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>
32 /* add this memory to iomem resource */
33 static struct resource *register_memory_resource(u64 start, u64 size)
36 res = kzalloc(sizeof(struct resource), GFP_KERNEL);
39 res->name = "System RAM";
41 res->end = start + size - 1;
42 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
43 if (request_resource(&iomem_resource, res) < 0) {
44 printk("System RAM resource %llx - %llx cannot be added\n",
45 (unsigned long long)res->start, (unsigned long long)res->end);
52 static void release_memory_resource(struct resource *res)
56 release_resource(res);
61 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
62 #ifndef CONFIG_SPARSEMEM_VMEMMAP
63 static void get_page_bootmem(unsigned long info, struct page *page, int magic)
65 atomic_set(&page->_mapcount, magic);
67 set_page_private(page, info);
68 atomic_inc(&page->_count);
71 void put_page_bootmem(struct page *page)
75 magic = atomic_read(&page->_mapcount);
78 if (atomic_dec_return(&page->_count) == 1) {
79 ClearPagePrivate(page);
80 set_page_private(page, 0);
81 reset_page_mapcount(page);
82 __free_pages_bootmem(page, 0);
87 void register_page_bootmem_info_section(unsigned long start_pfn)
89 unsigned long *usemap, mapsize, section_nr, i;
90 struct mem_section *ms;
91 struct page *page, *memmap;
93 if (!pfn_valid(start_pfn))
96 section_nr = pfn_to_section_nr(start_pfn);
97 ms = __nr_to_section(section_nr);
99 /* Get section's memmap address */
100 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
103 * Get page for the memmap's phys address
104 * XXX: need more consideration for sparse_vmemmap...
106 page = virt_to_page(memmap);
107 mapsize = sizeof(struct page) * PAGES_PER_SECTION;
108 mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
110 /* remember memmap's page */
111 for (i = 0; i < mapsize; i++, page++)
112 get_page_bootmem(section_nr, page, SECTION_INFO);
114 usemap = __nr_to_section(section_nr)->pageblock_flags;
115 page = virt_to_page(usemap);
117 mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT;
119 for (i = 0; i < mapsize; i++, page++)
120 get_page_bootmem(section_nr, page, MIX_INFO);
124 void register_page_bootmem_info_node(struct pglist_data *pgdat)
126 unsigned long i, pfn, end_pfn, nr_pages;
127 int node = pgdat->node_id;
131 nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
132 page = virt_to_page(pgdat);
134 for (i = 0; i < nr_pages; i++, page++)
135 get_page_bootmem(node, page, NODE_INFO);
137 zone = &pgdat->node_zones[0];
138 for (; zone < pgdat->node_zones + MAX_NR_ZONES - 1; zone++) {
139 if (zone->wait_table) {
140 nr_pages = zone->wait_table_hash_nr_entries
141 * sizeof(wait_queue_head_t);
142 nr_pages = PAGE_ALIGN(nr_pages) >> PAGE_SHIFT;
143 page = virt_to_page(zone->wait_table);
145 for (i = 0; i < nr_pages; i++, page++)
146 get_page_bootmem(node, page, NODE_INFO);
150 pfn = pgdat->node_start_pfn;
151 end_pfn = pfn + pgdat->node_spanned_pages;
153 /* register_section info */
154 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION)
155 register_page_bootmem_info_section(pfn);
158 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
160 static int __add_zone(struct zone *zone, unsigned long phys_start_pfn)
162 struct pglist_data *pgdat = zone->zone_pgdat;
163 int nr_pages = PAGES_PER_SECTION;
164 int nid = pgdat->node_id;
167 zone_type = zone - pgdat->node_zones;
168 if (!zone->wait_table) {
170 ret = init_currently_empty_zone(zone, phys_start_pfn,
171 nr_pages, MEMMAP_HOTPLUG);
175 memmap_init_zone(nr_pages, nid, zone_type,
176 phys_start_pfn, MEMMAP_HOTPLUG);
180 static int __add_section(struct zone *zone, unsigned long phys_start_pfn)
182 int nr_pages = PAGES_PER_SECTION;
185 if (pfn_valid(phys_start_pfn))
188 ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
193 ret = __add_zone(zone, phys_start_pfn);
198 return register_new_memory(__pfn_to_section(phys_start_pfn));
201 static int __remove_section(struct zone *zone, struct mem_section *ms)
204 struct pglist_data *pgdat = zone->zone_pgdat;
207 if (!valid_section(ms))
210 ret = unregister_memory_section(ms);
214 pgdat_resize_lock(pgdat, &flags);
215 sparse_remove_one_section(zone, ms);
216 pgdat_resize_unlock(pgdat, &flags);
221 * Reasonably generic function for adding memory. It is
222 * expected that archs that support memory hotplug will
223 * call this function after deciding the zone to which to
226 int __add_pages(struct zone *zone, unsigned long phys_start_pfn,
227 unsigned long nr_pages)
231 int start_sec, end_sec;
232 /* during initialize mem_map, align hot-added range to section */
233 start_sec = pfn_to_section_nr(phys_start_pfn);
234 end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
236 for (i = start_sec; i <= end_sec; i++) {
237 err = __add_section(zone, i << PFN_SECTION_SHIFT);
240 * EEXIST is finally dealt with by ioresource collision
241 * check. see add_memory() => register_memory_resource()
242 * Warning will be printed if there is collision.
