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/kthread.h>
41 #include <linux/random.h>
43 #include <dt_object.h>
44 #include <lprocfs_status.h>
45 #include <lustre_dlm.h>
46 #include <lustre_disk.h>
47 #include <lustre_log.h>
48 #include <lustre_nodemap.h>
49 #include <lustre_swab.h>
50 #include <obd_class.h>
51 #include <lustre_barrier.h>
53 #include "mgc_internal.h"
55 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
60 if (len > sizeof(resname)) {
61 CERROR("name too long: %s\n", name);
65 CERROR("missing name: %s\n", name);
68 memcpy(&resname, name, len);
70 /* Always use the same endianness for the resid */
71 memset(res_id, 0, sizeof(*res_id));
72 res_id->name[0] = cpu_to_le64(resname);
73 /* XXX: unfortunately, sptlprc and config llog share one lock */
76 case CONFIG_T_SPTLRPC:
79 case CONFIG_T_RECOVER:
81 case CONFIG_T_NODEMAP:
82 case CONFIG_T_BARRIER:
88 res_id->name[1] = cpu_to_le64(resname);
89 CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
90 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
94 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
96 /* fsname is at most 8 chars long, maybe contain "-".
97 * e.g. "lustre", "SUN-000" */
98 return mgc_name2resid(fsname, strlen(fsname), res_id, type);
100 EXPORT_SYMBOL(mgc_fsname2resid);
102 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int 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);
153 config_log_put(cld->cld_barrier);
154 config_log_put(cld->cld_recover);
155 config_log_put(cld->cld_params);
156 config_log_put(cld->cld_nodemap);
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,
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 atomic_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 *config_recover_log_add(struct obd_device *obd,
259 struct config_llog_instance *cfg,
260 struct super_block *sb)
262 struct config_llog_instance lcfg = *cfg;
263 struct lustre_sb_info *lsi = s2lsi(sb);
264 struct config_llog_data *cld;
270 /* for osp-on-ost, see lustre_start_osp() */
271 if (IS_MDT(lsi) && lcfg.cfg_instance)
274 /* We have to use different llog for clients and MDTs for DNE,
275 * where only clients are notified if one of DNE server restarts.
277 LASSERT(strlen(fsname) < sizeof(logname) / 2);
278 strncpy(logname, fsname, sizeof(logname));
279 if (IS_SERVER(lsi)) { /* mdt */
280 LASSERT(lcfg.cfg_instance == 0);
281 lcfg.cfg_instance = ll_get_cfg_instance(sb);
282 strncat(logname, "-mdtir", sizeof(logname));
284 LASSERT(lcfg.cfg_instance != 0);
285 strncat(logname, "-cliir", sizeof(logname));
288 cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
292 static struct config_llog_data *config_log_find_or_add(struct obd_device *obd,
293 char *logname, struct super_block *sb, int 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 struct config_llog_data *nodemap_cld = NULL;
323 struct config_llog_data *barrier_cld = NULL;
330 CDEBUG(D_MGC, "add config log %s-%016lx\n", logname,
334 * for each regular log, the depended sptlrpc log name is
335 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
337 ptr = strrchr(logname, '-');
338 if (ptr == NULL || ptr - logname > 8) {
339 CERROR("logname %s is too long\n", logname);
340 RETURN(ERR_PTR(-EINVAL));
343 memcpy(seclogname, logname, ptr - logname);
344 strcpy(seclogname + (ptr - logname), "-sptlrpc");
346 if (cfg->cfg_sub_clds & CONFIG_SUB_SPTLRPC) {
347 sptlrpc_cld = config_log_find_or_add(obd, seclogname, NULL,
348 CONFIG_T_SPTLRPC, cfg);
349 if (IS_ERR(sptlrpc_cld)) {
350 CERROR("%s: can't create sptlrpc log %s: rc = %ld\n",
351 obd->obd_name, seclogname, PTR_ERR(sptlrpc_cld));
356 if (!IS_MGS(lsi) && cfg->cfg_sub_clds & CONFIG_SUB_NODEMAP) {
357 nodemap_cld = config_log_find_or_add(obd, LUSTRE_NODEMAP_NAME,
358 NULL, CONFIG_T_NODEMAP,
360 if (IS_ERR(nodemap_cld)) {
361 rc = PTR_ERR(nodemap_cld);
362 CERROR("%s: cannot create nodemap log: rc = %d\n",
364 GOTO(out_sptlrpc, rc);
368 if (cfg->cfg_sub_clds & CONFIG_SUB_PARAMS) {
369 params_cld = config_log_find_or_add(obd, PARAMS_FILENAME, sb,
370 CONFIG_T_PARAMS, cfg);
371 if (IS_ERR(params_cld)) {
372 rc = PTR_ERR(params_cld);
373 CERROR("%s: can't create params log: rc = %d\n",
375 GOTO(out_nodemap, rc);
379 if (IS_MDT(s2lsi(sb)) && cfg->cfg_sub_clds & CONFIG_SUB_BARRIER) {
380 snprintf(seclogname + (ptr - logname), sizeof(seclogname) - 1,
381 "-%s", BARRIER_FILENAME);
382 barrier_cld = config_log_find_or_add(obd, seclogname, sb,
383 CONFIG_T_BARRIER, cfg);
384 if (IS_ERR(barrier_cld)) {
385 rc = PTR_ERR(barrier_cld);
386 CERROR("%s: can't create barrier log: rc = %d\n",
388 GOTO(out_params, rc);
392 cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
395 CERROR("%s: can't create log: rc = %d\n",
397 GOTO(out_barrier, rc = PTR_ERR(cld));
400 LASSERT(lsi->lsi_lmd);
401 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR) &&
402 cfg->cfg_sub_clds & CONFIG_SUB_RECOVER) {
403 struct config_llog_data *recover_cld;
405 ptr = strrchr(seclogname, '-');
409 CERROR("%s: sptlrpc log name not correct, %s: "
410 "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
411 GOTO(out_cld, rc = -EINVAL);
414 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
415 if (IS_ERR(recover_cld)) {
416 rc = PTR_ERR(recover_cld);
417 CERROR("%s: can't create recover log: rc = %d\n",
422 mutex_lock(&cld->cld_lock);
424 cld->cld_recover = recover_cld;
428 mutex_lock(&cld->cld_lock);
429 cld->cld_params = params_cld;
430 cld->cld_barrier = barrier_cld;
431 cld->cld_nodemap = nodemap_cld;
432 cld->cld_sptlrpc = sptlrpc_cld;
433 mutex_unlock(&cld->cld_lock);
440 config_log_put(barrier_cld);
442 config_log_put(params_cld);
444 config_log_put(nodemap_cld);
446 config_log_put(sptlrpc_cld);
451 DEFINE_MUTEX(llog_process_lock);
453 static inline void config_mark_cld_stop_nolock(struct config_llog_data *cld)
457 spin_lock(&config_list_lock);
458 cld->cld_stopping = 1;
459 spin_unlock(&config_list_lock);
461 CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
462 if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
463 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
466 static inline void config_mark_cld_stop(struct config_llog_data *cld)
469 mutex_lock(&cld->cld_lock);
470 config_mark_cld_stop_nolock(cld);
471 mutex_unlock(&cld->cld_lock);
475 /** Stop watching for updates on this log.
