]>
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 | ||
60 | #include <asm/tlbflush.h> | |
61 | ||
62 | #include "internal.h" | |
63 | ||
64 | static struct kmem_cache *anon_vma_cachep; | |
65 | static struct kmem_cache *anon_vma_chain_cachep; | |
66 | ||
67 | static inline struct anon_vma *anon_vma_alloc(void) | |
68 | { | |
69 | return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); | |
70 | } | |
71 | ||
72 | void anon_vma_free(struct anon_vma *anon_vma) | |
73 | { | |
74 | kmem_cache_free(anon_vma_cachep, anon_vma); | |
75 | } | |
76 | ||
77 | static inline struct anon_vma_chain *anon_vma_chain_alloc(void) | |
78 | { | |
79 | return kmem_cache_alloc(anon_vma_chain_cachep, GFP_KERNEL); | |
80 | } | |
81 | ||
82 | void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) | |
83 | { | |
84 | kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); | |
85 | } | |
86 | ||
87 | /** | |
88 | * anon_vma_prepare - attach an anon_vma to a memory region | |
89 | * @vma: the memory region in question | |
90 | * | |
91 | * This makes sure the memory mapping described by 'vma' has | |
92 | * an 'anon_vma' attached to it, so that we can associate the | |
93 | * anonymous pages mapped into it with that anon_vma. | |
94 | * | |
95 | * The common case will be that we already have one, but if | |
96 | * if not we either need to find an adjacent mapping that we | |
97 | * can re-use the anon_vma from (very common when the only | |
98 | * reason for splitting a vma has been mprotect()), or we | |
99 | * allocate a new one. | |
100 | * | |
101 | * Anon-vma allocations are very subtle, because we may have | |
102 | * optimistically looked up an anon_vma in page_lock_anon_vma() | |
103 | * and that may actually touch the spinlock even in the newly | |
104 | * allocated vma (it depends on RCU to make sure that the | |
105 | * anon_vma isn't actually destroyed). | |
106 | * | |
107 | * As a result, we need to do proper anon_vma locking even | |
108 | * for the new allocation. At the same time, we do not want | |
109 | * to do any locking for the common case of already having | |
110 | * an anon_vma. | |
111 | * | |
112 | * This must be called with the mmap_sem held for reading. | |
113 | */ | |
114 | int anon_vma_prepare(struct vm_area_struct *vma) | |
115 | { | |
116 | struct anon_vma *anon_vma = vma->anon_vma; | |
117 | struct anon_vma_chain *avc; | |
118 | ||
119 | might_sleep(); | |
120 | if (unlikely(!anon_vma)) { | |
121 | struct mm_struct *mm = vma->vm_mm; | |
122 | struct anon_vma *allocated; | |
123 | ||
124 | avc = anon_vma_chain_alloc(); | |
125 | if (!avc) | |
126 | goto out_enomem; | |
127 | ||
128 | anon_vma = find_mergeable_anon_vma(vma); | |
129 | allocated = NULL; | |
130 | if (!anon_vma) { | |
131 | anon_vma = anon_vma_alloc(); | |
132 | if (unlikely(!anon_vma)) | |
133 | goto out_enomem_free_avc; | |
134 | allocated = anon_vma; | |
135 | } | |
136 | ||
137 | spin_lock(&anon_vma->lock); | |
138 | /* page_table_lock to protect against threads */ | |
139 | spin_lock(&mm->page_table_lock); | |
140 | if (likely(!vma->anon_vma)) { | |
141 | vma->anon_vma = anon_vma; | |
142 | avc->anon_vma = anon_vma; | |
143 | avc->vma = vma; | |
144 | list_add(&avc->same_vma, &vma->anon_vma_chain); | |
145 | list_add(&avc->same_anon_vma, &anon_vma->head); | |
146 | allocated = NULL; | |
147 | avc = NULL; | |
148 | } | |
149 | spin_unlock(&mm->page_table_lock); | |
150 | spin_unlock(&anon_vma->lock); | |
151 | ||
152 | if (unlikely(allocated)) | |
153 | anon_vma_free(allocated); | |
154 | if (unlikely(avc)) | |
155 | anon_vma_chain_free(avc); | |
156 | } | |
157 | return 0; | |
158 | ||
159 | out_enomem_free_avc: | |
160 | anon_vma_chain_free(avc); | |
161 | out_enomem: | |
162 | return -ENOMEM; | |
163 | } | |
164 | ||
165 | static void anon_vma_chain_link(struct vm_area_struct *vma, | |
166 | struct anon_vma_chain *avc, | |
167 | struct anon_vma *anon_vma) | |
168 | { | |
169 | avc->vma = vma; | |
170 | avc->anon_vma = anon_vma; | |
171 | list_add(&avc->same_vma, &vma->anon_vma_chain); | |
172 | ||
173 | spin_lock(&anon_vma->lock); | |
174 | list_add_tail(&avc->same_anon_vma, &anon_vma->head); | |
175 | spin_unlock(&anon_vma->lock); | |
176 | } | |
177 | ||
178 | /* | |
179 | * Attach the anon_vmas from src to dst. | |
180 | * Returns 0 on success, -ENOMEM on failure. | |
181 | */ | |
182 | int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) | |
183 | { | |
184 | struct anon_vma_chain *avc, *pavc; | |
185 | ||
186 | list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { | |
187 | avc = anon_vma_chain_alloc(); | |
188 | if (!avc) | |
189 | goto enomem_failure; | |
190 | anon_vma_chain_link(dst, avc, pavc->anon_vma); | |
191 | } | |
192 | return 0; | |
193 | ||
194 | enomem_failure: | |
195 | unlink_anon_vmas(dst); | |
196 | return -ENOMEM; | |
197 | } | |
198 | ||
199 | /* | |
200 | * Attach vma to its own anon_vma, as well as to the anon_vmas that | |
201 | * the corresponding VMA in the parent process is attached to. | |
202 | * Returns 0 on success, non-zero on failure. | |
203 | */ | |
204 | int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) | |
205 | { | |
206 | struct anon_vma_chain *avc; | |
207 | struct anon_vma *anon_vma; | |
208 | ||
209 | /* Don't bother if the parent process has no anon_vma here. */ | |
210 | if (!pvma->anon_vma) | |
211 | return 0; | |
212 | ||
213 | /* | |
214 | * First, attach the new VMA to the parent VMA's anon_vmas, | |
215 | * so rmap can find non-COWed pages in child processes. | |
216 | */ | |
217 | if (anon_vma_clone(vma, pvma)) | |
218 | return -ENOMEM; | |
219 | ||
220 | /* Then add our own anon_vma. */ | |
221 | anon_vma = anon_vma_alloc(); | |
222 | if (!anon_vma) | |
223 | goto out_error; | |
224 | avc = anon_vma_chain_alloc(); | |
225 | if (!avc) | |
226 | goto out_error_free_anon_vma; | |
227 | anon_vma_chain_link(vma, avc, anon_vma); | |
228 | /* Mark this anon_vma as the one where our new (COWed) pages go. */ | |
229 | vma->anon_vma = anon_vma; | |
230 | ||
231 | return 0; | |
232 | ||
233 | out_error_free_anon_vma: | |
234 | anon_vma_free(anon_vma); | |
235 | out_error: | |
236 | unlink_anon_vmas(vma); | |
237 | return -ENOMEM; | |
238 | } | |
239 | ||
