Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * drivers/base/node.c - basic Node class support | |
3 | */ | |
4 | ||
5 | #include <linux/sysdev.h> | |
6 | #include <linux/module.h> | |
7 | #include <linux/init.h> | |
8 | #include <linux/mm.h> | |
c04fc586 | 9 | #include <linux/memory.h> |
fa25c503 | 10 | #include <linux/vmstat.h> |
1da177e4 LT |
11 | #include <linux/node.h> |
12 | #include <linux/hugetlb.h> | |
ed4a6d7f | 13 | #include <linux/compaction.h> |
1da177e4 LT |
14 | #include <linux/cpumask.h> |
15 | #include <linux/topology.h> | |
16 | #include <linux/nodemask.h> | |
76b67ed9 | 17 | #include <linux/cpu.h> |
bde631a5 | 18 | #include <linux/device.h> |
af936a16 | 19 | #include <linux/swap.h> |
18e5b539 | 20 | #include <linux/slab.h> |
1da177e4 | 21 | |
3701cde6 AK |
22 | static struct sysdev_class_attribute *node_state_attrs[]; |
23 | ||
1da177e4 | 24 | static struct sysdev_class node_class = { |
af5ca3f4 | 25 | .name = "node", |
3701cde6 | 26 | .attrs = node_state_attrs, |
1da177e4 LT |
27 | }; |
28 | ||
29 | ||
39106dcf | 30 | static ssize_t node_read_cpumap(struct sys_device *dev, int type, char *buf) |
1da177e4 LT |
31 | { |
32 | struct node *node_dev = to_node(dev); | |
a70f7302 | 33 | const struct cpumask *mask = cpumask_of_node(node_dev->sysdev.id); |
1da177e4 LT |
34 | int len; |
35 | ||
39106dcf MT |
36 | /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */ |
37 | BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1)); | |
1da177e4 | 38 | |
39106dcf | 39 | len = type? |
29c0177e RR |
40 | cpulist_scnprintf(buf, PAGE_SIZE-2, mask) : |
41 | cpumask_scnprintf(buf, PAGE_SIZE-2, mask); | |
c5f59f08 MT |
42 | buf[len++] = '\n'; |
43 | buf[len] = '\0'; | |
1da177e4 LT |
44 | return len; |
45 | } | |
46 | ||
4a0b2b4d AK |
47 | static inline ssize_t node_read_cpumask(struct sys_device *dev, |
48 | struct sysdev_attribute *attr, char *buf) | |
39106dcf MT |
49 | { |
50 | return node_read_cpumap(dev, 0, buf); | |
51 | } | |
4a0b2b4d AK |
52 | static inline ssize_t node_read_cpulist(struct sys_device *dev, |
53 | struct sysdev_attribute *attr, char *buf) | |
39106dcf MT |
54 | { |
55 | return node_read_cpumap(dev, 1, buf); | |
56 | } | |
57 | ||
58 | static SYSDEV_ATTR(cpumap, S_IRUGO, node_read_cpumask, NULL); | |
59 | static SYSDEV_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL); | |
1da177e4 LT |
60 | |
61 | #define K(x) ((x) << (PAGE_SHIFT - 10)) | |
4a0b2b4d AK |
62 | static ssize_t node_read_meminfo(struct sys_device * dev, |
63 | struct sysdev_attribute *attr, char * buf) | |
1da177e4 LT |
64 | { |
65 | int n; | |
66 | int nid = dev->id; | |
67 | struct sysinfo i; | |
1da177e4 LT |
68 | |
69 | si_meminfo_node(&i, nid); | |
7ee92255 | 70 | n = sprintf(buf, |
4f98a2fe RR |
71 | "Node %d MemTotal: %8lu kB\n" |
72 | "Node %d MemFree: %8lu kB\n" | |
73 | "Node %d MemUsed: %8lu kB\n" | |
74 | "Node %d Active: %8lu kB\n" | |
75 | "Node %d Inactive: %8lu kB\n" | |
76 | "Node %d Active(anon): %8lu kB\n" | |
77 | "Node %d Inactive(anon): %8lu kB\n" | |
78 | "Node %d Active(file): %8lu kB\n" | |
79 | "Node %d Inactive(file): %8lu kB\n" | |
