2 * linux/mm/memory_hotplug.c
7 #include <linux/stddef.h>
9 #include <linux/swap.h>
10 #include <linux/interrupt.h>
11 #include <linux/pagemap.h>
12 #include <linux/bootmem.h>
13 #include <linux/compiler.h>
14 #include <linux/module.h>
15 #include <linux/pagevec.h>
16 #include <linux/writeback.h>
17 #include <linux/slab.h>
18 #include <linux/sysctl.h>
19 #include <linux/cpu.h>
20 #include <linux/memory.h>
21 #include <linux/memory_hotplug.h>
22 #include <linux/highmem.h>
23 #include <linux/vmalloc.h>
24 #include <linux/ioport.h>
25 #include <linux/delay.h>
26 #include <linux/migrate.h>
27 #include <linux/page-isolation.h>
28 #include <linux/pfn.h>
29 #include <linux/suspend.h>
30 #include <linux/mm_inline.h>
31 #include <linux/firmware-map.h>
33 #include <asm/tlbflush.h>
37 DEFINE_MUTEX(mem_hotplug_mutex);
39 void lock_memory_hotplug(void)
41 mutex_lock(&mem_hotplug_mutex);
43 /* for exclusive hibernation if CONFIG_HIBERNATION=y */
47 void unlock_memory_hotplug(void)
49 unlock_system_sleep();
50 mutex_unlock(&mem_hotplug_mutex);
54 /* add this memory to iomem resource */
55 static struct resource *register_memory_resource(u64 start, u64 size)
58 res = kzalloc(sizeof(struct resource), GFP_KERNEL);
61 res->name = "System RAM";
63 res->end = start + size - 1;
64 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
65 if (request_resource(&iomem_resource, res) < 0) {
66 printk("System RAM resource %llx - %llx cannot be added\n",
67 (unsigned long long)res->start, (unsigned long long)res->end);
74 static void release_memory_resource(struct resource *res)
78 release_resource(res);
83 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
84 #ifndef CONFIG_SPARSEMEM_VMEMMAP
85 static void get_page_bootmem(unsigned long info, struct page *page, int type)
87 atomic_set(&page->_mapcount, type);
89 set_page_private(page, info);
90 atomic_inc(&page->_count);
93 /* reference to __meminit __free_pages_bootmem is valid
94 * so use __ref to tell modpost not to generate a warning */
95 void __ref put_page_bootmem(struct page *page)
99 type = atomic_read(&page->_mapcount);
102 if (atomic_dec_return(&page->_count) == 1) {
103 ClearPagePrivate(page);
104 set_page_private(page, 0);
105 reset_page_mapcount(page);
106 __free_pages_bootmem(page, 0);
111 static void register_page_bootmem_info_section(unsigned long start_pfn)
113 unsigned long *usemap, mapsize, section_nr, i;
114 struct mem_section *ms;
115 struct page *page, *memmap;
117 if (!pfn_valid(start_pfn))
120 section_nr = pfn_to_section_nr(start_pfn);
121 ms = __nr_to_section(section_nr);
123 /* Get section's memmap address */
124 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
127 * Get page for the memmap's phys address
128 * XXX: need more consideration for sparse_vmemmap...
