Whamcloud - gitweb
Branch HEAD
[fs/lustre-release.git] / lustre / obdclass / obd_mount.c
index 9cef323..72b5128 100644 (file)
@@ -521,10 +521,13 @@ static int server_start_mgs(struct super_block *sb)
 
         rc = server_register_mount(LUSTRE_MGS_OBDNAME, sb, mnt);
 
-        if (!rc &&
-            ((rc = lustre_start_simple(LUSTRE_MGS_OBDNAME, LUSTRE_MGS_NAME,
-                                       LUSTRE_MGS_OBDNAME, 0, 0))))
-                server_deregister_mount(LUSTRE_MGS_OBDNAME);
+        if (!rc) {
+                rc = lustre_start_simple(LUSTRE_MGS_OBDNAME, LUSTRE_MGS_NAME,
+                                         LUSTRE_MGS_OBDNAME, 0, 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 */
+        }
 
         if (rc)
                 LCONSOLE_ERROR_MSG(0x15e, "Failed to start MGS '%s' (%d). "
@@ -564,7 +567,6 @@ DECLARE_MUTEX(mgc_start_lock);
  */
 static int lustre_start_mgc(struct super_block *sb)
 {
-        struct lustre_handle mgc_conn = {0, };
         struct obd_connect_data *data = NULL;
         struct lustre_sb_info *lsi = s2lsi(sb);
         struct obd_device *obd;
@@ -572,7 +574,7 @@ static int lustre_start_mgc(struct super_block *sb)
         struct obd_uuid *uuid;
         class_uuid_t uuidc;
         lnet_nid_t nid;
-        char *mgcname, *niduuid;
+        char *mgcname, *niduuid, *mgssec;
         char *ptr;
         int recov_bk;
         int rc = 0, i = 0, j, len;
@@ -615,10 +617,18 @@ static int lustre_start_mgc(struct super_block *sb)
                 GOTO(out_free, rc = -ENOMEM);
         sprintf(mgcname, "%s%s", LUSTRE_MGC_OBDNAME, libcfs_nid2str(nid));
 
+        mgssec = lsi->lsi_lmd->lmd_mgssec ? lsi->lsi_lmd->lmd_mgssec : "";
+
         mutex_down(&mgc_start_lock);
 
         obd = class_name2obd(mgcname);
         if (obd && !obd->obd_stopping) {
+                rc = obd_set_info_async(obd->obd_self_export,
+                                        strlen(KEY_MGSSEC), KEY_MGSSEC,
+                                        strlen(mgssec), mgssec, NULL);
+                if (rc)
+                        GOTO(out_free, rc);
+
                 /* Re-using an existing MGC */
                 atomic_inc(&obd->u.cli.cl_mgc_refcount);
 
@@ -731,6 +741,12 @@ static int lustre_start_mgc(struct super_block *sb)
                 GOTO(out_free, rc = -ENOTCONN);
         }
 
+        rc = obd_set_info_async(obd->obd_self_export,
+                                strlen(KEY_MGSSEC), KEY_MGSSEC,
+                                strlen(mgssec), mgssec, NULL);
+        if (rc)
+                GOTO(out_free, rc);
+
         /* Keep a refcount of servers/clients who started with "mount",
            so we know when we can get rid of the mgc. */
         atomic_set(&obd->u.cli.cl_mgc_refcount, 1);
@@ -751,14 +767,13 @@ static int lustre_start_mgc(struct super_block *sb)
         data->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_FID |
                                   OBD_CONNECT_AT;
         data->ocd_version = LUSTRE_VERSION_CODE;
-        rc = obd_connect(NULL, &mgc_conn, obd, &(obd->obd_uuid), data, NULL);
+        rc = obd_connect(NULL, &exp, obd, &(obd->obd_uuid), data, NULL);
         OBD_FREE_PTR(data);
         if (rc) {
                 CERROR("connect failed %d\n", rc);
                 GOTO(out, rc);
         }
 
