drivers/base/node.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Basic Node interface support
   4  */
   5
   6 #include <linux/module.h>
   7 #include <linux/init.h>
   8 #include <linux/mm.h>
   9 #include <linux/memory.h>
  10 #include <linux/vmstat.h>
  11 #include <linux/notifier.h>
  12 #include <linux/node.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/compaction.h>
  15 #include <linux/cpumask.h>
  16 #include <linux/topology.h>
  17 #include <linux/nodemask.h>
  18 #include <linux/cpu.h>
  19 #include <linux/device.h>
  20 #include <linux/pm_runtime.h>
  21 #include <linux/swap.h>
  22 #include <linux/slab.h>
  23
  24 static struct bus_type node_subsys = {
  25         .name = "node",
  26         .dev_name = "node",
  27 };
  28
  29
  30 static ssize_t node_read_cpumap(struct device *dev, bool list, char *buf)
  31 {
  32         ssize_t n;
  33         cpumask_var_t mask;
  34         struct node *node_dev = to_node(dev);
  35
  36         /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
  37         BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1));
  38
  39         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
  40                 return 0;
  41
  42         cpumask_and(mask, cpumask_of_node(node_dev->dev.id), cpu_online_mask);
  43         n = cpumap_print_to_pagebuf(list, buf, mask);
  44         free_cpumask_var(mask);
  45
  46         return n;
  47 }
  48
  49 static inline ssize_t node_read_cpumask(struct device *dev,
  50                                 struct device_attribute *attr, char *buf)
  51 {
  52         return node_read_cpumap(dev, false, buf);
  53 }
  54 static inline ssize_t node_read_cpulist(struct device *dev,
  55                                 struct device_attribute *attr, char *buf)
  56 {
  57         return node_read_cpumap(dev, true, buf);
  58 }
  59
  60 static DEVICE_ATTR(cpumap,  S_IRUGO, node_read_cpumask, NULL);
  61 static DEVICE_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
  62
  63 /**
  64  * struct node_access_nodes - Access class device to hold user visible
  65  *                            relationships to other nodes.
  66  * @dev:        Device for this memory access class
  67  * @list_node:  List element in the node's access list
  68  * @access:     The access class rank
  69  * @hmem_attrs: Heterogeneous memory performance attributes
  70  */
  71 struct node_access_nodes {
  72         struct device           dev;
  73         struct list_head        list_node;
  74         unsigned                access;
  75 #ifdef CONFIG_HMEM_REPORTING
  76         struct node_hmem_attrs  hmem_attrs;
  77 #endif
  78 };
  79 #define to_access_nodes(dev) container_of(dev, struct node_access_nodes, dev)
  80
  81 static struct attribute *node_init_access_node_attrs[] = {
  82         NULL,
  83 };
  84
  85 static struct attribute *node_targ_access_node_attrs[] = {
  86         NULL,
  87 };
  88
  89 static const struct attribute_group initiators = {
  90         .name   = "initiators",
  91         .attrs  = node_init_access_node_attrs,
  92 };
  93
  94 static const struct attribute_group targets = {
  95         .name   = "targets",
  96         .attrs  = node_targ_access_node_attrs,
  97 };
  98
  99 static const struct attribute_group *node_access_node_groups[] = {
 100         &initiators,
 101         &targets,
 102         NULL,
 103 };
 104
 105 static void node_remove_accesses(struct node *node)
 106 {
 107         struct node_access_nodes *c, *cnext;
 108
 109         list_for_each_entry_safe(c, cnext, &node->access_list, list_node) {
 110                 list_del(&c->list_node);
 111                 device_unregister(&c->dev);
 112         }
 113 }
 114
 115 static void node_access_release(struct device *dev)
 116 {
 117         kfree(to_access_nodes(dev));
 118 }
 119
 120 static struct node_access_nodes *node_init_node_access(struct node *node,
 121                                                        unsigned access)
 122 {
 123         struct node_access_nodes *access_node;
 124         struct device *dev;
 125
 126         list_for_each_entry(access_node, &node->access_list, list_node)
 127                 if (access_node->access == access)
 128                         return access_node;
 129
 130         access_node = kzalloc(sizeof(*access_node), GFP_KERNEL);
 131         if (!access_node)
 132                 return NULL;
 133
 134         access_node->access = access;
 135         dev = &access_node->dev;
 136         dev->parent = &node->dev;
 137         dev->release = node_access_release;
 138         dev->groups = node_access_node_groups;
 139         if (dev_set_name(dev, "access%u", access))
 140                 goto free;
 141
 142         if (device_register(dev))
 143                 goto free_name;
 144
 145         pm_runtime_no_callbacks(dev);
 146         list_add_tail(&access_node->list_node, &node->access_list);
 147         return access_node;
 148 free_name:
 149         kfree_const(dev->kobj.name);
 150 free:
 151         kfree(access_node);
 152         return NULL;
 153 }
 154
 155 #ifdef CONFIG_HMEM_REPORTING
 156 #define ACCESS_ATTR(name)                                                  \
 157 static ssize_t name##_show(struct device *dev,                             \
 158                            struct device_attribute *attr,                  \
