]>
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 | * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share) | |
27 | * mapping->i_mmap_rwsem | |
28 | * anon_vma->rwsem | |
29 | * mm->page_table_lock or pte_lock | |
30 | * zone_lru_lock (in mark_page_accessed, isolate_lru_page) | |
31 | * swap_lock (in swap_duplicate, swap_info_get) | |
32 | * mmlist_lock (in mmput, drain_mmlist and others) | |
33 | * mapping->private_lock (in __set_page_dirty_buffers) | |
34 | * mem_cgroup_{begin,end}_page_stat (memcg->move_lock) | |
35 | * mapping->tree_lock (widely used) | |
36 | * inode->i_lock (in set_page_dirty's __mark_inode_dirty) | |
37 | * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty) | |
38 | * sb_lock (within inode_lock in fs/fs-writeback.c) | |
39 | * mapping->tree_lock (widely used, in set_page_dirty, | |
40 | * in arch-dependent flush_dcache_mmap_lock, | |
41 | * within bdi.wb->list_lock in __sync_single_inode) | |
42 | * | |
43 | * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon) | |
44 | * ->tasklist_lock | |
45 | * pte map lock | |
46 | */ | |
47 | ||
48 | #include <linux/mm.h> | |
49 | #include <linux/sched/mm.h> | |
50 | #include <linux/pagemap.h> | |
51 | #include <linux/swap.h> | |
52 | #include <linux/swapops.h> | |
53 | #include <linux/slab.h> | |
54 | #include <linux/init.h> | |
55 | #include <linux/ksm.h> | |
56 | #include <linux/rmap.h> | |
57 | #include <linux/rcupdate.h> | |
58 | #include <linux/export.h> | |
59 | #include <linux/memcontrol.h> | |
60 | #include <linux/mmu_notifier.h> | |
61 | #include <linux/migrate.h> | |
62 | #include <linux/hugetlb.h> | |
63 | #include <linux/backing-dev.h> | |
64 | #include <linux/page_idle.h> | |
65 | ||
66 | #include <asm/tlbflush.h> | |
67 | ||
68 | #include <trace/events/tlb.h> | |
69 | ||
70 | #include "internal.h" | |
71 | ||
72 | static struct kmem_cache *anon_vma_cachep; | |
73 | static struct kmem_cache *anon_vma_chain_cachep; | |
74 | ||
75 | static inline struct anon_vma *anon_vma_alloc(void) | |
76 | { | |
77 | struct anon_vma *anon_vma; | |
78 | ||
79 | anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); | |
80 | if (anon_vma) { | |
81 | atomic_set(&anon_vma->refcount, 1); | |
82 | anon_vma->degree = 1; /* Reference for first vma */ | |
83 | anon_vma->parent = anon_vma; | |
84 | /* | |
85 | * Initialise the anon_vma root to point to itself. If called | |
86 | * from fork, the root will be reset to the parents anon_vma. | |
87 | */ | |
88 | anon_vma->root = anon_vma; | |
89 | } | |
90 | ||
91 | return anon_vma; | |
92 | } | |
93 | ||
94 | static inline void anon_vma_free(struct anon_vma *anon_vma) | |
95 | { | |
96 | VM_BUG_ON(atomic_read(&anon_vma->refcount)); | |
97 | ||
98 | /* | |
99 | * Synchronize against page_lock_anon_vma_read() such that | |
100 | * we can safely hold the lock without the anon_vma getting | |
101 | * freed. | |
102 | * | |
103 | * Relies on the full mb implied by the atomic_dec_and_test() from | |
104 | * put_anon_vma() against the acquire barrier implied by | |
105 | * down_read_trylock() from page_lock_anon_vma_read(). This orders: | |
106 | * | |
107 | * page_lock_anon_vma_read() VS put_anon_vma() | |
108 | * down_read_trylock() atomic_dec_and_test() | |
109 | * LOCK MB | |
110 | * atomic_read() rwsem_is_locked() | |
111 | * | |
112 | * LOCK should suffice since the actual taking of the lock must | |
113 | * happen _before_ what follows. | |
114 | */ | |
115 | might_sleep(); | |
116 | if (rwsem_is_locked(&anon_vma->root->rwsem)) { | |
117 | anon_vma_lock_write(anon_vma); | |
118 | anon_vma_unlock_write(anon_vma); | |
119 | } | |
120 | ||
121 | kmem_cache_free(anon_vma_cachep, anon_vma); | |
122 | } | |
123 | ||
124 | static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp) | |
125 | { | |
126 | return kmem_cache_alloc(anon_vma_chain_cachep, gfp); | |
127 | } | |
128 | ||
129 | static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) | |
130 | { | |
131 | kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); | |
132 | } | |
133 | ||
134 | static void anon_vma_chain_link(struct vm_area_struct *vma, | |
135 | struct anon_vma_chain *avc, | |
136 | struct anon_vma *anon_vma) | |
137 | { | |
138 | avc->vma = vma; | |
139 | avc->anon_vma = anon_vma; | |
140 | list_add(&avc->same_vma, &vma->anon_vma_chain); | |
141 | anon_vma_interval_tree_insert(avc, &anon_vma->rb_root); | |
142 | } | |
143 | ||
144 | /** | |
145 | * __anon_vma_prepare - attach an anon_vma to a memory region | |
146 | * @vma: the memory region in question | |
147 | * | |
148 | * This makes sure the memory mapping described by 'vma' has | |
149 | * an 'anon_vma' attached to it, so that we can associate the | |
150 | * anonymous pages mapped into it with that anon_vma. | |
151 | * | |
152 | * The common case will be that we already have one, which | |
153 | * is handled inline by anon_vma_prepare(). But if | |
154 | * not we either need to find an adjacent mapping that we | |
155 | * can re-use the anon_vma from (very common when the only | |
156 | * reason for splitting a vma has been mprotect()), or we | |
157 | * allocate a new one. | |
158 | * | |
159 | * Anon-vma allocations are very subtle, because we may have | |
160 | * optimistically looked up an anon_vma in page_lock_anon_vma_read() | |
161 | * and that may actually touch the spinlock even in the newly | |
162 | * allocated vma (it depends on RCU to make sure that the | |
163 | * anon_vma isn't actually destroyed). | |
164 | * | |
165 | * As a result, we need to do proper anon_vma locking even | |
166 | * for the new allocation. At the same time, we do not want | |
167 | * to do any locking for the common case of already having | |
168 | * an anon_vma. | |
169 | * | |
170 | * This must be called with the mmap_sem held for reading. | |
171 | */ | |
172 | int __anon_vma_prepare(struct vm_area_struct *vma) | |
173 | { | |
174 | struct mm_struct *mm = vma->vm_mm; | |
175 | struct anon_vma *anon_vma, *allocated; | |
176 | struct anon_vma_chain *avc; | |
177 | ||
178 | might_sleep(); | |
179 | ||
180 | avc = anon_vma_chain_alloc(GFP_KERNEL); | |
181 | if (!avc) | |
182 | goto out_enomem; | |
183 | ||
184 | anon_vma = find_mergeable_anon_vma(vma); | |
185 | allocated = NULL; | |
186 | if (!anon_vma) { | |
187 | anon_vma = anon_vma_alloc(); | |
188 | if (unlikely(!anon_vma)) | |
189 | goto out_enomem_free_avc; | |
190 | allocated = anon_vma; | |
191 | } | |
192 | ||
193 | anon_vma_lock_write(anon_vma); | |
194 | /* page_table_lock to protect against threads */ | |
195 | spin_lock(&mm->page_table_lock); | |
196 | if (likely(!vma->anon_vma)) { | |
197 | vma->anon_vma = anon_vma; | |
198 | anon_vma_chain_link(vma, avc, anon_vma); | |
199 | /* vma reference or self-parent link for new root */ | |
200 | anon_vma->degree++; | |
201 | allocated = NULL; | |
202 | avc = NULL; | |
203 | } | |
204 | spin_unlock(&mm->page_table_lock); | |
205 | anon_vma_unlock_write(anon_vma); | |
206 | ||
207 | if (unlikely(allocated)) | |
208 | put_anon_vma(allocated); | |
209 | if (unlikely(avc)) | |
210 | anon_vma_chain_free(avc); | |
211 | ||
212 | return 0; | |
213 | ||
214 | out_enomem_free_avc: | |
215 | anon_vma_chain_free(avc); | |
216 | out_enomem: | |
217 | return -ENOMEM; | |
218 | } | |
219 | ||
220 | /* | |
221 | * This is a useful helper function for locking the anon_vma root as | |
222 | * we traverse the vma->anon_vma_chain, looping over anon_vma's that | |
223 | * have the same vma. | |
224 | * | |
225 | * Such anon_vma's should have the same root, so you'd expect to see | |
226 | * just a single mutex_lock for the whole traversal. | |
227 | */ | |
228 | static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma) | |
229 | { | |
230 | struct anon_vma *new_root = anon_vma->root; | |
231 | if (new_root != root) { | |
232 | if (WARN_ON_ONCE(root)) | |
233 | up_write(&root->rwsem); | |
234 | root = new_root; | |
235 | down_write(&root->rwsem); | |
236 | } | |
