blob: e4b713a6ed7ed791a3b8f5f30f1b464d85a3c9df [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
30 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
31 * swap_lock (in swap_duplicate, swap_info_get)
32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mapping->private_lock (in __set_page_dirty_buffers)
34 * 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/**
144 * anon_vma_prepare - attach an anon_vma to a memory region
145 * @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 *
151 * The common case will be that we already have one, but if
Figo.zhang23a07902010-12-27 15:14:06 +0100152 * not we either need to find an adjacent mapping that we
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700153 * can re-use the anon_vma from (very common when the only
154 * reason for splitting a vma has been mprotect()), or we
155 * allocate a new one.
156 *
157 * Anon-vma allocations are very subtle, because we may have
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000158 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700159 * and that may actually touch the spinlock even in the newly
160 * allocated vma (it depends on RCU to make sure that the
161 * anon_vma isn't actually destroyed).
162 *
163 * As a result, we need to do proper anon_vma locking even
164 * for the new allocation. At the same time, we do not want
165 * to do any locking for the common case of already having
166 * an anon_vma.
167 *
168 * This must be called with the mmap_sem held for reading.
169 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170int anon_vma_prepare(struct vm_area_struct *vma)
171{
172 struct anon_vma *anon_vma = vma->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800173 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174
175 might_sleep();
176 if (unlikely(!anon_vma)) {
177 struct mm_struct *mm = vma->vm_mm;
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700178 struct anon_vma *allocated;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179
Linus Torvaldsdd347392011-06-17 19:05:36 -0700180 avc = anon_vma_chain_alloc(GFP_KERNEL);
Rik van Riel5beb4932010-03-05 13:42:07 -0800181 if (!avc)
182 goto out_enomem;
183
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 anon_vma = find_mergeable_anon_vma(vma);
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700185 allocated = NULL;
186 if (!anon_vma) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187 anon_vma = anon_vma_alloc();
188 if (unlikely(!anon_vma))
Rik van Riel5beb4932010-03-05 13:42:07 -0800189 goto out_enomem_free_avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 allocated = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 }
192
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000193 anon_vma_lock_write(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194 /* page_table_lock to protect against threads */
195 spin_lock(&mm->page_table_lock);
196 if (likely(!vma->anon_vma)) {
197 vma->anon_vma = anon_vma;
Kautuk Consul6583a842012-03-21 16:34:01 -0700198 anon_vma_chain_link(vma, avc, anon_vma);
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800199 /* vma reference or self-parent link for new root */
200 anon_vma->degree++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201 allocated = NULL;
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400202 avc = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203 }
204 spin_unlock(&mm->page_table_lock);
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -0800205 anon_vma_unlock_write(anon_vma);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400206
207 if (unlikely(allocated))
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700208 put_anon_vma(allocated);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400209 if (unlikely(avc))
Rik van Riel5beb4932010-03-05 13:42:07 -0800210 anon_vma_chain_free(avc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211 }
212 return 0;
Rik van Riel5beb4932010-03-05 13:42:07 -0800213
214 out_enomem_free_avc:
215 anon_vma_chain_free(avc);
216 out_enomem:
217 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218}
219
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700220/*
221 * This is a useful helper function for locking the anon_vma root as
222 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
223 * have the same vma.
224 *
225 * Such anon_vma's should have the same root, so you'd expect to see
226 * just a single mutex_lock for the whole traversal.
227 */
228static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
229{
230 struct anon_vma *new_root = anon_vma->root;
231 if (new_root != root) {
232 if (WARN_ON_ONCE(root))
Ingo Molnar5a505082012-12-02 19:56:46 +0000233 up_write(&root->rwsem);
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700234 root = new_root;
Ingo Molnar5a505082012-12-02 19:56:46 +0000235 down_write(&root->rwsem);
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700236 }
237 return root;
238}
239
240static inline void unlock_anon_vma_root(struct anon_vma *root)
241{
242 if (root)
Ingo Molnar5a505082012-12-02 19:56:46 +0000243 up_write(&root->rwsem);
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700244}
245
Rik van Riel5beb4932010-03-05 13:42:07 -0800246/*
247 * Attach the anon_vmas from src to dst.
248 * Returns 0 on success, -ENOMEM on failure.
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800249 *
250 * If dst->anon_vma is NULL this function tries to find and reuse existing
251 * anon_vma which has no vmas and only one child anon_vma. This prevents
252 * degradation of anon_vma hierarchy to endless linear chain in case of
253 * constantly forking task. On the other hand, an anon_vma with more than one
254 * child isn't reused even if there was no alive vma, thus rmap walker has a
255 * good chance of avoiding scanning the whole hierarchy when it searches where
256 * page is mapped.
Rik van Riel5beb4932010-03-05 13:42:07 -0800257 */
258int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259{
Rik van Riel5beb4932010-03-05 13:42:07 -0800260 struct anon_vma_chain *avc, *pavc;
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700261 struct anon_vma *root = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262
Linus Torvalds646d87b2010-04-11 17:15:03 -0700263 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700264 struct anon_vma *anon_vma;
265
Linus Torvaldsdd347392011-06-17 19:05:36 -0700266 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
267 if (unlikely(!avc)) {
268 unlock_anon_vma_root(root);
269 root = NULL;
270 avc = anon_vma_chain_alloc(GFP_KERNEL);
271 if (!avc)
272 goto enomem_failure;
273 }
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700274 anon_vma = pavc->anon_vma;
275 root = lock_anon_vma_root(root, anon_vma);
276 anon_vma_chain_link(dst, avc, anon_vma);
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800277
278 /*
279 * Reuse existing anon_vma if its degree lower than two,
280 * that means it has no vma and only one anon_vma child.
281 *
282 * Do not chose parent anon_vma, otherwise first child
283 * will always reuse it. Root anon_vma is never reused:
284 * it has self-parent reference and at least one child.
285 */
286 if (!dst->anon_vma && anon_vma != src->anon_vma &&
287 anon_vma->degree < 2)
288 dst->anon_vma = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 }
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800290 if (dst->anon_vma)
291 dst->anon_vma->degree++;
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700292 unlock_anon_vma_root(root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800293 return 0;
294
295 enomem_failure:
Leon Yu3fe89b32015-03-25 15:55:11 -0700296 /*
297 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
298 * decremented in unlink_anon_vmas().
299 * We can safely do this because callers of anon_vma_clone() don't care
300 * about dst->anon_vma if anon_vma_clone() failed.
301 */
302 dst->anon_vma = NULL;
Rik van Riel5beb4932010-03-05 13:42:07 -0800303 unlink_anon_vmas(dst);
304 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305}
306
Rik van Riel5beb4932010-03-05 13:42:07 -0800307/*
308 * Attach vma to its own anon_vma, as well as to the anon_vmas that
309 * the corresponding VMA in the parent process is attached to.
310 * Returns 0 on success, non-zero on failure.
311 */
312int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313{
Rik van Riel5beb4932010-03-05 13:42:07 -0800314 struct anon_vma_chain *avc;
315 struct anon_vma *anon_vma;
Daniel Forrestc4ea95d2014-12-02 15:59:42 -0800316 int error;
Rik van Riel5beb4932010-03-05 13:42:07 -0800317
318 /* Don't bother if the parent process has no anon_vma here. */
319 if (!pvma->anon_vma)
320 return 0;
321
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800322 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
323 vma->anon_vma = NULL;
324
Rik van Riel5beb4932010-03-05 13:42:07 -0800325 /*
326 * First, attach the new VMA to the parent VMA's anon_vmas,
327 * so rmap can find non-COWed pages in child processes.
328 */
Daniel Forrestc4ea95d2014-12-02 15:59:42 -0800329 error = anon_vma_clone(vma, pvma);
330 if (error)
331 return error;
Rik van Riel5beb4932010-03-05 13:42:07 -0800332
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800333 /* An existing anon_vma has been reused, all done then. */
334 if (vma->anon_vma)
335 return 0;
336
Rik van Riel5beb4932010-03-05 13:42:07 -0800337 /* Then add our own anon_vma. */
338 anon_vma = anon_vma_alloc();
339 if (!anon_vma)
340 goto out_error;
Linus Torvaldsdd347392011-06-17 19:05:36 -0700341 avc = anon_vma_chain_alloc(GFP_KERNEL);
Rik van Riel5beb4932010-03-05 13:42:07 -0800342 if (!avc)
343 goto out_error_free_anon_vma;
Rik van Riel5c341ee12010-08-09 17:18:39 -0700344
345 /*
346 * The root anon_vma's spinlock is the lock actually used when we
347 * lock any of the anon_vmas in this anon_vma tree.
348 */
349 anon_vma->root = pvma->anon_vma->root;
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800350 anon_vma->parent = pvma->anon_vma;
Rik van Riel76545062010-08-09 17:18:41 -0700351 /*
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700352 * With refcounts, an anon_vma can stay around longer than the
353 * process it belongs to. The root anon_vma needs to be pinned until
354 * this anon_vma is freed, because the lock lives in the root.
