blob: 31df474d72f4a275ba0c87fb79a26c538fc661ca [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002 * Basic Node interface support
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 */
4
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/module.h>
6#include <linux/init.h>
7#include <linux/mm.h>
Gary Hadec04fc582009-01-06 14:39:14 -08008#include <linux/memory.h>
KOSAKI Motohirofa25c502011-05-24 17:11:28 -07009#include <linux/vmstat.h>
Andrew Morton6e259e72013-04-29 15:08:07 -070010#include <linux/notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070011#include <linux/node.h>
12#include <linux/hugetlb.h>
Mel Gormaned4a6d72010-05-24 14:32:29 -070013#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/cpumask.h>
15#include <linux/topology.h>
16#include <linux/nodemask.h>
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -070017#include <linux/cpu.h>
Lee Schermerhornbde631a2007-10-16 01:26:27 -070018#include <linux/device.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070019#include <linux/swap.h>
Tejun Heo18e5b532010-04-06 19:23:33 +090020#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021
Kay Sievers10fbcf42011-12-21 14:48:43 -080022static struct bus_type node_subsys = {
Kay Sieversaf5ca3f2007-12-20 02:09:39 +010023 .name = "node",
Kay Sievers10fbcf42011-12-21 14:48:43 -080024 .dev_name = "node",
Linus Torvalds1da177e2005-04-16 15:20:36 -070025};
26
27
Sudeep Holla5aaba362014-09-30 14:48:22 +010028static ssize_t node_read_cpumap(struct device *dev, bool list, char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -070029{
30 struct node *node_dev = to_node(dev);
Kay Sievers10fbcf42011-12-21 14:48:43 -080031 const struct cpumask *mask = cpumask_of_node(node_dev->dev.id);
Linus Torvalds1da177e2005-04-16 15:20:36 -070032
Mike Travis39106dc2008-04-08 11:43:03 -070033 /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
34 BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1));
Linus Torvalds1da177e2005-04-16 15:20:36 -070035
Sudeep Holla5aaba362014-09-30 14:48:22 +010036 return cpumap_print_to_pagebuf(list, buf, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -070037}
38
Kay Sievers10fbcf42011-12-21 14:48:43 -080039static inline ssize_t node_read_cpumask(struct device *dev,
40 struct device_attribute *attr, char *buf)
Mike Travis39106dc2008-04-08 11:43:03 -070041{
Sudeep Holla5aaba362014-09-30 14:48:22 +010042 return node_read_cpumap(dev, false, buf);
Mike Travis39106dc2008-04-08 11:43:03 -070043}
Kay Sievers10fbcf42011-12-21 14:48:43 -080044static inline ssize_t node_read_cpulist(struct device *dev,
45 struct device_attribute *attr, char *buf)
Mike Travis39106dc2008-04-08 11:43:03 -070046{
Sudeep Holla5aaba362014-09-30 14:48:22 +010047 return node_read_cpumap(dev, true, buf);
Mike Travis39106dc2008-04-08 11:43:03 -070048}
49
Kay Sievers10fbcf42011-12-21 14:48:43 -080050static DEVICE_ATTR(cpumap, S_IRUGO, node_read_cpumask, NULL);
51static DEVICE_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -070052
53#define K(x) ((x) << (PAGE_SHIFT - 10))
Kay Sievers10fbcf42011-12-21 14:48:43 -080054static ssize_t node_read_meminfo(struct device *dev,
55 struct device_attribute *attr, char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -070056{
57 int n;
58 int nid = dev->id;
59 struct sysinfo i;
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
61 si_meminfo_node(&i, nid);
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070062 n = sprintf(buf,
Rik van Riel4f98a2f2008-10-18 20:26:32 -070063 "Node %d MemTotal: %8lu kB\n"
64 "Node %d MemFree: %8lu kB\n"
65 "Node %d MemUsed: %8lu kB\n"
66 "Node %d Active: %8lu kB\n"
67 "Node %d Inactive: %8lu kB\n"
68 "Node %d Active(anon): %8lu kB\n"
69 "Node %d Inactive(anon): %8lu kB\n"
70 "Node %d Active(file): %8lu kB\n"
71 "Node %d Inactive(file): %8lu kB\n"
Nick Piggin5344b7e2008-10-18 20:26:51 -070072 "Node %d Unevictable: %8lu kB\n"
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070073 "Node %d Mlocked: %8lu kB\n",
74 nid, K(i.totalram),
75 nid, K(i.freeram),
76 nid, K(i.totalram - i.freeram),
77 nid, K(node_page_state(nid, NR_ACTIVE_ANON) +
78 node_page_state(nid, NR_ACTIVE_FILE)),
79 nid, K(node_page_state(nid, NR_INACTIVE_ANON) +
