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 #include <linux/selinux.h>
47 #include <linux/statfs.h>
48 #include <linux/version.h>
50 #include <llog_swab.h>
51 #include <lustre_disk.h>
52 #include <uapi/linux/lustre/lustre_ioctl.h>
53 #include <lustre_log.h>
54 #include <uapi/linux/lustre/lustre_param.h>
56 #include <obd_class.h>
58 /*********** mount lookup *********/
60 static DEFINE_MUTEX(lustre_mount_info_lock);
61 static LIST_HEAD(server_mount_info_list);
63 static struct lustre_mount_info *server_find_mount(const char *name)
65 struct list_head *tmp;
66 struct lustre_mount_info *lmi;
69 list_for_each(tmp, &server_mount_info_list) {
70 lmi = list_entry(tmp, struct lustre_mount_info,
72 if (strcmp(name, lmi->lmi_name) == 0)
78 /* we must register an obd for a mount before we call the setup routine.
79 *_setup will call lustre_get_mount to get the mnt struct
80 by obd_name, since we can't pass the pointer to setup. */
81 static int server_register_mount(const char *name, struct super_block *sb)
83 struct lustre_mount_info *lmi;
89 OBD_ALLOC(lmi, sizeof(*lmi));
92 OBD_ALLOC(name_cp, strlen(name) + 1);
94 OBD_FREE(lmi, sizeof(*lmi));
97 strcpy(name_cp, name);
99 mutex_lock(&lustre_mount_info_lock);
101 if (server_find_mount(name)) {
102 mutex_unlock(&lustre_mount_info_lock);
103 OBD_FREE(lmi, sizeof(*lmi));
104 OBD_FREE(name_cp, strlen(name) + 1);
105 CERROR("Already registered %s\n", name);
108 lmi->lmi_name = name_cp;
110 list_add(&lmi->lmi_list_chain, &server_mount_info_list);
112 mutex_unlock(&lustre_mount_info_lock);
114 CDEBUG(D_MOUNT, "register mount %p from %s\n", sb, name);
119 /* when an obd no longer needs a mount */
120 static int server_deregister_mount(const char *name)
122 struct lustre_mount_info *lmi;
125 mutex_lock(&lustre_mount_info_lock);
126 lmi = server_find_mount(name);
128 mutex_unlock(&lustre_mount_info_lock);
129 CERROR("%s not registered\n", name);
133 CDEBUG(D_MOUNT, "deregister mount %p from %s\n", lmi->lmi_sb, name);
135 OBD_FREE(lmi->lmi_name, strlen(lmi->lmi_name) + 1);
136 list_del(&lmi->lmi_list_chain);
137 OBD_FREE(lmi, sizeof(*lmi));
138 mutex_unlock(&lustre_mount_info_lock);
143 /* obd's look up a registered mount using their obdname. This is just
144 for initial obd setup to find the mount struct. It should not be
145 called every time you want to mntget. */
146 struct lustre_mount_info *server_get_mount(const char *name)
148 struct lustre_mount_info *lmi;
149 struct lustre_sb_info *lsi;
152 mutex_lock(&lustre_mount_info_lock);
153 lmi = server_find_mount(name);
154 mutex_unlock(&lustre_mount_info_lock);
156 CERROR("Can't find mount for %s\n", name);
159 lsi = s2lsi(lmi->lmi_sb);
161 atomic_inc(&lsi->lsi_mounts);
163 CDEBUG(D_MOUNT, "get mount %p from %s, refs=%d\n", lmi->lmi_sb,
164 name, atomic_read(&lsi->lsi_mounts));
168 EXPORT_SYMBOL(server_get_mount);
171 * server_put_mount: to be called from obd_cleanup methods
173 * @dereg_mnt: 0 or 1 depending on whether the mount is to be deregistered or
176 * The caller decides whether server_deregister_mount() needs to be called or
177 * not. Calling of server_deregister_mount() does not depend on refcounting on
178 * lsi because we could have say the mgs and mds on the same node and we
179 * unmount the mds, then the ref on the lsi would still be non-zero but we
180 * would still want to deregister the mds mount.
182 int server_put_mount(const char *name, bool dereg_mnt)
184 struct lustre_mount_info *lmi;
185 struct lustre_sb_info *lsi;
188 mutex_lock(&lustre_mount_info_lock);
189 lmi = server_find_mount(name);
190 mutex_unlock(&lustre_mount_info_lock);
192 CERROR("Can't find mount for %s\n", name);
195 lsi = s2lsi(lmi->lmi_sb);
197 CDEBUG(D_MOUNT, "put mount %p from %s, refs=%d\n",
198 lmi->lmi_sb, name, atomic_read(&lsi->lsi_mounts));
200 if (lustre_put_lsi(lmi->lmi_sb))
201 CDEBUG(D_MOUNT, "Last put of mount %p from %s\n",
205 /* this obd should never need the mount again */
206 server_deregister_mount(name);
210 EXPORT_SYMBOL(server_put_mount);
212 /* Set up a MGS to serve startup logs */
213 static int server_start_mgs(struct super_block *sb)
215 struct lustre_sb_info *lsi = s2lsi(sb);
216 struct lustre_mount_info *lmi;
220 /* It is impossible to have more than 1 MGS per node, since
221 MGC wouldn't know which to connect to */
222 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
224 lsi = s2lsi(lmi->lmi_sb);
225 LCONSOLE_ERROR_MSG(0x15d, "The MGS service was already started"
230 CDEBUG(D_CONFIG, "Start MGS service %s\n", LUSTRE_MGS_OBDNAME);
232 rc = server_register_mount(LUSTRE_MGS_OBDNAME, sb);
235 rc = lustre_start_simple(LUSTRE_MGS_OBDNAME, LUSTRE_MGS_NAME,
236 LUSTRE_MGS_OBDNAME, NULL, NULL,
237 lsi->lsi_osd_obdname, NULL);
238 /* server_deregister_mount() is not called previously, for lsi
239 * and other stuff can't be freed cleanly when mgs calls
240 * server_put_mount() in error handling case (see b=17758),
241 * this problem is caused by a bug in mgs_init0, which forgot
242 * calling server_put_mount in error case. */
245 server_deregister_mount(LUSTRE_MGS_OBDNAME);
249 LCONSOLE_ERROR_MSG(0x15e, "Failed to start MGS '%s' (%d). "
250 "Is the 'mgs' module loaded?\n",
251 LUSTRE_MGS_OBDNAME, rc);
255 static int server_stop_mgs(struct super_block *sb)
257 struct obd_device *obd;
259 struct lustre_mount_info *lmi;
262 /* Do not stop MGS if this device is not the running MGT */
263 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
264 if (lmi != NULL && lmi->lmi_sb != sb)
267 CDEBUG(D_MOUNT, "Stop MGS service %s\n", LUSTRE_MGS_OBDNAME);
269 /* There better be only one MGS */
270 obd = class_name2obd(LUSTRE_MGS_OBDNAME);
272 CDEBUG(D_CONFIG, "mgs %s not running\n", LUSTRE_MGS_OBDNAME);
276 /* The MGS should always stop when we say so */
278 rc = class_manual_cleanup(obd);
282 /* Since there's only one mgc per node, we have to change it's fs to get
283 access to the right disk. */
284 static int server_mgc_set_fs(const struct lu_env *env,
285 struct obd_device *mgc, struct super_block *sb)
287 struct lustre_sb_info *lsi = s2lsi(sb);
291 CDEBUG(D_MOUNT, "Set mgc disk for %s\n", lsi->lsi_lmd->lmd_dev);
293 /* cl_mgc_sem in mgc insures we sleep if the mgc_fs is busy */
294 rc = obd_set_info_async(env, mgc->obd_self_export,
295 sizeof(KEY_SET_FS), KEY_SET_FS,
296 sizeof(*sb), sb, NULL);
298 CERROR("can't set_fs %d\n", rc);
303 static int server_mgc_clear_fs(const struct lu_env *env,
304 struct obd_device *mgc)
309 CDEBUG(D_MOUNT, "Unassign mgc disk\n");
311 rc = obd_set_info_async(env, mgc->obd_self_export,
312 sizeof(KEY_CLEAR_FS), KEY_CLEAR_FS,
317 static inline bool is_mdc_device(const char *devname)
321 ptr = strrchr(devname, '-');
322 return ptr != NULL && strcmp(ptr, "-mdc") == 0;
