Commit | Line | Data |
---|---|---|
b20a3503 CL |
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> | |
cde53535 | 12 | * Christoph Lameter |
b20a3503 CL |
13 | */ |
14 | ||
15 | #include <linux/migrate.h> | |
16 | #include <linux/module.h> | |
17 | #include <linux/swap.h> | |
0697212a | 18 | #include <linux/swapops.h> |
b20a3503 | 19 | #include <linux/pagemap.h> |
e23ca00b | 20 | #include <linux/buffer_head.h> |
b20a3503 | 21 | #include <linux/mm_inline.h> |
b488893a | 22 | #include <linux/nsproxy.h> |
b20a3503 CL |
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> | |
04e62a29 | 28 | #include <linux/writeback.h> |
742755a1 CL |
29 | #include <linux/mempolicy.h> |
30 | #include <linux/vmalloc.h> | |
86c3a764 | 31 | #include <linux/security.h> |
8a9f3ccd | 32 | #include <linux/memcontrol.h> |
4f5ca265 | 33 | #include <linux/syscalls.h> |
b20a3503 CL |
34 | |
35 | #include "internal.h" | |
36 | ||
b20a3503 CL |
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); | |
3dd9fe8c | 55 | if (PageLRU(page) && get_page_unless_zero(page)) { |
b20a3503 | 56 | ret = 0; |
b20a3503 CL |
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 | /* | |
742755a1 CL |
70 | * migrate_prep() needs to be called before we start compiling a list of pages |
71 | * to be migrated using isolate_lru_page(). | |
b20a3503 CL |
72 | */ |
73 | int migrate_prep(void) | |
74 | { | |
b20a3503 CL |
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 | { | |
b20a3503 CL |
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) { | |
e24f0b8f | 113 | list_del(&page->lru); |
b20a3503 CL |
114 | move_to_lru(page); |
115 | count++; | |
116 | } | |
117 | return count; | |
118 | } | |
119 | ||
0697212a CL |
120 | /* |
121 | * Restore a potential migration pte to a working pte entry | |
122 | */ | |
04e62a29 | 123 | static void remove_migration_pte(struct vm_area_struct *vma, |
0697212a CL |
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; | |
04e62a29 CL |
133 | unsigned long addr = page_address_in_vma(new, vma); |
134 | ||
135 | if (addr == -EFAULT) | |
136 | return; | |
0697212a CL |
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 | ||
98837c7f HD |
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 | ||
0697212a CL |
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); | |
97ee0524 | 186 | flush_cache_page(vma, addr, pte_pfn(pte)); |
0697212a | 187 | set_pte_at(mm, addr, ptep, pte); |
04e62a29 CL |
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); | |
04e62a29 | 196 | |
0697212a CL |
197 | out: |
198 | pte_unmap_unlock(ptep, ptl); | |
199 | } | |
200 | ||
201 | /* | |
04e62a29 CL |
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 | /* | |
0697212a CL |
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 | */ | |
04e62a29 | 227 | static void remove_anon_migration_ptes(struct page *old, struct page *new) |
0697212a CL |
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) | |
04e62a29 | 245 | remove_migration_pte(vma, old, new); |
0697212a CL |
246 | |
247 | spin_unlock(&anon_vma->lock); | |
248 | } | |
249 | ||
04e62a29 CL |
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 | ||
0697212a CL |
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 | get_page(page); | |
289 | pte_unmap_unlock(ptep, ptl); | |
290 | wait_on_page_locked(page); | |
291 | put_page(page); | |
292 | return; | |
293 | out: | |
294 | pte_unmap_unlock(ptep, ptl); | |
295 | } | |
296 | ||
b20a3503 | 297 | /* |
c3fcf8a5 | 298 | * Replace the page in the mapping. |
5b5c7120 CL |
299 | * |
300 | * The number of remaining references must be: | |
301 | * 1 for anonymous pages without a mapping | |
302 | * 2 for pages with a mapping | |
303 | * 3 for pages with a mapping and PagePrivate set. | |
b20a3503 | 304 | */ |
2d1db3b1 CL |