5344b7e6 | 80 | "Node %d Unevictable: %8lu kB\n" |
7ee92255 KM |
81 | "Node %d Mlocked: %8lu kB\n", |
82 | nid, K(i.totalram), | |
83 | nid, K(i.freeram), | |
84 | nid, K(i.totalram - i.freeram), | |
85 | nid, K(node_page_state(nid, NR_ACTIVE_ANON) + | |
86 | node_page_state(nid, NR_ACTIVE_FILE)), | |
87 | nid, K(node_page_state(nid, NR_INACTIVE_ANON) + | |
88 | node_page_state(nid, NR_INACTIVE_FILE)), | |
89 | nid, K(node_page_state(nid, NR_ACTIVE_ANON)), | |
90 | nid, K(node_page_state(nid, NR_INACTIVE_ANON)), | |
91 | nid, K(node_page_state(nid, NR_ACTIVE_FILE)), | |
92 | nid, K(node_page_state(nid, NR_INACTIVE_FILE)), | |
93 | nid, K(node_page_state(nid, NR_UNEVICTABLE)), | |
94 | nid, K(node_page_state(nid, NR_MLOCK))); | |
95 | ||
182e8e23 | 96 | #ifdef CONFIG_HIGHMEM |
7ee92255 | 97 | n += sprintf(buf + n, |
4f98a2fe RR |
98 | "Node %d HighTotal: %8lu kB\n" |
99 | "Node %d HighFree: %8lu kB\n" | |
100 | "Node %d LowTotal: %8lu kB\n" | |
7ee92255 KM |
101 | "Node %d LowFree: %8lu kB\n", |
102 | nid, K(i.totalhigh), | |
103 | nid, K(i.freehigh), | |
104 | nid, K(i.totalram - i.totalhigh), | |
105 | nid, K(i.freeram - i.freehigh)); | |
182e8e23 | 106 | #endif |
7ee92255 | 107 | n += sprintf(buf + n, |
4f98a2fe RR |
108 | "Node %d Dirty: %8lu kB\n" |
109 | "Node %d Writeback: %8lu kB\n" | |
110 | "Node %d FilePages: %8lu kB\n" | |
111 | "Node %d Mapped: %8lu kB\n" | |
112 | "Node %d AnonPages: %8lu kB\n" | |
4b02108a | 113 | "Node %d Shmem: %8lu kB\n" |
c6a7f572 | 114 | "Node %d KernelStack: %8lu kB\n" |
4f98a2fe RR |
115 | "Node %d PageTables: %8lu kB\n" |
116 | "Node %d NFS_Unstable: %8lu kB\n" | |
117 | "Node %d Bounce: %8lu kB\n" | |
118 | "Node %d WritebackTmp: %8lu kB\n" | |
119 | "Node %d Slab: %8lu kB\n" | |
120 | "Node %d SReclaimable: %8lu kB\n" | |
05b258e9 DR |
121 | "Node %d SUnreclaim: %8lu kB\n" |
122 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
123 | "Node %d AnonHugePages: %8lu kB\n" | |
124 | #endif | |
125 | , | |
b1e7a8fd | 126 | nid, K(node_page_state(nid, NR_FILE_DIRTY)), |
ce866b34 | 127 | nid, K(node_page_state(nid, NR_WRITEBACK)), |
347ce434 | 128 | nid, K(node_page_state(nid, NR_FILE_PAGES)), |
65ba55f5 | 129 | nid, K(node_page_state(nid, NR_FILE_MAPPED)), |
05b258e9 DR |
130 | nid, K(node_page_state(nid, NR_ANON_PAGES) |
131 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
132 | + node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) * | |
133 | HPAGE_PMD_NR | |
134 | #endif | |
135 | ), | |
4b02108a | 136 | nid, K(node_page_state(nid, NR_SHMEM)), |
c6a7f572 KM |
137 | nid, node_page_state(nid, NR_KERNEL_STACK) * |
138 | THREAD_SIZE / 1024, | |
df849a15 | 139 | nid, K(node_page_state(nid, NR_PAGETABLE)), |
fd39fc85 | 140 | nid, K(node_page_state(nid, NR_UNSTABLE_NFS)), |
d2c5e30c | 141 | nid, K(node_page_state(nid, NR_BOUNCE)), |
fc3ba692 | 142 | nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)), |
972d1a7b CL |
143 | nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) + |
144 | node_page_state(nid, NR_SLAB_UNRECLAIMABLE)), | |
145 | nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)), | |
05b258e9 DR |
