X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fobdclass%2Fobd_mount_server.c;h=d3854db973f11ff36f9f8d4dfc1c39a7b824900b;hp=ef453b07b6115b66e3378c6373e0c6a9a82a19b3;hb=5c883ea2748ae9e430a9cd863a9b630b2a74440a;hpb=eba2a78067ef87a56242e47e626c584624fb1e34

diff --git a/lustre/obdclass/obd_mount_server.c b/lustre/obdclass/obd_mount_server.c
index ef453b0..d3854db 100644
--- a/lustre/obdclass/obd_mount_server.c
+++ b/lustre/obdclass/obd_mount_server.c
@@ -15,11 +15,7 @@
  *
  * You should have received a copy of the GNU General Public License
  * version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * http://www.gnu.org/licenses/gpl-2.0.html
  *
  * GPL HEADER END
  */
@@ -27,7 +23,7 @@
  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2013, 2014, Intel Corporation.
+ * Copyright (c) 2013, 2017, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -46,27 +42,26 @@
 #define PRINT_CMD CDEBUG
 #define PRINT_MASK (D_SUPER | D_CONFIG)
 
+#include <linux/types.h>
+#ifdef HAVE_LINUX_SELINUX_IS_ENABLED
+#include <linux/selinux.h>
+#endif
 #include <linux/statfs.h>
-#include <obd.h>
-#include <obd_class.h>
-#include <lustre/lustre_user.h>
 #include <linux/version.h>
-#include <lustre_ioctl.h>
-#include <lustre_log.h>
+#include <linux/delay.h>
+
+#include <llog_swab.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
+#include <uapi/linux/lustre/lustre_ioctl.h>
+#include <lustre_log.h>
+#include <uapi/linux/lustre/lustre_param.h>
+#include <obd.h>
+#include <obd_class.h>
 
 /*********** mount lookup *********/
 
 static DEFINE_MUTEX(lustre_mount_info_lock);
-static struct list_head server_mount_info_list =
-	LIST_HEAD_INIT(server_mount_info_list);
+static LIST_HEAD(server_mount_info_list);
 
 static struct lustre_mount_info *server_find_mount(const char *name)
 {
@@ -289,7 +284,8 @@ static int server_stop_mgs(struct super_block *sb)
 
 /* Since there's only one mgc per node, we have to change it's fs to get
    access to the right disk. */
-static int server_mgc_set_fs(struct obd_device *mgc, struct super_block *sb)
+static int server_mgc_set_fs(const struct lu_env *env,
+			     struct obd_device *mgc, struct super_block *sb)
 {
 	struct lustre_sb_info *lsi = s2lsi(sb);
 	int rc;
@@ -298,7 +294,7 @@ static int server_mgc_set_fs(struct obd_device *mgc, struct super_block *sb)
 	CDEBUG(D_MOUNT, "Set mgc disk for %s\n", lsi->lsi_lmd->lmd_dev);
 
 	/* cl_mgc_sem in mgc insures we sleep if the mgc_fs is busy */
-	rc = obd_set_info_async(NULL, mgc->obd_self_export,
+	rc = obd_set_info_async(env, mgc->obd_self_export,
 				sizeof(KEY_SET_FS), KEY_SET_FS,
 				sizeof(*sb), sb, NULL);
 	if (rc != 0)
@@ -307,14 +303,15 @@ static int server_mgc_set_fs(struct obd_device *mgc, struct super_block *sb)
 	RETURN(rc);
 }
 
-static int server_mgc_clear_fs(struct obd_device *mgc)
+static int server_mgc_clear_fs(const struct lu_env *env,
+			       struct obd_device *mgc)
 {
 	int rc;
 	ENTRY;
 
 	CDEBUG(D_MOUNT, "Unassign mgc disk\n");
 
-	rc = obd_set_info_async(NULL, mgc->obd_self_export,
+	rc = obd_set_info_async(env, mgc->obd_self_export,
 				sizeof(KEY_CLEAR_FS), KEY_CLEAR_FS,
 				0, NULL, NULL);
 	RETURN(rc);
@@ -382,15 +379,26 @@ cleanup:
 }
 EXPORT_SYMBOL(tgt_name2lwp_name);
 
-static struct list_head lwp_register_list =
-	LIST_HEAD_INIT(lwp_register_list);
-static DEFINE_MUTEX(lwp_register_list_lock);
+static LIST_HEAD(lwp_register_list);
+static DEFINE_SPINLOCK(lwp_register_list_lock);
+
+static void lustre_put_lwp_item(struct lwp_register_item *lri)
+{
+	if (atomic_dec_and_test(&lri->lri_ref)) {
+		LASSERT(list_empty(&lri->lri_list));
+
+		if (*lri->lri_exp != NULL)
+			class_export_put(*lri->lri_exp);
+		OBD_FREE_PTR(lri);
+	}
+}
 
 int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
 			     register_lwp_cb cb_func, void *cb_data)
 {
 	struct obd_device	 *lwp;
 	struct lwp_register_item *lri;
+	bool cb = false;
 	ENTRY;
 
 	LASSERTF(strlen(lwpname) < MTI_NAME_MAXLEN, "lwpname is too long %s\n",
@@ -401,20 +409,17 @@ int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
 	if (lri == NULL)
 		RETURN(-ENOMEM);
 
-	mutex_lock(&lwp_register_list_lock);
-
 	lwp = class_name2obd(lwpname);
 	if (lwp != NULL && lwp->obd_set_up == 1) {
 		struct obd_uuid *uuid;
 
