xref: /openbmc/linux/mm/memory_hotplug.c (revision 81d67439)
1 /*
2  *  linux/mm/memory_hotplug.c
3  *
4  *  Copyright (C)
5  */
6 
7 #include <linux/stddef.h>
8 #include <linux/mm.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>
32 
33 #include <asm/tlbflush.h>
34 
35 #include "internal.h"
36 
37 DEFINE_MUTEX(mem_hotplug_mutex);
38 
39 void lock_memory_hotplug(void)
40 {
41 	mutex_lock(&mem_hotplug_mutex);
42 
43 	/* for exclusive hibernation if CONFIG_HIBERNATION=y */
44 	lock_system_sleep();
45 }
46 
47 void unlock_memory_hotplug(void)
48 {
49 	unlock_system_sleep();
50 	mutex_unlock(&mem_hotplug_mutex);
51 }
52 
53 
54 /* add this memory to iomem resource */
55 static struct resource *register_memory_resource(u64 start, u64 size)
56 {
57 	struct resource *res;
58 	res = kzalloc(sizeof(struct resource), GFP_KERNEL);
59 	BUG_ON(!res);
60 
61 	res->name = "System RAM";
62 	res->start = start;
63 	res->end = start + size - 1;
64 	res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
65 	if (request_resource(&iomem_resource, res) < 0) {
66 		printk("System RAM resource %llx - %llx cannot be added\n",
67 		(unsigned long long)res->start, (unsigned long long)res->end);
68 		kfree(res);
69 		res = NULL;
70 	}
71 	return res;
72 }
73 
74 static void release_memory_resource(struct resource *res)
75 {
76 	if (!res)
77 		return;
78 	release_resource(res);
79 	kfree(res);
80 	return;
81 }
82 
83 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
84 #ifndef CONFIG_SPARSEMEM_VMEMMAP
85 static void get_page_bootmem(unsigned long info,  struct page *page,
86 			     unsigned long type)
87 {
88 	page->lru.next = (struct list_head *) type;
89 	SetPagePrivate(page);
90 	set_page_private(page, info);
91 	atomic_inc(&page->_count);
92 }
93 
94 /* reference to __meminit __free_pages_bootmem is valid
95  * so use __ref to tell modpost not to generate a warning */
96 void __ref put_page_bootmem(struct page *page)
97 {
98 	unsigned long type;
99 
100 	type = (unsigned long) page->lru.next;
101 	BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE ||
102 	       type > MEMORY_HOTPLUG_MAX_BOOTMEM_TYPE);
103 
104 	if (atomic_dec_return(&page->_count) == 1) {
105 		ClearPagePrivate(page);
106 		set_page_private(page, 0);
107 		INIT_LIST_HEAD(&page->lru);
108 		__free_pages_bootmem(page, 0);
109 	}
110 
111 }
112 
113 static void register_page_bootmem_info_section(unsigned long start_pfn)
114 {
115 	unsigned long *usemap, mapsize, section_nr, i;
116 	struct mem_section *ms;
117 	struct page *page, *memmap;
118 
119 	if (!pfn_valid(start_pfn))
120 		return;
121 
122 	section_nr = pfn_to_section_nr(start_pfn);
123 	ms = __nr_to_section(section_nr);
124 
125 	/* Get section's memmap address */
126 	memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
127 
128 	/*
129 	 * Get page for the memmap's phys address
130 	 * XXX: need more consideration for sparse_vmemmap...
