memory unplug: page offline
[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/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 /* add this memory to iomem resource */
33 static struct resource *register_memory_resource(u64 start, u64 size)
34 {
35         struct resource *res;
36         res = kzalloc(sizeof(struct resource), GFP_KERNEL);
37         BUG_ON(!res);
38
39         res->name = "System RAM";
40         res->start = start;
41         res->end = start + size - 1;
42         res->flags = IORESOURCE_MEM;
43         if (request_resource(&iomem_resource, res) < 0) {
44                 printk("System RAM resource %llx - %llx cannot be added\n",
45                 (unsigned long long)res->start, (unsigned long long)res->end);
46                 kfree(res);
47                 res = NULL;
48         }
49         return res;
50 }
51
52 static void release_memory_resource(struct resource *res)
53 {
54         if (!res)
55                 return;
56         release_resource(res);
57         kfree(res);
58         return;
59 }
60
61
62 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
63 static int __add_zone(struct zone *zone, unsigned long phys_start_pfn)
64 {
65         struct pglist_data *pgdat = zone->zone_pgdat;
66         int nr_pages = PAGES_PER_SECTION;
67         int nid = pgdat->node_id;
68         int zone_type;
69
70         zone_type = zone - pgdat->node_zones;
71         if (!zone->wait_table) {
72                 int ret = 0;
73                 ret = init_currently_empty_zone(zone, phys_start_pfn,
74                                                 nr_pages, MEMMAP_HOTPLUG);
75                 if (ret < 0)
76                         return ret;
77         }
78         memmap_init_zone(nr_pages, nid, zone_type,
79                          phys_start_pfn, MEMMAP_HOTPLUG);
80         return 0;
81 }
82
83 static int __add_section(struct zone *zone, unsigned long phys_start_pfn)
84 {
85         int nr_pages = PAGES_PER_SECTION;
86         int ret;
87
88         if (pfn_valid(phys_start_pfn))
89                 return -EEXIST;
90
91         ret = sparse_add_one_section(zone, phys_start_pfn, nr_pages);
92
93         if (ret < 0)
94                 return ret;
95
96         ret = __add_zone(zone, phys_start_pfn);
97
98         if (ret < 0)
99                 return ret;
100
101         return register_new_memory(__pfn_to_section(phys_start_pfn));
102 }
103
104 /*
105  * Reasonably generic function for adding memory.  It is
106  * expected that archs that support memory hotplug will
107  * call this function after deciding the zone to which to
108  * add the new pages.
109  */
110 int __add_pages(struct zone *zone, unsigned long phys_start_pfn,
111                  unsigned long nr_pages)
112 {
113         unsigned long i;
114         int err = 0;
115         int start_sec, end_sec;
116         /* during initialize mem_map, align hot-added range to section */
117         start_sec = pfn_to_section_nr(phys_start_pfn);
118         end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1);
119
120         for (i = start_sec; i <= end_sec; i++) {
121                 err = __add_section(zone, i << PFN_SECTION_SHIFT);
122
123                 /*
124                  * EEXIST is finally dealed with by ioresource collision
125                  * check. see add_memory() => register_memory_resource()
126                  * Warning will be printed if there is collision.
127                  */
128                 if (err && (err != -EEXIST))
129                         break;
130                 err = 0;
131         }
132
133         return err;
134 }
135 EXPORT_SYMBOL_GPL(__add_pages);
136
137 static void grow_zone_span(struct zone *zone,
138                 unsigned long start_pfn, unsigned long end_pfn)
139 {
140         unsigned long old_zone_end_pfn;
141
142         zone_span_writelock(zone);
143
144         old_zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
145         if (start_pfn < zone->zone_start_pfn)
146                 zone->zone_start_pfn = start_pfn;
147
148         zone->spanned_pages = max(old_zone_end_pfn, end_pfn) -
149                                 zone->zone_start_pfn;
150
151         zone_span_writeunlock(zone);
152 }
153
154 static void grow_pgdat_span(struct pglist_data *pgdat,
155                 unsigned long start_pfn, unsigned long end_pfn)
156 {
157         unsigned long old_pgdat_end_pfn =
158                 pgdat->node_start_pfn + pgdat->node_spanned_pages;
159
160         if (start_pfn < pgdat->node_start_pfn)
161                 pgdat->node_start_pfn = start_pfn;
162
163         pgdat->node_spanned_pages = max(old_pgdat_end_pfn, end_pfn) -
164                                         pgdat->node_start_pfn;
165 }
166
167 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages,
168                         void *arg)
169 {
170         unsigned long i;
171         unsigned long onlined_pages = *(unsigned long *)arg;
172         struct page *page;
173         if (PageReserved(pfn_to_page(start_pfn)))
174                 for (i = 0; i < nr_pages; i++) {
175                         page = pfn_to_page(start_pfn + i);
176                         online_page(page);
177                         onlined_pages++;
178                 }
179         *(unsigned long *)arg = onlined_pages;
180         return 0;
181 }
182
183
184 int online_pages(unsigned long pfn, unsigned long nr_pages)
185 {
186         unsigned long flags;
187         unsigned long onlined_pages = 0;
188         struct zone *zone;
189         int need_zonelists_rebuild = 0;
190
191         /*
192          * This doesn't need a lock to do pfn_to_page().
