11da177e4SLinus Torvalds /* 210fbcf4cSKay Sievers * Basic Node interface support 31da177e4SLinus Torvalds */ 41da177e4SLinus Torvalds 51da177e4SLinus Torvalds #include <linux/module.h> 61da177e4SLinus Torvalds #include <linux/init.h> 71da177e4SLinus Torvalds #include <linux/mm.h> 8c04fc586SGary Hade #include <linux/memory.h> 9fa25c503SKOSAKI Motohiro #include <linux/vmstat.h> 106e259e7dSAndrew Morton #include <linux/notifier.h> 111da177e4SLinus Torvalds #include <linux/node.h> 121da177e4SLinus Torvalds #include <linux/hugetlb.h> 13ed4a6d7fSMel Gorman #include <linux/compaction.h> 141da177e4SLinus Torvalds #include <linux/cpumask.h> 151da177e4SLinus Torvalds #include <linux/topology.h> 161da177e4SLinus Torvalds #include <linux/nodemask.h> 1776b67ed9SKAMEZAWA Hiroyuki #include <linux/cpu.h> 18bde631a5SLee Schermerhorn #include <linux/device.h> 19af936a16SLee Schermerhorn #include <linux/swap.h> 2018e5b539STejun Heo #include <linux/slab.h> 211da177e4SLinus Torvalds 2210fbcf4cSKay Sievers static struct bus_type node_subsys = { 23af5ca3f4SKay Sievers .name = "node", 2410fbcf4cSKay Sievers .dev_name = "node", 251da177e4SLinus Torvalds }; 261da177e4SLinus Torvalds 271da177e4SLinus Torvalds 285aaba363SSudeep Holla static ssize_t node_read_cpumap(struct device *dev, bool list, char *buf) 291da177e4SLinus Torvalds { 301da177e4SLinus Torvalds struct node *node_dev = to_node(dev); 3110fbcf4cSKay Sievers const struct cpumask *mask = cpumask_of_node(node_dev->dev.id); 321da177e4SLinus Torvalds 3339106dcfSMike Travis /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */ 3439106dcfSMike Travis BUILD_BUG_ON((NR_CPUS/32 * 9) > (PAGE_SIZE-1)); 351da177e4SLinus Torvalds 365aaba363SSudeep Holla return cpumap_print_to_pagebuf(list, buf, mask); 371da177e4SLinus Torvalds } 381da177e4SLinus Torvalds 3910fbcf4cSKay Sievers static inline ssize_t node_read_cpumask(struct device *dev, 4010fbcf4cSKay Sievers struct device_attribute *attr, char *buf) 4139106dcfSMike Travis { 425aaba363SSudeep Holla return node_read_cpumap(dev, false, buf); 4339106dcfSMike Travis } 4410fbcf4cSKay Sievers static inline ssize_t node_read_cpulist(struct device *dev, 4510fbcf4cSKay Sievers struct device_attribute *attr, char *buf) 4639106dcfSMike Travis { 475aaba363SSudeep Holla return node_read_cpumap(dev, true, buf); 4839106dcfSMike Travis } 4939106dcfSMike Travis 5010fbcf4cSKay Sievers static DEVICE_ATTR(cpumap, S_IRUGO, node_read_cpumask, NULL); 5110fbcf4cSKay Sievers static DEVICE_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL); 521da177e4SLinus Torvalds 531da177e4SLinus Torvalds #define K(x) ((x) << (PAGE_SHIFT - 10)) 5410fbcf4cSKay Sievers static ssize_t node_read_meminfo(struct device *dev, 5510fbcf4cSKay Sievers struct device_attribute *attr, char *buf) 561da177e4SLinus Torvalds { 571da177e4SLinus Torvalds int n; 581da177e4SLinus Torvalds int nid = dev->id; 591da177e4SLinus Torvalds struct sysinfo i; 601da177e4SLinus Torvalds 611da177e4SLinus Torvalds si_meminfo_node(&i, nid); 627ee92255SKOSAKI Motohiro n = sprintf(buf, 631da177e4SLinus Torvalds "Node %d MemTotal: %8lu kB\n" 641da177e4SLinus Torvalds "Node %d MemFree: %8lu kB\n" 651da177e4SLinus Torvalds "Node %d MemUsed: %8lu kB\n" 661da177e4SLinus Torvalds "Node %d Active: %8lu kB\n" 671da177e4SLinus Torvalds "Node %d Inactive: %8lu kB\n" 684f98a2feSRik van Riel "Node %d Active(anon): %8lu kB\n" 694f98a2feSRik van Riel "Node %d Inactive(anon): %8lu kB\n" 704f98a2feSRik van Riel "Node %d Active(file): %8lu kB\n" 714f98a2feSRik van Riel "Node %d Inactive(file): %8lu kB\n" 725344b7e6SNick Piggin "Node %d Unevictable: %8lu kB\n" 737ee92255SKOSAKI Motohiro "Node %d Mlocked: %8lu kB\n", 747ee92255SKOSAKI Motohiro nid, K(i.totalram), 757ee92255SKOSAKI Motohiro nid, K(i.freeram), 767ee92255SKOSAKI Motohiro nid, K(i.totalram - i.freeram), 777ee92255SKOSAKI Motohiro nid, K(node_page_state(nid, NR_ACTIVE_ANON) + 787ee92255SKOSAKI Motohiro node_page_state(nid, NR_ACTIVE_FILE)), 797ee92255SKOSAKI Motohiro nid, K(node_page_state(nid, NR_INACTIVE_ANON) + 807ee92255SKOSAKI Motohiro node_page_state(nid, NR_INACTIVE_FILE)), 817ee92255SKOSAKI Motohiro nid, K(node_page_state(nid, NR_ACTIVE_ANON)), 827ee92255SKOSAKI Motohiro nid, K(node_page_state(nid, NR_INACTIVE_ANON)), 837ee92255SKOSAKI Motohiro nid, K(node_page_state(nid, NR_ACTIVE_FILE)), 847ee92255SKOSAKI Motohiro nid, K(node_page_state(nid, NR_INACTIVE_FILE)), 857ee92255SKOSAKI Motohiro nid, K(node_page_state(nid, NR_UNEVICTABLE)), 