4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2013, 2015, Intel Corporation.
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_server.c
38 * 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)
49 #include <linux/types.h>
50 #include <linux/selinux.h>
51 #include <linux/statfs.h>
52 #include <linux/version.h>
53 #ifdef HAVE_KERNEL_LOCKED
54 #include <linux/smp_lock.h>
57 #include <lustre/lustre_idl.h>
58 #include <lustre/lustre_user.h>
60 #include <llog_swab.h>
61 #include <lustre_disk.h>
62 #include <lustre_ioctl.h>
63 #include <lustre_log.h>
64 #include <lustre_param.h>
66 #include <obd_class.h>
68 /*********** mount lookup *********/
70 static DEFINE_MUTEX(lustre_mount_info_lock);
71 static struct list_head server_mount_info_list =
72 LIST_HEAD_INIT(server_mount_info_list);
74 static struct lustre_mount_info *server_find_mount(const char *name)
76 struct list_head *tmp;
77 struct lustre_mount_info *lmi;
80 list_for_each(tmp, &server_mount_info_list) {
81 lmi = list_entry(tmp, struct lustre_mount_info,
83 if (strcmp(name, lmi->lmi_name) == 0)
89 /* we must register an obd for a mount before we call the setup routine.
90 *_setup will call lustre_get_mount to get the mnt struct
91 by obd_name, since we can't pass the pointer to setup. */
92 static int server_register_mount(const char *name, struct super_block *sb)
94 struct lustre_mount_info *lmi;
100 OBD_ALLOC(lmi, sizeof(*lmi));
103 OBD_ALLOC(name_cp, strlen(name) + 1);
105 OBD_FREE(lmi, sizeof(*lmi));
108 strcpy(name_cp, name);
110 mutex_lock(&lustre_mount_info_lock);
112 if (server_find_mount(name)) {
113 mutex_unlock(&lustre_mount_info_lock);
114 OBD_FREE(lmi, sizeof(*lmi));
115 OBD_FREE(name_cp, strlen(name) + 1);
116 CERROR("Already registered %s\n", name);
119 lmi->lmi_name = name_cp;
121 list_add(&lmi->lmi_list_chain, &server_mount_info_list);
123 mutex_unlock(&lustre_mount_info_lock);
125 CDEBUG(D_MOUNT, "register mount %p from %s\n", sb, name);
130 /* when an obd no longer needs a mount */
131 static int server_deregister_mount(const char *name)
133 struct lustre_mount_info *lmi;
136 mutex_lock(&lustre_mount_info_lock);
137 lmi = server_find_mount(name);
139 mutex_unlock(&lustre_mount_info_lock);
140 CERROR("%s not registered\n", name);
144 CDEBUG(D_MOUNT, "deregister mount %p from %s\n", lmi->lmi_sb, name);
146 OBD_FREE(lmi->lmi_name, strlen(lmi->lmi_name) + 1);
147 list_del(&lmi->lmi_list_chain);
148 OBD_FREE(lmi, sizeof(*lmi));
149 mutex_unlock(&lustre_mount_info_lock);
154 /* obd's look up a registered mount using their obdname. This is just
155 for initial obd setup to find the mount struct. It should not be
156 called every time you want to mntget. */
157 struct lustre_mount_info *server_get_mount(const char *name)
159 struct lustre_mount_info *lmi;
160 struct lustre_sb_info *lsi;
163 mutex_lock(&lustre_mount_info_lock);
164 lmi = server_find_mount(name);
165 mutex_unlock(&lustre_mount_info_lock);
167 CERROR("Can't find mount for %s\n", name);
170 lsi = s2lsi(lmi->lmi_sb);
172 atomic_inc(&lsi->lsi_mounts);
174 CDEBUG(D_MOUNT, "get mount %p from %s, refs=%d\n", lmi->lmi_sb,
175 name, atomic_read(&lsi->lsi_mounts));
179 EXPORT_SYMBOL(server_get_mount);
182 * server_put_mount: to be called from obd_cleanup methods
184 * @dereg_mnt: 0 or 1 depending on whether the mount is to be deregistered or
187 * The caller decides whether server_deregister_mount() needs to be called or
188 * not. Calling of server_deregister_mount() does not depend on refcounting on
189 * lsi because we could have say the mgs and mds on the same node and we
190 * unmount the mds, then the ref on the lsi would still be non-zero but we
191 * would still want to deregister the mds mount.
193 int server_put_mount(const char *name, bool dereg_mnt)
195 struct lustre_mount_info *lmi;
196 struct lustre_sb_info *lsi;
199 mutex_lock(&lustre_mount_info_lock);
200 lmi = server_find_mount(name);
201 mutex_unlock(&lustre_mount_info_lock);
203 CERROR("Can't find mount for %s\n", name);
206 lsi = s2lsi(lmi->lmi_sb);
208 CDEBUG(D_MOUNT, "put mount %p from %s, refs=%d\n",
209 lmi->lmi_sb, name, atomic_read(&lsi->lsi_mounts));
211 if (lustre_put_lsi(lmi->lmi_sb))
212 CDEBUG(D_MOUNT, "Last put of mount %p from %s\n",
216 /* this obd should never need the mount again */
217 server_deregister_mount(name);
221 EXPORT_SYMBOL(server_put_mount);
223 /* Set up a MGS to serve startup logs */
224 static int server_start_mgs(struct super_block *sb)
226 struct lustre_sb_info *lsi = s2lsi(sb);
227 struct lustre_mount_info *lmi;
231 /* It is impossible to have more than 1 MGS per node, since
232 MGC wouldn't know which to connect to */
233 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
235 lsi = s2lsi(lmi->lmi_sb);
236 LCONSOLE_ERROR_MSG(0x15d, "The MGS service was already started"
241 CDEBUG(D_CONFIG, "Start MGS service %s\n", LUSTRE_MGS_OBDNAME);
243 rc = server_register_mount(LUSTRE_MGS_OBDNAME, sb);
246 rc = lustre_start_simple(LUSTRE_MGS_OBDNAME, LUSTRE_MGS_NAME,
247 LUSTRE_MGS_OBDNAME, NULL, NULL,
248 lsi->lsi_osd_obdname, NULL);
249 /* server_deregister_mount() is not called previously, for lsi
250 * and other stuff can't be freed cleanly when mgs calls
251 * server_put_mount() in error handling case (see b=17758),
252 * this problem is caused by a bug in mgs_init0, which forgot
253 * calling server_put_mount in error case. */
256 server_deregister_mount(LUSTRE_MGS_OBDNAME);
