Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph...
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / drivers / base / node.c
blob5693ecee9a4052a339b9eca105ff3704fffe29b4
1 /*
2 * drivers/base/node.c - basic Node class support
3 */
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/vmstat.h>
11 #include <linux/node.h>
12 #include <linux/hugetlb.h>
13 #include <linux/compaction.h>
14 #include <linux/cpumask.h>
15 #include <linux/topology.h>
16 #include <linux/nodemask.h>
17 #include <linux/cpu.h>
18 #include <linux/device.h>
19 #include <linux/swap.h>
20 #include <linux/slab.h>
22 static struct sysdev_class_attribute *node_state_attrs[];
24 static struct sysdev_class node_class = {
25 .name = "node",
26 .attrs = node_state_attrs,
30 static ssize_t node_read_cpumap(struct sys_device *dev, int type, char *buf)
32 struct node *node_dev = to_node(dev);
33 const struct cpumask *mask = cpumask_of_node(node_dev->sysdev.id);
34 int len;
36 /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
37 BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1));
39 len = type?
40 cpulist_scnprintf(buf, PAGE_SIZE-2, mask) :
41 cpumask_scnprintf(buf, PAGE_SIZE-2, mask);
42 buf[len++] = '\n';
43 buf[len] = '\0';
44 return len;
47 static inline ssize_t node_read_cpumask(struct sys_device *dev,
48 struct sysdev_attribute *attr, char *buf)
50 return node_read_cpumap(dev, 0, buf);
52 static inline ssize_t node_read_cpulist(struct sys_device *dev,
53 struct sysdev_attribute *attr, char *buf)
55 return node_read_cpumap(dev, 1, buf);
58 static SYSDEV_ATTR(cpumap, S_IRUGO, node_read_cpumask, NULL);
59 static SYSDEV_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
61 #define K(x) ((x) << (PAGE_SHIFT - 10))
62 static ssize_t node_read_meminfo(struct sys_device * dev,
63 struct sysdev_attribute *attr, char * buf)
65 int n;
66 int nid = dev->id;
67 struct sysinfo i;
69 si_meminfo_node(&i, nid);
70 n = sprintf(buf,
71 "Node %d MemTotal: %8lu kB\n"
72 "Node %d MemFree: %8lu kB\n"
73 "Node %d MemUsed: %8lu kB\n"
74 "Node %d Active: %8lu kB\n"
75 "Node %d Inactive: %8lu kB\n"
76 "Node %d Active(anon): %8lu kB\n"
77 "Node %d Inactive(anon): %8lu kB\n"
78 "Node %d Active(file): %8lu kB\n"
79 "Node %d Inactive(file): %8lu kB\n"
80 "Node %d Unevictable: %8lu kB\n"
81 "Node %d Mlocked: %8lu kB\n",
82 nid, K(i.totalram),
83 nid, K(i.freeram),
84 nid, K(i.totalram - i.freeram),
85 nid, K(node_page_state(nid, NR_ACTIVE_ANON) +
86 node_page_state(nid, NR_ACTIVE_FILE)),
87 nid, K(node_page_state(nid, NR_INACTIVE_ANON) +
88 node_page_state(nid, NR_INACTIVE_FILE)),
89 nid, K(node_page_state(nid, NR_ACTIVE_ANON)),
90 nid, K(node_page_state(nid, NR_INACTIVE_ANON)),
91 nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
92 nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
93 nid, K(node_page_state(nid, NR_UNEVICTABLE)),
94 nid, K(node_page_state(nid, NR_MLOCK)));
96 #ifdef CONFIG_HIGHMEM
97 n += sprintf(buf + n,
98 "Node %d HighTotal: %8lu kB\n"
99 "Node %d HighFree: %8lu kB\n"
100 "Node %d LowTotal: %8lu kB\n"
101 "Node %d LowFree: %8lu kB\n",
102 nid, K(i.totalhigh),
103 nid, K(i.freehigh),
104 nid, K(i.totalram - i.totalhigh),
105 nid, K(i.freeram - i.freehigh));
106 #endif
107 n += sprintf(buf + n,
108 "Node %d Dirty: %8lu kB\n"
109 "Node %d Writeback: %8lu kB\n"
110 "Node %d FilePages: %8lu kB\n"
111 "Node %d Mapped: %8lu kB\n"
112 "Node %d AnonPages: %8lu kB\n"
113 "Node %d Shmem: %8lu kB\n"
114 "Node %d KernelStack: %8lu kB\n"
115 "Node %d PageTables: %8lu kB\n"
116 "Node %d NFS_Unstable: %8lu kB\n"
117 "Node %d Bounce: %8lu kB\n"
