mm: fix wrong vmap address calculations with odd NR_CPUS values
[linux-2.6.git] / mm / memory_hotplug.c
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 /*
38  * online_page_callback contains pointer to current page onlining function.
39  * Initially it is generic_online_page(). If it is required it could be
40  * changed by calling set_online_page_callback() for callback registration
41  * and restore_online_page_callback() for generic callback restore.
42  */
43
44 static void generic_online_page(struct page *page);
45
46 static online_page_callback_t online_page_callback = generic_online_page;
47
48 DEFINE_MUTEX(mem_hotplug_mutex);
49
50 void lock_memory_hotplug(void)
51 {
52         mutex_lock(&mem_hotplug_mutex);
53
54         /* for exclusive hibernation if CONFIG_HIBERNATION=y */
55         lock_system_sleep();
56 }
57
58 void unlock_memory_hotplug(void)
59 {
60         unlock_system_sleep();
61         mutex_unlock(&mem_hotplug_mutex);
62 }
63
64
65 /* add this memory to iomem resource */
66 static struct resource *register_memory_resource(u64 start, u64 size)
67 {
68         struct resource *res;
69         res = kzalloc(sizeof(struct resource), GFP_KERNEL);
70         BUG_ON(!res);
71
72         res->name = "System RAM";
73         res->start = start;
74         res->end = start + size - 1;
75         res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
76         if (request_resource(&iomem_resource, res) < 0) {
77                 printk("System RAM resource %llx - %llx cannot be added\n",
78                 (unsigned long long)res->start, (unsigned long long)res->end);
79                 kfree(res);
80                 res = NULL;
81         }
82         return res;
83 }
84
85 static void release_memory_resource(struct resource *res)
86 {
87         if (!res)
88                 return;
89         release_resource(res);
90         kfree(res);
91         return;
92 }
93
94 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
95 #ifndef CONFIG_SPARSEMEM_VMEMMAP
96 static void get_page_bootmem(unsigned long info,  struct page *page,
97                              unsigned long type)
98 {
99         page->lru.next = (struct list_head *) type;
100         SetPagePrivate(page);
101         set_page_private(page, info);
102         atomic_inc(&page->_count);
103 }
104
105 /* reference to __meminit __free_pages_bootmem is valid
106  * so use __ref to tell modpost not to generate a warning */
107 void __ref put_page_bootmem(struct page *page)
108 {
109         unsigned long type;
110
111         type = (unsigned long) page->lru.next;
112         BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE ||
113                type > MEMORY_HOTPLUG_MAX_BOOTMEM_TYPE);
114
115         if (atomic_dec_return(&page->_count) == 1) {
116                 ClearPagePrivate(page);
117                 set_page_private(page, 0);
118                 INIT_LIST_HEAD(&page->lru);
119                 __free_pages_bootmem(page, 0);
120         }
121
122 }
123
124 static void register_page_bootmem_info_section(unsigned long start_pfn)
125 {
126         unsigned long *usemap, mapsize, section_nr, i;
127         struct mem_section *ms;
128         struct page *page, *memmap;
129
130         if (!pfn_valid(start_pfn))
131                 return;
132
133         section_nr = pfn_to_section_nr(start_pfn);
134         ms = __nr_to_section(section_nr);
135
136         /* Get section's memmap address */
137         memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
138
139         /*
140          * Get page for the memmap's phys address
141          * XXX: need more consideration for sparse_vmemmap...
