4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lustre/mgc/mgc_request.c
33 * Author: Nathan Rutman <nathan@clusterfs.com>
36 #define DEBUG_SUBSYSTEM S_MGC
37 #define D_MGC D_CONFIG /*|D_WARNING*/
39 #include <linux/module.h>
40 #include <linux/random.h>
42 #include <lprocfs_status.h>
43 #include <lustre_dlm.h>
44 #include <lustre_disk.h>
45 #include <lustre_log.h>
46 #include <lustre_swab.h>
47 #include <obd_class.h>
49 #include "mgc_internal.h"
51 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
52 enum mgs_cfg_type type)
56 if (len > sizeof(resname)) {
57 CERROR("name too long: %s\n", name);
61 CERROR("missing name: %s\n", name);
64 memcpy(&resname, name, len);
66 /* Always use the same endianness for the resid */
67 memset(res_id, 0, sizeof(*res_id));
68 res_id->name[0] = cpu_to_le64(resname);
69 /* XXX: unfortunately, sptlprc and config llog share one lock */
71 case MGS_CFG_T_CONFIG:
72 case MGS_CFG_T_SPTLRPC:
75 case MGS_CFG_T_RECOVER:
76 case MGS_CFG_T_PARAMS:
77 #ifdef HAVE_SERVER_SUPPORT
78 case MGS_CFG_T_NODEMAP:
79 case MGS_CFG_T_BARRIER:
86 res_id->name[1] = cpu_to_le64(resname);
87 CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
88 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
92 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id,
93 enum mgs_cfg_type type)
95 /* fsname is at most 8 chars long, can contain "-". eg. lustre, lu-0 */
96 return mgc_name2resid(fsname, strlen(fsname), res_id, type);
98 EXPORT_SYMBOL(mgc_fsname2resid);
100 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id,
101 enum mgs_cfg_type type)
106 /* logname consists of "fsname-nodetype". eg. "lustre-MDT0001",
107 * "SUN-000-client" there is an exception: llog "params"
109 name_end = strrchr(logname, '-');
111 len = strlen(logname);
113 len = name_end - logname;
114 return mgc_name2resid(logname, len, res_id, type);
116 EXPORT_SYMBOL(mgc_logname2resid);
118 /********************** config llog list **********************/
119 static LIST_HEAD(config_llog_list);
120 static DEFINE_SPINLOCK(config_list_lock); /* protects config_llog_list */
122 /* Take a reference to a config log */
123 static int config_log_get(struct config_llog_data *cld)
126 refcount_inc(&cld->cld_refcount);
127 CDEBUG(D_INFO, "log %s (%p) refs %d\n", cld->cld_logname, cld,
128 refcount_read(&cld->cld_refcount));
132 /* Drop a reference to a config log. When no longer referenced, We can free the
135 static void config_log_put(struct config_llog_data *cld)
142 CDEBUG(D_INFO, "log %s(%p) refs %d\n", cld->cld_logname, cld,
143 refcount_read(&cld->cld_refcount));
145 /* spinlock to make sure no item with 0 refcount in the list */
146 if (refcount_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
147 list_del(&cld->cld_list_chain);
148 spin_unlock(&config_list_lock);
150 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
151 #ifdef HAVE_SERVER_SUPPORT
152 config_log_put(cld->cld_barrier);
153 config_log_put(cld->cld_nodemap);
155 config_log_put(cld->cld_recover);
156 config_log_put(cld->cld_params);
157 config_log_put(cld->cld_sptlrpc);
158 if (cld_is_sptlrpc(cld)) {
159 cld->cld_stopping = 1;
160 sptlrpc_conf_log_stop(cld->cld_logname);
163 class_export_put(cld->cld_mgcexp);
164 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
170 /* Find a config log by name */
172 struct config_llog_data *config_log_find(char *logname,
173 struct config_llog_instance *cfg)
175 struct config_llog_data *cld;
176 struct config_llog_data *found = NULL;
177 unsigned long cfg_instance;
180 LASSERT(logname != NULL);
182 cfg_instance = cfg ? cfg->cfg_instance : 0;
183 spin_lock(&config_list_lock);
184 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
185 /* check if cfg_instance is the one we want */
186 if (cfg_instance != cld->cld_cfg.cfg_instance)
189 /* instance may be NULL, should check name */
190 if (strcmp(logname, cld->cld_logname) == 0) {
192 config_log_get(found);
196 spin_unlock(&config_list_lock);
201 struct config_llog_data *do_config_log_add(struct obd_device *obd,
203 enum mgs_cfg_type type,
204 struct config_llog_instance *cfg,
205 struct super_block *sb)
207 struct config_llog_data *cld;
212 CDEBUG(D_MGC, "do adding config log %s-%016lx\n", logname,
213 cfg ? cfg->cfg_instance : 0);
215 OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
217 RETURN(ERR_PTR(-ENOMEM));
219 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
221 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
225 strcpy(cld->cld_logname, logname);
229 cld->cld_cfg.cfg_callback = class_config_llog_handler;
230 mutex_init(&cld->cld_lock);
231 cld->cld_cfg.cfg_last_idx = 0;
232 cld->cld_cfg.cfg_flags = 0;
233 cld->cld_cfg.cfg_sb = sb;
234 cld->cld_type = type;
235 refcount_set(&cld->cld_refcount, 1);
237 /* Keep the mgc around until we are done */
238 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
240 if (cld_is_sptlrpc(cld))
241 sptlrpc_conf_log_start(logname);
243 spin_lock(&config_list_lock);
244 list_add(&cld->cld_list_chain, &config_llog_list);
245 spin_unlock(&config_list_lock);
247 if (cld_is_sptlrpc(cld) || cld_is_nodemap(cld) || cld_is_barrier(cld)) {
248 rc = mgc_process_log(obd, cld);
249 if (rc && rc != -ENOENT)
250 CERROR("%s: failed processing log, type %d: rc = %d\n",
251 obd->obd_name, type, rc);
257 static struct config_llog_data *
258 config_recover_log_add(struct obd_device *obd, char *fsname,
259 struct config_llog_instance *cfg,
260 struct super_block *sb)
262 struct config_llog_instance lcfg = *cfg;
263 struct config_llog_data *cld;
264 bool is_server = IS_SERVER(s2lsi(sb));
267 #ifdef HAVE_SERVER_SUPPORT
268 if (IS_OST(s2lsi(sb)))
271 /* for osp-on-ost, see lustre_start_osp() */
272 if (IS_MDT(s2lsi(sb)) && lcfg.cfg_instance)
275 /* We have to use different llog for clients and MDTs for DNE,
276 * where only clients are notified if one of DNE server restarts.
