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 | * mm->mmap_sem |
25 | * page->flags PG_locked (lock_page) | |
3d48ae45 | 26 | * mapping->i_mmap_mutex |
2b575eb6 | 27 | * anon_vma->mutex |
82591e6e NP |
28 | * 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) | |
250df6ed | 33 | * inode->i_lock (in set_page_dirty's __mark_inode_dirty) |
f758eeab | 34 | * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty) |
82591e6e NP |
35 | * 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, | |
f758eeab | 38 | * within bdi.wb->list_lock in __sync_single_inode) |
6a46079c | 39 | * |
9b679320 PZ |
40 | * anon_vma->mutex,mapping->i_mutex (memory_failure, collect_procs_anon) |
41 | * ->tasklist_lock | |
6a46079c | 42 | * pte map lock |
1da177e4 LT |
43 | */ |
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> | |
5ad64688 | 51 | #include <linux/ksm.h> |
1da177e4 LT |
52 | #include <linux/rmap.h> |
53 | #include <linux/rcupdate.h> | |
b95f1b31 | 54 | #include <linux/export.h> |
8a9f3ccd | 55 | #include <linux/memcontrol.h> |
cddb8a5c | 56 | #include <linux/mmu_notifier.h> |
64cdd548 | 57 | #include <linux/migrate.h> |
0fe6e20b | 58 | #include <linux/hugetlb.h> |
1da177e4 LT |
59 | |
60 | #include <asm/tlbflush.h> | |
61 | ||
b291f000 NP |
62 | #include "internal.h" |
63 | ||
fdd2e5f8 | 64 | static struct kmem_cache *anon_vma_cachep; |
5beb4930 | 65 | static struct kmem_cache *anon_vma_chain_cachep; |
fdd2e5f8 AB |
66 | |
67 | static inline struct anon_vma *anon_vma_alloc(void) | |
68 | { | |
01d8b20d PZ |
69 | 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; | |
fdd2e5f8 AB |
82 | } |
83 | ||
01d8b20d | 84 | static inline void anon_vma_free(struct anon_vma *anon_vma) |
fdd2e5f8 | 85 | { |
01d8b20d | 86 | VM_BUG_ON(atomic_read(&anon_vma->refcount)); |
88c22088 PZ |
87 | |
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 | ||
fdd2e5f8 AB |
110 | kmem_cache_free(anon_vma_cachep, anon_vma); |
111 | } | |
1da177e4 | 112 | |
dd34739c | 113 | static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp) |
5beb4930 | 114 | { |
dd34739c | 115 | return kmem_cache_alloc(anon_vma_chain_cachep, gfp); |
5beb4930 RR |
116 | } |
117 | ||
e574b5fd | 118 | static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) |
5beb4930 RR |
119 | { |
120 | kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); | |
121 | } | |
122 | ||
d9d332e0 LT |
123 | /** |
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 | |
23a0790a | 132 | * not we either need to find an adjacent mapping that we |
d9d332e0 LT |
133 | * 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 | */ | |
1da177e4 LT |
150 | int anon_vma_prepare(struct vm_area_struct *vma) |
151 | { | |
152 | struct anon_vma *anon_vma = vma->anon_vma; | |
5beb4930 | 153 | struct anon_vma_chain *avc; |
1da177e4 LT |
154 | |
155 | might_sleep(); | |
156 | if (unlikely(!anon_vma)) { | |
157 | struct mm_struct *mm = vma->vm_mm; | |
d9d332e0 | 158 | struct anon_vma *allocated; |
1da177e4 | 159 | |
dd34739c | 160 | avc = anon_vma_chain_alloc(GFP_KERNEL); |
5beb4930 RR |
161 | if (!avc) |
162 | goto out_enomem; | |
163 | ||
1da177e4 | 164 | anon_vma = find_mergeable_anon_vma(vma); |
d9d332e0 LT |
165 | allocated = NULL; |
166 | if (!anon_vma) { | |
1da177e4 LT |
167 | anon_vma = anon_vma_alloc(); |
168 | if (unlikely(!anon_vma)) | |
5beb4930 | 169 | goto out_enomem_free_avc; |
1da177e4 | 170 | allocated = anon_vma; |
1da177e4 LT |
171 | } |
172 | ||
cba48b98 | 173 | anon_vma_lock(anon_vma); |
1da177e4 LT |
174 | /* 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; | |
5beb4930 RR |
178 | avc->anon_vma = anon_vma; |
179 | avc->vma = vma; | |
180 | list_add(&avc->same_vma, &vma->anon_vma_chain); | |
26ba0cb6 | 181 | list_add_tail(&avc->same_anon_vma, &anon_vma->head); |
1da177e4 | 182 | allocated = NULL; |
31f2b0eb | 183 | avc = NULL; |
1da177e4 LT |
184 | } |
185 | spin_unlock(&mm->page_table_lock); | |
cba48b98 | 186 | anon_vma_unlock(anon_vma); |
31f2b0eb ON |
187 | |
188 | if (unlikely(allocated)) | |
01d8b20d | 189 | put_anon_vma(allocated); |
31f2b0eb | 190 | if (unlikely(avc)) |
5beb4930 | 191 | anon_vma_chain_free(avc); |
1da177e4 LT |
192 | } |
193 | return 0; | |
5beb4930 RR |
194 | |
195 | out_enomem_free_avc: | |
196 | anon_vma_chain_free(avc); | |
197 | out_enomem: | |
198 | return -ENOMEM; | |
1da177e4 LT |
199 | } |
200 | ||
bb4aa396 LT |
201 | /* |
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 | */ | |
209 | static 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 | ||
221 | static inline void unlock_anon_vma_root(struct anon_vma *root) | |
222 | { | |
223 | if (root) | |
224 | mutex_unlock(&root->mutex); | |
225 | } | |
226 | ||
5beb4930 RR |
227 | static void anon_vma_chain_link(struct vm_area_struct *vma, |
228 | struct anon_vma_chain *avc, | |
229 | struct anon_vma *anon_vma) | |
1da177e4 | 230 | { |
5beb4930 RR |
231 | avc->vma = vma; |
232 | avc->anon_vma = anon_vma; | |
233 | list_add(&avc->same_vma, &vma->anon_vma_chain); | |
234 | ||
05759d38 AA |
235 | /* |
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 | */ | |
5beb4930 | 239 | list_add_tail(&avc->same_anon_vma, &anon_vma->head); |
1da177e4 LT |
240 | } |
241 | ||
5beb4930 RR |
242 | /* |
243 | * Attach the anon_vmas from src to dst. | |
244 | * Returns 0 on success, -ENOMEM on failure. | |
245 | */ | |
246 | int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) | |
1da177e4 | 247 | { |
5beb4930 | 248 | struct anon_vma_chain *avc, *pavc; |
bb4aa396 | 249 | struct anon_vma *root = NULL; |