477 static int config_log_end(char *logname, struct config_llog_instance *cfg)
479 struct config_llog_data *cld;
480 struct config_llog_data *cld_sptlrpc = NULL;
481 struct config_llog_data *cld_params = NULL;
482 struct config_llog_data *cld_recover = NULL;
483 struct config_llog_data *cld_nodemap = NULL;
484 struct config_llog_data *cld_barrier = NULL;
489 cld = config_log_find(logname, cfg);
493 mutex_lock(&cld->cld_lock);
495 * if cld_stopping is set, it means we didn't start the log thus
496 * not owning the start ref. this can happen after previous umount:
497 * the cld still hanging there waiting for lock cancel, and we
498 * remount again but failed in the middle and call log_end without
501 if (unlikely(cld->cld_stopping)) {
502 mutex_unlock(&cld->cld_lock);
503 /* drop the ref from the find */
508 cld_recover = cld->cld_recover;
509 cld->cld_recover = NULL;
510 cld_params = cld->cld_params;
511 cld->cld_params = NULL;
512 cld_nodemap = cld->cld_nodemap;
513 cld->cld_nodemap = NULL;
514 cld_barrier = cld->cld_barrier;
515 cld->cld_barrier = NULL;
516 cld_sptlrpc = cld->cld_sptlrpc;
517 cld->cld_sptlrpc = NULL;
519 config_mark_cld_stop_nolock(cld);
520 mutex_unlock(&cld->cld_lock);
522 config_mark_cld_stop(cld_recover);
523 config_log_put(cld_recover);
524 config_mark_cld_stop(cld_params);
525 config_log_put(cld_params);
526 config_mark_cld_stop(cld_barrier);
527 config_log_put(cld_barrier);
528 /* don't explicitly set cld_stopping on sptlrpc lock here, as other
529 * targets may be active, it will be done in config_log_put if necessary
531 config_log_put(cld_sptlrpc);
532 /* don't set cld_stopping on nm lock as other targets may be active */
533 config_log_put(cld_nodemap);
535 /* drop the ref from the find */
537 /* drop the start ref */
540 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
545 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
547 struct obd_device *obd = data;
548 struct obd_import *imp;
549 struct obd_connect_data *ocd;
550 struct config_llog_data *cld;
555 with_imp_locked(obd, imp, rc) {
556 ocd = &imp->imp_connect_data;
558 seq_printf(m, "imperative_recovery: %s\n",
559 OCD_HAS_FLAG(ocd, IMP_RECOV) ?
560 "ENABLED" : "DISABLED");
565 seq_printf(m, "client_state:\n");
567 spin_lock(&config_list_lock);
568 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
569 if (cld->cld_recover == NULL)
571 seq_printf(m, " - { client: %s, nidtbl_version: %u }\n",
573 cld->cld_recover->cld_cfg.cfg_last_idx);
575 spin_unlock(&config_list_lock);
580 /* reenqueue any lost locks */
581 #define RQ_RUNNING 0x1
585 #define RQ_PRECLEANUP 0x10
586 static int rq_state = 0;
587 static wait_queue_head_t rq_waitq;
588 static DECLARE_COMPLETION(rq_exit);
589 static DECLARE_COMPLETION(rq_start);
591 static void do_requeue(struct config_llog_data *cld)
596 LASSERT(atomic_read(&cld->cld_refcount) > 0);
599 * Do not run mgc_process_log on a disconnected export or an
600 * export which is being disconnected. Take the client
601 * semaphore to make the check non-racy.
603 down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
605 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
606 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
607 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
608 if (rc && rc != -ENOENT)
609 CERROR("failed processing log: %d\n", rc);
611 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
614 up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
619 /* this timeout represents how many seconds MGC should wait before
620 * requeue config and recover lock to the MGS. We need to randomize this
621 * in order to not flood the MGS.
623 #define MGC_TIMEOUT_MIN_SECONDS 5
624 #define MGC_TIMEOUT_RAND_CENTISEC 500
626 static int mgc_requeue_thread(void *data)
632 CDEBUG(D_MGC, "Starting requeue thread\n");
634 /* Keep trying failed locks periodically */
635 spin_lock(&config_list_lock);
636 rq_state |= RQ_RUNNING;
637 while (!(rq_state & RQ_STOP)) {
638 struct config_llog_data *cld, *cld_prev;
639 int rand = prandom_u32_max(MGC_TIMEOUT_RAND_CENTISEC);
642 /* Any new or requeued lostlocks will change the state */
643 rq_state &= ~(RQ_NOW | RQ_LATER);
644 spin_unlock(&config_list_lock);
651 /* Always wait a few seconds to allow the server who
652 * caused the lock revocation to finish its setup, plus some
653 * random so everyone doesn't try to reconnect at once.
655 to = cfs_time_seconds(MGC_TIMEOUT_MIN_SECONDS * 100 + rand);
656 /* rand is centi-seconds */
657 wait_event_idle_timeout(rq_waitq,
658 rq_state & (RQ_STOP | RQ_PRECLEANUP),
662 * iterate & processing through the list. for each cld, process
663 * its depending sptlrpc cld firstly (if any) and then itself.
665 * it's guaranteed any item in the list must have
666 * reference > 0; and if cld_lostlock is set, at
667 * least one reference is taken by the previous enqueue.
