mm: memcontrol: switch to native NR_FILE_PAGES and NR_SHMEM counters
[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 *
1b1dcc1b 23 * inode->i_mutex (while writing or truncating, not reading or faulting)
82591e6e 24 * mm->mmap_sem
c0d0381a 25 * page->flags PG_locked (lock_page) * (see huegtlbfs below)
88f306b6
KS
26 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
27 * mapping->i_mmap_rwsem
c0d0381a 28 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
88f306b6
KS
29 * anon_vma->rwsem
30 * mm->page_table_lock or pte_lock
f4b7e272 31 * pgdat->lru_lock (in mark_page_accessed, isolate_lru_page)
88f306b6
KS
32 * swap_lock (in swap_duplicate, swap_info_get)
33 * mmlist_lock (in mmput, drain_mmlist and others)
34 * mapping->private_lock (in __set_page_dirty_buffers)
35 * mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
b93b0163 36 * i_pages lock (widely used)
88f306b6
KS
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)
b93b0163 40 * i_pages lock (widely used, in set_page_dirty,
88f306b6
KS
41 * in arch-dependent flush_dcache_mmap_lock,
42 * within bdi.wb->list_lock in __sync_single_inode)
6a46079c 43 *
5a505085 44 * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon)
9b679320 45 * ->tasklist_lock
6a46079c 46 * pte map lock
c0d0381a
MK
47 *
48 * * hugetlbfs PageHuge() pages take locks in this order:
49 * mapping->i_mmap_rwsem
50 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
51 * page->flags PG_locked (lock_page)
1da177e4
LT
52 */
53
54#include <linux/mm.h>
6e84f315 55#include <linux/sched/mm.h>
29930025 56#include <linux/sched/task.h>
1da177e4
LT
57#include <linux/pagemap.h>
58#include <linux/swap.h>
59#include <linux/swapops.h>
60#include <linux/slab.h>
61#include <linux/init.h>
5ad64688 62#include <linux/ksm.h>
1da177e4
LT
63#include <linux/rmap.h>
64#include <linux/rcupdate.h>
b95f1b31 65#include <linux/export.h>
8a9f3ccd 66#include <linux/memcontrol.h>
cddb8a5c 67#include <linux/mmu_notifier.h>
64cdd548 68#include <linux/migrate.h>
0fe6e20b 69#include <linux/hugetlb.h>
444f84fd 70#include <linux/huge_mm.h>
ef5d437f 71#include <linux/backing-dev.h>
33c3fc71 72#include <linux/page_idle.h>
a5430dda 73#include <linux/memremap.h>
bce73e48 74#include <linux/userfaultfd_k.h>
1da177e4
LT
75
76#include <asm/tlbflush.h>
77
72b252ae
MG
78#include <trace/events/tlb.h>
79
b291f000
NP
80#include "internal.h"
81
fdd2e5f8 82static struct kmem_cache *anon_vma_cachep;
5beb4930 83static struct kmem_cache *anon_vma_chain_cachep;
fdd2e5f8
AB
84
85static inline struct anon_vma *anon_vma_alloc(void)
86{
01d8b20d
PZ
87 struct anon_vma *anon_vma;
88
89 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
90 if (anon_vma) {
91 atomic_set(&anon_vma->refcount, 1);
7a3ef208
KK
92 anon_vma->degree = 1; /* Reference for first vma */
93 anon_vma->parent = anon_vma;
01d8b20d
PZ
94 /*
95 * Initialise the anon_vma root to point to itself. If called
96 * from fork, the root will be reset to the parents anon_vma.
97 */
98 anon_vma->root = anon_vma;
99 }
100
101 return anon_vma;
fdd2e5f8
AB
102}
103
01d8b20d 104static inline void anon_vma_free(struct anon_vma *anon_vma)
fdd2e5f8 105{
01d8b20d 106 VM_BUG_ON(atomic_read(&anon_vma->refcount));
88c22088
PZ
107
108 /*
4fc3f1d6 109 * Synchronize against page_lock_anon_vma_read() such that
88c22088
PZ
110 * we can safely hold the lock without the anon_vma getting
111 * freed.
112 *
113 * Relies on the full mb implied by the atomic_dec_and_test() from
114 * put_anon_vma() against the acquire barrier implied by
4fc3f1d6 115 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
88c22088 116 *
4fc3f1d6
IM
117 * page_lock_anon_vma_read() VS put_anon_vma()
118 * down_read_trylock() atomic_dec_and_test()
88c22088 119 * LOCK MB
4fc3f1d6 120 * atomic_read() rwsem_is_locked()
88c22088
PZ
121 *
122 * LOCK should suffice since the actual taking of the lock must
123 * happen _before_ what follows.
124 */
7f39dda9 125 might_sleep();
5a505085 126 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
4fc3f1d6 127 anon_vma_lock_write(anon_vma);
08b52706 128 anon_vma_unlock_write(anon_vma);
88c22088
PZ
129 }
130
fdd2e5f8
AB
131 kmem_cache_free(anon_vma_cachep, anon_vma);
132}
1da177e4 133
dd34739c 134static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
5beb4930 135{
dd34739c 136 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
5beb4930
RR
137}
138
e574b5fd 139static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
5beb4930
RR
140{
141 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
142}
143
6583a843
KC
144static void anon_vma_chain_link(struct vm_area_struct *vma,
145 struct anon_vma_chain *avc,
146 struct anon_vma *anon_vma)
147{
148 avc->vma = vma;
149 avc->anon_vma = anon_vma;
150 list_add(&avc->same_vma, &vma->anon_vma_chain);
bf181b9f 151 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
6583a843
KC
152}
153
d9d332e0 154/**
d5a187da 155 * __anon_vma_prepare - attach an anon_vma to a memory region
d9d332e0
LT
156 * @vma: the memory region in question
157 *
158 * This makes sure the memory mapping described by 'vma' has
159 * an 'anon_vma' attached to it, so that we can associate the
160 * anonymous pages mapped into it with that anon_vma.
161 *
d5a187da
VB
162 * The common case will be that we already have one, which
163 * is handled inline by anon_vma_prepare(). But if
23a0790a 164 * not we either need to find an adjacent mapping that we
d9d332e0
LT
165 * can re-use the anon_vma from (very common when the only
166 * reason for splitting a vma has been mprotect()), or we
167 * allocate a new one.
168 *
169 * Anon-vma allocations are very subtle, because we may have
4fc3f1d6 170 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
d9d332e0
LT
171 * and that may actually touch the spinlock even in the newly
172 * allocated vma (it depends on RCU to make sure that the
173 * anon_vma isn't actually destroyed).
174 *
175 * As a result, we need to do proper anon_vma locking even
176 * for the new allocation. At the same time, we do not want
177 * to do any locking for the common case of already having
178 * an anon_vma.
179 *
180 * This must be called with the mmap_sem held for reading.
181 */
d5a187da 182int __anon_vma_prepare(struct vm_area_struct *vma)
1da177e4 183{
d5a187da
VB
184 struct mm_struct *mm = vma->vm_mm;
185 struct anon_vma *anon_vma, *allocated;
5beb4930 186 struct anon_vma_chain *avc;
1da177e4
LT
187
188 might_sleep();
1da177e4 189
d5a187da
VB
190 avc = anon_vma_chain_alloc(GFP_KERNEL);
191 if (!avc)
192 goto out_enomem;
193
194 anon_vma = find_mergeable_anon_vma(vma);
195 allocated = NULL;
196 if (!anon_vma) {
197 anon_vma = anon_vma_alloc();
198 if (unlikely(!anon_vma))
199 goto out_enomem_free_avc;
200 allocated = anon_vma;
201 }
5beb4930 202
d5a187da
VB
203 anon_vma_lock_write(anon_vma);
204 /* page_table_lock to protect against threads */
205 spin_lock(&mm->page_table_lock);
206 if (likely(!vma->anon_vma)) {
207 vma->anon_vma = anon_vma;
208 anon_vma_chain_link(vma, avc, anon_vma);
209 /* vma reference or self-parent link for new root */
210 anon_vma->degree++;
d9d332e0 211 allocated = NULL;
d5a187da
VB
212 avc = NULL;
213 }
214 spin_unlock(&mm->page_table_lock);
215 anon_vma_unlock_write(anon_vma);
1da177e4 216
d5a187da
VB
217 if (unlikely(allocated))
218 put_anon_vma(allocated);
219 if (unlikely(avc))
220 anon_vma_chain_free(avc);
31f2b0eb 221
1da177e4 222 return 0;
5beb4930
RR
223
224 out_enomem_free_avc:
225 anon_vma_chain_free(avc);
226 out_enomem:
227 return -ENOMEM;
1da177e4
LT
228}
229
bb4aa396
LT
230/*
231 * This is a useful helper function for locking the anon_vma root as
232 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
233 * have the same vma.
234 *
235 * Such anon_vma's should have the same root, so you'd expect to see
236 * just a single mutex_lock for the whole traversal.
237 */
238static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
239{
240 struct anon_vma *new_root = anon_vma->root;
241 if (new_root != root) {
242 if (WARN_ON_ONCE(root))
5a505085 243 up_write(&root->rwsem);
bb4aa396 244 root = new_root;
5a505085 245 down_write(&root->rwsem);
bb4aa396
LT
246 }
247 return root;
248}
249
250static inline void unlock_anon_vma_root(struct anon_vma *root)
251{
252 if (root)
5a505085 253 up_write(&root->rwsem);
bb4aa396
LT
254}
255
5beb4930
RR
256/*
257 * Attach the anon_vmas from src to dst.
258 * Returns 0 on success, -ENOMEM on failure.