 159                            char *buf)                                      \
 160 {                                                                          \
 161         return sprintf(buf, "%u\n", to_access_nodes(dev)->hmem_attrs.name); \
 162 }                                                                          \
 163 static DEVICE_ATTR_RO(name);
 164
 165 ACCESS_ATTR(read_bandwidth)
 166 ACCESS_ATTR(read_latency)
 167 ACCESS_ATTR(write_bandwidth)
 168 ACCESS_ATTR(write_latency)
 169
 170 static struct attribute *access_attrs[] = {
 171         &dev_attr_read_bandwidth.attr,
 172         &dev_attr_read_latency.attr,
 173         &dev_attr_write_bandwidth.attr,
 174         &dev_attr_write_latency.attr,
 175         NULL,
 176 };
 177
 178 /**
 179  * node_set_perf_attrs - Set the performance values for given access class
 180  * @nid: Node identifier to be set
 181  * @hmem_attrs: Heterogeneous memory performance attributes
 182  * @access: The access class the for the given attributes
 183  */
 184 void node_set_perf_attrs(unsigned int nid, struct node_hmem_attrs *hmem_attrs,
 185                          unsigned access)
 186 {
 187         struct node_access_nodes *c;
 188         struct node *node;
 189         int i;
 190
 191         if (WARN_ON_ONCE(!node_online(nid)))
 192                 return;
 193
 194         node = node_devices[nid];
 195         c = node_init_node_access(node, access);
 196         if (!c)
 197                 return;
 198
 199         c->hmem_attrs = *hmem_attrs;
 200         for (i = 0; access_attrs[i] != NULL; i++) {
 201                 if (sysfs_add_file_to_group(&c->dev.kobj, access_attrs[i],
 202                                             "initiators")) {
 203                         pr_info("failed to add performance attribute to node %d\n",
 204                                 nid);
 205                         break;
 206                 }
 207         }
 208 }
 209
 210 /**
 211  * struct node_cache_info - Internal tracking for memory node caches
 212  * @dev:        Device represeting the cache level
 213  * @node:       List element for tracking in the node
 214  * @cache_attrs:Attributes for this cache level
 215  */
 216 struct node_cache_info {
 217         struct device dev;
 218         struct list_head node;
 219         struct node_cache_attrs cache_attrs;
 220 };
 221 #define to_cache_info(device) container_of(device, struct node_cache_info, dev)
 222
 223 #define CACHE_ATTR(name, fmt)                                           \
 224 static ssize_t name##_show(struct device *dev,                          \
 225                            struct device_attribute *attr,               \
 226                            char *buf)                                   \
 227 {                                                                       \
 228         return sprintf(buf, fmt "\n", to_cache_info(dev)->cache_attrs.name);\
 229 }                                                                       \
 230 DEVICE_ATTR_RO(name);
 231
 232 CACHE_ATTR(size, "%llu")
 233 CACHE_ATTR(line_size, "%u")
 234 CACHE_ATTR(indexing, "%u")
 235 CACHE_ATTR(write_policy, "%u")
 236
 237 static struct attribute *cache_attrs[] = {
 238         &dev_attr_indexing.attr,
 239         &dev_attr_size.attr,
 240         &dev_attr_line_size.attr,
 241         &dev_attr_write_policy.attr,
 242         NULL,
 243 };
 244 ATTRIBUTE_GROUPS(cache);
 245
 246 static void node_cache_release(struct device *dev)
 247 {
 248         kfree(dev);
 249 }
 250
 251 static void node_cacheinfo_release(struct device *dev)
 252 {
 253         struct node_cache_info *info = to_cache_info(dev);
 254         kfree(info);
 255 }
 256
 257 static void node_init_cache_dev(struct node *node)
 258 {
 259         struct device *dev;
 260
 261         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
 262         if (!dev)
 263                 return;
 264
 265         dev->parent = &node->dev;
 266         dev->release = node_cache_release;
 267         if (dev_set_name(dev, "memory_side_cache"))
 268                 goto free_dev;
 269
 270         if (device_register(dev))
 271                 goto free_name;
 272
 273         pm_runtime_no_callbacks(dev);
 274         node->cache_dev = dev;
 275         return;
 276 free_name:
 277         kfree_const(dev->kobj.name);
 278 free_dev:
 279         kfree(dev);
 280 }
 281
 282 /**
 283  * node_add_cache() - add cache attribute to a memory node
 284  * @nid: Node identifier that has new cache attributes
 285  * @cache_attrs: Attributes for the cache being added
 286  */
 287 void node_add_cache(unsigned int nid, struct node_cache_attrs *cache_attrs)
 288 {
 289         struct node_cache_info *info;
 290         struct device *dev;
 291         struct node *node;
 292
 293         if (!node_online(nid) || !node_devices[nid])
 294                 return;
 295
 296         node = node_devices[nid];
 297         list_for_each_entry(info, &node->cache_attrs, node) {
 298                 if (info->cache_attrs.level == cache_attrs->level) {
 299                         dev_warn(&node->dev,
 300                                 "attempt to add duplicate cache level:%d\n",
 301                                 cache_attrs->level);
 302                         return;
 303                 }
 304         }
