blob: bc9f43bf7e29a46714cb1f220bb0d866c5d3eab8 [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
Kay Sievers10fbcf42011-12-21 14:48:43 -080028static ssize_t node_read_cpumap(struct device *dev, int type, 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 int len;
33
Mike Travis39106dc2008-04-08 11:43:03 -070034 /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
35 BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1));
Linus Torvalds1da177e2005-04-16 15:20:36 -070036
Mike Travis39106dc2008-04-08 11:43:03 -070037 len = type?
Rusty Russell29c01772008-12-13 21:20:25 +103038 cpulist_scnprintf(buf, PAGE_SIZE-2, mask) :
39 cpumask_scnprintf(buf, PAGE_SIZE-2, mask);
Mike Travisc5f59f02008-04-04 18:11:10 -070040 buf[len++] = '\n';
41 buf[len] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -070042 return len;
43}
44
Kay Sievers10fbcf42011-12-21 14:48:43 -080045static inline ssize_t node_read_cpumask(struct device *dev,
46 struct device_attribute *attr, char *buf)
Mike Travis39106dc2008-04-08 11:43:03 -070047{
48 return node_read_cpumap(dev, 0, buf);
49}
Kay Sievers10fbcf42011-12-21 14:48:43 -080050static inline ssize_t node_read_cpulist(struct device *dev,
51 struct device_attribute *attr, char *buf)
Mike Travis39106dc2008-04-08 11:43:03 -070052{
53 return node_read_cpumap(dev, 1, buf);
54}
55
Kay Sievers10fbcf42011-12-21 14:48:43 -080056static DEVICE_ATTR(cpumap, S_IRUGO, node_read_cpumask, NULL);
57static DEVICE_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -070058
59#define K(x) ((x) << (PAGE_SHIFT - 10))
Kay Sievers10fbcf42011-12-21 14:48:43 -080060static ssize_t node_read_meminfo(struct device *dev,
61 struct device_attribute *attr, char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -070062{
63 int n;
64 int nid = dev->id;
65 struct sysinfo i;
Linus Torvalds1da177e2005-04-16 15:20:36 -070066
67 si_meminfo_node(&i, nid);
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070068 n = sprintf(buf,
Rik van Riel4f98a2f2008-10-18 20:26:32 -070069 "Node %d MemTotal: %8lu kB\n"
70 "Node %d MemFree: %8lu kB\n"
71 "Node %d MemUsed: %8lu kB\n"
72 "Node %d Active: %8lu kB\n"
73 "Node %d Inactive: %8lu kB\n"
74 "Node %d Active(anon): %8lu kB\n"
75 "Node %d Inactive(anon): %8lu kB\n"
76 "Node %d Active(file): %8lu kB\n"
77 "Node %d Inactive(file): %8lu kB\n"
Nick Piggin5344b7e2008-10-18 20:26:51 -070078 "Node %d Unevictable: %8lu kB\n"
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070079 "Node %d Mlocked: %8lu kB\n",
80 nid, K(i.totalram),
81 nid, K(i.freeram),
82 nid, K(i.totalram - i.freeram),
83 nid, K(node_page_state(nid, NR_ACTIVE_ANON) +
84 node_page_state(nid, NR_ACTIVE_FILE)),
85 nid, K(node_page_state(nid, NR_INACTIVE_ANON) +
86 node_page_state(nid, NR_INACTIVE_FILE)),
87 nid, K(node_page_state(nid, NR_ACTIVE_ANON)),
88 nid, K(node_page_state(nid, NR_INACTIVE_ANON)),
89 nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
90 nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
91 nid, K(node_page_state(nid, NR_UNEVICTABLE)),
92 nid, K(node_page_state(nid, NR_MLOCK)));
93
Christoph Lameter182e8e22006-09-25 23:31:10 -070094#ifdef CONFIG_HIGHMEM
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070095 n += sprintf(buf + n,
Rik van Riel4f98a2f2008-10-18 20:26:32 -070096 "Node %d HighTotal: %8lu kB\n"
97 "Node %d HighFree: %8lu kB\n"
98 "Node %d LowTotal: %8lu kB\n"
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070099 "Node %d LowFree: %8lu kB\n",
100 nid, K(i.totalhigh),
101 nid, K(i.freehigh),
102 nid, K(i.totalram - i.totalhigh),
103 nid, K(i.freeram - i.freehigh));
Christoph Lameter182e8e22006-09-25 23:31:10 -0700104#endif
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -0700105 n += sprintf(buf + n,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700106 "Node %d Dirty: %8lu kB\n"
