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>
51 #ifdef HAVE_KERNEL_LOCKED
52 #include <linux/smp_lock.h>
55 #include <llog_swab.h>
56 #include <lustre_disk.h>
57 #include <uapi/linux/lustre/lustre_ioctl.h>
58 #include <lustre_log.h>
59 #include <uapi/linux/lustre/lustre_param.h>
61 #include <obd_class.h>
63 /*********** mount lookup *********/
65 static DEFINE_MUTEX(lustre_mount_info_lock);
66 static LIST_HEAD(server_mount_info_list);
68 static struct lustre_mount_info *server_find_mount(const char *name)
70 struct list_head *tmp;
71 struct lustre_mount_info *lmi;
74 list_for_each(tmp, &server_mount_info_list) {
75 lmi = list_entry(tmp, struct lustre_mount_info,
77 if (strcmp(name, lmi->lmi_name) == 0)
83 /* we must register an obd for a mount before we call the setup routine.
84 *_setup will call lustre_get_mount to get the mnt struct
85 by obd_name, since we can't pass the pointer to setup. */
86 static int server_register_mount(const char *name, struct super_block *sb)
88 struct lustre_mount_info *lmi;
94 OBD_ALLOC(lmi, sizeof(*lmi));
97 OBD_ALLOC(name_cp, strlen(name) + 1);
99 OBD_FREE(lmi, sizeof(*lmi));
102 strcpy(name_cp, name);
104 mutex_lock(&lustre_mount_info_lock);
106 if (server_find_mount(name)) {
107 mutex_unlock(&lustre_mount_info_lock);
108 OBD_FREE(lmi, sizeof(*lmi));
109 OBD_FREE(name_cp, strlen(name) + 1);
110 CERROR("Already registered %s\n", name);
113 lmi->lmi_name = name_cp;
115 list_add(&lmi->lmi_list_chain, &server_mount_info_list);
117 mutex_unlock(&lustre_mount_info_lock);
119 CDEBUG(D_MOUNT, "register mount %p from %s\n", sb, name);
124 /* when an obd no longer needs a mount */
125 static int server_deregister_mount(const char *name)
127 struct lustre_mount_info *lmi;
130 mutex_lock(&lustre_mount_info_lock);
131 lmi = server_find_mount(name);
133 mutex_unlock(&lustre_mount_info_lock);
134 CERROR("%s not registered\n", name);
138 CDEBUG(D_MOUNT, "deregister mount %p from %s\n", lmi->lmi_sb, name);
140 OBD_FREE(lmi->lmi_name, strlen(lmi->lmi_name) + 1);
141 list_del(&lmi->lmi_list_chain);
142 OBD_FREE(lmi, sizeof(*lmi));
143 mutex_unlock(&lustre_mount_info_lock);
148 /* obd's look up a registered mount using their obdname. This is just
149 for initial obd setup to find the mount struct. It should not be
150 called every time you want to mntget. */
151 struct lustre_mount_info *server_get_mount(const char *name)
153 struct lustre_mount_info *lmi;
154 struct lustre_sb_info *lsi;
157 mutex_lock(&lustre_mount_info_lock);
158 lmi = server_find_mount(name);
159 mutex_unlock(&lustre_mount_info_lock);
161 CERROR("Can't find mount for %s\n", name);
164 lsi = s2lsi(lmi->lmi_sb);
166 atomic_inc(&lsi->lsi_mounts);
168 CDEBUG(D_MOUNT, "get mount %p from %s, refs=%d\n", lmi->lmi_sb,
169 name, atomic_read(&lsi->lsi_mounts));
173 EXPORT_SYMBOL(server_get_mount);
176 * server_put_mount: to be called from obd_cleanup methods
178 * @dereg_mnt: 0 or 1 depending on whether the mount is to be deregistered or
181 * The caller decides whether server_deregister_mount() needs to be called or
182 * not. Calling of server_deregister_mount() does not depend on refcounting on
183 * lsi because we could have say the mgs and mds on the same node and we
184 * unmount the mds, then the ref on the lsi would still be non-zero but we
185 * would still want to deregister the mds mount.
187 int server_put_mount(const char *name, bool dereg_mnt)
189 struct lustre_mount_info *lmi;
190 struct lustre_sb_info *lsi;
193 mutex_lock(&lustre_mount_info_lock);
194 lmi = server_find_mount(name);
195 mutex_unlock(&lustre_mount_info_lock);
197 CERROR("Can't find mount for %s\n", name);
200 lsi = s2lsi(lmi->lmi_sb);
202 CDEBUG(D_MOUNT, "put mount %p from %s, refs=%d\n",
203 lmi->lmi_sb, name, atomic_read(&lsi->lsi_mounts));
205 if (lustre_put_lsi(lmi->lmi_sb))
206 CDEBUG(D_MOUNT, "Last put of mount %p from %s\n",
210 /* this obd should never need the mount again */
211 server_deregister_mount(name);
215 EXPORT_SYMBOL(server_put_mount);
217 /* Set up a MGS to serve startup logs */
218 static int server_start_mgs(struct super_block *sb)
220 struct lustre_sb_info *lsi = s2lsi(sb);
221 struct lustre_mount_info *lmi;
225 /* It is impossible to have more than 1 MGS per node, since
226 MGC wouldn't know which to connect to */
227 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
229 lsi = s2lsi(lmi->lmi_sb);
230 LCONSOLE_ERROR_MSG(0x15d, "The MGS service was already started"
235 CDEBUG(D_CONFIG, "Start MGS service %s\n", LUSTRE_MGS_OBDNAME);
237 rc = server_register_mount(LUSTRE_MGS_OBDNAME, sb);
240 rc = lustre_start_simple(LUSTRE_MGS_OBDNAME, LUSTRE_MGS_NAME,
241 LUSTRE_MGS_OBDNAME, NULL, NULL,
242 lsi->lsi_osd_obdname, NULL);
243 /* server_deregister_mount() is not called previously, for lsi
244 * and other stuff can't be freed cleanly when mgs calls
245 * server_put_mount() in error handling case (see b=17758),
246 * this problem is caused by a bug in mgs_init0, which forgot
247 * calling server_put_mount in error case. */
250 server_deregister_mount(LUSTRE_MGS_OBDNAME);
254 LCONSOLE_ERROR_MSG(0x15e, "Failed to start MGS '%s' (%d). "
255 "Is the 'mgs' module loaded?\n",
256 LUSTRE_MGS_OBDNAME, rc);
260 static int server_stop_mgs(struct super_block *sb)
262 struct obd_device *obd;
264 struct lustre_mount_info *lmi;
267 /* Do not stop MGS if this device is not the running MGT */
268 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
269 if (lmi != NULL && lmi->lmi_sb != sb)
272 CDEBUG(D_MOUNT, "Stop MGS service %s\n", LUSTRE_MGS_OBDNAME);
274 /* There better be only one MGS */
275 obd = class_name2obd(LUSTRE_MGS_OBDNAME);
277 CDEBUG(D_CONFIG, "mgs %s not running\n", LUSTRE_MGS_OBDNAME);
281 /* The MGS should always stop when we say so */
283 rc = class_manual_cleanup(obd);
287 /* Since there's only one mgc per node, we have to change it's fs to get
288 access to the right disk. */
289 static int server_mgc_set_fs(const struct lu_env *env,
290 struct obd_device *mgc, struct super_block *sb)
292 struct lustre_sb_info *lsi = s2lsi(sb);
296 CDEBUG(D_MOUNT, "Set mgc disk for %s\n", lsi->lsi_lmd->lmd_dev);
298 /* cl_mgc_sem in mgc insures we sleep if the mgc_fs is busy */
299 rc = obd_set_info_async(env, mgc->obd_self_export,
300 sizeof(KEY_SET_FS), KEY_SET_FS,
301 sizeof(*sb), sb, NULL);
303 CERROR("can't set_fs %d\n", rc);
308 static int server_mgc_clear_fs(const struct lu_env *env,
309 struct obd_device *mgc)
314 CDEBUG(D_MOUNT, "Unassign mgc disk\n");
316 rc = obd_set_info_async(env, mgc->obd_self_export,
317 sizeof(KEY_CLEAR_FS), KEY_CLEAR_FS,
322 static inline bool is_mdc_device(const char *devname)
326 ptr = strrchr(devname, '-');
327 return ptr != NULL && strcmp(ptr, "-mdc") == 0;
