blob: 8774791e28099be82b7f4b7062c45600e8aed4ff [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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
Hugh Dickins98f32602009-05-21 20:33:58 +010017 * Contributions by Hugh Dickins 2003, 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -070018 */
19
20/*
21 * Lock ordering in mm:
22 *
Jes Sorensen1b1dcc12006-01-09 15:59:24 -080023 * inode->i_mutex (while writing or truncating, not reading or faulting)
Nick Piggin82591e62006-10-19 23:29:10 -070024 * mm->mmap_sem
25 * page->flags PG_locked (lock_page)
Kirill A. Shutemov88f306b2016-01-15 16:57:31 -080026 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
27 * mapping->i_mmap_rwsem
28 * anon_vma->rwsem
29 * mm->page_table_lock or pte_lock
Mel Gormana52633d2016-07-28 15:45:28 -070030 * zone_lru_lock (in mark_page_accessed, isolate_lru_page)
Kirill A. Shutemov88f306b2016-01-15 16:57:31 -080031 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
34 * mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
35 * mapping->tree_lock (widely used)
36 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
37 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
38 * sb_lock (within inode_lock in fs/fs-writeback.c)
39 * mapping->tree_lock (widely used, in set_page_dirty,
40 * in arch-dependent flush_dcache_mmap_lock,
41 * within bdi.wb->list_lock in __sync_single_inode)
Andi Kleen6a460792009-09-16 11:50:15 +020042 *
Ingo Molnar5a505082012-12-02 19:56:46 +000043 * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon)
Peter Zijlstra9b679322011-06-27 16:18:09 -070044 * ->tasklist_lock
Andi Kleen6a460792009-09-16 11:50:15 +020045 * pte map lock
Linus Torvalds1da177e2005-04-16 15:20:36 -070046 */
47
48#include <linux/mm.h>
49#include <linux/pagemap.h>
50#include <linux/swap.h>
51#include <linux/swapops.h>
52#include <linux/slab.h>
53#include <linux/init.h>
Hugh Dickins5ad64682009-12-14 17:59:24 -080054#include <linux/ksm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070055#include <linux/rmap.h>
56#include <linux/rcupdate.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040057#include <linux/export.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080058#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070059#include <linux/mmu_notifier.h>
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -080060#include <linux/migrate.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090061#include <linux/hugetlb.h>
Jan Karaef5d4372012-10-25 13:37:31 -070062#include <linux/backing-dev.h>
Vladimir Davydov33c3fc72015-09-09 15:35:45 -070063#include <linux/page_idle.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070064
65#include <asm/tlbflush.h>
66
Mel Gorman72b252a2015-09-04 15:47:32 -070067#include <trace/events/tlb.h>
68
Nick Pigginb291f002008-10-18 20:26:44 -070069#include "internal.h"
70
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070071static struct kmem_cache *anon_vma_cachep;
Rik van Riel5beb4932010-03-05 13:42:07 -080072static struct kmem_cache *anon_vma_chain_cachep;
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070073
74static inline struct anon_vma *anon_vma_alloc(void)
75{
Peter Zijlstra01d8b202011-03-22 16:32:49 -070076 struct anon_vma *anon_vma;
77
78 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
79 if (anon_vma) {
80 atomic_set(&anon_vma->refcount, 1);
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -080081 anon_vma->degree = 1; /* Reference for first vma */
82 anon_vma->parent = anon_vma;
Peter Zijlstra01d8b202011-03-22 16:32:49 -070083 /*
84 * Initialise the anon_vma root to point to itself. If called
85 * from fork, the root will be reset to the parents anon_vma.
86 */
87 anon_vma->root = anon_vma;
88 }
89
90 return anon_vma;
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070091}
92
Peter Zijlstra01d8b202011-03-22 16:32:49 -070093static inline void anon_vma_free(struct anon_vma *anon_vma)
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070094{
Peter Zijlstra01d8b202011-03-22 16:32:49 -070095 VM_BUG_ON(atomic_read(&anon_vma->refcount));
Peter Zijlstra88c22082011-05-24 17:12:13 -070096
97 /*
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +000098 * Synchronize against page_lock_anon_vma_read() such that
Peter Zijlstra88c22082011-05-24 17:12:13 -070099 * we can safely hold the lock without the anon_vma getting
100 * freed.
101 *
102 * Relies on the full mb implied by the atomic_dec_and_test() from
103 * put_anon_vma() against the acquire barrier implied by
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000104 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
Peter Zijlstra88c22082011-05-24 17:12:13 -0700105 *
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000106 * page_lock_anon_vma_read() VS put_anon_vma()
107 * down_read_trylock() atomic_dec_and_test()
Peter Zijlstra88c22082011-05-24 17:12:13 -0700108 * LOCK MB
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000109 * atomic_read() rwsem_is_locked()
Peter Zijlstra88c22082011-05-24 17:12:13 -0700110 *
111 * LOCK should suffice since the actual taking of the lock must
112 * happen _before_ what follows.
113 */
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700114 might_sleep();
Ingo Molnar5a505082012-12-02 19:56:46 +0000115 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000116 anon_vma_lock_write(anon_vma);
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -0800117 anon_vma_unlock_write(anon_vma);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700118 }
119
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -0700120 kmem_cache_free(anon_vma_cachep, anon_vma);
121}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122
Linus Torvaldsdd347392011-06-17 19:05:36 -0700123static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
Rik van Riel5beb4932010-03-05 13:42:07 -0800124{
Linus Torvaldsdd347392011-06-17 19:05:36 -0700125 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
Rik van Riel5beb4932010-03-05 13:42:07 -0800126}
127
Namhyung Kime574b5f2010-10-26 14:22:02 -0700128static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
Rik van Riel5beb4932010-03-05 13:42:07 -0800129{
130 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
131}
132
Kautuk Consul6583a842012-03-21 16:34:01 -0700133static void anon_vma_chain_link(struct vm_area_struct *vma,
134 struct anon_vma_chain *avc,
135 struct anon_vma *anon_vma)
136{
137 avc->vma = vma;
138 avc->anon_vma = anon_vma;
139 list_add(&avc->same_vma, &vma->anon_vma_chain);
Michel Lespinassebf181b92012-10-08 16:31:39 -0700140 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
Kautuk Consul6583a842012-03-21 16:34:01 -0700141}
142
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700143/**
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800144 * __anon_vma_prepare - attach an anon_vma to a memory region
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700145 * @vma: the memory region in question
146 *
147 * This makes sure the memory mapping described by 'vma' has
148 * an 'anon_vma' attached to it, so that we can associate the
149 * anonymous pages mapped into it with that anon_vma.
150 *
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800151 * The common case will be that we already have one, which
152 * is handled inline by anon_vma_prepare(). But if
Figo.zhang23a07902010-12-27 15:14:06 +0100153 * not we either need to find an adjacent mapping that we
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700154 * can re-use the anon_vma from (very common when the only
155 * reason for splitting a vma has been mprotect()), or we
156 * allocate a new one.
157 *
158 * Anon-vma allocations are very subtle, because we may have
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000159 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700160 * and that may actually touch the spinlock even in the newly
161 * allocated vma (it depends on RCU to make sure that the
162 * anon_vma isn't actually destroyed).
163 *
164 * As a result, we need to do proper anon_vma locking even
165 * for the new allocation. At the same time, we do not want
166 * to do any locking for the common case of already having
167 * an anon_vma.
168 *
169 * This must be called with the mmap_sem held for reading.
170 */
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800171int __anon_vma_prepare(struct vm_area_struct *vma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172{
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800173 struct mm_struct *mm = vma->vm_mm;
174 struct anon_vma *anon_vma, *allocated;
Rik van Riel5beb4932010-03-05 13:42:07 -0800175 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176
177 might_sleep();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800179 avc = anon_vma_chain_alloc(GFP_KERNEL);
180 if (!avc)
181 goto out_enomem;
Rik van Riel5beb4932010-03-05 13:42:07 -0800182
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800183 anon_vma = find_mergeable_anon_vma(vma);
184 allocated = NULL;
185 if (!anon_vma) {
186 anon_vma = anon_vma_alloc();
187 if (unlikely(!anon_vma))
188 goto out_enomem_free_avc;
189 allocated = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 }
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800191
192 anon_vma_lock_write(anon_vma);
193 /* page_table_lock to protect against threads */
194 spin_lock(&mm->page_table_lock);
195 if (likely(!vma->anon_vma)) {
196 vma->anon_vma = anon_vma;
197 anon_vma_chain_link(vma, avc, anon_vma);
198 /* vma reference or self-parent link for new root */
199 anon_vma->degree++;
200 allocated = NULL;
201 avc = NULL;
202 }
203 spin_unlock(&mm->page_table_lock);
204 anon_vma_unlock_write(anon_vma);
205
206 if (unlikely(allocated))
207 put_anon_vma(allocated);
208 if (unlikely(avc))
209 anon_vma_chain_free(avc);
210
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211 return 0;
Rik van Riel5beb4932010-03-05 13:42:07 -0800212
213 out_enomem_free_avc:
214 anon_vma_chain_free(avc);
215 out_enomem:
216 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217}
218
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700219/*
220 * This is a useful helper function for locking the anon_vma root as
221 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
222 * have the same vma.
