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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2013, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/obdclass/obd_mount_server.c
34 * Server mount routines
36 * Author: Nathan Rutman <nathan@clusterfs.com>
40 #define DEBUG_SUBSYSTEM S_CLASS
41 #define D_MOUNT (D_SUPER | D_CONFIG /* | D_WARNING */)
42 #define PRINT_CMD CDEBUG
43 #define PRINT_MASK (D_SUPER | D_CONFIG)
45 #include <linux/types.h>
46 #ifdef HAVE_LINUX_SELINUX_IS_ENABLED
47 #include <linux/selinux.h>
49 #include <linux/statfs.h>
50 #include <linux/version.h>
52 #include <llog_swab.h>
53 #include <lustre_disk.h>
54 #include <uapi/linux/lustre/lustre_ioctl.h>
55 #include <lustre_log.h>
56 #include <uapi/linux/lustre/lustre_param.h>
58 #include <obd_class.h>
60 /*********** mount lookup *********/
62 static DEFINE_MUTEX(lustre_mount_info_lock);
63 static LIST_HEAD(server_mount_info_list);
65 static struct lustre_mount_info *server_find_mount(const char *name)
67 struct list_head *tmp;
68 struct lustre_mount_info *lmi;
71 list_for_each(tmp, &server_mount_info_list) {
72 lmi = list_entry(tmp, struct lustre_mount_info,
74 if (strcmp(name, lmi->lmi_name) == 0)
80 /* we must register an obd for a mount before we call the setup routine.
81 *_setup will call lustre_get_mount to get the mnt struct
82 by obd_name, since we can't pass the pointer to setup. */
83 static int server_register_mount(const char *name, struct super_block *sb)
85 struct lustre_mount_info *lmi;
91 OBD_ALLOC(lmi, sizeof(*lmi));
94 OBD_ALLOC(name_cp, strlen(name) + 1);
96 OBD_FREE(lmi, sizeof(*lmi));
99 strcpy(name_cp, name);
101 mutex_lock(&lustre_mount_info_lock);
103 if (server_find_mount(name)) {
104 mutex_unlock(&lustre_mount_info_lock);
105 OBD_FREE(lmi, sizeof(*lmi));
106 OBD_FREE(name_cp, strlen(name) + 1);
107 CERROR("Already registered %s\n", name);
110 lmi->lmi_name = name_cp;
112 list_add(&lmi->lmi_list_chain, &server_mount_info_list);
114 mutex_unlock(&lustre_mount_info_lock);
116 CDEBUG(D_MOUNT, "register mount %p from %s\n", sb, name);
121 /* when an obd no longer needs a mount */
122 static int server_deregister_mount(const char *name)
124 struct lustre_mount_info *lmi;
127 mutex_lock(&lustre_mount_info_lock);
128 lmi = server_find_mount(name);
130 mutex_unlock(&lustre_mount_info_lock);
131 CERROR("%s not registered\n", name);
135 CDEBUG(D_MOUNT, "deregister mount %p from %s\n", lmi->lmi_sb, name);
137 OBD_FREE(lmi->lmi_name, strlen(lmi->lmi_name) + 1);
138 list_del(&lmi->lmi_list_chain);
139 OBD_FREE(lmi, sizeof(*lmi));
140 mutex_unlock(&lustre_mount_info_lock);
145 /* obd's look up a registered mount using their obdname. This is just
146 for initial obd setup to find the mount struct. It should not be
147 called every time you want to mntget. */
148 struct lustre_mount_info *server_get_mount(const char *name)
150 struct lustre_mount_info *lmi;
151 struct lustre_sb_info *lsi;
154 mutex_lock(&lustre_mount_info_lock);
155 lmi = server_find_mount(name);
156 mutex_unlock(&lustre_mount_info_lock);
158 CERROR("Can't find mount for %s\n", name);
161 lsi = s2lsi(lmi->lmi_sb);
163 atomic_inc(&lsi->lsi_mounts);
165 CDEBUG(D_MOUNT, "get mount %p from %s, refs=%d\n", lmi->lmi_sb,
166 name, atomic_read(&lsi->lsi_mounts));
170 EXPORT_SYMBOL(server_get_mount);
173 * server_put_mount: to be called from obd_cleanup methods
175 * @dereg_mnt: 0 or 1 depending on whether the mount is to be deregistered or
178 * The caller decides whether server_deregister_mount() needs to be called or
179 * not. Calling of server_deregister_mount() does not depend on refcounting on
180 * lsi because we could have say the mgs and mds on the same node and we
181 * unmount the mds, then the ref on the lsi would still be non-zero but we
182 * would still want to deregister the mds mount.
184 int server_put_mount(const char *name, bool dereg_mnt)
186 struct lustre_mount_info *lmi;
187 struct lustre_sb_info *lsi;
190 mutex_lock(&lustre_mount_info_lock);
191 lmi = server_find_mount(name);
192 mutex_unlock(&lustre_mount_info_lock);
194 CERROR("Can't find mount for %s\n", name);
197 lsi = s2lsi(lmi->lmi_sb);
199 CDEBUG(D_MOUNT, "put mount %p from %s, refs=%d\n",
200 lmi->lmi_sb, name, atomic_read(&lsi->lsi_mounts));
202 if (lustre_put_lsi(lmi->lmi_sb))
203 CDEBUG(D_MOUNT, "Last put of mount %p from %s\n",
207 /* this obd should never need the mount again */
208 server_deregister_mount(name);
212 EXPORT_SYMBOL(server_put_mount);
214 /* Set up a MGS to serve startup logs */
215 static int server_start_mgs(struct super_block *sb)
217 struct lustre_sb_info *lsi = s2lsi(sb);
218 struct lustre_mount_info *lmi;
222 /* It is impossible to have more than 1 MGS per node, since
223 MGC wouldn't know which to connect to */
224 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
226 lsi = s2lsi(lmi->lmi_sb);
227 LCONSOLE_ERROR_MSG(0x15d, "The MGS service was already started"
232 CDEBUG(D_CONFIG, "Start MGS service %s\n", LUSTRE_MGS_OBDNAME);
234 rc = server_register_mount(LUSTRE_MGS_OBDNAME, sb);
237 rc = lustre_start_simple(LUSTRE_MGS_OBDNAME, LUSTRE_MGS_NAME,
238 LUSTRE_MGS_OBDNAME, NULL, NULL,
239 lsi->lsi_osd_obdname, NULL);
240 /* server_deregister_mount() is not called previously, for lsi
241 * and other stuff can't be freed cleanly when mgs calls
242 * server_put_mount() in error handling case (see b=17758),
243 * this problem is caused by a bug in mgs_init0, which forgot
244 * calling server_put_mount in error case. */
247 server_deregister_mount(LUSTRE_MGS_OBDNAME);
251 LCONSOLE_ERROR_MSG(0x15e, "Failed to start MGS '%s' (%d). "
252 "Is the 'mgs' module loaded?\n",
