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                        "Node %d ShmemHugePages: %8lu kB\n"
 117                        "Node %d ShmemPmdMapped: %8lu kB\n"
 118 #endif
 119                         ,
 120                        nid, K(node_page_state(nid, NR_FILE_DIRTY)),
 121                        nid, K(node_page_state(nid, NR_WRITEBACK)),
 122                        nid, K(node_page_state(nid, NR_FILE_PAGES)),
 123                        nid, K(node_page_state(nid, NR_FILE_MAPPED)),
 124                        nid, K(node_page_state(nid, NR_ANON_PAGES)),
 125                        nid, K(i.sharedram),
 126                        nid, node_page_state(nid, NR_KERNEL_STACK) *
 127                                 THREAD_SIZE / 1024,
 128                        nid, K(node_page_state(nid, NR_PAGETABLE)),
 129                        nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
 130                        nid, K(node_page_state(nid, NR_BOUNCE)),
 131                        nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
 132                        nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
 133                                 node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
 134                        nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
 135 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 136                        nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
 137                        nid, K(node_page_state(nid, NR_ANON_THPS) *
 138                                        HPAGE_PMD_NR),
 139                        nid, K(node_page_state(nid, NR_SHMEM_THPS) *
 140                                        HPAGE_PMD_NR),
 141                        nid, K(node_page_state(nid, NR_SHMEM_PMDMAPPED) *
 142                                        HPAGE_PMD_NR));
 143 #else
 144                        nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
 145 #endif
 146         n += hugetlb_report_node_meminfo(nid, buf + n);
 147         return n;
 148 }
 149
 150 #undef K
 151 static DEVICE_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
 152
 153 static ssize_t node_read_numastat(struct device *dev,
 154                                 struct device_attribute *attr, char *buf)
 155 {
 156         return sprintf(buf,
 157                        "numa_hit %lu\n"
 158                        "numa_miss %lu\n"
 159                        "numa_foreign %lu\n"
 160                        "interleave_hit %lu\n"
 161                        "local_node %lu\n"
 162                        "other_node %lu\n",
 163                        node_page_state(dev->id, NUMA_HIT),
 164                        node_page_state(dev->id, NUMA_MISS),
 165                        node_page_state(dev->id, NUMA_FOREIGN),
 166                        node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
 167                        node_page_state(dev->id, NUMA_LOCAL),
 168                        node_page_state(dev->id, NUMA_OTHER));
 169 }
 170 static DEVICE_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
 171
 172 static ssize_t node_read_vmstat(struct device *dev,
 173                                 struct device_attribute *attr, char *buf)
 174 {
 175         int nid = dev->id;
 176         int i;
 177         int n = 0;
 178
 179         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 180                 n += sprintf(buf+n, "%s %lu\n", vmstat_text[i],
 181                              node_page_state(nid, i));
 182
 183         return n;
 184 }
 185 static DEVICE_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
 186
 187 static ssize_t node_read_distance(struct device *dev,
 188                         struct device_attribute *attr, char *buf)
 189 {
 190         int nid = dev->id;
 191         int len = 0;
 192         int i;
 193
 194         /*
 195          * buf is currently PAGE_SIZE in length and each node needs 4 chars
 196          * at the most (distance + space or newline).
 197          */
 198         BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
 199
 200         for_each_online_node(i)
 201                 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
 202
 203         len += sprintf(buf + len, "\n");
 204         return len;
 205 }
 206 static DEVICE_ATTR(distance, S_IRUGO, node_read_distance, NULL);
 207
 208 static struct attribute *node_dev_attrs[] = {
 209         &dev_attr_cpumap.attr,
 210         &dev_attr_cpulist.attr,
 211         &dev_attr_meminfo.attr,
 212         &dev_attr_numastat.attr,
 213         &dev_attr_distance.attr,
 214         &dev_attr_vmstat.attr,
 215         NULL
 216 };
 217 ATTRIBUTE_GROUPS(node_dev);
 218
 219 #ifdef CONFIG_HUGETLBFS
 220 /*
 221  * hugetlbfs per node attributes registration interface:
 222  * When/if hugetlb[fs] subsystem initializes [sometime after this module],
 223  * it will register its per node attributes for all online nodes with
 224  * memory.  It will also call register_hugetlbfs_with_node(), below, to
 225  * register its attribute registration functions with this node driver.
