]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Memory Migration functionality - linux/mm/migration.c | |
3 | * | |
4 | * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter | |
5 | * | |
6 | * Page migration was first developed in the context of the memory hotplug | |
7 | * project. The main authors of the migration code are: | |
8 | * | |
9 | * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> | |
10 | * Hirokazu Takahashi <taka@valinux.co.jp> | |
11 | * Dave Hansen <haveblue@us.ibm.com> | |
12 | * Christoph Lameter | |
13 | */ | |
14 | ||
15 | #include <linux/migrate.h> | |
16 | #include <linux/module.h> | |
17 | #include <linux/swap.h> | |
18 | #include <linux/swapops.h> | |
19 | #include <linux/pagemap.h> | |
20 | #include <linux/buffer_head.h> | |
21 | #include <linux/mm_inline.h> | |
22 | #include <linux/nsproxy.h> | |
23 | #include <linux/pagevec.h> | |
24 | #include <linux/rmap.h> | |
25 | #include <linux/topology.h> | |
26 | #include <linux/cpu.h> | |
27 | #include <linux/cpuset.h> | |
28 | #include <linux/writeback.h> | |
29 | #include <linux/mempolicy.h> | |
30 | #include <linux/vmalloc.h> | |
31 | #include <linux/security.h> | |
32 | #include <linux/memcontrol.h> | |
33 | #include <linux/syscalls.h> | |
34 | ||
35 | #include "internal.h" | |
36 | ||
37 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) | |
38 | ||
39 | /* | |
40 | * Isolate one page from the LRU lists. If successful put it onto | |
41 | * the indicated list with elevated page count. | |
42 | * | |
43 | * Result: | |
44 | * -EBUSY: page not on LRU list | |
45 | * 0: page removed from LRU list and added to the specified list. | |
46 | */ | |
47 | int isolate_lru_page(struct page *page, struct list_head *pagelist) | |
48 | { | |
49 | int ret = -EBUSY; | |
50 | ||
51 | if (PageLRU(page)) { | |
52 | struct zone *zone = page_zone(page); | |
53 | ||
54 | spin_lock_irq(&zone->lru_lock); | |
55 | if (PageLRU(page) && get_page_unless_zero(page)) { | |
56 | ret = 0; | |
57 | ClearPageLRU(page); | |
58 | if (PageActive(page)) | |
59 | del_page_from_active_list(zone, page); | |
60 | else | |
61 | del_page_from_inactive_list(zone, page); | |
62 | list_add_tail(&page->lru, pagelist); | |
63 | } | |
64 | spin_unlock_irq(&zone->lru_lock); | |
65 | } | |
66 | return ret; | |
67 | } | |
68 | ||
69 | /* | |
70 | * migrate_prep() needs to be called before we start compiling a list of pages | |
71 | * to be migrated using isolate_lru_page(). | |
72 | */ | |
73 | int migrate_prep(void) | |
74 | { | |
75 | /* | |
76 | * Clear the LRU lists so pages can be isolated. | |
77 | * Note that pages may be moved off the LRU after we have | |
78 | * drained them. Those pages will fail to migrate like other | |
79 | * pages that may be busy. | |
80 | */ | |
81 | lru_add_drain_all(); | |
82 | ||
83 | return 0; | |
84 | } | |
85 | ||
86 | static inline void move_to_lru(struct page *page) | |
87 | { | |
88 | if (PageActive(page)) { | |
89 | /* | |
90 | * lru_cache_add_active checks that | |
91 | * the PG_active bit is off. | |
92 | */ | |
93 | ClearPageActive(page); | |
94 | lru_cache_add_active(page); | |
95 | } else { | |
96 | lru_cache_add(page); | |
97 | } | |
98 | put_page(page); | |
99 | } | |
100 | ||
101 | /* | |
102 | * Add isolated pages on the list back to the LRU. | |
103 | * | |
104 | * returns the number of pages put back. | |
105 | */ | |
106 | int putback_lru_pages(struct list_head *l) | |
107 | { | |
108 | struct page *page; | |
109 | struct page *page2; | |
110 | int count = 0; | |
111 | ||
112 | list_for_each_entry_safe(page, page2, l, lru) { | |
113 | list_del(&page->lru); | |
114 | move_to_lru(page); | |
115 | count++; | |
116 | } | |
117 | return count; | |
118 | } | |
119 | ||
120 | /* | |
121 | * Restore a potential migration pte to a working pte entry | |
122 | */ | |
123 | static void remove_migration_pte(struct vm_area_struct *vma, | |
124 | struct page *old, struct page *new) | |
125 | { | |
126 | struct mm_struct *mm = vma->vm_mm; | |
127 | swp_entry_t entry; | |
128 | pgd_t *pgd; | |
129 | pud_t *pud; | |
130 | pmd_t *pmd; | |
131 | pte_t *ptep, pte; | |
132 | spinlock_t *ptl; | |
133 | unsigned long addr = page_address_in_vma(new, vma); | |
134 | ||
135 | if (addr == -EFAULT) | |
136 | return; | |
137 | ||
138 | pgd = pgd_offset(mm, addr); | |
139 | if (!pgd_present(*pgd)) | |
140 | return; | |
141 | ||
142 | pud = pud_offset(pgd, addr); | |