5beb4930 | 250 | |
646d87b4 | 251 | list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { |
bb4aa396 LT |
252 | struct anon_vma *anon_vma; |
253 | ||
dd34739c LT |
254 | 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 | } | |
bb4aa396 LT |
262 | anon_vma = pavc->anon_vma; |
263 | root = lock_anon_vma_root(root, anon_vma); | |
264 | anon_vma_chain_link(dst, avc, anon_vma); | |
5beb4930 | 265 | } |
bb4aa396 | 266 | unlock_anon_vma_root(root); |
5beb4930 | 267 | return 0; |
1da177e4 | 268 | |
5beb4930 RR |
269 | enomem_failure: |
270 | unlink_anon_vmas(dst); | |
271 | return -ENOMEM; | |
1da177e4 LT |
272 | } |
273 | ||
948f017b AA |
274 | /* |
275 | * 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 | */ | |
304 | void 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 | ||
5beb4930 RR |
319 | /* |
320 | * 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 | */ | |
324 | int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) | |
1da177e4 | 325 | { |
5beb4930 RR |
326 | struct anon_vma_chain *avc; |
327 | struct anon_vma *anon_vma; | |
1da177e4 | 328 | |
5beb4930 RR |
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; | |
dd34739c | 344 | avc = anon_vma_chain_alloc(GFP_KERNEL); |
5beb4930 RR |
345 | if (!avc) |
346 | goto out_error_free_anon_vma; | |
5c341ee1 RR |
347 | |
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; | |
76545066 | 353 | /* |
01d8b20d PZ |
354 | * 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. | |
76545066 RR |
357 | */ |
358 | get_anon_vma(anon_vma->root); | |
5beb4930 RR |
359 | /* Mark this anon_vma as the one where our new (COWed) pages go. */ |
360 | vma->anon_vma = anon_vma; | |
bb4aa396 | 361 | anon_vma_lock(anon_vma); |
5c341ee1 | 362 | anon_vma_chain_link(vma, avc, anon_vma); |
bb4aa396 | 363 | anon_vma_unlock(anon_vma); |
5beb4930 RR |
364 | |
365 | return 0; | |
366 | ||
367 | out_error_free_anon_vma: | |
01d8b20d | 368 | put_anon_vma(anon_vma); |
5beb4930 | 369 | out_error: |
4946d54c | 370 | unlink_anon_vmas(vma); |
5beb4930 | 371 | return -ENOMEM; |
1da177e4 LT |
372 | } |
373 | ||
5beb4930 RR |
374 | void unlink_anon_vmas(struct vm_area_struct *vma) |
375 | { | |
376 | struct anon_vma_chain *avc, *next; | |
eee2acba | 377 | struct anon_vma *root = NULL; |
5beb4930 | 378 | |
5c341ee1 RR |
379 | /* |
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 | */ | |
5beb4930 | 383 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { |
eee2acba PZ |
384 | 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 | ||
5beb4930 RR |
411 | list_del(&avc->same_vma); |
412 | anon_vma_chain_free(avc); | |
413 | } | |
414 | } | |
415 | ||
51cc5068 | 416 | static void anon_vma_ctor(void *data) |
1da177e4 | 417 | { |
a35afb83 | 418 | struct anon_vma *anon_vma = data; |
1da177e4 | 419 | |
2b575eb6 | 420 | mutex_init(&anon_vma->mutex); |
83813267 | 421 | atomic_set(&anon_vma->refcount, 0); |
a35afb83 | 422 | INIT_LIST_HEAD(&anon_vma->head); |
1da177e4 LT |
423 | } |
424 | ||
425 | void __init anon_vma_init(void) | |
426 | { | |
427 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
20c2df83 | 428 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor); |
5beb4930 | 429 | anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC); |
1da177e4 LT |
430 | } |
431 | ||
432 | /* | |
6111e4ca PZ |
433 | * 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 | * | |
bc658c96 PZ |
442 | * 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 | * | |
6111e4ca PZ |
447 | * 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. | |
1da177e4 | 454 | */ |
746b18d4 | 455 | struct anon_vma *page_get_anon_vma(struct page *page) |
1da177e4 | 456 | { |
746b18d4 | 457 | struct anon_vma *anon_vma = NULL; |
1da177e4 LT |
458 | unsigned long anon_mapping; |
459 | ||
460 | rcu_read_lock(); | |
80e14822 | 461 | anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping); |
3ca7b3c5 | 462 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) |
1da177e4 LT |
463 | goto out; |
464 | if (!page_mapped(page)) | |
465 | goto out; | |
466 | ||
467 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
746b18d4 PZ |
468 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { |
469 | anon_vma = NULL; | |
470 | goto out; | |
471 | } | |
f1819427 HD |
472 | |
473 | /* | |
474 | * If this page is still mapped, then its anon_vma cannot have been | |
746b18d4 PZ |
475 | * 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). | |
f1819427 | 479 | */ |
746b18d4 PZ |
480 | if (!page_mapped(page)) { |
481 | put_anon_vma(anon_vma); | |
482 | anon_vma = NULL; | |
483 | } | |
1da177e4 LT |
484 | out: |
485 | rcu_read_unlock(); | |
746b18d4 PZ |
486 | |
487 | return anon_vma; | |
488 | } | |
489 | ||
88c22088 PZ |
490 | /* |
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 | */ | |
746b18d4 PZ |
497 | struct anon_vma *page_lock_anon_vma(struct page *page) |
498 | { | |
88c22088 | 499 | struct anon_vma *anon_vma = NULL; |
eee0f252 | 500 | struct anon_vma *root_anon_vma; |
88c22088 | 501 | unsigned long anon_mapping; |
746b18d4 | 502 | |
88c22088 PZ |
503 | 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; | |
509 | ||
510 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
eee0f252 HD |
511 | root_anon_vma = ACCESS_ONCE(anon_vma->root); |
512 | if (mutex_trylock(&root_anon_vma->mutex)) { | |
88c22088 | 513 | /* |
eee0f252 HD |
514 | * If the page is still mapped, then this anon_vma is still |
515 | * its anon_vma, and holding the mutex ensures that it will | |
bc658c96 | 516 | * not go away, see anon_vma_free(). |