223 *
224 * Such anon_vma's should have the same root, so you'd expect to see
225 * just a single mutex_lock for the whole traversal.
226 */
227static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
228{
229 struct anon_vma *new_root = anon_vma->root;
230 if (new_root != root) {
231 if (WARN_ON_ONCE(root))
Ingo Molnar5a505082012-12-02 19:56:46 +0000232 up_write(&root->rwsem);
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700233 root = new_root;
Ingo Molnar5a505082012-12-02 19:56:46 +0000234 down_write(&root->rwsem);
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700235 }
236 return root;
237}
238
239static inline void unlock_anon_vma_root(struct anon_vma *root)
240{
241 if (root)
Ingo Molnar5a505082012-12-02 19:56:46 +0000242 up_write(&root->rwsem);
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700243}
244
Rik van Riel5beb4932010-03-05 13:42:07 -0800245/*
246 * Attach the anon_vmas from src to dst.
247 * Returns 0 on success, -ENOMEM on failure.
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800248 *
249 * If dst->anon_vma is NULL this function tries to find and reuse existing
250 * anon_vma which has no vmas and only one child anon_vma. This prevents
251 * degradation of anon_vma hierarchy to endless linear chain in case of
252 * constantly forking task. On the other hand, an anon_vma with more than one
253 * child isn't reused even if there was no alive vma, thus rmap walker has a
254 * good chance of avoiding scanning the whole hierarchy when it searches where
255 * page is mapped.
Rik van Riel5beb4932010-03-05 13:42:07 -0800256 */
257int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258{
Rik van Riel5beb4932010-03-05 13:42:07 -0800259 struct anon_vma_chain *avc, *pavc;
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700260 struct anon_vma *root = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261
Linus Torvalds646d87b2010-04-11 17:15:03 -0700262 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700263 struct anon_vma *anon_vma;
264
Linus Torvaldsdd347392011-06-17 19:05:36 -0700265 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
266 if (unlikely(!avc)) {
267 unlock_anon_vma_root(root);
268 root = NULL;
269 avc = anon_vma_chain_alloc(GFP_KERNEL);
270 if (!avc)
271 goto enomem_failure;
272 }
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700273 anon_vma = pavc->anon_vma;
274 root = lock_anon_vma_root(root, anon_vma);
275 anon_vma_chain_link(dst, avc, anon_vma);
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800276
277 /*
278 * Reuse existing anon_vma if its degree lower than two,
279 * that means it has no vma and only one anon_vma child.
280 *
281 * Do not chose parent anon_vma, otherwise first child
282 * will always reuse it. Root anon_vma is never reused:
283 * it has self-parent reference and at least one child.
284 */
285 if (!dst->anon_vma && anon_vma != src->anon_vma &&
286 anon_vma->degree < 2)
287 dst->anon_vma = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288 }
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800289 if (dst->anon_vma)
290 dst->anon_vma->degree++;
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700291 unlock_anon_vma_root(root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800292 return 0;
293
294 enomem_failure:
Leon Yu3fe89b32015-03-25 15:55:11 -0700295 /*
296 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
297 * decremented in unlink_anon_vmas().
298 * We can safely do this because callers of anon_vma_clone() don't care
299 * about dst->anon_vma if anon_vma_clone() failed.
300 */
301 dst->anon_vma = NULL;
Rik van Riel5beb4932010-03-05 13:42:07 -0800302 unlink_anon_vmas(dst);
303 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304}
305
Rik van Riel5beb4932010-03-05 13:42:07 -0800306/*
307 * Attach vma to its own anon_vma, as well as to the anon_vmas that
308 * the corresponding VMA in the parent process is attached to.
309 * Returns 0 on success, non-zero on failure.
310 */
311int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312{
Rik van Riel5beb4932010-03-05 13:42:07 -0800313 struct anon_vma_chain *avc;
314 struct anon_vma *anon_vma;
Daniel Forrestc4ea95d2014-12-02 15:59:42 -0800315 int error;
Rik van Riel5beb4932010-03-05 13:42:07 -0800316
317 /* Don't bother if the parent process has no anon_vma here. */
318 if (!pvma->anon_vma)
319 return 0;
320
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800321 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
322 vma->anon_vma = NULL;
323
Rik van Riel5beb4932010-03-05 13:42:07 -0800324 /*
325 * First, attach the new VMA to the parent VMA's anon_vmas,
326 * so rmap can find non-COWed pages in child processes.
327 */
Daniel Forrestc4ea95d2014-12-02 15:59:42 -0800328 error = anon_vma_clone(vma, pvma);
329 if (error)
330 return error;
Rik van Riel5beb4932010-03-05 13:42:07 -0800331
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800332 /* An existing anon_vma has been reused, all done then. */
333 if (vma->anon_vma)
334 return 0;
335
Rik van Riel5beb4932010-03-05 13:42:07 -0800336 /* Then add our own anon_vma. */
337 anon_vma = anon_vma_alloc();
338 if (!anon_vma)
339 goto out_error;
Linus Torvaldsdd347392011-06-17 19:05:36 -0700340 avc = anon_vma_chain_alloc(GFP_KERNEL);
Rik van Riel5beb4932010-03-05 13:42:07 -0800341 if (!avc)
342 goto out_error_free_anon_vma;
Rik van Riel5c341ee12010-08-09 17:18:39 -0700343
344 /*
345 * The root anon_vma's spinlock is the lock actually used when we
346 * lock any of the anon_vmas in this anon_vma tree.
347 */
348 anon_vma->root = pvma->anon_vma->root;
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800349 anon_vma->parent = pvma->anon_vma;
Rik van Riel76545062010-08-09 17:18:41 -0700350 /*
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700351 * With refcounts, an anon_vma can stay around longer than the
352 * process it belongs to. The root anon_vma needs to be pinned until
353 * this anon_vma is freed, because the lock lives in the root.
Rik van Riel76545062010-08-09 17:18:41 -0700354 */
355 get_anon_vma(anon_vma->root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800356 /* Mark this anon_vma as the one where our new (COWed) pages go. */
357 vma->anon_vma = anon_vma;
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000358 anon_vma_lock_write(anon_vma);
Rik van Riel5c341ee12010-08-09 17:18:39 -0700359 anon_vma_chain_link(vma, avc, anon_vma);
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800360 anon_vma->parent->degree++;
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -0800361 anon_vma_unlock_write(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800362
363 return 0;
364
365 out_error_free_anon_vma:
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700366 put_anon_vma(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800367 out_error:
Rik van Riel4946d542010-04-05 12:13:33 -0400368 unlink_anon_vmas(vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800369 return -ENOMEM;
370}
371
Rik van Riel5beb4932010-03-05 13:42:07 -0800372void unlink_anon_vmas(struct vm_area_struct *vma)
373{
374 struct anon_vma_chain *avc, *next;
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200375 struct anon_vma *root = NULL;
Rik van Riel5beb4932010-03-05 13:42:07 -0800376
Rik van Riel5c341ee12010-08-09 17:18:39 -0700377 /*
378 * Unlink each anon_vma chained to the VMA. This list is ordered
379 * from newest to oldest, ensuring the root anon_vma gets freed last.
380 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800381 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200382 struct anon_vma *anon_vma = avc->anon_vma;
383
384 root = lock_anon_vma_root(root, anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -0700385 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200386
387 /*
388 * Leave empty anon_vmas on the list - we'll need
389 * to free them outside the lock.
390 */
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800391 if (RB_EMPTY_ROOT(&anon_vma->rb_root)) {
392 anon_vma->parent->degree--;
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200393 continue;
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800394 }
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200395
396 list_del(&avc->same_vma);
397 anon_vma_chain_free(avc);
398 }
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800399 if (vma->anon_vma)
400 vma->anon_vma->degree--;
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200401 unlock_anon_vma_root(root);
402
403 /*
404 * Iterate the list once more, it now only contains empty and unlinked
405 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
Ingo Molnar5a505082012-12-02 19:56:46 +0000406 * needing to write-acquire the anon_vma->root->rwsem.
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200407 */
408 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
409 struct anon_vma *anon_vma = avc->anon_vma;
410
Konstantin Khlebnikove4c58002016-05-19 17:11:46 -0700411 VM_WARN_ON(anon_vma->degree);
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200412 put_anon_vma(anon_vma);
413
Rik van Riel5beb4932010-03-05 13:42:07 -0800414 list_del(&avc->same_vma);
415 anon_vma_chain_free(avc);
416 }
417}
418
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700419static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420{
Christoph Lametera35afb82007-05-16 22:10:57 -0700421 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422
Ingo Molnar5a505082012-12-02 19:56:46 +0000423 init_rwsem(&anon_vma->rwsem);
Peter Zijlstra83813262011-03-22 16:32:48 -0700424 atomic_set(&anon_vma->refcount, 0);
Michel Lespinassebf181b92012-10-08 16:31:39 -0700425 anon_vma->rb_root = RB_ROOT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426}
427
428void __init anon_vma_init(void)
429{
430 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Vladimir Davydov5d097052016-01-14 15:18:21 -0800431 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
432 anon_vma_ctor);
433 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
434 SLAB_PANIC|SLAB_ACCOUNT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435}
436
437/*
Peter Zijlstra6111e4c2011-05-24 17:12:08 -0700438 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
439 *
440 * Since there is no serialization what so ever against page_remove_rmap()
441 * the best this function can do is return a locked anon_vma that might
442 * have been relevant to this page.