240 | static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain) | |
241 | { | |
242 | struct anon_vma *anon_vma = anon_vma_chain->anon_vma; | |
243 | int empty; | |
244 | ||
245 | /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */ | |
246 | if (!anon_vma) | |
247 | return; | |
248 | ||
249 | spin_lock(&anon_vma->lock); | |
250 | list_del(&anon_vma_chain->same_anon_vma); | |
251 | ||
252 | /* We must garbage collect the anon_vma if it's empty */ | |
253 | empty = list_empty(&anon_vma->head) && !ksm_refcount(anon_vma) && | |
254 | !migrate_refcount(anon_vma); | |
255 | spin_unlock(&anon_vma->lock); | |
256 | ||
257 | if (empty) | |
258 | anon_vma_free(anon_vma); | |
259 | } | |
260 | ||
261 | void unlink_anon_vmas(struct vm_area_struct *vma) | |
262 | { | |
263 | struct anon_vma_chain *avc, *next; | |
264 | ||
265 | /* Unlink each anon_vma chained to the VMA. */ | |
266 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { | |
267 | anon_vma_unlink(avc); | |
268 | list_del(&avc->same_vma); | |
269 | anon_vma_chain_free(avc); | |
270 | } | |
271 | } | |
272 | ||
273 | static void anon_vma_ctor(void *data) | |
274 | { | |
275 | struct anon_vma *anon_vma = data; | |
276 | ||
277 | spin_lock_init(&anon_vma->lock); | |
278 | ksm_refcount_init(anon_vma); | |
279 | migrate_refcount_init(anon_vma); | |
280 | INIT_LIST_HEAD(&anon_vma->head); | |
281 | } | |
282 | ||
283 | void __init anon_vma_init(void) | |
284 | { | |
285 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
286 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor); | |
287 | anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC); | |
288 | } | |
289 | ||
290 | /* | |
291 | * Getting a lock on a stable anon_vma from a page off the LRU is | |
292 | * tricky: page_lock_anon_vma rely on RCU to guard against the races. | |
293 | */ | |
294 | struct anon_vma *page_lock_anon_vma(struct page *page) | |
295 | { | |
296 | struct anon_vma *anon_vma; | |
297 | unsigned long anon_mapping; | |
298 | ||
299 | rcu_read_lock(); | |
300 | anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping); | |
301 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) | |
302 | goto out; | |
303 | if (!page_mapped(page)) | |
304 | goto out; | |
305 | ||
306 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
307 | spin_lock(&anon_vma->lock); | |
308 | return anon_vma; | |
309 | out: | |
310 | rcu_read_unlock(); | |
311 | return NULL; | |
312 | } | |
313 | ||
314 | void page_unlock_anon_vma(struct anon_vma *anon_vma) | |
315 | { | |
316 | spin_unlock(&anon_vma->lock); | |
317 | rcu_read_unlock(); | |
318 | } | |
319 | ||
320 | /* | |
321 | * At what user virtual address is page expected in @vma? | |
322 | * Returns virtual address or -EFAULT if page's index/offset is not | |
323 | * within the range mapped the @vma. | |
324 | */ | |
325 | static inline unsigned long | |
326 | vma_address(struct page *page, struct vm_area_struct *vma) | |
327 | { | |
328 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
329 | unsigned long address; | |
330 | ||
331 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
332 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { | |
333 | /* page should be within @vma mapping range */ | |
334 | return -EFAULT; | |
335 | } | |
336 | return address; | |
337 | } | |
338 | ||
339 | /* | |
340 | * At what user virtual address is page expected in vma? | |
341 | * Caller should check the page is actually part of the vma. | |
342 | */ | |
343 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) | |
344 | { | |
345 | if (PageAnon(page)) | |
346 | ; | |
347 | else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { | |
348 | if (!vma->vm_file || | |
349 | vma->vm_file->f_mapping != page->mapping) | |
350 | return -EFAULT; | |
351 | } else | |
352 | return -EFAULT; | |
353 | return vma_address(page, vma); | |
354 | } | |
355 | ||
356 | /* | |
357 | * Check that @page is mapped at @address into @mm. | |
358 | * | |
359 | * If @sync is false, page_check_address may perform a racy check to avoid | |
360 | * the page table lock when the pte is not present (helpful when reclaiming | |
361 | * highly shared pages). | |
362 | * | |
363 | * On success returns with pte mapped and locked. | |
364 | */ | |
365 | pte_t *page_check_address(struct page *page, struct mm_struct *mm, | |
366 | unsigned long address, spinlock_t **ptlp, int sync) | |
367 | { | |
368 | pgd_t *pgd; | |
369 | pud_t *pud; | |
370 | pmd_t *pmd; | |
371 | pte_t *pte; | |
372 | spinlock_t *ptl; | |
373 | ||
374 | pgd = pgd_offset(mm, address); | |
375 | if (!pgd_present(*pgd)) | |
376 | return NULL; | |
377 | ||
378 | pud = pud_offset(pgd, address); | |
379 | if (!pud_present(*pud)) | |
380 | return NULL; | |
381 | ||
382 | pmd = pmd_offset(pud, address); | |
383 | if (!pmd_present(*pmd)) | |
384 | return NULL; | |
385 | ||
386 | pte = pte_offset_map(pmd, address); | |
387 | /* Make a quick check before getting the lock */ | |
388 | if (!sync && !pte_present(*pte)) { | |
389 | pte_unmap(pte); | |
390 | return NULL; | |
391 | } | |
392 | ||
393 | ptl = pte_lockptr(mm, pmd); | |
394 | spin_lock(ptl); | |
395 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { | |
396 | *ptlp = ptl; | |
397 | return pte; | |
398 | } | |
399 | pte_unmap_unlock(pte, ptl); | |
400 | return NULL; | |
401 | } | |
402 | ||
403 | /** | |
404 | * page_mapped_in_vma - check whether a page is really mapped in a VMA | |
405 | * @page: the page to test | |
406 | * @vma: the VMA to test | |
407 | * | |
408 | * Returns 1 if the page is mapped into the page tables of the VMA, 0 | |
409 | * if the page is not mapped into the page tables of this VMA. Only | |
410 | * valid for normal file or anonymous VMAs. | |
411 | */ | |
412 | int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma) | |
413 | { | |
414 | unsigned long address; | |
415 | pte_t *pte; | |
416 | spinlock_t *ptl; | |
417 | ||
418 | address = vma_address(page, vma); | |
419 | if (address == -EFAULT) /* out of vma range */ | |
420 | return 0; | |
421 | pte = page_check_address(page, vma->vm_mm, address, &ptl, 1); | |
422 | if (!pte) /* the page is not in this mm */ | |
423 | return 0; | |
424 | pte_unmap_unlock(pte, ptl); | |
425 | ||
426 | return 1; | |
427 | } | |
428 | ||
429 | /* | |
