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/
31 * lustre/obdclass/obd_mount_server.c
33 * Server mount routines
35 * Author: Nathan Rutman <nathan@clusterfs.com>
39 #define DEBUG_SUBSYSTEM S_CLASS
40 #define D_MOUNT (D_SUPER | D_CONFIG /* | D_WARNING */)
41 #define PRINT_CMD CDEBUG
42 #define PRINT_MASK (D_SUPER | D_CONFIG)
44 #include <linux/types.h>
45 #ifdef HAVE_LINUX_SELINUX_IS_ENABLED
46 #include <linux/selinux.h>
48 #include <linux/statfs.h>
49 #include <linux/version.h>
50 #include <linux/delay.h>
52 #include <llog_swab.h>
53 #include <lustre_disk.h>
54 #include <uapi/linux/lustre/lustre_ioctl.h>
55 #include <lustre_log.h>
56 #include <uapi/linux/lustre/lustre_param.h>
58 #include <obd_class.h>
60 /*********** mount lookup *********/
62 static DEFINE_MUTEX(lustre_mount_info_lock);
63 static LIST_HEAD(server_mount_info_list);
65 static struct lustre_mount_info *server_find_mount(const char *name)
67 struct list_head *tmp;
68 struct lustre_mount_info *lmi;
71 list_for_each(tmp, &server_mount_info_list) {
72 lmi = list_entry(tmp, struct lustre_mount_info,
74 if (strcmp(name, lmi->lmi_name) == 0)
80 /* we must register an obd for a mount before we call the setup routine.
81 *_setup will call lustre_get_mount to get the mnt struct
82 by obd_name, since we can't pass the pointer to setup. */
83 static int server_register_mount(const char *name, struct super_block *sb)
85 struct lustre_mount_info *lmi;
91 OBD_ALLOC(lmi, sizeof(*lmi));
94 OBD_ALLOC(name_cp, strlen(name) + 1);
96 OBD_FREE(lmi, sizeof(*lmi));
99 strcpy(name_cp, name);
101 mutex_lock(&lustre_mount_info_lock);
103 if (server_find_mount(name)) {
104 mutex_unlock(&lustre_mount_info_lock);
105 OBD_FREE(lmi, sizeof(*lmi));
106 OBD_FREE(name_cp, strlen(name) + 1);
107 CERROR("Already registered %s\n", name);
110 lmi->lmi_name = name_cp;
112 list_add(&lmi->lmi_list_chain, &server_mount_info_list);
114 mutex_unlock(&lustre_mount_info_lock);
116 CDEBUG(D_MOUNT, "register mount %p from %s\n", sb, name);
121 /* when an obd no longer needs a mount */
122 static int server_deregister_mount(const char *name)
124 struct lustre_mount_info *lmi;
127 mutex_lock(&lustre_mount_info_lock);
128 lmi = server_find_mount(name);
130 mutex_unlock(&lustre_mount_info_lock);
131 CERROR("%s not registered\n", name);
135 CDEBUG(D_MOUNT, "deregister mount %p from %s\n", lmi->lmi_sb, name);
137 OBD_FREE(lmi->lmi_name, strlen(lmi->lmi_name) + 1);
138 list_del(&lmi->lmi_list_chain);
139 OBD_FREE(lmi, sizeof(*lmi));
140 mutex_unlock(&lustre_mount_info_lock);
142 OBD_RACE(OBD_FAIL_MDS_LLOG_UMOUNT_RACE);
146 /* obd's look up a registered mount using their obdname. This is just
147 for initial obd setup to find the mount struct. It should not be
148 called every time you want to mntget. */
149 struct lustre_mount_info *server_get_mount(const char *name)
151 struct lustre_mount_info *lmi;
152 struct lustre_sb_info *lsi;
155 mutex_lock(&lustre_mount_info_lock);
156 lmi = server_find_mount(name);
157 mutex_unlock(&lustre_mount_info_lock);
159 CERROR("Can't find mount for %s\n", name);
162 lsi = s2lsi(lmi->lmi_sb);
164 atomic_inc(&lsi->lsi_mounts);
166 CDEBUG(D_MOUNT, "get mount %p from %s, refs=%d\n", lmi->lmi_sb,
167 name, atomic_read(&lsi->lsi_mounts));
171 EXPORT_SYMBOL(server_get_mount);
174 * server_put_mount: to be called from obd_cleanup methods
176 * @dereg_mnt: 0 or 1 depending on whether the mount is to be deregistered or
179 * The caller decides whether server_deregister_mount() needs to be called or
180 * not. Calling of server_deregister_mount() does not depend on refcounting on
181 * lsi because we could have say the mgs and mds on the same node and we
182 * unmount the mds, then the ref on the lsi would still be non-zero but we
183 * would still want to deregister the mds mount.
185 int server_put_mount(const char *name, bool dereg_mnt)
187 struct lustre_mount_info *lmi;
188 struct lustre_sb_info *lsi;
191 mutex_lock(&lustre_mount_info_lock);
192 lmi = server_find_mount(name);
193 mutex_unlock(&lustre_mount_info_lock);
195 CERROR("Can't find mount for %s\n", name);
198 lsi = s2lsi(lmi->lmi_sb);
200 CDEBUG(D_MOUNT, "put mount %p from %s, refs=%d\n",
201 lmi->lmi_sb, name, atomic_read(&lsi->lsi_mounts));
203 if (lustre_put_lsi(lmi->lmi_sb))
204 CDEBUG(D_MOUNT, "Last put of mount %p from %s\n",
208 /* this obd should never need the mount again */
209 server_deregister_mount(name);
213 EXPORT_SYMBOL(server_put_mount);
215 /* Set up a MGS to serve startup logs */
216 static int server_start_mgs(struct super_block *sb)
218 struct lustre_sb_info *lsi = s2lsi(sb);
219 struct lustre_mount_info *lmi;
223 /* It is impossible to have more than 1 MGS per node, since
224 MGC wouldn't know which to connect to */
225 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
227 lsi = s2lsi(lmi->lmi_sb);
228 LCONSOLE_ERROR_MSG(0x15d, "The MGS service was already started"
233 CDEBUG(D_CONFIG, "Start MGS service %s\n", LUSTRE_MGS_OBDNAME);
235 rc = server_register_mount(LUSTRE_MGS_OBDNAME, sb);
238 rc = lustre_start_simple(LUSTRE_MGS_OBDNAME, LUSTRE_MGS_NAME,
239 LUSTRE_MGS_OBDNAME, NULL, NULL,
240 lsi->lsi_osd_obdname, NULL);
241 /* server_deregister_mount() is not called previously, for lsi
242 * and other stuff can't be freed cleanly when mgs calls
243 * server_put_mount() in error handling case (see b=17758),
244 * this problem is caused by a bug in mgs_init0, which forgot
245 * calling server_put_mount in error case. */
248 server_deregister_mount(LUSTRE_MGS_OBDNAME);
252 LCONSOLE_ERROR_MSG(0x15e, "Failed to start MGS '%s' (%d). "
253 "Is the 'mgs' module loaded?\n",
254 LUSTRE_MGS_OBDNAME, rc);
258 static int server_stop_mgs(struct super_block *sb)
260 struct obd_device *obd;
262 struct lustre_mount_info *lmi;
265 /* Do not stop MGS if this device is not the running MGT */
266 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
267 if (lmi != NULL && lmi->lmi_sb != sb)
270 CDEBUG(D_MOUNT, "Stop MGS service %s\n", LUSTRE_MGS_OBDNAME);
272 /* There better be only one MGS */
273 obd = class_name2obd(LUSTRE_MGS_OBDNAME);
275 CDEBUG(D_CONFIG, "mgs %s not running\n", LUSTRE_MGS_OBDNAME);
279 /* The MGS should always stop when we say so */
281 rc = class_manual_cleanup(obd);
285 /* Since there's only one mgc per node, we have to change it's fs to get
286 access to the right disk. */
287 static int server_mgc_set_fs(const struct lu_env *env,
288 struct obd_device *mgc, struct super_block *sb)
290 struct lustre_sb_info *lsi = s2lsi(sb);
294 CDEBUG(D_MOUNT, "Set mgc disk for %s\n", lsi->lsi_lmd->lmd_dev);
296 /* cl_mgc_sem in mgc insures we sleep if the mgc_fs is busy */
297 rc = obd_set_info_async(env, mgc->obd_self_export,
298 sizeof(KEY_SET_FS), KEY_SET_FS,
299 sizeof(*sb), sb, NULL);
301 CERROR("can't set_fs %d\n", rc);
306 static int server_mgc_clear_fs(const struct lu_env *env,
307 struct obd_device *mgc)
312 CDEBUG(D_MOUNT, "Unassign mgc disk\n");
