Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6
[pandora-kernel.git] / drivers / base / node.c
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>
9 #include <linux/memory.h>
10 #include <linux/vmstat.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
22 static struct sysdev_class_attribute *node_state_attrs[];
23
24 static struct sysdev_class node_class = {
25         .name = "node",
26         .attrs = node_state_attrs,
27 };
28
29
30 static ssize_t node_read_cpumap(struct sys_device *dev, int type, char *buf)
31 {
32         struct node *node_dev = to_node(dev);
33         const struct cpumask *mask = cpumask_of_node(node_dev->sysdev.id);
34         int len;
35
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));
38
39         len = type?
40                 cpulist_scnprintf(buf, PAGE_SIZE-2, mask) :
41                 cpumask_scnprintf(buf, PAGE_SIZE-2, mask);
42         buf[len++] = '\n';
43         buf[len] = '\0';
44         return len;
45 }
46
47 static inline ssize_t node_read_cpumask(struct sys_device *dev,
48                                 struct sysdev_attribute *attr, char *buf)
49 {
50         return node_read_cpumap(dev, 0, buf);
51 }
52 static inline ssize_t node_read_cpulist(struct sys_device *dev,
53                                 struct sysdev_attribute *attr, char *buf)
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);
60
61 #define K(x) ((x) << (PAGE_SHIFT - 10))
62 static ssize_t node_read_meminfo(struct sys_device * dev,
63                         struct sysdev_attribute *attr, char * buf)
64 {
65         int n;
66         int nid = dev->id;
67         struct sysinfo i;
68
69         si_meminfo_node(&i, nid);
70         n = sprintf(buf,
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"
80                        "Node %d Unevictable:    %8lu kB\n"
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
96 #ifdef CONFIG_HIGHMEM
97         n += sprintf(buf + n,
98                        "Node %d HighTotal:      %8lu kB\n"
99                        "Node %d HighFree:       %8lu kB\n"
100                        "Node %d LowTotal:       %8lu kB\n"
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));
106 #endif
107         n += sprintf(buf + n,
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"
113                        "Node %d Shmem:          %8lu kB\n"
114                        "Node %d KernelStack:    %8lu kB\n"
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"
121                        "Node %d SUnreclaim:     %8lu kB\n"
122 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
123                        "Node %d AnonHugePages:  %8lu kB\n"
124 #endif
125                         ,
126                        nid, K(node_page_state(nid, NR_FILE_DIRTY)),
127                        nid, K(node_page_state(nid, NR_WRITEBACK)),
128                        nid, K(node_page_state(nid, NR_FILE_PAGES)),
129                        nid, K(node_page_state(nid, NR_FILE_MAPPED)),
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                        ),
136                        nid, K(node_page_state(nid, NR_SHMEM)),
137                        nid, node_page_state(nid, NR_KERNEL_STACK) *
138                                 THREAD_SIZE / 1024,
139                        nid, K(node_page_state(nid, NR_PAGETABLE)),
140                        nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
141                        nid, K(node_page_state(nid, NR_BOUNCE)),
142                        nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
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)),
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                        );
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
160 static ssize_t node_read_numastat(struct sys_device * dev,
161                                 struct sysdev_attribute *attr, char * buf)
162 {
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",
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));
176 }
177 static SYSDEV_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
178
179 static ssize_t node_read_vmstat(struct sys_device *dev,
180                                 struct sysdev_attribute *attr, char *buf)
181 {
182         int nid = dev->id;
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;
191 }
192 static SYSDEV_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
193
194 static ssize_t node_read_distance(struct sys_device * dev,
195                         struct sysdev_attribute *attr, char * buf)
196 {
197         int nid = dev->id;
198         int len = 0;
199         int i;
200
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);
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
215 #ifdef CONFIG_HUGETLBFS
216 /*
217  * hugetlbfs per node attributes registration interface:
218  * When/if hugetlb[fs] subsystem initializes [sometime after this module],
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
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
228 static inline bool hugetlb_register_node(struct node *node)
229 {
230         if (__hugetlb_register_node &&
231                         node_state(node->sysdev.id, N_HIGH_MEMORY)) {
232                 __hugetlb_register_node(node);
233                 return true;
234         }
235         return false;
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
256
257 /*
258  * register_node - Setup a sysfs device for a node.
259  * @num - Node number to use when creating the device.
260  *
261  * Initialize and register the node device.