867ee92255SKOSAKI Motohiro nid, K(node_page_state(nid, NR_MLOCK))); 877ee92255SKOSAKI Motohiro 88182e8e23SChristoph Lameter #ifdef CONFIG_HIGHMEM 897ee92255SKOSAKI Motohiro n += sprintf(buf + n, 901da177e4SLinus Torvalds "Node %d HighTotal: %8lu kB\n" 911da177e4SLinus Torvalds "Node %d HighFree: %8lu kB\n" 921da177e4SLinus Torvalds "Node %d LowTotal: %8lu kB\n" 937ee92255SKOSAKI Motohiro "Node %d LowFree: %8lu kB\n", 947ee92255SKOSAKI Motohiro nid, K(i.totalhigh), 957ee92255SKOSAKI Motohiro nid, K(i.freehigh), 967ee92255SKOSAKI Motohiro nid, K(i.totalram - i.totalhigh), 977ee92255SKOSAKI Motohiro nid, K(i.freeram - i.freehigh)); 98182e8e23SChristoph Lameter #endif 997ee92255SKOSAKI Motohiro n += sprintf(buf + n, 100c07e02dbSMartin Hicks "Node %d Dirty: %8lu kB\n" 101c07e02dbSMartin Hicks "Node %d Writeback: %8lu kB\n" 102347ce434SChristoph Lameter "Node %d FilePages: %8lu kB\n" 103c07e02dbSMartin Hicks "Node %d Mapped: %8lu kB\n" 104f3dbd344SChristoph Lameter "Node %d AnonPages: %8lu kB\n" 1054b02108aSKOSAKI Motohiro "Node %d Shmem: %8lu kB\n" 106c6a7f572SKOSAKI Motohiro "Node %d KernelStack: %8lu kB\n" 107df849a15SChristoph Lameter "Node %d PageTables: %8lu kB\n" 108f5ef68daSAndrew Morton "Node %d NFS_Unstable: %8lu kB\n" 109d2c5e30cSChristoph Lameter "Node %d Bounce: %8lu kB\n" 110fc3ba692SMiklos Szeredi "Node %d WritebackTmp: %8lu kB\n" 111972d1a7bSChristoph Lameter "Node %d Slab: %8lu kB\n" 112972d1a7bSChristoph Lameter "Node %d SReclaimable: %8lu kB\n" 11305b258e9SDavid Rientjes "Node %d SUnreclaim: %8lu kB\n" 11405b258e9SDavid Rientjes #ifdef CONFIG_TRANSPARENT_HUGEPAGE 11505b258e9SDavid Rientjes "Node %d AnonHugePages: %8lu kB\n" 11605b258e9SDavid Rientjes #endif 11705b258e9SDavid Rientjes , 118b1e7a8fdSChristoph Lameter nid, K(node_page_state(nid, NR_FILE_DIRTY)), 119ce866b34SChristoph Lameter nid, K(node_page_state(nid, NR_WRITEBACK)), 120347ce434SChristoph Lameter nid, K(node_page_state(nid, NR_FILE_PAGES)), 12165ba55f5SChristoph Lameter nid, K(node_page_state(nid, NR_FILE_MAPPED)), 12291a13c28SClaudio Scordino nid, K(node_page_state(nid, NR_ANON_PAGES)), 123cc7452b6SRafael Aquini nid, K(i.sharedram), 124c6a7f572SKOSAKI Motohiro nid, node_page_state(nid, NR_KERNEL_STACK) * 125c6a7f572SKOSAKI Motohiro THREAD_SIZE / 1024, 126df849a15SChristoph Lameter nid, K(node_page_state(nid, NR_PAGETABLE)), 127fd39fc85SChristoph Lameter nid, K(node_page_state(nid, NR_UNSTABLE_NFS)), 128d2c5e30cSChristoph Lameter nid, K(node_page_state(nid, NR_BOUNCE)), 129fc3ba692SMiklos Szeredi nid, K(node_page_state(nid, NR_WRITEBACK_TEMP)), 130972d1a7bSChristoph Lameter nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) + 131972d1a7bSChristoph Lameter node_page_state(nid, NR_SLAB_UNRECLAIMABLE)), 132972d1a7bSChristoph Lameter nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)), 13305b258e9SDavid Rientjes #ifdef CONFIG_TRANSPARENT_HUGEPAGE 13491a13c28SClaudio Scordino nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)) 13505b258e9SDavid Rientjes , nid, 13605b258e9SDavid Rientjes K(node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) * 13791a13c28SClaudio Scordino HPAGE_PMD_NR)); 13891a13c28SClaudio Scordino #else 13991a13c28SClaudio Scordino nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE))); 14005b258e9SDavid Rientjes #endif 1411da177e4SLinus Torvalds n += hugetlb_report_node_meminfo(nid, buf + n); 1421da177e4SLinus Torvalds return n; 1431da177e4SLinus Torvalds } 1441da177e4SLinus Torvalds 1451da177e4SLinus Torvalds #undef K 14610fbcf4cSKay Sievers static DEVICE_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL); 1471da177e4SLinus Torvalds 14810fbcf4cSKay Sievers static ssize_t node_read_numastat(struct device *dev, 14910fbcf4cSKay Sievers struct device_attribute *attr, char *buf) 1501da177e4SLinus Torvalds { 1511da177e4SLinus Torvalds return sprintf(buf, 1521da177e4SLinus Torvalds "numa_hit %lu\n" 1531da177e4SLinus Torvalds "numa_miss %lu\n" 1541da177e4SLinus Torvalds "numa_foreign %lu\n" 1551da177e4SLinus Torvalds "interleave_hit %lu\n" 1561da177e4SLinus Torvalds "local_node %lu\n" 1571da177e4SLinus Torvalds "other_node %lu\n", 158ca889e6cSChristoph Lameter node_page_state(dev->id, NUMA_HIT), 159ca889e6cSChristoph Lameter node_page_state(dev->id, NUMA_MISS), 160ca889e6cSChristoph Lameter node_page_state(dev->id, NUMA_FOREIGN), 161ca889e6cSChristoph Lameter node_page_state(dev->id, NUMA_INTERLEAVE_HIT), 162ca889e6cSChristoph Lameter node_page_state(dev->id, NUMA_LOCAL), 163ca889e6cSChristoph