blob: c8454e06b6c84b424de3952c567fbbcc8264ff75 [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)
Peter Zijlstra3d48ae42011-05-24 17:12:06 -070026 * mapping->i_mmap_mutex
Peter Zijlstra2b575eb2011-05-24 17:12:11 -070027 * anon_vma->mutex
Nick Piggin82591e62006-10-19 23:29:10 -070028 * mm->page_table_lock or pte_lock
29 * zone->lru_lock (in mark_page_accessed, isolate_lru_page)
30 * swap_lock (in swap_duplicate, swap_info_get)
31 * mmlist_lock (in mmput, drain_mmlist and others)
32 * mapping->private_lock (in __set_page_dirty_buffers)
Dave Chinner250df6e2011-03-22 22:23:36 +110033 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
Christoph Hellwigf758eea2011-04-21 18:19:44 -060034 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
Nick Piggin82591e62006-10-19 23:29:10 -070035 * sb_lock (within inode_lock in fs/fs-writeback.c)
36 * mapping->tree_lock (widely used, in set_page_dirty,
37 * in arch-dependent flush_dcache_mmap_lock,
Christoph Hellwigf758eea2011-04-21 18:19:44 -060038 * within bdi.wb->list_lock in __sync_single_inode)
Andi Kleen6a460792009-09-16 11:50:15 +020039 *
Peter Zijlstra9b679322011-06-27 16:18:09 -070040 * anon_vma->mutex,mapping->i_mutex (memory_failure, collect_procs_anon)
41 * ->tasklist_lock
Andi Kleen6a460792009-09-16 11:50:15 +020042 * pte map lock
Linus Torvalds1da177e2005-04-16 15:20:36 -070043 */
44
45#include <linux/mm.h>
46#include <linux/pagemap.h>
47#include <linux/swap.h>
48#include <linux/swapops.h>
49#include <linux/slab.h>
50#include <linux/init.h>
Hugh Dickins5ad64682009-12-14 17:59:24 -080051#include <linux/ksm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070052#include <linux/rmap.h>
53#include <linux/rcupdate.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040054#include <linux/export.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080055#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070056#include <linux/mmu_notifier.h>
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -080057#include <linux/migrate.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090058#include <linux/hugetlb.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070059
60#include <asm/tlbflush.h>
61
Nick Pigginb291f002008-10-18 20:26:44 -070062#include "internal.h"
63
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070064static struct kmem_cache *anon_vma_cachep;
Rik van Riel5beb4932010-03-05 13:42:07 -080065static struct kmem_cache *anon_vma_chain_cachep;
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070066
67static inline struct anon_vma *anon_vma_alloc(void)
68{
Peter Zijlstra01d8b202011-03-22 16:32:49 -070069 struct anon_vma *anon_vma;
70
71 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
72 if (anon_vma) {
73 atomic_set(&anon_vma->refcount, 1);
74 /*
75 * Initialise the anon_vma root to point to itself. If called
76 * from fork, the root will be reset to the parents anon_vma.
77 */
78 anon_vma->root = anon_vma;
79 }
80
81 return anon_vma;
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070082}
83
Peter Zijlstra01d8b202011-03-22 16:32:49 -070084static inline void anon_vma_free(struct anon_vma *anon_vma)
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -070085{
Peter Zijlstra01d8b202011-03-22 16:32:49 -070086 VM_BUG_ON(atomic_read(&anon_vma->refcount));
Peter Zijlstra88c22082011-05-24 17:12:13 -070087
88 /*
89 * Synchronize against page_lock_anon_vma() such that
90 * we can safely hold the lock without the anon_vma getting
91 * freed.
92 *
93 * Relies on the full mb implied by the atomic_dec_and_test() from
94 * put_anon_vma() against the acquire barrier implied by
95 * mutex_trylock() from page_lock_anon_vma(). This orders:
96 *
97 * page_lock_anon_vma() VS put_anon_vma()
98 * mutex_trylock() atomic_dec_and_test()
99 * LOCK MB
100 * atomic_read() mutex_is_locked()
101 *
102 * LOCK should suffice since the actual taking of the lock must
103 * happen _before_ what follows.
104 */
105 if (mutex_is_locked(&anon_vma->root->mutex)) {
106 anon_vma_lock(anon_vma);
107 anon_vma_unlock(anon_vma);
108 }
109
Adrian Bunkfdd2e5f2008-10-18 20:28:38 -0700110 kmem_cache_free(anon_vma_cachep, anon_vma);
111}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112
Linus Torvaldsdd347392011-06-17 19:05:36 -0700113static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
Rik van Riel5beb4932010-03-05 13:42:07 -0800114{
Linus Torvaldsdd347392011-06-17 19:05:36 -0700115 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
Rik van Riel5beb4932010-03-05 13:42:07 -0800116}
117
Namhyung Kime574b5f2010-10-26 14:22:02 -0700118static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
Rik van Riel5beb4932010-03-05 13:42:07 -0800119{
120 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
121}
122
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700123/**
124 * anon_vma_prepare - attach an anon_vma to a memory region
125 * @vma: the memory region in question
126 *
127 * This makes sure the memory mapping described by 'vma' has
128 * an 'anon_vma' attached to it, so that we can associate the
129 * anonymous pages mapped into it with that anon_vma.
130 *
131 * The common case will be that we already have one, but if
Figo.zhang23a07902010-12-27 15:14:06 +0100132 * not we either need to find an adjacent mapping that we
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700133 * can re-use the anon_vma from (very common when the only
134 * reason for splitting a vma has been mprotect()), or we
135 * allocate a new one.
136 *
137 * Anon-vma allocations are very subtle, because we may have
138 * optimistically looked up an anon_vma in page_lock_anon_vma()
139 * and that may actually touch the spinlock even in the newly
140 * allocated vma (it depends on RCU to make sure that the
141 * anon_vma isn't actually destroyed).
142 *
143 * As a result, we need to do proper anon_vma locking even
144 * for the new allocation. At the same time, we do not want
145 * to do any locking for the common case of already having
146 * an anon_vma.
147 *
148 * This must be called with the mmap_sem held for reading.
149 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150int anon_vma_prepare(struct vm_area_struct *vma)
151{
152 struct anon_vma *anon_vma = vma->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800153 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154
155 might_sleep();
156 if (unlikely(!anon_vma)) {
157 struct mm_struct *mm = vma->vm_mm;
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700158 struct anon_vma *allocated;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159
Linus Torvaldsdd347392011-06-17 19:05:36 -0700160 avc = anon_vma_chain_alloc(GFP_KERNEL);
Rik van Riel5beb4932010-03-05 13:42:07 -0800161 if (!avc)
162 goto out_enomem;
163
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164 anon_vma = find_mergeable_anon_vma(vma);
Linus Torvaldsd9d332e2008-10-19 10:32:20 -0700165 allocated = NULL;
166 if (!anon_vma) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 anon_vma = anon_vma_alloc();
168 if (unlikely(!anon_vma))
Rik van Riel5beb4932010-03-05 13:42:07 -0800169 goto out_enomem_free_avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170 allocated = anon_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 }
172
Rik van Rielcba48b92010-08-09 17:18:38 -0700173 anon_vma_lock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 /* page_table_lock to protect against threads */
175 spin_lock(&mm->page_table_lock);
176 if (likely(!vma->anon_vma)) {
177 vma->anon_vma = anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800178 avc->anon_vma = anon_vma;
179 avc->vma = vma;
180 list_add(&avc->same_vma, &vma->anon_vma_chain);
Andrea Arcangeli26ba0cb2010-08-09 17:19:08 -0700181 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182 allocated = NULL;
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400183 avc = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 }
185 spin_unlock(&mm->page_table_lock);
Rik van Rielcba48b92010-08-09 17:18:38 -0700186 anon_vma_unlock(anon_vma);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400187
188 if (unlikely(allocated))
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700189 put_anon_vma(allocated);
Oleg Nesterov31f2b0e2010-04-23 13:18:01 -0400190 if (unlikely(avc))
Rik van Riel5beb4932010-03-05 13:42:07 -0800191 anon_vma_chain_free(avc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 }
193 return 0;
Rik van Riel5beb4932010-03-05 13:42:07 -0800194
195 out_enomem_free_avc:
196 anon_vma_chain_free(avc);
197 out_enomem:
198 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199}
200
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700201/*
202 * This is a useful helper function for locking the anon_vma root as
203 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
204 * have the same vma.
205 *
206 * Such anon_vma's should have the same root, so you'd expect to see
207 * just a single mutex_lock for the whole traversal.
208 */
209static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
210{
211 struct anon_vma *new_root = anon_vma->root;
212 if (new_root != root) {
213 if (WARN_ON_ONCE(root))
214 mutex_unlock(&root->mutex);
215 root = new_root;
216 mutex_lock(&root->mutex);
217 }
218 return root;
219}
220
221static inline void unlock_anon_vma_root(struct anon_vma *root)
222{
223 if (root)
224 mutex_unlock(&root->mutex);
225}
226
Rik van Riel5beb4932010-03-05 13:42:07 -0800227static void anon_vma_chain_link(struct vm_area_struct *vma,
228 struct anon_vma_chain *avc,
229 struct anon_vma *anon_vma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230{
Rik van Riel5beb4932010-03-05 13:42:07 -0800231 avc->vma = vma;
232 avc->anon_vma = anon_vma;
233 list_add(&avc->same_vma, &vma->anon_vma_chain);
234
Andrea Arcangeli05759d32011-01-13 15:46:53 -0800235 /*
236 * It's critical to add new vmas to the tail of the anon_vma,
237 * see comment in huge_memory.c:__split_huge_page().
