]>
Commit | Line | Data |
---|---|---|
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 | |
17 | * Contributions by Hugh Dickins 2003, 2004 | |
18 | */ | |
19 | ||
20 | /* | |
21 | * Lock ordering in mm: | |
22 | * | |
23 | * inode->i_mutex (while writing or truncating, not reading or faulting) | |
24 | * mm->mmap_sem | |
25 | * page->flags PG_locked (lock_page) | |
26 | * mapping->i_mmap_rwsem | |
27 | * anon_vma->rwsem | |
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) | |
33 | * inode->i_lock (in set_page_dirty's __mark_inode_dirty) | |
34 | * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty) | |
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, | |
38 | * within bdi.wb->list_lock in __sync_single_inode) | |
39 | * | |
40 | * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon) | |
41 | * ->tasklist_lock | |
42 | * pte map lock | |
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> | |
51 | #include <linux/ksm.h> | |
52 | #include <linux/rmap.h> | |
53 | #include <linux/rcupdate.h> | |
54 | #include <linux/export.h> | |
55 | #include <linux/memcontrol.h> | |
56 | #include <linux/mmu_notifier.h> | |
57 | #include <linux/migrate.h> | |
58 | #include <linux/hugetlb.h> | |
59 | #include <linux/backing-dev.h> | |
60 | ||
61 | #include <asm/tlbflush.h> | |
62 | ||
63 | #include "internal.h" | |
64 | ||
65 | static struct kmem_cache *anon_vma_cachep; | |
66 | static struct kmem_cache *anon_vma_chain_cachep; | |
67 | ||
68 | static inline struct anon_vma *anon_vma_alloc(void) | |
69 | { | |
70 | struct anon_vma *anon_vma; | |
71 | ||
72 | anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); | |
73 | if (anon_vma) { | |
74 | atomic_set(&anon_vma->refcount, 1); | |
75 | anon_vma->degree = 1; /* Reference for first vma */ | |
76 | anon_vma->parent = anon_vma; | |
77 | /* | |
78 | * Initialise the anon_vma root to point to itself. If called | |
79 | * from fork, the root will be reset to the parents anon_vma. | |
80 | */ | |
81 | anon_vma->root = anon_vma; | |
82 | } | |
83 | ||
84 | return anon_vma; | |
85 | } | |
86 | ||
87 | static inline void anon_vma_free(struct anon_vma *anon_vma) | |
88 | { | |
89 | VM_BUG_ON(atomic_read(&anon_vma->refcount)); | |
90 | ||
91 | /* | |
92 | * Synchronize against page_lock_anon_vma_read() such that | |
93 | * we can safely hold the lock without the anon_vma getting | |
94 | * freed. | |
95 | * | |
96 | * Relies on the full mb implied by the atomic_dec_and_test() from | |
97 | * put_anon_vma() against the acquire barrier implied by | |
98 | * down_read_trylock() from page_lock_anon_vma_read(). This orders: | |
99 | * | |
100 | * page_lock_anon_vma_read() VS put_anon_vma() | |
101 | * down_read_trylock() atomic_dec_and_test() | |
102 | * LOCK MB | |
103 | * atomic_read() rwsem_is_locked() | |
104 | * | |
105 | * LOCK should suffice since the actual taking of the lock must | |
106 | * happen _before_ what follows. | |
107 | */ | |
108 | might_sleep(); | |
109 | if (rwsem_is_locked(&anon_vma->root->rwsem)) { | |
110 | anon_vma_lock_write(anon_vma); | |
111 | anon_vma_unlock_write(anon_vma); | |
112 | } | |
113 | ||
114 | kmem_cache_free(anon_vma_cachep, anon_vma); | |
115 | } | |
116 | ||
117 | static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp) | |
118 | { | |
119 | return kmem_cache_alloc(anon_vma_chain_cachep, gfp); | |
120 | } | |
121 | ||
122 | static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) | |
123 | { | |
124 | kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); | |
125 | } | |
126 | ||
127 | static void anon_vma_chain_link(struct vm_area_struct *vma, | |
128 | struct anon_vma_chain *avc, | |
129 | struct anon_vma *anon_vma) | |
130 | { | |
131 | avc->vma = vma; | |
132 | avc->anon_vma = anon_vma; | |
133 | list_add(&avc->same_vma, &vma->anon_vma_chain); | |
134 | anon_vma_interval_tree_insert(avc, &anon_vma->rb_root); | |
135 | } | |
136 | ||
137 | /** | |
138 | * anon_vma_prepare - attach an anon_vma to a memory region | |
139 | * @vma: the memory region in question | |
140 | * | |
141 | * This makes sure the memory mapping described by 'vma' has | |
142 | * an 'anon_vma' attached to it, so that we can associate the | |
143 | * anonymous pages mapped into it with that anon_vma. | |
144 | * | |
145 | * The common case will be that we already have one, but if | |
146 | * not we either need to find an adjacent mapping that we | |
147 | * can re-use the anon_vma from (very common when the only | |
148 | * reason for splitting a vma has been mprotect()), or we | |
149 | * allocate a new one. | |
150 | * | |
151 | * Anon-vma allocations are very subtle, because we may have | |
152 | * optimistically looked up an anon_vma in page_lock_anon_vma_read() | |
153 | * and that may actually touch the spinlock even in the newly | |
154 | * allocated vma (it depends on RCU to make sure that the | |
155 | * anon_vma isn't actually destroyed). | |
156 | * | |
157 | * As a result, we need to do proper anon_vma locking even | |
158 | * for the new allocation. At the same time, we do not want | |
159 | * to do any locking for the common case of already having | |
160 | * an anon_vma. | |
161 | * | |
162 | * This must be called with the mmap_sem held for reading. | |
163 | */ | |
164 | int anon_vma_prepare(struct vm_area_struct *vma) | |
165 | { | |
166 | struct anon_vma *anon_vma = vma->anon_vma; | |
167 | struct anon_vma_chain *avc; | |
168 | ||
169 | might_sleep(); | |
170 | if (unlikely(!anon_vma)) { | |
171 | struct mm_struct *mm = vma->vm_mm; | |
172 | struct anon_vma *allocated; | |
173 | ||
174 | avc = anon_vma_chain_alloc(GFP_KERNEL); | |
175 | if (!avc) | |
176 | goto out_enomem; | |
177 | ||
178 | anon_vma = find_mergeable_anon_vma(vma); | |
179 | allocated = NULL; | |
180 | if (!anon_vma) { | |
181 | anon_vma = anon_vma_alloc(); | |
182 | if (unlikely(!anon_vma)) | |
183 | goto out_enomem_free_avc; | |
184 | allocated = anon_vma; | |
185 | } | |
186 | ||
187 | anon_vma_lock_write(anon_vma); | |
188 | /* page_table_lock to protect against threads */ | |
189 | spin_lock(&mm->page_table_lock); | |
190 | if (likely(!vma->anon_vma)) { | |
191 | vma->anon_vma = anon_vma; | |
192 | anon_vma_chain_link(vma, avc, anon_vma); | |
193 | /* vma reference or self-parent link for new root */ | |
194 | anon_vma->degree++; | |
195 | allocated = NULL; | |
196 | avc = NULL; | |
197 | } | |
198 | spin_unlock(&mm->page_table_lock); | |
199 | anon_vma_unlock_write(anon_vma); | |
200 | ||
201 | if (unlikely(allocated)) | |
202 | put_anon_vma(allocated); | |
203 | if (unlikely(avc)) | |
204 | anon_vma_chain_free(avc); | |
205 | } | |
206 | return 0; | |
207 | ||
208 | out_enomem_free_avc: | |
209 | anon_vma_chain_free(avc); | |
210 | out_enomem: | |
211 | return -ENOMEM; | |
212 | } | |
213 | ||
214 | /* | |
215 | * This is a useful helper function for locking the anon_vma root as | |
216 | * we traverse the vma->anon_vma_chain, looping over anon_vma's that | |
217 | * have the same vma. | |
218 | * | |
219 | * Such anon_vma's should have the same root, so you'd expect to see | |
