PM / Sleep: Add wakeup_source_activate and wakeup_source_deactivate tracepoints
[linux-2.6.git] / drivers / base / node.c
index 91d4087..90aa2a1 100644 (file)
@@ -1,30 +1,33 @@
 /*
- * drivers/base/node.c - basic Node class support
+ * Basic Node interface support
  */
 
-#include <linux/sysdev.h>
 #include <linux/module.h>
 #include <linux/init.h>
 #include <linux/mm.h>
 #include <linux/memory.h>
+#include <linux/vmstat.h>
 #include <linux/node.h>
 #include <linux/hugetlb.h>
+#include <linux/compaction.h>
 #include <linux/cpumask.h>
 #include <linux/topology.h>
 #include <linux/nodemask.h>
 #include <linux/cpu.h>
 #include <linux/device.h>
 #include <linux/swap.h>
+#include <linux/slab.h>
 
-static struct sysdev_class node_class = {
+static struct bus_type node_subsys = {
        .name = "node",
+       .dev_name = "node",
 };
 
 
-static ssize_t node_read_cpumap(struct sys_device *dev, int type, char *buf)
+static ssize_t node_read_cpumap(struct device *dev, int type, char *buf)
 {
        struct node *node_dev = to_node(dev);
-       const struct cpumask *mask = cpumask_of_node(node_dev->sysdev.id);
+       const struct cpumask *mask = cpumask_of_node(node_dev->dev.id);
        int len;
 
        /* 2008/04/07: buf currently PAGE_SIZE, need 9 chars per 32 bits. */
@@ -38,31 +41,30 @@ static ssize_t node_read_cpumap(struct sys_device *dev, int type, char *buf)
        return len;
 }
 
-static inline ssize_t node_read_cpumask(struct sys_device *dev,
-                               struct sysdev_attribute *attr, char *buf)
+static inline ssize_t node_read_cpumask(struct device *dev,
+                               struct device_attribute *attr, char *buf)
 {
        return node_read_cpumap(dev, 0, buf);
 }
-static inline ssize_t node_read_cpulist(struct sys_device *dev,
-                               struct sysdev_attribute *attr, char *buf)
+static inline ssize_t node_read_cpulist(struct device *dev,
+                               struct device_attribute *attr, char *buf)
 {
        return node_read_cpumap(dev, 1, buf);
 }
 
-static SYSDEV_ATTR(cpumap,  S_IRUGO, node_read_cpumask, NULL);
-static SYSDEV_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
+static DEVICE_ATTR(cpumap,  S_IRUGO, node_read_cpumask, NULL);
+static DEVICE_ATTR(cpulist, S_IRUGO, node_read_cpulist, NULL);
 
 #define K(x) ((x) << (PAGE_SHIFT - 10))
-static ssize_t node_read_meminfo(struct sys_device * dev,
-                       struct sysdev_attribute *attr, char * buf)
+static ssize_t node_read_meminfo(struct device *dev,
+                       struct device_attribute *attr, char *buf)
 {
        int n;
        int nid = dev->id;
        struct sysinfo i;
 
