2 * Basic Node interface support
5 #include <linux/module.h>
6 #include <linux/init.h>
8 #include <linux/memory.h>
9 #include <linux/vmstat.h>
10 #include <linux/notifier.h>
11 #include <linux/node.h>
12 #include <linux/hugetlb.h>
13 #include <linux/compaction.h>
14 #include <linux/cpumask.h>
15 #include <linux/topology.h>
16 #include <linux/nodemask.h>
17 #include <linux/cpu.h>
18 #include <linux/device.h>
19 #include <linux/swap.h>
20 #include <linux/slab.h>
22 static struct bus_type node_subsys
= {
28 static ssize_t
node_read_cpumap(struct device
*dev
, int type
, char *buf
)
30 struct node
*node_dev
= to_node(dev
);
31 const struct cpumask
*mask
= cpumask_of_node(node_dev
->dev
.id
);
34 /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
35 BUILD_BUG_ON((NR_CPUS
/32 * 9) > (PAGE_SIZE
-1));
38 cpulist_scnprintf(buf
, PAGE_SIZE
-2, mask
) :
39 cpumask_scnprintf(buf
, PAGE_SIZE
-2, mask
);
45 static inline ssize_t
node_read_cpumask(struct device
*dev
,
46 struct device_attribute
*attr
, char *buf
)
48 return node_read_cpumap(dev
, 0, buf
);
50 static inline ssize_t
node_read_cpulist(struct device
*dev
,
51 struct device_attribute
*attr
, char *buf
)
53 return node_read_cpumap(dev
, 1, buf
);
56 static DEVICE_ATTR(cpumap
, S_IRUGO
, node_read_cpumask
, NULL
);
57 static DEVICE_ATTR(cpulist
, S_IRUGO
, node_read_cpulist
, NULL
);
59 #define K(x) ((x) << (PAGE_SHIFT - 10))
60 static ssize_t
node_read_meminfo(struct device
*dev
,
61 struct device_attribute
*attr
, char *buf
)
67 si_meminfo_node(&i
, nid
);
69 "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"
78 "Node %d Unevictable: %8lu kB\n"
79 "Node %d Mlocked: %8lu kB\n",
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
)));
96 "Node %d HighTotal: %8lu kB\n"
97 "Node %d HighFree: %8lu kB\n"
98 "Node %d LowTotal: %8lu kB\n"
99 "Node %d LowFree: %8lu kB\n",
102 nid
, K(i
.totalram
- i
.totalhigh
),
103 nid
, K(i
.freeram
- i
.freehigh
));
105 n
+= sprintf(buf
+ n
,
106 "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"
111 "Node %d Shmem: %8lu kB\n"
112 "Node %d KernelStack: %8lu kB\n"
113 "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"
119 "Node %d SUnreclaim: %8lu kB\n"
120 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
121 "Node %d AnonHugePages: %8lu kB\n"
124 nid
, K(node_page_state(nid
, NR_FILE_DIRTY
)),
125 nid
, K(node_page_state(nid
, NR_WRITEBACK
)),
126 nid
, K(node_page_state(nid
, NR_FILE_PAGES
)),
127 nid
, K(node_page_state(nid
, NR_FILE_MAPPED
)),
128 nid
, K(node_page_state(nid
, NR_ANON_PAGES
)),
129 nid
, K(node_page_state(nid
, NR_SHMEM
)),
130 nid
, node_page_state(nid
, NR_KERNEL_STACK
) *
132 nid
, K(node_page_state(nid
, NR_PAGETABLE
)),
133 nid
, K(node_page_state(nid
, NR_UNSTABLE_NFS
)),
134 nid
, K(node_page_state(nid
, NR_BOUNCE
)),
135 nid
, K(node_page_state(nid
, NR_WRITEBACK_TEMP
)),
136 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
)),
139 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
140 nid
, K(node_page_state(nid
, NR_SLAB_UNRECLAIMABLE
))
142 K(node_page_state(nid
, NR_ANON_TRANSPARENT_HUGEPAGES
) *
145 nid
, K(node_page_state(nid
, NR_SLAB_UNRECLAIMABLE
)));
147 n
+= hugetlb_report_node_meminfo(nid
, buf
+ n
);
152 static DEVICE_ATTR(meminfo
, S_IRUGO
, node_read_meminfo
, NULL
);
154 static ssize_t
node_read_numastat(struct device
*dev
,
155 struct device_attribute
*attr
, char *buf
)
161 "interleave_hit %lu\n"
164 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
));
171 static DEVICE_ATTR(numastat
, S_IRUGO
, node_read_numastat
, NULL
);
173 static ssize_t
node_read_vmstat(struct device
*dev
,
174 struct device_attribute
*attr
, char *buf
)
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
));
186 static DEVICE_ATTR(vmstat
, S_IRUGO
, node_read_vmstat
, NULL
);
188 static ssize_t
node_read_distance(struct device
*dev
,
189 struct device_attribute
*attr
, char * buf
)
196 * buf is currently PAGE_SIZE in length and each node needs 4 chars
197 * at the most (distance + space or newline).
