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 CL |
21 | #include <linux/mm_inline.h> |
22 | #include <linux/pagevec.h> | |
23 | #include <linux/rmap.h> | |
24 | #include <linux/topology.h> | |
25 | #include <linux/cpu.h> | |
26 | #include <linux/cpuset.h> | |
04e62a29 | 27 | #include <linux/writeback.h> |
742755a1 CL |
28 | #include <linux/mempolicy.h> |
29 | #include <linux/vmalloc.h> | |
b20a3503 CL |
30 | |
31 | #include "internal.h" | |
32 | ||
b20a3503 CL |
33 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) |
34 | ||
35 | /* | |
36 | * Isolate one page from the LRU lists. If successful put it onto | |
37 | * the indicated list with elevated page count. | |
38 | * | |
39 | * Result: | |
40 | * -EBUSY: page not on LRU list | |
41 | * 0: page removed from LRU list and added to the specified list. | |
42 | */ | |
43 | int isolate_lru_page(struct page *page, struct list_head *pagelist) | |
44 | { | |
45 | int ret = -EBUSY; | |
46 | ||
47 | if (PageLRU(page)) { | |
48 | struct zone *zone = page_zone(page); | |
49 | ||
50 | spin_lock_irq(&zone->lru_lock); | |
51 | if (PageLRU(page)) { | |
52 | ret = 0; | |
53 | get_page(page); | |
54 | ClearPageLRU(page); | |
55 | if (PageActive(page)) | |
56 | del_page_from_active_list(zone, page); | |
57 | else | |
58 | del_page_from_inactive_list(zone, page); | |
59 | list_add_tail(&page->lru, pagelist); | |
60 | } | |
61 | spin_unlock_irq(&zone->lru_lock); | |
62 | } | |
63 | return ret; | |
64 | } | |
65 | ||
66 | /* | |
742755a1 CL |
67 | * migrate_prep() needs to be called before we start compiling a list of pages |
68 | * to be migrated using isolate_lru_page(). | |
b20a3503 CL |
69 | */ |
70 | int migrate_prep(void) | |
71 | { | |
b20a3503 CL |
72 | /* |
73 | * Clear the LRU lists so pages can be isolated. | |
74 | * Note that pages may be moved off the LRU after we have | |
75 | * drained them. Those pages will fail to migrate like other | |
76 | * pages that may be busy. | |
77 | */ | |
78 | lru_add_drain_all(); | |
79 | ||
80 | return 0; | |
81 | } | |
82 | ||
83 | static inline void move_to_lru(struct page *page) | |
84 | { | |
b20a3503 CL |
85 | if (PageActive(page)) { |
86 | /* | |
87 | * lru_cache_add_active checks that | |
88 | * the PG_active bit is off. | |
89 | */ | |
90 | ClearPageActive(page); | |
91 | lru_cache_add_active(page); | |
92 | } else { | |
93 | lru_cache_add(page); | |
94 | } | |
95 | put_page(page); | |
96 | } | |
97 | ||
98 | /* | |
99 | * Add isolated pages on the list back to the LRU. | |
100 | * | |
101 | * returns the number of pages put back. | |
102 | */ | |
103 | int putback_lru_pages(struct list_head *l) | |
104 | { | |
105 | struct page *page; | |
106 | struct page *page2; | |
107 | int count = 0; | |
108 | ||
109 | list_for_each_entry_safe(page, page2, l, lru) { | |
e24f0b8f | 110 | list_del(&page->lru); |
b20a3503 CL |
111 | move_to_lru(page); |
112 | count++; | |
113 | } | |
114 | return count; | |
115 | } | |
116 | ||
0697212a CL |
117 | static inline int is_swap_pte(pte_t pte) |
118 | { | |
119 | return !pte_none(pte) && !pte_present(pte) && !pte_file(pte); | |
120 | } | |
121 | ||
122 | /* | |
123 | * Restore a potential migration pte to a working pte entry | |
124 | */ | |
04e62a29 | 125 | static void remove_migration_pte(struct vm_area_struct *vma, |
0697212a CL |
126 | struct page *old, struct page *new) |
127 | { | |
128 | struct mm_struct *mm = vma->vm_mm; | |
129 | swp_entry_t entry; | |
130 | pgd_t *pgd; | |
131 | pud_t *pud; | |
132 | pmd_t *pmd; | |
133 | pte_t *ptep, pte; | |
134 | spinlock_t *ptl; | |
04e62a29 CL |
135 | unsigned long addr = page_address_in_vma(new, vma); |
136 | ||
137 | if (addr == -EFAULT) | |
138 | return; | |
