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(node_page_state(pgdat, 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 nid, K(node_page_state(pgdat, NR_SHMEM_THPS)), 466 nid, K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED) * 467 HPAGE_PMD_NR), 468 nid, K(node_page_state(pgdat, NR_FILE_THPS)), 469 nid, K(node_page_state(pgdat, NR_FILE_PMDMAPPED) * 470 HPAGE_PMD_NR) 471 #endif 472 ); 473 len += hugetlb_report_node_meminfo(buf, len, nid); 474 return len; 475 } 476 477 #undef K 478 static DEVICE_ATTR(meminfo, 0444, node_read_meminfo, NULL); 479 480 static ssize_t node_read_numastat(struct device *dev, 481 struct device_attribute *attr, char *buf) 482 { 483 return sysfs_emit(buf, 484 "numa_hit %lu\n" 485 "numa_miss %lu\n" 486 "numa_foreign %lu\n" 487 "interleave_hit %lu\n" 488 "local_node %lu\n" 489 "other_node %lu\n", 490 sum_zone_numa_state(dev->id, NUMA_HIT), 491 sum_zone_numa_state(dev->id, NUMA_MISS), 492 sum_zone_numa_state(dev->id, NUMA_FOREIGN), 493 sum_zone_numa_state(dev->id, NUMA_INTERLEAVE_HIT), 494 sum_zone_numa_state(dev->id, NUMA_LOCAL), 495 sum_zone_numa_state(dev->id, NUMA_OTHER)); 496 } 497 static DEVICE_ATTR(numastat, 0444, node_read_numastat, NULL); 498 499 static ssize_t node_read_vmstat(struct device *dev, 500 struct device_attribute *attr, char *buf) 501 { 502 int nid = dev->id; 503 struct pglist_data *pgdat = NODE_DATA(nid); 504 int i; 505 int len = 0; 506 507 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) 508 len += sysfs_emit_at(buf, len, "%s %lu\n", 509 zone_stat_name(i), 510 sum_zone_node_page_state(nid, i)); 511 512 #ifdef CONFIG_NUMA 513 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) 514 len += sysfs_emit_at(buf, len, "%s %lu\n", 515 numa_stat_name(i), 516 sum_zone_numa_state(nid, i)); 517 518 #endif 519 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) { 520 unsigned long pages = node_page_state_pages(pgdat, i); 521 522 if (vmstat_item_print_in_thp(i)) 523 pages /= HPAGE_PMD_NR; 524 len += sysfs_emit_at(buf, len, "%s %lu\n", node_stat_name(i), 525 pages); 526 } 527 528 return len; 529 } 530 static DEVICE_ATTR(vmstat, 0444, node_read_vmstat, NULL); 531 532 static ssize_t node_read_distance(struct device *dev, 533 struct device_attribute *attr, char *buf) 534 { 535 int nid = dev->id; 536 int len = 0; 537 int i; 538 539 /* 540 * buf is currently PAGE_SIZE in length and each node needs 4 chars 541 * at the most (distance + space or newline). 542 */ 543 BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE); 544 545 for_each_online_node(i) { 546 len += sysfs_emit_at(buf, len, "%s%d", 547 i ? " " : "", node_distance(nid, i)); 548 } 549 550 len += sysfs_emit_at(buf, len, "\n"); 551 return len; 552 } 553 static DEVICE_ATTR(distance, 0444, node_read_distance, NULL); 554 555 static struct attribute *node_dev_attrs[] = { 556 &dev_attr_cpumap.attr, 557 &dev_attr_cpulist.attr, 558 &dev_attr_meminfo.attr, 559 &dev_attr_numastat.attr, 560 &dev_attr_distance.attr, 561 &dev_attr_vmstat.attr, 562 NULL 563 }; 564 ATTRIBUTE_GROUPS(node_dev); 565 566 #ifdef CONFIG_HUGETLBFS 567 /* 568 * hugetlbfs per node attributes registration interface: 569 * When/if hugetlb[fs] subsystem initializes [sometime after this module], 570 * it will register its per node attributes for all online nodes with 571 * memory. It will also call register_hugetlbfs_with_node(), below, to 572 * register its attribute registration functions with this node driver. 