Rik van Riel76545062010-08-09 17:18:41 -0700355 */
356 get_anon_vma(anon_vma->root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800357 /* Mark this anon_vma as the one where our new (COWed) pages go. */
358 vma->anon_vma = anon_vma;
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000359 anon_vma_lock_write(anon_vma);
Rik van Riel5c341ee12010-08-09 17:18:39 -0700360 anon_vma_chain_link(vma, avc, anon_vma);
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800361 anon_vma->parent->degree++;
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -0800362 anon_vma_unlock_write(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800363
364 return 0;
365
366 out_error_free_anon_vma:
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700367 put_anon_vma(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800368 out_error:
Rik van Riel4946d542010-04-05 12:13:33 -0400369 unlink_anon_vmas(vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800370 return -ENOMEM;
371}
372
Rik van Riel5beb4932010-03-05 13:42:07 -0800373void unlink_anon_vmas(struct vm_area_struct *vma)
374{
375 struct anon_vma_chain *avc, *next;
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200376 struct anon_vma *root = NULL;
Rik van Riel5beb4932010-03-05 13:42:07 -0800377
Rik van Riel5c341ee12010-08-09 17:18:39 -0700378 /*
379 * Unlink each anon_vma chained to the VMA. This list is ordered
380 * from newest to oldest, ensuring the root anon_vma gets freed last.
381 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800382 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200383 struct anon_vma *anon_vma = avc->anon_vma;
384
385 root = lock_anon_vma_root(root, anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -0700386 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200387
388 /*
389 * Leave empty anon_vmas on the list - we'll need
390 * to free them outside the lock.
391 */
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800392 if (RB_EMPTY_ROOT(&anon_vma->rb_root)) {
393 anon_vma->parent->degree--;
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200394 continue;
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800395 }
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200396
397 list_del(&avc->same_vma);
398 anon_vma_chain_free(avc);
399 }
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800400 if (vma->anon_vma)
401 vma->anon_vma->degree--;
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200402 unlock_anon_vma_root(root);
403
404 /*
405 * Iterate the list once more, it now only contains empty and unlinked
406 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
Ingo Molnar5a505082012-12-02 19:56:46 +0000407 * needing to write-acquire the anon_vma->root->rwsem.
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200408 */
409 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
410 struct anon_vma *anon_vma = avc->anon_vma;
411
Konstantin Khlebnikove4c58002016-05-19 17:11:46 -0700412 VM_WARN_ON(anon_vma->degree);
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200413 put_anon_vma(anon_vma);
414
Rik van Riel5beb4932010-03-05 13:42:07 -0800415 list_del(&avc->same_vma);
416 anon_vma_chain_free(avc);
417 }
418}
419
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700420static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421{
Christoph Lametera35afb82007-05-16 22:10:57 -0700422 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423
Ingo Molnar5a505082012-12-02 19:56:46 +0000424 init_rwsem(&anon_vma->rwsem);
Peter Zijlstra83813262011-03-22 16:32:48 -0700425 atomic_set(&anon_vma->refcount, 0);
Michel Lespinassebf181b92012-10-08 16:31:39 -0700426 anon_vma->rb_root = RB_ROOT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427}
428
429void __init anon_vma_init(void)
430{
431 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Vladimir Davydov5d097052016-01-14 15:18:21 -0800432 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
433 anon_vma_ctor);
434 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
435 SLAB_PANIC|SLAB_ACCOUNT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436}
437
438/*
Peter Zijlstra6111e4c2011-05-24 17:12:08 -0700439 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
440 *
441 * Since there is no serialization what so ever against page_remove_rmap()
442 * the best this function can do is return a locked anon_vma that might
443 * have been relevant to this page.
444 *
445 * The page might have been remapped to a different anon_vma or the anon_vma
446 * returned may already be freed (and even reused).
447 *
Peter Zijlstrabc658c92011-05-29 10:33:44 +0200448 * In case it was remapped to a different anon_vma, the new anon_vma will be a
449 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
450 * ensure that any anon_vma obtained from the page will still be valid for as
451 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
452 *
Peter Zijlstra6111e4c2011-05-24 17:12:08 -0700453 * All users of this function must be very careful when walking the anon_vma
454 * chain and verify that the page in question is indeed mapped in it
455 * [ something equivalent to page_mapped_in_vma() ].
456 *
457 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
458 * that the anon_vma pointer from page->mapping is valid if there is a
459 * mapcount, we can dereference the anon_vma after observing those.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700461struct anon_vma *page_get_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462{
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700463 struct anon_vma *anon_vma = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 unsigned long anon_mapping;
465
466 rcu_read_lock();
Jason Low4db0c3c2015-04-15 16:14:08 -0700467 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800468 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 goto out;
470 if (!page_mapped(page))
471 goto out;
472
473 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700474 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
475 anon_vma = NULL;
476 goto out;
477 }
Hugh Dickinsf18194272010-08-25 23:12:54 -0700478
479 /*
480 * If this page is still mapped, then its anon_vma cannot have been
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700481 * freed. But if it has been unmapped, we have no security against the
482 * anon_vma structure being freed and reused (for another anon_vma:
483 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
484 * above cannot corrupt).
Hugh Dickinsf18194272010-08-25 23:12:54 -0700485 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700486 if (!page_mapped(page)) {
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700487 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700488 put_anon_vma(anon_vma);
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700489 return NULL;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700490 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491out:
492 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700493
494 return anon_vma;
495}
496
Peter Zijlstra88c22082011-05-24 17:12:13 -0700497/*
498 * Similar to page_get_anon_vma() except it locks the anon_vma.
499 *
500 * Its a little more complex as it tries to keep the fast path to a single
501 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
502 * reference like with page_get_anon_vma() and then block on the mutex.
503 */
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000504struct anon_vma *page_lock_anon_vma_read(struct page *page)
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700505{
Peter Zijlstra88c22082011-05-24 17:12:13 -0700506 struct anon_vma *anon_vma = NULL;
Hugh Dickinseee0f252011-05-28 13:20:21 -0700507 struct anon_vma *root_anon_vma;
Peter Zijlstra88c22082011-05-24 17:12:13 -0700508 unsigned long anon_mapping;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700509
Peter Zijlstra88c22082011-05-24 17:12:13 -0700510 rcu_read_lock();
Jason Low4db0c3c2015-04-15 16:14:08 -0700511 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700512 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
513 goto out;
514 if (!page_mapped(page))
515 goto out;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700516
Peter Zijlstra88c22082011-05-24 17:12:13 -0700517 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Jason Low4db0c3c2015-04-15 16:14:08 -0700518 root_anon_vma = READ_ONCE(anon_vma->root);
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000519 if (down_read_trylock(&root_anon_vma->rwsem)) {
Peter Zijlstra88c22082011-05-24 17:12:13 -0700520 /*
Hugh Dickinseee0f252011-05-28 13:20:21 -0700521 * If the page is still mapped, then this anon_vma is still
522 * its anon_vma, and holding the mutex ensures that it will
Peter Zijlstrabc658c92011-05-29 10:33:44 +0200523 * not go away, see anon_vma_free().
Peter Zijlstra88c22082011-05-24 17:12:13 -0700524 */
Hugh Dickinseee0f252011-05-28 13:20:21 -0700525 if (!page_mapped(page)) {
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000526 up_read(&root_anon_vma->rwsem);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700527 anon_vma = NULL;
528 }
529 goto out;
530 }
531
532 /* trylock failed, we got to sleep */
533 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
534 anon_vma = NULL;
535 goto out;
536 }
537
538 if (!page_mapped(page)) {
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700539 rcu_read_unlock();
Peter Zijlstra88c22082011-05-24 17:12:13 -0700540 put_anon_vma(anon_vma);
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700541 return NULL;
Peter Zijlstra88c22082011-05-24 17:12:13 -0700542 }
543
544 /* we pinned the anon_vma, its safe to sleep */
545 rcu_read_unlock();
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000546 anon_vma_lock_read(anon_vma);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700547
548 if (atomic_dec_and_test(&anon_vma->refcount)) {
549 /*
550 * Oops, we held the last refcount, release the lock
551 * and bail -- can't simply use put_anon_vma() because
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000552 * we'll deadlock on the anon_vma_lock_write() recursion.
Peter Zijlstra88c22082011-05-24 17:12:13 -0700553 */
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000554 anon_vma_unlock_read(anon_vma);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700555 __put_anon_vma(anon_vma);
556 anon_vma = NULL;
557 }
558
559 return anon_vma;
560
561out:
562 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700563 return anon_vma;
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800564}
565
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000566void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800567{
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000568 anon_vma_unlock_read(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569}
570
Mel Gorman72b252a2015-09-04 15:47:32 -0700571#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
Mel Gorman72b252a2015-09-04 15:47:32 -0700572/*
573 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
574 * important if a PTE was dirty when it was unmapped that it's flushed
575 * before any IO is initiated on the page to prevent lost writes. Similarly,
576 * it must be flushed before freeing to prevent data leakage.
