Whamcloud - gitweb
LU-5589 obdclass: fix NULL pointer dereference
[fs/lustre-release.git] / lustre / obdclass / obd_mount_server.c
index 01fd9e0..0059604 100644 (file)
@@ -27,7 +27,7 @@
  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2011, 2012, Intel Corporation.
+ * Copyright (c) 2013, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
 #define PRINT_CMD CDEBUG
 #define PRINT_MASK (D_SUPER | D_CONFIG)
 
+#include <linux/statfs.h>
 #include <obd.h>
-#include <lvfs.h>
-#include <lustre_fsfilt.h>
 #include <obd_class.h>
 #include <lustre/lustre_user.h>
 #include <linux/version.h>
+#include <lustre_ioctl.h>
 #include <lustre_log.h>
 #include <lustre_disk.h>
 #include <lustre_param.h>
 #ifdef HAVE_KERNEL_LOCKED
 #include <linux/smp_lock.h>
 #endif
+#ifdef HAVE_SELINUX_IS_ENABLED
+#include <linux/selinux.h>
+#endif
 
 /*********** mount lookup *********/
 
 DEFINE_MUTEX(lustre_mount_info_lock);
-static CFS_LIST_HEAD(server_mount_info_list);
+static struct list_head server_mount_info_list =
+       LIST_HEAD_INIT(server_mount_info_list);
 
 static struct lustre_mount_info *server_find_mount(const char *name)
 {
-       cfs_list_t *tmp;
+       struct list_head *tmp;
        struct lustre_mount_info *lmi;
        ENTRY;
 
-       cfs_list_for_each(tmp, &server_mount_info_list) {
-               lmi = cfs_list_entry(tmp, struct lustre_mount_info,
-                                    lmi_list_chain);
+       list_for_each(tmp, &server_mount_info_list) {
+               lmi = list_entry(tmp, struct lustre_mount_info,
+                                lmi_list_chain);
                if (strcmp(name, lmi->lmi_name) == 0)
                        RETURN(lmi);
        }
@@ -82,8 +86,7 @@ static struct lustre_mount_info *server_find_mount(const char *name)
 /* we must register an obd for a mount before we call the setup routine.
  *_setup will call lustre_get_mount to get the mnt struct
  by obd_name, since we can't pass the pointer to setup. */
-static int server_register_mount(const char *name, struct super_block *sb,
-                                struct vfsmount *mnt)
+static int server_register_mount(const char *name, struct super_block *sb)
 {
        struct lustre_mount_info *lmi;
        char *name_cp;
@@ -112,12 +115,11 @@ static int server_register_mount(const char *name, struct super_block *sb,
        }
        lmi->lmi_name = name_cp;
        lmi->lmi_sb = sb;
-       lmi->lmi_mnt = mnt;
-       cfs_list_add(&lmi->lmi_list_chain, &server_mount_info_list);
+       list_add(&lmi->lmi_list_chain, &server_mount_info_list);
 
        mutex_unlock(&lustre_mount_info_lock);
 
-       CDEBUG(D_MOUNT, "reg_mnt %p from %s\n", lmi->lmi_mnt, name);
+       CDEBUG(D_MOUNT, "register mount %p from %s\n", sb, name);
 
        RETURN(0);
 }
@@ -136,10 +138,10 @@ static int server_deregister_mount(const char *name)
                RETURN(-ENOENT);
        }
 
-       CDEBUG(D_MOUNT, "dereg_mnt %p from %s\n", lmi->lmi_mnt, name);
+       CDEBUG(D_MOUNT, "deregister mount %p from %s\n", lmi->lmi_sb, name);
 
        OBD_FREE(lmi->lmi_name, strlen(lmi->lmi_name) + 1);
-       cfs_list_del(&lmi->lmi_list_chain);
+       list_del(&lmi->lmi_list_chain);
        OBD_FREE(lmi, sizeof(*lmi));
        mutex_unlock(&lustre_mount_info_lock);
 
@@ -164,37 +166,28 @@ struct lustre_mount_info *server_get_mount(const char *name)
        }
        lsi = s2lsi(lmi->lmi_sb);
 
-       cfs_atomic_inc(&lsi->lsi_mounts);
+       atomic_inc(&lsi->lsi_mounts);
 
-       CDEBUG(D_MOUNT, "get_mnt %p from %s, refs=%d\n", lmi->lmi_mnt,
-              name, cfs_atomic_read(&lsi->lsi_mounts));
+       CDEBUG(D_MOUNT, "get mount %p from %s, refs=%d\n", lmi->lmi_sb,
+              name, atomic_read(&lsi->lsi_mounts));
 
        RETURN(lmi);
 }
 EXPORT_SYMBOL(server_get_mount);
 
-/*
- * Used by mdt to get mount_info from obdname.
- * There are no blocking when using the mount_info.
- * Do not use server_get_mount for this purpose.
+/**
+ * server_put_mount: to be called from obd_cleanup methods
+ * @name:      obd name
+ * @dereg_mnt: 0 or 1 depending on whether the mount is to be deregistered or
+ * not
+ *
+ * The caller decides whether server_deregister_mount() needs to be called or
+ * not. Calling of server_deregister_mount() does not depend on refcounting on
+ * lsi because we could have say the mgs and mds on the same node and we
+ * unmount the mds, then the ref on the lsi would still be non-zero but we
+ * would still want to deregister the mds mount.
  */
-struct lustre_mount_info *server_get_mount_2(const char *name)
-{
-       struct lustre_mount_info *lmi;
-       ENTRY;
-
-       mutex_lock(&lustre_mount_info_lock);
-       lmi = server_find_mount(name);
-       mutex_unlock(&lustre_mount_info_lock);
-       if (!lmi)
-               CERROR("Can't find mount for %s\n", name);
-
-       RETURN(lmi);
-}
-EXPORT_SYMBOL(server_get_mount_2);
-
-/* to be called from obd_cleanup methods */
-int server_put_mount(const char *name, struct vfsmount *mnt)
+int server_put_mount(const char *name, bool dereg_mnt)
 {
        struct lustre_mount_info *lmi;
        struct lustre_sb_info *lsi;
@@ -209,33 +202,25 @@ int server_put_mount(const char *name, struct vfsmount *mnt)
        }
        lsi = s2lsi(lmi->lmi_sb);
 