 305
 306         if (!node->cache_dev)
 307                 node_init_cache_dev(node);
 308         if (!node->cache_dev)
 309                 return;
 310
 311         info = kzalloc(sizeof(*info), GFP_KERNEL);
 312         if (!info)
 313                 return;
 314
 315         dev = &info->dev;
 316         dev->parent = node->cache_dev;
 317         dev->release = node_cacheinfo_release;
 318         dev->groups = cache_groups;
 319         if (dev_set_name(dev, "index%d", cache_attrs->level))
 320                 goto free_cache;
 321
 322         info->cache_attrs = *cache_attrs;
 323         if (device_register(dev)) {
 324                 dev_warn(&node->dev, "failed to add cache level:%d\n",
 325                          cache_attrs->level);
 326                 goto free_name;
 327         }
 328         pm_runtime_no_callbacks(dev);
 329         list_add_tail(&info->node, &node->cache_attrs);
 330         return;
 331 free_name:
 332         kfree_const(dev->kobj.name);
 333 free_cache:
 334         kfree(info);
 335 }
 336
 337 static void node_remove_caches(struct node *node)
 338 {
 339         struct node_cache_info *info, *next;
 340
 341         if (!node->cache_dev)
 342                 return;
 343
 344         list_for_each_entry_safe(info, next, &node->cache_attrs, node) {
 345                 list_del(&info->node);
 346                 device_unregister(&info->dev);
 347         }
 348         device_unregister(node->cache_dev);
 349 }
 350
 351 static void node_init_caches(unsigned int nid)
 352 {
 353         INIT_LIST_HEAD(&node_devices[nid]->cache_attrs);
 354 }
 355 #else
 356 static void node_init_caches(unsigned int nid) { }
 357 static void node_remove_caches(struct node *node) { }
 358 #endif
 359
 360 #define K(x) ((x) << (PAGE_SHIFT - 10))
 361 static ssize_t node_read_meminfo(struct device *dev,
 362                         struct device_attribute *attr, char *buf)
 363 {
 364         int n;
 365         int nid = dev->id;
 366         struct pglist_data *pgdat = NODE_DATA(nid);
 367         struct sysinfo i;
 368         unsigned long sreclaimable, sunreclaimable;
 369
 370         si_meminfo_node(&i, nid);
 371         sreclaimable = node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B);
 372         sunreclaimable = node_page_state_pages(pgdat, NR_SLAB_UNRECLAIMABLE_B);
 373         n = sprintf(buf,
 374                        "Node %d MemTotal:       %8lu kB\n"
 375                        "Node %d MemFree:        %8lu kB\n"
 376                        "Node %d MemUsed:        %8lu kB\n"
 377                        "Node %d Active:         %8lu kB\n"
 378                        "Node %d Inactive:       %8lu kB\n"
 379                        "Node %d Active(anon):   %8lu kB\n"
 380                        "Node %d Inactive(anon): %8lu kB\n"
 381                        "Node %d Active(file):   %8lu kB\n"
 382                        "Node %d Inactive(file): %8lu kB\n"
 383                        "Node %d Unevictable:    %8lu kB\n"
 384                        "Node %d Mlocked:        %8lu kB\n",
 385                        nid, K(i.totalram),
 386                        nid, K(i.freeram),
 387                        nid, K(i.totalram - i.freeram),
 388                        nid, K(node_page_state(pgdat, NR_ACTIVE_ANON) +
 389                                 node_page_state(pgdat, NR_ACTIVE_FILE)),
 390                        nid, K(node_page_state(pgdat, NR_INACTIVE_ANON) +
 391                                 node_page_state(pgdat, NR_INACTIVE_FILE)),
 392                        nid, K(node_page_state(pgdat, NR_ACTIVE_ANON)),
 393                        nid, K(node_page_state(pgdat, NR_INACTIVE_ANON)),
 394                        nid, K(node_page_state(pgdat, NR_ACTIVE_FILE)),
 395                        nid, K(node_page_state(pgdat, NR_INACTIVE_FILE)),
 396                        nid, K(node_page_state(pgdat, NR_UNEVICTABLE)),
 397                        nid, K(sum_zone_node_page_state(nid, NR_MLOCK)));
 398
 399 #ifdef CONFIG_HIGHMEM
 400         n += sprintf(buf + n,
 401                        "Node %d HighTotal:      %8lu kB\n"
 402                        "Node %d HighFree:       %8lu kB\n"
 403                        "Node %d LowTotal:       %8lu kB\n"
 404                        "Node %d LowFree:        %8lu kB\n",
 405                        nid, K(i.totalhigh),
 406                        nid, K(i.freehigh),
 407                        nid, K(i.totalram - i.totalhigh),
 408                        nid, K(i.freeram - i.freehigh));
 409 #endif
 410         n += sprintf(buf + n,
 411                        "Node %d Dirty:          %8lu kB\n"
 412                        "Node %d Writeback:      %8lu kB\n"
 413                        "Node %d FilePages:      %8lu kB\n"
 414                        "Node %d Mapped:         %8lu kB\n"
 415                        "Node %d AnonPages:      %8lu kB\n"
 416                        "Node %d Shmem:          %8lu kB\n"
 417                        "Node %d KernelStack:    %8lu kB\n"
 418 #ifdef CONFIG_SHADOW_CALL_STACK
 419                        "Node %d ShadowCallStack:%8lu kB\n"
 420 #endif
 421                        "Node %d PageTables:     %8lu kB\n"
 422                        "Node %d NFS_Unstable:   %8lu kB\n"
 423                        "Node %d Bounce:         %8lu kB\n"