Lameter node_page_state(dev->id, NUMA_OTHER)); 1641da177e4SLinus Torvalds } 16510fbcf4cSKay Sievers static DEVICE_ATTR(numastat, S_IRUGO, node_read_numastat, NULL); 1661da177e4SLinus Torvalds 16710fbcf4cSKay Sievers static ssize_t node_read_vmstat(struct device *dev, 16810fbcf4cSKay Sievers struct device_attribute *attr, char *buf) 1692ac39037SMichael Rubin { 1702ac39037SMichael Rubin int nid = dev->id; 171fa25c503SKOSAKI Motohiro int i; 172fa25c503SKOSAKI Motohiro int n = 0; 173fa25c503SKOSAKI Motohiro 174fa25c503SKOSAKI Motohiro for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) 175fa25c503SKOSAKI Motohiro n += sprintf(buf+n, "%s %lu\n", vmstat_text[i], 176fa25c503SKOSAKI Motohiro node_page_state(nid, i)); 177fa25c503SKOSAKI Motohiro 178fa25c503SKOSAKI Motohiro return n; 1792ac39037SMichael Rubin } 18010fbcf4cSKay Sievers static DEVICE_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL); 1812ac39037SMichael Rubin 18210fbcf4cSKay Sievers static ssize_t node_read_distance(struct device *dev, 18310fbcf4cSKay Sievers struct device_attribute *attr, char *buf) 1841da177e4SLinus Torvalds { 1851da177e4SLinus Torvalds int nid = dev->id; 1861da177e4SLinus Torvalds int len = 0; 1871da177e4SLinus Torvalds int i; 1881da177e4SLinus Torvalds 18912ee3c0aSDavid Rientjes /* 19012ee3c0aSDavid Rientjes * buf is currently PAGE_SIZE in length and each node needs 4 chars 19112ee3c0aSDavid Rientjes * at the most (distance + space or newline). 19212ee3c0aSDavid Rientjes */ 19312ee3c0aSDavid Rientjes BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE); 1941da177e4SLinus Torvalds 1951da177e4SLinus Torvalds for_each_online_node(i) 1961da177e4SLinus Torvalds len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i)); 1971da177e4SLinus Torvalds 1981da177e4SLinus Torvalds len += sprintf(buf + len, "\n"); 1991da177e4SLinus Torvalds return len; 2001da177e4SLinus Torvalds } 20110fbcf4cSKay Sievers static DEVICE_ATTR(distance, S_IRUGO, node_read_distance, NULL); 2021da177e4SLinus Torvalds 2033c9b8aafSTakashi Iwai static struct attribute *node_dev_attrs[] = { 2043c9b8aafSTakashi Iwai &dev_attr_cpumap.attr, 2053c9b8aafSTakashi Iwai &dev_attr_cpulist.attr, 2063c9b8aafSTakashi Iwai &dev_attr_meminfo.attr, 2073c9b8aafSTakashi Iwai &dev_attr_numastat.attr, 2083c9b8aafSTakashi Iwai &dev_attr_distance.attr, 2093c9b8aafSTakashi Iwai &dev_attr_vmstat.attr, 2103c9b8aafSTakashi Iwai NULL 2113c9b8aafSTakashi Iwai }; 2127ca7ec40SGreg Kroah-Hartman ATTRIBUTE_GROUPS(node_dev); 2133c9b8aafSTakashi Iwai 2149a305230SLee Schermerhorn #ifdef CONFIG_HUGETLBFS 2159a305230SLee Schermerhorn /* 2169a305230SLee Schermerhorn * hugetlbfs per node attributes registration interface: 2179a305230SLee Schermerhorn * When/if hugetlb[fs] subsystem initializes [sometime after this module], 2184faf8d95SLee Schermerhorn * it will register its per node attributes for all online nodes with 2194faf8d95SLee Schermerhorn * memory. It will also call register_hugetlbfs_with_node(), below, to 2209a305230SLee Schermerhorn * register its attribute registration functions with this node driver. 2219a305230SLee Schermerhorn * Once these hooks have been initialized, the node driver will call into 2229a305230SLee Schermerhorn * the hugetlb module to [un]register attributes for hot-plugged nodes. 2239a305230SLee Schermerhorn */ 2249a305230SLee Schermerhorn static node_registration_func_t __hugetlb_register_node; 2259a305230SLee Schermerhorn static node_registration_func_t __hugetlb_unregister_node; 2269a305230SLee Schermerhorn 22739da08cbSLee Schermerhorn static inline bool hugetlb_register_node(struct node *node) 2289a305230SLee Schermerhorn { 2294faf8d95SLee Schermerhorn if (__hugetlb_register_node && 2308cebfcd0SLai Jiangshan node_state(node->dev.id, N_MEMORY)) { 2319a305230SLee Schermerhorn __hugetlb_register_node(node); 23239da08cbSLee Schermerhorn return true; 23339da08cbSLee Schermerhorn } 23439da08cbSLee Schermerhorn return false; 2359a305230SLee Schermerhorn } 2369a305230SLee Schermerhorn 2379a305230SLee Schermerhorn static inline void hugetlb_unregister_node(struct node *node) 2389a305230SLee Schermerhorn { 2399a305230SLee Schermerhorn if (__hugetlb_unregister_node) 2409a305230SLee Schermerhorn __hugetlb_unregister_node(node); 2419a305230SLee Schermerhorn } 2429a305230SLee Schermerhorn 2439a305230SLee Schermerhorn void register_hugetlbfs_with_node(node_registration_func_t doregister, 2449a305230SLee Schermerhorn node_registration_func_t