199 BUILD_BUG_ON(MAX_NUMNODES
* 4 > PAGE_SIZE
);
201 for_each_online_node(i
)
202 len
+= sprintf(buf
+ len
, "%s%d", i
? " " : "", node_distance(nid
, i
));
204 len
+= sprintf(buf
+ len
, "\n");
207 static DEVICE_ATTR(distance
, S_IRUGO
, node_read_distance
, NULL
);
209 #ifdef CONFIG_HUGETLBFS
211 * hugetlbfs per node attributes registration interface:
212 * When/if hugetlb[fs] subsystem initializes [sometime after this module],
213 * it will register its per node attributes for all online nodes with
214 * memory. It will also call register_hugetlbfs_with_node(), below, to
215 * 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.
219 static node_registration_func_t __hugetlb_register_node
;
220 static node_registration_func_t __hugetlb_unregister_node
;
222 static inline bool hugetlb_register_node(struct node
*node
)
224 if (__hugetlb_register_node
&&
225 node_state(node
->dev
.id
, N_MEMORY
)) {
226 __hugetlb_register_node(node
);
232 static inline void hugetlb_unregister_node(struct node
*node
)
234 if (__hugetlb_unregister_node
)
235 __hugetlb_unregister_node(node
);
238 void register_hugetlbfs_with_node(node_registration_func_t doregister
,
239 node_registration_func_t unregister
)
241 __hugetlb_register_node
= doregister
;
242 __hugetlb_unregister_node
= unregister
;
245 static inline void hugetlb_register_node(struct node
*node
) {}
247 static inline void hugetlb_unregister_node(struct node
*node
) {}
250 static void node_device_release(struct device
*dev
)
252 struct node
*node
= to_node(dev
);
254 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HUGETLBFS)
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.
261 * The work is using node->node_work, so we should
262 * flush work before freeing the memory.
264 flush_work(&node
->node_work
);
270 * register_node - Setup a sysfs device for a node.
271 * @num - Node number to use when creating the device.
273 * Initialize and register the node device.
275 static int register_node(struct node
*node
, int num
, struct node
*parent
)
280 node
->dev
.bus
= &node_subsys
;
281 node
->dev
.release
= node_device_release
;
282 error
= device_register(&node
->dev
);
285 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
);
292 scan_unevictable_register_node(node
);
294 hugetlb_register_node(node
);
296 compaction_register_node(node
);
302 * unregister_node - unregister a node device
303 * @node: node going away
305 * Unregisters a node device @node. All the devices on the node must be
306 * unregistered before calling this function.