88c22088 | 517 | */ |
eee0f252 HD |
518 | if (!page_mapped(page)) { |
519 | mutex_unlock(&root_anon_vma->mutex); | |
88c22088 PZ |
520 | anon_vma = NULL; |
521 | } | |
522 | goto out; | |
523 | } | |
746b18d4 | 524 | |
88c22088 PZ |
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 | ||
554 | out: | |
555 | rcu_read_unlock(); | |
746b18d4 | 556 | return anon_vma; |
34bbd704 ON |
557 | } |
558 | ||
10be22df | 559 | void page_unlock_anon_vma(struct anon_vma *anon_vma) |
34bbd704 | 560 | { |
cba48b98 | 561 | anon_vma_unlock(anon_vma); |
1da177e4 LT |
562 | } |
563 | ||
564 | /* | |
3ad33b24 LS |
565 | * 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. | |
1da177e4 | 568 | */ |
71e3aac0 | 569 | inline unsigned long |
1da177e4 LT |
570 | vma_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 | ||
0fe6e20b NH |
575 | if (unlikely(is_vm_hugetlb_page(vma))) |
576 | pgoff = page->index << huge_page_order(page_hstate(page)); | |
1da177e4 LT |
577 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); |
578 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { | |
3ad33b24 | 579 | /* page should be within @vma mapping range */ |
1da177e4 LT |
580 | return -EFAULT; |
581 | } | |
582 | return address; | |
583 | } | |
584 | ||
585 | /* | |
bf89c8c8 | 586 | * At what user virtual address is page expected in vma? |
ab941e0f | 587 | * Caller should check the page is actually part of the vma. |
1da177e4 LT |
588 | */ |
589 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
590 | { | |
21d0d443 | 591 | if (PageAnon(page)) { |
4829b906 HD |
592 | 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) | |
21d0d443 AA |
599 | return -EFAULT; |
600 | } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { | |
ee498ed7 HD |
601 | if (!vma->vm_file || |
602 | vma->vm_file->f_mapping != page->mapping) | |
1da177e4 LT |
603 | return -EFAULT; |
604 | } else | |
605 | return -EFAULT; | |
606 | return vma_address(page, vma); | |
607 | } | |
608 | ||
81b4082d ND |
609 | /* |
610 | * Check that @page is mapped at @address into @mm. | |
611 | * | |
479db0bf NP |
612 | * 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 | * | |
b8072f09 | 616 | * On success returns with pte mapped and locked. |
81b4082d | 617 | */ |
e9a81a82 | 618 | pte_t *__page_check_address(struct page *page, struct mm_struct *mm, |
479db0bf | 619 | unsigned long address, spinlock_t **ptlp, int sync) |
81b4082d ND |
620 | { |
621 | pgd_t *pgd; | |
622 | pud_t *pud; | |
623 | pmd_t *pmd; | |
624 | pte_t *pte; | |
c0718806 | 625 | spinlock_t *ptl; |
81b4082d | 626 | |
0fe6e20b NH |
627 | if (unlikely(PageHuge(page))) { |
628 | pte = huge_pte_offset(mm, address); | |
629 | ptl = &mm->page_table_lock; | |
630 | goto check; | |
631 | } | |
632 | ||
81b4082d | 633 | pgd = pgd_offset(mm, address); |
c0718806 HD |
634 | 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; | |
71e3aac0 AA |
644 | if (pmd_trans_huge(*pmd)) |
645 | return NULL; | |
c0718806 HD |
646 | |
647 | pte = pte_offset_map(pmd, address); | |
648 | /* Make a quick check before getting the lock */ | |
479db0bf | 649 | if (!sync && !pte_present(*pte)) { |
c0718806 HD |
650 | pte_unmap(pte); |
651 | return NULL; | |
652 | } | |
653 | ||
4c21e2f2 | 654 | ptl = pte_lockptr(mm, pmd); |
0fe6e20b | 655 | check: |
c0718806 HD |
656 | spin_lock(ptl); |
657 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { | |
658 | *ptlp = ptl; | |
659 | return pte; | |
81b4082d | 660 | } |
c0718806 HD |
661 | pte_unmap_unlock(pte, ptl); |
662 | return NULL; | |
81b4082d ND |
663 | } |
664 | ||
b291f000 NP |
665 | /** |
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 | */ | |
6a46079c | 674 | int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma) |
b291f000 NP |
675 | { |
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 | ||
1da177e4 LT |
691 | /* |
692 | * Subfunctions of page_referenced: page_referenced_one called | |
693 | * repeatedly from either page_referenced_anon or page_referenced_file. | |
694 | */ | |
5ad64688 HD |
695 | int page_referenced_one(struct page *page, struct vm_area_struct *vma, |
696 | unsigned long address, unsigned int *mapcount, | |
697 | unsigned long *vm_flags) | |
1da177e4 LT |
698 | { |
699 | struct mm_struct *mm = vma->vm_mm; | |
1da177e4 LT |
700 | int referenced = 0; |
701 | ||
71e3aac0 AA |
702 | if (unlikely(PageTransHuge(page))) { |
703 | pmd_t *pmd; | |
704 | ||
705 | spin_lock(&mm->page_table_lock); | |
2da28bfd AA |
706 | /* |
707 | * rmap might return false positives; we must filter | |
708 | * these out using page_check_address_pmd(). | |
709 | */ | |
71e3aac0 AA |
710 | pmd = page_check_address_pmd(page, mm, address, |
711 | PAGE_CHECK_ADDRESS_PMD_FLAG); | |
2da28bfd AA |
712 | 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)) | |
71e3aac0 AA |
726 | referenced++; |
727 | spin_unlock(&mm->page_table_lock); | |
728 | } else { | |
729 | pte_t *pte; | |
730 | spinlock_t *ptl; | |
731 | ||
2da28bfd AA |
732 | /* |
733 | * rmap might return false positives; we must filter | |
734 | * these out using page_check_address(). | |
735 | */ | |
71e3aac0 AA |
736 | pte = page_check_address(page, mm, address, &ptl, 0); |
737 | if (!pte) | |
738 | goto out; | |
739 | ||
2da28bfd AA |
740 | 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 | ||
71e3aac0 AA |
747 | 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 | ||
2da28bfd AA |
761 | /* 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 | ||
c0718806 | 767 | (*mapcount)--; |