0697212a CL |
139 | |
140 | pgd = pgd_offset(mm, addr); | |
141 | if (!pgd_present(*pgd)) | |
142 | return; | |
143 | ||
144 | pud = pud_offset(pgd, addr); | |
145 | if (!pud_present(*pud)) | |
146 | return; | |
147 | ||
148 | pmd = pmd_offset(pud, addr); | |
149 | if (!pmd_present(*pmd)) | |
150 | return; | |
151 | ||
152 | ptep = pte_offset_map(pmd, addr); | |
153 | ||
154 | if (!is_swap_pte(*ptep)) { | |
155 | pte_unmap(ptep); | |
156 | return; | |
157 | } | |
158 | ||
159 | ptl = pte_lockptr(mm, pmd); | |
160 | spin_lock(ptl); | |
161 | pte = *ptep; | |
162 | if (!is_swap_pte(pte)) | |
163 | goto out; | |
164 | ||
165 | entry = pte_to_swp_entry(pte); | |
166 | ||
167 | if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old) | |
168 | goto out; | |
169 | ||
0697212a CL |
170 | get_page(new); |
171 | pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); | |
172 | if (is_write_migration_entry(entry)) | |
173 | pte = pte_mkwrite(pte); | |
174 | set_pte_at(mm, addr, ptep, pte); | |
04e62a29 CL |
175 | |
176 | if (PageAnon(new)) | |
177 | page_add_anon_rmap(new, vma, addr); | |
178 | else | |
179 | page_add_file_rmap(new); | |
180 | ||
181 | /* No need to invalidate - it was non-present before */ | |
182 | update_mmu_cache(vma, addr, pte); | |
183 | lazy_mmu_prot_update(pte); | |
184 | ||
0697212a CL |
185 | out: |
186 | pte_unmap_unlock(ptep, ptl); | |
187 | } | |
188 | ||
189 | /* | |
04e62a29 CL |
190 | * Note that remove_file_migration_ptes will only work on regular mappings, |
191 | * Nonlinear mappings do not use migration entries. | |
192 | */ | |
193 | static void remove_file_migration_ptes(struct page *old, struct page *new) | |
194 | { | |
195 | struct vm_area_struct *vma; | |
196 | struct address_space *mapping = page_mapping(new); | |
197 | struct prio_tree_iter iter; | |
198 | pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
199 | ||
200 | if (!mapping) | |
201 | return; | |
202 | ||
203 | spin_lock(&mapping->i_mmap_lock); | |
204 | ||
205 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) | |
206 | remove_migration_pte(vma, old, new); | |
207 | ||
208 | spin_unlock(&mapping->i_mmap_lock); | |
209 | } | |
210 | ||
211 | /* | |
0697212a CL |
212 | * Must hold mmap_sem lock on at least one of the vmas containing |
213 | * the page so that the anon_vma cannot vanish. | |
214 | */ | |
04e62a29 | 215 | static void remove_anon_migration_ptes(struct page *old, struct page *new) |
0697212a CL |
216 | { |
217 | struct anon_vma *anon_vma; | |
218 | struct vm_area_struct *vma; | |
219 | unsigned long mapping; | |
220 | ||
221 | mapping = (unsigned long)new->mapping; | |
222 | ||
223 | if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0) | |
224 | return; | |
225 | ||
226 | /* | |
227 | * We hold the mmap_sem lock. So no need to call page_lock_anon_vma. | |
228 | */ | |
229 | anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON); | |
230 | spin_lock(&anon_vma->lock); | |
231 | ||
232 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) | |
04e62a29 | 233 | remove_migration_pte(vma, old, new); |
0697212a CL |
234 | |
235 | spin_unlock(&anon_vma->lock); | |
236 | } | |
237 | ||
04e62a29 CL |
238 | /* |
239 | * Get rid of all migration entries and replace them by | |
240 | * references to the indicated page. | |
241 | */ | |
242 | static void remove_migration_ptes(struct page *old, struct page *new) | |
243 | { | |
244 | if (PageAnon(new)) | |
245 | remove_anon_migration_ptes(old, new); | |
246 | else | |
247 | remove_file_migration_ptes(old, new); | |
248 | } | |
249 | ||
0697212a CL |
250 | /* |
251 | * Something used the pte of a page under migration. We need to | |
252 | * get to the page and wait until migration is finished. | |