 		OBD_ALLOC_PTR(uuid);
 		if (uuid == NULL) {
-			mutex_unlock(&lwp_register_list_lock);
 			OBD_FREE_PTR(lri);
 			RETURN(-ENOMEM);
 		}
 		memcpy(uuid->uuid, lwpname, strlen(lwpname));
-		*exp = cfs_hash_lookup(lwp->obd_uuid_hash, uuid);
+		*exp = obd_uuid_lookup(lwp, uuid);
 		OBD_FREE_PTR(uuid);
 	}
 
@@ -423,31 +428,57 @@ int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
 	lri->lri_cb_func = cb_func;
 	lri->lri_cb_data = cb_data;
 	INIT_LIST_HEAD(&lri->lri_list);
+	/*
+	 * Initialize the lri_ref at 2, one will be released before
+	 * current function returned via lustre_put_lwp_item(), the
+	 * other will be released in lustre_deregister_lwp_item().
+	 */
+	atomic_set(&lri->lri_ref, 2);
+
+	spin_lock(&lwp_register_list_lock);
 	list_add(&lri->lri_list, &lwp_register_list);
+	if (*exp != NULL)
+		cb = true;
+	spin_unlock(&lwp_register_list_lock);
 
-	if (*exp != NULL && cb_func != NULL)
+	if (cb && cb_func != NULL)
 		cb_func(cb_data);
+	lustre_put_lwp_item(lri);
 
-	mutex_unlock(&lwp_register_list_lock);
 	RETURN(0);
 }
 EXPORT_SYMBOL(lustre_register_lwp_item);
 
 void lustre_deregister_lwp_item(struct obd_export **exp)
 {
-	struct lwp_register_item *lri, *tmp;
+	struct lwp_register_item *lri;
+	bool removed = false;
+	int repeat = 0;
 
-	mutex_lock(&lwp_register_list_lock);
-	list_for_each_entry_safe(lri, tmp, &lwp_register_list, lri_list) {
+	spin_lock(&lwp_register_list_lock);
+	list_for_each_entry(lri, &lwp_register_list, lri_list) {
 		if (exp == lri->lri_exp) {
-			if (*exp)
-				class_export_put(*exp);
-			list_del(&lri->lri_list);
-			OBD_FREE_PTR(lri);
+			list_del_init(&lri->lri_list);
+			removed = true;
 			break;
 		}
 	}
-	mutex_unlock(&lwp_register_list_lock);
+	spin_unlock(&lwp_register_list_lock);
+
+	if (!removed)
+		return;
+
+	/* See lustre_notify_lwp_list(), in some extreme race conditions,
+	 * the notify callback could be still on the fly, we need to wait
+	 * for the callback done before moving on to free the data used
+	 * by callback. */
+	while (atomic_read(&lri->lri_ref) > 1) {
+		CDEBUG(D_MOUNT, "lri reference count %u, repeat: %d\n",
+		       atomic_read(&lri->lri_ref), repeat);
+		repeat++;
+		schedule_timeout_interruptible(cfs_time_seconds(1));
+	}
+	lustre_put_lwp_item(lri);
 }
 EXPORT_SYMBOL(lustre_deregister_lwp_item);
 
@@ -474,16 +505,16 @@ struct obd_export *lustre_find_lwp_by_index(const char *dev, __u32 idx)
 	}
 
 	snprintf(lwp_name, sizeof(lwp_name), "%s-MDT%04x", fsname, idx);
-	spin_lock(&lsi->lsi_lwp_lock);
+	mutex_lock(&lsi->lsi_lwp_mutex);
 	list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
 		char *ptr = strstr(lwp->obd_name, lwp_name);
 
-		if (ptr != NULL) {
+		if (ptr != NULL && lwp->obd_lwp_export != NULL) {
 			exp = class_export_get(lwp->obd_lwp_export);
 			break;
 		}
 	}
-	spin_unlock(&lsi->lsi_lwp_lock);
+	mutex_unlock(&lsi->lsi_lwp_mutex);
 
 err_lmi:
 	server_put_mount(dev, false);
@@ -492,25 +523,39 @@ err_lmi:
 }
 EXPORT_SYMBOL(lustre_find_lwp_by_index);
 
-static void lustre_notify_lwp_list(struct obd_export *exp)
+void lustre_notify_lwp_list(struct obd_export *exp)
 {
-	struct lwp_register_item *lri, *tmp;
+	struct lwp_register_item *lri;
 	LASSERT(exp != NULL);
 
-	mutex_lock(&lwp_register_list_lock);
-	list_for_each_entry_safe(lri, tmp, &lwp_register_list, lri_list) {
+again:
+	spin_lock(&lwp_register_list_lock);
+	list_for_each_entry(lri, &lwp_register_list, lri_list) {
 		if (strcmp(exp->exp_obd->obd_name, lri->lri_name))
 			continue;
 		if (*lri->lri_exp != NULL)
 			continue;
 		*lri->lri_exp = class_export_get(exp);
-		if (lri->lri_cb_func != NULL)
-			lri->lri_cb_func(lri->lri_cb_data);
+		if (lri->lri_cb_func == NULL)
+			continue;
+		atomic_inc(&lri->lri_ref);
+		spin_unlock(&lwp_register_list_lock);
+
+		lri->lri_cb_func(lri->lri_cb_data);
+		lustre_put_lwp_item(lri);
+
+		/* Others may have changed the list after we unlock, we have
+		 * to rescan the list from the beginning. Usually, the list
+		 * 'lwp_register_list' is very short, and there is 'guard'
+		 * lri::lri_exp that will prevent the callback to be done
+		 * repeatedly. So rescanning the list has no problem. */
+		goto again;
 	}
-	mutex_unlock(&lwp_register_list_lock);
+	spin_unlock(&lwp_register_list_lock);
 }
+EXPORT_SYMBOL(lustre_notify_lwp_list);
 