131 	 */
132 	page = virt_to_page(memmap);
133 	mapsize = sizeof(struct page) * PAGES_PER_SECTION;
134 	mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
135 
136 	/* remember memmap's page */
137 	for (i = 0; i < mapsize; i++, page++)
138 		get_page_bootmem(section_nr, page, SECTION_INFO);
139 
140 	usemap = __nr_to_section(section_nr)->pageblock_flags;
141 	page = virt_to_page(usemap);
142 
143 	mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT;
144 
145 	for (i = 0; i < mapsize; i++, page++)
146 		get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
147 
148 }
149 
150 void register_page_bootmem_info_node(struct pglist_data *pgdat)
151 {
152 	unsigned long i, pfn, end_pfn, nr_pages;
153 	int node = pgdat->node_id;
154 	struct page *page;
155 	struct zone *zone;
156 
157 	nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
158 	page = virt_to_page(pgdat);
159 
160 	for (i = 0; i < nr_pages; i++, page++)
161 		get_page_bootmem(node, page, NODE_INFO);
162 
163 	zone = &pgdat->node_zones[0];
164 	for (; zone < pgdat->node_zones + MAX_NR_ZONES - 1; zone++) {
165 		if (zone->wait_table) {
166 			nr_pages = zone->wait_table_hash_nr_entries
167 				* sizeof(wait_queue_head_t);
168 			nr_pages = PAGE_ALIGN(nr_pages) >> PAGE_SHIFT;
169 			page = virt_to_page(zone->wait_table);
170 
171 			for (i = 0; i < nr_pages; i++, page++)
172 				get_page_bootmem(node, page, NODE_INFO);
173 		}
174 	}
175 
176 	pfn = pgdat->node_start_pfn;
177 	end_pfn = pfn + pgdat->node_spanned_pages;
178 
179 	/* register_section info */
180 	for (; pfn < end_pfn; pfn += PAGES_PER_SECTION)
181 		register_page_bootmem_info_section(pfn);
182 
183 }
184 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
185 
186 static void grow_zone_span(struct zone *zone, unsigned long start_pfn,
187 			   unsigned long end_pfn)
188 {
189 	unsigned long old_zone_end_pfn;
190 
191 	zone_span_writelock(zone);
192 
193 	old_zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
194 	if (start_pfn < zone->zone_start_pfn)
195 		zone->zone_start_pfn = start_pfn;
196 
197 	zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
198 				zone->zone_start_pfn;
199 
200 	zone_span_writeunlock(zone);
201 }
202 
203 static void grow_pgdat_span(struct pglist_data *pgdat, unsigned long start_pfn,
204 			    unsigned long end_pfn)
205 {
206 	unsigned long old_pgdat_end_pfn =
207 		pgdat->node_start_pfn + pgdat->node_spanned_pages;
208 
209 	if (start_pfn < pgdat->node_start_pfn)
210 		pgdat->node_start_pfn = start_pfn;
211 
212 	pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
213 					pgdat->node_start_pfn;
214 }
215 
216 static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
217 {
218 	struct pglist_data *pgdat = zone->zone_pgdat;
219 	int nr_pages = PAGES_PER_SECTION;
220 	int nid = pgdat->node_id;
221 	int zone_type;
222 	unsigned long flags;
223 
224 	zone_type = zone - pgdat->node_zones;
225 	if (!zone->wait_table) {
226 		int ret;
227 
228 		ret = init_currently_empty_zone(zone, phys_start_pfn,
229 						nr_pages, MEMMAP_HOTPLUG);
230 		if (ret)
231 			return ret;
232 	}
233 	pgdat_resize_lock(zone->zone_pgdat, &flags);
234 	grow_zone_span(zone, phys_start_pfn, phys_start_pfn + nr_pages);
235 	grow_pgdat_span(zone->zone_pgdat, phys_start_pfn,
236 			phys_start_pfn + nr_pages);
237 	pgdat_resize_unlock(zone->zone_pgdat, &flags);
238 	memmap_init_zone(nr_pages, nid, zone_type,
239 			 phys_start_pfn, MEMMAP_HOTPLUG);
240 	return 0;
241 }
242 
243 static int __meminit __add_section(int nid, struct zone *zone,
244 					unsigned long phys_start_pfn)
245 {
246 	int nr_pages = PAGES_PER_SECTION;
247 	int ret;
248 
249 	if (pfn_valid(phys_start_pfn))
250 		return -EEXIST;
251 
252 	ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
253 
254 	if (ret < 0)
255 		return ret;
256 
257 	ret = __add_zone(zone, phys_start_pfn);
258 
259 	if (ret < 0)
260 		return ret;
261 
262 	return register_new_memory(nid, __pfn_to_section(phys_start_pfn));
263 }
264 
265 #ifdef CONFIG_SPARSEMEM_VMEMMAP
266 static int __remove_section(struct zone *zone, struct mem_section *ms)
267 {
268 	/*
269 	 * XXX: Freeing memmap with vmemmap is not implement yet.
270 	 *      This should be removed later.