unregister) 2459a305230SLee Schermerhorn { 2469a305230SLee Schermerhorn __hugetlb_register_node = doregister; 2479a305230SLee Schermerhorn __hugetlb_unregister_node = unregister; 2489a305230SLee Schermerhorn } 2499a305230SLee Schermerhorn #else 2509a305230SLee Schermerhorn static inline void hugetlb_register_node(struct node *node) {} 2519a305230SLee Schermerhorn 2529a305230SLee Schermerhorn static inline void hugetlb_unregister_node(struct node *node) {} 2539a305230SLee Schermerhorn #endif 2549a305230SLee Schermerhorn 2558c7b5b4eSYasuaki Ishimatsu static void node_device_release(struct device *dev) 2568c7b5b4eSYasuaki Ishimatsu { 2578c7b5b4eSYasuaki Ishimatsu struct node *node = to_node(dev); 2588c7b5b4eSYasuaki Ishimatsu 2598c7b5b4eSYasuaki Ishimatsu #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HUGETLBFS) 2608c7b5b4eSYasuaki Ishimatsu /* 2618c7b5b4eSYasuaki Ishimatsu * We schedule the work only when a memory section is 2628c7b5b4eSYasuaki Ishimatsu * onlined/offlined on this node. When we come here, 2638c7b5b4eSYasuaki Ishimatsu * all the memory on this node has been offlined, 2648c7b5b4eSYasuaki Ishimatsu * so we won't enqueue new work to this work. 2658c7b5b4eSYasuaki Ishimatsu * 2668c7b5b4eSYasuaki Ishimatsu * The work is using node->node_work, so we should 2678c7b5b4eSYasuaki Ishimatsu * flush work before freeing the memory. 2688c7b5b4eSYasuaki Ishimatsu */ 2698c7b5b4eSYasuaki Ishimatsu flush_work(&node->node_work); 2708c7b5b4eSYasuaki Ishimatsu #endif 2718c7b5b4eSYasuaki Ishimatsu kfree(node); 2728c7b5b4eSYasuaki Ishimatsu } 2731da177e4SLinus Torvalds 2741da177e4SLinus Torvalds /* 275405ae7d3SRobert P. J. Day * register_node - Setup a sysfs device for a node. 2761da177e4SLinus Torvalds * @num - Node number to use when creating the device. 2771da177e4SLinus Torvalds * 2781da177e4SLinus Torvalds * Initialize and register the node device. 2791da177e4SLinus Torvalds */ 280fa264375SYasuaki Ishimatsu static int register_node(struct node *node, int num, struct node *parent) 2811da177e4SLinus Torvalds { 2821da177e4SLinus Torvalds int error; 2831da177e4SLinus Torvalds 28410fbcf4cSKay Sievers node->dev.id = num; 28510fbcf4cSKay Sievers node->dev.bus = &node_subsys; 2868c7b5b4eSYasuaki Ishimatsu node->dev.release = node_device_release; 2877ca7ec40SGreg Kroah-Hartman node->dev.groups = node_dev_groups; 28810fbcf4cSKay Sievers error = device_register(&node->dev); 2891da177e4SLinus Torvalds 2901da177e4SLinus Torvalds if (!error){ 2919a305230SLee Schermerhorn hugetlb_register_node(node); 292ed4a6d7fSMel Gorman 293ed4a6d7fSMel Gorman compaction_register_node(node); 2941da177e4SLinus Torvalds } 2951da177e4SLinus Torvalds return error; 2961da177e4SLinus Torvalds } 2971da177e4SLinus Torvalds 2984b45099bSKeiichiro Tokunaga /** 2994b45099bSKeiichiro Tokunaga * unregister_node - unregister a node device 3004b45099bSKeiichiro Tokunaga * @node: node going away 3014b45099bSKeiichiro Tokunaga * 3024b45099bSKeiichiro Tokunaga * Unregisters a node device @node. All the devices on the node must be 3034b45099bSKeiichiro Tokunaga * unregistered before calling this function. 3044b45099bSKeiichiro Tokunaga */ 3054b45099bSKeiichiro Tokunaga void unregister_node(struct node *node) 3064b45099bSKeiichiro Tokunaga { 3074faf8d95SLee Schermerhorn hugetlb_unregister_node(node); /* no-op, if memoryless node */ 308af936a16SLee Schermerhorn 30910fbcf4cSKay Sievers device_unregister(&node->dev); 3104b45099bSKeiichiro Tokunaga } 3114b45099bSKeiichiro Tokunaga 3128732794bSWen Congyang struct node *node_devices[MAX_NUMNODES]; 3130fc44159SYasunori Goto 31476b67ed9SKAMEZAWA Hiroyuki /* 31576b67ed9SKAMEZAWA Hiroyuki * register cpu under node 31676b67ed9SKAMEZAWA Hiroyuki */ 31776b67ed9SKAMEZAWA Hiroyuki int register_cpu_under_node(unsigned int cpu, unsigned int nid) 31876b67ed9SKAMEZAWA Hiroyuki { 3191830794aSAlex Chiang int ret; 3208a25a2fdSKay Sievers struct device *obj; 321f8246f31SAlex Chiang 322f8246f31SAlex Chiang if (!node_online(nid)) 323f8246f31SAlex Chiang return 0; 324f8246f31SAlex Chiang 3258a25a2fdSKay Sievers obj = get_cpu_device(cpu); 32676b67ed9SKAMEZAWA Hiroyuki if (!obj) 32776b67ed9SKAMEZAWA Hiroyuki return 0; 328f8246f31SAlex Chiang 3298732794bSWen Congyang ret = sysfs_create_link(&node_devices[nid]->dev.kobj, 33076b67ed9SKAMEZAWA Hiroyuki &obj->kobj, 33176b67ed9SKAMEZAWA Hiroyuki kobject_name(&obj->kobj)); 3321830794aSAlex Chiang if (ret) 3331830794aSAlex Chiang return ret; 3341830794aSAlex Chiang 3351830794aSAlex