253 LUSTRE_MGS_OBDNAME, rc);
257 static int server_stop_mgs(struct super_block *sb)
259 struct obd_device *obd;
261 struct lustre_mount_info *lmi;
264 /* Do not stop MGS if this device is not the running MGT */
265 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
266 if (lmi != NULL && lmi->lmi_sb != sb)
269 CDEBUG(D_MOUNT, "Stop MGS service %s\n", LUSTRE_MGS_OBDNAME);
271 /* There better be only one MGS */
272 obd = class_name2obd(LUSTRE_MGS_OBDNAME);
274 CDEBUG(D_CONFIG, "mgs %s not running\n", LUSTRE_MGS_OBDNAME);
278 /* The MGS should always stop when we say so */
280 rc = class_manual_cleanup(obd);
284 /* Since there's only one mgc per node, we have to change it's fs to get
285 access to the right disk. */
286 static int server_mgc_set_fs(const struct lu_env *env,
287 struct obd_device *mgc, struct super_block *sb)
289 struct lustre_sb_info *lsi = s2lsi(sb);
293 CDEBUG(D_MOUNT, "Set mgc disk for %s\n", lsi->lsi_lmd->lmd_dev);
295 /* cl_mgc_sem in mgc insures we sleep if the mgc_fs is busy */
296 rc = obd_set_info_async(env, mgc->obd_self_export,
297 sizeof(KEY_SET_FS), KEY_SET_FS,
298 sizeof(*sb), sb, NULL);
300 CERROR("can't set_fs %d\n", rc);
305 static int server_mgc_clear_fs(const struct lu_env *env,
306 struct obd_device *mgc)
311 CDEBUG(D_MOUNT, "Unassign mgc disk\n");
313 rc = obd_set_info_async(env, mgc->obd_self_export,
314 sizeof(KEY_CLEAR_FS), KEY_CLEAR_FS,
319 static inline bool is_mdc_device(const char *devname)
323 ptr = strrchr(devname, '-');
324 return ptr != NULL && strcmp(ptr, "-mdc") == 0;
327 static inline bool tgt_is_mdt(const char *tgtname, __u32 *idx)
331 type = server_name2index(tgtname, idx, NULL);
333 return type == LDD_F_SV_TYPE_MDT;
337 * Convert OST/MDT name(fsname-{MDT,OST}xxxx) to a lwp name with the @idx:yyyy
338 * (fsname-MDTyyyy-lwp-{MDT,OST}xxxx)
340 int tgt_name2lwp_name(const char *tgt_name, char *lwp_name, int len, __u32 idx)
347 OBD_ALLOC(fsname, MTI_NAME_MAXLEN);
351 rc = server_name2fsname(tgt_name, fsname, &tgt);
353 CERROR("%s: failed to get fsname from tgt_name: rc = %d\n",
358 if (*tgt != '-' && *tgt != ':') {
359 CERROR("%s: invalid tgt_name name!\n", tgt_name);
360 GOTO(cleanup, rc = -EINVAL);
364 if (strncmp(tgt, "OST", 3) != 0 && strncmp(tgt, "MDT", 3) != 0) {
365 CERROR("%s is not an OST or MDT target!\n", tgt_name);
366 GOTO(cleanup, rc = -EINVAL);
368 snprintf(lwp_name, len, "%s-MDT%04x-%s-%s",
369 fsname, idx, LUSTRE_LWP_NAME, tgt);
371 GOTO(cleanup, rc = 0);
375 OBD_FREE(fsname, MTI_NAME_MAXLEN);
379 EXPORT_SYMBOL(tgt_name2lwp_name);
381 static LIST_HEAD(lwp_register_list);
382 static DEFINE_SPINLOCK(lwp_register_list_lock);
384 static void lustre_put_lwp_item(struct lwp_register_item *lri)
386 if (atomic_dec_and_test(&lri->lri_ref)) {
387 LASSERT(list_empty(&lri->lri_list));
389 if (*lri->lri_exp != NULL)
390 class_export_put(*lri->lri_exp);
395 int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
396 register_lwp_cb cb_func, void *cb_data)
398 struct obd_device *lwp;
399 struct lwp_register_item *lri;
403 LASSERTF(strlen(lwpname) < MTI_NAME_MAXLEN, "lwpname is too long %s\n",
405 LASSERT(exp != NULL && *exp == NULL);
411 lwp = class_name2obd(lwpname);
412 if (lwp != NULL && lwp->obd_set_up == 1) {
413 struct obd_uuid *uuid;
420 memcpy(uuid->uuid, lwpname, strlen(lwpname));
421 *exp = cfs_hash_lookup(lwp->obd_uuid_hash, uuid);
425 memcpy(lri->lri_name, lwpname, strlen(lwpname));
427 lri->lri_cb_func = cb_func;
428 lri->lri_cb_data = cb_data;
429 INIT_LIST_HEAD(&lri->lri_list);
431 * Initialize the lri_ref at 2, one will be released before
432 * current function returned via lustre_put_lwp_item(), the
433 * other will be released in lustre_deregister_lwp_item().
435 atomic_set(&lri->lri_ref, 2);
437 spin_lock(&lwp_register_list_lock);
438 list_add(&lri->lri_list, &lwp_register_list);
441 spin_unlock(&lwp_register_list_lock);
443 if (cb && cb_func != NULL)
445 lustre_put_lwp_item(lri);
449 EXPORT_SYMBOL(lustre_register_lwp_item);
451 void lustre_deregister_lwp_item(struct obd_export **exp)
453 struct lwp_register_item *lri;
454 bool removed = false;
457 spin_lock(&lwp_register_list_lock);
458 list_for_each_entry(lri, &lwp_register_list, lri_list) {
459 if (exp == lri->lri_exp) {
460 list_del_init(&lri->lri_list);
465 spin_unlock(&lwp_register_list_lock);
470 /* See lustre_notify_lwp_list(), in some extreme race conditions,
471 * the notify callback could be still on the fly, we need to wait
472 * for the callback done before moving on to free the data used
474 while (atomic_read(&lri->lri_ref) > 1) {
475 CDEBUG(D_MOUNT, "lri reference count %u, repeat: %d\n",
476 atomic_read(&lri->lri_ref), repeat);
478 set_current_state(TASK_INTERRUPTIBLE);
479 schedule_timeout(cfs_time_seconds(1));
481 lustre_put_lwp_item(lri);
483 EXPORT_SYMBOL(lustre_deregister_lwp_item);
485 struct obd_export *lustre_find_lwp_by_index(const char *dev, __u32 idx)
487 struct lustre_mount_info *lmi;
488 struct lustre_sb_info *lsi;
489 struct obd_device *lwp;
490 struct obd_export *exp = NULL;
495 lmi = server_get_mount(dev);
499 lsi = s2lsi(lmi->lmi_sb);
500 rc = server_name2fsname(lsi->lsi_svname, fsname, NULL);
502 CERROR("%s: failed to get fsname: rc = %d\n",
503 lsi->lsi_svname, rc);
507 snprintf(lwp_name, sizeof(lwp_name), "%s-MDT%04x", fsname, idx);
508 spin_lock(&lsi->lsi_lwp_lock);
509 list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
510 char *ptr = strstr(lwp->obd_name, lwp_name);
512 if (ptr != NULL && lwp->obd_lwp_export != NULL) {
513 exp = class_export_get(lwp->obd_lwp_export);
517 spin_unlock(&lsi->lsi_lwp_lock);
520 server_put_mount(dev, false);
524 EXPORT_SYMBOL(lustre_find_lwp_by_index);