146 | nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)) |
147 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
148 | , nid, | |
149 | K(node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) * | |
150 | HPAGE_PMD_NR) | |
151 | #endif | |
152 | ); | |
1da177e4 LT |
153 | n += hugetlb_report_node_meminfo(nid, buf + n); |
154 | return n; | |
155 | } | |
156 | ||
157 | #undef K | |
158 | static SYSDEV_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL); | |
159 | ||
4a0b2b4d AK |
160 | static ssize_t node_read_numastat(struct sys_device * dev, |
161 | struct sysdev_attribute *attr, char * buf) | |
1da177e4 | 162 | { |
1da177e4 LT |
163 | return sprintf(buf, |
164 | "numa_hit %lu\n" | |
165 | "numa_miss %lu\n" | |
166 | "numa_foreign %lu\n" | |
167 | "interleave_hit %lu\n" | |
168 | "local_node %lu\n" | |
169 | "other_node %lu\n", | |
ca889e6c CL |
170 | node_page_state(dev->id, NUMA_HIT), |
171 | node_page_state(dev->id, NUMA_MISS), | |
172 | node_page_state(dev->id, NUMA_FOREIGN), | |
173 | node_page_state(dev->id, NUMA_INTERLEAVE_HIT), | |
174 | node_page_state(dev->id, NUMA_LOCAL), | |
175 | node_page_state(dev->id, NUMA_OTHER)); | |
1da177e4 LT |
176 | } |
177 | static SYSDEV_ATTR(numastat, S_IRUGO, node_read_numastat, NULL); | |
178 | ||
2ac39037 MR |
179 | static ssize_t node_read_vmstat(struct sys_device *dev, |
180 | struct sysdev_attribute *attr, char *buf) | |
181 | { | |
182 | int nid = dev->id; | |
fa25c503 KM |
183 | int i; |
184 | int n = 0; | |
185 | ||
186 | for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) | |
187 | n += sprintf(buf+n, "%s %lu\n", vmstat_text[i], | |
188 | node_page_state(nid, i)); | |
189 | ||
190 | return n; | |
2ac39037 MR |
191 | } |
192 | static SYSDEV_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL); | |
193 | ||
4a0b2b4d AK |
194 | static ssize_t node_read_distance(struct sys_device * dev, |
195 | struct sysdev_attribute *attr, char * buf) | |
1da177e4 LT |
196 | { |
197 | int nid = dev->id; | |
198 | int len = 0; | |
199 | int i; | |
200 | ||
12ee3c0a DR |
201 | /* |
202 | * buf is currently PAGE_SIZE in length and each node needs 4 chars | |
203 | * at the most (distance + space or newline). | |
204 | */ | |
205 | BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE); | |
1da177e4 LT |
206 | |
207 | for_each_online_node(i) | |
208 | len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i)); | |
209 | ||
210 | len += sprintf(buf + len, "\n"); | |
211 | return len; | |
212 | } | |
213 | static SYSDEV_ATTR(distance, S_IRUGO, node_read_distance, NULL); | |
214 | ||
9a305230 LS |
215 | #ifdef CONFIG_HUGETLBFS |
216 | /* | |
217 | * hugetlbfs per node attributes registration interface: | |
218 | * When/if hugetlb[fs] subsystem initializes [sometime after this module], | |
4faf8d95 LS |
219 | * it will register its per node attributes for all online nodes with |
220 | * memory. It will also call register_hugetlbfs_with_node(), below, to | |
9a305230 LS |
221 | * register its attribute registration functions with this node driver. |
222 | * Once these hooks have been initialized, the node driver will call into | |
