2 * drivers/base/node.c - basic Node class support
5 #include <linux/sysdev.h>
6 #include <linux/module.h>
7 #include <linux/init.h>
9 #include <linux/memory.h>
10 #include <linux/node.h>
11 #include <linux/hugetlb.h>
12 #include <linux/cpumask.h>
13 #include <linux/topology.h>
14 #include <linux/nodemask.h>
15 #include <linux/cpu.h>
16 #include <linux/device.h>
17 #include <linux/swap.h>
19 static struct sysdev_class node_class = {
24 static ssize_t node_read_cpumap(struct sys_device *dev, int type, char *buf)
26 struct node *node_dev = to_node(dev);
27 const struct cpumask *mask = cpumask_of_node(node_dev->sysdev.id);
30 /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
31 BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1));
34 cpulist_scnprintf(buf, PAGE_SIZE-2, mask) :
35 cpumask_scnprintf(buf, PAGE_SIZE-2, mask);
41 static inline ssize_t node_read_cpumask(struct sys_device *dev,
42 struct sysdev_attribute *attr, char *buf)
44 return node_read_cpumap(dev, 0, buf);
46 static inline ssize_t node_read_cpulist(struct sys_device *dev,
47 struct sysdev_attribute *attr, char *buf)
49 return node_read_cpumap(dev, 1, buf);
52 static SYSDEV_ATTR(cpumap, S_IRUGO, node_read_cpumask, NULL);
53 static SYSDEV_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
55 #define K(x) ((x) << (PAGE_SHIFT - 10))
56 static ssize_t node_read_meminfo(struct sys_device * dev,
57 struct sysdev_attribute *attr, char * buf)
63 si_meminfo_node(&i, nid);
66 "Node %d MemTotal: %8lu kB\n"
67 "Node %d MemFree: %8lu kB\n"
68 "Node %d MemUsed: %8lu kB\n"
69 "Node %d Active: %8lu kB\n"
70 "Node %d Inactive: %8lu kB\n"
71 "Node %d Active(anon): %8lu kB\n"
72 "Node %d Inactive(anon): %8lu kB\n"
73 "Node %d Active(file): %8lu kB\n"
74 "Node %d Inactive(file): %8lu kB\n"
75 "Node %d Unevictable: %8lu kB\n"
76 "Node %d Mlocked: %8lu kB\n"
78 "Node %d HighTotal: %8lu kB\n"
79 "Node %d HighFree: %8lu kB\n"
80 "Node %d LowTotal: %8lu kB\n"
81 "Node %d LowFree: %8lu kB\n"
83 "Node %d Dirty: %8lu kB\n"
84 "Node %d Writeback: %8lu kB\n"
85 "Node %d FilePages: %8lu kB\n"
86 "Node %d Mapped: %8lu kB\n"
87 "Node %d AnonPages: %8lu kB\n"
88 "Node %d KernelStack: %8lu kB\n"
89 "Node %d PageTables: %8lu kB\n"
90 "Node %d NFS_Unstable: %8lu kB\n"
91 "Node %d Bounce: %8lu kB\n"
92 "Node %d WritebackTmp: %8lu kB\n"
93 "Node %d Slab: %8lu kB\n"
94 "Node %d SReclaimable: %8lu kB\n"
95 "Node %d SUnreclaim: %8lu kB\n",
98 nid, K(i.totalram - i.freeram),
99 nid, K(node_page_state(nid, NR_ACTIVE_ANON) +
100 node_page_state(nid, NR_ACTIVE_FILE)),
101 nid, K(node_page_state(nid, NR_INACTIVE_ANON) +
102 node_page_state(nid, NR_INACTIVE_FILE)),
103 nid, K(node_page_state(nid, NR_ACTIVE_ANON)),
104 nid, K(node_page_state(nid, NR_INACTIVE_ANON)),
105 nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
106 nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
107 nid, K(node_page_state(nid, NR_UNEVICTABLE)),
108 nid, K(node_page_state(nid, NR_MLOCK)),
109 #ifdef CONFIG_HIGHMEM
112 nid, K(i.totalram - i.totalhigh),
113 nid, K(i.freeram - i.freehigh),
115 nid, K(node_page_state(nid, NR_FILE_DIRTY)),
116 nid, K(node_page_state(nid, NR_WRITEBACK)),
117 nid, K(node_page_state(nid, NR_FILE_PAGES)),
118 nid, K(node_page_state(nid, NR_FILE_MAPPED)),
119 nid, K(node_page_state(nid, NR_ANON_PAGES)),
120 nid, node_page_state(nid, NR_KERNEL_STACK) *
122 nid, K(node_page_state(nid, NR_PAGETABLE)),
123 nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
124 nid, K(node_page_state(nid, NR_BOUNCE)),
125 nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
126 nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
127 node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
128 nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