305 | static int migrate_page_move_mapping(struct address_space *mapping, |
306 | struct page *newpage, struct page *page) | |
b20a3503 | 307 | { |
7cf9c2c7 | 308 | void **pslot; |
b20a3503 | 309 | |
6c5240ae | 310 | if (!mapping) { |
0e8c7d0f | 311 | /* Anonymous page without mapping */ |
6c5240ae CL |
312 | if (page_count(page) != 1) |
313 | return -EAGAIN; | |
314 | return 0; | |
315 | } | |
316 | ||
b20a3503 CL |
317 | write_lock_irq(&mapping->tree_lock); |
318 | ||
7cf9c2c7 NP |
319 | pslot = radix_tree_lookup_slot(&mapping->page_tree, |
320 | page_index(page)); | |
b20a3503 | 321 | |
6c5240ae | 322 | if (page_count(page) != 2 + !!PagePrivate(page) || |
7cf9c2c7 | 323 | (struct page *)radix_tree_deref_slot(pslot) != page) { |
b20a3503 | 324 | write_unlock_irq(&mapping->tree_lock); |
e23ca00b | 325 | return -EAGAIN; |
b20a3503 CL |
326 | } |
327 | ||
328 | /* | |
329 | * Now we know that no one else is looking at the page. | |
b20a3503 | 330 | */ |
7cf9c2c7 | 331 | get_page(newpage); /* add cache reference */ |
6c5240ae | 332 | #ifdef CONFIG_SWAP |
b20a3503 CL |
333 | if (PageSwapCache(page)) { |
334 | SetPageSwapCache(newpage); | |
335 | set_page_private(newpage, page_private(page)); | |
336 | } | |
6c5240ae | 337 | #endif |
b20a3503 | 338 | |
7cf9c2c7 NP |
339 | radix_tree_replace_slot(pslot, newpage); |
340 | ||
341 | /* | |
342 | * Drop cache reference from old page. | |
343 | * We know this isn't the last reference. | |
344 | */ | |
b20a3503 | 345 | __put_page(page); |
7cf9c2c7 | 346 | |
0e8c7d0f CL |
347 | /* |
348 | * If moved to a different zone then also account | |
349 | * the page for that zone. Other VM counters will be | |
350 | * taken care of when we establish references to the | |
351 | * new page and drop references to the old page. | |
352 | * | |
353 | * Note that anonymous pages are accounted for | |
354 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
355 | * are mapped to swap space. | |
356 | */ | |
357 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
358 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
359 | ||
b20a3503 CL |
360 | write_unlock_irq(&mapping->tree_lock); |
361 | ||
362 | return 0; | |
363 | } | |
b20a3503 CL |
364 | |
365 | /* | |
366 | * Copy the page to its new location | |
367 | */ | |
e7340f73 | 368 | static void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 CL |
369 | { |
370 | copy_highpage(newpage, page); | |
371 | ||
372 | if (PageError(page)) | |
373 | SetPageError(newpage); | |
374 | if (PageReferenced(page)) | |
375 | SetPageReferenced(newpage); | |
376 | if (PageUptodate(page)) | |
377 | SetPageUptodate(newpage); | |
378 | if (PageActive(page)) | |
379 | SetPageActive(newpage); | |
380 | if (PageChecked(page)) | |
381 | SetPageChecked(newpage); | |
382 | if (PageMappedToDisk(page)) | |
383 | SetPageMappedToDisk(newpage); | |
384 | ||
385 | if (PageDirty(page)) { | |
386 | clear_page_dirty_for_io(page); | |
3a902c5f NP |
387 | /* |
388 | * Want to mark the page and the radix tree as dirty, and | |
389 | * redo the accounting that clear_page_dirty_for_io undid, | |
390 | * but we can't use set_page_dirty because that function | |
391 | * is actually a signal that all of the page has become dirty. | |
392 | * Wheras only part of our page may be dirty. | |
393 | */ | |
394 | __set_page_dirty_nobuffers(newpage); | |
b20a3503 CL |
395 | } |
396 | ||
6c5240ae | 397 | #ifdef CONFIG_SWAP |
b20a3503 | 398 | ClearPageSwapCache(page); |
6c5240ae | 399 | #endif |
b20a3503 CL |
400 | ClearPageActive(page); |
401 | ClearPagePrivate(page); | |
402 | set_page_private(page, 0); | |
403 | page->mapping = NULL; | |
404 | ||
405 | /* | |
406 | * If any waiters have accumulated on the new page then | |
407 | * wake them up. | |
408 | */ | |
