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 /* add this memory to iomem resource */
38 static struct resource *register_memory_resource(u64 start, u64 size)
41 res = kzalloc(sizeof(struct resource), GFP_KERNEL);
44 res->name = "System RAM";
46 res->end = start + size - 1;
47 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
48 if (request_resource(&iomem_resource, res) < 0) {
49 printk("System RAM resource %llx - %llx cannot be added\n",
50 (unsigned long long)res->start, (unsigned long long)res->end);
57 static void release_memory_resource(struct resource *res)
61 release_resource(res);
66 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
67 #ifndef CONFIG_SPARSEMEM_VMEMMAP
68 static void get_page_bootmem(unsigned long info, struct page *page, int type)
70 atomic_set(&page->_mapcount, type);
72 set_page_private(page, info);
73 atomic_inc(&page->_count);
76 /* reference to __meminit __free_pages_bootmem is valid
77 * so use __ref to tell modpost not to generate a warning */
78 void __ref put_page_bootmem(struct page *page)
82 type = atomic_read(&page->_mapcount);
85 if (atomic_dec_return(&page->_count) == 1) {
86 ClearPagePrivate(page);
87 set_page_private(page, 0);
88 reset_page_mapcount(page);
89 __free_pages_bootmem(page, 0);
94 static void register_page_bootmem_info_section(unsigned long start_pfn)
96 unsigned long *usemap, mapsize, section_nr, i;
97 struct mem_section *ms;
98 struct page *page, *memmap;
100 if (!pfn_valid(start_pfn))
103 section_nr = pfn_to_section_nr(start_pfn);
104 ms = __nr_to_section(section_nr);
106 /* Get section's memmap address */
107 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
110 * Get page for the memmap's phys address
111 * XXX: need more consideration for sparse_vmemmap...
113 page = virt_to_page(memmap);
114 mapsize = sizeof(struct page) * PAGES_PER_SECTION;
115 mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
117 /* remember memmap's page */
118 for (i = 0; i < mapsize; i++, page++)
119 get_page_bootmem(section_nr, page, SECTION_INFO);
121 usemap = __nr_to_section(section_nr)->pageblock_flags;
122 page = virt_to_page(usemap);
124 mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT;
126 for (i = 0; i < mapsize; i++, page++)
127 get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
131 void register_page_bootmem_info_node(struct pglist_data *pgdat)
133 unsigned long i, pfn, end_pfn, nr_pages;
134 int node = pgdat->node_id;
138 nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
139 page = virt_to_page(pgdat);
141 for (i = 0; i < nr_pages; i++, page++)
142 get_page_bootmem(node, page, NODE_INFO);
144 zone = &pgdat->node_zones[0];
145 for (; zone < pgdat->node_zones + MAX_NR_ZONES - 1; zone++) {
146 if (zone->wait_table) {
147 nr_pages = zone->wait_table_hash_nr_entries
148 * sizeof(wait_queue_head_t);
149 nr_pages = PAGE_ALIGN(nr_pages) >> PAGE_SHIFT;
150 page = virt_to_page(zone->wait_table);
152 for (i = 0; i < nr_pages; i++, page++)
153 get_page_bootmem(node, page, NODE_INFO);
157 pfn = pgdat->node_start_pfn;
158 end_pfn = pfn + pgdat->node_spanned_pages;
160 /* register_section info */
161 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION)
162 register_page_bootmem_info_section(pfn);
165 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
167 static void grow_zone_span(struct zone *zone, unsigned long start_pfn,
168 unsigned long end_pfn)
170 unsigned long old_zone_end_pfn;
172 zone_span_writelock(zone);
174 old_zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