244 if (err && (err != -EEXIST))
251 EXPORT_SYMBOL_GPL(__add_pages);
254 * __remove_pages() - remove sections of pages from a zone
255 * @zone: zone from which pages need to be removed
256 * @phys_start_pfn: starting pageframe (must be aligned to start of a section)
257 * @nr_pages: number of pages to remove (must be multiple of section size)
259 * Generic helper function to remove section mappings and sysfs entries
260 * for the section of the memory we are removing. Caller needs to make
261 * sure that pages are marked reserved and zones are adjust properly by
262 * calling offline_pages().
264 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
265 unsigned long nr_pages)
267 unsigned long i, ret = 0;
268 int sections_to_remove;
271 * We can only remove entire sections
273 BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
274 BUG_ON(nr_pages % PAGES_PER_SECTION);
276 release_mem_region(phys_start_pfn << PAGE_SHIFT, nr_pages * PAGE_SIZE);
278 sections_to_remove = nr_pages / PAGES_PER_SECTION;
279 for (i = 0; i < sections_to_remove; i++) {
280 unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION;
281 ret = __remove_section(zone, __pfn_to_section(pfn));
287 EXPORT_SYMBOL_GPL(__remove_pages);
289 static void grow_zone_span(struct zone *zone,
290 unsigned long start_pfn, unsigned long end_pfn)
292 unsigned long old_zone_end_pfn;
294 zone_span_writelock(zone);
296 old_zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
297 if (start_pfn < zone->zone_start_pfn)
298 zone->zone_start_pfn = start_pfn;
300 zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
301 zone->zone_start_pfn;
303 zone_span_writeunlock(zone);
306 static void grow_pgdat_span(struct pglist_data *pgdat,
307 unsigned long start_pfn, unsigned long end_pfn)
309 unsigned long old_pgdat_end_pfn =
310 pgdat->node_start_pfn + pgdat->node_spanned_pages;
312 if (start_pfn < pgdat->node_start_pfn)
313 pgdat->node_start_pfn = start_pfn;
315 pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
316 pgdat->node_start_pfn;
319 void online_page(struct page *page)
324 #ifdef CONFIG_HIGHMEM
325 if (PageHighMem(page))
329 #ifdef CONFIG_FLATMEM
330 max_mapnr = max(page_to_pfn(page), max_mapnr);
333 ClearPageReserved(page);
334 init_page_count(page);
338 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
342 unsigned long onlined_pages = *(unsigned long *)arg;
344 if (PageReserved(pfn_to_page(start_pfn)))
345 for (i = 0; i < nr_pages; i++) {
346 page = pfn_to_page(start_pfn + i);
350 *(unsigned long *)arg = onlined_pages;
355 int online_pages(unsigned long pfn, unsigned long nr_pages)
358 unsigned long onlined_pages = 0;
360 int need_zonelists_rebuild = 0;
363 struct memory_notify arg;
366 arg.nr_pages = nr_pages;
367 arg.status_change_nid = -1;
369 nid = page_to_nid(pfn_to_page(pfn));
370 if (node_present_pages(nid) == 0)
371 arg.status_change_nid = nid;
373 ret = memory_notify(MEM_GOING_ONLINE, &arg);
374 ret = notifier_to_errno(ret);
376 memory_notify(MEM_CANCEL_ONLINE, &arg);
380 * This doesn't need a lock to do pfn_to_page().