671 spin_lock(&config_list_lock);
672 rq_state &= ~RQ_PRECLEANUP;
673 list_for_each_entry(cld, &config_llog_list,
675 if (!cld->cld_lostlock || cld->cld_stopping)
678 /* hold reference to avoid being freed during
679 * subsequent processing. */
681 cld->cld_lostlock = 0;
682 spin_unlock(&config_list_lock);
684 config_log_put(cld_prev);
687 if (likely(!(rq_state & RQ_STOP))) {
689 spin_lock(&config_list_lock);
691 spin_lock(&config_list_lock);
695 spin_unlock(&config_list_lock);
696 config_log_put(cld_prev);
698 /* Wait a bit to see if anyone else needs a requeue */
699 wait_event_idle(rq_waitq, rq_state & (RQ_NOW | RQ_STOP));
700 spin_lock(&config_list_lock);
703 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
704 rq_state &= ~RQ_RUNNING;
705 spin_unlock(&config_list_lock);
709 CDEBUG(D_MGC, "Ending requeue thread\n");
713 /* Add a cld to the list to requeue. Start the requeue thread if needed.
714 We are responsible for dropping the config log reference from here on out. */
715 static void mgc_requeue_add(struct config_llog_data *cld)
720 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
721 cld->cld_logname, atomic_read(&cld->cld_refcount),
722 cld->cld_stopping, rq_state);
723 LASSERT(atomic_read(&cld->cld_refcount) > 0);
725 /* lets cancel an existent lock to mark cld as "lostlock" */
726 CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
727 if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
728 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
730 mutex_lock(&cld->cld_lock);
731 spin_lock(&config_list_lock);
732 if (!(rq_state & RQ_STOP) && !cld->cld_stopping) {
733 cld->cld_lostlock = 1;
737 spin_unlock(&config_list_lock);
738 mutex_unlock(&cld->cld_lock);
745 /********************** class fns **********************/
746 static int mgc_local_llog_init(const struct lu_env *env,
747 struct obd_device *obd,
748 struct obd_device *disk)
750 struct llog_ctxt *ctxt;
755 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
760 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
762 ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
768 static int mgc_local_llog_fini(const struct lu_env *env,
769 struct obd_device *obd)
771 struct llog_ctxt *ctxt;
775 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
776 llog_cleanup(env, ctxt);
781 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
782 struct super_block *sb)
784 struct lustre_sb_info *lsi = s2lsi(sb);
785 struct client_obd *cli = &obd->u.cli;
786 struct lu_fid rfid, fid;
787 struct dt_object *root, *dto;
793 LASSERT(lsi->lsi_dt_dev);
795 /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
796 mutex_lock(&cli->cl_mgc_mutex);
798 /* Setup the configs dir */
799 fid.f_seq = FID_SEQ_LOCAL_NAME;
802 rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
807 rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
811 root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
812 &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
813 if (unlikely(IS_ERR(root)))
814 GOTO(out_los, rc = PTR_ERR(root));
816 dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
818 S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
819 dt_object_put_nocache(env, root);
821 GOTO(out_los, rc = PTR_ERR(dto));
823 cli->cl_mgc_configs_dir = dto;
825 LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
826 rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
830 /* We take an obd ref to insure that we can't get to mgc_cleanup
831 * without calling mgc_fs_cleanup first. */
832 class_incref(obd, "mgc_fs", obd);
834 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
838 dt_object_put(env, cli->cl_mgc_configs_dir);
839 cli->cl_mgc_configs_dir = NULL;
843 local_oid_storage_fini(env, cli->cl_mgc_los);
844 cli->cl_mgc_los = NULL;
845 mutex_unlock(&cli->cl_mgc_mutex);
850 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
852 struct client_obd *cli = &obd->u.cli;
855 LASSERT(cli->cl_mgc_los != NULL);
857 mgc_local_llog_fini(env, obd);
859 dt_object_put_nocache(env, cli->cl_mgc_configs_dir);
860 cli->cl_mgc_configs_dir = NULL;
862 local_oid_storage_fini(env, cli->cl_mgc_los);
863 cli->cl_mgc_los = NULL;
865 class_decref(obd, "mgc_fs", obd);
866 mutex_unlock(&cli->cl_mgc_mutex);
871 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
873 struct llog_ctxt *ctxt;
878 /* setup only remote ctxt, the local disk context is switched per each
879 * filesystem during mgc_fs_setup() */
880 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
885 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
888 llog_initiator_connect(ctxt);
894 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
896 struct llog_ctxt *ctxt;
900 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
902 llog_cleanup(env, ctxt);
908 static atomic_t mgc_count = ATOMIC_INIT(0);
909 static int mgc_precleanup(struct obd_device *obd)
915 if (atomic_dec_and_test(&mgc_count)) {
916 LASSERT(rq_state & RQ_RUNNING);
917 /* stop requeue thread */
920 /* wakeup requeue thread to clean our cld */
921 temp = RQ_NOW | RQ_PRECLEANUP;
924 spin_lock(&config_list_lock);
926 spin_unlock(&config_list_lock);
930 wait_for_completion(&rq_exit);
931 obd_cleanup_client_import(obd);
933 rc = mgc_llog_fini(NULL, obd);
935 CERROR("failed to cleanup llogging subsystems\n");
940 static int mgc_cleanup(struct obd_device *obd)
945 /* COMPAT_146 - old config logs may have added profiles we don't
947 if (atomic_read(&obd->obd_type->typ_refcnt) <= 1)
948 /* Only for the last mgc */
949 class_del_profiles();
951 lprocfs_obd_cleanup(obd);
954 rc = client_obd_cleanup(obd);
958 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
960 struct task_struct *task;
964 rc = ptlrpcd_addref();
968 rc = client_obd_setup(obd, lcfg);
970 GOTO(err_decref, rc);
972 rc = mgc_llog_init(NULL, obd);
974 CERROR("failed to setup llogging subsystems\n");
975 GOTO(err_cleanup, rc);
978 rc = mgc_tunables_init(obd);
982 if (atomic_inc_return(&mgc_count) == 1) {
984 init_waitqueue_head(&rq_waitq);
986 /* start requeue thread */
987 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
990 CERROR("%s: cannot start requeue thread: rc = %d; "
991 "no more log updates\n",
995 /* rc is the task_struct pointer of mgc_requeue_thread. */
997 wait_for_completion(&rq_start);
1003 lprocfs_obd_cleanup(obd);
1005 client_obd_cleanup(obd);
1011 /* based on ll_mdc_blocking_ast */
1012 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1013 void *data, int flag)
1015 struct lustre_handle lockh;
1016 struct config_llog_data *cld = (struct config_llog_data *)data;
1021 case LDLM_CB_BLOCKING:
