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/cpuset.h>
26 #include <linux/delay.h>
27 #include <linux/migrate.h>
28 #include <linux/page-isolation.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 __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 __add_section(struct zone *zone, unsigned long phys_start_pfn)
221 int nr_pages = PAGES_PER_SECTION;
224 if (pfn_valid(phys_start_pfn))
227 ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
232 ret = __add_zone(zone, phys_start_pfn);
237 return register_new_memory(__pfn_to_section(phys_start_pfn));
240 #ifdef CONFIG_SPARSEMEM_VMEMMAP
241 static int __remove_section(struct zone *zone, struct mem_section *ms)
244 * XXX: Freeing memmap with vmemmap is not implement yet.
245 * This should be removed later.
250 static int __remove_section(struct zone *zone, struct mem_section *ms)
253 struct pglist_data *pgdat = zone->zone_pgdat;
256 if (!valid_section(ms))
259 ret = unregister_memory_section(ms);
263 pgdat_resize_lock(pgdat, &flags);
264 sparse_remove_one_section(zone, ms);
265 pgdat_resize_unlock(pgdat, &flags);
271 * Reasonably generic function for adding memory. It is
272 * expected that archs that support memory hotplug will
273 * call this function after deciding the zone to which to
276 int __add_pages(struct zone *zone, unsigned long phys_start_pfn,
277 unsigned long nr_pages)
281 int start_sec, end_sec;
282 /* during initialize mem_map, align hot-added range to section */
283 start_sec = pfn_to_section_nr(phys_start_pfn);
284 end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
286 for (i = start_sec; i <= end_sec; i++) {
287 err = __add_section(zone, i << PFN_SECTION_SHIFT);
290 * EEXIST is finally dealt with by ioresource collision
291 * check. see add_memory() => register_memory_resource()
292 * Warning will be printed if there is collision.
294 if (err && (err != -EEXIST))
301 EXPORT_SYMBOL_GPL(__add_pages);
304 * __remove_pages() - remove sections of pages from a zone
305 * @zone: zone from which pages need to be removed
306 * @phys_start_pfn: starting pageframe (must be aligned to start of a section)
307 * @nr_pages: number of pages to remove (must be multiple of section size)
309 * Generic helper function to remove section mappings and sysfs entries
310 * for the section of the memory we are removing. Caller needs to make
311 * sure that pages are marked reserved and zones are adjust properly by
312 * calling offline_pages().
314 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
315 unsigned long nr_pages)
317 unsigned long i, ret = 0;
318 int sections_to_remove;
321 * We can only remove entire sections
323 BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
324 BUG_ON(nr_pages % PAGES_PER_SECTION);
326 release_mem_region(phys_start_pfn << PAGE_SHIFT, nr_pages * PAGE_SIZE);
328 sections_to_remove = nr_pages / PAGES_PER_SECTION;
329 for (i = 0; i < sections_to_remove; i++) {
330 unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION;
331 ret = __remove_section(zone, __pfn_to_section(pfn));
337 EXPORT_SYMBOL_GPL(__remove_pages);
339 void online_page(struct page *page)
344 #ifdef CONFIG_HIGHMEM
345 if (PageHighMem(page))
349 #ifdef CONFIG_FLATMEM
350 max_mapnr = max(page_to_pfn(page), max_mapnr);
353 ClearPageReserved(page);
354 init_page_count(page);
358 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
362 unsigned long onlined_pages = *(unsigned long *)arg;
364 if (PageReserved(pfn_to_page(start_pfn)))
365 for (i = 0; i < nr_pages; i++) {
366 page = pfn_to_page(start_pfn + i);
370 *(unsigned long *)arg = onlined_pages;
375 int online_pages(unsigned long pfn, unsigned long nr_pages)
377 unsigned long onlined_pages = 0;
379 int need_zonelists_rebuild = 0;
382 struct memory_notify arg;
385 arg.nr_pages = nr_pages;
386 arg.status_change_nid = -1;
388 nid = page_to_nid(pfn_to_page(pfn));
389 if (node_present_pages(nid) == 0)
390 arg.status_change_nid = nid;
392 ret = memory_notify(MEM_GOING_ONLINE, &arg);
393 ret = notifier_to_errno(ret);
395 memory_notify(MEM_CANCEL_ONLINE, &arg);
399 * This doesn't need a lock to do pfn_to_page().
400 * The section can't be removed here because of the
401 * memory_block->state_mutex.
403 zone = page_zone(pfn_to_page(pfn));
405 * If this zone is not populated, then it is not in zonelist.
406 * This means the page allocator ignores this zone.
407 * So, zonelist must be updated after online.