107 "Node %d Writeback: %8lu kB\n"
108 "Node %d FilePages: %8lu kB\n"
109 "Node %d Mapped: %8lu kB\n"
110 "Node %d AnonPages: %8lu kB\n"
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700111 "Node %d Shmem: %8lu kB\n"
KOSAKI Motohiroc6a7f572009-09-21 17:01:32 -0700112 "Node %d KernelStack: %8lu kB\n"
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700113 "Node %d PageTables: %8lu kB\n"
114 "Node %d NFS_Unstable: %8lu kB\n"
115 "Node %d Bounce: %8lu kB\n"
116 "Node %d WritebackTmp: %8lu kB\n"
117 "Node %d Slab: %8lu kB\n"
118 "Node %d SReclaimable: %8lu kB\n"
David Rientjes05b258e2011-01-13 15:47:14 -0800119 "Node %d SUnreclaim: %8lu kB\n"
120#ifdef CONFIG_TRANSPARENT_HUGEPAGE
121 "Node %d AnonHugePages: %8lu kB\n"
122#endif
123 ,
Christoph Lameterb1e7a8f2006-06-30 01:55:39 -0700124 nid, K(node_page_state(nid, NR_FILE_DIRTY)),
Christoph Lameterce866b32006-06-30 01:55:40 -0700125 nid, K(node_page_state(nid, NR_WRITEBACK)),
Christoph Lameter347ce432006-06-30 01:55:35 -0700126 nid, K(node_page_state(nid, NR_FILE_PAGES)),
Christoph Lameter65ba55f2006-06-30 01:55:34 -0700127 nid, K(node_page_state(nid, NR_FILE_MAPPED)),
Claudio Scordino91a13c22011-11-17 11:08:32 +0100128 nid, K(node_page_state(nid, NR_ANON_PAGES)),
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700129 nid, K(node_page_state(nid, NR_SHMEM)),
KOSAKI Motohiroc6a7f572009-09-21 17:01:32 -0700130 nid, node_page_state(nid, NR_KERNEL_STACK) *
131 THREAD_SIZE / 1024,
Christoph Lameterdf849a12006-06-30 01:55:38 -0700132 nid, K(node_page_state(nid, NR_PAGETABLE)),
Christoph Lameterfd39fc82006-06-30 01:55:40 -0700133 nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
Christoph Lameterd2c5e302006-06-30 01:55:41 -0700134 nid, K(node_page_state(nid, NR_BOUNCE)),
Miklos Szeredifc3ba692008-04-30 00:54:38 -0700135 nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
Christoph Lameter972d1a72006-09-25 23:31:51 -0700136 nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
137 node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
138 nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
David Rientjes05b258e2011-01-13 15:47:14 -0800139#ifdef CONFIG_TRANSPARENT_HUGEPAGE
Claudio Scordino91a13c22011-11-17 11:08:32 +0100140 nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE))
David Rientjes05b258e2011-01-13 15:47:14 -0800141 , nid,
142 K(node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) *
Claudio Scordino91a13c22011-11-17 11:08:32 +0100143 HPAGE_PMD_NR));
144#else
145 nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
David Rientjes05b258e2011-01-13 15:47:14 -0800146#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147 n += hugetlb_report_node_meminfo(nid, buf + n);
148 return n;
149}
150
151#undef K
Kay Sievers10fbcf42011-12-21 14:48:43 -0800152static DEVICE_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153
Kay Sievers10fbcf42011-12-21 14:48:43 -0800154static ssize_t node_read_numastat(struct device *dev,
155 struct device_attribute *attr, char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157 return sprintf(buf,
158 "numa_hit %lu\n"
159 "numa_miss %lu\n"
160 "numa_foreign %lu\n"
161 "interleave_hit %lu\n"
162 "local_node %lu\n"
163 "other_node %lu\n",
Christoph Lameterca889e62006-06-30 01:55:44 -0700164 node_page_state(dev->id, NUMA_HIT),
165 node_page_state(dev->id, NUMA_MISS),
166 node_page_state(dev->id, NUMA_FOREIGN),
167 node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
168 node_page_state(dev->id, NUMA_LOCAL),
169 node_page_state(dev->id, NUMA_OTHER));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170}
Kay Sievers10fbcf42011-12-21 14:48:43 -0800171static DEVICE_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172
Kay Sievers10fbcf42011-12-21 14:48:43 -0800173static ssize_t node_read_vmstat(struct device *dev,