430 | * Subfunctions of page_referenced: page_referenced_one called | |
431 | * repeatedly from either page_referenced_anon or page_referenced_file. | |
432 | */ | |
433 | int page_referenced_one(struct page *page, struct vm_area_struct *vma, | |
434 | unsigned long address, unsigned int *mapcount, | |
435 | unsigned long *vm_flags) | |
436 | { | |
437 | struct mm_struct *mm = vma->vm_mm; | |
438 | pte_t *pte; | |
439 | spinlock_t *ptl; | |
440 | int referenced = 0; | |
441 | ||
442 | pte = page_check_address(page, mm, address, &ptl, 0); | |
443 | if (!pte) | |
444 | goto out; | |
445 | ||
446 | /* | |
447 | * Don't want to elevate referenced for mlocked page that gets this far, | |
448 | * in order that it progresses to try_to_unmap and is moved to the | |
449 | * unevictable list. | |
450 | */ | |
451 | if (vma->vm_flags & VM_LOCKED) { | |
452 | *mapcount = 1; /* break early from loop */ | |
453 | *vm_flags |= VM_LOCKED; | |
454 | goto out_unmap; | |
455 | } | |
456 | ||
457 | if (ptep_clear_flush_young_notify(vma, address, pte)) { | |
458 | /* | |
459 | * Don't treat a reference through a sequentially read | |
460 | * mapping as such. If the page has been used in | |
461 | * another mapping, we will catch it; if this other | |
462 | * mapping is already gone, the unmap path will have | |
463 | * set PG_referenced or activated the page. | |
464 | */ | |
465 | if (likely(!VM_SequentialReadHint(vma))) | |
466 | referenced++; | |
467 | } | |
468 | ||
469 | /* Pretend the page is referenced if the task has the | |
470 | swap token and is in the middle of a page fault. */ | |
471 | if (mm != current->mm && has_swap_token(mm) && | |
472 | rwsem_is_locked(&mm->mmap_sem)) | |
473 | referenced++; | |
474 | ||
475 | out_unmap: | |
476 | (*mapcount)--; | |
477 | pte_unmap_unlock(pte, ptl); | |
478 | ||
479 | if (referenced) | |
480 | *vm_flags |= vma->vm_flags; | |
481 | out: | |
482 | return referenced; | |
483 | } | |
484 | ||
485 | static int page_referenced_anon(struct page *page, | |
486 | struct mem_cgroup *mem_cont, | |
487 | unsigned long *vm_flags) | |
488 | { | |
489 | unsigned int mapcount; | |
490 | struct anon_vma *anon_vma; | |
491 | struct anon_vma_chain *avc; | |
492 | int referenced = 0; | |
493 | ||
494 | anon_vma = page_lock_anon_vma(page); | |
495 | if (!anon_vma) | |
496 | return referenced; | |
497 | ||
498 | mapcount = page_mapcount(page); | |
499 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { | |
500 | struct vm_area_struct *vma = avc->vma; | |
501 | unsigned long address = vma_address(page, vma); | |
502 | if (address == -EFAULT) | |
503 | continue; | |
504 | /* | |
505 | * If we are reclaiming on behalf of a cgroup, skip | |
506 | * counting on behalf of references from different | |
507 | * cgroups | |
508 | */ | |
509 | if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont)) | |
510 | continue; | |
511 | referenced += page_referenced_one(page, vma, address, | |
512 | &mapcount, vm_flags); | |
513 | if (!mapcount) | |
514 | break; | |
515 | } | |
516 | ||
517 | page_unlock_anon_vma(anon_vma); | |
518 | return referenced; | |
519 | } | |
520 | ||
521 | /** | |
522 | * page_referenced_file - referenced check for object-based rmap | |
523 | * @page: the page we're checking references on. | |
524 | * @mem_cont: target memory controller | |
525 | * @vm_flags: collect encountered vma->vm_flags who actually referenced the page | |
526 | * | |
527 | * For an object-based mapped page, find all the places it is mapped and | |
528 | * check/clear the referenced flag. This is done by following the page->mapping | |
529 | * pointer, then walking the chain of vmas it holds. It returns the number | |
530 | * of references it found. | |
531 | * | |
532 | * This function is only called from page_referenced for object-based pages. | |
533 | */ | |
534 | static int page_referenced_file(struct page *page, | |
535 | struct mem_cgroup *mem_cont, | |
536 | unsigned long *vm_flags) | |
537 | { | |
538 | unsigned int mapcount; | |
539 | struct address_space *mapping = page->mapping; | |
540 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
541 | struct vm_area_struct *vma; | |
542 | struct prio_tree_iter iter; | |
543 | int referenced = 0; | |
544 | ||
545 | /* | |
546 | * The caller's checks on page->mapping and !PageAnon have made | |
547 | * sure that this is a file page: the check for page->mapping | |
548 | * excludes the case just before it gets set on an anon page. | |
549 | */ | |
550 | BUG_ON(PageAnon(page)); | |
551 | ||
552 | /* | |
553 | * The page lock not only makes sure that page->mapping cannot | |
554 | * suddenly be NULLified by truncation, it makes sure that the | |
555 | * structure at mapping cannot be freed and reused yet, | |
556 | * so we can safely take mapping->i_mmap_lock. | |
557 | */ | |
558 | BUG_ON(!PageLocked(page)); | |
559 | ||
560 | spin_lock(&mapping->i_mmap_lock); | |
561 | ||
562 | /* | |
563 | * i_mmap_lock does not stabilize mapcount at all, but mapcount | |
564 | * is more likely to be accurate if we note it after spinning. | |
565 | */ | |
566 | mapcount = page_mapcount(page); | |
567 | ||
568 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
569 | unsigned long address = vma_address(page, vma); | |
570 | if (address == -EFAULT) | |
571 | continue; | |
572 | /* | |
573 | * If we are reclaiming on behalf of a cgroup, skip | |
574 | * counting on behalf of references from different | |
575 | * cgroups | |
576 | */ | |
577 | if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont)) | |
578 | continue; | |
579 | referenced += page_referenced_one(page, vma, address, | |
580 | &mapcount, vm_flags); | |
581 | if (!mapcount) | |
582 | break; | |
583 | } | |
584 | ||
585 | spin_unlock(&mapping->i_mmap_lock); | |
586 | return referenced; | |
587 | } | |
588 | ||
589 | /** | |
590 | * page_referenced - test if the page was referenced | |
591 | * @page: the page to test | |
592 | * @is_locked: caller holds lock on the page | |
593 | * @mem_cont: target memory controller | |
594 | * @vm_flags: collect encountered vma->vm_flags who actually referenced the page | |
595 | * | |
596 | * Quick test_and_clear_referenced for all mappings to a page, | |
597 | * returns the number of ptes which referenced the page. | |
598 | */ | |
