4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mgc/mgc_request.c
38 * Author: Nathan Rutman <nathan@clusterfs.com>
41 #define DEBUG_SUBSYSTEM S_MGC
42 #define D_MGC D_CONFIG /*|D_WARNING*/
44 #include <linux/module.h>
45 #include "../include/obd_class.h"
46 #include "../include/lustre_dlm.h"
47 #include "../include/lprocfs_status.h"
48 #include "../include/lustre_log.h"
49 #include "../include/lustre_disk.h"
50 #include "../include/dt_object.h"
52 #include "mgc_internal.h"
54 static int mgc_name2resid(char *name
, int len
, struct ldlm_res_id
*res_id
,
59 if (len
> sizeof(resname
)) {
60 CERROR("name too long: %s\n", name
);
64 CERROR("missing name: %s\n", name
);
67 memcpy(&resname
, name
, len
);
69 /* Always use the same endianness for the resid */
70 memset(res_id
, 0, sizeof(*res_id
));
71 res_id
->name
[0] = cpu_to_le64(resname
);
72 /* XXX: unfortunately, sptlprc and config llog share one lock */
75 case CONFIG_T_SPTLRPC
:
78 case CONFIG_T_RECOVER
:
85 res_id
->name
[1] = cpu_to_le64(resname
);
86 CDEBUG(D_MGC
, "log %s to resid %#llx/%#llx (%.8s)\n", name
,
87 res_id
->name
[0], res_id
->name
[1], (char *)&res_id
->name
[0]);
91 int mgc_fsname2resid(char *fsname
, struct ldlm_res_id
*res_id
, int type
)
93 /* fsname is at most 8 chars long, maybe contain "-".
94 * e.g. "lustre", "SUN-000" */
95 return mgc_name2resid(fsname
, strlen(fsname
), res_id
, type
);
97 EXPORT_SYMBOL(mgc_fsname2resid
);
99 int mgc_logname2resid(char *logname
, struct ldlm_res_id
*res_id
, int type
)
104 /* logname consists of "fsname-nodetype".
105 * e.g. "lustre-MDT0001", "SUN-000-client"
106 * there is an exception: llog "params" */
107 name_end
= strrchr(logname
, '-');
109 len
= strlen(logname
);
111 len
= name_end
- logname
;
112 return mgc_name2resid(logname
, len
, res_id
, type
);
115 /********************** config llog list **********************/
116 static LIST_HEAD(config_llog_list
);
117 static DEFINE_SPINLOCK(config_list_lock
);
119 /* Take a reference to a config log */
120 static int config_log_get(struct config_llog_data
*cld
)
122 atomic_inc(&cld
->cld_refcount
);
123 CDEBUG(D_INFO
, "log %s refs %d\n", cld
->cld_logname
,
124 atomic_read(&cld
->cld_refcount
));
128 /* Drop a reference to a config log. When no longer referenced,
129 we can free the config log data */
130 static void config_log_put(struct config_llog_data
*cld
)
132 CDEBUG(D_INFO
, "log %s refs %d\n", cld
->cld_logname
,
133 atomic_read(&cld
->cld_refcount
));
134 LASSERT(atomic_read(&cld
->cld_refcount
) > 0);
136 /* spinlock to make sure no item with 0 refcount in the list */
137 if (atomic_dec_and_lock(&cld
->cld_refcount
, &config_list_lock
)) {
138 list_del(&cld
->cld_list_chain
);
139 spin_unlock(&config_list_lock
);
141 CDEBUG(D_MGC
, "dropping config log %s\n", cld
->cld_logname
);
143 if (cld
->cld_recover
)
144 config_log_put(cld
->cld_recover
);
145 if (cld
->cld_sptlrpc
)
146 config_log_put(cld
->cld_sptlrpc
);
148 config_log_put(cld
->cld_params
);
149 if (cld_is_sptlrpc(cld
))
150 sptlrpc_conf_log_stop(cld
->cld_logname
);
152 class_export_put(cld
->cld_mgcexp
);
153 OBD_FREE(cld
, sizeof(*cld
) + strlen(cld
->cld_logname
) + 1);
157 /* Find a config log by name */
159 struct config_llog_data
*config_log_find(char *logname
,
160 struct config_llog_instance
*cfg
)
162 struct config_llog_data
*cld
;
163 struct config_llog_data
*found
= NULL
;
166 LASSERT(logname
!= NULL
);
168 instance
= cfg
? cfg
->cfg_instance
: NULL
;
169 spin_lock(&config_list_lock
);
170 list_for_each_entry(cld
, &config_llog_list
, cld_list_chain
) {
171 /* check if instance equals */
172 if (instance
!= cld
->cld_cfg
.cfg_instance
)
175 /* instance may be NULL, should check name */
176 if (strcmp(logname
, cld
->cld_logname
) == 0) {
182 atomic_inc(&found
->cld_refcount
);
183 LASSERT(found
->cld_stopping
== 0 || cld_is_sptlrpc(found
) == 0);
185 spin_unlock(&config_list_lock
);
190 struct config_llog_data
*do_config_log_add(struct obd_device
*obd
,
193 struct config_llog_instance
*cfg
,
194 struct super_block
*sb
)
196 struct config_llog_data
*cld
;
199 CDEBUG(D_MGC
, "do adding config log %s:%p\n", logname
,
200 cfg
? cfg
->cfg_instance
: 0);
202 OBD_ALLOC(cld
, sizeof(*cld
) + strlen(logname
) + 1);
204 return ERR_PTR(-ENOMEM
);
206 strcpy(cld
->cld_logname
, logname
);
210 cld
->cld_cfg
.cfg_callback
= class_config_llog_handler
;
211 mutex_init(&cld
->cld_lock
);
212 cld
->cld_cfg
.cfg_last_idx
= 0;
213 cld
->cld_cfg
.cfg_flags
= 0;
214 cld
->cld_cfg
.cfg_sb
= sb
;
215 cld
->cld_type
= type
;
216 atomic_set(&cld
->cld_refcount
, 1);
218 /* Keep the mgc around until we are done */
219 cld
->cld_mgcexp
= class_export_get(obd
->obd_self_export
);
221 if (cld_is_sptlrpc(cld
)) {
222 sptlrpc_conf_log_start(logname
);
223 cld
->cld_cfg
.cfg_obdname
= obd
->obd_name
;
226 rc
= mgc_logname2resid(logname
, &cld
->cld_resid
, type
);
228 spin_lock(&config_list_lock
);
229 list_add(&cld
->cld_list_chain
, &config_llog_list
);
230 spin_unlock(&config_list_lock
);
237 if (cld_is_sptlrpc(cld
)) {
238 rc
= mgc_process_log(obd
, cld
);
239 if (rc
&& rc
!= -ENOENT
)
240 CERROR("failed processing sptlrpc log: %d\n", rc
);
246 static struct config_llog_data
*config_recover_log_add(struct obd_device
*obd
,
248 struct config_llog_instance
*cfg
,
249 struct super_block
*sb
)
251 struct config_llog_instance lcfg
= *cfg
;
252 struct lustre_sb_info
*lsi
= s2lsi(sb
);
253 struct config_llog_data
*cld
;
259 /* for osp-on-ost, see lustre_start_osp() */
260 if (IS_MDT(lsi
) && lcfg
.cfg_instance
)
263 /* we have to use different llog for clients and mdts for cmd
264 * where only clients are notified if one of cmd server restarts */
265 LASSERT(strlen(fsname
) < sizeof(logname
) / 2);
266 strcpy(logname
, fsname
);
267 if (IS_SERVER(lsi
)) { /* mdt */
268 LASSERT(lcfg
.cfg_instance
== NULL
);
269 lcfg
.cfg_instance
= sb
;
270 strcat(logname
, "-mdtir");
272 LASSERT(lcfg
.cfg_instance
!= NULL
);
273 strcat(logname
, "-cliir");
276 cld
= do_config_log_add(obd
, logname
, CONFIG_T_RECOVER
, &lcfg
, sb
);
280 static struct config_llog_data
*config_params_log_add(struct obd_device
*obd
,
281 struct config_llog_instance
*cfg
, struct super_block
*sb
)
283 struct config_llog_instance lcfg
= *cfg
;
284 struct config_llog_data
*cld
;
286 lcfg
.cfg_instance
= sb
;
288 cld
= do_config_log_add(obd
, PARAMS_FILENAME
, CONFIG_T_PARAMS
,
294 /** Add this log to the list of active logs watched by an MGC.
