]>
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 | ||
b20a3503 | 39 | /* |
742755a1 CL |
40 | * migrate_prep() needs to be called before we start compiling a list of pages |
41 | * to be migrated using isolate_lru_page(). | |
b20a3503 CL |
42 | */ |
43 | int migrate_prep(void) | |
44 | { | |
b20a3503 CL |
45 | /* |
46 | * Clear the LRU lists so pages can be isolated. | |
47 | * Note that pages may be moved off the LRU after we have | |
48 | * drained them. Those pages will fail to migrate like other | |
49 | * pages that may be busy. | |
50 | */ | |
51 | lru_add_drain_all(); | |
52 | ||
53 | return 0; | |
54 | } | |
55 | ||
56 | static inline void move_to_lru(struct page *page) | |
57 | { | |
f04e9ebb | 58 | lru_cache_add_lru(page, page_lru(page)); |
b20a3503 CL |
59 | put_page(page); |
60 | } | |
61 | ||
62 | /* | |
63 | * Add isolated pages on the list back to the LRU. | |
64 | * | |
65 | * returns the number of pages put back. | |
66 | */ | |
67 | int putback_lru_pages(struct list_head *l) | |
68 | { | |
69 | struct page *page; | |
70 | struct page *page2; | |
71 | int count = 0; | |
72 | ||
73 | list_for_each_entry_safe(page, page2, l, lru) { | |
e24f0b8f | 74 | list_del(&page->lru); |
b20a3503 CL |
75 | move_to_lru(page); |
76 | count++; | |
77 | } | |
78 | return count; | |
79 | } | |
80 | ||
0697212a CL |
81 | /* |
82 | * Restore a potential migration pte to a working pte entry | |
83 | */ | |
04e62a29 | 84 | static void remove_migration_pte(struct vm_area_struct *vma, |
0697212a CL |
85 | struct page *old, struct page *new) |
86 | { | |
87 | struct mm_struct *mm = vma->vm_mm; | |
88 | swp_entry_t entry; | |
89 | pgd_t *pgd; | |
90 | pud_t *pud; | |
91 | pmd_t *pmd; | |
92 | pte_t *ptep, pte; | |
93 | spinlock_t *ptl; | |
04e62a29 CL |
94 | unsigned long addr = page_address_in_vma(new, vma); |
95 | ||
96 | if (addr == -EFAULT) | |
97 | return; | |
0697212a CL |
98 | |
99 | pgd = pgd_offset(mm, addr); | |
100 | if (!pgd_present(*pgd)) | |
101 | return; | |
102 | ||
103 | pud = pud_offset(pgd, addr); | |
104 | if (!pud_present(*pud)) | |
105 | return; | |
106 | ||
107 | pmd = pmd_offset(pud, addr); | |
108 | if (!pmd_present(*pmd)) | |
109 | return; | |
110 | ||
111 | ptep = pte_offset_map(pmd, addr); | |
112 | ||
113 | if (!is_swap_pte(*ptep)) { | |
114 | pte_unmap(ptep); | |
115 | return; | |
116 | } | |
117 | ||
118 | ptl = pte_lockptr(mm, pmd); | |
119 | spin_lock(ptl); | |
120 | pte = *ptep; | |
121 | if (!is_swap_pte(pte)) | |
122 | goto out; | |
123 | ||
124 | entry = pte_to_swp_entry(pte); | |
125 | ||
126 | if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old) | |
127 | goto out; | |
128 | ||
98837c7f HD |
129 | /* |
130 | * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge. | |
131 | * Failure is not an option here: we're now expected to remove every | |
132 | * migration pte, and will cause crashes otherwise. Normally this | |
133 | * is not an issue: mem_cgroup_prepare_migration bumped up the old | |
134 | * page_cgroup count for safety, that's now attached to the new page, | |
135 | * so this charge should just be another incrementation of the count, | |
136 | * to keep in balance with rmap.c's mem_cgroup_uncharging. But if | |
137 | * there's been a force_empty, those reference counts may no longer | |
138 | * be reliable, and this charge can actually fail: oh well, we don't | |
139 | * make the situation any worse by proceeding as if it had succeeded. | |
140 | */ | |
141 | mem_cgroup_charge(new, mm, GFP_ATOMIC); | |
142 | ||
0697212a CL |
143 | get_page(new); |
144 | pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); | |
145 | if (is_write_migration_entry(entry)) | |
146 | pte = pte_mkwrite(pte); | |
97ee0524 | 147 | flush_cache_page(vma, addr, pte_pfn(pte)); |
0697212a | 148 | set_pte_at(mm, addr, ptep, pte); |
04e62a29 CL |
149 | |
150 | if (PageAnon(new)) | |
151 | page_add_anon_rmap(new, vma, addr); | |
152 | else | |
153 | page_add_file_rmap(new); | |
154 | ||
155 | /* No need to invalidate - it was non-present before */ | |
156 | update_mmu_cache(vma, addr, pte); | |
04e62a29 | 157 | |
0697212a CL |
158 | out: |
159 | pte_unmap_unlock(ptep, ptl); | |
160 | } | |
161 | ||
162 | /* | |
04e62a29 CL |
163 | * Note that remove_file_migration_ptes will only work on regular mappings, |
164 | * Nonlinear mappings do not use migration entries. | |
165 | */ | |
