2 * Memory Migration functionality - linux/mm/migration.c
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
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>
15 #include <linux/migrate.h>
16 #include <linux/module.h>
17 #include <linux/swap.h>
18 #include <linux/pagemap.h>
19 #include <linux/buffer_head.h>
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>
30 /* The maximum number of pages to take off the LRU for migration */
31 #define MIGRATE_CHUNK_SIZE 256
33 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
36 * Isolate one page from the LRU lists. If successful put it onto
37 * the indicated list with elevated page count.
40 * -EBUSY: page not on LRU list
41 * 0: page removed from LRU list and added to the specified list.
43 int isolate_lru_page(struct page *page, struct list_head *pagelist)
48 struct zone *zone = page_zone(page);
50 spin_lock_irq(&zone->lru_lock);
56 del_page_from_active_list(zone, page);
58 del_page_from_inactive_list(zone, page);
59 list_add_tail(&page->lru, pagelist);
61 spin_unlock_irq(&zone->lru_lock);
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
71 int migrate_prep(void)
73 /* Must have swap device for migration */
74 if (nr_swap_pages <= 0)
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.
88 static inline void move_to_lru(struct page *page)
91 if (PageActive(page)) {
93 * lru_cache_add_active checks that
94 * the PG_active bit is off.
96 ClearPageActive(page);
97 lru_cache_add_active(page);
105 * Add isolated pages on the list back to the LRU.
107 * returns the number of pages put back.
109 int putback_lru_pages(struct list_head *l)
115 list_for_each_entry_safe(page, page2, l, lru) {
123 * Non migratable page
125 int fail_migrate_page(struct page *newpage, struct page *page)
129 EXPORT_SYMBOL(fail_migrate_page);
132 * swapout a single page
133 * page is locked upon entry, unlocked on exit
135 static int swap_page(struct page *page)
137 struct address_space *mapping = page_mapping(page);
139 if (page_mapped(page) && mapping)
140 if (try_to_unmap(page, 1) != SWAP_SUCCESS)
143 if (PageDirty(page)) {
144 /* Page is dirty, try to write it out here */
145 switch(pageout(page, mapping)) {
154 ; /* try to free the page below */
158 if (PagePrivate(page)) {
159 if (!try_to_release_page(page, GFP_KERNEL) ||
160 (!mapping && page_count(page) == 1))
164 if (remove_mapping(mapping, page)) {
178 * Remove references for a page and establish the new page with the correct
179 * basic settings to be able to stop accesses to the page.
181 int migrate_page_remove_references(struct page *newpage,
182 struct page *page, int nr_refs)
184 struct address_space *mapping = page_mapping(page);
185 struct page **radix_pointer;
188 * Avoid doing any of the following work if the page count
189 * indicates that the page is in use or truncate has removed
192 if (!mapping || page_mapcount(page) + nr_refs != page_count(page))
196 * Establish swap ptes for anonymous pages or destroy pte
199 * In order to reestablish file backed mappings the fault handlers
200 * will take the radix tree_lock which may then be used to stop
201 * processses from accessing this page until the new page is ready.
203 * A process accessing via a swap pte (an anonymous page) will take a
204 * page_lock on the old page which will block the process until the
205 * migration attempt is complete. At that time the PageSwapCache bit
206 * will be examined. If the page was migrated then the PageSwapCache
207 * bit will be clear and the operation to retrieve the page will be
208 * retried which will find the new page in the radix tree. Then a new
209 * direct mapping may be generated based on the radix tree contents.
211 * If the page was not migrated then the PageSwapCache bit
212 * is still set and the operation may continue.
214 if (try_to_unmap(page, 1) == SWAP_FAIL)
215 /* A vma has VM_LOCKED set -> permanent failure */
219 * Give up if we were unable to remove all mappings.