-static int lustre_lwp_connect(struct obd_device *lwp)
+static int lustre_lwp_connect(struct obd_device *lwp, bool is_mdt)
 {
 	struct lu_env		 env;
 	struct lu_context	 session_ctx;
@@ -536,11 +581,14 @@ static int lustre_lwp_connect(struct obd_device *lwp)
 
 	data->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_INDEX;
 	data->ocd_version = LUSTRE_VERSION_CODE;
-	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_LFSCK |
-		OBD_CONNECT_BULK_MBITS;
+	data->ocd_connect_flags |= OBD_CONNECT_FID | OBD_CONNECT_AT |
+		OBD_CONNECT_LRU_RESIZE | OBD_CONNECT_FULL20 |
+		OBD_CONNECT_LVB_TYPE | OBD_CONNECT_LIGHTWEIGHT |
+		OBD_CONNECT_LFSCK | OBD_CONNECT_BULK_MBITS;
+
+	if (is_mdt)
+		data->ocd_connect_flags |= OBD_CONNECT_MDS_MDS;
+
 	OBD_ALLOC_PTR(uuid);
 	if (uuid == NULL)
 		GOTO(out, rc = -ENOMEM);
@@ -560,7 +608,6 @@ static int lustre_lwp_connect(struct obd_device *lwp)
 		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);
@@ -625,12 +672,12 @@ static int lustre_lwp_setup(struct lustre_cfg *lcfg, struct lustre_sb_info *lsi,
 	obd = class_name2obd(lwpname);
 	LASSERT(obd != NULL);
 
-	rc = lustre_lwp_connect(obd);
+	rc = lustre_lwp_connect(obd, strstr(lsi->lsi_svname, "-MDT") != NULL);
 	if (rc == 0) {
 		obd->u.cli.cl_max_mds_easize = MAX_MD_SIZE;
-		spin_lock(&lsi->lsi_lwp_lock);
+		mutex_lock(&lsi->lsi_lwp_mutex);
 		list_add_tail(&obd->obd_lwp_list, &lsi->lsi_lwp_list);
-		spin_unlock(&lsi->lsi_lwp_lock);
+		mutex_unlock(&lsi->lsi_lwp_mutex);
 	} else {
 		CERROR("%s: connect failed: rc = %d\n", lwpname, rc);
 	}
@@ -707,15 +754,18 @@ static int lustre_lwp_add_conn(struct lustre_cfg *cfg,
 	lustre_cfg_bufs_set_string(bufs, 1,
 				   lustre_cfg_string(cfg, 1));
 
-	lcfg = lustre_cfg_new(LCFG_ADD_CONN, bufs);
-	if (lcfg == NULL)
+	OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
+	if (!lcfg)
 		GOTO(out_cfg, rc = -ENOMEM);
+	lustre_cfg_init(lcfg, LCFG_ADD_CONN, bufs);
+
 	rc = class_add_conn(lwp, lcfg);
 	if (rc)
 		CERROR("%s: can't add conn: rc = %d\n", lwpname, rc);
 
-	if (lcfg != NULL)
-		lustre_cfg_free(lcfg);
+	if (lcfg)
+		OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
+					      lcfg->lcfg_buflens));
 out_cfg:
 	if (bufs != NULL)
 		OBD_FREE_PTR(bufs);
@@ -747,7 +797,7 @@ static int client_lwp_config_process(const struct lu_env *env,
 				     struct llog_handle *handle,
 				     struct llog_rec_hdr *rec, void *data)
 {
-	struct config_llog_instance *clli = data;
+	struct config_llog_instance *cfg = data;
 	int			     cfg_len = rec->lrh_len;
 	char			    *cfg_buf = (char *) (rec + 1);
 	struct lustre_cfg	    *lcfg = NULL;
@@ -761,8 +811,9 @@ static int client_lwp_config_process(const struct lu_env *env,
 		RETURN(-EINVAL);
 	}
 
-	LASSERT(clli->cfg_sb != NULL);
-	lsi = s2lsi(clli->cfg_sb);
+	if (cfg->cfg_sb == NULL)
+		GOTO(out, rc = -EINVAL);
+	lsi = s2lsi(cfg->cfg_sb);
 
 	lcfg = (struct lustre_cfg *)cfg_buf;
 	if (lcfg->lcfg_version == __swab32(LUSTRE_CFG_VERSION)) {
@@ -784,35 +835,35 @@ static int client_lwp_config_process(const struct lu_env *env,
 		    marker->cm_flags & CM_EXCLUDE)
 			GOTO(out, rc = 0);
 
-		if (!tgt_is_mdt(marker->cm_tgtname, &clli->cfg_lwp_idx))
+		if (!tgt_is_mdt(marker->cm_tgtname, &cfg->cfg_lwp_idx))
 			GOTO(out, rc = 0);
 
