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, 2013, 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 <obd_class.h>
46 #include <lustre_dlm.h>
47 #include <lprocfs_status.h>
48 #include <lustre_log.h>
49 #include <lustre_disk.h>
50 #include <dt_object.h>
52 #include "mgc_internal.h"
54 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
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:
84 res_id->name[1] = cpu_to_le64(resname);
85 CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
86 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
90 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
92 /* fsname is at most 8 chars long, maybe contain "-".
93 * e.g. "lustre", "SUN-000" */
94 return mgc_name2resid(fsname, strlen(fsname), res_id, type);
96 EXPORT_SYMBOL(mgc_fsname2resid);
98 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
103 /* logname consists of "fsname-nodetype".
104 * e.g. "lustre-MDT0001", "SUN-000-client" */
105 name_end = strrchr(logname, '-');
107 len = name_end - logname;
108 return mgc_name2resid(logname, len, res_id, type);
111 /********************** config llog list **********************/
112 static CFS_LIST_HEAD(config_llog_list);
113 static DEFINE_SPINLOCK(config_list_lock);
115 /* Take a reference to a config log */
116 static int config_log_get(struct config_llog_data *cld)
119 cfs_atomic_inc(&cld->cld_refcount);
120 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
121 cfs_atomic_read(&cld->cld_refcount));
125 /* Drop a reference to a config log. When no longer referenced,
126 we can free the config log data */
127 static void config_log_put(struct config_llog_data *cld)
131 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
132 cfs_atomic_read(&cld->cld_refcount));
133 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
135 /* spinlock to make sure no item with 0 refcount in the list */
136 if (cfs_atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
137 cfs_list_del(&cld->cld_list_chain);
138 spin_unlock(&config_list_lock);
140 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
142 if (cld->cld_recover)
143 config_log_put(cld->cld_recover);
144 if (cld->cld_sptlrpc)
145 config_log_put(cld->cld_sptlrpc);
146 if (cld_is_sptlrpc(cld))
147 sptlrpc_conf_log_stop(cld->cld_logname);
149 class_export_put(cld->cld_mgcexp);
150 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
156 /* Find a config log by name */
158 struct config_llog_data *config_log_find(char *logname,
159 struct config_llog_instance *cfg)
161 struct config_llog_data *cld;
162 struct config_llog_data *found = NULL;
166 LASSERT(logname != NULL);
168 instance = cfg ? cfg->cfg_instance : NULL;
169 spin_lock(&config_list_lock);
170 cfs_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 cfs_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;
200 CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
201 cfg ? cfg->cfg_instance : 0);
203 OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
205 RETURN(ERR_PTR(-ENOMEM));
207 strcpy(cld->cld_logname, logname);
211 cld->cld_cfg.cfg_callback = class_config_llog_handler;
212 mutex_init(&cld->cld_lock);
213 cld->cld_cfg.cfg_last_idx = 0;
214 cld->cld_cfg.cfg_flags = 0;
215 cld->cld_cfg.cfg_sb = sb;
216 cld->cld_type = type;
217 cfs_atomic_set(&cld->cld_refcount, 1);
219 /* Keep the mgc around until we are done */
220 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
222 if (cld_is_sptlrpc(cld)) {
223 sptlrpc_conf_log_start(logname);
224 cld->cld_cfg.cfg_obdname = obd->obd_name;
227 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
229 spin_lock(&config_list_lock);
230 cfs_list_add(&cld->cld_list_chain, &config_llog_list);
231 spin_unlock(&config_list_lock);
238 if (cld_is_sptlrpc(cld)) {
239 rc = mgc_process_log(obd, cld);
240 if (rc && rc != -ENOENT)
241 CERROR("failed processing sptlrpc log: %d\n", rc);
247 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
249 struct config_llog_instance *cfg,
250 struct super_block *sb)
252 struct config_llog_instance lcfg = *cfg;
253 struct lustre_sb_info *lsi = s2lsi(sb);
254 struct config_llog_data *cld;
260 /* for osp-on-ost, see lustre_start_osp() */
261 if (IS_MDT(lsi) && lcfg.cfg_instance)
264 /* we have to use different llog for clients and mdts for cmd
265 * where only clients are notified if one of cmd server restarts */
266 LASSERT(strlen(fsname) < sizeof(logname) / 2);
267 strcpy(logname, fsname);
268 if (IS_SERVER(lsi)) { /* mdt */
269 LASSERT(lcfg.cfg_instance == NULL);
270 lcfg.cfg_instance = sb;
271 strcat(logname, "-mdtir");
273 LASSERT(lcfg.cfg_instance != NULL);
274 strcat(logname, "-cliir");
277 cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
282 /** Add this log to the list of active logs watched by an MGC.
283 * Active means we're watching for updates.
284 * We have one active log per "mount" - client instance or servername.
285 * Each instance may be at a different point in the log.