237 | return root; | |
238 | } | |
239 | ||
240 | static inline void unlock_anon_vma_root(struct anon_vma *root) | |
241 | { | |
242 | if (root) | |
243 | up_write(&root->rwsem); | |
244 | } | |
245 | ||
246 | /* | |
247 | * Attach the anon_vmas from src to dst. | |
248 | * Returns 0 on success, -ENOMEM on failure. | |
249 | * | |
250 | * If dst->anon_vma is NULL this function tries to find and reuse existing | |
251 | * anon_vma which has no vmas and only one child anon_vma. This prevents | |
252 | * degradation of anon_vma hierarchy to endless linear chain in case of | |
253 | * constantly forking task. On the other hand, an anon_vma with more than one | |
254 | * child isn't reused even if there was no alive vma, thus rmap walker has a | |
255 | * good chance of avoiding scanning the whole hierarchy when it searches where | |
256 | * page is mapped. | |
257 | */ | |
258 | int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) | |
259 | { | |
260 | struct anon_vma_chain *avc, *pavc; | |
261 | struct anon_vma *root = NULL; | |
262 | ||
263 | list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { | |
264 | struct anon_vma *anon_vma; | |
265 | ||
266 | avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN); | |
267 | if (unlikely(!avc)) { | |
268 | unlock_anon_vma_root(root); | |
269 | root = NULL; | |
270 | avc = anon_vma_chain_alloc(GFP_KERNEL); | |
271 | if (!avc) | |
272 | goto enomem_failure; | |
273 | } | |
274 | anon_vma = pavc->anon_vma; | |
275 | root = lock_anon_vma_root(root, anon_vma); | |
276 | anon_vma_chain_link(dst, avc, anon_vma); | |
277 | ||
278 | /* | |
279 | * Reuse existing anon_vma if its degree lower than two, | |
280 | * that means it has no vma and only one anon_vma child. | |
281 | * | |
282 | * Do not chose parent anon_vma, otherwise first child | |
283 | * will always reuse it. Root anon_vma is never reused: | |
284 | * it has self-parent reference and at least one child. | |
285 | */ | |
286 | if (!dst->anon_vma && anon_vma != src->anon_vma && | |
287 | anon_vma->degree < 2) | |
288 | dst->anon_vma = anon_vma; | |
289 | } | |
290 | if (dst->anon_vma) | |
291 | dst->anon_vma->degree++; | |
292 | unlock_anon_vma_root(root); | |
293 | return 0; | |
294 | ||
295 | enomem_failure: | |
296 | /* | |
297 | * dst->anon_vma is dropped here otherwise its degree can be incorrectly | |
298 | * decremented in unlink_anon_vmas(). | |
299 | * We can safely do this because callers of anon_vma_clone() don't care | |
300 | * about dst->anon_vma if anon_vma_clone() failed. | |
301 | */ | |
302 | dst->anon_vma = NULL; | |
303 | unlink_anon_vmas(dst); | |
304 | return -ENOMEM; | |
305 | } | |
306 | ||
307 | /* | |
308 | * Attach vma to its own anon_vma, as well as to the anon_vmas that | |
309 | * the corresponding VMA in the parent process is attached to. | |
310 | * Returns 0 on success, non-zero on failure. | |
311 | */ | |
312 | int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) | |
313 | { | |
314 | struct anon_vma_chain *avc; | |
315 | struct anon_vma *anon_vma; | |
316 | int error; | |
317 | ||
318 | /* Don't bother if the parent process has no anon_vma here. */ | |
319 | if (!pvma->anon_vma) | |
320 | return 0; | |
321 | ||
322 | /* Drop inherited anon_vma, we'll reuse existing or allocate new. */ | |
323 | vma->anon_vma = NULL; | |
324 | ||
325 | /* | |
326 | * First, attach the new VMA to the parent VMA's anon_vmas, | |
327 | * so rmap can find non-COWed pages in child processes. | |
328 | */ | |
329 | error = anon_vma_clone(vma, pvma); | |
330 | if (error) | |
331 | return error; | |
332 | ||
333 | /* An existing anon_vma has been reused, all done then. */ | |
334 | if (vma->anon_vma) | |
335 | return 0; | |
336 | ||
337 | /* Then add our own anon_vma. */ | |
338 | anon_vma = anon_vma_alloc(); | |
339 | if (!anon_vma) | |
340 | goto out_error; | |
341 | avc = anon_vma_chain_alloc(GFP_KERNEL); | |
342 | if (!avc) | |
343 | goto out_error_free_anon_vma; | |
344 | ||
345 | /* | |
346 | * The root anon_vma's spinlock is the lock actually used when we | |
347 | * lock any of the anon_vmas in this anon_vma tree. | |
348 | */ | |
349 | anon_vma->root = pvma->anon_vma->root; | |
350 | anon_vma->parent = pvma->anon_vma; | |
351 | /* | |
352 | * With refcounts, an anon_vma can stay around longer than the | |
353 | * process it belongs to. The root anon_vma needs to be pinned until | |
354 | * this anon_vma is freed, because the lock lives in the root. | |
355 | */ | |
356 | get_anon_vma(anon_vma->root); | |
357 | /* Mark this anon_vma as the one where our new (COWed) pages go. */ | |
358 | vma->anon_vma = anon_vma; | |
359 | anon_vma_lock_write(anon_vma); | |
360 | anon_vma_chain_link(vma, avc, anon_vma); | |
361 | anon_vma->parent->degree++; | |
362 | anon_vma_unlock_write(anon_vma); | |
363 | ||
364 | return 0; | |
365 | ||
366 | out_error_free_anon_vma: | |
367 | put_anon_vma(anon_vma); | |
368 | out_error: | |
369 | unlink_anon_vmas(vma); | |
370 | return -ENOMEM; | |
371 | } | |
372 | ||
373 | void unlink_anon_vmas(struct vm_area_struct *vma) | |
374 | { | |
375 | struct anon_vma_chain *avc, *next; | |
376 | struct anon_vma *root = NULL; | |
377 | ||
378 | /* | |
379 | * Unlink each anon_vma chained to the VMA. This list is ordered | |
380 | * from newest to oldest, ensuring the root anon_vma gets freed last. | |
381 | */ | |
382 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { | |
383 | struct anon_vma *anon_vma = avc->anon_vma; | |
384 | ||
385 | root = lock_anon_vma_root(root, anon_vma); | |
386 | anon_vma_interval_tree_remove(avc, &anon_vma->rb_root); | |
387 | ||
388 | /* | |
389 | * Leave empty anon_vmas on the list - we'll need | |
390 | * to free them outside the lock. | |
391 | */ | |
392 | if (RB_EMPTY_ROOT(&anon_vma->rb_root)) { | |
393 | anon_vma->parent->degree--; | |
394 | continue; | |
395 | } | |
396 | ||
397 | list_del(&avc->same_vma); | |
398 | anon_vma_chain_free(avc); | |
399 | } | |
400 | if (vma->anon_vma) | |
401 | vma->anon_vma->degree--; | |
402 | unlock_anon_vma_root(root); | |
403 | ||
404 | /* | |
405 | * Iterate the list once more, it now only contains empty and unlinked | |
406 | * anon_vmas, destroy them. Could not do before due to __put_anon_vma() | |
407 | * needing to write-acquire the anon_vma->root->rwsem. | |
408 | */ | |
409 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { | |
410 | struct anon_vma *anon_vma = avc->anon_vma; | |
411 | ||
412 | VM_WARN_ON(anon_vma->degree); | |
413 | put_anon_vma(anon_vma); | |
414 | ||
415 | list_del(&avc->same_vma); | |
416 | anon_vma_chain_free(avc); | |
417 | } | |
418 | } | |
419 | ||
420 | static void anon_vma_ctor(void *data) | |
421 | { | |
422 | struct anon_vma *anon_vma = data; | |
423 | ||
424 | init_rwsem(&anon_vma->rwsem); | |
425 | atomic_set(&anon_vma->refcount, 0); | |
426 | anon_vma->rb_root = RB_ROOT; | |
427 | } | |
428 | ||
429 | void __init anon_vma_init(void) | |
430 | { | |
431 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
432 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT, | |
433 | anon_vma_ctor); | |
434 | anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, | |
435 | SLAB_PANIC|SLAB_ACCOUNT); | |
436 | } | |
437 | ||
438 | /* | |
439 | * Getting a lock on a stable anon_vma from a page off the LRU is tricky! | |
440 | * | |
441 | * Since there is no serialization what so ever against page_remove_rmap() | |
442 | * the best this function can do is return a locked anon_vma that might | |
443 | * have been relevant to this page. | |
444 | * | |
445 | * The page might have been remapped to a different anon_vma or the anon_vma | |
446 | * returned may already be freed (and even reused). | |
447 | * | |
448 | * In case it was remapped to a different anon_vma, the new anon_vma will be a | |
449 | * child of the old anon_vma, and the anon_vma lifetime rules will therefore | |
450 | * ensure that any anon_vma obtained from the page will still be valid for as | |