1022 /* mgs wants the lock, give it up... */
1023 LDLM_DEBUG(lock, "MGC blocking CB");
1024 ldlm_lock2handle(lock, &lockh);
1025 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
1027 case LDLM_CB_CANCELING:
1028 /* We've given up the lock, prepare ourselves to update. */
1029 LDLM_DEBUG(lock, "MGC cancel CB");
1031 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
1032 PLDLMRES(lock->l_resource),
1033 (char *)&lock->l_resource->lr_name.name[0]);
1036 CDEBUG(D_INFO, "missing data, won't requeue\n");
1040 /* held at mgc_process_log(). */
1041 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1043 lock->l_ast_data = NULL;
1044 cld->cld_lockh.cookie = 0;
1045 /* Are we done with this log? */
1046 if (cld->cld_stopping) {
1047 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1049 config_log_put(cld);
1052 /* Make sure not to re-enqueue when the mgc is stopping
1053 (we get called from client_disconnect_export) */
1054 if (lock->l_conn_export == NULL ||
1055 lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1056 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1058 config_log_put(cld);
1062 /* Re-enqueue now */
1063 mgc_requeue_add(cld);
1064 config_log_put(cld);
1073 /* Not sure where this should go... */
1074 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1075 * that we can have a chance to try the secondary MGS if any. */
1076 #define MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1078 #define MGC_TARGET_REG_LIMIT 10
1079 #define MGC_TARGET_REG_LIMIT_MAX RECONNECT_DELAY_MAX
1080 #define MGC_SEND_PARAM_LIMIT 10
1082 /* Take a config lock so we can get cancel notifications */
1083 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1084 union ldlm_policy_data *policy, enum ldlm_mode mode,
1085 __u64 *flags, ldlm_glimpse_callback glimpse_callback,
1086 void *data, __u32 lvb_len, void *lvb_swabber,
1087 struct lustre_handle *lockh)
1089 struct config_llog_data *cld = (struct config_llog_data *)data;
1090 struct ldlm_enqueue_info einfo = {
1093 .ei_cb_bl = mgc_blocking_ast,
1094 .ei_cb_cp = ldlm_completion_ast,
1095 .ei_cb_gl = glimpse_callback,
1097 struct ptlrpc_request *req;
1098 int short_limit = cld_is_sptlrpc(cld);
1105 CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1106 cld->cld_resid.name[0]);
1108 /* We need a callback for every lockholder, so don't try to
1109 ldlm_lock_match (see rev 1.1.2.11.2.47) */
1110 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1111 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1116 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1117 ptlrpc_request_set_replen(req);
1119 /* check if this is server or client */
1120 if (cld->cld_cfg.cfg_sb) {
1121 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1122 if (lsi && IS_SERVER(lsi))
1125 /* Limit how long we will wait for the enqueue to complete */
1126 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1127 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1128 NULL, 0, LVB_T_NONE, lockh, 0);
1129 /* A failed enqueue should still call the mgc_blocking_ast,
1130 where it will be requeued if needed ("grant failed"). */
1131 ptlrpc_req_finished(req);
1135 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1136 struct lustre_handle *lockh)
1140 ldlm_lock_decref(lockh, mode);
1145 static void mgc_notify_active(struct obd_device *unused)
1147 /* wakeup mgc_requeue_thread to requeue mgc lock */
1148 spin_lock(&config_list_lock);
1150 spin_unlock(&config_list_lock);
1153 /* TODO: Help the MGS rebuild nidtbl. -jay */
1156 /* Send target_reg message to MGS */
1157 static int mgc_target_register(struct obd_export *exp,
1158 struct mgs_target_info *mti)
1160 struct ptlrpc_request *req;
1161 struct mgs_target_info *req_mti, *rep_mti;
1165 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1166 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1171 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1173 ptlrpc_req_finished(req);
1177 memcpy(req_mti, mti, sizeof(*req_mti));
1178 ptlrpc_request_set_replen(req);
1179 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1180 /* Limit how long we will wait for the enqueue to complete */
1181 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1183 /* if the target needs to regenerate the config log in MGS, it's better
1184 * to use some longer limit to let MGC have time to change connection to
1185 * another MGS (or try again with the same MGS) for the target (server)
1186 * will fail and exit if the request expired due to delay limit. */
1187 if (mti->mti_flags & (LDD_F_UPDATE | LDD_F_NEED_INDEX))
1188 req->rq_delay_limit = MGC_TARGET_REG_LIMIT_MAX;
1190 rc = ptlrpc_queue_wait(req);
1192 rep_mti = req_capsule_server_get(&req->rq_pill,
1193 &RMF_MGS_TARGET_INFO);
1194 memcpy(mti, rep_mti, sizeof(*rep_mti));
1195 CDEBUG(D_MGC, "register %s got index = %d\n",
1196 mti->mti_svname, mti->mti_stripe_index);
1198 ptlrpc_req_finished(req);
1203 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1204 u32 keylen, void *key,
1205 u32 vallen, void *val,
1206 struct ptlrpc_request_set *set)
1211 /* Turn off initial_recov after we try all backup servers once */
1212 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1213 struct obd_import *imp = class_exp2cliimp(exp);
1215 if (vallen != sizeof(int))
1217 value = *(int *)val;
1218 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1219 imp->imp_obd->obd_name, value,
1220 imp->imp_deactive, imp->imp_invalid,
1221 imp->imp_replayable, imp->imp_obd->obd_replayable,
1222 ptlrpc_import_state_name(imp->imp_state));
1223 /* Resurrect the import immediately if
1224 * 1. we previously got disconnected,
1225 * 2. value > 1 (at the same node with MGS)
1227 if (imp->imp_state == LUSTRE_IMP_DISCON || value > 1)
1228 ptlrpc_reconnect_import(imp);
1233 /* FIXME move this to mgc_process_config */
1234 if (KEY_IS(KEY_REGISTER_TARGET)) {
1235 struct mgs_target_info *mti;
1236 if (vallen != sizeof(struct mgs_target_info))
1238 mti = (struct mgs_target_info *)val;
1239 CDEBUG(D_MGC, "register_target %s %#x\n",
1240 mti->mti_svname, mti->mti_flags);
1241 rc = mgc_target_register(exp, mti);
1244 if (KEY_IS(KEY_SET_FS)) {
1245 struct super_block *sb = (struct super_block *)val;
1247 if (vallen != sizeof(struct super_block))
1250 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1253 if (KEY_IS(KEY_CLEAR_FS)) {
1256 rc = mgc_fs_cleanup(env, exp->exp_obd);
1259 if (KEY_IS(KEY_MGSSEC)) {
1260 struct client_obd *cli = &exp->exp_obd->u.cli;
1261 struct sptlrpc_flavor flvr;
1264 * empty string means using current flavor, if which haven't
1265 * been set yet, set it as null.