166 | static void remove_file_migration_ptes(struct page *old, struct page *new) | |
167 | { | |
168 | struct vm_area_struct *vma; | |
169 | struct address_space *mapping = page_mapping(new); | |
170 | struct prio_tree_iter iter; | |
171 | pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); | |
172 | ||
173 | if (!mapping) | |
174 | return; | |
175 | ||
176 | spin_lock(&mapping->i_mmap_lock); | |
177 | ||
178 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) | |
179 | remove_migration_pte(vma, old, new); | |
180 | ||
181 | spin_unlock(&mapping->i_mmap_lock); | |
182 | } | |
183 | ||
184 | /* | |
0697212a CL |
185 | * Must hold mmap_sem lock on at least one of the vmas containing |
186 | * the page so that the anon_vma cannot vanish. | |
187 | */ | |
04e62a29 | 188 | static void remove_anon_migration_ptes(struct page *old, struct page *new) |
0697212a CL |
189 | { |
190 | struct anon_vma *anon_vma; | |
191 | struct vm_area_struct *vma; | |
192 | unsigned long mapping; | |
193 | ||
194 | mapping = (unsigned long)new->mapping; | |
195 | ||
196 | if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0) | |
197 | return; | |
198 | ||
199 | /* | |
200 | * We hold the mmap_sem lock. So no need to call page_lock_anon_vma. | |
201 | */ | |
202 | anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON); | |
203 | spin_lock(&anon_vma->lock); | |
204 | ||
205 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) | |
04e62a29 | 206 | remove_migration_pte(vma, old, new); |
0697212a CL |
207 | |
208 | spin_unlock(&anon_vma->lock); | |
209 | } | |
210 | ||
04e62a29 CL |
211 | /* |
212 | * Get rid of all migration entries and replace them by | |
213 | * references to the indicated page. | |
214 | */ | |
215 | static void remove_migration_ptes(struct page *old, struct page *new) | |
216 | { | |
217 | if (PageAnon(new)) | |
218 | remove_anon_migration_ptes(old, new); | |
219 | else | |
220 | remove_file_migration_ptes(old, new); | |
221 | } | |
222 | ||
0697212a CL |
223 | /* |
224 | * Something used the pte of a page under migration. We need to | |
225 | * get to the page and wait until migration is finished. | |
226 | * When we return from this function the fault will be retried. | |
227 | * | |
228 | * This function is called from do_swap_page(). | |
229 | */ | |
230 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, | |
231 | unsigned long address) | |
232 | { | |
233 | pte_t *ptep, pte; | |
234 | spinlock_t *ptl; | |
235 | swp_entry_t entry; | |
236 | struct page *page; | |
237 | ||
238 | ptep = pte_offset_map_lock(mm, pmd, address, &ptl); | |
239 | pte = *ptep; | |
240 | if (!is_swap_pte(pte)) | |
241 | goto out; | |
242 | ||
243 | entry = pte_to_swp_entry(pte); | |
244 | if (!is_migration_entry(entry)) | |
245 | goto out; | |
246 | ||
247 | page = migration_entry_to_page(entry); | |
248 | ||
e286781d NP |
249 | /* |
250 | * Once radix-tree replacement of page migration started, page_count | |
251 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
252 | * against a page without get_page(). | |
253 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
254 | * will occur again. | |
255 | */ | |
256 | if (!get_page_unless_zero(page)) | |
257 | goto out; | |
0697212a CL |
258 | pte_unmap_unlock(ptep, ptl); |
259 | wait_on_page_locked(page); | |
260 | put_page(page); | |
261 | return; | |
262 | out: | |
263 | pte_unmap_unlock(ptep, ptl); | |
264 | } | |
265 | ||
b20a3503 | 266 | /* |
c3fcf8a5 | 267 | * Replace the page in the mapping. |
5b5c7120 CL |
268 | * |
269 | * The number of remaining references must be: | |
270 | * 1 for anonymous pages without a mapping | |
271 | * 2 for pages with a mapping | |
272 | * 3 for pages with a mapping and PagePrivate set. | |
b20a3503 | 273 | */ |
2d1db3b1 CL |
274 | static int migrate_page_move_mapping(struct address_space *mapping, |
275 | struct page *newpage, struct page *page) | |
b20a3503 | 276 | { |
e286781d | 277 | int expected_count; |
7cf9c2c7 | 278 | void **pslot; |
b20a3503 | 279 | |
6c5240ae | 280 | if (!mapping) { |
0e8c7d0f | 281 | /* Anonymous page without mapping */ |
6c5240ae CL |
282 | if (page_count(page) != 1) |
283 | return -EAGAIN; | |
284 | return 0; | |
285 | } | |
286 | ||
19fd6231 | 287 | spin_lock_irq(&mapping->tree_lock); |
b20a3503 | 288 | |
7cf9c2c7 NP |
289 | pslot = radix_tree_lookup_slot(&mapping->page_tree, |
290 | page_index(page)); | |
b20a3503 | 291 | |
