Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs-2.6
[pandora-kernel.git] / drivers / base / node.c
1 /*
2  * drivers/base/node.c - basic Node class support
3  */
4
5 #include <linux/sysdev.h>
6 #include <linux/module.h>
7 #include <linux/init.h>
8 #include <linux/mm.h>
9 #include <linux/memory.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 sysdev_class_attribute *node_state_attrs[];
22
23 static struct sysdev_class node_class = {
24         .name = "node",
25         .attrs = node_state_attrs,
26 };
27
28
29 static ssize_t node_read_cpumap(struct sys_device *dev, int type, char *buf)
30 {
31         struct node *node_dev = to_node(dev);
32         const struct cpumask *mask = cpumask_of_node(node_dev->sysdev.id);
33         int len;
34
35         /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
36         BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1));
37
38         len = type?
39                 cpulist_scnprintf(buf, PAGE_SIZE-2, mask) :
40                 cpumask_scnprintf(buf, PAGE_SIZE-2, mask);
41         buf[len++] = '\n';
42         buf[len] = '\0';
43         return len;
44 }
45
46 static inline ssize_t node_read_cpumask(struct sys_device *dev,
47                                 struct sysdev_attribute *attr, char *buf)
48 {
49         return node_read_cpumap(dev, 0, buf);
50 }
51 static inline ssize_t node_read_cpulist(struct sys_device *dev,
52                                 struct sysdev_attribute *attr, char *buf)
53 {
54         return node_read_cpumap(dev, 1, buf);
55 }
56
57 static SYSDEV_ATTR(cpumap,  S_IRUGO, node_read_cpumask, NULL);
58 static SYSDEV_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
59
60 #define K(x) ((x) << (PAGE_SHIFT - 10))
61 static ssize_t node_read_meminfo(struct sys_device * dev,
62                         struct sysdev_attribute *attr, char * buf)
63 {
64         int n;
65         int nid = dev->id;
66         struct sysinfo i;
67
68         si_meminfo_node(&i, nid);
69         n = sprintf(buf,
70                        "Node %d MemTotal:       %8lu kB\n"
71                        "Node %d MemFree:        %8lu kB\n"
72                        "Node %d MemUsed:        %8lu kB\n"
73                        "Node %d Active:         %8lu kB\n"
74                        "Node %d Inactive:       %8lu kB\n"
75                        "Node %d Active(anon):   %8lu kB\n"
76                        "Node %d Inactive(anon): %8lu kB\n"
77                        "Node %d Active(file):   %8lu kB\n"
78                        "Node %d Inactive(file): %8lu kB\n"
79                        "Node %d Unevictable:    %8lu kB\n"
80                        "Node %d Mlocked:        %8lu kB\n",
81                        nid, K(i.totalram),
82                        nid, K(i.freeram),
83                        nid, K(i.totalram - i.freeram),
84                        nid, K(node_page_state(nid, NR_ACTIVE_ANON) +
85                                 node_page_state(nid, NR_ACTIVE_FILE)),
86                        nid, K(node_page_state(nid, NR_INACTIVE_ANON) +
87                                 node_page_state(nid, NR_INACTIVE_FILE)),
88                        nid, K(node_page_state(nid, NR_ACTIVE_ANON)),
89                        nid, K(node_page_state(nid, NR_INACTIVE_ANON)),
90                        nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
91                        nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
92                        nid, K(node_page_state(nid, NR_UNEVICTABLE)),
93                        nid, K(node_page_state(nid, NR_MLOCK)));
94
95 #ifdef CONFIG_HIGHMEM
96         n += sprintf(buf + n,
97                        "Node %d HighTotal:      %8lu kB\n"
98                        "Node %d HighFree:       %8lu kB\n"
99                        "Node %d LowTotal:       %8lu kB\n"
100                        "Node %d LowFree:        %8lu kB\n",
101                        nid, K(i.totalhigh),
102                        nid, K(i.freehigh),
103                        nid, K(i.totalram - i.totalhigh),
104                        nid, K(i.freeram - i.freehigh));
105 #endif
106         n += sprintf(buf + n,
107                        "Node %d Dirty:          %8lu kB\n"
108                        "Node %d Writeback:      %8lu kB\n"
109                        "Node %d FilePages:      %8lu kB\n"
110                        "Node %d Mapped:         %8lu kB\n"
111                        "Node %d AnonPages:      %8lu kB\n"
112                        "Node %d Shmem:          %8lu kB\n"