142          */
143         page = virt_to_page(memmap);
144         mapsize = sizeof(struct page) * PAGES_PER_SECTION;
145         mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
146
147         /* remember memmap's page */
148         for (i = 0; i < mapsize; i++, page++)
149                 get_page_bootmem(section_nr, page, SECTION_INFO);
150
151         usemap = __nr_to_section(section_nr)->pageblock_flags;
152         page = virt_to_page(usemap);
153
154         mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT;
155
156         for (i = 0; i < mapsize; i++, page++)
157                 get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
158
159 }
160
161 void register_page_bootmem_info_node(struct pglist_data *pgdat)
162 {
163         unsigned long i, pfn, end_pfn, nr_pages;
164         int node = pgdat->node_id;
165         struct page *page;
166         struct zone *zone;
167
168         nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
169         page = virt_to_page(pgdat);
170
171         for (i = 0; i < nr_pages; i++, page++)
172                 get_page_bootmem(node, page, NODE_INFO);
173
174         zone = &pgdat->node_zones[0];
175         for (; zone < pgdat->node_zones + MAX_NR_ZONES - 1; zone++) {
176                 if (zone->wait_table) {
177                         nr_pages = zone->wait_table_hash_nr_entries
178                                 * sizeof(wait_queue_head_t);
179                         nr_pages = PAGE_ALIGN(nr_pages) >> PAGE_SHIFT;
180                         page = virt_to_page(zone->wait_table);
181
182                         for (i = 0; i < nr_pages; i++, page++)
183                                 get_page_bootmem(node, page, NODE_INFO);
184                 }
185         }
186
187         pfn = pgdat->node_start_pfn;
188         end_pfn = pfn + pgdat->node_spanned_pages;
189
190         /* register_section info */
191         for (; pfn < end_pfn; pfn += PAGES_PER_SECTION)
192                 register_page_bootmem_info_section(pfn);
193
194 }
195 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
196
197 static void grow_zone_span(struct zone *zone, unsigned long start_pfn,
198                            unsigned long end_pfn)
199 {
200         unsigned long old_zone_end_pfn;
201
202         zone_span_writelock(zone);
203
204         old_zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
205         if (start_pfn < zone->zone_start_pfn)
206                 zone->zone_start_pfn = start_pfn;
207
208         zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
209                                 zone->zone_start_pfn;
210
211         zone_span_writeunlock(zone);
212 }
213
214 static void grow_pgdat_span(struct pglist_data *pgdat, unsigned long start_pfn,
215                             unsigned long end_pfn)
216 {
217         unsigned long old_pgdat_end_pfn =
218                 pgdat->node_start_pfn + pgdat->node_spanned_pages;
219
220         if (start_pfn < pgdat->node_start_pfn)
221                 pgdat->node_start_pfn = start_pfn;
222
223         pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
224                                         pgdat->node_start_pfn;
225 }
226
227 static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
228 {
229         struct pglist_data *pgdat = zone->zone_pgdat;
230         int nr_pages = PAGES_PER_SECTION;
231         int nid = pgdat->node_id;
232         int zone_type;
233         unsigned long flags;
234
235         zone_type = zone - pgdat->node_zones;
236         if (!zone->wait_table) {
237                 int ret;
238
239                 ret = init_currently_empty_zone(zone, phys_start_pfn,
240                                                 nr_pages, MEMMAP_HOTPLUG);
241                 if (ret)
242                         return ret;
243         }
244         pgdat_resize_lock(zone->zone_pgdat, &flags);
245         grow_zone_span(zone, phys_start_pfn, phys_start_pfn + nr_pages);
246         grow_pgdat_span(zone->zone_pgdat, phys_start_pfn,
247                         phys_start_pfn + nr_pages);
248         pgdat_resize_unlock(zone->zone_pgdat, &flags);
249         memmap_init_zone(nr_pages, nid, zone_type,
250                          phys_start_pfn, MEMMAP_HOTPLUG);
251         return 0;
252 }
253
254 static int __meminit __add_section(int nid, struct zone *zone,
255                                         unsigned long phys_start_pfn)
256 {
257         int nr_pages = PAGES_PER_SECTION;
258         int ret;
259
260         if (pfn_valid(phys_start_pfn))
261                 return -EEXIST;
262
263         ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
264
265         if (ret < 0)
266                 return ret;
267
268         ret = __add_zone(zone, phys_start_pfn);
269
270         if (ret < 0)
271                 return ret;
272
273         return register_new_memory(nid, __pfn_to_section(phys_start_pfn));
274 }
275
276 #ifdef CONFIG_SPARSEMEM_VMEMMAP
277 static int __remove_section(struct zone *zone, struct mem_section *ms)
278 {
279         /*
280          * XXX: Freeing memmap with vmemmap is not implement yet.
281          *      This should be removed later.
282          */
283         return -EBUSY;
284 }
285 #else
286 static int __remove_section(struct zone *zone, struct mem_section *ms)
287 {
288         unsigned long flags;
289         struct pglist_data *pgdat = zone->zone_pgdat;
290         int ret = -EINVAL;
291
292         if (!valid_section(ms))
293                 return ret;
294
295         ret = unregister_memory_section(ms);
296         if (ret)
297                 return ret;
298
299         pgdat_resize_lock(pgdat, &flags);
300         sparse_remove_one_section(zone, ms);
301         pgdat_resize_unlock(pgdat, &flags);
302         return 0;
303 }
304 #endif
305
306 /*
307  * Reasonably generic function for adding memory.  It is
308  * expected that archs that support memory hotplug will
309  * call this function after deciding the zone to which to
310  * add the new pages.
