd, address);
724 if (!p4d_present(*p4d))
725 goto out;
726
727 pud = pud_offset(p4d, address);
728 if (!pud_present(*pud))
729 goto out;
730
731 pmd = pmd_offset(pud, address);
732 /*
733 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
734 * without holding anon_vma lock for write. So when looking for a
735 * genuine pmde (in which to find pte), test present and !THP together.
736 */
737 pmde = *pmd;
738 barrier();
739 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
740 pmd = NULL;
741 out:
742 return pmd;
743 }
744
745 struct page_referenced_arg {
746 int mapcount;
747 int referenced;
748 unsigned long vm_flags;
749 struct mem_cgroup *memcg;
750 };
751 /*
752 * arg: page_referenced_arg will be passed
753 */
754 static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
755 unsigned long address, void *arg)
756 {
757 struct page_referenced_arg *pra = arg;
758 struct page_vma_mapped_walk pvmw = {
759 .page = page,
760 .vma = vma,
761 .address = address,
762 };
763 int referenced = 0;
764
765 while (page_vma_mapped_walk(&pvmw)) {
766 address = pvmw.address;
767
768 if (vma->vm_flags & VM_LOCKED) {
769 page_vma_mapped_walk_done(&pvmw);
770 pra->vm_flags |= VM_LOCKED;
771 return false; /* To break the loop */
772 }
773
774 if (pvmw.pte) {
775 if (ptep_clear_flush_young_notify(vma, address,
776 pvmw.pte)) {
777 /*
778 * Don't treat a reference through
779 * a sequentially read mapping as such.
780 * If the page has been used in another mapping,
781 * we will catch it; if this other mapping is
782 * already gone, the unmap path will have set
783 * PG_referenced or activated the page.
784 */
785 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
786 referenced++;
787 }
788 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
789 if (pmdp_clear_flush_young_notify(vma, address,
790 pvmw.pmd))
791 referenced++;
792 } else {
793 /* unexpected pmd-mapped page? */
794 WARN_ON_ONCE(1);
795 }
796
797 pra->mapcount--;
798 }
799
800 if (referenced)
801 clear_page_idle(page);
802 if (test_and_clear_page_young(page))
803 referenced++;
804
805 if (referenced) {
806 pra->referenced++;
807 pra->vm_flags |= vma->vm_flags;
808 }
809
810 if (!pra->mapcount)
811 return false; /* To break the loop */
812
813 return true;
814 }
815
816 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
817 {
818 struct page_referenced_arg *pra = arg;
819 struct mem_cgroup *memcg = pra->memcg;
820
821 if (!mm_match_cgroup(vma->vm_mm, memcg))
822 return true;
823
824 return false;
825 }
826
827 /**
828 * page_referenced - test if the page was referenced
829 * @page: the page to test
830 * @is_locked: caller holds lock on the page
831 * @memcg: target memory cgroup
832 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
833 *
834 * Quick test_and_clear_referenced for all mappings to a page,
835 * returns the number of ptes which referenced the page.
836 */
837 int page_referenced(struct page *page,
838 int is_locked,
839 struct mem_cgroup *memcg,
840 unsigned long *vm_flags)
841 {
842 int we_locked = 0;
843 struct page_referenced_arg pra = {
844 .mapcount = total_mapcount(page),
845 .memcg = memcg,
846 };
847 struct rmap_walk_control rwc = {
848 .rmap_one = page_referenced_one,
849 .arg = (void *)&pra,
850 .anon_lock = page_lock_anon_vma_read,
851 };
852
853 *vm_flags = 0;
854 if (!pra.mapcount)
855 return 0;
856
857 if (!page_rmapping(page))
858 return 0;
859
860 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
861 we_locked = trylock_page(page);
862 if (!we_locked)
863 return 1;
864 }
865
866 /*
867 * If we are reclaiming on behalf of a cgroup, skip
868 * counting on behalf of references from different
869 * cgroups
870 */
871 if (memcg) {
872 rwc.invalid_vma = invalid_page_referenced_vma;
873 }
874
875 rmap_walk(page, &rwc);
876 *vm_flags = pra.vm_flags;
877
878 if (we_locked)
879 unlock_page(page);
880
881 return pra.referenced;
882 }
883
884 static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
885 unsigned long address, void *arg)
886 {
887 struct page_vma_mapped_walk pvmw = {
888 .page = page,
889 .vma = vma,
890 .address = address,
891 .flags = PVMW_SYNC,
892 };
893 struct mmu_notifier_range range;
894 int *cleaned = arg;
895
896 /*
897 * We have to assume the worse case ie pmd for invalidation. Note that
898 * the page can not be free from this function.