599 | int page_referenced(struct page *page, | |
600 | int is_locked, | |
601 | struct mem_cgroup *mem_cont, | |
602 | unsigned long *vm_flags) | |
603 | { | |
604 | int referenced = 0; | |
605 | int we_locked = 0; | |
606 | ||
607 | *vm_flags = 0; | |
608 | if (page_mapped(page) && page_rmapping(page)) { | |
609 | if (!is_locked && (!PageAnon(page) || PageKsm(page))) { | |
610 | we_locked = trylock_page(page); | |
611 | if (!we_locked) { | |
612 | referenced++; | |
613 | goto out; | |
614 | } | |
615 | } | |
616 | if (unlikely(PageKsm(page))) | |
617 | referenced += page_referenced_ksm(page, mem_cont, | |
618 | vm_flags); | |
619 | else if (PageAnon(page)) | |
620 | referenced += page_referenced_anon(page, mem_cont, | |
621 | vm_flags); | |
622 | else if (page->mapping) | |
623 | referenced += page_referenced_file(page, mem_cont, | |
624 | vm_flags); | |
625 | if (we_locked) | |
626 | unlock_page(page); | |
627 | } | |
628 | out: | |
629 | if (page_test_and_clear_young(page)) | |
630 | referenced++; | |
631 | ||
632 | return referenced; | |
633 | } | |
634 | ||
635 | static int page_mkclean_one(struct page *page, struct vm_area_struct *vma, | |
636 | unsigned long address) | |
637 | { | |
638 | struct mm_struct *mm = vma->vm_mm; | |
639 | pte_t *pte; | |
640 | spinlock_t *ptl; | |
641 | int ret = 0; | |
642 | ||
643 | pte = page_check_address(page, mm, address, &ptl, 1); | |
644 | if (!pte) | |
645 | goto out; | |
646 | ||
647 | if (pte_dirty(*pte) || pte_write(*pte)) { | |
648 | pte_t entry; | |
649 | ||
650 | flush_cache_page(vma, address, pte_pfn(*pte)); | |
651 | entry = ptep_clear_flush_notify(vma, address, pte); | |
652 | entry = pte_wrprotect(entry); | |
653 | entry = pte_mkclean(entry); | |
654 | set_pte_at(mm, address, pte, entry); | |
655 | ret = 1; | |
656 | } | |
657 | ||
658 | pte_unmap_unlock(pte, ptl); | |
659 | out: | |
660 | return ret; | |
661 | } | |
662 | ||
663 | static int page_mkclean_file(struct address_space *mapping, struct page *page) | |
664 | { | |
665 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
666 | struct vm_area_struct *vma; | |
667 | struct prio_tree_iter iter; | |
668 | int ret = 0; | |
669 | ||
670 | BUG_ON(PageAnon(page)); | |
671 | ||
672 | spin_lock(&mapping->i_mmap_lock); | |
673 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
674 | if (vma->vm_flags & VM_SHARED) { | |
675 | unsigned long address = vma_address(page, vma); | |
676 | if (address == -EFAULT) | |
677 | continue; | |
678 | ret += page_mkclean_one(page, vma, address); | |
679 | } | |
680 | } | |
681 | spin_unlock(&mapping->i_mmap_lock); | |
682 | return ret; | |
683 | } | |
684 | ||
685 | int page_mkclean(struct page *page) | |
686 | { | |
687 | int ret = 0; | |
688 | ||
689 | BUG_ON(!PageLocked(page)); | |
690 | ||
691 | if (page_mapped(page)) { | |
692 | struct address_space *mapping = page_mapping(page); | |
693 | if (mapping) { | |
694 | ret = page_mkclean_file(mapping, page); | |
695 | if (page_test_dirty(page)) { | |
696 | page_clear_dirty(page); | |
697 | ret = 1; | |
698 | } | |
699 | } | |
700 | } | |
701 | ||
702 | return ret; | |
703 | } | |
704 | EXPORT_SYMBOL_GPL(page_mkclean); | |
705 | ||
706 | /** | |
707 | * page_move_anon_rmap - move a page to our anon_vma | |
708 | * @page: the page to move to our anon_vma | |
709 | * @vma: the vma the page belongs to | |
710 | * @address: the user virtual address mapped | |
711 | * | |
712 | * When a page belongs exclusively to one process after a COW event, | |
713 | * that page can be moved into the anon_vma that belongs to just that | |
714 | * process, so the rmap code will not search the parent or sibling | |
715 | * processes. | |
716 | */ | |
717 | void page_move_anon_rmap(struct page *page, | |
718 | struct vm_area_struct *vma, unsigned long address) | |
719 | { | |
720 | struct anon_vma *anon_vma = vma->anon_vma; | |
721 | ||
722 | VM_BUG_ON(!PageLocked(page)); | |
723 | VM_BUG_ON(!anon_vma); | |
724 | VM_BUG_ON(page->index != linear_page_index(vma, address)); | |
725 | ||
726 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
727 | page->mapping = (struct address_space *) anon_vma; | |
728 | } | |
729 | ||
730 | /** | |
731 | * __page_set_anon_rmap - setup new anonymous rmap | |
732 | * @page: the page to add the mapping to | |
733 | * @vma: the vm area in which the mapping is added | |
734 | * @address: the user virtual address mapped | |
735 | * @exclusive: the page is exclusively owned by the current process | |
736 | */ | |
737 | static void __page_set_anon_rmap(struct page *page, | |
738 | struct vm_area_struct *vma, unsigned long address, int exclusive) | |
739 | { | |
740 | struct anon_vma *anon_vma = vma->anon_vma; | |
741 | ||
742 | BUG_ON(!anon_vma); | |
743 | ||
744 | /* | |
745 | * If the page isn't exclusively mapped into this vma, | |
746 | * we must use the _oldest_ possible anon_vma for the | |
747 | * page mapping! | |
748 | * | |
749 | * So take the last AVC chain entry in the vma, which is | |
750 | * the deepest ancestor, and use the anon_vma from that. | |
751 | */ | |
752 | if (!exclusive) { | |
753 | struct anon_vma_chain *avc; | |
754 | avc = list_entry(vma->anon_vma_chain.prev, struct anon_vma_chain, same_vma); | |
755 | anon_vma = avc->anon_vma; | |
756 | } | |
757 | ||
758 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
759 | page->mapping = (struct address_space *) anon_vma; | |
760 | page->index = linear_page_index(vma, address); | |
761 | } | |
762 | ||
763 | /** | |
764 | * __page_check_anon_rmap - sanity check anonymous rmap addition | |
765 | * @page: the page to add the mapping to | |
766 | * @vma: the vm area in which the mapping is added | |
767 | * @address: the user virtual address mapped | |
768 | */ | |
769 | static void __page_check_anon_rmap(struct page *page, | |
770 | struct vm_area_struct *vma, unsigned long address) | |
771 | { | |
772 | #ifdef CONFIG_DEBUG_VM | |
773 | /* | |
774 | * The page's anon-rmap details (mapping and index) are guaranteed to | |
775 | * be set up correctly at this point. | |
776 | * | |
777 | * We have exclusion against page_add_anon_rmap because the caller | |
778 | * always holds the page locked, except if called from page_dup_rmap, | |
779 | * in which case the page is already known to be setup. | |
