Merge branch 'stable-3.2' into pandora-3.2
[pandora-kernel.git] / drivers / base / node.c
1 /*
2  * Basic Node interface support
3  */
4
5 #include <linux/module.h>
6 #include <linux/init.h>
7 #include <linux/mm.h>
8 #include <linux/memory.h>
9 #include <linux/vmstat.h>
10 #include <linux/node.h>
11 #include <linux/hugetlb.h>
12 #include <linux/compaction.h>
13 #include <linux/cpumask.h>
14 #include <linux/topology.h>
15 #include <linux/nodemask.h>
16 #include <linux/cpu.h>
17 #include <linux/device.h>
18 #include <linux/swap.h>
19 #include <linux/slab.h>
20
21 static struct bus_type node_subsys = {
22         .name = "node",
23         .dev_name = "node",
24 };
25
26
27 static ssize_t node_read_cpumap(struct device *dev, int type, char *buf)
28 {
29         struct node *node_dev = to_node(dev);
30         const struct cpumask *mask = cpumask_of_node(node_dev->dev.id);
31         int len;
32
33         /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
34         BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1));
35
36         len = type?
37                 cpulist_scnprintf(buf, PAGE_SIZE-2, mask) :
38                 cpumask_scnprintf(buf, PAGE_SIZE-2, mask);
39         buf[len++] = '\n';
40         buf[len] = '\0';
41         return len;
42 }
43
44 static inline ssize_t node_read_cpumask(struct device *dev,
45                                 struct device_attribute *attr, char *buf)
46 {
47         return node_read_cpumap(dev, 0, buf);
48 }
49 static inline ssize_t node_read_cpulist(struct device *dev,
50                                 struct device_attribute *attr, char *buf)
51 {
52         return node_read_cpumap(dev, 1, buf);
53 }
54
55 static DEVICE_ATTR(cpumap,  S_IRUGO, node_read_cpumask, NULL);
56 static DEVICE_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
57
58 #define K(x) ((x) << (PAGE_SHIFT - 10))
59 static ssize_t node_read_meminfo(struct device *dev,
60                         struct device_attribute *attr, char *buf)
61 {
62         int n;
63         int nid = dev->id;
64         struct sysinfo i;
65
66         si_meminfo_node(&i, nid);
67         n = sprintf(buf,
68                        "Node %d MemTotal:       %8lu kB\n"
69                        "Node %d MemFree:        %8lu kB\n"
70                        "Node %d MemUsed:        %8lu kB\n"
71                        "Node %d Active:         %8lu kB\n"
72                        "Node %d Inactive:       %8lu kB\n"
73                        "Node %d Active(anon):   %8lu kB\n"
74                        "Node %d Inactive(anon): %8lu kB\n"
75                        "Node %d Active(file):   %8lu kB\n"
76                        "Node %d Inactive(file): %8lu kB\n"
77                        "Node %d Unevictable:    %8lu kB\n"
78                        "Node %d Mlocked:        %8lu kB\n",
79                        nid, K(i.totalram),
80                        nid, K(i.freeram),
81                        nid, K(i.totalram - i.freeram),
82                        nid, K(node_page_state(nid, NR_ACTIVE_ANON) +
83                                 node_page_state(nid, NR_ACTIVE_FILE)),
84                        nid, K(node_page_state(nid, NR_INACTIVE_ANON) +
85                                 node_page_state(nid, NR_INACTIVE_FILE)),
86                        nid, K(node_page_state(nid, NR_ACTIVE_ANON)),
87                        nid, K(node_page_state(nid, NR_INACTIVE_ANON)),
88                        nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
89                        nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
90                        nid, K(node_page_state(nid, NR_UNEVICTABLE)),
91                        nid, K(node_page_state(nid, NR_MLOCK)));
92
93 #ifdef CONFIG_HIGHMEM
94         n += sprintf(buf + n,
95                        "Node %d HighTotal:      %8lu kB\n"
96                        "Node %d HighFree:       %8lu kB\n"
97                        "Node %d LowTotal:       %8lu kB\n"
98                        "Node %d LowFree:        %8lu kB\n",
99                        nid, K(i.totalhigh),
100                        nid, K(i.freehigh),
101                        nid, K(i.totalram - i.totalhigh),
102                        nid, K(i.freeram - i.freehigh));
103 #endif
104         n += sprintf(buf + n,
105                        "Node %d Dirty:          %8lu kB\n"
106                        "Node %d Writeback:      %8lu kB\n"
107                        "Node %d FilePages:      %8lu kB\n"
108                        "Node %d Mapped:         %8lu kB\n"
109                        "Node %d AnonPages:      %8lu kB\n"
110                        "Node %d Shmem:          %8lu kB\n"
111                        "Node %d KernelStack:    %8lu kB\n"
112                        "Node %d PageTables:     %8lu kB\n"
113                        "Node %d NFS_Unstable:   %8lu kB\n"
114                        "Node %d Bounce:         %8lu kB\n"
115                        "Node %d WritebackTmp:   %8lu kB\n"
116                        "Node %d Slab:           %8lu kB\n"