287 static int config_log_add(struct obd_device *obd, char *logname,
288 struct config_llog_instance *cfg,
289 struct super_block *sb)
291 struct lustre_sb_info *lsi = s2lsi(sb);
292 struct config_llog_data *cld;
293 struct config_llog_data *sptlrpc_cld;
298 CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
301 * for each regular log, the depended sptlrpc log name is
302 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
304 ptr = strrchr(logname, '-');
305 if (ptr == NULL || ptr - logname > 8) {
306 CERROR("logname %s is too long\n", logname);
310 memcpy(seclogname, logname, ptr - logname);
311 strcpy(seclogname + (ptr - logname), "-sptlrpc");
313 sptlrpc_cld = config_log_find(seclogname, NULL);
314 if (sptlrpc_cld == NULL) {
315 sptlrpc_cld = do_config_log_add(obd, seclogname,
316 CONFIG_T_SPTLRPC, NULL, NULL);
317 if (IS_ERR(sptlrpc_cld)) {
318 CERROR("can't create sptlrpc log: %s\n", seclogname);
319 RETURN(PTR_ERR(sptlrpc_cld));
323 cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
325 CERROR("can't create log: %s\n", logname);
326 config_log_put(sptlrpc_cld);
327 RETURN(PTR_ERR(cld));
330 cld->cld_sptlrpc = sptlrpc_cld;
332 LASSERT(lsi->lsi_lmd);
333 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
334 struct config_llog_data *recover_cld;
335 *strrchr(seclogname, '-') = 0;
336 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
337 if (IS_ERR(recover_cld)) {
339 RETURN(PTR_ERR(recover_cld));
341 cld->cld_recover = recover_cld;
347 DEFINE_MUTEX(llog_process_lock);
349 /** Stop watching for updates on this log.
351 static int config_log_end(char *logname, struct config_llog_instance *cfg)
353 struct config_llog_data *cld;
354 struct config_llog_data *cld_sptlrpc = NULL;
355 struct config_llog_data *cld_recover = NULL;
359 cld = config_log_find(logname, cfg);
363 mutex_lock(&cld->cld_lock);
365 * if cld_stopping is set, it means we didn't start the log thus
366 * not owning the start ref. this can happen after previous umount:
367 * the cld still hanging there waiting for lock cancel, and we
368 * remount again but failed in the middle and call log_end without
371 if (unlikely(cld->cld_stopping)) {
372 mutex_unlock(&cld->cld_lock);
373 /* drop the ref from the find */
378 cld->cld_stopping = 1;
380 cld_recover = cld->cld_recover;
381 cld->cld_recover = NULL;
382 mutex_unlock(&cld->cld_lock);
385 mutex_lock(&cld_recover->cld_lock);
386 cld_recover->cld_stopping = 1;
387 mutex_unlock(&cld_recover->cld_lock);
388 config_log_put(cld_recover);
391 spin_lock(&config_list_lock);
392 cld_sptlrpc = cld->cld_sptlrpc;
393 cld->cld_sptlrpc = NULL;
394 spin_unlock(&config_list_lock);
397 config_log_put(cld_sptlrpc);
399 /* drop the ref from the find */
401 /* drop the start ref */
404 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
409 int lprocfs_mgc_rd_ir_state(char *page, char **start, off_t off,
410 int count, int *eof, void *data)
412 struct obd_device *obd = data;
413 struct obd_import *imp = obd->u.cli.cl_import;
414 struct obd_connect_data *ocd = &imp->imp_connect_data;
415 struct config_llog_data *cld;
419 rc = snprintf(page, count, "imperative_recovery: %s\n",
420 OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
421 rc += snprintf(page + rc, count - rc, "client_state:\n");
423 spin_lock(&config_list_lock);
424 cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
425 if (cld->cld_recover == NULL)
427 rc += snprintf(page + rc, count - rc,
428 " - { client: %s, nidtbl_version: %u }\n",
430 cld->cld_recover->cld_cfg.cfg_last_idx);
432 spin_unlock(&config_list_lock);
437 /* reenqueue any lost locks */
438 #define RQ_RUNNING 0x1
442 static int rq_state = 0;
443 static cfs_waitq_t rq_waitq;
444 static DECLARE_COMPLETION(rq_exit);
446 static void do_requeue(struct config_llog_data *cld)
449 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
451 /* Do not run mgc_process_log on a disconnected export or an
452 export which is being disconnected. Take the client
453 semaphore to make the check non-racy. */
454 down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
455 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
456 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
457 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
459 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
462 up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
467 /* this timeout represents how many seconds MGC should wait before
468 * requeue config and recover lock to the MGS. We need to randomize this
469 * in order to not flood the MGS.
471 #define MGC_TIMEOUT_MIN_SECONDS 5
472 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
474 static int mgc_requeue_thread(void *data)
479 CDEBUG(D_MGC, "Starting requeue thread\n");
481 /* Keep trying failed locks periodically */
482 spin_lock(&config_list_lock);
483 rq_state |= RQ_RUNNING;
485 struct l_wait_info lwi;
486 struct config_llog_data *cld, *cld_prev;
487 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
488 int stopped = !!(rq_state & RQ_STOP);
491 /* Any new or requeued lostlocks will change the state */
492 rq_state &= ~(RQ_NOW | RQ_LATER);
493 spin_unlock(&config_list_lock);
495 /* Always wait a few seconds to allow the server who
496 caused the lock revocation to finish its setup, plus some
497 random so everyone doesn't try to reconnect at once. */
498 to = MGC_TIMEOUT_MIN_SECONDS * CFS_HZ;
499 to += rand * CFS_HZ / 100; /* rand is centi-seconds */
500 lwi = LWI_TIMEOUT(to, NULL, NULL);
501 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi);
504 * iterate & processing through the list. for each cld, process
505 * its depending sptlrpc cld firstly (if any) and then itself.
507 * it's guaranteed any item in the list must have
508 * reference > 0; and if cld_lostlock is set, at
509 * least one reference is taken by the previous enqueue.