80 node_page_state(nid, NR_INACTIVE_FILE)),
81 nid, K(node_page_state(nid, NR_ACTIVE_ANON)),
82 nid, K(node_page_state(nid, NR_INACTIVE_ANON)),
83 nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
84 nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
85 nid, K(node_page_state(nid, NR_UNEVICTABLE)),
86 nid, K(node_page_state(nid, NR_MLOCK)));
87
Christoph Lameter182e8e22006-09-25 23:31:10 -070088#ifdef CONFIG_HIGHMEM
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070089 n += sprintf(buf + n,
Rik van Riel4f98a2f2008-10-18 20:26:32 -070090 "Node %d HighTotal: %8lu kB\n"
91 "Node %d HighFree: %8lu kB\n"
92 "Node %d LowTotal: %8lu kB\n"
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070093 "Node %d LowFree: %8lu kB\n",
94 nid, K(i.totalhigh),
95 nid, K(i.freehigh),
96 nid, K(i.totalram - i.totalhigh),
97 nid, K(i.freeram - i.freehigh));
Christoph Lameter182e8e22006-09-25 23:31:10 -070098#endif
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070099 n += sprintf(buf + n,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700100 "Node %d Dirty: %8lu kB\n"
101 "Node %d Writeback: %8lu kB\n"
102 "Node %d FilePages: %8lu kB\n"
103 "Node %d Mapped: %8lu kB\n"
104 "Node %d AnonPages: %8lu kB\n"
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700105 "Node %d Shmem: %8lu kB\n"
KOSAKI Motohiroc6a7f572009-09-21 17:01:32 -0700106 "Node %d KernelStack: %8lu kB\n"
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700107 "Node %d PageTables: %8lu kB\n"
108 "Node %d NFS_Unstable: %8lu kB\n"
109 "Node %d Bounce: %8lu kB\n"
110 "Node %d WritebackTmp: %8lu kB\n"
111 "Node %d Slab: %8lu kB\n"
112 "Node %d SReclaimable: %8lu kB\n"
David Rientjes05b258e2011-01-13 15:47:14 -0800113 "Node %d SUnreclaim: %8lu kB\n"
114#ifdef CONFIG_TRANSPARENT_HUGEPAGE
115 "Node %d AnonHugePages: %8lu kB\n"
116#endif
117 ,
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)),
Claudio Scordino91a13c22011-11-17 11:08:32 +0100122 nid, K(node_page_state(nid, NR_ANON_PAGES)),
Rafael Aquinicc7452b2014-08-06 16:06:38 -0700123 nid, K(i.sharedram),
KOSAKI Motohiroc6a7f572009-09-21 17:01:32 -0700124 nid, node_page_state(nid, NR_KERNEL_STACK) *
125 THREAD_SIZE / 1024,
Christoph Lameterdf849a12006-06-30 01:55:38 -0700126 nid, K(node_page_state(nid, NR_PAGETABLE)),
Christoph Lameterfd39fc82006-06-30 01:55:40 -0700127 nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
Christoph Lameterd2c5e302006-06-30 01:55:41 -0700128 nid, K(node_page_state(nid, NR_BOUNCE)),
Miklos Szeredifc3ba692008-04-30 00:54:38 -0700129 nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
Christoph Lameter972d1a72006-09-25 23:31:51 -0700130 nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
131 node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
132 nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
David Rientjes05b258e2011-01-13 15:47:14 -0800133#ifdef CONFIG_TRANSPARENT_HUGEPAGE
Claudio Scordino91a13c22011-11-17 11:08:32 +0100134 nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE))
David Rientjes05b258e2011-01-13 15:47:14 -0800135 , nid,
136 K(node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) *
Claudio Scordino91a13c22011-11-17 11:08:32 +0100137 HPAGE_PMD_NR));
138#else
139 nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
David Rientjes05b258e2011-01-13 15:47:14 -0800140#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 n += hugetlb_report_node_meminfo(nid, buf + n);
142 return n;
143}
144
145#undef K
Kay Sievers10fbcf42011-12-21 14:48:43 -0800146static DEVICE_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147
Kay Sievers10fbcf42011-12-21 14:48:43 -0800148static ssize_t node_read_numastat(struct device *dev,
149 struct device_attribute *attr, char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 return sprintf(buf,
152 "numa_hit %lu\n"
153 "numa_miss %lu\n"
154 "numa_foreign %lu\n"
155 "interleave_hit %lu\n"
156 "local_node %lu\n"
157 "other_node %lu\n",
Christoph Lameterca889e62006-06-30 01:55:44 -0700158 node_page_state(dev->id, NUMA_HIT),
159 node_page_state(dev->id, NUMA_MISS),
160 node_page_state(dev->id, NUMA_FOREIGN),
161 node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
162 node_page_state(dev->id, NUMA_LOCAL),