238 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800239 list_add_tail(&avc->same_anon_vma, &anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240}
241
Rik van Riel5beb4932010-03-05 13:42:07 -0800242/*
243 * Attach the anon_vmas from src to dst.
244 * Returns 0 on success, -ENOMEM on failure.
245 */
246int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247{
Rik van Riel5beb4932010-03-05 13:42:07 -0800248 struct anon_vma_chain *avc, *pavc;
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700249 struct anon_vma *root = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250
Linus Torvalds646d87b2010-04-11 17:15:03 -0700251 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700252 struct anon_vma *anon_vma;
253
Linus Torvaldsdd347392011-06-17 19:05:36 -0700254 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
255 if (unlikely(!avc)) {
256 unlock_anon_vma_root(root);
257 root = NULL;
258 avc = anon_vma_chain_alloc(GFP_KERNEL);
259 if (!avc)
260 goto enomem_failure;
261 }
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700262 anon_vma = pavc->anon_vma;
263 root = lock_anon_vma_root(root, anon_vma);
264 anon_vma_chain_link(dst, avc, anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 }
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700266 unlock_anon_vma_root(root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800267 return 0;
268
269 enomem_failure:
270 unlink_anon_vmas(dst);
271 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272}
273
Rik van Riel5beb4932010-03-05 13:42:07 -0800274/*
Andrea Arcangeli948f0172012-01-10 15:08:05 -0800275 * Some rmap walk that needs to find all ptes/hugepmds without false
276 * negatives (like migrate and split_huge_page) running concurrent
277 * with operations that copy or move pagetables (like mremap() and
278 * fork()) to be safe. They depend on the anon_vma "same_anon_vma"
279 * list to be in a certain order: the dst_vma must be placed after the
280 * src_vma in the list. This is always guaranteed by fork() but
281 * mremap() needs to call this function to enforce it in case the
282 * dst_vma isn't newly allocated and chained with the anon_vma_clone()
283 * function but just an extension of a pre-existing vma through
284 * vma_merge.
285 *
286 * NOTE: the same_anon_vma list can still be changed by other
287 * processes while mremap runs because mremap doesn't hold the
288 * anon_vma mutex to prevent modifications to the list while it
289 * runs. All we need to enforce is that the relative order of this
290 * process vmas isn't changing (we don't care about other vmas
291 * order). Each vma corresponds to an anon_vma_chain structure so
292 * there's no risk that other processes calling anon_vma_moveto_tail()
293 * and changing the same_anon_vma list under mremap() will screw with
294 * the relative order of this process vmas in the list, because we
295 * they can't alter the order of any vma that belongs to this
296 * process. And there can't be another anon_vma_moveto_tail() running
297 * concurrently with mremap() coming from this process because we hold
298 * the mmap_sem for the whole mremap(). fork() ordering dependency
299 * also shouldn't be affected because fork() only cares that the
300 * parent vmas are placed in the list before the child vmas and
301 * anon_vma_moveto_tail() won't reorder vmas from either the fork()
302 * parent or child.
303 */
304void anon_vma_moveto_tail(struct vm_area_struct *dst)
305{
306 struct anon_vma_chain *pavc;
307 struct anon_vma *root = NULL;
308
309 list_for_each_entry_reverse(pavc, &dst->anon_vma_chain, same_vma) {
310 struct anon_vma *anon_vma = pavc->anon_vma;
311 VM_BUG_ON(pavc->vma != dst);
312 root = lock_anon_vma_root(root, anon_vma);
313 list_del(&pavc->same_anon_vma);
314 list_add_tail(&pavc->same_anon_vma, &anon_vma->head);
315 }
316 unlock_anon_vma_root(root);
317}
318
319/*
Rik van Riel5beb4932010-03-05 13:42:07 -0800320 * Attach vma to its own anon_vma, as well as to the anon_vmas that
321 * the corresponding VMA in the parent process is attached to.
322 * Returns 0 on success, non-zero on failure.
323 */
324int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325{
Rik van Riel5beb4932010-03-05 13:42:07 -0800326 struct anon_vma_chain *avc;
327 struct anon_vma *anon_vma;
328
329 /* Don't bother if the parent process has no anon_vma here. */
330 if (!pvma->anon_vma)
331 return 0;
332
333 /*
334 * First, attach the new VMA to the parent VMA's anon_vmas,
335 * so rmap can find non-COWed pages in child processes.
336 */
337 if (anon_vma_clone(vma, pvma))
338 return -ENOMEM;
339
340 /* Then add our own anon_vma. */
341 anon_vma = anon_vma_alloc();
342 if (!anon_vma)
343 goto out_error;
Linus Torvaldsdd347392011-06-17 19:05:36 -0700344 avc = anon_vma_chain_alloc(GFP_KERNEL);
Rik van Riel5beb4932010-03-05 13:42:07 -0800345 if (!avc)
346 goto out_error_free_anon_vma;
Rik van Riel5c341ee2010-08-09 17:18:39 -0700347
348 /*
349 * The root anon_vma's spinlock is the lock actually used when we
350 * lock any of the anon_vmas in this anon_vma tree.
351 */
352 anon_vma->root = pvma->anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -0700353 /*
Peter Zijlstra01d8b202011-03-22 16:32:49 -0700354 * With refcounts, an anon_vma can stay around longer than the
355 * process it belongs to. The root anon_vma needs to be pinned until
356 * this anon_vma is freed, because the lock lives in the root.
Rik van Riel76545062010-08-09 17:18:41 -0700357 */
358 get_anon_vma(anon_vma->root);
Rik van Riel5beb4932010-03-05 13:42:07 -0800359 /* Mark this anon_vma as the one where our new (COWed) pages go. */
360 vma->anon_vma = anon_vma;
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700361 anon_vma_lock(anon_vma);
Rik van Riel5c341ee2010-08-09 17:18:39 -0700362 anon_vma_chain_link(vma, avc, anon_vma);
Linus Torvaldsbb4aa392011-06-16 20:44:51 -0700363 anon_vma_unlock(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 Riel5c341ee2010-08-09 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);
387 list_del(&avc->same_anon_vma);
388
389 /*
390 * Leave empty anon_vmas on the list - we'll need
391 * to free them outside the lock.
392 */
393 if (list_empty(&anon_vma->head))
394 continue;
395
396 list_del(&avc->same_vma);
397 anon_vma_chain_free(avc);
398 }
399 unlock_anon_vma_root(root);
400
401 /*
402 * Iterate the list once more, it now only contains empty and unlinked
403 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
404 * needing to acquire the anon_vma->root->mutex.
405 */
406 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
407 struct anon_vma *anon_vma = avc->anon_vma;
408
409 put_anon_vma(anon_vma);
410
Rik van Riel5beb4932010-03-05 13:42:07 -0800411 list_del(&avc->same_vma);
412 anon_vma_chain_free(avc);
413 }
414}
415
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700416static void anon_vma_ctor(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417{
Christoph Lametera35afb82007-05-16 22:10:57 -0700418 struct anon_vma *anon_vma = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419
Peter Zijlstra2b575eb2011-05-24 17:12:11 -0700420 mutex_init(&anon_vma->mutex);
Peter Zijlstra83813262011-03-22 16:32:48 -0700421 atomic_set(&anon_vma->refcount, 0);
Christoph Lametera35afb82007-05-16 22:10:57 -0700422 INIT_LIST_HEAD(&anon_vma->head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423}
424
425void __init anon_vma_init(void)
426{
427 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
Paul Mundt20c2df82007-07-20 10:11:58 +0900428 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
Rik van Riel5beb4932010-03-05 13:42:07 -0800429 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430}
431
432/*
Peter Zijlstra6111e4c2011-05-24 17:12:08 -0700433 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
434 *
435 * Since there is no serialization what so ever against page_remove_rmap()
436 * the best this function can do is return a locked anon_vma that might
437 * have been relevant to this page.
438 *
439 * The page might have been remapped to a different anon_vma or the anon_vma
440 * returned may already be freed (and even reused).
441 *
Peter Zijlstrabc658c92011-05-29 10:33:44 +0200442 * In case it was remapped to a different anon_vma, the new anon_vma will be a
443 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
444 * ensure that any anon_vma obtained from the page will still be valid for as
445 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
446 *
Peter Zijlstra6111e4c2011-05-24 17:12:08 -0700447 * All users of this function must be very careful when walking the anon_vma
448 * chain and verify that the page in question is indeed mapped in it
449 * [ something equivalent to page_mapped_in_vma() ].
450 *
451 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
452 * that the anon_vma pointer from page->mapping is valid if there is a
453 * mapcount, we can dereference the anon_vma after observing those.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700455struct anon_vma *page_get_anon_vma(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456{
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700457 struct anon_vma *anon_vma = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 unsigned long anon_mapping;
459
460 rcu_read_lock();
Hugh Dickins80e14822009-12-14 17:59:29 -0800461 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800462 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 goto out;
464 if (!page_mapped(page))
465 goto out;
466
467 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700468 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
469 anon_vma = NULL;
470 goto out;
471 }
Hugh Dickinsf1819422010-08-25 23:12:54 -0700472
473 /*
474 * If this page is still mapped, then its anon_vma cannot have been
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700475 * freed. But if it has been unmapped, we have no security against the
476 * anon_vma structure being freed and reused (for another anon_vma:
477 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
478 * above cannot corrupt).
