projects/linux·blob v4.1 · refs

687 lines · 18.0 KiB

1/*
2 * Basic Node interface support
3 */
4
5#include <linux/module.h>
6#include <linux/init.h>
7#include <linux/mm.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>
21
22static struct bus_type node_subsys = {
23 .name = "node",
24 .dev_name = "node",
25};
26
27
28static ssize_t node_read_cpumap(struct device *dev, bool list, char *buf)
29{
30 struct node *node_dev = to_node(dev);
31 const struct cpumask *mask = cpumask_of_node(node_dev->dev.id);
32
33 /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
34 BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1));
35
36 return cpumap_print_to_pagebuf(list, buf, mask);
37}
38
39static inline ssize_t node_read_cpumask(struct device *dev,
40 struct device_attribute *attr, char *buf)
41{
42 return node_read_cpumap(dev, false, buf);
43}
44static inline ssize_t node_read_cpulist(struct device *dev,
45 struct device_attribute *attr, char *buf)
46{
47 return node_read_cpumap(dev, true, buf);
48}
49
50static DEVICE_ATTR(cpumap, S_IRUGO, node_read_cpumask, NULL);
51static DEVICE_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
52
53#define K(x) ((x) << (PAGE_SHIFT - 10))
54static ssize_t node_read_meminfo(struct device *dev,
55 struct device_attribute *attr, char *buf)
56{
57 int n;
58 int nid = dev->id;
59 struct sysinfo i;
60
61 si_meminfo_node(&i, nid);
62 n = sprintf(buf,
63 "Node %d MemTotal: %8lu kB\n"
64 "Node %d MemFree: %8lu kB\n"
65 "Node %d MemUsed: %8lu kB\n"
66 "Node %d Active: %8lu kB\n"
67 "Node %d Inactive: %8lu kB\n"
68 "Node %d Active(anon): %8lu kB\n"
69 "Node %d Inactive(anon): %8lu kB\n"
70 "Node %d Active(file): %8lu kB\n"
71 "Node %d Inactive(file): %8lu kB\n"
72 "Node %d Unevictable: %8lu kB\n"
73 "Node %d Mlocked: %8lu kB\n",
74 nid, K(i.totalram),
75 nid, K(i.freeram),
76 nid, K(i.totalram - i.freeram),
77 nid, K(node_page_state(nid, NR_ACTIVE_ANON) +
78 node_page_state(nid, NR_ACTIVE_FILE)),
79 nid, K(node_page_state(nid, NR_INACTIVE_ANON) +
80 node_page_state(nid, NR_INACTIVE_FILE)),
81 nid, K(node_page_state(nid, NR_ACTIVE_ANON)),
82 nid, K(node_page_state(nid, NR_INACTIVE_ANON)),
83 nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
84 nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
85 nid, K(node_page_state(nid, NR_UNEVICTABLE)),
86 nid, K(node_page_state(nid, NR_MLOCK)));
87
88#ifdef CONFIG_HIGHMEM
89 n += sprintf(buf + n,
90 "Node %d HighTotal: %8lu kB\n"
91 "Node %d HighFree: %8lu kB\n"
92 "Node %d LowTotal: %8lu kB\n"
93 "Node %d LowFree: %8lu kB\n",
94 nid, K(i.totalhigh),
95 nid, K(i.freehigh),
96 nid, K(i.totalram - i.totalhigh),
97 nid, K(i.freeram - i.freehigh));
98#endif
99 n += sprintf(buf + n,
100 "Node %d Dirty: %8lu kB\n"
101 "Node %d Writeback: %8lu kB\n"
102 "Node %d FilePages: %8lu kB\n"
103 "Node %d Mapped: %8lu kB\n"
104 "Node %d AnonPages: %8lu kB\n"
105 "Node %d Shmem: %8lu kB\n"
106 "Node %d KernelStack: %8lu kB\n"
107 "Node %d PageTables: %8lu kB\n"
108 "Node %d NFS_Unstable: %8lu kB\n"
109 "Node %d Bounce: %8lu kB\n"
110 "Node %d WritebackTmp: %8lu kB\n"
111 "Node %d Slab: %8lu kB\n"
112 "Node %d SReclaimable: %8lu kB\n"
113 "Node %d SUnreclaim: %8lu kB\n"
114#ifdef CONFIG_TRANSPARENT_HUGEPAGE
115 "Node %d AnonHugePages: %8lu kB\n"
116#endif
117 ,
118 nid, K(node_page_state(nid, NR_FILE_DIRTY)),
119 nid, K(node_page_state(nid, NR_WRITEBACK)),
120 nid, K(node_page_state(nid, NR_FILE_PAGES)),
121 nid, K(node_page_state(nid, NR_FILE_MAPPED)),
122 nid, K(node_page_state(nid, NR_ANON_PAGES)),
123 nid, K(i.sharedram),
124 nid, node_page_state(nid, NR_KERNEL_STACK) *
125 THREAD_SIZE / 1024,
126 nid, K(node_page_state(nid, NR_PAGETABLE)),
127 nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
128 nid, K(node_page_state(nid, NR_BOUNCE)),
129 nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
130 nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
131 node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
132 nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
133#ifdef CONFIG_TRANSPARENT_HUGEPAGE
