blob: f6838015810f5610abe039daec170aa1da634422 [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>
Ingo Molnar6e84f312017-02-08 18:51:29 +010049#include <linux/sched/mm.h>
Ingo Molnar29930022017-02-08 18:51:36 +010050#include <linux/sched/task.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070051#include <linux/pagemap.h>
52#include <linux/swap.h>
53#include <linux/swapops.h>
54#include <linux/slab.h>
55#include <linux/init.h>
Hugh Dickins5ad64682009-12-14 17:59:24 -080056#include <linux/ksm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070057#include <linux/rmap.h>
58#include <linux/rcupdate.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040059#include <linux/export.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080060#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070061#include <linux/mmu_notifier.h>
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -080062#include <linux/migrate.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090063#include <linux/hugetlb.h>
Jan Karaef5d4372012-10-25 13:37:31 -070064#include <linux/backing-dev.h>
Vladimir Davydov33c3fc72015-09-09 15:35:45 -070065#include <linux/page_idle.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070066
67#include <asm/tlbflush.h>
68
Mel Gorman72b252a2015-09-04 15:47:32 -070069#include <trace/events/tlb.h>
70
Nick Pigginb291f002008-10-18 20:26:44 -070071#include "internal.h"
72
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070073static struct kmem_cache *anon_vma_cachep;
Rik van Riel5beb4932010-03-05 13:42:07 -080074static struct kmem_cache *anon_vma_chain_cachep;
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070075
76static inline struct anon_vma *anon_vma_alloc(void)
77{
Peter Zijlstra01d8b202011-03-22 16:32:49 -070078 struct anon_vma *anon_vma;
79
80 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
81 if (anon_vma) {
82 atomic_set(&anon_vma->refcount, 1);
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -080083 anon_vma->degree = 1; /* Reference for first vma */
84 anon_vma->parent = anon_vma;
Peter Zijlstra01d8b202011-03-22 16:32:49 -070085 /*
86 * Initialise the anon_vma root to point to itself. If called
87 * from fork, the root will be reset to the parents anon_vma.
88 */
89 anon_vma->root = anon_vma;
90 }
91
92 return anon_vma;
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070093}
94
Peter Zijlstra01d8b202011-03-22 16:32:49 -070095static inline void anon_vma_free(struct anon_vma *anon_vma)
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070096{
Peter Zijlstra01d8b202011-03-22 16:32:49 -070097 VM_BUG_ON(atomic_read(&anon_vma->refcount));
Peter Zijlstra88c22082011-05-24 17:12:13 -070098
99 /*
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000100 * Synchronize against page_lock_anon_vma_read() such that
Peter Zijlstra88c22082011-05-24 17:12:13 -0700101 * we can safely hold the lock without the anon_vma getting
102 * freed.
103 *
104 * Relies on the full mb implied by the atomic_dec_and_test() from
105 * put_anon_vma() against the acquire barrier implied by
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000106 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
Peter Zijlstra88c22082011-05-24 17:12:13 -0700107 *
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000108 * page_lock_anon_vma_read() VS put_anon_vma()
109 * down_read_trylock() atomic_dec_and_test()
Peter Zijlstra88c22082011-05-24 17:12:13 -0700110 * LOCK MB
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000111 * atomic_read() rwsem_is_locked()
Peter Zijlstra88c22082011-05-24 17:12:13 -0700112 *
113 * LOCK should suffice since the actual taking of the lock must
114 * happen _before_ what follows.
115 */
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700116 might_sleep();
Ingo Molnar5a505082012-12-02 19:56:46 +0000117 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000118 anon_vma_lock_write(anon_vma);
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -0800119 anon_vma_unlock_write(anon_vma);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700120 }
121
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -0700122 kmem_cache_free(anon_vma_cachep, anon_vma);
123}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124
Linus Torvaldsdd347392011-06-17 19:05:36 -0700125static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
Rik van Riel5beb4932010-03-05 13:42:07 -0800126{
Linus Torvaldsdd347392011-06-17 19:05:36 -0700127 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
Rik van Riel5beb4932010-03-05 13:42:07 -0800128}
129
Namhyung Kime574b5f2010-10-26 14:22:02 -0700130static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
Rik van Riel5beb4932010-03-05 13:42:07 -0800131{
132 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
133}
134
Kautuk Consul6583a842012-03-21 16:34:01 -0700135static void anon_vma_chain_link(struct vm_area_struct *vma,
136 struct anon_vma_chain *avc,
137 struct anon_vma *anon_vma)
138{
139 avc->vma = vma;
140 avc->anon_vma = anon_vma;
141 list_add(&avc->same_vma, &vma->anon_vma_chain);
Michel Lespinassebf181b92012-10-08 16:31:39 -0700142 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
Kautuk Consul6583a842012-03-21 16:34:01 -0700143}
144
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700145/**
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800146 * __anon_vma_prepare - attach an anon_vma to a memory region
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700147 * @vma: the memory region in question
148 *
149 * This makes sure the memory mapping described by 'vma' has
150 * an 'anon_vma' attached to it, so that we can associate the
151 * anonymous pages mapped into it with that anon_vma.
152 *
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800153 * The common case will be that we already have one, which
154 * is handled inline by anon_vma_prepare(). But if
Figo.zhang23a07902010-12-27 15:14:06 +0100155 * not we either need to find an adjacent mapping that we
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700156 * can re-use the anon_vma from (very common when the only
157 * reason for splitting a vma has been mprotect()), or we
158 * allocate a new one.
159 *
160 * Anon-vma allocations are very subtle, because we may have
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000161 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700162 * and that may actually touch the spinlock even in the newly
163 * allocated vma (it depends on RCU to make sure that the
164 * anon_vma isn't actually destroyed).
165 *
166 * As a result, we need to do proper anon_vma locking even
167 * for the new allocation. At the same time, we do not want
168 * to do any locking for the common case of already having
169 * an anon_vma.
170 *
171 * This must be called with the mmap_sem held for reading.
172 */
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800173int __anon_vma_prepare(struct vm_area_struct *vma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174{
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800175 struct mm_struct *mm = vma->vm_mm;
176 struct anon_vma *anon_vma, *allocated;
Rik van Riel5beb4932010-03-05 13:42:07 -0800177 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178
179 might_sleep();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800181 avc = anon_vma_chain_alloc(GFP_KERNEL);
182 if (!avc)
183 goto out_enomem;
Rik van Riel5beb4932010-03-05 13:42:07 -0800184
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800185 anon_vma = find_mergeable_anon_vma(vma);
186 allocated = NULL;
187 if (!anon_vma) {
188 anon_vma = anon_vma_alloc();
189 if (unlikely(!anon_vma))
190 goto out_enomem_free_avc;
191 allocated = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 }
Vlastimil Babkad5a187d2016-12-12 16:44:38 -0800193
194 anon_vma_lock_write(anon_vma);
195 /* page_table_lock to protect against threads */
196 spin_lock(&mm->page_table_lock);
197 if (likely(!vma->anon_vma)) {
198 vma->anon_vma = anon_vma;
199 anon_vma_chain_link(vma, avc, anon_vma);
200 /* vma reference or self-parent link for new root */
201 anon_vma->degree++;
202 allocated = NULL;
203 avc = NULL;
204 }
205 spin_unlock(&mm->page_table_lock);
206 anon_vma_unlock_write(anon_vma);
207
208 if (unlikely(allocated))
209 put_anon_vma(allocated);
210 if (unlikely(avc))
211 anon_vma_chain_free(avc);
212
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213 return 0;
Rik van Riel5beb4932010-03-05 13:42:07 -0800214
215 out_enomem_free_avc:
216 anon_vma_chain_free(avc);
217 out_enomem:
218 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219}
220
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700221/*
222 * This is a useful helper function for locking the anon_vma root as
223 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
224 * have the same vma.
225 *
226 * Such anon_vma's should have the same root, so you'd expect to see
227 * just a single mutex_lock for the whole traversal.
228 */
229static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
230{
231 struct anon_vma *new_root = anon_vma->root;
232 if (new_root != root) {
233 if (WARN_ON_ONCE(root))
Ingo Molnar5a505082012-12-02 19:56:46 +0000234 up_write(&root->rwsem);
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700235 root = new_root;
Ingo Molnar5a505082012-12-02 19:56:46 +0000236 down_write(&root->rwsem);
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700237 }
238 return root;
239}
240
241static inline void unlock_anon_vma_root(struct anon_vma *root)
242{
243 if (root)
Ingo Molnar5a505082012-12-02 19:56:46 +0000244 up_write(&root->rwsem);
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700245}
246
Rik van Riel5beb4932010-03-05 13:42:07 -0800247/*
248 * Attach the anon_vmas from src to dst.
249 * Returns 0 on success, -ENOMEM on failure.
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800250 *
251 * If dst->anon_vma is NULL this function tries to find and reuse existing
252 * anon_vma which has no vmas and only one child anon_vma. This prevents
253 * degradation of anon_vma hierarchy to endless linear chain in case of
254 * constantly forking task. On the other hand, an anon_vma with more than one
255 * child isn't reused even if there was no alive vma, thus rmap walker has a
256 * good chance of avoiding scanning the whole hierarchy when it searches where
257 * page is mapped.