        si_meminfo_node(&i, nid);
-
-       n = sprintf(buf, "\n"
+       n = sprintf(buf,
                       "Node %d MemTotal:       %8lu kB\n"
                       "Node %d MemFree:        %8lu kB\n"
                       "Node %d MemUsed:        %8lu kB\n"
@@ -73,25 +75,7 @@ static ssize_t node_read_meminfo(struct sys_device * dev,
                       "Node %d Active(file):   %8lu kB\n"
                       "Node %d Inactive(file): %8lu kB\n"
                       "Node %d Unevictable:    %8lu kB\n"
-                      "Node %d Mlocked:        %8lu kB\n"
-#ifdef CONFIG_HIGHMEM
-                      "Node %d HighTotal:      %8lu kB\n"
-                      "Node %d HighFree:       %8lu kB\n"
-                      "Node %d LowTotal:       %8lu kB\n"
-                      "Node %d LowFree:        %8lu kB\n"
-#endif
-                      "Node %d Dirty:          %8lu kB\n"
-                      "Node %d Writeback:      %8lu kB\n"
-                      "Node %d FilePages:      %8lu kB\n"
-                      "Node %d Mapped:         %8lu kB\n"
-                      "Node %d AnonPages:      %8lu kB\n"
-                      "Node %d PageTables:     %8lu kB\n"
-                      "Node %d NFS_Unstable:   %8lu kB\n"
-                      "Node %d Bounce:         %8lu kB\n"
-                      "Node %d WritebackTmp:   %8lu kB\n"
-                      "Node %d Slab:           %8lu kB\n"
-                      "Node %d SReclaimable:   %8lu kB\n"
-                      "Node %d SUnreclaim:     %8lu kB\n",
+                      "Node %d Mlocked:        %8lu kB\n",
                       nid, K(i.totalram),
                       nid, K(i.freeram),
                       nid, K(i.totalram - i.freeram),
@@ -104,18 +88,52 @@ static ssize_t node_read_meminfo(struct sys_device * dev,
                       nid, K(node_page_state(nid, NR_ACTIVE_FILE)),
                       nid, K(node_page_state(nid, NR_INACTIVE_FILE)),
                       nid, K(node_page_state(nid, NR_UNEVICTABLE)),
-                      nid, K(node_page_state(nid, NR_MLOCK)),
+                      nid, K(node_page_state(nid, NR_MLOCK)));
+
 #ifdef CONFIG_HIGHMEM
+       n += sprintf(buf + n,
+                      "Node %d HighTotal:      %8lu kB\n"
+                      "Node %d HighFree:       %8lu kB\n"
+                      "Node %d LowTotal:       %8lu kB\n"
+                      "Node %d LowFree:        %8lu kB\n",
                       nid, K(i.totalhigh),
                       nid, K(i.freehigh),
                       nid, K(i.totalram - i.totalhigh),
-                      nid, K(i.freeram - i.freehigh),
+                      nid, K(i.freeram - i.freehigh));
+#endif
+       n += sprintf(buf + n,
+                      "Node %d Dirty:          %8lu kB\n"
+                      "Node %d Writeback:      %8lu kB\n"
+                      "Node %d FilePages:      %8lu kB\n"
+                      "Node %d Mapped:         %8lu kB\n"
+                      "Node %d AnonPages:      %8lu kB\n"
+                      "Node %d Shmem:          %8lu kB\n"
+                      "Node %d KernelStack:    %8lu kB\n"
+                      "Node %d PageTables:     %8lu kB\n"
+                      "Node %d NFS_Unstable:   %8lu kB\n"
+                      "Node %d Bounce:         %8lu kB\n"
+                      "Node %d WritebackTmp:   %8lu kB\n"
+                      "Node %d Slab:           %8lu kB\n"
+                      "Node %d SReclaimable:   %8lu kB\n"
+                      "Node %d SUnreclaim:     %8lu kB\n"
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+                      "Node %d AnonHugePages:  %8lu kB\n"
 #endif
+                       ,
                       nid, K(node_page_state(nid, NR_FILE_DIRTY)),
                       nid, K(node_page_state(nid, NR_WRITEBACK)),
                       nid, K(node_page_state(nid, NR_FILE_PAGES)),
                       nid, K(node_page_state(nid, NR_FILE_MAPPED)),
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+                      nid, K(node_page_state(nid, NR_ANON_PAGES)
+                       + node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) *
+                       HPAGE_PMD_NR),
+#else
                       nid, K(node_page_state(nid, NR_ANON_PAGES)),
+#endif
+                      nid, K(node_page_state(nid, NR_SHMEM)),
+                      nid, node_page_state(nid, NR_KERNEL_STACK) *
+                               THREAD_SIZE / 1024,
                       nid, K(node_page_state(nid, NR_PAGETABLE)),
                       nid, K(node_page_state(nid, NR_UNSTABLE_NFS)),
                       nid, K(node_page_state(nid, NR_BOUNCE)),
@@ -123,16 +141,23 @@ static ssize_t node_read_meminfo(struct sys_device * dev,
                       nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE) +
                                node_page_state(nid, NR_SLAB_UNRECLAIMABLE)),
                       nid, K(node_page_state(nid, NR_SLAB_RECLAIMABLE)),
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+                      nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE))
+                       , nid,
+                       K(node_page_state(nid, NR_ANON_TRANSPARENT_HUGEPAGES) *
+                       HPAGE_PMD_NR));
+#else
                       nid, K(node_page_state(nid, NR_SLAB_UNRECLAIMABLE)));
+#endif
        n += hugetlb_report_node_meminfo(nid, buf + n);
        return n;
 }
 
