blob: f8f578a71b25c311c31c27989789d5e2fe53f9ef [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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>
Gary Hadec04fc582009-01-06 14:39:14 -08009#include <linux/memory.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/node.h>
11#include <linux/hugetlb.h>
12#include <linux/cpumask.h>
13#include <linux/topology.h>
14#include <linux/nodemask.h>
KAMEZAWA Hiroyuki76b67ed2006-06-27 02:53:41 -070015#include <linux/cpu.h>
Lee Schermerhornbde631a2007-10-16 01:26:27 -070016#include <linux/device.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070017#include <linux/swap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018
19static struct sysdev_class node_class = {
Kay Sieversaf5ca3f42007-12-20 02:09:39 +010020 .name = "node",
Linus Torvalds1da177e2005-04-16 15:20:36 -070021};
22
23
Mike Travis39106dc2008-04-08 11:43:03 -070024static ssize_t node_read_cpumap(struct sys_device *dev, int type, char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -070025{
26 struct node *node_dev = to_node(dev);
Mike Travisc5f59f02008-04-04 18:11:10 -070027 node_to_cpumask_ptr(mask, node_dev->sysdev.id);
Linus Torvalds1da177e2005-04-16 15:20:36 -070028 int len;
29
Mike Travis39106dc2008-04-08 11:43:03 -070030 /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
31 BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1));
Linus Torvalds1da177e2005-04-16 15:20:36 -070032
Mike Travis39106dc2008-04-08 11:43:03 -070033 len = type?
Rusty Russell29c01772008-12-13 21:20:25 +103034 cpulist_scnprintf(buf, PAGE_SIZE-2, mask) :
35 cpumask_scnprintf(buf, PAGE_SIZE-2, mask);
Mike Travisc5f59f02008-04-04 18:11:10 -070036 buf[len++] = '\n';
37 buf[len] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -070038 return len;
39}
40
Andi Kleen4a0b2b42008-07-01 18:48:41 +020041static inline ssize_t node_read_cpumask(struct sys_device *dev,
42 struct sysdev_attribute *attr, char *buf)
Mike Travis39106dc2008-04-08 11:43:03 -070043{
44 return node_read_cpumap(dev, 0, buf);
45}
Andi Kleen4a0b2b42008-07-01 18:48:41 +020046static inline ssize_t node_read_cpulist(struct sys_device *dev,
47 struct sysdev_attribute *attr, char *buf)
Mike Travis39106dc2008-04-08 11:43:03 -070048{
49 return node_read_cpumap(dev, 1, buf);
50}
51
52static SYSDEV_ATTR(cpumap, S_IRUGO, node_read_cpumask, NULL);
53static SYSDEV_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -070054
55#define K(x) ((x) << (PAGE_SHIFT - 10))
Andi Kleen4a0b2b42008-07-01 18:48:41 +020056static ssize_t node_read_meminfo(struct sys_device * dev,
57 struct sysdev_attribute *attr, char * buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -070058{
59 int n;
60 int nid = dev->id;
61 struct sysinfo i;
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
63 si_meminfo_node(&i, nid);
Martin Hicksc07e02d2005-09-03 15:55:11 -070064
Linus Torvalds1da177e2005-04-16 15:20:36 -070065 n = sprintf(buf, "\n"
Rik van Riel4f98a2f2008-10-18 20:26:32 -070066 "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"
Lee Schermerhorn7b854122008-10-18 20:26:40 -070075#ifdef CONFIG_UNEVICTABLE_LRU
