mm,thp,rmap: simplify compound page mapcount handling
[linux-block.git] / mm / rmap.c
CommitLineData
1da177e4
LT
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
98f32602 17 * Contributions by Hugh Dickins 2003, 2004
1da177e4
LT
18 */
19
20/*
21 * Lock ordering in mm:
22 *
9608703e 23 * inode->i_rwsem (while writing or truncating, not reading or faulting)
c1e8d7c6 24 * mm->mmap_lock
730633f0 25 * mapping->invalidate_lock (in filemap_fault)
3a47c54f 26 * page->flags PG_locked (lock_page)
8d9bfb26 27 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share, see hugetlbfs below)
730633f0 28 * mapping->i_mmap_rwsem
730633f0
JK
29 * anon_vma->rwsem
30 * mm->page_table_lock or pte_lock
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
e621900a
MWO
33 * mapping->private_lock (in block_dirty_folio)
34 * folio_lock_memcg move_lock (in block_dirty_folio)
730633f0 35 * i_pages lock (widely used)
e809c3fe 36 * lruvec->lru_lock (in folio_lruvec_lock_irq)
730633f0
JK
37 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
38 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
39 * sb_lock (within inode_lock in fs/fs-writeback.c)
40 * i_pages lock (widely used, in set_page_dirty,
41 * in arch-dependent flush_dcache_mmap_lock,
42 * within bdi.wb->list_lock in __sync_single_inode)
6a46079c 43 *
9608703e 44 * anon_vma->rwsem,mapping->i_mmap_rwsem (memory_failure, collect_procs_anon)
9b679320 45 * ->tasklist_lock
6a46079c 46 * pte map lock
c0d0381a 47 *
8d9bfb26
MK
48 * hugetlbfs PageHuge() take locks in this order:
49 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
50 * vma_lock (hugetlb specific lock for pmd_sharing)
51 * mapping->i_mmap_rwsem (also used for hugetlb pmd sharing)
52 * page->flags PG_locked (lock_page)
1da177e4
LT
53 */
54
55#include <linux/mm.h>
6e84f315 56#include <linux/sched/mm.h>
29930025 57#include <linux/sched/task.h>
1da177e4
LT
58#include <linux/pagemap.h>
59#include <linux/swap.h>
60#include <linux/swapops.h>
61#include <linux/slab.h>
62#include <linux/init.h>
5ad64688 63#include <linux/ksm.h>
1da177e4
LT
64#include <linux/rmap.h>
65#include <linux/rcupdate.h>
b95f1b31 66#include <linux/export.h>
8a9f3ccd 67#include <linux/memcontrol.h>
cddb8a5c 68#include <linux/mmu_notifier.h>
64cdd548 69#include <linux/migrate.h>
0fe6e20b 70#include <linux/hugetlb.h>
444f84fd 71#include <linux/huge_mm.h>
ef5d437f 72#include <linux/backing-dev.h>
33c3fc71 73#include <linux/page_idle.h>
a5430dda 74#include <linux/memremap.h>
bce73e48 75#include <linux/userfaultfd_k.h>
999dad82 76#include <linux/mm_inline.h>
1da177e4
LT
77
78#include <asm/tlbflush.h>
79
4cc79b33 80#define CREATE_TRACE_POINTS
72b252ae 81#include <trace/events/tlb.h>
4cc79b33 82#include <trace/events/migrate.h>
72b252ae 83
b291f000
NP
84#include "internal.h"
85
fdd2e5f8 86static struct kmem_cache *anon_vma_cachep;
5beb4930 87static struct kmem_cache *anon_vma_chain_cachep;
fdd2e5f8
AB
88
89static inline struct anon_vma *anon_vma_alloc(void)
90{
01d8b20d
PZ
91 struct anon_vma *anon_vma;
92
93 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
94 if (anon_vma) {
95 atomic_set(&anon_vma->refcount, 1);
2555283e
JH
96 anon_vma->num_children = 0;
97 anon_vma->num_active_vmas = 0;
7a3ef208 98 anon_vma->parent = anon_vma;
01d8b20d
PZ
99 /*
100 * Initialise the anon_vma root to point to itself. If called
101 * from fork, the root will be reset to the parents anon_vma.
102 */
103 anon_vma->root = anon_vma;
104 }
105
106 return anon_vma;
fdd2e5f8
AB
107}
108
01d8b20d 109static inline void anon_vma_free(struct anon_vma *anon_vma)
fdd2e5f8 110{
01d8b20d 111 VM_BUG_ON(atomic_read(&anon_vma->refcount));
88c22088
PZ
112
113 /*
2f031c6f 114 * Synchronize against folio_lock_anon_vma_read() such that
88c22088
PZ
115 * we can safely hold the lock without the anon_vma getting
116 * freed.
117 *
118 * Relies on the full mb implied by the atomic_dec_and_test() from
119 * put_anon_vma() against the acquire barrier implied by
2f031c6f 120 * down_read_trylock() from folio_lock_anon_vma_read(). This orders:
88c22088 121 *
2f031c6f 122 * folio_lock_anon_vma_read() VS put_anon_vma()
4fc3f1d6 123 * down_read_trylock() atomic_dec_and_test()
88c22088 124 * LOCK MB
4fc3f1d6 125 * atomic_read() rwsem_is_locked()
88c22088
PZ
126 *
127 * LOCK should suffice since the actual taking of the lock must
128 * happen _before_ what follows.
129 */
7f39dda9 130 might_sleep();
5a505085 131 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
4fc3f1d6 132 anon_vma_lock_write(anon_vma);
08b52706 133 anon_vma_unlock_write(anon_vma);
88c22088
PZ
134 }
135
fdd2e5f8
AB
136 kmem_cache_free(anon_vma_cachep, anon_vma);
137}
1da177e4 138
dd34739c 139static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
5beb4930 140{
dd34739c 141 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
5beb4930
RR
142}
143
e574b5fd 144static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
5beb4930
RR
145{
146 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
147}
148
6583a843
KC
149static void anon_vma_chain_link(struct vm_area_struct *vma,
150 struct anon_vma_chain *avc,
151 struct anon_vma *anon_vma)
152{
153 avc->vma = vma;
154 avc->anon_vma = anon_vma;
155 list_add(&avc->same_vma, &vma->anon_vma_chain);
bf181b9f 156 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
6583a843
KC
157}
158
d9d332e0 159/**
d5a187da 160 * __anon_vma_prepare - attach an anon_vma to a memory region
d9d332e0
LT
161 * @vma: the memory region in question
162 *
163 * This makes sure the memory mapping described by 'vma' has
164 * an 'anon_vma' attached to it, so that we can associate the
165 * anonymous pages mapped into it with that anon_vma.
166 *
d5a187da
VB
167 * The common case will be that we already have one, which
168 * is handled inline by anon_vma_prepare(). But if
23a0790a 169 * not we either need to find an adjacent mapping that we
d9d332e0
LT
170 * can re-use the anon_vma from (very common when the only
171 * reason for splitting a vma has been mprotect()), or we
172 * allocate a new one.
173 *
174 * Anon-vma allocations are very subtle, because we may have
2f031c6f 175 * optimistically looked up an anon_vma in folio_lock_anon_vma_read()
aaf1f990 176 * and that may actually touch the rwsem even in the newly
d9d332e0
LT
177 * allocated vma (it depends on RCU to make sure that the
178 * anon_vma isn't actually destroyed).
179 *
180 * As a result, we need to do proper anon_vma locking even
181 * for the new allocation. At the same time, we do not want
182 * to do any locking for the common case of already having
183 * an anon_vma.
184 *
c1e8d7c6 185 * This must be called with the mmap_lock held for reading.
d9d332e0 186 */
d5a187da 187int __anon_vma_prepare(struct vm_area_struct *vma)
1da177e4 188{
d5a187da
VB
189 struct mm_struct *mm = vma->vm_mm;
190 struct anon_vma *anon_vma, *allocated;
5beb4930 191 struct anon_vma_chain *avc;
1da177e4
LT
192
193 might_sleep();
1da177e4 194
d5a187da
VB
195 avc = anon_vma_chain_alloc(GFP_KERNEL);
196 if (!avc)
197 goto out_enomem;
198
199 anon_vma = find_mergeable_anon_vma(vma);
200 allocated = NULL;
201 if (!anon_vma) {
202 anon_vma = anon_vma_alloc();
203 if (unlikely(!anon_vma))
204 goto out_enomem_free_avc;
2555283e 205 anon_vma->num_children++; /* self-parent link for new root */
d5a187da
VB
206 allocated = anon_vma;
207 }
5beb4930 208
d5a187da
VB
209 anon_vma_lock_write(anon_vma);
210 /* page_table_lock to protect against threads */
211 spin_lock(&mm->page_table_lock);
212 if (likely(!vma->anon_vma)) {
213 vma->anon_vma = anon_vma;
214 anon_vma_chain_link(vma, avc, anon_vma);
2555283e 215 anon_vma->num_active_vmas++;
d9d332e0 216 allocated = NULL;
d5a187da
VB
217 avc = NULL;
218 }
219 spin_unlock(&mm->page_table_lock);
220 anon_vma_unlock_write(anon_vma);
1da177e4 221
d5a187da
VB
222 if (unlikely(allocated))
223 put_anon_vma(allocated);
224 if (unlikely(avc))
225 anon_vma_chain_free(avc);
31f2b0eb 226
1da177e4 227 return 0;
5beb4930
RR
228
229 out_enomem_free_avc:
230 anon_vma_chain_free(avc);
231 out_enomem:
232 return -ENOMEM;
1da177e4
LT
233}
234
bb4aa396
LT
235/*
236 * This is a useful helper function for locking the anon_vma root as
237 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
238 * have the same vma.
239 *
240 * Such anon_vma's should have the same root, so you'd expect to see
241 * just a single mutex_lock for the whole traversal.
242 */
243static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
244{
245 struct anon_vma *new_root = anon_vma->root;
246 if (new_root != root) {
247 if (WARN_ON_ONCE(root))
5a505085 248 up_write(&root->rwsem);
bb4aa396 249 root = new_root;
5a505085 250 down_write(&root->rwsem);
bb4aa396
LT
251 }
252 return root;
253}
254
255static inline void unlock_anon_vma_root(struct anon_vma *root)
256{
257 if (root)
5a505085 258 up_write(&root->rwsem);
bb4aa396
LT
259}
260
5beb4930
RR
261/*
262 * Attach the anon_vmas from src to dst.
263 * Returns 0 on success, -ENOMEM on failure.
7a3ef208 264 *
cb152a1a 265 * anon_vma_clone() is called by __vma_adjust(), __split_vma(), copy_vma() and
47b390d2
WY
266 * anon_vma_fork(). The first three want an exact copy of src, while the last
267 * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
268 * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
269 * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
270 *
271 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
272 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
273 * This prevents degradation of anon_vma hierarchy to endless linear chain in
274 * case of constantly forking task. On the other hand, an anon_vma with more
275 * than one child isn't reused even if there was no alive vma, thus rmap
276 * walker has a good chance of avoiding scanning the whole hierarchy when it
277 * searches where page is mapped.
5beb4930
RR
278 */
279int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
1da177e4 280{
5beb4930 281 struct anon_vma_chain *avc, *pavc;
bb4aa396 282 struct anon_vma *root = NULL;
5beb4930 283
646d87b4 284 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
bb4aa396
LT
285 struct anon_vma *anon_vma;
286
dd34739c
LT
287 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
288 if (unlikely(!avc)) {
289 unlock_anon_vma_root(root);
290 root = NULL;
291 avc = anon_vma_chain_alloc(GFP_KERNEL);
292 if (!avc)
293 goto enomem_failure;
294 }
bb4aa396
LT
295 anon_vma = pavc->anon_vma;
296 root = lock_anon_vma_root(root, anon_vma);
297 anon_vma_chain_link(dst, avc, anon_vma);
7a3ef208
KK
298
299 /*
2555283e
JH
300 * Reuse existing anon_vma if it has no vma and only one
301 * anon_vma child.
7a3ef208 302 *
2555283e 303 * Root anon_vma is never reused:
7a3ef208
KK
304 * it has self-parent reference and at least one child.
305 */
47b390d2 306 if (!dst->anon_vma && src->anon_vma &&
2555283e
JH
307 anon_vma->num_children < 2 &&
308 anon_vma->num_active_vmas == 0)
7a3ef208 309 dst->anon_vma = anon_vma;
5beb4930 310 }
7a3ef208 311 if (dst->anon_vma)
2555283e 312 dst->anon_vma->num_active_vmas++;
bb4aa396 313 unlock_anon_vma_root(root);
5beb4930 314 return 0;
1da177e4 315
5beb4930 316 enomem_failure:
3fe89b3e 317 /*
d8e454eb
MW
318 * dst->anon_vma is dropped here otherwise its num_active_vmas can
319 * be incorrectly decremented in unlink_anon_vmas().
3fe89b3e
LY
320 * We can safely do this because callers of anon_vma_clone() don't care
321 * about dst->anon_vma if anon_vma_clone() failed.
322 */
323 dst->anon_vma = NULL;
5beb4930
RR
324 unlink_anon_vmas(dst);
325 return -ENOMEM;
1da177e4
LT
326}
327
5beb4930
RR
328/*
329 * Attach vma to its own anon_vma, as well as to the anon_vmas that
330 * the corresponding VMA in the parent process is attached to.