330 static inline bool tgt_is_mdt(const char *tgtname, __u32 *idx)
334 type = server_name2index(tgtname, idx, NULL);
336 return type == LDD_F_SV_TYPE_MDT;
340 * Convert OST/MDT name(fsname-{MDT,OST}xxxx) to a lwp name with the @idx:yyyy
341 * (fsname-MDTyyyy-lwp-{MDT,OST}xxxx)
343 int tgt_name2lwp_name(const char *tgt_name, char *lwp_name, int len, __u32 idx)
350 OBD_ALLOC(fsname, MTI_NAME_MAXLEN);
354 rc = server_name2fsname(tgt_name, fsname, &tgt);
356 CERROR("%s: failed to get fsname from tgt_name: rc = %d\n",
361 if (*tgt != '-' && *tgt != ':') {
362 CERROR("%s: invalid tgt_name name!\n", tgt_name);
363 GOTO(cleanup, rc = -EINVAL);
367 if (strncmp(tgt, "OST", 3) != 0 && strncmp(tgt, "MDT", 3) != 0) {
368 CERROR("%s is not an OST or MDT target!\n", tgt_name);
369 GOTO(cleanup, rc = -EINVAL);
371 snprintf(lwp_name, len, "%s-MDT%04x-%s-%s",
372 fsname, idx, LUSTRE_LWP_NAME, tgt);
374 GOTO(cleanup, rc = 0);
378 OBD_FREE(fsname, MTI_NAME_MAXLEN);
382 EXPORT_SYMBOL(tgt_name2lwp_name);
384 static LIST_HEAD(lwp_register_list);
385 static DEFINE_SPINLOCK(lwp_register_list_lock);
387 static void lustre_put_lwp_item(struct lwp_register_item *lri)
389 if (atomic_dec_and_test(&lri->lri_ref)) {
390 LASSERT(list_empty(&lri->lri_list));
392 if (*lri->lri_exp != NULL)
393 class_export_put(*lri->lri_exp);
398 int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
399 register_lwp_cb cb_func, void *cb_data)
401 struct obd_device *lwp;
402 struct lwp_register_item *lri;
406 LASSERTF(strlen(lwpname) < MTI_NAME_MAXLEN, "lwpname is too long %s\n",
408 LASSERT(exp != NULL && *exp == NULL);
414 lwp = class_name2obd(lwpname);
415 if (lwp != NULL && lwp->obd_set_up == 1) {
416 struct obd_uuid *uuid;
423 memcpy(uuid->uuid, lwpname, strlen(lwpname));
424 *exp = cfs_hash_lookup(lwp->obd_uuid_hash, uuid);
428 memcpy(lri->lri_name, lwpname, strlen(lwpname));
430 lri->lri_cb_func = cb_func;
431 lri->lri_cb_data = cb_data;
432 INIT_LIST_HEAD(&lri->lri_list);
434 * Initialize the lri_ref at 2, one will be released before
435 * current function returned via lustre_put_lwp_item(), the
436 * other will be released in lustre_deregister_lwp_item().
438 atomic_set(&lri->lri_ref, 2);
440 spin_lock(&lwp_register_list_lock);
441 list_add(&lri->lri_list, &lwp_register_list);
444 spin_unlock(&lwp_register_list_lock);
446 if (cb && cb_func != NULL)
448 lustre_put_lwp_item(lri);
452 EXPORT_SYMBOL(lustre_register_lwp_item);
454 void lustre_deregister_lwp_item(struct obd_export **exp)
456 struct lwp_register_item *lri;
457 bool removed = false;
460 spin_lock(&lwp_register_list_lock);
461 list_for_each_entry(lri, &lwp_register_list, lri_list) {
462 if (exp == lri->lri_exp) {
463 list_del_init(&lri->lri_list);
468 spin_unlock(&lwp_register_list_lock);
473 /* See lustre_notify_lwp_list(), in some extreme race conditions,
474 * the notify callback could be still on the fly, we need to wait
475 * for the callback done before moving on to free the data used
477 while (atomic_read(&lri->lri_ref) > 1) {
478 CDEBUG(D_MOUNT, "lri reference count %u, repeat: %d\n",
479 atomic_read(&lri->lri_ref), repeat);
481 set_current_state(TASK_INTERRUPTIBLE);
482 schedule_timeout(cfs_time_seconds(1));
484 lustre_put_lwp_item(lri);
486 EXPORT_SYMBOL(lustre_deregister_lwp_item);
488 struct obd_export *lustre_find_lwp_by_index(const char *dev, __u32 idx)
490 struct lustre_mount_info *lmi;
491 struct lustre_sb_info *lsi;
492 struct obd_device *lwp;
493 struct obd_export *exp = NULL;
498 lmi = server_get_mount(dev);
502 lsi = s2lsi(lmi->lmi_sb);
503 rc = server_name2fsname(lsi->lsi_svname, fsname, NULL);
505 CERROR("%s: failed to get fsname: rc = %d\n",
506 lsi->lsi_svname, rc);
510 snprintf(lwp_name, sizeof(lwp_name), "%s-MDT%04x", fsname, idx);
511 mutex_lock(&lsi->lsi_lwp_mutex);
512 list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
513 char *ptr = strstr(lwp->obd_name, lwp_name);
515 if (ptr != NULL && lwp->obd_lwp_export != NULL) {
516 exp = class_export_get(lwp->obd_lwp_export);
520 mutex_unlock(&lsi->lsi_lwp_mutex);
523 server_put_mount(dev, false);
527 EXPORT_SYMBOL(lustre_find_lwp_by_index);
529 void lustre_notify_lwp_list(struct obd_export *exp)
531 struct lwp_register_item *lri;
532 LASSERT(exp != NULL);
535 spin_lock(&lwp_register_list_lock);
536 list_for_each_entry(lri, &lwp_register_list, lri_list) {
537 if (strcmp(exp->exp_obd->obd_name, lri->lri_name))
539 if (*lri->lri_exp != NULL)
541 *lri->lri_exp = class_export_get(exp);
542 if (lri->lri_cb_func == NULL)
544 atomic_inc(&lri->lri_ref);
545 spin_unlock(&lwp_register_list_lock);
547 lri->lri_cb_func(lri->lri_cb_data);
548 lustre_put_lwp_item(lri);
550 /* Others may have changed the list after we unlock, we have
551 * to rescan the list from the beginning. Usually, the list
552 * 'lwp_register_list' is very short, and there is 'guard'
553 * lri::lri_exp that will prevent the callback to be done
554 * repeatedly. So rescanning the list has no problem. */
557 spin_unlock(&lwp_register_list_lock);
559 EXPORT_SYMBOL(lustre_notify_lwp_list);
561 static int lustre_lwp_connect(struct obd_device *lwp, bool is_mdt)
564 struct lu_context session_ctx;
565 struct obd_export *exp;
566 struct obd_uuid *uuid = NULL;
567 struct obd_connect_data *data = NULL;
571 /* log has been fully processed, let clients connect */
572 rc = lu_env_init(&env, lwp->obd_lu_dev->ld_type->ldt_ctx_tags);
576 lu_context_init(&session_ctx, LCT_SERVER_SESSION);
577 session_ctx.lc_thread = NULL;
578 lu_context_enter(&session_ctx);
579 env.le_ses = &session_ctx;
583 GOTO(out, rc = -ENOMEM);
585 data->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_INDEX;
586 data->ocd_version = LUSTRE_VERSION_CODE;
587 data->ocd_connect_flags |= OBD_CONNECT_FID | OBD_CONNECT_AT |
588 OBD_CONNECT_LRU_RESIZE | OBD_CONNECT_FULL20 |
589 OBD_CONNECT_LVB_TYPE | OBD_CONNECT_LIGHTWEIGHT |
590 OBD_CONNECT_LFSCK | OBD_CONNECT_BULK_MBITS;
593 data->ocd_connect_flags |= OBD_CONNECT_MDS_MDS;
597 GOTO(out, rc = -ENOMEM);
599 if (strlen(lwp->obd_name) > sizeof(uuid->uuid)) {
600 CERROR("%s: Too long lwp name %s, max_size is %d\n",
601 lwp->obd_name, lwp->obd_name, (int)sizeof(uuid->uuid));
602 GOTO(out, rc = -EINVAL);
605 /* Use lwp name as the uuid, so we find the export by lwp name later */
606 memcpy(uuid->uuid, lwp->obd_name, strlen(lwp->obd_name));
607 rc = obd_connect(&env, &exp, lwp, uuid, data, NULL);
609 CERROR("%s: connect failed: rc = %d\n", lwp->obd_name, rc);
611 if (unlikely(lwp->obd_lwp_export != NULL))
612 class_export_put(lwp->obd_lwp_export);
613 lwp->obd_lwp_export = class_export_get(exp);
625 lu_context_exit(&session_ctx);
626 lu_context_fini(&session_ctx);
632 * lwp is used by slaves (Non-MDT0 targets) to manage the connection to MDT0,
633 * or from the OSTx to MDTy.