899 */
900 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
901 0, vma, vma->vm_mm, address,
902 min(vma->vm_end, address + page_size(page)));
903 mmu_notifier_invalidate_range_start(&range);
904
905 while (page_vma_mapped_walk(&pvmw)) {
906 int ret = 0;
907
908 address = pvmw.address;
909 if (pvmw.pte) {
910 pte_t entry;
911 pte_t *pte = pvmw.pte;
912
913 if (!pte_dirty(*pte) && !pte_write(*pte))
914 continue;
915
916 flush_cache_page(vma, address, pte_pfn(*pte));
917 entry = ptep_clear_flush(vma, address, pte);
918 entry = pte_wrprotect(entry);
919 entry = pte_mkclean(entry);
920 set_pte_at(vma->vm_mm, address, pte, entry);
921 ret = 1;
922 } else {
923 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
924 pmd_t *pmd = pvmw.pmd;
925 pmd_t entry;
926
927 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
928 continue;
929
930 flush_cache_page(vma, address, page_to_pfn(page));
931 entry = pmdp_invalidate(vma, address, pmd);
932 entry = pmd_wrprotect(entry);
933 entry = pmd_mkclean(entry);
934 set_pmd_at(vma->vm_mm, address, pmd, entry);
935 ret = 1;
936 #else
937 /* unexpected pmd-mapped page? */
938 WARN_ON_ONCE(1);
939 #endif
940 }
941
942 /*
943 * No need to call mmu_notifier_invalidate_range() as we are
944 * downgrading page table protection not changing it to point
945 * to a new page.
946 *
947 * See Documentation/vm/mmu_notifier.rst
948 */
949 if (ret)
950 (*cleaned)++;
951 }
952
953 mmu_notifier_invalidate_range_end(&range);
954
955 return true;
956 }
957
958 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
959 {
960 if (vma->vm_flags & VM_SHARED)
961 return false;
962
963 return true;
964 }
965
966 int page_mkclean(struct page *page)
967 {
968 int cleaned = 0;
969 struct address_space *mapping;
970 struct rmap_walk_control rwc = {
971 .arg = (void *)&cleaned,
972 .rmap_one = page_mkclean_one,
973 .invalid_vma = invalid_mkclean_vma,
974 };
975
976 BUG_ON(!PageLocked(page));
977
978 if (!page_mapped(page))
979 return 0;
980
981 mapping = page_mapping(page);
982 if (!mapping)
983 return 0;
984
985 rmap_walk(page, &rwc);
986
987 return cleaned;
988 }
989 EXPORT_SYMBOL_GPL(page_mkclean);
990
991 /**
992 * page_move_anon_rmap - move a page to our anon_vma
993 * @page: the page to move to our anon_vma
994 * @vma: the vma the page belongs to
995 *
996 * When a page belongs exclusively to one process after a COW event,
997 * that page can be moved into the anon_vma that belongs to just that
998 * process, so the rmap code will not search the parent or sibling
999 * processes.
1000 */
1001 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1002 {
1003 struct anon_vma *anon_vma = vma->anon_vma;
1004
1005 page = compound_head(page);
1006
1007 VM_BUG_ON_PAGE(!PageLocked(page), page);
1008 VM_BUG_ON_VMA(!anon_vma, vma);
1009
1010 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1011 /*
1012 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1013 * simultaneously, so a concurrent reader (eg page_referenced()'s
1014 * PageAnon()) will not see one without the other.
1015 */
1016 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1017 }
1018
1019 /**
1020 * __page_set_anon_rmap - set up new anonymous rmap
1021 * @page: Page or Hugepage to add to rmap
1022 * @vma: VM area to add page to.
1023 * @address: User virtual address of the mapping
1024 * @exclusive: the page is exclusively owned by the current process
1025 */
1026 static void __page_set_anon_rmap(struct page *page,
1027 struct vm_area_struct *vma, unsigned long address, int exclusive)
1028 {
1029 struct anon_vma *anon_vma = vma->anon_vma;
1030
1031 BUG_ON(!anon_vma);
1032
1033 if (PageAnon(page))
1034 return;
1035
1036 /*
1037 * If the page isn't exclusively mapped into this vma,
1038 * we must use the _oldest_ possible anon_vma for the
1039 * page mapping!
1040 */
1041 if (!exclusive)
1042 anon_vma = anon_vma->root;
1043
1044 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1045 page->mapping = (struct address_space *) anon_vma;
1046 page->index = linear_page_index(vma, address);
1047 }
1048
1049 /**
1050 * __page_check_anon_rmap - sanity check anonymous rmap addition
1051 * @page: the page to add the mapping to
1052 * @vma: the vm area in which the mapping is added
1053 * @address: the user virtual address mapped
1054 */
1055 static void __page_check_anon_rmap(struct page *page,
1056 struct vm_area_struct *vma, unsigned long address)
1057 {
1058 #ifdef CONFIG_DEBUG_VM
1059 /*
1060 * The page's anon-rmap details (mapping and index) are guaranteed to
1061 * be set up correctly at this point.
