Loading...
1/*
2 * mm/rmap.c - physical to virtual reverse mappings
3 *
4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5 * Released under the General Public License (GPL).
6 *
7 * Simple, low overhead reverse mapping scheme.
8 * Please try to keep this thing as modular as possible.
9 *
10 * Provides methods for unmapping each kind of mapped page:
11 * the anon methods track anonymous pages, and
12 * the file methods track pages belonging to an inode.
13 *
14 * Original design by Rik van Riel <riel@conectiva.com.br> 2001
15 * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
16 * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
17 * Contributions by Hugh Dickins 2003, 2004
18 */
19
20/*
21 * Lock ordering in mm:
22 *
23 * inode->i_rwsem (while writing or truncating, not reading or faulting)
24 * mm->mmap_lock
25 * mapping->invalidate_lock (in filemap_fault)
26 * page->flags PG_locked (lock_page)
27 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share, see hugetlbfs below)
28 * vma_start_write
29 * mapping->i_mmap_rwsem
30 * anon_vma->rwsem
31 * mm->page_table_lock or pte_lock
32 * swap_lock (in swap_duplicate, swap_info_get)
33 * mmlist_lock (in mmput, drain_mmlist and others)
34 * mapping->private_lock (in block_dirty_folio)
35 * folio_lock_memcg move_lock (in block_dirty_folio)
36 * i_pages lock (widely used)
37 * lruvec->lru_lock (in folio_lruvec_lock_irq)
38 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
39 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
40 * sb_lock (within inode_lock in fs/fs-writeback.c)
41 * i_pages lock (widely used, in set_page_dirty,
42 * in arch-dependent flush_dcache_mmap_lock,
43 * within bdi.wb->list_lock in __sync_single_inode)
44 *
45 * anon_vma->rwsem,mapping->i_mmap_rwsem (memory_failure, collect_procs_anon)
46 * ->tasklist_lock
47 * pte map lock
48 *
49 * hugetlbfs PageHuge() take locks in this order:
50 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
51 * vma_lock (hugetlb specific lock for pmd_sharing)
52 * mapping->i_mmap_rwsem (also used for hugetlb pmd sharing)
53 * page->flags PG_locked (lock_page)
54 */
55
56#include <linux/mm.h>
57#include <linux/sched/mm.h>
58#include <linux/sched/task.h>
59#include <linux/pagemap.h>
60#include <linux/swap.h>
61#include <linux/swapops.h>
62#include <linux/slab.h>
63#include <linux/init.h>
64#include <linux/ksm.h>
65#include <linux/rmap.h>
66#include <linux/rcupdate.h>
67#include <linux/export.h>
68#include <linux/memcontrol.h>
69#include <linux/mmu_notifier.h>
70#include <linux/migrate.h>
71#include <linux/hugetlb.h>
72#include <linux/huge_mm.h>
73#include <linux/backing-dev.h>
74#include <linux/page_idle.h>
75#include <linux/memremap.h>
76#include <linux/userfaultfd_k.h>
77#include <linux/mm_inline.h>
78
79#include <asm/tlbflush.h>
80
81#define CREATE_TRACE_POINTS
82#include <trace/events/tlb.h>
83#include <trace/events/migrate.h>
84
85#include "internal.h"
86
87static struct kmem_cache *anon_vma_cachep;
88static struct kmem_cache *anon_vma_chain_cachep;
89
90static inline struct anon_vma *anon_vma_alloc(void)
91{
92 struct anon_vma *anon_vma;
93
94 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
95 if (anon_vma) {
96 atomic_set(&anon_vma->refcount, 1);
97 anon_vma->num_children = 0;
98 anon_vma->num_active_vmas = 0;
99 anon_vma->parent = anon_vma;
100 /*
101 * Initialise the anon_vma root to point to itself. If called
102 * from fork, the root will be reset to the parents anon_vma.
103 */
104 anon_vma->root = anon_vma;
105 }
106
107 return anon_vma;
108}
109
110static inline void anon_vma_free(struct anon_vma *anon_vma)
111{
112 VM_BUG_ON(atomic_read(&anon_vma->refcount));
113
114 /*
115 * Synchronize against folio_lock_anon_vma_read() such that
116 * we can safely hold the lock without the anon_vma getting
117 * freed.
118 *
119 * Relies on the full mb implied by the atomic_dec_and_test() from
120 * put_anon_vma() against the acquire barrier implied by
121 * down_read_trylock() from folio_lock_anon_vma_read(). This orders:
122 *
123 * folio_lock_anon_vma_read() VS put_anon_vma()
124 * down_read_trylock() atomic_dec_and_test()
125 * LOCK MB
126 * atomic_read() rwsem_is_locked()
127 *
128 * LOCK should suffice since the actual taking of the lock must
129 * happen _before_ what follows.
130 */
131 might_sleep();
132 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
133 anon_vma_lock_write(anon_vma);
134 anon_vma_unlock_write(anon_vma);
135 }
136
137 kmem_cache_free(anon_vma_cachep, anon_vma);
138}
139
140static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
141{
142 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
143}
144
145static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
146{
147 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
148}
149
150static void anon_vma_chain_link(struct vm_area_struct *vma,
151 struct anon_vma_chain *avc,
152 struct anon_vma *anon_vma)
153{
154 avc->vma = vma;
155 avc->anon_vma = anon_vma;
156 list_add(&avc->same_vma, &vma->anon_vma_chain);
157 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
158}
159
160/**
161 * __anon_vma_prepare - attach an anon_vma to a memory region
162 * @vma: the memory region in question
163 *
164 * This makes sure the memory mapping described by 'vma' has
165 * an 'anon_vma' attached to it, so that we can associate the
166 * anonymous pages mapped into it with that anon_vma.
167 *
168 * The common case will be that we already have one, which
169 * is handled inline by anon_vma_prepare(). But if
170 * not we either need to find an adjacent mapping that we
171 * can re-use the anon_vma from (very common when the only
172 * reason for splitting a vma has been mprotect()), or we
173 * allocate a new one.
174 *
175 * Anon-vma allocations are very subtle, because we may have
176 * optimistically looked up an anon_vma in folio_lock_anon_vma_read()
177 * and that may actually touch the rwsem even in the newly
178 * allocated vma (it depends on RCU to make sure that the
179 * anon_vma isn't actually destroyed).
180 *
181 * As a result, we need to do proper anon_vma locking even
182 * for the new allocation. At the same time, we do not want
183 * to do any locking for the common case of already having
184 * an anon_vma.
185 *
186 * This must be called with the mmap_lock held for reading.
187 */
188int __anon_vma_prepare(struct vm_area_struct *vma)
189{
190 struct mm_struct *mm = vma->vm_mm;
191 struct anon_vma *anon_vma, *allocated;
192 struct anon_vma_chain *avc;
193
194 might_sleep();
195
196 avc = anon_vma_chain_alloc(GFP_KERNEL);
197 if (!avc)
198 goto out_enomem;
199
200 anon_vma = find_mergeable_anon_vma(vma);
201 allocated = NULL;
202 if (!anon_vma) {
203 anon_vma = anon_vma_alloc();
204 if (unlikely(!anon_vma))
205 goto out_enomem_free_avc;
206 anon_vma->num_children++; /* self-parent link for new root */
207 allocated = anon_vma;
208 }
209
210 anon_vma_lock_write(anon_vma);
211 /* page_table_lock to protect against threads */
212 spin_lock(&mm->page_table_lock);
213 if (likely(!vma->anon_vma)) {
214 vma->anon_vma = anon_vma;
215 anon_vma_chain_link(vma, avc, anon_vma);
216 anon_vma->num_active_vmas++;
217 allocated = NULL;
218 avc = NULL;
219 }
220 spin_unlock(&mm->page_table_lock);
221 anon_vma_unlock_write(anon_vma);
222
223 if (unlikely(allocated))
224 put_anon_vma(allocated);
225 if (unlikely(avc))
226 anon_vma_chain_free(avc);
227
228 return 0;
229
230 out_enomem_free_avc:
231 anon_vma_chain_free(avc);
232 out_enomem:
233 return -ENOMEM;
234}
235
236/*
237 * This is a useful helper function for locking the anon_vma root as
238 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
239 * have the same vma.
240 *
241 * Such anon_vma's should have the same root, so you'd expect to see
242 * just a single mutex_lock for the whole traversal.
243 */
244static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
245{
246 struct anon_vma *new_root = anon_vma->root;
247 if (new_root != root) {
248 if (WARN_ON_ONCE(root))
249 up_write(&root->rwsem);
250 root = new_root;
251 down_write(&root->rwsem);
252 }
253 return root;
254}
255
256static inline void unlock_anon_vma_root(struct anon_vma *root)
257{
258 if (root)
259 up_write(&root->rwsem);
260}
261
262/*
263 * Attach the anon_vmas from src to dst.
264 * Returns 0 on success, -ENOMEM on failure.
265 *
266 * anon_vma_clone() is called by vma_expand(), vma_merge(), __split_vma(),
267 * copy_vma() and anon_vma_fork(). The first four want an exact copy of src,
268 * while the last one, anon_vma_fork(), may try to reuse an existing anon_vma to
269 * prevent endless growth of anon_vma. Since dst->anon_vma is set to NULL before
270 * call, we can identify this case by checking (!dst->anon_vma &&
271 * src->anon_vma).
272 *
273 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
274 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
275 * This prevents degradation of anon_vma hierarchy to endless linear chain in
276 * case of constantly forking task. On the other hand, an anon_vma with more
277 * than one child isn't reused even if there was no alive vma, thus rmap
278 * walker has a good chance of avoiding scanning the whole hierarchy when it
279 * searches where page is mapped.
280 */
281int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
282{
283 struct anon_vma_chain *avc, *pavc;
284 struct anon_vma *root = NULL;
285
286 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
287 struct anon_vma *anon_vma;
288
289 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
290 if (unlikely(!avc)) {
291 unlock_anon_vma_root(root);
292 root = NULL;
293 avc = anon_vma_chain_alloc(GFP_KERNEL);
294 if (!avc)
295 goto enomem_failure;
296 }
297 anon_vma = pavc->anon_vma;
298 root = lock_anon_vma_root(root, anon_vma);
299 anon_vma_chain_link(dst, avc, anon_vma);
300
301 /*
302 * Reuse existing anon_vma if it has no vma and only one
303 * anon_vma child.
304 *
305 * Root anon_vma is never reused:
306 * it has self-parent reference and at least one child.
307 */
308 if (!dst->anon_vma && src->anon_vma &&
309 anon_vma->num_children < 2 &&
310 anon_vma->num_active_vmas == 0)
311 dst->anon_vma = anon_vma;
312 }
313 if (dst->anon_vma)
314 dst->anon_vma->num_active_vmas++;
315 unlock_anon_vma_root(root);
316 return 0;
317
318 enomem_failure:
319 /*
320 * dst->anon_vma is dropped here otherwise its num_active_vmas can
321 * be incorrectly decremented in unlink_anon_vmas().
322 * We can safely do this because callers of anon_vma_clone() don't care
323 * about dst->anon_vma if anon_vma_clone() failed.
324 */
325 dst->anon_vma = NULL;
326 unlink_anon_vmas(dst);
327 return -ENOMEM;
328}
329
330/*
331 * Attach vma to its own anon_vma, as well as to the anon_vmas that
332 * the corresponding VMA in the parent process is attached to.
333 * Returns 0 on success, non-zero on failure.
334 */
335int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
336{
337 struct anon_vma_chain *avc;
338 struct anon_vma *anon_vma;
339 int error;
340
341 /* Don't bother if the parent process has no anon_vma here. */
342 if (!pvma->anon_vma)
343 return 0;
344
345 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
346 vma->anon_vma = NULL;
347
348 /*
349 * First, attach the new VMA to the parent VMA's anon_vmas,
350 * so rmap can find non-COWed pages in child processes.
351 */
352 error = anon_vma_clone(vma, pvma);
353 if (error)
354 return error;
355
356 /* An existing anon_vma has been reused, all done then. */
357 if (vma->anon_vma)
358 return 0;
359
360 /* Then add our own anon_vma. */
361 anon_vma = anon_vma_alloc();
362 if (!anon_vma)
363 goto out_error;
364 anon_vma->num_active_vmas++;
365 avc = anon_vma_chain_alloc(GFP_KERNEL);
366 if (!avc)
367 goto out_error_free_anon_vma;
368
369 /*
370 * The root anon_vma's rwsem is the lock actually used when we
371 * lock any of the anon_vmas in this anon_vma tree.
372 */
373 anon_vma->root = pvma->anon_vma->root;
374 anon_vma->parent = pvma->anon_vma;
375 /*
376 * With refcounts, an anon_vma can stay around longer than the
377 * process it belongs to. The root anon_vma needs to be pinned until
378 * this anon_vma is freed, because the lock lives in the root.
379 */
380 get_anon_vma(anon_vma->root);
381 /* Mark this anon_vma as the one where our new (COWed) pages go. */
382 vma->anon_vma = anon_vma;
383 anon_vma_lock_write(anon_vma);
384 anon_vma_chain_link(vma, avc, anon_vma);
385 anon_vma->parent->num_children++;
386 anon_vma_unlock_write(anon_vma);
387
388 return 0;
389
390 out_error_free_anon_vma:
391 put_anon_vma(anon_vma);
392 out_error:
393 unlink_anon_vmas(vma);
394 return -ENOMEM;
395}
396
397void unlink_anon_vmas(struct vm_area_struct *vma)
398{
399 struct anon_vma_chain *avc, *next;
400 struct anon_vma *root = NULL;
401
402 /*
403 * Unlink each anon_vma chained to the VMA. This list is ordered
404 * from newest to oldest, ensuring the root anon_vma gets freed last.
405 */
406 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
407 struct anon_vma *anon_vma = avc->anon_vma;
408
409 root = lock_anon_vma_root(root, anon_vma);
410 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
411
412 /*
413 * Leave empty anon_vmas on the list - we'll need
414 * to free them outside the lock.
415 */
416 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
417 anon_vma->parent->num_children--;
418 continue;
419 }
420
421 list_del(&avc->same_vma);
422 anon_vma_chain_free(avc);
423 }
424 if (vma->anon_vma) {
425 vma->anon_vma->num_active_vmas--;
426
427 /*
428 * vma would still be needed after unlink, and anon_vma will be prepared
429 * when handle fault.
430 */
431 vma->anon_vma = NULL;
432 }
433 unlock_anon_vma_root(root);
434
435 /*
436 * Iterate the list once more, it now only contains empty and unlinked
437 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
438 * needing to write-acquire the anon_vma->root->rwsem.
439 */
440 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
441 struct anon_vma *anon_vma = avc->anon_vma;
442
443 VM_WARN_ON(anon_vma->num_children);
444 VM_WARN_ON(anon_vma->num_active_vmas);
445 put_anon_vma(anon_vma);
446
447 list_del(&avc->same_vma);
448 anon_vma_chain_free(avc);
449 }
450}
451
452static void anon_vma_ctor(void *data)
453{
454 struct anon_vma *anon_vma = data;
455
456 init_rwsem(&anon_vma->rwsem);
457 atomic_set(&anon_vma->refcount, 0);
458 anon_vma->rb_root = RB_ROOT_CACHED;
459}
460
461void __init anon_vma_init(void)
462{
463 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
464 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
465 anon_vma_ctor);
466 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
467 SLAB_PANIC|SLAB_ACCOUNT);
468}
469
470/*
471 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
472 *
473 * Since there is no serialization what so ever against folio_remove_rmap_*()
474 * the best this function can do is return a refcount increased anon_vma
475 * that might have been relevant to this page.
476 *
477 * The page might have been remapped to a different anon_vma or the anon_vma
478 * returned may already be freed (and even reused).
479 *
480 * In case it was remapped to a different anon_vma, the new anon_vma will be a
481 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
482 * ensure that any anon_vma obtained from the page will still be valid for as
483 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
484 *
485 * All users of this function must be very careful when walking the anon_vma
486 * chain and verify that the page in question is indeed mapped in it
487 * [ something equivalent to page_mapped_in_vma() ].
488 *
489 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
490 * folio_remove_rmap_*() that the anon_vma pointer from page->mapping is valid
491 * if there is a mapcount, we can dereference the anon_vma after observing
492 * those.
493 *
494 * NOTE: the caller should normally hold folio lock when calling this. If
495 * not, the caller needs to double check the anon_vma didn't change after
496 * taking the anon_vma lock for either read or write (UFFDIO_MOVE can modify it
497 * concurrently without folio lock protection). See folio_lock_anon_vma_read()
498 * which has already covered that, and comment above remap_pages().
499 */
500struct anon_vma *folio_get_anon_vma(struct folio *folio)
501{
502 struct anon_vma *anon_vma = NULL;
503 unsigned long anon_mapping;
504
505 rcu_read_lock();
506 anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
507 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
508 goto out;
509 if (!folio_mapped(folio))
510 goto out;
511
512 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
513 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
514 anon_vma = NULL;
515 goto out;
516 }
517
518 /*
519 * If this folio is still mapped, then its anon_vma cannot have been
520 * freed. But if it has been unmapped, we have no security against the
521 * anon_vma structure being freed and reused (for another anon_vma:
522 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
523 * above cannot corrupt).
524 */
525 if (!folio_mapped(folio)) {
526 rcu_read_unlock();
527 put_anon_vma(anon_vma);
528 return NULL;
529 }
530out:
531 rcu_read_unlock();
532
533 return anon_vma;
534}
535
536/*
537 * Similar to folio_get_anon_vma() except it locks the anon_vma.
538 *
539 * Its a little more complex as it tries to keep the fast path to a single
540 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
541 * reference like with folio_get_anon_vma() and then block on the mutex
542 * on !rwc->try_lock case.
543 */
544struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
545 struct rmap_walk_control *rwc)
546{
547 struct anon_vma *anon_vma = NULL;
548 struct anon_vma *root_anon_vma;
549 unsigned long anon_mapping;
550
551retry:
552 rcu_read_lock();
553 anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
554 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
555 goto out;
556 if (!folio_mapped(folio))
557 goto out;
558
559 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
560 root_anon_vma = READ_ONCE(anon_vma->root);
561 if (down_read_trylock(&root_anon_vma->rwsem)) {
562 /*
563 * folio_move_anon_rmap() might have changed the anon_vma as we
564 * might not hold the folio lock here.
565 */
566 if (unlikely((unsigned long)READ_ONCE(folio->mapping) !=
567 anon_mapping)) {
568 up_read(&root_anon_vma->rwsem);
569 rcu_read_unlock();
570 goto retry;
571 }
572
573 /*
574 * If the folio is still mapped, then this anon_vma is still
575 * its anon_vma, and holding the mutex ensures that it will
576 * not go away, see anon_vma_free().
577 */
578 if (!folio_mapped(folio)) {
579 up_read(&root_anon_vma->rwsem);
580 anon_vma = NULL;
581 }
582 goto out;
583 }
584
585 if (rwc && rwc->try_lock) {
586 anon_vma = NULL;
587 rwc->contended = true;
588 goto out;
589 }
590
591 /* trylock failed, we got to sleep */
592 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
593 anon_vma = NULL;
594 goto out;
595 }
596
597 if (!folio_mapped(folio)) {
598 rcu_read_unlock();
599 put_anon_vma(anon_vma);
600 return NULL;
601 }
602
603 /* we pinned the anon_vma, its safe to sleep */
604 rcu_read_unlock();
605 anon_vma_lock_read(anon_vma);
606
607 /*
608 * folio_move_anon_rmap() might have changed the anon_vma as we might
609 * not hold the folio lock here.
610 */
611 if (unlikely((unsigned long)READ_ONCE(folio->mapping) !=
612 anon_mapping)) {
613 anon_vma_unlock_read(anon_vma);
614 put_anon_vma(anon_vma);
615 anon_vma = NULL;
616 goto retry;
617 }
618
619 if (atomic_dec_and_test(&anon_vma->refcount)) {
620 /*
621 * Oops, we held the last refcount, release the lock
622 * and bail -- can't simply use put_anon_vma() because
623 * we'll deadlock on the anon_vma_lock_write() recursion.
624 */
625 anon_vma_unlock_read(anon_vma);
626 __put_anon_vma(anon_vma);
627 anon_vma = NULL;
628 }
629
630 return anon_vma;
631
632out:
633 rcu_read_unlock();
634 return anon_vma;
635}
636
637#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
638/*
639 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
640 * important if a PTE was dirty when it was unmapped that it's flushed
641 * before any IO is initiated on the page to prevent lost writes. Similarly,
642 * it must be flushed before freeing to prevent data leakage.
643 */
644void try_to_unmap_flush(void)
645{
646 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
647
648 if (!tlb_ubc->flush_required)
649 return;
650
651 arch_tlbbatch_flush(&tlb_ubc->arch);
652 tlb_ubc->flush_required = false;
653 tlb_ubc->writable = false;
654}
655
656/* Flush iff there are potentially writable TLB entries that can race with IO */
657void try_to_unmap_flush_dirty(void)
658{
659 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
660
661 if (tlb_ubc->writable)
662 try_to_unmap_flush();
663}
664
665/*
666 * Bits 0-14 of mm->tlb_flush_batched record pending generations.
