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,
88 page->lru.next = (struct list_head *) type;
90 set_page_private(page, info);
91 atomic_inc(&page->_count);
94 /* reference to __meminit __free_pages_bootmem is valid
95 * so use __ref to tell modpost not to generate a warning */
96 void __ref put_page_bootmem(struct page *page)
100 type = (unsigned long) page->lru.next;
101 BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE ||
102 type > MEMORY_HOTPLUG_MAX_BOOTMEM_TYPE);
104 if (atomic_dec_return(&page->_count) == 1) {
105 ClearPagePrivate(page);
106 set_page_private(page, 0);
107 INIT_LIST_HEAD(&page->lru);
108 __free_pages_bootmem(page, 0);
113 static void register_page_bootmem_info_section(unsigned long start_pfn)
115 unsigned long *usemap, mapsize, section_nr, i;
116 struct mem_section *ms;
117 struct page *page, *memmap;
119 section_nr = pfn_to_section_nr(start_pfn);
120 ms = __nr_to_section(section_nr);
122 /* Get section's memmap address */
123 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
126 * Get page for the memmap's phys address
127 * XXX: need more consideration for sparse_vmemmap...
129 page = virt_to_page(memmap);
130 mapsize = sizeof(struct page) * PAGES_PER_SECTION;
131 mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
133 /* remember memmap's page */
134 for (i = 0; i < mapsize; i++, page++)
135 get_page_bootmem(section_nr, page, SECTION_INFO);
137 usemap = __nr_to_section(section_nr)->pageblock_flags;
138 page = virt_to_page(usemap);
140 mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT;
142 for (i = 0; i < mapsize; i++, page++)
143 get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
147 void register_page_bootmem_info_node(struct pglist_data *pgdat)
149 unsigned long i, pfn, end_pfn, nr_pages;
150 int node = pgdat->node_id;
154 nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
155 page = virt_to_page(pgdat);
157 for (i = 0; i < nr_pages; i++, page++)
158 get_page_bootmem(node, page, NODE_INFO);
160 zone = &pgdat->node_zones[0];
161 for (; zone < pgdat->node_zones + MAX_NR_ZONES - 1; zone++) {
162 if (zone->wait_table) {
163 nr_pages = zone->wait_table_hash_nr_entries
164 * sizeof(wait_queue_head_t);
165 nr_pages = PAGE_ALIGN(nr_pages) >> PAGE_SHIFT;
166 page = virt_to_page(zone->wait_table);
168 for (i = 0; i < nr_pages; i++, page++)
169 get_page_bootmem(node, page, NODE_INFO);
173 pfn = pgdat->node_start_pfn;
174 end_pfn = pfn + pgdat->node_spanned_pages;
176 /* register_section info */
177 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
179 * Some platforms can assign the same pfn to multiple nodes - on
180 * node0 as well as nodeN. To avoid registering a pfn against
181 * multiple nodes we check that this pfn does not already
182 * reside in some other node.