260 LCONSOLE_ERROR_MSG(0x15e, "Failed to start MGS '%s' (%d). "
261 "Is the 'mgs' module loaded?\n",
262 LUSTRE_MGS_OBDNAME, rc);
266 static int server_stop_mgs(struct super_block *sb)
268 struct obd_device *obd;
270 struct lustre_mount_info *lmi;
273 /* Do not stop MGS if this device is not the running MGT */
274 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
275 if (lmi != NULL && lmi->lmi_sb != sb)
278 CDEBUG(D_MOUNT, "Stop MGS service %s\n", LUSTRE_MGS_OBDNAME);
280 /* There better be only one MGS */
281 obd = class_name2obd(LUSTRE_MGS_OBDNAME);
283 CDEBUG(D_CONFIG, "mgs %s not running\n", LUSTRE_MGS_OBDNAME);
287 /* The MGS should always stop when we say so */
289 rc = class_manual_cleanup(obd);
293 /* Since there's only one mgc per node, we have to change it's fs to get
294 access to the right disk. */
295 static int server_mgc_set_fs(const struct lu_env *env,
296 struct obd_device *mgc, struct super_block *sb)
298 struct lustre_sb_info *lsi = s2lsi(sb);
302 CDEBUG(D_MOUNT, "Set mgc disk for %s\n", lsi->lsi_lmd->lmd_dev);
304 /* cl_mgc_sem in mgc insures we sleep if the mgc_fs is busy */
305 rc = obd_set_info_async(env, mgc->obd_self_export,
306 sizeof(KEY_SET_FS), KEY_SET_FS,
307 sizeof(*sb), sb, NULL);
309 CERROR("can't set_fs %d\n", rc);
314 static int server_mgc_clear_fs(const struct lu_env *env,
315 struct obd_device *mgc)
320 CDEBUG(D_MOUNT, "Unassign mgc disk\n");
322 rc = obd_set_info_async(env, mgc->obd_self_export,
323 sizeof(KEY_CLEAR_FS), KEY_CLEAR_FS,
328 static inline bool is_mdc_device(const char *devname)
332 ptr = strrchr(devname, '-');
333 return ptr != NULL && strcmp(ptr, "-mdc") == 0;
336 static inline bool tgt_is_mdt(const char *tgtname, __u32 *idx)
340 type = server_name2index(tgtname, idx, NULL);
342 return type == LDD_F_SV_TYPE_MDT;
346 * Convert OST/MDT name(fsname-{MDT,OST}xxxx) to a lwp name with the @idx:yyyy
347 * (fsname-MDTyyyy-lwp-{MDT,OST}xxxx)
349 int tgt_name2lwp_name(const char *tgt_name, char *lwp_name, int len, __u32 idx)
356 OBD_ALLOC(fsname, MTI_NAME_MAXLEN);
360 rc = server_name2fsname(tgt_name, fsname, &tgt);
362 CERROR("%s: failed to get fsname from tgt_name: rc = %d\n",
367 if (*tgt != '-' && *tgt != ':') {
368 CERROR("%s: invalid tgt_name name!\n", tgt_name);
369 GOTO(cleanup, rc = -EINVAL);
373 if (strncmp(tgt, "OST", 3) != 0 && strncmp(tgt, "MDT", 3) != 0) {
374 CERROR("%s is not an OST or MDT target!\n", tgt_name);
375 GOTO(cleanup, rc = -EINVAL);
377 snprintf(lwp_name, len, "%s-MDT%04x-%s-%s",
378 fsname, idx, LUSTRE_LWP_NAME, tgt);
380 GOTO(cleanup, rc = 0);
384 OBD_FREE(fsname, MTI_NAME_MAXLEN);
388 EXPORT_SYMBOL(tgt_name2lwp_name);
390 static struct list_head lwp_register_list =
391 LIST_HEAD_INIT(lwp_register_list);
392 static DEFINE_SPINLOCK(lwp_register_list_lock);
394 static void lustre_put_lwp_item(struct lwp_register_item *lri)
396 if (atomic_dec_and_test(&lri->lri_ref)) {
397 LASSERT(list_empty(&lri->lri_list));
399 if (*lri->lri_exp != NULL)
400 class_export_put(*lri->lri_exp);
405 int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
406 register_lwp_cb cb_func, void *cb_data)
408 struct obd_device *lwp;
409 struct lwp_register_item *lri;
412 LASSERTF(strlen(lwpname) < MTI_NAME_MAXLEN, "lwpname is too long %s\n",
414 LASSERT(exp != NULL && *exp == NULL);
420 lwp = class_name2obd(lwpname);
421 if (lwp != NULL && lwp->obd_set_up == 1) {
422 struct obd_uuid *uuid;
429 memcpy(uuid->uuid, lwpname, strlen(lwpname));
430 *exp = cfs_hash_lookup(lwp->obd_uuid_hash, uuid);
434 memcpy(lri->lri_name, lwpname, strlen(lwpname));
436 lri->lri_cb_func = cb_func;
437 lri->lri_cb_data = cb_data;
438 INIT_LIST_HEAD(&lri->lri_list);
440 * Initialize the lri_ref at 2, one will be released before
441 * current function returned via lustre_put_lwp_item(), the
442 * other will be released in lustre_deregister_lwp_item().
444 atomic_set(&lri->lri_ref, 2);
446 spin_lock(&lwp_register_list_lock);
447 list_add(&lri->lri_list, &lwp_register_list);
448 spin_unlock(&lwp_register_list_lock);
450 if (*exp != NULL && cb_func != NULL)
452 lustre_put_lwp_item(lri);
456 EXPORT_SYMBOL(lustre_register_lwp_item);
458 void lustre_deregister_lwp_item(struct obd_export **exp)
460 struct lwp_register_item *lri;
462 spin_lock(&lwp_register_list_lock);
463 list_for_each_entry(lri, &lwp_register_list, lri_list) {
464 if (exp == lri->lri_exp) {
465 list_del_init(&lri->lri_list);
466 spin_unlock(&lwp_register_list_lock);
468 lustre_put_lwp_item(lri);
472 spin_unlock(&lwp_register_list_lock);
474 EXPORT_SYMBOL(lustre_deregister_lwp_item);
476 struct obd_export *lustre_find_lwp_by_index(const char *dev, __u32 idx)
478 struct lustre_mount_info *lmi;
479 struct lustre_sb_info *lsi;
480 struct obd_device *lwp;
481 struct obd_export *exp = NULL;
486 lmi = server_get_mount(dev);
490 lsi = s2lsi(lmi->lmi_sb);
491 rc = server_name2fsname(lsi->lsi_svname, fsname, NULL);
493 CERROR("%s: failed to get fsname: rc = %d\n",
494 lsi->lsi_svname, rc);
498 snprintf(lwp_name, sizeof(lwp_name), "%s-MDT%04x", fsname, idx);
499 spin_lock(&lsi->lsi_lwp_lock);
500 list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
501 char *ptr = strstr(lwp->obd_name, lwp_name);
503 if (ptr != NULL && lwp->obd_lwp_export != NULL) {
504 exp = class_export_get(lwp->obd_lwp_export);
508 spin_unlock(&lsi->lsi_lwp_lock);
511 server_put_mount(dev, false);
515 EXPORT_SYMBOL(lustre_find_lwp_by_index);