295 * Active means we're watching for updates.
296 * We have one active log per "mount" - client instance or servername.
297 * Each instance may be at a different point in the log.
299 static int config_log_add(struct obd_device
*obd
, char *logname
,
300 struct config_llog_instance
*cfg
,
301 struct super_block
*sb
)
303 struct lustre_sb_info
*lsi
= s2lsi(sb
);
304 struct config_llog_data
*cld
;
305 struct config_llog_data
*sptlrpc_cld
;
306 struct config_llog_data
*params_cld
;
311 CDEBUG(D_MGC
, "adding config log %s:%p\n", logname
, cfg
->cfg_instance
);
314 * for each regular log, the depended sptlrpc log name is
315 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
317 ptr
= strrchr(logname
, '-');
318 if (ptr
== NULL
|| ptr
- logname
> 8) {
319 CERROR("logname %s is too long\n", logname
);
323 memcpy(seclogname
, logname
, ptr
- logname
);
324 strcpy(seclogname
+ (ptr
- logname
), "-sptlrpc");
326 sptlrpc_cld
= config_log_find(seclogname
, NULL
);
327 if (sptlrpc_cld
== NULL
) {
328 sptlrpc_cld
= do_config_log_add(obd
, seclogname
,
329 CONFIG_T_SPTLRPC
, NULL
, NULL
);
330 if (IS_ERR(sptlrpc_cld
)) {
331 CERROR("can't create sptlrpc log: %s\n", seclogname
);
332 GOTO(out_err
, rc
= PTR_ERR(sptlrpc_cld
));
335 params_cld
= config_params_log_add(obd
, cfg
, sb
);
336 if (IS_ERR(params_cld
)) {
337 rc
= PTR_ERR(params_cld
);
338 CERROR("%s: can't create params log: rc = %d\n",
343 cld
= do_config_log_add(obd
, logname
, CONFIG_T_CONFIG
, cfg
, sb
);
345 CERROR("can't create log: %s\n", logname
);
346 GOTO(out_err2
, rc
= PTR_ERR(cld
));
349 cld
->cld_sptlrpc
= sptlrpc_cld
;
350 cld
->cld_params
= params_cld
;
352 LASSERT(lsi
->lsi_lmd
);
353 if (!(lsi
->lsi_lmd
->lmd_flags
& LMD_FLG_NOIR
)) {
354 struct config_llog_data
*recover_cld
;
355 *strrchr(seclogname
, '-') = 0;
356 recover_cld
= config_recover_log_add(obd
, seclogname
, cfg
, sb
);
357 if (IS_ERR(recover_cld
))
358 GOTO(out_err3
, rc
= PTR_ERR(recover_cld
));
359 cld
->cld_recover
= recover_cld
;
368 config_log_put(params_cld
);
371 config_log_put(sptlrpc_cld
);
377 DEFINE_MUTEX(llog_process_lock
);
379 /** Stop watching for updates on this log.
381 static int config_log_end(char *logname
, struct config_llog_instance
*cfg
)
383 struct config_llog_data
*cld
;
384 struct config_llog_data
*cld_sptlrpc
= NULL
;
385 struct config_llog_data
*cld_params
= NULL
;
386 struct config_llog_data
*cld_recover
= NULL
;
389 cld
= config_log_find(logname
, cfg
);
393 mutex_lock(&cld
->cld_lock
);
395 * if cld_stopping is set, it means we didn't start the log thus
396 * not owning the start ref. this can happen after previous umount:
397 * the cld still hanging there waiting for lock cancel, and we
398 * remount again but failed in the middle and call log_end without
401 if (unlikely(cld
->cld_stopping
)) {
402 mutex_unlock(&cld
->cld_lock
);
403 /* drop the ref from the find */
408 cld
->cld_stopping
= 1;
410 cld_recover
= cld
->cld_recover
;
411 cld
->cld_recover
= NULL
;
412 mutex_unlock(&cld
->cld_lock
);
415 mutex_lock(&cld_recover
->cld_lock
);
416 cld_recover
->cld_stopping
= 1;
417 mutex_unlock(&cld_recover
->cld_lock
);
418 config_log_put(cld_recover
);
421 spin_lock(&config_list_lock
);
422 cld_sptlrpc
= cld
->cld_sptlrpc
;
423 cld
->cld_sptlrpc
= NULL
;
424 cld_params
= cld
->cld_params
;
425 cld
->cld_params
= NULL
;
426 spin_unlock(&config_list_lock
);
429 config_log_put(cld_sptlrpc
);
432 mutex_lock(&cld_params
->cld_lock
);
433 cld_params
->cld_stopping
= 1;
434 mutex_unlock(&cld_params
->cld_lock
);
435 config_log_put(cld_params
);
438 /* drop the ref from the find */
440 /* drop the start ref */
443 CDEBUG(D_MGC
, "end config log %s (%d)\n", logname
? logname
: "client",
449 int lprocfs_mgc_rd_ir_state(struct seq_file
*m
, void *data
)
451 struct obd_device
*obd
= data
;
452 struct obd_import
*imp
= obd
->u
.cli
.cl_import
;
453 struct obd_connect_data
*ocd
= &imp
->imp_connect_data
;
454 struct config_llog_data
*cld
;
456 seq_printf(m
, "imperative_recovery: %s\n",
457 OCD_HAS_FLAG(ocd
, IMP_RECOV
) ? "ENABLED" : "DISABLED");
458 seq_printf(m
, "client_state:\n");
460 spin_lock(&config_list_lock
);
461 list_for_each_entry(cld
, &config_llog_list
, cld_list_chain
) {
462 if (cld
->cld_recover
== NULL
)
464 seq_printf(m
, " - { client: %s, nidtbl_version: %u }\n",
466 cld
->cld_recover
->cld_cfg
.cfg_last_idx
);
468 spin_unlock(&config_list_lock
);
474 /* reenqueue any lost locks */
475 #define RQ_RUNNING 0x1
479 static int rq_state
= 0;
480 static wait_queue_head_t rq_waitq
;
481 static DECLARE_COMPLETION(rq_exit
);
483 static void do_requeue(struct config_llog_data
*cld
)
485 LASSERT(atomic_read(&cld
->cld_refcount
) > 0);
487 /* Do not run mgc_process_log on a disconnected export or an
488 export which is being disconnected. Take the client
489 semaphore to make the check non-racy. */
490 down_read(&cld
->cld_mgcexp
->exp_obd
->u
.cli
.cl_sem
);
491 if (cld
->cld_mgcexp
->exp_obd
->u
.cli
.cl_conn_count
!= 0) {
492 CDEBUG(D_MGC
, "updating log %s\n", cld
->cld_logname
);
493 mgc_process_log(cld
->cld_mgcexp
->exp_obd
, cld
);
495 CDEBUG(D_MGC
, "disconnecting, won't update log %s\n",
498 up_read(&cld
->cld_mgcexp
->exp_obd
->u
.cli
.cl_sem
);
501 /* this timeout represents how many seconds MGC should wait before
502 * requeue config and recover lock to the MGS. We need to randomize this
503 * in order to not flood the MGS.
