]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Memory Migration functionality - linux/mm/migration.c | |
3 | * | |
4 | * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter | |
5 | * | |
6 | * Page migration was first developed in the context of the memory hotplug | |
7 | * project. The main authors of the migration code are: | |
8 | * | |
9 | * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> | |
10 | * Hirokazu Takahashi <taka@valinux.co.jp> | |
11 | * Dave Hansen <haveblue@us.ibm.com> | |
12 | * Christoph Lameter | |
13 | */ | |
14 | ||
15 | #include <linux/migrate.h> | |
16 | #include <linux/export.h> | |
17 | #include <linux/swap.h> | |
18 | #include <linux/swapops.h> | |
19 | #include <linux/pagemap.h> | |
20 | #include <linux/buffer_head.h> | |
21 | #include <linux/mm_inline.h> | |
22 | #include <linux/nsproxy.h> | |
23 | #include <linux/pagevec.h> | |
24 | #include <linux/ksm.h> | |
25 | #include <linux/rmap.h> | |
26 | #include <linux/topology.h> | |
27 | #include <linux/cpu.h> | |
28 | #include <linux/cpuset.h> | |
29 | #include <linux/writeback.h> | |
30 | #include <linux/mempolicy.h> | |
31 | #include <linux/vmalloc.h> | |
32 | #include <linux/security.h> | |
33 | #include <linux/memcontrol.h> | |
34 | #include <linux/syscalls.h> | |
35 | #include <linux/hugetlb.h> | |
36 | #include <linux/hugetlb_cgroup.h> | |
37 | #include <linux/gfp.h> | |
38 | #include <linux/balloon_compaction.h> | |
39 | ||
40 | #include <asm/tlbflush.h> | |
41 | ||
42 | #define CREATE_TRACE_POINTS | |
43 | #include <trace/events/migrate.h> | |
44 | ||
45 | #include "internal.h" | |
46 | ||
47 | /* | |
48 | * migrate_prep() needs to be called before we start compiling a list of pages | |
49 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is | |
50 | * undesirable, use migrate_prep_local() | |
51 | */ | |
52 | int migrate_prep(void) | |
53 | { | |
54 | /* | |
55 | * Clear the LRU lists so pages can be isolated. | |
56 | * Note that pages may be moved off the LRU after we have | |
57 | * drained them. Those pages will fail to migrate like other | |
58 | * pages that may be busy. | |
59 | */ | |
60 | lru_add_drain_all(); | |
61 | ||
62 | return 0; | |
63 | } | |
64 | ||
65 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ | |
66 | int migrate_prep_local(void) | |
67 | { | |
68 | lru_add_drain(); | |
69 | ||
70 | return 0; | |
71 | } | |
72 | ||
73 | /* | |
74 | * Add isolated pages on the list back to the LRU under page lock | |
75 | * to avoid leaking evictable pages back onto unevictable list. | |
76 | */ | |
77 | void putback_lru_pages(struct list_head *l) | |
78 | { | |
79 | struct page *page; | |
80 | struct page *page2; | |
81 | ||
82 | list_for_each_entry_safe(page, page2, l, lru) { | |
83 | list_del(&page->lru); | |
84 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
85 | page_is_file_cache(page)); | |
86 | putback_lru_page(page); | |
87 | } | |
88 | } | |
89 | ||
90 | /* | |
91 | * Put previously isolated pages back onto the appropriate lists | |
92 | * from where they were once taken off for compaction/migration. | |
93 | * | |
94 | * This function shall be used instead of putback_lru_pages(), | |
95 | * whenever the isolated pageset has been built by isolate_migratepages_range() | |
96 | */ | |
97 | void putback_movable_pages(struct list_head *l) | |
98 | { | |
99 | struct page *page; | |
100 | struct page *page2; | |
101 | ||
102 | list_for_each_entry_safe(page, page2, l, lru) { | |
103 | if (unlikely(PageHuge(page))) { | |
104 | putback_active_hugepage(page); | |
105 | continue; | |
106 | } | |
107 | list_del(&page->lru); | |
108 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
109 | page_is_file_cache(page)); | |
110 | if (unlikely(isolated_balloon_page(page))) | |
111 | balloon_page_putback(page); | |
112 | else | |
113 | putback_lru_page(page); | |
114 | } | |
115 | } | |
116 | ||
117 | /* | |
118 | * Restore a potential migration pte to a working pte entry | |
119 | */ | |
120 | static int remove_migration_pte(struct page *new, struct vm_area_struct *vma, | |
121 | unsigned long addr, void *old) | |
122 | { | |
123 | struct mm_struct *mm = vma->vm_mm; | |
124 | swp_entry_t entry; | |
125 | pmd_t *pmd; | |
126 | pte_t *ptep, pte; | |
127 | spinlock_t *ptl; | |
128 | ||
129 | if (unlikely(PageHuge(new))) { | |
130 | ptep = huge_pte_offset(mm, addr); | |
131 | if (!ptep) | |
132 | goto out; | |
133 | ptl = &mm->page_table_lock; | |
134 | } else { | |
135 | pmd = mm_find_pmd(mm, addr); | |
136 | if (!pmd) | |
137 | goto out; | |
138 | if (pmd_trans_huge(*pmd)) | |
139 | goto out; | |
140 | ||
141 | ptep = pte_offset_map(pmd, addr); | |
142 | ||
143 | /* | |
144 | * Peek to check is_swap_pte() before taking ptlock? No, we | |
145 | * can race mremap's move_ptes(), which skips anon_vma lock. | |
146 | */ | |
147 | ||
148 | ptl = pte_lockptr(mm, pmd); | |
149 | } | |
150 | ||
151 | spin_lock(ptl); | |
152 | pte = *ptep; | |
153 | if (!is_swap_pte(pte)) | |
154 | goto unlock; | |
155 | ||
156 | entry = pte_to_swp_entry(pte); | |
157 | ||
158 | if (!is_migration_entry(entry) || | |
159 | migration_entry_to_page(entry) != old) | |
160 | goto unlock; | |
161 | ||
162 | get_page(new); | |
163 | pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); | |
164 | if (is_write_migration_entry(entry)) | |
165 | pte = pte_mkwrite(pte); | |
166 | #ifdef CONFIG_HUGETLB_PAGE | |
167 | if (PageHuge(new)) { | |
168 | pte = pte_mkhuge(pte); | |
169 | pte = arch_make_huge_pte(pte, vma, new, 0); | |
170 | } | |
171 | #endif | |
172 | flush_dcache_page(new); | |
173 | set_pte_at(mm, addr, ptep, pte); | |
174 | ||
175 | if (PageHuge(new)) { | |
176 | if (PageAnon(new)) | |
177 | hugepage_add_anon_rmap(new, vma, addr); | |
178 | else | |
179 | page_dup_rmap(new); | |
180 | } else if (PageAnon(new)) | |
181 | page_add_anon_rmap(new, vma, addr); | |
182 | else | |
183 | page_add_file_rmap(new); | |
184 | ||
185 | /* No need to invalidate - it was non-present before */ | |
186 | update_mmu_cache(vma, addr, ptep); | |
187 | unlock: | |
188 | pte_unmap_unlock(ptep, ptl); | |
189 | out: | |
190 | return SWAP_AGAIN; | |
191 | } | |
192 | ||
193 | /* | |
194 | * Get rid of all migration entries and replace them by | |
195 | * references to the indicated page. | |
196 | */ | |
197 | static void remove_migration_ptes(struct page *old, struct page *new) | |
198 | { | |
199 | rmap_walk(new, remove_migration_pte, old); | |
200 | } | |
201 | ||
202 | /* | |
203 | * Something used the pte of a page under migration. We need to | |
204 | * get to the page and wait until migration is finished. | |
205 | * When we return from this function the fault will be retried. | |
206 | */ | |
207 | static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, | |
208 | spinlock_t *ptl) | |
209 | { | |
210 | pte_t pte; | |
211 | swp_entry_t entry; | |
212 | struct page *page; | |
213 | ||
214 | spin_lock(ptl); | |
215 | pte = *ptep; | |
216 | if (!is_swap_pte(pte)) | |
217 | goto out; | |
218 | ||
219 | entry = pte_to_swp_entry(pte); | |
220 | if (!is_migration_entry(entry)) | |
221 | goto out; | |
222 | ||
223 | page = migration_entry_to_page(entry); | |
224 | ||
225 | /* | |
226 | * Once radix-tree replacement of page migration started, page_count | |
227 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
228 | * against a page without get_page(). | |
229 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
230 | * will occur again. | |
231 | */ | |
232 | if (!get_page_unless_zero(page)) | |
233 | goto out; | |