130 page = virt_to_page(memmap);
131 mapsize = sizeof(struct page) * PAGES_PER_SECTION;
132 mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
134 /* remember memmap's page */
135 for (i = 0; i < mapsize; i++, page++)
136 get_page_bootmem(section_nr, page, SECTION_INFO);
138 usemap = __nr_to_section(section_nr)->pageblock_flags;
139 page = virt_to_page(usemap);
141 mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT;
143 for (i = 0; i < mapsize; i++, page++)
144 get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
148 void register_page_bootmem_info_node(struct pglist_data *pgdat)
150 unsigned long i, pfn, end_pfn, nr_pages;
151 int node = pgdat->node_id;
155 nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
156 page = virt_to_page(pgdat);
158 for (i = 0; i < nr_pages; i++, page++)
159 get_page_bootmem(node, page, NODE_INFO);
161 zone = &pgdat->node_zones[0];
162 for (; zone < pgdat->node_zones + MAX_NR_ZONES - 1; zone++) {
163 if (zone->wait_table) {
164 nr_pages = zone->wait_table_hash_nr_entries
165 * sizeof(wait_queue_head_t);
166 nr_pages = PAGE_ALIGN(nr_pages) >> PAGE_SHIFT;
167 page = virt_to_page(zone->wait_table);
169 for (i = 0; i < nr_pages; i++, page++)
170 get_page_bootmem(node, page, NODE_INFO);
174 pfn = pgdat->node_start_pfn;
175 end_pfn = pfn + pgdat->node_spanned_pages;
177 /* register_section info */
178 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION)
179 register_page_bootmem_info_section(pfn);
182 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
184 static void grow_zone_span(struct zone *zone, unsigned long start_pfn,
185 unsigned long end_pfn)
187 unsigned long old_zone_end_pfn;
189 zone_span_writelock(zone);
191 old_zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
192 if (start_pfn < zone->zone_start_pfn)
193 zone->zone_start_pfn = start_pfn;
195 zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
196 zone->zone_start_pfn;
198 zone_span_writeunlock(zone);
201 static void grow_pgdat_span(struct pglist_data *pgdat, unsigned long start_pfn,
202 unsigned long end_pfn)
204 unsigned long old_pgdat_end_pfn =
205 pgdat->node_start_pfn + pgdat->node_spanned_pages;
207 if (start_pfn < pgdat->node_start_pfn)
208 pgdat->node_start_pfn = start_pfn;
210 pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
211 pgdat->node_start_pfn;
214 static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
216 struct pglist_data *pgdat = zone->zone_pgdat;
217 int nr_pages = PAGES_PER_SECTION;
218 int nid = pgdat->node_id;
222 zone_type = zone - pgdat->node_zones;
223 if (!zone->wait_table) {
226 ret = init_currently_empty_zone(zone, phys_start_pfn,
227 nr_pages, MEMMAP_HOTPLUG);
231 pgdat_resize_lock(zone->zone_pgdat, &flags);
232 grow_zone_span(zone, phys_start_pfn, phys_start_pfn + nr_pages);
233 grow_pgdat_span(zone->zone_pgdat, phys_start_pfn,
234 phys_start_pfn + nr_pages);
235 pgdat_resize_unlock(zone->zone_pgdat, &flags);
236 memmap_init_zone(nr_pages, nid, zone_type,
237 phys_start_pfn, MEMMAP_HOTPLUG);
241 static int __meminit __add_section(int nid, struct zone *zone,
242 unsigned long phys_start_pfn)
244 int nr_pages = PAGES_PER_SECTION;
247 if (pfn_valid(phys_start_pfn))
250 ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
255 ret = __add_zone(zone, phys_start_pfn);
260 return register_new_memory(nid, __pfn_to_section(phys_start_pfn));
263 #ifdef CONFIG_SPARSEMEM_VMEMMAP
264 static int __remove_section(struct zone *zone, struct mem_section *ms)
267 * XXX: Freeing memmap with vmemmap is not implement yet.
268 * This should be removed later.
273 static int __remove_section(struct zone *zone, struct mem_section *ms)
276 struct pglist_data *pgdat = zone->zone_pgdat;
279 if (!valid_section(ms))
282 ret = unregister_memory_section(ms);
286 pgdat_resize_lock(pgdat, &flags);
287 sparse_remove_one_section(zone, ms);
288 pgdat_resize_unlock(pgdat, &flags);
294 * Reasonably generic function for adding memory. It is
295 * expected that archs that support memory hotplug will
296 * call this function after deciding the zone to which to
299 int __ref __add_pages(int nid, struct zone *zone, unsigned long phys_start_pfn,
300 unsigned long nr_pages)
304 int start_sec, end_sec;
305 /* during initialize mem_map, align hot-added range to section */
306 start_sec = pfn_to_section_nr(phys_start_pfn);
307 end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
309 for (i = start_sec; i <= end_sec; i++) {
310 err = __add_section(nid, zone, i << PFN_SECTION_SHIFT);
313 * EEXIST is finally dealt with by ioresource collision
314 * check. see add_memory() => register_memory_resource()
315 * Warning will be printed if there is collision.
317 if (err && (err != -EEXIST))
324 EXPORT_SYMBOL_GPL(__add_pages);
327 * __remove_pages() - remove sections of pages from a zone
328 * @zone: zone from which pages need to be removed
329 * @phys_start_pfn: starting pageframe (must be aligned to start of a section)
330 * @nr_pages: number of pages to remove (must be multiple of section size)
332 * Generic helper function to remove section mappings and sysfs entries
333 * for the section of the memory we are removing. Caller needs to make
334 * sure that pages are marked reserved and zones are adjust properly by
335 * calling offline_pages().