451 | * long as we observe page_mapped() [ hence all those page_mapped() tests ]. | |
452 | * | |
453 | * All users of this function must be very careful when walking the anon_vma | |
454 | * chain and verify that the page in question is indeed mapped in it | |
455 | * [ something equivalent to page_mapped_in_vma() ]. | |
456 | * | |
457 | * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap() | |
458 | * that the anon_vma pointer from page->mapping is valid if there is a | |
459 | * mapcount, we can dereference the anon_vma after observing those. | |
460 | */ | |
461 | struct anon_vma *page_get_anon_vma(struct page *page) | |
462 | { | |
463 | struct anon_vma *anon_vma = NULL; | |
464 | unsigned long anon_mapping; | |
465 | ||
466 | rcu_read_lock(); | |
467 | anon_mapping = (unsigned long)READ_ONCE(page->mapping); | |
468 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) | |
469 | goto out; | |
470 | if (!page_mapped(page)) | |
471 | goto out; | |
472 | ||
473 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
474 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { | |
475 | anon_vma = NULL; | |
476 | goto out; | |
477 | } | |
478 | ||
479 | /* | |
480 | * If this page is still mapped, then its anon_vma cannot have been | |
481 | * freed. But if it has been unmapped, we have no security against the | |
482 | * anon_vma structure being freed and reused (for another anon_vma: | |
483 | * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero() | |
484 | * above cannot corrupt). | |
485 | */ | |
486 | if (!page_mapped(page)) { | |
487 | rcu_read_unlock(); | |
488 | put_anon_vma(anon_vma); | |
489 | return NULL; | |
490 | } | |
491 | out: | |
492 | rcu_read_unlock(); | |
493 | ||
494 | return anon_vma; | |
495 | } | |
496 | ||
497 | /* | |
498 | * Similar to page_get_anon_vma() except it locks the anon_vma. | |
499 | * | |
500 | * Its a little more complex as it tries to keep the fast path to a single | |
501 | * atomic op -- the trylock. If we fail the trylock, we fall back to getting a | |
502 | * reference like with page_get_anon_vma() and then block on the mutex. | |
503 | */ | |
504 | struct anon_vma *page_lock_anon_vma_read(struct page *page) | |
505 | { | |
506 | struct anon_vma *anon_vma = NULL; | |
507 | struct anon_vma *root_anon_vma; | |
508 | unsigned long anon_mapping; | |
509 | ||
510 | rcu_read_lock(); | |
511 | anon_mapping = (unsigned long)READ_ONCE(page->mapping); | |
512 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) | |
513 | goto out; | |
514 | if (!page_mapped(page)) | |
515 | goto out; | |
516 | ||
517 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
518 | root_anon_vma = READ_ONCE(anon_vma->root); | |
519 | if (down_read_trylock(&root_anon_vma->rwsem)) { | |
520 | /* | |
521 | * If the page is still mapped, then this anon_vma is still | |
522 | * its anon_vma, and holding the mutex ensures that it will | |
523 | * not go away, see anon_vma_free(). | |
524 | */ | |
525 | if (!page_mapped(page)) { | |
526 | up_read(&root_anon_vma->rwsem); | |
527 | anon_vma = NULL; | |
528 | } | |
529 | goto out; | |
530 | } | |
531 | ||
532 | /* trylock failed, we got to sleep */ | |
533 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { | |
534 | anon_vma = NULL; | |
535 | goto out; | |
536 | } | |
537 | ||
538 | if (!page_mapped(page)) { | |
539 | rcu_read_unlock(); | |
540 | put_anon_vma(anon_vma); | |
541 | return NULL; | |
542 | } | |
543 | ||
544 | /* we pinned the anon_vma, its safe to sleep */ | |
545 | rcu_read_unlock(); | |
546 | anon_vma_lock_read(anon_vma); | |
547 | ||
548 | if (atomic_dec_and_test(&anon_vma->refcount)) { | |
549 | /* | |
550 | * Oops, we held the last refcount, release the lock | |
551 | * and bail -- can't simply use put_anon_vma() because | |
552 | * we'll deadlock on the anon_vma_lock_write() recursion. | |
553 | */ | |
554 | anon_vma_unlock_read(anon_vma); | |
555 | __put_anon_vma(anon_vma); | |
556 | anon_vma = NULL; | |
557 | } | |
558 | ||
559 | return anon_vma; | |
560 | ||
561 | out: | |
562 | rcu_read_unlock(); | |
563 | return anon_vma; | |
564 | } | |
565 | ||
566 | void page_unlock_anon_vma_read(struct anon_vma *anon_vma) | |
567 | { | |
568 | anon_vma_unlock_read(anon_vma); | |
569 | } | |
570 | ||
571 | #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH | |
572 | /* | |
573 | * Flush TLB entries for recently unmapped pages from remote CPUs. It is | |
574 | * important if a PTE was dirty when it was unmapped that it's flushed | |
575 | * before any IO is initiated on the page to prevent lost writes. Similarly, | |
576 | * it must be flushed before freeing to prevent data leakage. | |
577 | */ | |
578 | void try_to_unmap_flush(void) | |
579 | { | |
580 | struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; | |
581 | int cpu; | |
582 | ||
583 | if (!tlb_ubc->flush_required) | |
584 | return; | |
585 | ||
586 | cpu = get_cpu(); | |
587 | ||
588 | if (cpumask_test_cpu(cpu, &tlb_ubc->cpumask)) { | |
589 | count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL); | |
590 | local_flush_tlb(); | |
591 | trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL); | |
592 | } | |
593 | ||
594 | if (cpumask_any_but(&tlb_ubc->cpumask, cpu) < nr_cpu_ids) | |
595 | flush_tlb_others(&tlb_ubc->cpumask, NULL, 0, TLB_FLUSH_ALL); | |
596 | cpumask_clear(&tlb_ubc->cpumask); | |
597 | tlb_ubc->flush_required = false; | |
598 | tlb_ubc->writable = false; | |
599 | put_cpu(); | |
600 | } | |
601 | ||
602 | /* Flush iff there are potentially writable TLB entries that can race with IO */ | |
603 | void try_to_unmap_flush_dirty(void) | |
604 | { | |
605 | struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; | |
606 | ||
607 | if (tlb_ubc->writable) | |
608 | try_to_unmap_flush(); | |
609 | } | |
610 | ||
611 | static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable) | |
612 | { | |
613 | struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; | |
614 | ||
615 | cpumask_or(&tlb_ubc->cpumask, &tlb_ubc->cpumask, mm_cpumask(mm)); | |
616 | tlb_ubc->flush_required = true; | |
617 | ||
618 | /* | |
619 | * If the PTE was dirty then it's best to assume it's writable. The | |
620 | * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush() | |
621 | * before the page is queued for IO. | |
622 | */ | |
623 | if (writable) | |
624 | tlb_ubc->writable = true; | |
625 | } | |
626 | ||
627 | /* | |
628 | * Returns true if the TLB flush should be deferred to the end of a batch of | |
629 | * unmap operations to reduce IPIs. | |
630 | */ | |
631 | static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) | |
632 | { | |
633 | bool should_defer = false; | |
634 | ||
635 | if (!(flags & TTU_BATCH_FLUSH)) | |
636 | return false; | |
637 | ||
638 | /* If remote CPUs need to be flushed then defer batch the flush */ | |
639 | if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids) | |
640 | should_defer = true; | |
641 | put_cpu(); | |
642 | ||
643 | return should_defer; | |
644 | } | |
645 | #else | |
646 | static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable) | |
647 | { | |
648 | } | |
649 | ||
650 | static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) | |
651 | { | |
652 | return false; | |
653 | } | |
654 | #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */ | |
655 | ||
656 | /* | |
657 | * At what user virtual address is page expected in vma? | |
658 | * Caller should check the page is actually part of the vma. | |
659 | */ | |
660 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
661 | { | |
662 | unsigned long address; | |
663 | if (PageAnon(page)) { | |
664 | struct anon_vma *page__anon_vma = page_anon_vma(page); | |
665 | /* | |
666 | * Note: swapoff's unuse_vma() is more efficient with this | |
667 | * check, and needs it to match anon_vma when KSM is active. | |
668 | */ | |
669 | if (!vma->anon_vma || !page__anon_vma || | |
670 | vma->anon_vma->root != page__anon_vma->root) | |
