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, bool list, char *buf)
32 struct node *node_dev = to_node(dev);
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));
37 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
40 cpumask_and(mask, cpumask_of_node(node_dev->dev.id), cpu_online_mask);
41 n = cpumap_print_to_pagebuf(list, buf, mask);
42 free_cpumask_var(mask);
47 static inline ssize_t node_read_cpumask(struct device *dev,
48 struct device_attribute *attr, char *buf)
50 return node_read_cpumap(dev, false, buf);
52 static inline ssize_t node_read_cpulist(struct device *dev,
53 struct device_attribute *attr, char *buf)
55 return node_read_cpumap(dev, true, buf);
58 static DEVICE_ATTR(cpumap, S_IRUGO, node_read_cpumask, NULL);
59 static DEVICE_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
61 #define K(x) ((x) << (PAGE_SHIFT - 10))
62 static ssize_t node_read_meminfo(struct device *dev,
63 struct device_attribute *attr, char *buf)
67 struct pglist_data *pgdat = NODE_DATA(nid);
70 si_meminfo_node(&i, nid);
72 "Node %d MemTotal: %8lu kB\n"
73 "Node %d MemFree: %8lu kB\n"
74 "Node %d MemUsed: %8lu kB\n"
75 "Node %d Active: %8lu kB\n"
76 "Node %d Inactive: %8lu kB\n"
77 "Node %d Active(anon): %8lu kB\n"
78 "Node %d Inactive(anon): %8lu kB\n"
79 "Node %d Active(file): %8lu kB\n"
80 "Node %d Inactive(file): %8lu kB\n"
81 "Node %d Unevictable: %8lu kB\n"
82 "Node %d Mlocked: %8lu kB\n",
85 nid, K(i.totalram - i.freeram),
86 nid, K(node_page_state(pgdat, NR_ACTIVE_ANON) +
87 node_page_state(pgdat, NR_ACTIVE_FILE)),
88 nid, K(node_page_state(pgdat, NR_INACTIVE_ANON) +
89 node_page_state(pgdat, NR_INACTIVE_FILE)),
90 nid, K(node_page_state(pgdat, NR_ACTIVE_ANON)),
91 nid, K(node_page_state(pgdat, NR_INACTIVE_ANON)),
92 nid, K(node_page_state(pgdat, NR_ACTIVE_FILE)),
93 nid, K(node_page_state(pgdat, NR_INACTIVE_FILE)),
94 nid, K(node_page_state(pgdat, NR_UNEVICTABLE)),
95 nid, K(sum_zone_node_page_state(nid, NR_MLOCK)));
99 "Node %d HighTotal: %8lu kB\n"
100 "Node %d HighFree: %8lu kB\n"
101 "Node %d LowTotal: %8lu kB\n"
102 "Node %d LowFree: %8lu kB\n",
105 nid, K(i.totalram - i.totalhigh),
106 nid, K(i.freeram - i.freehigh));
108 n += sprintf(buf + n,
109 "Node %d Dirty: %8lu kB\n"
110 "Node %d Writeback: %8lu kB\n"
111 "Node %d FilePages: %8lu kB\n"
112 "Node %d Mapped: %8lu kB\n"
113 "Node %d AnonPages: %8lu kB\n"
114 "Node %d Shmem: %8lu kB\n"
115 "Node %d KernelStack: %8lu kB\n"
116 "Node %d PageTables: %8lu kB\n"
117 "Node %d NFS_Unstable: %8lu kB\n"
118 "Node %d Bounce: %8lu kB\n"
119 "Node %d WritebackTmp: %8lu kB\n"
120 "Node %d Slab: %8lu kB\n"
121 "Node %d SReclaimable: %8lu kB\n"
122 "Node %d SUnreclaim: %8lu kB\n"
123 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
124 "Node %d AnonHugePages: %8lu kB\n"
125 "Node %d ShmemHugePages: %8lu kB\n"
126 "Node %d ShmemPmdMapped: %8lu kB\n"
129 nid, K(node_page_state(pgdat, NR_FILE_DIRTY)),
130 nid, K(node_page_state(pgdat, NR_WRITEBACK)),
131 nid, K(node_page_state(pgdat, NR_FILE_PAGES)),
132 nid, K(node_page_state(pgdat, NR_FILE_MAPPED)),
133 nid, K(node_page_state(pgdat, NR_ANON_MAPPED)),
135 nid, sum_zone_node_page_state(nid, NR_KERNEL_STACK_KB),
136 nid, K(sum_zone_node_page_state(nid, NR_PAGETABLE)),
137 nid, K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
138 nid, K(sum_zone_node_page_state(nid, NR_BOUNCE)),
139 nid, K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
140 nid, K(node_page_state(pgdat, NR_SLAB_RECLAIMABLE) +
141 node_page_state(pgdat, NR_SLAB_UNRECLAIMABLE)),
142 nid, K(node_page_state(pgdat, NR_SLAB_RECLAIMABLE)),
143 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
144 nid, K(node_page_state(pgdat, NR_SLAB_UNRECLAIMABLE)),
145 nid, K(node_page_state(pgdat, NR_ANON_THPS) *
147 nid, K(node_page_state(pgdat, NR_SHMEM_THPS) *
149 nid, K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED) *
152 nid, K(node_page_state(pgdat, NR_SLAB_UNRECLAIMABLE)));
