blob: be3f141e53a42a4aad9d4b5678f06a3f343154d2 [file] [log] [blame]
Christoph Lameterb20a3502006-03-22 00:09:12 -08001/*
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>
Christoph Lametere23ca002006-04-10 22:52:57 -070019#include <linux/buffer_head.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080020#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
Christoph Lameterb20a3502006-03-22 00:09:12 -080030/* 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 */
43int 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 */
71int 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
88static 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 */
109int 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
122/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800123 * swapout a single page
124 * page is locked upon entry, unlocked on exit
125 */
126static 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
161unlock_retry:
162 unlock_page(page);
163
164retry:
165 return -EAGAIN;
166}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800167
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 */
Christoph Lametere7340f72006-06-23 02:03:29 -0700172static int migrate_page_remove_references(struct page *newpage,
Christoph Lameterb20a3502006-03-22 00:09:12 -0800173 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);
Christoph Lametere23ca002006-04-10 22:52:57 -0700224 return -EAGAIN;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800225 }
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}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800249
250/*
251 * Copy the page to its new location
252 */
Christoph Lametere7340f72006-06-23 02:03:29 -0700253static void migrate_page_copy(struct page *newpage, struct page *page)
Christoph Lameterb20a3502006-03-22 00:09:12 -0800254{
255 copy_highpage(newpage, page);
256
257 if (PageError(page))
258 SetPageError(newpage);
259 if (PageReferenced(page))
260 SetPageReferenced(newpage);
261 if (PageUptodate(page))
262 SetPageUptodate(newpage);
263 if (PageActive(page))
264 SetPageActive(newpage);
265 if (PageChecked(page))
266 SetPageChecked(newpage);
267 if (PageMappedToDisk(page))
268 SetPageMappedToDisk(newpage);
269
270 if (PageDirty(page)) {
271 clear_page_dirty_for_io(page);
272 set_page_dirty(newpage);
273 }
274
275 ClearPageSwapCache(page);
276 ClearPageActive(page);
277 ClearPagePrivate(page);
278 set_page_private(page, 0);
279 page->mapping = NULL;
280
281 /*
282 * If any waiters have accumulated on the new page then
283 * wake them up.
284 */
285 if (PageWriteback(newpage))
286 end_page_writeback(newpage);
287}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800288
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700289/************************************************************
290 * Migration functions
291 ***********************************************************/
292
293/* Always fail migration. Used for mappings that are not movable */
294int fail_migrate_page(struct page *newpage, struct page *page)
295{
296 return -EIO;
297}
298EXPORT_SYMBOL(fail_migrate_page);
299
Christoph Lameterb20a3502006-03-22 00:09:12 -0800300/*
301 * Common logic to directly migrate a single page suitable for
302 * pages that do not use PagePrivate.
303 *
304 * Pages are locked upon entry and exit.
305 */
306int migrate_page(struct page *newpage, struct page *page)
307{
308 int rc;
309
310 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
311
312 rc = migrate_page_remove_references(newpage, page, 2);
313
314 if (rc)
315 return rc;
316
317 migrate_page_copy(newpage, page);
318
319 /*
320 * Remove auxiliary swap entries and replace
321 * them with real ptes.
322 *
323 * Note that a real pte entry will allow processes that are not
324 * waiting on the page lock to use the new page via the page tables
325 * before the new page is unlocked.
326 */
327 remove_from_swap(newpage);
328 return 0;
329}
330EXPORT_SYMBOL(migrate_page);
331
332/*
Christoph Lameter1d8b85c2006-06-23 02:03:28 -0700333 * Migration function for pages with buffers. This function can only be used
334 * if the underlying filesystem guarantees that no other references to "page"
335 * exist.