278 LASSERT(strlen(fsname) < sizeof(logname) / 2);
279 strncpy(logname, fsname, sizeof(logname));
281 LASSERT(is_server ? lcfg.cfg_instance == 0 : lcfg.cfg_instance != 0);
283 lcfg.cfg_instance = ll_get_cfg_instance(sb);
284 scnprintf(logname, sizeof(logname), "%s-%s", fsname,
285 is_server ? "mdtir" : "cliir");
287 cld = do_config_log_add(obd, logname, MGS_CFG_T_RECOVER, &lcfg, sb);
291 static struct config_llog_data *
292 config_log_find_or_add(struct obd_device *obd, char *logname,
293 struct super_block *sb, enum mgs_cfg_type type,
294 struct config_llog_instance *cfg)
296 struct config_llog_instance lcfg = *cfg;
297 struct config_llog_data *cld;
299 /* Note class_config_llog_handler() depends on getting "obd" back */
300 lcfg.cfg_instance = sb ? ll_get_cfg_instance(sb) : (unsigned long)obd;
302 cld = config_log_find(logname, &lcfg);
303 if (unlikely(cld != NULL))
306 return do_config_log_add(obd, logname, type, &lcfg, sb);
309 /** Add this log to the list of active logs watched by an MGC.
310 * Active means we're watching for updates.
311 * We have one active log per "mount" - client instance or servername.
312 * Each instance may be at a different point in the log.
314 static struct config_llog_data *
315 config_log_add(struct obd_device *obd, char *logname,
316 struct config_llog_instance *cfg, struct super_block *sb)
318 struct lustre_sb_info *lsi = s2lsi(sb);
319 struct config_llog_data *cld = NULL;
320 struct config_llog_data *sptlrpc_cld = NULL;
321 struct config_llog_data *params_cld = NULL;
322 #ifdef HAVE_SERVER_SUPPORT
323 struct config_llog_data *nodemap_cld = NULL;
324 struct config_llog_data *barrier_cld = NULL;
333 CDEBUG(D_MGC, "add config log %s-%016lx\n", logname,
337 * for each regular log, the depended sptlrpc log name is
338 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
340 ptr = strrchr(logname, '-');
341 if (ptr == NULL || ptr - logname > 8) {
342 CERROR("logname %s is too long\n", logname);
343 RETURN(ERR_PTR(-EINVAL));
346 memcpy(seclogname, logname, ptr - logname);
347 strcpy(seclogname + (ptr - logname), "-sptlrpc");
349 if (cfg->cfg_sub_clds & CONFIG_SUB_SPTLRPC) {
350 sptlrpc_cld = config_log_find_or_add(obd, seclogname, NULL,
351 MGS_CFG_T_SPTLRPC, cfg);
352 if (IS_ERR(sptlrpc_cld)) {
353 rc = PTR_ERR(sptlrpc_cld);
354 CERROR("%s: can't create sptlrpc log %s: rc = %d\n",
355 obd->obd_name, seclogname, rc);
360 if (cfg->cfg_sub_clds & CONFIG_SUB_PARAMS) {
361 params_cld = config_log_find_or_add(obd, PARAMS_FILENAME, sb,
362 MGS_CFG_T_PARAMS, cfg);
363 if (IS_ERR(params_cld)) {
364 rc = PTR_ERR(params_cld);
365 CERROR("%s: can't create params log: rc = %d\n",
367 GOTO(out_sptlrpc, rc);
371 #ifdef HAVE_SERVER_SUPPORT
372 if (!IS_MGS(lsi) && cfg->cfg_sub_clds & CONFIG_SUB_NODEMAP) {
373 nodemap_cld = config_log_find_or_add(obd, LUSTRE_NODEMAP_NAME,
374 NULL, MGS_CFG_T_NODEMAP,
376 if (IS_ERR(nodemap_cld)) {
377 rc = PTR_ERR(nodemap_cld);
378 CERROR("%s: cannot create nodemap log: rc = %d\n",
380 GOTO(out_params, rc);
384 if (IS_MDT(s2lsi(sb)) && cfg->cfg_sub_clds & CONFIG_SUB_BARRIER) {
385 snprintf(seclogname + (ptr - logname), sizeof(seclogname) - 1,
386 "-%s", BARRIER_FILENAME);
387 barrier_cld = config_log_find_or_add(obd, seclogname, sb,
388 MGS_CFG_T_BARRIER, cfg);
389 if (IS_ERR(barrier_cld)) {
390 rc = PTR_ERR(barrier_cld);
391 CERROR("%s: can't create barrier log: rc = %d\n",
393 GOTO(out_nodemap, rc);
397 cld = do_config_log_add(obd, logname, MGS_CFG_T_CONFIG, cfg, sb);
400 CERROR("%s: can't create log: rc = %d\n",
402 #ifdef HAVE_SERVER_SUPPORT
403 GOTO(out_barrier, rc);
405 GOTO(out_params, rc);
409 LASSERT(lsi->lsi_lmd);
410 if (!test_bit(LMD_FLG_NOIR, lsi->lsi_lmd->lmd_flags) &&
411 cfg->cfg_sub_clds & CONFIG_SUB_RECOVER) {
412 struct config_llog_data *recover_cld;
414 ptr = strrchr(seclogname, '-');
418 CERROR("%s: sptlrpc log name not correct, %s: rc = %d\n",
419 obd->obd_name, seclogname, -EINVAL);
420 GOTO(out_cld, rc = -EINVAL);
423 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
424 if (IS_ERR(recover_cld)) {
425 rc = PTR_ERR(recover_cld);
426 CERROR("%s: can't create recover log: rc = %d\n",
431 mutex_lock(&cld->cld_lock);
433 cld->cld_recover = recover_cld;
437 mutex_lock(&cld->cld_lock);
438 #ifdef HAVE_SERVER_SUPPORT
439 cld->cld_barrier = barrier_cld;
440 cld->cld_nodemap = nodemap_cld;
442 cld->cld_params = params_cld;
443 cld->cld_sptlrpc = sptlrpc_cld;
444 mutex_unlock(&cld->cld_lock);
450 #ifdef HAVE_SERVER_SUPPORT
452 config_log_put(barrier_cld);
454 config_log_put(nodemap_cld);
457 config_log_put(params_cld);
459 config_log_put(sptlrpc_cld);
464 DEFINE_MUTEX(llog_process_lock);
466 static inline void config_mark_cld_stop_nolock(struct config_llog_data *cld)
470 spin_lock(&config_list_lock);
471 cld->cld_stopping = 1;
472 spin_unlock(&config_list_lock);
474 CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
475 if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
476 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
479 static inline void config_mark_cld_stop(struct config_llog_data *cld)
482 mutex_lock(&cld->cld_lock);
483 config_mark_cld_stop_nolock(cld);
484 mutex_unlock(&cld->cld_lock);
488 /** Stop watching for updates on this log.