Nick Piggin5344b7e2008-10-18 20:26:51 -070076 "Node %d Unevictable: %8lu kB\n"
77 "Node %d Mlocked: %8lu kB\n"
Lee Schermerhorn7b854122008-10-18 20:26:40 -070078#endif
Christoph Lameter182e8e22006-09-25 23:31:10 -070079#ifdef CONFIG_HIGHMEM
Rik van Riel4f98a2f2008-10-18 20:26:32 -070080 "Node %d HighTotal: %8lu kB\n"
81 "Node %d HighFree: %8lu kB\n"
82 "Node %d LowTotal: %8lu kB\n"
83 "Node %d LowFree: %8lu kB\n"
Christoph Lameter182e8e22006-09-25 23:31:10 -070084#endif
Rik van Riel4f98a2f2008-10-18 20:26:32 -070085 "Node %d Dirty: %8lu kB\n"
86 "Node %d Writeback: %8lu kB\n"
87 "Node %d FilePages: %8lu kB\n"
88 "Node %d Mapped: %8lu kB\n"
89 "Node %d AnonPages: %8lu kB\n"
90 "Node %d PageTables: %8lu kB\n"
91 "Node %d NFS_Unstable: %8lu kB\n"
92 "Node %d Bounce: %8lu kB\n"
93 "Node %d WritebackTmp: %8lu kB\n"
94 "Node %d Slab: %8lu kB\n"
95 "Node %d SReclaimable: %8lu kB\n"
96 "Node %d SUnreclaim: %8lu kB\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -070097 nid, K(i.totalram),
98 nid, K(i.freeram),
99 nid, K(i.totalram - i.freeram),
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700100 nid, K(node_page_state(nid, NR_ACTIVE_ANON) +
101 node_page_state(nid, NR_ACTIVE_FILE)),
102 nid, K(node_page_state(nid, NR_INACTIVE_ANON) +
103 node_page_state(nid, NR_INACTIVE_FILE)),
104 nid, K(node_page_state(nid, NR_ACTIVE_ANON)),
105 nid, K(node_page_state(nid, NR_INACTIVE_ANON)),
106 nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
107 nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
Lee Schermerhorn7b854122008-10-18 20:26:40 -0700108#ifdef CONFIG_UNEVICTABLE_LRU
109 nid, K(node_page_state(nid, NR_UNEVICTABLE)),
Nick Piggin5344b7e2008-10-18 20:26:51 -0700110 nid, K(node_page_state(nid, NR_MLOCK)),
Lee Schermerhorn7b854122008-10-18 20:26:40 -0700111#endif
Christoph Lameter182e8e22006-09-25 23:31:10 -0700112#ifdef CONFIG_HIGHMEM
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 nid, K(i.totalhigh),
114 nid, K(i.freehigh),
115 nid, K(i.totalram - i.totalhigh),
Martin Hicksc07e02d2005-09-03 15:55:11 -0700116 nid, K(i.freeram - i.freehigh),
Christoph Lameter182e8e22006-09-25 23:31:10 -0700117#endif
Christoph Lameterb1e7a8f2006-06-30 01:55:39 -0700118 nid, K(node_page_state(nid, NR_FILE_DIRTY)),
Christoph Lameterce866b32006-06-30 01:55:40 -0700119 nid, K(node_page_state(nid, NR_WRITEBACK)),
Christoph Lameter347ce432006-06-30 01:55:35 -0700120 nid, K(node_page_state(nid, NR_FILE_PAGES)),
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700121 nid, K(node_page_state(nid, NR_FILE_MAPPED)),
Christoph Lameterf3dbd342006-06-30 01:55:36 -0700122 nid, K(node_page_state(nid, NR_ANON_PAGES)),
Christoph Lameterdf849a12006-06-30 01:55:38 -0700123 nid, K(node_page_state(nid, NR_PAGETABLE)),
Christoph Lameterfd39fc852006-06-30 01:55:40 -0700124 nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
Christoph Lameterd2c5e302006-06-30 01:55:41 -0700125 nid, K(node_page_state(nid, NR_BOUNCE)),
Miklos Szeredifc3ba692008-04-30 00:54:38 -0700126 nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
Christoph Lameter972d1a72006-09-25 23:31:51 -0700127 nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