336 */
337int buffer_migrate_page(struct page *newpage, struct page *page)
338{
339 struct address_space *mapping = page->mapping;
340 struct buffer_head *bh, *head;
341 int rc;
342
343 if (!mapping)
344 return -EAGAIN;
345
346 if (!page_has_buffers(page))
347 return migrate_page(newpage, page);
348
349 head = page_buffers(page);
350
351 rc = migrate_page_remove_references(newpage, page, 3);
352
353 if (rc)
354 return rc;
355
356 bh = head;
357 do {
358 get_bh(bh);
359 lock_buffer(bh);
360 bh = bh->b_this_page;
361
362 } while (bh != head);
363
364 ClearPagePrivate(page);
365 set_page_private(newpage, page_private(page));
366 set_page_private(page, 0);
367 put_page(page);
368 get_page(newpage);
369
370 bh = head;
371 do {
372 set_bh_page(bh, newpage, bh_offset(bh));
373 bh = bh->b_this_page;
374
375 } while (bh != head);
376
377 SetPagePrivate(newpage);
378
379 migrate_page_copy(newpage, page);
380
381 bh = head;
382 do {
383 unlock_buffer(bh);
384 put_bh(bh);
385 bh = bh->b_this_page;
386
387 } while (bh != head);
388
389 return 0;
390}
391EXPORT_SYMBOL(buffer_migrate_page);
392
393/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800394 * migrate_pages
395 *
396 * Two lists are passed to this function. The first list
397 * contains the pages isolated from the LRU to be migrated.
398 * The second list contains new pages that the pages isolated
399 * can be moved to. If the second list is NULL then all
400 * pages are swapped out.
401 *
402 * The function returns after 10 attempts or if no pages
403 * are movable anymore because to has become empty
404 * or no retryable pages exist anymore.
405 *
406 * Return: Number of pages not migrated when "to" ran empty.
407 */
408int migrate_pages(struct list_head *from, struct list_head *to,
409 struct list_head *moved, struct list_head *failed)
410{
411 int retry;
412 int nr_failed = 0;
413 int pass = 0;
414 struct page *page;
415 struct page *page2;
416 int swapwrite = current->flags & PF_SWAPWRITE;
417 int rc;
418
419 if (!swapwrite)
420 current->flags |= PF_SWAPWRITE;
421
422redo:
423 retry = 0;
424
425 list_for_each_entry_safe(page, page2, from, lru) {
426 struct page *newpage = NULL;
427 struct address_space *mapping;
428
429 cond_resched();
430
431 rc = 0;
432 if (page_count(page) == 1)
433 /* page was freed from under us. So we are done. */
434 goto next;
435
436 if (to && list_empty(to))
437 break;
438
439 /*
440 * Skip locked pages during the first two passes to give the
441 * functions holding the lock time to release the page. Later we
442 * use lock_page() to have a higher chance of acquiring the
443 * lock.
444 */
445 rc = -EAGAIN;
446 if (pass > 2)
447 lock_page(page);
448 else
449 if (TestSetPageLocked(page))
450 goto next;
451
452 /*
453 * Only wait on writeback if we have already done a pass where
454 * we we may have triggered writeouts for lots of pages.
455 */
456 if (pass > 0) {
457 wait_on_page_writeback(page);
458 } else {
459 if (PageWriteback(page))
460 goto unlock_page;
461 }
462
463 /*
464 * Anonymous pages must have swap cache references otherwise
465 * the information contained in the page maps cannot be
466 * preserved.
467 */
468 if (PageAnon(page) && !PageSwapCache(page)) {
469 if (!add_to_swap(page, GFP_KERNEL)) {
470 rc = -ENOMEM;
471 goto unlock_page;
472 }
473 }
474
475 if (!to) {
476 rc = swap_page(page);
477 goto next;
478 }
479
480 newpage = lru_to_page(to);
481 lock_page(newpage);
482
483 /*
484 * Pages are properly locked and writeback is complete.
485 * Try to migrate the page.
486 */
487 mapping = page_mapping(page);
488 if (!mapping)
489 goto unlock_both;
490
491 if (mapping->a_ops->migratepage) {
492 /*
493 * Most pages have a mapping and most filesystems
494 * should provide a migration function. Anonymous
495 * pages are part of swap space which also has its
496 * own migration function. This is the most common
497 * path for page migration.
498 */
499 rc = mapping->a_ops->migratepage(newpage, page);
500 goto unlock_both;
501 }
502
Christoph Lameter4c28f812006-05-01 12:16:08 -0700503 /* Make sure the dirty bit is up to date */
504 if (try_to_unmap(page, 1) == SWAP_FAIL) {
505 rc = -EPERM;
506 goto unlock_both;
507 }
508
509 if (page_mapcount(page)) {
510 rc = -EAGAIN;
511 goto unlock_both;
512 }
513
Christoph Lameterb20a3502006-03-22 00:09:12 -0800514 /*
515 * Default handling if a filesystem does not provide
516 * a migration function. We can only migrate clean
517 * pages so try to write out any dirty pages first.
