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;
265 bool is_server = IS_SERVER(s2lsi(sb));
268 #ifdef HAVE_SERVER_SUPPORT
269 if (IS_OST(s2lsi(sb)))
272 /* for osp-on-ost, see lustre_start_osp() */
273 if (IS_MDT(s2lsi(sb)) && lcfg.cfg_instance)
276 /* We have to use different llog for clients and MDTs for DNE,
277 * where only clients are notified if one of DNE server restarts.
279 LASSERT(strlen(fsname) < sizeof(logname) / 2);
280 strncpy(logname, fsname, sizeof(logname));
282 LASSERT(is_server ? lcfg.cfg_instance == 0 : lcfg.cfg_instance != 0);
284 lcfg.cfg_instance = ll_get_cfg_instance(sb);
285 scnprintf(logname, sizeof(logname), "%s-%s", fsname,
286 is_server ? "mdtir" : "cliir");
288 cld = do_config_log_add(obd, logname, MGS_CFG_T_RECOVER, &lcfg, sb);
292 static struct config_llog_data *
293 config_log_find_or_add(struct obd_device *obd, char *logname,
294 struct super_block *sb, enum mgs_cfg_type type,
295 struct config_llog_instance *cfg)
297 struct config_llog_instance lcfg = *cfg;
298 struct config_llog_data *cld;
300 /* Note class_config_llog_handler() depends on getting "obd" back */
301 lcfg.cfg_instance = sb ? ll_get_cfg_instance(sb) : (unsigned long)obd;
303 cld = config_log_find(logname, &lcfg);
304 if (unlikely(cld != NULL))
307 return do_config_log_add(obd, logname, type, &lcfg, sb);
310 /** Add this log to the list of active logs watched by an MGC.
311 * Active means we're watching for updates.
312 * We have one active log per "mount" - client instance or servername.
313 * Each instance may be at a different point in the log.
315 static struct config_llog_data *
316 config_log_add(struct obd_device *obd, char *logname,
317 struct config_llog_instance *cfg, struct super_block *sb)
319 struct lustre_sb_info *lsi = s2lsi(sb);
320 struct config_llog_data *cld = NULL;
321 struct config_llog_data *sptlrpc_cld = NULL;
322 struct config_llog_data *params_cld = NULL;
323 #ifdef HAVE_SERVER_SUPPORT
324 struct config_llog_data *nodemap_cld = NULL;
325 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: "
419 "rc = %d\n", 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_printf(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
599 static int rq_state = 0;
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)
609 LASSERT(atomic_read(&cld->cld_refcount) > 0);
612 * Do not run mgc_process_log on a disconnected export or an
613 * export which is being disconnected. Take the client
614 * semaphore to make the check non-racy.
616 down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
618 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
619 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
620 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
621 if (rc && rc != -ENOENT)
622 CERROR("failed processing log: %d\n", rc);
624 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
627 up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
632 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 prandom_u32_max(cfs_time_seconds(to));
663 wait_event_idle_timeout(rq_waitq,
664 rq_state & (RQ_STOP | RQ_PRECLEANUP), to);
667 * iterate & processing through the list. for each cld, process
668 * its depending sptlrpc cld firstly (if any) and then itself.
670 * it's guaranteed any item in the list must have
671 * reference > 0; and if cld_lostlock is set, at
672 * least one reference is taken by the previous enqueue.
676 spin_lock(&config_list_lock);
677 rq_state &= ~RQ_PRECLEANUP;
678 list_for_each_entry(cld, &config_llog_list,
680 if (!cld->cld_lostlock || cld->cld_stopping)
683 /* hold reference to avoid being freed during
684 * subsequent processing. */
686 cld->cld_lostlock = 0;
687 spin_unlock(&config_list_lock);
689 config_log_put(cld_prev);
692 if (likely(!(rq_state & RQ_STOP))) {
694 spin_lock(&config_list_lock);
696 spin_lock(&config_list_lock);
700 spin_unlock(&config_list_lock);
701 config_log_put(cld_prev);
703 /* Wait a bit to see if anyone else needs a requeue */
704 wait_event_idle(rq_waitq, rq_state & (RQ_NOW | RQ_STOP));
705 spin_lock(&config_list_lock);
708 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
709 rq_state &= ~RQ_RUNNING;
710 spin_unlock(&config_list_lock);
714 CDEBUG(D_MGC, "Ending requeue thread\n");
718 /* Add a cld to the list to requeue. Start the requeue thread if needed.