325 static inline bool tgt_is_mdt(const char *tgtname, __u32 *idx)
329 type = server_name2index(tgtname, idx, NULL);
331 return type == LDD_F_SV_TYPE_MDT;
335 * Convert OST/MDT name(fsname-{MDT,OST}xxxx) to a lwp name with the @idx:yyyy
336 * (fsname-MDTyyyy-lwp-{MDT,OST}xxxx)
338 int tgt_name2lwp_name(const char *tgt_name, char *lwp_name, int len, __u32 idx)
345 OBD_ALLOC(fsname, MTI_NAME_MAXLEN);
349 rc = server_name2fsname(tgt_name, fsname, &tgt);
351 CERROR("%s: failed to get fsname from tgt_name: rc = %d\n",
356 if (*tgt != '-' && *tgt != ':') {
357 CERROR("%s: invalid tgt_name name!\n", tgt_name);
358 GOTO(cleanup, rc = -EINVAL);
362 if (strncmp(tgt, "OST", 3) != 0 && strncmp(tgt, "MDT", 3) != 0) {
363 CERROR("%s is not an OST or MDT target!\n", tgt_name);
364 GOTO(cleanup, rc = -EINVAL);
366 snprintf(lwp_name, len, "%s-MDT%04x-%s-%s",
367 fsname, idx, LUSTRE_LWP_NAME, tgt);
369 GOTO(cleanup, rc = 0);
373 OBD_FREE(fsname, MTI_NAME_MAXLEN);
377 EXPORT_SYMBOL(tgt_name2lwp_name);
379 static LIST_HEAD(lwp_register_list);
380 static DEFINE_SPINLOCK(lwp_register_list_lock);
382 static void lustre_put_lwp_item(struct lwp_register_item *lri)
384 if (atomic_dec_and_test(&lri->lri_ref)) {
385 LASSERT(list_empty(&lri->lri_list));
387 if (*lri->lri_exp != NULL)
388 class_export_put(*lri->lri_exp);
393 int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
394 register_lwp_cb cb_func, void *cb_data)
396 struct obd_device *lwp;
397 struct lwp_register_item *lri;
401 LASSERTF(strlen(lwpname) < MTI_NAME_MAXLEN, "lwpname is too long %s\n",
403 LASSERT(exp != NULL && *exp == NULL);
409 lwp = class_name2obd(lwpname);
410 if (lwp != NULL && lwp->obd_set_up == 1) {
411 struct obd_uuid *uuid;
418 memcpy(uuid->uuid, lwpname, strlen(lwpname));
419 *exp = cfs_hash_lookup(lwp->obd_uuid_hash, uuid);
423 memcpy(lri->lri_name, lwpname, strlen(lwpname));
425 lri->lri_cb_func = cb_func;
426 lri->lri_cb_data = cb_data;
427 INIT_LIST_HEAD(&lri->lri_list);
429 * Initialize the lri_ref at 2, one will be released before
430 * current function returned via lustre_put_lwp_item(), the
431 * other will be released in lustre_deregister_lwp_item().
433 atomic_set(&lri->lri_ref, 2);
435 spin_lock(&lwp_register_list_lock);
436 list_add(&lri->lri_list, &lwp_register_list);
439 spin_unlock(&lwp_register_list_lock);
441 if (cb && cb_func != NULL)
443 lustre_put_lwp_item(lri);
447 EXPORT_SYMBOL(lustre_register_lwp_item);
449 void lustre_deregister_lwp_item(struct obd_export **exp)
451 struct lwp_register_item *lri;
452 bool removed = false;
455 spin_lock(&lwp_register_list_lock);
456 list_for_each_entry(lri, &lwp_register_list, lri_list) {
457 if (exp == lri->lri_exp) {
458 list_del_init(&lri->lri_list);
463 spin_unlock(&lwp_register_list_lock);
468 /* See lustre_notify_lwp_list(), in some extreme race conditions,
469 * the notify callback could be still on the fly, we need to wait
470 * for the callback done before moving on to free the data used
472 while (atomic_read(&lri->lri_ref) > 1) {
473 CDEBUG(D_MOUNT, "lri reference count %u, repeat: %d\n",
474 atomic_read(&lri->lri_ref), repeat);
476 set_current_state(TASK_INTERRUPTIBLE);
477 schedule_timeout(cfs_time_seconds(1));
479 lustre_put_lwp_item(lri);
481 EXPORT_SYMBOL(lustre_deregister_lwp_item);
483 struct obd_export *lustre_find_lwp_by_index(const char *dev, __u32 idx)
485 struct lustre_mount_info *lmi;
486 struct lustre_sb_info *lsi;
487 struct obd_device *lwp;
488 struct obd_export *exp = NULL;
493 lmi = server_get_mount(dev);
497 lsi = s2lsi(lmi->lmi_sb);
498 rc = server_name2fsname(lsi->lsi_svname, fsname, NULL);
500 CERROR("%s: failed to get fsname: rc = %d\n",
501 lsi->lsi_svname, rc);
505 snprintf(lwp_name, sizeof(lwp_name), "%s-MDT%04x", fsname, idx);
506 spin_lock(&lsi->lsi_lwp_lock);
507 list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
508 char *ptr = strstr(lwp->obd_name, lwp_name);
510 if (ptr != NULL && lwp->obd_lwp_export != NULL) {
511 exp = class_export_get(lwp->obd_lwp_export);
515 spin_unlock(&lsi->lsi_lwp_lock);
518 server_put_mount(dev, false);
522 EXPORT_SYMBOL(lustre_find_lwp_by_index);
524 void lustre_notify_lwp_list(struct obd_export *exp)
526 struct lwp_register_item *lri;
527 LASSERT(exp != NULL);
530 spin_lock(&lwp_register_list_lock);
531 list_for_each_entry(lri, &lwp_register_list, lri_list) {
532 if (strcmp(exp->exp_obd->obd_name, lri->lri_name))
534 if (*lri->lri_exp != NULL)
536 *lri->lri_exp = class_export_get(exp);
537 if (lri->lri_cb_func == NULL)
539 atomic_inc(&lri->lri_ref);
540 spin_unlock(&lwp_register_list_lock);
542 lri->lri_cb_func(lri->lri_cb_data);
543 lustre_put_lwp_item(lri);
545 /* Others may have changed the list after we unlock, we have
546 * to rescan the list from the beginning. Usually, the list
547 * 'lwp_register_list' is very short, and there is 'guard'
548 * lri::lri_exp that will prevent the callback to be done
549 * repeatedly. So rescanning the list has no problem. */
552 spin_unlock(&lwp_register_list_lock);
554 EXPORT_SYMBOL(lustre_notify_lwp_list);
556 static int lustre_lwp_connect(struct obd_device *lwp, bool is_mdt)
559 struct lu_context session_ctx;
560 struct obd_export *exp;
561 struct obd_uuid *uuid = NULL;
562 struct obd_connect_data *data = NULL;
566 /* log has been fully processed, let clients connect */
567 rc = lu_env_init(&env, lwp->obd_lu_dev->ld_type->ldt_ctx_tags);
571 lu_context_init(&session_ctx, LCT_SERVER_SESSION);
572 session_ctx.lc_thread = NULL;
573 lu_context_enter(&session_ctx);
574 env.le_ses = &session_ctx;
578 GOTO(out, rc = -ENOMEM);
580 data->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_INDEX;
581 data->ocd_version = LUSTRE_VERSION_CODE;
582 data->ocd_connect_flags |= OBD_CONNECT_FID | OBD_CONNECT_AT |
583 OBD_CONNECT_LRU_RESIZE | OBD_CONNECT_FULL20 |
584 OBD_CONNECT_LVB_TYPE | OBD_CONNECT_LIGHTWEIGHT |
585 OBD_CONNECT_LFSCK | OBD_CONNECT_BULK_MBITS;
588 data->ocd_connect_flags |= OBD_CONNECT_MDS_MDS;
592 GOTO(out, rc = -ENOMEM);
594 if (strlen(lwp->obd_name) > sizeof(uuid->uuid)) {
595 CERROR("%s: Too long lwp name %s, max_size is %d\n",
596 lwp->obd_name, lwp->obd_name, (int)sizeof(uuid->uuid));
597 GOTO(out, rc = -EINVAL);
600 /* Use lwp name as the uuid, so we find the export by lwp name later */
601 memcpy(uuid->uuid, lwp->obd_name, strlen(lwp->obd_name));
602 rc = obd_connect(&env, &exp, lwp, uuid, data, NULL);
604 CERROR("%s: connect failed: rc = %d\n", lwp->obd_name, rc);
606 if (unlikely(lwp->obd_lwp_export != NULL))
607 class_export_put(lwp->obd_lwp_export);
608 lwp->obd_lwp_export = class_export_get(exp);
620 lu_context_exit(&session_ctx);
621 lu_context_fini(&session_ctx);
627 * lwp is used by slaves (Non-MDT0 targets) to manage the connection to MDT0,
628 * or from the OSTx to MDTy.