518 */
519 if (PageDirty(page)) {
520 switch (pageout(page, mapping)) {
521 case PAGE_KEEP:
522 case PAGE_ACTIVATE:
523 goto unlock_both;
524
525 case PAGE_SUCCESS:
526 unlock_page(newpage);
527 goto next;
528
529 case PAGE_CLEAN:
530 ; /* try to migrate the page below */
531 }
532 }
533
534 /*
535 * Buffers are managed in a filesystem specific way.
536 * We must have no buffers or drop them.
537 */
538 if (!page_has_buffers(page) ||
539 try_to_release_page(page, GFP_KERNEL)) {
540 rc = migrate_page(newpage, page);
541 goto unlock_both;
542 }
543
544 /*
545 * On early passes with mapped pages simply
546 * retry. There may be a lock held for some
547 * buffers that may go away. Later
548 * swap them out.
549 */
550 if (pass > 4) {
551 /*
552 * Persistently unable to drop buffers..... As a
553 * measure of last resort we fall back to
554 * swap_page().
555 */
556 unlock_page(newpage);
557 newpage = NULL;
558 rc = swap_page(page);
559 goto next;
560 }
561
562unlock_both:
563 unlock_page(newpage);
564
565unlock_page:
566 unlock_page(page);
567
568next:
569 if (rc == -EAGAIN) {
570 retry++;
571 } else if (rc) {
572 /* Permanent failure */
573 list_move(&page->lru, failed);
574 nr_failed++;
575 } else {
576 if (newpage) {
577 /* Successful migration. Return page to LRU */
578 move_to_lru(newpage);
579 }
580 list_move(&page->lru, moved);
581 }
582 }
583 if (retry && pass++ < 10)
584 goto redo;
585
586 if (!swapwrite)
587 current->flags &= ~PF_SWAPWRITE;
588
589 return nr_failed + retry;
590}
591
592/*
Christoph Lameterb20a3502006-03-22 00:09:12 -0800593 * Migrate the list 'pagelist' of pages to a certain destination.
594 *
595 * Specify destination with either non-NULL vma or dest_node >= 0
596 * Return the number of pages not migrated or error code
597 */
598int migrate_pages_to(struct list_head *pagelist,
599 struct vm_area_struct *vma, int dest)
600{
601 LIST_HEAD(newlist);
602 LIST_HEAD(moved);
603 LIST_HEAD(failed);
604 int err = 0;
605 unsigned long offset = 0;
606 int nr_pages;
607 struct page *page;
608 struct list_head *p;
609
610redo:
611 nr_pages = 0;
612 list_for_each(p, pagelist) {
613 if (vma) {
614 /*
615 * The address passed to alloc_page_vma is used to
616 * generate the proper interleave behavior. We fake
617 * the address here by an increasing offset in order
618 * to get the proper distribution of pages.
619 *
620 * No decision has been made as to which page
621 * a certain old page is moved to so we cannot
622 * specify the correct address.
623 */
624 page = alloc_page_vma(GFP_HIGHUSER, vma,
625 offset + vma->vm_start);
626 offset += PAGE_SIZE;
627 }
628 else
629 page = alloc_pages_node(dest, GFP_HIGHUSER, 0);
630
631 if (!page) {
632 err = -ENOMEM;
633 goto out;
634 }
635 list_add_tail(&page->lru, &newlist);
636 nr_pages++;
637 if (nr_pages > MIGRATE_CHUNK_SIZE)
638 break;
639 }
640 err = migrate_pages(pagelist, &newlist, &moved, &failed);
641
642 putback_lru_pages(&moved); /* Call release pages instead ?? */
643
644 if (err >= 0 && list_empty(&newlist) && !list_empty(pagelist))
645 goto redo;
646out:
647 /* Return leftover allocated pages */
648 while (!list_empty(&newlist)) {
649 page = list_entry(newlist.next, struct page, lru);
650 list_del(&page->lru);
651 __free_page(page);
652 }
653 list_splice(&failed, pagelist);
654 if (err < 0)
655 return err;
656
657 /* Calculate number of leftover pages */
658 nr_pages = 0;
659 list_for_each(p, pagelist)
660 nr_pages++;
661 return nr_pages;
662}