273f047e | 768 | |
6fe6b7e3 WF |
769 | if (referenced) |
770 | *vm_flags |= vma->vm_flags; | |
273f047e | 771 | out: |
1da177e4 LT |
772 | return referenced; |
773 | } | |
774 | ||
bed7161a | 775 | static int page_referenced_anon(struct page *page, |
72835c86 | 776 | struct mem_cgroup *memcg, |
6fe6b7e3 | 777 | unsigned long *vm_flags) |
1da177e4 LT |
778 | { |
779 | unsigned int mapcount; | |
780 | struct anon_vma *anon_vma; | |
5beb4930 | 781 | struct anon_vma_chain *avc; |
1da177e4 LT |
782 | 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); | |
5beb4930 RR |
789 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { |
790 | struct vm_area_struct *vma = avc->vma; | |
1cb1729b HD |
791 | unsigned long address = vma_address(page, vma); |
792 | if (address == -EFAULT) | |
793 | continue; | |
bed7161a BS |
794 | /* |
795 | * If we are reclaiming on behalf of a cgroup, skip | |
796 | * counting on behalf of references from different | |
797 | * cgroups | |
798 | */ | |
72835c86 | 799 | if (memcg && !mm_match_cgroup(vma->vm_mm, memcg)) |
bed7161a | 800 | continue; |
1cb1729b | 801 | referenced += page_referenced_one(page, vma, address, |
6fe6b7e3 | 802 | &mapcount, vm_flags); |
1da177e4 LT |
803 | if (!mapcount) |
804 | break; | |
805 | } | |
34bbd704 ON |
806 | |
807 | page_unlock_anon_vma(anon_vma); | |
1da177e4 LT |
808 | return referenced; |
809 | } | |
810 | ||
811 | /** | |
812 | * page_referenced_file - referenced check for object-based rmap | |
813 | * @page: the page we're checking references on. | |
72835c86 | 814 | * @memcg: target memory control group |
6fe6b7e3 | 815 | * @vm_flags: collect encountered vma->vm_flags who actually referenced the page |
1da177e4 LT |
816 | * |
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 | */ | |
bed7161a | 824 | static int page_referenced_file(struct page *page, |
72835c86 | 825 | struct mem_cgroup *memcg, |
6fe6b7e3 | 826 | unsigned long *vm_flags) |
1da177e4 LT |
827 | { |
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, | |
3d48ae45 | 846 | * so we can safely take mapping->i_mmap_mutex. |
1da177e4 LT |
847 | */ |
848 | BUG_ON(!PageLocked(page)); | |
849 | ||
3d48ae45 | 850 | mutex_lock(&mapping->i_mmap_mutex); |
1da177e4 LT |
851 | |
852 | /* | |
3d48ae45 | 853 | * i_mmap_mutex does not stabilize mapcount at all, but mapcount |
1da177e4 LT |
854 | * 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) { | |
1cb1729b HD |
859 | unsigned long address = vma_address(page, vma); |
860 | if (address == -EFAULT) | |
861 | continue; | |
bed7161a BS |
862 | /* |
863 | * If we are reclaiming on behalf of a cgroup, skip | |
864 | * counting on behalf of references from different | |
865 | * cgroups | |
866 | */ | |
72835c86 | 867 | if (memcg && !mm_match_cgroup(vma->vm_mm, memcg)) |
bed7161a | 868 | continue; |
1cb1729b | 869 | referenced += page_referenced_one(page, vma, address, |
6fe6b7e3 | 870 | &mapcount, vm_flags); |
1da177e4 LT |
871 | if (!mapcount) |
872 | break; | |
873 | } | |
874 | ||
3d48ae45 | 875 | mutex_unlock(&mapping->i_mmap_mutex); |
1da177e4 LT |
876 | 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 | |
72835c86 | 883 | * @memcg: target memory cgroup |
6fe6b7e3 | 884 | * @vm_flags: collect encountered vma->vm_flags who actually referenced the page |
1da177e4 LT |
885 | * |
886 | * Quick test_and_clear_referenced for all mappings to a page, | |
887 | * returns the number of ptes which referenced the page. | |
888 | */ | |
6fe6b7e3 WF |
889 | int page_referenced(struct page *page, |
890 | int is_locked, | |
72835c86 | 891 | struct mem_cgroup *memcg, |
6fe6b7e3 | 892 | unsigned long *vm_flags) |
1da177e4 LT |
893 | { |
894 | int referenced = 0; | |
5ad64688 | 895 | int we_locked = 0; |
1da177e4 | 896 | |
6fe6b7e3 | 897 | *vm_flags = 0; |
3ca7b3c5 | 898 | if (page_mapped(page) && page_rmapping(page)) { |
5ad64688 HD |
899 | 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))) | |
72835c86 | 907 | referenced += page_referenced_ksm(page, memcg, |
5ad64688 HD |
908 | vm_flags); |
909 | else if (PageAnon(page)) | |
72835c86 | 910 | referenced += page_referenced_anon(page, memcg, |
6fe6b7e3 | 911 | vm_flags); |
5ad64688 | 912 | else if (page->mapping) |
72835c86 | 913 | referenced += page_referenced_file(page, memcg, |
6fe6b7e3 | 914 | vm_flags); |
5ad64688 | 915 | if (we_locked) |
1da177e4 | 916 | unlock_page(page); |
50a15981 MS |
917 | |
918 | if (page_test_and_clear_young(page_to_pfn(page))) | |
919 | referenced++; | |
1da177e4 | 920 | } |
5ad64688 | 921 | out: |
1da177e4 LT |
922 | return referenced; |
923 | } | |
924 | ||
1cb1729b HD |
925 | static int page_mkclean_one(struct page *page, struct vm_area_struct *vma, |
926 | unsigned long address) | |
d08b3851 PZ |
927 | { |
928 | struct mm_struct *mm = vma->vm_mm; | |
c2fda5fe | 929 | pte_t *pte; |
d08b3851 PZ |
930 | spinlock_t *ptl; |
931 | int ret = 0; | |
932 | ||
479db0bf | 933 | pte = page_check_address(page, mm, address, &ptl, 1); |
d08b3851 PZ |
934 | if (!pte) |
935 | goto out; | |
936 | ||
c2fda5fe PZ |
937 | if (pte_dirty(*pte) || pte_write(*pte)) { |
938 | pte_t entry; | |
d08b3851 | 939 | |
c2fda5fe | 940 | flush_cache_page(vma, address, pte_pfn(*pte)); |
cddb8a5c | 941 | entry = ptep_clear_flush_notify(vma, address, pte); |
c2fda5fe PZ |
942 | entry = pte_wrprotect(entry); |
943 | entry = pte_mkclean(entry); | |
d6e88e67 | 944 | set_pte_at(mm, address, pte, entry); |
c2fda5fe PZ |
945 | ret = 1; |
946 | } | |
d08b3851 | 947 | |
d08b3851 PZ |
948 | pte_unmap_unlock(pte, ptl); |