163 node_page_state(dev->id, NUMA_OTHER));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164}
Kay Sievers10fbcf42011-12-21 14:48:43 -0800165static DEVICE_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166
Kay Sievers10fbcf42011-12-21 14:48:43 -0800167static ssize_t node_read_vmstat(struct device *dev,
168 struct device_attribute *attr, char *buf)
Michael Rubin2ac39032010-10-26 14:21:35 -0700169{
170 int nid = dev->id;
KOSAKI Motohirofa25c502011-05-24 17:11:28 -0700171 int i;
172 int n = 0;
173
174 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
175 n += sprintf(buf+n, "%s %lu\n", vmstat_text[i],
176 node_page_state(nid, i));
177
178 return n;
Michael Rubin2ac39032010-10-26 14:21:35 -0700179}
Kay Sievers10fbcf42011-12-21 14:48:43 -0800180static DEVICE_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
Michael Rubin2ac39032010-10-26 14:21:35 -0700181
Kay Sievers10fbcf42011-12-21 14:48:43 -0800182static ssize_t node_read_distance(struct device *dev,
Ana Nedelcu518d3f32015-03-08 12:48:48 +0200183 struct device_attribute *attr, char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184{
185 int nid = dev->id;
186 int len = 0;
187 int i;
188
David Rientjes12ee3c02010-03-10 14:50:21 -0800189 /*
190 * buf is currently PAGE_SIZE in length and each node needs 4 chars
191 * at the most (distance + space or newline).
192 */
193 BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194
195 for_each_online_node(i)
196 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
197
198 len += sprintf(buf + len, "\n");
199 return len;
200}
Kay Sievers10fbcf42011-12-21 14:48:43 -0800201static DEVICE_ATTR(distance, S_IRUGO, node_read_distance, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202
Takashi Iwai3c9b8aa2015-01-29 12:29:22 +0100203static struct attribute *node_dev_attrs[] = {
204 &dev_attr_cpumap.attr,
205 &dev_attr_cpulist.attr,
206 &dev_attr_meminfo.attr,
207 &dev_attr_numastat.attr,
208 &dev_attr_distance.attr,
209 &dev_attr_vmstat.attr,
210 NULL
211};
Greg Kroah-Hartman7ca7ec42015-03-25 13:47:17 +0100212ATTRIBUTE_GROUPS(node_dev);
Takashi Iwai3c9b8aa2015-01-29 12:29:22 +0100213
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800214#ifdef CONFIG_HUGETLBFS
215/*
216 * hugetlbfs per node attributes registration interface:
217 * When/if hugetlb[fs] subsystem initializes [sometime after this module],
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800218 * it will register its per node attributes for all online nodes with
219 * memory. It will also call register_hugetlbfs_with_node(), below, to
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800220 * 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 */
224static node_registration_func_t __hugetlb_register_node;
225static node_registration_func_t __hugetlb_unregister_node;
226
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800227static inline bool hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800228{
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800229 if (__hugetlb_register_node &&
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -0800230 node_state(node->dev.id, N_MEMORY)) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800231 __hugetlb_register_node(node);
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800232 return true;
233 }
234 return false;
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800235}
236
237static inline void hugetlb_unregister_node(struct node *node)
238{
239 if (__hugetlb_unregister_node)
240 __hugetlb_unregister_node(node);
241}
242
243void 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
250static inline void hugetlb_register_node(struct node *node) {}
251
252static inline void hugetlb_unregister_node(struct node *node) {}
253#endif
254
Yasuaki Ishimatsu8c7b5b42012-12-11 16:00:57 -0800255static void node_device_release(struct device *dev)
256{
257 struct node *node = to_node(dev);
258
259#if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HUGETLBFS)
260 /*
261 * We schedule the work only when a memory section is
262 * onlined/offlined on this node. When we come here,
263 * all the memory on this node has been offlined,
264 * so we won't enqueue new work to this work.