220 | * just a single mutex_lock for the whole traversal. | |
221 | */ | |
222 | static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma) | |
223 | { | |
224 | struct anon_vma *new_root = anon_vma->root; | |
225 | if (new_root != root) { | |
226 | if (WARN_ON_ONCE(root)) | |
227 | up_write(&root->rwsem); | |
228 | root = new_root; | |
229 | down_write(&root->rwsem); | |
230 | } | |
231 | return root; | |
232 | } | |
233 | ||
234 | static inline void unlock_anon_vma_root(struct anon_vma *root) | |
235 | { | |
236 | if (root) | |
237 | up_write(&root->rwsem); | |
238 | } | |
239 | ||
240 | /* | |
241 | * Attach the anon_vmas from src to dst. | |
242 | * Returns 0 on success, -ENOMEM on failure. | |
243 | * | |
244 | * If dst->anon_vma is NULL this function tries to find and reuse existing | |
245 | * anon_vma which has no vmas and only one child anon_vma. This prevents | |
246 | * degradation of anon_vma hierarchy to endless linear chain in case of | |
247 | * constantly forking task. On the other hand, an anon_vma with more than one | |
248 | * child isn't reused even if there was no alive vma, thus rmap walker has a | |
249 | * good chance of avoiding scanning the whole hierarchy when it searches where | |
250 | * page is mapped. | |
251 | */ | |
252 | int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) | |
253 | { | |
254 | struct anon_vma_chain *avc, *pavc; | |
255 | struct anon_vma *root = NULL; | |
256 | ||
257 | list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { | |
258 | struct anon_vma *anon_vma; | |
259 | ||
260 | avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN); | |
261 | if (unlikely(!avc)) { | |
262 | unlock_anon_vma_root(root); | |
263 | root = NULL; | |
264 | avc = anon_vma_chain_alloc(GFP_KERNEL); | |
265 | if (!avc) | |
266 | goto enomem_failure; | |
267 | } | |
268 | anon_vma = pavc->anon_vma; | |
269 | root = lock_anon_vma_root(root, anon_vma); | |
270 | anon_vma_chain_link(dst, avc, anon_vma); | |
271 | ||
272 | /* | |
273 | * Reuse existing anon_vma if its degree lower than two, | |
274 | * that means it has no vma and only one anon_vma child. | |
275 | * | |
276 | * Do not chose parent anon_vma, otherwise first child | |
277 | * will always reuse it. Root anon_vma is never reused: | |
278 | * it has self-parent reference and at least one child. | |
279 | */ | |
280 | if (!dst->anon_vma && anon_vma != src->anon_vma && | |
281 | anon_vma->degree < 2) | |
282 | dst->anon_vma = anon_vma; | |
283 | } | |
284 | if (dst->anon_vma) | |
285 | dst->anon_vma->degree++; | |
286 | unlock_anon_vma_root(root); | |
287 | return 0; | |
288 | ||
289 | enomem_failure: | |
290 | /* | |
291 | * dst->anon_vma is dropped here otherwise its degree can be incorrectly | |
292 | * decremented in unlink_anon_vmas(). | |
293 | * We can safely do this because callers of anon_vma_clone() don't care | |
294 | * about dst->anon_vma if anon_vma_clone() failed. | |
295 | */ | |
296 | dst->anon_vma = NULL; | |
297 | unlink_anon_vmas(dst); | |
298 | return -ENOMEM; | |
299 | } | |
300 | ||
301 | /* | |
302 | * Attach vma to its own anon_vma, as well as to the anon_vmas that | |
303 | * the corresponding VMA in the parent process is attached to. | |
304 | * Returns 0 on success, non-zero on failure. | |
305 | */ | |
306 | int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) | |
307 | { | |
308 | struct anon_vma_chain *avc; | |
309 | struct anon_vma *anon_vma; | |
310 | int error; | |
311 | ||
312 | /* Don't bother if the parent process has no anon_vma here. */ | |
313 | if (!pvma->anon_vma) | |
314 | return 0; | |
315 | ||
316 | /* Drop inherited anon_vma, we'll reuse existing or allocate new. */ | |
317 | vma->anon_vma = NULL; | |
318 | ||
319 | /* | |
320 | * First, attach the new VMA to the parent VMA's anon_vmas, | |
321 | * so rmap can find non-COWed pages in child processes. | |
322 | */ | |
323 | error = anon_vma_clone(vma, pvma); | |
324 | if (error) | |
325 | return error; | |
326 | ||
327 | /* An existing anon_vma has been reused, all done then. */ | |
328 | if (vma->anon_vma) | |
329 | return 0; | |
330 | ||
331 | /* Then add our own anon_vma. */ | |
332 | anon_vma = anon_vma_alloc(); | |
333 | if (!anon_vma) | |
334 | goto out_error; | |
335 | avc = anon_vma_chain_alloc(GFP_KERNEL); | |
336 | if (!avc) | |
337 | goto out_error_free_anon_vma; | |
338 | ||
339 | /* | |
340 | * The root anon_vma's spinlock is the lock actually used when we | |
341 | * lock any of the anon_vmas in this anon_vma tree. | |
342 | */ | |
343 | anon_vma->root = pvma->anon_vma->root; | |
344 | anon_vma->parent = pvma->anon_vma; | |
345 | /* | |
346 | * With refcounts, an anon_vma can stay around longer than the | |
347 | * process it belongs to. The root anon_vma needs to be pinned until | |
348 | * this anon_vma is freed, because the lock lives in the root. | |
349 | */ | |
350 | get_anon_vma(anon_vma->root); | |
351 | /* Mark this anon_vma as the one where our new (COWed) pages go. */ | |
352 | vma->anon_vma = anon_vma; | |
353 | anon_vma_lock_write(anon_vma); | |
354 | anon_vma_chain_link(vma, avc, anon_vma); | |
355 | anon_vma->parent->degree++; | |
356 | anon_vma_unlock_write(anon_vma); | |
357 | ||
358 | return 0; | |
359 | ||
360 | out_error_free_anon_vma: | |
361 | put_anon_vma(anon_vma); | |
362 | out_error: | |
363 | unlink_anon_vmas(vma); | |
364 | return -ENOMEM; | |
365 | } | |
366 | ||
367 | void unlink_anon_vmas(struct vm_area_struct *vma) | |
368 | { | |
369 | struct anon_vma_chain *avc, *next; | |
370 | struct anon_vma *root = NULL; | |
371 | ||
372 | /* | |
373 | * Unlink each anon_vma chained to the VMA. This list is ordered | |
374 | * from newest to oldest, ensuring the root anon_vma gets freed last. | |
375 | */ | |
376 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { | |
377 | struct anon_vma *anon_vma = avc->anon_vma; | |
378 | ||
379 | root = lock_anon_vma_root(root, anon_vma); | |
380 | anon_vma_interval_tree_remove(avc, &anon_vma->rb_root); | |
381 | ||
382 | /* | |
383 | * Leave empty anon_vmas on the list - we'll need | |
384 | * to free them outside the lock. | |
385 | */ | |
386 | if (RB_EMPTY_ROOT(&anon_vma->rb_root)) { | |
387 | anon_vma->parent->degree--; | |
388 | continue; | |
389 | } | |
390 | ||
391 | list_del(&avc->same_vma); | |
392 | anon_vma_chain_free(avc); | |
393 | } | |
394 | if (vma->anon_vma) | |
395 | vma->anon_vma->degree--; | |
396 | unlock_anon_vma_root(root); | |
397 | ||
398 | /* | |
399 | * Iterate the list once more, it now only contains empty and unlinked | |
400 | * anon_vmas, destroy them. Could not do before due to __put_anon_vma() | |
401 | * needing to write-acquire the anon_vma->root->rwsem. | |
402 | */ | |
403 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { | |
404 | struct anon_vma *anon_vma = avc->anon_vma; | |
405 | ||
406 | BUG_ON(anon_vma->degree); | |
407 | put_anon_vma(anon_vma); | |
408 | ||
409 | list_del(&avc->same_vma); | |
410 | anon_vma_chain_free(avc); | |
411 | } | |
412 | } | |