Chiang return sysfs_create_link(&obj->kobj, 3368732794bSWen Congyang &node_devices[nid]->dev.kobj, 3378732794bSWen Congyang kobject_name(&node_devices[nid]->dev.kobj)); 33876b67ed9SKAMEZAWA Hiroyuki } 33976b67ed9SKAMEZAWA Hiroyuki 34076b67ed9SKAMEZAWA Hiroyuki int unregister_cpu_under_node(unsigned int cpu, unsigned int nid) 34176b67ed9SKAMEZAWA Hiroyuki { 3428a25a2fdSKay Sievers struct device *obj; 343b9d52dadSAlex Chiang 344b9d52dadSAlex Chiang if (!node_online(nid)) 345b9d52dadSAlex Chiang return 0; 346b9d52dadSAlex Chiang 3478a25a2fdSKay Sievers obj = get_cpu_device(cpu); 348b9d52dadSAlex Chiang if (!obj) 349b9d52dadSAlex Chiang return 0; 350b9d52dadSAlex Chiang 3518732794bSWen Congyang sysfs_remove_link(&node_devices[nid]->dev.kobj, 35276b67ed9SKAMEZAWA Hiroyuki kobject_name(&obj->kobj)); 3531830794aSAlex Chiang sysfs_remove_link(&obj->kobj, 3548732794bSWen Congyang kobject_name(&node_devices[nid]->dev.kobj)); 355b9d52dadSAlex Chiang 35676b67ed9SKAMEZAWA Hiroyuki return 0; 35776b67ed9SKAMEZAWA Hiroyuki } 35876b67ed9SKAMEZAWA Hiroyuki 359c04fc586SGary Hade #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE 360c04fc586SGary Hade #define page_initialized(page) (page->lru.next) 361c04fc586SGary Hade 3623a80a7faSMel Gorman static int __init_refok get_nid_for_pfn(unsigned long pfn) 363c04fc586SGary Hade { 364c04fc586SGary Hade struct page *page; 365c04fc586SGary Hade 366c04fc586SGary Hade if (!pfn_valid_within(pfn)) 367c04fc586SGary Hade return -1; 3683a80a7faSMel Gorman #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT 3693a80a7faSMel Gorman if (system_state == SYSTEM_BOOTING) 3703a80a7faSMel Gorman return early_pfn_to_nid(pfn); 3713a80a7faSMel Gorman #endif 372c04fc586SGary Hade page = pfn_to_page(pfn); 373c04fc586SGary Hade if (!page_initialized(page)) 374c04fc586SGary Hade return -1; 375c04fc586SGary Hade return pfn_to_nid(pfn); 376c04fc586SGary Hade } 377c04fc586SGary Hade 378c04fc586SGary Hade /* register memory section under specified node if it spans that node */ 379c04fc586SGary Hade int register_mem_sect_under_node(struct memory_block *mem_blk, int nid) 380c04fc586SGary Hade { 381dee5d0d5SAlex Chiang int ret; 382c04fc586SGary Hade unsigned long pfn, sect_start_pfn, sect_end_pfn; 383c04fc586SGary Hade 384c04fc586SGary Hade if (!mem_blk) 385c04fc586SGary Hade return -EFAULT; 386c04fc586SGary Hade if (!node_online(nid)) 387c04fc586SGary Hade return 0; 388d3360164SNathan Fontenot 389d3360164SNathan Fontenot sect_start_pfn = section_nr_to_pfn(mem_blk->start_section_nr); 390d3360164SNathan Fontenot sect_end_pfn = section_nr_to_pfn(mem_blk->end_section_nr); 391d3360164SNathan Fontenot sect_end_pfn += PAGES_PER_SECTION - 1; 392c04fc586SGary Hade for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) { 393c04fc586SGary Hade int page_nid; 394c04fc586SGary Hade 39504697858SYinghai Lu /* 39604697858SYinghai Lu * memory block could have several absent sections from start. 39704697858SYinghai Lu * skip pfn range from absent section 39804697858SYinghai Lu */ 39904697858SYinghai Lu if (!pfn_present(pfn)) { 40004697858SYinghai Lu pfn = round_down(pfn + PAGES_PER_SECTION, 40104697858SYinghai Lu PAGES_PER_SECTION) - 1; 40204697858SYinghai Lu continue; 40304697858SYinghai Lu } 40404697858SYinghai Lu 405c04fc586SGary Hade page_nid = get_nid_for_pfn(pfn); 406c04fc586SGary Hade if (page_nid < 0) 407c04fc586SGary Hade continue; 408c04fc586SGary Hade if (page_nid != nid) 409c04fc586SGary Hade continue; 4108732794bSWen Congyang ret = sysfs_create_link_nowarn(&node_devices[nid]->dev.kobj, 41110fbcf4cSKay Sievers &mem_blk->dev.kobj, 41210fbcf4cSKay Sievers kobject_name(&mem_blk->dev.kobj)); 413dee5d0d5SAlex Chiang if (ret) 414dee5d0d5SAlex Chiang return ret; 415dee5d0d5SAlex Chiang 41610fbcf4cSKay Sievers return sysfs_create_link_nowarn(&mem_blk->dev.kobj, 4178732794bSWen Congyang &node_devices[nid]->dev.kobj, 4188732794bSWen Congyang kobject_name(&node_devices[nid]->dev.kobj)); 419c04fc586SGary Hade } 420c04fc586SGary Hade /* mem section does not span the specified node */ 421c04fc586SGary Hade return 0; 422c04fc586SGary Hade } 423c04fc586SGary Hade 424c04fc586SGary Hade /* unregister memory section under all nodes that it spans */ 425d3360164SNathan Fontenot int unregister_mem_sect_under_nodes(struct memory_block *mem_blk, 426d3360164SNathan Fontenot unsigned long phys_index) 427c04fc586SGary Hade { 4289ae49fabSDavid Rientjes NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL); 429c04fc586SGary