253 | * When we return from this function the fault will be retried. | |
254 | * | |
255 | * This function is called from do_swap_page(). | |
256 | */ | |
257 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, | |
258 | unsigned long address) | |
259 | { | |
260 | pte_t *ptep, pte; | |
261 | spinlock_t *ptl; | |
262 | swp_entry_t entry; | |
263 | struct page *page; | |
264 | ||
265 | ptep = pte_offset_map_lock(mm, pmd, address, &ptl); | |
266 | pte = *ptep; | |
267 | if (!is_swap_pte(pte)) | |
268 | goto out; | |
269 | ||
270 | entry = pte_to_swp_entry(pte); | |
271 | if (!is_migration_entry(entry)) | |
272 | goto out; | |
273 | ||
274 | page = migration_entry_to_page(entry); | |
275 | ||
276 | get_page(page); | |
277 | pte_unmap_unlock(ptep, ptl); | |
278 | wait_on_page_locked(page); | |
279 | put_page(page); | |
280 | return; | |
281 | out: | |
282 | pte_unmap_unlock(ptep, ptl); | |
283 | } | |
284 | ||
b20a3503 | 285 | /* |
c3fcf8a5 | 286 | * Replace the page in the mapping. |
5b5c7120 CL |
287 | * |
288 | * The number of remaining references must be: | |
289 | * 1 for anonymous pages without a mapping | |
290 | * 2 for pages with a mapping | |
291 | * 3 for pages with a mapping and PagePrivate set. | |
b20a3503 | 292 | */ |
2d1db3b1 CL |
293 | static int migrate_page_move_mapping(struct address_space *mapping, |
294 | struct page *newpage, struct page *page) | |
b20a3503 | 295 | { |
b20a3503 CL |
296 | struct page **radix_pointer; |
297 | ||
6c5240ae CL |
298 | if (!mapping) { |
299 | /* Anonymous page */ | |
300 | if (page_count(page) != 1) | |
301 | return -EAGAIN; | |
302 | return 0; | |
303 | } | |
304 | ||
b20a3503 CL |
305 | write_lock_irq(&mapping->tree_lock); |
306 | ||
307 | radix_pointer = (struct page **)radix_tree_lookup_slot( | |
308 | &mapping->page_tree, | |
309 | page_index(page)); | |
310 | ||
6c5240ae | 311 | if (page_count(page) != 2 + !!PagePrivate(page) || |
b20a3503 CL |
312 | *radix_pointer != page) { |
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 CL |
319 | */ |
320 | get_page(newpage); | |
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 CL |
327 | |
328 | *radix_pointer = newpage; | |
329 | __put_page(page); | |
330 | write_unlock_irq(&mapping->tree_lock); | |
331 | ||
332 | return 0; | |
333 | } | |
b20a3503 CL |
334 | |
335 | /* | |
336 | * Copy the page to its new location | |
337 | */ | |
e7340f73 | 338 | static void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 CL |
339 | { |
340 | copy_highpage(newpage, page); | |
341 | ||
342 | if (PageError(page)) | |
343 | SetPageError(newpage); | |
344 | if (PageReferenced(page)) | |
345 | SetPageReferenced(newpage); | |
346 | if (PageUptodate(page)) | |
347 | SetPageUptodate(newpage); | |
348 | if (PageActive(page)) | |
349 | SetPageActive(newpage); | |
350 | if (PageChecked(page)) | |
351 | SetPageChecked(newpage); | |
352 | if (PageMappedToDisk(page)) | |
353 | SetPageMappedToDisk(newpage); | |
354 | ||
355 | if (PageDirty(page)) { | |
356 | clear_page_dirty_for_io(page); | |
357 | set_page_dirty(newpage); | |
358 | } | |
359 | ||
6c5240ae | 360 | #ifdef CONFIG_SWAP |
b20a3503 | 361 | ClearPageSwapCache(page); |
6c5240ae | 362 | #endif |
b20a3503 CL |
363 | ClearPageActive(page); |
364 | ClearPagePrivate(page); | |
365 | set_page_private(page, 0); | |
366 | page->mapping = NULL; | |
367 | ||
368 | /* | |
369 | * If any waiters have accumulated on the new page then | |
370 | * wake them up. | |
371 | */ | |
372 | if (PageWriteback(newpage)) | |
373 | end_page_writeback(newpage); | |
374 | } | |
b20a3503 | 375 | |
1d8b85cc CL |
376 | /************************************************************ |