e286781d NP |
292 | expected_count = 2 + !!PagePrivate(page); |
293 | if (page_count(page) != expected_count || | |
7cf9c2c7 | 294 | (struct page *)radix_tree_deref_slot(pslot) != page) { |
19fd6231 | 295 | spin_unlock_irq(&mapping->tree_lock); |
e23ca00b | 296 | return -EAGAIN; |
b20a3503 CL |
297 | } |
298 | ||
e286781d | 299 | if (!page_freeze_refs(page, expected_count)) { |
19fd6231 | 300 | spin_unlock_irq(&mapping->tree_lock); |
e286781d NP |
301 | return -EAGAIN; |
302 | } | |
303 | ||
b20a3503 CL |
304 | /* |
305 | * Now we know that no one else is looking at the page. | |
b20a3503 | 306 | */ |
7cf9c2c7 | 307 | get_page(newpage); /* add cache reference */ |
6c5240ae | 308 | #ifdef CONFIG_SWAP |
b20a3503 CL |
309 | if (PageSwapCache(page)) { |
310 | SetPageSwapCache(newpage); | |
311 | set_page_private(newpage, page_private(page)); | |
312 | } | |
6c5240ae | 313 | #endif |
b20a3503 | 314 | |
7cf9c2c7 NP |
315 | radix_tree_replace_slot(pslot, newpage); |
316 | ||
e286781d | 317 | page_unfreeze_refs(page, expected_count); |
7cf9c2c7 NP |
318 | /* |
319 | * Drop cache reference from old page. | |
320 | * We know this isn't the last reference. | |
321 | */ | |
b20a3503 | 322 | __put_page(page); |
7cf9c2c7 | 323 | |
0e8c7d0f CL |
324 | /* |
325 | * If moved to a different zone then also account | |
326 | * the page for that zone. Other VM counters will be | |
327 | * taken care of when we establish references to the | |
328 | * new page and drop references to the old page. | |
329 | * | |
330 | * Note that anonymous pages are accounted for | |
331 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
332 | * are mapped to swap space. | |
333 | */ | |
334 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
335 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
336 | ||
19fd6231 NP |
337 | spin_unlock_irq(&mapping->tree_lock); |
338 | if (!PageSwapCache(newpage)) | |
69029cd5 | 339 | mem_cgroup_uncharge_cache_page(page); |
b20a3503 CL |
340 | |
341 | return 0; | |
342 | } | |
b20a3503 CL |
343 | |
344 | /* | |
345 | * Copy the page to its new location | |
346 | */ | |
e7340f73 | 347 | static void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 CL |
348 | { |
349 | copy_highpage(newpage, page); | |
350 | ||
351 | if (PageError(page)) | |
352 | SetPageError(newpage); | |
353 | if (PageReferenced(page)) | |
354 | SetPageReferenced(newpage); | |
355 | if (PageUptodate(page)) | |
356 | SetPageUptodate(newpage); | |
357 | if (PageActive(page)) | |
358 | SetPageActive(newpage); | |
359 | if (PageChecked(page)) | |
360 | SetPageChecked(newpage); | |
361 | if (PageMappedToDisk(page)) | |
362 | SetPageMappedToDisk(newpage); | |
363 | ||
364 | if (PageDirty(page)) { | |
365 | clear_page_dirty_for_io(page); | |
3a902c5f NP |
366 | /* |
367 | * Want to mark the page and the radix tree as dirty, and | |
368 | * redo the accounting that clear_page_dirty_for_io undid, | |
369 | * but we can't use set_page_dirty because that function | |
370 | * is actually a signal that all of the page has become dirty. | |
371 | * Wheras only part of our page may be dirty. | |
372 | */ | |
373 | __set_page_dirty_nobuffers(newpage); | |
b20a3503 CL |
374 | } |
375 | ||
6c5240ae | 376 | #ifdef CONFIG_SWAP |
b20a3503 | 377 | ClearPageSwapCache(page); |
6c5240ae | 378 | #endif |
b20a3503 CL |
379 | ClearPageActive(page); |
380 | ClearPagePrivate(page); | |
381 | set_page_private(page, 0); | |
382 | page->mapping = NULL; | |
383 | ||
384 | /* | |
385 | * If any waiters have accumulated on the new page then | |
386 | * wake them up. | |
387 | */ | |
388 | if (PageWriteback(newpage)) | |
389 | end_page_writeback(newpage); | |
390 | } | |
b20a3503 | 391 | |
1d8b85cc CL |
392 | /************************************************************ |
393 | * Migration functions | |
394 | ***********************************************************/ | |
395 | ||
396 | /* Always fail migration. Used for mappings that are not movable */ | |
2d1db3b1 CL |
397 | int fail_migrate_page(struct address_space *mapping, |
398 | struct page *newpage, struct page *page) | |
1d8b85cc CL |
399 | { |
400 | return -EIO; | |
401 | } | |
402 | EXPORT_SYMBOL(fail_migrate_page); | |
403 | ||
b20a3503 CL |
404 | /* |
405 | * Common logic to directly migrate a single page suitable for | |
406 | * pages that do not use PagePrivate. | |
407 | * | |
408 | * Pages are locked upon entry and exit. | |