1267 * if flavor has been set previously, check the asking flavor
1268 * must match the existing one.
1271 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1277 rc = sptlrpc_parse_flavor(val, &flvr);
1279 CERROR("invalid sptlrpc flavor %s to MGS\n",
1285 * caller already hold a mutex
1287 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1288 cli->cl_flvr_mgc = flvr;
1289 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1290 sizeof(flvr)) != 0) {
1293 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1295 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1296 "currently %s is in use\n",
1306 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1307 __u32 keylen, void *key, __u32 *vallen, void *val)
1311 if (KEY_IS(KEY_CONN_DATA)) {
1312 struct obd_import *imp = class_exp2cliimp(exp);
1313 struct obd_connect_data *data = val;
1315 if (*vallen == sizeof(*data)) {
1316 *data = imp->imp_connect_data;
1324 static int mgc_import_event(struct obd_device *obd,
1325 struct obd_import *imp,
1326 enum obd_import_event event)
1330 LASSERT(imp->imp_obd == obd);
1331 CDEBUG(D_MGC, "import event %#x\n", event);
1334 case IMP_EVENT_DISCON:
1335 /* MGC imports should not wait for recovery */
1336 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1337 ptlrpc_pinger_ir_down();
1339 case IMP_EVENT_INACTIVE:
1341 case IMP_EVENT_INVALIDATE: {
1342 struct ldlm_namespace *ns = obd->obd_namespace;
1343 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1346 case IMP_EVENT_ACTIVE:
1347 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1348 /* Clearing obd_no_recov allows us to continue pinging */
1349 obd->obd_no_recov = 0;
1350 mgc_notify_active(obd);
1351 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1352 ptlrpc_pinger_ir_up();
1356 case IMP_EVENT_DEACTIVATE:
1357 case IMP_EVENT_ACTIVATE:
1360 CERROR("Unknown import event %#x\n", event);
1367 CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1368 CONFIG_READ_NRPAGES = 4
1371 static int mgc_apply_recover_logs(struct obd_device *mgc,
1372 struct config_llog_data *cld,
1374 void *data, int datalen, bool mne_swab)
1376 struct config_llog_instance *cfg = &cld->cld_cfg;
1377 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1378 struct mgs_nidtbl_entry *entry;
1379 struct lustre_cfg *lcfg;
1380 struct lustre_cfg_bufs bufs;
1381 u64 prev_version = 0;
1388 unsigned long dynamic_nids;
1391 LASSERT(cfg->cfg_instance != 0);
1392 LASSERT(ll_get_cfg_instance(cfg->cfg_sb) == cfg->cfg_instance);
1394 /* get dynamic nids setting */
1395 dynamic_nids = mgc->obd_dynamic_nids;
1397 OBD_ALLOC(inst, PAGE_SIZE);
1401 if (!IS_SERVER(lsi)) {
1402 pos = snprintf(inst, PAGE_SIZE, "%016lx", cfg->cfg_instance);
1403 if (pos >= PAGE_SIZE) {
1404 OBD_FREE(inst, PAGE_SIZE);
1407 #ifdef HAVE_SERVER_SUPPORT
1409 LASSERT(IS_MDT(lsi));
1410 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1413 OBD_FREE(inst, PAGE_SIZE);
1417 #endif /* HAVE_SERVER_SUPPORT */
1422 bufsz = PAGE_SIZE - pos;
1424 while (datalen > 0) {
1425 int entry_len = sizeof(*entry);
1427 struct obd_device *obd;
1428 struct obd_import *imp;
1435 if (datalen < sizeof(*entry))
1438 entry = (typeof(entry))(data + off);
1441 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1443 if (entry->mne_nid_count == 0) /* at least one nid entry */
1445 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1448 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1449 if (datalen < entry_len) /* must have entry_len at least */
1452 /* Keep this swab for normal mixed endian handling. LU-1644 */
1454 lustre_swab_mgs_nidtbl_entry(entry);
1455 if (entry->mne_length > PAGE_SIZE) {
1456 CERROR("MNE too large (%u)\n", entry->mne_length);
1460 if (entry->mne_length < entry_len)
1463 off += entry->mne_length;
1464 datalen -= entry->mne_length;
1468 if (entry->mne_version > max_version) {
1469 CERROR("entry index(%lld) is over max_index(%lld)\n",
1470 entry->mne_version, max_version);
1474 if (prev_version >= entry->mne_version) {
1475 CERROR("index unsorted, prev %lld, now %lld\n",
1476 prev_version, entry->mne_version);
1479 prev_version = entry->mne_version;
1482 * Write a string with format "nid::instance" to
1483 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1486 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1487 memset(buf, 0, bufsz);
1491 /* lustre-OST0001-osc-<instance #> */
1492 strcpy(obdname, cld->cld_logname);
1493 cname = strrchr(obdname, '-');
1494 if (cname == NULL) {
1495 CERROR("mgc %s: invalid logname %s\n",
1496 mgc->obd_name, obdname);
1500 pos = cname - obdname;
1502 pos += sprintf(obdname + pos, "-%s%04x",
1503 is_ost ? "OST" : "MDT", entry->mne_index);
1505 cname = is_ost ? "osc" : "mdc",
1506 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1507 lustre_cfg_bufs_reset(&bufs, obdname);
1509 /* find the obd by obdname */
1510 obd = class_name2obd(obdname);
1512 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1513 mgc->obd_name, obdname);
1515 /* this is a safe race, when the ost is starting up...*/
1519 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1522 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1525 with_imp_locked(obd, imp, rc) {
1526 /* iterate all nids to find one */
1527 /* find uuid by nid */
1528 /* create import entries if they don't exist */
1529 rc = client_import_add_nids_to_conn(
1530 imp, entry->u.nids, entry->mne_nid_count,
1531 (struct obd_uuid *)uuid);
1533 if (rc == -ENOENT && dynamic_nids) {
1534 /* create a new connection for this import */
1536 libcfs_nid2str(entry->u.nids[0]);
1537 int prim_nid_len = strlen(primary_nid) + 1;
1538 struct obd_uuid server_uuid;
1540 if (prim_nid_len > UUID_MAX)
1542 strncpy(server_uuid.uuid, primary_nid,
1545 CDEBUG(D_INFO, "Adding a connection for %s\n",
1548 rc = client_import_dyn_add_conn(
