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 * swapout a single page
124 * page is locked upon entry, unlocked on exit
126 static int swap_page(struct page *page)
128 struct address_space *mapping = page_mapping(page);
130 if (page_mapped(page) && mapping)
131 if (try_to_unmap(page, 1) != SWAP_SUCCESS)
134 if (PageDirty(page)) {
135 /* Page is dirty, try to write it out here */
136 switch(pageout(page, mapping)) {
145 ; /* try to free the page below */
149 if (PagePrivate(page)) {
150 if (!try_to_release_page(page, GFP_KERNEL) ||
151 (!mapping && page_count(page) == 1))
155 if (remove_mapping(mapping, page)) {
169 * Replace the page in the mapping.
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.
176 static int migrate_page_move_mapping(struct page *newpage,
179 struct address_space *mapping = page_mapping(page);
180 struct page **radix_pointer;
185 write_lock_irq(&mapping->tree_lock);
187 radix_pointer = (struct page **)radix_tree_lookup_slot(
191 if (!page_mapping(page) ||
192 page_count(page) != 2 + !!PagePrivate(page) ||
193 *radix_pointer != page) {
194 write_unlock_irq(&mapping->tree_lock);
199 * Now we know that no one else is looking at the page.
201 * Certain minimal information about a page must be available
202 * in order for other subsystems to properly handle the page if they
203 * find it through the radix tree update before we are finished
207 newpage->index = page->index;
208 newpage->mapping = page->mapping;
209 if (PageSwapCache(page)) {
210 SetPageSwapCache(newpage);
211 set_page_private(newpage, page_private(page));
214 *radix_pointer = newpage;
216 write_unlock_irq(&mapping->tree_lock);
222 * Copy the page to its new location
224 static void migrate_page_copy(struct page *newpage, struct page *page)
226 copy_highpage(newpage, page);
229 SetPageError(newpage);
230 if (PageReferenced(page))
231 SetPageReferenced(newpage);
232 if (PageUptodate(page))
233 SetPageUptodate(newpage);
234 if (PageActive(page))
235 SetPageActive(newpage);
236 if (PageChecked(page))
237 SetPageChecked(newpage);
238 if (PageMappedToDisk(page))
239 SetPageMappedToDisk(newpage);
241 if (PageDirty(page)) {
242 clear_page_dirty_for_io(page);
243 set_page_dirty(newpage);
246 ClearPageSwapCache(page);
247 ClearPageActive(page);
248 ClearPagePrivate(page);
249 set_page_private(page, 0);
250 page->mapping = NULL;
253 * If any waiters have accumulated on the new page then
256 if (PageWriteback(newpage))
257 end_page_writeback(newpage);
260 /************************************************************
261 * Migration functions
262 ***********************************************************/
264 /* Always fail migration. Used for mappings that are not movable */
265 int fail_migrate_page(struct page *newpage, struct page *page)
269 EXPORT_SYMBOL(fail_migrate_page);
272 * Common logic to directly migrate a single page suitable for
273 * pages that do not use PagePrivate.
275 * Pages are locked upon entry and exit.
277 int migrate_page(struct page *newpage, struct page *page)
281 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
283 rc = migrate_page_move_mapping(newpage, page);
288 migrate_page_copy(newpage, page);
291 * Remove auxiliary swap entries and replace
292 * them with real ptes.
294 * Note that a real pte entry will allow processes that are not
295 * waiting on the page lock to use the new page via the page tables
296 * before the new page is unlocked.
298 remove_from_swap(newpage);
301 EXPORT_SYMBOL(migrate_page);
304 * Migration function for pages with buffers. This function can only be used
305 * if the underlying filesystem guarantees that no other references to "page"
308 int buffer_migrate_page(struct page *newpage, struct page *page)
310 struct address_space *mapping = page->mapping;
311 struct buffer_head *bh, *head;
317 if (!page_has_buffers(page))
318 return migrate_page(newpage, page);
320 head = page_buffers(page);
322 rc = migrate_page_move_mapping(newpage, page);
331 bh = bh->b_this_page;
333 } while (bh != head);
335 ClearPagePrivate(page);
336 set_page_private(newpage, page_private(page));
337 set_page_private(page, 0);
343 set_bh_page(bh, newpage, bh_offset(bh));
344 bh = bh->b_this_page;
346 } while (bh != head);
348 SetPagePrivate(newpage);
350 migrate_page_copy(newpage, page);
356 bh = bh->b_this_page;
358 } while (bh != head);
362 EXPORT_SYMBOL(buffer_migrate_page);
367 * Two lists are passed to this function. The first list
368 * contains the pages isolated from the LRU to be migrated.
369 * The second list contains new pages that the pages isolated
370 * can be moved to. If the second list is NULL then all
371 * pages are swapped out.
373 * The function returns after 10 attempts or if no pages
374 * are movable anymore because to has become empty
375 * or no retryable pages exist anymore.
377 * Return: Number of pages not migrated when "to" ran empty.