311  */
312 int __ref __add_pages(int nid, struct zone *zone, unsigned long phys_start_pfn,
313                         unsigned long nr_pages)
314 {
315         unsigned long i;
316         int err = 0;
317         int start_sec, end_sec;
318         /* during initialize mem_map, align hot-added range to section */
319         start_sec = pfn_to_section_nr(phys_start_pfn);
320         end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
321
322         for (i = start_sec; i <= end_sec; i++) {
323                 err = __add_section(nid, zone, i << PFN_SECTION_SHIFT);
324
325                 /*
326                  * EEXIST is finally dealt with by ioresource collision
327                  * check. see add_memory() => register_memory_resource()
328                  * Warning will be printed if there is collision.
329                  */
330                 if (err && (err != -EEXIST))
331                         break;
332                 err = 0;
333         }
334
335         return err;
336 }
337 EXPORT_SYMBOL_GPL(__add_pages);
338
339 /**
340  * __remove_pages() - remove sections of pages from a zone
341  * @zone: zone from which pages need to be removed
342  * @phys_start_pfn: starting pageframe (must be aligned to start of a section)
343  * @nr_pages: number of pages to remove (must be multiple of section size)
344  *
345  * Generic helper function to remove section mappings and sysfs entries
346  * for the section of the memory we are removing. Caller needs to make
347  * sure that pages are marked reserved and zones are adjust properly by
348  * calling offline_pages().
349  */
350 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
351                  unsigned long nr_pages)
352 {
353         unsigned long i, ret = 0;
354         int sections_to_remove;
355
356         /*
357          * We can only remove entire sections
358          */
359         BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
360         BUG_ON(nr_pages % PAGES_PER_SECTION);
361
362         sections_to_remove = nr_pages / PAGES_PER_SECTION;
363         for (i = 0; i < sections_to_remove; i++) {
364                 unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION;
365                 release_mem_region(pfn << PAGE_SHIFT,
366                                    PAGES_PER_SECTION << PAGE_SHIFT);
367                 ret = __remove_section(zone, __pfn_to_section(pfn));
368                 if (ret)
369                         break;
370         }
371         return ret;
372 }
373 EXPORT_SYMBOL_GPL(__remove_pages);
374
375 int set_online_page_callback(online_page_callback_t callback)
376 {
377         int rc = -EINVAL;
378
379         lock_memory_hotplug();
380
381         if (online_page_callback == generic_online_page) {
382                 online_page_callback = callback;
383                 rc = 0;
384         }
385
386         unlock_memory_hotplug();
387
388         return rc;
389 }
390 EXPORT_SYMBOL_GPL(set_online_page_callback);
391
392 int restore_online_page_callback(online_page_callback_t callback)
393 {
394         int rc = -EINVAL;
395
396         lock_memory_hotplug();
397
398         if (online_page_callback == callback) {
399                 online_page_callback = generic_online_page;
400                 rc = 0;
401         }
402
403         unlock_memory_hotplug();
404
405         return rc;
406 }
407 EXPORT_SYMBOL_GPL(restore_online_page_callback);
408
409 void __online_page_set_limits(struct page *page)
410 {
411         unsigned long pfn = page_to_pfn(page);
412
413         if (pfn >= num_physpages)
414                 num_physpages = pfn + 1;
415 }
416 EXPORT_SYMBOL_GPL(__online_page_set_limits);
417
418 void __online_page_increment_counters(struct page *page)
419 {
420         totalram_pages++;
421
422 #ifdef CONFIG_HIGHMEM
423         if (PageHighMem(page))
424                 totalhigh_pages++;
425 #endif
426 }
427 EXPORT_SYMBOL_GPL(__online_page_increment_counters);
428
429 void __online_page_free(struct page *page)
430 {
431         ClearPageReserved(page);
432         init_page_count(page);
433         __free_page(page);
434 }
435 EXPORT_SYMBOL_GPL(__online_page_free);
436
437 static void generic_online_page(struct page *page)
438 {
439         __online_page_set_limits(page);
440         __online_page_increment_counters(page);
441         __online_page_free(page);
442 }
443
444 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
445                         void *arg)
446 {
447         unsigned long i;