184 if (pfn_valid(pfn) && (pfn_to_nid(pfn) == node))
185 register_page_bootmem_info_section(pfn);
188 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
190 static void grow_zone_span(struct zone *zone, unsigned long start_pfn,
191 unsigned long end_pfn)
193 unsigned long old_zone_end_pfn;
195 zone_span_writelock(zone);
197 old_zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
198 if (start_pfn < zone->zone_start_pfn)
199 zone->zone_start_pfn = start_pfn;
201 zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
202 zone->zone_start_pfn;
204 zone_span_writeunlock(zone);
207 static void grow_pgdat_span(struct pglist_data *pgdat, unsigned long start_pfn,
208 unsigned long end_pfn)
210 unsigned long old_pgdat_end_pfn =
211 pgdat->node_start_pfn + pgdat->node_spanned_pages;
213 if (start_pfn < pgdat->node_start_pfn)
214 pgdat->node_start_pfn = start_pfn;
216 pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
217 pgdat->node_start_pfn;
220 static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
222 struct pglist_data *pgdat = zone->zone_pgdat;
223 int nr_pages = PAGES_PER_SECTION;
224 int nid = pgdat->node_id;
228 zone_type = zone - pgdat->node_zones;
229 if (!zone->wait_table) {
232 ret = init_currently_empty_zone(zone, phys_start_pfn,
233 nr_pages, MEMMAP_HOTPLUG);
237 pgdat_resize_lock(zone->zone_pgdat, &flags);
238 grow_zone_span(zone, phys_start_pfn, phys_start_pfn + nr_pages);
239 grow_pgdat_span(zone->zone_pgdat, phys_start_pfn,
240 phys_start_pfn + nr_pages);
241 pgdat_resize_unlock(zone->zone_pgdat, &flags);
242 memmap_init_zone(nr_pages, nid, zone_type,
243 phys_start_pfn, MEMMAP_HOTPLUG);
247 static int __meminit __add_section(int nid, struct zone *zone,
248 unsigned long phys_start_pfn)
250 int nr_pages = PAGES_PER_SECTION;
253 if (pfn_valid(phys_start_pfn))
256 ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
261 ret = __add_zone(zone, phys_start_pfn);
266 return register_new_memory(nid, __pfn_to_section(phys_start_pfn));
269 #ifdef CONFIG_SPARSEMEM_VMEMMAP
270 static int __remove_section(struct zone *zone, struct mem_section *ms)
273 * XXX: Freeing memmap with vmemmap is not implement yet.
274 * This should be removed later.
279 static int __remove_section(struct zone *zone, struct mem_section *ms)
282 struct pglist_data *pgdat = zone->zone_pgdat;
285 if (!valid_section(ms))
288 ret = unregister_memory_section(ms);
292 pgdat_resize_lock(pgdat, &flags);
293 sparse_remove_one_section(zone, ms);
294 pgdat_resize_unlock(pgdat, &flags);
300 * Reasonably generic function for adding memory. It is
301 * expected that archs that support memory hotplug will
302 * call this function after deciding the zone to which to
305 int __ref __add_pages(int nid, struct zone *zone, unsigned long phys_start_pfn,
306 unsigned long nr_pages)
310 int start_sec, end_sec;
311 /* during initialize mem_map, align hot-added range to section */
312 start_sec = pfn_to_section_nr(phys_start_pfn);
313 end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
315 for (i = start_sec; i <= end_sec; i++) {
316 err = __add_section(nid, zone, i << PFN_SECTION_SHIFT);
319 * EEXIST is finally dealt with by ioresource collision
320 * check. see add_memory() => register_memory_resource()
321 * Warning will be printed if there is collision.
323 if (err && (err != -EEXIST))
330 EXPORT_SYMBOL_GPL(__add_pages);
333 * __remove_pages() - remove sections of pages from a zone
334 * @zone: zone from which pages need to be removed
335 * @phys_start_pfn: starting pageframe (must be aligned to start of a section)
336 * @nr_pages: number of pages to remove (must be multiple of section size)
338 * Generic helper function to remove section mappings and sysfs entries
339 * for the section of the memory we are removing. Caller needs to make
340 * sure that pages are marked reserved and zones are adjust properly by
341 * calling offline_pages().
