mm: bugfix: set current->reclaim_state to NULL while returning from kswapd()
[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/export.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         section_nr = pfn_to_section_nr(start_pfn);
131         ms = __nr_to_section(section_nr);
132
133         /* Get section's memmap address */
134         memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
135
136         /*
137          * Get page for the memmap's phys address
138          * XXX: need more consideration for sparse_vmemmap...
139          */
140         page = virt_to_page(memmap);
141         mapsize = sizeof(struct page) * PAGES_PER_SECTION;
142         mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT;
143
144         /* remember memmap's page */
145         for (i = 0; i < mapsize; i++, page++)
146                 get_page_bootmem(section_nr, page, SECTION_INFO);
147
148         usemap = __nr_to_section(section_nr)->pageblock_flags;
149         page = virt_to_page(usemap);
150
151         mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT;
152
153         for (i = 0; i < mapsize; i++, page++)
154                 get_page_bootmem(section_nr, page, MIX_SECTION_INFO);
155
156 }
157
158 void register_page_bootmem_info_node(struct pglist_data *pgdat)
159 {
160         unsigned long i, pfn, end_pfn, nr_pages;
161         int node = pgdat->node_id;
162         struct page *page;
163         struct zone *zone;
164
165         nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT;
166         page = virt_to_page(pgdat);
167
168         for (i = 0; i < nr_pages; i++, page++)
169                 get_page_bootmem(node, page, NODE_INFO);
170
171         zone = &pgdat->node_zones[0];
172         for (; zone < pgdat->node_zones + MAX_NR_ZONES - 1; zone++) {
173                 if (zone->wait_table) {
174                         nr_pages = zone->wait_table_hash_nr_entries
175                                 * sizeof(wait_queue_head_t);
176                         nr_pages = PAGE_ALIGN(nr_pages) >> PAGE_SHIFT;
177                         page = virt_to_page(zone->wait_table);
178
179                         for (i = 0; i < nr_pages; i++, page++)
180                                 get_page_bootmem(node, page, NODE_INFO);
181                 }
182         }
183
184         pfn = pgdat->node_start_pfn;
185         end_pfn = pfn + pgdat->node_spanned_pages;
186
187         /* register_section info */
188         for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
189                 /*
190                  * Some platforms can assign the same pfn to multiple nodes - on
191                  * node0 as well as nodeN.  To avoid registering a pfn against
192                  * multiple nodes we check that this pfn does not already
193                  * reside in some other node.
194                  */
195                 if (pfn_valid(pfn) && (pfn_to_nid(pfn) == node))
196                         register_page_bootmem_info_section(pfn);
197         }
198 }
199 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
200
201 static void grow_zone_span(struct zone *zone, unsigned long start_pfn,
202                            unsigned long end_pfn)
203 {
204         unsigned long old_zone_end_pfn;
205
206         zone_span_writelock(zone);
207
208         old_zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
209         if (start_pfn < zone->zone_start_pfn)
210                 zone->zone_start_pfn = start_pfn;
211
212         zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
213                                 zone->zone_start_pfn;
214
215         zone_span_writeunlock(zone);
216 }
217
218 static void grow_pgdat_span(struct pglist_data *pgdat, unsigned long start_pfn,
219                             unsigned long end_pfn)
220 {
221         unsigned long old_pgdat_end_pfn =
222                 pgdat->node_start_pfn + pgdat->node_spanned_pages;
223
224         if (start_pfn < pgdat->node_start_pfn)
225                 pgdat->node_start_pfn = start_pfn;
226
227         pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
228                                         pgdat->node_start_pfn;
229 }
230
231 static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
232 {
233         struct pglist_data *pgdat = zone->zone_pgdat;