448         unsigned long onlined_pages = *(unsigned long *)arg;
449         struct page *page;
450         if (PageReserved(pfn_to_page(start_pfn)))
451                 for (i = 0; i < nr_pages; i++) {
452                         page = pfn_to_page(start_pfn + i);
453                         (*online_page_callback)(page);
454                         onlined_pages++;
455                 }
456         *(unsigned long *)arg = onlined_pages;
457         return 0;
458 }
459
460
461 int __ref online_pages(unsigned long pfn, unsigned long nr_pages)
462 {
463         unsigned long onlined_pages = 0;
464         struct zone *zone;
465         int need_zonelists_rebuild = 0;
466         int nid;
467         int ret;
468         struct memory_notify arg;
469
470         lock_memory_hotplug();
471         arg.start_pfn = pfn;
472         arg.nr_pages = nr_pages;
473         arg.status_change_nid = -1;
474
475         nid = page_to_nid(pfn_to_page(pfn));
476         if (node_present_pages(nid) == 0)
477                 arg.status_change_nid = nid;
478
479         ret = memory_notify(MEM_GOING_ONLINE, &arg);
480         ret = notifier_to_errno(ret);
481         if (ret) {
482                 memory_notify(MEM_CANCEL_ONLINE, &arg);
483                 unlock_memory_hotplug();
484                 return ret;
485         }
486         /*
487          * This doesn't need a lock to do pfn_to_page().
488          * The section can't be removed here because of the
489          * memory_block->state_mutex.
490          */
491         zone = page_zone(pfn_to_page(pfn));
492         /*
493          * If this zone is not populated, then it is not in zonelist.
494          * This means the page allocator ignores this zone.
495          * So, zonelist must be updated after online.
496          */
497         mutex_lock(&zonelists_mutex);
498         if (!populated_zone(zone))
499                 need_zonelists_rebuild = 1;
500
501         ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages,
502                 online_pages_range);
503         if (ret) {
504                 mutex_unlock(&zonelists_mutex);
505                 printk(KERN_DEBUG "online_pages %lx at %lx failed\n",
506                         nr_pages, pfn);
507                 memory_notify(MEM_CANCEL_ONLINE, &arg);
508                 unlock_memory_hotplug();
509                 return ret;
510         }
511
512         zone->present_pages += onlined_pages;
513         zone->zone_pgdat->node_present_pages += onlined_pages;
514         if (need_zonelists_rebuild)
515                 build_all_zonelists(zone);
516         else
517                 zone_pcp_update(zone);
518
519         mutex_unlock(&zonelists_mutex);
520
521         init_per_zone_wmark_min();
522
523         if (onlined_pages) {
524                 kswapd_run(zone_to_nid(zone));
525                 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
526         }
527
528         vm_total_pages = nr_free_pagecache_pages();
529
530         writeback_set_ratelimit();
531
532         if (onlined_pages)
533                 memory_notify(MEM_ONLINE, &arg);
534         unlock_memory_hotplug();
535
536         return 0;
537 }
538 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
539
540 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
541 static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
542 {
543         struct pglist_data *pgdat;
544         unsigned long zones_size[MAX_NR_ZONES] = {0};
545         unsigned long zholes_size[MAX_NR_ZONES] = {0};
546         unsigned long start_pfn = start >> PAGE_SHIFT;
547
548         pgdat = arch_alloc_nodedata(nid);
549         if (!pgdat)
550                 return NULL;
551
552         arch_refresh_nodedata(nid, pgdat);
553
554         /* we can use NODE_DATA(nid) from here */
555
556         /* init node's zones as empty zones, we don't have any present pages.*/
557         free_area_init_node(nid, zones_size, start_pfn, zholes_size);
558
559         /*
560          * The node we allocated has no zone fallback lists. For avoiding
561          * to access not-initialized zonelist, build here.
562          */
563         mutex_lock(&zonelists_mutex);
564         build_all_zonelists(NULL);
565         mutex_unlock(&zonelists_mutex);
566
567         return pgdat;
568 }
569
570 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
571 {
572         arch_refresh_nodedata(nid, NULL);
573         arch_free_nodedata(pgdat);
574         return;
575 }
576
577
578 /*
579  * called by cpu_up() to online a node without onlined memory.
