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_attribute *node_state_attrs[];
  20
  21 static struct sysdev_class node_class = {
  22         .name = "node",
  23         .attrs = node_state_attrs,
  24 };
  25
  26
  27 static ssize_t node_read_cpumap(struct sys_device *dev, int type, char *buf)
  28 {
  29         struct node *node_dev = to_node(dev);
  30         const struct cpumask *mask = cpumask_of_node(node_dev->sysdev.id);
  31         int len;
  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         len = type?
  37                 cpulist_scnprintf(buf, PAGE_SIZE-2, mask) :
  38                 cpumask_scnprintf(buf, PAGE_SIZE-2, mask);
  39         buf[len++] = '\n';
  40         buf[len] = '\0';
  41         return len;
  42 }
  43
  44 static inline ssize_t node_read_cpumask(struct sys_device *dev,
  45                                 struct sysdev_attribute *attr, char *buf)
  46 {
  47         return node_read_cpumap(dev, 0, buf);
  48 }
  49 static inline ssize_t node_read_cpulist(struct sys_device *dev,
  50                                 struct sysdev_attribute *attr, char *buf)
  51 {
  52         return node_read_cpumap(dev, 1, buf);
  53 }
  54
  55 static SYSDEV_ATTR(cpumap,  S_IRUGO, node_read_cpumask, NULL);
  56 static SYSDEV_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
  57
  58 #define K(x) ((x) << (PAGE_SHIFT - 10))
  59 static ssize_t node_read_meminfo(struct sys_device * dev,
  60                         struct sysdev_attribute *attr, char * buf)
  61 {
  62         int n;
  63         int nid = dev->id;
  64         struct sysinfo i;
  65
  66         si_meminfo_node(&i, nid);
  67
  68         n = sprintf(buf, "\n"
  69                        "Node %d MemTotal:       %8lu kB\n"
  70                        "Node %d MemFree:        %8lu kB\n"
  71                        "Node %d MemUsed:        %8lu kB\n"
  72                        "Node %d Active:         %8lu kB\n"
  73                        "Node %d Inactive:       %8lu kB\n"
  74                        "Node %d Active(anon):   %8lu kB\n"
  75                        "Node %d Inactive(anon): %8lu kB\n"
  76                        "Node %d Active(file):   %8lu kB\n"
  77                        "Node %d Inactive(file): %8lu kB\n"
  78                        "Node %d Unevictable:    %8lu kB\n"
  79                        "Node %d Mlocked:        %8lu kB\n"
  80 #ifdef CONFIG_HIGHMEM
  81                        "Node %d HighTotal:      %8lu kB\n"
  82                        "Node %d HighFree:       %8lu kB\n"
  83                        "Node %d LowTotal:       %8lu kB\n"
  84                        "Node %d LowFree:        %8lu kB\n"
  85 #endif
  86                        "Node %d Dirty:          %8lu kB\n"
  87                        "Node %d Writeback:      %8lu kB\n"
  88                        "Node %d FilePages:      %8lu kB\n"
  89                        "Node %d Mapped:         %8lu kB\n"
  90                        "Node %d AnonPages:      %8lu kB\n"
  91                        "Node %d Shmem:          %8lu kB\n"
  92                        "Node %d KernelStack:    %8lu kB\n"
  93                        "Node %d PageTables:     %8lu kB\n"
  94                        "Node %d NFS_Unstable:   %8lu kB\n"
  95                        "Node %d Bounce:         %8lu kB\n"
  96                        "Node %d WritebackTmp:   %8lu kB\n"
  97                        "Node %d Slab:           %8lu kB\n"
  98                        "Node %d SReclaimable:   %8lu kB\n"
  99                        "Node %d SUnreclaim:     %8lu kB\n",
 100                        nid, K(i.totalram),
 101                        nid, K(i.freeram),
 102                        nid, K(i.totalram - i.freeram),
 103                        nid, K(node_page_state(nid, NR_ACTIVE_ANON) +
 104                                 node_page_state(nid, NR_ACTIVE_FILE)),
 105                        nid, K(node_page_state(nid, NR_INACTIVE_ANON) +
 106                                 node_page_state(nid, NR_INACTIVE_FILE)),
 107                        nid, K(node_page_state(nid, NR_ACTIVE_ANON)),
 108                        nid, K(node_page_state(nid, NR_INACTIVE_ANON)),
 109                        nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
 110                        nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
 111                        nid, K(node_page_state(nid, NR_UNEVICTABLE)),
 112                        nid, K(node_page_state(nid, NR_MLOCK)),
 113 #ifdef CONFIG_HIGHMEM
 114                        nid, K(i.totalhigh),
 115                        nid, K(i.freehigh),
 116                        nid, K(i.totalram - i.totalhigh),
 117                        nid, K(i.freeram - i.freehigh),
