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>
30 #include <asm/tlbflush.h>
34 /* add this memory to iomem resource */
35 static struct resource *register_memory_resource(u64 start, u64 size)
38 res = kzalloc(sizeof(struct resource), GFP_KERNEL);
41 res->name = "System RAM";
43 res->end = start + size - 1;
44 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
45 if (request_resource(&iomem_resource, res) < 0) {
46 printk("System RAM resource %llx - %llx cannot be added\n",
47 (unsigned long long)res->start, (unsigned long long)res->end);
54 static void release_memory_resource(struct resource *res)
58 release_resource(res);
63 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
64 #ifndef CONFIG_SPARSEMEM_VMEMMAP
65 static void get_page_bootmem(unsigned long info, struct page *page, int type)
67 atomic_set(&page->_mapcount, type);
69 set_page_private(page, info);
70 atomic_inc(&page->_count);
73 void put_page_bootmem(struct page *page)
77 type = atomic_read(&page->_mapcount);
80 if (atomic_dec_return(&page->_count) == 1) {
81 ClearPagePrivate(page);
82 set_page_private(page, 0);
83 reset_page_mapcount(page);
84 __free_pages_bootmem(page, 0);
89 static void register_page_bootmem_info_section(unsigned long start_pfn)
91 unsigned long *usemap, mapsize, section_nr, i;
92 struct mem_section *ms;
93 struct page *page, *memmap;
95 if (!pfn_valid(start_pfn))
98 section_nr = pfn_to_section_nr(start_pfn);
99 ms = __nr_to_section(section_nr);
101 /* Get section's memmap address */
102 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
105 * Get page for the memmap's phys address
106 * XXX: need more consideration for sparse_vmemmap...
108 page = virt_to_page(memmap);
109 mapsize = sizeof(struct page) * PAGES_PER_SECTION;
110 mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
112 /* remember memmap's page */
113 for (i = 0; i < mapsize; i++, page++)
114 get_page_bootmem(section_nr, page, SECTION_INFO);
116 usemap = __nr_to_section(section_nr)->pageblock_flags;
117 page = virt_to_page(usemap);
119 mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT;
121 for (i = 0; i < mapsize; i++, page++)
122 get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
126 void register_page_bootmem_info_node(struct pglist_data *pgdat)
128 unsigned long i, pfn, end_pfn, nr_pages;
129 int node = pgdat->node_id;
133 nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
134 page = virt_to_page(pgdat);
136 for (i = 0; i < nr_pages; i++, page++)
137 get_page_bootmem(node, page, NODE_INFO);
139 zone = &pgdat->node_zones[0];
140 for (; zone < pgdat->node_zones + MAX_NR_ZONES - 1; zone++) {
141 if (zone->wait_table) {
142 nr_pages = zone->wait_table_hash_nr_entries
143 * sizeof(wait_queue_head_t);
144 nr_pages = PAGE_ALIGN(nr_pages) >> PAGE_SHIFT;
145 page = virt_to_page(zone->wait_table);
147 for (i = 0; i < nr_pages; i++, page++)
148 get_page_bootmem(node, page, NODE_INFO);
152 pfn = pgdat->node_start_pfn;
153 end_pfn = pfn + pgdat->node_spanned_pages;
155 /* register_section info */
156 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION)
157 register_page_bootmem_info_section(pfn);
160 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
162 static void grow_zone_span(struct zone *zone, unsigned long start_pfn,
163 unsigned long end_pfn)
165 unsigned long old_zone_end_pfn;
167 zone_span_writelock(zone);
169 old_zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
170 if (start_pfn < zone->zone_start_pfn)