413 | ||
414 | static void anon_vma_ctor(void *data) | |
415 | { | |
416 | struct anon_vma *anon_vma = data; | |
417 | ||
418 | init_rwsem(&anon_vma->rwsem); | |
419 | atomic_set(&anon_vma->refcount, 0); | |
420 | anon_vma->rb_root = RB_ROOT; | |
421 | } | |
422 | ||
423 | void __init anon_vma_init(void) | |
424 | { | |
425 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
426 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor); | |
427 | anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC); | |
428 | } | |
429 | ||
430 | /* | |
431 | * Getting a lock on a stable anon_vma from a page off the LRU is tricky! | |
432 | * | |
433 | * Since there is no serialization what so ever against page_remove_rmap() | |
434 | * the best this function can do is return a locked anon_vma that might | |
435 | * have been relevant to this page. | |
436 | * | |
437 | * The page might have been remapped to a different anon_vma or the anon_vma | |
438 | * returned may already be freed (and even reused). | |
439 | * | |
440 | * In case it was remapped to a different anon_vma, the new anon_vma will be a | |
441 | * child of the old anon_vma, and the anon_vma lifetime rules will therefore | |
442 | * ensure that any anon_vma obtained from the page will still be valid for as | |
443 | * long as we observe page_mapped() [ hence all those page_mapped() tests ]. | |
444 | * | |
445 | * All users of this function must be very careful when walking the anon_vma | |
446 | * chain and verify that the page in question is indeed mapped in it | |
447 | * [ something equivalent to page_mapped_in_vma() ]. | |
448 | * | |
449 | * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap() | |
450 | * that the anon_vma pointer from page->mapping is valid if there is a | |
451 | * mapcount, we can dereference the anon_vma after observing those. | |
452 | */ | |
453 | struct anon_vma *page_get_anon_vma(struct page *page) | |
454 | { | |
455 | struct anon_vma *anon_vma = NULL; | |
456 | unsigned long anon_mapping; | |
457 | ||
458 | rcu_read_lock(); | |
459 | anon_mapping = (unsigned long)READ_ONCE(page->mapping); | |
460 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) | |
461 | goto out; | |
462 | if (!page_mapped(page)) | |
463 | goto out; | |
464 | ||
465 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
466 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { | |
467 | anon_vma = NULL; | |
468 | goto out; | |
469 | } | |
470 | ||
471 | /* | |
472 | * If this page is still mapped, then its anon_vma cannot have been | |
473 | * freed. But if it has been unmapped, we have no security against the | |
474 | * anon_vma structure being freed and reused (for another anon_vma: | |
475 | * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero() | |
476 | * above cannot corrupt). | |
477 | */ | |
478 | if (!page_mapped(page)) { | |
479 | rcu_read_unlock(); | |
480 | put_anon_vma(anon_vma); | |
481 | return NULL; | |
482 | } | |
483 | out: | |
484 | rcu_read_unlock(); | |
485 | ||
486 | return anon_vma; | |
487 | } | |
488 | ||
489 | /* | |
490 | * Similar to page_get_anon_vma() except it locks the anon_vma. | |
491 | * | |
492 | * Its a little more complex as it tries to keep the fast path to a single | |
493 | * atomic op -- the trylock. If we fail the trylock, we fall back to getting a | |
494 | * reference like with page_get_anon_vma() and then block on the mutex. | |
495 | */ | |
496 | struct anon_vma *page_lock_anon_vma_read(struct page *page) | |
497 | { | |
498 | struct anon_vma *anon_vma = NULL; | |
499 | struct anon_vma *root_anon_vma; | |
500 | unsigned long anon_mapping; | |
501 | ||
502 | rcu_read_lock(); | |
503 | anon_mapping = (unsigned long)READ_ONCE(page->mapping); | |
504 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) | |
505 | goto out; | |
506 | if (!page_mapped(page)) | |
507 | goto out; | |
508 | ||
509 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
510 | root_anon_vma = READ_ONCE(anon_vma->root); | |
511 | if (down_read_trylock(&root_anon_vma->rwsem)) { | |
512 | /* | |
513 | * If the page is still mapped, then this anon_vma is still | |
514 | * its anon_vma, and holding the mutex ensures that it will | |
515 | * not go away, see anon_vma_free(). | |
516 | */ | |
517 | if (!page_mapped(page)) { | |
518 | up_read(&root_anon_vma->rwsem); | |
519 | anon_vma = NULL; | |
520 | } | |
521 | goto out; | |
522 | } | |
523 | ||
524 | /* trylock failed, we got to sleep */ | |
525 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { | |
526 | anon_vma = NULL; | |
527 | goto out; | |
528 | } | |
529 | ||
530 | if (!page_mapped(page)) { | |
531 | rcu_read_unlock(); | |
532 | put_anon_vma(anon_vma); | |
533 | return NULL; | |
534 | } | |
535 | ||
536 | /* we pinned the anon_vma, its safe to sleep */ | |
537 | rcu_read_unlock(); | |
538 | anon_vma_lock_read(anon_vma); | |
539 | ||
540 | if (atomic_dec_and_test(&anon_vma->refcount)) { | |
541 | /* | |
542 | * Oops, we held the last refcount, release the lock | |
543 | * and bail -- can't simply use put_anon_vma() because | |
544 | * we'll deadlock on the anon_vma_lock_write() recursion. | |
545 | */ | |
546 | anon_vma_unlock_read(anon_vma); | |
547 | __put_anon_vma(anon_vma); | |
548 | anon_vma = NULL; | |
549 | } | |
550 | ||
551 | return anon_vma; | |
552 | ||
553 | out: | |
554 | rcu_read_unlock(); | |
555 | return anon_vma; | |
556 | } | |
557 | ||
558 | void page_unlock_anon_vma_read(struct anon_vma *anon_vma) | |
559 | { | |
560 | anon_vma_unlock_read(anon_vma); | |
561 | } | |
562 | ||
563 | /* | |
564 | * At what user virtual address is page expected in @vma? | |
565 | */ | |
566 | static inline unsigned long | |
567 | __vma_address(struct page *page, struct vm_area_struct *vma) | |
568 | { | |
569 | pgoff_t pgoff = page_to_pgoff(page); | |
570 | return vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
571 | } | |
572 | ||
573 | inline unsigned long | |
574 | vma_address(struct page *page, struct vm_area_struct *vma) | |
575 | { | |
576 | unsigned long address = __vma_address(page, vma); | |
577 | ||
578 | /* page should be within @vma mapping range */ | |
579 | VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); | |
580 | ||
581 | return address; | |
582 | } | |
583 | ||
584 | /* | |
585 | * At what user virtual address is page expected in vma? | |
586 | * Caller should check the page is actually part of the vma. | |
587 | */ | |
588 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
589 | { | |
590 | unsigned long address; | |
591 | if (PageAnon(page)) { | |
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) | |
599 | return -EFAULT; | |
600 | } else if (page->mapping) { | |
601 | if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping) | |
602 | return -EFAULT; | |
603 | } else | |
604 | return -EFAULT; | |
605 | address = __vma_address(page, vma); | |
606 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) | |
607 | return -EFAULT; | |
608 | return address; | |
609 | } | |
610 | ||
611 | pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address) | |
