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>
32 #include <asm/tlbflush.h>
36 /* add this memory to iomem resource */
37 static struct resource *register_memory_resource(u64 start, u64 size)
40 res = kzalloc(sizeof(struct resource), GFP_KERNEL);
43 res->name = "System RAM";
45 res->end = start + size - 1;
46 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
47 if (request_resource(&iomem_resource, res) < 0) {
48 printk("System RAM resource %llx - %llx cannot be added\n",
49 (unsigned long long)res->start, (unsigned long long)res->end);
56 static void release_memory_resource(struct resource *res)
60 release_resource(res);
65 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
66 #ifndef CONFIG_SPARSEMEM_VMEMMAP
67 static void get_page_bootmem(unsigned long info, struct page *page, int type)
69 atomic_set(&page->_mapcount, type);
71 set_page_private(page, info);
72 atomic_inc(&page->_count);
75 /* reference to __meminit __free_pages_bootmem is valid
76 * so use __ref to tell modpost not to generate a warning */
77 void __ref put_page_bootmem(struct page *page)
81 type = atomic_read(&page->_mapcount);
84 if (atomic_dec_return(&page->_count) == 1) {
85 ClearPagePrivate(page);
86 set_page_private(page, 0);
87 reset_page_mapcount(page);
88 __free_pages_bootmem(page, 0);
93 static void register_page_bootmem_info_section(unsigned long start_pfn)
95 unsigned long *usemap, mapsize, section_nr, i;
96 struct mem_section *ms;
97 struct page *page, *memmap;
99 if (!pfn_valid(start_pfn))
102 section_nr = pfn_to_section_nr(start_pfn);
103 ms = __nr_to_section(section_nr);
105 /* Get section's memmap address */
106 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
109 * Get page for the memmap's phys address
110 * XXX: need more consideration for sparse_vmemmap...
112 page = virt_to_page(memmap);
113 mapsize = sizeof(struct page) * PAGES_PER_SECTION;
114 mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
116 /* remember memmap's page */
117 for (i = 0; i < mapsize; i++, page++)
118 get_page_bootmem(section_nr, page, SECTION_INFO);
120 usemap = __nr_to_section(section_nr)->pageblock_flags;
121 page = virt_to_page(usemap);
123 mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT;
125 for (i = 0; i < mapsize; i++, page++)
126 get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
130 void register_page_bootmem_info_node(struct pglist_data *pgdat)
132 unsigned long i, pfn, end_pfn, nr_pages;
133 int node = pgdat->node_id;
137 nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
138 page = virt_to_page(pgdat);
140 for (i = 0; i < nr_pages; i++, page++)
141 get_page_bootmem(node, page, NODE_INFO);
143 zone = &pgdat->node_zones[0];
144 for (; zone < pgdat->node_zones + MAX_NR_ZONES - 1; zone++) {
145 if (zone->wait_table) {
146 nr_pages = zone->wait_table_hash_nr_entries
147 * sizeof(wait_queue_head_t);
148 nr_pages = PAGE_ALIGN(nr_pages) >> PAGE_SHIFT;
149 page = virt_to_page(zone->wait_table);
151 for (i = 0; i < nr_pages; i++, page++)
152 get_page_bootmem(node, page, NODE_INFO);
156 pfn = pgdat->node_start_pfn;
157 end_pfn = pfn + pgdat->node_spanned_pages;
159 /* register_section info */
160 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION)
161 register_page_bootmem_info_section(pfn);
164 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
166 static void grow_zone_span(struct zone *zone, unsigned long start_pfn,
167 unsigned long end_pfn)
169 unsigned long old_zone_end_pfn;
171 zone_span_writelock(zone);
173 old_zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