143 | if (!pud_present(*pud)) | |
144 | return; | |
145 | ||
146 | pmd = pmd_offset(pud, addr); | |
147 | if (!pmd_present(*pmd)) | |
148 | return; | |
149 | ||
150 | ptep = pte_offset_map(pmd, addr); | |
151 | ||
152 | if (!is_swap_pte(*ptep)) { | |
153 | pte_unmap(ptep); | |
154 | return; | |
155 | } | |
156 | ||
157 | ptl = pte_lockptr(mm, pmd); | |
158 | spin_lock(ptl); | |
159 | pte = *ptep; | |
160 | if (!is_swap_pte(pte)) | |
161 | goto out; | |
162 | ||
163 | entry = pte_to_swp_entry(pte); | |
164 | ||
165 | if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old) | |
166 | goto out; | |
167 | ||
168 | /* | |
169 | * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge. | |
170 | * Failure is not an option here: we're now expected to remove every | |
171 | * migration pte, and will cause crashes otherwise. Normally this | |
172 | * is not an issue: mem_cgroup_prepare_migration bumped up the old | |
173 | * page_cgroup count for safety, that's now attached to the new page, | |
174 | * so this charge should just be another incrementation of the count, | |
175 | * to keep in balance with rmap.c's mem_cgroup_uncharging. But if | |
176 | * there's been a force_empty, those reference counts may no longer | |
177 | * be reliable, and this charge can actually fail: oh well, we don't | |
178 | * make the situation any worse by proceeding as if it had succeeded. | |
179 | */ | |
180 | mem_cgroup_charge(new, mm, GFP_ATOMIC); | |
181 | ||
182 | get_page(new); | |
183 | pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); | |
184 | if (is_write_migration_entry(entry)) | |
185 | pte = pte_mkwrite(pte); | |
186 | flush_cache_page(vma, addr, pte_pfn(pte)); | |
187 | set_pte_at(mm, addr, ptep, pte); | |
188 | ||
189 | if (PageAnon(new)) | |
190 | page_add_anon_rmap(new, vma, addr); | |
191 | else | |
192 | page_add_file_rmap(new); | |
193 | ||
194 | /* No need to invalidate - it was non-present before */ | |
195 | update_mmu_cache(vma, addr, pte); | |
196 | ||
197 | out: | |
198 | pte_unmap_unlock(ptep, ptl); | |
199 | } | |
200 | ||
201 | /* | |
202 | * Note that remove_file_migration_ptes will only work on regular mappings, | |
203 | * Nonlinear mappings do not use migration entries. | |
204 | */ | |
205 | static void remove_file_migration_ptes(struct page *old, struct page *new) | |
206 | { | |
207 | struct vm_area_struct *vma; | |
208 | struct address_space *mapping = page_mapping(new); | |
209 | struct prio_tree_iter iter; | |
210 | pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
211 | ||
212 | if (!mapping) | |
213 | return; | |
214 | ||
215 | spin_lock(&mapping->i_mmap_lock); | |
216 | ||
217 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) | |
218 | remove_migration_pte(vma, old, new); | |
219 | ||
220 | spin_unlock(&mapping->i_mmap_lock); | |
221 | } | |
222 | ||
223 | /* | |
224 | * Must hold mmap_sem lock on at least one of the vmas containing | |
225 | * the page so that the anon_vma cannot vanish. | |
226 | */ | |
227 | static void remove_anon_migration_ptes(struct page *old, struct page *new) | |
228 | { | |
229 | struct anon_vma *anon_vma; | |
230 | struct vm_area_struct *vma; | |
231 | unsigned long mapping; | |
232 | ||
233 | mapping = (unsigned long)new->mapping; | |
234 | ||
235 | if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0) | |
236 | return; | |
237 | ||
238 | /* | |
239 | * We hold the mmap_sem lock. So no need to call page_lock_anon_vma. | |
240 | */ | |
241 | anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON); | |
242 | spin_lock(&anon_vma->lock); | |
243 | ||
244 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) | |
245 | remove_migration_pte(vma, old, new); | |
246 | ||
247 | spin_unlock(&anon_vma->lock); | |
248 | } | |
249 | ||
250 | /* | |
251 | * Get rid of all migration entries and replace them by | |
252 | * references to the indicated page. | |
253 | */ | |
254 | static void remove_migration_ptes(struct page *old, struct page *new) | |
255 | { | |
256 | if (PageAnon(new)) | |
257 | remove_anon_migration_ptes(old, new); | |
258 | else | |
259 | remove_file_migration_ptes(old, new); | |
260 | } | |
261 | ||
262 | /* | |
263 | * Something used the pte of a page under migration. We need to | |
264 | * get to the page and wait until migration is finished. | |
265 | * When we return from this function the fault will be retried. | |
266 | * | |