337 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
338 unsigned long nr_pages)
340 unsigned long i, ret = 0;
341 int sections_to_remove;
344 * We can only remove entire sections
346 BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
347 BUG_ON(nr_pages % PAGES_PER_SECTION);
349 sections_to_remove = nr_pages / PAGES_PER_SECTION;
350 for (i = 0; i < sections_to_remove; i++) {
351 unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION;
352 release_mem_region(pfn << PAGE_SHIFT,
353 PAGES_PER_SECTION << PAGE_SHIFT);
354 ret = __remove_section(zone, __pfn_to_section(pfn));
360 EXPORT_SYMBOL_GPL(__remove_pages);
362 void online_page(struct page *page)
364 unsigned long pfn = page_to_pfn(page);
367 if (pfn >= num_physpages)
368 num_physpages = pfn + 1;
370 #ifdef CONFIG_HIGHMEM
371 if (PageHighMem(page))
375 #ifdef CONFIG_FLATMEM
376 max_mapnr = max(page_to_pfn(page), max_mapnr);
379 ClearPageReserved(page);
380 init_page_count(page);
384 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
388 unsigned long onlined_pages = *(unsigned long *)arg;
390 if (PageReserved(pfn_to_page(start_pfn)))
391 for (i = 0; i < nr_pages; i++) {
392 page = pfn_to_page(start_pfn + i);
396 *(unsigned long *)arg = onlined_pages;
401 int online_pages(unsigned long pfn, unsigned long nr_pages)
403 unsigned long onlined_pages = 0;
405 int need_zonelists_rebuild = 0;
408 struct memory_notify arg;
410 lock_memory_hotplug();
412 arg.nr_pages = nr_pages;
413 arg.status_change_nid = -1;
415 nid = page_to_nid(pfn_to_page(pfn));
416 if (node_present_pages(nid) == 0)
417 arg.status_change_nid = nid;
419 ret = memory_notify(MEM_GOING_ONLINE, &arg);
420 ret = notifier_to_errno(ret);
422 memory_notify(MEM_CANCEL_ONLINE, &arg);
423 unlock_memory_hotplug();
427 * This doesn't need a lock to do pfn_to_page().
428 * The section can't be removed here because of the
429 * memory_block->state_mutex.
431 zone = page_zone(pfn_to_page(pfn));
433 * If this zone is not populated, then it is not in zonelist.
434 * This means the page allocator ignores this zone.
435 * So, zonelist must be updated after online.
437 mutex_lock(&zonelists_mutex);
438 if (!populated_zone(zone))
439 need_zonelists_rebuild = 1;
441 ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages,
444 mutex_unlock(&zonelists_mutex);
445 printk(KERN_DEBUG "online_pages %lx at %lx failed\n",
447 memory_notify(MEM_CANCEL_ONLINE, &arg);
448 unlock_memory_hotplug();
452 zone->present_pages += onlined_pages;
453 zone->zone_pgdat->node_present_pages += onlined_pages;
454 if (need_zonelists_rebuild)
455 build_all_zonelists(zone);
457 zone_pcp_update(zone);
459 mutex_unlock(&zonelists_mutex);
460 setup_per_zone_wmarks();
461 calculate_zone_inactive_ratio(zone);
463 kswapd_run(zone_to_nid(zone));
464 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
467 vm_total_pages = nr_free_pagecache_pages();
469 writeback_set_ratelimit();
472 memory_notify(MEM_ONLINE, &arg);
473 unlock_memory_hotplug();
477 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
479 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
480 static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
482 struct pglist_data *pgdat;
483 unsigned long zones_size[MAX_NR_ZONES] = {0};
484 unsigned long zholes_size[MAX_NR_ZONES] = {0};
485 unsigned long start_pfn = start >> PAGE_SHIFT;
487 pgdat = arch_alloc_nodedata(nid);
491 arch_refresh_nodedata(nid, pgdat);
493 /* we can use NODE_DATA(nid) from here */
495 /* init node's zones as empty zones, we don't have any present pages.*/
496 free_area_init_node(nid, zones_size, start_pfn, zholes_size);
501 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
503 arch_refresh_nodedata(nid, NULL);
504 arch_free_nodedata(pgdat);
510 * called by cpu_up() to online a node without onlined memory.
