2 * drivers/base/node.c - basic Node class support
5 #include <linux/sysdev.h>
6 #include <linux/module.h>
7 #include <linux/init.h>
9 #include <linux/memory.h>
10 #include <linux/node.h>
11 #include <linux/hugetlb.h>
12 #include <linux/cpumask.h>
13 #include <linux/topology.h>
14 #include <linux/nodemask.h>
15 #include <linux/cpu.h>
16 #include <linux/device.h>
17 #include <linux/swap.h>
19 static struct sysdev_class node_class
= {
24 static ssize_t
node_read_cpumap(struct sys_device
*dev
, int type
, char *buf
)
26 struct node
*node_dev
= to_node(dev
);
27 const struct cpumask
*mask
= cpumask_of_node(node_dev
->sysdev
.id
);
30 /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
31 BUILD_BUG_ON((NR_CPUS
/32 * 9) > (PAGE_SIZE
-1));
34 cpulist_scnprintf(buf
, PAGE_SIZE
-2, mask
) :
35 cpumask_scnprintf(buf
, PAGE_SIZE
-2, mask
);
41 static inline ssize_t
node_read_cpumask(struct sys_device
*dev
,
42 struct sysdev_attribute
*attr
, char *buf
)
44 return node_read_cpumap(dev
, 0, buf
);
46 static inline ssize_t
node_read_cpulist(struct sys_device
*dev
,
47 struct sysdev_attribute
*attr
, char *buf
)
49 return node_read_cpumap(dev
, 1, buf
);
52 static SYSDEV_ATTR(cpumap
, S_IRUGO
, node_read_cpumask
, NULL
);
53 static SYSDEV_ATTR(cpulist
, S_IRUGO
, node_read_cpulist
, NULL
);
55 #define K(x) ((x) << (PAGE_SHIFT - 10))
56 static ssize_t
node_read_meminfo(struct sys_device
* dev
,
57 struct sysdev_attribute
*attr
, char * buf
)
63 si_meminfo_node(&i
, nid
);
66 "Node %d MemTotal: %8lu kB\n"
67 "Node %d MemFree: %8lu kB\n"
68 "Node %d MemUsed: %8lu kB\n"
69 "Node %d Active: %8lu kB\n"
70 "Node %d Inactive: %8lu kB\n"
71 "Node %d Active(anon): %8lu kB\n"
72 "Node %d Inactive(anon): %8lu kB\n"
73 "Node %d Active(file): %8lu kB\n"
74 "Node %d Inactive(file): %8lu kB\n"
75 "Node %d Unevictable: %8lu kB\n"
76 "Node %d Mlocked: %8lu kB\n"
78 "Node %d HighTotal: %8lu kB\n"
79 "Node %d HighFree: %8lu kB\n"
80 "Node %d LowTotal: %8lu kB\n"
81 "Node %d LowFree: %8lu kB\n"
83 "Node %d Dirty: %8lu kB\n"
84 "Node %d Writeback: %8lu kB\n"
85 "Node %d FilePages: %8lu kB\n"
86 "Node %d Mapped: %8lu kB\n"
87 "Node %d AnonPages: %8lu kB\n"
88 "Node %d Shmem: %8lu kB\n"
89 "Node %d KernelStack: %8lu kB\n"
90 "Node %d PageTables: %8lu kB\n"
91 "Node %d NFS_Unstable: %8lu kB\n"
92 "Node %d Bounce: %8lu kB\n"
93 "Node %d WritebackTmp: %8lu kB\n"
94 "Node %d Slab: %8lu kB\n"
95 "Node %d SReclaimable: %8lu kB\n"
96 "Node %d SUnreclaim: %8lu kB\n",
99 nid
, K(i
.totalram
- i
.freeram
),
100 nid
, K(node_page_state(nid
, NR_ACTIVE_ANON
) +
101 node_page_state(nid
, NR_ACTIVE_FILE
)),
102 nid
, K(node_page_state(nid
, NR_INACTIVE_ANON
) +
103 node_page_state(nid
, NR_INACTIVE_FILE
)),
104 nid
, K(node_page_state(nid
, NR_ACTIVE_ANON
)),