267 | * This function is called from do_swap_page(). | |
268 | */ | |
269 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, | |
270 | unsigned long address) | |
271 | { | |
272 | pte_t *ptep, pte; | |
273 | spinlock_t *ptl; | |
274 | swp_entry_t entry; | |
275 | struct page *page; | |
276 | ||
277 | ptep = pte_offset_map_lock(mm, pmd, address, &ptl); | |
278 | pte = *ptep; | |
279 | if (!is_swap_pte(pte)) | |
280 | goto out; | |
281 | ||
282 | entry = pte_to_swp_entry(pte); | |
283 | if (!is_migration_entry(entry)) | |
284 | goto out; | |
285 | ||
286 | page = migration_entry_to_page(entry); | |
287 | ||
288 | /* | |
289 | * Once radix-tree replacement of page migration started, page_count | |
290 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
291 | * against a page without get_page(). | |
292 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
293 | * will occur again. | |
294 | */ | |
295 | if (!get_page_unless_zero(page)) | |
296 | goto out; | |
297 | pte_unmap_unlock(ptep, ptl); | |
298 | wait_on_page_locked(page); | |
299 | put_page(page); | |
300 | return; | |
301 | out: | |
302 | pte_unmap_unlock(ptep, ptl); | |
303 | } | |
304 | ||
305 | /* | |
306 | * Replace the page in the mapping. | |
307 | * | |
308 | * The number of remaining references must be: | |
309 | * 1 for anonymous pages without a mapping | |
310 | * 2 for pages with a mapping | |
311 | * 3 for pages with a mapping and PagePrivate set. | |
312 | */ | |
313 | static int migrate_page_move_mapping(struct address_space *mapping, | |
314 | struct page *newpage, struct page *page) | |
315 | { | |
316 | int expected_count; | |
317 | void **pslot; | |
318 | ||
319 | if (!mapping) { | |
320 | /* Anonymous page without mapping */ | |
321 | if (page_count(page) != 1) | |
322 | return -EAGAIN; | |
323 | return 0; | |
324 | } | |
325 | ||
326 | write_lock_irq(&mapping->tree_lock); | |
327 | ||
328 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
329 | page_index(page)); | |
330 | ||
331 | expected_count = 2 + !!PagePrivate(page); | |
332 | if (page_count(page) != expected_count || | |
333 | (struct page *)radix_tree_deref_slot(pslot) != page) { | |
334 | write_unlock_irq(&mapping->tree_lock); | |
335 | return -EAGAIN; | |
336 | } | |
337 | ||
338 | if (!page_freeze_refs(page, expected_count)) { | |
339 | write_unlock_irq(&mapping->tree_lock); | |
340 | return -EAGAIN; | |
341 | } | |
342 | ||
343 | /* | |
344 | * Now we know that no one else is looking at the page. | |
345 | */ | |
346 | get_page(newpage); /* add cache reference */ | |
347 | #ifdef CONFIG_SWAP | |
348 | if (PageSwapCache(page)) { | |
349 | SetPageSwapCache(newpage); | |
350 | set_page_private(newpage, page_private(page)); | |
351 | } | |
352 | #endif | |
353 | ||
354 | radix_tree_replace_slot(pslot, newpage); | |
355 | ||
356 | page_unfreeze_refs(page, expected_count); | |
357 | /* | |
358 | * Drop cache reference from old page. | |
359 | * We know this isn't the last reference. | |
360 | */ | |
361 | __put_page(page); | |
362 | ||
363 | /* | |
364 | * If moved to a different zone then also account | |
365 | * the page for that zone. Other VM counters will be | |
366 | * taken care of when we establish references to the | |
367 | * new page and drop references to the old page. | |
368 | * | |
369 | * Note that anonymous pages are accounted for | |
370 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
371 | * are mapped to swap space. | |
372 | */ | |
373 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
374 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
375 | ||
376 | write_unlock_irq(&mapping->tree_lock); | |
377 | if (!PageSwapCache(newpage)) { | |
378 | mem_cgroup_uncharge_cache_page(page); | |
379 | } | |
380 | ||
381 | return 0; | |
382 | } | |
383 | ||
384 | /* | |
385 | * Copy the page to its new location | |
386 | */ | |
387 | static void migrate_page_copy(struct page *newpage, struct page *page) | |
388 | { | |
389 | copy_highpage(newpage, page); | |
390 | ||
391 | if (PageError(page)) | |
392 | SetPageError(newpage); | |
393 | if (PageReferenced(page)) | |
394 | SetPageReferenced(newpage); | |
395 | if (PageUptodate(page)) | |
396 | SetPageUptodate(newpage); | |
397 | if (PageActive(page)) | |
398 | SetPageActive(newpage); | |
399 | if (PageChecked(page)) | |
400 | SetPageChecked(newpage); | |
401 | if (PageMappedToDisk(page)) | |
402 | SetPageMappedToDisk(newpage); | |