630 static int lustre_lwp_setup(struct lustre_cfg *lcfg, struct lustre_sb_info *lsi,
633 struct obd_device *obd;
634 char *lwpname = NULL;
635 char *lwpuuid = NULL;
639 rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
642 CERROR("%s: Can't add uuid: rc =%d\n", lsi->lsi_svname, rc);
646 OBD_ALLOC(lwpname, MTI_NAME_MAXLEN);
648 GOTO(out, rc = -ENOMEM);
650 rc = tgt_name2lwp_name(lsi->lsi_svname, lwpname, MTI_NAME_MAXLEN, idx);
652 CERROR("%s: failed to generate lwp name: rc = %d\n",
653 lsi->lsi_svname, rc);
657 OBD_ALLOC(lwpuuid, MTI_NAME_MAXLEN);
659 GOTO(out, rc = -ENOMEM);
661 sprintf(lwpuuid, "%s_UUID", lwpname);
662 rc = lustre_start_simple(lwpname, LUSTRE_LWP_NAME,
663 lwpuuid, lustre_cfg_string(lcfg, 1),
666 CERROR("%s: setup up failed: rc %d\n", lwpname, rc);
670 obd = class_name2obd(lwpname);
671 LASSERT(obd != NULL);
673 rc = lustre_lwp_connect(obd, strstr(lsi->lsi_svname, "-MDT") != NULL);
675 obd->u.cli.cl_max_mds_easize = MAX_MD_SIZE;
676 spin_lock(&lsi->lsi_lwp_lock);
677 list_add_tail(&obd->obd_lwp_list, &lsi->lsi_lwp_list);
678 spin_unlock(&lsi->lsi_lwp_lock);
680 CERROR("%s: connect failed: rc = %d\n", lwpname, rc);
687 OBD_FREE(lwpname, MTI_NAME_MAXLEN);
689 OBD_FREE(lwpuuid, MTI_NAME_MAXLEN);
694 /* the caller is responsible for memory free */
695 static struct obd_device *lustre_find_lwp(struct lustre_sb_info *lsi,
696 char **lwpname, __u32 idx)
698 struct obd_device *lwp;
702 LASSERT(lwpname != NULL);
703 LASSERT(IS_OST(lsi) || IS_MDT(lsi));
705 OBD_ALLOC(*lwpname, MTI_NAME_MAXLEN);
706 if (*lwpname == NULL)
707 RETURN(ERR_PTR(-ENOMEM));
709 rc = tgt_name2lwp_name(lsi->lsi_svname, *lwpname, MTI_NAME_MAXLEN, idx);
711 CERROR("%s: failed to generate lwp name: rc = %d\n",
712 lsi->lsi_svname, rc);
713 GOTO(out, rc = -EINVAL);
716 lwp = class_name2obd(*lwpname);
720 if (*lwpname != NULL) {
721 OBD_FREE(*lwpname, MTI_NAME_MAXLEN);
727 RETURN(lwp != NULL ? lwp : ERR_PTR(-ENOENT));
730 static int lustre_lwp_add_conn(struct lustre_cfg *cfg,
731 struct lustre_sb_info *lsi, __u32 idx)
733 struct lustre_cfg_bufs *bufs = NULL;
734 struct lustre_cfg *lcfg = NULL;
735 char *lwpname = NULL;
736 struct obd_device *lwp;
740 lwp = lustre_find_lwp(lsi, &lwpname, idx);
742 CERROR("%s: can't find lwp device.\n", lsi->lsi_svname);
743 GOTO(out, rc = PTR_ERR(lwp));
745 LASSERT(lwpname != NULL);
749 GOTO(out, rc = -ENOMEM);
751 lustre_cfg_bufs_reset(bufs, lwpname);
752 lustre_cfg_bufs_set_string(bufs, 1,
753 lustre_cfg_string(cfg, 1));
755 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
757 GOTO(out_cfg, rc = -ENOMEM);
758 lustre_cfg_init(lcfg, LCFG_ADD_CONN, bufs);
760 rc = class_add_conn(lwp, lcfg);
762 CERROR("%s: can't add conn: rc = %d\n", lwpname, rc);
765 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
766 lcfg->lcfg_buflens));
772 OBD_FREE(lwpname, MTI_NAME_MAXLEN);
777 * Retrieve MDT nids from the client log, then start the lwp device.
778 * there are only two scenarios which would include mdt nid.