-        exp = class_conn2export(&mgc_conn);
         obd->u.cli.cl_mgc_mgsexp = exp;
 
 out:
@@ -791,6 +806,7 @@ static int lustre_stop_mgc(struct super_block *sb)
         lsi->lsi_mgc = NULL;
 
         mutex_down(&mgc_start_lock);
+        LASSERT(atomic_read(&obd->u.cli.cl_mgc_refcount) > 0);
         if (!atomic_dec_and_test(&obd->u.cli.cl_mgc_refcount)) {
                 /* This is not fatal, every client that stops
                    will call in here. */
@@ -826,7 +842,8 @@ static int lustre_stop_mgc(struct super_block *sb)
 
         /* Clean the nid uuids */
         if (!niduuid)
-                RETURN(-ENOMEM);
+                GOTO(out, rc = -ENOMEM);
+
         for (i = 0; i < lsi->lsi_lmd->lmd_mgs_failnodes; i++) {
                 sprintf(ptr, "_%x", i);
                 rc = do_lcfg(LUSTRE_MGC_OBDNAME, 0, LCFG_DEL_UUID,
@@ -835,10 +852,11 @@ static int lustre_stop_mgc(struct super_block *sb)
                         CERROR("del MDC UUID %s failed: rc = %d\n",
                                niduuid, rc);
         }
-        OBD_FREE(niduuid, len);
-        /* class_import_put will get rid of the additional connections */
-
 out:
+        if (niduuid)
+                OBD_FREE(niduuid, len);
+
+        /* class_import_put will get rid of the additional connections */
         mutex_up(&mgc_start_lock);
         RETURN(rc);
 }
@@ -1132,7 +1150,9 @@ static int server_start_targets(struct super_block *sb, struct vfsmount *mnt)
         if (rc) {
                 CERROR("failed to start server %s: %d\n",
                        lsi->lsi_ldd->ldd_svname, rc);
-                server_deregister_mount(lsi->lsi_ldd->ldd_svname);
+                /* Do NOT call server_deregister_mount() here. This makes it
+                 * impossible to find mount later in cleanup time and leaves
+                 * @lsi and othder stuff leaked. -umka */
                 GOTO(out_mgc, rc);
         }
 
@@ -1165,15 +1185,15 @@ out_mgc:
 
 struct lustre_sb_info *lustre_init_lsi(struct super_block *sb)
 {
-        struct lustre_sb_info *lsi = NULL;
+        struct lustre_sb_info *lsi;
         ENTRY;
 
-        OBD_ALLOC(lsi, sizeof(*lsi));
+        OBD_ALLOC_PTR(lsi);
         if (!lsi)
                 RETURN(NULL);
-        OBD_ALLOC(lsi->lsi_lmd, sizeof(*lsi->lsi_lmd));
+        OBD_ALLOC_PTR(lsi->lsi_lmd);
         if (!lsi->lsi_lmd) {
-                OBD_FREE(lsi, sizeof(*lsi));
+                OBD_FREE_PTR(lsi);
                 RETURN(NULL);
         }
 
@@ -1193,10 +1213,8 @@ static int lustre_free_lsi(struct super_block *sb)
         struct lustre_sb_info *lsi = s2lsi(sb);
         ENTRY;
 
-        if (!lsi)
-                RETURN(0);
-
-        CDEBUG(D_MOUNT, "Freeing lsi\n");
+        LASSERT(lsi != NULL);
+        CDEBUG(D_MOUNT, "Freeing lsi %p\n", lsi);
 