234         int nr_pages = PAGES_PER_SECTION;
235         int nid = pgdat->node_id;
236         int zone_type;
237         unsigned long flags;
238
239         zone_type = zone - pgdat->node_zones;
240         if (!zone->wait_table) {
241                 int ret;
242
243                 ret = init_currently_empty_zone(zone, phys_start_pfn,
244                                                 nr_pages, MEMMAP_HOTPLUG);
245                 if (ret)
246                         return ret;
247         }
248         pgdat_resize_lock(zone->zone_pgdat, &flags);
249         grow_zone_span(zone, phys_start_pfn, phys_start_pfn + nr_pages);
250         grow_pgdat_span(zone->zone_pgdat, phys_start_pfn,
251                         phys_start_pfn + nr_pages);
252         pgdat_resize_unlock(zone->zone_pgdat, &flags);
253         memmap_init_zone(nr_pages, nid, zone_type,
254                          phys_start_pfn, MEMMAP_HOTPLUG);
255         return 0;
256 }
257
258 static int __meminit __add_section(int nid, struct zone *zone,
259                                         unsigned long phys_start_pfn)
260 {
261         int nr_pages = PAGES_PER_SECTION;
262         int ret;
263
264         if (pfn_valid(phys_start_pfn))
265                 return -EEXIST;
266
267         ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
268
269         if (ret < 0)
270                 return ret;
271
272         ret = __add_zone(zone, phys_start_pfn);
273
274         if (ret < 0)
275                 return ret;
276
277         return register_new_memory(nid, __pfn_to_section(phys_start_pfn));
278 }
279
280 #ifdef CONFIG_SPARSEMEM_VMEMMAP
281 static int __remove_section(struct zone *zone, struct mem_section *ms)
282 {
283         /*
284          * XXX: Freeing memmap with vmemmap is not implement yet.
285          *      This should be removed later.
286          */
287         return -EBUSY;
288 }
289 #else
290 static int __remove_section(struct zone *zone, struct mem_section *ms)
291 {
292         unsigned long flags;
293         struct pglist_data *pgdat = zone->zone_pgdat;
294         int ret = -EINVAL;
295
296         if (!valid_section(ms))
297                 return ret;
298
299         ret = unregister_memory_section(ms);
300         if (ret)
301                 return ret;
302
303         pgdat_resize_lock(pgdat, &flags);
304         sparse_remove_one_section(zone, ms);
305         pgdat_resize_unlock(pgdat, &flags);
306         return 0;
307 }
308 #endif
309
310 /*
311  * Reasonably generic function for adding memory.  It is
312  * expected that archs that support memory hotplug will
313  * call this function after deciding the zone to which to
314  * add the new pages.
315  */
316 int __ref __add_pages(int nid, struct zone *zone, unsigned long phys_start_pfn,
317                         unsigned long nr_pages)
318 {
319         unsigned long i;
320         int err = 0;
321         int start_sec, end_sec;
322         /* during initialize mem_map, align hot-added range to section */
323         start_sec = pfn_to_section_nr(phys_start_pfn);
324         end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
325
326         for (i = start_sec; i <= end_sec; i++) {
327                 err = __add_section(nid, zone, i << PFN_SECTION_SHIFT);
328
329                 /*
330                  * EEXIST is finally dealt with by ioresource collision
331                  * check. see add_memory() => register_memory_resource()
332                  * Warning will be printed if there is collision.
333                  */
334                 if (err && (err != -EEXIST))
335                         break;
336                 err = 0;
337         }
338
339         return err;
340 }
341 EXPORT_SYMBOL_GPL(__add_pages);
342
343 /**
344  * __remove_pages() - remove sections of pages from a zone
345  * @zone: zone from which pages need to be removed
346  * @phys_start_pfn: starting pageframe (must be aligned to start of a section)
347  * @nr_pages: number of pages to remove (must be multiple of section size)
348  *
349  * Generic helper function to remove section mappings and sysfs entries
350  * for the section of the memory we are removing. Caller needs to make
351  * sure that pages are marked reserved and zones are adjust properly by
352  * calling offline_pages().
