1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdclass/obd_mount.c
38 * Client/server mount routines
40 * Author: Nathan Rutman <nathan@clusterfs.com>
44 #define DEBUG_SUBSYSTEM S_CLASS
45 #define D_MOUNT D_SUPER|D_CONFIG /*|D_WARNING */
46 #define PRINT_CMD CDEBUG
47 #define PRINT_MASK D_SUPER|D_CONFIG
51 #include <lustre_fsfilt.h>
52 #include <obd_class.h>
53 #include <lustre/lustre_user.h>
54 #include <linux/version.h>
55 #include <lustre_log.h>
56 #include <lustre_disk.h>
57 #include <lustre_param.h>
59 static int (*client_fill_super)(struct super_block *sb) = NULL;
60 static void (*kill_super_cb)(struct super_block *sb) = NULL;
62 /*********** mount lookup *********/
64 DECLARE_MUTEX(lustre_mount_info_lock);
65 struct list_head server_mount_info_list = LIST_HEAD_INIT(server_mount_info_list);
67 static struct lustre_mount_info *server_find_mount(char *name)
69 struct list_head *tmp;
70 struct lustre_mount_info *lmi;
73 list_for_each(tmp, &server_mount_info_list) {
74 lmi = list_entry(tmp, struct lustre_mount_info, lmi_list_chain);
75 if (strcmp(name, lmi->lmi_name) == 0)
81 struct lustre_mount_info *server_find_mount_locked(char *name)
83 struct lustre_mount_info *lmi;
85 down(&lustre_mount_info_lock);
86 lmi = server_find_mount(name);
87 up(&lustre_mount_info_lock);
92 /* we must register an obd for a mount before we call the setup routine.
93 *_setup will call lustre_get_mount to get the mnt struct
94 by obd_name, since we can't pass the pointer to setup. */
95 static int server_register_mount(char *name, struct super_block *sb,
98 struct lustre_mount_info *lmi;
105 OBD_ALLOC(lmi, sizeof(*lmi));
108 OBD_ALLOC(name_cp, strlen(name) + 1);
110 OBD_FREE(lmi, sizeof(*lmi));
113 strcpy(name_cp, name);
115 down(&lustre_mount_info_lock);
117 if (server_find_mount(name)) {
118 up(&lustre_mount_info_lock);
119 OBD_FREE(lmi, sizeof(*lmi));
120 OBD_FREE(name_cp, strlen(name) + 1);
121 CERROR("Already registered %s\n", name);
124 lmi->lmi_name = name_cp;
127 list_add(&lmi->lmi_list_chain, &server_mount_info_list);
129 up(&lustre_mount_info_lock);
131 CDEBUG(D_MOUNT, "reg_mnt %p from %s, vfscount=%d\n",
132 lmi->lmi_mnt, name, atomic_read(&lmi->lmi_mnt->mnt_count));
137 /* when an obd no longer needs a mount */
138 static int server_deregister_mount(char *name)
140 struct lustre_mount_info *lmi;
143 down(&lustre_mount_info_lock);
144 lmi = server_find_mount(name);
146 up(&lustre_mount_info_lock);
147 CERROR("%s not registered\n", name);
151 CDEBUG(D_MOUNT, "dereg_mnt %p from %s, vfscount=%d\n",
152 lmi->lmi_mnt, name, atomic_read(&lmi->lmi_mnt->mnt_count));
154 OBD_FREE(lmi->lmi_name, strlen(lmi->lmi_name) + 1);
155 list_del(&lmi->lmi_list_chain);
156 OBD_FREE(lmi, sizeof(*lmi));
157 up(&lustre_mount_info_lock);
162 /* obd's look up a registered mount using their obdname. This is just
163 for initial obd setup to find the mount struct. It should not be
164 called every time you want to mntget. */
165 struct lustre_mount_info *server_get_mount(char *name)
167 struct lustre_mount_info *lmi;
168 struct lustre_sb_info *lsi;
171 down(&lustre_mount_info_lock);
172 lmi = server_find_mount(name);
173 up(&lustre_mount_info_lock);
175 CERROR("Can't find mount for %s\n", name);
178 lsi = s2lsi(lmi->lmi_sb);
179 mntget(lmi->lmi_mnt);
180 atomic_inc(&lsi->lsi_mounts);
182 CDEBUG(D_MOUNT, "get_mnt %p from %s, refs=%d, vfscount=%d\n",
183 lmi->lmi_mnt, name, atomic_read(&lsi->lsi_mounts),
184 atomic_read(&lmi->lmi_mnt->mnt_count));
189 static void unlock_mntput(struct vfsmount *mnt)
191 if (kernel_locked()) {
200 static int lustre_put_lsi(struct super_block *sb);
202 /* to be called from obd_cleanup methods */
203 int server_put_mount(char *name, struct vfsmount *mnt)
205 struct lustre_mount_info *lmi;
206 struct lustre_sb_info *lsi;
207 int count = atomic_read(&mnt->mnt_count) - 1;
210 /* This might be the last one, can't deref after this */
213 down(&lustre_mount_info_lock);
214 lmi = server_find_mount(name);
215 up(&lustre_mount_info_lock);
217 CERROR("Can't find mount for %s\n", name);
220 LASSERT(lmi->lmi_mnt == mnt);
221 /* Note if lsi is poisoned, it's probably because server_wait_finished
222 didn't wait long enough and the sb was freed at the umount. */
223 lsi = s2lsi(lmi->lmi_sb);
225 CDEBUG(D_MOUNT, "put_mnt %p from %s, refs=%d, vfscount=%d\n",
226 lmi->lmi_mnt, name, atomic_read(&lsi->lsi_mounts), count);
228 if (lustre_put_lsi(lmi->lmi_sb)) {
229 CDEBUG(D_MOUNT, "Last put of mnt %p from %s, vfscount=%d\n",
230 lmi->lmi_mnt, name, count);
231 /* last mount is the One True Mount */
233 CERROR("%s: mount busy, vfscount=%d!\n", name, count);
236 /* this obd should never need the mount again */
237 server_deregister_mount(name);
242 /******* mount helper utilities *********/
244 static void ldd_print(struct lustre_disk_data *ldd)
246 PRINT_CMD(PRINT_MASK, " disk data:\n");
247 PRINT_CMD(PRINT_MASK, "server: %s\n", ldd->ldd_svname);
248 PRINT_CMD(PRINT_MASK, "uuid: %s\n", (char *)ldd->ldd_uuid);
249 PRINT_CMD(PRINT_MASK, "fs: %s\n", ldd->ldd_fsname);
250 PRINT_CMD(PRINT_MASK, "index: %04x\n", ldd->ldd_svindex);
251 PRINT_CMD(PRINT_MASK, "config: %d\n", ldd->ldd_config_ver);
252 PRINT_CMD(PRINT_MASK, "flags: %#x\n", ldd->ldd_flags);
253 PRINT_CMD(PRINT_MASK, "diskfs: %s\n", MT_STR(ldd));
254 PRINT_CMD(PRINT_MASK, "options: %s\n", ldd->ldd_mount_opts);
255 PRINT_CMD(PRINT_MASK, "params: %s\n", ldd->ldd_params);
256 PRINT_CMD(PRINT_MASK, "comment: %s\n", ldd->ldd_userdata);
259 static int ldd_parse(struct lvfs_run_ctxt *mount_ctxt,
260 struct lustre_disk_data *ldd)
262 struct lvfs_run_ctxt saved;
269 push_ctxt(&saved, mount_ctxt, NULL);
271 file = filp_open(MOUNT_DATA_FILE, O_RDONLY, 0644);
274 CERROR("cannot open %s: rc = %d\n", MOUNT_DATA_FILE, rc);
278 len = i_size_read(file->f_dentry->d_inode);
279 CDEBUG(D_MOUNT, "Have %s, size %lu\n", MOUNT_DATA_FILE, len);
280 if (len != sizeof(*ldd)) {
281 CERROR("disk data size does not match: see %lu expect "LPSZ"\n",
283 GOTO(out_close, rc = -EINVAL);
286 rc = lustre_fread(file, ldd, len, &off);
288 CERROR("error reading %s: read %d of %lu\n",
289 MOUNT_DATA_FILE, rc, len);
290 GOTO(out_close, rc = -EINVAL);
294 if (ldd->ldd_magic != LDD_MAGIC) {
295 /* FIXME add swabbing support */
296 CERROR("Bad magic in %s: %x!=%x\n", MOUNT_DATA_FILE,
297 ldd->ldd_magic, LDD_MAGIC);
298 GOTO(out_close, rc = -EINVAL);
301 if (ldd->ldd_feature_incompat & ~LDD_INCOMPAT_SUPP) {
302 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
304 ldd->ldd_feature_incompat & ~LDD_INCOMPAT_SUPP);
305 GOTO(out_close, rc = -EINVAL);
307 if (ldd->ldd_feature_rocompat & ~LDD_ROCOMPAT_SUPP) {
308 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
310 ldd->ldd_feature_rocompat & ~LDD_ROCOMPAT_SUPP);
311 /* Do something like remount filesystem read-only */
312 GOTO(out_close, rc = -EINVAL);
320 pop_ctxt(&saved, mount_ctxt, NULL);
324 static int ldd_write(struct lvfs_run_ctxt *mount_ctxt,
325 struct lustre_disk_data *ldd)
327 struct lvfs_run_ctxt saved;
330 unsigned long len = sizeof(struct lustre_disk_data);
334 LASSERT(ldd->ldd_magic == LDD_MAGIC);
336 ldd->ldd_config_ver++;
338 push_ctxt(&saved, mount_ctxt, NULL);
340 file = filp_open(MOUNT_DATA_FILE, O_RDWR, 0644);
343 CERROR("cannot open %s: rc = %d\n", MOUNT_DATA_FILE, rc);
347 rc = lustre_fwrite(file, ldd, len, &off);
349 CERROR("error writing %s: read %d of %lu\n",
350 MOUNT_DATA_FILE, rc, len);
351 GOTO(out_close, rc = -EINVAL);
360 pop_ctxt(&saved, mount_ctxt, NULL);
365 /**************** config llog ********************/
367 /* Get a config log from the MGS and process it.