517 void lustre_notify_lwp_list(struct obd_export *exp)
519 struct lwp_register_item *lri;
520 LASSERT(exp != NULL);
523 spin_lock(&lwp_register_list_lock);
524 list_for_each_entry(lri, &lwp_register_list, lri_list) {
525 if (strcmp(exp->exp_obd->obd_name, lri->lri_name))
527 if (*lri->lri_exp != NULL)
529 *lri->lri_exp = class_export_get(exp);
530 atomic_inc(&lri->lri_ref);
531 spin_unlock(&lwp_register_list_lock);
533 if (lri->lri_cb_func != NULL)
534 lri->lri_cb_func(lri->lri_cb_data);
535 lustre_put_lwp_item(lri);
537 /* Others may have changed the list after we unlock, we have
538 * to rescan the list from the beginning. Usually, the list
539 * 'lwp_register_list' is very short, and there is 'guard'
540 * lri::lri_exp that will prevent the callback to be done
541 * repeatedly. So rescanning the list has no problem. */
544 spin_unlock(&lwp_register_list_lock);
546 EXPORT_SYMBOL(lustre_notify_lwp_list);
548 static int lustre_lwp_connect(struct obd_device *lwp)
551 struct lu_context session_ctx;
552 struct obd_export *exp;
553 struct obd_uuid *uuid = NULL;
554 struct obd_connect_data *data = NULL;
558 /* log has been fully processed, let clients connect */
559 rc = lu_env_init(&env, lwp->obd_lu_dev->ld_type->ldt_ctx_tags);
563 lu_context_init(&session_ctx, LCT_SERVER_SESSION);
564 session_ctx.lc_thread = NULL;
565 lu_context_enter(&session_ctx);
566 env.le_ses = &session_ctx;
570 GOTO(out, rc = -ENOMEM);
572 data->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_INDEX;
573 data->ocd_version = LUSTRE_VERSION_CODE;
574 data->ocd_connect_flags |= OBD_CONNECT_MDS_MDS | OBD_CONNECT_FID |
575 OBD_CONNECT_AT | OBD_CONNECT_LRU_RESIZE |
576 OBD_CONNECT_FULL20 | OBD_CONNECT_LVB_TYPE |
577 OBD_CONNECT_LIGHTWEIGHT | OBD_CONNECT_LFSCK |
578 OBD_CONNECT_BULK_MBITS;
581 GOTO(out, rc = -ENOMEM);
583 if (strlen(lwp->obd_name) > sizeof(uuid->uuid)) {
584 CERROR("%s: Too long lwp name %s, max_size is %d\n",
585 lwp->obd_name, lwp->obd_name, (int)sizeof(uuid->uuid));
586 GOTO(out, rc = -EINVAL);
589 /* Use lwp name as the uuid, so we find the export by lwp name later */
590 memcpy(uuid->uuid, lwp->obd_name, strlen(lwp->obd_name));
591 rc = obd_connect(&env, &exp, lwp, uuid, data, NULL);
593 CERROR("%s: connect failed: rc = %d\n", lwp->obd_name, rc);
595 if (unlikely(lwp->obd_lwp_export != NULL))
596 class_export_put(lwp->obd_lwp_export);
597 lwp->obd_lwp_export = class_export_get(exp);
609 lu_context_exit(&session_ctx);
610 lu_context_fini(&session_ctx);
616 * lwp is used by slaves (Non-MDT0 targets) to manage the connection to MDT0,
617 * or from the OSTx to MDTy.
619 static int lustre_lwp_setup(struct lustre_cfg *lcfg, struct lustre_sb_info *lsi,
622 struct obd_device *obd;
623 char *lwpname = NULL;
624 char *lwpuuid = NULL;
628 rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
631 CERROR("%s: Can't add uuid: rc =%d\n", lsi->lsi_svname, rc);
635 OBD_ALLOC(lwpname, MTI_NAME_MAXLEN);
637 GOTO(out, rc = -ENOMEM);
639 rc = tgt_name2lwp_name(lsi->lsi_svname, lwpname, MTI_NAME_MAXLEN, idx);
641 CERROR("%s: failed to generate lwp name: rc = %d\n",
642 lsi->lsi_svname, rc);
646 OBD_ALLOC(lwpuuid, MTI_NAME_MAXLEN);
648 GOTO(out, rc = -ENOMEM);
650 sprintf(lwpuuid, "%s_UUID", lwpname);
651 rc = lustre_start_simple(lwpname, LUSTRE_LWP_NAME,
652 lwpuuid, lustre_cfg_string(lcfg, 1),
655 CERROR("%s: setup up failed: rc %d\n", lwpname, rc);
659 obd = class_name2obd(lwpname);
660 LASSERT(obd != NULL);
662 rc = lustre_lwp_connect(obd);
664 obd->u.cli.cl_max_mds_easize = MAX_MD_SIZE;
665 spin_lock(&lsi->lsi_lwp_lock);
666 list_add_tail(&obd->obd_lwp_list, &lsi->lsi_lwp_list);
667 spin_unlock(&lsi->lsi_lwp_lock);
669 CERROR("%s: connect failed: rc = %d\n", lwpname, rc);
676 OBD_FREE(lwpname, MTI_NAME_MAXLEN);
678 OBD_FREE(lwpuuid, MTI_NAME_MAXLEN);
683 /* the caller is responsible for memory free */
684 static struct obd_device *lustre_find_lwp(struct lustre_sb_info *lsi,
685 char **lwpname, __u32 idx)
687 struct obd_device *lwp;
691 LASSERT(lwpname != NULL);
692 LASSERT(IS_OST(lsi) || IS_MDT(lsi));
694 OBD_ALLOC(*lwpname, MTI_NAME_MAXLEN);
695 if (*lwpname == NULL)
696 RETURN(ERR_PTR(-ENOMEM));
698 rc = tgt_name2lwp_name(lsi->lsi_svname, *lwpname, MTI_NAME_MAXLEN, idx);
700 CERROR("%s: failed to generate lwp name: rc = %d\n",
701 lsi->lsi_svname, rc);
702 GOTO(out, rc = -EINVAL);
705 lwp = class_name2obd(*lwpname);
709 if (*lwpname != NULL) {
710 OBD_FREE(*lwpname, MTI_NAME_MAXLEN);
716 RETURN(lwp != NULL ? lwp : ERR_PTR(-ENOENT));
719 static int lustre_lwp_add_conn(struct lustre_cfg *cfg,
720 struct lustre_sb_info *lsi, __u32 idx)
722 struct lustre_cfg_bufs *bufs = NULL;
723 struct lustre_cfg *lcfg = NULL;
724 char *lwpname = NULL;
725 struct obd_device *lwp;
729 lwp = lustre_find_lwp(lsi, &lwpname, idx);
731 CERROR("%s: can't find lwp device.\n", lsi->lsi_svname);
732 GOTO(out, rc = PTR_ERR(lwp));
734 LASSERT(lwpname != NULL);
738 GOTO(out, rc = -ENOMEM);
740 lustre_cfg_bufs_reset(bufs, lwpname);
741 lustre_cfg_bufs_set_string(bufs, 1,
742 lustre_cfg_string(cfg, 1));
744 lcfg = lustre_cfg_new(LCFG_ADD_CONN, bufs);
746 GOTO(out_cfg, rc = -ENOMEM);
747 rc = class_add_conn(lwp, lcfg);
749 CERROR("%s: can't add conn: rc = %d\n", lwpname, rc);
752 lustre_cfg_free(lcfg);
758 OBD_FREE(lwpname, MTI_NAME_MAXLEN);
763 * Retrieve MDT nids from the client log, then start the lwp device.
764 * there are only two scenarios which would include mdt nid.