7a3ef208 259 *
47b390d2
WY
260 * anon_vma_clone() is called by __vma_split(), __split_vma(), copy_vma() and
261 * anon_vma_fork(). The first three want an exact copy of src, while the last
262 * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
263 * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
264 * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
265 *
266 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
267 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
268 * This prevents degradation of anon_vma hierarchy to endless linear chain in
269 * case of constantly forking task. On the other hand, an anon_vma with more
270 * than one child isn't reused even if there was no alive vma, thus rmap
271 * walker has a good chance of avoiding scanning the whole hierarchy when it
272 * searches where page is mapped.
5beb4930
RR
273 */
274int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
1da177e4 275{
5beb4930 276 struct anon_vma_chain *avc, *pavc;
bb4aa396 277 struct anon_vma *root = NULL;
5beb4930 278
646d87b4 279 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
bb4aa396
LT
280 struct anon_vma *anon_vma;
281
dd34739c
LT
282 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
283 if (unlikely(!avc)) {
284 unlock_anon_vma_root(root);
285 root = NULL;
286 avc = anon_vma_chain_alloc(GFP_KERNEL);
287 if (!avc)
288 goto enomem_failure;
289 }
bb4aa396
LT
290 anon_vma = pavc->anon_vma;
291 root = lock_anon_vma_root(root, anon_vma);
292 anon_vma_chain_link(dst, avc, anon_vma);
7a3ef208
KK
293
294 /*
295 * Reuse existing anon_vma if its degree lower than two,
296 * that means it has no vma and only one anon_vma child.
297 *
298 * Do not chose parent anon_vma, otherwise first child
299 * will always reuse it. Root anon_vma is never reused:
300 * it has self-parent reference and at least one child.
301 */
47b390d2
WY
302 if (!dst->anon_vma && src->anon_vma &&
303 anon_vma != src->anon_vma && anon_vma->degree < 2)
7a3ef208 304 dst->anon_vma = anon_vma;
5beb4930 305 }
7a3ef208
KK
306 if (dst->anon_vma)
307 dst->anon_vma->degree++;
bb4aa396 308 unlock_anon_vma_root(root);
5beb4930 309 return 0;
1da177e4 310
5beb4930 311 enomem_failure:
3fe89b3e
LY
312 /*
313 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
314 * decremented in unlink_anon_vmas().
315 * We can safely do this because callers of anon_vma_clone() don't care
316 * about dst->anon_vma if anon_vma_clone() failed.
317 */
318 dst->anon_vma = NULL;
5beb4930
RR
319 unlink_anon_vmas(dst);
320 return -ENOMEM;
1da177e4
LT
321}
322
5beb4930
RR
323/*
324 * Attach vma to its own anon_vma, as well as to the anon_vmas that
325 * the corresponding VMA in the parent process is attached to.
326 * Returns 0 on success, non-zero on failure.
327 */
328int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
1da177e4 329{
5beb4930
RR
330 struct anon_vma_chain *avc;
331 struct anon_vma *anon_vma;
c4ea95d7 332 int error;
1da177e4 333
5beb4930
RR
334 /* Don't bother if the parent process has no anon_vma here. */
335 if (!pvma->anon_vma)
336 return 0;
337
7a3ef208
KK
338 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
339 vma->anon_vma = NULL;
340
5beb4930
RR
341 /*
342 * First, attach the new VMA to the parent VMA's anon_vmas,
343 * so rmap can find non-COWed pages in child processes.
344 */
c4ea95d7
DF
345 error = anon_vma_clone(vma, pvma);
346 if (error)
347 return error;
5beb4930 348
7a3ef208
KK
349 /* An existing anon_vma has been reused, all done then. */
350 if (vma->anon_vma)
351 return 0;
352
5beb4930
RR
353 /* Then add our own anon_vma. */
354 anon_vma = anon_vma_alloc();
355 if (!anon_vma)
356 goto out_error;
dd34739c 357 avc = anon_vma_chain_alloc(GFP_KERNEL);
5beb4930
RR
358 if (!avc)
359 goto out_error_free_anon_vma;
5c341ee1
RR
360
361 /*
362 * The root anon_vma's spinlock is the lock actually used when we
363 * lock any of the anon_vmas in this anon_vma tree.
364 */
365 anon_vma->root = pvma->anon_vma->root;
7a3ef208 366 anon_vma->parent = pvma->anon_vma;
76545066 367 /*
01d8b20d
PZ
368 * With refcounts, an anon_vma can stay around longer than the
369 * process it belongs to. The root anon_vma needs to be pinned until
370 * this anon_vma is freed, because the lock lives in the root.
76545066
RR
371 */
372 get_anon_vma(anon_vma->root);
5beb4930
RR
373 /* Mark this anon_vma as the one where our new (COWed) pages go. */
374 vma->anon_vma = anon_vma;
4fc3f1d6 375 anon_vma_lock_write(anon_vma);
5c341ee1 376 anon_vma_chain_link(vma, avc, anon_vma);
7a3ef208 377 anon_vma->parent->degree++;
08b52706 378 anon_vma_unlock_write(anon_vma);
5beb4930
RR
379
380 return 0;
381
382 out_error_free_anon_vma:
01d8b20d 383 put_anon_vma(anon_vma);
5beb4930 384 out_error:
4946d54c 385 unlink_anon_vmas(vma);
5beb4930 386 return -ENOMEM;
1da177e4
LT
387}
388
5beb4930
RR
389void unlink_anon_vmas(struct vm_area_struct *vma)
390{
391 struct anon_vma_chain *avc, *next;
eee2acba 392 struct anon_vma *root = NULL;
5beb4930 393
5c341ee1
RR
394 /*
395 * Unlink each anon_vma chained to the VMA. This list is ordered
396 * from newest to oldest, ensuring the root anon_vma gets freed last.
397 */
5beb4930 398 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
eee2acba
PZ
399 struct anon_vma *anon_vma = avc->anon_vma;
400
401 root = lock_anon_vma_root(root, anon_vma);
bf181b9f 402 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
eee2acba
PZ
403
404 /*
405 * Leave empty anon_vmas on the list - we'll need
406 * to free them outside the lock.
407 */
f808c13f 408 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
7a3ef208 409 anon_vma->parent->degree--;
eee2acba 410 continue;
7a3ef208 411 }
eee2acba
PZ
412
413 list_del(&avc->same_vma);
414 anon_vma_chain_free(avc);
415 }
7a3ef208
KK
416 if (vma->anon_vma)
417 vma->anon_vma->degree--;
eee2acba
PZ
418 unlock_anon_vma_root(root);
419
420 /*
421 * Iterate the list once more, it now only contains empty and unlinked
422 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
5a505085 423 * needing to write-acquire the anon_vma->root->rwsem.
eee2acba
PZ
424 */
425 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
426 struct anon_vma *anon_vma = avc->anon_vma;
427
e4c5800a 428 VM_WARN_ON(anon_vma->degree);
eee2acba
PZ
429 put_anon_vma(anon_vma);
430
5beb4930
RR
431 list_del(&avc->same_vma);
432 anon_vma_chain_free(avc);
433 }
434}
435
51cc5068 436static void anon_vma_ctor(void *data)
1da177e4 437{
a35afb83 438 struct anon_vma *anon_vma = data;
1da177e4 439
5a505085 440 init_rwsem(&anon_vma->rwsem);
83813267 441 atomic_set(&anon_vma->refcount, 0);
f808c13f 442 anon_vma->rb_root = RB_ROOT_CACHED;
1da177e4
LT
443}
444
445void __init anon_vma_init(void)
446{
447 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
5f0d5a3a 448 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
5d097056
VD
449 anon_vma_ctor);
450 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
451 SLAB_PANIC|SLAB_ACCOUNT);
1da177e4
LT
452}
453
454/*
6111e4ca
PZ
455 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
456 *
457 * Since there is no serialization what so ever against page_remove_rmap()
458 * the best this function can do is return a locked anon_vma that might
459 * have been relevant to this page.
460 *
461 * The page might have been remapped to a different anon_vma or the anon_vma
462 * returned may already be freed (and even reused).
463 *
bc658c96
PZ
464 * In case it was remapped to a different anon_vma, the new anon_vma will be a
465 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
466 * ensure that any anon_vma obtained from the page will still be valid for as
467 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
468 *
6111e4ca
PZ
469 * All users of this function must be very careful when walking the anon_vma
470 * chain and verify that the page in question is indeed mapped in it
471 * [ something equivalent to page_mapped_in_vma() ].
472 *
091e4299
MC
473 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
474 * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
475 * if there is a mapcount, we can dereference the anon_vma after observing
476 * those.
1da177e4 477 */
746b18d4 478struct anon_vma *page_get_anon_vma(struct page *page)
1da177e4 479{
746b18d4 480 struct anon_vma *anon_vma = NULL;
1da177e4
LT
481 unsigned long anon_mapping;
482
483 rcu_read_lock();
4db0c3c2 484 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
3ca7b3c5 485 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
1da177e4
LT
486 goto out;
487 if (!page_mapped(page))
488 goto out;
489
490 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
746b18d4
PZ
491 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
492 anon_vma = NULL;
493 goto out;
494 }
f1819427
HD
495
496 /*
497 * If this page is still mapped, then its anon_vma cannot have been
746b18d4
PZ
498 * freed. But if it has been unmapped, we have no security against the
499 * anon_vma structure being freed and reused (for another anon_vma:
5f0d5a3a 500 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
746b18d4 501 * above cannot corrupt).
f1819427 502 */
746b18d4 503 if (!page_mapped(page)) {
7f39dda9 504 rcu_read_unlock();
746b18d4 505 put_anon_vma(anon_vma);
7f39dda9 506 return NULL;
746b18d4 507 }
1da177e4
LT
508out:
509 rcu_read_unlock();
746b18d4
PZ
510
511 return anon_vma;
512}
513
88c22088
PZ
514/*
515 * Similar to page_get_anon_vma() except it locks the anon_vma.
