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
  19 static struct sysdev_class node_class = {
  20         .name = "node",
  21 };
  22
  23
  24 static ssize_t node_read_cpumap(struct sys_device *dev, int type, char *buf)
  25 {
  26         struct node *node_dev = to_node(dev);
  27         const struct cpumask *mask = cpumask_of_node(node_dev->sysdev.id);
  28         int len;
  29
  30         /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
  31         BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1));
  32
  33         len = type?
  34                 cpulist_scnprintf(buf, PAGE_SIZE-2, mask) :
  35                 cpumask_scnprintf(buf, PAGE_SIZE-2, mask);
  36         buf[len++] = '\n';
  37         buf[len] = '\0';
  38         return len;
  39 }
  40
  41 static inline ssize_t node_read_cpumask(struct sys_device *dev,
  42                                 struct sysdev_attribute *attr, char *buf)
  43 {
  44         return node_read_cpumap(dev, 0, buf);
  45 }
  46 static inline ssize_t node_read_cpulist(struct sys_device *dev,
  47                                 struct sysdev_attribute *attr, char *buf)
  48 {
  49         return node_read_cpumap(dev, 1, buf);
  50 }
  51
  52 static SYSDEV_ATTR(cpumap,  S_IRUGO, node_read_cpumask, NULL);
  53 static SYSDEV_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
  54
  55 #define K(x) ((x) << (PAGE_SHIFT - 10))
  56 static ssize_t node_read_meminfo(struct sys_device * dev,
  57                         struct sysdev_attribute *attr, char * buf)
  58 {
  59         int n;
  60         int nid = dev->id;
  61         struct sysinfo i;
  62
  63         si_meminfo_node(&i, nid);
  64
  65         n = sprintf(buf, "\n"
  66                        "Node %d MemTotal:       %8lu kB\n"
  67                        "Node %d MemFree:        %8lu kB\n"
  68                        "Node %d MemUsed:        %8lu kB\n"
  69                        "Node %d Active:         %8lu kB\n"
  70                        "Node %d Inactive:       %8lu kB\n"
  71                        "Node %d Active(anon):   %8lu kB\n"
  72                        "Node %d Inactive(anon): %8lu kB\n"
  73                        "Node %d Active(file):   %8lu kB\n"
  74                        "Node %d Inactive(file): %8lu kB\n"
  75                        "Node %d Unevictable:    %8lu kB\n"
  76                        "Node %d Mlocked:        %8lu kB\n"
  77 #ifdef CONFIG_HIGHMEM
  78                        "Node %d HighTotal:      %8lu kB\n"
  79                        "Node %d HighFree:       %8lu kB\n"
  80                        "Node %d LowTotal:       %8lu kB\n"
  81                        "Node %d LowFree:        %8lu kB\n"
  82 #endif
  83                        "Node %d Dirty:          %8lu kB\n"
  84                        "Node %d Writeback:      %8lu kB\n"
  85                        "Node %d FilePages:      %8lu kB\n"
  86                        "Node %d Mapped:         %8lu kB\n"
  87                        "Node %d AnonPages:      %8lu kB\n"
  88                        "Node %d KernelStack:    %8lu kB\n"
  89                        "Node %d PageTables:     %8lu kB\n"
  90                        "Node %d NFS_Unstable:   %8lu kB\n"
  91                        "Node %d Bounce:         %8lu kB\n"
  92                        "Node %d WritebackTmp:   %8lu kB\n"
  93                        "Node %d Slab:           %8lu kB\n"
  94                        "Node %d SReclaimable:   %8lu kB\n"
  95                        "Node %d SUnreclaim:     %8lu kB\n",
  96                        nid, K(i.totalram),
  97                        nid, K(i.freeram),
  98                        nid, K(i.totalram - i.freeram),
  99                        nid, K(node_page_state(nid, NR_ACTIVE_ANON) +
 100                                 node_page_state(nid, NR_ACTIVE_FILE)),
 101                        nid, K(node_page_state(nid, NR_INACTIVE_ANON) +
 102                                 node_page_state(nid, NR_INACTIVE_FILE)),
 103                        nid, K(node_page_state(nid, NR_ACTIVE_ANON)),
 104                        nid, K(node_page_state(nid, NR_INACTIVE_ANON)),
 105                        nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
 106                        nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
 107                        nid, K(node_page_state(nid, NR_UNEVICTABLE)),