635 static int lustre_lwp_setup(struct lustre_cfg *lcfg, struct lustre_sb_info *lsi,
638 struct obd_device *obd;
639 char *lwpname = NULL;
640 char *lwpuuid = NULL;
644 rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
647 CERROR("%s: Can't add uuid: rc =%d\n", lsi->lsi_svname, rc);
651 OBD_ALLOC(lwpname, MTI_NAME_MAXLEN);
653 GOTO(out, rc = -ENOMEM);
655 rc = tgt_name2lwp_name(lsi->lsi_svname, lwpname, MTI_NAME_MAXLEN, idx);
657 CERROR("%s: failed to generate lwp name: rc = %d\n",
658 lsi->lsi_svname, rc);
662 OBD_ALLOC(lwpuuid, MTI_NAME_MAXLEN);
664 GOTO(out, rc = -ENOMEM);
666 sprintf(lwpuuid, "%s_UUID", lwpname);
667 rc = lustre_start_simple(lwpname, LUSTRE_LWP_NAME,
668 lwpuuid, lustre_cfg_string(lcfg, 1),
671 CERROR("%s: setup up failed: rc %d\n", lwpname, rc);
675 obd = class_name2obd(lwpname);
676 LASSERT(obd != NULL);
678 rc = lustre_lwp_connect(obd, strstr(lsi->lsi_svname, "-MDT") != NULL);
680 obd->u.cli.cl_max_mds_easize = MAX_MD_SIZE;
681 mutex_lock(&lsi->lsi_lwp_mutex);
682 list_add_tail(&obd->obd_lwp_list, &lsi->lsi_lwp_list);
683 mutex_unlock(&lsi->lsi_lwp_mutex);
685 CERROR("%s: connect failed: rc = %d\n", lwpname, rc);
692 OBD_FREE(lwpname, MTI_NAME_MAXLEN);
694 OBD_FREE(lwpuuid, MTI_NAME_MAXLEN);
699 /* the caller is responsible for memory free */
700 static struct obd_device *lustre_find_lwp(struct lustre_sb_info *lsi,
701 char **lwpname, __u32 idx)
703 struct obd_device *lwp;
707 LASSERT(lwpname != NULL);
708 LASSERT(IS_OST(lsi) || IS_MDT(lsi));
710 OBD_ALLOC(*lwpname, MTI_NAME_MAXLEN);
711 if (*lwpname == NULL)
712 RETURN(ERR_PTR(-ENOMEM));
714 rc = tgt_name2lwp_name(lsi->lsi_svname, *lwpname, MTI_NAME_MAXLEN, idx);
716 CERROR("%s: failed to generate lwp name: rc = %d\n",
717 lsi->lsi_svname, rc);
718 GOTO(out, rc = -EINVAL);
721 lwp = class_name2obd(*lwpname);
725 if (*lwpname != NULL) {
726 OBD_FREE(*lwpname, MTI_NAME_MAXLEN);
732 RETURN(lwp != NULL ? lwp : ERR_PTR(-ENOENT));
735 static int lustre_lwp_add_conn(struct lustre_cfg *cfg,
736 struct lustre_sb_info *lsi, __u32 idx)
738 struct lustre_cfg_bufs *bufs = NULL;
739 struct lustre_cfg *lcfg = NULL;
740 char *lwpname = NULL;
741 struct obd_device *lwp;
745 lwp = lustre_find_lwp(lsi, &lwpname, idx);
747 CERROR("%s: can't find lwp device.\n", lsi->lsi_svname);
748 GOTO(out, rc = PTR_ERR(lwp));
750 LASSERT(lwpname != NULL);
754 GOTO(out, rc = -ENOMEM);
756 lustre_cfg_bufs_reset(bufs, lwpname);
757 lustre_cfg_bufs_set_string(bufs, 1,
758 lustre_cfg_string(cfg, 1));
760 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
762 GOTO(out_cfg, rc = -ENOMEM);
763 lustre_cfg_init(lcfg, LCFG_ADD_CONN, bufs);
765 rc = class_add_conn(lwp, lcfg);
767 CERROR("%s: can't add conn: rc = %d\n", lwpname, rc);
770 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
771 lcfg->lcfg_buflens));
777 OBD_FREE(lwpname, MTI_NAME_MAXLEN);
782 * Retrieve MDT nids from the client log, then start the lwp device.
783 * there are only two scenarios which would include mdt nid.