271 	 */
272 	return -EBUSY;
273 }
274 #else
275 static int __remove_section(struct zone *zone, struct mem_section *ms)
276 {
277 	unsigned long flags;
278 	struct pglist_data *pgdat = zone->zone_pgdat;
279 	int ret = -EINVAL;
280 
281 	if (!valid_section(ms))
282 		return ret;
283 
284 	ret = unregister_memory_section(ms);
285 	if (ret)
286 		return ret;
287 
288 	pgdat_resize_lock(pgdat, &flags);
289 	sparse_remove_one_section(zone, ms);
290 	pgdat_resize_unlock(pgdat, &flags);
291 	return 0;
292 }
293 #endif
294 
295 /*
296  * Reasonably generic function for adding memory.  It is
297  * expected that archs that support memory hotplug will
298  * call this function after deciding the zone to which to
299  * add the new pages.
300  */
301 int __ref __add_pages(int nid, struct zone *zone, unsigned long phys_start_pfn,
302 			unsigned long nr_pages)
303 {
304 	unsigned long i;
305 	int err = 0;
306 	int start_sec, end_sec;
307 	/* during initialize mem_map, align hot-added range to section */
308 	start_sec = pfn_to_section_nr(phys_start_pfn);
309 	end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
310 
311 	for (i = start_sec; i <= end_sec; i++) {
312 		err = __add_section(nid, zone, i << PFN_SECTION_SHIFT);
313 
314 		/*
315 		 * EEXIST is finally dealt with by ioresource collision
316 		 * check. see add_memory() => register_memory_resource()
317 		 * Warning will be printed if there is collision.
318 		 */
319 		if (err && (err != -EEXIST))
320 			break;
321 		err = 0;
322 	}
323 
324 	return err;
325 }
326 EXPORT_SYMBOL_GPL(__add_pages);
327 
328 /**
329  * __remove_pages() - remove sections of pages from a zone
330  * @zone: zone from which pages need to be removed
331  * @phys_start_pfn: starting pageframe (must be aligned to start of a section)
332  * @nr_pages: number of pages to remove (must be multiple of section size)
333  *
334  * Generic helper function to remove section mappings and sysfs entries
335  * for the section of the memory we are removing. Caller needs to make
336  * sure that pages are marked reserved and zones are adjust properly by
337  * calling offline_pages().
338  */
339 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
340 		 unsigned long nr_pages)
341 {
342 	unsigned long i, ret = 0;
343 	int sections_to_remove;
344 
345 	/*
346 	 * We can only remove entire sections
347 	 */
348 	BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
349 	BUG_ON(nr_pages % PAGES_PER_SECTION);
350 
351 	sections_to_remove = nr_pages / PAGES_PER_SECTION;
352 	for (i = 0; i < sections_to_remove; i++) {
353 		unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION;
354 		release_mem_region(pfn << PAGE_SHIFT,
355 				   PAGES_PER_SECTION << PAGE_SHIFT);
356 		ret = __remove_section(zone, __pfn_to_section(pfn));
357 		if (ret)
358 			break;
359 	}
360 	return ret;
361 }
362 EXPORT_SYMBOL_GPL(__remove_pages);
363 
364 void online_page(struct page *page)
365 {
366 	unsigned long pfn = page_to_pfn(page);
367 
368 	totalram_pages++;
369 	if (pfn >= num_physpages)
370 		num_physpages = pfn + 1;
371 
372 #ifdef CONFIG_HIGHMEM
373 	if (PageHighMem(page))
374 		totalhigh_pages++;
375 #endif
376 
377 	ClearPageReserved(page);
378 	init_page_count(page);
379 	__free_page(page);
380 }
381 
382 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
383 			void *arg)
384 {
385 	unsigned long i;
386 	unsigned long onlined_pages = *(unsigned long *)arg;
387 	struct page *page;
388 	if (PageReserved(pfn_to_page(start_pfn)))
389 		for (i = 0; i < nr_pages; i++) {
390 			page = pfn_to_page(start_pfn + i);
391 			online_page(page);
392 			onlined_pages++;
393 		}
394 	*(unsigned long *)arg = onlined_pages;
395 	return 0;
396 }
397 
398 
399 int __ref online_pages(unsigned long pfn, unsigned long nr_pages)
400 {
401 	unsigned long onlined_pages = 0;
402 	struct zone *zone;
403 	int need_zonelists_rebuild = 0;
404 	int nid;
405 	int ret;
406 	struct memory_notify arg;
407 
408 	lock_memory_hotplug();
409 	arg.start_pfn = pfn;
410 	arg.nr_pages = nr_pages;
411 	arg.status_change_nid = -1;
412 
413 	nid = page_to_nid(pfn_to_page(pfn));
414 	if (node_present_pages(nid) == 0)
415 		arg.status_change_nid = nid;
416 
417 	ret = memory_notify(MEM_GOING_ONLINE, &arg);
418 	ret = notifier_to_errno(ret);
419 	if (ret) {
420 		memory_notify(MEM_CANCEL_ONLINE, &arg);
421 		unlock_memory_hotplug();
422 		return ret;
423 	}
424 	/*
425 	 * This doesn't need a lock to do pfn_to_page().
