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> | |
12 | * Christoph Lameter <clameter@sgi.com> | |
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> |
b20a3503 CL |
32 | |
33 | #include "internal.h" | |
34 | ||
b20a3503 CL |
35 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) |
36 | ||
37 | /* | |
38 | * Isolate one page from the LRU lists. If successful put it onto | |
39 | * the indicated list with elevated page count. | |
40 | * | |
41 | * Result: | |
42 | * -EBUSY: page not on LRU list | |
43 | * 0: page removed from LRU list and added to the specified list. | |
44 | */ | |
45 | int isolate_lru_page(struct page *page, struct list_head *pagelist) | |
46 | { | |
47 | int ret = -EBUSY; | |
48 | ||
49 | if (PageLRU(page)) { | |
50 | struct zone *zone = page_zone(page); | |
51 | ||
52 | spin_lock_irq(&zone->lru_lock); | |
3dd9fe8c | 53 | if (PageLRU(page) && get_page_unless_zero(page)) { |
b20a3503 | 54 | ret = 0; |
b20a3503 CL |
55 | ClearPageLRU(page); |
56 | if (PageActive(page)) | |
57 | del_page_from_active_list(zone, page); | |
58 | else | |
59 | del_page_from_inactive_list(zone, page); | |
60 | list_add_tail(&page->lru, pagelist); | |
61 | } | |
62 | spin_unlock_irq(&zone->lru_lock); | |
63 | } | |
64 | return ret; | |
65 | } | |
66 | ||
67 | /* | |
742755a1 CL |
68 | * migrate_prep() needs to be called before we start compiling a list of pages |
69 | * to be migrated using isolate_lru_page(). | |
b20a3503 CL |
70 | */ |
71 | int migrate_prep(void) | |
72 | { | |
b20a3503 CL |
73 | /* |
74 | * Clear the LRU lists so pages can be isolated. | |
75 | * Note that pages may be moved off the LRU after we have | |
76 | * drained them. Those pages will fail to migrate like other | |
77 | * pages that may be busy. | |
78 | */ | |
79 | lru_add_drain_all(); | |
80 | ||
81 | return 0; | |
82 | } | |
83 | ||
84 | static inline void move_to_lru(struct page *page) | |
85 | { | |
b20a3503 CL |
86 | if (PageActive(page)) { |
87 | /* | |
88 | * lru_cache_add_active checks that | |
89 | * the PG_active bit is off. | |
90 | */ | |
91 | ClearPageActive(page); | |
92 | lru_cache_add_active(page); | |
93 | } else { | |
94 | lru_cache_add(page); | |
95 | } | |
96 | put_page(page); | |
97 | } | |
98 | ||
99 | /* | |
100 | * Add isolated pages on the list back to the LRU. | |
101 | * | |
102 | * returns the number of pages put back. | |
103 | */ | |
104 | int putback_lru_pages(struct list_head *l) | |
105 | { | |
106 | struct page *page; | |
107 | struct page *page2; | |
108 | int count = 0; | |
109 | ||
110 | list_for_each_entry_safe(page, page2, l, lru) { | |
e24f0b8f | 111 | list_del(&page->lru); |
b20a3503 CL |
112 | move_to_lru(page); |
113 | count++; | |
114 | } | |
115 | return count; | |
116 | } | |
117 | ||
0697212a CL |
118 | static inline int is_swap_pte(pte_t pte) |
119 | { | |
120 | return !pte_none(pte) && !pte_present(pte) && !pte_file(pte); | |
121 | } | |
122 | ||
123 | /* | |
124 | * Restore a potential migration pte to a working pte entry | |
125 | */ | |
04e62a29 | 126 | static void remove_migration_pte(struct vm_area_struct *vma, |
0697212a CL |
127 | struct page *old, struct page *new) |
128 | { | |
129 | struct mm_struct *mm = vma->vm_mm; | |
130 | swp_entry_t entry; | |
131 | pgd_t *pgd; | |
132 | pud_t *pud; | |
133 | pmd_t *pmd; | |
134 | pte_t *ptep, pte; | |
135 | spinlock_t *ptl; | |
04e62a29 CL |
136 | unsigned long addr = page_address_in_vma(new, vma); |
137 | ||
138 | if (addr == -EFAULT) | |
139 | return; | |
0697212a CL |
140 | |
141 | pgd = pgd_offset(mm, addr); | |
142 | if (!pgd_present(*pgd)) | |
143 | return; | |
144 | ||
145 | pud = pud_offset(pgd, addr); | |
146 | if (!pud_present(*pud)) | |
147 | return; | |
148 | ||
149 | pmd = pmd_offset(pud, addr); | |
150 | if (!pmd_present(*pmd)) | |
151 | return; | |
152 | ||
153 | ptep = pte_offset_map(pmd, addr); | |
154 | ||
155 | if (!is_swap_pte(*ptep)) { | |
156 | pte_unmap(ptep); | |
157 | return; | |
158 | } | |
159 | ||
160 | ptl = pte_lockptr(mm, pmd); | |
161 | spin_lock(ptl); | |
162 | pte = *ptep; | |
163 | if (!is_swap_pte(pte)) | |
164 | goto out; | |
165 | ||
