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, maybe contain "-".
96 * e.g. "lustre", "SUN-000" */
97 return mgc_name2resid(fsname, strlen(fsname), res_id, type);
99 EXPORT_SYMBOL(mgc_fsname2resid);
101 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id,
102 enum mgs_cfg_type type)
107 /* logname consists of "fsname-nodetype".
108 * e.g. "lustre-MDT0001", "SUN-000-client"
109 * there is an exception: llog "params" */
110 name_end = strrchr(logname, '-');
112 len = strlen(logname);
114 len = name_end - logname;
115 return mgc_name2resid(logname, len, res_id, type);
117 EXPORT_SYMBOL(mgc_logname2resid);
119 /********************** config llog list **********************/
120 static LIST_HEAD(config_llog_list);
121 static DEFINE_SPINLOCK(config_list_lock); /* protects config_llog_list */
123 /* Take a reference to a config log */
124 static int config_log_get(struct config_llog_data *cld)
127 atomic_inc(&cld->cld_refcount);
128 CDEBUG(D_INFO, "log %s (%p) refs %d\n", cld->cld_logname, cld,
129 atomic_read(&cld->cld_refcount));
133 /* Drop a reference to a config log. When no longer referenced,
134 we can free the config log data */
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 atomic_read(&cld->cld_refcount));
144 LASSERT(atomic_read(&cld->cld_refcount) > 0);
146 /* spinlock to make sure no item with 0 refcount in the list */
147 if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
148 list_del(&cld->cld_list_chain);
149 spin_unlock(&config_list_lock);
151 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
152 #ifdef HAVE_SERVER_SUPPORT
153 config_log_put(cld->cld_barrier);
154 config_log_put(cld->cld_nodemap);
156 config_log_put(cld->cld_recover);
157 config_log_put(cld->cld_params);
158 config_log_put(cld->cld_sptlrpc);
159 if (cld_is_sptlrpc(cld)) {
160 cld->cld_stopping = 1;
161 sptlrpc_conf_log_stop(cld->cld_logname);
164 class_export_put(cld->cld_mgcexp);
165 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
171 /* Find a config log by name */
173 struct config_llog_data *config_log_find(char *logname,
174 struct config_llog_instance *cfg)
176 struct config_llog_data *cld;
177 struct config_llog_data *found = NULL;
178 unsigned long cfg_instance;
181 LASSERT(logname != NULL);
183 cfg_instance = cfg ? cfg->cfg_instance : 0;
184 spin_lock(&config_list_lock);
185 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
186 /* check if cfg_instance is the one we want */
187 if (cfg_instance != cld->cld_cfg.cfg_instance)
190 /* instance may be NULL, should check name */
191 if (strcmp(logname, cld->cld_logname) == 0) {
193 config_log_get(found);
197 spin_unlock(&config_list_lock);
202 struct config_llog_data *do_config_log_add(struct obd_device *obd,
204 enum mgs_cfg_type type,
205 struct config_llog_instance *cfg,
206 struct super_block *sb)
208 struct config_llog_data *cld;
213 CDEBUG(D_MGC, "do adding config log %s-%016lx\n", logname,
214 cfg ? cfg->cfg_instance : 0);
216 OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
218 RETURN(ERR_PTR(-ENOMEM));
220 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
222 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
226 strcpy(cld->cld_logname, logname);
230 cld->cld_cfg.cfg_callback = class_config_llog_handler;
231 mutex_init(&cld->cld_lock);
232 cld->cld_cfg.cfg_last_idx = 0;
233 cld->cld_cfg.cfg_flags = 0;
234 cld->cld_cfg.cfg_sb = sb;
235 cld->cld_type = type;
236 atomic_set(&cld->cld_refcount, 1);
238 /* Keep the mgc around until we are done */
239 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
241 if (cld_is_sptlrpc(cld))
242 sptlrpc_conf_log_start(logname);
244 spin_lock(&config_list_lock);
245 list_add(&cld->cld_list_chain, &config_llog_list);
246 spin_unlock(&config_list_lock);
248 if (cld_is_sptlrpc(cld) || cld_is_nodemap(cld) || cld_is_barrier(cld)) {
249 rc = mgc_process_log(obd, cld);
250 if (rc && rc != -ENOENT)
251 CERROR("%s: failed processing log, type %d: rc = %d\n",
252 obd->obd_name, type, rc);
258 static struct config_llog_data *
259 config_recover_log_add(struct obd_device *obd, char *fsname,
260 struct config_llog_instance *cfg,
261 struct super_block *sb)
263 struct config_llog_instance lcfg = *cfg;
264 struct config_llog_data *cld;
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));
280 if (IS_SERVER(s2lsi(sb))) { /* mdt */
281 LASSERT(lcfg.cfg_instance == 0);
282 lcfg.cfg_instance = ll_get_cfg_instance(sb);
283 strncat(logname, "-mdtir", sizeof(logname));
285 LASSERT(lcfg.cfg_instance != 0);
286 strncat(logname, "-cliir", sizeof(logname));
289 cld = do_config_log_add(obd, logname, MGS_CFG_T_RECOVER, &lcfg, sb);
293 static struct config_llog_data *
294 config_log_find_or_add(struct obd_device *obd, char *logname,
295 struct super_block *sb, enum mgs_cfg_type type,
296 struct config_llog_instance *cfg)
298 struct config_llog_instance lcfg = *cfg;
299 struct config_llog_data *cld;
301 /* Note class_config_llog_handler() depends on getting "obd" back */
302 lcfg.cfg_instance = sb ? ll_get_cfg_instance(sb) : (unsigned long)obd;
304 cld = config_log_find(logname, &lcfg);
305 if (unlikely(cld != NULL))
308 return do_config_log_add(obd, logname, type, &lcfg, sb);
311 /** Add this log to the list of active logs watched by an MGC.