171 zone->zone_start_pfn = start_pfn;
173 zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
174 zone->zone_start_pfn;
176 zone_span_writeunlock(zone);
179 static void grow_pgdat_span(struct pglist_data *pgdat, unsigned long start_pfn,
180 unsigned long end_pfn)
182 unsigned long old_pgdat_end_pfn =
183 pgdat->node_start_pfn + pgdat->node_spanned_pages;
185 if (start_pfn < pgdat->node_start_pfn)
186 pgdat->node_start_pfn = start_pfn;
188 pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
189 pgdat->node_start_pfn;
192 static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
194 struct pglist_data *pgdat = zone->zone_pgdat;
195 int nr_pages = PAGES_PER_SECTION;
196 int nid = pgdat->node_id;
200 zone_type = zone - pgdat->node_zones;
201 if (!zone->wait_table) {
204 ret = init_currently_empty_zone(zone, phys_start_pfn,
205 nr_pages, MEMMAP_HOTPLUG);
209 pgdat_resize_lock(zone->zone_pgdat, &flags);
210 grow_zone_span(zone, phys_start_pfn, phys_start_pfn + nr_pages);
211 grow_pgdat_span(zone->zone_pgdat, phys_start_pfn,
212 phys_start_pfn + nr_pages);
213 pgdat_resize_unlock(zone->zone_pgdat, &flags);
214 memmap_init_zone(nr_pages, nid, zone_type,
215 phys_start_pfn, MEMMAP_HOTPLUG);
219 static int __meminit __add_section(int nid, struct zone *zone,
220 unsigned long phys_start_pfn)
222 int nr_pages = PAGES_PER_SECTION;
225 if (pfn_valid(phys_start_pfn))
228 ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
233 ret = __add_zone(zone, phys_start_pfn);
238 return register_new_memory(nid, __pfn_to_section(phys_start_pfn));
241 #ifdef CONFIG_SPARSEMEM_VMEMMAP
242 static int __remove_section(struct zone *zone, struct mem_section *ms)
245 * XXX: Freeing memmap with vmemmap is not implement yet.
246 * This should be removed later.
251 static int __remove_section(struct zone *zone, struct mem_section *ms)
254 struct pglist_data *pgdat = zone->zone_pgdat;
257 if (!valid_section(ms))
260 ret = unregister_memory_section(ms);
264 pgdat_resize_lock(pgdat, &flags);
265 sparse_remove_one_section(zone, ms);
266 pgdat_resize_unlock(pgdat, &flags);
272 * Reasonably generic function for adding memory. It is
273 * expected that archs that support memory hotplug will
274 * call this function after deciding the zone to which to
277 int __ref __add_pages(int nid, struct zone *zone, unsigned long phys_start_pfn,
278 unsigned long nr_pages)
282 int start_sec, end_sec;
283 /* during initialize mem_map, align hot-added range to section */
284 start_sec = pfn_to_section_nr(phys_start_pfn);
285 end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
287 for (i = start_sec; i <= end_sec; i++) {
288 err = __add_section(nid, zone, i << PFN_SECTION_SHIFT);
291 * EEXIST is finally dealt with by ioresource collision
292 * check. see add_memory() => register_memory_resource()
293 * Warning will be printed if there is collision.
295 if (err && (err != -EEXIST))
302 EXPORT_SYMBOL_GPL(__add_pages);
305 * __remove_pages() - remove sections of pages from a zone
306 * @zone: zone from which pages need to be removed
307 * @phys_start_pfn: starting pageframe (must be aligned to start of a section)
308 * @nr_pages: number of pages to remove (must be multiple of section size)
310 * Generic helper function to remove section mappings and sysfs entries
311 * for the section of the memory we are removing. Caller needs to make
312 * sure that pages are marked reserved and zones are adjust properly by
313 * calling offline_pages().