 118 #endif
 119                        nid, K(node_page_state(nid, NR_FILE_DIRTY)),
 120                        nid, K(node_page_state(nid, NR_WRITEBACK)),
 121                        nid, K(node_page_state(nid, NR_FILE_PAGES)),
 122                        nid, K(node_page_state(nid, NR_FILE_MAPPED)),
 123                        nid, K(node_page_state(nid, NR_ANON_PAGES)),
 124                        nid, K(node_page_state(nid, NR_SHMEM)),
 125                        nid, node_page_state(nid, NR_KERNEL_STACK) *
 126                                 THREAD_SIZE / 1024,
 127                        nid, K(node_page_state(nid, NR_PAGETABLE)),
 128                        nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
 129                        nid, K(node_page_state(nid, NR_BOUNCE)),
 130                        nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
 131                        nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
 132                                 node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
 133                        nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
 134                        nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
 135         n += hugetlb_report_node_meminfo(nid, buf + n);
 136         return n;
 137 }
 138
 139 #undef K
 140 static SYSDEV_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
 141
 142 static ssize_t node_read_numastat(struct sys_device * dev,
 143                                 struct sysdev_attribute *attr, char * buf)
 144 {
 145         return sprintf(buf,
 146                        "numa_hit %lu\n"
 147                        "numa_miss %lu\n"
 148                        "numa_foreign %lu\n"
 149                        "interleave_hit %lu\n"
 150                        "local_node %lu\n"
 151                        "other_node %lu\n",
 152                        node_page_state(dev->id, NUMA_HIT),
 153                        node_page_state(dev->id, NUMA_MISS),
 154                        node_page_state(dev->id, NUMA_FOREIGN),
 155                        node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
 156                        node_page_state(dev->id, NUMA_LOCAL),
 157                        node_page_state(dev->id, NUMA_OTHER));
 158 }
 159 static SYSDEV_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
 160
 161 static ssize_t node_read_distance(struct sys_device * dev,
 162                         struct sysdev_attribute *attr, char * buf)
 163 {
 164         int nid = dev->id;
 165         int len = 0;
 166         int i;
 167
 168         /* buf currently PAGE_SIZE, need ~4 chars per node */
 169         BUILD_BUG_ON(MAX_NUMNODES*4 > PAGE_SIZE/2);
 170
 171         for_each_online_node(i)
 172                 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
 173
 174         len += sprintf(buf + len, "\n");
 175         return len;
 176 }
 177 static SYSDEV_ATTR(distance, S_IRUGO, node_read_distance, NULL);
 178
 179 #ifdef CONFIG_HUGETLBFS
 180 /*
 181  * hugetlbfs per node attributes registration interface:
 182  * When/if hugetlb[fs] subsystem initializes [sometime after this module],
 183  * it will register its per node attributes for all online nodes with
 184  * memory.  It will also call register_hugetlbfs_with_node(), below, to
 185  * register its attribute registration functions with this node driver.
 186  * Once these hooks have been initialized, the node driver will call into
 187  * the hugetlb module to [un]register attributes for hot-plugged nodes.
 188  */
 189 static node_registration_func_t __hugetlb_register_node;
 190 static node_registration_func_t __hugetlb_unregister_node;
 191
 192 static inline bool hugetlb_register_node(struct node *node)
 193 {
 194         if (__hugetlb_register_node &&
 195                         node_state(node->sysdev.id, N_HIGH_MEMORY)) {
 196                 __hugetlb_register_node(node);
 197                 return true;
 198         }
 199         return false;
 200 }
 201
 202 static inline void hugetlb_unregister_node(struct node *node)
 203 {
 204         if (__hugetlb_unregister_node)
 205                 __hugetlb_unregister_node(node);
 206 }
 207
 208 void register_hugetlbfs_with_node(node_registration_func_t doregister,
 209                                   node_registration_func_t unregister)
 210 {
 211         __hugetlb_register_node   = doregister;
 212         __hugetlb_unregister_node = unregister;
 213 }
 214 #else
 215 static inline void hugetlb_register_node(struct node *node) {}
 216
 217 static inline void hugetlb_unregister_node(struct node *node) {}
 218 #endif
 219
 220
 221 /*
 222  * register_node - Setup a sysfs device for a node.