Hade unsigned long pfn, sect_start_pfn, sect_end_pfn; 430c04fc586SGary Hade 4319ae49fabSDavid Rientjes if (!mem_blk) { 4329ae49fabSDavid Rientjes NODEMASK_FREE(unlinked_nodes); 433c04fc586SGary Hade return -EFAULT; 4349ae49fabSDavid Rientjes } 4359ae49fabSDavid Rientjes if (!unlinked_nodes) 4369ae49fabSDavid Rientjes return -ENOMEM; 4379ae49fabSDavid Rientjes nodes_clear(*unlinked_nodes); 438d3360164SNathan Fontenot 439d3360164SNathan Fontenot sect_start_pfn = section_nr_to_pfn(phys_index); 440c04fc586SGary Hade sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1; 441c04fc586SGary Hade for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) { 44247504980SRoel Kluin int nid; 443c04fc586SGary Hade 444c04fc586SGary Hade nid = get_nid_for_pfn(pfn); 445c04fc586SGary Hade if (nid < 0) 446c04fc586SGary Hade continue; 447c04fc586SGary Hade if (!node_online(nid)) 448c04fc586SGary Hade continue; 4499ae49fabSDavid Rientjes if (node_test_and_set(nid, *unlinked_nodes)) 450c04fc586SGary Hade continue; 4518732794bSWen Congyang sysfs_remove_link(&node_devices[nid]->dev.kobj, 45210fbcf4cSKay Sievers kobject_name(&mem_blk->dev.kobj)); 45310fbcf4cSKay Sievers sysfs_remove_link(&mem_blk->dev.kobj, 4548732794bSWen Congyang kobject_name(&node_devices[nid]->dev.kobj)); 455c04fc586SGary Hade } 4569ae49fabSDavid Rientjes NODEMASK_FREE(unlinked_nodes); 457c04fc586SGary Hade return 0; 458c04fc586SGary Hade } 459c04fc586SGary Hade 460c04fc586SGary Hade static int link_mem_sections(int nid) 461c04fc586SGary Hade { 462c04fc586SGary Hade unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn; 463c04fc586SGary Hade unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages; 464c04fc586SGary Hade unsigned long pfn; 46563d027a6SRobin Holt struct memory_block *mem_blk = NULL; 466c04fc586SGary Hade int err = 0; 467c04fc586SGary Hade 468c04fc586SGary Hade for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 469c04fc586SGary Hade unsigned long section_nr = pfn_to_section_nr(pfn); 470c04fc586SGary Hade struct mem_section *mem_sect; 471c04fc586SGary Hade int ret; 472c04fc586SGary Hade 473c04fc586SGary Hade if (!present_section_nr(section_nr)) 474c04fc586SGary Hade continue; 475c04fc586SGary Hade mem_sect = __nr_to_section(section_nr); 476321bf4edSYinghai Lu 477321bf4edSYinghai Lu /* same memblock ? */ 478321bf4edSYinghai Lu if (mem_blk) 479321bf4edSYinghai Lu if ((section_nr >= mem_blk->start_section_nr) && 480321bf4edSYinghai Lu (section_nr <= mem_blk->end_section_nr)) 481321bf4edSYinghai Lu continue; 482321bf4edSYinghai Lu 48363d027a6SRobin Holt mem_blk = find_memory_block_hinted(mem_sect, mem_blk); 484321bf4edSYinghai Lu 485c04fc586SGary Hade ret = register_mem_sect_under_node(mem_blk, nid); 486c04fc586SGary Hade if (!err) 487c04fc586SGary Hade err = ret; 488c04fc586SGary Hade 489c04fc586SGary Hade /* discard ref obtained in find_memory_block() */ 490c04fc586SGary Hade } 49163d027a6SRobin Holt 49263d027a6SRobin Holt if (mem_blk) 49310fbcf4cSKay Sievers kobject_put(&mem_blk->dev.kobj); 494c04fc586SGary Hade return err; 495c04fc586SGary Hade } 4964faf8d95SLee Schermerhorn 49739da08cbSLee Schermerhorn #ifdef CONFIG_HUGETLBFS 4984faf8d95SLee Schermerhorn /* 4994faf8d95SLee Schermerhorn * Handle per node hstate attribute [un]registration on transistions 5004faf8d95SLee Schermerhorn * to/from memoryless state. 5014faf8d95SLee Schermerhorn */ 50239da08cbSLee Schermerhorn static void node_hugetlb_work(struct work_struct *work) 50339da08cbSLee Schermerhorn { 50439da08cbSLee Schermerhorn struct node *node = container_of(work, struct node, node_work); 50539da08cbSLee Schermerhorn 50639da08cbSLee Schermerhorn /* 50739da08cbSLee Schermerhorn * We only get here when a node transitions to/from memoryless state. 50839da08cbSLee Schermerhorn * We can detect which transition occurred by examining whether the 50939da08cbSLee Schermerhorn * node has memory now. hugetlb_register_node() already check this 51039da08cbSLee Schermerhorn * so we try to register the attributes. If that fails, then the 51139da08cbSLee Schermerhorn * node has transitioned to memoryless, try to unregister the 51239da08cbSLee Schermerhorn * attributes. 