513 spin_lock(&config_list_lock);
514 cfs_list_for_each_entry(cld, &config_llog_list,
516 if (!cld->cld_lostlock)
519 spin_unlock(&config_list_lock);
521 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
523 /* Whether we enqueued again or not in mgc_process_log,
524 * we're done with the ref from the old enqueue */
526 config_log_put(cld_prev);
529 cld->cld_lostlock = 0;
530 if (likely(!stopped))
533 spin_lock(&config_list_lock);
535 spin_unlock(&config_list_lock);
537 config_log_put(cld_prev);
539 /* break after scanning the list so that we can drop
540 * refcount to losing lock clds */
541 if (unlikely(stopped)) {
542 spin_lock(&config_list_lock);
546 /* Wait a bit to see if anyone else needs a requeue */
547 lwi = (struct l_wait_info) { 0 };
548 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
550 spin_lock(&config_list_lock);
552 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
553 rq_state &= ~RQ_RUNNING;
554 spin_unlock(&config_list_lock);
558 CDEBUG(D_MGC, "Ending requeue thread\n");
562 /* Add a cld to the list to requeue. Start the requeue thread if needed.
563 We are responsible for dropping the config log reference from here on out. */
564 static void mgc_requeue_add(struct config_llog_data *cld)
568 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
569 cld->cld_logname, cfs_atomic_read(&cld->cld_refcount),
570 cld->cld_stopping, rq_state);
571 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
573 mutex_lock(&cld->cld_lock);
574 if (cld->cld_stopping || cld->cld_lostlock) {
575 mutex_unlock(&cld->cld_lock);
578 /* this refcount will be released in mgc_requeue_thread. */
580 cld->cld_lostlock = 1;
581 mutex_unlock(&cld->cld_lock);
583 /* Hold lock for rq_state */
584 spin_lock(&config_list_lock);
585 if (rq_state & RQ_STOP) {
586 spin_unlock(&config_list_lock);
587 cld->cld_lostlock = 0;
591 spin_unlock(&config_list_lock);
592 cfs_waitq_signal(&rq_waitq);
597 /********************** class fns **********************/
598 static int mgc_local_llog_init(const struct lu_env *env,
599 struct obd_device *obd,
600 struct obd_device *disk)
602 struct llog_ctxt *ctxt;
607 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
612 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
614 ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
620 static int mgc_local_llog_fini(const struct lu_env *env,
621 struct obd_device *obd)
623 struct llog_ctxt *ctxt;
627 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
628 llog_cleanup(env, ctxt);
633 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb)
635 struct lustre_sb_info *lsi = s2lsi(sb);
636 struct client_obd *cli = &obd->u.cli;
637 struct lu_fid rfid, fid;
638 struct dt_object *root, *dto;
645 LASSERT(lsi->lsi_dt_dev);
651 /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
652 down(&cli->cl_mgc_sem);
654 cfs_cleanup_group_info();
656 /* Setup the configs dir */
657 rc = lu_env_init(env, LCT_MG_THREAD);
661 fid.f_seq = FID_SEQ_LOCAL_NAME;
664 rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
669 rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
673 root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
674 &cli->cl_mgc_los->los_dev->dd_lu_dev);
675 if (unlikely(IS_ERR(root)))
676 GOTO(out_los, rc = PTR_ERR(root));
678 dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
680 S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
681 lu_object_put_nocache(env, &root->do_lu);
683 GOTO(out_los, rc = PTR_ERR(dto));
685 cli->cl_mgc_configs_dir = dto;
687 LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
688 rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
692 /* We take an obd ref to insure that we can't get to mgc_cleanup
693 * without calling mgc_fs_cleanup first. */
694 class_incref(obd, "mgc_fs", obd);
696 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
700 lu_object_put(env, &cli->cl_mgc_configs_dir->do_lu);
701 cli->cl_mgc_configs_dir = NULL;
705 local_oid_storage_fini(env, cli->cl_mgc_los);
706 cli->cl_mgc_los = NULL;
707 up(&cli->cl_mgc_sem);
716 static int mgc_fs_cleanup(struct obd_device *obd)
719 struct client_obd *cli = &obd->u.cli;
724 LASSERT(cli->cl_mgc_los != NULL);
726 rc = lu_env_init(&env, LCT_MG_THREAD);
730 mgc_local_llog_fini(&env, obd);
732 lu_object_put_nocache(&env, &cli->cl_mgc_configs_dir->do_lu);
733 cli->cl_mgc_configs_dir = NULL;
735 local_oid_storage_fini(&env, cli->cl_mgc_los);
736 cli->cl_mgc_los = NULL;
740 class_decref(obd, "mgc_fs", obd);
741 up(&cli->cl_mgc_sem);
746 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
748 struct llog_ctxt *ctxt;
753 /* setup only remote ctxt, the local disk context is switched per each
754 * filesystem during mgc_fs_setup() */
755 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
760 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
763 llog_initiator_connect(ctxt);
769 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
771 struct llog_ctxt *ctxt;
775 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
777 llog_cleanup(env, ctxt);
783 static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0);
784 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