223 | * the hugetlb module to [un]register attributes for hot-plugged nodes. | |
224 | */ | |
225 | static node_registration_func_t __hugetlb_register_node; | |
226 | static node_registration_func_t __hugetlb_unregister_node; | |
227 | ||
39da08cb | 228 | static inline bool hugetlb_register_node(struct node *node) |
9a305230 | 229 | { |
4faf8d95 | 230 | if (__hugetlb_register_node && |
39da08cb | 231 | node_state(node->sysdev.id, N_HIGH_MEMORY)) { |
9a305230 | 232 | __hugetlb_register_node(node); |
39da08cb LS |
233 | return true; |
234 | } | |
235 | return false; | |
9a305230 LS |
236 | } |
237 | ||
238 | static inline void hugetlb_unregister_node(struct node *node) | |
239 | { | |
240 | if (__hugetlb_unregister_node) | |
241 | __hugetlb_unregister_node(node); | |
242 | } | |
243 | ||
244 | void register_hugetlbfs_with_node(node_registration_func_t doregister, | |
245 | node_registration_func_t unregister) | |
246 | { | |
247 | __hugetlb_register_node = doregister; | |
248 | __hugetlb_unregister_node = unregister; | |
249 | } | |
250 | #else | |
251 | static inline void hugetlb_register_node(struct node *node) {} | |
252 | ||
253 | static inline void hugetlb_unregister_node(struct node *node) {} | |
254 | #endif | |
255 | ||
1da177e4 LT |
256 | |
257 | /* | |
405ae7d3 | 258 | * register_node - Setup a sysfs device for a node. |
1da177e4 LT |
259 | * @num - Node number to use when creating the device. |
260 | * | |
261 | * Initialize and register the node device. | |
262 | */ | |
4b45099b | 263 | int register_node(struct node *node, int num, struct node *parent) |
1da177e4 LT |
264 | { |
265 | int error; | |
266 | ||
267 | node->sysdev.id = num; | |
268 | node->sysdev.cls = &node_class; | |
269 | error = sysdev_register(&node->sysdev); | |
270 | ||
271 | if (!error){ | |
272 | sysdev_create_file(&node->sysdev, &attr_cpumap); | |
39106dcf | 273 | sysdev_create_file(&node->sysdev, &attr_cpulist); |
1da177e4 LT |
274 | sysdev_create_file(&node->sysdev, &attr_meminfo); |
275 | sysdev_create_file(&node->sysdev, &attr_numastat); | |
276 | sysdev_create_file(&node->sysdev, &attr_distance); | |
2ac39037 | 277 | sysdev_create_file(&node->sysdev, &attr_vmstat); |
af936a16 LS |
278 | |
279 | scan_unevictable_register_node(node); | |
4faf8d95 | 280 | |
9a305230 | 281 | hugetlb_register_node(node); |
ed4a6d7f MG |
282 | |
283 | compaction_register_node(node); | |
1da177e4 LT |
284 | } |
285 | return error; | |
286 | } | |
287 | ||
4b45099b KT |
288 | /** |
289 | * unregister_node - unregister a node device | |
290 | * @node: node going away | |
291 | * | |
292 | * Unregisters a node device @node. All the devices on the node must be | |
293 | * unregistered before calling this function. | |
294 | */ | |
295 | void unregister_node(struct node *node) | |
296 | { | |
297 | sysdev_remove_file(&node->sysdev, &attr_cpumap); | |
39106dcf | 298 | sysdev_remove_file(&node->sysdev, &attr_cpulist); |
4b45099b KT |
299 | sysdev_remove_file(&node->sysdev, &attr_meminfo); |
300 | sysdev_remove_file(&node->sysdev, &attr_numastat); | |
301 | sysdev_remove_file(&node->sysdev, &attr_distance); | |