573 * Once these hooks have been initialized, the node driver will call into 574 * the hugetlb module to [un]register attributes for hot-plugged nodes. 575 */ 576 static node_registration_func_t __hugetlb_register_node; 577 static node_registration_func_t __hugetlb_unregister_node; 578 579 static inline bool hugetlb_register_node(struct node *node) 580 { 581 if (__hugetlb_register_node && 582 node_state(node->dev.id, N_MEMORY)) { 583 __hugetlb_register_node(node); 584 return true; 585 } 586 return false; 587 } 588 589 static inline void hugetlb_unregister_node(struct node *node) 590 { 591 if (__hugetlb_unregister_node) 592 __hugetlb_unregister_node(node); 593 } 594 595 void register_hugetlbfs_with_node(node_registration_func_t doregister, 596 node_registration_func_t unregister) 597 { 598 __hugetlb_register_node = doregister; 599 __hugetlb_unregister_node = unregister; 600 } 601 #else 602 static inline void hugetlb_register_node(struct node *node) {} 603 604 static inline void hugetlb_unregister_node(struct node *node) {} 605 #endif 606 607 static void node_device_release(struct device *dev) 608 { 609 struct node *node = to_node(dev); 610 611 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HUGETLBFS) 612 /* 613 * We schedule the work only when a memory section is 614 * onlined/offlined on this node. When we come here, 615 * all the memory on this node has been offlined, 616 * so we won't enqueue new work to this work. 617 * 618 * The work is using node->node_work, so we should 619 * flush work before freeing the memory. 620 */ 621 flush_work(&node->node_work); 622 #endif 623 kfree(node); 624 } 625 626 /* 627 * register_node - Setup a sysfs device for a node. 628 * @num - Node number to use when creating the device. 629 * 630 * Initialize and register the node device. 631 */ 632 static int register_node(struct node *node, int num) 633 { 634 int error; 635 636 node->dev.id = num; 637 node->dev.bus = &node_subsys; 638 node->dev.release = node_device_release; 639 node->dev.groups = node_dev_groups; 640 error = device_register(&node->dev); 641 642 if (error) 643 put_device(&node->dev); 644 else { 645 hugetlb_register_node(node); 646 647 compaction_register_node(node); 648 } 649 return error; 650 } 651 652 /** 653 * unregister_node - unregister a node device 654 * @node: node going away 655 * 656 * Unregisters a node device @node. All the devices on the node must be 657 * unregistered before calling this function. 658 */ 659 void unregister_node(struct node *node) 660 { 661 hugetlb_unregister_node(node); /* no-op, if memoryless node */ 662 node_remove_accesses(node); 663 node_remove_caches(node); 664 device_unregister(&node->dev); 665 } 666 667 struct node *node_devices[MAX_NUMNODES]; 668 669 /* 670 * register cpu under node 671 */ 672 int register_cpu_under_node(unsigned int cpu, unsigned int nid) 673 { 674 int ret; 675 struct device *obj; 676 677 if (!node_online(nid)) 678 return 0; 679 680 obj = get_cpu_device(cpu); 681 if (!obj) 682 return 0; 683 684 ret = sysfs_create_link(&node_devices[nid]->dev.kobj, 685 &obj->kobj, 686 kobject_name(&obj->kobj)); 687 if (ret) 688 return ret; 689 690 return sysfs_create_link(&obj->kobj, 691 &node_devices[nid]->dev.kobj, 692 kobject_name(&node_devices[nid]->dev.kobj)); 693 } 694 695 /** 696 * register_memory_node_under_compute_node - link memory node to its compute 697 * node for a given access class. 698 * @mem_nid: Memory node number 699 * @cpu_nid: Cpu node number 700 * @access: Access class to register 701 * 702 * Description: 703 * For use with platforms that may have separate memory and compute nodes. 