315 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
316 unsigned long nr_pages)
318 unsigned long i, ret = 0;
319 int sections_to_remove;
322 * We can only remove entire sections
324 BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
325 BUG_ON(nr_pages % PAGES_PER_SECTION);
327 sections_to_remove = nr_pages / PAGES_PER_SECTION;
328 for (i = 0; i < sections_to_remove; i++) {
329 unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION;
330 release_mem_region(pfn << PAGE_SHIFT,
331 PAGES_PER_SECTION << PAGE_SHIFT);
332 ret = __remove_section(zone, __pfn_to_section(pfn));
338 EXPORT_SYMBOL_GPL(__remove_pages);
340 void online_page(struct page *page)
345 #ifdef CONFIG_HIGHMEM
346 if (PageHighMem(page))
350 #ifdef CONFIG_FLATMEM
351 max_mapnr = max(page_to_pfn(page), max_mapnr);
354 ClearPageReserved(page);
355 init_page_count(page);
359 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
363 unsigned long onlined_pages = *(unsigned long *)arg;
365 if (PageReserved(pfn_to_page(start_pfn)))
366 for (i = 0; i < nr_pages; i++) {
367 page = pfn_to_page(start_pfn + i);
371 *(unsigned long *)arg = onlined_pages;
376 int online_pages(unsigned long pfn, unsigned long nr_pages)
378 unsigned long onlined_pages = 0;
380 int need_zonelists_rebuild = 0;
383 struct memory_notify arg;
386 arg.nr_pages = nr_pages;
387 arg.status_change_nid = -1;
389 nid = page_to_nid(pfn_to_page(pfn));
390 if (node_present_pages(nid) == 0)
391 arg.status_change_nid = nid;
393 ret = memory_notify(MEM_GOING_ONLINE, &arg);
394 ret = notifier_to_errno(ret);
396 memory_notify(MEM_CANCEL_ONLINE, &arg);
400 * This doesn't need a lock to do pfn_to_page().
401 * The section can't be removed here because of the
402 * memory_block->state_mutex.
404 zone = page_zone(pfn_to_page(pfn));
406 * If this zone is not populated, then it is not in zonelist.
407 * This means the page allocator ignores this zone.
408 * So, zonelist must be updated after online.
410 if (!populated_zone(zone))
411 need_zonelists_rebuild = 1;
413 ret = walk_memory_resource(pfn, nr_pages, &onlined_pages,
416 printk(KERN_DEBUG "online_pages %lx at %lx failed\n",
418 memory_notify(MEM_CANCEL_ONLINE, &arg);
422 zone->present_pages += onlined_pages;
423 zone->zone_pgdat->node_present_pages += onlined_pages;
425 setup_per_zone_wmarks();
426 calculate_zone_inactive_ratio(zone);
428 kswapd_run(zone_to_nid(zone));
429 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
432 if (need_zonelists_rebuild)
433 build_all_zonelists();
435 vm_total_pages = nr_free_pagecache_pages();
437 writeback_set_ratelimit();
440 memory_notify(MEM_ONLINE, &arg);
444 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
446 static pg_data_t *hotadd_new_pgdat(int nid, u64 start)
448 struct pglist_data *pgdat;
449 unsigned long zones_size[MAX_NR_ZONES] = {0};
450 unsigned long zholes_size[MAX_NR_ZONES] = {0};
451 unsigned long start_pfn = start >> PAGE_SHIFT;
453 pgdat = arch_alloc_nodedata(nid);
457 arch_refresh_nodedata(nid, pgdat);
459 /* we can use NODE_DATA(nid) from here */
461 /* init node's zones as empty zones, we don't have any present pages.*/
462 free_area_init_node(nid, zones_size, start_pfn, zholes_size);
467 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
469 arch_refresh_nodedata(nid, NULL);
470 arch_free_nodedata(pgdat);
475 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
476 int __ref add_memory(int nid, u64 start, u64 size)
478 pg_data_t *pgdat = NULL;
480 struct resource *res;
483 res = register_memory_resource(start, size);
487 if (!node_online(nid)) {
488 pgdat = hotadd_new_pgdat(nid, start);
494 /* call arch's memory hotadd */
495 ret = arch_add_memory(nid, start, size);
500 /* we online node here. we can't roll back from here. */
501 node_set_online(nid);
504 ret = register_one_node(nid);
506 * If sysfs file of new node can't create, cpu on the node
507 * can't be hot-added. There is no rollback way now.