1062 *
1063 * We have exclusion against page_add_anon_rmap because the caller
1064 * always holds the page locked, except if called from page_dup_rmap,
1065 * in which case the page is already known to be setup.
1066 *
1067 * We have exclusion against page_add_new_anon_rmap because those pages
1068 * are initially only visible via the pagetables, and the pte is locked
1069 * over the call to page_add_new_anon_rmap.
1070 */
1071 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1072 BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
1073 #endif
1074 }
1075
1076 /**
1077 * page_add_anon_rmap - add pte mapping to an anonymous page
1078 * @page: the page to add the mapping to
1079 * @vma: the vm area in which the mapping is added
1080 * @address: the user virtual address mapped
1081 * @compound: charge the page as compound or small page
1082 *
1083 * The caller needs to hold the pte lock, and the page must be locked in
1084 * the anon_vma case: to serialize mapping,index checking after setting,
1085 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1086 * (but PageKsm is never downgraded to PageAnon).
1087 */
1088 void page_add_anon_rmap(struct page *page,
1089 struct vm_area_struct *vma, unsigned long address, bool compound)
1090 {
1091 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1092 }
1093
1094 /*
1095 * Special version of the above for do_swap_page, which often runs
1096 * into pages that are exclusively owned by the current process.
1097 * Everybody else should continue to use page_add_anon_rmap above.
1098 */
1099 void do_page_add_anon_rmap(struct page *page,
1100 struct vm_area_struct *vma, unsigned long address, int flags)
1101 {
1102 bool compound = flags & RMAP_COMPOUND;
1103 bool first;
1104
1105 if (compound) {
1106 atomic_t *mapcount;
1107 VM_BUG_ON_PAGE(!PageLocked(page), page);
1108 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1109 mapcount = compound_mapcount_ptr(page);
1110 first = atomic_inc_and_test(mapcount);
1111 } else {
1112 first = atomic_inc_and_test(&page->_mapcount);
1113 }
1114
1115 if (first) {
1116 int nr = compound ? hpage_nr_pages(page) : 1;
1117 /*
1118 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1119 * these counters are not modified in interrupt context, and
1120 * pte lock(a spinlock) is held, which implies preemption
1121 * disabled.
1122 */
1123 if (compound)
1124 __inc_node_page_state(page, NR_ANON_THPS);
1125 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1126 }
1127 if (unlikely(PageKsm(page)))
1128 return;
1129
1130 VM_BUG_ON_PAGE(!PageLocked(page), page);
1131
1132 /* address might be in next vma when migration races vma_adjust */
1133 if (first)
1134 __page_set_anon_rmap(page, vma, address,
1135 flags & RMAP_EXCLUSIVE);
1136 else
1137 __page_check_anon_rmap(page, vma, address);
1138 }
1139
1140 /**
1141 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1142 * @page: the page to add the mapping to
1143 * @vma: the vm area in which the mapping is added
1144 * @address: the user virtual address mapped
1145 * @compound: charge the page as compound or small page
1146 *
1147 * Same as page_add_anon_rmap but must only be called on *new* pages.
1148 * This means the inc-and-test can be bypassed.
1149 * Page does not have to be locked.
1150 */
1151 void page_add_new_anon_rmap(struct page *page,
1152 struct vm_area_struct *vma, unsigned long address, bool compound)
1153 {
1154 int nr = compound ? hpage_nr_pages(page) : 1;
1155
1156 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1157 __SetPageSwapBacked(page);
1158 if (compound) {
1159 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1160 /* increment count (starts at -1) */
1161 atomic_set(compound_mapcount_ptr(page), 0);
1162 __inc_node_page_state(page, NR_ANON_THPS);
1163 } else {
1164 /* Anon THP always mapped first with PMD */
1165 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1166 /* increment count (starts at -1) */
1167 atomic_set(&page->_mapcount, 0);
1168 }
1169 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1170 __page_set_anon_rmap(page, vma, address, 1);
1171 }
1172
1173 /**
1174 * page_add_file_rmap - add pte mapping to a file page
1175 * @page: the page to add the mapping to
1176 * @compound: charge the page as compound or small page
1177 *
1178 * The caller needs to hold the pte lock.