426 	 * The section can't be removed here because of the
427 	 * memory_block->state_mutex.
428 	 */
429 	zone = page_zone(pfn_to_page(pfn));
430 	/*
431 	 * If this zone is not populated, then it is not in zonelist.
432 	 * This means the page allocator ignores this zone.
433 	 * So, zonelist must be updated after online.
434 	 */
435 	mutex_lock(&zonelists_mutex);
436 	if (!populated_zone(zone))
437 		need_zonelists_rebuild = 1;
438 
439 	ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages,
440 		online_pages_range);
441 	if (ret) {
442 		mutex_unlock(&zonelists_mutex);
443 		printk(KERN_DEBUG "online_pages %lx at %lx failed\n",
444 			nr_pages, pfn);
445 		memory_notify(MEM_CANCEL_ONLINE, &arg);
446 		unlock_memory_hotplug();
447 		return ret;
448 	}
449 
450 	zone->present_pages += onlined_pages;
451 	zone->zone_pgdat->node_present_pages += onlined_pages;
452 	if (need_zonelists_rebuild)
453 		build_all_zonelists(zone);
454 	else
455 		zone_pcp_update(zone);
456 
457 	mutex_unlock(&zonelists_mutex);
458 
459 	init_per_zone_wmark_min();
460 
461 	if (onlined_pages) {
462 		kswapd_run(zone_to_nid(zone));
463 		node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
464 	}
465 
466 	vm_total_pages = nr_free_pagecache_pages();
467 
468 	writeback_set_ratelimit();
469 
470 	if (onlined_pages)
471 		memory_notify(MEM_ONLINE, &arg);
472 	unlock_memory_hotplug();
473 
474 	return 0;
475 }
476 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
477 
478 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
479 static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
480 {
481 	struct pglist_data *pgdat;
482 	unsigned long zones_size[MAX_NR_ZONES] = {0};
483 	unsigned long zholes_size[MAX_NR_ZONES] = {0};
484 	unsigned long start_pfn = start >> PAGE_SHIFT;
485 
486 	pgdat = arch_alloc_nodedata(nid);
487 	if (!pgdat)
488 		return NULL;
489 
490 	arch_refresh_nodedata(nid, pgdat);
491 
492 	/* we can use NODE_DATA(nid) from here */
493 
494 	/* init node's zones as empty zones, we don't have any present pages.*/
495 	free_area_init_node(nid, zones_size, start_pfn, zholes_size);
496 
497 	/*
498 	 * The node we allocated has no zone fallback lists. For avoiding
499 	 * to access not-initialized zonelist, build here.
500 	 */
501 	mutex_lock(&zonelists_mutex);
502 	build_all_zonelists(NULL);
503 	mutex_unlock(&zonelists_mutex);
504 
505 	return pgdat;
506 }
507 
508 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
509 {
510 	arch_refresh_nodedata(nid, NULL);
511 	arch_free_nodedata(pgdat);
512 	return;
513 }
514 
515 
516 /*
517  * called by cpu_up() to online a node without onlined memory.
