headers: utsname.h redux
[safe/jmp/linux-2.6] / fs / ocfs2 / stackglue.c
index 76ae4fc..3f2f1c4 100644 (file)
@@ -6,7 +6,7 @@
  * Code which implements an OCFS2 specific interface to underlying
  * cluster stacks.
  *
- * Copyright (C) 2007 Oracle.  All rights reserved.
+ * Copyright (C) 2007, 2009 Oracle.  All rights reserved.
  *
  * This program is free software; you can redistribute it and/or
  * modify it under the terms of the GNU General Public
 #include <linux/fs.h>
 #include <linux/kobject.h>
 #include <linux/sysfs.h>
+#include <linux/sysctl.h>
+
+#include "ocfs2_fs.h"
 
 #include "stackglue.h"
 
+#define OCFS2_STACK_PLUGIN_O2CB                "o2cb"
+#define OCFS2_STACK_PLUGIN_USER                "user"
+#define OCFS2_MAX_HB_CTL_PATH          256
+
 static struct ocfs2_locking_protocol *lproto;
 static DEFINE_SPINLOCK(ocfs2_stack_lock);
 static LIST_HEAD(ocfs2_stack_list);
+static char cluster_stack_name[OCFS2_STACK_LABEL_LEN + 1];
+static char ocfs2_hb_ctl_path[OCFS2_MAX_HB_CTL_PATH] = "/sbin/ocfs2_hb_ctl";
 
 /*
  * The stack currently in use.  If not null, active_stack->sp_count > 0,
@@ -53,38 +62,49 @@ static struct ocfs2_stack_plugin *ocfs2_stack_lookup(const char *name)
        return NULL;
 }
 
-static int ocfs2_stack_driver_request(const char *name)
+static int ocfs2_stack_driver_request(const char *stack_name,
+                                     const char *plugin_name)
 {
        int rc;
        struct ocfs2_stack_plugin *p;
 
        spin_lock(&ocfs2_stack_lock);
 
+       /*
+        * If the stack passed by the filesystem isn't the selected one,
+        * we can't continue.
+        */
+       if (strcmp(stack_name, cluster_stack_name)) {
+               rc = -EBUSY;
+               goto out;
+       }
+
        if (active_stack) {
                /*
                 * If the active stack isn't the one we want, it cannot
                 * be selected right now.
                 */
-               if (!strcmp(active_stack->sp_name, name))
+               if (!strcmp(active_stack->sp_name, plugin_name))
                        rc = 0;
                else
                        rc = -EBUSY;
                goto out;
        }
 
-       p = ocfs2_stack_lookup(name);
+       p = ocfs2_stack_lookup(plugin_name);
        if (!p || !try_module_get(p->sp_owner)) {
                rc = -ENOENT;
                goto out;
        }
 
-       /* Ok, the stack is pinned */
-       p->sp_count++;
        active_stack = p;
-
        rc = 0;
 
 out:
+       /* If we found it, pin it */
+       if (!rc)
+               active_stack->sp_count++;
+
        spin_unlock(&ocfs2_stack_lock);
        return rc;
 }
@@ -94,23 +114,42 @@ out:
  * there is no stack, it tries to load it.  It will fail if the stack still
  * cannot be found.  It will also fail if a different stack is in use.
  */
-static int ocfs2_stack_driver_get(const char *name)
+static int ocfs2_stack_driver_get(const char *stack_name)
 {
        int rc;
+       char *plugin_name = OCFS2_STACK_PLUGIN_O2CB;
+
+       /*
+        * Classic stack does not pass in a stack name.  This is
+        * compatible with older tools as well.
+        */
+       if (!stack_name || !*stack_name)
+               stack_name = OCFS2_STACK_PLUGIN_O2CB;
+
+       if (strlen(stack_name) != OCFS2_STACK_LABEL_LEN) {
+               printk(KERN_ERR
+                      "ocfs2 passed an invalid cluster stack label: \"%s\"\n",
+                      stack_name);
+               return -EINVAL;
+       }
 
