xref: /openbmc/linux/mm/memory_hotplug.c (revision 732a675a)
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/cpuset.h>
26 #include <linux/delay.h>
27 #include <linux/migrate.h>
28 #include <linux/page-isolation.h>
29 
30 #include <asm/tlbflush.h>
31 
32 #include "internal.h"
33 
34 /* add this memory to iomem resource */
35 static struct resource *register_memory_resource(u64 start, u64 size)
36 {
37 	struct resource *res;
38 	res = kzalloc(sizeof(struct resource), GFP_KERNEL);
39 	BUG_ON(!res);
40 
41 	res->name = "System RAM";
42 	res->start = start;
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);
48 		kfree(res);
49 		res = NULL;
50 	}
51 	return res;
52 }
53 
54 static void release_memory_resource(struct resource *res)
55 {
56 	if (!res)
57 		return;
58 	release_resource(res);
59 	kfree(res);
60 	return;
61 }
62 
63 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
64 #ifndef CONFIG_SPARSEMEM_VMEMMAP
65 static void get_page_bootmem(unsigned long info,  struct page *page, int magic)
66 {
67 	atomic_set(&page->_mapcount, magic);
68 	SetPagePrivate(page);
69 	set_page_private(page, info);
70 	atomic_inc(&page->_count);
71 }
72 
73 void put_page_bootmem(struct page *page)
74 {
75 	int magic;
76 
77 	magic = atomic_read(&page->_mapcount);
78 	BUG_ON(magic >= -1);
79 
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);
85 	}
86 
87 }
88 
89 void register_page_bootmem_info_section(unsigned long start_pfn)
90 {
91 	unsigned long *usemap, mapsize, section_nr, i;
92 	struct mem_section *ms;
93 	struct page *page, *memmap;
94 
95 	if (!pfn_valid(start_pfn))
96 		return;
97 
98 	section_nr = pfn_to_section_nr(start_pfn);
99 	ms = __nr_to_section(section_nr);
100 
101 	/* Get section's memmap address */
102 	memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
103 
104 	/*
105 	 * Get page for the memmap's phys address
106 	 * XXX: need more consideration for sparse_vmemmap...
107 	 */
108 	page = virt_to_page(memmap);
109 	mapsize = sizeof(struct page) * PAGES_PER_SECTION;
110 	mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
111 
112 	/* remember memmap's page */
113 	for (i = 0; i < mapsize; i++, page++)
114 		get_page_bootmem(section_nr, page, SECTION_INFO);
115 
116 	usemap = __nr_to_section(section_nr)->pageblock_flags;
117 	page = virt_to_page(usemap);
118 
119 	mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT;
120 
121 	for (i = 0; i < mapsize; i++, page++)
122 		get_page_bootmem(section_nr, page, MIX_INFO);
123 
124 }
125 
126 void register_page_bootmem_info_node(struct pglist_data *pgdat)
127 {
128 	unsigned long i, pfn, end_pfn, nr_pages;
129 	int node = pgdat->node_id;
130 	struct page *page;
131 	struct zone *zone;
132 
133 	nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
134 	page = virt_to_page(pgdat);
135 
136 	for (i = 0; i < nr_pages; i++, page++)
137 		get_page_bootmem(node, page, NODE_INFO);
138 
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);
146 
147 			for (i = 0; i < nr_pages; i++, page++)
148 				get_page_bootmem(node, page, NODE_INFO);
149 		}
150 	}
151 
152 	pfn = pgdat->node_start_pfn;
153 	end_pfn = pfn + pgdat->node_spanned_pages;
154 
155 	/* register_section info */
156 	for (; pfn < end_pfn; pfn += PAGES_PER_SECTION)
157 		register_page_bootmem_info_section(pfn);
158 
159 }
160 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
161 
162 static void grow_zone_span(struct zone *zone, unsigned long start_pfn,
163 			   unsigned long end_pfn)
164 {
165 	unsigned long old_zone_end_pfn;
166 
167 	zone_span_writelock(zone);
168 
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;
172 
173 	zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
174 				zone->zone_start_pfn;
175 
176 	zone_span_writeunlock(zone);
177 }
178 
179 static void grow_pgdat_span(struct pglist_data *pgdat, unsigned long start_pfn,
180 			    unsigned long end_pfn)
181 {
182 	unsigned long old_pgdat_end_pfn =
183 		pgdat->node_start_pfn + pgdat->node_spanned_pages;
184 
185 	if (start_pfn < pgdat->node_start_pfn)
186 		pgdat->node_start_pfn = start_pfn;
187 
188 	pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
189 					pgdat->node_start_pfn;
190 }
191 
192 static int __add_zone(struct zone *zone, unsigned long phys_start_pfn)
193 {
194 	struct pglist_data *pgdat = zone->zone_pgdat;
195 	int nr_pages = PAGES_PER_SECTION;
196 	int nid = pgdat->node_id;
197 	int zone_type;