577 */
578void try_to_unmap_flush(void)
579{
580 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
581 int cpu;
582
583 if (!tlb_ubc->flush_required)
584 return;
585
586 cpu = get_cpu();
587
Nadav Amit858eaaa72016-04-01 14:31:26 -0700588 if (cpumask_test_cpu(cpu, &tlb_ubc->cpumask)) {
589 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
590 local_flush_tlb();
591 trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
Mel Gorman72b252a2015-09-04 15:47:32 -0700592 }
Nadav Amit858eaaa72016-04-01 14:31:26 -0700593
594 if (cpumask_any_but(&tlb_ubc->cpumask, cpu) < nr_cpu_ids)
595 flush_tlb_others(&tlb_ubc->cpumask, NULL, 0, TLB_FLUSH_ALL);
Mel Gorman72b252a2015-09-04 15:47:32 -0700596 cpumask_clear(&tlb_ubc->cpumask);
597 tlb_ubc->flush_required = false;
Mel Gormand950c942015-09-04 15:47:35 -0700598 tlb_ubc->writable = false;
Mel Gorman72b252a2015-09-04 15:47:32 -0700599 put_cpu();
600}
601
Mel Gormand950c942015-09-04 15:47:35 -0700602/* Flush iff there are potentially writable TLB entries that can race with IO */
603void try_to_unmap_flush_dirty(void)
604{
605 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
606
607 if (tlb_ubc->writable)
608 try_to_unmap_flush();
609}
610
Mel Gorman72b252a2015-09-04 15:47:32 -0700611static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
Mel Gormand950c942015-09-04 15:47:35 -0700612 struct page *page, bool writable)
Mel Gorman72b252a2015-09-04 15:47:32 -0700613{
614 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
615
616 cpumask_or(&tlb_ubc->cpumask, &tlb_ubc->cpumask, mm_cpumask(mm));
617 tlb_ubc->flush_required = true;
Mel Gormand950c942015-09-04 15:47:35 -0700618
619 /*
620 * If the PTE was dirty then it's best to assume it's writable. The
621 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
622 * before the page is queued for IO.
623 */
624 if (writable)
625 tlb_ubc->writable = true;
Mel Gorman72b252a2015-09-04 15:47:32 -0700626}
627
628/*
629 * Returns true if the TLB flush should be deferred to the end of a batch of
630 * unmap operations to reduce IPIs.
631 */
632static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
633{
634 bool should_defer = false;
635
636 if (!(flags & TTU_BATCH_FLUSH))
637 return false;
638
639 /* If remote CPUs need to be flushed then defer batch the flush */
640 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
641 should_defer = true;
642 put_cpu();
643
644 return should_defer;
645}
646#else
647static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
Mel Gormand950c942015-09-04 15:47:35 -0700648 struct page *page, bool writable)
Mel Gorman72b252a2015-09-04 15:47:32 -0700649{
650}
651
652static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
653{
654 return false;
655}
656#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
657
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700659 * At what user virtual address is page expected in vma?
Naoya Horiguchiab941e02010-05-11 14:06:55 -0700660 * Caller should check the page is actually part of the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661 */
662unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
663{
Michel Lespinasse86c2ad12012-10-08 16:31:42 -0700664 unsigned long address;
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700665 if (PageAnon(page)) {
Hugh Dickins4829b902010-10-02 17:46:06 -0700666 struct anon_vma *page__anon_vma = page_anon_vma(page);
667 /*
668 * Note: swapoff's unuse_vma() is more efficient with this
669 * check, and needs it to match anon_vma when KSM is active.
670 */
671 if (!vma->anon_vma || !page__anon_vma ||
672 vma->anon_vma->root != page__anon_vma->root)
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700673 return -EFAULT;
Kirill A. Shutemov27ba0642015-02-10 14:09:59 -0800674 } else if (page->mapping) {
675 if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 return -EFAULT;
677 } else
678 return -EFAULT;
Michel Lespinasse86c2ad12012-10-08 16:31:42 -0700679 address = __vma_address(page, vma);
680 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
681 return -EFAULT;
682 return address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683}
684
Bob Liu62190492012-12-11 16:00:37 -0800685pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
686{
687 pgd_t *pgd;
688 pud_t *pud;
689 pmd_t *pmd = NULL;
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -0700690 pmd_t pmde;
Bob Liu62190492012-12-11 16:00:37 -0800691
692 pgd = pgd_offset(mm, address);
693 if (!pgd_present(*pgd))
694 goto out;
695
696 pud = pud_offset(pgd, address);
697 if (!pud_present(*pud))
698 goto out;
699
700 pmd = pmd_offset(pud, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -0700701 /*
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -0700702 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -0700703 * without holding anon_vma lock for write. So when looking for a
704 * genuine pmde (in which to find pte), test present and !THP together.
705 */
Christian Borntraegere37c6982014-12-07 21:41:33 +0100706 pmde = *pmd;
707 barrier();
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -0700708 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
Bob Liu62190492012-12-11 16:00:37 -0800709 pmd = NULL;
710out:
711 return pmd;
712}
713
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700715 * Check that @page is mapped at @address into @mm.
716 *
Nick Piggin479db0b2008-08-20 14:09:18 -0700717 * If @sync is false, page_check_address may perform a racy check to avoid
718 * the page table lock when the pte is not present (helpful when reclaiming
719 * highly shared pages).
720 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700721 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700722 */
Namhyung Kime9a81a82010-10-26 14:22:01 -0700723pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
Nick Piggin479db0b2008-08-20 14:09:18 -0700724 unsigned long address, spinlock_t **ptlp, int sync)
Nikita Danilov81b40822005-05-01 08:58:36 -0700725{
Nikita Danilov81b40822005-05-01 08:58:36 -0700726 pmd_t *pmd;
727 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700728 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700729
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900730 if (unlikely(PageHuge(page))) {
Jianguo Wu98398c32013-12-18 17:08:59 -0800731 /* when pud is not present, pte will be NULL */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900732 pte = huge_pte_offset(mm, address);
Jianguo Wu98398c32013-12-18 17:08:59 -0800733 if (!pte)
734 return NULL;
735
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -0800736 ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900737 goto check;
738 }
739
Bob Liu62190492012-12-11 16:00:37 -0800740 pmd = mm_find_pmd(mm, address);
741 if (!pmd)
Hugh Dickinsc0718802005-10-29 18:16:31 -0700742 return NULL;
743
Hugh Dickinsc0718802005-10-29 18:16:31 -0700744 pte = pte_offset_map(pmd, address);
745 /* Make a quick check before getting the lock */
Nick Piggin479db0b2008-08-20 14:09:18 -0700746 if (!sync && !pte_present(*pte)) {
Hugh Dickinsc0718802005-10-29 18:16:31 -0700747 pte_unmap(pte);
748 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700749 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700750
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700751 ptl = pte_lockptr(mm, pmd);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900752check:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700753 spin_lock(ptl);
754 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
755 *ptlp = ptl;
756 return pte;
757 }
758 pte_unmap_unlock(pte, ptl);
759 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700760}
761
Nick Pigginb291f002008-10-18 20:26:44 -0700762/**
763 * page_mapped_in_vma - check whether a page is really mapped in a VMA
764 * @page: the page to test
765 * @vma: the VMA to test
766 *
767 * Returns 1 if the page is mapped into the page tables of the VMA, 0
768 * if the page is not mapped into the page tables of this VMA. Only
769 * valid for normal file or anonymous VMAs.
770 */
Andi Kleen6a460792009-09-16 11:50:15 +0200771int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
Nick Pigginb291f002008-10-18 20:26:44 -0700772{
773 unsigned long address;
774 pte_t *pte;
775 spinlock_t *ptl;
776
Michel Lespinasse86c2ad12012-10-08 16:31:42 -0700777 address = __vma_address(page, vma);
778 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
Nick Pigginb291f002008-10-18 20:26:44 -0700779 return 0;
780 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
781 if (!pte) /* the page is not in this mm */
782 return 0;
783 pte_unmap_unlock(pte, ptl);
784
785 return 1;
786}
787
Vladimir Davydov8749cfe2016-01-15 16:54:45 -0800788#ifdef CONFIG_TRANSPARENT_HUGEPAGE
789/*
790 * Check that @page is mapped at @address into @mm. In contrast to
791 * page_check_address(), this function can handle transparent huge pages.
792 *
793 * On success returns true with pte mapped and locked. For PMD-mapped
794 * transparent huge pages *@ptep is set to NULL.