 226  * Once these hooks have been initialized, the node driver will call into
 227  * the hugetlb module to [un]register attributes for hot-plugged nodes.
 228  */
 229 static node_registration_func_t __hugetlb_register_node;
 230 static node_registration_func_t __hugetlb_unregister_node;
 231
 232 static inline bool hugetlb_register_node(struct node *node)
 233 {
 234         if (__hugetlb_register_node &&
 235                         node_state(node->dev.id, N_MEMORY)) {
 236                 __hugetlb_register_node(node);
 237                 return true;
 238         }
 239         return false;
 240 }
 241
 242 static inline void hugetlb_unregister_node(struct node *node)
 243 {
 244         if (__hugetlb_unregister_node)
 245                 __hugetlb_unregister_node(node);
 246 }
 247
 248 void register_hugetlbfs_with_node(node_registration_func_t doregister,
 249                                   node_registration_func_t unregister)
 250 {
 251         __hugetlb_register_node   = doregister;
 252         __hugetlb_unregister_node = unregister;
 253 }
 254 #else
 255 static inline void hugetlb_register_node(struct node *node) {}
 256
 257 static inline void hugetlb_unregister_node(struct node *node) {}
 258 #endif
 259
 260 static void node_device_release(struct device *dev)
 261 {
 262         struct node *node = to_node(dev);
 263
 264 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HUGETLBFS)
 265         /*
 266          * We schedule the work only when a memory section is
 267          * onlined/offlined on this node. When we come here,
 268          * all the memory on this node has been offlined,
 269          * so we won't enqueue new work to this work.
 270          *
 271          * The work is using node->node_work, so we should
 272          * flush work before freeing the memory.
 273          */
 274         flush_work(&node->node_work);
 275 #endif
 276         kfree(node);
 277 }
 278
 279 /*
 280  * register_node - Setup a sysfs device for a node.
 281  * @num - Node number to use when creating the device.
 282  *
 283  * Initialize and register the node device.
 284  */
 285 static int register_node(struct node *node, int num, struct node *parent)
 286 {
 287         int error;
 288
 289         node->dev.id = num;
 290         node->dev.bus = &node_subsys;
 291         node->dev.release = node_device_release;
 292         node->dev.groups = node_dev_groups;
 293         error = device_register(&node->dev);
 294
 295         if (!error){
 296                 hugetlb_register_node(node);
 297
 298                 compaction_register_node(node);
 299         }
 300         return error;
 301 }
 302
 303 /**
 304  * unregister_node - unregister a node device
 305  * @node: node going away
 306  *
 307  * Unregisters a node device @node.  All the devices on the node must be
 308  * unregistered before calling this function.