2ac39037 | 302 | sysdev_remove_file(&node->sysdev, &attr_vmstat); |
4b45099b | 303 | |
af936a16 | 304 | scan_unevictable_unregister_node(node); |
4faf8d95 | 305 | hugetlb_unregister_node(node); /* no-op, if memoryless node */ |
af936a16 | 306 | |
4b45099b KT |
307 | sysdev_unregister(&node->sysdev); |
308 | } | |
1da177e4 | 309 | |
0fc44159 YG |
310 | struct node node_devices[MAX_NUMNODES]; |
311 | ||
76b67ed9 KH |
312 | /* |
313 | * register cpu under node | |
314 | */ | |
315 | int register_cpu_under_node(unsigned int cpu, unsigned int nid) | |
316 | { | |
1830794a | 317 | int ret; |
f8246f31 | 318 | struct sys_device *obj; |
76b67ed9 | 319 | |
f8246f31 AC |
320 | if (!node_online(nid)) |
321 | return 0; | |
322 | ||
323 | obj = get_cpu_sysdev(cpu); | |
324 | if (!obj) | |
325 | return 0; | |
326 | ||
1830794a | 327 | ret = sysfs_create_link(&node_devices[nid].sysdev.kobj, |
f8246f31 AC |
328 | &obj->kobj, |
329 | kobject_name(&obj->kobj)); | |
1830794a AC |
330 | if (ret) |
331 | return ret; | |
332 | ||
333 | return sysfs_create_link(&obj->kobj, | |
334 | &node_devices[nid].sysdev.kobj, | |
335 | kobject_name(&node_devices[nid].sysdev.kobj)); | |
76b67ed9 KH |
336 | } |
337 | ||
338 | int unregister_cpu_under_node(unsigned int cpu, unsigned int nid) | |
339 | { | |
b9d52dad AC |
340 | struct sys_device *obj; |
341 | ||
342 | if (!node_online(nid)) | |
343 | return 0; | |
344 | ||
345 | obj = get_cpu_sysdev(cpu); | |
346 | if (!obj) | |
347 | return 0; | |
348 | ||
349 | sysfs_remove_link(&node_devices[nid].sysdev.kobj, | |
350 | kobject_name(&obj->kobj)); | |
1830794a AC |
351 | sysfs_remove_link(&obj->kobj, |
352 | kobject_name(&node_devices[nid].sysdev.kobj)); | |
b9d52dad | 353 | |
76b67ed9 KH |
354 | return 0; |
355 | } | |
356 | ||
c04fc586 GH |
357 | #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE |
358 | #define page_initialized(page) (page->lru.next) | |
359 | ||
360 | static int get_nid_for_pfn(unsigned long pfn) | |
361 | { | |
362 | struct page *page; | |
363 | ||
364 | if (!pfn_valid_within(pfn)) | |
365 | return -1; | |
366 | page = pfn_to_page(pfn); | |
367 | if (!page_initialized(page)) | |
368 | return -1; | |
369 | return pfn_to_nid(pfn); | |
370 | } | |
371 | ||
372 | /* register memory section under specified node if it spans that node */ | |
373 | int register_mem_sect_under_node(struct memory_block *mem_blk, int nid) | |
374 | { | |
dee5d0d5 | 375 | int ret; |
c04fc586 GH |
376 | unsigned long pfn, sect_start_pfn, sect_end_pfn; |
377 | ||
378 | if (!mem_blk) | |
379 | return -EFAULT; | |
380 | if (!node_online(nid)) | |
381 | return 0; | |
d3360164 NF |
382 | |
383 | sect_start_pfn = section_nr_to_pfn(mem_blk->start_section_nr); | |
384 | sect_end_pfn = section_nr_to_pfn(mem_blk->end_section_nr); | |
385 | sect_end_pfn += PAGES_PER_SECTION - 1; | |
c04fc586 GH |
386 | for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) { |
387 | int page_nid; | |
388 | ||
389 | page_nid = get_nid_for_pfn(pfn); | |
390 | if (page_nid < 0) | |
391 | continue; | |
392 | if (page_nid != nid) | |
393 | continue; | |
dee5d0d5 | 394 | ret = sysfs_create_link_nowarn(&node_devices[nid].sysdev.kobj, |