314 rc = obd_set_info_async(env, mgc->obd_self_export,
315 sizeof(KEY_CLEAR_FS), KEY_CLEAR_FS,
320 static inline bool is_mdc_device(const char *devname)
324 ptr = strrchr(devname, '-');
325 return ptr != NULL && strcmp(ptr, "-mdc") == 0;
328 static inline bool tgt_is_mdt(const char *tgtname, __u32 *idx)
332 type = server_name2index(tgtname, idx, NULL);
334 return type == LDD_F_SV_TYPE_MDT;
338 * Convert OST/MDT name(fsname-{MDT,OST}xxxx) to a lwp name with the @idx:yyyy
339 * (fsname-MDTyyyy-lwp-{MDT,OST}xxxx)
341 int tgt_name2lwp_name(const char *tgt_name, char *lwp_name, int len, __u32 idx)
348 OBD_ALLOC(fsname, MTI_NAME_MAXLEN);
352 rc = server_name2fsname(tgt_name, fsname, &tgt);
354 CERROR("%s: failed to get fsname from tgt_name: rc = %d\n",
359 if (*tgt != '-' && *tgt != ':') {
360 CERROR("%s: invalid tgt_name name!\n", tgt_name);
361 GOTO(cleanup, rc = -EINVAL);
365 if (strncmp(tgt, "OST", 3) != 0 && strncmp(tgt, "MDT", 3) != 0) {
366 CERROR("%s is not an OST or MDT target!\n", tgt_name);
367 GOTO(cleanup, rc = -EINVAL);
369 snprintf(lwp_name, len, "%s-MDT%04x-%s-%s",
370 fsname, idx, LUSTRE_LWP_NAME, tgt);
372 GOTO(cleanup, rc = 0);
376 OBD_FREE(fsname, MTI_NAME_MAXLEN);
380 EXPORT_SYMBOL(tgt_name2lwp_name);
382 static LIST_HEAD(lwp_register_list);
383 static DEFINE_SPINLOCK(lwp_register_list_lock);
385 static void lustre_put_lwp_item(struct lwp_register_item *lri)
387 if (atomic_dec_and_test(&lri->lri_ref)) {
388 LASSERT(list_empty(&lri->lri_list));
390 if (*lri->lri_exp != NULL)
391 class_export_put(*lri->lri_exp);
396 int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
397 register_lwp_cb cb_func, void *cb_data)
399 struct obd_device *lwp;
400 struct lwp_register_item *lri;
404 LASSERTF(strlen(lwpname) < MTI_NAME_MAXLEN, "lwpname is too long %s\n",
406 LASSERT(exp != NULL && *exp == NULL);
412 lwp = class_name2obd(lwpname);
413 if (lwp != NULL && lwp->obd_set_up == 1) {
414 struct obd_uuid *uuid;
421 memcpy(uuid->uuid, lwpname, strlen(lwpname));
422 *exp = obd_uuid_lookup(lwp, uuid);
426 memcpy(lri->lri_name, lwpname, strlen(lwpname));
428 lri->lri_cb_func = cb_func;
429 lri->lri_cb_data = cb_data;
430 INIT_LIST_HEAD(&lri->lri_list);
432 * Initialize the lri_ref at 2, one will be released before
433 * current function returned via lustre_put_lwp_item(), the
434 * other will be released in lustre_deregister_lwp_item().
436 atomic_set(&lri->lri_ref, 2);
438 spin_lock(&lwp_register_list_lock);
439 list_add(&lri->lri_list, &lwp_register_list);
442 spin_unlock(&lwp_register_list_lock);
444 if (cb && cb_func != NULL)
446 lustre_put_lwp_item(lri);
450 EXPORT_SYMBOL(lustre_register_lwp_item);
452 void lustre_deregister_lwp_item(struct obd_export **exp)
454 struct lwp_register_item *lri;
455 bool removed = false;
458 spin_lock(&lwp_register_list_lock);
459 list_for_each_entry(lri, &lwp_register_list, lri_list) {
460 if (exp == lri->lri_exp) {
461 list_del_init(&lri->lri_list);
466 spin_unlock(&lwp_register_list_lock);
471 /* See lustre_notify_lwp_list(), in some extreme race conditions,
472 * the notify callback could be still on the fly, we need to wait
473 * for the callback done before moving on to free the data used
475 while (atomic_read(&lri->lri_ref) > 1) {
476 CDEBUG(D_MOUNT, "lri reference count %u, repeat: %d\n",
477 atomic_read(&lri->lri_ref), repeat);
479 schedule_timeout_interruptible(cfs_time_seconds(1));
481 lustre_put_lwp_item(lri);
483 EXPORT_SYMBOL(lustre_deregister_lwp_item);
485 struct obd_export *lustre_find_lwp_by_index(const char *dev, __u32 idx)
487 struct lustre_mount_info *lmi;
488 struct lustre_sb_info *lsi;
489 struct obd_device *lwp;
490 struct obd_export *exp = NULL;
495 lmi = server_get_mount(dev);
499 lsi = s2lsi(lmi->lmi_sb);
500 rc = server_name2fsname(lsi->lsi_svname, fsname, NULL);
502 CERROR("%s: failed to get fsname: rc = %d\n",
503 lsi->lsi_svname, rc);
507 snprintf(lwp_name, sizeof(lwp_name), "%s-MDT%04x", fsname, idx);
508 mutex_lock(&lsi->lsi_lwp_mutex);
509 list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
510 char *ptr = strstr(lwp->obd_name, lwp_name);
512 if (ptr != NULL && lwp->obd_lwp_export != NULL) {
513 exp = class_export_get(lwp->obd_lwp_export);
517 mutex_unlock(&lsi->lsi_lwp_mutex);
520 server_put_mount(dev, false);
524 EXPORT_SYMBOL(lustre_find_lwp_by_index);
526 void lustre_notify_lwp_list(struct obd_export *exp)
528 struct lwp_register_item *lri;
529 LASSERT(exp != NULL);
532 spin_lock(&lwp_register_list_lock);
533 list_for_each_entry(lri, &lwp_register_list, lri_list) {
534 if (strcmp(exp->exp_obd->obd_name, lri->lri_name))
536 if (*lri->lri_exp != NULL)
538 *lri->lri_exp = class_export_get(exp);
539 if (lri->lri_cb_func == NULL)
541 atomic_inc(&lri->lri_ref);
542 spin_unlock(&lwp_register_list_lock);
544 lri->lri_cb_func(lri->lri_cb_data);
545 lustre_put_lwp_item(lri);
547 /* Others may have changed the list after we unlock, we have
548 * to rescan the list from the beginning. Usually, the list
549 * 'lwp_register_list' is very short, and there is 'guard'
550 * lri::lri_exp that will prevent the callback to be done
551 * repeatedly. So rescanning the list has no problem. */
554 spin_unlock(&lwp_register_list_lock);
556 EXPORT_SYMBOL(lustre_notify_lwp_list);
558 static int lustre_lwp_connect(struct obd_device *lwp, bool is_mdt)
561 struct lu_context session_ctx;
562 struct obd_export *exp;
563 struct obd_uuid *uuid = NULL;
564 struct obd_connect_data *data = NULL;
568 /* log has been fully processed, let clients connect */
569 rc = lu_env_init(&env, lwp->obd_lu_dev->ld_type->ldt_ctx_tags);
573 lu_context_init(&session_ctx, LCT_SERVER_SESSION);
574 session_ctx.lc_thread = NULL;
575 lu_context_enter(&session_ctx);
576 env.le_ses = &session_ctx;
580 GOTO(out, rc = -ENOMEM);
582 data->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_INDEX;
583 data->ocd_version = LUSTRE_VERSION_CODE;
584 data->ocd_connect_flags |= OBD_CONNECT_FID | OBD_CONNECT_AT |
585 OBD_CONNECT_LRU_RESIZE | OBD_CONNECT_FULL20 |
586 OBD_CONNECT_LVB_TYPE | OBD_CONNECT_LIGHTWEIGHT |
587 OBD_CONNECT_LFSCK | OBD_CONNECT_BULK_MBITS;
590 data->ocd_connect_flags |= OBD_CONNECT_MDS_MDS;
594 GOTO(out, rc = -ENOMEM);
596 if (strlen(lwp->obd_name) > sizeof(uuid->uuid)) {
597 CERROR("%s: Too long lwp name %s, max_size is %d\n",
598 lwp->obd_name, lwp->obd_name, (int)sizeof(uuid->uuid));
599 GOTO(out, rc = -EINVAL);
602 /* Use lwp name as the uuid, so we find the export by lwp name later */
603 memcpy(uuid->uuid, lwp->obd_name, strlen(lwp->obd_name));
604 rc = obd_connect(&env, &exp, lwp, uuid, data, NULL);
606 CERROR("%s: connect failed: rc = %d\n", lwp->obd_name, rc);
608 if (unlikely(lwp->obd_lwp_export != NULL))
609 class_export_put(lwp->obd_lwp_export);
610 lwp->obd_lwp_export = class_export_get(exp);
622 lu_context_exit(&session_ctx);
623 lu_context_fini(&session_ctx);
629 * lwp is used by slaves (Non-MDT0 targets) to manage the connection to MDT0,
630 * or from the OSTx to MDTy.