780 * marker 5 (flags=0x01, v2.1.54.0) lustre-MDTyyyy 'add mdc' xxx-
781 * add_uuid nid=192.168.122.162@tcp(0x20000c0a87aa2) 0: 1:192.168.122.162@tcp
782 * attach 0:lustre-MDTyyyy-mdc 1:mdc 2:lustre-clilmv_UUID
783 * setup 0:lustre-MDTyyyy-mdc 1:lustre-MDTyyyy_UUID 2:192.168.122.162@tcp
784 * add_uuid nid=192.168.172.1@tcp(0x20000c0a8ac01) 0: 1:192.168.172.1@tcp
785 * add_conn 0:lustre-MDTyyyy-mdc 1:192.168.172.1@tcp
786 * modify_mdc_tgts add 0:lustre-clilmv 1:lustre-MDTyyyy_UUID xxxx
787 * marker 5 (flags=0x02, v2.1.54.0) lustre-MDTyyyy 'add mdc' xxxx-
789 * marker 7 (flags=0x01, v2.1.54.0) lustre-MDTyyyy 'add failnid' xxxx-
790 * add_uuid nid=192.168.122.2@tcp(0x20000c0a87a02) 0: 1:192.168.122.2@tcp
791 * add_conn 0:lustre-MDTyyyy-mdc 1:192.168.122.2@tcp
792 * marker 7 (flags=0x02, v2.1.54.0) lustre-MDTyyyy 'add failnid' xxxx-
794 static int client_lwp_config_process(const struct lu_env *env,
795 struct llog_handle *handle,
796 struct llog_rec_hdr *rec, void *data)
798 struct config_llog_instance *cfg = data;
799 int cfg_len = rec->lrh_len;
800 char *cfg_buf = (char *) (rec + 1);
801 struct lustre_cfg *lcfg = NULL;
802 struct lustre_sb_info *lsi;
803 int rc = 0, swab = 0;
806 if (rec->lrh_type != OBD_CFG_REC) {
807 CERROR("Unknown llog record type %#x encountered\n",
812 if (cfg->cfg_sb == NULL)
813 GOTO(out, rc = -EINVAL);
814 lsi = s2lsi(cfg->cfg_sb);
816 lcfg = (struct lustre_cfg *)cfg_buf;
817 if (lcfg->lcfg_version == __swab32(LUSTRE_CFG_VERSION)) {
818 lustre_swab_lustre_cfg(lcfg);
822 rc = lustre_cfg_sanity_check(cfg_buf, cfg_len);
826 switch (lcfg->lcfg_command) {
828 struct cfg_marker *marker = lustre_cfg_buf(lcfg, 1);
830 lustre_swab_cfg_marker(marker, swab,
831 LUSTRE_CFG_BUFLEN(lcfg, 1));
832 if (marker->cm_flags & CM_SKIP ||
833 marker->cm_flags & CM_EXCLUDE)
836 if (!tgt_is_mdt(marker->cm_tgtname, &cfg->cfg_lwp_idx))
839 if (IS_MDT(lsi) && cfg->cfg_lwp_idx != 0)
842 if (!strncmp(marker->cm_comment, "add mdc", 7) ||
843 !strncmp(marker->cm_comment, "add failnid", 11)) {
844 if (marker->cm_flags & CM_START) {
845 cfg->cfg_flags = CFG_F_MARKER;
846 /* This hack is to differentiate the
847 * ADD_UUID is come from "add mdc" record
848 * or from "add failnid" record. */
849 if (!strncmp(marker->cm_comment,
851 cfg->cfg_flags |= CFG_F_SKIP;
852 } else if (marker->cm_flags & CM_END) {
858 case LCFG_ADD_UUID: {
859 if (cfg->cfg_flags == CFG_F_MARKER) {
860 rc = lustre_lwp_setup(lcfg, lsi, cfg->cfg_lwp_idx);
861 /* XXX: process only the first nid as
862 * we don't need another instance of lwp */
863 cfg->cfg_flags |= CFG_F_SKIP;
864 } else if (cfg->cfg_flags == (CFG_F_MARKER | CFG_F_SKIP)) {
865 rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
868 CERROR("%s: Fail to add uuid, rc:%d\n",
869 lsi->lsi_svname, rc);
873 case LCFG_ADD_CONN: {
874 char *devname = lustre_cfg_string(lcfg, 0);
878 if (!is_mdc_device(devname))
881 ptr = strrchr(devname, '-');
886 if (!tgt_is_mdt(devname, &idx)) {
892 if (IS_MDT(lsi) && idx != 0)
895 rc = lustre_lwp_add_conn(lcfg, lsi, idx);
905 static int lustre_disconnect_lwp(struct super_block *sb)
907 struct lustre_sb_info *lsi = s2lsi(sb);
908 struct obd_device *lwp;
909 char *logname = NULL;
910 struct lustre_cfg_bufs *bufs = NULL;
911 struct config_llog_instance *cfg = NULL;
916 if (likely(lsi->lsi_lwp_started)) {
917 OBD_ALLOC(logname, MTI_NAME_MAXLEN);
921 rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
923 CERROR("%s: failed to get fsname from svname: "
924 "rc = %d\n", lsi->lsi_svname, rc);
925 GOTO(out, rc = -EINVAL);
928 strcat(logname, "-client");
931 GOTO(out, rc = -ENOMEM);
934 cfg->cfg_instance = ll_get_cfg_instance(sb);
935 rc = lustre_end_log(sb, logname, cfg);
936 if (rc != 0 && rc != -ENOENT)
939 lsi->lsi_lwp_started = 0;
944 GOTO(out, rc = -ENOMEM);
946 list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
947 struct lustre_cfg *lcfg;
949 if (likely(lwp->obd_lwp_export != NULL)) {
950 class_export_put(lwp->obd_lwp_export);
951 lwp->obd_lwp_export = NULL;
954 lustre_cfg_bufs_reset(bufs, lwp->obd_name);
955 lustre_cfg_bufs_set_string(bufs, 1, NULL);
956 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount,
959 GOTO(out, rc = -ENOMEM);
960 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
962 /* Disconnect import first. NULL is passed for the '@env',
963 * since it will not be used. */
964 rc = lwp->obd_lu_dev->ld_ops->ldo_process_config(NULL,
965 lwp->obd_lu_dev, lcfg);
966 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
967 lcfg->lcfg_buflens));
968 if (rc != 0 && rc != -ETIMEDOUT) {
969 CERROR("%s: fail to disconnect LWP: rc = %d\n",
983 OBD_FREE(logname, MTI_NAME_MAXLEN);
985 return rc1 != 0 ? rc1 : rc;
989 * Stop the lwp for an OST/MDT target.
991 static int lustre_stop_lwp(struct super_block *sb)
993 struct lustre_sb_info *lsi = s2lsi(sb);
994 struct obd_device *lwp;
999 while (!list_empty(&lsi->lsi_lwp_list)) {
1000 lwp = list_entry(lsi->lsi_lwp_list.next, struct obd_device,
1002 list_del_init(&lwp->obd_lwp_list);
1004 rc = class_manual_cleanup(lwp);
1006 CERROR("%s: fail to stop LWP: rc = %d\n",
1012 RETURN(rc1 != 0 ? rc1 : rc);
1016 * Start the lwp(fsname-MDTyyyy-lwp-{MDT,OST}xxxx) for a MDT/OST or MDT target.