-		if (IS_MDT(lsi) && clli->cfg_lwp_idx != 0)
+		if (IS_MDT(lsi) && cfg->cfg_lwp_idx != 0)
 			GOTO(out, rc = 0);
 
 		if (!strncmp(marker->cm_comment, "add mdc", 7) ||
 		    !strncmp(marker->cm_comment, "add failnid", 11)) {
 			if (marker->cm_flags & CM_START) {
-				clli->cfg_flags = CFG_F_MARKER;
+				cfg->cfg_flags = CFG_F_MARKER;
 				/* This hack is to differentiate the
 				 * ADD_UUID is come from "add mdc" record
 				 * or from "add failnid" record. */
 				if (!strncmp(marker->cm_comment,
 					     "add failnid", 11))
-					clli->cfg_flags |= CFG_F_SKIP;
+					cfg->cfg_flags |= CFG_F_SKIP;
 			} else if (marker->cm_flags & CM_END) {
-				clli->cfg_flags = 0;
+				cfg->cfg_flags = 0;
 			}
 		}
 		break;
 	}
 	case LCFG_ADD_UUID: {
-		if (clli->cfg_flags == CFG_F_MARKER) {
-			rc = lustre_lwp_setup(lcfg, lsi, clli->cfg_lwp_idx);
+		if (cfg->cfg_flags == CFG_F_MARKER) {
+			rc = lustre_lwp_setup(lcfg, lsi, cfg->cfg_lwp_idx);
 			/* XXX: process only the first nid as
 			 * we don't need another instance of lwp */
-			clli->cfg_flags |= CFG_F_SKIP;
-		} else if (clli->cfg_flags == (CFG_F_MARKER | CFG_F_SKIP)) {
+			cfg->cfg_flags |= CFG_F_SKIP;
+		} else if (cfg->cfg_flags == (CFG_F_MARKER | CFG_F_SKIP)) {
 			rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
 					    lcfg->lcfg_nid);
 			if (rc)
@@ -829,6 +880,20 @@ static int client_lwp_config_process(const struct lu_env *env,
 		if (!is_mdc_device(devname))
 			break;
 
+		if (!(cfg->cfg_flags & CFG_F_MARKER)) {
+			CDEBUG(D_CONFIG, "Skipping add_conn for %s, rec %d\n",
+			       devname, rec->lrh_index);
+			break;
+		}
+
+		/* add_conn should follow by add_uuid. This
+		 * guarantee lwp device was created
+		 */
+		if (!(cfg->cfg_flags & CFG_F_SKIP)) {
+			CWARN("Error at config for %s rec %d, add_conn should follow by add_uuid\n",
+			      devname, rec->lrh_index);
+			break;
+		}
 		ptr = strrchr(devname, '-');
 		if (ptr == NULL)
 			break;
@@ -882,7 +947,7 @@ static int lustre_disconnect_lwp(struct super_block *sb)
 			GOTO(out, rc = -ENOMEM);
 
 		/* end log first */
-		cfg->cfg_instance = sb;
+		cfg->cfg_instance = ll_get_cfg_instance(sb);
 		rc = lustre_end_log(sb, logname, cfg);
 		if (rc != 0 && rc != -ENOENT)
 			GOTO(out, rc);
@@ -894,6 +959,7 @@ static int lustre_disconnect_lwp(struct super_block *sb)
 	if (bufs == NULL)
 		GOTO(out, rc = -ENOMEM);
 
+	mutex_lock(&lsi->lsi_lwp_mutex);
 	list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
 		struct lustre_cfg *lcfg;
 
@@ -904,21 +970,27 @@ static int lustre_disconnect_lwp(struct super_block *sb)
 
 		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);
+		OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount,
+					       bufs->lcfg_buflen));
+		if (!lcfg) {
+			rc = -ENOMEM;
+			break;
+		}
+		lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
 
 		/* 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);
+		OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
+					      lcfg->lcfg_buflens));
 		if (rc != 0 && rc != -ETIMEDOUT) {
 			CERROR("%s: fail to disconnect LWP: rc = %d\n",
 			       lwp->obd_name, rc);
 			rc1 = rc;
 		}
 	}
+	mutex_unlock(&lsi->lsi_lwp_mutex);
 
 	GOTO(out, rc);
 
@@ -944,18 +1016,23 @@ static int lustre_stop_lwp(struct super_block *sb)
 	int			 rc1 = 0;
 	ENTRY;
 
+	mutex_lock(&lsi->lsi_lwp_mutex);
 	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;
+		mutex_unlock(&lsi->lsi_lwp_mutex);
+
 		rc = class_manual_cleanup(lwp);
 		if (rc != 0) {
 			CERROR("%s: fail to stop LWP: rc = %d\n",
 			       lwp->obd_name, rc);
 			rc1 = rc;
 		}
+		mutex_lock(&lsi->lsi_lwp_mutex);
 	}
+	mutex_unlock(&lsi->lsi_lwp_mutex);
 