331 * Returns 0 on success, non-zero on failure.
332 */
333int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
1da177e4 334{
5beb4930
RR
335 struct anon_vma_chain *avc;
336 struct anon_vma *anon_vma;
c4ea95d7 337 int error;
1da177e4 338
5beb4930
RR
339 /* Don't bother if the parent process has no anon_vma here. */
340 if (!pvma->anon_vma)
341 return 0;
342
7a3ef208
KK
343 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
344 vma->anon_vma = NULL;
345
5beb4930
RR
346 /*
347 * First, attach the new VMA to the parent VMA's anon_vmas,
348 * so rmap can find non-COWed pages in child processes.
349 */
c4ea95d7
DF
350 error = anon_vma_clone(vma, pvma);
351 if (error)
352 return error;
5beb4930 353
7a3ef208
KK
354 /* An existing anon_vma has been reused, all done then. */
355 if (vma->anon_vma)
356 return 0;
357
5beb4930
RR
358 /* Then add our own anon_vma. */
359 anon_vma = anon_vma_alloc();
360 if (!anon_vma)
361 goto out_error;
2555283e 362 anon_vma->num_active_vmas++;
dd34739c 363 avc = anon_vma_chain_alloc(GFP_KERNEL);
5beb4930
RR
364 if (!avc)
365 goto out_error_free_anon_vma;
5c341ee1
RR
366
367 /*
aaf1f990 368 * The root anon_vma's rwsem is the lock actually used when we
5c341ee1
RR
369 * lock any of the anon_vmas in this anon_vma tree.
370 */
371 anon_vma->root = pvma->anon_vma->root;
7a3ef208 372 anon_vma->parent = pvma->anon_vma;
76545066 373 /*
01d8b20d
PZ
374 * With refcounts, an anon_vma can stay around longer than the
375 * process it belongs to. The root anon_vma needs to be pinned until
376 * this anon_vma is freed, because the lock lives in the root.
76545066
RR
377 */
378 get_anon_vma(anon_vma->root);
5beb4930
RR
379 /* Mark this anon_vma as the one where our new (COWed) pages go. */
380 vma->anon_vma = anon_vma;
4fc3f1d6 381 anon_vma_lock_write(anon_vma);
5c341ee1 382 anon_vma_chain_link(vma, avc, anon_vma);
2555283e 383 anon_vma->parent->num_children++;
08b52706 384 anon_vma_unlock_write(anon_vma);
5beb4930
RR
385
386 return 0;
387
388 out_error_free_anon_vma:
01d8b20d 389 put_anon_vma(anon_vma);
5beb4930 390 out_error:
4946d54c 391 unlink_anon_vmas(vma);
5beb4930 392 return -ENOMEM;
1da177e4
LT
393}
394
5beb4930
RR
395void unlink_anon_vmas(struct vm_area_struct *vma)
396{
397 struct anon_vma_chain *avc, *next;
eee2acba 398 struct anon_vma *root = NULL;
5beb4930 399
5c341ee1
RR
400 /*
401 * Unlink each anon_vma chained to the VMA. This list is ordered
402 * from newest to oldest, ensuring the root anon_vma gets freed last.
403 */
5beb4930 404 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
eee2acba
PZ
405 struct anon_vma *anon_vma = avc->anon_vma;
406
407 root = lock_anon_vma_root(root, anon_vma);
bf181b9f 408 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
eee2acba
PZ
409
410 /*
411 * Leave empty anon_vmas on the list - we'll need
412 * to free them outside the lock.
413 */
f808c13f 414 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
2555283e 415 anon_vma->parent->num_children--;
eee2acba 416 continue;
7a3ef208 417 }
eee2acba
PZ
418
419 list_del(&avc->same_vma);
420 anon_vma_chain_free(avc);
421 }
ee8ab190 422 if (vma->anon_vma) {
2555283e 423 vma->anon_vma->num_active_vmas--;
ee8ab190
LX
424
425 /*
426 * vma would still be needed after unlink, and anon_vma will be prepared
427 * when handle fault.
428 */
429 vma->anon_vma = NULL;
430 }
eee2acba
PZ
431 unlock_anon_vma_root(root);
432
433 /*
434 * Iterate the list once more, it now only contains empty and unlinked
435 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
5a505085 436 * needing to write-acquire the anon_vma->root->rwsem.
eee2acba
PZ
437 */
438 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
439 struct anon_vma *anon_vma = avc->anon_vma;
440
2555283e
JH
441 VM_WARN_ON(anon_vma->num_children);
442 VM_WARN_ON(anon_vma->num_active_vmas);
eee2acba
PZ
443 put_anon_vma(anon_vma);
444
5beb4930
RR
445 list_del(&avc->same_vma);
446 anon_vma_chain_free(avc);
447 }
448}
449
51cc5068 450static void anon_vma_ctor(void *data)
1da177e4 451{
a35afb83 452 struct anon_vma *anon_vma = data;
1da177e4 453
5a505085 454 init_rwsem(&anon_vma->rwsem);
83813267 455 atomic_set(&anon_vma->refcount, 0);
f808c13f 456 anon_vma->rb_root = RB_ROOT_CACHED;
1da177e4
LT
457}
458
459void __init anon_vma_init(void)
460{
461 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
5f0d5a3a 462 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
5d097056
VD
463 anon_vma_ctor);
464 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
465 SLAB_PANIC|SLAB_ACCOUNT);
1da177e4
LT
466}
467
468/*
6111e4ca
PZ
469 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
470 *
471 * Since there is no serialization what so ever against page_remove_rmap()
ad8a20cf
ML
472 * the best this function can do is return a refcount increased anon_vma
473 * that might have been relevant to this page.
6111e4ca
PZ
474 *
475 * The page might have been remapped to a different anon_vma or the anon_vma
476 * returned may already be freed (and even reused).
477 *
bc658c96
PZ
478 * In case it was remapped to a different anon_vma, the new anon_vma will be a
479 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
480 * ensure that any anon_vma obtained from the page will still be valid for as
481 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
482 *
6111e4ca
PZ
483 * All users of this function must be very careful when walking the anon_vma
484 * chain and verify that the page in question is indeed mapped in it
485 * [ something equivalent to page_mapped_in_vma() ].
486 *
091e4299
MC
487 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
488 * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
489 * if there is a mapcount, we can dereference the anon_vma after observing
490 * those.
1da177e4 491 */
29eea9b5 492struct anon_vma *folio_get_anon_vma(struct folio *folio)
1da177e4 493{
746b18d4 494 struct anon_vma *anon_vma = NULL;
1da177e4
LT
495 unsigned long anon_mapping;
496
497 rcu_read_lock();
29eea9b5 498 anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
3ca7b3c5 499 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
1da177e4 500 goto out;
29eea9b5 501 if (!folio_mapped(folio))
1da177e4
LT
502 goto out;
503
504 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
746b18d4
PZ
505 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
506 anon_vma = NULL;
507 goto out;
508 }
f1819427
HD
509
510 /*
29eea9b5 511 * If this folio is still mapped, then its anon_vma cannot have been
746b18d4
PZ
512 * freed. But if it has been unmapped, we have no security against the
513 * anon_vma structure being freed and reused (for another anon_vma:
5f0d5a3a 514 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
746b18d4 515 * above cannot corrupt).
f1819427 516 */
29eea9b5 517 if (!folio_mapped(folio)) {
7f39dda9 518 rcu_read_unlock();
746b18d4 519 put_anon_vma(anon_vma);
7f39dda9 520 return NULL;
746b18d4 521 }
1da177e4
LT
522out:
523 rcu_read_unlock();
746b18d4
PZ
524
525 return anon_vma;
526}
527
88c22088 528/*
29eea9b5 529 * Similar to folio_get_anon_vma() except it locks the anon_vma.
88c22088
PZ
530 *
531 * Its a little more complex as it tries to keep the fast path to a single
532 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
29eea9b5 533 * reference like with folio_get_anon_vma() and then block on the mutex
6d4675e6 534 * on !rwc->try_lock case.
88c22088 535 */
6d4675e6
MK
536struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
537 struct rmap_walk_control *rwc)
746b18d4 538{
88c22088 539 struct anon_vma *anon_vma = NULL;
eee0f252 540 struct anon_vma *root_anon_vma;
88c22088 541 unsigned long anon_mapping;
746b18d4 542
88c22088 543 rcu_read_lock();
9595d769 544 anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
88c22088
PZ
545 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
546 goto out;
9595d769 547 if (!folio_mapped(folio))
88c22088
PZ
548 goto out;
549
550 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
4db0c3c2 551 root_anon_vma = READ_ONCE(anon_vma->root);
4fc3f1d6 552 if (down_read_trylock(&root_anon_vma->rwsem)) {
88c22088 553 /*
9595d769 554 * If the folio is still mapped, then this anon_vma is still
eee0f252 555 * its anon_vma, and holding the mutex ensures that it will
bc658c96 556 * not go away, see anon_vma_free().
88c22088 557 */
9595d769 558 if (!folio_mapped(folio)) {
4fc3f1d6 559 up_read(&root_anon_vma->rwsem);
88c22088
PZ
560 anon_vma = NULL;
561 }
562 goto out;
563 }
746b18d4 564
6d4675e6
MK
565 if (rwc && rwc->try_lock) {
566 anon_vma = NULL;
567 rwc->contended = true;
568 goto out;
569 }
570
88c22088
PZ
571 /* trylock failed, we got to sleep */
572 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
573 anon_vma = NULL;
574 goto out;
575 }
576
9595d769 577 if (!folio_mapped(folio)) {
7f39dda9 578 rcu_read_unlock();
88c22088 579 put_anon_vma(anon_vma);
7f39dda9 580 return NULL;
88c22088
PZ
581 }
582
583 /* we pinned the anon_vma, its safe to sleep */
584 rcu_read_unlock();
4fc3f1d6 585 anon_vma_lock_read(anon_vma);
88c22088
PZ
586
587 if (atomic_dec_and_test(&anon_vma->refcount)) {
588 /*
589 * Oops, we held the last refcount, release the lock
590 * and bail -- can't simply use put_anon_vma() because
4fc3f1d6 591 * we'll deadlock on the anon_vma_lock_write() recursion.
88c22088 592 */
4fc3f1d6 593 anon_vma_unlock_read(anon_vma);
88c22088
PZ
594 __put_anon_vma(anon_vma);
595 anon_vma = NULL;
596 }
597
598 return anon_vma;
599
600out:
601 rcu_read_unlock();
746b18d4 602 return anon_vma;
34bbd704
ON
603}
604
72b252ae 605#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
72b252ae
MG
606/*
607 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
608 * important if a PTE was dirty when it was unmapped that it's flushed
609 * before any IO is initiated on the page to prevent lost writes. Similarly,
610 * it must be flushed before freeing to prevent data leakage.
611 */
612void try_to_unmap_flush(void)
613{
614 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
72b252ae
MG
615
616 if (!tlb_ubc->flush_required)
617 return;
618
e73ad5ff 619 arch_tlbbatch_flush(&tlb_ubc->arch);
72b252ae 620 tlb_ubc->flush_required = false;
d950c947 621 tlb_ubc->writable = false;
72b252ae
MG
622}
623
d950c947
MG
624/* Flush iff there are potentially writable TLB entries that can race with IO */
625void try_to_unmap_flush_dirty(void)
626{
627 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
628
629 if (tlb_ubc->writable)
630 try_to_unmap_flush();
631}
632
5ee2fa2f
HY
633/*
634 * Bits 0-14 of mm->tlb_flush_batched record pending generations.
635 * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations.
636 */
637#define TLB_FLUSH_BATCH_FLUSHED_SHIFT 16
638#define TLB_FLUSH_BATCH_PENDING_MASK \
639 ((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1)
640#define TLB_FLUSH_BATCH_PENDING_LARGE \
641 (TLB_FLUSH_BATCH_PENDING_MASK / 2)
642
c7ab0d2f 643static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
72b252ae
MG
644{
645 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
5ee2fa2f 646 int batch, nbatch;
72b252ae 647
e73ad5ff 648 arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
72b252ae 649 tlb_ubc->flush_required = true;
d950c947 650
3ea27719
MG
651 /*
652 * Ensure compiler does not re-order the setting of tlb_flush_batched
653 * before the PTE is cleared.
654 */
655 barrier();
5ee2fa2f
HY
656 batch = atomic_read(&mm->tlb_flush_batched);
657retry:
658 if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) {
659 /*
660 * Prevent `pending' from catching up with `flushed' because of
661 * overflow. Reset `pending' and `flushed' to be 1 and 0 if
662 * `pending' becomes large.
663 */
664 nbatch = atomic_cmpxchg(&mm->tlb_flush_batched, batch, 1);
665 if (nbatch != batch) {
666 batch = nbatch;
667 goto retry;
668 }
669 } else {
670 atomic_inc(&mm->tlb_flush_batched);
671 }
3ea27719 672
d950c947
MG
673 /*
674 * If the PTE was dirty then it's best to assume it's writable. The
675 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
676 * before the page is queued for IO.
677 */
678 if (writable)
679 tlb_ubc->writable = true;
72b252ae
MG
680}
681
682/*
683 * Returns true if the TLB flush should be deferred to the end of a batch of
684 * unmap operations to reduce IPIs.
