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 cpumap_show(struct device *dev,
  50                                   struct device_attribute *attr,
  51                                   char *buf)
  52 {
  53         return node_read_cpumap(dev, false, buf);
  54 }
  55
  56 static DEVICE_ATTR_RO(cpumap);
  57
  58 static inline ssize_t cpulist_show(struct device *dev,
  59                                    struct device_attribute *attr,
  60                                    char *buf)
  61 {
  62         return node_read_cpumap(dev, true, buf);
  63 }
  64
  65 static DEVICE_ATTR_RO(cpulist);
  66
  67 /**
  68  * struct node_access_nodes - Access class device to hold user visible
  69  *                            relationships to other nodes.
  70  * @dev:        Device for this memory access class
  71  * @list_node:  List element in the node's access list
  72  * @access:     The access class rank
  73  * @hmem_attrs: Heterogeneous memory performance attributes
  74  */
  75 struct node_access_nodes {
  76         struct device           dev;
  77         struct list_head        list_node;
  78         unsigned                access;
  79 #ifdef CONFIG_HMEM_REPORTING
  80         struct node_hmem_attrs  hmem_attrs;
  81 #endif
  82 };
  83 #define to_access_nodes(dev) container_of(dev, struct node_access_nodes, dev)
  84
  85 static struct attribute *node_init_access_node_attrs[] = {
  86         NULL,
  87 };
  88
  89 static struct attribute *node_targ_access_node_attrs[] = {
  90         NULL,
  91 };
  92
  93 static const struct attribute_group initiators = {
  94         .name   = "initiators",
  95         .attrs  = node_init_access_node_attrs,
  96 };
  97
  98 static const struct attribute_group targets = {
  99         .name   = "targets",
 100         .attrs  = node_targ_access_node_attrs,
 101 };
 102
 103 static const struct attribute_group *node_access_node_groups[] = {
 104         &initiators,
 105         &targets,
 106         NULL,
 107 };
 108
 109 static void node_remove_accesses(struct node *node)
 110 {
 111         struct node_access_nodes *c, *cnext;
 112
 113         list_for_each_entry_safe(c, cnext, &node->access_list, list_node) {
 114                 list_del(&c->list_node);
 115                 device_unregister(&c->dev);
 116         }
 117 }
 118
 119 static void node_access_release(struct device *dev)
 120 {
 121         kfree(to_access_nodes(dev));
 122 }
 123
 124 static struct node_access_nodes *node_init_node_access(struct node *node,
 125                                                        unsigned access)
 126 {
 127         struct node_access_nodes *access_node;
 128         struct device *dev;
 129
 130         list_for_each_entry(access_node, &node->access_list, list_node)
 131                 if (access_node->access == access)
 132                         return access_node;
 133
 134         access_node = kzalloc(sizeof(*access_node), GFP_KERNEL);
 135         if (!access_node)
 136                 return NULL;
 137
 138         access_node->access = access;
 139         dev = &access_node->dev;
 140         dev->parent = &node->dev;
 141         dev->release = node_access_release;
 142         dev->groups = node_access_node_groups;
 143         if (dev_set_name(dev, "access%u", access))
 144                 goto free;
 145
 146         if (device_register(dev))
 147                 goto free_name;
 148
 149         pm_runtime_no_callbacks(dev);
 150         list_add_tail(&access_node->list_node, &node->access_list);
 151         return access_node;
 152 free_name:
 153         kfree_const(dev->kobj.name);
 154 free:
 155         kfree(access_node);
 156         return NULL;
 157 }
 158
 159 #ifdef CONFIG_HMEM_REPORTING
 160 #define ACCESS_ATTR(name)                                               \
 161 static ssize_t name##_show(struct device *dev,                          \
 162                            struct device_attribute *attr,               \
 163                            char *buf)                                   \
 164 {                                                                       \
 165         return sysfs_emit(buf, "%u\n",                                  \
 166                           to_access_nodes(dev)->hmem_attrs.name);       \
 167 }                                                                       \
 168 static DEVICE_ATTR_RO(name)
 169
 170 ACCESS_ATTR(read_bandwidth);
 171 ACCESS_ATTR(read_latency);
 172 ACCESS_ATTR(write_bandwidth);
 173 ACCESS_ATTR(write_latency);
 174
 175 static struct attribute *access_attrs[] = {
 176         &dev_attr_read_bandwidth.attr,
 177         &dev_attr_read_latency.attr,
 178         &dev_attr_write_bandwidth.attr,
 179         &dev_attr_write_latency.attr,
 180         NULL,
 181 };
 182
 183 /**
 184  * node_set_perf_attrs - Set the performance values for given access class
 185  * @nid: Node identifier to be set
 186  * @hmem_attrs: Heterogeneous memory performance attributes
 187  * @access: The access class the for the given attributes
 188  */
 189 void node_set_perf_attrs(unsigned int nid, struct node_hmem_attrs *hmem_attrs,
 190                          unsigned access)
 191 {