1549 imp, &server_uuid, entry->u.nids[0], 1);
1551 CERROR("%s: Failed to add new connection with NID '%s' to import: rc = %d\n",
1552 obd->obd_name, primary_nid, rc);
1555 rc = client_import_add_nids_to_conn(
1557 entry->mne_nid_count,
1558 (struct obd_uuid *)uuid);
1560 CERROR("%s: failed to lookup UUID: rc = %d\n",
1567 if (rc == -ENODEV) {
1568 /* client does not connect to the OST yet */
1573 if (rc < 0 && rc != -ENOSPC) {
1574 CERROR("mgc: cannot find UUID by nid '%s': rc = %d\n",
1575 libcfs_nid2str(entry->u.nids[0]), rc);
1579 CDEBUG(D_INFO, "Found UUID '%s' by NID '%s'\n",
1580 uuid, libcfs_nid2str(entry->u.nids[0]));
1582 pos += strlen(uuid);
1583 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1584 LASSERT(pos < bufsz);
1586 lustre_cfg_bufs_set_string(&bufs, 1, params);
1588 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1594 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1596 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1597 prev_version, max_version, obdname, params);
1599 rc = class_process_config(lcfg);
1600 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1601 lcfg->lcfg_buflens));
1603 CDEBUG(D_INFO, "process config for %s error %d\n",
1606 /* continue, even one with error */
1609 OBD_FREE(inst, PAGE_SIZE);
1614 * This function is called if this client was notified for target restarting
1615 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1618 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1619 struct config_llog_data *cld)
1621 struct ptlrpc_connection *mgc_conn;
1622 struct ptlrpc_request *req = NULL;
1623 struct config_llog_instance *cfg = &cld->cld_cfg;
1624 struct mgs_config_body *body;
1625 struct mgs_config_res *res;
1626 struct nodemap_config *new_config = NULL;
1627 struct lu_nodemap *recent_nodemap = NULL;
1628 struct ptlrpc_bulk_desc *desc;
1629 struct page **pages = NULL;
1630 __u64 config_read_offset = 0;
1631 __u8 nodemap_cur_pass = 0;
1634 bool mne_swab = false;
1640 mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1642 /* don't need to get local config */
1643 if (cld_is_nodemap(cld) && LNetIsPeerLocal(mgc_conn->c_peer.nid))
1646 /* allocate buffer for bulk transfer.
1647 * if this is the first time for this mgs to read logs,
1648 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1649 * once; otherwise, it only reads increment of logs, this should be
1650 * small and CONFIG_READ_NRPAGES will be used.
1652 nrpages = CONFIG_READ_NRPAGES;
1653 if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1654 nrpages = CONFIG_READ_NRPAGES_INIT;
1656 OBD_ALLOC_PTR_ARRAY_LARGE(pages, nrpages);
1658 GOTO(out, rc = -ENOMEM);
1660 for (i = 0; i < nrpages; i++) {
1661 pages[i] = alloc_page(GFP_KERNEL);
1662 if (pages[i] == NULL)
1663 GOTO(out, rc = -ENOMEM);
1667 #ifdef HAVE_SERVER_SUPPORT
1668 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1669 new_config = nodemap_config_alloc();
1670 if (IS_ERR(new_config)) {
1671 rc = PTR_ERR(new_config);
1677 LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1678 LASSERT(mutex_is_locked(&cld->cld_lock));
1679 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1680 &RQF_MGS_CONFIG_READ);
1682 GOTO(out, rc = -ENOMEM);
1684 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1689 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1690 LASSERT(body != NULL);
1691 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1692 if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1693 >= sizeof(body->mcb_name))
1694 GOTO(out, rc = -E2BIG);
1695 if (cld_is_nodemap(cld))
1696 body->mcb_offset = config_read_offset;
1698 body->mcb_offset = cfg->cfg_last_idx + 1;
1699 body->mcb_type = cld->cld_type;
1700 body->mcb_bits = PAGE_SHIFT;
1701 body->mcb_units = nrpages;
1702 body->mcb_nm_cur_pass = nodemap_cur_pass;
1704 /* allocate bulk transfer descriptor */
1705 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1706 PTLRPC_BULK_PUT_SINK,
1708 &ptlrpc_bulk_kiov_pin_ops);
1710 GOTO(out, rc = -ENOMEM);
1712 for (i = 0; i < nrpages; i++)
1713 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1716 ptlrpc_request_set_replen(req);
1717 rc = ptlrpc_queue_wait(req);
1721 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1723 GOTO(out, rc = -EPROTO);
1725 if (cld_is_nodemap(cld)) {
1726 config_read_offset = res->mcr_offset;
1727 eof = config_read_offset == II_END_OFF;
1728 nodemap_cur_pass = res->mcr_nm_cur_pass;
1730 if (res->mcr_size < res->mcr_offset)
1731 GOTO(out, rc = -EINVAL);
1733 /* always update the index even though it might have errors with
1734 * handling the recover logs
1736 cfg->cfg_last_idx = res->mcr_offset;
1737 eof = res->mcr_offset == res->mcr_size;
1739 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1740 res->mcr_offset, eof == false);
1743 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1745 GOTO(out, rc = ealen);
1747 if (ealen > nrpages << PAGE_SHIFT)
1748 GOTO(out, rc = -EINVAL);
1750 if (ealen == 0) { /* no logs transferred */
1751 #ifdef HAVE_SERVER_SUPPORT
1752 /* config changed since first read RPC */
1753 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1754 CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1755 GOTO(out, rc = -EAGAIN);
1763 mne_swab = ptlrpc_rep_need_swab(req);
1765 /* When a nodemap config is received, we build a new nodemap config,
1766 * with new nodemap structs. We keep track of the most recently added
1767 * nodemap since the config is read ordered by nodemap_id, and so it
1768 * is likely that the next record will be related. Because access to
1769 * the nodemaps is single threaded until the nodemap_config is active,
1770 * we don't need to reference count with recent_nodemap, though
1771 * recent_nodemap should be set to NULL when the nodemap_config
1772 * is either destroyed or set active.