 #undef K
-static SYSDEV_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
+static DEVICE_ATTR(meminfo, S_IRUGO, node_read_meminfo, NULL);
 
-static ssize_t node_read_numastat(struct sys_device * dev,
-                               struct sysdev_attribute *attr, char * buf)
+static ssize_t node_read_numastat(struct device *dev,
+                               struct device_attribute *attr, char *buf)
 {
        return sprintf(buf,
                       "numa_hit %lu\n"
@@ -148,17 +173,35 @@ static ssize_t node_read_numastat(struct sys_device * dev,
                       node_page_state(dev->id, NUMA_LOCAL),
                       node_page_state(dev->id, NUMA_OTHER));
 }
-static SYSDEV_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
+static DEVICE_ATTR(numastat, S_IRUGO, node_read_numastat, NULL);
+
+static ssize_t node_read_vmstat(struct device *dev,
+                               struct device_attribute *attr, char *buf)
+{
+       int nid = dev->id;
+       int i;
+       int n = 0;
+
+       for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
+               n += sprintf(buf+n, "%s %lu\n", vmstat_text[i],
+                            node_page_state(nid, i));
+
+       return n;
+}
+static DEVICE_ATTR(vmstat, S_IRUGO, node_read_vmstat, NULL);
 
-static ssize_t node_read_distance(struct sys_device * dev,
-                       struct sysdev_attribute *attr, char * buf)
+static ssize_t node_read_distance(struct device *dev,
+                       struct device_attribute *attr, char * buf)
 {
        int nid = dev->id;
        int len = 0;
        int i;
 
-       /* buf currently PAGE_SIZE, need ~4 chars per node */
-       BUILD_BUG_ON(MAX_NUMNODES*4 > PAGE_SIZE/2);
+       /*
+        * buf is currently PAGE_SIZE in length and each node needs 4 chars
+        * at the most (distance + space or newline).
+        */
+       BUILD_BUG_ON(MAX_NUMNODES * 4 > PAGE_SIZE);
 
        for_each_online_node(i)
                len += sprintf(buf + len, "%s%d", i ? " " : "", node_distance(nid, i));
@@ -166,7 +209,48 @@ static ssize_t node_read_distance(struct sys_device * dev,
        len += sprintf(buf + len, "\n");
        return len;
 }
-static SYSDEV_ATTR(distance, S_IRUGO, node_read_distance, NULL);
+static DEVICE_ATTR(distance, S_IRUGO, node_read_distance, NULL);
+
+#ifdef CONFIG_HUGETLBFS
+/*
+ * hugetlbfs per node attributes registration interface:
+ * When/if hugetlb[fs] subsystem initializes [sometime after this module],
+ * it will register its per node attributes for all online nodes with
+ * memory.  It will also call register_hugetlbfs_with_node(), below, to
+ * register its attribute registration functions with this node driver.
+ * Once these hooks have been initialized, the node driver will call into
+ * the hugetlb module to [un]register attributes for hot-plugged nodes.
+ */
+static node_registration_func_t __hugetlb_register_node;
+static node_registration_func_t __hugetlb_unregister_node;
+
+static inline bool hugetlb_register_node(struct node *node)
+{
+       if (__hugetlb_register_node &&
+                       node_state(node->dev.id, N_HIGH_MEMORY)) {
+               __hugetlb_register_node(node);
+               return true;
+       }
+       return false;
+}
+
+static inline void hugetlb_unregister_node(struct node *node)
+{
+       if (__hugetlb_unregister_node)
+               __hugetlb_unregister_node(node);
+}
+
+void register_hugetlbfs_with_node(node_registration_func_t doregister,
+                                 node_registration_func_t unregister)
+{
+       __hugetlb_register_node   = doregister;
+       __hugetlb_unregister_node = unregister;
+}
+#else
+static inline void hugetlb_register_node(struct node *node) {}
+
+static inline void hugetlb_unregister_node(struct node *node) {}
+#endif
 