490 static int config_log_end(char *logname, struct config_llog_instance *cfg)
492 struct config_llog_data *cld;
493 struct config_llog_data *cld_sptlrpc = NULL;
494 struct config_llog_data *cld_params = NULL;
495 struct config_llog_data *cld_recover = NULL;
496 struct config_llog_data *cld_nodemap = NULL;
497 struct config_llog_data *cld_barrier = NULL;
502 cld = config_log_find(logname, cfg);
506 mutex_lock(&cld->cld_lock);
508 * if cld_stopping is set, it means we didn't start the log thus
509 * not owning the start ref. this can happen after previous umount:
510 * the cld still hanging there waiting for lock cancel, and we
511 * remount again but failed in the middle and call log_end without
514 if (unlikely(cld->cld_stopping)) {
515 mutex_unlock(&cld->cld_lock);
516 /* drop the ref from the find */
521 cld_recover = cld->cld_recover;
522 cld->cld_recover = NULL;
523 cld_params = cld->cld_params;
524 cld->cld_params = NULL;
525 cld_nodemap = cld->cld_nodemap;
526 cld->cld_nodemap = NULL;
527 cld_barrier = cld->cld_barrier;
528 cld->cld_barrier = NULL;
529 cld_sptlrpc = cld->cld_sptlrpc;
530 cld->cld_sptlrpc = NULL;
532 config_mark_cld_stop_nolock(cld);
533 mutex_unlock(&cld->cld_lock);
535 config_mark_cld_stop(cld_recover);
536 config_log_put(cld_recover);
537 config_mark_cld_stop(cld_params);
538 config_log_put(cld_params);
539 config_mark_cld_stop(cld_barrier);
540 config_log_put(cld_barrier);
541 /* don't explicitly set cld_stopping on sptlrpc lock here, as other
542 * targets may be active, it will be done in config_log_put if necessary
544 config_log_put(cld_sptlrpc);
545 /* don't set cld_stopping on nm lock as other targets may be active */
546 config_log_put(cld_nodemap);
548 /* drop the ref from the find */
550 /* drop the start ref */
553 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
558 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
560 struct obd_device *obd = data;
561 struct obd_import *imp;
562 struct obd_connect_data *ocd;
563 struct config_llog_data *cld;
568 with_imp_locked(obd, imp, rc) {
569 ocd = &imp->imp_connect_data;
571 seq_printf(m, "imperative_recovery: %s\n",
572 OCD_HAS_FLAG(ocd, IMP_RECOV) ?
573 "ENABLED" : "DISABLED");
578 seq_puts(m, "client_state:\n");
580 spin_lock(&config_list_lock);
581 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
582 if (cld->cld_recover == NULL)
584 seq_printf(m, " - { client: %s, nidtbl_version: %u }\n",
586 cld->cld_recover->cld_cfg.cfg_last_idx);
588 spin_unlock(&config_list_lock);
593 /* reenqueue any lost locks */
594 #define RQ_RUNNING 0x1
598 #define RQ_PRECLEANUP 0x10
600 static wait_queue_head_t rq_waitq;
601 static DECLARE_COMPLETION(rq_exit);
602 static DECLARE_COMPLETION(rq_start);
604 static void do_requeue(struct config_llog_data *cld)
611 * Do not run mgc_process_log on a disconnected export or an
612 * export which is being disconnected. Take the client
613 * semaphore to make the check non-racy.
615 down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
617 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
618 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
619 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
620 if (rc && rc != -ENOENT)
621 CERROR("failed processing log: %d\n", rc);
623 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
626 up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
631 static int mgc_requeue_thread(void *data)
638 CDEBUG(D_MGC, "Starting requeue thread\n");
640 /* Keep trying failed locks periodically */
641 spin_lock(&config_list_lock);
642 rq_state |= RQ_RUNNING;
643 while (!(rq_state & RQ_STOP)) {
644 struct config_llog_data *cld, *cld_prev;
647 /* Any new or requeued lostlocks will change the state */
648 rq_state &= ~(RQ_NOW | RQ_LATER);
649 spin_unlock(&config_list_lock);
656 /* Always wait a few seconds to allow the server who
657 * caused the lock revocation to finish its setup, plus some
658 * random so everyone doesn't try to reconnect at once.
660 to = mgc_requeue_timeout_min == 0 ? 1 : mgc_requeue_timeout_min;
661 to = cfs_time_seconds(mgc_requeue_timeout_min) +
662 get_random_u32_below(cfs_time_seconds(to));
663 wait_event_idle_timeout(rq_waitq,
664 rq_state & (RQ_STOP | RQ_PRECLEANUP),
668 * iterate & processing through the list. for each cld, process
669 * its depending sptlrpc cld firstly (if any) and then itself.
671 * it's guaranteed any item in the list must have
672 * reference > 0; and if cld_lostlock is set, at
673 * least one reference is taken by the previous enqueue.
677 spin_lock(&config_list_lock);
678 rq_state &= ~RQ_PRECLEANUP;
679 list_for_each_entry(cld, &config_llog_list,
681 if (!cld->cld_lostlock || cld->cld_stopping)
684 /* hold reference to avoid being freed during
685 * subsequent processing.
688 cld->cld_lostlock = 0;
689 spin_unlock(&config_list_lock);
691 config_log_put(cld_prev);
694 if (likely(!(rq_state & RQ_STOP))) {
696 spin_lock(&config_list_lock);
698 spin_lock(&config_list_lock);
702 spin_unlock(&config_list_lock);
703 config_log_put(cld_prev);
705 /* Wait a bit to see if anyone else needs a requeue */
706 wait_event_idle(rq_waitq, rq_state & (RQ_NOW | RQ_STOP));
707 spin_lock(&config_list_lock);
710 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
711 rq_state &= ~RQ_RUNNING;
712 spin_unlock(&config_list_lock);
716 CDEBUG(D_MGC, "Ending requeue thread\n");
720 /* Add a cld to the list to requeue. Start the requeue thread if needed.