154 n += hugetlb_report_node_meminfo(nid, buf + n);
159 static DEVICE_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
161 static ssize_t node_read_numastat(struct device *dev,
162 struct device_attribute *attr, char *buf)
168 "interleave_hit %lu\n"
171 sum_zone_numa_state(dev->id, NUMA_HIT),
172 sum_zone_numa_state(dev->id, NUMA_MISS),
173 sum_zone_numa_state(dev->id, NUMA_FOREIGN),
174 sum_zone_numa_state(dev->id, NUMA_INTERLEAVE_HIT),
175 sum_zone_numa_state(dev->id, NUMA_LOCAL),
176 sum_zone_numa_state(dev->id, NUMA_OTHER));
178 static DEVICE_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
180 static ssize_t node_read_vmstat(struct device *dev,
181 struct device_attribute *attr, char *buf)
184 struct pglist_data *pgdat = NODE_DATA(nid);
188 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
189 n += sprintf(buf+n, "%s %lu\n", vmstat_text[i],
190 sum_zone_node_page_state(nid, i));
193 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
194 n += sprintf(buf+n, "%s %lu\n",
195 vmstat_text[i + NR_VM_ZONE_STAT_ITEMS],
196 sum_zone_numa_state(nid, i));
199 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
200 n += sprintf(buf+n, "%s %lu\n",
201 vmstat_text[i + NR_VM_ZONE_STAT_ITEMS +
202 NR_VM_NUMA_STAT_ITEMS],
203 node_page_state(pgdat, i));
207 static DEVICE_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
209 static ssize_t node_read_distance(struct device *dev,
210 struct device_attribute *attr, char *buf)
217 * buf is currently PAGE_SIZE in length and each node needs 4 chars
218 * at the most (distance + space or newline).
220 BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
222 for_each_online_node(i)
223 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
225 len += sprintf(buf + len, "\n");
228 static DEVICE_ATTR(distance, S_IRUGO, node_read_distance, NULL);
230 static struct attribute *node_dev_attrs[] = {
231 &dev_attr_cpumap.attr,
232 &dev_attr_cpulist.attr,
233 &dev_attr_meminfo.attr,
234 &dev_attr_numastat.attr,
235 &dev_attr_distance.attr,
236 &dev_attr_vmstat.attr,
239 ATTRIBUTE_GROUPS(node_dev);
241 #ifdef CONFIG_HUGETLBFS
243 * hugetlbfs per node attributes registration interface:
244 * When/if hugetlb[fs] subsystem initializes [sometime after this module],
245 * it will register its per node attributes for all online nodes with
246 * memory. It will also call register_hugetlbfs_with_node(), below, to
247 * register its attribute registration functions with this node driver.
248 * Once these hooks have been initialized, the node driver will call into
249 * the hugetlb module to [un]register attributes for hot-plugged nodes.
251 static node_registration_func_t __hugetlb_register_node;
252 static node_registration_func_t __hugetlb_unregister_node;
254 static inline bool hugetlb_register_node(struct node *node)
256 if (__hugetlb_register_node &&
257 node_state(node->dev.id, N_MEMORY)) {
258 __hugetlb_register_node(node);
264 static inline void hugetlb_unregister_node(struct node *node)
266 if (__hugetlb_unregister_node)
267 __hugetlb_unregister_node(node);
270 void register_hugetlbfs_with_node(node_registration_func_t doregister,
271 node_registration_func_t unregister)
273 __hugetlb_register_node = doregister;
274 __hugetlb_unregister_node = unregister;
277 static inline void hugetlb_register_node(struct node *node) {}
279 static inline void hugetlb_unregister_node(struct node *node) {}
282 static void node_device_release(struct device *dev)
284 struct node *node = to_node(dev);
286 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HUGETLBFS)
288 * We schedule the work only when a memory section is
289 * onlined/offlined on this node. When we come here,
290 * all the memory on this node has been offlined,
291 * so we won't enqueue new work to this work.
293 * The work is using node->node_work, so we should
294 * flush work before freeing the memory.
296 flush_work(&node->node_work);
302 * register_node - Setup a sysfs device for a node.