174 if (start_pfn < zone->zone_start_pfn)
175 zone->zone_start_pfn = start_pfn;
177 zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
178 zone->zone_start_pfn;
180 zone_span_writeunlock(zone);
183 static void grow_pgdat_span(struct pglist_data *pgdat, unsigned long start_pfn,
184 unsigned long end_pfn)
186 unsigned long old_pgdat_end_pfn =
187 pgdat->node_start_pfn + pgdat->node_spanned_pages;
189 if (start_pfn < pgdat->node_start_pfn)
190 pgdat->node_start_pfn = start_pfn;
192 pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
193 pgdat->node_start_pfn;
196 static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
198 struct pglist_data *pgdat = zone->zone_pgdat;
199 int nr_pages = PAGES_PER_SECTION;
200 int nid = pgdat->node_id;
204 zone_type = zone - pgdat->node_zones;
205 if (!zone->wait_table) {
208 ret = init_currently_empty_zone(zone, phys_start_pfn,
209 nr_pages, MEMMAP_HOTPLUG);
213 pgdat_resize_lock(zone->zone_pgdat, &flags);
214 grow_zone_span(zone, phys_start_pfn, phys_start_pfn + nr_pages);
215 grow_pgdat_span(zone->zone_pgdat, phys_start_pfn,
216 phys_start_pfn + nr_pages);
217 pgdat_resize_unlock(zone->zone_pgdat, &flags);
218 memmap_init_zone(nr_pages, nid, zone_type,
219 phys_start_pfn, MEMMAP_HOTPLUG);
223 static int __meminit __add_section(int nid, struct zone *zone,
224 unsigned long phys_start_pfn)
226 int nr_pages = PAGES_PER_SECTION;
229 if (pfn_valid(phys_start_pfn))
232 ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
237 ret = __add_zone(zone, phys_start_pfn);
242 return register_new_memory(nid, __pfn_to_section(phys_start_pfn));
245 #ifdef CONFIG_SPARSEMEM_VMEMMAP
246 static int __remove_section(struct zone *zone, struct mem_section *ms)
249 * XXX: Freeing memmap with vmemmap is not implement yet.
250 * This should be removed later.
255 static int __remove_section(struct zone *zone, struct mem_section *ms)
258 struct pglist_data *pgdat = zone->zone_pgdat;
261 if (!valid_section(ms))
264 ret = unregister_memory_section(ms);
268 pgdat_resize_lock(pgdat, &flags);
269 sparse_remove_one_section(zone, ms);
270 pgdat_resize_unlock(pgdat, &flags);
276 * Reasonably generic function for adding memory. It is
277 * expected that archs that support memory hotplug will
278 * call this function after deciding the zone to which to
281 int __ref __add_pages(int nid, struct zone *zone, unsigned long phys_start_pfn,
282 unsigned long nr_pages)
286 int start_sec, end_sec;
287 /* during initialize mem_map, align hot-added range to section */
288 start_sec = pfn_to_section_nr(phys_start_pfn);
289 end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
291 for (i = start_sec; i <= end_sec; i++) {
292 err = __add_section(nid, zone, i << PFN_SECTION_SHIFT);
295 * EEXIST is finally dealt with by ioresource collision
296 * check. see add_memory() => register_memory_resource()
297 * Warning will be printed if there is collision.
299 if (err && (err != -EEXIST))
306 EXPORT_SYMBOL_GPL(__add_pages);
309 * __remove_pages() - remove sections of pages from a zone
310 * @zone: zone from which pages need to be removed
311 * @phys_start_pfn: starting pageframe (must be aligned to start of a section)
312 * @nr_pages: number of pages to remove (must be multiple of section size)
314 * Generic helper function to remove section mappings and sysfs entries
315 * for the section of the memory we are removing. Caller needs to make
316 * sure that pages are marked reserved and zones are adjust properly by
317 * calling offline_pages().