409 | */ | |
2d1db3b1 CL |
410 | int migrate_page(struct address_space *mapping, |
411 | struct page *newpage, struct page *page) | |
b20a3503 CL |
412 | { |
413 | int rc; | |
414 | ||
415 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
416 | ||
2d1db3b1 | 417 | rc = migrate_page_move_mapping(mapping, newpage, page); |
b20a3503 CL |
418 | |
419 | if (rc) | |
420 | return rc; | |
421 | ||
422 | migrate_page_copy(newpage, page); | |
b20a3503 CL |
423 | return 0; |
424 | } | |
425 | EXPORT_SYMBOL(migrate_page); | |
426 | ||
9361401e | 427 | #ifdef CONFIG_BLOCK |
1d8b85cc CL |
428 | /* |
429 | * Migration function for pages with buffers. This function can only be used | |
430 | * if the underlying filesystem guarantees that no other references to "page" | |
431 | * exist. | |
432 | */ | |
2d1db3b1 CL |
433 | int buffer_migrate_page(struct address_space *mapping, |
434 | struct page *newpage, struct page *page) | |
1d8b85cc | 435 | { |
1d8b85cc CL |
436 | struct buffer_head *bh, *head; |
437 | int rc; | |
438 | ||
1d8b85cc | 439 | if (!page_has_buffers(page)) |
2d1db3b1 | 440 | return migrate_page(mapping, newpage, page); |
1d8b85cc CL |
441 | |
442 | head = page_buffers(page); | |
443 | ||
2d1db3b1 | 444 | rc = migrate_page_move_mapping(mapping, newpage, page); |
1d8b85cc CL |
445 | |
446 | if (rc) | |
447 | return rc; | |
448 | ||
449 | bh = head; | |
450 | do { | |
451 | get_bh(bh); | |
452 | lock_buffer(bh); | |
453 | bh = bh->b_this_page; | |
454 | ||
455 | } while (bh != head); | |
456 | ||
457 | ClearPagePrivate(page); | |
458 | set_page_private(newpage, page_private(page)); | |
459 | set_page_private(page, 0); | |
460 | put_page(page); | |
461 | get_page(newpage); | |
462 | ||
463 | bh = head; | |
464 | do { | |
465 | set_bh_page(bh, newpage, bh_offset(bh)); | |
466 | bh = bh->b_this_page; | |
467 | ||
468 | } while (bh != head); | |
469 | ||
470 | SetPagePrivate(newpage); | |
471 | ||
472 | migrate_page_copy(newpage, page); | |
473 | ||
474 | bh = head; | |
475 | do { | |
476 | unlock_buffer(bh); | |
477 | put_bh(bh); | |
478 | bh = bh->b_this_page; | |
479 | ||
480 | } while (bh != head); | |
481 | ||
482 | return 0; | |
483 | } | |
484 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 485 | #endif |
1d8b85cc | 486 | |
04e62a29 CL |
487 | /* |
488 | * Writeback a page to clean the dirty state | |
489 | */ | |
490 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 491 | { |
04e62a29 CL |
492 | struct writeback_control wbc = { |
493 | .sync_mode = WB_SYNC_NONE, | |
494 | .nr_to_write = 1, | |
495 | .range_start = 0, | |
496 | .range_end = LLONG_MAX, | |
497 | .nonblocking = 1, | |
498 | .for_reclaim = 1 | |
499 | }; | |
500 | int rc; | |
501 | ||
502 | if (!mapping->a_ops->writepage) | |
503 | /* No write method for the address space */ | |
504 | return -EINVAL; | |
505 | ||
506 | if (!clear_page_dirty_for_io(page)) | |
507 | /* Someone else already triggered a write */ | |
508 | return -EAGAIN; | |
509 | ||
8351a6e4 | 510 | /* |
04e62a29 CL |
511 | * A dirty page may imply that the underlying filesystem has |
512 | * the page on some queue. So the page must be clean for | |
513 | * migration. Writeout may mean we loose the lock and the | |
514 | * page state is no longer what we checked for earlier. | |
515 | * At this point we know that the migration attempt cannot | |
516 | * be successful. | |
8351a6e4 | 517 | */ |
04e62a29 | 518 | remove_migration_ptes(page, page); |
8351a6e4 | 519 | |
04e62a29 CL |
520 | rc = mapping->a_ops->writepage(page, &wbc); |
521 | if (rc < 0) | |
522 | /* I/O Error writing */ | |
523 | return -EIO; | |
8351a6e4 | 524 | |
04e62a29 CL |
525 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
526 | /* unlocked. Relock */ | |
527 | lock_page(page); | |
528 | ||
529 | return -EAGAIN; | |
530 | } | |
531 | ||
532 | /* | |
533 | * Default handling if a filesystem does not provide a migration function. | |
534 | */ | |
535 | static int fallback_migrate_page(struct address_space *mapping, | |
536 | struct page *newpage, struct page *page) | |
537 | { | |
538 | if (PageDirty(page)) | |
539 | return writeout(mapping, page); | |
8351a6e4 CL |
540 | |
541 | /* | |
542 | * Buffers may be managed in a filesystem specific way. | |
543 | * We must have no buffers or drop them. | |
544 | */ | |
b398f6bf | 545 | if (PagePrivate(page) && |
8351a6e4 CL |
546 | !try_to_release_page(page, GFP_KERNEL)) |
