Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * mm/rmap.c - physical to virtual reverse mappings | |
3 | * | |
4 | * Copyright 2001, Rik van Riel <riel@conectiva.com.br> | |
5 | * Released under the General Public License (GPL). | |
6 | * | |
7 | * Simple, low overhead reverse mapping scheme. | |
8 | * Please try to keep this thing as modular as possible. | |
9 | * | |
10 | * Provides methods for unmapping each kind of mapped page: | |
11 | * the anon methods track anonymous pages, and | |
12 | * the file methods track pages belonging to an inode. | |
13 | * | |
14 | * Original design by Rik van Riel <riel@conectiva.com.br> 2001 | |
15 | * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004 | |
16 | * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004 | |
98f32602 | 17 | * Contributions by Hugh Dickins 2003, 2004 |
1da177e4 LT |
18 | */ |
19 | ||
20 | /* | |
21 | * Lock ordering in mm: | |
22 | * | |
1b1dcc1b | 23 | * inode->i_mutex (while writing or truncating, not reading or faulting) |
82591e6e NP |
24 | * inode->i_alloc_sem (vmtruncate_range) |
25 | * mm->mmap_sem | |
26 | * page->flags PG_locked (lock_page) | |
3d48ae45 | 27 | * mapping->i_mmap_mutex |
2b575eb6 | 28 | * anon_vma->mutex |
82591e6e NP |
29 | * mm->page_table_lock or pte_lock |
30 | * zone->lru_lock (in mark_page_accessed, isolate_lru_page) | |
31 | * swap_lock (in swap_duplicate, swap_info_get) | |
32 | * mmlist_lock (in mmput, drain_mmlist and others) | |
33 | * mapping->private_lock (in __set_page_dirty_buffers) | |
250df6ed | 34 | * inode->i_lock (in set_page_dirty's __mark_inode_dirty) |
a66979ab | 35 | * inode_wb_list_lock (in set_page_dirty's __mark_inode_dirty) |
82591e6e NP |
36 | * sb_lock (within inode_lock in fs/fs-writeback.c) |
37 | * mapping->tree_lock (widely used, in set_page_dirty, | |
38 | * in arch-dependent flush_dcache_mmap_lock, | |
a66979ab | 39 | * within inode_wb_list_lock in __sync_single_inode) |
6a46079c AK |
40 | * |
41 | * (code doesn't rely on that order so it could be switched around) | |
42 | * ->tasklist_lock | |
2b575eb6 | 43 | * anon_vma->mutex (memory_failure, collect_procs_anon) |
6a46079c | 44 | * pte map lock |
1da177e4 LT |
45 | */ |
46 | ||
47 | #include <linux/mm.h> | |
48 | #include <linux/pagemap.h> | |
49 | #include <linux/swap.h> | |
50 | #include <linux/swapops.h> | |
51 | #include <linux/slab.h> | |
52 | #include <linux/init.h> | |
5ad64688 | 53 | #include <linux/ksm.h> |
1da177e4 LT |
54 | #include <linux/rmap.h> |
55 | #include <linux/rcupdate.h> | |
a48d07af | 56 | #include <linux/module.h> |
8a9f3ccd | 57 | #include <linux/memcontrol.h> |
cddb8a5c | 58 | #include <linux/mmu_notifier.h> |
64cdd548 | 59 | #include <linux/migrate.h> |
0fe6e20b | 60 | #include <linux/hugetlb.h> |
1da177e4 LT |
61 | |
62 | #include <asm/tlbflush.h> | |
63 | ||
b291f000 NP |
64 | #include "internal.h" |
65 | ||
fdd2e5f8 | 66 | static struct kmem_cache *anon_vma_cachep; |
5beb4930 | 67 | static struct kmem_cache *anon_vma_chain_cachep; |
fdd2e5f8 AB |
68 | |
69 | static inline struct anon_vma *anon_vma_alloc(void) | |
70 | { | |
01d8b20d PZ |
71 | struct anon_vma *anon_vma; |
72 | ||
73 | anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); | |
74 | if (anon_vma) { | |
75 | atomic_set(&anon_vma->refcount, 1); | |
76 | /* | |
77 | * Initialise the anon_vma root to point to itself. If called | |
78 | * from fork, the root will be reset to the parents anon_vma. | |
79 | */ | |
80 | anon_vma->root = anon_vma; | |
81 | } | |
82 | ||
83 | return anon_vma; | |
fdd2e5f8 AB |
84 | } |
85 | ||
01d8b20d | 86 | static inline void anon_vma_free(struct anon_vma *anon_vma) |
fdd2e5f8 | 87 | { |
01d8b20d | 88 | VM_BUG_ON(atomic_read(&anon_vma->refcount)); |
88c22088 PZ |
89 | |
90 | /* | |
91 | * Synchronize against page_lock_anon_vma() such that | |
92 | * we can safely hold the lock without the anon_vma getting | |
93 | * freed. | |
94 | * | |
95 | * Relies on the full mb implied by the atomic_dec_and_test() from | |
96 | * put_anon_vma() against the acquire barrier implied by | |
97 | * mutex_trylock() from page_lock_anon_vma(). This orders: | |
98 | * | |
99 | * page_lock_anon_vma() VS put_anon_vma() | |
100 | * mutex_trylock() atomic_dec_and_test() | |
101 | * LOCK MB | |
102 | * atomic_read() mutex_is_locked() | |
103 | * | |
104 | * LOCK should suffice since the actual taking of the lock must | |
105 | * happen _before_ what follows. | |
106 | */ | |
107 | if (mutex_is_locked(&anon_vma->root->mutex)) { | |
108 | anon_vma_lock(anon_vma); | |
109 | anon_vma_unlock(anon_vma); | |
110 | } | |
111 | ||
fdd2e5f8 AB |
112 | kmem_cache_free(anon_vma_cachep, anon_vma); |
113 | } | |
1da177e4 | 114 | |
5beb4930 RR |
115 | static inline struct anon_vma_chain *anon_vma_chain_alloc(void) |
116 | { | |
117 | return kmem_cache_alloc(anon_vma_chain_cachep, GFP_KERNEL); | |
118 | } | |
119 | ||
e574b5fd | 120 | static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) |
5beb4930 RR |
121 | { |
122 | kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); | |
123 | } | |
124 | ||
d9d332e0 LT |
125 | /** |
126 | * anon_vma_prepare - attach an anon_vma to a memory region | |
127 | * @vma: the memory region in question | |
128 | * | |
129 | * This makes sure the memory mapping described by 'vma' has | |
130 | * an 'anon_vma' attached to it, so that we can associate the | |
131 | * anonymous pages mapped into it with that anon_vma. | |
132 | * | |
133 | * The common case will be that we already have one, but if | |
23a0790a | 134 | * not we either need to find an adjacent mapping that we |
d9d332e0 LT |
135 | * can re-use the anon_vma from (very common when the only |
136 | * reason for splitting a vma has been mprotect()), or we | |
137 | * allocate a new one. | |
138 | * | |
139 | * Anon-vma allocations are very subtle, because we may have | |
140 | * optimistically looked up an anon_vma in page_lock_anon_vma() | |
141 | * and that may actually touch the spinlock even in the newly | |
142 | * allocated vma (it depends on RCU to make sure that the | |
143 | * anon_vma isn't actually destroyed). | |
144 | * | |
145 | * As a result, we need to do proper anon_vma locking even | |
146 | * for the new allocation. At the same time, we do not want | |
147 | * to do any locking for the common case of already having | |
148 | * an anon_vma. | |
149 | * | |
150 | * This must be called with the mmap_sem held for reading. | |
151 | */ | |
1da177e4 LT |
152 | int anon_vma_prepare(struct vm_area_struct *vma) |
153 | { | |
154 | struct anon_vma *anon_vma = vma->anon_vma; | |
5beb4930 | 155 | struct anon_vma_chain *avc; |
1da177e4 LT |
156 | |
157 | might_sleep(); | |
158 | if (unlikely(!anon_vma)) { | |
159 | struct mm_struct *mm = vma->vm_mm; | |
d9d332e0 | 160 | struct anon_vma *allocated; |
1da177e4 | 161 | |
5beb4930 RR |
162 | avc = anon_vma_chain_alloc(); |
163 | if (!avc) | |
164 | goto out_enomem; | |
165 | ||
1da177e4 | 166 | anon_vma = find_mergeable_anon_vma(vma); |
d9d332e0 LT |
167 | allocated = NULL; |
168 | if (!anon_vma) { | |
1da177e4 LT |
169 | anon_vma = anon_vma_alloc(); |
170 | if (unlikely(!anon_vma)) | |
5beb4930 | 171 | goto out_enomem_free_avc; |
1da177e4 | 172 | allocated = anon_vma; |
1da177e4 LT |
173 | } |
174 | ||
cba48b98 | 175 | anon_vma_lock(anon_vma); |
1da177e4 LT |
176 | /* page_table_lock to protect against threads */ |
177 | spin_lock(&mm->page_table_lock); | |
178 | if (likely(!vma->anon_vma)) { | |
179 | vma->anon_vma = anon_vma; | |
5beb4930 RR |
180 | avc->anon_vma = anon_vma; |
181 | avc->vma = vma; | |
182 | list_add(&avc->same_vma, &vma->anon_vma_chain); | |
26ba0cb6 | 183 | list_add_tail(&avc->same_anon_vma, &anon_vma->head); |
1da177e4 | 184 | allocated = NULL; |
31f2b0eb | 185 | avc = NULL; |
1da177e4 LT |
186 | } |
187 | spin_unlock(&mm->page_table_lock); | |
cba48b98 | 188 | anon_vma_unlock(anon_vma); |
31f2b0eb ON |
189 | |
190 | if (unlikely(allocated)) | |
01d8b20d | 191 | put_anon_vma(allocated); |
31f2b0eb | 192 | if (unlikely(avc)) |
5beb4930 | 193 | anon_vma_chain_free(avc); |
1da177e4 LT |
194 | } |
195 | return 0; | |
5beb4930 RR |
196 | |
197 | out_enomem_free_avc: | |
198 | anon_vma_chain_free(avc); | |
199 | out_enomem: | |
200 | return -ENOMEM; | |
1da177e4 LT |
201 | } |
202 | ||
bb4aa396 LT |
203 | /* |
204 | * This is a useful helper function for locking the anon_vma root as | |
205 | * we traverse the vma->anon_vma_chain, looping over anon_vma's that | |
206 | * have the same vma. | |
207 | * | |
208 | * Such anon_vma's should have the same root, so you'd expect to see | |
209 | * just a single mutex_lock for the whole traversal. | |
210 | */ | |
211 | static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma) | |
212 | { | |
213 | struct anon_vma *new_root = anon_vma->root; | |
214 | if (new_root != root) { | |
215 | if (WARN_ON_ONCE(root)) | |
216 | mutex_unlock(&root->mutex); | |
217 | root = new_root; | |
218 | mutex_lock(&root->mutex); | |
219 | } | |