377 | * Migration functions | |
378 | ***********************************************************/ | |
379 | ||
380 | /* Always fail migration. Used for mappings that are not movable */ | |
2d1db3b1 CL |
381 | int fail_migrate_page(struct address_space *mapping, |
382 | struct page *newpage, struct page *page) | |
1d8b85cc CL |
383 | { |
384 | return -EIO; | |
385 | } | |
386 | EXPORT_SYMBOL(fail_migrate_page); | |
387 | ||
b20a3503 CL |
388 | /* |
389 | * Common logic to directly migrate a single page suitable for | |
390 | * pages that do not use PagePrivate. | |
391 | * | |
392 | * Pages are locked upon entry and exit. | |
393 | */ | |
2d1db3b1 CL |
394 | int migrate_page(struct address_space *mapping, |
395 | struct page *newpage, struct page *page) | |
b20a3503 CL |
396 | { |
397 | int rc; | |
398 | ||
399 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
400 | ||
2d1db3b1 | 401 | rc = migrate_page_move_mapping(mapping, newpage, page); |
b20a3503 CL |
402 | |
403 | if (rc) | |
404 | return rc; | |
405 | ||
406 | migrate_page_copy(newpage, page); | |
b20a3503 CL |
407 | return 0; |
408 | } | |
409 | EXPORT_SYMBOL(migrate_page); | |
410 | ||
1d8b85cc CL |
411 | /* |
412 | * Migration function for pages with buffers. This function can only be used | |
413 | * if the underlying filesystem guarantees that no other references to "page" | |
414 | * exist. | |
415 | */ | |
2d1db3b1 CL |
416 | int buffer_migrate_page(struct address_space *mapping, |
417 | struct page *newpage, struct page *page) | |
1d8b85cc | 418 | { |
1d8b85cc CL |
419 | struct buffer_head *bh, *head; |
420 | int rc; | |
421 | ||
1d8b85cc | 422 | if (!page_has_buffers(page)) |
2d1db3b1 | 423 | return migrate_page(mapping, newpage, page); |
1d8b85cc CL |
424 | |
425 | head = page_buffers(page); | |
426 | ||
2d1db3b1 | 427 | rc = migrate_page_move_mapping(mapping, newpage, page); |
1d8b85cc CL |
428 | |
429 | if (rc) | |
430 | return rc; | |
431 | ||
432 | bh = head; | |
433 | do { | |
434 | get_bh(bh); | |
435 | lock_buffer(bh); | |
436 | bh = bh->b_this_page; | |
437 | ||
438 | } while (bh != head); | |
439 | ||
440 | ClearPagePrivate(page); | |
441 | set_page_private(newpage, page_private(page)); | |
442 | set_page_private(page, 0); | |
443 | put_page(page); | |
444 | get_page(newpage); | |
445 | ||
446 | bh = head; | |
447 | do { | |
448 | set_bh_page(bh, newpage, bh_offset(bh)); | |
449 | bh = bh->b_this_page; | |
450 | ||
451 | } while (bh != head); | |
452 | ||
453 | SetPagePrivate(newpage); | |
454 | ||
455 | migrate_page_copy(newpage, page); | |
456 | ||
457 | bh = head; | |
458 | do { | |
459 | unlock_buffer(bh); | |
460 | put_bh(bh); | |
461 | bh = bh->b_this_page; | |
462 | ||
463 | } while (bh != head); | |
464 | ||
465 | return 0; | |
466 | } | |
467 | EXPORT_SYMBOL(buffer_migrate_page); | |
468 | ||
04e62a29 CL |
469 | /* |
470 | * Writeback a page to clean the dirty state | |
471 | */ | |
472 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 473 | { |
04e62a29 CL |
474 | struct writeback_control wbc = { |
475 | .sync_mode = WB_SYNC_NONE, | |
476 | .nr_to_write = 1, | |
477 | .range_start = 0, | |
478 | .range_end = LLONG_MAX, | |
479 | .nonblocking = 1, | |
480 | .for_reclaim = 1 | |
481 | }; | |
482 | int rc; | |
483 | ||
484 | if (!mapping->a_ops->writepage) | |
485 | /* No write method for the address space */ | |
486 | return -EINVAL; | |
487 | ||
488 | if (!clear_page_dirty_for_io(page)) | |
489 | /* Someone else already triggered a write */ | |
490 | return -EAGAIN; | |
491 | ||
8351a6e4 | 492 | /* |
04e62a29 CL |
493 | * A dirty page may imply that the underlying filesystem has |
494 | * the page on some queue. So the page must be clean for | |