221 if (page_mapcount(page))
224 write_lock_irq(&mapping->tree_lock);
226 radix_pointer = (struct page **)radix_tree_lookup_slot(
230 if (!page_mapping(page) || page_count(page) != nr_refs ||
231 *radix_pointer != page) {
232 write_unlock_irq(&mapping->tree_lock);
237 * Now we know that no one else is looking at the page.
239 * Certain minimal information about a page must be available
240 * in order for other subsystems to properly handle the page if they
241 * find it through the radix tree update before we are finished
245 newpage->index = page->index;
246 newpage->mapping = page->mapping;
247 if (PageSwapCache(page)) {
248 SetPageSwapCache(newpage);
249 set_page_private(newpage, page_private(page));
252 *radix_pointer = newpage;
254 write_unlock_irq(&mapping->tree_lock);
258 EXPORT_SYMBOL(migrate_page_remove_references);
261 * Copy the page to its new location
263 void migrate_page_copy(struct page *newpage, struct page *page)
265 copy_highpage(newpage, page);
268 SetPageError(newpage);
269 if (PageReferenced(page))
270 SetPageReferenced(newpage);
271 if (PageUptodate(page))
272 SetPageUptodate(newpage);
273 if (PageActive(page))
274 SetPageActive(newpage);
275 if (PageChecked(page))
276 SetPageChecked(newpage);
277 if (PageMappedToDisk(page))
278 SetPageMappedToDisk(newpage);
280 if (PageDirty(page)) {
281 clear_page_dirty_for_io(page);
282 set_page_dirty(newpage);
285 ClearPageSwapCache(page);
286 ClearPageActive(page);
287 ClearPagePrivate(page);
288 set_page_private(page, 0);
289 page->mapping = NULL;
292 * If any waiters have accumulated on the new page then
295 if (PageWriteback(newpage))
296 end_page_writeback(newpage);
298 EXPORT_SYMBOL(migrate_page_copy);
301 * Common logic to directly migrate a single page suitable for
302 * pages that do not use PagePrivate.
304 * Pages are locked upon entry and exit.
306 int migrate_page(struct page *newpage, struct page *page)
310 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
312 rc = migrate_page_remove_references(newpage, page, 2);
317 migrate_page_copy(newpage, page);
320 * Remove auxiliary swap entries and replace
321 * them with real ptes.
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.
327 remove_from_swap(newpage);
330 EXPORT_SYMBOL(migrate_page);
335 * Two lists are passed to this function. The first list
336 * contains the pages isolated from the LRU to be migrated.
337 * The second list contains new pages that the pages isolated
338 * can be moved to. If the second list is NULL then all
339 * pages are swapped out.
341 * The function returns after 10 attempts or if no pages
342 * are movable anymore because to has become empty
343 * or no retryable pages exist anymore.
345 * Return: Number of pages not migrated when "to" ran empty.
347 int migrate_pages(struct list_head *from, struct list_head *to,
348 struct list_head *moved, struct list_head *failed)
355 int swapwrite = current->flags & PF_SWAPWRITE;
359 current->flags |= PF_SWAPWRITE;
364 list_for_each_entry_safe(page, page2, from, lru) {
365 struct page *newpage = NULL;
366 struct address_space *mapping;
371 if (page_count(page) == 1)
372 /* page was freed from under us. So we are done. */
375 if (to && list_empty(to))
379 * Skip locked pages during the first two passes to give the
380 * functions holding the lock time to release the page. Later we
381 * use lock_page() to have a higher chance of acquiring the
388 if (TestSetPageLocked(page))
392 * Only wait on writeback if we have already done a pass where
393 * we we may have triggered writeouts for lots of pages.
396 wait_on_page_writeback(page);
398 if (PageWriteback(page))
403 * Anonymous pages must have swap cache references otherwise
404 * the information contained in the page maps cannot be
407 if (PageAnon(page) && !PageSwapCache(page)) {
408 if (!add_to_swap(page, GFP_KERNEL)) {
415 rc = swap_page(page);
419 newpage = lru_to_page(to);
423 * Pages are properly locked and writeback is complete.