128 node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
129 nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
130 nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 n += hugetlb_report_node_meminfo(nid, buf + n);
132 return n;
133}
134
135#undef K
136static SYSDEV_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
137
Andi Kleen4a0b2b42008-07-01 18:48:41 +0200138static ssize_t node_read_numastat(struct sys_device * dev,
139 struct sysdev_attribute *attr, char * buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 return sprintf(buf,
142 "numa_hit %lu\n"
143 "numa_miss %lu\n"
144 "numa_foreign %lu\n"
145 "interleave_hit %lu\n"
146 "local_node %lu\n"
147 "other_node %lu\n",
Christoph Lameterca889e62006-06-30 01:55:44 -0700148 node_page_state(dev->id, NUMA_HIT),
149 node_page_state(dev->id, NUMA_MISS),
150 node_page_state(dev->id, NUMA_FOREIGN),
151 node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
152 node_page_state(dev->id, NUMA_LOCAL),
153 node_page_state(dev->id, NUMA_OTHER));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154}
155static SYSDEV_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
156
Andi Kleen4a0b2b42008-07-01 18:48:41 +0200157static ssize_t node_read_distance(struct sys_device * dev,
158 struct sysdev_attribute *attr, char * buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159{
160 int nid = dev->id;
161 int len = 0;
162 int i;
163
164 /* buf currently PAGE_SIZE, need ~4 chars per node */
165 BUILD_BUG_ON(MAX_NUMNODES*4 > PAGE_SIZE/2);
166
167 for_each_online_node(i)
168 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
169
170 len += sprintf(buf + len, "\n");
171 return len;
172}
173static SYSDEV_ATTR(distance, S_IRUGO, node_read_distance, NULL);
174
175
176/*
Robert P. J. Day405ae7d2007-02-17 19:13:42 +0100177 * register_node - Setup a sysfs device for a node.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 * @num - Node number to use when creating the device.
179 *
180 * Initialize and register the node device.
181 */
Keiichiro Tokunaga4b450992005-05-08 21:28:53 +0900182int register_node(struct node *node, int num, struct node *parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183{
184 int error;
185
186 node->sysdev.id = num;
187 node->sysdev.cls = &node_class;
188 error = sysdev_register(&node->sysdev);
189
190 if (!error){
191 sysdev_create_file(&node->sysdev, &attr_cpumap);
Mike Travis39106dc2008-04-08 11:43:03 -0700192 sysdev_create_file(&node->sysdev, &attr_cpulist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193 sysdev_create_file(&node->sysdev, &attr_meminfo);
194 sysdev_create_file(&node->sysdev, &attr_numastat);
195 sysdev_create_file(&node->sysdev, &attr_distance);
Lee Schermerhornaf936a12008-10-18 20:26:53 -0700196
197 scan_unevictable_register_node(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198 }
199 return error;
200}
201
Keiichiro Tokunaga4b450992005-05-08 21:28:53 +0900202/**
203 * unregister_node - unregister a node device
204 * @node: node going away
205 *
206 * Unregisters a node device @node. All the devices on the node must be
207 * unregistered before calling this function.
