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, 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
  39 static 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 }
  44 static 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
  50 static DEVICE_ATTR(cpumap,  S_IRUGO, node_read_cpumask, NULL);
  51 static DEVICE_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
  52
  53 #define K(x) ((x) << (PAGE_SHIFT - 10))
  54 static 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
 146 static DEVICE_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
 147
 148 static 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 }
 165 static DEVICE_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
 166
 167 static 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 }
 180 static DEVICE_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
 181
 182 static 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 }
 201 static DEVICE_ATTR(distance, S_IRUGO, node_read_distance, NULL);
 202
 203 #ifdef CONFIG_HUGETLBFS
 204 /*
 205  * hugetlbfs per node attributes registration interface:
 206  * When/if hugetlb[fs] subsystem initializes [sometime after this module],
 207  * it will register its per node attributes for all online nodes with
 208  * memory.  It will also call register_hugetlbfs_with_node(), below, to
 209  * register its attribute registration functions with this node driver.
 210  * Once these hooks have been initialized, the node driver will call into
 211  * the hugetlb module to [un]register attributes for hot-plugged nodes.
 212  */
 213 static node_registration_func_t __hugetlb_register_node;
 214 static node_registration_func_t __hugetlb_unregister_node;
 215
 216 static inline bool hugetlb_register_node(struct node *node)
 217 {
 218         if (__hugetlb_register_node &&
 219                         node_state(node->dev.id, N_MEMORY)) {
 220                 __hugetlb_register_node(node);
 221                 return true;
 222         }
 223         return false;
 224 }
 225
 226 static inline void hugetlb_unregister_node(struct node *node)
 227 {
 228         if (__hugetlb_unregister_node)
 229                 __hugetlb_unregister_node(node);
 230 }
 231
 232 void register_hugetlbfs_with_node(node_registration_func_t doregister,
 233                                   node_registration_func_t unregister)
 234 {
 235         __hugetlb_register_node   = doregister;
 236         __hugetlb_unregister_node = unregister;
 237 }
 238 #else
 239 static inline void hugetlb_register_node(struct node *node) {}
 240
 241 static inline void hugetlb_unregister_node(struct node *node) {}
 242 #endif
 243
 244 static void node_device_release(struct device *dev)
 245 {
 246         struct node *node = to_node(dev);
 247
 248 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HUGETLBFS)
 249         /*
 250          * We schedule the work only when a memory section is
 251          * onlined/offlined on this node. When we come here,
 252          * all the memory on this node has been offlined,
 253          * so we won't enqueue new work to this work.
 254          *
 255          * The work is using node->node_work, so we should
 256          * flush work before freeing the memory.
 257          */
 258         flush_work(&node->node_work);
 259 #endif
 260         kfree(node);
 261 }
 262
 263 /*
 264  * register_node - Setup a sysfs device for a node.
 265  * @num - Node number to use when creating the device.
 266  *
 267  * Initialize and register the node device.
 268  */
 269 static int register_node(struct node *node, int num, struct node *parent)
 270 {
 271         int error;
 272
 273         node->dev.id = num;
 274         node->dev.bus = &node_subsys;
 275         node->dev.release = node_device_release;
 276         error = device_register(&node->dev);
 277
 278         if (!error){
 279                 device_create_file(&node->dev, &dev_attr_cpumap);
 280                 device_create_file(&node->dev, &dev_attr_cpulist);
 281                 device_create_file(&node->dev, &dev_attr_meminfo);
 282                 device_create_file(&node->dev, &dev_attr_numastat);
 283                 device_create_file(&node->dev, &dev_attr_distance);
 284                 device_create_file(&node->dev, &dev_attr_vmstat);
 285
 286                 hugetlb_register_node(node);
 287
 288                 compaction_register_node(node);
 289         }
 290         return error;
 291 }
 292
 293 /**
 294  * unregister_node - unregister a node device
 295  * @node: node going away
 296  *
 297  * Unregisters a node device @node.  All the devices on the node must be
 298  * unregistered before calling this function.