780 | * | |
781 | * We have exclusion against page_add_new_anon_rmap because those pages | |
782 | * are initially only visible via the pagetables, and the pte is locked | |
783 | * over the call to page_add_new_anon_rmap. | |
784 | */ | |
785 | BUG_ON(page->index != linear_page_index(vma, address)); | |
786 | #endif | |
787 | } | |
788 | ||
789 | /** | |
790 | * page_add_anon_rmap - add pte mapping to an anonymous page | |
791 | * @page: the page to add the mapping to | |
792 | * @vma: the vm area in which the mapping is added | |
793 | * @address: the user virtual address mapped | |
794 | * | |
795 | * The caller needs to hold the pte lock, and the page must be locked in | |
796 | * the anon_vma case: to serialize mapping,index checking after setting, | |
797 | * and to ensure that PageAnon is not being upgraded racily to PageKsm | |
798 | * (but PageKsm is never downgraded to PageAnon). | |
799 | */ | |
800 | void page_add_anon_rmap(struct page *page, | |
801 | struct vm_area_struct *vma, unsigned long address) | |
802 | { | |
803 | int first = atomic_inc_and_test(&page->_mapcount); | |
804 | if (first) | |
805 | __inc_zone_page_state(page, NR_ANON_PAGES); | |
806 | if (unlikely(PageKsm(page))) | |
807 | return; | |
808 | ||
809 | VM_BUG_ON(!PageLocked(page)); | |
810 | VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); | |
811 | if (first) | |
812 | __page_set_anon_rmap(page, vma, address, 0); | |
813 | else | |
814 | __page_check_anon_rmap(page, vma, address); | |
815 | } | |
816 | ||
817 | /** | |
818 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page | |
819 | * @page: the page to add the mapping to | |
820 | * @vma: the vm area in which the mapping is added | |
821 | * @address: the user virtual address mapped | |
822 | * | |
823 | * Same as page_add_anon_rmap but must only be called on *new* pages. | |
824 | * This means the inc-and-test can be bypassed. | |
825 | * Page does not have to be locked. | |
826 | */ | |
827 | void page_add_new_anon_rmap(struct page *page, | |
828 | struct vm_area_struct *vma, unsigned long address) | |
829 | { | |
830 | VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); | |
831 | SetPageSwapBacked(page); | |
832 | atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */ | |
833 | __inc_zone_page_state(page, NR_ANON_PAGES); | |
834 | __page_set_anon_rmap(page, vma, address, 1); | |
835 | if (page_evictable(page, vma)) | |
836 | lru_cache_add_lru(page, LRU_ACTIVE_ANON); | |
837 | else | |
838 | add_page_to_unevictable_list(page); | |
839 | } | |
840 | ||
841 | /** | |
842 | * page_add_file_rmap - add pte mapping to a file page | |
843 | * @page: the page to add the mapping to | |
844 | * | |
845 | * The caller needs to hold the pte lock. | |
846 | */ | |
847 | void page_add_file_rmap(struct page *page) | |
848 | { | |
849 | if (atomic_inc_and_test(&page->_mapcount)) { | |
850 | __inc_zone_page_state(page, NR_FILE_MAPPED); | |
851 | mem_cgroup_update_file_mapped(page, 1); | |
852 | } | |
853 | } | |
854 | ||
855 | /** | |
856 | * page_remove_rmap - take down pte mapping from a page | |
857 | * @page: page to remove mapping from | |
858 | * | |
859 | * The caller needs to hold the pte lock. | |
860 | */ | |
861 | void page_remove_rmap(struct page *page) | |
862 | { | |
863 | /* page still mapped by someone else? */ | |
864 | if (!atomic_add_negative(-1, &page->_mapcount)) | |
865 | return; | |
866 | ||
867 | /* | |
868 | * Now that the last pte has gone, s390 must transfer dirty | |
869 | * flag from storage key to struct page. We can usually skip | |
870 | * this if the page is anon, so about to be freed; but perhaps | |
871 | * not if it's in swapcache - there might be another pte slot | |
872 | * containing the swap entry, but page not yet written to swap. | |
873 | */ | |
874 | if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) { | |
875 | page_clear_dirty(page); | |
876 | set_page_dirty(page); | |
877 | } | |
878 | if (PageAnon(page)) { | |
879 | mem_cgroup_uncharge_page(page); | |
880 | __dec_zone_page_state(page, NR_ANON_PAGES); | |
881 | } else { | |
882 | __dec_zone_page_state(page, NR_FILE_MAPPED); | |
883 | mem_cgroup_update_file_mapped(page, -1); | |
884 | } | |
885 | /* | |
886 | * It would be tidy to reset the PageAnon mapping here, | |
887 | * but that might overwrite a racing page_add_anon_rmap | |
888 | * which increments mapcount after us but sets mapping | |
889 | * before us: so leave the reset to free_hot_cold_page, | |
890 | * and remember that it's only reliable while mapped. | |
891 | * Leaving it set also helps swapoff to reinstate ptes | |
892 | * faster for those pages still in swapcache. | |
893 | */ | |
894 | } | |
895 | ||
896 | /* | |
897 | * Subfunctions of try_to_unmap: try_to_unmap_one called | |
898 | * repeatedly from either try_to_unmap_anon or try_to_unmap_file. | |
899 | */ | |
900 | int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, | |
901 | unsigned long address, enum ttu_flags flags) | |
902 | { | |
903 | struct mm_struct *mm = vma->vm_mm; | |
904 | pte_t *pte; | |
905 | pte_t pteval; | |
906 | spinlock_t *ptl; | |
907 | int ret = SWAP_AGAIN; | |
908 | ||
909 | pte = page_check_address(page, mm, address, &ptl, 0); | |
910 | if (!pte) | |
911 | goto out; | |
912 | ||
913 | /* | |
914 | * If the page is mlock()d, we cannot swap it out. | |
915 | * If it's recently referenced (perhaps page_referenced | |
916 | * skipped over this mm) then we should reactivate it. | |
917 | */ | |
918 | if (!(flags & TTU_IGNORE_MLOCK)) { | |
919 | if (vma->vm_flags & VM_LOCKED) | |
920 | goto out_mlock; | |
921 | ||
922 | if (TTU_ACTION(flags) == TTU_MUNLOCK) | |
923 | goto out_unmap; | |
924 | } | |
925 | if (!(flags & TTU_IGNORE_ACCESS)) { | |
926 | if (ptep_clear_flush_young_notify(vma, address, pte)) { | |
927 | ret = SWAP_FAIL; | |
928 | goto out_unmap; | |
929 | } | |
930 | } | |
931 | ||
932 | /* Nuke the page table entry. */ | |
933 | flush_cache_page(vma, address, page_to_pfn(page)); | |
934 | pteval = ptep_clear_flush_notify(vma, address, pte); | |
935 | ||
936 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
937 | if (pte_dirty(pteval)) | |
938 | set_page_dirty(page); | |
939 | ||
940 | /* Update high watermark before we lower rss */ | |