174 struct device_attribute *attr, char *buf)
Michael Rubin2ac39032010-10-26 14:21:35 -0700175{
176 int nid = dev->id;
KOSAKI Motohirofa25c502011-05-24 17:11:28 -0700177 int i;
178 int n = 0;
179
180 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
181 n += sprintf(buf+n, "%s %lu\n", vmstat_text[i],
182 node_page_state(nid, i));
183
184 return n;
Michael Rubin2ac39032010-10-26 14:21:35 -0700185}
Kay Sievers10fbcf42011-12-21 14:48:43 -0800186static DEVICE_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
Michael Rubin2ac39032010-10-26 14:21:35 -0700187
Kay Sievers10fbcf42011-12-21 14:48:43 -0800188static ssize_t node_read_distance(struct device *dev,
189 struct device_attribute *attr, char * buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190{
191 int nid = dev->id;
192 int len = 0;
193 int i;
194
David Rientjes12ee3c02010-03-10 14:50:21 -0800195 /*
196 * buf is currently PAGE_SIZE in length and each node needs 4 chars
197 * at the most (distance + space or newline).
198 */
199 BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200
201 for_each_online_node(i)
202 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
203
204 len += sprintf(buf + len, "\n");
205 return len;
206}
Kay Sievers10fbcf42011-12-21 14:48:43 -0800207static DEVICE_ATTR(distance, S_IRUGO, node_read_distance, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800209#ifdef CONFIG_HUGETLBFS
210/*
211 * hugetlbfs per node attributes registration interface:
212 * When/if hugetlb[fs] subsystem initializes [sometime after this module],
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800213 * it will register its per node attributes for all online nodes with
214 * memory. It will also call register_hugetlbfs_with_node(), below, to
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800215 * register its attribute registration functions with this node driver.
216 * Once these hooks have been initialized, the node driver will call into
217 * the hugetlb module to [un]register attributes for hot-plugged nodes.
218 */
219static node_registration_func_t __hugetlb_register_node;
220static node_registration_func_t __hugetlb_unregister_node;
221
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800222static inline bool hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800223{
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800224 if (__hugetlb_register_node &&
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -0800225 node_state(node->dev.id, N_MEMORY)) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800226 __hugetlb_register_node(node);
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800227 return true;
228 }
229 return false;
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800230}
231
232static inline void hugetlb_unregister_node(struct node *node)
233{
234 if (__hugetlb_unregister_node)
235 __hugetlb_unregister_node(node);
236}
237
238void register_hugetlbfs_with_node(node_registration_func_t doregister,
239 node_registration_func_t unregister)
240{
241 __hugetlb_register_node = doregister;
242 __hugetlb_unregister_node = unregister;
243}
244#else
245static inline void hugetlb_register_node(struct node *node) {}
246
247static inline void hugetlb_unregister_node(struct node *node) {}
248#endif
249
Yasuaki Ishimatsu8c7b5b42012-12-11 16:00:57 -0800250static void node_device_release(struct device *dev)
251{
252 struct node *node = to_node(dev);
253
254#if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HUGETLBFS)
255 /*
256 * We schedule the work only when a memory section is
257 * onlined/offlined on this node. When we come here,
258 * all the memory on this node has been offlined,
259 * so we won't enqueue new work to this work.
260 *
261 * The work is using node->node_work, so we should
262 * flush work before freeing the memory.