303 * @num - Node number to use when creating the device.
305 * Initialize and register the node device.
307 static int register_node(struct node *node, int num)
312 node->dev.bus = &node_subsys;
313 node->dev.release = node_device_release;
314 node->dev.groups = node_dev_groups;
315 error = device_register(&node->dev);
318 hugetlb_register_node(node);
320 compaction_register_node(node);
326 * unregister_node - unregister a node device
327 * @node: node going away
329 * Unregisters a node device @node. All the devices on the node must be
330 * unregistered before calling this function.
332 void unregister_node(struct node *node)
334 hugetlb_unregister_node(node); /* no-op, if memoryless node */
336 device_unregister(&node->dev);
339 struct node *node_devices[MAX_NUMNODES];
342 * register cpu under node
344 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
349 if (!node_online(nid))
352 obj = get_cpu_device(cpu);
356 ret = sysfs_create_link(&node_devices[nid]->dev.kobj,
358 kobject_name(&obj->kobj));
362 return sysfs_create_link(&obj->kobj,
363 &node_devices[nid]->dev.kobj,
364 kobject_name(&node_devices[nid]->dev.kobj));
367 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
371 if (!node_online(nid))
374 obj = get_cpu_device(cpu);
378 sysfs_remove_link(&node_devices[nid]->dev.kobj,
379 kobject_name(&obj->kobj));
380 sysfs_remove_link(&obj->kobj,
381 kobject_name(&node_devices[nid]->dev.kobj));
386 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
387 static int __ref get_nid_for_pfn(unsigned long pfn)
389 if (!pfn_valid_within(pfn))
391 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
392 if (system_state < SYSTEM_RUNNING)
393 return early_pfn_to_nid(pfn);
395 return pfn_to_nid(pfn);
398 /* register memory section under specified node if it spans that node */
399 int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
402 unsigned long pfn, sect_start_pfn, sect_end_pfn;
406 if (!node_online(nid))
409 sect_start_pfn = section_nr_to_pfn(mem_blk->start_section_nr);
410 sect_end_pfn = section_nr_to_pfn(mem_blk->end_section_nr);
411 sect_end_pfn += PAGES_PER_SECTION - 1;
412 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
416 * memory block could have several absent sections from start.
417 * skip pfn range from absent section
419 if (!pfn_present(pfn)) {
420 pfn = round_down(pfn + PAGES_PER_SECTION,
421 PAGES_PER_SECTION) - 1;
425 page_nid = get_nid_for_pfn(pfn);
430 ret = sysfs_create_link_nowarn(&node_devices[nid]->dev.kobj,
432 kobject_name(&mem_blk->dev.kobj));
436 return sysfs_create_link_nowarn(&mem_blk->dev.kobj,
437 &node_devices[nid]->dev.kobj,
438 kobject_name(&node_devices[nid]->dev.kobj));
440 /* mem section does not span the specified node */
444 /* unregister memory section under all nodes that it spans */
445 int unregister_mem_sect_under_nodes(struct memory_block *mem_blk,
446 unsigned long phys_index)
448 NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
449 unsigned long pfn, sect_start_pfn, sect_end_pfn;
452 NODEMASK_FREE(unlinked_nodes);
457 nodes_clear(*unlinked_nodes);
459 sect_start_pfn = section_nr_to_pfn(phys_index);
460 sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
461 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
464 nid = get_nid_for_pfn(pfn);
467 if (!node_online(nid))
469 if (node_test_and_set(nid, *unlinked_nodes))
471 sysfs_remove_link(&node_devices[nid]->dev.kobj,
472 kobject_name(&mem_blk->dev.kobj));
473 sysfs_remove_link(&mem_blk->dev.kobj,
474 kobject_name(&node_devices[nid]->dev.kobj));
476 NODEMASK_FREE(unlinked_nodes);
480 int link_mem_sections(int nid, unsigned long start_pfn, unsigned long nr_pages)
482 unsigned long end_pfn = start_pfn + nr_pages;
484 struct memory_block *mem_blk = NULL;
487 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
488 unsigned long section_nr = pfn_to_section_nr(pfn);
489 struct mem_section *mem_sect;
492 if (!present_section_nr(section_nr))
494 mem_sect = __nr_to_section(section_nr);
496 /* same memblock ? */
498 if ((section_nr >= mem_blk->start_section_nr) &&
499 (section_nr <= mem_blk->end_section_nr))
502 mem_blk = find_memory_block_hinted(mem_sect, mem_blk);
504 ret = register_mem_sect_under_node(mem_blk, nid);
508 /* discard ref obtained in find_memory_block() */
512 kobject_put(&mem_blk->dev.kobj);
516 #ifdef CONFIG_HUGETLBFS
518 * Handle per node hstate attribute [un]registration on transistions
519 * to/from memoryless state.