526 void lustre_notify_lwp_list(struct obd_export *exp)
528 struct lwp_register_item *lri;
529 LASSERT(exp != NULL);
532 spin_lock(&lwp_register_list_lock);
533 list_for_each_entry(lri, &lwp_register_list, lri_list) {
534 if (strcmp(exp->exp_obd->obd_name, lri->lri_name))
536 if (*lri->lri_exp != NULL)
538 *lri->lri_exp = class_export_get(exp);
539 if (lri->lri_cb_func == NULL)
541 atomic_inc(&lri->lri_ref);
542 spin_unlock(&lwp_register_list_lock);
544 lri->lri_cb_func(lri->lri_cb_data);
545 lustre_put_lwp_item(lri);
547 /* Others may have changed the list after we unlock, we have
548 * to rescan the list from the beginning. Usually, the list
549 * 'lwp_register_list' is very short, and there is 'guard'
550 * lri::lri_exp that will prevent the callback to be done
551 * repeatedly. So rescanning the list has no problem. */
554 spin_unlock(&lwp_register_list_lock);
556 EXPORT_SYMBOL(lustre_notify_lwp_list);
558 static int lustre_lwp_connect(struct obd_device *lwp, bool is_mdt)
561 struct lu_context session_ctx;
562 struct obd_export *exp;
563 struct obd_uuid *uuid = NULL;
564 struct obd_connect_data *data = NULL;
568 /* log has been fully processed, let clients connect */
569 rc = lu_env_init(&env, lwp->obd_lu_dev->ld_type->ldt_ctx_tags);
573 lu_context_init(&session_ctx, LCT_SERVER_SESSION);
574 session_ctx.lc_thread = NULL;
575 lu_context_enter(&session_ctx);
576 env.le_ses = &session_ctx;
580 GOTO(out, rc = -ENOMEM);
582 data->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_INDEX;
583 data->ocd_version = LUSTRE_VERSION_CODE;
584 data->ocd_connect_flags |= OBD_CONNECT_FID | OBD_CONNECT_AT |
585 OBD_CONNECT_LRU_RESIZE | OBD_CONNECT_FULL20 |
586 OBD_CONNECT_LVB_TYPE | OBD_CONNECT_LIGHTWEIGHT |
587 OBD_CONNECT_LFSCK | OBD_CONNECT_BULK_MBITS;
590 data->ocd_connect_flags |= OBD_CONNECT_MDS_MDS;
594 GOTO(out, rc = -ENOMEM);
596 if (strlen(lwp->obd_name) > sizeof(uuid->uuid)) {
597 CERROR("%s: Too long lwp name %s, max_size is %d\n",
598 lwp->obd_name, lwp->obd_name, (int)sizeof(uuid->uuid));
599 GOTO(out, rc = -EINVAL);
602 /* Use lwp name as the uuid, so we find the export by lwp name later */
603 memcpy(uuid->uuid, lwp->obd_name, strlen(lwp->obd_name));
604 rc = obd_connect(&env, &exp, lwp, uuid, data, NULL);
606 CERROR("%s: connect failed: rc = %d\n", lwp->obd_name, rc);
608 if (unlikely(lwp->obd_lwp_export != NULL))
609 class_export_put(lwp->obd_lwp_export);
610 lwp->obd_lwp_export = class_export_get(exp);
622 lu_context_exit(&session_ctx);
623 lu_context_fini(&session_ctx);
629 * lwp is used by slaves (Non-MDT0 targets) to manage the connection to MDT0,
630 * or from the OSTx to MDTy.
632 static int lustre_lwp_setup(struct lustre_cfg *lcfg, struct lustre_sb_info *lsi,
635 struct obd_device *obd;
636 char *lwpname = NULL;
637 char *lwpuuid = NULL;
641 rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
644 CERROR("%s: Can't add uuid: rc =%d\n", lsi->lsi_svname, rc);
648 OBD_ALLOC(lwpname, MTI_NAME_MAXLEN);
650 GOTO(out, rc = -ENOMEM);
652 rc = tgt_name2lwp_name(lsi->lsi_svname, lwpname, MTI_NAME_MAXLEN, idx);
654 CERROR("%s: failed to generate lwp name: rc = %d\n",
655 lsi->lsi_svname, rc);
659 OBD_ALLOC(lwpuuid, MTI_NAME_MAXLEN);
661 GOTO(out, rc = -ENOMEM);
663 sprintf(lwpuuid, "%s_UUID", lwpname);
664 rc = lustre_start_simple(lwpname, LUSTRE_LWP_NAME,
665 lwpuuid, lustre_cfg_string(lcfg, 1),
668 CERROR("%s: setup up failed: rc %d\n", lwpname, rc);
672 obd = class_name2obd(lwpname);
673 LASSERT(obd != NULL);
675 rc = lustre_lwp_connect(obd, strstr(lsi->lsi_svname, "-MDT") != NULL);
677 obd->u.cli.cl_max_mds_easize = MAX_MD_SIZE;
678 spin_lock(&lsi->lsi_lwp_lock);
679 list_add_tail(&obd->obd_lwp_list, &lsi->lsi_lwp_list);
680 spin_unlock(&lsi->lsi_lwp_lock);
682 CERROR("%s: connect failed: rc = %d\n", lwpname, rc);
689 OBD_FREE(lwpname, MTI_NAME_MAXLEN);
691 OBD_FREE(lwpuuid, MTI_NAME_MAXLEN);
696 /* the caller is responsible for memory free */
697 static struct obd_device *lustre_find_lwp(struct lustre_sb_info *lsi,
698 char **lwpname, __u32 idx)
700 struct obd_device *lwp;
704 LASSERT(lwpname != NULL);
705 LASSERT(IS_OST(lsi) || IS_MDT(lsi));
707 OBD_ALLOC(*lwpname, MTI_NAME_MAXLEN);
708 if (*lwpname == NULL)
709 RETURN(ERR_PTR(-ENOMEM));
711 rc = tgt_name2lwp_name(lsi->lsi_svname, *lwpname, MTI_NAME_MAXLEN, idx);
713 CERROR("%s: failed to generate lwp name: rc = %d\n",
714 lsi->lsi_svname, rc);
715 GOTO(out, rc = -EINVAL);
718 lwp = class_name2obd(*lwpname);
722 if (*lwpname != NULL) {
723 OBD_FREE(*lwpname, MTI_NAME_MAXLEN);
729 RETURN(lwp != NULL ? lwp : ERR_PTR(-ENOENT));
732 static int lustre_lwp_add_conn(struct lustre_cfg *cfg,
733 struct lustre_sb_info *lsi, __u32 idx)
735 struct lustre_cfg_bufs *bufs = NULL;
736 struct lustre_cfg *lcfg = NULL;
737 char *lwpname = NULL;
738 struct obd_device *lwp;
742 lwp = lustre_find_lwp(lsi, &lwpname, idx);
744 CERROR("%s: can't find lwp device.\n", lsi->lsi_svname);
745 GOTO(out, rc = PTR_ERR(lwp));
747 LASSERT(lwpname != NULL);
751 GOTO(out, rc = -ENOMEM);
753 lustre_cfg_bufs_reset(bufs, lwpname);
754 lustre_cfg_bufs_set_string(bufs, 1,
755 lustre_cfg_string(cfg, 1));
757 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
759 GOTO(out_cfg, rc = -ENOMEM);
760 lustre_cfg_init(lcfg, LCFG_ADD_CONN, bufs);
762 rc = class_add_conn(lwp, lcfg);
764 CERROR("%s: can't add conn: rc = %d\n", lwpname, rc);
767 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
768 lcfg->lcfg_buflens));
774 OBD_FREE(lwpname, MTI_NAME_MAXLEN);
779 * Retrieve MDT nids from the client log, then start the lwp device.
780 * there are only two scenarios which would include mdt nid.