 424                        "Node %d WritebackTmp:   %8lu kB\n"
 425                        "Node %d KReclaimable:   %8lu kB\n"
 426                        "Node %d Slab:           %8lu kB\n"
 427                        "Node %d SReclaimable:   %8lu kB\n"
 428                        "Node %d SUnreclaim:     %8lu kB\n"
 429 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 430                        "Node %d AnonHugePages:  %8lu kB\n"
 431                        "Node %d ShmemHugePages: %8lu kB\n"
 432                        "Node %d ShmemPmdMapped: %8lu kB\n"
 433                        "Node %d FileHugePages: %8lu kB\n"
 434                        "Node %d FilePmdMapped: %8lu kB\n"
 435 #endif
 436                         ,
 437                        nid, K(node_page_state(pgdat, NR_FILE_DIRTY)),
 438                        nid, K(node_page_state(pgdat, NR_WRITEBACK)),
 439                        nid, K(node_page_state(pgdat, NR_FILE_PAGES)),
 440                        nid, K(node_page_state(pgdat, NR_FILE_MAPPED)),
 441                        nid, K(node_page_state(pgdat, NR_ANON_MAPPED)),
 442                        nid, K(i.sharedram),
 443                        nid, node_page_state(pgdat, NR_KERNEL_STACK_KB),
 444 #ifdef CONFIG_SHADOW_CALL_STACK
 445                        nid, node_page_state(pgdat, NR_KERNEL_SCS_KB),
 446 #endif
 447                        nid, K(sum_zone_node_page_state(nid, NR_PAGETABLE)),
 448                        nid, 0UL,
 449                        nid, K(sum_zone_node_page_state(nid, NR_BOUNCE)),
 450                        nid, K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
 451                        nid, K(sreclaimable +
 452                               node_page_state(pgdat, NR_KERNEL_MISC_RECLAIMABLE)),
 453                        nid, K(sreclaimable + sunreclaimable),
 454                        nid, K(sreclaimable),
 455                        nid, K(sunreclaimable)
 456 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 457                        ,
 458                        nid, K(node_page_state(pgdat, NR_ANON_THPS) *
 459                                        HPAGE_PMD_NR),
 460                        nid, K(node_page_state(pgdat, NR_SHMEM_THPS) *
 461                                        HPAGE_PMD_NR),
 462                        nid, K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED) *
 463                                        HPAGE_PMD_NR),
 464                        nid, K(node_page_state(pgdat, NR_FILE_THPS) *
 465                                        HPAGE_PMD_NR),
 466                        nid, K(node_page_state(pgdat, NR_FILE_PMDMAPPED) *
 467                                        HPAGE_PMD_NR)
 468 #endif
 469                        );
 470         n += hugetlb_report_node_meminfo(nid, buf + n);
 471         return n;
 472 }
 473
 474 #undef K
 475 static DEVICE_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
 476
 477 static ssize_t node_read_numastat(struct device *dev,
 478                                 struct device_attribute *attr, char *buf)
 479 {
 480         return sprintf(buf,
 481                        "numa_hit %lu\n"
 482                        "numa_miss %lu\n"
 483                        "numa_foreign %lu\n"
 484                        "interleave_hit %lu\n"
 485                        "local_node %lu\n"
 486                        "other_node %lu\n",
 487                        sum_zone_numa_state(dev->id, NUMA_HIT),
 488                        sum_zone_numa_state(dev->id, NUMA_MISS),
 489                        sum_zone_numa_state(dev->id, NUMA_FOREIGN),
 490                        sum_zone_numa_state(dev->id, NUMA_INTERLEAVE_HIT),
 491                        sum_zone_numa_state(dev->id, NUMA_LOCAL),
 492                        sum_zone_numa_state(dev->id, NUMA_OTHER));
 493 }
 494 static DEVICE_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
 495
 496 static ssize_t node_read_vmstat(struct device *dev,
 497                                 struct device_attribute *attr, char *buf)
 498 {
 499         int nid = dev->id;
 500         struct pglist_data *pgdat = NODE_DATA(nid);
 501         int i;
 502         int n = 0;
 503
 504         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 505                 n += sprintf(buf+n, "%s %lu\n", zone_stat_name(i),
 506                              sum_zone_node_page_state(nid, i));
 507
 508 #ifdef CONFIG_NUMA
 509         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 510                 n += sprintf(buf+n, "%s %lu\n", numa_stat_name(i),
 511                              sum_zone_numa_state(nid, i));
 512 #endif
 513
 514         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 515                 n += sprintf(buf+n, "%s %lu\n", node_stat_name(i),
 516                              node_page_state_pages(pgdat, i));
 517
 518         return n;
 519 }
 520 static DEVICE_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
 521
 522 static ssize_t node_read_distance(struct device *dev,
 523                         struct device_attribute *attr, char *buf)
 524 {
 525         int nid = dev->id;
 526         int len = 0;
 527         int i;
 528
 529         /*
 530          * buf is currently PAGE_SIZE in length and each node needs 4 chars
 531          * at the most (distance + space or newline).