685 */
686static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
687{
688 bool should_defer = false;
689
690 if (!(flags & TTU_BATCH_FLUSH))
691 return false;
692
693 /* If remote CPUs need to be flushed then defer batch the flush */
694 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
695 should_defer = true;
696 put_cpu();
697
698 return should_defer;
699}
3ea27719
MG
700
701/*
702 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
703 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
704 * operation such as mprotect or munmap to race between reclaim unmapping
705 * the page and flushing the page. If this race occurs, it potentially allows
706 * access to data via a stale TLB entry. Tracking all mm's that have TLB
707 * batching in flight would be expensive during reclaim so instead track
708 * whether TLB batching occurred in the past and if so then do a flush here
709 * if required. This will cost one additional flush per reclaim cycle paid
710 * by the first operation at risk such as mprotect and mumap.
711 *
712 * This must be called under the PTL so that an access to tlb_flush_batched
713 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
714 * via the PTL.
715 */
716void flush_tlb_batched_pending(struct mm_struct *mm)
717{
5ee2fa2f
HY
718 int batch = atomic_read(&mm->tlb_flush_batched);
719 int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK;
720 int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT;
3ea27719 721
5ee2fa2f
HY
722 if (pending != flushed) {
723 flush_tlb_mm(mm);
3ea27719 724 /*
5ee2fa2f
HY
725 * If the new TLB flushing is pending during flushing, leave
726 * mm->tlb_flush_batched as is, to avoid losing flushing.
3ea27719 727 */
5ee2fa2f
HY
728 atomic_cmpxchg(&mm->tlb_flush_batched, batch,
729 pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT));
3ea27719
MG
730 }
731}
72b252ae 732#else
c7ab0d2f 733static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
72b252ae
MG
734{
735}
736
737static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
738{
739 return false;
740}
741#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
742
1da177e4 743/*
bf89c8c8 744 * At what user virtual address is page expected in vma?
ab941e0f 745 * Caller should check the page is actually part of the vma.
1da177e4
LT
746 */
747unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
748{
e05b3453
MWO
749 struct folio *folio = page_folio(page);
750 if (folio_test_anon(folio)) {
751 struct anon_vma *page__anon_vma = folio_anon_vma(folio);
4829b906
HD
752 /*
753 * Note: swapoff's unuse_vma() is more efficient with this
754 * check, and needs it to match anon_vma when KSM is active.
755 */
756 if (!vma->anon_vma || !page__anon_vma ||
757 vma->anon_vma->root != page__anon_vma->root)
21d0d443 758 return -EFAULT;
31657170
JW
759 } else if (!vma->vm_file) {
760 return -EFAULT;
e05b3453 761 } else if (vma->vm_file->f_mapping != folio->mapping) {
1da177e4 762 return -EFAULT;
31657170 763 }
494334e4
HD
764
765 return vma_address(page, vma);
1da177e4
LT
766}
767
50722804
ZK
768/*
769 * Returns the actual pmd_t* where we expect 'address' to be mapped from, or
770 * NULL if it doesn't exist. No guarantees / checks on what the pmd_t*
771 * represents.
772 */
6219049a
BL
773pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
774{
775 pgd_t *pgd;
c2febafc 776 p4d_t *p4d;
6219049a
BL
777 pud_t *pud;
778 pmd_t *pmd = NULL;
779
780 pgd = pgd_offset(mm, address);
781 if (!pgd_present(*pgd))
782 goto out;
783
c2febafc
KS
784 p4d = p4d_offset(pgd, address);
785 if (!p4d_present(*p4d))
786 goto out;
787
788 pud = pud_offset(p4d, address);
6219049a
BL
789 if (!pud_present(*pud))
790 goto out;
791
792 pmd = pmd_offset(pud, address);
6219049a
BL
793out:
794 return pmd;
795}
796
b3ac0413 797struct folio_referenced_arg {
8749cfea
VD
798 int mapcount;
799 int referenced;
800 unsigned long vm_flags;
801 struct mem_cgroup *memcg;
802};
803/*
b3ac0413 804 * arg: folio_referenced_arg will be passed
8749cfea 805 */
2f031c6f
MWO
806static bool folio_referenced_one(struct folio *folio,
807 struct vm_area_struct *vma, unsigned long address, void *arg)
8749cfea 808{
b3ac0413
MWO
809 struct folio_referenced_arg *pra = arg;
810 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
8749cfea
VD
811 int referenced = 0;
812
8eaedede
KS
813 while (page_vma_mapped_walk(&pvmw)) {
814 address = pvmw.address;
b20ce5e0 815
47d4f3ee 816 if ((vma->vm_flags & VM_LOCKED) &&
b3ac0413 817 (!folio_test_large(folio) || !pvmw.pte)) {
47d4f3ee 818 /* Restore the mlock which got missed */
b3ac0413 819 mlock_vma_folio(folio, vma, !pvmw.pte);
8eaedede
KS
820 page_vma_mapped_walk_done(&pvmw);
821 pra->vm_flags |= VM_LOCKED;
e4b82222 822 return false; /* To break the loop */
8eaedede 823 }
71e3aac0 824
8eaedede 825 if (pvmw.pte) {
018ee47f
YZ
826 if (lru_gen_enabled() && pte_young(*pvmw.pte) &&
827 !(vma->vm_flags & (VM_SEQ_READ | VM_RAND_READ))) {
828 lru_gen_look_around(&pvmw);
829 referenced++;
830 }
831
8eaedede
KS
832 if (ptep_clear_flush_young_notify(vma, address,
833 pvmw.pte)) {
834 /*
835 * Don't treat a reference through
836 * a sequentially read mapping as such.
b3ac0413 837 * If the folio has been used in another mapping,
8eaedede
KS
838 * we will catch it; if this other mapping is
839 * already gone, the unmap path will have set
b3ac0413 840 * the referenced flag or activated the folio.
8eaedede
KS
841 */
842 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
843 referenced++;
844 }
845 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
846 if (pmdp_clear_flush_young_notify(vma, address,
847 pvmw.pmd))
8749cfea 848 referenced++;
8eaedede 849 } else {
b3ac0413 850 /* unexpected pmd-mapped folio? */
8eaedede 851 WARN_ON_ONCE(1);
8749cfea 852 }
8eaedede
KS
853
854 pra->mapcount--;
b20ce5e0 855 }
b20ce5e0 856
33c3fc71 857 if (referenced)
b3ac0413
MWO
858 folio_clear_idle(folio);
859 if (folio_test_clear_young(folio))
33c3fc71
VD
860 referenced++;
861
9f32624b
JK
862 if (referenced) {
863 pra->referenced++;
47d4f3ee 864 pra->vm_flags |= vma->vm_flags & ~VM_LOCKED;
1da177e4 865 }
34bbd704 866
9f32624b 867 if (!pra->mapcount)
e4b82222 868 return false; /* To break the loop */
9f32624b 869
e4b82222 870 return true;
1da177e4
LT
871}
872
b3ac0413 873static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg)
1da177e4 874{
b3ac0413 875 struct folio_referenced_arg *pra = arg;
9f32624b 876 struct mem_cgroup *memcg = pra->memcg;
1da177e4 877
9f32624b
JK
878 if (!mm_match_cgroup(vma->vm_mm, memcg))
879 return true;
1da177e4 880
9f32624b 881 return false;
1da177e4
LT
882}
883
884/**
b3ac0413
MWO
885 * folio_referenced() - Test if the folio was referenced.
886 * @folio: The folio to test.
887 * @is_locked: Caller holds lock on the folio.
72835c86 888 * @memcg: target memory cgroup
b3ac0413 889 * @vm_flags: A combination of all the vma->vm_flags which referenced the folio.
1da177e4 890 *
b3ac0413
MWO
891 * Quick test_and_clear_referenced for all mappings of a folio,
892 *
6d4675e6
MK
893 * Return: The number of mappings which referenced the folio. Return -1 if
894 * the function bailed out due to rmap lock contention.
1da177e4 895 */
b3ac0413
MWO
896int folio_referenced(struct folio *folio, int is_locked,
897 struct mem_cgroup *memcg, unsigned long *vm_flags)
1da177e4 898{
5ad64688 899 int we_locked = 0;
b3ac0413
MWO
900 struct folio_referenced_arg pra = {
901 .mapcount = folio_mapcount(folio),
9f32624b
JK
902 .memcg = memcg,
903 };
904 struct rmap_walk_control rwc = {
b3ac0413 905 .rmap_one = folio_referenced_one,
9f32624b 906 .arg = (void *)&pra,
2f031c6f 907 .anon_lock = folio_lock_anon_vma_read,
6d4675e6 908 .try_lock = true,
9f32624b 909 };
1da177e4 910
6fe6b7e3 911 *vm_flags = 0;
059d8442 912 if (!pra.mapcount)
9f32624b
JK
913 return 0;
914
b3ac0413 915 if (!folio_raw_mapping(folio))
9f32624b
JK
916 return 0;
917
b3ac0413
MWO
918 if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) {
919 we_locked = folio_trylock(folio);
9f32624b
JK
920 if (!we_locked)
921 return 1;
1da177e4 922 }
9f32624b
JK
923
924 /*
925 * If we are reclaiming on behalf of a cgroup, skip
926 * counting on behalf of references from different
927 * cgroups
928 */
929 if (memcg) {
b3ac0413 930 rwc.invalid_vma = invalid_folio_referenced_vma;
9f32624b
JK
931 }
932
2f031c6f 933 rmap_walk(folio, &rwc);
9f32624b
JK
934 *vm_flags = pra.vm_flags;
935
936 if (we_locked)
b3ac0413 937 folio_unlock(folio);
9f32624b 938
6d4675e6 939 return rwc.contended ? -1 : pra.referenced;
1da177e4
LT
940}
941
6a8e0596 942static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw)
d08b3851 943{
6a8e0596
MS
944 int cleaned = 0;
945 struct vm_area_struct *vma = pvmw->vma;
ac46d4f3 946 struct mmu_notifier_range range;
6a8e0596 947 unsigned long address = pvmw->address;
d08b3851 948
369ea824
JG
949 /*
950 * We have to assume the worse case ie pmd for invalidation. Note that
e83c09a2 951 * the folio can not be freed from this function.
369ea824 952 */
7269f999
JG
953 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
954 0, vma, vma->vm_mm, address,
6a8e0596 955 vma_address_end(pvmw));
ac46d4f3 956 mmu_notifier_invalidate_range_start(&range);
369ea824 957
6a8e0596 958 while (page_vma_mapped_walk(pvmw)) {
f27176cf 959 int ret = 0;
369ea824 960
6a8e0596
MS
961 address = pvmw->address;
962 if (pvmw->pte) {
f27176cf 963 pte_t entry;
6a8e0596 964 pte_t *pte = pvmw->pte;
f27176cf
KS
965
966 if (!pte_dirty(*pte) && !pte_write(*pte))
967 continue;
968
785373b4
LT
969 flush_cache_page(vma, address, pte_pfn(*pte));
970 entry = ptep_clear_flush(vma, address, pte);
f27176cf
KS
971 entry = pte_wrprotect(entry);
972 entry = pte_mkclean(entry);
785373b4 973 set_pte_at(vma->vm_mm, address, pte, entry);
f27176cf
KS
974 ret = 1;
975 } else {
396bcc52 976#ifdef CONFIG_TRANSPARENT_HUGEPAGE
6a8e0596 977 pmd_t *pmd = pvmw->pmd;
f27176cf
KS
978 pmd_t entry;
979
980 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
981 continue;
982
7f9c9b60
MS
983 flush_cache_range(vma, address,
984 address + HPAGE_PMD_SIZE);
024eee0e 985 entry = pmdp_invalidate(vma, address, pmd);
f27176cf
KS
986 entry = pmd_wrprotect(entry);
987 entry = pmd_mkclean(entry);
785373b4 988 set_pmd_at(vma->vm_mm, address, pmd, entry);
f27176cf
KS
989 ret = 1;
990#else
e83c09a2 991 /* unexpected pmd-mapped folio? */
f27176cf
KS
992 WARN_ON_ONCE(1);
993#endif
994 }
d08b3851 995
0f10851e
JG
996 /*
997 * No need to call mmu_notifier_invalidate_range() as we are
998 * downgrading page table protection not changing it to point
999 * to a new page.
1000 *
ee65728e 1001 * See Documentation/mm/mmu_notifier.rst
0f10851e
JG
1002 */
1003 if (ret)
6a8e0596 1004 cleaned++;
c2fda5fe 1005 }
d08b3851 1006
ac46d4f3 1007 mmu_notifier_invalidate_range_end(&range);
369ea824 1008
6a8e0596
MS
1009 return cleaned;
1010}
1011
1012static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma,
1013 unsigned long address, void *arg)
1014{
1015 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC);
1016 int *cleaned = arg;
1017
1018 *cleaned += page_vma_mkclean_one(&pvmw);
1019
e4b82222 1020 return true;
d08b3851
PZ
1021}
1022
9853a407 1023static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
d08b3851 1024{
9853a407 1025 if (vma->vm_flags & VM_SHARED)
871beb8c 1026 return false;
d08b3851 1027
871beb8c 1028 return true;
d08b3851
PZ
1029}
1030
d9c08e22 1031int folio_mkclean(struct folio *folio)
d08b3851 1032{
9853a407
JK
1033 int cleaned = 0;
1034 struct address_space *mapping;
1035 struct rmap_walk_control rwc = {
1036 .arg = (void *)&cleaned,
1037 .rmap_one = page_mkclean_one,
1038 .invalid_vma = invalid_mkclean_vma,
1039 };
d08b3851 1040
d9c08e22 1041 BUG_ON(!folio_test_locked(folio));
d08b3851 1042
d9c08e22 1043 if (!folio_mapped(folio))
9853a407
JK
1044 return 0;
1045
d9c08e22 1046 mapping = folio_mapping(folio);
9853a407
JK
1047 if (!mapping)
1048 return 0;
1049
2f031c6f 1050 rmap_walk(folio, &rwc);
d08b3851 1051
9853a407 1052 return cleaned;
d08b3851 1053}
d9c08e22 1054EXPORT_SYMBOL_GPL(folio_mkclean);
d08b3851 1055
6a8e0596
MS
1056/**
1057 * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of
1058 * [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff)
1059 * within the @vma of shared mappings. And since clean PTEs
1060 * should also be readonly, write protects them too.