308 void unregister_node(struct node
*node
)
310 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
);
317 scan_unevictable_unregister_node(node
);
318 hugetlb_unregister_node(node
); /* no-op, if memoryless node */
320 device_unregister(&node
->dev
);
323 struct node
*node_devices
[MAX_NUMNODES
];
326 * register cpu under node
328 int register_cpu_under_node(unsigned int cpu
, unsigned int nid
)
333 if (!node_online(nid
))
336 obj
= get_cpu_device(cpu
);
340 ret
= sysfs_create_link(&node_devices
[nid
]->dev
.kobj
,
342 kobject_name(&obj
->kobj
));
346 return sysfs_create_link(&obj
->kobj
,
347 &node_devices
[nid
]->dev
.kobj
,
348 kobject_name(&node_devices
[nid
]->dev
.kobj
));
351 int unregister_cpu_under_node(unsigned int cpu
, unsigned int nid
)
355 if (!node_online(nid
))
358 obj
= get_cpu_device(cpu
);
362 sysfs_remove_link(&node_devices
[nid
]->dev
.kobj
,
363 kobject_name(&obj
->kobj
));
364 sysfs_remove_link(&obj
->kobj
,
365 kobject_name(&node_devices
[nid
]->dev
.kobj
));
370 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
371 #define page_initialized(page) (page->lru.next)
373 static int get_nid_for_pfn(unsigned long pfn
)
377 if (!pfn_valid_within(pfn
))
379 page
= pfn_to_page(pfn
);
380 if (!page_initialized(page
))
382 return pfn_to_nid(pfn
);
385 /* register memory section under specified node if it spans that node */
386 int register_mem_sect_under_node(struct memory_block
*mem_blk
, int nid
)
389 unsigned long pfn
, sect_start_pfn
, sect_end_pfn
;
393 if (!node_online(nid
))
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;
399 for (pfn
= sect_start_pfn
; pfn
<= sect_end_pfn
; pfn
++) {
402 page_nid
= get_nid_for_pfn(pfn
);
407 ret
= sysfs_create_link_nowarn(&node_devices
[nid
]->dev
.kobj
,
409 kobject_name(&mem_blk
->dev
.kobj
));
413 return sysfs_create_link_nowarn(&mem_blk
->dev
.kobj
,
414 &node_devices
[nid
]->dev
.kobj
,
415 kobject_name(&node_devices
[nid
]->dev
.kobj
));
417 /* mem section does not span the specified node */
421 /* unregister memory section under all nodes that it spans */
422 int unregister_mem_sect_under_nodes(struct memory_block
*mem_blk
,
423 unsigned long phys_index
)
425 NODEMASK_ALLOC(nodemask_t
, unlinked_nodes
, GFP_KERNEL
);
426 unsigned long pfn
, sect_start_pfn
, sect_end_pfn
;
429 NODEMASK_FREE(unlinked_nodes
);
434 nodes_clear(*unlinked_nodes
);
436 sect_start_pfn
= section_nr_to_pfn(phys_index
);
437 sect_end_pfn
= sect_start_pfn
+ PAGES_PER_SECTION
- 1;
438 for (pfn
= sect_start_pfn
; pfn
<= sect_end_pfn
; pfn
++) {
441 nid
= get_nid_for_pfn(pfn
);
444 if (!node_online(nid
))
446 if (node_test_and_set(nid
, *unlinked_nodes
))
448 sysfs_remove_link(&node_devices
[nid
]->dev
.kobj
,
449 kobject_name(&mem_blk
->dev
.kobj
));
450 sysfs_remove_link(&mem_blk
->dev
.kobj
,
451 kobject_name(&node_devices
[nid
]->dev
.kobj
));
453 NODEMASK_FREE(unlinked_nodes
);
457 static int link_mem_sections(int nid
)
459 unsigned long start_pfn
= NODE_DATA(nid
)->node_start_pfn
;
460 unsigned long end_pfn
= start_pfn
+ NODE_DATA(nid
)->node_spanned_pages
;
462 struct memory_block
*mem_blk
= NULL
;
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
;
470 if (!present_section_nr(section_nr
))
472 mem_sect
= __nr_to_section(section_nr
);
474 /* same memblock ? */
476 if ((section_nr
>= mem_blk
->start_section_nr
) &&
477 (section_nr
<= mem_blk
->end_section_nr
))
480 mem_blk
= find_memory_block_hinted(mem_sect
, mem_blk
);
482 ret
= register_mem_sect_under_node(mem_blk
, nid
);
486 /* discard ref obtained in find_memory_block() */
490 kobject_put(&mem_blk
->dev
.kobj
);
494 #ifdef CONFIG_HUGETLBFS
496 * Handle per node hstate attribute [un]registration on transistions
497 * to/from memoryless state.
499 static void node_hugetlb_work(struct work_struct
*work
)
501 struct node
*node
= container_of(work
, struct node
, node_work
);
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
511 if (!hugetlb_register_node(node
))
512 hugetlb_unregister_node(node
);
515 static void init_node_hugetlb_work(int nid
)
517 INIT_WORK(&node_devices
[nid
]->node_work
, node_hugetlb_work
);
520 static int node_memory_callback(struct notifier_block
*self
,
521 unsigned long action
, void *arg
)
523 struct memory_notify
*mnb
= arg
;
524 int nid
= mnb
->status_change_nid
;
530 * offload per node hstate [un]registration to a work thread
531 * when transitioning to/from memoryless state.