175 if (start_pfn < zone->zone_start_pfn)
176 zone->zone_start_pfn = start_pfn;
178 zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
179 zone->zone_start_pfn;
181 zone_span_writeunlock(zone);
184 static void grow_pgdat_span(struct pglist_data *pgdat, unsigned long start_pfn,
185 unsigned long end_pfn)
187 unsigned long old_pgdat_end_pfn =
188 pgdat->node_start_pfn + pgdat->node_spanned_pages;
190 if (start_pfn < pgdat->node_start_pfn)
191 pgdat->node_start_pfn = start_pfn;
193 pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
194 pgdat->node_start_pfn;
197 static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
199 struct pglist_data *pgdat = zone->zone_pgdat;
200 int nr_pages = PAGES_PER_SECTION;
201 int nid = pgdat->node_id;
205 zone_type = zone - pgdat->node_zones;
206 if (!zone->wait_table) {
209 ret = init_currently_empty_zone(zone, phys_start_pfn,
210 nr_pages, MEMMAP_HOTPLUG);
214 pgdat_resize_lock(zone->zone_pgdat, &flags);
215 grow_zone_span(zone, phys_start_pfn, phys_start_pfn + nr_pages);
216 grow_pgdat_span(zone->zone_pgdat, phys_start_pfn,
217 phys_start_pfn + nr_pages);
218 pgdat_resize_unlock(zone->zone_pgdat, &flags);
219 memmap_init_zone(nr_pages, nid, zone_type,
220 phys_start_pfn, MEMMAP_HOTPLUG);
224 static int __meminit __add_section(int nid, struct zone *zone,
225 unsigned long phys_start_pfn)
227 int nr_pages = PAGES_PER_SECTION;
230 if (pfn_valid(phys_start_pfn))
233 ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
238 ret = __add_zone(zone, phys_start_pfn);
243 return register_new_memory(nid, __pfn_to_section(phys_start_pfn));
246 #ifdef CONFIG_SPARSEMEM_VMEMMAP
247 static int __remove_section(struct zone *zone, struct mem_section *ms)
250 * XXX: Freeing memmap with vmemmap is not implement yet.
251 * This should be removed later.
256 static int __remove_section(struct zone *zone, struct mem_section *ms)
259 struct pglist_data *pgdat = zone->zone_pgdat;
262 if (!valid_section(ms))
265 ret = unregister_memory_section(ms);
269 pgdat_resize_lock(pgdat, &flags);
270 sparse_remove_one_section(zone, ms);
271 pgdat_resize_unlock(pgdat, &flags);
277 * Reasonably generic function for adding memory. It is
278 * expected that archs that support memory hotplug will
279 * call this function after deciding the zone to which to
282 int __ref __add_pages(int nid, struct zone *zone, unsigned long phys_start_pfn,
283 unsigned long nr_pages)
287 int start_sec, end_sec;
288 /* during initialize mem_map, align hot-added range to section */
289 start_sec = pfn_to_section_nr(phys_start_pfn);
290 end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
292 for (i = start_sec; i <= end_sec; i++) {
293 err = __add_section(nid, zone, i << PFN_SECTION_SHIFT);
296 * EEXIST is finally dealt with by ioresource collision
297 * check. see add_memory() => register_memory_resource()
298 * Warning will be printed if there is collision.
300 if (err && (err != -EEXIST))
307 EXPORT_SYMBOL_GPL(__add_pages);
310 * __remove_pages() - remove sections of pages from a zone
311 * @zone: zone from which pages need to be removed
312 * @phys_start_pfn: starting pageframe (must be aligned to start of a section)
313 * @nr_pages: number of pages to remove (must be multiple of section size)
315 * Generic helper function to remove section mappings and sysfs entries
316 * for the section of the memory we are removing. Caller needs to make
317 * sure that pages are marked reserved and zones are adjust properly by
318 * calling offline_pages().