1179 */
1180 void page_add_file_rmap(struct page *page, bool compound)
1181 {
1182 int i, nr = 1;
1183
1184 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1185 lock_page_memcg(page);
1186 if (compound && PageTransHuge(page)) {
1187 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1188 if (atomic_inc_and_test(&page[i]._mapcount))
1189 nr++;
1190 }
1191 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1192 goto out;
1193 if (PageSwapBacked(page))
1194 __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1195 else
1196 __inc_node_page_state(page, NR_FILE_PMDMAPPED);
1197 } else {
1198 if (PageTransCompound(page) && page_mapping(page)) {
1199 VM_WARN_ON_ONCE(!PageLocked(page));
1200
1201 SetPageDoubleMap(compound_head(page));
1202 if (PageMlocked(page))
1203 clear_page_mlock(compound_head(page));
1204 }
1205 if (!atomic_inc_and_test(&page->_mapcount))
1206 goto out;
1207 }
1208 __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1209 out:
1210 unlock_page_memcg(page);
1211 }
1212
1213 static void page_remove_file_rmap(struct page *page, bool compound)
1214 {
1215 int i, nr = 1;
1216
1217 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1218 lock_page_memcg(page);
1219
1220 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1221 if (unlikely(PageHuge(page))) {
1222 /* hugetlb pages are always mapped with pmds */
1223 atomic_dec(compound_mapcount_ptr(page));
1224 goto out;
1225 }
1226
1227 /* page still mapped by someone else? */
1228 if (compound && PageTransHuge(page)) {
1229 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1230 if (atomic_add_negative(-1, &page[i]._mapcount))
1231 nr++;
1232 }
1233 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1234 goto out;
1235 if (PageSwapBacked(page))
1236 __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1237 else
1238 __dec_node_page_state(page, NR_FILE_PMDMAPPED);
1239 } else {
1240 if (!atomic_add_negative(-1, &page->_mapcount))
1241 goto out;
1242 }
1243
1244 /*
1245 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1246 * these counters are not modified in interrupt context, and
1247 * pte lock(a spinlock) is held, which implies preemption disabled.
1248 */
1249 __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1250
1251 if (unlikely(PageMlocked(page)))
1252 clear_page_mlock(page);
1253 out:
1254 unlock_page_memcg(page);
1255 }
1256
1257 static void page_remove_anon_compound_rmap(struct page *page)
1258 {
1259 int i, nr;
1260
1261 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1262 return;
1263
1264 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1265 if (unlikely(PageHuge(page)))
1266 return;
1267
1268 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1269 return;
1270
1271 __dec_node_page_state(page, NR_ANON_THPS);
1272
1273 if (TestClearPageDoubleMap(page)) {
1274 /*
1275 * Subpages can be mapped with PTEs too. Check how many of
1276 * themi are still mapped.
1277 */
1278 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1279 if (atomic_add_negative(-1, &page[i]._mapcount))
1280 nr++;
1281 }
1282 } else {
1283 nr = HPAGE_PMD_NR;
1284 }
1285
1286 if (unlikely(PageMlocked(page)))
1287 clear_page_mlock(page);
1288
1289 if (nr) {
1290 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
1291 deferred_split_huge_page(page);
1292 }
1293 }
1294
1295 /**
1296 * page_remove_rmap - take down pte mapping from a page
1297 * @page: page to remove mapping from
1298 * @compound: uncharge the page as compound or small page
1299 *
1300 * The caller needs to hold the pte lock.
1301 */
1302 void page_remove_rmap(struct page *page, bool compound)
1303 {
1304 if (!PageAnon(page))
1305 return page_remove_file_rmap(page, compound);
1306
1307 if (compound)
1308 return page_remove_anon_compound_rmap(page);
1309
1310 /* page still mapped by someone else? */
1311 if (!atomic_add_negative(-1, &page->_mapcount))
1312 return;
1313
1314 /*
1315 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1316 * these counters are not modified in interrupt context, and
1317 * pte lock(a spinlock) is held, which implies preemption disabled.
1318 */
1319 __dec_node_page_state(page, NR_ANON_MAPPED);
1320
1321 if (unlikely(PageMlocked(page)))
1322 clear_page_mlock(page);
1323
1324 if (PageTransCompound(page))
1325 deferred_split_huge_page(compound_head(page));
1326
1327 /*
1328 * It would be tidy to reset the PageAnon mapping here,
1329 * but that might overwrite a racing page_add_anon_rmap
1330 * which increments mapcount after us but sets mapping
1331 * before us: so leave the reset to free_unref_page,
1332 * and remember that it's only reliable while mapped.
1333 * Leaving it set also helps swapoff to reinstate ptes
1334 * faster for those pages still in swapcache.
1335 */
1336 }
1337
1338 /*
1339 * @arg: enum ttu_flags will be passed to this argument
1340 */
1341 static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1342 unsigned long address, void *arg)
1343 {
1344 struct mm_struct *mm = vma->vm_mm;
1345 struct page_vma_mapped_walk pvmw = {
1346 .page = page,
1347 .vma = vma,
1348 .address = address,
1349 };
1350 pte_t pteval;
1351 struct page *subpage;
1352 bool ret = true;
1353 struct mmu_notifier_range range;
1354 enum ttu_flags flags = (enum ttu_flags)arg;
1355
1356 /* munlock has nothing to gain from examining un-locked vmas */
1357 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1358 return true;
1359
1360 if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1361 is_zone_device_page(page) && !is_device_private_page(page))
1362 return true;
1363
1364 if (flags & TTU_SPLIT_HUGE_PMD) {
1365 split_huge_pmd_address(vma, address,
1366 flags & TTU_SPLIT_FREEZE, page);
1367 }
1368
1369 /*
1370 * For THP, we have to assume the worse case ie pmd for invalidation.