 299  */
 300 void unregister_node(struct node *node)
 301 {
 302         device_remove_file(&node->dev, &dev_attr_cpumap);
 303         device_remove_file(&node->dev, &dev_attr_cpulist);
 304         device_remove_file(&node->dev, &dev_attr_meminfo);
 305         device_remove_file(&node->dev, &dev_attr_numastat);
 306         device_remove_file(&node->dev, &dev_attr_distance);
 307         device_remove_file(&node->dev, &dev_attr_vmstat);
 308
 309         hugetlb_unregister_node(node);          /* no-op, if memoryless node */
 310
 311         device_unregister(&node->dev);
 312 }
 313
 314 struct node *node_devices[MAX_NUMNODES];
 315
 316 /*
 317  * register cpu under node
 318  */
 319 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
 320 {
 321         int ret;
 322         struct device *obj;
 323
 324         if (!node_online(nid))
 325                 return 0;
 326
 327         obj = get_cpu_device(cpu);
 328         if (!obj)
 329                 return 0;
 330
 331         ret = sysfs_create_link(&node_devices[nid]->dev.kobj,
 332                                 &obj->kobj,
 333                                 kobject_name(&obj->kobj));
 334         if (ret)
 335                 return ret;
 336
 337         return sysfs_create_link(&obj->kobj,
 338                                  &node_devices[nid]->dev.kobj,
 339                                  kobject_name(&node_devices[nid]->dev.kobj));
 340 }
 341
 342 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
 343 {
 344         struct device *obj;
 345
 346         if (!node_online(nid))
 347                 return 0;
 348
 349         obj = get_cpu_device(cpu);
 350         if (!obj)
 351                 return 0;
 352
 353         sysfs_remove_link(&node_devices[nid]->dev.kobj,
 354                           kobject_name(&obj->kobj));
 355         sysfs_remove_link(&obj->kobj,
 356                           kobject_name(&node_devices[nid]->dev.kobj));
 357
 358         return 0;
 359 }
 360
 361 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
 362 #define page_initialized(page)  (page->lru.next)
 363
 364 static int get_nid_for_pfn(unsigned long pfn)
 365 {
 366         struct page *page;
 367
 368         if (!pfn_valid_within(pfn))
 369                 return -1;
 370         page = pfn_to_page(pfn);
 371         if (!page_initialized(page))
 372                 return -1;
 373         return pfn_to_nid(pfn);
 374 }
 375
 376 /* register memory section under specified node if it spans that node */
 377 int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
 378 {
 379         int ret;
 380         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 381
 382         if (!mem_blk)
 383                 return -EFAULT;
 384         if (!node_online(nid))
 385                 return 0;
 386
 387         sect_start_pfn = section_nr_to_pfn(mem_blk->start_section_nr);
 388         sect_end_pfn = section_nr_to_pfn(mem_blk->end_section_nr);
 389         sect_end_pfn += PAGES_PER_SECTION - 1;
 390         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
 391                 int page_nid;
 392
 393                 page_nid = get_nid_for_pfn(pfn);
 394                 if (page_nid < 0)
 395                         continue;
 396                 if (page_nid != nid)
 397                         continue;
 398                 ret = sysfs_create_link_nowarn(&node_devices[nid]->dev.kobj,
 399                                         &mem_blk->dev.kobj,
 400                                         kobject_name(&mem_blk->dev.kobj));
 401                 if (ret)
 402                         return ret;
 403
 404                 return sysfs_create_link_nowarn(&mem_blk->dev.kobj,
 405                                 &node_devices[nid]->dev.kobj,
 406                                 kobject_name(&node_devices[nid]->dev.kobj));
 407         }
 408         /* mem section does not span the specified node */
 409         return 0;
 410 }
 411
 412 /* unregister memory section under all nodes that it spans */
 413 int unregister_mem_sect_under_nodes(struct memory_block *mem_blk,
 414                                     unsigned long phys_index)
 415 {
 416         NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
 417         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 418
 419         if (!mem_blk) {
 420                 NODEMASK_FREE(unlinked_nodes);
 421                 return -EFAULT;
 422         }
 423         if (!unlinked_nodes)
 424                 return -ENOMEM;
 425         nodes_clear(*unlinked_nodes);
 426
 427         sect_start_pfn = section_nr_to_pfn(phys_index);