198 	unsigned long flags;
199 
200 	zone_type = zone - pgdat->node_zones;
201 	if (!zone->wait_table) {
202 		int ret;
203 
204 		ret = init_currently_empty_zone(zone, phys_start_pfn,
205 						nr_pages, MEMMAP_HOTPLUG);
206 		if (ret)
207 			return ret;
208 	}
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);
216 	return 0;
217 }
218 
219 static int __add_section(struct zone *zone, unsigned long phys_start_pfn)
220 {
221 	int nr_pages = PAGES_PER_SECTION;
222 	int ret;
223 
224 	if (pfn_valid(phys_start_pfn))
225 		return -EEXIST;
226 
227 	ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
228 
229 	if (ret < 0)
230 		return ret;
231 
232 	ret = __add_zone(zone, phys_start_pfn);
233 
234 	if (ret < 0)
235 		return ret;
236 
237 	return register_new_memory(__pfn_to_section(phys_start_pfn));
238 }
239 
240 #ifdef CONFIG_SPARSEMEM_VMEMMAP
241 static int __remove_section(struct zone *zone, struct mem_section *ms)
242 {
243 	/*
244 	 * XXX: Freeing memmap with vmemmap is not implement yet.
245 	 *      This should be removed later.
246 	 */
247 	return -EBUSY;
248 }
249 #else
250 static int __remove_section(struct zone *zone, struct mem_section *ms)
251 {
252 	unsigned long flags;
253 	struct pglist_data *pgdat = zone->zone_pgdat;
254 	int ret = -EINVAL;
255 
256 	if (!valid_section(ms))
257 		return ret;
258 
259 	ret = unregister_memory_section(ms);
260 	if (ret)
261 		return ret;
262 
263 	pgdat_resize_lock(pgdat, &flags);
264 	sparse_remove_one_section(zone, ms);
265 	pgdat_resize_unlock(pgdat, &flags);
266 	return 0;
267 }
268 #endif
269 
270 /*
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
274  * add the new pages.
275  */
276 int __add_pages(struct zone *zone, unsigned long phys_start_pfn,
277 		 unsigned long nr_pages)
278 {
279 	unsigned long i;
280 	int err = 0;
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);
285 
286 	for (i = start_sec; i <= end_sec; i++) {
287 		err = __add_section(zone, i << PFN_SECTION_SHIFT);
288 
289 		/*
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.
293 		 */
294 		if (err && (err != -EEXIST))
295 			break;
296 		err = 0;
297 	}
298 
299 	return err;
300 }
301 EXPORT_SYMBOL_GPL(__add_pages);
302 
303 /**
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)
308  *
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().
313  */
314 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
315 		 unsigned long nr_pages)
316 {
317 	unsigned long i, ret = 0;
318 	int sections_to_remove;
319 
320 	/*
321 	 * We can only remove entire sections
322 	 */
323 	BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
324 	BUG_ON(nr_pages % PAGES_PER_SECTION);
325 
326 	release_mem_region(phys_start_pfn << PAGE_SHIFT, nr_pages * PAGE_SIZE);
327 
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));
332 		if (ret)
333 			break;
334 	}
335 	return ret;
336 }
337 EXPORT_SYMBOL_GPL(__remove_pages);
338 
339 void online_page(struct page *page)
340 {
341 	totalram_pages++;
342 	num_physpages++;
343 
344 #ifdef CONFIG_HIGHMEM
345 	if (PageHighMem(page))
346 		totalhigh_pages++;
347 #endif
348 
349 #ifdef CONFIG_FLATMEM
350 	max_mapnr = max(page_to_pfn(page), max_mapnr);
351 #endif
352 
353 	ClearPageReserved(page);
354 	init_page_count(page);
355 	__free_page(page);
356 }
357 
358 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
359 			void *arg)
360 {
361 	unsigned long i;
362 	unsigned long onlined_pages = *(unsigned long *)arg;
363 	struct page *page;
364 	if (PageReserved(pfn_to_page(start_pfn)))
365 		for (i = 0; i < nr_pages; i++) {
366 			page = pfn_to_page(start_pfn + i);
367 			online_page(page);
368 			onlined_pages++;
369 		}
370 	*(unsigned long *)arg = onlined_pages;
371 	return 0;
372 }
373 
374 
375 int online_pages(unsigned long pfn, unsigned long nr_pages)
376 {
377 	unsigned long onlined_pages = 0;
378 	struct zone *zone;
379 	int need_zonelists_rebuild = 0;
380 	int nid;
381 	int ret;
382 	struct memory_notify arg;
383 
384 	arg.start_pfn = pfn;
385 	arg.nr_pages = nr_pages;
386 	arg.status_change_nid = -1;
387 
388 	nid = page_to_nid(pfn_to_page(pfn));
389 	if (node_present_pages(nid) == 0)
390 		arg.status_change_nid = nid;
391 
392 	ret = memory_notify(MEM_GOING_ONLINE, &arg);
393 	ret = notifier_to_errno(ret);
394 	if (ret) {
395 		memory_notify(MEM_CANCEL_ONLINE, &arg);
396 		return ret;
397 	}
398 	/*
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.