-       rc = ocfs2_stack_driver_request(name);
+       /* Anything that isn't the classic stack is a user stack */
+       if (strcmp(stack_name, OCFS2_STACK_PLUGIN_O2CB))
+               plugin_name = OCFS2_STACK_PLUGIN_USER;
+
+       rc = ocfs2_stack_driver_request(stack_name, plugin_name);
        if (rc == -ENOENT) {
-               request_module("ocfs2_stack_%s", name);
-               rc = ocfs2_stack_driver_request(name);
+               request_module("ocfs2_stack_%s", plugin_name);
+               rc = ocfs2_stack_driver_request(stack_name, plugin_name);
        }
 
        if (rc == -ENOENT) {
                printk(KERN_ERR
                       "ocfs2: Cluster stack driver \"%s\" cannot be found\n",
-                      name);
+                      plugin_name);
        } else if (rc == -EBUSY) {
                printk(KERN_ERR
-                      "ocfs2: A different cluster stack driver is in use\n");
+                      "ocfs2: A different cluster stack is in use\n");
        }
 
        return rc;
@@ -193,13 +232,20 @@ void ocfs2_stack_glue_set_locking_protocol(struct ocfs2_locking_protocol *proto)
 EXPORT_SYMBOL_GPL(ocfs2_stack_glue_set_locking_protocol);
 
 
+/*
+ * The ocfs2_dlm_lock() and ocfs2_dlm_unlock() functions take
+ * "struct ocfs2_lock_res *astarg" instead of "void *astarg" because the
+ * underlying stack plugins need to pilfer the lksb off of the lock_res.
+ * If some other structure needs to be passed as an astarg, the plugins
+ * will need to be given a different avenue to the lksb.
+ */
 int ocfs2_dlm_lock(struct ocfs2_cluster_connection *conn,
                   int mode,
                   union ocfs2_dlm_lksb *lksb,
                   u32 flags,
                   void *name,
                   unsigned int namelen,
-                  void *astarg)
+                  struct ocfs2_lock_res *astarg)
 {
        BUG_ON(lproto == NULL);
 
@@ -211,7 +257,7 @@ EXPORT_SYMBOL_GPL(ocfs2_dlm_lock);
 int ocfs2_dlm_unlock(struct ocfs2_cluster_connection *conn,
                     union ocfs2_dlm_lksb *lksb,
                     u32 flags,
-                    void *astarg)
+                    struct ocfs2_lock_res *astarg)
 {
        BUG_ON(lproto == NULL);
 
@@ -225,11 +271,12 @@ int ocfs2_dlm_lock_status(union ocfs2_dlm_lksb *lksb)
 }
 EXPORT_SYMBOL_GPL(ocfs2_dlm_lock_status);
 
-/*
- * Why don't we cast to ocfs2_meta_lvb?  The "clean" answer is that we
- * don't cast at the glue level.  The real answer is that the header
- * ordering is nigh impossible.
- */
+int ocfs2_dlm_lvb_valid(union ocfs2_dlm_lksb *lksb)
+{
+       return active_stack->sp_ops->lvb_valid(lksb);
+}
+EXPORT_SYMBOL_GPL(ocfs2_dlm_lvb_valid);
+
 void *ocfs2_dlm_lvb(union ocfs2_dlm_lksb *lksb)
 {
        return active_stack->sp_ops->lock_lvb(lksb);
@@ -242,7 +289,28 @@ void ocfs2_dlm_dump_lksb(union ocfs2_dlm_lksb *lksb)
 }
 EXPORT_SYMBOL_GPL(ocfs2_dlm_dump_lksb);
 