134 nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE))
135 , nid,
136 K(node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) *
137 HPAGE_PMD_NR));
138#else
139 nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
140#endif
141 n += hugetlb_report_node_meminfo(nid, buf + n);
142 return n;
143}
144
145#undef K
146static DEVICE_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
147
148static ssize_t node_read_numastat(struct device *dev,
149 struct device_attribute *attr, char *buf)
150{
151 return sprintf(buf,
152 "numa_hit %lu\n"
153 "numa_miss %lu\n"
154 "numa_foreign %lu\n"
155 "interleave_hit %lu\n"
156 "local_node %lu\n"
157 "other_node %lu\n",
158 node_page_state(dev->id, NUMA_HIT),
159 node_page_state(dev->id, NUMA_MISS),
160 node_page_state(dev->id, NUMA_FOREIGN),
161 node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
162 node_page_state(dev->id, NUMA_LOCAL),
163 node_page_state(dev->id, NUMA_OTHER));
164}
165static DEVICE_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
166
167static ssize_t node_read_vmstat(struct device *dev,
168 struct device_attribute *attr, char *buf)
169{
170 int nid = dev->id;
171 int i;
172 int n = 0;
173
174 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
175 n += sprintf(buf+n, "%s %lu\n", vmstat_text[i],
176 node_page_state(nid, i));
177
178 return n;
179}
180static DEVICE_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
181
182static ssize_t node_read_distance(struct device *dev,
183 struct device_attribute *attr, char *buf)
184{
185 int nid = dev->id;
186 int len = 0;
187 int i;
188
189 /*
190 * buf is currently PAGE_SIZE in length and each node needs 4 chars
191 * at the most (distance + space or newline).
192 */
193 BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
194
195 for_each_online_node(i)
196 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
197
198 len += sprintf(buf + len, "\n");
199 return len;
200}
201static DEVICE_ATTR(distance, S_IRUGO, node_read_distance, NULL);
202
203static struct attribute *node_dev_attrs[] = {
204 &dev_attr_cpumap.attr,
205 &dev_attr_cpulist.attr,
206 &dev_attr_meminfo.attr,
207 &dev_attr_numastat.attr,
208 &dev_attr_distance.attr,
209 &dev_attr_vmstat.attr,
210 NULL
211};
212ATTRIBUTE_GROUPS(node_dev);
213
214#ifdef CONFIG_HUGETLBFS
215/*
216 * hugetlbfs per node attributes registration interface:
217 * When/if hugetlb[fs] subsystem initializes [sometime after this module],
218 * it will register its per node attributes for all online nodes with
219 * memory. It will also call register_hugetlbfs_with_node(), below, to
220 * register its attribute registration functions with this node driver.
221 * Once these hooks have been initialized, the node driver will call into
222 * the hugetlb module to [un]register attributes for hot-plugged nodes.
223 */
224static node_registration_func_t __hugetlb_register_node;
225static node_registration_func_t __hugetlb_unregister_node;
226
227static inline bool hugetlb_register_node(struct node *node)
228{
229 if (__hugetlb_register_node &&
230 node_state(node->dev.id, N_MEMORY)) {
231 __hugetlb_register_node(node);
232 return true;
233 }
234 return false;
235}
236
237static inline void hugetlb_unregister_node(struct node *node)
238{
239 if (__hugetlb_unregister_node)
240 __hugetlb_unregister_node(node);
241}
242
243void register_hugetlbfs_with_node(node_registration_func_t doregister,
244 node_registration_func_t unregister)
245{
246 __hugetlb_register_node = doregister;
247 __hugetlb_unregister_node = unregister;
248}
249#else
250static inline void hugetlb_register_node(struct node *node) {}
251
252static inline void hugetlb_unregister_node(struct node *node) {}
253#endif
254
255static void node_device_release(struct device *dev)
256{
257 struct node *node = to_node(dev);
258
259#if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HUGETLBFS)
260 /*
261 * We schedule the work only when a memory section is
262 * onlined/offlined on this node. When we come here,
263 * all the memory on this node has been offlined,
264 * so we won't enqueue new work to this work.