 532          */
 533         BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
 534
 535         for_each_online_node(i)
 536                 len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
 537
 538         len += sprintf(buf + len, "\n");
 539         return len;
 540 }
 541 static DEVICE_ATTR(distance, S_IRUGO, node_read_distance, NULL);
 542
 543 static struct attribute *node_dev_attrs[] = {
 544         &dev_attr_cpumap.attr,
 545         &dev_attr_cpulist.attr,
 546         &dev_attr_meminfo.attr,
 547         &dev_attr_numastat.attr,
 548         &dev_attr_distance.attr,
 549         &dev_attr_vmstat.attr,
 550         NULL
 551 };
 552 ATTRIBUTE_GROUPS(node_dev);
 553
 554 #ifdef CONFIG_HUGETLBFS
 555 /*
 556  * hugetlbfs per node attributes registration interface:
 557  * When/if hugetlb[fs] subsystem initializes [sometime after this module],
 558  * it will register its per node attributes for all online nodes with
 559  * memory.  It will also call register_hugetlbfs_with_node(), below, to
 560  * register its attribute registration functions with this node driver.
 561  * Once these hooks have been initialized, the node driver will call into
 562  * the hugetlb module to [un]register attributes for hot-plugged nodes.
 563  */
 564 static node_registration_func_t __hugetlb_register_node;
 565 static node_registration_func_t __hugetlb_unregister_node;
 566
 567 static inline bool hugetlb_register_node(struct node *node)
 568 {
 569         if (__hugetlb_register_node &&
 570                         node_state(node->dev.id, N_MEMORY)) {
 571                 __hugetlb_register_node(node);
 572                 return true;
 573         }
 574         return false;
 575 }
 576
 577 static inline void hugetlb_unregister_node(struct node *node)
 578 {
 579         if (__hugetlb_unregister_node)
 580                 __hugetlb_unregister_node(node);
 581 }
 582
 583 void register_hugetlbfs_with_node(node_registration_func_t doregister,
 584                                   node_registration_func_t unregister)
 585 {
 586         __hugetlb_register_node   = doregister;
 587         __hugetlb_unregister_node = unregister;
 588 }
 589 #else
 590 static inline void hugetlb_register_node(struct node *node) {}
 591
 592 static inline void hugetlb_unregister_node(struct node *node) {}
 593 #endif
 594
 595 static void node_device_release(struct device *dev)
 596 {
 597         struct node *node = to_node(dev);
 598
 599 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HUGETLBFS)
 600         /*
 601          * We schedule the work only when a memory section is
 602          * onlined/offlined on this node. When we come here,
 603          * all the memory on this node has been offlined,
 604          * so we won't enqueue new work to this work.
 605          *
 606          * The work is using node->node_work, so we should
 607          * flush work before freeing the memory.
 608          */
 609         flush_work(&node->node_work);
 610 #endif
 611         kfree(node);
 612 }
 613
 614 /*
 615  * register_node - Setup a sysfs device for a node.
 616  * @num - Node number to use when creating the device.
 617  *
 618  * Initialize and register the node device.
 619  */
 620 static int register_node(struct node *node, int num)
 621 {
 622         int error;
 623
 624         node->dev.id = num;
 625         node->dev.bus = &node_subsys;
 626         node->dev.release = node_device_release;
 627         node->dev.groups = node_dev_groups;
 628         error = device_register(&node->dev);
 629
 630         if (error)
 631                 put_device(&node->dev);
 632         else {
 633                 hugetlb_register_node(node);
 634
 635                 compaction_register_node(node);
 636         }
 637         return error;
 638 }
 639
 640 /**
 641  * unregister_node - unregister a node device
 642  * @node: node going away
 643  *
 644  * Unregisters a node device @node.  All the devices on the node must be
 645  * unregistered before calling this function.