795 */
796bool page_check_address_transhuge(struct page *page, struct mm_struct *mm,
797 unsigned long address, pmd_t **pmdp,
798 pte_t **ptep, spinlock_t **ptlp)
799{
800 pgd_t *pgd;
801 pud_t *pud;
802 pmd_t *pmd;
803 pte_t *pte;
804 spinlock_t *ptl;
805
806 if (unlikely(PageHuge(page))) {
807 /* when pud is not present, pte will be NULL */
808 pte = huge_pte_offset(mm, address);
809 if (!pte)
810 return false;
811
812 ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
813 pmd = NULL;
814 goto check_pte;
815 }
816
817 pgd = pgd_offset(mm, address);
818 if (!pgd_present(*pgd))
819 return false;
820 pud = pud_offset(pgd, address);
821 if (!pud_present(*pud))
822 return false;
823 pmd = pmd_offset(pud, address);
824
825 if (pmd_trans_huge(*pmd)) {
826 ptl = pmd_lock(mm, pmd);
827 if (!pmd_present(*pmd))
828 goto unlock_pmd;
829 if (unlikely(!pmd_trans_huge(*pmd))) {
830 spin_unlock(ptl);
831 goto map_pte;
832 }
833
834 if (pmd_page(*pmd) != page)
835 goto unlock_pmd;
836
837 pte = NULL;
838 goto found;
839unlock_pmd:
840 spin_unlock(ptl);
841 return false;
842 } else {
843 pmd_t pmde = *pmd;
844
845 barrier();
846 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
847 return false;
848 }
849map_pte:
850 pte = pte_offset_map(pmd, address);
851 if (!pte_present(*pte)) {
852 pte_unmap(pte);
853 return false;
854 }
855
856 ptl = pte_lockptr(mm, pmd);
857check_pte:
858 spin_lock(ptl);
859
860 if (!pte_present(*pte)) {
861 pte_unmap_unlock(pte, ptl);
862 return false;
863 }
864
865 /* THP can be referenced by any subpage */
866 if (pte_pfn(*pte) - page_to_pfn(page) >= hpage_nr_pages(page)) {
867 pte_unmap_unlock(pte, ptl);
868 return false;
869 }
870found:
871 *ptep = pte;
872 *pmdp = pmd;
873 *ptlp = ptl;
874 return true;
875}
876#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
877
Joonsoo Kim9f326242014-01-21 15:49:53 -0800878struct page_referenced_arg {
879 int mapcount;
880 int referenced;
881 unsigned long vm_flags;
882 struct mem_cgroup *memcg;
883};
Nikita Danilov81b40822005-05-01 08:58:36 -0700884/*
Joonsoo Kim9f326242014-01-21 15:49:53 -0800885 * arg: page_referenced_arg will be passed
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 */
Kirill A. Shutemovac769502014-06-04 16:08:17 -0700887static int page_referenced_one(struct page *page, struct vm_area_struct *vma,
Joonsoo Kim9f326242014-01-21 15:49:53 -0800888 unsigned long address, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889{
890 struct mm_struct *mm = vma->vm_mm;
Joonsoo Kim9f326242014-01-21 15:49:53 -0800891 struct page_referenced_arg *pra = arg;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800892 pmd_t *pmd;
893 pte_t *pte;
Vladimir Davydov8749cfe2016-01-15 16:54:45 -0800894 spinlock_t *ptl;
895 int referenced = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896
Vladimir Davydov8749cfe2016-01-15 16:54:45 -0800897 if (!page_check_address_transhuge(page, mm, address, &pmd, &pte, &ptl))
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800898 return SWAP_AGAIN;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800899
900 if (vma->vm_flags & VM_LOCKED) {
Vladimir Davydov8749cfe2016-01-15 16:54:45 -0800901 if (pte)
902 pte_unmap(pte);
903 spin_unlock(ptl);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800904 pra->vm_flags |= VM_LOCKED;
905 return SWAP_FAIL; /* To break the loop */
906 }
907
Vladimir Davydov8749cfe2016-01-15 16:54:45 -0800908 if (pte) {
909 if (ptep_clear_flush_young_notify(vma, address, pte)) {
910 /*
911 * Don't treat a reference through a sequentially read
912 * mapping as such. If the page has been used in
913 * another mapping, we will catch it; if this other
914 * mapping is already gone, the unmap path will have
915 * set PG_referenced or activated the page.
916 */
917 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
918 referenced++;
919 }
920 pte_unmap(pte);
921 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
922 if (pmdp_clear_flush_young_notify(vma, address, pmd))
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800923 referenced++;
Vladimir Davydov8749cfe2016-01-15 16:54:45 -0800924 } else {
925 /* unexpected pmd-mapped page? */
926 WARN_ON_ONCE(1);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800927 }
Vladimir Davydov8749cfe2016-01-15 16:54:45 -0800928 spin_unlock(ptl);
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800929
Vladimir Davydov33c3fc72015-09-09 15:35:45 -0700930 if (referenced)
931 clear_page_idle(page);
932 if (test_and_clear_page_young(page))
933 referenced++;
934
Joonsoo Kim9f326242014-01-21 15:49:53 -0800935 if (referenced) {
936 pra->referenced++;
937 pra->vm_flags |= vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800939
Joonsoo Kim9f326242014-01-21 15:49:53 -0800940 pra->mapcount--;
941 if (!pra->mapcount)
942 return SWAP_SUCCESS; /* To break the loop */
943
944 return SWAP_AGAIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945}
946
Joonsoo Kim9f326242014-01-21 15:49:53 -0800947static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948{
Joonsoo Kim9f326242014-01-21 15:49:53 -0800949 struct page_referenced_arg *pra = arg;
950 struct mem_cgroup *memcg = pra->memcg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951
Joonsoo Kim9f326242014-01-21 15:49:53 -0800952 if (!mm_match_cgroup(vma->vm_mm, memcg))
953 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954
Joonsoo Kim9f326242014-01-21 15:49:53 -0800955 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956}
957
958/**
959 * page_referenced - test if the page was referenced
960 * @page: the page to test
961 * @is_locked: caller holds lock on the page
Johannes Weiner72835c82012-01-12 17:18:32 -0800962 * @memcg: target memory cgroup
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700963 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 *
965 * Quick test_and_clear_referenced for all mappings to a page,
966 * returns the number of ptes which referenced the page.
967 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700968int page_referenced(struct page *page,
969 int is_locked,
Johannes Weiner72835c82012-01-12 17:18:32 -0800970 struct mem_cgroup *memcg,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700971 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972{
Joonsoo Kim9f326242014-01-21 15:49:53 -0800973 int ret;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800974 int we_locked = 0;
Joonsoo Kim9f326242014-01-21 15:49:53 -0800975 struct page_referenced_arg pra = {
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800976 .mapcount = total_mapcount(page),
Joonsoo Kim9f326242014-01-21 15:49:53 -0800977 .memcg = memcg,
978 };
979 struct rmap_walk_control rwc = {
980 .rmap_one = page_referenced_one,
981 .arg = (void *)&pra,
982 .anon_lock = page_lock_anon_vma_read,
983 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700985 *vm_flags = 0;
Joonsoo Kim9f326242014-01-21 15:49:53 -0800986 if (!page_mapped(page))
987 return 0;
988
989 if (!page_rmapping(page))
990 return 0;
991
992 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
993 we_locked = trylock_page(page);
994 if (!we_locked)
995 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 }
Joonsoo Kim9f326242014-01-21 15:49:53 -0800997
998 /*
999 * If we are reclaiming on behalf of a cgroup, skip
1000 * counting on behalf of references from different
1001 * cgroups
1002 */
1003 if (memcg) {
1004 rwc.invalid_vma = invalid_page_referenced_vma;
1005 }
1006
1007 ret = rmap_walk(page, &rwc);
1008 *vm_flags = pra.vm_flags;
1009
1010 if (we_locked)
1011 unlock_page(page);
1012
1013 return pra.referenced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014}
1015
Hugh Dickins1cb17292009-12-14 17:59:01 -08001016static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
Joonsoo Kim9853a402014-01-21 15:49:55 -08001017 unsigned long address, void *arg)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001018{
1019 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +01001020 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001021 spinlock_t *ptl;
1022 int ret = 0;
Joonsoo Kim9853a402014-01-21 15:49:55 -08001023 int *cleaned = arg;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001024
Nick Piggin479db0b2008-08-20 14:09:18 -07001025 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001026 if (!pte)
1027 goto out;
1028
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +01001029 if (pte_dirty(*pte) || pte_write(*pte)) {
1030 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001031
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +01001032 flush_cache_page(vma, address, pte_pfn(*pte));
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001033 entry = ptep_clear_flush(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +01001034 entry = pte_wrprotect(entry);
1035 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -08001036 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +01001037 ret = 1;
1038 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001039
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001040 pte_unmap_unlock(pte, ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001041
Joonsoo Kim9853a402014-01-21 15:49:55 -08001042 if (ret) {
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001043 mmu_notifier_invalidate_page(mm, address);
Joonsoo Kim9853a402014-01-21 15:49:55 -08001044 (*cleaned)++;
1045 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001046out:
Joonsoo Kim9853a402014-01-21 15:49:55 -08001047 return SWAP_AGAIN;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001048}
1049
Joonsoo Kim9853a402014-01-21 15:49:55 -08001050static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001051{
Joonsoo Kim9853a402014-01-21 15:49:55 -08001052 if (vma->vm_flags & VM_SHARED)
Fengguang Wu871beb82014-01-23 15:53:41 -08001053 return false;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001054
Fengguang Wu871beb82014-01-23 15:53:41 -08001055 return true;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001056}
1057
1058int page_mkclean(struct page *page)
1059{
Joonsoo Kim9853a402014-01-21 15:49:55 -08001060 int cleaned = 0;
1061 struct address_space *mapping;
1062 struct rmap_walk_control rwc = {
1063 .arg = (void *)&cleaned,
1064 .rmap_one = page_mkclean_one,
1065 .invalid_vma = invalid_mkclean_vma,
1066 };
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001067
1068 BUG_ON(!PageLocked(page));
1069
Joonsoo Kim9853a402014-01-21 15:49:55 -08001070 if (!page_mapped(page))
1071 return 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001072
Joonsoo Kim9853a402014-01-21 15:49:55 -08001073 mapping = page_mapping(page);
1074 if (!mapping)
1075 return 0;
1076
1077 rmap_walk(page, &rwc);
1078
1079 return cleaned;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001080}
Jaya Kumar60b59be2007-05-08 00:37:37 -07001081EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001082
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083/**
Rik van Rielc44b6742010-03-05 13:42:09 -08001084 * page_move_anon_rmap - move a page to our anon_vma
1085 * @page: the page to move to our anon_vma
1086 * @vma: the vma the page belongs to
1087 * @address: the user virtual address mapped
1088 *
1089 * When a page belongs exclusively to one process after a COW event,
1090 * that page can be moved into the anon_vma that belongs to just that
1091 * process, so the rmap code will not search the parent or sibling
1092 * processes.