166 | entry = pte_to_swp_entry(pte); | |
167 | ||
168 | if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old) | |
169 | goto out; | |
170 | ||
0697212a CL |
171 | get_page(new); |
172 | pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); | |
173 | if (is_write_migration_entry(entry)) | |
174 | pte = pte_mkwrite(pte); | |
97ee0524 | 175 | flush_cache_page(vma, addr, pte_pfn(pte)); |
0697212a | 176 | set_pte_at(mm, addr, ptep, pte); |
04e62a29 CL |
177 | |
178 | if (PageAnon(new)) | |
179 | page_add_anon_rmap(new, vma, addr); | |
180 | else | |
181 | page_add_file_rmap(new); | |
182 | ||
183 | /* No need to invalidate - it was non-present before */ | |
184 | update_mmu_cache(vma, addr, pte); | |
04e62a29 | 185 | |
0697212a CL |
186 | out: |
187 | pte_unmap_unlock(ptep, ptl); | |
188 | } | |
189 | ||
190 | /* | |
04e62a29 CL |
191 | * Note that remove_file_migration_ptes will only work on regular mappings, |
192 | * Nonlinear mappings do not use migration entries. | |
193 | */ | |
194 | static void remove_file_migration_ptes(struct page *old, struct page *new) | |
195 | { | |
196 | struct vm_area_struct *vma; | |
197 | struct address_space *mapping = page_mapping(new); | |
198 | struct prio_tree_iter iter; | |
199 | pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
200 | ||
201 | if (!mapping) | |
202 | return; | |
203 | ||
204 | spin_lock(&mapping->i_mmap_lock); | |
205 | ||
206 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) | |
207 | remove_migration_pte(vma, old, new); | |
208 | ||
209 | spin_unlock(&mapping->i_mmap_lock); | |
210 | } | |
211 | ||
212 | /* | |
0697212a CL |
213 | * Must hold mmap_sem lock on at least one of the vmas containing |
214 | * the page so that the anon_vma cannot vanish. | |
215 | */ | |
04e62a29 | 216 | static void remove_anon_migration_ptes(struct page *old, struct page *new) |
0697212a CL |
217 | { |
218 | struct anon_vma *anon_vma; | |
219 | struct vm_area_struct *vma; | |
220 | unsigned long mapping; | |
221 | ||
222 | mapping = (unsigned long)new->mapping; | |
223 | ||
224 | if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0) | |
225 | return; | |
226 | ||
227 | /* | |
228 | * We hold the mmap_sem lock. So no need to call page_lock_anon_vma. | |
229 | */ | |
230 | anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON); | |
231 | spin_lock(&anon_vma->lock); | |
232 | ||
233 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) | |
04e62a29 | 234 | remove_migration_pte(vma, old, new); |
0697212a CL |
235 | |
236 | spin_unlock(&anon_vma->lock); | |
237 | } | |
238 | ||
04e62a29 CL |
239 | /* |
240 | * Get rid of all migration entries and replace them by | |
241 | * references to the indicated page. | |
242 | */ | |
243 | static void remove_migration_ptes(struct page *old, struct page *new) | |
244 | { | |
245 | if (PageAnon(new)) | |
246 | remove_anon_migration_ptes(old, new); | |
247 | else | |
248 | remove_file_migration_ptes(old, new); | |
249 | } | |
250 | ||
0697212a CL |
251 | /* |
252 | * Something used the pte of a page under migration. We need to | |
253 | * get to the page and wait until migration is finished. | |
254 | * When we return from this function the fault will be retried. | |
255 | * | |
256 | * This function is called from do_swap_page(). | |
257 | */ | |
258 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, | |
259 | unsigned long address) | |
260 | { | |
261 | pte_t *ptep, pte; | |
262 | spinlock_t *ptl; | |
263 | swp_entry_t entry; | |
264 | struct page *page; | |
265 | ||
266 | ptep = pte_offset_map_lock(mm, pmd, address, &ptl); | |
267 | pte = *ptep; | |
268 | if (!is_swap_pte(pte)) | |
269 | goto out; | |
270 | ||
271 | entry = pte_to_swp_entry(pte); | |
272 | if (!is_migration_entry(entry)) | |
273 | goto out; | |
274 | ||
275 | page = migration_entry_to_page(entry); | |
276 | ||
277 | get_page(page); | |
278 | pte_unmap_unlock(ptep, ptl); | |
279 | wait_on_page_locked(page); | |
280 | put_page(page); | |
281 | return; | |
282 | out: | |
283 | pte_unmap_unlock(ptep, ptl); | |
284 | } | |
285 | ||
b20a3503 | 286 | /* |
c3fcf8a5 | 287 | * Replace the page in the mapping. |