505 #define MGC_TIMEOUT_MIN_SECONDS 5
506 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
508 static int mgc_requeue_thread(void *data
)
512 CDEBUG(D_MGC
, "Starting requeue thread\n");
514 /* Keep trying failed locks periodically */
515 spin_lock(&config_list_lock
);
516 rq_state
|= RQ_RUNNING
;
518 struct l_wait_info lwi
;
519 struct config_llog_data
*cld
, *cld_prev
;
520 int rand
= cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC
;
521 int stopped
= !!(rq_state
& RQ_STOP
);
524 /* Any new or requeued lostlocks will change the state */
525 rq_state
&= ~(RQ_NOW
| RQ_LATER
);
526 spin_unlock(&config_list_lock
);
528 /* Always wait a few seconds to allow the server who
529 caused the lock revocation to finish its setup, plus some
530 random so everyone doesn't try to reconnect at once. */
531 to
= MGC_TIMEOUT_MIN_SECONDS
* HZ
;
532 to
+= rand
* HZ
/ 100; /* rand is centi-seconds */
533 lwi
= LWI_TIMEOUT(to
, NULL
, NULL
);
534 l_wait_event(rq_waitq
, rq_state
& RQ_STOP
, &lwi
);
537 * iterate & processing through the list. for each cld, process
538 * its depending sptlrpc cld firstly (if any) and then itself.
540 * it's guaranteed any item in the list must have
541 * reference > 0; and if cld_lostlock is set, at
542 * least one reference is taken by the previous enqueue.
546 spin_lock(&config_list_lock
);
547 list_for_each_entry(cld
, &config_llog_list
,
549 if (!cld
->cld_lostlock
)
552 spin_unlock(&config_list_lock
);
554 LASSERT(atomic_read(&cld
->cld_refcount
) > 0);
556 /* Whether we enqueued again or not in mgc_process_log,
557 * we're done with the ref from the old enqueue */
559 config_log_put(cld_prev
);
562 cld
->cld_lostlock
= 0;
563 if (likely(!stopped
))
566 spin_lock(&config_list_lock
);
568 spin_unlock(&config_list_lock
);
570 config_log_put(cld_prev
);
572 /* break after scanning the list so that we can drop
573 * refcount to losing lock clds */
574 if (unlikely(stopped
)) {
575 spin_lock(&config_list_lock
);
579 /* Wait a bit to see if anyone else needs a requeue */
580 lwi
= (struct l_wait_info
) { 0 };
581 l_wait_event(rq_waitq
, rq_state
& (RQ_NOW
| RQ_STOP
),
583 spin_lock(&config_list_lock
);
585 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
586 rq_state
&= ~RQ_RUNNING
;
587 spin_unlock(&config_list_lock
);
591 CDEBUG(D_MGC
, "Ending requeue thread\n");
595 /* Add a cld to the list to requeue. Start the requeue thread if needed.
596 We are responsible for dropping the config log reference from here on out. */
597 static void mgc_requeue_add(struct config_llog_data
*cld
)
599 CDEBUG(D_INFO
, "log %s: requeue (r=%d sp=%d st=%x)\n",
600 cld
->cld_logname
, atomic_read(&cld
->cld_refcount
),
601 cld
->cld_stopping
, rq_state
);
602 LASSERT(atomic_read(&cld
->cld_refcount
) > 0);
604 mutex_lock(&cld
->cld_lock
);
605 if (cld
->cld_stopping
|| cld
->cld_lostlock
) {
606 mutex_unlock(&cld
->cld_lock
);
609 /* this refcount will be released in mgc_requeue_thread. */
611 cld
->cld_lostlock
= 1;
612 mutex_unlock(&cld
->cld_lock
);
614 /* Hold lock for rq_state */
615 spin_lock(&config_list_lock
);
616 if (rq_state
& RQ_STOP
) {
617 spin_unlock(&config_list_lock
);
618 cld
->cld_lostlock
= 0;
622 spin_unlock(&config_list_lock
);
627 /********************** class fns **********************/
628 static int mgc_local_llog_init(const struct lu_env
*env
,
629 struct obd_device
*obd
,
630 struct obd_device
*disk
)
632 struct llog_ctxt
*ctxt
;
635 rc
= llog_setup(env
, obd
, &obd
->obd_olg
, LLOG_CONFIG_ORIG_CTXT
, disk
,
640 ctxt
= llog_get_context(obd
, LLOG_CONFIG_ORIG_CTXT
);
642 ctxt
->loc_dir
= obd
->u
.cli
.cl_mgc_configs_dir
;
648 static int mgc_local_llog_fini(const struct lu_env
*env
,
649 struct obd_device
*obd
)
651 struct llog_ctxt
*ctxt
;
653 ctxt
= llog_get_context(obd
, LLOG_CONFIG_ORIG_CTXT
);
654 llog_cleanup(env
, ctxt
);
659 static int mgc_fs_setup(struct obd_device
*obd
, struct super_block
*sb
)
661 struct lustre_sb_info
*lsi
= s2lsi(sb
);
662 struct client_obd
*cli
= &obd
->u
.cli
;
663 struct lu_fid rfid
, fid
;
664 struct dt_object
*root
, *dto
;
669 LASSERT(lsi
->lsi_dt_dev
);
675 /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
676 mutex_lock(&cli
->cl_mgc_mutex
);
678 cfs_cleanup_group_info();
680 /* Setup the configs dir */
681 rc
= lu_env_init(env
, LCT_MG_THREAD
);
685 fid
.f_seq
= FID_SEQ_LOCAL_NAME
;
688 rc
= local_oid_storage_init(env
, lsi
->lsi_dt_dev
, &fid
,
693 rc
= dt_root_get(env
, lsi
->lsi_dt_dev
, &rfid
);
697 root
= dt_locate_at(env
, lsi
->lsi_dt_dev
, &rfid
,
698 &cli
->cl_mgc_los
->los_dev
->dd_lu_dev
);
699 if (unlikely(IS_ERR(root
)))
700 GOTO(out_los
, rc
= PTR_ERR(root
));
702 dto
= local_file_find_or_create(env
, cli
->cl_mgc_los
, root
,
704 S_IFDIR
| S_IRUGO
| S_IWUSR
| S_IXUGO
);
705 lu_object_put_nocache(env
, &root
->do_lu
);
707 GOTO(out_los
, rc
= PTR_ERR(dto
));
709 cli
->cl_mgc_configs_dir
= dto
;
711 LASSERT(lsi
->lsi_osd_exp
->exp_obd
->obd_lvfs_ctxt
.dt
);
712 rc
= mgc_local_llog_init(env
, obd
, lsi
->lsi_osd_exp
->exp_obd
);
716 /* We take an obd ref to insure that we can't get to mgc_cleanup
717 * without calling mgc_fs_cleanup first. */
718 class_incref(obd
, "mgc_fs", obd
);
720 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
723 lu_object_put(env
, &cli
->cl_mgc_configs_dir
->do_lu
);
724 cli
->cl_mgc_configs_dir
= NULL
;
728 local_oid_storage_fini(env
, cli
->cl_mgc_los
);
729 cli
->cl_mgc_los
= NULL
;
730 mutex_unlock(&cli
->cl_mgc_mutex
);
739 static int mgc_fs_cleanup(struct obd_device
*obd
)
742 struct client_obd
*cli
= &obd
->u
.cli
;
745 LASSERT(cli
->cl_mgc_los
!= NULL
);
747 rc
= lu_env_init(&env
, LCT_MG_THREAD
);
751 mgc_local_llog_fini(&env
, obd