-       CDEBUG(D_MOUNT, "put_mnt %p from %s, refs=%d\n",
-              lmi->lmi_mnt, name, cfs_atomic_read(&lsi->lsi_mounts));
+       CDEBUG(D_MOUNT, "put mount %p from %s, refs=%d\n",
+              lmi->lmi_sb, name, atomic_read(&lsi->lsi_mounts));
 
        if (lustre_put_lsi(lmi->lmi_sb))
-               CDEBUG(D_MOUNT, "Last put of mnt %p from %s\n",
-                      lmi->lmi_mnt, name);
+               CDEBUG(D_MOUNT, "Last put of mount %p from %s\n",
+                      lmi->lmi_sb, name);
 
-       /* this obd should never need the mount again */
-       server_deregister_mount(name);
+       if (dereg_mnt)
+               /* this obd should never need the mount again */
+               server_deregister_mount(name);
 
        RETURN(0);
 }
 EXPORT_SYMBOL(server_put_mount);
 
-/* Corresponding to server_get_mount_2 */
-int server_put_mount_2(const char *name, struct vfsmount *mnt)
-{
-       ENTRY;
-       RETURN(0);
-}
-EXPORT_SYMBOL(server_put_mount_2);
-
 /* Set up a MGS to serve startup logs */
 static int server_start_mgs(struct super_block *sb)
 {
        struct lustre_sb_info    *lsi = s2lsi(sb);
-       struct vfsmount   *mnt = lsi->lsi_srv_mnt;
        struct lustre_mount_info *lmi;
        int    rc = 0;
        ENTRY;
@@ -252,15 +237,20 @@ static int server_start_mgs(struct super_block *sb)
 
        CDEBUG(D_CONFIG, "Start MGS service %s\n", LUSTRE_MGS_OBDNAME);
 
-       rc = server_register_mount(LUSTRE_MGS_OBDNAME, sb, mnt);
+       rc = server_register_mount(LUSTRE_MGS_OBDNAME, sb);
 
        if (!rc) {
                rc = lustre_start_simple(LUSTRE_MGS_OBDNAME, LUSTRE_MGS_NAME,
                                         LUSTRE_MGS_OBDNAME, 0, 0,
                                         lsi->lsi_osd_obdname, 0);
-               /* Do NOT call server_deregister_mount() here. This leads to
-                * inability cleanup cleanly and free lsi and other stuff when
-                * mgs calls server_put_mount() in error handling case. -umka */
+               /* server_deregister_mount() is not called previously, for lsi
+                * and other stuff can't be freed cleanly when mgs calls
+                * server_put_mount() in error handling case (see b=17758),
+                * this problem is caused by a bug in mgs_init0, which forgot
+                * calling server_put_mount in error case. */
+
+               if (rc)
+                       server_deregister_mount(LUSTRE_MGS_OBDNAME);
        }
 
        if (rc)
@@ -324,54 +314,28 @@ static int server_mgc_clear_fs(struct obd_device *mgc)
        RETURN(rc);
 }
 
-static int is_mdc_device(const char *devname)
+static inline bool is_mdc_device(const char *devname)
 {
        char *ptr;
 
        ptr = strrchr(devname, '-');
-       if (ptr != NULL && strcmp(ptr, "-mdc") == 0)
-               return 1;
-
-       return 0;
+       return ptr != NULL && strcmp(ptr, "-mdc") == 0;
 }
 
-static inline int tgt_is_mdt0(const char *tgtname)
+static inline bool tgt_is_mdt(const char *tgtname, __u32 *idx)
 {
-       __u32 idx;
-       int   type;
+       int type;
 
-       type = server_name2index(tgtname, &idx, NULL);
-       if (type != LDD_F_SV_TYPE_MDT)
-               return 0;
+       type = server_name2index(tgtname, idx, NULL);
 
-       return idx == 0;
-}
-
-static inline int is_mdc_for_mdt0(const char *devname)
-{
-       char   *ptr;
-
-       if (!is_mdc_device(devname))
-               return 0;
-
-       ptr = strrchr(devname, '-');
-       if (ptr == NULL)
-               return 0;
-
-       *ptr = 0;
-       if (tgt_is_mdt0(devname)) {
-               *ptr = '-';
-               return 1;
-       }
-       *ptr = '-';
-       return 0;
+       return type == LDD_F_SV_TYPE_MDT;
 }
 
 /**
- * Convert OST/MDT name(fsname-OSTxxxx) to a lwp name
- * (fsname-MDT0000-lwp-OSTxxxx)
+ * Convert OST/MDT name(fsname-{MDT,OST}xxxx) to a lwp name with the @idx:yyyy
+ * (fsname-MDTyyyy-lwp-{MDT,OST}xxxx)
  **/
-int tgt_name2lwpname(const char *svname, char *lwpname)
+int tgt_name2lwp_name(const char *tgt_name, char *lwp_name, int len, __u32 idx)
 {
        char            *fsname;
        const char      *tgt;
@@ -382,32 +346,38 @@ int tgt_name2lwpname(const char *svname, char *lwpname)
        if (fsname == NULL)
                RETURN(-ENOMEM);
 
-       rc = server_name2fsname(svname, fsname, &tgt);
+       rc = server_name2fsname(tgt_name, fsname, &tgt);
        if (rc != 0) {
-               CERROR("%s: failed to get fsname from svname. %d\n",
-                      svname, rc);
+               CERROR("%s: failed to get fsname from tgt_name: rc = %d\n",
+                      tgt_name, rc);
                GOTO(cleanup, rc);
        }
 
        if (*tgt != '-' && *tgt != ':') {
-               CERROR("%s: invalid svname name!\n", svname);
+               CERROR("%s: invalid tgt_name name!\n", tgt_name);
                GOTO(cleanup, rc = -EINVAL);
        }
 
        tgt++;
        if (strncmp(tgt, "OST", 3) != 0 && strncmp(tgt, "MDT", 3) != 0) {
-               CERROR("%s is not an OST or MDT target!\n", svname);
+               CERROR("%s is not an OST or MDT target!\n", tgt_name);
                GOTO(cleanup, rc = -EINVAL);
        }
-       sprintf(lwpname, "%s-MDT0000-%s-%s", fsname, LUSTRE_LWP_NAME, tgt);
+       snprintf(lwp_name, len, "%s-MDT%04x-%s-%s",
+                fsname, idx, LUSTRE_LWP_NAME, tgt);
+
+       GOTO(cleanup, rc = 0);
+
 cleanup:
        if (fsname != NULL)
                OBD_FREE(fsname, MTI_NAME_MAXLEN);
-       RETURN(rc);
+
+       return rc;
 }
-EXPORT_SYMBOL(tgt_name2lwpname);
+EXPORT_SYMBOL(tgt_name2lwp_name);
 