117                        "Node %d SReclaimable:   %8lu kB\n"
118                        "Node %d SUnreclaim:     %8lu kB\n"
119 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
120                        "Node %d AnonHugePages:  %8lu kB\n"
121 #endif
122                         ,
123                        nid, K(node_page_state(nid, NR_FILE_DIRTY)),
124                        nid, K(node_page_state(nid, NR_WRITEBACK)),
125                        nid, K(node_page_state(nid, NR_FILE_PAGES)),
126                        nid, K(node_page_state(nid, NR_FILE_MAPPED)),
127 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
128                        nid, K(node_page_state(nid, NR_ANON_PAGES)
129                         + node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) *
130                         HPAGE_PMD_NR),
131 #else
132                        nid, K(node_page_state(nid, NR_ANON_PAGES)),
133 #endif
134                        nid, K(node_page_state(nid, NR_SHMEM)),
135                        nid, node_page_state(nid, NR_KERNEL_STACK) *
136                                 THREAD_SIZE / 1024,
137                        nid, K(node_page_state(nid, NR_PAGETABLE)),
138                        nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
139                        nid, K(node_page_state(nid, NR_BOUNCE)),
140                        nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
141                        nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
142                                 node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
143                        nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
144 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
145                        nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE))
146                         , nid,
147                         K(node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) *
148                         HPAGE_PMD_NR));
149 #else
150                        nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
151 #endif
152         n += hugetlb_report_node_meminfo(nid, buf + n);
153         return n;
154 }
155
156 #undef K
157 static DEVICE_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
158
159 static ssize_t node_read_numastat(struct device *dev,
160                                 struct device_attribute *attr, char *buf)
161 {
162         return sprintf(buf,
163                        "numa_hit %lu\n"
164                        "numa_miss %lu\n"
165                        "numa_foreign %lu\n"
166                        "interleave_hit %lu\n"
167                        "local_node %lu\n"
168                        "other_node %lu\n",
169                        node_page_state(dev->id, NUMA_HIT),
170                        node_page_state(dev->id, NUMA_MISS),
171                        node_page_state(dev->id, NUMA_FOREIGN),
172                        node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
173                        node_page_state(dev->id, NUMA_LOCAL),
174                        node_page_state(dev->id, NUMA_OTHER));
175 }
176 static DEVICE_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
177
178 static ssize_t node_read_vmstat(struct device *dev,
179                                 struct device_attribute *attr, char *buf)
180 {
181         int nid = dev->id;
182         int i;
183         int n = 0;
184
185         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
186                 n += sprintf(buf+n, "%s %lu\n", vmstat_text[i],
187                              node_page_state(nid, i));
188
189         return n;
190 }
191 static DEVICE_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
192
193 static ssize_t node_read_distance(struct device *dev,
194                         struct device_attribute *attr, char * buf)
195 {
196         int nid = dev->id;
197         int len = 0;
198         int i;
199
200         /*
201          * buf is currently PAGE_SIZE in length and each node needs 4 chars
202          * at the most (distance + space or newline).
203          */
204         BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
205
206         for_each_online_node(i)
207                 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
208
209         len += sprintf(buf + len, "\n");
210         return len;
211 }
212 static DEVICE_ATTR(distance, S_IRUGO, node_read_distance, NULL);
213
214 #ifdef CONFIG_HUGETLBFS
215 /*
216  * hugetlbfs per node attributes registration interface:
217  * When/if hugetlb[fs] subsystem initializes [sometime after this module],
218  * it will register its per node attributes for all online nodes with
219  * memory.  It will also call register_hugetlbfs_with_node(), below, to
220  * register its attribute registration functions with this node driver.
221  * Once these hooks have been initialized, the node driver will call into
222  * the hugetlb module to [un]register attributes for hot-plugged nodes.