1371 * For hugetlb, it could be much worse if we need to do pud
1372 * invalidation in the case of pmd sharing.
1373 *
1374 * Note that the page can not be free in this function as call of
1375 * try_to_unmap() must hold a reference on the page.
1376 */
1377 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1378 address,
1379 min(vma->vm_end, address + page_size(page)));
1380 if (PageHuge(page)) {
1381 /*
1382 * If sharing is possible, start and end will be adjusted
1383 * accordingly.
1384 */
1385 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1386 &range.end);
1387 }
1388 mmu_notifier_invalidate_range_start(&range);
1389
1390 while (page_vma_mapped_walk(&pvmw)) {
1391 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1392 /* PMD-mapped THP migration entry */
1393 if (!pvmw.pte && (flags & TTU_MIGRATION)) {
1394 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
1395
1396 set_pmd_migration_entry(&pvmw, page);
1397 continue;
1398 }
1399 #endif
1400
1401 /*
1402 * If the page is mlock()d, we cannot swap it out.
1403 * If it's recently referenced (perhaps page_referenced
1404 * skipped over this mm) then we should reactivate it.
1405 */
1406 if (!(flags & TTU_IGNORE_MLOCK)) {
1407 if (vma->vm_flags & VM_LOCKED) {
1408 /* PTE-mapped THP are never mlocked */
1409 if (!PageTransCompound(page)) {
1410 /*
1411 * Holding pte lock, we do *not* need
1412 * mmap_sem here
1413 */
1414 mlock_vma_page(page);
1415 }
1416 ret = false;
1417 page_vma_mapped_walk_done(&pvmw);
1418 break;
1419 }
1420 if (flags & TTU_MUNLOCK)
1421 continue;
1422 }
1423
1424 /* Unexpected PMD-mapped THP? */
1425 VM_BUG_ON_PAGE(!pvmw.pte, page);
1426
1427 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1428 address = pvmw.address;
1429
1430 if (PageHuge(page)) {
1431 if (huge_pmd_unshare(mm, &address, pvmw.pte)) {
1432 /*
1433 * huge_pmd_unshare unmapped an entire PMD
1434 * page. There is no way of knowing exactly
1435 * which PMDs may be cached for this mm, so
1436 * we must flush them all. start/end were
1437 * already adjusted above to cover this range.
1438 */
1439 flush_cache_range(vma, range.start, range.end);
1440 flush_tlb_range(vma, range.start, range.end);
1441 mmu_notifier_invalidate_range(mm, range.start,
1442 range.end);
1443
1444 /*
1445 * The ref count of the PMD page was dropped
1446 * which is part of the way map counting
1447 * is done for shared PMDs. Return 'true'
1448 * here. When there is no other sharing,
1449 * huge_pmd_unshare returns false and we will
1450 * unmap the actual page and drop map count
1451 * to zero.
1452 */
1453 page_vma_mapped_walk_done(&pvmw);
1454 break;
1455 }
1456 }
1457
1458 if (IS_ENABLED(CONFIG_MIGRATION) &&
1459 (flags & TTU_MIGRATION) &&
1460 is_zone_device_page(page)) {
1461 swp_entry_t entry;
1462 pte_t swp_pte;
1463
1464 pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte);
1465
1466 /*
1467 * Store the pfn of the page in a special migration
1468 * pte. do_swap_page() will wait until the migration
1469 * pte is removed and then restart fault handling.
1470 */
1471 entry = make_migration_entry(page, 0);
1472 swp_pte = swp_entry_to_pte(entry);
1473 if (pte_soft_dirty(pteval))
1474 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1475 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1476 /*
1477 * No need to invalidate here it will synchronize on
1478 * against the special swap migration pte.
1479 *
1480 * The assignment to subpage above was computed from a
1481 * swap PTE which results in an invalid pointer.
1482 * Since only PAGE_SIZE pages can currently be
1483 * migrated, just set it to page. This will need to be
1484 * changed when hugepage migrations to device private
1485 * memory are supported.
1486 */
1487 subpage = page;
1488 goto discard;
1489 }
1490
1491 if (!(flags & TTU_IGNORE_ACCESS)) {
1492 if (ptep_clear_flush_young_notify(vma, address,
1493 pvmw.pte)) {
1494 ret = false;
1495 page_vma_mapped_walk_done(&pvmw);
1496 break;
1497 }
1498 }
1499
1500 /* Nuke the page table entry. */
1501 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1502 if (should_defer_flush(mm, flags)) {
1503 /*
1504 * We clear the PTE but do not flush so potentially
1505 * a remote CPU could still be writing to the page.
1506 * If the entry was previously clean then the
1507 * architecture must guarantee that a clear->dirty
1508 * transition on a cached TLB entry is written through
1509 * and traps if the PTE is unmapped.