5b5c7120 CL |
288 | * |
289 | * The number of remaining references must be: | |
290 | * 1 for anonymous pages without a mapping | |
291 | * 2 for pages with a mapping | |
292 | * 3 for pages with a mapping and PagePrivate set. | |
b20a3503 | 293 | */ |
2d1db3b1 CL |
294 | static int migrate_page_move_mapping(struct address_space *mapping, |
295 | struct page *newpage, struct page *page) | |
b20a3503 | 296 | { |
7cf9c2c7 | 297 | void **pslot; |
b20a3503 | 298 | |
6c5240ae | 299 | if (!mapping) { |
0e8c7d0f | 300 | /* Anonymous page without mapping */ |
6c5240ae CL |
301 | if (page_count(page) != 1) |
302 | return -EAGAIN; | |
303 | return 0; | |
304 | } | |
305 | ||
b20a3503 CL |
306 | write_lock_irq(&mapping->tree_lock); |
307 | ||
7cf9c2c7 NP |
308 | pslot = radix_tree_lookup_slot(&mapping->page_tree, |
309 | page_index(page)); | |
b20a3503 | 310 | |
6c5240ae | 311 | if (page_count(page) != 2 + !!PagePrivate(page) || |
7cf9c2c7 | 312 | (struct page *)radix_tree_deref_slot(pslot) != page) { |
b20a3503 | 313 | write_unlock_irq(&mapping->tree_lock); |
e23ca00b | 314 | return -EAGAIN; |
b20a3503 CL |
315 | } |
316 | ||
317 | /* | |
318 | * Now we know that no one else is looking at the page. | |
b20a3503 | 319 | */ |
7cf9c2c7 | 320 | get_page(newpage); /* add cache reference */ |
6c5240ae | 321 | #ifdef CONFIG_SWAP |
b20a3503 CL |
322 | if (PageSwapCache(page)) { |
323 | SetPageSwapCache(newpage); | |
324 | set_page_private(newpage, page_private(page)); | |
325 | } | |
6c5240ae | 326 | #endif |
b20a3503 | 327 | |
7cf9c2c7 NP |
328 | radix_tree_replace_slot(pslot, newpage); |
329 | ||
330 | /* | |
331 | * Drop cache reference from old page. | |
332 | * We know this isn't the last reference. | |
333 | */ | |
b20a3503 | 334 | __put_page(page); |
7cf9c2c7 | 335 | |
0e8c7d0f CL |
336 | /* |
337 | * If moved to a different zone then also account | |
338 | * the page for that zone. Other VM counters will be | |
339 | * taken care of when we establish references to the | |
340 | * new page and drop references to the old page. | |
341 | * | |
342 | * Note that anonymous pages are accounted for | |
343 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
344 | * are mapped to swap space. | |
345 | */ | |
346 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
347 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
348 | ||
b20a3503 CL |
349 | write_unlock_irq(&mapping->tree_lock); |
350 | ||
351 | return 0; | |
352 | } | |
b20a3503 CL |
353 | |
354 | /* | |
355 | * Copy the page to its new location | |
356 | */ | |
e7340f73 | 357 | static void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 CL |
358 | { |
359 | copy_highpage(newpage, page); | |
360 | ||
361 | if (PageError(page)) | |
362 | SetPageError(newpage); | |
363 | if (PageReferenced(page)) | |
364 | SetPageReferenced(newpage); | |
365 | if (PageUptodate(page)) | |
366 | SetPageUptodate(newpage); | |
367 | if (PageActive(page)) | |
368 | SetPageActive(newpage); | |
369 | if (PageChecked(page)) | |
370 | SetPageChecked(newpage); | |
371 | if (PageMappedToDisk(page)) | |
372 | SetPageMappedToDisk(newpage); | |
373 | ||
374 | if (PageDirty(page)) { | |
375 | clear_page_dirty_for_io(page); | |
376 | set_page_dirty(newpage); | |
377 | } | |
378 | ||
6c5240ae | 379 | #ifdef CONFIG_SWAP |
b20a3503 | 380 | ClearPageSwapCache(page); |
6c5240ae | 381 | #endif |
b20a3503 CL |
382 | ClearPageActive(page); |
383 | ClearPagePrivate(page); | |
384 | set_page_private(page, 0); | |
385 | page->mapping = NULL; | |
386 | ||
387 | /* | |
388 | * If any waiters have accumulated on the new page then | |
389 | * wake them up. | |
390 | */ | |
391 | if (PageWriteback(newpage)) | |
392 | end_page_writeback(newpage); | |
393 | } | |
b20a3503 | 394 | |
1d8b85cc CL |
395 | /************************************************************ |
396 | * Migration functions | |
397 | ***********************************************************/ | |
398 | ||
399 | /* Always fail migration. Used for mappings that are not movable */ | |
2d1db3b1 CL |
400 | int fail_migrate_page(struct address_space *mapping, |
401 | struct page *newpage, struct page *page) | |