263 */
264 flush_work(&node->node_work);
265#endif
266 kfree(node);
267}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268
269/*
Robert P. J. Day405ae7d2007-02-17 19:13:42 +0100270 * register_node - Setup a sysfs device for a node.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271 * @num - Node number to use when creating the device.
272 *
273 * Initialize and register the node device.
274 */
Yasuaki Ishimatsufa264372012-12-11 16:02:52 -0800275static int register_node(struct node *node, int num, struct node *parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276{
277 int error;
278
Kay Sievers10fbcf42011-12-21 14:48:43 -0800279 node->dev.id = num;
280 node->dev.bus = &node_subsys;
Yasuaki Ishimatsu8c7b5b42012-12-11 16:00:57 -0800281 node->dev.release = node_device_release;
Kay Sievers10fbcf42011-12-21 14:48:43 -0800282 error = device_register(&node->dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283
284 if (!error){
Kay Sievers10fbcf42011-12-21 14:48:43 -0800285 device_create_file(&node->dev, &dev_attr_cpumap);
286 device_create_file(&node->dev, &dev_attr_cpulist);
287 device_create_file(&node->dev, &dev_attr_meminfo);
288 device_create_file(&node->dev, &dev_attr_numastat);
289 device_create_file(&node->dev, &dev_attr_distance);
290 device_create_file(&node->dev, &dev_attr_vmstat);
Lee Schermerhornaf936a12008-10-18 20:26:53 -0700291
292 scan_unevictable_register_node(node);
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800293
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800294 hugetlb_register_node(node);
Mel Gormaned4a6d72010-05-24 14:32:29 -0700295
296 compaction_register_node(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297 }
298 return error;
299}
300
Keiichiro Tokunaga4b450992005-05-08 21:28:53 +0900301/**
302 * unregister_node - unregister a node device
303 * @node: node going away
304 *
305 * Unregisters a node device @node. All the devices on the node must be
306 * unregistered before calling this function.
307 */
308void unregister_node(struct node *node)
309{
Kay Sievers10fbcf42011-12-21 14:48:43 -0800310 device_remove_file(&node->dev, &dev_attr_cpumap);
311 device_remove_file(&node->dev, &dev_attr_cpulist);
312 device_remove_file(&node->dev, &dev_attr_meminfo);
313 device_remove_file(&node->dev, &dev_attr_numastat);
314 device_remove_file(&node->dev, &dev_attr_distance);
315 device_remove_file(&node->dev, &dev_attr_vmstat);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316
Lee Schermerhornaf936a12008-10-18 20:26:53 -0700317 scan_unevictable_unregister_node(node);
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800318 hugetlb_unregister_node(node); /* no-op, if memoryless node */
Lee Schermerhornaf936a12008-10-18 20:26:53 -0700319
Kay Sievers10fbcf42011-12-21 14:48:43 -0800320 device_unregister(&node->dev);
Keiichiro Tokunaga4b450992005-05-08 21:28:53 +0900321}
322
Wen Congyang87327942012-12-11 16:00:56 -0800323struct node *node_devices[MAX_NUMNODES];
Yasunori Goto0fc44152006-06-27 02:53:38 -0700324
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700325/*
326 * register cpu under node
327 */
328int register_cpu_under_node(unsigned int cpu, unsigned int nid)
329{
Alex Chiang18307942009-12-14 17:59:08 -0800330 int ret;
Kay Sievers8a25a2f2011-12-21 14:29:42 -0800331 struct device *obj;
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700332
Alex Chiangf8246f32009-12-14 17:59:06 -0800333 if (!node_online(nid))
334 return 0;
335
Kay Sievers8a25a2f2011-12-21 14:29:42 -0800336 obj = get_cpu_device(cpu);