508 * So, check by BUG_ON() to catch it reluctantly..
515 /* rollback pgdat allocation and others */
517 rollback_node_hotadd(nid, pgdat);
519 release_memory_resource(res);
523 EXPORT_SYMBOL_GPL(add_memory);
525 #ifdef CONFIG_MEMORY_HOTREMOVE
527 * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
528 * set and the size of the free page is given by page_order(). Using this,
529 * the function determines if the pageblock contains only free pages.
530 * Due to buddy contraints, a free page at least the size of a pageblock will
531 * be located at the start of the pageblock
533 static inline int pageblock_free(struct page *page)
535 return PageBuddy(page) && page_order(page) >= pageblock_order;
538 /* Return the start of the next active pageblock after a given page */
539 static struct page *next_active_pageblock(struct page *page)
541 int pageblocks_stride;
543 /* Ensure the starting page is pageblock-aligned */
544 BUG_ON(page_to_pfn(page) & (pageblock_nr_pages - 1));
546 /* Move forward by at least 1 * pageblock_nr_pages */
547 pageblocks_stride = 1;
549 /* If the entire pageblock is free, move to the end of free page */
550 if (pageblock_free(page))
551 pageblocks_stride += page_order(page) - pageblock_order;
553 return page + (pageblocks_stride * pageblock_nr_pages);
556 /* Checks if this range of memory is likely to be hot-removable. */
557 int is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
560 struct page *page = pfn_to_page(start_pfn);
561 struct page *end_page = page + nr_pages;
563 /* Check the starting page of each pageblock within the range */
564 for (; page < end_page; page = next_active_pageblock(page)) {
565 type = get_pageblock_migratetype(page);
568 * A pageblock containing MOVABLE or free pages is considered
571 if (type != MIGRATE_MOVABLE && !pageblock_free(page))
575 * A pageblock starting with a PageReserved page is not
576 * considered removable.
578 if (PageReserved(page))
582 /* All pageblocks in the memory block are likely to be hot-removable */
587 * Confirm all pages in a range [start, end) is belongs to the same zone.
589 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
592 struct zone *zone = NULL;
595 for (pfn = start_pfn;
597 pfn += MAX_ORDER_NR_PAGES) {
599 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
600 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
602 if (i == MAX_ORDER_NR_PAGES)
604 page = pfn_to_page(pfn + i);
605 if (zone && page_zone(page) != zone)
607 zone = page_zone(page);
613 * Scanning pfn is much easier than scanning lru list.
614 * Scan pfn from start to end and Find LRU page.
616 int scan_lru_pages(unsigned long start, unsigned long end)
620 for (pfn = start; pfn < end; pfn++) {
621 if (pfn_valid(pfn)) {
622 page = pfn_to_page(pfn);
631 hotremove_migrate_alloc(struct page *page, unsigned long private, int **x)
633 /* This should be improooooved!! */
634 return alloc_page(GFP_HIGHUSER_MOVABLE);
637 #define NR_OFFLINE_AT_ONCE_PAGES (256)
639 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
643 int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
648 for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
651 page = pfn_to_page(pfn);
652 if (!page_count(page))
655 * We can skip free pages. And we can only deal with pages on
658 ret = isolate_lru_page(page);
659 if (!ret) { /* Success */
660 list_add_tail(&page->lru, &source);
663 /* Becasue we don't have big zone->lock. we should
664 check this again here. */
665 if (page_count(page))
667 #ifdef CONFIG_DEBUG_VM
668 printk(KERN_INFO "removing from LRU failed"
670 pfn, page_count(page), page->flags);
676 if (!list_empty(&source))
677 putback_lru_pages(&source);
681 if (list_empty(&source))
683 /* this function returns # of failed pages */
684 ret = migrate_pages(&source, hotremove_migrate_alloc, 0);
691 * remove from free_area[] and mark all as Reserved.