785 * marker 5 (flags=0x01, v2.1.54.0) lustre-MDTyyyy 'add mdc' xxx-
786 * add_uuid nid=192.168.122.162@tcp(0x20000c0a87aa2) 0: 1:192.168.122.162@tcp
787 * attach 0:lustre-MDTyyyy-mdc 1:mdc 2:lustre-clilmv_UUID
788 * setup 0:lustre-MDTyyyy-mdc 1:lustre-MDTyyyy_UUID 2:192.168.122.162@tcp
789 * add_uuid nid=192.168.172.1@tcp(0x20000c0a8ac01) 0: 1:192.168.172.1@tcp
790 * add_conn 0:lustre-MDTyyyy-mdc 1:192.168.172.1@tcp
791 * modify_mdc_tgts add 0:lustre-clilmv 1:lustre-MDTyyyy_UUID xxxx
792 * marker 5 (flags=0x02, v2.1.54.0) lustre-MDTyyyy 'add mdc' xxxx-
794 * marker 7 (flags=0x01, v2.1.54.0) lustre-MDTyyyy 'add failnid' xxxx-
795 * add_uuid nid=192.168.122.2@tcp(0x20000c0a87a02) 0: 1:192.168.122.2@tcp
796 * add_conn 0:lustre-MDTyyyy-mdc 1:192.168.122.2@tcp
797 * marker 7 (flags=0x02, v2.1.54.0) lustre-MDTyyyy 'add failnid' xxxx-
799 static int client_lwp_config_process(const struct lu_env *env,
800 struct llog_handle *handle,
801 struct llog_rec_hdr *rec, void *data)
803 struct config_llog_instance *cfg = data;
804 int cfg_len = rec->lrh_len;
805 char *cfg_buf = (char *) (rec + 1);
806 struct lustre_cfg *lcfg = NULL;
807 struct lustre_sb_info *lsi;
808 int rc = 0, swab = 0;
811 if (rec->lrh_type != OBD_CFG_REC) {
812 CERROR("Unknown llog record type %#x encountered\n",
817 if (cfg->cfg_sb == NULL)
818 GOTO(out, rc = -EINVAL);
819 lsi = s2lsi(cfg->cfg_sb);
821 lcfg = (struct lustre_cfg *)cfg_buf;
822 if (lcfg->lcfg_version == __swab32(LUSTRE_CFG_VERSION)) {
823 lustre_swab_lustre_cfg(lcfg);
827 rc = lustre_cfg_sanity_check(cfg_buf, cfg_len);
831 switch (lcfg->lcfg_command) {
833 struct cfg_marker *marker = lustre_cfg_buf(lcfg, 1);
835 lustre_swab_cfg_marker(marker, swab,
836 LUSTRE_CFG_BUFLEN(lcfg, 1));
837 if (marker->cm_flags & CM_SKIP ||
838 marker->cm_flags & CM_EXCLUDE)
841 if (!tgt_is_mdt(marker->cm_tgtname, &cfg->cfg_lwp_idx))
844 if (IS_MDT(lsi) && cfg->cfg_lwp_idx != 0)
847 if (!strncmp(marker->cm_comment, "add mdc", 7) ||
848 !strncmp(marker->cm_comment, "add failnid", 11)) {
849 if (marker->cm_flags & CM_START) {
850 cfg->cfg_flags = CFG_F_MARKER;
851 /* This hack is to differentiate the
852 * ADD_UUID is come from "add mdc" record
853 * or from "add failnid" record. */
854 if (!strncmp(marker->cm_comment,
856 cfg->cfg_flags |= CFG_F_SKIP;
857 } else if (marker->cm_flags & CM_END) {
863 case LCFG_ADD_UUID: {
864 if (cfg->cfg_flags == CFG_F_MARKER) {
865 rc = lustre_lwp_setup(lcfg, lsi, cfg->cfg_lwp_idx);
866 /* XXX: process only the first nid as
867 * we don't need another instance of lwp */
868 cfg->cfg_flags |= CFG_F_SKIP;
869 } else if (cfg->cfg_flags == (CFG_F_MARKER | CFG_F_SKIP)) {
870 rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
873 CERROR("%s: Fail to add uuid, rc:%d\n",
874 lsi->lsi_svname, rc);
878 case LCFG_ADD_CONN: {
879 char *devname = lustre_cfg_string(lcfg, 0);
883 if (!is_mdc_device(devname))
886 ptr = strrchr(devname, '-');
891 if (!tgt_is_mdt(devname, &idx)) {
897 if (IS_MDT(lsi) && idx != 0)
900 rc = lustre_lwp_add_conn(lcfg, lsi, idx);
910 static int lustre_disconnect_lwp(struct super_block *sb)
912 struct lustre_sb_info *lsi = s2lsi(sb);
913 struct obd_device *lwp;
914 char *logname = NULL;
915 struct lustre_cfg_bufs *bufs = NULL;
916 struct config_llog_instance *cfg = NULL;
921 if (likely(lsi->lsi_lwp_started)) {
922 OBD_ALLOC(logname, MTI_NAME_MAXLEN);
926 rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
928 CERROR("%s: failed to get fsname from svname: "
929 "rc = %d\n", lsi->lsi_svname, rc);
930 GOTO(out, rc = -EINVAL);
933 strcat(logname, "-client");
936 GOTO(out, rc = -ENOMEM);
939 cfg->cfg_instance = ll_get_cfg_instance(sb);
940 rc = lustre_end_log(sb, logname, cfg);
941 if (rc != 0 && rc != -ENOENT)
944 lsi->lsi_lwp_started = 0;
949 GOTO(out, rc = -ENOMEM);
951 mutex_lock(&lsi->lsi_lwp_mutex);
952 list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
953 struct lustre_cfg *lcfg;
955 if (likely(lwp->obd_lwp_export != NULL)) {
956 class_export_put(lwp->obd_lwp_export);
957 lwp->obd_lwp_export = NULL;
960 lustre_cfg_bufs_reset(bufs, lwp->obd_name);
961 lustre_cfg_bufs_set_string(bufs, 1, NULL);
962 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount,
968 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
970 /* Disconnect import first. NULL is passed for the '@env',
971 * since it will not be used. */
972 rc = lwp->obd_lu_dev->ld_ops->ldo_process_config(NULL,
973 lwp->obd_lu_dev, lcfg);
974 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
975 lcfg->lcfg_buflens));
976 if (rc != 0 && rc != -ETIMEDOUT) {
977 CERROR("%s: fail to disconnect LWP: rc = %d\n",
982 mutex_unlock(&lsi->lsi_lwp_mutex);
992 OBD_FREE(logname, MTI_NAME_MAXLEN);
994 return rc1 != 0 ? rc1 : rc;
998 * Stop the lwp for an OST/MDT target.
1000 static int lustre_stop_lwp(struct super_block *sb)
1002 struct lustre_sb_info *lsi = s2lsi(sb);
1003 struct obd_device *lwp;
1008 mutex_lock(&lsi->lsi_lwp_mutex);
1009 while (!list_empty(&lsi->lsi_lwp_list)) {
1010 lwp = list_entry(lsi->lsi_lwp_list.next, struct obd_device,
1012 list_del_init(&lwp->obd_lwp_list);
1014 mutex_unlock(&lsi->lsi_lwp_mutex);
1016 rc = class_manual_cleanup(lwp);
1018 CERROR("%s: fail to stop LWP: rc = %d\n",
1022 mutex_lock(&lsi->lsi_lwp_mutex);
1024 mutex_unlock(&lsi->lsi_lwp_mutex);
1026 RETURN(rc1 != 0 ? rc1 : rc);
1030 * Start the lwp(fsname-MDTyyyy-lwp-{MDT,OST}xxxx) for a MDT/OST or MDT target.