c04fc586 GH |
395 | &mem_blk->sysdev.kobj, |
396 | kobject_name(&mem_blk->sysdev.kobj)); | |
dee5d0d5 AC |
397 | if (ret) |
398 | return ret; | |
399 | ||
400 | return sysfs_create_link_nowarn(&mem_blk->sysdev.kobj, | |
401 | &node_devices[nid].sysdev.kobj, | |
402 | kobject_name(&node_devices[nid].sysdev.kobj)); | |
c04fc586 GH |
403 | } |
404 | /* mem section does not span the specified node */ | |
405 | return 0; | |
406 | } | |
407 | ||
408 | /* unregister memory section under all nodes that it spans */ | |
d3360164 NF |
409 | int unregister_mem_sect_under_nodes(struct memory_block *mem_blk, |
410 | unsigned long phys_index) | |
c04fc586 | 411 | { |
9ae49fab | 412 | NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL); |
c04fc586 GH |
413 | unsigned long pfn, sect_start_pfn, sect_end_pfn; |
414 | ||
9ae49fab DR |
415 | if (!mem_blk) { |
416 | NODEMASK_FREE(unlinked_nodes); | |
c04fc586 | 417 | return -EFAULT; |
9ae49fab DR |
418 | } |
419 | if (!unlinked_nodes) | |
420 | return -ENOMEM; | |
421 | nodes_clear(*unlinked_nodes); | |
d3360164 NF |
422 | |
423 | sect_start_pfn = section_nr_to_pfn(phys_index); | |
c04fc586 GH |
424 | sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1; |
425 | for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) { | |
47504980 | 426 | int nid; |
c04fc586 GH |
427 | |
428 | nid = get_nid_for_pfn(pfn); | |
429 | if (nid < 0) | |
430 | continue; | |
431 | if (!node_online(nid)) | |
432 | continue; | |
9ae49fab | 433 | if (node_test_and_set(nid, *unlinked_nodes)) |
c04fc586 GH |
434 | continue; |
435 | sysfs_remove_link(&node_devices[nid].sysdev.kobj, | |
436 | kobject_name(&mem_blk->sysdev.kobj)); | |
dee5d0d5 AC |
437 | sysfs_remove_link(&mem_blk->sysdev.kobj, |
438 | kobject_name(&node_devices[nid].sysdev.kobj)); | |
c04fc586 | 439 | } |
9ae49fab | 440 | NODEMASK_FREE(unlinked_nodes); |
c04fc586 GH |
441 | return 0; |
442 | } | |
443 | ||
444 | static int link_mem_sections(int nid) | |
445 | { | |
446 | unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn; | |
447 | unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages; | |
448 | unsigned long pfn; | |
63d027a6 | 449 | struct memory_block *mem_blk = NULL; |
c04fc586 GH |
450 | int err = 0; |
451 | ||
452 | for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) { | |
453 | unsigned long section_nr = pfn_to_section_nr(pfn); | |
454 | struct mem_section *mem_sect; | |
c04fc586 GH |
455 | int ret; |
456 | ||
457 | if (!present_section_nr(section_nr)) | |
458 | continue; | |
459 | mem_sect = __nr_to_section(section_nr); | |
63d027a6 | 460 | mem_blk = find_memory_block_hinted(mem_sect, mem_blk); |
c04fc586 GH |
461 | ret = register_mem_sect_under_node(mem_blk, nid); |
462 | if (!err) | |
463 | err = ret; | |
464 | ||
465 | /* discard ref obtained in find_memory_block() */ | |
c04fc586 | 466 | } |
63d027a6 RH |
467 | |
468 | if (mem_blk) | |
469 | kobject_put(&mem_blk->sysdev.kobj); | |
c04fc586 GH |
470 | return err; |
471 | } | |
4faf8d95 | 472 | |
39da08cb | 473 | #ifdef CONFIG_HUGETLBFS |
4faf8d95 LS |
474 | /* |
475 | * Handle per node hstate attribute [un]registration on transistions | |