265 *
266 * The work is using node->node_work, so we should
267 * flush work before freeing the memory.
268 */
269 flush_work(&node->node_work);
270#endif
271 kfree(node);
272}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
274/*
Robert P. J. Day405ae7d2007-02-17 19:13:42 +0100275 * register_node - Setup a sysfs device for a node.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 * @num - Node number to use when creating the device.
277 *
278 * Initialize and register the node device.
279 */
Yasuaki Ishimatsufa264372012-12-11 16:02:52 -0800280static int register_node(struct node *node, int num, struct node *parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281{
282 int error;
283
Kay Sievers10fbcf42011-12-21 14:48:43 -0800284 node->dev.id = num;
285 node->dev.bus = &node_subsys;
Yasuaki Ishimatsu8c7b5b42012-12-11 16:00:57 -0800286 node->dev.release = node_device_release;
Greg Kroah-Hartman7ca7ec42015-03-25 13:47:17 +0100287 node->dev.groups = node_dev_groups;
Kay Sievers10fbcf42011-12-21 14:48:43 -0800288 error = device_register(&node->dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289
290 if (!error){
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800291 hugetlb_register_node(node);
Mel Gormaned4a6d72010-05-24 14:32:29 -0700292
293 compaction_register_node(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294 }
295 return error;
296}
297
Keiichiro Tokunaga4b450992005-05-08 21:28:53 +0900298/**
299 * unregister_node - unregister a node device
300 * @node: node going away
301 *
302 * Unregisters a node device @node. All the devices on the node must be
303 * unregistered before calling this function.
304 */
305void unregister_node(struct node *node)
306{
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800307 hugetlb_unregister_node(node); /* no-op, if memoryless node */
Lee Schermerhornaf936a12008-10-18 20:26:53 -0700308
Kay Sievers10fbcf42011-12-21 14:48:43 -0800309 device_unregister(&node->dev);
Keiichiro Tokunaga4b450992005-05-08 21:28:53 +0900310}
311
Wen Congyang87327942012-12-11 16:00:56 -0800312struct node *node_devices[MAX_NUMNODES];
Yasunori Goto0fc44152006-06-27 02:53:38 -0700313
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700314/*
315 * register cpu under node
316 */
317int register_cpu_under_node(unsigned int cpu, unsigned int nid)
318{
Alex Chiang18307942009-12-14 17:59:08 -0800319 int ret;
Kay Sievers8a25a2f2011-12-21 14:29:42 -0800320 struct device *obj;
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700321
Alex Chiangf8246f32009-12-14 17:59:06 -0800322 if (!node_online(nid))
323 return 0;
324
Kay Sievers8a25a2f2011-12-21 14:29:42 -0800325 obj = get_cpu_device(cpu);
Alex Chiangf8246f32009-12-14 17:59:06 -0800326 if (!obj)
327 return 0;
328
Wen Congyang87327942012-12-11 16:00:56 -0800329 ret = sysfs_create_link(&node_devices[nid]->dev.kobj,
Alex Chiangf8246f32009-12-14 17:59:06 -0800330 &obj->kobj,
331 kobject_name(&obj->kobj));
Alex Chiang18307942009-12-14 17:59:08 -0800332 if (ret)
333 return ret;
334
335 return sysfs_create_link(&obj->kobj,
Wen Congyang87327942012-12-11 16:00:56 -0800336 &node_devices[nid]->dev.kobj,
337 kobject_name(&node_devices[nid]->dev.kobj));
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700338}
339
340int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
341{
Kay Sievers8a25a2f2011-12-21 14:29:42 -0800342 struct device *obj;