Alex Chiangf8246f32009-12-14 17:59:06 -0800337 if (!obj)
338 return 0;
339
Wen Congyang87327942012-12-11 16:00:56 -0800340 ret = sysfs_create_link(&node_devices[nid]->dev.kobj,
Alex Chiangf8246f32009-12-14 17:59:06 -0800341 &obj->kobj,
342 kobject_name(&obj->kobj));
Alex Chiang18307942009-12-14 17:59:08 -0800343 if (ret)
344 return ret;
345
346 return sysfs_create_link(&obj->kobj,
Wen Congyang87327942012-12-11 16:00:56 -0800347 &node_devices[nid]->dev.kobj,
348 kobject_name(&node_devices[nid]->dev.kobj));
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700349}
350
351int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
352{
Kay Sievers8a25a2f2011-12-21 14:29:42 -0800353 struct device *obj;
Alex Chiangb9d52da2009-12-14 17:59:07 -0800354
355 if (!node_online(nid))
356 return 0;
357
Kay Sievers8a25a2f2011-12-21 14:29:42 -0800358 obj = get_cpu_device(cpu);
Alex Chiangb9d52da2009-12-14 17:59:07 -0800359 if (!obj)
360 return 0;
361
Wen Congyang87327942012-12-11 16:00:56 -0800362 sysfs_remove_link(&node_devices[nid]->dev.kobj,
Alex Chiangb9d52da2009-12-14 17:59:07 -0800363 kobject_name(&obj->kobj));
Alex Chiang18307942009-12-14 17:59:08 -0800364 sysfs_remove_link(&obj->kobj,
Wen Congyang87327942012-12-11 16:00:56 -0800365 kobject_name(&node_devices[nid]->dev.kobj));
Alex Chiangb9d52da2009-12-14 17:59:07 -0800366
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700367 return 0;
368}
369
Gary Hadec04fc582009-01-06 14:39:14 -0800370#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
371#define page_initialized(page) (page->lru.next)
372
373static int get_nid_for_pfn(unsigned long pfn)
374{
375 struct page *page;
376
377 if (!pfn_valid_within(pfn))
378 return -1;
379 page = pfn_to_page(pfn);
380 if (!page_initialized(page))
381 return -1;
382 return pfn_to_nid(pfn);
383}
384
385/* register memory section under specified node if it spans that node */
386int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
387{
Alex Chiangdee5d0d2009-12-14 17:59:05 -0800388 int ret;
Gary Hadec04fc582009-01-06 14:39:14 -0800389 unsigned long pfn, sect_start_pfn, sect_end_pfn;
390
391 if (!mem_blk)
392 return -EFAULT;
393 if (!node_online(nid))
394 return 0;
Nathan Fontenotd3360162011-01-20 10:44:29 -0600395
396 sect_start_pfn = section_nr_to_pfn(mem_blk->start_section_nr);
397 sect_end_pfn = section_nr_to_pfn(mem_blk->end_section_nr);
398 sect_end_pfn += PAGES_PER_SECTION - 1;
Gary Hadec04fc582009-01-06 14:39:14 -0800399 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
400 int page_nid;
401
402 page_nid = get_nid_for_pfn(pfn);
403 if (page_nid < 0)
404 continue;
405 if (page_nid != nid)
406 continue;
Wen Congyang87327942012-12-11 16:00:56 -0800407 ret = sysfs_create_link_nowarn(&node_devices[nid]->dev.kobj,
Kay Sievers10fbcf42011-12-21 14:48:43 -0800408 &mem_blk->dev.kobj,
409 kobject_name(&mem_blk->dev.kobj));
Alex Chiangdee5d0d2009-12-14 17:59:05 -0800410 if (ret)
411 return ret;
412
Kay Sievers10fbcf42011-12-21 14:48:43 -0800413 return sysfs_create_link_nowarn(&mem_blk->dev.kobj,
Wen Congyang87327942012-12-11 16:00:56 -0800414 &node_devices[nid]->dev.kobj,
415 kobject_name(&node_devices[nid]->dev.kobj));
Gary Hadec04fc582009-01-06 14:39:14 -0800416 }
417 /* mem section does not span the specified node */
418 return 0;
419}
420
421/* unregister memory section under all nodes that it spans */
Nathan Fontenotd3360162011-01-20 10:44:29 -0600422int unregister_mem_sect_under_nodes(struct memory_block *mem_blk,
423 unsigned long phys_index)
Gary Hadec04fc582009-01-06 14:39:14 -0800424{