 
 /*
@@ -179,18 +263,23 @@ int register_node(struct node *node, int num, struct node *parent)
 {
        int error;
 
-       node->sysdev.id = num;
-       node->sysdev.cls = &node_class;
-       error = sysdev_register(&node->sysdev);
+       node->dev.id = num;
+       node->dev.bus = &node_subsys;
+       error = device_register(&node->dev);
 
        if (!error){
-               sysdev_create_file(&node->sysdev, &attr_cpumap);
-               sysdev_create_file(&node->sysdev, &attr_cpulist);
-               sysdev_create_file(&node->sysdev, &attr_meminfo);
-               sysdev_create_file(&node->sysdev, &attr_numastat);
-               sysdev_create_file(&node->sysdev, &attr_distance);
+               device_create_file(&node->dev, &dev_attr_cpumap);
+               device_create_file(&node->dev, &dev_attr_cpulist);
+               device_create_file(&node->dev, &dev_attr_meminfo);
+               device_create_file(&node->dev, &dev_attr_numastat);
+               device_create_file(&node->dev, &dev_attr_distance);
+               device_create_file(&node->dev, &dev_attr_vmstat);
 
                scan_unevictable_register_node(node);
+
+               hugetlb_register_node(node);
+
+               compaction_register_node(node);
        }
        return error;
 }
@@ -204,15 +293,17 @@ int register_node(struct node *node, int num, struct node *parent)
  */
 void unregister_node(struct node *node)
 {
-       sysdev_remove_file(&node->sysdev, &attr_cpumap);
-       sysdev_remove_file(&node->sysdev, &attr_cpulist);
-       sysdev_remove_file(&node->sysdev, &attr_meminfo);
-       sysdev_remove_file(&node->sysdev, &attr_numastat);
-       sysdev_remove_file(&node->sysdev, &attr_distance);
+       device_remove_file(&node->dev, &dev_attr_cpumap);
+       device_remove_file(&node->dev, &dev_attr_cpulist);
+       device_remove_file(&node->dev, &dev_attr_meminfo);
+       device_remove_file(&node->dev, &dev_attr_numastat);
+       device_remove_file(&node->dev, &dev_attr_distance);
+       device_remove_file(&node->dev, &dev_attr_vmstat);
 
        scan_unevictable_unregister_node(node);
+       hugetlb_unregister_node(node);          /* no-op, if memoryless node */
 
-       sysdev_unregister(&node->sysdev);
+       device_unregister(&node->dev);
 }
 
 struct node node_devices[MAX_NUMNODES];
@@ -222,26 +313,43 @@ struct node node_devices[MAX_NUMNODES];
  */
 int register_cpu_under_node(unsigned int cpu, unsigned int nid)
 {
-       if (node_online(nid)) {
-               struct sys_device *obj = get_cpu_sysdev(cpu);
-               if (!obj)
-                       return 0;
-               return sysfs_create_link(&node_devices[nid].sysdev.kobj,
-                                        &obj->kobj,
-                                        kobject_name(&obj->kobj));
-        }
+       int ret;
+       struct device *obj;
 
-       return 0;
+       if (!node_online(nid))
+               return 0;
+
+       obj = get_cpu_device(cpu);
+       if (!obj)
+               return 0;
+
+       ret = sysfs_create_link(&node_devices[nid].dev.kobj,
+                               &obj->kobj,
+                               kobject_name(&obj->kobj));
+       if (ret)
+               return ret;
+
+       return sysfs_create_link(&obj->kobj,
+                                &node_devices[nid].dev.kobj,
+                                kobject_name(&node_devices[nid].dev.kobj));
 }
 
 int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
 {
-       if (node_online(nid)) {
-               struct sys_device *obj = get_cpu_sysdev(cpu);
-               if (obj)
-                       sysfs_remove_link(&node_devices[nid].sysdev.kobj,
-                                        kobject_name(&obj->kobj));
-       }
+       struct device *obj;
+
+       if (!node_online(nid))
+               return 0;
+
+       obj = get_cpu_device(cpu);
+       if (!obj)
+               return 0;
+
+       sysfs_remove_link(&node_devices[nid].dev.kobj,
+                         kobject_name(&obj->kobj));
+       sysfs_remove_link(&obj->kobj,
+                         kobject_name(&node_devices[nid].dev.kobj));
+
        return 0;
 }
 