495 | * migration. Writeout may mean we loose the lock and the | |
496 | * page state is no longer what we checked for earlier. | |
497 | * At this point we know that the migration attempt cannot | |
498 | * be successful. | |
8351a6e4 | 499 | */ |
04e62a29 | 500 | remove_migration_ptes(page, page); |
8351a6e4 | 501 | |
04e62a29 CL |
502 | rc = mapping->a_ops->writepage(page, &wbc); |
503 | if (rc < 0) | |
504 | /* I/O Error writing */ | |
505 | return -EIO; | |
8351a6e4 | 506 | |
04e62a29 CL |
507 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
508 | /* unlocked. Relock */ | |
509 | lock_page(page); | |
510 | ||
511 | return -EAGAIN; | |
512 | } | |
513 | ||
514 | /* | |
515 | * Default handling if a filesystem does not provide a migration function. | |
516 | */ | |
517 | static int fallback_migrate_page(struct address_space *mapping, | |
518 | struct page *newpage, struct page *page) | |
519 | { | |
520 | if (PageDirty(page)) | |
521 | return writeout(mapping, page); | |
8351a6e4 CL |
522 | |
523 | /* | |
524 | * Buffers may be managed in a filesystem specific way. | |
525 | * We must have no buffers or drop them. | |
526 | */ | |
527 | if (page_has_buffers(page) && | |
528 | !try_to_release_page(page, GFP_KERNEL)) | |
529 | return -EAGAIN; | |
530 | ||
531 | return migrate_page(mapping, newpage, page); | |
532 | } | |
533 | ||
e24f0b8f CL |
534 | /* |
535 | * Move a page to a newly allocated page | |
536 | * The page is locked and all ptes have been successfully removed. | |
537 | * | |
538 | * The new page will have replaced the old page if this function | |
539 | * is successful. | |
540 | */ | |
541 | static int move_to_new_page(struct page *newpage, struct page *page) | |
542 | { | |
543 | struct address_space *mapping; | |
544 | int rc; | |
545 | ||
546 | /* | |
547 | * Block others from accessing the page when we get around to | |
548 | * establishing additional references. We are the only one | |
549 | * holding a reference to the new page at this point. | |
550 | */ | |
551 | if (TestSetPageLocked(newpage)) | |
552 | BUG(); | |
553 | ||
554 | /* Prepare mapping for the new page.*/ | |
555 | newpage->index = page->index; | |
556 | newpage->mapping = page->mapping; | |
557 | ||
558 | mapping = page_mapping(page); | |
559 | if (!mapping) | |
560 | rc = migrate_page(mapping, newpage, page); | |
561 | else if (mapping->a_ops->migratepage) | |
562 | /* | |
563 | * Most pages have a mapping and most filesystems | |
564 | * should provide a migration function. Anonymous | |
565 | * pages are part of swap space which also has its | |
566 | * own migration function. This is the most common | |
567 | * path for page migration. | |
568 | */ | |
569 | rc = mapping->a_ops->migratepage(mapping, | |
570 | newpage, page); | |
571 | else | |
572 | rc = fallback_migrate_page(mapping, newpage, page); | |
573 | ||
574 | if (!rc) | |
575 | remove_migration_ptes(page, newpage); | |
576 | else | |
577 | newpage->mapping = NULL; | |
578 | ||
579 | unlock_page(newpage); | |
580 | ||
581 | return rc; | |
582 | } | |
583 | ||
584 | /* | |
585 | * Obtain the lock on page, remove all ptes and migrate the page | |
586 | * to the newly allocated page in newpage. | |
587 | */ | |
95a402c3 CL |
588 | static int unmap_and_move(new_page_t get_new_page, unsigned long private, |
589 | struct page *page, int force) | |
e24f0b8f CL |
590 | { |
591 | int rc = 0; | |
742755a1 CL |
592 | int *result = NULL; |
593 | struct page *newpage = get_new_page(page, private, &result); | |
95a402c3 CL |
594 | |
595 | if (!newpage) | |
596 | return -ENOMEM; | |
e24f0b8f CL |
597 | |
598 | if (page_count(page) == 1) | |
599 | /* page was freed from under us. So we are done. */ | |