319 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
320 unsigned long nr_pages)
322 unsigned long i, ret = 0;
323 int sections_to_remove;
326 * We can only remove entire sections
328 BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
329 BUG_ON(nr_pages % PAGES_PER_SECTION);
331 sections_to_remove = nr_pages / PAGES_PER_SECTION;
332 for (i = 0; i < sections_to_remove; i++) {
333 unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION;
334 release_mem_region(pfn << PAGE_SHIFT,
335 PAGES_PER_SECTION << PAGE_SHIFT);
336 ret = __remove_section(zone, __pfn_to_section(pfn));
342 EXPORT_SYMBOL_GPL(__remove_pages);
344 void online_page(struct page *page)
346 unsigned long pfn = page_to_pfn(page);
349 if (pfn >= num_physpages)
350 num_physpages = pfn + 1;
352 #ifdef CONFIG_HIGHMEM
353 if (PageHighMem(page))
357 #ifdef CONFIG_FLATMEM
358 max_mapnr = max(page_to_pfn(page), max_mapnr);
361 ClearPageReserved(page);
362 init_page_count(page);
366 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
370 unsigned long onlined_pages = *(unsigned long *)arg;
372 if (PageReserved(pfn_to_page(start_pfn)))
373 for (i = 0; i < nr_pages; i++) {
374 page = pfn_to_page(start_pfn + i);
378 *(unsigned long *)arg = onlined_pages;
383 int online_pages(unsigned long pfn, unsigned long nr_pages)
385 unsigned long onlined_pages = 0;
387 int need_zonelists_rebuild = 0;
390 struct memory_notify arg;
393 arg.nr_pages = nr_pages;
394 arg.status_change_nid = -1;
396 nid = page_to_nid(pfn_to_page(pfn));
397 if (node_present_pages(nid) == 0)
398 arg.status_change_nid = nid;
400 ret = memory_notify(MEM_GOING_ONLINE, &arg);
401 ret = notifier_to_errno(ret);
403 memory_notify(MEM_CANCEL_ONLINE, &arg);
407 * This doesn't need a lock to do pfn_to_page().
408 * The section can't be removed here because of the
409 * memory_block->state_mutex.
411 zone = page_zone(pfn_to_page(pfn));
413 * If this zone is not populated, then it is not in zonelist.
414 * This means the page allocator ignores this zone.
415 * So, zonelist must be updated after online.
417 if (!populated_zone(zone))
418 need_zonelists_rebuild = 1;
420 ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages,
423 printk(KERN_DEBUG "online_pages %lx at %lx failed\n",
425 memory_notify(MEM_CANCEL_ONLINE, &arg);
429 zone->present_pages += onlined_pages;
430 zone->zone_pgdat->node_present_pages += onlined_pages;
432 zone_pcp_update(zone);
433 setup_per_zone_wmarks();
434 calculate_zone_inactive_ratio(zone);
436 kswapd_run(zone_to_nid(zone));
437 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
440 if (need_zonelists_rebuild)
441 build_all_zonelists();
443 vm_total_pages = nr_free_pagecache_pages();
445 writeback_set_ratelimit();
448 memory_notify(MEM_ONLINE, &arg);
452 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
454 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
455 static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
457 struct pglist_data *pgdat;
458 unsigned long zones_size[MAX_NR_ZONES] = {0};
459 unsigned long zholes_size[MAX_NR_ZONES] = {0};
460 unsigned long start_pfn = start >> PAGE_SHIFT;
462 pgdat = arch_alloc_nodedata(nid);
466 arch_refresh_nodedata(nid, pgdat);
468 /* we can use NODE_DATA(nid) from here */
470 /* init node's zones as empty zones, we don't have any present pages.*/
471 free_area_init_node(nid, zones_size, start_pfn, zholes_size);
476 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
478 arch_refresh_nodedata(nid, NULL);
479 arch_free_nodedata(pgdat);
484 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
485 int __ref add_memory(int nid, u64 start, u64 size)
487 pg_data_t *pgdat = NULL;
489 struct resource *res;
494 res = register_memory_resource(start, size);
499 if (!node_online(nid)) {
500 pgdat = hotadd_new_pgdat(nid, start);
507 /* call arch's memory hotadd */
508 ret = arch_add_memory(nid, start, size);
513 /* we online node here. we can't roll back from here. */
514 node_set_online(nid);
517 ret = register_one_node(nid);
519 * If sysfs file of new node can't create, cpu on the node
520 * can't be hot-added. There is no rollback way now.