547 | return -EAGAIN; | |
548 | ||
549 | return migrate_page(mapping, newpage, page); | |
550 | } | |
551 | ||
e24f0b8f CL |
552 | /* |
553 | * Move a page to a newly allocated page | |
554 | * The page is locked and all ptes have been successfully removed. | |
555 | * | |
556 | * The new page will have replaced the old page if this function | |
557 | * is successful. | |
558 | */ | |
559 | static int move_to_new_page(struct page *newpage, struct page *page) | |
560 | { | |
561 | struct address_space *mapping; | |
562 | int rc; | |
563 | ||
564 | /* | |
565 | * Block others from accessing the page when we get around to | |
566 | * establishing additional references. We are the only one | |
567 | * holding a reference to the new page at this point. | |
568 | */ | |
529ae9aa | 569 | if (!trylock_page(newpage)) |
e24f0b8f CL |
570 | BUG(); |
571 | ||
572 | /* Prepare mapping for the new page.*/ | |
573 | newpage->index = page->index; | |
574 | newpage->mapping = page->mapping; | |
575 | ||
576 | mapping = page_mapping(page); | |
577 | if (!mapping) | |
578 | rc = migrate_page(mapping, newpage, page); | |
579 | else if (mapping->a_ops->migratepage) | |
580 | /* | |
581 | * Most pages have a mapping and most filesystems | |
582 | * should provide a migration function. Anonymous | |
583 | * pages are part of swap space which also has its | |
584 | * own migration function. This is the most common | |
585 | * path for page migration. | |
586 | */ | |
587 | rc = mapping->a_ops->migratepage(mapping, | |
588 | newpage, page); | |
589 | else | |
590 | rc = fallback_migrate_page(mapping, newpage, page); | |
591 | ||
ae41be37 | 592 | if (!rc) { |
e24f0b8f | 593 | remove_migration_ptes(page, newpage); |
ae41be37 | 594 | } else |
e24f0b8f CL |
595 | newpage->mapping = NULL; |
596 | ||
597 | unlock_page(newpage); | |
598 | ||
599 | return rc; | |
600 | } | |
601 | ||
602 | /* | |
603 | * Obtain the lock on page, remove all ptes and migrate the page | |
604 | * to the newly allocated page in newpage. | |
605 | */ | |
95a402c3 CL |
606 | static int unmap_and_move(new_page_t get_new_page, unsigned long private, |
607 | struct page *page, int force) | |
e24f0b8f CL |
608 | { |
609 | int rc = 0; | |
742755a1 CL |
610 | int *result = NULL; |
611 | struct page *newpage = get_new_page(page, private, &result); | |
989f89c5 | 612 | int rcu_locked = 0; |
ae41be37 | 613 | int charge = 0; |
95a402c3 CL |
614 | |
615 | if (!newpage) | |
616 | return -ENOMEM; | |
e24f0b8f CL |
617 | |
618 | if (page_count(page) == 1) | |
619 | /* page was freed from under us. So we are done. */ | |
95a402c3 | 620 | goto move_newpage; |
e24f0b8f | 621 | |
e8589cc1 KH |
622 | charge = mem_cgroup_prepare_migration(page, newpage); |
623 | if (charge == -ENOMEM) { | |
624 | rc = -ENOMEM; | |
625 | goto move_newpage; | |
626 | } | |
627 | /* prepare cgroup just returns 0 or -ENOMEM */ | |
628 | BUG_ON(charge); | |
629 | ||
e24f0b8f | 630 | rc = -EAGAIN; |
529ae9aa | 631 | if (!trylock_page(page)) { |
e24f0b8f | 632 | if (!force) |
95a402c3 | 633 | goto move_newpage; |
e24f0b8f CL |
634 | lock_page(page); |
635 | } | |
636 | ||
637 | if (PageWriteback(page)) { | |
638 | if (!force) | |
639 | goto unlock; | |
640 | wait_on_page_writeback(page); | |
641 | } | |
e24f0b8f | 642 | /* |
dc386d4d KH |
643 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
644 | * we cannot notice that anon_vma is freed while we migrates a page. | |
645 | * This rcu_read_lock() delays freeing anon_vma pointer until the end | |
646 | * of migration. File cache pages are no problem because of page_lock() | |
989f89c5 KH |
647 | * File Caches may use write_page() or lock_page() in migration, then, |
648 | * just care Anon page here. | |
dc386d4d | 649 | */ |
989f89c5 KH |
650 | if (PageAnon(page)) { |
651 | rcu_read_lock(); | |
652 | rcu_locked = 1; | |
653 | } | |
62e1c553 | 654 | |
dc386d4d | 655 | /* |
62e1c553 SL |
656 | * Corner case handling: |
657 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
658 | * and treated as swapcache but it has no rmap yet. | |
659 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
660 | * trigger a BUG. So handle it here. | |
661 | * 2. An orphaned page (see truncate_complete_page) might have | |
662 | * fs-private metadata. The page can be picked up due to memory | |
663 | * offlining. Everywhere else except page reclaim, the page is | |