Rik van Riel5beb4932010-03-05 13:42:07 -0800258 */
259int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260{
Rik van Riel5beb4932010-03-05 13:42:07 -0800261 struct anon_vma_chain *avc, *pavc;
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700262 struct anon_vma *root = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263
Linus Torvalds646d87b2010-04-11 17:15:03 -0700264 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700265 struct anon_vma *anon_vma;
266
Linus Torvaldsdd347392011-06-17 19:05:36 -0700267 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
268 if (unlikely(!avc)) {
269 unlock_anon_vma_root(root);
270 root = NULL;
271 avc = anon_vma_chain_alloc(GFP_KERNEL);
272 if (!avc)
273 goto enomem_failure;
274 }
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700275 anon_vma = pavc->anon_vma;
276 root = lock_anon_vma_root(root, anon_vma);
277 anon_vma_chain_link(dst, avc, anon_vma);
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800278
279 /*
280 * Reuse existing anon_vma if its degree lower than two,
281 * that means it has no vma and only one anon_vma child.
282 *
283 * Do not chose parent anon_vma, otherwise first child
284 * will always reuse it. Root anon_vma is never reused:
285 * it has self-parent reference and at least one child.
286 */
287 if (!dst->anon_vma && anon_vma != src->anon_vma &&
288 anon_vma->degree < 2)
289 dst->anon_vma = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 }
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800291 if (dst->anon_vma)
292 dst->anon_vma->degree++;
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700293 unlock_anon_vma_root(root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800294 return 0;
295
296 enomem_failure:
Leon Yu3fe89b32015-03-25 15:55:11 -0700297 /*
298 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
299 * decremented in unlink_anon_vmas().
300 * We can safely do this because callers of anon_vma_clone() don't care
301 * about dst->anon_vma if anon_vma_clone() failed.
302 */
303 dst->anon_vma = NULL;
Rik van Riel5beb4932010-03-05 13:42:07 -0800304 unlink_anon_vmas(dst);
305 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306}
307
Rik van Riel5beb4932010-03-05 13:42:07 -0800308/*
309 * Attach vma to its own anon_vma, as well as to the anon_vmas that
310 * the corresponding VMA in the parent process is attached to.
311 * Returns 0 on success, non-zero on failure.
312 */
313int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314{
Rik van Riel5beb4932010-03-05 13:42:07 -0800315 struct anon_vma_chain *avc;
316 struct anon_vma *anon_vma;
Daniel Forrestc4ea95d2014-12-02 15:59:42 -0800317 int error;
Rik van Riel5beb4932010-03-05 13:42:07 -0800318
319 /* Don't bother if the parent process has no anon_vma here. */
320 if (!pvma->anon_vma)
321 return 0;
322
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800323 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
324 vma->anon_vma = NULL;
325
Rik van Riel5beb4932010-03-05 13:42:07 -0800326 /*
327 * First, attach the new VMA to the parent VMA's anon_vmas,
328 * so rmap can find non-COWed pages in child processes.
329 */
Daniel Forrestc4ea95d2014-12-02 15:59:42 -0800330 error = anon_vma_clone(vma, pvma);
331 if (error)
332 return error;
Rik van Riel5beb4932010-03-05 13:42:07 -0800333
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800334 /* An existing anon_vma has been reused, all done then. */
335 if (vma->anon_vma)
336 return 0;
337
Rik van Riel5beb4932010-03-05 13:42:07 -0800338 /* Then add our own anon_vma. */
339 anon_vma = anon_vma_alloc();
340 if (!anon_vma)
341 goto out_error;
Linus Torvaldsdd347392011-06-17 19:05:36 -0700342 avc = anon_vma_chain_alloc(GFP_KERNEL);
Rik van Riel5beb4932010-03-05 13:42:07 -0800343 if (!avc)
344 goto out_error_free_anon_vma;
Rik van Riel5c341ee12010-08-09 17:18:39 -0700345
346 /*
347 * The root anon_vma's spinlock is the lock actually used when we
348 * lock any of the anon_vmas in this anon_vma tree.
349 */
350 anon_vma->root = pvma->anon_vma->root;
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800351 anon_vma->parent = pvma->anon_vma;
Rik van Riel76545062010-08-09 17:18:41 -0700352 /*
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700353 * With refcounts, an anon_vma can stay around longer than the
354 * process it belongs to. The root anon_vma needs to be pinned until
355 * this anon_vma is freed, because the lock lives in the root.
Rik van Riel76545062010-08-09 17:18:41 -0700356 */
357 get_anon_vma(anon_vma->root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800358 /* Mark this anon_vma as the one where our new (COWed) pages go. */
359 vma->anon_vma = anon_vma;
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000360 anon_vma_lock_write(anon_vma);
Rik van Riel5c341ee12010-08-09 17:18:39 -0700361 anon_vma_chain_link(vma, avc, anon_vma);
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800362 anon_vma->parent->degree++;
Konstantin Khlebnikov08b52702013-02-22 16:34:40 -0800363 anon_vma_unlock_write(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800364
365 return 0;
366
367 out_error_free_anon_vma:
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700368 put_anon_vma(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800369 out_error:
Rik van Riel4946d542010-04-05 12:13:33 -0400370 unlink_anon_vmas(vma);
Rik van Riel5beb4932010-03-05 13:42:07 -0800371 return -ENOMEM;
372}
373
Rik van Riel5beb4932010-03-05 13:42:07 -0800374void unlink_anon_vmas(struct vm_area_struct *vma)
375{
376 struct anon_vma_chain *avc, *next;
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200377 struct anon_vma *root = NULL;
Rik van Riel5beb4932010-03-05 13:42:07 -0800378
Rik van Riel5c341ee12010-08-09 17:18:39 -0700379 /*
380 * Unlink each anon_vma chained to the VMA. This list is ordered
381 * from newest to oldest, ensuring the root anon_vma gets freed last.
382 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800383 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200384 struct anon_vma *anon_vma = avc->anon_vma;
385
386 root = lock_anon_vma_root(root, anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -0700387 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200388
389 /*
390 * Leave empty anon_vmas on the list - we'll need
391 * to free them outside the lock.
392 */
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800393 if (RB_EMPTY_ROOT(&anon_vma->rb_root)) {
394 anon_vma->parent->degree--;
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200395 continue;
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800396 }
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200397
398 list_del(&avc->same_vma);
399 anon_vma_chain_free(avc);
400 }
Konstantin Khlebnikov7a3ef202015-01-08 14:32:15 -0800401 if (vma->anon_vma)
402 vma->anon_vma->degree--;
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200403 unlock_anon_vma_root(root);
404
405 /*
406 * Iterate the list once more, it now only contains empty and unlinked
407 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
Ingo Molnar5a505082012-12-02 19:56:46 +0000408 * needing to write-acquire the anon_vma->root->rwsem.
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200409 */
410 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
411 struct anon_vma *anon_vma = avc->anon_vma;
412
Konstantin Khlebnikove4c58002016-05-19 17:11:46 -0700413 VM_WARN_ON(anon_vma->degree);
Peter Zijlstraeee2acb2011-06-17 13:54:23 +0200414 put_anon_vma(anon_vma);
415
Rik van Riel5beb4932010-03-05 13:42:07 -0800416 list_del(&avc->same_vma);
417 anon_vma_chain_free(avc);
418 }
419}
420
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700421static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422{
Christoph Lametera35afb82007-05-16 22:10:57 -0700423 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424
Ingo Molnar5a505082012-12-02 19:56:46 +0000425 init_rwsem(&anon_vma->rwsem);
Peter Zijlstra83813262011-03-22 16:32:48 -0700426 atomic_set(&anon_vma->refcount, 0);
Michel Lespinassebf181b92012-10-08 16:31:39 -0700427 anon_vma->rb_root = RB_ROOT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428}
429
430void __init anon_vma_init(void)
431{
432 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Vladimir Davydov5d097052016-01-14 15:18:21 -0800433 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
434 anon_vma_ctor);
435 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
436 SLAB_PANIC|SLAB_ACCOUNT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437}
438
439/*
Peter Zijlstra6111e4c2011-05-24 17:12:08 -0700440 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
441 *
442 * Since there is no serialization what so ever against page_remove_rmap()
443 * the best this function can do is return a locked anon_vma that might
444 * have been relevant to this page.
445 *
446 * The page might have been remapped to a different anon_vma or the anon_vma
447 * returned may already be freed (and even reused).
448 *
Peter Zijlstrabc658c92011-05-29 10:33:44 +0200449 * In case it was remapped to a different anon_vma, the new anon_vma will be a
450 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
451 * ensure that any anon_vma obtained from the page will still be valid for as
452 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
453 *
Peter Zijlstra6111e4c2011-05-24 17:12:08 -0700454 * All users of this function must be very careful when walking the anon_vma
455 * chain and verify that the page in question is indeed mapped in it
456 * [ something equivalent to page_mapped_in_vma() ].