790 case OBD_CLEANUP_EARLY:
792 case OBD_CLEANUP_EXPORTS:
793 if (cfs_atomic_dec_and_test(&mgc_count)) {
795 /* stop requeue thread */
796 spin_lock(&config_list_lock);
797 running = rq_state & RQ_RUNNING;
800 spin_unlock(&config_list_lock);
802 cfs_waitq_signal(&rq_waitq);
803 wait_for_completion(&rq_exit);
806 obd_cleanup_client_import(obd);
807 rc = mgc_llog_fini(NULL, obd);
809 CERROR("failed to cleanup llogging subsystems\n");
815 static int mgc_cleanup(struct obd_device *obd)
820 /* COMPAT_146 - old config logs may have added profiles we don't
822 if (obd->obd_type->typ_refcnt <= 1)
823 /* Only for the last mgc */
824 class_del_profiles();
826 lprocfs_obd_cleanup(obd);
829 rc = client_obd_cleanup(obd);
833 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
835 struct lprocfs_static_vars lvars;
841 rc = client_obd_setup(obd, lcfg);
843 GOTO(err_decref, rc);
845 rc = mgc_llog_init(NULL, obd);
847 CERROR("failed to setup llogging subsystems\n");
848 GOTO(err_cleanup, rc);
851 lprocfs_mgc_init_vars(&lvars);
852 lprocfs_obd_setup(obd, lvars.obd_vars);
853 sptlrpc_lprocfs_cliobd_attach(obd);
855 if (cfs_atomic_inc_return(&mgc_count) == 1) {
857 cfs_waitq_init(&rq_waitq);
859 /* start requeue thread */
860 rc = PTR_ERR(kthread_run(mgc_requeue_thread, NULL,
862 if (IS_ERR_VALUE(rc)) {
863 CERROR("%s: Cannot start requeue thread (%d),"
864 "no more log updates!\n",
866 GOTO(err_cleanup, rc);
868 /* rc is the task_struct pointer of mgc_requeue_thread. */
875 client_obd_cleanup(obd);
881 /* based on ll_mdc_blocking_ast */
882 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
883 void *data, int flag)
885 struct lustre_handle lockh;
886 struct config_llog_data *cld = (struct config_llog_data *)data;
891 case LDLM_CB_BLOCKING:
892 /* mgs wants the lock, give it up... */
893 LDLM_DEBUG(lock, "MGC blocking CB");
894 ldlm_lock2handle(lock, &lockh);
895 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
897 case LDLM_CB_CANCELING:
898 /* We've given up the lock, prepare ourselves to update. */
899 LDLM_DEBUG(lock, "MGC cancel CB");
901 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
902 PLDLMRES(lock->l_resource),
903 (char *)&lock->l_resource->lr_name.name[0]);
906 CDEBUG(D_INFO, "missing data, won't requeue\n");
910 /* held at mgc_process_log(). */
911 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
912 /* Are we done with this log? */
913 if (cld->cld_stopping) {
914 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
919 /* Make sure not to re-enqueue when the mgc is stopping
920 (we get called from client_disconnect_export) */
921 if (!lock->l_conn_export ||
922 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
923 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
930 mgc_requeue_add(cld);
940 /* Not sure where this should go... */
941 #define MGC_ENQUEUE_LIMIT 50
942 #define MGC_TARGET_REG_LIMIT 10
943 #define MGC_SEND_PARAM_LIMIT 10
945 /* Send parameter to MGS*/
946 static int mgc_set_mgs_param(struct obd_export *exp,
947 struct mgs_send_param *msp)
949 struct ptlrpc_request *req;
950 struct mgs_send_param *req_msp, *rep_msp;
954 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
955 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
960 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
962 ptlrpc_req_finished(req);
966 memcpy(req_msp, msp, sizeof(*req_msp));
967 ptlrpc_request_set_replen(req);
969 /* Limit how long we will wait for the enqueue to complete */
970 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
971 rc = ptlrpc_queue_wait(req);
973 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
974 memcpy(msp, rep_msp, sizeof(*rep_msp));
977 ptlrpc_req_finished(req);
982 /* Take a config lock so we can get cancel notifications */
983 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
984 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
985 __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
986 void *data, __u32 lvb_len, void *lvb_swabber,
987 struct lustre_handle *lockh)
989 struct config_llog_data *cld = (struct config_llog_data *)data;
990 struct ldlm_enqueue_info einfo = {
993 .ei_cb_bl = mgc_blocking_ast,
994 .ei_cb_cp = ldlm_completion_ast,
996 struct ptlrpc_request *req;
997 int short_limit = cld_is_sptlrpc(cld);
1001 CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
1002 cld->cld_resid.name[0]);
1004 /* We need a callback for every lockholder, so don't try to
1005 ldlm_lock_match (see rev 1.1.2.11.2.47) */
1006 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1007 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1012 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1013 ptlrpc_request_set_replen(req);
1015 /* check if this is server or client */
1016 if (cld->cld_cfg.cfg_sb) {
1017 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1018 if (lsi && IS_SERVER(lsi))
1021 /* Limit how long we will wait for the enqueue to complete */
1022 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1023 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1024 NULL, 0, LVB_T_NONE, lockh, 0);