664 | * invisible to the vm, so the page can not be migrated. So try to | |
665 | * free the metadata, so the page can be freed. | |
e24f0b8f | 666 | */ |
62e1c553 SL |
667 | if (!page->mapping) { |
668 | if (!PageAnon(page) && PagePrivate(page)) { | |
669 | /* | |
670 | * Go direct to try_to_free_buffers() here because | |
671 | * a) that's what try_to_release_page() would do anyway | |
672 | * b) we may be under rcu_read_lock() here, so we can't | |
673 | * use GFP_KERNEL which is what try_to_release_page() | |
674 | * needs to be effective. | |
675 | */ | |
676 | try_to_free_buffers(page); | |
677 | } | |
dc386d4d | 678 | goto rcu_unlock; |
62e1c553 SL |
679 | } |
680 | ||
dc386d4d | 681 | /* Establish migration ptes or remove ptes */ |
e6a1530d | 682 | try_to_unmap(page, 1); |
dc386d4d | 683 | |
e6a1530d CL |
684 | if (!page_mapped(page)) |
685 | rc = move_to_new_page(newpage, page); | |
e24f0b8f | 686 | |
e8589cc1 | 687 | if (rc) |
e24f0b8f | 688 | remove_migration_ptes(page, page); |
dc386d4d | 689 | rcu_unlock: |
989f89c5 KH |
690 | if (rcu_locked) |
691 | rcu_read_unlock(); | |
e6a1530d | 692 | |
e24f0b8f | 693 | unlock: |
dc386d4d | 694 | |
e24f0b8f | 695 | unlock_page(page); |
95a402c3 | 696 | |
e24f0b8f | 697 | if (rc != -EAGAIN) { |
aaa994b3 CL |
698 | /* |
699 | * A page that has been migrated has all references | |
700 | * removed and will be freed. A page that has not been | |
701 | * migrated will have kepts its references and be | |
702 | * restored. | |
703 | */ | |
704 | list_del(&page->lru); | |
705 | move_to_lru(page); | |
e24f0b8f | 706 | } |
95a402c3 CL |
707 | |
708 | move_newpage: | |
e8589cc1 KH |
709 | if (!charge) |
710 | mem_cgroup_end_migration(newpage); | |
95a402c3 CL |
711 | /* |
712 | * Move the new page to the LRU. If migration was not successful | |
713 | * then this will free the page. | |
714 | */ | |
715 | move_to_lru(newpage); | |
742755a1 CL |
716 | if (result) { |
717 | if (rc) | |
718 | *result = rc; | |
719 | else | |
720 | *result = page_to_nid(newpage); | |
721 | } | |
e24f0b8f CL |
722 | return rc; |
723 | } | |
724 | ||
b20a3503 CL |
725 | /* |
726 | * migrate_pages | |
727 | * | |
95a402c3 CL |
728 | * The function takes one list of pages to migrate and a function |
729 | * that determines from the page to be migrated and the private data | |
730 | * the target of the move and allocates the page. | |
b20a3503 CL |
731 | * |
732 | * The function returns after 10 attempts or if no pages | |
733 | * are movable anymore because to has become empty | |
aaa994b3 | 734 | * or no retryable pages exist anymore. All pages will be |
e9534b3f | 735 | * returned to the LRU or freed. |
b20a3503 | 736 | * |
95a402c3 | 737 | * Return: Number of pages not migrated or error code. |
b20a3503 | 738 | */ |
95a402c3 CL |
739 | int migrate_pages(struct list_head *from, |
740 | new_page_t get_new_page, unsigned long private) | |
b20a3503 | 741 | { |
e24f0b8f | 742 | int retry = 1; |
b20a3503 CL |
743 | int nr_failed = 0; |
744 | int pass = 0; | |
745 | struct page *page; | |
746 | struct page *page2; | |
747 | int swapwrite = current->flags & PF_SWAPWRITE; | |
748 | int rc; | |
749 | ||
750 | if (!swapwrite) | |
751 | current->flags |= PF_SWAPWRITE; | |
752 | ||
e24f0b8f CL |
753 | for(pass = 0; pass < 10 && retry; pass++) { |
754 | retry = 0; | |
b20a3503 | 755 | |
e24f0b8f | 756 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 757 | cond_resched(); |
2d1db3b1 | 758 | |
95a402c3 CL |
759 | rc = unmap_and_move(get_new_page, private, |
760 | page, pass > 2); | |
2d1db3b1 | 761 | |
e24f0b8f | 762 | switch(rc) { |
95a402c3 CL |
763 | case -ENOMEM: |
764 | goto out; | |
e24f0b8f | 765 | case -EAGAIN: |
2d1db3b1 | 766 | retry++; |
e24f0b8f CL |
767 | break; |
768 | case 0: | |
e24f0b8f CL |
769 | break; |
770 | default: | |
2d1db3b1 | 771 | /* Permanent failure */ |
2d1db3b1 | 772 | nr_failed++; |
e24f0b8f | 773 | break; |
2d1db3b1 | 774 | } |
b20a3503 CL |
775 | } |
776 | } | |
95a402c3 CL |
777 | rc = 0; |
778 | out: | |
b20a3503 CL |
779 | if (!swapwrite) |
780 | current->flags &= ~PF_SWAPWRITE; | |
781 | ||
aaa994b3 | 782 | putback_lru_pages(from); |
b20a3503 | 783 | |
95a402c3 CL |
784 | if (rc) |
785 | return rc; | |
b20a3503 | 786 | |
95a402c3 | 787 | return nr_failed + retry; |
b20a3503 | 788 | } |
95a402c3 | 789 | |
742755a1 CL |
790 | #ifdef CONFIG_NUMA |
791 | /* | |