443 *
444 * The page might have been remapped to a different anon_vma or the anon_vma
445 * returned may already be freed (and even reused).
446 *
Peter Zijlstrabc658c92011-05-29 10:33:44 +0200447 * In case it was remapped to a different anon_vma, the new anon_vma will be a
448 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
449 * ensure that any anon_vma obtained from the page will still be valid for as
450 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
451 *
Peter Zijlstra6111e4c2011-05-24 17:12:08 -0700452 * All users of this function must be very careful when walking the anon_vma
453 * chain and verify that the page in question is indeed mapped in it
454 * [ something equivalent to page_mapped_in_vma() ].
455 *
456 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
457 * that the anon_vma pointer from page->mapping is valid if there is a
458 * mapcount, we can dereference the anon_vma after observing those.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700460struct anon_vma *page_get_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461{
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700462 struct anon_vma *anon_vma = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 unsigned long anon_mapping;
464
465 rcu_read_lock();
Jason Low4db0c3c2015-04-15 16:14:08 -0700466 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800467 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468 goto out;
469 if (!page_mapped(page))
470 goto out;
471
472 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700473 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
474 anon_vma = NULL;
475 goto out;
476 }
Hugh Dickinsf18194272010-08-25 23:12:54 -0700477
478 /*
479 * If this page is still mapped, then its anon_vma cannot have been
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700480 * freed. But if it has been unmapped, we have no security against the
481 * anon_vma structure being freed and reused (for another anon_vma:
482 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
483 * above cannot corrupt).
Hugh Dickinsf18194272010-08-25 23:12:54 -0700484 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700485 if (!page_mapped(page)) {
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700486 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700487 put_anon_vma(anon_vma);
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700488 return NULL;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700489 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490out:
491 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700492
493 return anon_vma;
494}
495
Peter Zijlstra88c22082011-05-24 17:12:13 -0700496/*
497 * Similar to page_get_anon_vma() except it locks the anon_vma.
498 *
499 * Its a little more complex as it tries to keep the fast path to a single
500 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
501 * reference like with page_get_anon_vma() and then block on the mutex.
502 */
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000503struct anon_vma *page_lock_anon_vma_read(struct page *page)
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700504{
Peter Zijlstra88c22082011-05-24 17:12:13 -0700505 struct anon_vma *anon_vma = NULL;
Hugh Dickinseee0f252011-05-28 13:20:21 -0700506 struct anon_vma *root_anon_vma;
Peter Zijlstra88c22082011-05-24 17:12:13 -0700507 unsigned long anon_mapping;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700508
Peter Zijlstra88c22082011-05-24 17:12:13 -0700509 rcu_read_lock();
Jason Low4db0c3c2015-04-15 16:14:08 -0700510 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700511 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
512 goto out;
513 if (!page_mapped(page))
514 goto out;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700515
Peter Zijlstra88c22082011-05-24 17:12:13 -0700516 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Jason Low4db0c3c2015-04-15 16:14:08 -0700517 root_anon_vma = READ_ONCE(anon_vma->root);
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000518 if (down_read_trylock(&root_anon_vma->rwsem)) {
Peter Zijlstra88c22082011-05-24 17:12:13 -0700519 /*
Hugh Dickinseee0f252011-05-28 13:20:21 -0700520 * If the page is still mapped, then this anon_vma is still
521 * its anon_vma, and holding the mutex ensures that it will
Peter Zijlstrabc658c92011-05-29 10:33:44 +0200522 * not go away, see anon_vma_free().
Peter Zijlstra88c22082011-05-24 17:12:13 -0700523 */
Hugh Dickinseee0f252011-05-28 13:20:21 -0700524 if (!page_mapped(page)) {
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000525 up_read(&root_anon_vma->rwsem);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700526 anon_vma = NULL;
527 }
528 goto out;
529 }
530
531 /* trylock failed, we got to sleep */
532 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
533 anon_vma = NULL;
534 goto out;
535 }
536
537 if (!page_mapped(page)) {
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700538 rcu_read_unlock();
Peter Zijlstra88c22082011-05-24 17:12:13 -0700539 put_anon_vma(anon_vma);
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700540 return NULL;
Peter Zijlstra88c22082011-05-24 17:12:13 -0700541 }
542
543 /* we pinned the anon_vma, its safe to sleep */
544 rcu_read_unlock();
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000545 anon_vma_lock_read(anon_vma);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700546
547 if (atomic_dec_and_test(&anon_vma->refcount)) {
548 /*
549 * Oops, we held the last refcount, release the lock
550 * and bail -- can't simply use put_anon_vma() because
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000551 * we'll deadlock on the anon_vma_lock_write() recursion.
Peter Zijlstra88c22082011-05-24 17:12:13 -0700552 */
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000553 anon_vma_unlock_read(anon_vma);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700554 __put_anon_vma(anon_vma);
555 anon_vma = NULL;
556 }
557
558 return anon_vma;
559
560out:
561 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700562 return anon_vma;
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800563}
564
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000565void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800566{
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000567 anon_vma_unlock_read(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568}
569
Mel Gorman72b252a2015-09-04 15:47:32 -0700570#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
Mel Gorman72b252a2015-09-04 15:47:32 -0700571/*
572 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
573 * important if a PTE was dirty when it was unmapped that it's flushed
574 * before any IO is initiated on the page to prevent lost writes. Similarly,
575 * it must be flushed before freeing to prevent data leakage.
576 */
577void try_to_unmap_flush(void)
578{
579 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
580 int cpu;
581
582 if (!tlb_ubc->flush_required)
583 return;
584
585 cpu = get_cpu();
586
Nadav Amit858eaaa72016-04-01 14:31:26 -0700587 if (cpumask_test_cpu(cpu, &tlb_ubc->cpumask)) {
588 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
589 local_flush_tlb();
590 trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
Mel Gorman72b252a2015-09-04 15:47:32 -0700591 }
Nadav Amit858eaaa72016-04-01 14:31:26 -0700592
593 if (cpumask_any_but(&tlb_ubc->cpumask, cpu) < nr_cpu_ids)
594 flush_tlb_others(&tlb_ubc->cpumask, NULL, 0, TLB_FLUSH_ALL);
Mel Gorman72b252a2015-09-04 15:47:32 -0700595 cpumask_clear(&tlb_ubc->cpumask);
596 tlb_ubc->flush_required = false;
Mel Gormand950c942015-09-04 15:47:35 -0700597 tlb_ubc->writable = false;
Mel Gorman72b252a2015-09-04 15:47:32 -0700598 put_cpu();
599}
600
Mel Gormand950c942015-09-04 15:47:35 -0700601/* Flush iff there are potentially writable TLB entries that can race with IO */
602void try_to_unmap_flush_dirty(void)
603{
604 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
605
606 if (tlb_ubc->writable)
607 try_to_unmap_flush();
608}
609
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -0800610static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
Mel Gorman72b252a2015-09-04 15:47:32 -0700611{
612 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
613
614 cpumask_or(&tlb_ubc->cpumask, &tlb_ubc->cpumask, mm_cpumask(mm));
615 tlb_ubc->flush_required = true;
Mel Gormand950c942015-09-04 15:47:35 -0700616
617 /*
618 * If the PTE was dirty then it's best to assume it's writable. The
619 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
620 * before the page is queued for IO.
621 */
622 if (writable)
623 tlb_ubc->writable = true;
Mel Gorman72b252a2015-09-04 15:47:32 -0700624}
625
626/*
627 * Returns true if the TLB flush should be deferred to the end of a batch of
628 * unmap operations to reduce IPIs.
629 */
630static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
631{
632 bool should_defer = false;
633
634 if (!(flags & TTU_BATCH_FLUSH))
635 return false;
636
637 /* If remote CPUs need to be flushed then defer batch the flush */
638 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
639 should_defer = true;
640 put_cpu();
641
642 return should_defer;
643}
644#else
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -0800645static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
Mel Gorman72b252a2015-09-04 15:47:32 -0700646{
647}
648
649static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
650{
651 return false;
652}
653#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
654
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700656 * At what user virtual address is page expected in vma?
Naoya Horiguchiab941e02010-05-11 14:06:55 -0700657 * Caller should check the page is actually part of the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 */
659unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
660{
Michel Lespinasse86c2ad12012-10-08 16:31:42 -0700661 unsigned long address;
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700662 if (PageAnon(page)) {
Hugh Dickins4829b902010-10-02 17:46:06 -0700663 struct anon_vma *page__anon_vma = page_anon_vma(page);
664 /*
665 * Note: swapoff's unuse_vma() is more efficient with this
666 * check, and needs it to match anon_vma when KSM is active.
