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/node.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/cpumask.h>
  13 #include <linux/topology.h>
  14 #include <linux/nodemask.h>
  15 #include <linux/cpu.h>
  16 #include <linux/device.h>
  17 #include <linux/swap.h>
  18 #include <linux/slab.h>
  19
  20 static struct sysdev_class_attribute *node_state_attrs[];
  21
  22 static struct sysdev_class node_class = {
  23         .name = "node",
  24         .attrs = node_state_attrs,
  25 };
  26
  27
  28 static ssize_t node_read_cpumap(struct sys_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->sysdev.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 sys_device *dev,
  46                                 struct sysdev_attribute *attr, char *buf)
  47 {
  48         return node_read_cpumap(dev, 0, buf);
  49 }
  50 static inline ssize_t node_read_cpulist(struct sys_device *dev,
  51                                 struct sysdev_attribute *attr, char *buf)
  52 {
  53         return node_read_cpumap(dev, 1, buf);
  54 }
  55
  56 static SYSDEV_ATTR(cpumap,  S_IRUGO, node_read_cpumask, NULL);
  57 static SYSDEV_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 sys_device * dev,
  61                         struct sysdev_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
  69         n = sprintf(buf, "\n"
  70                        "Node %d MemTotal:       %8lu kB\n"
  71                        "Node %d MemFree:        %8lu kB\n"
  72                        "Node %d MemUsed:        %8lu kB\n"
  73                        "Node %d Active:         %8lu kB\n"
  74                        "Node %d Inactive:       %8lu kB\n"
  75                        "Node %d Active(anon):   %8lu kB\n"
  76                        "Node %d Inactive(anon): %8lu kB\n"
  77                        "Node %d Active(file):   %8lu kB\n"
  78                        "Node %d Inactive(file): %8lu kB\n"
  79                        "Node %d Unevictable:    %8lu kB\n"
  80                        "Node %d Mlocked:        %8lu kB\n"
  81 #ifdef CONFIG_HIGHMEM
  82                        "Node %d HighTotal:      %8lu kB\n"
  83                        "Node %d HighFree:       %8lu kB\n"
  84                        "Node %d LowTotal:       %8lu kB\n"
  85                        "Node %d LowFree:        %8lu kB\n"
  86 #endif
  87                        "Node %d Dirty:          %8lu kB\n"
  88                        "Node %d Writeback:      %8lu kB\n"
  89                        "Node %d FilePages:      %8lu kB\n"
  90                        "Node %d Mapped:         %8lu kB\n"
  91                        "Node %d AnonPages:      %8lu kB\n"
  92                        "Node %d Shmem:          %8lu kB\n"
  93                        "Node %d KernelStack:    %8lu kB\n"
  94                        "Node %d PageTables:     %8lu kB\n"
  95                        "Node %d NFS_Unstable:   %8lu kB\n"
  96                        "Node %d Bounce:         %8lu kB\n"
  97                        "Node %d WritebackTmp:   %8lu kB\n"
  98                        "Node %d Slab:           %8lu kB\n"
  99                        "Node %d SReclaimable:   %8lu kB\n"
 100                        "Node %d SUnreclaim:     %8lu kB\n",
 101                        nid, K(i.totalram),
 102                        nid, K(i.freeram),
 103                        nid, K(i.totalram - i.freeram),
 104                        nid, K(node_page_state(nid, NR_ACTIVE_ANON) +
 105                                 node_page_state(nid, NR_ACTIVE_FILE)),
 106                        nid, K(node_page_state(nid, NR_INACTIVE_ANON) +
 107                                 node_page_state(nid, NR_INACTIVE_FILE)),
 108                        nid, K(node_page_state(nid, NR_ACTIVE_ANON)),
 109                        nid, K(node_page_state(nid, NR_INACTIVE_ANON)),
 110                        nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
 111                        nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
 112                        nid, K(node_page_state(nid, NR_UNEVICTABLE)),
 113                        nid, K(node_page_state(nid, NR_MLOCK)),
 114 #ifdef CONFIG_HIGHMEM