312 * Active means we're watching for updates.
313 * We have one active log per "mount" - client instance or servername.
314 * Each instance may be at a different point in the log.
316 static struct config_llog_data *
317 config_log_add(struct obd_device *obd, char *logname,
318 struct config_llog_instance *cfg, struct super_block *sb)
320 struct lustre_sb_info *lsi = s2lsi(sb);
321 struct config_llog_data *cld = NULL;
322 struct config_llog_data *sptlrpc_cld = NULL;
323 struct config_llog_data *params_cld = NULL;
324 #ifdef HAVE_SERVER_SUPPORT
325 struct config_llog_data *nodemap_cld = NULL;
326 struct config_llog_data *barrier_cld = NULL;
334 CDEBUG(D_MGC, "add config log %s-%016lx\n", logname,
338 * for each regular log, the depended sptlrpc log name is
339 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
341 ptr = strrchr(logname, '-');
342 if (ptr == NULL || ptr - logname > 8) {
343 CERROR("logname %s is too long\n", logname);
344 RETURN(ERR_PTR(-EINVAL));
347 memcpy(seclogname, logname, ptr - logname);
348 strcpy(seclogname + (ptr - logname), "-sptlrpc");
350 if (cfg->cfg_sub_clds & CONFIG_SUB_SPTLRPC) {
351 sptlrpc_cld = config_log_find_or_add(obd, seclogname, NULL,
352 MGS_CFG_T_SPTLRPC, cfg);
353 if (IS_ERR(sptlrpc_cld)) {
354 rc = PTR_ERR(sptlrpc_cld);
355 CERROR("%s: can't create sptlrpc log %s: rc = %d\n",
356 obd->obd_name, seclogname, rc);
361 if (cfg->cfg_sub_clds & CONFIG_SUB_PARAMS) {
362 params_cld = config_log_find_or_add(obd, PARAMS_FILENAME, sb,
363 MGS_CFG_T_PARAMS, cfg);
364 if (IS_ERR(params_cld)) {
365 rc = PTR_ERR(params_cld);
366 CERROR("%s: can't create params log: rc = %d\n",
368 GOTO(out_sptlrpc, rc);
372 #ifdef HAVE_SERVER_SUPPORT
373 if (!IS_MGS(lsi) && cfg->cfg_sub_clds & CONFIG_SUB_NODEMAP) {
374 nodemap_cld = config_log_find_or_add(obd, LUSTRE_NODEMAP_NAME,
375 NULL, MGS_CFG_T_NODEMAP,
377 if (IS_ERR(nodemap_cld)) {
378 rc = PTR_ERR(nodemap_cld);
379 CERROR("%s: cannot create nodemap log: rc = %d\n",
381 GOTO(out_params, rc);
385 if (IS_MDT(s2lsi(sb)) && cfg->cfg_sub_clds & CONFIG_SUB_BARRIER) {
386 snprintf(seclogname + (ptr - logname), sizeof(seclogname) - 1,
387 "-%s", BARRIER_FILENAME);
388 barrier_cld = config_log_find_or_add(obd, seclogname, sb,
389 MGS_CFG_T_BARRIER, cfg);
390 if (IS_ERR(barrier_cld)) {
391 rc = PTR_ERR(barrier_cld);
392 CERROR("%s: can't create barrier log: rc = %d\n",
394 GOTO(out_nodemap, rc);
398 cld = do_config_log_add(obd, logname, MGS_CFG_T_CONFIG, cfg, sb);
401 CERROR("%s: can't create log: rc = %d\n",
403 #ifdef HAVE_SERVER_SUPPORT
404 GOTO(out_barrier, rc);
406 GOTO(out_params, rc);
410 LASSERT(lsi->lsi_lmd);
411 if (!test_bit(LMD_FLG_NOIR, lsi->lsi_lmd->lmd_flags) &&
412 cfg->cfg_sub_clds & CONFIG_SUB_RECOVER) {
413 struct config_llog_data *recover_cld;
415 ptr = strrchr(seclogname, '-');
419 CERROR("%s: sptlrpc log name not correct, %s: "
420 "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
421 GOTO(out_cld, rc = -EINVAL);
424 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
425 if (IS_ERR(recover_cld)) {
426 rc = PTR_ERR(recover_cld);
427 CERROR("%s: can't create recover log: rc = %d\n",
432 mutex_lock(&cld->cld_lock);
434 cld->cld_recover = recover_cld;
438 mutex_lock(&cld->cld_lock);
439 #ifdef HAVE_SERVER_SUPPORT
440 cld->cld_barrier = barrier_cld;
441 cld->cld_nodemap = nodemap_cld;
443 cld->cld_params = params_cld;
444 cld->cld_sptlrpc = sptlrpc_cld;
445 mutex_unlock(&cld->cld_lock);
451 #ifdef HAVE_SERVER_SUPPORT
453 config_log_put(barrier_cld);
455 config_log_put(nodemap_cld);
458 config_log_put(params_cld);
460 config_log_put(sptlrpc_cld);
465 DEFINE_MUTEX(llog_process_lock);
467 static inline void config_mark_cld_stop_nolock(struct config_llog_data *cld)
471 spin_lock(&config_list_lock);
472 cld->cld_stopping = 1;
473 spin_unlock(&config_list_lock);
475 CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
476 if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
477 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
480 static inline void config_mark_cld_stop(struct config_llog_data *cld)
483 mutex_lock(&cld->cld_lock);
484 config_mark_cld_stop_nolock(cld);
485 mutex_unlock(&cld->cld_lock);
489 /** Stop watching for updates on this log.