113                        "Node %d KernelStack:    %8lu kB\n"
114                        "Node %d PageTables:     %8lu kB\n"
115                        "Node %d NFS_Unstable:   %8lu kB\n"
116                        "Node %d Bounce:         %8lu kB\n"
117                        "Node %d WritebackTmp:   %8lu kB\n"
118                        "Node %d Slab:           %8lu kB\n"
119                        "Node %d SReclaimable:   %8lu kB\n"
120                        "Node %d SUnreclaim:     %8lu kB\n",
121                        nid, K(node_page_state(nid, NR_FILE_DIRTY)),
122                        nid, K(node_page_state(nid, NR_WRITEBACK)),
123                        nid, K(node_page_state(nid, NR_FILE_PAGES)),
124                        nid, K(node_page_state(nid, NR_FILE_MAPPED)),
125                        nid, K(node_page_state(nid, NR_ANON_PAGES)),
126                        nid, K(node_page_state(nid, NR_SHMEM)),
127                        nid, node_page_state(nid, NR_KERNEL_STACK) *
128                                 THREAD_SIZE / 1024,
129                        nid, K(node_page_state(nid, NR_PAGETABLE)),
130                        nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
131                        nid, K(node_page_state(nid, NR_BOUNCE)),
132                        nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
133                        nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
134                                 node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
135                        nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
136                        nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
137         n += hugetlb_report_node_meminfo(nid, buf + n);
138         return n;
139 }
140
141 #undef K
142 static SYSDEV_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
143
144 static ssize_t node_read_numastat(struct sys_device * dev,
145                                 struct sysdev_attribute *attr, char * buf)
146 {
147         return sprintf(buf,
148                        "numa_hit %lu\n"
149                        "numa_miss %lu\n"
150                        "numa_foreign %lu\n"
151                        "interleave_hit %lu\n"
152                        "local_node %lu\n"
153                        "other_node %lu\n",
154                        node_page_state(dev->id, NUMA_HIT),
155                        node_page_state(dev->id, NUMA_MISS),
156                        node_page_state(dev->id, NUMA_FOREIGN),
157                        node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
158                        node_page_state(dev->id, NUMA_LOCAL),
159                        node_page_state(dev->id, NUMA_OTHER));
160 }
161 static SYSDEV_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
162
163 static ssize_t node_read_vmstat(struct sys_device *dev,
164                                 struct sysdev_attribute *attr, char *buf)
165 {
166         int nid = dev->id;
167         return sprintf(buf,
168                 "nr_written %lu\n"
169                 "nr_dirtied %lu\n",
170                 node_page_state(nid, NR_WRITTEN),
171                 node_page_state(nid, NR_DIRTIED));
172 }
173 static SYSDEV_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
174
175 static ssize_t node_read_distance(struct sys_device * dev,
176                         struct sysdev_attribute *attr, char * buf)
177 {
178         int nid = dev->id;
179         int len = 0;
180         int i;
181
182         /*
183          * buf is currently PAGE_SIZE in length and each node needs 4 chars
184          * at the most (distance + space or newline).
185          */
186         BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
187
188         for_each_online_node(i)
189                 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
190
191         len += sprintf(buf + len, "\n");
192         return len;
193 }
194 static SYSDEV_ATTR(distance, S_IRUGO, node_read_distance, NULL);
195
196 #ifdef CONFIG_HUGETLBFS
197 /*
198  * hugetlbfs per node attributes registration interface:
199  * When/if hugetlb[fs] subsystem initializes [sometime after this module],
200  * it will register its per node attributes for all online nodes with
201  * memory.  It will also call register_hugetlbfs_with_node(), below, to
202  * register its attribute registration functions with this node driver.
203  * Once these hooks have been initialized, the node driver will call into
204  * the hugetlb module to [un]register attributes for hot-plugged nodes.