265 *
266 * The work is using node->node_work, so we should
267 * flush work before freeing the memory.
268 */
269 flush_work(&node->node_work);
270#endif
271 kfree(node);
272}
273
274/*
275 * register_node - Setup a sysfs device for a node.
276 * @num - Node number to use when creating the device.
277 *
278 * Initialize and register the node device.
279 */
280static int register_node(struct node *node, int num, struct node *parent)
281{
282 int error;
283
284 node->dev.id = num;
285 node->dev.bus = &node_subsys;
286 node->dev.release = node_device_release;
287 node->dev.groups = node_dev_groups;
288 error = device_register(&node->dev);
289
290 if (!error){
291 hugetlb_register_node(node);
292
293 compaction_register_node(node);
294 }
295 return error;
296}
297
298/**
299 * unregister_node - unregister a node device
300 * @node: node going away
301 *
302 * Unregisters a node device @node. All the devices on the node must be
303 * unregistered before calling this function.
304 */
305void unregister_node(struct node *node)
306{
307 hugetlb_unregister_node(node); /* no-op, if memoryless node */
308
309 device_unregister(&node->dev);
310}
311
312struct node *node_devices[MAX_NUMNODES];
313
314/*
315 * register cpu under node
316 */
317int register_cpu_under_node(unsigned int cpu, unsigned int nid)
318{
319 int ret;
320 struct device *obj;
321
322 if (!node_online(nid))
323 return 0;
324
325 obj = get_cpu_device(cpu);
326 if (!obj)
327 return 0;
328
329 ret = sysfs_create_link(&node_devices[nid]->dev.kobj,
330 &obj->kobj,
331 kobject_name(&obj->kobj));
332 if (ret)
333 return ret;
334
335 return sysfs_create_link(&obj->kobj,
336 &node_devices[nid]->dev.kobj,
337 kobject_name(&node_devices[nid]->dev.kobj));
338}
339
340int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
341{
342 struct device *obj;
343
344 if (!node_online(nid))
345 return 0;
346
347 obj = get_cpu_device(cpu);
348 if (!obj)
349 return 0;
350
351 sysfs_remove_link(&node_devices[nid]->dev.kobj,
352 kobject_name(&obj->kobj));
353 sysfs_remove_link(&obj->kobj,
354 kobject_name(&node_devices[nid]->dev.kobj));
355
356 return 0;
357}
358
359#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
360#define page_initialized(page) (page->lru.next)
361
362static int get_nid_for_pfn(unsigned long pfn)
363{
364 struct page *page;
365
366 if (!pfn_valid_within(pfn))
367 return -1;
368 page = pfn_to_page(pfn);
369 if (!page_initialized(page))
370 return -1;
371 return pfn_to_nid(pfn);
372}
373
374/* register memory section under specified node if it spans that node */
375int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
376{
377 int ret;
378 unsigned long pfn, sect_start_pfn, sect_end_pfn;
379
380 if (!mem_blk)
381 return -EFAULT;
382 if (!node_online(nid))
383 return 0;
384
385 sect_start_pfn = section_nr_to_pfn(mem_blk->start_section_nr);
386 sect_end_pfn = section_nr_to_pfn(mem_blk->end_section_nr);
387 sect_end_pfn += PAGES_PER_SECTION - 1;
388 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
389 int page_nid;
390
391 page_nid = get_nid_for_pfn(pfn);
392 if (page_nid < 0)
393 continue;
394 if (page_nid != nid)
395 continue;
396 ret = sysfs_create_link_nowarn(&node_devices[nid]->dev.kobj,
397 &mem_blk->dev.kobj,
398 kobject_name(&mem_blk->dev.kobj));
399 if (ret)
400 return ret;
401
402 return sysfs_create_link_nowarn(&mem_blk->dev.kobj,
403 &node_devices[nid]->dev.kobj,