1093 */
1094void page_move_anon_rmap(struct page *page,
1095 struct vm_area_struct *vma, unsigned long address)
1096{
1097 struct anon_vma *anon_vma = vma->anon_vma;
1098
Sasha Levin309381fea2014-01-23 15:52:54 -08001099 VM_BUG_ON_PAGE(!PageLocked(page), page);
Sasha Levin81d1b092014-10-09 15:28:10 -07001100 VM_BUG_ON_VMA(!anon_vma, vma);
Kirill A. Shutemov0798d3c2016-05-27 14:27:38 -07001101 if (IS_ENABLED(CONFIG_DEBUG_VM) && PageTransHuge(page))
1102 address &= HPAGE_PMD_MASK;
Sasha Levin309381fea2014-01-23 15:52:54 -08001103 VM_BUG_ON_PAGE(page->index != linear_page_index(vma, address), page);
Rik van Rielc44b6742010-03-05 13:42:09 -08001104
1105 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
Vladimir Davydov414e2fb2015-06-24 16:56:56 -07001106 /*
1107 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1108 * simultaneously, so a concurrent reader (eg page_referenced()'s
1109 * PageAnon()) will not see one without the other.
1110 */
1111 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
Rik van Rielc44b6742010-03-05 13:42:09 -08001112}
1113
1114/**
Andi Kleen4e1c1972010-09-22 12:43:56 +02001115 * __page_set_anon_rmap - set up new anonymous rmap
1116 * @page: Page to add to rmap
1117 * @vma: VM area to add page to.
1118 * @address: User virtual address of the mapping
Rik van Riele8a03fe2010-04-14 17:59:28 -04001119 * @exclusive: the page is exclusively owned by the current process
Nick Piggin9617d952006-01-06 00:11:12 -08001120 */
1121static void __page_set_anon_rmap(struct page *page,
Rik van Riele8a03fe2010-04-14 17:59:28 -04001122 struct vm_area_struct *vma, unsigned long address, int exclusive)
Nick Piggin9617d952006-01-06 00:11:12 -08001123{
Rik van Riele8a03fe2010-04-14 17:59:28 -04001124 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -08001125
Rik van Riele8a03fe2010-04-14 17:59:28 -04001126 BUG_ON(!anon_vma);
Linus Torvaldsea900022010-04-12 12:44:29 -07001127
Andi Kleen4e1c1972010-09-22 12:43:56 +02001128 if (PageAnon(page))
1129 return;
1130
Linus Torvaldsea900022010-04-12 12:44:29 -07001131 /*
Rik van Riele8a03fe2010-04-14 17:59:28 -04001132 * If the page isn't exclusively mapped into this vma,
1133 * we must use the _oldest_ possible anon_vma for the
1134 * page mapping!
Linus Torvaldsea900022010-04-12 12:44:29 -07001135 */
Andi Kleen4e1c1972010-09-22 12:43:56 +02001136 if (!exclusive)
Andrea Arcangeli288468c2010-08-09 17:19:09 -07001137 anon_vma = anon_vma->root;
Linus Torvaldsea900022010-04-12 12:44:29 -07001138
Nick Piggin9617d952006-01-06 00:11:12 -08001139 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1140 page->mapping = (struct address_space *) anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -08001141 page->index = linear_page_index(vma, address);
Nick Piggin9617d952006-01-06 00:11:12 -08001142}
1143
1144/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -07001145 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -07001146 * @page: the page to add the mapping to
1147 * @vma: the vm area in which the mapping is added
1148 * @address: the user virtual address mapped
1149 */
1150static void __page_check_anon_rmap(struct page *page,
1151 struct vm_area_struct *vma, unsigned long address)
1152{
1153#ifdef CONFIG_DEBUG_VM
1154 /*
1155 * The page's anon-rmap details (mapping and index) are guaranteed to
1156 * be set up correctly at this point.
1157 *
1158 * We have exclusion against page_add_anon_rmap because the caller
1159 * always holds the page locked, except if called from page_dup_rmap,
1160 * in which case the page is already known to be setup.
1161 *
1162 * We have exclusion against page_add_new_anon_rmap because those pages
1163 * are initially only visible via the pagetables, and the pte is locked
1164 * over the call to page_add_new_anon_rmap.
1165 */
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001166 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001167 BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
Nick Pigginc97a9e12007-05-16 22:11:21 -07001168#endif
1169}
1170
1171/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172 * page_add_anon_rmap - add pte mapping to an anonymous page
1173 * @page: the page to add the mapping to
1174 * @vma: the vm area in which the mapping is added
1175 * @address: the user virtual address mapped
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001176 * @compound: charge the page as compound or small page
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 *
Hugh Dickins5ad64682009-12-14 17:59:24 -08001178 * The caller needs to hold the pte lock, and the page must be locked in
Hugh Dickins80e148222009-12-14 17:59:29 -08001179 * the anon_vma case: to serialize mapping,index checking after setting,
1180 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1181 * (but PageKsm is never downgraded to PageAnon).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 */
1183void page_add_anon_rmap(struct page *page,
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001184 struct vm_area_struct *vma, unsigned long address, bool compound)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185{
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001186 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
Rik van Rielad8c2ee2010-08-09 17:19:48 -07001187}
1188
1189/*
1190 * Special version of the above for do_swap_page, which often runs
1191 * into pages that are exclusively owned by the current process.
1192 * Everybody else should continue to use page_add_anon_rmap above.
1193 */
1194void do_page_add_anon_rmap(struct page *page,
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001195 struct vm_area_struct *vma, unsigned long address, int flags)
Rik van Rielad8c2ee2010-08-09 17:19:48 -07001196{
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001197 bool compound = flags & RMAP_COMPOUND;
1198 bool first;
1199
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001200 if (compound) {
1201 atomic_t *mapcount;
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001202 VM_BUG_ON_PAGE(!PageLocked(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001203 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1204 mapcount = compound_mapcount_ptr(page);
1205 first = atomic_inc_and_test(mapcount);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001206 } else {
1207 first = atomic_inc_and_test(&page->_mapcount);
1208 }
1209
Andrea Arcangeli79134172011-01-13 15:46:58 -08001210 if (first) {
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001211 int nr = compound ? hpage_nr_pages(page) : 1;
Jianyu Zhanbea04b02014-06-04 16:09:51 -07001212 /*
1213 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1214 * these counters are not modified in interrupt context, and
1215 * pte lock(a spinlock) is held, which implies preemption
1216 * disabled.
1217 */
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001218 if (compound) {
Andrea Arcangeli79134172011-01-13 15:46:58 -08001219 __inc_zone_page_state(page,
1220 NR_ANON_TRANSPARENT_HUGEPAGES);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001221 }
1222 __mod_zone_page_state(page_zone(page), NR_ANON_PAGES, nr);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001223 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001224 if (unlikely(PageKsm(page)))
1225 return;
1226
Sasha Levin309381fea2014-01-23 15:52:54 -08001227 VM_BUG_ON_PAGE(!PageLocked(page), page);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001228
Hugh Dickins5dbe0af2011-05-28 13:17:04 -07001229 /* address might be in next vma when migration races vma_adjust */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001230 if (first)
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001231 __page_set_anon_rmap(page, vma, address,
1232 flags & RMAP_EXCLUSIVE);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -07001233 else
Nick Pigginc97a9e12007-05-16 22:11:21 -07001234 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235}
1236
Randy Dunlap43d8eac2008-03-19 17:00:43 -07001237/**
Nick Piggin9617d952006-01-06 00:11:12 -08001238 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1239 * @page: the page to add the mapping to
1240 * @vma: the vm area in which the mapping is added
1241 * @address: the user virtual address mapped
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001242 * @compound: charge the page as compound or small page
Nick Piggin9617d952006-01-06 00:11:12 -08001243 *
1244 * Same as page_add_anon_rmap but must only be called on *new* pages.