 646  */
 647 void unregister_node(struct node *node)
 648 {
 649         hugetlb_unregister_node(node);          /* no-op, if memoryless node */
 650         node_remove_accesses(node);
 651         node_remove_caches(node);
 652         device_unregister(&node->dev);
 653 }
 654
 655 struct node *node_devices[MAX_NUMNODES];
 656
 657 /*
 658  * register cpu under node
 659  */
 660 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
 661 {
 662         int ret;
 663         struct device *obj;
 664
 665         if (!node_online(nid))
 666                 return 0;
 667
 668         obj = get_cpu_device(cpu);
 669         if (!obj)
 670                 return 0;
 671
 672         ret = sysfs_create_link(&node_devices[nid]->dev.kobj,
 673                                 &obj->kobj,
 674                                 kobject_name(&obj->kobj));
 675         if (ret)
 676                 return ret;
 677
 678         return sysfs_create_link(&obj->kobj,
 679                                  &node_devices[nid]->dev.kobj,
 680                                  kobject_name(&node_devices[nid]->dev.kobj));
 681 }
 682
 683 /**
 684  * register_memory_node_under_compute_node - link memory node to its compute
 685  *                                           node for a given access class.
 686  * @mem_nid:    Memory node number
 687  * @cpu_nid:    Cpu  node number
 688  * @access:     Access class to register
 689  *
 690  * Description:
 691  *      For use with platforms that may have separate memory and compute nodes.
 692  *      This function will export node relationships linking which memory
 693  *      initiator nodes can access memory targets at a given ranked access
 694  *      class.
 695  */
 696 int register_memory_node_under_compute_node(unsigned int mem_nid,
 697                                             unsigned int cpu_nid,
 698                                             unsigned access)
 699 {
 700         struct node *init_node, *targ_node;
 701         struct node_access_nodes *initiator, *target;
 702         int ret;
 703
 704         if (!node_online(cpu_nid) || !node_online(mem_nid))
 705                 return -ENODEV;
 706
 707         init_node = node_devices[cpu_nid];
 708         targ_node = node_devices[mem_nid];
 709         initiator = node_init_node_access(init_node, access);
 710         target = node_init_node_access(targ_node, access);
 711         if (!initiator || !target)
 712                 return -ENOMEM;
 713
 714         ret = sysfs_add_link_to_group(&initiator->dev.kobj, "targets",
 715                                       &targ_node->dev.kobj,
 716                                       dev_name(&targ_node->dev));
 717         if (ret)
 718                 return ret;
 719
 720         ret = sysfs_add_link_to_group(&target->dev.kobj, "initiators",
 721                                       &init_node->dev.kobj,
 722                                       dev_name(&init_node->dev));
 723         if (ret)
 724                 goto err;
 725
 726         return 0;
 727  err:
 728         sysfs_remove_link_from_group(&initiator->dev.kobj, "targets",
 729                                      dev_name(&targ_node->dev));
 730         return ret;
 731 }
 732
 733 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
 734 {
 735         struct device *obj;
 736
 737         if (!node_online(nid))
 738                 return 0;
 739
 740         obj = get_cpu_device(cpu);
 741         if (!obj)
 742                 return 0;
 743
 744         sysfs_remove_link(&node_devices[nid]->dev.kobj,
 745                           kobject_name(&obj->kobj));
 746         sysfs_remove_link(&obj->kobj,
 747                           kobject_name(&node_devices[nid]->dev.kobj));
 748
 749         return 0;
 750 }
 751
 752 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
 753 static int __ref get_nid_for_pfn(unsigned long pfn)
 754 {
 755         if (!pfn_valid_within(pfn))
 756                 return -1;
 757 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
 758         if (system_state < SYSTEM_RUNNING)
 759                 return early_pfn_to_nid(pfn);
 760 #endif
 761         return pfn_to_nid(pfn);
 762 }
 763
 764 static int do_register_memory_block_under_node(int nid,
 765                                                struct memory_block *mem_blk)
 766 {
 767         int ret;
 768
 769         /*
 770          * If this memory block spans multiple nodes, we only indicate
 771          * the last processed node.
 772          */
 773         mem_blk->nid = nid;
 774
 775         ret = sysfs_create_link_nowarn(&node_devices[nid]->dev.kobj,
 776                                        &mem_blk->dev.kobj,
 777                                        kobject_name(&mem_blk->dev.kobj));
 778         if (ret)
 779                 return ret;
 780
 781         return sysfs_create_link_nowarn(&mem_blk->dev.kobj,
 782                                 &node_devices[nid]->dev.kobj,
 783                                 kobject_name(&node_devices[nid]->dev.kobj));
 784 }
 785
 786 /* register memory section under specified node if it spans that node */
 787 static int register_mem_block_under_node_early(struct memory_block *mem_blk,
 788                                                void *arg)
 789 {
 790         unsigned long memory_block_pfns = memory_block_size_bytes() / PAGE_SIZE;
 791         unsigned long start_pfn = section_nr_to_pfn(mem_blk->start_section_nr);
 792         unsigned long end_pfn = start_pfn + memory_block_pfns - 1;
 793         int nid = *(int *)arg;
 794         unsigned long pfn;
 795
 796         for (pfn = start_pfn; pfn <= end_pfn; pfn++) {
 797                 int page_nid;
 798
 799                 /*
 800                  * memory block could have several absent sections from start.