782 * marker 5 (flags=0x01, v2.1.54.0) lustre-MDTyyyy 'add mdc' xxx-
783 * add_uuid nid=192.168.122.162@tcp(0x20000c0a87aa2) 0: 1:192.168.122.162@tcp
784 * attach 0:lustre-MDTyyyy-mdc 1:mdc 2:lustre-clilmv_UUID
785 * setup 0:lustre-MDTyyyy-mdc 1:lustre-MDTyyyy_UUID 2:192.168.122.162@tcp
786 * add_uuid nid=192.168.172.1@tcp(0x20000c0a8ac01) 0: 1:192.168.172.1@tcp
787 * add_conn 0:lustre-MDTyyyy-mdc 1:192.168.172.1@tcp
788 * modify_mdc_tgts add 0:lustre-clilmv 1:lustre-MDTyyyy_UUID xxxx
789 * marker 5 (flags=0x02, v2.1.54.0) lustre-MDTyyyy 'add mdc' xxxx-
791 * marker 7 (flags=0x01, v2.1.54.0) lustre-MDTyyyy 'add failnid' xxxx-
792 * add_uuid nid=192.168.122.2@tcp(0x20000c0a87a02) 0: 1:192.168.122.2@tcp
793 * add_conn 0:lustre-MDTyyyy-mdc 1:192.168.122.2@tcp
794 * marker 7 (flags=0x02, v2.1.54.0) lustre-MDTyyyy 'add failnid' xxxx-
796 static int client_lwp_config_process(const struct lu_env *env,
797 struct llog_handle *handle,
798 struct llog_rec_hdr *rec, void *data)
800 struct config_llog_instance *cfg = data;
801 int cfg_len = rec->lrh_len;
802 char *cfg_buf = (char *) (rec + 1);
803 struct lustre_cfg *lcfg = NULL;
804 struct lustre_sb_info *lsi;
805 int rc = 0, swab = 0;
808 if (rec->lrh_type != OBD_CFG_REC) {
809 CERROR("Unknown llog record type %#x encountered\n",
814 if (cfg->cfg_sb == NULL)
815 GOTO(out, rc = -EINVAL);
816 lsi = s2lsi(cfg->cfg_sb);
818 lcfg = (struct lustre_cfg *)cfg_buf;
819 if (lcfg->lcfg_version == __swab32(LUSTRE_CFG_VERSION)) {
820 lustre_swab_lustre_cfg(lcfg);
824 rc = lustre_cfg_sanity_check(cfg_buf, cfg_len);
828 switch (lcfg->lcfg_command) {
830 struct cfg_marker *marker = lustre_cfg_buf(lcfg, 1);
832 lustre_swab_cfg_marker(marker, swab,
833 LUSTRE_CFG_BUFLEN(lcfg, 1));
834 if (marker->cm_flags & CM_SKIP ||
835 marker->cm_flags & CM_EXCLUDE)
838 if (!tgt_is_mdt(marker->cm_tgtname, &cfg->cfg_lwp_idx))
841 if (IS_MDT(lsi) && cfg->cfg_lwp_idx != 0)
844 if (!strncmp(marker->cm_comment, "add mdc", 7) ||
845 !strncmp(marker->cm_comment, "add failnid", 11)) {
846 if (marker->cm_flags & CM_START) {
847 cfg->cfg_flags = CFG_F_MARKER;
848 /* This hack is to differentiate the
849 * ADD_UUID is come from "add mdc" record
850 * or from "add failnid" record. */
851 if (!strncmp(marker->cm_comment,
853 cfg->cfg_flags |= CFG_F_SKIP;
854 } else if (marker->cm_flags & CM_END) {
860 case LCFG_ADD_UUID: {
861 if (cfg->cfg_flags == CFG_F_MARKER) {
862 rc = lustre_lwp_setup(lcfg, lsi, cfg->cfg_lwp_idx);
863 /* XXX: process only the first nid as
864 * we don't need another instance of lwp */
865 cfg->cfg_flags |= CFG_F_SKIP;
866 } else if (cfg->cfg_flags == (CFG_F_MARKER | CFG_F_SKIP)) {
867 rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
870 CERROR("%s: Fail to add uuid, rc:%d\n",
871 lsi->lsi_svname, rc);
875 case LCFG_ADD_CONN: {
876 char *devname = lustre_cfg_string(lcfg, 0);
880 if (!is_mdc_device(devname))
883 ptr = strrchr(devname, '-');
888 if (!tgt_is_mdt(devname, &idx)) {
894 if (IS_MDT(lsi) && idx != 0)
897 rc = lustre_lwp_add_conn(lcfg, lsi, idx);
907 static int lustre_disconnect_lwp(struct super_block *sb)
909 struct lustre_sb_info *lsi = s2lsi(sb);
910 struct obd_device *lwp;
911 char *logname = NULL;
912 struct lustre_cfg_bufs *bufs = NULL;
913 struct config_llog_instance *cfg = NULL;
918 if (likely(lsi->lsi_lwp_started)) {
919 OBD_ALLOC(logname, MTI_NAME_MAXLEN);
923 rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
925 CERROR("%s: failed to get fsname from svname: "
926 "rc = %d\n", lsi->lsi_svname, rc);
927 GOTO(out, rc = -EINVAL);
930 strcat(logname, "-client");
933 GOTO(out, rc = -ENOMEM);
936 cfg->cfg_instance = ll_get_cfg_instance(sb);
937 rc = lustre_end_log(sb, logname, cfg);
938 if (rc != 0 && rc != -ENOENT)
941 lsi->lsi_lwp_started = 0;
946 GOTO(out, rc = -ENOMEM);
948 list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
949 struct lustre_cfg *lcfg;
951 if (likely(lwp->obd_lwp_export != NULL)) {
952 class_export_put(lwp->obd_lwp_export);
953 lwp->obd_lwp_export = NULL;
956 lustre_cfg_bufs_reset(bufs, lwp->obd_name);
957 lustre_cfg_bufs_set_string(bufs, 1, NULL);
958 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount,
961 GOTO(out, rc = -ENOMEM);
962 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
964 /* Disconnect import first. NULL is passed for the '@env',
965 * since it will not be used. */
966 rc = lwp->obd_lu_dev->ld_ops->ldo_process_config(NULL,
967 lwp->obd_lu_dev, lcfg);
968 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
969 lcfg->lcfg_buflens));
970 if (rc != 0 && rc != -ETIMEDOUT) {
971 CERROR("%s: fail to disconnect LWP: rc = %d\n",
985 OBD_FREE(logname, MTI_NAME_MAXLEN);
987 return rc1 != 0 ? rc1 : rc;
991 * Stop the lwp for an OST/MDT target.
993 static int lustre_stop_lwp(struct super_block *sb)
995 struct lustre_sb_info *lsi = s2lsi(sb);
996 struct obd_device *lwp;
1001 while (!list_empty(&lsi->lsi_lwp_list)) {
1002 lwp = list_entry(lsi->lsi_lwp_list.next, struct obd_device,
1004 list_del_init(&lwp->obd_lwp_list);
1006 rc = class_manual_cleanup(lwp);
1008 CERROR("%s: fail to stop LWP: rc = %d\n",
1014 RETURN(rc1 != 0 ? rc1 : rc);
1018 * Start the lwp(fsname-MDTyyyy-lwp-{MDT,OST}xxxx) for a MDT/OST or MDT target.