223  */
224 static node_registration_func_t __hugetlb_register_node;
225 static node_registration_func_t __hugetlb_unregister_node;
226
227 static inline bool hugetlb_register_node(struct node *node)
228 {
229         if (__hugetlb_register_node &&
230                         node_state(node->dev.id, N_HIGH_MEMORY)) {
231                 __hugetlb_register_node(node);
232                 return true;
233         }
234         return false;
235 }
236
237 static inline void hugetlb_unregister_node(struct node *node)
238 {
239         if (__hugetlb_unregister_node)
240                 __hugetlb_unregister_node(node);
241 }
242
243 void register_hugetlbfs_with_node(node_registration_func_t doregister,
244                                   node_registration_func_t unregister)
245 {
246         __hugetlb_register_node   = doregister;
247         __hugetlb_unregister_node = unregister;
248 }
249 #else
250 static inline void hugetlb_register_node(struct node *node) {}
251
252 static inline void hugetlb_unregister_node(struct node *node) {}
253 #endif
254
255
256 /*
257  * register_node - Setup a sysfs device for a node.
258  * @num - Node number to use when creating the device.
259  *
260  * Initialize and register the node device.
261  */
262 int register_node(struct node *node, int num, struct node *parent)
263 {
264         int error;
265
266         node->dev.id = num;
267         node->dev.bus = &node_subsys;
268         error = device_register(&node->dev);
269
270         if (!error){
271                 device_create_file(&node->dev, &dev_attr_cpumap);
272                 device_create_file(&node->dev, &dev_attr_cpulist);
273                 device_create_file(&node->dev, &dev_attr_meminfo);
274                 device_create_file(&node->dev, &dev_attr_numastat);
275                 device_create_file(&node->dev, &dev_attr_distance);
276                 device_create_file(&node->dev, &dev_attr_vmstat);
277
278                 scan_unevictable_register_node(node);
279
280                 hugetlb_register_node(node);
281
282                 compaction_register_node(node);
283         }
284         return error;
285 }
286
287 /**
288  * unregister_node - unregister a node device
289  * @node: node going away
290  *
291  * Unregisters a node device @node.  All the devices on the node must be
292  * unregistered before calling this function.
293  */
294 void unregister_node(struct node *node)
295 {
296         device_remove_file(&node->dev, &dev_attr_cpumap);
297         device_remove_file(&node->dev, &dev_attr_cpulist);
298         device_remove_file(&node->dev, &dev_attr_meminfo);
299         device_remove_file(&node->dev, &dev_attr_numastat);
300         device_remove_file(&node->dev, &dev_attr_distance);
301         device_remove_file(&node->dev, &dev_attr_vmstat);
302
303         scan_unevictable_unregister_node(node);
304         hugetlb_unregister_node(node);          /* no-op, if memoryless node */
305
306         device_unregister(&node->dev);
307 }
308
309 struct node node_devices[MAX_NUMNODES];
310
311 /*
312  * register cpu under node
313  */
314 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
315 {
316         int ret;
317         struct sys_device *obj;
318
319         if (!node_online(nid))
320                 return 0;
321
322         obj = get_cpu_sysdev(cpu);
323         if (!obj)
324                 return 0;
325
326         ret = sysfs_create_link(&node_devices[nid].dev.kobj,
327                                 &obj->kobj,
328                                 kobject_name(&obj->kobj));
329         if (ret)
330                 return ret;
331
332         return sysfs_create_link(&obj->kobj,
333                                  &node_devices[nid].dev.kobj,
334                                  kobject_name(&node_devices[nid].dev.kobj));
335 }
336
337 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
338 {
339         struct sys_device *obj;
340
341         if (!node_online(nid))
342                 return 0;
343
344         obj = get_cpu_sysdev(cpu);
345         if (!obj)
346                 return 0;
347
348         sysfs_remove_link(&node_devices[nid].dev.kobj,
349                           kobject_name(&obj->kobj));
350         sysfs_remove_link(&obj->kobj,
351                           kobject_name(&node_devices[nid].dev.kobj));
352
353         return 0;
354 }
355
356 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
357 #define page_initialized(page)  (page->lru.next)
358
359 static int get_nid_for_pfn(unsigned long pfn)
360 {
361         struct page *page;
362
363         if (!pfn_valid_within(pfn))
364                 return -1;
365         page = pfn_to_page(pfn);
366         if (!page_initialized(page))
367                 return -1;
368         return pfn_to_nid(pfn);
369 }
370
371 /* register memory section under specified node if it spans that node */
372 int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
373 {
374         int ret;
375         unsigned long pfn, sect_start_pfn, sect_end_pfn;
376
377         if (!mem_blk)
378                 return -EFAULT;
379         if (!node_online(nid))
380                 return 0;
381
