blob: 5548f96860162904e7af380eb89239211d9d2cc9 [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;
Mel Gorman599d0c92016-07-28 15:45:31 -070059 struct pglist_data *pgdat = NODE_DATA(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -070060 struct sysinfo i;
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
62 si_meminfo_node(&i, nid);
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070063 n = sprintf(buf,
Rik van Riel4f98a2f2008-10-18 20:26:32 -070064 "Node %d MemTotal: %8lu kB\n"
65 "Node %d MemFree: %8lu kB\n"
66 "Node %d MemUsed: %8lu kB\n"
67 "Node %d Active: %8lu kB\n"
68 "Node %d Inactive: %8lu kB\n"
69 "Node %d Active(anon): %8lu kB\n"
70 "Node %d Inactive(anon): %8lu kB\n"
71 "Node %d Active(file): %8lu kB\n"
72 "Node %d Inactive(file): %8lu kB\n"
Nick Piggin5344b7e2008-10-18 20:26:51 -070073 "Node %d Unevictable: %8lu kB\n"
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070074 "Node %d Mlocked: %8lu kB\n",
75 nid, K(i.totalram),
76 nid, K(i.freeram),
77 nid, K(i.totalram - i.freeram),
Mel Gorman599d0c92016-07-28 15:45:31 -070078 nid, K(node_page_state(pgdat, NR_ACTIVE_ANON) +
79 node_page_state(pgdat, NR_ACTIVE_FILE)),
80 nid, K(node_page_state(pgdat, NR_INACTIVE_ANON) +
81 node_page_state(pgdat, NR_INACTIVE_FILE)),
82 nid, K(node_page_state(pgdat, NR_ACTIVE_ANON)),
83 nid, K(node_page_state(pgdat, NR_INACTIVE_ANON)),
84 nid, K(node_page_state(pgdat, NR_ACTIVE_FILE)),
85 nid, K(node_page_state(pgdat, NR_INACTIVE_FILE)),
86 nid, K(node_page_state(pgdat, NR_UNEVICTABLE)),
Mel Gorman75ef7182016-07-28 15:45:24 -070087 nid, K(sum_zone_node_page_state(nid, NR_MLOCK)));
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070088
Christoph Lameter182e8e22006-09-25 23:31:10 -070089#ifdef CONFIG_HIGHMEM
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070090 n += sprintf(buf + n,
Rik van Riel4f98a2f2008-10-18 20:26:32 -070091 "Node %d HighTotal: %8lu kB\n"
92 "Node %d HighFree: %8lu kB\n"
93 "Node %d LowTotal: %8lu kB\n"
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -070094 "Node %d LowFree: %8lu kB\n",
95 nid, K(i.totalhigh),
96 nid, K(i.freehigh),
97 nid, K(i.totalram - i.totalhigh),
98 nid, K(i.freeram - i.freehigh));
Christoph Lameter182e8e22006-09-25 23:31:10 -070099#endif
KOSAKI Motohiro7ee92252010-08-09 17:19:50 -0700100 n += sprintf(buf + n,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700101 "Node %d Dirty: %8lu kB\n"
102 "Node %d Writeback: %8lu kB\n"
103 "Node %d FilePages: %8lu kB\n"
104 "Node %d Mapped: %8lu kB\n"
105 "Node %d AnonPages: %8lu kB\n"
KOSAKI Motohiro4b021082009-09-21 17:01:33 -0700106 "Node %d Shmem: %8lu kB\n"
KOSAKI Motohiroc6a7f572009-09-21 17:01:32 -0700107 "Node %d KernelStack: %8lu kB\n"
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700108 "Node %d PageTables: %8lu kB\n"
109 "Node %d NFS_Unstable: %8lu kB\n"
110 "Node %d Bounce: %8lu kB\n"
111 "Node %d WritebackTmp: %8lu kB\n"
112 "Node %d Slab: %8lu kB\n"
113 "Node %d SReclaimable: %8lu kB\n"
David Rientjes05b258e2011-01-13 15:47:14 -0800114 "Node %d SUnreclaim: %8lu kB\n"
115#ifdef CONFIG_TRANSPARENT_HUGEPAGE
116 "Node %d AnonHugePages: %8lu kB\n"