105 nid
, K(node_page_state(nid
, NR_INACTIVE_ANON
)),
106 nid
, K(node_page_state(nid
, NR_ACTIVE_FILE
)),
107 nid
, K(node_page_state(nid
, NR_INACTIVE_FILE
)),
108 nid
, K(node_page_state(nid
, NR_UNEVICTABLE
)),
109 nid
, K(node_page_state(nid
, NR_MLOCK
)),
110 #ifdef CONFIG_HIGHMEM
113 nid
, K(i
.totalram
- i
.totalhigh
),
114 nid
, K(i
.freeram
- i
.freehigh
),
116 nid
, K(node_page_state(nid
, NR_FILE_DIRTY
)),
117 nid
, K(node_page_state(nid
, NR_WRITEBACK
)),
118 nid
, K(node_page_state(nid
, NR_FILE_PAGES
)),
119 nid
, K(node_page_state(nid
, NR_FILE_MAPPED
)),
120 nid
, K(node_page_state(nid
, NR_ANON_PAGES
)),
121 nid
, K(node_page_state(nid
, NR_SHMEM
)),
122 nid
, node_page_state(nid
, NR_KERNEL_STACK
) *
124 nid
, K(node_page_state(nid
, NR_PAGETABLE
)),
125 nid
, K(node_page_state(nid
, NR_UNSTABLE_NFS
)),
126 nid
, K(node_page_state(nid
, NR_BOUNCE
)),
127 nid
, K(node_page_state(nid
, NR_WRITEBACK_TEMP
)),
128 nid
, K(node_page_state(nid
, NR_SLAB_RECLAIMABLE
) +
129 node_page_state(nid
, NR_SLAB_UNRECLAIMABLE
)),
130 nid
, K(node_page_state(nid
, NR_SLAB_RECLAIMABLE
)),
131 nid
, K(node_page_state(nid
, NR_SLAB_UNRECLAIMABLE
)));
132 n
+= hugetlb_report_node_meminfo(nid
, buf
+ n
);
137 static SYSDEV_ATTR(meminfo
, S_IRUGO
, node_read_meminfo
, NULL
);
139 static ssize_t
node_read_numastat(struct sys_device
* dev
,
140 struct sysdev_attribute
*attr
, char * buf
)
146 "interleave_hit %lu\n"
149 node_page_state(dev
->id
, NUMA_HIT
),
150 node_page_state(dev
->id
, NUMA_MISS
),
151 node_page_state(dev
->id
, NUMA_FOREIGN
),
152 node_page_state(dev
->id
, NUMA_INTERLEAVE_HIT
),
153 node_page_state(dev
->id
, NUMA_LOCAL
),
154 node_page_state(dev
->id
, NUMA_OTHER
));
156 static SYSDEV_ATTR(numastat
, S_IRUGO
, node_read_numastat
, NULL
);
158 static ssize_t
node_read_distance(struct sys_device
* dev
,
159 struct sysdev_attribute
*attr
, char * buf
)
165 /* buf currently PAGE_SIZE, need ~4 chars per node */
166 BUILD_BUG_ON(MAX_NUMNODES
*4 > PAGE_SIZE
/2);
168 for_each_online_node(i
)
169 len
+= sprintf(buf
+ len
, "%s%d", i
? " " : "", node_distance(nid
, i
));
171 len
+= sprintf(buf
+ len
, "\n");
174 static SYSDEV_ATTR(distance
, S_IRUGO
, node_read_distance
, NULL
);
176 #ifdef CONFIG_HUGETLBFS
178 * hugetlbfs per node attributes registration interface:
179 * When/if hugetlb[fs] subsystem initializes [sometime after this module],
180 * it will register its per node attributes for all online nodes with
181 * memory. It will also call register_hugetlbfs_with_node(), below, to
182 * register its attribute registration functions with this node driver.
183 * Once these hooks have been initialized, the node driver will call into
184 * the hugetlb module to [un]register attributes for hot-plugged nodes.