-static CFS_LIST_HEAD(lwp_register_list);
+static struct list_head lwp_register_list =
+       LIST_HEAD_INIT(lwp_register_list);
 DEFINE_MUTEX(lwp_register_list_lock);
 
 int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
@@ -446,8 +416,8 @@ int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
        lri->lri_exp = exp;
        lri->lri_cb_func = cb_func;
        lri->lri_cb_data = cb_data;
-       CFS_INIT_LIST_HEAD(&lri->lri_list);
-       cfs_list_add(&lri->lri_list, &lwp_register_list);
+       INIT_LIST_HEAD(&lri->lri_list);
+       list_add(&lri->lri_list, &lwp_register_list);
 
        if (*exp != NULL && cb_func != NULL)
                cb_func(cb_data);
@@ -462,11 +432,11 @@ void lustre_deregister_lwp_item(struct obd_export **exp)
        struct lwp_register_item *lri, *tmp;
 
        mutex_lock(&lwp_register_list_lock);
-       cfs_list_for_each_entry_safe(lri, tmp, &lwp_register_list, lri_list) {
+       list_for_each_entry_safe(lri, tmp, &lwp_register_list, lri_list) {
                if (exp == lri->lri_exp) {
                        if (*exp)
                                class_export_put(*exp);
-                       cfs_list_del(&lri->lri_list);
+                       list_del(&lri->lri_list);
                        OBD_FREE_PTR(lri);
                        break;
                }
@@ -475,13 +445,54 @@ void lustre_deregister_lwp_item(struct obd_export **exp)
 }
 EXPORT_SYMBOL(lustre_deregister_lwp_item);
 
+struct obd_export *lustre_find_lwp_by_index(const char *dev, __u32 idx)
+{
+       struct lustre_mount_info *lmi;
+       struct lustre_sb_info    *lsi;
+       struct obd_device        *lwp;
+       struct obd_export        *exp = NULL;
+       char                      fsname[16];
+       char                      lwp_name[24];
+       int                       rc;
+
+       lmi = server_get_mount(dev);
+       if (lmi == NULL)
+               return NULL;
+
+       lsi = s2lsi(lmi->lmi_sb);
+       rc = server_name2fsname(lsi->lsi_svname, fsname, NULL);
+       if (rc != 0) {
+               CERROR("%s: failed to get fsname: rc = %d\n",
+                      lsi->lsi_svname, rc);
+               goto err_lmi;
+       }
+
+       snprintf(lwp_name, sizeof(lwp_name), "%s-MDT%04x", fsname, idx);
+       spin_lock(&lsi->lsi_lwp_lock);
+       list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
+               char *ptr = strstr(lwp->obd_name, lwp_name);
+
+               if (ptr != NULL) {
+                       exp = class_export_get(lwp->obd_lwp_export);
+                       break;
+               }
+       }
+       spin_unlock(&lsi->lsi_lwp_lock);
+
+err_lmi:
+       server_put_mount(dev, false);
+
+       return exp;
+}
+EXPORT_SYMBOL(lustre_find_lwp_by_index);
+
 static void lustre_notify_lwp_list(struct obd_export *exp)
 {
        struct lwp_register_item *lri, *tmp;
        LASSERT(exp != NULL);
 
        mutex_lock(&lwp_register_list_lock);
-       cfs_list_for_each_entry_safe(lri, tmp, &lwp_register_list, lri_list) {
+       list_for_each_entry_safe(lri, tmp, &lwp_register_list, lri_list) {
                if (strcmp(exp->exp_obd->obd_name, lri->lri_name))
                        continue;
                if (*lri->lri_exp != NULL)
@@ -508,7 +519,7 @@ static int lustre_lwp_connect(struct obd_device *lwp)
        if (rc != 0)
                RETURN(rc);
 
-       lu_context_init(&session_ctx, LCT_SESSION);
+       lu_context_init(&session_ctx, LCT_SERVER_SESSION);
        session_ctx.lc_thread = NULL;
        lu_context_enter(&session_ctx);
        env.le_ses = &session_ctx;
@@ -522,7 +533,7 @@ static int lustre_lwp_connect(struct obd_device *lwp)
        data->ocd_connect_flags |= OBD_CONNECT_MDS_MDS | OBD_CONNECT_FID |
                OBD_CONNECT_AT | OBD_CONNECT_LRU_RESIZE |
                OBD_CONNECT_FULL20 | OBD_CONNECT_LVB_TYPE |
-               OBD_CONNECT_LIGHTWEIGHT;
+               OBD_CONNECT_LIGHTWEIGHT | OBD_CONNECT_LFSCK;
        OBD_ALLOC_PTR(uuid);
        if (uuid == NULL)
                GOTO(out, rc = -ENOMEM);
@@ -536,10 +547,16 @@ static int lustre_lwp_connect(struct obd_device *lwp)
        /* Use lwp name as the uuid, so we find the export by lwp name later */
        memcpy(uuid->uuid, lwp->obd_name, strlen(lwp->obd_name));
        rc = obd_connect(&env, &exp, lwp, uuid, data, NULL);
-       if (rc != 0)
+       if (rc != 0) {
                CERROR("%s: connect failed: rc = %d\n", lwp->obd_name, rc);
-       else
+       } else {
+               if (unlikely(lwp->obd_lwp_export != NULL))
+                       class_export_put(lwp->obd_lwp_export);
+               lwp->obd_lwp_export = class_export_get(exp);
                lustre_notify_lwp_list(exp);
+       }
+
+       GOTO(out, rc);
 
 out:
        if (data != NULL)
@@ -551,16 +568,16 @@ out:
        lu_context_exit(&session_ctx);
        lu_context_fini(&session_ctx);
 
