Commit | Line | Data |
---|---|---|
b20a3503 CL |
1 | /* |
2 | * Memory Migration functionality - linux/mm/migration.c | |
3 | * | |
4 | * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter | |
5 | * | |
6 | * Page migration was first developed in the context of the memory hotplug | |
7 | * project. The main authors of the migration code are: | |
8 | * | |
9 | * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> | |
10 | * Hirokazu Takahashi <taka@valinux.co.jp> | |
11 | * Dave Hansen <haveblue@us.ibm.com> | |
12 | * Christoph Lameter <clameter@sgi.com> | |
13 | */ | |
14 | ||
15 | #include <linux/migrate.h> | |
16 | #include <linux/module.h> | |
17 | #include <linux/swap.h> | |
18 | #include <linux/pagemap.h> | |
e23ca00b | 19 | #include <linux/buffer_head.h> |
b20a3503 CL |
20 | #include <linux/mm_inline.h> |
21 | #include <linux/pagevec.h> | |
22 | #include <linux/rmap.h> | |
23 | #include <linux/topology.h> | |
24 | #include <linux/cpu.h> | |
25 | #include <linux/cpuset.h> | |
26 | #include <linux/swapops.h> | |
27 | ||
28 | #include "internal.h" | |
29 | ||
b20a3503 CL |
30 | /* The maximum number of pages to take off the LRU for migration */ |
31 | #define MIGRATE_CHUNK_SIZE 256 | |
32 | ||
33 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) | |
34 | ||
35 | /* | |
36 | * Isolate one page from the LRU lists. If successful put it onto | |
37 | * the indicated list with elevated page count. | |
38 | * | |
39 | * Result: | |
40 | * -EBUSY: page not on LRU list | |
41 | * 0: page removed from LRU list and added to the specified list. | |
42 | */ | |
43 | int isolate_lru_page(struct page *page, struct list_head *pagelist) | |
44 | { | |
45 | int ret = -EBUSY; | |
46 | ||
47 | if (PageLRU(page)) { | |
48 | struct zone *zone = page_zone(page); | |
49 | ||
50 | spin_lock_irq(&zone->lru_lock); | |
51 | if (PageLRU(page)) { | |
52 | ret = 0; | |
53 | get_page(page); | |
54 | ClearPageLRU(page); | |
55 | if (PageActive(page)) | |
56 | del_page_from_active_list(zone, page); | |
57 | else | |
58 | del_page_from_inactive_list(zone, page); | |
59 | list_add_tail(&page->lru, pagelist); | |
60 | } | |
61 | spin_unlock_irq(&zone->lru_lock); | |
62 | } | |
63 | return ret; | |
64 | } | |
65 | ||
66 | /* | |
67 | * migrate_prep() needs to be called after we have compiled the list of pages | |
68 | * to be migrated using isolate_lru_page() but before we begin a series of calls | |
69 | * to migrate_pages(). | |
70 | */ | |
71 | int migrate_prep(void) | |
72 | { | |
73 | /* Must have swap device for migration */ | |
74 | if (nr_swap_pages <= 0) | |
75 | return -ENODEV; | |
76 | ||
77 | /* | |
78 | * Clear the LRU lists so pages can be isolated. | |
79 | * Note that pages may be moved off the LRU after we have | |
80 | * drained them. Those pages will fail to migrate like other | |
81 | * pages that may be busy. | |
82 | */ | |
83 | lru_add_drain_all(); | |
84 | ||
85 | return 0; | |
86 | } | |
87 | ||
88 | static inline void move_to_lru(struct page *page) | |
89 | { | |
90 | list_del(&page->lru); | |
91 | if (PageActive(page)) { | |
92 | /* | |
93 | * lru_cache_add_active checks that | |
94 | * the PG_active bit is off. | |
95 | */ | |
96 | ClearPageActive(page); | |
97 | lru_cache_add_active(page); | |
98 | } else { | |
99 | lru_cache_add(page); | |
100 | } | |
101 | put_page(page); | |
102 | } | |
103 | ||
104 | /* | |
105 | * Add isolated pages on the list back to the LRU. | |
106 | * | |
107 | * returns the number of pages put back. | |