1245 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -07001246 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -08001247 */
1248void page_add_new_anon_rmap(struct page *page,
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001249 struct vm_area_struct *vma, unsigned long address, bool compound)
Nick Piggin9617d952006-01-06 00:11:12 -08001250{
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001251 int nr = compound ? hpage_nr_pages(page) : 1;
1252
Sasha Levin81d1b092014-10-09 15:28:10 -07001253 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
Hugh Dickinsfa9949d2016-05-19 17:12:41 -07001254 __SetPageSwapBacked(page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001255 if (compound) {
1256 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001257 /* increment count (starts at -1) */
1258 atomic_set(compound_mapcount_ptr(page), 0);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001259 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001260 } else {
1261 /* Anon THP always mapped first with PMD */
1262 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1263 /* increment count (starts at -1) */
1264 atomic_set(&page->_mapcount, 0);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001265 }
1266 __mod_zone_page_state(page_zone(page), NR_ANON_PAGES, nr);
Rik van Riele8a03fe2010-04-14 17:59:28 -04001267 __page_set_anon_rmap(page, vma, address, 1);
Nick Piggin9617d952006-01-06 00:11:12 -08001268}
1269
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270/**
1271 * page_add_file_rmap - add pte mapping to a file page
1272 * @page: the page to add the mapping to
1273 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -07001274 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 */
1276void page_add_file_rmap(struct page *page)
1277{
Johannes Weiner62cccb82016-03-15 14:57:22 -07001278 lock_page_memcg(page);
Balbir Singhd69b0422009-06-17 16:26:34 -07001279 if (atomic_inc_and_test(&page->_mapcount)) {
Christoph Lameter65ba55f2006-06-30 01:55:34 -07001280 __inc_zone_page_state(page, NR_FILE_MAPPED);
Johannes Weiner62cccb82016-03-15 14:57:22 -07001281 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
Balbir Singhd69b0422009-06-17 16:26:34 -07001282 }
Johannes Weiner62cccb82016-03-15 14:57:22 -07001283 unlock_page_memcg(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284}
1285
Johannes Weiner8186eb62014-10-29 14:50:51 -07001286static void page_remove_file_rmap(struct page *page)
1287{
Johannes Weiner62cccb82016-03-15 14:57:22 -07001288 lock_page_memcg(page);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001289
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001290 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1291 if (unlikely(PageHuge(page))) {
1292 /* hugetlb pages are always mapped with pmds */
1293 atomic_dec(compound_mapcount_ptr(page));
1294 goto out;
1295 }
1296
Johannes Weiner8186eb62014-10-29 14:50:51 -07001297 /* page still mapped by someone else? */
1298 if (!atomic_add_negative(-1, &page->_mapcount))
1299 goto out;
1300
Johannes Weiner8186eb62014-10-29 14:50:51 -07001301 /*
1302 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1303 * these counters are not modified in interrupt context, and
1304 * pte lock(a spinlock) is held, which implies preemption disabled.
1305 */
1306 __dec_zone_page_state(page, NR_FILE_MAPPED);
Johannes Weiner62cccb82016-03-15 14:57:22 -07001307 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001308
1309 if (unlikely(PageMlocked(page)))
1310 clear_page_mlock(page);
1311out:
Johannes Weiner62cccb82016-03-15 14:57:22 -07001312 unlock_page_memcg(page);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001313}
1314
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001315static void page_remove_anon_compound_rmap(struct page *page)
1316{
1317 int i, nr;
1318
1319 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1320 return;
1321
1322 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1323 if (unlikely(PageHuge(page)))
1324 return;
1325
1326 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1327 return;
1328
1329 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1330
1331 if (TestClearPageDoubleMap(page)) {
1332 /*
1333 * Subpages can be mapped with PTEs too. Check how many of
1334 * themi are still mapped.
1335 */
1336 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1337 if (atomic_add_negative(-1, &page[i]._mapcount))
1338 nr++;
1339 }
1340 } else {
1341 nr = HPAGE_PMD_NR;
1342 }
1343
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001344 if (unlikely(PageMlocked(page)))
1345 clear_page_mlock(page);
1346
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001347 if (nr) {
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001348 __mod_zone_page_state(page_zone(page), NR_ANON_PAGES, -nr);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001349 deferred_split_huge_page(page);
1350 }
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001351}
1352
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353/**
1354 * page_remove_rmap - take down pte mapping from a page
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001355 * @page: page to remove mapping from
1356 * @compound: uncharge the page as compound or small page
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -07001358 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 */
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001360void page_remove_rmap(struct page *page, bool compound)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361{
Johannes Weiner8186eb62014-10-29 14:50:51 -07001362 if (!PageAnon(page)) {
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001363 VM_BUG_ON_PAGE(compound && !PageHuge(page), page);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001364 page_remove_file_rmap(page);
1365 return;
1366 }
KAMEZAWA Hiroyuki89c06bd2012-03-21 16:34:25 -07001367
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001368 if (compound)
1369 return page_remove_anon_compound_rmap(page);
1370
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001371 /* page still mapped by someone else? */
1372 if (!atomic_add_negative(-1, &page->_mapcount))
Johannes Weiner8186eb62014-10-29 14:50:51 -07001373 return;
1374
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001375 /*
Jianyu Zhanbea04b02014-06-04 16:09:51 -07001376 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1377 * these counters are not modified in interrupt context, and
Jianyu Zhanbea04b02014-06-04 16:09:51 -07001378 * pte lock(a spinlock) is held, which implies preemption disabled.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001379 */
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001380 __dec_zone_page_state(page, NR_ANON_PAGES);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001381
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001382 if (unlikely(PageMlocked(page)))
1383 clear_page_mlock(page);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001384
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001385 if (PageTransCompound(page))
1386 deferred_split_huge_page(compound_head(page));
1387
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001388 /*
1389 * It would be tidy to reset the PageAnon mapping here,
1390 * but that might overwrite a racing page_add_anon_rmap
1391 * which increments mapcount after us but sets mapping
1392 * before us: so leave the reset to free_hot_cold_page,
1393 * and remember that it's only reliable while mapped.
1394 * Leaving it set also helps swapoff to reinstate ptes
1395 * faster for those pages still in swapcache.
1396 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397}
1398
Minchan Kim854e9ed2016-01-15 16:54:53 -08001399struct rmap_private {
1400 enum ttu_flags flags;
1401 int lazyfreed;
1402};
1403
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404/*
Joonsoo Kim52629502014-01-21 15:49:50 -08001405 * @arg: enum ttu_flags will be passed to this argument
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 */
Kirill A. Shutemovac769502014-06-04 16:08:17 -07001407static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
Joonsoo Kim52629502014-01-21 15:49:50 -08001408 unsigned long address, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409{
1410 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411 pte_t *pte;
1412 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001413 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414 int ret = SWAP_AGAIN;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001415 struct rmap_private *rp = arg;
1416 enum ttu_flags flags = rp->flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417
Hugh Dickinsb87537d9e2015-11-05 18:49:33 -08001418 /* munlock has nothing to gain from examining un-locked vmas */
1419 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1420 goto out;
1421
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001422 if (flags & TTU_SPLIT_HUGE_PMD) {
1423 split_huge_pmd_address(vma, address,
1424 flags & TTU_MIGRATION, page);
1425 /* check if we have anything to do after split */
1426 if (page_mapcount(page) == 0)
1427 goto out;
1428 }
1429
Naoya Horiguchi55bda432016-07-14 12:07:35 -07001430 pte = page_check_address(page, mm, address, &ptl,
1431 PageTransCompound(page));
Hugh Dickinsc0718802005-10-29 18:16:31 -07001432 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -07001433 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434
1435 /*
1436 * If the page is mlock()d, we cannot swap it out.
1437 * If it's recently referenced (perhaps page_referenced
1438 * skipped over this mm) then we should reactivate it.
1439 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001440 if (!(flags & TTU_IGNORE_MLOCK)) {
Hugh Dickinsb87537d9e2015-11-05 18:49:33 -08001441 if (vma->vm_flags & VM_LOCKED) {
1442 /* Holding pte lock, we do *not* need mmap_sem here */
1443 mlock_vma_page(page);
1444 ret = SWAP_MLOCK;
1445 goto out_unmap;
1446 }
Konstantin Khlebnikovdaa5ba72014-06-04 16:10:52 -07001447 if (flags & TTU_MUNLOCK)
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001448 goto out_unmap;
Andi Kleen14fa31b2009-09-16 11:50:10 +02001449 }
1450 if (!(flags & TTU_IGNORE_ACCESS)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001451 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1452 ret = SWAP_FAIL;
1453 goto out_unmap;
1454 }
1455 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 /* Nuke the page table entry. */
1458 flush_cache_page(vma, address, page_to_pfn(page));
Mel Gorman72b252a2015-09-04 15:47:32 -07001459 if (should_defer_flush(mm, flags)) {
1460 /*
1461 * We clear the PTE but do not flush so potentially a remote
1462 * CPU could still be writing to the page. If the entry was
1463 * previously clean then the architecture must guarantee that
1464 * a clear->dirty transition on a cached TLB entry is written
1465 * through and traps if the PTE is unmapped.