         /* someone didn't call server_put_mount. */
         LASSERT(atomic_read(&lsi->lsi_mounts) == 0);
@@ -1211,6 +1229,9 @@ static int lustre_free_lsi(struct super_block *sb)
                 if (lsi->lsi_lmd->lmd_profile != NULL)
                         OBD_FREE(lsi->lsi_lmd->lmd_profile,
                                  strlen(lsi->lsi_lmd->lmd_profile) + 1);
+                if (lsi->lsi_lmd->lmd_mgssec != NULL)
+                        OBD_FREE(lsi->lsi_lmd->lmd_mgssec,
+                                 strlen(lsi->lsi_lmd->lmd_mgssec) + 1);
                 if (lsi->lsi_lmd->lmd_opts != NULL)
                         OBD_FREE(lsi->lsi_lmd->lmd_opts,
                                  strlen(lsi->lsi_lmd->lmd_opts) + 1);
@@ -1235,10 +1256,9 @@ static int lustre_put_lsi(struct super_block *sb)
         struct lustre_sb_info *lsi = s2lsi(sb);
         ENTRY;
 
-        LASSERT(lsi);
+        LASSERT(lsi != NULL);
 
         CDEBUG(D_MOUNT, "put %p %d\n", sb, atomic_read(&lsi->lsi_mounts));
-
         if (atomic_dec_and_test(&lsi->lsi_mounts)) {
                 lustre_free_lsi(sb);
                 RETURN(1);
@@ -1336,6 +1356,10 @@ static struct vfsmount *server_kernel_mount(struct super_block *sb)
                 GOTO(out_free, rc);
         }
 
+        if (lmd->lmd_flags & LMD_FLG_ABORT_RECOV)
+                simple_truncate(mnt->mnt_sb->s_root, mnt, LAST_RCVD,
+                                LR_CLIENT_START);
+
         OBD_PAGE_FREE(__page);
         lsi->lsi_ldd = ldd;   /* freed at lsi cleanup */
         CDEBUG(D_SUPER, "%s: mnt = %p\n", lmd->lmd_dev, mnt);
@@ -1358,8 +1382,8 @@ static void server_wait_finished(struct vfsmount *mnt)
         init_waitqueue_head(&waitq);
 
         while ((atomic_read(&mnt->mnt_count) > 1) && (retries > 0)) {
-                LCONSOLE_WARN("Mount still busy with %d refs, waiting for "
-                              "%d secs...\n",
+                LCONSOLE_WARN("%s: Mount still busy with %d refs, waiting for "
+                              "%d secs...\n", mnt->mnt_devname,
                               atomic_read(&mnt->mnt_count), retries);
 
                 /* Wait for a bit */
@@ -1368,8 +1392,8 @@ static void server_wait_finished(struct vfsmount *mnt)
                 l_wait_event(waitq, 0, &lwi);
         }
         if (atomic_read(&mnt->mnt_count) > 1) {
-                CERROR("Mount %p is still busy (%d refs), giving up.\n",
-                       mnt, atomic_read(&mnt->mnt_count));
+                CERROR("%s: Mount still busy (%d refs), giving up.\n",
+                       mnt->mnt_devname, atomic_read(&mnt->mnt_count));
         }
 }
 
@@ -1382,7 +1406,6 @@ static void server_put_super(struct super_block *sb)
         int tmpname_sz;
         int lddflags = lsi->lsi_ldd->ldd_flags;
         int lsiflags = lsi->lsi_flags;
-        int rc;
         ENTRY;
 
         LASSERT(lsiflags & LSI_SERVER);
@@ -1391,6 +1414,8 @@ static void server_put_super(struct super_block *sb)
         OBD_ALLOC(tmpname, tmpname_sz);
         memcpy(tmpname, lsi->lsi_ldd->ldd_svname, tmpname_sz);
         CDEBUG(D_MOUNT, "server put_super %s\n", tmpname);
+        if (IS_MDT(lsi->lsi_ldd) && (lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC))
+                snprintf(tmpname, tmpname_sz, "MGS");
 
         /* Stop the target */
         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
@@ -1427,17 +1452,13 @@ static void server_put_super(struct super_block *sb)
         /* If they wanted the mgs to stop separately from the mdt, they
            should have put it on a different device. */
         if (IS_MGS(lsi->lsi_ldd)) {
-                /* stop the mgc before the mgs so the connection gets cleaned
-                   up */
-                lustre_stop_mgc(sb);
                 /* if MDS start with --nomgs, don't stop MGS then */
                 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS))
                         server_stop_mgs(sb);
         }
 