424 * Try to migrate the page.
426 mapping = page_mapping(page);
430 if (mapping->a_ops->migratepage) {
432 * Most pages have a mapping and most filesystems
433 * should provide a migration function. Anonymous
434 * pages are part of swap space which also has its
435 * own migration function. This is the most common
436 * path for page migration.
438 rc = mapping->a_ops->migratepage(newpage, page);
442 /* Make sure the dirty bit is up to date */
443 if (try_to_unmap(page, 1) == SWAP_FAIL) {
448 if (page_mapcount(page)) {
454 * Default handling if a filesystem does not provide
455 * a migration function. We can only migrate clean
456 * pages so try to write out any dirty pages first.
458 if (PageDirty(page)) {
459 switch (pageout(page, mapping)) {
465 unlock_page(newpage);
469 ; /* try to migrate the page below */
474 * Buffers are managed in a filesystem specific way.
475 * We must have no buffers or drop them.
477 if (!page_has_buffers(page) ||
478 try_to_release_page(page, GFP_KERNEL)) {
479 rc = migrate_page(newpage, page);
484 * On early passes with mapped pages simply
485 * retry. There may be a lock held for some
486 * buffers that may go away. Later
491 * Persistently unable to drop buffers..... As a
492 * measure of last resort we fall back to
495 unlock_page(newpage);
497 rc = swap_page(page);
502 unlock_page(newpage);
511 /* Permanent failure */
512 list_move(&page->lru, failed);
516 /* Successful migration. Return page to LRU */
517 move_to_lru(newpage);
519 list_move(&page->lru, moved);
522 if (retry && pass++ < 10)
526 current->flags &= ~PF_SWAPWRITE;
528 return nr_failed + retry;
532 * Migration function for pages with buffers. This function can only be used
533 * if the underlying filesystem guarantees that no other references to "page"
536 int buffer_migrate_page(struct page *newpage, struct page *page)
538 struct address_space *mapping = page->mapping;
539 struct buffer_head *bh, *head;
545 if (!page_has_buffers(page))
546 return migrate_page(newpage, page);
548 head = page_buffers(page);
550 rc = migrate_page_remove_references(newpage, page, 3);
559 bh = bh->b_this_page;
561 } while (bh != head);
563 ClearPagePrivate(page);
564 set_page_private(newpage, page_private(page));
565 set_page_private(page, 0);
571 set_bh_page(bh, newpage, bh_offset(bh));
572 bh = bh->b_this_page;
574 } while (bh != head);
576 SetPagePrivate(newpage);
578 migrate_page_copy(newpage, page);
584 bh = bh->b_this_page;
586 } while (bh != head);
590 EXPORT_SYMBOL(buffer_migrate_page);
593 * Migrate the list 'pagelist' of pages to a certain destination.
595 * Specify destination with either non-NULL vma or dest_node >= 0
596 * Return the number of pages not migrated or error code
598 int migrate_pages_to(struct list_head *pagelist,
599 struct vm_area_struct *vma, int dest)
605 unsigned long offset = 0;
612 list_for_each(p, pagelist) {
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.
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.
624 page = alloc_page_vma(GFP_HIGHUSER, vma,
625 offset + vma->vm_start);
629 page = alloc_pages_node(dest, GFP_HIGHUSER, 0);
635 list_add_tail(&page->lru, &newlist);
637 if (nr_pages > MIGRATE_CHUNK_SIZE)
640 err = migrate_pages(pagelist, &newlist, &moved, &failed);
642 putback_lru_pages(&moved); /* Call release pages instead ?? */
644 if (err >= 0 && list_empty(&newlist) && !list_empty(pagelist))
647 /* Return leftover allocated pages */
648 while (!list_empty(&newlist)) {
649 page = list_entry(newlist.next, struct page, lru);
650 list_del(&page->lru);
653 list_splice(&failed, pagelist);
657 /* Calculate number of leftover pages */
659 list_for_each(p, pagelist)