Kirill A. Shutemov65c45372016-07-26 15:26:10 -0700117 "Node %d ShmemHugePages: %8lu kB\n"
118 "Node %d ShmemPmdMapped: %8lu kB\n"
David Rientjes05b258e2011-01-13 15:47:14 -0800119#endif
120 ,
Mel Gorman11fb9982016-07-28 15:46:20 -0700121 nid, K(node_page_state(pgdat, NR_FILE_DIRTY)),
122 nid, K(node_page_state(pgdat, NR_WRITEBACK)),
123 nid, K(node_page_state(pgdat, NR_FILE_PAGES)),
Mel Gorman50658e22016-07-28 15:46:14 -0700124 nid, K(node_page_state(pgdat, NR_FILE_MAPPED)),
Mel Gorman4b9d0fa2016-07-28 15:46:17 -0700125 nid, K(node_page_state(pgdat, NR_ANON_MAPPED)),
Rafael Aquinicc7452b2014-08-06 16:06:38 -0700126 nid, K(i.sharedram),
Andy Lutomirskid30dd8b2016-07-28 15:48:14 -0700127 nid, sum_zone_node_page_state(nid, NR_KERNEL_STACK_KB),
Mel Gorman75ef7182016-07-28 15:45:24 -0700128 nid, K(sum_zone_node_page_state(nid, NR_PAGETABLE)),
Mel Gorman11fb9982016-07-28 15:46:20 -0700129 nid, K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
Mel Gorman75ef7182016-07-28 15:45:24 -0700130 nid, K(sum_zone_node_page_state(nid, NR_BOUNCE)),
Mel Gorman11fb9982016-07-28 15:46:20 -0700131 nid, K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
Mel Gorman75ef7182016-07-28 15:45:24 -0700132 nid, K(sum_zone_node_page_state(nid, NR_SLAB_RECLAIMABLE) +
133 sum_zone_node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
134 nid, K(sum_zone_node_page_state(nid, NR_SLAB_RECLAIMABLE)),
David Rientjes05b258e2011-01-13 15:47:14 -0800135#ifdef CONFIG_TRANSPARENT_HUGEPAGE
Mel Gorman75ef7182016-07-28 15:45:24 -0700136 nid, K(sum_zone_node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
Mel Gorman11fb9982016-07-28 15:46:20 -0700137 nid, K(node_page_state(pgdat, NR_ANON_THPS) *
Kirill A. Shutemov65c45372016-07-26 15:26:10 -0700138 HPAGE_PMD_NR),
Mel Gorman11fb9982016-07-28 15:46:20 -0700139 nid, K(node_page_state(pgdat, NR_SHMEM_THPS) *
Kirill A. Shutemov65c45372016-07-26 15:26:10 -0700140 HPAGE_PMD_NR),
Mel Gorman11fb9982016-07-28 15:46:20 -0700141 nid, K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED) *
Kirill A. Shutemov65c45372016-07-26 15:26:10 -0700142 HPAGE_PMD_NR));
Claudio Scordino91a13c22011-11-17 11:08:32 +0100143#else
Mel Gorman75ef7182016-07-28 15:45:24 -0700144 nid, K(sum_zone_node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
David Rientjes05b258e2011-01-13 15:47:14 -0800145#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 n += hugetlb_report_node_meminfo(nid, buf + n);
147 return n;
148}
149
150#undef K
Kay Sievers10fbcf42011-12-21 14:48:43 -0800151static DEVICE_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152
Kay Sievers10fbcf42011-12-21 14:48:43 -0800153static ssize_t node_read_numastat(struct device *dev,
154 struct device_attribute *attr, char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156 return sprintf(buf,
157 "numa_hit %lu\n"
158 "numa_miss %lu\n"
159 "numa_foreign %lu\n"
160 "interleave_hit %lu\n"
161 "local_node %lu\n"
162 "other_node %lu\n",
Mel Gorman75ef7182016-07-28 15:45:24 -0700163 sum_zone_node_page_state(dev->id, NUMA_HIT),
164 sum_zone_node_page_state(dev->id, NUMA_MISS),
165 sum_zone_node_page_state(dev->id, NUMA_FOREIGN),
166 sum_zone_node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