Hugh Dickinsf1819422010-08-25 23:12:54 -0700479 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700480 if (!page_mapped(page)) {
481 put_anon_vma(anon_vma);
482 anon_vma = NULL;
483 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484out:
485 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700486
487 return anon_vma;
488}
489
Peter Zijlstra88c22082011-05-24 17:12:13 -0700490/*
491 * Similar to page_get_anon_vma() except it locks the anon_vma.
492 *
493 * Its a little more complex as it tries to keep the fast path to a single
494 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
495 * reference like with page_get_anon_vma() and then block on the mutex.
496 */
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700497struct anon_vma *page_lock_anon_vma(struct page *page)
498{
Peter Zijlstra88c22082011-05-24 17:12:13 -0700499 struct anon_vma *anon_vma = NULL;
Hugh Dickinseee0f252011-05-28 13:20:21 -0700500 struct anon_vma *root_anon_vma;
Peter Zijlstra88c22082011-05-24 17:12:13 -0700501 unsigned long anon_mapping;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700502
Peter Zijlstra88c22082011-05-24 17:12:13 -0700503 rcu_read_lock();
504 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
505 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
506 goto out;
507 if (!page_mapped(page))
508 goto out;
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700509
Peter Zijlstra88c22082011-05-24 17:12:13 -0700510 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
Hugh Dickinseee0f252011-05-28 13:20:21 -0700511 root_anon_vma = ACCESS_ONCE(anon_vma->root);
512 if (mutex_trylock(&root_anon_vma->mutex)) {
Peter Zijlstra88c22082011-05-24 17:12:13 -0700513 /*
Hugh Dickinseee0f252011-05-28 13:20:21 -0700514 * If the page is still mapped, then this anon_vma is still
515 * its anon_vma, and holding the mutex ensures that it will
Peter Zijlstrabc658c92011-05-29 10:33:44 +0200516 * not go away, see anon_vma_free().
Peter Zijlstra88c22082011-05-24 17:12:13 -0700517 */
Hugh Dickinseee0f252011-05-28 13:20:21 -0700518 if (!page_mapped(page)) {
519 mutex_unlock(&root_anon_vma->mutex);
Peter Zijlstra88c22082011-05-24 17:12:13 -0700520 anon_vma = NULL;
521 }
522 goto out;
523 }
524
525 /* trylock failed, we got to sleep */
526 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
527 anon_vma = NULL;
528 goto out;
529 }
530
531 if (!page_mapped(page)) {
532 put_anon_vma(anon_vma);
533 anon_vma = NULL;
534 goto out;
535 }
536
537 /* we pinned the anon_vma, its safe to sleep */
538 rcu_read_unlock();
539 anon_vma_lock(anon_vma);
540
541 if (atomic_dec_and_test(&anon_vma->refcount)) {
542 /*
543 * Oops, we held the last refcount, release the lock
544 * and bail -- can't simply use put_anon_vma() because
545 * we'll deadlock on the anon_vma_lock() recursion.
546 */
547 anon_vma_unlock(anon_vma);
548 __put_anon_vma(anon_vma);
549 anon_vma = NULL;
550 }
551
552 return anon_vma;
553
554out:
555 rcu_read_unlock();
Peter Zijlstra746b18d2011-05-24 17:12:10 -0700556 return anon_vma;
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800557}
558
Andi Kleen10be22d2009-09-16 11:50:04 +0200559void page_unlock_anon_vma(struct anon_vma *anon_vma)
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800560{
Rik van Rielcba48b92010-08-09 17:18:38 -0700561 anon_vma_unlock(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562}
563
564/*
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800565 * At what user virtual address is page expected in @vma?
566 * Returns virtual address or -EFAULT if page's index/offset is not
567 * within the range mapped the @vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800569inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570vma_address(struct page *page, struct vm_area_struct *vma)
571{
572 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
573 unsigned long address;
574
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900575 if (unlikely(is_vm_hugetlb_page(vma)))
576 pgoff = page->index << huge_page_order(page_hstate(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
578 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800579 /* page should be within @vma mapping range */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 return -EFAULT;
581 }
582 return address;
583}
584
585/*
Huang Shijiebf89c8c2009-10-01 15:44:04 -0700586 * At what user virtual address is page expected in vma?
Naoya Horiguchiab941e02010-05-11 14:06:55 -0700587 * Caller should check the page is actually part of the vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 */
589unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
590{
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700591 if (PageAnon(page)) {
Hugh Dickins4829b902010-10-02 17:46:06 -0700592 struct anon_vma *page__anon_vma = page_anon_vma(page);
593 /*
594 * Note: swapoff's unuse_vma() is more efficient with this
595 * check, and needs it to match anon_vma when KSM is active.
596 */
597 if (!vma->anon_vma || !page__anon_vma ||
598 vma->anon_vma->root != page__anon_vma->root)
Andrea Arcangeli21d0d442010-08-09 17:19:10 -0700599 return -EFAULT;
600 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800601 if (!vma->vm_file ||
602 vma->vm_file->f_mapping != page->mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603 return -EFAULT;
604 } else
605 return -EFAULT;
606 return vma_address(page, vma);
607}
608
609/*
Nikita Danilov81b40822005-05-01 08:58:36 -0700610 * Check that @page is mapped at @address into @mm.
611 *
Nick Piggin479db0b2008-08-20 14:09:18 -0700612 * If @sync is false, page_check_address may perform a racy check to avoid
613 * the page table lock when the pte is not present (helpful when reclaiming
614 * highly shared pages).
615 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -0700616 * On success returns with pte mapped and locked.
Nikita Danilov81b40822005-05-01 08:58:36 -0700617 */
Namhyung Kime9a81a82010-10-26 14:22:01 -0700618pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
Nick Piggin479db0b2008-08-20 14:09:18 -0700619 unsigned long address, spinlock_t **ptlp, int sync)
Nikita Danilov81b40822005-05-01 08:58:36 -0700620{
621 pgd_t *pgd;
622 pud_t *pud;
623 pmd_t *pmd;
624 pte_t *pte;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700625 spinlock_t *ptl;
Nikita Danilov81b40822005-05-01 08:58:36 -0700626
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900627 if (unlikely(PageHuge(page))) {
628 pte = huge_pte_offset(mm, address);
629 ptl = &mm->page_table_lock;
630 goto check;
631 }
632
Nikita Danilov81b40822005-05-01 08:58:36 -0700633 pgd = pgd_offset(mm, address);
Hugh Dickinsc0718802005-10-29 18:16:31 -0700634 if (!pgd_present(*pgd))
635 return NULL;
636
637 pud = pud_offset(pgd, address);
638 if (!pud_present(*pud))
639 return NULL;
640
641 pmd = pmd_offset(pud, address);
642 if (!pmd_present(*pmd))
643 return NULL;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800644 if (pmd_trans_huge(*pmd))
645 return NULL;
Hugh Dickinsc0718802005-10-29 18:16:31 -0700646
647 pte = pte_offset_map(pmd, address);
648 /* Make a quick check before getting the lock */
Nick Piggin479db0b2008-08-20 14:09:18 -0700649 if (!sync && !pte_present(*pte)) {
Hugh Dickinsc0718802005-10-29 18:16:31 -0700650 pte_unmap(pte);
651 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700652 }
Hugh Dickinsc0718802005-10-29 18:16:31 -0700653
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700654 ptl = pte_lockptr(mm, pmd);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900655check:
Hugh Dickinsc0718802005-10-29 18:16:31 -0700656 spin_lock(ptl);
657 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
658 *ptlp = ptl;
659 return pte;
660 }
661 pte_unmap_unlock(pte, ptl);
662 return NULL;
Nikita Danilov81b40822005-05-01 08:58:36 -0700663}
664
Nick Pigginb291f002008-10-18 20:26:44 -0700665/**
666 * page_mapped_in_vma - check whether a page is really mapped in a VMA
667 * @page: the page to test
668 * @vma: the VMA to test
669 *
670 * Returns 1 if the page is mapped into the page tables of the VMA, 0
671 * if the page is not mapped into the page tables of this VMA. Only
672 * valid for normal file or anonymous VMAs.
673 */
Andi Kleen6a460792009-09-16 11:50:15 +0200674int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
Nick Pigginb291f002008-10-18 20:26:44 -0700675{
676 unsigned long address;
677 pte_t *pte;
678 spinlock_t *ptl;
679
680 address = vma_address(page, vma);
681 if (address == -EFAULT) /* out of vma range */
682 return 0;
683 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
684 if (!pte) /* the page is not in this mm */
685 return 0;
686 pte_unmap_unlock(pte, ptl);
687
688 return 1;
689}
690
Nikita Danilov81b40822005-05-01 08:58:36 -0700691/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 * Subfunctions of page_referenced: page_referenced_one called
693 * repeatedly from either page_referenced_anon or page_referenced_file.
694 */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800695int page_referenced_one(struct page *page, struct vm_area_struct *vma,
696 unsigned long address, unsigned int *mapcount,
697 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698{
699 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 int referenced = 0;
701
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800702 if (unlikely(PageTransHuge(page))) {
703 pmd_t *pmd;
704
705 spin_lock(&mm->page_table_lock);
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800706 /*
707 * rmap might return false positives; we must filter
708 * these out using page_check_address_pmd().