941 | update_hiwater_rss(mm); | |
942 | ||
943 | if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) { | |
944 | if (PageAnon(page)) | |
945 | dec_mm_counter(mm, MM_ANONPAGES); | |
946 | else | |
947 | dec_mm_counter(mm, MM_FILEPAGES); | |
948 | set_pte_at(mm, address, pte, | |
949 | swp_entry_to_pte(make_hwpoison_entry(page))); | |
950 | } else if (PageAnon(page)) { | |
951 | swp_entry_t entry = { .val = page_private(page) }; | |
952 | ||
953 | if (PageSwapCache(page)) { | |
954 | /* | |
955 | * Store the swap location in the pte. | |
956 | * See handle_pte_fault() ... | |
957 | */ | |
958 | if (swap_duplicate(entry) < 0) { | |
959 | set_pte_at(mm, address, pte, pteval); | |
960 | ret = SWAP_FAIL; | |
961 | goto out_unmap; | |
962 | } | |
963 | if (list_empty(&mm->mmlist)) { | |
964 | spin_lock(&mmlist_lock); | |
965 | if (list_empty(&mm->mmlist)) | |
966 | list_add(&mm->mmlist, &init_mm.mmlist); | |
967 | spin_unlock(&mmlist_lock); | |
968 | } | |
969 | dec_mm_counter(mm, MM_ANONPAGES); | |
970 | inc_mm_counter(mm, MM_SWAPENTS); | |
971 | } else if (PAGE_MIGRATION) { | |
972 | /* | |
973 | * Store the pfn of the page in a special migration | |
974 | * pte. do_swap_page() will wait until the migration | |
975 | * pte is removed and then restart fault handling. | |
976 | */ | |
977 | BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION); | |
978 | entry = make_migration_entry(page, pte_write(pteval)); | |
979 | } | |
980 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
981 | BUG_ON(pte_file(*pte)); | |
982 | } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) { | |
983 | /* Establish migration entry for a file page */ | |
984 | swp_entry_t entry; | |
985 | entry = make_migration_entry(page, pte_write(pteval)); | |
986 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); | |
987 | } else | |
988 | dec_mm_counter(mm, MM_FILEPAGES); | |
989 | ||
990 | page_remove_rmap(page); | |
991 | page_cache_release(page); | |
992 | ||
993 | out_unmap: | |
994 | pte_unmap_unlock(pte, ptl); | |
995 | out: | |
996 | return ret; | |
997 | ||
998 | out_mlock: | |
999 | pte_unmap_unlock(pte, ptl); | |
1000 | ||
1001 | ||
1002 | /* | |
1003 | * We need mmap_sem locking, Otherwise VM_LOCKED check makes | |
1004 | * unstable result and race. Plus, We can't wait here because | |
1005 | * we now hold anon_vma->lock or mapping->i_mmap_lock. | |
1006 | * if trylock failed, the page remain in evictable lru and later | |
1007 | * vmscan could retry to move the page to unevictable lru if the | |
1008 | * page is actually mlocked. | |
1009 | */ | |
1010 | if (down_read_trylock(&vma->vm_mm->mmap_sem)) { | |
1011 | if (vma->vm_flags & VM_LOCKED) { | |
1012 | mlock_vma_page(page); | |
1013 | ret = SWAP_MLOCK; | |
1014 | } | |
1015 | up_read(&vma->vm_mm->mmap_sem); | |
1016 | } | |
1017 | return ret; | |
1018 | } | |
1019 | ||
1020 | /* | |
1021 | * objrmap doesn't work for nonlinear VMAs because the assumption that | |
1022 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. | |
1023 | * Consequently, given a particular page and its ->index, we cannot locate the | |
1024 | * ptes which are mapping that page without an exhaustive linear search. | |
1025 | * | |
1026 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which | |
1027 | * maps the file to which the target page belongs. The ->vm_private_data field | |
1028 | * holds the current cursor into that scan. Successive searches will circulate | |
1029 | * around the vma's virtual address space. | |
1030 | * | |
1031 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, | |
1032 | * more scanning pressure is placed against them as well. Eventually pages | |
1033 | * will become fully unmapped and are eligible for eviction. | |
1034 | * | |
1035 | * For very sparsely populated VMAs this is a little inefficient - chances are | |
1036 | * there there won't be many ptes located within the scan cluster. In this case | |
1037 | * maybe we could scan further - to the end of the pte page, perhaps. | |
1038 | * | |
1039 | * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can | |
1040 | * acquire it without blocking. If vma locked, mlock the pages in the cluster, | |
1041 | * rather than unmapping them. If we encounter the "check_page" that vmscan is | |
1042 | * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN. | |
1043 | */ | |
1044 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) | |
1045 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) | |
1046 | ||
1047 | static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount, | |
1048 | struct vm_area_struct *vma, struct page *check_page) | |
1049 | { | |
1050 | struct mm_struct *mm = vma->vm_mm; | |
1051 | pgd_t *pgd; | |
1052 | pud_t *pud; | |
1053 | pmd_t *pmd; | |
1054 | pte_t *pte; | |
1055 | pte_t pteval; | |
1056 | spinlock_t *ptl; | |
1057 | struct page *page; | |
1058 | unsigned long address; | |
1059 | unsigned long end; | |
1060 | int ret = SWAP_AGAIN; | |
1061 | int locked_vma = 0; | |
1062 | ||
1063 | address = (vma->vm_start + cursor) & CLUSTER_MASK; | |
1064 | end = address + CLUSTER_SIZE; | |
1065 | if (address < vma->vm_start) | |
1066 | address = vma->vm_start; | |
1067 | if (end > vma->vm_end) | |
1068 | end = vma->vm_end; | |
1069 | ||
1070 | pgd = pgd_offset(mm, address); | |
1071 | if (!pgd_present(*pgd)) | |
1072 | return ret; | |
1073 | ||
1074 | pud = pud_offset(pgd, address); | |
1075 | if (!pud_present(*pud)) | |
1076 | return ret; | |
1077 | ||
1078 | pmd = pmd_offset(pud, address); | |
1079 | if (!pmd_present(*pmd)) | |
1080 | return ret; | |
1081 | ||
1082 | /* | |
1083 | * If we can acquire the mmap_sem for read, and vma is VM_LOCKED, | |
1084 | * keep the sem while scanning the cluster for mlocking pages. | |
1085 | */ | |
1086 | if (down_read_trylock(&vma->vm_mm->mmap_sem)) { | |
1087 | locked_vma = (vma->vm_flags & VM_LOCKED); | |
1088 | if (!locked_vma) | |
1089 | up_read(&vma->vm_mm->mmap_sem); /* don't need it */ | |
1090 | } | |
1091 | ||
1092 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
1093 | ||
1094 | /* Update high watermark before we lower rss */ | |
1095 | update_hiwater_rss(mm); | |
1096 | ||
1097 | for (; address < end; pte++, address += PAGE_SIZE) { | |