167 sum_zone_node_page_state(dev->id, NUMA_LOCAL),
168 sum_zone_node_page_state(dev->id, NUMA_OTHER));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169}
Kay Sievers10fbcf42011-12-21 14:48:43 -0800170static DEVICE_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171
Kay Sievers10fbcf42011-12-21 14:48:43 -0800172static ssize_t node_read_vmstat(struct device *dev,
173 struct device_attribute *attr, char *buf)
Michael Rubin2ac39032010-10-26 14:21:35 -0700174{
175 int nid = dev->id;
Mel Gorman75ef7182016-07-28 15:45:24 -0700176 struct pglist_data *pgdat = NODE_DATA(nid);
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],
Mel Gorman75ef7182016-07-28 15:45:24 -0700182 sum_zone_node_page_state(nid, i));
183
184 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
185 n += sprintf(buf+n, "%s %lu\n",
186 vmstat_text[i + NR_VM_ZONE_STAT_ITEMS],
187 node_page_state(pgdat, i));
KOSAKI Motohirofa25c502011-05-24 17:11:28 -0700188
189 return n;
Michael Rubin2ac39032010-10-26 14:21:35 -0700190}
Kay Sievers10fbcf42011-12-21 14:48:43 -0800191static DEVICE_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
Michael Rubin2ac39032010-10-26 14:21:35 -0700192
Kay Sievers10fbcf42011-12-21 14:48:43 -0800193static ssize_t node_read_distance(struct device *dev,
Ana Nedelcu518d3f32015-03-08 12:48:48 +0200194 struct device_attribute *attr, char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195{
196 int nid = dev->id;
197 int len = 0;
198 int i;
199
David Rientjes12ee3c02010-03-10 14:50:21 -0800200 /*
201 * buf is currently PAGE_SIZE in length and each node needs 4 chars
202 * at the most (distance + space or newline).
203 */
204 BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205
206 for_each_online_node(i)
207 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
208
209 len += sprintf(buf + len, "\n");
210 return len;
211}
Kay Sievers10fbcf42011-12-21 14:48:43 -0800212static DEVICE_ATTR(distance, S_IRUGO, node_read_distance, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213
Takashi Iwai3c9b8aa2015-01-29 12:29:22 +0100214static struct attribute *node_dev_attrs[] = {
215 &dev_attr_cpumap.attr,
216 &dev_attr_cpulist.attr,
217 &dev_attr_meminfo.attr,
218 &dev_attr_numastat.attr,
219 &dev_attr_distance.attr,
220 &dev_attr_vmstat.attr,
221 NULL
222};
Greg Kroah-Hartman7ca7ec42015-03-25 13:47:17 +0100223ATTRIBUTE_GROUPS(node_dev);
Takashi Iwai3c9b8aa2015-01-29 12:29:22 +0100224
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800225#ifdef CONFIG_HUGETLBFS
226/*
227 * hugetlbfs per node attributes registration interface:
228 * When/if hugetlb[fs] subsystem initializes [sometime after this module],
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800229 * it will register its per node attributes for all online nodes with
230 * memory. It will also call register_hugetlbfs_with_node(), below, to
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800231 * register its attribute registration functions with this node driver.
232 * Once these hooks have been initialized, the node driver will call into
233 * the hugetlb module to [un]register attributes for hot-plugged nodes.