457 *
458 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
459 * that the anon_vma pointer from page->mapping is valid if there is a
460 * mapcount, we can dereference the anon_vma after observing those.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700462struct anon_vma *page_get_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463{
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700464 struct anon_vma *anon_vma = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 unsigned long anon_mapping;
466
467 rcu_read_lock();
Jason Low4db0c3c2015-04-15 16:14:08 -0700468 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800469 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470 goto out;
471 if (!page_mapped(page))
472 goto out;
473
474 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700475 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
476 anon_vma = NULL;
477 goto out;
478 }
Hugh Dickinsf18194272010-08-25 23:12:54 -0700479
480 /*
481 * If this page is still mapped, then its anon_vma cannot have been
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700482 * freed. But if it has been unmapped, we have no security against the
483 * anon_vma structure being freed and reused (for another anon_vma:
484 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
485 * above cannot corrupt).
Hugh Dickinsf18194272010-08-25 23:12:54 -0700486 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700487 if (!page_mapped(page)) {
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700488 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700489 put_anon_vma(anon_vma);
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700490 return NULL;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700491 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492out:
493 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700494
495 return anon_vma;
496}
497
Peter Zijlstra88c22082011-05-24 17:12:13 -0700498/*
499 * Similar to page_get_anon_vma() except it locks the anon_vma.
500 *
501 * Its a little more complex as it tries to keep the fast path to a single
502 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
503 * reference like with page_get_anon_vma() and then block on the mutex.
504 */
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000505struct anon_vma *page_lock_anon_vma_read(struct page *page)
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700506{
Peter Zijlstra88c22082011-05-24 17:12:13 -0700507 struct anon_vma *anon_vma = NULL;
Hugh Dickinseee0f252011-05-28 13:20:21 -0700508 struct anon_vma *root_anon_vma;
Peter Zijlstra88c22082011-05-24 17:12:13 -0700509 unsigned long anon_mapping;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700510
Peter Zijlstra88c22082011-05-24 17:12:13 -0700511 rcu_read_lock();
Jason Low4db0c3c2015-04-15 16:14:08 -0700512 anon_mapping = (unsigned long)READ_ONCE(page->mapping);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700513 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
514 goto out;
515 if (!page_mapped(page))
516 goto out;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700517
Peter Zijlstra88c22082011-05-24 17:12:13 -0700518 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Jason Low4db0c3c2015-04-15 16:14:08 -0700519 root_anon_vma = READ_ONCE(anon_vma->root);
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000520 if (down_read_trylock(&root_anon_vma->rwsem)) {
Peter Zijlstra88c22082011-05-24 17:12:13 -0700521 /*
Hugh Dickinseee0f252011-05-28 13:20:21 -0700522 * If the page is still mapped, then this anon_vma is still
523 * its anon_vma, and holding the mutex ensures that it will
Peter Zijlstrabc658c92011-05-29 10:33:44 +0200524 * not go away, see anon_vma_free().
Peter Zijlstra88c22082011-05-24 17:12:13 -0700525 */
Hugh Dickinseee0f252011-05-28 13:20:21 -0700526 if (!page_mapped(page)) {
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000527 up_read(&root_anon_vma->rwsem);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700528 anon_vma = NULL;
529 }
530 goto out;
531 }
532
533 /* trylock failed, we got to sleep */
534 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
535 anon_vma = NULL;
536 goto out;
537 }
538
539 if (!page_mapped(page)) {
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700540 rcu_read_unlock();
Peter Zijlstra88c22082011-05-24 17:12:13 -0700541 put_anon_vma(anon_vma);
Hugh Dickins7f39dda2014-06-04 16:05:33 -0700542 return NULL;
Peter Zijlstra88c22082011-05-24 17:12:13 -0700543 }
544
545 /* we pinned the anon_vma, its safe to sleep */
546 rcu_read_unlock();
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000547 anon_vma_lock_read(anon_vma);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700548
549 if (atomic_dec_and_test(&anon_vma->refcount)) {
550 /*
551 * Oops, we held the last refcount, release the lock
552 * and bail -- can't simply use put_anon_vma() because
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000553 * we'll deadlock on the anon_vma_lock_write() recursion.
Peter Zijlstra88c22082011-05-24 17:12:13 -0700554 */
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000555 anon_vma_unlock_read(anon_vma);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700556 __put_anon_vma(anon_vma);
557 anon_vma = NULL;
558 }
559
560 return anon_vma;
561
562out:
563 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700564 return anon_vma;
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800565}
566
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000567void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800568{
Ingo Molnar4fc3f1d2012-12-02 19:56:50 +0000569 anon_vma_unlock_read(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570}
571
Mel Gorman72b252a2015-09-04 15:47:32 -0700572#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
Mel Gorman72b252a2015-09-04 15:47:32 -0700573/*
574 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
575 * important if a PTE was dirty when it was unmapped that it's flushed
576 * before any IO is initiated on the page to prevent lost writes. Similarly,
577 * it must be flushed before freeing to prevent data leakage.
578 */
579void try_to_unmap_flush(void)
580{
581 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
582 int cpu;
583
584 if (!tlb_ubc->flush_required)
585 return;
586
587 cpu = get_cpu();
588
Nadav Amit858eaaa72016-04-01 14:31:26 -0700589 if (cpumask_test_cpu(cpu, &tlb_ubc->cpumask)) {
590 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
591 local_flush_tlb();
592 trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
Mel Gorman72b252a2015-09-04 15:47:32 -0700593 }
Nadav Amit858eaaa72016-04-01 14:31:26 -0700594
595 if (cpumask_any_but(&tlb_ubc->cpumask, cpu) < nr_cpu_ids)
596 flush_tlb_others(&tlb_ubc->cpumask, NULL, 0, TLB_FLUSH_ALL);
Mel Gorman72b252a2015-09-04 15:47:32 -0700597 cpumask_clear(&tlb_ubc->cpumask);
598 tlb_ubc->flush_required = false;
Mel Gormand950c942015-09-04 15:47:35 -0700599 tlb_ubc->writable = false;
Mel Gorman72b252a2015-09-04 15:47:32 -0700600 put_cpu();
601}
602
Mel Gormand950c942015-09-04 15:47:35 -0700603/* Flush iff there are potentially writable TLB entries that can race with IO */
604void try_to_unmap_flush_dirty(void)
605{
606 struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
607
608 if (tlb_ubc->writable)
609 try_to_unmap_flush();
610}
611
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -0800612static void set_tlb_ubc_flush_pending(struct mm_struct *mm, 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
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -0800647static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
Mel Gorman72b252a2015-09-04 15:47:32 -0700648{
649}
650
651static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
652{
653 return false;
654}
655#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
656
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700658 * At what user virtual address is page expected in vma?
Naoya Horiguchiab941e02010-05-11 14:06:55 -0700659 * Caller should check the page is actually part of the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 */
661unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
662{
Michel Lespinasse86c2ad12012-10-08 16:31:42 -0700663 unsigned long address;
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700664 if (PageAnon(page)) {
Hugh Dickins4829b902010-10-02 17:46:06 -0700665 struct anon_vma *page__anon_vma = page_anon_vma(page);
666 /*
667 * Note: swapoff's unuse_vma() is more efficient with this
668 * check, and needs it to match anon_vma when KSM is active.
669 */
670 if (!vma->anon_vma || !page__anon_vma ||
671 vma->anon_vma->root != page__anon_vma->root)
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700672 return -EFAULT;
Kirill A. Shutemov27ba0642015-02-10 14:09:59 -0800673 } else if (page->mapping) {
674 if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 return -EFAULT;
676 } else
677 return -EFAULT;
Michel Lespinasse86c2ad12012-10-08 16:31:42 -0700678 address = __vma_address(page, vma);
679 if (unlikely(address < vma->vm_start || address >= vma->vm_end))
680 return -EFAULT;
681 return address;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682}
683
Bob Liu62190492012-12-11 16:00:37 -0800684pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
685{
686 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +0300687 p4d_t *p4d;
Bob Liu62190492012-12-11 16:00:37 -0800688 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
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +0300696 p4d = p4d_offset(pgd, address);
697 if (!p4d_present(*p4d))
698 goto out;
699
700 pud = pud_offset(p4d, address);
Bob Liu62190492012-12-11 16:00:37 -0800701 if (!pud_present(*pud))
702 goto out;
703
704 pmd = pmd_offset(pud, address);
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -0700705 /*
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -0700706 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -0700707 * without holding anon_vma lock for write. So when looking for a
708 * genuine pmde (in which to find pte), test present and !THP together.