1025 /* A failed enqueue should still call the mgc_blocking_ast,
1026 where it will be requeued if needed ("grant failed"). */
1027 ptlrpc_req_finished(req);
1031 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
1032 __u32 mode, struct lustre_handle *lockh)
1036 ldlm_lock_decref(lockh, mode);
1041 static void mgc_notify_active(struct obd_device *unused)
1043 /* wakeup mgc_requeue_thread to requeue mgc lock */
1044 spin_lock(&config_list_lock);
1046 spin_unlock(&config_list_lock);
1047 cfs_waitq_signal(&rq_waitq);
1049 /* TODO: Help the MGS rebuild nidtbl. -jay */
1052 /* Send target_reg message to MGS */
1053 static int mgc_target_register(struct obd_export *exp,
1054 struct mgs_target_info *mti)
1056 struct ptlrpc_request *req;
1057 struct mgs_target_info *req_mti, *rep_mti;
1061 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1062 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1067 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1069 ptlrpc_req_finished(req);
1073 memcpy(req_mti, mti, sizeof(*req_mti));
1074 ptlrpc_request_set_replen(req);
1075 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1076 /* Limit how long we will wait for the enqueue to complete */
1077 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1079 rc = ptlrpc_queue_wait(req);
1081 rep_mti = req_capsule_server_get(&req->rq_pill,
1082 &RMF_MGS_TARGET_INFO);
1083 memcpy(mti, rep_mti, sizeof(*rep_mti));
1084 CDEBUG(D_MGC, "register %s got index = %d\n",
1085 mti->mti_svname, mti->mti_stripe_index);
1087 ptlrpc_req_finished(req);
1092 int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1093 obd_count keylen, void *key, obd_count vallen,
1094 void *val, struct ptlrpc_request_set *set)
1099 /* Turn off initial_recov after we try all backup servers once */
1100 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1101 struct obd_import *imp = class_exp2cliimp(exp);
1103 if (vallen != sizeof(int))
1105 value = *(int *)val;
1106 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1107 imp->imp_obd->obd_name, value,
1108 imp->imp_deactive, imp->imp_invalid,
1109 imp->imp_replayable, imp->imp_obd->obd_replayable,
1110 ptlrpc_import_state_name(imp->imp_state));
1111 /* Resurrect if we previously died */
1112 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1113 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1114 ptlrpc_reconnect_import(imp);
1117 /* FIXME move this to mgc_process_config */
1118 if (KEY_IS(KEY_REGISTER_TARGET)) {
1119 struct mgs_target_info *mti;
1120 if (vallen != sizeof(struct mgs_target_info))
1122 mti = (struct mgs_target_info *)val;
1123 CDEBUG(D_MGC, "register_target %s %#x\n",
1124 mti->mti_svname, mti->mti_flags);
1125 rc = mgc_target_register(exp, mti);
1128 if (KEY_IS(KEY_SET_FS)) {
1129 struct super_block *sb = (struct super_block *)val;
1131 if (vallen != sizeof(struct super_block))
1134 rc = mgc_fs_setup(exp->exp_obd, sb);
1137 if (KEY_IS(KEY_CLEAR_FS)) {
1140 rc = mgc_fs_cleanup(exp->exp_obd);
1143 if (KEY_IS(KEY_SET_INFO)) {
1144 struct mgs_send_param *msp;
1146 msp = (struct mgs_send_param *)val;
1147 rc = mgc_set_mgs_param(exp, msp);
1150 if (KEY_IS(KEY_MGSSEC)) {
1151 struct client_obd *cli = &exp->exp_obd->u.cli;
1152 struct sptlrpc_flavor flvr;
1155 * empty string means using current flavor, if which haven't
1156 * been set yet, set it as null.
1158 * if flavor has been set previously, check the asking flavor
1159 * must match the existing one.
1162 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1168 rc = sptlrpc_parse_flavor(val, &flvr);
1170 CERROR("invalid sptlrpc flavor %s to MGS\n",
1176 * caller already hold a mutex
1178 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1179 cli->cl_flvr_mgc = flvr;
1180 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1181 sizeof(flvr)) != 0) {
1184 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1186 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1187 "currently %s is in use\n",
1197 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1198 __u32 keylen, void *key, __u32 *vallen, void *val,
1199 struct lov_stripe_md *unused)
1203 if (KEY_IS(KEY_CONN_DATA)) {
1204 struct obd_import *imp = class_exp2cliimp(exp);
1205 struct obd_connect_data *data = val;
1207 if (*vallen == sizeof(*data)) {
1208 *data = imp->imp_connect_data;
1216 static int mgc_import_event(struct obd_device *obd,
1217 struct obd_import *imp,
1218 enum obd_import_event event)
1222 LASSERT(imp->imp_obd == obd);
1223 CDEBUG(D_MGC, "import event %#x\n", event);
1226 case IMP_EVENT_DISCON:
1227 /* MGC imports should not wait for recovery */
1228 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1229 ptlrpc_pinger_ir_down();
1231 case IMP_EVENT_INACTIVE:
1233 case IMP_EVENT_INVALIDATE: {
1234 struct ldlm_namespace *ns = obd->obd_namespace;
1235 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1238 case IMP_EVENT_ACTIVE:
1239 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1240 /* Clearing obd_no_recov allows us to continue pinging */
1241 obd->obd_no_recov = 0;
1242 mgc_notify_active(obd);
1243 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1244 ptlrpc_pinger_ir_up();