521 * So, check by BUG_ON() to catch it reluctantly..
529 /* rollback pgdat allocation and others */
531 rollback_node_hotadd(nid, pgdat);
533 release_memory_resource(res);
536 unlock_system_sleep();
539 EXPORT_SYMBOL_GPL(add_memory);
541 #ifdef CONFIG_MEMORY_HOTREMOVE
543 * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
544 * set and the size of the free page is given by page_order(). Using this,
545 * the function determines if the pageblock contains only free pages.
546 * Due to buddy contraints, a free page at least the size of a pageblock will
547 * be located at the start of the pageblock
549 static inline int pageblock_free(struct page *page)
551 return PageBuddy(page) && page_order(page) >= pageblock_order;
554 /* Return the start of the next active pageblock after a given page */
555 static struct page *next_active_pageblock(struct page *page)
557 int pageblocks_stride;
559 /* Ensure the starting page is pageblock-aligned */
560 BUG_ON(page_to_pfn(page) & (pageblock_nr_pages - 1));
562 /* Move forward by at least 1 * pageblock_nr_pages */
563 pageblocks_stride = 1;
565 /* If the entire pageblock is free, move to the end of free page */
566 if (pageblock_free(page))
567 pageblocks_stride += page_order(page) - pageblock_order;
569 return page + (pageblocks_stride * pageblock_nr_pages);
572 /* Checks if this range of memory is likely to be hot-removable. */
573 int is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
576 struct page *page = pfn_to_page(start_pfn);
577 struct page *end_page = page + nr_pages;
579 /* Check the starting page of each pageblock within the range */
580 for (; page < end_page; page = next_active_pageblock(page)) {
581 type = get_pageblock_migratetype(page);
584 * A pageblock containing MOVABLE or free pages is considered
587 if (type != MIGRATE_MOVABLE && !pageblock_free(page))
591 * A pageblock starting with a PageReserved page is not
592 * considered removable.
594 if (PageReserved(page))
598 /* All pageblocks in the memory block are likely to be hot-removable */
603 * Confirm all pages in a range [start, end) is belongs to the same zone.
605 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
608 struct zone *zone = NULL;
611 for (pfn = start_pfn;
613 pfn += MAX_ORDER_NR_PAGES) {
615 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
616 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
618 if (i == MAX_ORDER_NR_PAGES)
620 page = pfn_to_page(pfn + i);
621 if (zone && page_zone(page) != zone)
623 zone = page_zone(page);
629 * Scanning pfn is much easier than scanning lru list.
630 * Scan pfn from start to end and Find LRU page.
632 int scan_lru_pages(unsigned long start, unsigned long end)
636 for (pfn = start; pfn < end; pfn++) {
637 if (pfn_valid(pfn)) {
638 page = pfn_to_page(pfn);
647 hotremove_migrate_alloc(struct page *page, unsigned long private, int **x)
649 /* This should be improooooved!! */
650 return alloc_page(GFP_HIGHUSER_MOVABLE);
653 #define NR_OFFLINE_AT_ONCE_PAGES (256)
655 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
659 int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
664 for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
667 page = pfn_to_page(pfn);
668 if (!page_count(page))
671 * We can skip free pages. And we can only deal with pages on
674 ret = isolate_lru_page(page);
675 if (!ret) { /* Success */
676 list_add_tail(&page->lru, &source);
678 inc_zone_page_state(page, NR_ISOLATED_ANON +
679 page_is_file_cache(page));
682 /* Becasue we don't have big zone->lock. we should
683 check this again here. */
684 if (page_count(page))
686 #ifdef CONFIG_DEBUG_VM
687 printk(KERN_INFO "removing from LRU failed"
689 pfn, page_count(page), page->flags);
695 if (!list_empty(&source))
696 putback_lru_pages(&source);
700 if (list_empty(&source))
702 /* this function returns # of failed pages */
703 ret = migrate_pages(&source, hotremove_migrate_alloc, 0, 1);
710 * remove from free_area[] and mark all as Reserved.
