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 (c) 2007, 2010, Oracle and/or its affiliates. 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 CFS_DECLARE_MUTEX(lustre_mount_info_lock);
65 static CFS_LIST_HEAD(server_mount_info_list);
67 static struct lustre_mount_info *server_find_mount(const char *name)
70 struct lustre_mount_info *lmi;
73 cfs_list_for_each(tmp, &server_mount_info_list) {
74 lmi = cfs_list_entry(tmp, struct lustre_mount_info,
76 if (strcmp(name, lmi->lmi_name) == 0)
82 /* we must register an obd for a mount before we call the setup routine.
83 *_setup will call lustre_get_mount to get the mnt struct
84 by obd_name, since we can't pass the pointer to setup. */
85 static int server_register_mount(const char *name, struct super_block *sb,
88 struct lustre_mount_info *lmi;
95 OBD_ALLOC(lmi, sizeof(*lmi));
98 OBD_ALLOC(name_cp, strlen(name) + 1);
100 OBD_FREE(lmi, sizeof(*lmi));
103 strcpy(name_cp, name);
105 cfs_down(&lustre_mount_info_lock);
107 if (server_find_mount(name)) {
108 cfs_up(&lustre_mount_info_lock);
109 OBD_FREE(lmi, sizeof(*lmi));
110 OBD_FREE(name_cp, strlen(name) + 1);
111 CERROR("Already registered %s\n", name);
114 lmi->lmi_name = name_cp;
117 cfs_list_add(&lmi->lmi_list_chain, &server_mount_info_list);
119 cfs_up(&lustre_mount_info_lock);
121 CDEBUG(D_MOUNT, "reg_mnt %p from %s, vfscount=%d\n",
122 lmi->lmi_mnt, name, cfs_atomic_read(&lmi->lmi_mnt->mnt_count));
127 /* when an obd no longer needs a mount */
128 static int server_deregister_mount(const char *name)
130 struct lustre_mount_info *lmi;
133 cfs_down(&lustre_mount_info_lock);
134 lmi = server_find_mount(name);
136 cfs_up(&lustre_mount_info_lock);
137 CERROR("%s not registered\n", name);
141 CDEBUG(D_MOUNT, "dereg_mnt %p from %s, vfscount=%d\n",
142 lmi->lmi_mnt, name, cfs_atomic_read(&lmi->lmi_mnt->mnt_count));
144 OBD_FREE(lmi->lmi_name, strlen(lmi->lmi_name) + 1);
145 cfs_list_del(&lmi->lmi_list_chain);
146 OBD_FREE(lmi, sizeof(*lmi));
147 cfs_up(&lustre_mount_info_lock);
152 /* obd's look up a registered mount using their obdname. This is just
153 for initial obd setup to find the mount struct. It should not be
154 called every time you want to mntget. */
155 struct lustre_mount_info *server_get_mount(const char *name)
157 struct lustre_mount_info *lmi;
158 struct lustre_sb_info *lsi;
161 cfs_down(&lustre_mount_info_lock);
162 lmi = server_find_mount(name);
163 cfs_up(&lustre_mount_info_lock);
165 CERROR("Can't find mount for %s\n", name);
168 lsi = s2lsi(lmi->lmi_sb);
169 mntget(lmi->lmi_mnt);
170 cfs_atomic_inc(&lsi->lsi_mounts);
172 CDEBUG(D_MOUNT, "get_mnt %p from %s, refs=%d, vfscount=%d\n",
173 lmi->lmi_mnt, name, cfs_atomic_read(&lsi->lsi_mounts),
174 cfs_atomic_read(&lmi->lmi_mnt->mnt_count));
180 * Used by mdt to get mount_info from obdname.
181 * There are no blocking when using the mount_info.
182 * Do not use server_get_mount for this purpose.
184 struct lustre_mount_info *server_get_mount_2(const char *name)
186 struct lustre_mount_info *lmi;
189 cfs_down(&lustre_mount_info_lock);
190 lmi = server_find_mount(name);
191 cfs_up(&lustre_mount_info_lock);
193 CERROR("Can't find mount for %s\n", name);
198 static void unlock_mntput(struct vfsmount *mnt)
200 if (kernel_locked()) {
209 static int lustre_put_lsi(struct super_block *sb);
211 /* to be called from obd_cleanup methods */
212 int server_put_mount(const char *name, struct vfsmount *mnt)
214 struct lustre_mount_info *lmi;
215 struct lustre_sb_info *lsi;
216 int count = atomic_read(&mnt->mnt_count) - 1;
219 /* This might be the last one, can't deref after this */
222 cfs_down(&lustre_mount_info_lock);
223 lmi = server_find_mount(name);
224 cfs_up(&lustre_mount_info_lock);
226 CERROR("Can't find mount for %s\n", name);
229 lsi = s2lsi(lmi->lmi_sb);
230 LASSERT(lmi->lmi_mnt == mnt);
232 CDEBUG(D_MOUNT, "put_mnt %p from %s, refs=%d, vfscount=%d\n",
233 lmi->lmi_mnt, name, cfs_atomic_read(&lsi->lsi_mounts), count);
235 if (lustre_put_lsi(lmi->lmi_sb)) {
236 CDEBUG(D_MOUNT, "Last put of mnt %p from %s, vfscount=%d\n",
237 lmi->lmi_mnt, name, count);
238 /* last mount is the One True Mount */
240 CERROR("%s: mount busy, vfscount=%d!\n", name, count);
243 /* this obd should never need the mount again */
244 server_deregister_mount(name);
249 /* Corresponding to server_get_mount_2 */
250 int server_put_mount_2(const char *name, struct vfsmount *mnt)
256 /******* mount helper utilities *********/
259 static void ldd_print(struct lustre_disk_data *ldd)
261 PRINT_CMD(PRINT_MASK, " disk data:\n");
262 PRINT_CMD(PRINT_MASK, "server: %s\n", ldd->ldd_svname);
263 PRINT_CMD(PRINT_MASK, "uuid: %s\n", (char *)ldd->ldd_uuid);
264 PRINT_CMD(PRINT_MASK, "fs: %s\n", ldd->ldd_fsname);
265 PRINT_CMD(PRINT_MASK, "index: %04x\n", ldd->ldd_svindex);
266 PRINT_CMD(PRINT_MASK, "config: %d\n", ldd->ldd_config_ver);
267 PRINT_CMD(PRINT_MASK, "flags: %#x\n", ldd->ldd_flags);
268 PRINT_CMD(PRINT_MASK, "diskfs: %s\n", MT_STR(ldd));
269 PRINT_CMD(PRINT_MASK, "options: %s\n", ldd->ldd_mount_opts);
270 PRINT_CMD(PRINT_MASK, "params: %s\n", ldd->ldd_params);
271 PRINT_CMD(PRINT_MASK, "comment: %s\n", ldd->ldd_userdata);
275 static int ldd_parse(struct lvfs_run_ctxt *mount_ctxt,
276 struct lustre_disk_data *ldd)
278 struct lvfs_run_ctxt saved;
285 push_ctxt(&saved, mount_ctxt, NULL);
287 file = filp_open(MOUNT_DATA_FILE, O_RDONLY, 0644);
290 CERROR("cannot open %s: rc = %d\n", MOUNT_DATA_FILE, rc);
294 len = i_size_read(file->f_dentry->d_inode);
295 CDEBUG(D_MOUNT, "Have %s, size %lu\n", MOUNT_DATA_FILE, len);
296 if (len != sizeof(*ldd)) {
297 CERROR("disk data size does not match: see %lu expect %u\n",
298 len, (int)sizeof(*ldd));
299 GOTO(out_close, rc = -EINVAL);
302 rc = lustre_fread(file, ldd, len, &off);
304 CERROR("error reading %s: read %d of %lu\n",
305 MOUNT_DATA_FILE, rc, len);
306 GOTO(out_close, rc = -EINVAL);
310 if (ldd->ldd_magic != LDD_MAGIC) {
311 /* FIXME add swabbing support */
312 CERROR("Bad magic in %s: %x!=%x\n", MOUNT_DATA_FILE,
313 ldd->ldd_magic, LDD_MAGIC);
314 GOTO(out_close, rc = -EINVAL);
317 if (ldd->ldd_feature_incompat & ~LDD_INCOMPAT_SUPP) {
318 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
320 ldd->ldd_feature_incompat & ~LDD_INCOMPAT_SUPP);
321 GOTO(out_close, rc = -EINVAL);
323 if (ldd->ldd_feature_rocompat & ~LDD_ROCOMPAT_SUPP) {
324 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
326 ldd->ldd_feature_rocompat & ~LDD_ROCOMPAT_SUPP);
327 /* Do something like remount filesystem read-only */
328 GOTO(out_close, rc = -EINVAL);
334 pop_ctxt(&saved, mount_ctxt, NULL);
338 static int ldd_write(struct lvfs_run_ctxt *mount_ctxt,
339 struct lustre_disk_data *ldd)
341 struct lvfs_run_ctxt saved;
344 unsigned long len = sizeof(struct lustre_disk_data);
348 LASSERT(ldd->ldd_magic == LDD_MAGIC);
350 ldd->ldd_config_ver++;
352 push_ctxt(&saved, mount_ctxt, NULL);
354 file = filp_open(MOUNT_DATA_FILE, O_RDWR, 0644);
357 CERROR("cannot open %s: rc = %d\n", MOUNT_DATA_FILE, rc);
361 rc = lustre_fwrite(file, ldd, len, &off);
363 CERROR("error writing %s: read %d of %lu\n",
364 MOUNT_DATA_FILE, rc, len);
365 GOTO(out_close, rc = -EINVAL);
373 pop_ctxt(&saved, mount_ctxt, NULL);
378 /**************** config llog ********************/
380 /** Get a config log from the MGS and process it.