1061 * @pfn: start pfn.
1062 * @nr_pages: number of physically contiguous pages srarting with @pfn.
1063 * @pgoff: page offset that the @pfn mapped with.
1064 * @vma: vma that @pfn mapped within.
1065 *
1066 * Returns the number of cleaned PTEs (including PMDs).
1067 */
1068int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff,
1069 struct vm_area_struct *vma)
1070{
1071 struct page_vma_mapped_walk pvmw = {
1072 .pfn = pfn,
1073 .nr_pages = nr_pages,
1074 .pgoff = pgoff,
1075 .vma = vma,
1076 .flags = PVMW_SYNC,
1077 };
1078
1079 if (invalid_mkclean_vma(vma, NULL))
1080 return 0;
1081
1082 pvmw.address = vma_pgoff_address(pgoff, nr_pages, vma);
1083 VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma);
1084
1085 return page_vma_mkclean_one(&pvmw);
1086}
1087
cb67f428
HD
1088/*
1089 * When mapping a THP's first pmd, or unmapping its last pmd, if that THP
1090 * also has pte mappings, then those must be discounted: in order to maintain
1091 * NR_ANON_MAPPED and NR_FILE_MAPPED statistics exactly, without any drift,
1092 * and to decide when an anon THP should be put on the deferred split queue.
1093 */
1094static int nr_subpages_unmapped(struct page *head, int nr_subpages)
1095{
1096 int nr = nr_subpages;
1097 int i;
1098
1099 /* Discount those subpages mapped by pte */
1100 for (i = 0; i < nr_subpages; i++)
1101 if (atomic_read(&head[i]._mapcount) >= 0)
1102 nr--;
1103 return nr;
1104}
1105
c44b6743
RR
1106/**
1107 * page_move_anon_rmap - move a page to our anon_vma
1108 * @page: the page to move to our anon_vma
1109 * @vma: the vma the page belongs to
c44b6743
RR
1110 *
1111 * When a page belongs exclusively to one process after a COW event,
1112 * that page can be moved into the anon_vma that belongs to just that
1113 * process, so the rmap code will not search the parent or sibling
1114 * processes.
1115 */
5a49973d 1116void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
c44b6743 1117{
595af4c9
MWO
1118 void *anon_vma = vma->anon_vma;
1119 struct folio *folio = page_folio(page);
5a49973d 1120
595af4c9 1121 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
81d1b09c 1122 VM_BUG_ON_VMA(!anon_vma, vma);
c44b6743 1123
595af4c9 1124 anon_vma += PAGE_MAPPING_ANON;
414e2fb8
VD
1125 /*
1126 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
b3ac0413
MWO
1127 * simultaneously, so a concurrent reader (eg folio_referenced()'s
1128 * folio_test_anon()) will not see one without the other.
414e2fb8 1129 */
595af4c9
MWO
1130 WRITE_ONCE(folio->mapping, anon_vma);
1131 SetPageAnonExclusive(page);
c44b6743
RR
1132}
1133
9617d95e 1134/**
4e1c1975 1135 * __page_set_anon_rmap - set up new anonymous rmap
451b9514 1136 * @page: Page or Hugepage to add to rmap
4e1c1975
AK
1137 * @vma: VM area to add page to.
1138 * @address: User virtual address of the mapping
e8a03feb 1139 * @exclusive: the page is exclusively owned by the current process
9617d95e
NP
1140 */
1141static void __page_set_anon_rmap(struct page *page,
e8a03feb 1142 struct vm_area_struct *vma, unsigned long address, int exclusive)
9617d95e 1143{
e8a03feb 1144 struct anon_vma *anon_vma = vma->anon_vma;
ea90002b 1145
e8a03feb 1146 BUG_ON(!anon_vma);
ea90002b 1147
4e1c1975 1148 if (PageAnon(page))
6c287605 1149 goto out;
4e1c1975 1150
ea90002b 1151 /*
e8a03feb
RR
1152 * If the page isn't exclusively mapped into this vma,
1153 * we must use the _oldest_ possible anon_vma for the
1154 * page mapping!
ea90002b 1155 */
4e1c1975 1156 if (!exclusive)
288468c3 1157 anon_vma = anon_vma->root;
9617d95e 1158
16f5e707
AS
1159 /*
1160 * page_idle does a lockless/optimistic rmap scan on page->mapping.
1161 * Make sure the compiler doesn't split the stores of anon_vma and
1162 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1163 * could mistake the mapping for a struct address_space and crash.
1164 */
9617d95e 1165 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
16f5e707 1166 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
9617d95e 1167 page->index = linear_page_index(vma, address);
6c287605
DH
1168out:
1169 if (exclusive)
1170 SetPageAnonExclusive(page);
9617d95e
NP
1171}
1172
c97a9e10 1173/**
43d8eac4 1174 * __page_check_anon_rmap - sanity check anonymous rmap addition
c97a9e10
NP
1175 * @page: the page to add the mapping to
1176 * @vma: the vm area in which the mapping is added
1177 * @address: the user virtual address mapped
1178 */
1179static void __page_check_anon_rmap(struct page *page,
1180 struct vm_area_struct *vma, unsigned long address)
1181{
e05b3453 1182 struct folio *folio = page_folio(page);
c97a9e10
NP
1183 /*
1184 * The page's anon-rmap details (mapping and index) are guaranteed to
1185 * be set up correctly at this point.
1186 *
1187 * We have exclusion against page_add_anon_rmap because the caller
90aaca85 1188 * always holds the page locked.
c97a9e10
NP
1189 *
1190 * We have exclusion against page_add_new_anon_rmap because those pages
1191 * are initially only visible via the pagetables, and the pte is locked
1192 * over the call to page_add_new_anon_rmap.
1193 */
e05b3453
MWO
1194 VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root,
1195 folio);
30c46382
YS
1196 VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1197 page);
c97a9e10
NP
1198}
1199
1da177e4
LT
1200/**
1201 * page_add_anon_rmap - add pte mapping to an anonymous page
1202 * @page: the page to add the mapping to
1203 * @vma: the vm area in which the mapping is added
1204 * @address: the user virtual address mapped
f1e2db12 1205 * @flags: the rmap flags
1da177e4 1206 *
5ad64688 1207 * The caller needs to hold the pte lock, and the page must be locked in
80e14822
HD
1208 * the anon_vma case: to serialize mapping,index checking after setting,
1209 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1210 * (but PageKsm is never downgraded to PageAnon).
1da177e4
LT
1211 */
1212void page_add_anon_rmap(struct page *page,
14f9135d 1213 struct vm_area_struct *vma, unsigned long address, rmap_t flags)
1da177e4 1214{
cb67f428 1215 int nr, nr_pages;
53f9263b
KS
1216 bool compound = flags & RMAP_COMPOUND;
1217 bool first;
1218
be5d0a74
JW
1219 if (unlikely(PageKsm(page)))
1220 lock_page_memcg(page);
1221 else
1222 VM_BUG_ON_PAGE(!PageLocked(page), page);
1223
cb67f428 1224 if (compound && PageTransHuge(page)) {
e9b61f19 1225 atomic_t *mapcount;
53f9263b 1226 VM_BUG_ON_PAGE(!PageLocked(page), page);
e9b61f19
KS
1227 mapcount = compound_mapcount_ptr(page);
1228 first = atomic_inc_and_test(mapcount);
cb67f428
HD
1229
1230 nr = nr_pages = thp_nr_pages(page);
1231 if (first && head_subpages_mapcount(page))
1232 nr = nr_subpages_unmapped(page, nr_pages);
53f9263b 1233 } else {
cb67f428
HD
1234 nr = 1;
1235 if (PageTransCompound(page)) {
1236 struct page *head = compound_head(page);
1237
1238 atomic_inc(subpages_mapcount_ptr(head));
1239 nr = !head_compound_mapcount(head);
1240 }
53f9263b
KS
1241 first = atomic_inc_and_test(&page->_mapcount);
1242 }
cb67f428 1243
6c287605
DH
1244 VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
1245 VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
53f9263b 1246
79134171 1247 if (first) {
65c45377 1248 if (compound)
cb67f428 1249 __mod_lruvec_page_state(page, NR_ANON_THPS, nr_pages);
be5d0a74 1250 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
79134171 1251 }
5ad64688 1252
cea86fe2 1253 if (unlikely(PageKsm(page)))
be5d0a74 1254 unlock_page_memcg(page);
53f9263b 1255
5dbe0af4 1256 /* address might be in next vma when migration races vma_adjust */
cea86fe2 1257 else if (first)
d281ee61 1258 __page_set_anon_rmap(page, vma, address,
14f9135d 1259 !!(flags & RMAP_EXCLUSIVE));
69029cd5 1260 else
c97a9e10 1261 __page_check_anon_rmap(page, vma, address);
cea86fe2
HD
1262
1263 mlock_vma_page(page, vma, compound);
1da177e4
LT
1264}
1265
43d8eac4 1266/**
40f2bbf7 1267 * page_add_new_anon_rmap - add mapping to a new anonymous page
9617d95e
NP
1268 * @page: the page to add the mapping to
1269 * @vma: the vm area in which the mapping is added
1270 * @address: the user virtual address mapped
40f2bbf7
DH
1271 *
1272 * If it's a compound page, it is accounted as a compound page. As the page
1273 * is new, it's assume to get mapped exclusively by a single process.
9617d95e
NP
1274 *
1275 * Same as page_add_anon_rmap but must only be called on *new* pages.
1276 * This means the inc-and-test can be bypassed.
c97a9e10 1277 * Page does not have to be locked.
9617d95e
NP
1278 */
1279void page_add_new_anon_rmap(struct page *page,
40f2bbf7 1280 struct vm_area_struct *vma, unsigned long address)
9617d95e 1281{
40f2bbf7 1282 const bool compound = PageCompound(page);
6c357848 1283 int nr = compound ? thp_nr_pages(page) : 1;
d281ee61 1284
81d1b09c 1285 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
fa9949da 1286 __SetPageSwapBacked(page);
d281ee61
KS
1287 if (compound) {
1288 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
53f9263b
KS
1289 /* increment count (starts at -1) */
1290 atomic_set(compound_mapcount_ptr(page), 0);
69473e5d 1291 __mod_lruvec_page_state(page, NR_ANON_THPS, nr);
53f9263b 1292 } else {
53f9263b
KS
1293 /* increment count (starts at -1) */
1294 atomic_set(&page->_mapcount, 0);
d281ee61 1295 }
be5d0a74 1296 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
e8a03feb 1297 __page_set_anon_rmap(page, vma, address, 1);
9617d95e
NP
1298}
1299
1da177e4
LT
1300/**
1301 * page_add_file_rmap - add pte mapping to a file page
cea86fe2
HD
1302 * @page: the page to add the mapping to
1303 * @vma: the vm area in which the mapping is added
1304 * @compound: charge the page as compound or small page
1da177e4 1305 *
b8072f09 1306 * The caller needs to hold the pte lock.