1098 | if (!pte_present(*pte)) | |
1099 | continue; | |
1100 | page = vm_normal_page(vma, address, *pte); | |
1101 | BUG_ON(!page || PageAnon(page)); | |
1102 | ||
1103 | if (locked_vma) { | |
1104 | mlock_vma_page(page); /* no-op if already mlocked */ | |
1105 | if (page == check_page) | |
1106 | ret = SWAP_MLOCK; | |
1107 | continue; /* don't unmap */ | |
1108 | } | |
1109 | ||
1110 | if (ptep_clear_flush_young_notify(vma, address, pte)) | |
1111 | continue; | |
1112 | ||
1113 | /* Nuke the page table entry. */ | |
1114 | flush_cache_page(vma, address, pte_pfn(*pte)); | |
1115 | pteval = ptep_clear_flush_notify(vma, address, pte); | |
1116 | ||
1117 | /* If nonlinear, store the file page offset in the pte. */ | |
1118 | if (page->index != linear_page_index(vma, address)) | |
1119 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); | |
1120 | ||
1121 | /* Move the dirty bit to the physical page now the pte is gone. */ | |
1122 | if (pte_dirty(pteval)) | |
1123 | set_page_dirty(page); | |
1124 | ||
1125 | page_remove_rmap(page); | |
1126 | page_cache_release(page); | |
1127 | dec_mm_counter(mm, MM_FILEPAGES); | |
1128 | (*mapcount)--; | |
1129 | } | |
1130 | pte_unmap_unlock(pte - 1, ptl); | |
1131 | if (locked_vma) | |
1132 | up_read(&vma->vm_mm->mmap_sem); | |
1133 | return ret; | |
1134 | } | |
1135 | ||
1136 | /** | |
1137 | * try_to_unmap_anon - unmap or unlock anonymous page using the object-based | |
1138 | * rmap method | |
1139 | * @page: the page to unmap/unlock | |
1140 | * @flags: action and flags | |
1141 | * | |
1142 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
1143 | * contained in the anon_vma struct it points to. | |
1144 | * | |
1145 | * This function is only called from try_to_unmap/try_to_munlock for | |
1146 | * anonymous pages. | |
1147 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma | |
1148 | * where the page was found will be held for write. So, we won't recheck | |
1149 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be | |
1150 | * 'LOCKED. | |
1151 | */ | |
1152 | static int try_to_unmap_anon(struct page *page, enum ttu_flags flags) | |
1153 | { | |
1154 | struct anon_vma *anon_vma; | |
1155 | struct anon_vma_chain *avc; | |
1156 | int ret = SWAP_AGAIN; | |
1157 | ||
1158 | anon_vma = page_lock_anon_vma(page); | |
1159 | if (!anon_vma) | |
1160 | return ret; | |
1161 | ||
1162 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { | |
1163 | struct vm_area_struct *vma = avc->vma; | |
1164 | unsigned long address = vma_address(page, vma); | |
1165 | if (address == -EFAULT) | |
1166 | continue; | |
1167 | ret = try_to_unmap_one(page, vma, address, flags); | |
1168 | if (ret != SWAP_AGAIN || !page_mapped(page)) | |
1169 | break; | |
1170 | } | |
1171 | ||
1172 | page_unlock_anon_vma(anon_vma); | |
1173 | return ret; | |
1174 | } | |
1175 | ||
1176 | /** | |
1177 | * try_to_unmap_file - unmap/unlock file page using the object-based rmap method | |
1178 | * @page: the page to unmap/unlock | |
1179 | * @flags: action and flags | |
1180 | * | |
1181 | * Find all the mappings of a page using the mapping pointer and the vma chains | |
1182 | * contained in the address_space struct it points to. | |
1183 | * | |
1184 | * This function is only called from try_to_unmap/try_to_munlock for | |
1185 | * object-based pages. | |
1186 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma | |
1187 | * where the page was found will be held for write. So, we won't recheck | |
1188 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be | |
1189 | * 'LOCKED. | |
1190 | */ | |
1191 | static int try_to_unmap_file(struct page *page, enum ttu_flags flags) | |
1192 | { | |
1193 | struct address_space *mapping = page->mapping; | |
1194 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
1195 | struct vm_area_struct *vma; | |
1196 | struct prio_tree_iter iter; | |
1197 | int ret = SWAP_AGAIN; | |
1198 | unsigned long cursor; | |
1199 | unsigned long max_nl_cursor = 0; | |
1200 | unsigned long max_nl_size = 0; | |
1201 | unsigned int mapcount; | |
1202 | ||
1203 | spin_lock(&mapping->i_mmap_lock); | |
1204 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
1205 | unsigned long address = vma_address(page, vma); | |
1206 | if (address == -EFAULT) | |
1207 | continue; | |
1208 | ret = try_to_unmap_one(page, vma, address, flags); | |
1209 | if (ret != SWAP_AGAIN || !page_mapped(page)) | |
1210 | goto out; | |
1211 | } | |
1212 | ||
1213 | if (list_empty(&mapping->i_mmap_nonlinear)) | |
1214 | goto out; | |
1215 | ||
1216 | /* | |
1217 | * We don't bother to try to find the munlocked page in nonlinears. | |
1218 | * It's costly. Instead, later, page reclaim logic may call | |
1219 | * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily. | |
1220 | */ | |
1221 | if (TTU_ACTION(flags) == TTU_MUNLOCK) | |
1222 | goto out; | |
1223 | ||
1224 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
1225 | shared.vm_set.list) { | |
1226 | cursor = (unsigned long) vma->vm_private_data; | |
1227 | if (cursor > max_nl_cursor) | |
1228 | max_nl_cursor = cursor; | |
1229 | cursor = vma->vm_end - vma->vm_start; | |
1230 | if (cursor > max_nl_size) | |
1231 | max_nl_size = cursor; | |
1232 | } | |
1233 | ||
1234 | if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */ | |
1235 | ret = SWAP_FAIL; | |
1236 | goto out; | |
1237 | } | |
1238 | ||
1239 | /* | |
1240 | * We don't try to search for this page in the nonlinear vmas, | |
1241 | * and page_referenced wouldn't have found it anyway. Instead | |
1242 | * just walk the nonlinear vmas trying to age and unmap some. | |
1243 | * The mapcount of the page we came in with is irrelevant, | |
1244 | * but even so use it as a guide to how hard we should try? | |
1245 | */ | |
1246 | mapcount = page_mapcount(page); | |
1247 | if (!mapcount) | |
1248 | goto out; | |
1249 | cond_resched_lock(&mapping->i_mmap_lock); | |
1250 | ||
1251 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; | |
1252 | if (max_nl_cursor == 0) | |
1253 | max_nl_cursor = CLUSTER_SIZE; | |
1254 | ||
1255 | do { | |
1256 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, | |
1257 | shared.vm_set.list) { | |
1258 | cursor = (unsigned long) vma->vm_private_data; | |
1259 | while ( cursor < max_nl_cursor && | |