403 | ||
404 | if (PageDirty(page)) { | |
405 | clear_page_dirty_for_io(page); | |
406 | /* | |
407 | * Want to mark the page and the radix tree as dirty, and | |
408 | * redo the accounting that clear_page_dirty_for_io undid, | |
409 | * but we can't use set_page_dirty because that function | |
410 | * is actually a signal that all of the page has become dirty. | |
411 | * Wheras only part of our page may be dirty. | |
412 | */ | |
413 | __set_page_dirty_nobuffers(newpage); | |
414 | } | |
415 | ||
416 | #ifdef CONFIG_SWAP | |
417 | ClearPageSwapCache(page); | |
418 | #endif | |
419 | ClearPageActive(page); | |
420 | ClearPagePrivate(page); | |
421 | set_page_private(page, 0); | |
422 | page->mapping = NULL; | |
423 | ||
424 | /* | |
425 | * If any waiters have accumulated on the new page then | |
426 | * wake them up. | |
427 | */ | |
428 | if (PageWriteback(newpage)) | |
429 | end_page_writeback(newpage); | |
430 | } | |
431 | ||
432 | /************************************************************ | |
433 | * Migration functions | |
434 | ***********************************************************/ | |
435 | ||
436 | /* Always fail migration. Used for mappings that are not movable */ | |
437 | int fail_migrate_page(struct address_space *mapping, | |
438 | struct page *newpage, struct page *page) | |
439 | { | |
440 | return -EIO; | |
441 | } | |
442 | EXPORT_SYMBOL(fail_migrate_page); | |
443 | ||
444 | /* | |
445 | * Common logic to directly migrate a single page suitable for | |
446 | * pages that do not use PagePrivate. | |
447 | * | |
448 | * Pages are locked upon entry and exit. | |
449 | */ | |
450 | int migrate_page(struct address_space *mapping, | |
451 | struct page *newpage, struct page *page) | |
452 | { | |
453 | int rc; | |
454 | ||
455 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
456 | ||
457 | rc = migrate_page_move_mapping(mapping, newpage, page); | |
458 | ||
459 | if (rc) | |
460 | return rc; | |
461 | ||
462 | migrate_page_copy(newpage, page); | |
463 | return 0; | |
464 | } | |
465 | EXPORT_SYMBOL(migrate_page); | |
466 | ||
467 | #ifdef CONFIG_BLOCK | |
468 | /* | |
469 | * Migration function for pages with buffers. This function can only be used | |
470 | * if the underlying filesystem guarantees that no other references to "page" | |
471 | * exist. | |
472 | */ | |
473 | int buffer_migrate_page(struct address_space *mapping, | |
474 | struct page *newpage, struct page *page) | |
475 | { | |
476 | struct buffer_head *bh, *head; | |
477 | int rc; | |
478 | ||
479 | if (!page_has_buffers(page)) | |
480 | return migrate_page(mapping, newpage, page); | |
481 | ||
482 | head = page_buffers(page); | |
483 | ||
484 | rc = migrate_page_move_mapping(mapping, newpage, page); | |
485 | ||
486 | if (rc) | |
487 | return rc; | |
488 | ||
489 | bh = head; | |
490 | do { | |
491 | get_bh(bh); | |
492 | lock_buffer(bh); | |
493 | bh = bh->b_this_page; | |
494 | ||
495 | } while (bh != head); | |
496 | ||
497 | ClearPagePrivate(page); | |
498 | set_page_private(newpage, page_private(page)); | |
499 | set_page_private(page, 0); | |
500 | put_page(page); | |
501 | get_page(newpage); | |
502 | ||
503 | bh = head; | |
504 | do { | |
505 | set_bh_page(bh, newpage, bh_offset(bh)); | |
506 | bh = bh->b_this_page; | |
507 | ||
508 | } while (bh != head); | |
509 | ||
510 | SetPagePrivate(newpage); | |
511 | ||
512 | migrate_page_copy(newpage, page); | |
513 | ||
514 | bh = head; | |
515 | do { | |
516 | unlock_buffer(bh); | |
517 | put_bh(bh); | |
518 | bh = bh->b_this_page; | |
519 | ||
520 | } while (bh != head); | |
521 | ||
522 | return 0; | |
523 | } | |
524 | EXPORT_SYMBOL(buffer_migrate_page); | |
525 | #endif | |
526 | ||
527 | /* | |
528 | * Writeback a page to clean the dirty state | |
529 | */ | |
530 | static int writeout(struct address_space *mapping, struct page *page) | |
531 | { | |
532 | struct writeback_control wbc = { | |
533 | .sync_mode = WB_SYNC_NONE, | |
534 | .nr_to_write = 1, | |
535 | .range_start = 0, | |
536 | .range_end = LLONG_MAX, | |
537 | .nonblocking = 1, | |
538 | .for_reclaim = 1 | |
539 | }; | |
540 | int rc; | |
541 | ||
542 | if (!mapping->a_ops->writepage) | |
543 | /* No write method for the address space */ | |
544 | return -EINVAL; | |
545 | ||
546 | if (!clear_page_dirty_for_io(page)) | |
547 | /* Someone else already triggered a write */ | |
548 | return -EAGAIN; | |
549 | ||