 309  */
 310 void unregister_node(struct node *node)
 311 {
 312         hugetlb_unregister_node(node);          /* no-op, if memoryless node */
 313
 314         device_unregister(&node->dev);
 315 }
 316
 317 struct node *node_devices[MAX_NUMNODES];
 318
 319 /*
 320  * register cpu under node
 321  */
 322 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
 323 {
 324         int ret;
 325         struct device *obj;
 326
 327         if (!node_online(nid))
 328                 return 0;
 329
 330         obj = get_cpu_device(cpu);
 331         if (!obj)
 332                 return 0;
 333
 334         ret = sysfs_create_link(&node_devices[nid]->dev.kobj,
 335                                 &obj->kobj,
 336                                 kobject_name(&obj->kobj));
 337         if (ret)
 338                 return ret;
 339
 340         return sysfs_create_link(&obj->kobj,
 341                                  &node_devices[nid]->dev.kobj,
 342                                  kobject_name(&node_devices[nid]->dev.kobj));
 343 }
 344
 345 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
 346 {
 347         struct device *obj;
 348
 349         if (!node_online(nid))
 350                 return 0;
 351
 352         obj = get_cpu_device(cpu);
 353         if (!obj)
 354                 return 0;
 355
 356         sysfs_remove_link(&node_devices[nid]->dev.kobj,
 357                           kobject_name(&obj->kobj));
 358         sysfs_remove_link(&obj->kobj,
 359                           kobject_name(&node_devices[nid]->dev.kobj));
 360
 361         return 0;
 362 }
 363
 364 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
 365 #define page_initialized(page)  (page->lru.next)
 366
 367 static int __init_refok get_nid_for_pfn(unsigned long pfn)
 368 {
 369         struct page *page;
 370
 371         if (!pfn_valid_within(pfn))
 372                 return -1;
 373 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
 374         if (system_state == SYSTEM_BOOTING)
 375                 return early_pfn_to_nid(pfn);
 376 #endif
 377         page = pfn_to_page(pfn);
 378         if (!page_initialized(page))
 379                 return -1;
 380         return pfn_to_nid(pfn);
 381 }
 382
 383 /* register memory section under specified node if it spans that node */
 384 int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
 385 {
 386         int ret;
 387         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 388
 389         if (!mem_blk)
 390                 return -EFAULT;
 391         if (!node_online(nid))
 392                 return 0;
 393
 394         sect_start_pfn = section_nr_to_pfn(mem_blk->start_section_nr);
 395         sect_end_pfn = section_nr_to_pfn(mem_blk->end_section_nr);
 396         sect_end_pfn += PAGES_PER_SECTION - 1;
 397         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
 398                 int page_nid;
 399
 400                 /*
 401                  * memory block could have several absent sections from start.
 402                  * skip pfn range from absent section
 403                  */
 404                 if (!pfn_present(pfn)) {
 405                         pfn = round_down(pfn + PAGES_PER_SECTION,
 406                                          PAGES_PER_SECTION) - 1;
 407                         continue;
 408                 }
 409
 410                 page_nid = get_nid_for_pfn(pfn);
 411                 if (page_nid < 0)
 412                         continue;
 413                 if (page_nid != nid)
 414                         continue;
 415                 ret = sysfs_create_link_nowarn(&node_devices[nid]->dev.kobj,
 416                                         &mem_blk->dev.kobj,
 417                                         kobject_name(&mem_blk->dev.kobj));
 418                 if (ret)
 419                         return ret;
 420
 421                 return sysfs_create_link_nowarn(&mem_blk->dev.kobj,
 422                                 &node_devices[nid]->dev.kobj,
 423                                 kobject_name(&node_devices[nid]->dev.kobj));
 424         }
 425         /* mem section does not span the specified node */
 426         return 0;
 427 }
 428
 429 /* unregister memory section under all nodes that it spans */
 430 int unregister_mem_sect_under_nodes(struct memory_block *mem_blk,
 431                                     unsigned long phys_index)
 432 {
 433         NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
 434         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 435
 436         if (!mem_blk) {
 437                 NODEMASK_FREE(unlinked_nodes);
 438                 return -EFAULT;
 439         }
 440         if (!unlinked_nodes)
 441                 return -ENOMEM;
 442         nodes_clear(*unlinked_nodes);
 443
 444         sect_start_pfn = section_nr_to_pfn(phys_index);
 445         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
 446         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
 447                 int nid;
 448
 449                 nid = get_nid_for_pfn(pfn);
 450                 if (nid < 0)
 451                         continue;
 452                 if (!node_online(nid))
 453                         continue;
 454                 if (node_test_and_set(nid, *unlinked_nodes))
 455                         continue;
 456                 sysfs_remove_link(&node_devices[nid]->dev.kobj,
 457                          kobject_name(&mem_blk->dev.kobj));
 458                 sysfs_remove_link(&mem_blk->dev.kobj,
 459                          kobject_name(&node_devices[nid]->dev.kobj));
 460         }
 461         NODEMASK_FREE(unlinked_nodes);
 462         return 0;
 463 }
 464
 465 static int link_mem_sections(int nid)
 466 {
 467         unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
 468         unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
 469         unsigned long pfn;
 470         struct memory_block *mem_blk = NULL;
 471         int err = 0;
 472
 473         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 474                 unsigned long section_nr = pfn_to_section_nr(pfn);
 475                 struct mem_section *mem_sect;
 476                 int ret;
 477
 478                 if (!present_section_nr(section_nr))
 479                         continue;
 480                 mem_sect = __nr_to_section(section_nr);
 481
 482                 /* same memblock ? */
 483                 if (mem_blk)
 484                         if ((section_nr >= mem_blk->start_section_nr) &&
 485                             (section_nr <= mem_blk->end_section_nr))
 486                                 continue;
 487
 488                 mem_blk = find_memory_block_hinted(mem_sect, mem_blk);
 489
 490                 ret = register_mem_sect_under_node(mem_blk, nid);
 491                 if (!err)
 492                         err = ret;
 493
 494                 /* discard ref obtained in find_memory_block() */
 495         }
 496
 497         if (mem_blk)
 498                 kobject_put(&mem_blk->dev.kobj);
 499         return err;
 500 }
 501
 502 #ifdef CONFIG_HUGETLBFS
 503 /*
 504  * Handle per node hstate attribute [un]registration on transistions
 505  * to/from memoryless state.