David Rientjes9ae49fa2009-12-14 17:59:46 -0800425 NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
Gary Hadec04fc582009-01-06 14:39:14 -0800426 unsigned long pfn, sect_start_pfn, sect_end_pfn;
427
David Rientjes9ae49fa2009-12-14 17:59:46 -0800428 if (!mem_blk) {
429 NODEMASK_FREE(unlinked_nodes);
Gary Hadec04fc582009-01-06 14:39:14 -0800430 return -EFAULT;
David Rientjes9ae49fa2009-12-14 17:59:46 -0800431 }
432 if (!unlinked_nodes)
433 return -ENOMEM;
434 nodes_clear(*unlinked_nodes);
Nathan Fontenotd3360162011-01-20 10:44:29 -0600435
436 sect_start_pfn = section_nr_to_pfn(phys_index);
Gary Hadec04fc582009-01-06 14:39:14 -0800437 sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
438 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
Roel Kluin47504982009-03-10 12:55:45 -0700439 int nid;
Gary Hadec04fc582009-01-06 14:39:14 -0800440
441 nid = get_nid_for_pfn(pfn);
442 if (nid < 0)
443 continue;
444 if (!node_online(nid))
445 continue;
David Rientjes9ae49fa2009-12-14 17:59:46 -0800446 if (node_test_and_set(nid, *unlinked_nodes))
Gary Hadec04fc582009-01-06 14:39:14 -0800447 continue;
Wen Congyang87327942012-12-11 16:00:56 -0800448 sysfs_remove_link(&node_devices[nid]->dev.kobj,
Kay Sievers10fbcf42011-12-21 14:48:43 -0800449 kobject_name(&mem_blk->dev.kobj));
450 sysfs_remove_link(&mem_blk->dev.kobj,
Wen Congyang87327942012-12-11 16:00:56 -0800451 kobject_name(&node_devices[nid]->dev.kobj));
Gary Hadec04fc582009-01-06 14:39:14 -0800452 }
David Rientjes9ae49fa2009-12-14 17:59:46 -0800453 NODEMASK_FREE(unlinked_nodes);
Gary Hadec04fc582009-01-06 14:39:14 -0800454 return 0;
455}
456
457static int link_mem_sections(int nid)
458{
459 unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
460 unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
461 unsigned long pfn;
Robin Holt63d027a2010-09-29 14:00:56 -0500462 struct memory_block *mem_blk = NULL;
Gary Hadec04fc582009-01-06 14:39:14 -0800463 int err = 0;
464
465 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
466 unsigned long section_nr = pfn_to_section_nr(pfn);
467 struct mem_section *mem_sect;
Gary Hadec04fc582009-01-06 14:39:14 -0800468 int ret;
469
470 if (!present_section_nr(section_nr))
471 continue;
472 mem_sect = __nr_to_section(section_nr);
Yinghai Lu321bf4e2012-01-30 13:57:12 -0800473
474 /* same memblock ? */
475 if (mem_blk)
476 if ((section_nr >= mem_blk->start_section_nr) &&
477 (section_nr <= mem_blk->end_section_nr))
478 continue;
479
Robin Holt63d027a2010-09-29 14:00:56 -0500480 mem_blk = find_memory_block_hinted(mem_sect, mem_blk);
Yinghai Lu321bf4e2012-01-30 13:57:12 -0800481
Gary Hadec04fc582009-01-06 14:39:14 -0800482 ret = register_mem_sect_under_node(mem_blk, nid);
483 if (!err)
484 err = ret;
485
486 /* discard ref obtained in find_memory_block() */
Gary Hadec04fc582009-01-06 14:39:14 -0800487 }
Robin Holt63d027a2010-09-29 14:00:56 -0500488
489 if (mem_blk)
Kay Sievers10fbcf42011-12-21 14:48:43 -0800490 kobject_put(&mem_blk->dev.kobj);
Gary Hadec04fc582009-01-06 14:39:14 -0800491 return err;
492}
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800493
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800494#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800495/*
496 * Handle per node hstate attribute [un]registration on transistions
497 * to/from memoryless state.
498 */
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800499static void node_hugetlb_work(struct work_struct *work)
500{
501 struct node *node = container_of(work, struct node, node_work);
502
503 /*
504 * We only get here when a node transitions to/from memoryless state.