-       RETURN(rc);
+       return rc;
 }
 
 /**
- * lwp is used by slaves (Non-MDT0 targets) to manage the connection
- * to MDT0.
+ * lwp is used by slaves (Non-MDT0 targets) to manage the connection to MDT0,
+ * or from the OSTx to MDTy.
  **/
-static int lustre_lwp_setup(struct lustre_cfg *lcfg, struct lustre_sb_info *lsi)
+static int lustre_lwp_setup(struct lustre_cfg *lcfg, struct lustre_sb_info *lsi,
+                           __u32 idx)
 {
-       struct obd_connect_data *data = NULL;
        struct obd_device       *obd;
        char                    *lwpname = NULL;
        char                    *lwpuuid = NULL;
@@ -569,18 +586,18 @@ static int lustre_lwp_setup(struct lustre_cfg *lcfg, struct lustre_sb_info *lsi)
 
        rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
                            lcfg->lcfg_nid);
-       if (rc) {
+       if (rc != 0) {
                CERROR("%s: Can't add uuid: rc =%d\n", lsi->lsi_svname, rc);
-               GOTO(out, rc);
+               RETURN(rc);
        }
 
        OBD_ALLOC(lwpname, MTI_NAME_MAXLEN);
        if (lwpname == NULL)
                GOTO(out, rc = -ENOMEM);
 
-       rc = tgt_name2lwpname(lsi->lsi_svname, lwpname);
+       rc = tgt_name2lwp_name(lsi->lsi_svname, lwpname, MTI_NAME_MAXLEN, idx);
        if (rc != 0) {
-               CERROR("%s: failed to generate lwp name. %d\n",
+               CERROR("%s: failed to generate lwp name: rc = %d\n",
                       lsi->lsi_svname, rc);
                GOTO(out, rc);
        }
@@ -602,22 +619,29 @@ static int lustre_lwp_setup(struct lustre_cfg *lcfg, struct lustre_sb_info *lsi)
        LASSERT(obd != NULL);
 
        rc = lustre_lwp_connect(obd);
-       if (rc != 0)
+       if (rc == 0) {
+               obd->u.cli.cl_max_mds_easize = MAX_MD_SIZE;
+               spin_lock(&lsi->lsi_lwp_lock);
+               list_add_tail(&obd->obd_lwp_list, &lsi->lsi_lwp_list);
+               spin_unlock(&lsi->lsi_lwp_lock);
+       } else {
                CERROR("%s: connect failed: rc = %d\n", lwpname, rc);
+       }
+
+       GOTO(out, rc);
+
 out:
-       if (data != NULL)
-               OBD_FREE_PTR(data);
        if (lwpname != NULL)
                OBD_FREE(lwpname, MTI_NAME_MAXLEN);
        if (lwpuuid != NULL)
                OBD_FREE(lwpuuid, MTI_NAME_MAXLEN);
 
-       RETURN(rc);
+       return rc;
 }
 
 /* the caller is responsible for memory free */
 static struct obd_device *lustre_find_lwp(struct lustre_sb_info *lsi,
-                                         char **lwpname, char **logname)
+                                         char **lwpname, __u32 idx)
 {
        struct obd_device       *lwp;
        int                      rc = 0;
@@ -630,22 +654,9 @@ static struct obd_device *lustre_find_lwp(struct lustre_sb_info *lsi,
        if (*lwpname == NULL)
                RETURN(ERR_PTR(-ENOMEM));
 
-       if (logname != NULL) {
-               OBD_ALLOC(*logname, MTI_NAME_MAXLEN);
-               if (*logname == NULL)
-                       GOTO(out, rc = -ENOMEM);
-               rc = server_name2fsname(lsi->lsi_svname, *lwpname, NULL);
-               if (rc != 0) {
-                       CERROR("%s: failed to get fsname from svname. %d\n",
-                              lsi->lsi_svname, rc);
-                       GOTO(out, rc = -EINVAL);
-               }
-               sprintf(*logname, "%s-client", *lwpname);
-       }
-
-       rc = tgt_name2lwpname(lsi->lsi_svname, *lwpname);
+       rc = tgt_name2lwp_name(lsi->lsi_svname, *lwpname, MTI_NAME_MAXLEN, idx);
        if (rc != 0) {
-               CERROR("%s: failed to generate lwp name. %d\n",
+               CERROR("%s: failed to generate lwp name: rc = %d\n",
                       lsi->lsi_svname, rc);
                GOTO(out, rc = -EINVAL);
        }
@@ -658,10 +669,6 @@ out:
                        OBD_FREE(*lwpname, MTI_NAME_MAXLEN);
                        *lwpname = NULL;
                }
-               if (logname != NULL && *logname != NULL) {
-                       OBD_FREE(*logname, MTI_NAME_MAXLEN);
-                       *logname = NULL;
-               }
                lwp = ERR_PTR(rc);
        }
 
@@ -669,7 +676,7 @@ out:
 }
 
 static int lustre_lwp_add_conn(struct lustre_cfg *cfg,
-                              struct lustre_sb_info *lsi)
+                              struct lustre_sb_info *lsi, __u32 idx)
 {
        struct lustre_cfg_bufs *bufs = NULL;
        struct lustre_cfg      *lcfg = NULL;
@@ -678,7 +685,7 @@ static int lustre_lwp_add_conn(struct lustre_cfg *cfg,
        int                     rc;
        ENTRY;
 
-       lwp = lustre_find_lwp(lsi, &lwpname, NULL);
+       lwp = lustre_find_lwp(lsi, &lwpname, idx);
        if (IS_ERR(lwp)) {
                CERROR("%s: can't find lwp device.\n", lsi->lsi_svname);
                GOTO(out, rc = PTR_ERR(lwp));
@@ -694,16 +701,18 @@ static int lustre_lwp_add_conn(struct lustre_cfg *cfg,
                                   lustre_cfg_string(cfg, 1));
 
        lcfg = lustre_cfg_new(LCFG_ADD_CONN, bufs);
-
+       if (lcfg == NULL)
+               GOTO(out_cfg, rc = -ENOMEM);
        rc = class_add_conn(lwp, lcfg);
        if (rc)
                CERROR("%s: can't add conn: rc = %d\n", lwpname, rc);
 