1510 */
1511 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1512
1513 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1514 } else {
1515 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1516 }
1517
1518 /* Move the dirty bit to the page. Now the pte is gone. */
1519 if (pte_dirty(pteval))
1520 set_page_dirty(page);
1521
1522 /* Update high watermark before we lower rss */
1523 update_hiwater_rss(mm);
1524
1525 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1526 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1527 if (PageHuge(page)) {
1528 hugetlb_count_sub(compound_nr(page), mm);
1529 set_huge_swap_pte_at(mm, address,
1530 pvmw.pte, pteval,
1531 vma_mmu_pagesize(vma));
1532 } else {
1533 dec_mm_counter(mm, mm_counter(page));
1534 set_pte_at(mm, address, pvmw.pte, pteval);
1535 }
1536
1537 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1538 /*
1539 * The guest indicated that the page content is of no
1540 * interest anymore. Simply discard the pte, vmscan
1541 * will take care of the rest.
1542 * A future reference will then fault in a new zero
1543 * page. When userfaultfd is active, we must not drop
1544 * this page though, as its main user (postcopy
1545 * migration) will not expect userfaults on already
1546 * copied pages.
1547 */
1548 dec_mm_counter(mm, mm_counter(page));
1549 /* We have to invalidate as we cleared the pte */
1550 mmu_notifier_invalidate_range(mm, address,
1551 address + PAGE_SIZE);
1552 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1553 (flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) {
1554 swp_entry_t entry;
1555 pte_t swp_pte;
1556
1557 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1558 set_pte_at(mm, address, pvmw.pte, pteval);
1559 ret = false;
1560 page_vma_mapped_walk_done(&pvmw);
1561 break;
1562 }
1563
1564 /*
1565 * Store the pfn of the page in a special migration
1566 * pte. do_swap_page() will wait until the migration
1567 * pte is removed and then restart fault handling.
1568 */
1569 entry = make_migration_entry(subpage,
1570 pte_write(pteval));
1571 swp_pte = swp_entry_to_pte(entry);
1572 if (pte_soft_dirty(pteval))
1573 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1574 set_pte_at(mm, address, pvmw.pte, swp_pte);
1575 /*
1576 * No need to invalidate here it will synchronize on
1577 * against the special swap migration pte.
1578 */
1579 } else if (PageAnon(page)) {
1580 swp_entry_t entry = { .val = page_private(subpage) };
1581 pte_t swp_pte;
1582 /*
1583 * Store the swap location in the pte.
1584 * See handle_pte_fault() ...
1585 */
1586 if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1587 WARN_ON_ONCE(1);
1588 ret = false;
1589 /* We have to invalidate as we cleared the pte */
1590 mmu_notifier_invalidate_range(mm, address,
1591 address + PAGE_SIZE);
1592 page_vma_mapped_walk_done(&pvmw);
1593 break;
1594 }
1595
1596 /* MADV_FREE page check */
1597 if (!PageSwapBacked(page)) {
1598 if (!PageDirty(page)) {
1599 /* Invalidate as we cleared the pte */
1600 mmu_notifier_invalidate_range(mm,
1601 address, address + PAGE_SIZE);
1602 dec_mm_counter(mm, MM_ANONPAGES);
1603 goto discard;
1604 }
1605
1606 /*
1607 * If the page was redirtied, it cannot be
1608 * discarded. Remap the page to page table.
1609 */
1610 set_pte_at(mm, address, pvmw.pte, pteval);
1611 SetPageSwapBacked(page);
1612 ret = false;
1613 page_vma_mapped_walk_done(&pvmw);
1614 break;
1615 }
1616
1617 if (swap_duplicate(entry) < 0) {
1618 set_pte_at(mm, address, pvmw.pte, pteval);
1619 ret = false;
1620 page_vma_mapped_walk_done(&pvmw);
1621 break;
1622 }
1623 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1624 set_pte_at(mm, address, pvmw.pte, pteval);
1625 ret = false;
1626 page_vma_mapped_walk_done(&pvmw);
1627 break;
1628 }
1629 if (list_empty(&mm->mmlist)) {
1630 spin_lock(&mmlist_lock);
1631 if (list_empty(&mm->mmlist))
1632 list_add(&mm->mmlist, &init_mm.mmlist);
1633 spin_unlock(&mmlist_lock);
1634 }
1635 dec_mm_counter(mm, MM_ANONPAGES);
1636 inc_mm_counter(mm, MM_SWAPENTS);
1637 swp_pte = swp_entry_to_pte(entry);
1638 if (pte_soft_dirty(pteval))
1639 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1640 set_pte_at(mm, address, pvmw.pte, swp_pte);
1641 /* Invalidate as we cleared the pte */
1642 mmu_notifier_invalidate_range(mm, address,
1643 address + PAGE_SIZE);
1644 } else {
1645 /*
1646 * This is a locked file-backed page, thus it cannot
1647 * be removed from the page cache and replaced by a new
1648 * page before mmu_notifier_invalidate_range_end, so no
1649 * concurrent thread might update its page table to
1650 * point at new page while a device still is using this
1651 * page.