721 * We are responsible for dropping the config log reference from here on out.
723 static void mgc_requeue_add(struct config_llog_data *cld)
729 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
730 cld->cld_logname, refcount_read(&cld->cld_refcount),
731 cld->cld_stopping, rq_state);
733 /* lets cancel an existent lock to mark cld as "lostlock" */
734 CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
735 if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
736 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
738 mutex_lock(&cld->cld_lock);
739 spin_lock(&config_list_lock);
740 if (!(rq_state & RQ_STOP) && !cld->cld_stopping) {
741 cld->cld_lostlock = 1;
745 spin_unlock(&config_list_lock);
746 mutex_unlock(&cld->cld_lock);
753 /********************** class fns **********************/
754 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
756 struct llog_ctxt *ctxt;
761 /* setup only remote ctxt, the local disk context is switched per each
762 * filesystem during mgc_fs_setup()
764 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
769 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
772 llog_initiator_connect(ctxt);
778 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
780 struct llog_ctxt *ctxt;
784 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
786 llog_cleanup(env, ctxt);
792 static atomic_t mgc_count = ATOMIC_INIT(0);
793 static int mgc_precleanup(struct obd_device *obd)
800 if (atomic_dec_and_test(&mgc_count)) {
801 LASSERT(rq_state & RQ_RUNNING);
802 /* stop requeue thread */
805 /* wakeup requeue thread to clean our cld */
806 temp = RQ_NOW | RQ_PRECLEANUP;
809 spin_lock(&config_list_lock);
811 spin_unlock(&config_list_lock);
815 wait_for_completion(&rq_exit);
816 obd_cleanup_client_import(obd);
818 rc = mgc_llog_fini(NULL, obd);
820 CERROR("failed to cleanup llogging subsystems\n");
825 static int mgc_cleanup(struct obd_device *obd)
831 /* COMPAT_146 - old config logs may have added profiles secretly */
832 if (atomic_read(&obd->obd_type->typ_refcnt) <= 1)
833 /* Only for the last mgc */
834 class_del_profiles();
836 lprocfs_obd_cleanup(obd);
839 rc = client_obd_cleanup(obd);
843 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
845 struct task_struct *task;
850 rc = ptlrpcd_addref();
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 rc = mgc_tunables_init(obd);
868 if (atomic_inc_return(&mgc_count) == 1) {
870 init_waitqueue_head(&rq_waitq);
872 /* start requeue thread */
873 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
876 CERROR("%s: cannot start requeue thread: rc = %d; "
877 "no more log updates\n",
881 /* rc is the task_struct pointer of mgc_requeue_thread. */
883 wait_for_completion(&rq_start);
889 lprocfs_obd_cleanup(obd);
891 client_obd_cleanup(obd);
897 /* based on ll_mdc_blocking_ast */
898 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
899 void *data, int flag)
901 struct lustre_handle lockh;
902 struct config_llog_data *cld = (struct config_llog_data *)data;
908 case LDLM_CB_BLOCKING:
909 /* mgs wants the lock, give it up... */
910 LDLM_DEBUG(lock, "MGC blocking CB");
911 ldlm_lock2handle(lock, &lockh);
912 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
914 case LDLM_CB_CANCELING:
915 /* We've given up the lock, prepare ourselves to update. */
916 LDLM_DEBUG(lock, "MGC cancel CB");
918 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
919 PLDLMRES(lock->l_resource),
920 (char *)&lock->l_resource->lr_name.name[0]);
923 CDEBUG(D_INFO, "missing data, won't requeue\n");
927 lock->l_ast_data = NULL;
928 cld->cld_lockh.cookie = 0;
929 /* Are we done with this log? */
930 if (cld->cld_stopping) {
931 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
936 /* Make sure not to re-enqueue when the mgc is stopping
937 * (we get called from client_disconnect_export)
939 if (lock->l_conn_export == NULL ||
940 lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
941 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
948 mgc_requeue_add(cld);
958 /* Take a config lock so we can get cancel notifications */
959 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
960 union ldlm_policy_data *policy, enum ldlm_mode mode,
961 __u64 *flags, ldlm_glimpse_callback glimpse_callback,
962 void *data, __u32 lvb_len, void *lvb_swabber,
963 struct lustre_handle *lockh)
965 struct config_llog_data *cld = (struct config_llog_data *)data;
966 struct ldlm_enqueue_info einfo = {
969 .ei_cb_bl = mgc_blocking_ast,
970 .ei_cb_cp = ldlm_completion_ast,
971 .ei_cb_gl = glimpse_callback,
973 struct ptlrpc_request *req;
974 int short_limit = cld_is_sptlrpc(cld);
982 CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
983 cld->cld_resid.name[0]);
985 /* We need a callback for every lockholder, so don't try to
986 * ldlm_lock_match (see rev 1.1.2.11.2.47)
988 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
989 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
994 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
995 ptlrpc_request_set_replen(req);
997 /* check if this is server or client */
998 if (cld->cld_cfg.cfg_sb &&
999 IS_SERVER(s2lsi(cld->cld_cfg.cfg_sb)))
1002 /* Limit how long we will wait for the enqueue to complete */
1003 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT(exp->exp_obd);
1004 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1005 NULL, 0, LVB_T_NONE, lockh, 0);
1006 /* A failed enqueue should still call the mgc_blocking_ast,
1007 * where it will be requeued if needed ("grant failed").