1020 static int lustre_start_lwp(struct super_block *sb)
1022 struct lustre_sb_info *lsi = s2lsi(sb);
1023 struct config_llog_instance *cfg = NULL;
1028 if (unlikely(lsi->lsi_lwp_started))
1031 OBD_ALLOC(logname, MTI_NAME_MAXLEN);
1032 if (logname == NULL)
1035 rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
1037 CERROR("%s: failed to get fsname from svname: rc = %d\n",
1038 lsi->lsi_svname, rc);
1039 GOTO(out, rc = -EINVAL);
1042 strcat(logname, "-client");
1045 GOTO(out, rc = -ENOMEM);
1047 cfg->cfg_callback = client_lwp_config_process;
1048 cfg->cfg_instance = ll_get_cfg_instance(sb);
1049 rc = lustre_process_log(sb, logname, cfg);
1050 /* need to remove config llog from mgc */
1051 lsi->lsi_lwp_started = 1;
1056 OBD_FREE(logname, MTI_NAME_MAXLEN);
1063 static DEFINE_MUTEX(server_start_lock);
1065 /* Stop MDS/OSS if nobody is using them */
1066 static int server_stop_servers(int lsiflags)
1068 struct obd_device *obd = NULL;
1069 struct obd_type *type = NULL;
1073 mutex_lock(&server_start_lock);
1075 /* Either an MDT or an OST or neither */
1076 /* if this was an MDT, and there are no more MDT's, clean up the MDS */
1077 if (lsiflags & LDD_F_SV_TYPE_MDT) {
1078 obd = class_name2obd(LUSTRE_MDS_OBDNAME);
1080 type = class_search_type(LUSTRE_MDT_NAME);
1083 /* if this was an OST, and there are no more OST's, clean up the OSS */
1084 if (lsiflags & LDD_F_SV_TYPE_OST) {
1085 obd = class_name2obd(LUSTRE_OSS_OBDNAME);
1087 type = class_search_type(LUSTRE_OST_NAME);
1090 if (obd != NULL && (type == NULL || type->typ_refcnt == 0)) {
1092 /* obd_fail doesn't mean much on a server obd */
1093 rc = class_manual_cleanup(obd);
1096 mutex_unlock(&server_start_lock);
1101 int server_mti_print(const char *title, struct mgs_target_info *mti)
1103 PRINT_CMD(PRINT_MASK, "mti %s\n", title);
1104 PRINT_CMD(PRINT_MASK, "server: %s\n", mti->mti_svname);
1105 PRINT_CMD(PRINT_MASK, "fs: %s\n", mti->mti_fsname);
1106 PRINT_CMD(PRINT_MASK, "uuid: %s\n", mti->mti_uuid);
1107 PRINT_CMD(PRINT_MASK, "ver: %d flags: %#x\n",
1108 mti->mti_config_ver, mti->mti_flags);
1112 /* Generate data for registration */
1113 static int server_lsi2mti(struct lustre_sb_info *lsi,
1114 struct mgs_target_info *mti)
1116 struct lnet_process_id id;
1121 if (!IS_SERVER(lsi))
1124 if (strlcpy(mti->mti_svname, lsi->lsi_svname, sizeof(mti->mti_svname))
1125 >= sizeof(mti->mti_svname))
1128 mti->mti_nid_count = 0;
1129 while (LNetGetId(i++, &id) != -ENOENT) {
1130 if (LNET_NETTYP(LNET_NIDNET(id.nid)) == LOLND)
1133 /* server use --servicenode param, only allow specified
1134 * nids be registered */
1135 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_NO_PRIMNODE) != 0 &&
1136 class_match_nid(lsi->lsi_lmd->lmd_params,
1137 PARAM_FAILNODE, id.nid) < 1)
1140 /* match specified network */
1141 if (!class_match_net(lsi->lsi_lmd->lmd_params,
1142 PARAM_NETWORK, LNET_NIDNET(id.nid)))
1145 mti->mti_nids[mti->mti_nid_count] = id.nid;
1146 mti->mti_nid_count++;
1147 if (mti->mti_nid_count >= MTI_NIDS_MAX) {
1148 CWARN("Only using first %d nids for %s\n",
1149 mti->mti_nid_count, mti->mti_svname);
1154 if (mti->mti_nid_count == 0) {
1155 CERROR("Failed to get NID for server %s, please check whether "
1156 "the target is specifed with improper --servicenode or "
1157 "--network options.\n", mti->mti_svname);
1161 mti->mti_lustre_ver = LUSTRE_VERSION_CODE;
1162 mti->mti_config_ver = 0;
1164 rc = server_name2fsname(lsi->lsi_svname, mti->mti_fsname, NULL);
1168 rc = server_name2index(lsi->lsi_svname, &mti->mti_stripe_index, NULL);
1171 /* Orion requires index to be set */
1172 LASSERT(!(rc & LDD_F_NEED_INDEX));
1173 /* keep only LDD flags */
1174 mti->mti_flags = lsi->lsi_flags & LDD_F_MASK;
1175 if (mti->mti_flags & (LDD_F_WRITECONF | LDD_F_VIRGIN))
1176 mti->mti_flags |= LDD_F_UPDATE;
1177 cplen = strlcpy(mti->mti_params, lsi->lsi_lmd->lmd_params,
1178 sizeof(mti->mti_params));
1179 if (cplen >= sizeof(mti->mti_params))
1184 /* Register an old or new target with the MGS. If needed MGS will construct
1185 startup logs and assign index */
1186 static int server_register_target(struct lustre_sb_info *lsi)
1188 struct obd_device *mgc = lsi->lsi_mgc;
1189 struct mgs_target_info *mti = NULL;
1197 if (!IS_SERVER(lsi))
1204 rc = server_lsi2mti(lsi, mti);
1208 CDEBUG(D_MOUNT, "Registration %s, fs=%s, %s, index=%04x, flags=%#x\n",
1209 mti->mti_svname, mti->mti_fsname,
1210 libcfs_nid2str(mti->mti_nids[0]), mti->mti_stripe_index,
1213 /* if write_conf is true, the registration must succeed */
1214 writeconf = !!(lsi->lsi_flags & (LDD_F_NEED_INDEX | LDD_F_UPDATE));
1215 mti->mti_flags |= LDD_F_OPC_REG;
1218 /* Register the target */
1219 /* FIXME use mgc_process_config instead */
1220 rc = obd_set_info_async(NULL, mgc->u.cli.cl_mgc_mgsexp,
1221 sizeof(KEY_REGISTER_TARGET),
1222 KEY_REGISTER_TARGET,
1223 sizeof(*mti), mti, NULL);
1225 if (mti->mti_flags & LDD_F_ERROR) {
1226 LCONSOLE_ERROR_MSG(0x160,
1227 "%s: the MGS refuses to allow this server "
1228 "to start: rc = %d. Please see messages on "
1229 "the MGS.\n", lsi->lsi_svname, rc);
1230 } else if (writeconf) {
1231 if ((rc == -ESHUTDOWN || rc == -EIO) && ++tried < 5) {
1232 /* The connection with MGS is not established.
1233 * Try again after 2 seconds. Interruptable. */
1234 set_current_state(TASK_INTERRUPTIBLE);
1235 schedule_timeout(cfs_time_seconds(2));
1236 set_current_state(TASK_RUNNING);
1237 if (!signal_pending(current))
1241 LCONSOLE_ERROR_MSG(0x15f,
1242 "%s: cannot register this server with the MGS: "
1243 "rc = %d. Is the MGS running?\n",
1244 lsi->lsi_svname, rc);
1246 CDEBUG(D_HA, "%s: error registering with the MGS: "
1247 "rc = %d (not fatal)\n", lsi->lsi_svname, rc);
1248 /* reset the error code for non-fatal error. */
1261 * Notify the MGS that this target is ready.
1262 * Used by IR - if the MGS receives this message, it will notify clients.