766 * marker 5 (flags=0x01, v2.1.54.0) lustre-MDTyyyy 'add mdc' xxx-
767 * add_uuid nid=192.168.122.162@tcp(0x20000c0a87aa2) 0: 1:192.168.122.162@tcp
768 * attach 0:lustre-MDTyyyy-mdc 1:mdc 2:lustre-clilmv_UUID
769 * setup 0:lustre-MDTyyyy-mdc 1:lustre-MDTyyyy_UUID 2:192.168.122.162@tcp
770 * add_uuid nid=192.168.172.1@tcp(0x20000c0a8ac01) 0: 1:192.168.172.1@tcp
771 * add_conn 0:lustre-MDTyyyy-mdc 1:192.168.172.1@tcp
772 * modify_mdc_tgts add 0:lustre-clilmv 1:lustre-MDTyyyy_UUID xxxx
773 * marker 5 (flags=0x02, v2.1.54.0) lustre-MDTyyyy 'add mdc' xxxx-
775 * marker 7 (flags=0x01, v2.1.54.0) lustre-MDTyyyy 'add failnid' xxxx-
776 * add_uuid nid=192.168.122.2@tcp(0x20000c0a87a02) 0: 1:192.168.122.2@tcp
777 * add_conn 0:lustre-MDTyyyy-mdc 1:192.168.122.2@tcp
778 * marker 7 (flags=0x02, v2.1.54.0) lustre-MDTyyyy 'add failnid' xxxx-
780 static int client_lwp_config_process(const struct lu_env *env,
781 struct llog_handle *handle,
782 struct llog_rec_hdr *rec, void *data)
784 struct config_llog_instance *cfg = data;
785 int cfg_len = rec->lrh_len;
786 char *cfg_buf = (char *) (rec + 1);
787 struct lustre_cfg *lcfg = NULL;
788 struct lustre_sb_info *lsi;
789 int rc = 0, swab = 0;
792 if (rec->lrh_type != OBD_CFG_REC) {
793 CERROR("Unknown llog record type %#x encountered\n",
798 if (cfg->cfg_sb == NULL)
799 GOTO(out, rc = -EINVAL);
800 lsi = s2lsi(cfg->cfg_sb);
802 lcfg = (struct lustre_cfg *)cfg_buf;
803 if (lcfg->lcfg_version == __swab32(LUSTRE_CFG_VERSION)) {
804 lustre_swab_lustre_cfg(lcfg);
808 rc = lustre_cfg_sanity_check(cfg_buf, cfg_len);
812 switch (lcfg->lcfg_command) {
814 struct cfg_marker *marker = lustre_cfg_buf(lcfg, 1);
816 lustre_swab_cfg_marker(marker, swab,
817 LUSTRE_CFG_BUFLEN(lcfg, 1));
818 if (marker->cm_flags & CM_SKIP ||
819 marker->cm_flags & CM_EXCLUDE)
822 if (!tgt_is_mdt(marker->cm_tgtname, &cfg->cfg_lwp_idx))
825 if (IS_MDT(lsi) && cfg->cfg_lwp_idx != 0)
828 if (!strncmp(marker->cm_comment, "add mdc", 7) ||
829 !strncmp(marker->cm_comment, "add failnid", 11)) {
830 if (marker->cm_flags & CM_START) {
831 cfg->cfg_flags = CFG_F_MARKER;
832 /* This hack is to differentiate the
833 * ADD_UUID is come from "add mdc" record
834 * or from "add failnid" record. */
835 if (!strncmp(marker->cm_comment,
837 cfg->cfg_flags |= CFG_F_SKIP;
838 } else if (marker->cm_flags & CM_END) {
844 case LCFG_ADD_UUID: {
845 if (cfg->cfg_flags == CFG_F_MARKER) {
846 rc = lustre_lwp_setup(lcfg, lsi, cfg->cfg_lwp_idx);
847 /* XXX: process only the first nid as
848 * we don't need another instance of lwp */
849 cfg->cfg_flags |= CFG_F_SKIP;
850 } else if (cfg->cfg_flags == (CFG_F_MARKER | CFG_F_SKIP)) {
851 rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
854 CERROR("%s: Fail to add uuid, rc:%d\n",
855 lsi->lsi_svname, rc);
859 case LCFG_ADD_CONN: {
860 char *devname = lustre_cfg_string(lcfg, 0);
864 if (!is_mdc_device(devname))
867 ptr = strrchr(devname, '-');
872 if (!tgt_is_mdt(devname, &idx)) {
878 if (IS_MDT(lsi) && idx != 0)
881 rc = lustre_lwp_add_conn(lcfg, lsi, idx);
891 static int lustre_disconnect_lwp(struct super_block *sb)
893 struct lustre_sb_info *lsi = s2lsi(sb);
894 struct obd_device *lwp;
895 char *logname = NULL;
896 struct lustre_cfg_bufs *bufs = NULL;
897 struct config_llog_instance *cfg = NULL;
902 if (likely(lsi->lsi_lwp_started)) {
903 OBD_ALLOC(logname, MTI_NAME_MAXLEN);
907 rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
909 CERROR("%s: failed to get fsname from svname: "
910 "rc = %d\n", lsi->lsi_svname, rc);
911 GOTO(out, rc = -EINVAL);
914 strcat(logname, "-client");
917 GOTO(out, rc = -ENOMEM);
920 cfg->cfg_instance = sb;
921 rc = lustre_end_log(sb, logname, cfg);
922 if (rc != 0 && rc != -ENOENT)
925 lsi->lsi_lwp_started = 0;
930 GOTO(out, rc = -ENOMEM);
932 list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
933 struct lustre_cfg *lcfg;
935 if (likely(lwp->obd_lwp_export != NULL)) {
936 class_export_put(lwp->obd_lwp_export);
937 lwp->obd_lwp_export = NULL;
940 lustre_cfg_bufs_reset(bufs, lwp->obd_name);
941 lustre_cfg_bufs_set_string(bufs, 1, NULL);
942 lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
944 GOTO(out, rc = -ENOMEM);
946 /* Disconnect import first. NULL is passed for the '@env',
947 * since it will not be used. */
948 rc = lwp->obd_lu_dev->ld_ops->ldo_process_config(NULL,
949 lwp->obd_lu_dev, lcfg);
950 lustre_cfg_free(lcfg);
951 if (rc != 0 && rc != -ETIMEDOUT) {
952 CERROR("%s: fail to disconnect LWP: rc = %d\n",
966 OBD_FREE(logname, MTI_NAME_MAXLEN);
968 return rc1 != 0 ? rc1 : rc;
972 * Stop the lwp for an OST/MDT target.
974 static int lustre_stop_lwp(struct super_block *sb)
976 struct lustre_sb_info *lsi = s2lsi(sb);
977 struct obd_device *lwp;
982 while (!list_empty(&lsi->lsi_lwp_list)) {
983 lwp = list_entry(lsi->lsi_lwp_list.next, struct obd_device,
985 list_del_init(&lwp->obd_lwp_list);
987 rc = class_manual_cleanup(lwp);
989 CERROR("%s: fail to stop LWP: rc = %d\n",
995 RETURN(rc1 != 0 ? rc1 : rc);
999 * Start the lwp(fsname-MDTyyyy-lwp-{MDT,OST}xxxx) for a MDT/OST or MDT target.