709 */
Christian Borntraegere37c6982014-12-07 21:41:33 +0100710 pmde = *pmd;
711 barrier();
Hugh Dickinsf72e7dc2014-06-23 13:22:05 -0700712 if (!pmd_present(pmde) || pmd_trans_huge(pmde))
Bob Liu62190492012-12-11 16:00:37 -0800713 pmd = NULL;
714out:
715 return pmd;
716}
717
Joonsoo Kim9f326242014-01-21 15:49:53 -0800718struct page_referenced_arg {
719 int mapcount;
720 int referenced;
721 unsigned long vm_flags;
722 struct mem_cgroup *memcg;
723};
Nikita Danilov81b40822005-05-01 08:58:36 -0700724/*
Joonsoo Kim9f326242014-01-21 15:49:53 -0800725 * arg: page_referenced_arg will be passed
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726 */
Kirill A. Shutemovac769502014-06-04 16:08:17 -0700727static int page_referenced_one(struct page *page, struct vm_area_struct *vma,
Joonsoo Kim9f326242014-01-21 15:49:53 -0800728 unsigned long address, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729{
Joonsoo Kim9f326242014-01-21 15:49:53 -0800730 struct page_referenced_arg *pra = arg;
Kirill A. Shutemov8eaeded2017-02-24 14:57:48 -0800731 struct page_vma_mapped_walk pvmw = {
732 .page = page,
733 .vma = vma,
734 .address = address,
735 };
Vladimir Davydov8749cfe2016-01-15 16:54:45 -0800736 int referenced = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737
Kirill A. Shutemov8eaeded2017-02-24 14:57:48 -0800738 while (page_vma_mapped_walk(&pvmw)) {
739 address = pvmw.address;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800740
Kirill A. Shutemov8eaeded2017-02-24 14:57:48 -0800741 if (vma->vm_flags & VM_LOCKED) {
742 page_vma_mapped_walk_done(&pvmw);
743 pra->vm_flags |= VM_LOCKED;
744 return SWAP_FAIL; /* To break the loop */
Vladimir Davydov8749cfe2016-01-15 16:54:45 -0800745 }
Kirill A. Shutemov8eaeded2017-02-24 14:57:48 -0800746
747 if (pvmw.pte) {
748 if (ptep_clear_flush_young_notify(vma, address,
749 pvmw.pte)) {
750 /*
751 * Don't treat a reference through
752 * a sequentially read mapping as such.
753 * If the page has been used in another mapping,
754 * we will catch it; if this other mapping is
755 * already gone, the unmap path will have set
756 * PG_referenced or activated the page.
757 */
758 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
759 referenced++;
760 }
761 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
762 if (pmdp_clear_flush_young_notify(vma, address,
763 pvmw.pmd))
764 referenced++;
765 } else {
766 /* unexpected pmd-mapped page? */
767 WARN_ON_ONCE(1);
768 }
769
770 pra->mapcount--;
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800771 }
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800772
Vladimir Davydov33c3fc72015-09-09 15:35:45 -0700773 if (referenced)
774 clear_page_idle(page);
775 if (test_and_clear_page_young(page))
776 referenced++;
777
Joonsoo Kim9f326242014-01-21 15:49:53 -0800778 if (referenced) {
779 pra->referenced++;
780 pra->vm_flags |= vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800782
Joonsoo Kim9f326242014-01-21 15:49:53 -0800783 if (!pra->mapcount)
784 return SWAP_SUCCESS; /* To break the loop */
785
786 return SWAP_AGAIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787}
788
Joonsoo Kim9f326242014-01-21 15:49:53 -0800789static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790{
Joonsoo Kim9f326242014-01-21 15:49:53 -0800791 struct page_referenced_arg *pra = arg;
792 struct mem_cgroup *memcg = pra->memcg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793
Joonsoo Kim9f326242014-01-21 15:49:53 -0800794 if (!mm_match_cgroup(vma->vm_mm, memcg))
795 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796
Joonsoo Kim9f326242014-01-21 15:49:53 -0800797 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798}
799
800/**
801 * page_referenced - test if the page was referenced
802 * @page: the page to test
803 * @is_locked: caller holds lock on the page
Johannes Weiner72835c82012-01-12 17:18:32 -0800804 * @memcg: target memory cgroup
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700805 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806 *
807 * Quick test_and_clear_referenced for all mappings to a page,
808 * returns the number of ptes which referenced the page.
809 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700810int page_referenced(struct page *page,
811 int is_locked,
Johannes Weiner72835c82012-01-12 17:18:32 -0800812 struct mem_cgroup *memcg,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700813 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814{
Joonsoo Kim9f326242014-01-21 15:49:53 -0800815 int ret;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800816 int we_locked = 0;
Joonsoo Kim9f326242014-01-21 15:49:53 -0800817 struct page_referenced_arg pra = {
Kirill A. Shutemovb20ce5e2016-01-15 16:54:37 -0800818 .mapcount = total_mapcount(page),
Joonsoo Kim9f326242014-01-21 15:49:53 -0800819 .memcg = memcg,
820 };
821 struct rmap_walk_control rwc = {
822 .rmap_one = page_referenced_one,
823 .arg = (void *)&pra,
824 .anon_lock = page_lock_anon_vma_read,
825 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700827 *vm_flags = 0;
Joonsoo Kim9f326242014-01-21 15:49:53 -0800828 if (!page_mapped(page))
829 return 0;
830
831 if (!page_rmapping(page))
832 return 0;
833
834 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
835 we_locked = trylock_page(page);
836 if (!we_locked)
837 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838 }
Joonsoo Kim9f326242014-01-21 15:49:53 -0800839
840 /*
841 * If we are reclaiming on behalf of a cgroup, skip
842 * counting on behalf of references from different
843 * cgroups
844 */
845 if (memcg) {
846 rwc.invalid_vma = invalid_page_referenced_vma;
847 }
848
849 ret = rmap_walk(page, &rwc);
850 *vm_flags = pra.vm_flags;
851
852 if (we_locked)
853 unlock_page(page);
854
855 return pra.referenced;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856}
857
Hugh Dickins1cb17292009-12-14 17:59:01 -0800858static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
Joonsoo Kim9853a402014-01-21 15:49:55 -0800859 unsigned long address, void *arg)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700860{
Kirill A. Shutemovf27176c2017-02-24 14:57:57 -0800861 struct page_vma_mapped_walk pvmw = {
862 .page = page,
863 .vma = vma,
864 .address = address,
865 .flags = PVMW_SYNC,
866 };
Joonsoo Kim9853a402014-01-21 15:49:55 -0800867 int *cleaned = arg;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700868
Kirill A. Shutemovf27176c2017-02-24 14:57:57 -0800869 while (page_vma_mapped_walk(&pvmw)) {
870 int ret = 0;
871 address = pvmw.address;
872 if (pvmw.pte) {
873 pte_t entry;
874 pte_t *pte = pvmw.pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700875
Kirill A. Shutemovf27176c2017-02-24 14:57:57 -0800876 if (!pte_dirty(*pte) && !pte_write(*pte))
877 continue;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700878
Kirill A. Shutemovf27176c2017-02-24 14:57:57 -0800879 flush_cache_page(vma, address, pte_pfn(*pte));
880 entry = ptep_clear_flush(vma, address, pte);
881 entry = pte_wrprotect(entry);
882 entry = pte_mkclean(entry);
883 set_pte_at(vma->vm_mm, address, pte, entry);
884 ret = 1;
885 } else {
886#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
887 pmd_t *pmd = pvmw.pmd;
888 pmd_t entry;
889
890 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
891 continue;
892
893 flush_cache_page(vma, address, page_to_pfn(page));
894 entry = pmdp_huge_clear_flush(vma, address, pmd);
895 entry = pmd_wrprotect(entry);
896 entry = pmd_mkclean(entry);
897 set_pmd_at(vma->vm_mm, address, pmd, entry);
898 ret = 1;
899#else
900 /* unexpected pmd-mapped page? */
901 WARN_ON_ONCE(1);
902#endif
903 }
904
905 if (ret) {
906 mmu_notifier_invalidate_page(vma->vm_mm, address);
907 (*cleaned)++;
908 }
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100909 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700910
Joonsoo Kim9853a402014-01-21 15:49:55 -0800911 return SWAP_AGAIN;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700912}
913
Joonsoo Kim9853a402014-01-21 15:49:55 -0800914static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700915{
Joonsoo Kim9853a402014-01-21 15:49:55 -0800916 if (vma->vm_flags & VM_SHARED)
Fengguang Wu871beb82014-01-23 15:53:41 -0800917 return false;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700918
Fengguang Wu871beb82014-01-23 15:53:41 -0800919 return true;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700920}
921
922int page_mkclean(struct page *page)
923{
Joonsoo Kim9853a402014-01-21 15:49:55 -0800924 int cleaned = 0;
925 struct address_space *mapping;
926 struct rmap_walk_control rwc = {
927 .arg = (void *)&cleaned,
928 .rmap_one = page_mkclean_one,
929 .invalid_vma = invalid_mkclean_vma,
930 };
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700931
932 BUG_ON(!PageLocked(page));
933
Joonsoo Kim9853a402014-01-21 15:49:55 -0800934 if (!page_mapped(page))
935 return 0;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700936
Joonsoo Kim9853a402014-01-21 15:49:55 -0800937 mapping = page_mapping(page);
938 if (!mapping)
939 return 0;
940
941 rmap_walk(page, &rwc);
942
943 return cleaned;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700944}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700945EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700946
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947/**
Rik van Rielc44b6742010-03-05 13:42:09 -0800948 * page_move_anon_rmap - move a page to our anon_vma
949 * @page: the page to move to our anon_vma
950 * @vma: the vma the page belongs to
Rik van Rielc44b6742010-03-05 13:42:09 -0800951 *
952 * When a page belongs exclusively to one process after a COW event,
953 * that page can be moved into the anon_vma that belongs to just that
954 * process, so the rmap code will not search the parent or sibling
955 * processes.
