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