1da177e4 1307 */
cea86fe2
HD
1308void page_add_file_rmap(struct page *page,
1309 struct vm_area_struct *vma, bool compound)
1da177e4 1310{
cb67f428 1311 int nr = 0;
dd78fedd
KS
1312
1313 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
62cccb8c 1314 lock_page_memcg(page);
dd78fedd 1315 if (compound && PageTransHuge(page)) {
cb67f428 1316 int nr_pages;
a1528e21 1317
dd78fedd
KS
1318 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1319 goto out;
bd55b0c2 1320
cb67f428
HD
1321 nr = nr_pages = thp_nr_pages(page);
1322 if (head_subpages_mapcount(page))
1323 nr = nr_subpages_unmapped(page, nr_pages);
bd55b0c2 1324
99cb0dbd 1325 if (PageSwapBacked(page))
a1528e21
MS
1326 __mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED,
1327 nr_pages);
99cb0dbd 1328 else
380780e7
MS
1329 __mod_lruvec_page_state(page, NR_FILE_PMDMAPPED,
1330 nr_pages);
dd78fedd 1331 } else {
cb67f428
HD
1332 bool pmd_mapped = false;
1333
1334 if (PageTransCompound(page)) {
1335 struct page *head = compound_head(page);
1336
1337 atomic_inc(subpages_mapcount_ptr(head));
1338 pmd_mapped = head_compound_mapcount(head);
9a73f61b 1339 }
cb67f428 1340 if (atomic_inc_and_test(&page->_mapcount) && !pmd_mapped)
5d543f13 1341 nr++;
d69b042f 1342 }
dd78fedd 1343out:
5d543f13
HD
1344 if (nr)
1345 __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
62cccb8c 1346 unlock_page_memcg(page);
cea86fe2
HD
1347
1348 mlock_vma_page(page, vma, compound);
1da177e4
LT
1349}
1350
dd78fedd 1351static void page_remove_file_rmap(struct page *page, bool compound)
8186eb6a 1352{
cb67f428 1353 int nr = 0;
dd78fedd 1354
57dea93a 1355 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
8186eb6a 1356
53f9263b
KS
1357 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1358 if (unlikely(PageHuge(page))) {
1359 /* hugetlb pages are always mapped with pmds */
1360 atomic_dec(compound_mapcount_ptr(page));
be5d0a74 1361 return;
53f9263b 1362 }
8186eb6a 1363
53f9263b 1364 /* page still mapped by someone else? */
dd78fedd 1365 if (compound && PageTransHuge(page)) {
cb67f428 1366 int nr_pages;
a1528e21 1367
dd78fedd 1368 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
cb67f428
HD
1369 return;
1370
1371 nr = nr_pages = thp_nr_pages(page);
1372 if (head_subpages_mapcount(page))
1373 nr = nr_subpages_unmapped(page, nr_pages);
1374
99cb0dbd 1375 if (PageSwapBacked(page))
a1528e21
MS
1376 __mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED,
1377 -nr_pages);
99cb0dbd 1378 else
380780e7
MS
1379 __mod_lruvec_page_state(page, NR_FILE_PMDMAPPED,
1380 -nr_pages);
dd78fedd 1381 } else {
cb67f428
HD
1382 bool pmd_mapped = false;
1383
1384 if (PageTransCompound(page)) {
1385 struct page *head = compound_head(page);
1386
1387 atomic_dec(subpages_mapcount_ptr(head));
1388 pmd_mapped = head_compound_mapcount(head);
1389 }
1390 if (atomic_add_negative(-1, &page->_mapcount) && !pmd_mapped)
5d543f13 1391 nr++;
dd78fedd 1392 }
cb67f428 1393
5d543f13
HD
1394 if (nr)
1395 __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
8186eb6a
JW
1396}
1397
53f9263b
KS
1398static void page_remove_anon_compound_rmap(struct page *page)
1399{
cb67f428 1400 int nr, nr_pages;
53f9263b
KS
1401
1402 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1403 return;
1404
1405 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1406 if (unlikely(PageHuge(page)))
1407 return;
1408
1409 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1410 return;
1411
cb67f428
HD
1412 nr = nr_pages = thp_nr_pages(page);
1413 __mod_lruvec_page_state(page, NR_ANON_THPS, -nr);
53f9263b 1414
cb67f428
HD
1415 if (head_subpages_mapcount(page)) {
1416 nr = nr_subpages_unmapped(page, nr_pages);
f1fe80d4
KS
1417
1418 /*
1419 * Queue the page for deferred split if at least one small
1420 * page of the compound page is unmapped, but at least one
1421 * small page is still mapped.
1422 */
cb67f428 1423 if (nr && nr < nr_pages)
f1fe80d4 1424 deferred_split_huge_page(page);
53f9263b
KS
1425 }
1426
f1fe80d4 1427 if (nr)
be5d0a74 1428 __mod_lruvec_page_state(page, NR_ANON_MAPPED, -nr);
53f9263b
KS
1429}
1430
1da177e4
LT
1431/**
1432 * page_remove_rmap - take down pte mapping from a page
d281ee61 1433 * @page: page to remove mapping from
cea86fe2 1434 * @vma: the vm area from which the mapping is removed
d281ee61 1435 * @compound: uncharge the page as compound or small page
1da177e4 1436 *
b8072f09 1437 * The caller needs to hold the pte lock.
1da177e4 1438 */
cea86fe2
HD
1439void page_remove_rmap(struct page *page,
1440 struct vm_area_struct *vma, bool compound)
1da177e4 1441{
cb67f428
HD
1442 bool pmd_mapped = false;
1443
be5d0a74 1444 lock_page_memcg(page);
89c06bd5 1445
be5d0a74
JW
1446 if (!PageAnon(page)) {
1447 page_remove_file_rmap(page, compound);
1448 goto out;
1449 }
1450
1451 if (compound) {
1452 page_remove_anon_compound_rmap(page);
1453 goto out;
1454 }
53f9263b 1455
cb67f428
HD
1456 if (PageTransCompound(page)) {
1457 struct page *head = compound_head(page);
1458
1459 atomic_dec(subpages_mapcount_ptr(head));
1460 pmd_mapped = head_compound_mapcount(head);
1461 }
1462
b904dcfe 1463 /* page still mapped by someone else? */
cb67f428 1464 if (!atomic_add_negative(-1, &page->_mapcount) || pmd_mapped)
be5d0a74 1465 goto out;
8186eb6a 1466
be5d0a74 1467 __dec_lruvec_page_state(page, NR_ANON_MAPPED);
8186eb6a 1468
9a982250
KS
1469 if (PageTransCompound(page))
1470 deferred_split_huge_page(compound_head(page));
1471
b904dcfe
KM
1472 /*
1473 * It would be tidy to reset the PageAnon mapping here,
1474 * but that might overwrite a racing page_add_anon_rmap
1475 * which increments mapcount after us but sets mapping
2d4894b5 1476 * before us: so leave the reset to free_unref_page,
b904dcfe
KM
1477 * and remember that it's only reliable while mapped.
1478 * Leaving it set also helps swapoff to reinstate ptes
1479 * faster for those pages still in swapcache.
1480 */
be5d0a74
JW
1481out:
1482 unlock_page_memcg(page);
cea86fe2
HD
1483
1484 munlock_vma_page(page, vma, compound);
1da177e4
LT
1485}
1486
1487/*
52629506 1488 * @arg: enum ttu_flags will be passed to this argument
1da177e4 1489 */
2f031c6f 1490static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
52629506 1491 unsigned long address, void *arg)
1da177e4
LT
1492{
1493 struct mm_struct *mm = vma->vm_mm;
869f7ee6 1494 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1da177e4 1495 pte_t pteval;
c7ab0d2f 1496 struct page *subpage;
6c287605 1497 bool anon_exclusive, ret = true;
ac46d4f3 1498 struct mmu_notifier_range range;
4708f318 1499 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1da177e4 1500
732ed558
HD
1501 /*
1502 * When racing against e.g. zap_pte_range() on another cpu,
1503 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1fb08ac6 1504 * try_to_unmap() may return before page_mapped() has become false,
732ed558
HD
1505 * if page table locking is skipped: use TTU_SYNC to wait for that.
1506 */
1507 if (flags & TTU_SYNC)
1508 pvmw.flags = PVMW_SYNC;
1509
a98a2f0c 1510 if (flags & TTU_SPLIT_HUGE_PMD)
af28a988 1511 split_huge_pmd_address(vma, address, false, folio);
fec89c10 1512
369ea824 1513 /*
017b1660
MK
1514 * For THP, we have to assume the worse case ie pmd for invalidation.
1515 * For hugetlb, it could be much worse if we need to do pud
1516 * invalidation in the case of pmd sharing.
1517 *
869f7ee6
MWO
1518 * Note that the folio can not be freed in this function as call of
1519 * try_to_unmap() must hold a reference on the folio.
369ea824 1520 */
2aff7a47 1521 range.end = vma_address_end(&pvmw);
7269f999 1522 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
494334e4 1523 address, range.end);
869f7ee6 1524 if (folio_test_hugetlb(folio)) {
017b1660
MK
1525 /*
1526 * If sharing is possible, start and end will be adjusted
1527 * accordingly.
1528 */
ac46d4f3
JG
1529 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1530 &range.end);
017b1660 1531 }
ac46d4f3 1532 mmu_notifier_invalidate_range_start(&range);
369ea824 1533
c7ab0d2f 1534 while (page_vma_mapped_walk(&pvmw)) {
cea86fe2 1535 /* Unexpected PMD-mapped THP? */
869f7ee6 1536 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
cea86fe2 1537
c7ab0d2f 1538 /*
869f7ee6 1539 * If the folio is in an mlock()d vma, we must not swap it out.
c7ab0d2f 1540 */
efdb6720
HD
1541 if (!(flags & TTU_IGNORE_MLOCK) &&
1542 (vma->vm_flags & VM_LOCKED)) {
cea86fe2 1543 /* Restore the mlock which got missed */
869f7ee6 1544 mlock_vma_folio(folio, vma, false);
efdb6720
HD
1545 page_vma_mapped_walk_done(&pvmw);
1546 ret = false;
1547 break;
b87537d9 1548 }
c7ab0d2f 1549
869f7ee6
MWO
1550 subpage = folio_page(folio,
1551 pte_pfn(*pvmw.pte) - folio_pfn(folio));
785373b4 1552 address = pvmw.address;
6c287605
DH
1553 anon_exclusive = folio_test_anon(folio) &&
1554 PageAnonExclusive(subpage);
785373b4 1555
dfc7ab57 1556 if (folio_test_hugetlb(folio)) {
0506c31d
BW
1557 bool anon = folio_test_anon(folio);
1558
a00a8759
BW
1559 /*
1560 * The try_to_unmap() is only passed a hugetlb page
1561 * in the case where the hugetlb page is poisoned.
1562 */
1563 VM_BUG_ON_PAGE(!PageHWPoison(subpage), subpage);
54205e9c
BW
1564 /*
1565 * huge_pmd_unshare may unmap an entire PMD page.
1566 * There is no way of knowing exactly which PMDs may
1567 * be cached for this mm, so we must flush them all.
1568 * start/end were already adjusted above to cover this
1569 * range.
1570 */
1571 flush_cache_range(vma, range.start, range.end);
1572
0506c31d
BW
1573 /*
1574 * To call huge_pmd_unshare, i_mmap_rwsem must be
1575 * held in write mode. Caller needs to explicitly
1576 * do this outside rmap routines.
40549ba8
MK
1577 *
1578 * We also must hold hugetlb vma_lock in write mode.
1579 * Lock order dictates acquiring vma_lock BEFORE
1580 * i_mmap_rwsem. We can only try lock here and fail
1581 * if unsuccessful.
0506c31d 1582 */
40549ba8
MK
1583 if (!anon) {
1584 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1585 if (!hugetlb_vma_trylock_write(vma)) {
1586 page_vma_mapped_walk_done(&pvmw);
1587 ret = false;
1588 break;
1589 }
1590 if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1591 hugetlb_vma_unlock_write(vma);
1592 flush_tlb_range(vma,
1593 range.start, range.end);
1594 mmu_notifier_invalidate_range(mm,
1595 range.start, range.end);
1596 /*
1597 * The ref count of the PMD page was
1598 * dropped which is part of the way map
1599 * counting is done for shared PMDs.
1600 * Return 'true' here. When there is
1601 * no other sharing, huge_pmd_unshare
1602 * returns false and we will unmap the
1603 * actual page and drop map count
1604 * to zero.
1605 */
1606 page_vma_mapped_walk_done(&pvmw);
1607 break;
1608 }
1609 hugetlb_vma_unlock_write(vma);
017b1660 1610 }
a00a8759 1611 pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
54205e9c
BW
1612 } else {
1613 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
088b8aa5
DH
1614 /* Nuke the page table entry. */
1615 if (should_defer_flush(mm, flags)) {
a00a8759
BW
1616 /*
1617 * We clear the PTE but do not flush so potentially
1618 * a remote CPU could still be writing to the folio.
1619 * If the entry was previously clean then the
1620 * architecture must guarantee that a clear->dirty
1621 * transition on a cached TLB entry is written through
1622 * and traps if the PTE is unmapped.
1623 */
1624 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
c7ab0d2f 1625
a00a8759
BW
1626 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1627 } else {
1628 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1629 }
c7ab0d2f 1630 }
72b252ae 1631
999dad82
PX
1632 /*
1633 * Now the pte is cleared. If this pte was uffd-wp armed,
1634 * we may want to replace a none pte with a marker pte if
1635 * it's file-backed, so we don't lose the tracking info.
1636 */
1637 pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval);
1638
869f7ee6 1639 /* Set the dirty flag on the folio now the pte is gone. */
c7ab0d2f 1640 if (pte_dirty(pteval))
869f7ee6 1641 folio_mark_dirty(folio);
1da177e4 1642
c7ab0d2f
KS
1643 /* Update high watermark before we lower rss */
1644 update_hiwater_rss(mm);
1da177e4 1645
da358d5c 1646 if (PageHWPoison(subpage) && !(flags & TTU_IGNORE_HWPOISON)) {
5fd27b8e 1647 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
869f7ee6
MWO
1648 if (folio_test_hugetlb(folio)) {
1649 hugetlb_count_sub(folio_nr_pages(folio), mm);
18f39629 1650 set_huge_pte_at(mm, address, pvmw.pte, pteval);
c7ab0d2f 1651 } else {
869f7ee6 1652 dec_mm_counter(mm, mm_counter(&folio->page));
785373b4 1653 set_pte_at(mm, address, pvmw.pte, pteval);
c7ab0d2f 1654 }
365e9c87 1655
bce73e48 1656 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
c7ab0d2f
KS
1657 /*
1658 * The guest indicated that the page content is of no
1659 * interest anymore. Simply discard the pte, vmscan
1660 * will take care of the rest.
bce73e48
CB
1661 * A future reference will then fault in a new zero
1662 * page. When userfaultfd is active, we must not drop
1663 * this page though, as its main user (postcopy
1664 * migration) will not expect userfaults on already
1665 * copied pages.
c7ab0d2f 1666 */
869f7ee6 1667 dec_mm_counter(mm, mm_counter(&folio->page));
0f10851e
JG
1668 /* We have to invalidate as we cleared the pte */
1669 mmu_notifier_invalidate_range(mm, address,
1670 address + PAGE_SIZE);
869f7ee6 1671 } else if (folio_test_anon(folio)) {
c7ab0d2f
KS
1672 swp_entry_t entry = { .val = page_private(subpage) };
1673 pte_t swp_pte;
1674 /*
1675 * Store the swap location in the pte.