381 * This func is called for both clients and servers.
382 * Continue to process new statements appended to the logs
383 * (whenever the config lock is revoked) until lustre_end_log
385 * @param sb The superblock is used by the MGC to write to the local copy of
387 * @param logname The name of the llog to replicate from the MGS
388 * @param cfg Since the same mgc may be used to follow multiple config logs
389 * (e.g. ost1, ost2, client), the config_llog_instance keeps the state for
390 * this log, and is added to the mgc's list of logs to follow.
392 int lustre_process_log(struct super_block *sb, char *logname,
393 struct config_llog_instance *cfg)
395 struct lustre_cfg *lcfg;
396 struct lustre_cfg_bufs bufs;
397 struct lustre_sb_info *lsi = s2lsi(sb);
398 struct obd_device *mgc = lsi->lsi_mgc;
405 /* mgc_process_config */
406 lustre_cfg_bufs_reset(&bufs, mgc->obd_name);
407 lustre_cfg_bufs_set_string(&bufs, 1, logname);
408 lustre_cfg_bufs_set(&bufs, 2, cfg, sizeof(*cfg));
409 lustre_cfg_bufs_set(&bufs, 3, &sb, sizeof(sb));
410 lcfg = lustre_cfg_new(LCFG_LOG_START, &bufs);
411 rc = obd_process_config(mgc, sizeof(*lcfg), lcfg);
412 lustre_cfg_free(lcfg);
415 LCONSOLE_ERROR_MSG(0x15b, "%s: The configuration from log '%s'"
416 "failed from the MGS (%d). Make sure this "
417 "client and the MGS are running compatible "
418 "versions of Lustre.\n",
419 mgc->obd_name, logname, rc);
422 LCONSOLE_ERROR_MSG(0x15c, "%s: The configuration from log '%s' "
423 "failed (%d). This may be the result of "
424 "communication errors between this node and "
425 "the MGS, a bad configuration, or other "
426 "errors. See the syslog for more "
427 "information.\n", mgc->obd_name, logname,
430 /* class_obd_list(); */
434 /* Stop watching this config log for updates */
435 int lustre_end_log(struct super_block *sb, char *logname,
436 struct config_llog_instance *cfg)
438 struct lustre_cfg *lcfg;
439 struct lustre_cfg_bufs bufs;
440 struct lustre_sb_info *lsi = s2lsi(sb);
441 struct obd_device *mgc = lsi->lsi_mgc;
448 /* mgc_process_config */
449 lustre_cfg_bufs_reset(&bufs, mgc->obd_name);
450 lustre_cfg_bufs_set_string(&bufs, 1, logname);
452 lustre_cfg_bufs_set(&bufs, 2, cfg, sizeof(*cfg));
453 lcfg = lustre_cfg_new(LCFG_LOG_END, &bufs);
454 rc = obd_process_config(mgc, sizeof(*lcfg), lcfg);
455 lustre_cfg_free(lcfg);
459 /**************** obd start *******************/
461 /** lustre_cfg_bufs are a holdover from 1.4; we can still set these up from
462 * lctl (and do for echo cli/srv.
464 int do_lcfg(char *cfgname, lnet_nid_t nid, int cmd,
465 char *s1, char *s2, char *s3, char *s4)
467 struct lustre_cfg_bufs bufs;
468 struct lustre_cfg * lcfg = NULL;
471 CDEBUG(D_TRACE, "lcfg %s %#x %s %s %s %s\n", cfgname,
472 cmd, s1, s2, s3, s4);
474 lustre_cfg_bufs_reset(&bufs, cfgname);
476 lustre_cfg_bufs_set_string(&bufs, 1, s1);
478 lustre_cfg_bufs_set_string(&bufs, 2, s2);
480 lustre_cfg_bufs_set_string(&bufs, 3, s3);
482 lustre_cfg_bufs_set_string(&bufs, 4, s4);
484 lcfg = lustre_cfg_new(cmd, &bufs);
485 lcfg->lcfg_nid = nid;
486 rc = class_process_config(lcfg);
487 lustre_cfg_free(lcfg);
491 /** Call class_attach and class_setup. These methods in turn call
492 * obd type-specific methods.
494 static int lustre_start_simple(char *obdname, char *type, char *uuid,
498 CDEBUG(D_MOUNT, "Starting obd %s (typ=%s)\n", obdname, type);
500 rc = do_lcfg(obdname, 0, LCFG_ATTACH, type, uuid, 0, 0);
502 CERROR("%s attach error %d\n", obdname, rc);
505 rc = do_lcfg(obdname, 0, LCFG_SETUP, s1, s2, 0, 0);
507 CERROR("%s setup error %d\n", obdname, rc);
508 do_lcfg(obdname, 0, LCFG_DETACH, 0, 0, 0, 0);
513 /* Set up a MGS to serve startup logs */
514 static int server_start_mgs(struct super_block *sb)
516 struct lustre_sb_info *lsi = s2lsi(sb);
517 struct vfsmount *mnt = lsi->lsi_srv_mnt;
518 struct lustre_mount_info *lmi;
523 /* It is impossible to have more than 1 MGS per node, since
524 MGC wouldn't know which to connect to */
525 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
527 lsi = s2lsi(lmi->lmi_sb);
528 LCONSOLE_ERROR_MSG(0x15d, "The MGS service was already started"
530 lsi->lsi_ldd->ldd_svname);
534 CDEBUG(D_CONFIG, "Start MGS service %s\n", LUSTRE_MGS_OBDNAME);
536 rc = server_register_mount(LUSTRE_MGS_OBDNAME, sb, mnt);
539 rc = lustre_start_simple(LUSTRE_MGS_OBDNAME, LUSTRE_MGS_NAME,
540 LUSTRE_MGS_OBDNAME, 0, 0);
541 /* Do NOT call server_deregister_mount() here. This leads to
542 * inability cleanup cleanly and free lsi and other stuff when
543 * mgs calls server_put_mount() in error handling case. -umka */
547 LCONSOLE_ERROR_MSG(0x15e, "Failed to start MGS '%s' (%d). "
548 "Is the 'mgs' module loaded?\n",
549 LUSTRE_MGS_OBDNAME, rc);
553 static int server_stop_mgs(struct super_block *sb)
555 struct obd_device *obd;
559 CDEBUG(D_MOUNT, "Stop MGS service %s\n", LUSTRE_MGS_OBDNAME);
561 /* There better be only one MGS */
562 obd = class_name2obd(LUSTRE_MGS_OBDNAME);
564 CDEBUG(D_CONFIG, "mgs %s not running\n", LUSTRE_MGS_OBDNAME);
568 /* The MGS should always stop when we say so */
570 rc = class_manual_cleanup(obd);
574 CFS_DECLARE_MUTEX(mgc_start_lock);
576 /** Set up a mgc obd to process startup logs
578 * \param sb [in] super block of the mgc obd
580 * \retval 0 success, otherwise error code
582 static int lustre_start_mgc(struct super_block *sb)
584 struct obd_connect_data *data = NULL;
585 struct lustre_sb_info *lsi = s2lsi(sb);
586 struct obd_device *obd;
587 struct obd_export *exp;
588 struct obd_uuid *uuid;
591 char *mgcname, *niduuid, *mgssec;
594 int rc = 0, i = 0, j, len;
597 LASSERT(lsi->lsi_lmd);
599 /* Find the first non-lo MGS nid for our MGC name */
600 if (lsi->lsi_flags & LSI_SERVER) {
601 ptr = lsi->lsi_ldd->ldd_params;
602 /* Use mgsnode= nids */
603 if ((class_find_param(ptr, PARAM_MGSNODE, &ptr) == 0) &&
604 (class_parse_nid(ptr, &nid, &ptr) == 0)) {
606 } else if (IS_MGS(lsi->lsi_ldd)) {
607 lnet_process_id_t id;
608 while ((rc = LNetGetId(i++, &id)) != -ENOENT) {
609 if (LNET_NETTYP(LNET_NIDNET(id.nid)) == LOLND)
616 } else { /* client */
617 /* Use nids from mount line: uml1,1@elan:uml2,2@elan:/lustre */
618 ptr = lsi->lsi_lmd->lmd_dev;
619 if (class_parse_nid(ptr, &nid, &ptr) == 0)
623 CERROR("No valid MGS nids found.\n");
627 len = strlen(LUSTRE_MGC_OBDNAME) + strlen(libcfs_nid2str(nid)) + 1;
628 OBD_ALLOC(mgcname, len);
629 OBD_ALLOC(niduuid, len + 2);
630 if (!mgcname || !niduuid)
631 GOTO(out_free, rc = -ENOMEM);
632 sprintf(mgcname, "%s%s", LUSTRE_MGC_OBDNAME, libcfs_nid2str(nid));
634 mgssec = lsi->lsi_lmd->lmd_mgssec ? lsi->lsi_lmd->lmd_mgssec : "";
636 cfs_mutex_down(&mgc_start_lock);
638 obd = class_name2obd(mgcname);
639 if (obd && !obd->obd_stopping) {
640 rc = obd_set_info_async(obd->obd_self_export,
641 strlen(KEY_MGSSEC), KEY_MGSSEC,
642 strlen(mgssec), mgssec, NULL);
646 /* Re-using an existing MGC */
647 cfs_atomic_inc(&obd->u.cli.cl_mgc_refcount);
650 /* If we are restarting the MGS, don't try to keep the MGC's
651 old connection, or registration will fail. */
652 if ((lsi->lsi_flags & LSI_SERVER) && IS_MGS(lsi->lsi_ldd)) {
653 CDEBUG(D_MOUNT, "New MGS with live MGC\n");
657 /* Try all connections, but only once (again).