51339da08cbSLee Schermerhorn */ 51439da08cbSLee Schermerhorn if (!hugetlb_register_node(node)) 51539da08cbSLee Schermerhorn hugetlb_unregister_node(node); 51639da08cbSLee Schermerhorn } 51739da08cbSLee Schermerhorn 51839da08cbSLee Schermerhorn static void init_node_hugetlb_work(int nid) 51939da08cbSLee Schermerhorn { 5208732794bSWen Congyang INIT_WORK(&node_devices[nid]->node_work, node_hugetlb_work); 52139da08cbSLee Schermerhorn } 5224faf8d95SLee Schermerhorn 5234faf8d95SLee Schermerhorn static int node_memory_callback(struct notifier_block *self, 5244faf8d95SLee Schermerhorn unsigned long action, void *arg) 5254faf8d95SLee Schermerhorn { 5264faf8d95SLee Schermerhorn struct memory_notify *mnb = arg; 5274faf8d95SLee Schermerhorn int nid = mnb->status_change_nid; 5284faf8d95SLee Schermerhorn 5294faf8d95SLee Schermerhorn switch (action) { 53039da08cbSLee Schermerhorn case MEM_ONLINE: 53139da08cbSLee Schermerhorn case MEM_OFFLINE: 53239da08cbSLee Schermerhorn /* 53339da08cbSLee Schermerhorn * offload per node hstate [un]registration to a work thread 53439da08cbSLee Schermerhorn * when transitioning to/from memoryless state. 53539da08cbSLee Schermerhorn */ 5364faf8d95SLee Schermerhorn if (nid != NUMA_NO_NODE) 5378732794bSWen Congyang schedule_work(&node_devices[nid]->node_work); 5384faf8d95SLee Schermerhorn break; 53939da08cbSLee Schermerhorn 5404faf8d95SLee Schermerhorn case MEM_GOING_ONLINE: 5414faf8d95SLee Schermerhorn case MEM_GOING_OFFLINE: 5424faf8d95SLee Schermerhorn case MEM_CANCEL_ONLINE: 5434faf8d95SLee Schermerhorn case MEM_CANCEL_OFFLINE: 5444faf8d95SLee Schermerhorn default: 5454faf8d95SLee Schermerhorn break; 5464faf8d95SLee Schermerhorn } 5474faf8d95SLee Schermerhorn 5484faf8d95SLee Schermerhorn return NOTIFY_OK; 5494faf8d95SLee Schermerhorn } 55039da08cbSLee Schermerhorn #endif /* CONFIG_HUGETLBFS */ 55139da08cbSLee Schermerhorn #else /* !CONFIG_MEMORY_HOTPLUG_SPARSE */ 5524faf8d95SLee Schermerhorn 55339da08cbSLee Schermerhorn static int link_mem_sections(int nid) { return 0; } 55439da08cbSLee Schermerhorn #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */ 55539da08cbSLee Schermerhorn 55639da08cbSLee Schermerhorn #if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \ 55739da08cbSLee Schermerhorn !defined(CONFIG_HUGETLBFS) 5584faf8d95SLee Schermerhorn static inline int node_memory_callback(struct notifier_block *self, 5594faf8d95SLee Schermerhorn unsigned long action, void *arg) 5604faf8d95SLee Schermerhorn { 5614faf8d95SLee Schermerhorn return NOTIFY_OK; 5624faf8d95SLee Schermerhorn } 56339da08cbSLee Schermerhorn 56439da08cbSLee Schermerhorn static void init_node_hugetlb_work(int nid) { } 56539da08cbSLee Schermerhorn 56639da08cbSLee Schermerhorn #endif 567c04fc586SGary Hade 5680fc44159SYasunori Goto int register_one_node(int nid) 5690fc44159SYasunori Goto { 5700fc44159SYasunori Goto int error = 0; 57176b67ed9SKAMEZAWA Hiroyuki int cpu; 5720fc44159SYasunori Goto 5730fc44159SYasunori Goto if (node_online(nid)) { 5740fc44159SYasunori Goto int p_node = parent_node(nid); 5750fc44159SYasunori Goto struct node *parent = NULL; 5760fc44159SYasunori Goto 5770fc44159SYasunori Goto if (p_node != nid) 5788732794bSWen Congyang parent = node_devices[p_node]; 5790fc44159SYasunori Goto 5808732794bSWen Congyang node_devices[nid] = kzalloc(sizeof(struct node), GFP_KERNEL); 5818732794bSWen Congyang if (!node_devices[nid]) 5828732794bSWen Congyang return -ENOMEM; 5838732794bSWen Congyang 5848732794bSWen Congyang error = register_node(node_devices[nid], nid, parent); 58576b67ed9SKAMEZAWA Hiroyuki 58676b67ed9SKAMEZAWA Hiroyuki /* link cpu under this node */ 58776b67ed9SKAMEZAWA Hiroyuki for_each_present_cpu(cpu) { 58876b67ed9SKAMEZAWA Hiroyuki if (cpu_to_node(cpu) == nid) 58976b67ed9SKAMEZAWA Hiroyuki register_cpu_under_node(cpu, nid); 59076b67ed9SKAMEZAWA Hiroyuki } 591c04fc586SGary Hade 592c04fc586SGary Hade /* link memory sections under this node */ 593c04fc586SGary Hade error = link_mem_sections(nid); 59439da08cbSLee Schermerhorn 59539da08cbSLee Schermerhorn /* initialize work queue for memory hot plug */ 59639da08cbSLee Schermerhorn init_node_hugetlb_work(nid); 5970fc44159SYasunori Goto } 5980fc44159SYasunori Goto 5990fc44159SYasunori Goto return error; 6000fc44159SYasunori Goto 6010fc44159SYasunori Goto } 6020fc44159SYasunori Goto 6030fc44159SYasunori Goto void unregister_one_node(int nid) 6040fc44159SYasunori Goto { 60592d585efSXishi Qiu if (!node_devices[nid]) 60692d585efSXishi Qiu return; 60792d585efSXishi Qiu 6088732794bSWen Congyang unregister_node(node_devices[nid]); 6098732794bSWen Congyang node_devices[nid] = NULL; 6100fc44159SYasunori Goto } 6110fc44159SYasunori Goto 612bde631a5SLee Schermerhorn /* 613bde631a5SLee Schermerhorn * node states attributes 614bde631a5SLee Schermerhorn */ 