667 * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations.
668 */
669#define TLB_FLUSH_BATCH_FLUSHED_SHIFT 16
670#define TLB_FLUSH_BATCH_PENDING_MASK \
671 ((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1)
672#define TLB_FLUSH_BATCH_PENDING_LARGE \
673 (TLB_FLUSH_BATCH_PENDING_MASK / 2)
674
675static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval,
676 unsigned long uaddr)
677{
678 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
679 int batch;
680 bool writable = pte_dirty(pteval);
681
682 if (!pte_accessible(mm, pteval))
683 return;
684
685 arch_tlbbatch_add_pending(&tlb_ubc->arch, mm, uaddr);
686 tlb_ubc->flush_required = true;
687
688 /*
689 * Ensure compiler does not re-order the setting of tlb_flush_batched
690 * before the PTE is cleared.
691 */
692 barrier();
693 batch = atomic_read(&mm->tlb_flush_batched);
694retry:
695 if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) {
696 /*
697 * Prevent `pending' from catching up with `flushed' because of
698 * overflow. Reset `pending' and `flushed' to be 1 and 0 if
699 * `pending' becomes large.
700 */
701 if (!atomic_try_cmpxchg(&mm->tlb_flush_batched, &batch, 1))
702 goto retry;
703 } else {
704 atomic_inc(&mm->tlb_flush_batched);
705 }
706
707 /*
708 * If the PTE was dirty then it's best to assume it's writable. The
709 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
710 * before the page is queued for IO.
711 */
712 if (writable)
713 tlb_ubc->writable = true;
714}
715
716/*
717 * Returns true if the TLB flush should be deferred to the end of a batch of
718 * unmap operations to reduce IPIs.
719 */
720static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
721{
722 if (!(flags & TTU_BATCH_FLUSH))
723 return false;
724
725 return arch_tlbbatch_should_defer(mm);
726}
727
728/*
729 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
730 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
731 * operation such as mprotect or munmap to race between reclaim unmapping
732 * the page and flushing the page. If this race occurs, it potentially allows
733 * access to data via a stale TLB entry. Tracking all mm's that have TLB
734 * batching in flight would be expensive during reclaim so instead track
735 * whether TLB batching occurred in the past and if so then do a flush here
736 * if required. This will cost one additional flush per reclaim cycle paid
737 * by the first operation at risk such as mprotect and mumap.
738 *
739 * This must be called under the PTL so that an access to tlb_flush_batched
740 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
741 * via the PTL.
742 */
743void flush_tlb_batched_pending(struct mm_struct *mm)
744{
745 int batch = atomic_read(&mm->tlb_flush_batched);
746 int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK;
747 int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT;
748
749 if (pending != flushed) {
750 arch_flush_tlb_batched_pending(mm);
751 /*
752 * If the new TLB flushing is pending during flushing, leave
753 * mm->tlb_flush_batched as is, to avoid losing flushing.
754 */
755 atomic_cmpxchg(&mm->tlb_flush_batched, batch,
756 pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT));
757 }
758}
759#else
760static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval,
761 unsigned long uaddr)
762{
763}
764
765static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
766{
767 return false;
768}
769#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
770
771/*
772 * At what user virtual address is page expected in vma?
773 * Caller should check the page is actually part of the vma.
774 */
775unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
776{
777 struct folio *folio = page_folio(page);
778 if (folio_test_anon(folio)) {
779 struct anon_vma *page__anon_vma = folio_anon_vma(folio);
780 /*
781 * Note: swapoff's unuse_vma() is more efficient with this
782 * check, and needs it to match anon_vma when KSM is active.
783 */
784 if (!vma->anon_vma || !page__anon_vma ||
785 vma->anon_vma->root != page__anon_vma->root)
786 return -EFAULT;
787 } else if (!vma->vm_file) {
788 return -EFAULT;
789 } else if (vma->vm_file->f_mapping != folio->mapping) {
790 return -EFAULT;
791 }
792
793 return vma_address(page, vma);
794}
795
796/*
797 * Returns the actual pmd_t* where we expect 'address' to be mapped from, or
798 * NULL if it doesn't exist. No guarantees / checks on what the pmd_t*
799 * represents.
800 */
801pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
802{
803 pgd_t *pgd;
804 p4d_t *p4d;
805 pud_t *pud;
806 pmd_t *pmd = NULL;
807
808 pgd = pgd_offset(mm, address);
809 if (!pgd_present(*pgd))
810 goto out;
811
812 p4d = p4d_offset(pgd, address);
813 if (!p4d_present(*p4d))
814 goto out;
815
816 pud = pud_offset(p4d, address);
817 if (!pud_present(*pud))
818 goto out;
819
820 pmd = pmd_offset(pud, address);
821out:
822 return pmd;
823}
824
825struct folio_referenced_arg {
826 int mapcount;
827 int referenced;
828 unsigned long vm_flags;
829 struct mem_cgroup *memcg;
830};
831
832/*
833 * arg: folio_referenced_arg will be passed
834 */
835static bool folio_referenced_one(struct folio *folio,
836 struct vm_area_struct *vma, unsigned long address, void *arg)
837{
838 struct folio_referenced_arg *pra = arg;
839 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
840 int referenced = 0;
841 unsigned long start = address, ptes = 0;
842
843 while (page_vma_mapped_walk(&pvmw)) {
844 address = pvmw.address;
845
846 if (vma->vm_flags & VM_LOCKED) {
847 if (!folio_test_large(folio) || !pvmw.pte) {
848 /* Restore the mlock which got missed */
849 mlock_vma_folio(folio, vma);
850 page_vma_mapped_walk_done(&pvmw);
851 pra->vm_flags |= VM_LOCKED;
852 return false; /* To break the loop */
853 }
854 /*
855 * For large folio fully mapped to VMA, will
856 * be handled after the pvmw loop.
857 *
858 * For large folio cross VMA boundaries, it's
859 * expected to be picked by page reclaim. But
860 * should skip reference of pages which are in
861 * the range of VM_LOCKED vma. As page reclaim
862 * should just count the reference of pages out
863 * the range of VM_LOCKED vma.
864 */
865 ptes++;
866 pra->mapcount--;
867 continue;
868 }
869
870 if (pvmw.pte) {
871 if (lru_gen_enabled() &&
872 pte_young(ptep_get(pvmw.pte))) {
873 lru_gen_look_around(&pvmw);
874 referenced++;
875 }
876
877 if (ptep_clear_flush_young_notify(vma, address,
878 pvmw.pte))
879 referenced++;
880 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
881 if (pmdp_clear_flush_young_notify(vma, address,
882 pvmw.pmd))
883 referenced++;
884 } else {
885 /* unexpected pmd-mapped folio? */
886 WARN_ON_ONCE(1);
887 }
888
889 pra->mapcount--;
890 }
891
892 if ((vma->vm_flags & VM_LOCKED) &&
893 folio_test_large(folio) &&
894 folio_within_vma(folio, vma)) {
895 unsigned long s_align, e_align;
896
897 s_align = ALIGN_DOWN(start, PMD_SIZE);
898 e_align = ALIGN_DOWN(start + folio_size(folio) - 1, PMD_SIZE);
899
900 /* folio doesn't cross page table boundary and fully mapped */
901 if ((s_align == e_align) && (ptes == folio_nr_pages(folio))) {
902 /* Restore the mlock which got missed */
903 mlock_vma_folio(folio, vma);
904 pra->vm_flags |= VM_LOCKED;
905 return false; /* To break the loop */
906 }
907 }
908
909 if (referenced)
910 folio_clear_idle(folio);
911 if (folio_test_clear_young(folio))
912 referenced++;
913
914 if (referenced) {
915 pra->referenced++;
916 pra->vm_flags |= vma->vm_flags & ~VM_LOCKED;
917 }
918
919 if (!pra->mapcount)
920 return false; /* To break the loop */
921
922 return true;
923}
924
925static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg)
926{
927 struct folio_referenced_arg *pra = arg;
928 struct mem_cgroup *memcg = pra->memcg;
929
930 /*
931 * Ignore references from this mapping if it has no recency. If the
932 * folio has been used in another mapping, we will catch it; if this
933 * other mapping is already gone, the unmap path will have set the
934 * referenced flag or activated the folio in zap_pte_range().
935 */
936 if (!vma_has_recency(vma))
937 return true;
938
939 /*
940 * If we are reclaiming on behalf of a cgroup, skip counting on behalf
941 * of references from different cgroups.
942 */
943 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
944 return true;
945
946 return false;
947}
948
949/**
950 * folio_referenced() - Test if the folio was referenced.
951 * @folio: The folio to test.
952 * @is_locked: Caller holds lock on the folio.
953 * @memcg: target memory cgroup
954 * @vm_flags: A combination of all the vma->vm_flags which referenced the folio.
955 *
956 * Quick test_and_clear_referenced for all mappings of a folio,
957 *
958 * Return: The number of mappings which referenced the folio. Return -1 if
959 * the function bailed out due to rmap lock contention.
960 */
961int folio_referenced(struct folio *folio, int is_locked,
962 struct mem_cgroup *memcg, unsigned long *vm_flags)
963{
964 int we_locked = 0;
965 struct folio_referenced_arg pra = {
966 .mapcount = folio_mapcount(folio),
967 .memcg = memcg,
968 };
969 struct rmap_walk_control rwc = {
970 .rmap_one = folio_referenced_one,
971 .arg = (void *)&pra,
972 .anon_lock = folio_lock_anon_vma_read,
973 .try_lock = true,
974 .invalid_vma = invalid_folio_referenced_vma,
975 };
976
977 *vm_flags = 0;
978 if (!pra.mapcount)
979 return 0;
980
981 if (!folio_raw_mapping(folio))
982 return 0;
983
984 if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) {
985 we_locked = folio_trylock(folio);
986 if (!we_locked)
987 return 1;
988 }
989
990 rmap_walk(folio, &rwc);
991 *vm_flags = pra.vm_flags;
992
993 if (we_locked)
994 folio_unlock(folio);
995
996 return rwc.contended ? -1 : pra.referenced;
997}
998
999static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw)
1000{
1001 int cleaned = 0;
1002 struct vm_area_struct *vma = pvmw->vma;
1003 struct mmu_notifier_range range;
1004 unsigned long address = pvmw->address;
1005
1006 /*
1007 * We have to assume the worse case ie pmd for invalidation. Note that
1008 * the folio can not be freed from this function.
1009 */
1010 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 0,
1011 vma->vm_mm, address, vma_address_end(pvmw));
1012 mmu_notifier_invalidate_range_start(&range);
1013
1014 while (page_vma_mapped_walk(pvmw)) {
1015 int ret = 0;
1016
1017 address = pvmw->address;
1018 if (pvmw->pte) {
1019 pte_t *pte = pvmw->pte;
1020 pte_t entry = ptep_get(pte);
1021
1022 if (!pte_dirty(entry) && !pte_write(entry))
1023 continue;
1024
1025 flush_cache_page(vma, address, pte_pfn(entry));
1026 entry = ptep_clear_flush(vma, address, pte);
1027 entry = pte_wrprotect(entry);
1028 entry = pte_mkclean(entry);
1029 set_pte_at(vma->vm_mm, address, pte, entry);
1030 ret = 1;
1031 } else {
1032#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1033 pmd_t *pmd = pvmw->pmd;
1034 pmd_t entry;
1035
1036 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
1037 continue;
1038
1039 flush_cache_range(vma, address,
1040 address + HPAGE_PMD_SIZE);
1041 entry = pmdp_invalidate(vma, address, pmd);
1042 entry = pmd_wrprotect(entry);
1043 entry = pmd_mkclean(entry);
1044 set_pmd_at(vma->vm_mm, address, pmd, entry);
1045 ret = 1;
1046#else
1047 /* unexpected pmd-mapped folio? */
1048 WARN_ON_ONCE(1);
1049#endif
1050 }
1051
1052 if (ret)
1053 cleaned++;
1054 }
1055
1056 mmu_notifier_invalidate_range_end(&range);
1057
1058 return cleaned;
1059}
1060
1061static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma,
1062 unsigned long address, void *arg)
1063{
1064 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC);
1065 int *cleaned = arg;
1066
1067 *cleaned += page_vma_mkclean_one(&pvmw);
1068
1069 return true;
1070}
1071
1072static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1073{
1074 if (vma->vm_flags & VM_SHARED)
1075 return false;
1076
1077 return true;
1078}
1079
1080int folio_mkclean(struct folio *folio)
1081{
1082 int cleaned = 0;
1083 struct address_space *mapping;
1084 struct rmap_walk_control rwc = {
1085 .arg = (void *)&cleaned,
1086 .rmap_one = page_mkclean_one,
1087 .invalid_vma = invalid_mkclean_vma,
1088 };
1089
1090 BUG_ON(!folio_test_locked(folio));
1091
1092 if (!folio_mapped(folio))
1093 return 0;
1094
1095 mapping = folio_mapping(folio);
1096 if (!mapping)
1097 return 0;
1098
1099 rmap_walk(folio, &rwc);
1100
1101 return cleaned;
1102}
1103EXPORT_SYMBOL_GPL(folio_mkclean);
1104
1105/**
1106 * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of
1107 * [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff)
1108 * within the @vma of shared mappings. And since clean PTEs
1109 * should also be readonly, write protects them too.
1110 * @pfn: start pfn.
1111 * @nr_pages: number of physically contiguous pages srarting with @pfn.
1112 * @pgoff: page offset that the @pfn mapped with.
1113 * @vma: vma that @pfn mapped within.
1114 *
1115 * Returns the number of cleaned PTEs (including PMDs).
1116 */
1117int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff,
1118 struct vm_area_struct *vma)
1119{
1120 struct page_vma_mapped_walk pvmw = {
1121 .pfn = pfn,
1122 .nr_pages = nr_pages,
1123 .pgoff = pgoff,
1124 .vma = vma,
1125 .flags = PVMW_SYNC,
1126 };
1127
1128 if (invalid_mkclean_vma(vma, NULL))
1129 return 0;
1130
1131 pvmw.address = vma_pgoff_address(pgoff, nr_pages, vma);
1132 VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma);
1133
1134 return page_vma_mkclean_one(&pvmw);
1135}
1136
1137int folio_total_mapcount(struct folio *folio)
1138{
1139 int mapcount = folio_entire_mapcount(folio);
1140 int nr_pages;
1141 int i;
1142
1143 /* In the common case, avoid the loop when no pages mapped by PTE */
1144 if (folio_nr_pages_mapped(folio) == 0)
1145 return mapcount;
1146 /*
1147 * Add all the PTE mappings of those pages mapped by PTE.
1148 * Limit the loop to folio_nr_pages_mapped()?
1149 * Perhaps: given all the raciness, that may be a good or a bad idea.
1150 */
1151 nr_pages = folio_nr_pages(folio);
1152 for (i = 0; i < nr_pages; i++)
1153 mapcount += atomic_read(&folio_page(folio, i)->_mapcount);
1154
1155 /* But each of those _mapcounts was based on -1 */
1156 mapcount += nr_pages;
1157 return mapcount;
1158}
1159
1160static __always_inline unsigned int __folio_add_rmap(struct folio *folio,
1161 struct page *page, int nr_pages, enum rmap_level level,
1162 int *nr_pmdmapped)
1163{
1164 atomic_t *mapped = &folio->_nr_pages_mapped;
1165 int first, nr = 0;
1166
1167 __folio_rmap_sanity_checks(folio, page, nr_pages, level);
1168
1169 switch (level) {
1170 case RMAP_LEVEL_PTE:
1171 do {
1172 first = atomic_inc_and_test(&page->_mapcount);
1173 if (first && folio_test_large(folio)) {
1174 first = atomic_inc_return_relaxed(mapped);
1175 first = (first < ENTIRELY_MAPPED);
1176 }
1177
1178 if (first)
1179 nr++;
1180 } while (page++, --nr_pages > 0);
1181 break;
1182 case RMAP_LEVEL_PMD:
1183 first = atomic_inc_and_test(&folio->_entire_mapcount);
1184 if (first) {
1185 nr = atomic_add_return_relaxed(ENTIRELY_MAPPED, mapped);
1186 if (likely(nr < ENTIRELY_MAPPED + ENTIRELY_MAPPED)) {
1187 *nr_pmdmapped = folio_nr_pages(folio);
1188 nr = *nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
1189 /* Raced ahead of a remove and another add? */
1190 if (unlikely(nr < 0))
1191 nr = 0;
1192 } else {
1193 /* Raced ahead of a remove of ENTIRELY_MAPPED */
1194 nr = 0;
1195 }
1196 }
1197 break;
1198 }
1199 return nr;
1200}
1201
1202/**
1203 * folio_move_anon_rmap - move a folio to our anon_vma
1204 * @folio: The folio to move to our anon_vma
1205 * @vma: The vma the folio belongs to
1206 *
1207 * When a folio belongs exclusively to one process after a COW event,
1208 * that folio can be moved into the anon_vma that belongs to just that
1209 * process, so the rmap code will not search the parent or sibling processes.
1210 */
1211void folio_move_anon_rmap(struct folio *folio, struct vm_area_struct *vma)
1212{
1213 void *anon_vma = vma->anon_vma;
1214
1215 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1216 VM_BUG_ON_VMA(!anon_vma, vma);
1217
1218 anon_vma += PAGE_MAPPING_ANON;
1219 /*
1220 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1221 * simultaneously, so a concurrent reader (eg folio_referenced()'s
1222 * folio_test_anon()) will not see one without the other.
1223 */
1224 WRITE_ONCE(folio->mapping, anon_vma);
1225}
1226
1227/**
1228 * __folio_set_anon - set up a new anonymous rmap for a folio
1229 * @folio: The folio to set up the new anonymous rmap for.
1230 * @vma: VM area to add the folio to.
1231 * @address: User virtual address of the mapping
1232 * @exclusive: Whether the folio is exclusive to the process.
1233 */
1234static void __folio_set_anon(struct folio *folio, struct vm_area_struct *vma,
1235 unsigned long address, bool exclusive)
1236{
1237 struct anon_vma *anon_vma = vma->anon_vma;
1238
1239 BUG_ON(!anon_vma);
1240
1241 /*
1242 * If the folio isn't exclusive to this vma, we must use the _oldest_
1243 * possible anon_vma for the folio mapping!
1244 */
1245 if (!exclusive)
1246 anon_vma = anon_vma->root;
1247
1248 /*
1249 * page_idle does a lockless/optimistic rmap scan on folio->mapping.
1250 * Make sure the compiler doesn't split the stores of anon_vma and
1251 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1252 * could mistake the mapping for a struct address_space and crash.
1253 */
1254 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1255 WRITE_ONCE(folio->mapping, (struct address_space *) anon_vma);
1256 folio->index = linear_page_index(vma, address);
1257}
1258
1259/**
1260 * __page_check_anon_rmap - sanity check anonymous rmap addition
1261 * @folio: The folio containing @page.
1262 * @page: the page to check the mapping of
1263 * @vma: the vm area in which the mapping is added
1264 * @address: the user virtual address mapped
1265 */
1266static void __page_check_anon_rmap(struct folio *folio, struct page *page,
1267 struct vm_area_struct *vma, unsigned long address)
1268{
1269 /*
1270 * The page's anon-rmap details (mapping and index) are guaranteed to
1271 * be set up correctly at this point.
1272 *
1273 * We have exclusion against folio_add_anon_rmap_*() because the caller
1274 * always holds the page locked.
1275 *
1276 * We have exclusion against folio_add_new_anon_rmap because those pages
1277 * are initially only visible via the pagetables, and the pte is locked
1278 * over the call to folio_add_new_anon_rmap.
1279 */
1280 VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root,
1281 folio);
1282 VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1283 page);
1284}
1285
1286static __always_inline void __folio_add_anon_rmap(struct folio *folio,
1287 struct page *page, int nr_pages, struct vm_area_struct *vma,
1288 unsigned long address, rmap_t flags, enum rmap_level level)
1289{
1290 int i, nr, nr_pmdmapped = 0;
1291
1292 nr = __folio_add_rmap(folio, page, nr_pages, level, &nr_pmdmapped);
1293 if (nr_pmdmapped)
1294 __lruvec_stat_mod_folio(folio, NR_ANON_THPS, nr_pmdmapped);
1295 if (nr)
1296 __lruvec_stat_mod_folio(folio, NR_ANON_MAPPED, nr);
1297
1298 if (unlikely(!folio_test_anon(folio))) {
1299 VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
1300 /*
1301 * For a PTE-mapped large folio, we only know that the single
1302 * PTE is exclusive. Further, __folio_set_anon() might not get
1303 * folio->index right when not given the address of the head
1304 * page.