792 | * Move a list of individual pages | |
793 | */ | |
794 | struct page_to_node { | |
795 | unsigned long addr; | |
796 | struct page *page; | |
797 | int node; | |
798 | int status; | |
799 | }; | |
800 | ||
801 | static struct page *new_page_node(struct page *p, unsigned long private, | |
802 | int **result) | |
803 | { | |
804 | struct page_to_node *pm = (struct page_to_node *)private; | |
805 | ||
806 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
807 | pm++; | |
808 | ||
809 | if (pm->node == MAX_NUMNODES) | |
810 | return NULL; | |
811 | ||
812 | *result = &pm->status; | |
813 | ||
769848c0 MG |
814 | return alloc_pages_node(pm->node, |
815 | GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0); | |
742755a1 CL |
816 | } |
817 | ||
818 | /* | |
819 | * Move a set of pages as indicated in the pm array. The addr | |
820 | * field must be set to the virtual address of the page to be moved | |
821 | * and the node number must contain a valid target node. | |
822 | */ | |
823 | static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm, | |
824 | int migrate_all) | |
825 | { | |
826 | int err; | |
827 | struct page_to_node *pp; | |
828 | LIST_HEAD(pagelist); | |
829 | ||
830 | down_read(&mm->mmap_sem); | |
831 | ||
832 | /* | |
833 | * Build a list of pages to migrate | |
834 | */ | |
835 | migrate_prep(); | |
836 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { | |
837 | struct vm_area_struct *vma; | |
838 | struct page *page; | |
839 | ||
840 | /* | |
841 | * A valid page pointer that will not match any of the | |
842 | * pages that will be moved. | |
843 | */ | |
844 | pp->page = ZERO_PAGE(0); | |
845 | ||
846 | err = -EFAULT; | |
847 | vma = find_vma(mm, pp->addr); | |
0dc952dc | 848 | if (!vma || !vma_migratable(vma)) |
742755a1 CL |
849 | goto set_status; |
850 | ||
851 | page = follow_page(vma, pp->addr, FOLL_GET); | |
89f5b7da LT |
852 | |
853 | err = PTR_ERR(page); | |
854 | if (IS_ERR(page)) | |
855 | goto set_status; | |
856 | ||
742755a1 CL |
857 | err = -ENOENT; |
858 | if (!page) | |
859 | goto set_status; | |
860 | ||
861 | if (PageReserved(page)) /* Check for zero page */ | |
862 | goto put_and_set; | |
863 | ||
864 | pp->page = page; | |
865 | err = page_to_nid(page); | |
866 | ||
867 | if (err == pp->node) | |
868 | /* | |
869 | * Node already in the right place | |
870 | */ | |
871 | goto put_and_set; | |
872 | ||
873 | err = -EACCES; | |
874 | if (page_mapcount(page) > 1 && | |
875 | !migrate_all) | |
876 | goto put_and_set; | |
877 | ||
62695a84 NP |
878 | err = isolate_lru_page(page); |
879 | if (!err) | |
880 | list_add_tail(&page->lru, &pagelist); | |
742755a1 CL |
881 | put_and_set: |
882 | /* | |
883 | * Either remove the duplicate refcount from | |
884 | * isolate_lru_page() or drop the page ref if it was | |
885 | * not isolated. | |
886 | */ | |
887 | put_page(page); | |
888 | set_status: | |
889 | pp->status = err; | |
890 | } | |
891 | ||
892 | if (!list_empty(&pagelist)) | |
893 | err = migrate_pages(&pagelist, new_page_node, | |
894 | (unsigned long)pm); | |
895 | else | |
896 | err = -ENOENT; | |
897 | ||
898 | up_read(&mm->mmap_sem); | |
899 | return err; | |
900 | } | |
901 | ||
902 | /* | |
903 | * Determine the nodes of a list of pages. The addr in the pm array | |
904 | * must have been set to the virtual address of which we want to determine | |
905 | * the node number. | |
906 | */ | |
907 | static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm) | |
908 | { | |
909 | down_read(&mm->mmap_sem); | |
910 | ||
911 | for ( ; pm->node != MAX_NUMNODES; pm++) { | |
912 | struct vm_area_struct *vma; | |
913 | struct page *page; | |
914 | int err; | |
915 | ||
916 | err = -EFAULT; | |
917 | vma = find_vma(mm, pm->addr); | |
918 | if (!vma) | |
919 | goto set_status; | |
920 | ||
921 | page = follow_page(vma, pm->addr, 0); | |
89f5b7da LT |
922 | |
923 | err = PTR_ERR(page); | |
924 | if (IS_ERR(page)) | |
925 | goto set_status; | |
926 | ||
742755a1 CL |
927 | err = -ENOENT; |
928 | /* Use PageReserved to check for zero page */ | |
929 | if (!page || PageReserved(page)) | |
930 | goto set_status; | |
931 | ||
932 | err = page_to_nid(page); | |
933 | set_status: | |
934 | pm->status = err; | |
935 | } | |
936 | ||
937 | up_read(&mm->mmap_sem); | |
938 | return 0; | |
939 | } | |
940 | ||
941 | /* | |
942 | * Move a list of pages in the address space of the currently executing | |