1009 ptlrpc_req_finished(req);
1013 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1014 struct lustre_handle *lockh)
1018 ldlm_lock_decref(lockh, mode);
1023 static void mgc_notify_active(struct obd_device *unused)
1025 /* wakeup mgc_requeue_thread to requeue mgc lock */
1026 spin_lock(&config_list_lock);
1028 spin_unlock(&config_list_lock);
1031 /* TODO: Help the MGS rebuild nidtbl. -jay */
1034 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1035 u32 keylen, void *key,
1036 u32 vallen, void *val,
1037 struct ptlrpc_request_set *set)
1043 /* Turn off initial_recov after we try all backup servers once */
1044 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1045 struct obd_import *imp = class_exp2cliimp(exp);
1048 if (vallen != sizeof(int))
1050 value = *(int *)val;
1051 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1052 imp->imp_obd->obd_name, value,
1053 imp->imp_deactive, imp->imp_invalid,
1054 imp->imp_replayable, imp->imp_obd->obd_replayable,
1055 ptlrpc_import_state_name(imp->imp_state));
1056 /* Resurrect the import immediately if
1057 * 1. we previously got disconnected,
1058 * 2. value > 1 (at the same node with MGS)
1061 struct adaptive_timeout *at;
1063 at = &imp->imp_at.iat_net_latency;
1064 at_reset(at, INITIAL_CONNECT_TIMEOUT);
1067 if (imp->imp_state == LUSTRE_IMP_DISCON || value > 1)
1068 ptlrpc_reconnect_import(imp);
1073 if (KEY_IS(KEY_MGSSEC)) {
1074 struct client_obd *cli = &exp->exp_obd->u.cli;
1075 struct sptlrpc_flavor flvr;
1078 * empty string means using current flavor, if which haven't
1079 * been set yet, set it as null.
1081 * if flavor has been set previously, check the asking flavor
1082 * must match the existing one.
1085 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1091 rc = sptlrpc_parse_flavor(val, &flvr);
1093 CERROR("invalid sptlrpc flavor %s to MGS\n",
1098 /* caller already hold a mutex */
1099 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1100 cli->cl_flvr_mgc = flvr;
1101 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1102 sizeof(flvr)) != 0) {
1105 sptlrpc_flavor2name(&cli->cl_flvr_mgc, str,
1107 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but currently %s is in use\n",
1114 #ifdef HAVE_SERVER_SUPPORT
1115 rc = mgc_set_info_async_server(env, exp, keylen, key, vallen, val, set);
1120 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1121 __u32 keylen, void *key, __u32 *vallen, void *val)
1125 if (KEY_IS(KEY_CONN_DATA)) {
1126 struct obd_import *imp = class_exp2cliimp(exp);
1127 struct obd_connect_data *data = val;
1129 if (*vallen == sizeof(*data)) {
1130 *data = imp->imp_connect_data;
1138 static int mgc_import_event(struct obd_device *obd,
1139 struct obd_import *imp,
1140 enum obd_import_event event)
1144 LASSERT(imp->imp_obd == obd);
1145 CDEBUG(D_MGC, "import event %#x\n", event);
1148 case IMP_EVENT_DISCON:
1149 /* MGC imports should not wait for recovery */
1150 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1151 ptlrpc_pinger_ir_down();
1153 case IMP_EVENT_INACTIVE:
1155 case IMP_EVENT_INVALIDATE: {
1156 struct ldlm_namespace *ns = obd->obd_namespace;
1158 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1161 case IMP_EVENT_ACTIVE:
1162 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1163 /* Clearing obd_no_recov allows us to continue pinging */
1164 obd->obd_no_recov = 0;
1165 mgc_notify_active(obd);
1166 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1167 ptlrpc_pinger_ir_up();
1171 case IMP_EVENT_DEACTIVATE:
1172 case IMP_EVENT_ACTIVATE:
1175 CERROR("Unknown import event %#x\n", event);
1181 static int mgc_apply_recover_logs(struct obd_device *mgc,
1182 struct config_llog_data *cld,
1184 void *data, int datalen, bool mne_swab)
1186 struct config_llog_instance *cfg = &cld->cld_cfg;
1187 struct lustre_cfg *lcfg;
1188 struct lustre_cfg_bufs bufs;
1189 u64 prev_version = 0;
1190 char inst[MTI_NAME_MAXLEN + 1];
1196 unsigned long dynamic_nids;
1199 LASSERT(cfg->cfg_instance != 0);
1200 LASSERT(ll_get_cfg_instance(cfg->cfg_sb) == cfg->cfg_instance);
1202 /* get dynamic nids setting */
1203 dynamic_nids = mgc->obd_dynamic_nids;
1205 if (!IS_SERVER(s2lsi(cfg->cfg_sb))) {
1206 pos = snprintf(inst, sizeof(inst), "%016lx", cfg->cfg_instance);
1207 if (pos >= PAGE_SIZE)
1209 #ifdef HAVE_SERVER_SUPPORT
1211 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1213 LASSERT(IS_MDT(lsi));
1214 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1218 #endif /* HAVE_SERVER_SUPPORT */
1221 OBD_ALLOC(buf, PAGE_SIZE);
1227 while (datalen > 0) {
1228 struct mgs_nidtbl_entry *entry = (data + off);
1229 struct lnet_nid *nidlist = NULL;
1230 int entry_len = sizeof(*entry);
1231 struct obd_device *obd;
1232 struct obd_import *imp;
1233 struct obd_uuid *uuid;
1241 if (datalen < entry_len) /* really short on data */
1244 /* swab non nid data */
1246 lustre_swab_mgs_nidtbl_entry_header(entry);
1248 if (entry->mne_nid_count == 0) /* at least one nid entry */
1251 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1252 if (datalen < entry_len) /* must have entry_len at least */
1255 if (entry->mne_length > PAGE_SIZE) {
1256 CERROR("MNE too large (%u)\n", entry->mne_length);
1260 /* improper mne_lenth */
1261 if (entry->mne_length < entry_len)
1264 /* entry length reports larger than all the data passed in */