108 | */ | |
109 | int putback_lru_pages(struct list_head *l) | |
110 | { | |
111 | struct page *page; | |
112 | struct page *page2; | |
113 | int count = 0; | |
114 | ||
115 | list_for_each_entry_safe(page, page2, l, lru) { | |
116 | move_to_lru(page); | |
117 | count++; | |
118 | } | |
119 | return count; | |
120 | } | |
121 | ||
b20a3503 CL |
122 | /* |
123 | * swapout a single page | |
124 | * page is locked upon entry, unlocked on exit | |
125 | */ | |
126 | static int swap_page(struct page *page) | |
127 | { | |
128 | struct address_space *mapping = page_mapping(page); | |
129 | ||
130 | if (page_mapped(page) && mapping) | |
131 | if (try_to_unmap(page, 1) != SWAP_SUCCESS) | |
132 | goto unlock_retry; | |
133 | ||
134 | if (PageDirty(page)) { | |
135 | /* Page is dirty, try to write it out here */ | |
136 | switch(pageout(page, mapping)) { | |
137 | case PAGE_KEEP: | |
138 | case PAGE_ACTIVATE: | |
139 | goto unlock_retry; | |
140 | ||
141 | case PAGE_SUCCESS: | |
142 | goto retry; | |
143 | ||
144 | case PAGE_CLEAN: | |
145 | ; /* try to free the page below */ | |
146 | } | |
147 | } | |
148 | ||
149 | if (PagePrivate(page)) { | |
150 | if (!try_to_release_page(page, GFP_KERNEL) || | |
151 | (!mapping && page_count(page) == 1)) | |
152 | goto unlock_retry; | |
153 | } | |
154 | ||
155 | if (remove_mapping(mapping, page)) { | |
156 | /* Success */ | |
157 | unlock_page(page); | |
158 | return 0; | |
159 | } | |
160 | ||
161 | unlock_retry: | |
162 | unlock_page(page); | |
163 | ||
164 | retry: | |
165 | return -EAGAIN; | |
166 | } | |
b20a3503 CL |
167 | |
168 | /* | |
169 | * Remove references for a page and establish the new page with the correct | |
170 | * basic settings to be able to stop accesses to the page. | |
171 | */ | |
172 | int migrate_page_remove_references(struct page *newpage, | |
173 | struct page *page, int nr_refs) | |
174 | { | |
175 | struct address_space *mapping = page_mapping(page); | |
176 | struct page **radix_pointer; | |
177 | ||
178 | /* | |
179 | * Avoid doing any of the following work if the page count | |
180 | * indicates that the page is in use or truncate has removed | |
181 | * the page. | |
182 | */ | |
183 | if (!mapping || page_mapcount(page) + nr_refs != page_count(page)) | |
184 | return -EAGAIN; | |
185 | ||
186 | /* | |
187 | * Establish swap ptes for anonymous pages or destroy pte | |
188 | * maps for files. | |
189 | * | |
190 | * In order to reestablish file backed mappings the fault handlers | |
191 | * will take the radix tree_lock which may then be used to stop | |
192 | * processses from accessing this page until the new page is ready. | |
193 | * | |
194 | * A process accessing via a swap pte (an anonymous page) will take a | |
195 | * page_lock on the old page which will block the process until the | |
196 | * migration attempt is complete. At that time the PageSwapCache bit | |
197 | * will be examined. If the page was migrated then the PageSwapCache | |
198 | * bit will be clear and the operation to retrieve the page will be | |
199 | * retried which will find the new page in the radix tree. Then a new | |
200 | * direct mapping may be generated based on the radix tree contents. | |
201 | * | |
202 | * If the page was not migrated then the PageSwapCache bit | |
203 | * is still set and the operation may continue. | |
204 | */ | |
205 | if (try_to_unmap(page, 1) == SWAP_FAIL) | |
206 | /* A vma has VM_LOCKED set -> permanent failure */ | |
207 | return -EPERM; | |
208 | ||
209 | /* | |