205  */
206 static node_registration_func_t __hugetlb_register_node;
207 static node_registration_func_t __hugetlb_unregister_node;
208
209 static inline bool hugetlb_register_node(struct node *node)
210 {
211         if (__hugetlb_register_node &&
212                         node_state(node->sysdev.id, N_HIGH_MEMORY)) {
213                 __hugetlb_register_node(node);
214                 return true;
215         }
216         return false;
217 }
218
219 static inline void hugetlb_unregister_node(struct node *node)
220 {
221         if (__hugetlb_unregister_node)
222                 __hugetlb_unregister_node(node);
223 }
224
225 void register_hugetlbfs_with_node(node_registration_func_t doregister,
226                                   node_registration_func_t unregister)
227 {
228         __hugetlb_register_node   = doregister;
229         __hugetlb_unregister_node = unregister;
230 }
231 #else
232 static inline void hugetlb_register_node(struct node *node) {}
233
234 static inline void hugetlb_unregister_node(struct node *node) {}
235 #endif
236
237
238 /*
239  * register_node - Setup a sysfs device for a node.
240  * @num - Node number to use when creating the device.
241  *
242  * Initialize and register the node device.
243  */
244 int register_node(struct node *node, int num, struct node *parent)
245 {
246         int error;
247
248         node->sysdev.id = num;
249         node->sysdev.cls = &node_class;
250         error = sysdev_register(&node->sysdev);
251
252         if (!error){
253                 sysdev_create_file(&node->sysdev, &attr_cpumap);
254                 sysdev_create_file(&node->sysdev, &attr_cpulist);
255                 sysdev_create_file(&node->sysdev, &attr_meminfo);
256                 sysdev_create_file(&node->sysdev, &attr_numastat);
257                 sysdev_create_file(&node->sysdev, &attr_distance);
258                 sysdev_create_file(&node->sysdev, &attr_vmstat);
259
260                 scan_unevictable_register_node(node);
261
262                 hugetlb_register_node(node);
263
264                 compaction_register_node(node);
265         }
266         return error;
267 }
268
269 /**
270  * unregister_node - unregister a node device
271  * @node: node going away
272  *
273  * Unregisters a node device @node.  All the devices on the node must be
274  * unregistered before calling this function.
275  */
276 void unregister_node(struct node *node)
277 {
278         sysdev_remove_file(&node->sysdev, &attr_cpumap);
279         sysdev_remove_file(&node->sysdev, &attr_cpulist);
280         sysdev_remove_file(&node->sysdev, &attr_meminfo);
281         sysdev_remove_file(&node->sysdev, &attr_numastat);
282         sysdev_remove_file(&node->sysdev, &attr_distance);
283         sysdev_remove_file(&node->sysdev, &attr_vmstat);
284
285         scan_unevictable_unregister_node(node);
286         hugetlb_unregister_node(node);          /* no-op, if memoryless node */
287
288         sysdev_unregister(&node->sysdev);
289 }
290
291 struct node node_devices[MAX_NUMNODES];
292
293 /*
294  * register cpu under node
295  */
296 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
297 {
298         int ret;
299         struct sys_device *obj;
300
301         if (!node_online(nid))
302                 return 0;
303
304         obj = get_cpu_sysdev(cpu);
305         if (!obj)
306                 return 0;
307
308         ret = sysfs_create_link(&node_devices[nid].sysdev.kobj,
309                                 &obj->kobj,
310                                 kobject_name(&obj->kobj));
311         if (ret)
312                 return ret;
313
314         return sysfs_create_link(&obj->kobj,
315                                  &node_devices[nid].sysdev.kobj,
316                                  kobject_name(&node_devices[nid].sysdev.kobj));
317 }
318
319 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
320 {
321         struct sys_device *obj;
322
323         if (!node_online(nid))
324                 return 0;
325
326         obj = get_cpu_sysdev(cpu);
327         if (!obj)
328                 return 0;
329
330         sysfs_remove_link(&node_devices[nid].sysdev.kobj,
331                           kobject_name(&obj->kobj));
332         sysfs_remove_link(&obj->kobj,
333                           kobject_name(&node_devices[nid].sysdev.kobj));
334
335         return 0;
336 }
337
338 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
339 #define page_initialized(page)  (page->lru.next)
340
341 static int get_nid_for_pfn(unsigned long pfn)
342 {
343         struct page *page;
344
345         if (!pfn_valid_within(pfn))
346                 return -1;
347         page = pfn_to_page(pfn);
348         if (!page_initialized(page))
349                 return -1;
350         return pfn_to_nid(pfn);
351 }
352
353 /* register memory section under specified node if it spans that node */
354 int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
355 {
356         int ret;
357         unsigned long pfn, sect_start_pfn, sect_end_pfn;
358
359         if (!mem_blk)
360                 return -EFAULT;