 	RETURN(rc1 != 0 ? rc1 : rc);
 }
@@ -991,7 +1068,7 @@ static int lustre_start_lwp(struct super_block *sb)
 		GOTO(out, rc = -ENOMEM);
 
 	cfg->cfg_callback = client_lwp_config_process;
-	cfg->cfg_instance = sb;
+	cfg->cfg_instance = ll_get_cfg_instance(sb);
 	rc = lustre_process_log(sb, logname, cfg);
 	/* need to remove config llog from mgc */
 	lsi->lsi_lwp_started = 1;
@@ -1014,6 +1091,7 @@ static int server_stop_servers(int lsiflags)
 	struct obd_device *obd = NULL;
 	struct obd_type *type = NULL;
 	int rc = 0;
+	bool type_last;
 	ENTRY;
 
 	mutex_lock(&server_start_lock);
@@ -1022,23 +1100,32 @@ static int server_stop_servers(int lsiflags)
 	/* if this was an MDT, and there are no more MDT's, clean up the MDS */
 	if (lsiflags & LDD_F_SV_TYPE_MDT) {
 		obd = class_name2obd(LUSTRE_MDS_OBDNAME);
-		if (obd != NULL)
-			type = class_search_type(LUSTRE_MDT_NAME);
-	}
-
+		type = class_search_type(LUSTRE_MDT_NAME);
+	} else if (lsiflags & LDD_F_SV_TYPE_OST) {
 	/* if this was an OST, and there are no more OST's, clean up the OSS */
-	if (lsiflags & LDD_F_SV_TYPE_OST) {
 		obd = class_name2obd(LUSTRE_OSS_OBDNAME);
-		if (obd != NULL)
-			type = class_search_type(LUSTRE_OST_NAME);
+		type = class_search_type(LUSTRE_OST_NAME);
 	}
 
-	if (obd != NULL && (type == NULL || type->typ_refcnt == 0)) {
+	/* server_stop_servers is a pair of server_start_targets
+	 * Here we put type which was taken at server_start_targets.
+	 * If type is NULL then there is a wrong logic around type or
+	 * type reference. */
+	LASSERTF(type, "Server flags %d, obd %s\n", lsiflags,
+		 obd ? obd->obd_name : "NULL");
+
+	type_last = (atomic_read(&type->typ_refcnt) == 1);
+
+	class_put_type(type);
+	if (obd != NULL && type_last) {
 		obd->obd_force = 1;
 		/* obd_fail doesn't mean much on a server obd */
 		rc = class_manual_cleanup(obd);
 	}
 
+	/* put reference taken by class_search_type */
+	kobject_put(&type->typ_kobj);
+
 	mutex_unlock(&server_start_lock);
 
 	RETURN(rc);
@@ -1059,7 +1146,7 @@ int server_mti_print(const char *title, struct mgs_target_info *mti)
 static int server_lsi2mti(struct lustre_sb_info *lsi,
 			  struct mgs_target_info *mti)
 {
-	lnet_process_id_t id;
+	struct lnet_process_id id;
 	int rc, i = 0;
 	int cplen = 0;
 	ENTRY;
@@ -1097,6 +1184,13 @@ static int server_lsi2mti(struct lustre_sb_info *lsi,
 		}
 	}
 
+	if (mti->mti_nid_count == 0) {
+		CERROR("Failed to get NID for server %s, please check whether "
+		       "the target is specifed with improper --servicenode or "
+		       "--network options.\n", mti->mti_svname);
+		RETURN(-EINVAL);
+	}
+
 	mti->mti_lustre_ver = LUSTRE_VERSION_CODE;
 	mti->mti_config_ver = 0;
 
@@ -1128,6 +1222,7 @@ static int server_register_target(struct lustre_sb_info *lsi)
 	struct mgs_target_info *mti = NULL;
 	bool writeconf;
 	int rc;
+	int tried = 0;
 	ENTRY;
 
 	LASSERT(mgc);
@@ -1152,6 +1247,7 @@ static int server_register_target(struct lustre_sb_info *lsi)
 	writeconf = !!(lsi->lsi_flags & (LDD_F_NEED_INDEX | LDD_F_UPDATE));
 	mti->mti_flags |= LDD_F_OPC_REG;
 
+again:
 	/* Register the target */
 	/* FIXME use mgc_process_config instead */
 	rc = obd_set_info_async(NULL, mgc->u.cli.cl_mgc_mgsexp,
@@ -1165,6 +1261,15 @@ static int server_register_target(struct lustre_sb_info *lsi)
 				"to start: rc = %d. Please see messages on "
 				"the MGS.\n", lsi->lsi_svname, rc);
 		} else if (writeconf) {
+			if ((rc == -ESHUTDOWN || rc == -EIO) && ++tried < 5) {
+				/* The connection with MGS is not established.
+				 * Try again after 2 seconds. Interruptable. */
+				schedule_timeout_interruptible(
+					cfs_time_seconds(2));
+				if (!signal_pending(current))
+					goto again;
+			}
+
 			LCONSOLE_ERROR_MSG(0x15f,
 				"%s: cannot register this server with the MGS: "
 				"rc = %d. Is the MGS running?\n",
@@ -1236,55 +1341,65 @@ static int server_start_targets(struct super_block *sb)
 	struct obd_device *obd;
 	struct lustre_sb_info *lsi = s2lsi(sb);
 	struct config_llog_instance cfg;
-	struct lu_env env;
+	struct lu_env mgc_env;
 	struct lu_device *dev;
+	char *name_service, *obd_name_service = NULL;
+	struct obd_type *type = NULL;
 	int rc;
 	ENTRY;
 
 	CDEBUG(D_MOUNT, "starting target %s\n", lsi->lsi_svname);
 