1305 */
1306 VM_WARN_ON_FOLIO(folio_test_large(folio) &&
1307 level != RMAP_LEVEL_PMD, folio);
1308 __folio_set_anon(folio, vma, address,
1309 !!(flags & RMAP_EXCLUSIVE));
1310 } else if (likely(!folio_test_ksm(folio))) {
1311 __page_check_anon_rmap(folio, page, vma, address);
1312 }
1313
1314 if (flags & RMAP_EXCLUSIVE) {
1315 switch (level) {
1316 case RMAP_LEVEL_PTE:
1317 for (i = 0; i < nr_pages; i++)
1318 SetPageAnonExclusive(page + i);
1319 break;
1320 case RMAP_LEVEL_PMD:
1321 SetPageAnonExclusive(page);
1322 break;
1323 }
1324 }
1325 for (i = 0; i < nr_pages; i++) {
1326 struct page *cur_page = page + i;
1327
1328 /* While PTE-mapping a THP we have a PMD and a PTE mapping. */
1329 VM_WARN_ON_FOLIO((atomic_read(&cur_page->_mapcount) > 0 ||
1330 (folio_test_large(folio) &&
1331 folio_entire_mapcount(folio) > 1)) &&
1332 PageAnonExclusive(cur_page), folio);
1333 }
1334
1335 /*
1336 * For large folio, only mlock it if it's fully mapped to VMA. It's
1337 * not easy to check whether the large folio is fully mapped to VMA
1338 * here. Only mlock normal 4K folio and leave page reclaim to handle
1339 * large folio.
1340 */
1341 if (!folio_test_large(folio))
1342 mlock_vma_folio(folio, vma);
1343}
1344
1345/**
1346 * folio_add_anon_rmap_ptes - add PTE mappings to a page range of an anon folio
1347 * @folio: The folio to add the mappings to
1348 * @page: The first page to add
1349 * @nr_pages: The number of pages which will be mapped
1350 * @vma: The vm area in which the mappings are added
1351 * @address: The user virtual address of the first page to map
1352 * @flags: The rmap flags
1353 *
1354 * The page range of folio is defined by [first_page, first_page + nr_pages)
1355 *
1356 * The caller needs to hold the page table lock, and the page must be locked in
1357 * the anon_vma case: to serialize mapping,index checking after setting,
1358 * and to ensure that an anon folio is not being upgraded racily to a KSM folio
1359 * (but KSM folios are never downgraded).
1360 */
1361void folio_add_anon_rmap_ptes(struct folio *folio, struct page *page,
1362 int nr_pages, struct vm_area_struct *vma, unsigned long address,
1363 rmap_t flags)
1364{
1365 __folio_add_anon_rmap(folio, page, nr_pages, vma, address, flags,
1366 RMAP_LEVEL_PTE);
1367}
1368
1369/**
1370 * folio_add_anon_rmap_pmd - add a PMD mapping to a page range of an anon folio
1371 * @folio: The folio to add the mapping to
1372 * @page: The first page to add
1373 * @vma: The vm area in which the mapping is added
1374 * @address: The user virtual address of the first page to map
1375 * @flags: The rmap flags
1376 *
1377 * The page range of folio is defined by [first_page, first_page + HPAGE_PMD_NR)
1378 *
1379 * The caller needs to hold the page table lock, and the page must be locked in
1380 * the anon_vma case: to serialize mapping,index checking after setting.
1381 */
1382void folio_add_anon_rmap_pmd(struct folio *folio, struct page *page,
1383 struct vm_area_struct *vma, unsigned long address, rmap_t flags)
1384{
1385#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1386 __folio_add_anon_rmap(folio, page, HPAGE_PMD_NR, vma, address, flags,
1387 RMAP_LEVEL_PMD);
1388#else
1389 WARN_ON_ONCE(true);
1390#endif
1391}
1392
1393/**
1394 * folio_add_new_anon_rmap - Add mapping to a new anonymous folio.
1395 * @folio: The folio to add the mapping to.
1396 * @vma: the vm area in which the mapping is added
1397 * @address: the user virtual address mapped
1398 *
1399 * Like folio_add_anon_rmap_*() but must only be called on *new* folios.
1400 * This means the inc-and-test can be bypassed.
1401 * The folio does not have to be locked.
1402 *
1403 * If the folio is pmd-mappable, it is accounted as a THP. As the folio
1404 * is new, it's assumed to be mapped exclusively by a single process.
1405 */
1406void folio_add_new_anon_rmap(struct folio *folio, struct vm_area_struct *vma,
1407 unsigned long address)
1408{
1409 int nr = folio_nr_pages(folio);
1410
1411 VM_WARN_ON_FOLIO(folio_test_hugetlb(folio), folio);
1412 VM_BUG_ON_VMA(address < vma->vm_start ||
1413 address + (nr << PAGE_SHIFT) > vma->vm_end, vma);
1414 __folio_set_swapbacked(folio);
1415 __folio_set_anon(folio, vma, address, true);
1416
1417 if (likely(!folio_test_large(folio))) {
1418 /* increment count (starts at -1) */
1419 atomic_set(&folio->_mapcount, 0);
1420 SetPageAnonExclusive(&folio->page);
1421 } else if (!folio_test_pmd_mappable(folio)) {
1422 int i;
1423
1424 for (i = 0; i < nr; i++) {
1425 struct page *page = folio_page(folio, i);
1426
1427 /* increment count (starts at -1) */
1428 atomic_set(&page->_mapcount, 0);
1429 SetPageAnonExclusive(page);
1430 }
1431
1432 atomic_set(&folio->_nr_pages_mapped, nr);
1433 } else {
1434 /* increment count (starts at -1) */
1435 atomic_set(&folio->_entire_mapcount, 0);
1436 atomic_set(&folio->_nr_pages_mapped, ENTIRELY_MAPPED);
1437 SetPageAnonExclusive(&folio->page);
1438 __lruvec_stat_mod_folio(folio, NR_ANON_THPS, nr);
1439 }
1440
1441 __lruvec_stat_mod_folio(folio, NR_ANON_MAPPED, nr);
1442}
1443
1444static __always_inline void __folio_add_file_rmap(struct folio *folio,
1445 struct page *page, int nr_pages, struct vm_area_struct *vma,
1446 enum rmap_level level)
1447{
1448 int nr, nr_pmdmapped = 0;
1449
1450 VM_WARN_ON_FOLIO(folio_test_anon(folio), folio);
1451
1452 nr = __folio_add_rmap(folio, page, nr_pages, level, &nr_pmdmapped);
1453 if (nr_pmdmapped)
1454 __lruvec_stat_mod_folio(folio, folio_test_swapbacked(folio) ?
1455 NR_SHMEM_PMDMAPPED : NR_FILE_PMDMAPPED, nr_pmdmapped);
1456 if (nr)
1457 __lruvec_stat_mod_folio(folio, NR_FILE_MAPPED, nr);
1458
1459 /* See comments in folio_add_anon_rmap_*() */
1460 if (!folio_test_large(folio))
1461 mlock_vma_folio(folio, vma);
1462}
1463
1464/**
1465 * folio_add_file_rmap_ptes - add PTE mappings to a page range of a folio
1466 * @folio: The folio to add the mappings to
1467 * @page: The first page to add
1468 * @nr_pages: The number of pages that will be mapped using PTEs
1469 * @vma: The vm area in which the mappings are added
1470 *
1471 * The page range of the folio is defined by [page, page + nr_pages)
1472 *
1473 * The caller needs to hold the page table lock.
1474 */
1475void folio_add_file_rmap_ptes(struct folio *folio, struct page *page,
1476 int nr_pages, struct vm_area_struct *vma)
1477{
1478 __folio_add_file_rmap(folio, page, nr_pages, vma, RMAP_LEVEL_PTE);
1479}
1480
1481/**
1482 * folio_add_file_rmap_pmd - add a PMD mapping to a page range of a folio
1483 * @folio: The folio to add the mapping to
1484 * @page: The first page to add
1485 * @vma: The vm area in which the mapping is added
1486 *
1487 * The page range of the folio is defined by [page, page + HPAGE_PMD_NR)
1488 *
1489 * The caller needs to hold the page table lock.
1490 */
1491void folio_add_file_rmap_pmd(struct folio *folio, struct page *page,
1492 struct vm_area_struct *vma)
1493{
1494#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1495 __folio_add_file_rmap(folio, page, HPAGE_PMD_NR, vma, RMAP_LEVEL_PMD);
1496#else
1497 WARN_ON_ONCE(true);
1498#endif
1499}
1500
1501static __always_inline void __folio_remove_rmap(struct folio *folio,
1502 struct page *page, int nr_pages, struct vm_area_struct *vma,
1503 enum rmap_level level)
1504{
1505 atomic_t *mapped = &folio->_nr_pages_mapped;
1506 int last, nr = 0, nr_pmdmapped = 0;
1507 enum node_stat_item idx;
1508
1509 __folio_rmap_sanity_checks(folio, page, nr_pages, level);
1510
1511 switch (level) {
1512 case RMAP_LEVEL_PTE:
1513 do {
1514 last = atomic_add_negative(-1, &page->_mapcount);
1515 if (last && folio_test_large(folio)) {
1516 last = atomic_dec_return_relaxed(mapped);
1517 last = (last < ENTIRELY_MAPPED);
1518 }
1519
1520 if (last)
1521 nr++;
1522 } while (page++, --nr_pages > 0);
1523 break;
1524 case RMAP_LEVEL_PMD:
1525 last = atomic_add_negative(-1, &folio->_entire_mapcount);
1526 if (last) {
1527 nr = atomic_sub_return_relaxed(ENTIRELY_MAPPED, mapped);
1528 if (likely(nr < ENTIRELY_MAPPED)) {
1529 nr_pmdmapped = folio_nr_pages(folio);
1530 nr = nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
1531 /* Raced ahead of another remove and an add? */
1532 if (unlikely(nr < 0))
1533 nr = 0;
1534 } else {
1535 /* An add of ENTIRELY_MAPPED raced ahead */
1536 nr = 0;
1537 }
1538 }
1539 break;
1540 }
1541
1542 if (nr_pmdmapped) {
1543 if (folio_test_anon(folio))
1544 idx = NR_ANON_THPS;
1545 else if (folio_test_swapbacked(folio))
1546 idx = NR_SHMEM_PMDMAPPED;
1547 else
1548 idx = NR_FILE_PMDMAPPED;
1549 __lruvec_stat_mod_folio(folio, idx, -nr_pmdmapped);
1550 }
1551 if (nr) {
1552 idx = folio_test_anon(folio) ? NR_ANON_MAPPED : NR_FILE_MAPPED;
1553 __lruvec_stat_mod_folio(folio, idx, -nr);
1554
1555 /*
1556 * Queue anon large folio for deferred split if at least one
1557 * page of the folio is unmapped and at least one page
1558 * is still mapped.
1559 */
1560 if (folio_test_large(folio) && folio_test_anon(folio))
1561 if (level == RMAP_LEVEL_PTE || nr < nr_pmdmapped)
1562 deferred_split_folio(folio);
1563 }
1564
1565 /*
1566 * It would be tidy to reset folio_test_anon mapping when fully
1567 * unmapped, but that might overwrite a racing folio_add_anon_rmap_*()
1568 * which increments mapcount after us but sets mapping before us:
1569 * so leave the reset to free_pages_prepare, and remember that
1570 * it's only reliable while mapped.
1571 */
1572
1573 munlock_vma_folio(folio, vma);
1574}
1575
1576/**
1577 * folio_remove_rmap_ptes - remove PTE mappings from a page range of a folio
1578 * @folio: The folio to remove the mappings from
1579 * @page: The first page to remove
1580 * @nr_pages: The number of pages that will be removed from the mapping
1581 * @vma: The vm area from which the mappings are removed
1582 *
1583 * The page range of the folio is defined by [page, page + nr_pages)
1584 *
1585 * The caller needs to hold the page table lock.
1586 */
1587void folio_remove_rmap_ptes(struct folio *folio, struct page *page,
1588 int nr_pages, struct vm_area_struct *vma)
1589{
1590 __folio_remove_rmap(folio, page, nr_pages, vma, RMAP_LEVEL_PTE);
1591}
1592
1593/**
1594 * folio_remove_rmap_pmd - remove a PMD mapping from a page range of a folio
1595 * @folio: The folio to remove the mapping from
1596 * @page: The first page to remove
1597 * @vma: The vm area from which the mapping is removed
1598 *
1599 * The page range of the folio is defined by [page, page + HPAGE_PMD_NR)
1600 *
1601 * The caller needs to hold the page table lock.
1602 */
1603void folio_remove_rmap_pmd(struct folio *folio, struct page *page,
1604 struct vm_area_struct *vma)
1605{
1606#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1607 __folio_remove_rmap(folio, page, HPAGE_PMD_NR, vma, RMAP_LEVEL_PMD);
1608#else
1609 WARN_ON_ONCE(true);
1610#endif
1611}
1612
1613/*
1614 * @arg: enum ttu_flags will be passed to this argument
1615 */
1616static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
1617 unsigned long address, void *arg)
1618{
1619 struct mm_struct *mm = vma->vm_mm;
1620 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1621 pte_t pteval;
1622 struct page *subpage;
1623 bool anon_exclusive, ret = true;
1624 struct mmu_notifier_range range;
1625 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1626 unsigned long pfn;
1627 unsigned long hsz = 0;
1628
1629 /*
1630 * When racing against e.g. zap_pte_range() on another cpu,
1631 * in between its ptep_get_and_clear_full() and folio_remove_rmap_*(),
1632 * try_to_unmap() may return before page_mapped() has become false,
1633 * if page table locking is skipped: use TTU_SYNC to wait for that.
1634 */
1635 if (flags & TTU_SYNC)
1636 pvmw.flags = PVMW_SYNC;
1637
1638 if (flags & TTU_SPLIT_HUGE_PMD)
1639 split_huge_pmd_address(vma, address, false, folio);
1640
1641 /*
1642 * For THP, we have to assume the worse case ie pmd for invalidation.
1643 * For hugetlb, it could be much worse if we need to do pud
1644 * invalidation in the case of pmd sharing.
1645 *
1646 * Note that the folio can not be freed in this function as call of
1647 * try_to_unmap() must hold a reference on the folio.
1648 */
1649 range.end = vma_address_end(&pvmw);
1650 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1651 address, range.end);
1652 if (folio_test_hugetlb(folio)) {
1653 /*
1654 * If sharing is possible, start and end will be adjusted
1655 * accordingly.
1656 */
1657 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1658 &range.end);
1659
1660 /* We need the huge page size for set_huge_pte_at() */
1661 hsz = huge_page_size(hstate_vma(vma));
1662 }
1663 mmu_notifier_invalidate_range_start(&range);
1664
1665 while (page_vma_mapped_walk(&pvmw)) {
1666 /* Unexpected PMD-mapped THP? */
1667 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1668
1669 /*
1670 * If the folio is in an mlock()d vma, we must not swap it out.
1671 */
1672 if (!(flags & TTU_IGNORE_MLOCK) &&
1673 (vma->vm_flags & VM_LOCKED)) {
1674 /* Restore the mlock which got missed */
1675 if (!folio_test_large(folio))
1676 mlock_vma_folio(folio, vma);
1677 page_vma_mapped_walk_done(&pvmw);
1678 ret = false;
1679 break;
1680 }
1681
1682 pfn = pte_pfn(ptep_get(pvmw.pte));
1683 subpage = folio_page(folio, pfn - folio_pfn(folio));
1684 address = pvmw.address;
1685 anon_exclusive = folio_test_anon(folio) &&
1686 PageAnonExclusive(subpage);
1687
1688 if (folio_test_hugetlb(folio)) {
1689 bool anon = folio_test_anon(folio);
1690
1691 /*
1692 * The try_to_unmap() is only passed a hugetlb page
1693 * in the case where the hugetlb page is poisoned.
1694 */
1695 VM_BUG_ON_PAGE(!PageHWPoison(subpage), subpage);
1696 /*
1697 * huge_pmd_unshare may unmap an entire PMD page.
1698 * There is no way of knowing exactly which PMDs may
1699 * be cached for this mm, so we must flush them all.
1700 * start/end were already adjusted above to cover this
1701 * range.
1702 */
1703 flush_cache_range(vma, range.start, range.end);
1704
1705 /*
1706 * To call huge_pmd_unshare, i_mmap_rwsem must be
1707 * held in write mode. Caller needs to explicitly
1708 * do this outside rmap routines.
1709 *
1710 * We also must hold hugetlb vma_lock in write mode.
1711 * Lock order dictates acquiring vma_lock BEFORE
1712 * i_mmap_rwsem. We can only try lock here and fail
1713 * if unsuccessful.
1714 */
1715 if (!anon) {
1716 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1717 if (!hugetlb_vma_trylock_write(vma)) {
1718 page_vma_mapped_walk_done(&pvmw);
1719 ret = false;
1720 break;
1721 }
1722 if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1723 hugetlb_vma_unlock_write(vma);
1724 flush_tlb_range(vma,
1725 range.start, range.end);
1726 /*
1727 * The ref count of the PMD page was
1728 * dropped which is part of the way map
1729 * counting is done for shared PMDs.
1730 * Return 'true' here. When there is
1731 * no other sharing, huge_pmd_unshare
1732 * returns false and we will unmap the
1733 * actual page and drop map count
1734 * to zero.
1735 */
1736 page_vma_mapped_walk_done(&pvmw);
1737 break;
1738 }
1739 hugetlb_vma_unlock_write(vma);
1740 }
1741 pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
1742 } else {
1743 flush_cache_page(vma, address, pfn);
1744 /* Nuke the page table entry. */
1745 if (should_defer_flush(mm, flags)) {
1746 /*
1747 * We clear the PTE but do not flush so potentially
1748 * a remote CPU could still be writing to the folio.
1749 * If the entry was previously clean then the
1750 * architecture must guarantee that a clear->dirty
1751 * transition on a cached TLB entry is written through
1752 * and traps if the PTE is unmapped.
1753 */
1754 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1755
1756 set_tlb_ubc_flush_pending(mm, pteval, address);
1757 } else {
1758 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1759 }
1760 }
1761
1762 /*
1763 * Now the pte is cleared. If this pte was uffd-wp armed,
1764 * we may want to replace a none pte with a marker pte if
1765 * it's file-backed, so we don't lose the tracking info.
1766 */
1767 pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval);
1768
1769 /* Set the dirty flag on the folio now the pte is gone. */
1770 if (pte_dirty(pteval))
1771 folio_mark_dirty(folio);
1772
1773 /* Update high watermark before we lower rss */
1774 update_hiwater_rss(mm);
1775
1776 if (PageHWPoison(subpage) && (flags & TTU_HWPOISON)) {
1777 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1778 if (folio_test_hugetlb(folio)) {
1779 hugetlb_count_sub(folio_nr_pages(folio), mm);
1780 set_huge_pte_at(mm, address, pvmw.pte, pteval,
1781 hsz);
1782 } else {
1783 dec_mm_counter(mm, mm_counter(&folio->page));
1784 set_pte_at(mm, address, pvmw.pte, pteval);
1785 }
1786
1787 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1788 /*
1789 * The guest indicated that the page content is of no
1790 * interest anymore. Simply discard the pte, vmscan
1791 * will take care of the rest.
1792 * A future reference will then fault in a new zero
1793 * page. When userfaultfd is active, we must not drop
1794 * this page though, as its main user (postcopy
1795 * migration) will not expect userfaults on already
1796 * copied pages.
1797 */
1798 dec_mm_counter(mm, mm_counter(&folio->page));
1799 } else if (folio_test_anon(folio)) {
1800 swp_entry_t entry = page_swap_entry(subpage);
1801 pte_t swp_pte;
1802 /*
1803 * Store the swap location in the pte.
1804 * See handle_pte_fault() ...
1805 */
1806 if (unlikely(folio_test_swapbacked(folio) !=
1807 folio_test_swapcache(folio))) {
1808 WARN_ON_ONCE(1);
1809 ret = false;
1810 page_vma_mapped_walk_done(&pvmw);
1811 break;
1812 }
1813
1814 /* MADV_FREE page check */
1815 if (!folio_test_swapbacked(folio)) {
1816 int ref_count, map_count;
1817
1818 /*
1819 * Synchronize with gup_pte_range():
1820 * - clear PTE; barrier; read refcount
1821 * - inc refcount; barrier; read PTE
1822 */
1823 smp_mb();
1824
1825 ref_count = folio_ref_count(folio);
1826 map_count = folio_mapcount(folio);
1827
1828 /*
1829 * Order reads for page refcount and dirty flag
1830 * (see comments in __remove_mapping()).
1831 */
1832 smp_rmb();
1833
1834 /*
1835 * The only page refs must be one from isolation
1836 * plus the rmap(s) (dropped by discard:).
1837 */
1838 if (ref_count == 1 + map_count &&
1839 !folio_test_dirty(folio)) {
1840 dec_mm_counter(mm, MM_ANONPAGES);
1841 goto discard;
1842 }
1843
1844 /*
1845 * If the folio was redirtied, it cannot be
1846 * discarded. Remap the page to page table.