 192         struct node_access_nodes *c;
 193         struct node *node;
 194         int i;
 195
 196         if (WARN_ON_ONCE(!node_online(nid)))
 197                 return;
 198
 199         node = node_devices[nid];
 200         c = node_init_node_access(node, access);
 201         if (!c)
 202                 return;
 203
 204         c->hmem_attrs = *hmem_attrs;
 205         for (i = 0; access_attrs[i] != NULL; i++) {
 206                 if (sysfs_add_file_to_group(&c->dev.kobj, access_attrs[i],
 207                                             "initiators")) {
 208                         pr_info("failed to add performance attribute to node %d\n",
 209                                 nid);
 210                         break;
 211                 }
 212         }
 213 }
 214
 215 /**
 216  * struct node_cache_info - Internal tracking for memory node caches
 217  * @dev:        Device represeting the cache level
 218  * @node:       List element for tracking in the node
 219  * @cache_attrs:Attributes for this cache level
 220  */
 221 struct node_cache_info {
 222         struct device dev;
 223         struct list_head node;
 224         struct node_cache_attrs cache_attrs;
 225 };
 226 #define to_cache_info(device) container_of(device, struct node_cache_info, dev)
 227
 228 #define CACHE_ATTR(name, fmt)                                           \
 229 static ssize_t name##_show(struct device *dev,                          \
 230                            struct device_attribute *attr,               \
 231                            char *buf)                                   \
 232 {                                                                       \
 233         return sysfs_emit(buf, fmt "\n",                                \
 234                           to_cache_info(dev)->cache_attrs.name);        \
 235 }                                                                       \
 236 DEVICE_ATTR_RO(name);
 237
 238 CACHE_ATTR(size, "%llu")
 239 CACHE_ATTR(line_size, "%u")
 240 CACHE_ATTR(indexing, "%u")
 241 CACHE_ATTR(write_policy, "%u")
 242
 243 static struct attribute *cache_attrs[] = {
 244         &dev_attr_indexing.attr,
 245         &dev_attr_size.attr,
 246         &dev_attr_line_size.attr,
 247         &dev_attr_write_policy.attr,
 248         NULL,
 249 };
 250 ATTRIBUTE_GROUPS(cache);
 251
 252 static void node_cache_release(struct device *dev)
 253 {
 254         kfree(dev);
 255 }
 256
 257 static void node_cacheinfo_release(struct device *dev)
 258 {
 259         struct node_cache_info *info = to_cache_info(dev);
 260         kfree(info);
 261 }
 262
 263 static void node_init_cache_dev(struct node *node)
 264 {
 265         struct device *dev;
 266
 267         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
 268         if (!dev)
 269                 return;
 270
 271         dev->parent = &node->dev;
 272         dev->release = node_cache_release;
 273         if (dev_set_name(dev, "memory_side_cache"))
 274                 goto free_dev;
 275
 276         if (device_register(dev))
 277                 goto free_name;
 278
 279         pm_runtime_no_callbacks(dev);
 280         node->cache_dev = dev;
 281         return;
 282 free_name:
 283         kfree_const(dev->kobj.name);
 284 free_dev:
 285         kfree(dev);
 286 }
 287
 288 /**
 289  * node_add_cache() - add cache attribute to a memory node
 290  * @nid: Node identifier that has new cache attributes
 291  * @cache_attrs: Attributes for the cache being added
 292  */
 293 void node_add_cache(unsigned int nid, struct node_cache_attrs *cache_attrs)
 294 {
 295         struct node_cache_info *info;
 296         struct device *dev;
 297         struct node *node;
 298
 299         if (!node_online(nid) || !node_devices[nid])
 300                 return;
 301
 302         node = node_devices[nid];
 303         list_for_each_entry(info, &node->cache_attrs, node) {
 304                 if (info->cache_attrs.level == cache_attrs->level) {
 305                         dev_warn(&node->dev,
 306                                 "attempt to add duplicate cache level:%d\n",
 307                                 cache_attrs->level);
 308                         return;
 309                 }
 310         }
 311
 312         if (!node->cache_dev)
 313                 node_init_cache_dev(node);
 314         if (!node->cache_dev)
 315                 return;
 316
 317         info = kzalloc(sizeof(*info), GFP_KERNEL);
 318         if (!info)
 319                 return;
 320
 321         dev = &info->dev;
 322         dev->parent = node->cache_dev;
 323         dev->release = node_cacheinfo_release;
 324         dev->groups = cache_groups;
 325         if (dev_set_name(dev, "index%d", cache_attrs->level))
 326                 goto free_cache;
 327
 328         info->cache_attrs = *cache_attrs;
 329         if (device_register(dev)) {
 330                 dev_warn(&node->dev, "failed to add cache level:%d\n",
 331                          cache_attrs->level);
 332                 goto free_name;
 333         }
 334         pm_runtime_no_callbacks(dev);
 335         list_add_tail(&info->node, &node->cache_attrs);
 336         return;
 337 free_name:
 338         kfree_const(dev->kobj.name);
 339 free_cache:
 340         kfree(info);
 341 }
 342
 343 static void node_remove_caches(struct node *node)
 344 {
 345         struct node_cache_info *info, *next;
 346
 347         if (!node->cache_dev)
 348                 return;
 349
 350         list_for_each_entry_safe(info, next, &node->cache_attrs, node) {
 351                 list_del(&info->node);
 352                 device_unregister(&info->dev);
 353         }
 354         device_unregister(node->cache_dev);
 355 }
 356
 357 static void node_init_caches(unsigned int nid)
 358 {
 359         INIT_LIST_HEAD(&node_devices[nid]->cache_attrs);
 360 }
 361 #else
 362 static void node_init_caches(unsigned int nid) { }
 363 static void node_remove_caches(struct node *node) { }
 364 #endif
 365
 366 #define K(x) ((x) << (PAGE_SHIFT - 10))
 367 static ssize_t node_read_meminfo(struct device *dev,
 368                         struct device_attribute *attr, char *buf)
 369 {
 370         int len = 0;
 371         int nid = dev->id;
 372         struct pglist_data *pgdat = NODE_DATA(nid);
 373         struct sysinfo i;
 374         unsigned long sreclaimable, sunreclaimable;
 375
 376         si_meminfo_node(&i, nid);
 377         sreclaimable = node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B);
 378         sunreclaimable = node_page_state_pages(pgdat, NR_SLAB_UNRECLAIMABLE_B);
 379         len = sysfs_emit_at(buf, len,
 380                             "Node %d MemTotal:       %8lu kB\n"
 381                             "Node %d MemFree:        %8lu kB\n"
 382                             "Node %d MemUsed:        %8lu kB\n"
 383                             "Node %d Active:         %8lu kB\n"
 384                             "Node %d Inactive:       %8lu kB\n"
 385                             "Node %d Active(anon):   %8lu kB\n"
 386                             "Node %d Inactive(anon): %8lu kB\n"
 387                             "Node %d Active(file):   %8lu kB\n"
 388                             "Node %d Inactive(file): %8lu kB\n"
 389                             "Node %d Unevictable:    %8lu kB\n"
 390                             "Node %d Mlocked:        %8lu kB\n",
 391                             nid, K(i.totalram),
 392                             nid, K(i.freeram),
 393                             nid, K(i.totalram - i.freeram),
 394                             nid, K(node_page_state(pgdat, NR_ACTIVE_ANON) +
 395                                    node_page_state(pgdat, NR_ACTIVE_FILE)),
 396                             nid, K(node_page_state(pgdat, NR_INACTIVE_ANON) +
 397                                    node_page_state(pgdat, NR_INACTIVE_FILE)),
 398                             nid, K(node_page_state(pgdat, NR_ACTIVE_ANON)),
 399                             nid, K(node_page_state(pgdat, NR_INACTIVE_ANON)),
 400                             nid, K(node_page_state(pgdat, NR_ACTIVE_FILE)),
 401                             nid, K(node_page_state(pgdat, NR_INACTIVE_FILE)),
 402                             nid, K(node_page_state(pgdat, NR_UNEVICTABLE)),
 403                             nid, K(sum_zone_node_page_state(nid, NR_MLOCK)));
 404
 405 #ifdef CONFIG_HIGHMEM
 406         len += sysfs_emit_at(buf, len,
 407                              "Node %d HighTotal:      %8lu kB\n"
 408                              "Node %d HighFree:       %8lu kB\n"
 409                              "Node %d LowTotal:       %8lu kB\n"
 410                              "Node %d LowFree:        %8lu kB\n",
 411                              nid, K(i.totalhigh),
 412                              nid, K(i.freehigh),
 413                              nid, K(i.totalram - i.totalhigh),
 414                              nid, K(i.freeram - i.freehigh));
 415 #endif
 416         len += sysfs_emit_at(buf, len,
 417                              "Node %d Dirty:          %8lu kB\n"
 418                              "Node %d Writeback:      %8lu kB\n"
 419                              "Node %d FilePages:      %8lu kB\n"
 420                              "Node %d Mapped:         %8lu kB\n"
 421                              "Node %d AnonPages:      %8lu kB\n"
 422                              "Node %d Shmem:          %8lu kB\n"
 423                              "Node %d KernelStack:    %8lu kB\n"
 424 #ifdef CONFIG_SHADOW_CALL_STACK
 425                              "Node %d ShadowCallStack:%8lu kB\n"
 426 #endif
 427                              "Node %d PageTables:     %8lu kB\n"
 428                              "Node %d NFS_Unstable:   %8lu kB\n"
 429                              "Node %d Bounce:         %8lu kB\n"
 430                              "Node %d WritebackTmp:   %8lu kB\n"
 431                              "Node %d KReclaimable:   %8lu kB\n"
 432                              "Node %d Slab:           %8lu kB\n"
 433                              "Node %d SReclaimable:   %8lu kB\n"
 434                              "Node %d SUnreclaim:     %8lu kB\n"
 435 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 436                              "Node %d AnonHugePages:  %8lu kB\n"
 437                              "Node %d ShmemHugePages: %8lu kB\n"