949 | out: | |
950 | return ret; | |
951 | } | |
952 | ||
953 | static 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 | ||
3d48ae45 | 962 | mutex_lock(&mapping->i_mmap_mutex); |
d08b3851 | 963 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { |
1cb1729b HD |
964 | 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 | } | |
d08b3851 | 970 | } |
3d48ae45 | 971 | mutex_unlock(&mapping->i_mmap_mutex); |
d08b3851 PZ |
972 | return ret; |
973 | } | |
974 | ||
975 | int 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); | |
ce7e9fae | 983 | if (mapping) { |
d08b3851 | 984 | ret = page_mkclean_file(mapping, page); |
2d42552d | 985 | if (page_test_and_clear_dirty(page_to_pfn(page), 1)) |
ce7e9fae | 986 | ret = 1; |
6c210482 | 987 | } |
d08b3851 PZ |
988 | } |
989 | ||
990 | return ret; | |
991 | } | |
60b59bea | 992 | EXPORT_SYMBOL_GPL(page_mkclean); |
d08b3851 | 993 | |
c44b6743 RR |
994 | /** |
995 | * 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 | */ | |
1005 | void 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 | ||
9617d95e | 1018 | /** |
4e1c1975 AK |
1019 | * __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 | |
e8a03feb | 1023 | * @exclusive: the page is exclusively owned by the current process |
9617d95e NP |
1024 | */ |
1025 | static void __page_set_anon_rmap(struct page *page, | |
e8a03feb | 1026 | struct vm_area_struct *vma, unsigned long address, int exclusive) |
9617d95e | 1027 | { |
e8a03feb | 1028 | struct anon_vma *anon_vma = vma->anon_vma; |
ea90002b | 1029 | |
e8a03feb | 1030 | BUG_ON(!anon_vma); |
ea90002b | 1031 | |
4e1c1975 AK |
1032 | if (PageAnon(page)) |
1033 | return; | |
1034 | ||
ea90002b | 1035 | /* |
e8a03feb RR |
1036 | * If the page isn't exclusively mapped into this vma, |
1037 | * we must use the _oldest_ possible anon_vma for the | |
1038 | * page mapping! | |
ea90002b | 1039 | */ |
4e1c1975 | 1040 | if (!exclusive) |
288468c3 | 1041 | anon_vma = anon_vma->root; |
9617d95e | 1042 | |
9617d95e NP |
1043 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; |
1044 | page->mapping = (struct address_space *) anon_vma; | |
9617d95e | 1045 | page->index = linear_page_index(vma, address); |
9617d95e NP |
1046 | } |
1047 | ||
c97a9e10 | 1048 | /** |
43d8eac4 | 1049 | * __page_check_anon_rmap - sanity check anonymous rmap addition |
c97a9e10 NP |
1050 | * @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 | */ | |
1054 | static 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 | */ | |
44ab57a0 | 1070 | BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root); |
c97a9e10 NP |
1071 | BUG_ON(page->index != linear_page_index(vma, address)); |
1072 | #endif | |
1073 | } | |
1074 | ||
1da177e4 LT |
1075 | /** |
1076 | * 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 | * | |
5ad64688 | 1081 | * The caller needs to hold the pte lock, and the page must be locked in |
80e14822 HD |
1082 | * 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). | |
1da177e4 LT |
1085 | */ |
1086 | void page_add_anon_rmap(struct page *page, | |
1087 | struct vm_area_struct *vma, unsigned long address) | |
ad8c2ee8 RR |
1088 | { |
1089 | 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 | */ | |
1097 | void do_page_add_anon_rmap(struct page *page, | |
1098 | struct vm_area_struct *vma, unsigned long address, int exclusive) | |
1da177e4 | 1099 | { |
5ad64688 | 1100 | int first = atomic_inc_and_test(&page->_mapcount); |
79134171 AA |
1101 | 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 | } | |
5ad64688 HD |
1108 | if (unlikely(PageKsm(page))) |
1109 | return; | |
1110 | ||
c97a9e10 | 1111 | VM_BUG_ON(!PageLocked(page)); |
5dbe0af4 | 1112 | /* address might be in next vma when migration races vma_adjust */ |
5ad64688 | 1113 | if (first) |
ad8c2ee8 | 1114 | __page_set_anon_rmap(page, vma, address, exclusive); |
69029cd5 | 1115 | else |
c97a9e10 | 1116 | __page_check_anon_rmap(page, vma, address); |
1da177e4 LT |
1117 | } |
1118 | ||
43d8eac4 | 1119 | /** |
9617d95e NP |
1120 | * 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. | |
c97a9e10 | 1127 | * Page does not have to be locked. |
9617d95e NP |
1128 | */ |
1129 | void page_add_new_anon_rmap(struct page *page, | |
1130 | struct vm_area_struct *vma, unsigned long address) | |
1131 | { | |
b5934c53 | 1132 | VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); |
cbf84b7a HD |
1133 | SetPageSwapBacked(page); |
1134 | atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */ | |
79134171 AA |
1135 | if (!PageTransHuge(page)) |
1136 | __inc_zone_page_state(page, NR_ANON_PAGES); | |
1137 | else | |
1138 | __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES); | |
e8a03feb | 1139 | __page_set_anon_rmap(page, vma, address, 1); |
b5934c53 | 1140 | if (page_evictable(page, vma)) |
cbf84b7a | 1141 | lru_cache_add_lru(page, LRU_ACTIVE_ANON); |
b5934c53 HD |
1142 | else |
1143 | add_page_to_unevictable_list(page); | |
9617d95e NP |
1144 | } |
1145 | ||
1da177e4 LT |
1146 | /** |
1147 | * page_add_file_rmap - add pte mapping to a file page | |
1148 | * @page: the page to add the mapping to | |
1149 | * | |
b8072f09 | 1150 | * The caller needs to hold the pte lock. |
1da177e4 LT |
1151 | */ |
1152 | void page_add_file_rmap(struct page *page) | |
1153 | { | |
d69b042f | 1154 | if (atomic_inc_and_test(&page->_mapcount)) { |
65ba55f5 | 1155 | __inc_zone_page_state(page, NR_FILE_MAPPED); |
2a7106f2 | 1156 | mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED); |
d69b042f | 1157 | } |
1da177e4 LT |
1158 | } |
1159 | ||