234 */
235static node_registration_func_t __hugetlb_register_node;
236static node_registration_func_t __hugetlb_unregister_node;
237
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800238static inline bool hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800239{
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800240 if (__hugetlb_register_node &&
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -0800241 node_state(node->dev.id, N_MEMORY)) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800242 __hugetlb_register_node(node);
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800243 return true;
244 }
245 return false;
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800246}
247
248static inline void hugetlb_unregister_node(struct node *node)
249{
250 if (__hugetlb_unregister_node)
251 __hugetlb_unregister_node(node);
252}
253
254void register_hugetlbfs_with_node(node_registration_func_t doregister,
255 node_registration_func_t unregister)
256{
257 __hugetlb_register_node = doregister;
258 __hugetlb_unregister_node = unregister;
259}
260#else
261static inline void hugetlb_register_node(struct node *node) {}
262
263static inline void hugetlb_unregister_node(struct node *node) {}
264#endif
265
Yasuaki Ishimatsu8c7b5b42012-12-11 16:00:57 -0800266static void node_device_release(struct device *dev)
267{
268 struct node *node = to_node(dev);
269
270#if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HUGETLBFS)
271 /*
272 * We schedule the work only when a memory section is
273 * onlined/offlined on this node. When we come here,
274 * all the memory on this node has been offlined,
275 * so we won't enqueue new work to this work.
276 *
277 * The work is using node->node_work, so we should
278 * flush work before freeing the memory.
279 */
280 flush_work(&node->node_work);
281#endif
282 kfree(node);
283}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284
285/*
Robert P. J. Day405ae7d2007-02-17 19:13:42 +0100286 * register_node - Setup a sysfs device for a node.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 * @num - Node number to use when creating the device.
288 *
289 * Initialize and register the node device.
290 */
Yasuaki Ishimatsufa264372012-12-11 16:02:52 -0800291static int register_node(struct node *node, int num, struct node *parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292{
293 int error;
294
Kay Sievers10fbcf42011-12-21 14:48:43 -0800295 node->dev.id = num;
296 node->dev.bus = &node_subsys;
Yasuaki Ishimatsu8c7b5b42012-12-11 16:00:57 -0800297 node->dev.release = node_device_release;
Greg Kroah-Hartman7ca7ec42015-03-25 13:47:17 +0100298 node->dev.groups = node_dev_groups;
Kay Sievers10fbcf42011-12-21 14:48:43 -0800299 error = device_register(&node->dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300
301 if (!error){
Lee Schermerhorn9a305232009-12-14 17:58:25 -0800302 hugetlb_register_node(node);
Mel Gormaned4a6d72010-05-24 14:32:29 -0700303
304 compaction_register_node(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 }
306 return error;
307}
308
Keiichiro Tokunaga4b450992005-05-08 21:28:53 +0900309/**
310 * unregister_node - unregister a node device
311 * @node: node going away
312 *
313 * Unregisters a node device @node. All the devices on the node must be
314 * unregistered before calling this function.
315 */
316void unregister_node(struct node *node)
317{
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
Fabian Frederickbd721ea2016-08-02 14:03:33 -0700373static int __ref get_nid_for_pfn(unsigned long pfn)
Gary Hadec04fc582009-01-06 14:39:14 -0800374{
375 struct page *page;
376
377 if (!pfn_valid_within(pfn))
378 return -1;
Mel Gorman3a80a7f2015-06-30 14:57:02 -0700379#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
380 if (system_state == SYSTEM_BOOTING)
381 return early_pfn_to_nid(pfn);
382#endif
Gary Hadec04fc582009-01-06 14:39:14 -0800383 page = pfn_to_page(pfn);
384 if (!page_initialized(page))
385 return -1;
386 return pfn_to_nid(pfn);
387}
388
389/* register memory section under specified node if it spans that node */
390int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
391{
Alex Chiangdee5d0d2009-12-14 17:59:05 -0800392 int ret;
Gary Hadec04fc582009-01-06 14:39:14 -0800393 unsigned long pfn, sect_start_pfn, sect_end_pfn;
394
395 if (!mem_blk)
396 return -EFAULT;
397 if (!node_online(nid))
398 return 0;
Nathan Fontenotd3360162011-01-20 10:44:29 -0600399
400 sect_start_pfn = section_nr_to_pfn(mem_blk->start_section_nr);
401 sect_end_pfn = section_nr_to_pfn(mem_blk->end_section_nr);
402 sect_end_pfn += PAGES_PER_SECTION - 1;
Gary Hadec04fc582009-01-06 14:39:14 -0800403 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
404 int page_nid;
405
Yinghai Lu04697852015-09-04 15:42:39 -0700406 /*
407 * memory block could have several absent sections from start.