1001 static int lustre_start_lwp(struct super_block *sb)
1003 struct lustre_sb_info *lsi = s2lsi(sb);
1004 struct config_llog_instance *cfg = NULL;
1009 if (unlikely(lsi->lsi_lwp_started))
1012 OBD_ALLOC(logname, MTI_NAME_MAXLEN);
1013 if (logname == NULL)
1016 rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
1018 CERROR("%s: failed to get fsname from svname: rc = %d\n",
1019 lsi->lsi_svname, rc);
1020 GOTO(out, rc = -EINVAL);
1023 strcat(logname, "-client");
1026 GOTO(out, rc = -ENOMEM);
1028 cfg->cfg_callback = client_lwp_config_process;
1029 cfg->cfg_instance = sb;
1030 rc = lustre_process_log(sb, logname, cfg);
1031 /* need to remove config llog from mgc */
1032 lsi->lsi_lwp_started = 1;
1037 OBD_FREE(logname, MTI_NAME_MAXLEN);
1044 static DEFINE_MUTEX(server_start_lock);
1046 /* Stop MDS/OSS if nobody is using them */
1047 static int server_stop_servers(int lsiflags)
1049 struct obd_device *obd = NULL;
1050 struct obd_type *type = NULL;
1054 mutex_lock(&server_start_lock);
1056 /* Either an MDT or an OST or neither */
1057 /* if this was an MDT, and there are no more MDT's, clean up the MDS */
1058 if (lsiflags & LDD_F_SV_TYPE_MDT) {
1059 obd = class_name2obd(LUSTRE_MDS_OBDNAME);
1061 type = class_search_type(LUSTRE_MDT_NAME);
1064 /* if this was an OST, and there are no more OST's, clean up the OSS */
1065 if (lsiflags & LDD_F_SV_TYPE_OST) {
1066 obd = class_name2obd(LUSTRE_OSS_OBDNAME);
1068 type = class_search_type(LUSTRE_OST_NAME);
1071 if (obd != NULL && (type == NULL || type->typ_refcnt == 0)) {
1073 /* obd_fail doesn't mean much on a server obd */
1074 rc = class_manual_cleanup(obd);
1077 mutex_unlock(&server_start_lock);
1082 int server_mti_print(const char *title, struct mgs_target_info *mti)
1084 PRINT_CMD(PRINT_MASK, "mti %s\n", title);
1085 PRINT_CMD(PRINT_MASK, "server: %s\n", mti->mti_svname);
1086 PRINT_CMD(PRINT_MASK, "fs: %s\n", mti->mti_fsname);
1087 PRINT_CMD(PRINT_MASK, "uuid: %s\n", mti->mti_uuid);
1088 PRINT_CMD(PRINT_MASK, "ver: %d flags: %#x\n",
1089 mti->mti_config_ver, mti->mti_flags);
1093 /* Generate data for registration */
1094 static int server_lsi2mti(struct lustre_sb_info *lsi,
1095 struct mgs_target_info *mti)
1097 lnet_process_id_t id;
1102 if (!IS_SERVER(lsi))
1105 if (strlcpy(mti->mti_svname, lsi->lsi_svname, sizeof(mti->mti_svname))
1106 >= sizeof(mti->mti_svname))
1109 mti->mti_nid_count = 0;
1110 while (LNetGetId(i++, &id) != -ENOENT) {
1111 if (LNET_NETTYP(LNET_NIDNET(id.nid)) == LOLND)
1114 /* server use --servicenode param, only allow specified
1115 * nids be registered */
1116 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_NO_PRIMNODE) != 0 &&
1117 class_match_nid(lsi->lsi_lmd->lmd_params,
1118 PARAM_FAILNODE, id.nid) < 1)
1121 /* match specified network */
1122 if (!class_match_net(lsi->lsi_lmd->lmd_params,
1123 PARAM_NETWORK, LNET_NIDNET(id.nid)))
1126 mti->mti_nids[mti->mti_nid_count] = id.nid;
1127 mti->mti_nid_count++;
1128 if (mti->mti_nid_count >= MTI_NIDS_MAX) {
1129 CWARN("Only using first %d nids for %s\n",
1130 mti->mti_nid_count, mti->mti_svname);
1135 if (mti->mti_nid_count == 0) {
1136 CERROR("Failed to get NID for server %s, please check whether "
1137 "the target is specifed with improper --servicenode or "
1138 "--network options.\n", mti->mti_svname);
1142 mti->mti_lustre_ver = LUSTRE_VERSION_CODE;
1143 mti->mti_config_ver = 0;
1145 rc = server_name2fsname(lsi->lsi_svname, mti->mti_fsname, NULL);
1149 rc = server_name2index(lsi->lsi_svname, &mti->mti_stripe_index, NULL);
1152 /* Orion requires index to be set */
1153 LASSERT(!(rc & LDD_F_NEED_INDEX));
1154 /* keep only LDD flags */
1155 mti->mti_flags = lsi->lsi_flags & LDD_F_MASK;
1156 if (mti->mti_flags & (LDD_F_WRITECONF | LDD_F_VIRGIN))
1157 mti->mti_flags |= LDD_F_UPDATE;
1158 cplen = strlcpy(mti->mti_params, lsi->lsi_lmd->lmd_params,
1159 sizeof(mti->mti_params));
1160 if (cplen >= sizeof(mti->mti_params))
1165 /* Register an old or new target with the MGS. If needed MGS will construct
1166 startup logs and assign index */
1167 static int server_register_target(struct lustre_sb_info *lsi)
1169 struct obd_device *mgc = lsi->lsi_mgc;
1170 struct mgs_target_info *mti = NULL;
1177 if (!IS_SERVER(lsi))
1184 rc = server_lsi2mti(lsi, mti);
1188 CDEBUG(D_MOUNT, "Registration %s, fs=%s, %s, index=%04x, flags=%#x\n",
1189 mti->mti_svname, mti->mti_fsname,
1190 libcfs_nid2str(mti->mti_nids[0]), mti->mti_stripe_index,
1193 /* if write_conf is true, the registration must succeed */
1194 writeconf = !!(lsi->lsi_flags & (LDD_F_NEED_INDEX | LDD_F_UPDATE));
1195 mti->mti_flags |= LDD_F_OPC_REG;
1197 /* Register the target */
1198 /* FIXME use mgc_process_config instead */
1199 rc = obd_set_info_async(NULL, mgc->u.cli.cl_mgc_mgsexp,
1200 sizeof(KEY_REGISTER_TARGET),
1201 KEY_REGISTER_TARGET,
1202 sizeof(*mti), mti, NULL);
1204 if (mti->mti_flags & LDD_F_ERROR) {
1205 LCONSOLE_ERROR_MSG(0x160,
1206 "%s: the MGS refuses to allow this server "
1207 "to start: rc = %d. Please see messages on "
1208 "the MGS.\n", lsi->lsi_svname, rc);
1209 } else if (writeconf) {
1210 LCONSOLE_ERROR_MSG(0x15f,
1211 "%s: cannot register this server with the MGS: "
1212 "rc = %d. Is the MGS running?\n",
1213 lsi->lsi_svname, rc);
1215 CDEBUG(D_HA, "%s: error registering with the MGS: "
1216 "rc = %d (not fatal)\n", lsi->lsi_svname, rc);
1217 /* reset the error code for non-fatal error. */
1230 * Notify the MGS that this target is ready.