 438                              "Node %d ShmemPmdMapped: %8lu kB\n"
 439                              "Node %d FileHugePages: %8lu kB\n"
 440                              "Node %d FilePmdMapped: %8lu kB\n"
 441 #endif
 442                              ,
 443                              nid, K(node_page_state(pgdat, NR_FILE_DIRTY)),
 444                              nid, K(node_page_state(pgdat, NR_WRITEBACK)),
 445                              nid, K(node_page_state(pgdat, NR_FILE_PAGES)),
 446                              nid, K(node_page_state(pgdat, NR_FILE_MAPPED)),
 447                              nid, K(node_page_state(pgdat, NR_ANON_MAPPED)),
 448                              nid, K(i.sharedram),
 449                              nid, node_page_state(pgdat, NR_KERNEL_STACK_KB),
 450 #ifdef CONFIG_SHADOW_CALL_STACK
 451                              nid, node_page_state(pgdat, NR_KERNEL_SCS_KB),
 452 #endif
 453                              nid, K(sum_zone_node_page_state(nid, NR_PAGETABLE)),
 454                              nid, 0UL,
 455                              nid, K(sum_zone_node_page_state(nid, NR_BOUNCE)),
 456                              nid, K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
 457                              nid, K(sreclaimable +
 458                                     node_page_state(pgdat, NR_KERNEL_MISC_RECLAIMABLE)),
 459                              nid, K(sreclaimable + sunreclaimable),
 460                              nid, K(sreclaimable),
 461                              nid, K(sunreclaimable)
 462 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 463                              ,
 464                              nid, K(node_page_state(pgdat, NR_ANON_THPS) *
 465                                     HPAGE_PMD_NR),
 466                              nid, K(node_page_state(pgdat, NR_SHMEM_THPS) *
 467                                     HPAGE_PMD_NR),
 468                              nid, K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED) *
 469                                     HPAGE_PMD_NR),
 470                              nid, K(node_page_state(pgdat, NR_FILE_THPS) *
 471                                     HPAGE_PMD_NR),
 472                              nid, K(node_page_state(pgdat, NR_FILE_PMDMAPPED) *
 473                                     HPAGE_PMD_NR)
 474 #endif
 475                             );
 476         len += hugetlb_report_node_meminfo(buf, len, nid);
 477         return len;
 478 }
 479
 480 #undef K
 481 static DEVICE_ATTR(meminfo, 0444, node_read_meminfo, NULL);
 482
 483 static ssize_t node_read_numastat(struct device *dev,
 484                                   struct device_attribute *attr, char *buf)
 485 {
 486         return sysfs_emit(buf,
 487                           "numa_hit %lu\n"
 488                           "numa_miss %lu\n"
 489                           "numa_foreign %lu\n"
 490                           "interleave_hit %lu\n"
 491                           "local_node %lu\n"
 492                           "other_node %lu\n",
 493                           sum_zone_numa_state(dev->id, NUMA_HIT),
 494                           sum_zone_numa_state(dev->id, NUMA_MISS),
 495                           sum_zone_numa_state(dev->id, NUMA_FOREIGN),
 496                           sum_zone_numa_state(dev->id, NUMA_INTERLEAVE_HIT),
 497                           sum_zone_numa_state(dev->id, NUMA_LOCAL),
 498                           sum_zone_numa_state(dev->id, NUMA_OTHER));
 499 }
 500 static DEVICE_ATTR(numastat, 0444, node_read_numastat, NULL);
 501
 502 static ssize_t node_read_vmstat(struct device *dev,
 503                                 struct device_attribute *attr, char *buf)
 504 {
 505         int nid = dev->id;
 506         struct pglist_data *pgdat = NODE_DATA(nid);
 507         int i;
 508         int len = 0;
 509
 510         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 511                 len += sysfs_emit_at(buf, len, "%s %lu\n",
 512                                      zone_stat_name(i),
 513                                      sum_zone_node_page_state(nid, i));
 514
 515 #ifdef CONFIG_NUMA
 516         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 517                 len += sysfs_emit_at(buf, len, "%s %lu\n",
 518                                      numa_stat_name(i),
 519                                      sum_zone_numa_state(nid, i));
 520
 521 #endif
 522         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 523                 len += sysfs_emit_at(buf, len, "%s %lu\n",
 524                                      node_stat_name(i),
 525                                      node_page_state_pages(pgdat, i));
 526
 527         return len;
 528 }
 529 static DEVICE_ATTR(vmstat, 0444, node_read_vmstat, NULL);
 530
 531 static ssize_t node_read_distance(struct device *dev,
 532                                   struct device_attribute *attr, char *buf)
 533 {
 534         int nid = dev->id;
 535         int len = 0;
 536         int i;
 537
 538         /*
 539          * buf is currently PAGE_SIZE in length and each node needs 4 chars
 540          * at the most (distance + space or newline).