220 | return root; | |
221 | } | |
222 | ||
223 | static inline void unlock_anon_vma_root(struct anon_vma *root) | |
224 | { | |
225 | if (root) | |
226 | mutex_unlock(&root->mutex); | |
227 | } | |
228 | ||
5beb4930 RR |
229 | static void anon_vma_chain_link(struct vm_area_struct *vma, |
230 | struct anon_vma_chain *avc, | |
231 | struct anon_vma *anon_vma) | |
1da177e4 | 232 | { |
5beb4930 RR |
233 | avc->vma = vma; |
234 | avc->anon_vma = anon_vma; | |
235 | list_add(&avc->same_vma, &vma->anon_vma_chain); | |
236 | ||
05759d38 AA |
237 | /* |
238 | * It's critical to add new vmas to the tail of the anon_vma, | |
239 | * see comment in huge_memory.c:__split_huge_page(). | |
240 | */ | |
5beb4930 | 241 | list_add_tail(&avc->same_anon_vma, &anon_vma->head); |
1da177e4 LT |
242 | } |
243 | ||
5beb4930 RR |
244 | /* |
245 | * Attach the anon_vmas from src to dst. | |
246 | * Returns 0 on success, -ENOMEM on failure. | |
247 | */ | |
248 | int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) | |
1da177e4 | 249 | { |
5beb4930 | 250 | struct anon_vma_chain *avc, *pavc; |
bb4aa396 | 251 | struct anon_vma *root = NULL; |
5beb4930 | 252 | |
646d87b4 | 253 | list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { |
bb4aa396 LT |
254 | struct anon_vma *anon_vma; |
255 | ||
5beb4930 RR |
256 | avc = anon_vma_chain_alloc(); |
257 | if (!avc) | |
258 | goto enomem_failure; | |
bb4aa396 LT |
259 | anon_vma = pavc->anon_vma; |
260 | root = lock_anon_vma_root(root, anon_vma); | |
261 | anon_vma_chain_link(dst, avc, anon_vma); | |
5beb4930 | 262 | } |
bb4aa396 | 263 | unlock_anon_vma_root(root); |
5beb4930 | 264 | return 0; |
1da177e4 | 265 | |
5beb4930 | 266 | enomem_failure: |
bb4aa396 | 267 | unlock_anon_vma_root(root); |
5beb4930 RR |
268 | unlink_anon_vmas(dst); |
269 | return -ENOMEM; | |
1da177e4 LT |
270 | } |
271 | ||
5beb4930 RR |
272 | /* |
273 | * Attach vma to its own anon_vma, as well as to the anon_vmas that | |
274 | * the corresponding VMA in the parent process is attached to. | |
275 | * Returns 0 on success, non-zero on failure. | |
276 | */ | |
277 | int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) | |
1da177e4 | 278 | { |
5beb4930 RR |
279 | struct anon_vma_chain *avc; |
280 | struct anon_vma *anon_vma; | |
1da177e4 | 281 | |
5beb4930 RR |
282 | /* Don't bother if the parent process has no anon_vma here. */ |
283 | if (!pvma->anon_vma) | |
284 | return 0; | |
285 | ||
286 | /* | |
287 | * First, attach the new VMA to the parent VMA's anon_vmas, | |
288 | * so rmap can find non-COWed pages in child processes. | |
289 | */ | |
290 | if (anon_vma_clone(vma, pvma)) | |
291 | return -ENOMEM; | |
292 | ||
293 | /* Then add our own anon_vma. */ | |
294 | anon_vma = anon_vma_alloc(); | |
295 | if (!anon_vma) | |
296 | goto out_error; | |
297 | avc = anon_vma_chain_alloc(); | |
298 | if (!avc) | |
299 | goto out_error_free_anon_vma; | |
5c341ee1 RR |
300 | |
301 | /* | |
302 | * The root anon_vma's spinlock is the lock actually used when we | |
303 | * lock any of the anon_vmas in this anon_vma tree. | |
304 | */ | |
305 | anon_vma->root = pvma->anon_vma->root; | |
76545066 | 306 | /* |
01d8b20d PZ |
307 | * With refcounts, an anon_vma can stay around longer than the |
308 | * process it belongs to. The root anon_vma needs to be pinned until | |
309 | * this anon_vma is freed, because the lock lives in the root. | |
76545066 RR |
310 | */ |
311 | get_anon_vma(anon_vma->root); | |
5beb4930 RR |
312 | /* Mark this anon_vma as the one where our new (COWed) pages go. */ |
313 | vma->anon_vma = anon_vma; | |
bb4aa396 | 314 | anon_vma_lock(anon_vma); |
5c341ee1 | 315 | anon_vma_chain_link(vma, avc, anon_vma); |
bb4aa396 | 316 | anon_vma_unlock(anon_vma); |
5beb4930 RR |
317 | |
318 | return 0; | |
319 | ||
320 | out_error_free_anon_vma: | |
01d8b20d | 321 | put_anon_vma(anon_vma); |
5beb4930 | 322 | out_error: |
4946d54c | 323 | unlink_anon_vmas(vma); |
5beb4930 | 324 | return -ENOMEM; |
1da177e4 LT |
325 | } |
326 | ||
5beb4930 | 327 | static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain) |
1da177e4 | 328 | { |
5beb4930 | 329 | struct anon_vma *anon_vma = anon_vma_chain->anon_vma; |
1da177e4 LT |
330 | int empty; |
331 | ||
5beb4930 | 332 | /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */ |
1da177e4 LT |
333 | if (!anon_vma) |
334 | return; | |
335 | ||
cba48b98 | 336 | anon_vma_lock(anon_vma); |
5beb4930 | 337 | list_del(&anon_vma_chain->same_anon_vma); |
1da177e4 LT |
338 | |
339 | /* We must garbage collect the anon_vma if it's empty */ | |
01d8b20d | 340 | empty = list_empty(&anon_vma->head); |
cba48b98 | 341 | anon_vma_unlock(anon_vma); |
1da177e4 | 342 | |
01d8b20d PZ |
343 | if (empty) |
344 | put_anon_vma(anon_vma); | |
1da177e4 LT |
345 | } |
346 | ||
5beb4930 RR |
347 | void unlink_anon_vmas(struct vm_area_struct *vma) |
348 | { | |
349 | struct anon_vma_chain *avc, *next; | |
350 | ||
5c341ee1 RR |
351 | /* |
352 | * Unlink each anon_vma chained to the VMA. This list is ordered | |
353 | * from newest to oldest, ensuring the root anon_vma gets freed last. | |
354 | */ | |
5beb4930 RR |
355 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { |
356 | anon_vma_unlink(avc); | |
357 | list_del(&avc->same_vma); | |
358 | anon_vma_chain_free(avc); | |
359 | } | |
360 | } | |
361 | ||
51cc5068 | 362 | static void anon_vma_ctor(void *data) |
1da177e4 | 363 | { |
a35afb83 | 364 | struct anon_vma *anon_vma = data; |
1da177e4 | 365 | |
2b575eb6 | 366 | mutex_init(&anon_vma->mutex); |
83813267 | 367 | atomic_set(&anon_vma->refcount, 0); |
a35afb83 | 368 | INIT_LIST_HEAD(&anon_vma->head); |
1da177e4 LT |
369 | } |
370 | ||
371 | void __init anon_vma_init(void) | |
372 | { | |
373 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
20c2df83 | 374 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor); |
5beb4930 | 375 | anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC); |
1da177e4 LT |
376 | } |
377 | ||
378 | /* | |
6111e4ca PZ |
379 | * Getting a lock on a stable anon_vma from a page off the LRU is tricky! |
380 | * | |
381 | * Since there is no serialization what so ever against page_remove_rmap() | |
382 | * the best this function can do is return a locked anon_vma that might | |
383 | * have been relevant to this page. | |
384 | * | |
385 | * The page might have been remapped to a different anon_vma or the anon_vma | |
386 | * returned may already be freed (and even reused). | |
387 | * | |
bc658c96 PZ |
388 | * In case it was remapped to a different anon_vma, the new anon_vma will be a |
389 | * child of the old anon_vma, and the anon_vma lifetime rules will therefore | |
390 | * ensure that any anon_vma obtained from the page will still be valid for as | |
391 | * long as we observe page_mapped() [ hence all those page_mapped() tests ]. | |
392 | * | |
6111e4ca PZ |
393 | * All users of this function must be very careful when walking the anon_vma |
394 | * chain and verify that the page in question is indeed mapped in it | |
395 | * [ something equivalent to page_mapped_in_vma() ]. | |
396 | * | |
397 | * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap() | |
398 | * that the anon_vma pointer from page->mapping is valid if there is a | |
399 | * mapcount, we can dereference the anon_vma after observing those. | |
1da177e4 | 400 | */ |
746b18d4 | 401 | struct anon_vma *page_get_anon_vma(struct page *page) |
1da177e4 | 402 | { |
746b18d4 | 403 | struct anon_vma *anon_vma = NULL; |
1da177e4 LT |
404 | unsigned long anon_mapping; |
405 | ||
406 | rcu_read_lock(); | |
80e14822 | 407 | anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping); |
3ca7b3c5 | 408 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) |
1da177e4 LT |
409 | goto out; |
410 | if (!page_mapped(page)) | |
411 | goto out; | |
412 | ||
413 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
746b18d4 PZ |
414 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { |
415 | anon_vma = NULL; | |
416 | goto out; | |
417 | } | |
f1819427 HD |
418 | |
419 | /* | |
420 | * If this page is still mapped, then its anon_vma cannot have been | |
746b18d4 PZ |
421 | * freed. But if it has been unmapped, we have no security against the |
422 | * anon_vma structure being freed and reused (for another anon_vma: | |
423 | * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero() | |
424 | * above cannot corrupt). | |
f1819427 | 425 | */ |
746b18d4 PZ |
426 | if (!page_mapped(page)) { |
427 | put_anon_vma(anon_vma); | |
428 | anon_vma = NULL; | |
429 | } | |
1da177e4 LT |
430 | out: |
431 | rcu_read_unlock(); | |
746b18d4 PZ |
432 | |
433 | return anon_vma; | |
434 | } | |
435 | ||
88c22088 PZ |
436 | /* |
437 | * Similar to page_get_anon_vma() except it locks the anon_vma. | |
438 | * | |
439 | * Its a little more complex as it tries to keep the fast path to a single | |
440 | * atomic op -- the trylock. If we fail the trylock, we fall back to getting a | |
441 | * reference like with page_get_anon_vma() and then block on the mutex. | |