1774 for (i = 0; i < nrpages && ealen > 0; i++) {
1778 ptr = kmap(pages[i]);
1779 if (cld_is_nodemap(cld))
1780 rc2 = nodemap_process_idx_pages(new_config, ptr,
1783 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1790 CWARN("%s: error processing %s log %s: rc = %d\n",
1792 cld_is_nodemap(cld) ? "nodemap" : "recovery",
1795 GOTO(out, rc = rc2);
1803 ptlrpc_req_finished(req);
1807 if (rc == 0 && !eof)
1810 #ifdef HAVE_SERVER_SUPPORT
1811 if (new_config != NULL) {
1812 /* recent_nodemap cannot be used after set_active/dealloc */
1814 nodemap_config_set_active_mgc(new_config);
1816 nodemap_config_dealloc(new_config);
1821 for (i = 0; i < nrpages; i++) {
1822 if (pages[i] == NULL)
1824 __free_page(pages[i]);
1826 OBD_FREE_PTR_ARRAY_LARGE(pages, nrpages);
1831 static int mgc_barrier_glimpse_ast(struct ldlm_lock *lock, void *data)
1833 struct config_llog_data *cld = lock->l_ast_data;
1837 if (cld->cld_stopping)
1840 rc = barrier_handler(s2lsi(cld->cld_cfg.cfg_sb)->lsi_dt_dev,
1841 (struct ptlrpc_request *)data);
1846 /* Copy a remote log locally */
1847 static int mgc_llog_local_copy(const struct lu_env *env,
1848 struct obd_device *obd,
1849 struct llog_ctxt *rctxt,
1850 struct llog_ctxt *lctxt, char *logname)
1858 * - copy it to backup using llog_backup()
1859 * - copy remote llog to logname using llog_backup()
1860 * - if failed then move bakup to logname again
1863 OBD_ALLOC(temp_log, strlen(logname) + 2);
1866 sprintf(temp_log, "%sT", logname);
1868 /* make a copy of local llog at first */
1869 rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1870 if (rc < 0 && rc != -ENOENT)
1872 /* copy remote llog to the local copy */
1873 rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1874 if (rc == -ENOENT) {
1875 /* no remote llog, delete local one too */
1876 llog_erase(env, lctxt, NULL, logname);
1877 } else if (rc < 0) {
1878 /* error during backup, get local one back from the copy */
1879 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1881 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1882 obd->obd_name, logname, rc);
1884 llog_erase(env, lctxt, NULL, temp_log);
1885 OBD_FREE(temp_log, strlen(logname) + 2);
1889 /* local_only means it cannot get remote llogs */
1890 static int mgc_process_cfg_log(struct obd_device *mgc,
1891 struct config_llog_data *cld, int local_only)
1893 struct llog_ctxt *ctxt, *lctxt = NULL;
1894 struct client_obd *cli = &mgc->u.cli;
1895 struct lustre_sb_info *lsi = NULL;
1902 LASSERT(mutex_is_locked(&cld->cld_lock));
1904 if (cld->cld_cfg.cfg_sb)
1905 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1911 rc = lu_env_init(env, LCT_MG_THREAD);
1915 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1918 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1920 /* Copy the setup log locally if we can. Don't mess around if we're
1921 * running an MGS though (logs are already local). */
1922 if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1923 cli->cl_mgc_configs_dir != NULL &&
1924 lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1926 if (!local_only && !lsi->lsi_dt_dev->dd_rdonly) {
1927 /* Only try to copy log if we have the lock. */
1928 CDEBUG(D_INFO, "%s: copy local log %s\n",
1929 mgc->obd_name, cld->cld_logname);
1931 rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1934 lsi->lsi_flags &= ~LDD_F_NO_LOCAL_LOGS;
1936 if (local_only || rc) {
1937 if (unlikely(lsi->lsi_flags & LDD_F_NO_LOCAL_LOGS)
1939 CWARN("%s: local log %s are not valid and/or remote logs are not accessbile rc = %d\n",
1940 mgc->obd_name, cld->cld_logname, rc);
1941 GOTO(out_pop, rc = -EIO);
1944 if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1945 llog_is_empty(env, lctxt, cld->cld_logname)) {
1946 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS log %s and no local copy.\n",
1948 GOTO(out_pop, rc = -ENOENT);
1950 CDEBUG(D_MGC, "%s: Failed to get MGS log %s, using local copy for now, will try to update later.\n",
1951 mgc->obd_name, cld->cld_logname);
1954 /* Now, whether we copied or not, start using the local llog.
1955 * If we failed to copy, we'll start using whatever the old
1957 llog_ctxt_put(ctxt);
1961 if (local_only) /* no local log at client side */
1962 GOTO(out_pop, rc = -EIO);
1966 if (lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1967 lsi->lsi_dt_dev->dd_rdonly) {
1968 struct llog_ctxt *rctxt;
1970 /* Under readonly mode, we may have no local copy or local
1971 * copy is incomplete, so try to use remote llog firstly. */
1972 rctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1975 rc = class_config_parse_llog(env, rctxt, cld->cld_logname,
1977 llog_ctxt_put(rctxt);
1980 if (rc && rc != -ENOENT)
1981 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1985 * update settings on existing OBDs.
1986 * the logname must be <fsname>-sptlrpc
1988 if (rc == 0 && cld_is_sptlrpc(cld))
1989 class_notify_sptlrpc_conf(cld->cld_logname,
1990 strlen(cld->cld_logname) -
1991 strlen("-sptlrpc"));
1995 __llog_ctxt_put(env, ctxt);
1997 __llog_ctxt_put(env, lctxt);
2005 static bool mgc_import_in_recovery(struct obd_import *imp)
2007 bool in_recovery = true;
2009 spin_lock(&imp->imp_lock);
2010 if (imp->imp_state == LUSTRE_IMP_FULL ||
2011 imp->imp_state == LUSTRE_IMP_CLOSED)
2012 in_recovery = false;
2013 spin_unlock(&imp->imp_lock);
2019 * Get a configuration log from the MGS and process it.
