drivers/base/node.c

   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
  22 static struct bus_type node_subsys = {
  23         .name = "node",
  24         .dev_name = "node",
  25 };
  26
  27
  28 static ssize_t node_read_cpumap(struct device *dev, int type, char *buf)
  29 {
  30         struct node *node_dev = to_node(dev);
  31         const struct cpumask *mask = cpumask_of_node(node_dev->dev.id);
  32         int len;
  33
  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));
  36
  37         len = type?
  38                 cpulist_scnprintf(buf, PAGE_SIZE-2, mask) :
  39                 cpumask_scnprintf(buf, PAGE_SIZE-2, mask);
  40         buf[len++] = '\n';
  41         buf[len] = '\0';
  42         return len;
  43 }
  44
  45 static inline ssize_t node_read_cpumask(struct device *dev,
  46                                 struct device_attribute *attr, char *buf)
  47 {
  48         return node_read_cpumap(dev, 0, buf);
  49 }
  50 static inline ssize_t node_read_cpulist(struct device *dev,
  51                                 struct device_attribute *attr, char *buf)
  52 {
  53         return node_read_cpumap(dev, 1, buf);
  54 }
  55
  56 static DEVICE_ATTR(cpumap,  S_IRUGO, node_read_cpumask, NULL);
  57 static DEVICE_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
  58
  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)
  62 {
  63         int n;
  64         int nid = dev->id;
  65         struct sysinfo i;
  66
  67         si_meminfo_node(&i, nid);
  68         n = sprintf(buf,
  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",
  80                        nid, K(i.totalram),
  81                        nid, K(i.freeram),
  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)));
  93
  94 #ifdef CONFIG_HIGHMEM
  95         n += sprintf(buf + n,
  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",
 100                        nid, K(i.totalhigh),
 101                        nid, K(i.freehigh),
 102                        nid, K(i.totalram - i.totalhigh),
 103                        nid, K(i.freeram - i.freehigh));
 104 #endif
 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"
 122 #endif
 123                         ,
 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) *
 131                                 THREAD_SIZE / 1024,
 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))
 141                         , nid,
 142                         K(node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) *
 143                         HPAGE_PMD_NR));
 144 #else
 145                        nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
 146 #endif
 147         n += hugetlb_report_node_meminfo(nid, buf + n);
 148         return n;
 149 }
 150
 151 #undef K
 152 static DEVICE_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
 153
 154 static ssize_t node_read_numastat(struct device *dev,
 155                                 struct device_attribute *attr, char *buf)
 156 {
 157         return sprintf(buf,
 158                        "numa_hit %lu\n"
 159                        "numa_miss %lu\n"
 160                        "numa_foreign %lu\n"
 161                        "interleave_hit %lu\n"
 162                        "local_node %lu\n"
 163                        "other_node %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));
 170 }
 171 static DEVICE_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
 172
 173 static ssize_t node_read_vmstat(struct device *dev,
 174                                 struct device_attribute *attr, char *buf)
 175 {
 176         int nid = dev->id;
 177         int i;
 178         int n = 0;
 179
 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));
 183
 184         return n;
 185 }
 186 static DEVICE_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
 187
 188 static ssize_t node_read_distance(struct device *dev,
 189                         struct device_attribute *attr, char * buf)
 190 {
 191         int nid = dev->id;
 192         int len = 0;
 193         int i;
 194
 195         /*
 196          * buf is currently PAGE_SIZE in length and each node needs 4 chars
 197          * at the most (distance + space or newline).
 198          */
 199         BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
 200
 201         for_each_online_node(i)
 202                 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
 203
 204         len += sprintf(buf + len, "\n");
 205         return len;
 206 }
 207 static DEVICE_ATTR(distance, S_IRUGO, node_read_distance, NULL);
 208
 209 #ifdef CONFIG_HUGETLBFS
 210 /*
 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.
 218  */
 219 static node_registration_func_t __hugetlb_register_node;
 220 static node_registration_func_t __hugetlb_unregister_node;
 221
 222 static inline bool hugetlb_register_node(struct node *node)
 223 {
 224         if (__hugetlb_register_node &&
 225                         node_state(node->dev.id, N_MEMORY)) {
 226                 __hugetlb_register_node(node);
 227                 return true;
 228         }
 229         return false;
 230 }
 231
 232 static inline void hugetlb_unregister_node(struct node *node)
 233 {
 234         if (__hugetlb_unregister_node)
 235                 __hugetlb_unregister_node(node);
 236 }
 237
 238 void register_hugetlbfs_with_node(node_registration_func_t doregister,
 239                                   node_registration_func_t unregister)
 240 {
 241         __hugetlb_register_node   = doregister;
 242         __hugetlb_unregister_node = unregister;
 243 }
 244 #else
 245 static inline void hugetlb_register_node(struct node *node) {}
 246
 247 static inline void hugetlb_unregister_node(struct node *node) {}
 248 #endif
 249
 250 static void node_device_release(struct device *dev)
 251 {
 252         struct node *node = to_node(dev);
 253
 254 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HUGETLBFS)
 255         /*
 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.