612 | { | |
613 | pgd_t *pgd; | |
614 | pud_t *pud; | |
615 | pmd_t *pmd = NULL; | |
616 | pmd_t pmde; | |
617 | ||
618 | pgd = pgd_offset(mm, address); | |
619 | if (!pgd_present(*pgd)) | |
620 | goto out; | |
621 | ||
622 | pud = pud_offset(pgd, address); | |
623 | if (!pud_present(*pud)) | |
624 | goto out; | |
625 | ||
626 | pmd = pmd_offset(pud, address); | |
627 | /* | |
628 | * Some THP functions use the sequence pmdp_clear_flush(), set_pmd_at() | |
629 | * without holding anon_vma lock for write. So when looking for a | |
630 | * genuine pmde (in which to find pte), test present and !THP together. | |
631 | */ | |
632 | pmde = *pmd; | |
633 | barrier(); | |
634 | if (!pmd_present(pmde) || pmd_trans_huge(pmde)) | |
635 | pmd = NULL; | |
636 | out: | |
637 | return pmd; | |
638 | } | |
639 | ||
640 | /* | |
641 | * Check that @page is mapped at @address into @mm. | |
642 | * | |
643 | * If @sync is false, page_check_address may perform a racy check to avoid | |
644 | * the page table lock when the pte is not present (helpful when reclaiming | |
645 | * highly shared pages). | |
646 | * | |
647 | * On success returns with pte mapped and locked. | |
648 | */ | |
649 | pte_t *__page_check_address(struct page *page, struct mm_struct *mm, | |
650 | unsigned long address, spinlock_t **ptlp, int sync) | |
651 | { | |
652 | pmd_t *pmd; | |
653 | pte_t *pte; | |
654 | spinlock_t *ptl; | |
655 | ||
656 | if (unlikely(PageHuge(page))) { | |
657 | /* when pud is not present, pte will be NULL */ | |
658 | pte = huge_pte_offset(mm, address); | |
659 | if (!pte) | |
660 | return NULL; | |
661 | ||
662 | ptl = huge_pte_lockptr(page_hstate(page), mm, pte); | |
663 | goto check; | |
664 | } | |
665 | ||
666 | pmd = mm_find_pmd(mm, address); | |
667 | if (!pmd) | |
668 | return NULL; | |
669 | ||
670 | pte = pte_offset_map(pmd, address); | |
671 | /* Make a quick check before getting the lock */ | |
672 | if (!sync && !pte_present(*pte)) { | |
673 | pte_unmap(pte); | |
674 | return NULL; | |
675 | } | |
676 | ||
677 | ptl = pte_lockptr(mm, pmd); | |
678 | check: | |
679 | spin_lock(ptl); | |
680 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { | |
681 | *ptlp = ptl; | |
682 | return pte; | |
683 | } | |
684 | pte_unmap_unlock(pte, ptl); | |
685 | return NULL; | |
686 | } | |
687 | ||
688 | /** | |
689 | * page_mapped_in_vma - check whether a page is really mapped in a VMA | |
690 | * @page: the page to test | |
691 | * @vma: the VMA to test | |
692 | * | |
693 | * Returns 1 if the page is mapped into the page tables of the VMA, 0 | |
694 | * if the page is not mapped into the page tables of this VMA. Only | |
695 | * valid for normal file or anonymous VMAs. | |
696 | */ | |
697 | int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma) | |
698 | { | |
699 | unsigned long address; | |
700 | pte_t *pte; | |
701 | spinlock_t *ptl; | |
702 | ||
703 | address = __vma_address(page, vma); | |
704 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) | |
705 | return 0; | |
706 | pte = page_check_address(page, vma->vm_mm, address, &ptl, 1); | |
707 | if (!pte) /* the page is not in this mm */ | |
708 | return 0; | |
709 | pte_unmap_unlock(pte, ptl); | |
710 | ||
711 | return 1; | |
712 | } | |
713 | ||
714 | struct page_referenced_arg { | |
715 | int mapcount; | |
716 | int referenced; | |
717 | unsigned long vm_flags; | |
718 | struct mem_cgroup *memcg; | |
719 | }; | |
720 | /* | |
721 | * arg: page_referenced_arg will be passed | |
722 | */ | |
723 | static int page_referenced_one(struct page *page, struct vm_area_struct *vma, | |
724 | unsigned long address, void *arg) | |
725 | { | |
726 | struct mm_struct *mm = vma->vm_mm; | |
727 | spinlock_t *ptl; | |
728 | int referenced = 0; | |
729 | struct page_referenced_arg *pra = arg; | |
730 | ||
731 | if (unlikely(PageTransHuge(page))) { | |
732 | pmd_t *pmd; | |
733 | ||
734 | /* | |
735 | * rmap might return false positives; we must filter | |
736 | * these out using page_check_address_pmd(). | |
737 | */ | |
738 | pmd = page_check_address_pmd(page, mm, address, | |
739 | PAGE_CHECK_ADDRESS_PMD_FLAG, &ptl); | |
740 | if (!pmd) | |
741 | return SWAP_AGAIN; | |
742 | ||
743 | if (vma->vm_flags & VM_LOCKED) { | |
744 | spin_unlock(ptl); | |
745 | pra->vm_flags |= VM_LOCKED; | |
746 | return SWAP_FAIL; /* To break the loop */ | |
747 | } | |
748 | ||
749 | /* go ahead even if the pmd is pmd_trans_splitting() */ | |
750 | if (pmdp_clear_flush_young_notify(vma, address, pmd)) | |
751 | referenced++; | |
752 | spin_unlock(ptl); | |
753 | } else { | |
754 | pte_t *pte; | |
755 | ||
756 | /* | |
757 | * rmap might return false positives; we must filter | |
758 | * these out using page_check_address(). | |
759 | */ | |
760 | pte = page_check_address(page, mm, address, &ptl, 0); | |
761 | if (!pte) | |
762 | return SWAP_AGAIN; | |
763 | ||
764 | if (vma->vm_flags & VM_LOCKED) { | |
765 | pte_unmap_unlock(pte, ptl); | |
766 | pra->vm_flags |= VM_LOCKED; | |
767 | return SWAP_FAIL; /* To break the loop */ | |
768 | } | |
769 | ||
770 | if (ptep_clear_flush_young_notify(vma, address, pte)) { | |
771 | /* | |
772 | * Don't treat a reference through a sequentially read | |
773 | * mapping as such. If the page has been used in | |
774 | * another mapping, we will catch it; if this other | |
775 | * mapping is already gone, the unmap path will have | |
776 | * set PG_referenced or activated the page. | |
777 | */ | |
778 | if (likely(!(vma->vm_flags & VM_SEQ_READ))) | |
779 | referenced++; | |
780 | } | |
781 | pte_unmap_unlock(pte, ptl); | |
782 | } | |
783 | ||
784 | if (referenced) { | |
785 | pra->referenced++; | |
786 | pra->vm_flags |= vma->vm_flags; | |
787 | } | |
788 | ||
789 | pra->mapcount--; | |
790 | if (!pra->mapcount) | |
791 | return SWAP_SUCCESS; /* To break the loop */ | |
792 | ||
793 | return SWAP_AGAIN; | |
794 | } | |
795 | ||
796 | static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg) | |
797 | { | |
798 | struct page_referenced_arg *pra = arg; | |
799 | struct mem_cgroup *memcg = pra->memcg; | |
800 | ||
801 | if (!mm_match_cgroup(vma->vm_mm, memcg)) | |
802 | return true; | |
803 | ||
804 | return false; | |
805 | } | |
806 | ||
807 | /** | |
808 | * page_referenced - test if the page was referenced | |
809 | * @page: the page to test | |
810 | * @is_locked: caller holds lock on the page | |
811 | * @memcg: target memory cgroup | |
812 | * @vm_flags: collect encountered vma->vm_flags who actually referenced the page | |
813 | * | |
814 | * Quick test_and_clear_referenced for all mappings to a page, | |
815 | * returns the number of ptes which referenced the page. | |
816 | */ | |
817 | int page_referenced(struct page *page, | |
818 | int is_locked, | |
819 | struct mem_cgroup *memcg, | |
820 | unsigned long *vm_flags) | |
821 | { | |
822 | int ret; | |
823 | int we_locked = 0; | |
824 | struct page_referenced_arg pra = { | |
825 | .mapcount = page_mapcount(page), | |
826 | .memcg = memcg, | |
827 | }; | |
828 | struct rmap_walk_control rwc = { | |