320 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
321 unsigned long nr_pages)
323 unsigned long i, ret = 0;
324 int sections_to_remove;
327 * We can only remove entire sections
329 BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
330 BUG_ON(nr_pages % PAGES_PER_SECTION);
332 sections_to_remove = nr_pages / PAGES_PER_SECTION;
333 for (i = 0; i < sections_to_remove; i++) {
334 unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION;
335 release_mem_region(pfn << PAGE_SHIFT,
336 PAGES_PER_SECTION << PAGE_SHIFT);
337 ret = __remove_section(zone, __pfn_to_section(pfn));
343 EXPORT_SYMBOL_GPL(__remove_pages);
345 void online_page(struct page *page)
347 unsigned long pfn = page_to_pfn(page);
350 if (pfn >= num_physpages)
351 num_physpages = pfn + 1;
353 #ifdef CONFIG_HIGHMEM
354 if (PageHighMem(page))
358 #ifdef CONFIG_FLATMEM
359 max_mapnr = max(page_to_pfn(page), max_mapnr);
362 ClearPageReserved(page);
363 init_page_count(page);
367 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
371 unsigned long onlined_pages = *(unsigned long *)arg;
373 if (PageReserved(pfn_to_page(start_pfn)))
374 for (i = 0; i < nr_pages; i++) {
375 page = pfn_to_page(start_pfn + i);
379 *(unsigned long *)arg = onlined_pages;
384 int online_pages(unsigned long pfn, unsigned long nr_pages)
386 unsigned long onlined_pages = 0;
388 int need_zonelists_rebuild = 0;
391 struct memory_notify arg;
394 arg.nr_pages = nr_pages;
395 arg.status_change_nid = -1;
397 nid = page_to_nid(pfn_to_page(pfn));
398 if (node_present_pages(nid) == 0)
399 arg.status_change_nid = nid;
401 ret = memory_notify(MEM_GOING_ONLINE, &arg);
402 ret = notifier_to_errno(ret);
404 memory_notify(MEM_CANCEL_ONLINE, &arg);
408 * This doesn't need a lock to do pfn_to_page().
409 * The section can't be removed here because of the
410 * memory_block->state_mutex.
412 zone = page_zone(pfn_to_page(pfn));
414 * If this zone is not populated, then it is not in zonelist.
415 * This means the page allocator ignores this zone.
416 * So, zonelist must be updated after online.
418 if (!populated_zone(zone))
419 need_zonelists_rebuild = 1;
421 ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages,
424 printk(KERN_DEBUG "online_pages %lx at %lx failed\n",
426 memory_notify(MEM_CANCEL_ONLINE, &arg);
430 zone->present_pages += onlined_pages;
431 zone->zone_pgdat->node_present_pages += onlined_pages;
433 zone_pcp_update(zone);
434 setup_per_zone_wmarks();
435 calculate_zone_inactive_ratio(zone);
437 kswapd_run(zone_to_nid(zone));
438 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
441 if (need_zonelists_rebuild)
442 build_all_zonelists();
444 vm_total_pages = nr_free_pagecache_pages();
446 writeback_set_ratelimit();
449 memory_notify(MEM_ONLINE, &arg);
453 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
455 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
456 static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
458 struct pglist_data *pgdat;
459 unsigned long zones_size[MAX_NR_ZONES] = {0};
460 unsigned long zholes_size[MAX_NR_ZONES] = {0};
461 unsigned long start_pfn = start >> PAGE_SHIFT;
463 pgdat = arch_alloc_nodedata(nid);
467 arch_refresh_nodedata(nid, pgdat);
469 /* we can use NODE_DATA(nid) from here */
471 /* init node's zones as empty zones, we don't have any present pages.*/
472 free_area_init_node(nid, zones_size, start_pfn, zholes_size);
477 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
479 arch_refresh_nodedata(nid, NULL);
480 arch_free_nodedata(pgdat);
485 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
486 int __ref add_memory(int nid, u64 start, u64 size)
488 pg_data_t *pgdat = NULL;
490 struct resource *res;
495 res = register_memory_resource(start, size);
500 if (!node_online(nid)) {
501 pgdat = hotadd_new_pgdat(nid, start);
508 /* call arch's memory hotadd */
509 ret = arch_add_memory(nid, start, size);
514 /* we online node here. we can't roll back from here. */
515 node_set_online(nid);
518 ret = register_one_node(nid);
520 * If sysfs file of new node can't create, cpu on the node
521 * can't be hot-added. There is no rollback way now.