518  */
519 int mem_online_node(int nid)
520 {
521 	pg_data_t	*pgdat;
522 	int	ret;
523 
524 	lock_memory_hotplug();
525 	pgdat = hotadd_new_pgdat(nid, 0);
526 	if (!pgdat) {
527 		ret = -ENOMEM;
528 		goto out;
529 	}
530 	node_set_online(nid);
531 	ret = register_one_node(nid);
532 	BUG_ON(ret);
533 
534 out:
535 	unlock_memory_hotplug();
536 	return ret;
537 }
538 
539 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
540 int __ref add_memory(int nid, u64 start, u64 size)
541 {
542 	pg_data_t *pgdat = NULL;
543 	int new_pgdat = 0;
544 	struct resource *res;
545 	int ret;
546 
547 	lock_memory_hotplug();
548 
549 	res = register_memory_resource(start, size);
550 	ret = -EEXIST;
551 	if (!res)
552 		goto out;
553 
554 	if (!node_online(nid)) {
555 		pgdat = hotadd_new_pgdat(nid, start);
556 		ret = -ENOMEM;
557 		if (!pgdat)
558 			goto out;
559 		new_pgdat = 1;
560 	}
561 
562 	/* call arch's memory hotadd */
563 	ret = arch_add_memory(nid, start, size);
564 
565 	if (ret < 0)
566 		goto error;
567 
568 	/* we online node here. we can't roll back from here. */
569 	node_set_online(nid);
570 
571 	if (new_pgdat) {
572 		ret = register_one_node(nid);
573 		/*
574 		 * If sysfs file of new node can't create, cpu on the node
575 		 * can't be hot-added. There is no rollback way now.
576 		 * So, check by BUG_ON() to catch it reluctantly..
577 		 */
578 		BUG_ON(ret);
579 	}
580 
581 	/* create new memmap entry */
582 	firmware_map_add_hotplug(start, start + size, "System RAM");
583 
584 	goto out;
585 
586 error:
587 	/* rollback pgdat allocation and others */
588 	if (new_pgdat)
589 		rollback_node_hotadd(nid, pgdat);
590 	if (res)
591 		release_memory_resource(res);
592 
593 out:
594 	unlock_memory_hotplug();
595 	return ret;
596 }
597 EXPORT_SYMBOL_GPL(add_memory);
598 
599 #ifdef CONFIG_MEMORY_HOTREMOVE
600 /*
601  * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
602  * set and the size of the free page is given by page_order(). Using this,
603  * the function determines if the pageblock contains only free pages.
604  * Due to buddy contraints, a free page at least the size of a pageblock will
605  * be located at the start of the pageblock
606  */
607 static inline int pageblock_free(struct page *page)
608 {
609 	return PageBuddy(page) && page_order(page) >= pageblock_order;
610 }
611 
612 /* Return the start of the next active pageblock after a given page */
613 static struct page *next_active_pageblock(struct page *page)
614 {
615 	/* Ensure the starting page is pageblock-aligned */
616 	BUG_ON(page_to_pfn(page) & (pageblock_nr_pages - 1));
617 
618 	/* If the entire pageblock is free, move to the end of free page */
619 	if (pageblock_free(page)) {
620 		int order;
621 		/* be careful. we don't have locks, page_order can be changed.*/
622 		order = page_order(page);
623 		if ((order < MAX_ORDER) && (order >= pageblock_order))
624 			return page + (1 << order);
625 	}
626 
627 	return page + pageblock_nr_pages;
628 }
629 
630 /* Checks if this range of memory is likely to be hot-removable. */
631 int is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
632 {
633 	struct page *page = pfn_to_page(start_pfn);
634 	struct page *end_page = page + nr_pages;
635 
636 	/* Check the starting page of each pageblock within the range */
637 	for (; page < end_page; page = next_active_pageblock(page)) {
638 		if (!is_pageblock_removable_nolock(page))
639 			return 0;
640 		cond_resched();
641 	}
642 
643 	/* All pageblocks in the memory block are likely to be hot-removable */
644 	return 1;
645 }
646 
647 /*
648  * Confirm all pages in a range [start, end) is belongs to the same zone.
649  */
650 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
651 {
652 	unsigned long pfn;
653 	struct zone *zone = NULL;
654 	struct page *page;
655 	int i;
656 	for (pfn = start_pfn;
657 	     pfn < end_pfn;
658 	     pfn += MAX_ORDER_NR_PAGES) {
659 		i = 0;
660 		/* This is just a CONFIG_HOLES_IN_ZONE check.*/
661 		while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
662 			i++;
663 		if (i == MAX_ORDER_NR_PAGES)
664 			continue;
665 		page = pfn_to_page(pfn + i);
666 		if (zone && page_zone(page) != zone)
667 			return 0;
668 		zone = page_zone(page);
669 	}
670 	return 1;
671 }
672 
673 /*
674  * Scanning pfn is much easier than scanning lru list.