632 static int lustre_lwp_setup(struct lustre_cfg *lcfg, struct lustre_sb_info *lsi,
635 struct obd_device *obd;
636 char *lwpname = NULL;
637 char *lwpuuid = NULL;
641 rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
644 CERROR("%s: Can't add uuid: rc =%d\n", lsi->lsi_svname, rc);
648 OBD_ALLOC(lwpname, MTI_NAME_MAXLEN);
650 GOTO(out, rc = -ENOMEM);
652 rc = tgt_name2lwp_name(lsi->lsi_svname, lwpname, MTI_NAME_MAXLEN, idx);
654 CERROR("%s: failed to generate lwp name: rc = %d\n",
655 lsi->lsi_svname, rc);
659 OBD_ALLOC(lwpuuid, MTI_NAME_MAXLEN);
661 GOTO(out, rc = -ENOMEM);
663 sprintf(lwpuuid, "%s_UUID", lwpname);
664 rc = lustre_start_simple(lwpname, LUSTRE_LWP_NAME,
665 lwpuuid, lustre_cfg_string(lcfg, 1),
668 CERROR("%s: setup up failed: rc %d\n", lwpname, rc);
672 obd = class_name2obd(lwpname);
673 LASSERT(obd != NULL);
675 rc = lustre_lwp_connect(obd, strstr(lsi->lsi_svname, "-MDT") != NULL);
677 obd->u.cli.cl_max_mds_easize = MAX_MD_SIZE;
678 mutex_lock(&lsi->lsi_lwp_mutex);
679 list_add_tail(&obd->obd_lwp_list, &lsi->lsi_lwp_list);
680 mutex_unlock(&lsi->lsi_lwp_mutex);
682 CERROR("%s: connect failed: rc = %d\n", lwpname, rc);
689 OBD_FREE(lwpname, MTI_NAME_MAXLEN);
691 OBD_FREE(lwpuuid, MTI_NAME_MAXLEN);
696 /* the caller is responsible for memory free */
697 static struct obd_device *lustre_find_lwp(struct lustre_sb_info *lsi,
698 char **lwpname, __u32 idx)
700 struct obd_device *lwp;
704 LASSERT(lwpname != NULL);
705 LASSERT(IS_OST(lsi) || IS_MDT(lsi));
707 OBD_ALLOC(*lwpname, MTI_NAME_MAXLEN);
708 if (*lwpname == NULL)
709 RETURN(ERR_PTR(-ENOMEM));
711 rc = tgt_name2lwp_name(lsi->lsi_svname, *lwpname, MTI_NAME_MAXLEN, idx);
713 CERROR("%s: failed to generate lwp name: rc = %d\n",
714 lsi->lsi_svname, rc);
715 GOTO(out, rc = -EINVAL);
718 lwp = class_name2obd(*lwpname);
722 if (*lwpname != NULL) {
723 OBD_FREE(*lwpname, MTI_NAME_MAXLEN);
729 RETURN(lwp != NULL ? lwp : ERR_PTR(-ENOENT));
732 static int lustre_lwp_add_conn(struct lustre_cfg *cfg,
733 struct lustre_sb_info *lsi, __u32 idx)
735 struct lustre_cfg_bufs *bufs = NULL;
736 struct lustre_cfg *lcfg = NULL;
737 char *lwpname = NULL;
738 struct obd_device *lwp;
742 lwp = lustre_find_lwp(lsi, &lwpname, idx);
744 CERROR("%s: can't find lwp device.\n", lsi->lsi_svname);
745 GOTO(out, rc = PTR_ERR(lwp));
747 LASSERT(lwpname != NULL);
751 GOTO(out, rc = -ENOMEM);
753 lustre_cfg_bufs_reset(bufs, lwpname);
754 lustre_cfg_bufs_set_string(bufs, 1,
755 lustre_cfg_string(cfg, 1));
757 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
759 GOTO(out_cfg, rc = -ENOMEM);
760 lustre_cfg_init(lcfg, LCFG_ADD_CONN, bufs);
762 rc = class_add_conn(lwp, lcfg);
764 CERROR("%s: can't add conn: rc = %d\n", lwpname, rc);
767 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
768 lcfg->lcfg_buflens));
774 OBD_FREE(lwpname, MTI_NAME_MAXLEN);
779 * Retrieve MDT nids from the client log, then start the lwp device.
780 * there are only two scenarios which would include mdt nid.
