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