516 *
517 * Its a little more complex as it tries to keep the fast path to a single
518 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
519 * reference like with page_get_anon_vma() and then block on the mutex.
520 */
4fc3f1d6 521struct anon_vma *page_lock_anon_vma_read(struct page *page)
746b18d4 522{
88c22088 523 struct anon_vma *anon_vma = NULL;
eee0f252 524 struct anon_vma *root_anon_vma;
88c22088 525 unsigned long anon_mapping;
746b18d4 526
88c22088 527 rcu_read_lock();
4db0c3c2 528 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
88c22088
PZ
529 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
530 goto out;
531 if (!page_mapped(page))
532 goto out;
533
534 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
4db0c3c2 535 root_anon_vma = READ_ONCE(anon_vma->root);
4fc3f1d6 536 if (down_read_trylock(&root_anon_vma->rwsem)) {
88c22088 537 /*
eee0f252
HD
538 * If the page is still mapped, then this anon_vma is still
539 * its anon_vma, and holding the mutex ensures that it will
bc658c96 540 * not go away, see anon_vma_free().
88c22088 541 */
eee0f252 542 if (!page_mapped(page)) {
4fc3f1d6 543 up_read(&root_anon_vma->rwsem);
88c22088
PZ
544 anon_vma = NULL;
545 }
546 goto out;
547 }
746b18d4 548
88c22088
PZ
549 /* trylock failed, we got to sleep */
550 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
551 anon_vma = NULL;
552 goto out;
553 }
554
555 if (!page_mapped(page)) {
7f39dda9 556 rcu_read_unlock();
88c22088 557 put_anon_vma(anon_vma);
7f39dda9 558 return NULL;
88c22088
PZ
559 }
560
561 /* we pinned the anon_vma, its safe to sleep */
562 rcu_read_unlock();
4fc3f1d6 563 anon_vma_lock_read(anon_vma);
88c22088
PZ
564
565 if (atomic_dec_and_test(&anon_vma->refcount)) {
566 /*
567 * Oops, we held the last refcount, release the lock
568 * and bail -- can't simply use put_anon_vma() because
4fc3f1d6 569 * we'll deadlock on the anon_vma_lock_write() recursion.
88c22088 570 */
4fc3f1d6 571 anon_vma_unlock_read(anon_vma);
88c22088
PZ
572 __put_anon_vma(anon_vma);
573 anon_vma = NULL;
574 }
575
576 return anon_vma;
577
578out:
579 rcu_read_unlock();
746b18d4 580 return anon_vma;
34bbd704
ON
581}
582
4fc3f1d6 583void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
34bbd704 584{
4fc3f1d6 585 anon_vma_unlock_read(anon_vma);
1da177e4
LT
586}
587
72b252ae 588#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
72b252ae
MG
589/*
590 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
591 * important if a PTE was dirty when it was unmapped that it's flushed
592 * before any IO is initiated on the page to prevent lost writes. Similarly,
593 * it must be flushed before freeing to prevent data leakage.
594 */
595void try_to_unmap_flush(void)
596{
597 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
72b252ae
MG
598
599 if (!tlb_ubc->flush_required)
600 return;
601
e73ad5ff 602 arch_tlbbatch_flush(&tlb_ubc->arch);
72b252ae 603 tlb_ubc->flush_required = false;
d950c947 604 tlb_ubc->writable = false;
72b252ae
MG
605}
606
d950c947
MG
607/* Flush iff there are potentially writable TLB entries that can race with IO */
608void try_to_unmap_flush_dirty(void)
609{
610 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
611
612 if (tlb_ubc->writable)
613 try_to_unmap_flush();
614}
615
c7ab0d2f 616static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
72b252ae
MG
617{
618 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
619
e73ad5ff 620 arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
72b252ae 621 tlb_ubc->flush_required = true;
d950c947 622
3ea27719
MG
623 /*
624 * Ensure compiler does not re-order the setting of tlb_flush_batched
625 * before the PTE is cleared.
626 */
627 barrier();
628 mm->tlb_flush_batched = true;
629
d950c947
MG
630 /*
631 * If the PTE was dirty then it's best to assume it's writable. The
632 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
633 * before the page is queued for IO.
634 */
635 if (writable)
636 tlb_ubc->writable = true;
72b252ae
MG
637}
638
639/*
640 * Returns true if the TLB flush should be deferred to the end of a batch of
641 * unmap operations to reduce IPIs.
642 */
643static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
644{
645 bool should_defer = false;
646
647 if (!(flags & TTU_BATCH_FLUSH))
648 return false;
649
650 /* If remote CPUs need to be flushed then defer batch the flush */
651 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
652 should_defer = true;
653 put_cpu();
654
655 return should_defer;
656}
3ea27719
MG
657
658/*
659 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
660 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
661 * operation such as mprotect or munmap to race between reclaim unmapping
662 * the page and flushing the page. If this race occurs, it potentially allows
663 * access to data via a stale TLB entry. Tracking all mm's that have TLB
664 * batching in flight would be expensive during reclaim so instead track
665 * whether TLB batching occurred in the past and if so then do a flush here
666 * if required. This will cost one additional flush per reclaim cycle paid
667 * by the first operation at risk such as mprotect and mumap.
668 *
669 * This must be called under the PTL so that an access to tlb_flush_batched
670 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
671 * via the PTL.
672 */
673void flush_tlb_batched_pending(struct mm_struct *mm)
674{
675 if (mm->tlb_flush_batched) {
676 flush_tlb_mm(mm);
677
678 /*
679 * Do not allow the compiler to re-order the clearing of
680 * tlb_flush_batched before the tlb is flushed.
681 */
682 barrier();
683 mm->tlb_flush_batched = false;
684 }
685}
72b252ae 686#else
c7ab0d2f 687static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
72b252ae
MG
688{
689}
690
691static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
692{
693 return false;
694}
695#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
696
1da177e4 697/*
bf89c8c8 698 * At what user virtual address is page expected in vma?
ab941e0f 699 * Caller should check the page is actually part of the vma.
1da177e4
LT
700 */
701unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
702{
86c2ad19 703 unsigned long address;
21d0d443 704 if (PageAnon(page)) {
4829b906
HD
705 struct anon_vma *page__anon_vma = page_anon_vma(page);
706 /*
707 * Note: swapoff's unuse_vma() is more efficient with this
708 * check, and needs it to match anon_vma when KSM is active.
709 */
710 if (!vma->anon_vma || !page__anon_vma ||
711 vma->anon_vma->root != page__anon_vma->root)
21d0d443 712 return -EFAULT;
27ba0644
KS
713 } else if (page->mapping) {
714 if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
1da177e4
LT
715 return -EFAULT;
716 } else
717 return -EFAULT;
86c2ad19
ML
718 address = __vma_address(page, vma);
719 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
720 return -EFAULT;
721 return address;
1da177e4
LT
722}
723
6219049a
BL
724pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
725{
726 pgd_t *pgd;
c2febafc 727 p4d_t *p4d;
6219049a
BL
728 pud_t *pud;
729 pmd_t *pmd = NULL;
f72e7dcd 730 pmd_t pmde;
6219049a
BL
731
732 pgd = pgd_offset(mm, address);
733 if (!pgd_present(*pgd))
734 goto out;
735
c2febafc
KS
736 p4d = p4d_offset(pgd, address);
737 if (!p4d_present(*p4d))
738 goto out;
739
740 pud = pud_offset(p4d, address);
6219049a
BL
741 if (!pud_present(*pud))
742 goto out;
743
744 pmd = pmd_offset(pud, address);
f72e7dcd 745 /*
8809aa2d 746 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
f72e7dcd
HD
747 * without holding anon_vma lock for write. So when looking for a
748 * genuine pmde (in which to find pte), test present and !THP together.
749 */
e37c6982
CB
750 pmde = *pmd;
751 barrier();
f72e7dcd 752 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
6219049a
BL
753 pmd = NULL;
754out:
755 return pmd;
756}
757
8749cfea
VD
758struct page_referenced_arg {
759 int mapcount;
760 int referenced;
761 unsigned long vm_flags;
762 struct mem_cgroup *memcg;
763};
764/*
765 * arg: page_referenced_arg will be passed
766 */
e4b82222 767static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
8749cfea
VD
768 unsigned long address, void *arg)
769{
8749cfea 770 struct page_referenced_arg *pra = arg;
8eaedede
KS
771 struct page_vma_mapped_walk pvmw = {
772 .page = page,
773 .vma = vma,
774 .address = address,
775 };
8749cfea
VD
776 int referenced = 0;
777
8eaedede
KS
778 while (page_vma_mapped_walk(&pvmw)) {
779 address = pvmw.address;
b20ce5e0 780
8eaedede
KS
781 if (vma->vm_flags & VM_LOCKED) {
782 page_vma_mapped_walk_done(&pvmw);
783 pra->vm_flags |= VM_LOCKED;
e4b82222 784 return false; /* To break the loop */
8eaedede 785 }
71e3aac0 786
8eaedede
KS
787 if (pvmw.pte) {
788 if (ptep_clear_flush_young_notify(vma, address,
789 pvmw.pte)) {
790 /*
791 * Don't treat a reference through
792 * a sequentially read mapping as such.
793 * If the page has been used in another mapping,
794 * we will catch it; if this other mapping is
795 * already gone, the unmap path will have set
796 * PG_referenced or activated the page.