409 | if (PageWriteback(newpage)) | |
410 | end_page_writeback(newpage); | |
411 | } | |
b20a3503 | 412 | |
1d8b85cc CL |
413 | /************************************************************ |
414 | * Migration functions | |
415 | ***********************************************************/ | |
416 | ||
417 | /* Always fail migration. Used for mappings that are not movable */ | |
2d1db3b1 CL |
418 | int fail_migrate_page(struct address_space *mapping, |
419 | struct page *newpage, struct page *page) | |
1d8b85cc CL |
420 | { |
421 | return -EIO; | |
422 | } | |
423 | EXPORT_SYMBOL(fail_migrate_page); | |
424 | ||
b20a3503 CL |
425 | /* |
426 | * Common logic to directly migrate a single page suitable for | |
427 | * pages that do not use PagePrivate. | |
428 | * | |
429 | * Pages are locked upon entry and exit. | |
430 | */ | |
2d1db3b1 CL |
431 | int migrate_page(struct address_space *mapping, |
432 | struct page *newpage, struct page *page) | |
b20a3503 CL |
433 | { |
434 | int rc; | |
435 | ||
436 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
437 | ||
2d1db3b1 | 438 | rc = migrate_page_move_mapping(mapping, newpage, page); |
b20a3503 CL |
439 | |
440 | if (rc) | |
441 | return rc; | |
442 | ||
443 | migrate_page_copy(newpage, page); | |
b20a3503 CL |
444 | return 0; |
445 | } | |
446 | EXPORT_SYMBOL(migrate_page); | |
447 | ||
9361401e | 448 | #ifdef CONFIG_BLOCK |
1d8b85cc CL |
449 | /* |
450 | * Migration function for pages with buffers. This function can only be used | |
451 | * if the underlying filesystem guarantees that no other references to "page" | |
452 | * exist. | |
453 | */ | |
2d1db3b1 CL |
454 | int buffer_migrate_page(struct address_space *mapping, |
455 | struct page *newpage, struct page *page) | |
1d8b85cc | 456 | { |
1d8b85cc CL |
457 | struct buffer_head *bh, *head; |
458 | int rc; | |
459 | ||
1d8b85cc | 460 | if (!page_has_buffers(page)) |
2d1db3b1 | 461 | return migrate_page(mapping, newpage, page); |
1d8b85cc CL |
462 | |
463 | head = page_buffers(page); | |
464 | ||
2d1db3b1 | 465 | rc = migrate_page_move_mapping(mapping, newpage, page); |
1d8b85cc CL |
466 | |
467 | if (rc) | |
468 | return rc; | |
469 | ||
470 | bh = head; | |
471 | do { | |
472 | get_bh(bh); | |
473 | lock_buffer(bh); | |
474 | bh = bh->b_this_page; | |
475 | ||
476 | } while (bh != head); | |
477 | ||
478 | ClearPagePrivate(page); | |
479 | set_page_private(newpage, page_private(page)); | |
480 | set_page_private(page, 0); | |
481 | put_page(page); | |
482 | get_page(newpage); | |
483 | ||
484 | bh = head; | |
485 | do { | |
486 | set_bh_page(bh, newpage, bh_offset(bh)); | |
487 | bh = bh->b_this_page; | |
488 | ||
489 | } while (bh != head); | |
490 | ||
491 | SetPagePrivate(newpage); | |
492 | ||
493 | migrate_page_copy(newpage, page); | |
494 | ||
495 | bh = head; | |
496 | do { | |
497 | unlock_buffer(bh); | |
498 | put_bh(bh); | |
499 | bh = bh->b_this_page; | |
500 | ||
501 | } while (bh != head); | |
502 | ||
503 | return 0; | |
504 | } | |
505 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 506 | #endif |
1d8b85cc | 507 | |
04e62a29 CL |
508 | /* |
509 | * Writeback a page to clean the dirty state | |
510 | */ | |
511 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 512 | { |
04e62a29 CL |
513 | struct writeback_control wbc = { |
514 | .sync_mode = WB_SYNC_NONE, | |
515 | .nr_to_write = 1, | |
516 | .range_start = 0, | |
517 | .range_end = LLONG_MAX, | |
518 | .nonblocking = 1, | |
519 | .for_reclaim = 1 | |
520 | }; | |
521 | int rc; | |
522 | ||
523 | if (!mapping->a_ops->writepage) | |
524 | /* No write method for the address space */ | |
525 | return -EINVAL; | |
526 | ||
527 | if (!clear_page_dirty_for_io(page)) | |
528 | /* Someone else already triggered a write */ | |