+	LASSERTF(IS_MDT(lsi) || IS_OST(lsi), "designed for MDT or OST only\n");
+
 	if (IS_MDT(lsi)) {
-		/* make sure the MDS is started */
-		mutex_lock(&server_start_lock);
-		obd = class_name2obd(LUSTRE_MDS_OBDNAME);
-		if (!obd) {
-			rc = lustre_start_simple(LUSTRE_MDS_OBDNAME,
-						 LUSTRE_MDS_NAME,
-						 LUSTRE_MDS_OBDNAME"_uuid",
-						 NULL, NULL, NULL, NULL);
-			if (rc) {
-				mutex_unlock(&server_start_lock);
-				CERROR("failed to start MDS: %d\n", rc);
-				RETURN(rc);
-			}
-		}
-		mutex_unlock(&server_start_lock);
+		obd_name_service = LUSTRE_MDS_OBDNAME;
+		name_service = LUSTRE_MDS_NAME;
+	} else {
+		obd_name_service = LUSTRE_OSS_OBDNAME;
+		name_service = LUSTRE_OSS_NAME;
 	}
 
-	/* If we're an OST, make sure the global OSS is running */
-	if (IS_OST(lsi)) {
-		/* make sure OSS is started */
-		mutex_lock(&server_start_lock);
-		obd = class_name2obd(LUSTRE_OSS_OBDNAME);
-		if (!obd) {
-			rc = lustre_start_simple(LUSTRE_OSS_OBDNAME,
-						 LUSTRE_OSS_NAME,
-						 LUSTRE_OSS_OBDNAME"_uuid",
-						 NULL, NULL, NULL, NULL);
-			if (rc) {
-				mutex_unlock(&server_start_lock);
-				CERROR("failed to start OSS: %d\n", rc);
-				RETURN(rc);
-			}
+	/* make sure MDS/OSS is started */
+	mutex_lock(&server_start_lock);
+	obd = class_name2obd(obd_name_service);
+	if (!obd) {
+		rc = lustre_start_simple(obd_name_service, name_service,
+					 (IS_MDT(lsi) ?
+					  LUSTRE_MDS_OBDNAME"_uuid" :
+					  LUSTRE_OSS_OBDNAME"_uuid"),
+					 NULL, NULL, NULL, NULL);
+		if (rc) {
+			mutex_unlock(&server_start_lock);
+			CERROR("failed to start %s: %d\n",
+			       obd_name_service, rc);
+			RETURN(rc);
 		}
+	}
+	/* hold a type reference and put it at server_stop_servers */
+	type = class_get_type(IS_MDT(lsi) ?
+			      LUSTRE_MDT_NAME : LUSTRE_OST_NAME);
+	if (!type) {
 		mutex_unlock(&server_start_lock);
+		GOTO(out_stop_service, rc = -ENODEV);
 	}
+	lsi->lsi_server_started = 1;
+	mutex_unlock(&server_start_lock);
+	if (OBD_FAIL_PRECHECK(OBD_FAIL_OBD_STOP_MDS_RACE) &&
+	    IS_MDT(lsi)) {
+		OBD_RACE(OBD_FAIL_OBD_STOP_MDS_RACE);
+		msleep(2 * MSEC_PER_SEC);
+	}
+
+	rc = lu_env_init(&mgc_env, LCT_MG_THREAD);
+	if (rc != 0)
+		GOTO(out_stop_service, rc);
 
 	/* 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. */
-	rc = server_mgc_set_fs(lsi->lsi_mgc, sb);
+	rc = server_mgc_set_fs(&mgc_env, lsi->lsi_mgc, sb);
 	if (rc)
-		GOTO(out_stop_service, rc);
+		GOTO(out_env, rc);
 
 	/* Register with MGS */
 	rc = server_register_target(lsi);
@@ -1300,6 +1415,7 @@ static int server_start_targets(struct super_block *sb)
 	/* Start targets using the llog named for the target */
 	memset(&cfg, 0, sizeof(cfg));
 	cfg.cfg_callback = class_config_llog_handler;
+	cfg.cfg_sub_clds = CONFIG_SUB_SERVER;
 	rc = lustre_process_log(sb, lsi->lsi_svname, &cfg);
 	if (rc) {
 		CERROR("failed to start server %s: %d\n",
@@ -1330,12 +1446,11 @@ static int server_start_targets(struct super_block *sb)
 	/* calculate recovery timeout, do it after lustre_process_log */
 	server_calc_timeout(lsi, obd);
 
-	/* log has been fully processed */
-	obd_notify(obd, NULL, OBD_NOTIFY_CONFIG, (void *)CONFIG_LOG);
-
 	/* log has been fully processed, let clients connect */
 	dev = obd->obd_lu_dev;
 	if (dev && dev->ld_ops->ldo_prepare) {
+		struct lu_env env;
+
 		rc = lu_env_init(&env, dev->ld_type->ldt_ctx_tags);
 		if (rc == 0) {
 			struct lu_context  session_ctx;
@@ -1363,11 +1478,13 @@ static int server_start_targets(struct super_block *sb)
 
 out_mgc:
 	/* Release the mgc fs for others to use */
-	server_mgc_clear_fs(lsi->lsi_mgc);
-
+	server_mgc_clear_fs(&mgc_env, lsi->lsi_mgc);
+out_env:
+	lu_env_fini(&mgc_env);
 out_stop_service:
-	if (rc != 0)
-		server_stop_servers(lsi->lsi_flags);
+	/* in case of error upper function call
+	 * server_put_super->server_stop_servers()
+	 */
 
 	RETURN(rc);
 }
@@ -1453,6 +1570,7 @@ static void server_put_super(struct super_block *sb)
 	char *tmpname, *extraname = NULL;
 	int tmpname_sz;
 	int lsiflags = lsi->lsi_flags;
+	bool stop_servers = lsi->lsi_server_started;
 	ENTRY;
 