709 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800710 pmd = page_check_address_pmd(page, mm, address,
711 PAGE_CHECK_ADDRESS_PMD_FLAG);
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800712 if (!pmd) {
713 spin_unlock(&mm->page_table_lock);
714 goto out;
715 }
716
717 if (vma->vm_flags & VM_LOCKED) {
718 spin_unlock(&mm->page_table_lock);
719 *mapcount = 0; /* break early from loop */
720 *vm_flags |= VM_LOCKED;
721 goto out;
722 }
723
724 /* go ahead even if the pmd is pmd_trans_splitting() */
725 if (pmdp_clear_flush_young_notify(vma, address, pmd))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800726 referenced++;
727 spin_unlock(&mm->page_table_lock);
728 } else {
729 pte_t *pte;
730 spinlock_t *ptl;
731
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800732 /*
733 * rmap might return false positives; we must filter
734 * these out using page_check_address().
735 */
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800736 pte = page_check_address(page, mm, address, &ptl, 0);
737 if (!pte)
738 goto out;
739
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800740 if (vma->vm_flags & VM_LOCKED) {
741 pte_unmap_unlock(pte, ptl);
742 *mapcount = 0; /* break early from loop */
743 *vm_flags |= VM_LOCKED;
744 goto out;
745 }
746
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800747 if (ptep_clear_flush_young_notify(vma, address, pte)) {
748 /*
749 * Don't treat a reference through a sequentially read
750 * mapping as such. If the page has been used in
751 * another mapping, we will catch it; if this other
752 * mapping is already gone, the unmap path will have
753 * set PG_referenced or activated the page.
754 */
755 if (likely(!VM_SequentialReadHint(vma)))
756 referenced++;
757 }
758 pte_unmap_unlock(pte, ptl);
759 }
760
Andrea Arcangeli2da28bf2011-03-11 14:58:29 -0800761 /* Pretend the page is referenced if the task has the
762 swap token and is in the middle of a page fault. */
763 if (mm != current->mm && has_swap_token(mm) &&
764 rwsem_is_locked(&mm->mmap_sem))
765 referenced++;
766
Hugh Dickinsc0718802005-10-29 18:16:31 -0700767 (*mapcount)--;
Huang Shijie273f0472009-12-14 17:58:51 -0800768
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700769 if (referenced)
770 *vm_flags |= vma->vm_flags;
Huang Shijie273f0472009-12-14 17:58:51 -0800771out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 return referenced;
773}
774
Balbir Singhbed71612008-02-07 00:14:01 -0800775static int page_referenced_anon(struct page *page,
Johannes Weiner72835c82012-01-12 17:18:32 -0800776 struct mem_cgroup *memcg,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700777 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778{
779 unsigned int mapcount;
780 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -0800781 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782 int referenced = 0;
783
784 anon_vma = page_lock_anon_vma(page);
785 if (!anon_vma)
786 return referenced;
787
788 mapcount = page_mapcount(page);
Rik van Riel5beb4932010-03-05 13:42:07 -0800789 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
790 struct vm_area_struct *vma = avc->vma;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800791 unsigned long address = vma_address(page, vma);
792 if (address == -EFAULT)
793 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800794 /*
795 * If we are reclaiming on behalf of a cgroup, skip
796 * counting on behalf of references from different
797 * cgroups
798 */
Johannes Weiner72835c82012-01-12 17:18:32 -0800799 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
Balbir Singhbed71612008-02-07 00:14:01 -0800800 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800801 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700802 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 if (!mapcount)
804 break;
805 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -0800806
807 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 return referenced;
809}
810
811/**
812 * page_referenced_file - referenced check for object-based rmap
813 * @page: the page we're checking references on.
Johannes Weiner72835c82012-01-12 17:18:32 -0800814 * @memcg: target memory control group
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700815 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 *
817 * For an object-based mapped page, find all the places it is mapped and
818 * check/clear the referenced flag. This is done by following the page->mapping
819 * pointer, then walking the chain of vmas it holds. It returns the number
820 * of references it found.
821 *
822 * This function is only called from page_referenced for object-based pages.
823 */
Balbir Singhbed71612008-02-07 00:14:01 -0800824static int page_referenced_file(struct page *page,
Johannes Weiner72835c82012-01-12 17:18:32 -0800825 struct mem_cgroup *memcg,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700826 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827{
828 unsigned int mapcount;
829 struct address_space *mapping = page->mapping;
830 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
831 struct vm_area_struct *vma;
832 struct prio_tree_iter iter;
833 int referenced = 0;
834
835 /*
836 * The caller's checks on page->mapping and !PageAnon have made
837 * sure that this is a file page: the check for page->mapping
838 * excludes the case just before it gets set on an anon page.
839 */
840 BUG_ON(PageAnon(page));
841
842 /*
843 * The page lock not only makes sure that page->mapping cannot
844 * suddenly be NULLified by truncation, it makes sure that the
845 * structure at mapping cannot be freed and reused yet,
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700846 * so we can safely take mapping->i_mmap_mutex.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847 */
848 BUG_ON(!PageLocked(page));
849
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700850 mutex_lock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851
852 /*
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700853 * i_mmap_mutex does not stabilize mapcount at all, but mapcount
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 * is more likely to be accurate if we note it after spinning.
855 */
856 mapcount = page_mapcount(page);
857
858 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800859 unsigned long address = vma_address(page, vma);
860 if (address == -EFAULT)
861 continue;
Balbir Singhbed71612008-02-07 00:14:01 -0800862 /*
863 * If we are reclaiming on behalf of a cgroup, skip
864 * counting on behalf of references from different
865 * cgroups
866 */
Johannes Weiner72835c82012-01-12 17:18:32 -0800867 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
Balbir Singhbed71612008-02-07 00:14:01 -0800868 continue;
Hugh Dickins1cb17292009-12-14 17:59:01 -0800869 referenced += page_referenced_one(page, vma, address,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700870 &mapcount, vm_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871 if (!mapcount)
872 break;
873 }
874
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700875 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 return referenced;
877}
878
879/**
880 * page_referenced - test if the page was referenced
881 * @page: the page to test
882 * @is_locked: caller holds lock on the page
Johannes Weiner72835c82012-01-12 17:18:32 -0800883 * @memcg: target memory cgroup
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700884 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 *
886 * Quick test_and_clear_referenced for all mappings to a page,
887 * returns the number of ptes which referenced the page.
888 */
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700889int page_referenced(struct page *page,
890 int is_locked,
Johannes Weiner72835c82012-01-12 17:18:32 -0800891 struct mem_cgroup *memcg,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700892 unsigned long *vm_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893{
894 int referenced = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800895 int we_locked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700897 *vm_flags = 0;
Hugh Dickins3ca7b3c2009-12-14 17:58:57 -0800898 if (page_mapped(page) && page_rmapping(page)) {
Hugh Dickins5ad64682009-12-14 17:59:24 -0800899 if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
900 we_locked = trylock_page(page);
901 if (!we_locked) {
902 referenced++;
903 goto out;
904 }
905 }
906 if (unlikely(PageKsm(page)))
Johannes Weiner72835c82012-01-12 17:18:32 -0800907 referenced += page_referenced_ksm(page, memcg,
Hugh Dickins5ad64682009-12-14 17:59:24 -0800908 vm_flags);
909 else if (PageAnon(page))
Johannes Weiner72835c82012-01-12 17:18:32 -0800910 referenced += page_referenced_anon(page, memcg,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700911 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800912 else if (page->mapping)
Johannes Weiner72835c82012-01-12 17:18:32 -0800913 referenced += page_referenced_file(page, memcg,
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700914 vm_flags);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800915 if (we_locked)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 unlock_page(page);
Martin Schwidefsky50a15982011-07-24 10:47:59 +0200917
918 if (page_test_and_clear_young(page_to_pfn(page)))
919 referenced++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 }
Hugh Dickins5ad64682009-12-14 17:59:24 -0800921out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 return referenced;
923}
924
Hugh Dickins1cb17292009-12-14 17:59:01 -0800925static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
926 unsigned long address)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700927{
928 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100929 pte_t *pte;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700930 spinlock_t *ptl;
931 int ret = 0;
932
Nick Piggin479db0b2008-08-20 14:09:18 -0700933 pte = page_check_address(page, mm, address, &ptl, 1);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700934 if (!pte)
935 goto out;
936
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100937 if (pte_dirty(*pte) || pte_write(*pte)) {
938 pte_t entry;
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700939
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100940 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700941 entry = ptep_clear_flush_notify(vma, address, pte);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100942 entry = pte_wrprotect(entry);
943 entry = pte_mkclean(entry);
Al Virod6e88e62006-12-29 16:48:35 -0800944 set_pte_at(mm, address, pte, entry);
Peter Zijlstrac2fda5f2006-12-22 14:25:52 +0100945 ret = 1;
946 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700947
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700948 pte_unmap_unlock(pte, ptl);
949out:
950 return ret;
951}
952
953static int page_mkclean_file(struct address_space *mapping, struct page *page)
954{
955 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
956 struct vm_area_struct *vma;
957 struct prio_tree_iter iter;
958 int ret = 0;
959
960 BUG_ON(PageAnon(page));
961
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700962 mutex_lock(&mapping->i_mmap_mutex);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700963 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -0800964 if (vma->vm_flags & VM_SHARED) {
965 unsigned long address = vma_address(page, vma);
966 if (address == -EFAULT)
967 continue;
968 ret += page_mkclean_one(page, vma, address);
969 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700970 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -0700971 mutex_unlock(&mapping->i_mmap_mutex);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700972 return ret;
973}
974
975int page_mkclean(struct page *page)
976{
977 int ret = 0;
978
979 BUG_ON(!PageLocked(page));
980
981 if (page_mapped(page)) {
982 struct address_space *mapping = page_mapping(page);
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100983 if (mapping) {
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700984 ret = page_mkclean_file(mapping, page);
Martin Schwidefsky2d425522011-05-23 10:24:39 +0200985 if (page_test_and_clear_dirty(page_to_pfn(page), 1))
Christian Borntraegerce7e9fa2007-11-20 11:13:36 +0100986 ret = 1;
Martin Schwidefsky6c210482007-04-27 16:01:57 +0200987 }
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700988 }
989
990 return ret;
991}
Jaya Kumar60b59be2007-05-08 00:37:37 -0700992EXPORT_SYMBOL_GPL(page_mkclean);
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700993
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994/**
Rik van Rielc44b6742010-03-05 13:42:09 -0800995 * page_move_anon_rmap - move a page to our anon_vma
996 * @page: the page to move to our anon_vma
997 * @vma: the vma the page belongs to
998 * @address: the user virtual address mapped
999 *
1000 * When a page belongs exclusively to one process after a COW event,
1001 * that page can be moved into the anon_vma that belongs to just that
1002 * process, so the rmap code will not search the parent or sibling
1003 * processes.