1160 | /** | |
1161 | * page_remove_rmap - take down pte mapping from a page | |
1162 | * @page: page to remove mapping from | |
1163 | * | |
b8072f09 | 1164 | * The caller needs to hold the pte lock. |
1da177e4 | 1165 | */ |
edc315fd | 1166 | void page_remove_rmap(struct page *page) |
1da177e4 | 1167 | { |
b904dcfe KM |
1168 | /* 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 | */ | |
2d42552d MS |
1179 | if ((!PageAnon(page) || PageSwapCache(page)) && |
1180 | page_test_and_clear_dirty(page_to_pfn(page), 1)) | |
b904dcfe | 1181 | set_page_dirty(page); |
0fe6e20b NH |
1182 | /* |
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; | |
b904dcfe KM |
1188 | if (PageAnon(page)) { |
1189 | mem_cgroup_uncharge_page(page); | |
79134171 AA |
1190 | 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); | |
b904dcfe KM |
1195 | } else { |
1196 | __dec_zone_page_state(page, NR_FILE_MAPPED); | |
2a7106f2 | 1197 | mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED); |
b904dcfe | 1198 | } |
b904dcfe KM |
1199 | /* |
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 | */ | |
1da177e4 LT |
1208 | } |
1209 | ||
1210 | /* | |
1211 | * Subfunctions of try_to_unmap: try_to_unmap_one called | |
99ef0315 | 1212 | * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file. |
1da177e4 | 1213 | */ |
5ad64688 HD |
1214 | int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, |
1215 | unsigned long address, enum ttu_flags flags) | |
1da177e4 LT |
1216 | { |
1217 | struct mm_struct *mm = vma->vm_mm; | |
1da177e4 LT |
1218 | pte_t *pte; |
1219 | pte_t pteval; | |
c0718806 | 1220 | spinlock_t *ptl; |
1da177e4 LT |
1221 | int ret = SWAP_AGAIN; |
1222 | ||
479db0bf | 1223 | pte = page_check_address(page, mm, address, &ptl, 0); |
c0718806 | 1224 | if (!pte) |
81b4082d | 1225 | goto out; |
1da177e4 LT |
1226 | |
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 | */ | |
14fa31b8 | 1232 | if (!(flags & TTU_IGNORE_MLOCK)) { |
caed0f48 KM |
1233 | if (vma->vm_flags & VM_LOCKED) |
1234 | goto out_mlock; | |
1235 | ||
af8e3354 | 1236 | if (TTU_ACTION(flags) == TTU_MUNLOCK) |
53f79acb | 1237 | goto out_unmap; |
14fa31b8 AK |
1238 | } |
1239 | if (!(flags & TTU_IGNORE_ACCESS)) { | |
b291f000 NP |
1240 | if (ptep_clear_flush_young_notify(vma, address, pte)) { |
1241 | ret = SWAP_FAIL; | |
1242 | goto out_unmap; | |
1243 | } | |
1244 | } | |
1da177e4 | 1245 | |
1da177e4 LT |
1246 | /* Nuke the page table entry. */ |
1247 | flush_cache_page(vma, address, page_to_pfn(page)); | |
cddb8a5c | 1248 | pteval = ptep_clear_flush_notify(vma, address, pte); |
1da177e4 LT |
1249 | |
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 | ||
365e9c87 HD |
1254 | /* Update high watermark before we lower rss */ |
1255 | update_hiwater_rss(mm); | |
1256 | ||
888b9f7c AK |
1257 | if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) { |
1258 | if (PageAnon(page)) | |
d559db08 | 1259 | dec_mm_counter(mm, MM_ANONPAGES); |
888b9f7c | 1260 | else |
d559db08 | 1261 | dec_mm_counter(mm, MM_FILEPAGES); |
888b9f7c AK |
1262 | set_pte_at(mm, address, pte, |
1263 | swp_entry_to_pte(make_hwpoison_entry(page))); | |
1264 | } else if (PageAnon(page)) { | |
4c21e2f2 | 1265 | swp_entry_t entry = { .val = page_private(page) }; |
0697212a CL |
1266 | |
1267 | if (PageSwapCache(page)) { | |
1268 | /* | |
1269 | * Store the swap location in the pte. | |
1270 | * See handle_pte_fault() ... | |
1271 | */ | |
570a335b HD |
1272 | if (swap_duplicate(entry) < 0) { |
1273 | set_pte_at(mm, address, pte, pteval); | |
1274 | ret = SWAP_FAIL; | |
1275 | goto out_unmap; | |
1276 | } | |
0697212a CL |
1277 | 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 | } | |
d559db08 | 1283 | dec_mm_counter(mm, MM_ANONPAGES); |
b084d435 | 1284 | inc_mm_counter(mm, MM_SWAPENTS); |
64cdd548 | 1285 | } else if (PAGE_MIGRATION) { |
0697212a CL |
1286 | /* |
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 | */ | |
14fa31b8 | 1291 | BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION); |
0697212a | 1292 | entry = make_migration_entry(page, pte_write(pteval)); |
1da177e4 LT |
1293 | } |
1294 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
1295 | BUG_ON(pte_file(*pte)); | |
14fa31b8 | 1296 | } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) { |
04e62a29 CL |
1297 | /* 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 | |
d559db08 | 1302 | dec_mm_counter(mm, MM_FILEPAGES); |
1da177e4 | 1303 | |
edc315fd | 1304 | page_remove_rmap(page); |
1da177e4 LT |
1305 | page_cache_release(page); |
1306 | ||
1307 | out_unmap: | |
c0718806 | 1308 | pte_unmap_unlock(pte, ptl); |
caed0f48 KM |
1309 | out: |
1310 | return ret; | |
53f79acb | 1311 | |
caed0f48 KM |
1312 | out_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 | |
2b575eb6 | 1319 | * we now hold anon_vma->mutex or mapping->i_mmap_mutex. |
caed0f48 KM |
1320 | * 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; | |
53f79acb | 1328 | } |
caed0f48 | 1329 | up_read(&vma->vm_mm->mmap_sem); |
53f79acb | 1330 | } |
1da177e4 LT |
1331 | return ret; |
1332 | } | |
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. | |
b291f000 NP |
1352 | * |
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. | |
1da177e4 LT |
1357 | */ |
1358 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) | |
1359 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) | |
1360 | ||
b291f000 NP |
1361 | static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount, |
1362 | struct vm_area_struct *vma, struct page *check_page) | |