1676 * See handle_pte_fault() ...
1677 */
869f7ee6
MWO
1678 if (unlikely(folio_test_swapbacked(folio) !=
1679 folio_test_swapcache(folio))) {
eb94a878 1680 WARN_ON_ONCE(1);
83612a94 1681 ret = false;
369ea824 1682 /* We have to invalidate as we cleared the pte */
0f10851e
JG
1683 mmu_notifier_invalidate_range(mm, address,
1684 address + PAGE_SIZE);
eb94a878
MK
1685 page_vma_mapped_walk_done(&pvmw);
1686 break;
1687 }
c7ab0d2f 1688
802a3a92 1689 /* MADV_FREE page check */
869f7ee6 1690 if (!folio_test_swapbacked(folio)) {
6c8e2a25
MFO
1691 int ref_count, map_count;
1692
1693 /*
1694 * Synchronize with gup_pte_range():
1695 * - clear PTE; barrier; read refcount
1696 * - inc refcount; barrier; read PTE
1697 */
1698 smp_mb();
1699
1700 ref_count = folio_ref_count(folio);
1701 map_count = folio_mapcount(folio);
1702
1703 /*
1704 * Order reads for page refcount and dirty flag
1705 * (see comments in __remove_mapping()).
1706 */
1707 smp_rmb();
1708
1709 /*
1710 * The only page refs must be one from isolation
1711 * plus the rmap(s) (dropped by discard:).
1712 */
1713 if (ref_count == 1 + map_count &&
1714 !folio_test_dirty(folio)) {
0f10851e
JG
1715 /* Invalidate as we cleared the pte */
1716 mmu_notifier_invalidate_range(mm,
1717 address, address + PAGE_SIZE);
802a3a92
SL
1718 dec_mm_counter(mm, MM_ANONPAGES);
1719 goto discard;
1720 }
1721
1722 /*
869f7ee6 1723 * If the folio was redirtied, it cannot be
802a3a92
SL
1724 * discarded. Remap the page to page table.
1725 */
785373b4 1726 set_pte_at(mm, address, pvmw.pte, pteval);
869f7ee6 1727 folio_set_swapbacked(folio);
e4b82222 1728 ret = false;
802a3a92
SL
1729 page_vma_mapped_walk_done(&pvmw);
1730 break;
c7ab0d2f 1731 }
854e9ed0 1732
c7ab0d2f 1733 if (swap_duplicate(entry) < 0) {
785373b4 1734 set_pte_at(mm, address, pvmw.pte, pteval);
e4b82222 1735 ret = false;
c7ab0d2f
KS
1736 page_vma_mapped_walk_done(&pvmw);
1737 break;
1738 }
ca827d55 1739 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
322842ea 1740 swap_free(entry);
ca827d55
KA
1741 set_pte_at(mm, address, pvmw.pte, pteval);
1742 ret = false;
1743 page_vma_mapped_walk_done(&pvmw);
1744 break;
1745 }
088b8aa5
DH
1746
1747 /* See page_try_share_anon_rmap(): clear PTE first. */
6c287605
DH
1748 if (anon_exclusive &&
1749 page_try_share_anon_rmap(subpage)) {
1750 swap_free(entry);
1751 set_pte_at(mm, address, pvmw.pte, pteval);
1752 ret = false;
1753 page_vma_mapped_walk_done(&pvmw);
1754 break;
1755 }
1756 /*
1493a191
DH
1757 * Note: We *don't* remember if the page was mapped
1758 * exclusively in the swap pte if the architecture
1759 * doesn't support __HAVE_ARCH_PTE_SWP_EXCLUSIVE. In
1760 * that case, swapin code has to re-determine that
1761 * manually and might detect the page as possibly
1762 * shared, for example, if there are other references on
1763 * the page or if the page is under writeback. We made
1764 * sure that there are no GUP pins on the page that
1765 * would rely on it, so for GUP pins this is fine.
6c287605 1766 */
c7ab0d2f
KS
1767 if (list_empty(&mm->mmlist)) {
1768 spin_lock(&mmlist_lock);
1769 if (list_empty(&mm->mmlist))
1770 list_add(&mm->mmlist, &init_mm.mmlist);
1771 spin_unlock(&mmlist_lock);
1772 }
854e9ed0 1773 dec_mm_counter(mm, MM_ANONPAGES);
c7ab0d2f
KS
1774 inc_mm_counter(mm, MM_SWAPENTS);
1775 swp_pte = swp_entry_to_pte(entry);
1493a191
DH
1776 if (anon_exclusive)
1777 swp_pte = pte_swp_mkexclusive(swp_pte);
c7ab0d2f
KS
1778 if (pte_soft_dirty(pteval))
1779 swp_pte = pte_swp_mksoft_dirty(swp_pte);
f45ec5ff
PX
1780 if (pte_uffd_wp(pteval))
1781 swp_pte = pte_swp_mkuffd_wp(swp_pte);
785373b4 1782 set_pte_at(mm, address, pvmw.pte, swp_pte);
0f10851e
JG
1783 /* Invalidate as we cleared the pte */
1784 mmu_notifier_invalidate_range(mm, address,
1785 address + PAGE_SIZE);
1786 } else {
1787 /*
869f7ee6
MWO
1788 * This is a locked file-backed folio,
1789 * so it cannot be removed from the page
1790 * cache and replaced by a new folio before
1791 * mmu_notifier_invalidate_range_end, so no
1792 * concurrent thread might update its page table
1793 * to point at a new folio while a device is
1794 * still using this folio.
0f10851e 1795 *
ee65728e 1796 * See Documentation/mm/mmu_notifier.rst
0f10851e 1797 */
869f7ee6 1798 dec_mm_counter(mm, mm_counter_file(&folio->page));
0f10851e 1799 }
854e9ed0 1800discard:
0f10851e
JG
1801 /*
1802 * No need to call mmu_notifier_invalidate_range() it has be
1803 * done above for all cases requiring it to happen under page
1804 * table lock before mmu_notifier_invalidate_range_end()
1805 *
ee65728e 1806 * See Documentation/mm/mmu_notifier.rst
0f10851e 1807 */
869f7ee6 1808 page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
b7435507 1809 if (vma->vm_flags & VM_LOCKED)
adb11e78 1810 mlock_page_drain_local();
869f7ee6 1811 folio_put(folio);
c7ab0d2f 1812 }
369ea824 1813
ac46d4f3 1814 mmu_notifier_invalidate_range_end(&range);
369ea824 1815
caed0f48 1816 return ret;
1da177e4
LT
1817}
1818
52629506
JK
1819static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1820{
222100ee 1821 return vma_is_temporary_stack(vma);
52629506
JK
1822}
1823
f3ad032c 1824static int folio_not_mapped(struct folio *folio)
52629506 1825{
2f031c6f 1826 return !folio_mapped(folio);
2a52bcbc 1827}
52629506 1828
1da177e4 1829/**
869f7ee6
MWO
1830 * try_to_unmap - Try to remove all page table mappings to a folio.
1831 * @folio: The folio to unmap.
14fa31b8 1832 * @flags: action and flags
1da177e4
LT
1833 *
1834 * Tries to remove all the page table entries which are mapping this
869f7ee6
MWO
1835 * folio. It is the caller's responsibility to check if the folio is
1836 * still mapped if needed (use TTU_SYNC to prevent accounting races).
1da177e4 1837 *
869f7ee6 1838 * Context: Caller must hold the folio lock.
1da177e4 1839 */
869f7ee6 1840void try_to_unmap(struct folio *folio, enum ttu_flags flags)
1da177e4 1841{
52629506
JK
1842 struct rmap_walk_control rwc = {
1843 .rmap_one = try_to_unmap_one,
802a3a92 1844 .arg = (void *)flags,
f3ad032c 1845 .done = folio_not_mapped,
2f031c6f 1846 .anon_lock = folio_lock_anon_vma_read,
52629506 1847 };
1da177e4 1848
a98a2f0c 1849 if (flags & TTU_RMAP_LOCKED)
2f031c6f 1850 rmap_walk_locked(folio, &rwc);
a98a2f0c 1851 else
2f031c6f 1852 rmap_walk(folio, &rwc);
a98a2f0c
AP
1853}
1854
1855/*
1856 * @arg: enum ttu_flags will be passed to this argument.
1857 *
1858 * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
64b586d1 1859 * containing migration entries.
a98a2f0c 1860 */
2f031c6f 1861static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
a98a2f0c
AP
1862 unsigned long address, void *arg)
1863{
1864 struct mm_struct *mm = vma->vm_mm;
4b8554c5 1865 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
a98a2f0c
AP
1866 pte_t pteval;
1867 struct page *subpage;
6c287605 1868 bool anon_exclusive, ret = true;
a98a2f0c
AP
1869 struct mmu_notifier_range range;
1870 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1871
a98a2f0c
AP
1872 /*
1873 * When racing against e.g. zap_pte_range() on another cpu,
1874 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1875 * try_to_migrate() may return before page_mapped() has become false,
1876 * if page table locking is skipped: use TTU_SYNC to wait for that.
1877 */
1878 if (flags & TTU_SYNC)
1879 pvmw.flags = PVMW_SYNC;
1880
1881 /*
1882 * unmap_page() in mm/huge_memory.c is the only user of migration with
1883 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
1884 */
1885 if (flags & TTU_SPLIT_HUGE_PMD)
af28a988 1886 split_huge_pmd_address(vma, address, true, folio);
a98a2f0c
AP
1887
1888 /*
1889 * For THP, we have to assume the worse case ie pmd for invalidation.
1890 * For hugetlb, it could be much worse if we need to do pud
1891 * invalidation in the case of pmd sharing.
1892 *
1893 * Note that the page can not be free in this function as call of
1894 * try_to_unmap() must hold a reference on the page.
1895 */
2aff7a47 1896 range.end = vma_address_end(&pvmw);
a98a2f0c
AP
1897 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1898 address, range.end);
4b8554c5 1899 if (folio_test_hugetlb(folio)) {
a98a2f0c
AP
1900 /*
1901 * If sharing is possible, start and end will be adjusted
1902 * accordingly.
1903 */
1904 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1905 &range.end);
1906 }
1907 mmu_notifier_invalidate_range_start(&range);
1908
1909 while (page_vma_mapped_walk(&pvmw)) {
1910#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1911 /* PMD-mapped THP migration entry */
1912 if (!pvmw.pte) {
4b8554c5
MWO
1913 subpage = folio_page(folio,
1914 pmd_pfn(*pvmw.pmd) - folio_pfn(folio));
1915 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
1916 !folio_test_pmd_mappable(folio), folio);
a98a2f0c 1917
7f5abe60
DH
1918 if (set_pmd_migration_entry(&pvmw, subpage)) {
1919 ret = false;
1920 page_vma_mapped_walk_done(&pvmw);
1921 break;
1922 }
a98a2f0c
AP
1923 continue;
1924 }
1925#endif
1926
1927 /* Unexpected PMD-mapped THP? */
4b8554c5 1928 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
a98a2f0c 1929
1118234e
DH
1930 if (folio_is_zone_device(folio)) {
1931 /*
1932 * Our PTE is a non-present device exclusive entry and
1933 * calculating the subpage as for the common case would
1934 * result in an invalid pointer.
1935 *
1936 * Since only PAGE_SIZE pages can currently be
1937 * migrated, just set it to page. This will need to be
1938 * changed when hugepage migrations to device private
1939 * memory are supported.
1940 */
1941 VM_BUG_ON_FOLIO(folio_nr_pages(folio) > 1, folio);
1942 subpage = &folio->page;
1943 } else {
1944 subpage = folio_page(folio,
1945 pte_pfn(*pvmw.pte) - folio_pfn(folio));
1946 }
a98a2f0c 1947 address = pvmw.address;
6c287605
DH
1948 anon_exclusive = folio_test_anon(folio) &&
1949 PageAnonExclusive(subpage);
a98a2f0c 1950
dfc7ab57 1951 if (folio_test_hugetlb(folio)) {
0506c31d
BW
1952 bool anon = folio_test_anon(folio);
1953
54205e9c
BW
1954 /*
1955 * huge_pmd_unshare may unmap an entire PMD page.
1956 * There is no way of knowing exactly which PMDs may
1957 * be cached for this mm, so we must flush them all.
1958 * start/end were already adjusted above to cover this
1959 * range.
1960 */
1961 flush_cache_range(vma, range.start, range.end);
1962
0506c31d
BW
1963 /*
1964 * To call huge_pmd_unshare, i_mmap_rwsem must be
1965 * held in write mode. Caller needs to explicitly
1966 * do this outside rmap routines.
40549ba8
MK
1967 *
1968 * We also must hold hugetlb vma_lock in write mode.