491 static int config_log_end(char *logname, struct config_llog_instance *cfg)
493 struct config_llog_data *cld;
494 struct config_llog_data *cld_sptlrpc = NULL;
495 struct config_llog_data *cld_params = NULL;
496 struct config_llog_data *cld_recover = NULL;
497 struct config_llog_data *cld_nodemap = NULL;
498 struct config_llog_data *cld_barrier = NULL;
503 cld = config_log_find(logname, cfg);
507 mutex_lock(&cld->cld_lock);
509 * if cld_stopping is set, it means we didn't start the log thus
510 * not owning the start ref. this can happen after previous umount:
511 * the cld still hanging there waiting for lock cancel, and we
512 * remount again but failed in the middle and call log_end without
515 if (unlikely(cld->cld_stopping)) {
516 mutex_unlock(&cld->cld_lock);
517 /* drop the ref from the find */
522 cld_recover = cld->cld_recover;
523 cld->cld_recover = NULL;
524 cld_params = cld->cld_params;
525 cld->cld_params = NULL;
526 cld_nodemap = cld->cld_nodemap;
527 cld->cld_nodemap = NULL;
528 cld_barrier = cld->cld_barrier;
529 cld->cld_barrier = NULL;
530 cld_sptlrpc = cld->cld_sptlrpc;
531 cld->cld_sptlrpc = NULL;
533 config_mark_cld_stop_nolock(cld);
534 mutex_unlock(&cld->cld_lock);
536 config_mark_cld_stop(cld_recover);
537 config_log_put(cld_recover);
538 config_mark_cld_stop(cld_params);
539 config_log_put(cld_params);
540 config_mark_cld_stop(cld_barrier);
541 config_log_put(cld_barrier);
542 /* don't explicitly set cld_stopping on sptlrpc lock here, as other
543 * targets may be active, it will be done in config_log_put if necessary
545 config_log_put(cld_sptlrpc);
546 /* don't set cld_stopping on nm lock as other targets may be active */
547 config_log_put(cld_nodemap);
549 /* drop the ref from the find */
551 /* drop the start ref */
554 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
559 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
561 struct obd_device *obd = data;
562 struct obd_import *imp;
563 struct obd_connect_data *ocd;
564 struct config_llog_data *cld;
569 with_imp_locked(obd, imp, rc) {
570 ocd = &imp->imp_connect_data;
572 seq_printf(m, "imperative_recovery: %s\n",
573 OCD_HAS_FLAG(ocd, IMP_RECOV) ?
574 "ENABLED" : "DISABLED");
579 seq_printf(m, "client_state:\n");
581 spin_lock(&config_list_lock);
582 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
583 if (cld->cld_recover == NULL)
585 seq_printf(m, " - { client: %s, nidtbl_version: %u }\n",
587 cld->cld_recover->cld_cfg.cfg_last_idx);
589 spin_unlock(&config_list_lock);
594 /* reenqueue any lost locks */
595 #define RQ_RUNNING 0x1
599 #define RQ_PRECLEANUP 0x10
600 static int rq_state = 0;
601 static wait_queue_head_t rq_waitq;
602 static DECLARE_COMPLETION(rq_exit);
603 static DECLARE_COMPLETION(rq_start);
605 static void do_requeue(struct config_llog_data *cld)
610 LASSERT(atomic_read(&cld->cld_refcount) > 0);
613 * Do not run mgc_process_log on a disconnected export or an
614 * export which is being disconnected. Take the client
615 * semaphore to make the check non-racy.
617 down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
619 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
620 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
621 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
622 if (rc && rc != -ENOENT)
623 CERROR("failed processing log: %d\n", rc);
625 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
628 up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
633 static int mgc_requeue_thread(void *data)
639 CDEBUG(D_MGC, "Starting requeue thread\n");
641 /* Keep trying failed locks periodically */
642 spin_lock(&config_list_lock);
643 rq_state |= RQ_RUNNING;
644 while (!(rq_state & RQ_STOP)) {
645 struct config_llog_data *cld, *cld_prev;
648 /* Any new or requeued lostlocks will change the state */
649 rq_state &= ~(RQ_NOW | RQ_LATER);
650 spin_unlock(&config_list_lock);
657 /* Always wait a few seconds to allow the server who
658 * caused the lock revocation to finish its setup, plus some
659 * random so everyone doesn't try to reconnect at once.
661 to = mgc_requeue_timeout_min == 0 ? 1 : mgc_requeue_timeout_min;
662 to = cfs_time_seconds(mgc_requeue_timeout_min) +
663 prandom_u32_max(cfs_time_seconds(to));
664 wait_event_idle_timeout(rq_waitq,
665 rq_state & (RQ_STOP | RQ_PRECLEANUP), to);
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. */
687 cld->cld_lostlock = 0;
688 spin_unlock(&config_list_lock);
690 config_log_put(cld_prev);
693 if (likely(!(rq_state & RQ_STOP))) {
695 spin_lock(&config_list_lock);
697 spin_lock(&config_list_lock);
701 spin_unlock(&config_list_lock);
702 config_log_put(cld_prev);
704 /* Wait a bit to see if anyone else needs a requeue */
705 wait_event_idle(rq_waitq, rq_state & (RQ_NOW | RQ_STOP));
706 spin_lock(&config_list_lock);
709 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
710 rq_state &= ~RQ_RUNNING;
711 spin_unlock(&config_list_lock);
715 CDEBUG(D_MGC, "Ending requeue thread\n");
719 /* Add a cld to the list to requeue. Start the requeue thread if needed.