 260          *
 261          * The work is using node->node_work, so we should
 262          * flush work before freeing the memory.
 263          */
 264         flush_work(&node->node_work);
 265 #endif
 266         kfree(node);
 267 }
 268
 269 /*
 270  * register_node - Setup a sysfs device for a node.
 271  * @num - Node number to use when creating the device.
 272  *
 273  * Initialize and register the node device.
 274  */
 275 static int register_node(struct node *node, int num, struct node *parent)
 276 {
 277         int error;
 278
 279         node->dev.id = num;
 280         node->dev.bus = &node_subsys;
 281         node->dev.release = node_device_release;
 282         error = device_register(&node->dev);
 283
 284         if (!error){
 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);
 291
 292                 scan_unevictable_register_node(node);
 293
 294                 hugetlb_register_node(node);
 295
 296                 compaction_register_node(node);
 297         }
 298         return error;
 299 }
 300
 301 /**
 302  * unregister_node - unregister a node device
 303  * @node: node going away
 304  *
 305  * Unregisters a node device @node.  All the devices on the node must be
 306  * unregistered before calling this function.
 307  */
 308 void unregister_node(struct node *node)
 309 {
 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);
 316
 317         scan_unevictable_unregister_node(node);
 318         hugetlb_unregister_node(node);          /* no-op, if memoryless node */
 319
 320         device_unregister(&node->dev);
 321 }
 322
 323 struct node *node_devices[MAX_NUMNODES];
 324
 325 /*
 326  * register cpu under node
 327  */
 328 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
 329 {
 330         int ret;
 331         struct device *obj;
 332
 333         if (!node_online(nid))
 334                 return 0;
 335
 336         obj = get_cpu_device(cpu);
 337         if (!obj)
 338                 return 0;
 339
 340         ret = sysfs_create_link(&node_devices[nid]->dev.kobj,
 341                                 &obj->kobj,
 342                                 kobject_name(&obj->kobj));
 343         if (ret)
 344                 return ret;
 345
 346         return sysfs_create_link(&obj->kobj,
 347                                  &node_devices[nid]->dev.kobj,
 348                                  kobject_name(&node_devices[nid]->dev.kobj));
 349 }
 350
 351 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
 352 {
 353         struct device *obj;
 354
 355         if (!node_online(nid))
 356                 return 0;
 357
 358         obj = get_cpu_device(cpu);
 359         if (!obj)
 360                 return 0;
 361
 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));
 366
 367         return 0;
 368 }
 369
 370 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
 371 #define page_initialized(page)  (page->lru.next)
 372
 373 static int get_nid_for_pfn(unsigned long pfn)
 374 {
 375         struct page *page;
 376
 377         if (!pfn_valid_within(pfn))
 378                 return -1;
 379         page = pfn_to_page(pfn);
 380         if (!page_initialized(page))
 381                 return -1;
 382         return pfn_to_nid(pfn);
 383 }
 384
 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)
 387 {
 388         int ret;
 389         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 390
 391         if (!mem_blk)
 392                 return -EFAULT;
 393         if (!node_online(nid))
 394                 return 0;
 395
 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++) {
 400                 int page_nid;
 401
 402                 /*
 403                  * memory block could have several absent sections from start.