671 | return -EFAULT; | |
672 | } else if (page->mapping) { | |
673 | if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping) | |
674 | return -EFAULT; | |
675 | } else | |
676 | return -EFAULT; | |
677 | address = __vma_address(page, vma); | |
678 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) | |
679 | return -EFAULT; | |
680 | return address; | |
681 | } | |
682 | ||
683 | pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address) | |
684 | { | |
685 | pgd_t *pgd; | |
686 | pud_t *pud; | |
687 | pmd_t *pmd = NULL; | |
688 | pmd_t pmde; | |
689 | ||
690 | pgd = pgd_offset(mm, address); | |
691 | if (!pgd_present(*pgd)) | |
692 | goto out; | |
693 | ||
694 | pud = pud_offset(pgd, address); | |
695 | if (!pud_present(*pud)) | |
696 | goto out; | |
697 | ||
698 | pmd = pmd_offset(pud, address); | |
699 | /* | |
700 | * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at() | |
701 | * without holding anon_vma lock for write. So when looking for a | |
702 | * genuine pmde (in which to find pte), test present and !THP together. | |
703 | */ | |
704 | pmde = *pmd; | |
705 | barrier(); | |
706 | if (!pmd_present(pmde) || pmd_trans_huge(pmde)) | |
707 | pmd = NULL; | |
708 | out: | |
709 | return pmd; | |
710 | } | |
711 | ||
712 | struct page_referenced_arg { | |
713 | int mapcount; | |
714 | int referenced; | |
715 | unsigned long vm_flags; | |
716 | struct mem_cgroup *memcg; | |
717 | }; | |
718 | /* | |
719 | * arg: page_referenced_arg will be passed | |
720 | */ | |
721 | static int page_referenced_one(struct page *page, struct vm_area_struct *vma, | |
722 | unsigned long address, void *arg) | |
723 | { | |
724 | struct page_referenced_arg *pra = arg; | |
725 | struct page_vma_mapped_walk pvmw = { | |
726 | .page = page, | |
727 | .vma = vma, | |
728 | .address = address, | |
729 | }; | |
730 | int referenced = 0; | |
731 | ||
732 | while (page_vma_mapped_walk(&pvmw)) { | |
733 | address = pvmw.address; | |
734 | ||
735 | if (vma->vm_flags & VM_LOCKED) { | |
736 | page_vma_mapped_walk_done(&pvmw); | |
737 | pra->vm_flags |= VM_LOCKED; | |
738 | return SWAP_FAIL; /* To break the loop */ | |
739 | } | |
740 | ||
741 | if (pvmw.pte) { | |
742 | if (ptep_clear_flush_young_notify(vma, address, | |
743 | pvmw.pte)) { | |
744 | /* | |
745 | * Don't treat a reference through | |
746 | * a sequentially read mapping as such. | |
747 | * If the page has been used in another mapping, | |
748 | * we will catch it; if this other mapping is | |
749 | * already gone, the unmap path will have set | |
750 | * PG_referenced or activated the page. | |
751 | */ | |
752 | if (likely(!(vma->vm_flags & VM_SEQ_READ))) | |
753 | referenced++; | |
754 | } | |
755 | } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) { | |
756 | if (pmdp_clear_flush_young_notify(vma, address, | |
757 | pvmw.pmd)) | |
758 | referenced++; | |
759 | } else { | |
760 | /* unexpected pmd-mapped page? */ | |
761 | WARN_ON_ONCE(1); | |
762 | } | |
763 | ||
764 | pra->mapcount--; | |
765 | } | |
766 | ||
767 | if (referenced) | |
768 | clear_page_idle(page); | |
769 | if (test_and_clear_page_young(page)) | |
770 | referenced++; | |
771 | ||
772 | if (referenced) { | |
773 | pra->referenced++; | |
774 | pra->vm_flags |= vma->vm_flags; | |
775 | } | |
776 | ||
777 | if (!pra->mapcount) | |
778 | return SWAP_SUCCESS; /* To break the loop */ | |
779 | ||
780 | return SWAP_AGAIN; | |
781 | } | |
782 | ||
783 | static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg) | |
784 | { | |
785 | struct page_referenced_arg *pra = arg; | |
786 | struct mem_cgroup *memcg = pra->memcg; | |
787 | ||
788 | if (!mm_match_cgroup(vma->vm_mm, memcg)) | |
789 | return true; | |
790 | ||
791 | return false; | |
792 | } | |
793 | ||
794 | /** | |
795 | * page_referenced - test if the page was referenced | |
796 | * @page: the page to test | |
797 | * @is_locked: caller holds lock on the page | |
798 | * @memcg: target memory cgroup | |
799 | * @vm_flags: collect encountered vma->vm_flags who actually referenced the page | |
800 | * | |
801 | * Quick test_and_clear_referenced for all mappings to a page, | |
802 | * returns the number of ptes which referenced the page. | |
803 | */ | |
804 | int page_referenced(struct page *page, | |
805 | int is_locked, | |
806 | struct mem_cgroup *memcg, | |
807 | unsigned long *vm_flags) | |
808 | { | |
809 | int ret; | |
810 | int we_locked = 0; | |
811 | struct page_referenced_arg pra = { | |
812 | .mapcount = total_mapcount(page), | |
813 | .memcg = memcg, | |
814 | }; | |
815 | struct rmap_walk_control rwc = { | |
816 | .rmap_one = page_referenced_one, | |
817 | .arg = (void *)&pra, | |
818 | .anon_lock = page_lock_anon_vma_read, | |
819 | }; | |
820 | ||
821 | *vm_flags = 0; | |
822 | if (!page_mapped(page)) | |
823 | return 0; | |
824 | ||
825 | if (!page_rmapping(page)) | |
826 | return 0; | |
827 | ||
828 | if (!is_locked && (!PageAnon(page) || PageKsm(page))) { | |
829 | we_locked = trylock_page(page); | |
830 | if (!we_locked) | |
831 | return 1; | |
832 | } | |
833 | ||
834 | /* | |
835 | * If we are reclaiming on behalf of a cgroup, skip | |
836 | * counting on behalf of references from different | |
837 | * cgroups | |
838 | */ | |
839 | if (memcg) { | |
840 | rwc.invalid_vma = invalid_page_referenced_vma; | |
841 | } | |
842 | ||
843 | ret = rmap_walk(page, &rwc); | |
844 | *vm_flags = pra.vm_flags; | |
845 | ||
846 | if (we_locked) | |
847 | unlock_page(page); | |
848 | ||
849 | return pra.referenced; | |
850 | } | |
851 | ||
852 | static int page_mkclean_one(struct page *page, struct vm_area_struct *vma, | |
853 | unsigned long address, void *arg) | |
854 | { | |
855 | struct page_vma_mapped_walk pvmw = { | |
856 | .page = page, | |
857 | .vma = vma, | |
858 | .address = address, | |
859 | .flags = PVMW_SYNC, | |
860 | }; | |
861 | int *cleaned = arg; | |
862 | ||
863 | while (page_vma_mapped_walk(&pvmw)) { | |
864 | int ret = 0; | |
865 | address = pvmw.address; | |
866 | if (pvmw.pte) { | |
867 | pte_t entry; | |
868 | pte_t *pte = pvmw.pte; | |
869 | ||
870 | if (!pte_dirty(*pte) && !pte_write(*pte)) | |
871 | continue; | |
872 | ||
873 | flush_cache_page(vma, address, pte_pfn(*pte)); | |
874 | entry = ptep_clear_flush(vma, address, pte); | |
875 | entry = pte_wrprotect(entry); | |
876 | entry = pte_mkclean(entry); | |
877 | set_pte_at(vma->vm_mm, address, pte, entry); | |
878 | ret = 1; | |
879 | } else { | |
880 | #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE | |
881 | pmd_t *pmd = pvmw.pmd; | |
882 | pmd_t entry; | |
883 | ||
884 | if (!pmd_dirty(*pmd) && !pmd_write(*pmd)) | |
885 | continue; | |
886 | ||
887 | flush_cache_page(vma, address, page_to_pfn(page)); | |
888 | entry = pmdp_huge_clear_flush(vma, address, pmd); | |
889 | entry = pmd_wrprotect(entry); | |
890 | entry = pmd_mkclean(entry); | |
891 | set_pmd_at(vma->vm_mm, address, pmd, entry); | |
892 | ret = 1; | |
893 | #else | |
894 | /* unexpected pmd-mapped page? */ | |
895 | WARN_ON_ONCE(1); | |
896 | #endif | |
897 | } | |
898 | ||
899 | if (ret) { | |
900 | mmu_notifier_invalidate_page(vma->vm_mm, address); | |
901 | (*cleaned)++; | |
902 | } | |
903 | } | |
904 | ||
905 | return SWAP_AGAIN; | |
906 | } | |
907 | ||
908 | static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg) | |
909 | { | |
910 | if (vma->vm_flags & VM_SHARED) | |
911 | return false; | |
912 | ||
913 | return true; | |
914 | } | |
915 | ||
916 | int page_mkclean(struct page *page) | |
917 | { | |
918 | int cleaned = 0; | |
919 | struct address_space *mapping; | |
920 | struct rmap_walk_control rwc = { | |
921 | .arg = (void *)&cleaned, | |
922 | .rmap_one = page_mkclean_one, | |
923 | .invalid_vma = invalid_mkclean_vma, | |
924 | }; | |
925 | ||
926 | BUG_ON(!PageLocked(page)); | |
927 | ||
928 | if (!page_mapped(page)) | |
929 | return 0; | |
930 | ||
931 | mapping = page_mapping(page); | |
932 | if (!mapping) | |
933 | return 0; | |
934 | ||