1248 case IMP_EVENT_DEACTIVATE:
1249 case IMP_EVENT_ACTIVATE:
1252 CERROR("Unknown import event %#x\n", event);
1259 CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_CACHE_SHIFT),
1260 CONFIG_READ_NRPAGES = 4
1263 static int mgc_apply_recover_logs(struct obd_device *mgc,
1264 struct config_llog_data *cld,
1266 void *data, int datalen, bool mne_swab)
1268 struct config_llog_instance *cfg = &cld->cld_cfg;
1269 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1270 struct mgs_nidtbl_entry *entry;
1271 struct lustre_cfg *lcfg;
1272 struct lustre_cfg_bufs bufs;
1273 u64 prev_version = 0;
1282 LASSERT(cfg->cfg_instance != NULL);
1283 LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1285 OBD_ALLOC(inst, PAGE_CACHE_SIZE);
1289 if (!IS_SERVER(lsi)) {
1290 pos = snprintf(inst, PAGE_CACHE_SIZE, "%p", cfg->cfg_instance);
1291 if (pos >= PAGE_CACHE_SIZE) {
1292 OBD_FREE(inst, PAGE_CACHE_SIZE);
1296 LASSERT(IS_MDT(lsi));
1297 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1300 OBD_FREE(inst, PAGE_CACHE_SIZE);
1308 bufsz = PAGE_CACHE_SIZE - pos;
1310 while (datalen > 0) {
1311 int entry_len = sizeof(*entry);
1313 struct obd_device *obd;
1320 if (datalen < sizeof(*entry))
1323 entry = (typeof(entry))(data + off);
1326 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1328 if (entry->mne_nid_count == 0) /* at least one nid entry */
1330 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1333 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1334 if (datalen < entry_len) /* must have entry_len at least */
1337 /* Keep this swab for normal mixed endian handling. LU-1644 */
1339 lustre_swab_mgs_nidtbl_entry(entry);
1340 if (entry->mne_length > PAGE_CACHE_SIZE) {
1341 CERROR("MNE too large (%u)\n", entry->mne_length);
1345 if (entry->mne_length < entry_len)
1348 off += entry->mne_length;
1349 datalen -= entry->mne_length;
1353 if (entry->mne_version > max_version) {
1354 CERROR("entry index(%lld) is over max_index(%lld)\n",
1355 entry->mne_version, max_version);
1359 if (prev_version >= entry->mne_version) {
1360 CERROR("index unsorted, prev %lld, now %lld\n",
1361 prev_version, entry->mne_version);
1364 prev_version = entry->mne_version;
1367 * Write a string with format "nid::instance" to
1368 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1371 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1372 memset(buf, 0, bufsz);
1376 /* lustre-OST0001-osc-<instance #> */
1377 strcpy(obdname, cld->cld_logname);
1378 cname = strrchr(obdname, '-');
1379 if (cname == NULL) {
1380 CERROR("mgc %s: invalid logname %s\n",
1381 mgc->obd_name, obdname);
1385 pos = cname - obdname;
1387 pos += sprintf(obdname + pos, "-%s%04x",
1388 is_ost ? "OST" : "MDT", entry->mne_index);
1390 cname = is_ost ? "osc" : "mdc",
1391 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1392 lustre_cfg_bufs_reset(&bufs, obdname);
1394 /* find the obd by obdname */
1395 obd = class_name2obd(obdname);
1397 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1398 mgc->obd_name, obdname);
1400 /* this is a safe race, when the ost is starting up...*/
1404 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1407 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1410 down_read(&obd->u.cli.cl_sem);
1411 if (obd->u.cli.cl_import == NULL) {
1412 /* client does not connect to the OST yet */
1413 up_read(&obd->u.cli.cl_sem);
1418 /* TODO: iterate all nids to find one */
1419 /* find uuid by nid */
1420 rc = client_import_find_conn(obd->u.cli.cl_import,
1422 (struct obd_uuid *)uuid);
1423 up_read(&obd->u.cli.cl_sem);
1425 CERROR("mgc: cannot find uuid by nid %s\n",
1426 libcfs_nid2str(entry->u.nids[0]));
1430 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1431 uuid, libcfs_nid2str(entry->u.nids[0]));
1433 pos += strlen(uuid);
1434 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1435 LASSERT(pos < bufsz);
1437 lustre_cfg_bufs_set_string(&bufs, 1, params);
1440 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1442 CERROR("mgc: cannot allocate memory\n");
1446 CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n",
1447 prev_version, max_version, obdname, params);
1449 rc = class_process_config(lcfg);
1450 lustre_cfg_free(lcfg);
1452 CDEBUG(D_INFO, "process config for %s error %d\n",
1455 /* continue, even one with error */
1458 OBD_FREE(inst, PAGE_CACHE_SIZE);
1463 * This function is called if this client was notified for target restarting
1464 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1466 static int mgc_process_recover_log(struct obd_device *obd,
1467 struct config_llog_data *cld)
1469 struct ptlrpc_request *req = NULL;
1470 struct config_llog_instance *cfg = &cld->cld_cfg;
1471 struct mgs_config_body *body;
1472 struct mgs_config_res *res;
1473 struct ptlrpc_bulk_desc *desc;
1474 struct page **pages;
1477 bool mne_swab = false;
1483 /* allocate buffer for bulk transfer.
1484 * if this is the first time for this mgs to read logs,
1485 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1486 * once; otherwise, it only reads increment of logs, this should be
1487 * small and CONFIG_READ_NRPAGES will be used.