1d8b85cc CL |
402 | { |
403 | return -EIO; | |
404 | } | |
405 | EXPORT_SYMBOL(fail_migrate_page); | |
406 | ||
b20a3503 CL |
407 | /* |
408 | * Common logic to directly migrate a single page suitable for | |
409 | * pages that do not use PagePrivate. | |
410 | * | |
411 | * Pages are locked upon entry and exit. | |
412 | */ | |
2d1db3b1 CL |
413 | int migrate_page(struct address_space *mapping, |
414 | struct page *newpage, struct page *page) | |
b20a3503 CL |
415 | { |
416 | int rc; | |
417 | ||
418 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
419 | ||
2d1db3b1 | 420 | rc = migrate_page_move_mapping(mapping, newpage, page); |
b20a3503 CL |
421 | |
422 | if (rc) | |
423 | return rc; | |
424 | ||
425 | migrate_page_copy(newpage, page); | |
b20a3503 CL |
426 | return 0; |
427 | } | |
428 | EXPORT_SYMBOL(migrate_page); | |
429 | ||
9361401e | 430 | #ifdef CONFIG_BLOCK |
1d8b85cc CL |
431 | /* |
432 | * Migration function for pages with buffers. This function can only be used | |
433 | * if the underlying filesystem guarantees that no other references to "page" | |
434 | * exist. | |
435 | */ | |
2d1db3b1 CL |
436 | int buffer_migrate_page(struct address_space *mapping, |
437 | struct page *newpage, struct page *page) | |
1d8b85cc | 438 | { |
1d8b85cc CL |
439 | struct buffer_head *bh, *head; |
440 | int rc; | |
441 | ||
1d8b85cc | 442 | if (!page_has_buffers(page)) |
2d1db3b1 | 443 | return migrate_page(mapping, newpage, page); |
1d8b85cc CL |
444 | |
445 | head = page_buffers(page); | |
446 | ||
2d1db3b1 | 447 | rc = migrate_page_move_mapping(mapping, newpage, page); |
1d8b85cc CL |
448 | |
449 | if (rc) | |
450 | return rc; | |
451 | ||
452 | bh = head; | |
453 | do { | |
454 | get_bh(bh); | |
455 | lock_buffer(bh); | |
456 | bh = bh->b_this_page; | |
457 | ||
458 | } while (bh != head); | |
459 | ||
460 | ClearPagePrivate(page); | |
461 | set_page_private(newpage, page_private(page)); | |
462 | set_page_private(page, 0); | |
463 | put_page(page); | |
464 | get_page(newpage); | |
465 | ||
466 | bh = head; | |
467 | do { | |
468 | set_bh_page(bh, newpage, bh_offset(bh)); | |
469 | bh = bh->b_this_page; | |
470 | ||
471 | } while (bh != head); | |
472 | ||
473 | SetPagePrivate(newpage); | |
474 | ||
475 | migrate_page_copy(newpage, page); | |
476 | ||
477 | bh = head; | |
478 | do { | |
479 | unlock_buffer(bh); | |
480 | put_bh(bh); | |
481 | bh = bh->b_this_page; | |
482 | ||
483 | } while (bh != head); | |
484 | ||
485 | return 0; | |
486 | } | |
487 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 488 | #endif |
1d8b85cc | 489 | |
04e62a29 CL |
490 | /* |
491 | * Writeback a page to clean the dirty state | |
492 | */ | |
493 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 494 | { |
04e62a29 CL |
495 | struct writeback_control wbc = { |
496 | .sync_mode = WB_SYNC_NONE, | |
497 | .nr_to_write = 1, | |
498 | .range_start = 0, | |
499 | .range_end = LLONG_MAX, | |
500 | .nonblocking = 1, | |
501 | .for_reclaim = 1 | |
502 | }; | |
503 | int rc; | |
504 | ||
505 | if (!mapping->a_ops->writepage) | |
506 | /* No write method for the address space */ | |
507 | return -EINVAL; | |
508 | ||
509 | if (!clear_page_dirty_for_io(page)) | |
510 | /* Someone else already triggered a write */ | |
511 | return -EAGAIN; | |
512 | ||
8351a6e4 | 513 | /* |
04e62a29 CL |
514 | * A dirty page may imply that the underlying filesystem has |
515 | * the page on some queue. So the page must be clean for | |
516 | * migration. Writeout may mean we loose the lock and the | |
517 | * page state is no longer what we checked for earlier. | |
518 | * At this point we know that the migration attempt cannot | |
519 | * be successful. | |
8351a6e4 | 520 | */ |
04e62a29 | 521 | remove_migration_ptes(page, page); |
8351a6e4 | 522 | |
04e62a29 CL |
523 | rc = mapping->a_ops->writepage(page, &wbc); |
524 | if (rc < 0) | |
525 | /* I/O Error writing */ | |
526 | return -EIO; | |
8351a6e4 | 527 | |
04e62a29 CL |
528 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
529 | /* unlocked. Relock */ | |
530 | lock_page(page); | |
531 | ||
532 | return -EAGAIN; | |