118 "Node %d WritebackTmp: %8lu kB\n"
119 "Node %d Slab: %8lu kB\n"
120 "Node %d SReclaimable: %8lu kB\n"
121 "Node %d SUnreclaim: %8lu kB\n"
122 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
123 "Node %d AnonHugePages: %8lu kB\n"
124 #endif
126 nid, K(node_page_state(nid, NR_FILE_DIRTY)),
127 nid, K(node_page_state(nid, NR_WRITEBACK)),
128 nid, K(node_page_state(nid, NR_FILE_PAGES)),
129 nid, K(node_page_state(nid, NR_FILE_MAPPED)),
130 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
131 nid, K(node_page_state(nid, NR_ANON_PAGES)
132 + node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) *
133 HPAGE_PMD_NR),
134 #else
135 nid, K(node_page_state(nid, NR_ANON_PAGES)),
136 #endif
137 nid, K(node_page_state(nid, NR_SHMEM)),
138 nid, node_page_state(nid, NR_KERNEL_STACK) *
139 THREAD_SIZE / 1024,
140 nid, K(node_page_state(nid, NR_PAGETABLE)),
141 nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
142 nid, K(node_page_state(nid, NR_BOUNCE)),
143 nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
144 nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
145 node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
146 nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
147 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
148 nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE))
149 , nid,
150 K(node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) *
151 HPAGE_PMD_NR));
152 #else
153 nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
154 #endif
155 n += hugetlb_report_node_meminfo(nid, buf + n);
156 return n;
159 #undef K
160 static SYSDEV_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
162 static ssize_t node_read_numastat(struct sys_device * dev,
163 struct sysdev_attribute *attr, char * buf)
165 return sprintf(buf,
166 "numa_hit %lu\n"
167 "numa_miss %lu\n"
168 "numa_foreign %lu\n"
169 "interleave_hit %lu\n"
170 "local_node %lu\n"
171 "other_node %lu\n",
172 node_page_state(dev->id, NUMA_HIT),
173 node_page_state(dev->id, NUMA_MISS),
174 node_page_state(dev->id, NUMA_FOREIGN),
175 node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
176 node_page_state(dev->id, NUMA_LOCAL),
177 node_page_state(dev->id, NUMA_OTHER));
179 static SYSDEV_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
181 static ssize_t node_read_vmstat(struct sys_device *dev,
182 struct sysdev_attribute *attr, char *buf)
184 int nid = dev->id;
185 int i;
186 int n = 0;
188 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
189 n += sprintf(buf+n, "%s %lu\n", vmstat_text[i],
190 node_page_state(nid, i));
192 return n;
194 static SYSDEV_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
196 static ssize_t node_read_distance(struct sys_device * dev,
197 struct sysdev_attribute *attr, char * buf)
199 int nid = dev->id;
200 int len = 0;
201 int i;
204 * buf is currently PAGE_SIZE in length and each node needs 4 chars
205 * at the most (distance + space or newline).
207 BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
209 for_each_online_node(i)
210 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
212 len += sprintf(buf + len, "\n");
213 return len;
215 static SYSDEV_ATTR(distance, S_IRUGO, node_read_distance, NULL);
217 #ifdef CONFIG_HUGETLBFS
219 * hugetlbfs per node attributes registration interface:
220 * When/if hugetlb[fs] subsystem initializes [sometime after this module],
221 * it will register its per node attributes for all online nodes with
222 * memory. It will also call register_hugetlbfs_with_node(), below, to
223 * register its attribute registration functions with this node driver.