956 */
Hugh Dickins5a499732016-07-14 12:07:38 -0700957void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
Rik van Rielc44b6742010-03-05 13:42:09 -0800958{
959 struct anon_vma *anon_vma = vma->anon_vma;
960
Hugh Dickins5a499732016-07-14 12:07:38 -0700961 page = compound_head(page);
962
Sasha Levin309381fea2014-01-23 15:52:54 -0800963 VM_BUG_ON_PAGE(!PageLocked(page), page);
Sasha Levin81d1b092014-10-09 15:28:10 -0700964 VM_BUG_ON_VMA(!anon_vma, vma);
Rik van Rielc44b6742010-03-05 13:42:09 -0800965
966 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
Vladimir Davydov414e2fb2015-06-24 16:56:56 -0700967 /*
968 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
969 * simultaneously, so a concurrent reader (eg page_referenced()'s
970 * PageAnon()) will not see one without the other.
971 */
972 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
Rik van Rielc44b6742010-03-05 13:42:09 -0800973}
974
975/**
Andi Kleen4e1c1972010-09-22 12:43:56 +0200976 * __page_set_anon_rmap - set up new anonymous rmap
977 * @page: Page to add to rmap
978 * @vma: VM area to add page to.
979 * @address: User virtual address of the mapping
Rik van Riele8a03fe2010-04-14 17:59:28 -0400980 * @exclusive: the page is exclusively owned by the current process
Nick Piggin9617d952006-01-06 00:11:12 -0800981 */
982static void __page_set_anon_rmap(struct page *page,
Rik van Riele8a03fe2010-04-14 17:59:28 -0400983 struct vm_area_struct *vma, unsigned long address, int exclusive)
Nick Piggin9617d952006-01-06 00:11:12 -0800984{
Rik van Riele8a03fe2010-04-14 17:59:28 -0400985 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -0800986
Rik van Riele8a03fe2010-04-14 17:59:28 -0400987 BUG_ON(!anon_vma);
Linus Torvaldsea900022010-04-12 12:44:29 -0700988
Andi Kleen4e1c1972010-09-22 12:43:56 +0200989 if (PageAnon(page))
990 return;
991
Linus Torvaldsea900022010-04-12 12:44:29 -0700992 /*
Rik van Riele8a03fe2010-04-14 17:59:28 -0400993 * If the page isn't exclusively mapped into this vma,
994 * we must use the _oldest_ possible anon_vma for the
995 * page mapping!
Linus Torvaldsea900022010-04-12 12:44:29 -0700996 */
Andi Kleen4e1c1972010-09-22 12:43:56 +0200997 if (!exclusive)
Andrea Arcangeli288468c2010-08-09 17:19:09 -0700998 anon_vma = anon_vma->root;
Linus Torvaldsea900022010-04-12 12:44:29 -0700999
Nick Piggin9617d952006-01-06 00:11:12 -08001000 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1001 page->mapping = (struct address_space *) anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -08001002 page->index = linear_page_index(vma, address);
Nick Piggin9617d952006-01-06 00:11:12 -08001003}
1004
1005/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -07001006 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -07001007 * @page: the page to add the mapping to
1008 * @vma: the vm area in which the mapping is added
1009 * @address: the user virtual address mapped
1010 */
1011static void __page_check_anon_rmap(struct page *page,
1012 struct vm_area_struct *vma, unsigned long address)
1013{
1014#ifdef CONFIG_DEBUG_VM
1015 /*
1016 * The page's anon-rmap details (mapping and index) are guaranteed to
1017 * be set up correctly at this point.
1018 *
1019 * We have exclusion against page_add_anon_rmap because the caller
1020 * always holds the page locked, except if called from page_dup_rmap,
1021 * in which case the page is already known to be setup.
1022 *
1023 * We have exclusion against page_add_new_anon_rmap because those pages
1024 * are initially only visible via the pagetables, and the pte is locked
1025 * over the call to page_add_new_anon_rmap.
1026 */
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001027 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001028 BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
Nick Pigginc97a9e12007-05-16 22:11:21 -07001029#endif
1030}
1031
1032/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001033 * page_add_anon_rmap - add pte mapping to an anonymous page
1034 * @page: the page to add the mapping to
1035 * @vma: the vm area in which the mapping is added
1036 * @address: the user virtual address mapped
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001037 * @compound: charge the page as compound or small page
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 *
Hugh Dickins5ad64682009-12-14 17:59:24 -08001039 * The caller needs to hold the pte lock, and the page must be locked in
Hugh Dickins80e148222009-12-14 17:59:29 -08001040 * the anon_vma case: to serialize mapping,index checking after setting,
1041 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1042 * (but PageKsm is never downgraded to PageAnon).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 */
1044void page_add_anon_rmap(struct page *page,
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001045 struct vm_area_struct *vma, unsigned long address, bool compound)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046{
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001047 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
Rik van Rielad8c2ee2010-08-09 17:19:48 -07001048}
1049
1050/*
1051 * Special version of the above for do_swap_page, which often runs
1052 * into pages that are exclusively owned by the current process.
1053 * Everybody else should continue to use page_add_anon_rmap above.
1054 */
1055void do_page_add_anon_rmap(struct page *page,
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001056 struct vm_area_struct *vma, unsigned long address, int flags)
Rik van Rielad8c2ee2010-08-09 17:19:48 -07001057{
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001058 bool compound = flags & RMAP_COMPOUND;
1059 bool first;
1060
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001061 if (compound) {
1062 atomic_t *mapcount;
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001063 VM_BUG_ON_PAGE(!PageLocked(page), page);
Kirill A. Shutemove9b61f12016-01-15 16:54:10 -08001064 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1065 mapcount = compound_mapcount_ptr(page);
1066 first = atomic_inc_and_test(mapcount);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001067 } else {
1068 first = atomic_inc_and_test(&page->_mapcount);
1069 }
1070
Andrea Arcangeli79134172011-01-13 15:46:58 -08001071 if (first) {
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001072 int nr = compound ? hpage_nr_pages(page) : 1;
Jianyu Zhanbea04b02014-06-04 16:09:51 -07001073 /*
1074 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1075 * these counters are not modified in interrupt context, and
1076 * pte lock(a spinlock) is held, which implies preemption
1077 * disabled.
1078 */
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07001079 if (compound)
Mel Gorman11fb9982016-07-28 15:46:20 -07001080 __inc_node_page_state(page, NR_ANON_THPS);
Mel Gorman4b9d0fa2016-07-28 15:46:17 -07001081 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001082 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001083 if (unlikely(PageKsm(page)))
1084 return;
1085
Sasha Levin309381fea2014-01-23 15:52:54 -08001086 VM_BUG_ON_PAGE(!PageLocked(page), page);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001087
Hugh Dickins5dbe0af2011-05-28 13:17:04 -07001088 /* address might be in next vma when migration races vma_adjust */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001089 if (first)
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001090 __page_set_anon_rmap(page, vma, address,
1091 flags & RMAP_EXCLUSIVE);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -07001092 else
Nick Pigginc97a9e12007-05-16 22:11:21 -07001093 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094}
1095
Randy Dunlap43d8eac2008-03-19 17:00:43 -07001096/**
Nick Piggin9617d952006-01-06 00:11:12 -08001097 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1098 * @page: the page to add the mapping to
1099 * @vma: the vm area in which the mapping is added
1100 * @address: the user virtual address mapped
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001101 * @compound: charge the page as compound or small page
Nick Piggin9617d952006-01-06 00:11:12 -08001102 *
1103 * Same as page_add_anon_rmap but must only be called on *new* pages.