719 We are responsible for dropping the config log reference from here on out. */
720 static void mgc_requeue_add(struct config_llog_data *cld)
725 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
726 cld->cld_logname, atomic_read(&cld->cld_refcount),
727 cld->cld_stopping, rq_state);
728 LASSERT(atomic_read(&cld->cld_refcount) > 0);
730 /* lets cancel an existent lock to mark cld as "lostlock" */
731 CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
732 if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
733 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
735 mutex_lock(&cld->cld_lock);
736 spin_lock(&config_list_lock);
737 if (!(rq_state & RQ_STOP) && !cld->cld_stopping) {
738 cld->cld_lostlock = 1;
742 spin_unlock(&config_list_lock);
743 mutex_unlock(&cld->cld_lock);
750 /********************** class fns **********************/
751 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
753 struct llog_ctxt *ctxt;
758 /* setup only remote ctxt, the local disk context is switched per each
759 * filesystem during mgc_fs_setup() */
760 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
765 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
768 llog_initiator_connect(ctxt);
774 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
776 struct llog_ctxt *ctxt;
780 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
782 llog_cleanup(env, ctxt);
788 static atomic_t mgc_count = ATOMIC_INIT(0);
789 static int mgc_precleanup(struct obd_device *obd)
795 if (atomic_dec_and_test(&mgc_count)) {
796 LASSERT(rq_state & RQ_RUNNING);
797 /* stop requeue thread */
800 /* wakeup requeue thread to clean our cld */
801 temp = RQ_NOW | RQ_PRECLEANUP;
804 spin_lock(&config_list_lock);
806 spin_unlock(&config_list_lock);
810 wait_for_completion(&rq_exit);
811 obd_cleanup_client_import(obd);
813 rc = mgc_llog_fini(NULL, obd);
815 CERROR("failed to cleanup llogging subsystems\n");
820 static int mgc_cleanup(struct obd_device *obd)
825 /* COMPAT_146 - old config logs may have added profiles we don't
827 if (atomic_read(&obd->obd_type->typ_refcnt) <= 1)
828 /* Only for the last mgc */
829 class_del_profiles();
831 lprocfs_obd_cleanup(obd);
834 rc = client_obd_cleanup(obd);
838 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
840 struct task_struct *task;
844 rc = ptlrpcd_addref();
848 rc = client_obd_setup(obd, lcfg);
850 GOTO(err_decref, rc);
852 rc = mgc_llog_init(NULL, obd);
854 CERROR("failed to setup llogging subsystems\n");
855 GOTO(err_cleanup, rc);
858 rc = mgc_tunables_init(obd);
862 if (atomic_inc_return(&mgc_count) == 1) {
864 init_waitqueue_head(&rq_waitq);
866 /* start requeue thread */
867 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
870 CERROR("%s: cannot start requeue thread: rc = %d; "
871 "no more log updates\n",
875 /* rc is the task_struct pointer of mgc_requeue_thread. */
877 wait_for_completion(&rq_start);
883 lprocfs_obd_cleanup(obd);
885 client_obd_cleanup(obd);
891 /* based on ll_mdc_blocking_ast */
892 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
893 void *data, int flag)
895 struct lustre_handle lockh;
896 struct config_llog_data *cld = (struct config_llog_data *)data;
901 case LDLM_CB_BLOCKING:
902 /* mgs wants the lock, give it up... */
903 LDLM_DEBUG(lock, "MGC blocking CB");
904 ldlm_lock2handle(lock, &lockh);
905 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
907 case LDLM_CB_CANCELING:
908 /* We've given up the lock, prepare ourselves to update. */
909 LDLM_DEBUG(lock, "MGC cancel CB");
911 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
912 PLDLMRES(lock->l_resource),
913 (char *)&lock->l_resource->lr_name.name[0]);