 115                        nid, K(i.totalhigh),
 116                        nid, K(i.freehigh),
 117                        nid, K(i.totalram - i.totalhigh),
 118                        nid, K(i.freeram - i.freehigh),
 119 #endif
 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(node_page_state(nid, NR_SHMEM)),
 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                        nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
 136         n += hugetlb_report_node_meminfo(nid, buf + n);
 137         return n;
 138 }
 139
 140 #undef K
 141 static SYSDEV_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
 142
 143 static ssize_t node_read_numastat(struct sys_device * dev,
 144                                 struct sysdev_attribute *attr, char * buf)
 145 {
 146         return sprintf(buf,
 147                        "numa_hit %lu\n"
 148                        "numa_miss %lu\n"
 149                        "numa_foreign %lu\n"
 150                        "interleave_hit %lu\n"
 151                        "local_node %lu\n"
 152                        "other_node %lu\n",
 153                        node_page_state(dev->id, NUMA_HIT),
 154                        node_page_state(dev->id, NUMA_MISS),
 155                        node_page_state(dev->id, NUMA_FOREIGN),
 156                        node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
 157                        node_page_state(dev->id, NUMA_LOCAL),
 158                        node_page_state(dev->id, NUMA_OTHER));
 159 }
 160 static SYSDEV_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
 161
 162 static ssize_t node_read_distance(struct sys_device * dev,
 163                         struct sysdev_attribute *attr, char * buf)
 164 {
 165         int nid = dev->id;
 166         int len = 0;
 167         int i;
 168
 169         /*
 170          * buf is currently PAGE_SIZE in length and each node needs 4 chars
 171          * at the most (distance + space or newline).
 172          */
 173         BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
 174
 175         for_each_online_node(i)
 176                 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
 177
 178         len += sprintf(buf + len, "\n");
 179         return len;
 180 }
 181 static SYSDEV_ATTR(distance, S_IRUGO, node_read_distance, NULL);
 182
 183 #ifdef CONFIG_HUGETLBFS
 184 /*
 185  * hugetlbfs per node attributes registration interface:
 186  * When/if hugetlb[fs] subsystem initializes [sometime after this module],
 187  * it will register its per node attributes for all online nodes with
 188  * memory.  It will also call register_hugetlbfs_with_node(), below, to
 189  * register its attribute registration functions with this node driver.
 190  * Once these hooks have been initialized, the node driver will call into
 191  * the hugetlb module to [un]register attributes for hot-plugged nodes.
 192  */
 193 static node_registration_func_t __hugetlb_register_node;
 194 static node_registration_func_t __hugetlb_unregister_node;
 195
 196 static inline bool hugetlb_register_node(struct node *node)
 197 {
 198         if (__hugetlb_register_node &&
 199                         node_state(node->sysdev.id, N_HIGH_MEMORY)) {
 200                 __hugetlb_register_node(node);
 201                 return true;
 202         }
 203         return false;
 204 }
 205
 206 static inline void hugetlb_unregister_node(struct node *node)
 207 {
 208         if (__hugetlb_unregister_node)
 209                 __hugetlb_unregister_node(node);
 210 }
 211
 212 void register_hugetlbfs_with_node(node_registration_func_t doregister,
 213                                   node_registration_func_t unregister)
 214 {
 215         __hugetlb_register_node   = doregister;
 216         __hugetlb_unregister_node = unregister;
 217 }
 218 #else
 219 static inline void hugetlb_register_node(struct node *node) {}
 220
 221 static inline void hugetlb_unregister_node(struct node *node) {}
 222 #endif
 223
 224
 225 /*
 226  * register_node - Setup a sysfs device for a node.
 227  * @num - Node number to use when creating the device.
 228  *
 229  * Initialize and register the node device.