505 * We can detect which transition occurred by examining whether the
506 * node has memory now. hugetlb_register_node() already check this
507 * so we try to register the attributes. If that fails, then the
508 * node has transitioned to memoryless, try to unregister the
509 * attributes.
510 */
511 if (!hugetlb_register_node(node))
512 hugetlb_unregister_node(node);
513}
514
515static void init_node_hugetlb_work(int nid)
516{
Wen Congyang87327942012-12-11 16:00:56 -0800517 INIT_WORK(&node_devices[nid]->node_work, node_hugetlb_work);
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800518}
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800519
520static int node_memory_callback(struct notifier_block *self,
521 unsigned long action, void *arg)
522{
523 struct memory_notify *mnb = arg;
524 int nid = mnb->status_change_nid;
525
526 switch (action) {
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800527 case MEM_ONLINE:
528 case MEM_OFFLINE:
529 /*
530 * offload per node hstate [un]registration to a work thread
531 * when transitioning to/from memoryless state.
532 */
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800533 if (nid != NUMA_NO_NODE)
Wen Congyang87327942012-12-11 16:00:56 -0800534 schedule_work(&node_devices[nid]->node_work);
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800535 break;
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800536
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800537 case MEM_GOING_ONLINE:
538 case MEM_GOING_OFFLINE:
539 case MEM_CANCEL_ONLINE:
540 case MEM_CANCEL_OFFLINE:
541 default:
542 break;
543 }
544
545 return NOTIFY_OK;
546}
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800547#endif /* CONFIG_HUGETLBFS */
548#else /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800549
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800550static int link_mem_sections(int nid) { return 0; }
551#endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
552
553#if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
554 !defined(CONFIG_HUGETLBFS)
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800555static inline int node_memory_callback(struct notifier_block *self,
556 unsigned long action, void *arg)
557{
558 return NOTIFY_OK;
559}
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800560
561static void init_node_hugetlb_work(int nid) { }
562
563#endif
Gary Hadec04fc582009-01-06 14:39:14 -0800564
Yasunori Goto0fc44152006-06-27 02:53:38 -0700565int register_one_node(int nid)
566{
567 int error = 0;
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700568 int cpu;
Yasunori Goto0fc44152006-06-27 02:53:38 -0700569
570 if (node_online(nid)) {
571 int p_node = parent_node(nid);
572 struct node *parent = NULL;
573
574 if (p_node != nid)
Wen Congyang87327942012-12-11 16:00:56 -0800575 parent = node_devices[p_node];
Yasunori Goto0fc44152006-06-27 02:53:38 -0700576
Wen Congyang87327942012-12-11 16:00:56 -0800577 node_devices[nid] = kzalloc(sizeof(struct node), GFP_KERNEL);
578 if (!node_devices[nid])
579 return -ENOMEM;
580
581 error = register_node(node_devices[nid], nid, parent);
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700582
583 /* link cpu under this node */
584 for_each_present_cpu(cpu) {
585 if (cpu_to_node(cpu) == nid)
586 register_cpu_under_node(cpu, nid);
587 }
Gary Hadec04fc582009-01-06 14:39:14 -0800588
589 /* link memory sections under this node */
590 error = link_mem_sections(nid);
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800591
592 /* initialize work queue for memory hot plug */
593 init_node_hugetlb_work(nid);
Yasunori Goto0fc44152006-06-27 02:53:38 -0700594 }
595
596 return error;
597
598}
599
600void unregister_one_node(int nid)
601{
Wen Congyang87327942012-12-11 16:00:56 -0800602 unregister_node(node_devices[nid]);
Wen Congyang87327942012-12-11 16:00:56 -0800603 node_devices[nid] = NULL;
Yasunori Goto0fc44152006-06-27 02:53:38 -0700604}
605
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700606/*
607 * node states attributes
608 */
609
610static ssize_t print_nodes_state(enum node_states state, char *buf)
611{
612 int n;
613
Ryota Ozakif6238812012-05-29 15:06:20 -0700614 n = nodelist_scnprintf(buf, PAGE_SIZE-2, node_states[state]);