-out:
-       if (bufs != NULL)
-               OBD_FREE_PTR(bufs);
        if (lcfg != NULL)
                lustre_cfg_free(lcfg);
+out_cfg:
+       if (bufs != NULL)
+               OBD_FREE_PTR(bufs);
+out:
        if (lwpname != NULL)
                OBD_FREE(lwpname, MTI_NAME_MAXLEN);
        RETURN(rc);
@@ -713,19 +722,19 @@ out:
  * Retrieve MDT nids from the client log, then start the lwp device.
  * there are only two scenarios which would include mdt nid.
  * 1.
- * marker   5 (flags=0x01, v2.1.54.0) lustre-MDT0000  'add mdc' xxx-
+ * marker   5 (flags=0x01, v2.1.54.0) lustre-MDTyyyy  'add mdc' xxx-
  * add_uuid  nid=192.168.122.162@tcp(0x20000c0a87aa2)  0:  1:192.168.122.162@tcp
- * attach    0:lustre-MDT0000-mdc  1:mdc  2:lustre-clilmv_UUID
- * setup     0:lustre-MDT0000-mdc  1:lustre-MDT0000_UUID  2:192.168.122.162@tcp
+ * attach    0:lustre-MDTyyyy-mdc  1:mdc  2:lustre-clilmv_UUID
+ * setup     0:lustre-MDTyyyy-mdc  1:lustre-MDTyyyy_UUID  2:192.168.122.162@tcp
  * add_uuid  nid=192.168.172.1@tcp(0x20000c0a8ac01)  0:  1:192.168.172.1@tcp
- * add_conn  0:lustre-MDT0000-mdc  1:192.168.172.1@tcp
- * modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID xxxx
- * marker   5 (flags=0x02, v2.1.54.0) lustre-MDT0000  'add mdc' xxxx-
+ * add_conn  0:lustre-MDTyyyy-mdc  1:192.168.172.1@tcp
+ * modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDTyyyy_UUID xxxx
+ * marker   5 (flags=0x02, v2.1.54.0) lustre-MDTyyyy  'add mdc' xxxx-
  * 2.
- * marker   7 (flags=0x01, v2.1.54.0) lustre-MDT0000  'add failnid' xxxx-
+ * marker   7 (flags=0x01, v2.1.54.0) lustre-MDTyyyy  'add failnid' xxxx-
  * add_uuid  nid=192.168.122.2@tcp(0x20000c0a87a02)  0:  1:192.168.122.2@tcp
- * add_conn  0:lustre-MDT0000-mdc  1:192.168.122.2@tcp
- * marker   7 (flags=0x02, v2.1.54.0) lustre-MDT0000  'add failnid' xxxx-
+ * add_conn  0:lustre-MDTyyyy-mdc  1:192.168.122.2@tcp
+ * marker   7 (flags=0x02, v2.1.54.0) lustre-MDTyyyy  'add failnid' xxxx-
  **/
 static int client_lwp_config_process(const struct lu_env *env,
                                     struct llog_handle *handle,
@@ -768,7 +777,10 @@ static int client_lwp_config_process(const struct lu_env *env,
                    marker->cm_flags & CM_EXCLUDE)
                        GOTO(out, rc = 0);
 
-               if (!tgt_is_mdt0(marker->cm_tgtname))
+               if (!tgt_is_mdt(marker->cm_tgtname, &clli->cfg_lwp_idx))
+                       GOTO(out, rc = 0);
+
+               if (IS_MDT(lsi) && clli->cfg_lwp_idx != 0)
                        GOTO(out, rc = 0);
 
                if (!strncmp(marker->cm_comment, "add mdc", 7) ||
@@ -789,7 +801,7 @@ static int client_lwp_config_process(const struct lu_env *env,
        }
        case LCFG_ADD_UUID: {
                if (clli->cfg_flags == CFG_F_MARKER) {
-                       rc = lustre_lwp_setup(lcfg, lsi);
+                       rc = lustre_lwp_setup(lcfg, lsi, clli->cfg_lwp_idx);
                        /* XXX: process only the first nid as
                         * we don't need another instance of lwp */
                        clli->cfg_flags |= CFG_F_SKIP;
@@ -803,8 +815,28 @@ static int client_lwp_config_process(const struct lu_env *env,
                break;
        }
        case LCFG_ADD_CONN: {
-               if (is_mdc_for_mdt0(lustre_cfg_string(lcfg, 0)))
-                       rc = lustre_lwp_add_conn(lcfg, lsi);
+               char *devname = lustre_cfg_string(lcfg, 0);
+               char *ptr;
+               __u32 idx     = 0;
+
+               if (!is_mdc_device(devname))
+                       break;
+
+               ptr = strrchr(devname, '-');
+               if (ptr == NULL)
+                       break;
+
+               *ptr = 0;
+               if (!tgt_is_mdt(devname, &idx)) {
+                       *ptr = '-';
+                       break;
+               }
+               *ptr = '-';
+
+               if (IS_MDT(lsi) && idx != 0)
+                       break;
+
+               rc = lustre_lwp_add_conn(lcfg, lsi, idx);
                break;
        }
        default:
@@ -816,67 +848,82 @@ out:
 
 static int lustre_disconnect_lwp(struct super_block *sb)
 {
-       struct lustre_sb_info           *lsi = s2lsi(sb);
+       struct lustre_sb_info           *lsi     = s2lsi(sb);
        struct obd_device               *lwp;
-       char                            *lwpname = NULL;
        char                            *logname = NULL;
-       struct lustre_cfg               *lcfg = NULL;
-       struct lustre_cfg_bufs          *bufs = NULL;
-       struct config_llog_instance     *cfg = NULL;
-       int                              rc;
+       struct lustre_cfg_bufs          *bufs    = NULL;
+       struct config_llog_instance     *cfg     = NULL;
+       int                              rc      = 0;
+       int                              rc1     = 0;
        ENTRY;
 