1466 */
1467 pteval = ptep_get_and_clear(mm, address, pte);
1468
Mel Gormand950c942015-09-04 15:47:35 -07001469 set_tlb_ubc_flush_pending(mm, page, pte_dirty(pteval));
Mel Gorman72b252a2015-09-04 15:47:32 -07001470 } else {
1471 pteval = ptep_clear_flush(vma, address, pte);
1472 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473
1474 /* Move the dirty bit to the physical page now the pte is gone. */
1475 if (pte_dirty(pteval))
1476 set_page_dirty(page);
1477
Hugh Dickins365e9c872005-10-29 18:16:18 -07001478 /* Update high watermark before we lower rss */
1479 update_hiwater_rss(mm);
1480
Andi Kleen888b9f72009-09-16 11:50:11 +02001481 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08001482 if (PageHuge(page)) {
1483 hugetlb_count_sub(1 << compound_order(page), mm);
1484 } else {
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001485 dec_mm_counter(mm, mm_counter(page));
Naoya Horiguchi5f24ae52012-12-12 13:52:30 -08001486 }
Andi Kleen888b9f72009-09-16 11:50:11 +02001487 set_pte_at(mm, address, pte,
Naoya Horiguchi5f24ae52012-12-12 13:52:30 -08001488 swp_entry_to_pte(make_hwpoison_entry(page)));
Konstantin Weitz45961722013-04-17 13:59:32 +02001489 } else if (pte_unused(pteval)) {
1490 /*
1491 * The guest indicated that the page content is of no
1492 * interest anymore. Simply discard the pte, vmscan
1493 * will take care of the rest.
1494 */
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001495 dec_mm_counter(mm, mm_counter(page));
Hugh Dickins470f1192015-11-05 18:49:59 -08001496 } else if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION)) {
1497 swp_entry_t entry;
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07001498 pte_t swp_pte;
Hugh Dickins470f1192015-11-05 18:49:59 -08001499 /*
1500 * Store the pfn of the page in a special migration
1501 * pte. do_swap_page() will wait until the migration
1502 * pte is removed and then restart fault handling.
1503 */
1504 entry = make_migration_entry(page, pte_write(pteval));
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07001505 swp_pte = swp_entry_to_pte(entry);
1506 if (pte_soft_dirty(pteval))
1507 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1508 set_pte_at(mm, address, pte, swp_pte);
Hugh Dickins470f1192015-11-05 18:49:59 -08001509 } else if (PageAnon(page)) {
1510 swp_entry_t entry = { .val = page_private(page) };
1511 pte_t swp_pte;
1512 /*
1513 * Store the swap location in the pte.
1514 * See handle_pte_fault() ...
1515 */
1516 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
Minchan Kim854e9ed2016-01-15 16:54:53 -08001517
1518 if (!PageDirty(page) && (flags & TTU_LZFREE)) {
1519 /* It's a freeable page by MADV_FREE */
1520 dec_mm_counter(mm, MM_ANONPAGES);
1521 rp->lazyfreed++;
1522 goto discard;
1523 }
1524
Hugh Dickins470f1192015-11-05 18:49:59 -08001525 if (swap_duplicate(entry) < 0) {
1526 set_pte_at(mm, address, pte, pteval);
1527 ret = SWAP_FAIL;
1528 goto out_unmap;
1529 }
1530 if (list_empty(&mm->mmlist)) {
1531 spin_lock(&mmlist_lock);
1532 if (list_empty(&mm->mmlist))
1533 list_add(&mm->mmlist, &init_mm.mmlist);
1534 spin_unlock(&mmlist_lock);
1535 }
1536 dec_mm_counter(mm, MM_ANONPAGES);
1537 inc_mm_counter(mm, MM_SWAPENTS);
1538 swp_pte = swp_entry_to_pte(entry);
1539 if (pte_soft_dirty(pteval))
1540 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1541 set_pte_at(mm, address, pte, swp_pte);
Christoph Lameter04e62a22006-06-23 02:03:38 -07001542 } else
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001543 dec_mm_counter(mm, mm_counter_file(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544
Minchan Kim854e9ed2016-01-15 16:54:53 -08001545discard:
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001546 page_remove_rmap(page, PageHuge(page));
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001547 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548
1549out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -07001550 pte_unmap_unlock(pte, ptl);
Hugh Dickinsb87537d9e2015-11-05 18:49:33 -08001551 if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK))
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001552 mmu_notifier_invalidate_page(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553out:
1554 return ret;
1555}
1556
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001557bool is_vma_temporary_stack(struct vm_area_struct *vma)
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001558{
1559 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1560
1561 if (!maybe_stack)
1562 return false;
1563
1564 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1565 VM_STACK_INCOMPLETE_SETUP)
1566 return true;
1567
1568 return false;
1569}
1570
Joonsoo Kim52629502014-01-21 15:49:50 -08001571static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1572{
1573 return is_vma_temporary_stack(vma);
1574}
1575
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001576static int page_mapcount_is_zero(struct page *page)
Joonsoo Kim52629502014-01-21 15:49:50 -08001577{
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001578 return !page_mapcount(page);
1579}
Joonsoo Kim52629502014-01-21 15:49:50 -08001580
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581/**
1582 * try_to_unmap - try to remove all page table mappings to a page
1583 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001584 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585 *
1586 * Tries to remove all the page table entries which are mapping this
1587 * page, used in the pageout path. Caller must hold the page lock.
1588 * Return values are:
1589 *
1590 * SWAP_SUCCESS - we succeeded in removing all mappings
1591 * SWAP_AGAIN - we missed a mapping, try again later
1592 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001593 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001595int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596{
1597 int ret;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001598 struct rmap_private rp = {
1599 .flags = flags,
1600 .lazyfreed = 0,
1601 };
1602
Joonsoo Kim52629502014-01-21 15:49:50 -08001603 struct rmap_walk_control rwc = {
1604 .rmap_one = try_to_unmap_one,
Minchan Kim854e9ed2016-01-15 16:54:53 -08001605 .arg = &rp,
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001606 .done = page_mapcount_is_zero,
Joonsoo Kim52629502014-01-21 15:49:50 -08001607 .anon_lock = page_lock_anon_vma_read,
1608 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609
Joonsoo Kim52629502014-01-21 15:49:50 -08001610 /*
1611 * During exec, a temporary VMA is setup and later moved.
1612 * The VMA is moved under the anon_vma lock but not the
1613 * page tables leading to a race where migration cannot
1614 * find the migration ptes. Rather than increasing the
1615 * locking requirements of exec(), migration skips
1616 * temporary VMAs until after exec() completes.
1617 */
Konstantin Khlebnikovdaa5ba72014-06-04 16:10:52 -07001618 if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page))
Joonsoo Kim52629502014-01-21 15:49:50 -08001619 rwc.invalid_vma = invalid_migration_vma;
1620
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001621 if (flags & TTU_RMAP_LOCKED)
1622 ret = rmap_walk_locked(page, &rwc);
1623 else
1624 ret = rmap_walk(page, &rwc);
Joonsoo Kim52629502014-01-21 15:49:50 -08001625
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001626 if (ret != SWAP_MLOCK && !page_mapcount(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627 ret = SWAP_SUCCESS;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001628 if (rp.lazyfreed && !PageDirty(page))
1629 ret = SWAP_LZFREE;
1630 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 return ret;
1632}
Nikita Danilov81b40822005-05-01 08:58:36 -07001633
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001634static int page_not_mapped(struct page *page)
1635{
1636 return !page_mapped(page);
1637};
1638
Nick Pigginb291f002008-10-18 20:26:44 -07001639/**
1640 * try_to_munlock - try to munlock a page
1641 * @page: the page to be munlocked
1642 *
1643 * Called from munlock code. Checks all of the VMAs mapping the page
1644 * to make sure nobody else has this page mlocked. The page will be
1645 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1646 *
1647 * Return values are:
1648 *
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001649 * SWAP_AGAIN - no vma is holding page mlocked, or,
Nick Pigginb291f002008-10-18 20:26:44 -07001650 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
Hugh Dickins5ad64682009-12-14 17:59:24 -08001651 * SWAP_FAIL - page cannot be located at present
Nick Pigginb291f002008-10-18 20:26:44 -07001652 * SWAP_MLOCK - page is now mlocked.