512 int mem_online_node(int nid)
517 lock_memory_hotplug();
518 pgdat = hotadd_new_pgdat(nid, 0);
523 node_set_online(nid);
524 ret = register_one_node(nid);
528 unlock_memory_hotplug();
532 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
533 int __ref add_memory(int nid, u64 start, u64 size)
535 pg_data_t *pgdat = NULL;
537 struct resource *res;
540 lock_memory_hotplug();
542 res = register_memory_resource(start, size);
547 if (!node_online(nid)) {
548 pgdat = hotadd_new_pgdat(nid, start);
555 /* call arch's memory hotadd */
556 ret = arch_add_memory(nid, start, size);
561 /* we online node here. we can't roll back from here. */
562 node_set_online(nid);
565 ret = register_one_node(nid);
567 * If sysfs file of new node can't create, cpu on the node
568 * can't be hot-added. There is no rollback way now.
569 * So, check by BUG_ON() to catch it reluctantly..
574 /* create new memmap entry */
575 firmware_map_add_hotplug(start, start + size, "System RAM");
580 /* rollback pgdat allocation and others */
582 rollback_node_hotadd(nid, pgdat);
584 release_memory_resource(res);
587 unlock_memory_hotplug();
590 EXPORT_SYMBOL_GPL(add_memory);
592 #ifdef CONFIG_MEMORY_HOTREMOVE
594 * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
595 * set and the size of the free page is given by page_order(). Using this,
596 * the function determines if the pageblock contains only free pages.
597 * Due to buddy contraints, a free page at least the size of a pageblock will
598 * be located at the start of the pageblock
600 static inline int pageblock_free(struct page *page)
602 return PageBuddy(page) && page_order(page) >= pageblock_order;
605 /* Return the start of the next active pageblock after a given page */
606 static struct page *next_active_pageblock(struct page *page)
608 /* Ensure the starting page is pageblock-aligned */
609 BUG_ON(page_to_pfn(page) & (pageblock_nr_pages - 1));
611 /* If the entire pageblock is free, move to the end of free page */
612 if (pageblock_free(page)) {
614 /* be careful. we don't have locks, page_order can be changed.*/
615 order = page_order(page);
616 if ((order < MAX_ORDER) && (order >= pageblock_order))
617 return page + (1 << order);
620 return page + pageblock_nr_pages;
623 /* Checks if this range of memory is likely to be hot-removable. */
624 int is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
626 struct page *page = pfn_to_page(start_pfn);
627 struct page *end_page = page + nr_pages;
629 /* Check the starting page of each pageblock within the range */
630 for (; page < end_page; page = next_active_pageblock(page)) {
631 if (!is_pageblock_removable_nolock(page))
636 /* All pageblocks in the memory block are likely to be hot-removable */
641 * Confirm all pages in a range [start, end) is belongs to the same zone.
643 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
646 struct zone *zone = NULL;
649 for (pfn = start_pfn;
651 pfn += MAX_ORDER_NR_PAGES) {
653 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
654 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
656 if (i == MAX_ORDER_NR_PAGES)
658 page = pfn_to_page(pfn + i);
659 if (zone && page_zone(page) != zone)
661 zone = page_zone(page);
667 * Scanning pfn is much easier than scanning lru list.
668 * Scan pfn from start to end and Find LRU page.
670 static unsigned long scan_lru_pages(unsigned long start, unsigned long end)
674 for (pfn = start; pfn < end; pfn++) {
675 if (pfn_valid(pfn)) {
676 page = pfn_to_page(pfn);
685 hotremove_migrate_alloc(struct page *page, unsigned long private, int **x)
687 /* This should be improooooved!! */
688 return alloc_page(GFP_HIGHUSER_MOVABLE);
691 #define NR_OFFLINE_AT_ONCE_PAGES (256)
693 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
697 int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
702 for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
705 page = pfn_to_page(pfn);
706 if (!page_count(page))
709 * We can skip free pages. And we can only deal with pages on
712 ret = isolate_lru_page(page);
713 if (!ret) { /* Success */
714 list_add_tail(&page->lru, &source);
716 inc_zone_page_state(page, NR_ISOLATED_ANON +
717 page_is_file_cache(page));
720 #ifdef CONFIG_DEBUG_VM
721 printk(KERN_ALERT "removing pfn %lx from LRU failed\n",
725 /* Becasue we don't have big zone->lock. we should
726 check this again here. */
727 if (page_count(page)) {
734 if (!list_empty(&source)) {
736 putback_lru_pages(&source);
739 /* this function returns # of failed pages */
740 ret = migrate_pages(&source, hotremove_migrate_alloc, 0, 1);
742 putback_lru_pages(&source);
749 * remove from free_area[] and mark all as Reserved.