1018 static int lustre_start_lwp(struct super_block *sb)
1020 struct lustre_sb_info *lsi = s2lsi(sb);
1021 struct config_llog_instance *cfg = NULL;
1026 if (unlikely(lsi->lsi_lwp_started))
1029 OBD_ALLOC(logname, MTI_NAME_MAXLEN);
1030 if (logname == NULL)
1033 rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
1035 CERROR("%s: failed to get fsname from svname: rc = %d\n",
1036 lsi->lsi_svname, rc);
1037 GOTO(out, rc = -EINVAL);
1040 strcat(logname, "-client");
1043 GOTO(out, rc = -ENOMEM);
1045 cfg->cfg_callback = client_lwp_config_process;
1046 cfg->cfg_instance = ll_get_cfg_instance(sb);
1047 rc = lustre_process_log(sb, logname, cfg);
1048 /* need to remove config llog from mgc */
1049 lsi->lsi_lwp_started = 1;
1054 OBD_FREE(logname, MTI_NAME_MAXLEN);
1061 static DEFINE_MUTEX(server_start_lock);
1063 /* Stop MDS/OSS if nobody is using them */
1064 static int server_stop_servers(int lsiflags)
1066 struct obd_device *obd = NULL;
1067 struct obd_type *type = NULL;
1071 mutex_lock(&server_start_lock);
1073 /* Either an MDT or an OST or neither */
1074 /* if this was an MDT, and there are no more MDT's, clean up the MDS */
1075 if (lsiflags & LDD_F_SV_TYPE_MDT) {
1076 obd = class_name2obd(LUSTRE_MDS_OBDNAME);
1078 type = class_search_type(LUSTRE_MDT_NAME);
1081 /* if this was an OST, and there are no more OST's, clean up the OSS */
1082 if (lsiflags & LDD_F_SV_TYPE_OST) {
1083 obd = class_name2obd(LUSTRE_OSS_OBDNAME);
1085 type = class_search_type(LUSTRE_OST_NAME);
1088 if (obd != NULL && (type == NULL || type->typ_refcnt == 0)) {
1090 /* obd_fail doesn't mean much on a server obd */
1091 rc = class_manual_cleanup(obd);
1094 mutex_unlock(&server_start_lock);
1099 int server_mti_print(const char *title, struct mgs_target_info *mti)
1101 PRINT_CMD(PRINT_MASK, "mti %s\n", title);
1102 PRINT_CMD(PRINT_MASK, "server: %s\n", mti->mti_svname);
1103 PRINT_CMD(PRINT_MASK, "fs: %s\n", mti->mti_fsname);
1104 PRINT_CMD(PRINT_MASK, "uuid: %s\n", mti->mti_uuid);
1105 PRINT_CMD(PRINT_MASK, "ver: %d flags: %#x\n",
1106 mti->mti_config_ver, mti->mti_flags);
1110 /* Generate data for registration */
1111 static int server_lsi2mti(struct lustre_sb_info *lsi,
1112 struct mgs_target_info *mti)
1114 struct lnet_process_id id;
1119 if (!IS_SERVER(lsi))
1122 if (strlcpy(mti->mti_svname, lsi->lsi_svname, sizeof(mti->mti_svname))
1123 >= sizeof(mti->mti_svname))
1126 mti->mti_nid_count = 0;
1127 while (LNetGetId(i++, &id) != -ENOENT) {
1128 if (LNET_NETTYP(LNET_NIDNET(id.nid)) == LOLND)
1131 /* server use --servicenode param, only allow specified
1132 * nids be registered */
1133 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_NO_PRIMNODE) != 0 &&
1134 class_match_nid(lsi->lsi_lmd->lmd_params,
1135 PARAM_FAILNODE, id.nid) < 1)
1138 /* match specified network */
1139 if (!class_match_net(lsi->lsi_lmd->lmd_params,
1140 PARAM_NETWORK, LNET_NIDNET(id.nid)))
1143 mti->mti_nids[mti->mti_nid_count] = id.nid;
1144 mti->mti_nid_count++;
1145 if (mti->mti_nid_count >= MTI_NIDS_MAX) {
1146 CWARN("Only using first %d nids for %s\n",
1147 mti->mti_nid_count, mti->mti_svname);
1152 if (mti->mti_nid_count == 0) {
1153 CERROR("Failed to get NID for server %s, please check whether "
1154 "the target is specifed with improper --servicenode or "
1155 "--network options.\n", mti->mti_svname);
1159 mti->mti_lustre_ver = LUSTRE_VERSION_CODE;
1160 mti->mti_config_ver = 0;
1162 rc = server_name2fsname(lsi->lsi_svname, mti->mti_fsname, NULL);
1166 rc = server_name2index(lsi->lsi_svname, &mti->mti_stripe_index, NULL);
1169 /* Orion requires index to be set */
1170 LASSERT(!(rc & LDD_F_NEED_INDEX));
1171 /* keep only LDD flags */
1172 mti->mti_flags = lsi->lsi_flags & LDD_F_MASK;
1173 if (mti->mti_flags & (LDD_F_WRITECONF | LDD_F_VIRGIN))
1174 mti->mti_flags |= LDD_F_UPDATE;
1175 cplen = strlcpy(mti->mti_params, lsi->lsi_lmd->lmd_params,
1176 sizeof(mti->mti_params));
1177 if (cplen >= sizeof(mti->mti_params))
1182 /* Register an old or new target with the MGS. If needed MGS will construct
1183 startup logs and assign index */
1184 static int server_register_target(struct lustre_sb_info *lsi)
1186 struct obd_device *mgc = lsi->lsi_mgc;
1187 struct mgs_target_info *mti = NULL;
1195 if (!IS_SERVER(lsi))
1202 rc = server_lsi2mti(lsi, mti);
1206 CDEBUG(D_MOUNT, "Registration %s, fs=%s, %s, index=%04x, flags=%#x\n",
1207 mti->mti_svname, mti->mti_fsname,
1208 libcfs_nid2str(mti->mti_nids[0]), mti->mti_stripe_index,
1211 /* if write_conf is true, the registration must succeed */
1212 writeconf = !!(lsi->lsi_flags & (LDD_F_NEED_INDEX | LDD_F_UPDATE));
1213 mti->mti_flags |= LDD_F_OPC_REG;
1216 /* Register the target */
1217 /* FIXME use mgc_process_config instead */
1218 rc = obd_set_info_async(NULL, mgc->u.cli.cl_mgc_mgsexp,
1219 sizeof(KEY_REGISTER_TARGET),
1220 KEY_REGISTER_TARGET,
1221 sizeof(*mti), mti, NULL);
1223 if (mti->mti_flags & LDD_F_ERROR) {
1224 LCONSOLE_ERROR_MSG(0x160,
1225 "%s: the MGS refuses to allow this server "
1226 "to start: rc = %d. Please see messages on "
1227 "the MGS.\n", lsi->lsi_svname, rc);
1228 } else if (writeconf) {
1229 if ((rc == -ESHUTDOWN || rc == -EIO) && ++tried < 5) {
1230 /* The connection with MGS is not established.
1231 * Try again after 2 seconds. Interruptable. */
1232 set_current_state(TASK_INTERRUPTIBLE);
1233 schedule_timeout(cfs_time_seconds(2));
1234 set_current_state(TASK_RUNNING);
1235 if (!signal_pending(current))
1239 LCONSOLE_ERROR_MSG(0x15f,
1240 "%s: cannot register this server with the MGS: "
1241 "rc = %d. Is the MGS running?\n",
1242 lsi->lsi_svname, rc);
1244 CDEBUG(D_HA, "%s: error registering with the MGS: "
1245 "rc = %d (not fatal)\n", lsi->lsi_svname, rc);
1246 /* reset the error code for non-fatal error. */
1259 * Notify the MGS that this target is ready.
1260 * Used by IR - if the MGS receives this message, it will notify clients.