1004 */
1005void page_move_anon_rmap(struct page *page,
1006 struct vm_area_struct *vma, unsigned long address)
1007{
1008 struct anon_vma *anon_vma = vma->anon_vma;
1009
1010 VM_BUG_ON(!PageLocked(page));
1011 VM_BUG_ON(!anon_vma);
1012 VM_BUG_ON(page->index != linear_page_index(vma, address));
1013
1014 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1015 page->mapping = (struct address_space *) anon_vma;
1016}
1017
1018/**
Andi Kleen4e1c1972010-09-22 12:43:56 +02001019 * __page_set_anon_rmap - set up new anonymous rmap
1020 * @page: Page to add to rmap
1021 * @vma: VM area to add page to.
1022 * @address: User virtual address of the mapping
Rik van Riele8a03fe2010-04-14 17:59:28 -04001023 * @exclusive: the page is exclusively owned by the current process
Nick Piggin9617d952006-01-06 00:11:12 -08001024 */
1025static void __page_set_anon_rmap(struct page *page,
Rik van Riele8a03fe2010-04-14 17:59:28 -04001026 struct vm_area_struct *vma, unsigned long address, int exclusive)
Nick Piggin9617d952006-01-06 00:11:12 -08001027{
Rik van Riele8a03fe2010-04-14 17:59:28 -04001028 struct anon_vma *anon_vma = vma->anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -08001029
Rik van Riele8a03fe2010-04-14 17:59:28 -04001030 BUG_ON(!anon_vma);
Linus Torvaldsea900022010-04-12 12:44:29 -07001031
Andi Kleen4e1c1972010-09-22 12:43:56 +02001032 if (PageAnon(page))
1033 return;
1034
Linus Torvaldsea900022010-04-12 12:44:29 -07001035 /*
Rik van Riele8a03fe2010-04-14 17:59:28 -04001036 * If the page isn't exclusively mapped into this vma,
1037 * we must use the _oldest_ possible anon_vma for the
1038 * page mapping!
Linus Torvaldsea900022010-04-12 12:44:29 -07001039 */
Andi Kleen4e1c1972010-09-22 12:43:56 +02001040 if (!exclusive)
Andrea Arcangeli288468c2010-08-09 17:19:09 -07001041 anon_vma = anon_vma->root;
Linus Torvaldsea900022010-04-12 12:44:29 -07001042
Nick Piggin9617d952006-01-06 00:11:12 -08001043 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1044 page->mapping = (struct address_space *) anon_vma;
Nick Piggin9617d952006-01-06 00:11:12 -08001045 page->index = linear_page_index(vma, address);
Nick Piggin9617d952006-01-06 00:11:12 -08001046}
1047
1048/**
Randy Dunlap43d8eac2008-03-19 17:00:43 -07001049 * __page_check_anon_rmap - sanity check anonymous rmap addition
Nick Pigginc97a9e12007-05-16 22:11:21 -07001050 * @page: the page to add the mapping to
1051 * @vma: the vm area in which the mapping is added
1052 * @address: the user virtual address mapped
1053 */
1054static void __page_check_anon_rmap(struct page *page,
1055 struct vm_area_struct *vma, unsigned long address)
1056{
1057#ifdef CONFIG_DEBUG_VM
1058 /*
1059 * The page's anon-rmap details (mapping and index) are guaranteed to
1060 * be set up correctly at this point.
1061 *
1062 * We have exclusion against page_add_anon_rmap because the caller
1063 * always holds the page locked, except if called from page_dup_rmap,
1064 * in which case the page is already known to be setup.
1065 *
1066 * We have exclusion against page_add_new_anon_rmap because those pages
1067 * are initially only visible via the pagetables, and the pte is locked
1068 * over the call to page_add_new_anon_rmap.
1069 */
Andrea Arcangeli44ab57a2010-08-09 17:19:10 -07001070 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
Nick Pigginc97a9e12007-05-16 22:11:21 -07001071 BUG_ON(page->index != linear_page_index(vma, address));
1072#endif
1073}
1074
1075/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 * page_add_anon_rmap - add pte mapping to an anonymous page
1077 * @page: the page to add the mapping to
1078 * @vma: the vm area in which the mapping is added
1079 * @address: the user virtual address mapped
1080 *
Hugh Dickins5ad64682009-12-14 17:59:24 -08001081 * The caller needs to hold the pte lock, and the page must be locked in
Hugh Dickins80e14822009-12-14 17:59:29 -08001082 * the anon_vma case: to serialize mapping,index checking after setting,
1083 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1084 * (but PageKsm is never downgraded to PageAnon).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085 */
1086void page_add_anon_rmap(struct page *page,
1087 struct vm_area_struct *vma, unsigned long address)
1088{
Rik van Rielad8c2ee2010-08-09 17:19:48 -07001089 do_page_add_anon_rmap(page, vma, address, 0);
1090}
1091
1092/*
1093 * Special version of the above for do_swap_page, which often runs
1094 * into pages that are exclusively owned by the current process.
1095 * Everybody else should continue to use page_add_anon_rmap above.
1096 */
1097void do_page_add_anon_rmap(struct page *page,
1098 struct vm_area_struct *vma, unsigned long address, int exclusive)
1099{
Hugh Dickins5ad64682009-12-14 17:59:24 -08001100 int first = atomic_inc_and_test(&page->_mapcount);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001101 if (first) {
1102 if (!PageTransHuge(page))
1103 __inc_zone_page_state(page, NR_ANON_PAGES);
1104 else
1105 __inc_zone_page_state(page,
1106 NR_ANON_TRANSPARENT_HUGEPAGES);
1107 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001108 if (unlikely(PageKsm(page)))
1109 return;
1110
Nick Pigginc97a9e12007-05-16 22:11:21 -07001111 VM_BUG_ON(!PageLocked(page));
Hugh Dickins5dbe0af2011-05-28 13:17:04 -07001112 /* address might be in next vma when migration races vma_adjust */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001113 if (first)
Rik van Rielad8c2ee2010-08-09 17:19:48 -07001114 __page_set_anon_rmap(page, vma, address, exclusive);
KAMEZAWA Hiroyuki69029cd2008-07-25 01:47:14 -07001115 else
Nick Pigginc97a9e12007-05-16 22:11:21 -07001116 __page_check_anon_rmap(page, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117}
1118
Randy Dunlap43d8eac2008-03-19 17:00:43 -07001119/**
Nick Piggin9617d952006-01-06 00:11:12 -08001120 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1121 * @page: the page to add the mapping to
1122 * @vma: the vm area in which the mapping is added
1123 * @address: the user virtual address mapped
1124 *
1125 * Same as page_add_anon_rmap but must only be called on *new* pages.
1126 * This means the inc-and-test can be bypassed.
Nick Pigginc97a9e12007-05-16 22:11:21 -07001127 * Page does not have to be locked.