1847 */
1848 set_pte_at(mm, address, pvmw.pte, pteval);
1849 folio_set_swapbacked(folio);
1850 ret = false;
1851 page_vma_mapped_walk_done(&pvmw);
1852 break;
1853 }
1854
1855 if (swap_duplicate(entry) < 0) {
1856 set_pte_at(mm, address, pvmw.pte, pteval);
1857 ret = false;
1858 page_vma_mapped_walk_done(&pvmw);
1859 break;
1860 }
1861 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1862 swap_free(entry);
1863 set_pte_at(mm, address, pvmw.pte, pteval);
1864 ret = false;
1865 page_vma_mapped_walk_done(&pvmw);
1866 break;
1867 }
1868
1869 /* See folio_try_share_anon_rmap(): clear PTE first. */
1870 if (anon_exclusive &&
1871 folio_try_share_anon_rmap_pte(folio, subpage)) {
1872 swap_free(entry);
1873 set_pte_at(mm, address, pvmw.pte, pteval);
1874 ret = false;
1875 page_vma_mapped_walk_done(&pvmw);
1876 break;
1877 }
1878 if (list_empty(&mm->mmlist)) {
1879 spin_lock(&mmlist_lock);
1880 if (list_empty(&mm->mmlist))
1881 list_add(&mm->mmlist, &init_mm.mmlist);
1882 spin_unlock(&mmlist_lock);
1883 }
1884 dec_mm_counter(mm, MM_ANONPAGES);
1885 inc_mm_counter(mm, MM_SWAPENTS);
1886 swp_pte = swp_entry_to_pte(entry);
1887 if (anon_exclusive)
1888 swp_pte = pte_swp_mkexclusive(swp_pte);
1889 if (pte_soft_dirty(pteval))
1890 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1891 if (pte_uffd_wp(pteval))
1892 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1893 set_pte_at(mm, address, pvmw.pte, swp_pte);
1894 } else {
1895 /*
1896 * This is a locked file-backed folio,
1897 * so it cannot be removed from the page
1898 * cache and replaced by a new folio before
1899 * mmu_notifier_invalidate_range_end, so no
1900 * concurrent thread might update its page table
1901 * to point at a new folio while a device is
1902 * still using this folio.
1903 *
1904 * See Documentation/mm/mmu_notifier.rst
1905 */
1906 dec_mm_counter(mm, mm_counter_file(&folio->page));
1907 }
1908discard:
1909 if (unlikely(folio_test_hugetlb(folio)))
1910 hugetlb_remove_rmap(folio);
1911 else
1912 folio_remove_rmap_pte(folio, subpage, vma);
1913 if (vma->vm_flags & VM_LOCKED)
1914 mlock_drain_local();
1915 folio_put(folio);
1916 }
1917
1918 mmu_notifier_invalidate_range_end(&range);
1919
1920 return ret;
1921}
1922
1923static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1924{
1925 return vma_is_temporary_stack(vma);
1926}
1927
1928static int folio_not_mapped(struct folio *folio)
1929{
1930 return !folio_mapped(folio);
1931}
1932
1933/**
1934 * try_to_unmap - Try to remove all page table mappings to a folio.
1935 * @folio: The folio to unmap.
1936 * @flags: action and flags
1937 *
1938 * Tries to remove all the page table entries which are mapping this
1939 * folio. It is the caller's responsibility to check if the folio is
1940 * still mapped if needed (use TTU_SYNC to prevent accounting races).
1941 *
1942 * Context: Caller must hold the folio lock.
1943 */
1944void try_to_unmap(struct folio *folio, enum ttu_flags flags)
1945{
1946 struct rmap_walk_control rwc = {
1947 .rmap_one = try_to_unmap_one,
1948 .arg = (void *)flags,
1949 .done = folio_not_mapped,
1950 .anon_lock = folio_lock_anon_vma_read,
1951 };
1952
1953 if (flags & TTU_RMAP_LOCKED)
1954 rmap_walk_locked(folio, &rwc);
1955 else
1956 rmap_walk(folio, &rwc);
1957}
1958
1959/*
1960 * @arg: enum ttu_flags will be passed to this argument.
1961 *
1962 * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
1963 * containing migration entries.
1964 */
1965static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
1966 unsigned long address, void *arg)
1967{
1968 struct mm_struct *mm = vma->vm_mm;
1969 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1970 pte_t pteval;
1971 struct page *subpage;
1972 bool anon_exclusive, ret = true;
1973 struct mmu_notifier_range range;
1974 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1975 unsigned long pfn;
1976 unsigned long hsz = 0;
1977
1978 /*
1979 * When racing against e.g. zap_pte_range() on another cpu,
1980 * in between its ptep_get_and_clear_full() and folio_remove_rmap_*(),
1981 * try_to_migrate() may return before page_mapped() has become false,
1982 * if page table locking is skipped: use TTU_SYNC to wait for that.
1983 */
1984 if (flags & TTU_SYNC)
1985 pvmw.flags = PVMW_SYNC;
1986
1987 /*
1988 * unmap_page() in mm/huge_memory.c is the only user of migration with
1989 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
1990 */
1991 if (flags & TTU_SPLIT_HUGE_PMD)
1992 split_huge_pmd_address(vma, address, true, folio);
1993
1994 /*
1995 * For THP, we have to assume the worse case ie pmd for invalidation.
1996 * For hugetlb, it could be much worse if we need to do pud
1997 * invalidation in the case of pmd sharing.
1998 *
1999 * Note that the page can not be free in this function as call of
2000 * try_to_unmap() must hold a reference on the page.
2001 */
2002 range.end = vma_address_end(&pvmw);
2003 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2004 address, range.end);
2005 if (folio_test_hugetlb(folio)) {
2006 /*
2007 * If sharing is possible, start and end will be adjusted
2008 * accordingly.
2009 */
2010 adjust_range_if_pmd_sharing_possible(vma, &range.start,
2011 &range.end);
2012
2013 /* We need the huge page size for set_huge_pte_at() */
2014 hsz = huge_page_size(hstate_vma(vma));
2015 }
2016 mmu_notifier_invalidate_range_start(&range);
2017
2018 while (page_vma_mapped_walk(&pvmw)) {
2019#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2020 /* PMD-mapped THP migration entry */
2021 if (!pvmw.pte) {
2022 subpage = folio_page(folio,
2023 pmd_pfn(*pvmw.pmd) - folio_pfn(folio));
2024 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
2025 !folio_test_pmd_mappable(folio), folio);
2026
2027 if (set_pmd_migration_entry(&pvmw, subpage)) {
2028 ret = false;
2029 page_vma_mapped_walk_done(&pvmw);
2030 break;
2031 }
2032 continue;
2033 }
2034#endif
2035
2036 /* Unexpected PMD-mapped THP? */
2037 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2038
2039 pfn = pte_pfn(ptep_get(pvmw.pte));
2040
2041 if (folio_is_zone_device(folio)) {
2042 /*
2043 * Our PTE is a non-present device exclusive entry and
2044 * calculating the subpage as for the common case would
2045 * result in an invalid pointer.
2046 *
2047 * Since only PAGE_SIZE pages can currently be
2048 * migrated, just set it to page. This will need to be
2049 * changed when hugepage migrations to device private
2050 * memory are supported.
2051 */
2052 VM_BUG_ON_FOLIO(folio_nr_pages(folio) > 1, folio);
2053 subpage = &folio->page;
2054 } else {
2055 subpage = folio_page(folio, pfn - folio_pfn(folio));
2056 }
2057 address = pvmw.address;
2058 anon_exclusive = folio_test_anon(folio) &&
2059 PageAnonExclusive(subpage);
2060
2061 if (folio_test_hugetlb(folio)) {
2062 bool anon = folio_test_anon(folio);
2063
2064 /*
2065 * huge_pmd_unshare may unmap an entire PMD page.
2066 * There is no way of knowing exactly which PMDs may
2067 * be cached for this mm, so we must flush them all.
2068 * start/end were already adjusted above to cover this
2069 * range.
2070 */
2071 flush_cache_range(vma, range.start, range.end);
2072
2073 /*
2074 * To call huge_pmd_unshare, i_mmap_rwsem must be
2075 * held in write mode. Caller needs to explicitly
2076 * do this outside rmap routines.
2077 *
2078 * We also must hold hugetlb vma_lock in write mode.
2079 * Lock order dictates acquiring vma_lock BEFORE
2080 * i_mmap_rwsem. We can only try lock here and
2081 * fail if unsuccessful.
2082 */
2083 if (!anon) {
2084 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
2085 if (!hugetlb_vma_trylock_write(vma)) {
2086 page_vma_mapped_walk_done(&pvmw);
2087 ret = false;
2088 break;
2089 }
2090 if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
2091 hugetlb_vma_unlock_write(vma);
2092 flush_tlb_range(vma,
2093 range.start, range.end);
2094
2095 /*
2096 * The ref count of the PMD page was
2097 * dropped which is part of the way map
2098 * counting is done for shared PMDs.
2099 * Return 'true' here. When there is
2100 * no other sharing, huge_pmd_unshare
2101 * returns false and we will unmap the
2102 * actual page and drop map count
2103 * to zero.
2104 */
2105 page_vma_mapped_walk_done(&pvmw);
2106 break;
2107 }
2108 hugetlb_vma_unlock_write(vma);
2109 }
2110 /* Nuke the hugetlb page table entry */
2111 pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
2112 } else {
2113 flush_cache_page(vma, address, pfn);
2114 /* Nuke the page table entry. */
2115 if (should_defer_flush(mm, flags)) {
2116 /*
2117 * We clear the PTE but do not flush so potentially
2118 * a remote CPU could still be writing to the folio.
2119 * If the entry was previously clean then the
2120 * architecture must guarantee that a clear->dirty
2121 * transition on a cached TLB entry is written through
2122 * and traps if the PTE is unmapped.
2123 */
2124 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
2125
2126 set_tlb_ubc_flush_pending(mm, pteval, address);
2127 } else {
2128 pteval = ptep_clear_flush(vma, address, pvmw.pte);
2129 }
2130 }
2131
2132 /* Set the dirty flag on the folio now the pte is gone. */
2133 if (pte_dirty(pteval))
2134 folio_mark_dirty(folio);
2135
2136 /* Update high watermark before we lower rss */
2137 update_hiwater_rss(mm);
2138
2139 if (folio_is_device_private(folio)) {
2140 unsigned long pfn = folio_pfn(folio);
2141 swp_entry_t entry;
2142 pte_t swp_pte;
2143
2144 if (anon_exclusive)
2145 WARN_ON_ONCE(folio_try_share_anon_rmap_pte(folio,
2146 subpage));
2147
2148 /*
2149 * Store the pfn of the page in a special migration
2150 * pte. do_swap_page() will wait until the migration
2151 * pte is removed and then restart fault handling.
2152 */
2153 entry = pte_to_swp_entry(pteval);
2154 if (is_writable_device_private_entry(entry))
2155 entry = make_writable_migration_entry(pfn);
2156 else if (anon_exclusive)
2157 entry = make_readable_exclusive_migration_entry(pfn);
2158 else
2159 entry = make_readable_migration_entry(pfn);
2160 swp_pte = swp_entry_to_pte(entry);
2161
2162 /*
2163 * pteval maps a zone device page and is therefore
2164 * a swap pte.
2165 */
2166 if (pte_swp_soft_dirty(pteval))
2167 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2168 if (pte_swp_uffd_wp(pteval))
2169 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2170 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
2171 trace_set_migration_pte(pvmw.address, pte_val(swp_pte),
2172 compound_order(&folio->page));
2173 /*
2174 * No need to invalidate here it will synchronize on
2175 * against the special swap migration pte.
2176 */
2177 } else if (PageHWPoison(subpage)) {
2178 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
2179 if (folio_test_hugetlb(folio)) {
2180 hugetlb_count_sub(folio_nr_pages(folio), mm);
2181 set_huge_pte_at(mm, address, pvmw.pte, pteval,
2182 hsz);
2183 } else {
2184 dec_mm_counter(mm, mm_counter(&folio->page));
2185 set_pte_at(mm, address, pvmw.pte, pteval);
2186 }
2187
2188 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
2189 /*
2190 * The guest indicated that the page content is of no
2191 * interest anymore. Simply discard the pte, vmscan
2192 * will take care of the rest.
2193 * A future reference will then fault in a new zero
2194 * page. When userfaultfd is active, we must not drop
2195 * this page though, as its main user (postcopy
2196 * migration) will not expect userfaults on already
2197 * copied pages.
2198 */
2199 dec_mm_counter(mm, mm_counter(&folio->page));
2200 } else {
2201 swp_entry_t entry;
2202 pte_t swp_pte;
2203
2204 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
2205 if (folio_test_hugetlb(folio))
2206 set_huge_pte_at(mm, address, pvmw.pte,
2207 pteval, hsz);
2208 else
2209 set_pte_at(mm, address, pvmw.pte, pteval);
2210 ret = false;
2211 page_vma_mapped_walk_done(&pvmw);
2212 break;
2213 }
2214 VM_BUG_ON_PAGE(pte_write(pteval) && folio_test_anon(folio) &&
2215 !anon_exclusive, subpage);
2216
2217 /* See folio_try_share_anon_rmap_pte(): clear PTE first. */
2218 if (folio_test_hugetlb(folio)) {
2219 if (anon_exclusive &&
2220 hugetlb_try_share_anon_rmap(folio)) {
2221 set_huge_pte_at(mm, address, pvmw.pte,
2222 pteval, hsz);
2223 ret = false;
2224 page_vma_mapped_walk_done(&pvmw);
2225 break;
2226 }
2227 } else if (anon_exclusive &&
2228 folio_try_share_anon_rmap_pte(folio, subpage)) {
2229 set_pte_at(mm, address, pvmw.pte, pteval);
2230 ret = false;
2231 page_vma_mapped_walk_done(&pvmw);
2232 break;
2233 }
2234
2235 /*
2236 * Store the pfn of the page in a special migration
2237 * pte. do_swap_page() will wait until the migration
2238 * pte is removed and then restart fault handling.
2239 */
2240 if (pte_write(pteval))
2241 entry = make_writable_migration_entry(
2242 page_to_pfn(subpage));
2243 else if (anon_exclusive)
2244 entry = make_readable_exclusive_migration_entry(
2245 page_to_pfn(subpage));
2246 else
2247 entry = make_readable_migration_entry(
2248 page_to_pfn(subpage));
2249 if (pte_young(pteval))
2250 entry = make_migration_entry_young(entry);
2251 if (pte_dirty(pteval))
2252 entry = make_migration_entry_dirty(entry);
2253 swp_pte = swp_entry_to_pte(entry);
2254 if (pte_soft_dirty(pteval))
2255 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2256 if (pte_uffd_wp(pteval))
2257 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2258 if (folio_test_hugetlb(folio))
2259 set_huge_pte_at(mm, address, pvmw.pte, swp_pte,
2260 hsz);
2261 else
2262 set_pte_at(mm, address, pvmw.pte, swp_pte);
2263 trace_set_migration_pte(address, pte_val(swp_pte),
2264 compound_order(&folio->page));
2265 /*
2266 * No need to invalidate here it will synchronize on
2267 * against the special swap migration pte.
2268 */
2269 }
2270
2271 if (unlikely(folio_test_hugetlb(folio)))
2272 hugetlb_remove_rmap(folio);
2273 else
2274 folio_remove_rmap_pte(folio, subpage, vma);
2275 if (vma->vm_flags & VM_LOCKED)
2276 mlock_drain_local();
2277 folio_put(folio);
2278 }
2279
2280 mmu_notifier_invalidate_range_end(&range);
2281
2282 return ret;
2283}
2284
2285/**
2286 * try_to_migrate - try to replace all page table mappings with swap entries
2287 * @folio: the folio to replace page table entries for
2288 * @flags: action and flags
2289 *
2290 * Tries to remove all the page table entries which are mapping this folio and
2291 * replace them with special swap entries. Caller must hold the folio lock.
2292 */
2293void try_to_migrate(struct folio *folio, enum ttu_flags flags)
2294{
2295 struct rmap_walk_control rwc = {
2296 .rmap_one = try_to_migrate_one,
2297 .arg = (void *)flags,
2298 .done = folio_not_mapped,
2299 .anon_lock = folio_lock_anon_vma_read,
2300 };
2301
2302 /*
2303 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
2304 * TTU_SPLIT_HUGE_PMD, TTU_SYNC, and TTU_BATCH_FLUSH flags.
2305 */
2306 if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2307 TTU_SYNC | TTU_BATCH_FLUSH)))
2308 return;
2309
2310 if (folio_is_zone_device(folio) &&
2311 (!folio_is_device_private(folio) && !folio_is_device_coherent(folio)))
2312 return;
2313
2314 /*
2315 * During exec, a temporary VMA is setup and later moved.
2316 * The VMA is moved under the anon_vma lock but not the
2317 * page tables leading to a race where migration cannot
2318 * find the migration ptes. Rather than increasing the
2319 * locking requirements of exec(), migration skips
2320 * temporary VMAs until after exec() completes.
2321 */
2322 if (!folio_test_ksm(folio) && folio_test_anon(folio))
2323 rwc.invalid_vma = invalid_migration_vma;
2324
2325 if (flags & TTU_RMAP_LOCKED)
2326 rmap_walk_locked(folio, &rwc);
2327 else
2328 rmap_walk(folio, &rwc);
2329}
2330
2331#ifdef CONFIG_DEVICE_PRIVATE
2332struct make_exclusive_args {
2333 struct mm_struct *mm;
2334 unsigned long address;
2335 void *owner;
2336 bool valid;
2337};
2338
2339static bool page_make_device_exclusive_one(struct folio *folio,
2340 struct vm_area_struct *vma, unsigned long address, void *priv)
2341{
2342 struct mm_struct *mm = vma->vm_mm;
2343 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
2344 struct make_exclusive_args *args = priv;
2345 pte_t pteval;
2346 struct page *subpage;
2347 bool ret = true;
2348 struct mmu_notifier_range range;
2349 swp_entry_t entry;
2350 pte_t swp_pte;
2351 pte_t ptent;
2352
2353 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0,
2354 vma->vm_mm, address, min(vma->vm_end,
2355 address + folio_size(folio)),
2356 args->owner);
2357 mmu_notifier_invalidate_range_start(&range);
2358
2359 while (page_vma_mapped_walk(&pvmw)) {
2360 /* Unexpected PMD-mapped THP? */
2361 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2362
2363 ptent = ptep_get(pvmw.pte);
2364 if (!pte_present(ptent)) {
2365 ret = false;
2366 page_vma_mapped_walk_done(&pvmw);
2367 break;
2368 }
2369
2370 subpage = folio_page(folio,
2371 pte_pfn(ptent) - folio_pfn(folio));
2372 address = pvmw.address;
2373
2374 /* Nuke the page table entry. */
2375 flush_cache_page(vma, address, pte_pfn(ptent));
2376 pteval = ptep_clear_flush(vma, address, pvmw.pte);
2377
2378 /* Set the dirty flag on the folio now the pte is gone. */
2379 if (pte_dirty(pteval))
2380 folio_mark_dirty(folio);
2381
2382 /*
2383 * Check that our target page is still mapped at the expected
2384 * address.
2385 */
2386 if (args->mm == mm && args->address == address &&
2387 pte_write(pteval))
2388 args->valid = true;
2389
2390 /*
2391 * Store the pfn of the page in a special migration
2392 * pte. do_swap_page() will wait until the migration
2393 * pte is removed and then restart fault handling.
2394 */
2395 if (pte_write(pteval))
2396 entry = make_writable_device_exclusive_entry(
2397 page_to_pfn(subpage));
2398 else
2399 entry = make_readable_device_exclusive_entry(
2400 page_to_pfn(subpage));
2401 swp_pte = swp_entry_to_pte(entry);
2402 if (pte_soft_dirty(pteval))
2403 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2404 if (pte_uffd_wp(pteval))
2405 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2406
2407 set_pte_at(mm, address, pvmw.pte, swp_pte);
2408
2409 /*
2410 * There is a reference on the page for the swap entry which has
2411 * been removed, so shouldn't take another.
2412 */
2413 folio_remove_rmap_pte(folio, subpage, vma);
2414 }
2415
2416 mmu_notifier_invalidate_range_end(&range);
2417
2418 return ret;
2419}
2420
2421/**
2422 * folio_make_device_exclusive - Mark the folio exclusively owned by a device.
2423 * @folio: The folio to replace page table entries for.
2424 * @mm: The mm_struct where the folio is expected to be mapped.
2425 * @address: Address where the folio is expected to be mapped.
2426 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2427 *
2428 * Tries to remove all the page table entries which are mapping this
2429 * folio and replace them with special device exclusive swap entries to
2430 * grant a device exclusive access to the folio.
2431 *
2432 * Context: Caller must hold the folio lock.
2433 * Return: false if the page is still mapped, or if it could not be unmapped
2434 * from the expected address. Otherwise returns true (success).
2435 */
2436static bool folio_make_device_exclusive(struct folio *folio,
2437 struct mm_struct *mm, unsigned long address, void *owner)
2438{
2439 struct make_exclusive_args args = {
2440 .mm = mm,
2441 .address = address,
2442 .owner = owner,
2443 .valid = false,
2444 };
2445 struct rmap_walk_control rwc = {
2446 .rmap_one = page_make_device_exclusive_one,
2447 .done = folio_not_mapped,
2448 .anon_lock = folio_lock_anon_vma_read,
2449 .arg = &args,
2450 };
2451
2452 /*
2453 * Restrict to anonymous folios for now to avoid potential writeback
2454 * issues.