-int ocfs2_cluster_connect(const char *group,
+int ocfs2_stack_supports_plocks(void)
+{
+       return active_stack && active_stack->sp_ops->plock;
+}
+EXPORT_SYMBOL_GPL(ocfs2_stack_supports_plocks);
+
+/*
+ * ocfs2_plock() can only be safely called if
+ * ocfs2_stack_supports_plocks() returned true
+ */
+int ocfs2_plock(struct ocfs2_cluster_connection *conn, u64 ino,
+               struct file *file, int cmd, struct file_lock *fl)
+{
+       WARN_ON_ONCE(active_stack->sp_ops->plock == NULL);
+       if (active_stack->sp_ops->plock)
+               return active_stack->sp_ops->plock(conn, ino, file, cmd, fl);
+       return -EOPNOTSUPP;
+}
+EXPORT_SYMBOL_GPL(ocfs2_plock);
+
+int ocfs2_cluster_connect(const char *stack_name,
+                         const char *group,
                          int grouplen,
                          void (*recovery_handler)(int node_num,
                                                   void *recovery_data),
@@ -277,7 +345,7 @@ int ocfs2_cluster_connect(const char *group,
        new_conn->cc_version = lproto->lp_max_version;
 
        /* This will pin the stack driver if successful */
-       rc = ocfs2_stack_driver_get("o2cb");
+       rc = ocfs2_stack_driver_get(stack_name);
        if (rc)
                goto out_free;
 
@@ -306,7 +374,7 @@ int ocfs2_cluster_disconnect(struct ocfs2_cluster_connection *conn,
 
        BUG_ON(conn == NULL);
 
-       ret = active_stack->sp_ops->disconnect(conn, hangup_pending);
+       ret = active_stack->sp_ops->disconnect(conn);
 
        /* XXX Should we free it anyway? */
        if (!ret) {
@@ -319,12 +387,48 @@ int ocfs2_cluster_disconnect(struct ocfs2_cluster_connection *conn,
 }
 EXPORT_SYMBOL_GPL(ocfs2_cluster_disconnect);
 
+/*
+ * Leave the group for this filesystem.  This is executed by a userspace
+ * program (stored in ocfs2_hb_ctl_path).
+ */
+static void ocfs2_leave_group(const char *group)
+{
+       int ret;
+       char *argv[5], *envp[3];
+
+       argv[0] = ocfs2_hb_ctl_path;
+       argv[1] = "-K";
+       argv[2] = "-u";
+       argv[3] = (char *)group;
+       argv[4] = NULL;
+
+       /* minimal command environment taken from cpu_run_sbin_hotplug */
+       envp[0] = "HOME=/";
+       envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
+       envp[2] = NULL;
+
+       ret = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_PROC);
+       if (ret < 0) {
+               printk(KERN_ERR
+                      "ocfs2: Error %d running user helper "
+                      "\"%s %s %s %s\"\n",
+                      ret, argv[0], argv[1], argv[2], argv[3]);
+       }
+}
+
+/*
+ * Hangup is a required post-umount.  ocfs2-tools software expects the
+ * filesystem to call "ocfs2_hb_ctl" during unmount.  This happens
+ * regardless of whether the DLM got started, so we can't do it
+ * in ocfs2_cluster_disconnect().  The ocfs2_leave_group() function does
+ * the actual work.
+ */
 void ocfs2_cluster_hangup(const char *group, int grouplen)
 {
        BUG_ON(group == NULL);
        BUG_ON(group[grouplen] != '\0');
 
-       active_stack->sp_ops->hangup(group, grouplen);
+       ocfs2_leave_group(group);
 
        /* cluster_disconnect() was called with hangup_pending==1 */
        ocfs2_stack_driver_put();
@@ -416,10 +520,61 @@ static struct kobj_attribute ocfs2_attr_active_cluster_plugin =
        __ATTR(active_cluster_plugin, S_IFREG | S_IRUGO,
               ocfs2_active_cluster_plugin_show, NULL);
 