Alex Chiangb9d52da2009-12-14 17:59:07 -0800343
344 if (!node_online(nid))
345 return 0;
346
Kay Sievers8a25a2f2011-12-21 14:29:42 -0800347 obj = get_cpu_device(cpu);
Alex Chiangb9d52da2009-12-14 17:59:07 -0800348 if (!obj)
349 return 0;
350
Wen Congyang87327942012-12-11 16:00:56 -0800351 sysfs_remove_link(&node_devices[nid]->dev.kobj,
Alex Chiangb9d52da2009-12-14 17:59:07 -0800352 kobject_name(&obj->kobj));
Alex Chiang18307942009-12-14 17:59:08 -0800353 sysfs_remove_link(&obj->kobj,
Wen Congyang87327942012-12-11 16:00:56 -0800354 kobject_name(&node_devices[nid]->dev.kobj));
Alex Chiangb9d52da2009-12-14 17:59:07 -0800355
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700356 return 0;
357}
358
Gary Hadec04fc582009-01-06 14:39:14 -0800359#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
360#define page_initialized(page) (page->lru.next)
361
Mel Gorman3a80a7f2015-06-30 14:57:02 -0700362static int __init_refok get_nid_for_pfn(unsigned long pfn)
Gary Hadec04fc582009-01-06 14:39:14 -0800363{
364 struct page *page;
365
366 if (!pfn_valid_within(pfn))
367 return -1;
Mel Gorman3a80a7f2015-06-30 14:57:02 -0700368#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
369 if (system_state == SYSTEM_BOOTING)
370 return early_pfn_to_nid(pfn);
371#endif
Gary Hadec04fc582009-01-06 14:39:14 -0800372 page = pfn_to_page(pfn);
373 if (!page_initialized(page))
374 return -1;
375 return pfn_to_nid(pfn);
376}
377
378/* register memory section under specified node if it spans that node */
379int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
380{
Alex Chiangdee5d0d2009-12-14 17:59:05 -0800381 int ret;
Gary Hadec04fc582009-01-06 14:39:14 -0800382 unsigned long pfn, sect_start_pfn, sect_end_pfn;
383
384 if (!mem_blk)
385 return -EFAULT;
386 if (!node_online(nid))
387 return 0;
Nathan Fontenotd3360162011-01-20 10:44:29 -0600388
389 sect_start_pfn = section_nr_to_pfn(mem_blk->start_section_nr);
390 sect_end_pfn = section_nr_to_pfn(mem_blk->end_section_nr);
391 sect_end_pfn += PAGES_PER_SECTION - 1;
Gary Hadec04fc582009-01-06 14:39:14 -0800392 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
393 int page_nid;
394
395 page_nid = get_nid_for_pfn(pfn);
396 if (page_nid < 0)
397 continue;
398 if (page_nid != nid)
399 continue;
Wen Congyang87327942012-12-11 16:00:56 -0800400 ret = sysfs_create_link_nowarn(&node_devices[nid]->dev.kobj,
Kay Sievers10fbcf42011-12-21 14:48:43 -0800401 &mem_blk->dev.kobj,
402 kobject_name(&mem_blk->dev.kobj));
Alex Chiangdee5d0d2009-12-14 17:59:05 -0800403 if (ret)
404 return ret;
405
Kay Sievers10fbcf42011-12-21 14:48:43 -0800406 return sysfs_create_link_nowarn(&mem_blk->dev.kobj,
Wen Congyang87327942012-12-11 16:00:56 -0800407 &node_devices[nid]->dev.kobj,
408 kobject_name(&node_devices[nid]->dev.kobj));
Gary Hadec04fc582009-01-06 14:39:14 -0800409 }
410 /* mem section does not span the specified node */
411 return 0;
412}
413
414/* unregister memory section under all nodes that it spans */
Nathan Fontenotd3360162011-01-20 10:44:29 -0600415int unregister_mem_sect_under_nodes(struct memory_block *mem_blk,
416 unsigned long phys_index)
Gary Hadec04fc582009-01-06 14:39:14 -0800417{
David Rientjes9ae49fa2009-12-14 17:59:46 -0800418 NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
Gary Hadec04fc582009-01-06 14:39:14 -0800419 unsigned long pfn, sect_start_pfn, sect_end_pfn;
420