 230  */
 231 int register_node(struct node *node, int num, struct node *parent)
 232 {
 233         int error;
 234
 235         node->sysdev.id = num;
 236         node->sysdev.cls = &node_class;
 237         error = sysdev_register(&node->sysdev);
 238
 239         if (!error){
 240                 sysdev_create_file(&node->sysdev, &attr_cpumap);
 241                 sysdev_create_file(&node->sysdev, &attr_cpulist);
 242                 sysdev_create_file(&node->sysdev, &attr_meminfo);
 243                 sysdev_create_file(&node->sysdev, &attr_numastat);
 244                 sysdev_create_file(&node->sysdev, &attr_distance);
 245
 246                 scan_unevictable_register_node(node);
 247
 248                 hugetlb_register_node(node);
 249         }
 250         return error;
 251 }
 252
 253 /**
 254  * unregister_node - unregister a node device
 255  * @node: node going away
 256  *
 257  * Unregisters a node device @node.  All the devices on the node must be
 258  * unregistered before calling this function.
 259  */
 260 void unregister_node(struct node *node)
 261 {
 262         sysdev_remove_file(&node->sysdev, &attr_cpumap);
 263         sysdev_remove_file(&node->sysdev, &attr_cpulist);
 264         sysdev_remove_file(&node->sysdev, &attr_meminfo);
 265         sysdev_remove_file(&node->sysdev, &attr_numastat);
 266         sysdev_remove_file(&node->sysdev, &attr_distance);
 267
 268         scan_unevictable_unregister_node(node);
 269         hugetlb_unregister_node(node);          /* no-op, if memoryless node */
 270
 271         sysdev_unregister(&node->sysdev);
 272 }
 273
 274 struct node node_devices[MAX_NUMNODES];
 275
 276 /*
 277  * register cpu under node
 278  */
 279 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
 280 {
 281         int ret;
 282         struct sys_device *obj;
 283
 284         if (!node_online(nid))
 285                 return 0;
 286
 287         obj = get_cpu_sysdev(cpu);
 288         if (!obj)
 289                 return 0;
 290
 291         ret = sysfs_create_link(&node_devices[nid].sysdev.kobj,
 292                                 &obj->kobj,
 293                                 kobject_name(&obj->kobj));
 294         if (ret)
 295                 return ret;
 296
 297         return sysfs_create_link(&obj->kobj,
 298                                  &node_devices[nid].sysdev.kobj,
 299                                  kobject_name(&node_devices[nid].sysdev.kobj));
 300 }
 301
 302 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
 303 {
 304         struct sys_device *obj;
 305
 306         if (!node_online(nid))
 307                 return 0;
 308
 309         obj = get_cpu_sysdev(cpu);
 310         if (!obj)
 311                 return 0;
 312
 313         sysfs_remove_link(&node_devices[nid].sysdev.kobj,
 314                           kobject_name(&obj->kobj));
 315         sysfs_remove_link(&obj->kobj,
 316                           kobject_name(&node_devices[nid].sysdev.kobj));
 317
 318         return 0;
 319 }
 320
 321 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
 322 #define page_initialized(page)  (page->lru.next)
 323
 324 static int get_nid_for_pfn(unsigned long pfn)
 325 {
 326         struct page *page;
 327
 328         if (!pfn_valid_within(pfn))
 329                 return -1;
 330         page = pfn_to_page(pfn);
 331         if (!page_initialized(page))
 332                 return -1;
 333         return pfn_to_nid(pfn);
 334 }
 335
 336 /* register memory section under specified node if it spans that node */
 337 int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
 338 {
 339         int ret;
 340         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 341
 342         if (!mem_blk)
 343                 return -EFAULT;
 344         if (!node_online(nid))
 345                 return 0;
 346         sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
 347         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
 348         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
 349                 int page_nid;
 350
 351                 page_nid = get_nid_for_pfn(pfn);
 352                 if (page_nid < 0)
 353                         continue;
 354                 if (page_nid != nid)
 355                         continue;
 356                 ret = sysfs_create_link_nowarn(&node_devices[nid].sysdev.kobj,
 357                                         &mem_blk->sysdev.kobj,