404 kobject_name(&node_devices[nid]->dev.kobj));
405 }
406 /* mem section does not span the specified node */
407 return 0;
408}
409
410/* unregister memory section under all nodes that it spans */
411int unregister_mem_sect_under_nodes(struct memory_block *mem_blk,
412 unsigned long phys_index)
413{
414 NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
415 unsigned long pfn, sect_start_pfn, sect_end_pfn;
416
417 if (!mem_blk) {
418 NODEMASK_FREE(unlinked_nodes);
419 return -EFAULT;
420 }
421 if (!unlinked_nodes)
422 return -ENOMEM;
423 nodes_clear(*unlinked_nodes);
424
425 sect_start_pfn = section_nr_to_pfn(phys_index);
426 sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
427 for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
428 int nid;
429
430 nid = get_nid_for_pfn(pfn);
431 if (nid < 0)
432 continue;
433 if (!node_online(nid))
434 continue;
435 if (node_test_and_set(nid, *unlinked_nodes))
436 continue;
437 sysfs_remove_link(&node_devices[nid]->dev.kobj,
438 kobject_name(&mem_blk->dev.kobj));
439 sysfs_remove_link(&mem_blk->dev.kobj,
440 kobject_name(&node_devices[nid]->dev.kobj));
441 }
442 NODEMASK_FREE(unlinked_nodes);
443 return 0;
444}
445
446static int link_mem_sections(int nid)
447{
448 unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
449 unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
450 unsigned long pfn;
451 struct memory_block *mem_blk = NULL;
452 int err = 0;
453
454 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
455 unsigned long section_nr = pfn_to_section_nr(pfn);
456 struct mem_section *mem_sect;
457 int ret;
458
459 if (!present_section_nr(section_nr))
460 continue;
461 mem_sect = __nr_to_section(section_nr);
462
463 /* same memblock ? */
464 if (mem_blk)
465 if ((section_nr >= mem_blk->start_section_nr) &&
466 (section_nr <= mem_blk->end_section_nr))
467 continue;
468
469 mem_blk = find_memory_block_hinted(mem_sect, mem_blk);
470
471 ret = register_mem_sect_under_node(mem_blk, nid);
472 if (!err)
473 err = ret;
474
475 /* discard ref obtained in find_memory_block() */
476 }
477
478 if (mem_blk)
479 kobject_put(&mem_blk->dev.kobj);
480 return err;
481}
482
483#ifdef CONFIG_HUGETLBFS
484/*
485 * Handle per node hstate attribute [un]registration on transistions
486 * to/from memoryless state.
487 */
488static void node_hugetlb_work(struct work_struct *work)
489{
490 struct node *node = container_of(work, struct node, node_work);
491
492 /*
493 * We only get here when a node transitions to/from memoryless state.
494 * We can detect which transition occurred by examining whether the
495 * node has memory now. hugetlb_register_node() already check this
496 * so we try to register the attributes. If that fails, then the
497 * node has transitioned to memoryless, try to unregister the
498 * attributes.
499 */
500 if (!hugetlb_register_node(node))
501 hugetlb_unregister_node(node);
502}
503
504static void init_node_hugetlb_work(int nid)
505{
506 INIT_WORK(&node_devices[nid]->node_work, node_hugetlb_work);
507}
508
509static int node_memory_callback(struct notifier_block *self,
510 unsigned long action, void *arg)
511{
512 struct memory_notify *mnb = arg;
513 int nid = mnb->status_change_nid;
514
515 switch (action) {
516 case MEM_ONLINE:
517 case MEM_OFFLINE:
518 /*
519 * offload per node hstate [un]registration to a work thread
520 * when transitioning to/from memoryless state.