 801                  * skip pfn range from absent section
 802                  */
 803                 if (!pfn_in_present_section(pfn)) {
 804                         pfn = round_down(pfn + PAGES_PER_SECTION,
 805                                          PAGES_PER_SECTION) - 1;
 806                         continue;
 807                 }
 808
 809                 /*
 810                  * We need to check if page belongs to nid only at the boot
 811                  * case because node's ranges can be interleaved.
 812                  */
 813                 page_nid = get_nid_for_pfn(pfn);
 814                 if (page_nid < 0)
 815                         continue;
 816                 if (page_nid != nid)
 817                         continue;
 818
 819                 return do_register_memory_block_under_node(nid, mem_blk);
 820         }
 821         /* mem section does not span the specified node */
 822         return 0;
 823 }
 824
 825 /*
 826  * During hotplug we know that all pages in the memory block belong to the same
 827  * node.
 828  */
 829 static int register_mem_block_under_node_hotplug(struct memory_block *mem_blk,
 830                                                  void *arg)
 831 {
 832         int nid = *(int *)arg;
 833
 834         return do_register_memory_block_under_node(nid, mem_blk);
 835 }
 836
 837 /*
 838  * Unregister a memory block device under the node it spans. Memory blocks
 839  * with multiple nodes cannot be offlined and therefore also never be removed.
 840  */
 841 void unregister_memory_block_under_nodes(struct memory_block *mem_blk)
 842 {
 843         if (mem_blk->nid == NUMA_NO_NODE)
 844                 return;
 845
 846         sysfs_remove_link(&node_devices[mem_blk->nid]->dev.kobj,
 847                           kobject_name(&mem_blk->dev.kobj));
 848         sysfs_remove_link(&mem_blk->dev.kobj,
 849                           kobject_name(&node_devices[mem_blk->nid]->dev.kobj));
 850 }
 851
 852 int link_mem_sections(int nid, unsigned long start_pfn, unsigned long end_pfn,
 853                       enum meminit_context context)
 854 {
 855         walk_memory_blocks_func_t func;
 856
 857         if (context == MEMINIT_HOTPLUG)
 858                 func = register_mem_block_under_node_hotplug;
 859         else
 860                 func = register_mem_block_under_node_early;
 861
 862         return walk_memory_blocks(PFN_PHYS(start_pfn),
 863                                   PFN_PHYS(end_pfn - start_pfn), (void *)&nid,
 864                                   func);
 865 }
 866
 867 #ifdef CONFIG_HUGETLBFS
 868 /*
 869  * Handle per node hstate attribute [un]registration on transistions
 870  * to/from memoryless state.
 871  */
 872 static void node_hugetlb_work(struct work_struct *work)
 873 {
 874         struct node *node = container_of(work, struct node, node_work);
 875
 876         /*
 877          * We only get here when a node transitions to/from memoryless state.
 878          * We can detect which transition occurred by examining whether the
 879          * node has memory now.  hugetlb_register_node() already check this
 880          * so we try to register the attributes.  If that fails, then the
 881          * node has transitioned to memoryless, try to unregister the
 882          * attributes.
 883          */
 884         if (!hugetlb_register_node(node))
 885                 hugetlb_unregister_node(node);
 886 }
 887
 888 static void init_node_hugetlb_work(int nid)
 889 {
 890         INIT_WORK(&node_devices[nid]->node_work, node_hugetlb_work);
 891 }
 892
 893 static int node_memory_callback(struct notifier_block *self,
 894                                 unsigned long action, void *arg)
 895 {
 896         struct memory_notify *mnb = arg;
 897         int nid = mnb->status_change_nid;
 898
 899         switch (action) {
 900         case MEM_ONLINE:
 901         case MEM_OFFLINE:
 902                 /*
 903                  * offload per node hstate [un]registration to a work thread
 904                  * when transitioning to/from memoryless state.