379 int migrate_pages(struct list_head *from, struct list_head *to,
380 struct list_head *moved, struct list_head *failed)
387 int swapwrite = current->flags & PF_SWAPWRITE;
391 current->flags |= PF_SWAPWRITE;
396 list_for_each_entry_safe(page, page2, from, lru) {
397 struct page *newpage = NULL;
398 struct address_space *mapping;
403 if (page_count(page) == 1)
404 /* page was freed from under us. So we are done. */
407 if (to && list_empty(to))
411 * Skip locked pages during the first two passes to give the
412 * functions holding the lock time to release the page. Later we
413 * use lock_page() to have a higher chance of acquiring the
420 if (TestSetPageLocked(page))
424 * Only wait on writeback if we have already done a pass where
425 * we we may have triggered writeouts for lots of pages.
428 wait_on_page_writeback(page);
430 if (PageWriteback(page))
435 * Anonymous pages must have swap cache references otherwise
436 * the information contained in the page maps cannot be
439 if (PageAnon(page) && !PageSwapCache(page)) {
440 if (!add_to_swap(page, GFP_KERNEL)) {
447 rc = swap_page(page);
451 newpage = lru_to_page(to);
455 * Establish swap ptes for anonymous pages or destroy pte
458 * In order to reestablish file backed mappings the fault handlers
459 * will take the radix tree_lock which may then be used to stop
460 * processses from accessing this page until the new page is ready.
462 * A process accessing via a swap pte (an anonymous page) will take a
463 * page_lock on the old page which will block the process until the
464 * migration attempt is complete. At that time the PageSwapCache bit
465 * will be examined. If the page was migrated then the PageSwapCache
466 * bit will be clear and the operation to retrieve the page will be
467 * retried which will find the new page in the radix tree. Then a new
468 * direct mapping may be generated based on the radix tree contents.
470 * If the page was not migrated then the PageSwapCache bit
471 * is still set and the operation may continue.
474 if (try_to_unmap(page, 1) == SWAP_FAIL)
475 /* A vma has VM_LOCKED set -> permanent failure */
479 if (page_mapped(page))
482 * Pages are properly locked and writeback is complete.
483 * Try to migrate the page.
485 mapping = page_mapping(page);
489 if (mapping->a_ops->migratepage) {
491 * Most pages have a mapping and most filesystems
492 * should provide a migration function. Anonymous
493 * pages are part of swap space which also has its
494 * own migration function. This is the most common
495 * path for page migration.
497 rc = mapping->a_ops->migratepage(newpage, page);
502 * Default handling if a filesystem does not provide
503 * a migration function. We can only migrate clean
504 * pages so try to write out any dirty pages first.
506 if (PageDirty(page)) {
507 switch (pageout(page, mapping)) {
513 unlock_page(newpage);
517 ; /* try to migrate the page below */
522 * Buffers are managed in a filesystem specific way.
523 * We must have no buffers or drop them.
525 if (!page_has_buffers(page) ||
526 try_to_release_page(page, GFP_KERNEL)) {
527 rc = migrate_page(newpage, page);
532 * On early passes with mapped pages simply
533 * retry. There may be a lock held for some
534 * buffers that may go away. Later
539 * Persistently unable to drop buffers..... As a
540 * measure of last resort we fall back to
543 unlock_page(newpage);
545 rc = swap_page(page);
550 unlock_page(newpage);
559 /* Permanent failure */
560 list_move(&page->lru, failed);
564 /* Successful migration. Return page to LRU */
565 move_to_lru(newpage);
567 list_move(&page->lru, moved);
570 if (retry && pass++ < 10)
574 current->flags &= ~PF_SWAPWRITE;
576 return nr_failed + retry;
580 * Migrate the list 'pagelist' of pages to a certain destination.
582 * Specify destination with either non-NULL vma or dest_node >= 0
583 * Return the number of pages not migrated or error code
585 int migrate_pages_to(struct list_head *pagelist,
586 struct vm_area_struct *vma, int dest)
592 unsigned long offset = 0;
599 list_for_each(p, pagelist) {
602 * The address passed to alloc_page_vma is used to
603 * generate the proper interleave behavior. We fake
604 * the address here by an increasing offset in order
605 * to get the proper distribution of pages.
607 * No decision has been made as to which page
608 * a certain old page is moved to so we cannot
609 * specify the correct address.
611 page = alloc_page_vma(GFP_HIGHUSER, vma,
612 offset + vma->vm_start);
616 page = alloc_pages_node(dest, GFP_HIGHUSER, 0);
622 list_add_tail(&page->lru, &newlist);
624 if (nr_pages > MIGRATE_CHUNK_SIZE)
627 err = migrate_pages(pagelist, &newlist, &moved, &failed);
629 putback_lru_pages(&moved); /* Call release pages instead ?? */
631 if (err >= 0 && list_empty(&newlist) && !list_empty(pagelist))
634 /* Return leftover allocated pages */
635 while (!list_empty(&newlist)) {
636 page = list_entry(newlist.next, struct page, lru);
637 list_del(&page->lru);
640 list_splice(&failed, pagelist);
644 /* Calculate number of leftover pages */
646 list_for_each(p, pagelist)