         /* Clean the mgc and sb */
-        rc = lustre_common_put_super(sb);
-        /* FIXME how can I report a failure to umount? */
+        lustre_common_put_super(sb);
 
         /* Wait for the targets to really clean up - can't exit (and let the
            sb get destroyed) while the mount is still in use */
@@ -1586,9 +1607,9 @@ static int server_fill_super(struct super_block *sb)
         if (IS_ERR(mnt)) {
                 rc = PTR_ERR(mnt);
                 CERROR("Unable to mount device %s: %d\n",
-                      lsi->lsi_lmd->lmd_dev, rc);
+                       lsi->lsi_lmd->lmd_dev, rc);
                 lustre_put_lsi(sb);
-                GOTO(out, rc);
+                RETURN(rc);
         }
         lsi->lsi_srv_mnt = mnt;
 
@@ -1602,13 +1623,13 @@ static int server_fill_super(struct super_block *sb)
                                    "running. Double-mount may have compromised"
                                    " the disk journal.\n",
                                    lsi->lsi_ldd->ldd_svname);
-                unlock_mntput(mnt);
                 lustre_put_lsi(sb);
-                GOTO(out, rc = -EALREADY);
+                unlock_mntput(mnt);
+                RETURN(-EALREADY);
         }
 
-        /* start MGS before MGC */
-        if (IS_MGS(lsi->lsi_ldd) && !(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS)) {
+        /* Start MGS before MGC */
+        if (IS_MGS(lsi->lsi_ldd) && !(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS)){
                 rc = server_start_mgs(sb);
                 if (rc)
                         GOTO(out_mnt, rc);
@@ -1642,14 +1663,17 @@ static int server_fill_super(struct super_block *sb)
                 GOTO(out_mnt, rc);
 
         LCONSOLE_WARN("Server %s on device %s has started\n",
-                      lsi->lsi_ldd->ldd_svname, lsi->lsi_lmd->lmd_dev);
+                      ((lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
+                       (IS_MDT(lsi->lsi_ldd))) ? "MGS" : lsi->lsi_ldd->ldd_svname,
+                      lsi->lsi_lmd->lmd_dev);
 
         RETURN(0);
-
 out_mnt:
+        /* We jump here in case of failure while starting targets or MGS.
+         * In this case we can't just put @mnt and have to do real cleanup
+         * with stoping targets, etc. */
         server_put_super(sb);
-out:
-        RETURN(rc);
+        return rc;
 }
 
 /* Get the index from the obd name.
@@ -1660,16 +1684,27 @@ int server_name2index(char *svname, __u32 *idx, char **endptr)
 {
         unsigned long index;
         int rc;
-        char *dash = strchr(svname, '-');
+        char *dash = strrchr(svname, '-');
         if (!dash)
                 return(-EINVAL);
 
+        /* intepret <fsname>-MDTXXXXX-mdc as mdt, the better way is to pass
+         * in the fsname, then determine the server index */
+        if (!strcmp(LUSTRE_MDC_NAME, dash + 1)) {
+                dash--;
+                for (; dash > svname && *dash != '-'; dash--);
+                if (dash == svname)
+                        return(-EINVAL);
+        }
+
         if (strncmp(dash + 1, "MDT", 3) == 0)
                 rc = LDD_F_SV_TYPE_MDT;
         else if (strncmp(dash + 1, "OST", 3) == 0)
                 rc = LDD_F_SV_TYPE_OST;
         else
                 return(-EINVAL);
+        if (strcmp(dash + 4, "all") == 0)
+                return rc | LDD_F_SV_ALL;
 
         index = simple_strtoul(dash + 4, endptr, 16);
         *idx = index;
@@ -1699,6 +1734,7 @@ int lustre_common_put_super(struct super_block *sb)
         }
         /* Drop a ref to the mounted disk */
         lustre_put_lsi(sb);
+        lu_types_stop();
         RETURN(rc);
 }
 