234 | pte_unmap_unlock(ptep, ptl); | |
235 | wait_on_page_locked(page); | |
236 | put_page(page); | |
237 | return; | |
238 | out: | |
239 | pte_unmap_unlock(ptep, ptl); | |
240 | } | |
241 | ||
242 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, | |
243 | unsigned long address) | |
244 | { | |
245 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
246 | pte_t *ptep = pte_offset_map(pmd, address); | |
247 | __migration_entry_wait(mm, ptep, ptl); | |
248 | } | |
249 | ||
250 | void migration_entry_wait_huge(struct mm_struct *mm, pte_t *pte) | |
251 | { | |
252 | spinlock_t *ptl = &(mm)->page_table_lock; | |
253 | __migration_entry_wait(mm, pte, ptl); | |
254 | } | |
255 | ||
256 | #ifdef CONFIG_BLOCK | |
257 | /* Returns true if all buffers are successfully locked */ | |
258 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
259 | enum migrate_mode mode) | |
260 | { | |
261 | struct buffer_head *bh = head; | |
262 | ||
263 | /* Simple case, sync compaction */ | |
264 | if (mode != MIGRATE_ASYNC) { | |
265 | do { | |
266 | get_bh(bh); | |
267 | lock_buffer(bh); | |
268 | bh = bh->b_this_page; | |
269 | ||
270 | } while (bh != head); | |
271 | ||
272 | return true; | |
273 | } | |
274 | ||
275 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
276 | do { | |
277 | get_bh(bh); | |
278 | if (!trylock_buffer(bh)) { | |
279 | /* | |
280 | * We failed to lock the buffer and cannot stall in | |
281 | * async migration. Release the taken locks | |
282 | */ | |
283 | struct buffer_head *failed_bh = bh; | |
284 | put_bh(failed_bh); | |
285 | bh = head; | |
286 | while (bh != failed_bh) { | |
287 | unlock_buffer(bh); | |
288 | put_bh(bh); | |
289 | bh = bh->b_this_page; | |
290 | } | |
291 | return false; | |
292 | } | |
293 | ||
294 | bh = bh->b_this_page; | |
295 | } while (bh != head); | |
296 | return true; | |
297 | } | |
298 | #else | |
299 | static inline bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
300 | enum migrate_mode mode) | |
301 | { | |
302 | return true; | |
303 | } | |
304 | #endif /* CONFIG_BLOCK */ | |
305 | ||
306 | /* | |
307 | * Replace the page in the mapping. | |
308 | * | |
309 | * The number of remaining references must be: | |
310 | * 1 for anonymous pages without a mapping | |
311 | * 2 for pages with a mapping | |
312 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. | |
313 | */ | |
314 | int migrate_page_move_mapping(struct address_space *mapping, | |
315 | struct page *newpage, struct page *page, | |
316 | struct buffer_head *head, enum migrate_mode mode) | |
317 | { | |
318 | int expected_count = 0; | |
319 | void **pslot; | |
320 | ||
321 | if (!mapping) { | |
322 | /* Anonymous page without mapping */ | |
323 | if (page_count(page) != 1) | |
324 | return -EAGAIN; | |
325 | return MIGRATEPAGE_SUCCESS; | |
326 | } | |
327 | ||
328 | spin_lock_irq(&mapping->tree_lock); | |
329 | ||
330 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
331 | page_index(page)); | |
332 | ||
333 | expected_count = 2 + page_has_private(page); | |
334 | if (page_count(page) != expected_count || | |
335 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { | |
336 | spin_unlock_irq(&mapping->tree_lock); | |
337 | return -EAGAIN; | |
338 | } | |
339 | ||
340 | if (!page_freeze_refs(page, expected_count)) { | |
341 | spin_unlock_irq(&mapping->tree_lock); | |
342 | return -EAGAIN; | |
343 | } | |
344 | ||
345 | /* | |
346 | * In the async migration case of moving a page with buffers, lock the | |
347 | * buffers using trylock before the mapping is moved. If the mapping | |
348 | * was moved, we later failed to lock the buffers and could not move | |
349 | * the mapping back due to an elevated page count, we would have to | |
350 | * block waiting on other references to be dropped. | |
351 | */ | |
352 | if (mode == MIGRATE_ASYNC && head && | |
353 | !buffer_migrate_lock_buffers(head, mode)) { | |
354 | page_unfreeze_refs(page, expected_count); | |
355 | spin_unlock_irq(&mapping->tree_lock); | |
356 | return -EAGAIN; | |
357 | } | |
358 | ||
359 | /* | |
360 | * Now we know that no one else is looking at the page. | |
361 | */ | |
362 | get_page(newpage); /* add cache reference */ | |
363 | if (PageSwapCache(page)) { | |
364 | SetPageSwapCache(newpage); | |
365 | set_page_private(newpage, page_private(page)); | |
366 | } | |
367 | ||
368 | radix_tree_replace_slot(pslot, newpage); | |
369 | ||
370 | /* | |
371 | * Drop cache reference from old page by unfreezing | |
372 | * to one less reference. | |
373 | * We know this isn't the last reference. | |
374 | */ | |
375 | page_unfreeze_refs(page, expected_count - 1); | |
376 | ||
377 | /* | |
378 | * If moved to a different zone then also account | |
379 | * the page for that zone. Other VM counters will be | |
380 | * taken care of when we establish references to the | |
381 | * new page and drop references to the old page. | |
382 | * | |
383 | * Note that anonymous pages are accounted for | |
384 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
385 | * are mapped to swap space. | |
386 | */ | |
387 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
388 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
389 | if (!PageSwapCache(page) && PageSwapBacked(page)) { | |
390 | __dec_zone_page_state(page, NR_SHMEM); | |
391 | __inc_zone_page_state(newpage, NR_SHMEM); | |
392 | } | |
393 | spin_unlock_irq(&mapping->tree_lock); | |
394 | ||
395 | return MIGRATEPAGE_SUCCESS; | |
396 | } | |
397 | ||
398 | /* | |
399 | * The expected number of remaining references is the same as that | |
400 | * of migrate_page_move_mapping(). | |
401 | */ | |
402 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
403 | struct page *newpage, struct page *page) | |
404 | { | |
405 | int expected_count; | |
406 | void **pslot; | |
407 | ||
408 | if (!mapping) { | |
409 | if (page_count(page) != 1) | |
410 | return -EAGAIN; | |
411 | return MIGRATEPAGE_SUCCESS; | |
412 | } | |
413 | ||
414 | spin_lock_irq(&mapping->tree_lock); | |
415 | ||
416 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
417 | page_index(page)); | |
418 | ||
419 | expected_count = 2 + page_has_private(page); | |
420 | if (page_count(page) != expected_count || | |
421 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { | |
422 | spin_unlock_irq(&mapping->tree_lock); | |
423 | return -EAGAIN; | |
424 | } | |
425 | ||
426 | if (!page_freeze_refs(page, expected_count)) { | |
427 | spin_unlock_irq(&mapping->tree_lock); | |
428 | return -EAGAIN; | |
429 | } | |
430 | ||
431 | get_page(newpage); | |
432 | ||
433 | radix_tree_replace_slot(pslot, newpage); | |
434 | ||
435 | page_unfreeze_refs(page, expected_count - 1); | |
436 | ||
437 | spin_unlock_irq(&mapping->tree_lock); | |
438 | return MIGRATEPAGE_SUCCESS; | |
439 | } | |
440 | ||
441 | /* | |
442 | * Copy the page to its new location | |
443 | */ | |
444 | void migrate_page_copy(struct page *newpage, struct page *page) | |
445 | { | |
446 | if (PageHuge(page) || PageTransHuge(page)) | |
447 | copy_huge_page(newpage, page); | |
448 | else | |
449 | copy_highpage(newpage, page); | |
450 | ||
451 | if (PageError(page)) | |
452 | SetPageError(newpage); | |
453 | if (PageReferenced(page)) | |
454 | SetPageReferenced(newpage); | |
455 | if (PageUptodate(page)) | |
456 | SetPageUptodate(newpage); | |
457 | if (TestClearPageActive(page)) { | |
458 | VM_BUG_ON(PageUnevictable(page)); | |
459 | SetPageActive(newpage); | |
460 | } else if (TestClearPageUnevictable(page)) | |
461 | SetPageUnevictable(newpage); | |