224 * Once these hooks have been initialized, the node driver will call into
225 * the hugetlb module to [un]register attributes for hot-plugged nodes.
227 static node_registration_func_t __hugetlb_register_node;
228 static node_registration_func_t __hugetlb_unregister_node;
230 static inline bool hugetlb_register_node(struct node *node)
232 if (__hugetlb_register_node &&
233 node_state(node->sysdev.id, N_HIGH_MEMORY)) {
234 __hugetlb_register_node(node);
235 return true;
237 return false;
240 static inline void hugetlb_unregister_node(struct node *node)
242 if (__hugetlb_unregister_node)
243 __hugetlb_unregister_node(node);
246 void register_hugetlbfs_with_node(node_registration_func_t doregister,
247 node_registration_func_t unregister)
249 __hugetlb_register_node = doregister;
250 __hugetlb_unregister_node = unregister;
252 #else
253 static inline void hugetlb_register_node(struct node *node) {}
255 static inline void hugetlb_unregister_node(struct node *node) {}
256 #endif
260 * register_node - Setup a sysfs device for a node.
261 * @num - Node number to use when creating the device.
263 * Initialize and register the node device.
265 int register_node(struct node *node, int num, struct node *parent)
267 int error;
269 node->sysdev.id = num;
270 node->sysdev.cls = &node_class;
271 error = sysdev_register(&node->sysdev);
273 if (!error){
274 sysdev_create_file(&node->sysdev, &attr_cpumap);
275 sysdev_create_file(&node->sysdev, &attr_cpulist);
276 sysdev_create_file(&node->sysdev, &attr_meminfo);
277 sysdev_create_file(&node->sysdev, &attr_numastat);
278 sysdev_create_file(&node->sysdev, &attr_distance);
279 sysdev_create_file(&node->sysdev, &attr_vmstat);
281 scan_unevictable_register_node(node);
283 hugetlb_register_node(node);
285 compaction_register_node(node);
287 return error;
291 * unregister_node - unregister a node device
292 * @node: node going away
294 * Unregisters a node device @node. All the devices on the node must be
295 * unregistered before calling this function.
297 void unregister_node(struct node *node)
299 sysdev_remove_file(&node->sysdev, &attr_cpumap);
300 sysdev_remove_file(&node->sysdev, &attr_cpulist);
301 sysdev_remove_file(&node->sysdev, &attr_meminfo);
302 sysdev_remove_file(&node->sysdev, &attr_numastat);
303 sysdev_remove_file(&node->sysdev, &attr_distance);
304 sysdev_remove_file(&node->sysdev, &attr_vmstat);
306 scan_unevictable_unregister_node(node);
307 hugetlb_unregister_node(node); /* no-op, if memoryless node */
309 sysdev_unregister(&node->sysdev);
312 struct node node_devices[MAX_NUMNODES];
315 * register cpu under node
317 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
319 int ret;
320 struct sys_device *obj;
322 if (!node_online(nid))
323 return 0;
325 obj = get_cpu_sysdev(cpu);
326 if (!obj)
327 return 0;
329 ret = sysfs_create_link(&node_devices[nid].sysdev.kobj,
330 &obj->kobj,
331 kobject_name(&obj->kobj));
332 if (ret)
333 return ret;
335 return sysfs_create_link(&obj->kobj,
336 &node_devices[nid].sysdev.kobj,
337 kobject_name(&node_devices[nid].sysdev.kobj));
340 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
342 struct sys_device *obj;
344 if (!node_online(nid))
345 return 0;
347 obj = get_cpu_sysdev(cpu);
348 if (!obj)
349 return 0;
351 sysfs_remove_link(&node_devices[nid].sysdev.kobj,
352 kobject_name(&obj->kobj));
353 sysfs_remove_link(&obj->kobj,
354 kobject_name(&node_devices[nid].sysdev.kobj));
356 return 0;
359 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
360 #define page_initialized(page) (page->lru.next)
362 static int get_nid_for_pfn(unsigned long pfn)