343 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
344 unsigned long nr_pages)
346 unsigned long i, ret = 0;
347 int sections_to_remove;
350 * We can only remove entire sections
352 BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
353 BUG_ON(nr_pages % PAGES_PER_SECTION);
355 sections_to_remove = nr_pages / PAGES_PER_SECTION;
356 for (i = 0; i < sections_to_remove; i++) {
357 unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION;
358 release_mem_region(pfn << PAGE_SHIFT,
359 PAGES_PER_SECTION << PAGE_SHIFT);
360 ret = __remove_section(zone, __pfn_to_section(pfn));
366 EXPORT_SYMBOL_GPL(__remove_pages);
368 void online_page(struct page *page)
370 unsigned long pfn = page_to_pfn(page);
373 if (pfn >= num_physpages)
374 num_physpages = pfn + 1;
376 #ifdef CONFIG_HIGHMEM
377 if (PageHighMem(page))
381 ClearPageReserved(page);
382 init_page_count(page);
386 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
390 unsigned long onlined_pages = *(unsigned long *)arg;
392 if (PageReserved(pfn_to_page(start_pfn)))
393 for (i = 0; i < nr_pages; i++) {
394 page = pfn_to_page(start_pfn + i);
398 *(unsigned long *)arg = onlined_pages;
403 int __ref online_pages(unsigned long pfn, unsigned long nr_pages)
405 unsigned long onlined_pages = 0;
407 int need_zonelists_rebuild = 0;
410 struct memory_notify arg;
412 lock_memory_hotplug();
414 arg.nr_pages = nr_pages;
415 arg.status_change_nid = -1;
417 nid = page_to_nid(pfn_to_page(pfn));
418 if (node_present_pages(nid) == 0)
419 arg.status_change_nid = nid;
421 ret = memory_notify(MEM_GOING_ONLINE, &arg);
422 ret = notifier_to_errno(ret);
424 memory_notify(MEM_CANCEL_ONLINE, &arg);
425 unlock_memory_hotplug();
429 * This doesn't need a lock to do pfn_to_page().
430 * The section can't be removed here because of the
431 * memory_block->state_mutex.
433 zone = page_zone(pfn_to_page(pfn));
435 * If this zone is not populated, then it is not in zonelist.
436 * This means the page allocator ignores this zone.
437 * So, zonelist must be updated after online.
439 mutex_lock(&zonelists_mutex);
440 if (!populated_zone(zone))
441 need_zonelists_rebuild = 1;
443 ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages,
446 mutex_unlock(&zonelists_mutex);
447 printk(KERN_DEBUG "online_pages %lx at %lx failed\n",
449 memory_notify(MEM_CANCEL_ONLINE, &arg);
450 unlock_memory_hotplug();
454 zone->present_pages += onlined_pages;
455 zone->zone_pgdat->node_present_pages += onlined_pages;
456 if (need_zonelists_rebuild)
457 build_all_zonelists(zone);
459 zone_pcp_update(zone);
461 mutex_unlock(&zonelists_mutex);
463 init_per_zone_wmark_min();
466 kswapd_run(zone_to_nid(zone));
467 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
470 vm_total_pages = nr_free_pagecache_pages();
472 writeback_set_ratelimit();
475 memory_notify(MEM_ONLINE, &arg);
476 unlock_memory_hotplug();
480 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
482 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
483 static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
485 struct pglist_data *pgdat;
486 unsigned long zones_size[MAX_NR_ZONES] = {0};
487 unsigned long zholes_size[MAX_NR_ZONES] = {0};
488 unsigned long start_pfn = start >> PAGE_SHIFT;
490 pgdat = arch_alloc_nodedata(nid);
494 arch_refresh_nodedata(nid, pgdat);
496 /* we can use NODE_DATA(nid) from here */
498 /* init node's zones as empty zones, we don't have any present pages.*/
499 free_area_init_node(nid, zones_size, start_pfn, zholes_size);
502 * The node we allocated has no zone fallback lists. For avoiding
503 * to access not-initialized zonelist, build here.
505 mutex_lock(&zonelists_mutex);
506 build_all_zonelists(NULL);
507 mutex_unlock(&zonelists_mutex);
512 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
514 arch_refresh_nodedata(nid, NULL);
515 arch_free_nodedata(pgdat);
521 * called by cpu_up() to online a node without onlined memory.
523 int mem_online_node(int nid)
528 lock_memory_hotplug();
529 pgdat = hotadd_new_pgdat(nid, 0);
534 node_set_online(nid);
535 ret = register_one_node(nid);
539 unlock_memory_hotplug();
543 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
544 int __ref add_memory(int nid, u64 start, u64 size)
546 pg_data_t *pgdat = NULL;
548 struct resource *res;
551 lock_memory_hotplug();
553 res = register_memory_resource(start, size);
558 if (!node_online(nid)) {
559 pgdat = hotadd_new_pgdat(nid, start);
566 /* call arch's memory hotadd */
567 ret = arch_add_memory(nid, start, size);
572 /* we online node here. we can't roll back from here. */
573 node_set_online(nid);
576 ret = register_one_node(nid);
578 * If sysfs file of new node can't create, cpu on the node
579 * can't be hot-added. There is no rollback way now.