658 We don't want to block another target from starting
659 (using its local copy of the log), but we do want to connect
660 if at all possible. */
662 CDEBUG(D_MOUNT, "%s: Set MGC reconnect %d\n", mgcname,recov_bk);
663 rc = obd_set_info_async(obd->obd_self_export,
664 sizeof(KEY_INIT_RECOV_BACKUP),
665 KEY_INIT_RECOV_BACKUP,
666 sizeof(recov_bk), &recov_bk, NULL);
670 CDEBUG(D_MOUNT, "Start MGC '%s'\n", mgcname);
672 /* Add the primary nids for the MGS */
674 sprintf(niduuid, "%s_%x", mgcname, i);
675 if (lsi->lsi_flags & LSI_SERVER) {
676 ptr = lsi->lsi_ldd->ldd_params;
677 if (IS_MGS(lsi->lsi_ldd)) {
678 /* Use local nids (including LO) */
679 lnet_process_id_t id;
680 while ((rc = LNetGetId(i++, &id)) != -ENOENT) {
681 rc = do_lcfg(mgcname, id.nid,
682 LCFG_ADD_UUID, niduuid, 0,0,0);
685 /* Use mgsnode= nids */
686 if (class_find_param(ptr, PARAM_MGSNODE, &ptr) != 0) {
687 CERROR("No MGS nids given.\n");
688 GOTO(out_free, rc = -EINVAL);
690 while (class_parse_nid(ptr, &nid, &ptr) == 0) {
691 rc = do_lcfg(mgcname, nid,
692 LCFG_ADD_UUID, niduuid, 0,0,0);
696 } else { /* client */
697 /* Use nids from mount line: uml1,1@elan:uml2,2@elan:/lustre */
698 ptr = lsi->lsi_lmd->lmd_dev;
699 while (class_parse_nid(ptr, &nid, &ptr) == 0) {
700 rc = do_lcfg(mgcname, nid,
701 LCFG_ADD_UUID, niduuid, 0,0,0);
703 /* Stop at the first failover nid */
709 CERROR("No valid MGS nids found.\n");
710 GOTO(out_free, rc = -EINVAL);
712 lsi->lsi_lmd->lmd_mgs_failnodes = 1;
714 /* Random uuid for MGC allows easier reconnects */
716 ll_generate_random_uuid(uuidc);
717 class_uuid_unparse(uuidc, uuid);
720 rc = lustre_start_simple(mgcname, LUSTRE_MGC_NAME,
721 (char *)uuid->uuid, LUSTRE_MGS_OBDNAME,
727 /* Add any failover MGS nids */
729 while ((*ptr == ':' ||
730 class_find_param(ptr, PARAM_MGSNODE, &ptr) == 0)) {
731 /* New failover node */
732 sprintf(niduuid, "%s_%x", mgcname, i);
734 while (class_parse_nid(ptr, &nid, &ptr) == 0) {
736 rc = do_lcfg(mgcname, nid,
737 LCFG_ADD_UUID, niduuid, 0,0,0);
742 rc = do_lcfg(mgcname, 0, LCFG_ADD_CONN,
750 lsi->lsi_lmd->lmd_mgs_failnodes = i;
752 obd = class_name2obd(mgcname);
754 CERROR("Can't find mgcobd %s\n", mgcname);
755 GOTO(out_free, rc = -ENOTCONN);
758 rc = obd_set_info_async(obd->obd_self_export,
759 strlen(KEY_MGSSEC), KEY_MGSSEC,
760 strlen(mgssec), mgssec, NULL);
764 /* Keep a refcount of servers/clients who started with "mount",
765 so we know when we can get rid of the mgc. */
766 cfs_atomic_set(&obd->u.cli.cl_mgc_refcount, 1);
768 /* Try all connections, but only once. */
770 rc = obd_set_info_async(obd->obd_self_export,
771 sizeof(KEY_INIT_RECOV_BACKUP),
772 KEY_INIT_RECOV_BACKUP,
773 sizeof(recov_bk), &recov_bk, NULL);
776 CWARN("can't set %s %d\n", KEY_INIT_RECOV_BACKUP, rc);
777 /* We connect to the MGS at setup, and don't disconnect until cleanup */
780 GOTO(out, rc = -ENOMEM);
781 data->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_FID |
782 OBD_CONNECT_AT | OBD_CONNECT_FULL20;
783 data->ocd_version = LUSTRE_VERSION_CODE;
784 rc = obd_connect(NULL, &exp, obd, &(obd->obd_uuid), data, NULL);
787 CERROR("connect failed %d\n", rc);
791 obd->u.cli.cl_mgc_mgsexp = exp;
794 /* Keep the mgc info in the sb. Note that many lsi's can point
798 cfs_mutex_up(&mgc_start_lock);
801 OBD_FREE(mgcname, len);
803 OBD_FREE(niduuid, len + 2);
807 static int lustre_stop_mgc(struct super_block *sb)
809 struct lustre_sb_info *lsi = s2lsi(sb);
810 struct obd_device *obd;
811 char *niduuid = 0, *ptr = 0;
812 int i, rc = 0, len = 0;
822 cfs_mutex_down(&mgc_start_lock);
823 LASSERT(cfs_atomic_read(&obd->u.cli.cl_mgc_refcount) > 0);
824 if (!cfs_atomic_dec_and_test(&obd->u.cli.cl_mgc_refcount)) {
825 /* This is not fatal, every client that stops
826 will call in here. */
827 CDEBUG(D_MOUNT, "mgc still has %d references.\n",
828 cfs_atomic_read(&obd->u.cli.cl_mgc_refcount));
829 GOTO(out, rc = -EBUSY);
832 /* The MGC has no recoverable data in any case.