916 CDEBUG(D_INFO, "missing data, won't requeue\n");
920 /* held at mgc_process_log(). */
921 LASSERT(atomic_read(&cld->cld_refcount) > 0);
923 lock->l_ast_data = NULL;
924 cld->cld_lockh.cookie = 0;
925 /* Are we done with this log? */
926 if (cld->cld_stopping) {
927 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
932 /* Make sure not to re-enqueue when the mgc is stopping
933 (we get called from client_disconnect_export) */
934 if (lock->l_conn_export == NULL ||
935 lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
936 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
943 mgc_requeue_add(cld);
953 /* Take a config lock so we can get cancel notifications */
954 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
955 union ldlm_policy_data *policy, enum ldlm_mode mode,
956 __u64 *flags, ldlm_glimpse_callback glimpse_callback,
957 void *data, __u32 lvb_len, void *lvb_swabber,
958 struct lustre_handle *lockh)
960 struct config_llog_data *cld = (struct config_llog_data *)data;
961 struct ldlm_enqueue_info einfo = {
964 .ei_cb_bl = mgc_blocking_ast,
965 .ei_cb_cp = ldlm_completion_ast,
966 .ei_cb_gl = glimpse_callback,
968 struct ptlrpc_request *req;
969 int short_limit = cld_is_sptlrpc(cld);
976 CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
977 cld->cld_resid.name[0]);
979 /* We need a callback for every lockholder, so don't try to
980 ldlm_lock_match (see rev 1.1.2.11.2.47) */
981 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
982 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
987 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
988 ptlrpc_request_set_replen(req);
990 /* check if this is server or client */
991 if (cld->cld_cfg.cfg_sb &&
992 IS_SERVER(s2lsi(cld->cld_cfg.cfg_sb)))
995 /* Limit how long we will wait for the enqueue to complete */
996 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
997 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
998 NULL, 0, LVB_T_NONE, lockh, 0);
999 /* A failed enqueue should still call the mgc_blocking_ast,
1000 where it will be requeued if needed ("grant failed"). */
1001 ptlrpc_req_finished(req);
1005 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1006 struct lustre_handle *lockh)
1010 ldlm_lock_decref(lockh, mode);
1015 static void mgc_notify_active(struct obd_device *unused)
1017 /* wakeup mgc_requeue_thread to requeue mgc lock */
1018 spin_lock(&config_list_lock);
1020 spin_unlock(&config_list_lock);
1023 /* TODO: Help the MGS rebuild nidtbl. -jay */
1026 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1027 u32 keylen, void *key,
1028 u32 vallen, void *val,
1029 struct ptlrpc_request_set *set)
1034 /* Turn off initial_recov after we try all backup servers once */
1035 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1036 struct obd_import *imp = class_exp2cliimp(exp);
1038 if (vallen != sizeof(int))
1040 value = *(int *)val;
1041 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1042 imp->imp_obd->obd_name, value,
1043 imp->imp_deactive, imp->imp_invalid,
1044 imp->imp_replayable, imp->imp_obd->obd_replayable,
1045 ptlrpc_import_state_name(imp->imp_state));
1046 /* Resurrect the import immediately if
1047 * 1. we previously got disconnected,
1048 * 2. value > 1 (at the same node with MGS)
1050 if (imp->imp_state == LUSTRE_IMP_DISCON || value > 1)
1051 ptlrpc_reconnect_import(imp);
1056 if (KEY_IS(KEY_MGSSEC)) {
1057 struct client_obd *cli = &exp->exp_obd->u.cli;
1058 struct sptlrpc_flavor flvr;
1061 * empty string means using current flavor, if which haven't
1062 * been set yet, set it as null.
1064 * if flavor has been set previously, check the asking flavor
1065 * must match the existing one.