580 * So, check by BUG_ON() to catch it reluctantly..
585 /* create new memmap entry */
586 firmware_map_add_hotplug(start, start + size, "System RAM");
591 /* rollback pgdat allocation and others */
593 rollback_node_hotadd(nid, pgdat);
595 release_memory_resource(res);
598 unlock_memory_hotplug();
601 EXPORT_SYMBOL_GPL(add_memory);
603 #ifdef CONFIG_MEMORY_HOTREMOVE
605 * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
606 * set and the size of the free page is given by page_order(). Using this,
607 * the function determines if the pageblock contains only free pages.
608 * Due to buddy contraints, a free page at least the size of a pageblock will
609 * be located at the start of the pageblock
611 static inline int pageblock_free(struct page *page)
613 return PageBuddy(page) && page_order(page) >= pageblock_order;
616 /* Return the start of the next active pageblock after a given page */
617 static struct page *next_active_pageblock(struct page *page)
619 /* Ensure the starting page is pageblock-aligned */
620 BUG_ON(page_to_pfn(page) & (pageblock_nr_pages - 1));
622 /* If the entire pageblock is free, move to the end of free page */
623 if (pageblock_free(page)) {
625 /* be careful. we don't have locks, page_order can be changed.*/
626 order = page_order(page);
627 if ((order < MAX_ORDER) && (order >= pageblock_order))
628 return page + (1 << order);
631 return page + pageblock_nr_pages;
634 /* Checks if this range of memory is likely to be hot-removable. */
635 int is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
637 struct page *page = pfn_to_page(start_pfn);
638 struct page *end_page = page + nr_pages;
640 /* Check the starting page of each pageblock within the range */
641 for (; page < end_page; page = next_active_pageblock(page)) {
642 if (!is_pageblock_removable_nolock(page))
647 /* All pageblocks in the memory block are likely to be hot-removable */
652 * Confirm all pages in a range [start, end) is belongs to the same zone.
654 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
657 struct zone *zone = NULL;
660 for (pfn = start_pfn;
662 pfn += MAX_ORDER_NR_PAGES) {
664 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
665 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
667 if (i == MAX_ORDER_NR_PAGES)
669 page = pfn_to_page(pfn + i);
670 if (zone && page_zone(page) != zone)
672 zone = page_zone(page);
678 * Scanning pfn is much easier than scanning lru list.
679 * Scan pfn from start to end and Find LRU page.
681 static unsigned long scan_lru_pages(unsigned long start, unsigned long end)
685 for (pfn = start; pfn < end; pfn++) {
686 if (pfn_valid(pfn)) {
687 page = pfn_to_page(pfn);
696 hotremove_migrate_alloc(struct page *page, unsigned long private, int **x)
698 /* This should be improooooved!! */
699 return alloc_page(GFP_HIGHUSER_MOVABLE);
702 #define NR_OFFLINE_AT_ONCE_PAGES (256)
704 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
708 int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
713 for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
716 page = pfn_to_page(pfn);
717 if (!get_page_unless_zero(page))
720 * We can skip free pages. And we can only deal with pages on
723 ret = isolate_lru_page(page);
724 if (!ret) { /* Success */
726 list_add_tail(&page->lru, &source);
728 inc_zone_page_state(page, NR_ISOLATED_ANON +
729 page_is_file_cache(page));
732 #ifdef CONFIG_DEBUG_VM
733 printk(KERN_ALERT "removing pfn %lx from LRU failed\n",
738 /* Because we don't have big zone->lock. we should
739 check this again here. */
740 if (page_count(page)) {
747 if (!list_empty(&source)) {
749 putback_lru_pages(&source);
752 /* this function returns # of failed pages */
753 ret = migrate_pages(&source, hotremove_migrate_alloc, 0,
756 putback_lru_pages(&source);
763 * remove from free_area[] and mark all as Reserved.