1262 static int server_notify_target(struct super_block *sb, struct obd_device *obd)
1264 struct lustre_sb_info *lsi = s2lsi(sb);
1265 struct obd_device *mgc = lsi->lsi_mgc;
1266 struct mgs_target_info *mti = NULL;
1272 if (!(IS_SERVER(lsi)))
1278 rc = server_lsi2mti(lsi, mti);
1282 mti->mti_instance = obd->u.obt.obt_instance;
1283 mti->mti_flags |= LDD_F_OPC_READY;
1285 /* FIXME use mgc_process_config instead */
1286 rc = obd_set_info_async(NULL, mgc->u.cli.cl_mgc_mgsexp,
1287 sizeof(KEY_REGISTER_TARGET),
1288 KEY_REGISTER_TARGET,
1289 sizeof(*mti), mti, NULL);
1291 /* Imperative recovery: if the mgs informs us to use IR? */
1292 if (!rc && !(mti->mti_flags & LDD_F_ERROR) &&
1293 (mti->mti_flags & LDD_F_IR_CAPABLE))
1294 lsi->lsi_flags |= LDD_F_IR_CAPABLE;
1303 /** Start server targets: MDTs and OSTs
1305 static int server_start_targets(struct super_block *sb)
1307 struct obd_device *obd;
1308 struct lustre_sb_info *lsi = s2lsi(sb);
1309 struct config_llog_instance cfg;
1310 struct lu_env mgc_env;
1311 struct lu_device *dev;
1315 CDEBUG(D_MOUNT, "starting target %s\n", lsi->lsi_svname);
1318 /* make sure the MDS is started */
1319 mutex_lock(&server_start_lock);
1320 obd = class_name2obd(LUSTRE_MDS_OBDNAME);
1322 rc = lustre_start_simple(LUSTRE_MDS_OBDNAME,
1324 LUSTRE_MDS_OBDNAME"_uuid",
1325 NULL, NULL, NULL, NULL);
1327 mutex_unlock(&server_start_lock);
1328 CERROR("failed to start MDS: %d\n", rc);
1332 mutex_unlock(&server_start_lock);
1335 /* If we're an OST, make sure the global OSS is running */
1337 /* make sure OSS is started */
1338 mutex_lock(&server_start_lock);
1339 obd = class_name2obd(LUSTRE_OSS_OBDNAME);
1341 rc = lustre_start_simple(LUSTRE_OSS_OBDNAME,
1343 LUSTRE_OSS_OBDNAME"_uuid",
1344 NULL, NULL, NULL, NULL);
1346 mutex_unlock(&server_start_lock);
1347 CERROR("failed to start OSS: %d\n", rc);
1351 mutex_unlock(&server_start_lock);
1354 rc = lu_env_init(&mgc_env, LCT_MG_THREAD);
1356 GOTO(out_stop_service, rc);
1358 /* Set the mgc fs to our server disk. This allows the MGC to
1359 * read and write configs locally, in case it can't talk to the MGS. */
1360 rc = server_mgc_set_fs(&mgc_env, lsi->lsi_mgc, sb);
1364 /* Register with MGS */
1365 rc = server_register_target(lsi);
1369 /* Let the target look up the mount using the target's name
1370 (we can't pass the sb or mnt through class_process_config.) */
1371 rc = server_register_mount(lsi->lsi_svname, sb);
1375 /* Start targets using the llog named for the target */
1376 memset(&cfg, 0, sizeof(cfg));
1377 cfg.cfg_callback = class_config_llog_handler;
1378 cfg.cfg_sub_clds = CONFIG_SUB_SERVER;
1379 rc = lustre_process_log(sb, lsi->lsi_svname, &cfg);
1381 CERROR("failed to start server %s: %d\n",
1382 lsi->lsi_svname, rc);
1383 /* Do NOT call server_deregister_mount() here. This makes it
1384 * impossible to find mount later in cleanup time and leaves
1385 * @lsi and othder stuff leaked. -umka */
1389 obd = class_name2obd(lsi->lsi_svname);
1391 CERROR("no server named %s was started\n", lsi->lsi_svname);
1392 GOTO(out_mgc, rc = -ENXIO);
1395 if (IS_OST(lsi) || IS_MDT(lsi)) {
1396 rc = lustre_start_lwp(sb);
1398 CERROR("%s: failed to start LWP: %d\n",
1399 lsi->lsi_svname, rc);
1404 server_notify_target(sb, obd);
1406 /* calculate recovery timeout, do it after lustre_process_log */
1407 server_calc_timeout(lsi, obd);
1409 /* log has been fully processed, let clients connect */
1410 dev = obd->obd_lu_dev;
1411 if (dev && dev->ld_ops->ldo_prepare) {
1414 rc = lu_env_init(&env, dev->ld_type->ldt_ctx_tags);
1416 struct lu_context session_ctx;
1418 lu_context_init(&session_ctx, LCT_SERVER_SESSION);
1419 session_ctx.lc_thread = NULL;
1420 lu_context_enter(&session_ctx);
1421 env.le_ses = &session_ctx;
1423 rc = dev->ld_ops->ldo_prepare(&env, NULL, dev);
1426 lu_context_exit(&session_ctx);
1427 lu_context_fini(&session_ctx);
1431 /* abort recovery only on the complete stack:
1432 * many devices can be involved */
1433 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_ABORT_RECOV) &&
1434 (OBP(obd, iocontrol))) {
1435 obd_iocontrol(OBD_IOC_ABORT_RECOVERY, obd->obd_self_export, 0,
1440 /* Release the mgc fs for others to use */
1441 server_mgc_clear_fs(&mgc_env, lsi->lsi_mgc);
1443 lu_env_fini(&mgc_env);
1446 server_stop_servers(lsi->lsi_flags);
1451 static int lsi_prepare(struct lustre_sb_info *lsi)
1453 const char *osd_type;
1460 LASSERT(lsi->lsi_lmd);
1462 /* The server name is given as a mount line option */
1463 if (lsi->lsi_lmd->lmd_profile == NULL) {
1464 LCONSOLE_ERROR("Can't determine server name\n");
1468 /* Determine osd type */
1469 if (lsi->lsi_lmd->lmd_osd_type == NULL) {
1470 osd_type = LUSTRE_OSD_LDISKFS_NAME;
1473 osd_type = lsi->lsi_lmd->lmd_osd_type;
1474 fstype = lsi->lsi_lmd->lmd_osd_type;
1477 if (strlen(lsi->lsi_lmd->lmd_profile) >= sizeof(lsi->lsi_svname) ||
1478 strlen(osd_type) >= sizeof(lsi->lsi_osd_type) ||
1479 strlen(fstype) >= sizeof(lsi->lsi_fstype))
1480 RETURN(-ENAMETOOLONG);
1482 strlcpy(lsi->lsi_svname, lsi->lsi_lmd->lmd_profile,
1483 sizeof(lsi->lsi_svname));
1484 strlcpy(lsi->lsi_osd_type, osd_type, sizeof(lsi->lsi_osd_type));
1485 /* XXX: a temp. solution for components using ldiskfs
1486 * to be removed in one of the subsequent patches */
1487 strlcpy(lsi->lsi_fstype, fstype, sizeof(lsi->lsi_fstype));
1489 /* Determine server type */
1490 rc = server_name2index(lsi->lsi_svname, &index, NULL);
1492 if (lsi->lsi_lmd->lmd_flags & LMD_FLG_MGS) {
1493 /* Assume we're a bare MGS */
1495 lsi->lsi_lmd->lmd_flags |= LMD_FLG_NOSVC;
1497 LCONSOLE_ERROR("Can't determine server type of '%s'\n",
1502 lsi->lsi_flags |= rc;
1504 /* Add mount line flags that used to be in ldd:
1505 * writeconf, mgs, anything else?
1507 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_WRITECONF) ?
1508 LDD_F_WRITECONF : 0;
1509 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_VIRGIN) ?
1511 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_UPDATE) ?
1513 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_MGS) ?
1514 LDD_F_SV_TYPE_MGS : 0;
1515 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_NO_PRIMNODE) ?
1516 LDD_F_NO_PRIMNODE : 0;
1521 /*************** server mount ******************/
1523 /** Start the shutdown of servers at umount.
1525 static void server_put_super(struct super_block *sb)
1527 struct lustre_sb_info *lsi = s2lsi(sb);
1528 struct obd_device *obd;
1529 char *tmpname, *extraname = NULL;
1531 int lsiflags = lsi->lsi_flags;
1534 LASSERT(IS_SERVER(lsi));
1536 tmpname_sz = strlen(lsi->lsi_svname) + 1;
1537 OBD_ALLOC(tmpname, tmpname_sz);
1538 memcpy(tmpname, lsi->lsi_svname, tmpname_sz);
1539 CDEBUG(D_MOUNT, "server put_super %s\n", tmpname);
1540 if (IS_MDT(lsi) && (lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC))
1541 snprintf(tmpname, tmpname_sz, "MGS");
1543 /* disconnect the lwp first to drain off the inflight request */
1544 if (IS_OST(lsi) || IS_MDT(lsi)) {
1547 rc = lustre_disconnect_lwp(sb);
1548 if (rc != 0 && rc != -ETIMEDOUT &&
1549 rc != -ENOTCONN && rc != -ESHUTDOWN)
1550 CWARN("%s: failed to disconnect lwp: rc= %d\n",
1554 /* Stop the target */
1555 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1556 (IS_MDT(lsi) || IS_OST(lsi))) {
1557 struct lustre_profile *lprof = NULL;
1559 /* tell the mgc to drop the config log */
1560 lustre_end_log(sb, lsi->lsi_svname, NULL);
1562 /* COMPAT_146 - profile may get deleted in mgc_cleanup.