442 | */ | |
746b18d4 PZ |
443 | struct anon_vma *page_lock_anon_vma(struct page *page) |
444 | { | |
88c22088 | 445 | struct anon_vma *anon_vma = NULL; |
eee0f252 | 446 | struct anon_vma *root_anon_vma; |
88c22088 | 447 | unsigned long anon_mapping; |
746b18d4 | 448 | |
88c22088 PZ |
449 | rcu_read_lock(); |
450 | anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping); | |
451 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) | |
452 | goto out; | |
453 | if (!page_mapped(page)) | |
454 | goto out; | |
455 | ||
456 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
eee0f252 HD |
457 | root_anon_vma = ACCESS_ONCE(anon_vma->root); |
458 | if (mutex_trylock(&root_anon_vma->mutex)) { | |
88c22088 | 459 | /* |
eee0f252 HD |
460 | * If the page is still mapped, then this anon_vma is still |
461 | * its anon_vma, and holding the mutex ensures that it will | |
bc658c96 | 462 | * not go away, see anon_vma_free(). |
88c22088 | 463 | */ |
eee0f252 HD |
464 | if (!page_mapped(page)) { |
465 | mutex_unlock(&root_anon_vma->mutex); | |
88c22088 PZ |
466 | anon_vma = NULL; |
467 | } | |
468 | goto out; | |
469 | } | |
746b18d4 | 470 | |
88c22088 PZ |
471 | /* trylock failed, we got to sleep */ |
472 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { | |
473 | anon_vma = NULL; | |
474 | goto out; | |
475 | } | |
476 | ||
477 | if (!page_mapped(page)) { | |
478 | put_anon_vma(anon_vma); | |
479 | anon_vma = NULL; | |
480 | goto out; | |
481 | } | |
482 | ||
483 | /* we pinned the anon_vma, its safe to sleep */ | |
484 | rcu_read_unlock(); | |
485 | anon_vma_lock(anon_vma); | |
486 | ||
487 | if (atomic_dec_and_test(&anon_vma->refcount)) { | |
488 | /* | |
489 | * Oops, we held the last refcount, release the lock | |
490 | * and bail -- can't simply use put_anon_vma() because | |
491 | * we'll deadlock on the anon_vma_lock() recursion. | |
492 | */ | |
493 | anon_vma_unlock(anon_vma); | |
494 | __put_anon_vma(anon_vma); | |
495 | anon_vma = NULL; | |
496 | } | |
497 | ||
498 | return anon_vma; | |
499 | ||
500 | out: | |
501 | rcu_read_unlock(); | |
746b18d4 | 502 | return anon_vma; |
34bbd704 ON |
503 | } |
504 | ||
10be22df | 505 | void page_unlock_anon_vma(struct anon_vma *anon_vma) |
34bbd704 | 506 | { |
cba48b98 | 507 | anon_vma_unlock(anon_vma); |
1da177e4 LT |
508 | } |
509 | ||
510 | /* | |
3ad33b24 LS |
511 | * At what user virtual address is page expected in @vma? |
512 | * Returns virtual address or -EFAULT if page's index/offset is not | |
513 | * within the range mapped the @vma. | |
1da177e4 | 514 | */ |
71e3aac0 | 515 | inline unsigned long |
1da177e4 LT |
516 | vma_address(struct page *page, struct vm_area_struct *vma) |
517 | { | |
518 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
519 | unsigned long address; | |
520 | ||
0fe6e20b NH |
521 | if (unlikely(is_vm_hugetlb_page(vma))) |
522 | pgoff = page->index << huge_page_order(page_hstate(page)); | |
1da177e4 LT |
523 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); |
524 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { | |
3ad33b24 | 525 | /* page should be within @vma mapping range */ |
1da177e4 LT |
526 | return -EFAULT; |
527 | } | |
528 | return address; | |
529 | } | |
530 | ||
531 | /* | |
bf89c8c8 | 532 | * At what user virtual address is page expected in vma? |
ab941e0f | 533 | * Caller should check the page is actually part of the vma. |
1da177e4 LT |
534 | */ |
535 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
536 | { | |
21d0d443 | 537 | if (PageAnon(page)) { |
4829b906 HD |
538 | struct anon_vma *page__anon_vma = page_anon_vma(page); |
539 | /* | |
540 | * Note: swapoff's unuse_vma() is more efficient with this | |
541 | * check, and needs it to match anon_vma when KSM is active. | |
542 | */ | |
543 | if (!vma->anon_vma || !page__anon_vma || | |
544 | vma->anon_vma->root != page__anon_vma->root) | |
21d0d443 AA |
545 | return -EFAULT; |
546 | } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { | |
ee498ed7 HD |
547 | if (!vma->vm_file || |
548 | vma->vm_file->f_mapping != page->mapping) | |
1da177e4 LT |
549 | return -EFAULT; |
550 | } else | |
551 | return -EFAULT; | |
552 | return vma_address(page, vma); | |
553 | } | |
554 | ||
81b4082d ND |
555 | /* |
556 | * Check that @page is mapped at @address into @mm. | |
557 | * | |
479db0bf NP |
558 | * If @sync is false, page_check_address may perform a racy check to avoid |
559 | * the page table lock when the pte is not present (helpful when reclaiming | |
560 | * highly shared pages). | |
561 | * | |
b8072f09 | 562 | * On success returns with pte mapped and locked. |
81b4082d | 563 | */ |
e9a81a82 | 564 | pte_t *__page_check_address(struct page *page, struct mm_struct *mm, |
479db0bf | 565 | unsigned long address, spinlock_t **ptlp, int sync) |
81b4082d ND |
566 | { |
567 | pgd_t *pgd; | |
568 | pud_t *pud; | |
569 | pmd_t *pmd; | |
570 | pte_t *pte; | |
c0718806 | 571 | spinlock_t *ptl; |
81b4082d | 572 | |
0fe6e20b NH |
573 | if (unlikely(PageHuge(page))) { |
574 | pte = huge_pte_offset(mm, address); | |
575 | ptl = &mm->page_table_lock; | |
576 | goto check; | |
577 | } | |
578 | ||
81b4082d | 579 | pgd = pgd_offset(mm, address); |
c0718806 HD |
580 | if (!pgd_present(*pgd)) |
581 | return NULL; | |
582 | ||
583 | pud = pud_offset(pgd, address); | |
584 | if (!pud_present(*pud)) | |
585 | return NULL; | |
586 | ||
587 | pmd = pmd_offset(pud, address); | |
588 | if (!pmd_present(*pmd)) | |
589 | return NULL; | |
71e3aac0 AA |
590 | if (pmd_trans_huge(*pmd)) |
591 | return NULL; | |
c0718806 HD |
592 | |
593 | pte = pte_offset_map(pmd, address); | |
594 | /* Make a quick check before getting the lock */ | |
479db0bf | 595 | if (!sync && !pte_present(*pte)) { |
c0718806 HD |
596 | pte_unmap(pte); |
597 | return NULL; | |
598 | } | |
599 | ||
4c21e2f2 | 600 | ptl = pte_lockptr(mm, pmd); |
0fe6e20b | 601 | check: |
c0718806 HD |
602 | spin_lock(ptl); |
603 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { | |
604 | *ptlp = ptl; | |
605 | return pte; | |
81b4082d | 606 | } |
c0718806 HD |
607 | pte_unmap_unlock(pte, ptl); |
608 | return NULL; | |
81b4082d ND |
609 | } |
610 | ||
b291f000 NP |
611 | /** |
612 | * page_mapped_in_vma - check whether a page is really mapped in a VMA | |
613 | * @page: the page to test | |
614 | * @vma: the VMA to test | |
615 | * | |
616 | * Returns 1 if the page is mapped into the page tables of the VMA, 0 | |
617 | * if the page is not mapped into the page tables of this VMA. Only | |
618 | * valid for normal file or anonymous VMAs. | |
619 | */ | |
6a46079c | 620 | int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma) |
b291f000 NP |
621 | { |
622 | unsigned long address; | |
623 | pte_t *pte; | |
624 | spinlock_t *ptl; | |
625 | ||
626 | address = vma_address(page, vma); | |
627 | if (address == -EFAULT) /* out of vma range */ | |
628 | return 0; | |
629 | pte = page_check_address(page, vma->vm_mm, address, &ptl, 1); | |
630 | if (!pte) /* the page is not in this mm */ | |
631 | return 0; | |
632 | pte_unmap_unlock(pte, ptl); | |
633 | ||
634 | return 1; | |
635 | } | |
636 | ||
1da177e4 LT |
637 | /* |
638 | * Subfunctions of page_referenced: page_referenced_one called | |
639 | * repeatedly from either page_referenced_anon or page_referenced_file. | |
640 | */ | |
5ad64688 HD |
641 | int page_referenced_one(struct page *page, struct vm_area_struct *vma, |
642 | unsigned long address, unsigned int *mapcount, | |
643 | unsigned long *vm_flags) | |
1da177e4 LT |
644 | { |
645 | struct mm_struct *mm = vma->vm_mm; | |
1da177e4 LT |
646 | int referenced = 0; |
647 | ||
71e3aac0 AA |
648 | if (unlikely(PageTransHuge(page))) { |
649 | pmd_t *pmd; | |
650 | ||
651 | spin_lock(&mm->page_table_lock); | |
2da28bfd AA |
652 | /* |
653 | * rmap might return false positives; we must filter | |
654 | * these out using page_check_address_pmd(). | |
655 | */ | |
71e3aac0 AA |
656 | pmd = page_check_address_pmd(page, mm, address, |
657 | PAGE_CHECK_ADDRESS_PMD_FLAG); | |
2da28bfd AA |
658 | if (!pmd) { |
659 | spin_unlock(&mm->page_table_lock); | |
660 | goto out; | |
661 | } | |
662 | ||
663 | if (vma->vm_flags & VM_LOCKED) { | |
664 | spin_unlock(&mm->page_table_lock); | |
665 | *mapcount = 0; /* break early from loop */ | |
666 | *vm_flags |= VM_LOCKED; | |
667 | goto out; | |
668 | } | |
669 | ||
670 | /* go ahead even if the pmd is pmd_trans_splitting() */ | |
671 | if (pmdp_clear_flush_young_notify(vma, address, pmd)) | |
71e3aac0 AA |
672 | referenced++; |
673 | spin_unlock(&mm->page_table_lock); | |
674 | } else { | |
675 | pte_t *pte; | |
676 | spinlock_t *ptl; | |
677 | ||
2da28bfd AA |
678 | /* |
679 | * rmap might return false positives; we must filter | |
680 | * these out using page_check_address(). | |
681 | */ | |
71e3aac0 AA |
682 | pte = page_check_address(page, mm, address, &ptl, 0); |
683 | if (!pte) | |
684 | goto out; | |
685 | ||
2da28bfd AA |
686 | if (vma->vm_flags & VM_LOCKED) { |