1231 * Used by IR - if the MGS receives this message, it will notify clients.
1233 static int server_notify_target(struct super_block *sb, struct obd_device *obd)
1235 struct lustre_sb_info *lsi = s2lsi(sb);
1236 struct obd_device *mgc = lsi->lsi_mgc;
1237 struct mgs_target_info *mti = NULL;
1243 if (!(IS_SERVER(lsi)))
1249 rc = server_lsi2mti(lsi, mti);
1253 mti->mti_instance = obd->u.obt.obt_instance;
1254 mti->mti_flags |= LDD_F_OPC_READY;
1256 /* FIXME use mgc_process_config instead */
1257 rc = obd_set_info_async(NULL, mgc->u.cli.cl_mgc_mgsexp,
1258 sizeof(KEY_REGISTER_TARGET),
1259 KEY_REGISTER_TARGET,
1260 sizeof(*mti), mti, NULL);
1262 /* Imperative recovery: if the mgs informs us to use IR? */
1263 if (!rc && !(mti->mti_flags & LDD_F_ERROR) &&
1264 (mti->mti_flags & LDD_F_IR_CAPABLE))
1265 lsi->lsi_flags |= LDD_F_IR_CAPABLE;
1274 /** Start server targets: MDTs and OSTs
1276 static int server_start_targets(struct super_block *sb)
1278 struct obd_device *obd;
1279 struct lustre_sb_info *lsi = s2lsi(sb);
1280 struct config_llog_instance cfg;
1281 struct lu_env mgc_env;
1282 struct lu_device *dev;
1286 CDEBUG(D_MOUNT, "starting target %s\n", lsi->lsi_svname);
1289 /* make sure the MDS is started */
1290 mutex_lock(&server_start_lock);
1291 obd = class_name2obd(LUSTRE_MDS_OBDNAME);
1293 rc = lustre_start_simple(LUSTRE_MDS_OBDNAME,
1295 LUSTRE_MDS_OBDNAME"_uuid",
1296 NULL, NULL, NULL, NULL);
1298 mutex_unlock(&server_start_lock);
1299 CERROR("failed to start MDS: %d\n", rc);
1303 mutex_unlock(&server_start_lock);
1306 /* If we're an OST, make sure the global OSS is running */
1308 /* make sure OSS is started */
1309 mutex_lock(&server_start_lock);
1310 obd = class_name2obd(LUSTRE_OSS_OBDNAME);
1312 rc = lustre_start_simple(LUSTRE_OSS_OBDNAME,
1314 LUSTRE_OSS_OBDNAME"_uuid",
1315 NULL, NULL, NULL, NULL);
1317 mutex_unlock(&server_start_lock);
1318 CERROR("failed to start OSS: %d\n", rc);
1322 mutex_unlock(&server_start_lock);
1325 rc = lu_env_init(&mgc_env, LCT_MG_THREAD);
1327 GOTO(out_stop_service, rc);
1329 /* Set the mgc fs to our server disk. This allows the MGC to
1330 * read and write configs locally, in case it can't talk to the MGS. */
1331 rc = server_mgc_set_fs(&mgc_env, lsi->lsi_mgc, sb);
1335 /* Register with MGS */
1336 rc = server_register_target(lsi);
1340 /* Let the target look up the mount using the target's name
1341 (we can't pass the sb or mnt through class_process_config.) */
1342 rc = server_register_mount(lsi->lsi_svname, sb);
1346 /* Start targets using the llog named for the target */
1347 memset(&cfg, 0, sizeof(cfg));
1348 cfg.cfg_callback = class_config_llog_handler;
1349 rc = lustre_process_log(sb, lsi->lsi_svname, &cfg);
1351 CERROR("failed to start server %s: %d\n",
1352 lsi->lsi_svname, rc);
1353 /* Do NOT call server_deregister_mount() here. This makes it
1354 * impossible to find mount later in cleanup time and leaves
1355 * @lsi and othder stuff leaked. -umka */
1359 obd = class_name2obd(lsi->lsi_svname);
1361 CERROR("no server named %s was started\n", lsi->lsi_svname);
1362 GOTO(out_mgc, rc = -ENXIO);
1365 if (IS_OST(lsi) || IS_MDT(lsi)) {
1366 rc = lustre_start_lwp(sb);
1368 CERROR("%s: failed to start LWP: %d\n",
1369 lsi->lsi_svname, rc);
1374 server_notify_target(sb, obd);
1376 /* calculate recovery timeout, do it after lustre_process_log */
1377 server_calc_timeout(lsi, obd);
1379 /* log has been fully processed */
1380 obd_notify(obd, NULL, OBD_NOTIFY_CONFIG, (void *)CONFIG_LOG);
1382 /* log has been fully processed, let clients connect */
1383 dev = obd->obd_lu_dev;
1384 if (dev && dev->ld_ops->ldo_prepare) {
1387 rc = lu_env_init(&env, dev->ld_type->ldt_ctx_tags);
1389 struct lu_context session_ctx;
1391 lu_context_init(&session_ctx, LCT_SERVER_SESSION);
1392 session_ctx.lc_thread = NULL;
1393 lu_context_enter(&session_ctx);
1394 env.le_ses = &session_ctx;
1396 rc = dev->ld_ops->ldo_prepare(&env, NULL, dev);
1399 lu_context_exit(&session_ctx);
1400 lu_context_fini(&session_ctx);
1404 /* abort recovery only on the complete stack:
1405 * many devices can be involved */
1406 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_ABORT_RECOV) &&
1407 (OBP(obd, iocontrol))) {
1408 obd_iocontrol(OBD_IOC_ABORT_RECOVERY, obd->obd_self_export, 0,
1413 /* Release the mgc fs for others to use */
1414 server_mgc_clear_fs(&mgc_env, lsi->lsi_mgc);
1416 lu_env_fini(&mgc_env);
1419 server_stop_servers(lsi->lsi_flags);
1424 static int lsi_prepare(struct lustre_sb_info *lsi)
1426 const char *osd_type;
1433 LASSERT(lsi->lsi_lmd);
1435 /* The server name is given as a mount line option */
1436 if (lsi->lsi_lmd->lmd_profile == NULL) {
1437 LCONSOLE_ERROR("Can't determine server name\n");
1441 /* Determine osd type */
1442 if (lsi->lsi_lmd->lmd_osd_type == NULL) {
1443 osd_type = LUSTRE_OSD_LDISKFS_NAME;
1446 osd_type = lsi->lsi_lmd->lmd_osd_type;
1447 fstype = lsi->lsi_lmd->lmd_osd_type;
1450 if (strlen(lsi->lsi_lmd->lmd_profile) >= sizeof(lsi->lsi_svname) ||
1451 strlen(osd_type) >= sizeof(lsi->lsi_osd_type) ||
1452 strlen(fstype) >= sizeof(lsi->lsi_fstype))
1453 RETURN(-ENAMETOOLONG);
1455 strlcpy(lsi->lsi_svname, lsi->lsi_lmd->lmd_profile,
1456 sizeof(lsi->lsi_svname));
1457 strlcpy(lsi->lsi_osd_type, osd_type, sizeof(lsi->lsi_osd_type));
1458 /* XXX: a temp. solution for components using ldiskfs
1459 * to be removed in one of the subsequent patches */
1460 strlcpy(lsi->lsi_fstype, fstype, sizeof(lsi->lsi_fstype));
1462 /* Determine server type */
1463 rc = server_name2index(lsi->lsi_svname, &index, NULL);
1465 if (lsi->lsi_lmd->lmd_flags & LMD_FLG_MGS) {
1466 /* Assume we're a bare MGS */
1468 lsi->lsi_lmd->lmd_flags |= LMD_FLG_NOSVC;
1470 LCONSOLE_ERROR("Can't determine server type of '%s'\n",
1475 lsi->lsi_flags |= rc;
1477 /* Add mount line flags that used to be in ldd:
1478 * writeconf, mgs, anything else?