580  */
581 int mem_online_node(int nid)
582 {
583         pg_data_t       *pgdat;
584         int     ret;
585
586         lock_memory_hotplug();
587         pgdat = hotadd_new_pgdat(nid, 0);
588         if (!pgdat) {
589                 ret = -ENOMEM;
590                 goto out;
591         }
592         node_set_online(nid);
593         ret = register_one_node(nid);
594         BUG_ON(ret);
595
596 out:
597         unlock_memory_hotplug();
598         return ret;
599 }
600
601 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
602 int __ref add_memory(int nid, u64 start, u64 size)
603 {
604         pg_data_t *pgdat = NULL;
605         int new_pgdat = 0;
606         struct resource *res;
607         int ret;
608
609         lock_memory_hotplug();
610
611         res = register_memory_resource(start, size);
612         ret = -EEXIST;
613         if (!res)
614                 goto out;
615
616         if (!node_online(nid)) {
617                 pgdat = hotadd_new_pgdat(nid, start);
618                 ret = -ENOMEM;
619                 if (!pgdat)
620                         goto out;
621                 new_pgdat = 1;
622         }
623
624         /* call arch's memory hotadd */
625         ret = arch_add_memory(nid, start, size);
626
627         if (ret < 0)
628                 goto error;
629
630         /* we online node here. we can't roll back from here. */
631         node_set_online(nid);
632
633         if (new_pgdat) {
634                 ret = register_one_node(nid);
635                 /*
636                  * If sysfs file of new node can't create, cpu on the node
637                  * can't be hot-added. There is no rollback way now.
638                  * So, check by BUG_ON() to catch it reluctantly..
639                  */
640                 BUG_ON(ret);
641         }
642
643         /* create new memmap entry */
644         firmware_map_add_hotplug(start, start + size, "System RAM");
645
646         goto out;
647
648 error:
649         /* rollback pgdat allocation and others */
650         if (new_pgdat)
651                 rollback_node_hotadd(nid, pgdat);
652         if (res)
653                 release_memory_resource(res);
654
655 out:
656         unlock_memory_hotplug();
657         return ret;
658 }
659 EXPORT_SYMBOL_GPL(add_memory);
660
661 #ifdef CONFIG_MEMORY_HOTREMOVE
662 /*
663  * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
664  * set and the size of the free page is given by page_order(). Using this,
665  * the function determines if the pageblock contains only free pages.
666  * Due to buddy contraints, a free page at least the size of a pageblock will
667  * be located at the start of the pageblock
668  */
669 static inline int pageblock_free(struct page *page)
670 {
671         return PageBuddy(page) && page_order(page) >= pageblock_order;
672 }
673
674 /* Return the start of the next active pageblock after a given page */
675 static struct page *next_active_pageblock(struct page *page)
676 {
677         /* Ensure the starting page is pageblock-aligned */
678         BUG_ON(page_to_pfn(page) & (pageblock_nr_pages - 1));
679
680         /* If the entire pageblock is free, move to the end of free page */
681         if (pageblock_free(page)) {
682                 int order;
683                 /* be careful. we don't have locks, page_order can be changed.*/
684                 order = page_order(page);
685                 if ((order < MAX_ORDER) && (order >= pageblock_order))
686                         return page + (1 << order);
687         }
688
689         return page + pageblock_nr_pages;
690 }
691
692 /* Checks if this range of memory is likely to be hot-removable. */
693 int is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
694 {
695         struct page *page = pfn_to_page(start_pfn);
696         struct page *end_page = page + nr_pages;
697
698         /* Check the starting page of each pageblock within the range */
699         for (; page < end_page; page = next_active_pageblock(page)) {
700                 if (!is_pageblock_removable_nolock(page))
701                         return 0;
702                 cond_resched();
703         }
704
705         /* All pageblocks in the memory block are likely to be hot-removable */
706         return 1;
707 }
708
709 /*
710  * Confirm all pages in a range [start, end) is belongs to the same zone.