186 static node_registration_func_t __hugetlb_register_node
;
187 static node_registration_func_t __hugetlb_unregister_node
;
189 static inline bool hugetlb_register_node(struct node
*node
)
191 if (__hugetlb_register_node
&&
192 node_state(node
->sysdev
.id
, N_HIGH_MEMORY
)) {
193 __hugetlb_register_node(node
);
199 static inline void hugetlb_unregister_node(struct node
*node
)
201 if (__hugetlb_unregister_node
)
202 __hugetlb_unregister_node(node
);
205 void register_hugetlbfs_with_node(node_registration_func_t doregister
,
206 node_registration_func_t unregister
)
208 __hugetlb_register_node
= doregister
;
209 __hugetlb_unregister_node
= unregister
;
212 static inline void hugetlb_register_node(struct node
*node
) {}
214 static inline void hugetlb_unregister_node(struct node
*node
) {}
219 * register_node - Setup a sysfs device for a node.
220 * @num - Node number to use when creating the device.
222 * Initialize and register the node device.
224 int register_node(struct node
*node
, int num
, struct node
*parent
)
228 node
->sysdev
.id
= num
;
229 node
->sysdev
.cls
= &node_class
;
230 error
= sysdev_register(&node
->sysdev
);
233 sysdev_create_file(&node
->sysdev
, &attr_cpumap
);
234 sysdev_create_file(&node
->sysdev
, &attr_cpulist
);
235 sysdev_create_file(&node
->sysdev
, &attr_meminfo
);
236 sysdev_create_file(&node
->sysdev
, &attr_numastat
);
237 sysdev_create_file(&node
->sysdev
, &attr_distance
);
239 scan_unevictable_register_node(node
);
241 hugetlb_register_node(node
);
247 * unregister_node - unregister a node device
248 * @node: node going away
250 * Unregisters a node device @node. All the devices on the node must be
251 * unregistered before calling this function.
253 void unregister_node(struct node
*node
)
255 sysdev_remove_file(&node
->sysdev
, &attr_cpumap
);
256 sysdev_remove_file(&node
->sysdev
, &attr_cpulist
);
257 sysdev_remove_file(&node
->sysdev
, &attr_meminfo
);
258 sysdev_remove_file(&node
->sysdev
, &attr_numastat
);
259 sysdev_remove_file(&node
->sysdev
, &attr_distance
);
261 scan_unevictable_unregister_node(node
);
262 hugetlb_unregister_node(node
); /* no-op, if memoryless node */
264 sysdev_unregister(&node
->sysdev
);
267 struct node node_devices
[MAX_NUMNODES
];
270 * register cpu under node
272 int register_cpu_under_node(unsigned int cpu
, unsigned int nid
)
275 struct sys_device
*obj
;
277 if (!node_online(nid
))
280 obj
= get_cpu_sysdev(cpu
);
284 ret
= sysfs_create_link(&node_devices
[nid
].sysdev
.kobj
,
286 kobject_name(&obj
->kobj
));
290 return sysfs_create_link(&obj
->kobj
,
291 &node_devices
[nid
].sysdev
.kobj
,
292 kobject_name(&node_devices
[nid
].sysdev
.kobj
));
295 int unregister_cpu_under_node(unsigned int cpu
, unsigned int nid
)
297 struct sys_device
*obj
;
299 if (!node_online(nid
))
302 obj
= get_cpu_sysdev(cpu
);
306 sysfs_remove_link(&node_devices
[nid
].sysdev
.kobj
,
307 kobject_name(&obj
->kobj
));
308 sysfs_remove_link(&obj
->kobj
,
309 kobject_name(&node_devices
[nid
].sysdev
.kobj
));
314 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
315 #define page_initialized(page) (page->lru.next)
317 static int get_nid_for_pfn(unsigned long pfn
)
321 if (!pfn_valid_within(pfn
))
323 page
= pfn_to_page(pfn
);
324 if (!page_initialized(page
))
326 return pfn_to_nid(pfn
);
329 /* register memory section under specified node if it spans that node */