 223  * @num - Node number to use when creating the device.
 224  *
 225  * Initialize and register the node device.
 226  */
 227 int register_node(struct node *node, int num, struct node *parent)
 228 {
 229         int error;
 230
 231         node->sysdev.id = num;
 232         node->sysdev.cls = &node_class;
 233         error = sysdev_register(&node->sysdev);
 234
 235         if (!error){
 236                 sysdev_create_file(&node->sysdev, &attr_cpumap);
 237                 sysdev_create_file(&node->sysdev, &attr_cpulist);
 238                 sysdev_create_file(&node->sysdev, &attr_meminfo);
 239                 sysdev_create_file(&node->sysdev, &attr_numastat);
 240                 sysdev_create_file(&node->sysdev, &attr_distance);
 241
 242                 scan_unevictable_register_node(node);
 243
 244                 hugetlb_register_node(node);
 245         }
 246         return error;
 247 }
 248
 249 /**
 250  * unregister_node - unregister a node device
 251  * @node: node going away
 252  *
 253  * Unregisters a node device @node.  All the devices on the node must be
 254  * unregistered before calling this function.
 255  */
 256 void unregister_node(struct node *node)
 257 {
 258         sysdev_remove_file(&node->sysdev, &attr_cpumap);
 259         sysdev_remove_file(&node->sysdev, &attr_cpulist);
 260         sysdev_remove_file(&node->sysdev, &attr_meminfo);
 261         sysdev_remove_file(&node->sysdev, &attr_numastat);
 262         sysdev_remove_file(&node->sysdev, &attr_distance);
 263
 264         scan_unevictable_unregister_node(node);
 265         hugetlb_unregister_node(node);          /* no-op, if memoryless node */
 266
 267         sysdev_unregister(&node->sysdev);
 268 }
 269
 270 struct node node_devices[MAX_NUMNODES];
 271
 272 /*
 273  * register cpu under node
 274  */
 275 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
 276 {
 277         int ret;
 278         struct sys_device *obj;
 279
 280         if (!node_online(nid))
 281                 return 0;
 282
 283         obj = get_cpu_sysdev(cpu);
 284         if (!obj)
 285                 return 0;
 286
 287         ret = sysfs_create_link(&node_devices[nid].sysdev.kobj,
 288                                 &obj->kobj,
 289                                 kobject_name(&obj->kobj));
 290         if (ret)
 291                 return ret;
 292
 293         return sysfs_create_link(&obj->kobj,
 294                                  &node_devices[nid].sysdev.kobj,
 295                                  kobject_name(&node_devices[nid].sysdev.kobj));
 296 }
 297
 298 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
 299 {
 300         struct sys_device *obj;
 301
 302         if (!node_online(nid))
 303                 return 0;
 304
 305         obj = get_cpu_sysdev(cpu);
 306         if (!obj)
 307                 return 0;
 308
 309         sysfs_remove_link(&node_devices[nid].sysdev.kobj,
 310                           kobject_name(&obj->kobj));
 311         sysfs_remove_link(&obj->kobj,
 312                           kobject_name(&node_devices[nid].sysdev.kobj));
 313
 314         return 0;
 315 }
 316
 317 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
 318 #define page_initialized(page)  (page->lru.next)
 319
 320 static int get_nid_for_pfn(unsigned long pfn)
 321 {
 322         struct page *page;
 323
 324         if (!pfn_valid_within(pfn))
 325                 return -1;
 326         page = pfn_to_page(pfn);
 327         if (!page_initialized(page))
 328                 return -1;
 329         return pfn_to_nid(pfn);
 330 }
 331
 332 /* register memory section under specified node if it spans that node */
 333 int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
 334 {
 335         int ret;
 336         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 337
 338         if (!mem_blk)
 339                 return -EFAULT;
 340         if (!node_online(nid))
 341                 return 0;
 342         sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
 343         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
 344         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
 345                 int page_nid;
 346
 347                 page_nid = get_nid_for_pfn(pfn);
 348                 if (page_nid < 0)
 349                         continue;
 350                 if (page_nid != nid)
 351                         continue;
 352                 ret = sysfs_create_link_nowarn(&node_devices[nid].sysdev.kobj,
 353                                         &mem_blk->sysdev.kobj,