1068 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1074 rc = sptlrpc_parse_flavor(val, &flvr);
1076 CERROR("invalid sptlrpc flavor %s to MGS\n",
1082 * caller already hold a mutex
1084 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1085 cli->cl_flvr_mgc = flvr;
1086 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1087 sizeof(flvr)) != 0) {
1090 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1092 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1093 "currently %s is in use\n",
1100 #ifdef HAVE_SERVER_SUPPORT
1101 rc = mgc_set_info_async_server(env, exp, keylen, key, vallen, val, set);
1106 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1107 __u32 keylen, void *key, __u32 *vallen, void *val)
1111 if (KEY_IS(KEY_CONN_DATA)) {
1112 struct obd_import *imp = class_exp2cliimp(exp);
1113 struct obd_connect_data *data = val;
1115 if (*vallen == sizeof(*data)) {
1116 *data = imp->imp_connect_data;
1124 static int mgc_import_event(struct obd_device *obd,
1125 struct obd_import *imp,
1126 enum obd_import_event event)
1130 LASSERT(imp->imp_obd == obd);
1131 CDEBUG(D_MGC, "import event %#x\n", event);
1134 case IMP_EVENT_DISCON:
1135 /* MGC imports should not wait for recovery */
1136 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1137 ptlrpc_pinger_ir_down();
1139 case IMP_EVENT_INACTIVE:
1141 case IMP_EVENT_INVALIDATE: {
1142 struct ldlm_namespace *ns = obd->obd_namespace;
1143 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1146 case IMP_EVENT_ACTIVE:
1147 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1148 /* Clearing obd_no_recov allows us to continue pinging */
1149 obd->obd_no_recov = 0;
1150 mgc_notify_active(obd);
1151 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1152 ptlrpc_pinger_ir_up();
1156 case IMP_EVENT_DEACTIVATE:
1157 case IMP_EVENT_ACTIVATE:
1160 CERROR("Unknown import event %#x\n", event);
1166 static int mgc_apply_recover_logs(struct obd_device *mgc,
1167 struct config_llog_data *cld,
1169 void *data, int datalen, bool mne_swab)
1171 struct config_llog_instance *cfg = &cld->cld_cfg;
1172 struct mgs_nidtbl_entry *entry;
1173 struct lustre_cfg *lcfg;
1174 struct lustre_cfg_bufs bufs;
1175 u64 prev_version = 0;
1176 char inst[MTI_NAME_MAXLEN + 1];
1182 unsigned long dynamic_nids;
1185 LASSERT(cfg->cfg_instance != 0);
1186 LASSERT(ll_get_cfg_instance(cfg->cfg_sb) == cfg->cfg_instance);
1188 /* get dynamic nids setting */
1189 dynamic_nids = mgc->obd_dynamic_nids;
1191 if (!IS_SERVER(s2lsi(cfg->cfg_sb))) {
1192 pos = snprintf(inst, sizeof(inst), "%016lx", cfg->cfg_instance);
1193 if (pos >= PAGE_SIZE)
1195 #ifdef HAVE_SERVER_SUPPORT
1197 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1199 LASSERT(IS_MDT(lsi));
1200 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1204 #endif /* HAVE_SERVER_SUPPORT */
1207 OBD_ALLOC(buf, PAGE_SIZE);
1213 while (datalen > 0) {
1214 int entry_len = sizeof(*entry);
1216 struct obd_device *obd;
1217 struct obd_import *imp;
1224 if (datalen < sizeof(*entry))
1227 entry = (typeof(entry))(data + off);
1230 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1232 if (entry->mne_nid_count == 0) /* at least one nid entry */
1234 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1237 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1238 if (datalen < entry_len) /* must have entry_len at least */
1241 /* Keep this swab for normal mixed endian handling. LU-1644 */
1243 lustre_swab_mgs_nidtbl_entry(entry);
1244 if (entry->mne_length > PAGE_SIZE) {
1245 CERROR("MNE too large (%u)\n", entry->mne_length);
1249 if (entry->mne_length < entry_len)
1252 off += entry->mne_length;
1253 datalen -= entry->mne_length;
1257 if (entry->mne_version > max_version) {
1258 CERROR("entry index(%lld) is over max_index(%lld)\n",
1259 entry->mne_version, max_version);
1263 if (prev_version >= entry->mne_version) {