782 * marker 5 (flags=0x01, v2.1.54.0) lustre-MDTyyyy 'add mdc' xxx-
783 * add_uuid nid=192.168.122.162@tcp(0x20000c0a87aa2) 0: 1:192.168.122.162@tcp
784 * attach 0:lustre-MDTyyyy-mdc 1:mdc 2:lustre-clilmv_UUID
785 * setup 0:lustre-MDTyyyy-mdc 1:lustre-MDTyyyy_UUID 2:192.168.122.162@tcp
786 * add_uuid nid=192.168.172.1@tcp(0x20000c0a8ac01) 0: 1:192.168.172.1@tcp
787 * add_conn 0:lustre-MDTyyyy-mdc 1:192.168.172.1@tcp
788 * modify_mdc_tgts add 0:lustre-clilmv 1:lustre-MDTyyyy_UUID xxxx
789 * marker 5 (flags=0x02, v2.1.54.0) lustre-MDTyyyy 'add mdc' xxxx-
791 * marker 7 (flags=0x01, v2.1.54.0) lustre-MDTyyyy 'add failnid' xxxx-
792 * add_uuid nid=192.168.122.2@tcp(0x20000c0a87a02) 0: 1:192.168.122.2@tcp
793 * add_conn 0:lustre-MDTyyyy-mdc 1:192.168.122.2@tcp
794 * marker 7 (flags=0x02, v2.1.54.0) lustre-MDTyyyy 'add failnid' xxxx-
796 static int client_lwp_config_process(const struct lu_env *env,
797 struct llog_handle *handle,
798 struct llog_rec_hdr *rec, void *data)
800 struct config_llog_instance *cfg = data;
801 int cfg_len = rec->lrh_len;
802 char *cfg_buf = (char *) (rec + 1);
803 struct lustre_cfg *lcfg = NULL;
804 struct lustre_sb_info *lsi;
805 int rc = 0, swab = 0;
808 if (rec->lrh_type != OBD_CFG_REC) {
809 CERROR("Unknown llog record type %#x encountered\n",
814 if (cfg->cfg_sb == NULL)
815 GOTO(out, rc = -EINVAL);
816 lsi = s2lsi(cfg->cfg_sb);
818 lcfg = (struct lustre_cfg *)cfg_buf;
819 if (lcfg->lcfg_version == __swab32(LUSTRE_CFG_VERSION)) {
820 lustre_swab_lustre_cfg(lcfg);
824 rc = lustre_cfg_sanity_check(cfg_buf, cfg_len);
828 switch (lcfg->lcfg_command) {
830 struct cfg_marker *marker = lustre_cfg_buf(lcfg, 1);
832 lustre_swab_cfg_marker(marker, swab,
833 LUSTRE_CFG_BUFLEN(lcfg, 1));
834 if (marker->cm_flags & CM_SKIP ||
835 marker->cm_flags & CM_EXCLUDE)
838 if (!tgt_is_mdt(marker->cm_tgtname, &cfg->cfg_lwp_idx))
841 if (IS_MDT(lsi) && cfg->cfg_lwp_idx != 0)
844 if (!strncmp(marker->cm_comment, "add mdc", 7) ||
845 !strncmp(marker->cm_comment, "add failnid", 11)) {
846 if (marker->cm_flags & CM_START) {
847 cfg->cfg_flags = CFG_F_MARKER;
848 /* This hack is to differentiate the
849 * ADD_UUID is come from "add mdc" record
850 * or from "add failnid" record. */
851 if (!strncmp(marker->cm_comment,
853 cfg->cfg_flags |= CFG_F_SKIP;
854 } else if (marker->cm_flags & CM_END) {
860 case LCFG_ADD_UUID: {
861 if (cfg->cfg_flags == CFG_F_MARKER) {
862 rc = lustre_lwp_setup(lcfg, lsi, cfg->cfg_lwp_idx);
863 /* XXX: process only the first nid as
864 * we don't need another instance of lwp */
865 cfg->cfg_flags |= CFG_F_SKIP;
866 } else if (cfg->cfg_flags == (CFG_F_MARKER | CFG_F_SKIP)) {
867 rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
870 CERROR("%s: Fail to add uuid, rc:%d\n",
871 lsi->lsi_svname, rc);
875 case LCFG_ADD_CONN: {
876 char *devname = lustre_cfg_string(lcfg, 0);
880 if (!is_mdc_device(devname))
883 if (!(cfg->cfg_flags & CFG_F_MARKER)) {
884 CDEBUG(D_CONFIG, "Skipping add_conn for %s, rec %d\n",
885 devname, rec->lrh_index);
889 /* add_conn should follow by add_uuid. This
890 * guarantee lwp device was created
892 if (!(cfg->cfg_flags & CFG_F_SKIP)) {
893 CWARN("Error at config for %s rec %d, add_conn should follow by add_uuid\n",
894 devname, rec->lrh_index);
897 ptr = strrchr(devname, '-');
902 if (!tgt_is_mdt(devname, &idx)) {
908 if (IS_MDT(lsi) && idx != 0)
911 rc = lustre_lwp_add_conn(lcfg, lsi, idx);
921 static int lustre_disconnect_lwp(struct super_block *sb)
923 struct lustre_sb_info *lsi = s2lsi(sb);
924 struct obd_device *lwp;
925 char *logname = NULL;
926 struct lustre_cfg_bufs *bufs = NULL;
927 struct config_llog_instance *cfg = NULL;
932 if (likely(lsi->lsi_lwp_started)) {
933 OBD_ALLOC(logname, MTI_NAME_MAXLEN);
937 rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
939 CERROR("%s: failed to get fsname from svname: "
940 "rc = %d\n", lsi->lsi_svname, rc);
941 GOTO(out, rc = -EINVAL);
944 strcat(logname, "-client");
947 GOTO(out, rc = -ENOMEM);
950 cfg->cfg_instance = ll_get_cfg_instance(sb);
951 rc = lustre_end_log(sb, logname, cfg);
952 if (rc != 0 && rc != -ENOENT)
955 lsi->lsi_lwp_started = 0;
960 GOTO(out, rc = -ENOMEM);
962 mutex_lock(&lsi->lsi_lwp_mutex);
963 list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
964 struct lustre_cfg *lcfg;
966 if (likely(lwp->obd_lwp_export != NULL)) {
967 class_export_put(lwp->obd_lwp_export);
968 lwp->obd_lwp_export = NULL;
971 lustre_cfg_bufs_reset(bufs, lwp->obd_name);
972 lustre_cfg_bufs_set_string(bufs, 1, NULL);
973 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount,
979 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
981 /* Disconnect import first. NULL is passed for the '@env',
982 * since it will not be used. */
983 rc = lwp->obd_lu_dev->ld_ops->ldo_process_config(NULL,
984 lwp->obd_lu_dev, lcfg);
985 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
986 lcfg->lcfg_buflens));
987 if (rc != 0 && rc != -ETIMEDOUT) {
988 CERROR("%s: fail to disconnect LWP: rc = %d\n",
993 mutex_unlock(&lsi->lsi_lwp_mutex);
1002 if (logname != NULL)
1003 OBD_FREE(logname, MTI_NAME_MAXLEN);
1005 return rc1 != 0 ? rc1 : rc;
1009 * Stop the lwp for an OST/MDT target.
1011 static int lustre_stop_lwp(struct super_block *sb)
1013 struct lustre_sb_info *lsi = s2lsi(sb);
1014 struct obd_device *lwp;
1019 mutex_lock(&lsi->lsi_lwp_mutex);
1020 while (!list_empty(&lsi->lsi_lwp_list)) {
1021 lwp = list_entry(lsi->lsi_lwp_list.next, struct obd_device,
1023 list_del_init(&lwp->obd_lwp_list);
1025 mutex_unlock(&lsi->lsi_lwp_mutex);
1027 rc = class_manual_cleanup(lwp);
1029 CERROR("%s: fail to stop LWP: rc = %d\n",
1033 mutex_lock(&lsi->lsi_lwp_mutex);
1035 mutex_unlock(&lsi->lsi_lwp_mutex);
1037 RETURN(rc1 != 0 ? rc1 : rc);
1041 * Start the lwp(fsname-MDTyyyy-lwp-{MDT,OST}xxxx) for a MDT/OST or MDT target.
1043 static int lustre_start_lwp(struct super_block *sb)
1045 struct lustre_sb_info *lsi = s2lsi(sb);
1046 struct config_llog_instance *cfg = NULL;
1051 if (unlikely(lsi->lsi_lwp_started))
1054 OBD_ALLOC(logname, MTI_NAME_MAXLEN);
1055 if (logname == NULL)
1058 rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
1060 CERROR("%s: failed to get fsname from svname: rc = %d\n",
1061 lsi->lsi_svname, rc);
1062 GOTO(out, rc = -EINVAL);
1065 strcat(logname, "-client");
1068 GOTO(out, rc = -ENOMEM);
1070 cfg->cfg_callback = client_lwp_config_process;
1071 cfg->cfg_instance = ll_get_cfg_instance(sb);
1072 rc = lustre_process_log(sb, logname, cfg);
1073 /* need to remove config llog from mgc */
1074 lsi->lsi_lwp_started = 1;
1079 OBD_FREE(logname, MTI_NAME_MAXLEN);
1086 static DEFINE_MUTEX(server_start_lock);
1088 /* Stop MDS/OSS if nobody is using them */
1089 static int server_stop_servers(int lsiflags)
1091 struct obd_device *obd = NULL;
1092 struct obd_type *type = NULL;
1097 mutex_lock(&server_start_lock);
1099 /* Either an MDT or an OST or neither */
1100 /* if this was an MDT, and there are no more MDT's, clean up the MDS */
1101 if (lsiflags & LDD_F_SV_TYPE_MDT) {
1102 obd = class_name2obd(LUSTRE_MDS_OBDNAME);
1103 type = class_search_type(LUSTRE_MDT_NAME);
1104 } else if (lsiflags & LDD_F_SV_TYPE_OST) {
1105 /* if this was an OST, and there are no more OST's, clean up the OSS */
1106 obd = class_name2obd(LUSTRE_OSS_OBDNAME);
1107 type = class_search_type(LUSTRE_OST_NAME);
1110 /* server_stop_servers is a pair of server_start_targets
1111 * Here we put type which was taken at server_start_targets.
