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/compaction.h>
  13 #include <linux/cpumask.h>
  14 #include <linux/topology.h>
  15 #include <linux/nodemask.h>
  16 #include <linux/cpu.h>
  17 #include <linux/device.h>
  18 #include <linux/swap.h>
  19 #include <linux/slab.h>
  20
  21 static struct sysdev_class_attribute *node_state_attrs[];
  22
  23 static struct sysdev_class node_class = {
  24         .name = "node",
  25         .attrs = node_state_attrs,
  26 };
  27
  28
  29 static ssize_t node_read_cpumap(struct sys_device *dev, int type, char *buf)
  30 {
  31         struct node *node_dev = to_node(dev);
  32         const struct cpumask *mask = cpumask_of_node(node_dev->sysdev.id);
  33         int len;
  34
  35         /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
  36         BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1));
  37
  38         len = type?
  39                 cpulist_scnprintf(buf, PAGE_SIZE-2, mask) :
  40                 cpumask_scnprintf(buf, PAGE_SIZE-2, mask);
  41         buf[len++] = '\n';
  42         buf[len] = '\0';
  43         return len;
  44 }
  45
  46 static inline ssize_t node_read_cpumask(struct sys_device *dev,
  47                                 struct sysdev_attribute *attr, char *buf)
  48 {
  49         return node_read_cpumap(dev, 0, buf);
  50 }
  51 static inline ssize_t node_read_cpulist(struct sys_device *dev,
  52                                 struct sysdev_attribute *attr, char *buf)
  53 {
  54         return node_read_cpumap(dev, 1, buf);
  55 }
  56
  57 static SYSDEV_ATTR(cpumap,  S_IRUGO, node_read_cpumask, NULL);
  58 static SYSDEV_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
  59
  60 #define K(x) ((x) << (PAGE_SHIFT - 10))
  61 static ssize_t node_read_meminfo(struct sys_device * dev,
  62                         struct sysdev_attribute *attr, char * buf)
  63 {
  64         int n;
  65         int nid = dev->id;
  66         struct sysinfo i;
  67
  68         si_meminfo_node(&i, nid);
  69         n = sprintf(buf,
  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                        nid, K(i.totalram),
  82                        nid, K(i.freeram),
  83                        nid, K(i.totalram - i.freeram),
  84                        nid, K(node_page_state(nid, NR_ACTIVE_ANON) +
  85                                 node_page_state(nid, NR_ACTIVE_FILE)),
  86                        nid, K(node_page_state(nid, NR_INACTIVE_ANON) +
  87                                 node_page_state(nid, NR_INACTIVE_FILE)),
  88                        nid, K(node_page_state(nid, NR_ACTIVE_ANON)),
  89                        nid, K(node_page_state(nid, NR_INACTIVE_ANON)),
  90                        nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
  91                        nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
  92                        nid, K(node_page_state(nid, NR_UNEVICTABLE)),
  93                        nid, K(node_page_state(nid, NR_MLOCK)));
  94
  95 #ifdef CONFIG_HIGHMEM
  96         n += sprintf(buf + n,
  97                        "Node %d HighTotal:      %8lu kB\n"
  98                        "Node %d HighFree:       %8lu kB\n"
  99                        "Node %d LowTotal:       %8lu kB\n"
 100                        "Node %d LowFree:        %8lu kB\n",
 101                        nid, K(i.totalhigh),
 102                        nid, K(i.freehigh),
 103                        nid, K(i.totalram - i.totalhigh),
 104                        nid, K(i.freeram - i.freehigh));
 105 #endif
 106         n += sprintf(buf + n,
 107                        "Node %d Dirty:          %8lu kB\n"
 108                        "Node %d Writeback:      %8lu kB\n"
 109                        "Node %d FilePages:      %8lu kB\n"
 110                        "Node %d Mapped:         %8lu kB\n"
 111                        "Node %d AnonPages:      %8lu kB\n"
 112                        "Node %d Shmem:          %8lu kB\n"
 113                        "Node %d KernelStack:    %8lu kB\n"
 114                        "Node %d PageTables:     %8lu kB\n"
 115                        "Node %d NFS_Unstable:   %8lu kB\n"
 116                        "Node %d Bounce:         %8lu kB\n"
 117                        "Node %d WritebackTmp:   %8lu kB\n"
 118                        "Node %d Slab:           %8lu kB\n"
 119                        "Node %d SReclaimable:   %8lu kB\n"
 120                        "Node %d SUnreclaim:     %8lu kB\n",
 121                        nid, K(node_page_state(nid, NR_FILE_DIRTY)),
 122                        nid, K(node_page_state(nid, NR_WRITEBACK)),