1489 nrpages = CONFIG_READ_NRPAGES;
1490 if (cfg->cfg_last_idx == 0) /* the first time */
1491 nrpages = CONFIG_READ_NRPAGES_INIT;
1493 OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1495 GOTO(out, rc = -ENOMEM);
1497 for (i = 0; i < nrpages; i++) {
1498 pages[i] = alloc_page(GFP_IOFS);
1499 if (pages[i] == NULL)
1500 GOTO(out, rc = -ENOMEM);
1504 LASSERT(cld_is_recover(cld));
1505 LASSERT(mutex_is_locked(&cld->cld_lock));
1506 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1507 &RQF_MGS_CONFIG_READ);
1509 GOTO(out, rc = -ENOMEM);
1511 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1516 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1517 LASSERT(body != NULL);
1518 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1519 if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1520 >= sizeof(body->mcb_name))
1521 GOTO(out, rc = -E2BIG);
1522 body->mcb_offset = cfg->cfg_last_idx + 1;
1523 body->mcb_type = cld->cld_type;
1524 body->mcb_bits = PAGE_CACHE_SHIFT;
1525 body->mcb_units = nrpages;
1527 /* allocate bulk transfer descriptor */
1528 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1, BULK_PUT_SINK,
1531 GOTO(out, rc = -ENOMEM);
1533 for (i = 0; i < nrpages; i++)
1534 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
1536 ptlrpc_request_set_replen(req);
1537 rc = ptlrpc_queue_wait(req);
1541 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1542 if (res->mcr_size < res->mcr_offset)
1543 GOTO(out, rc = -EINVAL);
1545 /* always update the index even though it might have errors with
1546 * handling the recover logs */
1547 cfg->cfg_last_idx = res->mcr_offset;
1548 eof = res->mcr_offset == res->mcr_size;
1550 CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n",
1551 res->mcr_offset, eof == false);
1553 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1555 GOTO(out, rc = ealen);
1557 if (ealen > nrpages << PAGE_CACHE_SHIFT)
1558 GOTO(out, rc = -EINVAL);
1560 if (ealen == 0) { /* no logs transferred */
1566 mne_swab = !!ptlrpc_rep_need_swab(req);
1567 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1568 /* This import flag means the server did an extra swab of IR MNE
1569 * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1570 if (unlikely(req->rq_import->imp_need_mne_swab))
1571 mne_swab = !mne_swab;
1573 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1576 for (i = 0; i < nrpages && ealen > 0; i++) {
1580 ptr = kmap(pages[i]);
1581 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1582 min_t(int, ealen, PAGE_CACHE_SIZE),
1586 CWARN("Process recover log %s error %d\n",
1587 cld->cld_logname, rc2);
1591 ealen -= PAGE_CACHE_SIZE;
1596 ptlrpc_req_finished(req);
1598 if (rc == 0 && !eof)
1602 for (i = 0; i < nrpages; i++) {
1603 if (pages[i] == NULL)
1605 __free_page(pages[i]);
1607 OBD_FREE(pages, sizeof(*pages) * nrpages);
1612 /* Copy a remote log locally */
1613 static int mgc_llog_local_copy(const struct lu_env *env,
1614 struct obd_device *obd,
1615 struct llog_ctxt *rctxt,
1616 struct llog_ctxt *lctxt, char *logname)
1624 * - copy it to backup using llog_backup()
1625 * - copy remote llog to logname using llog_backup()
1626 * - if failed then move bakup to logname again
1629 OBD_ALLOC(temp_log, strlen(logname) + 1);
1632 sprintf(temp_log, "%sT", logname);
1634 /* make a copy of local llog at first */
1635 rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1636 if (rc < 0 && rc != -ENOENT)
1638 /* copy remote llog to the local copy */
1639 rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1640 if (rc == -ENOENT) {
1641 /* no remote llog, delete local one too */
1642 llog_erase(env, lctxt, NULL, logname);
1643 } else if (rc < 0) {
1644 /* error during backup, get local one back from the copy */
1645 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1647 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1648 obd->obd_name, logname, rc);
1650 llog_erase(env, lctxt, NULL, temp_log);
1651 OBD_FREE(temp_log, strlen(logname) + 1);
1655 /* local_only means it cannot get remote llogs */
1656 static int mgc_process_cfg_log(struct obd_device *mgc,
1657 struct config_llog_data *cld, int local_only)
1659 struct llog_ctxt *ctxt, *lctxt = NULL;
1660 struct client_obd *cli = &mgc->u.cli;
1661 struct lustre_sb_info *lsi = NULL;
1663 bool sptlrpc_started = false;
1669 LASSERT(mutex_is_locked(&cld->cld_lock));
1672 * local copy of sptlrpc log is controlled elsewhere, don't try to
1675 if (cld_is_sptlrpc(cld) && local_only)
1678 if (cld->cld_cfg.cfg_sb)
1679 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1685 rc = lu_env_init(env, LCT_MG_THREAD);
1689 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1692 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1694 /* Copy the setup log locally if we can. Don't mess around if we're
1695 * running an MGS though (logs are already local). */
1696 if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1697 cli->cl_mgc_configs_dir != NULL &&
1698 lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1701 /* Only try to copy log if we have the lock. */
1702 rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1704 if (local_only || rc) {
1705 if (llog_is_empty(env, lctxt, cld->cld_logname)) {
1706 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1707 "log %s and no local copy."
1708 "\n", cld->cld_logname);
1709 GOTO(out_pop, rc = -ENOTCONN);
1711 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1712 "copy for now, will try to update later.\n",
1715 /* Now, whether we copied or not, start using the local llog.
1716 * If we failed to copy, we'll start using whatever the old
1718 llog_ctxt_put(ctxt);
1722 if (local_only) /* no local log at client side */
1723 GOTO(out_pop, rc = -EIO);
1726 if (cld_is_sptlrpc(cld)) {
1727 sptlrpc_conf_log_update_begin(cld->cld_logname);
1728 sptlrpc_started = true;
1731 /* logname and instance info should be the same, so use our
1732 * copy of the instance for the update. The cfg_last_idx will
1733 * be updated here. */
1734 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1739 __llog_ctxt_put(env, ctxt);
1741 __llog_ctxt_put(env, lctxt);
1744 * update settings on existing OBDs. doing it inside
1745 * of llog_process_lock so no device is attaching/detaching
1747 * the logname must be <fsname>-sptlrpc
1749 if (sptlrpc_started) {
1750 LASSERT(cld_is_sptlrpc(cld));
1751 sptlrpc_conf_log_update_end(cld->cld_logname);
1752 class_notify_sptlrpc_conf(cld->cld_logname,
1753 strlen(cld->cld_logname) -
1754 strlen("-sptlrpc"));
1763 /** Get a config log from the MGS and process it.