667 */
668 if (!vma->anon_vma || !page__anon_vma ||
669 vma->anon_vma->root != page__anon_vma->root)
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700670 return -EFAULT;
Kirill A. Shutemov27ba0642015-02-10 14:09:59 -0800671 } else if (page->mapping) {
672 if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 return -EFAULT;
674 } else
675 return -EFAULT;
Michel Lespinasse86c2ad12012-10-08 16:31:42 -0700676 address = __vma_address(page, vma);
677 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
678 return -EFAULT;
679 return address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680}
681
Bob Liu62190492012-12-11 16:00:37 -0800682pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
683{
684 pgd_t *pgd;
685 pud_t *pud;
686 pmd_t *pmd = NULL;
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -0700687 pmd_t pmde;
Bob Liu62190492012-12-11 16:00:37 -0800688
689 pgd = pgd_offset(mm, address);
690 if (!pgd_present(*pgd))
691 goto out;
692
693 pud = pud_offset(pgd, address);
694 if (!pud_present(*pud))
695 goto out;
696
697 pmd = pmd_offset(pud, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -0700698 /*
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -0700699 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -0700700 * without holding anon_vma lock for write. So when looking for a
701 * genuine pmde (in which to find pte), test present and !THP together.
702 */
Christian Borntraegere37c6982014-12-07 21:41:33 +0100703 pmde = *pmd;
704 barrier();
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -0700705 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
Bob Liu62190492012-12-11 16:00:37 -0800706 pmd = NULL;
707out:
708 return pmd;
709}
710
Joonsoo Kim9f326242014-01-21 15:49:53 -0800711struct page_referenced_arg {
712 int mapcount;
713 int referenced;
714 unsigned long vm_flags;
715 struct mem_cgroup *memcg;
716};
Nikita Danilov81b40822005-05-01 08:58:36 -0700717/*
Joonsoo Kim9f326242014-01-21 15:49:53 -0800718 * arg: page_referenced_arg will be passed
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719 */
Kirill A. Shutemovac769502014-06-04 16:08:17 -0700720static int page_referenced_one(struct page *page, struct vm_area_struct *vma,
Joonsoo Kim9f326242014-01-21 15:49:53 -0800721 unsigned long address, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722{
Joonsoo Kim9f326242014-01-21 15:49:53 -0800723 struct page_referenced_arg *pra = arg;
Kirill A. Shutemov8eaeded2017-02-24 14:57:48 -0800724 struct page_vma_mapped_walk pvmw = {
725 .page = page,
726 .vma = vma,
727 .address = address,
728 };
Vladimir Davydov8749cfe2016-01-15 16:54:45 -0800729 int referenced = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730
Kirill A. Shutemov8eaeded2017-02-24 14:57:48 -0800731 while (page_vma_mapped_walk(&pvmw)) {
732 address = pvmw.address;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800733
Kirill A. Shutemov8eaeded2017-02-24 14:57:48 -0800734 if (vma->vm_flags & VM_LOCKED) {
735 page_vma_mapped_walk_done(&pvmw);
736 pra->vm_flags |= VM_LOCKED;
737 return SWAP_FAIL; /* To break the loop */
Vladimir Davydov8749cfe2016-01-15 16:54:45 -0800738 }
Kirill A. Shutemov8eaeded2017-02-24 14:57:48 -0800739
740 if (pvmw.pte) {
741 if (ptep_clear_flush_young_notify(vma, address,
742 pvmw.pte)) {
743 /*
744 * Don't treat a reference through
745 * a sequentially read mapping as such.
746 * If the page has been used in another mapping,
747 * we will catch it; if this other mapping is
748 * already gone, the unmap path will have set
749 * PG_referenced or activated the page.
750 */
751 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
752 referenced++;
753 }
754 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
755 if (pmdp_clear_flush_young_notify(vma, address,
756 pvmw.pmd))
757 referenced++;
758 } else {
759 /* unexpected pmd-mapped page? */
760 WARN_ON_ONCE(1);
761 }
762
763 pra->mapcount--;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800764 }
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800765
Vladimir Davydov33c3fc72015-09-09 15:35:45 -0700766 if (referenced)
767 clear_page_idle(page);
768 if (test_and_clear_page_young(page))
769 referenced++;
770
Joonsoo Kim9f326242014-01-21 15:49:53 -0800771 if (referenced) {
772 pra->referenced++;
773 pra->vm_flags |= vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800775
Joonsoo Kim9f326242014-01-21 15:49:53 -0800776 if (!pra->mapcount)
777 return SWAP_SUCCESS; /* To break the loop */
778
779 return SWAP_AGAIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780}
781
Joonsoo Kim9f326242014-01-21 15:49:53 -0800782static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783{
Joonsoo Kim9f326242014-01-21 15:49:53 -0800784 struct page_referenced_arg *pra = arg;
785 struct mem_cgroup *memcg = pra->memcg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786
Joonsoo Kim9f326242014-01-21 15:49:53 -0800787 if (!mm_match_cgroup(vma->vm_mm, memcg))
788 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789
Joonsoo Kim9f326242014-01-21 15:49:53 -0800790 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791}
792
793/**
794 * page_referenced - test if the page was referenced
795 * @page: the page to test
796 * @is_locked: caller holds lock on the page
Johannes Weiner72835c82012-01-12 17:18:32 -0800797 * @memcg: target memory cgroup
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700798 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 *
800 * Quick test_and_clear_referenced for all mappings to a page,
801 * returns the number of ptes which referenced the page.
802 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700803int page_referenced(struct page *page,
804 int is_locked,
Johannes Weiner72835c82012-01-12 17:18:32 -0800805 struct mem_cgroup *memcg,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700806 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807{
Joonsoo Kim9f326242014-01-21 15:49:53 -0800808 int ret;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800809 int we_locked = 0;
Joonsoo Kim9f326242014-01-21 15:49:53 -0800810 struct page_referenced_arg pra = {
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800811 .mapcount = total_mapcount(page),
Joonsoo Kim9f326242014-01-21 15:49:53 -0800812 .memcg = memcg,
813 };
814 struct rmap_walk_control rwc = {
815 .rmap_one = page_referenced_one,
816 .arg = (void *)&pra,
817 .anon_lock = page_lock_anon_vma_read,
818 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700820 *vm_flags = 0;
Joonsoo Kim9f326242014-01-21 15:49:53 -0800821 if (!page_mapped(page))
822 return 0;
823
824 if (!page_rmapping(page))
825 return 0;
826
827 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
828 we_locked = trylock_page(page);
829 if (!we_locked)
830 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 }
Joonsoo Kim9f326242014-01-21 15:49:53 -0800832
833 /*
834 * If we are reclaiming on behalf of a cgroup, skip
835 * counting on behalf of references from different
836 * cgroups
837 */
838 if (memcg) {
839 rwc.invalid_vma = invalid_page_referenced_vma;
840 }
841
842 ret = rmap_walk(page, &rwc);
843 *vm_flags = pra.vm_flags;
844
845 if (we_locked)
846 unlock_page(page);
847
848 return pra.referenced;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849}
850
Hugh Dickins1cb17292009-12-14 17:59:01 -0800851static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
Joonsoo Kim9853a402014-01-21 15:49:55 -0800852 unsigned long address, void *arg)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700853{
Kirill A. Shutemovf27176c2017-02-24 14:57:57 -0800854 struct page_vma_mapped_walk pvmw = {
855 .page = page,
856 .vma = vma,
857 .address = address,
858 .flags = PVMW_SYNC,
859 };
Joonsoo Kim9853a402014-01-21 15:49:55 -0800860 int *cleaned = arg;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700861
Kirill A. Shutemovf27176c2017-02-24 14:57:57 -0800862 while (page_vma_mapped_walk(&pvmw)) {
863 int ret = 0;
864 address = pvmw.address;
865 if (pvmw.pte) {
866 pte_t entry;
867 pte_t *pte = pvmw.pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700868
Kirill A. Shutemovf27176c2017-02-24 14:57:57 -0800869 if (!pte_dirty(*pte) && !pte_write(*pte))
870 continue;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700871
Kirill A. Shutemovf27176c2017-02-24 14:57:57 -0800872 flush_cache_page(vma, address, pte_pfn(*pte));
873 entry = ptep_clear_flush(vma, address, pte);
874 entry = pte_wrprotect(entry);
875 entry = pte_mkclean(entry);
876 set_pte_at(vma->vm_mm, address, pte, entry);
877 ret = 1;
878 } else {
879#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
880 pmd_t *pmd = pvmw.pmd;
881 pmd_t entry;
882
883 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
884 continue;
885
886 flush_cache_page(vma, address, page_to_pfn(page));
887 entry = pmdp_huge_clear_flush(vma, address, pmd);
888 entry = pmd_wrprotect(entry);
889 entry = pmd_mkclean(entry);
890 set_pmd_at(vma->vm_mm, address, pmd, entry);
891 ret = 1;
892#else
893 /* unexpected pmd-mapped page? */
894 WARN_ON_ONCE(1);
895#endif
896 }
897
898 if (ret) {
899 mmu_notifier_invalidate_page(vma->vm_mm, address);
900 (*cleaned)++;
901 }
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100902 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700903
Joonsoo Kim9853a402014-01-21 15:49:55 -0800904 return SWAP_AGAIN;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700905}
906
Joonsoo Kim9853a402014-01-21 15:49:55 -0800907static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700908{
Joonsoo Kim9853a402014-01-21 15:49:55 -0800909 if (vma->vm_flags & VM_SHARED)
Fengguang Wu871beb82014-01-23 15:53:41 -0800910 return false;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700911
Fengguang Wu871beb82014-01-23 15:53:41 -0800912 return true;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700913}
914
915int page_mkclean(struct page *page)
916{
Joonsoo Kim9853a402014-01-21 15:49:55 -0800917 int cleaned = 0;
918 struct address_space *mapping;
919 struct rmap_walk_control rwc = {
920 .arg = (void *)&cleaned,
921 .rmap_one = page_mkclean_one,
922 .invalid_vma = invalid_mkclean_vma,
923 };
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700924
925 BUG_ON(!PageLocked(page));
926
Joonsoo Kim9853a402014-01-21 15:49:55 -0800927 if (!page_mapped(page))
928 return 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700929
Joonsoo Kim9853a402014-01-21 15:49:55 -0800930 mapping = page_mapping(page);
931 if (!mapping)
932 return 0;
933
934 rmap_walk(page, &rwc);
935
936 return cleaned;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700937}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700938EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700939
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940/**
Rik van Rielc44b6742010-03-05 13:42:09 -0800941 * page_move_anon_rmap - move a page to our anon_vma
942 * @page: the page to move to our anon_vma
943 * @vma: the vma the page belongs to
Rik van Rielc44b6742010-03-05 13:42:09 -0800944 *
945 * When a page belongs exclusively to one process after a COW event,
946 * that page can be moved into the anon_vma that belongs to just that
947 * process, so the rmap code will not search the parent or sibling
948 * processes.