521 static void node_hugetlb_work(struct work_struct *work)
523 struct node *node = container_of(work, struct node, node_work);
526 * We only get here when a node transitions to/from memoryless state.
527 * We can detect which transition occurred by examining whether the
528 * node has memory now. hugetlb_register_node() already check this
529 * so we try to register the attributes. If that fails, then the
530 * node has transitioned to memoryless, try to unregister the
533 if (!hugetlb_register_node(node))
534 hugetlb_unregister_node(node);
537 static void init_node_hugetlb_work(int nid)
539 INIT_WORK(&node_devices[nid]->node_work, node_hugetlb_work);
542 static int node_memory_callback(struct notifier_block *self,
543 unsigned long action, void *arg)
545 struct memory_notify *mnb = arg;
546 int nid = mnb->status_change_nid;
552 * offload per node hstate [un]registration to a work thread
553 * when transitioning to/from memoryless state.
555 if (nid != NUMA_NO_NODE)
556 schedule_work(&node_devices[nid]->node_work);
559 case MEM_GOING_ONLINE:
560 case MEM_GOING_OFFLINE:
561 case MEM_CANCEL_ONLINE:
562 case MEM_CANCEL_OFFLINE:
569 #endif /* CONFIG_HUGETLBFS */
570 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
572 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
573 !defined(CONFIG_HUGETLBFS)
574 static inline int node_memory_callback(struct notifier_block *self,
575 unsigned long action, void *arg)
580 static void init_node_hugetlb_work(int nid) { }
584 int __register_one_node(int nid)
589 node_devices[nid] = kzalloc(sizeof(struct node), GFP_KERNEL);
590 if (!node_devices[nid])
593 error = register_node(node_devices[nid], nid);
595 /* link cpu under this node */
596 for_each_present_cpu(cpu) {
597 if (cpu_to_node(cpu) == nid)
598 register_cpu_under_node(cpu, nid);
601 /* initialize work queue for memory hot plug */
602 init_node_hugetlb_work(nid);
607 void unregister_one_node(int nid)
609 if (!node_devices[nid])
612 unregister_node(node_devices[nid]);
613 node_devices[nid] = NULL;
617 * node states attributes
620 static ssize_t print_nodes_state(enum node_states state, char *buf)
624 n = scnprintf(buf, PAGE_SIZE - 1, "%*pbl",
625 nodemask_pr_args(&node_states[state]));
632 struct device_attribute attr;
633 enum node_states state;
636 static ssize_t show_node_state(struct device *dev,
637 struct device_attribute *attr, char *buf)
639 struct node_attr *na = container_of(attr, struct node_attr, attr);
640 return print_nodes_state(na->state, buf);
643 #define _NODE_ATTR(name, state) \
644 { __ATTR(name, 0444, show_node_state, NULL), state }
646 static struct node_attr node_state_attr[] = {
647 [N_POSSIBLE] = _NODE_ATTR(possible, N_POSSIBLE),
648 [N_ONLINE] = _NODE_ATTR(online, N_ONLINE),
649 [N_NORMAL_MEMORY] = _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
650 #ifdef CONFIG_HIGHMEM
651 [N_HIGH_MEMORY] = _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
653 [N_MEMORY] = _NODE_ATTR(has_memory, N_MEMORY),
654 [N_CPU] = _NODE_ATTR(has_cpu, N_CPU),
657 static struct attribute *node_state_attrs[] = {
658 &node_state_attr[N_POSSIBLE].attr.attr,
659 &node_state_attr[N_ONLINE].attr.attr,
660 &node_state_attr[N_NORMAL_MEMORY].attr.attr,
661 #ifdef CONFIG_HIGHMEM
662 &node_state_attr[N_HIGH_MEMORY].attr.attr,
664 &node_state_attr[N_MEMORY].attr.attr,
665 &node_state_attr[N_CPU].attr.attr,
669 static struct attribute_group memory_root_attr_group = {
670 .attrs = node_state_attrs,
673 static const struct attribute_group *cpu_root_attr_groups[] = {
674 &memory_root_attr_group,
678 #define NODE_CALLBACK_PRI 2 /* lower than SLAB */
679 static int __init register_node_type(void)
683 BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
684 BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
686 ret = subsys_system_register(&node_subsys, cpu_root_attr_groups);
688 static struct notifier_block node_memory_callback_nb = {
689 .notifier_call = node_memory_callback,
690 .priority = NODE_CALLBACK_PRI,
692 register_hotmemory_notifier(&node_memory_callback_nb);
696 * Note: we're not going to unregister the node class if we fail
697 * to register the node state class attribute files.
701 postcore_initcall(register_node_type);