353  */
354 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn,
355                  unsigned long nr_pages)
356 {
357         unsigned long i, ret = 0;
358         int sections_to_remove;
359
360         /*
361          * We can only remove entire sections
362          */
363         BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK);
364         BUG_ON(nr_pages % PAGES_PER_SECTION);
365
366         sections_to_remove = nr_pages / PAGES_PER_SECTION;
367         for (i = 0; i < sections_to_remove; i++) {
368                 unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION;
369                 release_mem_region(pfn << PAGE_SHIFT,
370                                    PAGES_PER_SECTION << PAGE_SHIFT);
371                 ret = __remove_section(zone, __pfn_to_section(pfn));
372                 if (ret)
373                         break;
374         }
375         return ret;
376 }
377 EXPORT_SYMBOL_GPL(__remove_pages);
378
379 int set_online_page_callback(online_page_callback_t callback)
380 {
381         int rc = -EINVAL;
382
383         lock_memory_hotplug();
384
385         if (online_page_callback == generic_online_page) {
386                 online_page_callback = callback;
387                 rc = 0;
388         }
389
390         unlock_memory_hotplug();
391
392         return rc;
393 }
394 EXPORT_SYMBOL_GPL(set_online_page_callback);
395
396 int restore_online_page_callback(online_page_callback_t callback)
397 {
398         int rc = -EINVAL;
399
400         lock_memory_hotplug();
401
402         if (online_page_callback == callback) {
403                 online_page_callback = generic_online_page;
404                 rc = 0;
405         }
406
407         unlock_memory_hotplug();
408
409         return rc;
410 }
411 EXPORT_SYMBOL_GPL(restore_online_page_callback);
412
413 void __online_page_set_limits(struct page *page)
414 {
415         unsigned long pfn = page_to_pfn(page);
416
417         if (pfn >= num_physpages)
418                 num_physpages = pfn + 1;
419 }
420 EXPORT_SYMBOL_GPL(__online_page_set_limits);
421
422 void __online_page_increment_counters(struct page *page)
423 {
424         totalram_pages++;
425
426 #ifdef CONFIG_HIGHMEM
427         if (PageHighMem(page))
428                 totalhigh_pages++;
429 #endif
430 }
431 EXPORT_SYMBOL_GPL(__online_page_increment_counters);
432
433 void __online_page_free(struct page *page)
434 {
435         ClearPageReserved(page);
436         init_page_count(page);
437         __free_page(page);
438 }
439 EXPORT_SYMBOL_GPL(__online_page_free);
440
441 static void generic_online_page(struct page *page)
442 {
443         __online_page_set_limits(page);
444         __online_page_increment_counters(page);
445         __online_page_free(page);
446 }
447
448 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
449                         void *arg)
450 {
451         unsigned long i;
452         unsigned long onlined_pages = *(unsigned long *)arg;
453         struct page *page;
454         if (PageReserved(pfn_to_page(start_pfn)))
455                 for (i = 0; i < nr_pages; i++) {
456                         page = pfn_to_page(start_pfn + i);
457                         (*online_page_callback)(page);
458                         onlined_pages++;
459                 }
460         *(unsigned long *)arg = onlined_pages;
461         return 0;
462 }
463
464
465 int __ref online_pages(unsigned long pfn, unsigned long nr_pages)
466 {
467         unsigned long onlined_pages = 0;
468         struct zone *zone;
469         int need_zonelists_rebuild = 0;
470         int nid;
471         int ret;
472         struct memory_notify arg;
473
474         lock_memory_hotplug();
475         arg.start_pfn = pfn;
476         arg.nr_pages = nr_pages;
477         arg.status_change_nid = -1;
478
479         nid = page_to_nid(pfn_to_page(pfn));
480         if (node_present_pages(nid) == 0)
481                 arg.status_change_nid = nid;
482
483         ret = memory_notify(MEM_GOING_ONLINE, &arg);
484         ret = notifier_to_errno(ret);
485         if (ret) {
486                 memory_notify(MEM_CANCEL_ONLINE, &arg);
487                 unlock_memory_hotplug();
488                 return ret;
489         }
490         /*
491          * This doesn't need a lock to do pfn_to_page().