262  */
263 int register_node(struct node *node, int num, struct node *parent)
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);
273                 sysdev_create_file(&node->sysdev, &attr_cpulist);
274                 sysdev_create_file(&node->sysdev, &attr_meminfo);
275                 sysdev_create_file(&node->sysdev, &attr_numastat);
276                 sysdev_create_file(&node->sysdev, &attr_distance);
277                 sysdev_create_file(&node->sysdev, &attr_vmstat);
278
279                 scan_unevictable_register_node(node);
280
281                 hugetlb_register_node(node);
282
283                 compaction_register_node(node);
284         }
285         return error;
286 }
287
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);
298         sysdev_remove_file(&node->sysdev, &attr_cpulist);
299         sysdev_remove_file(&node->sysdev, &attr_meminfo);
300         sysdev_remove_file(&node->sysdev, &attr_numastat);
301         sysdev_remove_file(&node->sysdev, &attr_distance);
302         sysdev_remove_file(&node->sysdev, &attr_vmstat);
303
304         scan_unevictable_unregister_node(node);
305         hugetlb_unregister_node(node);          /* no-op, if memoryless node */
306
307         sysdev_unregister(&node->sysdev);
308 }
309
310 struct node node_devices[MAX_NUMNODES];
311
312 /*
313  * register cpu under node
314  */
315 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
316 {
317         int ret;
318         struct sys_device *obj;
319
320         if (!node_online(nid))
321                 return 0;
322
323         obj = get_cpu_sysdev(cpu);
324         if (!obj)
325                 return 0;
326
327         ret = sysfs_create_link(&node_devices[nid].sysdev.kobj,
328                                 &obj->kobj,
329                                 kobject_name(&obj->kobj));
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));
336 }
337
338 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
339 {
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));
351         sysfs_remove_link(&obj->kobj,
352                           kobject_name(&node_devices[nid].sysdev.kobj));
353
354         return 0;
355 }
356
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 {
375         int ret;
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;
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;
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;
394                 ret = sysfs_create_link_nowarn(&node_devices[nid].sysdev.kobj,
395                                         &mem_blk->sysdev.kobj,
396                                         kobject_name(&mem_blk->sysdev.kobj));
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));
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 */
409 int unregister_mem_sect_under_nodes(struct memory_block *mem_blk,
410                                     unsigned long phys_index)
411 {
412         NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
413         unsigned long pfn, sect_start_pfn, sect_end_pfn;
414
415         if (!mem_blk) {
416                 NODEMASK_FREE(unlinked_nodes);
417                 return -EFAULT;
418         }
419         if (!unlinked_nodes)
420                 return -ENOMEM;
421         nodes_clear(*unlinked_nodes);
422
423         sect_start_pfn = section_nr_to_pfn(phys_index);
424         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
425         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
426                 int nid;
427
428                 nid = get_nid_for_pfn(pfn);
429                 if (nid < 0)
430                         continue;
431                 if (!node_online(nid))
432                         continue;
433                 if (node_test_and_set(nid, *unlinked_nodes))
434                         continue;
435                 sysfs_remove_link(&node_devices[nid].sysdev.kobj,
436                          kobject_name(&mem_blk->sysdev.kobj));
437                 sysfs_remove_link(&mem_blk->sysdev.kobj,
438                          kobject_name(&node_devices[nid].sysdev.kobj));
439         }
440         NODEMASK_FREE(unlinked_nodes);
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;
449         struct memory_block *mem_blk = NULL;
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;
455                 int ret;
456
457                 if (!present_section_nr(section_nr))
458                         continue;
459                 mem_sect = __nr_to_section(section_nr);
460                 mem_blk = find_memory_block_hinted(mem_sect, mem_blk);
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() */
466         }
467
468         if (mem_blk)
469                 kobject_put(&mem_blk->sysdev.kobj);
470         return err;
471 }
472
473 #ifdef CONFIG_HUGETLBFS
474 /*
475  * Handle per node hstate attribute [un]registration on transistions
476  * to/from memoryless state.
477  */
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 }
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) {
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                  */
512                 if (nid != NUMA_NO_NODE)
513                         schedule_work(&node_devices[nid].node_work);
514                 break;
515
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 }
526 #endif  /* CONFIG_HUGETLBFS */
527 #else   /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
528
529 static int link_mem_sections(int nid) { return 0; }
530 #endif  /* CONFIG_MEMORY_HOTPLUG_SPARSE */
531
532 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
533     !defined(CONFIG_HUGETLBFS)
534 static inline int node_memory_callback(struct notifier_block *self,
535                                 unsigned long action, void *arg)
536 {
537         return NOTIFY_OK;
538 }
539
540 static void init_node_hugetlb_work(int nid) { }
541
542 #endif
543
544 int register_one_node(int nid)
545 {
546         int error = 0;
547         int cpu;
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);
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                 }
563
564                 /* link memory sections under this node */
565                 error = link_mem_sections(nid);
566
567                 /* initialize work queue for memory hot plug */
568                 init_node_hugetlb_work(nid);
569         }
570
571         return error;
572
573 }
574
575 void unregister_one_node(int nid)
576 {
577         unregister_node(&node_devices[nid]);
578 }
579
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
596 struct node_attr {
597         struct sysdev_class_attribute attr;
598         enum node_states state;
599 };
600
601 static ssize_t show_node_state(struct sysdev_class *class,
602                                struct sysdev_class_attribute *attr, char *buf)
603 {
604         struct node_attr *na = container_of(attr, struct node_attr, attr);
605         return print_nodes_state(na->state, buf);
606 }
607
608 #define _NODE_ATTR(name, state) \
609         { _SYSDEV_CLASS_ATTR(name, 0444, show_node_state, NULL), state }
610
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),
616 #ifdef CONFIG_HIGHMEM
617         _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
618 #endif
619 };
620
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 };
631
632 #define NODE_CALLBACK_PRI       2       /* lower than SLAB */
633 static int __init register_node_type(void)
634 {
635         int ret;
636
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
640         ret = sysdev_class_register(&node_class);
641         if (!ret) {
642                 hotplug_memory_notifier(node_memory_callback,
643                                         NODE_CALLBACK_PRI);
644         }
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;
651 }
652 postcore_initcall(register_node_type);