522 * So, check by BUG_ON() to catch it reluctantly..
527 /* create new memmap entry */
528 firmware_map_add_hotplug(start, start + size, "System RAM");
533 /* rollback pgdat allocation and others */
535 rollback_node_hotadd(nid, pgdat);
537 release_memory_resource(res);
540 unlock_system_sleep();
543 EXPORT_SYMBOL_GPL(add_memory);
545 #ifdef CONFIG_MEMORY_HOTREMOVE
547 * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
548 * set and the size of the free page is given by page_order(). Using this,
549 * the function determines if the pageblock contains only free pages.
550 * Due to buddy contraints, a free page at least the size of a pageblock will
551 * be located at the start of the pageblock
553 static inline int pageblock_free(struct page *page)
555 return PageBuddy(page) && page_order(page) >= pageblock_order;
558 /* Return the start of the next active pageblock after a given page */
559 static struct page *next_active_pageblock(struct page *page)
561 int pageblocks_stride;
563 /* Ensure the starting page is pageblock-aligned */
564 BUG_ON(page_to_pfn(page) & (pageblock_nr_pages - 1));
566 /* Move forward by at least 1 * pageblock_nr_pages */
567 pageblocks_stride = 1;
569 /* If the entire pageblock is free, move to the end of free page */
570 if (pageblock_free(page))
571 pageblocks_stride += page_order(page) - pageblock_order;
573 return page + (pageblocks_stride * pageblock_nr_pages);
576 /* Checks if this range of memory is likely to be hot-removable. */
577 int is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
580 struct page *page = pfn_to_page(start_pfn);
581 struct page *end_page = page + nr_pages;
583 /* Check the starting page of each pageblock within the range */
584 for (; page < end_page; page = next_active_pageblock(page)) {
585 type = get_pageblock_migratetype(page);
588 * A pageblock containing MOVABLE or free pages is considered
591 if (type != MIGRATE_MOVABLE && !pageblock_free(page))
595 * A pageblock starting with a PageReserved page is not
596 * considered removable.
598 if (PageReserved(page))
602 /* All pageblocks in the memory block are likely to be hot-removable */
607 * Confirm all pages in a range [start, end) is belongs to the same zone.
609 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
612 struct zone *zone = NULL;
615 for (pfn = start_pfn;
617 pfn += MAX_ORDER_NR_PAGES) {
619 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
620 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
622 if (i == MAX_ORDER_NR_PAGES)
624 page = pfn_to_page(pfn + i);
625 if (zone && page_zone(page) != zone)
627 zone = page_zone(page);
633 * Scanning pfn is much easier than scanning lru list.
634 * Scan pfn from start to end and Find LRU page.
636 int scan_lru_pages(unsigned long start, unsigned long end)
640 for (pfn = start; pfn < end; pfn++) {
641 if (pfn_valid(pfn)) {
642 page = pfn_to_page(pfn);
651 hotremove_migrate_alloc(struct page *page, unsigned long private, int **x)
653 /* This should be improooooved!! */
654 return alloc_page(GFP_HIGHUSER_MOVABLE);
657 #define NR_OFFLINE_AT_ONCE_PAGES (256)
659 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
663 int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
668 for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
671 page = pfn_to_page(pfn);
672 if (!page_count(page))
675 * We can skip free pages. And we can only deal with pages on
678 ret = isolate_lru_page(page);
679 if (!ret) { /* Success */
680 list_add_tail(&page->lru, &source);
682 inc_zone_page_state(page, NR_ISOLATED_ANON +
683 page_is_file_cache(page));
686 /* Becasue we don't have big zone->lock. we should
687 check this again here. */
688 if (page_count(page))
690 #ifdef CONFIG_DEBUG_VM
691 printk(KERN_ALERT "removing pfn %lx from LRU failed\n",
699 if (!list_empty(&source))
700 putback_lru_pages(&source);
704 if (list_empty(&source))
706 /* this function returns # of failed pages */
707 ret = migrate_pages(&source, hotremove_migrate_alloc, 0, 1);
714 * remove from free_area[] and mark all as Reserved.