368 This func is called for both clients and servers.
369 Continue to process new statements appended to the logs
370 (whenever the config lock is revoked) until lustre_end_log
372 int lustre_process_log(struct super_block *sb, char *logname,
373 struct config_llog_instance *cfg)
375 struct lustre_cfg *lcfg;
376 struct lustre_cfg_bufs bufs;
377 struct lustre_sb_info *lsi = s2lsi(sb);
378 struct obd_device *mgc = lsi->lsi_mgc;
385 /* mgc_process_config */
386 lustre_cfg_bufs_reset(&bufs, mgc->obd_name);
387 lustre_cfg_bufs_set_string(&bufs, 1, logname);
388 lustre_cfg_bufs_set(&bufs, 2, cfg, sizeof(*cfg));
389 lustre_cfg_bufs_set(&bufs, 3, &sb, sizeof(sb));
390 lcfg = lustre_cfg_new(LCFG_LOG_START, &bufs);
391 rc = obd_process_config(mgc, sizeof(*lcfg), lcfg);
392 lustre_cfg_free(lcfg);
395 LCONSOLE_ERROR_MSG(0x15b, "%s: The configuration from log '%s' "
396 "failed (%d). Make sure this client and "
397 "the MGS are running compatible versions of "
399 mgc->obd_name, logname, rc);
402 LCONSOLE_ERROR_MSG(0x15c, "%s: The configuration from log '%s' "
403 "failed (%d). This may be the result of "
404 "communication errors between this node and "
405 "the MGS, a bad configuration, or other "
406 "errors. See the syslog for more "
407 "information.\n", mgc->obd_name, logname,
410 /* class_obd_list(); */
414 /* Stop watching this config log for updates */
415 int lustre_end_log(struct super_block *sb, char *logname,
416 struct config_llog_instance *cfg)
418 struct lustre_cfg *lcfg;
419 struct lustre_cfg_bufs bufs;
420 struct lustre_sb_info *lsi = s2lsi(sb);
421 struct obd_device *mgc = lsi->lsi_mgc;
428 /* mgc_process_config */
429 lustre_cfg_bufs_reset(&bufs, mgc->obd_name);
430 lustre_cfg_bufs_set_string(&bufs, 1, logname);
432 lustre_cfg_bufs_set(&bufs, 2, cfg, sizeof(*cfg));
433 lcfg = lustre_cfg_new(LCFG_LOG_END, &bufs);
434 rc = obd_process_config(mgc, sizeof(*lcfg), lcfg);
435 lustre_cfg_free(lcfg);
439 /**************** obd start *******************/
441 int do_lcfg(char *cfgname, lnet_nid_t nid, int cmd,
442 char *s1, char *s2, char *s3, char *s4)
444 struct lustre_cfg_bufs bufs;
445 struct lustre_cfg * lcfg = NULL;
448 CDEBUG(D_TRACE, "lcfg %s %#x %s %s %s %s\n", cfgname,
449 cmd, s1, s2, s3, s4);
451 lustre_cfg_bufs_reset(&bufs, cfgname);
453 lustre_cfg_bufs_set_string(&bufs, 1, s1);
455 lustre_cfg_bufs_set_string(&bufs, 2, s2);
457 lustre_cfg_bufs_set_string(&bufs, 3, s3);
459 lustre_cfg_bufs_set_string(&bufs, 4, s4);
461 lcfg = lustre_cfg_new(cmd, &bufs);
462 lcfg->lcfg_nid = nid;
463 rc = class_process_config(lcfg);
464 lustre_cfg_free(lcfg);
468 static int lustre_start_simple(char *obdname, char *type, char *uuid,
472 CDEBUG(D_MOUNT, "Starting obd %s (typ=%s)\n", obdname, type);
474 rc = do_lcfg(obdname, 0, LCFG_ATTACH, type, uuid, 0, 0);
476 CERROR("%s attach error %d\n", obdname, rc);
479 rc = do_lcfg(obdname, 0, LCFG_SETUP, s1, s2, 0, 0);
481 CERROR("%s setup error %d\n", obdname, rc);
482 do_lcfg(obdname, 0, LCFG_DETACH, 0, 0, 0, 0);
487 /* Set up a MGS to serve startup logs */
488 static int server_start_mgs(struct super_block *sb)
490 struct lustre_sb_info *lsi = s2lsi(sb);
491 struct vfsmount *mnt = lsi->lsi_srv_mnt;
492 struct lustre_mount_info *lmi;
497 /* It is impossible to have more than 1 MGS per node, since
498 MGC wouldn't know which to connect to */
499 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
501 lsi = s2lsi(lmi->lmi_sb);
502 LCONSOLE_ERROR_MSG(0x15d, "The MGS service was already started "
503 "from server %s\n", lsi->lsi_ldd->ldd_svname);
507 CDEBUG(D_CONFIG, "Start MGS service %s\n", LUSTRE_MGS_OBDNAME);
509 rc = server_register_mount(LUSTRE_MGS_OBDNAME, sb, mnt);
512 rc = lustre_start_simple(LUSTRE_MGS_OBDNAME, LUSTRE_MGS_NAME,
513 LUSTRE_MGS_OBDNAME, 0, 0);
514 /* Do NOT call server_deregister_mount() here. This leads to
515 * inability cleanup cleanly and free lsi and other stuff when
516 * mgs calls server_put_mount() in error handling case. -umka */
520 LCONSOLE_ERROR_MSG(0x15e, "Failed to start MGS '%s' (%d). Is "
521 "the 'mgs' module loaded?\n",
522 LUSTRE_MGS_OBDNAME, rc);
526 static int server_stop_mgs(struct super_block *sb)
528 struct obd_device *obd;
532 CDEBUG(D_MOUNT, "Stop MGS service %s\n", LUSTRE_MGS_OBDNAME);
534 /* There better be only one MGS */
535 obd = class_name2obd(LUSTRE_MGS_OBDNAME);
537 CDEBUG(D_CONFIG, "mgs %s not running\n", LUSTRE_MGS_OBDNAME);
541 /* The MGS should always stop when we say so */
543 rc = class_manual_cleanup(obd);
547 DECLARE_MUTEX(mgc_start_lock);
549 /** Set up a mgc obd to process startup logs
551 * \param sb [in] super block of the mgc obd
553 * \retval 0 success, otherwise error code
555 static int lustre_start_mgc(struct super_block *sb)
557 struct lustre_handle mgc_conn = {0, };
558 struct obd_connect_data *data = NULL;
559 struct lustre_sb_info *lsi = s2lsi(sb);
560 struct obd_device *obd;
561 struct obd_export *exp;
562 struct obd_uuid *uuid;
565 char *mgcname, *niduuid;
568 int rc = 0, i = 0, j, len;
571 LASSERT(lsi->lsi_lmd);
573 /* Find the first non-lo MGS nid for our MGC name */
574 if (lsi->lsi_flags & LSI_SERVER) {
575 ptr = lsi->lsi_ldd->ldd_params;
576 /* Use mgsnode= nids */
577 if ((class_find_param(ptr, PARAM_MGSNODE, &ptr) == 0) &&
578 (class_parse_nid(ptr, &nid, &ptr) == 0)) {
580 } else if (IS_MGS(lsi->lsi_ldd)) {
581 lnet_process_id_t id;
582 while ((rc = LNetGetId(i++, &id)) != -ENOENT) {
583 if (LNET_NETTYP(LNET_NIDNET(id.nid)) == LOLND)
590 } else { /* client */
591 /* Use nids from mount line: uml1,1@elan:uml2,2@elan:/lustre */
592 ptr = lsi->lsi_lmd->lmd_dev;
593 if (class_parse_nid(ptr, &nid, &ptr) == 0)
597 CERROR("No valid MGS nids found.\n");
601 len = strlen(LUSTRE_MGC_OBDNAME) + strlen(libcfs_nid2str(nid)) + 1;
602 OBD_ALLOC(mgcname, len);
603 OBD_ALLOC(niduuid, len + 2);
604 if (!mgcname || !niduuid)
605 GOTO(out_free, rc = -ENOMEM);
606 sprintf(mgcname, "%s%s", LUSTRE_MGC_OBDNAME, libcfs_nid2str(nid));
608 mutex_down(&mgc_start_lock);
610 obd = class_name2obd(mgcname);
611 if (obd && !obd->obd_stopping) {
612 /* Re-using an existing MGC */
613 atomic_inc(&obd->u.cli.cl_mgc_refcount);
616 /* If we are restarting the MGS, don't try to keep the MGC's
617 old connection, or registration will fail. */
618 if ((lsi->lsi_flags & LSI_SERVER) && IS_MGS(lsi->lsi_ldd)) {
619 CDEBUG(D_MOUNT, "New MGS with live MGC\n");
623 /* Try all connections, but only once (again).