David Rientjes9ae49fa2009-12-14 17:59:46 -0800421 if (!mem_blk) {
422 NODEMASK_FREE(unlinked_nodes);
Gary Hadec04fc582009-01-06 14:39:14 -0800423 return -EFAULT;
David Rientjes9ae49fa2009-12-14 17:59:46 -0800424 }
425 if (!unlinked_nodes)
426 return -ENOMEM;
427 nodes_clear(*unlinked_nodes);
Nathan Fontenotd3360162011-01-20 10:44:29 -0600428
429 sect_start_pfn = section_nr_to_pfn(phys_index);
Gary Hadec04fc582009-01-06 14:39:14 -0800430 sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
431 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
Roel Kluin47504982009-03-10 12:55:45 -0700432 int nid;
Gary Hadec04fc582009-01-06 14:39:14 -0800433
434 nid = get_nid_for_pfn(pfn);
435 if (nid < 0)
436 continue;
437 if (!node_online(nid))
438 continue;
David Rientjes9ae49fa2009-12-14 17:59:46 -0800439 if (node_test_and_set(nid, *unlinked_nodes))
Gary Hadec04fc582009-01-06 14:39:14 -0800440 continue;
Wen Congyang87327942012-12-11 16:00:56 -0800441 sysfs_remove_link(&node_devices[nid]->dev.kobj,
Kay Sievers10fbcf42011-12-21 14:48:43 -0800442 kobject_name(&mem_blk->dev.kobj));
443 sysfs_remove_link(&mem_blk->dev.kobj,
Wen Congyang87327942012-12-11 16:00:56 -0800444 kobject_name(&node_devices[nid]->dev.kobj));
Gary Hadec04fc582009-01-06 14:39:14 -0800445 }
David Rientjes9ae49fa2009-12-14 17:59:46 -0800446 NODEMASK_FREE(unlinked_nodes);
Gary Hadec04fc582009-01-06 14:39:14 -0800447 return 0;
448}
449
450static int link_mem_sections(int nid)
451{
452 unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
453 unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
454 unsigned long pfn;
Robin Holt63d027a2010-09-29 14:00:56 -0500455 struct memory_block *mem_blk = NULL;
Gary Hadec04fc582009-01-06 14:39:14 -0800456 int err = 0;
457
458 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
459 unsigned long section_nr = pfn_to_section_nr(pfn);
460 struct mem_section *mem_sect;
Gary Hadec04fc582009-01-06 14:39:14 -0800461 int ret;
462
463 if (!present_section_nr(section_nr))
464 continue;
465 mem_sect = __nr_to_section(section_nr);
Yinghai Lu321bf4e2012-01-30 13:57:12 -0800466
467 /* same memblock ? */
468 if (mem_blk)
469 if ((section_nr >= mem_blk->start_section_nr) &&
470 (section_nr <= mem_blk->end_section_nr))
471 continue;
472
Robin Holt63d027a2010-09-29 14:00:56 -0500473 mem_blk = find_memory_block_hinted(mem_sect, mem_blk);
Yinghai Lu321bf4e2012-01-30 13:57:12 -0800474
Gary Hadec04fc582009-01-06 14:39:14 -0800475 ret = register_mem_sect_under_node(mem_blk, nid);
476 if (!err)
477 err = ret;
478
479 /* discard ref obtained in find_memory_block() */
Gary Hadec04fc582009-01-06 14:39:14 -0800480 }
Robin Holt63d027a2010-09-29 14:00:56 -0500481
482 if (mem_blk)
Kay Sievers10fbcf42011-12-21 14:48:43 -0800483 kobject_put(&mem_blk->dev.kobj);
Gary Hadec04fc582009-01-06 14:39:14 -0800484 return err;
485}
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800486
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800487#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800488/*
489 * Handle per node hstate attribute [un]registration on transistions
490 * to/from memoryless state.
491 */
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800492static void node_hugetlb_work(struct work_struct *work)
493{
494 struct node *node = container_of(work, struct node, node_work);
495
496 /*
497 * We only get here when a node transitions to/from memoryless state.