409 if (!populated_zone(zone))
410 need_zonelists_rebuild = 1;
412 ret = walk_memory_resource(pfn, nr_pages, &onlined_pages,
415 printk(KERN_DEBUG "online_pages %lx at %lx failed\n",
417 memory_notify(MEM_CANCEL_ONLINE, &arg);
421 zone->present_pages += onlined_pages;
422 zone->zone_pgdat->node_present_pages += onlined_pages;
424 setup_per_zone_pages_min();
426 kswapd_run(zone_to_nid(zone));
427 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
430 if (need_zonelists_rebuild)
431 build_all_zonelists();
433 vm_total_pages = nr_free_pagecache_pages();
435 writeback_set_ratelimit();
438 memory_notify(MEM_ONLINE, &arg);
442 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
444 static pg_data_t *hotadd_new_pgdat(int nid, u64 start)
446 struct pglist_data *pgdat;
447 unsigned long zones_size[MAX_NR_ZONES] = {0};
448 unsigned long zholes_size[MAX_NR_ZONES] = {0};
449 unsigned long start_pfn = start >> PAGE_SHIFT;
451 pgdat = arch_alloc_nodedata(nid);
455 arch_refresh_nodedata(nid, pgdat);
457 /* we can use NODE_DATA(nid) from here */
459 /* init node's zones as empty zones, we don't have any present pages.*/
460 free_area_init_node(nid, zones_size, start_pfn, zholes_size);
465 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
467 arch_refresh_nodedata(nid, NULL);
468 arch_free_nodedata(pgdat);
473 int add_memory(int nid, u64 start, u64 size)
475 pg_data_t *pgdat = NULL;
477 struct resource *res;
480 res = register_memory_resource(start, size);
484 if (!node_online(nid)) {
485 pgdat = hotadd_new_pgdat(nid, start);
491 /* call arch's memory hotadd */
492 ret = arch_add_memory(nid, start, size);
497 /* we online node here. we can't roll back from here. */
498 node_set_online(nid);
500 cpuset_track_online_nodes();
503 ret = register_one_node(nid);
505 * If sysfs file of new node can't create, cpu on the node
506 * can't be hot-added. There is no rollback way now.
507 * So, check by BUG_ON() to catch it reluctantly..
514 /* rollback pgdat allocation and others */
516 rollback_node_hotadd(nid, pgdat);
518 release_memory_resource(res);
522 EXPORT_SYMBOL_GPL(add_memory);
524 #ifdef CONFIG_MEMORY_HOTREMOVE
526 * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
527 * set and the size of the free page is given by page_order(). Using this,
528 * the function determines if the pageblock contains only free pages.
529 * Due to buddy contraints, a free page at least the size of a pageblock will
530 * be located at the start of the pageblock
532 static inline int pageblock_free(struct page *page)
534 return PageBuddy(page) && page_order(page) >= pageblock_order;
537 /* Return the start of the next active pageblock after a given page */
538 static struct page *next_active_pageblock(struct page *page)
540 int pageblocks_stride;
542 /* Ensure the starting page is pageblock-aligned */
543 BUG_ON(page_to_pfn(page) & (pageblock_nr_pages - 1));
545 /* Move forward by at least 1 * pageblock_nr_pages */
546 pageblocks_stride = 1;
548 /* If the entire pageblock is free, move to the end of free page */
549 if (pageblock_free(page))
550 pageblocks_stride += page_order(page) - pageblock_order;
552 return page + (pageblocks_stride * pageblock_nr_pages);
555 /* Checks if this range of memory is likely to be hot-removable. */
556 int is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
559 struct page *page = pfn_to_page(start_pfn);
560 struct page *end_page = page + nr_pages;
562 /* Check the starting page of each pageblock within the range */
563 for (; page < end_page; page = next_active_pageblock(page)) {
564 type = get_pageblock_migratetype(page);
567 * A pageblock containing MOVABLE or free pages is considered
570 if (type != MIGRATE_MOVABLE && !pageblock_free(page))
574 * A pageblock starting with a PageReserved page is not
575 * considered removable.
577 if (PageReserved(page))
581 /* All pageblocks in the memory block are likely to be hot-removable */
586 * Confirm all pages in a range [start, end) is belongs to the same zone.
588 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
591 struct zone *zone = NULL;
594 for (pfn = start_pfn;
596 pfn += MAX_ORDER_NR_PAGES) {
598 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
599 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
601 if (i == MAX_ORDER_NR_PAGES)
603 page = pfn_to_page(pfn + i);
604 if (zone && page_zone(page) != zone)
606 zone = page_zone(page);
612 * Scanning pfn is much easier than scanning lru list.
613 * Scan pfn from start to end and Find LRU page.