);
753 lu_object_put_nocache(&env
, &cli
->cl_mgc_configs_dir
->do_lu
);
754 cli
->cl_mgc_configs_dir
= NULL
;
756 local_oid_storage_fini(&env
, cli
->cl_mgc_los
);
757 cli
->cl_mgc_los
= NULL
;
761 class_decref(obd
, "mgc_fs", obd
);
762 mutex_unlock(&cli
->cl_mgc_mutex
);
767 static int mgc_llog_init(const struct lu_env
*env
, struct obd_device
*obd
)
769 struct llog_ctxt
*ctxt
;
772 /* setup only remote ctxt, the local disk context is switched per each
773 * filesystem during mgc_fs_setup() */
774 rc
= llog_setup(env
, obd
, &obd
->obd_olg
, LLOG_CONFIG_REPL_CTXT
, obd
,
779 ctxt
= llog_get_context(obd
, LLOG_CONFIG_REPL_CTXT
);
782 llog_initiator_connect(ctxt
);
788 static int mgc_llog_fini(const struct lu_env
*env
, struct obd_device
*obd
)
790 struct llog_ctxt
*ctxt
;
792 ctxt
= llog_get_context(obd
, LLOG_CONFIG_REPL_CTXT
);
794 llog_cleanup(env
, ctxt
);
799 static atomic_t mgc_count
= ATOMIC_INIT(0);
800 static int mgc_precleanup(struct obd_device
*obd
, enum obd_cleanup_stage stage
)
805 case OBD_CLEANUP_EARLY
:
807 case OBD_CLEANUP_EXPORTS
:
808 if (atomic_dec_and_test(&mgc_count
)) {
810 /* stop requeue thread */
811 spin_lock(&config_list_lock
);
812 running
= rq_state
& RQ_RUNNING
;
815 spin_unlock(&config_list_lock
);
818 wait_for_completion(&rq_exit
);
821 obd_cleanup_client_import(obd
);
822 rc
= mgc_llog_fini(NULL
, obd
);
824 CERROR("failed to cleanup llogging subsystems\n");
830 static int mgc_cleanup(struct obd_device
*obd
)
834 /* COMPAT_146 - old config logs may have added profiles we don't
836 if (obd
->obd_type
->typ_refcnt
<= 1)
837 /* Only for the last mgc */
838 class_del_profiles();
840 lprocfs_obd_cleanup(obd
);
843 rc
= client_obd_cleanup(obd
);
847 static int mgc_setup(struct obd_device
*obd
, struct lustre_cfg
*lcfg
)
849 struct lprocfs_static_vars lvars
;
854 rc
= client_obd_setup(obd
, lcfg
);
856 GOTO(err_decref
, rc
);
858 rc
= mgc_llog_init(NULL
, obd
);
860 CERROR("failed to setup llogging subsystems\n");
861 GOTO(err_cleanup
, rc
);
864 lprocfs_mgc_init_vars(&lvars
);
865 lprocfs_obd_setup(obd
, lvars
.obd_vars
);
866 sptlrpc_lprocfs_cliobd_attach(obd
);
868 if (atomic_inc_return(&mgc_count
) == 1) {
870 init_waitqueue_head(&rq_waitq
);
872 /* start requeue thread */
873 rc
= PTR_ERR(kthread_run(mgc_requeue_thread
, NULL
,
875 if (IS_ERR_VALUE(rc
)) {
876 CERROR("%s: Cannot start requeue thread (%d),"
877 "no more log updates!\n",
879 GOTO(err_cleanup
, rc
);
881 /* rc is the task_struct pointer of mgc_requeue_thread. */
888 client_obd_cleanup(obd
);
894 /* based on ll_mdc_blocking_ast */
895 static int mgc_blocking_ast(struct ldlm_lock
*lock
, struct ldlm_lock_desc
*desc
,
896 void *data
, int flag
)
898 struct lustre_handle lockh
;
899 struct config_llog_data
*cld
= (struct config_llog_data
*)data
;
903 case LDLM_CB_BLOCKING
:
904 /* mgs wants the lock, give it up... */
905 LDLM_DEBUG(lock
, "MGC blocking CB");
906 ldlm_lock2handle(lock
, &lockh
);
907 rc
= ldlm_cli_cancel(&lockh
, LCF_ASYNC
);
909 case LDLM_CB_CANCELING
:
910 /* We've given up the lock, prepare ourselves to update. */
911 LDLM_DEBUG(lock
, "MGC cancel CB");
913 CDEBUG(D_MGC
, "Lock res "DLDLMRES
" (%.8s)\n",
914 PLDLMRES(lock
->l_resource
),
915 (char *)&lock
->l_resource
->lr_name
.name
[0]);
918 CDEBUG(D_INFO
, "missing data, won't requeue\n");
922 /* held at mgc_process_log(). */
923 LASSERT(atomic_read(&cld
->cld_refcount
) > 0);
924 /* Are we done with this log? */
925 if (cld
->cld_stopping
) {
926 CDEBUG(D_MGC
, "log %s: stopping, won't requeue\n",
931 /* Make sure not to re-enqueue when the mgc is stopping
932 (we get called from client_disconnect_export) */
933 if (!lock
->l_conn_export
||
934 !lock
->l_conn_export
->exp_obd
->u
.cli
.cl_conn_count
) {
935 CDEBUG(D_MGC
, "log %.8s: disconnecting, won't requeue\n",
942 mgc_requeue_add(cld
);
952 /* Not sure where this should go... */
953 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
954 * that we can have a chance to try the secondary MGS if any. */
955 #define MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
957 #define MGC_TARGET_REG_LIMIT 10
958 #define MGC_SEND_PARAM_LIMIT 10
960 /* Send parameter to MGS*/
961 static int mgc_set_mgs_param(struct obd_export
*exp
,
962 struct mgs_send_param
*msp
)
964 struct ptlrpc_request
*req
;
965 struct mgs_send_param
*req_msp
, *rep_msp
;
968 req
= ptlrpc_request_alloc_pack(class_exp2cliimp(exp
),
969 &RQF_MGS_SET_INFO
, LUSTRE_MGS_VERSION
,
974 req_msp
= req_capsule_client_get(&req
->rq_pill
, &RMF_MGS_SEND_PARAM
);
976 ptlrpc_req_finished(req
);
980 memcpy(req_msp
, msp
, sizeof(*req_msp
));
981 ptlrpc_request_set_replen(req
);
983 /* Limit how long we will wait for the enqueue to complete */
984 req
->rq_delay_limit
= MGC_SEND_PARAM_LIMIT
;
985 rc
= ptlrpc_queue_wait(req
);
987 rep_msp
= req_capsule_server_get(&req
->rq_pill
, &RMF_MGS_SEND_PARAM
);
988 memcpy(msp
, rep_msp
, sizeof(*rep_msp
));
991 ptlrpc_req_finished(req
);
996 /* Take a config lock so we can get cancel notifications */
997 static int mgc_enqueue(struct obd_export
*exp
, struct lov_stripe_md
*lsm
,
998 __u32 type
, ldlm_policy_data_t
*policy
, __u32 mode
,
999 __u64
*flags
, void *bl_cb
, void *cp_cb
, void *gl_cb
,
1000 void *data
, __u32 lvb_len
, void *lvb_swabber
,
1001 struct lustre_handle
*lockh
)
1003 struct config_llog_data
*cld
= (struct config_llog_data
*)data
;
1004 struct ldlm_enqueue_info einfo
= {
1007 .ei_cb_bl
= mgc_blocking_ast
,
1008 .ei_cb_cp
= ldlm_completion_ast
,
1010 struct ptlrpc_request
*req
;
1011 int short_limit
= cld_is_sptlrpc(cld
);
1014 CDEBUG(D_MGC
, "Enqueue for %s (res %#llx)\n", cld
->cld_logname
,
1015 cld
->cld_resid
.name
[0]);