 506  */
 507 static void node_hugetlb_work(struct work_struct *work)
 508 {
 509         struct node *node = container_of(work, struct node, node_work);
 510
 511         /*
 512          * We only get here when a node transitions to/from memoryless state.
 513          * We can detect which transition occurred by examining whether the
 514          * node has memory now.  hugetlb_register_node() already check this
 515          * so we try to register the attributes.  If that fails, then the
 516          * node has transitioned to memoryless, try to unregister the
 517          * attributes.
 518          */
 519         if (!hugetlb_register_node(node))
 520                 hugetlb_unregister_node(node);
 521 }
 522
 523 static void init_node_hugetlb_work(int nid)
 524 {
 525         INIT_WORK(&node_devices[nid]->node_work, node_hugetlb_work);
 526 }
 527
 528 static int node_memory_callback(struct notifier_block *self,
 529                                 unsigned long action, void *arg)
 530 {
 531         struct memory_notify *mnb = arg;
 532         int nid = mnb->status_change_nid;
 533
 534         switch (action) {
 535         case MEM_ONLINE:
 536         case MEM_OFFLINE:
 537                 /*
 538                  * offload per node hstate [un]registration to a work thread
 539                  * when transitioning to/from memoryless state.
 540                  */
 541                 if (nid != NUMA_NO_NODE)
 542                         schedule_work(&node_devices[nid]->node_work);
 543                 break;
 544
 545         case MEM_GOING_ONLINE:
 546         case MEM_GOING_OFFLINE:
 547         case MEM_CANCEL_ONLINE:
 548         case MEM_CANCEL_OFFLINE:
 549         default:
 550                 break;
 551         }
 552
 553         return NOTIFY_OK;
 554 }
 555 #endif  /* CONFIG_HUGETLBFS */
 556 #else   /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
 557
 558 static int link_mem_sections(int nid) { return 0; }
 559 #endif  /* CONFIG_MEMORY_HOTPLUG_SPARSE */
 560
 561 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
 562     !defined(CONFIG_HUGETLBFS)
 563 static inline int node_memory_callback(struct notifier_block *self,
 564                                 unsigned long action, void *arg)
 565 {
 566         return NOTIFY_OK;
 567 }
 568
 569 static void init_node_hugetlb_work(int nid) { }
 570
 571 #endif
 572
 573 int register_one_node(int nid)
 574 {
 575         int error = 0;
 576         int cpu;
 577
 578         if (node_online(nid)) {
 579                 int p_node = parent_node(nid);
 580                 struct node *parent = NULL;
 581
 582                 if (p_node != nid)
 583                         parent = node_devices[p_node];
 584
 585                 node_devices[nid] = kzalloc(sizeof(struct node), GFP_KERNEL);
 586                 if (!node_devices[nid])
 587                         return -ENOMEM;
 588
 589                 error = register_node(node_devices[nid], nid, parent);
 590
 591                 /* link cpu under this node */
 592                 for_each_present_cpu(cpu) {
 593                         if (cpu_to_node(cpu) == nid)
 594                                 register_cpu_under_node(cpu, nid);
 595                 }
 596
 597                 /* link memory sections under this node */
 598                 error = link_mem_sections(nid);
 599
 600                 /* initialize work queue for memory hot plug */