492          * The section can't be removed here because of the
493          * memory_block->state_mutex.
494          */
495         zone = page_zone(pfn_to_page(pfn));
496         /*
497          * If this zone is not populated, then it is not in zonelist.
498          * This means the page allocator ignores this zone.
499          * So, zonelist must be updated after online.
500          */
501         mutex_lock(&zonelists_mutex);
502         if (!populated_zone(zone))
503                 need_zonelists_rebuild = 1;
504
505         ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages,
506                 online_pages_range);
507         if (ret) {
508                 mutex_unlock(&zonelists_mutex);
509                 printk(KERN_DEBUG "online_pages %lx at %lx failed\n",
510                         nr_pages, pfn);
511                 memory_notify(MEM_CANCEL_ONLINE, &arg);
512                 unlock_memory_hotplug();
513                 return ret;
514         }
515
516         zone->present_pages += onlined_pages;
517         zone->zone_pgdat->node_present_pages += onlined_pages;
518         if (need_zonelists_rebuild)
519                 build_all_zonelists(zone);
520         else
521                 zone_pcp_update(zone);
522
523         mutex_unlock(&zonelists_mutex);
524
525         init_per_zone_wmark_min();
526
527         if (onlined_pages) {
528                 kswapd_run(zone_to_nid(zone));
529                 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
530         }
531
532         vm_total_pages = nr_free_pagecache_pages();
533
534         writeback_set_ratelimit();
535
536         if (onlined_pages)
537                 memory_notify(MEM_ONLINE, &arg);
538         unlock_memory_hotplug();
539
540         return 0;
541 }
542 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
543
544 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
545 static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
546 {
547         struct pglist_data *pgdat;
548         unsigned long zones_size[MAX_NR_ZONES] = {0};
549         unsigned long zholes_size[MAX_NR_ZONES] = {0};
550         unsigned long start_pfn = start >> PAGE_SHIFT;
551
552         pgdat = arch_alloc_nodedata(nid);
553         if (!pgdat)
554                 return NULL;
555
556         arch_refresh_nodedata(nid, pgdat);
557
558         /* we can use NODE_DATA(nid) from here */
559
560         /* init node's zones as empty zones, we don't have any present pages.*/
561         free_area_init_node(nid, zones_size, start_pfn, zholes_size);
562
563         /*
564          * The node we allocated has no zone fallback lists. For avoiding
565          * to access not-initialized zonelist, build here.
566          */
567         mutex_lock(&zonelists_mutex);
568         build_all_zonelists(NULL);
569         mutex_unlock(&zonelists_mutex);
570
571         return pgdat;
572 }
573
574 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
575 {
576         arch_refresh_nodedata(nid, NULL);
577         arch_free_nodedata(pgdat);
578         return;
579 }
580
581
582 /*
583  * called by cpu_up() to online a node without onlined memory.
584  */
585 int mem_online_node(int nid)
586 {
587         pg_data_t       *pgdat;
588         int     ret;
589
590         lock_memory_hotplug();
591         pgdat = hotadd_new_pgdat(nid, 0);
592         if (!pgdat) {
593                 ret = -ENOMEM;
594                 goto out;
595         }
596         node_set_online(nid);
597         ret = register_one_node(nid);
598         BUG_ON(ret);
599
600 out:
601         unlock_memory_hotplug();
602         return ret;
603 }
604
605 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
606 int __ref add_memory(int nid, u64 start, u64 size)
607 {
608         pg_data_t *pgdat = NULL;
609         int new_pgdat = 0;
610         struct resource *res;
611         int ret;
612
613         lock_memory_hotplug();
614
615         res = register_memory_resource(start, size);
616         ret = -EEXIST;
617         if (!res)
618                 goto out;
619
620         if (!node_online(nid)) {
621                 pgdat = hotadd_new_pgdat(nid, start);
622                 ret = -ENOMEM;
623                 if (!pgdat)
624                         goto out;
625                 new_pgdat = 1;
626         }
627
628         /* call arch's memory hotadd */
629         ret = arch_add_memory(nid, start, size);
630
631         if (ret < 0)
632                 goto error;
633
634         /* we online node here. we can't roll back from here. */
635         node_set_online(nid);
636
637         if (new_pgdat) {
638                 ret = register_one_node(nid);
639                 /*
640                  * If sysfs file of new node can't create, cpu on the node
641                  * can't be hot-added. There is no rollback way now.