476 | * to/from memoryless state. | |
477 | */ | |
39da08cb LS |
478 | static void node_hugetlb_work(struct work_struct *work) |
479 | { | |
480 | struct node *node = container_of(work, struct node, node_work); | |
481 | ||
482 | /* | |
483 | * We only get here when a node transitions to/from memoryless state. | |
484 | * We can detect which transition occurred by examining whether the | |
485 | * node has memory now. hugetlb_register_node() already check this | |
486 | * so we try to register the attributes. If that fails, then the | |
487 | * node has transitioned to memoryless, try to unregister the | |
488 | * attributes. | |
489 | */ | |
490 | if (!hugetlb_register_node(node)) | |
491 | hugetlb_unregister_node(node); | |
492 | } | |
493 | ||
494 | static void init_node_hugetlb_work(int nid) | |
495 | { | |
496 | INIT_WORK(&node_devices[nid].node_work, node_hugetlb_work); | |
497 | } | |
4faf8d95 LS |
498 | |
499 | static int node_memory_callback(struct notifier_block *self, | |
500 | unsigned long action, void *arg) | |
501 | { | |
502 | struct memory_notify *mnb = arg; | |
503 | int nid = mnb->status_change_nid; | |
504 | ||
505 | switch (action) { | |
39da08cb LS |
506 | case MEM_ONLINE: |
507 | case MEM_OFFLINE: | |
508 | /* | |
509 | * offload per node hstate [un]registration to a work thread | |
510 | * when transitioning to/from memoryless state. | |
511 | */ | |
4faf8d95 | 512 | if (nid != NUMA_NO_NODE) |
39da08cb | 513 | schedule_work(&node_devices[nid].node_work); |
4faf8d95 | 514 | break; |
39da08cb | 515 | |
4faf8d95 LS |
516 | case MEM_GOING_ONLINE: |
517 | case MEM_GOING_OFFLINE: | |
518 | case MEM_CANCEL_ONLINE: | |
519 | case MEM_CANCEL_OFFLINE: | |
520 | default: | |
521 | break; | |
522 | } | |
523 | ||
524 | return NOTIFY_OK; | |
525 | } | |
39da08cb LS |
526 | #endif /* CONFIG_HUGETLBFS */ |
527 | #else /* !CONFIG_MEMORY_HOTPLUG_SPARSE */ | |
528 | ||
c04fc586 | 529 | static int link_mem_sections(int nid) { return 0; } |
39da08cb | 530 | #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */ |
4faf8d95 | 531 | |
39da08cb LS |
532 | #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \ |
533 | !defined(CONFIG_HUGETLBFS) | |
4faf8d95 LS |
534 | static inline int node_memory_callback(struct notifier_block *self, |
535 | unsigned long action, void *arg) | |
536 | { | |
537 | return NOTIFY_OK; | |
538 | } | |
39da08cb LS |
539 | |
540 | static void init_node_hugetlb_work(int nid) { } | |
541 | ||
542 | #endif | |
c04fc586 | 543 | |
0fc44159 YG |
544 | int register_one_node(int nid) |
545 | { | |
546 | int error = 0; | |
76b67ed9 | 547 | int cpu; |
0fc44159 YG |
548 | |
549 | if (node_online(nid)) { | |
550 | int p_node = parent_node(nid); | |
551 | struct node *parent = NULL; | |
552 | ||
553 | if (p_node != nid) | |
554 | parent = &node_devices[p_node]; | |
555 | ||
556 | error = register_node(&node_devices[nid], nid, parent); | |
76b67ed9 KH |
557 | |
558 | /* link cpu under this node */ | |
559 | for_each_present_cpu(cpu) { | |
560 | if (cpu_to_node(cpu) == nid) | |
561 | register_cpu_under_node(cpu, nid); | |