462 | if (PageChecked(page)) | |
463 | SetPageChecked(newpage); | |
464 | if (PageMappedToDisk(page)) | |
465 | SetPageMappedToDisk(newpage); | |
466 | ||
467 | if (PageDirty(page)) { | |
468 | clear_page_dirty_for_io(page); | |
469 | /* | |
470 | * Want to mark the page and the radix tree as dirty, and | |
471 | * redo the accounting that clear_page_dirty_for_io undid, | |
472 | * but we can't use set_page_dirty because that function | |
473 | * is actually a signal that all of the page has become dirty. | |
474 | * Whereas only part of our page may be dirty. | |
475 | */ | |
476 | if (PageSwapBacked(page)) | |
477 | SetPageDirty(newpage); | |
478 | else | |
479 | __set_page_dirty_nobuffers(newpage); | |
480 | } | |
481 | ||
482 | mlock_migrate_page(newpage, page); | |
483 | ksm_migrate_page(newpage, page); | |
484 | /* | |
485 | * Please do not reorder this without considering how mm/ksm.c's | |
486 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
487 | */ | |
488 | ClearPageSwapCache(page); | |
489 | ClearPagePrivate(page); | |
490 | set_page_private(page, 0); | |
491 | ||
492 | /* | |
493 | * If any waiters have accumulated on the new page then | |
494 | * wake them up. | |
495 | */ | |
496 | if (PageWriteback(newpage)) | |
497 | end_page_writeback(newpage); | |
498 | } | |
499 | ||
500 | /************************************************************ | |
501 | * Migration functions | |
502 | ***********************************************************/ | |
503 | ||
504 | /* Always fail migration. Used for mappings that are not movable */ | |
505 | int fail_migrate_page(struct address_space *mapping, | |
506 | struct page *newpage, struct page *page) | |
507 | { | |
508 | return -EIO; | |
509 | } | |
510 | EXPORT_SYMBOL(fail_migrate_page); | |
511 | ||
512 | /* | |
513 | * Common logic to directly migrate a single page suitable for | |
514 | * pages that do not use PagePrivate/PagePrivate2. | |
515 | * | |
516 | * Pages are locked upon entry and exit. | |
517 | */ | |
518 | int migrate_page(struct address_space *mapping, | |
519 | struct page *newpage, struct page *page, | |
520 | enum migrate_mode mode) | |
521 | { | |
522 | int rc; | |
523 | ||
524 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
525 | ||
526 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode); | |
527 | ||
528 | if (rc != MIGRATEPAGE_SUCCESS) | |
529 | return rc; | |
530 | ||
531 | migrate_page_copy(newpage, page); | |
532 | return MIGRATEPAGE_SUCCESS; | |
533 | } | |
534 | EXPORT_SYMBOL(migrate_page); | |
535 | ||
536 | #ifdef CONFIG_BLOCK | |
537 | /* | |
538 | * Migration function for pages with buffers. This function can only be used | |
539 | * if the underlying filesystem guarantees that no other references to "page" | |
540 | * exist. | |
541 | */ | |
542 | int buffer_migrate_page(struct address_space *mapping, | |
543 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
544 | { | |
545 | struct buffer_head *bh, *head; | |
546 | int rc; | |
547 | ||
548 | if (!page_has_buffers(page)) | |
549 | return migrate_page(mapping, newpage, page, mode); | |
550 | ||
551 | head = page_buffers(page); | |
552 | ||
553 | rc = migrate_page_move_mapping(mapping, newpage, page, head, mode); | |
554 | ||
555 | if (rc != MIGRATEPAGE_SUCCESS) | |
556 | return rc; | |
557 | ||
558 | /* | |
559 | * In the async case, migrate_page_move_mapping locked the buffers | |
560 | * with an IRQ-safe spinlock held. In the sync case, the buffers | |
561 | * need to be locked now | |
562 | */ | |
563 | if (mode != MIGRATE_ASYNC) | |
564 | BUG_ON(!buffer_migrate_lock_buffers(head, mode)); | |
565 | ||
566 | ClearPagePrivate(page); | |
567 | set_page_private(newpage, page_private(page)); | |
568 | set_page_private(page, 0); | |
569 | put_page(page); | |
570 | get_page(newpage); | |
571 | ||
572 | bh = head; | |
573 | do { | |
574 | set_bh_page(bh, newpage, bh_offset(bh)); | |
575 | bh = bh->b_this_page; | |
576 | ||
577 | } while (bh != head); | |
578 | ||
579 | SetPagePrivate(newpage); | |
580 | ||
581 | migrate_page_copy(newpage, page); | |
582 | ||
583 | bh = head; | |
584 | do { | |
585 | unlock_buffer(bh); | |
586 | put_bh(bh); | |
587 | bh = bh->b_this_page; | |
588 | ||
589 | } while (bh != head); | |
590 | ||
591 | return MIGRATEPAGE_SUCCESS; | |
592 | } | |
593 | EXPORT_SYMBOL(buffer_migrate_page); | |
594 | #endif | |
595 | ||
596 | /* | |
597 | * Writeback a page to clean the dirty state | |
598 | */ | |
599 | static int writeout(struct address_space *mapping, struct page *page) | |
600 | { | |
601 | struct writeback_control wbc = { | |
602 | .sync_mode = WB_SYNC_NONE, | |
603 | .nr_to_write = 1, | |
604 | .range_start = 0, | |
605 | .range_end = LLONG_MAX, | |
606 | .for_reclaim = 1 | |
607 | }; | |
608 | int rc; | |
609 | ||
610 | if (!mapping->a_ops->writepage) | |
611 | /* No write method for the address space */ | |
612 | return -EINVAL; | |
613 | ||
614 | if (!clear_page_dirty_for_io(page)) | |
615 | /* Someone else already triggered a write */ | |
616 | return -EAGAIN; | |
617 | ||
618 | /* | |
619 | * A dirty page may imply that the underlying filesystem has | |
620 | * the page on some queue. So the page must be clean for | |
621 | * migration. Writeout may mean we loose the lock and the | |
622 | * page state is no longer what we checked for earlier. | |
623 | * At this point we know that the migration attempt cannot | |
624 | * be successful. | |
625 | */ | |
626 | remove_migration_ptes(page, page); | |
627 | ||
628 | rc = mapping->a_ops->writepage(page, &wbc); | |
629 | ||
630 | if (rc != AOP_WRITEPAGE_ACTIVATE) | |
631 | /* unlocked. Relock */ | |
632 | lock_page(page); | |
633 | ||
634 | return (rc < 0) ? -EIO : -EAGAIN; | |
635 | } | |
636 | ||
637 | /* | |
638 | * Default handling if a filesystem does not provide a migration function. | |
639 | */ | |
640 | static int fallback_migrate_page(struct address_space *mapping, | |
641 | struct page *newpage, struct page *page, enum migrate_mode mode) | |
642 | { | |
643 | if (PageDirty(page)) { | |
644 | /* Only writeback pages in full synchronous migration */ | |
645 | if (mode != MIGRATE_SYNC) | |
646 | return -EBUSY; | |
647 | return writeout(mapping, page); | |
648 | } | |
649 | ||
650 | /* | |
651 | * Buffers may be managed in a filesystem specific way. | |
652 | * We must have no buffers or drop them. | |
653 | */ | |
654 | if (page_has_private(page) && | |
655 | !try_to_release_page(page, GFP_KERNEL)) | |
656 | return -EAGAIN; | |
657 | ||
658 | return migrate_page(mapping, newpage, page, mode); | |
659 | } | |
660 | ||
661 | /* | |
662 | * Move a page to a newly allocated page | |
663 | * The page is locked and all ptes have been successfully removed. | |
664 | * | |
665 | * The new page will have replaced the old page if this function | |
666 | * is successful. | |
667 | * | |
668 | * Return value: | |
669 | * < 0 - error code | |
670 | * MIGRATEPAGE_SUCCESS - success | |
671 | */ | |
672 | static int move_to_new_page(struct page *newpage, struct page *page, | |
673 | int remap_swapcache, enum migrate_mode mode) | |
674 | { | |
675 | struct address_space *mapping; | |
676 | int rc; | |
677 | ||
678 | /* | |
679 | * Block others from accessing the page when we get around to | |
680 | * establishing additional references. We are the only one | |
681 | * holding a reference to the new page at this point. | |
682 | */ | |
683 | if (!trylock_page(newpage)) | |
684 | BUG(); | |
685 | ||
686 | /* Prepare mapping for the new page.*/ | |
687 | newpage->index = page->index; | |