 404                  * skip pfn range from absent section
 405                  */
 406                 if (!pfn_present(pfn)) {
 407                         pfn = round_down(pfn + PAGES_PER_SECTION,
 408                                          PAGES_PER_SECTION) - 1;
 409                         continue;
 410                 }
 411
 412                 page_nid = get_nid_for_pfn(pfn);
 413                 if (page_nid < 0)
 414                         continue;
 415                 if (page_nid != nid)
 416                         continue;
 417                 ret = sysfs_create_link_nowarn(&node_devices[nid]->dev.kobj,
 418                                         &mem_blk->dev.kobj,
 419                                         kobject_name(&mem_blk->dev.kobj));
 420                 if (ret)
 421                         return ret;
 422
 423                 return sysfs_create_link_nowarn(&mem_blk->dev.kobj,
 424                                 &node_devices[nid]->dev.kobj,
 425                                 kobject_name(&node_devices[nid]->dev.kobj));
 426         }
 427         /* mem section does not span the specified node */
 428         return 0;
 429 }
 430
 431 /* unregister memory section under all nodes that it spans */
 432 int unregister_mem_sect_under_nodes(struct memory_block *mem_blk,
 433                                     unsigned long phys_index)
 434 {
 435         NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
 436         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 437
 438         if (!mem_blk) {
 439                 NODEMASK_FREE(unlinked_nodes);
 440                 return -EFAULT;
 441         }
 442         if (!unlinked_nodes)
 443                 return -ENOMEM;
 444         nodes_clear(*unlinked_nodes);
 445
 446         sect_start_pfn = section_nr_to_pfn(phys_index);
 447         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
 448         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
 449                 int nid;
 450
 451                 nid = get_nid_for_pfn(pfn);
 452                 if (nid < 0)
 453                         continue;
 454                 if (!node_online(nid))
 455                         continue;
 456                 if (node_test_and_set(nid, *unlinked_nodes))
 457                         continue;
 458                 sysfs_remove_link(&node_devices[nid]->dev.kobj,
 459                          kobject_name(&mem_blk->dev.kobj));
 460                 sysfs_remove_link(&mem_blk->dev.kobj,
 461                          kobject_name(&node_devices[nid]->dev.kobj));
 462         }
 463         NODEMASK_FREE(unlinked_nodes);
 464         return 0;
 465 }
 466
 467 static int link_mem_sections(int nid)
 468 {
 469         unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
 470         unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
 471         unsigned long pfn;
 472         struct memory_block *mem_blk = NULL;
 473         int err = 0;
 474
 475         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 476                 unsigned long section_nr = pfn_to_section_nr(pfn);
 477                 struct mem_section *mem_sect;
 478                 int ret;
 479
 480                 if (!present_section_nr(section_nr))
 481                         continue;
 482                 mem_sect = __nr_to_section(section_nr);
 483
 484                 /* same memblock ? */
 485                 if (mem_blk)
 486                         if ((section_nr >= mem_blk->start_section_nr) &&
 487                             (section_nr <= mem_blk->end_section_nr))
 488                                 continue;
 489
 490                 mem_blk = find_memory_block_hinted(mem_sect, mem_blk);
 491
 492                 ret = register_mem_sect_under_node(mem_blk, nid);
 493                 if (!err)
 494                         err = ret;
 495
 496                 /* discard ref obtained in find_memory_block() */
 497         }
 498
 499         if (mem_blk)
 500                 kobject_put(&mem_blk->dev.kobj);
 501         return err;
 502 }
 503
 504 #ifdef CONFIG_HUGETLBFS
 505 /*
 506  * Handle per node hstate attribute [un]registration on transistions
 507  * to/from memoryless state.
 508  */
 509 static void node_hugetlb_work(struct work_struct *work)
 510 {
 511         struct node *node = container_of(work, struct node, node_work);
 512
 513         /*
 514          * We only get here when a node transitions to/from memoryless state.
 515          * We can detect which transition occurred by examining whether the
 516          * node has memory now.  hugetlb_register_node() already check this
 517          * so we try to register the attributes.  If that fails, then the
 518          * node has transitioned to memoryless, try to unregister the
 519          * attributes.
 520          */
 521         if (!hugetlb_register_node(node))
 522                 hugetlb_unregister_node(node);
 523 }
 524
 525 static void init_node_hugetlb_work(int nid)
 526 {
 527         INIT_WORK(&node_devices[nid]->node_work, node_hugetlb_work);
 528 }
 529
 530 static int node_memory_callback(struct notifier_block *self,
 531                                 unsigned long action, void *arg)
 532 {
 533         struct memory_notify *mnb = arg;
 534         int nid = mnb->status_change_nid;
 535
 536         switch (action) {
 537         case MEM_ONLINE:
 538         case MEM_OFFLINE:
 539                 /*
 540                  * offload per node hstate [un]registration to a work thread
 541                  * when transitioning to/from memoryless state.