364 struct page *page;
366 if (!pfn_valid_within(pfn))
367 return -1;
368 page = pfn_to_page(pfn);
369 if (!page_initialized(page))
370 return -1;
371 return pfn_to_nid(pfn);
374 /* register memory section under specified node if it spans that node */
375 int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
377 int ret;
378 unsigned long pfn, sect_start_pfn, sect_end_pfn;
380 if (!mem_blk)
381 return -EFAULT;
382 if (!node_online(nid))
383 return 0;
385 sect_start_pfn = section_nr_to_pfn(mem_blk->start_section_nr);
386 sect_end_pfn = section_nr_to_pfn(mem_blk->end_section_nr);
387 sect_end_pfn += PAGES_PER_SECTION - 1;
388 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
389 int page_nid;
391 page_nid = get_nid_for_pfn(pfn);
392 if (page_nid < 0)
393 continue;
394 if (page_nid != nid)
395 continue;
396 ret = sysfs_create_link_nowarn(&node_devices[nid].sysdev.kobj,
397 &mem_blk->sysdev.kobj,
398 kobject_name(&mem_blk->sysdev.kobj));
399 if (ret)
400 return ret;
402 return sysfs_create_link_nowarn(&mem_blk->sysdev.kobj,
403 &node_devices[nid].sysdev.kobj,
404 kobject_name(&node_devices[nid].sysdev.kobj));
406 /* mem section does not span the specified node */
407 return 0;
410 /* unregister memory section under all nodes that it spans */
411 int unregister_mem_sect_under_nodes(struct memory_block *mem_blk,
412 unsigned long phys_index)
414 NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
415 unsigned long pfn, sect_start_pfn, sect_end_pfn;
417 if (!mem_blk) {
418 NODEMASK_FREE(unlinked_nodes);
419 return -EFAULT;
421 if (!unlinked_nodes)
422 return -ENOMEM;
423 nodes_clear(*unlinked_nodes);
425 sect_start_pfn = section_nr_to_pfn(phys_index);
426 sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
427 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
428 int nid;
430 nid = get_nid_for_pfn(pfn);
431 if (nid < 0)
432 continue;
433 if (!node_online(nid))
434 continue;
435 if (node_test_and_set(nid, *unlinked_nodes))
436 continue;
437 sysfs_remove_link(&node_devices[nid].sysdev.kobj,
438 kobject_name(&mem_blk->sysdev.kobj));
439 sysfs_remove_link(&mem_blk->sysdev.kobj,
440 kobject_name(&node_devices[nid].sysdev.kobj));
442 NODEMASK_FREE(unlinked_nodes);
443 return 0;
446 static int link_mem_sections(int nid)
448 unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
449 unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
450 unsigned long pfn;
451 struct memory_block *mem_blk = NULL;
452 int err = 0;
454 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
455 unsigned long section_nr = pfn_to_section_nr(pfn);
456 struct mem_section *mem_sect;
457 int ret;
459 if (!present_section_nr(section_nr))
460 continue;
461 mem_sect = __nr_to_section(section_nr);
462 mem_blk = find_memory_block_hinted(mem_sect, mem_blk);
463 ret = register_mem_sect_under_node(mem_blk, nid);
464 if (!err)
465 err = ret;
467 /* discard ref obtained in find_memory_block() */
470 if (mem_blk)
471 kobject_put(&mem_blk->sysdev.kobj);
472 return err;
475 #ifdef CONFIG_HUGETLBFS
477 * Handle per node hstate attribute [un]registration on transistions
478 * to/from memoryless state.
480 static void node_hugetlb_work(struct work_struct *work)
482 struct node *node = container_of(work, struct node, node_work);
485 * We only get here when a node transitions to/from memoryless state.
486 * We can detect which transition occurred by examining whether the
487 * node has memory now. hugetlb_register_node() already check this
488 * so we try to register the attributes. If that fails, then the
489 * node has transitioned to memoryless, try to unregister the
490 * attributes.
