Whamcloud - gitweb
LU-8066 obd_type: discard obd_types linked list.
[fs/lustre-release.git] / lustre / obdclass / obd_mount_server.c
index c65cb53..5b5cb20 100644 (file)
 #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 <linux/version.h>
-#ifdef HAVE_KERNEL_LOCKED
-#include <linux/smp_lock.h>
-#endif
+#include <linux/delay.h>
 
 #include <llog_swab.h>
 #include <lustre_disk.h>
@@ -506,7 +506,7 @@ 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);
 
@@ -515,7 +515,7 @@ struct obd_export *lustre_find_lwp_by_index(const char *dev, __u32 idx)
                        break;
                }
        }
-       spin_unlock(&lsi->lsi_lwp_lock);
+       mutex_unlock(&lsi->lsi_lwp_mutex);
 
 err_lmi:
        server_put_mount(dev, false);
@@ -556,7 +556,7 @@ again:
 }
 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;
@@ -582,11 +582,14 @@ 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);
@@ -670,12 +673,12 @@ 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);
        }
@@ -931,7 +934,7 @@ 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);
@@ -943,6 +946,7 @@ 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;
 
@@ -955,8 +959,10 @@ static int lustre_disconnect_lwp(struct super_block *sb)
                lustre_cfg_bufs_set_string(bufs, 1, NULL);
                OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount,
                                               bufs->lcfg_buflen));
-               if (!lcfg)
-                       GOTO(out, rc = -ENOMEM);
+               if (!lcfg) {
+                       rc = -ENOMEM;
+                       break;
+               }
                lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
 
                /* Disconnect import first. NULL is passed for the '@env',
@@ -971,6 +977,7 @@ static int lustre_disconnect_lwp(struct super_block *sb)
                        rc1 = rc;
                }
        }
+       mutex_unlock(&lsi->lsi_lwp_mutex);
 
        GOTO(out, rc);
 
@@ -996,18 +1003,23 @@ 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);
 }
@@ -1043,7 +1055,7 @@ 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;
@@ -1066,6 +1078,7 @@ 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);
@@ -1074,23 +1087,32 @@ 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 = (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);
@@ -1230,8 +1252,7 @@ again:
                                /* The connection with MGS is not established.
                                 * Try again after 2 seconds. Interruptable. */
                                set_current_state(TASK_INTERRUPTIBLE);
-                               schedule_timeout(
-                                       msecs_to_jiffies(MSEC_PER_SEC) * 2);
+                               schedule_timeout(cfs_time_seconds(2));
                                set_current_state(TASK_RUNNING);
                                if (!signal_pending(current))
                                        goto again;
@@ -1310,46 +1331,48 @@ static int server_start_targets(struct super_block *sb)
        struct config_llog_instance cfg;
        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);
                }
-               mutex_unlock(&server_start_lock);
+       }
+       /* hold a type reference and put it at server_stop_servers */
+       type = class_get_type(IS_MDT(lsi) ?
+                             LUSTRE_MDT_NAME : LUSTRE_OST_NAME);
+       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);
@@ -1443,8 +1466,9 @@ out_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);
 }
@@ -1530,6 +1554,7 @@ 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));
@@ -1582,6 +1607,13 @@ 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);
@@ -1612,7 +1644,8 @@ 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) {
@@ -1766,6 +1799,43 @@ 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,
@@ -1774,6 +1844,10 @@ static const struct inode_operations server_inode_operations = {
        .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
 
@@ -1788,7 +1862,7 @@ 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);
@@ -1802,6 +1876,7 @@ 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);
@@ -1814,10 +1889,10 @@ 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,
@@ -1827,7 +1902,7 @@ 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) {