687 | pte_unmap_unlock(pte, ptl); | |
688 | *mapcount = 0; /* break early from loop */ | |
689 | *vm_flags |= VM_LOCKED; | |
690 | goto out; | |
691 | } | |
692 | ||
71e3aac0 AA |
693 | if (ptep_clear_flush_young_notify(vma, address, pte)) { |
694 | /* | |
695 | * Don't treat a reference through a sequentially read | |
696 | * mapping as such. If the page has been used in | |
697 | * another mapping, we will catch it; if this other | |
698 | * mapping is already gone, the unmap path will have | |
699 | * set PG_referenced or activated the page. | |
700 | */ | |
701 | if (likely(!VM_SequentialReadHint(vma))) | |
702 | referenced++; | |
703 | } | |
704 | pte_unmap_unlock(pte, ptl); | |
705 | } | |
706 | ||
2da28bfd AA |
707 | /* Pretend the page is referenced if the task has the |
708 | swap token and is in the middle of a page fault. */ | |
709 | if (mm != current->mm && has_swap_token(mm) && | |
710 | rwsem_is_locked(&mm->mmap_sem)) | |
711 | referenced++; | |
712 | ||
c0718806 | 713 | (*mapcount)--; |
273f047e | 714 | |
6fe6b7e3 WF |
715 | if (referenced) |
716 | *vm_flags |= vma->vm_flags; | |
273f047e | 717 | out: |
1da177e4 LT |
718 | return referenced; |
719 | } | |
720 | ||
bed7161a | 721 | static int page_referenced_anon(struct page *page, |
6fe6b7e3 WF |
722 | struct mem_cgroup *mem_cont, |
723 | unsigned long *vm_flags) | |
1da177e4 LT |
724 | { |
725 | unsigned int mapcount; | |
726 | struct anon_vma *anon_vma; | |
5beb4930 | 727 | struct anon_vma_chain *avc; |
1da177e4 LT |
728 | int referenced = 0; |
729 | ||
730 | anon_vma = page_lock_anon_vma(page); | |
731 | if (!anon_vma) | |
732 | return referenced; | |
733 | ||
734 | mapcount = page_mapcount(page); | |
5beb4930 RR |
735 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { |
736 | struct vm_area_struct *vma = avc->vma; | |
1cb1729b HD |
737 | unsigned long address = vma_address(page, vma); |
738 | if (address == -EFAULT) | |
739 | continue; | |
bed7161a BS |
740 | /* |
741 | * If we are reclaiming on behalf of a cgroup, skip | |
742 | * counting on behalf of references from different | |
743 | * cgroups | |
744 | */ | |
bd845e38 | 745 | if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont)) |
bed7161a | 746 | continue; |
1cb1729b | 747 | referenced += page_referenced_one(page, vma, address, |
6fe6b7e3 | 748 | &mapcount, vm_flags); |
1da177e4 LT |
749 | if (!mapcount) |
750 | break; | |
751 | } | |
34bbd704 ON |
752 | |
753 | page_unlock_anon_vma(anon_vma); | |
1da177e4 LT |
754 | return referenced; |
755 | } | |
756 | ||
757 | /** | |
758 | * page_referenced_file - referenced check for object-based rmap | |
759 | * @page: the page we're checking references on. | |
43d8eac4 | 760 | * @mem_cont: target memory controller |
6fe6b7e3 | 761 | * @vm_flags: collect encountered vma->vm_flags who actually referenced the page |
1da177e4 LT |
762 | * |
763 | * For an object-based mapped page, find all the places it is mapped and | |
764 | * check/clear the referenced flag. This is done by following the page->mapping | |
765 | * pointer, then walking the chain of vmas it holds. It returns the number | |
766 | * of references it found. | |
767 | * | |
768 | * This function is only called from page_referenced for object-based pages. | |
769 | */ | |
bed7161a | 770 | static int page_referenced_file(struct page *page, |
6fe6b7e3 WF |
771 | struct mem_cgroup *mem_cont, |
772 | unsigned long *vm_flags) | |
1da177e4 LT |
773 | { |
774 | unsigned int mapcount; | |
775 | struct address_space *mapping = page->mapping; | |
776 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
777 | struct vm_area_struct *vma; | |
778 | struct prio_tree_iter iter; | |
779 | int referenced = 0; | |
780 | ||
781 | /* | |
782 | * The caller's checks on page->mapping and !PageAnon have made | |
783 | * sure that this is a file page: the check for page->mapping | |
784 | * excludes the case just before it gets set on an anon page. | |
785 | */ | |
786 | BUG_ON(PageAnon(page)); | |
787 | ||
788 | /* | |
789 | * The page lock not only makes sure that page->mapping cannot | |
790 | * suddenly be NULLified by truncation, it makes sure that the | |
791 | * structure at mapping cannot be freed and reused yet, | |
3d48ae45 | 792 | * so we can safely take mapping->i_mmap_mutex. |
1da177e4 LT |
793 | */ |
794 | BUG_ON(!PageLocked(page)); | |
795 | ||
3d48ae45 | 796 | mutex_lock(&mapping->i_mmap_mutex); |
1da177e4 LT |
797 | |
798 | /* | |
3d48ae45 | 799 | * i_mmap_mutex does not stabilize mapcount at all, but mapcount |
1da177e4 LT |
800 | * is more likely to be accurate if we note it after spinning. |
801 | */ | |
802 | mapcount = page_mapcount(page); | |
803 | ||
804 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
1cb1729b HD |
805 | unsigned long address = vma_address(page, vma); |
806 | if (address == -EFAULT) | |
807 | continue; | |
bed7161a BS |
808 | /* |
809 | * If we are reclaiming on behalf of a cgroup, skip | |
810 | * counting on behalf of references from different | |
811 | * cgroups | |
812 | */ | |
bd845e38 | 813 | if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont)) |
bed7161a | 814 | continue; |
1cb1729b | 815 | referenced += page_referenced_one(page, vma, address, |
6fe6b7e3 | 816 | &mapcount, vm_flags); |
1da177e4 LT |
817 | if (!mapcount) |
818 | break; | |
819 | } | |
820 | ||
3d48ae45 | 821 | mutex_unlock(&mapping->i_mmap_mutex); |
1da177e4 LT |
822 | return referenced; |
823 | } | |
824 | ||
825 | /** | |
826 | * page_referenced - test if the page was referenced | |
827 | * @page: the page to test | |
828 | * @is_locked: caller holds lock on the page | |
43d8eac4 | 829 | * @mem_cont: target memory controller |
6fe6b7e3 | 830 | * @vm_flags: collect encountered vma->vm_flags who actually referenced the page |
1da177e4 LT |
831 | * |
832 | * Quick test_and_clear_referenced for all mappings to a page, | |
833 | * returns the number of ptes which referenced the page. | |
834 | */ | |
6fe6b7e3 WF |
835 | int page_referenced(struct page *page, |
836 | int is_locked, | |
837 | struct mem_cgroup *mem_cont, | |
838 | unsigned long *vm_flags) | |
1da177e4 LT |
839 | { |
840 | int referenced = 0; | |
5ad64688 | 841 | int we_locked = 0; |
1da177e4 | 842 | |
6fe6b7e3 | 843 | *vm_flags = 0; |
3ca7b3c5 | 844 | if (page_mapped(page) && page_rmapping(page)) { |
5ad64688 HD |
845 | if (!is_locked && (!PageAnon(page) || PageKsm(page))) { |
846 | we_locked = trylock_page(page); | |
847 | if (!we_locked) { | |
848 | referenced++; | |
849 | goto out; | |
850 | } | |
851 | } | |
852 | if (unlikely(PageKsm(page))) | |
853 | referenced += page_referenced_ksm(page, mem_cont, | |
854 | vm_flags); | |
855 | else if (PageAnon(page)) | |
6fe6b7e3 WF |
856 | referenced += page_referenced_anon(page, mem_cont, |
857 | vm_flags); | |
5ad64688 | 858 | else if (page->mapping) |
6fe6b7e3 WF |
859 | referenced += page_referenced_file(page, mem_cont, |
860 | vm_flags); | |
5ad64688 | 861 | if (we_locked) |
1da177e4 | 862 | unlock_page(page); |
1da177e4 | 863 | } |
5ad64688 | 864 | out: |
2d42552d | 865 | if (page_test_and_clear_young(page_to_pfn(page))) |
5b7baf05 CB |
866 | referenced++; |
867 | ||
1da177e4 LT |
868 | return referenced; |
869 | } | |
870 | ||
1cb1729b HD |
871 | static int page_mkclean_one(struct page *page, struct vm_area_struct *vma, |
872 | unsigned long address) | |
d08b3851 PZ |
873 | { |
874 | struct mm_struct *mm = vma->vm_mm; | |
c2fda5fe | 875 | pte_t *pte; |
d08b3851 PZ |
876 | spinlock_t *ptl; |
877 | int ret = 0; | |
878 | ||
479db0bf | 879 | pte = page_check_address(page, mm, address, &ptl, 1); |
d08b3851 PZ |
880 | if (!pte) |
881 | goto out; | |
882 | ||
c2fda5fe PZ |
883 | if (pte_dirty(*pte) || pte_write(*pte)) { |
884 | pte_t entry; | |
d08b3851 | 885 | |
c2fda5fe | 886 | flush_cache_page(vma, address, pte_pfn(*pte)); |
cddb8a5c | 887 | entry = ptep_clear_flush_notify(vma, address, pte); |
c2fda5fe PZ |
888 | entry = pte_wrprotect(entry); |
889 | entry = pte_mkclean(entry); | |
d6e88e67 | 890 | set_pte_at(mm, address, pte, entry); |
c2fda5fe PZ |
891 | ret = 1; |
892 | } | |
d08b3851 | 893 | |
d08b3851 PZ |
894 | pte_unmap_unlock(pte, ptl); |
895 | out: | |
896 | return ret; | |
897 | } | |
898 | ||
899 | static int page_mkclean_file(struct address_space *mapping, struct page *page) | |
900 | { | |
901 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
902 | struct vm_area_struct *vma; | |
903 | struct prio_tree_iter iter; | |
904 | int ret = 0; | |
905 | ||
906 | BUG_ON(PageAnon(page)); | |
907 | ||
3d48ae45 | 908 | mutex_lock(&mapping->i_mmap_mutex); |
d08b3851 | 909 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { |
1cb1729b HD |
910 | if (vma->vm_flags & VM_SHARED) { |
911 | unsigned long address = vma_address(page, vma); | |
912 | if (address == -EFAULT) | |
913 | continue; | |
914 | ret += page_mkclean_one(page, vma, address); | |
915 | } | |
d08b3851 | 916 | } |
3d48ae45 | 917 | mutex_unlock(&mapping->i_mmap_mutex); |
d08b3851 PZ |
918 | return ret; |