935 | rmap_walk(page, &rwc); | |
936 | ||
937 | return cleaned; | |
938 | } | |
939 | EXPORT_SYMBOL_GPL(page_mkclean); | |
940 | ||
941 | /** | |
942 | * page_move_anon_rmap - move a page to our anon_vma | |
943 | * @page: the page to move to our anon_vma | |
944 | * @vma: the vma the page belongs to | |
945 | * | |
946 | * When a page belongs exclusively to one process after a COW event, | |
947 | * that page can be moved into the anon_vma that belongs to just that | |
948 | * process, so the rmap code will not search the parent or sibling | |
949 | * processes. | |
950 | */ | |
951 | void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma) | |
952 | { | |
953 | struct anon_vma *anon_vma = vma->anon_vma; | |
954 | ||
955 | page = compound_head(page); | |
956 | ||
957 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
958 | VM_BUG_ON_VMA(!anon_vma, vma); | |
959 | ||
960 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
961 | /* | |
962 | * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written | |
963 | * simultaneously, so a concurrent reader (eg page_referenced()'s | |
964 | * PageAnon()) will not see one without the other. | |
965 | */ | |
966 | WRITE_ONCE(page->mapping, (struct address_space *) anon_vma); | |
967 | } | |
968 | ||
969 | /** | |
970 | * __page_set_anon_rmap - set up new anonymous rmap | |
971 | * @page: Page to add to rmap | |
972 | * @vma: VM area to add page to. | |
973 | * @address: User virtual address of the mapping | |
974 | * @exclusive: the page is exclusively owned by the current process | |
975 | */ | |
976 | static void __page_set_anon_rmap(struct page *page, | |
977 | struct vm_area_struct *vma, unsigned long address, int exclusive) | |
978 | { | |
979 | struct anon_vma *anon_vma = vma->anon_vma; | |
980 | ||
981 | BUG_ON(!anon_vma); | |
982 | ||
983 | if (PageAnon(page)) | |
984 | return; | |
985 | ||
986 | /* | |
987 | * If the page isn't exclusively mapped into this vma, | |
988 | * we must use the _oldest_ possible anon_vma for the | |
989 | * page mapping! | |
990 | */ | |
991 | if (!exclusive) | |
992 | anon_vma = anon_vma->root; | |
993 | ||
994 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
995 | page->mapping = (struct address_space *) anon_vma; | |
996 | page->index = linear_page_index(vma, address); | |
997 | } | |
998 | ||
999 | /** | |
1000 | * __page_check_anon_rmap - sanity check anonymous rmap addition | |
1001 | * @page: the page to add the mapping to | |
1002 | * @vma: the vm area in which the mapping is added | |
1003 | * @address: the user virtual address mapped | |
1004 | */ | |
1005 | static void __page_check_anon_rmap(struct page *page, | |
1006 | struct vm_area_struct *vma, unsigned long address) | |
1007 | { | |
1008 | #ifdef CONFIG_DEBUG_VM | |
1009 | /* | |
1010 | * The page's anon-rmap details (mapping and index) are guaranteed to | |
1011 | * be set up correctly at this point. | |
1012 | * | |
1013 | * We have exclusion against page_add_anon_rmap because the caller | |
1014 | * always holds the page locked, except if called from page_dup_rmap, | |
1015 | * in which case the page is already known to be setup. | |
1016 | * | |
1017 | * We have exclusion against page_add_new_anon_rmap because those pages | |
1018 | * are initially only visible via the pagetables, and the pte is locked | |
1019 | * over the call to page_add_new_anon_rmap. | |
1020 | */ | |
1021 | BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root); | |
1022 | BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address)); | |
1023 | #endif | |
1024 | } | |
1025 | ||
1026 | /** | |
1027 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
1028 | * @page: the page to add the mapping to | |
1029 | * @vma: the vm area in which the mapping is added | |
1030 | * @address: the user virtual address mapped | |
1031 | * @compound: charge the page as compound or small page | |
1032 | * | |
1033 | * The caller needs to hold the pte lock, and the page must be locked in | |
1034 | * the anon_vma case: to serialize mapping,index checking after setting, | |
1035 | * and to ensure that PageAnon is not being upgraded racily to PageKsm | |
1036 | * (but PageKsm is never downgraded to PageAnon). | |
1037 | */ | |
1038 | void page_add_anon_rmap(struct page *page, | |
1039 | struct vm_area_struct *vma, unsigned long address, bool compound) | |
1040 | { | |
1041 | do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0); | |
1042 | } | |
1043 | ||
1044 | /* | |
1045 | * Special version of the above for do_swap_page, which often runs | |
1046 | * into pages that are exclusively owned by the current process. | |
1047 | * Everybody else should continue to use page_add_anon_rmap above. | |
1048 | */ | |
1049 | void do_page_add_anon_rmap(struct page *page, | |
1050 | struct vm_area_struct *vma, unsigned long address, int flags) | |
1051 | { | |
1052 | bool compound = flags & RMAP_COMPOUND; | |
1053 | bool first; | |
1054 | ||
1055 | if (compound) { | |
1056 | atomic_t *mapcount; | |
1057 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
1058 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
1059 | mapcount = compound_mapcount_ptr(page); | |
1060 | first = atomic_inc_and_test(mapcount); | |
1061 | } else { | |
1062 | first = atomic_inc_and_test(&page->_mapcount); | |
1063 | } | |
1064 | ||
1065 | if (first) { | |
1066 | int nr = compound ? hpage_nr_pages(page) : 1; | |
1067 | /* | |
1068 | * We use the irq-unsafe __{inc|mod}_zone_page_stat because | |
1069 | * these counters are not modified in interrupt context, and | |
1070 | * pte lock(a spinlock) is held, which implies preemption | |
1071 | * disabled. | |
1072 | */ | |
1073 | if (compound) | |
1074 | __inc_node_page_state(page, NR_ANON_THPS); | |
1075 | __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr); | |
1076 | } | |
1077 | if (unlikely(PageKsm(page))) | |
1078 | return; | |
1079 | ||
1080 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
1081 | ||
1082 | /* address might be in next vma when migration races vma_adjust */ | |
1083 | if (first) | |
1084 | __page_set_anon_rmap(page, vma, address, | |
1085 | flags & RMAP_EXCLUSIVE); | |
1086 | else | |
1087 | __page_check_anon_rmap(page, vma, address); | |
1088 | } | |
1089 | ||
1090 | /** | |
1091 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page | |
1092 | * @page: the page to add the mapping to | |
1093 | * @vma: the vm area in which the mapping is added | |
1094 | * @address: the user virtual address mapped | |
1095 | * @compound: charge the page as compound or small page | |
1096 | * | |
1097 | * Same as page_add_anon_rmap but must only be called on *new* pages. | |
1098 | * This means the inc-and-test can be bypassed. | |
1099 | * Page does not have to be locked. | |
1100 | */ | |
1101 | void page_add_new_anon_rmap(struct page *page, | |
1102 | struct vm_area_struct *vma, unsigned long address, bool compound) | |
1103 | { | |
1104 | int nr = compound ? hpage_nr_pages(page) : 1; | |
1105 | ||
1106 | VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); | |
1107 | __SetPageSwapBacked(page); | |
1108 | if (compound) { | |
1109 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
1110 | /* increment count (starts at -1) */ | |
1111 | atomic_set(compound_mapcount_ptr(page), 0); | |
1112 | __inc_node_page_state(page, NR_ANON_THPS); | |
1113 | } else { | |
1114 | /* Anon THP always mapped first with PMD */ | |
1115 | VM_BUG_ON_PAGE(PageTransCompound(page), page); | |
1116 | /* increment count (starts at -1) */ | |
1117 | atomic_set(&page->_mapcount, 0); | |
1118 | } | |
1119 | __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr); | |
1120 | __page_set_anon_rmap(page, vma, address, 1); | |
1121 | } | |
1122 | ||
1123 | /** | |
1124 | * page_add_file_rmap - add pte mapping to a file page | |
1125 | * @page: the page to add the mapping to | |
1126 | * | |
1127 | * The caller needs to hold the pte lock. | |
1128 | */ | |