Nick Piggin9617d952006-01-06 00:11:12 -08001128 */
1129void page_add_new_anon_rmap(struct page *page,
1130 struct vm_area_struct *vma, unsigned long address)
1131{
Hugh Dickinsb5934c52009-01-06 14:39:25 -08001132 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
Hugh Dickinscbf84b72009-01-06 14:39:27 -08001133 SetPageSwapBacked(page);
1134 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
Andrea Arcangeli79134172011-01-13 15:46:58 -08001135 if (!PageTransHuge(page))
1136 __inc_zone_page_state(page, NR_ANON_PAGES);
1137 else
1138 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
Rik van Riele8a03fe2010-04-14 17:59:28 -04001139 __page_set_anon_rmap(page, vma, address, 1);
Hugh Dickinsb5934c52009-01-06 14:39:25 -08001140 if (page_evictable(page, vma))
Hugh Dickinscbf84b72009-01-06 14:39:27 -08001141 lru_cache_add_lru(page, LRU_ACTIVE_ANON);
Hugh Dickinsb5934c52009-01-06 14:39:25 -08001142 else
1143 add_page_to_unevictable_list(page);
Nick Piggin9617d952006-01-06 00:11:12 -08001144}
1145
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146/**
1147 * page_add_file_rmap - add pte mapping to a file page
1148 * @page: the page to add the mapping to
1149 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -07001150 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 */
1152void page_add_file_rmap(struct page *page)
1153{
Balbir Singhd69b0422009-06-17 16:26:34 -07001154 if (atomic_inc_and_test(&page->_mapcount)) {
Christoph Lameter65ba55f2006-06-30 01:55:34 -07001155 __inc_zone_page_state(page, NR_FILE_MAPPED);
Greg Thelen2a7106f2011-01-13 15:47:37 -08001156 mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
Balbir Singhd69b0422009-06-17 16:26:34 -07001157 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158}
1159
1160/**
1161 * page_remove_rmap - take down pte mapping from a page
1162 * @page: page to remove mapping from
1163 *
Hugh Dickinsb8072f02005-10-29 18:16:41 -07001164 * The caller needs to hold the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165 */
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001166void page_remove_rmap(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167{
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001168 /* page still mapped by someone else? */
1169 if (!atomic_add_negative(-1, &page->_mapcount))
1170 return;
1171
1172 /*
1173 * Now that the last pte has gone, s390 must transfer dirty
1174 * flag from storage key to struct page. We can usually skip
1175 * this if the page is anon, so about to be freed; but perhaps
1176 * not if it's in swapcache - there might be another pte slot
1177 * containing the swap entry, but page not yet written to swap.
1178 */
Martin Schwidefsky2d425522011-05-23 10:24:39 +02001179 if ((!PageAnon(page) || PageSwapCache(page)) &&
1180 page_test_and_clear_dirty(page_to_pfn(page), 1))
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001181 set_page_dirty(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001182 /*
1183 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1184 * and not charged by memcg for now.
1185 */
1186 if (unlikely(PageHuge(page)))
1187 return;
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001188 if (PageAnon(page)) {
1189 mem_cgroup_uncharge_page(page);
Andrea Arcangeli79134172011-01-13 15:46:58 -08001190 if (!PageTransHuge(page))
1191 __dec_zone_page_state(page, NR_ANON_PAGES);
1192 else
1193 __dec_zone_page_state(page,
1194 NR_ANON_TRANSPARENT_HUGEPAGES);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001195 } else {
1196 __dec_zone_page_state(page, NR_FILE_MAPPED);
Greg Thelen2a7106f2011-01-13 15:47:37 -08001197 mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001198 }
KOSAKI Motohirob904dcf2009-09-21 17:01:28 -07001199 /*
1200 * It would be tidy to reset the PageAnon mapping here,
1201 * but that might overwrite a racing page_add_anon_rmap
1202 * which increments mapcount after us but sets mapping
1203 * before us: so leave the reset to free_hot_cold_page,
1204 * and remember that it's only reliable while mapped.
1205 * Leaving it set also helps swapoff to reinstate ptes
1206 * faster for those pages still in swapcache.
1207 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208}
1209
1210/*
1211 * Subfunctions of try_to_unmap: try_to_unmap_one called
Wanlong Gao99ef0312011-10-31 17:08:51 -07001212 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001214int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1215 unsigned long address, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216{
1217 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 pte_t *pte;
1219 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001220 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 int ret = SWAP_AGAIN;
1222
Nick Piggin479db0b2008-08-20 14:09:18 -07001223 pte = page_check_address(page, mm, address, &ptl, 0);
Hugh Dickinsc0718802005-10-29 18:16:31 -07001224 if (!pte)
Nikita Danilov81b40822005-05-01 08:58:36 -07001225 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226
1227 /*
1228 * If the page is mlock()d, we cannot swap it out.
1229 * If it's recently referenced (perhaps page_referenced
1230 * skipped over this mm) then we should reactivate it.
1231 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001232 if (!(flags & TTU_IGNORE_MLOCK)) {
KOSAKI Motohirocaed0f42009-12-14 17:59:45 -08001233 if (vma->vm_flags & VM_LOCKED)
1234 goto out_mlock;
1235
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001236 if (TTU_ACTION(flags) == TTU_MUNLOCK)
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001237 goto out_unmap;
Andi Kleen14fa31b2009-09-16 11:50:10 +02001238 }
1239 if (!(flags & TTU_IGNORE_ACCESS)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001240 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1241 ret = SWAP_FAIL;
1242 goto out_unmap;
1243 }
1244 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 /* Nuke the page table entry. */
1247 flush_cache_page(vma, address, page_to_pfn(page));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001248 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249
1250 /* Move the dirty bit to the physical page now the pte is gone. */
1251 if (pte_dirty(pteval))
1252 set_page_dirty(page);
1253
Hugh Dickins365e9c872005-10-29 18:16:18 -07001254 /* Update high watermark before we lower rss */
1255 update_hiwater_rss(mm);
1256
Andi Kleen888b9f72009-09-16 11:50:11 +02001257 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1258 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001259 dec_mm_counter(mm, MM_ANONPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001260 else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001261 dec_mm_counter(mm, MM_FILEPAGES);
Andi Kleen888b9f72009-09-16 11:50:11 +02001262 set_pte_at(mm, address, pte,
1263 swp_entry_to_pte(make_hwpoison_entry(page)));
1264 } else if (PageAnon(page)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001265 swp_entry_t entry = { .val = page_private(page) };
Christoph Lameter06972122006-06-23 02:03:35 -07001266
1267 if (PageSwapCache(page)) {
1268 /*
1269 * Store the swap location in the pte.
1270 * See handle_pte_fault() ...
1271 */
Hugh Dickins570a3352009-12-14 17:58:46 -08001272 if (swap_duplicate(entry) < 0) {
1273 set_pte_at(mm, address, pte, pteval);
1274 ret = SWAP_FAIL;
1275 goto out_unmap;
1276 }
Christoph Lameter06972122006-06-23 02:03:35 -07001277 if (list_empty(&mm->mmlist)) {
1278 spin_lock(&mmlist_lock);
1279 if (list_empty(&mm->mmlist))
1280 list_add(&mm->mmlist, &init_mm.mmlist);
1281 spin_unlock(&mmlist_lock);
1282 }
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001283 dec_mm_counter(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001284 inc_mm_counter(mm, MM_SWAPENTS);
KOSAKI Motohiro64cdd542009-01-06 14:39:16 -08001285 } else if (PAGE_MIGRATION) {
Christoph Lameter06972122006-06-23 02:03:35 -07001286 /*
1287 * Store the pfn of the page in a special migration
1288 * pte. do_swap_page() will wait until the migration
1289 * pte is removed and then restart fault handling.
1290 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001291 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
Christoph Lameter06972122006-06-23 02:03:35 -07001292 entry = make_migration_entry(page, pte_write(pteval));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 }
1294 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1295 BUG_ON(pte_file(*pte));
Andi Kleen14fa31b2009-09-16 11:50:10 +02001296 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
Christoph Lameter04e62a22006-06-23 02:03:38 -07001297 /* Establish migration entry for a file page */
1298 swp_entry_t entry;
1299 entry = make_migration_entry(page, pte_write(pteval));
1300 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1301 } else
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001302 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001304 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 page_cache_release(page);
1306
1307out_unmap:
Hugh Dickinsc0718802005-10-29 18:16:31 -07001308 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309out:
1310 return ret;
KOSAKI Motohirocaed0f42009-12-14 17:59:45 -08001311
1312out_mlock:
1313 pte_unmap_unlock(pte, ptl);
1314
1315
1316 /*
1317 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1318 * unstable result and race. Plus, We can't wait here because
Peter Zijlstra2b575eb2011-05-24 17:12:11 -07001319 * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
KOSAKI Motohirocaed0f42009-12-14 17:59:45 -08001320 * if trylock failed, the page remain in evictable lru and later
1321 * vmscan could retry to move the page to unevictable lru if the
1322 * page is actually mlocked.
1323 */
1324 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1325 if (vma->vm_flags & VM_LOCKED) {
1326 mlock_vma_page(page);
1327 ret = SWAP_MLOCK;
1328 }
1329 up_read(&vma->vm_mm->mmap_sem);
1330 }
1331 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332}
1333
1334/*
1335 * objrmap doesn't work for nonlinear VMAs because the assumption that
1336 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1337 * Consequently, given a particular page and its ->index, we cannot locate the
1338 * ptes which are mapping that page without an exhaustive linear search.
1339 *
1340 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1341 * maps the file to which the target page belongs. The ->vm_private_data field
1342 * holds the current cursor into that scan. Successive searches will circulate
1343 * around the vma's virtual address space.
1344 *
1345 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1346 * more scanning pressure is placed against them as well. Eventually pages
1347 * will become fully unmapped and are eligible for eviction.
1348 *
1349 * For very sparsely populated VMAs this is a little inefficient - chances are
1350 * there there won't be many ptes located within the scan cluster. In this case
1351 * maybe we could scan further - to the end of the pte page, perhaps.