1969 * Lock order dictates acquiring vma_lock BEFORE
1970 * i_mmap_rwsem. We can only try lock here and
1971 * fail if unsuccessful.
0506c31d 1972 */
40549ba8
MK
1973 if (!anon) {
1974 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1975 if (!hugetlb_vma_trylock_write(vma)) {
1976 page_vma_mapped_walk_done(&pvmw);
1977 ret = false;
1978 break;
1979 }
1980 if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1981 hugetlb_vma_unlock_write(vma);
1982 flush_tlb_range(vma,
1983 range.start, range.end);
1984 mmu_notifier_invalidate_range(mm,
1985 range.start, range.end);
1986
1987 /*
1988 * The ref count of the PMD page was
1989 * dropped which is part of the way map
1990 * counting is done for shared PMDs.
1991 * Return 'true' here. When there is
1992 * no other sharing, huge_pmd_unshare
1993 * returns false and we will unmap the
1994 * actual page and drop map count
1995 * to zero.
1996 */
1997 page_vma_mapped_walk_done(&pvmw);
1998 break;
1999 }
2000 hugetlb_vma_unlock_write(vma);
a98a2f0c 2001 }
5d4af619
BW
2002 /* Nuke the hugetlb page table entry */
2003 pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
54205e9c
BW
2004 } else {
2005 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
5d4af619
BW
2006 /* Nuke the page table entry. */
2007 pteval = ptep_clear_flush(vma, address, pvmw.pte);
a98a2f0c
AP
2008 }
2009
4b8554c5 2010 /* Set the dirty flag on the folio now the pte is gone. */
a98a2f0c 2011 if (pte_dirty(pteval))
4b8554c5 2012 folio_mark_dirty(folio);
a98a2f0c
AP
2013
2014 /* Update high watermark before we lower rss */
2015 update_hiwater_rss(mm);
2016
f25cbb7a 2017 if (folio_is_device_private(folio)) {
4b8554c5 2018 unsigned long pfn = folio_pfn(folio);
a98a2f0c
AP
2019 swp_entry_t entry;
2020 pte_t swp_pte;
2021
6c287605
DH
2022 if (anon_exclusive)
2023 BUG_ON(page_try_share_anon_rmap(subpage));
2024
a98a2f0c
AP
2025 /*
2026 * Store the pfn of the page in a special migration
2027 * pte. do_swap_page() will wait until the migration
2028 * pte is removed and then restart fault handling.
2029 */
3d88705c
AP
2030 entry = pte_to_swp_entry(pteval);
2031 if (is_writable_device_private_entry(entry))
2032 entry = make_writable_migration_entry(pfn);
6c287605
DH
2033 else if (anon_exclusive)
2034 entry = make_readable_exclusive_migration_entry(pfn);
3d88705c
AP
2035 else
2036 entry = make_readable_migration_entry(pfn);
a98a2f0c
AP
2037 swp_pte = swp_entry_to_pte(entry);
2038
2039 /*
2040 * pteval maps a zone device page and is therefore
2041 * a swap pte.
2042 */
2043 if (pte_swp_soft_dirty(pteval))
2044 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2045 if (pte_swp_uffd_wp(pteval))
2046 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2047 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
4cc79b33
AK
2048 trace_set_migration_pte(pvmw.address, pte_val(swp_pte),
2049 compound_order(&folio->page));
a98a2f0c
AP
2050 /*
2051 * No need to invalidate here it will synchronize on
2052 * against the special swap migration pte.
a98a2f0c 2053 */
da358d5c 2054 } else if (PageHWPoison(subpage)) {
a98a2f0c 2055 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
4b8554c5
MWO
2056 if (folio_test_hugetlb(folio)) {
2057 hugetlb_count_sub(folio_nr_pages(folio), mm);
18f39629 2058 set_huge_pte_at(mm, address, pvmw.pte, pteval);
a98a2f0c 2059 } else {
4b8554c5 2060 dec_mm_counter(mm, mm_counter(&folio->page));
a98a2f0c
AP
2061 set_pte_at(mm, address, pvmw.pte, pteval);
2062 }
2063
2064 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
2065 /*
2066 * The guest indicated that the page content is of no
2067 * interest anymore. Simply discard the pte, vmscan
2068 * will take care of the rest.
2069 * A future reference will then fault in a new zero
2070 * page. When userfaultfd is active, we must not drop
2071 * this page though, as its main user (postcopy
2072 * migration) will not expect userfaults on already
2073 * copied pages.
2074 */
4b8554c5 2075 dec_mm_counter(mm, mm_counter(&folio->page));
a98a2f0c
AP
2076 /* We have to invalidate as we cleared the pte */
2077 mmu_notifier_invalidate_range(mm, address,
2078 address + PAGE_SIZE);
2079 } else {
2080 swp_entry_t entry;
2081 pte_t swp_pte;
2082
2083 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
5d4af619
BW
2084 if (folio_test_hugetlb(folio))
2085 set_huge_pte_at(mm, address, pvmw.pte, pteval);
2086 else
2087 set_pte_at(mm, address, pvmw.pte, pteval);
a98a2f0c
AP
2088 ret = false;
2089 page_vma_mapped_walk_done(&pvmw);
2090 break;
2091 }
6c287605
DH
2092 VM_BUG_ON_PAGE(pte_write(pteval) && folio_test_anon(folio) &&
2093 !anon_exclusive, subpage);
088b8aa5
DH
2094
2095 /* See page_try_share_anon_rmap(): clear PTE first. */
6c287605
DH
2096 if (anon_exclusive &&
2097 page_try_share_anon_rmap(subpage)) {
5d4af619
BW
2098 if (folio_test_hugetlb(folio))
2099 set_huge_pte_at(mm, address, pvmw.pte, pteval);
2100 else
2101 set_pte_at(mm, address, pvmw.pte, pteval);
6c287605
DH
2102 ret = false;
2103 page_vma_mapped_walk_done(&pvmw);
2104 break;
2105 }
a98a2f0c
AP
2106
2107 /*
2108 * Store the pfn of the page in a special migration
2109 * pte. do_swap_page() will wait until the migration
2110 * pte is removed and then restart fault handling.
2111 */
2112 if (pte_write(pteval))
2113 entry = make_writable_migration_entry(
2114 page_to_pfn(subpage));
6c287605
DH
2115 else if (anon_exclusive)
2116 entry = make_readable_exclusive_migration_entry(
2117 page_to_pfn(subpage));
a98a2f0c
AP
2118 else
2119 entry = make_readable_migration_entry(
2120 page_to_pfn(subpage));
2e346877
PX
2121 if (pte_young(pteval))
2122 entry = make_migration_entry_young(entry);
2123 if (pte_dirty(pteval))
2124 entry = make_migration_entry_dirty(entry);
a98a2f0c
AP
2125 swp_pte = swp_entry_to_pte(entry);
2126 if (pte_soft_dirty(pteval))
2127 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2128 if (pte_uffd_wp(pteval))
2129 swp_pte = pte_swp_mkuffd_wp(swp_pte);
5d4af619 2130 if (folio_test_hugetlb(folio))
18f39629 2131 set_huge_pte_at(mm, address, pvmw.pte, swp_pte);
5d4af619
BW
2132 else
2133 set_pte_at(mm, address, pvmw.pte, swp_pte);
4cc79b33
AK
2134 trace_set_migration_pte(address, pte_val(swp_pte),
2135 compound_order(&folio->page));
a98a2f0c
AP
2136 /*
2137 * No need to invalidate here it will synchronize on
2138 * against the special swap migration pte.
2139 */
2140 }
2141
2142 /*
2143 * No need to call mmu_notifier_invalidate_range() it has be
2144 * done above for all cases requiring it to happen under page
2145 * table lock before mmu_notifier_invalidate_range_end()
2146 *
ee65728e 2147 * See Documentation/mm/mmu_notifier.rst
a98a2f0c 2148 */
4b8554c5 2149 page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
b7435507 2150 if (vma->vm_flags & VM_LOCKED)
adb11e78 2151 mlock_page_drain_local();
4b8554c5 2152 folio_put(folio);
a98a2f0c
AP
2153 }
2154
2155 mmu_notifier_invalidate_range_end(&range);
2156
2157 return ret;
2158}
2159
2160/**
2161 * try_to_migrate - try to replace all page table mappings with swap entries
4b8554c5 2162 * @folio: the folio to replace page table entries for
a98a2f0c
AP
2163 * @flags: action and flags
2164 *
4b8554c5
MWO
2165 * Tries to remove all the page table entries which are mapping this folio and
2166 * replace them with special swap entries. Caller must hold the folio lock.
a98a2f0c 2167 */
4b8554c5 2168void try_to_migrate(struct folio *folio, enum ttu_flags flags)
a98a2f0c
AP
2169{
2170 struct rmap_walk_control rwc = {
2171 .rmap_one = try_to_migrate_one,
2172 .arg = (void *)flags,
f3ad032c 2173 .done = folio_not_mapped,
2f031c6f 2174 .anon_lock = folio_lock_anon_vma_read,
a98a2f0c
AP
2175 };
2176
2177 /*
2178 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
2179 * TTU_SPLIT_HUGE_PMD and TTU_SYNC flags.
2180 */
2181 if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2182 TTU_SYNC)))
2183 return;
2184
f25cbb7a
AS
2185 if (folio_is_zone_device(folio) &&
2186 (!folio_is_device_private(folio) && !folio_is_device_coherent(folio)))
6c855fce
HD
2187 return;
2188
52629506
JK
2189 /*
2190 * During exec, a temporary VMA is setup and later moved.
2191 * The VMA is moved under the anon_vma lock but not the
2192 * page tables leading to a race where migration cannot
2193 * find the migration ptes. Rather than increasing the
2194 * locking requirements of exec(), migration skips
2195 * temporary VMAs until after exec() completes.
2196 */
4b8554c5 2197 if (!folio_test_ksm(folio) && folio_test_anon(folio))
52629506
JK
2198 rwc.invalid_vma = invalid_migration_vma;
2199
2a52bcbc 2200 if (flags & TTU_RMAP_LOCKED)
2f031c6f 2201 rmap_walk_locked(folio, &rwc);
2a52bcbc 2202 else
2f031c6f 2203 rmap_walk(folio, &rwc);
b291f000 2204}
e9995ef9 2205
b756a3b5
AP
2206#ifdef CONFIG_DEVICE_PRIVATE
2207struct make_exclusive_args {
2208 struct mm_struct *mm;
2209 unsigned long address;
2210 void *owner;
2211 bool valid;
2212};
2213
2f031c6f 2214static bool page_make_device_exclusive_one(struct folio *folio,
b756a3b5
AP
2215 struct vm_area_struct *vma, unsigned long address, void *priv)
2216{
2217 struct mm_struct *mm = vma->vm_mm;
0d251485 2218 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
b756a3b5
AP
2219 struct make_exclusive_args *args = priv;
2220 pte_t pteval;
2221 struct page *subpage;
2222 bool ret = true;
2223 struct mmu_notifier_range range;
2224 swp_entry_t entry;
2225 pte_t swp_pte;
2226
2227 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
2228 vma->vm_mm, address, min(vma->vm_end,
0d251485
MWO
2229 address + folio_size(folio)),
2230 args->owner);
b756a3b5
AP
2231 mmu_notifier_invalidate_range_start(&range);
2232
2233 while (page_vma_mapped_walk(&pvmw)) {
2234 /* Unexpected PMD-mapped THP? */
0d251485 2235 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
b756a3b5
AP
2236
2237 if (!pte_present(*pvmw.pte)) {
2238 ret = false;
2239 page_vma_mapped_walk_done(&pvmw);
2240 break;
2241 }
2242
0d251485
MWO
2243 subpage = folio_page(folio,
2244 pte_pfn(*pvmw.pte) - folio_pfn(folio));
b756a3b5
AP
2245 address = pvmw.address;
2246
2247 /* Nuke the page table entry. */
2248 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
2249 pteval = ptep_clear_flush(vma, address, pvmw.pte);
2250
0d251485 2251 /* Set the dirty flag on the folio now the pte is gone. */
b756a3b5 2252 if (pte_dirty(pteval))
0d251485 2253 folio_mark_dirty(folio);
b756a3b5
AP
2254
2255 /*
2256 * Check that our target page is still mapped at the expected
2257 * address.
2258 */
2259 if (args->mm == mm && args->address == address &&
2260 pte_write(pteval))
2261 args->valid = true;
2262
2263 /*
2264 * Store the pfn of the page in a special migration
2265 * pte. do_swap_page() will wait until the migration
2266 * pte is removed and then restart fault handling.
2267 */
2268 if (pte_write(pteval))
2269 entry = make_writable_device_exclusive_entry(
2270 page_to_pfn(subpage));
2271 else
2272 entry = make_readable_device_exclusive_entry(
2273 page_to_pfn(subpage));
2274 swp_pte = swp_entry_to_pte(entry);
2275 if (pte_soft_dirty(pteval))
2276 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2277 if (pte_uffd_wp(pteval))
2278 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2279
2280 set_pte_at(mm, address, pvmw.pte, swp_pte);
2281
2282 /*
2283 * There is a reference on the page for the swap entry which has
2284 * been removed, so shouldn't take another.