1129 | void page_add_file_rmap(struct page *page, bool compound) | |
1130 | { | |
1131 | int i, nr = 1; | |
1132 | ||
1133 | VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page); | |
1134 | lock_page_memcg(page); | |
1135 | if (compound && PageTransHuge(page)) { | |
1136 | for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) { | |
1137 | if (atomic_inc_and_test(&page[i]._mapcount)) | |
1138 | nr++; | |
1139 | } | |
1140 | if (!atomic_inc_and_test(compound_mapcount_ptr(page))) | |
1141 | goto out; | |
1142 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
1143 | __inc_node_page_state(page, NR_SHMEM_PMDMAPPED); | |
1144 | } else { | |
1145 | if (PageTransCompound(page) && page_mapping(page)) { | |
1146 | VM_WARN_ON_ONCE(!PageLocked(page)); | |
1147 | ||
1148 | SetPageDoubleMap(compound_head(page)); | |
1149 | if (PageMlocked(page)) | |
1150 | clear_page_mlock(compound_head(page)); | |
1151 | } | |
1152 | if (!atomic_inc_and_test(&page->_mapcount)) | |
1153 | goto out; | |
1154 | } | |
1155 | __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, nr); | |
1156 | mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED); | |
1157 | out: | |
1158 | unlock_page_memcg(page); | |
1159 | } | |
1160 | ||
1161 | static void page_remove_file_rmap(struct page *page, bool compound) | |
1162 | { | |
1163 | int i, nr = 1; | |
1164 | ||
1165 | VM_BUG_ON_PAGE(compound && !PageHead(page), page); | |
1166 | lock_page_memcg(page); | |
1167 | ||
1168 | /* Hugepages are not counted in NR_FILE_MAPPED for now. */ | |
1169 | if (unlikely(PageHuge(page))) { | |
1170 | /* hugetlb pages are always mapped with pmds */ | |
1171 | atomic_dec(compound_mapcount_ptr(page)); | |
1172 | goto out; | |
1173 | } | |
1174 | ||
1175 | /* page still mapped by someone else? */ | |
1176 | if (compound && PageTransHuge(page)) { | |
1177 | for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) { | |
1178 | if (atomic_add_negative(-1, &page[i]._mapcount)) | |
1179 | nr++; | |
1180 | } | |
1181 | if (!atomic_add_negative(-1, compound_mapcount_ptr(page))) | |
1182 | goto out; | |
1183 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
1184 | __dec_node_page_state(page, NR_SHMEM_PMDMAPPED); | |
1185 | } else { | |
1186 | if (!atomic_add_negative(-1, &page->_mapcount)) | |
1187 | goto out; | |
1188 | } | |
1189 | ||
1190 | /* | |
1191 | * We use the irq-unsafe __{inc|mod}_zone_page_state because | |
1192 | * these counters are not modified in interrupt context, and | |
1193 | * pte lock(a spinlock) is held, which implies preemption disabled. | |
1194 | */ | |
1195 | __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, -nr); | |
1196 | mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED); | |
1197 | ||
1198 | if (unlikely(PageMlocked(page))) | |
1199 | clear_page_mlock(page); | |
1200 | out: | |
1201 | unlock_page_memcg(page); | |
1202 | } | |
1203 | ||
1204 | static void page_remove_anon_compound_rmap(struct page *page) | |
1205 | { | |
1206 | int i, nr; | |
1207 | ||
1208 | if (!atomic_add_negative(-1, compound_mapcount_ptr(page))) | |
1209 | return; | |
1210 | ||
1211 | /* Hugepages are not counted in NR_ANON_PAGES for now. */ | |
1212 | if (unlikely(PageHuge(page))) | |
1213 | return; | |
1214 | ||
1215 | if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) | |
1216 | return; | |
1217 | ||
1218 | __dec_node_page_state(page, NR_ANON_THPS); | |
1219 | ||
1220 | if (TestClearPageDoubleMap(page)) { | |
1221 | /* | |
1222 | * Subpages can be mapped with PTEs too. Check how many of | |
1223 | * themi are still mapped. | |
1224 | */ | |
1225 | for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) { | |
1226 | if (atomic_add_negative(-1, &page[i]._mapcount)) | |
1227 | nr++; | |
1228 | } | |
1229 | } else { | |
1230 | nr = HPAGE_PMD_NR; | |
1231 | } | |
1232 | ||
1233 | if (unlikely(PageMlocked(page))) | |
1234 | clear_page_mlock(page); | |
1235 | ||
1236 | if (nr) { | |
1237 | __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr); | |
1238 | deferred_split_huge_page(page); | |
1239 | } | |
1240 | } | |
1241 | ||
1242 | /** | |
1243 | * page_remove_rmap - take down pte mapping from a page | |
1244 | * @page: page to remove mapping from | |
1245 | * @compound: uncharge the page as compound or small page | |
1246 | * | |
1247 | * The caller needs to hold the pte lock. | |
1248 | */ | |
1249 | void page_remove_rmap(struct page *page, bool compound) | |
1250 | { | |
1251 | if (!PageAnon(page)) | |
1252 | return page_remove_file_rmap(page, compound); | |
1253 | ||
1254 | if (compound) | |
1255 | return page_remove_anon_compound_rmap(page); | |
1256 | ||
1257 | /* page still mapped by someone else? */ | |
1258 | if (!atomic_add_negative(-1, &page->_mapcount)) | |
1259 | return; | |
1260 | ||
1261 | /* | |
1262 | * We use the irq-unsafe __{inc|mod}_zone_page_stat because | |
1263 | * these counters are not modified in interrupt context, and | |
1264 | * pte lock(a spinlock) is held, which implies preemption disabled. | |
1265 | */ | |
1266 | __dec_node_page_state(page, NR_ANON_MAPPED); | |
1267 | ||
1268 | if (unlikely(PageMlocked(page))) | |
1269 | clear_page_mlock(page); | |
1270 | ||
1271 | if (PageTransCompound(page)) | |
1272 | deferred_split_huge_page(compound_head(page)); | |
1273 | ||
1274 | /* | |
1275 | * It would be tidy to reset the PageAnon mapping here, | |
1276 | * but that might overwrite a racing page_add_anon_rmap | |
1277 | * which increments mapcount after us but sets mapping | |
1278 | * before us: so leave the reset to free_hot_cold_page, | |
1279 | * and remember that it's only reliable while mapped. | |
1280 | * Leaving it set also helps swapoff to reinstate ptes | |
1281 | * faster for those pages still in swapcache. | |
1282 | */ | |
1283 | } | |
1284 | ||
1285 | struct rmap_private { | |
1286 | enum ttu_flags flags; | |
1287 | int lazyfreed; | |
1288 | }; | |
1289 | ||
1290 | /* | |
1291 | * @arg: enum ttu_flags will be passed to this argument | |
1292 | */ | |
1293 | static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, | |
1294 | unsigned long address, void *arg) | |
1295 | { | |
1296 | struct mm_struct *mm = vma->vm_mm; | |
1297 | struct page_vma_mapped_walk pvmw = { | |
1298 | .page = page, | |
1299 | .vma = vma, | |
1300 | .address = address, | |
1301 | }; | |
1302 | pte_t pteval; | |
1303 | struct page *subpage; | |
1304 | int ret = SWAP_AGAIN; | |
1305 | struct rmap_private *rp = arg; | |
1306 | enum ttu_flags flags = rp->flags; | |
1307 | ||
1308 | /* munlock has nothing to gain from examining un-locked vmas */ | |
1309 | if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED)) | |
1310 | return SWAP_AGAIN; | |
1311 | ||
1312 | if (flags & TTU_SPLIT_HUGE_PMD) { | |
1313 | split_huge_pmd_address(vma, address, | |
1314 | flags & TTU_MIGRATION, page); | |
1315 | } | |
1316 | ||
1317 | while (page_vma_mapped_walk(&pvmw)) { | |
1318 | subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte); | |
1319 | address = pvmw.address; | |
1320 | ||
1321 | /* Unexpected PMD-mapped THP? */ | |
1322 | VM_BUG_ON_PAGE(!pvmw.pte, page); | |
1323 | ||
1324 | /* | |
1325 | * If the page is mlock()d, we cannot swap it out. | |
1326 | * If it's recently referenced (perhaps page_referenced | |
1327 | * skipped over this mm) then we should reactivate it. | |
1328 | */ | |
1329 | if (!(flags & TTU_IGNORE_MLOCK)) { | |
1330 | if (vma->vm_flags & VM_LOCKED) { | |
1331 | /* PTE-mapped THP are never mlocked */ | |
1332 | if (!PageTransCompound(page)) { | |
1333 | /* | |
1334 | * Holding pte lock, we do *not* need | |
1335 | * mmap_sem here | |
1336 | */ | |
1337 | mlock_vma_page(page); | |
1338 | } | |
1339 | ret = SWAP_MLOCK; | |
1340 | page_vma_mapped_walk_done(&pvmw); | |
1341 | break; | |
1342 | } | |
1343 | if (flags & TTU_MUNLOCK) | |
1344 | continue; | |
1345 | } | |
1346 | ||
1347 | if (!(flags & TTU_IGNORE_ACCESS)) { | |
1348 | if (ptep_clear_flush_young_notify(vma, address, | |
1349 | pvmw.pte)) { | |