713 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
716 __offline_isolated_pages(start, start + nr_pages);
721 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
723 walk_system_ram_range(start_pfn, end_pfn - start_pfn, NULL,
724 offline_isolated_pages_cb);
728 * Check all pages in range, recoreded as memory resource, are isolated.
731 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
735 long offlined = *(long *)data;
736 ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
739 *(long *)data += offlined;
744 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
749 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, &offlined,
750 check_pages_isolated_cb);
752 offlined = (long)ret;
756 static int offline_pages(unsigned long start_pfn,
757 unsigned long end_pfn, unsigned long timeout)
759 unsigned long pfn, nr_pages, expire;
761 int ret, drain, retry_max, node;
763 struct memory_notify arg;
765 BUG_ON(start_pfn >= end_pfn);
766 /* at least, alignment against pageblock is necessary */
767 if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
769 if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
771 /* This makes hotplug much easier...and readable.
772 we assume this for now. .*/
773 if (!test_pages_in_a_zone(start_pfn, end_pfn))
778 zone = page_zone(pfn_to_page(start_pfn));
779 node = zone_to_nid(zone);
780 nr_pages = end_pfn - start_pfn;
782 /* set above range as isolated */
783 ret = start_isolate_page_range(start_pfn, end_pfn);
787 arg.start_pfn = start_pfn;
788 arg.nr_pages = nr_pages;
789 arg.status_change_nid = -1;
790 if (nr_pages >= node_present_pages(node))
791 arg.status_change_nid = node;
793 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
794 ret = notifier_to_errno(ret);
799 expire = jiffies + timeout;
803 /* start memory hot removal */
805 if (time_after(jiffies, expire))
808 if (signal_pending(current))
813 flush_scheduled_work();
818 pfn = scan_lru_pages(start_pfn, end_pfn);
819 if (pfn) { /* We have page on LRU */
820 ret = do_migrate_range(pfn, end_pfn);
826 if (--retry_max == 0)
833 /* drain all zone's lru pagevec, this is asyncronous... */
835 flush_scheduled_work();
837 /* drain pcp pages , this is synchrouns. */
840 offlined_pages = check_pages_isolated(start_pfn, end_pfn);
841 if (offlined_pages < 0) {
845 printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
846 /* Ok, all of our target is islaoted.
847 We cannot do rollback at this point. */
848 offline_isolated_pages(start_pfn, end_pfn);
849 /* reset pagetype flags and makes migrate type to be MOVABLE */
850 undo_isolate_page_range(start_pfn, end_pfn);
851 /* removal success */
852 zone->present_pages -= offlined_pages;
853 zone->zone_pgdat->node_present_pages -= offlined_pages;
854 totalram_pages -= offlined_pages;
856 setup_per_zone_wmarks();
857 calculate_zone_inactive_ratio(zone);
858 if (!node_present_pages(node)) {
859 node_clear_state(node, N_HIGH_MEMORY);
863 vm_total_pages = nr_free_pagecache_pages();
864 writeback_set_ratelimit();
866 memory_notify(MEM_OFFLINE, &arg);
867 unlock_system_sleep();
871 printk(KERN_INFO "memory offlining %lx to %lx failed\n",
873 memory_notify(MEM_CANCEL_OFFLINE, &arg);
874 /* pushback to free area */
875 undo_isolate_page_range(start_pfn, end_pfn);
878 unlock_system_sleep();
882 int remove_memory(u64 start, u64 size)
884 unsigned long start_pfn, end_pfn;
886 start_pfn = PFN_DOWN(start);
887 end_pfn = start_pfn + PFN_DOWN(size);
888 return offline_pages(start_pfn, end_pfn, 120 * HZ);
891 int remove_memory(u64 start, u64 size)
895 #endif /* CONFIG_MEMORY_HOTREMOVE */
896 EXPORT_SYMBOL_GPL(remove_memory);