193          * The section can't be removed here because of the
194          * memory_block->state_sem.
195          */
196         zone = page_zone(pfn_to_page(pfn));
197         pgdat_resize_lock(zone->zone_pgdat, &flags);
198         grow_zone_span(zone, pfn, pfn + nr_pages);
199         grow_pgdat_span(zone->zone_pgdat, pfn, pfn + nr_pages);
200         pgdat_resize_unlock(zone->zone_pgdat, &flags);
201
202         /*
203          * If this zone is not populated, then it is not in zonelist.
204          * This means the page allocator ignores this zone.
205          * So, zonelist must be updated after online.
206          */
207         if (!populated_zone(zone))
208                 need_zonelists_rebuild = 1;
209
210         walk_memory_resource(pfn, nr_pages, &onlined_pages,
211                 online_pages_range);
212         zone->present_pages += onlined_pages;
213         zone->zone_pgdat->node_present_pages += onlined_pages;
214
215         setup_per_zone_pages_min();
216         if (onlined_pages) {
217                 kswapd_run(zone_to_nid(zone));
218                 node_set_state(zone_to_nid(zone), N_HIGH_MEMORY);
219         }
220
221         if (need_zonelists_rebuild)
222                 build_all_zonelists();
223         vm_total_pages = nr_free_pagecache_pages();
224         writeback_set_ratelimit();
225         return 0;
226 }
227 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
228
229 static pg_data_t *hotadd_new_pgdat(int nid, u64 start)
230 {
231         struct pglist_data *pgdat;
232         unsigned long zones_size[MAX_NR_ZONES] = {0};
233         unsigned long zholes_size[MAX_NR_ZONES] = {0};
234         unsigned long start_pfn = start >> PAGE_SHIFT;
235
236         pgdat = arch_alloc_nodedata(nid);
237         if (!pgdat)
238                 return NULL;
239
240         arch_refresh_nodedata(nid, pgdat);
241
242         /* we can use NODE_DATA(nid) from here */
243
244         /* init node's zones as empty zones, we don't have any present pages.*/
245         free_area_init_node(nid, pgdat, zones_size, start_pfn, zholes_size);
246
247         return pgdat;
248 }
249
250 static void rollback_node_hotadd(int nid, pg_data_t *pgdat)
251 {
252         arch_refresh_nodedata(nid, NULL);
253         arch_free_nodedata(pgdat);
254         return;
255 }
256
257
258 int add_memory(int nid, u64 start, u64 size)
259 {
260         pg_data_t *pgdat = NULL;
261         int new_pgdat = 0;
262         struct resource *res;
263         int ret;
264
265         res = register_memory_resource(start, size);
266         if (!res)
267                 return -EEXIST;
268
269         if (!node_online(nid)) {
270                 pgdat = hotadd_new_pgdat(nid, start);
271                 if (!pgdat)
272                         return -ENOMEM;
273                 new_pgdat = 1;
274         }
275
276         /* call arch's memory hotadd */
277         ret = arch_add_memory(nid, start, size);
278
279         if (ret < 0)
280                 goto error;
281
282         /* we online node here. we can't roll back from here. */
283         node_set_online(nid);
284
285         cpuset_track_online_nodes();
286
287         if (new_pgdat) {
288                 ret = register_one_node(nid);
289                 /*
290                  * If sysfs file of new node can't create, cpu on the node
291                  * can't be hot-added. There is no rollback way now.