720 We are responsible for dropping the config log reference from here on out. */
721 static void mgc_requeue_add(struct config_llog_data *cld)
726 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
727 cld->cld_logname, atomic_read(&cld->cld_refcount),
728 cld->cld_stopping, rq_state);
729 LASSERT(atomic_read(&cld->cld_refcount) > 0);
731 /* lets cancel an existent lock to mark cld as "lostlock" */
732 CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
733 if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
734 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
736 mutex_lock(&cld->cld_lock);
737 spin_lock(&config_list_lock);
738 if (!(rq_state & RQ_STOP) && !cld->cld_stopping) {
739 cld->cld_lostlock = 1;
743 spin_unlock(&config_list_lock);
744 mutex_unlock(&cld->cld_lock);
751 /********************** class fns **********************/
752 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
754 struct llog_ctxt *ctxt;
759 /* setup only remote ctxt, the local disk context is switched per each
760 * filesystem during mgc_fs_setup() */
761 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
766 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
769 llog_initiator_connect(ctxt);
775 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
777 struct llog_ctxt *ctxt;
781 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
783 llog_cleanup(env, ctxt);
789 static atomic_t mgc_count = ATOMIC_INIT(0);
790 static int mgc_precleanup(struct obd_device *obd)
796 if (atomic_dec_and_test(&mgc_count)) {
797 LASSERT(rq_state & RQ_RUNNING);
798 /* stop requeue thread */
801 /* wakeup requeue thread to clean our cld */
802 temp = RQ_NOW | RQ_PRECLEANUP;
805 spin_lock(&config_list_lock);
807 spin_unlock(&config_list_lock);
811 wait_for_completion(&rq_exit);
812 obd_cleanup_client_import(obd);
814 rc = mgc_llog_fini(NULL, obd);
816 CERROR("failed to cleanup llogging subsystems\n");
821 static int mgc_cleanup(struct obd_device *obd)
826 /* COMPAT_146 - old config logs may have added profiles we don't
828 if (atomic_read(&obd->obd_type->typ_refcnt) <= 1)
829 /* Only for the last mgc */
830 class_del_profiles();
832 lprocfs_obd_cleanup(obd);
835 rc = client_obd_cleanup(obd);
839 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
841 struct task_struct *task;
845 rc = ptlrpcd_addref();
849 rc = client_obd_setup(obd, lcfg);
851 GOTO(err_decref, rc);
853 rc = mgc_llog_init(NULL, obd);
855 CERROR("failed to setup llogging subsystems\n");
856 GOTO(err_cleanup, rc);
859 rc = mgc_tunables_init(obd);
863 if (atomic_inc_return(&mgc_count) == 1) {
865 init_waitqueue_head(&rq_waitq);
867 /* start requeue thread */
868 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
871 CERROR("%s: cannot start requeue thread: rc = %d; "
872 "no more log updates\n",
876 /* rc is the task_struct pointer of mgc_requeue_thread. */
878 wait_for_completion(&rq_start);
884 lprocfs_obd_cleanup(obd);
886 client_obd_cleanup(obd);
892 /* based on ll_mdc_blocking_ast */
893 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
894 void *data, int flag)
896 struct lustre_handle lockh;
897 struct config_llog_data *cld = (struct config_llog_data *)data;
902 case LDLM_CB_BLOCKING:
903 /* mgs wants the lock, give it up... */
904 LDLM_DEBUG(lock, "MGC blocking CB");
905 ldlm_lock2handle(lock, &lockh);
906 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
908 case LDLM_CB_CANCELING:
909 /* We've given up the lock, prepare ourselves to update. */
910 LDLM_DEBUG(lock, "MGC cancel CB");
912 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
913 PLDLMRES(lock->l_resource),
914 (char *)&lock->l_resource->lr_name.name[0]);
917 CDEBUG(D_INFO, "missing data, won't requeue\n");
921 /* held at mgc_process_log(). */
922 LASSERT(atomic_read(&cld->cld_refcount) > 0);
924 lock->l_ast_data = NULL;
925 cld->cld_lockh.cookie = 0;
926 /* Are we done with this log? */
927 if (cld->cld_stopping) {
928 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
933 /* Make sure not to re-enqueue when the mgc is stopping
934 (we get called from client_disconnect_export) */
935 if (lock->l_conn_export == NULL ||
936 lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
937 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
944 mgc_requeue_add(cld);
954 /* Take a config lock so we can get cancel notifications */
955 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
956 union ldlm_policy_data *policy, enum ldlm_mode mode,
957 __u64 *flags, ldlm_glimpse_callback glimpse_callback,
958 void *data, __u32 lvb_len, void *lvb_swabber,
959 struct lustre_handle *lockh)
961 struct config_llog_data *cld = (struct config_llog_data *)data;
962 struct ldlm_enqueue_info einfo = {
965 .ei_cb_bl = mgc_blocking_ast,
966 .ei_cb_cp = ldlm_completion_ast,
967 .ei_cb_gl = glimpse_callback,
969 struct ptlrpc_request *req;
970 int short_limit = cld_is_sptlrpc(cld);
977 CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
978 cld->cld_resid.name[0]);
980 /* We need a callback for every lockholder, so don't try to
981 ldlm_lock_match (see rev 1.1.2.11.2.47) */
982 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
983 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
988 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
989 ptlrpc_request_set_replen(req);
991 /* check if this is server or client */
992 if (cld->cld_cfg.cfg_sb &&
993 IS_SERVER(s2lsi(cld->cld_cfg.cfg_sb)))
996 /* Limit how long we will wait for the enqueue to complete */
997 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
998 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
999 NULL, 0, LVB_T_NONE, lockh, 0);
1000 /* A failed enqueue should still call the mgc_blocking_ast,
1001 where it will be requeued if needed ("grant failed"). */
1002 ptlrpc_req_finished(req);
1006 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1007 struct lustre_handle *lockh)
1011 ldlm_lock_decref(lockh, mode);
1016 static void mgc_notify_active(struct obd_device *unused)
1018 /* wakeup mgc_requeue_thread to requeue mgc lock */
1019 spin_lock(&config_list_lock);
1021 spin_unlock(&config_list_lock);
1024 /* TODO: Help the MGS rebuild nidtbl. -jay */
1027 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1028 u32 keylen, void *key,
1029 u32 vallen, void *val,
1030 struct ptlrpc_request_set *set)
1035 /* Turn off initial_recov after we try all backup servers once */
1036 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1037 struct obd_import *imp = class_exp2cliimp(exp);
1039 if (vallen != sizeof(int))
1041 value = *(int *)val;
1042 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1043 imp->imp_obd->obd_name, value,
1044 imp->imp_deactive, imp->imp_invalid,
1045 imp->imp_replayable, imp->imp_obd->obd_replayable,
1046 ptlrpc_import_state_name(imp->imp_state));
1047 /* Resurrect the import immediately if
1048 * 1. we previously got disconnected,
1049 * 2. value > 1 (at the same node with MGS)
1051 if (imp->imp_state == LUSTRE_IMP_DISCON || value > 1)
1052 ptlrpc_reconnect_import(imp);
1057 if (KEY_IS(KEY_MGSSEC)) {
1058 struct client_obd *cli = &exp->exp_obd->u.cli;
1059 struct sptlrpc_flavor flvr;
1062 * empty string means using current flavor, if which haven't
1063 * been set yet, set it as null.