-       lwp = lustre_find_lwp(lsi, &lwpname, &logname);
-       if (IS_ERR(lwp) && PTR_ERR(lwp) != -ENOENT)
-               GOTO(out, rc = PTR_ERR(lwp));
+       if (likely(lsi->lsi_lwp_started)) {
+               OBD_ALLOC(logname, MTI_NAME_MAXLEN);
+               if (logname == NULL)
+                       RETURN(-ENOMEM);
 
-       LASSERT(lwpname != NULL);
-       LASSERT(logname != NULL);
+               rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
+               if (rc != 0) {
+                       CERROR("%s: failed to get fsname from svname: "
+                              "rc = %d\n", lsi->lsi_svname, rc);
+                       GOTO(out, rc = -EINVAL);
+               }
 
-       OBD_ALLOC_PTR(cfg);
-       if (cfg == NULL)
-               GOTO(out, rc = -ENOMEM);
+               strcat(logname, "-client");
+               OBD_ALLOC_PTR(cfg);
+               if (cfg == NULL)
+                       GOTO(out, rc = -ENOMEM);
 
-       /* end log first */
-       cfg->cfg_instance = sb;
-       rc = lustre_end_log(sb, logname, cfg);
-       if (rc != 0) {
-               CERROR("%s: Can't end config log %s.\n", lwpname, logname);
-               GOTO(out, rc);
-       }
+               /* end log first */
+               cfg->cfg_instance = sb;
+               rc = lustre_end_log(sb, logname, cfg);
+               if (rc != 0)
+                       GOTO(out, rc);
 
-       if (PTR_ERR(lwp) == -ENOENT) {
-               CDEBUG(D_CONFIG, "%s: lwp device wasn't started.\n",
-                      lsi->lsi_svname);
-               GOTO(out, rc = 0);
+               lsi->lsi_lwp_started = 0;
        }
 
        OBD_ALLOC_PTR(bufs);
        if (bufs == NULL)
                GOTO(out, rc = -ENOMEM);
 
-       lustre_cfg_bufs_reset(bufs, lwp->obd_name);
-       lustre_cfg_bufs_set_string(bufs, 1, NULL);
-       lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
-       if (!lcfg)
-               GOTO(out, rc = -ENOMEM);
+       list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
+               struct lustre_cfg *lcfg;
 
-       /* Disconnect import first. NULL is passed for the '@env', since
-        * it will not be used. */
-       rc = lwp->obd_lu_dev->ld_ops->ldo_process_config(NULL, lwp->obd_lu_dev,
-                                                        lcfg);
-out:
-       if (lcfg)
+               if (likely(lwp->obd_lwp_export != NULL)) {
+                       class_export_put(lwp->obd_lwp_export);
+                       lwp->obd_lwp_export = NULL;
+               }
+
+               lustre_cfg_bufs_reset(bufs, lwp->obd_name);
+               lustre_cfg_bufs_set_string(bufs, 1, NULL);
+               lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
+               if (lcfg == NULL)
+                       GOTO(out, rc = -ENOMEM);
+
+               /* Disconnect import first. NULL is passed for the '@env',
+                * since it will not be used. */
+               rc = lwp->obd_lu_dev->ld_ops->ldo_process_config(NULL,
+                                                       lwp->obd_lu_dev, lcfg);
                lustre_cfg_free(lcfg);
-       if (bufs)
+               if (rc != 0 && rc != -ETIMEDOUT) {
+                       CERROR("%s: fail to disconnect LWP: rc = %d\n",
+                              lwp->obd_name, rc);
+                       rc1 = rc;
+               }
+       }
+
+       GOTO(out, rc);
+
+out:
+       if (bufs != NULL)
                OBD_FREE_PTR(bufs);
-       if (cfg)
+       if (cfg != NULL)
                OBD_FREE_PTR(cfg);
-       if (lwpname)
-               OBD_FREE(lwpname, MTI_NAME_MAXLEN);
-       if (logname)
+       if (logname != NULL)
                OBD_FREE(logname, MTI_NAME_MAXLEN);
-       RETURN(rc);
+
+       return rc1 != 0 ? rc1 : rc;
 }
 
 /**
@@ -885,69 +932,71 @@ out:
 static int lustre_stop_lwp(struct super_block *sb)
 {
        struct lustre_sb_info   *lsi = s2lsi(sb);
-       struct obd_device       *lwp = NULL;
-       char                    *lwpname = NULL;
-       int                      rc = 0;
+       struct obd_device       *lwp;
+       int                      rc  = 0;
+       int                      rc1 = 0;
        ENTRY;
 
-       lwp = lustre_find_lwp(lsi, &lwpname, NULL);
-       if (IS_ERR(lwp)) {
-               CDEBUG(PTR_ERR(lwp) == -ENOENT ? D_CONFIG : D_ERROR,
-                      "%s: lwp wasn't started.\n", lsi->lsi_svname);
-               GOTO(out, rc = 0);
+       while (!list_empty(&lsi->lsi_lwp_list)) {
+               lwp = list_entry(lsi->lsi_lwp_list.next, struct obd_device,
+                                obd_lwp_list);
+               list_del_init(&lwp->obd_lwp_list);
+               lwp->obd_force = 1;
+               rc = class_manual_cleanup(lwp);
+               if (rc != 0) {
+                       CERROR("%s: fail to stop LWP: rc = %d\n",
+                              lwp->obd_name, rc);
+                       rc1 = rc;
+               }
        }
 
-       lwp->obd_force = 1;
-       rc = class_manual_cleanup(lwp);
-
-out:
-       if (lwpname != NULL)
-               OBD_FREE(lwpname, MTI_NAME_MAXLEN);
-       RETURN(rc);
+       RETURN(rc1 != 0 ? rc1 : rc);
 }
 
 /**
- * Start the lwp(fsname-MDT0000-lwp-OSTxxxx) for an OST or MDT target,
- * which would be used to establish connection from OST to MDT0.
+ * Start the lwp(fsname-MDTyyyy-lwp-{MDT,OST}xxxx) for a MDT/OST or MDT target.
  **/
 static int lustre_start_lwp(struct super_block *sb)
 {
        struct lustre_sb_info       *lsi = s2lsi(sb);
        struct config_llog_instance *cfg = NULL;
-       struct obd_device           *lwp;
-       char                        *lwpname = NULL;
-       char                        *logname = NULL;
+       char                        *logname;
        int                          rc;
        ENTRY;
 