 123                        nid, K(node_page_state(nid, NR_FILE_PAGES)),
 124                        nid, K(node_page_state(nid, NR_FILE_MAPPED)),
 125                        nid, K(node_page_state(nid, NR_ANON_PAGES)),
 126                        nid, K(node_page_state(nid, NR_SHMEM)),
 127                        nid, node_page_state(nid, NR_KERNEL_STACK) *
 128                                 THREAD_SIZE / 1024,
 129                        nid, K(node_page_state(nid, NR_PAGETABLE)),
 130                        nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
 131                        nid, K(node_page_state(nid, NR_BOUNCE)),
 132                        nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)),
 133                        nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
 134                                 node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
 135                        nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
 136                        nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
 137         n += hugetlb_report_node_meminfo(nid, buf + n);
 138         return n;
 139 }
 140
 141 #undef K
 142 static SYSDEV_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
 143
 144 static ssize_t node_read_numastat(struct sys_device * dev,
 145                                 struct sysdev_attribute *attr, char * buf)
 146 {
 147         return sprintf(buf,
 148                        "numa_hit %lu\n"
 149                        "numa_miss %lu\n"
 150                        "numa_foreign %lu\n"
 151                        "interleave_hit %lu\n"
 152                        "local_node %lu\n"
 153                        "other_node %lu\n",
 154                        node_page_state(dev->id, NUMA_HIT),
 155                        node_page_state(dev->id, NUMA_MISS),
 156                        node_page_state(dev->id, NUMA_FOREIGN),
 157                        node_page_state(dev->id, NUMA_INTERLEAVE_HIT),
 158                        node_page_state(dev->id, NUMA_LOCAL),
 159                        node_page_state(dev->id, NUMA_OTHER));
 160 }
 161 static SYSDEV_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
 162
 163 static ssize_t node_read_distance(struct sys_device * dev,
 164                         struct sysdev_attribute *attr, char * buf)
 165 {
 166         int nid = dev->id;
 167         int len = 0;
 168         int i;
 169
 170         /*
 171          * buf is currently PAGE_SIZE in length and each node needs 4 chars
 172          * at the most (distance + space or newline).
 173          */
 174         BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
 175
 176         for_each_online_node(i)
 177                 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
 178
 179         len += sprintf(buf + len, "\n");
 180         return len;
 181 }
 182 static SYSDEV_ATTR(distance, S_IRUGO, node_read_distance, NULL);
 183
 184 #ifdef CONFIG_HUGETLBFS
 185 /*
 186  * hugetlbfs per node attributes registration interface:
 187  * When/if hugetlb[fs] subsystem initializes [sometime after this module],
 188  * it will register its per node attributes for all online nodes with
 189  * memory.  It will also call register_hugetlbfs_with_node(), below, to
 190  * register its attribute registration functions with this node driver.
 191  * Once these hooks have been initialized, the node driver will call into
 192  * the hugetlb module to [un]register attributes for hot-plugged nodes.
 193  */
 194 static node_registration_func_t __hugetlb_register_node;
 195 static node_registration_func_t __hugetlb_unregister_node;
 196
 197 static inline bool hugetlb_register_node(struct node *node)
 198 {
 199         if (__hugetlb_register_node &&
 200                         node_state(node->sysdev.id, N_HIGH_MEMORY)) {
 201                 __hugetlb_register_node(node);
 202                 return true;
 203         }
 204         return false;
 205 }
 206
 207 static inline void hugetlb_unregister_node(struct node *node)
 208 {
 209         if (__hugetlb_unregister_node)
 210                 __hugetlb_unregister_node(node);
 211 }
 212
 213 void register_hugetlbfs_with_node(node_registration_func_t doregister,
 214                                   node_registration_func_t unregister)
 215 {
 216         __hugetlb_register_node   = doregister;
 217         __hugetlb_unregister_node = unregister;
 218 }
 219 #else
 220 static inline void hugetlb_register_node(struct node *node) {}
 221
 222 static inline void hugetlb_unregister_node(struct node *node) {}
 223 #endif
 224
 225
 226 /*
 227  * register_node - Setup a sysfs device for a node.