@@ -1802,6 +1838,31 @@ static int lmd_make_exclusion(struct lustre_mount_data *lmd, char *ptr)
         RETURN(rc);
 }
 
+static int lmd_parse_mgssec(struct lustre_mount_data *lmd, char *ptr)
+{
+        char   *tail;
+        int     length;
+
+        if (lmd->lmd_mgssec != NULL) {
+                OBD_FREE(lmd->lmd_mgssec, strlen(lmd->lmd_mgssec) + 1);
+                lmd->lmd_mgssec = NULL;
+        }
+
+        tail = strchr(ptr, ',');
+        if (tail == NULL)
+                length = strlen(ptr);
+        else
+                length = tail - ptr;
+
+        OBD_ALLOC(lmd->lmd_mgssec, length + 1);
+        if (lmd->lmd_mgssec == NULL)
+                return -ENOMEM;
+
+        memcpy(lmd->lmd_mgssec, ptr, length);
+        lmd->lmd_mgssec[length] = '\0';
+        return 0;
+}
+
 /* mount -v -t lustre uml1:uml2:/lustre-client /mnt/lustre */
 static int lmd_parse(char *options, struct lustre_mount_data *lmd)
 {
@@ -1849,6 +1910,11 @@ static int lmd_parse(char *options, struct lustre_mount_data *lmd)
                 } else if (strncmp(s1, "nomgs", 5) == 0) {
                         lmd->lmd_flags |= LMD_FLG_NOMGS;
                         clear++;
+                } else if (strncmp(s1, "mgssec=", 7) == 0) {
+                        rc = lmd_parse_mgssec(lmd, s1 + 7);
+                        if (rc)
+                                goto invalid;
+                        clear++;
                 /* ost exclusion list */
                 } else if (strncmp(s1, "exclude=", 8) == 0) {
                         rc = lmd_make_exclusion(lmd, s1 + 7);
@@ -1942,10 +2008,16 @@ int lustre_fill_super(struct super_block *sb, void *data, int silent)
                 RETURN(-ENOMEM);
         lmd = lsi->lsi_lmd;
 
+        /*
+         * Disable lockdep during mount, because mount locking patterns are
+         * `special'.
+         */
+        lockdep_off();
+
         /* Figure out the lmd from the mount options */
         if (lmd_parse((char *)data, lmd)) {
                 lustre_put_lsi(sb);
-                RETURN(-EINVAL);
+                GOTO(out, rc = -EINVAL);
         }
 
         if (lmd_is_client(lmd)) {
@@ -1954,12 +2026,13 @@ int lustre_fill_super(struct super_block *sb, void *data, int silent)
                         LCONSOLE_ERROR_MSG(0x165, "Nothing registered for "
                                            "client mount! Is the 'lustre' "
                                            "module loaded?\n");
+                        lustre_put_lsi(sb);
                         rc = -ENODEV;
                 } else {
                         rc = lustre_start_mgc(sb);
                         if (rc) {
-                                lustre_stop_mgc(sb);
-                                goto out;
+                                lustre_put_lsi(sb);
+                                GOTO(out, rc);
                         }
                         /* Connect and start */
                         /* (should always be ll_fill_super) */
@@ -1976,14 +2049,19 @@ int lustre_fill_super(struct super_block *sb, void *data, int silent)
                 /* s_f_s will call server_put_super on failure */
         }
 
+        /* If error happens in fill_super() call, @lsi will be killed there.
+         * This is why we do not put it here. */
+        GOTO(out, rc);
 out:
-        if (rc){
+        if (rc) {
                 CERROR("Unable to mount %s (%d)\n",
                        s2lsi(sb) ? lmd->lmd_dev : "", rc);
         } else {
-                CDEBUG(D_SUPER, "mount %s complete\n", lmd->lmd_dev);
+                CDEBUG(D_SUPER, "Mount %s complete\n",
+                       lmd->lmd_dev);
         }
-        RETURN(rc);
+        lockdep_on();
+        return rc;
 }