642                  * So, check by BUG_ON() to catch it reluctantly..
643                  */
644                 BUG_ON(ret);
645         }
646
647         /* create new memmap entry */
648         firmware_map_add_hotplug(start, start + size, "System RAM");
649
650         goto out;
651
652 error:
653         /* rollback pgdat allocation and others */
654         if (new_pgdat)
655                 rollback_node_hotadd(nid, pgdat);
656         if (res)
657                 release_memory_resource(res);
658
659 out:
660         unlock_memory_hotplug();
661         return ret;
662 }
663 EXPORT_SYMBOL_GPL(add_memory);
664
665 #ifdef CONFIG_MEMORY_HOTREMOVE
666 /*
667  * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy
668  * set and the size of the free page is given by page_order(). Using this,
669  * the function determines if the pageblock contains only free pages.
670  * Due to buddy contraints, a free page at least the size of a pageblock will
671  * be located at the start of the pageblock
672  */
673 static inline int pageblock_free(struct page *page)
674 {
675         return PageBuddy(page) && page_order(page) >= pageblock_order;
676 }
677
678 /* Return the start of the next active pageblock after a given page */
679 static struct page *next_active_pageblock(struct page *page)
680 {
681         /* Ensure the starting page is pageblock-aligned */
682         BUG_ON(page_to_pfn(page) & (pageblock_nr_pages - 1));
683
684         /* If the entire pageblock is free, move to the end of free page */
685         if (pageblock_free(page)) {
686                 int order;
687                 /* be careful. we don't have locks, page_order can be changed.*/
688                 order = page_order(page);
689                 if ((order < MAX_ORDER) && (order >= pageblock_order))
690                         return page + (1 << order);
691         }
692
693         return page + pageblock_nr_pages;
694 }
695
696 /* Checks if this range of memory is likely to be hot-removable. */
697 int is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages)
698 {
699         struct page *page = pfn_to_page(start_pfn);
700         struct page *end_page = page + nr_pages;
701
702         /* Check the starting page of each pageblock within the range */
703         for (; page < end_page; page = next_active_pageblock(page)) {
704                 if (!is_pageblock_removable_nolock(page))
705                         return 0;
706                 cond_resched();
707         }
708
709         /* All pageblocks in the memory block are likely to be hot-removable */
710         return 1;
711 }
712
713 /*
714  * Confirm all pages in a range [start, end) is belongs to the same zone.
715  */
716 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
717 {
718         unsigned long pfn;
719         struct zone *zone = NULL;
720         struct page *page;
721         int i;
722         for (pfn = start_pfn;
723              pfn < end_pfn;
724              pfn += MAX_ORDER_NR_PAGES) {
725                 i = 0;
726                 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
727                 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
728                         i++;
729                 if (i == MAX_ORDER_NR_PAGES)
730                         continue;
731                 page = pfn_to_page(pfn + i);
732                 if (zone && page_zone(page) != zone)
733                         return 0;
734                 zone = page_zone(page);
735         }
736         return 1;
737 }
738
739 /*
740  * Scanning pfn is much easier than scanning lru list.
741  * Scan pfn from start to end and Find LRU page.