688 | newpage->mapping = page->mapping; | |
689 | if (PageSwapBacked(page)) | |
690 | SetPageSwapBacked(newpage); | |
691 | ||
692 | mapping = page_mapping(page); | |
693 | if (!mapping) | |
694 | rc = migrate_page(mapping, newpage, page, mode); | |
695 | else if (mapping->a_ops->migratepage) | |
696 | /* | |
697 | * Most pages have a mapping and most filesystems provide a | |
698 | * migratepage callback. Anonymous pages are part of swap | |
699 | * space which also has its own migratepage callback. This | |
700 | * is the most common path for page migration. | |
701 | */ | |
702 | rc = mapping->a_ops->migratepage(mapping, | |
703 | newpage, page, mode); | |
704 | else | |
705 | rc = fallback_migrate_page(mapping, newpage, page, mode); | |
706 | ||
707 | if (rc != MIGRATEPAGE_SUCCESS) { | |
708 | newpage->mapping = NULL; | |
709 | } else { | |
710 | if (remap_swapcache) | |
711 | remove_migration_ptes(page, newpage); | |
712 | page->mapping = NULL; | |
713 | } | |
714 | ||
715 | unlock_page(newpage); | |
716 | ||
717 | return rc; | |
718 | } | |
719 | ||
720 | static int __unmap_and_move(struct page *page, struct page *newpage, | |
721 | int force, enum migrate_mode mode) | |
722 | { | |
723 | int rc = -EAGAIN; | |
724 | int remap_swapcache = 1; | |
725 | struct mem_cgroup *mem; | |
726 | struct anon_vma *anon_vma = NULL; | |
727 | ||
728 | if (!trylock_page(page)) { | |
729 | if (!force || mode == MIGRATE_ASYNC) | |
730 | goto out; | |
731 | ||
732 | /* | |
733 | * It's not safe for direct compaction to call lock_page. | |
734 | * For example, during page readahead pages are added locked | |
735 | * to the LRU. Later, when the IO completes the pages are | |
736 | * marked uptodate and unlocked. However, the queueing | |
737 | * could be merging multiple pages for one bio (e.g. | |
738 | * mpage_readpages). If an allocation happens for the | |
739 | * second or third page, the process can end up locking | |
740 | * the same page twice and deadlocking. Rather than | |
741 | * trying to be clever about what pages can be locked, | |
742 | * avoid the use of lock_page for direct compaction | |
743 | * altogether. | |
744 | */ | |
745 | if (current->flags & PF_MEMALLOC) | |
746 | goto out; | |
747 | ||
748 | lock_page(page); | |
749 | } | |
750 | ||
751 | /* charge against new page */ | |
752 | mem_cgroup_prepare_migration(page, newpage, &mem); | |
753 | ||
754 | if (PageWriteback(page)) { | |
755 | /* | |
756 | * Only in the case of a full synchronous migration is it | |
757 | * necessary to wait for PageWriteback. In the async case, | |
758 | * the retry loop is too short and in the sync-light case, | |
759 | * the overhead of stalling is too much | |
760 | */ | |
761 | if (mode != MIGRATE_SYNC) { | |
762 | rc = -EBUSY; | |
763 | goto uncharge; | |
764 | } | |
765 | if (!force) | |
766 | goto uncharge; | |
767 | wait_on_page_writeback(page); | |
768 | } | |
769 | /* | |
770 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, | |
771 | * we cannot notice that anon_vma is freed while we migrates a page. | |
772 | * This get_anon_vma() delays freeing anon_vma pointer until the end | |
773 | * of migration. File cache pages are no problem because of page_lock() | |
774 | * File Caches may use write_page() or lock_page() in migration, then, | |
775 | * just care Anon page here. | |
776 | */ | |
777 | if (PageAnon(page) && !PageKsm(page)) { | |
778 | /* | |
779 | * Only page_lock_anon_vma_read() understands the subtleties of | |
780 | * getting a hold on an anon_vma from outside one of its mms. | |
781 | */ | |
782 | anon_vma = page_get_anon_vma(page); | |
783 | if (anon_vma) { | |
784 | /* | |
785 | * Anon page | |
786 | */ | |
787 | } else if (PageSwapCache(page)) { | |
788 | /* | |
789 | * We cannot be sure that the anon_vma of an unmapped | |
790 | * swapcache page is safe to use because we don't | |
791 | * know in advance if the VMA that this page belonged | |
792 | * to still exists. If the VMA and others sharing the | |
793 | * data have been freed, then the anon_vma could | |
794 | * already be invalid. | |
795 | * | |
796 | * To avoid this possibility, swapcache pages get | |
797 | * migrated but are not remapped when migration | |
798 | * completes | |
799 | */ | |
800 | remap_swapcache = 0; | |
801 | } else { | |
802 | goto uncharge; | |
803 | } | |
804 | } | |
805 | ||
806 | if (unlikely(balloon_page_movable(page))) { | |
807 | /* | |
808 | * A ballooned page does not need any special attention from | |
809 | * physical to virtual reverse mapping procedures. | |
810 | * Skip any attempt to unmap PTEs or to remap swap cache, | |
811 | * in order to avoid burning cycles at rmap level, and perform | |
812 | * the page migration right away (proteced by page lock). | |
813 | */ | |
814 | rc = balloon_page_migrate(newpage, page, mode); | |
815 | goto uncharge; | |
816 | } | |
817 | ||
818 | /* | |
819 | * Corner case handling: | |
820 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
821 | * and treated as swapcache but it has no rmap yet. | |
822 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
823 | * trigger a BUG. So handle it here. | |
824 | * 2. An orphaned page (see truncate_complete_page) might have | |
825 | * fs-private metadata. The page can be picked up due to memory | |
826 | * offlining. Everywhere else except page reclaim, the page is | |
827 | * invisible to the vm, so the page can not be migrated. So try to | |
828 | * free the metadata, so the page can be freed. | |
829 | */ | |
830 | if (!page->mapping) { | |
831 | VM_BUG_ON(PageAnon(page)); | |
832 | if (page_has_private(page)) { | |
833 | try_to_free_buffers(page); | |
834 | goto uncharge; | |
835 | } | |
836 | goto skip_unmap; | |
837 | } | |
838 | ||
839 | /* Establish migration ptes or remove ptes */ | |
840 | try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
841 | ||
842 | skip_unmap: | |
843 | if (!page_mapped(page)) | |
844 | rc = move_to_new_page(newpage, page, remap_swapcache, mode); | |
845 | ||
846 | if (rc && remap_swapcache) | |
847 | remove_migration_ptes(page, page); | |
848 | ||
849 | /* Drop an anon_vma reference if we took one */ | |
850 | if (anon_vma) | |
851 | put_anon_vma(anon_vma); | |
852 | ||
853 | uncharge: | |
854 | mem_cgroup_end_migration(mem, page, newpage, | |
855 | (rc == MIGRATEPAGE_SUCCESS || | |
856 | rc == MIGRATEPAGE_BALLOON_SUCCESS)); | |
857 | unlock_page(page); | |
858 | out: | |
859 | return rc; | |
860 | } | |
861 | ||
862 | /* | |
863 | * Obtain the lock on page, remove all ptes and migrate the page | |
864 | * to the newly allocated page in newpage. | |
865 | */ | |
866 | static int unmap_and_move(new_page_t get_new_page, unsigned long private, | |
867 | struct page *page, int force, enum migrate_mode mode) | |
868 | { | |
869 | int rc = 0; | |
870 | int *result = NULL; | |
871 | struct page *newpage = get_new_page(page, private, &result); | |
872 | ||
873 | if (!newpage) | |
874 | return -ENOMEM; | |
875 | ||
876 | if (page_count(page) == 1) { | |
877 | /* page was freed from under us. So we are done. */ | |
878 | goto out; | |
879 | } | |
880 | ||
881 | if (unlikely(PageTransHuge(page))) | |
882 | if (unlikely(split_huge_page(page))) | |
883 | goto out; | |
884 | ||
885 | rc = __unmap_and_move(page, newpage, force, mode); | |
886 | ||
887 | if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) { | |
888 | /* | |
889 | * A ballooned page has been migrated already. | |
890 | * Now, it's the time to wrap-up counters, | |
891 | * handle the page back to Buddy and return. | |