1065 * if flavor has been set previously, check the asking flavor
1066 * must match the existing one.
1069 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1075 rc = sptlrpc_parse_flavor(val, &flvr);
1077 CERROR("invalid sptlrpc flavor %s to MGS\n",
1083 * caller already hold a mutex
1085 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1086 cli->cl_flvr_mgc = flvr;
1087 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1088 sizeof(flvr)) != 0) {
1091 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1093 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1094 "currently %s is in use\n",
1101 #ifdef HAVE_SERVER_SUPPORT
1102 rc = mgc_set_info_async_server(env, exp, keylen, key, vallen, val, set);
1107 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1108 __u32 keylen, void *key, __u32 *vallen, void *val)
1112 if (KEY_IS(KEY_CONN_DATA)) {
1113 struct obd_import *imp = class_exp2cliimp(exp);
1114 struct obd_connect_data *data = val;
1116 if (*vallen == sizeof(*data)) {
1117 *data = imp->imp_connect_data;
1125 static int mgc_import_event(struct obd_device *obd,
1126 struct obd_import *imp,
1127 enum obd_import_event event)
1131 LASSERT(imp->imp_obd == obd);
1132 CDEBUG(D_MGC, "import event %#x\n", event);
1135 case IMP_EVENT_DISCON:
1136 /* MGC imports should not wait for recovery */
1137 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1138 ptlrpc_pinger_ir_down();
1140 case IMP_EVENT_INACTIVE:
1142 case IMP_EVENT_INVALIDATE: {
1143 struct ldlm_namespace *ns = obd->obd_namespace;
1144 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1147 case IMP_EVENT_ACTIVE:
1148 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1149 /* Clearing obd_no_recov allows us to continue pinging */
1150 obd->obd_no_recov = 0;
1151 mgc_notify_active(obd);
1152 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1153 ptlrpc_pinger_ir_up();
1157 case IMP_EVENT_DEACTIVATE:
1158 case IMP_EVENT_ACTIVATE:
1161 CERROR("Unknown import event %#x\n", event);
1167 static int mgc_apply_recover_logs(struct obd_device *mgc,
1168 struct config_llog_data *cld,
1170 void *data, int datalen, bool mne_swab)
1172 struct config_llog_instance *cfg = &cld->cld_cfg;
1173 struct mgs_nidtbl_entry *entry;
1174 struct lustre_cfg *lcfg;
1175 struct lustre_cfg_bufs bufs;
1176 u64 prev_version = 0;
1177 char inst[MTI_NAME_MAXLEN + 1];
1183 unsigned long dynamic_nids;
1186 LASSERT(cfg->cfg_instance != 0);
1187 LASSERT(ll_get_cfg_instance(cfg->cfg_sb) == cfg->cfg_instance);
1189 /* get dynamic nids setting */
1190 dynamic_nids = mgc->obd_dynamic_nids;
1192 if (!IS_SERVER(s2lsi(cfg->cfg_sb))) {
1193 pos = snprintf(inst, sizeof(inst), "%016lx", cfg->cfg_instance);
1194 if (pos >= PAGE_SIZE)
1196 #ifdef HAVE_SERVER_SUPPORT
1198 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1200 LASSERT(IS_MDT(lsi));
1201 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1205 #endif /* HAVE_SERVER_SUPPORT */
1208 OBD_ALLOC(buf, PAGE_SIZE);
1214 while (datalen > 0) {
1215 int entry_len = sizeof(*entry);
1217 struct obd_device *obd;
1218 struct obd_import *imp;
1225 if (datalen < sizeof(*entry))
1228 entry = (typeof(entry))(data + off);
1231 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1233 if (entry->mne_nid_count == 0) /* at least one nid entry */
1235 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1238 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1239 if (datalen < entry_len) /* must have entry_len at least */
1242 /* Keep this swab for normal mixed endian handling. LU-1644 */
1244 lustre_swab_mgs_nidtbl_entry(entry);
1245 if (entry->mne_length > PAGE_SIZE) {
1246 CERROR("MNE too large (%u)\n", entry->mne_length);
1250 if (entry->mne_length < entry_len)
1253 off += entry->mne_length;
1254 datalen -= entry->mne_length;
1258 if (entry->mne_version > max_version) {
1259 CERROR("entry index(%lld) is over max_index(%lld)\n",
1260 entry->mne_version, max_version);
1264 if (prev_version >= entry->mne_version) {
1265 CERROR("index unsorted, prev %lld, now %lld\n",
1266 prev_version, entry->mne_version);
1269 prev_version = entry->mne_version;
1272 * Write a string with format "nid::instance" to
1273 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1276 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1277 memset(buf, 0, bufsz);
1281 /* lustre-OST0001-osc-<instance #> */
1282 strcpy(obdname, cld->cld_logname);
1283 cname = strrchr(obdname, '-');
1284 if (cname == NULL) {
1285 CERROR("mgc %s: invalid logname %s\n",
1286 mgc->obd_name, obdname);
1290 pos = cname - obdname;
1292 pos += sprintf(obdname + pos, "-%s%04x",
1293 is_ost ? "OST" : "MDT", entry->mne_index);
1295 cname = is_ost ? "osc" : "mdc",
1296 pos += snprintf(obdname + pos, bufsz, "-%s-%s", cname, inst);
1297 lustre_cfg_bufs_reset(&bufs, obdname);
1299 /* find the obd by obdname */
1300 obd = class_name2obd(obdname);
1302 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1303 mgc->obd_name, obdname);
1305 /* this is a safe race, when the ost is starting up...*/
1309 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1312 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1315 with_imp_locked(obd, imp, rc) {
1316 /* iterate all nids to find one */
1317 /* find uuid by nid */
1318 /* create import entries if they don't exist */
1319 rc = client_import_add_nids_to_conn(
1320 imp, entry->u.nids, entry->mne_nid_count,
1321 (struct obd_uuid *)uuid);