402 	 */
403 	zone = page_zone(pfn_to_page(pfn));
404 	/*
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.
408 	 */
409 	if (!populated_zone(zone))
410 		need_zonelists_rebuild = 1;
411 
412 	ret = walk_memory_resource(pfn, nr_pages, &onlined_pages,
413 		online_pages_range);
414 	if (ret) {
415 		printk(KERN_DEBUG "online_pages %lx at %lx failed\n",
416 			nr_pages, pfn);
417 		memory_notify(MEM_CANCEL_ONLINE, &arg);
418 		return ret;
419 	}
420 
421 	zone->present_pages += onlined_pages;
422 	zone->zone_pgdat->node_present_pages += onlined_pages;
423 
424 	setup_per_zone_pages_min();
425 	if (onlined_pages) {
426 		kswapd_run(zone_to_nid(zone));
427 		node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
428 	}
429 
430 	if (need_zonelists_rebuild)
431 		build_all_zonelists();
432 	vm_total_pages = nr_free_pagecache_pages();
433 	writeback_set_ratelimit();
434 
435 	if (onlined_pages)
436 		memory_notify(MEM_ONLINE, &arg);
437 
438 	return 0;
439 }
440 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
441 
442 static pg_data_t *hotadd_new_pgdat(int nid, u64 start)
443 {
444 	struct pglist_data *pgdat;
445 	unsigned long zones_size[MAX_NR_ZONES] = {0};
446 	unsigned long zholes_size[MAX_NR_ZONES] = {0};
447 	unsigned long start_pfn = start >> PAGE_SHIFT;
448 
449 	pgdat = arch_alloc_nodedata(nid);
450 	if (!pgdat)
451 		return NULL;
452 
453 	arch_refresh_nodedata(nid, pgdat);
454 
455 	/* we can use NODE_DATA(nid) from here */
456 
457 	/* init node's zones as empty zones, we don't have any present pages.*/
458 	free_area_init_node(nid, pgdat, zones_size, start_pfn, zholes_size);
459 
460 	return pgdat;
461 }
462 
463 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
464 {
465 	arch_refresh_nodedata(nid, NULL);
466 	arch_free_nodedata(pgdat);
467 	return;
468 }
469 
470 
471 int add_memory(int nid, u64 start, u64 size)
472 {
473 	pg_data_t *pgdat = NULL;
474 	int new_pgdat = 0;
475 	struct resource *res;
476 	int ret;
477 
478 	res = register_memory_resource(start, size);
479 	if (!res)
480 		return -EEXIST;
481 
482 	if (!node_online(nid)) {
483 		pgdat = hotadd_new_pgdat(nid, start);
484 		if (!pgdat)
485 			return -ENOMEM;
486 		new_pgdat = 1;
487 	}
488 
489 	/* call arch's memory hotadd */
490 	ret = arch_add_memory(nid, start, size);
491 
492 	if (ret < 0)
493 		goto error;
494 
495 	/* we online node here. we can't roll back from here. */
496 	node_set_online(nid);
497 
498 	cpuset_track_online_nodes();
499 
500 	if (new_pgdat) {
501 		ret = register_one_node(nid);
502 		/*
503 		 * If sysfs file of new node can't create, cpu on the node
504 		 * can't be hot-added. There is no rollback way now.
505 		 * So, check by BUG_ON() to catch it reluctantly..