1112 * If type is NULL then there is a wrong logic around type or
1113 * type reference. */
1114 LASSERTF(type, "Server flags %d, obd %s\n", lsiflags,
1115 obd ? obd->obd_name : "NULL");
1117 type_last = (atomic_read(&type->typ_refcnt) == 1);
1119 class_put_type(type);
1120 if (obd != NULL && type_last) {
1122 /* obd_fail doesn't mean much on a server obd */
1123 rc = class_manual_cleanup(obd);
1126 /* put reference taken by class_search_type */
1127 kobject_put(&type->typ_kobj);
1129 mutex_unlock(&server_start_lock);
1134 int server_mti_print(const char *title, struct mgs_target_info *mti)
1136 PRINT_CMD(PRINT_MASK, "mti %s\n", title);
1137 PRINT_CMD(PRINT_MASK, "server: %s\n", mti->mti_svname);
1138 PRINT_CMD(PRINT_MASK, "fs: %s\n", mti->mti_fsname);
1139 PRINT_CMD(PRINT_MASK, "uuid: %s\n", mti->mti_uuid);
1140 PRINT_CMD(PRINT_MASK, "ver: %d flags: %#x\n",
1141 mti->mti_config_ver, mti->mti_flags);
1145 /* Generate data for registration */
1146 static int server_lsi2mti(struct lustre_sb_info *lsi,
1147 struct mgs_target_info *mti)
1149 struct lnet_process_id id;
1154 if (!IS_SERVER(lsi))
1157 if (strlcpy(mti->mti_svname, lsi->lsi_svname, sizeof(mti->mti_svname))
1158 >= sizeof(mti->mti_svname))
1161 mti->mti_nid_count = 0;
1162 while (LNetGetId(i++, &id) != -ENOENT) {
1163 if (id.nid == LNET_NID_LO_0)
1166 /* server use --servicenode param, only allow specified
1167 * nids be registered */
1168 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_NO_PRIMNODE) != 0 &&
1169 class_match_nid(lsi->lsi_lmd->lmd_params,
1170 PARAM_FAILNODE, id.nid) < 1)
1173 /* match specified network */
1174 if (!class_match_net(lsi->lsi_lmd->lmd_params,
1175 PARAM_NETWORK, LNET_NIDNET(id.nid)))
1178 mti->mti_nids[mti->mti_nid_count] = id.nid;
1179 mti->mti_nid_count++;
1180 if (mti->mti_nid_count >= MTI_NIDS_MAX) {
1181 CWARN("Only using first %d nids for %s\n",
1182 mti->mti_nid_count, mti->mti_svname);
1187 if (mti->mti_nid_count == 0) {
1188 CERROR("Failed to get NID for server %s, please check whether "
1189 "the target is specifed with improper --servicenode or "
1190 "--network options.\n", mti->mti_svname);
1194 mti->mti_lustre_ver = LUSTRE_VERSION_CODE;
1195 mti->mti_config_ver = 0;
1197 rc = server_name2fsname(lsi->lsi_svname, mti->mti_fsname, NULL);
1201 rc = server_name2index(lsi->lsi_svname, &mti->mti_stripe_index, NULL);
1204 /* Orion requires index to be set */
1205 LASSERT(!(rc & LDD_F_NEED_INDEX));
1206 /* keep only LDD flags */
1207 mti->mti_flags = lsi->lsi_flags & LDD_F_MASK;
1208 if (mti->mti_flags & (LDD_F_WRITECONF | LDD_F_VIRGIN))
1209 mti->mti_flags |= LDD_F_UPDATE;
1210 cplen = strlcpy(mti->mti_params, lsi->lsi_lmd->lmd_params,
1211 sizeof(mti->mti_params));
1212 if (cplen >= sizeof(mti->mti_params))
1217 /* Register an old or new target with the MGS. If needed MGS will construct
1218 startup logs and assign index */
1219 static int server_register_target(struct lustre_sb_info *lsi)
1221 struct obd_device *mgc = lsi->lsi_mgc;
1222 struct mgs_target_info *mti = NULL;
1230 if (!IS_SERVER(lsi))
1237 rc = server_lsi2mti(lsi, mti);
1241 CDEBUG(D_MOUNT, "Registration %s, fs=%s, %s, index=%04x, flags=%#x\n",
1242 mti->mti_svname, mti->mti_fsname,
1243 libcfs_nid2str(mti->mti_nids[0]), mti->mti_stripe_index,
1246 /* we cannot ignore registration failure if MGS logs must be updated. */
1247 must_succeed = !!(lsi->lsi_flags &
1248 (LDD_F_NEED_INDEX | LDD_F_UPDATE | LDD_F_WRITECONF |
1250 mti->mti_flags |= LDD_F_OPC_REG;
1253 /* Register the target */
1254 /* FIXME use mgc_process_config instead */
1255 rc = obd_set_info_async(NULL, mgc->u.cli.cl_mgc_mgsexp,
1256 sizeof(KEY_REGISTER_TARGET),
1257 KEY_REGISTER_TARGET,
1258 sizeof(*mti), mti, NULL);
1260 if (mti->mti_flags & LDD_F_ERROR) {
1261 LCONSOLE_ERROR_MSG(0x160,
1262 "%s: the MGS refuses to allow this server "
1263 "to start: rc = %d. Please see messages on "
1264 "the MGS.\n", lsi->lsi_svname, rc);
1265 } else if (must_succeed) {
1266 if ((rc == -ESHUTDOWN || rc == -EIO) && ++tried < 5) {
1267 /* The connection with MGS is not established.
1268 * Try again after 2 seconds. Interruptable. */
1269 schedule_timeout_interruptible(
1270 cfs_time_seconds(2));
1271 if (!signal_pending(current))
1275 LCONSOLE_ERROR_MSG(0x15f,
1276 "%s: cannot register this server with the MGS: "
1277 "rc = %d. Is the MGS running?\n",
1278 lsi->lsi_svname, rc);
1280 CDEBUG(D_HA, "%s: error registering with the MGS: "
1281 "rc = %d (not fatal)\n", lsi->lsi_svname, rc);
1282 /* reset the error code for non-fatal error. */
1295 * Notify the MGS that this target is ready.
1296 * Used by IR - if the MGS receives this message, it will notify clients.
1298 static int server_notify_target(struct super_block *sb, struct obd_device *obd)
1300 struct lustre_sb_info *lsi = s2lsi(sb);
1301 struct obd_device *mgc = lsi->lsi_mgc;
1302 struct mgs_target_info *mti = NULL;
1308 if (!(IS_SERVER(lsi)))
1314 rc = server_lsi2mti(lsi, mti);
1318 mti->mti_instance = obd->u.obt.obt_instance;
1319 mti->mti_flags |= LDD_F_OPC_READY;
1321 /* FIXME use mgc_process_config instead */
1322 rc = obd_set_info_async(NULL, mgc->u.cli.cl_mgc_mgsexp,
1323 sizeof(KEY_REGISTER_TARGET),
1324 KEY_REGISTER_TARGET,
1325 sizeof(*mti), mti, NULL);
1327 /* Imperative recovery: if the mgs informs us to use IR? */
1328 if (!rc && !(mti->mti_flags & LDD_F_ERROR) &&
1329 (mti->mti_flags & LDD_F_IR_CAPABLE))
1330 lsi->lsi_flags |= LDD_F_IR_CAPABLE;
1339 /** Start server targets: MDTs and OSTs
1341 static int server_start_targets(struct super_block *sb)
1343 struct obd_device *obd;
1344 struct lustre_sb_info *lsi = s2lsi(sb);
1345 struct config_llog_instance cfg;
1346 struct lu_env mgc_env;
1347 struct lu_device *dev;
1348 char *name_service, *obd_name_service = NULL;
1349 struct obd_type *type = NULL;
1353 CDEBUG(D_MOUNT, "starting target %s\n", lsi->lsi_svname);
1355 LASSERTF(IS_MDT(lsi) || IS_OST(lsi), "designed for MDT or OST only\n");
1358 obd_name_service = LUSTRE_MDS_OBDNAME;
1359 name_service = LUSTRE_MDS_NAME;
1361 obd_name_service = LUSTRE_OSS_OBDNAME;
1362 name_service = LUSTRE_OSS_NAME;
1365 /* make sure MDS/OSS is started */
1366 mutex_lock(&server_start_lock);
1367 obd = class_name2obd(obd_name_service);
1369 rc = lustre_start_simple(obd_name_service, name_service,
1371 LUSTRE_MDS_OBDNAME"_uuid" :
1372 LUSTRE_OSS_OBDNAME"_uuid"),
1373 NULL, NULL, NULL, NULL);
1375 mutex_unlock(&server_start_lock);
1376 CERROR("failed to start %s: %d\n",
1377 obd_name_service, rc);
1381 /* hold a type reference and put it at server_stop_servers */
1382 type = class_get_type(IS_MDT(lsi) ?