624 We don't want to block another target from starting
625 (using its local copy of the log), but we do want to connect
626 if at all possible. */
628 CDEBUG(D_MOUNT, "%s: Set MGC reconnect %d\n", mgcname,recov_bk);
629 rc = obd_set_info_async(obd->obd_self_export,
630 sizeof(KEY_INIT_RECOV_BACKUP),
631 KEY_INIT_RECOV_BACKUP,
632 sizeof(recov_bk), &recov_bk, NULL);
636 CDEBUG(D_MOUNT, "Start MGC '%s'\n", mgcname);
638 /* Add the primary nids for the MGS */
640 sprintf(niduuid, "%s_%x", mgcname, i);
641 if (lsi->lsi_flags & LSI_SERVER) {
642 ptr = lsi->lsi_ldd->ldd_params;
643 if (IS_MGS(lsi->lsi_ldd)) {
644 /* Use local nids (including LO) */
645 lnet_process_id_t id;
646 while ((rc = LNetGetId(i++, &id)) != -ENOENT) {
647 rc = do_lcfg(mgcname, id.nid,
648 LCFG_ADD_UUID, niduuid, 0,0,0);
651 /* Use mgsnode= nids */
652 if (class_find_param(ptr, PARAM_MGSNODE, &ptr) != 0) {
653 CERROR("No MGS nids given.\n");
654 GOTO(out_free, rc = -EINVAL);
656 while (class_parse_nid(ptr, &nid, &ptr) == 0) {
657 rc = do_lcfg(mgcname, nid,
658 LCFG_ADD_UUID, niduuid, 0,0,0);
662 } else { /* client */
663 /* Use nids from mount line: uml1,1@elan:uml2,2@elan:/lustre */
664 ptr = lsi->lsi_lmd->lmd_dev;
665 while (class_parse_nid(ptr, &nid, &ptr) == 0) {
666 rc = do_lcfg(mgcname, nid,
667 LCFG_ADD_UUID, niduuid, 0,0,0);
669 /* Stop at the first failover nid */
675 CERROR("No valid MGS nids found.\n");
676 GOTO(out_free, rc = -EINVAL);
678 lsi->lsi_lmd->lmd_mgs_failnodes = 1;
680 /* Random uuid for MGC allows easier reconnects */
682 ll_generate_random_uuid(uuidc);
683 class_uuid_unparse(uuidc, uuid);
686 rc = lustre_start_simple(mgcname, LUSTRE_MGC_NAME,
687 (char *)uuid->uuid, LUSTRE_MGS_OBDNAME,
693 /* Add any failover MGS nids */
695 while ((*ptr == ':' ||
696 class_find_param(ptr, PARAM_MGSNODE, &ptr) == 0)) {
697 /* New failover node */
698 sprintf(niduuid, "%s_%x", mgcname, i);
700 while (class_parse_nid(ptr, &nid, &ptr) == 0) {
702 rc = do_lcfg(mgcname, nid,
703 LCFG_ADD_UUID, niduuid, 0,0,0);
708 rc = do_lcfg(mgcname, 0, LCFG_ADD_CONN,
716 lsi->lsi_lmd->lmd_mgs_failnodes = i;
718 obd = class_name2obd(mgcname);
720 CERROR("Can't find mgcobd %s\n", mgcname);
721 GOTO(out_free, rc = -ENOTCONN);
724 /* Keep a refcount of servers/clients who started with "mount",
725 so we know when we can get rid of the mgc. */
726 atomic_set(&obd->u.cli.cl_mgc_refcount, 1);
728 /* Try all connections, but only once. */
730 rc = obd_set_info_async(obd->obd_self_export,
731 sizeof(KEY_INIT_RECOV_BACKUP),
732 KEY_INIT_RECOV_BACKUP,
733 sizeof(recov_bk), &recov_bk, NULL);
736 CWARN("can't set %s %d\n", KEY_INIT_RECOV_BACKUP, rc);
740 GOTO(out, rc = -ENOMEM);
741 data->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_AT |
743 data->ocd_version = LUSTRE_VERSION_CODE;
744 /* We connect to the MGS at setup, and don't disconnect until cleanup */
745 rc = obd_connect(&mgc_conn, obd, &(obd->obd_uuid), data, &exp);
748 CERROR("connect failed %d\n", rc);
751 obd->u.cli.cl_mgc_mgsexp = exp;
754 /* Keep the mgc info in the sb. Note that many lsi's can point
758 mutex_up(&mgc_start_lock);
761 OBD_FREE(mgcname, len);
763 OBD_FREE(niduuid, len + 2);
767 static int lustre_stop_mgc(struct super_block *sb)
769 struct lustre_sb_info *lsi = s2lsi(sb);
770 struct obd_device *obd;
771 char *niduuid = 0, *ptr = 0;
772 int i, rc = 0, len = 0;
782 mutex_down(&mgc_start_lock);
783 if (!atomic_dec_and_test(&obd->u.cli.cl_mgc_refcount)) {
784 /* This is not fatal, every client that stops
785 will call in here. */
786 CDEBUG(D_MOUNT, "mgc still has %d references.\n",
787 atomic_read(&obd->u.cli.cl_mgc_refcount));
788 GOTO(out, rc = -EBUSY);
791 /* The MGC has no recoverable data in any case.