 228  * @num - Node number to use when creating the device.
 229  *
 230  * Initialize and register the node device.
 231  */
 232 int register_node(struct node *node, int num, struct node *parent)
 233 {
 234         int error;
 235
 236         node->sysdev.id = num;
 237         node->sysdev.cls = &node_class;
 238         error = sysdev_register(&node->sysdev);
 239
 240         if (!error){
 241                 sysdev_create_file(&node->sysdev, &attr_cpumap);
 242                 sysdev_create_file(&node->sysdev, &attr_cpulist);
 243                 sysdev_create_file(&node->sysdev, &attr_meminfo);
 244                 sysdev_create_file(&node->sysdev, &attr_numastat);
 245                 sysdev_create_file(&node->sysdev, &attr_distance);
 246
 247                 scan_unevictable_register_node(node);
 248
 249                 hugetlb_register_node(node);
 250
 251                 compaction_register_node(node);
 252         }
 253         return error;
 254 }
 255
 256 /**
 257  * unregister_node - unregister a node device
 258  * @node: node going away
 259  *
 260  * Unregisters a node device @node.  All the devices on the node must be
 261  * unregistered before calling this function.
 262  */
 263 void unregister_node(struct node *node)
 264 {
 265         sysdev_remove_file(&node->sysdev, &attr_cpumap);
 266         sysdev_remove_file(&node->sysdev, &attr_cpulist);
 267         sysdev_remove_file(&node->sysdev, &attr_meminfo);
 268         sysdev_remove_file(&node->sysdev, &attr_numastat);
 269         sysdev_remove_file(&node->sysdev, &attr_distance);
 270
 271         scan_unevictable_unregister_node(node);
 272         hugetlb_unregister_node(node);          /* no-op, if memoryless node */
 273
 274         sysdev_unregister(&node->sysdev);
 275 }
 276
 277 struct node node_devices[MAX_NUMNODES];
 278
 279 /*
 280  * register cpu under node
 281  */
 282 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
 283 {
 284         int ret;
 285         struct sys_device *obj;
 286
 287         if (!node_online(nid))
 288                 return 0;
 289
 290         obj = get_cpu_sysdev(cpu);
 291         if (!obj)
 292                 return 0;
 293
 294         ret = sysfs_create_link(&node_devices[nid].sysdev.kobj,
 295                                 &obj->kobj,
 296                                 kobject_name(&obj->kobj));
 297         if (ret)
 298                 return ret;
 299
 300         return sysfs_create_link(&obj->kobj,
 301                                  &node_devices[nid].sysdev.kobj,
 302                                  kobject_name(&node_devices[nid].sysdev.kobj));
 303 }
 304
 305 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
 306 {
 307         struct sys_device *obj;
 308
 309         if (!node_online(nid))
 310                 return 0;
 311
 312         obj = get_cpu_sysdev(cpu);
 313         if (!obj)
 314                 return 0;
 315
 316         sysfs_remove_link(&node_devices[nid].sysdev.kobj,
 317                           kobject_name(&obj->kobj));
 318         sysfs_remove_link(&obj->kobj,
 319                           kobject_name(&node_devices[nid].sysdev.kobj));
 320
 321         return 0;
 322 }
 323
 324 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
 325 #define page_initialized(page)  (page->lru.next)
 326
 327 static int get_nid_for_pfn(unsigned long pfn)
 328 {
 329         struct page *page;
 330
 331         if (!pfn_valid_within(pfn))
 332                 return -1;
 333         page = pfn_to_page(pfn);
 334         if (!page_initialized(page))
 335                 return -1;
 336         return pfn_to_nid(pfn);
 337 }
 338
 339 /* register memory section under specified node if it spans that node */
 340 int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
 341 {
 342         int ret;
 343         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 344
 345         if (!mem_blk)
 346                 return -EFAULT;
 347         if (!node_online(nid))
 348                 return 0;
 349         sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
 350         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
 351         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
 352                 int page_nid;
 353
 354                 page_nid = get_nid_for_pfn(pfn);
 355                 if (page_nid < 0)
 356                         continue;
 357                 if (page_nid != nid)
 358                         continue;
 359                 ret = sysfs_create_link_nowarn(&node_devices[nid].sysdev.kobj,