833 * force shotdown set in umount_begin */
834 obd->obd_no_recov = 1;
836 if (obd->u.cli.cl_mgc_mgsexp) {
837 /* An error is not fatal, if we are unable to send the
838 disconnect mgs ping evictor cleans up the export */
839 rc = obd_disconnect(obd->u.cli.cl_mgc_mgsexp);
841 CDEBUG(D_MOUNT, "disconnect failed %d\n", rc);
844 /* Save the obdname for cleaning the nid uuids, which are
846 len = strlen(obd->obd_name) + 6;
847 OBD_ALLOC(niduuid, len);
849 strcpy(niduuid, obd->obd_name);
850 ptr = niduuid + strlen(niduuid);
853 rc = class_manual_cleanup(obd);
857 /* Clean the nid uuids */
859 GOTO(out, rc = -ENOMEM);
861 for (i = 0; i < lsi->lsi_lmd->lmd_mgs_failnodes; i++) {
862 sprintf(ptr, "_%x", i);
863 rc = do_lcfg(LUSTRE_MGC_OBDNAME, 0, LCFG_DEL_UUID,
866 CERROR("del MDC UUID %s failed: rc = %d\n",
871 OBD_FREE(niduuid, len);
873 /* class_import_put will get rid of the additional connections */
874 cfs_mutex_up(&mgc_start_lock);
878 /* Since there's only one mgc per node, we have to change it's fs to get
879 access to the right disk. */
880 static int server_mgc_set_fs(struct obd_device *mgc, struct super_block *sb)
882 struct lustre_sb_info *lsi = s2lsi(sb);
886 CDEBUG(D_MOUNT, "Set mgc disk for %s\n", lsi->lsi_lmd->lmd_dev);
888 /* cl_mgc_sem in mgc insures we sleep if the mgc_fs is busy */
889 rc = obd_set_info_async(mgc->obd_self_export,
890 sizeof(KEY_SET_FS), KEY_SET_FS,
891 sizeof(*sb), sb, NULL);
893 CERROR("can't set_fs %d\n", rc);
899 static int server_mgc_clear_fs(struct obd_device *mgc)
904 CDEBUG(D_MOUNT, "Unassign mgc disk\n");
906 rc = obd_set_info_async(mgc->obd_self_export,
907 sizeof(KEY_CLEAR_FS), KEY_CLEAR_FS,
912 CFS_DECLARE_MUTEX(server_start_lock);
914 /* Stop MDS/OSS if nobody is using them */
915 static int server_stop_servers(int lddflags, int lsiflags)
917 struct obd_device *obd = NULL;
918 struct obd_type *type = NULL;
922 cfs_mutex_down(&server_start_lock);
924 /* Either an MDT or an OST or neither */
925 /* if this was an MDT, and there are no more MDT's, clean up the MDS */
926 if ((lddflags & LDD_F_SV_TYPE_MDT) &&
927 (obd = class_name2obd(LUSTRE_MDS_OBDNAME))) {
928 /*FIXME pre-rename, should eventually be LUSTRE_MDT_NAME*/
929 type = class_search_type(LUSTRE_MDS_NAME);
931 /* if this was an OST, and there are no more OST's, clean up the OSS */
932 if ((lddflags & LDD_F_SV_TYPE_OST) &&
933 (obd = class_name2obd(LUSTRE_OSS_OBDNAME))) {
934 type = class_search_type(LUSTRE_OST_NAME);
937 if (obd && (!type || !type->typ_refcnt)) {
940 /* obd_fail doesn't mean much on a server obd */
941 err = class_manual_cleanup(obd);
946 cfs_mutex_up(&server_start_lock);
951 int server_mti_print(char *title, struct mgs_target_info *mti)
953 PRINT_CMD(PRINT_MASK, "mti %s\n", title);
954 PRINT_CMD(PRINT_MASK, "server: %s\n", mti->mti_svname);
955 PRINT_CMD(PRINT_MASK, "fs: %s\n", mti->mti_fsname);
956 PRINT_CMD(PRINT_MASK, "uuid: %s\n", mti->mti_uuid);
957 PRINT_CMD(PRINT_MASK, "ver: %d flags: %#x\n",
958 mti->mti_config_ver, mti->mti_flags);
962 static int server_sb2mti(struct super_block *sb, struct mgs_target_info *mti)
964 struct lustre_sb_info *lsi = s2lsi(sb);
965 struct lustre_disk_data *ldd = lsi->lsi_ldd;
966 lnet_process_id_t id;
970 if (!(lsi->lsi_flags & LSI_SERVER))
973 strncpy(mti->mti_fsname, ldd->ldd_fsname,
974 sizeof(mti->mti_fsname));
975 strncpy(mti->mti_svname, ldd->ldd_svname,
976 sizeof(mti->mti_svname));
978 mti->mti_nid_count = 0;
979 while (LNetGetId(i++, &id) != -ENOENT) {
980 if (LNET_NETTYP(LNET_NIDNET(id.nid)) == LOLND)
983 /* server use --servicenode param, only allow specified
984 * nids be registered */
985 if ((ldd->ldd_flags & LDD_F_NO_PRIMNODE) != 0 &&
986 class_match_nid(ldd->ldd_params,
987 PARAM_FAILNODE, id.nid) < 1)
990 /* match specified network */
991 if (!class_match_net(ldd->ldd_params,
992 PARAM_NETWORK, LNET_NIDNET(id.nid)))
995 mti->mti_nids[mti->mti_nid_count] = id.nid;
996 mti->mti_nid_count++;
997 if (mti->mti_nid_count >= MTI_NIDS_MAX) {
998 CWARN("Only using first %d nids for %s\n",
999 mti->mti_nid_count, mti->mti_svname);
1004 mti->mti_lustre_ver = LUSTRE_VERSION_CODE;
1005 mti->mti_config_ver = 0;
1006 if (lsi->lsi_lmd->lmd_flags & LMD_FLG_WRITECONF)
1007 ldd->ldd_flags |= LDD_F_WRITECONF;
1008 mti->mti_flags = ldd->ldd_flags;
1009 mti->mti_stripe_index = ldd->ldd_svindex;
1010 memcpy(mti->mti_uuid, ldd->ldd_uuid, sizeof(mti->mti_uuid));
1011 if (strlen(ldd->ldd_params) > sizeof(mti->mti_params)) {
1012 CERROR("params too big for mti\n");
1015 memcpy(mti->mti_params, ldd->ldd_params, sizeof(mti->mti_params));
1019 /* Register an old or new target with the MGS. If needed MGS will construct
1020 startup logs and assign index */
1021 int server_register_target(struct super_block *sb)
1023 struct lustre_sb_info *lsi = s2lsi(sb);
1024 struct obd_device *mgc = lsi->lsi_mgc;
1025 struct lustre_disk_data *ldd = lsi->lsi_ldd;
1026 struct mgs_target_info *mti = NULL;
1032 if (!(lsi->lsi_flags & LSI_SERVER))
1038 rc = server_sb2mti(sb, mti);
1042 CDEBUG(D_MOUNT, "Registration %s, fs=%s, %s, index=%04x, flags=%#x\n",
1043 mti->mti_svname, mti->mti_fsname,
1044 libcfs_nid2str(mti->mti_nids[0]), mti->mti_stripe_index,
1047 /* Register the target */
1048 /* FIXME use mgc_process_config instead */
1049 rc = obd_set_info_async(mgc->u.cli.cl_mgc_mgsexp,
1050 sizeof(KEY_REGISTER_TARGET), KEY_REGISTER_TARGET,
1051 sizeof(*mti), mti, NULL);
1055 /* Always update our flags */
1056 ldd->ldd_flags = mti->mti_flags & ~LDD_F_REWRITE_LDD;
1058 /* If this flag is set, it means the MGS wants us to change our
1059 on-disk data. (So far this means just the index.) */
1060 if (mti->mti_flags & LDD_F_REWRITE_LDD) {
1063 CDEBUG(D_MOUNT, "Changing on-disk index from %#x to %#x "
1064 "for %s\n", ldd->ldd_svindex, mti->mti_stripe_index,
1066 ldd->ldd_svindex = mti->mti_stripe_index;
1067 strncpy(ldd->ldd_svname, mti->mti_svname,
1068 sizeof(ldd->ldd_svname));
1069 /* or ldd_make_sv_name(ldd); */
1070 ldd_write(&mgc->obd_lvfs_ctxt, ldd);
1071 err = fsfilt_set_label(mgc, lsi->lsi_srv_mnt->mnt_sb,
1074 CERROR("Label set error %d\n", err);
1075 label = fsfilt_get_label(mgc, lsi->lsi_srv_mnt->mnt_sb);
1077 CDEBUG(D_MOUNT, "Disk label changed to %s\n", label);
1079 /* Flush the new ldd to disk */
1080 fsfilt_sync(mgc, lsi->lsi_srv_mnt->mnt_sb);
1089 /** Start server targets: MDTs and OSTs
1091 static int server_start_targets(struct super_block *sb, struct vfsmount *mnt)
1093 struct obd_device *obd;
1094 struct lustre_sb_info *lsi = s2lsi(sb);
1095 struct config_llog_instance cfg;
1099 CDEBUG(D_MOUNT, "starting target %s\n", lsi->lsi_ldd->ldd_svname);
1102 /* If we're an MDT, make sure the global MDS is running */
1103 if (lsi->lsi_ldd->ldd_flags & LDD_F_SV_TYPE_MDT) {
1104 /* make sure the MDS is started */
1105 cfs_mutex_down(&server_start_lock);
1106 obd = class_name2obd(LUSTRE_MDS_OBDNAME);
1108 rc = lustre_start_simple(LUSTRE_MDS_OBDNAME,
1109 /* FIXME pre-rename, should eventually be LUSTRE_MDS_NAME */
1111 LUSTRE_MDS_OBDNAME"_uuid",
1114 cfs_mutex_up(&server_start_lock);
1115 CERROR("failed to start MDS: %d\n", rc);
1119 cfs_mutex_up(&server_start_lock);
1123 /* If we're an OST, make sure the global OSS is running */
1124 if (lsi->lsi_ldd->ldd_flags & LDD_F_SV_TYPE_OST) {
1125 /* make sure OSS is started */
1126 cfs_mutex_down(&server_start_lock);
1127 obd = class_name2obd(LUSTRE_OSS_OBDNAME);
1129 rc = lustre_start_simple(LUSTRE_OSS_OBDNAME,