129 nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
130 n += hugetlb_report_node_meminfo(nid, buf + n);
135 static SYSDEV_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
137 static ssize_t node_read_numastat(struct sys_device * dev,
138 struct sysdev_attribute *attr, char * buf)
144 "interleave_hit %lu\n"
147 node_page_state(dev->id, NUMA_HIT),
148 node_page_state(dev->id, NUMA_MISS),
149 node_page_state(dev->id, NUMA_FOREIGN),
150 node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
151 node_page_state(dev->id, NUMA_LOCAL),
152 node_page_state(dev->id, NUMA_OTHER));
154 static SYSDEV_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
156 static ssize_t node_read_distance(struct sys_device * dev,
157 struct sysdev_attribute *attr, char * buf)
163 /* buf currently PAGE_SIZE, need ~4 chars per node */
164 BUILD_BUG_ON(MAX_NUMNODES*4 > PAGE_SIZE/2);
166 for_each_online_node(i)
167 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
169 len += sprintf(buf + len, "\n");
172 static SYSDEV_ATTR(distance, S_IRUGO, node_read_distance, NULL);
176 * register_node - Setup a sysfs device for a node.
177 * @num - Node number to use when creating the device.
179 * Initialize and register the node device.
181 int register_node(struct node *node, int num, struct node *parent)
185 node->sysdev.id = num;
186 node->sysdev.cls = &node_class;
187 error = sysdev_register(&node->sysdev);
190 sysdev_create_file(&node->sysdev, &attr_cpumap);
191 sysdev_create_file(&node->sysdev, &attr_cpulist);
192 sysdev_create_file(&node->sysdev, &attr_meminfo);
193 sysdev_create_file(&node->sysdev, &attr_numastat);
194 sysdev_create_file(&node->sysdev, &attr_distance);
196 scan_unevictable_register_node(node);
202 * unregister_node - unregister a node device
203 * @node: node going away
205 * Unregisters a node device @node. All the devices on the node must be
206 * unregistered before calling this function.
208 void unregister_node(struct node *node)
210 sysdev_remove_file(&node->sysdev, &attr_cpumap);
211 sysdev_remove_file(&node->sysdev, &attr_cpulist);
212 sysdev_remove_file(&node->sysdev, &attr_meminfo);
213 sysdev_remove_file(&node->sysdev, &attr_numastat);
214 sysdev_remove_file(&node->sysdev, &attr_distance);
216 scan_unevictable_unregister_node(node);
218 sysdev_unregister(&node->sysdev);
221 struct node node_devices[MAX_NUMNODES];
224 * register cpu under node
226 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
228 if (node_online(nid)) {
229 struct sys_device *obj = get_cpu_sysdev(cpu);
232 return sysfs_create_link(&node_devices[nid].sysdev.kobj,
234 kobject_name(&obj->kobj));
240 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
242 if (node_online(nid)) {
243 struct sys_device *obj = get_cpu_sysdev(cpu);
245 sysfs_remove_link(&node_devices[nid].sysdev.kobj,
246 kobject_name(&obj->kobj));
251 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
252 #define page_initialized(page) (page->lru.next)
254 static int get_nid_for_pfn(unsigned long pfn)
258 if (!pfn_valid_within(pfn))
260 page = pfn_to_page(pfn);
261 if (!page_initialized(page))
263 return pfn_to_nid(pfn);
266 /* register memory section under specified node if it spans that node */
267 int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
269 unsigned long pfn, sect_start_pfn, sect_end_pfn;
273 if (!node_online(nid))
275 sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
276 sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
277 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
280 page_nid = get_nid_for_pfn(pfn);
285 return sysfs_create_link_nowarn(&node_devices[nid].sysdev.kobj,
286 &mem_blk->sysdev.kobj,
287 kobject_name(&mem_blk->sysdev.kobj));