949 */
Hugh Dickins5a499732016-07-14 12:07:38 -0700950void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
Rik van Rielc44b6742010-03-05 13:42:09 -0800951{
952 struct anon_vma *anon_vma = vma->anon_vma;
953
Hugh Dickins5a499732016-07-14 12:07:38 -0700954 page = compound_head(page);
955
Sasha Levin309381fea2014-01-23 15:52:54 -0800956 VM_BUG_ON_PAGE(!PageLocked(page), page);
Sasha Levin81d1b092014-10-09 15:28:10 -0700957 VM_BUG_ON_VMA(!anon_vma, vma);
Rik van Rielc44b6742010-03-05 13:42:09 -0800958
959 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
Vladimir Davydov414e2fb2015-06-24 16:56:56 -0700960 /*
961 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
962 * simultaneously, so a concurrent reader (eg page_referenced()'s
963 * PageAnon()) will not see one without the other.
964 */
965 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
Rik van Rielc44b6742010-03-05 13:42:09 -0800966}
967
968/**
Andi Kleen4e1c1972010-09-22 12:43:56 +0200969 * __page_set_anon_rmap - set up new anonymous rmap
970 * @page: Page to add to rmap
971 * @vma: VM area to add page to.
972 * @address: User virtual address of the mapping
Rik van Riele8a03fe2010-04-14 17:59:28 -0400973 * @exclusive: the page is exclusively owned by the current process
Nick Piggin9617d952006-01-06 00:11:12 -0800974 */
975static void __page_set_anon_rmap(struct page *page,
Rik van Riele8a03fe2010-04-14 17:59:28 -0400976 struct vm_area_struct *vma, unsigned long address, int exclusive)
Nick Piggin9617d952006-01-06 00:11:12 -0800977{
Rik van Riele8a03fe2010-04-14 17:59:28 -0400978 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800979
Rik van Riele8a03fe2010-04-14 17:59:28 -0400980 BUG_ON(!anon_vma);
Linus Torvaldsea900022010-04-12 12:44:29 -0700981
Andi Kleen4e1c1972010-09-22 12:43:56 +0200982 if (PageAnon(page))
983 return;
984
Linus Torvaldsea900022010-04-12 12:44:29 -0700985 /*
Rik van Riele8a03fe2010-04-14 17:59:28 -0400986 * If the page isn't exclusively mapped into this vma,
987 * we must use the _oldest_ possible anon_vma for the
988 * page mapping!
Linus Torvaldsea900022010-04-12 12:44:29 -0700989 */
Andi Kleen4e1c1972010-09-22 12:43:56 +0200990 if (!exclusive)
Andrea Arcangeli288468c2010-08-09 17:19:09 -0700991 anon_vma = anon_vma->root;
Linus Torvaldsea900022010-04-12 12:44:29 -0700992
Nick Piggin9617d952006-01-06 00:11:12 -0800993 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
994 page->mapping = (struct address_space *) anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800995 page->index = linear_page_index(vma, address);
Nick Piggin9617d952006-01-06 00:11:12 -0800996}
997
998/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -0700999 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -07001000 * @page: the page to add the mapping to
1001 * @vma: the vm area in which the mapping is added
1002 * @address: the user virtual address mapped
1003 */
1004static void __page_check_anon_rmap(struct page *page,
1005 struct vm_area_struct *vma, unsigned long address)
1006{
1007#ifdef CONFIG_DEBUG_VM
1008 /*
1009 * The page's anon-rmap details (mapping and index) are guaranteed to
1010 * be set up correctly at this point.
1011 *
1012 * We have exclusion against page_add_anon_rmap because the caller
1013 * always holds the page locked, except if called from page_dup_rmap,
1014 * in which case the page is already known to be setup.
1015 *
1016 * We have exclusion against page_add_new_anon_rmap because those pages
1017 * are initially only visible via the pagetables, and the pte is locked
1018 * over the call to page_add_new_anon_rmap.
1019 */
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001020 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001021 BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
Nick Pigginc97a9e12007-05-16 22:11:21 -07001022#endif
1023}
1024
1025/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026 * page_add_anon_rmap - add pte mapping to an anonymous page
1027 * @page: the page to add the mapping to
1028 * @vma: the vm area in which the mapping is added
1029 * @address: the user virtual address mapped
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001030 * @compound: charge the page as compound or small page
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 *
Hugh Dickins5ad64682009-12-14 17:59:24 -08001032 * The caller needs to hold the pte lock, and the page must be locked in
Hugh Dickins80e148222009-12-14 17:59:29 -08001033 * the anon_vma case: to serialize mapping,index checking after setting,
1034 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1035 * (but PageKsm is never downgraded to PageAnon).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 */
1037void page_add_anon_rmap(struct page *page,
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001038 struct vm_area_struct *vma, unsigned long address, bool compound)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039{
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001040 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
Rik van Rielad8c2ee2010-08-09 17:19:48 -07001041}
1042
1043/*
1044 * Special version of the above for do_swap_page, which often runs
1045 * into pages that are exclusively owned by the current process.
1046 * Everybody else should continue to use page_add_anon_rmap above.
1047 */
1048void do_page_add_anon_rmap(struct page *page,
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001049 struct vm_area_struct *vma, unsigned long address, int flags)
Rik van Rielad8c2ee2010-08-09 17:19:48 -07001050{
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001051 bool compound = flags & RMAP_COMPOUND;
1052 bool first;
1053
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001054 if (compound) {
1055 atomic_t *mapcount;
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001056 VM_BUG_ON_PAGE(!PageLocked(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001057 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1058 mapcount = compound_mapcount_ptr(page);
1059 first = atomic_inc_and_test(mapcount);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001060 } else {
1061 first = atomic_inc_and_test(&page->_mapcount);
1062 }
1063
Andrea Arcangeli79134172011-01-13 15:46:58 -08001064 if (first) {
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001065 int nr = compound ? hpage_nr_pages(page) : 1;
Jianyu Zhanbea04b02014-06-04 16:09:51 -07001066 /*
1067 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1068 * these counters are not modified in interrupt context, and
1069 * pte lock(a spinlock) is held, which implies preemption
1070 * disabled.
1071 */
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07001072 if (compound)
Mel Gorman11fb9982016-07-28 15:46:20 -07001073 __inc_node_page_state(page, NR_ANON_THPS);
Mel Gorman4b9d0fa2016-07-28 15:46:17 -07001074 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001075 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001076 if (unlikely(PageKsm(page)))
1077 return;
1078
Sasha Levin309381fea2014-01-23 15:52:54 -08001079 VM_BUG_ON_PAGE(!PageLocked(page), page);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001080
Hugh Dickins5dbe0af2011-05-28 13:17:04 -07001081 /* address might be in next vma when migration races vma_adjust */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001082 if (first)
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001083 __page_set_anon_rmap(page, vma, address,
1084 flags & RMAP_EXCLUSIVE);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -07001085 else
Nick Pigginc97a9e12007-05-16 22:11:21 -07001086 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087}
1088
Randy Dunlap43d8eac2008-03-19 17:00:43 -07001089/**
Nick Piggin9617d952006-01-06 00:11:12 -08001090 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1091 * @page: the page to add the mapping to
1092 * @vma: the vm area in which the mapping is added
1093 * @address: the user virtual address mapped
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001094 * @compound: charge the page as compound or small page
Nick Piggin9617d952006-01-06 00:11:12 -08001095 *
1096 * Same as page_add_anon_rmap but must only be called on *new* pages.