1104 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -07001105 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -08001106 */
1107void page_add_new_anon_rmap(struct page *page,
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001108 struct vm_area_struct *vma, unsigned long address, bool compound)
Nick Piggin9617d952006-01-06 00:11:12 -08001109{
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001110 int nr = compound ? hpage_nr_pages(page) : 1;
1111
Sasha Levin81d1b092014-10-09 15:28:10 -07001112 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
Hugh Dickinsfa9949d2016-05-19 17:12:41 -07001113 __SetPageSwapBacked(page);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001114 if (compound) {
1115 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001116 /* increment count (starts at -1) */
1117 atomic_set(compound_mapcount_ptr(page), 0);
Mel Gorman11fb9982016-07-28 15:46:20 -07001118 __inc_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001119 } else {
1120 /* Anon THP always mapped first with PMD */
1121 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1122 /* increment count (starts at -1) */
1123 atomic_set(&page->_mapcount, 0);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001124 }
Mel Gorman4b9d0fa2016-07-28 15:46:17 -07001125 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
Rik van Riele8a03fe2010-04-14 17:59:28 -04001126 __page_set_anon_rmap(page, vma, address, 1);
Nick Piggin9617d952006-01-06 00:11:12 -08001127}
1128
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129/**
1130 * page_add_file_rmap - add pte mapping to a file page
1131 * @page: the page to add the mapping to
1132 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -07001133 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 */
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001135void page_add_file_rmap(struct page *page, bool compound)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001137 int i, nr = 1;
1138
1139 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
Johannes Weiner62cccb82016-03-15 14:57:22 -07001140 lock_page_memcg(page);
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001141 if (compound && PageTransHuge(page)) {
1142 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1143 if (atomic_inc_and_test(&page[i]._mapcount))
1144 nr++;
1145 }
1146 if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1147 goto out;
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07001148 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
Mel Gorman11fb9982016-07-28 15:46:20 -07001149 __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001150 } else {
Kirill A. Shutemovc8efc392016-08-10 16:27:52 -07001151 if (PageTransCompound(page) && page_mapping(page)) {
1152 VM_WARN_ON_ONCE(!PageLocked(page));
1153
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001154 SetPageDoubleMap(compound_head(page));
1155 if (PageMlocked(page))
1156 clear_page_mlock(compound_head(page));
1157 }
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001158 if (!atomic_inc_and_test(&page->_mapcount))
1159 goto out;
Balbir Singhd69b0422009-06-17 16:26:34 -07001160 }
Mel Gorman50658e22016-07-28 15:46:14 -07001161 __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, nr);
Johannes Weiner553af432017-03-31 15:11:50 -07001162 mem_cgroup_update_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED, nr);
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001163out:
Johannes Weiner62cccb82016-03-15 14:57:22 -07001164 unlock_page_memcg(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165}
1166
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001167static void page_remove_file_rmap(struct page *page, bool compound)
Johannes Weiner8186eb62014-10-29 14:50:51 -07001168{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001169 int i, nr = 1;
1170
Steve Capper57dea932016-08-10 16:27:55 -07001171 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
Johannes Weiner62cccb82016-03-15 14:57:22 -07001172 lock_page_memcg(page);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001173
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001174 /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1175 if (unlikely(PageHuge(page))) {
1176 /* hugetlb pages are always mapped with pmds */
1177 atomic_dec(compound_mapcount_ptr(page));
1178 goto out;
1179 }
1180
Johannes Weiner8186eb62014-10-29 14:50:51 -07001181 /* page still mapped by someone else? */
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001182 if (compound && PageTransHuge(page)) {
1183 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1184 if (atomic_add_negative(-1, &page[i]._mapcount))
1185 nr++;
1186 }
1187 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1188 goto out;
Kirill A. Shutemov65c45372016-07-26 15:26:10 -07001189 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
Mel Gorman11fb9982016-07-28 15:46:20 -07001190 __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001191 } else {
1192 if (!atomic_add_negative(-1, &page->_mapcount))
1193 goto out;
1194 }
Johannes Weiner8186eb62014-10-29 14:50:51 -07001195
Johannes Weiner8186eb62014-10-29 14:50:51 -07001196 /*
Mel Gorman50658e22016-07-28 15:46:14 -07001197 * We use the irq-unsafe __{inc|mod}_zone_page_state because
Johannes Weiner8186eb62014-10-29 14:50:51 -07001198 * these counters are not modified in interrupt context, and
1199 * pte lock(a spinlock) is held, which implies preemption disabled.
1200 */
Mel Gorman50658e22016-07-28 15:46:14 -07001201 __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, -nr);
Johannes Weiner553af432017-03-31 15:11:50 -07001202 mem_cgroup_update_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED, -nr);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001203
1204 if (unlikely(PageMlocked(page)))
1205 clear_page_mlock(page);
1206out:
Johannes Weiner62cccb82016-03-15 14:57:22 -07001207 unlock_page_memcg(page);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001208}
1209
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001210static void page_remove_anon_compound_rmap(struct page *page)
1211{
1212 int i, nr;
1213
1214 if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1215 return;
1216
1217 /* Hugepages are not counted in NR_ANON_PAGES for now. */
1218 if (unlikely(PageHuge(page)))
1219 return;
1220
1221 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1222 return;
1223
Mel Gorman11fb9982016-07-28 15:46:20 -07001224 __dec_node_page_state(page, NR_ANON_THPS);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001225
1226 if (TestClearPageDoubleMap(page)) {
1227 /*
1228 * Subpages can be mapped with PTEs too. Check how many of
1229 * themi are still mapped.
1230 */
1231 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1232 if (atomic_add_negative(-1, &page[i]._mapcount))
1233 nr++;
1234 }
1235 } else {
1236 nr = HPAGE_PMD_NR;
1237 }
1238
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08001239 if (unlikely(PageMlocked(page)))
1240 clear_page_mlock(page);
1241
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001242 if (nr) {
Mel Gorman4b9d0fa2016-07-28 15:46:17 -07001243 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001244 deferred_split_huge_page(page);
1245 }
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001246}
1247
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248/**
1249 * page_remove_rmap - take down pte mapping from a page
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001250 * @page: page to remove mapping from
1251 * @compound: uncharge the page as compound or small page
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -07001253 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 */
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001255void page_remove_rmap(struct page *page, bool compound)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256{
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001257 if (!PageAnon(page))
1258 return page_remove_file_rmap(page, compound);
KAMEZAWA Hiroyuki89c06bd2012-03-21 16:34:25 -07001259
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001260 if (compound)
1261 return page_remove_anon_compound_rmap(page);
1262
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001263 /* page still mapped by someone else? */
1264 if (!atomic_add_negative(-1, &page->_mapcount))
Johannes Weiner8186eb62014-10-29 14:50:51 -07001265 return;
1266
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001267 /*
Jianyu Zhanbea04b02014-06-04 16:09:51 -07001268 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1269 * these counters are not modified in interrupt context, and
Jianyu Zhanbea04b02014-06-04 16:09:51 -07001270 * pte lock(a spinlock) is held, which implies preemption disabled.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001271 */
Mel Gorman4b9d0fa2016-07-28 15:46:17 -07001272 __dec_node_page_state(page, NR_ANON_MAPPED);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001273
Hugh Dickinse6c509f2012-10-08 16:33:19 -07001274 if (unlikely(PageMlocked(page)))
1275 clear_page_mlock(page);
Johannes Weiner8186eb62014-10-29 14:50:51 -07001276
Kirill A. Shutemov9a982252016-01-15 16:54:17 -08001277 if (PageTransCompound(page))
1278 deferred_split_huge_page(compound_head(page));
1279
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001280 /*
1281 * It would be tidy to reset the PageAnon mapping here,
1282 * but that might overwrite a racing page_add_anon_rmap
1283 * which increments mapcount after us but sets mapping
1284 * before us: so leave the reset to free_hot_cold_page,
1285 * and remember that it's only reliable while mapped.
1286 * Leaving it set also helps swapoff to reinstate ptes
1287 * faster for those pages still in swapcache.
1288 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289}
1290
Minchan Kim854e9ed2016-01-15 16:54:53 -08001291struct rmap_private {
1292 enum ttu_flags flags;
1293 int lazyfreed;
1294};
1295
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296/*
Joonsoo Kim52629502014-01-21 15:49:50 -08001297 * @arg: enum ttu_flags will be passed to this argument
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298 */
Kirill A. Shutemovac769502014-06-04 16:08:17 -07001299static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
Joonsoo Kim52629502014-01-21 15:49:50 -08001300 unsigned long address, void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301{
1302 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001303 struct page_vma_mapped_walk pvmw = {
1304 .page = page,
1305 .vma = vma,
1306 .address = address,
1307 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 pte_t pteval;
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001309 struct page *subpage;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 int ret = SWAP_AGAIN;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001311 struct rmap_private *rp = arg;
1312 enum ttu_flags flags = rp->flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313
Hugh Dickinsb87537d9e2015-11-05 18:49:33 -08001314 /* munlock has nothing to gain from examining un-locked vmas */
1315 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001316 return SWAP_AGAIN;
Hugh Dickinsb87537d9e2015-11-05 18:49:33 -08001317
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001318 if (flags & TTU_SPLIT_HUGE_PMD) {
1319 split_huge_pmd_address(vma, address,
1320 flags & TTU_MIGRATION, page);
Kirill A. Shutemovfec89c12016-03-17 14:20:10 -07001321 }
1322
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001323 while (page_vma_mapped_walk(&pvmw)) {
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001324 /*
1325 * If the page is mlock()d, we cannot swap it out.
1326 * If it's recently referenced (perhaps page_referenced
1327 * skipped over this mm) then we should reactivate it.