1323 if (rc == -ENOENT && dynamic_nids) {
1324 /* create a new connection for this import */
1326 libcfs_nid2str(entry->u.nids[0]);
1327 int prim_nid_len = strlen(primary_nid) + 1;
1328 struct obd_uuid server_uuid;
1330 if (prim_nid_len > UUID_MAX)
1332 strncpy(server_uuid.uuid, primary_nid,
1335 CDEBUG(D_INFO, "Adding a connection for %s\n",
1338 rc = client_import_dyn_add_conn(
1339 imp, &server_uuid, entry->u.nids[0], 1);
1341 CERROR("%s: Failed to add new connection with NID '%s' to import: rc = %d\n",
1342 obd->obd_name, primary_nid, rc);
1345 rc = client_import_add_nids_to_conn(
1347 entry->mne_nid_count,
1348 (struct obd_uuid *)uuid);
1350 CERROR("%s: failed to lookup UUID: rc = %d\n",
1357 if (rc == -ENODEV) {
1358 /* client does not connect to the OST yet */
1363 if (rc < 0 && rc != -ENOSPC) {
1364 CERROR("mgc: cannot find UUID by nid '%s': rc = %d\n",
1365 libcfs_nid2str(entry->u.nids[0]), rc);
1369 CDEBUG(D_INFO, "Found UUID '%s' by NID '%s'\n",
1370 uuid, libcfs_nid2str(entry->u.nids[0]));
1372 pos += strlen(uuid);
1373 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1374 LASSERT(pos < bufsz);
1376 lustre_cfg_bufs_set_string(&bufs, 1, params);
1378 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1384 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1386 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1387 prev_version, max_version, obdname, params);
1389 rc = class_process_config(lcfg);
1390 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1391 lcfg->lcfg_buflens));
1393 CDEBUG(D_INFO, "process config for %s error %d\n",
1396 /* continue, even one with error */
1399 OBD_FREE(buf, PAGE_SIZE);
1405 * This function is called if this client was notified for target restarting
1406 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1409 static int mgc_process_recover_log(struct obd_device *obd,
1410 struct config_llog_data *cld)
1412 struct ptlrpc_request *req = NULL;
1413 struct config_llog_instance *cfg = &cld->cld_cfg;
1414 struct mgs_config_body *body;
1415 struct mgs_config_res *res;
1416 struct ptlrpc_bulk_desc *desc;
1417 struct page **pages = NULL;
1420 bool mne_swab = false;
1427 /* allocate buffer for bulk transfer.
1428 * if this is the first time for this mgs to read logs,
1429 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1430 * once; otherwise, it only reads increment of logs, this should be
1431 * small and CONFIG_READ_NRPAGES will be used.
1433 nrpages = CONFIG_READ_NRPAGES;
1434 if (cfg->cfg_last_idx == 0)
1435 nrpages = CONFIG_READ_NRPAGES_INIT;
1437 OBD_ALLOC_PTR_ARRAY_LARGE(pages, nrpages);
1439 GOTO(out, rc = -ENOMEM);
1441 for (i = 0; i < nrpages; i++) {
1442 pages[i] = alloc_page(GFP_KERNEL);
1443 if (pages[i] == NULL)
1444 GOTO(out, rc = -ENOMEM);
1448 LASSERT(cld_is_recover(cld));
1449 LASSERT(mutex_is_locked(&cld->cld_lock));
1450 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1451 &RQF_MGS_CONFIG_READ);
1453 GOTO(out, rc = -ENOMEM);
1455 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1460 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1461 LASSERT(body != NULL);
1462 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1463 if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1464 >= sizeof(body->mcb_name))
1465 GOTO(out, rc = -E2BIG);
1466 body->mcb_offset = cfg->cfg_last_idx + 1;
1467 body->mcb_type = cld->cld_type;
1468 body->mcb_bits = PAGE_SHIFT;
1469 body->mcb_units = nrpages;
1471 /* allocate bulk transfer descriptor */
1472 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1473 PTLRPC_BULK_PUT_SINK,
1475 &ptlrpc_bulk_kiov_pin_ops);
1477 GOTO(out, rc = -ENOMEM);
1479 for (i = 0; i < nrpages; i++)
1480 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1483 ptlrpc_request_set_replen(req);
1484 rc = ptlrpc_queue_wait(req);
1488 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1490 GOTO(out, rc = -EPROTO);
1492 if (res->mcr_size < res->mcr_offset)
1493 GOTO(out, rc = -EINVAL);
1495 /* always update the index even though it might have errors with
1496 * handling the recover logs
1498 cfg->cfg_last_idx = res->mcr_offset;
1499 eof = res->mcr_offset == res->mcr_size;
1501 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1502 res->mcr_offset, eof == false);
1504 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1506 GOTO(out, rc = ealen);
1508 if (ealen > nrpages << PAGE_SHIFT)
1509 GOTO(out, rc = -EINVAL);
1511 if (ealen == 0) { /* no logs transferred */
1517 mne_swab = req_capsule_rep_need_swab(&req->rq_pill);
1519 for (i = 0; i < nrpages && ealen > 0; i++) {
1523 ptr = kmap(pages[i]);
1524 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1525 min_t(int, ealen, PAGE_SIZE),
1529 CWARN("%s: error processing %s log recovery: rc = %d\n",
1533 GOTO(out, rc = rc2);
1541 ptlrpc_req_finished(req);
1545 if (rc == 0 && !eof)
1549 for (i = 0; i < nrpages; i++) {
1550 if (pages[i] == NULL)
1552 __free_page(pages[i]);
1554 OBD_FREE_PTR_ARRAY_LARGE(pages, nrpages);
1559 /* local_only means it cannot get remote llogs */
1560 static int mgc_process_cfg_log(struct obd_device *mgc,
1561 struct config_llog_data *cld, int local_only)
1563 struct llog_ctxt *ctxt;
1564 struct lustre_sb_info *lsi = NULL;
1570 LASSERT(mutex_is_locked(&cld->cld_lock));
1572 if (cld->cld_cfg.cfg_sb)
1573 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1579 rc = lu_env_init(env, LCT_MG_THREAD);
1583 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1585 #ifdef HAVE_SERVER_SUPPORT