919 | } | |
920 | ||
921 | int page_mkclean(struct page *page) | |
922 | { | |
923 | int ret = 0; | |
924 | ||
925 | BUG_ON(!PageLocked(page)); | |
926 | ||
927 | if (page_mapped(page)) { | |
928 | struct address_space *mapping = page_mapping(page); | |
ce7e9fae | 929 | if (mapping) { |
d08b3851 | 930 | ret = page_mkclean_file(mapping, page); |
2d42552d | 931 | if (page_test_and_clear_dirty(page_to_pfn(page), 1)) |
ce7e9fae | 932 | ret = 1; |
6c210482 | 933 | } |
d08b3851 PZ |
934 | } |
935 | ||
936 | return ret; | |
937 | } | |
60b59bea | 938 | EXPORT_SYMBOL_GPL(page_mkclean); |
d08b3851 | 939 | |
c44b6743 RR |
940 | /** |
941 | * page_move_anon_rmap - move a page to our anon_vma | |
942 | * @page: the page to move to our anon_vma | |
943 | * @vma: the vma the page belongs to | |
944 | * @address: the user virtual address mapped | |
945 | * | |
946 | * When a page belongs exclusively to one process after a COW event, | |
947 | * that page can be moved into the anon_vma that belongs to just that | |
948 | * process, so the rmap code will not search the parent or sibling | |
949 | * processes. | |
950 | */ | |
951 | void page_move_anon_rmap(struct page *page, | |
952 | struct vm_area_struct *vma, unsigned long address) | |
953 | { | |
954 | struct anon_vma *anon_vma = vma->anon_vma; | |
955 | ||
956 | VM_BUG_ON(!PageLocked(page)); | |
957 | VM_BUG_ON(!anon_vma); | |
958 | VM_BUG_ON(page->index != linear_page_index(vma, address)); | |
959 | ||
960 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
961 | page->mapping = (struct address_space *) anon_vma; | |
962 | } | |
963 | ||
9617d95e | 964 | /** |
4e1c1975 AK |
965 | * __page_set_anon_rmap - set up new anonymous rmap |
966 | * @page: Page to add to rmap | |
967 | * @vma: VM area to add page to. | |
968 | * @address: User virtual address of the mapping | |
e8a03feb | 969 | * @exclusive: the page is exclusively owned by the current process |
9617d95e NP |
970 | */ |
971 | static void __page_set_anon_rmap(struct page *page, | |
e8a03feb | 972 | struct vm_area_struct *vma, unsigned long address, int exclusive) |
9617d95e | 973 | { |
e8a03feb | 974 | struct anon_vma *anon_vma = vma->anon_vma; |
ea90002b | 975 | |
e8a03feb | 976 | BUG_ON(!anon_vma); |
ea90002b | 977 | |
4e1c1975 AK |
978 | if (PageAnon(page)) |
979 | return; | |
980 | ||
ea90002b | 981 | /* |
e8a03feb RR |
982 | * If the page isn't exclusively mapped into this vma, |
983 | * we must use the _oldest_ possible anon_vma for the | |
984 | * page mapping! | |
ea90002b | 985 | */ |
4e1c1975 | 986 | if (!exclusive) |
288468c3 | 987 | anon_vma = anon_vma->root; |
9617d95e | 988 | |
9617d95e NP |
989 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; |
990 | page->mapping = (struct address_space *) anon_vma; | |
9617d95e | 991 | page->index = linear_page_index(vma, address); |
9617d95e NP |
992 | } |
993 | ||
c97a9e10 | 994 | /** |
43d8eac4 | 995 | * __page_check_anon_rmap - sanity check anonymous rmap addition |
c97a9e10 NP |
996 | * @page: the page to add the mapping to |
997 | * @vma: the vm area in which the mapping is added | |
998 | * @address: the user virtual address mapped | |
999 | */ | |
1000 | static void __page_check_anon_rmap(struct page *page, | |
1001 | struct vm_area_struct *vma, unsigned long address) | |
1002 | { | |
1003 | #ifdef CONFIG_DEBUG_VM | |
1004 | /* | |
1005 | * The page's anon-rmap details (mapping and index) are guaranteed to | |
1006 | * be set up correctly at this point. | |
1007 | * | |
1008 | * We have exclusion against page_add_anon_rmap because the caller | |
1009 | * always holds the page locked, except if called from page_dup_rmap, | |
1010 | * in which case the page is already known to be setup. | |
1011 | * | |
1012 | * We have exclusion against page_add_new_anon_rmap because those pages | |
1013 | * are initially only visible via the pagetables, and the pte is locked | |
1014 | * over the call to page_add_new_anon_rmap. | |
1015 | */ | |
44ab57a0 | 1016 | BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root); |
c97a9e10 NP |
1017 | BUG_ON(page->index != linear_page_index(vma, address)); |
1018 | #endif | |
1019 | } | |
1020 | ||
1da177e4 LT |
1021 | /** |
1022 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
1023 | * @page: the page to add the mapping to | |
1024 | * @vma: the vm area in which the mapping is added | |
1025 | * @address: the user virtual address mapped | |
1026 | * | |
5ad64688 | 1027 | * The caller needs to hold the pte lock, and the page must be locked in |
80e14822 HD |
1028 | * the anon_vma case: to serialize mapping,index checking after setting, |
1029 | * and to ensure that PageAnon is not being upgraded racily to PageKsm | |
1030 | * (but PageKsm is never downgraded to PageAnon). | |
1da177e4 LT |
1031 | */ |
1032 | void page_add_anon_rmap(struct page *page, | |
1033 | struct vm_area_struct *vma, unsigned long address) | |
ad8c2ee8 RR |
1034 | { |
1035 | do_page_add_anon_rmap(page, vma, address, 0); | |
1036 | } | |
1037 | ||
1038 | /* | |
1039 | * Special version of the above for do_swap_page, which often runs | |
1040 | * into pages that are exclusively owned by the current process. | |
1041 | * Everybody else should continue to use page_add_anon_rmap above. | |
1042 | */ | |
1043 | void do_page_add_anon_rmap(struct page *page, | |
1044 | struct vm_area_struct *vma, unsigned long address, int exclusive) | |
1da177e4 | 1045 | { |
5ad64688 | 1046 | int first = atomic_inc_and_test(&page->_mapcount); |
79134171 AA |
1047 | if (first) { |
1048 | if (!PageTransHuge(page)) | |
1049 | __inc_zone_page_state(page, NR_ANON_PAGES); | |
1050 | else | |
1051 | __inc_zone_page_state(page, | |
1052 | NR_ANON_TRANSPARENT_HUGEPAGES); | |
1053 | } | |
5ad64688 HD |
1054 | if (unlikely(PageKsm(page))) |
1055 | return; | |
1056 | ||
c97a9e10 | 1057 | VM_BUG_ON(!PageLocked(page)); |
5dbe0af4 | 1058 | /* address might be in next vma when migration races vma_adjust */ |
5ad64688 | 1059 | if (first) |
ad8c2ee8 | 1060 | __page_set_anon_rmap(page, vma, address, exclusive); |
69029cd5 | 1061 | else |
c97a9e10 | 1062 | __page_check_anon_rmap(page, vma, address); |
1da177e4 LT |
1063 | } |
1064 | ||
43d8eac4 | 1065 | /** |
9617d95e NP |
1066 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page |
1067 | * @page: the page to add the mapping to | |
1068 | * @vma: the vm area in which the mapping is added | |
1069 | * @address: the user virtual address mapped | |
1070 | * | |
1071 | * Same as page_add_anon_rmap but must only be called on *new* pages. | |
1072 | * This means the inc-and-test can be bypassed. | |
c97a9e10 | 1073 | * Page does not have to be locked. |
9617d95e NP |
1074 | */ |
1075 | void page_add_new_anon_rmap(struct page *page, | |
1076 | struct vm_area_struct *vma, unsigned long address) | |
1077 | { | |
b5934c53 | 1078 | VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); |
cbf84b7a HD |
1079 | SetPageSwapBacked(page); |
1080 | atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */ | |
79134171 AA |
1081 | if (!PageTransHuge(page)) |
1082 | __inc_zone_page_state(page, NR_ANON_PAGES); | |
1083 | else | |
1084 | __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES); | |
e8a03feb | 1085 | __page_set_anon_rmap(page, vma, address, 1); |
b5934c53 | 1086 | if (page_evictable(page, vma)) |
cbf84b7a | 1087 | lru_cache_add_lru(page, LRU_ACTIVE_ANON); |
b5934c53 HD |
1088 | else |
1089 | add_page_to_unevictable_list(page); | |
9617d95e NP |
1090 | } |
1091 | ||
1da177e4 LT |
1092 | /** |
1093 | * page_add_file_rmap - add pte mapping to a file page | |
1094 | * @page: the page to add the mapping to | |
1095 | * | |
b8072f09 | 1096 | * The caller needs to hold the pte lock. |
1da177e4 LT |
1097 | */ |
1098 | void page_add_file_rmap(struct page *page) | |
1099 | { | |
d69b042f | 1100 | if (atomic_inc_and_test(&page->_mapcount)) { |
65ba55f5 | 1101 | __inc_zone_page_state(page, NR_FILE_MAPPED); |
2a7106f2 | 1102 | mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED); |
d69b042f | 1103 | } |
1da177e4 LT |
1104 | } |
1105 | ||
1106 | /** | |
1107 | * page_remove_rmap - take down pte mapping from a page | |
1108 | * @page: page to remove mapping from | |
1109 | * | |
b8072f09 | 1110 | * The caller needs to hold the pte lock. |
1da177e4 | 1111 | */ |
edc315fd | 1112 | void page_remove_rmap(struct page *page) |
1da177e4 | 1113 | { |
b904dcfe KM |
1114 | /* page still mapped by someone else? */ |
1115 | if (!atomic_add_negative(-1, &page->_mapcount)) | |
1116 | return; | |
1117 | ||
1118 | /* | |
1119 | * Now that the last pte has gone, s390 must transfer dirty | |
1120 | * flag from storage key to struct page. We can usually skip | |
1121 | * this if the page is anon, so about to be freed; but perhaps | |
1122 | * not if it's in swapcache - there might be another pte slot | |
1123 | * containing the swap entry, but page not yet written to swap. | |
1124 | */ | |
2d42552d MS |
1125 | if ((!PageAnon(page) || PageSwapCache(page)) && |