797 */
798 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
799 referenced++;
800 }
801 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
802 if (pmdp_clear_flush_young_notify(vma, address,
803 pvmw.pmd))
8749cfea 804 referenced++;
8eaedede
KS
805 } else {
806 /* unexpected pmd-mapped page? */
807 WARN_ON_ONCE(1);
8749cfea 808 }
8eaedede
KS
809
810 pra->mapcount--;
b20ce5e0 811 }
b20ce5e0 812
33c3fc71
VD
813 if (referenced)
814 clear_page_idle(page);
815 if (test_and_clear_page_young(page))
816 referenced++;
817
9f32624b
JK
818 if (referenced) {
819 pra->referenced++;
820 pra->vm_flags |= vma->vm_flags;
1da177e4 821 }
34bbd704 822
9f32624b 823 if (!pra->mapcount)
e4b82222 824 return false; /* To break the loop */
9f32624b 825
e4b82222 826 return true;
1da177e4
LT
827}
828
9f32624b 829static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
1da177e4 830{
9f32624b
JK
831 struct page_referenced_arg *pra = arg;
832 struct mem_cgroup *memcg = pra->memcg;
1da177e4 833
9f32624b
JK
834 if (!mm_match_cgroup(vma->vm_mm, memcg))
835 return true;
1da177e4 836
9f32624b 837 return false;
1da177e4
LT
838}
839
840/**
841 * page_referenced - test if the page was referenced
842 * @page: the page to test
843 * @is_locked: caller holds lock on the page
72835c86 844 * @memcg: target memory cgroup
6fe6b7e3 845 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
1da177e4
LT
846 *
847 * Quick test_and_clear_referenced for all mappings to a page,
848 * returns the number of ptes which referenced the page.
849 */
6fe6b7e3
WF
850int page_referenced(struct page *page,
851 int is_locked,
72835c86 852 struct mem_cgroup *memcg,
6fe6b7e3 853 unsigned long *vm_flags)
1da177e4 854{
5ad64688 855 int we_locked = 0;
9f32624b 856 struct page_referenced_arg pra = {
b20ce5e0 857 .mapcount = total_mapcount(page),
9f32624b
JK
858 .memcg = memcg,
859 };
860 struct rmap_walk_control rwc = {
861 .rmap_one = page_referenced_one,
862 .arg = (void *)&pra,
863 .anon_lock = page_lock_anon_vma_read,
864 };
1da177e4 865
6fe6b7e3 866 *vm_flags = 0;
059d8442 867 if (!pra.mapcount)
9f32624b
JK
868 return 0;
869
870 if (!page_rmapping(page))
871 return 0;
872
873 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
874 we_locked = trylock_page(page);
875 if (!we_locked)
876 return 1;
1da177e4 877 }
9f32624b
JK
878
879 /*
880 * If we are reclaiming on behalf of a cgroup, skip
881 * counting on behalf of references from different
882 * cgroups
883 */
884 if (memcg) {
885 rwc.invalid_vma = invalid_page_referenced_vma;
886 }
887
c24f386c 888 rmap_walk(page, &rwc);
9f32624b
JK
889 *vm_flags = pra.vm_flags;
890
891 if (we_locked)
892 unlock_page(page);
893
894 return pra.referenced;
1da177e4
LT
895}
896
e4b82222 897static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
9853a407 898 unsigned long address, void *arg)
d08b3851 899{
f27176cf
KS
900 struct page_vma_mapped_walk pvmw = {
901 .page = page,
902 .vma = vma,
903 .address = address,
904 .flags = PVMW_SYNC,
905 };
ac46d4f3 906 struct mmu_notifier_range range;
9853a407 907 int *cleaned = arg;
d08b3851 908
369ea824
JG
909 /*
910 * We have to assume the worse case ie pmd for invalidation. Note that
911 * the page can not be free from this function.
912 */
7269f999
JG
913 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
914 0, vma, vma->vm_mm, address,
a50b854e 915 min(vma->vm_end, address + page_size(page)));
ac46d4f3 916 mmu_notifier_invalidate_range_start(&range);
369ea824 917
f27176cf
KS
918 while (page_vma_mapped_walk(&pvmw)) {
919 int ret = 0;
369ea824 920
1f18b296 921 address = pvmw.address;
f27176cf
KS
922 if (pvmw.pte) {
923 pte_t entry;
924 pte_t *pte = pvmw.pte;
925
926 if (!pte_dirty(*pte) && !pte_write(*pte))
927 continue;
928
785373b4
LT
929 flush_cache_page(vma, address, pte_pfn(*pte));
930 entry = ptep_clear_flush(vma, address, pte);
f27176cf
KS
931 entry = pte_wrprotect(entry);
932 entry = pte_mkclean(entry);
785373b4 933 set_pte_at(vma->vm_mm, address, pte, entry);
f27176cf
KS
934 ret = 1;
935 } else {
396bcc52 936#ifdef CONFIG_TRANSPARENT_HUGEPAGE
f27176cf
KS
937 pmd_t *pmd = pvmw.pmd;
938 pmd_t entry;
939
940 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
941 continue;
942
785373b4 943 flush_cache_page(vma, address, page_to_pfn(page));
024eee0e 944 entry = pmdp_invalidate(vma, address, pmd);
f27176cf
KS
945 entry = pmd_wrprotect(entry);
946 entry = pmd_mkclean(entry);
785373b4 947 set_pmd_at(vma->vm_mm, address, pmd, entry);
f27176cf
KS
948 ret = 1;
949#else
950 /* unexpected pmd-mapped page? */
951 WARN_ON_ONCE(1);
952#endif
953 }
d08b3851 954
0f10851e
JG
955 /*
956 * No need to call mmu_notifier_invalidate_range() as we are
957 * downgrading page table protection not changing it to point
958 * to a new page.
959 *
ad56b738 960 * See Documentation/vm/mmu_notifier.rst
0f10851e
JG
961 */
962 if (ret)
f27176cf 963 (*cleaned)++;
c2fda5fe 964 }
d08b3851 965
ac46d4f3 966 mmu_notifier_invalidate_range_end(&range);
369ea824 967
e4b82222 968 return true;
d08b3851
PZ
969}
970
9853a407 971static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
d08b3851 972{
9853a407 973 if (vma->vm_flags & VM_SHARED)
871beb8c 974 return false;
d08b3851 975
871beb8c 976 return true;
d08b3851
PZ
977}
978
979int page_mkclean(struct page *page)
980{
9853a407
JK
981 int cleaned = 0;
982 struct address_space *mapping;
983 struct rmap_walk_control rwc = {
984 .arg = (void *)&cleaned,
985 .rmap_one = page_mkclean_one,
986 .invalid_vma = invalid_mkclean_vma,
987 };
d08b3851
PZ
988
989 BUG_ON(!PageLocked(page));
990
9853a407
JK
991 if (!page_mapped(page))
992 return 0;
993
994 mapping = page_mapping(page);
995 if (!mapping)
996 return 0;
997
998 rmap_walk(page, &rwc);
d08b3851 999
9853a407 1000 return cleaned;
d08b3851 1001}
60b59bea 1002EXPORT_SYMBOL_GPL(page_mkclean);
d08b3851 1003
c44b6743
RR
1004/**
1005 * page_move_anon_rmap - move a page to our anon_vma
1006 * @page: the page to move to our anon_vma
1007 * @vma: the vma the page belongs to
c44b6743
RR
1008 *
1009 * When a page belongs exclusively to one process after a COW event,
1010 * that page can be moved into the anon_vma that belongs to just that
1011 * process, so the rmap code will not search the parent or sibling
1012 * processes.
1013 */
5a49973d 1014void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
c44b6743
RR
1015{
1016 struct anon_vma *anon_vma = vma->anon_vma;
1017
5a49973d
HD
1018 page = compound_head(page);
1019
309381fe 1020 VM_BUG_ON_PAGE(!PageLocked(page), page);
81d1b09c 1021 VM_BUG_ON_VMA(!anon_vma, vma);
c44b6743
RR
1022
1023 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
414e2fb8
VD
1024 /*
1025 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1026 * simultaneously, so a concurrent reader (eg page_referenced()'s
1027 * PageAnon()) will not see one without the other.
1028 */
1029 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
c44b6743
RR
1030}
1031
9617d95e 1032/**
4e1c1975 1033 * __page_set_anon_rmap - set up new anonymous rmap
451b9514 1034 * @page: Page or Hugepage to add to rmap
4e1c1975
AK
1035 * @vma: VM area to add page to.
1036 * @address: User virtual address of the mapping
e8a03feb 1037 * @exclusive: the page is exclusively owned by the current process
9617d95e
NP
1038 */
1039static void __page_set_anon_rmap(struct page *page,
e8a03feb 1040 struct vm_area_struct *vma, unsigned long address, int exclusive)
9617d95e 1041{
e8a03feb 1042 struct anon_vma *anon_vma = vma->anon_vma;
ea90002b 1043
e8a03feb 1044 BUG_ON(!anon_vma);
ea90002b 1045
4e1c1975
AK
1046 if (PageAnon(page))
1047 return;
1048
ea90002b 1049 /*
e8a03feb
RR
1050 * If the page isn't exclusively mapped into this vma,
1051 * we must use the _oldest_ possible anon_vma for the
1052 * page mapping!
ea90002b 1053 */
4e1c1975 1054 if (!exclusive)
288468c3 1055 anon_vma = anon_vma->root;
9617d95e 1056
9617d95e
NP
1057 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1058 page->mapping = (struct address_space *) anon_vma;
9617d95e 1059 page->index = linear_page_index(vma, address);
9617d95e
NP
1060}
1061
c97a9e10 1062/**
43d8eac4 1063 * __page_check_anon_rmap - sanity check anonymous rmap addition
c97a9e10
NP
1064 * @page: the page to add the mapping to
1065 * @vma: the vm area in which the mapping is added
1066 * @address: the user virtual address mapped
1067 */
1068static void __page_check_anon_rmap(struct page *page,
1069 struct vm_area_struct *vma, unsigned long address)
1070{
c97a9e10
NP
1071 /*
1072 * The page's anon-rmap details (mapping and index) are guaranteed to
1073 * be set up correctly at this point.