694 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
697 __offline_isolated_pages(start, start + nr_pages);
702 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
704 walk_memory_resource(start_pfn, end_pfn - start_pfn, NULL,
705 offline_isolated_pages_cb);
709 * Check all pages in range, recoreded as memory resource, are isolated.
712 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
716 long offlined = *(long *)data;
717 ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
720 *(long *)data += offlined;
725 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
730 ret = walk_memory_resource(start_pfn, end_pfn - start_pfn, &offlined,
731 check_pages_isolated_cb);
733 offlined = (long)ret;
737 int offline_pages(unsigned long start_pfn,
738 unsigned long end_pfn, unsigned long timeout)
740 unsigned long pfn, nr_pages, expire;
742 int ret, drain, retry_max, node;
744 struct memory_notify arg;
746 BUG_ON(start_pfn >= end_pfn);
747 /* at least, alignment against pageblock is necessary */
748 if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
750 if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
752 /* This makes hotplug much easier...and readable.
753 we assume this for now. .*/
754 if (!test_pages_in_a_zone(start_pfn, end_pfn))
757 zone = page_zone(pfn_to_page(start_pfn));
758 node = zone_to_nid(zone);
759 nr_pages = end_pfn - start_pfn;
761 /* set above range as isolated */
762 ret = start_isolate_page_range(start_pfn, end_pfn);
766 arg.start_pfn = start_pfn;
767 arg.nr_pages = nr_pages;
768 arg.status_change_nid = -1;
769 if (nr_pages >= node_present_pages(node))
770 arg.status_change_nid = node;
772 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
773 ret = notifier_to_errno(ret);
778 expire = jiffies + timeout;
782 /* start memory hot removal */
784 if (time_after(jiffies, expire))
787 if (signal_pending(current))
792 flush_scheduled_work();
797 pfn = scan_lru_pages(start_pfn, end_pfn);
798 if (pfn) { /* We have page on LRU */
799 ret = do_migrate_range(pfn, end_pfn);
805 if (--retry_max == 0)
812 /* drain all zone's lru pagevec, this is asyncronous... */
814 flush_scheduled_work();
816 /* drain pcp pages , this is synchrouns. */
819 offlined_pages = check_pages_isolated(start_pfn, end_pfn);
820 if (offlined_pages < 0) {
824 printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
825 /* Ok, all of our target is islaoted.
826 We cannot do rollback at this point. */
827 offline_isolated_pages(start_pfn, end_pfn);
828 /* reset pagetype flags and makes migrate type to be MOVABLE */
829 undo_isolate_page_range(start_pfn, end_pfn);
830 /* removal success */
831 zone->present_pages -= offlined_pages;
832 zone->zone_pgdat->node_present_pages -= offlined_pages;
833 totalram_pages -= offlined_pages;
834 num_physpages -= offlined_pages;
836 setup_per_zone_wmarks();
837 calculate_zone_inactive_ratio(zone);
839 vm_total_pages = nr_free_pagecache_pages();
840 writeback_set_ratelimit();
842 memory_notify(MEM_OFFLINE, &arg);
846 printk(KERN_INFO "memory offlining %lx to %lx failed\n",
848 memory_notify(MEM_CANCEL_OFFLINE, &arg);
849 /* pushback to free area */
850 undo_isolate_page_range(start_pfn, end_pfn);
855 int remove_memory(u64 start, u64 size)
857 unsigned long start_pfn, end_pfn;
859 start_pfn = PFN_DOWN(start);
860 end_pfn = start_pfn + PFN_DOWN(size);
861 return offline_pages(start_pfn, end_pfn, 120 * HZ);
864 int remove_memory(u64 start, u64 size)
868 #endif /* CONFIG_MEMORY_HOTREMOVE */
869 EXPORT_SYMBOL_GPL(remove_memory);