 541          */
 542         BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
 543
 544         for_each_online_node(i) {
 545                 len += sysfs_emit_at(buf, len, "%s%d",
 546                                      i ? " " : "", node_distance(nid, i));
 547         }
 548
 549         len += sysfs_emit_at(buf, len, "\n");
 550         return len;
 551 }
 552 static DEVICE_ATTR(distance, 0444, node_read_distance, NULL);
 553
 554 static struct attribute *node_dev_attrs[] = {
 555         &dev_attr_cpumap.attr,
 556         &dev_attr_cpulist.attr,
 557         &dev_attr_meminfo.attr,
 558         &dev_attr_numastat.attr,
 559         &dev_attr_distance.attr,
 560         &dev_attr_vmstat.attr,
 561         NULL
 562 };
 563 ATTRIBUTE_GROUPS(node_dev);
 564
 565 #ifdef CONFIG_HUGETLBFS
 566 /*
 567  * hugetlbfs per node attributes registration interface:
 568  * When/if hugetlb[fs] subsystem initializes [sometime after this module],
 569  * it will register its per node attributes for all online nodes with
 570  * memory.  It will also call register_hugetlbfs_with_node(), below, to
 571  * register its attribute registration functions with this node driver.
 572  * Once these hooks have been initialized, the node driver will call into
 573  * the hugetlb module to [un]register attributes for hot-plugged nodes.
 574  */
 575 static node_registration_func_t __hugetlb_register_node;
 576 static node_registration_func_t __hugetlb_unregister_node;
 577
 578 static inline bool hugetlb_register_node(struct node *node)
 579 {
 580         if (__hugetlb_register_node &&
 581                         node_state(node->dev.id, N_MEMORY)) {
 582                 __hugetlb_register_node(node);
 583                 return true;
 584         }
 585         return false;
 586 }
 587
 588 static inline void hugetlb_unregister_node(struct node *node)
 589 {
 590         if (__hugetlb_unregister_node)
 591                 __hugetlb_unregister_node(node);
 592 }
 593
 594 void register_hugetlbfs_with_node(node_registration_func_t doregister,
 595                                   node_registration_func_t unregister)
 596 {
 597         __hugetlb_register_node   = doregister;
 598         __hugetlb_unregister_node = unregister;
 599 }
 600 #else
 601 static inline void hugetlb_register_node(struct node *node) {}
 602
 603 static inline void hugetlb_unregister_node(struct node *node) {}
 604 #endif
 605
 606 static void node_device_release(struct device *dev)
 607 {
 608         struct node *node = to_node(dev);
 609
 610 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HUGETLBFS)
 611         /*
 612          * We schedule the work only when a memory section is
 613          * onlined/offlined on this node. When we come here,
 614          * all the memory on this node has been offlined,
 615          * so we won't enqueue new work to this work.
 616          *
 617          * The work is using node->node_work, so we should
 618          * flush work before freeing the memory.
 619          */
 620         flush_work(&node->node_work);
 621 #endif
 622         kfree(node);
 623 }
 624
 625 /*
 626  * register_node - Setup a sysfs device for a node.
 627  * @num - Node number to use when creating the device.
 628  *
 629  * Initialize and register the node device.
 630  */
 631 static int register_node(struct node *node, int num)
 632 {
 633         int error;
 634
 635         node->dev.id = num;
 636         node->dev.bus = &node_subsys;
 637         node->dev.release = node_device_release;
 638         node->dev.groups = node_dev_groups;
 639         error = device_register(&node->dev);
 640
 641         if (error)
 642                 put_device(&node->dev);
 643         else {
 644                 hugetlb_register_node(node);
 645
 646                 compaction_register_node(node);
 647         }
 648         return error;
 649 }
 650
 651 /**
 652  * unregister_node - unregister a node device
 653  * @node: node going away
 654  *
 655  * Unregisters a node device @node.  All the devices on the node must be
 656  * unregistered before calling this function.