704 * This function will export node relationships linking which memory 705 * initiator nodes can access memory targets at a given ranked access 706 * class. 707 */ 708 int register_memory_node_under_compute_node(unsigned int mem_nid, 709 unsigned int cpu_nid, 710 unsigned access) 711 { 712 struct node *init_node, *targ_node; 713 struct node_access_nodes *initiator, *target; 714 int ret; 715 716 if (!node_online(cpu_nid) || !node_online(mem_nid)) 717 return -ENODEV; 718 719 init_node = node_devices[cpu_nid]; 720 targ_node = node_devices[mem_nid]; 721 initiator = node_init_node_access(init_node, access); 722 target = node_init_node_access(targ_node, access); 723 if (!initiator || !target) 724 return -ENOMEM; 725 726 ret = sysfs_add_link_to_group(&initiator->dev.kobj, "targets", 727 &targ_node->dev.kobj, 728 dev_name(&targ_node->dev)); 729 if (ret) 730 return ret; 731 732 ret = sysfs_add_link_to_group(&target->dev.kobj, "initiators", 733 &init_node->dev.kobj, 734 dev_name(&init_node->dev)); 735 if (ret) 736 goto err; 737 738 return 0; 739 err: 740 sysfs_remove_link_from_group(&initiator->dev.kobj, "targets", 741 dev_name(&targ_node->dev)); 742 return ret; 743 } 744 745 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid) 746 { 747 struct device *obj; 748 749 if (!node_online(nid)) 750 return 0; 751 752 obj = get_cpu_device(cpu); 753 if (!obj) 754 return 0; 755 756 sysfs_remove_link(&node_devices[nid]->dev.kobj, 757 kobject_name(&obj->kobj)); 758 sysfs_remove_link(&obj->kobj, 759 kobject_name(&node_devices[nid]->dev.kobj)); 760 761 return 0; 762 } 763 764 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE 765 static int __ref get_nid_for_pfn(unsigned long pfn) 766 { 767 if (!pfn_valid_within(pfn)) 768 return -1; 769 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT 770 if (system_state < SYSTEM_RUNNING) 771 return early_pfn_to_nid(pfn); 772 #endif 773 return pfn_to_nid(pfn); 774 } 775 776 static void do_register_memory_block_under_node(int nid, 777 struct memory_block *mem_blk) 778 { 779 int ret; 780 781 /* 782 * If this memory block spans multiple nodes, we only indicate 783 * the last processed node. 784 */ 785 mem_blk->nid = nid; 786 787 ret = sysfs_create_link_nowarn(&node_devices[nid]->dev.kobj, 788 &mem_blk->dev.kobj, 789 kobject_name(&mem_blk->dev.kobj)); 790 if (ret && ret != -EEXIST) 791 dev_err_ratelimited(&node_devices[nid]->dev, 792 "can't create link to %s in sysfs (%d)\n", 793 kobject_name(&mem_blk->dev.kobj), ret); 794 795 ret = sysfs_create_link_nowarn(&mem_blk->dev.kobj, 796 &node_devices[nid]->dev.kobj, 797 kobject_name(&node_devices[nid]->dev.kobj)); 798 if (ret && ret != -EEXIST) 799 dev_err_ratelimited(&mem_blk->dev, 800 "can't create link to %s in sysfs (%d)\n", 801 kobject_name(&node_devices[nid]->dev.kobj), 802 ret); 803 } 804 805 /* register memory section under specified node if it spans that node */ 806 static int register_mem_block_under_node_early(struct memory_block *mem_blk, 807 void *arg) 808 { 809 unsigned long memory_block_pfns = memory_block_size_bytes() / PAGE_SIZE; 810 unsigned long start_pfn = section_nr_to_pfn(mem_blk->start_section_nr); 811 unsigned long end_pfn = start_pfn + memory_block_pfns - 1; 812 int nid = *(int *)arg; 813 unsigned long pfn; 814 815 for (pfn = start_pfn; pfn <= end_pfn; pfn++) { 816 int page_nid; 817 818 /* 819 * memory block could have several absent sections from start. 