533 | } | |
534 | ||
535 | /* | |
536 | * Default handling if a filesystem does not provide a migration function. | |
537 | */ | |
538 | static int fallback_migrate_page(struct address_space *mapping, | |
539 | struct page *newpage, struct page *page) | |
540 | { | |
541 | if (PageDirty(page)) | |
542 | return writeout(mapping, page); | |
8351a6e4 CL |
543 | |
544 | /* | |
545 | * Buffers may be managed in a filesystem specific way. | |
546 | * We must have no buffers or drop them. | |
547 | */ | |
b398f6bf | 548 | if (PagePrivate(page) && |
8351a6e4 CL |
549 | !try_to_release_page(page, GFP_KERNEL)) |
550 | return -EAGAIN; | |
551 | ||
552 | return migrate_page(mapping, newpage, page); | |
553 | } | |
554 | ||
e24f0b8f CL |
555 | /* |
556 | * Move a page to a newly allocated page | |
557 | * The page is locked and all ptes have been successfully removed. | |
558 | * | |
559 | * The new page will have replaced the old page if this function | |
560 | * is successful. | |
561 | */ | |
562 | static int move_to_new_page(struct page *newpage, struct page *page) | |
563 | { | |
564 | struct address_space *mapping; | |
565 | int rc; | |
566 | ||
567 | /* | |
568 | * Block others from accessing the page when we get around to | |
569 | * establishing additional references. We are the only one | |
570 | * holding a reference to the new page at this point. | |
571 | */ | |
572 | if (TestSetPageLocked(newpage)) | |
573 | BUG(); | |
574 | ||
575 | /* Prepare mapping for the new page.*/ | |
576 | newpage->index = page->index; | |
577 | newpage->mapping = page->mapping; | |
578 | ||
579 | mapping = page_mapping(page); | |
580 | if (!mapping) | |
581 | rc = migrate_page(mapping, newpage, page); | |
582 | else if (mapping->a_ops->migratepage) | |
583 | /* | |
584 | * Most pages have a mapping and most filesystems | |
585 | * should provide a migration function. Anonymous | |
586 | * pages are part of swap space which also has its | |
587 | * own migration function. This is the most common | |
588 | * path for page migration. | |
589 | */ | |
590 | rc = mapping->a_ops->migratepage(mapping, | |
591 | newpage, page); | |
592 | else | |
593 | rc = fallback_migrate_page(mapping, newpage, page); | |
594 | ||
595 | if (!rc) | |
596 | remove_migration_ptes(page, newpage); | |
597 | else | |
598 | newpage->mapping = NULL; | |
599 | ||
600 | unlock_page(newpage); | |
601 | ||
602 | return rc; | |
603 | } | |
604 | ||
605 | /* | |
606 | * Obtain the lock on page, remove all ptes and migrate the page | |
607 | * to the newly allocated page in newpage. | |
608 | */ | |
95a402c3 CL |
609 | static int unmap_and_move(new_page_t get_new_page, unsigned long private, |
610 | struct page *page, int force) | |
e24f0b8f CL |
611 | { |
612 | int rc = 0; | |
742755a1 CL |
613 | int *result = NULL; |
614 | struct page *newpage = get_new_page(page, private, &result); | |
989f89c5 | 615 | int rcu_locked = 0; |
95a402c3 CL |
616 | |
617 | if (!newpage) | |
618 | return -ENOMEM; | |
e24f0b8f CL |
619 | |
620 | if (page_count(page) == 1) | |
621 | /* page was freed from under us. So we are done. */ | |
95a402c3 | 622 | goto move_newpage; |
e24f0b8f CL |
623 | |
624 | rc = -EAGAIN; | |
625 | if (TestSetPageLocked(page)) { | |
626 | if (!force) | |
95a402c3 | 627 | goto move_newpage; |
e24f0b8f CL |
628 | lock_page(page); |
629 | } | |
630 | ||
631 | if (PageWriteback(page)) { | |
632 | if (!force) | |
633 | goto unlock; | |
634 | wait_on_page_writeback(page); | |
635 | } | |
e24f0b8f | 636 | /* |
dc386d4d KH |
637 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
638 | * we cannot notice that anon_vma is freed while we migrates a page. | |
639 | * This rcu_read_lock() delays freeing anon_vma pointer until the end | |
640 | * of migration. File cache pages are no problem because of page_lock() | |
989f89c5 KH |
641 | * File Caches may use write_page() or lock_page() in migration, then, |
642 | * just care Anon page here. | |
dc386d4d | 643 | */ |
989f89c5 KH |
644 | if (PageAnon(page)) { |
645 | rcu_read_lock(); | |
646 | rcu_locked = 1; | |
647 | } | |
dc386d4d KH |
648 | /* |