792 * force shotdown set in umount_begin */
793 obd->obd_no_recov = 1;
795 if (obd->u.cli.cl_mgc_mgsexp) {
796 /* An error is not fatal, if we are unable to send the
797 disconnect mgs ping evictor cleans up the export */
798 rc = obd_disconnect(obd->u.cli.cl_mgc_mgsexp);
800 CDEBUG(D_MOUNT, "disconnect failed %d\n", rc);
803 /* Save the obdname for cleaning the nid uuids, which are
805 len = strlen(obd->obd_name) + 6;
806 OBD_ALLOC(niduuid, len);
808 strcpy(niduuid, obd->obd_name);
809 ptr = niduuid + strlen(niduuid);
812 rc = class_manual_cleanup(obd);
816 /* Clean the nid uuids */
818 GOTO(out, rc = -ENOMEM);
820 for (i = 0; i < lsi->lsi_lmd->lmd_mgs_failnodes; i++) {
821 sprintf(ptr, "_%x", i);
822 rc = do_lcfg(LUSTRE_MGC_OBDNAME, 0, LCFG_DEL_UUID,
825 CERROR("del MDC UUID %s failed: rc = %d\n",
830 OBD_FREE(niduuid, len);
832 /* class_import_put will get rid of the additional connections */
833 mutex_up(&mgc_start_lock);
837 /* Since there's only one mgc per node, we have to change it's fs to get
838 access to the right disk. */
839 static int server_mgc_set_fs(struct obd_device *mgc, struct super_block *sb)
841 struct lustre_sb_info *lsi = s2lsi(sb);
845 CDEBUG(D_MOUNT, "Set mgc disk for %s\n", lsi->lsi_lmd->lmd_dev);
847 /* cl_mgc_sem in mgc insures we sleep if the mgc_fs is busy */
848 rc = obd_set_info_async(mgc->obd_self_export,
849 sizeof(KEY_SET_FS), KEY_SET_FS,
850 sizeof(*sb), sb, NULL);
852 CERROR("can't set_fs %d\n", rc);
858 static int server_mgc_clear_fs(struct obd_device *mgc)
863 CDEBUG(D_MOUNT, "Unassign mgc disk\n");
865 rc = obd_set_info_async(mgc->obd_self_export,
866 sizeof(KEY_CLEAR_FS), KEY_CLEAR_FS,
871 DECLARE_MUTEX(server_start_lock);
873 /* Stop MDS/OSS if nobody is using them */
874 static int server_stop_servers(int lddflags, int lsiflags)
876 struct obd_device *obd = NULL;
877 struct obd_type *type = NULL;
881 mutex_down(&server_start_lock);
883 /* Either an MDT or an OST or neither */
884 /* if this was an MDT, and there are no more MDT's, clean up the MDS */
885 if ((lddflags & LDD_F_SV_TYPE_MDT) &&
886 (obd = class_name2obd(LUSTRE_MDS_OBDNAME))) {
887 /*FIXME pre-rename, should eventually be LUSTRE_MDT_NAME*/
888 type = class_search_type(LUSTRE_MDS_NAME);
890 /* if this was an OST, and there are no more OST's, clean up the OSS */
891 if ((lddflags & LDD_F_SV_TYPE_OST) &&
892 (obd = class_name2obd(LUSTRE_OSS_OBDNAME))) {
893 type = class_search_type(LUSTRE_OST_NAME);
896 if (obd && (!type || !type->typ_refcnt)) {
899 /* obd_fail doesn't mean much on a server obd */
900 err = class_manual_cleanup(obd);
905 mutex_up(&server_start_lock);
910 int server_mti_print(char *title, struct mgs_target_info *mti)
912 PRINT_CMD(PRINT_MASK, "mti %s\n", title);
913 PRINT_CMD(PRINT_MASK, "server: %s\n", mti->mti_svname);
914 PRINT_CMD(PRINT_MASK, "fs: %s\n", mti->mti_fsname);
915 PRINT_CMD(PRINT_MASK, "uuid: %s\n", mti->mti_uuid);
916 PRINT_CMD(PRINT_MASK, "ver: %d flags: %#x\n",
917 mti->mti_config_ver, mti->mti_flags);
921 static int server_sb2mti(struct super_block *sb, struct mgs_target_info *mti)
923 struct lustre_sb_info *lsi = s2lsi(sb);
924 struct lustre_disk_data *ldd = lsi->lsi_ldd;
925 lnet_process_id_t id;
929 if (!(lsi->lsi_flags & LSI_SERVER))
932 strncpy(mti->mti_fsname, ldd->ldd_fsname,
933 sizeof(mti->mti_fsname));
934 strncpy(mti->mti_svname, ldd->ldd_svname,
935 sizeof(mti->mti_svname));
937 mti->mti_nid_count = 0;
938 while (LNetGetId(i++, &id) != -ENOENT) {
939 if (LNET_NETTYP(LNET_NIDNET(id.nid)) == LOLND)
941 mti->mti_nids[mti->mti_nid_count] = id.nid;
942 mti->mti_nid_count++;
943 if (mti->mti_nid_count >= MTI_NIDS_MAX) {
944 CWARN("Only using first %d nids for %s\n",
945 mti->mti_nid_count, mti->mti_svname);
950 mti->mti_lustre_ver = LUSTRE_VERSION_CODE;
951 mti->mti_config_ver = 0;
952 mti->mti_flags = ldd->ldd_flags;
953 mti->mti_stripe_index = ldd->ldd_svindex;
954 memcpy(mti->mti_uuid, ldd->ldd_uuid, sizeof(mti->mti_uuid));
955 if (strlen(ldd->ldd_params) > sizeof(mti->mti_params)) {
956 CERROR("params too big for mti\n");
959 memcpy(mti->mti_params, ldd->ldd_params, sizeof(mti->mti_params));
963 /* Register an old or new target with the MGS. If needed MGS will construct
964 startup logs and assign index */
965 int server_register_target(struct super_block *sb)
967 struct lustre_sb_info *lsi = s2lsi(sb);
968 struct obd_device *mgc = lsi->lsi_mgc;
969 struct lustre_disk_data *ldd = lsi->lsi_ldd;
970 struct mgs_target_info *mti = NULL;
976 if (!(lsi->lsi_flags & LSI_SERVER))
982 rc = server_sb2mti(sb, mti);
986 CDEBUG(D_MOUNT, "Registration %s, fs=%s, %s, index=%04x, flags=%#x\n",
987 mti->mti_svname, mti->mti_fsname,
988 libcfs_nid2str(mti->mti_nids[0]), mti->mti_stripe_index,
991 /* Register the target */
992 /* FIXME use mgc_process_config instead */
993 rc = obd_set_info_async(mgc->u.cli.cl_mgc_mgsexp,
994 sizeof(KEY_REGISTER_TARGET), KEY_REGISTER_TARGET,
995 sizeof(*mti), mti, NULL);
999 /* Always update our flags */
1000 ldd->ldd_flags = mti->mti_flags & ~LDD_F_REWRITE_LDD;
1002 /* If this flag is set, it means the MGS wants us to change our
1003 on-disk data. (So far this means just the index.) */
1004 if (mti->mti_flags & LDD_F_REWRITE_LDD) {
1007 CDEBUG(D_MOUNT, "Changing on-disk index from %#x to %#x "
1008 "for %s\n", ldd->ldd_svindex, mti->mti_stripe_index,
1010 ldd->ldd_svindex = mti->mti_stripe_index;
1011 strncpy(ldd->ldd_svname, mti->mti_svname,
1012 sizeof(ldd->ldd_svname));
1013 /* or ldd_make_sv_name(ldd); */
1014 ldd_write(&mgc->obd_lvfs_ctxt, ldd);
1016 err = fsfilt_set_label(mgc, lsi->lsi_srv_mnt->mnt_sb,
1019 CERROR("Label set error %d\n", err);
1020 label = fsfilt_get_label(mgc, lsi->lsi_srv_mnt->mnt_sb);
1022 CDEBUG(D_MOUNT, "Disk label changed to %s\n", label);
1024 /* Flush the new ldd to disk */
1025 fsfilt_sync(mgc, lsi->lsi_srv_mnt->mnt_sb);
1035 static int server_start_targets(struct super_block *sb, struct vfsmount *mnt)
1037 struct obd_device *obd;
1038 struct lustre_sb_info *lsi = s2lsi(sb);
1039 struct config_llog_instance cfg;
1043 CDEBUG(D_MOUNT, "starting target %s\n", lsi->lsi_ldd->ldd_svname);
1045 /* If we're an MDT, make sure the global MDS is running */
1046 if (lsi->lsi_ldd->ldd_flags & LDD_F_SV_TYPE_MDT) {
1047 /* make sure the MDS is started */
1048 mutex_down(&server_start_lock);
1049 obd = class_name2obd(LUSTRE_MDS_OBDNAME);
1051 rc = lustre_start_simple(LUSTRE_MDS_OBDNAME,
1052 /* FIXME pre-rename, should eventually be LUSTRE_MDS_NAME */
1054 LUSTRE_MDS_OBDNAME"_uuid",
1057 mutex_up(&server_start_lock);
1058 CERROR("failed to start MDS: %d\n", rc);
1062 mutex_up(&server_start_lock);
1065 /* If we're an OST, make sure the global OSS is running */
1066 if (lsi->lsi_ldd->ldd_flags & LDD_F_SV_TYPE_OST) {
1067 /* make sure OSS is started */
1068 mutex_down(&server_start_lock);
1069 obd = class_name2obd(LUSTRE_OSS_OBDNAME);
1071 rc = lustre_start_simple(LUSTRE_OSS_OBDNAME,
1073 LUSTRE_OSS_OBDNAME"_uuid",
1076 mutex_up(&server_start_lock);
1077 CERROR("failed to start OSS: %d\n", rc);
1081 mutex_up(&server_start_lock);
1084 /* Set the mgc fs to our server disk. This allows the MGC
1085 to read and write configs locally. */