 360                                         &mem_blk->sysdev.kobj,
 361                                         kobject_name(&mem_blk->sysdev.kobj));
 362                 if (ret)
 363                         return ret;
 364
 365                 return sysfs_create_link_nowarn(&mem_blk->sysdev.kobj,
 366                                 &node_devices[nid].sysdev.kobj,
 367                                 kobject_name(&node_devices[nid].sysdev.kobj));
 368         }
 369         /* mem section does not span the specified node */
 370         return 0;
 371 }
 372
 373 /* unregister memory section under all nodes that it spans */
 374 int unregister_mem_sect_under_nodes(struct memory_block *mem_blk)
 375 {
 376         NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
 377         unsigned long pfn, sect_start_pfn, sect_end_pfn;
 378
 379         if (!mem_blk) {
 380                 NODEMASK_FREE(unlinked_nodes);
 381                 return -EFAULT;
 382         }
 383         if (!unlinked_nodes)
 384                 return -ENOMEM;
 385         nodes_clear(*unlinked_nodes);
 386         sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
 387         sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
 388         for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
 389                 int nid;
 390
 391                 nid = get_nid_for_pfn(pfn);
 392                 if (nid < 0)
 393                         continue;
 394                 if (!node_online(nid))
 395                         continue;
 396                 if (node_test_and_set(nid, *unlinked_nodes))
 397                         continue;
 398                 sysfs_remove_link(&node_devices[nid].sysdev.kobj,
 399                          kobject_name(&mem_blk->sysdev.kobj));
 400                 sysfs_remove_link(&mem_blk->sysdev.kobj,
 401                          kobject_name(&node_devices[nid].sysdev.kobj));
 402         }
 403         NODEMASK_FREE(unlinked_nodes);
 404         return 0;
 405 }
 406
 407 static int link_mem_sections(int nid)
 408 {
 409         unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
 410         unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
 411         unsigned long pfn;
 412         int err = 0;
 413
 414         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 415                 unsigned long section_nr = pfn_to_section_nr(pfn);
 416                 struct mem_section *mem_sect;
 417                 struct memory_block *mem_blk;
 418                 int ret;
 419
 420                 if (!present_section_nr(section_nr))
 421                         continue;
 422                 mem_sect = __nr_to_section(section_nr);
 423                 mem_blk = find_memory_block(mem_sect);
 424                 ret = register_mem_sect_under_node(mem_blk, nid);
 425                 if (!err)
 426                         err = ret;
 427
 428                 /* discard ref obtained in find_memory_block() */
 429                 kobject_put(&mem_blk->sysdev.kobj);
 430         }
 431         return err;
 432 }
 433
 434 #ifdef CONFIG_HUGETLBFS
 435 /*
 436  * Handle per node hstate attribute [un]registration on transistions
 437  * to/from memoryless state.
 438  */
 439 static void node_hugetlb_work(struct work_struct *work)
 440 {
 441         struct node *node = container_of(work, struct node, node_work);
 442
 443         /*
 444          * We only get here when a node transitions to/from memoryless state.
 445          * We can detect which transition occurred by examining whether the
 446          * node has memory now.  hugetlb_register_node() already check this
 447          * so we try to register the attributes.  If that fails, then the
 448          * node has transitioned to memoryless, try to unregister the
 449          * attributes.
 450          */
 451         if (!hugetlb_register_node(node))
 452                 hugetlb_unregister_node(node);
 453 }
 454
 455 static void init_node_hugetlb_work(int nid)
 456 {
 457         INIT_WORK(&node_devices[nid].node_work, node_hugetlb_work);
 458 }
 459
 460 static int node_memory_callback(struct notifier_block *self,
 461                                 unsigned long action, void *arg)
 462 {
 463         struct memory_notify *mnb = arg;
 464         int nid = mnb->status_change_nid;
 465
 466         switch (action) {
 467         case MEM_ONLINE:
 468         case MEM_OFFLINE:
 469                 /*
 470                  * offload per node hstate [un]registration to a work thread
 471                  * when transitioning to/from memoryless state.