1032 static int lustre_start_lwp(struct super_block *sb)
1034 struct lustre_sb_info *lsi = s2lsi(sb);
1035 struct config_llog_instance *cfg = NULL;
1040 if (unlikely(lsi->lsi_lwp_started))
1043 OBD_ALLOC(logname, MTI_NAME_MAXLEN);
1044 if (logname == NULL)
1047 rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
1049 CERROR("%s: failed to get fsname from svname: rc = %d\n",
1050 lsi->lsi_svname, rc);
1051 GOTO(out, rc = -EINVAL);
1054 strcat(logname, "-client");
1057 GOTO(out, rc = -ENOMEM);
1059 cfg->cfg_callback = client_lwp_config_process;
1060 cfg->cfg_instance = ll_get_cfg_instance(sb);
1061 rc = lustre_process_log(sb, logname, cfg);
1062 /* need to remove config llog from mgc */
1063 lsi->lsi_lwp_started = 1;
1068 OBD_FREE(logname, MTI_NAME_MAXLEN);
1075 static DEFINE_MUTEX(server_start_lock);
1077 /* Stop MDS/OSS if nobody is using them */
1078 static int server_stop_servers(int lsiflags)
1080 struct obd_device *obd = NULL;
1081 struct obd_type *type = NULL;
1085 mutex_lock(&server_start_lock);
1087 /* Either an MDT or an OST or neither */
1088 /* if this was an MDT, and there are no more MDT's, clean up the MDS */
1089 if (lsiflags & LDD_F_SV_TYPE_MDT) {
1090 obd = class_name2obd(LUSTRE_MDS_OBDNAME);
1092 type = class_search_type(LUSTRE_MDT_NAME);
1095 /* if this was an OST, and there are no more OST's, clean up the OSS */
1096 if (lsiflags & LDD_F_SV_TYPE_OST) {
1097 obd = class_name2obd(LUSTRE_OSS_OBDNAME);
1099 type = class_search_type(LUSTRE_OST_NAME);
1102 if (obd != NULL && (type == NULL || type->typ_refcnt == 0)) {
1104 /* obd_fail doesn't mean much on a server obd */
1105 rc = class_manual_cleanup(obd);
1108 mutex_unlock(&server_start_lock);
1113 int server_mti_print(const char *title, struct mgs_target_info *mti)
1115 PRINT_CMD(PRINT_MASK, "mti %s\n", title);
1116 PRINT_CMD(PRINT_MASK, "server: %s\n", mti->mti_svname);
1117 PRINT_CMD(PRINT_MASK, "fs: %s\n", mti->mti_fsname);
1118 PRINT_CMD(PRINT_MASK, "uuid: %s\n", mti->mti_uuid);
1119 PRINT_CMD(PRINT_MASK, "ver: %d flags: %#x\n",
1120 mti->mti_config_ver, mti->mti_flags);
1124 /* Generate data for registration */
1125 static int server_lsi2mti(struct lustre_sb_info *lsi,
1126 struct mgs_target_info *mti)
1128 struct lnet_process_id id;
1133 if (!IS_SERVER(lsi))
1136 if (strlcpy(mti->mti_svname, lsi->lsi_svname, sizeof(mti->mti_svname))
1137 >= sizeof(mti->mti_svname))
1140 mti->mti_nid_count = 0;
1141 while (LNetGetId(i++, &id) != -ENOENT) {
1142 if (id.nid == LNET_NID_LO_0)
1145 /* server use --servicenode param, only allow specified
1146 * nids be registered */
1147 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_NO_PRIMNODE) != 0 &&
1148 class_match_nid(lsi->lsi_lmd->lmd_params,
1149 PARAM_FAILNODE, id.nid) < 1)
1152 /* match specified network */
1153 if (!class_match_net(lsi->lsi_lmd->lmd_params,
1154 PARAM_NETWORK, LNET_NIDNET(id.nid)))
1157 mti->mti_nids[mti->mti_nid_count] = id.nid;
1158 mti->mti_nid_count++;
1159 if (mti->mti_nid_count >= MTI_NIDS_MAX) {
1160 CWARN("Only using first %d nids for %s\n",
1161 mti->mti_nid_count, mti->mti_svname);
1166 if (mti->mti_nid_count == 0) {
1167 CERROR("Failed to get NID for server %s, please check whether "
1168 "the target is specifed with improper --servicenode or "
1169 "--network options.\n", mti->mti_svname);
1173 mti->mti_lustre_ver = LUSTRE_VERSION_CODE;
1174 mti->mti_config_ver = 0;
1176 rc = server_name2fsname(lsi->lsi_svname, mti->mti_fsname, NULL);
1180 rc = server_name2index(lsi->lsi_svname, &mti->mti_stripe_index, NULL);
1183 /* Orion requires index to be set */
1184 LASSERT(!(rc & LDD_F_NEED_INDEX));
1185 /* keep only LDD flags */
1186 mti->mti_flags = lsi->lsi_flags & LDD_F_MASK;
1187 if (mti->mti_flags & (LDD_F_WRITECONF | LDD_F_VIRGIN))
1188 mti->mti_flags |= LDD_F_UPDATE;
1189 cplen = strlcpy(mti->mti_params, lsi->lsi_lmd->lmd_params,
1190 sizeof(mti->mti_params));
1191 if (cplen >= sizeof(mti->mti_params))
1196 /* Register an old or new target with the MGS. If needed MGS will construct
1197 startup logs and assign index */
1198 static int server_register_target(struct lustre_sb_info *lsi)
1200 struct obd_device *mgc = lsi->lsi_mgc;
1201 struct mgs_target_info *mti = NULL;
1209 if (!IS_SERVER(lsi))
1216 rc = server_lsi2mti(lsi, mti);
1220 CDEBUG(D_MOUNT, "Registration %s, fs=%s, %s, index=%04x, flags=%#x\n",
1221 mti->mti_svname, mti->mti_fsname,
1222 libcfs_nid2str(mti->mti_nids[0]), mti->mti_stripe_index,
1225 /* if write_conf is true, the registration must succeed */
1226 writeconf = !!(lsi->lsi_flags & (LDD_F_NEED_INDEX | LDD_F_UPDATE));
1227 mti->mti_flags |= LDD_F_OPC_REG;
1230 /* Register the target */
1231 /* FIXME use mgc_process_config instead */
1232 rc = obd_set_info_async(NULL, mgc->u.cli.cl_mgc_mgsexp,
1233 sizeof(KEY_REGISTER_TARGET),
1234 KEY_REGISTER_TARGET,
1235 sizeof(*mti), mti, NULL);
1237 if (mti->mti_flags & LDD_F_ERROR) {
1238 LCONSOLE_ERROR_MSG(0x160,
1239 "%s: the MGS refuses to allow this server "
1240 "to start: rc = %d. Please see messages on "
1241 "the MGS.\n", lsi->lsi_svname, rc);
1242 } else if (writeconf) {
1243 if ((rc == -ESHUTDOWN || rc == -EIO) && ++tried < 5) {
1244 /* The connection with MGS is not established.
1245 * Try again after 2 seconds. Interruptable. */
1246 set_current_state(TASK_INTERRUPTIBLE);
1248 msecs_to_jiffies(MSEC_PER_SEC) * 2);
1249 set_current_state(TASK_RUNNING);
1250 if (!signal_pending(current))
1254 LCONSOLE_ERROR_MSG(0x15f,
1255 "%s: cannot register this server with the MGS: "
1256 "rc = %d. Is the MGS running?\n",
1257 lsi->lsi_svname, rc);
1259 CDEBUG(D_HA, "%s: error registering with the MGS: "
1260 "rc = %d (not fatal)\n", lsi->lsi_svname, rc);
1261 /* reset the error code for non-fatal error. */
1274 * Notify the MGS that this target is ready.
1275 * Used by IR - if the MGS receives this message, it will notify clients.