649 | * This is a corner case handling. | |
650 | * When a new swap-cache is read into, it is linked to LRU | |
651 | * and treated as swapcache but has no rmap yet. | |
652 | * Calling try_to_unmap() against a page->mapping==NULL page is | |
653 | * BUG. So handle it here. | |
e24f0b8f | 654 | */ |
dc386d4d KH |
655 | if (!page->mapping) |
656 | goto rcu_unlock; | |
657 | /* Establish migration ptes or remove ptes */ | |
e6a1530d | 658 | try_to_unmap(page, 1); |
dc386d4d | 659 | |
e6a1530d CL |
660 | if (!page_mapped(page)) |
661 | rc = move_to_new_page(newpage, page); | |
e24f0b8f CL |
662 | |
663 | if (rc) | |
664 | remove_migration_ptes(page, page); | |
dc386d4d | 665 | rcu_unlock: |
989f89c5 KH |
666 | if (rcu_locked) |
667 | rcu_read_unlock(); | |
e6a1530d | 668 | |
e24f0b8f | 669 | unlock: |
dc386d4d | 670 | |
e24f0b8f | 671 | unlock_page(page); |
95a402c3 | 672 | |
e24f0b8f | 673 | if (rc != -EAGAIN) { |
aaa994b3 CL |
674 | /* |
675 | * A page that has been migrated has all references | |
676 | * removed and will be freed. A page that has not been | |
677 | * migrated will have kepts its references and be | |
678 | * restored. | |
679 | */ | |
680 | list_del(&page->lru); | |
681 | move_to_lru(page); | |
e24f0b8f | 682 | } |
95a402c3 CL |
683 | |
684 | move_newpage: | |
685 | /* | |
686 | * Move the new page to the LRU. If migration was not successful | |
687 | * then this will free the page. | |
688 | */ | |
689 | move_to_lru(newpage); | |
742755a1 CL |
690 | if (result) { |
691 | if (rc) | |
692 | *result = rc; | |
693 | else | |
694 | *result = page_to_nid(newpage); | |
695 | } | |
e24f0b8f CL |
696 | return rc; |
697 | } | |
698 | ||
b20a3503 CL |
699 | /* |
700 | * migrate_pages | |
701 | * | |
95a402c3 CL |
702 | * The function takes one list of pages to migrate and a function |
703 | * that determines from the page to be migrated and the private data | |
704 | * the target of the move and allocates the page. | |
b20a3503 CL |
705 | * |
706 | * The function returns after 10 attempts or if no pages | |
707 | * are movable anymore because to has become empty | |
aaa994b3 CL |
708 | * or no retryable pages exist anymore. All pages will be |
709 | * retruned to the LRU or freed. | |
b20a3503 | 710 | * |
95a402c3 | 711 | * Return: Number of pages not migrated or error code. |
b20a3503 | 712 | */ |
95a402c3 CL |
713 | int migrate_pages(struct list_head *from, |
714 | new_page_t get_new_page, unsigned long private) | |
b20a3503 | 715 | { |
e24f0b8f | 716 | int retry = 1; |
b20a3503 CL |
717 | int nr_failed = 0; |
718 | int pass = 0; | |
719 | struct page *page; | |
720 | struct page *page2; | |
721 | int swapwrite = current->flags & PF_SWAPWRITE; | |
722 | int rc; | |
723 | ||
724 | if (!swapwrite) | |
725 | current->flags |= PF_SWAPWRITE; | |
726 | ||
e24f0b8f CL |
727 | for(pass = 0; pass < 10 && retry; pass++) { |
728 | retry = 0; | |
b20a3503 | 729 | |
e24f0b8f | 730 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 731 | cond_resched(); |
2d1db3b1 | 732 | |
95a402c3 CL |
733 | rc = unmap_and_move(get_new_page, private, |
734 | page, pass > 2); | |
2d1db3b1 | 735 | |
e24f0b8f | 736 | switch(rc) { |
95a402c3 CL |
737 | case -ENOMEM: |
738 | goto out; | |
e24f0b8f | 739 | case -EAGAIN: |
2d1db3b1 | 740 | retry++; |
e24f0b8f CL |
741 | break; |
742 | case 0: | |
e24f0b8f CL |
743 | break; |
744 | default: | |
2d1db3b1 | 745 | /* Permanent failure */ |
2d1db3b1 | 746 | nr_failed++; |
e24f0b8f | 747 | break; |
2d1db3b1 | 748 | } |
b20a3503 CL |
749 | } |
750 | } | |
95a402c3 CL |
751 | rc = 0; |
752 | out: | |
b20a3503 CL |
753 | if (!swapwrite) |
754 | current->flags &= ~PF_SWAPWRITE; | |
755 | ||
aaa994b3 | 756 | putback_lru_pages(from); |
b20a3503 | 757 | |
95a402c3 CL |
758 | if (rc) |
759 | return rc; | |
b20a3503 | 760 | |
95a402c3 | 761 | return nr_failed + retry; |
b20a3503 | 762 | } |
95a402c3 | 763 | |
742755a1 CL |
764 | #ifdef CONFIG_NUMA |
765 | /* | |
766 | * Move a list of individual pages | |
767 | */ | |
768 | struct page_to_node { | |
769 | unsigned long addr; | |
770 | struct page *page; | |
771 | int node; | |
772 | int status; | |
773 | }; | |