1074 *
1075 * We have exclusion against page_add_anon_rmap because the caller
1076 * always holds the page locked, except if called from page_dup_rmap,
1077 * in which case the page is already known to be setup.
1078 *
1079 * We have exclusion against page_add_new_anon_rmap because those pages
1080 * are initially only visible via the pagetables, and the pte is locked
1081 * over the call to page_add_new_anon_rmap.
1082 */
30c46382
YS
1083 VM_BUG_ON_PAGE(page_anon_vma(page)->root != vma->anon_vma->root, page);
1084 VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1085 page);
c97a9e10
NP
1086}
1087
1da177e4
LT
1088/**
1089 * page_add_anon_rmap - add pte mapping to an anonymous page
1090 * @page: the page to add the mapping to
1091 * @vma: the vm area in which the mapping is added
1092 * @address: the user virtual address mapped
d281ee61 1093 * @compound: charge the page as compound or small page
1da177e4 1094 *
5ad64688 1095 * The caller needs to hold the pte lock, and the page must be locked in
80e14822
HD
1096 * the anon_vma case: to serialize mapping,index checking after setting,
1097 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1098 * (but PageKsm is never downgraded to PageAnon).
1da177e4
LT
1099 */
1100void page_add_anon_rmap(struct page *page,
d281ee61 1101 struct vm_area_struct *vma, unsigned long address, bool compound)
ad8c2ee8 1102{
d281ee61 1103 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
ad8c2ee8
RR
1104}
1105
1106/*
1107 * Special version of the above for do_swap_page, which often runs
1108 * into pages that are exclusively owned by the current process.
1109 * Everybody else should continue to use page_add_anon_rmap above.
1110 */
1111void do_page_add_anon_rmap(struct page *page,
d281ee61 1112 struct vm_area_struct *vma, unsigned long address, int flags)
1da177e4 1113{
53f9263b
KS
1114 bool compound = flags & RMAP_COMPOUND;
1115 bool first;
1116
e9b61f19
KS
1117 if (compound) {
1118 atomic_t *mapcount;
53f9263b 1119 VM_BUG_ON_PAGE(!PageLocked(page), page);
e9b61f19
KS
1120 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1121 mapcount = compound_mapcount_ptr(page);
1122 first = atomic_inc_and_test(mapcount);
53f9263b
KS
1123 } else {
1124 first = atomic_inc_and_test(&page->_mapcount);
1125 }
1126
79134171 1127 if (first) {
d281ee61 1128 int nr = compound ? hpage_nr_pages(page) : 1;
bea04b07
JZ
1129 /*
1130 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1131 * these counters are not modified in interrupt context, and
1132 * pte lock(a spinlock) is held, which implies preemption
1133 * disabled.
1134 */
65c45377 1135 if (compound)
11fb9989 1136 __inc_node_page_state(page, NR_ANON_THPS);
4b9d0fab 1137 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
79134171 1138 }
5ad64688
HD
1139 if (unlikely(PageKsm(page)))
1140 return;
1141
309381fe 1142 VM_BUG_ON_PAGE(!PageLocked(page), page);
53f9263b 1143
5dbe0af4 1144 /* address might be in next vma when migration races vma_adjust */
5ad64688 1145 if (first)
d281ee61
KS
1146 __page_set_anon_rmap(page, vma, address,
1147 flags & RMAP_EXCLUSIVE);
69029cd5 1148 else
c97a9e10 1149 __page_check_anon_rmap(page, vma, address);
1da177e4
LT
1150}
1151
43d8eac4 1152/**
9617d95e
NP
1153 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1154 * @page: the page to add the mapping to
1155 * @vma: the vm area in which the mapping is added
1156 * @address: the user virtual address mapped
d281ee61 1157 * @compound: charge the page as compound or small page
9617d95e
NP
1158 *
1159 * Same as page_add_anon_rmap but must only be called on *new* pages.
1160 * This means the inc-and-test can be bypassed.
c97a9e10 1161 * Page does not have to be locked.
9617d95e
NP
1162 */
1163void page_add_new_anon_rmap(struct page *page,
d281ee61 1164 struct vm_area_struct *vma, unsigned long address, bool compound)
9617d95e 1165{
d281ee61
KS
1166 int nr = compound ? hpage_nr_pages(page) : 1;
1167
81d1b09c 1168 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
fa9949da 1169 __SetPageSwapBacked(page);
d281ee61
KS
1170 if (compound) {
1171 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
53f9263b
KS
1172 /* increment count (starts at -1) */
1173 atomic_set(compound_mapcount_ptr(page), 0);
47e29d32
JH
1174 if (hpage_pincount_available(page))
1175 atomic_set(compound_pincount_ptr(page), 0);
1176
11fb9989 1177 __inc_node_page_state(page, NR_ANON_THPS);
53f9263b
KS
1178 } else {
1179 /* Anon THP always mapped first with PMD */
1180 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1181 /* increment count (starts at -1) */
1182 atomic_set(&page->_mapcount, 0);
d281ee61 1183 }
4b9d0fab 1184 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
e8a03feb 1185 __page_set_anon_rmap(page, vma, address, 1);
9617d95e
NP
1186}
1187
1da177e4
LT
1188/**
1189 * page_add_file_rmap - add pte mapping to a file page
1190 * @page: the page to add the mapping to
e8b098fc 1191 * @compound: charge the page as compound or small page
1da177e4 1192 *
b8072f09 1193 * The caller needs to hold the pte lock.
1da177e4 1194 */
dd78fedd 1195void page_add_file_rmap(struct page *page, bool compound)
1da177e4 1196{
dd78fedd
KS
1197 int i, nr = 1;
1198
1199 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
62cccb8c 1200 lock_page_memcg(page);
dd78fedd
KS
1201 if (compound && PageTransHuge(page)) {
1202 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1203 if (atomic_inc_and_test(&page[i]._mapcount))
1204 nr++;
1205 }
1206 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1207 goto out;
99cb0dbd
SL
1208 if (PageSwapBacked(page))
1209 __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1210 else
1211 __inc_node_page_state(page, NR_FILE_PMDMAPPED);
dd78fedd 1212 } else {
c8efc390
KS
1213 if (PageTransCompound(page) && page_mapping(page)) {
1214 VM_WARN_ON_ONCE(!PageLocked(page));
1215
9a73f61b
KS
1216 SetPageDoubleMap(compound_head(page));
1217 if (PageMlocked(page))
1218 clear_page_mlock(compound_head(page));
1219 }
dd78fedd
KS
1220 if (!atomic_inc_and_test(&page->_mapcount))
1221 goto out;
d69b042f 1222 }
00f3ca2c 1223 __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
dd78fedd 1224out:
62cccb8c 1225 unlock_page_memcg(page);
1da177e4
LT
1226}
1227
dd78fedd 1228static void page_remove_file_rmap(struct page *page, bool compound)
8186eb6a 1229{
dd78fedd
KS
1230 int i, nr = 1;
1231
57dea93a 1232 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
62cccb8c 1233 lock_page_memcg(page);
8186eb6a 1234
53f9263b
KS
1235 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1236 if (unlikely(PageHuge(page))) {
1237 /* hugetlb pages are always mapped with pmds */
1238 atomic_dec(compound_mapcount_ptr(page));
8186eb6a 1239 goto out;
53f9263b 1240 }
8186eb6a 1241
53f9263b 1242 /* page still mapped by someone else? */
dd78fedd
KS
1243 if (compound && PageTransHuge(page)) {
1244 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1245 if (atomic_add_negative(-1, &page[i]._mapcount))
1246 nr++;
1247 }
1248 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1249 goto out;
99cb0dbd
SL
1250 if (PageSwapBacked(page))
1251 __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1252 else
1253 __dec_node_page_state(page, NR_FILE_PMDMAPPED);
dd78fedd
KS
1254 } else {
1255 if (!atomic_add_negative(-1, &page->_mapcount))
1256 goto out;
1257 }
8186eb6a
JW
1258
1259 /*
00f3ca2c 1260 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
8186eb6a
JW
1261 * these counters are not modified in interrupt context, and
1262 * pte lock(a spinlock) is held, which implies preemption disabled.
1263 */
00f3ca2c 1264 __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
8186eb6a
JW
1265
1266 if (unlikely(PageMlocked(page)))
1267 clear_page_mlock(page);
1268out:
62cccb8c 1269 unlock_page_memcg(page);
8186eb6a
JW
1270}
1271
53f9263b
KS
1272static void page_remove_anon_compound_rmap(struct page *page)
1273{
1274 int i, nr;
1275
1276 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1277 return;
1278
1279 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1280 if (unlikely(PageHuge(page)))
1281 return;
1282
1283 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1284 return;
1285
11fb9989 1286 __dec_node_page_state(page, NR_ANON_THPS);
53f9263b
KS
1287
1288 if (TestClearPageDoubleMap(page)) {
1289 /*
1290 * Subpages can be mapped with PTEs too. Check how many of
f1fe80d4 1291 * them are still mapped.
53f9263b
KS
1292 */
1293 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1294 if (atomic_add_negative(-1, &page[i]._mapcount))
1295 nr++;
1296 }
f1fe80d4
KS
1297
1298 /*
1299 * Queue the page for deferred split if at least one small
1300 * page of the compound page is unmapped, but at least one
1301 * small page is still mapped.