210 | * Give up if we were unable to remove all mappings. | |
211 | */ | |
212 | if (page_mapcount(page)) | |
213 | return -EAGAIN; | |
214 | ||
215 | write_lock_irq(&mapping->tree_lock); | |
216 | ||
217 | radix_pointer = (struct page **)radix_tree_lookup_slot( | |
218 | &mapping->page_tree, | |
219 | page_index(page)); | |
220 | ||
221 | if (!page_mapping(page) || page_count(page) != nr_refs || | |
222 | *radix_pointer != page) { | |
223 | write_unlock_irq(&mapping->tree_lock); | |
e23ca00b | 224 | return -EAGAIN; |
b20a3503 CL |
225 | } |
226 | ||
227 | /* | |
228 | * Now we know that no one else is looking at the page. | |
229 | * | |
230 | * Certain minimal information about a page must be available | |
231 | * in order for other subsystems to properly handle the page if they | |
232 | * find it through the radix tree update before we are finished | |
233 | * copying the page. | |
234 | */ | |
235 | get_page(newpage); | |
236 | newpage->index = page->index; | |
237 | newpage->mapping = page->mapping; | |
238 | if (PageSwapCache(page)) { | |
239 | SetPageSwapCache(newpage); | |
240 | set_page_private(newpage, page_private(page)); | |
241 | } | |
242 | ||
243 | *radix_pointer = newpage; | |
244 | __put_page(page); | |
245 | write_unlock_irq(&mapping->tree_lock); | |
246 | ||
247 | return 0; | |
248 | } | |
249 | EXPORT_SYMBOL(migrate_page_remove_references); | |
250 | ||
251 | /* | |
252 | * Copy the page to its new location | |
253 | */ | |
254 | void migrate_page_copy(struct page *newpage, struct page *page) | |
255 | { | |
256 | copy_highpage(newpage, page); | |
257 | ||
258 | if (PageError(page)) | |
259 | SetPageError(newpage); | |
260 | if (PageReferenced(page)) | |
261 | SetPageReferenced(newpage); | |
262 | if (PageUptodate(page)) | |
263 | SetPageUptodate(newpage); | |
264 | if (PageActive(page)) | |
265 | SetPageActive(newpage); | |
266 | if (PageChecked(page)) | |
267 | SetPageChecked(newpage); | |
268 | if (PageMappedToDisk(page)) | |
269 | SetPageMappedToDisk(newpage); | |
270 | ||
271 | if (PageDirty(page)) { | |
272 | clear_page_dirty_for_io(page); | |
273 | set_page_dirty(newpage); | |
274 | } | |
275 | ||
276 | ClearPageSwapCache(page); | |
277 | ClearPageActive(page); | |
278 | ClearPagePrivate(page); | |
279 | set_page_private(page, 0); | |
280 | page->mapping = NULL; | |
281 | ||
282 | /* | |
283 | * If any waiters have accumulated on the new page then | |
284 | * wake them up. | |
285 | */ | |
286 | if (PageWriteback(newpage)) | |
287 | end_page_writeback(newpage); | |
288 | } | |
289 | EXPORT_SYMBOL(migrate_page_copy); | |
290 | ||
1d8b85cc CL |
291 | /************************************************************ |
292 | * Migration functions | |
293 | ***********************************************************/ | |
294 | ||
295 | /* Always fail migration. Used for mappings that are not movable */ | |
296 | int fail_migrate_page(struct page *newpage, struct page *page) | |
297 | { | |
298 | return -EIO; | |
299 | } | |
300 | EXPORT_SYMBOL(fail_migrate_page); | |
301 | ||
b20a3503 CL |
302 | /* |
303 | * Common logic to directly migrate a single page suitable for | |
304 | * pages that do not use PagePrivate. | |
305 | * | |
306 | * Pages are locked upon entry and exit. | |
307 | */ | |
308 | int migrate_page(struct page *newpage, struct page *page) | |
309 | { | |
310 | int rc; | |
311 | ||
312 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
313 | ||
314 | rc = migrate_page_remove_references(newpage, page, 2); | |