943 | * process. | |
944 | */ | |
945 | asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages, | |
946 | const void __user * __user *pages, | |
947 | const int __user *nodes, | |
948 | int __user *status, int flags) | |
949 | { | |
950 | int err = 0; | |
951 | int i; | |
952 | struct task_struct *task; | |
953 | nodemask_t task_nodes; | |
954 | struct mm_struct *mm; | |
955 | struct page_to_node *pm = NULL; | |
956 | ||
957 | /* Check flags */ | |
958 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
959 | return -EINVAL; | |
960 | ||
961 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
962 | return -EPERM; | |
963 | ||
964 | /* Find the mm_struct */ | |
965 | read_lock(&tasklist_lock); | |
228ebcbe | 966 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 CL |
967 | if (!task) { |
968 | read_unlock(&tasklist_lock); | |
969 | return -ESRCH; | |
970 | } | |
971 | mm = get_task_mm(task); | |
972 | read_unlock(&tasklist_lock); | |
973 | ||
974 | if (!mm) | |
975 | return -EINVAL; | |
976 | ||
977 | /* | |
978 | * Check if this process has the right to modify the specified | |
979 | * process. The right exists if the process has administrative | |
980 | * capabilities, superuser privileges or the same | |
981 | * userid as the target process. | |
982 | */ | |
983 | if ((current->euid != task->suid) && (current->euid != task->uid) && | |
984 | (current->uid != task->suid) && (current->uid != task->uid) && | |
985 | !capable(CAP_SYS_NICE)) { | |
986 | err = -EPERM; | |
987 | goto out2; | |
988 | } | |
989 | ||
86c3a764 DQ |
990 | err = security_task_movememory(task); |
991 | if (err) | |
992 | goto out2; | |
993 | ||
994 | ||
742755a1 CL |
995 | task_nodes = cpuset_mems_allowed(task); |
996 | ||
997 | /* Limit nr_pages so that the multiplication may not overflow */ | |
998 | if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) { | |
999 | err = -E2BIG; | |
1000 | goto out2; | |
1001 | } | |
1002 | ||
1003 | pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node)); | |
1004 | if (!pm) { | |
1005 | err = -ENOMEM; | |
1006 | goto out2; | |
1007 | } | |
1008 | ||
1009 | /* | |
1010 | * Get parameters from user space and initialize the pm | |
1011 | * array. Return various errors if the user did something wrong. | |
1012 | */ | |
1013 | for (i = 0; i < nr_pages; i++) { | |
9d966d49 | 1014 | const void __user *p; |
742755a1 CL |
1015 | |
1016 | err = -EFAULT; | |
1017 | if (get_user(p, pages + i)) | |
1018 | goto out; | |
1019 | ||
1020 | pm[i].addr = (unsigned long)p; | |
1021 | if (nodes) { | |
1022 | int node; | |
1023 | ||
1024 | if (get_user(node, nodes + i)) | |
1025 | goto out; | |
1026 | ||
1027 | err = -ENODEV; | |
56bbd65d | 1028 | if (!node_state(node, N_HIGH_MEMORY)) |
742755a1 CL |
1029 | goto out; |
1030 | ||
1031 | err = -EACCES; | |
1032 | if (!node_isset(node, task_nodes)) | |
1033 | goto out; | |
1034 | ||
1035 | pm[i].node = node; | |
8ce08464 SR |
1036 | } else |
1037 | pm[i].node = 0; /* anything to not match MAX_NUMNODES */ | |
742755a1 CL |
1038 | } |
1039 | /* End marker */ | |
1040 | pm[nr_pages].node = MAX_NUMNODES; | |
1041 | ||
1042 | if (nodes) | |
1043 | err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL); | |
1044 | else | |
1045 | err = do_pages_stat(mm, pm); | |
1046 | ||
1047 | if (err >= 0) | |
1048 | /* Return status information */ | |
1049 | for (i = 0; i < nr_pages; i++) | |
1050 | if (put_user(pm[i].status, status + i)) | |
1051 | err = -EFAULT; | |
1052 | ||
1053 | out: | |
1054 | vfree(pm); | |
1055 | out2: | |
1056 | mmput(mm); | |
1057 | return err; | |
1058 | } | |
742755a1 | 1059 | |
7b2259b3 CL |
1060 | /* |
1061 | * Call migration functions in the vma_ops that may prepare | |
1062 | * memory in a vm for migration. migration functions may perform | |
1063 | * the migration for vmas that do not have an underlying page struct. | |
1064 | */ | |
1065 | int migrate_vmas(struct mm_struct *mm, const nodemask_t *to, | |
1066 | const nodemask_t *from, unsigned long flags) | |
1067 | { | |
1068 | struct vm_area_struct *vma; | |
1069 | int err = 0; | |
1070 | ||
1071 | for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) { | |
1072 | if (vma->vm_ops && vma->vm_ops->migrate) { | |
1073 | err = vma->vm_ops->migrate(vma, to, from, flags); | |
1074 | if (err) | |
1075 | break; | |
1076 | } | |
1077 | } | |
1078 | return err; | |
1079 | } | |
83d1674a | 1080 | #endif |