2455 */
2456 if (!folio_test_anon(folio))
2457 return false;
2458
2459 rmap_walk(folio, &rwc);
2460
2461 return args.valid && !folio_mapcount(folio);
2462}
2463
2464/**
2465 * make_device_exclusive_range() - Mark a range for exclusive use by a device
2466 * @mm: mm_struct of associated target process
2467 * @start: start of the region to mark for exclusive device access
2468 * @end: end address of region
2469 * @pages: returns the pages which were successfully marked for exclusive access
2470 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2471 *
2472 * Returns: number of pages found in the range by GUP. A page is marked for
2473 * exclusive access only if the page pointer is non-NULL.
2474 *
2475 * This function finds ptes mapping page(s) to the given address range, locks
2476 * them and replaces mappings with special swap entries preventing userspace CPU
2477 * access. On fault these entries are replaced with the original mapping after
2478 * calling MMU notifiers.
2479 *
2480 * A driver using this to program access from a device must use a mmu notifier
2481 * critical section to hold a device specific lock during programming. Once
2482 * programming is complete it should drop the page lock and reference after
2483 * which point CPU access to the page will revoke the exclusive access.
2484 */
2485int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
2486 unsigned long end, struct page **pages,
2487 void *owner)
2488{
2489 long npages = (end - start) >> PAGE_SHIFT;
2490 long i;
2491
2492 npages = get_user_pages_remote(mm, start, npages,
2493 FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
2494 pages, NULL);
2495 if (npages < 0)
2496 return npages;
2497
2498 for (i = 0; i < npages; i++, start += PAGE_SIZE) {
2499 struct folio *folio = page_folio(pages[i]);
2500 if (PageTail(pages[i]) || !folio_trylock(folio)) {
2501 folio_put(folio);
2502 pages[i] = NULL;
2503 continue;
2504 }
2505
2506 if (!folio_make_device_exclusive(folio, mm, start, owner)) {
2507 folio_unlock(folio);
2508 folio_put(folio);
2509 pages[i] = NULL;
2510 }
2511 }
2512
2513 return npages;
2514}
2515EXPORT_SYMBOL_GPL(make_device_exclusive_range);
2516#endif
2517
2518void __put_anon_vma(struct anon_vma *anon_vma)
2519{
2520 struct anon_vma *root = anon_vma->root;
2521
2522 anon_vma_free(anon_vma);
2523 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
2524 anon_vma_free(root);
2525}
2526
2527static struct anon_vma *rmap_walk_anon_lock(struct folio *folio,
2528 struct rmap_walk_control *rwc)
2529{
2530 struct anon_vma *anon_vma;
2531
2532 if (rwc->anon_lock)
2533 return rwc->anon_lock(folio, rwc);
2534
2535 /*
2536 * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read()
2537 * because that depends on page_mapped(); but not all its usages
2538 * are holding mmap_lock. Users without mmap_lock are required to
2539 * take a reference count to prevent the anon_vma disappearing
2540 */
2541 anon_vma = folio_anon_vma(folio);
2542 if (!anon_vma)
2543 return NULL;
2544
2545 if (anon_vma_trylock_read(anon_vma))
2546 goto out;
2547
2548 if (rwc->try_lock) {
2549 anon_vma = NULL;
2550 rwc->contended = true;
2551 goto out;
2552 }
2553
2554 anon_vma_lock_read(anon_vma);
2555out:
2556 return anon_vma;
2557}
2558
2559/*
2560 * rmap_walk_anon - do something to anonymous page using the object-based
2561 * rmap method
2562 * @folio: the folio to be handled
2563 * @rwc: control variable according to each walk type
2564 * @locked: caller holds relevant rmap lock
2565 *
2566 * Find all the mappings of a folio using the mapping pointer and the vma
2567 * chains contained in the anon_vma struct it points to.
2568 */
2569static void rmap_walk_anon(struct folio *folio,
2570 struct rmap_walk_control *rwc, bool locked)
2571{
2572 struct anon_vma *anon_vma;
2573 pgoff_t pgoff_start, pgoff_end;
2574 struct anon_vma_chain *avc;
2575
2576 if (locked) {
2577 anon_vma = folio_anon_vma(folio);
2578 /* anon_vma disappear under us? */
2579 VM_BUG_ON_FOLIO(!anon_vma, folio);
2580 } else {
2581 anon_vma = rmap_walk_anon_lock(folio, rwc);
2582 }
2583 if (!anon_vma)
2584 return;
2585
2586 pgoff_start = folio_pgoff(folio);
2587 pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2588 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
2589 pgoff_start, pgoff_end) {
2590 struct vm_area_struct *vma = avc->vma;
2591 unsigned long address = vma_address(&folio->page, vma);
2592
2593 VM_BUG_ON_VMA(address == -EFAULT, vma);
2594 cond_resched();
2595
2596 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2597 continue;
2598
2599 if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2600 break;
2601 if (rwc->done && rwc->done(folio))
2602 break;
2603 }
2604
2605 if (!locked)
2606 anon_vma_unlock_read(anon_vma);
2607}
2608
2609/*
2610 * rmap_walk_file - do something to file page using the object-based rmap method
2611 * @folio: the folio to be handled
2612 * @rwc: control variable according to each walk type
2613 * @locked: caller holds relevant rmap lock
2614 *
2615 * Find all the mappings of a folio using the mapping pointer and the vma chains
2616 * contained in the address_space struct it points to.
2617 */
2618static void rmap_walk_file(struct folio *folio,
2619 struct rmap_walk_control *rwc, bool locked)
2620{
2621 struct address_space *mapping = folio_mapping(folio);
2622 pgoff_t pgoff_start, pgoff_end;
2623 struct vm_area_struct *vma;
2624
2625 /*
2626 * The page lock not only makes sure that page->mapping cannot
2627 * suddenly be NULLified by truncation, it makes sure that the
2628 * structure at mapping cannot be freed and reused yet,
2629 * so we can safely take mapping->i_mmap_rwsem.
2630 */
2631 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2632
2633 if (!mapping)
2634 return;
2635
2636 pgoff_start = folio_pgoff(folio);
2637 pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2638 if (!locked) {
2639 if (i_mmap_trylock_read(mapping))
2640 goto lookup;
2641
2642 if (rwc->try_lock) {
2643 rwc->contended = true;
2644 return;
2645 }
2646
2647 i_mmap_lock_read(mapping);
2648 }
2649lookup:
2650 vma_interval_tree_foreach(vma, &mapping->i_mmap,
2651 pgoff_start, pgoff_end) {
2652 unsigned long address = vma_address(&folio->page, vma);
2653
2654 VM_BUG_ON_VMA(address == -EFAULT, vma);
2655 cond_resched();
2656
2657 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2658 continue;
2659
2660 if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2661 goto done;
2662 if (rwc->done && rwc->done(folio))
2663 goto done;
2664 }
2665
2666done:
2667 if (!locked)
2668 i_mmap_unlock_read(mapping);
2669}
2670
2671void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc)
2672{
2673 if (unlikely(folio_test_ksm(folio)))
2674 rmap_walk_ksm(folio, rwc);
2675 else if (folio_test_anon(folio))
2676 rmap_walk_anon(folio, rwc, false);
2677 else
2678 rmap_walk_file(folio, rwc, false);
2679}
2680
2681/* Like rmap_walk, but caller holds relevant rmap lock */
2682void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc)
2683{
2684 /* no ksm support for now */
2685 VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio);
2686 if (folio_test_anon(folio))
2687 rmap_walk_anon(folio, rwc, true);
2688 else
2689 rmap_walk_file(folio, rwc, true);
2690}
2691
2692#ifdef CONFIG_HUGETLB_PAGE
2693/*
2694 * The following two functions are for anonymous (private mapped) hugepages.
2695 * Unlike common anonymous pages, anonymous hugepages have no accounting code
2696 * and no lru code, because we handle hugepages differently from common pages.
2697 */
2698void hugetlb_add_anon_rmap(struct folio *folio, struct vm_area_struct *vma,
2699 unsigned long address, rmap_t flags)
2700{
2701 VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio);
2702 VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
2703
2704 atomic_inc(&folio->_entire_mapcount);
2705 if (flags & RMAP_EXCLUSIVE)
2706 SetPageAnonExclusive(&folio->page);
2707 VM_WARN_ON_FOLIO(folio_entire_mapcount(folio) > 1 &&
2708 PageAnonExclusive(&folio->page), folio);
2709}
2710
2711void hugetlb_add_new_anon_rmap(struct folio *folio,
2712 struct vm_area_struct *vma, unsigned long address)
2713{
2714 VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio);
2715
2716 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2717 /* increment count (starts at -1) */
2718 atomic_set(&folio->_entire_mapcount, 0);
2719 folio_clear_hugetlb_restore_reserve(folio);
2720 __folio_set_anon(folio, vma, address, true);
2721 SetPageAnonExclusive(&folio->page);
2722}
2723#endif /* CONFIG_HUGETLB_PAGE */
1/*
2 * mm/rmap.c - physical to virtual reverse mappings
3 *
4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5 * Released under the General Public License (GPL).
6 *
7 * Simple, low overhead reverse mapping scheme.
8 * Please try to keep this thing as modular as possible.
9 *
10 * Provides methods for unmapping each kind of mapped page:
11 * the anon methods track anonymous pages, and
12 * the file methods track pages belonging to an inode.
13 *
14 * Original design by Rik van Riel <riel@conectiva.com.br> 2001
15 * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
16 * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
17 * Contributions by Hugh Dickins 2003, 2004
18 */
19
20/*
21 * Lock ordering in mm:
22 *
23 * inode->i_mutex (while writing or truncating, not reading or faulting)
24 * mm->mmap_lock
25 * page->flags PG_locked (lock_page) * (see huegtlbfs below)
26 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
27 * mapping->i_mmap_rwsem
28 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
29 * anon_vma->rwsem
30 * mm->page_table_lock or pte_lock
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
34 * lock_page_memcg move_lock (in __set_page_dirty_buffers)
35 * i_pages lock (widely used)
36 * lruvec->lru_lock (in lock_page_lruvec_irq)
37 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
38 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
39 * sb_lock (within inode_lock in fs/fs-writeback.c)
40 * i_pages lock (widely used, in set_page_dirty,
41 * in arch-dependent flush_dcache_mmap_lock,
42 * within bdi.wb->list_lock in __sync_single_inode)
43 *
44 * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon)
45 * ->tasklist_lock
46 * pte map lock
47 *
48 * * hugetlbfs PageHuge() pages take locks in this order:
49 * mapping->i_mmap_rwsem
50 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
51 * page->flags PG_locked (lock_page)
52 */
53
54#include <linux/mm.h>
55#include <linux/sched/mm.h>
56#include <linux/sched/task.h>
57#include <linux/pagemap.h>
58#include <linux/swap.h>
59#include <linux/swapops.h>
60#include <linux/slab.h>
61#include <linux/init.h>
62#include <linux/ksm.h>
63#include <linux/rmap.h>
64#include <linux/rcupdate.h>
65#include <linux/export.h>
66#include <linux/memcontrol.h>
67#include <linux/mmu_notifier.h>
68#include <linux/migrate.h>
69#include <linux/hugetlb.h>
70#include <linux/huge_mm.h>
71#include <linux/backing-dev.h>
72#include <linux/page_idle.h>
73#include <linux/memremap.h>
74#include <linux/userfaultfd_k.h>
75
76#include <asm/tlbflush.h>
77
78#include <trace/events/tlb.h>
79
80#include "internal.h"
81
82static struct kmem_cache *anon_vma_cachep;
83static struct kmem_cache *anon_vma_chain_cachep;
84
85static inline struct anon_vma *anon_vma_alloc(void)
86{
87 struct anon_vma *anon_vma;
88
89 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
90 if (anon_vma) {
91 atomic_set(&anon_vma->refcount, 1);
92 anon_vma->degree = 1; /* Reference for first vma */
93 anon_vma->parent = anon_vma;
94 /*
95 * Initialise the anon_vma root to point to itself. If called
96 * from fork, the root will be reset to the parents anon_vma.
97 */
98 anon_vma->root = anon_vma;
99 }
100
101 return anon_vma;
102}
103
104static inline void anon_vma_free(struct anon_vma *anon_vma)
105{
106 VM_BUG_ON(atomic_read(&anon_vma->refcount));
107
108 /*
109 * Synchronize against page_lock_anon_vma_read() such that
110 * we can safely hold the lock without the anon_vma getting
111 * freed.
112 *
113 * Relies on the full mb implied by the atomic_dec_and_test() from
114 * put_anon_vma() against the acquire barrier implied by
115 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
116 *
117 * page_lock_anon_vma_read() VS put_anon_vma()
118 * down_read_trylock() atomic_dec_and_test()
119 * LOCK MB
120 * atomic_read() rwsem_is_locked()
121 *
122 * LOCK should suffice since the actual taking of the lock must
123 * happen _before_ what follows.
124 */
125 might_sleep();
126 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
127 anon_vma_lock_write(anon_vma);
128 anon_vma_unlock_write(anon_vma);
129 }
130
131 kmem_cache_free(anon_vma_cachep, anon_vma);
132}
133
134static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
135{
136 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
137}
138
139static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
140{
141 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
142}
143
144static void anon_vma_chain_link(struct vm_area_struct *vma,
145 struct anon_vma_chain *avc,
146 struct anon_vma *anon_vma)
147{
148 avc->vma = vma;
149 avc->anon_vma = anon_vma;
150 list_add(&avc->same_vma, &vma->anon_vma_chain);
151 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
152}
153
154/**
155 * __anon_vma_prepare - attach an anon_vma to a memory region
156 * @vma: the memory region in question
157 *
158 * This makes sure the memory mapping described by 'vma' has
159 * an 'anon_vma' attached to it, so that we can associate the
160 * anonymous pages mapped into it with that anon_vma.
161 *
162 * The common case will be that we already have one, which
163 * is handled inline by anon_vma_prepare(). But if
164 * not we either need to find an adjacent mapping that we
165 * can re-use the anon_vma from (very common when the only
166 * reason for splitting a vma has been mprotect()), or we
167 * allocate a new one.
168 *
169 * Anon-vma allocations are very subtle, because we may have
170 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
171 * and that may actually touch the rwsem even in the newly
172 * allocated vma (it depends on RCU to make sure that the
173 * anon_vma isn't actually destroyed).
174 *
175 * As a result, we need to do proper anon_vma locking even
176 * for the new allocation. At the same time, we do not want
177 * to do any locking for the common case of already having
178 * an anon_vma.
179 *
180 * This must be called with the mmap_lock held for reading.
181 */
182int __anon_vma_prepare(struct vm_area_struct *vma)
183{
184 struct mm_struct *mm = vma->vm_mm;
185 struct anon_vma *anon_vma, *allocated;
186 struct anon_vma_chain *avc;
187
188 might_sleep();
189
190 avc = anon_vma_chain_alloc(GFP_KERNEL);
191 if (!avc)
192 goto out_enomem;
193
194 anon_vma = find_mergeable_anon_vma(vma);
195 allocated = NULL;
196 if (!anon_vma) {
197 anon_vma = anon_vma_alloc();
198 if (unlikely(!anon_vma))
199 goto out_enomem_free_avc;
200 allocated = anon_vma;
201 }
202
203 anon_vma_lock_write(anon_vma);
204 /* page_table_lock to protect against threads */
205 spin_lock(&mm->page_table_lock);
206 if (likely(!vma->anon_vma)) {
207 vma->anon_vma = anon_vma;
208 anon_vma_chain_link(vma, avc, anon_vma);
209 /* vma reference or self-parent link for new root */
210 anon_vma->degree++;
211 allocated = NULL;
212 avc = NULL;
213 }
214 spin_unlock(&mm->page_table_lock);
215 anon_vma_unlock_write(anon_vma);
216
217 if (unlikely(allocated))
218 put_anon_vma(allocated);
219 if (unlikely(avc))
220 anon_vma_chain_free(avc);
221
222 return 0;
223
224 out_enomem_free_avc:
225 anon_vma_chain_free(avc);
226 out_enomem:
227 return -ENOMEM;
228}
229
230/*
231 * This is a useful helper function for locking the anon_vma root as
232 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
233 * have the same vma.
234 *
235 * Such anon_vma's should have the same root, so you'd expect to see
236 * just a single mutex_lock for the whole traversal.
237 */
238static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
239{
240 struct anon_vma *new_root = anon_vma->root;
241 if (new_root != root) {
242 if (WARN_ON_ONCE(root))
243 up_write(&root->rwsem);
244 root = new_root;
245 down_write(&root->rwsem);
246 }
247 return root;
248}
249
250static inline void unlock_anon_vma_root(struct anon_vma *root)
251{
252 if (root)
253 up_write(&root->rwsem);
254}
255
256/*
257 * Attach the anon_vmas from src to dst.
258 * Returns 0 on success, -ENOMEM on failure.
259 *
260 * anon_vma_clone() is called by __vma_adjust(), __split_vma(), copy_vma() and
261 * anon_vma_fork(). The first three want an exact copy of src, while the last
262 * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
263 * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
264 * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
265 *
266 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
267 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
268 * This prevents degradation of anon_vma hierarchy to endless linear chain in
269 * case of constantly forking task. On the other hand, an anon_vma with more
270 * than one child isn't reused even if there was no alive vma, thus rmap
271 * walker has a good chance of avoiding scanning the whole hierarchy when it
272 * searches where page is mapped.
273 */
274int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
275{
276 struct anon_vma_chain *avc, *pavc;
277 struct anon_vma *root = NULL;
278
279 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
280 struct anon_vma *anon_vma;
281
282 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
283 if (unlikely(!avc)) {
284 unlock_anon_vma_root(root);
285 root = NULL;
286 avc = anon_vma_chain_alloc(GFP_KERNEL);
287 if (!avc)
288 goto enomem_failure;
289 }
290 anon_vma = pavc->anon_vma;
291 root = lock_anon_vma_root(root, anon_vma);
292 anon_vma_chain_link(dst, avc, anon_vma);
293
294 /*
295 * Reuse existing anon_vma if its degree lower than two,
296 * that means it has no vma and only one anon_vma child.
297 *
298 * Do not chose parent anon_vma, otherwise first child
299 * will always reuse it. Root anon_vma is never reused:
300 * it has self-parent reference and at least one child.
301 */
302 if (!dst->anon_vma && src->anon_vma &&
303 anon_vma != src->anon_vma && anon_vma->degree < 2)
304 dst->anon_vma = anon_vma;
305 }
306 if (dst->anon_vma)
307 dst->anon_vma->degree++;
308 unlock_anon_vma_root(root);
309 return 0;
310
311 enomem_failure:
312 /*
313 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
314 * decremented in unlink_anon_vmas().
315 * We can safely do this because callers of anon_vma_clone() don't care
316 * about dst->anon_vma if anon_vma_clone() failed.
317 */
318 dst->anon_vma = NULL;
319 unlink_anon_vmas(dst);
320 return -ENOMEM;
321}
322
323/*
324 * Attach vma to its own anon_vma, as well as to the anon_vmas that
325 * the corresponding VMA in the parent process is attached to.
326 * Returns 0 on success, non-zero on failure.
327 */
328int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
329{
330 struct anon_vma_chain *avc;
331 struct anon_vma *anon_vma;
332 int error;
333
334 /* Don't bother if the parent process has no anon_vma here. */
335 if (!pvma->anon_vma)
336 return 0;
337
338 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
339 vma->anon_vma = NULL;
340
341 /*
342 * First, attach the new VMA to the parent VMA's anon_vmas,
343 * so rmap can find non-COWed pages in child processes.
344 */
345 error = anon_vma_clone(vma, pvma);
346 if (error)
347 return error;
348
349 /* An existing anon_vma has been reused, all done then. */
350 if (vma->anon_vma)
351 return 0;
352
353 /* Then add our own anon_vma. */
354 anon_vma = anon_vma_alloc();
355 if (!anon_vma)
356 goto out_error;
357 avc = anon_vma_chain_alloc(GFP_KERNEL);
358 if (!avc)
359 goto out_error_free_anon_vma;
360
361 /*
362 * The root anon_vma's rwsem is the lock actually used when we
363 * lock any of the anon_vmas in this anon_vma tree.
364 */
365 anon_vma->root = pvma->anon_vma->root;
366 anon_vma->parent = pvma->anon_vma;
367 /*
368 * With refcounts, an anon_vma can stay around longer than the
369 * process it belongs to. The root anon_vma needs to be pinned until
370 * this anon_vma is freed, because the lock lives in the root.
371 */
372 get_anon_vma(anon_vma->root);
373 /* Mark this anon_vma as the one where our new (COWed) pages go. */
374 vma->anon_vma = anon_vma;
375 anon_vma_lock_write(anon_vma);
376 anon_vma_chain_link(vma, avc, anon_vma);
377 anon_vma->parent->degree++;
378 anon_vma_unlock_write(anon_vma);
379
380 return 0;
381
382 out_error_free_anon_vma:
383 put_anon_vma(anon_vma);
384 out_error:
385 unlink_anon_vmas(vma);
386 return -ENOMEM;
387}
388
389void unlink_anon_vmas(struct vm_area_struct *vma)
390{
391 struct anon_vma_chain *avc, *next;
392 struct anon_vma *root = NULL;
393
394 /*
395 * Unlink each anon_vma chained to the VMA. This list is ordered
396 * from newest to oldest, ensuring the root anon_vma gets freed last.