@@ -263,14 +371,17 @@ static int get_nid_for_pfn(unsigned long pfn)
 /* register memory section under specified node if it spans that node */
 int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
 {
+       int ret;
        unsigned long pfn, sect_start_pfn, sect_end_pfn;
 
        if (!mem_blk)
                return -EFAULT;
        if (!node_online(nid))
                return 0;
-       sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
-       sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
+
+       sect_start_pfn = section_nr_to_pfn(mem_blk->start_section_nr);
+       sect_end_pfn = section_nr_to_pfn(mem_blk->end_section_nr);
+       sect_end_pfn += PAGES_PER_SECTION - 1;
        for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
                int page_nid;
 
@@ -279,24 +390,36 @@ int register_mem_sect_under_node(struct memory_block *mem_blk, int nid)
                        continue;
                if (page_nid != nid)
                        continue;
-               return sysfs_create_link_nowarn(&node_devices[nid].sysdev.kobj,
-                                       &mem_blk->sysdev.kobj,
-                                       kobject_name(&mem_blk->sysdev.kobj));
+               ret = sysfs_create_link_nowarn(&node_devices[nid].dev.kobj,
+                                       &mem_blk->dev.kobj,
+                                       kobject_name(&mem_blk->dev.kobj));
+               if (ret)
+                       return ret;
+
+               return sysfs_create_link_nowarn(&mem_blk->dev.kobj,
+                               &node_devices[nid].dev.kobj,
+                               kobject_name(&node_devices[nid].dev.kobj));
        }
        /* mem section does not span the specified node */
        return 0;
 }
 
 /* unregister memory section under all nodes that it spans */
-int unregister_mem_sect_under_nodes(struct memory_block *mem_blk)
+int unregister_mem_sect_under_nodes(struct memory_block *mem_blk,
+                                   unsigned long phys_index)
 {
-       nodemask_t unlinked_nodes;
+       NODEMASK_ALLOC(nodemask_t, unlinked_nodes, GFP_KERNEL);
        unsigned long pfn, sect_start_pfn, sect_end_pfn;
 
-       if (!mem_blk)
+       if (!mem_blk) {
+               NODEMASK_FREE(unlinked_nodes);
                return -EFAULT;
-       nodes_clear(unlinked_nodes);
-       sect_start_pfn = section_nr_to_pfn(mem_blk->phys_index);
+       }
+       if (!unlinked_nodes)
+               return -ENOMEM;
+       nodes_clear(*unlinked_nodes);
+
+       sect_start_pfn = section_nr_to_pfn(phys_index);
        sect_end_pfn = sect_start_pfn + PAGES_PER_SECTION - 1;
        for (pfn = sect_start_pfn; pfn <= sect_end_pfn; pfn++) {
                int nid;
@@ -306,11 +429,14 @@ int unregister_mem_sect_under_nodes(struct memory_block *mem_blk)
                        continue;
                if (!node_online(nid))
                        continue;
-               if (node_test_and_set(nid, unlinked_nodes))
+               if (node_test_and_set(nid, *unlinked_nodes))
                        continue;
-               sysfs_remove_link(&node_devices[nid].sysdev.kobj,
-                        kobject_name(&mem_blk->sysdev.kobj));
+               sysfs_remove_link(&node_devices[nid].dev.kobj,
+                        kobject_name(&mem_blk->dev.kobj));
+               sysfs_remove_link(&mem_blk->dev.kobj,
+                        kobject_name(&node_devices[nid].dev.kobj));
        }
+       NODEMASK_FREE(unlinked_nodes);
        return 0;
 }
 