892 | */ | |
893 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
894 | page_is_file_cache(page)); | |
895 | balloon_page_free(page); | |
896 | return MIGRATEPAGE_SUCCESS; | |
897 | } | |
898 | out: | |
899 | if (rc != -EAGAIN) { | |
900 | /* | |
901 | * A page that has been migrated has all references | |
902 | * removed and will be freed. A page that has not been | |
903 | * migrated will have kepts its references and be | |
904 | * restored. | |
905 | */ | |
906 | list_del(&page->lru); | |
907 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
908 | page_is_file_cache(page)); | |
909 | putback_lru_page(page); | |
910 | } | |
911 | /* | |
912 | * Move the new page to the LRU. If migration was not successful | |
913 | * then this will free the page. | |
914 | */ | |
915 | putback_lru_page(newpage); | |
916 | if (result) { | |
917 | if (rc) | |
918 | *result = rc; | |
919 | else | |
920 | *result = page_to_nid(newpage); | |
921 | } | |
922 | return rc; | |
923 | } | |
924 | ||
925 | /* | |
926 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
927 | * | |
928 | * This function doesn't wait the completion of hugepage I/O | |
929 | * because there is no race between I/O and migration for hugepage. | |
930 | * Note that currently hugepage I/O occurs only in direct I/O | |
931 | * where no lock is held and PG_writeback is irrelevant, | |
932 | * and writeback status of all subpages are counted in the reference | |
933 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
934 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
935 | * This means that when we try to migrate hugepage whose subpages are | |
936 | * doing direct I/O, some references remain after try_to_unmap() and | |
937 | * hugepage migration fails without data corruption. | |
938 | * | |
939 | * There is also no race when direct I/O is issued on the page under migration, | |
940 | * because then pte is replaced with migration swap entry and direct I/O code | |
941 | * will wait in the page fault for migration to complete. | |
942 | */ | |
943 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
944 | unsigned long private, struct page *hpage, | |
945 | int force, enum migrate_mode mode) | |
946 | { | |
947 | int rc = 0; | |
948 | int *result = NULL; | |
949 | struct page *new_hpage = get_new_page(hpage, private, &result); | |
950 | struct anon_vma *anon_vma = NULL; | |
951 | ||
952 | /* | |
953 | * Movability of hugepages depends on architectures and hugepage size. | |
954 | * This check is necessary because some callers of hugepage migration | |
955 | * like soft offline and memory hotremove don't walk through page | |
956 | * tables or check whether the hugepage is pmd-based or not before | |
957 | * kicking migration. | |
958 | */ | |
959 | if (!hugepage_migration_support(page_hstate(hpage))) | |
960 | return -ENOSYS; | |
961 | ||
962 | if (!new_hpage) | |
963 | return -ENOMEM; | |
964 | ||
965 | rc = -EAGAIN; | |
966 | ||
967 | if (!trylock_page(hpage)) { | |
968 | if (!force || mode != MIGRATE_SYNC) | |
969 | goto out; | |
970 | lock_page(hpage); | |
971 | } | |
972 | ||
973 | if (PageAnon(hpage)) | |
974 | anon_vma = page_get_anon_vma(hpage); | |
975 | ||
976 | try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
977 | ||
978 | if (!page_mapped(hpage)) | |
979 | rc = move_to_new_page(new_hpage, hpage, 1, mode); | |
980 | ||
981 | if (rc) | |
982 | remove_migration_ptes(hpage, hpage); | |
983 | ||
984 | if (anon_vma) | |
985 | put_anon_vma(anon_vma); | |
986 | ||
987 | if (!rc) | |
988 | hugetlb_cgroup_migrate(hpage, new_hpage); | |
989 | ||
990 | unlock_page(hpage); | |
991 | out: | |
992 | if (rc != -EAGAIN) | |
993 | putback_active_hugepage(hpage); | |
994 | put_page(new_hpage); | |
995 | if (result) { | |
996 | if (rc) | |
997 | *result = rc; | |
998 | else | |
999 | *result = page_to_nid(new_hpage); | |
1000 | } | |
1001 | return rc; | |
1002 | } | |
1003 | ||
1004 | /* | |
1005 | * migrate_pages - migrate the pages specified in a list, to the free pages | |
1006 | * supplied as the target for the page migration | |
1007 | * | |
1008 | * @from: The list of pages to be migrated. | |
1009 | * @get_new_page: The function used to allocate free pages to be used | |
1010 | * as the target of the page migration. | |
1011 | * @private: Private data to be passed on to get_new_page() | |
1012 | * @mode: The migration mode that specifies the constraints for | |
1013 | * page migration, if any. | |
1014 | * @reason: The reason for page migration. | |
1015 | * | |
1016 | * The function returns after 10 attempts or if no pages are movable any more | |
1017 | * because the list has become empty or no retryable pages exist any more. | |
1018 | * The caller should call putback_lru_pages() to return pages to the LRU | |
1019 | * or free list only if ret != 0. | |
1020 | * | |
1021 | * Returns the number of pages that were not migrated, or an error code. | |
1022 | */ | |
1023 | int migrate_pages(struct list_head *from, new_page_t get_new_page, | |
1024 | unsigned long private, enum migrate_mode mode, int reason) | |
1025 | { | |
1026 | int retry = 1; | |
1027 | int nr_failed = 0; | |
1028 | int nr_succeeded = 0; | |
1029 | int pass = 0; | |
1030 | struct page *page; | |
1031 | struct page *page2; | |
1032 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1033 | int rc; | |
1034 | ||
1035 | if (!swapwrite) | |
1036 | current->flags |= PF_SWAPWRITE; | |
1037 | ||
1038 | for(pass = 0; pass < 10 && retry; pass++) { | |
1039 | retry = 0; | |
1040 | ||
1041 | list_for_each_entry_safe(page, page2, from, lru) { | |
1042 | cond_resched(); | |
1043 | ||
1044 | if (PageHuge(page)) | |
1045 | rc = unmap_and_move_huge_page(get_new_page, | |
1046 | private, page, pass > 2, mode); | |
1047 | else | |
1048 | rc = unmap_and_move(get_new_page, private, | |
1049 | page, pass > 2, mode); | |
1050 | ||
1051 | switch(rc) { | |
1052 | case -ENOMEM: | |
1053 | goto out; | |
1054 | case -EAGAIN: | |
1055 | retry++; | |
1056 | break; | |
1057 | case MIGRATEPAGE_SUCCESS: | |
1058 | nr_succeeded++; | |
1059 | break; | |
1060 | default: | |
1061 | /* Permanent failure */ | |
1062 | nr_failed++; | |
1063 | break; | |
1064 | } | |
1065 | } | |
1066 | } | |
1067 | rc = nr_failed + retry; | |
1068 | out: | |
1069 | if (nr_succeeded) | |
1070 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1071 | if (nr_failed) | |
1072 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
1073 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); | |
1074 | ||
1075 | if (!swapwrite) | |
1076 | current->flags &= ~PF_SWAPWRITE; | |
1077 | ||
1078 | return rc; | |
1079 | } | |
1080 | ||
1081 | #ifdef CONFIG_NUMA | |
1082 | /* | |
1083 | * Move a list of individual pages | |
1084 | */ | |
1085 | struct page_to_node { | |
1086 | unsigned long addr; | |
1087 | struct page *page; | |
1088 | int node; | |
1089 | int status; | |
1090 | }; | |
1091 | ||
1092 | static struct page *new_page_node(struct page *p, unsigned long private, | |
1093 | int **result) | |
1094 | { | |
1095 | struct page_to_node *pm = (struct page_to_node *)private; | |
1096 | ||
1097 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
1098 | pm++; | |
1099 | ||
1100 | if (pm->node == MAX_NUMNODES) | |
1101 | return NULL; | |
1102 | ||
1103 | *result = &pm->status; | |
1104 | ||
1105 | if (PageHuge(p)) | |
1106 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1107 | pm->node); | |
1108 | else | |
1109 | return alloc_pages_exact_node(pm->node, | |
1110 | GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0); | |