562 | } | |
c04fc586 GH |
563 | |
564 | /* link memory sections under this node */ | |
565 | error = link_mem_sections(nid); | |
39da08cb LS |
566 | |
567 | /* initialize work queue for memory hot plug */ | |
568 | init_node_hugetlb_work(nid); | |
0fc44159 YG |
569 | } |
570 | ||
571 | return error; | |
572 | ||
573 | } | |
574 | ||
575 | void unregister_one_node(int nid) | |
576 | { | |
577 | unregister_node(&node_devices[nid]); | |
578 | } | |
579 | ||
bde631a5 LS |
580 | /* |
581 | * node states attributes | |
582 | */ | |
583 | ||
584 | static ssize_t print_nodes_state(enum node_states state, char *buf) | |
585 | { | |
586 | int n; | |
587 | ||
588 | n = nodelist_scnprintf(buf, PAGE_SIZE, node_states[state]); | |
589 | if (n > 0 && PAGE_SIZE > n + 1) { | |
590 | *(buf + n++) = '\n'; | |
591 | *(buf + n++) = '\0'; | |
592 | } | |
593 | return n; | |
594 | } | |
595 | ||
b15f562f AK |
596 | struct node_attr { |
597 | struct sysdev_class_attribute attr; | |
598 | enum node_states state; | |
599 | }; | |
bde631a5 | 600 | |
b15f562f AK |
601 | static ssize_t show_node_state(struct sysdev_class *class, |
602 | struct sysdev_class_attribute *attr, char *buf) | |
bde631a5 | 603 | { |
b15f562f AK |
604 | struct node_attr *na = container_of(attr, struct node_attr, attr); |
605 | return print_nodes_state(na->state, buf); | |
bde631a5 LS |
606 | } |
607 | ||
b15f562f AK |
608 | #define _NODE_ATTR(name, state) \ |
609 | { _SYSDEV_CLASS_ATTR(name, 0444, show_node_state, NULL), state } | |
bde631a5 | 610 | |
b15f562f AK |
611 | static struct node_attr node_state_attr[] = { |
612 | _NODE_ATTR(possible, N_POSSIBLE), | |
613 | _NODE_ATTR(online, N_ONLINE), | |
614 | _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY), | |
615 | _NODE_ATTR(has_cpu, N_CPU), | |
bde631a5 | 616 | #ifdef CONFIG_HIGHMEM |
b15f562f | 617 | _NODE_ATTR(has_high_memory, N_HIGH_MEMORY), |
bde631a5 | 618 | #endif |
bde631a5 LS |
619 | }; |
620 | ||
3701cde6 AK |
621 | static struct sysdev_class_attribute *node_state_attrs[] = { |
622 | &node_state_attr[0].attr, | |
623 | &node_state_attr[1].attr, | |
624 | &node_state_attr[2].attr, | |
625 | &node_state_attr[3].attr, | |
626 | #ifdef CONFIG_HIGHMEM | |
627 | &node_state_attr[4].attr, | |
628 | #endif | |
629 | NULL | |
630 | }; | |
bde631a5 | 631 | |
4faf8d95 | 632 | #define NODE_CALLBACK_PRI 2 /* lower than SLAB */ |
4b45099b | 633 | static int __init register_node_type(void) |
1da177e4 | 634 | { |
bde631a5 LS |
635 | int ret; |
636 | ||
3701cde6 AK |
637 | BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES); |
638 | BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES); | |
639 | ||
bde631a5 | 640 | ret = sysdev_class_register(&node_class); |
4faf8d95 | 641 | if (!ret) { |
4faf8d95 LS |
642 | hotplug_memory_notifier(node_memory_callback, |
643 | NODE_CALLBACK_PRI); | |
644 | } | |
bde631a5 LS |
645 | |
646 | /* | |
647 | * Note: we're not going to unregister the node class if we fail | |
648 | * to register the node state class attribute files. | |
649 | */ | |
650 | return ret; | |
1da177e4 LT |
651 | } |
652 | postcore_initcall(register_node_type); |