1586 rc = mgc_process_server_cfg_log(env, &ctxt, lsi, mgc, cld,
1588 if (rc == -EIO && local_only)
1590 if (rc && rc != -ENOENT)
1595 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1599 * update settings on existing OBDs.
1600 * the logname must be <fsname>-sptlrpc
1602 if (rc == 0 && cld_is_sptlrpc(cld))
1603 class_notify_sptlrpc_conf(cld->cld_logname,
1604 strlen(cld->cld_logname) -
1605 strlen("-sptlrpc"));
1609 __llog_ctxt_put(env, ctxt);
1616 static bool mgc_import_in_recovery(struct obd_import *imp)
1618 bool in_recovery = true;
1620 spin_lock(&imp->imp_lock);
1621 if (imp->imp_state == LUSTRE_IMP_FULL ||
1622 imp->imp_state == LUSTRE_IMP_CLOSED)
1623 in_recovery = false;
1624 spin_unlock(&imp->imp_lock);
1630 * Get a configuration log from the MGS and process it.
1632 * This function is called for both clients and servers to process the
1633 * configuration log from the MGS. The MGC enqueues a DLM lock on the
1634 * log from the MGS, and if the lock gets revoked the MGC will be notified
1635 * by the lock cancellation callback that the config log has changed,
1636 * and will enqueue another MGS lock on it, and then continue processing
1637 * the new additions to the end of the log.
1639 * Since the MGC import is not replayable, if the import is being evicted
1640 * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
1641 * the log until recovery is finished or the import is closed.
1643 * Make a local copy of the log before parsing it if appropriate (non-MGS
1644 * server) so that the server can start even when the MGS is down.
1646 * There shouldn't be multiple processes running process_log at once --
1647 * sounds like badness. It actually might be fine, as long as they're not
1648 * trying to update from the same log simultaneously, in which case we
1649 * should use a per-log semaphore instead of cld_lock.
1651 * \param[in] mgc MGC device by which to fetch the configuration log
1652 * \param[in] cld log processing state (stored in lock callback data)
1654 * \retval 0 on success
1655 * \retval negative errno on failure
1657 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1659 struct lustre_handle lockh = { 0 };
1660 __u64 flags = LDLM_FL_NO_LRU;
1665 LASSERT(cld != NULL);
1667 /* I don't want multiple processes running process_log at once --
1668 sounds like badness. It actually might be fine, as long as
1669 we're not trying to update from the same log
1670 simultaneously (in which case we should use a per-log sem.) */
1672 mutex_lock(&cld->cld_lock);
1673 if (cld->cld_stopping) {
1674 mutex_unlock(&cld->cld_lock);
1678 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1680 CDEBUG(D_MGC, "Process log %s-%016lx from %d\n", cld->cld_logname,
1681 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1683 /* Get the cfg lock on the llog */
1684 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
1686 cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
1687 cld, 0, NULL, &lockh);
1689 /* Get the cld, it will be released in mgc_blocking_ast. */
1690 config_log_get(cld);
1691 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1692 LASSERT(!lustre_handle_is_used(&cld->cld_lockh));
1694 cld->cld_lockh = lockh;
1696 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1697 cld->cld_lockh.cookie = 0;
1699 if (rcl == -ESHUTDOWN &&
1700 atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
1701 struct obd_import *imp;
1702 long timeout = cfs_time_seconds(obd_timeout);
1704 mutex_unlock(&cld->cld_lock);
1705 imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
1707 /* Let's force the pinger, and wait the import to be
1708 * connected, note: since mgc import is non-replayable,
1709 * and even the import state is disconnected, it does
1710 * not mean the "recovery" is stopped, so we will keep
1711 * waitting until timeout or the import state is
1713 ptlrpc_pinger_force(imp);
1715 wait_event_idle_timeout(imp->imp_recovery_waitq,
1716 !mgc_import_in_recovery(imp),
1719 if (imp->imp_state == LUSTRE_IMP_FULL) {
1723 mutex_lock(&cld->cld_lock);
1724 /* unlock/lock mutex, so check stopping again */
1725 if (cld->cld_stopping) {
1726 mutex_unlock(&cld->cld_lock);
1729 spin_lock(&config_list_lock);
1730 cld->cld_lostlock = 1;
1731 spin_unlock(&config_list_lock);
1734 /* mark cld_lostlock so that it will requeue
1735 * after MGC becomes available. */
1736 spin_lock(&config_list_lock);
1737 cld->cld_lostlock = 1;
1738 spin_unlock(&config_list_lock);
1742 if (cld_is_recover(cld) && !rcl)
1743 rc = mgc_process_recover_log(mgc, cld);
1744 #ifdef HAVE_SERVER_SUPPORT
1745 else if (cld_is_nodemap(cld)) {
1749 rc = mgc_process_nodemap_log(mgc, cld);
1752 else if (!cld_is_barrier(cld))
1753 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1755 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1756 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1758 /* Now drop the lock so MGS can revoke it */
1760 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
1762 CERROR("Can't drop cfg lock: %d\n", rcl);
1764 mutex_unlock(&cld->cld_lock);
1766 /* requeue nodemap lock immediately if transfer was interrupted */
1767 if ((cld_is_nodemap(cld) && rc == -EAGAIN) ||
1768 (cld_is_recover(cld) && rc)) {
1769 if (cld_is_recover(cld))
1770 CWARN("%s: IR log %s failed, not fatal: rc = %d\n",
1771 mgc->obd_name, cld->cld_logname, rc);
1772 mgc_requeue_add(cld);
1780 /* Called from lustre_process_log.