615bde631a5SLee Schermerhorn 616bde631a5SLee Schermerhorn static ssize_t print_nodes_state(enum node_states state, char *buf) 617bde631a5SLee Schermerhorn { 618bde631a5SLee Schermerhorn int n; 619bde631a5SLee Schermerhorn 620f799b1a7STejun Heo n = scnprintf(buf, PAGE_SIZE - 1, "%*pbl", 621f799b1a7STejun Heo nodemask_pr_args(&node_states[state])); 622f6238818SRyota Ozaki buf[n++] = '\n'; 623f6238818SRyota Ozaki buf[n] = '\0'; 624bde631a5SLee Schermerhorn return n; 625bde631a5SLee Schermerhorn } 626bde631a5SLee Schermerhorn 627b15f562fSAndi Kleen struct node_attr { 62810fbcf4cSKay Sievers struct device_attribute attr; 629b15f562fSAndi Kleen enum node_states state; 630b15f562fSAndi Kleen }; 631b15f562fSAndi Kleen 63210fbcf4cSKay Sievers static ssize_t show_node_state(struct device *dev, 63310fbcf4cSKay Sievers struct device_attribute *attr, char *buf) 634bde631a5SLee Schermerhorn { 635b15f562fSAndi Kleen struct node_attr *na = container_of(attr, struct node_attr, attr); 636b15f562fSAndi Kleen return print_nodes_state(na->state, buf); 637bde631a5SLee Schermerhorn } 638bde631a5SLee Schermerhorn 639b15f562fSAndi Kleen #define _NODE_ATTR(name, state) \ 64010fbcf4cSKay Sievers { __ATTR(name, 0444, show_node_state, NULL), state } 641bde631a5SLee Schermerhorn 642b15f562fSAndi Kleen static struct node_attr node_state_attr[] = { 643fcf07d22SLai Jiangshan [N_POSSIBLE] = _NODE_ATTR(possible, N_POSSIBLE), 644fcf07d22SLai Jiangshan [N_ONLINE] = _NODE_ATTR(online, N_ONLINE), 645fcf07d22SLai Jiangshan [N_NORMAL_MEMORY] = _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY), 646bde631a5SLee Schermerhorn #ifdef CONFIG_HIGHMEM 647fcf07d22SLai Jiangshan [N_HIGH_MEMORY] = _NODE_ATTR(has_high_memory, N_HIGH_MEMORY), 648bde631a5SLee Schermerhorn #endif 64920b2f52bSLai Jiangshan #ifdef CONFIG_MOVABLE_NODE 65020b2f52bSLai Jiangshan [N_MEMORY] = _NODE_ATTR(has_memory, N_MEMORY), 65120b2f52bSLai Jiangshan #endif 652fcf07d22SLai Jiangshan [N_CPU] = _NODE_ATTR(has_cpu, N_CPU), 653bde631a5SLee Schermerhorn }; 654bde631a5SLee Schermerhorn 65510fbcf4cSKay Sievers static struct attribute *node_state_attrs[] = { 656fcf07d22SLai Jiangshan &node_state_attr[N_POSSIBLE].attr.attr, 657fcf07d22SLai Jiangshan &node_state_attr[N_ONLINE].attr.attr, 658fcf07d22SLai Jiangshan &node_state_attr[N_NORMAL_MEMORY].attr.attr, 6593701cde6SAndi Kleen #ifdef CONFIG_HIGHMEM 660fcf07d22SLai Jiangshan &node_state_attr[N_HIGH_MEMORY].attr.attr, 6613701cde6SAndi Kleen #endif 66220b2f52bSLai Jiangshan #ifdef CONFIG_MOVABLE_NODE 66320b2f52bSLai Jiangshan &node_state_attr[N_MEMORY].attr.attr, 66420b2f52bSLai Jiangshan #endif 665fcf07d22SLai Jiangshan &node_state_attr[N_CPU].attr.attr, 6663701cde6SAndi Kleen NULL 6673701cde6SAndi Kleen }; 668bde631a5SLee Schermerhorn 66910fbcf4cSKay Sievers static struct attribute_group memory_root_attr_group = { 67010fbcf4cSKay Sievers .attrs = node_state_attrs, 67110fbcf4cSKay Sievers }; 67210fbcf4cSKay Sievers 67310fbcf4cSKay Sievers static const struct attribute_group *cpu_root_attr_groups[] = { 67410fbcf4cSKay Sievers &memory_root_attr_group, 67510fbcf4cSKay Sievers NULL, 67610fbcf4cSKay Sievers }; 67710fbcf4cSKay Sievers 6784faf8d95SLee Schermerhorn #define NODE_CALLBACK_PRI 2 /* lower than SLAB */ 6794b45099bSKeiichiro Tokunaga static int __init register_node_type(void) 6801da177e4SLinus Torvalds { 681bde631a5SLee Schermerhorn int ret; 682bde631a5SLee Schermerhorn 6833701cde6SAndi Kleen BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES); 6843701cde6SAndi Kleen BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES); 6853701cde6SAndi Kleen 68610fbcf4cSKay Sievers ret = subsys_system_register(&node_subsys, cpu_root_attr_groups); 6874faf8d95SLee Schermerhorn if (!ret) { 6886e259e7dSAndrew Morton static struct notifier_block node_memory_callback_nb = { 6896e259e7dSAndrew Morton .notifier_call = node_memory_callback, 6906e259e7dSAndrew Morton .priority = NODE_CALLBACK_PRI, 6916e259e7dSAndrew Morton }; 6926e259e7dSAndrew Morton register_hotmemory_notifier(&node_memory_callback_nb); 6934faf8d95SLee Schermerhorn } 694bde631a5SLee Schermerhorn 695bde631a5SLee Schermerhorn /* 696bde631a5SLee Schermerhorn * Note: we're not going to unregister the node class if we fail 697bde631a5SLee Schermerhorn * to register the node state class attribute files. 698bde631a5SLee Schermerhorn */ 699bde631a5SLee Schermerhorn return ret; 7001da177e4SLinus Torvalds } 7011da177e4SLinus Torvalds postcore_initcall(register_node_type); 702