1350 | ret = SWAP_FAIL; | |
1351 | page_vma_mapped_walk_done(&pvmw); | |
1352 | break; | |
1353 | } | |
1354 | } | |
1355 | ||
1356 | /* Nuke the page table entry. */ | |
1357 | flush_cache_page(vma, address, pte_pfn(*pvmw.pte)); | |
1358 | if (should_defer_flush(mm, flags)) { | |
1359 | /* | |
1360 | * We clear the PTE but do not flush so potentially | |
1361 | * a remote CPU could still be writing to the page. | |
1362 | * If the entry was previously clean then the | |
1363 | * architecture must guarantee that a clear->dirty | |
1364 | * transition on a cached TLB entry is written through | |
1365 | * and traps if the PTE is unmapped. | |
1366 | */ | |
1367 | pteval = ptep_get_and_clear(mm, address, pvmw.pte); | |
1368 | ||
1369 | set_tlb_ubc_flush_pending(mm, pte_dirty(pteval)); | |
1370 | } else { | |
1371 | pteval = ptep_clear_flush(vma, address, pvmw.pte); | |
1372 | } | |
1373 | ||
1374 | /* Move the dirty bit to the page. Now the pte is gone. */ | |
1375 | if (pte_dirty(pteval)) | |
1376 | set_page_dirty(page); | |
1377 | ||
1378 | /* Update high watermark before we lower rss */ | |
1379 | update_hiwater_rss(mm); | |
1380 | ||
1381 | if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) { | |
1382 | if (PageHuge(page)) { | |
1383 | int nr = 1 << compound_order(page); | |
1384 | hugetlb_count_sub(nr, mm); | |
1385 | } else { | |
1386 | dec_mm_counter(mm, mm_counter(page)); | |
1387 | } | |
1388 | ||
1389 | pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); | |
1390 | set_pte_at(mm, address, pvmw.pte, pteval); | |
1391 | } else if (pte_unused(pteval)) { | |
1392 | /* | |
1393 | * The guest indicated that the page content is of no | |
1394 | * interest anymore. Simply discard the pte, vmscan | |
1395 | * will take care of the rest. | |
1396 | */ | |
1397 | dec_mm_counter(mm, mm_counter(page)); | |
1398 | } else if (IS_ENABLED(CONFIG_MIGRATION) && | |
1399 | (flags & TTU_MIGRATION)) { | |
1400 | swp_entry_t entry; | |
1401 | pte_t swp_pte; | |
1402 | /* | |
1403 | * Store the pfn of the page in a special migration | |
1404 | * pte. do_swap_page() will wait until the migration | |
1405 | * pte is removed and then restart fault handling. | |
1406 | */ | |
1407 | entry = make_migration_entry(subpage, | |
1408 | pte_write(pteval)); | |
1409 | swp_pte = swp_entry_to_pte(entry); | |
1410 | if (pte_soft_dirty(pteval)) | |
1411 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
1412 | set_pte_at(mm, address, pvmw.pte, swp_pte); | |
1413 | } else if (PageAnon(page)) { | |
1414 | swp_entry_t entry = { .val = page_private(subpage) }; | |
1415 | pte_t swp_pte; | |
1416 | /* | |
1417 | * Store the swap location in the pte. | |
1418 | * See handle_pte_fault() ... | |
1419 | */ | |
1420 | VM_BUG_ON_PAGE(!PageSwapCache(page), page); | |
1421 | ||
1422 | if (!PageDirty(page) && (flags & TTU_LZFREE)) { | |
1423 | /* It's a freeable page by MADV_FREE */ | |
1424 | dec_mm_counter(mm, MM_ANONPAGES); | |
1425 | rp->lazyfreed++; | |
1426 | goto discard; | |
1427 | } | |
1428 | ||
1429 | if (swap_duplicate(entry) < 0) { | |
1430 | set_pte_at(mm, address, pvmw.pte, pteval); | |
1431 | ret = SWAP_FAIL; | |
1432 | page_vma_mapped_walk_done(&pvmw); | |
1433 | break; | |
1434 | } | |
1435 | if (list_empty(&mm->mmlist)) { | |
1436 | spin_lock(&mmlist_lock); | |
1437 | if (list_empty(&mm->mmlist)) | |
1438 | list_add(&mm->mmlist, &init_mm.mmlist); | |
1439 | spin_unlock(&mmlist_lock); | |
1440 | } | |
1441 | dec_mm_counter(mm, MM_ANONPAGES); | |
1442 | inc_mm_counter(mm, MM_SWAPENTS); | |
1443 | swp_pte = swp_entry_to_pte(entry); | |
1444 | if (pte_soft_dirty(pteval)) | |
1445 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
1446 | set_pte_at(mm, address, pvmw.pte, swp_pte); | |
1447 | } else | |
1448 | dec_mm_counter(mm, mm_counter_file(page)); | |
1449 | discard: | |
1450 | page_remove_rmap(subpage, PageHuge(page)); | |
1451 | put_page(page); | |
1452 | mmu_notifier_invalidate_page(mm, address); | |
1453 | } | |
1454 | return ret; | |
1455 | } | |
1456 | ||
1457 | bool is_vma_temporary_stack(struct vm_area_struct *vma) | |
1458 | { | |
1459 | int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP); | |
1460 | ||
1461 | if (!maybe_stack) | |
1462 | return false; | |
1463 | ||
1464 | if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) == | |
1465 | VM_STACK_INCOMPLETE_SETUP) | |
1466 | return true; | |
1467 | ||
1468 | return false; | |
1469 | } | |
1470 | ||
1471 | static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg) | |
1472 | { | |
1473 | return is_vma_temporary_stack(vma); | |
1474 | } | |
1475 | ||
1476 | static int page_mapcount_is_zero(struct page *page) | |
1477 | { | |
1478 | return !total_mapcount(page); | |
1479 | } | |
1480 | ||
1481 | /** | |
1482 | * try_to_unmap - try to remove all page table mappings to a page | |
1483 | * @page: the page to get unmapped | |
1484 | * @flags: action and flags | |
1485 | * | |
1486 | * Tries to remove all the page table entries which are mapping this | |
1487 | * page, used in the pageout path. Caller must hold the page lock. | |
1488 | * Return values are: | |
1489 | * | |
1490 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
1491 | * SWAP_AGAIN - we missed a mapping, try again later | |
1492 | * SWAP_FAIL - the page is unswappable | |
1493 | * SWAP_MLOCK - page is mlocked. | |
1494 | */ | |
1495 | int try_to_unmap(struct page *page, enum ttu_flags flags) | |
1496 | { | |
1497 | int ret; | |
1498 | struct rmap_private rp = { | |
1499 | .flags = flags, | |
1500 | .lazyfreed = 0, | |
1501 | }; | |
1502 | ||
1503 | struct rmap_walk_control rwc = { | |
1504 | .rmap_one = try_to_unmap_one, | |
1505 | .arg = &rp, | |
1506 | .done = page_mapcount_is_zero, | |
1507 | .anon_lock = page_lock_anon_vma_read, | |
1508 | }; | |
1509 | ||
1510 | /* | |
1511 | * During exec, a temporary VMA is setup and later moved. | |
1512 | * The VMA is moved under the anon_vma lock but not the | |
1513 | * page tables leading to a race where migration cannot | |
1514 | * find the migration ptes. Rather than increasing the | |
1515 | * locking requirements of exec(), migration skips | |
1516 | * temporary VMAs until after exec() completes. | |
1517 | */ | |
1518 | if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page)) | |
1519 | rwc.invalid_vma = invalid_migration_vma; | |
1520 | ||
1521 | if (flags & TTU_RMAP_LOCKED) | |
1522 | ret = rmap_walk_locked(page, &rwc); | |
1523 | else | |
1524 | ret = rmap_walk(page, &rwc); | |
1525 | ||
1526 | if (ret != SWAP_MLOCK && !page_mapcount(page)) { | |
1527 | ret = SWAP_SUCCESS; | |
1528 | if (rp.lazyfreed && !PageDirty(page)) | |
1529 | ret = SWAP_LZFREE; | |
1530 | } | |
1531 | return ret; | |
1532 | } | |
1533 | ||
1534 | static int page_not_mapped(struct page *page) | |
1535 | { | |
1536 | return !page_mapped(page); | |
1537 | }; | |
1538 | ||
1539 | /** | |
1540 | * try_to_munlock - try to munlock a page | |
1541 | * @page: the page to be munlocked | |
1542 | * | |
1543 | * Called from munlock code. Checks all of the VMAs mapping the page | |
1544 | * to make sure nobody else has this page mlocked. The page will be | |
1545 | * returned with PG_mlocked cleared if no other vmas have it mlocked. | |
1546 | * | |
1547 | * Return values are: | |
1548 | * | |
1549 | * SWAP_AGAIN - no vma is holding page mlocked, or, | |
1550 | * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem | |
1551 | * SWAP_FAIL - page cannot be located at present | |
1552 | * SWAP_MLOCK - page is now mlocked. | |
1553 | */ | |
1554 | int try_to_munlock(struct page *page) | |
1555 | { | |
1556 | int ret; | |
1557 | struct rmap_private rp = { | |
1558 | .flags = TTU_MUNLOCK, | |
1559 | .lazyfreed = 0, | |
1560 | }; | |
1561 | ||
1562 | struct rmap_walk_control rwc = { | |
1563 | .rmap_one = try_to_unmap_one, | |
1564 | .arg = &rp, | |
1565 | .done = page_not_mapped, | |
1566 | .anon_lock = page_lock_anon_vma_read, | |
1567 | ||
1568 | }; | |
1569 | ||