498 * We can detect which transition occurred by examining whether the
499 * node has memory now. hugetlb_register_node() already check this
500 * so we try to register the attributes. If that fails, then the
501 * node has transitioned to memoryless, try to unregister the
502 * attributes.
503 */
504 if (!hugetlb_register_node(node))
505 hugetlb_unregister_node(node);
506}
507
508static void init_node_hugetlb_work(int nid)
509{
Wen Congyang87327942012-12-11 16:00:56 -0800510 INIT_WORK(&node_devices[nid]->node_work, node_hugetlb_work);
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800511}
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800512
513static int node_memory_callback(struct notifier_block *self,
514 unsigned long action, void *arg)
515{
516 struct memory_notify *mnb = arg;
517 int nid = mnb->status_change_nid;
518
519 switch (action) {
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800520 case MEM_ONLINE:
521 case MEM_OFFLINE:
522 /*
523 * offload per node hstate [un]registration to a work thread
524 * when transitioning to/from memoryless state.
525 */
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800526 if (nid != NUMA_NO_NODE)
Wen Congyang87327942012-12-11 16:00:56 -0800527 schedule_work(&node_devices[nid]->node_work);
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800528 break;
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800529
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800530 case MEM_GOING_ONLINE:
531 case MEM_GOING_OFFLINE:
532 case MEM_CANCEL_ONLINE:
533 case MEM_CANCEL_OFFLINE:
534 default:
535 break;
536 }
537
538 return NOTIFY_OK;
539}
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800540#endif /* CONFIG_HUGETLBFS */
541#else /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800542
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800543static int link_mem_sections(int nid) { return 0; }
544#endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
545
546#if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
547 !defined(CONFIG_HUGETLBFS)
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800548static inline int node_memory_callback(struct notifier_block *self,
549 unsigned long action, void *arg)
550{
551 return NOTIFY_OK;
552}
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800553
554static void init_node_hugetlb_work(int nid) { }
555
556#endif
Gary Hadec04fc582009-01-06 14:39:14 -0800557
Yasunori Goto0fc44152006-06-27 02:53:38 -0700558int register_one_node(int nid)
559{
560 int error = 0;
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700561 int cpu;
Yasunori Goto0fc44152006-06-27 02:53:38 -0700562
563 if (node_online(nid)) {
564 int p_node = parent_node(nid);
565 struct node *parent = NULL;
566
567 if (p_node != nid)
Wen Congyang87327942012-12-11 16:00:56 -0800568 parent = node_devices[p_node];
Yasunori Goto0fc44152006-06-27 02:53:38 -0700569
Wen Congyang87327942012-12-11 16:00:56 -0800570 node_devices[nid] = kzalloc(sizeof(struct node), GFP_KERNEL);
571 if (!node_devices[nid])
572 return -ENOMEM;
573
574 error = register_node(node_devices[nid], nid, parent);
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700575
576 /* link cpu under this node */
577 for_each_present_cpu(cpu) {
578 if (cpu_to_node(cpu) == nid)
579 register_cpu_under_node(cpu, nid);
580 }
Gary Hadec04fc582009-01-06 14:39:14 -0800581
582 /* link memory sections under this node */
583 error = link_mem_sections(nid);
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800584
585 /* initialize work queue for memory hot plug */
586 init_node_hugetlb_work(nid);
Yasunori Goto0fc44152006-06-27 02:53:38 -0700587 }
588
589 return error;
590
591}
592
593void unregister_one_node(int nid)
594{
Xishi Qiu92d585e2014-03-06 17:18:21 +0800595 if (!node_devices[nid])
596 return;
597
Wen Congyang87327942012-12-11 16:00:56 -0800598 unregister_node(node_devices[nid]);
Wen Congyang87327942012-12-11 16:00:56 -0800599 node_devices[nid] = NULL;
Yasunori Goto0fc44152006-06-27 02:53:38 -0700600}
601
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700602/*
603 * node states attributes
604 */
605
606static ssize_t print_nodes_state(enum node_states state, char *buf)
607{
608 int n;
609
Tejun Heof799b1a2015-02-13 14:37:56 -0800610 n = scnprintf(buf, PAGE_SIZE - 1, "%*pbl",