829 | .rmap_one = page_referenced_one, | |
830 | .arg = (void *)&pra, | |
831 | .anon_lock = page_lock_anon_vma_read, | |
832 | }; | |
833 | ||
834 | *vm_flags = 0; | |
835 | if (!page_mapped(page)) | |
836 | return 0; | |
837 | ||
838 | if (!page_rmapping(page)) | |
839 | return 0; | |
840 | ||
841 | if (!is_locked && (!PageAnon(page) || PageKsm(page))) { | |
842 | we_locked = trylock_page(page); | |
843 | if (!we_locked) | |
844 | return 1; | |
845 | } | |
846 | ||
847 | /* | |
848 | * If we are reclaiming on behalf of a cgroup, skip | |
849 | * counting on behalf of references from different | |
850 | * cgroups | |
851 | */ | |
852 | if (memcg) { | |
853 | rwc.invalid_vma = invalid_page_referenced_vma; | |
854 | } | |
855 | ||
856 | ret = rmap_walk(page, &rwc); | |
857 | *vm_flags = pra.vm_flags; | |
858 | ||
859 | if (we_locked) | |
860 | unlock_page(page); | |
861 | ||
862 | return pra.referenced; | |
863 | } | |
864 | ||
865 | static int page_mkclean_one(struct page *page, struct vm_area_struct *vma, | |
866 | unsigned long address, void *arg) | |
867 | { | |
868 | struct mm_struct *mm = vma->vm_mm; | |
869 | pte_t *pte; | |
870 | spinlock_t *ptl; | |
871 | int ret = 0; | |
872 | int *cleaned = arg; | |
873 | ||
874 | pte = page_check_address(page, mm, address, &ptl, 1); | |
875 | if (!pte) | |
876 | goto out; | |
877 | ||
878 | if (pte_dirty(*pte) || pte_write(*pte)) { | |
879 | pte_t entry; | |
880 | ||
881 | flush_cache_page(vma, address, pte_pfn(*pte)); | |
882 | entry = ptep_clear_flush(vma, address, pte); | |
883 | entry = pte_wrprotect(entry); | |
884 | entry = pte_mkclean(entry); | |
885 | set_pte_at(mm, address, pte, entry); | |
886 | ret = 1; | |
887 | } | |
888 | ||
889 | pte_unmap_unlock(pte, ptl); | |
890 | ||
891 | if (ret) { | |
892 | mmu_notifier_invalidate_page(mm, address); | |
893 | (*cleaned)++; | |
894 | } | |
895 | out: | |
896 | return SWAP_AGAIN; | |
897 | } | |
898 | ||
899 | static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg) | |
900 | { | |
901 | if (vma->vm_flags & VM_SHARED) | |
902 | return false; | |
903 | ||
904 | return true; | |
905 | } | |
906 | ||
907 | int page_mkclean(struct page *page) | |
908 | { | |
909 | int cleaned = 0; | |
910 | struct address_space *mapping; | |
911 | struct rmap_walk_control rwc = { | |
912 | .arg = (void *)&cleaned, | |
913 | .rmap_one = page_mkclean_one, | |
914 | .invalid_vma = invalid_mkclean_vma, | |
915 | }; | |
916 | ||
917 | BUG_ON(!PageLocked(page)); | |
918 | ||
919 | if (!page_mapped(page)) | |
920 | return 0; | |
921 | ||
922 | mapping = page_mapping(page); | |
923 | if (!mapping) | |
924 | return 0; | |
925 | ||
926 | rmap_walk(page, &rwc); | |
927 | ||
928 | return cleaned; | |
929 | } | |
930 | EXPORT_SYMBOL_GPL(page_mkclean); | |
931 | ||
932 | /** | |
933 | * page_move_anon_rmap - move a page to our anon_vma | |
934 | * @page: the page to move to our anon_vma | |
935 | * @vma: the vma the page belongs to | |
936 | * @address: the user virtual address mapped | |
937 | * | |
938 | * When a page belongs exclusively to one process after a COW event, | |
939 | * that page can be moved into the anon_vma that belongs to just that | |
940 | * process, so the rmap code will not search the parent or sibling | |
941 | * processes. | |
942 | */ | |
943 | void page_move_anon_rmap(struct page *page, | |
944 | struct vm_area_struct *vma, unsigned long address) | |
945 | { | |
946 | struct anon_vma *anon_vma = vma->anon_vma; | |
947 | ||
948 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
949 | VM_BUG_ON_VMA(!anon_vma, vma); | |
950 | VM_BUG_ON_PAGE(page->index != linear_page_index(vma, address), page); | |
951 | ||
952 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
953 | page->mapping = (struct address_space *) anon_vma; | |
954 | } | |
955 | ||
956 | /** | |
957 | * __page_set_anon_rmap - set up new anonymous rmap | |
958 | * @page: Page to add to rmap | |
959 | * @vma: VM area to add page to. | |
960 | * @address: User virtual address of the mapping | |
961 | * @exclusive: the page is exclusively owned by the current process | |
962 | */ | |
963 | static void __page_set_anon_rmap(struct page *page, | |
964 | struct vm_area_struct *vma, unsigned long address, int exclusive) | |
965 | { | |
966 | struct anon_vma *anon_vma = vma->anon_vma; | |
967 | ||
968 | BUG_ON(!anon_vma); | |
969 | ||
970 | if (PageAnon(page)) | |
971 | return; | |
972 | ||
973 | /* | |
974 | * If the page isn't exclusively mapped into this vma, | |
975 | * we must use the _oldest_ possible anon_vma for the | |
976 | * page mapping! | |
977 | */ | |
978 | if (!exclusive) | |
979 | anon_vma = anon_vma->root; | |
980 | ||
981 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
982 | page->mapping = (struct address_space *) anon_vma; | |
983 | page->index = linear_page_index(vma, address); | |
984 | } | |
985 | ||
986 | /** | |
987 | * __page_check_anon_rmap - sanity check anonymous rmap addition | |
988 | * @page: the page to add the mapping to | |
989 | * @vma: the vm area in which the mapping is added | |
990 | * @address: the user virtual address mapped | |
991 | */ | |
992 | static void __page_check_anon_rmap(struct page *page, | |
993 | struct vm_area_struct *vma, unsigned long address) | |
994 | { | |
995 | #ifdef CONFIG_DEBUG_VM | |
996 | /* | |
997 | * The page's anon-rmap details (mapping and index) are guaranteed to | |
998 | * be set up correctly at this point. | |
999 | * | |
1000 | * We have exclusion against page_add_anon_rmap because the caller | |
1001 | * always holds the page locked, except if called from page_dup_rmap, | |
1002 | * in which case the page is already known to be setup. | |
1003 | * | |
1004 | * We have exclusion against page_add_new_anon_rmap because those pages | |
1005 | * are initially only visible via the pagetables, and the pte is locked | |
1006 | * over the call to page_add_new_anon_rmap. | |
1007 | */ | |
1008 | BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root); | |
1009 | BUG_ON(page->index != linear_page_index(vma, address)); | |
1010 | #endif | |
1011 | } | |
1012 | ||
1013 | /** | |
1014 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
1015 | * @page: the page to add the mapping to | |
1016 | * @vma: the vm area in which the mapping is added | |
1017 | * @address: the user virtual address mapped | |
1018 | * | |
1019 | * The caller needs to hold the pte lock, and the page must be locked in | |
1020 | * the anon_vma case: to serialize mapping,index checking after setting, | |
1021 | * and to ensure that PageAnon is not being upgraded racily to PageKsm | |
1022 | * (but PageKsm is never downgraded to PageAnon). | |
1023 | */ | |
1024 | void page_add_anon_rmap(struct page *page, | |
1025 | struct vm_area_struct *vma, unsigned long address) | |
1026 | { | |
1027 | do_page_add_anon_rmap(page, vma, address, 0); | |
1028 | } | |
1029 | ||
1030 | /* | |
1031 | * Special version of the above for do_swap_page, which often runs | |
1032 | * into pages that are exclusively owned by the current process. | |