492 if (!hugetlb_register_node(node))
493 hugetlb_unregister_node(node);
496 static void init_node_hugetlb_work(int nid)
498 INIT_WORK(&node_devices[nid].node_work, node_hugetlb_work);
501 static int node_memory_callback(struct notifier_block *self,
502 unsigned long action, void *arg)
504 struct memory_notify *mnb = arg;
505 int nid = mnb->status_change_nid;
507 switch (action) {
508 case MEM_ONLINE:
509 case MEM_OFFLINE:
511 * offload per node hstate [un]registration to a work thread
512 * when transitioning to/from memoryless state.
514 if (nid != NUMA_NO_NODE)
515 schedule_work(&node_devices[nid].node_work);
516 break;
518 case MEM_GOING_ONLINE:
519 case MEM_GOING_OFFLINE:
520 case MEM_CANCEL_ONLINE:
521 case MEM_CANCEL_OFFLINE:
522 default:
523 break;
526 return NOTIFY_OK;
528 #endif /* CONFIG_HUGETLBFS */
529 #else /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
531 static int link_mem_sections(int nid) { return 0; }
532 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
534 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
535 !defined(CONFIG_HUGETLBFS)
536 static inline int node_memory_callback(struct notifier_block *self,
537 unsigned long action, void *arg)
539 return NOTIFY_OK;
542 static void init_node_hugetlb_work(int nid) { }
544 #endif
546 int register_one_node(int nid)
548 int error = 0;
549 int cpu;
551 if (node_online(nid)) {
552 int p_node = parent_node(nid);
553 struct node *parent = NULL;
555 if (p_node != nid)
556 parent = &node_devices[p_node];
558 error = register_node(&node_devices[nid], nid, parent);
560 /* link cpu under this node */
561 for_each_present_cpu(cpu) {
562 if (cpu_to_node(cpu) == nid)
563 register_cpu_under_node(cpu, nid);
566 /* link memory sections under this node */
567 error = link_mem_sections(nid);
569 /* initialize work queue for memory hot plug */
570 init_node_hugetlb_work(nid);
573 return error;
577 void unregister_one_node(int nid)
579 unregister_node(&node_devices[nid]);
583 * node states attributes
586 static ssize_t print_nodes_state(enum node_states state, char *buf)
588 int n;
590 n = nodelist_scnprintf(buf, PAGE_SIZE, node_states[state]);
591 if (n > 0 && PAGE_SIZE > n + 1) {
592 *(buf + n++) = '\n';
593 *(buf + n++) = '\0';
595 return n;
598 struct node_attr {
599 struct sysdev_class_attribute attr;
600 enum node_states state;
603 static ssize_t show_node_state(struct sysdev_class *class,
604 struct sysdev_class_attribute *attr, char *buf)
606 struct node_attr *na = container_of(attr, struct node_attr, attr);
607 return print_nodes_state(na->state, buf);
610 #define _NODE_ATTR(name, state) \
611 { _SYSDEV_CLASS_ATTR(name, 0444, show_node_state, NULL), state }
613 static struct node_attr node_state_attr[] = {
614 _NODE_ATTR(possible, N_POSSIBLE),
615 _NODE_ATTR(online, N_ONLINE),
616 _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
617 _NODE_ATTR(has_cpu, N_CPU),
618 #ifdef CONFIG_HIGHMEM
619 _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
620 #endif
623 static struct sysdev_class_attribute *node_state_attrs[] = {
624 &node_state_attr[0].attr,
625 &node_state_attr[1].attr,
626 &node_state_attr[2].attr,
627 &node_state_attr[3].attr,
628 #ifdef CONFIG_HIGHMEM
629 &node_state_attr[4].attr,
630 #endif
631 NULL
634 #define NODE_CALLBACK_PRI 2 /* lower than SLAB */
635 static int __init register_node_type(void)
637 int ret;
639 BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
640 BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
642 ret = sysdev_class_register(&node_class);
643 if (!ret) {
644 hotplug_memory_notifier(node_memory_callback,
645 NODE_CALLBACK_PRI);
649 * Note: we're not going to unregister the node class if we fail
650 * to register the node state class attribute files.
652 return ret;
654 postcore_initcall(register_node_type);