 657  */
 658 void unregister_node(struct node *node)
 659 {
 660         hugetlb_unregister_node(node);          /* no-op, if memoryless node */
 661         node_remove_accesses(node);
 662         node_remove_caches(node);
 663         device_unregister(&node->dev);
 664 }
 665
 666 struct node *node_devices[MAX_NUMNODES];
 667
 668 /*
 669  * register cpu under node
 670  */
 671 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
 672 {
 673         int ret;
 674         struct device *obj;
 675
 676         if (!node_online(nid))
 677                 return 0;
 678
 679         obj = get_cpu_device(cpu);
 680         if (!obj)
 681                 return 0;
 682
 683         ret = sysfs_create_link(&node_devices[nid]->dev.kobj,
 684                                 &obj->kobj,
 685                                 kobject_name(&obj->kobj));
 686         if (ret)
 687                 return ret;
 688
 689         return sysfs_create_link(&obj->kobj,
 690                                  &node_devices[nid]->dev.kobj,
 691                                  kobject_name(&node_devices[nid]->dev.kobj));
 692 }
 693
 694 /**
 695  * register_memory_node_under_compute_node - link memory node to its compute
 696  *                                           node for a given access class.
 697  * @mem_nid:    Memory node number
 698  * @cpu_nid:    Cpu  node number
 699  * @access:     Access class to register
 700  *
 701  * Description:
 702  *      For use with platforms that may have separate memory and compute nodes.
 703  *      This function will export node relationships linking which memory
 704  *      initiator nodes can access memory targets at a given ranked access
 705  *      class.
 706  */
 707 int register_memory_node_under_compute_node(unsigned int mem_nid,
 708                                             unsigned int cpu_nid,
 709                                             unsigned access)
 710 {
 711         struct node *init_node, *targ_node;
 712         struct node_access_nodes *initiator, *target;
 713         int ret;
 714
 715         if (!node_online(cpu_nid) || !node_online(mem_nid))
 716                 return -ENODEV;
 717
 718         init_node = node_devices[cpu_nid];
 719         targ_node = node_devices[mem_nid];
 720         initiator = node_init_node_access(init_node, access);
 721         target = node_init_node_access(targ_node, access);
 722         if (!initiator || !target)
 723                 return -ENOMEM;
 724
 725         ret = sysfs_add_link_to_group(&initiator->dev.kobj, "targets",
 726                                       &targ_node->dev.kobj,
 727                                       dev_name(&targ_node->dev));
 728         if (ret)
 729                 return ret;
 730
 731         ret = sysfs_add_link_to_group(&target->dev.kobj, "initiators",
 732                                       &init_node->dev.kobj,
 733                                       dev_name(&init_node->dev));
 734         if (ret)
 735                 goto err;
 736
 737         return 0;
 738  err:
 739         sysfs_remove_link_from_group(&initiator->dev.kobj, "targets",
 740                                      dev_name(&targ_node->dev));
 741         return ret;
 742 }
 743
 744 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
 745 {
 746         struct device *obj;
 747
 748         if (!node_online(nid))
 749                 return 0;
 750
 751         obj = get_cpu_device(cpu);
 752         if (!obj)
 753                 return 0;
 754
 755         sysfs_remove_link(&node_devices[nid]->dev.kobj,
 756                           kobject_name(&obj->kobj));
 757         sysfs_remove_link(&obj->kobj,
 758                           kobject_name(&node_devices[nid]->dev.kobj));
 759
 760         return 0;
 761 }
 762
 763 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
 764 static int __ref get_nid_for_pfn(unsigned long pfn)
 765 {
 766         if (!pfn_valid_within(pfn))
 767                 return -1;
 768 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
 769         if (system_state < SYSTEM_RUNNING)
 770                 return early_pfn_to_nid(pfn);
 771 #endif
 772         return pfn_to_nid(pfn);
 773 }
 774
 775 static void do_register_memory_block_under_node(int nid,
 776                                                 struct memory_block *mem_blk)
 777 {
 778         int ret;
 779
 780         /*
 781          * If this memory block spans multiple nodes, we only indicate
 782          * the last processed node.
 783          */
 784         mem_blk->nid = nid;
 785
 786         ret = sysfs_create_link_nowarn(&node_devices[nid]->dev.kobj,
 787                                        &mem_blk->dev.kobj,
 788                                        kobject_name(&mem_blk->dev.kobj));
 789         if (ret && ret != -EEXIST)
 790                 dev_err_ratelimited(&node_devices[nid]->dev,
 791                                     "can't create link to %s in sysfs (%d)\n",
 792                                     kobject_name(&mem_blk->dev.kobj), ret);
 793
 794         ret = sysfs_create_link_nowarn(&mem_blk->dev.kobj,
 795                                 &node_devices[nid]->dev.kobj,
 796                                 kobject_name(&node_devices[nid]->dev.kobj));
 797         if (ret && ret != -EEXIST)
 798                 dev_err_ratelimited(&mem_blk->dev,
 799                                     "can't create link to %s in sysfs (%d)\n",
 800                                     kobject_name(&node_devices[nid]->dev.kobj),
 801                                     ret);
 802 }
 803
 804 /* register memory section under specified node if it spans that node */
 805 static int register_mem_block_under_node_early(struct memory_block *mem_blk,
 806                                                void *arg)
 807 {
 808         unsigned long memory_block_pfns = memory_block_size_bytes() / PAGE_SIZE;
 809         unsigned long start_pfn = section_nr_to_pfn(mem_blk->start_section_nr);
 810         unsigned long end_pfn = start_pfn + memory_block_pfns - 1;
 811         int nid = *(int *)arg;
 812         unsigned long pfn;
 813
 814         for (pfn = start_pfn; pfn <= end_pfn; pfn++) {
 815                 int page_nid;
 816
 817                 /*
 818                  * memory block could have several absent sections from start.