1265 if (datalen < entry->mne_length)
1268 /* Looks OK. Can process this entry? else move to next entry */
1269 off += entry->mne_length;
1270 datalen -= entry->mne_length;
1274 if (entry->mne_version > max_version) {
1275 CERROR("entry index(%lld) is over max_index(%lld)\n",
1276 entry->mne_version, max_version);
1280 if (prev_version >= entry->mne_version) {
1281 CERROR("index unsorted, prev %lld, now %lld\n",
1282 prev_version, entry->mne_version);
1285 prev_version = entry->mne_version;
1287 if (entry->mne_nid_type == 0) {
1288 struct lnet_nid *nid;
1291 OBD_ALLOC_PTR_ARRAY(nidlist, entry->mne_nid_count);
1297 /* Keep this nid data swab for normal mixed
1298 * endian handling. LU-1644
1301 lustre_swab_mgs_nidtbl_entry_content(entry);
1303 /* Turn old NID format to newer format. */
1305 for (i = 0; i < entry->mne_nid_count; i++) {
1306 lnet_nid4_to_nid(entry->u.nids[i], nid);
1307 nid += sizeof(struct lnet_nid);
1310 /* Handle the case if struct lnet_nid is expanded in
1311 * the future. The MGS should prevent this but just
1314 if (entry->mne_nid_size > sizeof(struct lnet_nid))
1317 nidlist = entry->u.nidlist;
1321 * Write a string with format "nid::instance" to
1322 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1324 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1325 memset(buf, 0, bufsz);
1329 /* lustre-OST0001-osc-<instance #> */
1330 strcpy(obdname, cld->cld_logname);
1331 cname = strrchr(obdname, '-');
1332 if (cname == NULL) {
1333 CERROR("mgc %s: invalid logname %s\n",
1334 mgc->obd_name, obdname);
1335 if (entry->mne_nid_type == 0)
1336 OBD_FREE_PTR_ARRAY(nidlist,
1337 entry->mne_nid_count);
1341 pos = cname - obdname;
1343 pos += sprintf(obdname + pos, "-%s%04x",
1344 is_ost ? "OST" : "MDT", entry->mne_index);
1346 cname = is_ost ? "osc" : "mdc",
1347 pos += snprintf(obdname + pos, bufsz, "-%s-%s", cname,
1349 lustre_cfg_bufs_reset(&bufs, obdname);
1351 /* find the obd by obdname */
1352 obd = class_name2obd(obdname);
1354 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1355 mgc->obd_name, obdname);
1357 /* this is a safe race, when the ost is starting up...*/
1361 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1364 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1365 uuid = (struct obd_uuid *)(buf + pos);
1367 with_imp_locked(obd, imp, rc) {
1368 struct obd_uuid server_uuid;
1372 /* iterate all nids to find one */
1373 /* find uuid by nid */
1374 /* create import entries if they don't exist */
1375 rc = client_import_add_nids_to_conn(imp, nidlist,
1376 entry->mne_nid_count,
1377 entry->mne_nid_size,
1379 if (rc != -ENOENT || !dynamic_nids)
1382 /* create a new connection for this import */
1383 primary_nid = libcfs_nidstr(&nidlist[0]);
1384 prim_nid_len = strlen(primary_nid) + 1;
1385 if (prim_nid_len > UUID_MAX)
1388 strncpy(server_uuid.uuid, primary_nid,
1391 CDEBUG(D_INFO, "Adding a connection for %s\n",
1394 rc = client_import_dyn_add_conn(imp, &server_uuid,
1397 CERROR("%s: Failed to add new connection with NID '%s' to import: rc = %d\n",
1398 obd->obd_name, primary_nid, rc);
1402 rc = client_import_add_nids_to_conn(imp, nidlist,
1403 entry->mne_nid_count,
1404 entry->mne_nid_size,
1407 CERROR("%s: failed to lookup UUID: rc = %d\n",
1412 if (rc == -ENODEV) {
1413 /* client does not connect to the OST yet */
1418 if (rc < 0 && rc != -ENOSPC) {
1419 CERROR("mgc: cannot find UUID by nid '%s': rc = %d\n",
1420 libcfs_nidstr(&nidlist[0]), rc);
1422 /* For old NID format case the nidlist was allocated. */
1423 if (entry->mne_nid_type == 0)
1424 OBD_FREE_PTR_ARRAY(nidlist,
1425 entry->mne_nid_count);
1429 CDEBUG(D_INFO, "Found UUID '%s' by NID '%s'\n",
1430 uuid->uuid, libcfs_nidstr(&nidlist[0]));
1432 pos += strlen(uuid->uuid);
1433 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1434 LASSERT(pos < bufsz);
1436 lustre_cfg_bufs_set_string(&bufs, 1, params);
1438 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1444 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1446 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1447 prev_version, max_version, obdname, params);
1449 rc = class_process_config(lcfg);
1450 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1451 lcfg->lcfg_buflens));
1453 CDEBUG(D_INFO, "process config for %s error %d\n",
1456 /* continue, even one with error */
1458 /* For old NID format case the nidlist was allocated. */
1459 if (entry->mne_nid_type == 0)
1460 OBD_FREE_PTR_ARRAY(nidlist, entry->mne_nid_count);
1463 OBD_FREE(buf, PAGE_SIZE);
1469 * This function is called if this client was notified for target restarting
1470 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1473 static int mgc_process_recover_log(struct obd_device *obd,
1474 struct config_llog_data *cld)
1476 struct ptlrpc_request *req = NULL;
1477 struct config_llog_instance *cfg = &cld->cld_cfg;
1478 struct mgs_config_body *body;
1479 struct mgs_config_res *res;
1480 struct ptlrpc_bulk_desc *desc;
1481 struct page **pages = NULL;
1484 bool mne_swab = false;
1491 /* allocate buffer for bulk transfer.
1492 * if this is the first time for this mgs to read logs,
1493 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1494 * once; otherwise, it only reads increment of logs, this should be
1495 * small and CONFIG_READ_NRPAGES will be used.