774 | ||
775 | static struct page *new_page_node(struct page *p, unsigned long private, | |
776 | int **result) | |
777 | { | |
778 | struct page_to_node *pm = (struct page_to_node *)private; | |
779 | ||
780 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
781 | pm++; | |
782 | ||
783 | if (pm->node == MAX_NUMNODES) | |
784 | return NULL; | |
785 | ||
786 | *result = &pm->status; | |
787 | ||
769848c0 MG |
788 | return alloc_pages_node(pm->node, |
789 | GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0); | |
742755a1 CL |
790 | } |
791 | ||
792 | /* | |
793 | * Move a set of pages as indicated in the pm array. The addr | |
794 | * field must be set to the virtual address of the page to be moved | |
795 | * and the node number must contain a valid target node. | |
796 | */ | |
797 | static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm, | |
798 | int migrate_all) | |
799 | { | |
800 | int err; | |
801 | struct page_to_node *pp; | |
802 | LIST_HEAD(pagelist); | |
803 | ||
804 | down_read(&mm->mmap_sem); | |
805 | ||
806 | /* | |
807 | * Build a list of pages to migrate | |
808 | */ | |
809 | migrate_prep(); | |
810 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { | |
811 | struct vm_area_struct *vma; | |
812 | struct page *page; | |
813 | ||
814 | /* | |
815 | * A valid page pointer that will not match any of the | |
816 | * pages that will be moved. | |
817 | */ | |
818 | pp->page = ZERO_PAGE(0); | |
819 | ||
820 | err = -EFAULT; | |
821 | vma = find_vma(mm, pp->addr); | |
0dc952dc | 822 | if (!vma || !vma_migratable(vma)) |
742755a1 CL |
823 | goto set_status; |
824 | ||
825 | page = follow_page(vma, pp->addr, FOLL_GET); | |
826 | err = -ENOENT; | |
827 | if (!page) | |
828 | goto set_status; | |
829 | ||
830 | if (PageReserved(page)) /* Check for zero page */ | |
831 | goto put_and_set; | |
832 | ||
833 | pp->page = page; | |
834 | err = page_to_nid(page); | |
835 | ||
836 | if (err == pp->node) | |
837 | /* | |
838 | * Node already in the right place | |
839 | */ | |
840 | goto put_and_set; | |
841 | ||
842 | err = -EACCES; | |
843 | if (page_mapcount(page) > 1 && | |
844 | !migrate_all) | |
845 | goto put_and_set; | |
846 | ||
847 | err = isolate_lru_page(page, &pagelist); | |
848 | put_and_set: | |
849 | /* | |
850 | * Either remove the duplicate refcount from | |
851 | * isolate_lru_page() or drop the page ref if it was | |
852 | * not isolated. | |
853 | */ | |
854 | put_page(page); | |
855 | set_status: | |
856 | pp->status = err; | |
857 | } | |
858 | ||
859 | if (!list_empty(&pagelist)) | |
860 | err = migrate_pages(&pagelist, new_page_node, | |
861 | (unsigned long)pm); | |
862 | else | |
863 | err = -ENOENT; | |
864 | ||
865 | up_read(&mm->mmap_sem); | |
866 | return err; | |
867 | } | |
868 | ||
869 | /* | |
870 | * Determine the nodes of a list of pages. The addr in the pm array | |
871 | * must have been set to the virtual address of which we want to determine | |
872 | * the node number. | |
873 | */ | |
874 | static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm) | |
875 | { | |
876 | down_read(&mm->mmap_sem); | |
877 | ||
878 | for ( ; pm->node != MAX_NUMNODES; pm++) { | |
879 | struct vm_area_struct *vma; | |
880 | struct page *page; | |
881 | int err; | |
882 | ||
883 | err = -EFAULT; | |
884 | vma = find_vma(mm, pm->addr); | |
885 | if (!vma) | |
886 | goto set_status; | |
887 | ||
888 | page = follow_page(vma, pm->addr, 0); | |
889 | err = -ENOENT; | |
890 | /* Use PageReserved to check for zero page */ | |
891 | if (!page || PageReserved(page)) | |
892 | goto set_status; | |
893 | ||
894 | err = page_to_nid(page); | |
895 | set_status: | |
896 | pm->status = err; | |
897 | } | |
898 | ||
899 | up_read(&mm->mmap_sem); | |
900 | return 0; | |
901 | } | |
902 | ||
903 | /* | |
904 | * Move a list of pages in the address space of the currently executing | |
905 | * process. | |
906 | */ | |
907 | asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages, | |
908 | const void __user * __user *pages, | |
909 | const int __user *nodes, | |
910 | int __user *status, int flags) | |
911 | { | |
912 | int err = 0; | |