-       lwp = lustre_find_lwp(lsi, &lwpname, &logname);
+       if (unlikely(lsi->lsi_lwp_started))
+               RETURN(0);
 
-       /* the lwp device already stared */
-       if (lwp && !IS_ERR(lwp))
-               GOTO(out, rc = 0);
-
-       if (PTR_ERR(lwp) != -ENOENT)
-               GOTO(out, rc = PTR_ERR(lwp));
+       OBD_ALLOC(logname, MTI_NAME_MAXLEN);
+       if (logname == NULL)
+               RETURN(-ENOMEM);
 
-       LASSERT(lwpname != NULL);
-       LASSERT(logname != NULL);
+       rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
+       if (rc != 0) {
+               CERROR("%s: failed to get fsname from svname: rc = %d\n",
+                      lsi->lsi_svname, rc);
+               GOTO(out, rc = -EINVAL);
+       }
 
+       strcat(logname, "-client");
        OBD_ALLOC_PTR(cfg);
        if (cfg == NULL)
                GOTO(out, rc = -ENOMEM);
 
        cfg->cfg_callback = client_lwp_config_process;
        cfg->cfg_instance = sb;
-
        rc = lustre_process_log(sb, logname, cfg);
+       if (rc == 0)
+               lsi->lsi_lwp_started = 1;
+
+       GOTO(out, rc);
+
 out:
-       if (lwpname != NULL)
-               OBD_FREE(lwpname, MTI_NAME_MAXLEN);
-       if (logname != NULL)
-               OBD_FREE(logname, MTI_NAME_MAXLEN);
+       OBD_FREE(logname, MTI_NAME_MAXLEN);
        if (cfg != NULL)
                OBD_FREE_PTR(cfg);
-       RETURN(rc);
+
+       return rc;
 }
 
 DEFINE_MUTEX(server_start_lock);
@@ -978,13 +1027,9 @@ static int server_stop_servers(int lsiflags)
        }
 
        if (obd != NULL && (type == NULL || type->typ_refcnt == 0)) {
-               int err;
-
                obd->obd_force = 1;
                /* obd_fail doesn't mean much on a server obd */
-               err = class_manual_cleanup(obd);
-               if (rc != 0)
-                       rc = err;
+               rc = class_manual_cleanup(obd);
        }
 
        mutex_unlock(&server_start_lock);
@@ -1180,7 +1225,7 @@ out:
 
 /** Start server targets: MDTs and OSTs
  */
-static int server_start_targets(struct super_block *sb, struct vfsmount *mnt)
+static int server_start_targets(struct super_block *sb)
 {
        struct obd_device *obd;
        struct lustre_sb_info *lsi = s2lsi(sb);
@@ -1231,11 +1276,9 @@ static int server_start_targets(struct super_block *sb, struct vfsmount *mnt)
 
        /* Set the mgc fs to our server disk.  This allows the MGC to
         * read and write configs locally, in case it can't talk to the MGS. */
-       if (lsi->lsi_srv_mnt) {
-               rc = server_mgc_set_fs(lsi->lsi_mgc, sb);
-               if (rc)
-                       GOTO(out_stop_service, rc);
-       }
+       rc = server_mgc_set_fs(lsi->lsi_mgc, sb);
+       if (rc)
+               GOTO(out_stop_service, rc);
 
        /* Register with MGS */
        rc = server_register_target(lsi);
@@ -1244,7 +1287,7 @@ static int server_start_targets(struct super_block *sb, struct vfsmount *mnt)
 
        /* Let the target look up the mount using the target's name
           (we can't pass the sb or mnt through class_process_config.) */
-       rc = server_register_mount(lsi->lsi_svname, sb, mnt);
+       rc = server_register_mount(lsi->lsi_svname, sb);
        if (rc)
                GOTO(out_mgc, rc);
 
@@ -1291,7 +1334,7 @@ static int server_start_targets(struct super_block *sb, struct vfsmount *mnt)
                if (rc == 0) {
                        struct lu_context  session_ctx;
 
-                       lu_context_init(&session_ctx, LCT_SESSION);
+                       lu_context_init(&session_ctx, LCT_SERVER_SESSION);
                        session_ctx.lc_thread = NULL;
                        lu_context_enter(&session_ctx);
                        env.le_ses = &session_ctx;
@@ -1314,8 +1357,7 @@ static int server_start_targets(struct super_block *sb, struct vfsmount *mnt)
 
 out_mgc:
        /* Release the mgc fs for others to use */
-       if (lsi->lsi_srv_mnt)
-               server_mgc_clear_fs(lsi->lsi_mgc);
+       server_mgc_clear_fs(lsi->lsi_mgc);
 
 out_stop_service:
        if (rc != 0)
@@ -1326,6 +1368,8 @@ out_stop_service:
 
 static int lsi_prepare(struct lustre_sb_info *lsi)
 {
+       const char *osd_type;
+       const char *fstype;
        __u32 index;
        int rc;
        ENTRY;
@@ -1339,28 +1383,26 @@ static int lsi_prepare(struct lustre_sb_info *lsi)
                RETURN(-EINVAL);
        }
 
-       if (strlen(lsi->lsi_lmd->lmd_profile) >= sizeof(lsi->lsi_svname))
-               RETURN(-ENAMETOOLONG);
-
-       strcpy(lsi->lsi_svname, lsi->lsi_lmd->lmd_profile);
-
        /* Determine osd type */
-       if (lsi->lsi_lmd->lmd_osd_type != NULL) {
-               if (strlen(lsi->lsi_lmd->lmd_osd_type) >=
-                   sizeof(lsi->lsi_osd_type))
-                       RETURN(-ENAMETOOLONG);
-
-               strcpy(lsi->lsi_osd_type, lsi->lsi_lmd->lmd_osd_type);
+       if (lsi->lsi_lmd->lmd_osd_type == NULL) {
+               osd_type = LUSTRE_OSD_LDISKFS_NAME;
+               fstype = "ldiskfs";
        } else {
-               strcpy(lsi->lsi_osd_type, LUSTRE_OSD_LDISKFS_NAME);
+               osd_type = lsi->lsi_lmd->lmd_osd_type;
+               fstype = lsi->lsi_lmd->lmd_osd_type;
        }
 