Nick Pigginb291f002008-10-18 20:26:44 -07001352 *
1353 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can
1354 * acquire it without blocking. If vma locked, mlock the pages in the cluster,
1355 * rather than unmapping them. If we encounter the "check_page" that vmscan is
1356 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 */
1358#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE)
1359#define CLUSTER_MASK (~(CLUSTER_SIZE - 1))
1360
Nick Pigginb291f002008-10-18 20:26:44 -07001361static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1362 struct vm_area_struct *vma, struct page *check_page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363{
1364 struct mm_struct *mm = vma->vm_mm;
1365 pgd_t *pgd;
1366 pud_t *pud;
1367 pmd_t *pmd;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001368 pte_t *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 pte_t pteval;
Hugh Dickinsc0718802005-10-29 18:16:31 -07001370 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 struct page *page;
1372 unsigned long address;
1373 unsigned long end;
Nick Pigginb291f002008-10-18 20:26:44 -07001374 int ret = SWAP_AGAIN;
1375 int locked_vma = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 address = (vma->vm_start + cursor) & CLUSTER_MASK;
1378 end = address + CLUSTER_SIZE;
1379 if (address < vma->vm_start)
1380 address = vma->vm_start;
1381 if (end > vma->vm_end)
1382 end = vma->vm_end;
1383
1384 pgd = pgd_offset(mm, address);
1385 if (!pgd_present(*pgd))
Nick Pigginb291f002008-10-18 20:26:44 -07001386 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387
1388 pud = pud_offset(pgd, address);
1389 if (!pud_present(*pud))
Nick Pigginb291f002008-10-18 20:26:44 -07001390 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391
1392 pmd = pmd_offset(pud, address);
1393 if (!pmd_present(*pmd))
Nick Pigginb291f002008-10-18 20:26:44 -07001394 return ret;
1395
1396 /*
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001397 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
Nick Pigginb291f002008-10-18 20:26:44 -07001398 * keep the sem while scanning the cluster for mlocking pages.
1399 */
Hugh Dickinsaf8e3352009-12-14 17:58:59 -08001400 if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
Nick Pigginb291f002008-10-18 20:26:44 -07001401 locked_vma = (vma->vm_flags & VM_LOCKED);
1402 if (!locked_vma)
1403 up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1404 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001405
1406 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407
Hugh Dickins365e9c872005-10-29 18:16:18 -07001408 /* Update high watermark before we lower rss */
1409 update_hiwater_rss(mm);
1410
Hugh Dickinsc0718802005-10-29 18:16:31 -07001411 for (; address < end; pte++, address += PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 if (!pte_present(*pte))
1413 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -08001414 page = vm_normal_page(vma, address, *pte);
1415 BUG_ON(!page || PageAnon(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416
Nick Pigginb291f002008-10-18 20:26:44 -07001417 if (locked_vma) {
1418 mlock_vma_page(page); /* no-op if already mlocked */
1419 if (page == check_page)
1420 ret = SWAP_MLOCK;
1421 continue; /* don't unmap */
1422 }
1423
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001424 if (ptep_clear_flush_young_notify(vma, address, pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425 continue;
1426
1427 /* Nuke the page table entry. */
Ben Collinseca35132005-11-29 11:45:26 -08001428 flush_cache_page(vma, address, pte_pfn(*pte));
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001429 pteval = ptep_clear_flush_notify(vma, address, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430
1431 /* If nonlinear, store the file page offset in the pte. */
1432 if (page->index != linear_page_index(vma, address))
1433 set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1434
1435 /* Move the dirty bit to the physical page now the pte is gone. */
1436 if (pte_dirty(pteval))
1437 set_page_dirty(page);
1438
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001439 page_remove_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 page_cache_release(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001441 dec_mm_counter(mm, MM_FILEPAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442 (*mapcount)--;
1443 }
Hugh Dickinsc0718802005-10-29 18:16:31 -07001444 pte_unmap_unlock(pte - 1, ptl);
Nick Pigginb291f002008-10-18 20:26:44 -07001445 if (locked_vma)
1446 up_read(&vma->vm_mm->mmap_sem);
1447 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448}
1449
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001450bool is_vma_temporary_stack(struct vm_area_struct *vma)
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001451{
1452 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1453
1454 if (!maybe_stack)
1455 return false;
1456
1457 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1458 VM_STACK_INCOMPLETE_SETUP)
1459 return true;
1460
1461 return false;
1462}
1463
Nick Pigginb291f002008-10-18 20:26:44 -07001464/**
1465 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1466 * rmap method
1467 * @page: the page to unmap/unlock
Huang Shijie8051be52009-12-14 17:58:50 -08001468 * @flags: action and flags
Nick Pigginb291f002008-10-18 20:26:44 -07001469 *
1470 * Find all the mappings of a page using the mapping pointer and the vma chains
1471 * contained in the anon_vma struct it points to.
1472 *
1473 * This function is only called from try_to_unmap/try_to_munlock for
1474 * anonymous pages.
1475 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1476 * where the page was found will be held for write. So, we won't recheck
1477 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1478 * 'LOCKED.
1479 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001480static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481{
1482 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001483 struct anon_vma_chain *avc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484 int ret = SWAP_AGAIN;
Nick Pigginb291f002008-10-18 20:26:44 -07001485
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486 anon_vma = page_lock_anon_vma(page);
1487 if (!anon_vma)
1488 return ret;
1489
Rik van Riel5beb4932010-03-05 13:42:07 -08001490 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1491 struct vm_area_struct *vma = avc->vma;
Mel Gormana8bef8ff2010-05-24 14:32:24 -07001492 unsigned long address;
1493
1494 /*
1495 * During exec, a temporary VMA is setup and later moved.
1496 * The VMA is moved under the anon_vma lock but not the
1497 * page tables leading to a race where migration cannot
1498 * find the migration ptes. Rather than increasing the
1499 * locking requirements of exec(), migration skips
1500 * temporary VMAs until after exec() completes.
1501 */
1502 if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1503 is_vma_temporary_stack(vma))
1504 continue;
1505
1506 address = vma_address(page, vma);
Hugh Dickins1cb17292009-12-14 17:59:01 -08001507 if (address == -EFAULT)
1508 continue;
1509 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001510 if (ret != SWAP_AGAIN || !page_mapped(page))
1511 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512 }
Oleg Nesterov34bbd702007-02-28 20:13:49 -08001513
1514 page_unlock_anon_vma(anon_vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515 return ret;
1516}
1517
1518/**
Nick Pigginb291f002008-10-18 20:26:44 -07001519 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1520 * @page: the page to unmap/unlock
Andi Kleen14fa31b2009-09-16 11:50:10 +02001521 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522 *
1523 * Find all the mappings of a page using the mapping pointer and the vma chains
1524 * contained in the address_space struct it points to.
1525 *
Nick Pigginb291f002008-10-18 20:26:44 -07001526 * This function is only called from try_to_unmap/try_to_munlock for
1527 * object-based pages.
1528 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1529 * where the page was found will be held for write. So, we won't recheck
1530 * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1531 * 'LOCKED.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001533static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534{
1535 struct address_space *mapping = page->mapping;
1536 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1537 struct vm_area_struct *vma;
1538 struct prio_tree_iter iter;
1539 int ret = SWAP_AGAIN;
1540 unsigned long cursor;
1541 unsigned long max_nl_cursor = 0;
1542 unsigned long max_nl_size = 0;
1543 unsigned int mapcount;
1544
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001545 mutex_lock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
Hugh Dickins1cb17292009-12-14 17:59:01 -08001547 unsigned long address = vma_address(page, vma);
1548 if (address == -EFAULT)
1549 continue;
1550 ret = try_to_unmap_one(page, vma, address, flags);
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001551 if (ret != SWAP_AGAIN || !page_mapped(page))
1552 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553 }
1554
1555 if (list_empty(&mapping->i_mmap_nonlinear))
1556 goto out;
1557
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001558 /*
1559 * We don't bother to try to find the munlocked page in nonlinears.
1560 * It's costly. Instead, later, page reclaim logic may call
1561 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1562 */
1563 if (TTU_ACTION(flags) == TTU_MUNLOCK)
1564 goto out;
1565
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1567 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001568 cursor = (unsigned long) vma->vm_private_data;
1569 if (cursor > max_nl_cursor)
1570 max_nl_cursor = cursor;
1571 cursor = vma->vm_end - vma->vm_start;
1572 if (cursor > max_nl_size)
1573 max_nl_size = cursor;
1574 }
1575
Nick Pigginb291f002008-10-18 20:26:44 -07001576 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577 ret = SWAP_FAIL;
1578 goto out;
1579 }
1580
1581 /*
1582 * We don't try to search for this page in the nonlinear vmas,
1583 * and page_referenced wouldn't have found it anyway. Instead
1584 * just walk the nonlinear vmas trying to age and unmap some.
1585 * The mapcount of the page we came in with is irrelevant,
1586 * but even so use it as a guide to how hard we should try?