 358                                         kobject_name(&mem_blk->sysdev.kobj));
 359                 if (ret)
 360                         return ret;
 361
 362                 return sysfs_create_link_nowarn(&mem_blk->sysdev.kobj,
 363                                 &node_devices[nid].sysdev.kobj,
 364                                 kobject_name(&node_devices[nid].sysdev.kobj));
 365         }
 366         /* mem section does not span the specified node */
 367         return 0;
 368 }
 369
 370 /* unregister memory section under all nodes that it spans */
 371 int unregister_mem_sect_under_nodes(struct memory_block *mem_blk)
 372 {
 373         NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
 374         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 375
 376         if (!mem_blk) {
 377                 NODEMASK_FREE(unlinked_nodes);
 378                 return -EFAULT;
 379         }
 380         if (!unlinked_nodes)
 381                 return -ENOMEM;
 382         nodes_clear(*unlinked_nodes);
 383         sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
 384         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
 385         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
 386                 int nid;
 387
 388                 nid = get_nid_for_pfn(pfn);
 389                 if (nid < 0)
 390                         continue;
 391                 if (!node_online(nid))
 392                         continue;
 393                 if (node_test_and_set(nid, *unlinked_nodes))
 394                         continue;
 395                 sysfs_remove_link(&node_devices[nid].sysdev.kobj,
 396                          kobject_name(&mem_blk->sysdev.kobj));
 397                 sysfs_remove_link(&mem_blk->sysdev.kobj,
 398                          kobject_name(&node_devices[nid].sysdev.kobj));
 399         }
 400         NODEMASK_FREE(unlinked_nodes);
 401         return 0;
 402 }
 403
 404 static int link_mem_sections(int nid)
 405 {
 406         unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
 407         unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
 408         unsigned long pfn;
 409         int err = 0;
 410
 411         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 412                 unsigned long section_nr = pfn_to_section_nr(pfn);
 413                 struct mem_section *mem_sect;
 414                 struct memory_block *mem_blk;
 415                 int ret;
 416
 417                 if (!present_section_nr(section_nr))
 418                         continue;
 419                 mem_sect = __nr_to_section(section_nr);
 420                 mem_blk = find_memory_block(mem_sect);
 421                 ret = register_mem_sect_under_node(mem_blk, nid);
 422                 if (!err)
 423                         err = ret;
 424
 425                 /* discard ref obtained in find_memory_block() */
 426                 kobject_put(&mem_blk->sysdev.kobj);
 427         }
 428         return err;
 429 }
 430
 431 #ifdef CONFIG_HUGETLBFS
 432 /*
 433  * Handle per node hstate attribute [un]registration on transistions
 434  * to/from memoryless state.
 435  */
 436 static void node_hugetlb_work(struct work_struct *work)
 437 {
 438         struct node *node = container_of(work, struct node, node_work);
 439
 440         /*
 441          * We only get here when a node transitions to/from memoryless state.
 442          * We can detect which transition occurred by examining whether the
 443          * node has memory now.  hugetlb_register_node() already check this
 444          * so we try to register the attributes.  If that fails, then the
 445          * node has transitioned to memoryless, try to unregister the
 446          * attributes.
 447          */
 448         if (!hugetlb_register_node(node))
 449                 hugetlb_unregister_node(node);
 450 }
 451
 452 static void init_node_hugetlb_work(int nid)
 453 {
 454         INIT_WORK(&node_devices[nid].node_work, node_hugetlb_work);
 455 }
 456
 457 static int node_memory_callback(struct notifier_block *self,
 458                                 unsigned long action, void *arg)
 459 {
 460         struct memory_notify *mnb = arg;
 461         int nid = mnb->status_change_nid;
 462
 463         switch (action) {
 464         case MEM_ONLINE:
 465         case MEM_OFFLINE:
 466                 /*
 467                  * offload per node hstate [un]registration to a work thread
 468                  * when transitioning to/from memoryless state.