1383 LUSTRE_MDT_NAME : LUSTRE_OST_NAME);
1385 mutex_unlock(&server_start_lock);
1386 GOTO(out_stop_service, rc = -ENODEV);
1388 lsi->lsi_server_started = 1;
1389 mutex_unlock(&server_start_lock);
1390 if (OBD_FAIL_PRECHECK(OBD_FAIL_OBD_STOP_MDS_RACE) &&
1392 OBD_RACE(OBD_FAIL_OBD_STOP_MDS_RACE);
1393 msleep(2 * MSEC_PER_SEC);
1396 rc = lu_env_init(&mgc_env, LCT_MG_THREAD);
1398 GOTO(out_stop_service, rc);
1400 /* Set the mgc fs to our server disk. This allows the MGC to
1401 * read and write configs locally, in case it can't talk to the MGS. */
1402 rc = server_mgc_set_fs(&mgc_env, lsi->lsi_mgc, sb);
1406 /* Register with MGS */
1407 rc = server_register_target(lsi);
1411 /* Let the target look up the mount using the target's name
1412 (we can't pass the sb or mnt through class_process_config.) */
1413 rc = server_register_mount(lsi->lsi_svname, sb);
1417 /* Start targets using the llog named for the target */
1418 memset(&cfg, 0, sizeof(cfg));
1419 cfg.cfg_callback = class_config_llog_handler;
1420 cfg.cfg_sub_clds = CONFIG_SUB_SERVER;
1421 rc = lustre_process_log(sb, lsi->lsi_svname, &cfg);
1423 CERROR("failed to start server %s: %d\n",
1424 lsi->lsi_svname, rc);
1425 /* Do NOT call server_deregister_mount() here. This makes it
1426 * impossible to find mount later in cleanup time and leaves
1427 * @lsi and othder stuff leaked. -umka */
1431 obd = class_name2obd(lsi->lsi_svname);
1433 CERROR("no server named %s was started\n", lsi->lsi_svname);
1434 GOTO(out_mgc, rc = -ENXIO);
1437 if (IS_OST(lsi) || IS_MDT(lsi)) {
1438 rc = lustre_start_lwp(sb);
1440 CERROR("%s: failed to start LWP: %d\n",
1441 lsi->lsi_svname, rc);
1446 server_notify_target(sb, obd);
1448 /* calculate recovery timeout, do it after lustre_process_log */
1449 server_calc_timeout(lsi, obd);
1451 /* log has been fully processed, let clients connect */
1452 dev = obd->obd_lu_dev;
1453 if (dev && dev->ld_ops->ldo_prepare) {
1456 rc = lu_env_init(&env, dev->ld_type->ldt_ctx_tags);
1458 struct lu_context session_ctx;
1460 lu_context_init(&session_ctx, LCT_SERVER_SESSION);
1461 session_ctx.lc_thread = NULL;
1462 lu_context_enter(&session_ctx);
1463 env.le_ses = &session_ctx;
1465 rc = dev->ld_ops->ldo_prepare(&env, NULL, dev);
1468 lu_context_exit(&session_ctx);
1469 lu_context_fini(&session_ctx);
1473 /* abort recovery only on the complete stack:
1474 * many devices can be involved */
1475 if ((lsi->lsi_lmd->lmd_flags &
1476 (LMD_FLG_ABORT_RECOV | LMD_FLG_ABORT_RECOV_MDT)) &&
1477 (OBP(obd, iocontrol))) {
1478 struct obd_ioctl_data karg = {
1479 .ioc_type = lsi->lsi_lmd->lmd_flags,
1482 obd_iocontrol(OBD_IOC_ABORT_RECOVERY, obd->obd_self_export, 0,
1487 /* Release the mgc fs for others to use */
1488 server_mgc_clear_fs(&mgc_env, lsi->lsi_mgc);
1490 lu_env_fini(&mgc_env);
1492 /* in case of error upper function call
1493 * server_put_super->server_stop_servers()
1499 static int lsi_prepare(struct lustre_sb_info *lsi)
1501 const char *osd_type;
1508 LASSERT(lsi->lsi_lmd);
1510 /* The server name is given as a mount line option */
1511 if (lsi->lsi_lmd->lmd_profile == NULL) {
1512 LCONSOLE_ERROR("Can't determine server name\n");
1516 /* Determine osd type */
1517 if (lsi->lsi_lmd->lmd_osd_type == NULL) {
1518 osd_type = LUSTRE_OSD_LDISKFS_NAME;
1521 osd_type = lsi->lsi_lmd->lmd_osd_type;
1522 fstype = lsi->lsi_lmd->lmd_osd_type;
1525 if (strlen(lsi->lsi_lmd->lmd_profile) >= sizeof(lsi->lsi_svname) ||
1526 strlen(osd_type) >= sizeof(lsi->lsi_osd_type) ||
1527 strlen(fstype) >= sizeof(lsi->lsi_fstype))
1528 RETURN(-ENAMETOOLONG);
1530 strlcpy(lsi->lsi_svname, lsi->lsi_lmd->lmd_profile,
1531 sizeof(lsi->lsi_svname));
1532 strlcpy(lsi->lsi_osd_type, osd_type, sizeof(lsi->lsi_osd_type));
1533 /* XXX: a temp. solution for components using ldiskfs
1534 * to be removed in one of the subsequent patches */
1535 strlcpy(lsi->lsi_fstype, fstype, sizeof(lsi->lsi_fstype));
1537 /* Determine server type */
1538 rc = server_name2index(lsi->lsi_svname, &index, NULL);
1540 if (lsi->lsi_lmd->lmd_flags & LMD_FLG_MGS) {
1541 /* Assume we're a bare MGS */
1543 lsi->lsi_lmd->lmd_flags |= LMD_FLG_NOSVC;
1545 LCONSOLE_ERROR("Can't determine server type of '%s'\n",
1550 lsi->lsi_flags |= rc;
1552 /* Add mount line flags that used to be in ldd:
1553 * writeconf, mgs, anything else?
1555 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_WRITECONF) ?
1556 LDD_F_WRITECONF : 0;
1557 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_NO_LOCAL_LOGS) ?
1558 LDD_F_NO_LOCAL_LOGS : 0;
1559 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_VIRGIN) ?
1561 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_UPDATE) ?
1563 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_MGS) ?
1564 LDD_F_SV_TYPE_MGS : 0;
1565 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_NO_PRIMNODE) ?
1566 LDD_F_NO_PRIMNODE : 0;
1571 /*************** server mount ******************/
1573 /** Start the shutdown of servers at umount.
