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 | /* | |
c3fcf8a5 | 169 | * Replace the page in the mapping. |
5b5c7120 CL |
170 | * |
171 | * The number of remaining references must be: | |
172 | * 1 for anonymous pages without a mapping | |
173 | * 2 for pages with a mapping | |
174 | * 3 for pages with a mapping and PagePrivate set. | |
b20a3503 | 175 | */ |
2d1db3b1 CL |
176 | static int migrate_page_move_mapping(struct address_space *mapping, |
177 | struct page *newpage, struct page *page) | |
b20a3503 | 178 | { |
b20a3503 CL |
179 | struct page **radix_pointer; |
180 | ||
b20a3503 CL |
181 | write_lock_irq(&mapping->tree_lock); |
182 | ||
183 | radix_pointer = (struct page **)radix_tree_lookup_slot( | |
184 | &mapping->page_tree, | |
185 | page_index(page)); | |
186 | ||
5b5c7120 CL |
187 | if (!page_mapping(page) || |
188 | page_count(page) != 2 + !!PagePrivate(page) || | |
b20a3503 CL |
189 | *radix_pointer != page) { |
190 | write_unlock_irq(&mapping->tree_lock); | |
e23ca00b | 191 | return -EAGAIN; |
b20a3503 CL |
192 | } |
193 | ||
194 | /* | |
195 | * Now we know that no one else is looking at the page. | |
b20a3503 CL |
196 | */ |
197 | get_page(newpage); | |
b20a3503 CL |
198 | if (PageSwapCache(page)) { |
199 | SetPageSwapCache(newpage); | |
200 | set_page_private(newpage, page_private(page)); | |
201 | } | |
202 | ||
203 | *radix_pointer = newpage; | |
204 | __put_page(page); | |
205 | write_unlock_irq(&mapping->tree_lock); | |
206 | ||
207 | return 0; | |
208 | } | |
b20a3503 CL |
209 | |
210 | /* | |
211 | * Copy the page to its new location | |
212 | */ | |
e7340f73 | 213 | static void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 CL |
214 | { |
215 | copy_highpage(newpage, page); | |
216 | ||
217 | if (PageError(page)) | |
218 | SetPageError(newpage); | |
219 | if (PageReferenced(page)) | |
220 | SetPageReferenced(newpage); | |
221 | if (PageUptodate(page)) | |
222 | SetPageUptodate(newpage); | |
223 | if (PageActive(page)) | |
224 | SetPageActive(newpage); | |
225 | if (PageChecked(page)) | |
226 | SetPageChecked(newpage); | |
227 | if (PageMappedToDisk(page)) | |
228 | SetPageMappedToDisk(newpage); | |
229 | ||
230 | if (PageDirty(page)) { | |
231 | clear_page_dirty_for_io(page); | |
232 | set_page_dirty(newpage); | |
233 | } | |
234 | ||
235 | ClearPageSwapCache(page); | |
236 | ClearPageActive(page); | |
237 | ClearPagePrivate(page); | |
238 | set_page_private(page, 0); | |
239 | page->mapping = NULL; | |
240 | ||
241 | /* | |
242 | * If any waiters have accumulated on the new page then | |
243 | * wake them up. | |
244 | */ | |
245 | if (PageWriteback(newpage)) | |
246 | end_page_writeback(newpage); | |
247 | } | |
b20a3503 | 248 | |
1d8b85cc CL |
249 | /************************************************************ |
250 | * Migration functions | |
251 | ***********************************************************/ | |
252 | ||
253 | /* Always fail migration. Used for mappings that are not movable */ | |
2d1db3b1 CL |
254 | int fail_migrate_page(struct address_space *mapping, |
255 | struct page *newpage, struct page *page) | |
1d8b85cc CL |
256 | { |
257 | return -EIO; | |
258 | } | |
259 | EXPORT_SYMBOL(fail_migrate_page); | |
260 | ||
b20a3503 CL |
261 | /* |
262 | * Common logic to directly migrate a single page suitable for | |
263 | * pages that do not use PagePrivate. | |
264 | * | |
265 | * Pages are locked upon entry and exit. | |
266 | */ | |
2d1db3b1 CL |
267 | int migrate_page(struct address_space *mapping, |
268 | struct page *newpage, struct page *page) | |
b20a3503 CL |
269 | { |
270 | int rc; | |
271 | ||
272 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
273 | ||
2d1db3b1 | 274 | rc = migrate_page_move_mapping(mapping, newpage, page); |
b20a3503 CL |
275 | |
276 | if (rc) | |
277 | return rc; | |
278 | ||
279 | migrate_page_copy(newpage, page); | |