315 | ||
316 | if (rc) | |
317 | return rc; | |
318 | ||
319 | migrate_page_copy(newpage, page); | |
320 | ||
321 | /* | |
322 | * Remove auxiliary swap entries and replace | |
323 | * them with real ptes. | |
324 | * | |
325 | * Note that a real pte entry will allow processes that are not | |
326 | * waiting on the page lock to use the new page via the page tables | |
327 | * before the new page is unlocked. | |
328 | */ | |
329 | remove_from_swap(newpage); | |
330 | return 0; | |
331 | } | |
332 | EXPORT_SYMBOL(migrate_page); | |
333 | ||
1d8b85cc CL |
334 | /* |
335 | * Migration function for pages with buffers. This function can only be used | |
336 | * if the underlying filesystem guarantees that no other references to "page" | |
337 | * exist. | |
338 | */ | |
339 | int buffer_migrate_page(struct page *newpage, struct page *page) | |
340 | { | |
341 | struct address_space *mapping = page->mapping; | |
342 | struct buffer_head *bh, *head; | |
343 | int rc; | |
344 | ||
345 | if (!mapping) | |
346 | return -EAGAIN; | |
347 | ||
348 | if (!page_has_buffers(page)) | |
349 | return migrate_page(newpage, page); | |
350 | ||
351 | head = page_buffers(page); | |
352 | ||
353 | rc = migrate_page_remove_references(newpage, page, 3); | |
354 | ||
355 | if (rc) | |
356 | return rc; | |
357 | ||
358 | bh = head; | |
359 | do { | |
360 | get_bh(bh); | |
361 | lock_buffer(bh); | |
362 | bh = bh->b_this_page; | |
363 | ||
364 | } while (bh != head); | |
365 | ||
366 | ClearPagePrivate(page); | |
367 | set_page_private(newpage, page_private(page)); | |
368 | set_page_private(page, 0); | |
369 | put_page(page); | |
370 | get_page(newpage); | |
371 | ||
372 | bh = head; | |
373 | do { | |
374 | set_bh_page(bh, newpage, bh_offset(bh)); | |
375 | bh = bh->b_this_page; | |
376 | ||
377 | } while (bh != head); | |
378 | ||
379 | SetPagePrivate(newpage); | |
380 | ||
381 | migrate_page_copy(newpage, page); | |
382 | ||
383 | bh = head; | |
384 | do { | |
385 | unlock_buffer(bh); | |
386 | put_bh(bh); | |
387 | bh = bh->b_this_page; | |
388 | ||
389 | } while (bh != head); | |
390 | ||
391 | return 0; | |
392 | } | |
393 | EXPORT_SYMBOL(buffer_migrate_page); | |
394 | ||
b20a3503 CL |
395 | /* |
396 | * migrate_pages | |
397 | * | |
398 | * Two lists are passed to this function. The first list | |
399 | * contains the pages isolated from the LRU to be migrated. | |
400 | * The second list contains new pages that the pages isolated | |
401 | * can be moved to. If the second list is NULL then all | |
402 | * pages are swapped out. | |
403 | * | |
404 | * The function returns after 10 attempts or if no pages | |
405 | * are movable anymore because to has become empty | |
406 | * or no retryable pages exist anymore. | |
407 | * | |
408 | * Return: Number of pages not migrated when "to" ran empty. | |
409 | */ | |
410 | int migrate_pages(struct list_head *from, struct list_head *to, | |
411 | struct list_head *moved, struct list_head *failed) | |
412 | { | |
413 | int retry; | |
414 | int nr_failed = 0; | |
415 | int pass = 0; | |
416 | struct page *page; | |
417 | struct page *page2; | |
418 | int swapwrite = current->flags & PF_SWAPWRITE; | |
419 | int rc; | |
420 | ||
421 | if (!swapwrite) | |
422 | current->flags |= PF_SWAPWRITE; | |
423 | ||
424 | redo: | |
425 | retry = 0; | |
426 | ||
427 | list_for_each_entry_safe(page, page2, from, lru) { | |
428 | struct page *newpage = NULL; | |
429 | struct address_space *mapping; | |
430 | ||
431 | cond_resched(); | |