1017 /* We need a callback for every lockholder, so don't try to
1018 ldlm_lock_match (see rev 1.1.2.11.2.47) */
1019 req
= ptlrpc_request_alloc_pack(class_exp2cliimp(exp
),
1020 &RQF_LDLM_ENQUEUE
, LUSTRE_DLM_VERSION
,
1025 req_capsule_set_size(&req
->rq_pill
, &RMF_DLM_LVB
, RCL_SERVER
, 0);
1026 ptlrpc_request_set_replen(req
);
1028 /* check if this is server or client */
1029 if (cld
->cld_cfg
.cfg_sb
) {
1030 struct lustre_sb_info
*lsi
= s2lsi(cld
->cld_cfg
.cfg_sb
);
1031 if (lsi
&& IS_SERVER(lsi
))
1034 /* Limit how long we will wait for the enqueue to complete */
1035 req
->rq_delay_limit
= short_limit
? 5 : MGC_ENQUEUE_LIMIT
;
1036 rc
= ldlm_cli_enqueue(exp
, &req
, &einfo
, &cld
->cld_resid
, NULL
, flags
,
1037 NULL
, 0, LVB_T_NONE
, lockh
, 0);
1038 /* A failed enqueue should still call the mgc_blocking_ast,
1039 where it will be requeued if needed ("grant failed"). */
1040 ptlrpc_req_finished(req
);
1044 static int mgc_cancel(struct obd_export
*exp
, struct lov_stripe_md
*md
,
1045 __u32 mode
, struct lustre_handle
*lockh
)
1047 ldlm_lock_decref(lockh
, mode
);
1052 static void mgc_notify_active(struct obd_device
*unused
)
1054 /* wakeup mgc_requeue_thread to requeue mgc lock */
1055 spin_lock(&config_list_lock
);
1057 spin_unlock(&config_list_lock
);
1060 /* TODO: Help the MGS rebuild nidtbl. -jay */
1063 /* Send target_reg message to MGS */
1064 static int mgc_target_register(struct obd_export
*exp
,
1065 struct mgs_target_info
*mti
)
1067 struct ptlrpc_request
*req
;
1068 struct mgs_target_info
*req_mti
, *rep_mti
;
1071 req
= ptlrpc_request_alloc_pack(class_exp2cliimp(exp
),
1072 &RQF_MGS_TARGET_REG
, LUSTRE_MGS_VERSION
,
1077 req_mti
= req_capsule_client_get(&req
->rq_pill
, &RMF_MGS_TARGET_INFO
);
1079 ptlrpc_req_finished(req
);
1083 memcpy(req_mti
, mti
, sizeof(*req_mti
));
1084 ptlrpc_request_set_replen(req
);
1085 CDEBUG(D_MGC
, "register %s\n", mti
->mti_svname
);
1086 /* Limit how long we will wait for the enqueue to complete */
1087 req
->rq_delay_limit
= MGC_TARGET_REG_LIMIT
;
1089 rc
= ptlrpc_queue_wait(req
);
1091 rep_mti
= req_capsule_server_get(&req
->rq_pill
,
1092 &RMF_MGS_TARGET_INFO
);
1093 memcpy(mti
, rep_mti
, sizeof(*rep_mti
));
1094 CDEBUG(D_MGC
, "register %s got index = %d\n",
1095 mti
->mti_svname
, mti
->mti_stripe_index
);
1097 ptlrpc_req_finished(req
);
1102 int mgc_set_info_async(const struct lu_env
*env
, struct obd_export
*exp
,
1103 obd_count keylen
, void *key
, obd_count vallen
,
1104 void *val
, struct ptlrpc_request_set
*set
)
1108 /* Turn off initial_recov after we try all backup servers once */
1109 if (KEY_IS(KEY_INIT_RECOV_BACKUP
)) {
1110 struct obd_import
*imp
= class_exp2cliimp(exp
);
1112 if (vallen
!= sizeof(int))
1114 value
= *(int *)val
;
1115 CDEBUG(D_MGC
, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1116 imp
->imp_obd
->obd_name
, value
,
1117 imp
->imp_deactive
, imp
->imp_invalid
,
1118 imp
->imp_replayable
, imp
->imp_obd
->obd_replayable
,
1119 ptlrpc_import_state_name(imp
->imp_state
));
1120 /* Resurrect if we previously died */
1121 if ((imp
->imp_state
!= LUSTRE_IMP_FULL
&&
1122 imp
->imp_state
!= LUSTRE_IMP_NEW
) || value
> 1)
1123 ptlrpc_reconnect_import(imp
);
1126 /* FIXME move this to mgc_process_config */
1127 if (KEY_IS(KEY_REGISTER_TARGET
)) {
1128 struct mgs_target_info
*mti
;
1129 if (vallen
!= sizeof(struct mgs_target_info
))
1131 mti
= (struct mgs_target_info
*)val
;
1132 CDEBUG(D_MGC
, "register_target %s %#x\n",
1133 mti
->mti_svname
, mti
->mti_flags
);
1134 rc
= mgc_target_register(exp
, mti
);
1137 if (KEY_IS(KEY_SET_FS
)) {
1138 struct super_block
*sb
= (struct super_block
*)val
;
1140 if (vallen
!= sizeof(struct super_block
))
1143 rc
= mgc_fs_setup(exp
->exp_obd
, sb
);
1145 CERROR("set_fs got %d\n", rc
);
1149 if (KEY_IS(KEY_CLEAR_FS
)) {
1152 rc
= mgc_fs_cleanup(exp
->exp_obd
);
1154 CERROR("clear_fs got %d\n", rc
);
1158 if (KEY_IS(KEY_SET_INFO
)) {
1159 struct mgs_send_param
*msp
;
1161 msp
= (struct mgs_send_param
*)val
;
1162 rc
= mgc_set_mgs_param(exp
, msp
);
1165 if (KEY_IS(KEY_MGSSEC
)) {
1166 struct client_obd
*cli
= &exp
->exp_obd
->u
.cli
;
1167 struct sptlrpc_flavor flvr
;
1170 * empty string means using current flavor, if which haven't
1171 * been set yet, set it as null.
1173 * if flavor has been set previously, check the asking flavor
1174 * must match the existing one.
1177 if (cli
->cl_flvr_mgc
.sf_rpc
!= SPTLRPC_FLVR_INVALID
)
1183 rc
= sptlrpc_parse_flavor(val
, &flvr
);
1185 CERROR("invalid sptlrpc flavor %s to MGS\n",
1191 * caller already hold a mutex
1193 if (cli
->cl_flvr_mgc
.sf_rpc
== SPTLRPC_FLVR_INVALID
) {
1194 cli
->cl_flvr_mgc
= flvr
;
1195 } else if (memcmp(&cli
->cl_flvr_mgc
, &flvr
,
1196 sizeof(flvr
)) != 0) {
1199 sptlrpc_flavor2name(&cli
->cl_flvr_mgc
,
1201 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1202 "currently %s is in use\n",
1212 static int mgc_get_info(const struct lu_env
*env
, struct obd_export
*exp
,
1213 __u32 keylen
, void *key
, __u32
*vallen
, void *val
,
1214 struct lov_stripe_md
*unused
)
1218 if (KEY_IS(KEY_CONN_DATA
)) {
1219 struct obd_import
*imp
= class_exp2cliimp(exp
);
1220 struct obd_connect_data
*data
= val
;
1222 if (*vallen
== sizeof(*data
)) {
1223 *data
= imp
->imp_connect_data
;
1231 static int mgc_import_event(struct obd_device
*obd
,
1232 struct obd_import
*imp
,
1233 enum obd_import_event event
)
1237 LASSERT(imp
->imp_obd
== obd
);
1238 CDEBUG(D_MGC
, "import event %#x\n", event
);
1241 case IMP_EVENT_DISCON
:
1242 /* MGC imports should not wait for recovery */
1243 if (OCD_HAS_FLAG(&imp
->imp_connect_data
, IMP_RECOV
))
1244 ptlrpc_pinger_ir_down();
1246 case IMP_EVENT_INACTIVE
:
1248 case IMP_EVENT_INVALIDATE
: {
1249 struct ldlm_namespace
*ns
= obd
->obd_namespace
;
1250 ldlm_namespace_cleanup(ns
, LDLM_FL_LOCAL_ONLY
);
1253 case IMP_EVENT_ACTIVE
:
1254 CDEBUG(D_INFO
, "%s: Reactivating import\n", obd
->obd_name
);
1255 /* Clearing obd_no_recov allows us to continue pinging */
1256 obd
->obd_no_recov
= 0;
1257 mgc_notify_active(obd
);
1258 if (OCD_HAS_FLAG(&imp
->imp_connect_data
, IMP_RECOV
))
1259 ptlrpc_pinger_ir_up();
1263 case IMP_EVENT_DEACTIVATE
:
1264 case IMP_EVENT_ACTIVATE
:
1267 CERROR("Unknown import event %#x\n", event
);
1274 CONFIG_READ_NRPAGES_INIT
= 1 << (20 - PAGE_CACHE_SHIFT
),
1275 CONFIG_READ_NRPAGES
= 4
1278 static int mgc_apply_recover_logs(struct obd_device
*mgc
,
1279 struct config_llog_data
*cld
,
1281 void *data
, int datalen
, bool mne_swab
)
1283 struct config_llog_instance
*cfg
= &cld
->cld_cfg
;
1284 struct lustre_sb_info
*lsi
= s2lsi(cfg
->cfg_sb
);
1285 struct mgs_nidtbl_entry
*entry
;
1286 struct lustre_cfg
*lcfg
;
1287 struct lustre_cfg_bufs bufs
;
1288 u64 prev_version
= 0;
1296 LASSERT(cfg
->cfg_instance
!= NULL
);
1297 LASSERT(cfg
->cfg_sb
== cfg
->cfg_instance
);
1299 OBD_ALLOC(inst
, PAGE_CACHE_SIZE
);
1303 if (!IS_SERVER(lsi
)) {
1304 pos
= snprintf(inst
, PAGE_CACHE_SIZE
, "%p", cfg
->cfg_instance
);
1305 if (pos
>= PAGE_CACHE_SIZE
) {
1306 OBD_FREE(inst
, PAGE_CACHE_SIZE
);
1310 LASSERT(IS_MDT(lsi
));
1311 rc
= server_name2svname(lsi
->lsi_svname
, inst
, NULL
,
1314 OBD_FREE(inst
, PAGE_CACHE_SIZE
);
1322 bufsz
= PAGE_CACHE_SIZE
- pos
;
1324 while (datalen
> 0) {
1325 int entry_len
= sizeof(*entry
);
1327 struct obd_device
*obd
;
1334 if (datalen
< sizeof(*entry
))
1337 entry
= (typeof(entry
))(data
+ off
);
1340 if (entry
->mne_nid_type
!= 0) /* only support type 0 for ipv4 */
1342 if (entry
->mne_nid_count
== 0) /* at least one nid entry */
1344 if (entry
->mne_nid_size
!= sizeof(lnet_nid_t
))
1347 entry_len
+= entry
->mne_nid_count
* entry
->mne_nid_size
;
1348 if (datalen
< entry_len
) /* must have entry_len at least */
1351 /* Keep this swab for normal mixed endian handling. LU-1644 */
1353 lustre_swab_mgs_nidtbl_entry(entry
);
1354 if (entry
->mne_length
> PAGE_CACHE_SIZE
) {
1355 CERROR("MNE too large (%u)\n", entry
->mne_length
);
1359 if (entry
->mne_length
< entry_len
)
1362 off
+= entry
->mne_length
;
1363 datalen
-= entry
->mne_length
;
1367 if (entry
->mne_version
> max_version
) {
1368 CERROR("entry index(%lld) is over max_index(%lld)\n",
1369 entry
->mne_version
, max_version
);
1373 if (prev_version
>= entry
->mne_version
) {
1374 CERROR("index unsorted, prev %lld, now %lld\n",
1375 prev_version
, entry
->mne_version
);
1378 prev_version
= entry
->mne_version
;
1381 * Write a string with format "nid::instance" to
1382 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1385 is_ost
= entry
->mne_type
== LDD_F_SV_TYPE_OST
;
1386 memset(buf
, 0, bufsz
);
1390 /* lustre-OST0001-osc-<instance #> */
1391 strcpy(obdname
, cld
->cld_logname
);
1392 cname
= strrchr(obdname
, '-');
1393 if (cname
== NULL
) {
1394 CERROR("mgc %s: invalid logname %s\n",
1395 mgc
->obd_name
, obdname
);
1399 pos
= cname
- obdname
;
1401 pos
+= sprintf(obdname
+ pos
, "-%s%04x",
1402 is_ost
? "OST" : "MDT", entry
->mne_index
);
1404 cname
= is_ost
? "osc" : "mdc",
1405 pos
+= sprintf(obdname
+ pos
, "-%s-%s", cname
, inst
);
1406 lustre_cfg_bufs_reset(&bufs
, obdname
);
1408 /* find the obd by obdname */
1409 obd
= class_name2obd(obdname
);
1411 CDEBUG(D_INFO
, "mgc %s: cannot find obdname %s\n",
1412 mgc
->obd_name
, obdname
);
1414 /* this is a safe race, when the ost is starting up...*/
1418 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1421 pos
+= sprintf(params
, "%s.import=%s", cname
, "connection=");
1424 down_read(&obd
->u
.cli
.cl_sem
);
1425 if (obd
->u
.cli
.cl_import
== NULL
) {
1426 /* client does not connect to the OST yet */
1427 up_read(&obd
->u
.cli
.cl_sem
);
1432 /* TODO: iterate all nids to find one */
1433 /* find uuid by nid */
1434 rc
= client_import_find_conn(obd
->u
.cli
.cl_import
,
1436 (struct obd_uuid
*)uuid
);
1437 up_read(&obd
->u
.cli
.cl_sem
);
1439 CERROR("mgc: cannot find uuid by nid %s\n",
1440 libcfs_nid2str(entry
->u
.nids
[0]));
1444 CDEBUG(D_INFO
, "Find uuid %s by nid %s\n",
1445 uuid
, libcfs_nid2str(entry
->u
.nids
[0]));
1447 pos
+= strlen(uuid
);
1448 pos
+= sprintf(buf
+ pos
, "::%u", entry
->mne_instance
);
1449 LASSERT(pos
< bufsz
);
1451 lustre_cfg_bufs_set_string(&bufs
, 1, params
);
1454 lcfg
= lustre_cfg_new(LCFG_PARAM
, &bufs
);
1456 CERROR("mgc: cannot allocate memory\n");
1460 CDEBUG(D_INFO
, "ir apply logs %lld/%lld for %s -> %s\n",
1461 prev_version
, max_version
, obdname
, params
);
1463 rc
= class_process_config(lcfg
);
1464 lustre_cfg_free(lcfg
);
1466 CDEBUG(D_INFO
, "process config for %s error %d\n",
1469 /* continue, even one with error */
1472 OBD_FREE(inst
, PAGE_CACHE_SIZE
);
1477 * This function is called if this client was notified for target restarting
1478 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1480 static int mgc_process_recover_log(struct obd_device
*obd
,
1481 struct config_llog_data
*cld
)
1483 struct ptlrpc_request
*req
= NULL
;
1484 struct config_llog_instance
*cfg
= &cld
->cld_cfg
;
1485 struct mgs_config_body
*body
;
1486 struct mgs_config_res
*res
;
1487 struct ptlrpc_bulk_desc
*desc
;
1488 struct page
**pages
;
1491 bool mne_swab
= false;
1496 /* allocate buffer for bulk transfer.
1497 * if this is the first time for this mgs to read logs,
1498 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1499 * once; otherwise, it only reads increment of logs, this should be
1500 * small and CONFIG_READ_NRPAGES will be used.