1264 CERROR("index unsorted, prev %lld, now %lld\n",
1265 prev_version, entry->mne_version);
1268 prev_version = entry->mne_version;
1271 * Write a string with format "nid::instance" to
1272 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1275 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1276 memset(buf, 0, bufsz);
1280 /* lustre-OST0001-osc-<instance #> */
1281 strcpy(obdname, cld->cld_logname);
1282 cname = strrchr(obdname, '-');
1283 if (cname == NULL) {
1284 CERROR("mgc %s: invalid logname %s\n",
1285 mgc->obd_name, obdname);
1289 pos = cname - obdname;
1291 pos += sprintf(obdname + pos, "-%s%04x",
1292 is_ost ? "OST" : "MDT", entry->mne_index);
1294 cname = is_ost ? "osc" : "mdc",
1295 pos += snprintf(obdname + pos, bufsz, "-%s-%s", cname, inst);
1296 lustre_cfg_bufs_reset(&bufs, obdname);
1298 /* find the obd by obdname */
1299 obd = class_name2obd(obdname);
1301 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1302 mgc->obd_name, obdname);
1304 /* this is a safe race, when the ost is starting up...*/
1308 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1311 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1314 with_imp_locked(obd, imp, rc) {
1315 /* iterate all nids to find one */
1316 /* find uuid by nid */
1317 /* create import entries if they don't exist */
1318 rc = client_import_add_nids_to_conn(
1319 imp, entry->u.nids, entry->mne_nid_count,
1320 (struct obd_uuid *)uuid);
1322 if (rc == -ENOENT && dynamic_nids) {
1323 /* create a new connection for this import */
1325 libcfs_nid2str(entry->u.nids[0]);
1326 int prim_nid_len = strlen(primary_nid) + 1;
1327 struct obd_uuid server_uuid;
1329 if (prim_nid_len > UUID_MAX)
1331 strncpy(server_uuid.uuid, primary_nid,
1334 CDEBUG(D_INFO, "Adding a connection for %s\n",
1337 rc = client_import_dyn_add_conn(
1338 imp, &server_uuid, entry->u.nids[0], 1);
1340 CERROR("%s: Failed to add new connection with NID '%s' to import: rc = %d\n",
1341 obd->obd_name, primary_nid, rc);
1344 rc = client_import_add_nids_to_conn(
1346 entry->mne_nid_count,
1347 (struct obd_uuid *)uuid);
1349 CERROR("%s: failed to lookup UUID: rc = %d\n",
1356 if (rc == -ENODEV) {
1357 /* client does not connect to the OST yet */
1362 if (rc < 0 && rc != -ENOSPC) {
1363 CERROR("mgc: cannot find UUID by nid '%s': rc = %d\n",
1364 libcfs_nid2str(entry->u.nids[0]), rc);
1368 CDEBUG(D_INFO, "Found UUID '%s' by NID '%s'\n",
1369 uuid, libcfs_nid2str(entry->u.nids[0]));
1371 pos += strlen(uuid);
1372 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1373 LASSERT(pos < bufsz);
1375 lustre_cfg_bufs_set_string(&bufs, 1, params);
1377 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1383 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1385 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1386 prev_version, max_version, obdname, params);
1388 rc = class_process_config(lcfg);
1389 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1390 lcfg->lcfg_buflens));
1392 CDEBUG(D_INFO, "process config for %s error %d\n",
1395 /* continue, even one with error */
1398 OBD_FREE(buf, PAGE_SIZE);
1404 * This function is called if this client was notified for target restarting
1405 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1408 static int mgc_process_recover_log(struct obd_device *obd,
1409 struct config_llog_data *cld)
1411 struct ptlrpc_request *req = NULL;
1412 struct config_llog_instance *cfg = &cld->cld_cfg;
1413 struct mgs_config_body *body;
1414 struct mgs_config_res *res;
1415 struct ptlrpc_bulk_desc *desc;
1416 struct page **pages = NULL;
1419 bool mne_swab = false;
1426 /* allocate buffer for bulk transfer.
1427 * if this is the first time for this mgs to read logs,
1428 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1429 * once; otherwise, it only reads increment of logs, this should be
1430 * small and CONFIG_READ_NRPAGES will be used.