529 | return -EAGAIN; | |
530 | ||
8351a6e4 | 531 | /* |
04e62a29 CL |
532 | * A dirty page may imply that the underlying filesystem has |
533 | * the page on some queue. So the page must be clean for | |
534 | * migration. Writeout may mean we loose the lock and the | |
535 | * page state is no longer what we checked for earlier. | |
536 | * At this point we know that the migration attempt cannot | |
537 | * be successful. | |
8351a6e4 | 538 | */ |
04e62a29 | 539 | remove_migration_ptes(page, page); |
8351a6e4 | 540 | |
04e62a29 CL |
541 | rc = mapping->a_ops->writepage(page, &wbc); |
542 | if (rc < 0) | |
543 | /* I/O Error writing */ | |
544 | return -EIO; | |
8351a6e4 | 545 | |
04e62a29 CL |
546 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
547 | /* unlocked. Relock */ | |
548 | lock_page(page); | |
549 | ||
550 | return -EAGAIN; | |
551 | } | |
552 | ||
553 | /* | |
554 | * Default handling if a filesystem does not provide a migration function. | |
555 | */ | |
556 | static int fallback_migrate_page(struct address_space *mapping, | |
557 | struct page *newpage, struct page *page) | |
558 | { | |
559 | if (PageDirty(page)) | |
560 | return writeout(mapping, page); | |
8351a6e4 CL |
561 | |
562 | /* | |
563 | * Buffers may be managed in a filesystem specific way. | |
564 | * We must have no buffers or drop them. | |
565 | */ | |
b398f6bf | 566 | if (PagePrivate(page) && |
8351a6e4 CL |
567 | !try_to_release_page(page, GFP_KERNEL)) |
568 | return -EAGAIN; | |
569 | ||
570 | return migrate_page(mapping, newpage, page); | |
571 | } | |
572 | ||
e24f0b8f CL |
573 | /* |
574 | * Move a page to a newly allocated page | |
575 | * The page is locked and all ptes have been successfully removed. | |
576 | * | |
577 | * The new page will have replaced the old page if this function | |
578 | * is successful. | |
579 | */ | |
580 | static int move_to_new_page(struct page *newpage, struct page *page) | |
581 | { | |
582 | struct address_space *mapping; | |
583 | int rc; | |
584 | ||
585 | /* | |
586 | * Block others from accessing the page when we get around to | |
587 | * establishing additional references. We are the only one | |
588 | * holding a reference to the new page at this point. | |
589 | */ | |
590 | if (TestSetPageLocked(newpage)) | |
591 | BUG(); | |
592 | ||
593 | /* Prepare mapping for the new page.*/ | |
594 | newpage->index = page->index; | |
595 | newpage->mapping = page->mapping; | |
596 | ||
597 | mapping = page_mapping(page); | |
598 | if (!mapping) | |
599 | rc = migrate_page(mapping, newpage, page); | |
600 | else if (mapping->a_ops->migratepage) | |
601 | /* | |
602 | * Most pages have a mapping and most filesystems | |
603 | * should provide a migration function. Anonymous | |
604 | * pages are part of swap space which also has its | |
605 | * own migration function. This is the most common | |
606 | * path for page migration. | |
607 | */ | |
608 | rc = mapping->a_ops->migratepage(mapping, | |
609 | newpage, page); | |
610 | else | |
611 | rc = fallback_migrate_page(mapping, newpage, page); | |
612 | ||
ae41be37 KH |
613 | if (!rc) { |
614 | mem_cgroup_page_migration(page, newpage); | |
e24f0b8f | 615 | remove_migration_ptes(page, newpage); |
ae41be37 | 616 | } else |
e24f0b8f CL |
617 | newpage->mapping = NULL; |
618 | ||
619 | unlock_page(newpage); | |
620 | ||
621 | return rc; | |
622 | } | |
623 | ||
624 | /* | |
625 | * Obtain the lock on page, remove all ptes and migrate the page | |
626 | * to the newly allocated page in newpage. | |
627 | */ | |
95a402c3 CL |
628 | static int unmap_and_move(new_page_t get_new_page, unsigned long private, |
629 | struct page *page, int force) | |
e24f0b8f CL |
630 | { |
631 | int rc = 0; | |
742755a1 CL |
632 | int *result = NULL; |