1097 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -07001098 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -08001099 */
1100void page_add_new_anon_rmap(struct page *page,
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001101 struct vm_area_struct *vma, unsigned long address, bool compound)
Nick Piggin9617d952006-01-06 00:11:12 -08001102{
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001103 int nr = compound ? hpage_nr_pages(page) : 1;
1104
Sasha Levin81d1b092014-10-09 15:28:10 -07001105 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
Hugh Dickinsfa9949d2016-05-19 17:12:41 -07001106 __SetPageSwapBacked(page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001107 if (compound) {
1108 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001109 /* increment count (starts at -1) */
1110 atomic_set(compound_mapcount_ptr(page), 0);
Mel Gorman11fb9982016-07-28 15:46:20 -07001111 __inc_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001112 } else {
1113 /* Anon THP always mapped first with PMD */
1114 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1115 /* increment count (starts at -1) */
1116 atomic_set(&page->_mapcount, 0);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001117 }
Mel Gorman4b9d0fa2016-07-28 15:46:17 -07001118 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
Rik van Riele8a03fe2010-04-14 17:59:28 -04001119 __page_set_anon_rmap(page, vma, address, 1);
Nick Piggin9617d952006-01-06 00:11:12 -08001120}
1121
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122/**
1123 * page_add_file_rmap - add pte mapping to a file page
1124 * @page: the page to add the mapping to
1125 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -07001126 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 */
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001128void page_add_file_rmap(struct page *page, bool compound)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001130 int i, nr = 1;
1131
1132 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
Johannes Weiner62cccb82016-03-15 14:57:22 -07001133 lock_page_memcg(page);
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001134 if (compound && PageTransHuge(page)) {
1135 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1136 if (atomic_inc_and_test(&page[i]._mapcount))
1137 nr++;
1138 }
1139 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1140 goto out;
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07001141 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
Mel Gorman11fb9982016-07-28 15:46:20 -07001142 __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001143 } else {
Kirill A. Shutemovc8efc392016-08-10 16:27:52 -07001144 if (PageTransCompound(page) && page_mapping(page)) {
1145 VM_WARN_ON_ONCE(!PageLocked(page));
1146
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001147 SetPageDoubleMap(compound_head(page));
1148 if (PageMlocked(page))
1149 clear_page_mlock(compound_head(page));
1150 }
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001151 if (!atomic_inc_and_test(&page->_mapcount))
1152 goto out;
Balbir Singhd69b0422009-06-17 16:26:34 -07001153 }
Mel Gorman50658e22016-07-28 15:46:14 -07001154 __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, nr);
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001155 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
1156out:
Johannes Weiner62cccb82016-03-15 14:57:22 -07001157 unlock_page_memcg(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158}
1159
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001160static void page_remove_file_rmap(struct page *page, bool compound)
Johannes Weiner8186eb62014-10-29 14:50:51 -07001161{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001162 int i, nr = 1;
1163
Steve Capper57dea932016-08-10 16:27:55 -07001164 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
Johannes Weiner62cccb82016-03-15 14:57:22 -07001165 lock_page_memcg(page);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001166
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001167 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1168 if (unlikely(PageHuge(page))) {
1169 /* hugetlb pages are always mapped with pmds */
1170 atomic_dec(compound_mapcount_ptr(page));
1171 goto out;
1172 }
1173
Johannes Weiner8186eb62014-10-29 14:50:51 -07001174 /* page still mapped by someone else? */
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001175 if (compound && PageTransHuge(page)) {
1176 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1177 if (atomic_add_negative(-1, &page[i]._mapcount))
1178 nr++;
1179 }
1180 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1181 goto out;
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07001182 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
Mel Gorman11fb9982016-07-28 15:46:20 -07001183 __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001184 } else {
1185 if (!atomic_add_negative(-1, &page->_mapcount))
1186 goto out;
1187 }
Johannes Weiner8186eb62014-10-29 14:50:51 -07001188
Johannes Weiner8186eb62014-10-29 14:50:51 -07001189 /*
Mel Gorman50658e22016-07-28 15:46:14 -07001190 * We use the irq-unsafe __{inc|mod}_zone_page_state because
Johannes Weiner8186eb62014-10-29 14:50:51 -07001191 * these counters are not modified in interrupt context, and
1192 * pte lock(a spinlock) is held, which implies preemption disabled.
1193 */
Mel Gorman50658e22016-07-28 15:46:14 -07001194 __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, -nr);
Johannes Weiner62cccb82016-03-15 14:57:22 -07001195 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001196
1197 if (unlikely(PageMlocked(page)))
1198 clear_page_mlock(page);
1199out:
Johannes Weiner62cccb82016-03-15 14:57:22 -07001200 unlock_page_memcg(page);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001201}
1202
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001203static void page_remove_anon_compound_rmap(struct page *page)
1204{
1205 int i, nr;
1206
1207 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1208 return;
1209
1210 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1211 if (unlikely(PageHuge(page)))
1212 return;
1213
1214 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1215 return;
1216
Mel Gorman11fb9982016-07-28 15:46:20 -07001217 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001218
1219 if (TestClearPageDoubleMap(page)) {
1220 /*
1221 * Subpages can be mapped with PTEs too. Check how many of
1222 * themi are still mapped.
1223 */
1224 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1225 if (atomic_add_negative(-1, &page[i]._mapcount))
1226 nr++;
1227 }
1228 } else {
1229 nr = HPAGE_PMD_NR;
1230 }
1231
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001232 if (unlikely(PageMlocked(page)))
1233 clear_page_mlock(page);
1234
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001235 if (nr) {
Mel Gorman4b9d0fa2016-07-28 15:46:17 -07001236 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001237 deferred_split_huge_page(page);
1238 }
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001239}
1240
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241/**
1242 * page_remove_rmap - take down pte mapping from a page
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001243 * @page: page to remove mapping from
1244 * @compound: uncharge the page as compound or small page
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -07001246 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247 */
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001248void page_remove_rmap(struct page *page, bool compound)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001250 if (!PageAnon(page))
1251 return page_remove_file_rmap(page, compound);
KAMEZAWA Hiroyuki89c06bd2012-03-21 16:34:25 -07001252
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001253 if (compound)
1254 return page_remove_anon_compound_rmap(page);
1255
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001256 /* page still mapped by someone else? */
1257 if (!atomic_add_negative(-1, &page->_mapcount))
Johannes Weiner8186eb62014-10-29 14:50:51 -07001258 return;
1259
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001260 /*
Jianyu Zhanbea04b02014-06-04 16:09:51 -07001261 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1262 * these counters are not modified in interrupt context, and
Jianyu Zhanbea04b02014-06-04 16:09:51 -07001263 * pte lock(a spinlock) is held, which implies preemption disabled.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001264 */
Mel Gorman4b9d0fa2016-07-28 15:46:17 -07001265 __dec_node_page_state(page, NR_ANON_MAPPED);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001266
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001267 if (unlikely(PageMlocked(page)))
1268 clear_page_mlock(page);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001269
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001270 if (PageTransCompound(page))
1271 deferred_split_huge_page(compound_head(page));
1272
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001273 /*
1274 * It would be tidy to reset the PageAnon mapping here,
1275 * but that might overwrite a racing page_add_anon_rmap
1276 * which increments mapcount after us but sets mapping
1277 * before us: so leave the reset to free_hot_cold_page,
1278 * and remember that it's only reliable while mapped.
1279 * Leaving it set also helps swapoff to reinstate ptes
1280 * faster for those pages still in swapcache.
1281 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282}
1283
Minchan Kim854e9ed2016-01-15 16:54:53 -08001284struct rmap_private {
1285 enum ttu_flags flags;
1286 int lazyfreed;
1287};
1288
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289/*
Joonsoo Kim52629502014-01-21 15:49:50 -08001290 * @arg: enum ttu_flags will be passed to this argument
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291 */
Kirill A. Shutemovac769502014-06-04 16:08:17 -07001292static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
Joonsoo Kim52629502014-01-21 15:49:50 -08001293 unsigned long address, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294{
1295 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001296 struct page_vma_mapped_walk pvmw = {
1297 .page = page,
1298 .vma = vma,
1299 .address = address,
1300 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 pte_t pteval;
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001302 struct page *subpage;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303 int ret = SWAP_AGAIN;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001304 struct rmap_private *rp = arg;
1305 enum ttu_flags flags = rp->flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306
Hugh Dickinsb87537d9e2015-11-05 18:49:33 -08001307 /* munlock has nothing to gain from examining un-locked vmas */
1308 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001309 return SWAP_AGAIN;
Hugh Dickinsb87537d9e2015-11-05 18:49:33 -08001310
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001311 if (flags & TTU_SPLIT_HUGE_PMD) {
1312 split_huge_pmd_address(vma, address,
1313 flags & TTU_MIGRATION, page);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001314 }
1315
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001316 while (page_vma_mapped_walk(&pvmw)) {
1317 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1318 address = pvmw.address;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001320 /* Unexpected PMD-mapped THP? */
1321 VM_BUG_ON_PAGE(!pvmw.pte, page);
1322
1323 /*
1324 * If the page is mlock()d, we cannot swap it out.
1325 * If it's recently referenced (perhaps page_referenced
1326 * skipped over this mm) then we should reactivate it.