1432 nrpages = CONFIG_READ_NRPAGES;
1433 if (cfg->cfg_last_idx == 0)
1434 nrpages = CONFIG_READ_NRPAGES_INIT;
1436 OBD_ALLOC_PTR_ARRAY_LARGE(pages, nrpages);
1438 GOTO(out, rc = -ENOMEM);
1440 for (i = 0; i < nrpages; i++) {
1441 pages[i] = alloc_page(GFP_KERNEL);
1442 if (pages[i] == NULL)
1443 GOTO(out, rc = -ENOMEM);
1447 LASSERT(cld_is_recover(cld));
1448 LASSERT(mutex_is_locked(&cld->cld_lock));
1449 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1450 &RQF_MGS_CONFIG_READ);
1452 GOTO(out, rc = -ENOMEM);
1454 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1459 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1460 LASSERT(body != NULL);
1461 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1462 if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1463 >= sizeof(body->mcb_name))
1464 GOTO(out, rc = -E2BIG);
1465 body->mcb_offset = cfg->cfg_last_idx + 1;
1466 body->mcb_type = cld->cld_type;
1467 body->mcb_bits = PAGE_SHIFT;
1468 body->mcb_units = nrpages;
1470 /* allocate bulk transfer descriptor */
1471 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1472 PTLRPC_BULK_PUT_SINK,
1474 &ptlrpc_bulk_kiov_pin_ops);
1476 GOTO(out, rc = -ENOMEM);
1478 for (i = 0; i < nrpages; i++)
1479 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1482 ptlrpc_request_set_replen(req);
1483 rc = ptlrpc_queue_wait(req);
1487 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1489 GOTO(out, rc = -EPROTO);
1491 if (res->mcr_size < res->mcr_offset)
1492 GOTO(out, rc = -EINVAL);
1494 /* always update the index even though it might have errors with
1495 * handling the recover logs
1497 cfg->cfg_last_idx = res->mcr_offset;
1498 eof = res->mcr_offset == res->mcr_size;
1500 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1501 res->mcr_offset, eof == false);
1503 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1505 GOTO(out, rc = ealen);
1507 if (ealen > nrpages << PAGE_SHIFT)
1508 GOTO(out, rc = -EINVAL);
1510 if (ealen == 0) { /* no logs transferred */
1516 mne_swab = req_capsule_rep_need_swab(&req->rq_pill);
1518 for (i = 0; i < nrpages && ealen > 0; i++) {
1522 ptr = kmap(pages[i]);
1523 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1524 min_t(int, ealen, PAGE_SIZE),
1528 CWARN("%s: error processing %s log recovery: rc = %d\n",
1532 GOTO(out, rc = rc2);
1540 ptlrpc_req_finished(req);
1544 if (rc == 0 && !eof)
1548 for (i = 0; i < nrpages; i++) {
1549 if (pages[i] == NULL)
1551 __free_page(pages[i]);
1553 OBD_FREE_PTR_ARRAY_LARGE(pages, nrpages);
1558 /* local_only means it cannot get remote llogs */
1559 static int mgc_process_cfg_log(struct obd_device *mgc,
1560 struct config_llog_data *cld, int local_only)
1562 struct llog_ctxt *ctxt;
1563 struct lustre_sb_info *lsi = NULL;
1569 LASSERT(mutex_is_locked(&cld->cld_lock));
1571 if (cld->cld_cfg.cfg_sb)
1572 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1578 rc = lu_env_init(env, LCT_MG_THREAD);
1582 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1584 #ifdef HAVE_SERVER_SUPPORT
1585 rc = mgc_process_server_cfg_log(env, &ctxt, lsi, mgc, cld,
1587 if (rc == -EIO && local_only)
1589 if (rc && rc != -ENOENT)
1594 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1598 * update settings on existing OBDs.
1599 * the logname must be <fsname>-sptlrpc
1601 if (rc == 0 && cld_is_sptlrpc(cld))
1602 class_notify_sptlrpc_conf(cld->cld_logname,
1603 strlen(cld->cld_logname) -
1604 strlen("-sptlrpc"));
1608 __llog_ctxt_put(env, ctxt);
1615 static bool mgc_import_in_recovery(struct obd_import *imp)
1617 bool in_recovery = true;
1619 spin_lock(&imp->imp_lock);
1620 if (imp->imp_state == LUSTRE_IMP_FULL ||
1621 imp->imp_state == LUSTRE_IMP_CLOSED)
1622 in_recovery = false;
1623 spin_unlock(&imp->imp_lock);
1629 * Get a configuration log from the MGS and process it.
1631 * This function is called for both clients and servers to process the
1632 * configuration log from the MGS. The MGC enqueues a DLM lock on the
1633 * log from the MGS, and if the lock gets revoked the MGC will be notified
1634 * by the lock cancellation callback that the config log has changed,
1635 * and will enqueue another MGS lock on it, and then continue processing
1636 * the new additions to the end of the log.