1328 */
1329 if (!(flags & TTU_IGNORE_MLOCK)) {
1330 if (vma->vm_flags & VM_LOCKED) {
1331 /* PTE-mapped THP are never mlocked */
1332 if (!PageTransCompound(page)) {
1333 /*
1334 * Holding pte lock, we do *not* need
1335 * mmap_sem here
1336 */
1337 mlock_vma_page(page);
1338 }
1339 ret = SWAP_MLOCK;
1340 page_vma_mapped_walk_done(&pvmw);
1341 break;
Kirill A. Shutemov9a73f612016-07-26 15:25:53 -07001342 }
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001343 if (flags & TTU_MUNLOCK)
1344 continue;
Hugh Dickinsb87537d9e2015-11-05 18:49:33 -08001345 }
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001346
Kirill A. Shutemov83462422017-03-09 16:17:20 -08001347 /* Unexpected PMD-mapped THP? */
1348 VM_BUG_ON_PAGE(!pvmw.pte, page);
1349
1350 subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1351 address = pvmw.address;
1352
1353
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001354 if (!(flags & TTU_IGNORE_ACCESS)) {
1355 if (ptep_clear_flush_young_notify(vma, address,
1356 pvmw.pte)) {
1357 ret = SWAP_FAIL;
1358 page_vma_mapped_walk_done(&pvmw);
1359 break;
1360 }
Nick Pigginb291f002008-10-18 20:26:44 -07001361 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001363 /* Nuke the page table entry. */
1364 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1365 if (should_defer_flush(mm, flags)) {
1366 /*
1367 * We clear the PTE but do not flush so potentially
1368 * a remote CPU could still be writing to the page.
1369 * If the entry was previously clean then the
1370 * architecture must guarantee that a clear->dirty
1371 * transition on a cached TLB entry is written through
1372 * and traps if the PTE is unmapped.
1373 */
1374 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
Mel Gorman72b252a2015-09-04 15:47:32 -07001375
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001376 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08001377 } else {
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001378 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1379 }
1380
1381 /* Move the dirty bit to the page. Now the pte is gone. */
1382 if (pte_dirty(pteval))
1383 set_page_dirty(page);
1384
1385 /* Update high watermark before we lower rss */
1386 update_hiwater_rss(mm);
1387
1388 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1389 if (PageHuge(page)) {
1390 int nr = 1 << compound_order(page);
1391 hugetlb_count_sub(nr, mm);
1392 } else {
1393 dec_mm_counter(mm, mm_counter(page));
1394 }
1395
1396 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1397 set_pte_at(mm, address, pvmw.pte, pteval);
1398 } else if (pte_unused(pteval)) {
1399 /*
1400 * The guest indicated that the page content is of no
1401 * interest anymore. Simply discard the pte, vmscan
1402 * will take care of the rest.
1403 */
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001404 dec_mm_counter(mm, mm_counter(page));
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001405 } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1406 (flags & TTU_MIGRATION)) {
1407 swp_entry_t entry;
1408 pte_t swp_pte;
1409 /*
1410 * Store the pfn of the page in a special migration
1411 * pte. do_swap_page() will wait until the migration
1412 * pte is removed and then restart fault handling.
1413 */
1414 entry = make_migration_entry(subpage,
1415 pte_write(pteval));
1416 swp_pte = swp_entry_to_pte(entry);
1417 if (pte_soft_dirty(pteval))
1418 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1419 set_pte_at(mm, address, pvmw.pte, swp_pte);
1420 } else if (PageAnon(page)) {
1421 swp_entry_t entry = { .val = page_private(subpage) };
1422 pte_t swp_pte;
1423 /*
1424 * Store the swap location in the pte.
1425 * See handle_pte_fault() ...
1426 */
1427 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
Minchan Kim854e9ed2016-01-15 16:54:53 -08001428
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001429 if (!PageDirty(page) && (flags & TTU_LZFREE)) {
1430 /* It's a freeable page by MADV_FREE */
1431 dec_mm_counter(mm, MM_ANONPAGES);
1432 rp->lazyfreed++;
1433 goto discard;
1434 }
1435
1436 if (swap_duplicate(entry) < 0) {
1437 set_pte_at(mm, address, pvmw.pte, pteval);
1438 ret = SWAP_FAIL;
1439 page_vma_mapped_walk_done(&pvmw);
1440 break;
1441 }
1442 if (list_empty(&mm->mmlist)) {
1443 spin_lock(&mmlist_lock);
1444 if (list_empty(&mm->mmlist))
1445 list_add(&mm->mmlist, &init_mm.mmlist);
1446 spin_unlock(&mmlist_lock);
1447 }
Minchan Kim854e9ed2016-01-15 16:54:53 -08001448 dec_mm_counter(mm, MM_ANONPAGES);
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001449 inc_mm_counter(mm, MM_SWAPENTS);
1450 swp_pte = swp_entry_to_pte(entry);
1451 if (pte_soft_dirty(pteval))
1452 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1453 set_pte_at(mm, address, pvmw.pte, swp_pte);
1454 } else
1455 dec_mm_counter(mm, mm_counter_file(page));
Minchan Kim854e9ed2016-01-15 16:54:53 -08001456discard:
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001457 page_remove_rmap(subpage, PageHuge(page));
1458 put_page(page);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001459 mmu_notifier_invalidate_page(mm, address);
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001460 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461 return ret;
1462}
1463
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001464bool is_vma_temporary_stack(struct vm_area_struct *vma)
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001465{
1466 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1467
1468 if (!maybe_stack)
1469 return false;
1470
1471 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1472 VM_STACK_INCOMPLETE_SETUP)
1473 return true;
1474
1475 return false;
1476}
1477
Joonsoo Kim52629502014-01-21 15:49:50 -08001478static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1479{
1480 return is_vma_temporary_stack(vma);
1481}
1482
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001483static int page_mapcount_is_zero(struct page *page)
Joonsoo Kim52629502014-01-21 15:49:50 -08001484{
Kirill A. Shutemovc7ab0d22017-02-24 14:58:01 -08001485 return !total_mapcount(page);
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001486}
Joonsoo Kim52629502014-01-21 15:49:50 -08001487
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488/**
1489 * try_to_unmap - try to remove all page table mappings to a page
1490 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001491 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492 *
1493 * Tries to remove all the page table entries which are mapping this
1494 * page, used in the pageout path. Caller must hold the page lock.
1495 * Return values are:
1496 *
1497 * SWAP_SUCCESS - we succeeded in removing all mappings
1498 * SWAP_AGAIN - we missed a mapping, try again later
1499 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001500 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001502int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503{
1504 int ret;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001505 struct rmap_private rp = {
1506 .flags = flags,
1507 .lazyfreed = 0,
1508 };
1509
Joonsoo Kim52629502014-01-21 15:49:50 -08001510 struct rmap_walk_control rwc = {
1511 .rmap_one = try_to_unmap_one,
Minchan Kim854e9ed2016-01-15 16:54:53 -08001512 .arg = &rp,
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001513 .done = page_mapcount_is_zero,
Joonsoo Kim52629502014-01-21 15:49:50 -08001514 .anon_lock = page_lock_anon_vma_read,
1515 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516
Joonsoo Kim52629502014-01-21 15:49:50 -08001517 /*
1518 * During exec, a temporary VMA is setup and later moved.
1519 * The VMA is moved under the anon_vma lock but not the
1520 * page tables leading to a race where migration cannot
1521 * find the migration ptes. Rather than increasing the
1522 * locking requirements of exec(), migration skips
1523 * temporary VMAs until after exec() completes.
1524 */
Konstantin Khlebnikovdaa5ba72014-06-04 16:10:52 -07001525 if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page))
Joonsoo Kim52629502014-01-21 15:49:50 -08001526 rwc.invalid_vma = invalid_migration_vma;
1527
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001528 if (flags & TTU_RMAP_LOCKED)
1529 ret = rmap_walk_locked(page, &rwc);
1530 else
1531 ret = rmap_walk(page, &rwc);
Joonsoo Kim52629502014-01-21 15:49:50 -08001532
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001533 if (ret != SWAP_MLOCK && !page_mapcount(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534 ret = SWAP_SUCCESS;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001535 if (rp.lazyfreed && !PageDirty(page))
1536 ret = SWAP_LZFREE;
1537 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538 return ret;
1539}
Nikita Danilov81b40822005-05-01 08:58:36 -07001540
Kirill A. Shutemov2a52bcb2016-03-17 14:20:04 -07001541static int page_not_mapped(struct page *page)
1542{
1543 return !page_mapped(page);
1544};
1545
Nick Pigginb291f002008-10-18 20:26:44 -07001546/**
1547 * try_to_munlock - try to munlock a page
1548 * @page: the page to be munlocked
1549 *
1550 * Called from munlock code. Checks all of the VMAs mapping the page
1551 * to make sure nobody else has this page mlocked. The page will be
1552 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1553 *
1554 * Return values are:
1555 *
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001556 * SWAP_AGAIN - no vma is holding page mlocked, or,
Nick Pigginb291f002008-10-18 20:26:44 -07001557 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
Hugh Dickins5ad64682009-12-14 17:59:24 -08001558 * SWAP_FAIL - page cannot be located at present
Nick Pigginb291f002008-10-18 20:26:44 -07001559 * SWAP_MLOCK - page is now mlocked.