1131 LUSTRE_OSS_OBDNAME"_uuid",
1134 cfs_mutex_up(&server_start_lock);
1135 CERROR("failed to start OSS: %d\n", rc);
1139 cfs_mutex_up(&server_start_lock);
1142 /* Set the mgc fs to our server disk. This allows the MGC to
1143 * read and write configs locally, in case it can't talk to the MGS. */
1144 rc = server_mgc_set_fs(lsi->lsi_mgc, sb);
1148 /* Register with MGS */
1149 rc = server_register_target(sb);
1150 if (rc && (lsi->lsi_ldd->ldd_flags &
1151 (LDD_F_NEED_INDEX | LDD_F_UPDATE | LDD_F_UPGRADE14))){
1152 CERROR("Required registration failed for %s: %d\n",
1153 lsi->lsi_ldd->ldd_svname, rc);
1155 LCONSOLE_ERROR_MSG(0x15f, "Communication error with "
1156 "the MGS. Is the MGS running?\n");
1160 if (rc == -EINVAL) {
1161 LCONSOLE_ERROR_MSG(0x160, "The MGS is refusing to allow this "
1162 "server (%s) to start. Please see messages"
1163 " on the MGS node.\n",
1164 lsi->lsi_ldd->ldd_svname);
1167 /* non-fatal error of registeration with MGS */
1169 CDEBUG(D_MOUNT, "Cannot register with MGS: %d\n", rc);
1171 /* Let the target look up the mount using the target's name
1172 (we can't pass the sb or mnt through class_process_config.) */
1173 rc = server_register_mount(lsi->lsi_ldd->ldd_svname, sb, mnt);
1177 /* Start targets using the llog named for the target */
1178 memset(&cfg, 0, sizeof(cfg));
1179 rc = lustre_process_log(sb, lsi->lsi_ldd->ldd_svname, &cfg);
1181 CERROR("failed to start server %s: %d\n",
1182 lsi->lsi_ldd->ldd_svname, rc);
1183 /* Do NOT call server_deregister_mount() here. This makes it
1184 * impossible to find mount later in cleanup time and leaves
1185 * @lsi and othder stuff leaked. -umka */
1190 /* Release the mgc fs for others to use */
1191 server_mgc_clear_fs(lsi->lsi_mgc);
1194 obd = class_name2obd(lsi->lsi_ldd->ldd_svname);
1196 CERROR("no server named %s was started\n",
1197 lsi->lsi_ldd->ldd_svname);
1201 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_ABORT_RECOV) &&
1202 (OBP(obd, iocontrol))) {
1203 obd_iocontrol(OBD_IOC_ABORT_RECOVERY,
1204 obd->obd_self_export, 0, NULL, NULL);
1207 /* log has been fully processed */
1208 obd_notify(obd, NULL, OBD_NOTIFY_CONFIG, (void *)CONFIG_LOG);
1214 /***************** lustre superblock **************/
1216 struct lustre_sb_info *lustre_init_lsi(struct super_block *sb)
1218 struct lustre_sb_info *lsi;
1224 OBD_ALLOC_PTR(lsi->lsi_lmd);
1225 if (!lsi->lsi_lmd) {
1230 lsi->lsi_lmd->lmd_exclude_count = 0;
1231 lsi->lsi_lmd->lmd_recovery_time_soft = 0;
1232 lsi->lsi_lmd->lmd_recovery_time_hard = 0;
1233 s2lsi_nocast(sb) = lsi;
1234 /* we take 1 extra ref for our setup */
1235 cfs_atomic_set(&lsi->lsi_mounts, 1);
1237 /* Default umount style */
1238 lsi->lsi_flags = LSI_UMOUNT_FAILOVER;
1243 static int lustre_free_lsi(struct super_block *sb)
1245 struct lustre_sb_info *lsi = s2lsi(sb);
1248 LASSERT(lsi != NULL);
1249 CDEBUG(D_MOUNT, "Freeing lsi %p\n", lsi);
1251 /* someone didn't call server_put_mount. */
1252 LASSERT(cfs_atomic_read(&lsi->lsi_mounts) == 0);
1254 if (lsi->lsi_ldd != NULL)
1255 OBD_FREE(lsi->lsi_ldd, sizeof(*lsi->lsi_ldd));
1257 if (lsi->lsi_lmd != NULL) {
1258 if (lsi->lsi_lmd->lmd_dev != NULL)
1259 OBD_FREE(lsi->lsi_lmd->lmd_dev,
1260 strlen(lsi->lsi_lmd->lmd_dev) + 1);
1261 if (lsi->lsi_lmd->lmd_profile != NULL)
1262 OBD_FREE(lsi->lsi_lmd->lmd_profile,
1263 strlen(lsi->lsi_lmd->lmd_profile) + 1);
1264 if (lsi->lsi_lmd->lmd_mgssec != NULL)
1265 OBD_FREE(lsi->lsi_lmd->lmd_mgssec,
1266 strlen(lsi->lsi_lmd->lmd_mgssec) + 1);
1267 if (lsi->lsi_lmd->lmd_opts != NULL)
1268 OBD_FREE(lsi->lsi_lmd->lmd_opts,
1269 strlen(lsi->lsi_lmd->lmd_opts) + 1);
1270 if (lsi->lsi_lmd->lmd_exclude_count)
1271 OBD_FREE(lsi->lsi_lmd->lmd_exclude,
1272 sizeof(lsi->lsi_lmd->lmd_exclude[0]) *
1273 lsi->lsi_lmd->lmd_exclude_count);
1274 OBD_FREE(lsi->lsi_lmd, sizeof(*lsi->lsi_lmd));
1277 LASSERT(lsi->lsi_llsbi == NULL);
1278 OBD_FREE(lsi, sizeof(*lsi));
1279 s2lsi_nocast(sb) = NULL;
1284 /* The lsi has one reference for every server that is using the disk -
1285 e.g. MDT, MGS, and potentially MGC */
1286 static int lustre_put_lsi(struct super_block *sb)
1288 struct lustre_sb_info *lsi = s2lsi(sb);
1291 LASSERT(lsi != NULL);
1293 CDEBUG(D_MOUNT, "put %p %d\n", sb, cfs_atomic_read(&lsi->lsi_mounts));
1294 if (cfs_atomic_dec_and_test(&lsi->lsi_mounts)) {
1295 lustre_free_lsi(sb);
1301 /*************** server mount ******************/
1303 /** Kernel mount using mount options in MOUNT_DATA_FILE.
1304 * Since this file lives on the disk, we pre-mount using a common
1305 * type, read the file, then re-mount using the type specified in the
1308 static struct vfsmount *server_kernel_mount(struct super_block *sb)
1310 struct lvfs_run_ctxt mount_ctxt;
1311 struct lustre_sb_info *lsi = s2lsi(sb);
1312 struct lustre_disk_data *ldd;
1313 struct lustre_mount_data *lmd = lsi->lsi_lmd;
1314 struct vfsmount *mnt;
1315 char *options = NULL;
1316 unsigned long page, s_flags;
1317 struct page *__page;
1321 OBD_ALLOC(ldd, sizeof(*ldd));
1323 RETURN(ERR_PTR(-ENOMEM));
1325 /* In the past, we have always used flags = 0.
1326 Note ext3/ldiskfs can't be mounted ro. */
1327 s_flags = sb->s_flags;
1329 /* allocate memory for options */
1330 OBD_PAGE_ALLOC(__page, CFS_ALLOC_STD);
1332 GOTO(out_free, rc = -ENOMEM);
1333 page = (unsigned long)cfs_page_address(__page);
1334 options = (char *)page;
1335 memset(options, 0, CFS_PAGE_SIZE);
1337 /* mount-line options must be added for pre-mount because it may
1338 * contain mount options such as journal_dev which are required
1339 * to mount successfuly the underlying filesystem */
1340 if (lmd->lmd_opts && (*(lmd->lmd_opts) != 0))
1341 strncat(options, lmd->lmd_opts, CFS_PAGE_SIZE - 1);
1343 /* Pre-mount ldiskfs to read the MOUNT_DATA_FILE */
1344 CDEBUG(D_MOUNT, "Pre-mount ldiskfs %s\n", lmd->lmd_dev);
1345 mnt = ll_kern_mount("ldiskfs", s_flags, lmd->lmd_dev, (void *)options);
1348 CERROR("premount %s:%#lx ldiskfs failed: %d "
1349 "Is the ldiskfs module available?\n",
1350 lmd->lmd_dev, s_flags, rc );
1354 OBD_SET_CTXT_MAGIC(&mount_ctxt);
1355 mount_ctxt.pwdmnt = mnt;
1356 mount_ctxt.pwd = mnt->mnt_root;
1357 mount_ctxt.fs = get_ds();
1359 rc = ldd_parse(&mount_ctxt, ldd);
1363 CERROR("premount parse options failed: rc = %d\n", rc);
1367 /* Done with our pre-mount, now do the real mount. */
1369 /* Glom up mount options */
1370 memset(options, 0, CFS_PAGE_SIZE);
1371 strncpy(options, ldd->ldd_mount_opts, CFS_PAGE_SIZE - 2);
1373 /* Add in any mount-line options */
1374 if (lmd->lmd_opts && (*(lmd->lmd_opts) != 0)) {
1375 int len = CFS_PAGE_SIZE - strlen(options) - 2;
1377 strcat(options, ",");
1378 strncat(options, lmd->lmd_opts, len);
1381 /* Special permanent mount flags */
1383 s_flags |= MS_NOATIME | MS_NODIRATIME;
1385 CDEBUG(D_MOUNT, "kern_mount: %s %s %s\n",
1386 MT_STR(ldd), lmd->lmd_dev, options);
1387 mnt = ll_kern_mount(MT_STR(ldd), s_flags, lmd->lmd_dev,
1391 CERROR("ll_kern_mount failed: rc = %d\n", rc);
1395 if (lmd->lmd_flags & LMD_FLG_ABORT_RECOV)
1396 simple_truncate(mnt->mnt_sb->s_root, mnt, LAST_RCVD,
1399 OBD_PAGE_FREE(__page);
1400 lsi->lsi_ldd = ldd; /* freed at lsi cleanup */
1401 CDEBUG(D_SUPER, "%s: mnt = %p\n", lmd->lmd_dev, mnt);
1406 OBD_PAGE_FREE(__page);
1407 OBD_FREE(ldd, sizeof(*ldd));
1408 lsi->lsi_ldd = NULL;
1409 RETURN(ERR_PTR(rc));
1412 /** Wait here forever until the mount refcount is 0 before completing umount,
1413 * else we risk dereferencing a null pointer.