1264 static int server_notify_target(struct super_block *sb, struct obd_device *obd)
1266 struct lustre_sb_info *lsi = s2lsi(sb);
1267 struct obd_device *mgc = lsi->lsi_mgc;
1268 struct mgs_target_info *mti = NULL;
1274 if (!(IS_SERVER(lsi)))
1280 rc = server_lsi2mti(lsi, mti);
1284 mti->mti_instance = obd->u.obt.obt_instance;
1285 mti->mti_flags |= LDD_F_OPC_READY;
1287 /* FIXME use mgc_process_config instead */
1288 rc = obd_set_info_async(NULL, mgc->u.cli.cl_mgc_mgsexp,
1289 sizeof(KEY_REGISTER_TARGET),
1290 KEY_REGISTER_TARGET,
1291 sizeof(*mti), mti, NULL);
1293 /* Imperative recovery: if the mgs informs us to use IR? */
1294 if (!rc && !(mti->mti_flags & LDD_F_ERROR) &&
1295 (mti->mti_flags & LDD_F_IR_CAPABLE))
1296 lsi->lsi_flags |= LDD_F_IR_CAPABLE;
1305 /** Start server targets: MDTs and OSTs
1307 static int server_start_targets(struct super_block *sb)
1309 struct obd_device *obd;
1310 struct lustre_sb_info *lsi = s2lsi(sb);
1311 struct config_llog_instance cfg;
1312 struct lu_env mgc_env;
1313 struct lu_device *dev;
1317 CDEBUG(D_MOUNT, "starting target %s\n", lsi->lsi_svname);
1320 /* make sure the MDS is started */
1321 mutex_lock(&server_start_lock);
1322 obd = class_name2obd(LUSTRE_MDS_OBDNAME);
1324 rc = lustre_start_simple(LUSTRE_MDS_OBDNAME,
1326 LUSTRE_MDS_OBDNAME"_uuid",
1327 NULL, NULL, NULL, NULL);
1329 mutex_unlock(&server_start_lock);
1330 CERROR("failed to start MDS: %d\n", rc);
1334 mutex_unlock(&server_start_lock);
1337 /* If we're an OST, make sure the global OSS is running */
1339 /* make sure OSS is started */
1340 mutex_lock(&server_start_lock);
1341 obd = class_name2obd(LUSTRE_OSS_OBDNAME);
1343 rc = lustre_start_simple(LUSTRE_OSS_OBDNAME,
1345 LUSTRE_OSS_OBDNAME"_uuid",
1346 NULL, NULL, NULL, NULL);
1348 mutex_unlock(&server_start_lock);
1349 CERROR("failed to start OSS: %d\n", rc);
1353 mutex_unlock(&server_start_lock);
1356 rc = lu_env_init(&mgc_env, LCT_MG_THREAD);
1358 GOTO(out_stop_service, rc);
1360 /* Set the mgc fs to our server disk. This allows the MGC to
1361 * read and write configs locally, in case it can't talk to the MGS. */
1362 rc = server_mgc_set_fs(&mgc_env, lsi->lsi_mgc, sb);
1366 /* Register with MGS */
1367 rc = server_register_target(lsi);
1371 /* Let the target look up the mount using the target's name
1372 (we can't pass the sb or mnt through class_process_config.) */
1373 rc = server_register_mount(lsi->lsi_svname, sb);
1377 /* Start targets using the llog named for the target */
1378 memset(&cfg, 0, sizeof(cfg));
1379 cfg.cfg_callback = class_config_llog_handler;
1380 cfg.cfg_sub_clds = CONFIG_SUB_SERVER;
1381 rc = lustre_process_log(sb, lsi->lsi_svname, &cfg);
1383 CERROR("failed to start server %s: %d\n",
1384 lsi->lsi_svname, rc);
1385 /* Do NOT call server_deregister_mount() here. This makes it
1386 * impossible to find mount later in cleanup time and leaves
1387 * @lsi and othder stuff leaked. -umka */
1391 obd = class_name2obd(lsi->lsi_svname);
1393 CERROR("no server named %s was started\n", lsi->lsi_svname);
1394 GOTO(out_mgc, rc = -ENXIO);
1397 if (IS_OST(lsi) || IS_MDT(lsi)) {
1398 rc = lustre_start_lwp(sb);
1400 CERROR("%s: failed to start LWP: %d\n",
1401 lsi->lsi_svname, rc);
1406 server_notify_target(sb, obd);
1408 /* calculate recovery timeout, do it after lustre_process_log */
1409 server_calc_timeout(lsi, obd);
1411 /* log has been fully processed, let clients connect */
1412 dev = obd->obd_lu_dev;
1413 if (dev && dev->ld_ops->ldo_prepare) {
1416 rc = lu_env_init(&env, dev->ld_type->ldt_ctx_tags);
1418 struct lu_context session_ctx;
1420 lu_context_init(&session_ctx, LCT_SERVER_SESSION);
1421 session_ctx.lc_thread = NULL;
1422 lu_context_enter(&session_ctx);
1423 env.le_ses = &session_ctx;
1425 rc = dev->ld_ops->ldo_prepare(&env, NULL, dev);
1428 lu_context_exit(&session_ctx);
1429 lu_context_fini(&session_ctx);
1433 /* abort recovery only on the complete stack:
1434 * many devices can be involved */
1435 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_ABORT_RECOV) &&
1436 (OBP(obd, iocontrol))) {
1437 obd_iocontrol(OBD_IOC_ABORT_RECOVERY, obd->obd_self_export, 0,
1442 /* Release the mgc fs for others to use */
1443 server_mgc_clear_fs(&mgc_env, lsi->lsi_mgc);
1445 lu_env_fini(&mgc_env);
1448 server_stop_servers(lsi->lsi_flags);
1453 static int lsi_prepare(struct lustre_sb_info *lsi)
1455 const char *osd_type;
1462 LASSERT(lsi->lsi_lmd);
1464 /* The server name is given as a mount line option */
1465 if (lsi->lsi_lmd->lmd_profile == NULL) {
1466 LCONSOLE_ERROR("Can't determine server name\n");
1470 /* Determine osd type */
1471 if (lsi->lsi_lmd->lmd_osd_type == NULL) {
1472 osd_type = LUSTRE_OSD_LDISKFS_NAME;
1475 osd_type = lsi->lsi_lmd->lmd_osd_type;
1476 fstype = lsi->lsi_lmd->lmd_osd_type;
1479 if (strlen(lsi->lsi_lmd->lmd_profile) >= sizeof(lsi->lsi_svname) ||
1480 strlen(osd_type) >= sizeof(lsi->lsi_osd_type) ||
1481 strlen(fstype) >= sizeof(lsi->lsi_fstype))
1482 RETURN(-ENAMETOOLONG);
1484 strlcpy(lsi->lsi_svname, lsi->lsi_lmd->lmd_profile,
1485 sizeof(lsi->lsi_svname));
1486 strlcpy(lsi->lsi_osd_type, osd_type, sizeof(lsi->lsi_osd_type));
1487 /* XXX: a temp. solution for components using ldiskfs
1488 * to be removed in one of the subsequent patches */
1489 strlcpy(lsi->lsi_fstype, fstype, sizeof(lsi->lsi_fstype));
1491 /* Determine server type */
1492 rc = server_name2index(lsi->lsi_svname, &index, NULL);
1494 if (lsi->lsi_lmd->lmd_flags & LMD_FLG_MGS) {
1495 /* Assume we're a bare MGS */
1497 lsi->lsi_lmd->lmd_flags |= LMD_FLG_NOSVC;
1499 LCONSOLE_ERROR("Can't determine server type of '%s'\n",
1504 lsi->lsi_flags |= rc;
1506 /* Add mount line flags that used to be in ldd:
1507 * writeconf, mgs, anything else?
1509 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_WRITECONF) ?
1510 LDD_F_WRITECONF : 0;
1511 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_VIRGIN) ?
1513 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_UPDATE) ?
1515 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_MGS) ?
1516 LDD_F_SV_TYPE_MGS : 0;
1517 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_NO_PRIMNODE) ?
1518 LDD_F_NO_PRIMNODE : 0;
1523 /*************** server mount ******************/
1525 /** Start the shutdown of servers at umount.
1527 static void server_put_super(struct super_block *sb)
1529 struct lustre_sb_info *lsi = s2lsi(sb);
1530 struct obd_device *obd;
1531 char *tmpname, *extraname = NULL;
1533 int lsiflags = lsi->lsi_flags;
1536 LASSERT(IS_SERVER(lsi));
1538 tmpname_sz = strlen(lsi->lsi_svname) + 1;
1539 OBD_ALLOC(tmpname, tmpname_sz);
1540 memcpy(tmpname, lsi->lsi_svname, tmpname_sz);
1541 CDEBUG(D_MOUNT, "server put_super %s\n", tmpname);
1542 if (IS_MDT(lsi) && (lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC))
1543 snprintf(tmpname, tmpname_sz, "MGS");
1545 /* disconnect the lwp first to drain off the inflight request */
1546 if (IS_OST(lsi) || IS_MDT(lsi)) {
1549 rc = lustre_disconnect_lwp(sb);
1550 if (rc != 0 && rc != -ETIMEDOUT &&
1551 rc != -ENOTCONN && rc != -ESHUTDOWN)
1552 CWARN("%s: failed to disconnect lwp: rc= %d\n",
1556 /* Stop the target */
1557 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1558 (IS_MDT(lsi) || IS_OST(lsi))) {
1559 struct lustre_profile *lprof = NULL;
1561 /* tell the mgc to drop the config log */
1562 lustre_end_log(sb, lsi->lsi_svname, NULL);
1564 /* COMPAT_146 - profile may get deleted in mgc_cleanup.