208 */
209void unregister_node(struct node *node)
210{
211 sysdev_remove_file(&node->sysdev, &attr_cpumap);
Mike Travis39106dc2008-04-08 11:43:03 -0700212 sysdev_remove_file(&node->sysdev, &attr_cpulist);
Keiichiro Tokunaga4b450992005-05-08 21:28:53 +0900213 sysdev_remove_file(&node->sysdev, &attr_meminfo);
214 sysdev_remove_file(&node->sysdev, &attr_numastat);
215 sysdev_remove_file(&node->sysdev, &attr_distance);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216
Lee Schermerhornaf936a12008-10-18 20:26:53 -0700217 scan_unevictable_unregister_node(node);
218
Keiichiro Tokunaga4b450992005-05-08 21:28:53 +0900219 sysdev_unregister(&node->sysdev);
220}
221
Yasunori Goto0fc44152006-06-27 02:53:38 -0700222struct node node_devices[MAX_NUMNODES];
223
KAMEZAWA Hiroyuki76b67ed2006-06-27 02:53:41 -0700224/*
225 * register cpu under node
226 */
227int register_cpu_under_node(unsigned int cpu, unsigned int nid)
228{
229 if (node_online(nid)) {
230 struct sys_device *obj = get_cpu_sysdev(cpu);
231 if (!obj)
232 return 0;
233 return sysfs_create_link(&node_devices[nid].sysdev.kobj,
234 &obj->kobj,
235 kobject_name(&obj->kobj));
236 }
237
238 return 0;
239}
240
241int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
242{
243 if (node_online(nid)) {
244 struct sys_device *obj = get_cpu_sysdev(cpu);
245 if (obj)
246 sysfs_remove_link(&node_devices[nid].sysdev.kobj,
247 kobject_name(&obj->kobj));
248 }
249 return 0;
250}
251
Gary Hadec04fc582009-01-06 14:39:14 -0800252#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
253#define page_initialized(page) (page->lru.next)
254
255static int get_nid_for_pfn(unsigned long pfn)
256{
257 struct page *page;
258
259 if (!pfn_valid_within(pfn))
260 return -1;
261 page = pfn_to_page(pfn);
262 if (!page_initialized(page))
263 return -1;
264 return pfn_to_nid(pfn);
265}
266
267/* register memory section under specified node if it spans that node */
268int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
269{
270 unsigned long pfn, sect_start_pfn, sect_end_pfn;
271
272 if (!mem_blk)
273 return -EFAULT;
274 if (!node_online(nid))
275 return 0;
276 sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
277 sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
278 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
279 int page_nid;
280
281 page_nid = get_nid_for_pfn(pfn);
282 if (page_nid < 0)
283 continue;
284 if (page_nid != nid)
285 continue;
286 return sysfs_create_link_nowarn(&node_devices[nid].sysdev.kobj,
287 &mem_blk->sysdev.kobj,
288 kobject_name(&mem_blk->sysdev.kobj));
289 }
290 /* mem section does not span the specified node */
291 return 0;
292}
293
294/* unregister memory section under all nodes that it spans */
295int unregister_mem_sect_under_nodes(struct memory_block *mem_blk)
296{
297 nodemask_t unlinked_nodes;
298 unsigned long pfn, sect_start_pfn, sect_end_pfn;
299
300 if (!mem_blk)
301 return -EFAULT;
302 nodes_clear(unlinked_nodes);
303 sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
304 sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
305 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
Roel Kluin47504982009-03-10 12:55:45 -0700306 int nid;
Gary Hadec04fc582009-01-06 14:39:14 -0800307
308 nid = get_nid_for_pfn(pfn);
309 if (nid < 0)
310 continue;
311 if (!node_online(nid))
312 continue;
313 if (node_test_and_set(nid, unlinked_nodes))
314 continue;
315 sysfs_remove_link(&node_devices[nid].sysdev.kobj,
316 kobject_name(&mem_blk->sysdev.kobj));
317 }
318 return 0;
319}
320
321static int link_mem_sections(int nid)
322{
323 unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
324 unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
325 unsigned long pfn;
326 int err = 0;
327
328 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
329 unsigned long section_nr = pfn_to_section_nr(pfn);
330 struct mem_section *mem_sect;
331 struct memory_block *mem_blk;
332 int ret;
333
334 if (!present_section_nr(section_nr))