1414 * LNET may take e.g. 165s before killing zombies.
1416 static void server_wait_finished(struct vfsmount *mnt)
1420 cfs_sigset_t blocked;
1422 cfs_waitq_init(&waitq);
1424 while (atomic_read(&mnt->mnt_count) > 1) {
1425 if (waited && (waited % 30 == 0))
1426 LCONSOLE_WARN("Mount still busy with %d refs after "
1428 atomic_read(&mnt->mnt_count),
1430 /* Cannot use l_event_wait() for an interruptible sleep. */
1432 blocked = cfs_block_sigsinv(sigmask(SIGKILL));
1433 cfs_waitq_wait_event_interruptible_timeout(
1435 (atomic_read(&mnt->mnt_count) == 1),
1436 cfs_time_seconds(3),
1438 cfs_block_sigs(blocked);
1440 LCONSOLE_EMERG("Danger: interrupted umount %s with "
1443 atomic_read(&mnt->mnt_count));
1450 /** Start the shutdown of servers at umount.
1452 static void server_put_super(struct super_block *sb)
1454 struct lustre_sb_info *lsi = s2lsi(sb);
1455 struct obd_device *obd;
1456 struct vfsmount *mnt = lsi->lsi_srv_mnt;
1457 char *tmpname, *extraname = NULL;
1459 int lddflags = lsi->lsi_ldd->ldd_flags;
1460 int lsiflags = lsi->lsi_flags;
1463 LASSERT(lsiflags & LSI_SERVER);
1465 tmpname_sz = strlen(lsi->lsi_ldd->ldd_svname) + 1;
1466 OBD_ALLOC(tmpname, tmpname_sz);
1467 memcpy(tmpname, lsi->lsi_ldd->ldd_svname, tmpname_sz);
1468 CDEBUG(D_MOUNT, "server put_super %s\n", tmpname);
1469 if (IS_MDT(lsi->lsi_ldd) && (lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC))
1470 snprintf(tmpname, tmpname_sz, "MGS");
1472 /* Stop the target */
1473 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1474 (IS_MDT(lsi->lsi_ldd) || IS_OST(lsi->lsi_ldd))) {
1475 struct lustre_profile *lprof = NULL;
1477 /* tell the mgc to drop the config log */
1478 lustre_end_log(sb, lsi->lsi_ldd->ldd_svname, NULL);
1480 /* COMPAT_146 - profile may get deleted in mgc_cleanup.
1481 If there are any setup/cleanup errors, save the lov
1482 name for safety cleanup later. */
1483 lprof = class_get_profile(lsi->lsi_ldd->ldd_svname);
1484 if (lprof && lprof->lp_dt) {
1485 OBD_ALLOC(extraname, strlen(lprof->lp_dt) + 1);
1486 strcpy(extraname, lprof->lp_dt);
1489 obd = class_name2obd(lsi->lsi_ldd->ldd_svname);
1491 CDEBUG(D_MOUNT, "stopping %s\n", obd->obd_name);
1492 if (lsi->lsi_flags & LSI_UMOUNT_FAILOVER)
1494 /* We can't seem to give an error return code
1495 * to .put_super, so we better make sure we clean up! */
1497 class_manual_cleanup(obd);
1499 CERROR("no obd %s\n", lsi->lsi_ldd->ldd_svname);
1500 server_deregister_mount(lsi->lsi_ldd->ldd_svname);
1504 /* If they wanted the mgs to stop separately from the mdt, they
1505 should have put it on a different device. */
1506 if (IS_MGS(lsi->lsi_ldd)) {
1507 /* if MDS start with --nomgs, don't stop MGS then */
1508 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS))
1509 server_stop_mgs(sb);
1512 /* Clean the mgc and sb */
1513 lustre_common_put_super(sb);
1515 /* Wait for the targets to really clean up - can't exit (and let the
1516 sb get destroyed) while the mount is still in use */
1517 server_wait_finished(mnt);
1519 /* drop the One True Mount */
1522 /* Stop the servers (MDS, OSS) if no longer needed. We must wait
1523 until the target is really gone so that our type refcount check
1525 server_stop_servers(lddflags, lsiflags);
1527 /* In case of startup or cleanup err, stop related obds */
1529 obd = class_name2obd(extraname);
1531 CWARN("Cleaning orphaned obd %s\n", extraname);
1533 class_manual_cleanup(obd);
1535 OBD_FREE(extraname, strlen(extraname) + 1);
1538 LCONSOLE_WARN("server umount %s complete\n", tmpname);
1539 OBD_FREE(tmpname, tmpname_sz);
1543 /** Called only for 'umount -f'
1545 #ifdef HAVE_UMOUNTBEGIN_VFSMOUNT
1546 static void server_umount_begin(struct vfsmount *vfsmnt, int flags)
1548 struct super_block *sb = vfsmnt->mnt_sb;
1550 static void server_umount_begin(struct super_block *sb)
1553 struct lustre_sb_info *lsi = s2lsi(sb);
1556 #ifdef HAVE_UMOUNTBEGIN_VFSMOUNT
1557 if (!(flags & MNT_FORCE)) {
1563 CDEBUG(D_MOUNT, "umount -f\n");
1564 /* umount = failover
1566 no third way to do non-force, non-failover */
1567 lsi->lsi_flags &= ~LSI_UMOUNT_FAILOVER;
1568 lsi->lsi_flags |= LSI_UMOUNT_FORCE;
1572 #ifndef HAVE_STATFS_DENTRY_PARAM
1573 static int server_statfs (struct super_block *sb, cfs_kstatfs_t *buf)
1576 static int server_statfs (struct dentry *dentry, cfs_kstatfs_t *buf)
1578 struct super_block *sb = dentry->d_sb;
1580 struct vfsmount *mnt = s2lsi(sb)->lsi_srv_mnt;
1583 if (mnt && mnt->mnt_sb && mnt->mnt_sb->s_op->statfs) {
1584 #ifdef HAVE_STATFS_DENTRY_PARAM
1585 int rc = mnt->mnt_sb->s_op->statfs(mnt->mnt_root, buf);
1587 int rc = mnt->mnt_sb->s_op->statfs(mnt->mnt_sb, buf);
1590 buf->f_type = sb->s_magic;
1596 buf->f_type = sb->s_magic;
1597 buf->f_bsize = sb->s_blocksize;
1603 buf->f_namelen = NAME_MAX;
1607 /** The operations we support directly on the superblock:
1608 * mount, umount, and df.
1610 static struct super_operations server_ops =
1612 .put_super = server_put_super,
1613 .umount_begin = server_umount_begin, /* umount -f */
1614 .statfs = server_statfs,
1617 #define log2(n) cfs_ffz(~(n))
1618 #define LUSTRE_SUPER_MAGIC 0x0BD00BD1
1620 static int server_fill_super_common(struct super_block *sb)
1622 struct inode *root = 0;
1625 CDEBUG(D_MOUNT, "Server sb, dev=%d\n", (int)sb->s_dev);
1627 sb->s_blocksize = 4096;
1628 sb->s_blocksize_bits = log2(sb->s_blocksize);
1629 sb->s_magic = LUSTRE_SUPER_MAGIC;
1630 sb->s_maxbytes = 0; //PAGE_CACHE_MAXBYTES;
1631 sb->s_flags |= MS_RDONLY;
1632 sb->s_op = &server_ops;
1634 root = new_inode(sb);
1636 CERROR("Can't make root inode\n");
1640 /* returns -EIO for every operation */
1641 /* make_bad_inode(root); -- badness - can't umount */
1642 /* apparently we need to be a directory for the mount to finish */
1643 root->i_mode = S_IFDIR;
1645 sb->s_root = d_alloc_root(root);
1647 CERROR("Can't make root dentry\n");
1655 /** Fill in the superblock info for a Lustre server.