717 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
720 __offline_isolated_pages(start, start + nr_pages);
725 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
727 walk_system_ram_range(start_pfn, end_pfn - start_pfn, NULL,
728 offline_isolated_pages_cb);
732 * Check all pages in range, recoreded as memory resource, are isolated.
735 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
739 long offlined = *(long *)data;
740 ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
743 *(long *)data += offlined;
748 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
753 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, &offlined,
754 check_pages_isolated_cb);
756 offlined = (long)ret;
760 static int offline_pages(unsigned long start_pfn,
761 unsigned long end_pfn, unsigned long timeout)
763 unsigned long pfn, nr_pages, expire;
765 int ret, drain, retry_max, node;
767 struct memory_notify arg;
769 BUG_ON(start_pfn >= end_pfn);
770 /* at least, alignment against pageblock is necessary */
771 if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
773 if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
775 /* This makes hotplug much easier...and readable.
776 we assume this for now. .*/
777 if (!test_pages_in_a_zone(start_pfn, end_pfn))
782 zone = page_zone(pfn_to_page(start_pfn));
783 node = zone_to_nid(zone);
784 nr_pages = end_pfn - start_pfn;
786 /* set above range as isolated */
787 ret = start_isolate_page_range(start_pfn, end_pfn);
791 arg.start_pfn = start_pfn;
792 arg.nr_pages = nr_pages;
793 arg.status_change_nid = -1;
794 if (nr_pages >= node_present_pages(node))
795 arg.status_change_nid = node;
797 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
798 ret = notifier_to_errno(ret);
803 expire = jiffies + timeout;
807 /* start memory hot removal */
809 if (time_after(jiffies, expire))
812 if (signal_pending(current))
817 flush_scheduled_work();
822 pfn = scan_lru_pages(start_pfn, end_pfn);
823 if (pfn) { /* We have page on LRU */
824 ret = do_migrate_range(pfn, end_pfn);
830 if (--retry_max == 0)
837 /* drain all zone's lru pagevec, this is asyncronous... */
839 flush_scheduled_work();
841 /* drain pcp pages , this is synchrouns. */
844 offlined_pages = check_pages_isolated(start_pfn, end_pfn);
845 if (offlined_pages < 0) {
849 printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
850 /* Ok, all of our target is islaoted.
851 We cannot do rollback at this point. */
852 offline_isolated_pages(start_pfn, end_pfn);
853 /* reset pagetype flags and makes migrate type to be MOVABLE */
854 undo_isolate_page_range(start_pfn, end_pfn);
855 /* removal success */
856 zone->present_pages -= offlined_pages;
857 zone->zone_pgdat->node_present_pages -= offlined_pages;
858 totalram_pages -= offlined_pages;
860 setup_per_zone_wmarks();
861 calculate_zone_inactive_ratio(zone);
862 if (!node_present_pages(node)) {
863 node_clear_state(node, N_HIGH_MEMORY);
867 vm_total_pages = nr_free_pagecache_pages();
868 writeback_set_ratelimit();
870 memory_notify(MEM_OFFLINE, &arg);
871 unlock_system_sleep();
875 printk(KERN_INFO "memory offlining %lx to %lx failed\n",
877 memory_notify(MEM_CANCEL_OFFLINE, &arg);
878 /* pushback to free area */
879 undo_isolate_page_range(start_pfn, end_pfn);
882 unlock_system_sleep();
886 int remove_memory(u64 start, u64 size)
888 unsigned long start_pfn, end_pfn;
890 start_pfn = PFN_DOWN(start);
891 end_pfn = start_pfn + PFN_DOWN(size);
892 return offline_pages(start_pfn, end_pfn, 120 * HZ);
895 int remove_memory(u64 start, u64 size)
899 #endif /* CONFIG_MEMORY_HOTREMOVE */
900 EXPORT_SYMBOL_GPL(remove_memory);