1563 If there are any setup/cleanup errors, save the lov
1564 name for safety cleanup later. */
1565 lprof = class_get_profile(lsi->lsi_svname);
1566 if (lprof != NULL) {
1567 if (lprof->lp_dt != NULL) {
1568 OBD_ALLOC(extraname, strlen(lprof->lp_dt) + 1);
1569 strncpy(extraname, lprof->lp_dt,
1570 strlen(lprof->lp_dt) + 1);
1572 class_put_profile(lprof);
1575 obd = class_name2obd(lsi->lsi_svname);
1577 CDEBUG(D_MOUNT, "stopping %s\n", obd->obd_name);
1578 if (lsiflags & LSI_UMOUNT_FAILOVER)
1580 /* We can't seem to give an error return code
1581 * to .put_super, so we better make sure we clean up! */
1583 class_manual_cleanup(obd);
1585 CERROR("no obd %s\n", lsi->lsi_svname);
1586 server_deregister_mount(lsi->lsi_svname);
1590 /* If they wanted the mgs to stop separately from the mdt, they
1591 should have put it on a different device. */
1593 /* if MDS start with --nomgs, don't stop MGS then */
1594 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS))
1595 server_stop_mgs(sb);
1598 if (IS_OST(lsi) || IS_MDT(lsi)) {
1599 if (lustre_stop_lwp(sb) < 0)
1600 CERROR("%s: failed to stop lwp!\n", tmpname);
1603 /* Clean the mgc and sb */
1604 lustre_common_put_super(sb);
1606 /* wait till all in-progress cleanups are done
1607 * specifically we're interested in ofd cleanup
1609 obd_zombie_barrier();
1611 /* Stop the servers (MDS, OSS) if no longer needed. We must wait
1612 until the target is really gone so that our type refcount check
1614 server_stop_servers(lsiflags);
1616 /* In case of startup or cleanup err, stop related obds */
1618 obd = class_name2obd(extraname);
1620 CWARN("Cleaning orphaned obd %s\n", extraname);
1622 class_manual_cleanup(obd);
1624 OBD_FREE(extraname, strlen(extraname) + 1);
1627 LCONSOLE_WARN("server umount %s complete\n", tmpname);
1628 OBD_FREE(tmpname, tmpname_sz);
1632 /** Called only for 'umount -f'
1634 static void server_umount_begin(struct super_block *sb)
1636 struct lustre_sb_info *lsi = s2lsi(sb);
1639 CDEBUG(D_MOUNT, "umount -f\n");
1640 /* umount = failover
1642 no third way to do non-force, non-failover */
1643 lsi->lsi_flags &= ~LSI_UMOUNT_FAILOVER;
1647 static int server_statfs(struct dentry *dentry, struct kstatfs *buf)
1649 struct super_block *sb = dentry->d_sb;
1650 struct lustre_sb_info *lsi = s2lsi(sb);
1651 struct obd_statfs statfs;
1655 if (lsi->lsi_dt_dev) {
1656 rc = dt_statfs(NULL, lsi->lsi_dt_dev, &statfs);
1658 statfs_unpack(buf, &statfs);
1659 buf->f_type = sb->s_magic;
1665 buf->f_type = sb->s_magic;
1666 buf->f_bsize = sb->s_blocksize;
1672 buf->f_namelen = NAME_MAX;
1676 #ifdef HAVE_SUPEROPS_USE_DENTRY
1677 int server_show_options(struct seq_file *seq, struct dentry *dentry)
1679 int server_show_options(struct seq_file *seq, struct vfsmount *vfs)
1682 struct lustre_sb_info *lsi;
1683 struct lustre_mount_data *lmd;
1685 #ifdef HAVE_SUPEROPS_USE_DENTRY
1686 LASSERT(seq != NULL && dentry != NULL);
1687 lsi = s2lsi(dentry->d_sb);
1689 LASSERT(seq != NULL && vfs != NULL);
1690 lsi = s2lsi(vfs->mnt_sb);
1694 seq_printf(seq, ",svname=%s", lmd->lmd_profile);
1696 if (lmd->lmd_flags & LMD_FLG_ABORT_RECOV)
1697 seq_puts(seq, ",abort_recov");
1699 if (lmd->lmd_flags & LMD_FLG_NOIR)
1700 seq_puts(seq, ",noir");
1702 if (lmd->lmd_flags & LMD_FLG_NOSVC)
1703 seq_puts(seq, ",nosvc");
1705 if (lmd->lmd_flags & LMD_FLG_NOMGS)
1706 seq_puts(seq, ",nomgs");
1708 if (lmd->lmd_flags & LMD_FLG_NOSCRUB)
1709 seq_puts(seq, ",noscrub");
1710 if (lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
1711 seq_puts(seq, ",skip_lfsck");
1713 if (lmd->lmd_flags & LMD_FLG_DEV_RDONLY)
1714 seq_puts(seq, ",rdonly_dev");
1716 if (lmd->lmd_flags & LMD_FLG_MGS)
1717 seq_puts(seq, ",mgs");
1719 if (lmd->lmd_mgs != NULL)
1720 seq_printf(seq, ",mgsnode=%s", lmd->lmd_mgs);
1722 if (lmd->lmd_osd_type != NULL)
1723 seq_printf(seq, ",osd=%s", lmd->lmd_osd_type);
1725 if (lmd->lmd_opts != NULL) {
1727 seq_puts(seq, lmd->lmd_opts);
1733 /** The operations we support directly on the superblock:
1734 * mount, umount, and df.