1560 */
1561int try_to_munlock(struct page *page)
1562{
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001563 int ret;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001564 struct rmap_private rp = {
1565 .flags = TTU_MUNLOCK,
1566 .lazyfreed = 0,
1567 };
1568
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001569 struct rmap_walk_control rwc = {
1570 .rmap_one = try_to_unmap_one,
Minchan Kim854e9ed2016-01-15 16:54:53 -08001571 .arg = &rp,
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001572 .done = page_not_mapped,
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001573 .anon_lock = page_lock_anon_vma_read,
1574
1575 };
1576
Sasha Levin309381fea2014-01-23 15:52:54 -08001577 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
Nick Pigginb291f002008-10-18 20:26:44 -07001578
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001579 ret = rmap_walk(page, &rwc);
1580 return ret;
Nick Pigginb291f002008-10-18 20:26:44 -07001581}
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001582
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001583void __put_anon_vma(struct anon_vma *anon_vma)
Rik van Riel76545062010-08-09 17:18:41 -07001584{
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001585 struct anon_vma *root = anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -07001586
Andrey Ryabinin624483f2014-06-06 19:09:30 +04001587 anon_vma_free(anon_vma);
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001588 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1589 anon_vma_free(root);
Rik van Riel76545062010-08-09 17:18:41 -07001590}
Rik van Riel76545062010-08-09 17:18:41 -07001591
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001592static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1593 struct rmap_walk_control *rwc)
Joonsoo Kimfaecd8d2014-01-21 15:49:46 -08001594{
1595 struct anon_vma *anon_vma;
1596
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001597 if (rwc->anon_lock)
1598 return rwc->anon_lock(page);
1599
Joonsoo Kimfaecd8d2014-01-21 15:49:46 -08001600 /*
1601 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1602 * because that depends on page_mapped(); but not all its usages
1603 * are holding mmap_sem. Users without mmap_sem are required to
1604 * take a reference count to prevent the anon_vma disappearing
1605 */
1606 anon_vma = page_anon_vma(page);
1607 if (!anon_vma)
1608 return NULL;
1609
1610 anon_vma_lock_read(anon_vma);
1611 return anon_vma;
1612}
1613
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001614/*
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001615 * rmap_walk_anon - do something to anonymous page using the object-based
1616 * rmap method
1617 * @page: the page to be handled
1618 * @rwc: control variable according to each walk type
1619 *
1620 * Find all the mappings of a page using the mapping pointer and the vma chains
1621 * contained in the anon_vma struct it points to.
1622 *
1623 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1624 * where the page was found will be held for write. So, we won't recheck
1625 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1626 * LOCKED.
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001627 */
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001628static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1629 bool locked)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001630{
1631 struct anon_vma *anon_vma;
Kirill A. Shutemova8fa41ad2017-02-24 14:57:54 -08001632 pgoff_t pgoff_start, pgoff_end;
Rik van Riel5beb4932010-03-05 13:42:07 -08001633 struct anon_vma_chain *avc;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001634 int ret = SWAP_AGAIN;
1635
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001636 if (locked) {
1637 anon_vma = page_anon_vma(page);
1638 /* anon_vma disappear under us? */
1639 VM_BUG_ON_PAGE(!anon_vma, page);
1640 } else {
1641 anon_vma = rmap_walk_anon_lock(page, rwc);
1642 }
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001643 if (!anon_vma)
1644 return ret;
Joonsoo Kimfaecd8d2014-01-21 15:49:46 -08001645
Kirill A. Shutemova8fa41ad2017-02-24 14:57:54 -08001646 pgoff_start = page_to_pgoff(page);
1647 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1648 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1649 pgoff_start, pgoff_end) {
Rik van Riel5beb4932010-03-05 13:42:07 -08001650 struct vm_area_struct *vma = avc->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001651 unsigned long address = vma_address(page, vma);
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001652
Andrea Arcangeliad126952015-11-05 18:49:07 -08001653 cond_resched();
1654
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001655 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1656 continue;
1657
Joonsoo Kim051ac832014-01-21 15:49:48 -08001658 ret = rwc->rmap_one(page, vma, address, rwc->arg);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001659 if (ret != SWAP_AGAIN)
1660 break;
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001661 if (rwc->done && rwc->done(page))
1662 break;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001663 }
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001664
1665 if (!locked)
1666 anon_vma_unlock_read(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001667 return ret;
1668}
1669
Joonsoo Kime8351ac2014-01-21 15:49:52 -08001670/*
1671 * rmap_walk_file - do something to file page using the object-based rmap method
1672 * @page: the page to be handled
1673 * @rwc: control variable according to each walk type
1674 *
1675 * Find all the mappings of a page using the mapping pointer and the vma chains
1676 * contained in the address_space struct it points to.
1677 *
1678 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1679 * where the page was found will be held for write. So, we won't recheck
1680 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1681 * LOCKED.
1682 */
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001683static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1684 bool locked)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001685{
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001686 struct address_space *mapping = page_mapping(page);
Kirill A. Shutemova8fa41ad2017-02-24 14:57:54 -08001687 pgoff_t pgoff_start, pgoff_end;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001688 struct vm_area_struct *vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001689 int ret = SWAP_AGAIN;
1690
Joonsoo Kim9f326242014-01-21 15:49:53 -08001691 /*
1692 * The page lock not only makes sure that page->mapping cannot
1693 * suddenly be NULLified by truncation, it makes sure that the
1694 * structure at mapping cannot be freed and reused yet,
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08001695 * so we can safely take mapping->i_mmap_rwsem.
Joonsoo Kim9f326242014-01-21 15:49:53 -08001696 */
Sasha Levin81d1b092014-10-09 15:28:10 -07001697 VM_BUG_ON_PAGE(!PageLocked(page), page);
Joonsoo Kim9f326242014-01-21 15:49:53 -08001698
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001699 if (!mapping)
1700 return ret;
Davidlohr Bueso3dec0ba2014-12-12 16:54:27 -08001701
Kirill A. Shutemova8fa41ad2017-02-24 14:57:54 -08001702 pgoff_start = page_to_pgoff(page);
1703 pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001704 if (!locked)
1705 i_mmap_lock_read(mapping);
Kirill A. Shutemova8fa41ad2017-02-24 14:57:54 -08001706 vma_interval_tree_foreach(vma, &mapping->i_mmap,
1707 pgoff_start, pgoff_end) {
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001708 unsigned long address = vma_address(page, vma);
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001709
Andrea Arcangeliad126952015-11-05 18:49:07 -08001710 cond_resched();
1711
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001712 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1713 continue;
1714
Joonsoo Kim051ac832014-01-21 15:49:48 -08001715 ret = rwc->rmap_one(page, vma, address, rwc->arg);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001716 if (ret != SWAP_AGAIN)
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001717 goto done;
1718 if (rwc->done && rwc->done(page))
1719 goto done;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001720 }
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001721
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001722done:
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001723 if (!locked)
1724 i_mmap_unlock_read(mapping);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001725 return ret;
1726}
1727
Joonsoo Kim051ac832014-01-21 15:49:48 -08001728int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001729{
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001730 if (unlikely(PageKsm(page)))
Joonsoo Kim051ac832014-01-21 15:49:48 -08001731 return rmap_walk_ksm(page, rwc);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001732 else if (PageAnon(page))
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001733 return rmap_walk_anon(page, rwc, false);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001734 else
Kirill A. Shutemovb9773192016-03-17 14:20:01 -07001735 return rmap_walk_file(page, rwc, false);
1736}
1737
1738/* Like rmap_walk, but caller holds relevant rmap lock */
1739int rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1740{
1741 /* no ksm support for now */
1742 VM_BUG_ON_PAGE(PageKsm(page), page);
1743 if (PageAnon(page))
1744 return rmap_walk_anon(page, rwc, true);
1745 else
1746 return rmap_walk_file(page, rwc, true);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001747}
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001748
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001749#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001750/*
1751 * The following three functions are for anonymous (private mapped) hugepages.
1752 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1753 * and no lru code, because we handle hugepages differently from common pages.
1754 */
1755static void __hugepage_set_anon_rmap(struct page *page,
1756 struct vm_area_struct *vma, unsigned long address, int exclusive)
1757{
1758 struct anon_vma *anon_vma = vma->anon_vma;
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001759
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001760 BUG_ON(!anon_vma);
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001761
1762 if (PageAnon(page))
1763 return;
1764 if (!exclusive)
1765 anon_vma = anon_vma->root;
1766
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001767 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1768 page->mapping = (struct address_space *) anon_vma;
1769 page->index = linear_page_index(vma, address);
1770}
1771
1772void hugepage_add_anon_rmap(struct page *page,
1773 struct vm_area_struct *vma, unsigned long address)
1774{
1775 struct anon_vma *anon_vma = vma->anon_vma;
1776 int first;
Naoya Horiguchia850ea32010-09-10 13:23:06 +09001777
1778 BUG_ON(!PageLocked(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001779 BUG_ON(!anon_vma);
Hugh Dickins5dbe0af2011-05-28 13:17:04 -07001780 /* address might be in next vma when migration races vma_adjust */
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001781 first = atomic_inc_and_test(compound_mapcount_ptr(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001782 if (first)
1783 __hugepage_set_anon_rmap(page, vma, address, 0);
1784}
1785
1786void hugepage_add_new_anon_rmap(struct page *page,
1787 struct vm_area_struct *vma, unsigned long address)
1788{
1789 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08001790 atomic_set(compound_mapcount_ptr(page), 0);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001791 __hugepage_set_anon_rmap(page, vma, address, 1);
1792}
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001793#endif /* CONFIG_HUGETLB_PAGE */