1111 | } | |
1112 | ||
1113 | /* | |
1114 | * Move a set of pages as indicated in the pm array. The addr | |
1115 | * field must be set to the virtual address of the page to be moved | |
1116 | * and the node number must contain a valid target node. | |
1117 | * The pm array ends with node = MAX_NUMNODES. | |
1118 | */ | |
1119 | static int do_move_page_to_node_array(struct mm_struct *mm, | |
1120 | struct page_to_node *pm, | |
1121 | int migrate_all) | |
1122 | { | |
1123 | int err; | |
1124 | struct page_to_node *pp; | |
1125 | LIST_HEAD(pagelist); | |
1126 | ||
1127 | down_read(&mm->mmap_sem); | |
1128 | ||
1129 | /* | |
1130 | * Build a list of pages to migrate | |
1131 | */ | |
1132 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { | |
1133 | struct vm_area_struct *vma; | |
1134 | struct page *page; | |
1135 | ||
1136 | err = -EFAULT; | |
1137 | vma = find_vma(mm, pp->addr); | |
1138 | if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) | |
1139 | goto set_status; | |
1140 | ||
1141 | page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT); | |
1142 | ||
1143 | err = PTR_ERR(page); | |
1144 | if (IS_ERR(page)) | |
1145 | goto set_status; | |
1146 | ||
1147 | err = -ENOENT; | |
1148 | if (!page) | |
1149 | goto set_status; | |
1150 | ||
1151 | /* Use PageReserved to check for zero page */ | |
1152 | if (PageReserved(page)) | |
1153 | goto put_and_set; | |
1154 | ||
1155 | pp->page = page; | |
1156 | err = page_to_nid(page); | |
1157 | ||
1158 | if (err == pp->node) | |
1159 | /* | |
1160 | * Node already in the right place | |
1161 | */ | |
1162 | goto put_and_set; | |
1163 | ||
1164 | err = -EACCES; | |
1165 | if (page_mapcount(page) > 1 && | |
1166 | !migrate_all) | |
1167 | goto put_and_set; | |
1168 | ||
1169 | if (PageHuge(page)) { | |
1170 | isolate_huge_page(page, &pagelist); | |
1171 | goto put_and_set; | |
1172 | } | |
1173 | ||
1174 | err = isolate_lru_page(page); | |
1175 | if (!err) { | |
1176 | list_add_tail(&page->lru, &pagelist); | |
1177 | inc_zone_page_state(page, NR_ISOLATED_ANON + | |
1178 | page_is_file_cache(page)); | |
1179 | } | |
1180 | put_and_set: | |
1181 | /* | |
1182 | * Either remove the duplicate refcount from | |
1183 | * isolate_lru_page() or drop the page ref if it was | |
1184 | * not isolated. | |
1185 | */ | |
1186 | put_page(page); | |
1187 | set_status: | |
1188 | pp->status = err; | |
1189 | } | |
1190 | ||
1191 | err = 0; | |
1192 | if (!list_empty(&pagelist)) { | |
1193 | err = migrate_pages(&pagelist, new_page_node, | |
1194 | (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL); | |
1195 | if (err) | |
1196 | putback_movable_pages(&pagelist); | |
1197 | } | |
1198 | ||
1199 | up_read(&mm->mmap_sem); | |
1200 | return err; | |
1201 | } | |
1202 | ||
1203 | /* | |
1204 | * Migrate an array of page address onto an array of nodes and fill | |
1205 | * the corresponding array of status. | |
1206 | */ | |
1207 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, | |
1208 | unsigned long nr_pages, | |
1209 | const void __user * __user *pages, | |
1210 | const int __user *nodes, | |
1211 | int __user *status, int flags) | |
1212 | { | |
1213 | struct page_to_node *pm; | |
1214 | unsigned long chunk_nr_pages; | |
1215 | unsigned long chunk_start; | |
1216 | int err; | |
1217 | ||
1218 | err = -ENOMEM; | |
1219 | pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); | |
1220 | if (!pm) | |
1221 | goto out; | |
1222 | ||
1223 | migrate_prep(); | |
1224 | ||
1225 | /* | |
1226 | * Store a chunk of page_to_node array in a page, | |
1227 | * but keep the last one as a marker | |
1228 | */ | |
1229 | chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; | |
1230 | ||
1231 | for (chunk_start = 0; | |
1232 | chunk_start < nr_pages; | |
1233 | chunk_start += chunk_nr_pages) { | |
1234 | int j; | |
1235 | ||
1236 | if (chunk_start + chunk_nr_pages > nr_pages) | |
1237 | chunk_nr_pages = nr_pages - chunk_start; | |
1238 | ||
1239 | /* fill the chunk pm with addrs and nodes from user-space */ | |
1240 | for (j = 0; j < chunk_nr_pages; j++) { | |
1241 | const void __user *p; | |
1242 | int node; | |
1243 | ||
1244 | err = -EFAULT; | |
1245 | if (get_user(p, pages + j + chunk_start)) | |
1246 | goto out_pm; | |
1247 | pm[j].addr = (unsigned long) p; | |
1248 | ||
1249 | if (get_user(node, nodes + j + chunk_start)) | |
1250 | goto out_pm; | |
1251 | ||
1252 | err = -ENODEV; | |
1253 | if (node < 0 || node >= MAX_NUMNODES) | |
1254 | goto out_pm; | |
1255 | ||
1256 | if (!node_state(node, N_MEMORY)) | |
1257 | goto out_pm; | |
1258 | ||
1259 | err = -EACCES; | |
1260 | if (!node_isset(node, task_nodes)) | |
1261 | goto out_pm; | |
1262 | ||
1263 | pm[j].node = node; | |
1264 | } | |
1265 | ||
1266 | /* End marker for this chunk */ | |
1267 | pm[chunk_nr_pages].node = MAX_NUMNODES; | |
1268 | ||
1269 | /* Migrate this chunk */ | |
1270 | err = do_move_page_to_node_array(mm, pm, | |
1271 | flags & MPOL_MF_MOVE_ALL); | |
1272 | if (err < 0) | |
1273 | goto out_pm; | |
1274 | ||
1275 | /* Return status information */ | |
1276 | for (j = 0; j < chunk_nr_pages; j++) | |
1277 | if (put_user(pm[j].status, status + j + chunk_start)) { | |
1278 | err = -EFAULT; | |
1279 | goto out_pm; | |
1280 | } | |
1281 | } | |
1282 | err = 0; | |
1283 | ||
1284 | out_pm: | |
1285 | free_page((unsigned long)pm); | |
1286 | out: | |
1287 | return err; | |
1288 | } | |
1289 | ||
1290 | /* | |
1291 | * Determine the nodes of an array of pages and store it in an array of status. | |
1292 | */ | |
1293 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, | |
1294 | const void __user **pages, int *status) | |
1295 | { | |
1296 | unsigned long i; | |
1297 | ||
1298 | down_read(&mm->mmap_sem); | |
1299 | ||
1300 | for (i = 0; i < nr_pages; i++) { | |
1301 | unsigned long addr = (unsigned long)(*pages); | |
1302 | struct vm_area_struct *vma; | |
1303 | struct page *page; | |
1304 | int err = -EFAULT; | |
1305 | ||
1306 | vma = find_vma(mm, addr); | |
1307 | if (!vma || addr < vma->vm_start) | |
1308 | goto set_status; | |
1309 | ||
1310 | page = follow_page(vma, addr, 0); | |
1311 | ||
1312 | err = PTR_ERR(page); | |
1313 | if (IS_ERR(page)) | |
1314 | goto set_status; | |
1315 | ||
1316 | err = -ENOENT; | |
1317 | /* Use PageReserved to check for zero page */ | |
1318 | if (!page || PageReserved(page)) | |
1319 | goto set_status; | |
1320 | ||
1321 | err = page_to_nid(page); | |
1322 | set_status: | |
1323 | *status = err; | |
1324 | ||
1325 | pages++; | |
1326 | status++; | |
1327 | } | |
1328 | ||
1329 | up_read(&mm->mmap_sem); | |
1330 | } | |
1331 | ||
1332 | /* | |
1333 | * Determine the nodes of a user array of pages and store it in | |
1334 | * a user array of status. | |
1335 | */ | |
1336 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1337 | const void __user * __user *pages, | |
1338 | int __user *status) | |
1339 | { | |
1340 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1341 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1342 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
1343 | ||
1344 | while (nr_pages) { | |
1345 | unsigned long chunk_nr; | |
1346 | ||
1347 | chunk_nr = nr_pages; | |
1348 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1349 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1350 | ||
1351 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1352 | break; | |
1353 | ||
1354 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1355 | ||
1356 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) | |