@@ -319,30 +445,108 @@ static int link_mem_sections(int nid)
        unsigned long start_pfn = NODE_DATA(nid)->node_start_pfn;
        unsigned long end_pfn = start_pfn + NODE_DATA(nid)->node_spanned_pages;
        unsigned long pfn;
+       struct memory_block *mem_blk = NULL;
        int err = 0;
 
        for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
                unsigned long section_nr = pfn_to_section_nr(pfn);
                struct mem_section *mem_sect;
-               struct memory_block *mem_blk;
                int ret;
 
                if (!present_section_nr(section_nr))
                        continue;
                mem_sect = __nr_to_section(section_nr);
-               mem_blk = find_memory_block(mem_sect);
+
+               /* same memblock ? */
+               if (mem_blk)
+                       if ((section_nr >= mem_blk->start_section_nr) &&
+                           (section_nr <= mem_blk->end_section_nr))
+                               continue;
+
+               mem_blk = find_memory_block_hinted(mem_sect, mem_blk);
+
                ret = register_mem_sect_under_node(mem_blk, nid);
                if (!err)
                        err = ret;
 
                /* discard ref obtained in find_memory_block() */
-               kobject_put(&mem_blk->sysdev.kobj);
        }
+
+       if (mem_blk)
+               kobject_put(&mem_blk->dev.kobj);
        return err;
 }
-#else
+
+#ifdef CONFIG_HUGETLBFS
+/*
+ * Handle per node hstate attribute [un]registration on transistions
+ * to/from memoryless state.
+ */
+static void node_hugetlb_work(struct work_struct *work)
+{
+       struct node *node = container_of(work, struct node, node_work);
+
+       /*
+        * We only get here when a node transitions to/from memoryless state.
+        * We can detect which transition occurred by examining whether the
+        * node has memory now.  hugetlb_register_node() already check this
+        * so we try to register the attributes.  If that fails, then the
+        * node has transitioned to memoryless, try to unregister the
+        * attributes.
+        */
+       if (!hugetlb_register_node(node))
+               hugetlb_unregister_node(node);
+}
+
+static void init_node_hugetlb_work(int nid)
+{
+       INIT_WORK(&node_devices[nid].node_work, node_hugetlb_work);
+}
+
+static int node_memory_callback(struct notifier_block *self,
+                               unsigned long action, void *arg)
+{
+       struct memory_notify *mnb = arg;
+       int nid = mnb->status_change_nid;
+
+       switch (action) {
+       case MEM_ONLINE:
+       case MEM_OFFLINE:
+               /*
+                * offload per node hstate [un]registration to a work thread
+                * when transitioning to/from memoryless state.
+                */
+               if (nid != NUMA_NO_NODE)
+                       schedule_work(&node_devices[nid].node_work);
+               break;
+
+       case MEM_GOING_ONLINE:
+       case MEM_GOING_OFFLINE:
+       case MEM_CANCEL_ONLINE:
+       case MEM_CANCEL_OFFLINE:
+       default:
+               break;
+       }
+
+       return NOTIFY_OK;
+}
+#endif /* CONFIG_HUGETLBFS */
+#else  /* !CONFIG_MEMORY_HOTPLUG_SPARSE */
+
 static int link_mem_sections(int nid) { return 0; }
-#endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
+#endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */
+
+#if !defined(CONFIG_MEMORY_HOTPLUG_SPARSE) || \
+    !defined(CONFIG_HUGETLBFS)
+static inline int node_memory_callback(struct notifier_block *self,
+                               unsigned long action, void *arg)
+{
+       return NOTIFY_OK;
+}
+
+static void init_node_hugetlb_work(int nid) { }
+
+#endif
 
 int register_one_node(int nid)
 {
@@ -366,6 +570,9 @@ int register_one_node(int nid)
 
                /* link memory sections under this node */
                error = link_mem_sections(nid);
+
+               /* initialize work queue for memory hot plug */
+               init_node_hugetlb_work(nid);
        }
 
        return error;
@@ -393,75 +600,64 @@ static ssize_t print_nodes_state(enum node_states state, char *buf)
        return n;
 }
 