 469                  */
 470                 if (nid != NUMA_NO_NODE)
 471                         schedule_work(&node_devices[nid].node_work);
 472                 break;
 473
 474         case MEM_GOING_ONLINE:
 475         case MEM_GOING_OFFLINE:
 476         case MEM_CANCEL_ONLINE:
 477         case MEM_CANCEL_OFFLINE:
 478         default:
 479                 break;
 480         }
 481
 482         return NOTIFY_OK;
 483 }
 484 #endif  /* CONFIG_HUGETLBFS */
 485 #else   /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
 486
 487 static int link_mem_sections(int nid) { return 0; }
 488 #endif  /* CONFIG_MEMORY_HOTPLUG_SPARSE */
 489
 490 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
 491     !defined(CONFIG_HUGETLBFS)
 492 static inline int node_memory_callback(struct notifier_block *self,
 493                                 unsigned long action, void *arg)
 494 {
 495         return NOTIFY_OK;
 496 }
 497
 498 static void init_node_hugetlb_work(int nid) { }
 499
 500 #endif
 501
 502 int register_one_node(int nid)
 503 {
 504         int error = 0;
 505         int cpu;
 506
 507         if (node_online(nid)) {
 508                 int p_node = parent_node(nid);
 509                 struct node *parent = NULL;
 510
 511                 if (p_node != nid)
 512                         parent = &node_devices[p_node];
 513
 514                 error = register_node(&node_devices[nid], nid, parent);
 515
 516                 /* link cpu under this node */
 517                 for_each_present_cpu(cpu) {
 518                         if (cpu_to_node(cpu) == nid)
 519                                 register_cpu_under_node(cpu, nid);
 520                 }
 521
 522                 /* link memory sections under this node */
 523                 error = link_mem_sections(nid);
 524
 525                 /* initialize work queue for memory hot plug */
 526                 init_node_hugetlb_work(nid);
 527         }
 528
 529         return error;
 530
 531 }
 532
 533 void unregister_one_node(int nid)
 534 {
 535         unregister_node(&node_devices[nid]);
 536 }
 537
 538 /*
 539  * node states attributes
 540  */
 541
 542 static ssize_t print_nodes_state(enum node_states state, char *buf)
 543 {
 544         int n;
 545
 546         n = nodelist_scnprintf(buf, PAGE_SIZE, node_states[state]);
 547         if (n > 0 && PAGE_SIZE > n + 1) {
 548                 *(buf + n++) = '\n';
 549                 *(buf + n++) = '\0';
 550         }
 551         return n;
 552 }
 553
 554 struct node_attr {
 555         struct sysdev_class_attribute attr;
 556         enum node_states state;
 557 };
 558
 559 static ssize_t show_node_state(struct sysdev_class *class,
 560                                struct sysdev_class_attribute *attr, char *buf)
 561 {
 562         struct node_attr *na = container_of(attr, struct node_attr, attr);
 563         return print_nodes_state(na->state, buf);
 564 }
 565
 566 #define _NODE_ATTR(name, state) \
 567         { _SYSDEV_CLASS_ATTR(name, 0444, show_node_state, NULL), state }
 568
 569 static struct node_attr node_state_attr[] = {
 570         _NODE_ATTR(possible, N_POSSIBLE),
 571         _NODE_ATTR(online, N_ONLINE),
 572         _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
 573         _NODE_ATTR(has_cpu, N_CPU),
 574 #ifdef CONFIG_HIGHMEM
 575         _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
 576 #endif
 577 };
 578
 579 static struct sysdev_class_attribute *node_state_attrs[] = {
 580         &node_state_attr[0].attr,
 581         &node_state_attr[1].attr,
 582         &node_state_attr[2].attr,
 583         &node_state_attr[3].attr,
 584 #ifdef CONFIG_HIGHMEM
 585         &node_state_attr[4].attr,
 586 #endif
 587         NULL
 588 };
 589
 590 #define NODE_CALLBACK_PRI       2       /* lower than SLAB */
 591 static int __init register_node_type(void)
 592 {
 593         int ret;
 594
 595         BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
 596         BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
 597
 598         ret = sysdev_class_register(&node_class);
 599         if (!ret) {
 600                 hotplug_memory_notifier(node_memory_callback,
 601                                         NODE_CALLBACK_PRI);
 602         }
 603
 604         /*
 605          * Note:  we're not going to unregister the node class if we fail
 606          * to register the node state class attribute files.
 607          */
 608         return ret;
 609 }
 610 postcore_initcall(register_node_type);