397 */
398 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
399 struct anon_vma *anon_vma = avc->anon_vma;
400
401 root = lock_anon_vma_root(root, anon_vma);
402 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
403
404 /*
405 * Leave empty anon_vmas on the list - we'll need
406 * to free them outside the lock.
407 */
408 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
409 anon_vma->parent->degree--;
410 continue;
411 }
412
413 list_del(&avc->same_vma);
414 anon_vma_chain_free(avc);
415 }
416 if (vma->anon_vma) {
417 vma->anon_vma->degree--;
418
419 /*
420 * vma would still be needed after unlink, and anon_vma will be prepared
421 * when handle fault.
422 */
423 vma->anon_vma = NULL;
424 }
425 unlock_anon_vma_root(root);
426
427 /*
428 * Iterate the list once more, it now only contains empty and unlinked
429 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
430 * needing to write-acquire the anon_vma->root->rwsem.
431 */
432 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
433 struct anon_vma *anon_vma = avc->anon_vma;
434
435 VM_WARN_ON(anon_vma->degree);
436 put_anon_vma(anon_vma);
437
438 list_del(&avc->same_vma);
439 anon_vma_chain_free(avc);
440 }
441}
442
443static void anon_vma_ctor(void *data)
444{
445 struct anon_vma *anon_vma = data;
446
447 init_rwsem(&anon_vma->rwsem);
448 atomic_set(&anon_vma->refcount, 0);
449 anon_vma->rb_root = RB_ROOT_CACHED;
450}
451
452void __init anon_vma_init(void)
453{
454 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
455 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
456 anon_vma_ctor);
457 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
458 SLAB_PANIC|SLAB_ACCOUNT);
459}
460
461/*
462 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
463 *
464 * Since there is no serialization what so ever against page_remove_rmap()
465 * the best this function can do is return a refcount increased anon_vma
466 * that might have been relevant to this page.
467 *
468 * The page might have been remapped to a different anon_vma or the anon_vma
469 * returned may already be freed (and even reused).
470 *
471 * In case it was remapped to a different anon_vma, the new anon_vma will be a
472 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
473 * ensure that any anon_vma obtained from the page will still be valid for as
474 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
475 *
476 * All users of this function must be very careful when walking the anon_vma
477 * chain and verify that the page in question is indeed mapped in it
478 * [ something equivalent to page_mapped_in_vma() ].
479 *
480 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
481 * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
482 * if there is a mapcount, we can dereference the anon_vma after observing
483 * those.
484 */
485struct anon_vma *page_get_anon_vma(struct page *page)
486{
487 struct anon_vma *anon_vma = NULL;
488 unsigned long anon_mapping;
489
490 rcu_read_lock();
491 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
492 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
493 goto out;
494 if (!page_mapped(page))
495 goto out;
496
497 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
498 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
499 anon_vma = NULL;
500 goto out;
501 }
502
503 /*
504 * If this page is still mapped, then its anon_vma cannot have been
505 * freed. But if it has been unmapped, we have no security against the
506 * anon_vma structure being freed and reused (for another anon_vma:
507 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
508 * above cannot corrupt).
509 */
510 if (!page_mapped(page)) {
511 rcu_read_unlock();
512 put_anon_vma(anon_vma);
513 return NULL;
514 }
515out:
516 rcu_read_unlock();
517
518 return anon_vma;
519}
520
521/*
522 * Similar to page_get_anon_vma() except it locks the anon_vma.
523 *
524 * Its a little more complex as it tries to keep the fast path to a single
525 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
526 * reference like with page_get_anon_vma() and then block on the mutex.
527 */
528struct anon_vma *page_lock_anon_vma_read(struct page *page)
529{
530 struct anon_vma *anon_vma = NULL;
531 struct anon_vma *root_anon_vma;
532 unsigned long anon_mapping;
533
534 rcu_read_lock();
535 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
536 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
537 goto out;
538 if (!page_mapped(page))
539 goto out;
540
541 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
542 root_anon_vma = READ_ONCE(anon_vma->root);
543 if (down_read_trylock(&root_anon_vma->rwsem)) {
544 /*
545 * If the page is still mapped, then this anon_vma is still
546 * its anon_vma, and holding the mutex ensures that it will
547 * not go away, see anon_vma_free().
548 */
549 if (!page_mapped(page)) {
550 up_read(&root_anon_vma->rwsem);
551 anon_vma = NULL;
552 }
553 goto out;
554 }
555
556 /* trylock failed, we got to sleep */
557 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
558 anon_vma = NULL;
559 goto out;
560 }
561
562 if (!page_mapped(page)) {
563 rcu_read_unlock();
564 put_anon_vma(anon_vma);
565 return NULL;
566 }
567
568 /* we pinned the anon_vma, its safe to sleep */
569 rcu_read_unlock();
570 anon_vma_lock_read(anon_vma);
571
572 if (atomic_dec_and_test(&anon_vma->refcount)) {
573 /*
574 * Oops, we held the last refcount, release the lock
575 * and bail -- can't simply use put_anon_vma() because
576 * we'll deadlock on the anon_vma_lock_write() recursion.
577 */
578 anon_vma_unlock_read(anon_vma);
579 __put_anon_vma(anon_vma);
580 anon_vma = NULL;
581 }
582
583 return anon_vma;
584
585out:
586 rcu_read_unlock();
587 return anon_vma;
588}
589
590void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
591{
592 anon_vma_unlock_read(anon_vma);
593}
594
595#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
596/*
597 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
598 * important if a PTE was dirty when it was unmapped that it's flushed
599 * before any IO is initiated on the page to prevent lost writes. Similarly,
600 * it must be flushed before freeing to prevent data leakage.
601 */
602void try_to_unmap_flush(void)
603{
604 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
605
606 if (!tlb_ubc->flush_required)
607 return;
608
609 arch_tlbbatch_flush(&tlb_ubc->arch);
610 tlb_ubc->flush_required = false;
611 tlb_ubc->writable = false;
612}
613
614/* Flush iff there are potentially writable TLB entries that can race with IO */
615void try_to_unmap_flush_dirty(void)
616{
617 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
618
619 if (tlb_ubc->writable)
620 try_to_unmap_flush();
621}
622
623static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
624{
625 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
626
627 arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
628 tlb_ubc->flush_required = true;
629
630 /*
631 * Ensure compiler does not re-order the setting of tlb_flush_batched
632 * before the PTE is cleared.
633 */
634 barrier();
635 mm->tlb_flush_batched = true;
636
637 /*
638 * If the PTE was dirty then it's best to assume it's writable. The
639 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
640 * before the page is queued for IO.
641 */
642 if (writable)
643 tlb_ubc->writable = true;
644}
645
646/*
647 * Returns true if the TLB flush should be deferred to the end of a batch of
648 * unmap operations to reduce IPIs.
649 */
650static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
651{
652 bool should_defer = false;
653
654 if (!(flags & TTU_BATCH_FLUSH))
655 return false;
656
657 /* If remote CPUs need to be flushed then defer batch the flush */
658 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
659 should_defer = true;
660 put_cpu();
661
662 return should_defer;
663}
664
665/*
666 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
667 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
668 * operation such as mprotect or munmap to race between reclaim unmapping
669 * the page and flushing the page. If this race occurs, it potentially allows
670 * access to data via a stale TLB entry. Tracking all mm's that have TLB
671 * batching in flight would be expensive during reclaim so instead track
672 * whether TLB batching occurred in the past and if so then do a flush here
673 * if required. This will cost one additional flush per reclaim cycle paid
674 * by the first operation at risk such as mprotect and mumap.
675 *
676 * This must be called under the PTL so that an access to tlb_flush_batched
677 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
678 * via the PTL.
679 */
680void flush_tlb_batched_pending(struct mm_struct *mm)
681{
682 if (data_race(mm->tlb_flush_batched)) {
683 flush_tlb_mm(mm);
684
685 /*
686 * Do not allow the compiler to re-order the clearing of
687 * tlb_flush_batched before the tlb is flushed.
688 */
689 barrier();
690 mm->tlb_flush_batched = false;
691 }
692}
693#else
694static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
695{
696}
697
698static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
699{
700 return false;
701}
702#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
703
704/*
705 * At what user virtual address is page expected in vma?
706 * Caller should check the page is actually part of the vma.
707 */
708unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
709{
710 if (PageAnon(page)) {
711 struct anon_vma *page__anon_vma = page_anon_vma(page);
712 /*
713 * Note: swapoff's unuse_vma() is more efficient with this
714 * check, and needs it to match anon_vma when KSM is active.
715 */
716 if (!vma->anon_vma || !page__anon_vma ||
717 vma->anon_vma->root != page__anon_vma->root)
718 return -EFAULT;
719 } else if (!vma->vm_file) {
720 return -EFAULT;
721 } else if (vma->vm_file->f_mapping != compound_head(page)->mapping) {
722 return -EFAULT;
723 }
724
725 return vma_address(page, vma);
726}
727
728pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
729{
730 pgd_t *pgd;
731 p4d_t *p4d;
732 pud_t *pud;
733 pmd_t *pmd = NULL;
734 pmd_t pmde;
735
736 pgd = pgd_offset(mm, address);
737 if (!pgd_present(*pgd))
738 goto out;
739
740 p4d = p4d_offset(pgd, address);
741 if (!p4d_present(*p4d))
742 goto out;
743
744 pud = pud_offset(p4d, address);
745 if (!pud_present(*pud))
746 goto out;
747
748 pmd = pmd_offset(pud, address);
749 /*
750 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
751 * without holding anon_vma lock for write. So when looking for a
752 * genuine pmde (in which to find pte), test present and !THP together.
753 */
754 pmde = *pmd;
755 barrier();
756 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
757 pmd = NULL;
758out:
759 return pmd;
760}
761
762struct page_referenced_arg {
763 int mapcount;
764 int referenced;
765 unsigned long vm_flags;
766 struct mem_cgroup *memcg;
767};
768/*
769 * arg: page_referenced_arg will be passed
770 */
771static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
772 unsigned long address, void *arg)
773{
774 struct page_referenced_arg *pra = arg;
775 struct page_vma_mapped_walk pvmw = {
776 .page = page,
777 .vma = vma,
778 .address = address,
779 };
780 int referenced = 0;
781
782 while (page_vma_mapped_walk(&pvmw)) {
783 address = pvmw.address;
784
785 if (vma->vm_flags & VM_LOCKED) {
786 page_vma_mapped_walk_done(&pvmw);
787 pra->vm_flags |= VM_LOCKED;
788 return false; /* To break the loop */
789 }
790
791 if (pvmw.pte) {
792 if (ptep_clear_flush_young_notify(vma, address,
793 pvmw.pte)) {
794 /*
795 * Don't treat a reference through
796 * a sequentially read mapping as such.
797 * If the page has been used in another mapping,
798 * we will catch it; if this other mapping is
799 * already gone, the unmap path will have set
800 * PG_referenced or activated the page.
801 */
802 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
803 referenced++;
804 }
805 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
806 if (pmdp_clear_flush_young_notify(vma, address,
807 pvmw.pmd))
808 referenced++;
809 } else {
810 /* unexpected pmd-mapped page? */
811 WARN_ON_ONCE(1);
812 }
813
814 pra->mapcount--;
815 }
816
817 if (referenced)
818 clear_page_idle(page);
819 if (test_and_clear_page_young(page))
820 referenced++;
821
822 if (referenced) {
823 pra->referenced++;
824 pra->vm_flags |= vma->vm_flags;
825 }
826
827 if (!pra->mapcount)
828 return false; /* To break the loop */
829
830 return true;
831}
832
833static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
834{
835 struct page_referenced_arg *pra = arg;
836 struct mem_cgroup *memcg = pra->memcg;
837
838 if (!mm_match_cgroup(vma->vm_mm, memcg))
839 return true;
840
841 return false;
842}
843
844/**
845 * page_referenced - test if the page was referenced
846 * @page: the page to test
847 * @is_locked: caller holds lock on the page
848 * @memcg: target memory cgroup
849 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
850 *
851 * Quick test_and_clear_referenced for all mappings to a page,
852 * returns the number of ptes which referenced the page.
853 */
854int page_referenced(struct page *page,
855 int is_locked,
856 struct mem_cgroup *memcg,
857 unsigned long *vm_flags)
858{
859 int we_locked = 0;
860 struct page_referenced_arg pra = {
861 .mapcount = total_mapcount(page),
862 .memcg = memcg,
863 };
864 struct rmap_walk_control rwc = {
865 .rmap_one = page_referenced_one,
866 .arg = (void *)&pra,
867 .anon_lock = page_lock_anon_vma_read,
868 };
869
870 *vm_flags = 0;
871 if (!pra.mapcount)
872 return 0;
873
874 if (!page_rmapping(page))
875 return 0;
876
877 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
878 we_locked = trylock_page(page);
879 if (!we_locked)
880 return 1;
881 }
882
883 /*
884 * If we are reclaiming on behalf of a cgroup, skip
885 * counting on behalf of references from different
886 * cgroups
887 */
888 if (memcg) {
889 rwc.invalid_vma = invalid_page_referenced_vma;
890 }
891
892 rmap_walk(page, &rwc);
893 *vm_flags = pra.vm_flags;
894
895 if (we_locked)
896 unlock_page(page);
897
898 return pra.referenced;
899}
900
901static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
902 unsigned long address, void *arg)
903{
904 struct page_vma_mapped_walk pvmw = {
905 .page = page,
906 .vma = vma,
907 .address = address,
908 .flags = PVMW_SYNC,
909 };
910 struct mmu_notifier_range range;
911 int *cleaned = arg;
912
913 /*
914 * We have to assume the worse case ie pmd for invalidation. Note that
915 * the page can not be free from this function.
916 */
917 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
918 0, vma, vma->vm_mm, address,
919 vma_address_end(page, vma));
920 mmu_notifier_invalidate_range_start(&range);
921
922 while (page_vma_mapped_walk(&pvmw)) {
923 int ret = 0;
924
925 address = pvmw.address;
926 if (pvmw.pte) {
927 pte_t entry;
928 pte_t *pte = pvmw.pte;
929
930 if (!pte_dirty(*pte) && !pte_write(*pte))
931 continue;
932
933 flush_cache_page(vma, address, pte_pfn(*pte));
934 entry = ptep_clear_flush(vma, address, pte);
935 entry = pte_wrprotect(entry);
936 entry = pte_mkclean(entry);
937 set_pte_at(vma->vm_mm, address, pte, entry);
938 ret = 1;
939 } else {
940#ifdef CONFIG_TRANSPARENT_HUGEPAGE
941 pmd_t *pmd = pvmw.pmd;
942 pmd_t entry;
943
944 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
945 continue;
946
947 flush_cache_page(vma, address, page_to_pfn(page));
948 entry = pmdp_invalidate(vma, address, pmd);
949 entry = pmd_wrprotect(entry);
950 entry = pmd_mkclean(entry);
951 set_pmd_at(vma->vm_mm, address, pmd, entry);
952 ret = 1;
953#else
954 /* unexpected pmd-mapped page? */
955 WARN_ON_ONCE(1);
956#endif
957 }
958
959 /*
960 * No need to call mmu_notifier_invalidate_range() as we are
961 * downgrading page table protection not changing it to point
962 * to a new page.
963 *
964 * See Documentation/vm/mmu_notifier.rst
965 */
966 if (ret)
967 (*cleaned)++;
968 }
969
970 mmu_notifier_invalidate_range_end(&range);
971
972 return true;
973}
974
975static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
976{
977 if (vma->vm_flags & VM_SHARED)
978 return false;
979
980 return true;
981}
982
983int page_mkclean(struct page *page)
984{
985 int cleaned = 0;
986 struct address_space *mapping;
987 struct rmap_walk_control rwc = {
988 .arg = (void *)&cleaned,
989 .rmap_one = page_mkclean_one,
990 .invalid_vma = invalid_mkclean_vma,
991 };
992
993 BUG_ON(!PageLocked(page));
994
995 if (!page_mapped(page))
996 return 0;
997
998 mapping = page_mapping(page);
999 if (!mapping)
1000 return 0;
1001
1002 rmap_walk(page, &rwc);
1003
1004 return cleaned;
1005}
1006EXPORT_SYMBOL_GPL(page_mkclean);
1007
1008/**
1009 * page_move_anon_rmap - move a page to our anon_vma
1010 * @page: the page to move to our anon_vma
1011 * @vma: the vma the page belongs to
1012 *
1013 * When a page belongs exclusively to one process after a COW event,
1014 * that page can be moved into the anon_vma that belongs to just that
1015 * process, so the rmap code will not search the parent or sibling
1016 * processes.
1017 */
1018void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1019{
1020 struct anon_vma *anon_vma = vma->anon_vma;
1021
1022 page = compound_head(page);
1023
1024 VM_BUG_ON_PAGE(!PageLocked(page), page);
1025 VM_BUG_ON_VMA(!anon_vma, vma);
1026
1027 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1028 /*
1029 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1030 * simultaneously, so a concurrent reader (eg page_referenced()'s
1031 * PageAnon()) will not see one without the other.
1032 */
1033 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1034}
1035
1036/**
1037 * __page_set_anon_rmap - set up new anonymous rmap
1038 * @page: Page or Hugepage to add to rmap
1039 * @vma: VM area to add page to.
1040 * @address: User virtual address of the mapping
1041 * @exclusive: the page is exclusively owned by the current process
1042 */
1043static void __page_set_anon_rmap(struct page *page,
1044 struct vm_area_struct *vma, unsigned long address, int exclusive)
1045{
1046 struct anon_vma *anon_vma = vma->anon_vma;
1047
1048 BUG_ON(!anon_vma);
1049
1050 if (PageAnon(page))
1051 return;
1052
1053 /*
1054 * If the page isn't exclusively mapped into this vma,
1055 * we must use the _oldest_ possible anon_vma for the
1056 * page mapping!
1057 */
1058 if (!exclusive)
1059 anon_vma = anon_vma->root;
1060
1061 /*
1062 * page_idle does a lockless/optimistic rmap scan on page->mapping.
1063 * Make sure the compiler doesn't split the stores of anon_vma and
1064 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1065 * could mistake the mapping for a struct address_space and crash.
1066 */
1067 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1068 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1069 page->index = linear_page_index(vma, address);
1070}
1071
1072/**
1073 * __page_check_anon_rmap - sanity check anonymous rmap addition
1074 * @page: the page to add the mapping to
1075 * @vma: the vm area in which the mapping is added
1076 * @address: the user virtual address mapped
1077 */
1078static void __page_check_anon_rmap(struct page *page,
1079 struct vm_area_struct *vma, unsigned long address)
1080{
1081 /*
1082 * The page's anon-rmap details (mapping and index) are guaranteed to
1083 * be set up correctly at this point.
1084 *
1085 * We have exclusion against page_add_anon_rmap because the caller
1086 * always holds the page locked.
1087 *
1088 * We have exclusion against page_add_new_anon_rmap because those pages
1089 * are initially only visible via the pagetables, and the pte is locked
1090 * over the call to page_add_new_anon_rmap.
1091 */
1092 VM_BUG_ON_PAGE(page_anon_vma(page)->root != vma->anon_vma->root, page);
1093 VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1094 page);
1095}
1096
1097/**
1098 * page_add_anon_rmap - add pte mapping to an anonymous page
1099 * @page: the page to add the mapping to
1100 * @vma: the vm area in which the mapping is added
1101 * @address: the user virtual address mapped
1102 * @compound: charge the page as compound or small page
1103 *
1104 * The caller needs to hold the pte lock, and the page must be locked in
1105 * the anon_vma case: to serialize mapping,index checking after setting,
1106 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1107 * (but PageKsm is never downgraded to PageAnon).
1108 */
1109void page_add_anon_rmap(struct page *page,
1110 struct vm_area_struct *vma, unsigned long address, bool compound)
1111{
1112 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1113}
1114
1115/*
1116 * Special version of the above for do_swap_page, which often runs
1117 * into pages that are exclusively owned by the current process.
1118 * Everybody else should continue to use page_add_anon_rmap above.
1119 */
1120void do_page_add_anon_rmap(struct page *page,
1121 struct vm_area_struct *vma, unsigned long address, int flags)
1122{
1123 bool compound = flags & RMAP_COMPOUND;
1124 bool first;
1125
1126 if (unlikely(PageKsm(page)))
1127 lock_page_memcg(page);
1128 else
1129 VM_BUG_ON_PAGE(!PageLocked(page), page);
1130
1131 if (compound) {
1132 atomic_t *mapcount;
1133 VM_BUG_ON_PAGE(!PageLocked(page), page);
1134 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1135 mapcount = compound_mapcount_ptr(page);
1136 first = atomic_inc_and_test(mapcount);
1137 } else {
1138 first = atomic_inc_and_test(&page->_mapcount);
1139 }
1140
1141 if (first) {
1142 int nr = compound ? thp_nr_pages(page) : 1;
1143 /*
1144 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1145 * these counters are not modified in interrupt context, and
1146 * pte lock(a spinlock) is held, which implies preemption
1147 * disabled.