408 * skip pfn range from absent section
409 */
410 if (!pfn_present(pfn)) {
411 pfn = round_down(pfn + PAGES_PER_SECTION,
412 PAGES_PER_SECTION) - 1;
413 continue;
414 }
415
Gary Hadec04fc582009-01-06 14:39:14 -0800416 page_nid = get_nid_for_pfn(pfn);
417 if (page_nid < 0)
418 continue;
419 if (page_nid != nid)
420 continue;
Wen Congyang87327942012-12-11 16:00:56 -0800421 ret = sysfs_create_link_nowarn(&node_devices[nid]->dev.kobj,
Kay Sievers10fbcf42011-12-21 14:48:43 -0800422 &mem_blk->dev.kobj,
423 kobject_name(&mem_blk->dev.kobj));
Alex Chiangdee5d0d2009-12-14 17:59:05 -0800424 if (ret)
425 return ret;
426
Kay Sievers10fbcf42011-12-21 14:48:43 -0800427 return sysfs_create_link_nowarn(&mem_blk->dev.kobj,
Wen Congyang87327942012-12-11 16:00:56 -0800428 &node_devices[nid]->dev.kobj,
429 kobject_name(&node_devices[nid]->dev.kobj));
Gary Hadec04fc582009-01-06 14:39:14 -0800430 }
431 /* mem section does not span the specified node */
432 return 0;
433}
434
435/* unregister memory section under all nodes that it spans */
Nathan Fontenotd3360162011-01-20 10:44:29 -0600436int unregister_mem_sect_under_nodes(struct memory_block *mem_blk,
437 unsigned long phys_index)
Gary Hadec04fc582009-01-06 14:39:14 -0800438{
David Rientjes9ae49fa2009-12-14 17:59:46 -0800439 NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
Gary Hadec04fc582009-01-06 14:39:14 -0800440 unsigned long pfn, sect_start_pfn, sect_end_pfn;
441
David Rientjes9ae49fa2009-12-14 17:59:46 -0800442 if (!mem_blk) {
443 NODEMASK_FREE(unlinked_nodes);
Gary Hadec04fc582009-01-06 14:39:14 -0800444 return -EFAULT;
David Rientjes9ae49fa2009-12-14 17:59:46 -0800445 }
446 if (!unlinked_nodes)
447 return -ENOMEM;
448 nodes_clear(*unlinked_nodes);
Nathan Fontenotd3360162011-01-20 10:44:29 -0600449
450 sect_start_pfn = section_nr_to_pfn(phys_index);
Gary Hadec04fc582009-01-06 14:39:14 -0800451 sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
452 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
Roel Kluin47504982009-03-10 12:55:45 -0700453 int nid;
Gary Hadec04fc582009-01-06 14:39:14 -0800454
455 nid = get_nid_for_pfn(pfn);
456 if (nid < 0)
457 continue;
458 if (!node_online(nid))
459 continue;
David Rientjes9ae49fa2009-12-14 17:59:46 -0800460 if (node_test_and_set(nid, *unlinked_nodes))
Gary Hadec04fc582009-01-06 14:39:14 -0800461 continue;
Wen Congyang87327942012-12-11 16:00:56 -0800462 sysfs_remove_link(&node_devices[nid]->dev.kobj,
Kay Sievers10fbcf42011-12-21 14:48:43 -0800463 kobject_name(&mem_blk->dev.kobj));
464 sysfs_remove_link(&mem_blk->dev.kobj,
Wen Congyang87327942012-12-11 16:00:56 -0800465 kobject_name(&node_devices[nid]->dev.kobj));
Gary Hadec04fc582009-01-06 14:39:14 -0800466 }
David Rientjes9ae49fa2009-12-14 17:59:46 -0800467 NODEMASK_FREE(unlinked_nodes);
Gary Hadec04fc582009-01-06 14:39:14 -0800468 return 0;
469}
470
471static int link_mem_sections(int nid)
472{
473 unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