550 | /* | |
551 | * A dirty page may imply that the underlying filesystem has | |
552 | * the page on some queue. So the page must be clean for | |
553 | * migration. Writeout may mean we loose the lock and the | |
554 | * page state is no longer what we checked for earlier. | |
555 | * At this point we know that the migration attempt cannot | |
556 | * be successful. | |
557 | */ | |
558 | remove_migration_ptes(page, page); | |
559 | ||
560 | rc = mapping->a_ops->writepage(page, &wbc); | |
561 | if (rc < 0) | |
562 | /* I/O Error writing */ | |
563 | return -EIO; | |
564 | ||
565 | if (rc != AOP_WRITEPAGE_ACTIVATE) | |
566 | /* unlocked. Relock */ | |
567 | lock_page(page); | |
568 | ||
569 | return -EAGAIN; | |
570 | } | |
571 | ||
572 | /* | |
573 | * Default handling if a filesystem does not provide a migration function. | |
574 | */ | |
575 | static int fallback_migrate_page(struct address_space *mapping, | |
576 | struct page *newpage, struct page *page) | |
577 | { | |
578 | if (PageDirty(page)) | |
579 | return writeout(mapping, page); | |
580 | ||
581 | /* | |
582 | * Buffers may be managed in a filesystem specific way. | |
583 | * We must have no buffers or drop them. | |
584 | */ | |
585 | if (PagePrivate(page) && | |
586 | !try_to_release_page(page, GFP_KERNEL)) | |
587 | return -EAGAIN; | |
588 | ||
589 | return migrate_page(mapping, newpage, page); | |
590 | } | |
591 | ||
592 | /* | |
593 | * Move a page to a newly allocated page | |
594 | * The page is locked and all ptes have been successfully removed. | |
595 | * | |
596 | * The new page will have replaced the old page if this function | |
597 | * is successful. | |
598 | */ | |
599 | static int move_to_new_page(struct page *newpage, struct page *page) | |
600 | { | |
601 | struct address_space *mapping; | |
602 | int rc; | |
603 | ||
604 | /* | |
605 | * Block others from accessing the page when we get around to | |
606 | * establishing additional references. We are the only one | |
607 | * holding a reference to the new page at this point. | |
608 | */ | |
609 | if (TestSetPageLocked(newpage)) | |
610 | BUG(); | |
611 | ||
612 | /* Prepare mapping for the new page.*/ | |
613 | newpage->index = page->index; | |
614 | newpage->mapping = page->mapping; | |
615 | ||
616 | mapping = page_mapping(page); | |
617 | if (!mapping) | |
618 | rc = migrate_page(mapping, newpage, page); | |
619 | else if (mapping->a_ops->migratepage) | |
620 | /* | |
621 | * Most pages have a mapping and most filesystems | |
622 | * should provide a migration function. Anonymous | |
623 | * pages are part of swap space which also has its | |
624 | * own migration function. This is the most common | |
625 | * path for page migration. | |
626 | */ | |
627 | rc = mapping->a_ops->migratepage(mapping, | |
628 | newpage, page); | |
629 | else | |
630 | rc = fallback_migrate_page(mapping, newpage, page); | |
631 | ||
632 | if (!rc) { | |
633 | remove_migration_ptes(page, newpage); | |
634 | } else | |
635 | newpage->mapping = NULL; | |
636 | ||
637 | unlock_page(newpage); | |
638 | ||
639 | return rc; | |
640 | } | |
641 | ||
642 | /* | |
643 | * Obtain the lock on page, remove all ptes and migrate the page | |
644 | * to the newly allocated page in newpage. | |
645 | */ | |
646 | static int unmap_and_move(new_page_t get_new_page, unsigned long private, | |
647 | struct page *page, int force) | |
648 | { | |
649 | int rc = 0; | |
650 | int *result = NULL; | |
651 | struct page *newpage = get_new_page(page, private, &result); | |
652 | int rcu_locked = 0; | |
653 | int charge = 0; | |
654 | ||
655 | if (!newpage) | |
656 | return -ENOMEM; | |
657 | ||
658 | if (page_count(page) == 1) | |
659 | /* page was freed from under us. So we are done. */ | |
660 | goto move_newpage; | |
661 | ||
662 | charge = mem_cgroup_prepare_migration(page, newpage); | |
663 | if (charge == -ENOMEM) { | |
664 | rc = -ENOMEM; | |
665 | goto move_newpage; | |
666 | } | |
667 | /* prepare cgroup just returns 0 or -ENOMEM */ | |
668 | BUG_ON(charge); | |
669 | ||
670 | rc = -EAGAIN; | |
671 | if (TestSetPageLocked(page)) { | |
672 | if (!force) | |
673 | goto move_newpage; | |
674 | lock_page(page); | |
675 | } | |
676 | ||
677 | if (PageWriteback(page)) { | |
678 | if (!force) | |
679 | goto unlock; | |
680 | wait_on_page_writeback(page); | |
681 | } | |
682 | /* | |
683 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, | |