1656 * Mount the device with the correct options.
1657 * Read the on-disk config file.
1658 * Start the services.
1660 static int server_fill_super(struct super_block *sb)
1662 struct lustre_sb_info *lsi = s2lsi(sb);
1663 struct vfsmount *mnt;
1667 /* the One True Mount */
1668 mnt = server_kernel_mount(sb);
1671 CERROR("Unable to mount device %s: %d\n",
1672 lsi->lsi_lmd->lmd_dev, rc);
1676 lsi->lsi_srv_mnt = mnt;
1678 LASSERT(lsi->lsi_ldd);
1679 CDEBUG(D_MOUNT, "Found service %s for fs '%s' on device %s\n",
1680 lsi->lsi_ldd->ldd_svname, lsi->lsi_ldd->ldd_fsname,
1681 lsi->lsi_lmd->lmd_dev);
1683 if (class_name2obd(lsi->lsi_ldd->ldd_svname)) {
1684 LCONSOLE_ERROR_MSG(0x161, "The target named %s is already "
1685 "running. Double-mount may have compromised"
1686 " the disk journal.\n",
1687 lsi->lsi_ldd->ldd_svname);
1693 /* Start MGS before MGC */
1694 if (IS_MGS(lsi->lsi_ldd) && !(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS)){
1695 rc = server_start_mgs(sb);
1700 /* Start MGC before servers */
1701 rc = lustre_start_mgc(sb);
1705 /* Set up all obd devices for service */
1706 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1707 (IS_OST(lsi->lsi_ldd) || IS_MDT(lsi->lsi_ldd))) {
1708 rc = server_start_targets(sb, mnt);
1710 CERROR("Unable to start targets: %d\n", rc);
1713 /* FIXME overmount client here,
1714 or can we just start a client log and client_fill_super on this sb?
1715 We need to make sure server_put_super gets called too - ll_put_super
1716 calls lustre_common_put_super; check there for LSI_SERVER flag,
1718 Probably should start client from new thread so we can return.
1719 Client will not finish until all servers are connected.
1720 Note - MGS-only server does NOT get a client, since there is no
1721 lustre fs associated - the MGS is for all lustre fs's */
1724 rc = server_fill_super_common(sb);
1730 /* We jump here in case of failure while starting targets or MGS.
1731 * In this case we can't just put @mnt and have to do real cleanup
1732 * with stoping targets, etc. */
1733 server_put_super(sb);
1737 /* Get the index from the obd name.
1738 rc = server type, or
1740 if endptr isn't NULL it is set to end of name */
1741 int server_name2index(char *svname, __u32 *idx, char **endptr)
1743 unsigned long index;
1745 char *dash = strrchr(svname, '-');
1749 /* intepret <fsname>-MDTXXXXX-mdc as mdt, the better way is to pass
1750 * in the fsname, then determine the server index */
1751 if (!strcmp(LUSTRE_MDC_NAME, dash + 1)) {
1753 for (; dash > svname && *dash != '-'; dash--);
1758 if (strncmp(dash + 1, "MDT", 3) == 0)
1759 rc = LDD_F_SV_TYPE_MDT;
1760 else if (strncmp(dash + 1, "OST", 3) == 0)
1761 rc = LDD_F_SV_TYPE_OST;
1764 if (strcmp(dash + 4, "all") == 0)
1765 return rc | LDD_F_SV_ALL;
1767 index = simple_strtoul(dash + 4, endptr, 16);
1772 /*************** mount common betweeen server and client ***************/
1775 int lustre_common_put_super(struct super_block *sb)
1780 CDEBUG(D_MOUNT, "dropping sb %p\n", sb);
1782 /* Drop a ref to the MGC */
1783 rc = lustre_stop_mgc(sb);
1784 if (rc && (rc != -ENOENT)) {
1786 CERROR("Can't stop MGC: %d\n", rc);
1789 /* BUSY just means that there's some other obd that
1790 needs the mgc. Let him clean it up. */
1791 CDEBUG(D_MOUNT, "MGC still in use\n");
1793 /* Drop a ref to the mounted disk */
1799 static void lmd_print(struct lustre_mount_data *lmd)
1803 PRINT_CMD(PRINT_MASK, " mount data:\n");
1804 if (lmd_is_client(lmd))
1805 PRINT_CMD(PRINT_MASK, "profile: %s\n", lmd->lmd_profile);
1806 PRINT_CMD(PRINT_MASK, "device: %s\n", lmd->lmd_dev);
1807 PRINT_CMD(PRINT_MASK, "flags: %x\n", lmd->lmd_flags);
1810 PRINT_CMD(PRINT_MASK, "options: %s\n", lmd->lmd_opts);
1812 if (lmd->lmd_recovery_time_soft)
1813 PRINT_CMD(PRINT_MASK, "recovery time soft: %d\n",
1814 lmd->lmd_recovery_time_soft);
1816 if (lmd->lmd_recovery_time_hard)
1817 PRINT_CMD(PRINT_MASK, "recovery time hard: %d\n",
1818 lmd->lmd_recovery_time_hard);
1820 for (i = 0; i < lmd->lmd_exclude_count; i++) {
1821 PRINT_CMD(PRINT_MASK, "exclude %d: OST%04x\n", i,
1822 lmd->lmd_exclude[i]);
1826 /* Is this server on the exclusion list */
1827 int lustre_check_exclusion(struct super_block *sb, char *svname)
1829 struct lustre_sb_info *lsi = s2lsi(sb);
1830 struct lustre_mount_data *lmd = lsi->lsi_lmd;
1835 rc = server_name2index(svname, &index, NULL);
1836 if (rc != LDD_F_SV_TYPE_OST)
1837 /* Only exclude OSTs */
1840 CDEBUG(D_MOUNT, "Check exclusion %s (%d) in %d of %s\n", svname,
1841 index, lmd->lmd_exclude_count, lmd->lmd_dev);
1843 for(i = 0; i < lmd->lmd_exclude_count; i++) {
1844 if (index == lmd->lmd_exclude[i]) {
1845 CWARN("Excluding %s (on exclusion list)\n", svname);
1852 /* mount -v -o exclude=lustre-OST0001:lustre-OST0002 -t lustre ... */
1853 static int lmd_make_exclusion(struct lustre_mount_data *lmd, char *ptr)
1855 char *s1 = ptr, *s2;
1856 __u32 index, *exclude_list;
1860 /* The shortest an ost name can be is 8 chars: -OST0000.