330 int register_mem_sect_under_node(struct memory_block
*mem_blk
, int nid
)
333 unsigned long pfn
, sect_start_pfn
, sect_end_pfn
;
337 if (!node_online(nid
))
339 sect_start_pfn
= section_nr_to_pfn(mem_blk
->phys_index
);
340 sect_end_pfn
= sect_start_pfn
+ PAGES_PER_SECTION
- 1;
341 for (pfn
= sect_start_pfn
; pfn
<= sect_end_pfn
; pfn
++) {
344 page_nid
= get_nid_for_pfn(pfn
);
349 ret
= sysfs_create_link_nowarn(&node_devices
[nid
].sysdev
.kobj
,
350 &mem_blk
->sysdev
.kobj
,
351 kobject_name(&mem_blk
->sysdev
.kobj
));
355 return sysfs_create_link_nowarn(&mem_blk
->sysdev
.kobj
,
356 &node_devices
[nid
].sysdev
.kobj
,
357 kobject_name(&node_devices
[nid
].sysdev
.kobj
));
359 /* mem section does not span the specified node */
363 /* unregister memory section under all nodes that it spans */
364 int unregister_mem_sect_under_nodes(struct memory_block
*mem_blk
)
366 NODEMASK_ALLOC(nodemask_t
, unlinked_nodes
, GFP_KERNEL
);
367 unsigned long pfn
, sect_start_pfn
, sect_end_pfn
;
370 NODEMASK_FREE(unlinked_nodes
);
375 nodes_clear(*unlinked_nodes
);
376 sect_start_pfn
= section_nr_to_pfn(mem_blk
->phys_index
);
377 sect_end_pfn
= sect_start_pfn
+ PAGES_PER_SECTION
- 1;
378 for (pfn
= sect_start_pfn
; pfn
<= sect_end_pfn
; pfn
++) {
381 nid
= get_nid_for_pfn(pfn
);
384 if (!node_online(nid
))
386 if (node_test_and_set(nid
, *unlinked_nodes
))
388 sysfs_remove_link(&node_devices
[nid
].sysdev
.kobj
,
389 kobject_name(&mem_blk
->sysdev
.kobj
));
390 sysfs_remove_link(&mem_blk
->sysdev
.kobj
,
391 kobject_name(&node_devices
[nid
].sysdev
.kobj
));
393 NODEMASK_FREE(unlinked_nodes
);
397 static int link_mem_sections(int nid
)
399 unsigned long start_pfn
= NODE_DATA(nid
)->node_start_pfn
;
400 unsigned long end_pfn
= start_pfn
+ NODE_DATA(nid
)->node_spanned_pages
;
404 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= PAGES_PER_SECTION
) {
405 unsigned long section_nr
= pfn_to_section_nr(pfn
);
406 struct mem_section
*mem_sect
;
407 struct memory_block
*mem_blk
;
410 if (!present_section_nr(section_nr
))
412 mem_sect
= __nr_to_section(section_nr
);
413 mem_blk
= find_memory_block(mem_sect
);
414 ret
= register_mem_sect_under_node(mem_blk
, nid
);
418 /* discard ref obtained in find_memory_block() */
419 kobject_put(&mem_blk
->sysdev
.kobj
);
424 #ifdef CONFIG_HUGETLBFS
426 * Handle per node hstate attribute [un]registration on transistions
427 * to/from memoryless state.
429 static void node_hugetlb_work(struct work_struct
*work
)
431 struct node
*node
= container_of(work
, struct node
, node_work
);
434 * We only get here when a node transitions to/from memoryless state.
435 * We can detect which transition occurred by examining whether the
436 * node has memory now. hugetlb_register_node() already check this
437 * so we try to register the attributes. If that fails, then the
438 * node has transitioned to memoryless, try to unregister the
441 if (!hugetlb_register_node(node
))
442 hugetlb_unregister_node(node
);
445 static void init_node_hugetlb_work(int nid
)
447 INIT_WORK(&node_devices
[nid
].node_work
, node_hugetlb_work
);
450 static int node_memory_callback(struct notifier_block
*self
,
451 unsigned long action
, void *arg
)
453 struct memory_notify
*mnb
= arg
;
454 int nid
= mnb
->status_change_nid
;
460 * offload per node hstate [un]registration to a work thread
461 * when transitioning to/from memoryless state.