633 | struct page *newpage = get_new_page(page, private, &result); | |
989f89c5 | 634 | int rcu_locked = 0; |
ae41be37 | 635 | int charge = 0; |
95a402c3 CL |
636 | |
637 | if (!newpage) | |
638 | return -ENOMEM; | |
e24f0b8f CL |
639 | |
640 | if (page_count(page) == 1) | |
641 | /* page was freed from under us. So we are done. */ | |
95a402c3 | 642 | goto move_newpage; |
e24f0b8f CL |
643 | |
644 | rc = -EAGAIN; | |
645 | if (TestSetPageLocked(page)) { | |
646 | if (!force) | |
95a402c3 | 647 | goto move_newpage; |
e24f0b8f CL |
648 | lock_page(page); |
649 | } | |
650 | ||
651 | if (PageWriteback(page)) { | |
652 | if (!force) | |
653 | goto unlock; | |
654 | wait_on_page_writeback(page); | |
655 | } | |
e24f0b8f | 656 | /* |
dc386d4d KH |
657 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
658 | * we cannot notice that anon_vma is freed while we migrates a page. | |
659 | * This rcu_read_lock() delays freeing anon_vma pointer until the end | |
660 | * of migration. File cache pages are no problem because of page_lock() | |
989f89c5 KH |
661 | * File Caches may use write_page() or lock_page() in migration, then, |
662 | * just care Anon page here. | |
dc386d4d | 663 | */ |
989f89c5 KH |
664 | if (PageAnon(page)) { |
665 | rcu_read_lock(); | |
666 | rcu_locked = 1; | |
667 | } | |
62e1c553 | 668 | |
dc386d4d | 669 | /* |
62e1c553 SL |
670 | * Corner case handling: |
671 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
672 | * and treated as swapcache but it has no rmap yet. | |
673 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
674 | * trigger a BUG. So handle it here. | |
675 | * 2. An orphaned page (see truncate_complete_page) might have | |
676 | * fs-private metadata. The page can be picked up due to memory | |
677 | * offlining. Everywhere else except page reclaim, the page is | |
678 | * invisible to the vm, so the page can not be migrated. So try to | |
679 | * free the metadata, so the page can be freed. | |
e24f0b8f | 680 | */ |
62e1c553 SL |
681 | if (!page->mapping) { |
682 | if (!PageAnon(page) && PagePrivate(page)) { | |
683 | /* | |
684 | * Go direct to try_to_free_buffers() here because | |
685 | * a) that's what try_to_release_page() would do anyway | |
686 | * b) we may be under rcu_read_lock() here, so we can't | |
687 | * use GFP_KERNEL which is what try_to_release_page() | |
688 | * needs to be effective. | |
689 | */ | |
690 | try_to_free_buffers(page); | |
691 | } | |
dc386d4d | 692 | goto rcu_unlock; |
62e1c553 SL |
693 | } |
694 | ||
ae41be37 | 695 | charge = mem_cgroup_prepare_migration(page); |
dc386d4d | 696 | /* Establish migration ptes or remove ptes */ |
e6a1530d | 697 | try_to_unmap(page, 1); |
dc386d4d | 698 | |
e6a1530d CL |
699 | if (!page_mapped(page)) |
700 | rc = move_to_new_page(newpage, page); | |
e24f0b8f | 701 | |
ae41be37 | 702 | if (rc) { |
e24f0b8f | 703 | remove_migration_ptes(page, page); |
ae41be37 KH |
704 | if (charge) |
705 | mem_cgroup_end_migration(page); | |
706 | } else if (charge) | |
707 | mem_cgroup_end_migration(newpage); | |
dc386d4d | 708 | rcu_unlock: |
989f89c5 KH |
709 | if (rcu_locked) |
710 | rcu_read_unlock(); | |
e6a1530d | 711 | |
e24f0b8f | 712 | unlock: |
dc386d4d | 713 | |
e24f0b8f | 714 | unlock_page(page); |
95a402c3 | 715 | |
e24f0b8f | 716 | if (rc != -EAGAIN) { |
aaa994b3 CL |
717 | /* |
718 | * A page that has been migrated has all references | |
719 | * removed and will be freed. A page that has not been | |
720 | * migrated will have kepts its references and be | |
721 | * restored. | |
722 | */ | |
723 | list_del(&page->lru); | |
724 | move_to_lru(page); | |
e24f0b8f | 725 | } |
95a402c3 CL |
726 | |
727 | move_newpage: | |
728 | /* | |
729 | * Move the new page to the LRU. If migration was not successful | |