506 		 */
507 		BUG_ON(ret);
508 	}
509 
510 	return ret;
511 error:
512 	/* rollback pgdat allocation and others */
513 	if (new_pgdat)
514 		rollback_node_hotadd(nid, pgdat);
515 	if (res)
516 		release_memory_resource(res);
517 
518 	return ret;
519 }
520 EXPORT_SYMBOL_GPL(add_memory);
521 
522 #ifdef CONFIG_MEMORY_HOTREMOVE
523 /*
524  * Confirm all pages in a range [start, end) is belongs to the same zone.
525  */
526 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
527 {
528 	unsigned long pfn;
529 	struct zone *zone = NULL;
530 	struct page *page;
531 	int i;
532 	for (pfn = start_pfn;
533 	     pfn < end_pfn;
534 	     pfn += MAX_ORDER_NR_PAGES) {
535 		i = 0;
536 		/* This is just a CONFIG_HOLES_IN_ZONE check.*/
537 		while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
538 			i++;
539 		if (i == MAX_ORDER_NR_PAGES)
540 			continue;
541 		page = pfn_to_page(pfn + i);
542 		if (zone && page_zone(page) != zone)
543 			return 0;
544 		zone = page_zone(page);
545 	}
546 	return 1;
547 }
548 
549 /*
550  * Scanning pfn is much easier than scanning lru list.
551  * Scan pfn from start to end and Find LRU page.
552  */
553 int scan_lru_pages(unsigned long start, unsigned long end)
554 {
555 	unsigned long pfn;
556 	struct page *page;
557 	for (pfn = start; pfn < end; pfn++) {
558 		if (pfn_valid(pfn)) {
559 			page = pfn_to_page(pfn);
560 			if (PageLRU(page))
561 				return pfn;
562 		}
563 	}
564 	return 0;
565 }
566 
567 static struct page *
568 hotremove_migrate_alloc(struct page *page,
569 			unsigned long private,
570 			int **x)
571 {
572 	/* This should be improoooooved!! */
573 	return alloc_page(GFP_HIGHUSER_PAGECACHE);
574 }
575 
576 
577 #define NR_OFFLINE_AT_ONCE_PAGES	(256)
578 static int
579 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
580 {
581 	unsigned long pfn;
582 	struct page *page;
583 	int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
584 	int not_managed = 0;
585 	int ret = 0;
586 	LIST_HEAD(source);
587 
588 	for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
589 		if (!pfn_valid(pfn))
590 			continue;
591 		page = pfn_to_page(pfn);
592 		if (!page_count(page))
593 			continue;
594 		/*
595 		 * We can skip free pages. And we can only deal with pages on
596 		 * LRU.
597 		 */
598 		ret = isolate_lru_page(page, &source);
599 		if (!ret) { /* Success */
600 			move_pages--;
601 		} else {
602 			/* Becasue we don't have big zone->lock. we should
603 			   check this again here. */
604 			if (page_count(page))
605 				not_managed++;
606 #ifdef CONFIG_DEBUG_VM
607 			printk(KERN_INFO "removing from LRU failed"
608 					 " %lx/%d/%lx\n",
609 				pfn, page_count(page), page->flags);
610 #endif
611 		}
612 	}
613 	ret = -EBUSY;
614 	if (not_managed) {
615 		if (!list_empty(&source))
616 			putback_lru_pages(&source);
617 		goto out;
618 	}
619 	ret = 0;
620 	if (list_empty(&source))
621 		goto out;
622 	/* this function returns # of failed pages */
623 	ret = migrate_pages(&source, hotremove_migrate_alloc, 0);
624 
625 out:
626 	return ret;
627 }
628 
629 /*
630  * remove from free_area[] and mark all as Reserved.
631  */
632 static int
633 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
634 			void *data)
635 {
636 	__offline_isolated_pages(start, start + nr_pages);
637 	return 0;
638 }
639 
640 static void
641 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
642 {
643 	walk_memory_resource(start_pfn, end_pfn - start_pfn, NULL,
644 				offline_isolated_pages_cb);
645 }
646 
647 /*
648  * Check all pages in range, recoreded as memory resource, are isolated.