1302 */
1303 if (nr && nr < HPAGE_PMD_NR)
1304 deferred_split_huge_page(page);
53f9263b
KS
1305 } else {
1306 nr = HPAGE_PMD_NR;
1307 }
1308
e90309c9
KS
1309 if (unlikely(PageMlocked(page)))
1310 clear_page_mlock(page);
1311
f1fe80d4 1312 if (nr)
4b9d0fab 1313 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
53f9263b
KS
1314}
1315
1da177e4
LT
1316/**
1317 * page_remove_rmap - take down pte mapping from a page
d281ee61
KS
1318 * @page: page to remove mapping from
1319 * @compound: uncharge the page as compound or small page
1da177e4 1320 *
b8072f09 1321 * The caller needs to hold the pte lock.
1da177e4 1322 */
d281ee61 1323void page_remove_rmap(struct page *page, bool compound)
1da177e4 1324{
dd78fedd
KS
1325 if (!PageAnon(page))
1326 return page_remove_file_rmap(page, compound);
89c06bd5 1327
53f9263b
KS
1328 if (compound)
1329 return page_remove_anon_compound_rmap(page);
1330
b904dcfe
KM
1331 /* page still mapped by someone else? */
1332 if (!atomic_add_negative(-1, &page->_mapcount))
8186eb6a
JW
1333 return;
1334
0fe6e20b 1335 /*
bea04b07
JZ
1336 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1337 * these counters are not modified in interrupt context, and
bea04b07 1338 * pte lock(a spinlock) is held, which implies preemption disabled.
0fe6e20b 1339 */
4b9d0fab 1340 __dec_node_page_state(page, NR_ANON_MAPPED);
8186eb6a 1341
e6c509f8
HD
1342 if (unlikely(PageMlocked(page)))
1343 clear_page_mlock(page);
8186eb6a 1344
9a982250
KS
1345 if (PageTransCompound(page))
1346 deferred_split_huge_page(compound_head(page));
1347
b904dcfe
KM
1348 /*
1349 * It would be tidy to reset the PageAnon mapping here,
1350 * but that might overwrite a racing page_add_anon_rmap
1351 * which increments mapcount after us but sets mapping
2d4894b5 1352 * before us: so leave the reset to free_unref_page,
b904dcfe
KM
1353 * and remember that it's only reliable while mapped.
1354 * Leaving it set also helps swapoff to reinstate ptes
1355 * faster for those pages still in swapcache.
1356 */
1da177e4
LT
1357}
1358
1359/*
52629506 1360 * @arg: enum ttu_flags will be passed to this argument
1da177e4 1361 */
e4b82222 1362static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
52629506 1363 unsigned long address, void *arg)
1da177e4
LT
1364{
1365 struct mm_struct *mm = vma->vm_mm;
c7ab0d2f
KS
1366 struct page_vma_mapped_walk pvmw = {
1367 .page = page,
1368 .vma = vma,
1369 .address = address,
1370 };
1da177e4 1371 pte_t pteval;
c7ab0d2f 1372 struct page *subpage;
785373b4 1373 bool ret = true;
ac46d4f3 1374 struct mmu_notifier_range range;
4708f318 1375 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1da177e4 1376
b87537d9
HD
1377 /* munlock has nothing to gain from examining un-locked vmas */
1378 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
e4b82222 1379 return true;
b87537d9 1380
a5430dda
JG
1381 if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1382 is_zone_device_page(page) && !is_device_private_page(page))
1383 return true;
1384
fec89c10
KS
1385 if (flags & TTU_SPLIT_HUGE_PMD) {
1386 split_huge_pmd_address(vma, address,
b5ff8161 1387 flags & TTU_SPLIT_FREEZE, page);
fec89c10
KS
1388 }
1389
369ea824 1390 /*
017b1660
MK
1391 * For THP, we have to assume the worse case ie pmd for invalidation.
1392 * For hugetlb, it could be much worse if we need to do pud
1393 * invalidation in the case of pmd sharing.
1394 *
1395 * Note that the page can not be free in this function as call of
1396 * try_to_unmap() must hold a reference on the page.
369ea824 1397 */
7269f999 1398 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
6f4f13e8 1399 address,
a50b854e 1400 min(vma->vm_end, address + page_size(page)));
017b1660
MK
1401 if (PageHuge(page)) {
1402 /*
1403 * If sharing is possible, start and end will be adjusted
1404 * accordingly.
c0d0381a
MK
1405 *
1406 * If called for a huge page, caller must hold i_mmap_rwsem
1407 * in write mode as it is possible to call huge_pmd_unshare.
017b1660 1408 */
ac46d4f3
JG
1409 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1410 &range.end);
017b1660 1411 }
ac46d4f3 1412 mmu_notifier_invalidate_range_start(&range);
369ea824 1413
c7ab0d2f 1414 while (page_vma_mapped_walk(&pvmw)) {
616b8371
ZY
1415#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1416 /* PMD-mapped THP migration entry */
1417 if (!pvmw.pte && (flags & TTU_MIGRATION)) {
1418 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
1419
616b8371
ZY
1420 set_pmd_migration_entry(&pvmw, page);
1421 continue;
1422 }
1423#endif
1424
c7ab0d2f
KS
1425 /*
1426 * If the page is mlock()d, we cannot swap it out.
1427 * If it's recently referenced (perhaps page_referenced
1428 * skipped over this mm) then we should reactivate it.
1429 */
1430 if (!(flags & TTU_IGNORE_MLOCK)) {
1431 if (vma->vm_flags & VM_LOCKED) {
1432 /* PTE-mapped THP are never mlocked */
1433 if (!PageTransCompound(page)) {
1434 /*
1435 * Holding pte lock, we do *not* need
1436 * mmap_sem here
1437 */
1438 mlock_vma_page(page);
1439 }
e4b82222 1440 ret = false;
c7ab0d2f
KS
1441 page_vma_mapped_walk_done(&pvmw);
1442 break;
9a73f61b 1443 }
c7ab0d2f
KS
1444 if (flags & TTU_MUNLOCK)
1445 continue;
b87537d9 1446 }
c7ab0d2f 1447
8346242a
KS
1448 /* Unexpected PMD-mapped THP? */
1449 VM_BUG_ON_PAGE(!pvmw.pte, page);
1450
1451 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
785373b4
LT
1452 address = pvmw.address;
1453
017b1660 1454 if (PageHuge(page)) {
c0d0381a
MK
1455 /*
1456 * To call huge_pmd_unshare, i_mmap_rwsem must be
1457 * held in write mode. Caller needs to explicitly
1458 * do this outside rmap routines.
1459 */
1460 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
017b1660
MK
1461 if (huge_pmd_unshare(mm, &address, pvmw.pte)) {
1462 /*
1463 * huge_pmd_unshare unmapped an entire PMD
1464 * page. There is no way of knowing exactly
1465 * which PMDs may be cached for this mm, so
1466 * we must flush them all. start/end were
1467 * already adjusted above to cover this range.
1468 */
ac46d4f3
JG
1469 flush_cache_range(vma, range.start, range.end);
1470 flush_tlb_range(vma, range.start, range.end);
1471 mmu_notifier_invalidate_range(mm, range.start,
1472 range.end);
017b1660
MK
1473
1474 /*
1475 * The ref count of the PMD page was dropped
1476 * which is part of the way map counting
1477 * is done for shared PMDs. Return 'true'
1478 * here. When there is no other sharing,
1479 * huge_pmd_unshare returns false and we will
1480 * unmap the actual page and drop map count
1481 * to zero.
1482 */
1483 page_vma_mapped_walk_done(&pvmw);
1484 break;
1485 }
1486 }
8346242a 1487
a5430dda
JG
1488 if (IS_ENABLED(CONFIG_MIGRATION) &&
1489 (flags & TTU_MIGRATION) &&
1490 is_zone_device_page(page)) {
1491 swp_entry_t entry;
1492 pte_t swp_pte;
1493
1494 pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte);
1495
1496 /*
1497 * Store the pfn of the page in a special migration
1498 * pte. do_swap_page() will wait until the migration
1499 * pte is removed and then restart fault handling.
1500 */
1501 entry = make_migration_entry(page, 0);
1502 swp_pte = swp_entry_to_pte(entry);
1503 if (pte_soft_dirty(pteval))
1504 swp_pte = pte_swp_mksoft_dirty(swp_pte);
f45ec5ff
PX
1505 if (pte_uffd_wp(pteval))
1506 swp_pte = pte_swp_mkuffd_wp(swp_pte);
a5430dda 1507 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
0f10851e
JG
1508 /*
1509 * No need to invalidate here it will synchronize on
1510 * against the special swap migration pte.
1de13ee5
RC
1511 *
1512 * The assignment to subpage above was computed from a
1513 * swap PTE which results in an invalid pointer.
1514 * Since only PAGE_SIZE pages can currently be
1515 * migrated, just set it to page. This will need to be
1516 * changed when hugepage migrations to device private
1517 * memory are supported.
0f10851e 1518 */
1de13ee5 1519 subpage = page;
a5430dda
JG
1520 goto discard;
1521 }
1522
c7ab0d2f 1523 if (!(flags & TTU_IGNORE_ACCESS)) {
785373b4 1524 if (ptep_clear_flush_young_notify(vma, address,
c7ab0d2f 1525 pvmw.pte)) {
e4b82222 1526 ret = false;
c7ab0d2f
KS
1527 page_vma_mapped_walk_done(&pvmw);
1528 break;
1529 }
b291f000 1530 }
1da177e4 1531
c7ab0d2f 1532 /* Nuke the page table entry. */
785373b4 1533 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
c7ab0d2f
KS
1534 if (should_defer_flush(mm, flags)) {
1535 /*
1536 * We clear the PTE but do not flush so potentially
1537 * a remote CPU could still be writing to the page.