1033 | * Everybody else should continue to use page_add_anon_rmap above. | |
1034 | */ | |
1035 | void do_page_add_anon_rmap(struct page *page, | |
1036 | struct vm_area_struct *vma, unsigned long address, int exclusive) | |
1037 | { | |
1038 | int first = atomic_inc_and_test(&page->_mapcount); | |
1039 | if (first) { | |
1040 | /* | |
1041 | * We use the irq-unsafe __{inc|mod}_zone_page_stat because | |
1042 | * these counters are not modified in interrupt context, and | |
1043 | * pte lock(a spinlock) is held, which implies preemption | |
1044 | * disabled. | |
1045 | */ | |
1046 | if (PageTransHuge(page)) | |
1047 | __inc_zone_page_state(page, | |
1048 | NR_ANON_TRANSPARENT_HUGEPAGES); | |
1049 | __mod_zone_page_state(page_zone(page), NR_ANON_PAGES, | |
1050 | hpage_nr_pages(page)); | |
1051 | } | |
1052 | if (unlikely(PageKsm(page))) | |
1053 | return; | |
1054 | ||
1055 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
1056 | /* address might be in next vma when migration races vma_adjust */ | |
1057 | if (first) | |
1058 | __page_set_anon_rmap(page, vma, address, exclusive); | |
1059 | else | |
1060 | __page_check_anon_rmap(page, vma, address); | |
1061 | } | |
1062 | ||
1063 | /** | |
1064 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page | |
1065 | * @page: the page to add the mapping to | |
1066 | * @vma: the vm area in which the mapping is added | |
1067 | * @address: the user virtual address mapped | |
1068 | * | |
1069 | * Same as page_add_anon_rmap but must only be called on *new* pages. | |
1070 | * This means the inc-and-test can be bypassed. | |
1071 | * Page does not have to be locked. | |
1072 | */ | |
1073 | void page_add_new_anon_rmap(struct page *page, | |
1074 | struct vm_area_struct *vma, unsigned long address) | |
1075 | { | |
1076 | VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); | |
1077 | SetPageSwapBacked(page); | |
1078 | atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */ | |
1079 | if (PageTransHuge(page)) | |
1080 | __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES); | |
1081 | __mod_zone_page_state(page_zone(page), NR_ANON_PAGES, | |
1082 | hpage_nr_pages(page)); | |
1083 | __page_set_anon_rmap(page, vma, address, 1); | |
1084 | } | |
1085 | ||
1086 | /** | |
1087 | * page_add_file_rmap - add pte mapping to a file page | |
1088 | * @page: the page to add the mapping to | |
1089 | * | |
1090 | * The caller needs to hold the pte lock. | |
1091 | */ | |
1092 | void page_add_file_rmap(struct page *page) | |
1093 | { | |
1094 | struct mem_cgroup *memcg; | |
1095 | ||
1096 | memcg = mem_cgroup_begin_page_stat(page); | |
1097 | if (atomic_inc_and_test(&page->_mapcount)) { | |
1098 | __inc_zone_page_state(page, NR_FILE_MAPPED); | |
1099 | mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED); | |
1100 | } | |
1101 | mem_cgroup_end_page_stat(memcg); | |
1102 | } | |
1103 | ||
1104 | static void page_remove_file_rmap(struct page *page) | |
1105 | { | |
1106 | struct mem_cgroup *memcg; | |
1107 | ||
1108 | memcg = mem_cgroup_begin_page_stat(page); | |
1109 | ||
1110 | /* page still mapped by someone else? */ | |
1111 | if (!atomic_add_negative(-1, &page->_mapcount)) | |
1112 | goto out; | |
1113 | ||
1114 | /* Hugepages are not counted in NR_FILE_MAPPED for now. */ | |
1115 | if (unlikely(PageHuge(page))) | |
1116 | goto out; | |
1117 | ||
1118 | /* | |
1119 | * We use the irq-unsafe __{inc|mod}_zone_page_stat because | |
1120 | * these counters are not modified in interrupt context, and | |
1121 | * pte lock(a spinlock) is held, which implies preemption disabled. | |
1122 | */ | |
1123 | __dec_zone_page_state(page, NR_FILE_MAPPED); | |
1124 | mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED); | |
1125 | ||
1126 | if (unlikely(PageMlocked(page))) | |
1127 | clear_page_mlock(page); | |
1128 | out: | |
1129 | mem_cgroup_end_page_stat(memcg); | |
1130 | } | |
1131 | ||
1132 | /** | |
1133 | * page_remove_rmap - take down pte mapping from a page | |
1134 | * @page: page to remove mapping from | |
1135 | * | |
1136 | * The caller needs to hold the pte lock. | |
1137 | */ | |
1138 | void page_remove_rmap(struct page *page) | |
1139 | { | |
1140 | if (!PageAnon(page)) { | |
1141 | page_remove_file_rmap(page); | |
1142 | return; | |
1143 | } | |
1144 | ||
1145 | /* page still mapped by someone else? */ | |
1146 | if (!atomic_add_negative(-1, &page->_mapcount)) | |
1147 | return; | |
1148 | ||
1149 | /* Hugepages are not counted in NR_ANON_PAGES for now. */ | |
1150 | if (unlikely(PageHuge(page))) | |
1151 | return; | |
1152 | ||
1153 | /* | |
1154 | * We use the irq-unsafe __{inc|mod}_zone_page_stat because | |
1155 | * these counters are not modified in interrupt context, and | |
1156 | * pte lock(a spinlock) is held, which implies preemption disabled. | |
1157 | */ | |
1158 | if (PageTransHuge(page)) | |
1159 | __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES); | |
1160 | ||
1161 | __mod_zone_page_state(page_zone(page), NR_ANON_PAGES, | |
1162 | -hpage_nr_pages(page)); | |
1163 | ||
1164 | if (unlikely(PageMlocked(page))) | |
1165 | clear_page_mlock(page); | |
1166 | ||
1167 | /* | |
1168 | * It would be tidy to reset the PageAnon mapping here, | |
1169 | * but that might overwrite a racing page_add_anon_rmap | |
1170 | * which increments mapcount after us but sets mapping | |
1171 | * before us: so leave the reset to free_hot_cold_page, | |
1172 | * and remember that it's only reliable while mapped. | |
1173 | * Leaving it set also helps swapoff to reinstate ptes | |
1174 | * faster for those pages still in swapcache. | |
1175 | */ | |
1176 | } | |
1177 | ||
1178 | /* | |
1179 | * @arg: enum ttu_flags will be passed to this argument | |
1180 | */ | |
1181 | static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, | |
1182 | unsigned long address, void *arg) | |
1183 | { | |
1184 | struct mm_struct *mm = vma->vm_mm; | |
1185 | pte_t *pte; | |
1186 | pte_t pteval; | |
1187 | spinlock_t *ptl; | |
1188 | int ret = SWAP_AGAIN; | |
1189 | enum ttu_flags flags = (enum ttu_flags)arg; | |
1190 | ||
1191 | pte = page_check_address(page, mm, address, &ptl, 0); | |
1192 | if (!pte) | |
1193 | goto out; | |
1194 | ||
1195 | /* | |
1196 | * If the page is mlock()d, we cannot swap it out. | |
1197 | * If it's recently referenced (perhaps page_referenced | |
1198 | * skipped over this mm) then we should reactivate it. | |
1199 | */ | |
1200 | if (!(flags & TTU_IGNORE_MLOCK)) { | |
1201 | if (vma->vm_flags & VM_LOCKED) | |
1202 | goto out_mlock; | |
1203 | ||
1204 | if (flags & TTU_MUNLOCK) | |
1205 | goto out_unmap; | |
1206 | } | |
1207 | if (!(flags & TTU_IGNORE_ACCESS)) { | |
1208 | if (ptep_clear_flush_young_notify(vma, address, pte)) { | |
1209 | ret = SWAP_FAIL; | |
1210 | goto out_unmap; | |
1211 | } | |
1212 | } | |
1213 | ||
1214 | /* Nuke the page table entry. */ | |
1215 | flush_cache_page(vma, address, page_to_pfn(page)); | |
1216 | pteval = ptep_clear_flush(vma, address, pte); | |
1217 | ||