-       /* XXX: a temp. solution for components using fsfilt
+       if (strlen(lsi->lsi_lmd->lmd_profile) >= sizeof(lsi->lsi_svname) ||
+           strlen(osd_type) >= sizeof(lsi->lsi_osd_type) ||
+           strlen(fstype) >= sizeof(lsi->lsi_fstype))
+               RETURN(-ENAMETOOLONG);
+
+       strlcpy(lsi->lsi_svname, lsi->lsi_lmd->lmd_profile,
+               sizeof(lsi->lsi_svname));
+       strlcpy(lsi->lsi_osd_type, osd_type, sizeof(lsi->lsi_osd_type));
+       /* XXX: a temp. solution for components using ldiskfs
         *      to be removed in one of the subsequent patches */
-       if (!strcmp(lsi->lsi_lmd->lmd_osd_type, "osd-ldiskfs"))
-               strcpy(lsi->lsi_fstype, "ldiskfs");
-       else
-               strcpy(lsi->lsi_fstype, lsi->lsi_lmd->lmd_osd_type);
+       strlcpy(lsi->lsi_fstype, fstype, sizeof(lsi->lsi_fstype));
 
        /* Determine server type */
        rc = server_name2index(lsi->lsi_svname, &index, NULL);
@@ -1421,9 +1463,10 @@ static void server_put_super(struct super_block *sb)
                int     rc;
 
                rc = lustre_disconnect_lwp(sb);
-               if (rc && rc != ETIMEDOUT)
-                       CERROR("%s: failed to disconnect lwp. (rc=%d)\n",
-                              tmpname, rc);
+               if (rc != 0 && rc != -ETIMEDOUT &&
+                   rc != -ENOTCONN && rc != -ESHUTDOWN)
+                       CWARN("%s: failed to disconnect lwp: rc= %d\n",
+                             tmpname, rc);
        }
 
        /* Stop the target */
@@ -1553,6 +1596,35 @@ static struct super_operations server_ops = {
        .statfs         = server_statfs,
 };
 
+/*
+ * Xattr support for Lustre servers
+ */
+static ssize_t lustre_getxattr(struct dentry *dentry, const char *name,
+                               void *buffer, size_t size)
+{
+       if (!selinux_is_enabled())
+               return -EOPNOTSUPP;
+       return -ENODATA;
+}
+
+static int lustre_setxattr(struct dentry *dentry, const char *name,
+                           const void *value, size_t size, int flags)
+{
+       return -EOPNOTSUPP;
+}
+
+static ssize_t lustre_listxattr(struct dentry *d_entry, char *name,
+                               size_t size)
+{
+       return -EOPNOTSUPP;
+}
+
+const struct inode_operations server_inode_operations = {
+       .setxattr       = lustre_setxattr,
+       .getxattr       = lustre_getxattr,
+       .listxattr      = lustre_listxattr,
+};
+
 #define log2(n) ffz(~(n))
 #define LUSTRE_SUPER_MAGIC 0x0BD00BD1
 
@@ -1580,7 +1652,7 @@ static int server_fill_super_common(struct super_block *sb)
        /* make_bad_inode(root); -- badness - can't umount */
        /* apparently we need to be a directory for the mount to finish */
        root->i_mode = S_IFDIR;
-
+       root->i_op = &server_inode_operations;
        sb->s_root = d_make_root(root);
        if (!sb->s_root) {
                CERROR("%s: can't make root dentry\n", sb->s_id);
@@ -1619,26 +1691,39 @@ static int osd_start(struct lustre_sb_info *lsi, unsigned long mflags)
                        GOTO(out, rc);
                obd = class_name2obd(lsi->lsi_osd_obdname);
                LASSERT(obd);
+       } else {
+               CDEBUG(D_MOUNT, "%s already started\n", lsi->lsi_osd_obdname);
+               /* but continue setup to allow special case of MDT and internal
+                * MGT being started separately. */
+               if (!((IS_MGS(lsi) && (lsi->lsi_lmd->lmd_flags &
+                                     LMD_FLG_NOMGS)) ||
+                    (IS_MDT(lsi) && (lsi->lsi_lmd->lmd_flags &
+                                     LMD_FLG_NOSVC))))
+                       RETURN(-EALREADY);
        }
 
        rc = obd_connect(NULL, &lsi->lsi_osd_exp,
                         obd, &obd->obd_uuid, NULL, NULL);
+
+       OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 10);
+
        if (rc) {
                obd->obd_force = 1;
                class_manual_cleanup(obd);
                lsi->lsi_dt_dev = NULL;
+               RETURN(rc);
        }
 
-       /* XXX: to keep support old components relying on lsi_srv_mnt
-        *      we get this info from OSD just started */
        LASSERT(obd->obd_lu_dev);
+       lu_device_get(obd->obd_lu_dev);
        lsi->lsi_dt_dev = lu2dt_dev(obd->obd_lu_dev);
        LASSERT(lsi->lsi_dt_dev);
 
-       dt_conf_get(NULL, lsi->lsi_dt_dev, &p);
-
-       lsi->lsi_srv_mnt = p.ddp_mnt;
+       /* set disk context for llog usage */
+       OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
+       obd->obd_lvfs_ctxt.dt = lsi->lsi_dt_dev;
 
+       dt_conf_get(NULL, lsi->lsi_dt_dev, &p);
 out:
        RETURN(rc);
 }
@@ -1694,7 +1779,7 @@ int server_fill_super(struct super_block *sb)
        /* Set up all obd devices for service */
        if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
            (IS_OST(lsi) || IS_MDT(lsi))) {
-               rc = server_start_targets(sb, lsi->lsi_srv_mnt);
+               rc = server_start_targets(sb);
                if (rc < 0) {
                        CERROR("Unable to start targets: %d\n", rc);
                        GOTO(out_mnt, rc);