280 | ||
281 | /* | |
282 | * Remove auxiliary swap entries and replace | |
283 | * them with real ptes. | |
284 | * | |
285 | * Note that a real pte entry will allow processes that are not | |
286 | * waiting on the page lock to use the new page via the page tables | |
287 | * before the new page is unlocked. | |
288 | */ | |
289 | remove_from_swap(newpage); | |
290 | return 0; | |
291 | } | |
292 | EXPORT_SYMBOL(migrate_page); | |
293 | ||
1d8b85cc CL |
294 | /* |
295 | * Migration function for pages with buffers. This function can only be used | |
296 | * if the underlying filesystem guarantees that no other references to "page" | |
297 | * exist. | |
298 | */ | |
2d1db3b1 CL |
299 | int buffer_migrate_page(struct address_space *mapping, |
300 | struct page *newpage, struct page *page) | |
1d8b85cc | 301 | { |
1d8b85cc CL |
302 | struct buffer_head *bh, *head; |
303 | int rc; | |
304 | ||
1d8b85cc | 305 | if (!page_has_buffers(page)) |
2d1db3b1 | 306 | return migrate_page(mapping, newpage, page); |
1d8b85cc CL |
307 | |
308 | head = page_buffers(page); | |
309 | ||
2d1db3b1 | 310 | rc = migrate_page_move_mapping(mapping, newpage, page); |
1d8b85cc CL |
311 | |
312 | if (rc) | |
313 | return rc; | |
314 | ||
315 | bh = head; | |
316 | do { | |
317 | get_bh(bh); | |
318 | lock_buffer(bh); | |
319 | bh = bh->b_this_page; | |
320 | ||
321 | } while (bh != head); | |
322 | ||
323 | ClearPagePrivate(page); | |
324 | set_page_private(newpage, page_private(page)); | |
325 | set_page_private(page, 0); | |
326 | put_page(page); | |
327 | get_page(newpage); | |
328 | ||
329 | bh = head; | |
330 | do { | |
331 | set_bh_page(bh, newpage, bh_offset(bh)); | |
332 | bh = bh->b_this_page; | |
333 | ||
334 | } while (bh != head); | |
335 | ||
336 | SetPagePrivate(newpage); | |
337 | ||
338 | migrate_page_copy(newpage, page); | |
339 | ||
340 | bh = head; | |
341 | do { | |
342 | unlock_buffer(bh); | |
343 | put_bh(bh); | |
344 | bh = bh->b_this_page; | |
345 | ||
346 | } while (bh != head); | |
347 | ||
348 | return 0; | |
349 | } | |
350 | EXPORT_SYMBOL(buffer_migrate_page); | |
351 | ||
8351a6e4 CL |
352 | static int fallback_migrate_page(struct address_space *mapping, |
353 | struct page *newpage, struct page *page) | |
354 | { | |
355 | /* | |
356 | * Default handling if a filesystem does not provide | |
357 | * a migration function. We can only migrate clean | |
358 | * pages so try to write out any dirty pages first. | |
359 | */ | |
360 | if (PageDirty(page)) { | |
361 | switch (pageout(page, mapping)) { | |
362 | case PAGE_KEEP: | |
363 | case PAGE_ACTIVATE: | |
364 | return -EAGAIN; | |
365 | ||
366 | case PAGE_SUCCESS: | |
367 | /* Relock since we lost the lock */ | |
368 | lock_page(page); | |
369 | /* Must retry since page state may have changed */ | |
370 | return -EAGAIN; | |
371 | ||
372 | case PAGE_CLEAN: | |
373 | ; /* try to migrate the page below */ | |
374 | } | |
375 | } | |
376 | ||
377 | /* | |
378 | * Buffers may be managed in a filesystem specific way. | |
379 | * We must have no buffers or drop them. | |
380 | */ | |
381 | if (page_has_buffers(page) && | |
382 | !try_to_release_page(page, GFP_KERNEL)) | |
383 | return -EAGAIN; | |
384 | ||
385 | return migrate_page(mapping, newpage, page); | |
386 | } | |
387 | ||
b20a3503 CL |
388 | /* |
389 | * migrate_pages | |
390 | * | |
391 | * Two lists are passed to this function. The first list | |
392 | * contains the pages isolated from the LRU to be migrated. | |
393 | * The second list contains new pages that the pages isolated | |
394 | * can be moved to. If the second list is NULL then all | |
395 | * pages are swapped out. | |
396 | * | |
397 | * The function returns after 10 attempts or if no pages | |
398 | * are movable anymore because to has become empty | |
399 | * or no retryable pages exist anymore. | |
400 | * | |