1781 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1782 * any services, and adds it to the list logs to watch (follow).
1784 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
1786 struct lustre_cfg *lcfg = buf;
1787 struct config_llog_instance *cfg = NULL;
1792 #ifdef HAVE_SERVER_SUPPORT
1793 rc = mgc_process_config_server(obd, len, buf);
1798 switch (lcfg->lcfg_command) {
1799 case LCFG_SPTLRPC_CONF: {
1800 rc = sptlrpc_process_config(lcfg);
1803 case LCFG_LOG_START: {
1804 struct config_llog_data *cld;
1805 struct super_block *sb;
1807 logname = lustre_cfg_string(lcfg, 1);
1808 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1809 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1811 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1814 /* We're only called through here on the initial mount */
1815 cld = config_log_add(obd, logname, cfg, sb);
1821 rc = mgc_process_log(obd, cld);
1822 if (rc == 0 && cld->cld_recover != NULL) {
1823 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1824 imp_connect_data, IMP_RECOV)) {
1825 rc = mgc_process_log(obd, cld->cld_recover);
1827 struct config_llog_data *cir;
1829 mutex_lock(&cld->cld_lock);
1830 cir = cld->cld_recover;
1831 cld->cld_recover = NULL;
1832 mutex_unlock(&cld->cld_lock);
1833 config_log_put(cir);
1837 CERROR("Cannot process recover llog %d\n", rc);
1840 if (rc == 0 && cld->cld_params != NULL) {
1841 rc = mgc_process_log(obd, cld->cld_params);
1842 if (rc == -ENOENT) {
1844 "There is no params config file yet\n");
1847 /* params log is optional */
1849 CERROR("%s: can't process params llog: rc = %d\n",
1855 case LCFG_LOG_END: {
1856 logname = lustre_cfg_string(lcfg, 1);
1858 if (lcfg->lcfg_bufcount >= 2)
1859 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1860 rc = config_log_end(logname, cfg);
1864 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1872 static const struct obd_ops mgc_obd_ops = {
1873 .o_owner = THIS_MODULE,
1874 .o_setup = mgc_setup,
1875 .o_precleanup = mgc_precleanup,
1876 .o_cleanup = mgc_cleanup,
1877 .o_add_conn = client_import_add_conn,
1878 .o_del_conn = client_import_del_conn,
1879 .o_connect = client_connect_import,
1880 .o_disconnect = client_disconnect_export,
1881 .o_set_info_async = mgc_set_info_async,
1882 .o_get_info = mgc_get_info,
1883 .o_import_event = mgc_import_event,
1884 .o_process_config = mgc_process_config,
1887 static int mgc_param_requeue_timeout_min_set(const char *val,
1888 cfs_kernel_param_arg_t *kp)
1893 rc = kstrtouint(val, 0, &num);
1899 mgc_requeue_timeout_min = num;
1904 static const struct kernel_param_ops param_ops_requeue_timeout_min = {
1905 .set = mgc_param_requeue_timeout_min_set,
1906 .get = param_get_uint,
1909 #define param_check_requeue_timeout_min(name, p) \
1910 __param_check(name, p, unsigned int)
1912 unsigned int mgc_requeue_timeout_min = MGC_TIMEOUT_MIN_SECONDS;
1913 #ifdef HAVE_KERNEL_PARAM_OPS
1914 module_param(mgc_requeue_timeout_min, requeue_timeout_min, 0644);
1916 module_param_call(mgc_requeue_timeout_min, mgc_param_requeue_timeout_min_set,
1917 param_get_uint, ¶m_ops_requeue_timeout_min, 0644);
1919 MODULE_PARM_DESC(mgc_requeue_timeout_min, "Minimal requeue time to refresh logs");
1921 static int __init mgc_init(void)
1923 return class_register_type(&mgc_obd_ops, NULL, false,
1924 LUSTRE_MGC_NAME, NULL);
1927 static void __exit mgc_exit(void)
1929 class_unregister_type(LUSTRE_MGC_NAME);
1932 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1933 MODULE_DESCRIPTION("Lustre Management Client");
1934 MODULE_VERSION(LUSTRE_VERSION_STRING);
1935 MODULE_LICENSE("GPL");
1937 module_init(mgc_init);
1938 module_exit(mgc_exit);