1575 static void server_put_super(struct super_block *sb)
1577 struct lustre_sb_info *lsi = s2lsi(sb);
1578 struct obd_device *obd;
1579 char *tmpname, *extraname = NULL;
1581 int lsiflags = lsi->lsi_flags;
1582 bool stop_servers = lsi->lsi_server_started;
1585 LASSERT(IS_SERVER(lsi));
1587 tmpname_sz = strlen(lsi->lsi_svname) + 1;
1588 OBD_ALLOC(tmpname, tmpname_sz);
1589 memcpy(tmpname, lsi->lsi_svname, tmpname_sz);
1590 CDEBUG(D_MOUNT, "server put_super %s\n", tmpname);
1591 if (IS_MDT(lsi) && (lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC))
1592 snprintf(tmpname, tmpname_sz, "MGS");
1594 /* disconnect the lwp first to drain off the inflight request */
1595 if (IS_OST(lsi) || IS_MDT(lsi)) {
1598 rc = lustre_disconnect_lwp(sb);
1599 if (rc != 0 && rc != -ETIMEDOUT &&
1600 rc != -ENOTCONN && rc != -ESHUTDOWN)
1601 CWARN("%s: failed to disconnect lwp: rc= %d\n",
1605 /* Stop the target */
1606 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1607 (IS_MDT(lsi) || IS_OST(lsi))) {
1608 struct lustre_profile *lprof = NULL;
1610 /* tell the mgc to drop the config log */
1611 lustre_end_log(sb, lsi->lsi_svname, NULL);
1613 /* COMPAT_146 - profile may get deleted in mgc_cleanup.
1614 If there are any setup/cleanup errors, save the lov
1615 name for safety cleanup later. */
1616 lprof = class_get_profile(lsi->lsi_svname);
1617 if (lprof != NULL) {
1618 if (lprof->lp_dt != NULL) {
1619 OBD_ALLOC(extraname, strlen(lprof->lp_dt) + 1);
1620 strncpy(extraname, lprof->lp_dt,
1621 strlen(lprof->lp_dt) + 1);
1623 class_put_profile(lprof);
1626 obd = class_name2obd(lsi->lsi_svname);
1628 CDEBUG(D_MOUNT, "stopping %s\n", obd->obd_name);
1629 if (lsiflags & LSI_UMOUNT_FAILOVER)
1631 /* We can't seem to give an error return code
1632 * to .put_super, so we better make sure we clean up! */
1634 class_manual_cleanup(obd);
1635 if (OBD_FAIL_PRECHECK(OBD_FAIL_OBD_STOP_MDS_RACE)) {
1637 server_name2index(lsi->lsi_svname, &idx, NULL);
1638 /* sleeping for MDT0001 */
1640 OBD_RACE(OBD_FAIL_OBD_STOP_MDS_RACE);
1643 CERROR("no obd %s\n", lsi->lsi_svname);
1644 server_deregister_mount(lsi->lsi_svname);
1648 /* If they wanted the mgs to stop separately from the mdt, they
1649 should have put it on a different device. */
1651 /* if MDS start with --nomgs, don't stop MGS then */
1652 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS))
1653 server_stop_mgs(sb);
1656 if (IS_OST(lsi) || IS_MDT(lsi)) {
1657 if (lustre_stop_lwp(sb) < 0)
1658 CERROR("%s: failed to stop lwp!\n", tmpname);
1661 /* Clean the mgc and sb */
1662 lustre_common_put_super(sb);
1664 /* wait till all in-progress cleanups are done
1665 * specifically we're interested in ofd cleanup
1667 obd_zombie_barrier();
1669 /* Stop the servers (MDS, OSS) if no longer needed. We must wait
1670 until the target is really gone so that our type refcount check
1673 server_stop_servers(lsiflags);
1675 /* In case of startup or cleanup err, stop related obds */
1677 obd = class_name2obd(extraname);
1679 CWARN("Cleaning orphaned obd %s\n", extraname);
1681 class_manual_cleanup(obd);
1683 OBD_FREE(extraname, strlen(extraname) + 1);
1686 LCONSOLE_WARN("server umount %s complete\n", tmpname);
1687 OBD_FREE(tmpname, tmpname_sz);
1691 /** Called only for 'umount -f'
1693 static void server_umount_begin(struct super_block *sb)
1695 struct lustre_sb_info *lsi = s2lsi(sb);
1698 CDEBUG(D_MOUNT, "umount -f\n");
1699 /* umount = failover
1701 no third way to do non-force, non-failover */
1702 lsi->lsi_flags &= ~LSI_UMOUNT_FAILOVER;
1706 static int server_statfs(struct dentry *dentry, struct kstatfs *buf)
1708 struct super_block *sb = dentry->d_sb;
1709 struct lustre_sb_info *lsi = s2lsi(sb);
1710 struct obd_statfs statfs;
1714 if (lsi->lsi_dt_dev) {
1715 rc = dt_statfs(NULL, lsi->lsi_dt_dev, &statfs);
1717 statfs_unpack(buf, &statfs);
1718 buf->f_type = sb->s_magic;
1724 buf->f_type = sb->s_magic;
1725 buf->f_bsize = sb->s_blocksize;
1731 buf->f_namelen = NAME_MAX;
1735 int server_show_options(struct seq_file *seq, struct dentry *dentry)
1737 struct lustre_sb_info *lsi;
1738 struct lustre_mount_data *lmd;
1740 LASSERT(seq != NULL && dentry != NULL);
1741 lsi = s2lsi(dentry->d_sb);
1743 seq_printf(seq, ",svname=%s", lmd->lmd_profile);
1745 if (lmd->lmd_flags & LMD_FLG_ABORT_RECOV)
1746 seq_puts(seq, ",abort_recov");
1748 if (lmd->lmd_flags & LMD_FLG_NOIR)
1749 seq_puts(seq, ",noir");
1751 if (lmd->lmd_flags & LMD_FLG_NOSVC)
1752 seq_puts(seq, ",nosvc");
1754 if (lmd->lmd_flags & LMD_FLG_NOMGS)
1755 seq_puts(seq, ",nomgs");
1757 if (lmd->lmd_flags & LMD_FLG_NOSCRUB)
1758 seq_puts(seq, ",noscrub");
1759 if (lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
1760 seq_puts(seq, ",skip_lfsck");
1762 if (lmd->lmd_flags & LMD_FLG_DEV_RDONLY)
1763 seq_puts(seq, ",rdonly_dev");
1765 if (lmd->lmd_flags & LMD_FLG_MGS)
1766 seq_puts(seq, ",mgs");
1768 if (lmd->lmd_mgs != NULL)
1769 seq_printf(seq, ",mgsnode=%s", lmd->lmd_mgs);
1771 if (lmd->lmd_osd_type != NULL)
1772 seq_printf(seq, ",osd=%s", lmd->lmd_osd_type);
1774 if (lmd->lmd_opts != NULL) {
1776 seq_puts(seq, lmd->lmd_opts);
1782 /** The operations we support directly on the superblock:
1783 * mount, umount, and df.