 819                  * skip pfn range from absent section
 820                  */
 821                 if (!pfn_in_present_section(pfn)) {
 822                         pfn = round_down(pfn + PAGES_PER_SECTION,
 823                                          PAGES_PER_SECTION) - 1;
 824                         continue;
 825                 }
 826
 827                 /*
 828                  * We need to check if page belongs to nid only at the boot
 829                  * case because node's ranges can be interleaved.
 830                  */
 831                 page_nid = get_nid_for_pfn(pfn);
 832                 if (page_nid < 0)
 833                         continue;
 834                 if (page_nid != nid)
 835                         continue;
 836
 837                 do_register_memory_block_under_node(nid, mem_blk);
 838                 return 0;
 839         }
 840         /* mem section does not span the specified node */
 841         return 0;
 842 }
 843
 844 /*
 845  * During hotplug we know that all pages in the memory block belong to the same
 846  * node.
 847  */
 848 static int register_mem_block_under_node_hotplug(struct memory_block *mem_blk,
 849                                                  void *arg)
 850 {
 851         int nid = *(int *)arg;
 852
 853         do_register_memory_block_under_node(nid, mem_blk);
 854         return 0;
 855 }
 856
 857 /*
 858  * Unregister a memory block device under the node it spans. Memory blocks
 859  * with multiple nodes cannot be offlined and therefore also never be removed.
 860  */
 861 void unregister_memory_block_under_nodes(struct memory_block *mem_blk)
 862 {
 863         if (mem_blk->nid == NUMA_NO_NODE)
 864                 return;
 865
 866         sysfs_remove_link(&node_devices[mem_blk->nid]->dev.kobj,
 867                           kobject_name(&mem_blk->dev.kobj));
 868         sysfs_remove_link(&mem_blk->dev.kobj,
 869                           kobject_name(&node_devices[mem_blk->nid]->dev.kobj));
 870 }
 871
 872 void link_mem_sections(int nid, unsigned long start_pfn, unsigned long end_pfn,
 873                        enum meminit_context context)
 874 {
 875         walk_memory_blocks_func_t func;
 876
 877         if (context == MEMINIT_HOTPLUG)
 878                 func = register_mem_block_under_node_hotplug;
 879         else
 880                 func = register_mem_block_under_node_early;
 881
 882         walk_memory_blocks(PFN_PHYS(start_pfn), PFN_PHYS(end_pfn - start_pfn),
 883                            (void *)&nid, func);
 884         return;
 885 }
 886
 887 #ifdef CONFIG_HUGETLBFS
 888 /*
 889  * Handle per node hstate attribute [un]registration on transistions
 890  * to/from memoryless state.
 891  */
 892 static void node_hugetlb_work(struct work_struct *work)
 893 {
 894         struct node *node = container_of(work, struct node, node_work);
 895
 896         /*
 897          * We only get here when a node transitions to/from memoryless state.
 898          * We can detect which transition occurred by examining whether the
 899          * node has memory now.  hugetlb_register_node() already check this
 900          * so we try to register the attributes.  If that fails, then the
 901          * node has transitioned to memoryless, try to unregister the
 902          * attributes.
 903          */
 904         if (!hugetlb_register_node(node))
 905                 hugetlb_unregister_node(node);
 906 }
 907
 908 static void init_node_hugetlb_work(int nid)
 909 {
 910         INIT_WORK(&node_devices[nid]->node_work, node_hugetlb_work);
 911 }
 912
 913 static int node_memory_callback(struct notifier_block *self,
 914                                 unsigned long action, void *arg)
 915 {
 916         struct memory_notify *mnb = arg;
 917         int nid = mnb->status_change_nid;
 918
 919         switch (action) {
 920         case MEM_ONLINE:
 921         case MEM_OFFLINE:
 922                 /*
 923                  * offload per node hstate [un]registration to a work thread
 924                  * when transitioning to/from memoryless state.