432 | ||
433 | rc = 0; | |
434 | if (page_count(page) == 1) | |
435 | /* page was freed from under us. So we are done. */ | |
436 | goto next; | |
437 | ||
438 | if (to && list_empty(to)) | |
439 | break; | |
440 | ||
441 | /* | |
442 | * Skip locked pages during the first two passes to give the | |
443 | * functions holding the lock time to release the page. Later we | |
444 | * use lock_page() to have a higher chance of acquiring the | |
445 | * lock. | |
446 | */ | |
447 | rc = -EAGAIN; | |
448 | if (pass > 2) | |
449 | lock_page(page); | |
450 | else | |
451 | if (TestSetPageLocked(page)) | |
452 | goto next; | |
453 | ||
454 | /* | |
455 | * Only wait on writeback if we have already done a pass where | |
456 | * we we may have triggered writeouts for lots of pages. | |
457 | */ | |
458 | if (pass > 0) { | |
459 | wait_on_page_writeback(page); | |
460 | } else { | |
461 | if (PageWriteback(page)) | |
462 | goto unlock_page; | |
463 | } | |
464 | ||
465 | /* | |
466 | * Anonymous pages must have swap cache references otherwise | |
467 | * the information contained in the page maps cannot be | |
468 | * preserved. | |
469 | */ | |
470 | if (PageAnon(page) && !PageSwapCache(page)) { | |
471 | if (!add_to_swap(page, GFP_KERNEL)) { | |
472 | rc = -ENOMEM; | |
473 | goto unlock_page; | |
474 | } | |
475 | } | |
476 | ||
477 | if (!to) { | |
478 | rc = swap_page(page); | |
479 | goto next; | |
480 | } | |
481 | ||
482 | newpage = lru_to_page(to); | |
483 | lock_page(newpage); | |
484 | ||
485 | /* | |
486 | * Pages are properly locked and writeback is complete. | |
487 | * Try to migrate the page. | |
488 | */ | |
489 | mapping = page_mapping(page); | |
490 | if (!mapping) | |
491 | goto unlock_both; | |
492 | ||
493 | if (mapping->a_ops->migratepage) { | |
494 | /* | |
495 | * Most pages have a mapping and most filesystems | |
496 | * should provide a migration function. Anonymous | |
497 | * pages are part of swap space which also has its | |
498 | * own migration function. This is the most common | |
499 | * path for page migration. | |
500 | */ | |
501 | rc = mapping->a_ops->migratepage(newpage, page); | |
502 | goto unlock_both; | |
503 | } | |
504 | ||
4c28f811 CL |
505 | /* Make sure the dirty bit is up to date */ |
506 | if (try_to_unmap(page, 1) == SWAP_FAIL) { | |
507 | rc = -EPERM; | |
508 | goto unlock_both; | |
509 | } | |
510 | ||
511 | if (page_mapcount(page)) { | |
512 | rc = -EAGAIN; | |
513 | goto unlock_both; | |
514 | } | |
515 | ||
b20a3503 CL |
516 | /* |
517 | * Default handling if a filesystem does not provide | |
518 | * a migration function. We can only migrate clean | |
519 | * pages so try to write out any dirty pages first. | |
520 | */ | |
521 | if (PageDirty(page)) { | |
522 | switch (pageout(page, mapping)) { | |
523 | case PAGE_KEEP: | |
524 | case PAGE_ACTIVATE: | |
525 | goto unlock_both; | |
526 | ||
527 | case PAGE_SUCCESS: | |
528 | unlock_page(newpage); | |
529 | goto next; | |
530 | ||
531 | case PAGE_CLEAN: | |
532 | ; /* try to migrate the page below */ | |
533 | } | |
534 | } | |
535 | ||
536 | /* | |
537 | * Buffers are managed in a filesystem specific way. | |
538 | * We must have no buffers or drop them. | |
539 | */ | |
540 | if (!page_has_buffers(page) || | |
541 | try_to_release_page(page, GFP_KERNEL)) { | |
542 | rc = migrate_page(newpage, page); | |
543 | goto unlock_both; | |
544 | } | |
545 | ||
546 | /* | |
547 | * On early passes with mapped pages simply | |