382         sect_start_pfn = section_nr_to_pfn(mem_blk->start_section_nr);
383         sect_end_pfn = section_nr_to_pfn(mem_blk->end_section_nr);
384         sect_end_pfn += PAGES_PER_SECTION - 1;
385         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
386                 int page_nid;
387
388                 page_nid = get_nid_for_pfn(pfn);
389                 if (page_nid < 0)
390                         continue;
391                 if (page_nid != nid)
392                         continue;
393                 ret = sysfs_create_link_nowarn(&node_devices[nid].dev.kobj,
394                                         &mem_blk->dev.kobj,
395                                         kobject_name(&mem_blk->dev.kobj));
396                 if (ret)
397                         return ret;
398
399                 return sysfs_create_link_nowarn(&mem_blk->dev.kobj,
400                                 &node_devices[nid].dev.kobj,
401                                 kobject_name(&node_devices[nid].dev.kobj));
402         }
403         /* mem section does not span the specified node */
404         return 0;
405 }
406
407 /* unregister memory section under all nodes that it spans */
408 int unregister_mem_sect_under_nodes(struct memory_block *mem_blk,
409                                     unsigned long phys_index)
410 {
411         NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
412         unsigned long pfn, sect_start_pfn, sect_end_pfn;
413
414         if (!mem_blk) {
415                 NODEMASK_FREE(unlinked_nodes);
416                 return -EFAULT;
417         }
418         if (!unlinked_nodes)
419                 return -ENOMEM;
420         nodes_clear(*unlinked_nodes);
421
422         sect_start_pfn = section_nr_to_pfn(phys_index);
423         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
424         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
425                 int nid;
426
427                 nid = get_nid_for_pfn(pfn);
428                 if (nid < 0)
429                         continue;
430                 if (!node_online(nid))
431                         continue;
432                 if (node_test_and_set(nid, *unlinked_nodes))
433                         continue;
434                 sysfs_remove_link(&node_devices[nid].dev.kobj,
435                          kobject_name(&mem_blk->dev.kobj));
436                 sysfs_remove_link(&mem_blk->dev.kobj,
437                          kobject_name(&node_devices[nid].dev.kobj));
438         }
439         NODEMASK_FREE(unlinked_nodes);
440         return 0;
441 }
442
443 static int link_mem_sections(int nid)
444 {
445         unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
446         unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
447         unsigned long pfn;
448         struct memory_block *mem_blk = NULL;
449         int err = 0;
450
451         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
452                 unsigned long section_nr = pfn_to_section_nr(pfn);
453                 struct mem_section *mem_sect;
454                 int ret;
455
456                 if (!present_section_nr(section_nr))
457                         continue;
458                 mem_sect = __nr_to_section(section_nr);
459                 mem_blk = find_memory_block_hinted(mem_sect, mem_blk);
460                 ret = register_mem_sect_under_node(mem_blk, nid);
461                 if (!err)
462                         err = ret;
463
464                 /* discard ref obtained in find_memory_block() */
465         }
466
467         if (mem_blk)
468                 kobject_put(&mem_blk->dev.kobj);
469         return err;
470 }
471
472 #ifdef CONFIG_HUGETLBFS
473 /*
474  * Handle per node hstate attribute [un]registration on transistions
475  * to/from memoryless state.
476  */
477 static void node_hugetlb_work(struct work_struct *work)
478 {
479         struct node *node = container_of(work, struct node, node_work);
480
481         /*
482          * We only get here when a node transitions to/from memoryless state.
483          * We can detect which transition occurred by examining whether the
484          * node has memory now.  hugetlb_register_node() already check this
485          * so we try to register the attributes.  If that fails, then the
486          * node has transitioned to memoryless, try to unregister the
487          * attributes.
488          */
489         if (!hugetlb_register_node(node))
490                 hugetlb_unregister_node(node);
491 }
492
493 static void init_node_hugetlb_work(int nid)
494 {
495         INIT_WORK(&node_devices[nid].node_work, node_hugetlb_work);
496 }
497
498 static int node_memory_callback(struct notifier_block *self,
499                                 unsigned long action, void *arg)
500 {
501         struct memory_notify *mnb = arg;
502         int nid = mnb->status_change_nid;
503
504         switch (action) {
505         case MEM_ONLINE:
506         case MEM_OFFLINE:
507                 /*
508                  * offload per node hstate [un]registration to a work thread
509                  * when transitioning to/from memoryless state.