1785 static const struct super_operations server_ops = {
1786 .put_super = server_put_super,
1787 .umount_begin = server_umount_begin, /* umount -f */
1788 .statfs = server_statfs,
1789 .show_options = server_show_options,
1793 * Xattr support for Lustre servers
1795 #ifdef HAVE_IOP_XATTR
1796 static ssize_t lustre_getxattr(struct dentry *dentry, const char *name,
1797 void *buffer, size_t size)
1799 if (!selinux_is_enabled())
1804 static int lustre_setxattr(struct dentry *dentry, const char *name,
1805 const void *value, size_t size, int flags)
1811 static ssize_t lustre_listxattr(struct dentry *d_entry, char *name,
1817 static bool is_cmd_supported(unsigned int command)
1829 static long server_ioctl(struct file *filp, unsigned int command,
1832 struct file active_filp;
1833 struct inode *inode = file_inode(filp);
1834 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
1835 struct super_block *dd_sb = dt_mnt_sb_get(lsi->lsi_dt_dev);
1836 struct inode *active_inode;
1837 int err = -EOPNOTSUPP;
1839 if (IS_ERR(dd_sb) || !is_cmd_supported(command))
1842 active_inode = igrab(dd_sb->s_root->d_inode);
1846 active_filp.f_inode = active_inode;
1847 if (active_inode->i_fop && active_inode->i_fop->unlocked_ioctl)
1848 err = active_inode->i_fop->unlocked_ioctl(&active_filp,
1854 static const struct inode_operations server_inode_operations = {
1855 #ifdef HAVE_IOP_XATTR
1856 .setxattr = lustre_setxattr,
1857 .getxattr = lustre_getxattr,
1859 .listxattr = lustre_listxattr,
1862 static const struct file_operations server_file_operations = {
1863 .unlocked_ioctl = server_ioctl,
1866 #define log2(n) ffz(~(n))
1867 #define LUSTRE_SUPER_MAGIC 0x0BD00BD1
1869 static int server_fill_super_common(struct super_block *sb)
1871 struct inode *root = NULL;
1874 CDEBUG(D_MOUNT, "Server sb, dev=%d\n", (int)sb->s_dev);
1876 sb->s_blocksize = 4096;
1877 sb->s_blocksize_bits = log2(sb->s_blocksize);
1878 sb->s_magic = LUSTRE_SUPER_MAGIC;
1879 sb->s_maxbytes = 0; /* we don't allow file IO on server mountpoints */
1880 sb->s_flags |= SB_RDONLY;
1881 sb->s_op = &server_ops;
1883 root = new_inode(sb);
1885 CERROR("Can't make root inode\n");
1889 /* returns -EIO for every operation */
1890 /* make_bad_inode(root); -- badness - can't umount */
1891 /* apparently we need to be a directory for the mount to finish */
1892 root->i_mode = S_IFDIR;
1893 root->i_op = &server_inode_operations;
1894 root->i_fop = &server_file_operations;
1895 sb->s_root = d_make_root(root);
1897 CERROR("%s: can't make root dentry\n", sb->s_id);
1904 static int osd_start(struct lustre_sb_info *lsi, unsigned long mflags)
1906 struct lustre_mount_data *lmd = lsi->lsi_lmd;
1907 struct obd_device *obd;
1908 struct dt_device_param p;
1909 char flagstr[20 + 1 + 10 + 1];
1914 "Attempting to start %s, type=%s, lsifl=%x, mountfl=%lx\n",
1915 lsi->lsi_svname, lsi->lsi_osd_type, lsi->lsi_flags, mflags);
1917 sprintf(lsi->lsi_osd_obdname, "%s-osd", lsi->lsi_svname);
1918 strcpy(lsi->lsi_osd_uuid, lsi->lsi_osd_obdname);
1919 strcat(lsi->lsi_osd_uuid, "_UUID");
1920 snprintf(flagstr, sizeof(flagstr), "%lu:%u", mflags, lmd->lmd_flags);
1922 obd = class_name2obd(lsi->lsi_osd_obdname);
1924 rc = lustre_start_simple(lsi->lsi_osd_obdname,
1926 lsi->lsi_osd_uuid, lmd->lmd_dev,
1927 flagstr, lsi->lsi_lmd->lmd_opts,
1931 obd = class_name2obd(lsi->lsi_osd_obdname);
1934 CDEBUG(D_MOUNT, "%s already started\n", lsi->lsi_osd_obdname);
1935 /* but continue setup to allow special case of MDT and internal
1936 * MGT being started separately. */
1937 if (!((IS_MGS(lsi) && (lsi->lsi_lmd->lmd_flags &
1939 (IS_MDT(lsi) && (lsi->lsi_lmd->lmd_flags &
1944 rc = obd_connect(NULL, &lsi->lsi_osd_exp,
1945 obd, &obd->obd_uuid, NULL, NULL);
1949 class_manual_cleanup(obd);
1950 lsi->lsi_dt_dev = NULL;
1954 LASSERT(obd->obd_lu_dev);
1955 lu_device_get(obd->obd_lu_dev);
1956 lsi->lsi_dt_dev = lu2dt_dev(obd->obd_lu_dev);
1957 LASSERT(lsi->lsi_dt_dev);
1959 /* set disk context for llog usage */
1960 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1961 obd->obd_lvfs_ctxt.dt = lsi->lsi_dt_dev;
1963 dt_conf_get(NULL, lsi->lsi_dt_dev, &p);
1968 /** Fill in the superblock info for a Lustre server.
1969 * Mount the device with the correct options.
1970 * Read the on-disk config file.
1971 * Start the services.
1973 int server_fill_super(struct super_block *sb)
1975 struct lustre_sb_info *lsi = s2lsi(sb);
1979 /* to simulate target mount race */
1980 OBD_RACE(OBD_FAIL_TGT_MOUNT_RACE);
1982 rc = lsi_prepare(lsi);
1988 /* Start low level OSD */
1989 rc = osd_start(lsi, sb->s_flags);
1991 CERROR("Unable to start osd on %s: %d\n",
1992 lsi->lsi_lmd->lmd_dev, rc);
1997 CDEBUG(D_MOUNT, "Found service %s on device %s\n",
1998 lsi->lsi_svname, lsi->lsi_lmd->lmd_dev);
2000 if (class_name2obd(lsi->lsi_svname)) {
2001 LCONSOLE_ERROR_MSG(0x161, "The target named %s is already "
2002 "running. Double-mount may have compromised"
2003 " the disk journal.\n",
2009 /* Start MGS before MGC */
2010 if (IS_MGS(lsi) && !(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS)) {
2011 rc = server_start_mgs(sb);
2016 /* Start MGC before servers */
2017 rc = lustre_start_mgc(sb);
2021 /* Set up all obd devices for service */
2022 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
2023 (IS_OST(lsi) || IS_MDT(lsi))) {
2024 rc = server_start_targets(sb);
2026 CERROR("Unable to start targets: %d\n", rc);
2029 /* FIXME overmount client here, or can we just start a
2030 * client log and client_fill_super on this sb? We
2031 * need to make sure server_put_super gets called too
2032 * - ll_put_super calls lustre_common_put_super; check
2033 * there for LSI_SERVER flag, call s_p_s if so.
2035 * Probably should start client from new thread so we
2036 * can return. Client will not finish until all
2037 * servers are connected. Note - MGS-only server does
2038 * NOT get a client, since there is no lustre fs
2039 * associated - the MGS is for all lustre fs's */
2042 rc = server_fill_super_common(sb);
2048 /* We jump here in case of failure while starting targets or MGS.
2049 * In this case we can't just put @mnt and have to do real cleanup
2050 * with stoping targets, etc. */
2051 server_put_super(sb);
2054 EXPORT_SYMBOL(server_fill_super);
2057 * Calculate timeout value for a target.
2059 void server_calc_timeout(struct lustre_sb_info *lsi, struct obd_device *obd)
2061 struct lustre_mount_data *lmd;
2065 bool has_ir = !!(lsi->lsi_flags & LDD_F_IR_CAPABLE);
2066 int min = OBD_RECOVERY_TIME_MIN;
2068 LASSERT(IS_SERVER(lsi));
2072 soft = lmd->lmd_recovery_time_soft;
2073 hard = lmd->lmd_recovery_time_hard;
2074 has_ir = has_ir && !(lmd->lmd_flags & LMD_FLG_NOIR);
2075 obd->obd_no_ir = !has_ir;
2079 soft = OBD_RECOVERY_TIME_SOFT;
2081 hard = OBD_RECOVERY_TIME_HARD;
2083 /* target may have ir_factor configured. */
2084 factor = OBD_IR_FACTOR_DEFAULT;
2085 if (obd->obd_recovery_ir_factor)
2086 factor = obd->obd_recovery_ir_factor;
2089 int new_soft = soft;
2091 /* adjust timeout value by imperative recovery */
2092 new_soft = (soft * factor) / OBD_IR_FACTOR_MAX;
2093 /* make sure the timeout is not too short */
2094 new_soft = max(min, new_soft);
2096 LCONSOLE_INFO("%s: Imperative Recovery enabled, recovery "
2097 "window shrunk from %d-%d down to %d-%d\n",
2098 obd->obd_name, soft, hard, new_soft, hard);
2102 LCONSOLE_INFO("%s: Imperative Recovery not enabled, recovery "
2103 "window %d-%d\n", obd->obd_name, soft, hard);
2107 obd->obd_recovery_timeout = max_t(time64_t, obd->obd_recovery_timeout,
2109 obd->obd_recovery_time_hard = hard;
2110 obd->obd_recovery_ir_factor = factor;