1652 *
1653 * See Documentation/vm/mmu_notifier.rst
1654 */
1655 dec_mm_counter(mm, mm_counter_file(page));
1656 }
1657 discard:
1658 /*
1659 * No need to call mmu_notifier_invalidate_range() it has be
1660 * done above for all cases requiring it to happen under page
1661 * table lock before mmu_notifier_invalidate_range_end()
1662 *
1663 * See Documentation/vm/mmu_notifier.rst
1664 */
1665 page_remove_rmap(subpage, PageHuge(page));
1666 put_page(page);
1667 }
1668
1669 mmu_notifier_invalidate_range_end(&range);
1670
1671 return ret;
1672 }
1673
1674 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1675 {
1676 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1677
1678 if (!maybe_stack)
1679 return false;
1680
1681 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1682 VM_STACK_INCOMPLETE_SETUP)
1683 return true;
1684
1685 return false;
1686 }
1687
1688 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1689 {
1690 return is_vma_temporary_stack(vma);
1691 }
1692
1693 static int page_mapcount_is_zero(struct page *page)
1694 {
1695 return !total_mapcount(page);
1696 }
1697
1698 /**
1699 * try_to_unmap - try to remove all page table mappings to a page
1700 * @page: the page to get unmapped
1701 * @flags: action and flags
1702 *
1703 * Tries to remove all the page table entries which are mapping this
1704 * page, used in the pageout path. Caller must hold the page lock.
1705 *
1706 * If unmap is successful, return true. Otherwise, false.
1707 */
1708 bool try_to_unmap(struct page *page, enum ttu_flags flags)
1709 {
1710 struct rmap_walk_control rwc = {
1711 .rmap_one = try_to_unmap_one,
1712 .arg = (void *)flags,
1713 .done = page_mapcount_is_zero,
1714 .anon_lock = page_lock_anon_vma_read,
1715 };
1716
1717 /*
1718 * During exec, a temporary VMA is setup and later moved.
1719 * The VMA is moved under the anon_vma lock but not the
1720 * page tables leading to a race where migration cannot
1721 * find the migration ptes. Rather than increasing the
1722 * locking requirements of exec(), migration skips
1723 * temporary VMAs until after exec() completes.
1724 */
1725 if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))
1726 && !PageKsm(page) && PageAnon(page))
1727 rwc.invalid_vma = invalid_migration_vma;
1728
1729 if (flags & TTU_RMAP_LOCKED)
1730 rmap_walk_locked(page, &rwc);
1731 else
1732 rmap_walk(page, &rwc);
1733
1734 return !page_mapcount(page) ? true : false;
1735 }
1736
1737 static int page_not_mapped(struct page *page)
1738 {
1739 return !page_mapped(page);
1740 };
1741
1742 /**
1743 * try_to_munlock - try to munlock a page
1744 * @page: the page to be munlocked
1745 *
1746 * Called from munlock code. Checks all of the VMAs mapping the page
1747 * to make sure nobody else has this page mlocked. The page will be
1748 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1749 */
1750
1751 void try_to_munlock(struct page *page)
1752 {
1753 struct rmap_walk_control rwc = {
1754 .rmap_one = try_to_unmap_one,
1755 .arg = (void *)TTU_MUNLOCK,
1756 .done = page_not_mapped,
1757 .anon_lock = page_lock_anon_vma_read,
1758
1759 };
1760
1761 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1762 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
1763
1764 rmap_walk(page, &rwc);
1765 }
1766
1767 void __put_anon_vma(struct anon_vma *anon_vma)
1768 {
1769 struct anon_vma *root = anon_vma->root;
1770
1771 anon_vma_free(anon_vma);
1772 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1773 anon_vma_free(root);
1774 }
1775
1776 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1777 struct rmap_walk_control *rwc)
1778 {
1779 struct anon_vma *anon_vma;
1780
1781 if (rwc->anon_lock)
1782 return rwc->anon_lock(page);
1783
1784 /*
1785 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1786 * because that depends on page_mapped(); but not all its usages
1787 * are holding mmap_sem. Users without mmap_sem are required to
1788 * take a reference count to prevent the anon_vma disappearing
1789 */
1790 anon_vma = page_anon_vma(page);
1791 if (!anon_vma)
1792 return NULL;
1793
1794 anon_vma_lock_read(anon_vma);
1795 return anon_vma;
1796 }
1797
1798 /*
1799 * rmap_walk_anon - do something to anonymous page using the object-based
1800 * rmap method
1801 * @page: the page to be handled
1802 * @rwc: control variable according to each walk type
1803 *
1804 * Find all the mappings of a page using the mapping pointer and the vma chains
1805 * contained in the anon_vma struct it points to.