820 * skip pfn range from absent section 821 */ 822 if (!pfn_in_present_section(pfn)) { 823 pfn = round_down(pfn + PAGES_PER_SECTION, 824 PAGES_PER_SECTION) - 1; 825 continue; 826 } 827 828 /* 829 * We need to check if page belongs to nid only at the boot 830 * case because node's ranges can be interleaved. 831 */ 832 page_nid = get_nid_for_pfn(pfn); 833 if (page_nid < 0) 834 continue; 835 if (page_nid != nid) 836 continue; 837 838 do_register_memory_block_under_node(nid, mem_blk); 839 return 0; 840 } 841 /* mem section does not span the specified node */ 842 return 0; 843 } 844 845 /* 846 * During hotplug we know that all pages in the memory block belong to the same 847 * node. 848 */ 849 static int register_mem_block_under_node_hotplug(struct memory_block *mem_blk, 850 void *arg) 851 { 852 int nid = *(int *)arg; 853 854 do_register_memory_block_under_node(nid, mem_blk); 855 return 0; 856 } 857 858 /* 859 * Unregister a memory block device under the node it spans. Memory blocks 860 * with multiple nodes cannot be offlined and therefore also never be removed. 861 */ 862 void unregister_memory_block_under_nodes(struct memory_block *mem_blk) 863 { 864 if (mem_blk->nid == NUMA_NO_NODE) 865 return; 866 867 sysfs_remove_link(&node_devices[mem_blk->nid]->dev.kobj, 868 kobject_name(&mem_blk->dev.kobj)); 869 sysfs_remove_link(&mem_blk->dev.kobj, 870 kobject_name(&node_devices[mem_blk->nid]->dev.kobj)); 871 } 872 873 void link_mem_sections(int nid, unsigned long start_pfn, unsigned long end_pfn, 874 enum meminit_context context) 875 { 876 walk_memory_blocks_func_t func; 877 878 if (context == MEMINIT_HOTPLUG) 879 func = register_mem_block_under_node_hotplug; 880 else 881 func = register_mem_block_under_node_early; 882 883 walk_memory_blocks(PFN_PHYS(start_pfn), PFN_PHYS(end_pfn - start_pfn), 884 (void *)&nid, func); 885 return; 886 } 887 888 #ifdef CONFIG_HUGETLBFS 889 /* 890 * Handle per node hstate attribute [un]registration on transistions 891 * to/from memoryless state. 892 */ 893 static void node_hugetlb_work(struct work_struct *work) 894 { 895 struct node *node = container_of(work, struct node, node_work); 896 897 /* 898 * We only get here when a node transitions to/from memoryless state. 899 * We can detect which transition occurred by examining whether the 900 * node has memory now. hugetlb_register_node() already check this 901 * so we try to register the attributes. If that fails, then the 902 * node has transitioned to memoryless, try to unregister the 903 * attributes. 904 */ 905 if (!hugetlb_register_node(node)) 906 hugetlb_unregister_node(node); 907 } 908 909 static void init_node_hugetlb_work(int nid) 910 { 911 INIT_WORK(&node_devices[nid]->node_work, node_hugetlb_work); 912 } 913 914 static int node_memory_callback(struct notifier_block *self, 915 unsigned long action, void *arg) 916 { 917 struct memory_notify *mnb = arg; 918 int nid = mnb->status_change_nid; 919 920 switch (action) { 921 case MEM_ONLINE: 922 case MEM_OFFLINE: 923 /* 924 * offload per node hstate [un]registration to a work thread 925 * when transitioning to/from memoryless state. 