1126 | page_test_and_clear_dirty(page_to_pfn(page), 1)) | |
b904dcfe | 1127 | set_page_dirty(page); |
0fe6e20b NH |
1128 | /* |
1129 | * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED | |
1130 | * and not charged by memcg for now. | |
1131 | */ | |
1132 | if (unlikely(PageHuge(page))) | |
1133 | return; | |
b904dcfe KM |
1134 | if (PageAnon(page)) { |
1135 | mem_cgroup_uncharge_page(page); | |
79134171 AA |
1136 | if (!PageTransHuge(page)) |
1137 | __dec_zone_page_state(page, NR_ANON_PAGES); | |
1138 | else | |
1139 | __dec_zone_page_state(page, | |
1140 | NR_ANON_TRANSPARENT_HUGEPAGES); | |
b904dcfe KM |
1141 | } else { |
1142 | __dec_zone_page_state(page, NR_FILE_MAPPED); | |
2a7106f2 | 1143 | mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED); |
b904dcfe | 1144 | } |
b904dcfe KM |
1145 | /* |
1146 | * It would be tidy to reset the PageAnon mapping here, | |
1147 | * but that might overwrite a racing page_add_anon_rmap | |
1148 | * which increments mapcount after us but sets mapping | |
1149 | * before us: so leave the reset to free_hot_cold_page, | |
1150 | * and remember that it's only reliable while mapped. | |
1151 | * Leaving it set also helps swapoff to reinstate ptes | |
1152 | * faster for those pages still in swapcache. | |
1153 | */ | |
1da177e4 LT |
1154 | } |
1155 | ||
1156 | /* | |
1157 | * Subfunctions of try_to_unmap: try_to_unmap_one called | |
1158 | * repeatedly from either try_to_unmap_anon or try_to_unmap_file. | |
1159 | */ | |
5ad64688 HD |
1160 | int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, |
1161 | unsigned long address, enum ttu_flags flags) | |
1da177e4 LT |
1162 | { |
1163 | struct mm_struct *mm = vma->vm_mm; | |
1da177e4 LT |
1164 | pte_t *pte; |
1165 | pte_t pteval; | |
c0718806 | 1166 | spinlock_t *ptl; |
1da177e4 LT |
1167 | int ret = SWAP_AGAIN; |
1168 | ||
479db0bf | 1169 | pte = page_check_address(page, mm, address, &ptl, 0); |
c0718806 | 1170 | if (!pte) |
81b4082d | 1171 | goto out; |
1da177e4 LT |
1172 | |
1173 | /* | |
1174 | * If the page is mlock()d, we cannot swap it out. | |
1175 | * If it's recently referenced (perhaps page_referenced | |
1176 | * skipped over this mm) then we should reactivate it. | |
1177 | */ | |
14fa31b8 | 1178 | if (!(flags & TTU_IGNORE_MLOCK)) { |
caed0f48 KM |
1179 | if (vma->vm_flags & VM_LOCKED) |
1180 | goto out_mlock; | |
1181 | ||
af8e3354 | 1182 | if (TTU_ACTION(flags) == TTU_MUNLOCK) |
53f79acb | 1183 | goto out_unmap; |
14fa31b8 AK |
1184 | } |
1185 | if (!(flags & TTU_IGNORE_ACCESS)) { | |
b291f000 NP |
1186 | if (ptep_clear_flush_young_notify(vma, address, pte)) { |
1187 | ret = SWAP_FAIL; | |
1188 | goto out_unmap; | |
1189 | } | |
1190 | } | |
1da177e4 | 1191 | |
1da177e4 LT |
1192 | /* Nuke the page table entry. */ |
1193 | flush_cache_page(vma, address, page_to_pfn(page)); | |
cddb8a5c | 1194 | pteval = ptep_clear_flush_notify(vma, address, pte); |
1da177e4 LT |
1195 | |
1196 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
1197 | if (pte_dirty(pteval)) | |
1198 | set_page_dirty(page); | |
1199 | ||
365e9c87 HD |
1200 | /* Update high watermark before we lower rss */ |
1201 | update_hiwater_rss(mm); | |
1202 | ||
888b9f7c AK |
1203 | if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) { |
1204 | if (PageAnon(page)) | |
d559db08 | 1205 | dec_mm_counter(mm, MM_ANONPAGES); |
888b9f7c | 1206 | else |
d559db08 | 1207 | dec_mm_counter(mm, MM_FILEPAGES); |
888b9f7c AK |
1208 | set_pte_at(mm, address, pte, |
1209 | swp_entry_to_pte(make_hwpoison_entry(page))); | |
1210 | } else if (PageAnon(page)) { | |
4c21e2f2 | 1211 | swp_entry_t entry = { .val = page_private(page) }; |
0697212a CL |
1212 | |
1213 | if (PageSwapCache(page)) { | |
1214 | /* | |
1215 | * Store the swap location in the pte. | |
1216 | * See handle_pte_fault() ... | |
1217 | */ | |
570a335b HD |
1218 | if (swap_duplicate(entry) < 0) { |
1219 | set_pte_at(mm, address, pte, pteval); | |
1220 | ret = SWAP_FAIL; | |
1221 | goto out_unmap; | |
1222 | } | |
0697212a CL |
1223 | if (list_empty(&mm->mmlist)) { |
1224 | spin_lock(&mmlist_lock); | |
1225 | if (list_empty(&mm->mmlist)) | |
1226 | list_add(&mm->mmlist, &init_mm.mmlist); | |
1227 | spin_unlock(&mmlist_lock); | |
1228 | } | |
d559db08 | 1229 | dec_mm_counter(mm, MM_ANONPAGES); |
b084d435 | 1230 | inc_mm_counter(mm, MM_SWAPENTS); |
64cdd548 | 1231 | } else if (PAGE_MIGRATION) { |
0697212a CL |
1232 | /* |
1233 | * Store the pfn of the page in a special migration | |
1234 | * pte. do_swap_page() will wait until the migration | |
1235 | * pte is removed and then restart fault handling. | |
1236 | */ | |
14fa31b8 | 1237 | BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION); |
0697212a | 1238 | entry = make_migration_entry(page, pte_write(pteval)); |
1da177e4 LT |
1239 | } |
1240 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
1241 | BUG_ON(pte_file(*pte)); | |
14fa31b8 | 1242 | } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) { |
04e62a29 CL |
1243 | /* Establish migration entry for a file page */ |
1244 | swp_entry_t entry; | |
1245 | entry = make_migration_entry(page, pte_write(pteval)); | |
1246 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
1247 | } else | |
d559db08 | 1248 | dec_mm_counter(mm, MM_FILEPAGES); |
1da177e4 | 1249 | |
edc315fd | 1250 | page_remove_rmap(page); |
1da177e4 LT |
1251 | page_cache_release(page); |
1252 | ||
1253 | out_unmap: | |
c0718806 | 1254 | pte_unmap_unlock(pte, ptl); |
caed0f48 KM |
1255 | out: |
1256 | return ret; | |
53f79acb | 1257 | |
caed0f48 KM |
1258 | out_mlock: |
1259 | pte_unmap_unlock(pte, ptl); | |
1260 | ||
1261 | ||
1262 | /* | |
1263 | * We need mmap_sem locking, Otherwise VM_LOCKED check makes | |
1264 | * unstable result and race. Plus, We can't wait here because | |
2b575eb6 | 1265 | * we now hold anon_vma->mutex or mapping->i_mmap_mutex. |
caed0f48 KM |
1266 | * if trylock failed, the page remain in evictable lru and later |
1267 | * vmscan could retry to move the page to unevictable lru if the | |
1268 | * page is actually mlocked. | |
1269 | */ | |
1270 | if (down_read_trylock(&vma->vm_mm->mmap_sem)) { | |
1271 | if (vma->vm_flags & VM_LOCKED) { | |
1272 | mlock_vma_page(page); | |
1273 | ret = SWAP_MLOCK; | |
53f79acb | 1274 | } |
caed0f48 | 1275 | up_read(&vma->vm_mm->mmap_sem); |
53f79acb | 1276 | } |
1da177e4 LT |
1277 | return ret; |
1278 | } | |
1279 | ||
1280 | /* | |
1281 | * objrmap doesn't work for nonlinear VMAs because the assumption that | |
1282 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. | |
1283 | * Consequently, given a particular page and its ->index, we cannot locate the | |
1284 | * ptes which are mapping that page without an exhaustive linear search. | |
1285 | * | |
1286 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which | |
1287 | * maps the file to which the target page belongs. The ->vm_private_data field | |
1288 | * holds the current cursor into that scan. Successive searches will circulate | |
1289 | * around the vma's virtual address space. | |
1290 | * | |
1291 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, | |
1292 | * more scanning pressure is placed against them as well. Eventually pages | |
1293 | * will become fully unmapped and are eligible for eviction. | |
1294 | * | |
1295 | * For very sparsely populated VMAs this is a little inefficient - chances are | |
1296 | * there there won't be many ptes located within the scan cluster. In this case | |
1297 | * maybe we could scan further - to the end of the pte page, perhaps. | |
b291f000 NP |
1298 | * |
1299 | * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can | |
1300 | * acquire it without blocking. If vma locked, mlock the pages in the cluster, | |
1301 | * rather than unmapping them. If we encounter the "check_page" that vmscan is | |
1302 | * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN. | |
1da177e4 LT |
1303 | */ |
1304 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) | |
1305 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) | |
1306 | ||
b291f000 NP |
1307 | static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount, |
1308 | struct vm_area_struct *vma, struct page *check_page) | |
1da177e4 LT |
1309 | { |
1310 | struct mm_struct *mm = vma->vm_mm; | |
1311 | pgd_t *pgd; | |
1312 | pud_t *pud; | |
1313 | pmd_t *pmd; | |
c0718806 | 1314 | pte_t *pte; |
1da177e4 | 1315 | pte_t pteval; |
c0718806 | 1316 | spinlock_t *ptl; |
1da177e4 LT |
1317 | struct page *page; |
1318 | unsigned long address; | |
1319 | unsigned long end; | |
b291f000 NP |
1320 | int ret = SWAP_AGAIN; |
1321 | int locked_vma = 0; | |
1da177e4 | 1322 | |
1da177e4 LT |
1323 | address = (vma->vm_start + cursor) & CLUSTER_MASK; |
1324 | end = address + CLUSTER_SIZE; | |
1325 | if (address < vma->vm_start) | |
1326 | address = vma->vm_start; | |
1327 | if (end > vma->vm_end) | |
1328 | end = vma->vm_end; | |
1329 | ||
1330 | pgd = pgd_offset(mm, address); | |
1331 | if (!pgd_present(*pgd)) | |