 	LASSERT(IS_SERVER(lsi));
@@ -1505,6 +1623,13 @@ static void server_put_super(struct super_block *sb)
 			 * to .put_super, so we better make sure we clean up! */
 			obd->obd_force = 1;
 			class_manual_cleanup(obd);
+			if (OBD_FAIL_PRECHECK(OBD_FAIL_OBD_STOP_MDS_RACE)) {
+				int idx;
+				server_name2index(lsi->lsi_svname, &idx, NULL);
+				/* sleeping for MDT0001 */
+				if (idx == 1)
+					OBD_RACE(OBD_FAIL_OBD_STOP_MDS_RACE);
+			}
 		} else {
 			CERROR("no obd %s\n", lsi->lsi_svname);
 			server_deregister_mount(lsi->lsi_svname);
@@ -1535,7 +1660,8 @@ static void server_put_super(struct super_block *sb)
 	/* Stop the servers (MDS, OSS) if no longer needed.  We must wait
 	   until the target is really gone so that our type refcount check
 	   is right. */
-	server_stop_servers(lsiflags);
+	if (stop_servers)
+		server_stop_servers(lsiflags);
 
 	/* In case of startup or cleanup err, stop related obds */
 	if (extraname) {
@@ -1597,6 +1723,63 @@ static int server_statfs(struct dentry *dentry, struct kstatfs *buf)
 	RETURN(0);
 }
 
+#ifdef HAVE_SUPEROPS_USE_DENTRY
+int server_show_options(struct seq_file *seq, struct dentry *dentry)
+#else
+int server_show_options(struct seq_file *seq, struct vfsmount *vfs)
+#endif
+{
+	struct lustre_sb_info *lsi;
+	struct lustre_mount_data *lmd;
+
+#ifdef HAVE_SUPEROPS_USE_DENTRY
+	LASSERT(seq != NULL && dentry != NULL);
+	lsi = s2lsi(dentry->d_sb);
+#else
+	LASSERT(seq != NULL && vfs != NULL);
+	lsi = s2lsi(vfs->mnt_sb);
+#endif
+
+	lmd = lsi->lsi_lmd;
+	seq_printf(seq, ",svname=%s", lmd->lmd_profile);
+
+	if  (lmd->lmd_flags & LMD_FLG_ABORT_RECOV)
+		seq_puts(seq, ",abort_recov");
+
+	if (lmd->lmd_flags & LMD_FLG_NOIR)
+		seq_puts(seq, ",noir");
+
+	if (lmd->lmd_flags & LMD_FLG_NOSVC)
+		seq_puts(seq, ",nosvc");
+
+	if (lmd->lmd_flags & LMD_FLG_NOMGS)
+		seq_puts(seq, ",nomgs");
+
+	if (lmd->lmd_flags & LMD_FLG_NOSCRUB)
+		seq_puts(seq, ",noscrub");
+	if (lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
+		seq_puts(seq, ",skip_lfsck");
+
+	if (lmd->lmd_flags & LMD_FLG_DEV_RDONLY)
+		seq_puts(seq, ",rdonly_dev");
+
+	if (lmd->lmd_flags & LMD_FLG_MGS)
+		seq_puts(seq, ",mgs");
+
+	if (lmd->lmd_mgs != NULL)
+		seq_printf(seq, ",mgsnode=%s", lmd->lmd_mgs);
+
+	if (lmd->lmd_osd_type != NULL)
+		seq_printf(seq, ",osd=%s", lmd->lmd_osd_type);
+
+	if (lmd->lmd_opts != NULL) {
+		seq_putc(seq, ',');
+		seq_puts(seq, lmd->lmd_opts);
+	}
+
+	RETURN(0);
+}
+
 /** The operations we support directly on the superblock:
  * mount, umount, and df.
  */
@@ -1604,11 +1787,13 @@ static struct super_operations server_ops = {
 	.put_super	= server_put_super,
 	.umount_begin	= server_umount_begin, /* umount -f */
 	.statfs		= server_statfs,
+	.show_options	= server_show_options,
 };
 
 /*
  * Xattr support for Lustre servers
  */
+#ifdef HAVE_IOP_XATTR
 static ssize_t lustre_getxattr(struct dentry *dentry, const char *name,
 				void *buffer, size_t size)
 {
@@ -1622,6 +1807,7 @@ static int lustre_setxattr(struct dentry *dentry, const char *name,
 {
 	return -EOPNOTSUPP;
 }
+#endif
 
 static ssize_t lustre_listxattr(struct dentry *d_entry, char *name,
 				size_t size)
@@ -1629,12 +1815,55 @@ static ssize_t lustre_listxattr(struct dentry *d_entry, char *name,
 	return -EOPNOTSUPP;
 }
 