615 int scan_lru_pages(unsigned long start, unsigned long end)
619 for (pfn = start; pfn < end; pfn++) {
620 if (pfn_valid(pfn)) {
621 page = pfn_to_page(pfn);
630 hotremove_migrate_alloc(struct page *page,
631 unsigned long private,
634 /* This should be improoooooved!! */
635 return alloc_page(GFP_HIGHUSER_PAGECACHE);
639 #define NR_OFFLINE_AT_ONCE_PAGES (256)
641 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
645 int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
650 for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
653 page = pfn_to_page(pfn);
654 if (!page_count(page))
657 * We can skip free pages. And we can only deal with pages on
660 ret = isolate_lru_page(page, &source);
661 if (!ret) { /* Success */
664 /* Becasue we don't have big zone->lock. we should
665 check this again here. */
666 if (page_count(page))
668 #ifdef CONFIG_DEBUG_VM
669 printk(KERN_INFO "removing from LRU failed"
671 pfn, page_count(page), page->flags);
677 if (!list_empty(&source))
678 putback_lru_pages(&source);
682 if (list_empty(&source))
684 /* this function returns # of failed pages */
685 ret = migrate_pages(&source, hotremove_migrate_alloc, 0);
692 * remove from free_area[] and mark all as Reserved.
695 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
698 __offline_isolated_pages(start, start + nr_pages);
703 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
705 walk_memory_resource(start_pfn, end_pfn - start_pfn, NULL,
706 offline_isolated_pages_cb);
710 * Check all pages in range, recoreded as memory resource, are isolated.
713 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
717 long offlined = *(long *)data;
718 ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
721 *(long *)data += offlined;
726 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
731 ret = walk_memory_resource(start_pfn, end_pfn - start_pfn, &offlined,
732 check_pages_isolated_cb);
734 offlined = (long)ret;
738 int offline_pages(unsigned long start_pfn,
739 unsigned long end_pfn, unsigned long timeout)
741 unsigned long pfn, nr_pages, expire;
743 int ret, drain, retry_max, node;
745 struct memory_notify arg;
747 BUG_ON(start_pfn >= end_pfn);
748 /* at least, alignment against pageblock is necessary */
749 if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
751 if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
753 /* This makes hotplug much easier...and readable.
754 we assume this for now. .*/
755 if (!test_pages_in_a_zone(start_pfn, end_pfn))
758 zone = page_zone(pfn_to_page(start_pfn));
759 node = zone_to_nid(zone);
760 nr_pages = end_pfn - start_pfn;
762 /* set above range as isolated */
763 ret = start_isolate_page_range(start_pfn, end_pfn);
767 arg.start_pfn = start_pfn;
768 arg.nr_pages = nr_pages;
769 arg.status_change_nid = -1;
770 if (nr_pages >= node_present_pages(node))
771 arg.status_change_nid = node;
773 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
774 ret = notifier_to_errno(ret);
779 expire = jiffies + timeout;
783 /* start memory hot removal */
785 if (time_after(jiffies, expire))
788 if (signal_pending(current))
793 flush_scheduled_work();
798 pfn = scan_lru_pages(start_pfn, end_pfn);
799 if (pfn) { /* We have page on LRU */
800 ret = do_migrate_range(pfn, end_pfn);
806 if (--retry_max == 0)
813 /* drain all zone's lru pagevec, this is asyncronous... */
815 flush_scheduled_work();
817 /* drain pcp pages , this is synchrouns. */
820 offlined_pages = check_pages_isolated(start_pfn, end_pfn);
821 if (offlined_pages < 0) {
825 printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
826 /* Ok, all of our target is islaoted.
827 We cannot do rollback at this point. */
828 offline_isolated_pages(start_pfn, end_pfn);
829 /* reset pagetype flags and makes migrate type to be MOVABLE */
830 undo_isolate_page_range(start_pfn, end_pfn);
831 /* removal success */
832 zone->present_pages -= offlined_pages;
833 zone->zone_pgdat->node_present_pages -= offlined_pages;
834 totalram_pages -= offlined_pages;
835 num_physpages -= offlined_pages;
837 vm_total_pages = nr_free_pagecache_pages();
838 writeback_set_ratelimit();
840 memory_notify(MEM_OFFLINE, &arg);
844 printk(KERN_INFO "memory offlining %lx to %lx failed\n",
846 memory_notify(MEM_CANCEL_OFFLINE, &arg);
847 /* pushback to free area */
848 undo_isolate_page_range(start_pfn, end_pfn);
853 int remove_memory(u64 start, u64 size)
857 EXPORT_SYMBOL_GPL(remove_memory);
858 #endif /* CONFIG_MEMORY_HOTREMOVE */