b291f000 | 1332 | return ret; |
1da177e4 LT |
1333 | |
1334 | pud = pud_offset(pgd, address); | |
1335 | if (!pud_present(*pud)) | |
b291f000 | 1336 | return ret; |
1da177e4 LT |
1337 | |
1338 | pmd = pmd_offset(pud, address); | |
1339 | if (!pmd_present(*pmd)) | |
b291f000 NP |
1340 | return ret; |
1341 | ||
1342 | /* | |
af8e3354 | 1343 | * If we can acquire the mmap_sem for read, and vma is VM_LOCKED, |
b291f000 NP |
1344 | * keep the sem while scanning the cluster for mlocking pages. |
1345 | */ | |
af8e3354 | 1346 | if (down_read_trylock(&vma->vm_mm->mmap_sem)) { |
b291f000 NP |
1347 | locked_vma = (vma->vm_flags & VM_LOCKED); |
1348 | if (!locked_vma) | |
1349 | up_read(&vma->vm_mm->mmap_sem); /* don't need it */ | |
1350 | } | |
c0718806 HD |
1351 | |
1352 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
1da177e4 | 1353 | |
365e9c87 HD |
1354 | /* Update high watermark before we lower rss */ |
1355 | update_hiwater_rss(mm); | |
1356 | ||
c0718806 | 1357 | for (; address < end; pte++, address += PAGE_SIZE) { |
1da177e4 LT |
1358 | if (!pte_present(*pte)) |
1359 | continue; | |
6aab341e LT |
1360 | page = vm_normal_page(vma, address, *pte); |
1361 | BUG_ON(!page || PageAnon(page)); | |
1da177e4 | 1362 | |
b291f000 NP |
1363 | if (locked_vma) { |
1364 | mlock_vma_page(page); /* no-op if already mlocked */ | |
1365 | if (page == check_page) | |
1366 | ret = SWAP_MLOCK; | |
1367 | continue; /* don't unmap */ | |
1368 | } | |
1369 | ||
cddb8a5c | 1370 | if (ptep_clear_flush_young_notify(vma, address, pte)) |
1da177e4 LT |
1371 | continue; |
1372 | ||
1373 | /* Nuke the page table entry. */ | |
eca35133 | 1374 | flush_cache_page(vma, address, pte_pfn(*pte)); |
cddb8a5c | 1375 | pteval = ptep_clear_flush_notify(vma, address, pte); |
1da177e4 LT |
1376 | |
1377 | /* If nonlinear, store the file page offset in the pte. */ | |
1378 | if (page->index != linear_page_index(vma, address)) | |
1379 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); | |
1380 | ||
1381 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
1382 | if (pte_dirty(pteval)) | |
1383 | set_page_dirty(page); | |
1384 | ||
edc315fd | 1385 | page_remove_rmap(page); |
1da177e4 | 1386 | page_cache_release(page); |
d559db08 | 1387 | dec_mm_counter(mm, MM_FILEPAGES); |
1da177e4 LT |
1388 | (*mapcount)--; |
1389 | } | |
c0718806 | 1390 | pte_unmap_unlock(pte - 1, ptl); |
b291f000 NP |
1391 | if (locked_vma) |
1392 | up_read(&vma->vm_mm->mmap_sem); | |
1393 | return ret; | |
1da177e4 LT |
1394 | } |
1395 | ||
71e3aac0 | 1396 | bool is_vma_temporary_stack(struct vm_area_struct *vma) |
a8bef8ff MG |
1397 | { |
1398 | int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP); | |
1399 | ||
1400 | if (!maybe_stack) | |
1401 | return false; | |
1402 | ||
1403 | if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) == | |
1404 | VM_STACK_INCOMPLETE_SETUP) | |
1405 | return true; | |
1406 | ||
1407 | return false; | |
1408 | } | |
1409 | ||
b291f000 NP |
1410 | /** |
1411 | * try_to_unmap_anon - unmap or unlock anonymous page using the object-based | |
1412 | * rmap method | |
1413 | * @page: the page to unmap/unlock | |
8051be5e | 1414 | * @flags: action and flags |
b291f000 NP |
1415 | * |
1416 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
1417 | * contained in the anon_vma struct it points to. | |
1418 | * | |
1419 | * This function is only called from try_to_unmap/try_to_munlock for | |
1420 | * anonymous pages. | |
1421 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma | |
1422 | * where the page was found will be held for write. So, we won't recheck | |
1423 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be | |
1424 | * 'LOCKED. | |
1425 | */ | |
14fa31b8 | 1426 | static int try_to_unmap_anon(struct page *page, enum ttu_flags flags) |
1da177e4 LT |
1427 | { |
1428 | struct anon_vma *anon_vma; | |
5beb4930 | 1429 | struct anon_vma_chain *avc; |
1da177e4 | 1430 | int ret = SWAP_AGAIN; |
b291f000 | 1431 | |
1da177e4 LT |
1432 | anon_vma = page_lock_anon_vma(page); |
1433 | if (!anon_vma) | |
1434 | return ret; | |
1435 | ||
5beb4930 RR |
1436 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { |
1437 | struct vm_area_struct *vma = avc->vma; | |
a8bef8ff MG |
1438 | unsigned long address; |
1439 | ||
1440 | /* | |
1441 | * During exec, a temporary VMA is setup and later moved. | |
1442 | * The VMA is moved under the anon_vma lock but not the | |
1443 | * page tables leading to a race where migration cannot | |
1444 | * find the migration ptes. Rather than increasing the | |
1445 | * locking requirements of exec(), migration skips | |
1446 | * temporary VMAs until after exec() completes. | |
1447 | */ | |
1448 | if (PAGE_MIGRATION && (flags & TTU_MIGRATION) && | |
1449 | is_vma_temporary_stack(vma)) | |
1450 | continue; | |
1451 | ||
1452 | address = vma_address(page, vma); | |
1cb1729b HD |
1453 | if (address == -EFAULT) |
1454 | continue; | |
1455 | ret = try_to_unmap_one(page, vma, address, flags); | |
53f79acb HD |
1456 | if (ret != SWAP_AGAIN || !page_mapped(page)) |
1457 | break; | |
1da177e4 | 1458 | } |
34bbd704 ON |
1459 | |
1460 | page_unlock_anon_vma(anon_vma); | |
1da177e4 LT |
1461 | return ret; |
1462 | } | |
1463 | ||
1464 | /** | |
b291f000 NP |
1465 | * try_to_unmap_file - unmap/unlock file page using the object-based rmap method |
1466 | * @page: the page to unmap/unlock | |
14fa31b8 | 1467 | * @flags: action and flags |
1da177e4 LT |
1468 | * |
1469 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
1470 | * contained in the address_space struct it points to. | |
1471 | * | |
b291f000 NP |
1472 | * This function is only called from try_to_unmap/try_to_munlock for |
1473 | * object-based pages. | |
1474 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma | |
1475 | * where the page was found will be held for write. So, we won't recheck | |
1476 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be | |
1477 | * 'LOCKED. | |
1da177e4 | 1478 | */ |
14fa31b8 | 1479 | static int try_to_unmap_file(struct page *page, enum ttu_flags flags) |
1da177e4 LT |
1480 | { |
1481 | struct address_space *mapping = page->mapping; | |
1482 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
1483 | struct vm_area_struct *vma; | |
1484 | struct prio_tree_iter iter; | |
1485 | int ret = SWAP_AGAIN; | |
1486 | unsigned long cursor; | |
1487 | unsigned long max_nl_cursor = 0; | |
1488 | unsigned long max_nl_size = 0; | |
1489 | unsigned int mapcount; | |
1490 | ||
3d48ae45 | 1491 | mutex_lock(&mapping->i_mmap_mutex); |
1da177e4 | 1492 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { |
1cb1729b HD |
1493 | unsigned long address = vma_address(page, vma); |
1494 | if (address == -EFAULT) | |
1495 | continue; | |
1496 | ret = try_to_unmap_one(page, vma, address, flags); | |
53f79acb HD |
1497 | if (ret != SWAP_AGAIN || !page_mapped(page)) |
1498 | goto out; | |
1da177e4 LT |
1499 | } |
1500 | ||
1501 | if (list_empty(&mapping->i_mmap_nonlinear)) | |
1502 | goto out; | |
1503 | ||
53f79acb HD |
1504 | /* |
1505 | * We don't bother to try to find the munlocked page in nonlinears. | |
1506 | * It's costly. Instead, later, page reclaim logic may call | |
1507 | * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily. | |
1508 | */ | |
1509 | if (TTU_ACTION(flags) == TTU_MUNLOCK) | |
1510 | goto out; | |
1511 | ||
1da177e4 LT |
1512 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, |
1513 | shared.vm_set.list) { | |
1da177e4 LT |
1514 | cursor = (unsigned long) vma->vm_private_data; |
1515 | if (cursor > max_nl_cursor) | |
1516 | max_nl_cursor = cursor; | |
1517 | cursor = vma->vm_end - vma->vm_start; | |
1518 | if (cursor > max_nl_size) | |
1519 | max_nl_size = cursor; | |
1520 | } | |
1521 | ||
b291f000 | 1522 | if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */ |
1da177e4 LT |
1523 | ret = SWAP_FAIL; |
1524 | goto out; | |
1525 | } | |
1526 | ||
1527 | /* | |
1528 | * We don't try to search for this page in the nonlinear vmas, | |
1529 | * and page_referenced wouldn't have found it anyway. Instead | |
1530 | * just walk the nonlinear vmas trying to age and unmap some. | |
1531 | * The mapcount of the page we came in with is irrelevant, | |
1532 | * but even so use it as a guide to how hard we should try? | |
1533 | */ | |
1534 | mapcount = page_mapcount(page); | |
1535 | if (!mapcount) | |
1536 | goto out; | |
3d48ae45 | 1537 | cond_resched(); |
1da177e4 LT |
1538 | |
1539 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; | |
1540 | if (max_nl_cursor == 0) | |
1541 | max_nl_cursor = CLUSTER_SIZE; | |
1542 | ||
1543 | do { | |
1544 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
1545 | shared.vm_set.list) { | |
1da177e4 | 1546 | cursor = (unsigned long) vma->vm_private_data; |
839b9685 | 1547 | while ( cursor < max_nl_cursor && |
1da177e4 | 1548 | cursor < vma->vm_end - vma->vm_start) { |
53f79acb HD |
1549 | if (try_to_unmap_cluster(cursor, &mapcount, |
1550 | vma, page) == SWAP_MLOCK) | |
1551 | ret = SWAP_MLOCK; | |