1861 We don't actually know the fsname at this time, so in fact
1862 a user could specify any fsname. */
1863 devmax = strlen(ptr) / 8 + 1;
1865 /* temp storage until we figure out how many we have */
1866 OBD_ALLOC(exclude_list, sizeof(index) * devmax);
1870 /* we enter this fn pointing at the '=' */
1871 while (*s1 && *s1 != ' ' && *s1 != ',') {
1873 rc = server_name2index(s1, &index, &s2);
1875 CERROR("Can't parse server name '%s'\n", s1);
1878 if (rc == LDD_F_SV_TYPE_OST)
1879 exclude_list[lmd->lmd_exclude_count++] = index;
1881 CDEBUG(D_MOUNT, "ignoring exclude %.7s\n", s1);
1883 /* now we are pointing at ':' (next exclude)
1884 or ',' (end of excludes) */
1885 if (lmd->lmd_exclude_count >= devmax)
1888 if (rc >= 0) /* non-err */
1891 if (lmd->lmd_exclude_count) {
1892 /* permanent, freed in lustre_free_lsi */
1893 OBD_ALLOC(lmd->lmd_exclude, sizeof(index) *
1894 lmd->lmd_exclude_count);
1895 if (lmd->lmd_exclude) {
1896 memcpy(lmd->lmd_exclude, exclude_list,
1897 sizeof(index) * lmd->lmd_exclude_count);
1900 lmd->lmd_exclude_count = 0;
1903 OBD_FREE(exclude_list, sizeof(index) * devmax);
1907 static int lmd_parse_mgssec(struct lustre_mount_data *lmd, char *ptr)
1912 if (lmd->lmd_mgssec != NULL) {
1913 OBD_FREE(lmd->lmd_mgssec, strlen(lmd->lmd_mgssec) + 1);
1914 lmd->lmd_mgssec = NULL;
1917 tail = strchr(ptr, ',');
1919 length = strlen(ptr);
1921 length = tail - ptr;
1923 OBD_ALLOC(lmd->lmd_mgssec, length + 1);
1924 if (lmd->lmd_mgssec == NULL)
1927 memcpy(lmd->lmd_mgssec, ptr, length);
1928 lmd->lmd_mgssec[length] = '\0';
1932 /** Parse mount line options
1933 * e.g. mount -v -t lustre -o abort_recov uml1:uml2:/lustre-client /mnt/lustre
1934 * dev is passed as device=uml1:/lustre by mount.lustre
1936 static int lmd_parse(char *options, struct lustre_mount_data *lmd)
1938 char *s1, *s2, *devname = NULL;
1939 struct lustre_mount_data *raw = (struct lustre_mount_data *)options;
1945 LCONSOLE_ERROR_MSG(0x162, "Missing mount data: check that "
1946 "/sbin/mount.lustre is installed.\n");
1950 /* Options should be a string - try to detect old lmd data */
1951 if ((raw->lmd_magic & 0xffffff00) == (LMD_MAGIC & 0xffffff00)) {
1952 LCONSOLE_ERROR_MSG(0x163, "You're using an old version of "
1953 "/sbin/mount.lustre. Please install "
1954 "version %s\n", LUSTRE_VERSION_STRING);
1957 lmd->lmd_magic = LMD_MAGIC;
1959 /* Set default flags here */
1964 int time_min = 2 * (CONNECTION_SWITCH_MAX +
1965 2 * INITIAL_CONNECT_TIMEOUT);
1967 /* Skip whitespace and extra commas */
1968 while (*s1 == ' ' || *s1 == ',')
1971 /* Client options are parsed in ll_options: eg. flock,
1974 /* Parse non-ldiskfs options here. Rather than modifying
1975 ldiskfs, we just zero these out here */
1976 if (strncmp(s1, "abort_recov", 11) == 0) {
1977 lmd->lmd_flags |= LMD_FLG_ABORT_RECOV;
1979 } else if (strncmp(s1, "recovery_time_soft=", 19) == 0) {
1980 lmd->lmd_recovery_time_soft = max_t(int,
1981 simple_strtoul(s1 + 19, NULL, 10), time_min);
1983 } else if (strncmp(s1, "recovery_time_hard=", 19) == 0) {
1984 lmd->lmd_recovery_time_hard = max_t(int,
1985 simple_strtoul(s1 + 19, NULL, 10), time_min);
1987 } else if (strncmp(s1, "nosvc", 5) == 0) {
1988 lmd->lmd_flags |= LMD_FLG_NOSVC;
1990 } else if (strncmp(s1, "nomgs", 5) == 0) {
1991 lmd->lmd_flags |= LMD_FLG_NOMGS;
1993 } else if (strncmp(s1, "writeconf", 9) == 0) {
1994 lmd->lmd_flags |= LMD_FLG_WRITECONF;
1996 } else if (strncmp(s1, "mgssec=", 7) == 0) {
1997 rc = lmd_parse_mgssec(lmd, s1 + 7);
2001 /* ost exclusion list */
2002 } else if (strncmp(s1, "exclude=", 8) == 0) {
2003 rc = lmd_make_exclusion(lmd, s1 + 7);
2008 /* Linux 2.4 doesn't pass the device, so we stuck it at the
2009 end of the options. */
2010 else if (strncmp(s1, "device=", 7) == 0) {
2012 /* terminate options right before device. device
2013 must be the last one. */
2019 s2 = strchr(s1, ',');
2027 memmove(s1, s2, strlen(s2) + 1);
2033 LCONSOLE_ERROR_MSG(0x164, "Can't find the device name "
2034 "(need mount option 'device=...')\n");
2038 s1 = strstr(devname, ":/");
2041 lmd->lmd_flags = LMD_FLG_CLIENT;
2042 /* Remove leading /s from fsname */
2043 while (*++s1 == '/') ;
2044 /* Freed in lustre_free_lsi */
2045 OBD_ALLOC(lmd->lmd_profile, strlen(s1) + 8);
2046 if (!lmd->lmd_profile)
2048 sprintf(lmd->lmd_profile, "%s-client", s1);
2051 /* Freed in lustre_free_lsi */
2052 OBD_ALLOC(lmd->lmd_dev, strlen(devname) + 1);
2055 strcpy(lmd->lmd_dev, devname);
2057 /* Save mount options */
2058 s1 = options + strlen(options) - 1;
2059 while (s1 >= options && (*s1 == ',' || *s1 == ' '))
2061 if (*options != 0) {
2062 /* Freed in lustre_free_lsi */
2063 OBD_ALLOC(lmd->lmd_opts, strlen(options) + 1);
2066 strcpy(lmd->lmd_opts, options);
2070 lmd->lmd_magic = LMD_MAGIC;
2075 CERROR("Bad mount options %s\n", options);
2080 /** This is the entry point for the mount call into Lustre.
2081 * This is called when a server or client is mounted,
2082 * and this is where we start setting things up.
2083 * @param data Mount options (e.g. -o flock,abort_recov)
2085 int lustre_fill_super(struct super_block *sb, void *data, int silent)
2087 struct lustre_mount_data *lmd;
2088 struct lustre_sb_info *lsi;
2092 CDEBUG(D_MOUNT|D_VFSTRACE, "VFS Op: sb %p\n", sb);
2094 lsi = lustre_init_lsi(sb);
2100 * Disable lockdep during mount, because mount locking patterns are
2105 /* Figure out the lmd from the mount options */
2106 if (lmd_parse((char *)data, lmd)) {
2108 GOTO(out, rc = -EINVAL);
2111 if (lmd_is_client(lmd)) {
2112 CDEBUG(D_MOUNT, "Mounting client %s\n", lmd->lmd_profile);
2113 if (!client_fill_super) {
2114 LCONSOLE_ERROR_MSG(0x165, "Nothing registered for "
2115 "client mount! Is the 'lustre' "
2116 "module loaded?\n");
2120 rc = lustre_start_mgc(sb);
2125 /* Connect and start */
2126 /* (should always be ll_fill_super) */
2127 rc = (*client_fill_super)(sb);
2128 /* c_f_s will call lustre_common_put_super on failure */
2131 CDEBUG(D_MOUNT, "Mounting server from %s\n", lmd->lmd_dev);
2132 lsi->lsi_flags |= LSI_SERVER;
2133 rc = server_fill_super(sb);
2134 /* s_f_s calls lustre_start_mgc after the mount because we need
2135 the MGS nids which are stored on disk. Plus, we may
2136 need to start the MGS first. */
2137 /* s_f_s will call server_put_super on failure */
2140 /* If error happens in fill_super() call, @lsi will be killed there.
2141 * This is why we do not put it here. */
2145 CERROR("Unable to mount %s (%d)\n",
2146 s2lsi(sb) ? lmd->lmd_dev : "", rc);
2148 CDEBUG(D_SUPER, "Mount %s complete\n",
2156 /* We can't call ll_fill_super by name because it lives in a module that
2157 must be loaded after this one. */
2158 void lustre_register_client_fill_super(int (*cfs)(struct super_block *sb))
2160 client_fill_super = cfs;
2163 void lustre_register_kill_super_cb(void (*cfs)(struct super_block *sb))
2165 kill_super_cb = cfs;
2168 /***************** FS registration ******************/
2170 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,18))
2171 struct super_block * lustre_get_sb(struct file_system_type *fs_type,
2172 int flags, const char *devname, void * data)
2174 return get_sb_nodev(fs_type, flags, data, lustre_fill_super);
2177 int lustre_get_sb(struct file_system_type *fs_type,
2178 int flags, const char *devname, void * data,
2179 struct vfsmount *mnt)
2181 return get_sb_nodev(fs_type, flags, data, lustre_fill_super, mnt);
2185 void lustre_kill_super(struct super_block *sb)
2187 struct lustre_sb_info *lsi = s2lsi(sb);
2189 if (kill_super_cb && lsi && !(lsi->lsi_flags & LSI_SERVER))
2190 (*kill_super_cb)(sb);
2192 kill_anon_super(sb);
2195 /** Register the "lustre" fs type
2197 struct file_system_type lustre_fs_type = {
2198 .owner = THIS_MODULE,
2200 .get_sb = lustre_get_sb,
2201 .kill_sb = lustre_kill_super,
2202 .fs_flags = FS_BINARY_MOUNTDATA | FS_REQUIRES_DEV |
2203 #ifdef FS_HAS_FIEMAP
2206 LL_RENAME_DOES_D_MOVE,
2209 int lustre_register_fs(void)
2211 return register_filesystem(&lustre_fs_type);
2214 int lustre_unregister_fs(void)
2216 return unregister_filesystem(&lustre_fs_type);
2219 EXPORT_SYMBOL(lustre_register_client_fill_super);
2220 EXPORT_SYMBOL(lustre_register_kill_super_cb);
2221 EXPORT_SYMBOL(lustre_common_put_super);
2222 EXPORT_SYMBOL(lustre_process_log);
2223 EXPORT_SYMBOL(lustre_end_log);
2224 EXPORT_SYMBOL(server_get_mount);
2225 EXPORT_SYMBOL(server_get_mount_2);
2226 EXPORT_SYMBOL(server_put_mount);
2227 EXPORT_SYMBOL(server_put_mount_2);
2228 EXPORT_SYMBOL(server_register_target);
2229 EXPORT_SYMBOL(server_name2index);
2230 EXPORT_SYMBOL(server_mti_print);
2231 EXPORT_SYMBOL(do_lcfg);