1480 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_WRITECONF) ?
1481 LDD_F_WRITECONF : 0;
1482 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_VIRGIN) ?
1484 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_UPDATE) ?
1486 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_MGS) ?
1487 LDD_F_SV_TYPE_MGS : 0;
1488 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_NO_PRIMNODE) ?
1489 LDD_F_NO_PRIMNODE : 0;
1494 /*************** server mount ******************/
1496 /** Start the shutdown of servers at umount.
1498 static void server_put_super(struct super_block *sb)
1500 struct lustre_sb_info *lsi = s2lsi(sb);
1501 struct obd_device *obd;
1502 char *tmpname, *extraname = NULL;
1504 int lsiflags = lsi->lsi_flags;
1507 LASSERT(IS_SERVER(lsi));
1509 tmpname_sz = strlen(lsi->lsi_svname) + 1;
1510 OBD_ALLOC(tmpname, tmpname_sz);
1511 memcpy(tmpname, lsi->lsi_svname, tmpname_sz);
1512 CDEBUG(D_MOUNT, "server put_super %s\n", tmpname);
1513 if (IS_MDT(lsi) && (lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC))
1514 snprintf(tmpname, tmpname_sz, "MGS");
1516 /* disconnect the lwp first to drain off the inflight request */
1517 if (IS_OST(lsi) || IS_MDT(lsi)) {
1520 rc = lustre_disconnect_lwp(sb);
1521 if (rc != 0 && rc != -ETIMEDOUT &&
1522 rc != -ENOTCONN && rc != -ESHUTDOWN)
1523 CWARN("%s: failed to disconnect lwp: rc= %d\n",
1527 /* Stop the target */
1528 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1529 (IS_MDT(lsi) || IS_OST(lsi))) {
1530 struct lustre_profile *lprof = NULL;
1532 /* tell the mgc to drop the config log */
1533 lustre_end_log(sb, lsi->lsi_svname, NULL);
1535 /* COMPAT_146 - profile may get deleted in mgc_cleanup.
1536 If there are any setup/cleanup errors, save the lov
1537 name for safety cleanup later. */
1538 lprof = class_get_profile(lsi->lsi_svname);
1539 if (lprof != NULL) {
1540 if (lprof->lp_dt != NULL) {
1541 OBD_ALLOC(extraname, strlen(lprof->lp_dt) + 1);
1542 strncpy(extraname, lprof->lp_dt,
1543 strlen(lprof->lp_dt) + 1);
1545 class_put_profile(lprof);
1548 obd = class_name2obd(lsi->lsi_svname);
1550 CDEBUG(D_MOUNT, "stopping %s\n", obd->obd_name);
1551 if (lsiflags & LSI_UMOUNT_FAILOVER)
1553 /* We can't seem to give an error return code
1554 * to .put_super, so we better make sure we clean up! */
1556 class_manual_cleanup(obd);
1558 CERROR("no obd %s\n", lsi->lsi_svname);
1559 server_deregister_mount(lsi->lsi_svname);
1563 /* If they wanted the mgs to stop separately from the mdt, they
1564 should have put it on a different device. */
1566 /* if MDS start with --nomgs, don't stop MGS then */
1567 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS))
1568 server_stop_mgs(sb);
1571 if (IS_OST(lsi) || IS_MDT(lsi)) {
1572 if (lustre_stop_lwp(sb) < 0)
1573 CERROR("%s: failed to stop lwp!\n", tmpname);
1576 /* Clean the mgc and sb */
1577 lustre_common_put_super(sb);
1579 /* wait till all in-progress cleanups are done
1580 * specifically we're interested in ofd cleanup
1582 obd_zombie_barrier();
1584 /* Stop the servers (MDS, OSS) if no longer needed. We must wait
1585 until the target is really gone so that our type refcount check
1587 server_stop_servers(lsiflags);
1589 /* In case of startup or cleanup err, stop related obds */
1591 obd = class_name2obd(extraname);
1593 CWARN("Cleaning orphaned obd %s\n", extraname);
1595 class_manual_cleanup(obd);
1597 OBD_FREE(extraname, strlen(extraname) + 1);
1600 LCONSOLE_WARN("server umount %s complete\n", tmpname);
1601 OBD_FREE(tmpname, tmpname_sz);
1605 /** Called only for 'umount -f'
1607 static void server_umount_begin(struct super_block *sb)
1609 struct lustre_sb_info *lsi = s2lsi(sb);
1612 CDEBUG(D_MOUNT, "umount -f\n");
1613 /* umount = failover
1615 no third way to do non-force, non-failover */
1616 lsi->lsi_flags &= ~LSI_UMOUNT_FAILOVER;
1620 static int server_statfs(struct dentry *dentry, struct kstatfs *buf)
1622 struct super_block *sb = dentry->d_sb;
1623 struct lustre_sb_info *lsi = s2lsi(sb);
1624 struct obd_statfs statfs;
1628 if (lsi->lsi_dt_dev) {
1629 rc = dt_statfs(NULL, lsi->lsi_dt_dev, &statfs);
1631 statfs_unpack(buf, &statfs);
1632 buf->f_type = sb->s_magic;
1638 buf->f_type = sb->s_magic;
1639 buf->f_bsize = sb->s_blocksize;
1645 buf->f_namelen = NAME_MAX;
1649 /** The operations we support directly on the superblock:
1650 * mount, umount, and df.