474 unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
475 unsigned long pfn;
Robin Holt63d027a2010-09-29 14:00:56 -0500476 struct memory_block *mem_blk = NULL;
Gary Hadec04fc582009-01-06 14:39:14 -0800477 int err = 0;
478
479 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
480 unsigned long section_nr = pfn_to_section_nr(pfn);
481 struct mem_section *mem_sect;
Gary Hadec04fc582009-01-06 14:39:14 -0800482 int ret;
483
484 if (!present_section_nr(section_nr))
485 continue;
486 mem_sect = __nr_to_section(section_nr);
Yinghai Lu321bf4e2012-01-30 13:57:12 -0800487
488 /* same memblock ? */
489 if (mem_blk)
490 if ((section_nr >= mem_blk->start_section_nr) &&
491 (section_nr <= mem_blk->end_section_nr))
492 continue;
493
Robin Holt63d027a2010-09-29 14:00:56 -0500494 mem_blk = find_memory_block_hinted(mem_sect, mem_blk);
Yinghai Lu321bf4e2012-01-30 13:57:12 -0800495
Gary Hadec04fc582009-01-06 14:39:14 -0800496 ret = register_mem_sect_under_node(mem_blk, nid);
497 if (!err)
498 err = ret;
499
500 /* discard ref obtained in find_memory_block() */
Gary Hadec04fc582009-01-06 14:39:14 -0800501 }
Robin Holt63d027a2010-09-29 14:00:56 -0500502
503 if (mem_blk)
Kay Sievers10fbcf42011-12-21 14:48:43 -0800504 kobject_put(&mem_blk->dev.kobj);
Gary Hadec04fc582009-01-06 14:39:14 -0800505 return err;
506}
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800507
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800508#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800509/*
510 * Handle per node hstate attribute [un]registration on transistions
511 * to/from memoryless state.
512 */
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800513static void node_hugetlb_work(struct work_struct *work)
514{
515 struct node *node = container_of(work, struct node, node_work);
516
517 /*
518 * We only get here when a node transitions to/from memoryless state.
519 * We can detect which transition occurred by examining whether the
520 * node has memory now. hugetlb_register_node() already check this
521 * so we try to register the attributes. If that fails, then the
522 * node has transitioned to memoryless, try to unregister the
523 * attributes.
524 */
525 if (!hugetlb_register_node(node))
526 hugetlb_unregister_node(node);
527}
528
529static void init_node_hugetlb_work(int nid)
530{
Wen Congyang87327942012-12-11 16:00:56 -0800531 INIT_WORK(&node_devices[nid]->node_work, node_hugetlb_work);
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800532}
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800533
534static int node_memory_callback(struct notifier_block *self,
535 unsigned long action, void *arg)
536{
537 struct memory_notify *mnb = arg;
538 int nid = mnb->status_change_nid;
539
540 switch (action) {
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800541 case MEM_ONLINE:
542 case MEM_OFFLINE:
543 /*
544 * offload per node hstate [un]registration to a work thread
545 * when transitioning to/from memoryless state.