1086 rc = server_mgc_set_fs(lsi->lsi_mgc, sb);
1090 /* Register with MGS */
1091 rc = server_register_target(sb);
1092 if (rc && (lsi->lsi_ldd->ldd_flags &
1093 (LDD_F_NEED_INDEX | LDD_F_UPDATE | LDD_F_UPGRADE14))){
1094 CERROR("Required registration failed for %s: %d\n",
1095 lsi->lsi_ldd->ldd_svname, rc);
1097 LCONSOLE_ERROR_MSG(0x15f, "Communication error with the"
1098 " MGS. Is the MGS running?\n");
1102 if (rc == -EINVAL) {
1103 LCONSOLE_ERROR_MSG(0x160, "The MGS is refusing to allow this "
1104 "server (%s) to start. Please see messages "
1105 "on the MGS node.\n",
1106 lsi->lsi_ldd->ldd_svname);
1109 /* non-fatal error of registeration with MGS */
1111 CDEBUG(D_MOUNT, "Cannot register with MGS: %d\n", rc);
1113 /* Let the target look up the mount using the target's name
1114 (we can't pass the sb or mnt through class_process_config.) */
1115 rc = server_register_mount(lsi->lsi_ldd->ldd_svname, sb, mnt);
1119 /* Start targets using the llog named for the target */
1120 memset(&cfg, 0, sizeof(cfg));
1121 rc = lustre_process_log(sb, lsi->lsi_ldd->ldd_svname, &cfg);
1123 CERROR("failed to start server %s: %d\n",
1124 lsi->lsi_ldd->ldd_svname, rc);
1125 /* Do NOT call server_deregister_mount() here. This makes it
1126 * impossible to find mount later in cleanup time and leaves
1127 * @lsi and othder stuff leaked. -umka */
1132 /* Release the mgc fs for others to use */
1133 server_mgc_clear_fs(lsi->lsi_mgc);
1136 obd = class_name2obd(lsi->lsi_ldd->ldd_svname);
1138 CERROR("no server named %s was started\n",
1139 lsi->lsi_ldd->ldd_svname);
1143 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_ABORT_RECOV) &&
1144 (OBP(obd, iocontrol))) {
1145 obd_iocontrol(OBD_IOC_ABORT_RECOVERY,
1146 obd->obd_self_export, 0, NULL, NULL);
1149 /* log has been fully processed */
1150 obd_notify(obd, NULL, OBD_NOTIFY_CONFIG, (void *)CONFIG_LOG);
1156 /***************** lustre superblock **************/
1158 struct lustre_sb_info *lustre_init_lsi(struct super_block *sb)
1160 struct lustre_sb_info *lsi = NULL;
1163 OBD_ALLOC(lsi, sizeof(*lsi));
1166 OBD_ALLOC(lsi->lsi_lmd, sizeof(*lsi->lsi_lmd));
1167 if (!lsi->lsi_lmd) {
1168 OBD_FREE(lsi, sizeof(*lsi));
1172 lsi->lsi_lmd->lmd_exclude_count = 0;
1173 lsi->lsi_lmd->lmd_recovery_time_soft = 0;
1174 lsi->lsi_lmd->lmd_recovery_time_hard = 0;
1175 s2lsi_nocast(sb) = lsi;
1176 /* we take 1 extra ref for our setup */
1177 atomic_set(&lsi->lsi_mounts, 1);
1179 /* Default umount style */
1180 lsi->lsi_flags = LSI_UMOUNT_FAILOVER;
1184 static int lustre_free_lsi(struct super_block *sb)
1186 struct lustre_sb_info *lsi = s2lsi(sb);
1189 LASSERT(lsi != NULL);
1190 CDEBUG(D_MOUNT, "Freeing lsi %p\n", lsi);
1192 /* someone didn't call server_put_mount. */
1193 LASSERT(atomic_read(&lsi->lsi_mounts) == 0);
1195 if (lsi->lsi_ldd != NULL)
1196 OBD_FREE(lsi->lsi_ldd, sizeof(*lsi->lsi_ldd));
1198 if (lsi->lsi_lmd != NULL) {
1199 if (lsi->lsi_lmd->lmd_dev != NULL)
1200 OBD_FREE(lsi->lsi_lmd->lmd_dev,
1201 strlen(lsi->lsi_lmd->lmd_dev) + 1);
1202 if (lsi->lsi_lmd->lmd_profile != NULL)
1203 OBD_FREE(lsi->lsi_lmd->lmd_profile,
1204 strlen(lsi->lsi_lmd->lmd_profile) + 1);
1205 if (lsi->lsi_lmd->lmd_opts != NULL)
1206 OBD_FREE(lsi->lsi_lmd->lmd_opts,
1207 strlen(lsi->lsi_lmd->lmd_opts) + 1);
1208 if (lsi->lsi_lmd->lmd_exclude_count)
1209 OBD_FREE(lsi->lsi_lmd->lmd_exclude,
1210 sizeof(lsi->lsi_lmd->lmd_exclude[0]) *
1211 lsi->lsi_lmd->lmd_exclude_count);
1212 OBD_FREE(lsi->lsi_lmd, sizeof(*lsi->lsi_lmd));
1215 LASSERT(lsi->lsi_llsbi == NULL);
1217 OBD_FREE(lsi, sizeof(*lsi));
1218 s2lsi_nocast(sb) = NULL;
1223 /* The lsi has one reference for every server that is using the disk -
1224 e.g. MDT, MGS, and potentially MGC */
1225 static int lustre_put_lsi(struct super_block *sb)
1227 struct lustre_sb_info *lsi = s2lsi(sb);
1230 LASSERT(lsi != NULL);
1232 CDEBUG(D_MOUNT, "put %p %d\n", sb, atomic_read(&lsi->lsi_mounts));
1233 if (atomic_dec_and_test(&lsi->lsi_mounts)) {
1234 lustre_free_lsi(sb);
1240 /*************** server mount ******************/
1242 /* Kernel mount using mount options in MOUNT_DATA_FILE */
1243 static struct vfsmount *server_kernel_mount(struct super_block *sb)
1245 struct lvfs_run_ctxt mount_ctxt;
1246 struct lustre_sb_info *lsi = s2lsi(sb);
1247 struct lustre_disk_data *ldd;
1248 struct lustre_mount_data *lmd = lsi->lsi_lmd;
1249 struct vfsmount *mnt;
1250 char *options = NULL;
1251 unsigned long page, s_flags;
1252 struct page *__page;
1256 OBD_ALLOC(ldd, sizeof(*ldd));
1258 RETURN(ERR_PTR(-ENOMEM));
1260 /* In the past, we have always used flags = 0.
1261 Note ext3/ldiskfs can't be mounted ro. */
1262 s_flags = sb->s_flags;
1264 /* allocate memory for options */
1265 OBD_PAGE_ALLOC(__page, CFS_ALLOC_STD);
1267 GOTO(out_free, rc = -ENOMEM);
1268 page = (unsigned long)cfs_page_address(__page);
1269 options = (char *)page;
1270 memset(options, 0, CFS_PAGE_SIZE);
1272 /* mount-line options must be added for pre-mount because it may
1273 * contain mount options such as journal_dev which are required
1274 * to mount successfuly the underlying filesystem */
1275 if (lmd->lmd_opts && (*(lmd->lmd_opts) != 0))
1276 strncat(options, lmd->lmd_opts, CFS_PAGE_SIZE - 1);
1278 /* Pre-mount ldiskfs to read the MOUNT_DATA_FILE */
1279 CDEBUG(D_MOUNT, "Pre-mount ldiskfs %s\n", lmd->lmd_dev);
1280 mnt = ll_kern_mount("ldiskfs", s_flags, lmd->lmd_dev, (void *)options);
1283 /* 2.6 kernels: if ldiskfs fails, try ldiskfs2 */
1284 mnt = ll_kern_mount("ldiskfs2", s_flags, lmd->lmd_dev,
1287 int rc2 = PTR_ERR(mnt);
1288 CERROR("premount %s:%#lx ldiskfs failed: %d, ldiskfs2 "
1289 "failed: %d. Is the ldiskfs module available?\n",
1290 lmd->lmd_dev, s_flags, rc, rc2);
1295 OBD_SET_CTXT_MAGIC(&mount_ctxt);
1296 mount_ctxt.pwdmnt = mnt;
1297 mount_ctxt.pwd = mnt->mnt_root;
1298 mount_ctxt.fs = get_ds();
1300 rc = ldd_parse(&mount_ctxt, ldd);
1304 CERROR("premount parse options failed: rc = %d\n", rc);
1308 /* Done with our pre-mount, now do the real mount. */
1310 /* Glom up mount options */
1311 memset(options, 0, CFS_PAGE_SIZE);
1312 strncpy(options, ldd->ldd_mount_opts, CFS_PAGE_SIZE - 2);
1314 /* Add in any mount-line options */
1315 if (lmd->lmd_opts && (*(lmd->lmd_opts) != 0)) {
1316 int len = CFS_PAGE_SIZE - strlen(options) - 2;
1318 strcat(options, ",");
1319 strncat(options, lmd->lmd_opts, len);
1322 /* Special permanent mount flags */
1324 s_flags |= MS_NOATIME | MS_NODIRATIME;
1326 CDEBUG(D_MOUNT, "kern_mount: %s %s %s\n",
1327 MT_STR(ldd), lmd->lmd_dev, options);
1328 mnt = ll_kern_mount(MT_STR(ldd), s_flags, lmd->lmd_dev,
1332 CERROR("ll_kern_mount failed: rc = %d\n", rc);
1336 if (lmd->lmd_flags & LMD_FLG_ABORT_RECOV)
1337 simple_truncate(mnt->mnt_sb->s_root, mnt, LAST_RCVD,
1340 OBD_PAGE_FREE(__page);
1341 lsi->lsi_ldd = ldd; /* freed at lsi cleanup */
1342 CDEBUG(D_SUPER, "%s: mnt = %p\n", lmd->lmd_dev, mnt);
1347 OBD_PAGE_FREE(__page);
1348 OBD_FREE(ldd, sizeof(*ldd));
1349 lsi->lsi_ldd = NULL;
1350 RETURN(ERR_PTR(rc));
1353 /* Wait here forever until the mount refcount is 0 before completing umount,
1354 * else we risk dereferencing a null pointer.