752 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
755 __offline_isolated_pages(start, start + nr_pages);
760 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
762 walk_system_ram_range(start_pfn, end_pfn - start_pfn, NULL,
763 offline_isolated_pages_cb);
767 * Check all pages in range, recoreded as memory resource, are isolated.
770 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
774 long offlined = *(long *)data;
775 ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
778 *(long *)data += offlined;
783 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
788 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, &offlined,
789 check_pages_isolated_cb);
791 offlined = (long)ret;
795 static int offline_pages(unsigned long start_pfn,
796 unsigned long end_pfn, unsigned long timeout)
798 unsigned long pfn, nr_pages, expire;
800 int ret, drain, retry_max, node;
802 struct memory_notify arg;
804 BUG_ON(start_pfn >= end_pfn);
805 /* at least, alignment against pageblock is necessary */
806 if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
808 if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
810 /* This makes hotplug much easier...and readable.
811 we assume this for now. .*/
812 if (!test_pages_in_a_zone(start_pfn, end_pfn))
815 lock_memory_hotplug();
817 zone = page_zone(pfn_to_page(start_pfn));
818 node = zone_to_nid(zone);
819 nr_pages = end_pfn - start_pfn;
821 /* set above range as isolated */
822 ret = start_isolate_page_range(start_pfn, end_pfn);
826 arg.start_pfn = start_pfn;
827 arg.nr_pages = nr_pages;
828 arg.status_change_nid = -1;
829 if (nr_pages >= node_present_pages(node))
830 arg.status_change_nid = node;
832 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
833 ret = notifier_to_errno(ret);
838 expire = jiffies + timeout;
842 /* start memory hot removal */
844 if (time_after(jiffies, expire))
847 if (signal_pending(current))
856 pfn = scan_lru_pages(start_pfn, end_pfn);
857 if (pfn) { /* We have page on LRU */
858 ret = do_migrate_range(pfn, end_pfn);
864 if (--retry_max == 0)
871 /* drain all zone's lru pagevec, this is asyncronous... */
874 /* drain pcp pages , this is synchrouns. */
877 offlined_pages = check_pages_isolated(start_pfn, end_pfn);
878 if (offlined_pages < 0) {
882 printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
883 /* Ok, all of our target is islaoted.
884 We cannot do rollback at this point. */
885 offline_isolated_pages(start_pfn, end_pfn);
886 /* reset pagetype flags and makes migrate type to be MOVABLE */
887 undo_isolate_page_range(start_pfn, end_pfn);
888 /* removal success */
889 zone->present_pages -= offlined_pages;
890 zone->zone_pgdat->node_present_pages -= offlined_pages;
891 totalram_pages -= offlined_pages;
893 setup_per_zone_wmarks();
894 calculate_zone_inactive_ratio(zone);
895 if (!node_present_pages(node)) {
896 node_clear_state(node, N_HIGH_MEMORY);
900 vm_total_pages = nr_free_pagecache_pages();
901 writeback_set_ratelimit();
903 memory_notify(MEM_OFFLINE, &arg);
904 unlock_memory_hotplug();
908 printk(KERN_INFO "memory offlining %lx to %lx failed\n",
910 memory_notify(MEM_CANCEL_OFFLINE, &arg);
911 /* pushback to free area */
912 undo_isolate_page_range(start_pfn, end_pfn);
915 unlock_memory_hotplug();
919 int remove_memory(u64 start, u64 size)
921 unsigned long start_pfn, end_pfn;
923 start_pfn = PFN_DOWN(start);
924 end_pfn = start_pfn + PFN_DOWN(size);
925 return offline_pages(start_pfn, end_pfn, 120 * HZ);
928 int remove_memory(u64 start, u64 size)
932 #endif /* CONFIG_MEMORY_HOTREMOVE */
933 EXPORT_SYMBOL_GPL(remove_memory);