1218 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
1219 | if (pte_dirty(pteval)) | |
1220 | set_page_dirty(page); | |
1221 | ||
1222 | /* Update high watermark before we lower rss */ | |
1223 | update_hiwater_rss(mm); | |
1224 | ||
1225 | if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) { | |
1226 | if (!PageHuge(page)) { | |
1227 | if (PageAnon(page)) | |
1228 | dec_mm_counter(mm, MM_ANONPAGES); | |
1229 | else | |
1230 | dec_mm_counter(mm, MM_FILEPAGES); | |
1231 | } | |
1232 | set_pte_at(mm, address, pte, | |
1233 | swp_entry_to_pte(make_hwpoison_entry(page))); | |
1234 | } else if (pte_unused(pteval)) { | |
1235 | /* | |
1236 | * The guest indicated that the page content is of no | |
1237 | * interest anymore. Simply discard the pte, vmscan | |
1238 | * will take care of the rest. | |
1239 | */ | |
1240 | if (PageAnon(page)) | |
1241 | dec_mm_counter(mm, MM_ANONPAGES); | |
1242 | else | |
1243 | dec_mm_counter(mm, MM_FILEPAGES); | |
1244 | } else if (PageAnon(page)) { | |
1245 | swp_entry_t entry = { .val = page_private(page) }; | |
1246 | pte_t swp_pte; | |
1247 | ||
1248 | if (PageSwapCache(page)) { | |
1249 | /* | |
1250 | * Store the swap location in the pte. | |
1251 | * See handle_pte_fault() ... | |
1252 | */ | |
1253 | if (swap_duplicate(entry) < 0) { | |
1254 | set_pte_at(mm, address, pte, pteval); | |
1255 | ret = SWAP_FAIL; | |
1256 | goto out_unmap; | |
1257 | } | |
1258 | if (list_empty(&mm->mmlist)) { | |
1259 | spin_lock(&mmlist_lock); | |
1260 | if (list_empty(&mm->mmlist)) | |
1261 | list_add(&mm->mmlist, &init_mm.mmlist); | |
1262 | spin_unlock(&mmlist_lock); | |
1263 | } | |
1264 | dec_mm_counter(mm, MM_ANONPAGES); | |
1265 | inc_mm_counter(mm, MM_SWAPENTS); | |
1266 | } else if (IS_ENABLED(CONFIG_MIGRATION)) { | |
1267 | /* | |
1268 | * Store the pfn of the page in a special migration | |
1269 | * pte. do_swap_page() will wait until the migration | |
1270 | * pte is removed and then restart fault handling. | |
1271 | */ | |
1272 | BUG_ON(!(flags & TTU_MIGRATION)); | |
1273 | entry = make_migration_entry(page, pte_write(pteval)); | |
1274 | } | |
1275 | swp_pte = swp_entry_to_pte(entry); | |
1276 | if (pte_soft_dirty(pteval)) | |
1277 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
1278 | set_pte_at(mm, address, pte, swp_pte); | |
1279 | } else if (IS_ENABLED(CONFIG_MIGRATION) && | |
1280 | (flags & TTU_MIGRATION)) { | |
1281 | /* Establish migration entry for a file page */ | |
1282 | swp_entry_t entry; | |
1283 | entry = make_migration_entry(page, pte_write(pteval)); | |
1284 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
1285 | } else | |
1286 | dec_mm_counter(mm, MM_FILEPAGES); | |
1287 | ||
1288 | page_remove_rmap(page); | |
1289 | page_cache_release(page); | |
1290 | ||
1291 | out_unmap: | |
1292 | pte_unmap_unlock(pte, ptl); | |
1293 | if (ret != SWAP_FAIL && !(flags & TTU_MUNLOCK)) | |
1294 | mmu_notifier_invalidate_page(mm, address); | |
1295 | out: | |
1296 | return ret; | |
1297 | ||
1298 | out_mlock: | |
1299 | pte_unmap_unlock(pte, ptl); | |
1300 | ||
1301 | ||
1302 | /* | |
1303 | * We need mmap_sem locking, Otherwise VM_LOCKED check makes | |
1304 | * unstable result and race. Plus, We can't wait here because | |
1305 | * we now hold anon_vma->rwsem or mapping->i_mmap_rwsem. | |
1306 | * if trylock failed, the page remain in evictable lru and later | |
1307 | * vmscan could retry to move the page to unevictable lru if the | |
1308 | * page is actually mlocked. | |
1309 | */ | |
1310 | if (down_read_trylock(&vma->vm_mm->mmap_sem)) { | |
1311 | if (vma->vm_flags & VM_LOCKED) { | |
1312 | mlock_vma_page(page); | |
1313 | ret = SWAP_MLOCK; | |
1314 | } | |
1315 | up_read(&vma->vm_mm->mmap_sem); | |
1316 | } | |
1317 | return ret; | |
1318 | } | |
1319 | ||
1320 | bool is_vma_temporary_stack(struct vm_area_struct *vma) | |
1321 | { | |
1322 | int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP); | |
1323 | ||
1324 | if (!maybe_stack) | |
1325 | return false; | |
1326 | ||
1327 | if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) == | |
1328 | VM_STACK_INCOMPLETE_SETUP) | |
1329 | return true; | |
1330 | ||
1331 | return false; | |
1332 | } | |
1333 | ||
1334 | static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg) | |
1335 | { | |
1336 | return is_vma_temporary_stack(vma); | |
1337 | } | |
1338 | ||
1339 | static int page_not_mapped(struct page *page) | |
1340 | { | |
1341 | return !page_mapped(page); | |
1342 | }; | |
1343 | ||
1344 | /** | |
1345 | * try_to_unmap - try to remove all page table mappings to a page | |
1346 | * @page: the page to get unmapped | |
1347 | * @flags: action and flags | |
1348 | * | |
1349 | * Tries to remove all the page table entries which are mapping this | |
1350 | * page, used in the pageout path. Caller must hold the page lock. | |
1351 | * Return values are: | |
1352 | * | |
1353 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
1354 | * SWAP_AGAIN - we missed a mapping, try again later | |
1355 | * SWAP_FAIL - the page is unswappable | |
1356 | * SWAP_MLOCK - page is mlocked. | |
1357 | */ | |
1358 | int try_to_unmap(struct page *page, enum ttu_flags flags) | |
1359 | { | |
1360 | int ret; | |
1361 | struct rmap_walk_control rwc = { | |
1362 | .rmap_one = try_to_unmap_one, | |
1363 | .arg = (void *)flags, | |
1364 | .done = page_not_mapped, | |
1365 | .anon_lock = page_lock_anon_vma_read, | |
1366 | }; | |
1367 | ||
1368 | VM_BUG_ON_PAGE(!PageHuge(page) && PageTransHuge(page), page); | |
1369 | ||
1370 | /* | |
1371 | * During exec, a temporary VMA is setup and later moved. | |
1372 | * The VMA is moved under the anon_vma lock but not the | |
1373 | * page tables leading to a race where migration cannot | |
1374 | * find the migration ptes. Rather than increasing the | |
1375 | * locking requirements of exec(), migration skips | |
1376 | * temporary VMAs until after exec() completes. | |
1377 | */ | |
1378 | if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page)) | |
1379 | rwc.invalid_vma = invalid_migration_vma; | |
1380 | ||
1381 | ret = rmap_walk(page, &rwc); | |
1382 | ||
1383 | if (ret != SWAP_MLOCK && !page_mapped(page)) | |
1384 | ret = SWAP_SUCCESS; | |
1385 | return ret; | |
1386 | } | |
1387 | ||
1388 | /** | |
1389 | * try_to_munlock - try to munlock a page | |
1390 | * @page: the page to be munlocked | |
1391 | * | |
1392 | * Called from munlock code. Checks all of the VMAs mapping the page | |
1393 | * to make sure nobody else has this page mlocked. The page will be | |
1394 | * returned with PG_mlocked cleared if no other vmas have it mlocked. | |
1395 | * | |
1396 | * Return values are: | |
1397 | * | |
1398 | * SWAP_AGAIN - no vma is holding page mlocked, or, | |
1399 | * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem | |
1400 | * SWAP_FAIL - page cannot be located at present | |
1401 | * SWAP_MLOCK - page is now mlocked. | |
1402 | */ | |
1403 | int try_to_munlock(struct page *page) | |
1404 | { | |
1405 | int ret; | |
1406 | struct rmap_walk_control rwc = { | |