1355 * LNET may take e.g. 165s before killing zombies.
1357 static void server_wait_finished(struct vfsmount *mnt)
1361 cfs_sigset_t blocked;
1363 cfs_waitq_init(&waitq);
1365 while (cfs_atomic_read(&mnt->mnt_count) > 1) {
1366 if (waited && (waited % 30 == 0))
1367 LCONSOLE_WARN("Mount still busy with %d refs after "
1369 atomic_read(&mnt->mnt_count),
1371 /* Cannot use l_event_wait() for an interruptible sleep. */
1373 blocked = l_w_e_set_sigs(sigmask(SIGKILL));
1374 rc = cfs_waitq_wait_event_interruptible_timeout(
1376 (cfs_atomic_read(&mnt->mnt_count) == 1),
1377 cfs_time_seconds(3));
1378 cfs_block_sigs(blocked);
1380 LCONSOLE_EMERG("Danger: interrupted umount %p with "
1382 mnt, atomic_read(&mnt->mnt_count));
1388 static void server_put_super(struct super_block *sb)
1390 struct lustre_sb_info *lsi = s2lsi(sb);
1391 struct obd_device *obd;
1392 struct vfsmount *mnt = lsi->lsi_srv_mnt;
1393 char *tmpname, *extraname = NULL;
1395 int lddflags = lsi->lsi_ldd->ldd_flags;
1396 int lsiflags = lsi->lsi_flags;
1400 LASSERT(lsiflags & LSI_SERVER);
1402 tmpname_sz = strlen(lsi->lsi_ldd->ldd_svname) + 1;
1403 OBD_ALLOC(tmpname, tmpname_sz);
1404 memcpy(tmpname, lsi->lsi_ldd->ldd_svname, tmpname_sz);
1405 CDEBUG(D_MOUNT, "server put_super %s\n", tmpname);
1406 if (IS_MDT(lsi->lsi_ldd) && (lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC))
1407 snprintf(tmpname, tmpname_sz, "MGS");
1409 /* Stop the target */
1410 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1411 (IS_MDT(lsi->lsi_ldd) || IS_OST(lsi->lsi_ldd))) {
1412 struct lustre_profile *lprof = NULL;
1414 /* tell the mgc to drop the config log */
1415 lustre_end_log(sb, lsi->lsi_ldd->ldd_svname, NULL);
1417 /* COMPAT_146 - profile may get deleted in mgc_cleanup.
1418 If there are any setup/cleanup errors, save the lov
1419 name for safety cleanup later. */
1420 lprof = class_get_profile(lsi->lsi_ldd->ldd_svname);
1421 if (lprof && lprof->lp_osc) {
1422 OBD_ALLOC(extraname, strlen(lprof->lp_osc) + 1);
1423 strcpy(extraname, lprof->lp_osc);
1426 obd = class_name2obd(lsi->lsi_ldd->ldd_svname);
1428 CDEBUG(D_MOUNT, "stopping %s\n", obd->obd_name);
1429 if (lsi->lsi_flags & LSI_UMOUNT_FAILOVER)
1431 /* We can't seem to give an error return code
1432 * to .put_super, so we better make sure we clean up! */
1434 class_manual_cleanup(obd);
1436 CERROR("no obd %s\n", lsi->lsi_ldd->ldd_svname);
1437 server_deregister_mount(lsi->lsi_ldd->ldd_svname);
1442 /* If they wanted the mgs to stop separately from the mdt, they
1443 should have put it on a different device. */
1444 if (IS_MGS(lsi->lsi_ldd)) {
1445 /* stop the mgc before the mgs so the connection gets cleaned
1447 lustre_stop_mgc(sb);
1448 /* if MDS start with --nomgs, don't stop MGS then */
1449 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS))
1450 server_stop_mgs(sb);
1453 /* Clean the mgc and sb */
1454 rc = lustre_common_put_super(sb);
1455 /* FIXME how can I report a failure to umount? */
1457 /* Wait for the targets to really clean up - can't exit (and let the
1458 sb get destroyed) while the mount is still in use */
1459 server_wait_finished(mnt);
1461 /* drop the One True Mount */
1464 /* Stop the servers (MDS, OSS) if no longer needed. We must wait
1465 until the target is really gone so that our type refcount check
1467 server_stop_servers(lddflags, lsiflags);
1469 /* In case of startup or cleanup err, stop related obds */
1471 obd = class_name2obd(extraname);
1473 CWARN("Cleaning orphaned obd %s\n", extraname);
1475 class_manual_cleanup(obd);
1477 OBD_FREE(extraname, strlen(extraname) + 1);
1480 LCONSOLE_WARN("server umount %s complete\n", tmpname);
1481 OBD_FREE(tmpname, tmpname_sz);
1484 #ifdef HAVE_UMOUNTBEGIN_VFSMOUNT
1485 static void server_umount_begin(struct vfsmount *vfsmnt, int flags)
1487 struct super_block *sb = vfsmnt->mnt_sb;
1489 static void server_umount_begin(struct super_block *sb)
1492 struct lustre_sb_info *lsi = s2lsi(sb);
1495 #ifdef HAVE_UMOUNTBEGIN_VFSMOUNT
1496 if (!(flags & MNT_FORCE)) {
1502 CDEBUG(D_MOUNT, "umount -f\n");
1503 /* umount = failover
1505 no third way to do non-force, non-failover */
1506 lsi->lsi_flags &= ~LSI_UMOUNT_FAILOVER;
1507 lsi->lsi_flags |= LSI_UMOUNT_FORCE;
1511 #ifndef HAVE_STATFS_DENTRY_PARAM
1512 static int server_statfs (struct super_block *sb, struct kstatfs *buf)
1515 static int server_statfs (struct dentry *dentry, struct kstatfs *buf)
1517 struct super_block *sb = dentry->d_sb;
1519 struct vfsmount *mnt = s2lsi(sb)->lsi_srv_mnt;
1522 if (mnt && mnt->mnt_sb && mnt->mnt_sb->s_op->statfs) {
1523 #ifdef HAVE_STATFS_DENTRY_PARAM
1524 int rc = mnt->mnt_sb->s_op->statfs(mnt->mnt_root, buf);
1526 int rc = mnt->mnt_sb->s_op->statfs(mnt->mnt_sb, buf);
1529 buf->f_type = sb->s_magic;
1535 buf->f_type = sb->s_magic;
1536 buf->f_bsize = sb->s_blocksize;
1542 buf->f_namelen = NAME_MAX;
1546 static int server_show_options(struct seq_file *seq, struct vfsmount *vfsmnt)
1548 struct vfsmount *mnt = s2lsi(vfsmnt->mnt_sb)->lsi_srv_mnt;
1551 if (mnt && mnt->mnt_sb && mnt->mnt_sb->s_op->show_options) {
1552 int rc = mnt->mnt_sb->s_op->show_options(seq, mnt);
1558 static struct super_operations server_ops =
1560 .put_super = server_put_super,
1561 .umount_begin = server_umount_begin, /* umount -f */
1562 .statfs = server_statfs,
1563 .show_options = server_show_options,
1566 #define log2(n) ffz(~(n))
1567 #define LUSTRE_SUPER_MAGIC 0x0BD00BD1
1569 static int server_fill_super_common(struct super_block *sb)
1571 struct inode *root = 0;
1574 CDEBUG(D_MOUNT, "Server sb, dev=%d\n", (int)sb->s_dev);
1576 sb->s_blocksize = 4096;
1577 sb->s_blocksize_bits = log2(sb->s_blocksize);
1578 sb->s_magic = LUSTRE_SUPER_MAGIC;
1579 sb->s_maxbytes = 0; //PAGE_CACHE_MAXBYTES;
1580 sb->s_flags |= MS_RDONLY;
1581 sb->s_op = &server_ops;
1583 root = new_inode(sb);
1585 CERROR("Can't make root inode\n");
1589 /* returns -EIO for every operation */
1590 /* make_bad_inode(root); -- badness - can't umount */
1591 /* apparently we need to be a directory for the mount to finish */
1592 root->i_mode = S_IFDIR;
1594 sb->s_root = d_alloc_root(root);
1596 CERROR("Can't make root dentry\n");
1604 static int server_fill_super(struct super_block *sb)
1606 struct lustre_sb_info *lsi = s2lsi(sb);
1607 struct vfsmount *mnt;
1611 /* the One True Mount */
1612 mnt = server_kernel_mount(sb);
1615 CERROR("Unable to mount device %s: %d\n",
1616 lsi->lsi_lmd->lmd_dev, rc);
1620 lsi->lsi_srv_mnt = mnt;
1622 LASSERT(lsi->lsi_ldd);
1623 CDEBUG(D_MOUNT, "Found service %s for fs '%s' on device %s\n",
1624 lsi->lsi_ldd->ldd_svname, lsi->lsi_ldd->ldd_fsname,
1625 lsi->lsi_lmd->lmd_dev);
1627 if (class_name2obd(lsi->lsi_ldd->ldd_svname)) {
1628 LCONSOLE_ERROR_MSG(0x161, "The target named %s is already "
1629 "running. Double-mount may have compromised "
1630 "the disk journal.\n",
1631 lsi->lsi_ldd->ldd_svname);
1637 /* Start MGS before MGC */
1638 if (IS_MGS(lsi->lsi_ldd) && !(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS)) {
1639 rc = server_start_mgs(sb);
1644 rc = lustre_start_mgc(sb);
1648 /* Set up all obd devices for service */
1649 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1650 (IS_OST(lsi->lsi_ldd) || IS_MDT(lsi->lsi_ldd))) {
1651 rc = server_start_targets(sb, mnt);
1653 CERROR("Unable to start targets: %d\n", rc);
1656 /* FIXME overmount client here,
1657 or can we just start a client log and client_fill_super on this sb?