95a402c3 | 600 | goto move_newpage; |
e24f0b8f CL |
601 | |
602 | rc = -EAGAIN; | |
603 | if (TestSetPageLocked(page)) { | |
604 | if (!force) | |
95a402c3 | 605 | goto move_newpage; |
e24f0b8f CL |
606 | lock_page(page); |
607 | } | |
608 | ||
609 | if (PageWriteback(page)) { | |
610 | if (!force) | |
611 | goto unlock; | |
612 | wait_on_page_writeback(page); | |
613 | } | |
614 | ||
615 | /* | |
616 | * Establish migration ptes or remove ptes | |
617 | */ | |
618 | if (try_to_unmap(page, 1) != SWAP_FAIL) { | |
619 | if (!page_mapped(page)) | |
620 | rc = move_to_new_page(newpage, page); | |
621 | } else | |
622 | /* A vma has VM_LOCKED set -> permanent failure */ | |
623 | rc = -EPERM; | |
624 | ||
625 | if (rc) | |
626 | remove_migration_ptes(page, page); | |
627 | unlock: | |
628 | unlock_page(page); | |
95a402c3 | 629 | |
e24f0b8f | 630 | if (rc != -EAGAIN) { |
aaa994b3 CL |
631 | /* |
632 | * A page that has been migrated has all references | |
633 | * removed and will be freed. A page that has not been | |
634 | * migrated will have kepts its references and be | |
635 | * restored. | |
636 | */ | |
637 | list_del(&page->lru); | |
638 | move_to_lru(page); | |
e24f0b8f | 639 | } |
95a402c3 CL |
640 | |
641 | move_newpage: | |
642 | /* | |
643 | * Move the new page to the LRU. If migration was not successful | |
644 | * then this will free the page. | |
645 | */ | |
646 | move_to_lru(newpage); | |
742755a1 CL |
647 | if (result) { |
648 | if (rc) | |
649 | *result = rc; | |
650 | else | |
651 | *result = page_to_nid(newpage); | |
652 | } | |
e24f0b8f CL |
653 | return rc; |
654 | } | |
655 | ||
b20a3503 CL |
656 | /* |
657 | * migrate_pages | |
658 | * | |
95a402c3 CL |
659 | * The function takes one list of pages to migrate and a function |
660 | * that determines from the page to be migrated and the private data | |
661 | * the target of the move and allocates the page. | |
b20a3503 CL |
662 | * |
663 | * The function returns after 10 attempts or if no pages | |
664 | * are movable anymore because to has become empty | |
aaa994b3 CL |
665 | * or no retryable pages exist anymore. All pages will be |
666 | * retruned to the LRU or freed. | |
b20a3503 | 667 | * |
95a402c3 | 668 | * Return: Number of pages not migrated or error code. |
b20a3503 | 669 | */ |
95a402c3 CL |
670 | int migrate_pages(struct list_head *from, |
671 | new_page_t get_new_page, unsigned long private) | |
b20a3503 | 672 | { |
e24f0b8f | 673 | int retry = 1; |
b20a3503 CL |
674 | int nr_failed = 0; |
675 | int pass = 0; | |
676 | struct page *page; | |
677 | struct page *page2; | |
678 | int swapwrite = current->flags & PF_SWAPWRITE; | |
679 | int rc; | |
680 | ||
681 | if (!swapwrite) | |
682 | current->flags |= PF_SWAPWRITE; | |
683 | ||
e24f0b8f CL |
684 | for(pass = 0; pass < 10 && retry; pass++) { |
685 | retry = 0; | |
b20a3503 | 686 | |
e24f0b8f | 687 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 688 | cond_resched(); |
2d1db3b1 | 689 | |
95a402c3 CL |
690 | rc = unmap_and_move(get_new_page, private, |
691 | page, pass > 2); | |
2d1db3b1 | 692 | |
e24f0b8f | 693 | switch(rc) { |
95a402c3 CL |
694 | case -ENOMEM: |
695 | goto out; | |
e24f0b8f | 696 | case -EAGAIN: |
2d1db3b1 | 697 | retry++; |
e24f0b8f CL |
698 | break; |
699 | case 0: | |
e24f0b8f CL |
700 | break; |
701 | default: | |
2d1db3b1 | 702 | /* Permanent failure */ |
2d1db3b1 | 703 | nr_failed++; |
e24f0b8f | 704 | break; |
2d1db3b1 | 705 | } |
b20a3503 CL |
706 | } |
707 | } | |
95a402c3 CL |
708 | rc = 0; |
709 | out: | |
b20a3503 CL |
710 | if (!swapwrite) |
711 | current->flags &= ~PF_SWAPWRITE; | |
712 | ||
aaa994b3 | 713 | putback_lru_pages(from); |
b20a3503 | 714 | |
95a402c3 CL |