 354                                         kobject_name(&mem_blk->sysdev.kobj));
 355                 if (ret)
 356                         return ret;
 357
 358                 return sysfs_create_link_nowarn(&mem_blk->sysdev.kobj,
 359                                 &node_devices[nid].sysdev.kobj,
 360                                 kobject_name(&node_devices[nid].sysdev.kobj));
 361         }
 362         /* mem section does not span the specified node */
 363         return 0;
 364 }
 365
 366 /* unregister memory section under all nodes that it spans */
 367 int unregister_mem_sect_under_nodes(struct memory_block *mem_blk)
 368 {
 369         NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
 370         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 371
 372         if (!mem_blk) {
 373                 NODEMASK_FREE(unlinked_nodes);
 374                 return -EFAULT;
 375         }
 376         if (!unlinked_nodes)
 377                 return -ENOMEM;
 378         nodes_clear(*unlinked_nodes);
 379         sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
 380         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
 381         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
 382                 int nid;
 383
 384                 nid = get_nid_for_pfn(pfn);
 385                 if (nid < 0)
 386                         continue;
 387                 if (!node_online(nid))
 388                         continue;
 389                 if (node_test_and_set(nid, *unlinked_nodes))
 390                         continue;
 391                 sysfs_remove_link(&node_devices[nid].sysdev.kobj,
 392                          kobject_name(&mem_blk->sysdev.kobj));
 393                 sysfs_remove_link(&mem_blk->sysdev.kobj,
 394                          kobject_name(&node_devices[nid].sysdev.kobj));
 395         }
 396         NODEMASK_FREE(unlinked_nodes);
 397         return 0;
 398 }
 399
 400 static int link_mem_sections(int nid)
 401 {
 402         unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
 403         unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
 404         unsigned long pfn;
 405         int err = 0;
 406
 407         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 408                 unsigned long section_nr = pfn_to_section_nr(pfn);
 409                 struct mem_section *mem_sect;
 410                 struct memory_block *mem_blk;
 411                 int ret;
 412
 413                 if (!present_section_nr(section_nr))
 414                         continue;
 415                 mem_sect = __nr_to_section(section_nr);
 416                 mem_blk = find_memory_block(mem_sect);
 417                 ret = register_mem_sect_under_node(mem_blk, nid);
 418                 if (!err)
 419                         err = ret;
 420
 421                 /* discard ref obtained in find_memory_block() */
 422                 kobject_put(&mem_blk->sysdev.kobj);
 423         }
 424         return err;
 425 }
 426
 427 #ifdef CONFIG_HUGETLBFS
 428 /*
 429  * Handle per node hstate attribute [un]registration on transistions
 430  * to/from memoryless state.
 431  */
 432 static void node_hugetlb_work(struct work_struct *work)
 433 {
 434         struct node *node = container_of(work, struct node, node_work);
 435
 436         /*
 437          * We only get here when a node transitions to/from memoryless state.
 438          * We can detect which transition occurred by examining whether the
 439          * node has memory now.  hugetlb_register_node() already check this
 440          * so we try to register the attributes.  If that fails, then the
 441          * node has transitioned to memoryless, try to unregister the
 442          * attributes.
 443          */
 444         if (!hugetlb_register_node(node))
 445                 hugetlb_unregister_node(node);
 446 }
 447
 448 static void init_node_hugetlb_work(int nid)
 449 {
 450         INIT_WORK(&node_devices[nid].node_work, node_hugetlb_work);
 451 }
 452
 453 static int node_memory_callback(struct notifier_block *self,
 454                                 unsigned long action, void *arg)
 455 {
 456         struct memory_notify *mnb = arg;
 457         int nid = mnb->status_change_nid;
 458
 459         switch (action) {
 460         case MEM_ONLINE:
 461         case MEM_OFFLINE:
 462                 /*
 463                  * offload per node hstate [un]registration to a work thread
 464                  * when transitioning to/from memoryless state.