1565 If there are any setup/cleanup errors, save the lov
1566 name for safety cleanup later. */
1567 lprof = class_get_profile(lsi->lsi_svname);
1568 if (lprof != NULL) {
1569 if (lprof->lp_dt != NULL) {
1570 OBD_ALLOC(extraname, strlen(lprof->lp_dt) + 1);
1571 strncpy(extraname, lprof->lp_dt,
1572 strlen(lprof->lp_dt) + 1);
1574 class_put_profile(lprof);
1577 obd = class_name2obd(lsi->lsi_svname);
1579 CDEBUG(D_MOUNT, "stopping %s\n", obd->obd_name);
1580 if (lsiflags & LSI_UMOUNT_FAILOVER)
1582 /* We can't seem to give an error return code
1583 * to .put_super, so we better make sure we clean up! */
1585 class_manual_cleanup(obd);
1587 CERROR("no obd %s\n", lsi->lsi_svname);
1588 server_deregister_mount(lsi->lsi_svname);
1592 /* If they wanted the mgs to stop separately from the mdt, they
1593 should have put it on a different device. */
1595 /* if MDS start with --nomgs, don't stop MGS then */
1596 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS))
1597 server_stop_mgs(sb);
1600 if (IS_OST(lsi) || IS_MDT(lsi)) {
1601 if (lustre_stop_lwp(sb) < 0)
1602 CERROR("%s: failed to stop lwp!\n", tmpname);
1605 /* Clean the mgc and sb */
1606 lustre_common_put_super(sb);
1608 /* wait till all in-progress cleanups are done
1609 * specifically we're interested in ofd cleanup
1611 obd_zombie_barrier();
1613 /* Stop the servers (MDS, OSS) if no longer needed. We must wait
1614 until the target is really gone so that our type refcount check
1616 server_stop_servers(lsiflags);
1618 /* In case of startup or cleanup err, stop related obds */
1620 obd = class_name2obd(extraname);
1622 CWARN("Cleaning orphaned obd %s\n", extraname);
1624 class_manual_cleanup(obd);
1626 OBD_FREE(extraname, strlen(extraname) + 1);
1629 LCONSOLE_WARN("server umount %s complete\n", tmpname);
1630 OBD_FREE(tmpname, tmpname_sz);
1634 /** Called only for 'umount -f'
1636 static void server_umount_begin(struct super_block *sb)
1638 struct lustre_sb_info *lsi = s2lsi(sb);
1641 CDEBUG(D_MOUNT, "umount -f\n");
1642 /* umount = failover
1644 no third way to do non-force, non-failover */
1645 lsi->lsi_flags &= ~LSI_UMOUNT_FAILOVER;
1649 static int server_statfs(struct dentry *dentry, struct kstatfs *buf)
1651 struct super_block *sb = dentry->d_sb;
1652 struct lustre_sb_info *lsi = s2lsi(sb);
1653 struct obd_statfs statfs;
1657 if (lsi->lsi_dt_dev) {
1658 rc = dt_statfs(NULL, lsi->lsi_dt_dev, &statfs);
1660 statfs_unpack(buf, &statfs);
1661 buf->f_type = sb->s_magic;
1667 buf->f_type = sb->s_magic;
1668 buf->f_bsize = sb->s_blocksize;
1674 buf->f_namelen = NAME_MAX;
1678 #ifdef HAVE_SUPEROPS_USE_DENTRY
1679 int server_show_options(struct seq_file *seq, struct dentry *dentry)
1681 int server_show_options(struct seq_file *seq, struct vfsmount *vfs)
1684 struct lustre_sb_info *lsi;
1685 struct lustre_mount_data *lmd;
1687 #ifdef HAVE_SUPEROPS_USE_DENTRY
1688 LASSERT(seq != NULL && dentry != NULL);
1689 lsi = s2lsi(dentry->d_sb);
1691 LASSERT(seq != NULL && vfs != NULL);
1692 lsi = s2lsi(vfs->mnt_sb);
1696 seq_printf(seq, ",svname=%s", lmd->lmd_profile);
1698 if (lmd->lmd_flags & LMD_FLG_ABORT_RECOV)
1699 seq_puts(seq, ",abort_recov");
1701 if (lmd->lmd_flags & LMD_FLG_NOIR)
1702 seq_puts(seq, ",noir");
1704 if (lmd->lmd_flags & LMD_FLG_NOSVC)
1705 seq_puts(seq, ",nosvc");
1707 if (lmd->lmd_flags & LMD_FLG_NOMGS)
1708 seq_puts(seq, ",nomgs");
1710 if (lmd->lmd_flags & LMD_FLG_NOSCRUB)
1711 seq_puts(seq, ",noscrub");
1712 if (lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
1713 seq_puts(seq, ",skip_lfsck");
1715 if (lmd->lmd_flags & LMD_FLG_DEV_RDONLY)
1716 seq_puts(seq, ",rdonly_dev");
1718 if (lmd->lmd_flags & LMD_FLG_MGS)
1719 seq_puts(seq, ",mgs");
1721 if (lmd->lmd_mgs != NULL)
1722 seq_printf(seq, ",mgsnode=%s", lmd->lmd_mgs);
1724 if (lmd->lmd_osd_type != NULL)
1725 seq_printf(seq, ",osd=%s", lmd->lmd_osd_type);
1727 if (lmd->lmd_opts != NULL) {
1729 seq_puts(seq, lmd->lmd_opts);
1735 /** The operations we support directly on the superblock:
1736 * mount, umount, and df.