361         if (!node_online(nid))
362                 return 0;
363         sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
364         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
365         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
366                 int page_nid;
367
368                 page_nid = get_nid_for_pfn(pfn);
369                 if (page_nid < 0)
370                         continue;
371                 if (page_nid != nid)
372                         continue;
373                 ret = sysfs_create_link_nowarn(&node_devices[nid].sysdev.kobj,
374                                         &mem_blk->sysdev.kobj,
375                                         kobject_name(&mem_blk->sysdev.kobj));
376                 if (ret)
377                         return ret;
378
379                 return sysfs_create_link_nowarn(&mem_blk->sysdev.kobj,
380                                 &node_devices[nid].sysdev.kobj,
381                                 kobject_name(&node_devices[nid].sysdev.kobj));
382         }
383         /* mem section does not span the specified node */
384         return 0;
385 }
386
387 /* unregister memory section under all nodes that it spans */
388 int unregister_mem_sect_under_nodes(struct memory_block *mem_blk)
389 {
390         NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
391         unsigned long pfn, sect_start_pfn, sect_end_pfn;
392
393         if (!mem_blk) {
394                 NODEMASK_FREE(unlinked_nodes);
395                 return -EFAULT;
396         }
397         if (!unlinked_nodes)
398                 return -ENOMEM;
399         nodes_clear(*unlinked_nodes);
400         sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
401         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
402         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
403                 int nid;
404
405                 nid = get_nid_for_pfn(pfn);
406                 if (nid < 0)
407                         continue;
408                 if (!node_online(nid))
409                         continue;
410                 if (node_test_and_set(nid, *unlinked_nodes))
411                         continue;
412                 sysfs_remove_link(&node_devices[nid].sysdev.kobj,
413                          kobject_name(&mem_blk->sysdev.kobj));
414                 sysfs_remove_link(&mem_blk->sysdev.kobj,
415                          kobject_name(&node_devices[nid].sysdev.kobj));
416         }
417         NODEMASK_FREE(unlinked_nodes);
418         return 0;
419 }
420
421 static int link_mem_sections(int nid)
422 {
423         unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
424         unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
425         unsigned long pfn;
426         struct memory_block *mem_blk = NULL;
427         int err = 0;
428
429         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
430                 unsigned long section_nr = pfn_to_section_nr(pfn);
431                 struct mem_section *mem_sect;
432                 int ret;
433
434                 if (!present_section_nr(section_nr))
435                         continue;
436                 mem_sect = __nr_to_section(section_nr);
437                 mem_blk = find_memory_block_hinted(mem_sect, mem_blk);
438                 ret = register_mem_sect_under_node(mem_blk, nid);
439                 if (!err)
440                         err = ret;
441
442                 /* discard ref obtained in find_memory_block() */
443         }
444
445         if (mem_blk)
446                 kobject_put(&mem_blk->sysdev.kobj);
447         return err;
448 }
449
450 #ifdef CONFIG_HUGETLBFS
451 /*
452  * Handle per node hstate attribute [un]registration on transistions
453  * to/from memoryless state.
454  */
455 static void node_hugetlb_work(struct work_struct *work)
456 {
457         struct node *node = container_of(work, struct node, node_work);
458
459         /*
460          * We only get here when a node transitions to/from memoryless state.
461          * We can detect which transition occurred by examining whether the
462          * node has memory now.  hugetlb_register_node() already check this
463          * so we try to register the attributes.  If that fails, then the
464          * node has transitioned to memoryless, try to unregister the
465          * attributes.
466          */
467         if (!hugetlb_register_node(node))
468                 hugetlb_unregister_node(node);
469 }
470
471 static void init_node_hugetlb_work(int nid)
472 {
473         INIT_WORK(&node_devices[nid].node_work, node_hugetlb_work);
474 }
475
476 static int node_memory_callback(struct notifier_block *self,
477                                 unsigned long action, void *arg)
478 {
479         struct memory_notify *mnb = arg;
480         int nid = mnb->status_change_nid;
481
482         switch (action) {
483         case MEM_ONLINE:
484         case MEM_OFFLINE:
485                 /*
486                  * offload per node hstate [un]registration to a work thread
487                  * when transitioning to/from memoryless state.