715 | if (rc) |
716 | return rc; | |
b20a3503 | 717 | |
95a402c3 | 718 | return nr_failed + retry; |
b20a3503 | 719 | } |
95a402c3 | 720 | |
742755a1 CL |
721 | #ifdef CONFIG_NUMA |
722 | /* | |
723 | * Move a list of individual pages | |
724 | */ | |
725 | struct page_to_node { | |
726 | unsigned long addr; | |
727 | struct page *page; | |
728 | int node; | |
729 | int status; | |
730 | }; | |
731 | ||
732 | static struct page *new_page_node(struct page *p, unsigned long private, | |
733 | int **result) | |
734 | { | |
735 | struct page_to_node *pm = (struct page_to_node *)private; | |
736 | ||
737 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
738 | pm++; | |
739 | ||
740 | if (pm->node == MAX_NUMNODES) | |
741 | return NULL; | |
742 | ||
743 | *result = &pm->status; | |
744 | ||
745 | return alloc_pages_node(pm->node, GFP_HIGHUSER, 0); | |
746 | } | |
747 | ||
748 | /* | |
749 | * Move a set of pages as indicated in the pm array. The addr | |
750 | * field must be set to the virtual address of the page to be moved | |
751 | * and the node number must contain a valid target node. | |
752 | */ | |
753 | static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm, | |
754 | int migrate_all) | |
755 | { | |
756 | int err; | |
757 | struct page_to_node *pp; | |
758 | LIST_HEAD(pagelist); | |
759 | ||
760 | down_read(&mm->mmap_sem); | |
761 | ||
762 | /* | |
763 | * Build a list of pages to migrate | |
764 | */ | |
765 | migrate_prep(); | |
766 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { | |
767 | struct vm_area_struct *vma; | |
768 | struct page *page; | |
769 | ||
770 | /* | |
771 | * A valid page pointer that will not match any of the | |
772 | * pages that will be moved. | |
773 | */ | |
774 | pp->page = ZERO_PAGE(0); | |
775 | ||
776 | err = -EFAULT; | |
777 | vma = find_vma(mm, pp->addr); | |
778 | if (!vma) | |
779 | goto set_status; | |
780 | ||
781 | page = follow_page(vma, pp->addr, FOLL_GET); | |
782 | err = -ENOENT; | |
783 | if (!page) | |
784 | goto set_status; | |
785 | ||
786 | if (PageReserved(page)) /* Check for zero page */ | |
787 | goto put_and_set; | |
788 | ||
789 | pp->page = page; | |
790 | err = page_to_nid(page); | |
791 | ||
792 | if (err == pp->node) | |
793 | /* | |
794 | * Node already in the right place | |
795 | */ | |
796 | goto put_and_set; | |
797 | ||
798 | err = -EACCES; | |
799 | if (page_mapcount(page) > 1 && | |
800 | !migrate_all) | |
801 | goto put_and_set; | |
802 | ||
803 | err = isolate_lru_page(page, &pagelist); | |
804 | put_and_set: | |
805 | /* | |
806 | * Either remove the duplicate refcount from | |
807 | * isolate_lru_page() or drop the page ref if it was | |
808 | * not isolated. | |
809 | */ | |
810 | put_page(page); | |
811 | set_status: | |
812 | pp->status = err; | |
813 | } | |
814 | ||
815 | if (!list_empty(&pagelist)) | |
816 | err = migrate_pages(&pagelist, new_page_node, | |
817 | (unsigned long)pm); | |
818 | else | |
819 | err = -ENOENT; | |
820 | ||
821 | up_read(&mm->mmap_sem); | |
822 | return err; | |
823 | } | |
824 | ||
825 | /* | |
826 | * Determine the nodes of a list of pages. The addr in the pm array | |
827 | * must have been set to the virtual address of which we want to determine | |
828 | * the node number. | |
829 | */ | |
830 | static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm) | |
831 | { | |
832 | down_read(&mm->mmap_sem); | |
833 | ||
834 | for ( ; pm->node != MAX_NUMNODES; pm++) { | |
835 | struct vm_area_struct *vma; | |
836 | struct page *page; | |
837 | int err; | |
838 | ||
839 | err = -EFAULT; | |
840 | vma = find_vma(mm, pm->addr); | |
841 | if (!vma) | |
842 | goto set_status; | |
843 | ||
844 | page = follow_page(vma, pm->addr, 0); | |