 542                  */
 543                 if (nid != NUMA_NO_NODE)
 544                         schedule_work(&node_devices[nid]->node_work);
 545                 break;
 546
 547         case MEM_GOING_ONLINE:
 548         case MEM_GOING_OFFLINE:
 549         case MEM_CANCEL_ONLINE:
 550         case MEM_CANCEL_OFFLINE:
 551         default:
 552                 break;
 553         }
 554
 555         return NOTIFY_OK;
 556 }
 557 #endif  /* CONFIG_HUGETLBFS */
 558 #else   /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
 559
 560 static int link_mem_sections(int nid) { return 0; }
 561 #endif  /* CONFIG_MEMORY_HOTPLUG_SPARSE */
 562
 563 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
 564     !defined(CONFIG_HUGETLBFS)
 565 static inline int node_memory_callback(struct notifier_block *self,
 566                                 unsigned long action, void *arg)
 567 {
 568         return NOTIFY_OK;
 569 }
 570
 571 static void init_node_hugetlb_work(int nid) { }
 572
 573 #endif
 574
 575 int register_one_node(int nid)
 576 {
 577         int error = 0;
 578         int cpu;
 579
 580         if (node_online(nid)) {
 581                 int p_node = parent_node(nid);
 582                 struct node *parent = NULL;
 583
 584                 if (p_node != nid)
 585                         parent = node_devices[p_node];
 586
 587                 node_devices[nid] = kzalloc(sizeof(struct node), GFP_KERNEL);
 588                 if (!node_devices[nid])
 589                         return -ENOMEM;
 590
 591                 error = register_node(node_devices[nid], nid, parent);
 592
 593                 /* link cpu under this node */
 594                 for_each_present_cpu(cpu) {
 595                         if (cpu_to_node(cpu) == nid)
 596                                 register_cpu_under_node(cpu, nid);
 597                 }
 598
 599                 /* link memory sections under this node */
 600                 error = link_mem_sections(nid);
 601
 602                 /* initialize work queue for memory hot plug */
 603                 init_node_hugetlb_work(nid);
 604         }
 605
 606         return error;
 607
 608 }
 609
 610 void unregister_one_node(int nid)
 611 {
 612         if (!node_devices[nid])
 613                 return;
 614
 615         unregister_node(node_devices[nid]);
 616         node_devices[nid] = NULL;
 617 }
 618
 619 /*
 620  * node states attributes
 621  */
 622
 623 static ssize_t print_nodes_state(enum node_states state, char *buf)
 624 {
 625         int n;
 626
 627         n = nodelist_scnprintf(buf, PAGE_SIZE-2, node_states[state]);
 628         buf[n++] = '\n';
 629         buf[n] = '\0';
 630         return n;
 631 }
 632
 633 struct node_attr {
 634         struct device_attribute attr;
 635         enum node_states state;
 636 };
 637
 638 static ssize_t show_node_state(struct device *dev,
 639                                struct device_attribute *attr, char *buf)
 640 {
 641         struct node_attr *na = container_of(attr, struct node_attr, attr);
 642         return print_nodes_state(na->state, buf);
 643 }
 644
 645 #define _NODE_ATTR(name, state) \
 646         { __ATTR(name, 0444, show_node_state, NULL), state }
 647
 648 static struct node_attr node_state_attr[] = {
 649         [N_POSSIBLE] = _NODE_ATTR(possible, N_POSSIBLE),
 650         [N_ONLINE] = _NODE_ATTR(online, N_ONLINE),
 651         [N_NORMAL_MEMORY] = _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
 652 #ifdef CONFIG_HIGHMEM
 653         [N_HIGH_MEMORY] = _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
 654 #endif
 655 #ifdef CONFIG_MOVABLE_NODE
 656         [N_MEMORY] = _NODE_ATTR(has_memory, N_MEMORY),
 657 #endif
 658         [N_CPU] = _NODE_ATTR(has_cpu, N_CPU),
 659 };
 660
 661 static struct attribute *node_state_attrs[] = {
 662         &node_state_attr[N_POSSIBLE].attr.attr,
 663         &node_state_attr[N_ONLINE].attr.attr,
 664         &node_state_attr[N_NORMAL_MEMORY].attr.attr,
 665 #ifdef CONFIG_HIGHMEM
 666         &node_state_attr[N_HIGH_MEMORY].attr.attr,
 667 #endif
 668 #ifdef CONFIG_MOVABLE_NODE
 669         &node_state_attr[N_MEMORY].attr.attr,
 670 #endif
 671         &node_state_attr[N_CPU].attr.attr,
 672         NULL
 673 };
 674
 675 static struct attribute_group memory_root_attr_group = {
 676         .attrs = node_state_attrs,
 677 };
 678
 679 static const struct attribute_group *cpu_root_attr_groups[] = {
 680         &memory_root_attr_group,
 681         NULL,
 682 };
 683
 684 #define NODE_CALLBACK_PRI       2       /* lower than SLAB */
 685 static int __init register_node_type(void)
 686 {
 687         int ret;
 688
 689         BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
 690         BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
 691
 692         ret = subsys_system_register(&node_subsys, cpu_root_attr_groups);
 693         if (!ret) {
 694                 static struct notifier_block node_memory_callback_nb = {
 695                         .notifier_call = node_memory_callback,
 696                         .priority = NODE_CALLBACK_PRI,
 697                 };
 698                 register_hotmemory_notifier(&node_memory_callback_nb);
 699         }
 700
 701         /*
 702          * Note:  we're not going to unregister the node class if we fail
 703          * to register the node state class attribute files.
 704          */
 705         return ret;
 706 }
 707 postcore_initcall(register_node_type);