1587 */
1588 mapcount = page_mapcount(page);
1589 if (!mapcount)
1590 goto out;
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001591 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592
1593 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1594 if (max_nl_cursor == 0)
1595 max_nl_cursor = CLUSTER_SIZE;
1596
1597 do {
1598 list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1599 shared.vm_set.list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600 cursor = (unsigned long) vma->vm_private_data;
Hugh Dickins839b9682005-09-03 15:54:43 -07001601 while ( cursor < max_nl_cursor &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 cursor < vma->vm_end - vma->vm_start) {
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001603 if (try_to_unmap_cluster(cursor, &mapcount,
1604 vma, page) == SWAP_MLOCK)
1605 ret = SWAP_MLOCK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606 cursor += CLUSTER_SIZE;
1607 vma->vm_private_data = (void *) cursor;
1608 if ((int)mapcount <= 0)
1609 goto out;
1610 }
1611 vma->vm_private_data = (void *) max_nl_cursor;
1612 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001613 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614 max_nl_cursor += CLUSTER_SIZE;
1615 } while (max_nl_cursor <= max_nl_size);
1616
1617 /*
1618 * Don't loop forever (perhaps all the remaining pages are
1619 * in locked vmas). Reset cursor on all unreserved nonlinear
1620 * vmas, now forgetting on which ones it had fallen behind.
1621 */
Hugh Dickins101d2be2005-11-21 21:32:16 -08001622 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1623 vma->vm_private_data = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624out:
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001625 mutex_unlock(&mapping->i_mmap_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 return ret;
1627}
1628
1629/**
1630 * try_to_unmap - try to remove all page table mappings to a page
1631 * @page: the page to get unmapped
Andi Kleen14fa31b2009-09-16 11:50:10 +02001632 * @flags: action and flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633 *
1634 * Tries to remove all the page table entries which are mapping this
1635 * page, used in the pageout path. Caller must hold the page lock.
1636 * Return values are:
1637 *
1638 * SWAP_SUCCESS - we succeeded in removing all mappings
1639 * SWAP_AGAIN - we missed a mapping, try again later
1640 * SWAP_FAIL - the page is unswappable
Nick Pigginb291f002008-10-18 20:26:44 -07001641 * SWAP_MLOCK - page is mlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 */
Andi Kleen14fa31b2009-09-16 11:50:10 +02001643int try_to_unmap(struct page *page, enum ttu_flags flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644{
1645 int ret;
1646
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647 BUG_ON(!PageLocked(page));
Andrea Arcangeli91600e92011-01-13 15:47:16 -08001648 VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649
Hugh Dickins5ad64682009-12-14 17:59:24 -08001650 if (unlikely(PageKsm(page)))
1651 ret = try_to_unmap_ksm(page, flags);
1652 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001653 ret = try_to_unmap_anon(page, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001655 ret = try_to_unmap_file(page, flags);
Nick Pigginb291f002008-10-18 20:26:44 -07001656 if (ret != SWAP_MLOCK && !page_mapped(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657 ret = SWAP_SUCCESS;
1658 return ret;
1659}
Nikita Danilov81b40822005-05-01 08:58:36 -07001660
Nick Pigginb291f002008-10-18 20:26:44 -07001661/**
1662 * try_to_munlock - try to munlock a page
1663 * @page: the page to be munlocked
1664 *
1665 * Called from munlock code. Checks all of the VMAs mapping the page
1666 * to make sure nobody else has this page mlocked. The page will be
1667 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1668 *
1669 * Return values are:
1670 *
Hugh Dickins53f79ac2009-12-14 17:58:58 -08001671 * SWAP_AGAIN - no vma is holding page mlocked, or,
Nick Pigginb291f002008-10-18 20:26:44 -07001672 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem
Hugh Dickins5ad64682009-12-14 17:59:24 -08001673 * SWAP_FAIL - page cannot be located at present
Nick Pigginb291f002008-10-18 20:26:44 -07001674 * SWAP_MLOCK - page is now mlocked.
1675 */
1676int try_to_munlock(struct page *page)
1677{
1678 VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1679
Hugh Dickins5ad64682009-12-14 17:59:24 -08001680 if (unlikely(PageKsm(page)))
1681 return try_to_unmap_ksm(page, TTU_MUNLOCK);
1682 else if (PageAnon(page))
Andi Kleen14fa31b2009-09-16 11:50:10 +02001683 return try_to_unmap_anon(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001684 else
Andi Kleen14fa31b2009-09-16 11:50:10 +02001685 return try_to_unmap_file(page, TTU_MUNLOCK);
Nick Pigginb291f002008-10-18 20:26:44 -07001686}
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001687
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001688void __put_anon_vma(struct anon_vma *anon_vma)
Rik van Riel76545062010-08-09 17:18:41 -07001689{
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001690 struct anon_vma *root = anon_vma->root;
Rik van Riel76545062010-08-09 17:18:41 -07001691
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001692 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1693 anon_vma_free(root);
Rik van Riel76545062010-08-09 17:18:41 -07001694
Peter Zijlstra01d8b202011-03-22 16:32:49 -07001695 anon_vma_free(anon_vma);
Rik van Riel76545062010-08-09 17:18:41 -07001696}
Rik van Riel76545062010-08-09 17:18:41 -07001697
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001698#ifdef CONFIG_MIGRATION
1699/*
1700 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1701 * Called by migrate.c to remove migration ptes, but might be used more later.
1702 */
1703static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1704 struct vm_area_struct *, unsigned long, void *), void *arg)
1705{
1706 struct anon_vma *anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001707 struct anon_vma_chain *avc;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001708 int ret = SWAP_AGAIN;
1709
1710 /*
1711 * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1712 * because that depends on page_mapped(); but not all its usages
Mel Gorman3f6c8272010-05-24 14:32:17 -07001713 * are holding mmap_sem. Users without mmap_sem are required to
1714 * take a reference count to prevent the anon_vma disappearing
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001715 */
1716 anon_vma = page_anon_vma(page);
1717 if (!anon_vma)
1718 return ret;
Rik van Rielcba48b92010-08-09 17:18:38 -07001719 anon_vma_lock(anon_vma);
Rik van Riel5beb4932010-03-05 13:42:07 -08001720 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1721 struct vm_area_struct *vma = avc->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001722 unsigned long address = vma_address(page, vma);
1723 if (address == -EFAULT)
1724 continue;
1725 ret = rmap_one(page, vma, address, arg);
1726 if (ret != SWAP_AGAIN)
1727 break;
1728 }
Rik van Rielcba48b92010-08-09 17:18:38 -07001729 anon_vma_unlock(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001730 return ret;
1731}
1732
1733static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1734 struct vm_area_struct *, unsigned long, void *), void *arg)
1735{
1736 struct address_space *mapping = page->mapping;
1737 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1738 struct vm_area_struct *vma;
1739 struct prio_tree_iter iter;
1740 int ret = SWAP_AGAIN;
1741
1742 if (!mapping)
1743 return ret;
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001744 mutex_lock(&mapping->i_mmap_mutex);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001745 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1746 unsigned long address = vma_address(page, vma);
1747 if (address == -EFAULT)
1748 continue;
1749 ret = rmap_one(page, vma, address, arg);
1750 if (ret != SWAP_AGAIN)
1751 break;
1752 }
1753 /*
1754 * No nonlinear handling: being always shared, nonlinear vmas
1755 * never contain migration ptes. Decide what to do about this
1756 * limitation to linear when we need rmap_walk() on nonlinear.
1757 */
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07001758 mutex_unlock(&mapping->i_mmap_mutex);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001759 return ret;
1760}
1761
1762int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1763 struct vm_area_struct *, unsigned long, void *), void *arg)
1764{
1765 VM_BUG_ON(!PageLocked(page));
1766
1767 if (unlikely(PageKsm(page)))
1768 return rmap_walk_ksm(page, rmap_one, arg);
1769 else if (PageAnon(page))
1770 return rmap_walk_anon(page, rmap_one, arg);
1771 else
1772 return rmap_walk_file(page, rmap_one, arg);
1773}
1774#endif /* CONFIG_MIGRATION */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001775
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001776#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001777/*
1778 * The following three functions are for anonymous (private mapped) hugepages.
1779 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1780 * and no lru code, because we handle hugepages differently from common pages.
1781 */
1782static void __hugepage_set_anon_rmap(struct page *page,
1783 struct vm_area_struct *vma, unsigned long address, int exclusive)
1784{
1785 struct anon_vma *anon_vma = vma->anon_vma;
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001786
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001787 BUG_ON(!anon_vma);
Naoya Horiguchi433abed2010-09-10 13:23:03 +09001788
1789 if (PageAnon(page))
1790 return;
1791 if (!exclusive)
1792 anon_vma = anon_vma->root;
1793
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001794 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1795 page->mapping = (struct address_space *) anon_vma;
1796 page->index = linear_page_index(vma, address);
1797}
1798
1799void hugepage_add_anon_rmap(struct page *page,
1800 struct vm_area_struct *vma, unsigned long address)
1801{
1802 struct anon_vma *anon_vma = vma->anon_vma;
1803 int first;
Naoya Horiguchia850ea32010-09-10 13:23:06 +09001804
1805 BUG_ON(!PageLocked(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001806 BUG_ON(!anon_vma);
Hugh Dickins5dbe0af2011-05-28 13:17:04 -07001807 /* address might be in next vma when migration races vma_adjust */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001808 first = atomic_inc_and_test(&page->_mapcount);
1809 if (first)
1810 __hugepage_set_anon_rmap(page, vma, address, 0);
1811}
1812
1813void hugepage_add_new_anon_rmap(struct page *page,
1814 struct vm_area_struct *vma, unsigned long address)
1815{
1816 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1817 atomic_set(&page->_mapcount, 0);
1818 __hugepage_set_anon_rmap(page, vma, address, 1);
1819}
Naoya Horiguchie3390f62010-06-15 13:18:13 +09001820#endif /* CONFIG_HUGETLB_PAGE */