730 | * then this will free the page. | |
731 | */ | |
732 | move_to_lru(newpage); | |
742755a1 CL |
733 | if (result) { |
734 | if (rc) | |
735 | *result = rc; | |
736 | else | |
737 | *result = page_to_nid(newpage); | |
738 | } | |
e24f0b8f CL |
739 | return rc; |
740 | } | |
741 | ||
b20a3503 CL |
742 | /* |
743 | * migrate_pages | |
744 | * | |
95a402c3 CL |
745 | * The function takes one list of pages to migrate and a function |
746 | * that determines from the page to be migrated and the private data | |
747 | * the target of the move and allocates the page. | |
b20a3503 CL |
748 | * |
749 | * The function returns after 10 attempts or if no pages | |
750 | * are movable anymore because to has become empty | |
aaa994b3 | 751 | * or no retryable pages exist anymore. All pages will be |
e9534b3f | 752 | * returned to the LRU or freed. |
b20a3503 | 753 | * |
95a402c3 | 754 | * Return: Number of pages not migrated or error code. |
b20a3503 | 755 | */ |
95a402c3 CL |
756 | int migrate_pages(struct list_head *from, |
757 | new_page_t get_new_page, unsigned long private) | |
b20a3503 | 758 | { |
e24f0b8f | 759 | int retry = 1; |
b20a3503 CL |
760 | int nr_failed = 0; |
761 | int pass = 0; | |
762 | struct page *page; | |
763 | struct page *page2; | |
764 | int swapwrite = current->flags & PF_SWAPWRITE; | |
765 | int rc; | |
766 | ||
767 | if (!swapwrite) | |
768 | current->flags |= PF_SWAPWRITE; | |
769 | ||
e24f0b8f CL |
770 | for(pass = 0; pass < 10 && retry; pass++) { |
771 | retry = 0; | |
b20a3503 | 772 | |
e24f0b8f | 773 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 774 | cond_resched(); |
2d1db3b1 | 775 | |
95a402c3 CL |
776 | rc = unmap_and_move(get_new_page, private, |
777 | page, pass > 2); | |
2d1db3b1 | 778 | |
e24f0b8f | 779 | switch(rc) { |
95a402c3 CL |
780 | case -ENOMEM: |
781 | goto out; | |
e24f0b8f | 782 | case -EAGAIN: |
2d1db3b1 | 783 | retry++; |
e24f0b8f CL |
784 | break; |
785 | case 0: | |
e24f0b8f CL |
786 | break; |
787 | default: | |
2d1db3b1 | 788 | /* Permanent failure */ |
2d1db3b1 | 789 | nr_failed++; |
e24f0b8f | 790 | break; |
2d1db3b1 | 791 | } |
b20a3503 CL |
792 | } |
793 | } | |
95a402c3 CL |
794 | rc = 0; |
795 | out: | |
b20a3503 CL |
796 | if (!swapwrite) |
797 | current->flags &= ~PF_SWAPWRITE; | |
798 | ||
aaa994b3 | 799 | putback_lru_pages(from); |
b20a3503 | 800 | |
95a402c3 CL |
801 | if (rc) |
802 | return rc; | |
b20a3503 | 803 | |
95a402c3 | 804 | return nr_failed + retry; |
b20a3503 | 805 | } |
95a402c3 | 806 | |
742755a1 CL |
807 | #ifdef CONFIG_NUMA |
808 | /* | |
809 | * Move a list of individual pages | |
810 | */ | |
811 | struct page_to_node { | |
812 | unsigned long addr; | |
813 | struct page *page; | |
814 | int node; | |
815 | int status; | |
816 | }; | |
817 | ||
818 | static struct page *new_page_node(struct page *p, unsigned long private, | |
819 | int **result) | |
820 | { | |
821 | struct page_to_node *pm = (struct page_to_node *)private; | |
822 | ||
823 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
824 | pm++; | |
825 | ||
826 | if (pm->node == MAX_NUMNODES) | |
827 | return NULL; | |
828 | ||
829 | *result = &pm->status; | |
830 | ||
769848c0 MG |
831 | return alloc_pages_node(pm->node, |
832 | GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0); | |
742755a1 CL |
833 | } |
834 | ||
835 | /* | |
836 | * Move a set of pages as indicated in the pm array. The addr | |
837 | * field must be set to the virtual address of the page to be moved | |
838 | * and the node number must contain a valid target node. | |
839 | */ | |
840 | static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm, | |
841 | int migrate_all) | |
842 | { | |
843 | int err; | |
844 | struct page_to_node *pp; | |
845 | LIST_HEAD(pagelist); | |