-static ssize_t print_nodes_possible(struct sysdev_class *class, char *buf)
-{
-       return print_nodes_state(N_POSSIBLE, buf);
-}
-
-static ssize_t print_nodes_online(struct sysdev_class *class, char *buf)
-{
-       return print_nodes_state(N_ONLINE, buf);
-}
-
-static ssize_t print_nodes_has_normal_memory(struct sysdev_class *class,
-                                               char *buf)
-{
-       return print_nodes_state(N_NORMAL_MEMORY, buf);
-}
+struct node_attr {
+       struct device_attribute attr;
+       enum node_states state;
+};
 
-static ssize_t print_nodes_has_cpu(struct sysdev_class *class, char *buf)
+static ssize_t show_node_state(struct device *dev,
+                              struct device_attribute *attr, char *buf)
 {
-       return print_nodes_state(N_CPU, buf);
+       struct node_attr *na = container_of(attr, struct node_attr, attr);
+       return print_nodes_state(na->state, buf);
 }
 
-static SYSDEV_CLASS_ATTR(possible, 0444, print_nodes_possible, NULL);
-static SYSDEV_CLASS_ATTR(online, 0444, print_nodes_online, NULL);
-static SYSDEV_CLASS_ATTR(has_normal_memory, 0444, print_nodes_has_normal_memory,
-                                                                       NULL);
-static SYSDEV_CLASS_ATTR(has_cpu, 0444, print_nodes_has_cpu, NULL);
+#define _NODE_ATTR(name, state) \
+       { __ATTR(name, 0444, show_node_state, NULL), state }
 
+static struct node_attr node_state_attr[] = {
+       _NODE_ATTR(possible, N_POSSIBLE),
+       _NODE_ATTR(online, N_ONLINE),
+       _NODE_ATTR(has_normal_memory, N_NORMAL_MEMORY),
+       _NODE_ATTR(has_cpu, N_CPU),
 #ifdef CONFIG_HIGHMEM
-static ssize_t print_nodes_has_high_memory(struct sysdev_class *class,
-                                                char *buf)
-{
-       return print_nodes_state(N_HIGH_MEMORY, buf);
-}
-
-static SYSDEV_CLASS_ATTR(has_high_memory, 0444, print_nodes_has_high_memory,
-                                                                        NULL);
+       _NODE_ATTR(has_high_memory, N_HIGH_MEMORY),
 #endif
+};
 
-struct sysdev_class_attribute *node_state_attr[] = {
-       &attr_possible,
-       &attr_online,
-       &attr_has_normal_memory,
+static struct attribute *node_state_attrs[] = {
+       &node_state_attr[0].attr.attr,
+       &node_state_attr[1].attr.attr,
+       &node_state_attr[2].attr.attr,
+       &node_state_attr[3].attr.attr,
 #ifdef CONFIG_HIGHMEM
-       &attr_has_high_memory,
+       &node_state_attr[4].attr.attr,
 #endif
-       &attr_has_cpu,
+       NULL
 };
 
-static int node_states_init(void)
-{
-       int i;
-       int err = 0;
+static struct attribute_group memory_root_attr_group = {
+       .attrs = node_state_attrs,
+};
 
-       for (i = 0;  i < NR_NODE_STATES; i++) {
-               int ret;
-               ret = sysdev_class_create_file(&node_class, node_state_attr[i]);
-               if (!err)
-                       err = ret;
-       }
-       return err;
-}
+static const struct attribute_group *cpu_root_attr_groups[] = {
+       &memory_root_attr_group,
+       NULL,
+};
 
+#define NODE_CALLBACK_PRI      2       /* lower than SLAB */
 static int __init register_node_type(void)
 {
        int ret;
 
-       ret = sysdev_class_register(&node_class);
-       if (!ret)
-               ret = node_states_init();
+       BUILD_BUG_ON(ARRAY_SIZE(node_state_attr) != NR_NODE_STATES);
+       BUILD_BUG_ON(ARRAY_SIZE(node_state_attrs)-1 != NR_NODE_STATES);
+
+       ret = subsys_system_register(&node_subsys, cpu_root_attr_groups);
+       if (!ret) {
+               hotplug_memory_notifier(node_memory_callback,
+                                       NODE_CALLBACK_PRI);
+       }
 
        /*
         * Note:  we're not going to unregister the node class if we fail