742  */
743 static unsigned long scan_lru_pages(unsigned long start, unsigned long end)
744 {
745         unsigned long pfn;
746         struct page *page;
747         for (pfn = start; pfn < end; pfn++) {
748                 if (pfn_valid(pfn)) {
749                         page = pfn_to_page(pfn);
750                         if (PageLRU(page))
751                                 return pfn;
752                 }
753         }
754         return 0;
755 }
756
757 static struct page *
758 hotremove_migrate_alloc(struct page *page, unsigned long private, int **x)
759 {
760         /* This should be improooooved!! */
761         return alloc_page(GFP_HIGHUSER_MOVABLE);
762 }
763
764 #define NR_OFFLINE_AT_ONCE_PAGES        (256)
765 static int
766 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
767 {
768         unsigned long pfn;
769         struct page *page;
770         int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
771         int not_managed = 0;
772         int ret = 0;
773         LIST_HEAD(source);
774
775         for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
776                 if (!pfn_valid(pfn))
777                         continue;
778                 page = pfn_to_page(pfn);
779                 if (!get_page_unless_zero(page))
780                         continue;
781                 /*
782                  * We can skip free pages. And we can only deal with pages on
783                  * LRU.
784                  */
785                 ret = isolate_lru_page(page);
786                 if (!ret) { /* Success */
787                         put_page(page);
788                         list_add_tail(&page->lru, &source);
789                         move_pages--;
790                         inc_zone_page_state(page, NR_ISOLATED_ANON +
791                                             page_is_file_cache(page));
792
793                 } else {
794 #ifdef CONFIG_DEBUG_VM
795                         printk(KERN_ALERT "removing pfn %lx from LRU failed\n",
796                                pfn);
797                         dump_page(page);
798 #endif
799                         put_page(page);
800                         /* Because we don't have big zone->lock. we should
801                            check this again here. */
802                         if (page_count(page)) {
803                                 not_managed++;
804                                 ret = -EBUSY;
805                                 break;
806                         }
807                 }
808         }
809         if (!list_empty(&source)) {
810                 if (not_managed) {
811                         putback_lru_pages(&source);
812                         goto out;
813                 }
814                 /* this function returns # of failed pages */
815                 ret = migrate_pages(&source, hotremove_migrate_alloc, 0,
816                                                         true, MIGRATE_SYNC);
817                 if (ret)
818                         putback_lru_pages(&source);
819         }
820 out:
821         return ret;
822 }
823
824 /*
825  * remove from free_area[] and mark all as Reserved.
826  */
827 static int
828 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
829                         void *data)
830 {
831         __offline_isolated_pages(start, start + nr_pages);
832         return 0;
833 }
834
835 static void
836 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
837 {
838         walk_system_ram_range(start_pfn, end_pfn - start_pfn, NULL,
839                                 offline_isolated_pages_cb);
840 }
841
842 /*
843  * Check all pages in range, recoreded as memory resource, are isolated.
844  */
845 static int
846 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
847                         void *data)
848 {
849         int ret;
850         long offlined = *(long *)data;
851         ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
852         offlined = nr_pages;
853         if (!ret)
854                 *(long *)data += offlined;
855         return ret;
856 }
857
858 static long
859 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
860 {
861         long offlined = 0;
862         int ret;
863
864         ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, &offlined,
865                         check_pages_isolated_cb);
866         if (ret < 0)
867                 offlined = (long)ret;
868         return offlined;
869 }
870
871 static int __ref offline_pages(unsigned long start_pfn,
872                   unsigned long end_pfn, unsigned long timeout)
873 {
874         unsigned long pfn, nr_pages, expire;
875         long offlined_pages;
876         int ret, drain, retry_max, node;
877         struct zone *zone;
878         struct memory_notify arg;
879
880         BUG_ON(start_pfn >= end_pfn);
881         /* at least, alignment against pageblock is necessary */
882         if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
883                 return -EINVAL;
884         if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
885                 return -EINVAL;
886         /* This makes hotplug much easier...and readable.
