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