 472                  */
 473                 if (nid != NUMA_NO_NODE)
 474                         schedule_work(&node_devices[nid].node_work);
 475                 break;
 476
 477         case MEM_GOING_ONLINE:
 478         case MEM_GOING_OFFLINE:
 479         case MEM_CANCEL_ONLINE:
 480         case MEM_CANCEL_OFFLINE:
 481         default:
 482                 break;
 483         }
 484
 485         return NOTIFY_OK;
 486 }
 487 #endif  /* CONFIG_HUGETLBFS */
 488 #else   /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
 489
 490 static int link_mem_sections(int nid) { return 0; }
 491 #endif  /* CONFIG_MEMORY_HOTPLUG_SPARSE */
 492
 493 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
 494     !defined(CONFIG_HUGETLBFS)
 495 static inline int node_memory_callback(struct notifier_block *self,
 496                                 unsigned long action, void *arg)
 497 {
 498         return NOTIFY_OK;
 499 }
 500
 501 static void init_node_hugetlb_work(int nid) { }
 502
 503 #endif
 504
 505 int register_one_node(int nid)
 506 {
 507         int error = 0;
 508         int cpu;
 509
 510         if (node_online(nid)) {
 511                 int p_node = parent_node(nid);
 512                 struct node *parent = NULL;
 513
 514                 if (p_node != nid)
 515                         parent = &node_devices[p_node];
 516
 517                 error = register_node(&node_devices[nid], nid, parent);
 518
 519                 /* link cpu under this node */
 520                 for_each_present_cpu(cpu) {
 521                         if (cpu_to_node(cpu) == nid)
 522                                 register_cpu_under_node(cpu, nid);
 523                 }
 524
 525                 /* link memory sections under this node */
 526                 error = link_mem_sections(nid);
 527
 528                 /* initialize work queue for memory hot plug */
 529                 init_node_hugetlb_work(nid);
 530         }
 531
 532         return error;
 533
 534 }
 535
 536 void unregister_one_node(int nid)
 537 {
 538         unregister_node(&node_devices[nid]);
 539 }
 540
 541 /*
 542  * node states attributes
 543  */
 544
 545 static ssize_t print_nodes_state(enum node_states state, char *buf)
 546 {
 547         int n;
 548
 549         n = nodelist_scnprintf(buf, PAGE_SIZE, node_states[state]);
 550         if (n > 0 && PAGE_SIZE > n + 1) {
 551                 *(buf + n++) = '\n';
 552                 *(buf + n++) = '\0';
 553         }
 554         return n;
 555 }
 556
 557 struct node_attr {
 558         struct sysdev_class_attribute attr;
 559         enum node_states state;
 560 };
 561
 562 static ssize_t show_node_state(struct sysdev_class *class,
 563                                struct sysdev_class_attribute *attr, char *buf)
 564 {
 565         struct node_attr *na = container_of(attr, struct node_attr, attr);
 566         return print_nodes_state(na->state, buf);
 567 }
 568
 569 #define _NODE_ATTR(name, state) \
 570         { _SYSDEV_CLASS_ATTR(name, 0444, show_node_state, NULL), state }
 571
 572 static struct node_attr node_state_attr[] = {
 573         _NODE_ATTR(possible, N_POSSIBLE),
 574         _NODE_ATTR(online, N_ONLINE),
 575         _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
 576         _NODE_ATTR(has_cpu, N_CPU),
 577 #ifdef CONFIG_HIGHMEM
 578         _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
 579 #endif
 580 };
 581
 582 static struct sysdev_class_attribute *node_state_attrs[] = {
 583         &node_state_attr[0].attr,
 584         &node_state_attr[1].attr,
 585         &node_state_attr[2].attr,
 586         &node_state_attr[3].attr,
 587 #ifdef CONFIG_HIGHMEM
 588         &node_state_attr[4].attr,
 589 #endif
 590         NULL
 591 };
 592
 593 #define NODE_CALLBACK_PRI       2       /* lower than SLAB */
 594 static int __init register_node_type(void)
 595 {
 596         int ret;
 597
 598         BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
 599         BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
 600
 601         ret = sysdev_class_register(&node_class);
 602         if (!ret) {
 603                 hotplug_memory_notifier(node_memory_callback,
 604                                         NODE_CALLBACK_PRI);
 605         }
 606
 607         /*
 608          * Note:  we're not going to unregister the node class if we fail
 609          * to register the node state class attribute files.
 610          */
 611         return ret;
 612 }
 613 postcore_initcall(register_node_type);