1327 */
1328 if (!(flags & TTU_IGNORE_MLOCK)) {
1329 if (vma->vm_flags & VM_LOCKED) {
1330 /* PTE-mapped THP are never mlocked */
1331 if (!PageTransCompound(page)) {
1332 /*
1333 * Holding pte lock, we do *not* need
1334 * mmap_sem here
1335 */
1336 mlock_vma_page(page);
1337 }
1338 ret = SWAP_MLOCK;
1339 page_vma_mapped_walk_done(&pvmw);
1340 break;
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001341 }
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001342 if (flags & TTU_MUNLOCK)
1343 continue;
Hugh Dickinsb87537d9e2015-11-05 18:49:33 -08001344 }
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001345
1346 if (!(flags & TTU_IGNORE_ACCESS)) {
1347 if (ptep_clear_flush_young_notify(vma, address,
1348 pvmw.pte)) {
1349 ret = SWAP_FAIL;
1350 page_vma_mapped_walk_done(&pvmw);
1351 break;
1352 }
Nick Pigginb291f002008-10-18 20:26:44 -07001353 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001355 /* Nuke the page table entry. */
1356 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1357 if (should_defer_flush(mm, flags)) {
1358 /*
1359 * We clear the PTE but do not flush so potentially
1360 * a remote CPU could still be writing to the page.
1361 * If the entry was previously clean then the
1362 * architecture must guarantee that a clear->dirty
1363 * transition on a cached TLB entry is written through
1364 * and traps if the PTE is unmapped.
1365 */
1366 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
Mel Gorman72b252a2015-09-04 15:47:32 -07001367
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001368 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08001369 } else {
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001370 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1371 }
1372
1373 /* Move the dirty bit to the page. Now the pte is gone. */
1374 if (pte_dirty(pteval))
1375 set_page_dirty(page);
1376
1377 /* Update high watermark before we lower rss */
1378 update_hiwater_rss(mm);
1379
1380 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1381 if (PageHuge(page)) {
1382 int nr = 1 << compound_order(page);
1383 hugetlb_count_sub(nr, mm);
1384 } else {
1385 dec_mm_counter(mm, mm_counter(page));
1386 }
1387
1388 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1389 set_pte_at(mm, address, pvmw.pte, pteval);
1390 } else if (pte_unused(pteval)) {
1391 /*
1392 * The guest indicated that the page content is of no
1393 * interest anymore. Simply discard the pte, vmscan
1394 * will take care of the rest.
1395 */
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001396 dec_mm_counter(mm, mm_counter(page));
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001397 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1398 (flags & TTU_MIGRATION)) {
1399 swp_entry_t entry;
1400 pte_t swp_pte;
1401 /*
1402 * Store the pfn of the page in a special migration
1403 * pte. do_swap_page() will wait until the migration
1404 * pte is removed and then restart fault handling.
1405 */
1406 entry = make_migration_entry(subpage,
1407 pte_write(pteval));
1408 swp_pte = swp_entry_to_pte(entry);
1409 if (pte_soft_dirty(pteval))
1410 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1411 set_pte_at(mm, address, pvmw.pte, swp_pte);
1412 } else if (PageAnon(page)) {
1413 swp_entry_t entry = { .val = page_private(subpage) };
1414 pte_t swp_pte;
1415 /*
1416 * Store the swap location in the pte.
1417 * See handle_pte_fault() ...
1418 */
1419 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
Minchan Kim854e9ed2016-01-15 16:54:53 -08001420
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001421 if (!PageDirty(page) && (flags & TTU_LZFREE)) {
1422 /* It's a freeable page by MADV_FREE */
1423 dec_mm_counter(mm, MM_ANONPAGES);
1424 rp->lazyfreed++;
1425 goto discard;
1426 }
1427
1428 if (swap_duplicate(entry) < 0) {
1429 set_pte_at(mm, address, pvmw.pte, pteval);
1430 ret = SWAP_FAIL;
1431 page_vma_mapped_walk_done(&pvmw);
1432 break;
1433 }
1434 if (list_empty(&mm->mmlist)) {
1435 spin_lock(&mmlist_lock);
1436 if (list_empty(&mm->mmlist))
1437 list_add(&mm->mmlist, &init_mm.mmlist);
1438 spin_unlock(&mmlist_lock);
1439 }
Minchan Kim854e9ed2016-01-15 16:54:53 -08001440 dec_mm_counter(mm, MM_ANONPAGES);
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001441 inc_mm_counter(mm, MM_SWAPENTS);
1442 swp_pte = swp_entry_to_pte(entry);
1443 if (pte_soft_dirty(pteval))
1444 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1445 set_pte_at(mm, address, pvmw.pte, swp_pte);
1446 } else
1447 dec_mm_counter(mm, mm_counter_file(page));
Minchan Kim854e9ed2016-01-15 16:54:53 -08001448discard:
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001449 page_remove_rmap(subpage, PageHuge(page));
1450 put_page(page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001451 mmu_notifier_invalidate_page(mm, address);
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001452 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453 return ret;
1454}
1455
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001456bool is_vma_temporary_stack(struct vm_area_struct *vma)
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001457{
1458 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1459
1460 if (!maybe_stack)
1461 return false;
1462
1463 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1464 VM_STACK_INCOMPLETE_SETUP)
1465 return true;
1466
1467 return false;
1468}
1469
Joonsoo Kim52629502014-01-21 15:49:50 -08001470static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1471{
1472 return is_vma_temporary_stack(vma);
1473}
1474
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001475static int page_mapcount_is_zero(struct page *page)
Joonsoo Kim52629502014-01-21 15:49:50 -08001476{
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001477 return !total_mapcount(page);
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001478}
Joonsoo Kim52629502014-01-21 15:49:50 -08001479
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480/**
1481 * try_to_unmap - try to remove all page table mappings to a page
1482 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001483 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484 *
1485 * Tries to remove all the page table entries which are mapping this
1486 * page, used in the pageout path. Caller must hold the page lock.
1487 * Return values are:
1488 *
1489 * SWAP_SUCCESS - we succeeded in removing all mappings
1490 * SWAP_AGAIN - we missed a mapping, try again later
1491 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001492 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001494int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495{
1496 int ret;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001497 struct rmap_private rp = {
1498 .flags = flags,
1499 .lazyfreed = 0,
1500 };
1501
Joonsoo Kim52629502014-01-21 15:49:50 -08001502 struct rmap_walk_control rwc = {
1503 .rmap_one = try_to_unmap_one,
Minchan Kim854e9ed2016-01-15 16:54:53 -08001504 .arg = &rp,
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001505 .done = page_mapcount_is_zero,
Joonsoo Kim52629502014-01-21 15:49:50 -08001506 .anon_lock = page_lock_anon_vma_read,
1507 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508
Joonsoo Kim52629502014-01-21 15:49:50 -08001509 /*
1510 * During exec, a temporary VMA is setup and later moved.
1511 * The VMA is moved under the anon_vma lock but not the
1512 * page tables leading to a race where migration cannot
1513 * find the migration ptes. Rather than increasing the
1514 * locking requirements of exec(), migration skips
1515 * temporary VMAs until after exec() completes.
1516 */
Konstantin Khlebnikovdaa5ba72014-06-04 16:10:52 -07001517 if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page))
Joonsoo Kim52629502014-01-21 15:49:50 -08001518 rwc.invalid_vma = invalid_migration_vma;
1519
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001520 if (flags & TTU_RMAP_LOCKED)
1521 ret = rmap_walk_locked(page, &rwc);
1522 else
1523 ret = rmap_walk(page, &rwc);
Joonsoo Kim52629502014-01-21 15:49:50 -08001524
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001525 if (ret != SWAP_MLOCK && !page_mapcount(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526 ret = SWAP_SUCCESS;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001527 if (rp.lazyfreed && !PageDirty(page))
1528 ret = SWAP_LZFREE;
1529 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 return ret;
1531}
Nikita Danilov81b40822005-05-01 08:58:36 -07001532
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001533static int page_not_mapped(struct page *page)
1534{
1535 return !page_mapped(page);
1536};
1537
Nick Pigginb291f002008-10-18 20:26:44 -07001538/**
1539 * try_to_munlock - try to munlock a page
1540 * @page: the page to be munlocked
1541 *
1542 * Called from munlock code. Checks all of the VMAs mapping the page
1543 * to make sure nobody else has this page mlocked. The page will be
1544 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1545 *
1546 * Return values are:
1547 *
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001548 * SWAP_AGAIN - no vma is holding page mlocked, or,
Nick Pigginb291f002008-10-18 20:26:44 -07001549 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
Hugh Dickins5ad64682009-12-14 17:59:24 -08001550 * SWAP_FAIL - page cannot be located at present
Nick Pigginb291f002008-10-18 20:26:44 -07001551 * SWAP_MLOCK - page is now mlocked.