546 */
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800547 if (nid != NUMA_NO_NODE)
Wen Congyang87327942012-12-11 16:00:56 -0800548 schedule_work(&node_devices[nid]->node_work);
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800549 break;
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800550
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800551 case MEM_GOING_ONLINE:
552 case MEM_GOING_OFFLINE:
553 case MEM_CANCEL_ONLINE:
554 case MEM_CANCEL_OFFLINE:
555 default:
556 break;
557 }
558
559 return NOTIFY_OK;
560}
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800561#endif /* CONFIG_HUGETLBFS */
562#else /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800563
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800564static int link_mem_sections(int nid) { return 0; }
565#endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
566
567#if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
568 !defined(CONFIG_HUGETLBFS)
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800569static inline int node_memory_callback(struct notifier_block *self,
570 unsigned long action, void *arg)
571{
572 return NOTIFY_OK;
573}
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800574
575static void init_node_hugetlb_work(int nid) { }
576
577#endif
Gary Hadec04fc582009-01-06 14:39:14 -0800578
Yasunori Goto0fc44152006-06-27 02:53:38 -0700579int register_one_node(int nid)
580{
581 int error = 0;
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700582 int cpu;
Yasunori Goto0fc44152006-06-27 02:53:38 -0700583
584 if (node_online(nid)) {
585 int p_node = parent_node(nid);
586 struct node *parent = NULL;
587
588 if (p_node != nid)
Wen Congyang87327942012-12-11 16:00:56 -0800589 parent = node_devices[p_node];
Yasunori Goto0fc44152006-06-27 02:53:38 -0700590
Wen Congyang87327942012-12-11 16:00:56 -0800591 node_devices[nid] = kzalloc(sizeof(struct node), GFP_KERNEL);
592 if (!node_devices[nid])
593 return -ENOMEM;
594
595 error = register_node(node_devices[nid], nid, parent);
KAMEZAWA Hiroyuki76b67ed92006-06-27 02:53:41 -0700596
597 /* link cpu under this node */
598 for_each_present_cpu(cpu) {
599 if (cpu_to_node(cpu) == nid)
600 register_cpu_under_node(cpu, nid);
601 }
Gary Hadec04fc582009-01-06 14:39:14 -0800602
603 /* link memory sections under this node */
604 error = link_mem_sections(nid);
Lee Schermerhorn39da08c2009-12-14 17:58:36 -0800605
606 /* initialize work queue for memory hot plug */
607 init_node_hugetlb_work(nid);
Yasunori Goto0fc44152006-06-27 02:53:38 -0700608 }
609
610 return error;
611
612}
613
614void unregister_one_node(int nid)
615{
Xishi Qiu92d585e2014-03-06 17:18:21 +0800616 if (!node_devices[nid])
617 return;
618
Wen Congyang87327942012-12-11 16:00:56 -0800619 unregister_node(node_devices[nid]);
Wen Congyang87327942012-12-11 16:00:56 -0800620 node_devices[nid] = NULL;
Yasunori Goto0fc44152006-06-27 02:53:38 -0700621}
622
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700623/*
624 * node states attributes
625 */
626
627static ssize_t print_nodes_state(enum node_states state, char *buf)
628{
629 int n;
630
Tejun Heof799b1a2015-02-13 14:37:56 -0800631 n = scnprintf(buf, PAGE_SIZE - 1, "%*pbl",
632 nodemask_pr_args(&node_states[state]));
Ryota Ozakif6238812012-05-29 15:06:20 -0700633 buf[n++] = '\n';
634 buf[n] = '\0';
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700635 return n;
636}
637
Andi Kleenb15f5622010-01-05 12:47:59 +0100638struct node_attr {
Kay Sievers10fbcf42011-12-21 14:48:43 -0800639 struct device_attribute attr;
Andi Kleenb15f5622010-01-05 12:47:59 +0100640 enum node_states state;
641};
642
Kay Sievers10fbcf42011-12-21 14:48:43 -0800643static ssize_t show_node_state(struct device *dev,