1407 | .rmap_one = try_to_unmap_one, | |
1408 | .arg = (void *)TTU_MUNLOCK, | |
1409 | .done = page_not_mapped, | |
1410 | .anon_lock = page_lock_anon_vma_read, | |
1411 | ||
1412 | }; | |
1413 | ||
1414 | VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page); | |
1415 | ||
1416 | ret = rmap_walk(page, &rwc); | |
1417 | return ret; | |
1418 | } | |
1419 | ||
1420 | void __put_anon_vma(struct anon_vma *anon_vma) | |
1421 | { | |
1422 | struct anon_vma *root = anon_vma->root; | |
1423 | ||
1424 | anon_vma_free(anon_vma); | |
1425 | if (root != anon_vma && atomic_dec_and_test(&root->refcount)) | |
1426 | anon_vma_free(root); | |
1427 | } | |
1428 | ||
1429 | static struct anon_vma *rmap_walk_anon_lock(struct page *page, | |
1430 | struct rmap_walk_control *rwc) | |
1431 | { | |
1432 | struct anon_vma *anon_vma; | |
1433 | ||
1434 | if (rwc->anon_lock) | |
1435 | return rwc->anon_lock(page); | |
1436 | ||
1437 | /* | |
1438 | * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read() | |
1439 | * because that depends on page_mapped(); but not all its usages | |
1440 | * are holding mmap_sem. Users without mmap_sem are required to | |
1441 | * take a reference count to prevent the anon_vma disappearing | |
1442 | */ | |
1443 | anon_vma = page_anon_vma(page); | |
1444 | if (!anon_vma) | |
1445 | return NULL; | |
1446 | ||
1447 | anon_vma_lock_read(anon_vma); | |
1448 | return anon_vma; | |
1449 | } | |
1450 | ||
1451 | /* | |
1452 | * rmap_walk_anon - do something to anonymous page using the object-based | |
1453 | * rmap method | |
1454 | * @page: the page to be handled | |
1455 | * @rwc: control variable according to each walk type | |
1456 | * | |
1457 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
1458 | * contained in the anon_vma struct it points to. | |
1459 | * | |
1460 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma | |
1461 | * where the page was found will be held for write. So, we won't recheck | |
1462 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be | |
1463 | * LOCKED. | |
1464 | */ | |
1465 | static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc) | |
1466 | { | |
1467 | struct anon_vma *anon_vma; | |
1468 | pgoff_t pgoff; | |
1469 | struct anon_vma_chain *avc; | |
1470 | int ret = SWAP_AGAIN; | |
1471 | ||
1472 | anon_vma = rmap_walk_anon_lock(page, rwc); | |
1473 | if (!anon_vma) | |
1474 | return ret; | |
1475 | ||
1476 | pgoff = page_to_pgoff(page); | |
1477 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { | |
1478 | struct vm_area_struct *vma = avc->vma; | |
1479 | unsigned long address = vma_address(page, vma); | |
1480 | ||
1481 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) | |
1482 | continue; | |
1483 | ||
1484 | ret = rwc->rmap_one(page, vma, address, rwc->arg); | |
1485 | if (ret != SWAP_AGAIN) | |
1486 | break; | |
1487 | if (rwc->done && rwc->done(page)) | |
1488 | break; | |
1489 | } | |
1490 | anon_vma_unlock_read(anon_vma); | |
1491 | return ret; | |
1492 | } | |
1493 | ||
1494 | /* | |
1495 | * rmap_walk_file - do something to file page using the object-based rmap method | |
1496 | * @page: the page to be handled | |
1497 | * @rwc: control variable according to each walk type | |
1498 | * | |
1499 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
1500 | * contained in the address_space struct it points to. | |
1501 | * | |
1502 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma | |
1503 | * where the page was found will be held for write. So, we won't recheck | |
1504 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be | |
1505 | * LOCKED. | |
1506 | */ | |
1507 | static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc) | |
1508 | { | |
1509 | struct address_space *mapping = page->mapping; | |
1510 | pgoff_t pgoff; | |
1511 | struct vm_area_struct *vma; | |
1512 | int ret = SWAP_AGAIN; | |
1513 | ||
1514 | /* | |
1515 | * The page lock not only makes sure that page->mapping cannot | |
1516 | * suddenly be NULLified by truncation, it makes sure that the | |
1517 | * structure at mapping cannot be freed and reused yet, | |
1518 | * so we can safely take mapping->i_mmap_rwsem. | |
1519 | */ | |
1520 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
1521 | ||
1522 | if (!mapping) | |
1523 | return ret; | |
1524 | ||
1525 | pgoff = page_to_pgoff(page); | |
1526 | i_mmap_lock_read(mapping); | |
1527 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { | |
1528 | unsigned long address = vma_address(page, vma); | |
1529 | ||
1530 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) | |
1531 | continue; | |
1532 | ||
1533 | ret = rwc->rmap_one(page, vma, address, rwc->arg); | |
1534 | if (ret != SWAP_AGAIN) | |
1535 | goto done; | |
1536 | if (rwc->done && rwc->done(page)) | |
1537 | goto done; | |
1538 | } | |
1539 | ||
1540 | done: | |
1541 | i_mmap_unlock_read(mapping); | |
1542 | return ret; | |
1543 | } | |
1544 | ||
1545 | int rmap_walk(struct page *page, struct rmap_walk_control *rwc) | |
1546 | { | |
1547 | if (unlikely(PageKsm(page))) | |
1548 | return rmap_walk_ksm(page, rwc); | |
1549 | else if (PageAnon(page)) | |
1550 | return rmap_walk_anon(page, rwc); | |
1551 | else | |
1552 | return rmap_walk_file(page, rwc); | |
1553 | } | |
1554 | ||
1555 | #ifdef CONFIG_HUGETLB_PAGE | |
1556 | /* | |
1557 | * The following three functions are for anonymous (private mapped) hugepages. | |
1558 | * Unlike common anonymous pages, anonymous hugepages have no accounting code | |
1559 | * and no lru code, because we handle hugepages differently from common pages. | |
1560 | */ | |
1561 | static void __hugepage_set_anon_rmap(struct page *page, | |
1562 | struct vm_area_struct *vma, unsigned long address, int exclusive) | |
1563 | { | |
1564 | struct anon_vma *anon_vma = vma->anon_vma; | |
1565 | ||
1566 | BUG_ON(!anon_vma); | |
1567 | ||
1568 | if (PageAnon(page)) | |
1569 | return; | |
1570 | if (!exclusive) | |
1571 | anon_vma = anon_vma->root; | |
1572 | ||
1573 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
1574 | page->mapping = (struct address_space *) anon_vma; | |
1575 | page->index = linear_page_index(vma, address); | |
1576 | } | |
1577 | ||
1578 | void hugepage_add_anon_rmap(struct page *page, | |
1579 | struct vm_area_struct *vma, unsigned long address) | |
1580 | { | |
1581 | struct anon_vma *anon_vma = vma->anon_vma; | |
1582 | int first; | |
1583 | ||
1584 | BUG_ON(!PageLocked(page)); | |
1585 | BUG_ON(!anon_vma); | |
1586 | /* address might be in next vma when migration races vma_adjust */ | |
1587 | first = atomic_inc_and_test(&page->_mapcount); | |
1588 | if (first) | |
1589 | __hugepage_set_anon_rmap(page, vma, address, 0); | |
1590 | } | |
1591 | ||
1592 | void hugepage_add_new_anon_rmap(struct page *page, | |
1593 | struct vm_area_struct *vma, unsigned long address) | |
1594 | { | |
1595 | BUG_ON(address < vma->vm_start || address >= vma->vm_end); | |
1596 | atomic_set(&page->_mapcount, 0); | |
1597 | __hugepage_set_anon_rmap(page, vma, address, 1); | |
1598 | } | |
1599 | #endif /* CONFIG_HUGETLB_PAGE */ |