684 | * we cannot notice that anon_vma is freed while we migrates a page. | |
685 | * This rcu_read_lock() delays freeing anon_vma pointer until the end | |
686 | * of migration. File cache pages are no problem because of page_lock() | |
687 | * File Caches may use write_page() or lock_page() in migration, then, | |
688 | * just care Anon page here. | |
689 | */ | |
690 | if (PageAnon(page)) { | |
691 | rcu_read_lock(); | |
692 | rcu_locked = 1; | |
693 | } | |
694 | ||
695 | /* | |
696 | * Corner case handling: | |
697 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
698 | * and treated as swapcache but it has no rmap yet. | |
699 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
700 | * trigger a BUG. So handle it here. | |
701 | * 2. An orphaned page (see truncate_complete_page) might have | |
702 | * fs-private metadata. The page can be picked up due to memory | |
703 | * offlining. Everywhere else except page reclaim, the page is | |
704 | * invisible to the vm, so the page can not be migrated. So try to | |
705 | * free the metadata, so the page can be freed. | |
706 | */ | |
707 | if (!page->mapping) { | |
708 | if (!PageAnon(page) && PagePrivate(page)) { | |
709 | /* | |
710 | * Go direct to try_to_free_buffers() here because | |
711 | * a) that's what try_to_release_page() would do anyway | |
712 | * b) we may be under rcu_read_lock() here, so we can't | |
713 | * use GFP_KERNEL which is what try_to_release_page() | |
714 | * needs to be effective. | |
715 | */ | |
716 | try_to_free_buffers(page); | |
717 | } | |
718 | goto rcu_unlock; | |
719 | } | |
720 | ||
721 | /* Establish migration ptes or remove ptes */ | |
722 | try_to_unmap(page, 1); | |
723 | ||
724 | if (!page_mapped(page)) | |
725 | rc = move_to_new_page(newpage, page); | |
726 | ||
727 | if (rc) | |
728 | remove_migration_ptes(page, page); | |
729 | rcu_unlock: | |
730 | if (rcu_locked) | |
731 | rcu_read_unlock(); | |
732 | ||
733 | unlock: | |
734 | ||
735 | unlock_page(page); | |
736 | ||
737 | if (rc != -EAGAIN) { | |
738 | /* | |
739 | * A page that has been migrated has all references | |
740 | * removed and will be freed. A page that has not been | |
741 | * migrated will have kepts its references and be | |
742 | * restored. | |
743 | */ | |
744 | list_del(&page->lru); | |
745 | move_to_lru(page); | |
746 | } | |
747 | ||
748 | move_newpage: | |
749 | if (!charge) | |
750 | mem_cgroup_end_migration(newpage); | |
751 | /* | |
752 | * Move the new page to the LRU. If migration was not successful | |
753 | * then this will free the page. | |
754 | */ | |
755 | move_to_lru(newpage); | |
756 | if (result) { | |
757 | if (rc) | |
758 | *result = rc; | |
759 | else | |
760 | *result = page_to_nid(newpage); | |
761 | } | |
762 | return rc; | |
763 | } | |
764 | ||
765 | /* | |
766 | * migrate_pages | |
767 | * | |
768 | * The function takes one list of pages to migrate and a function | |
769 | * that determines from the page to be migrated and the private data | |
770 | * the target of the move and allocates the page. | |
771 | * | |
772 | * The function returns after 10 attempts or if no pages | |
773 | * are movable anymore because to has become empty | |
774 | * or no retryable pages exist anymore. All pages will be | |
775 | * returned to the LRU or freed. | |
776 | * | |
777 | * Return: Number of pages not migrated or error code. | |
778 | */ | |
779 | int migrate_pages(struct list_head *from, | |
780 | new_page_t get_new_page, unsigned long private) | |
781 | { | |
782 | int retry = 1; | |
783 | int nr_failed = 0; | |
784 | int pass = 0; | |
785 | struct page *page; | |
786 | struct page *page2; | |
787 | int swapwrite = current->flags & PF_SWAPWRITE; | |
788 | int rc; | |
789 | ||
790 | if (!swapwrite) | |
791 | current->flags |= PF_SWAPWRITE; | |
792 | ||
793 | for(pass = 0; pass < 10 && retry; pass++) { | |
794 | retry = 0; | |
795 | ||
796 | list_for_each_entry_safe(page, page2, from, lru) { | |
797 | cond_resched(); | |
798 | ||
799 | rc = unmap_and_move(get_new_page, private, | |
800 | page, pass > 2); | |
801 | ||
802 | switch(rc) { | |
803 | case -ENOMEM: | |
804 | goto out; | |
805 | case -EAGAIN: | |
806 | retry++; | |
807 | break; | |
808 | case 0: | |
809 | break; | |
810 | default: | |
811 | /* Permanent failure */ | |
812 | nr_failed++; | |
813 | break; | |
814 | } | |
815 | } | |
816 | } | |
817 | rc = 0; | |
818 | out: | |