1806 *
1807 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1808 * where the page was found will be held for write. So, we won't recheck
1809 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1810 * LOCKED.
1811 */
1812 static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1813 bool locked)
1814 {
1815 struct anon_vma *anon_vma;
1816 pgoff_t pgoff_start, pgoff_end;
1817 struct anon_vma_chain *avc;
1818
1819 if (locked) {
1820 anon_vma = page_anon_vma(page);
1821 /* anon_vma disappear under us? */
1822 VM_BUG_ON_PAGE(!anon_vma, page);
1823 } else {
1824 anon_vma = rmap_walk_anon_lock(page, rwc);
1825 }
1826 if (!anon_vma)
1827 return;
1828
1829 pgoff_start = page_to_pgoff(page);
1830 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1831 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1832 pgoff_start, pgoff_end) {
1833 struct vm_area_struct *vma = avc->vma;
1834 unsigned long address = vma_address(page, vma);
1835
1836 cond_resched();
1837
1838 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1839 continue;
1840
1841 if (!rwc->rmap_one(page, vma, address, rwc->arg))
1842 break;
1843 if (rwc->done && rwc->done(page))
1844 break;
1845 }
1846
1847 if (!locked)
1848 anon_vma_unlock_read(anon_vma);
1849 }
1850
1851 /*
1852 * rmap_walk_file - do something to file page using the object-based rmap method
1853 * @page: the page to be handled
1854 * @rwc: control variable according to each walk type
1855 *
1856 * Find all the mappings of a page using the mapping pointer and the vma chains
1857 * contained in the address_space struct it points to.
1858 *
1859 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1860 * where the page was found will be held for write. So, we won't recheck
1861 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1862 * LOCKED.
1863 */
1864 static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1865 bool locked)
1866 {
1867 struct address_space *mapping = page_mapping(page);
1868 pgoff_t pgoff_start, pgoff_end;
1869 struct vm_area_struct *vma;
1870
1871 /*
1872 * The page lock not only makes sure that page->mapping cannot
1873 * suddenly be NULLified by truncation, it makes sure that the
1874 * structure at mapping cannot be freed and reused yet,
1875 * so we can safely take mapping->i_mmap_rwsem.
1876 */
1877 VM_BUG_ON_PAGE(!PageLocked(page), page);
1878
1879 if (!mapping)
1880 return;
1881
1882 pgoff_start = page_to_pgoff(page);
1883 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1884 if (!locked)
1885 i_mmap_lock_read(mapping);
1886 vma_interval_tree_foreach(vma, &mapping->i_mmap,
1887 pgoff_start, pgoff_end) {
1888 unsigned long address = vma_address(page, vma);
1889
1890 cond_resched();
1891
1892 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1893 continue;
1894
1895 if (!rwc->rmap_one(page, vma, address, rwc->arg))
1896 goto done;
1897 if (rwc->done && rwc->done(page))
1898 goto done;
1899 }
1900
1901 done:
1902 if (!locked)
1903 i_mmap_unlock_read(mapping);
1904 }
1905
1906 void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1907 {
1908 if (unlikely(PageKsm(page)))
1909 rmap_walk_ksm(page, rwc);
1910 else if (PageAnon(page))
1911 rmap_walk_anon(page, rwc, false);
1912 else
1913 rmap_walk_file(page, rwc, false);
1914 }
1915
1916 /* Like rmap_walk, but caller holds relevant rmap lock */
1917 void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1918 {
1919 /* no ksm support for now */
1920 VM_BUG_ON_PAGE(PageKsm(page), page);
1921 if (PageAnon(page))
1922 rmap_walk_anon(page, rwc, true);
1923 else
1924 rmap_walk_file(page, rwc, true);
1925 }
1926
1927 #ifdef CONFIG_HUGETLB_PAGE
1928 /*
1929 * The following two functions are for anonymous (private mapped) hugepages.
1930 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1931 * and no lru code, because we handle hugepages differently from common pages.
1932 */
1933 void hugepage_add_anon_rmap(struct page *page,
1934 struct vm_area_struct *vma, unsigned long address)
1935 {
1936 struct anon_vma *anon_vma = vma->anon_vma;
1937 int first;
1938
1939 BUG_ON(!PageLocked(page));
1940 BUG_ON(!anon_vma);
1941 /* address might be in next vma when migration races vma_adjust */
1942 first = atomic_inc_and_test(compound_mapcount_ptr(page));
1943 if (first)
1944 __page_set_anon_rmap(page, vma, address, 0);
1945 }
1946
1947 void hugepage_add_new_anon_rmap(struct page *page,
1948 struct vm_area_struct *vma, unsigned long address)
1949 {
1950 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1951 atomic_set(compound_mapcount_ptr(page), 0);
1952 __page_set_anon_rmap(page, vma, address, 1);
1953 }
1954 #endif /* CONFIG_HUGETLB_PAGE */