1652 static struct super_operations server_ops = {
1653 .put_super = server_put_super,
1654 .umount_begin = server_umount_begin, /* umount -f */
1655 .statfs = server_statfs,
1659 * Xattr support for Lustre servers
1661 static ssize_t lustre_getxattr(struct dentry *dentry, const char *name,
1662 void *buffer, size_t size)
1664 if (!selinux_is_enabled())
1669 static int lustre_setxattr(struct dentry *dentry, const char *name,
1670 const void *value, size_t size, int flags)
1675 static ssize_t lustre_listxattr(struct dentry *d_entry, char *name,
1681 static const struct inode_operations server_inode_operations = {
1682 .setxattr = lustre_setxattr,
1683 .getxattr = lustre_getxattr,
1684 .listxattr = lustre_listxattr,
1687 #define log2(n) ffz(~(n))
1688 #define LUSTRE_SUPER_MAGIC 0x0BD00BD1
1690 static int server_fill_super_common(struct super_block *sb)
1692 struct inode *root = NULL;
1695 CDEBUG(D_MOUNT, "Server sb, dev=%d\n", (int)sb->s_dev);
1697 sb->s_blocksize = 4096;
1698 sb->s_blocksize_bits = log2(sb->s_blocksize);
1699 sb->s_magic = LUSTRE_SUPER_MAGIC;
1700 sb->s_maxbytes = 0; /* we don't allow file IO on server mountpoints */
1701 sb->s_flags |= MS_RDONLY;
1702 sb->s_op = &server_ops;
1704 root = new_inode(sb);
1706 CERROR("Can't make root inode\n");
1710 /* returns -EIO for every operation */
1711 /* make_bad_inode(root); -- badness - can't umount */
1712 /* apparently we need to be a directory for the mount to finish */
1713 root->i_mode = S_IFDIR;
1714 root->i_op = &server_inode_operations;
1715 sb->s_root = d_make_root(root);
1717 CERROR("%s: can't make root dentry\n", sb->s_id);
1724 static int osd_start(struct lustre_sb_info *lsi, unsigned long mflags)
1726 struct lustre_mount_data *lmd = lsi->lsi_lmd;
1727 struct obd_device *obd;
1728 struct dt_device_param p;
1734 "Attempting to start %s, type=%s, lsifl=%x, mountfl=%lx\n",
1735 lsi->lsi_svname, lsi->lsi_osd_type, lsi->lsi_flags, mflags);
1737 sprintf(lsi->lsi_osd_obdname, "%s-osd", lsi->lsi_svname);
1738 strcpy(lsi->lsi_osd_uuid, lsi->lsi_osd_obdname);
1739 strcat(lsi->lsi_osd_uuid, "_UUID");
1740 sprintf(flagstr, "%lu:%lu", mflags, (unsigned long) lmd->lmd_flags);
1742 obd = class_name2obd(lsi->lsi_osd_obdname);
1744 rc = lustre_start_simple(lsi->lsi_osd_obdname,
1746 lsi->lsi_osd_uuid, lmd->lmd_dev,
1747 flagstr, lsi->lsi_lmd->lmd_opts,
1751 obd = class_name2obd(lsi->lsi_osd_obdname);
1754 CDEBUG(D_MOUNT, "%s already started\n", lsi->lsi_osd_obdname);
1755 /* but continue setup to allow special case of MDT and internal
1756 * MGT being started separately. */
1757 if (!((IS_MGS(lsi) && (lsi->lsi_lmd->lmd_flags &
1759 (IS_MDT(lsi) && (lsi->lsi_lmd->lmd_flags &
1764 rc = obd_connect(NULL, &lsi->lsi_osd_exp,
1765 obd, &obd->obd_uuid, NULL, NULL);
1769 class_manual_cleanup(obd);
1770 lsi->lsi_dt_dev = NULL;
1774 LASSERT(obd->obd_lu_dev);
1775 lu_device_get(obd->obd_lu_dev);
1776 lsi->lsi_dt_dev = lu2dt_dev(obd->obd_lu_dev);
1777 LASSERT(lsi->lsi_dt_dev);
1779 /* set disk context for llog usage */
1780 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1781 obd->obd_lvfs_ctxt.dt = lsi->lsi_dt_dev;
1783 dt_conf_get(NULL, lsi->lsi_dt_dev, &p);
1788 /** Fill in the superblock info for a Lustre server.
1789 * Mount the device with the correct options.
1790 * Read the on-disk config file.
1791 * Start the services.
1793 int server_fill_super(struct super_block *sb)
1795 struct lustre_sb_info *lsi = s2lsi(sb);
1799 /* to simulate target mount race */
1800 OBD_RACE(OBD_FAIL_TGT_MOUNT_RACE);
1802 rc = lsi_prepare(lsi);
1806 /* Start low level OSD */
1807 rc = osd_start(lsi, sb->s_flags);
1809 CERROR("Unable to start osd on %s: %d\n",
1810 lsi->lsi_lmd->lmd_dev, rc);
1815 CDEBUG(D_MOUNT, "Found service %s on device %s\n",
1816 lsi->lsi_svname, lsi->lsi_lmd->lmd_dev);
1818 if (class_name2obd(lsi->lsi_svname)) {
1819 LCONSOLE_ERROR_MSG(0x161, "The target named %s is already "
1820 "running. Double-mount may have compromised"
1821 " the disk journal.\n",
1827 /* Start MGS before MGC */
1828 if (IS_MGS(lsi) && !(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS)) {
1829 rc = server_start_mgs(sb);
1834 /* Start MGC before servers */
1835 rc = lustre_start_mgc(sb);
1839 /* Set up all obd devices for service */
1840 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1841 (IS_OST(lsi) || IS_MDT(lsi))) {
1842 rc = server_start_targets(sb);
1844 CERROR("Unable to start targets: %d\n", rc);
1847 /* FIXME overmount client here, or can we just start a
1848 * client log and client_fill_super on this sb? We
1849 * need to make sure server_put_super gets called too
1850 * - ll_put_super calls lustre_common_put_super; check
1851 * there for LSI_SERVER flag, call s_p_s if so.
1853 * Probably should start client from new thread so we
1854 * can return. Client will not finish until all
1855 * servers are connected. Note - MGS-only server does
1856 * NOT get a client, since there is no lustre fs
1857 * associated - the MGS is for all lustre fs's */
1860 rc = server_fill_super_common(sb);
1866 /* We jump here in case of failure while starting targets or MGS.
1867 * In this case we can't just put @mnt and have to do real cleanup
1868 * with stoping targets, etc. */
1869 server_put_super(sb);
1874 * Calculate timeout value for a target.
1876 void server_calc_timeout(struct lustre_sb_info *lsi, struct obd_device *obd)
1878 struct lustre_mount_data *lmd;
1882 bool has_ir = !!(lsi->lsi_flags & LDD_F_IR_CAPABLE);
1883 int min = OBD_RECOVERY_TIME_MIN;
1885 LASSERT(IS_SERVER(lsi));
1889 soft = lmd->lmd_recovery_time_soft;
1890 hard = lmd->lmd_recovery_time_hard;
1891 has_ir = has_ir && !(lmd->lmd_flags & LMD_FLG_NOIR);
1892 obd->obd_no_ir = !has_ir;
1896 soft = OBD_RECOVERY_TIME_SOFT;
1898 hard = OBD_RECOVERY_TIME_HARD;
1900 /* target may have ir_factor configured. */
1901 factor = OBD_IR_FACTOR_DEFAULT;
1902 if (obd->obd_recovery_ir_factor)
1903 factor = obd->obd_recovery_ir_factor;
1906 int new_soft = soft;
1907 int new_hard = hard;
1909 /* adjust timeout value by imperative recovery */
1911 new_soft = (soft * factor) / OBD_IR_FACTOR_MAX;
1912 new_hard = (hard * factor) / OBD_IR_FACTOR_MAX;
1914 /* make sure the timeout is not too short */
1915 new_soft = max(min, new_soft);
1916 new_hard = max(new_soft, new_hard);
1918 LCONSOLE_INFO("%s: Imperative Recovery enabled, recovery "
1919 "window shrunk from %d-%d down to %d-%d\n",
1920 obd->obd_name, soft, hard, new_soft, new_hard);
1927 obd->obd_recovery_timeout = max(obd->obd_recovery_timeout, soft);
1928 obd->obd_recovery_time_hard = hard;
1929 obd->obd_recovery_ir_factor = factor;