611 nodemask_pr_args(&node_states[state]));
Ryota Ozakif6238812012-05-29 15:06:20 -0700612 buf[n++] = '\n';
613 buf[n] = '\0';
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700614 return n;
615}
616
Andi Kleenb15f5622010-01-05 12:47:59 +0100617struct node_attr {
Kay Sievers10fbcf42011-12-21 14:48:43 -0800618 struct device_attribute attr;
Andi Kleenb15f5622010-01-05 12:47:59 +0100619 enum node_states state;
620};
621
Kay Sievers10fbcf42011-12-21 14:48:43 -0800622static ssize_t show_node_state(struct device *dev,
623 struct device_attribute *attr, char *buf)
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700624{
Andi Kleenb15f5622010-01-05 12:47:59 +0100625 struct node_attr *na = container_of(attr, struct node_attr, attr);
626 return print_nodes_state(na->state, buf);
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700627}
628
Andi Kleenb15f5622010-01-05 12:47:59 +0100629#define _NODE_ATTR(name, state) \
Kay Sievers10fbcf42011-12-21 14:48:43 -0800630 { __ATTR(name, 0444, show_node_state, NULL), state }
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700631
Andi Kleenb15f5622010-01-05 12:47:59 +0100632static struct node_attr node_state_attr[] = {
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800633 [N_POSSIBLE] = _NODE_ATTR(possible, N_POSSIBLE),
634 [N_ONLINE] = _NODE_ATTR(online, N_ONLINE),
635 [N_NORMAL_MEMORY] = _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700636#ifdef CONFIG_HIGHMEM
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800637 [N_HIGH_MEMORY] = _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700638#endif
Lai Jiangshan20b2f522012-12-12 13:52:00 -0800639#ifdef CONFIG_MOVABLE_NODE
640 [N_MEMORY] = _NODE_ATTR(has_memory, N_MEMORY),
641#endif
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800642 [N_CPU] = _NODE_ATTR(has_cpu, N_CPU),
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700643};
644
Kay Sievers10fbcf42011-12-21 14:48:43 -0800645static struct attribute *node_state_attrs[] = {
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800646 &node_state_attr[N_POSSIBLE].attr.attr,
647 &node_state_attr[N_ONLINE].attr.attr,
648 &node_state_attr[N_NORMAL_MEMORY].attr.attr,
Andi Kleen3701cde2010-01-05 12:48:04 +0100649#ifdef CONFIG_HIGHMEM
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800650 &node_state_attr[N_HIGH_MEMORY].attr.attr,
Andi Kleen3701cde2010-01-05 12:48:04 +0100651#endif
Lai Jiangshan20b2f522012-12-12 13:52:00 -0800652#ifdef CONFIG_MOVABLE_NODE
653 &node_state_attr[N_MEMORY].attr.attr,
654#endif
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800655 &node_state_attr[N_CPU].attr.attr,
Andi Kleen3701cde2010-01-05 12:48:04 +0100656 NULL
657};
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700658
Kay Sievers10fbcf42011-12-21 14:48:43 -0800659static struct attribute_group memory_root_attr_group = {
660 .attrs = node_state_attrs,
661};
662
663static const struct attribute_group *cpu_root_attr_groups[] = {
664 &memory_root_attr_group,
665 NULL,
666};
667
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800668#define NODE_CALLBACK_PRI 2 /* lower than SLAB */
Keiichiro Tokunaga4b450992005-05-08 21:28:53 +0900669static int __init register_node_type(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670{
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700671 int ret;
672
Andi Kleen3701cde2010-01-05 12:48:04 +0100673 BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
674 BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
675
Kay Sievers10fbcf42011-12-21 14:48:43 -0800676 ret = subsys_system_register(&node_subsys, cpu_root_attr_groups);
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800677 if (!ret) {
Andrew Morton6e259e72013-04-29 15:08:07 -0700678 static struct notifier_block node_memory_callback_nb = {
679 .notifier_call = node_memory_callback,
680 .priority = NODE_CALLBACK_PRI,
681 };
682 register_hotmemory_notifier(&node_memory_callback_nb);
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800683 }
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700684
685 /*
686 * Note: we're not going to unregister the node class if we fail
687 * to register the node state class attribute files.
688 */
689 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690}
691postcore_initcall(register_node_type);