845 | err = -ENOENT; | |
846 | /* Use PageReserved to check for zero page */ | |
847 | if (!page || PageReserved(page)) | |
848 | goto set_status; | |
849 | ||
850 | err = page_to_nid(page); | |
851 | set_status: | |
852 | pm->status = err; | |
853 | } | |
854 | ||
855 | up_read(&mm->mmap_sem); | |
856 | return 0; | |
857 | } | |
858 | ||
859 | /* | |
860 | * Move a list of pages in the address space of the currently executing | |
861 | * process. | |
862 | */ | |
863 | asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages, | |
864 | const void __user * __user *pages, | |
865 | const int __user *nodes, | |
866 | int __user *status, int flags) | |
867 | { | |
868 | int err = 0; | |
869 | int i; | |
870 | struct task_struct *task; | |
871 | nodemask_t task_nodes; | |
872 | struct mm_struct *mm; | |
873 | struct page_to_node *pm = NULL; | |
874 | ||
875 | /* Check flags */ | |
876 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
877 | return -EINVAL; | |
878 | ||
879 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
880 | return -EPERM; | |
881 | ||
882 | /* Find the mm_struct */ | |
883 | read_lock(&tasklist_lock); | |
884 | task = pid ? find_task_by_pid(pid) : current; | |
885 | if (!task) { | |
886 | read_unlock(&tasklist_lock); | |
887 | return -ESRCH; | |
888 | } | |
889 | mm = get_task_mm(task); | |
890 | read_unlock(&tasklist_lock); | |
891 | ||
892 | if (!mm) | |
893 | return -EINVAL; | |
894 | ||
895 | /* | |
896 | * Check if this process has the right to modify the specified | |
897 | * process. The right exists if the process has administrative | |
898 | * capabilities, superuser privileges or the same | |
899 | * userid as the target process. | |
900 | */ | |
901 | if ((current->euid != task->suid) && (current->euid != task->uid) && | |
902 | (current->uid != task->suid) && (current->uid != task->uid) && | |
903 | !capable(CAP_SYS_NICE)) { | |
904 | err = -EPERM; | |
905 | goto out2; | |
906 | } | |
907 | ||
908 | task_nodes = cpuset_mems_allowed(task); | |
909 | ||
910 | /* Limit nr_pages so that the multiplication may not overflow */ | |
911 | if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) { | |
912 | err = -E2BIG; | |
913 | goto out2; | |
914 | } | |
915 | ||
916 | pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node)); | |
917 | if (!pm) { | |
918 | err = -ENOMEM; | |
919 | goto out2; | |
920 | } | |
921 | ||
922 | /* | |
923 | * Get parameters from user space and initialize the pm | |
924 | * array. Return various errors if the user did something wrong. | |
925 | */ | |
926 | for (i = 0; i < nr_pages; i++) { | |
927 | const void *p; | |
928 | ||
929 | err = -EFAULT; | |
930 | if (get_user(p, pages + i)) | |
931 | goto out; | |
932 | ||
933 | pm[i].addr = (unsigned long)p; | |
934 | if (nodes) { | |
935 | int node; | |
936 | ||
937 | if (get_user(node, nodes + i)) | |
938 | goto out; | |
939 | ||
940 | err = -ENODEV; | |
941 | if (!node_online(node)) | |
942 | goto out; | |
943 | ||
944 | err = -EACCES; | |
945 | if (!node_isset(node, task_nodes)) | |
946 | goto out; | |
947 | ||
948 | pm[i].node = node; | |
949 | } | |
950 | } | |
951 | /* End marker */ | |
952 | pm[nr_pages].node = MAX_NUMNODES; | |
953 | ||
954 | if (nodes) | |
955 | err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL); | |
956 | else | |
957 | err = do_pages_stat(mm, pm); | |
958 | ||
959 | if (err >= 0) | |
960 | /* Return status information */ | |
961 | for (i = 0; i < nr_pages; i++) | |
962 | if (put_user(pm[i].status, status + i)) | |
963 | err = -EFAULT; | |
964 | ||
965 | out: | |
966 | vfree(pm); | |
967 | out2: | |
968 | mmput(mm); | |
969 | return err; | |
970 | } | |
971 | #endif | |
972 |