 428         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
 429         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
 430                 int nid;
 431
 432                 nid = get_nid_for_pfn(pfn);
 433                 if (nid < 0)
 434                         continue;
 435                 if (!node_online(nid))
 436                         continue;
 437                 if (node_test_and_set(nid, *unlinked_nodes))
 438                         continue;
 439                 sysfs_remove_link(&node_devices[nid]->dev.kobj,
 440                          kobject_name(&mem_blk->dev.kobj));
 441                 sysfs_remove_link(&mem_blk->dev.kobj,
 442                          kobject_name(&node_devices[nid]->dev.kobj));
 443         }
 444         NODEMASK_FREE(unlinked_nodes);
 445         return 0;
 446 }
 447
 448 static int link_mem_sections(int nid)
 449 {
 450         unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
 451         unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
 452         unsigned long pfn;
 453         struct memory_block *mem_blk = NULL;
 454         int err = 0;
 455
 456         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 457                 unsigned long section_nr = pfn_to_section_nr(pfn);
 458                 struct mem_section *mem_sect;
 459                 int ret;
 460
 461                 if (!present_section_nr(section_nr))
 462                         continue;
 463                 mem_sect = __nr_to_section(section_nr);
 464
 465                 /* same memblock ? */
 466                 if (mem_blk)
 467                         if ((section_nr >= mem_blk->start_section_nr) &&
 468                             (section_nr <= mem_blk->end_section_nr))
 469                                 continue;
 470
 471                 mem_blk = find_memory_block_hinted(mem_sect, mem_blk);
 472
 473                 ret = register_mem_sect_under_node(mem_blk, nid);
 474                 if (!err)
 475                         err = ret;
 476
 477                 /* discard ref obtained in find_memory_block() */
 478         }
 479
 480         if (mem_blk)
 481                 kobject_put(&mem_blk->dev.kobj);
 482         return err;
 483 }
 484
 485 #ifdef CONFIG_HUGETLBFS
 486 /*
 487  * Handle per node hstate attribute [un]registration on transistions
 488  * to/from memoryless state.
 489  */
 490 static void node_hugetlb_work(struct work_struct *work)
 491 {
 492         struct node *node = container_of(work, struct node, node_work);
 493
 494         /*
 495          * We only get here when a node transitions to/from memoryless state.
 496          * We can detect which transition occurred by examining whether the
 497          * node has memory now.  hugetlb_register_node() already check this
 498          * so we try to register the attributes.  If that fails, then the
 499          * node has transitioned to memoryless, try to unregister the
 500          * attributes.
 501          */
 502         if (!hugetlb_register_node(node))
 503                 hugetlb_unregister_node(node);
 504 }
 505
 506 static void init_node_hugetlb_work(int nid)
 507 {
 508         INIT_WORK(&node_devices[nid]->node_work, node_hugetlb_work);
 509 }
 510
 511 static int node_memory_callback(struct notifier_block *self,
 512                                 unsigned long action, void *arg)
 513 {
 514         struct memory_notify *mnb = arg;
 515         int nid = mnb->status_change_nid;
 516
 517         switch (action) {
 518         case MEM_ONLINE:
 519         case MEM_OFFLINE:
 520                 /*
 521                  * offload per node hstate [un]registration to a work thread
 522                  * when transitioning to/from memoryless state.
 523                  */
 524                 if (nid != NUMA_NO_NODE)
 525                         schedule_work(&node_devices[nid]->node_work);
 526                 break;
 527
 528         case MEM_GOING_ONLINE:
 529         case MEM_GOING_OFFLINE:
 530         case MEM_CANCEL_ONLINE:
 531         case MEM_CANCEL_OFFLINE:
 532         default:
 533                 break;
 534         }
 535
 536         return NOTIFY_OK;
 537 }
 538 #endif  /* CONFIG_HUGETLBFS */
 539 #else   /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
 540
 541 static int link_mem_sections(int nid) { return 0; }
 542 #endif  /* CONFIG_MEMORY_HOTPLUG_SPARSE */
 543
 544 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
 545     !defined(CONFIG_HUGETLBFS)
 546 static inline int node_memory_callback(struct notifier_block *self,
 547                                 unsigned long action, void *arg)