1736 static struct super_operations server_ops = {
1737 .put_super = server_put_super,
1738 .umount_begin = server_umount_begin, /* umount -f */
1739 .statfs = server_statfs,
1740 .show_options = server_show_options,
1744 * Xattr support for Lustre servers
1746 #ifdef HAVE_IOP_XATTR
1747 static ssize_t lustre_getxattr(struct dentry *dentry, const char *name,
1748 void *buffer, size_t size)
1750 if (!selinux_is_enabled())
1755 static int lustre_setxattr(struct dentry *dentry, const char *name,
1756 const void *value, size_t size, int flags)
1762 static ssize_t lustre_listxattr(struct dentry *d_entry, char *name,
1768 static bool is_cmd_supported(unsigned int command)
1780 static long server_ioctl(struct file *filp, unsigned int command,
1783 struct file active_filp;
1784 struct inode *inode = file_inode(filp);
1785 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
1786 struct super_block *dd_sb = dt_mnt_sb_get(lsi->lsi_dt_dev);
1787 struct inode *active_inode;
1788 int err = -EOPNOTSUPP;
1790 if (IS_ERR(dd_sb) || !is_cmd_supported(command))
1793 active_inode = igrab(dd_sb->s_root->d_inode);
1797 active_filp.f_inode = active_inode;
1798 if (active_inode->i_fop && active_inode->i_fop->unlocked_ioctl)
1799 err = active_inode->i_fop->unlocked_ioctl(&active_filp,
1805 static const struct inode_operations server_inode_operations = {
1806 #ifdef HAVE_IOP_XATTR
1807 .setxattr = lustre_setxattr,
1808 .getxattr = lustre_getxattr,
1810 .listxattr = lustre_listxattr,
1813 static const struct file_operations server_file_operations = {
1814 .unlocked_ioctl = server_ioctl,
1817 #define log2(n) ffz(~(n))
1818 #define LUSTRE_SUPER_MAGIC 0x0BD00BD1
1820 static int server_fill_super_common(struct super_block *sb)
1822 struct inode *root = NULL;
1825 CDEBUG(D_MOUNT, "Server sb, dev=%d\n", (int)sb->s_dev);
1827 sb->s_blocksize = 4096;
1828 sb->s_blocksize_bits = log2(sb->s_blocksize);
1829 sb->s_magic = LUSTRE_SUPER_MAGIC;
1830 sb->s_maxbytes = 0; /* we don't allow file IO on server mountpoints */
1831 sb->s_flags |= SB_RDONLY;
1832 sb->s_op = &server_ops;
1834 root = new_inode(sb);
1836 CERROR("Can't make root inode\n");
1840 /* returns -EIO for every operation */
1841 /* make_bad_inode(root); -- badness - can't umount */
1842 /* apparently we need to be a directory for the mount to finish */
1843 root->i_mode = S_IFDIR;
1844 root->i_op = &server_inode_operations;
1845 root->i_fop = &server_file_operations;
1846 sb->s_root = d_make_root(root);
1848 CERROR("%s: can't make root dentry\n", sb->s_id);
1855 static int osd_start(struct lustre_sb_info *lsi, unsigned long mflags)
1857 struct lustre_mount_data *lmd = lsi->lsi_lmd;
1858 struct obd_device *obd;
1859 struct dt_device_param p;
1860 char flagstr[20 + 1 + 10 + 1];
1865 "Attempting to start %s, type=%s, lsifl=%x, mountfl=%lx\n",
1866 lsi->lsi_svname, lsi->lsi_osd_type, lsi->lsi_flags, mflags);
1868 sprintf(lsi->lsi_osd_obdname, "%s-osd", lsi->lsi_svname);
1869 strcpy(lsi->lsi_osd_uuid, lsi->lsi_osd_obdname);
1870 strcat(lsi->lsi_osd_uuid, "_UUID");
1871 snprintf(flagstr, sizeof(flagstr), "%lu:%u", mflags, lmd->lmd_flags);
1873 obd = class_name2obd(lsi->lsi_osd_obdname);
1875 rc = lustre_start_simple(lsi->lsi_osd_obdname,
1877 lsi->lsi_osd_uuid, lmd->lmd_dev,
1878 flagstr, lsi->lsi_lmd->lmd_opts,
1882 obd = class_name2obd(lsi->lsi_osd_obdname);
1885 CDEBUG(D_MOUNT, "%s already started\n", lsi->lsi_osd_obdname);
1886 /* but continue setup to allow special case of MDT and internal
1887 * MGT being started separately. */
1888 if (!((IS_MGS(lsi) && (lsi->lsi_lmd->lmd_flags &
1890 (IS_MDT(lsi) && (lsi->lsi_lmd->lmd_flags &
1895 rc = obd_connect(NULL, &lsi->lsi_osd_exp,
1896 obd, &obd->obd_uuid, NULL, NULL);
1900 class_manual_cleanup(obd);
1901 lsi->lsi_dt_dev = NULL;
1905 LASSERT(obd->obd_lu_dev);
1906 lu_device_get(obd->obd_lu_dev);
1907 lsi->lsi_dt_dev = lu2dt_dev(obd->obd_lu_dev);
1908 LASSERT(lsi->lsi_dt_dev);
1910 /* set disk context for llog usage */
1911 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1912 obd->obd_lvfs_ctxt.dt = lsi->lsi_dt_dev;
1914 dt_conf_get(NULL, lsi->lsi_dt_dev, &p);
1919 /** Fill in the superblock info for a Lustre server.
1920 * Mount the device with the correct options.
1921 * Read the on-disk config file.
1922 * Start the services.
1924 int server_fill_super(struct super_block *sb)
1926 struct lustre_sb_info *lsi = s2lsi(sb);
1930 /* to simulate target mount race */
1931 OBD_RACE(OBD_FAIL_TGT_MOUNT_RACE);
1933 rc = lsi_prepare(lsi);
1939 /* Start low level OSD */
1940 rc = osd_start(lsi, sb->s_flags);
1942 CERROR("Unable to start osd on %s: %d\n",
1943 lsi->lsi_lmd->lmd_dev, rc);
1948 CDEBUG(D_MOUNT, "Found service %s on device %s\n",
1949 lsi->lsi_svname, lsi->lsi_lmd->lmd_dev);
1951 if (class_name2obd(lsi->lsi_svname)) {
1952 LCONSOLE_ERROR_MSG(0x161, "The target named %s is already "
1953 "running. Double-mount may have compromised"
1954 " the disk journal.\n",
1960 /* Start MGS before MGC */
1961 if (IS_MGS(lsi) && !(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS)) {
1962 rc = server_start_mgs(sb);
1967 /* Start MGC before servers */
1968 rc = lustre_start_mgc(sb);
1972 /* Set up all obd devices for service */
1973 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1974 (IS_OST(lsi) || IS_MDT(lsi))) {
1975 rc = server_start_targets(sb);
1977 CERROR("Unable to start targets: %d\n", rc);
1980 /* FIXME overmount client here, or can we just start a
1981 * client log and client_fill_super on this sb? We
1982 * need to make sure server_put_super gets called too
1983 * - ll_put_super calls lustre_common_put_super; check
1984 * there for LSI_SERVER flag, call s_p_s if so.
1986 * Probably should start client from new thread so we
1987 * can return. Client will not finish until all
1988 * servers are connected. Note - MGS-only server does
1989 * NOT get a client, since there is no lustre fs
1990 * associated - the MGS is for all lustre fs's */
1993 rc = server_fill_super_common(sb);
1999 /* We jump here in case of failure while starting targets or MGS.
2000 * In this case we can't just put @mnt and have to do real cleanup
2001 * with stoping targets, etc. */
2002 server_put_super(sb);
2007 * Calculate timeout value for a target.
2009 void server_calc_timeout(struct lustre_sb_info *lsi, struct obd_device *obd)
2011 struct lustre_mount_data *lmd;
2015 bool has_ir = !!(lsi->lsi_flags & LDD_F_IR_CAPABLE);
2016 int min = OBD_RECOVERY_TIME_MIN;
2018 LASSERT(IS_SERVER(lsi));
2022 soft = lmd->lmd_recovery_time_soft;
2023 hard = lmd->lmd_recovery_time_hard;
2024 has_ir = has_ir && !(lmd->lmd_flags & LMD_FLG_NOIR);
2025 obd->obd_no_ir = !has_ir;
2029 soft = OBD_RECOVERY_TIME_SOFT;
2031 hard = OBD_RECOVERY_TIME_HARD;
2033 /* target may have ir_factor configured. */
2034 factor = OBD_IR_FACTOR_DEFAULT;
2035 if (obd->obd_recovery_ir_factor)
2036 factor = obd->obd_recovery_ir_factor;
2039 int new_soft = soft;
2041 /* adjust timeout value by imperative recovery */
2042 new_soft = (soft * factor) / OBD_IR_FACTOR_MAX;
2043 /* make sure the timeout is not too short */
2044 new_soft = max(min, new_soft);
2046 LCONSOLE_INFO("%s: Imperative Recovery enabled, recovery "
2047 "window shrunk from %d-%d down to %d-%d\n",
2048 obd->obd_name, soft, hard, new_soft, hard);
2052 LCONSOLE_INFO("%s: Imperative Recovery not enabled, recovery "
2053 "window %d-%d\n", obd->obd_name, soft, hard);
2057 obd->obd_recovery_timeout = max_t(time64_t, obd->obd_recovery_timeout,
2059 obd->obd_recovery_time_hard = hard;
2060 obd->obd_recovery_ir_factor = factor;