711  */
712 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
713 {
714         unsigned long pfn;
715         struct zone *zone = NULL;
716         struct page *page;
717         int i;
718         for (pfn = start_pfn;
719              pfn < end_pfn;
720              pfn += MAX_ORDER_NR_PAGES) {
721                 i = 0;
722                 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
723                 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
724                         i++;
725                 if (i == MAX_ORDER_NR_PAGES)
726                         continue;
727                 page = pfn_to_page(pfn + i);
728                 if (zone && page_zone(page) != zone)
729                         return 0;
730                 zone = page_zone(page);
731         }
732         return 1;
733 }
734
735 /*
736  * Scanning pfn is much easier than scanning lru list.
737  * Scan pfn from start to end and Find LRU page.
738  */
739 static unsigned long scan_lru_pages(unsigned long start, unsigned long end)
740 {
741         unsigned long pfn;
742         struct page *page;
743         for (pfn = start; pfn < end; pfn++) {
744                 if (pfn_valid(pfn)) {
745                         page = pfn_to_page(pfn);
746                         if (PageLRU(page))
747                                 return pfn;
748                 }
749         }
750         return 0;
751 }
752
753 static struct page *
754 hotremove_migrate_alloc(struct page *page, unsigned long private, int **x)
755 {
756         /* This should be improooooved!! */
757         return alloc_page(GFP_HIGHUSER_MOVABLE);
758 }
759
760 #define NR_OFFLINE_AT_ONCE_PAGES        (256)
761 static int
762 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
763 {
764         unsigned long pfn;
765         struct page *page;
766         int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
767         int not_managed = 0;
768         int ret = 0;
769         LIST_HEAD(source);
770
771         for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
772                 if (!pfn_valid(pfn))
773                         continue;
774                 page = pfn_to_page(pfn);
775                 if (!get_page_unless_zero(page))
776                         continue;
777                 /*
778                  * We can skip free pages. And we can only deal with pages on
779                  * LRU.
780                  */
781                 ret = isolate_lru_page(page);
782                 if (!ret) { /* Success */
783                         put_page(page);
784                         list_add_tail(&page->lru, &source);
785                         move_pages--;
786                         inc_zone_page_state(page, NR_ISOLATED_ANON +
787                                             page_is_file_cache(page));
788
789                 } else {
790 #ifdef CONFIG_DEBUG_VM
791                         printk(KERN_ALERT "removing pfn %lx from LRU failed\n",
792                                pfn);
793                         dump_page(page);
794 #endif
795                         put_page(page);
796                         /* Because we don't have big zone->lock. we should
797                            check this again here. */
798                         if (page_count(page)) {
799                                 not_managed++;
800                                 ret = -EBUSY;
801                                 break;
802                         }
803                 }
804         }
805         if (!list_empty(&source)) {
806                 if (not_managed) {
807                         putback_lru_pages(&source);
808                         goto out;
809                 }
810                 /* this function returns # of failed pages */
811                 ret = migrate_pages(&source, hotremove_migrate_alloc, 0,
812                                                                 true, true);
813                 if (ret)
814                         putback_lru_pages(&source);
815         }
816 out:
817         return ret;
818 }
819
820 /*
821  * remove from free_area[] and mark all as Reserved.
822  */
823 static int
824 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
825                         void *data)
826 {
827         __offline_isolated_pages(start, start + nr_pages);
828         return 0;
829 }
830
831 static void
832 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
833 {
834         walk_system_ram_range(start_pfn, end_pfn - start_pfn, NULL,
835                                 offline_isolated_pages_cb);
836 }
837
838 /*
839  * Check all pages in range, recoreded as memory resource, are isolated.
840  */
841 static int
842 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
843                         void *data)
844 {
845         int ret;
846         long offlined = *(long *)data;
847         ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
848         offlined = nr_pages;
849         if (!ret)
850                 *(long *)data += offlined;
851         return ret;
852 }
853
854 static long
855 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
856 {
857         long offlined = 0;
858         int ret;
859
860         ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, &offlined,
861                         check_pages_isolated_cb);
862         if (ret < 0)
863                 offlined = (long)ret;
864         return offlined;
865 }
866
867 static int __ref offline_pages(unsigned long start_pfn,
868                   unsigned long end_pfn, unsigned long timeout)
869 {
870         unsigned long pfn, nr_pages, expire;
871         long offlined_pages;
872         int ret, drain, retry_max, node;
873         struct zone *zone;
874         struct memory_notify arg;
875
876         BUG_ON(start_pfn >= end_pfn);
877         /* at least, alignment against pageblock is necessary */
878         if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
879                 return -EINVAL;
880         if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
881                 return -EINVAL;
882         /* This makes hotplug much easier...and readable.