 465                  */
 466                 if (nid != NUMA_NO_NODE)
 467                         schedule_work(&node_devices[nid].node_work);
 468                 break;
 469
 470         case MEM_GOING_ONLINE:
 471         case MEM_GOING_OFFLINE:
 472         case MEM_CANCEL_ONLINE:
 473         case MEM_CANCEL_OFFLINE:
 474         default:
 475                 break;
 476         }
 477
 478         return NOTIFY_OK;
 479 }
 480 #endif  /* CONFIG_HUGETLBFS */
 481 #else   /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
 482
 483 static int link_mem_sections(int nid) { return 0; }
 484 #endif  /* CONFIG_MEMORY_HOTPLUG_SPARSE */
 485
 486 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
 487     !defined(CONFIG_HUGETLBFS)
 488 static inline int node_memory_callback(struct notifier_block *self,
 489                                 unsigned long action, void *arg)
 490 {
 491         return NOTIFY_OK;
 492 }
 493
 494 static void init_node_hugetlb_work(int nid) { }
 495
 496 #endif
 497
 498 int register_one_node(int nid)
 499 {
 500         int error = 0;
 501         int cpu;
 502
 503         if (node_online(nid)) {
 504                 int p_node = parent_node(nid);
 505                 struct node *parent = NULL;
 506
 507                 if (p_node != nid)
 508                         parent = &node_devices[p_node];
 509
 510                 error = register_node(&node_devices[nid], nid, parent);
 511
 512                 /* link cpu under this node */
 513                 for_each_present_cpu(cpu) {
 514                         if (cpu_to_node(cpu) == nid)
 515                                 register_cpu_under_node(cpu, nid);
 516                 }
 517
 518                 /* link memory sections under this node */
 519                 error = link_mem_sections(nid);
 520
 521                 /* initialize work queue for memory hot plug */
 522                 init_node_hugetlb_work(nid);
 523         }
 524
 525         return error;
 526
 527 }
 528
 529 void unregister_one_node(int nid)
 530 {
 531         unregister_node(&node_devices[nid]);
 532 }
 533
 534 /*
 535  * node states attributes
 536  */
 537
 538 static ssize_t print_nodes_state(enum node_states state, char *buf)
 539 {
 540         int n;
 541
 542         n = nodelist_scnprintf(buf, PAGE_SIZE, node_states[state]);
 543         if (n > 0 && PAGE_SIZE > n + 1) {
 544                 *(buf + n++) = '\n';
 545                 *(buf + n++) = '\0';
 546         }
 547         return n;
 548 }
 549
 550 struct node_attr {
 551         struct sysdev_class_attribute attr;
 552         enum node_states state;
 553 };
 554
 555 static ssize_t show_node_state(struct sysdev_class *class,
 556                                struct sysdev_class_attribute *attr, char *buf)
 557 {
 558         struct node_attr *na = container_of(attr, struct node_attr, attr);
 559         return print_nodes_state(na->state, buf);
 560 }
 561
 562 #define _NODE_ATTR(name, state) \
 563         { _SYSDEV_CLASS_ATTR(name, 0444, show_node_state, NULL), state }
 564
 565 static struct node_attr node_state_attr[] = {
 566         _NODE_ATTR(possible, N_POSSIBLE),
 567         _NODE_ATTR(online, N_ONLINE),
 568         _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
 569         _NODE_ATTR(has_cpu, N_CPU),
 570 #ifdef CONFIG_HIGHMEM
 571         _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
 572 #endif
 573 };
 574
 575 static struct sysdev_class_attribute *node_state_attrs[] = {
 576         &node_state_attr[0].attr,
 577         &node_state_attr[1].attr,
 578         &node_state_attr[2].attr,
 579         &node_state_attr[3].attr,
 580 #ifdef CONFIG_HIGHMEM
 581         &node_state_attr[4].attr,
 582 #endif
 583         NULL
 584 };
 585
 586 #define NODE_CALLBACK_PRI       2       /* lower than SLAB */
 587 static int __init register_node_type(void)
 588 {
 589         int ret;
 590
 591         BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
 592         BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
 593
 594         ret = sysdev_class_register(&node_class);
 595         if (!ret) {
 596                 hotplug_memory_notifier(node_memory_callback,
 597                                         NODE_CALLBACK_PRI);
 598         }
 599
 600         /*
 601          * Note:  we're not going to unregister the node class if we fail
 602          * to register the node state class attribute files.
 603          */
 604         return ret;
 605 }
 606 postcore_initcall(register_node_type);