766 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
769 __offline_isolated_pages(start, start + nr_pages);
774 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
776 walk_system_ram_range(start_pfn, end_pfn - start_pfn, NULL,
777 offline_isolated_pages_cb);
781 * Check all pages in range, recoreded as memory resource, are isolated.
784 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
788 long offlined = *(long *)data;
789 ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
792 *(long *)data += offlined;
797 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
802 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, &offlined,
803 check_pages_isolated_cb);
805 offlined = (long)ret;
809 static int __ref offline_pages(unsigned long start_pfn,
810 unsigned long end_pfn, unsigned long timeout)
812 unsigned long pfn, nr_pages, expire;
814 int ret, drain, retry_max, node;
816 struct memory_notify arg;
818 BUG_ON(start_pfn >= end_pfn);
819 /* at least, alignment against pageblock is necessary */
820 if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
822 if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
824 /* This makes hotplug much easier...and readable.
825 we assume this for now. .*/
826 if (!test_pages_in_a_zone(start_pfn, end_pfn))
829 lock_memory_hotplug();
831 zone = page_zone(pfn_to_page(start_pfn));
832 node = zone_to_nid(zone);
833 nr_pages = end_pfn - start_pfn;
835 /* set above range as isolated */
836 ret = start_isolate_page_range(start_pfn, end_pfn);
840 arg.start_pfn = start_pfn;
841 arg.nr_pages = nr_pages;
842 arg.status_change_nid = -1;
843 if (nr_pages >= node_present_pages(node))
844 arg.status_change_nid = node;
846 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
847 ret = notifier_to_errno(ret);
852 expire = jiffies + timeout;
856 /* start memory hot removal */
858 if (time_after(jiffies, expire))
861 if (signal_pending(current))
870 pfn = scan_lru_pages(start_pfn, end_pfn);
871 if (pfn) { /* We have page on LRU */
872 ret = do_migrate_range(pfn, end_pfn);
878 if (--retry_max == 0)
885 /* drain all zone's lru pagevec, this is asyncronous... */
888 /* drain pcp pages , this is synchrouns. */
891 offlined_pages = check_pages_isolated(start_pfn, end_pfn);
892 if (offlined_pages < 0) {
896 printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
897 /* Ok, all of our target is islaoted.
898 We cannot do rollback at this point. */
899 offline_isolated_pages(start_pfn, end_pfn);
900 /* reset pagetype flags and makes migrate type to be MOVABLE */
901 undo_isolate_page_range(start_pfn, end_pfn);
902 /* removal success */
903 zone->present_pages -= offlined_pages;
904 zone->zone_pgdat->node_present_pages -= offlined_pages;
905 totalram_pages -= offlined_pages;
907 init_per_zone_wmark_min();
909 if (!node_present_pages(node)) {
910 node_clear_state(node, N_HIGH_MEMORY);
914 vm_total_pages = nr_free_pagecache_pages();
915 writeback_set_ratelimit();
917 memory_notify(MEM_OFFLINE, &arg);
918 unlock_memory_hotplug();
922 printk(KERN_INFO "memory offlining %lx to %lx failed\n",
924 memory_notify(MEM_CANCEL_OFFLINE, &arg);
925 /* pushback to free area */
926 undo_isolate_page_range(start_pfn, end_pfn);
929 unlock_memory_hotplug();
933 int remove_memory(u64 start, u64 size)
935 unsigned long start_pfn, end_pfn;
937 start_pfn = PFN_DOWN(start);
938 end_pfn = start_pfn + PFN_DOWN(size);
939 return offline_pages(start_pfn, end_pfn, 120 * HZ);
942 int remove_memory(u64 start, u64 size)
946 #endif /* CONFIG_MEMORY_HOTREMOVE */
947 EXPORT_SYMBOL_GPL(remove_memory);