1da177e4 LT |
1552 | cursor += CLUSTER_SIZE; |
1553 | vma->vm_private_data = (void *) cursor; | |
1554 | if ((int)mapcount <= 0) | |
1555 | goto out; | |
1556 | } | |
1557 | vma->vm_private_data = (void *) max_nl_cursor; | |
1558 | } | |
3d48ae45 | 1559 | cond_resched(); |
1da177e4 LT |
1560 | max_nl_cursor += CLUSTER_SIZE; |
1561 | } while (max_nl_cursor <= max_nl_size); | |
1562 | ||
1563 | /* | |
1564 | * Don't loop forever (perhaps all the remaining pages are | |
1565 | * in locked vmas). Reset cursor on all unreserved nonlinear | |
1566 | * vmas, now forgetting on which ones it had fallen behind. | |
1567 | */ | |
101d2be7 HD |
1568 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) |
1569 | vma->vm_private_data = NULL; | |
1da177e4 | 1570 | out: |
3d48ae45 | 1571 | mutex_unlock(&mapping->i_mmap_mutex); |
1da177e4 LT |
1572 | return ret; |
1573 | } | |
1574 | ||
1575 | /** | |
1576 | * try_to_unmap - try to remove all page table mappings to a page | |
1577 | * @page: the page to get unmapped | |
14fa31b8 | 1578 | * @flags: action and flags |
1da177e4 LT |
1579 | * |
1580 | * Tries to remove all the page table entries which are mapping this | |
1581 | * page, used in the pageout path. Caller must hold the page lock. | |
1582 | * Return values are: | |
1583 | * | |
1584 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
1585 | * SWAP_AGAIN - we missed a mapping, try again later | |
1586 | * SWAP_FAIL - the page is unswappable | |
b291f000 | 1587 | * SWAP_MLOCK - page is mlocked. |
1da177e4 | 1588 | */ |
14fa31b8 | 1589 | int try_to_unmap(struct page *page, enum ttu_flags flags) |
1da177e4 LT |
1590 | { |
1591 | int ret; | |
1592 | ||
1da177e4 | 1593 | BUG_ON(!PageLocked(page)); |
91600e9e | 1594 | VM_BUG_ON(!PageHuge(page) && PageTransHuge(page)); |
1da177e4 | 1595 | |
5ad64688 HD |
1596 | if (unlikely(PageKsm(page))) |
1597 | ret = try_to_unmap_ksm(page, flags); | |
1598 | else if (PageAnon(page)) | |
14fa31b8 | 1599 | ret = try_to_unmap_anon(page, flags); |
1da177e4 | 1600 | else |
14fa31b8 | 1601 | ret = try_to_unmap_file(page, flags); |
b291f000 | 1602 | if (ret != SWAP_MLOCK && !page_mapped(page)) |
1da177e4 LT |
1603 | ret = SWAP_SUCCESS; |
1604 | return ret; | |
1605 | } | |
81b4082d | 1606 | |
b291f000 NP |
1607 | /** |
1608 | * try_to_munlock - try to munlock a page | |
1609 | * @page: the page to be munlocked | |
1610 | * | |
1611 | * Called from munlock code. Checks all of the VMAs mapping the page | |
1612 | * to make sure nobody else has this page mlocked. The page will be | |
1613 | * returned with PG_mlocked cleared if no other vmas have it mlocked. | |
1614 | * | |
1615 | * Return values are: | |
1616 | * | |
53f79acb | 1617 | * SWAP_AGAIN - no vma is holding page mlocked, or, |
b291f000 | 1618 | * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem |
5ad64688 | 1619 | * SWAP_FAIL - page cannot be located at present |
b291f000 NP |
1620 | * SWAP_MLOCK - page is now mlocked. |
1621 | */ | |
1622 | int try_to_munlock(struct page *page) | |
1623 | { | |
1624 | VM_BUG_ON(!PageLocked(page) || PageLRU(page)); | |
1625 | ||
5ad64688 HD |
1626 | if (unlikely(PageKsm(page))) |
1627 | return try_to_unmap_ksm(page, TTU_MUNLOCK); | |
1628 | else if (PageAnon(page)) | |
14fa31b8 | 1629 | return try_to_unmap_anon(page, TTU_MUNLOCK); |
b291f000 | 1630 | else |
14fa31b8 | 1631 | return try_to_unmap_file(page, TTU_MUNLOCK); |
b291f000 | 1632 | } |
e9995ef9 | 1633 | |
01d8b20d | 1634 | void __put_anon_vma(struct anon_vma *anon_vma) |
76545066 | 1635 | { |
01d8b20d | 1636 | struct anon_vma *root = anon_vma->root; |
76545066 | 1637 | |
01d8b20d PZ |
1638 | if (root != anon_vma && atomic_dec_and_test(&root->refcount)) |
1639 | anon_vma_free(root); | |
76545066 | 1640 | |
01d8b20d | 1641 | anon_vma_free(anon_vma); |
76545066 | 1642 | } |
76545066 | 1643 | |
e9995ef9 HD |
1644 | #ifdef CONFIG_MIGRATION |
1645 | /* | |
1646 | * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file(): | |
1647 | * Called by migrate.c to remove migration ptes, but might be used more later. | |
1648 | */ | |
1649 | static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *, | |
1650 | struct vm_area_struct *, unsigned long, void *), void *arg) | |
1651 | { | |
1652 | struct anon_vma *anon_vma; | |
5beb4930 | 1653 | struct anon_vma_chain *avc; |
e9995ef9 HD |
1654 | int ret = SWAP_AGAIN; |
1655 | ||
1656 | /* | |
1657 | * Note: remove_migration_ptes() cannot use page_lock_anon_vma() | |
1658 | * because that depends on page_mapped(); but not all its usages | |
3f6c8272 MG |
1659 | * are holding mmap_sem. Users without mmap_sem are required to |
1660 | * take a reference count to prevent the anon_vma disappearing | |
e9995ef9 HD |
1661 | */ |
1662 | anon_vma = page_anon_vma(page); | |
1663 | if (!anon_vma) | |
1664 | return ret; | |
cba48b98 | 1665 | anon_vma_lock(anon_vma); |
5beb4930 RR |
1666 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { |
1667 | struct vm_area_struct *vma = avc->vma; | |
e9995ef9 HD |
1668 | unsigned long address = vma_address(page, vma); |
1669 | if (address == -EFAULT) | |
1670 | continue; | |
1671 | ret = rmap_one(page, vma, address, arg); | |
1672 | if (ret != SWAP_AGAIN) | |
1673 | break; | |
1674 | } | |
cba48b98 | 1675 | anon_vma_unlock(anon_vma); |
e9995ef9 HD |
1676 | return ret; |
1677 | } | |
1678 | ||
1679 | static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *, | |
1680 | struct vm_area_struct *, unsigned long, void *), void *arg) | |
1681 | { | |
1682 | struct address_space *mapping = page->mapping; | |
1683 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
1684 | struct vm_area_struct *vma; | |
1685 | struct prio_tree_iter iter; | |
1686 | int ret = SWAP_AGAIN; | |
1687 | ||
1688 | if (!mapping) | |
1689 | return ret; | |
3d48ae45 | 1690 | mutex_lock(&mapping->i_mmap_mutex); |
e9995ef9 HD |
1691 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { |
1692 | unsigned long address = vma_address(page, vma); | |
1693 | if (address == -EFAULT) | |
1694 | continue; | |
1695 | ret = rmap_one(page, vma, address, arg); | |
1696 | if (ret != SWAP_AGAIN) | |
1697 | break; | |
1698 | } | |
1699 | /* | |
1700 | * No nonlinear handling: being always shared, nonlinear vmas | |
1701 | * never contain migration ptes. Decide what to do about this | |
1702 | * limitation to linear when we need rmap_walk() on nonlinear. | |
1703 | */ | |
3d48ae45 | 1704 | mutex_unlock(&mapping->i_mmap_mutex); |
e9995ef9 HD |
1705 | return ret; |
1706 | } | |
1707 | ||
1708 | int rmap_walk(struct page *page, int (*rmap_one)(struct page *, | |
1709 | struct vm_area_struct *, unsigned long, void *), void *arg) | |
1710 | { | |
1711 | VM_BUG_ON(!PageLocked(page)); | |
1712 | ||
1713 | if (unlikely(PageKsm(page))) | |
1714 | return rmap_walk_ksm(page, rmap_one, arg); | |
1715 | else if (PageAnon(page)) | |
1716 | return rmap_walk_anon(page, rmap_one, arg); | |
1717 | else | |
1718 | return rmap_walk_file(page, rmap_one, arg); | |
1719 | } | |
1720 | #endif /* CONFIG_MIGRATION */ | |
0fe6e20b | 1721 | |
e3390f67 | 1722 | #ifdef CONFIG_HUGETLB_PAGE |
0fe6e20b NH |
1723 | /* |
1724 | * The following three functions are for anonymous (private mapped) hugepages. | |
1725 | * Unlike common anonymous pages, anonymous hugepages have no accounting code | |
1726 | * and no lru code, because we handle hugepages differently from common pages. | |
1727 | */ | |
1728 | static void __hugepage_set_anon_rmap(struct page *page, | |
1729 | struct vm_area_struct *vma, unsigned long address, int exclusive) | |
1730 | { | |
1731 | struct anon_vma *anon_vma = vma->anon_vma; | |
433abed6 | 1732 | |
0fe6e20b | 1733 | BUG_ON(!anon_vma); |
433abed6 NH |
1734 | |
1735 | if (PageAnon(page)) | |
1736 | return; | |
1737 | if (!exclusive) | |
1738 | anon_vma = anon_vma->root; | |
1739 | ||
0fe6e20b NH |
1740 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; |
1741 | page->mapping = (struct address_space *) anon_vma; | |
1742 | page->index = linear_page_index(vma, address); | |
1743 | } | |
1744 | ||
1745 | void hugepage_add_anon_rmap(struct page *page, | |
1746 | struct vm_area_struct *vma, unsigned long address) | |
1747 | { | |
1748 | struct anon_vma *anon_vma = vma->anon_vma; | |
1749 | int first; | |
a850ea30 NH |
1750 | |
1751 | BUG_ON(!PageLocked(page)); | |
0fe6e20b | 1752 | BUG_ON(!anon_vma); |
5dbe0af4 | 1753 | /* address might be in next vma when migration races vma_adjust */ |
0fe6e20b NH |
1754 | first = atomic_inc_and_test(&page->_mapcount); |
1755 | if (first) | |
1756 | __hugepage_set_anon_rmap(page, vma, address, 0); | |
1757 | } | |
1758 | ||
1759 | void hugepage_add_new_anon_rmap(struct page *page, | |
1760 | struct vm_area_struct *vma, unsigned long address) | |
1761 | { | |
1762 | BUG_ON(address < vma->vm_start || address >= vma->vm_end); | |
1763 | atomic_set(&page->_mapcount, 0); | |
1764 | __hugepage_set_anon_rmap(page, vma, address, 1); | |
1765 | } | |
e3390f67 | 1766 | #endif /* CONFIG_HUGETLB_PAGE */ |