1764 * This func is called for both clients and servers.
1765 * Copy the log locally before parsing it if appropriate (non-MGS server)
1767 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1769 struct lustre_handle lockh = { 0 };
1770 __u64 flags = LDLM_FL_NO_LRU;
1776 /* I don't want multiple processes running process_log at once --
1777 sounds like badness. It actually might be fine, as long as
1778 we're not trying to update from the same log
1779 simultaneously (in which case we should use a per-log sem.) */
1780 mutex_lock(&cld->cld_lock);
1781 if (cld->cld_stopping) {
1782 mutex_unlock(&cld->cld_lock);
1786 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1788 CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1789 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1791 /* Get the cfg lock on the llog */
1792 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1793 LCK_CR, &flags, NULL, NULL, NULL,
1794 cld, 0, NULL, &lockh);
1796 /* Get the cld, it will be released in mgc_blocking_ast. */
1797 config_log_get(cld);
1798 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1801 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1803 /* mark cld_lostlock so that it will requeue
1804 * after MGC becomes available. */
1805 cld->cld_lostlock = 1;
1806 /* Get extra reference, it will be put in requeue thread */
1807 config_log_get(cld);
1811 if (cld_is_recover(cld)) {
1812 rc = 0; /* this is not a fatal error for recover log */
1814 rc = mgc_process_recover_log(mgc, cld);
1816 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1819 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1820 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1822 mutex_unlock(&cld->cld_lock);
1824 /* Now drop the lock so MGS can revoke it */
1826 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1829 CERROR("Can't drop cfg lock: %d\n", rcl);
1836 /** Called from lustre_process_log.
1837 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1838 * any services, and adds it to the list logs to watch (follow).
1840 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1842 struct lustre_cfg *lcfg = buf;
1843 struct config_llog_instance *cfg = NULL;
1848 switch(lcfg->lcfg_command) {
1849 case LCFG_LOV_ADD_OBD: {
1850 /* Overloading this cfg command: register a new target */
1851 struct mgs_target_info *mti;
1853 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1854 sizeof(struct mgs_target_info))
1855 GOTO(out, rc = -EINVAL);
1857 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1858 CDEBUG(D_MGC, "add_target %s %#x\n",
1859 mti->mti_svname, mti->mti_flags);
1860 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1863 case LCFG_LOV_DEL_OBD:
1864 /* Unregister has no meaning at the moment. */
1865 CERROR("lov_del_obd unimplemented\n");
1868 case LCFG_SPTLRPC_CONF: {
1869 rc = sptlrpc_process_config(lcfg);
1872 case LCFG_LOG_START: {
1873 struct config_llog_data *cld;
1874 struct super_block *sb;
1876 logname = lustre_cfg_string(lcfg, 1);
1877 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1878 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1880 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1883 /* We're only called through here on the initial mount */
1884 rc = config_log_add(obd, logname, cfg, sb);
1887 cld = config_log_find(logname, cfg);
1894 /* FIXME only set this for old logs! Right now this forces
1895 us to always skip the "inside markers" check */
1896 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1898 rc = mgc_process_log(obd, cld);
1899 if (rc == 0 && cld->cld_recover != NULL) {
1900 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1901 imp_connect_data, IMP_RECOV)) {
1902 rc = mgc_process_log(obd, cld->cld_recover);
1904 struct config_llog_data *cir = cld->cld_recover;
1905 cld->cld_recover = NULL;
1906 config_log_put(cir);
1909 CERROR("Cannot process recover llog %d\n", rc);
1911 config_log_put(cld);
1915 case LCFG_LOG_END: {
1916 logname = lustre_cfg_string(lcfg, 1);
1918 if (lcfg->lcfg_bufcount >= 2)
1919 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1921 rc = config_log_end(logname, cfg);
1925 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1926 GOTO(out, rc = -EINVAL);
1934 struct obd_ops mgc_obd_ops = {
1935 .o_owner = THIS_MODULE,
1936 .o_setup = mgc_setup,
1937 .o_precleanup = mgc_precleanup,
1938 .o_cleanup = mgc_cleanup,
1939 .o_add_conn = client_import_add_conn,
1940 .o_del_conn = client_import_del_conn,
1941 .o_connect = client_connect_import,
1942 .o_disconnect = client_disconnect_export,
1943 //.o_enqueue = mgc_enqueue,
1944 .o_cancel = mgc_cancel,
1945 //.o_iocontrol = mgc_iocontrol,
1946 .o_set_info_async = mgc_set_info_async,
1947 .o_get_info = mgc_get_info,
1948 .o_import_event = mgc_import_event,
1949 .o_process_config = mgc_process_config,
1952 int __init mgc_init(void)
1954 return class_register_type(&mgc_obd_ops, NULL, NULL,
1955 LUSTRE_MGC_NAME, NULL);
1959 static void /*__exit*/ mgc_exit(void)
1961 class_unregister_type(LUSTRE_MGC_NAME);
1964 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1965 MODULE_DESCRIPTION("Lustre Management Client");
1966 MODULE_LICENSE("GPL");
1968 module_init(mgc_init);
1969 module_exit(mgc_exit);