1277 static int server_notify_target(struct super_block *sb, struct obd_device *obd)
1279 struct lustre_sb_info *lsi = s2lsi(sb);
1280 struct obd_device *mgc = lsi->lsi_mgc;
1281 struct mgs_target_info *mti = NULL;
1287 if (!(IS_SERVER(lsi)))
1293 rc = server_lsi2mti(lsi, mti);
1297 mti->mti_instance = obd->u.obt.obt_instance;
1298 mti->mti_flags |= LDD_F_OPC_READY;
1300 /* FIXME use mgc_process_config instead */
1301 rc = obd_set_info_async(NULL, mgc->u.cli.cl_mgc_mgsexp,
1302 sizeof(KEY_REGISTER_TARGET),
1303 KEY_REGISTER_TARGET,
1304 sizeof(*mti), mti, NULL);
1306 /* Imperative recovery: if the mgs informs us to use IR? */
1307 if (!rc && !(mti->mti_flags & LDD_F_ERROR) &&
1308 (mti->mti_flags & LDD_F_IR_CAPABLE))
1309 lsi->lsi_flags |= LDD_F_IR_CAPABLE;
1318 /** Start server targets: MDTs and OSTs
1320 static int server_start_targets(struct super_block *sb)
1322 struct obd_device *obd;
1323 struct lustre_sb_info *lsi = s2lsi(sb);
1324 struct config_llog_instance cfg;
1325 struct lu_env mgc_env;
1326 struct lu_device *dev;
1330 CDEBUG(D_MOUNT, "starting target %s\n", lsi->lsi_svname);
1333 /* make sure the MDS is started */
1334 mutex_lock(&server_start_lock);
1335 obd = class_name2obd(LUSTRE_MDS_OBDNAME);
1337 rc = lustre_start_simple(LUSTRE_MDS_OBDNAME,
1339 LUSTRE_MDS_OBDNAME"_uuid",
1340 NULL, NULL, NULL, NULL);
1342 mutex_unlock(&server_start_lock);
1343 CERROR("failed to start MDS: %d\n", rc);
1347 mutex_unlock(&server_start_lock);
1350 /* If we're an OST, make sure the global OSS is running */
1352 /* make sure OSS is started */
1353 mutex_lock(&server_start_lock);
1354 obd = class_name2obd(LUSTRE_OSS_OBDNAME);
1356 rc = lustre_start_simple(LUSTRE_OSS_OBDNAME,
1358 LUSTRE_OSS_OBDNAME"_uuid",
1359 NULL, NULL, NULL, NULL);
1361 mutex_unlock(&server_start_lock);
1362 CERROR("failed to start OSS: %d\n", rc);
1366 mutex_unlock(&server_start_lock);
1369 rc = lu_env_init(&mgc_env, LCT_MG_THREAD);
1371 GOTO(out_stop_service, rc);
1373 /* Set the mgc fs to our server disk. This allows the MGC to
1374 * read and write configs locally, in case it can't talk to the MGS. */
1375 rc = server_mgc_set_fs(&mgc_env, lsi->lsi_mgc, sb);
1379 /* Register with MGS */
1380 rc = server_register_target(lsi);
1384 /* Let the target look up the mount using the target's name
1385 (we can't pass the sb or mnt through class_process_config.) */
1386 rc = server_register_mount(lsi->lsi_svname, sb);
1390 /* Start targets using the llog named for the target */
1391 memset(&cfg, 0, sizeof(cfg));
1392 cfg.cfg_callback = class_config_llog_handler;
1393 cfg.cfg_sub_clds = CONFIG_SUB_SERVER;
1394 rc = lustre_process_log(sb, lsi->lsi_svname, &cfg);
1396 CERROR("failed to start server %s: %d\n",
1397 lsi->lsi_svname, rc);
1398 /* Do NOT call server_deregister_mount() here. This makes it
1399 * impossible to find mount later in cleanup time and leaves
1400 * @lsi and othder stuff leaked. -umka */
1404 obd = class_name2obd(lsi->lsi_svname);
1406 CERROR("no server named %s was started\n", lsi->lsi_svname);
1407 GOTO(out_mgc, rc = -ENXIO);
1410 if (IS_OST(lsi) || IS_MDT(lsi)) {
1411 rc = lustre_start_lwp(sb);
1413 CERROR("%s: failed to start LWP: %d\n",
1414 lsi->lsi_svname, rc);
1419 server_notify_target(sb, obd);
1421 /* calculate recovery timeout, do it after lustre_process_log */
1422 server_calc_timeout(lsi, obd);
1424 /* log has been fully processed, let clients connect */
1425 dev = obd->obd_lu_dev;
1426 if (dev && dev->ld_ops->ldo_prepare) {
1429 rc = lu_env_init(&env, dev->ld_type->ldt_ctx_tags);
1431 struct lu_context session_ctx;
1433 lu_context_init(&session_ctx, LCT_SERVER_SESSION);
1434 session_ctx.lc_thread = NULL;
1435 lu_context_enter(&session_ctx);
1436 env.le_ses = &session_ctx;
1438 rc = dev->ld_ops->ldo_prepare(&env, NULL, dev);
1441 lu_context_exit(&session_ctx);
1442 lu_context_fini(&session_ctx);
1446 /* abort recovery only on the complete stack:
1447 * many devices can be involved */
1448 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_ABORT_RECOV) &&
1449 (OBP(obd, iocontrol))) {
1450 obd_iocontrol(OBD_IOC_ABORT_RECOVERY, obd->obd_self_export, 0,
1455 /* Release the mgc fs for others to use */
1456 server_mgc_clear_fs(&mgc_env, lsi->lsi_mgc);
1458 lu_env_fini(&mgc_env);
1461 server_stop_servers(lsi->lsi_flags);
1466 static int lsi_prepare(struct lustre_sb_info *lsi)
1468 const char *osd_type;
1475 LASSERT(lsi->lsi_lmd);
1477 /* The server name is given as a mount line option */
1478 if (lsi->lsi_lmd->lmd_profile == NULL) {
1479 LCONSOLE_ERROR("Can't determine server name\n");
1483 /* Determine osd type */
1484 if (lsi->lsi_lmd->lmd_osd_type == NULL) {
1485 osd_type = LUSTRE_OSD_LDISKFS_NAME;
1488 osd_type = lsi->lsi_lmd->lmd_osd_type;
1489 fstype = lsi->lsi_lmd->lmd_osd_type;
1492 if (strlen(lsi->lsi_lmd->lmd_profile) >= sizeof(lsi->lsi_svname) ||
1493 strlen(osd_type) >= sizeof(lsi->lsi_osd_type) ||
1494 strlen(fstype) >= sizeof(lsi->lsi_fstype))
1495 RETURN(-ENAMETOOLONG);
1497 strlcpy(lsi->lsi_svname, lsi->lsi_lmd->lmd_profile,
1498 sizeof(lsi->lsi_svname));
1499 strlcpy(lsi->lsi_osd_type, osd_type, sizeof(lsi->lsi_osd_type));
1500 /* XXX: a temp. solution for components using ldiskfs
1501 * to be removed in one of the subsequent patches */
1502 strlcpy(lsi->lsi_fstype, fstype, sizeof(lsi->lsi_fstype));
1504 /* Determine server type */
1505 rc = server_name2index(lsi->lsi_svname, &index, NULL);
1507 if (lsi->lsi_lmd->lmd_flags & LMD_FLG_MGS) {
1508 /* Assume we're a bare MGS */
1510 lsi->lsi_lmd->lmd_flags |= LMD_FLG_NOSVC;
1512 LCONSOLE_ERROR("Can't determine server type of '%s'\n",
1517 lsi->lsi_flags |= rc;
1519 /* Add mount line flags that used to be in ldd:
1520 * writeconf, mgs, anything else?
1522 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_WRITECONF) ?
1523 LDD_F_WRITECONF : 0;
1524 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_VIRGIN) ?
1526 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_UPDATE) ?
1528 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_MGS) ?
1529 LDD_F_SV_TYPE_MGS : 0;
1530 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_NO_PRIMNODE) ?
1531 LDD_F_NO_PRIMNODE : 0;
1536 /*************** server mount ******************/
1538 /** Start the shutdown of servers at umount.
1540 static void server_put_super(struct super_block *sb)
1542 struct lustre_sb_info *lsi = s2lsi(sb);
1543 struct obd_device *obd;
1544 char *tmpname, *extraname = NULL;
1546 int lsiflags = lsi->lsi_flags;
1549 LASSERT(IS_SERVER(lsi));
1551 tmpname_sz = strlen(lsi->lsi_svname) + 1;
1552 OBD_ALLOC(tmpname, tmpname_sz);
1553 memcpy(tmpname, lsi->lsi_svname, tmpname_sz);
1554 CDEBUG(D_MOUNT, "server put_super %s\n", tmpname);
1555 if (IS_MDT(lsi) && (lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC))
1556 snprintf(tmpname, tmpname_sz, "MGS");
1558 /* disconnect the lwp first to drain off the inflight request */
1559 if (IS_OST(lsi) || IS_MDT(lsi)) {
1562 rc = lustre_disconnect_lwp(sb);
1563 if (rc != 0 && rc != -ETIMEDOUT &&
1564 rc != -ENOTCONN && rc != -ESHUTDOWN)
1565 CWARN("%s: failed to disconnect lwp: rc= %d\n",
1569 /* Stop the target */
1570 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1571 (IS_MDT(lsi) || IS_OST(lsi))) {
1572 struct lustre_profile *lprof = NULL;
1574 /* tell the mgc to drop the config log */
1575 lustre_end_log(sb, lsi->lsi_svname, NULL);
1577 /* COMPAT_146 - profile may get deleted in mgc_cleanup.