381 * The section can't be removed here because of the
382 * memory_block->state_mutex.
384 zone = page_zone(pfn_to_page(pfn));
385 pgdat_resize_lock(zone->zone_pgdat, &flags);
386 grow_zone_span(zone, pfn, pfn + nr_pages);
387 grow_pgdat_span(zone->zone_pgdat, pfn, pfn + nr_pages);
388 pgdat_resize_unlock(zone->zone_pgdat, &flags);
391 * If this zone is not populated, then it is not in zonelist.
392 * This means the page allocator ignores this zone.
393 * So, zonelist must be updated after online.
395 if (!populated_zone(zone))
396 need_zonelists_rebuild = 1;
398 walk_memory_resource(pfn, nr_pages, &onlined_pages,
400 zone->present_pages += onlined_pages;
401 zone->zone_pgdat->node_present_pages += onlined_pages;
403 setup_per_zone_pages_min();
405 kswapd_run(zone_to_nid(zone));
406 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
409 if (need_zonelists_rebuild)
410 build_all_zonelists();
411 vm_total_pages = nr_free_pagecache_pages();
412 writeback_set_ratelimit();
415 memory_notify(MEM_ONLINE, &arg);
419 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
421 static pg_data_t *hotadd_new_pgdat(int nid, u64 start)
423 struct pglist_data *pgdat;
424 unsigned long zones_size[MAX_NR_ZONES] = {0};
425 unsigned long zholes_size[MAX_NR_ZONES] = {0};
426 unsigned long start_pfn = start >> PAGE_SHIFT;
428 pgdat = arch_alloc_nodedata(nid);
432 arch_refresh_nodedata(nid, pgdat);
434 /* we can use NODE_DATA(nid) from here */
436 /* init node's zones as empty zones, we don't have any present pages.*/
437 free_area_init_node(nid, pgdat, zones_size, start_pfn, zholes_size);
442 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
444 arch_refresh_nodedata(nid, NULL);
445 arch_free_nodedata(pgdat);
450 int add_memory(int nid, u64 start, u64 size)
452 pg_data_t *pgdat = NULL;
454 struct resource *res;
457 res = register_memory_resource(start, size);
461 if (!node_online(nid)) {
462 pgdat = hotadd_new_pgdat(nid, start);
468 /* call arch's memory hotadd */
469 ret = arch_add_memory(nid, start, size);
474 /* we online node here. we can't roll back from here. */
475 node_set_online(nid);
477 cpuset_track_online_nodes();
480 ret = register_one_node(nid);
482 * If sysfs file of new node can't create, cpu on the node
483 * can't be hot-added. There is no rollback way now.
484 * So, check by BUG_ON() to catch it reluctantly..
491 /* rollback pgdat allocation and others */
493 rollback_node_hotadd(nid, pgdat);
495 release_memory_resource(res);
499 EXPORT_SYMBOL_GPL(add_memory);
501 #ifdef CONFIG_MEMORY_HOTREMOVE
503 * Confirm all pages in a range [start, end) is belongs to the same zone.
505 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
508 struct zone *zone = NULL;
511 for (pfn = start_pfn;
513 pfn += MAX_ORDER_NR_PAGES) {
515 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
516 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
518 if (i == MAX_ORDER_NR_PAGES)
520 page = pfn_to_page(pfn + i);
521 if (zone && page_zone(page) != zone)
523 zone = page_zone(page);
529 * Scanning pfn is much easier than scanning lru list.
530 * Scan pfn from start to end and Find LRU page.