1497 nrpages = CONFIG_READ_NRPAGES;
1498 if (cfg->cfg_last_idx == 0)
1499 nrpages = CONFIG_READ_NRPAGES_INIT;
1501 OBD_ALLOC_PTR_ARRAY_LARGE(pages, nrpages);
1503 GOTO(out, rc = -ENOMEM);
1505 for (i = 0; i < nrpages; i++) {
1506 pages[i] = alloc_page(GFP_KERNEL);
1507 if (pages[i] == NULL)
1508 GOTO(out, rc = -ENOMEM);
1512 LASSERT(cld_is_recover(cld));
1513 LASSERT(mutex_is_locked(&cld->cld_lock));
1514 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1515 &RQF_MGS_CONFIG_READ);
1517 GOTO(out, rc = -ENOMEM);
1519 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1524 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1525 LASSERT(body != NULL);
1526 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1527 rc = strscpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name));
1530 body->mcb_offset = cfg->cfg_last_idx + 1;
1531 body->mcb_type = cld->cld_type;
1532 body->mcb_bits = PAGE_SHIFT;
1533 body->mcb_units = nrpages;
1534 body->mcb_rec_nid_size = sizeof(struct lnet_nid);
1536 /* allocate bulk transfer descriptor */
1537 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1538 PTLRPC_BULK_PUT_SINK,
1540 &ptlrpc_bulk_kiov_pin_ops);
1542 GOTO(out, rc = -ENOMEM);
1544 for (i = 0; i < nrpages; i++)
1545 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1548 ptlrpc_request_set_replen(req);
1549 rc = ptlrpc_queue_wait(req);
1553 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1555 GOTO(out, rc = -EPROTO);
1557 if (res->mcr_size < res->mcr_offset)
1558 GOTO(out, rc = -EINVAL);
1560 /* always update the index even though it might have errors with
1561 * handling the recover logs
1563 cfg->cfg_last_idx = res->mcr_offset;
1564 eof = res->mcr_offset == res->mcr_size;
1566 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1567 res->mcr_offset, eof == false);
1569 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1571 GOTO(out, rc = ealen);
1573 if (ealen > nrpages << PAGE_SHIFT)
1574 GOTO(out, rc = -EINVAL);
1576 if (ealen == 0) { /* no logs transferred */
1582 mne_swab = req_capsule_rep_need_swab(&req->rq_pill);
1584 for (i = 0; i < nrpages && ealen > 0; i++) {
1588 ptr = kmap(pages[i]);
1589 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1590 min_t(int, ealen, PAGE_SIZE),
1594 CWARN("%s: error processing %s log recovery: rc = %d\n",
1598 GOTO(out, rc = rc2);
1606 ptlrpc_req_finished(req);
1610 if (rc == 0 && !eof)
1614 for (i = 0; i < nrpages; i++) {
1615 if (pages[i] == NULL)
1617 __free_page(pages[i]);
1619 OBD_FREE_PTR_ARRAY_LARGE(pages, nrpages);
1624 /* local_only means it cannot get remote llogs */
1625 static int mgc_process_cfg_log(struct obd_device *mgc,
1626 struct config_llog_data *cld, int local_only)
1628 struct llog_ctxt *ctxt;
1629 struct lustre_sb_info *lsi = NULL;
1635 LASSERT(mutex_is_locked(&cld->cld_lock));
1637 if (cld->cld_cfg.cfg_sb)
1638 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1644 rc = lu_env_init(env, LCT_MG_THREAD);
1648 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1650 #ifdef HAVE_SERVER_SUPPORT
1651 rc = mgc_process_server_cfg_log(env, &ctxt, lsi, mgc, cld,
1653 if (rc == -EIO && local_only)
1655 if (rc && rc != -ENOENT)
1658 GOTO(out_pop, rc = -EIO);
1660 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1664 * update settings on existing OBDs.
1665 * the logname must be <fsname>-sptlrpc
1667 if (rc == 0 && cld_is_sptlrpc(cld))
1668 class_notify_sptlrpc_conf(cld->cld_logname,
1669 strlen(cld->cld_logname) -
1670 strlen("-sptlrpc"));
1674 __llog_ctxt_put(env, ctxt);
1681 static bool mgc_import_in_recovery(struct obd_import *imp)
1683 bool in_recovery = true;
1685 spin_lock(&imp->imp_lock);
1686 if (imp->imp_state == LUSTRE_IMP_FULL ||
1687 imp->imp_state == LUSTRE_IMP_CLOSED)
1688 in_recovery = false;
1689 spin_unlock(&imp->imp_lock);
1695 * Get a configuration log from the MGS and process it.
1697 * This function is called for both clients and servers to process the
1698 * configuration log from the MGS. The MGC enqueues a DLM lock on the
1699 * log from the MGS, and if the lock gets revoked the MGC will be notified
1700 * by the lock cancellation callback that the config log has changed,
1701 * and will enqueue another MGS lock on it, and then continue processing
1702 * the new additions to the end of the log.
1704 * Since the MGC import is not replayable, if the import is being evicted
1705 * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
1706 * the log until recovery is finished or the import is closed.
1708 * Make a local copy of the log before parsing it if appropriate (non-MGS
1709 * server) so that the server can start even when the MGS is down.
1711 * There shouldn't be multiple processes running process_log at once --
1712 * sounds like badness. It actually might be fine, as long as they're not
1713 * trying to update from the same log simultaneously, in which case we
1714 * should use a per-log semaphore instead of cld_lock.
1716 * \param[in] mgc MGC device by which to fetch the configuration log
1717 * \param[in] cld log processing state (stored in lock callback data)
1719 * \retval 0 on success
1720 * \retval negative errno on failure
1722 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1724 struct lustre_handle lockh = { 0 };
1725 __u64 flags = LDLM_FL_NO_LRU;
1731 LASSERT(cld != NULL);
1733 /* I don't want multiple processes running process_log at once --
1734 * sounds like badness. It actually might be fine, as long as
1735 * we're not trying to update from the same log
1736 * simultaneously (in which case we should use a per-log sem.)
1739 mutex_lock(&cld->cld_lock);
1740 if (cld->cld_stopping) {
1741 mutex_unlock(&cld->cld_lock);
1745 CFS_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1747 CDEBUG(D_MGC, "Process log %s-%016lx from %d\n", cld->cld_logname,
1748 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1750 /* Get the cfg lock on the llog */
1751 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
1753 cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
1754 cld, 0, NULL, &lockh);
1756 /* Get the cld, it will be released in mgc_blocking_ast. */
1757 config_log_get(cld);
1758 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1759 LASSERT(!lustre_handle_is_used(&cld->cld_lockh));
1761 cld->cld_lockh = lockh;
1763 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1764 cld->cld_lockh.cookie = 0;
1766 if (rcl == -ESHUTDOWN &&
1767 atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
1768 struct obd_import *imp;
1769 long timeout = cfs_time_seconds(obd_timeout);
1771 mutex_unlock(&cld->cld_lock);
1772 imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
1774 /* Let's force the pinger, and wait the import to be
1775 * connected, note: since mgc import is non-replayable,
1776 * and even the import state is disconnected, it does
1777 * not mean the "recovery" is stopped, so we will keep
1778 * waitting until timeout or the import state is
1781 ptlrpc_pinger_force(imp);
1783 wait_event_idle_timeout(imp->imp_recovery_waitq,
1784 !mgc_import_in_recovery(imp),
1787 if (imp->imp_state == LUSTRE_IMP_FULL) {
1791 mutex_lock(&cld->cld_lock);
1792 /* unlock/lock mutex, so check stopping again */
1793 if (cld->cld_stopping) {
1794 mutex_unlock(&cld->cld_lock);
1797 spin_lock(&config_list_lock);
1798 cld->cld_lostlock = 1;
1799 spin_unlock(&config_list_lock);
1802 /* mark cld_lostlock so that it will requeue
1803 * after MGC becomes available.