644 struct device_attribute *attr, char *buf)
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700645{
Andi Kleenb15f5622010-01-05 12:47:59 +0100646 struct node_attr *na = container_of(attr, struct node_attr, attr);
647 return print_nodes_state(na->state, buf);
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700648}
649
Andi Kleenb15f5622010-01-05 12:47:59 +0100650#define _NODE_ATTR(name, state) \
Kay Sievers10fbcf42011-12-21 14:48:43 -0800651 { __ATTR(name, 0444, show_node_state, NULL), state }
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700652
Andi Kleenb15f5622010-01-05 12:47:59 +0100653static struct node_attr node_state_attr[] = {
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800654 [N_POSSIBLE] = _NODE_ATTR(possible, N_POSSIBLE),
655 [N_ONLINE] = _NODE_ATTR(online, N_ONLINE),
656 [N_NORMAL_MEMORY] = _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700657#ifdef CONFIG_HIGHMEM
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800658 [N_HIGH_MEMORY] = _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700659#endif
Lai Jiangshan20b2f522012-12-12 13:52:00 -0800660#ifdef CONFIG_MOVABLE_NODE
661 [N_MEMORY] = _NODE_ATTR(has_memory, N_MEMORY),
662#endif
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800663 [N_CPU] = _NODE_ATTR(has_cpu, N_CPU),
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700664};
665
Kay Sievers10fbcf42011-12-21 14:48:43 -0800666static struct attribute *node_state_attrs[] = {
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800667 &node_state_attr[N_POSSIBLE].attr.attr,
668 &node_state_attr[N_ONLINE].attr.attr,
669 &node_state_attr[N_NORMAL_MEMORY].attr.attr,
Andi Kleen3701cde2010-01-05 12:48:04 +0100670#ifdef CONFIG_HIGHMEM
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800671 &node_state_attr[N_HIGH_MEMORY].attr.attr,
Andi Kleen3701cde2010-01-05 12:48:04 +0100672#endif
Lai Jiangshan20b2f522012-12-12 13:52:00 -0800673#ifdef CONFIG_MOVABLE_NODE
674 &node_state_attr[N_MEMORY].attr.attr,
675#endif
Lai Jiangshanfcf07d22012-12-11 16:03:13 -0800676 &node_state_attr[N_CPU].attr.attr,
Andi Kleen3701cde2010-01-05 12:48:04 +0100677 NULL
678};
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700679
Kay Sievers10fbcf42011-12-21 14:48:43 -0800680static struct attribute_group memory_root_attr_group = {
681 .attrs = node_state_attrs,
682};
683
684static const struct attribute_group *cpu_root_attr_groups[] = {
685 &memory_root_attr_group,
686 NULL,
687};
688
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800689#define NODE_CALLBACK_PRI 2 /* lower than SLAB */
Keiichiro Tokunaga4b450992005-05-08 21:28:53 +0900690static int __init register_node_type(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691{
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700692 int ret;
693
Andi Kleen3701cde2010-01-05 12:48:04 +0100694 BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
695 BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
696
Kay Sievers10fbcf42011-12-21 14:48:43 -0800697 ret = subsys_system_register(&node_subsys, cpu_root_attr_groups);
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800698 if (!ret) {
Andrew Morton6e259e72013-04-29 15:08:07 -0700699 static struct notifier_block node_memory_callback_nb = {
700 .notifier_call = node_memory_callback,
701 .priority = NODE_CALLBACK_PRI,
702 };
703 register_hotmemory_notifier(&node_memory_callback_nb);
Lee Schermerhorn4faf8d92009-12-14 17:58:35 -0800704 }
Lee Schermerhornbde631a2007-10-16 01:26:27 -0700705
706 /*
707 * Note: we're not going to unregister the node class if we fail
708 * to register the node state class attribute files.
709 */
710 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711}
712postcore_initcall(register_node_type);