401 | * Return: Number of pages not migrated when "to" ran empty. | |
402 | */ | |
403 | int migrate_pages(struct list_head *from, struct list_head *to, | |
404 | struct list_head *moved, struct list_head *failed) | |
405 | { | |
406 | int retry; | |
407 | int nr_failed = 0; | |
408 | int pass = 0; | |
409 | struct page *page; | |
410 | struct page *page2; | |
411 | int swapwrite = current->flags & PF_SWAPWRITE; | |
412 | int rc; | |
413 | ||
414 | if (!swapwrite) | |
415 | current->flags |= PF_SWAPWRITE; | |
416 | ||
417 | redo: | |
418 | retry = 0; | |
419 | ||
420 | list_for_each_entry_safe(page, page2, from, lru) { | |
421 | struct page *newpage = NULL; | |
422 | struct address_space *mapping; | |
423 | ||
424 | cond_resched(); | |
425 | ||
426 | rc = 0; | |
427 | if (page_count(page) == 1) | |
428 | /* page was freed from under us. So we are done. */ | |
429 | goto next; | |
430 | ||
431 | if (to && list_empty(to)) | |
432 | break; | |
433 | ||
434 | /* | |
435 | * Skip locked pages during the first two passes to give the | |
436 | * functions holding the lock time to release the page. Later we | |
437 | * use lock_page() to have a higher chance of acquiring the | |
438 | * lock. | |
439 | */ | |
440 | rc = -EAGAIN; | |
441 | if (pass > 2) | |
442 | lock_page(page); | |
443 | else | |
444 | if (TestSetPageLocked(page)) | |
445 | goto next; | |
446 | ||
447 | /* | |
448 | * Only wait on writeback if we have already done a pass where | |
449 | * we we may have triggered writeouts for lots of pages. | |
450 | */ | |
451 | if (pass > 0) { | |
452 | wait_on_page_writeback(page); | |
453 | } else { | |
454 | if (PageWriteback(page)) | |
455 | goto unlock_page; | |
456 | } | |
457 | ||
458 | /* | |
459 | * Anonymous pages must have swap cache references otherwise | |
460 | * the information contained in the page maps cannot be | |
461 | * preserved. | |
462 | */ | |
463 | if (PageAnon(page) && !PageSwapCache(page)) { | |
464 | if (!add_to_swap(page, GFP_KERNEL)) { | |
465 | rc = -ENOMEM; | |
466 | goto unlock_page; | |
467 | } | |
468 | } | |
469 | ||
470 | if (!to) { | |
471 | rc = swap_page(page); | |
472 | goto next; | |
473 | } | |
474 | ||
c3fcf8a5 CL |
475 | /* |
476 | * Establish swap ptes for anonymous pages or destroy pte | |
477 | * maps for files. | |
478 | * | |
479 | * In order to reestablish file backed mappings the fault handlers | |
480 | * will take the radix tree_lock which may then be used to stop | |
481 | * processses from accessing this page until the new page is ready. | |
482 | * | |
483 | * A process accessing via a swap pte (an anonymous page) will take a | |
484 | * page_lock on the old page which will block the process until the | |
485 | * migration attempt is complete. At that time the PageSwapCache bit | |
486 | * will be examined. If the page was migrated then the PageSwapCache | |
487 | * bit will be clear and the operation to retrieve the page will be | |
488 | * retried which will find the new page in the radix tree. Then a new | |
489 | * direct mapping may be generated based on the radix tree contents. | |
490 | * | |
491 | * If the page was not migrated then the PageSwapCache bit | |
492 | * is still set and the operation may continue. | |
493 | */ | |
494 | rc = -EPERM; | |
495 | if (try_to_unmap(page, 1) == SWAP_FAIL) | |
496 | /* A vma has VM_LOCKED set -> permanent failure */ | |
2d1db3b1 | 497 | goto unlock_page; |
c3fcf8a5 CL |
498 | |
499 | rc = -EAGAIN; | |
500 | if (page_mapped(page)) | |
2d1db3b1 CL |
501 | goto unlock_page; |
502 | ||
503 | newpage = lru_to_page(to); | |
504 | lock_page(newpage); | |
505 | /* Prepare mapping for the new page.*/ | |
506 | newpage->index = page->index; | |
507 | newpage->mapping = page->mapping; | |
508 | ||
b20a3503 CL |
509 | /* |
510 | * Pages are properly locked and writeback is complete. | |
511 | * Try to migrate the page. | |