 548 {
 549         return NOTIFY_OK;
 550 }
 551
 552 static void init_node_hugetlb_work(int nid) { }
 553
 554 #endif
 555
 556 int register_one_node(int nid)
 557 {
 558         int error = 0;
 559         int cpu;
 560
 561         if (node_online(nid)) {
 562                 int p_node = parent_node(nid);
 563                 struct node *parent = NULL;
 564
 565                 if (p_node != nid)
 566                         parent = node_devices[p_node];
 567
 568                 node_devices[nid] = kzalloc(sizeof(struct node), GFP_KERNEL);
 569                 if (!node_devices[nid])
 570                         return -ENOMEM;
 571
 572                 error = register_node(node_devices[nid], nid, parent);
 573
 574                 /* link cpu under this node */
 575                 for_each_present_cpu(cpu) {
 576                         if (cpu_to_node(cpu) == nid)
 577                                 register_cpu_under_node(cpu, nid);
 578                 }
 579
 580                 /* link memory sections under this node */
 581                 error = link_mem_sections(nid);
 582
 583                 /* initialize work queue for memory hot plug */
 584                 init_node_hugetlb_work(nid);
 585         }
 586
 587         return error;
 588
 589 }
 590
 591 void unregister_one_node(int nid)
 592 {
 593         if (!node_devices[nid])
 594                 return;
 595
 596         unregister_node(node_devices[nid]);
 597         node_devices[nid] = NULL;
 598 }
 599
 600 /*
 601  * node states attributes
 602  */
 603
 604 static ssize_t print_nodes_state(enum node_states state, char *buf)
 605 {
 606         int n;
 607
 608         n = nodelist_scnprintf(buf, PAGE_SIZE-2, node_states[state]);
 609         buf[n++] = '\n';
 610         buf[n] = '\0';
 611         return n;
 612 }
 613
 614 struct node_attr {
 615         struct device_attribute attr;
 616         enum node_states state;
 617 };
 618
 619 static ssize_t show_node_state(struct device *dev,
 620                                struct device_attribute *attr, char *buf)
 621 {
 622         struct node_attr *na = container_of(attr, struct node_attr, attr);
 623         return print_nodes_state(na->state, buf);
 624 }
 625
 626 #define _NODE_ATTR(name, state) \
 627         { __ATTR(name, 0444, show_node_state, NULL), state }
 628
 629 static struct node_attr node_state_attr[] = {
 630         [N_POSSIBLE] = _NODE_ATTR(possible, N_POSSIBLE),
 631         [N_ONLINE] = _NODE_ATTR(online, N_ONLINE),
 632         [N_NORMAL_MEMORY] = _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
 633 #ifdef CONFIG_HIGHMEM
 634         [N_HIGH_MEMORY] = _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
 635 #endif
 636 #ifdef CONFIG_MOVABLE_NODE
 637         [N_MEMORY] = _NODE_ATTR(has_memory, N_MEMORY),
 638 #endif
 639         [N_CPU] = _NODE_ATTR(has_cpu, N_CPU),
 640 };
 641
 642 static struct attribute *node_state_attrs[] = {
 643         &node_state_attr[N_POSSIBLE].attr.attr,
 644         &node_state_attr[N_ONLINE].attr.attr,
 645         &node_state_attr[N_NORMAL_MEMORY].attr.attr,
 646 #ifdef CONFIG_HIGHMEM
 647         &node_state_attr[N_HIGH_MEMORY].attr.attr,
 648 #endif
 649 #ifdef CONFIG_MOVABLE_NODE
 650         &node_state_attr[N_MEMORY].attr.attr,
 651 #endif
 652         &node_state_attr[N_CPU].attr.attr,
 653         NULL
 654 };
 655
 656 static struct attribute_group memory_root_attr_group = {
 657         .attrs = node_state_attrs,
 658 };
 659
 660 static const struct attribute_group *cpu_root_attr_groups[] = {
 661         &memory_root_attr_group,
 662         NULL,
 663 };
 664
 665 #define NODE_CALLBACK_PRI       2       /* lower than SLAB */
 666 static int __init register_node_type(void)
 667 {
 668         int ret;
 669
 670         BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
 671         BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
 672
 673         ret = subsys_system_register(&node_subsys, cpu_root_attr_groups);
 674         if (!ret) {
 675                 static struct notifier_block node_memory_callback_nb = {
 676                         .notifier_call = node_memory_callback,
 677                         .priority = NODE_CALLBACK_PRI,
 678                 };
 679                 register_hotmemory_notifier(&node_memory_callback_nb);
 680         }
 681
 682         /*
 683          * Note:  we're not going to unregister the node class if we fail
 684          * to register the node state class attribute files.
 685          */
 686         return ret;
 687 }
 688 postcore_initcall(register_node_type);