1570 | VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page); | |
1571 | ||
1572 | ret = rmap_walk(page, &rwc); | |
1573 | return ret; | |
1574 | } | |
1575 | ||
1576 | void __put_anon_vma(struct anon_vma *anon_vma) | |
1577 | { | |
1578 | struct anon_vma *root = anon_vma->root; | |
1579 | ||
1580 | anon_vma_free(anon_vma); | |
1581 | if (root != anon_vma && atomic_dec_and_test(&root->refcount)) | |
1582 | anon_vma_free(root); | |
1583 | } | |
1584 | ||
1585 | static struct anon_vma *rmap_walk_anon_lock(struct page *page, | |
1586 | struct rmap_walk_control *rwc) | |
1587 | { | |
1588 | struct anon_vma *anon_vma; | |
1589 | ||
1590 | if (rwc->anon_lock) | |
1591 | return rwc->anon_lock(page); | |
1592 | ||
1593 | /* | |
1594 | * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read() | |
1595 | * because that depends on page_mapped(); but not all its usages | |
1596 | * are holding mmap_sem. Users without mmap_sem are required to | |
1597 | * take a reference count to prevent the anon_vma disappearing | |
1598 | */ | |
1599 | anon_vma = page_anon_vma(page); | |
1600 | if (!anon_vma) | |
1601 | return NULL; | |
1602 | ||
1603 | anon_vma_lock_read(anon_vma); | |
1604 | return anon_vma; | |
1605 | } | |
1606 | ||
1607 | /* | |
1608 | * rmap_walk_anon - do something to anonymous page using the object-based | |
1609 | * rmap method | |
1610 | * @page: the page to be handled | |
1611 | * @rwc: control variable according to each walk type | |
1612 | * | |
1613 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
1614 | * contained in the anon_vma struct it points to. | |
1615 | * | |
1616 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma | |
1617 | * where the page was found will be held for write. So, we won't recheck | |
1618 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be | |
1619 | * LOCKED. | |
1620 | */ | |
1621 | static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc, | |
1622 | bool locked) | |
1623 | { | |
1624 | struct anon_vma *anon_vma; | |
1625 | pgoff_t pgoff_start, pgoff_end; | |
1626 | struct anon_vma_chain *avc; | |
1627 | int ret = SWAP_AGAIN; | |
1628 | ||
1629 | if (locked) { | |
1630 | anon_vma = page_anon_vma(page); | |
1631 | /* anon_vma disappear under us? */ | |
1632 | VM_BUG_ON_PAGE(!anon_vma, page); | |
1633 | } else { | |
1634 | anon_vma = rmap_walk_anon_lock(page, rwc); | |
1635 | } | |
1636 | if (!anon_vma) | |
1637 | return ret; | |
1638 | ||
1639 | pgoff_start = page_to_pgoff(page); | |
1640 | pgoff_end = pgoff_start + hpage_nr_pages(page) - 1; | |
1641 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, | |
1642 | pgoff_start, pgoff_end) { | |
1643 | struct vm_area_struct *vma = avc->vma; | |
1644 | unsigned long address = vma_address(page, vma); | |
1645 | ||
1646 | cond_resched(); | |
1647 | ||
1648 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) | |
1649 | continue; | |
1650 | ||
1651 | ret = rwc->rmap_one(page, vma, address, rwc->arg); | |
1652 | if (ret != SWAP_AGAIN) | |
1653 | break; | |
1654 | if (rwc->done && rwc->done(page)) | |
1655 | break; | |
1656 | } | |
1657 | ||
1658 | if (!locked) | |
1659 | anon_vma_unlock_read(anon_vma); | |
1660 | return ret; | |
1661 | } | |
1662 | ||
1663 | /* | |
1664 | * rmap_walk_file - do something to file page using the object-based rmap method | |
1665 | * @page: the page to be handled | |
1666 | * @rwc: control variable according to each walk type | |
1667 | * | |
1668 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
1669 | * contained in the address_space struct it points to. | |
1670 | * | |
1671 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma | |
1672 | * where the page was found will be held for write. So, we won't recheck | |
1673 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be | |
1674 | * LOCKED. | |
1675 | */ | |
1676 | static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc, | |
1677 | bool locked) | |
1678 | { | |
1679 | struct address_space *mapping = page_mapping(page); | |
1680 | pgoff_t pgoff_start, pgoff_end; | |
1681 | struct vm_area_struct *vma; | |
1682 | int ret = SWAP_AGAIN; | |
1683 | ||
1684 | /* | |
1685 | * The page lock not only makes sure that page->mapping cannot | |
1686 | * suddenly be NULLified by truncation, it makes sure that the | |
1687 | * structure at mapping cannot be freed and reused yet, | |
1688 | * so we can safely take mapping->i_mmap_rwsem. | |
1689 | */ | |
1690 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
1691 | ||
1692 | if (!mapping) | |
1693 | return ret; | |
1694 | ||
1695 | pgoff_start = page_to_pgoff(page); | |
1696 | pgoff_end = pgoff_start + hpage_nr_pages(page) - 1; | |
1697 | if (!locked) | |
1698 | i_mmap_lock_read(mapping); | |
1699 | vma_interval_tree_foreach(vma, &mapping->i_mmap, | |
1700 | pgoff_start, pgoff_end) { | |
1701 | unsigned long address = vma_address(page, vma); | |
1702 | ||
1703 | cond_resched(); | |
1704 | ||
1705 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) | |
1706 | continue; | |
1707 | ||
1708 | ret = rwc->rmap_one(page, vma, address, rwc->arg); | |
1709 | if (ret != SWAP_AGAIN) | |
1710 | goto done; | |
1711 | if (rwc->done && rwc->done(page)) | |
1712 | goto done; | |
1713 | } | |
1714 | ||
1715 | done: | |
1716 | if (!locked) | |
1717 | i_mmap_unlock_read(mapping); | |
1718 | return ret; | |
1719 | } | |
1720 | ||
1721 | int rmap_walk(struct page *page, struct rmap_walk_control *rwc) | |
1722 | { | |
1723 | if (unlikely(PageKsm(page))) | |
1724 | return rmap_walk_ksm(page, rwc); | |
1725 | else if (PageAnon(page)) | |
1726 | return rmap_walk_anon(page, rwc, false); | |
1727 | else | |
1728 | return rmap_walk_file(page, rwc, false); | |
1729 | } | |
1730 | ||
1731 | /* Like rmap_walk, but caller holds relevant rmap lock */ | |
1732 | int rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc) | |
1733 | { | |
1734 | /* no ksm support for now */ | |
1735 | VM_BUG_ON_PAGE(PageKsm(page), page); | |
1736 | if (PageAnon(page)) | |
1737 | return rmap_walk_anon(page, rwc, true); | |
1738 | else | |
1739 | return rmap_walk_file(page, rwc, true); | |
1740 | } | |
1741 | ||
1742 | #ifdef CONFIG_HUGETLB_PAGE | |
1743 | /* | |
1744 | * The following three functions are for anonymous (private mapped) hugepages. | |
1745 | * Unlike common anonymous pages, anonymous hugepages have no accounting code | |
1746 | * and no lru code, because we handle hugepages differently from common pages. | |
1747 | */ | |
1748 | static void __hugepage_set_anon_rmap(struct page *page, | |
1749 | struct vm_area_struct *vma, unsigned long address, int exclusive) | |
1750 | { | |
1751 | struct anon_vma *anon_vma = vma->anon_vma; | |
1752 | ||
1753 | BUG_ON(!anon_vma); | |
1754 | ||
1755 | if (PageAnon(page)) | |
1756 | return; | |
1757 | if (!exclusive) | |
1758 | anon_vma = anon_vma->root; | |
1759 | ||
1760 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
1761 | page->mapping = (struct address_space *) anon_vma; | |
1762 | page->index = linear_page_index(vma, address); | |
1763 | } | |
1764 | ||
1765 | void hugepage_add_anon_rmap(struct page *page, | |
1766 | struct vm_area_struct *vma, unsigned long address) | |
1767 | { | |
1768 | struct anon_vma *anon_vma = vma->anon_vma; | |
1769 | int first; | |
1770 | ||
1771 | BUG_ON(!PageLocked(page)); | |
1772 | BUG_ON(!anon_vma); | |
1773 | /* address might be in next vma when migration races vma_adjust */ | |
1774 | first = atomic_inc_and_test(compound_mapcount_ptr(page)); | |
1775 | if (first) | |
1776 | __hugepage_set_anon_rmap(page, vma, address, 0); | |
1777 | } | |
1778 | ||
1779 | void hugepage_add_new_anon_rmap(struct page *page, | |
1780 | struct vm_area_struct *vma, unsigned long address) | |
1781 | { | |
1782 | BUG_ON(address < vma->vm_start || address >= vma->vm_end); | |
1783 | atomic_set(compound_mapcount_ptr(page), 0); | |
1784 | __hugepage_set_anon_rmap(page, vma, address, 1); | |
1785 | } | |
1786 | #endif /* CONFIG_HUGETLB_PAGE */ |