548 | * retry. There may be a lock held for some | |
549 | * buffers that may go away. Later | |
550 | * swap them out. | |
551 | */ | |
552 | if (pass > 4) { | |
553 | /* | |
554 | * Persistently unable to drop buffers..... As a | |
555 | * measure of last resort we fall back to | |
556 | * swap_page(). | |
557 | */ | |
558 | unlock_page(newpage); | |
559 | newpage = NULL; | |
560 | rc = swap_page(page); | |
561 | goto next; | |
562 | } | |
563 | ||
564 | unlock_both: | |
565 | unlock_page(newpage); | |
566 | ||
567 | unlock_page: | |
568 | unlock_page(page); | |
569 | ||
570 | next: | |
571 | if (rc == -EAGAIN) { | |
572 | retry++; | |
573 | } else if (rc) { | |
574 | /* Permanent failure */ | |
575 | list_move(&page->lru, failed); | |
576 | nr_failed++; | |
577 | } else { | |
578 | if (newpage) { | |
579 | /* Successful migration. Return page to LRU */ | |
580 | move_to_lru(newpage); | |
581 | } | |
582 | list_move(&page->lru, moved); | |
583 | } | |
584 | } | |
585 | if (retry && pass++ < 10) | |
586 | goto redo; | |
587 | ||
588 | if (!swapwrite) | |
589 | current->flags &= ~PF_SWAPWRITE; | |
590 | ||
591 | return nr_failed + retry; | |
592 | } | |
593 | ||
b20a3503 CL |
594 | /* |
595 | * Migrate the list 'pagelist' of pages to a certain destination. | |
596 | * | |
597 | * Specify destination with either non-NULL vma or dest_node >= 0 | |
598 | * Return the number of pages not migrated or error code | |
599 | */ | |
600 | int migrate_pages_to(struct list_head *pagelist, | |
601 | struct vm_area_struct *vma, int dest) | |
602 | { | |
603 | LIST_HEAD(newlist); | |
604 | LIST_HEAD(moved); | |
605 | LIST_HEAD(failed); | |
606 | int err = 0; | |
607 | unsigned long offset = 0; | |
608 | int nr_pages; | |
609 | struct page *page; | |
610 | struct list_head *p; | |
611 | ||
612 | redo: | |
613 | nr_pages = 0; | |
614 | list_for_each(p, pagelist) { | |
615 | if (vma) { | |
616 | /* | |
617 | * The address passed to alloc_page_vma is used to | |
618 | * generate the proper interleave behavior. We fake | |
619 | * the address here by an increasing offset in order | |
620 | * to get the proper distribution of pages. | |
621 | * | |
622 | * No decision has been made as to which page | |
623 | * a certain old page is moved to so we cannot | |
624 | * specify the correct address. | |
625 | */ | |
626 | page = alloc_page_vma(GFP_HIGHUSER, vma, | |
627 | offset + vma->vm_start); | |
628 | offset += PAGE_SIZE; | |
629 | } | |
630 | else | |
631 | page = alloc_pages_node(dest, GFP_HIGHUSER, 0); | |
632 | ||
633 | if (!page) { | |
634 | err = -ENOMEM; | |
635 | goto out; | |
636 | } | |
637 | list_add_tail(&page->lru, &newlist); | |
638 | nr_pages++; | |
639 | if (nr_pages > MIGRATE_CHUNK_SIZE) | |
640 | break; | |
641 | } | |
642 | err = migrate_pages(pagelist, &newlist, &moved, &failed); | |
643 | ||
644 | putback_lru_pages(&moved); /* Call release pages instead ?? */ | |
645 | ||
646 | if (err >= 0 && list_empty(&newlist) && !list_empty(pagelist)) | |
647 | goto redo; | |
648 | out: | |
649 | /* Return leftover allocated pages */ | |
650 | while (!list_empty(&newlist)) { | |
651 | page = list_entry(newlist.next, struct page, lru); | |
652 | list_del(&page->lru); | |
653 | __free_page(page); | |
654 | } | |
655 | list_splice(&failed, pagelist); | |
656 | if (err < 0) | |
657 | return err; | |
658 | ||
659 | /* Calculate number of leftover pages */ | |
660 | nr_pages = 0; | |
661 | list_for_each(p, pagelist) | |
662 | nr_pages++; | |
663 | return nr_pages; | |
664 | } |