1357 | break; | |
1358 | ||
1359 | pages += chunk_nr; | |
1360 | status += chunk_nr; | |
1361 | nr_pages -= chunk_nr; | |
1362 | } | |
1363 | return nr_pages ? -EFAULT : 0; | |
1364 | } | |
1365 | ||
1366 | /* | |
1367 | * Move a list of pages in the address space of the currently executing | |
1368 | * process. | |
1369 | */ | |
1370 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, | |
1371 | const void __user * __user *, pages, | |
1372 | const int __user *, nodes, | |
1373 | int __user *, status, int, flags) | |
1374 | { | |
1375 | const struct cred *cred = current_cred(), *tcred; | |
1376 | struct task_struct *task; | |
1377 | struct mm_struct *mm; | |
1378 | int err; | |
1379 | nodemask_t task_nodes; | |
1380 | ||
1381 | /* Check flags */ | |
1382 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1383 | return -EINVAL; | |
1384 | ||
1385 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1386 | return -EPERM; | |
1387 | ||
1388 | /* Find the mm_struct */ | |
1389 | rcu_read_lock(); | |
1390 | task = pid ? find_task_by_vpid(pid) : current; | |
1391 | if (!task) { | |
1392 | rcu_read_unlock(); | |
1393 | return -ESRCH; | |
1394 | } | |
1395 | get_task_struct(task); | |
1396 | ||
1397 | /* | |
1398 | * Check if this process has the right to modify the specified | |
1399 | * process. The right exists if the process has administrative | |
1400 | * capabilities, superuser privileges or the same | |
1401 | * userid as the target process. | |
1402 | */ | |
1403 | tcred = __task_cred(task); | |
1404 | if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && | |
1405 | !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && | |
1406 | !capable(CAP_SYS_NICE)) { | |
1407 | rcu_read_unlock(); | |
1408 | err = -EPERM; | |
1409 | goto out; | |
1410 | } | |
1411 | rcu_read_unlock(); | |
1412 | ||
1413 | err = security_task_movememory(task); | |
1414 | if (err) | |
1415 | goto out; | |
1416 | ||
1417 | task_nodes = cpuset_mems_allowed(task); | |
1418 | mm = get_task_mm(task); | |
1419 | put_task_struct(task); | |
1420 | ||
1421 | if (!mm) | |
1422 | return -EINVAL; | |
1423 | ||
1424 | if (nodes) | |
1425 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1426 | nodes, status, flags); | |
1427 | else | |
1428 | err = do_pages_stat(mm, nr_pages, pages, status); | |
1429 | ||
1430 | mmput(mm); | |
1431 | return err; | |
1432 | ||
1433 | out: | |
1434 | put_task_struct(task); | |
1435 | return err; | |
1436 | } | |
1437 | ||
1438 | /* | |
1439 | * Call migration functions in the vma_ops that may prepare | |
1440 | * memory in a vm for migration. migration functions may perform | |
1441 | * the migration for vmas that do not have an underlying page struct. | |
1442 | */ | |
1443 | int migrate_vmas(struct mm_struct *mm, const nodemask_t *to, | |
1444 | const nodemask_t *from, unsigned long flags) | |
1445 | { | |
1446 | struct vm_area_struct *vma; | |
1447 | int err = 0; | |
1448 | ||
1449 | for (vma = mm->mmap; vma && !err; vma = vma->vm_next) { | |
1450 | if (vma->vm_ops && vma->vm_ops->migrate) { | |
1451 | err = vma->vm_ops->migrate(vma, to, from, flags); | |
1452 | if (err) | |
1453 | break; | |
1454 | } | |
1455 | } | |
1456 | return err; | |
1457 | } | |
1458 | ||
1459 | #ifdef CONFIG_NUMA_BALANCING | |
1460 | /* | |
1461 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1462 | * pages. Currently it only checks the watermarks which crude | |
1463 | */ | |
1464 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
1465 | unsigned long nr_migrate_pages) | |
1466 | { | |
1467 | int z; | |
1468 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1469 | struct zone *zone = pgdat->node_zones + z; | |
1470 | ||
1471 | if (!populated_zone(zone)) | |
1472 | continue; | |
1473 | ||
1474 | if (!zone_reclaimable(zone)) | |
1475 | continue; | |
1476 | ||
1477 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ | |
1478 | if (!zone_watermark_ok(zone, 0, | |
1479 | high_wmark_pages(zone) + | |
1480 | nr_migrate_pages, | |
1481 | 0, 0)) | |
1482 | continue; | |
1483 | return true; | |
1484 | } | |
1485 | return false; | |
1486 | } | |
1487 | ||
1488 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
1489 | unsigned long data, | |
1490 | int **result) | |
1491 | { | |
1492 | int nid = (int) data; | |
1493 | struct page *newpage; | |
1494 | ||
1495 | newpage = alloc_pages_exact_node(nid, | |
1496 | (GFP_HIGHUSER_MOVABLE | GFP_THISNODE | | |
1497 | __GFP_NOMEMALLOC | __GFP_NORETRY | | |
1498 | __GFP_NOWARN) & | |
1499 | ~GFP_IOFS, 0); | |
1500 | if (newpage) | |
1501 | page_cpupid_xchg_last(newpage, page_cpupid_last(page)); | |
1502 | ||
1503 | return newpage; | |
1504 | } | |
1505 | ||
1506 | /* | |
1507 | * page migration rate limiting control. | |
1508 | * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs | |
1509 | * window of time. Default here says do not migrate more than 1280M per second. | |
1510 | * If a node is rate-limited then PTE NUMA updates are also rate-limited. However | |
1511 | * as it is faults that reset the window, pte updates will happen unconditionally | |
1512 | * if there has not been a fault since @pteupdate_interval_millisecs after the | |
1513 | * throttle window closed. | |
1514 | */ | |
1515 | static unsigned int migrate_interval_millisecs __read_mostly = 100; | |
1516 | static unsigned int pteupdate_interval_millisecs __read_mostly = 1000; | |
1517 | static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); | |
1518 | ||
1519 | /* Returns true if NUMA migration is currently rate limited */ | |
1520 | bool migrate_ratelimited(int node) | |
1521 | { | |
1522 | pg_data_t *pgdat = NODE_DATA(node); | |
1523 | ||
1524 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window + | |
1525 | msecs_to_jiffies(pteupdate_interval_millisecs))) | |
1526 | return false; | |
1527 | ||
1528 | if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages) | |
1529 | return false; | |
1530 | ||
1531 | return true; | |
1532 | } | |
1533 | ||
1534 | /* Returns true if the node is migrate rate-limited after the update */ | |
1535 | bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages) | |
1536 | { | |
1537 | bool rate_limited = false; | |
1538 | ||
1539 | /* | |
1540 | * Rate-limit the amount of data that is being migrated to a node. | |
1541 | * Optimal placement is no good if the memory bus is saturated and | |
1542 | * all the time is being spent migrating! | |
1543 | */ | |
1544 | spin_lock(&pgdat->numabalancing_migrate_lock); | |
1545 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { | |
1546 | pgdat->numabalancing_migrate_nr_pages = 0; | |
1547 | pgdat->numabalancing_migrate_next_window = jiffies + | |
1548 | msecs_to_jiffies(migrate_interval_millisecs); | |
1549 | } | |
1550 | if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) | |
1551 | rate_limited = true; | |
1552 | else | |
1553 | pgdat->numabalancing_migrate_nr_pages += nr_pages; | |
1554 | spin_unlock(&pgdat->numabalancing_migrate_lock); | |
1555 | ||
1556 | return rate_limited; | |
1557 | } | |
1558 | ||
1559 | int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) | |
1560 | { | |
1561 | int page_lru; | |
1562 | ||
1563 | VM_BUG_ON(compound_order(page) && !PageTransHuge(page)); | |
1564 | ||
1565 | /* Avoid migrating to a node that is nearly full */ | |
1566 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) | |
1567 | return 0; | |
1568 | ||
1569 | if (isolate_lru_page(page)) | |
1570 | return 0; | |
1571 | ||
1572 | /* | |