1658 We need to make sure server_put_super gets called too - ll_put_super
1659 calls lustre_common_put_super; check there for LSI_SERVER flag,
1661 Probably should start client from new thread so we can return.
1662 Client will not finish until all servers are connected.
1663 Note - MGS-only server does NOT get a client, since there is no
1664 lustre fs associated - the MGS is for all lustre fs's */
1667 rc = server_fill_super_common(sb);
1671 LCONSOLE_WARN("Server %s on device %s has started\n",
1672 ((lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1673 (IS_MDT(lsi->lsi_ldd))) ? "MGS" : lsi->lsi_ldd->ldd_svname,
1674 lsi->lsi_lmd->lmd_dev);
1678 /* We jump here in case of failure while starting targets or MGS.
1679 * In this case we can't just put @mnt and have to do real cleanup
1680 * with stoping targets, etc. */
1681 server_put_super(sb);
1685 /* Get the index from the obd name.
1686 rc = server type, or
1688 if endptr isn't NULL it is set to end of name */
1689 int server_name2index(char *svname, __u32 *idx, char **endptr)
1691 unsigned long index;
1693 char *dash = strrchr(svname, '-');
1697 if (strncmp(dash + 1, "MDT", 3) == 0)
1698 rc = LDD_F_SV_TYPE_MDT;
1699 else if (strncmp(dash + 1, "OST", 3) == 0)
1700 rc = LDD_F_SV_TYPE_OST;
1704 index = simple_strtoul(dash + 4, endptr, 16);
1709 /*************** mount common betweeen server and client ***************/
1712 int lustre_common_put_super(struct super_block *sb)
1717 CDEBUG(D_MOUNT, "dropping sb %p\n", sb);
1719 /* Drop a ref to the MGC */
1720 rc = lustre_stop_mgc(sb);
1721 if (rc && (rc != -ENOENT)) {
1723 CERROR("Can't stop MGC: %d\n", rc);
1726 /* BUSY just means that there's some other obd that
1727 needs the mgc. Let him clean it up. */
1728 CDEBUG(D_MOUNT, "MGC still in use\n");
1730 /* Drop a ref to the mounted disk */
1735 static void lmd_print(struct lustre_mount_data *lmd)
1739 PRINT_CMD(PRINT_MASK, " mount data:\n");
1740 if (lmd_is_client(lmd))
1741 PRINT_CMD(PRINT_MASK, "profile: %s\n", lmd->lmd_profile);
1742 PRINT_CMD(PRINT_MASK, "device: %s\n", lmd->lmd_dev);
1743 PRINT_CMD(PRINT_MASK, "flags: %x\n", lmd->lmd_flags);
1746 PRINT_CMD(PRINT_MASK, "options: %s\n", lmd->lmd_opts);
1748 if (lmd->lmd_recovery_time_soft)
1749 PRINT_CMD(PRINT_MASK, "recovery time soft: %d\n",
1750 lmd->lmd_recovery_time_soft);
1752 if (lmd->lmd_recovery_time_hard)
1753 PRINT_CMD(PRINT_MASK, "recovery time hard: %d\n",
1754 lmd->lmd_recovery_time_hard);
1756 for (i = 0; i < lmd->lmd_exclude_count; i++) {
1757 PRINT_CMD(PRINT_MASK, "exclude %d: OST%04x\n", i,
1758 lmd->lmd_exclude[i]);
1762 /* Is this server on the exclusion list */
1763 int lustre_check_exclusion(struct super_block *sb, char *svname)
1765 struct lustre_sb_info *lsi = s2lsi(sb);
1766 struct lustre_mount_data *lmd = lsi->lsi_lmd;
1771 rc = server_name2index(svname, &index, NULL);
1772 if (rc != LDD_F_SV_TYPE_OST)
1773 /* Only exclude OSTs */
1776 CDEBUG(D_MOUNT, "Check exclusion %s (%d) in %d of %s\n", svname,
1777 index, lmd->lmd_exclude_count, lmd->lmd_dev);
1779 for(i = 0; i < lmd->lmd_exclude_count; i++) {
1780 if (index == lmd->lmd_exclude[i]) {
1781 CWARN("Excluding %s (on exclusion list)\n", svname);
1788 /* mount -v -o exclude=lustre-OST0001:lustre-OST0002 -t lustre ... */
1789 static int lmd_make_exclusion(struct lustre_mount_data *lmd, char *ptr)
1791 char *s1 = ptr, *s2;
1792 __u32 index, *exclude_list;
1796 /* The shortest an ost name can be is 8 chars: -OST0000.