335 continue;
336 mem_sect = __nr_to_section(section_nr);
337 mem_blk = find_memory_block(mem_sect);
338 ret = register_mem_sect_under_node(mem_blk, nid);
339 if (!err)
340 err = ret;
341
342 /* discard ref obtained in find_memory_block() */
343 kobject_put(&mem_blk->sysdev.kobj);
344 }
345 return err;
346}
347#else
348static int link_mem_sections(int nid) { return 0; }
349#endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
350
Yasunori Goto0fc44152006-06-27 02:53:38 -0700351int register_one_node(int nid)
352{
353 int error = 0;
KAMEZAWA Hiroyuki76b67ed2006-06-27 02:53:41 -0700354 int cpu;
Yasunori Goto0fc44152006-06-27 02:53:38 -0700355
356 if (node_online(nid)) {
357 int p_node = parent_node(nid);
358 struct node *parent = NULL;
359
360 if (p_node != nid)
361 parent = &node_devices[p_node];
362
363 error = register_node(&node_devices[nid], nid, parent);
KAMEZAWA Hiroyuki76b67ed2006-06-27 02:53:41 -0700364
365 /* link cpu under this node */
366 for_each_present_cpu(cpu) {
367 if (cpu_to_node(cpu) == nid)
368 register_cpu_under_node(cpu, nid);
369 }
Gary Hadec04fc582009-01-06 14:39:14 -0800370
371 /* link memory sections under this node */
372 error = link_mem_sections(nid);
Yasunori Goto0fc44152006-06-27 02:53:38 -0700373 }
374
375 return error;
376
377}
378
379void unregister_one_node(int nid)
380{
381 unregister_node(&node_devices[nid]);
382}
383
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700384/*
385 * node states attributes
386 */
387
388static ssize_t print_nodes_state(enum node_states state, char *buf)
389{
390 int n;
391
392 n = nodelist_scnprintf(buf, PAGE_SIZE, node_states[state]);
393 if (n > 0 && PAGE_SIZE > n + 1) {
394 *(buf + n++) = '\n';
395 *(buf + n++) = '\0';
396 }
397 return n;
398}
399
400static ssize_t print_nodes_possible(struct sysdev_class *class, char *buf)
401{
402 return print_nodes_state(N_POSSIBLE, buf);
403}
404
405static ssize_t print_nodes_online(struct sysdev_class *class, char *buf)
406{
407 return print_nodes_state(N_ONLINE, buf);
408}
409
410static ssize_t print_nodes_has_normal_memory(struct sysdev_class *class,
411 char *buf)
412{
413 return print_nodes_state(N_NORMAL_MEMORY, buf);
414}
415
416static ssize_t print_nodes_has_cpu(struct sysdev_class *class, char *buf)
417{
418 return print_nodes_state(N_CPU, buf);
419}
420
421static SYSDEV_CLASS_ATTR(possible, 0444, print_nodes_possible, NULL);
422static SYSDEV_CLASS_ATTR(online, 0444, print_nodes_online, NULL);
423static SYSDEV_CLASS_ATTR(has_normal_memory, 0444, print_nodes_has_normal_memory,
424 NULL);
425static SYSDEV_CLASS_ATTR(has_cpu, 0444, print_nodes_has_cpu, NULL);
426
427#ifdef CONFIG_HIGHMEM
428static ssize_t print_nodes_has_high_memory(struct sysdev_class *class,
429 char *buf)
430{
431 return print_nodes_state(N_HIGH_MEMORY, buf);
432}
433
434static SYSDEV_CLASS_ATTR(has_high_memory, 0444, print_nodes_has_high_memory,
435 NULL);
436#endif
437
438struct sysdev_class_attribute *node_state_attr[] = {
439 &attr_possible,
440 &attr_online,
441 &attr_has_normal_memory,
442#ifdef CONFIG_HIGHMEM
443 &attr_has_high_memory,
444#endif
445 &attr_has_cpu,
446};
447
448static int node_states_init(void)
449{
450 int i;
451 int err = 0;
452
453 for (i = 0; i < NR_NODE_STATES; i++) {
454 int ret;
455 ret = sysdev_class_create_file(&node_class, node_state_attr[i]);
456 if (!err)
457 err = ret;
458 }
459 return err;
460}
461
Keiichiro Tokunaga4b450992005-05-08 21:28:53 +0900462static int __init register_node_type(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463{
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700464 int ret;
465
466 ret = sysdev_class_register(&node_class);
467 if (!ret)
468 ret = node_states_init();
469
470 /*
471 * Note: we're not going to unregister the node class if we fail
472 * to register the node state class attribute files.
473 */
474 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475}
476postcore_initcall(register_node_type);