1578 If there are any setup/cleanup errors, save the lov
1579 name for safety cleanup later. */
1580 lprof = class_get_profile(lsi->lsi_svname);
1581 if (lprof != NULL) {
1582 if (lprof->lp_dt != NULL) {
1583 OBD_ALLOC(extraname, strlen(lprof->lp_dt) + 1);
1584 strncpy(extraname, lprof->lp_dt,
1585 strlen(lprof->lp_dt) + 1);
1587 class_put_profile(lprof);
1590 obd = class_name2obd(lsi->lsi_svname);
1592 CDEBUG(D_MOUNT, "stopping %s\n", obd->obd_name);
1593 if (lsiflags & LSI_UMOUNT_FAILOVER)
1595 /* We can't seem to give an error return code
1596 * to .put_super, so we better make sure we clean up! */
1598 class_manual_cleanup(obd);
1600 CERROR("no obd %s\n", lsi->lsi_svname);
1601 server_deregister_mount(lsi->lsi_svname);
1605 /* If they wanted the mgs to stop separately from the mdt, they
1606 should have put it on a different device. */
1608 /* if MDS start with --nomgs, don't stop MGS then */
1609 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS))
1610 server_stop_mgs(sb);
1613 if (IS_OST(lsi) || IS_MDT(lsi)) {
1614 if (lustre_stop_lwp(sb) < 0)
1615 CERROR("%s: failed to stop lwp!\n", tmpname);
1618 /* Clean the mgc and sb */
1619 lustre_common_put_super(sb);
1621 /* wait till all in-progress cleanups are done
1622 * specifically we're interested in ofd cleanup
1624 obd_zombie_barrier();
1626 /* Stop the servers (MDS, OSS) if no longer needed. We must wait
1627 until the target is really gone so that our type refcount check
1629 server_stop_servers(lsiflags);
1631 /* In case of startup or cleanup err, stop related obds */
1633 obd = class_name2obd(extraname);
1635 CWARN("Cleaning orphaned obd %s\n", extraname);
1637 class_manual_cleanup(obd);
1639 OBD_FREE(extraname, strlen(extraname) + 1);
1642 LCONSOLE_WARN("server umount %s complete\n", tmpname);
1643 OBD_FREE(tmpname, tmpname_sz);
1647 /** Called only for 'umount -f'
1649 static void server_umount_begin(struct super_block *sb)
1651 struct lustre_sb_info *lsi = s2lsi(sb);
1654 CDEBUG(D_MOUNT, "umount -f\n");
1655 /* umount = failover
1657 no third way to do non-force, non-failover */
1658 lsi->lsi_flags &= ~LSI_UMOUNT_FAILOVER;
1662 static int server_statfs(struct dentry *dentry, struct kstatfs *buf)
1664 struct super_block *sb = dentry->d_sb;
1665 struct lustre_sb_info *lsi = s2lsi(sb);
1666 struct obd_statfs statfs;
1670 if (lsi->lsi_dt_dev) {
1671 rc = dt_statfs(NULL, lsi->lsi_dt_dev, &statfs);
1673 statfs_unpack(buf, &statfs);
1674 buf->f_type = sb->s_magic;
1680 buf->f_type = sb->s_magic;
1681 buf->f_bsize = sb->s_blocksize;
1687 buf->f_namelen = NAME_MAX;
1691 #ifdef HAVE_SUPEROPS_USE_DENTRY
1692 int server_show_options(struct seq_file *seq, struct dentry *dentry)
1694 int server_show_options(struct seq_file *seq, struct vfsmount *vfs)
1697 struct lustre_sb_info *lsi;
1698 struct lustre_mount_data *lmd;
1700 #ifdef HAVE_SUPEROPS_USE_DENTRY
1701 LASSERT(seq != NULL && dentry != NULL);
1702 lsi = s2lsi(dentry->d_sb);
1704 LASSERT(seq != NULL && vfs != NULL);
1705 lsi = s2lsi(vfs->mnt_sb);
1709 seq_printf(seq, ",svname=%s", lmd->lmd_profile);
1711 if (lmd->lmd_flags & LMD_FLG_ABORT_RECOV)
1712 seq_puts(seq, ",abort_recov");
1714 if (lmd->lmd_flags & LMD_FLG_NOIR)
1715 seq_puts(seq, ",noir");
1717 if (lmd->lmd_flags & LMD_FLG_NOSVC)
1718 seq_puts(seq, ",nosvc");
1720 if (lmd->lmd_flags & LMD_FLG_NOMGS)
1721 seq_puts(seq, ",nomgs");
1723 if (lmd->lmd_flags & LMD_FLG_NOSCRUB)
1724 seq_puts(seq, ",noscrub");
1725 if (lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
1726 seq_puts(seq, ",skip_lfsck");
1728 if (lmd->lmd_flags & LMD_FLG_DEV_RDONLY)
1729 seq_puts(seq, ",rdonly_dev");
1731 if (lmd->lmd_flags & LMD_FLG_MGS)
1732 seq_puts(seq, ",mgs");
1734 if (lmd->lmd_mgs != NULL)
1735 seq_printf(seq, ",mgsnode=%s", lmd->lmd_mgs);
1737 if (lmd->lmd_osd_type != NULL)
1738 seq_printf(seq, ",osd=%s", lmd->lmd_osd_type);
1740 if (lmd->lmd_opts != NULL) {
1742 seq_puts(seq, lmd->lmd_opts);
1748 /** The operations we support directly on the superblock:
1749 * mount, umount, and df.