1638 * Since the MGC import is not replayable, if the import is being evicted
1639 * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
1640 * the log until recovery is finished or the import is closed.
1642 * Make a local copy of the log before parsing it if appropriate (non-MGS
1643 * server) so that the server can start even when the MGS is down.
1645 * There shouldn't be multiple processes running process_log at once --
1646 * sounds like badness. It actually might be fine, as long as they're not
1647 * trying to update from the same log simultaneously, in which case we
1648 * should use a per-log semaphore instead of cld_lock.
1650 * \param[in] mgc MGC device by which to fetch the configuration log
1651 * \param[in] cld log processing state (stored in lock callback data)
1653 * \retval 0 on success
1654 * \retval negative errno on failure
1656 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1658 struct lustre_handle lockh = { 0 };
1659 __u64 flags = LDLM_FL_NO_LRU;
1664 LASSERT(cld != NULL);
1666 /* I don't want multiple processes running process_log at once --
1667 sounds like badness. It actually might be fine, as long as
1668 we're not trying to update from the same log
1669 simultaneously (in which case we should use a per-log sem.) */
1671 mutex_lock(&cld->cld_lock);
1672 if (cld->cld_stopping) {
1673 mutex_unlock(&cld->cld_lock);
1677 CFS_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1679 CDEBUG(D_MGC, "Process log %s-%016lx from %d\n", cld->cld_logname,
1680 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1682 /* Get the cfg lock on the llog */
1683 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
1685 cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
1686 cld, 0, NULL, &lockh);
1688 /* Get the cld, it will be released in mgc_blocking_ast. */
1689 config_log_get(cld);
1690 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1691 LASSERT(!lustre_handle_is_used(&cld->cld_lockh));
1693 cld->cld_lockh = lockh;
1695 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1696 cld->cld_lockh.cookie = 0;
1698 if (rcl == -ESHUTDOWN &&
1699 atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
1700 struct obd_import *imp;
1701 long timeout = cfs_time_seconds(obd_timeout);
1703 mutex_unlock(&cld->cld_lock);
1704 imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
1706 /* Let's force the pinger, and wait the import to be
1707 * connected, note: since mgc import is non-replayable,
1708 * and even the import state is disconnected, it does
1709 * not mean the "recovery" is stopped, so we will keep
1710 * waitting until timeout or the import state is
1712 ptlrpc_pinger_force(imp);
1714 wait_event_idle_timeout(imp->imp_recovery_waitq,
1715 !mgc_import_in_recovery(imp),
1718 if (imp->imp_state == LUSTRE_IMP_FULL) {
1722 mutex_lock(&cld->cld_lock);
1723 /* unlock/lock mutex, so check stopping again */
1724 if (cld->cld_stopping) {
1725 mutex_unlock(&cld->cld_lock);
1728 spin_lock(&config_list_lock);
1729 cld->cld_lostlock = 1;
1730 spin_unlock(&config_list_lock);
1733 /* mark cld_lostlock so that it will requeue
1734 * after MGC becomes available. */
1735 spin_lock(&config_list_lock);
1736 cld->cld_lostlock = 1;
1737 spin_unlock(&config_list_lock);
1741 if (cld_is_recover(cld) && !rcl)
1742 rc = mgc_process_recover_log(mgc, cld);
1743 #ifdef HAVE_SERVER_SUPPORT
1744 else if (cld_is_nodemap(cld)) {
1748 rc = mgc_process_nodemap_log(mgc, cld);
1751 else if (!cld_is_barrier(cld))
1752 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1754 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1755 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1757 /* Now drop the lock so MGS can revoke it */
1759 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
1761 CERROR("Can't drop cfg lock: %d\n", rcl);
1763 mutex_unlock(&cld->cld_lock);
1765 /* requeue nodemap lock immediately if transfer was interrupted */
1766 if ((cld_is_nodemap(cld) && rc == -EAGAIN) ||
1767 (cld_is_recover(cld) && rc)) {
1768 if (cld_is_recover(cld))
1769 CWARN("%s: IR log %s failed, not fatal: rc = %d\n",
1770 mgc->obd_name, cld->cld_logname, rc);
1771 mgc_requeue_add(cld);
1779 /* Called from lustre_process_log.