510                  */
511                 if (nid != NUMA_NO_NODE)
512                         schedule_work(&node_devices[nid].node_work);
513                 break;
514
515         case MEM_GOING_ONLINE:
516         case MEM_GOING_OFFLINE:
517         case MEM_CANCEL_ONLINE:
518         case MEM_CANCEL_OFFLINE:
519         default:
520                 break;
521         }
522
523         return NOTIFY_OK;
524 }
525 #endif  /* CONFIG_HUGETLBFS */
526 #else   /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
527
528 static int link_mem_sections(int nid) { return 0; }
529 #endif  /* CONFIG_MEMORY_HOTPLUG_SPARSE */
530
531 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
532     !defined(CONFIG_HUGETLBFS)
533 static inline int node_memory_callback(struct notifier_block *self,
534                                 unsigned long action, void *arg)
535 {
536         return NOTIFY_OK;
537 }
538
539 static void init_node_hugetlb_work(int nid) { }
540
541 #endif
542
543 int register_one_node(int nid)
544 {
545         int error = 0;
546         int cpu;
547
548         if (node_online(nid)) {
549                 int p_node = parent_node(nid);
550                 struct node *parent = NULL;
551
552                 if (p_node != nid)
553                         parent = &node_devices[p_node];
554
555                 error = register_node(&node_devices[nid], nid, parent);
556
557                 /* link cpu under this node */
558                 for_each_present_cpu(cpu) {
559                         if (cpu_to_node(cpu) == nid)
560                                 register_cpu_under_node(cpu, nid);
561                 }
562
563                 /* link memory sections under this node */
564                 error = link_mem_sections(nid);
565
566                 /* initialize work queue for memory hot plug */
567                 init_node_hugetlb_work(nid);
568         }
569
570         return error;
571
572 }
573
574 void unregister_one_node(int nid)
575 {
576         unregister_node(&node_devices[nid]);
577 }
578
579 /*
580  * node states attributes
581  */
582
583 static ssize_t print_nodes_state(enum node_states state, char *buf)
584 {
585         int n;
586
587         n = nodelist_scnprintf(buf, PAGE_SIZE, node_states[state]);
588         if (n > 0 && PAGE_SIZE > n + 1) {
589                 *(buf + n++) = '\n';
590                 *(buf + n++) = '\0';
591         }
592         return n;
593 }
594
595 struct node_attr {
596         struct device_attribute attr;
597         enum node_states state;
598 };
599
600 static ssize_t show_node_state(struct device *dev,
601                                struct device_attribute *attr, char *buf)
602 {
603         struct node_attr *na = container_of(attr, struct node_attr, attr);
604         return print_nodes_state(na->state, buf);
605 }
606
607 #define _NODE_ATTR(name, state) \
608         { __ATTR(name, 0444, show_node_state, NULL), state }
609
610 static struct node_attr node_state_attr[] = {
611         _NODE_ATTR(possible, N_POSSIBLE),
612         _NODE_ATTR(online, N_ONLINE),
613         _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
614         _NODE_ATTR(has_cpu, N_CPU),
615 #ifdef CONFIG_HIGHMEM
616         _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
617 #endif
618 };
619
620 static struct attribute *node_state_attrs[] = {
621         &node_state_attr[0].attr.attr,
622         &node_state_attr[1].attr.attr,
623         &node_state_attr[2].attr.attr,
624         &node_state_attr[3].attr.attr,
625 #ifdef CONFIG_HIGHMEM
626         &node_state_attr[4].attr.attr,
627 #endif
628         NULL
629 };
630
631 static struct attribute_group memory_root_attr_group = {
632         .attrs = node_state_attrs,
633 };
634
635 static const struct attribute_group *cpu_root_attr_groups[] = {
636         &memory_root_attr_group,
637         NULL,
638 };
639
640 #define NODE_CALLBACK_PRI       2       /* lower than SLAB */
641 static int __init register_node_type(void)
642 {
643         int ret;
644
645         BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
646         BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
647
648         ret = subsys_system_register(&node_subsys, cpu_root_attr_groups);
649         if (!ret) {
650                 hotplug_memory_notifier(node_memory_callback,
651                                         NODE_CALLBACK_PRI);
652         }
653
654         /*
655          * Note:  we're not going to unregister the node class if we fail
656          * to register the node state class attribute files.
657          */
658         return ret;
659 }
660 postcore_initcall(register_node_type);