488                  */
489                 if (nid != NUMA_NO_NODE)
490                         schedule_work(&node_devices[nid].node_work);
491                 break;
492
493         case MEM_GOING_ONLINE:
494         case MEM_GOING_OFFLINE:
495         case MEM_CANCEL_ONLINE:
496         case MEM_CANCEL_OFFLINE:
497         default:
498                 break;
499         }
500
501         return NOTIFY_OK;
502 }
503 #endif  /* CONFIG_HUGETLBFS */
504 #else   /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
505
506 static int link_mem_sections(int nid) { return 0; }
507 #endif  /* CONFIG_MEMORY_HOTPLUG_SPARSE */
508
509 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
510     !defined(CONFIG_HUGETLBFS)
511 static inline int node_memory_callback(struct notifier_block *self,
512                                 unsigned long action, void *arg)
513 {
514         return NOTIFY_OK;
515 }
516
517 static void init_node_hugetlb_work(int nid) { }
518
519 #endif
520
521 int register_one_node(int nid)
522 {
523         int error = 0;
524         int cpu;
525
526         if (node_online(nid)) {
527                 int p_node = parent_node(nid);
528                 struct node *parent = NULL;
529
530                 if (p_node != nid)
531                         parent = &node_devices[p_node];
532
533                 error = register_node(&node_devices[nid], nid, parent);
534
535                 /* link cpu under this node */
536                 for_each_present_cpu(cpu) {
537                         if (cpu_to_node(cpu) == nid)
538                                 register_cpu_under_node(cpu, nid);
539                 }
540
541                 /* link memory sections under this node */
542                 error = link_mem_sections(nid);
543
544                 /* initialize work queue for memory hot plug */
545                 init_node_hugetlb_work(nid);
546         }
547
548         return error;
549
550 }
551
552 void unregister_one_node(int nid)
553 {
554         unregister_node(&node_devices[nid]);
555 }
556
557 /*
558  * node states attributes
559  */
560
561 static ssize_t print_nodes_state(enum node_states state, char *buf)
562 {
563         int n;
564
565         n = nodelist_scnprintf(buf, PAGE_SIZE, node_states[state]);
566         if (n > 0 && PAGE_SIZE > n + 1) {
567                 *(buf + n++) = '\n';
568                 *(buf + n++) = '\0';
569         }
570         return n;
571 }
572
573 struct node_attr {
574         struct sysdev_class_attribute attr;
575         enum node_states state;
576 };
577
578 static ssize_t show_node_state(struct sysdev_class *class,
579                                struct sysdev_class_attribute *attr, char *buf)
580 {
581         struct node_attr *na = container_of(attr, struct node_attr, attr);
582         return print_nodes_state(na->state, buf);
583 }
584
585 #define _NODE_ATTR(name, state) \
586         { _SYSDEV_CLASS_ATTR(name, 0444, show_node_state, NULL), state }
587
588 static struct node_attr node_state_attr[] = {
589         _NODE_ATTR(possible, N_POSSIBLE),
590         _NODE_ATTR(online, N_ONLINE),
591         _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
592         _NODE_ATTR(has_cpu, N_CPU),
593 #ifdef CONFIG_HIGHMEM
594         _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
595 #endif
596 };
597
598 static struct sysdev_class_attribute *node_state_attrs[] = {
599         &node_state_attr[0].attr,
600         &node_state_attr[1].attr,
601         &node_state_attr[2].attr,
602         &node_state_attr[3].attr,
603 #ifdef CONFIG_HIGHMEM
604         &node_state_attr[4].attr,
605 #endif
606         NULL
607 };
608
609 #define NODE_CALLBACK_PRI       2       /* lower than SLAB */
610 static int __init register_node_type(void)
611 {
612         int ret;
613
614         BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
615         BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
616
617         ret = sysdev_class_register(&node_class);
618         if (!ret) {
619                 hotplug_memory_notifier(node_memory_callback,
620                                         NODE_CALLBACK_PRI);
621         }
622
623         /*
624          * Note:  we're not going to unregister the node class if we fail
625          * to register the node state class attribute files.
626          */
627         return ret;
628 }
629 postcore_initcall(register_node_type);