292                  * So, check by BUG_ON() to catch it reluctantly..
293                  */
294                 BUG_ON(ret);
295         }
296
297         return ret;
298 error:
299         /* rollback pgdat allocation and others */
300         if (new_pgdat)
301                 rollback_node_hotadd(nid, pgdat);
302         if (res)
303                 release_memory_resource(res);
304
305         return ret;
306 }
307 EXPORT_SYMBOL_GPL(add_memory);
308
309 #ifdef CONFIG_MEMORY_HOTREMOVE
310 /*
311  * Confirm all pages in a range [start, end) is belongs to the same zone.
312  */
313 static int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn)
314 {
315         unsigned long pfn;
316         struct zone *zone = NULL;
317         struct page *page;
318         int i;
319         for (pfn = start_pfn;
320              pfn < end_pfn;
321              pfn += MAX_ORDER_NR_PAGES) {
322                 i = 0;
323                 /* This is just a CONFIG_HOLES_IN_ZONE check.*/
324                 while ((i < MAX_ORDER_NR_PAGES) && !pfn_valid_within(pfn + i))
325                         i++;
326                 if (i == MAX_ORDER_NR_PAGES)
327                         continue;
328                 page = pfn_to_page(pfn + i);
329                 if (zone && page_zone(page) != zone)
330                         return 0;
331                 zone = page_zone(page);
332         }
333         return 1;
334 }
335
336 /*
337  * Scanning pfn is much easier than scanning lru list.
338  * Scan pfn from start to end and Find LRU page.
339  */
340 int scan_lru_pages(unsigned long start, unsigned long end)
341 {
342         unsigned long pfn;
343         struct page *page;
344         for (pfn = start; pfn < end; pfn++) {
345                 if (pfn_valid(pfn)) {
346                         page = pfn_to_page(pfn);
347                         if (PageLRU(page))
348                                 return pfn;
349                 }
350         }
351         return 0;
352 }
353
354 static struct page *
355 hotremove_migrate_alloc(struct page *page,
356                         unsigned long private,
357                         int **x)
358 {
359         /* This should be improoooooved!! */
360         return alloc_page(GFP_HIGHUSER_PAGECACHE);
361 }
362
363
364 #define NR_OFFLINE_AT_ONCE_PAGES        (256)
365 static int
366 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
367 {
368         unsigned long pfn;
369         struct page *page;
370         int move_pages = NR_OFFLINE_AT_ONCE_PAGES;
371         int not_managed = 0;
372         int ret = 0;
373         LIST_HEAD(source);
374
375         for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) {
376                 if (!pfn_valid(pfn))
377                         continue;
378                 page = pfn_to_page(pfn);
379                 if (!page_count(page))
380                         continue;
381                 /*
382                  * We can skip free pages. And we can only deal with pages on
383                  * LRU.
384                  */
385                 ret = isolate_lru_page(page, &source);
386                 if (!ret) { /* Success */
387                         move_pages--;
388                 } else {
389                         /* Becasue we don't have big zone->lock. we should
390                            check this again here. */
391                         if (page_count(page))
392                                 not_managed++;
393 #ifdef CONFIG_DEBUG_VM
394                         printk(KERN_INFO "removing from LRU failed"
395                                          " %lx/%d/%lx\n",
396                                 pfn, page_count(page), page->flags);
397 #endif
398                 }
399         }
400         ret = -EBUSY;
401         if (not_managed) {
402                 if (!list_empty(&source))
403                         putback_lru_pages(&source);
404                 goto out;
405         }
406         ret = 0;
407         if (list_empty(&source))
408                 goto out;
409         /* this function returns # of failed pages */
410         ret = migrate_pages(&source, hotremove_migrate_alloc, 0);
411
412 out:
413         return ret;
414 }
415
416 /*
417  * remove from free_area[] and mark all as Reserved.
418  */
419 static int
420 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages,
421                         void *data)
422 {
423         __offline_isolated_pages(start, start + nr_pages);
424         return 0;
425 }
426
427 static void
428 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
429 {
430         walk_memory_resource(start_pfn, end_pfn - start_pfn, NULL,
431                                 offline_isolated_pages_cb);
432 }
433
434 /*
435  * Check all pages in range, recoreded as memory resource, are isolated.