 108                        nid, K(node_page_state(nid, NR_MLOCK)),
 109 #ifdef CONFIG_HIGHMEM
 110                        nid, K(i.totalhigh),
 111                        nid, K(i.freehigh),
 112                        nid, K(i.totalram - i.totalhigh),
 113                        nid, K(i.freeram - i.freehigh),
 114 #endif
 115                        nid, K(node_page_state(nid, NR_FILE_DIRTY)),
 116                        nid, K(node_page_state(nid, NR_WRITEBACK)),
 117                        nid, K(node_page_state(nid, NR_FILE_PAGES)),
 118                        nid, K(node_page_state(nid, NR_FILE_MAPPED)),
 119                        nid, K(node_page_state(nid, NR_ANON_PAGES)),
 120                        nid, node_page_state(nid, NR_KERNEL_STACK) *
 121                                 THREAD_SIZE / 1024,
 122                        nid, K(node_page_state(nid, NR_PAGETABLE)),
 123                        nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
 124                        nid, K(node_page_state(nid, NR_BOUNCE)),
 125                        nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
 126                        nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
 127                                 node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
 128                        nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
 129                        nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
 130         n += hugetlb_report_node_meminfo(nid, buf + n);
 131         return n;
 132 }
 133
 134 #undef K
 135 static SYSDEV_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
 136
 137 static ssize_t node_read_numastat(struct sys_device * dev,
 138                                 struct sysdev_attribute *attr, char * buf)
 139 {
 140         return sprintf(buf,
 141                        "numa_hit %lu\n"
 142                        "numa_miss %lu\n"
 143                        "numa_foreign %lu\n"
 144                        "interleave_hit %lu\n"
 145                        "local_node %lu\n"
 146                        "other_node %lu\n",
 147                        node_page_state(dev->id, NUMA_HIT),
 148                        node_page_state(dev->id, NUMA_MISS),
 149                        node_page_state(dev->id, NUMA_FOREIGN),
 150                        node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
 151                        node_page_state(dev->id, NUMA_LOCAL),
 152                        node_page_state(dev->id, NUMA_OTHER));
 153 }
 154 static SYSDEV_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
 155
 156 static ssize_t node_read_distance(struct sys_device * dev,
 157                         struct sysdev_attribute *attr, char * buf)
 158 {
 159         int nid = dev->id;
 160         int len = 0;
 161         int i;
 162
 163         /* buf currently PAGE_SIZE, need ~4 chars per node */
 164         BUILD_BUG_ON(MAX_NUMNODES*4 > PAGE_SIZE/2);
 165
 166         for_each_online_node(i)
 167                 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
 168
 169         len += sprintf(buf + len, "\n");
 170         return len;
 171 }
 172 static SYSDEV_ATTR(distance, S_IRUGO, node_read_distance, NULL);
 173
 174
 175 /*
 176  * register_node - Setup a sysfs device for a node.
 177  * @num - Node number to use when creating the device.
 178  *
 179  * Initialize and register the node device.
 180  */
 181 int register_node(struct node *node, int num, struct node *parent)
 182 {
 183         int error;
 184
 185         node->sysdev.id = num;
 186         node->sysdev.cls = &node_class;
 187         error = sysdev_register(&node->sysdev);
 188
 189         if (!error){
 190                 sysdev_create_file(&node->sysdev, &attr_cpumap);
 191                 sysdev_create_file(&node->sysdev, &attr_cpulist);
 192                 sysdev_create_file(&node->sysdev, &attr_meminfo);
 193                 sysdev_create_file(&node->sysdev, &attr_numastat);
 194                 sysdev_create_file(&node->sysdev, &attr_distance);
 195
 196                 scan_unevictable_register_node(node);
 197         }
 198         return error;
 199 }
 200
 201 /**
 202  * unregister_node - unregister a node device
 203  * @node: node going away
 204  *
 205  * Unregisters a node device @node.  All the devices on the node must be
 206  * unregistered before calling this function.