1502 nrpages
= CONFIG_READ_NRPAGES
;
1503 if (cfg
->cfg_last_idx
== 0) /* the first time */
1504 nrpages
= CONFIG_READ_NRPAGES_INIT
;
1506 OBD_ALLOC(pages
, sizeof(*pages
) * nrpages
);
1508 GOTO(out
, rc
= -ENOMEM
);
1510 for (i
= 0; i
< nrpages
; i
++) {
1511 pages
[i
] = alloc_page(GFP_IOFS
);
1512 if (pages
[i
] == NULL
)
1513 GOTO(out
, rc
= -ENOMEM
);
1517 LASSERT(cld_is_recover(cld
));
1518 LASSERT(mutex_is_locked(&cld
->cld_lock
));
1519 req
= ptlrpc_request_alloc(class_exp2cliimp(cld
->cld_mgcexp
),
1520 &RQF_MGS_CONFIG_READ
);
1522 GOTO(out
, rc
= -ENOMEM
);
1524 rc
= ptlrpc_request_pack(req
, LUSTRE_MGS_VERSION
, MGS_CONFIG_READ
);
1529 body
= req_capsule_client_get(&req
->rq_pill
, &RMF_MGS_CONFIG_BODY
);
1530 LASSERT(body
!= NULL
);
1531 LASSERT(sizeof(body
->mcb_name
) > strlen(cld
->cld_logname
));
1532 if (strlcpy(body
->mcb_name
, cld
->cld_logname
, sizeof(body
->mcb_name
))
1533 >= sizeof(body
->mcb_name
))
1534 GOTO(out
, rc
= -E2BIG
);
1535 body
->mcb_offset
= cfg
->cfg_last_idx
+ 1;
1536 body
->mcb_type
= cld
->cld_type
;
1537 body
->mcb_bits
= PAGE_CACHE_SHIFT
;
1538 body
->mcb_units
= nrpages
;
1540 /* allocate bulk transfer descriptor */
1541 desc
= ptlrpc_prep_bulk_imp(req
, nrpages
, 1, BULK_PUT_SINK
,
1544 GOTO(out
, rc
= -ENOMEM
);
1546 for (i
= 0; i
< nrpages
; i
++)
1547 ptlrpc_prep_bulk_page_pin(desc
, pages
[i
], 0, PAGE_CACHE_SIZE
);
1549 ptlrpc_request_set_replen(req
);
1550 rc
= ptlrpc_queue_wait(req
);
1554 res
= req_capsule_server_get(&req
->rq_pill
, &RMF_MGS_CONFIG_RES
);
1555 if (res
->mcr_size
< res
->mcr_offset
)
1556 GOTO(out
, rc
= -EINVAL
);
1558 /* always update the index even though it might have errors with
1559 * handling the recover logs */
1560 cfg
->cfg_last_idx
= res
->mcr_offset
;
1561 eof
= res
->mcr_offset
== res
->mcr_size
;
1563 CDEBUG(D_INFO
, "Latest version %lld, more %d.\n",
1564 res
->mcr_offset
, eof
== false);
1566 ealen
= sptlrpc_cli_unwrap_bulk_read(req
, req
->rq_bulk
, 0);
1568 GOTO(out
, rc
= ealen
);
1570 if (ealen
> nrpages
<< PAGE_CACHE_SHIFT
)
1571 GOTO(out
, rc
= -EINVAL
);
1573 if (ealen
== 0) { /* no logs transferred */
1579 mne_swab
= !!ptlrpc_rep_need_swab(req
);
1580 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1581 /* This import flag means the server did an extra swab of IR MNE
1582 * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1583 if (unlikely(req
->rq_import
->imp_need_mne_swab
))
1584 mne_swab
= !mne_swab
;
1586 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1589 for (i
= 0; i
< nrpages
&& ealen
> 0; i
++) {
1593 ptr
= kmap(pages
[i
]);
1594 rc2
= mgc_apply_recover_logs(obd
, cld
, res
->mcr_offset
, ptr
,
1595 min_t(int, ealen
, PAGE_CACHE_SIZE
),
1599 CWARN("Process recover log %s error %d\n",
1600 cld
->cld_logname
, rc2
);
1604 ealen
-= PAGE_CACHE_SIZE
;
1609 ptlrpc_req_finished(req
);
1611 if (rc
== 0 && !eof
)
1615 for (i
= 0; i
< nrpages
; i
++) {
1616 if (pages
[i
] == NULL
)
1618 __free_page(pages
[i
]);
1620 OBD_FREE(pages
, sizeof(*pages
) * nrpages
);
1625 static int mgc_llog_local_copy(const struct lu_env
*env
,
1626 struct obd_device
*obd
,
1627 struct llog_ctxt
*rctxt
,
1628 struct llog_ctxt
*lctxt
, char *logname
)
1636 * - copy it to backup using llog_backup()
1637 * - copy remote llog to logname using llog_backup()
1638 * - if failed then move backup to logname again
1641 OBD_ALLOC(temp_log
, strlen(logname
) + 1);
1644 sprintf(temp_log
, "%sT", logname
);
1646 /* make a copy of local llog at first */
1647 rc
= llog_backup(env
, obd
, lctxt
, lctxt
, logname
, temp_log
);
1648 if (rc
< 0 && rc
!= -ENOENT
)
1650 /* copy remote llog to the local copy */
1651 rc
= llog_backup(env
, obd
, rctxt
, lctxt
, logname
, logname
);
1652 if (rc
== -ENOENT
) {
1653 /* no remote llog, delete local one too */
1654 llog_erase(env
, lctxt
, NULL
, logname
);
1655 } else if (rc
< 0) {
1656 /* error during backup, get local one back from the copy */
1657 llog_backup(env
, obd
, lctxt
, lctxt
, temp_log
, logname
);
1659 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1660 obd
->obd_name
, logname
, rc
);
1662 llog_erase(env
, lctxt
, NULL
, temp_log
);
1663 OBD_FREE(temp_log
, strlen(logname
) + 1);
1667 /* local_only means it cannot get remote llogs */
1668 static int mgc_process_cfg_log(struct obd_device
*mgc
,
1669 struct config_llog_data
*cld
, int local_only
)
1671 struct llog_ctxt
*ctxt
, *lctxt
= NULL
;
1672 struct dt_object
*cl_mgc_dir
= mgc
->u
.cli
.cl_mgc_configs_dir
;
1673 struct lustre_sb_info
*lsi
= NULL
;
1675 bool sptlrpc_started
= false;
1679 LASSERT(mutex_is_locked(&cld
->cld_lock
));
1682 * local copy of sptlrpc log is controlled elsewhere, don't try to
1685 if (cld_is_sptlrpc(cld
) && local_only
)
1688 if (cld
->cld_cfg
.cfg_sb
)
1689 lsi
= s2lsi(cld
->cld_cfg
.cfg_sb
);
1695 rc
= lu_env_init(env
, LCT_MG_THREAD
);
1699 ctxt
= llog_get_context(mgc
, LLOG_CONFIG_REPL_CTXT
);
1702 lctxt
= llog_get_context(mgc
, LLOG_CONFIG_ORIG_CTXT
);
1704 /* Copy the setup log locally if we can. Don't mess around if we're
1705 * running an MGS though (logs are already local). */
1706 if (lctxt
&& lsi
&& IS_SERVER(lsi
) && !IS_MGS(lsi
) &&
1707 cl_mgc_dir
!= NULL
&&
1708 lu2dt_dev(cl_mgc_dir
->do_lu
.lo_dev
) == lsi
->lsi_dt_dev
) {
1710 /* Only try to copy log if we have the lock. */
1711 rc
= mgc_llog_local_copy(env
, mgc
, ctxt
, lctxt
,
1713 if (local_only
|| rc
) {
1714 if (llog_is_empty(env
, lctxt
, cld
->cld_logname
)) {
1715 LCONSOLE_ERROR_MSG(0x13a,
1716 "Failed to get MGS log %s and no local copy.\n",
1718 GOTO(out_pop
, rc
= -ENOENT
);
1721 "Failed to get MGS log %s, using local copy for now, will try to update later.\n",
1724 /* Now, whether we copied or not, start using the local llog.
1725 * If we failed to copy, we'll start using whatever the old
1727 llog_ctxt_put(ctxt
);
1731 if (local_only
) /* no local log at client side */
1732 GOTO(out_pop
, rc
= -EIO
);
1735 if (cld_is_sptlrpc(cld
)) {
1736 sptlrpc_conf_log_update_begin(cld
->cld_logname
);
1737 sptlrpc_started
= true;
1740 /* logname and instance info should be the same, so use our
1741 * copy of the instance for the update. The cfg_last_idx will
1742 * be updated here. */
1743 rc
= class_config_parse_llog(env
, ctxt
, cld
->cld_logname
,
1747 __llog_ctxt_put(env
, ctxt
);
1749 __llog_ctxt_put(env
, lctxt
);
1752 * update settings on existing OBDs. doing it inside
1753 * of llog_process_lock so no device is attaching/detaching
1755 * the logname must be <fsname>-sptlrpc
1757 if (sptlrpc_started
) {
1758 LASSERT(cld_is_sptlrpc(cld
));
1759 sptlrpc_conf_log_update_end(cld
->cld_logname
);
1760 class_notify_sptlrpc_conf(cld
->cld_logname
,
1761 strlen(cld
->cld_logname
) -
1762 strlen("-sptlrpc"));
1771 /** Get a config log from the MGS and process it.