1805 spin_lock(&config_list_lock);
1806 cld->cld_lostlock = 1;
1807 spin_unlock(&config_list_lock);
1811 if (cld_is_recover(cld) && !rcl)
1812 rc = mgc_process_recover_log(mgc, cld);
1813 #ifdef HAVE_SERVER_SUPPORT
1814 else if (cld_is_nodemap(cld)) {
1818 rc = mgc_process_nodemap_log(mgc, cld);
1821 else if (!cld_is_barrier(cld))
1822 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1824 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1825 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1826 if (rc != -ETIMEDOUT && rc != -EIO && rc != -EAGAIN) {
1827 cld->cld_processed = 1;
1831 /* Now drop the lock so MGS can revoke it */
1833 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
1835 CERROR("Can't drop cfg lock: %d\n", rcl);
1837 mutex_unlock(&cld->cld_lock);
1839 /* requeue nodemap lock immediately if transfer was interrupted */
1840 if ((cld_is_nodemap(cld) && rc == -EAGAIN) ||
1841 (cld_is_recover(cld) && rc)) {
1842 if (cld_is_recover(cld))
1843 CWARN("%s: IR log %s failed, not fatal: rc = %d\n",
1844 mgc->obd_name, cld->cld_logname, rc);
1845 mgc_requeue_add(cld);
1853 /* Called from lustre_process_log.
1854 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1855 * any services, and adds it to the list logs to watch (follow).
1857 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
1859 struct lustre_cfg *lcfg = buf;
1860 struct config_llog_instance *cfg = NULL;
1865 #ifdef HAVE_SERVER_SUPPORT
1866 rc = mgc_process_config_server(obd, len, buf);
1871 switch (lcfg->lcfg_command) {
1872 case LCFG_SPTLRPC_CONF: {
1873 rc = sptlrpc_process_config(lcfg);
1876 case LCFG_LOG_START: {
1877 struct config_llog_data *cld;
1878 struct super_block *sb;
1880 logname = lustre_cfg_string(lcfg, 1);
1881 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1882 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1884 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1887 /* We're only called through here on the initial mount */
1888 cld = config_log_add(obd, logname, cfg, sb);
1894 /* if it exists, the sptlrpc config log really needs to be
1895 * correctly processed before processing other logs,
1896 * otherwise client might use incorrect sec flavor
1898 if (cld->cld_sptlrpc && !cld->cld_sptlrpc->cld_processed) {
1899 unsigned int timeout = 120;
1901 /* we do not want to wait forever,
1902 * we prefer a (excessively) long timeout
1904 timeout = max(20 * mgc_requeue_timeout_min, timeout);
1905 wait_event_idle_timeout(rq_waitq,
1906 cld->cld_sptlrpc->cld_processed,
1907 cfs_time_seconds(timeout));
1910 rc = mgc_process_log(obd, cld);
1911 if (rc == 0 && cld->cld_recover != NULL) {
1912 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1913 imp_connect_data, IMP_RECOV)) {
1914 rc = mgc_process_log(obd, cld->cld_recover);
1916 struct config_llog_data *cir;
1918 mutex_lock(&cld->cld_lock);
1919 cir = cld->cld_recover;
1920 cld->cld_recover = NULL;
1921 mutex_unlock(&cld->cld_lock);
1922 config_log_put(cir);
1926 CERROR("Cannot process recover llog %d\n", rc);
1929 if (rc == 0 && cld->cld_params != NULL) {
1930 rc = mgc_process_log(obd, cld->cld_params);
1931 if (rc == -ENOENT) {
1933 "There is no params config file yet\n");
1936 /* params log is optional */
1938 CERROR("%s: can't process params llog: rc = %d\n",
1944 case LCFG_LOG_END: {
1945 logname = lustre_cfg_string(lcfg, 1);
1947 if (lcfg->lcfg_bufcount >= 2)
1948 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1949 rc = config_log_end(logname, cfg);
1953 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1961 static const struct obd_ops mgc_obd_ops = {
1962 .o_owner = THIS_MODULE,
1963 .o_setup = mgc_setup,
1964 .o_precleanup = mgc_precleanup,
1965 .o_cleanup = mgc_cleanup,
1966 .o_add_conn = client_import_add_conn,
1967 .o_del_conn = client_import_del_conn,
1968 .o_connect = client_connect_import,
1969 .o_disconnect = client_disconnect_export,
1970 .o_set_info_async = mgc_set_info_async,
1971 .o_get_info = mgc_get_info,
1972 .o_import_event = mgc_import_event,
1973 .o_process_config = mgc_process_config,
1976 static int mgc_param_requeue_timeout_min_set(const char *val,
1977 cfs_kernel_param_arg_t *kp)
1982 rc = kstrtouint(val, 0, &num);
1988 mgc_requeue_timeout_min = num;
1993 static const struct kernel_param_ops param_ops_requeue_timeout_min = {
1994 .set = mgc_param_requeue_timeout_min_set,
1995 .get = param_get_uint,
1998 #define param_check_requeue_timeout_min(name, p) \
1999 __param_check(name, p, unsigned int)
2001 unsigned int mgc_requeue_timeout_min = MGC_TIMEOUT_MIN_SECONDS;
2002 #ifdef HAVE_KERNEL_PARAM_OPS
2003 module_param(mgc_requeue_timeout_min, requeue_timeout_min, 0644);
2005 module_param_call(mgc_requeue_timeout_min, mgc_param_requeue_timeout_min_set,
2006 param_get_uint, ¶m_ops_requeue_timeout_min, 0644);
2008 MODULE_PARM_DESC(mgc_requeue_timeout_min, "Minimal requeue time to refresh logs");
2010 static int __init mgc_init(void)
2014 rc = libcfs_setup();
2018 return class_register_type(&mgc_obd_ops, NULL, false,
2019 LUSTRE_MGC_NAME, NULL);
2022 static void __exit mgc_exit(void)
2024 class_unregister_type(LUSTRE_MGC_NAME);
2027 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2028 MODULE_DESCRIPTION("Lustre Management Client");
2029 MODULE_VERSION(LUSTRE_VERSION_STRING);
2030 MODULE_LICENSE("GPL");
2032 module_init(mgc_init);
2033 module_exit(mgc_exit);