289 /* mem section does not span the specified node */
293 /* unregister memory section under all nodes that it spans */
294 int unregister_mem_sect_under_nodes(struct memory_block *mem_blk)
296 nodemask_t unlinked_nodes;
297 unsigned long pfn, sect_start_pfn, sect_end_pfn;
301 nodes_clear(unlinked_nodes);
302 sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
303 sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
304 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
307 nid = get_nid_for_pfn(pfn);
310 if (!node_online(nid))
312 if (node_test_and_set(nid, unlinked_nodes))
314 sysfs_remove_link(&node_devices[nid].sysdev.kobj,
315 kobject_name(&mem_blk->sysdev.kobj));
320 static int link_mem_sections(int nid)
322 unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
323 unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
327 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
328 unsigned long section_nr = pfn_to_section_nr(pfn);
329 struct mem_section *mem_sect;
330 struct memory_block *mem_blk;
333 if (!present_section_nr(section_nr))
335 mem_sect = __nr_to_section(section_nr);
336 mem_blk = find_memory_block(mem_sect);
337 ret = register_mem_sect_under_node(mem_blk, nid);
341 /* discard ref obtained in find_memory_block() */
342 kobject_put(&mem_blk->sysdev.kobj);
347 static int link_mem_sections(int nid) { return 0; }
348 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
350 int register_one_node(int nid)
355 if (node_online(nid)) {
356 int p_node = parent_node(nid);
357 struct node *parent = NULL;
360 parent = &node_devices[p_node];
362 error = register_node(&node_devices[nid], nid, parent);
364 /* link cpu under this node */
365 for_each_present_cpu(cpu) {
366 if (cpu_to_node(cpu) == nid)
367 register_cpu_under_node(cpu, nid);
370 /* link memory sections under this node */
371 error = link_mem_sections(nid);
378 void unregister_one_node(int nid)
380 unregister_node(&node_devices[nid]);
384 * node states attributes
387 static ssize_t print_nodes_state(enum node_states state, char *buf)
391 n = nodelist_scnprintf(buf, PAGE_SIZE, node_states[state]);
392 if (n > 0 && PAGE_SIZE > n + 1) {
399 static ssize_t print_nodes_possible(struct sysdev_class *class, char *buf)
401 return print_nodes_state(N_POSSIBLE, buf);
404 static ssize_t print_nodes_online(struct sysdev_class *class, char *buf)
406 return print_nodes_state(N_ONLINE, buf);
409 static ssize_t print_nodes_has_normal_memory(struct sysdev_class *class,
412 return print_nodes_state(N_NORMAL_MEMORY, buf);
415 static ssize_t print_nodes_has_cpu(struct sysdev_class *class, char *buf)
417 return print_nodes_state(N_CPU, buf);
420 static SYSDEV_CLASS_ATTR(possible, 0444, print_nodes_possible, NULL);
421 static SYSDEV_CLASS_ATTR(online, 0444, print_nodes_online, NULL);
422 static SYSDEV_CLASS_ATTR(has_normal_memory, 0444, print_nodes_has_normal_memory,
424 static SYSDEV_CLASS_ATTR(has_cpu, 0444, print_nodes_has_cpu, NULL);
426 #ifdef CONFIG_HIGHMEM
427 static ssize_t print_nodes_has_high_memory(struct sysdev_class *class,
430 return print_nodes_state(N_HIGH_MEMORY, buf);
433 static SYSDEV_CLASS_ATTR(has_high_memory, 0444, print_nodes_has_high_memory,
437 struct sysdev_class_attribute *node_state_attr[] = {
440 &attr_has_normal_memory,
441 #ifdef CONFIG_HIGHMEM
442 &attr_has_high_memory,
447 static int node_states_init(void)
452 for (i = 0; i < NR_NODE_STATES; i++) {
454 ret = sysdev_class_create_file(&node_class, node_state_attr[i]);
461 static int __init register_node_type(void)
465 ret = sysdev_class_register(&node_class);
467 ret = node_states_init();
470 * Note: we're not going to unregister the node class if we fail
471 * to register the node state class attribute files.
475 postcore_initcall(register_node_type);