1772 * This func is called for both clients and servers.
1773 * Copy the log locally before parsing it if appropriate (non-MGS server)
1775 int mgc_process_log(struct obd_device
*mgc
, struct config_llog_data
*cld
)
1777 struct lustre_handle lockh
= { 0 };
1778 __u64 flags
= LDLM_FL_NO_LRU
;
1783 /* I don't want multiple processes running process_log at once --
1784 sounds like badness. It actually might be fine, as long as
1785 we're not trying to update from the same log
1786 simultaneously (in which case we should use a per-log sem.) */
1787 mutex_lock(&cld
->cld_lock
);
1788 if (cld
->cld_stopping
) {
1789 mutex_unlock(&cld
->cld_lock
);
1793 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG
, 20);
1795 CDEBUG(D_MGC
, "Process log %s:%p from %d\n", cld
->cld_logname
,
1796 cld
->cld_cfg
.cfg_instance
, cld
->cld_cfg
.cfg_last_idx
+ 1);
1798 /* Get the cfg lock on the llog */
1799 rcl
= mgc_enqueue(mgc
->u
.cli
.cl_mgc_mgsexp
, NULL
, LDLM_PLAIN
, NULL
,
1800 LCK_CR
, &flags
, NULL
, NULL
, NULL
,
1801 cld
, 0, NULL
, &lockh
);
1803 /* Get the cld, it will be released in mgc_blocking_ast. */
1804 config_log_get(cld
);
1805 rc
= ldlm_lock_set_data(&lockh
, (void *)cld
);
1808 CDEBUG(D_MGC
, "Can't get cfg lock: %d\n", rcl
);
1810 /* mark cld_lostlock so that it will requeue
1811 * after MGC becomes available. */
1812 cld
->cld_lostlock
= 1;
1813 /* Get extra reference, it will be put in requeue thread */
1814 config_log_get(cld
);
1818 if (cld_is_recover(cld
)) {
1819 rc
= 0; /* this is not a fatal error for recover log */
1821 rc
= mgc_process_recover_log(mgc
, cld
);
1823 rc
= mgc_process_cfg_log(mgc
, cld
, rcl
!= 0);
1826 CDEBUG(D_MGC
, "%s: configuration from log '%s' %sed (%d).\n",
1827 mgc
->obd_name
, cld
->cld_logname
, rc
? "fail" : "succeed", rc
);
1829 mutex_unlock(&cld
->cld_lock
);
1831 /* Now drop the lock so MGS can revoke it */
1833 rcl
= mgc_cancel(mgc
->u
.cli
.cl_mgc_mgsexp
, NULL
,
1836 CERROR("Can't drop cfg lock: %d\n", rcl
);
1843 /** Called from lustre_process_log.
1844 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1845 * any services, and adds it to the list logs to watch (follow).
1847 static int mgc_process_config(struct obd_device
*obd
, obd_count len
, void *buf
)
1849 struct lustre_cfg
*lcfg
= buf
;
1850 struct config_llog_instance
*cfg
= NULL
;
1854 switch (lcfg
->lcfg_command
) {
1855 case LCFG_LOV_ADD_OBD
: {
1856 /* Overloading this cfg command: register a new target */
1857 struct mgs_target_info
*mti
;
1859 if (LUSTRE_CFG_BUFLEN(lcfg
, 1) !=
1860 sizeof(struct mgs_target_info
))
1861 GOTO(out
, rc
= -EINVAL
);
1863 mti
= (struct mgs_target_info
*)lustre_cfg_buf(lcfg
, 1);
1864 CDEBUG(D_MGC
, "add_target %s %#x\n",
1865 mti
->mti_svname
, mti
->mti_flags
);
1866 rc
= mgc_target_register(obd
->u
.cli
.cl_mgc_mgsexp
, mti
);
1869 case LCFG_LOV_DEL_OBD
:
1870 /* Unregister has no meaning at the moment. */
1871 CERROR("lov_del_obd unimplemented\n");
1874 case LCFG_SPTLRPC_CONF
: {
1875 rc
= sptlrpc_process_config(lcfg
);
1878 case LCFG_LOG_START
: {
1879 struct config_llog_data
*cld
;
1880 struct super_block
*sb
;
1882 logname
= lustre_cfg_string(lcfg
, 1);
1883 cfg
= (struct config_llog_instance
*)lustre_cfg_buf(lcfg
, 2);
1884 sb
= *(struct super_block
**)lustre_cfg_buf(lcfg
, 3);
1886 CDEBUG(D_MGC
, "parse_log %s from %d\n", logname
,
1889 /* We're only called through here on the initial mount */
1890 rc
= config_log_add(obd
, logname
, cfg
, sb
);
1893 cld
= config_log_find(logname
, cfg
);
1900 /* FIXME only set this for old logs! Right now this forces
1901 us to always skip the "inside markers" check */
1902 cld
->cld_cfg
.cfg_flags
|= CFG_F_COMPAT146
;
1904 rc
= mgc_process_log(obd
, cld
);
1905 if (rc
== 0 && cld
->cld_recover
!= NULL
) {
1906 if (OCD_HAS_FLAG(&obd
->u
.cli
.cl_import
->
1907 imp_connect_data
, IMP_RECOV
)) {
1908 rc
= mgc_process_log(obd
, cld
->cld_recover
);
1910 struct config_llog_data
*cir
= cld
->cld_recover
;
1911 cld
->cld_recover
= NULL
;
1912 config_log_put(cir
);
1915 CERROR("Cannot process recover llog %d\n", rc
);
1918 if (rc
== 0 && cld
->cld_params
!= NULL
) {
1919 rc
= mgc_process_log(obd
, cld
->cld_params
);
1920 if (rc
== -ENOENT
) {
1922 "There is no params config file yet\n");
1925 /* params log is optional */
1928 "%s: can't process params llog: rc = %d\n",
1931 config_log_put(cld
);
1935 case LCFG_LOG_END
: {
1936 logname
= lustre_cfg_string(lcfg
, 1);
1938 if (lcfg
->lcfg_bufcount
>= 2)
1939 cfg
= (struct config_llog_instance
*)lustre_cfg_buf(
1941 rc
= config_log_end(logname
, cfg
);
1945 CERROR("Unknown command: %d\n", lcfg
->lcfg_command
);
1946 GOTO(out
, rc
= -EINVAL
);
1954 struct obd_ops mgc_obd_ops
= {
1955 .o_owner
= THIS_MODULE
,
1956 .o_setup
= mgc_setup
,
1957 .o_precleanup
= mgc_precleanup
,
1958 .o_cleanup
= mgc_cleanup
,
1959 .o_add_conn
= client_import_add_conn
,
1960 .o_del_conn
= client_import_del_conn
,
1961 .o_connect
= client_connect_import
,
1962 .o_disconnect
= client_disconnect_export
,
1963 /* .o_enqueue = mgc_enqueue, */
1964 .o_cancel
= mgc_cancel
,
1965 /* .o_iocontrol = mgc_iocontrol, */
1966 .o_set_info_async
= mgc_set_info_async
,
1967 .o_get_info
= mgc_get_info
,
1968 .o_import_event
= mgc_import_event
,
1969 .o_process_config
= mgc_process_config
,
1972 int __init
mgc_init(void)
1974 return class_register_type(&mgc_obd_ops
, NULL
, NULL
,
1975 LUSTRE_MGC_NAME
, NULL
);
1978 static void /*__exit*/ mgc_exit(void)
1980 class_unregister_type(LUSTRE_MGC_NAME
);
1983 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1984 MODULE_DESCRIPTION("Lustre Management Client");
1985 MODULE_LICENSE("GPL");
1987 module_init(mgc_init
);
1988 module_exit(mgc_exit
);