533 if (nid
!= NUMA_NO_NODE
)
534 schedule_work(&node_devices
[nid
]->node_work
);
537 case MEM_GOING_ONLINE
:
538 case MEM_GOING_OFFLINE
:
539 case MEM_CANCEL_ONLINE
:
540 case MEM_CANCEL_OFFLINE
:
547 #endif /* CONFIG_HUGETLBFS */
548 #else /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
550 static int link_mem_sections(int nid
) { return 0; }
551 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
553 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
554 !defined(CONFIG_HUGETLBFS)
555 static inline int node_memory_callback(struct notifier_block
*self
,
556 unsigned long action
, void *arg
)
561 static void init_node_hugetlb_work(int nid
) { }
565 int register_one_node(int nid
)
570 if (node_online(nid
)) {
571 int p_node
= parent_node(nid
);
572 struct node
*parent
= NULL
;
575 parent
= node_devices
[p_node
];
577 node_devices
[nid
] = kzalloc(sizeof(struct node
), GFP_KERNEL
);
578 if (!node_devices
[nid
])
581 error
= register_node(node_devices
[nid
], nid
, parent
);
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
);
589 /* link memory sections under this node */
590 error
= link_mem_sections(nid
);
592 /* initialize work queue for memory hot plug */
593 init_node_hugetlb_work(nid
);
600 void unregister_one_node(int nid
)
602 if (!node_devices
[nid
])
605 unregister_node(node_devices
[nid
]);
606 kfree(node_devices
[nid
]);
607 node_devices
[nid
] = NULL
;
611 * node states attributes
614 static ssize_t
print_nodes_state(enum node_states state
, char *buf
)
618 n
= nodelist_scnprintf(buf
, PAGE_SIZE
-2, node_states
[state
]);
625 struct device_attribute attr
;
626 enum node_states state
;
629 static ssize_t
show_node_state(struct device
*dev
,
630 struct device_attribute
*attr
, char *buf
)
632 struct node_attr
*na
= container_of(attr
, struct node_attr
, attr
);
633 return print_nodes_state(na
->state
, buf
);
636 #define _NODE_ATTR(name, state) \
637 { __ATTR(name, 0444, show_node_state, NULL), state }
639 static struct node_attr node_state_attr
[] = {
640 [N_POSSIBLE
] = _NODE_ATTR(possible
, N_POSSIBLE
),
641 [N_ONLINE
] = _NODE_ATTR(online
, N_ONLINE
),
642 [N_NORMAL_MEMORY
] = _NODE_ATTR(has_normal_memory
, N_NORMAL_MEMORY
),
643 #ifdef CONFIG_HIGHMEM
644 [N_HIGH_MEMORY
] = _NODE_ATTR(has_high_memory
, N_HIGH_MEMORY
),
646 #ifdef CONFIG_MOVABLE_NODE
647 [N_MEMORY
] = _NODE_ATTR(has_memory
, N_MEMORY
),
649 [N_CPU
] = _NODE_ATTR(has_cpu
, N_CPU
),
652 static struct attribute
*node_state_attrs
[] = {
653 &node_state_attr
[N_POSSIBLE
].attr
.attr
,
654 &node_state_attr
[N_ONLINE
].attr
.attr
,
655 &node_state_attr
[N_NORMAL_MEMORY
].attr
.attr
,
656 #ifdef CONFIG_HIGHMEM
657 &node_state_attr
[N_HIGH_MEMORY
].attr
.attr
,
659 #ifdef CONFIG_MOVABLE_NODE
660 &node_state_attr
[N_MEMORY
].attr
.attr
,
662 &node_state_attr
[N_CPU
].attr
.attr
,
666 static struct attribute_group memory_root_attr_group
= {
667 .attrs
= node_state_attrs
,
670 static const struct attribute_group
*cpu_root_attr_groups
[] = {
671 &memory_root_attr_group
,
675 #define NODE_CALLBACK_PRI 2 /* lower than SLAB */
676 static int __init
register_node_type(void)
680 BUILD_BUG_ON(ARRAY_SIZE(node_state_attr
) != NR_NODE_STATES
);
681 BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs
)-1 != NR_NODE_STATES
);
683 ret
= subsys_system_register(&node_subsys
, cpu_root_attr_groups
);
685 static struct notifier_block node_memory_callback_nb
= {
686 .notifier_call
= node_memory_callback
,
687 .priority
= NODE_CALLBACK_PRI
,
689 register_hotmemory_notifier(&node_memory_callback_nb
);
693 * Note: we're not going to unregister the node class if we fail
694 * to register the node state class attribute files.
698 postcore_initcall(register_node_type
);