 925                  */
 926                 if (nid != NUMA_NO_NODE)
 927                         schedule_work(&node_devices[nid]->node_work);
 928                 break;
 929
 930         case MEM_GOING_ONLINE:
 931         case MEM_GOING_OFFLINE:
 932         case MEM_CANCEL_ONLINE:
 933         case MEM_CANCEL_OFFLINE:
 934         default:
 935                 break;
 936         }
 937
 938         return NOTIFY_OK;
 939 }
 940 #endif  /* CONFIG_HUGETLBFS */
 941 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
 942
 943 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
 944     !defined(CONFIG_HUGETLBFS)
 945 static inline int node_memory_callback(struct notifier_block *self,
 946                                 unsigned long action, void *arg)
 947 {
 948         return NOTIFY_OK;
 949 }
 950
 951 static void init_node_hugetlb_work(int nid) { }
 952
 953 #endif
 954
 955 int __register_one_node(int nid)
 956 {
 957         int error;
 958         int cpu;
 959
 960         node_devices[nid] = kzalloc(sizeof(struct node), GFP_KERNEL);
 961         if (!node_devices[nid])
 962                 return -ENOMEM;
 963
 964         error = register_node(node_devices[nid], nid);
 965
 966         /* link cpu under this node */
 967         for_each_present_cpu(cpu) {
 968                 if (cpu_to_node(cpu) == nid)
 969                         register_cpu_under_node(cpu, nid);
 970         }
 971
 972         INIT_LIST_HEAD(&node_devices[nid]->access_list);
 973         /* initialize work queue for memory hot plug */
 974         init_node_hugetlb_work(nid);
 975         node_init_caches(nid);
 976
 977         return error;
 978 }
 979
 980 void unregister_one_node(int nid)
 981 {
 982         if (!node_devices[nid])
 983                 return;
 984
 985         unregister_node(node_devices[nid]);
 986         node_devices[nid] = NULL;
 987 }
 988
 989 /*
 990  * node states attributes
 991  */
 992
 993 struct node_attr {
 994         struct device_attribute attr;
 995         enum node_states state;
 996 };
 997
 998 static ssize_t show_node_state(struct device *dev,
 999                                struct device_attribute *attr, char *buf)
1000 {
1001         struct node_attr *na = container_of(attr, struct node_attr, attr);
1002
1003         return sysfs_emit(buf, "%*pbl\n",
1004                           nodemask_pr_args(&node_states[na->state]));
1005 }
1006
1007 #define _NODE_ATTR(name, state) \
1008         { __ATTR(name, 0444, show_node_state, NULL), state }
1009
1010 static struct node_attr node_state_attr[] = {
1011         [N_POSSIBLE] = _NODE_ATTR(possible, N_POSSIBLE),
1012         [N_ONLINE] = _NODE_ATTR(online, N_ONLINE),
1013         [N_NORMAL_MEMORY] = _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
1014 #ifdef CONFIG_HIGHMEM
1015         [N_HIGH_MEMORY] = _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
1016 #endif
1017         [N_MEMORY] = _NODE_ATTR(has_memory, N_MEMORY),
1018         [N_CPU] = _NODE_ATTR(has_cpu, N_CPU),
1019         [N_GENERIC_INITIATOR] = _NODE_ATTR(has_generic_initiator,
1020                                            N_GENERIC_INITIATOR),
1021 };
1022
1023 static struct attribute *node_state_attrs[] = {
1024         &node_state_attr[N_POSSIBLE].attr.attr,
1025         &node_state_attr[N_ONLINE].attr.attr,
1026         &node_state_attr[N_NORMAL_MEMORY].attr.attr,
1027 #ifdef CONFIG_HIGHMEM
1028         &node_state_attr[N_HIGH_MEMORY].attr.attr,
1029 #endif
1030         &node_state_attr[N_MEMORY].attr.attr,
1031         &node_state_attr[N_CPU].attr.attr,
1032         &node_state_attr[N_GENERIC_INITIATOR].attr.attr,
1033         NULL
1034 };
1035
1036 static struct attribute_group memory_root_attr_group = {
1037         .attrs = node_state_attrs,
1038 };
1039
1040 static const struct attribute_group *cpu_root_attr_groups[] = {
1041         &memory_root_attr_group,
1042         NULL,
1043 };
1044
1045 #define NODE_CALLBACK_PRI       2       /* lower than SLAB */
1046 static int __init register_node_type(void)
1047 {
1048         int ret;
1049
1050         BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
1051         BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
1052
1053         ret = subsys_system_register(&node_subsys, cpu_root_attr_groups);
1054         if (!ret) {
1055                 static struct notifier_block node_memory_callback_nb = {
1056                         .notifier_call = node_memory_callback,
1057                         .priority = NODE_CALLBACK_PRI,
1058                 };
1059                 register_hotmemory_notifier(&node_memory_callback_nb);
1060         }
1061
1062         /*
1063          * Note:  we're not going to unregister the node class if we fail
1064          * to register the node state class attribute files.
1065          */
1066         return ret;
1067 }
1068 postcore_initcall(register_node_type);