1573 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1574 | * check on page_count(), so we must do it here, now that the page | |
1575 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1576 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1577 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1578 | */ | |
1579 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1580 | putback_lru_page(page); | |
1581 | return 0; | |
1582 | } | |
1583 | ||
1584 | page_lru = page_is_file_cache(page); | |
1585 | mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru, | |
1586 | hpage_nr_pages(page)); | |
1587 | ||
1588 | /* | |
1589 | * Isolating the page has taken another reference, so the | |
1590 | * caller's reference can be safely dropped without the page | |
1591 | * disappearing underneath us during migration. | |
1592 | */ | |
1593 | put_page(page); | |
1594 | return 1; | |
1595 | } | |
1596 | ||
1597 | /* | |
1598 | * Attempt to migrate a misplaced page to the specified destination | |
1599 | * node. Caller is expected to have an elevated reference count on | |
1600 | * the page that will be dropped by this function before returning. | |
1601 | */ | |
1602 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, | |
1603 | int node) | |
1604 | { | |
1605 | pg_data_t *pgdat = NODE_DATA(node); | |
1606 | int isolated; | |
1607 | int nr_remaining; | |
1608 | LIST_HEAD(migratepages); | |
1609 | ||
1610 | /* | |
1611 | * Don't migrate file pages that are mapped in multiple processes | |
1612 | * with execute permissions as they are probably shared libraries. | |
1613 | */ | |
1614 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && | |
1615 | (vma->vm_flags & VM_EXEC)) | |
1616 | goto out; | |
1617 | ||
1618 | /* | |
1619 | * Rate-limit the amount of data that is being migrated to a node. | |
1620 | * Optimal placement is no good if the memory bus is saturated and | |
1621 | * all the time is being spent migrating! | |
1622 | */ | |
1623 | if (numamigrate_update_ratelimit(pgdat, 1)) | |
1624 | goto out; | |
1625 | ||
1626 | isolated = numamigrate_isolate_page(pgdat, page); | |
1627 | if (!isolated) | |
1628 | goto out; | |
1629 | ||
1630 | list_add(&page->lru, &migratepages); | |
1631 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, | |
1632 | node, MIGRATE_ASYNC, MR_NUMA_MISPLACED); | |
1633 | if (nr_remaining) { | |
1634 | putback_lru_pages(&migratepages); | |
1635 | isolated = 0; | |
1636 | } else | |
1637 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
1638 | BUG_ON(!list_empty(&migratepages)); | |
1639 | return isolated; | |
1640 | ||
1641 | out: | |
1642 | put_page(page); | |
1643 | return 0; | |
1644 | } | |
1645 | #endif /* CONFIG_NUMA_BALANCING */ | |
1646 | ||
1647 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) | |
1648 | /* | |
1649 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
1650 | * before returning. | |
1651 | */ | |
1652 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, | |
1653 | struct vm_area_struct *vma, | |
1654 | pmd_t *pmd, pmd_t entry, | |
1655 | unsigned long address, | |
1656 | struct page *page, int node) | |
1657 | { | |
1658 | unsigned long haddr = address & HPAGE_PMD_MASK; | |
1659 | pg_data_t *pgdat = NODE_DATA(node); | |
1660 | int isolated = 0; | |
1661 | struct page *new_page = NULL; | |
1662 | struct mem_cgroup *memcg = NULL; | |
1663 | int page_lru = page_is_file_cache(page); | |
1664 | ||
1665 | /* | |
1666 | * Rate-limit the amount of data that is being migrated to a node. | |
1667 | * Optimal placement is no good if the memory bus is saturated and | |
1668 | * all the time is being spent migrating! | |
1669 | */ | |
1670 | if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR)) | |
1671 | goto out_dropref; | |
1672 | ||
1673 | new_page = alloc_pages_node(node, | |
1674 | (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER); | |
1675 | if (!new_page) | |
1676 | goto out_fail; | |
1677 | ||
1678 | page_cpupid_xchg_last(new_page, page_cpupid_last(page)); | |
1679 | ||
1680 | isolated = numamigrate_isolate_page(pgdat, page); | |
1681 | if (!isolated) { | |
1682 | put_page(new_page); | |
1683 | goto out_fail; | |
1684 | } | |
1685 | ||
1686 | /* Prepare a page as a migration target */ | |
1687 | __set_page_locked(new_page); | |
1688 | SetPageSwapBacked(new_page); | |
1689 | ||
1690 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
1691 | new_page->mapping = page->mapping; | |
1692 | new_page->index = page->index; | |
1693 | migrate_page_copy(new_page, page); | |
1694 | WARN_ON(PageLRU(new_page)); | |
1695 | ||
1696 | /* Recheck the target PMD */ | |
1697 | spin_lock(&mm->page_table_lock); | |
1698 | if (unlikely(!pmd_same(*pmd, entry))) { | |
1699 | spin_unlock(&mm->page_table_lock); | |
1700 | ||
1701 | /* Reverse changes made by migrate_page_copy() */ | |
1702 | if (TestClearPageActive(new_page)) | |
1703 | SetPageActive(page); | |
1704 | if (TestClearPageUnevictable(new_page)) | |
1705 | SetPageUnevictable(page); | |
1706 | mlock_migrate_page(page, new_page); | |
1707 | ||
1708 | unlock_page(new_page); | |
1709 | put_page(new_page); /* Free it */ | |
1710 | ||
1711 | /* Retake the callers reference and putback on LRU */ | |
1712 | get_page(page); | |
1713 | putback_lru_page(page); | |
1714 | mod_zone_page_state(page_zone(page), | |
1715 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); | |
1716 | goto out_fail; | |
1717 | } | |
1718 | ||
1719 | /* | |
1720 | * Traditional migration needs to prepare the memcg charge | |
1721 | * transaction early to prevent the old page from being | |
1722 | * uncharged when installing migration entries. Here we can | |
1723 | * save the potential rollback and start the charge transfer | |
1724 | * only when migration is already known to end successfully. | |
1725 | */ | |
1726 | mem_cgroup_prepare_migration(page, new_page, &memcg); | |
1727 | ||
1728 | entry = mk_pmd(new_page, vma->vm_page_prot); | |
1729 | entry = pmd_mknonnuma(entry); | |
1730 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1731 | entry = pmd_mkhuge(entry); | |
1732 | ||
1733 | pmdp_clear_flush(vma, haddr, pmd); | |
1734 | set_pmd_at(mm, haddr, pmd, entry); | |
1735 | page_add_new_anon_rmap(new_page, vma, haddr); | |
1736 | update_mmu_cache_pmd(vma, address, &entry); | |
1737 | page_remove_rmap(page); | |
1738 | /* | |
1739 | * Finish the charge transaction under the page table lock to | |
1740 | * prevent split_huge_page() from dividing up the charge | |
1741 | * before it's fully transferred to the new page. | |
1742 | */ | |
1743 | mem_cgroup_end_migration(memcg, page, new_page, true); | |
1744 | spin_unlock(&mm->page_table_lock); | |
1745 | ||
1746 | unlock_page(new_page); | |
1747 | unlock_page(page); | |
1748 | put_page(page); /* Drop the rmap reference */ | |
1749 | put_page(page); /* Drop the LRU isolation reference */ | |
1750 | ||
1751 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
1752 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
1753 | ||
1754 | mod_zone_page_state(page_zone(page), | |
1755 | NR_ISOLATED_ANON + page_lru, | |
1756 | -HPAGE_PMD_NR); | |
1757 | return isolated; | |
1758 | ||
1759 | out_fail: | |
1760 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
1761 | out_dropref: | |
1762 | entry = pmd_mknonnuma(entry); | |
1763 | set_pmd_at(mm, haddr, pmd, entry); | |
1764 | update_mmu_cache_pmd(vma, address, &entry); | |
1765 | ||
1766 | unlock_page(page); | |
1767 | put_page(page); | |
1768 | return 0; | |
1769 | } | |
1770 | #endif /* CONFIG_NUMA_BALANCING */ | |
1771 | ||
1772 | #endif /* CONFIG_NUMA */ |