887            we assume this for now. .*/
888         if (!test_pages_in_a_zone(start_pfn, end_pfn))
889                 return -EINVAL;
890
891         lock_memory_hotplug();
892
893         zone = page_zone(pfn_to_page(start_pfn));
894         node = zone_to_nid(zone);
895         nr_pages = end_pfn - start_pfn;
896
897         /* set above range as isolated */
898         ret = start_isolate_page_range(start_pfn, end_pfn);
899         if (ret)
900                 goto out;
901
902         arg.start_pfn = start_pfn;
903         arg.nr_pages = nr_pages;
904         arg.status_change_nid = -1;
905         if (nr_pages >= node_present_pages(node))
906                 arg.status_change_nid = node;
907
908         ret = memory_notify(MEM_GOING_OFFLINE, &arg);
909         ret = notifier_to_errno(ret);
910         if (ret)
911                 goto failed_removal;
912
913         pfn = start_pfn;
914         expire = jiffies + timeout;
915         drain = 0;
916         retry_max = 5;
917 repeat:
918         /* start memory hot removal */
919         ret = -EAGAIN;
920         if (time_after(jiffies, expire))
921                 goto failed_removal;
922         ret = -EINTR;
923         if (signal_pending(current))
924                 goto failed_removal;
925         ret = 0;
926         if (drain) {
927                 lru_add_drain_all();
928                 cond_resched();
929                 drain_all_pages();
930         }
931
932         pfn = scan_lru_pages(start_pfn, end_pfn);
933         if (pfn) { /* We have page on LRU */
934                 ret = do_migrate_range(pfn, end_pfn);
935                 if (!ret) {
936                         drain = 1;
937                         goto repeat;
938                 } else {
939                         if (ret < 0)
940                                 if (--retry_max == 0)
941                                         goto failed_removal;
942                         yield();
943                         drain = 1;
944                         goto repeat;
945                 }
946         }
947         /* drain all zone's lru pagevec, this is asyncronous... */
948         lru_add_drain_all();
949         yield();
950         /* drain pcp pages , this is synchrouns. */
951         drain_all_pages();
952         /* check again */
953         offlined_pages = check_pages_isolated(start_pfn, end_pfn);
954         if (offlined_pages < 0) {
955                 ret = -EBUSY;
956                 goto failed_removal;
957         }
958         printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
959         /* Ok, all of our target is islaoted.
960            We cannot do rollback at this point. */
961         offline_isolated_pages(start_pfn, end_pfn);
962         /* reset pagetype flags and makes migrate type to be MOVABLE */
963         undo_isolate_page_range(start_pfn, end_pfn);
964         /* removal success */
965         zone->present_pages -= offlined_pages;
966         zone->zone_pgdat->node_present_pages -= offlined_pages;
967         totalram_pages -= offlined_pages;
968
969         init_per_zone_wmark_min();
970
971         if (!node_present_pages(node)) {
972                 node_clear_state(node, N_HIGH_MEMORY);
973                 kswapd_stop(node);
974         }
975
976         vm_total_pages = nr_free_pagecache_pages();
977         writeback_set_ratelimit();
978
979         memory_notify(MEM_OFFLINE, &arg);
980         unlock_memory_hotplug();
981         return 0;
982
983 failed_removal:
984         printk(KERN_INFO "memory offlining %lx to %lx failed\n",
985                 start_pfn, end_pfn);
986         memory_notify(MEM_CANCEL_OFFLINE, &arg);
987         /* pushback to free area */
988         undo_isolate_page_range(start_pfn, end_pfn);
989
990 out:
991         unlock_memory_hotplug();
992         return ret;
993 }
994
995 int remove_memory(u64 start, u64 size)
996 {
997         unsigned long start_pfn, end_pfn;
998
999         start_pfn = PFN_DOWN(start);
1000         end_pfn = start_pfn + PFN_DOWN(size);
1001         return offline_pages(start_pfn, end_pfn, 120 * HZ);
1002 }
1003 #else
1004 int remove_memory(u64 start, u64 size)
1005 {
1006         return -EINVAL;
1007 }
1008 #endif /* CONFIG_MEMORY_HOTREMOVE */
1009 EXPORT_SYMBOL_GPL(remove_memory);