 207  */
 208 void unregister_node(struct node *node)
 209 {
 210         sysdev_remove_file(&node->sysdev, &attr_cpumap);
 211         sysdev_remove_file(&node->sysdev, &attr_cpulist);
 212         sysdev_remove_file(&node->sysdev, &attr_meminfo);
 213         sysdev_remove_file(&node->sysdev, &attr_numastat);
 214         sysdev_remove_file(&node->sysdev, &attr_distance);
 215
 216         scan_unevictable_unregister_node(node);
 217
 218         sysdev_unregister(&node->sysdev);
 219 }
 220
 221 struct node node_devices[MAX_NUMNODES];
 222
 223 /*
 224  * register cpu under node
 225  */
 226 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
 227 {
 228         if (node_online(nid)) {
 229                 struct sys_device *obj = get_cpu_sysdev(cpu);
 230                 if (!obj)
 231                         return 0;
 232                 return sysfs_create_link(&node_devices[nid].sysdev.kobj,
 233                                          &obj->kobj,
 234                                          kobject_name(&obj->kobj));
 235          }
 236
 237         return 0;
 238 }
 239
 240 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
 241 {
 242         if (node_online(nid)) {
 243                 struct sys_device *obj = get_cpu_sysdev(cpu);
 244                 if (obj)
 245                         sysfs_remove_link(&node_devices[nid].sysdev.kobj,
 246                                          kobject_name(&obj->kobj));
 247         }
 248         return 0;
 249 }
 250
 251 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
 252 #define page_initialized(page)  (page->lru.next)
 253
 254 static int get_nid_for_pfn(unsigned long pfn)
 255 {
 256         struct page *page;
 257
 258         if (!pfn_valid_within(pfn))
 259                 return -1;
 260         page = pfn_to_page(pfn);
 261         if (!page_initialized(page))
 262                 return -1;
 263         return pfn_to_nid(pfn);
 264 }
 265
 266 /* register memory section under specified node if it spans that node */
 267 int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
 268 {
 269         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 270
 271         if (!mem_blk)
 272                 return -EFAULT;
 273         if (!node_online(nid))
 274                 return 0;
 275         sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
 276         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
 277         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
 278                 int page_nid;
 279
 280                 page_nid = get_nid_for_pfn(pfn);
 281                 if (page_nid < 0)
 282                         continue;
 283                 if (page_nid != nid)
 284                         continue;
 285                 return sysfs_create_link_nowarn(&node_devices[nid].sysdev.kobj,
 286                                         &mem_blk->sysdev.kobj,
 287                                         kobject_name(&mem_blk->sysdev.kobj));
 288         }
 289         /* mem section does not span the specified node */
 290         return 0;
 291 }
 292
 293 /* unregister memory section under all nodes that it spans */
 294 int unregister_mem_sect_under_nodes(struct memory_block *mem_blk)
 295 {
 296         nodemask_t unlinked_nodes;
 297         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 298
 299         if (!mem_blk)
 300                 return -EFAULT;
 301         nodes_clear(unlinked_nodes);
 302         sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
 303         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
 304         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
 305                 int nid;
 306
 307                 nid = get_nid_for_pfn(pfn);
 308                 if (nid < 0)
 309                         continue;
 310                 if (!node_online(nid))
 311                         continue;
 312                 if (node_test_and_set(nid, unlinked_nodes))
 313                         continue;
 314                 sysfs_remove_link(&node_devices[nid].sysdev.kobj,
 315                          kobject_name(&mem_blk->sysdev.kobj));
 316         }
 317         return 0;
 318 }
 319
 320 static int link_mem_sections(int nid)
 321 {
 322         unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
 323         unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
 324         unsigned long pfn;
 325         int err = 0;
 326
 327         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 328                 unsigned long section_nr = pfn_to_section_nr(pfn);
 329                 struct mem_section *mem_sect;
 330                 struct memory_block *mem_blk;