846 | ||
847 | down_read(&mm->mmap_sem); | |
848 | ||
849 | /* | |
850 | * Build a list of pages to migrate | |
851 | */ | |
852 | migrate_prep(); | |
853 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { | |
854 | struct vm_area_struct *vma; | |
855 | struct page *page; | |
856 | ||
857 | /* | |
858 | * A valid page pointer that will not match any of the | |
859 | * pages that will be moved. | |
860 | */ | |
861 | pp->page = ZERO_PAGE(0); | |
862 | ||
863 | err = -EFAULT; | |
864 | vma = find_vma(mm, pp->addr); | |
0dc952dc | 865 | if (!vma || !vma_migratable(vma)) |
742755a1 CL |
866 | goto set_status; |
867 | ||
868 | page = follow_page(vma, pp->addr, FOLL_GET); | |
89f5b7da LT |
869 | |
870 | err = PTR_ERR(page); | |
871 | if (IS_ERR(page)) | |
872 | goto set_status; | |
873 | ||
742755a1 CL |
874 | err = -ENOENT; |
875 | if (!page) | |
876 | goto set_status; | |
877 | ||
878 | if (PageReserved(page)) /* Check for zero page */ | |
879 | goto put_and_set; | |
880 | ||
881 | pp->page = page; | |
882 | err = page_to_nid(page); | |
883 | ||
884 | if (err == pp->node) | |
885 | /* | |
886 | * Node already in the right place | |
887 | */ | |
888 | goto put_and_set; | |
889 | ||
890 | err = -EACCES; | |
891 | if (page_mapcount(page) > 1 && | |
892 | !migrate_all) | |
893 | goto put_and_set; | |
894 | ||
895 | err = isolate_lru_page(page, &pagelist); | |
896 | put_and_set: | |
897 | /* | |
898 | * Either remove the duplicate refcount from | |
899 | * isolate_lru_page() or drop the page ref if it was | |
900 | * not isolated. | |
901 | */ | |
902 | put_page(page); | |
903 | set_status: | |
904 | pp->status = err; | |
905 | } | |
906 | ||
907 | if (!list_empty(&pagelist)) | |
908 | err = migrate_pages(&pagelist, new_page_node, | |
909 | (unsigned long)pm); | |
910 | else | |
911 | err = -ENOENT; | |
912 | ||
913 | up_read(&mm->mmap_sem); | |
914 | return err; | |
915 | } | |
916 | ||
917 | /* | |
918 | * Determine the nodes of a list of pages. The addr in the pm array | |
919 | * must have been set to the virtual address of which we want to determine | |
920 | * the node number. | |
921 | */ | |
922 | static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm) | |
923 | { | |
924 | down_read(&mm->mmap_sem); | |
925 | ||
926 | for ( ; pm->node != MAX_NUMNODES; pm++) { | |
927 | struct vm_area_struct *vma; | |
928 | struct page *page; | |
929 | int err; | |
930 | ||
931 | err = -EFAULT; | |
932 | vma = find_vma(mm, pm->addr); | |
933 | if (!vma) | |
934 | goto set_status; | |
935 | ||
936 | page = follow_page(vma, pm->addr, 0); | |
89f5b7da LT |
937 | |
938 | err = PTR_ERR(page); | |
939 | if (IS_ERR(page)) | |
940 | goto set_status; | |
941 | ||
742755a1 CL |
942 | err = -ENOENT; |
943 | /* Use PageReserved to check for zero page */ | |
944 | if (!page || PageReserved(page)) | |
945 | goto set_status; | |
946 | ||
947 | err = page_to_nid(page); | |
948 | set_status: | |
949 | pm->status = err; | |
950 | } | |
951 | ||
952 | up_read(&mm->mmap_sem); | |
953 | return 0; | |
954 | } | |
955 | ||
956 | /* | |
957 | * Move a list of pages in the address space of the currently executing | |
958 | * process. | |
959 | */ | |
960 | asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages, | |
961 | const void __user * __user *pages, | |
962 | const int __user *nodes, | |
963 | int __user *status, int flags) | |
964 | { | |
965 | int err = 0; | |
966 | int i; | |
967 | struct task_struct *task; | |
968 | nodemask_t task_nodes; | |
969 | struct mm_struct *mm; | |
970 | struct page_to_node *pm = NULL; | |
971 | ||
972 | /* Check flags */ | |
973 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
974 | return -EINVAL; | |
975 | ||
976 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
977 | return -EPERM; | |