649  */
650 static int
651 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
652 			void *data)
653 {
654 	int ret;
655 	long offlined = *(long *)data;
656 	ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
657 	offlined = nr_pages;
658 	if (!ret)
659 		*(long *)data += offlined;
660 	return ret;
661 }
662 
663 static long
664 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
665 {
666 	long offlined = 0;
667 	int ret;
668 
669 	ret = walk_memory_resource(start_pfn, end_pfn - start_pfn, &offlined,
670 			check_pages_isolated_cb);
671 	if (ret < 0)
672 		offlined = (long)ret;
673 	return offlined;
674 }
675 
676 int offline_pages(unsigned long start_pfn,
677 		  unsigned long end_pfn, unsigned long timeout)
678 {
679 	unsigned long pfn, nr_pages, expire;
680 	long offlined_pages;
681 	int ret, drain, retry_max, node;
682 	struct zone *zone;
683 	struct memory_notify arg;
684 
685 	BUG_ON(start_pfn >= end_pfn);
686 	/* at least, alignment against pageblock is necessary */
687 	if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
688 		return -EINVAL;
689 	if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
690 		return -EINVAL;
691 	/* This makes hotplug much easier...and readable.
692 	   we assume this for now. .*/
693 	if (!test_pages_in_a_zone(start_pfn, end_pfn))
694 		return -EINVAL;
695 
696 	zone = page_zone(pfn_to_page(start_pfn));
697 	node = zone_to_nid(zone);
698 	nr_pages = end_pfn - start_pfn;
699 
700 	/* set above range as isolated */
701 	ret = start_isolate_page_range(start_pfn, end_pfn);
702 	if (ret)
703 		return ret;
704 
705 	arg.start_pfn = start_pfn;
706 	arg.nr_pages = nr_pages;
707 	arg.status_change_nid = -1;
708 	if (nr_pages >= node_present_pages(node))
709 		arg.status_change_nid = node;
710 
711 	ret = memory_notify(MEM_GOING_OFFLINE, &arg);
712 	ret = notifier_to_errno(ret);
713 	if (ret)
714 		goto failed_removal;
715 
716 	pfn = start_pfn;
717 	expire = jiffies + timeout;
718 	drain = 0;
719 	retry_max = 5;
720 repeat:
721 	/* start memory hot removal */
722 	ret = -EAGAIN;
723 	if (time_after(jiffies, expire))
724 		goto failed_removal;
725 	ret = -EINTR;
726 	if (signal_pending(current))
727 		goto failed_removal;
728 	ret = 0;
729 	if (drain) {
730 		lru_add_drain_all();
731 		flush_scheduled_work();
732 		cond_resched();
733 		drain_all_pages();
734 	}
735 
736 	pfn = scan_lru_pages(start_pfn, end_pfn);
737 	if (pfn) { /* We have page on LRU */
738 		ret = do_migrate_range(pfn, end_pfn);
739 		if (!ret) {
740 			drain = 1;
741 			goto repeat;
742 		} else {
743 			if (ret < 0)
744 				if (--retry_max == 0)
745 					goto failed_removal;
746 			yield();
747 			drain = 1;
748 			goto repeat;
749 		}
750 	}
751 	/* drain all zone's lru pagevec, this is asyncronous... */
752 	lru_add_drain_all();
753 	flush_scheduled_work();
754 	yield();
755 	/* drain pcp pages , this is synchrouns. */
756 	drain_all_pages();
757 	/* check again */
758 	offlined_pages = check_pages_isolated(start_pfn, end_pfn);
759 	if (offlined_pages < 0) {
760 		ret = -EBUSY;
761 		goto failed_removal;
762 	}
763 	printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
764 	/* Ok, all of our target is islaoted.
765 	   We cannot do rollback at this point. */
766 	offline_isolated_pages(start_pfn, end_pfn);
767 	/* reset pagetype flags and makes migrate type to be MOVABLE */
768 	undo_isolate_page_range(start_pfn, end_pfn);
769 	/* removal success */
770 	zone->present_pages -= offlined_pages;
771 	zone->zone_pgdat->node_present_pages -= offlined_pages;
772 	totalram_pages -= offlined_pages;
773 	num_physpages -= offlined_pages;
774 
775 	vm_total_pages = nr_free_pagecache_pages();
776 	writeback_set_ratelimit();
777 
778 	memory_notify(MEM_OFFLINE, &arg);
779 	return 0;
780 
781 failed_removal:
782 	printk(KERN_INFO "memory offlining %lx to %lx failed\n",
783 		start_pfn, end_pfn);
784 	memory_notify(MEM_CANCEL_OFFLINE, &arg);
785 	/* pushback to free area */
786 	undo_isolate_page_range(start_pfn, end_pfn);
787 
788 	return ret;
789 }
790 #else
791 int remove_memory(u64 start, u64 size)
792 {
793 	return -EINVAL;
794 }
795 EXPORT_SYMBOL_GPL(remove_memory);
796 #endif /* CONFIG_MEMORY_HOTREMOVE */
797