+static bool is_cmd_supported(unsigned int command)
+{
+	switch (command) {
+	case FITRIM:
+		return true;
+	default:
+		return false;
+	}
+
+	return false;
+}
+
+static long server_ioctl(struct file *filp, unsigned int command,
+			 unsigned long arg)
+{
+	struct file active_filp;
+	struct inode *inode = file_inode(filp);
+	struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
+	struct super_block *dd_sb = dt_mnt_sb_get(lsi->lsi_dt_dev);
+	struct inode *active_inode;
+	int err = -EOPNOTSUPP;
+
+	if (IS_ERR(dd_sb) || !is_cmd_supported(command))
+		return err;
+
+	active_inode = igrab(dd_sb->s_root->d_inode);
+	if (!active_inode)
+		return -EACCES;
+
+	active_filp.f_inode = active_inode;
+	if (active_inode->i_fop && active_inode->i_fop->unlocked_ioctl)
+		err = active_inode->i_fop->unlocked_ioctl(&active_filp,
+							  command, arg);
+	iput(active_inode);
+	return err;
+}
+
 static const struct inode_operations server_inode_operations = {
+#ifdef HAVE_IOP_XATTR
 	.setxattr       = lustre_setxattr,
 	.getxattr       = lustre_getxattr,
+#endif
 	.listxattr      = lustre_listxattr,
 };
 
+static const struct file_operations server_file_operations = {
+	.unlocked_ioctl = server_ioctl,
+};
+
 #define log2(n) ffz(~(n))
 #define LUSTRE_SUPER_MAGIC 0x0BD00BD1
 
@@ -1649,7 +1878,7 @@ static int server_fill_super_common(struct super_block *sb)
 	sb->s_blocksize_bits = log2(sb->s_blocksize);
 	sb->s_magic = LUSTRE_SUPER_MAGIC;
 	sb->s_maxbytes = 0; /* we don't allow file IO on server mountpoints */
-	sb->s_flags |= MS_RDONLY;
+	sb->s_flags |= SB_RDONLY;
 	sb->s_op = &server_ops;
 
 	root = new_inode(sb);
@@ -1663,6 +1892,7 @@ static int server_fill_super_common(struct super_block *sb)
 	/* apparently we need to be a directory for the mount to finish */
 	root->i_mode = S_IFDIR;
 	root->i_op = &server_inode_operations;
+	root->i_fop = &server_file_operations;
 	sb->s_root = d_make_root(root);
 	if (!sb->s_root) {
 		CERROR("%s: can't make root dentry\n", sb->s_id);
@@ -1675,10 +1905,10 @@ static int server_fill_super_common(struct super_block *sb)
 static int osd_start(struct lustre_sb_info *lsi, unsigned long mflags)
 {
 	struct lustre_mount_data *lmd = lsi->lsi_lmd;
-	struct obd_device	 *obd;
-	struct dt_device_param    p;
-	char			  flagstr[16];
-	int			  rc;
+	struct obd_device *obd;
+	struct dt_device_param p;
+	char flagstr[20 + 1 + 10 + 1];
+	int rc;
 	ENTRY;
 
 	CDEBUG(D_MOUNT,
@@ -1688,7 +1918,7 @@ static int osd_start(struct lustre_sb_info *lsi, unsigned long mflags)
 	sprintf(lsi->lsi_osd_obdname, "%s-osd", lsi->lsi_svname);
 	strcpy(lsi->lsi_osd_uuid, lsi->lsi_osd_obdname);
 	strcat(lsi->lsi_osd_uuid, "_UUID");
-	sprintf(flagstr, "%lu:%lu", mflags, (unsigned long) lmd->lmd_flags);
+	snprintf(flagstr, sizeof(flagstr), "%lu:%u", mflags, lmd->lmd_flags);
 
 	obd = class_name2obd(lsi->lsi_osd_obdname);
 	if (obd == NULL) {
@@ -1715,8 +1945,6 @@ static int osd_start(struct lustre_sb_info *lsi, unsigned long mflags)
 	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);
@@ -1749,9 +1977,14 @@ int server_fill_super(struct super_block *sb)
 	int rc;
 	ENTRY;
 
+	/* to simulate target mount race */
+	OBD_RACE(OBD_FAIL_TGT_MOUNT_RACE);
+
 	rc = lsi_prepare(lsi);
-	if (rc)
+	if (rc) {
+		lustre_put_lsi(sb);
 		RETURN(rc);
+	}
 
 	/* Start low level OSD */
 	rc = osd_start(lsi, sb->s_flags);
@@ -1854,27 +2087,25 @@ void server_calc_timeout(struct lustre_sb_info *lsi, struct obd_device *obd)
 
 	if (has_ir) {
 		int new_soft = soft;
-		int new_hard = hard;
 
 		/* adjust timeout value by imperative recovery */
-
 		new_soft = (soft * factor) / OBD_IR_FACTOR_MAX;
-		new_hard = (hard * factor) / OBD_IR_FACTOR_MAX;
-
 		/* make sure the timeout is not too short */
 		new_soft = max(min, new_soft);
-		new_hard = max(new_soft, new_hard);
 
 		LCONSOLE_INFO("%s: Imperative Recovery enabled, recovery "
 			      "window shrunk from %d-%d down to %d-%d\n",
-			      obd->obd_name, soft, hard, new_soft, new_hard);
+			      obd->obd_name, soft, hard, new_soft, hard);
 
 		soft = new_soft;
-		hard = new_hard;
+	} else {
+		LCONSOLE_INFO("%s: Imperative Recovery not enabled, recovery "
+			      "window %d-%d\n", obd->obd_name, soft, hard);
 	}
 
 	/* we're done */
-	obd->obd_recovery_timeout   = max(obd->obd_recovery_timeout, soft);
+	obd->obd_recovery_timeout = max_t(time64_t, obd->obd_recovery_timeout,
+					  soft);
 	obd->obd_recovery_time_hard = hard;
 	obd->obd_recovery_ir_factor = factor;
 }