2285 */
cea86fe2 2286 page_remove_rmap(subpage, vma, false);
b756a3b5
AP
2287 }
2288
2289 mmu_notifier_invalidate_range_end(&range);
2290
2291 return ret;
2292}
2293
2294/**
0d251485
MWO
2295 * folio_make_device_exclusive - Mark the folio exclusively owned by a device.
2296 * @folio: The folio to replace page table entries for.
2297 * @mm: The mm_struct where the folio is expected to be mapped.
2298 * @address: Address where the folio is expected to be mapped.
b756a3b5
AP
2299 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2300 *
0d251485
MWO
2301 * Tries to remove all the page table entries which are mapping this
2302 * folio and replace them with special device exclusive swap entries to
2303 * grant a device exclusive access to the folio.
b756a3b5 2304 *
0d251485
MWO
2305 * Context: Caller must hold the folio lock.
2306 * Return: false if the page is still mapped, or if it could not be unmapped
b756a3b5
AP
2307 * from the expected address. Otherwise returns true (success).
2308 */
0d251485
MWO
2309static bool folio_make_device_exclusive(struct folio *folio,
2310 struct mm_struct *mm, unsigned long address, void *owner)
b756a3b5
AP
2311{
2312 struct make_exclusive_args args = {
2313 .mm = mm,
2314 .address = address,
2315 .owner = owner,
2316 .valid = false,
2317 };
2318 struct rmap_walk_control rwc = {
2319 .rmap_one = page_make_device_exclusive_one,
f3ad032c 2320 .done = folio_not_mapped,
2f031c6f 2321 .anon_lock = folio_lock_anon_vma_read,
b756a3b5
AP
2322 .arg = &args,
2323 };
2324
2325 /*
0d251485
MWO
2326 * Restrict to anonymous folios for now to avoid potential writeback
2327 * issues.
b756a3b5 2328 */
0d251485 2329 if (!folio_test_anon(folio))
b756a3b5
AP
2330 return false;
2331
2f031c6f 2332 rmap_walk(folio, &rwc);
b756a3b5 2333
0d251485 2334 return args.valid && !folio_mapcount(folio);
b756a3b5
AP
2335}
2336
2337/**
2338 * make_device_exclusive_range() - Mark a range for exclusive use by a device
dd062302 2339 * @mm: mm_struct of associated target process
b756a3b5
AP
2340 * @start: start of the region to mark for exclusive device access
2341 * @end: end address of region
2342 * @pages: returns the pages which were successfully marked for exclusive access
2343 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2344 *
2345 * Returns: number of pages found in the range by GUP. A page is marked for
2346 * exclusive access only if the page pointer is non-NULL.
2347 *
2348 * This function finds ptes mapping page(s) to the given address range, locks
2349 * them and replaces mappings with special swap entries preventing userspace CPU
2350 * access. On fault these entries are replaced with the original mapping after
2351 * calling MMU notifiers.
2352 *
2353 * A driver using this to program access from a device must use a mmu notifier
2354 * critical section to hold a device specific lock during programming. Once
2355 * programming is complete it should drop the page lock and reference after
2356 * which point CPU access to the page will revoke the exclusive access.
2357 */
2358int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
2359 unsigned long end, struct page **pages,
2360 void *owner)
2361{
2362 long npages = (end - start) >> PAGE_SHIFT;
2363 long i;
2364
2365 npages = get_user_pages_remote(mm, start, npages,
2366 FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
2367 pages, NULL, NULL);
2368 if (npages < 0)
2369 return npages;
2370
2371 for (i = 0; i < npages; i++, start += PAGE_SIZE) {
0d251485
MWO
2372 struct folio *folio = page_folio(pages[i]);
2373 if (PageTail(pages[i]) || !folio_trylock(folio)) {
2374 folio_put(folio);
b756a3b5
AP
2375 pages[i] = NULL;
2376 continue;
2377 }
2378
0d251485
MWO
2379 if (!folio_make_device_exclusive(folio, mm, start, owner)) {
2380 folio_unlock(folio);
2381 folio_put(folio);
b756a3b5
AP
2382 pages[i] = NULL;
2383 }
2384 }
2385
2386 return npages;
2387}
2388EXPORT_SYMBOL_GPL(make_device_exclusive_range);
2389#endif
2390
01d8b20d 2391void __put_anon_vma(struct anon_vma *anon_vma)
76545066 2392{
01d8b20d 2393 struct anon_vma *root = anon_vma->root;
76545066 2394
624483f3 2395 anon_vma_free(anon_vma);
01d8b20d
PZ
2396 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
2397 anon_vma_free(root);
76545066 2398}
76545066 2399
2f031c6f 2400static struct anon_vma *rmap_walk_anon_lock(struct folio *folio,
6d4675e6 2401 struct rmap_walk_control *rwc)
faecd8dd
JK
2402{
2403 struct anon_vma *anon_vma;
2404
0dd1c7bb 2405 if (rwc->anon_lock)
6d4675e6 2406 return rwc->anon_lock(folio, rwc);
0dd1c7bb 2407
faecd8dd 2408 /*
2f031c6f 2409 * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read()
faecd8dd 2410 * because that depends on page_mapped(); but not all its usages
c1e8d7c6 2411 * are holding mmap_lock. Users without mmap_lock are required to
faecd8dd
JK
2412 * take a reference count to prevent the anon_vma disappearing
2413 */
e05b3453 2414 anon_vma = folio_anon_vma(folio);
faecd8dd
JK
2415 if (!anon_vma)
2416 return NULL;
2417
6d4675e6
MK
2418 if (anon_vma_trylock_read(anon_vma))
2419 goto out;
2420
2421 if (rwc->try_lock) {
2422 anon_vma = NULL;
2423 rwc->contended = true;
2424 goto out;
2425 }
2426
faecd8dd 2427 anon_vma_lock_read(anon_vma);
6d4675e6 2428out:
faecd8dd
JK
2429 return anon_vma;
2430}
2431
e9995ef9 2432/*
e8351ac9
JK
2433 * rmap_walk_anon - do something to anonymous page using the object-based
2434 * rmap method
2435 * @page: the page to be handled
2436 * @rwc: control variable according to each walk type
2437 *
2438 * Find all the mappings of a page using the mapping pointer and the vma chains
2439 * contained in the anon_vma struct it points to.
e9995ef9 2440 */
84fbbe21 2441static void rmap_walk_anon(struct folio *folio,
6d4675e6 2442 struct rmap_walk_control *rwc, bool locked)
e9995ef9
HD
2443{
2444 struct anon_vma *anon_vma;
a8fa41ad 2445 pgoff_t pgoff_start, pgoff_end;
5beb4930 2446 struct anon_vma_chain *avc;
e9995ef9 2447
b9773199 2448 if (locked) {
e05b3453 2449 anon_vma = folio_anon_vma(folio);
b9773199 2450 /* anon_vma disappear under us? */
e05b3453 2451 VM_BUG_ON_FOLIO(!anon_vma, folio);
b9773199 2452 } else {
2f031c6f 2453 anon_vma = rmap_walk_anon_lock(folio, rwc);
b9773199 2454 }
e9995ef9 2455 if (!anon_vma)
1df631ae 2456 return;
faecd8dd 2457
2f031c6f
MWO
2458 pgoff_start = folio_pgoff(folio);
2459 pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
a8fa41ad
KS
2460 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
2461 pgoff_start, pgoff_end) {
5beb4930 2462 struct vm_area_struct *vma = avc->vma;
2f031c6f 2463 unsigned long address = vma_address(&folio->page, vma);
0dd1c7bb 2464
494334e4 2465 VM_BUG_ON_VMA(address == -EFAULT, vma);
ad12695f
AA
2466 cond_resched();
2467
0dd1c7bb
JK
2468 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2469 continue;
2470
2f031c6f 2471 if (!rwc->rmap_one(folio, vma, address, rwc->arg))
e9995ef9 2472 break;
2f031c6f 2473 if (rwc->done && rwc->done(folio))
0dd1c7bb 2474 break;
e9995ef9 2475 }
b9773199
KS
2476
2477 if (!locked)
2478 anon_vma_unlock_read(anon_vma);
e9995ef9
HD
2479}
2480
e8351ac9
JK
2481/*
2482 * rmap_walk_file - do something to file page using the object-based rmap method
2483 * @page: the page to be handled
2484 * @rwc: control variable according to each walk type
2485 *
2486 * Find all the mappings of a page using the mapping pointer and the vma chains
2487 * contained in the address_space struct it points to.
e8351ac9 2488 */
84fbbe21 2489static void rmap_walk_file(struct folio *folio,
6d4675e6 2490 struct rmap_walk_control *rwc, bool locked)
e9995ef9 2491{
2f031c6f 2492 struct address_space *mapping = folio_mapping(folio);
a8fa41ad 2493 pgoff_t pgoff_start, pgoff_end;
e9995ef9 2494 struct vm_area_struct *vma;
e9995ef9 2495
9f32624b
JK
2496 /*
2497 * The page lock not only makes sure that page->mapping cannot
2498 * suddenly be NULLified by truncation, it makes sure that the
2499 * structure at mapping cannot be freed and reused yet,
c8c06efa 2500 * so we can safely take mapping->i_mmap_rwsem.
9f32624b 2501 */
2f031c6f 2502 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
9f32624b 2503
e9995ef9 2504 if (!mapping)
1df631ae 2505 return;
3dec0ba0 2506
2f031c6f
MWO
2507 pgoff_start = folio_pgoff(folio);
2508 pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
6d4675e6
MK
2509 if (!locked) {
2510 if (i_mmap_trylock_read(mapping))
2511 goto lookup;
2512
2513 if (rwc->try_lock) {
2514 rwc->contended = true;
2515 return;
2516 }
2517
b9773199 2518 i_mmap_lock_read(mapping);
6d4675e6
MK
2519 }
2520lookup:
a8fa41ad
KS
2521 vma_interval_tree_foreach(vma, &mapping->i_mmap,
2522 pgoff_start, pgoff_end) {
2f031c6f 2523 unsigned long address = vma_address(&folio->page, vma);
0dd1c7bb 2524
494334e4 2525 VM_BUG_ON_VMA(address == -EFAULT, vma);
ad12695f
AA
2526 cond_resched();
2527
0dd1c7bb
JK
2528 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2529 continue;
2530
2f031c6f 2531 if (!rwc->rmap_one(folio, vma, address, rwc->arg))
0dd1c7bb 2532 goto done;
2f031c6f 2533 if (rwc->done && rwc->done(folio))
0dd1c7bb 2534 goto done;
e9995ef9 2535 }
0dd1c7bb 2536
0dd1c7bb 2537done:
b9773199
KS
2538 if (!locked)
2539 i_mmap_unlock_read(mapping);
e9995ef9
HD
2540}
2541
6d4675e6 2542void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc)
e9995ef9 2543{
2f031c6f
MWO
2544 if (unlikely(folio_test_ksm(folio)))
2545 rmap_walk_ksm(folio, rwc);
2546 else if (folio_test_anon(folio))
2547 rmap_walk_anon(folio, rwc, false);
b9773199 2548 else
2f031c6f 2549 rmap_walk_file(folio, rwc, false);
b9773199
KS
2550}
2551
2552/* Like rmap_walk, but caller holds relevant rmap lock */
6d4675e6 2553void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc)
b9773199
KS
2554{
2555 /* no ksm support for now */
2f031c6f
MWO
2556 VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio);
2557 if (folio_test_anon(folio))
2558 rmap_walk_anon(folio, rwc, true);
e9995ef9 2559 else
2f031c6f 2560 rmap_walk_file(folio, rwc, true);
e9995ef9 2561}
0fe6e20b 2562
e3390f67 2563#ifdef CONFIG_HUGETLB_PAGE
0fe6e20b 2564/*
451b9514 2565 * The following two functions are for anonymous (private mapped) hugepages.
0fe6e20b
NH
2566 * Unlike common anonymous pages, anonymous hugepages have no accounting code
2567 * and no lru code, because we handle hugepages differently from common pages.
28c5209d
DH
2568 *
2569 * RMAP_COMPOUND is ignored.
0fe6e20b 2570 */
28c5209d
DH
2571void hugepage_add_anon_rmap(struct page *page, struct vm_area_struct *vma,
2572 unsigned long address, rmap_t flags)
0fe6e20b
NH
2573{
2574 struct anon_vma *anon_vma = vma->anon_vma;
2575 int first;
a850ea30
NH
2576
2577 BUG_ON(!PageLocked(page));
0fe6e20b 2578 BUG_ON(!anon_vma);
5dbe0af4 2579 /* address might be in next vma when migration races vma_adjust */
53f9263b 2580 first = atomic_inc_and_test(compound_mapcount_ptr(page));
6c287605
DH
2581 VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
2582 VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
0fe6e20b 2583 if (first)
28c5209d
DH
2584 __page_set_anon_rmap(page, vma, address,
2585 !!(flags & RMAP_EXCLUSIVE));
0fe6e20b
NH
2586}
2587
2588void hugepage_add_new_anon_rmap(struct page *page,
2589 struct vm_area_struct *vma, unsigned long address)
2590{
2591 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
cb67f428 2592 /* increment count (starts at -1) */
53f9263b 2593 atomic_set(compound_mapcount_ptr(page), 0);
4781593d 2594 ClearHPageRestoreReserve(page);
451b9514 2595 __page_set_anon_rmap(page, vma, address, 1);
0fe6e20b 2596}
e3390f67 2597#endif /* CONFIG_HUGETLB_PAGE */