1797 We don't actually know the fsname at this time, so in fact
1798 a user could specify any fsname. */
1799 devmax = strlen(ptr) / 8 + 1;
1801 /* temp storage until we figure out how many we have */
1802 OBD_ALLOC(exclude_list, sizeof(index) * devmax);
1806 /* we enter this fn pointing at the '=' */
1807 while (*s1 && *s1 != ' ' && *s1 != ',') {
1809 rc = server_name2index(s1, &index, &s2);
1811 CERROR("Can't parse server name '%s'\n", s1);
1814 if (rc == LDD_F_SV_TYPE_OST)
1815 exclude_list[lmd->lmd_exclude_count++] = index;
1817 CDEBUG(D_MOUNT, "ignoring exclude %.7s\n", s1);
1819 /* now we are pointing at ':' (next exclude)
1820 or ',' (end of excludes) */
1822 if (lmd->lmd_exclude_count >= devmax)
1825 if (rc >= 0) /* non-err */
1828 if (lmd->lmd_exclude_count) {
1829 /* permanent, freed in lustre_free_lsi */
1830 OBD_ALLOC(lmd->lmd_exclude, sizeof(index) *
1831 lmd->lmd_exclude_count);
1832 if (lmd->lmd_exclude) {
1833 memcpy(lmd->lmd_exclude, exclude_list,
1834 sizeof(index) * lmd->lmd_exclude_count);
1837 lmd->lmd_exclude_count = 0;
1840 OBD_FREE(exclude_list, sizeof(index) * devmax);
1844 /* mount -v -t lustre uml1:uml2:/lustre-client /mnt/lustre */
1845 static int lmd_parse(char *options, struct lustre_mount_data *lmd)
1847 char *s1, *s2, *devname = NULL;
1848 struct lustre_mount_data *raw = (struct lustre_mount_data *)options;
1854 LCONSOLE_ERROR_MSG(0x162, "Missing mount data: check that "
1855 "/sbin/mount.lustre is installed.\n");
1859 /* Options should be a string - try to detect old lmd data */
1860 if ((raw->lmd_magic & 0xffffff00) == (LMD_MAGIC & 0xffffff00)) {
1861 LCONSOLE_ERROR_MSG(0x163, "You're using an old version of "
1862 "/sbin/mount.lustre. Please install "
1863 "version %s\n", LUSTRE_VERSION_STRING);
1866 lmd->lmd_magic = LMD_MAGIC;
1868 /* Set default flags here */
1873 int time_min = 2 * (CONNECTION_SWITCH_MAX +
1874 2 * INITIAL_CONNECT_TIMEOUT);
1876 /* Skip whitespace and extra commas */
1877 while (*s1 == ' ' || *s1 == ',')
1880 /* Client options are parsed in ll_options: eg. flock,
1883 /* Parse non-ldiskfs options here. Rather than modifying
1884 ldiskfs, we just zero these out here */
1885 if (strncmp(s1, "abort_recov", 11) == 0) {
1886 lmd->lmd_flags |= LMD_FLG_ABORT_RECOV;
1888 } else if (strncmp(s1, "recovery_time_soft=", 19) == 0) {
1889 lmd->lmd_recovery_time_soft = max_t(int,
1890 simple_strtoul(s1 + 19, NULL, 10), time_min);
1892 } else if (strncmp(s1, "recovery_time_hard=", 19) == 0) {
1893 lmd->lmd_recovery_time_hard = max_t(int,
1894 simple_strtoul(s1 + 19, NULL, 10), time_min);
1896 } else if (strncmp(s1, "nosvc", 5) == 0) {
1897 lmd->lmd_flags |= LMD_FLG_NOSVC;
1899 } else if (strncmp(s1, "nomgs", 5) == 0) {
1900 lmd->lmd_flags |= LMD_FLG_NOMGS;
1902 /* ost exclusion list */
1903 } else if (strncmp(s1, "exclude=", 8) == 0) {
1904 rc = lmd_make_exclusion(lmd, s1 + 7);
1910 /* Linux 2.4 doesn't pass the device, so we stuck it at the
1911 end of the options. */
1912 else if (strncmp(s1, "device=", 7) == 0) {
1914 /* terminate options right before device. device
1915 must be the last one. */
1918 } else if (strncmp(s1, "loop=", 5) == 0) {
1924 s2 = strchr(s1, ',');
1932 memmove(s1, s2, strlen(s2) + 1);
1938 LCONSOLE_ERROR_MSG(0x164, "Can't find the device name "
1939 "(need mount option 'device=...')\n");
1943 s1 = strstr(devname, ":/");
1946 lmd->lmd_flags = LMD_FLG_CLIENT;
1947 /* Remove leading /s from fsname */
1948 while (*++s1 == '/') ;
1949 /* Freed in lustre_free_lsi */
1950 OBD_ALLOC(lmd->lmd_profile, strlen(s1) + 8);
1951 if (!lmd->lmd_profile)
1953 sprintf(lmd->lmd_profile, "%s-client", s1);
1956 /* Freed in lustre_free_lsi */
1957 OBD_ALLOC(lmd->lmd_dev, strlen(devname) + 1);
1960 strcpy(lmd->lmd_dev, devname);
1962 /* Save mount options */
1963 s1 = options + strlen(options) - 1;
1964 while (s1 >= options && (*s1 == ',' || *s1 == ' '))
1966 if (*options != 0) {
1967 /* Freed in lustre_free_lsi */
1968 OBD_ALLOC(lmd->lmd_opts, strlen(options) + 1);
1971 strcpy(lmd->lmd_opts, options);
1974 lmd->lmd_magic = LMD_MAGIC;
1980 CERROR("Bad mount options %s\n", options);
1986 int lustre_fill_super(struct super_block *sb, void *data, int silent)
1988 struct lustre_mount_data *lmd;
1989 struct lustre_sb_info *lsi;
1993 CDEBUG(D_MOUNT|D_VFSTRACE, "VFS Op: sb %p\n", sb);
1995 lsi = lustre_init_lsi(sb);
2000 /* Figure out the lmd from the mount options */
2001 if (lmd_parse((char *)data, lmd)) {
2003 GOTO(out, rc = -EINVAL);
2006 if (lmd_is_client(lmd)) {
2007 CDEBUG(D_MOUNT, "Mounting client %s\n", lmd->lmd_profile);
2008 if (!client_fill_super) {
2009 LCONSOLE_ERROR_MSG(0x165, "Nothing registered for "
2010 "client mount! Is the 'lustre' "
2011 "module loaded?\n");
2015 rc = lustre_start_mgc(sb);
2017 lustre_stop_mgc(sb);
2021 /* Connect and start */
2022 /* (should always be ll_fill_super) */
2023 rc = (*client_fill_super)(sb);
2024 /* c_f_s will call lustre_common_put_super on failure */
2027 CDEBUG(D_MOUNT, "Mounting server from %s\n", lmd->lmd_dev);
2028 lsi->lsi_flags |= LSI_SERVER;
2029 rc = server_fill_super(sb);
2030 /* s_f_s calls lustre_start_mgc after the mount because we need
2031 the MGS nids which are stored on disk. Plus, we may
2032 need to start the MGS first. */
2033 /* s_f_s will call server_put_super on failure */
2036 /* If error happens in fill_super() call, @lsi will be killed there.
2037 * This is why we do not put it here. */
2041 CERROR("Unable to mount %s (%d)\n",
2042 s2lsi(sb) ? lmd->lmd_dev : "", rc);
2044 CDEBUG(D_SUPER, "Mount %s complete\n",
2051 /* We can't call ll_fill_super by name because it lives in a module that
2052 must be loaded after this one. */
2053 void lustre_register_client_fill_super(int (*cfs)(struct super_block *sb))
2055 client_fill_super = cfs;
2058 void lustre_register_kill_super_cb(void (*cfs)(struct super_block *sb))
2060 kill_super_cb = cfs;
2063 /***************** FS registration ******************/
2066 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,18))
2067 struct super_block * lustre_get_sb(struct file_system_type *fs_type,
2068 int flags, const char *devname, void * data)
2070 /* calls back in fill super */
2071 /* we could append devname= onto options (*data) here,
2072 but 2.4 doesn't get devname. So we do it in mount_lustre.c */
2073 return get_sb_nodev(fs_type, flags, data, lustre_fill_super);
2076 int lustre_get_sb(struct file_system_type *fs_type,
2077 int flags, const char *devname, void * data,
2078 struct vfsmount *mnt)
2080 /* calls back in fill super */
2081 /* we could append devname= onto options (*data) here,
2082 but 2.4 doesn't get devname. So we do it in mount_lustre.c */
2083 return get_sb_nodev(fs_type, flags, data, lustre_fill_super, mnt);
2087 void lustre_kill_super(struct super_block *sb)
2089 struct lustre_sb_info *lsi = s2lsi(sb);
2091 if (kill_super_cb && lsi && !(lsi->lsi_flags & LSI_SERVER))
2092 (*kill_super_cb)(sb);
2094 kill_anon_super(sb);
2097 struct file_system_type lustre_fs_type = {
2098 .owner = THIS_MODULE,
2100 .get_sb = lustre_get_sb,
2101 .kill_sb = lustre_kill_super,
2102 .fs_flags = FS_BINARY_MOUNTDATA | FS_REQUIRES_DEV |
2103 #ifdef FS_HAS_FIEMAP
2106 LL_RENAME_DOES_D_MOVE,
2109 int lustre_register_fs(void)
2111 return register_filesystem(&lustre_fs_type);
2114 int lustre_unregister_fs(void)
2116 return unregister_filesystem(&lustre_fs_type);
2119 EXPORT_SYMBOL(lustre_register_client_fill_super);
2120 EXPORT_SYMBOL(lustre_register_kill_super_cb);
2121 EXPORT_SYMBOL(lustre_common_put_super);
2122 EXPORT_SYMBOL(lustre_process_log);
2123 EXPORT_SYMBOL(lustre_end_log);
2124 EXPORT_SYMBOL(server_find_mount_locked);
2125 EXPORT_SYMBOL(server_get_mount);
2126 EXPORT_SYMBOL(server_put_mount);
2127 EXPORT_SYMBOL(server_register_target);
2128 EXPORT_SYMBOL(server_name2index);
2129 EXPORT_SYMBOL(server_mti_print);
2130 EXPORT_SYMBOL(do_lcfg);