1780 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1781 * any services, and adds it to the list logs to watch (follow).
1783 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
1785 struct lustre_cfg *lcfg = buf;
1786 struct config_llog_instance *cfg = NULL;
1791 #ifdef HAVE_SERVER_SUPPORT
1792 rc = mgc_process_config_server(obd, len, buf);
1797 switch (lcfg->lcfg_command) {
1798 case LCFG_SPTLRPC_CONF: {
1799 rc = sptlrpc_process_config(lcfg);
1802 case LCFG_LOG_START: {
1803 struct config_llog_data *cld;
1804 struct super_block *sb;
1806 logname = lustre_cfg_string(lcfg, 1);
1807 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1808 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1810 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1813 /* We're only called through here on the initial mount */
1814 cld = config_log_add(obd, logname, cfg, sb);
1820 rc = mgc_process_log(obd, cld);
1821 if (rc == 0 && cld->cld_recover != NULL) {
1822 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1823 imp_connect_data, IMP_RECOV)) {
1824 rc = mgc_process_log(obd, cld->cld_recover);
1826 struct config_llog_data *cir;
1828 mutex_lock(&cld->cld_lock);
1829 cir = cld->cld_recover;
1830 cld->cld_recover = NULL;
1831 mutex_unlock(&cld->cld_lock);
1832 config_log_put(cir);
1836 CERROR("Cannot process recover llog %d\n", rc);
1839 if (rc == 0 && cld->cld_params != NULL) {
1840 rc = mgc_process_log(obd, cld->cld_params);
1841 if (rc == -ENOENT) {
1843 "There is no params config file yet\n");
1846 /* params log is optional */
1848 CERROR("%s: can't process params llog: rc = %d\n",
1854 case LCFG_LOG_END: {
1855 logname = lustre_cfg_string(lcfg, 1);
1857 if (lcfg->lcfg_bufcount >= 2)
1858 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1859 rc = config_log_end(logname, cfg);
1863 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1871 static const struct obd_ops mgc_obd_ops = {
1872 .o_owner = THIS_MODULE,
1873 .o_setup = mgc_setup,
1874 .o_precleanup = mgc_precleanup,
1875 .o_cleanup = mgc_cleanup,
1876 .o_add_conn = client_import_add_conn,
1877 .o_del_conn = client_import_del_conn,
1878 .o_connect = client_connect_import,
1879 .o_disconnect = client_disconnect_export,
1880 .o_set_info_async = mgc_set_info_async,
1881 .o_get_info = mgc_get_info,
1882 .o_import_event = mgc_import_event,
1883 .o_process_config = mgc_process_config,
1886 static int mgc_param_requeue_timeout_min_set(const char *val,
1887 cfs_kernel_param_arg_t *kp)
1892 rc = kstrtouint(val, 0, &num);
1898 mgc_requeue_timeout_min = num;
1903 static const struct kernel_param_ops param_ops_requeue_timeout_min = {
1904 .set = mgc_param_requeue_timeout_min_set,
1905 .get = param_get_uint,
1908 #define param_check_requeue_timeout_min(name, p) \
1909 __param_check(name, p, unsigned int)
1911 unsigned int mgc_requeue_timeout_min = MGC_TIMEOUT_MIN_SECONDS;
1912 #ifdef HAVE_KERNEL_PARAM_OPS
1913 module_param(mgc_requeue_timeout_min, requeue_timeout_min, 0644);
1915 module_param_call(mgc_requeue_timeout_min, mgc_param_requeue_timeout_min_set,
1916 param_get_uint, ¶m_ops_requeue_timeout_min, 0644);
1918 MODULE_PARM_DESC(mgc_requeue_timeout_min, "Minimal requeue time to refresh logs");
1920 static int __init mgc_init(void)
1922 return class_register_type(&mgc_obd_ops, NULL, false,
1923 LUSTRE_MGC_NAME, NULL);
1926 static void __exit mgc_exit(void)
1928 class_unregister_type(LUSTRE_MGC_NAME);
1931 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1932 MODULE_DESCRIPTION("Lustre Management Client");
1933 MODULE_VERSION(LUSTRE_VERSION_STRING);
1934 MODULE_LICENSE("GPL");
1936 module_init(mgc_init);
1937 module_exit(mgc_exit);