819 | if (!swapwrite) | |
820 | current->flags &= ~PF_SWAPWRITE; | |
821 | ||
822 | putback_lru_pages(from); | |
823 | ||
824 | if (rc) | |
825 | return rc; | |
826 | ||
827 | return nr_failed + retry; | |
828 | } | |
829 | ||
830 | #ifdef CONFIG_NUMA | |
831 | /* | |
832 | * Move a list of individual pages | |
833 | */ | |
834 | struct page_to_node { | |
835 | unsigned long addr; | |
836 | struct page *page; | |
837 | int node; | |
838 | int status; | |
839 | }; | |
840 | ||
841 | static struct page *new_page_node(struct page *p, unsigned long private, | |
842 | int **result) | |
843 | { | |
844 | struct page_to_node *pm = (struct page_to_node *)private; | |
845 | ||
846 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
847 | pm++; | |
848 | ||
849 | if (pm->node == MAX_NUMNODES) | |
850 | return NULL; | |
851 | ||
852 | *result = &pm->status; | |
853 | ||
854 | return alloc_pages_node(pm->node, | |
855 | GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0); | |
856 | } | |
857 | ||
858 | /* | |
859 | * Move a set of pages as indicated in the pm array. The addr | |
860 | * field must be set to the virtual address of the page to be moved | |
861 | * and the node number must contain a valid target node. | |
862 | */ | |
863 | static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm, | |
864 | int migrate_all) | |
865 | { | |
866 | int err; | |
867 | struct page_to_node *pp; | |
868 | LIST_HEAD(pagelist); | |
869 | ||
870 | down_read(&mm->mmap_sem); | |
871 | ||
872 | /* | |
873 | * Build a list of pages to migrate | |
874 | */ | |
875 | migrate_prep(); | |
876 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { | |
877 | struct vm_area_struct *vma; | |
878 | struct page *page; | |
879 | ||
880 | /* | |
881 | * A valid page pointer that will not match any of the | |
882 | * pages that will be moved. | |
883 | */ | |
884 | pp->page = ZERO_PAGE(0); | |
885 | ||
886 | err = -EFAULT; | |
887 | vma = find_vma(mm, pp->addr); | |
888 | if (!vma || !vma_migratable(vma)) | |
889 | goto set_status; | |
890 | ||
891 | page = follow_page(vma, pp->addr, FOLL_GET); | |
892 | ||
893 | err = PTR_ERR(page); | |
894 | if (IS_ERR(page)) | |
895 | goto set_status; | |
896 | ||
897 | err = -ENOENT; | |
898 | if (!page) | |
899 | goto set_status; | |
900 | ||
901 | if (PageReserved(page)) /* Check for zero page */ | |
902 | goto put_and_set; | |
903 | ||
904 | pp->page = page; | |
905 | err = page_to_nid(page); | |
906 | ||
907 | if (err == pp->node) | |
908 | /* | |
909 | * Node already in the right place | |
910 | */ | |
911 | goto put_and_set; | |
912 | ||
913 | err = -EACCES; | |
914 | if (page_mapcount(page) > 1 && | |
915 | !migrate_all) | |
916 | goto put_and_set; | |
917 | ||
918 | err = isolate_lru_page(page, &pagelist); | |
919 | put_and_set: | |
920 | /* | |
921 | * Either remove the duplicate refcount from | |
922 | * isolate_lru_page() or drop the page ref if it was | |
923 | * not isolated. | |
924 | */ | |
925 | put_page(page); | |
926 | set_status: | |
927 | pp->status = err; | |
928 | } | |
929 | ||
930 | if (!list_empty(&pagelist)) | |
931 | err = migrate_pages(&pagelist, new_page_node, | |
932 | (unsigned long)pm); | |
933 | else | |
934 | err = -ENOENT; | |
935 | ||
936 | up_read(&mm->mmap_sem); | |
937 | return err; | |
938 | } | |
939 | ||
940 | /* | |
941 | * Determine the nodes of a list of pages. The addr in the pm array | |
942 | * must have been set to the virtual address of which we want to determine | |
943 | * the node number. | |
944 | */ | |
945 | static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm) | |
946 | { | |
947 | down_read(&mm->mmap_sem); | |
948 | ||
949 | for ( ; pm->node != MAX_NUMNODES; pm++) { | |
950 | struct vm_area_struct *vma; | |
951 | struct page *page; | |
952 | int err; | |
953 | ||
954 | err = -EFAULT; | |
955 | vma = find_vma(mm, pm->addr); | |
956 | if (!vma) | |
957 | goto set_status; | |
958 | ||
959 | page = follow_page(vma, pm->addr, 0); | |
960 | ||
961 | err = PTR_ERR(page); | |
962 | if (IS_ERR(page)) | |
963 | goto set_status; | |
964 | ||
965 | err = -ENOENT; | |
966 | /* Use PageReserved to check for zero page */ | |
967 | if (!page || PageReserved(page)) | |
968 | goto set_status; | |
969 | ||
970 | err = page_to_nid(page); | |
971 | set_status: | |
972 | pm->status = err; | |
973 | } | |
974 | ||
975 | up_read(&mm->mmap_sem); | |
976 | return 0; | |
977 | } | |