1da177e4 LT |
1363 | { |
1364 | struct mm_struct *mm = vma->vm_mm; | |
1365 | pgd_t *pgd; | |
1366 | pud_t *pud; | |
1367 | pmd_t *pmd; | |
c0718806 | 1368 | pte_t *pte; |
1da177e4 | 1369 | pte_t pteval; |
c0718806 | 1370 | spinlock_t *ptl; |
1da177e4 LT |
1371 | struct page *page; |
1372 | unsigned long address; | |
1373 | unsigned long end; | |
b291f000 NP |
1374 | int ret = SWAP_AGAIN; |
1375 | int locked_vma = 0; | |
1da177e4 | 1376 | |
1da177e4 LT |
1377 | 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)) | |
b291f000 | 1386 | return ret; |
1da177e4 LT |
1387 | |
1388 | pud = pud_offset(pgd, address); | |
1389 | if (!pud_present(*pud)) | |
b291f000 | 1390 | return ret; |
1da177e4 LT |
1391 | |
1392 | pmd = pmd_offset(pud, address); | |
1393 | if (!pmd_present(*pmd)) | |
b291f000 NP |
1394 | return ret; |
1395 | ||
1396 | /* | |
af8e3354 | 1397 | * If we can acquire the mmap_sem for read, and vma is VM_LOCKED, |
b291f000 NP |
1398 | * keep the sem while scanning the cluster for mlocking pages. |
1399 | */ | |
af8e3354 | 1400 | if (down_read_trylock(&vma->vm_mm->mmap_sem)) { |
b291f000 NP |
1401 | locked_vma = (vma->vm_flags & VM_LOCKED); |
1402 | if (!locked_vma) | |
1403 | up_read(&vma->vm_mm->mmap_sem); /* don't need it */ | |
1404 | } | |
c0718806 HD |
1405 | |
1406 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
1da177e4 | 1407 | |
365e9c87 HD |
1408 | /* Update high watermark before we lower rss */ |
1409 | update_hiwater_rss(mm); | |
1410 | ||
c0718806 | 1411 | for (; address < end; pte++, address += PAGE_SIZE) { |
1da177e4 LT |
1412 | if (!pte_present(*pte)) |
1413 | continue; | |
6aab341e LT |
1414 | page = vm_normal_page(vma, address, *pte); |
1415 | BUG_ON(!page || PageAnon(page)); | |
1da177e4 | 1416 | |
b291f000 NP |
1417 | 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 | ||
cddb8a5c | 1424 | if (ptep_clear_flush_young_notify(vma, address, pte)) |
1da177e4 LT |
1425 | continue; |
1426 | ||
1427 | /* Nuke the page table entry. */ | |
eca35133 | 1428 | flush_cache_page(vma, address, pte_pfn(*pte)); |
cddb8a5c | 1429 | pteval = ptep_clear_flush_notify(vma, address, pte); |
1da177e4 LT |
1430 | |
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 | ||
edc315fd | 1439 | page_remove_rmap(page); |
1da177e4 | 1440 | page_cache_release(page); |
d559db08 | 1441 | dec_mm_counter(mm, MM_FILEPAGES); |
1da177e4 LT |
1442 | (*mapcount)--; |
1443 | } | |
c0718806 | 1444 | pte_unmap_unlock(pte - 1, ptl); |
b291f000 NP |
1445 | if (locked_vma) |
1446 | up_read(&vma->vm_mm->mmap_sem); | |
1447 | return ret; | |
1da177e4 LT |
1448 | } |
1449 | ||
71e3aac0 | 1450 | bool is_vma_temporary_stack(struct vm_area_struct *vma) |
a8bef8ff MG |
1451 | { |
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 | ||
b291f000 NP |
1464 | /** |
1465 | * try_to_unmap_anon - unmap or unlock anonymous page using the object-based | |
1466 | * rmap method | |
1467 | * @page: the page to unmap/unlock | |
8051be5e | 1468 | * @flags: action and flags |
b291f000 NP |
1469 | * |
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 | */ | |
14fa31b8 | 1480 | static int try_to_unmap_anon(struct page *page, enum ttu_flags flags) |
1da177e4 LT |
1481 | { |
1482 | struct anon_vma *anon_vma; | |
5beb4930 | 1483 | struct anon_vma_chain *avc; |
1da177e4 | 1484 | int ret = SWAP_AGAIN; |
b291f000 | 1485 | |
1da177e4 LT |
1486 | anon_vma = page_lock_anon_vma(page); |
1487 | if (!anon_vma) | |
1488 | return ret; | |
1489 | ||
5beb4930 RR |
1490 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { |
1491 | struct vm_area_struct *vma = avc->vma; | |
a8bef8ff MG |
1492 | 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); | |
1cb1729b HD |
1507 | if (address == -EFAULT) |
1508 | continue; | |
1509 | ret = try_to_unmap_one(page, vma, address, flags); | |
53f79acb HD |
1510 | if (ret != SWAP_AGAIN || !page_mapped(page)) |
1511 | break; | |
1da177e4 | 1512 | } |
34bbd704 ON |
1513 | |
1514 | page_unlock_anon_vma(anon_vma); | |
1da177e4 LT |
1515 | return ret; |
1516 | } | |
1517 | ||
1518 | /** | |
b291f000 NP |
1519 | * try_to_unmap_file - unmap/unlock file page using the object-based rmap method |
1520 | * @page: the page to unmap/unlock | |
14fa31b8 | 1521 | * @flags: action and flags |
1da177e4 LT |
1522 | * |
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 | * | |
b291f000 NP |
1526 | * 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. | |
1da177e4 | 1532 | */ |
14fa31b8 | 1533 | static int try_to_unmap_file(struct page *page, enum ttu_flags flags) |
1da177e4 LT |
1534 | { |
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 | ||
3d48ae45 | 1545 | mutex_lock(&mapping->i_mmap_mutex); |
1da177e4 | 1546 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { |
1cb1729b HD |
1547 | unsigned long address = vma_address(page, vma); |
1548 | if (address == -EFAULT) | |
1549 | continue; | |
1550 | ret = try_to_unmap_one(page, vma, address, flags); | |
53f79acb HD |
1551 | if (ret != SWAP_AGAIN || !page_mapped(page)) |
1552 | goto out; | |
1da177e4 LT |
1553 | } |
1554 | ||
1555 | if (list_empty(&mapping->i_mmap_nonlinear)) | |
1556 | goto out; | |
1557 | ||
53f79acb HD |
1558 | /* |
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 | ||
1da177e4 LT |
1566 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, |
1567 | shared.vm_set.list) { | |
1da177e4 LT |
1568 | 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 | ||
b291f000 | 1576 | if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */ |
1da177e4 LT |