675  * Scan pfn from start to end and Find LRU page.
676  */
677 static unsigned long scan_lru_pages(unsigned long start, unsigned long end)
678 {
679 	unsigned long pfn;
680 	struct page *page;
681 	for (pfn = start; pfn < end; pfn++) {
682 		if (pfn_valid(pfn)) {
683 			page = pfn_to_page(pfn);
684 			if (PageLRU(page))
685 				return pfn;
686 		}
687 	}
688 	return 0;
689 }
690 
691 static struct page *
692 hotremove_migrate_alloc(struct page *page, unsigned long private, int **x)
693 {
694 	/* This should be improooooved!! */
695 	return alloc_page(GFP_HIGHUSER_MOVABLE);
696 }
697 
698 #define NR_OFFLINE_AT_ONCE_PAGES	(256)
699 static int
700 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
701 {
702 	unsigned long pfn;
703 	struct page *page;
704 	int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
705 	int not_managed = 0;
706 	int ret = 0;
707 	LIST_HEAD(source);
708 
709 	for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
710 		if (!pfn_valid(pfn))
711 			continue;
712 		page = pfn_to_page(pfn);
713 		if (!get_page_unless_zero(page))
714 			continue;
715 		/*
716 		 * We can skip free pages. And we can only deal with pages on
717 		 * LRU.
718 		 */
719 		ret = isolate_lru_page(page);
720 		if (!ret) { /* Success */
721 			put_page(page);
722 			list_add_tail(&page->lru, &source);
723 			move_pages--;
724 			inc_zone_page_state(page, NR_ISOLATED_ANON +
725 					    page_is_file_cache(page));
726 
727 		} else {
728 #ifdef CONFIG_DEBUG_VM
729 			printk(KERN_ALERT "removing pfn %lx from LRU failed\n",
730 			       pfn);
731 			dump_page(page);
732 #endif
733 			put_page(page);
734 			/* Because we don't have big zone->lock. we should
735 			   check this again here. */
736 			if (page_count(page)) {
737 				not_managed++;
738 				ret = -EBUSY;
739 				break;
740 			}
741 		}
742 	}
743 	if (!list_empty(&source)) {
744 		if (not_managed) {
745 			putback_lru_pages(&source);
746 			goto out;
747 		}
748 		/* this function returns # of failed pages */
749 		ret = migrate_pages(&source, hotremove_migrate_alloc, 0,
750 								true, true);
751 		if (ret)
752 			putback_lru_pages(&source);
753 	}
754 out:
755 	return ret;
756 }
757 
758 /*
759  * remove from free_area[] and mark all as Reserved.
760  */
761 static int
762 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
763 			void *data)
764 {
765 	__offline_isolated_pages(start, start + nr_pages);
766 	return 0;
767 }
768 
769 static void
770 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
771 {
772 	walk_system_ram_range(start_pfn, end_pfn - start_pfn, NULL,
773 				offline_isolated_pages_cb);
774 }
775 
776 /*
777  * Check all pages in range, recoreded as memory resource, are isolated.
778  */
779 static int
780 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
781 			void *data)
782 {
783 	int ret;
784 	long offlined = *(long *)data;
785 	ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
786 	offlined = nr_pages;
787 	if (!ret)
788 		*(long *)data += offlined;
789 	return ret;
790 }
791 
792 static long
793 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
794 {
795 	long offlined = 0;
796 	int ret;
797 
798 	ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, &offlined,
799 			check_pages_isolated_cb);
800 	if (ret < 0)
801 		offlined = (long)ret;
802 	return offlined;
803 }
804 
805 static int __ref offline_pages(unsigned long start_pfn,
806 		  unsigned long end_pfn, unsigned long timeout)
807 {
808 	unsigned long pfn, nr_pages, expire;
809 	long offlined_pages;
810 	int ret, drain, retry_max, node;
811 	struct zone *zone;
812 	struct memory_notify arg;
813 
814 	BUG_ON(start_pfn >= end_pfn);
815 	/* at least, alignment against pageblock is necessary */
816 	if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
817 		return -EINVAL;
818 	if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
819 		return -EINVAL;
820 	/* This makes hotplug much easier...and readable.