978 | ||
979 | /* | |
980 | * Move a list of pages in the address space of the currently executing | |
981 | * process. | |
982 | */ | |
983 | asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages, | |
984 | const void __user * __user *pages, | |
985 | const int __user *nodes, | |
986 | int __user *status, int flags) | |
987 | { | |
988 | int err = 0; | |
989 | int i; | |
990 | struct task_struct *task; | |
991 | nodemask_t task_nodes; | |
992 | struct mm_struct *mm; | |
993 | struct page_to_node *pm = NULL; | |
994 | ||
995 | /* Check flags */ | |
996 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
997 | return -EINVAL; | |
998 | ||
999 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1000 | return -EPERM; | |
1001 | ||
1002 | /* Find the mm_struct */ | |
1003 | read_lock(&tasklist_lock); | |
1004 | task = pid ? find_task_by_vpid(pid) : current; | |
1005 | if (!task) { | |
1006 | read_unlock(&tasklist_lock); | |
1007 | return -ESRCH; | |
1008 | } | |
1009 | mm = get_task_mm(task); | |
1010 | read_unlock(&tasklist_lock); | |
1011 | ||
1012 | if (!mm) | |
1013 | return -EINVAL; | |
1014 | ||
1015 | /* | |
1016 | * Check if this process has the right to modify the specified | |
1017 | * process. The right exists if the process has administrative | |
1018 | * capabilities, superuser privileges or the same | |
1019 | * userid as the target process. | |
1020 | */ | |
1021 | if ((current->euid != task->suid) && (current->euid != task->uid) && | |
1022 | (current->uid != task->suid) && (current->uid != task->uid) && | |
1023 | !capable(CAP_SYS_NICE)) { | |
1024 | err = -EPERM; | |
1025 | goto out2; | |
1026 | } | |
1027 | ||
1028 | err = security_task_movememory(task); | |
1029 | if (err) | |
1030 | goto out2; | |
1031 | ||
1032 | ||
1033 | task_nodes = cpuset_mems_allowed(task); | |
1034 | ||
1035 | /* Limit nr_pages so that the multiplication may not overflow */ | |
1036 | if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) { | |
1037 | err = -E2BIG; | |
1038 | goto out2; | |
1039 | } | |
1040 | ||
1041 | pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node)); | |
1042 | if (!pm) { | |
1043 | err = -ENOMEM; | |
1044 | goto out2; | |
1045 | } | |
1046 | ||
1047 | /* | |
1048 | * Get parameters from user space and initialize the pm | |
1049 | * array. Return various errors if the user did something wrong. | |
1050 | */ | |
1051 | for (i = 0; i < nr_pages; i++) { | |
1052 | const void __user *p; | |
1053 | ||
1054 | err = -EFAULT; | |
1055 | if (get_user(p, pages + i)) | |
1056 | goto out; | |
1057 | ||
1058 | pm[i].addr = (unsigned long)p; | |
1059 | if (nodes) { | |
1060 | int node; | |
1061 | ||
1062 | if (get_user(node, nodes + i)) | |
1063 | goto out; | |
1064 | ||
1065 | err = -ENODEV; | |
1066 | if (!node_state(node, N_HIGH_MEMORY)) | |
1067 | goto out; | |
1068 | ||
1069 | err = -EACCES; | |
1070 | if (!node_isset(node, task_nodes)) | |
1071 | goto out; | |
1072 | ||
1073 | pm[i].node = node; | |
1074 | } else | |
1075 | pm[i].node = 0; /* anything to not match MAX_NUMNODES */ | |
1076 | } | |
1077 | /* End marker */ | |
1078 | pm[nr_pages].node = MAX_NUMNODES; | |
1079 | ||
1080 | if (nodes) | |
1081 | err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL); | |
1082 | else | |
1083 | err = do_pages_stat(mm, pm); | |
1084 | ||
1085 | if (err >= 0) | |
1086 | /* Return status information */ | |
1087 | for (i = 0; i < nr_pages; i++) | |
1088 | if (put_user(pm[i].status, status + i)) | |
1089 | err = -EFAULT; | |
1090 | ||
1091 | out: | |
1092 | vfree(pm); | |
1093 | out2: | |
1094 | mmput(mm); | |
1095 | return err; | |
1096 | } | |
1097 | ||
1098 | /* | |
1099 | * Call migration functions in the vma_ops that may prepare | |
1100 | * memory in a vm for migration. migration functions may perform | |
1101 | * the migration for vmas that do not have an underlying page struct. | |
1102 | */ | |
1103 | int migrate_vmas(struct mm_struct *mm, const nodemask_t *to, | |
1104 | const nodemask_t *from, unsigned long flags) | |
1105 | { | |
1106 | struct vm_area_struct *vma; | |
1107 | int err = 0; | |
1108 | ||
1109 | for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) { | |
1110 | if (vma->vm_ops && vma->vm_ops->migrate) { | |
1111 | err = vma->vm_ops->migrate(vma, to, from, flags); | |
1112 | if (err) | |
1113 | break; | |
1114 | } | |
1115 | } | |
1116 | return err; | |
1117 | } | |
1118 | #endif |