1738 static struct super_operations server_ops = {
1739 .put_super = server_put_super,
1740 .umount_begin = server_umount_begin, /* umount -f */
1741 .statfs = server_statfs,
1742 .show_options = server_show_options,
1746 * Xattr support for Lustre servers
1748 #ifdef HAVE_IOP_XATTR
1749 static ssize_t lustre_getxattr(struct dentry *dentry, const char *name,
1750 void *buffer, size_t size)
1752 if (!selinux_is_enabled())
1757 static int lustre_setxattr(struct dentry *dentry, const char *name,
1758 const void *value, size_t size, int flags)
1764 static ssize_t lustre_listxattr(struct dentry *d_entry, char *name,
1770 static bool is_cmd_supported(unsigned int command)
1782 static long server_ioctl(struct file *filp, unsigned int command,
1785 struct file active_filp;
1786 struct inode *inode = file_inode(filp);
1787 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
1788 struct super_block *dd_sb = dt_mnt_sb_get(lsi->lsi_dt_dev);
1789 struct inode *active_inode;
1790 int err = -EOPNOTSUPP;
1792 if (IS_ERR(dd_sb) || !is_cmd_supported(command))
1795 active_inode = igrab(dd_sb->s_root->d_inode);
1799 active_filp.f_inode = active_inode;
1800 if (active_inode->i_fop && active_inode->i_fop->unlocked_ioctl)
1801 err = active_inode->i_fop->unlocked_ioctl(&active_filp,
1807 static const struct inode_operations server_inode_operations = {
1808 #ifdef HAVE_IOP_XATTR
1809 .setxattr = lustre_setxattr,
1810 .getxattr = lustre_getxattr,
1812 .listxattr = lustre_listxattr,
1815 static const struct file_operations server_file_operations = {
1816 .unlocked_ioctl = server_ioctl,
1819 #define log2(n) ffz(~(n))
1820 #define LUSTRE_SUPER_MAGIC 0x0BD00BD1
1822 static int server_fill_super_common(struct super_block *sb)
1824 struct inode *root = NULL;
1827 CDEBUG(D_MOUNT, "Server sb, dev=%d\n", (int)sb->s_dev);
1829 sb->s_blocksize = 4096;
1830 sb->s_blocksize_bits = log2(sb->s_blocksize);
1831 sb->s_magic = LUSTRE_SUPER_MAGIC;
1832 sb->s_maxbytes = 0; /* we don't allow file IO on server mountpoints */
1833 sb->s_flags |= SB_RDONLY;
1834 sb->s_op = &server_ops;
1836 root = new_inode(sb);
1838 CERROR("Can't make root inode\n");
1842 /* returns -EIO for every operation */
1843 /* make_bad_inode(root); -- badness - can't umount */
1844 /* apparently we need to be a directory for the mount to finish */
1845 root->i_mode = S_IFDIR;
1846 root->i_op = &server_inode_operations;
1847 root->i_fop = &server_file_operations;
1848 sb->s_root = d_make_root(root);
1850 CERROR("%s: can't make root dentry\n", sb->s_id);
1857 static int osd_start(struct lustre_sb_info *lsi, unsigned long mflags)
1859 struct lustre_mount_data *lmd = lsi->lsi_lmd;
1860 struct obd_device *obd;
1861 struct dt_device_param p;
1862 char flagstr[20 + 1 + 10 + 1];
1867 "Attempting to start %s, type=%s, lsifl=%x, mountfl=%lx\n",
1868 lsi->lsi_svname, lsi->lsi_osd_type, lsi->lsi_flags, mflags);
1870 sprintf(lsi->lsi_osd_obdname, "%s-osd", lsi->lsi_svname);
1871 strcpy(lsi->lsi_osd_uuid, lsi->lsi_osd_obdname);
1872 strcat(lsi->lsi_osd_uuid, "_UUID");
1873 snprintf(flagstr, sizeof(flagstr), "%lu:%u", mflags, lmd->lmd_flags);
1875 obd = class_name2obd(lsi->lsi_osd_obdname);
1877 rc = lustre_start_simple(lsi->lsi_osd_obdname,
1879 lsi->lsi_osd_uuid, lmd->lmd_dev,
1880 flagstr, lsi->lsi_lmd->lmd_opts,
1884 obd = class_name2obd(lsi->lsi_osd_obdname);
1887 CDEBUG(D_MOUNT, "%s already started\n", lsi->lsi_osd_obdname);
1888 /* but continue setup to allow special case of MDT and internal
1889 * MGT being started separately. */
1890 if (!((IS_MGS(lsi) && (lsi->lsi_lmd->lmd_flags &
1892 (IS_MDT(lsi) && (lsi->lsi_lmd->lmd_flags &
1897 rc = obd_connect(NULL, &lsi->lsi_osd_exp,
1898 obd, &obd->obd_uuid, NULL, NULL);
1902 class_manual_cleanup(obd);
1903 lsi->lsi_dt_dev = NULL;
1907 LASSERT(obd->obd_lu_dev);
1908 lu_device_get(obd->obd_lu_dev);
1909 lsi->lsi_dt_dev = lu2dt_dev(obd->obd_lu_dev);
1910 LASSERT(lsi->lsi_dt_dev);
1912 /* set disk context for llog usage */
1913 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1914 obd->obd_lvfs_ctxt.dt = lsi->lsi_dt_dev;
1916 dt_conf_get(NULL, lsi->lsi_dt_dev, &p);
1921 /** Fill in the superblock info for a Lustre server.
1922 * Mount the device with the correct options.
1923 * Read the on-disk config file.
1924 * Start the services.
1926 int server_fill_super(struct super_block *sb)
1928 struct lustre_sb_info *lsi = s2lsi(sb);
1932 /* to simulate target mount race */
1933 OBD_RACE(OBD_FAIL_TGT_MOUNT_RACE);
1935 rc = lsi_prepare(lsi);
1941 /* Start low level OSD */
1942 rc = osd_start(lsi, sb->s_flags);
1944 CERROR("Unable to start osd on %s: %d\n",
1945 lsi->lsi_lmd->lmd_dev, rc);
1950 CDEBUG(D_MOUNT, "Found service %s on device %s\n",
1951 lsi->lsi_svname, lsi->lsi_lmd->lmd_dev);
1953 if (class_name2obd(lsi->lsi_svname)) {
1954 LCONSOLE_ERROR_MSG(0x161, "The target named %s is already "
1955 "running. Double-mount may have compromised"
1956 " the disk journal.\n",
1962 /* Start MGS before MGC */
1963 if (IS_MGS(lsi) && !(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS)) {
1964 rc = server_start_mgs(sb);
1969 /* Start MGC before servers */
1970 rc = lustre_start_mgc(sb);
1974 /* Set up all obd devices for service */
1975 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1976 (IS_OST(lsi) || IS_MDT(lsi))) {
1977 rc = server_start_targets(sb);
1979 CERROR("Unable to start targets: %d\n", rc);
1982 /* FIXME overmount client here, or can we just start a
1983 * client log and client_fill_super on this sb? We
1984 * need to make sure server_put_super gets called too
1985 * - ll_put_super calls lustre_common_put_super; check
1986 * there for LSI_SERVER flag, call s_p_s if so.
1988 * Probably should start client from new thread so we
1989 * can return. Client will not finish until all
1990 * servers are connected. Note - MGS-only server does
1991 * NOT get a client, since there is no lustre fs
1992 * associated - the MGS is for all lustre fs's */
1995 rc = server_fill_super_common(sb);
2001 /* We jump here in case of failure while starting targets or MGS.
2002 * In this case we can't just put @mnt and have to do real cleanup
2003 * with stoping targets, etc. */
2004 server_put_super(sb);
2009 * Calculate timeout value for a target.
2011 void server_calc_timeout(struct lustre_sb_info *lsi, struct obd_device *obd)
2013 struct lustre_mount_data *lmd;
2017 bool has_ir = !!(lsi->lsi_flags & LDD_F_IR_CAPABLE);
2018 int min = OBD_RECOVERY_TIME_MIN;
2020 LASSERT(IS_SERVER(lsi));
2024 soft = lmd->lmd_recovery_time_soft;
2025 hard = lmd->lmd_recovery_time_hard;
2026 has_ir = has_ir && !(lmd->lmd_flags & LMD_FLG_NOIR);
2027 obd->obd_no_ir = !has_ir;
2031 soft = OBD_RECOVERY_TIME_SOFT;
2033 hard = OBD_RECOVERY_TIME_HARD;
2035 /* target may have ir_factor configured. */
2036 factor = OBD_IR_FACTOR_DEFAULT;
2037 if (obd->obd_recovery_ir_factor)
2038 factor = obd->obd_recovery_ir_factor;
2041 int new_soft = soft;
2043 /* adjust timeout value by imperative recovery */
2044 new_soft = (soft * factor) / OBD_IR_FACTOR_MAX;
2045 /* make sure the timeout is not too short */
2046 new_soft = max(min, new_soft);
2048 LCONSOLE_INFO("%s: Imperative Recovery enabled, recovery "
2049 "window shrunk from %d-%d down to %d-%d\n",
2050 obd->obd_name, soft, hard, new_soft, hard);
2054 LCONSOLE_INFO("%s: Imperative Recovery not enabled, recovery "
2055 "window %d-%d\n", obd->obd_name, soft, hard);
2059 obd->obd_recovery_timeout = max_t(time64_t, obd->obd_recovery_timeout,
2061 obd->obd_recovery_time_hard = hard;
2062 obd->obd_recovery_ir_factor = factor;