1260 | cursor < vma->vm_end - vma->vm_start) { | |
1261 | if (try_to_unmap_cluster(cursor, &mapcount, | |
1262 | vma, page) == SWAP_MLOCK) | |
1263 | ret = SWAP_MLOCK; | |
1264 | cursor += CLUSTER_SIZE; | |
1265 | vma->vm_private_data = (void *) cursor; | |
1266 | if ((int)mapcount <= 0) | |
1267 | goto out; | |
1268 | } | |
1269 | vma->vm_private_data = (void *) max_nl_cursor; | |
1270 | } | |
1271 | cond_resched_lock(&mapping->i_mmap_lock); | |
1272 | max_nl_cursor += CLUSTER_SIZE; | |
1273 | } while (max_nl_cursor <= max_nl_size); | |
1274 | ||
1275 | /* | |
1276 | * Don't loop forever (perhaps all the remaining pages are | |
1277 | * in locked vmas). Reset cursor on all unreserved nonlinear | |
1278 | * vmas, now forgetting on which ones it had fallen behind. | |
1279 | */ | |
1280 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) | |
1281 | vma->vm_private_data = NULL; | |
1282 | out: | |
1283 | spin_unlock(&mapping->i_mmap_lock); | |
1284 | return ret; | |
1285 | } | |
1286 | ||
1287 | /** | |
1288 | * try_to_unmap - try to remove all page table mappings to a page | |
1289 | * @page: the page to get unmapped | |
1290 | * @flags: action and flags | |
1291 | * | |
1292 | * Tries to remove all the page table entries which are mapping this | |
1293 | * page, used in the pageout path. Caller must hold the page lock. | |
1294 | * Return values are: | |
1295 | * | |
1296 | * SWAP_SUCCESS - we succeeded in removing all mappings | |
1297 | * SWAP_AGAIN - we missed a mapping, try again later | |
1298 | * SWAP_FAIL - the page is unswappable | |
1299 | * SWAP_MLOCK - page is mlocked. | |
1300 | */ | |
1301 | int try_to_unmap(struct page *page, enum ttu_flags flags) | |
1302 | { | |
1303 | int ret; | |
1304 | ||
1305 | BUG_ON(!PageLocked(page)); | |
1306 | ||
1307 | if (unlikely(PageKsm(page))) | |
1308 | ret = try_to_unmap_ksm(page, flags); | |
1309 | else if (PageAnon(page)) | |
1310 | ret = try_to_unmap_anon(page, flags); | |
1311 | else | |
1312 | ret = try_to_unmap_file(page, flags); | |
1313 | if (ret != SWAP_MLOCK && !page_mapped(page)) | |
1314 | ret = SWAP_SUCCESS; | |
1315 | return ret; | |
1316 | } | |
1317 | ||
1318 | /** | |
1319 | * try_to_munlock - try to munlock a page | |
1320 | * @page: the page to be munlocked | |
1321 | * | |
1322 | * Called from munlock code. Checks all of the VMAs mapping the page | |
1323 | * to make sure nobody else has this page mlocked. The page will be | |
1324 | * returned with PG_mlocked cleared if no other vmas have it mlocked. | |
1325 | * | |
1326 | * Return values are: | |
1327 | * | |
1328 | * SWAP_AGAIN - no vma is holding page mlocked, or, | |
1329 | * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem | |
1330 | * SWAP_FAIL - page cannot be located at present | |
1331 | * SWAP_MLOCK - page is now mlocked. | |
1332 | */ | |
1333 | int try_to_munlock(struct page *page) | |
1334 | { | |
1335 | VM_BUG_ON(!PageLocked(page) || PageLRU(page)); | |
1336 | ||
1337 | if (unlikely(PageKsm(page))) | |
1338 | return try_to_unmap_ksm(page, TTU_MUNLOCK); | |
1339 | else if (PageAnon(page)) | |
1340 | return try_to_unmap_anon(page, TTU_MUNLOCK); | |
1341 | else | |
1342 | return try_to_unmap_file(page, TTU_MUNLOCK); | |
1343 | } | |
1344 | ||
1345 | #ifdef CONFIG_MIGRATION | |
1346 | /* | |
1347 | * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file(): | |
1348 | * Called by migrate.c to remove migration ptes, but might be used more later. | |
1349 | */ | |
1350 | static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *, | |
1351 | struct vm_area_struct *, unsigned long, void *), void *arg) | |
1352 | { | |
1353 | struct anon_vma *anon_vma; | |
1354 | struct anon_vma_chain *avc; | |
1355 | int ret = SWAP_AGAIN; | |
1356 | ||
1357 | /* | |
1358 | * Note: remove_migration_ptes() cannot use page_lock_anon_vma() | |
1359 | * because that depends on page_mapped(); but not all its usages | |
1360 | * are holding mmap_sem. Users without mmap_sem are required to | |
1361 | * take a reference count to prevent the anon_vma disappearing | |
1362 | */ | |
1363 | anon_vma = page_anon_vma(page); | |
1364 | if (!anon_vma) | |
1365 | return ret; | |
1366 | spin_lock(&anon_vma->lock); | |
1367 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { | |
1368 | struct vm_area_struct *vma = avc->vma; | |
1369 | unsigned long address = vma_address(page, vma); | |
1370 | if (address == -EFAULT) | |
1371 | continue; | |
1372 | ret = rmap_one(page, vma, address, arg); | |
1373 | if (ret != SWAP_AGAIN) | |
1374 | break; | |
1375 | } | |
1376 | spin_unlock(&anon_vma->lock); | |
1377 | return ret; | |
1378 | } | |
1379 | ||
1380 | static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *, | |
1381 | struct vm_area_struct *, unsigned long, void *), void *arg) | |
1382 | { | |
1383 | struct address_space *mapping = page->mapping; | |
1384 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
1385 | struct vm_area_struct *vma; | |
1386 | struct prio_tree_iter iter; | |
1387 | int ret = SWAP_AGAIN; | |
1388 | ||
1389 | if (!mapping) | |
1390 | return ret; | |
1391 | spin_lock(&mapping->i_mmap_lock); | |
1392 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
1393 | unsigned long address = vma_address(page, vma); | |
1394 | if (address == -EFAULT) | |
1395 | continue; | |
1396 | ret = rmap_one(page, vma, address, arg); | |
1397 | if (ret != SWAP_AGAIN) | |
1398 | break; | |
1399 | } | |
1400 | /* | |
1401 | * No nonlinear handling: being always shared, nonlinear vmas | |
1402 | * never contain migration ptes. Decide what to do about this | |
1403 | * limitation to linear when we need rmap_walk() on nonlinear. | |
1404 | */ | |
1405 | spin_unlock(&mapping->i_mmap_lock); | |
1406 | return ret; | |
1407 | } | |
1408 | ||
1409 | int rmap_walk(struct page *page, int (*rmap_one)(struct page *, | |
1410 | struct vm_area_struct *, unsigned long, void *), void *arg) | |
1411 | { | |
1412 | VM_BUG_ON(!PageLocked(page)); | |
1413 | ||
1414 | if (unlikely(PageKsm(page))) | |
1415 | return rmap_walk_ksm(page, rmap_one, arg); | |
1416 | else if (PageAnon(page)) | |
1417 | return rmap_walk_anon(page, rmap_one, arg); | |
1418 | else | |
1419 | return rmap_walk_file(page, rmap_one, arg); | |
1420 | } | |
1421 | #endif /* CONFIG_MIGRATION */ |