1552 */
1553int try_to_munlock(struct page *page)
1554{
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001555 int ret;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001556 struct rmap_private rp = {
1557 .flags = TTU_MUNLOCK,
1558 .lazyfreed = 0,
1559 };
1560
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001561 struct rmap_walk_control rwc = {
1562 .rmap_one = try_to_unmap_one,
Minchan Kim854e9ed2016-01-15 16:54:53 -08001563 .arg = &rp,
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001564 .done = page_not_mapped,
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001565 .anon_lock = page_lock_anon_vma_read,
1566
1567 };
1568
Sasha Levin309381fea2014-01-23 15:52:54 -08001569 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
Nick Pigginb291f002008-10-18 20:26:44 -07001570
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001571 ret = rmap_walk(page, &rwc);
1572 return ret;
Nick Pigginb291f002008-10-18 20:26:44 -07001573}
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001574
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001575void __put_anon_vma(struct anon_vma *anon_vma)
Rik van Riel76545062010-08-09 17:18:41 -07001576{
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001577 struct anon_vma *root = anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -07001578
Andrey Ryabinin624483f2014-06-06 19:09:30 +04001579 anon_vma_free(anon_vma);
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001580 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1581 anon_vma_free(root);
Rik van Riel76545062010-08-09 17:18:41 -07001582}
Rik van Riel76545062010-08-09 17:18:41 -07001583
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001584static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1585 struct rmap_walk_control *rwc)
Joonsoo Kimfaecd8d2014-01-21 15:49:46 -08001586{
1587 struct anon_vma *anon_vma;
1588
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001589 if (rwc->anon_lock)
1590 return rwc->anon_lock(page);
1591
Joonsoo Kimfaecd8d2014-01-21 15:49:46 -08001592 /*
1593 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1594 * because that depends on page_mapped(); but not all its usages
1595 * are holding mmap_sem. Users without mmap_sem are required to
1596 * take a reference count to prevent the anon_vma disappearing
1597 */
1598 anon_vma = page_anon_vma(page);
1599 if (!anon_vma)
1600 return NULL;
1601
1602 anon_vma_lock_read(anon_vma);
1603 return anon_vma;
1604}
1605
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001606/*
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001607 * rmap_walk_anon - do something to anonymous page using the object-based
1608 * rmap method
1609 * @page: the page to be handled
1610 * @rwc: control variable according to each walk type
1611 *
1612 * Find all the mappings of a page using the mapping pointer and the vma chains
1613 * contained in the anon_vma struct it points to.
1614 *
1615 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1616 * where the page was found will be held for write. So, we won't recheck
1617 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1618 * LOCKED.
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001619 */
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001620static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1621 bool locked)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001622{
1623 struct anon_vma *anon_vma;
Kirill A. Shutemova8fa41ad2017-02-24 14:57:54 -08001624 pgoff_t pgoff_start, pgoff_end;
Rik van Riel5beb4932010-03-05 13:42:07 -08001625 struct anon_vma_chain *avc;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001626 int ret = SWAP_AGAIN;
1627
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001628 if (locked) {
1629 anon_vma = page_anon_vma(page);
1630 /* anon_vma disappear under us? */
1631 VM_BUG_ON_PAGE(!anon_vma, page);
1632 } else {
1633 anon_vma = rmap_walk_anon_lock(page, rwc);
1634 }
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001635 if (!anon_vma)
1636 return ret;
Joonsoo Kimfaecd8d2014-01-21 15:49:46 -08001637
Kirill A. Shutemova8fa41ad2017-02-24 14:57:54 -08001638 pgoff_start = page_to_pgoff(page);
1639 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1640 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1641 pgoff_start, pgoff_end) {
Rik van Riel5beb4932010-03-05 13:42:07 -08001642 struct vm_area_struct *vma = avc->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001643 unsigned long address = vma_address(page, vma);
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001644
Andrea Arcangeliad126952015-11-05 18:49:07 -08001645 cond_resched();
1646
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001647 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1648 continue;
1649
Joonsoo Kim051ac832014-01-21 15:49:48 -08001650 ret = rwc->rmap_one(page, vma, address, rwc->arg);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001651 if (ret != SWAP_AGAIN)
1652 break;
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001653 if (rwc->done && rwc->done(page))
1654 break;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001655 }
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001656
1657 if (!locked)
1658 anon_vma_unlock_read(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001659 return ret;
1660}
1661
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001662/*
1663 * rmap_walk_file - do something to file page using the object-based rmap method
1664 * @page: the page to be handled
1665 * @rwc: control variable according to each walk type
1666 *
1667 * Find all the mappings of a page using the mapping pointer and the vma chains
1668 * contained in the address_space struct it points to.
1669 *
1670 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1671 * where the page was found will be held for write. So, we won't recheck
1672 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1673 * LOCKED.
1674 */
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001675static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1676 bool locked)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001677{
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001678 struct address_space *mapping = page_mapping(page);
Kirill A. Shutemova8fa41ad2017-02-24 14:57:54 -08001679 pgoff_t pgoff_start, pgoff_end;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001680 struct vm_area_struct *vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001681 int ret = SWAP_AGAIN;
1682
Joonsoo Kim9f326242014-01-21 15:49:53 -08001683 /*
1684 * The page lock not only makes sure that page->mapping cannot
1685 * suddenly be NULLified by truncation, it makes sure that the
1686 * structure at mapping cannot be freed and reused yet,
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08001687 * so we can safely take mapping->i_mmap_rwsem.
Joonsoo Kim9f326242014-01-21 15:49:53 -08001688 */
Sasha Levin81d1b092014-10-09 15:28:10 -07001689 VM_BUG_ON_PAGE(!PageLocked(page), page);
Joonsoo Kim9f326242014-01-21 15:49:53 -08001690
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001691 if (!mapping)
1692 return ret;
Davidlohr Bueso3dec0ba2014-12-12 16:54:27 -08001693
Kirill A. Shutemova8fa41ad2017-02-24 14:57:54 -08001694 pgoff_start = page_to_pgoff(page);
1695 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001696 if (!locked)
1697 i_mmap_lock_read(mapping);
Kirill A. Shutemova8fa41ad2017-02-24 14:57:54 -08001698 vma_interval_tree_foreach(vma, &mapping->i_mmap,
1699 pgoff_start, pgoff_end) {
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001700 unsigned long address = vma_address(page, vma);
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001701
Andrea Arcangeliad126952015-11-05 18:49:07 -08001702 cond_resched();
1703
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001704 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1705 continue;
1706
Joonsoo Kim051ac832014-01-21 15:49:48 -08001707 ret = rwc->rmap_one(page, vma, address, rwc->arg);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001708 if (ret != SWAP_AGAIN)
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001709 goto done;
1710 if (rwc->done && rwc->done(page))
1711 goto done;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001712 }
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001713
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001714done:
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001715 if (!locked)
1716 i_mmap_unlock_read(mapping);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001717 return ret;
1718}
1719
Joonsoo Kim051ac832014-01-21 15:49:48 -08001720int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001721{
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001722 if (unlikely(PageKsm(page)))
Joonsoo Kim051ac832014-01-21 15:49:48 -08001723 return rmap_walk_ksm(page, rwc);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001724 else if (PageAnon(page))
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001725 return rmap_walk_anon(page, rwc, false);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001726 else
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001727 return rmap_walk_file(page, rwc, false);
1728}
1729
1730/* Like rmap_walk, but caller holds relevant rmap lock */
1731int rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1732{
1733 /* no ksm support for now */
1734 VM_BUG_ON_PAGE(PageKsm(page), page);
1735 if (PageAnon(page))
1736 return rmap_walk_anon(page, rwc, true);
1737 else
1738 return rmap_walk_file(page, rwc, true);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001739}
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001740
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001741#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001742/*
1743 * The following three functions are for anonymous (private mapped) hugepages.
1744 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1745 * and no lru code, because we handle hugepages differently from common pages.
1746 */
1747static void __hugepage_set_anon_rmap(struct page *page,
1748 struct vm_area_struct *vma, unsigned long address, int exclusive)
1749{
1750 struct anon_vma *anon_vma = vma->anon_vma;
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001751
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001752 BUG_ON(!anon_vma);
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001753
1754 if (PageAnon(page))
1755 return;
1756 if (!exclusive)
1757 anon_vma = anon_vma->root;
1758
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001759 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1760 page->mapping = (struct address_space *) anon_vma;
1761 page->index = linear_page_index(vma, address);
1762}
1763
1764void hugepage_add_anon_rmap(struct page *page,
1765 struct vm_area_struct *vma, unsigned long address)
1766{
1767 struct anon_vma *anon_vma = vma->anon_vma;
1768 int first;
Naoya Horiguchia850ea32010-09-10 13:23:06 +09001769
1770 BUG_ON(!PageLocked(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001771 BUG_ON(!anon_vma);
Hugh Dickins5dbe0af2011-05-28 13:17:04 -07001772 /* address might be in next vma when migration races vma_adjust */
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001773 first = atomic_inc_and_test(compound_mapcount_ptr(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001774 if (first)
1775 __hugepage_set_anon_rmap(page, vma, address, 0);
1776}
1777
1778void hugepage_add_new_anon_rmap(struct page *page,
1779 struct vm_area_struct *vma, unsigned long address)
1780{
1781 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001782 atomic_set(compound_mapcount_ptr(page), 0);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001783 __hugepage_set_anon_rmap(page, vma, address, 1);
1784}
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001785#endif /* CONFIG_HUGETLB_PAGE */