 905                  */
 906                 if (nid != NUMA_NO_NODE)
 907                         schedule_work(&node_devices[nid]->node_work);
 908                 break;
 909
 910         case MEM_GOING_ONLINE:
 911         case MEM_GOING_OFFLINE:
 912         case MEM_CANCEL_ONLINE:
 913         case MEM_CANCEL_OFFLINE:
 914         default:
 915                 break;
 916         }
 917
 918         return NOTIFY_OK;
 919 }
 920 #endif  /* CONFIG_HUGETLBFS */
 921 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
 922
 923 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
 924     !defined(CONFIG_HUGETLBFS)
 925 static inline int node_memory_callback(struct notifier_block *self,
 926                                 unsigned long action, void *arg)
 927 {
 928         return NOTIFY_OK;
 929 }
 930
 931 static void init_node_hugetlb_work(int nid) { }
 932
 933 #endif
 934
 935 int __register_one_node(int nid)
 936 {
 937         int error;
 938         int cpu;
 939
 940         node_devices[nid] = kzalloc(sizeof(struct node), GFP_KERNEL);
 941         if (!node_devices[nid])
 942                 return -ENOMEM;
 943
 944         error = register_node(node_devices[nid], nid);
 945
 946         /* link cpu under this node */
 947         for_each_present_cpu(cpu) {
 948                 if (cpu_to_node(cpu) == nid)
 949                         register_cpu_under_node(cpu, nid);
 950         }
 951
 952         INIT_LIST_HEAD(&node_devices[nid]->access_list);
 953         /* initialize work queue for memory hot plug */
 954         init_node_hugetlb_work(nid);
 955         node_init_caches(nid);
 956
 957         return error;
 958 }
 959
 960 void unregister_one_node(int nid)
 961 {
 962         if (!node_devices[nid])
 963                 return;
 964
 965         unregister_node(node_devices[nid]);
 966         node_devices[nid] = NULL;
 967 }
 968
 969 /*
 970  * node states attributes
 971  */
 972
 973 static ssize_t print_nodes_state(enum node_states state, char *buf)
 974 {
 975         int n;
 976
 977         n = scnprintf(buf, PAGE_SIZE - 1, "%*pbl",
 978                       nodemask_pr_args(&node_states[state]));
 979         buf[n++] = '\n';
 980         buf[n] = '\0';
 981         return n;
 982 }
 983
 984 struct node_attr {
 985         struct device_attribute attr;
 986         enum node_states state;
 987 };
 988
 989 static ssize_t show_node_state(struct device *dev,
 990                                struct device_attribute *attr, char *buf)
 991 {
 992         struct node_attr *na = container_of(attr, struct node_attr, attr);
 993         return print_nodes_state(na->state, buf);
 994 }
 995
 996 #define _NODE_ATTR(name, state) \
 997         { __ATTR(name, 0444, show_node_state, NULL), state }
 998
 999 static struct node_attr node_state_attr[] = {
1000         [N_POSSIBLE] = _NODE_ATTR(possible, N_POSSIBLE),
1001         [N_ONLINE] = _NODE_ATTR(online, N_ONLINE),
1002         [N_NORMAL_MEMORY] = _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
1003 #ifdef CONFIG_HIGHMEM
1004         [N_HIGH_MEMORY] = _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
1005 #endif
1006         [N_MEMORY] = _NODE_ATTR(has_memory, N_MEMORY),
1007         [N_CPU] = _NODE_ATTR(has_cpu, N_CPU),
1008         [N_GENERIC_INITIATOR] = _NODE_ATTR(has_generic_initiator,
1009                                            N_GENERIC_INITIATOR),
1010 };
1011
1012 static struct attribute *node_state_attrs[] = {
1013         &node_state_attr[N_POSSIBLE].attr.attr,
1014         &node_state_attr[N_ONLINE].attr.attr,
1015         &node_state_attr[N_NORMAL_MEMORY].attr.attr,
1016 #ifdef CONFIG_HIGHMEM
1017         &node_state_attr[N_HIGH_MEMORY].attr.attr,
1018 #endif
1019         &node_state_attr[N_MEMORY].attr.attr,
1020         &node_state_attr[N_CPU].attr.attr,
1021         &node_state_attr[N_GENERIC_INITIATOR].attr.attr,
1022         NULL
1023 };
1024
1025 static struct attribute_group memory_root_attr_group = {
1026         .attrs = node_state_attrs,
1027 };
1028
1029 static const struct attribute_group *cpu_root_attr_groups[] = {
1030         &memory_root_attr_group,
1031         NULL,
1032 };
1033
1034 #define NODE_CALLBACK_PRI       2       /* lower than SLAB */
1035 static int __init register_node_type(void)
1036 {
1037         int ret;
1038
1039         BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
1040         BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
1041
1042         ret = subsys_system_register(&node_subsys, cpu_root_attr_groups);
1043         if (!ret) {
1044                 static struct notifier_block node_memory_callback_nb = {
1045                         .notifier_call = node_memory_callback,
1046                         .priority = NODE_CALLBACK_PRI,
1047                 };
1048                 register_hotmemory_notifier(&node_memory_callback_nb);
1049         }
1050
1051         /*
1052          * Note:  we're not going to unregister the node class if we fail
1053          * to register the node state class attribute files.
1054          */
1055         return ret;
1056 }
1057 postcore_initcall(register_node_type);