1148 */
1149 if (compound)
1150 __mod_lruvec_page_state(page, NR_ANON_THPS, nr);
1151 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1152 }
1153
1154 if (unlikely(PageKsm(page))) {
1155 unlock_page_memcg(page);
1156 return;
1157 }
1158
1159 /* address might be in next vma when migration races vma_adjust */
1160 if (first)
1161 __page_set_anon_rmap(page, vma, address,
1162 flags & RMAP_EXCLUSIVE);
1163 else
1164 __page_check_anon_rmap(page, vma, address);
1165}
1166
1167/**
1168 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1169 * @page: the page to add the mapping to
1170 * @vma: the vm area in which the mapping is added
1171 * @address: the user virtual address mapped
1172 * @compound: charge the page as compound or small page
1173 *
1174 * Same as page_add_anon_rmap but must only be called on *new* pages.
1175 * This means the inc-and-test can be bypassed.
1176 * Page does not have to be locked.
1177 */
1178void page_add_new_anon_rmap(struct page *page,
1179 struct vm_area_struct *vma, unsigned long address, bool compound)
1180{
1181 int nr = compound ? thp_nr_pages(page) : 1;
1182
1183 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1184 __SetPageSwapBacked(page);
1185 if (compound) {
1186 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1187 /* increment count (starts at -1) */
1188 atomic_set(compound_mapcount_ptr(page), 0);
1189 if (hpage_pincount_available(page))
1190 atomic_set(compound_pincount_ptr(page), 0);
1191
1192 __mod_lruvec_page_state(page, NR_ANON_THPS, nr);
1193 } else {
1194 /* Anon THP always mapped first with PMD */
1195 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1196 /* increment count (starts at -1) */
1197 atomic_set(&page->_mapcount, 0);
1198 }
1199 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1200 __page_set_anon_rmap(page, vma, address, 1);
1201}
1202
1203/**
1204 * page_add_file_rmap - add pte mapping to a file page
1205 * @page: the page to add the mapping to
1206 * @compound: charge the page as compound or small page
1207 *
1208 * The caller needs to hold the pte lock.
1209 */
1210void page_add_file_rmap(struct page *page, bool compound)
1211{
1212 int i, nr = 1;
1213
1214 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1215 lock_page_memcg(page);
1216 if (compound && PageTransHuge(page)) {
1217 int nr_pages = thp_nr_pages(page);
1218
1219 for (i = 0, nr = 0; i < nr_pages; i++) {
1220 if (atomic_inc_and_test(&page[i]._mapcount))
1221 nr++;
1222 }
1223 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1224 goto out;
1225 if (PageSwapBacked(page))
1226 __mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED,
1227 nr_pages);
1228 else
1229 __mod_lruvec_page_state(page, NR_FILE_PMDMAPPED,
1230 nr_pages);
1231 } else {
1232 if (PageTransCompound(page) && page_mapping(page)) {
1233 VM_WARN_ON_ONCE(!PageLocked(page));
1234
1235 SetPageDoubleMap(compound_head(page));
1236 if (PageMlocked(page))
1237 clear_page_mlock(compound_head(page));
1238 }
1239 if (!atomic_inc_and_test(&page->_mapcount))
1240 goto out;
1241 }
1242 __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1243out:
1244 unlock_page_memcg(page);
1245}
1246
1247static void page_remove_file_rmap(struct page *page, bool compound)
1248{
1249 int i, nr = 1;
1250
1251 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1252
1253 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1254 if (unlikely(PageHuge(page))) {
1255 /* hugetlb pages are always mapped with pmds */
1256 atomic_dec(compound_mapcount_ptr(page));
1257 return;
1258 }
1259
1260 /* page still mapped by someone else? */
1261 if (compound && PageTransHuge(page)) {
1262 int nr_pages = thp_nr_pages(page);
1263
1264 for (i = 0, nr = 0; i < nr_pages; i++) {
1265 if (atomic_add_negative(-1, &page[i]._mapcount))
1266 nr++;
1267 }
1268 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1269 return;
1270 if (PageSwapBacked(page))
1271 __mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED,
1272 -nr_pages);
1273 else
1274 __mod_lruvec_page_state(page, NR_FILE_PMDMAPPED,
1275 -nr_pages);
1276 } else {
1277 if (!atomic_add_negative(-1, &page->_mapcount))
1278 return;
1279 }
1280
1281 /*
1282 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1283 * these counters are not modified in interrupt context, and
1284 * pte lock(a spinlock) is held, which implies preemption disabled.
1285 */
1286 __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1287
1288 if (unlikely(PageMlocked(page)))
1289 clear_page_mlock(page);
1290}
1291
1292static void page_remove_anon_compound_rmap(struct page *page)
1293{
1294 int i, nr;
1295
1296 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1297 return;
1298
1299 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1300 if (unlikely(PageHuge(page)))
1301 return;
1302
1303 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1304 return;
1305
1306 __mod_lruvec_page_state(page, NR_ANON_THPS, -thp_nr_pages(page));
1307
1308 if (TestClearPageDoubleMap(page)) {
1309 /*
1310 * Subpages can be mapped with PTEs too. Check how many of
1311 * them are still mapped.
1312 */
1313 for (i = 0, nr = 0; i < thp_nr_pages(page); i++) {
1314 if (atomic_add_negative(-1, &page[i]._mapcount))
1315 nr++;
1316 }
1317
1318 /*
1319 * Queue the page for deferred split if at least one small
1320 * page of the compound page is unmapped, but at least one
1321 * small page is still mapped.
1322 */
1323 if (nr && nr < thp_nr_pages(page))
1324 deferred_split_huge_page(page);
1325 } else {
1326 nr = thp_nr_pages(page);
1327 }
1328
1329 if (unlikely(PageMlocked(page)))
1330 clear_page_mlock(page);
1331
1332 if (nr)
1333 __mod_lruvec_page_state(page, NR_ANON_MAPPED, -nr);
1334}
1335
1336/**
1337 * page_remove_rmap - take down pte mapping from a page
1338 * @page: page to remove mapping from
1339 * @compound: uncharge the page as compound or small page
1340 *
1341 * The caller needs to hold the pte lock.
1342 */
1343void page_remove_rmap(struct page *page, bool compound)
1344{
1345 lock_page_memcg(page);
1346
1347 if (!PageAnon(page)) {
1348 page_remove_file_rmap(page, compound);
1349 goto out;
1350 }
1351
1352 if (compound) {
1353 page_remove_anon_compound_rmap(page);
1354 goto out;
1355 }
1356
1357 /* page still mapped by someone else? */
1358 if (!atomic_add_negative(-1, &page->_mapcount))
1359 goto out;
1360
1361 /*
1362 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1363 * these counters are not modified in interrupt context, and
1364 * pte lock(a spinlock) is held, which implies preemption disabled.
1365 */
1366 __dec_lruvec_page_state(page, NR_ANON_MAPPED);
1367
1368 if (unlikely(PageMlocked(page)))
1369 clear_page_mlock(page);
1370
1371 if (PageTransCompound(page))
1372 deferred_split_huge_page(compound_head(page));
1373
1374 /*
1375 * It would be tidy to reset the PageAnon mapping here,
1376 * but that might overwrite a racing page_add_anon_rmap
1377 * which increments mapcount after us but sets mapping
1378 * before us: so leave the reset to free_unref_page,
1379 * and remember that it's only reliable while mapped.
1380 * Leaving it set also helps swapoff to reinstate ptes
1381 * faster for those pages still in swapcache.
1382 */
1383out:
1384 unlock_page_memcg(page);
1385}
1386
1387/*
1388 * @arg: enum ttu_flags will be passed to this argument
1389 */
1390static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1391 unsigned long address, void *arg)
1392{
1393 struct mm_struct *mm = vma->vm_mm;
1394 struct page_vma_mapped_walk pvmw = {
1395 .page = page,
1396 .vma = vma,
1397 .address = address,
1398 };
1399 pte_t pteval;
1400 struct page *subpage;
1401 bool ret = true;
1402 struct mmu_notifier_range range;
1403 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1404
1405 /*
1406 * When racing against e.g. zap_pte_range() on another cpu,
1407 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1408 * try_to_unmap() may return before page_mapped() has become false,
1409 * if page table locking is skipped: use TTU_SYNC to wait for that.
1410 */
1411 if (flags & TTU_SYNC)
1412 pvmw.flags = PVMW_SYNC;
1413
1414 if (flags & TTU_SPLIT_HUGE_PMD)
1415 split_huge_pmd_address(vma, address, false, page);
1416
1417 /*
1418 * For THP, we have to assume the worse case ie pmd for invalidation.
1419 * For hugetlb, it could be much worse if we need to do pud
1420 * invalidation in the case of pmd sharing.
1421 *
1422 * Note that the page can not be free in this function as call of
1423 * try_to_unmap() must hold a reference on the page.
1424 */
1425 range.end = PageKsm(page) ?
1426 address + PAGE_SIZE : vma_address_end(page, vma);
1427 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1428 address, range.end);
1429 if (PageHuge(page)) {
1430 /*
1431 * If sharing is possible, start and end will be adjusted
1432 * accordingly.
1433 */
1434 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1435 &range.end);
1436 }
1437 mmu_notifier_invalidate_range_start(&range);
1438
1439 while (page_vma_mapped_walk(&pvmw)) {
1440 /*
1441 * If the page is mlock()d, we cannot swap it out.
1442 */
1443 if (!(flags & TTU_IGNORE_MLOCK) &&
1444 (vma->vm_flags & VM_LOCKED)) {
1445 /*
1446 * PTE-mapped THP are never marked as mlocked: so do
1447 * not set it on a DoubleMap THP, nor on an Anon THP
1448 * (which may still be PTE-mapped after DoubleMap was
1449 * cleared). But stop unmapping even in those cases.
1450 */
1451 if (!PageTransCompound(page) || (PageHead(page) &&
1452 !PageDoubleMap(page) && !PageAnon(page)))
1453 mlock_vma_page(page);
1454 page_vma_mapped_walk_done(&pvmw);
1455 ret = false;
1456 break;
1457 }
1458
1459 /* Unexpected PMD-mapped THP? */
1460 VM_BUG_ON_PAGE(!pvmw.pte, page);
1461
1462 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1463 address = pvmw.address;
1464
1465 if (PageHuge(page) && !PageAnon(page)) {
1466 /*
1467 * To call huge_pmd_unshare, i_mmap_rwsem must be
1468 * held in write mode. Caller needs to explicitly
1469 * do this outside rmap routines.
1470 */
1471 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1472 if (huge_pmd_unshare(mm, vma, &address, pvmw.pte)) {
1473 /*
1474 * huge_pmd_unshare unmapped an entire PMD
1475 * page. There is no way of knowing exactly
1476 * which PMDs may be cached for this mm, so
1477 * we must flush them all. start/end were
1478 * already adjusted above to cover this range.
1479 */
1480 flush_cache_range(vma, range.start, range.end);
1481 flush_tlb_range(vma, range.start, range.end);
1482 mmu_notifier_invalidate_range(mm, range.start,
1483 range.end);
1484
1485 /*
1486 * The ref count of the PMD page was dropped
1487 * which is part of the way map counting
1488 * is done for shared PMDs. Return 'true'
1489 * here. When there is no other sharing,
1490 * huge_pmd_unshare returns false and we will
1491 * unmap the actual page and drop map count
1492 * to zero.
1493 */
1494 page_vma_mapped_walk_done(&pvmw);
1495 break;
1496 }
1497 }
1498
1499 /* Nuke the page table entry. */
1500 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1501 if (should_defer_flush(mm, flags)) {
1502 /*
1503 * We clear the PTE but do not flush so potentially
1504 * a remote CPU could still be writing to the page.
1505 * If the entry was previously clean then the
1506 * architecture must guarantee that a clear->dirty
1507 * transition on a cached TLB entry is written through
1508 * and traps if the PTE is unmapped.
1509 */
1510 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1511
1512 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1513 } else {
1514 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1515 }
1516
1517 /* Move the dirty bit to the page. Now the pte is gone. */
1518 if (pte_dirty(pteval))
1519 set_page_dirty(page);
1520
1521 /* Update high watermark before we lower rss */
1522 update_hiwater_rss(mm);
1523
1524 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1525 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1526 if (PageHuge(page)) {
1527 hugetlb_count_sub(compound_nr(page), mm);
1528 set_huge_swap_pte_at(mm, address,
1529 pvmw.pte, pteval,
1530 vma_mmu_pagesize(vma));
1531 } else {
1532 dec_mm_counter(mm, mm_counter(page));
1533 set_pte_at(mm, address, pvmw.pte, pteval);
1534 }
1535
1536 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1537 /*
1538 * The guest indicated that the page content is of no
1539 * interest anymore. Simply discard the pte, vmscan
1540 * will take care of the rest.
1541 * A future reference will then fault in a new zero
1542 * page. When userfaultfd is active, we must not drop
1543 * this page though, as its main user (postcopy
1544 * migration) will not expect userfaults on already
1545 * copied pages.
1546 */
1547 dec_mm_counter(mm, mm_counter(page));
1548 /* We have to invalidate as we cleared the pte */
1549 mmu_notifier_invalidate_range(mm, address,
1550 address + PAGE_SIZE);
1551 } else if (PageAnon(page)) {
1552 swp_entry_t entry = { .val = page_private(subpage) };
1553 pte_t swp_pte;
1554 /*
1555 * Store the swap location in the pte.
1556 * See handle_pte_fault() ...
1557 */
1558 if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1559 WARN_ON_ONCE(1);
1560 ret = false;
1561 /* We have to invalidate as we cleared the pte */
1562 mmu_notifier_invalidate_range(mm, address,
1563 address + PAGE_SIZE);
1564 page_vma_mapped_walk_done(&pvmw);
1565 break;
1566 }
1567
1568 /* MADV_FREE page check */
1569 if (!PageSwapBacked(page)) {
1570 if (!PageDirty(page)) {
1571 /* Invalidate as we cleared the pte */
1572 mmu_notifier_invalidate_range(mm,
1573 address, address + PAGE_SIZE);
1574 dec_mm_counter(mm, MM_ANONPAGES);
1575 goto discard;
1576 }
1577
1578 /*
1579 * If the page was redirtied, it cannot be
1580 * discarded. Remap the page to page table.
1581 */
1582 set_pte_at(mm, address, pvmw.pte, pteval);
1583 SetPageSwapBacked(page);
1584 ret = false;
1585 page_vma_mapped_walk_done(&pvmw);
1586 break;
1587 }
1588
1589 if (swap_duplicate(entry) < 0) {
1590 set_pte_at(mm, address, pvmw.pte, pteval);
1591 ret = false;
1592 page_vma_mapped_walk_done(&pvmw);
1593 break;
1594 }
1595 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1596 set_pte_at(mm, address, pvmw.pte, pteval);
1597 ret = false;
1598 page_vma_mapped_walk_done(&pvmw);
1599 break;
1600 }
1601 if (list_empty(&mm->mmlist)) {
1602 spin_lock(&mmlist_lock);
1603 if (list_empty(&mm->mmlist))
1604 list_add(&mm->mmlist, &init_mm.mmlist);
1605 spin_unlock(&mmlist_lock);
1606 }
1607 dec_mm_counter(mm, MM_ANONPAGES);
1608 inc_mm_counter(mm, MM_SWAPENTS);
1609 swp_pte = swp_entry_to_pte(entry);
1610 if (pte_soft_dirty(pteval))
1611 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1612 if (pte_uffd_wp(pteval))
1613 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1614 set_pte_at(mm, address, pvmw.pte, swp_pte);
1615 /* Invalidate as we cleared the pte */
1616 mmu_notifier_invalidate_range(mm, address,
1617 address + PAGE_SIZE);
1618 } else {
1619 /*
1620 * This is a locked file-backed page, thus it cannot
1621 * be removed from the page cache and replaced by a new
1622 * page before mmu_notifier_invalidate_range_end, so no
1623 * concurrent thread might update its page table to
1624 * point at new page while a device still is using this
1625 * page.
1626 *
1627 * See Documentation/vm/mmu_notifier.rst
1628 */
1629 dec_mm_counter(mm, mm_counter_file(page));
1630 }
1631discard:
1632 /*
1633 * No need to call mmu_notifier_invalidate_range() it has be
1634 * done above for all cases requiring it to happen under page
1635 * table lock before mmu_notifier_invalidate_range_end()
1636 *
1637 * See Documentation/vm/mmu_notifier.rst
1638 */
1639 page_remove_rmap(subpage, PageHuge(page));
1640 put_page(page);
1641 }
1642
1643 mmu_notifier_invalidate_range_end(&range);
1644
1645 return ret;
1646}
1647
1648static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1649{
1650 return vma_is_temporary_stack(vma);
1651}
1652
1653static int page_not_mapped(struct page *page)
1654{
1655 return !page_mapped(page);
1656}
1657
1658/**
1659 * try_to_unmap - try to remove all page table mappings to a page
1660 * @page: the page to get unmapped
1661 * @flags: action and flags
1662 *
1663 * Tries to remove all the page table entries which are mapping this
1664 * page, used in the pageout path. Caller must hold the page lock.
1665 *
1666 * It is the caller's responsibility to check if the page is still
1667 * mapped when needed (use TTU_SYNC to prevent accounting races).
1668 */
1669void try_to_unmap(struct page *page, enum ttu_flags flags)
1670{
1671 struct rmap_walk_control rwc = {
1672 .rmap_one = try_to_unmap_one,
1673 .arg = (void *)flags,
1674 .done = page_not_mapped,
1675 .anon_lock = page_lock_anon_vma_read,
1676 };
1677
1678 if (flags & TTU_RMAP_LOCKED)
1679 rmap_walk_locked(page, &rwc);
1680 else
1681 rmap_walk(page, &rwc);
1682}
1683
1684/*
1685 * @arg: enum ttu_flags will be passed to this argument.
1686 *
1687 * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
1688 * containing migration entries.
1689 */
1690static bool try_to_migrate_one(struct page *page, struct vm_area_struct *vma,
1691 unsigned long address, void *arg)
1692{
1693 struct mm_struct *mm = vma->vm_mm;
1694 struct page_vma_mapped_walk pvmw = {
1695 .page = page,
1696 .vma = vma,
1697 .address = address,
1698 };
1699 pte_t pteval;
1700 struct page *subpage;
1701 bool ret = true;
1702 struct mmu_notifier_range range;
1703 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1704
1705 /*
1706 * When racing against e.g. zap_pte_range() on another cpu,
1707 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1708 * try_to_migrate() may return before page_mapped() has become false,
1709 * if page table locking is skipped: use TTU_SYNC to wait for that.
1710 */
1711 if (flags & TTU_SYNC)
1712 pvmw.flags = PVMW_SYNC;
1713
1714 /*
1715 * unmap_page() in mm/huge_memory.c is the only user of migration with
1716 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
1717 */
1718 if (flags & TTU_SPLIT_HUGE_PMD)
1719 split_huge_pmd_address(vma, address, true, page);
1720
1721 /*
1722 * For THP, we have to assume the worse case ie pmd for invalidation.
1723 * For hugetlb, it could be much worse if we need to do pud
1724 * invalidation in the case of pmd sharing.
1725 *
1726 * Note that the page can not be free in this function as call of
1727 * try_to_unmap() must hold a reference on the page.
1728 */
1729 range.end = PageKsm(page) ?
1730 address + PAGE_SIZE : vma_address_end(page, vma);
1731 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1732 address, range.end);
1733 if (PageHuge(page)) {
1734 /*
1735 * If sharing is possible, start and end will be adjusted
1736 * accordingly.
1737 */
1738 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1739 &range.end);
1740 }
1741 mmu_notifier_invalidate_range_start(&range);
1742
1743 while (page_vma_mapped_walk(&pvmw)) {
1744#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1745 /* PMD-mapped THP migration entry */
1746 if (!pvmw.pte) {
1747 VM_BUG_ON_PAGE(PageHuge(page) ||
1748 !PageTransCompound(page), page);
1749
1750 set_pmd_migration_entry(&pvmw, page);
1751 continue;
1752 }
1753#endif
1754
1755 /* Unexpected PMD-mapped THP? */
1756 VM_BUG_ON_PAGE(!pvmw.pte, page);
1757
1758 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1759 address = pvmw.address;
1760
1761 if (PageHuge(page) && !PageAnon(page)) {
1762 /*
1763 * To call huge_pmd_unshare, i_mmap_rwsem must be
1764 * held in write mode. Caller needs to explicitly
1765 * do this outside rmap routines.
1766 */
1767 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1768 if (huge_pmd_unshare(mm, vma, &address, pvmw.pte)) {
1769 /*
1770 * huge_pmd_unshare unmapped an entire PMD
1771 * page. There is no way of knowing exactly
1772 * which PMDs may be cached for this mm, so
1773 * we must flush them all. start/end were
1774 * already adjusted above to cover this range.