978 | ||
979 | /* Find the mm_struct */ | |
980 | read_lock(&tasklist_lock); | |
228ebcbe | 981 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 CL |
982 | if (!task) { |
983 | read_unlock(&tasklist_lock); | |
984 | return -ESRCH; | |
985 | } | |
986 | mm = get_task_mm(task); | |
987 | read_unlock(&tasklist_lock); | |
988 | ||
989 | if (!mm) | |
990 | return -EINVAL; | |
991 | ||
992 | /* | |
993 | * Check if this process has the right to modify the specified | |
994 | * process. The right exists if the process has administrative | |
995 | * capabilities, superuser privileges or the same | |
996 | * userid as the target process. | |
997 | */ | |
998 | if ((current->euid != task->suid) && (current->euid != task->uid) && | |
999 | (current->uid != task->suid) && (current->uid != task->uid) && | |
1000 | !capable(CAP_SYS_NICE)) { | |
1001 | err = -EPERM; | |
1002 | goto out2; | |
1003 | } | |
1004 | ||
86c3a764 DQ |
1005 | err = security_task_movememory(task); |
1006 | if (err) | |
1007 | goto out2; | |
1008 | ||
1009 | ||
742755a1 CL |
1010 | task_nodes = cpuset_mems_allowed(task); |
1011 | ||
1012 | /* Limit nr_pages so that the multiplication may not overflow */ | |
1013 | if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) { | |
1014 | err = -E2BIG; | |
1015 | goto out2; | |
1016 | } | |
1017 | ||
1018 | pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node)); | |
1019 | if (!pm) { | |
1020 | err = -ENOMEM; | |
1021 | goto out2; | |
1022 | } | |
1023 | ||
1024 | /* | |
1025 | * Get parameters from user space and initialize the pm | |
1026 | * array. Return various errors if the user did something wrong. | |
1027 | */ | |
1028 | for (i = 0; i < nr_pages; i++) { | |
9d966d49 | 1029 | const void __user *p; |
742755a1 CL |
1030 | |
1031 | err = -EFAULT; | |
1032 | if (get_user(p, pages + i)) | |
1033 | goto out; | |
1034 | ||
1035 | pm[i].addr = (unsigned long)p; | |
1036 | if (nodes) { | |
1037 | int node; | |
1038 | ||
1039 | if (get_user(node, nodes + i)) | |
1040 | goto out; | |
1041 | ||
1042 | err = -ENODEV; | |
56bbd65d | 1043 | if (!node_state(node, N_HIGH_MEMORY)) |
742755a1 CL |
1044 | goto out; |
1045 | ||
1046 | err = -EACCES; | |
1047 | if (!node_isset(node, task_nodes)) | |
1048 | goto out; | |
1049 | ||
1050 | pm[i].node = node; | |
8ce08464 SR |
1051 | } else |
1052 | pm[i].node = 0; /* anything to not match MAX_NUMNODES */ | |
742755a1 CL |
1053 | } |
1054 | /* End marker */ | |
1055 | pm[nr_pages].node = MAX_NUMNODES; | |
1056 | ||
1057 | if (nodes) | |
1058 | err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL); | |
1059 | else | |
1060 | err = do_pages_stat(mm, pm); | |
1061 | ||
1062 | if (err >= 0) | |
1063 | /* Return status information */ | |
1064 | for (i = 0; i < nr_pages; i++) | |
1065 | if (put_user(pm[i].status, status + i)) | |
1066 | err = -EFAULT; | |
1067 | ||
1068 | out: | |
1069 | vfree(pm); | |
1070 | out2: | |
1071 | mmput(mm); | |
1072 | return err; | |
1073 | } | |
1074 | #endif | |
1075 | ||
7b2259b3 CL |
1076 | /* |
1077 | * Call migration functions in the vma_ops that may prepare | |
1078 | * memory in a vm for migration. migration functions may perform | |
1079 | * the migration for vmas that do not have an underlying page struct. | |
1080 | */ | |
1081 | int migrate_vmas(struct mm_struct *mm, const nodemask_t *to, | |
1082 | const nodemask_t *from, unsigned long flags) | |
1083 | { | |
1084 | struct vm_area_struct *vma; | |
1085 | int err = 0; | |
1086 | ||
1087 | for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) { | |
1088 | if (vma->vm_ops && vma->vm_ops->migrate) { | |
1089 | err = vma->vm_ops->migrate(vma, to, from, flags); | |
1090 | if (err) | |
1091 | break; | |
1092 | } | |
1093 | } | |
1094 | return err; | |
1095 | } |