436  */
437 static int
438 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages,
439                         void *data)
440 {
441         int ret;
442         long offlined = *(long *)data;
443         ret = test_pages_isolated(start_pfn, start_pfn + nr_pages);
444         offlined = nr_pages;
445         if (!ret)
446                 *(long *)data += offlined;
447         return ret;
448 }
449
450 static long
451 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
452 {
453         long offlined = 0;
454         int ret;
455
456         ret = walk_memory_resource(start_pfn, end_pfn - start_pfn, &offlined,
457                         check_pages_isolated_cb);
458         if (ret < 0)
459                 offlined = (long)ret;
460         return offlined;
461 }
462
463 extern void drain_all_local_pages(void);
464
465 int offline_pages(unsigned long start_pfn,
466                   unsigned long end_pfn, unsigned long timeout)
467 {
468         unsigned long pfn, nr_pages, expire;
469         long offlined_pages;
470         int ret, drain, retry_max;
471         struct zone *zone;
472
473         BUG_ON(start_pfn >= end_pfn);
474         /* at least, alignment against pageblock is necessary */
475         if (!IS_ALIGNED(start_pfn, pageblock_nr_pages))
476                 return -EINVAL;
477         if (!IS_ALIGNED(end_pfn, pageblock_nr_pages))
478                 return -EINVAL;
479         /* This makes hotplug much easier...and readable.
480            we assume this for now. .*/
481         if (!test_pages_in_a_zone(start_pfn, end_pfn))
482                 return -EINVAL;
483         /* set above range as isolated */
484         ret = start_isolate_page_range(start_pfn, end_pfn);
485         if (ret)
486                 return ret;
487         nr_pages = end_pfn - start_pfn;
488         pfn = start_pfn;
489         expire = jiffies + timeout;
490         drain = 0;
491         retry_max = 5;
492 repeat:
493         /* start memory hot removal */
494         ret = -EAGAIN;
495         if (time_after(jiffies, expire))
496                 goto failed_removal;
497         ret = -EINTR;
498         if (signal_pending(current))
499                 goto failed_removal;
500         ret = 0;
501         if (drain) {
502                 lru_add_drain_all();
503                 flush_scheduled_work();
504                 cond_resched();
505                 drain_all_local_pages();
506         }
507
508         pfn = scan_lru_pages(start_pfn, end_pfn);
509         if (pfn) { /* We have page on LRU */
510                 ret = do_migrate_range(pfn, end_pfn);
511                 if (!ret) {
512                         drain = 1;
513                         goto repeat;
514                 } else {
515                         if (ret < 0)
516                                 if (--retry_max == 0)
517                                         goto failed_removal;
518                         yield();
519                         drain = 1;
520                         goto repeat;
521                 }
522         }
523         /* drain all zone's lru pagevec, this is asyncronous... */
524         lru_add_drain_all();
525         flush_scheduled_work();
526         yield();
527         /* drain pcp pages , this is synchrouns. */
528         drain_all_local_pages();
529         /* check again */
530         offlined_pages = check_pages_isolated(start_pfn, end_pfn);
531         if (offlined_pages < 0) {
532                 ret = -EBUSY;
533                 goto failed_removal;
534         }
535         printk(KERN_INFO "Offlined Pages %ld\n", offlined_pages);
536         /* Ok, all of our target is islaoted.
537            We cannot do rollback at this point. */
538         offline_isolated_pages(start_pfn, end_pfn);
539         /* reset pagetype flags */
540         start_isolate_page_range(start_pfn, end_pfn);
541         /* removal success */
542         zone = page_zone(pfn_to_page(start_pfn));
543         zone->present_pages -= offlined_pages;
544         zone->zone_pgdat->node_present_pages -= offlined_pages;
545         totalram_pages -= offlined_pages;
546         num_physpages -= offlined_pages;
547         vm_total_pages = nr_free_pagecache_pages();
548         writeback_set_ratelimit();
549         return 0;
550
551 failed_removal:
552         printk(KERN_INFO "memory offlining %lx to %lx failed\n",
553                 start_pfn, end_pfn);
554         /* pushback to free area */
555         undo_isolate_page_range(start_pfn, end_pfn);
556         return ret;
557 }
558 #endif /* CONFIG_MEMORY_HOTREMOVE */