1653 */
1654int try_to_munlock(struct page *page)
1655{
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001656 int ret;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001657 struct rmap_private rp = {
1658 .flags = TTU_MUNLOCK,
1659 .lazyfreed = 0,
1660 };
1661
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001662 struct rmap_walk_control rwc = {
1663 .rmap_one = try_to_unmap_one,
Minchan Kim854e9ed2016-01-15 16:54:53 -08001664 .arg = &rp,
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001665 .done = page_not_mapped,
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001666 .anon_lock = page_lock_anon_vma_read,
1667
1668 };
1669
Sasha Levin309381fea2014-01-23 15:52:54 -08001670 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
Nick Pigginb291f002008-10-18 20:26:44 -07001671
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001672 ret = rmap_walk(page, &rwc);
1673 return ret;
Nick Pigginb291f002008-10-18 20:26:44 -07001674}
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001675
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001676void __put_anon_vma(struct anon_vma *anon_vma)
Rik van Riel76545062010-08-09 17:18:41 -07001677{
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001678 struct anon_vma *root = anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -07001679
Andrey Ryabinin624483f2014-06-06 19:09:30 +04001680 anon_vma_free(anon_vma);
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001681 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1682 anon_vma_free(root);
Rik van Riel76545062010-08-09 17:18:41 -07001683}
Rik van Riel76545062010-08-09 17:18:41 -07001684
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001685static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1686 struct rmap_walk_control *rwc)
Joonsoo Kimfaecd8d2014-01-21 15:49:46 -08001687{
1688 struct anon_vma *anon_vma;
1689
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001690 if (rwc->anon_lock)
1691 return rwc->anon_lock(page);
1692
Joonsoo Kimfaecd8d2014-01-21 15:49:46 -08001693 /*
1694 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1695 * because that depends on page_mapped(); but not all its usages
1696 * are holding mmap_sem. Users without mmap_sem are required to
1697 * take a reference count to prevent the anon_vma disappearing
1698 */
1699 anon_vma = page_anon_vma(page);
1700 if (!anon_vma)
1701 return NULL;
1702
1703 anon_vma_lock_read(anon_vma);
1704 return anon_vma;
1705}
1706
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001707/*
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001708 * rmap_walk_anon - do something to anonymous page using the object-based
1709 * rmap method
1710 * @page: the page to be handled
1711 * @rwc: control variable according to each walk type
1712 *
1713 * Find all the mappings of a page using the mapping pointer and the vma chains
1714 * contained in the anon_vma struct it points to.
1715 *
1716 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1717 * where the page was found will be held for write. So, we won't recheck
1718 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1719 * LOCKED.
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001720 */
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001721static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1722 bool locked)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001723{
1724 struct anon_vma *anon_vma;
Davidlohr Buesob258d862014-12-12 16:55:04 -08001725 pgoff_t pgoff;
Rik van Riel5beb4932010-03-05 13:42:07 -08001726 struct anon_vma_chain *avc;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001727 int ret = SWAP_AGAIN;
1728
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001729 if (locked) {
1730 anon_vma = page_anon_vma(page);
1731 /* anon_vma disappear under us? */
1732 VM_BUG_ON_PAGE(!anon_vma, page);
1733 } else {
1734 anon_vma = rmap_walk_anon_lock(page, rwc);
1735 }
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001736 if (!anon_vma)
1737 return ret;
Joonsoo Kimfaecd8d2014-01-21 15:49:46 -08001738
Davidlohr Buesob258d862014-12-12 16:55:04 -08001739 pgoff = page_to_pgoff(page);
Michel Lespinassebf181b92012-10-08 16:31:39 -07001740 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
Rik van Riel5beb4932010-03-05 13:42:07 -08001741 struct vm_area_struct *vma = avc->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001742 unsigned long address = vma_address(page, vma);
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001743
Andrea Arcangeliad126952015-11-05 18:49:07 -08001744 cond_resched();
1745
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001746 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1747 continue;
1748
Joonsoo Kim051ac832014-01-21 15:49:48 -08001749 ret = rwc->rmap_one(page, vma, address, rwc->arg);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001750 if (ret != SWAP_AGAIN)
1751 break;
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001752 if (rwc->done && rwc->done(page))
1753 break;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001754 }
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001755
1756 if (!locked)
1757 anon_vma_unlock_read(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001758 return ret;
1759}
1760
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001761/*
1762 * rmap_walk_file - do something to file page using the object-based rmap method
1763 * @page: the page to be handled
1764 * @rwc: control variable according to each walk type
1765 *
1766 * Find all the mappings of a page using the mapping pointer and the vma chains
1767 * contained in the address_space struct it points to.
1768 *
1769 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1770 * where the page was found will be held for write. So, we won't recheck
1771 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1772 * LOCKED.
1773 */
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001774static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1775 bool locked)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001776{
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001777 struct address_space *mapping = page_mapping(page);
Davidlohr Buesob258d862014-12-12 16:55:04 -08001778 pgoff_t pgoff;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001779 struct vm_area_struct *vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001780 int ret = SWAP_AGAIN;
1781
Joonsoo Kim9f326242014-01-21 15:49:53 -08001782 /*
1783 * The page lock not only makes sure that page->mapping cannot
1784 * suddenly be NULLified by truncation, it makes sure that the
1785 * structure at mapping cannot be freed and reused yet,
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08001786 * so we can safely take mapping->i_mmap_rwsem.
Joonsoo Kim9f326242014-01-21 15:49:53 -08001787 */
Sasha Levin81d1b092014-10-09 15:28:10 -07001788 VM_BUG_ON_PAGE(!PageLocked(page), page);
Joonsoo Kim9f326242014-01-21 15:49:53 -08001789
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001790 if (!mapping)
1791 return ret;
Davidlohr Bueso3dec0ba2014-12-12 16:54:27 -08001792
Davidlohr Buesob258d862014-12-12 16:55:04 -08001793 pgoff = page_to_pgoff(page);
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001794 if (!locked)
1795 i_mmap_lock_read(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07001796 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001797 unsigned long address = vma_address(page, vma);
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001798
Andrea Arcangeliad126952015-11-05 18:49:07 -08001799 cond_resched();
1800
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001801 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1802 continue;
1803
Joonsoo Kim051ac832014-01-21 15:49:48 -08001804 ret = rwc->rmap_one(page, vma, address, rwc->arg);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001805 if (ret != SWAP_AGAIN)
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001806 goto done;
1807 if (rwc->done && rwc->done(page))
1808 goto done;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001809 }
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001810
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001811done:
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001812 if (!locked)
1813 i_mmap_unlock_read(mapping);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001814 return ret;
1815}
1816
Joonsoo Kim051ac832014-01-21 15:49:48 -08001817int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001818{
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001819 if (unlikely(PageKsm(page)))
Joonsoo Kim051ac832014-01-21 15:49:48 -08001820 return rmap_walk_ksm(page, rwc);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001821 else if (PageAnon(page))
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001822 return rmap_walk_anon(page, rwc, false);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001823 else
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001824 return rmap_walk_file(page, rwc, false);
1825}
1826
1827/* Like rmap_walk, but caller holds relevant rmap lock */
1828int rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1829{
1830 /* no ksm support for now */
1831 VM_BUG_ON_PAGE(PageKsm(page), page);
1832 if (PageAnon(page))
1833 return rmap_walk_anon(page, rwc, true);
1834 else
1835 return rmap_walk_file(page, rwc, true);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001836}
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001837
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001838#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001839/*
1840 * The following three functions are for anonymous (private mapped) hugepages.
1841 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1842 * and no lru code, because we handle hugepages differently from common pages.
1843 */
1844static void __hugepage_set_anon_rmap(struct page *page,
1845 struct vm_area_struct *vma, unsigned long address, int exclusive)
1846{
1847 struct anon_vma *anon_vma = vma->anon_vma;
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001848
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001849 BUG_ON(!anon_vma);
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001850
1851 if (PageAnon(page))
1852 return;
1853 if (!exclusive)
1854 anon_vma = anon_vma->root;
1855
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001856 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1857 page->mapping = (struct address_space *) anon_vma;
1858 page->index = linear_page_index(vma, address);
1859}
1860
1861void hugepage_add_anon_rmap(struct page *page,
1862 struct vm_area_struct *vma, unsigned long address)
1863{
1864 struct anon_vma *anon_vma = vma->anon_vma;
1865 int first;
Naoya Horiguchia850ea32010-09-10 13:23:06 +09001866
1867 BUG_ON(!PageLocked(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001868 BUG_ON(!anon_vma);
Hugh Dickins5dbe0af2011-05-28 13:17:04 -07001869 /* address might be in next vma when migration races vma_adjust */
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001870 first = atomic_inc_and_test(compound_mapcount_ptr(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001871 if (first)
1872 __hugepage_set_anon_rmap(page, vma, address, 0);
1873}
1874
1875void hugepage_add_new_anon_rmap(struct page *page,
1876 struct vm_area_struct *vma, unsigned long address)
1877{
1878 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001879 atomic_set(compound_mapcount_ptr(page), 0);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001880 __hugepage_set_anon_rmap(page, vma, address, 1);
1881}
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001882#endif /* CONFIG_HUGETLB_PAGE */