821 	   we assume this for now. .*/
822 	if (!test_pages_in_a_zone(start_pfn, end_pfn))
823 		return -EINVAL;
824 
825 	lock_memory_hotplug();
826 
827 	zone = page_zone(pfn_to_page(start_pfn));
828 	node = zone_to_nid(zone);
829 	nr_pages = end_pfn - start_pfn;
830 
831 	/* set above range as isolated */
832 	ret = start_isolate_page_range(start_pfn, end_pfn);
833 	if (ret)
834 		goto out;
835 
836 	arg.start_pfn = start_pfn;
837 	arg.nr_pages = nr_pages;
838 	arg.status_change_nid = -1;
839 	if (nr_pages >= node_present_pages(node))
840 		arg.status_change_nid = node;
841 
842 	ret = memory_notify(MEM_GOING_OFFLINE, &arg);
843 	ret = notifier_to_errno(ret);
844 	if (ret)
845 		goto failed_removal;
846 
847 	pfn = start_pfn;
848 	expire = jiffies + timeout;
849 	drain = 0;
850 	retry_max = 5;
851 repeat:
852 	/* start memory hot removal */
853 	ret = -EAGAIN;
854 	if (time_after(jiffies, expire))
855 		goto failed_removal;
856 	ret = -EINTR;
857 	if (signal_pending(current))
858 		goto failed_removal;
859 	ret = 0;
860 	if (drain) {
861 		lru_add_drain_all();
862 		cond_resched();
863 		drain_all_pages();
864 	}
865 
866 	pfn = scan_lru_pages(start_pfn, end_pfn);
867 	if (pfn) { /* We have page on LRU */
868 		ret = do_migrate_range(pfn, end_pfn);
869 		if (!ret) {
870 			drain = 1;
871 			goto repeat;
872 		} else {
873 			if (ret < 0)
874 				if (--retry_max == 0)
875 					goto failed_removal;
876 			yield();
877 			drain = 1;
878 			goto repeat;
879 		}
880 	}
881 	/* drain all zone's lru pagevec, this is asyncronous... */
882 	lru_add_drain_all();
883 	yield();
884 	/* drain pcp pages , this is synchrouns. */
885 	drain_all_pages();
886 	/* check again */
887 	offlined_pages = check_pages_isolated(start_pfn, end_pfn);
888 	if (offlined_pages < 0) {
889 		ret = -EBUSY;
890 		goto failed_removal;
891 	}
892 	printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
893 	/* Ok, all of our target is islaoted.
894 	   We cannot do rollback at this point. */
895 	offline_isolated_pages(start_pfn, end_pfn);
896 	/* reset pagetype flags and makes migrate type to be MOVABLE */
897 	undo_isolate_page_range(start_pfn, end_pfn);
898 	/* removal success */
899 	zone->present_pages -= offlined_pages;
900 	zone->zone_pgdat->node_present_pages -= offlined_pages;
901 	totalram_pages -= offlined_pages;
902 
903 	init_per_zone_wmark_min();
904 
905 	if (!node_present_pages(node)) {
906 		node_clear_state(node, N_HIGH_MEMORY);
907 		kswapd_stop(node);
908 	}
909 
910 	vm_total_pages = nr_free_pagecache_pages();
911 	writeback_set_ratelimit();
912 
913 	memory_notify(MEM_OFFLINE, &arg);
914 	unlock_memory_hotplug();
915 	return 0;
916 
917 failed_removal:
918 	printk(KERN_INFO "memory offlining %lx to %lx failed\n",
919 		start_pfn, end_pfn);
920 	memory_notify(MEM_CANCEL_OFFLINE, &arg);
921 	/* pushback to free area */
922 	undo_isolate_page_range(start_pfn, end_pfn);
923 
924 out:
925 	unlock_memory_hotplug();
926 	return ret;
927 }
928 
929 int remove_memory(u64 start, u64 size)
930 {
931 	unsigned long start_pfn, end_pfn;
932 
933 	start_pfn = PFN_DOWN(start);
934 	end_pfn = start_pfn + PFN_DOWN(size);
935 	return offline_pages(start_pfn, end_pfn, 120 * HZ);
936 }
937 #else
938 int remove_memory(u64 start, u64 size)
939 {
940 	return -EINVAL;
941 }
942 #endif /* CONFIG_MEMORY_HOTREMOVE */
943 EXPORT_SYMBOL_GPL(remove_memory);
944