883            we assume this for now. .*/
884         if (!test_pages_in_a_zone(start_pfn, end_pfn))
885                 return -EINVAL;
886
887         lock_memory_hotplug();
888
889         zone = page_zone(pfn_to_page(start_pfn));
890         node = zone_to_nid(zone);
891         nr_pages = end_pfn - start_pfn;
892
893         /* set above range as isolated */
894         ret = start_isolate_page_range(start_pfn, end_pfn);
895         if (ret)
896                 goto out;
897
898         arg.start_pfn = start_pfn;
899         arg.nr_pages = nr_pages;
900         arg.status_change_nid = -1;
901         if (nr_pages >= node_present_pages(node))
902                 arg.status_change_nid = node;
903
904         ret = memory_notify(MEM_GOING_OFFLINE, &arg);
905         ret = notifier_to_errno(ret);
906         if (ret)
907                 goto failed_removal;
908
909         pfn = start_pfn;
910         expire = jiffies + timeout;
911         drain = 0;
912         retry_max = 5;
913 repeat:
914         /* start memory hot removal */
915         ret = -EAGAIN;
916         if (time_after(jiffies, expire))
917                 goto failed_removal;
918         ret = -EINTR;
919         if (signal_pending(current))
920                 goto failed_removal;
921         ret = 0;
922         if (drain) {
923                 lru_add_drain_all();
924                 cond_resched();
925                 drain_all_pages();
926         }
927
928         pfn = scan_lru_pages(start_pfn, end_pfn);
929         if (pfn) { /* We have page on LRU */
930                 ret = do_migrate_range(pfn, end_pfn);
931                 if (!ret) {
932                         drain = 1;
933                         goto repeat;
934                 } else {
935                         if (ret < 0)
936                                 if (--retry_max == 0)
937                                         goto failed_removal;
938                         yield();
939                         drain = 1;
940                         goto repeat;
941                 }
942         }
943         /* drain all zone's lru pagevec, this is asyncronous... */
944         lru_add_drain_all();
945         yield();
946         /* drain pcp pages , this is synchrouns. */
947         drain_all_pages();
948         /* check again */
949         offlined_pages = check_pages_isolated(start_pfn, end_pfn);
950         if (offlined_pages < 0) {
951                 ret = -EBUSY;
952                 goto failed_removal;
953         }
954         printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
955         /* Ok, all of our target is islaoted.
956            We cannot do rollback at this point. */
957         offline_isolated_pages(start_pfn, end_pfn);
958         /* reset pagetype flags and makes migrate type to be MOVABLE */
959         undo_isolate_page_range(start_pfn, end_pfn);
960         /* removal success */
961         zone->present_pages -= offlined_pages;
962         zone->zone_pgdat->node_present_pages -= offlined_pages;
963         totalram_pages -= offlined_pages;
964
965         init_per_zone_wmark_min();
966
967         if (!node_present_pages(node)) {
968                 node_clear_state(node, N_HIGH_MEMORY);
969                 kswapd_stop(node);
970         }
971
972         vm_total_pages = nr_free_pagecache_pages();
973         writeback_set_ratelimit();
974
975         memory_notify(MEM_OFFLINE, &arg);
976         unlock_memory_hotplug();
977         return 0;
978
979 failed_removal:
980         printk(KERN_INFO "memory offlining %lx to %lx failed\n",
981                 start_pfn, end_pfn);
982         memory_notify(MEM_CANCEL_OFFLINE, &arg);
983         /* pushback to free area */
984         undo_isolate_page_range(start_pfn, end_pfn);
985
986 out:
987         unlock_memory_hotplug();
988         return ret;
989 }
990
991 int remove_memory(u64 start, u64 size)
992 {
993         unsigned long start_pfn, end_pfn;
994
995         start_pfn = PFN_DOWN(start);
996         end_pfn = start_pfn + PFN_DOWN(size);
997         return offline_pages(start_pfn, end_pfn, 120 * HZ);
998 }
999 #else
1000 int remove_memory(u64 start, u64 size)
1001 {
1002         return -EINVAL;
1003 }
1004 #endif /* CONFIG_MEMORY_HOTREMOVE */
1005 EXPORT_SYMBOL_GPL(remove_memory);