 601                 init_node_hugetlb_work(nid);
 602         }
 603
 604         return error;
 605
 606 }
 607
 608 void unregister_one_node(int nid)
 609 {
 610         if (!node_devices[nid])
 611                 return;
 612
 613         unregister_node(node_devices[nid]);
 614         node_devices[nid] = NULL;
 615 }
 616
 617 /*
 618  * node states attributes
 619  */
 620
 621 static ssize_t print_nodes_state(enum node_states state, char *buf)
 622 {
 623         int n;
 624
 625         n = scnprintf(buf, PAGE_SIZE - 1, "%*pbl",
 626                       nodemask_pr_args(&node_states[state]));
 627         buf[n++] = '\n';
 628         buf[n] = '\0';
 629         return n;
 630 }
 631
 632 struct node_attr {
 633         struct device_attribute attr;
 634         enum node_states state;
 635 };
 636
 637 static ssize_t show_node_state(struct device *dev,
 638                                struct device_attribute *attr, char *buf)
 639 {
 640         struct node_attr *na = container_of(attr, struct node_attr, attr);
 641         return print_nodes_state(na->state, buf);
 642 }
 643
 644 #define _NODE_ATTR(name, state) \
 645         { __ATTR(name, 0444, show_node_state, NULL), state }
 646
 647 static struct node_attr node_state_attr[] = {
 648         [N_POSSIBLE] = _NODE_ATTR(possible, N_POSSIBLE),
 649         [N_ONLINE] = _NODE_ATTR(online, N_ONLINE),
 650         [N_NORMAL_MEMORY] = _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
 651 #ifdef CONFIG_HIGHMEM
 652         [N_HIGH_MEMORY] = _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
 653 #endif
 654 #ifdef CONFIG_MOVABLE_NODE
 655         [N_MEMORY] = _NODE_ATTR(has_memory, N_MEMORY),
 656 #endif
 657         [N_CPU] = _NODE_ATTR(has_cpu, N_CPU),
 658 };
 659
 660 static struct attribute *node_state_attrs[] = {
 661         &node_state_attr[N_POSSIBLE].attr.attr,
 662         &node_state_attr[N_ONLINE].attr.attr,
 663         &node_state_attr[N_NORMAL_MEMORY].attr.attr,
 664 #ifdef CONFIG_HIGHMEM
 665         &node_state_attr[N_HIGH_MEMORY].attr.attr,
 666 #endif
 667 #ifdef CONFIG_MOVABLE_NODE
 668         &node_state_attr[N_MEMORY].attr.attr,
 669 #endif
 670         &node_state_attr[N_CPU].attr.attr,
 671         NULL
 672 };
 673
 674 static struct attribute_group memory_root_attr_group = {
 675         .attrs = node_state_attrs,
 676 };
 677
 678 static const struct attribute_group *cpu_root_attr_groups[] = {
 679         &memory_root_attr_group,
 680         NULL,
 681 };
 682
 683 #define NODE_CALLBACK_PRI       2       /* lower than SLAB */
 684 static int __init register_node_type(void)
 685 {
 686         int ret;
 687
 688         BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
 689         BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
 690
 691         ret = subsys_system_register(&node_subsys, cpu_root_attr_groups);
 692         if (!ret) {
 693                 static struct notifier_block node_memory_callback_nb = {
 694                         .notifier_call = node_memory_callback,
 695                         .priority = NODE_CALLBACK_PRI,
 696                 };
 697                 register_hotmemory_notifier(&node_memory_callback_nb);
 698         }
 699
 700         /*
 701          * Note:  we're not going to unregister the node class if we fail
 702          * to register the node state class attribute files.
 703          */
 704         return ret;
 705 }
 706 postcore_initcall(register_node_type);