 331                 int ret;
 332
 333                 if (!present_section_nr(section_nr))
 334                         continue;
 335                 mem_sect = __nr_to_section(section_nr);
 336                 mem_blk = find_memory_block(mem_sect);
 337                 ret = register_mem_sect_under_node(mem_blk, nid);
 338                 if (!err)
 339                         err = ret;
 340
 341                 /* discard ref obtained in find_memory_block() */
 342                 kobject_put(&mem_blk->sysdev.kobj);
 343         }
 344         return err;
 345 }
 346 #else
 347 static int link_mem_sections(int nid) { return 0; }
 348 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
 349
 350 int register_one_node(int nid)
 351 {
 352         int error = 0;
 353         int cpu;
 354
 355         if (node_online(nid)) {
 356                 int p_node = parent_node(nid);
 357                 struct node *parent = NULL;
 358
 359                 if (p_node != nid)
 360                         parent = &node_devices[p_node];
 361
 362                 error = register_node(&node_devices[nid], nid, parent);
 363
 364                 /* link cpu under this node */
 365                 for_each_present_cpu(cpu) {
 366                         if (cpu_to_node(cpu) == nid)
 367                                 register_cpu_under_node(cpu, nid);
 368                 }
 369
 370                 /* link memory sections under this node */
 371                 error = link_mem_sections(nid);
 372         }
 373
 374         return error;
 375
 376 }
 377
 378 void unregister_one_node(int nid)
 379 {
 380         unregister_node(&node_devices[nid]);
 381 }
 382
 383 /*
 384  * node states attributes
 385  */
 386
 387 static ssize_t print_nodes_state(enum node_states state, char *buf)
 388 {
 389         int n;
 390
 391         n = nodelist_scnprintf(buf, PAGE_SIZE, node_states[state]);
 392         if (n > 0 && PAGE_SIZE > n + 1) {
 393                 *(buf + n++) = '\n';
 394                 *(buf + n++) = '\0';
 395         }
 396         return n;
 397 }
 398
 399 static ssize_t print_nodes_possible(struct sysdev_class *class, char *buf)
 400 {
 401         return print_nodes_state(N_POSSIBLE, buf);
 402 }
 403
 404 static ssize_t print_nodes_online(struct sysdev_class *class, char *buf)
 405 {
 406         return print_nodes_state(N_ONLINE, buf);
 407 }
 408
 409 static ssize_t print_nodes_has_normal_memory(struct sysdev_class *class,
 410                                                 char *buf)
 411 {
 412         return print_nodes_state(N_NORMAL_MEMORY, buf);
 413 }
 414
 415 static ssize_t print_nodes_has_cpu(struct sysdev_class *class, char *buf)
 416 {
 417         return print_nodes_state(N_CPU, buf);
 418 }
 419
 420 static SYSDEV_CLASS_ATTR(possible, 0444, print_nodes_possible, NULL);
 421 static SYSDEV_CLASS_ATTR(online, 0444, print_nodes_online, NULL);
 422 static SYSDEV_CLASS_ATTR(has_normal_memory, 0444, print_nodes_has_normal_memory,
 423                                                                         NULL);
 424 static SYSDEV_CLASS_ATTR(has_cpu, 0444, print_nodes_has_cpu, NULL);
 425
 426 #ifdef CONFIG_HIGHMEM
 427 static ssize_t print_nodes_has_high_memory(struct sysdev_class *class,
 428                                                  char *buf)
 429 {
 430         return print_nodes_state(N_HIGH_MEMORY, buf);
 431 }
 432
 433 static SYSDEV_CLASS_ATTR(has_high_memory, 0444, print_nodes_has_high_memory,
 434                                                                          NULL);
 435 #endif
 436
 437 struct sysdev_class_attribute *node_state_attr[] = {
 438         &attr_possible,
 439         &attr_online,
 440         &attr_has_normal_memory,
 441 #ifdef CONFIG_HIGHMEM
 442         &attr_has_high_memory,
 443 #endif
 444         &attr_has_cpu,
 445 };
 446
 447 static int node_states_init(void)
 448 {
 449         int i;
 450         int err = 0;
 451
 452         for (i = 0;  i < NR_NODE_STATES; i++) {
 453                 int ret;
 454                 ret = sysdev_class_create_file(&node_class, node_state_attr[i]);
 455                 if (!err)
 456                         err = ret;
 457         }
 458         return err;
 459 }
 460
 461 static int __init register_node_type(void)
 462 {
 463         int ret;
 464
 465         ret = sysdev_class_register(&node_class);
 466         if (!ret)
 467                 ret = node_states_init();
 468
 469         /*
 470          * Note:  we're not going to unregister the node class if we fail
 471          * to register the node state class attribute files.
 472          */
 473         return ret;
 474 }
 475 postcore_initcall(register_node_type);