1751 static struct super_operations server_ops = {
1752 .put_super = server_put_super,
1753 .umount_begin = server_umount_begin, /* umount -f */
1754 .statfs = server_statfs,
1755 .show_options = server_show_options,
1759 * Xattr support for Lustre servers
1761 #ifdef HAVE_IOP_XATTR
1762 static ssize_t lustre_getxattr(struct dentry *dentry, const char *name,
1763 void *buffer, size_t size)
1765 if (!selinux_is_enabled())
1770 static int lustre_setxattr(struct dentry *dentry, const char *name,
1771 const void *value, size_t size, int flags)
1777 static ssize_t lustre_listxattr(struct dentry *d_entry, char *name,
1783 static bool is_cmd_supported(unsigned int command)
1795 static long server_ioctl(struct file *filp, unsigned int command,
1798 struct file active_filp;
1799 struct inode *inode = file_inode(filp);
1800 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
1801 struct super_block *dd_sb = dt_mnt_sb_get(lsi->lsi_dt_dev);
1802 struct inode *active_inode;
1803 int err = -EOPNOTSUPP;
1805 if (IS_ERR(dd_sb) || !is_cmd_supported(command))
1808 active_inode = igrab(dd_sb->s_root->d_inode);
1812 active_filp.f_inode = active_inode;
1813 if (active_inode->i_fop && active_inode->i_fop->unlocked_ioctl)
1814 err = active_inode->i_fop->unlocked_ioctl(&active_filp,
1820 static const struct inode_operations server_inode_operations = {
1821 #ifdef HAVE_IOP_XATTR
1822 .setxattr = lustre_setxattr,
1823 .getxattr = lustre_getxattr,
1825 .listxattr = lustre_listxattr,
1828 static const struct file_operations server_file_operations = {
1829 .unlocked_ioctl = server_ioctl,
1832 #define log2(n) ffz(~(n))
1833 #define LUSTRE_SUPER_MAGIC 0x0BD00BD1
1835 static int server_fill_super_common(struct super_block *sb)
1837 struct inode *root = NULL;
1840 CDEBUG(D_MOUNT, "Server sb, dev=%d\n", (int)sb->s_dev);
1842 sb->s_blocksize = 4096;
1843 sb->s_blocksize_bits = log2(sb->s_blocksize);
1844 sb->s_magic = LUSTRE_SUPER_MAGIC;
1845 sb->s_maxbytes = 0; /* we don't allow file IO on server mountpoints */
1846 sb->s_flags |= MS_RDONLY;
1847 sb->s_op = &server_ops;
1849 root = new_inode(sb);
1851 CERROR("Can't make root inode\n");
1855 /* returns -EIO for every operation */
1856 /* make_bad_inode(root); -- badness - can't umount */
1857 /* apparently we need to be a directory for the mount to finish */
1858 root->i_mode = S_IFDIR;
1859 root->i_op = &server_inode_operations;
1860 root->i_fop = &server_file_operations;
1861 sb->s_root = d_make_root(root);
1863 CERROR("%s: can't make root dentry\n", sb->s_id);
1870 static int osd_start(struct lustre_sb_info *lsi, unsigned long mflags)
1872 struct lustre_mount_data *lmd = lsi->lsi_lmd;
1873 struct obd_device *obd;
1874 struct dt_device_param p;
1875 char flagstr[20 + 1 + 10 + 1];
1880 "Attempting to start %s, type=%s, lsifl=%x, mountfl=%lx\n",
1881 lsi->lsi_svname, lsi->lsi_osd_type, lsi->lsi_flags, mflags);
1883 sprintf(lsi->lsi_osd_obdname, "%s-osd", lsi->lsi_svname);
1884 strcpy(lsi->lsi_osd_uuid, lsi->lsi_osd_obdname);
1885 strcat(lsi->lsi_osd_uuid, "_UUID");
1886 snprintf(flagstr, sizeof(flagstr), "%lu:%u", mflags, lmd->lmd_flags);
1888 obd = class_name2obd(lsi->lsi_osd_obdname);
1890 rc = lustre_start_simple(lsi->lsi_osd_obdname,
1892 lsi->lsi_osd_uuid, lmd->lmd_dev,
1893 flagstr, lsi->lsi_lmd->lmd_opts,
1897 obd = class_name2obd(lsi->lsi_osd_obdname);
1900 CDEBUG(D_MOUNT, "%s already started\n", lsi->lsi_osd_obdname);
1901 /* but continue setup to allow special case of MDT and internal
1902 * MGT being started separately. */
1903 if (!((IS_MGS(lsi) && (lsi->lsi_lmd->lmd_flags &
1905 (IS_MDT(lsi) && (lsi->lsi_lmd->lmd_flags &
1910 rc = obd_connect(NULL, &lsi->lsi_osd_exp,
1911 obd, &obd->obd_uuid, NULL, NULL);
1915 class_manual_cleanup(obd);
1916 lsi->lsi_dt_dev = NULL;
1920 LASSERT(obd->obd_lu_dev);
1921 lu_device_get(obd->obd_lu_dev);
1922 lsi->lsi_dt_dev = lu2dt_dev(obd->obd_lu_dev);
1923 LASSERT(lsi->lsi_dt_dev);
1925 /* set disk context for llog usage */
1926 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1927 obd->obd_lvfs_ctxt.dt = lsi->lsi_dt_dev;
1929 dt_conf_get(NULL, lsi->lsi_dt_dev, &p);
1934 /** Fill in the superblock info for a Lustre server.
1935 * Mount the device with the correct options.
1936 * Read the on-disk config file.
1937 * Start the services.
1939 int server_fill_super(struct super_block *sb)
1941 struct lustre_sb_info *lsi = s2lsi(sb);
1945 /* to simulate target mount race */
1946 OBD_RACE(OBD_FAIL_TGT_MOUNT_RACE);
1948 rc = lsi_prepare(lsi);
1954 /* Start low level OSD */
1955 rc = osd_start(lsi, sb->s_flags);
1957 CERROR("Unable to start osd on %s: %d\n",
1958 lsi->lsi_lmd->lmd_dev, rc);
1963 CDEBUG(D_MOUNT, "Found service %s on device %s\n",
1964 lsi->lsi_svname, lsi->lsi_lmd->lmd_dev);
1966 if (class_name2obd(lsi->lsi_svname)) {
1967 LCONSOLE_ERROR_MSG(0x161, "The target named %s is already "
1968 "running. Double-mount may have compromised"
1969 " the disk journal.\n",
1975 /* Start MGS before MGC */
1976 if (IS_MGS(lsi) && !(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS)) {
1977 rc = server_start_mgs(sb);
1982 /* Start MGC before servers */
1983 rc = lustre_start_mgc(sb);
1987 /* Set up all obd devices for service */
1988 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1989 (IS_OST(lsi) || IS_MDT(lsi))) {
1990 rc = server_start_targets(sb);
1992 CERROR("Unable to start targets: %d\n", rc);
1995 /* FIXME overmount client here, or can we just start a
1996 * client log and client_fill_super on this sb? We
1997 * need to make sure server_put_super gets called too
1998 * - ll_put_super calls lustre_common_put_super; check
1999 * there for LSI_SERVER flag, call s_p_s if so.
2001 * Probably should start client from new thread so we
2002 * can return. Client will not finish until all
2003 * servers are connected. Note - MGS-only server does
2004 * NOT get a client, since there is no lustre fs
2005 * associated - the MGS is for all lustre fs's */
2008 rc = server_fill_super_common(sb);
2014 /* We jump here in case of failure while starting targets or MGS.
2015 * In this case we can't just put @mnt and have to do real cleanup
2016 * with stoping targets, etc. */
2017 server_put_super(sb);
2022 * Calculate timeout value for a target.
2024 void server_calc_timeout(struct lustre_sb_info *lsi, struct obd_device *obd)
2026 struct lustre_mount_data *lmd;
2030 bool has_ir = !!(lsi->lsi_flags & LDD_F_IR_CAPABLE);
2031 int min = OBD_RECOVERY_TIME_MIN;
2033 LASSERT(IS_SERVER(lsi));
2037 soft = lmd->lmd_recovery_time_soft;
2038 hard = lmd->lmd_recovery_time_hard;
2039 has_ir = has_ir && !(lmd->lmd_flags & LMD_FLG_NOIR);
2040 obd->obd_no_ir = !has_ir;
2044 soft = OBD_RECOVERY_TIME_SOFT;
2046 hard = OBD_RECOVERY_TIME_HARD;
2048 /* target may have ir_factor configured. */
2049 factor = OBD_IR_FACTOR_DEFAULT;
2050 if (obd->obd_recovery_ir_factor)
2051 factor = obd->obd_recovery_ir_factor;
2054 int new_soft = soft;
2056 /* adjust timeout value by imperative recovery */
2057 new_soft = (soft * factor) / OBD_IR_FACTOR_MAX;
2058 /* make sure the timeout is not too short */
2059 new_soft = max(min, new_soft);
2061 LCONSOLE_INFO("%s: Imperative Recovery enabled, recovery "
2062 "window shrunk from %d-%d down to %d-%d\n",
2063 obd->obd_name, soft, hard, new_soft, hard);
2067 LCONSOLE_INFO("%s: Imperative Recovery not enabled, recovery "
2068 "window %d-%d\n", obd->obd_name, soft, hard);
2072 obd->obd_recovery_timeout = max_t(time64_t, obd->obd_recovery_timeout,
2074 obd->obd_recovery_time_hard = hard;
2075 obd->obd_recovery_ir_factor = factor;