532 int scan_lru_pages(unsigned long start, unsigned long end)
536 for (pfn = start; pfn < end; pfn++) {
537 if (pfn_valid(pfn)) {
538 page = pfn_to_page(pfn);
547 hotremove_migrate_alloc(struct page *page,
548 unsigned long private,
551 /* This should be improoooooved!! */
552 return alloc_page(GFP_HIGHUSER_PAGECACHE);
556 #define NR_OFFLINE_AT_ONCE_PAGES (256)
558 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
562 int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
567 for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
570 page = pfn_to_page(pfn);
571 if (!page_count(page))
574 * We can skip free pages. And we can only deal with pages on
577 ret = isolate_lru_page(page, &source);
578 if (!ret) { /* Success */
581 /* Becasue we don't have big zone->lock. we should
582 check this again here. */
583 if (page_count(page))
585 #ifdef CONFIG_DEBUG_VM
586 printk(KERN_INFO "removing from LRU failed"
588 pfn, page_count(page), page->flags);
594 if (!list_empty(&source))
595 putback_lru_pages(&source);
599 if (list_empty(&source))
601 /* this function returns # of failed pages */
602 ret = migrate_pages(&source, hotremove_migrate_alloc, 0);
609 * remove from free_area[] and mark all as Reserved.
612 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
615 __offline_isolated_pages(start, start + nr_pages);
620 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
622 walk_memory_resource(start_pfn, end_pfn - start_pfn, NULL,
623 offline_isolated_pages_cb);
627 * Check all pages in range, recoreded as memory resource, are isolated.
630 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
634 long offlined = *(long *)data;
635 ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
638 *(long *)data += offlined;
643 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
648 ret = walk_memory_resource(start_pfn, end_pfn - start_pfn, &offlined,
649 check_pages_isolated_cb);
651 offlined = (long)ret;
655 int offline_pages(unsigned long start_pfn,
656 unsigned long end_pfn, unsigned long timeout)
658 unsigned long pfn, nr_pages, expire;
660 int ret, drain, retry_max, node;
662 struct memory_notify arg;
664 BUG_ON(start_pfn >= end_pfn);
665 /* at least, alignment against pageblock is necessary */
666 if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
668 if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
670 /* This makes hotplug much easier...and readable.
671 we assume this for now. .*/
672 if (!test_pages_in_a_zone(start_pfn, end_pfn))
675 zone = page_zone(pfn_to_page(start_pfn));
676 node = zone_to_nid(zone);
677 nr_pages = end_pfn - start_pfn;
679 /* set above range as isolated */
680 ret = start_isolate_page_range(start_pfn, end_pfn);
684 arg.start_pfn = start_pfn;
685 arg.nr_pages = nr_pages;
686 arg.status_change_nid = -1;
687 if (nr_pages >= node_present_pages(node))
688 arg.status_change_nid = node;
690 ret = memory_notify(MEM_GOING_OFFLINE, &arg);
691 ret = notifier_to_errno(ret);
696 expire = jiffies + timeout;
700 /* start memory hot removal */
702 if (time_after(jiffies, expire))
705 if (signal_pending(current))
710 flush_scheduled_work();
715 pfn = scan_lru_pages(start_pfn, end_pfn);
716 if (pfn) { /* We have page on LRU */
717 ret = do_migrate_range(pfn, end_pfn);
723 if (--retry_max == 0)
730 /* drain all zone's lru pagevec, this is asyncronous... */
732 flush_scheduled_work();
734 /* drain pcp pages , this is synchrouns. */
737 offlined_pages = check_pages_isolated(start_pfn, end_pfn);
738 if (offlined_pages < 0) {
742 printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
743 /* Ok, all of our target is islaoted.
744 We cannot do rollback at this point. */
745 offline_isolated_pages(start_pfn, end_pfn);
746 /* reset pagetype flags and makes migrate type to be MOVABLE */
747 undo_isolate_page_range(start_pfn, end_pfn);
748 /* removal success */
749 zone->present_pages -= offlined_pages;
750 zone->zone_pgdat->node_present_pages -= offlined_pages;
751 totalram_pages -= offlined_pages;
752 num_physpages -= offlined_pages;
754 vm_total_pages = nr_free_pagecache_pages();
755 writeback_set_ratelimit();
757 memory_notify(MEM_OFFLINE, &arg);
761 printk(KERN_INFO "memory offlining %lx to %lx failed\n",
763 memory_notify(MEM_CANCEL_OFFLINE, &arg);
764 /* pushback to free area */
765 undo_isolate_page_range(start_pfn, end_pfn);
770 int remove_memory(u64 start, u64 size)
774 EXPORT_SYMBOL_GPL(remove_memory);
775 #endif /* CONFIG_MEMORY_HOTREMOVE */