521 */
522 if (nid != NUMA_NO_NODE)
523 schedule_work(&node_devices[nid]->node_work);
524 break;
525
526 case MEM_GOING_ONLINE:
527 case MEM_GOING_OFFLINE:
528 case MEM_CANCEL_ONLINE:
529 case MEM_CANCEL_OFFLINE:
530 default:
531 break;
532 }
533
534 return NOTIFY_OK;
535}
536#endif /* CONFIG_HUGETLBFS */
537#else /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
538
539static int link_mem_sections(int nid) { return 0; }
540#endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
541
542#if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
543 !defined(CONFIG_HUGETLBFS)
544static inline int node_memory_callback(struct notifier_block *self,
545 unsigned long action, void *arg)
546{
547 return NOTIFY_OK;
548}
549
550static void init_node_hugetlb_work(int nid) { }
551
552#endif
553
554int register_one_node(int nid)
555{
556 int error = 0;
557 int cpu;
558
559 if (node_online(nid)) {
560 int p_node = parent_node(nid);
561 struct node *parent = NULL;
562
563 if (p_node != nid)
564 parent = node_devices[p_node];
565
566 node_devices[nid] = kzalloc(sizeof(struct node), GFP_KERNEL);
567 if (!node_devices[nid])
568 return -ENOMEM;
569
570 error = register_node(node_devices[nid], nid, parent);
571
572 /* link cpu under this node */
573 for_each_present_cpu(cpu) {
574 if (cpu_to_node(cpu) == nid)
575 register_cpu_under_node(cpu, nid);
576 }
577
578 /* link memory sections under this node */
579 error = link_mem_sections(nid);
580
581 /* initialize work queue for memory hot plug */
582 init_node_hugetlb_work(nid);
583 }
584
585 return error;
586
587}
588
589void unregister_one_node(int nid)
590{
591 if (!node_devices[nid])
592 return;
593
594 unregister_node(node_devices[nid]);
595 node_devices[nid] = NULL;
596}
597
598/*
599 * node states attributes
600 */
601
602static ssize_t print_nodes_state(enum node_states state, char *buf)
603{
604 int n;
605
606 n = scnprintf(buf, PAGE_SIZE - 1, "%*pbl",
607 nodemask_pr_args(&node_states[state]));
608 buf[n++] = '\n';
609 buf[n] = '\0';
610 return n;
611}
612
613struct node_attr {
614 struct device_attribute attr;
615 enum node_states state;
616};
617
618static ssize_t show_node_state(struct device *dev,
619 struct device_attribute *attr, char *buf)
620{
621 struct node_attr *na = container_of(attr, struct node_attr, attr);
622 return print_nodes_state(na->state, buf);
623}
624
625#define _NODE_ATTR(name, state) \
626 { __ATTR(name, 0444, show_node_state, NULL), state }
627
628static struct node_attr node_state_attr[] = {
629 [N_POSSIBLE] = _NODE_ATTR(possible, N_POSSIBLE),
630 [N_ONLINE] = _NODE_ATTR(online, N_ONLINE),
631 [N_NORMAL_MEMORY] = _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
632#ifdef CONFIG_HIGHMEM
633 [N_HIGH_MEMORY] = _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
634#endif
635#ifdef CONFIG_MOVABLE_NODE
636 [N_MEMORY] = _NODE_ATTR(has_memory, N_MEMORY),
637#endif
638 [N_CPU] = _NODE_ATTR(has_cpu, N_CPU),
639};
640
641static struct attribute *node_state_attrs[] = {
642 &node_state_attr[N_POSSIBLE].attr.attr,
643 &node_state_attr[N_ONLINE].attr.attr,
644 &node_state_attr[N_NORMAL_MEMORY].attr.attr,
645#ifdef CONFIG_HIGHMEM
646 &node_state_attr[N_HIGH_MEMORY].attr.attr,
647#endif
648#ifdef CONFIG_MOVABLE_NODE
649 &node_state_attr[N_MEMORY].attr.attr,
650#endif
651 &node_state_attr[N_CPU].attr.attr,
652 NULL
653};
654
655static struct attribute_group memory_root_attr_group = {
656 .attrs = node_state_attrs,
657};
658
659static const struct attribute_group *cpu_root_attr_groups[] = {
660 &memory_root_attr_group,
661 NULL,
662};
663
664#define NODE_CALLBACK_PRI 2 /* lower than SLAB */
665static int __init register_node_type(void)
666{
667 int ret;
668
669 BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
670 BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
671
672 ret = subsys_system_register(&node_subsys, cpu_root_attr_groups);
673 if (!ret) {
674 static struct notifier_block node_memory_callback_nb = {
675 .notifier_call = node_memory_callback,
676 .priority = NODE_CALLBACK_PRI,
677 };
678 register_hotmemory_notifier(&node_memory_callback_nb);
679 }
680
681 /*
682 * Note: we're not going to unregister the node class if we fail
683 * to register the node state class attribute files.
684 */
685 return ret;
686}
687postcore_initcall(register_node_type);