463 if (nid
!= NUMA_NO_NODE
)
464 schedule_work(&node_devices
[nid
].node_work
);
467 case MEM_GOING_ONLINE
:
468 case MEM_GOING_OFFLINE
:
469 case MEM_CANCEL_ONLINE
:
470 case MEM_CANCEL_OFFLINE
:
477 #endif /* CONFIG_HUGETLBFS */
478 #else /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
480 static int link_mem_sections(int nid
) { return 0; }
481 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
483 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
484 !defined(CONFIG_HUGETLBFS)
485 static inline int node_memory_callback(struct notifier_block
*self
,
486 unsigned long action
, void *arg
)
491 static void init_node_hugetlb_work(int nid
) { }
495 int register_one_node(int nid
)
500 if (node_online(nid
)) {
501 int p_node
= parent_node(nid
);
502 struct node
*parent
= NULL
;
505 parent
= &node_devices
[p_node
];
507 error
= register_node(&node_devices
[nid
], nid
, parent
);
509 /* link cpu under this node */
510 for_each_present_cpu(cpu
) {
511 if (cpu_to_node(cpu
) == nid
)
512 register_cpu_under_node(cpu
, nid
);
515 /* link memory sections under this node */
516 error
= link_mem_sections(nid
);
518 /* initialize work queue for memory hot plug */
519 init_node_hugetlb_work(nid
);
526 void unregister_one_node(int nid
)
528 unregister_node(&node_devices
[nid
]);
532 * node states attributes
535 static ssize_t
print_nodes_state(enum node_states state
, char *buf
)
539 n
= nodelist_scnprintf(buf
, PAGE_SIZE
, node_states
[state
]);
540 if (n
> 0 && PAGE_SIZE
> n
+ 1) {
547 static ssize_t
print_nodes_possible(struct sysdev_class
*class, char *buf
)
549 return print_nodes_state(N_POSSIBLE
, buf
);
552 static ssize_t
print_nodes_online(struct sysdev_class
*class, char *buf
)
554 return print_nodes_state(N_ONLINE
, buf
);
557 static ssize_t
print_nodes_has_normal_memory(struct sysdev_class
*class,
560 return print_nodes_state(N_NORMAL_MEMORY
, buf
);
563 static ssize_t
print_nodes_has_cpu(struct sysdev_class
*class, char *buf
)
565 return print_nodes_state(N_CPU
, buf
);
568 static SYSDEV_CLASS_ATTR(possible
, 0444, print_nodes_possible
, NULL
);
569 static SYSDEV_CLASS_ATTR(online
, 0444, print_nodes_online
, NULL
);
570 static SYSDEV_CLASS_ATTR(has_normal_memory
, 0444, print_nodes_has_normal_memory
,
572 static SYSDEV_CLASS_ATTR(has_cpu
, 0444, print_nodes_has_cpu
, NULL
);
574 #ifdef CONFIG_HIGHMEM
575 static ssize_t
print_nodes_has_high_memory(struct sysdev_class
*class,
578 return print_nodes_state(N_HIGH_MEMORY
, buf
);
581 static SYSDEV_CLASS_ATTR(has_high_memory
, 0444, print_nodes_has_high_memory
,
585 struct sysdev_class_attribute
*node_state_attr
[] = {
588 &attr_has_normal_memory
,
589 #ifdef CONFIG_HIGHMEM
590 &attr_has_high_memory
,
595 static int node_states_init(void)
600 for (i
= 0; i
< NR_NODE_STATES
; i
++) {
602 ret
= sysdev_class_create_file(&node_class
, node_state_attr
[i
]);
609 #define NODE_CALLBACK_PRI 2 /* lower than SLAB */
610 static int __init
register_node_type(void)
614 ret
= sysdev_class_register(&node_class
);
616 ret
= node_states_init();
617 hotplug_memory_notifier(node_memory_callback
,
622 * Note: we're not going to unregister the node class if we fail
623 * to register the node state class attribute files.
627 postcore_initcall(register_node_type
);