512 | */ | |
513 | mapping = page_mapping(page); | |
514 | if (!mapping) | |
515 | goto unlock_both; | |
516 | ||
8351a6e4 | 517 | if (mapping->a_ops->migratepage) |
b20a3503 CL |
518 | /* |
519 | * Most pages have a mapping and most filesystems | |
520 | * should provide a migration function. Anonymous | |
521 | * pages are part of swap space which also has its | |
522 | * own migration function. This is the most common | |
523 | * path for page migration. | |
524 | */ | |
2d1db3b1 CL |
525 | rc = mapping->a_ops->migratepage(mapping, |
526 | newpage, page); | |
8351a6e4 CL |
527 | else |
528 | rc = fallback_migrate_page(mapping, newpage, page); | |
b20a3503 CL |
529 | |
530 | unlock_both: | |
531 | unlock_page(newpage); | |
532 | ||
533 | unlock_page: | |
534 | unlock_page(page); | |
535 | ||
536 | next: | |
2d1db3b1 CL |
537 | if (rc) { |
538 | if (newpage) | |
539 | newpage->mapping = NULL; | |
540 | ||
541 | if (rc == -EAGAIN) | |
542 | retry++; | |
543 | else { | |
544 | /* Permanent failure */ | |
545 | list_move(&page->lru, failed); | |
546 | nr_failed++; | |
547 | } | |
b20a3503 CL |
548 | } else { |
549 | if (newpage) { | |
550 | /* Successful migration. Return page to LRU */ | |
551 | move_to_lru(newpage); | |
552 | } | |
553 | list_move(&page->lru, moved); | |
554 | } | |
555 | } | |
556 | if (retry && pass++ < 10) | |
557 | goto redo; | |
558 | ||
559 | if (!swapwrite) | |
560 | current->flags &= ~PF_SWAPWRITE; | |
561 | ||
562 | return nr_failed + retry; | |
563 | } | |
564 | ||
b20a3503 CL |
565 | /* |
566 | * Migrate the list 'pagelist' of pages to a certain destination. | |
567 | * | |
568 | * Specify destination with either non-NULL vma or dest_node >= 0 | |
569 | * Return the number of pages not migrated or error code | |
570 | */ | |
571 | int migrate_pages_to(struct list_head *pagelist, | |
572 | struct vm_area_struct *vma, int dest) | |
573 | { | |
574 | LIST_HEAD(newlist); | |
575 | LIST_HEAD(moved); | |
576 | LIST_HEAD(failed); | |
577 | int err = 0; | |
578 | unsigned long offset = 0; | |
579 | int nr_pages; | |
580 | struct page *page; | |
581 | struct list_head *p; | |
582 | ||
583 | redo: | |
584 | nr_pages = 0; | |
585 | list_for_each(p, pagelist) { | |
586 | if (vma) { | |
587 | /* | |
588 | * The address passed to alloc_page_vma is used to | |
589 | * generate the proper interleave behavior. We fake | |
590 | * the address here by an increasing offset in order | |
591 | * to get the proper distribution of pages. | |
592 | * | |
593 | * No decision has been made as to which page | |
594 | * a certain old page is moved to so we cannot | |
595 | * specify the correct address. | |
596 | */ | |
597 | page = alloc_page_vma(GFP_HIGHUSER, vma, | |
598 | offset + vma->vm_start); | |
599 | offset += PAGE_SIZE; | |
600 | } | |
601 | else | |
602 | page = alloc_pages_node(dest, GFP_HIGHUSER, 0); | |
603 | ||
604 | if (!page) { | |
605 | err = -ENOMEM; | |
606 | goto out; | |
607 | } | |
608 | list_add_tail(&page->lru, &newlist); | |
609 | nr_pages++; | |
610 | if (nr_pages > MIGRATE_CHUNK_SIZE) | |
611 | break; | |
612 | } | |
613 | err = migrate_pages(pagelist, &newlist, &moved, &failed); | |
614 | ||
615 | putback_lru_pages(&moved); /* Call release pages instead ?? */ | |
616 | ||
617 | if (err >= 0 && list_empty(&newlist) && !list_empty(pagelist)) | |
618 | goto redo; | |
619 | out: | |
620 | /* Return leftover allocated pages */ | |
621 | while (!list_empty(&newlist)) { | |
622 | page = list_entry(newlist.next, struct page, lru); | |
623 | list_del(&page->lru); | |
624 | __free_page(page); | |
625 | } | |
626 | list_splice(&failed, pagelist); | |
627 | if (err < 0) | |
628 | return err; | |
629 | ||
630 | /* Calculate number of leftover pages */ | |
631 | nr_pages = 0; | |
632 | list_for_each(p, pagelist) | |
633 | nr_pages++; | |
634 | return nr_pages; | |
635 | } |