926 */ 927 if (nid != NUMA_NO_NODE) 928 schedule_work(&node_devices[nid]->node_work); 929 break; 930 931 case MEM_GOING_ONLINE: 932 case MEM_GOING_OFFLINE: 933 case MEM_CANCEL_ONLINE: 934 case MEM_CANCEL_OFFLINE: 935 default: 936 break; 937 } 938 939 return NOTIFY_OK; 940 } 941 #endif /* CONFIG_HUGETLBFS */ 942 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */ 943 944 #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \ 945 !defined(CONFIG_HUGETLBFS) 946 static inline int node_memory_callback(struct notifier_block *self, 947 unsigned long action, void *arg) 948 { 949 return NOTIFY_OK; 950 } 951 952 static void init_node_hugetlb_work(int nid) { } 953 954 #endif 955 956 int __register_one_node(int nid) 957 { 958 int error; 959 int cpu; 960 961 node_devices[nid] = kzalloc(sizeof(struct node), GFP_KERNEL); 962 if (!node_devices[nid]) 963 return -ENOMEM; 964 965 error = register_node(node_devices[nid], nid); 966 967 /* link cpu under this node */ 968 for_each_present_cpu(cpu) { 969 if (cpu_to_node(cpu) == nid) 970 register_cpu_under_node(cpu, nid); 971 } 972 973 INIT_LIST_HEAD(&node_devices[nid]->access_list); 974 /* initialize work queue for memory hot plug */ 975 init_node_hugetlb_work(nid); 976 node_init_caches(nid); 977 978 return error; 979 } 980 981 void unregister_one_node(int nid) 982 { 983 if (!node_devices[nid]) 984 return; 985 986 unregister_node(node_devices[nid]); 987 node_devices[nid] = NULL; 988 } 989 990 /* 991 * node states attributes 992 */ 993 994 struct node_attr { 995 struct device_attribute attr; 996 enum node_states state; 997 }; 998 999 static ssize_t show_node_state(struct device *dev, 1000 struct device_attribute *attr, char *buf) 1001 { 1002 struct node_attr *na = container_of(attr, struct node_attr, attr); 1003 1004 return sysfs_emit(buf, "%*pbl\n", 1005 nodemask_pr_args(&node_states[na->state])); 1006 } 1007 1008 #define _NODE_ATTR(name, state) \ 1009 { __ATTR(name, 0444, show_node_state, NULL), state } 1010 1011 static struct node_attr node_state_attr[] = { 1012 [N_POSSIBLE] = _NODE_ATTR(possible, N_POSSIBLE), 1013 [N_ONLINE] = _NODE_ATTR(online, N_ONLINE), 1014 [N_NORMAL_MEMORY] = _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY), 1015 #ifdef CONFIG_HIGHMEM 1016 [N_HIGH_MEMORY] = _NODE_ATTR(has_high_memory, N_HIGH_MEMORY), 1017 #endif 1018 [N_MEMORY] = _NODE_ATTR(has_memory, N_MEMORY), 1019 [N_CPU] = _NODE_ATTR(has_cpu, N_CPU), 1020 [N_GENERIC_INITIATOR] = _NODE_ATTR(has_generic_initiator, 1021 N_GENERIC_INITIATOR), 1022 }; 1023 1024 static struct attribute *node_state_attrs[] = { 1025 &node_state_attr[N_POSSIBLE].attr.attr, 1026 &node_state_attr[N_ONLINE].attr.attr, 1027 &node_state_attr[N_NORMAL_MEMORY].attr.attr, 1028 #ifdef CONFIG_HIGHMEM 1029 &node_state_attr[N_HIGH_MEMORY].attr.attr, 1030 #endif 1031 &node_state_attr[N_MEMORY].attr.attr, 1032 &node_state_attr[N_CPU].attr.attr, 1033 &node_state_attr[N_GENERIC_INITIATOR].attr.attr, 1034 NULL 1035 }; 1036 1037 static struct attribute_group memory_root_attr_group = { 1038 .attrs = node_state_attrs, 1039 }; 1040 1041 static const struct attribute_group *cpu_root_attr_groups[] = { 1042 &memory_root_attr_group, 1043 NULL, 1044 }; 1045 1046 #define NODE_CALLBACK_PRI 2 /* lower than SLAB */ 1047 static int __init register_node_type(void) 1048 { 1049 int ret; 1050 1051 BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES); 1052 BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES); 1053 1054 ret = subsys_system_register(&node_subsys, cpu_root_attr_groups); 1055 if (!ret) { 1056 static struct notifier_block node_memory_callback_nb = { 1057 .notifier_call = node_memory_callback, 1058 .priority = NODE_CALLBACK_PRI, 1059 }; 1060 register_hotmemory_notifier(&node_memory_callback_nb); 1061 } 1062 1063 /* 1064 * Note: we're not going to unregister the node class if we fail 1065 * to register the node state class attribute files. 1066 */ 1067 return ret; 1068 } 1069 postcore_initcall(register_node_type); 1070