615 buf[n++] = '\n';
616 buf[n] = '\0';
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700617 return n;
618}
619
Andi Kleenb15f5622010-01-05 12:47:59 +0100620struct node_attr {
Kay Sievers10fbcf42011-12-21 14:48:43 -0800621 struct device_attribute attr;
Andi Kleenb15f5622010-01-05 12:47:59 +0100622 enum node_states state;
623};
624
Kay Sievers10fbcf42011-12-21 14:48:43 -0800625static ssize_t show_node_state(struct device *dev,
626 struct device_attribute *attr, char *buf)
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700627{
Andi Kleenb15f5622010-01-05 12:47:59 +0100628 struct node_attr *na = container_of(attr, struct node_attr, attr);
629 return print_nodes_state(na->state, buf);
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700630}
631
Andi Kleenb15f5622010-01-05 12:47:59 +0100632#define _NODE_ATTR(name, state) \
Kay Sievers10fbcf42011-12-21 14:48:43 -0800633 { __ATTR(name, 0444, show_node_state, NULL), state }
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700634
Andi Kleenb15f5622010-01-05 12:47:59 +0100635static struct node_attr node_state_attr[] = {
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800636 [N_POSSIBLE] = _NODE_ATTR(possible, N_POSSIBLE),
637 [N_ONLINE] = _NODE_ATTR(online, N_ONLINE),
638 [N_NORMAL_MEMORY] = _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700639#ifdef CONFIG_HIGHMEM
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800640 [N_HIGH_MEMORY] = _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700641#endif
Lai Jiangshan20b2f522012-12-12 13:52:00 -0800642#ifdef CONFIG_MOVABLE_NODE
643 [N_MEMORY] = _NODE_ATTR(has_memory, N_MEMORY),
644#endif
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800645 [N_CPU] = _NODE_ATTR(has_cpu, N_CPU),
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700646};
647
Kay Sievers10fbcf42011-12-21 14:48:43 -0800648static struct attribute *node_state_attrs[] = {
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800649 &node_state_attr[N_POSSIBLE].attr.attr,
650 &node_state_attr[N_ONLINE].attr.attr,
651 &node_state_attr[N_NORMAL_MEMORY].attr.attr,
Andi Kleen3701cde2010-01-05 12:48:04 +0100652#ifdef CONFIG_HIGHMEM
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800653 &node_state_attr[N_HIGH_MEMORY].attr.attr,
Andi Kleen3701cde2010-01-05 12:48:04 +0100654#endif
Lai Jiangshan20b2f522012-12-12 13:52:00 -0800655#ifdef CONFIG_MOVABLE_NODE
656 &node_state_attr[N_MEMORY].attr.attr,
657#endif
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800658 &node_state_attr[N_CPU].attr.attr,
Andi Kleen3701cde2010-01-05 12:48:04 +0100659 NULL
660};
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700661
Kay Sievers10fbcf42011-12-21 14:48:43 -0800662static struct attribute_group memory_root_attr_group = {
663 .attrs = node_state_attrs,
664};
665
666static const struct attribute_group *cpu_root_attr_groups[] = {
667 &memory_root_attr_group,
668 NULL,
669};
670
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800671#define NODE_CALLBACK_PRI 2 /* lower than SLAB */
Keiichiro Tokunaga4b450992005-05-08 21:28:53 +0900672static int __init register_node_type(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673{
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700674 int ret;
675
Andi Kleen3701cde2010-01-05 12:48:04 +0100676 BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
677 BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
678
Kay Sievers10fbcf42011-12-21 14:48:43 -0800679 ret = subsys_system_register(&node_subsys, cpu_root_attr_groups);
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800680 if (!ret) {
Andrew Morton6e259e72013-04-29 15:08:07 -0700681 static struct notifier_block node_memory_callback_nb = {
682 .notifier_call = node_memory_callback,
683 .priority = NODE_CALLBACK_PRI,
684 };
685 register_hotmemory_notifier(&node_memory_callback_nb);
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800686 }
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700687
688 /*
689 * Note: we're not going to unregister the node class if we fail
690 * to register the node state class attribute files.
691 */
692 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693}
694postcore_initcall(register_node_type);