1577 | 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; | |
3d48ae45 | 1591 | cond_resched(); |
1da177e4 LT |
1592 | |
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) { | |
1da177e4 | 1600 | cursor = (unsigned long) vma->vm_private_data; |
839b9685 | 1601 | while ( cursor < max_nl_cursor && |
1da177e4 | 1602 | cursor < vma->vm_end - vma->vm_start) { |
53f79acb HD |
1603 | if (try_to_unmap_cluster(cursor, &mapcount, |
1604 | vma, page) == SWAP_MLOCK) | |
1605 | ret = SWAP_MLOCK; | |
1da177e4 LT |
1606 | 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 | } | |
3d48ae45 | 1613 | cond_resched(); |
1da177e4 LT |
1614 | 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 | */ | |
101d2be7 HD |
1622 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) |
1623 | vma->vm_private_data = NULL; | |
1da177e4 | 1624 | out: |
3d48ae45 | 1625 | mutex_unlock(&mapping->i_mmap_mutex); |
1da177e4 LT |
1626 | 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 | |
14fa31b8 | 1632 | * @flags: action and flags |
1da177e4 LT |
1633 | * |
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 | |
b291f000 | 1641 | * SWAP_MLOCK - page is mlocked. |
1da177e4 | 1642 | */ |
14fa31b8 | 1643 | int try_to_unmap(struct page *page, enum ttu_flags flags) |
1da177e4 LT |
1644 | { |
1645 | int ret; | |
1646 | ||
1da177e4 | 1647 | BUG_ON(!PageLocked(page)); |
91600e9e | 1648 | VM_BUG_ON(!PageHuge(page) && PageTransHuge(page)); |
1da177e4 | 1649 | |
5ad64688 HD |
1650 | if (unlikely(PageKsm(page))) |
1651 | ret = try_to_unmap_ksm(page, flags); | |
1652 | else if (PageAnon(page)) | |
14fa31b8 | 1653 | ret = try_to_unmap_anon(page, flags); |
1da177e4 | 1654 | else |
14fa31b8 | 1655 | ret = try_to_unmap_file(page, flags); |
b291f000 | 1656 | if (ret != SWAP_MLOCK && !page_mapped(page)) |
1da177e4 LT |
1657 | ret = SWAP_SUCCESS; |
1658 | return ret; | |
1659 | } | |
81b4082d | 1660 | |
b291f000 NP |
1661 | /** |
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 | * | |
53f79acb | 1671 | * SWAP_AGAIN - no vma is holding page mlocked, or, |
b291f000 | 1672 | * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem |
5ad64688 | 1673 | * SWAP_FAIL - page cannot be located at present |
b291f000 NP |
1674 | * SWAP_MLOCK - page is now mlocked. |
1675 | */ | |
1676 | int try_to_munlock(struct page *page) | |
1677 | { | |
1678 | VM_BUG_ON(!PageLocked(page) || PageLRU(page)); | |
1679 | ||
5ad64688 HD |
1680 | if (unlikely(PageKsm(page))) |
1681 | return try_to_unmap_ksm(page, TTU_MUNLOCK); | |
1682 | else if (PageAnon(page)) | |
14fa31b8 | 1683 | return try_to_unmap_anon(page, TTU_MUNLOCK); |
b291f000 | 1684 | else |
14fa31b8 | 1685 | return try_to_unmap_file(page, TTU_MUNLOCK); |
b291f000 | 1686 | } |
e9995ef9 | 1687 | |
01d8b20d | 1688 | void __put_anon_vma(struct anon_vma *anon_vma) |
76545066 | 1689 | { |
01d8b20d | 1690 | struct anon_vma *root = anon_vma->root; |
76545066 | 1691 | |
01d8b20d PZ |
1692 | if (root != anon_vma && atomic_dec_and_test(&root->refcount)) |
1693 | anon_vma_free(root); | |
76545066 | 1694 | |
01d8b20d | 1695 | anon_vma_free(anon_vma); |
76545066 | 1696 | } |
76545066 | 1697 | |
e9995ef9 HD |
1698 | #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 | */ | |
1703 | static 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; | |
5beb4930 | 1707 | struct anon_vma_chain *avc; |
e9995ef9 HD |
1708 | 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 | |
3f6c8272 MG |
1713 | * are holding mmap_sem. Users without mmap_sem are required to |
1714 | * take a reference count to prevent the anon_vma disappearing | |
e9995ef9 HD |
1715 | */ |
1716 | anon_vma = page_anon_vma(page); | |
1717 | if (!anon_vma) | |
1718 | return ret; | |
cba48b98 | 1719 | anon_vma_lock(anon_vma); |
5beb4930 RR |
1720 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { |
1721 | struct vm_area_struct *vma = avc->vma; | |
e9995ef9 HD |
1722 | 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 | } | |
cba48b98 | 1729 | anon_vma_unlock(anon_vma); |
e9995ef9 HD |
1730 | return ret; |
1731 | } | |
1732 | ||
1733 | static 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; | |
3d48ae45 | 1744 | mutex_lock(&mapping->i_mmap_mutex); |
e9995ef9 HD |
1745 | 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 | */ | |
3d48ae45 | 1758 | mutex_unlock(&mapping->i_mmap_mutex); |
e9995ef9 HD |
1759 | return ret; |
1760 | } | |
1761 | ||
1762 | int 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 */ | |
0fe6e20b | 1775 | |
e3390f67 | 1776 | #ifdef CONFIG_HUGETLB_PAGE |
0fe6e20b NH |
1777 | /* |
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 | */ | |
1782 | static 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; | |
433abed6 | 1786 | |
0fe6e20b | 1787 | BUG_ON(!anon_vma); |
433abed6 NH |
1788 | |
1789 | if (PageAnon(page)) | |
1790 | return; | |
1791 | if (!exclusive) | |
1792 | anon_vma = anon_vma->root; | |
1793 | ||
0fe6e20b NH |
1794 | 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 | ||
1799 | void 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; | |
a850ea30 NH |
1804 | |
1805 | BUG_ON(!PageLocked(page)); | |
0fe6e20b | 1806 | BUG_ON(!anon_vma); |
5dbe0af4 | 1807 | /* address might be in next vma when migration races vma_adjust */ |
0fe6e20b NH |
1808 | first = atomic_inc_and_test(&page->_mapcount); |
1809 | if (first) | |
1810 | __hugepage_set_anon_rmap(page, vma, address, 0); | |
1811 | } | |
1812 | ||
1813 | void 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 | } | |
e3390f67 | 1820 | #endif /* CONFIG_HUGETLB_PAGE */ |