1538 * If the entry was previously clean then the
1539 * architecture must guarantee that a clear->dirty
1540 * transition on a cached TLB entry is written through
1541 * and traps if the PTE is unmapped.
1542 */
785373b4 1543 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
c7ab0d2f
KS
1544
1545 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1546 } else {
785373b4 1547 pteval = ptep_clear_flush(vma, address, pvmw.pte);
c7ab0d2f 1548 }
72b252ae 1549
c7ab0d2f
KS
1550 /* Move the dirty bit to the page. Now the pte is gone. */
1551 if (pte_dirty(pteval))
1552 set_page_dirty(page);
1da177e4 1553
c7ab0d2f
KS
1554 /* Update high watermark before we lower rss */
1555 update_hiwater_rss(mm);
1da177e4 1556
c7ab0d2f 1557 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
5fd27b8e 1558 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
c7ab0d2f 1559 if (PageHuge(page)) {
d8c6546b 1560 hugetlb_count_sub(compound_nr(page), mm);
785373b4 1561 set_huge_swap_pte_at(mm, address,
5fd27b8e
PA
1562 pvmw.pte, pteval,
1563 vma_mmu_pagesize(vma));
c7ab0d2f
KS
1564 } else {
1565 dec_mm_counter(mm, mm_counter(page));
785373b4 1566 set_pte_at(mm, address, pvmw.pte, pteval);
c7ab0d2f 1567 }
365e9c87 1568
bce73e48 1569 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
c7ab0d2f
KS
1570 /*
1571 * The guest indicated that the page content is of no
1572 * interest anymore. Simply discard the pte, vmscan
1573 * will take care of the rest.
bce73e48
CB
1574 * A future reference will then fault in a new zero
1575 * page. When userfaultfd is active, we must not drop
1576 * this page though, as its main user (postcopy
1577 * migration) will not expect userfaults on already
1578 * copied pages.
c7ab0d2f 1579 */
eca56ff9 1580 dec_mm_counter(mm, mm_counter(page));
0f10851e
JG
1581 /* We have to invalidate as we cleared the pte */
1582 mmu_notifier_invalidate_range(mm, address,
1583 address + PAGE_SIZE);
c7ab0d2f 1584 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
b5ff8161 1585 (flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) {
c7ab0d2f
KS
1586 swp_entry_t entry;
1587 pte_t swp_pte;
ca827d55
KA
1588
1589 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1590 set_pte_at(mm, address, pvmw.pte, pteval);
1591 ret = false;
1592 page_vma_mapped_walk_done(&pvmw);
1593 break;
1594 }
1595
c7ab0d2f
KS
1596 /*
1597 * Store the pfn of the page in a special migration
1598 * pte. do_swap_page() will wait until the migration
1599 * pte is removed and then restart fault handling.
1600 */
1601 entry = make_migration_entry(subpage,
1602 pte_write(pteval));
1603 swp_pte = swp_entry_to_pte(entry);
1604 if (pte_soft_dirty(pteval))
1605 swp_pte = pte_swp_mksoft_dirty(swp_pte);
f45ec5ff
PX
1606 if (pte_uffd_wp(pteval))
1607 swp_pte = pte_swp_mkuffd_wp(swp_pte);
785373b4 1608 set_pte_at(mm, address, pvmw.pte, swp_pte);
0f10851e
JG
1609 /*
1610 * No need to invalidate here it will synchronize on
1611 * against the special swap migration pte.
1612 */
c7ab0d2f
KS
1613 } else if (PageAnon(page)) {
1614 swp_entry_t entry = { .val = page_private(subpage) };
1615 pte_t swp_pte;
1616 /*
1617 * Store the swap location in the pte.
1618 * See handle_pte_fault() ...
1619 */
eb94a878
MK
1620 if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1621 WARN_ON_ONCE(1);
83612a94 1622 ret = false;
369ea824 1623 /* We have to invalidate as we cleared the pte */
0f10851e
JG
1624 mmu_notifier_invalidate_range(mm, address,
1625 address + PAGE_SIZE);
eb94a878
MK
1626 page_vma_mapped_walk_done(&pvmw);
1627 break;
1628 }
c7ab0d2f 1629
802a3a92
SL
1630 /* MADV_FREE page check */
1631 if (!PageSwapBacked(page)) {
1632 if (!PageDirty(page)) {
0f10851e
JG
1633 /* Invalidate as we cleared the pte */
1634 mmu_notifier_invalidate_range(mm,
1635 address, address + PAGE_SIZE);
802a3a92
SL
1636 dec_mm_counter(mm, MM_ANONPAGES);
1637 goto discard;
1638 }
1639
1640 /*
1641 * If the page was redirtied, it cannot be
1642 * discarded. Remap the page to page table.
1643 */
785373b4 1644 set_pte_at(mm, address, pvmw.pte, pteval);
18863d3a 1645 SetPageSwapBacked(page);
e4b82222 1646 ret = false;
802a3a92
SL
1647 page_vma_mapped_walk_done(&pvmw);
1648 break;
c7ab0d2f 1649 }
854e9ed0 1650
c7ab0d2f 1651 if (swap_duplicate(entry) < 0) {
785373b4 1652 set_pte_at(mm, address, pvmw.pte, pteval);
e4b82222 1653 ret = false;
c7ab0d2f
KS
1654 page_vma_mapped_walk_done(&pvmw);
1655 break;
1656 }
ca827d55
KA
1657 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1658 set_pte_at(mm, address, pvmw.pte, pteval);
1659 ret = false;
1660 page_vma_mapped_walk_done(&pvmw);
1661 break;
1662 }
c7ab0d2f
KS
1663 if (list_empty(&mm->mmlist)) {
1664 spin_lock(&mmlist_lock);
1665 if (list_empty(&mm->mmlist))
1666 list_add(&mm->mmlist, &init_mm.mmlist);
1667 spin_unlock(&mmlist_lock);
1668 }
854e9ed0 1669 dec_mm_counter(mm, MM_ANONPAGES);
c7ab0d2f
KS
1670 inc_mm_counter(mm, MM_SWAPENTS);
1671 swp_pte = swp_entry_to_pte(entry);
1672 if (pte_soft_dirty(pteval))
1673 swp_pte = pte_swp_mksoft_dirty(swp_pte);
f45ec5ff
PX
1674 if (pte_uffd_wp(pteval))
1675 swp_pte = pte_swp_mkuffd_wp(swp_pte);
785373b4 1676 set_pte_at(mm, address, pvmw.pte, swp_pte);
0f10851e
JG
1677 /* Invalidate as we cleared the pte */
1678 mmu_notifier_invalidate_range(mm, address,
1679 address + PAGE_SIZE);
1680 } else {
1681 /*
906f9cdf
HD
1682 * This is a locked file-backed page, thus it cannot
1683 * be removed from the page cache and replaced by a new
1684 * page before mmu_notifier_invalidate_range_end, so no
0f10851e
JG
1685 * concurrent thread might update its page table to
1686 * point at new page while a device still is using this
1687 * page.
1688 *
ad56b738 1689 * See Documentation/vm/mmu_notifier.rst
0f10851e 1690 */
c7ab0d2f 1691 dec_mm_counter(mm, mm_counter_file(page));
0f10851e 1692 }
854e9ed0 1693discard:
0f10851e
JG
1694 /*
1695 * No need to call mmu_notifier_invalidate_range() it has be
1696 * done above for all cases requiring it to happen under page
1697 * table lock before mmu_notifier_invalidate_range_end()
1698 *
ad56b738 1699 * See Documentation/vm/mmu_notifier.rst
0f10851e 1700 */
c7ab0d2f
KS
1701 page_remove_rmap(subpage, PageHuge(page));
1702 put_page(page);
c7ab0d2f 1703 }
369ea824 1704
ac46d4f3 1705 mmu_notifier_invalidate_range_end(&range);
369ea824 1706
caed0f48 1707 return ret;
1da177e4
LT
1708}
1709
52629506
JK
1710static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1711{
222100ee 1712 return vma_is_temporary_stack(vma);
52629506
JK
1713}
1714
2a52bcbc 1715static int page_mapcount_is_zero(struct page *page)
52629506 1716{
c7ab0d2f 1717 return !total_mapcount(page);
2a52bcbc 1718}
52629506 1719
1da177e4
LT
1720/**
1721 * try_to_unmap - try to remove all page table mappings to a page
1722 * @page: the page to get unmapped
14fa31b8 1723 * @flags: action and flags
1da177e4
LT
1724 *
1725 * Tries to remove all the page table entries which are mapping this
1726 * page, used in the pageout path. Caller must hold the page lock.
1da177e4 1727 *
666e5a40 1728 * If unmap is successful, return true. Otherwise, false.
1da177e4 1729 */
666e5a40 1730bool try_to_unmap(struct page *page, enum ttu_flags flags)
1da177e4 1731{
52629506
JK
1732 struct rmap_walk_control rwc = {
1733 .rmap_one = try_to_unmap_one,
802a3a92 1734 .arg = (void *)flags,
2a52bcbc 1735 .done = page_mapcount_is_zero,
52629506
JK
1736 .anon_lock = page_lock_anon_vma_read,
1737 };
1da177e4 1738
52629506
JK
1739 /*
1740 * During exec, a temporary VMA is setup and later moved.
1741 * The VMA is moved under the anon_vma lock but not the
1742 * page tables leading to a race where migration cannot
1743 * find the migration ptes. Rather than increasing the
1744 * locking requirements of exec(), migration skips
1745 * temporary VMAs until after exec() completes.