+static ssize_t ocfs2_cluster_stack_show(struct kobject *kobj,
+                                       struct kobj_attribute *attr,
+                                       char *buf)
+{
+       ssize_t ret;
+       spin_lock(&ocfs2_stack_lock);
+       ret = snprintf(buf, PAGE_SIZE, "%s\n", cluster_stack_name);
+       spin_unlock(&ocfs2_stack_lock);
+
+       return ret;
+}
+
+static ssize_t ocfs2_cluster_stack_store(struct kobject *kobj,
+                                        struct kobj_attribute *attr,
+                                        const char *buf, size_t count)
+{
+       size_t len = count;
+       ssize_t ret;
+
+       if (len == 0)
+               return len;
+
+       if (buf[len - 1] == '\n')
+               len--;
+
+       if ((len != OCFS2_STACK_LABEL_LEN) ||
+           (strnlen(buf, len) != len))
+               return -EINVAL;
+
+       spin_lock(&ocfs2_stack_lock);
+       if (active_stack) {
+               if (!strncmp(buf, cluster_stack_name, len))
+                       ret = count;
+               else
+                       ret = -EBUSY;
+       } else {
+               memcpy(cluster_stack_name, buf, len);
+               ret = count;
+       }
+       spin_unlock(&ocfs2_stack_lock);
+
+       return ret;
+}
+
+
+static struct kobj_attribute ocfs2_attr_cluster_stack =
+       __ATTR(cluster_stack, S_IFREG | S_IRUGO | S_IWUSR,
+              ocfs2_cluster_stack_show,
+              ocfs2_cluster_stack_store);
+
 static struct attribute *ocfs2_attrs[] = {
        &ocfs2_attr_max_locking_protocol.attr,
        &ocfs2_attr_loaded_cluster_plugins.attr,
        &ocfs2_attr_active_cluster_plugin.attr,
+       &ocfs2_attr_cluster_stack.attr,
        NULL,
 };
 
@@ -453,8 +608,83 @@ error:
        return ret;
 }
 
+/*
+ * Sysctl bits
+ *
+ * The sysctl lives at /proc/sys/fs/ocfs2/nm/hb_ctl_path.  The 'nm' doesn't
+ * make as much sense in a multiple cluster stack world, but it's safer
+ * and easier to preserve the name.
+ */
+
+#define FS_OCFS2_NM            1
+
+static ctl_table ocfs2_nm_table[] = {
+       {
+               .ctl_name       = 1,
+               .procname       = "hb_ctl_path",
+               .data           = ocfs2_hb_ctl_path,
+               .maxlen         = OCFS2_MAX_HB_CTL_PATH,
+               .mode           = 0644,
+               .proc_handler   = &proc_dostring,
+               .strategy       = &sysctl_string,
+       },
+       { .ctl_name = 0 }
+};
+
+static ctl_table ocfs2_mod_table[] = {
+       {
+               .ctl_name       = FS_OCFS2_NM,
+               .procname       = "nm",
+               .data           = NULL,
+               .maxlen         = 0,
+               .mode           = 0555,
+               .child          = ocfs2_nm_table
+       },
+       { .ctl_name = 0}
+};
+
+static ctl_table ocfs2_kern_table[] = {
+       {
+               .ctl_name       = FS_OCFS2,
+               .procname       = "ocfs2",
+               .data           = NULL,
+               .maxlen         = 0,
+               .mode           = 0555,
+               .child          = ocfs2_mod_table
+       },
+       { .ctl_name = 0}
+};
+
+static ctl_table ocfs2_root_table[] = {
+       {
+               .ctl_name       = CTL_FS,
+               .procname       = "fs",
+               .data           = NULL,
+               .maxlen         = 0,
+               .mode           = 0555,
+               .child          = ocfs2_kern_table
+       },
+       { .ctl_name = 0 }
+};
+
+static struct ctl_table_header *ocfs2_table_header = NULL;
+
+
+/*
+ * Initialization
+ */
+
 static int __init ocfs2_stack_glue_init(void)
 {
+       strcpy(cluster_stack_name, OCFS2_STACK_PLUGIN_O2CB);
+
+       ocfs2_table_header = register_sysctl_table(ocfs2_root_table);
+       if (!ocfs2_table_header) {
+               printk(KERN_ERR
+                      "ocfs2 stack glue: unable to register sysctl\n");
+               return -ENOMEM; /* or something. */
+       }
+
        return ocfs2_sysfs_init();
 }
 
@@ -462,6 +692,8 @@ static void __exit ocfs2_stack_glue_exit(void)
 {
        lproto = NULL;
        ocfs2_sysfs_exit();
+       if (ocfs2_table_header)
+               unregister_sysctl_table(ocfs2_table_header);
 }
 
 MODULE_AUTHOR("Oracle");