1775 */
1776 flush_cache_range(vma, range.start, range.end);
1777 flush_tlb_range(vma, range.start, range.end);
1778 mmu_notifier_invalidate_range(mm, range.start,
1779 range.end);
1780
1781 /*
1782 * The ref count of the PMD page was dropped
1783 * which is part of the way map counting
1784 * is done for shared PMDs. Return 'true'
1785 * here. When there is no other sharing,
1786 * huge_pmd_unshare returns false and we will
1787 * unmap the actual page and drop map count
1788 * to zero.
1789 */
1790 page_vma_mapped_walk_done(&pvmw);
1791 break;
1792 }
1793 }
1794
1795 /* Nuke the page table entry. */
1796 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1797 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1798
1799 /* Move the dirty bit to the page. Now the pte is gone. */
1800 if (pte_dirty(pteval))
1801 set_page_dirty(page);
1802
1803 /* Update high watermark before we lower rss */
1804 update_hiwater_rss(mm);
1805
1806 if (is_zone_device_page(page)) {
1807 swp_entry_t entry;
1808 pte_t swp_pte;
1809
1810 /*
1811 * Store the pfn of the page in a special migration
1812 * pte. do_swap_page() will wait until the migration
1813 * pte is removed and then restart fault handling.
1814 */
1815 entry = make_readable_migration_entry(
1816 page_to_pfn(page));
1817 swp_pte = swp_entry_to_pte(entry);
1818
1819 /*
1820 * pteval maps a zone device page and is therefore
1821 * a swap pte.
1822 */
1823 if (pte_swp_soft_dirty(pteval))
1824 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1825 if (pte_swp_uffd_wp(pteval))
1826 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1827 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1828 /*
1829 * No need to invalidate here it will synchronize on
1830 * against the special swap migration pte.
1831 *
1832 * The assignment to subpage above was computed from a
1833 * swap PTE which results in an invalid pointer.
1834 * Since only PAGE_SIZE pages can currently be
1835 * migrated, just set it to page. This will need to be
1836 * changed when hugepage migrations to device private
1837 * memory are supported.
1838 */
1839 subpage = page;
1840 } else if (PageHWPoison(page)) {
1841 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1842 if (PageHuge(page)) {
1843 hugetlb_count_sub(compound_nr(page), mm);
1844 set_huge_swap_pte_at(mm, address,
1845 pvmw.pte, pteval,
1846 vma_mmu_pagesize(vma));
1847 } else {
1848 dec_mm_counter(mm, mm_counter(page));
1849 set_pte_at(mm, address, pvmw.pte, pteval);
1850 }
1851
1852 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1853 /*
1854 * The guest indicated that the page content is of no
1855 * interest anymore. Simply discard the pte, vmscan
1856 * will take care of the rest.
1857 * A future reference will then fault in a new zero
1858 * page. When userfaultfd is active, we must not drop
1859 * this page though, as its main user (postcopy
1860 * migration) will not expect userfaults on already
1861 * copied pages.
1862 */
1863 dec_mm_counter(mm, mm_counter(page));
1864 /* We have to invalidate as we cleared the pte */
1865 mmu_notifier_invalidate_range(mm, address,
1866 address + PAGE_SIZE);
1867 } else {
1868 swp_entry_t entry;
1869 pte_t swp_pte;
1870
1871 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1872 set_pte_at(mm, address, pvmw.pte, pteval);
1873 ret = false;
1874 page_vma_mapped_walk_done(&pvmw);
1875 break;
1876 }
1877
1878 /*
1879 * Store the pfn of the page in a special migration
1880 * pte. do_swap_page() will wait until the migration
1881 * pte is removed and then restart fault handling.
1882 */
1883 if (pte_write(pteval))
1884 entry = make_writable_migration_entry(
1885 page_to_pfn(subpage));
1886 else
1887 entry = make_readable_migration_entry(
1888 page_to_pfn(subpage));
1889
1890 swp_pte = swp_entry_to_pte(entry);
1891 if (pte_soft_dirty(pteval))
1892 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1893 if (pte_uffd_wp(pteval))
1894 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1895 set_pte_at(mm, address, pvmw.pte, swp_pte);
1896 /*
1897 * No need to invalidate here it will synchronize on
1898 * against the special swap migration pte.
1899 */
1900 }
1901
1902 /*
1903 * No need to call mmu_notifier_invalidate_range() it has be
1904 * done above for all cases requiring it to happen under page
1905 * table lock before mmu_notifier_invalidate_range_end()
1906 *
1907 * See Documentation/vm/mmu_notifier.rst
1908 */
1909 page_remove_rmap(subpage, PageHuge(page));
1910 put_page(page);
1911 }
1912
1913 mmu_notifier_invalidate_range_end(&range);
1914
1915 return ret;
1916}
1917
1918/**
1919 * try_to_migrate - try to replace all page table mappings with swap entries
1920 * @page: the page to replace page table entries for
1921 * @flags: action and flags
1922 *
1923 * Tries to remove all the page table entries which are mapping this page and
1924 * replace them with special swap entries. Caller must hold the page lock.
1925 */
1926void try_to_migrate(struct page *page, enum ttu_flags flags)
1927{
1928 struct rmap_walk_control rwc = {
1929 .rmap_one = try_to_migrate_one,
1930 .arg = (void *)flags,
1931 .done = page_not_mapped,
1932 .anon_lock = page_lock_anon_vma_read,
1933 };
1934
1935 /*
1936 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
1937 * TTU_SPLIT_HUGE_PMD and TTU_SYNC flags.
1938 */
1939 if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
1940 TTU_SYNC)))
1941 return;
1942
1943 if (is_zone_device_page(page) && !is_device_private_page(page))
1944 return;
1945
1946 /*
1947 * During exec, a temporary VMA is setup and later moved.
1948 * The VMA is moved under the anon_vma lock but not the
1949 * page tables leading to a race where migration cannot
1950 * find the migration ptes. Rather than increasing the
1951 * locking requirements of exec(), migration skips
1952 * temporary VMAs until after exec() completes.
1953 */
1954 if (!PageKsm(page) && PageAnon(page))
1955 rwc.invalid_vma = invalid_migration_vma;
1956
1957 if (flags & TTU_RMAP_LOCKED)
1958 rmap_walk_locked(page, &rwc);
1959 else
1960 rmap_walk(page, &rwc);
1961}
1962
1963/*
1964 * Walks the vma's mapping a page and mlocks the page if any locked vma's are
1965 * found. Once one is found the page is locked and the scan can be terminated.
1966 */
1967static bool page_mlock_one(struct page *page, struct vm_area_struct *vma,
1968 unsigned long address, void *unused)
1969{
1970 struct page_vma_mapped_walk pvmw = {
1971 .page = page,
1972 .vma = vma,
1973 .address = address,
1974 };
1975
1976 /* An un-locked vma doesn't have any pages to lock, continue the scan */
1977 if (!(vma->vm_flags & VM_LOCKED))
1978 return true;
1979
1980 while (page_vma_mapped_walk(&pvmw)) {
1981 /*
1982 * Need to recheck under the ptl to serialise with
1983 * __munlock_pagevec_fill() after VM_LOCKED is cleared in
1984 * munlock_vma_pages_range().
1985 */
1986 if (vma->vm_flags & VM_LOCKED) {
1987 /*
1988 * PTE-mapped THP are never marked as mlocked; but
1989 * this function is never called on a DoubleMap THP,
1990 * nor on an Anon THP (which may still be PTE-mapped
1991 * after DoubleMap was cleared).
1992 */
1993 mlock_vma_page(page);
1994 /*
1995 * No need to scan further once the page is marked
1996 * as mlocked.
1997 */
1998 page_vma_mapped_walk_done(&pvmw);
1999 return false;
2000 }
2001 }
2002
2003 return true;
2004}
2005
2006/**
2007 * page_mlock - try to mlock a page
2008 * @page: the page to be mlocked
2009 *
2010 * Called from munlock code. Checks all of the VMAs mapping the page and mlocks
2011 * the page if any are found. The page will be returned with PG_mlocked cleared
2012 * if it is not mapped by any locked vmas.
2013 */
2014void page_mlock(struct page *page)
2015{
2016 struct rmap_walk_control rwc = {
2017 .rmap_one = page_mlock_one,
2018 .done = page_not_mapped,
2019 .anon_lock = page_lock_anon_vma_read,
2020
2021 };
2022
2023 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
2024 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
2025
2026 /* Anon THP are only marked as mlocked when singly mapped */
2027 if (PageTransCompound(page) && PageAnon(page))
2028 return;
2029
2030 rmap_walk(page, &rwc);
2031}
2032
2033#ifdef CONFIG_DEVICE_PRIVATE
2034struct make_exclusive_args {
2035 struct mm_struct *mm;
2036 unsigned long address;
2037 void *owner;
2038 bool valid;
2039};
2040
2041static bool page_make_device_exclusive_one(struct page *page,
2042 struct vm_area_struct *vma, unsigned long address, void *priv)
2043{
2044 struct mm_struct *mm = vma->vm_mm;
2045 struct page_vma_mapped_walk pvmw = {
2046 .page = page,
2047 .vma = vma,
2048 .address = address,
2049 };
2050 struct make_exclusive_args *args = priv;
2051 pte_t pteval;
2052 struct page *subpage;
2053 bool ret = true;
2054 struct mmu_notifier_range range;
2055 swp_entry_t entry;
2056 pte_t swp_pte;
2057
2058 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
2059 vma->vm_mm, address, min(vma->vm_end,
2060 address + page_size(page)), args->owner);
2061 mmu_notifier_invalidate_range_start(&range);
2062
2063 while (page_vma_mapped_walk(&pvmw)) {
2064 /* Unexpected PMD-mapped THP? */
2065 VM_BUG_ON_PAGE(!pvmw.pte, page);
2066
2067 if (!pte_present(*pvmw.pte)) {
2068 ret = false;
2069 page_vma_mapped_walk_done(&pvmw);
2070 break;
2071 }
2072
2073 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
2074 address = pvmw.address;
2075
2076 /* Nuke the page table entry. */
2077 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
2078 pteval = ptep_clear_flush(vma, address, pvmw.pte);
2079
2080 /* Move the dirty bit to the page. Now the pte is gone. */
2081 if (pte_dirty(pteval))
2082 set_page_dirty(page);
2083
2084 /*
2085 * Check that our target page is still mapped at the expected
2086 * address.
2087 */
2088 if (args->mm == mm && args->address == address &&
2089 pte_write(pteval))
2090 args->valid = true;
2091
2092 /*
2093 * Store the pfn of the page in a special migration
2094 * pte. do_swap_page() will wait until the migration
2095 * pte is removed and then restart fault handling.
2096 */
2097 if (pte_write(pteval))
2098 entry = make_writable_device_exclusive_entry(
2099 page_to_pfn(subpage));
2100 else
2101 entry = make_readable_device_exclusive_entry(
2102 page_to_pfn(subpage));
2103 swp_pte = swp_entry_to_pte(entry);
2104 if (pte_soft_dirty(pteval))
2105 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2106 if (pte_uffd_wp(pteval))
2107 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2108
2109 set_pte_at(mm, address, pvmw.pte, swp_pte);
2110
2111 /*
2112 * There is a reference on the page for the swap entry which has
2113 * been removed, so shouldn't take another.
2114 */
2115 page_remove_rmap(subpage, false);
2116 }
2117
2118 mmu_notifier_invalidate_range_end(&range);
2119
2120 return ret;
2121}
2122
2123/**
2124 * page_make_device_exclusive - mark the page exclusively owned by a device
2125 * @page: the page to replace page table entries for
2126 * @mm: the mm_struct where the page is expected to be mapped
2127 * @address: address where the page is expected to be mapped
2128 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2129 *
2130 * Tries to remove all the page table entries which are mapping this page and
2131 * replace them with special device exclusive swap entries to grant a device
2132 * exclusive access to the page. Caller must hold the page lock.
2133 *
2134 * Returns false if the page is still mapped, or if it could not be unmapped
2135 * from the expected address. Otherwise returns true (success).
2136 */
2137static bool page_make_device_exclusive(struct page *page, struct mm_struct *mm,
2138 unsigned long address, void *owner)
2139{
2140 struct make_exclusive_args args = {
2141 .mm = mm,
2142 .address = address,
2143 .owner = owner,
2144 .valid = false,
2145 };
2146 struct rmap_walk_control rwc = {
2147 .rmap_one = page_make_device_exclusive_one,
2148 .done = page_not_mapped,
2149 .anon_lock = page_lock_anon_vma_read,
2150 .arg = &args,
2151 };
2152
2153 /*
2154 * Restrict to anonymous pages for now to avoid potential writeback
2155 * issues. Also tail pages shouldn't be passed to rmap_walk so skip
2156 * those.
2157 */
2158 if (!PageAnon(page) || PageTail(page))
2159 return false;
2160
2161 rmap_walk(page, &rwc);
2162
2163 return args.valid && !page_mapcount(page);
2164}
2165
2166/**
2167 * make_device_exclusive_range() - Mark a range for exclusive use by a device
2168 * @mm: mm_struct of assoicated target process
2169 * @start: start of the region to mark for exclusive device access
2170 * @end: end address of region
2171 * @pages: returns the pages which were successfully marked for exclusive access
2172 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2173 *
2174 * Returns: number of pages found in the range by GUP. A page is marked for
2175 * exclusive access only if the page pointer is non-NULL.
2176 *
2177 * This function finds ptes mapping page(s) to the given address range, locks
2178 * them and replaces mappings with special swap entries preventing userspace CPU
2179 * access. On fault these entries are replaced with the original mapping after
2180 * calling MMU notifiers.
2181 *
2182 * A driver using this to program access from a device must use a mmu notifier
2183 * critical section to hold a device specific lock during programming. Once
2184 * programming is complete it should drop the page lock and reference after
2185 * which point CPU access to the page will revoke the exclusive access.
2186 */
2187int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
2188 unsigned long end, struct page **pages,
2189 void *owner)
2190{
2191 long npages = (end - start) >> PAGE_SHIFT;
2192 long i;
2193
2194 npages = get_user_pages_remote(mm, start, npages,
2195 FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
2196 pages, NULL, NULL);
2197 if (npages < 0)
2198 return npages;
2199
2200 for (i = 0; i < npages; i++, start += PAGE_SIZE) {
2201 if (!trylock_page(pages[i])) {
2202 put_page(pages[i]);
2203 pages[i] = NULL;
2204 continue;
2205 }
2206
2207 if (!page_make_device_exclusive(pages[i], mm, start, owner)) {
2208 unlock_page(pages[i]);
2209 put_page(pages[i]);
2210 pages[i] = NULL;
2211 }
2212 }
2213
2214 return npages;
2215}
2216EXPORT_SYMBOL_GPL(make_device_exclusive_range);
2217#endif
2218
2219void __put_anon_vma(struct anon_vma *anon_vma)
2220{
2221 struct anon_vma *root = anon_vma->root;
2222
2223 anon_vma_free(anon_vma);
2224 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
2225 anon_vma_free(root);
2226}
2227
2228static struct anon_vma *rmap_walk_anon_lock(struct page *page,
2229 struct rmap_walk_control *rwc)
2230{
2231 struct anon_vma *anon_vma;
2232
2233 if (rwc->anon_lock)
2234 return rwc->anon_lock(page);
2235
2236 /*
2237 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
2238 * because that depends on page_mapped(); but not all its usages
2239 * are holding mmap_lock. Users without mmap_lock are required to
2240 * take a reference count to prevent the anon_vma disappearing
2241 */
2242 anon_vma = page_anon_vma(page);
2243 if (!anon_vma)
2244 return NULL;
2245
2246 anon_vma_lock_read(anon_vma);
2247 return anon_vma;
2248}
2249
2250/*
2251 * rmap_walk_anon - do something to anonymous page using the object-based
2252 * rmap method
2253 * @page: the page to be handled
2254 * @rwc: control variable according to each walk type
2255 *
2256 * Find all the mappings of a page using the mapping pointer and the vma chains
2257 * contained in the anon_vma struct it points to.
2258 *
2259 * When called from page_mlock(), the mmap_lock of the mm containing the vma
2260 * where the page was found will be held for write. So, we won't recheck
2261 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
2262 * LOCKED.
2263 */
2264static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
2265 bool locked)
2266{
2267 struct anon_vma *anon_vma;
2268 pgoff_t pgoff_start, pgoff_end;
2269 struct anon_vma_chain *avc;
2270
2271 if (locked) {
2272 anon_vma = page_anon_vma(page);
2273 /* anon_vma disappear under us? */
2274 VM_BUG_ON_PAGE(!anon_vma, page);
2275 } else {
2276 anon_vma = rmap_walk_anon_lock(page, rwc);
2277 }
2278 if (!anon_vma)
2279 return;
2280
2281 pgoff_start = page_to_pgoff(page);
2282 pgoff_end = pgoff_start + thp_nr_pages(page) - 1;
2283 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
2284 pgoff_start, pgoff_end) {
2285 struct vm_area_struct *vma = avc->vma;
2286 unsigned long address = vma_address(page, vma);
2287
2288 VM_BUG_ON_VMA(address == -EFAULT, vma);
2289 cond_resched();
2290
2291 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2292 continue;
2293
2294 if (!rwc->rmap_one(page, vma, address, rwc->arg))
2295 break;
2296 if (rwc->done && rwc->done(page))
2297 break;
2298 }
2299
2300 if (!locked)
2301 anon_vma_unlock_read(anon_vma);
2302}
2303
2304/*
2305 * rmap_walk_file - do something to file page using the object-based rmap method
2306 * @page: the page to be handled
2307 * @rwc: control variable according to each walk type
2308 *
2309 * Find all the mappings of a page using the mapping pointer and the vma chains
2310 * contained in the address_space struct it points to.
2311 *
2312 * When called from page_mlock(), the mmap_lock of the mm containing the vma
2313 * where the page was found will be held for write. So, we won't recheck
2314 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
2315 * LOCKED.
2316 */
2317static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
2318 bool locked)
2319{
2320 struct address_space *mapping = page_mapping(page);
2321 pgoff_t pgoff_start, pgoff_end;
2322 struct vm_area_struct *vma;
2323
2324 /*
2325 * The page lock not only makes sure that page->mapping cannot
2326 * suddenly be NULLified by truncation, it makes sure that the
2327 * structure at mapping cannot be freed and reused yet,
2328 * so we can safely take mapping->i_mmap_rwsem.
2329 */
2330 VM_BUG_ON_PAGE(!PageLocked(page), page);
2331
2332 if (!mapping)
2333 return;
2334
2335 pgoff_start = page_to_pgoff(page);
2336 pgoff_end = pgoff_start + thp_nr_pages(page) - 1;
2337 if (!locked)
2338 i_mmap_lock_read(mapping);
2339 vma_interval_tree_foreach(vma, &mapping->i_mmap,
2340 pgoff_start, pgoff_end) {
2341 unsigned long address = vma_address(page, vma);
2342
2343 VM_BUG_ON_VMA(address == -EFAULT, vma);
2344 cond_resched();
2345
2346 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2347 continue;
2348
2349 if (!rwc->rmap_one(page, vma, address, rwc->arg))
2350 goto done;
2351 if (rwc->done && rwc->done(page))
2352 goto done;
2353 }
2354
2355done:
2356 if (!locked)
2357 i_mmap_unlock_read(mapping);
2358}
2359
2360void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
2361{
2362 if (unlikely(PageKsm(page)))
2363 rmap_walk_ksm(page, rwc);
2364 else if (PageAnon(page))
2365 rmap_walk_anon(page, rwc, false);
2366 else
2367 rmap_walk_file(page, rwc, false);
2368}
2369
2370/* Like rmap_walk, but caller holds relevant rmap lock */
2371void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
2372{
2373 /* no ksm support for now */
2374 VM_BUG_ON_PAGE(PageKsm(page), page);
2375 if (PageAnon(page))
2376 rmap_walk_anon(page, rwc, true);
2377 else
2378 rmap_walk_file(page, rwc, true);
2379}
2380
2381#ifdef CONFIG_HUGETLB_PAGE
2382/*
2383 * The following two functions are for anonymous (private mapped) hugepages.
2384 * Unlike common anonymous pages, anonymous hugepages have no accounting code
2385 * and no lru code, because we handle hugepages differently from common pages.
2386 */
2387void hugepage_add_anon_rmap(struct page *page,
2388 struct vm_area_struct *vma, unsigned long address)
2389{
2390 struct anon_vma *anon_vma = vma->anon_vma;
2391 int first;
2392
2393 BUG_ON(!PageLocked(page));
2394 BUG_ON(!anon_vma);
2395 /* address might be in next vma when migration races vma_adjust */
2396 first = atomic_inc_and_test(compound_mapcount_ptr(page));
2397 if (first)
2398 __page_set_anon_rmap(page, vma, address, 0);
2399}
2400
2401void hugepage_add_new_anon_rmap(struct page *page,
2402 struct vm_area_struct *vma, unsigned long address)
2403{
2404 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2405 atomic_set(compound_mapcount_ptr(page), 0);
2406 if (hpage_pincount_available(page))
2407 atomic_set(compound_pincount_ptr(page), 0);
2408
2409 __page_set_anon_rmap(page, vma, address, 1);
2410}
2411#endif /* CONFIG_HUGETLB_PAGE */