913 | int i; | |
914 | struct task_struct *task; | |
915 | nodemask_t task_nodes; | |
916 | struct mm_struct *mm; | |
917 | struct page_to_node *pm = NULL; | |
918 | ||
919 | /* Check flags */ | |
920 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
921 | return -EINVAL; | |
922 | ||
923 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
924 | return -EPERM; | |
925 | ||
926 | /* Find the mm_struct */ | |
927 | read_lock(&tasklist_lock); | |
b488893a PE |
928 | task = pid ? |
929 | find_task_by_pid_ns(pid, current->nsproxy->pid_ns) : current; | |
742755a1 CL |
930 | if (!task) { |
931 | read_unlock(&tasklist_lock); | |
932 | return -ESRCH; | |
933 | } | |
934 | mm = get_task_mm(task); | |
935 | read_unlock(&tasklist_lock); | |
936 | ||
937 | if (!mm) | |
938 | return -EINVAL; | |
939 | ||
940 | /* | |
941 | * Check if this process has the right to modify the specified | |
942 | * process. The right exists if the process has administrative | |
943 | * capabilities, superuser privileges or the same | |
944 | * userid as the target process. | |
945 | */ | |
946 | if ((current->euid != task->suid) && (current->euid != task->uid) && | |
947 | (current->uid != task->suid) && (current->uid != task->uid) && | |
948 | !capable(CAP_SYS_NICE)) { | |
949 | err = -EPERM; | |
950 | goto out2; | |
951 | } | |
952 | ||
86c3a764 DQ |
953 | err = security_task_movememory(task); |
954 | if (err) | |
955 | goto out2; | |
956 | ||
957 | ||
742755a1 CL |
958 | task_nodes = cpuset_mems_allowed(task); |
959 | ||
960 | /* Limit nr_pages so that the multiplication may not overflow */ | |
961 | if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) { | |
962 | err = -E2BIG; | |
963 | goto out2; | |
964 | } | |
965 | ||
966 | pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node)); | |
967 | if (!pm) { | |
968 | err = -ENOMEM; | |
969 | goto out2; | |
970 | } | |
971 | ||
972 | /* | |
973 | * Get parameters from user space and initialize the pm | |
974 | * array. Return various errors if the user did something wrong. | |
975 | */ | |
976 | for (i = 0; i < nr_pages; i++) { | |
9d966d49 | 977 | const void __user *p; |
742755a1 CL |
978 | |
979 | err = -EFAULT; | |
980 | if (get_user(p, pages + i)) | |
981 | goto out; | |
982 | ||
983 | pm[i].addr = (unsigned long)p; | |
984 | if (nodes) { | |
985 | int node; | |
986 | ||
987 | if (get_user(node, nodes + i)) | |
988 | goto out; | |
989 | ||
990 | err = -ENODEV; | |
56bbd65d | 991 | if (!node_state(node, N_HIGH_MEMORY)) |
742755a1 CL |
992 | goto out; |
993 | ||
994 | err = -EACCES; | |
995 | if (!node_isset(node, task_nodes)) | |
996 | goto out; | |
997 | ||
998 | pm[i].node = node; | |
8ce08464 SR |
999 | } else |
1000 | pm[i].node = 0; /* anything to not match MAX_NUMNODES */ | |
742755a1 CL |
1001 | } |
1002 | /* End marker */ | |
1003 | pm[nr_pages].node = MAX_NUMNODES; | |
1004 | ||
1005 | if (nodes) | |
1006 | err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL); | |
1007 | else | |
1008 | err = do_pages_stat(mm, pm); | |
1009 | ||
1010 | if (err >= 0) | |
1011 | /* Return status information */ | |
1012 | for (i = 0; i < nr_pages; i++) | |
1013 | if (put_user(pm[i].status, status + i)) | |
1014 | err = -EFAULT; | |
1015 | ||
1016 | out: | |
1017 | vfree(pm); | |
1018 | out2: | |
1019 | mmput(mm); | |
1020 | return err; | |
1021 | } | |
1022 | #endif | |
1023 | ||
7b2259b3 CL |
1024 | /* |
1025 | * Call migration functions in the vma_ops that may prepare | |
1026 | * memory in a vm for migration. migration functions may perform | |
1027 | * the migration for vmas that do not have an underlying page struct. | |
1028 | */ | |
1029 | int migrate_vmas(struct mm_struct *mm, const nodemask_t *to, | |
1030 | const nodemask_t *from, unsigned long flags) | |
1031 | { | |
1032 | struct vm_area_struct *vma; | |
1033 | int err = 0; | |
1034 | ||
1035 | for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) { | |
1036 | if (vma->vm_ops && vma->vm_ops->migrate) { | |
1037 | err = vma->vm_ops->migrate(vma, to, from, flags); | |
1038 | if (err) | |
1039 | break; | |
1040 | } | |
1041 | } | |
1042 | return err; | |
1043 | } |