1746 */
b5ff8161
NH
1747 if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))
1748 && !PageKsm(page) && PageAnon(page))
52629506
JK
1749 rwc.invalid_vma = invalid_migration_vma;
1750
2a52bcbc 1751 if (flags & TTU_RMAP_LOCKED)
33fc80e2 1752 rmap_walk_locked(page, &rwc);
2a52bcbc 1753 else
33fc80e2 1754 rmap_walk(page, &rwc);
52629506 1755
666e5a40 1756 return !page_mapcount(page) ? true : false;
1da177e4 1757}
81b4082d 1758
2a52bcbc
KS
1759static int page_not_mapped(struct page *page)
1760{
1761 return !page_mapped(page);
1762};
1763
b291f000
NP
1764/**
1765 * try_to_munlock - try to munlock a page
1766 * @page: the page to be munlocked
1767 *
1768 * Called from munlock code. Checks all of the VMAs mapping the page
1769 * to make sure nobody else has this page mlocked. The page will be
1770 * returned with PG_mlocked cleared if no other vmas have it mlocked.
b291f000 1771 */
854e9ed0 1772
192d7232
MK
1773void try_to_munlock(struct page *page)
1774{
e8351ac9
JK
1775 struct rmap_walk_control rwc = {
1776 .rmap_one = try_to_unmap_one,
802a3a92 1777 .arg = (void *)TTU_MUNLOCK,
e8351ac9 1778 .done = page_not_mapped,
e8351ac9
JK
1779 .anon_lock = page_lock_anon_vma_read,
1780
1781 };
1782
309381fe 1783 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
192d7232 1784 VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
b291f000 1785
192d7232 1786 rmap_walk(page, &rwc);
b291f000 1787}
e9995ef9 1788
01d8b20d 1789void __put_anon_vma(struct anon_vma *anon_vma)
76545066 1790{
01d8b20d 1791 struct anon_vma *root = anon_vma->root;
76545066 1792
624483f3 1793 anon_vma_free(anon_vma);
01d8b20d
PZ
1794 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1795 anon_vma_free(root);
76545066 1796}
76545066 1797
0dd1c7bb
JK
1798static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1799 struct rmap_walk_control *rwc)
faecd8dd
JK
1800{
1801 struct anon_vma *anon_vma;
1802
0dd1c7bb
JK
1803 if (rwc->anon_lock)
1804 return rwc->anon_lock(page);
1805
faecd8dd
JK
1806 /*
1807 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1808 * because that depends on page_mapped(); but not all its usages
1809 * are holding mmap_sem. Users without mmap_sem are required to
1810 * take a reference count to prevent the anon_vma disappearing
1811 */
1812 anon_vma = page_anon_vma(page);
1813 if (!anon_vma)
1814 return NULL;
1815
1816 anon_vma_lock_read(anon_vma);
1817 return anon_vma;
1818}
1819
e9995ef9 1820/*
e8351ac9
JK
1821 * rmap_walk_anon - do something to anonymous page using the object-based
1822 * rmap method
1823 * @page: the page to be handled
1824 * @rwc: control variable according to each walk type
1825 *
1826 * Find all the mappings of a page using the mapping pointer and the vma chains
1827 * contained in the anon_vma struct it points to.
1828 *
1829 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1830 * where the page was found will be held for write. So, we won't recheck
1831 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1832 * LOCKED.
e9995ef9 1833 */
1df631ae 1834static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
b9773199 1835 bool locked)
e9995ef9
HD
1836{
1837 struct anon_vma *anon_vma;
a8fa41ad 1838 pgoff_t pgoff_start, pgoff_end;
5beb4930 1839 struct anon_vma_chain *avc;
e9995ef9 1840
b9773199
KS
1841 if (locked) {
1842 anon_vma = page_anon_vma(page);
1843 /* anon_vma disappear under us? */
1844 VM_BUG_ON_PAGE(!anon_vma, page);
1845 } else {
1846 anon_vma = rmap_walk_anon_lock(page, rwc);
1847 }
e9995ef9 1848 if (!anon_vma)
1df631ae 1849 return;
faecd8dd 1850
a8fa41ad
KS
1851 pgoff_start = page_to_pgoff(page);
1852 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1853 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1854 pgoff_start, pgoff_end) {
5beb4930 1855 struct vm_area_struct *vma = avc->vma;
e9995ef9 1856 unsigned long address = vma_address(page, vma);
0dd1c7bb 1857
ad12695f
AA
1858 cond_resched();
1859
0dd1c7bb
JK
1860 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1861 continue;
1862
e4b82222 1863 if (!rwc->rmap_one(page, vma, address, rwc->arg))
e9995ef9 1864 break;
0dd1c7bb
JK
1865 if (rwc->done && rwc->done(page))
1866 break;
e9995ef9 1867 }
b9773199
KS
1868
1869 if (!locked)
1870 anon_vma_unlock_read(anon_vma);
e9995ef9
HD
1871}
1872
e8351ac9
JK
1873/*
1874 * rmap_walk_file - do something to file page using the object-based rmap method
1875 * @page: the page to be handled
1876 * @rwc: control variable according to each walk type
1877 *
1878 * Find all the mappings of a page using the mapping pointer and the vma chains
1879 * contained in the address_space struct it points to.
1880 *
1881 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1882 * where the page was found will be held for write. So, we won't recheck
1883 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1884 * LOCKED.
1885 */
1df631ae 1886static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
b9773199 1887 bool locked)
e9995ef9 1888{
b9773199 1889 struct address_space *mapping = page_mapping(page);
a8fa41ad 1890 pgoff_t pgoff_start, pgoff_end;
e9995ef9 1891 struct vm_area_struct *vma;
e9995ef9 1892
9f32624b
JK
1893 /*
1894 * The page lock not only makes sure that page->mapping cannot
1895 * suddenly be NULLified by truncation, it makes sure that the
1896 * structure at mapping cannot be freed and reused yet,
c8c06efa 1897 * so we can safely take mapping->i_mmap_rwsem.
9f32624b 1898 */
81d1b09c 1899 VM_BUG_ON_PAGE(!PageLocked(page), page);
9f32624b 1900
e9995ef9 1901 if (!mapping)
1df631ae 1902 return;
3dec0ba0 1903
a8fa41ad
KS
1904 pgoff_start = page_to_pgoff(page);
1905 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
b9773199
KS
1906 if (!locked)
1907 i_mmap_lock_read(mapping);
a8fa41ad
KS
1908 vma_interval_tree_foreach(vma, &mapping->i_mmap,
1909 pgoff_start, pgoff_end) {
e9995ef9 1910 unsigned long address = vma_address(page, vma);
0dd1c7bb 1911
ad12695f
AA
1912 cond_resched();
1913
0dd1c7bb
JK
1914 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1915 continue;
1916
e4b82222 1917 if (!rwc->rmap_one(page, vma, address, rwc->arg))
0dd1c7bb
JK
1918 goto done;
1919 if (rwc->done && rwc->done(page))
1920 goto done;
e9995ef9 1921 }
0dd1c7bb 1922
0dd1c7bb 1923done:
b9773199
KS
1924 if (!locked)
1925 i_mmap_unlock_read(mapping);
e9995ef9
HD
1926}
1927
1df631ae 1928void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
e9995ef9 1929{
e9995ef9 1930 if (unlikely(PageKsm(page)))
1df631ae 1931 rmap_walk_ksm(page, rwc);
e9995ef9 1932 else if (PageAnon(page))
1df631ae 1933 rmap_walk_anon(page, rwc, false);
b9773199 1934 else
1df631ae 1935 rmap_walk_file(page, rwc, false);
b9773199
KS
1936}
1937
1938/* Like rmap_walk, but caller holds relevant rmap lock */
1df631ae 1939void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
b9773199
KS
1940{
1941 /* no ksm support for now */
1942 VM_BUG_ON_PAGE(PageKsm(page), page);
1943 if (PageAnon(page))
1df631ae 1944 rmap_walk_anon(page, rwc, true);
e9995ef9 1945 else
1df631ae 1946 rmap_walk_file(page, rwc, true);
e9995ef9 1947}
0fe6e20b 1948
e3390f67 1949#ifdef CONFIG_HUGETLB_PAGE
0fe6e20b 1950/*
451b9514 1951 * The following two functions are for anonymous (private mapped) hugepages.
0fe6e20b
NH
1952 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1953 * and no lru code, because we handle hugepages differently from common pages.
1954 */
0fe6e20b
NH
1955void hugepage_add_anon_rmap(struct page *page,
1956 struct vm_area_struct *vma, unsigned long address)
1957{
1958 struct anon_vma *anon_vma = vma->anon_vma;
1959 int first;
a850ea30
NH
1960
1961 BUG_ON(!PageLocked(page));
0fe6e20b 1962 BUG_ON(!anon_vma);
5dbe0af4 1963 /* address might be in next vma when migration races vma_adjust */
53f9263b 1964 first = atomic_inc_and_test(compound_mapcount_ptr(page));
0fe6e20b 1965 if (first)
451b9514 1966 __page_set_anon_rmap(page, vma, address, 0);
0fe6e20b
NH
1967}
1968
1969void hugepage_add_new_anon_rmap(struct page *page,
1970 struct vm_area_struct *vma, unsigned long address)
1971{
1972 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
53f9263b 1973 atomic_set(compound_mapcount_ptr(page), 0);
47e29d32
JH
1974 if (hpage_pincount_available(page))
1975 atomic_set(compound_pincount_ptr(page), 0);
1976
451b9514 1977 __page_set_anon_rmap(page, vma, address, 1);
0fe6e20b 1978}
e3390f67 1979#endif /* CONFIG_HUGETLB_PAGE */