2021 * This function is called for both clients and servers to process the
2022 * configuration log from the MGS. The MGC enqueues a DLM lock on the
2023 * log from the MGS, and if the lock gets revoked the MGC will be notified
2024 * by the lock cancellation callback that the config log has changed,
2025 * and will enqueue another MGS lock on it, and then continue processing
2026 * the new additions to the end of the log.
2028 * Since the MGC import is not replayable, if the import is being evicted
2029 * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
2030 * the log until recovery is finished or the import is closed.
2032 * Make a local copy of the log before parsing it if appropriate (non-MGS
2033 * server) so that the server can start even when the MGS is down.
2035 * There shouldn't be multiple processes running process_log at once --
2036 * sounds like badness. It actually might be fine, as long as they're not
2037 * trying to update from the same log simultaneously, in which case we
2038 * should use a per-log semaphore instead of cld_lock.
2040 * \param[in] mgc MGC device by which to fetch the configuration log
2041 * \param[in] cld log processing state (stored in lock callback data)
2043 * \retval 0 on success
2044 * \retval negative errno on failure
2046 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
2048 struct lustre_handle lockh = { 0 };
2049 __u64 flags = LDLM_FL_NO_LRU;
2054 LASSERT(cld != NULL);
2056 /* I don't want multiple processes running process_log at once --
2057 sounds like badness. It actually might be fine, as long as
2058 we're not trying to update from the same log
2059 simultaneously (in which case we should use a per-log sem.) */
2061 mutex_lock(&cld->cld_lock);
2062 if (cld->cld_stopping) {
2063 mutex_unlock(&cld->cld_lock);
2067 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2069 CDEBUG(D_MGC, "Process log %s-%016lx from %d\n", cld->cld_logname,
2070 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2072 /* Get the cfg lock on the llog */
2073 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2075 cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
2076 cld, 0, NULL, &lockh);
2078 /* Get the cld, it will be released in mgc_blocking_ast. */
2079 config_log_get(cld);
2080 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2081 LASSERT(!lustre_handle_is_used(&cld->cld_lockh));
2083 cld->cld_lockh = lockh;
2085 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2086 cld->cld_lockh.cookie = 0;
2088 if (rcl == -ESHUTDOWN &&
2089 atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2090 struct obd_import *imp;
2091 long timeout = cfs_time_seconds(obd_timeout);
2093 mutex_unlock(&cld->cld_lock);
2094 imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2096 /* Let's force the pinger, and wait the import to be
2097 * connected, note: since mgc import is non-replayable,
2098 * and even the import state is disconnected, it does
2099 * not mean the "recovery" is stopped, so we will keep
2100 * waitting until timeout or the import state is
2102 ptlrpc_pinger_force(imp);
2104 wait_event_idle_timeout(imp->imp_recovery_waitq,
2105 !mgc_import_in_recovery(imp),
2108 if (imp->imp_state == LUSTRE_IMP_FULL) {
2112 mutex_lock(&cld->cld_lock);
2113 /* unlock/lock mutex, so check stopping again */
2114 if (cld->cld_stopping) {
2115 mutex_unlock(&cld->cld_lock);
2118 spin_lock(&config_list_lock);
2119 cld->cld_lostlock = 1;
2120 spin_unlock(&config_list_lock);
2123 /* mark cld_lostlock so that it will requeue
2124 * after MGC becomes available. */
2125 spin_lock(&config_list_lock);
2126 cld->cld_lostlock = 1;
2127 spin_unlock(&config_list_lock);
2131 if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2133 rc = mgc_process_recover_nodemap_log(mgc, cld);
2134 else if (cld_is_nodemap(cld))
2137 } else if (!cld_is_barrier(cld)) {
2138 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2141 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2142 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2144 /* Now drop the lock so MGS can revoke it */
2146 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2148 CERROR("Can't drop cfg lock: %d\n", rcl);
2150 mutex_unlock(&cld->cld_lock);
2152 /* requeue nodemap lock immediately if transfer was interrupted */
2153 if ((cld_is_nodemap(cld) && rc == -EAGAIN) ||
2154 (cld_is_recover(cld) && rc)) {
2155 if (cld_is_recover(cld))
2156 CWARN("%s: IR log %s failed, not fatal: rc = %d\n",
2157 mgc->obd_name, cld->cld_logname, rc);
2158 mgc_requeue_add(cld);
2166 /** Called from lustre_process_log.
2167 * LCFG_LOG_START gets the config log from the MGS, processes it to start
2168 * any services, and adds it to the list logs to watch (follow).
2170 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2172 struct lustre_cfg *lcfg = buf;
2173 struct config_llog_instance *cfg = NULL;
2178 switch(lcfg->lcfg_command) {
2179 case LCFG_LOV_ADD_OBD: {
2180 /* Overloading this cfg command: register a new target */
2181 struct mgs_target_info *mti;
2183 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2184 sizeof(struct mgs_target_info))
2185 GOTO(out, rc = -EINVAL);
2187 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2188 CDEBUG(D_MGC, "add_target %s %#x\n",
2189 mti->mti_svname, mti->mti_flags);
2190 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2193 case LCFG_LOV_DEL_OBD:
2194 /* Unregister has no meaning at the moment. */
2195 CERROR("lov_del_obd unimplemented\n");
2198 case LCFG_SPTLRPC_CONF: {
2199 rc = sptlrpc_process_config(lcfg);
2202 case LCFG_LOG_START: {
2203 struct config_llog_data *cld;
2204 struct super_block *sb;
2206 logname = lustre_cfg_string(lcfg, 1);
2207 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2208 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2210 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2213 /* We're only called through here on the initial mount */
2214 cld = config_log_add(obd, logname, cfg, sb);
2220 rc = mgc_process_log(obd, cld);
2221 if (rc == 0 && cld->cld_recover != NULL) {
2222 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2223 imp_connect_data, IMP_RECOV)) {
2224 rc = mgc_process_log(obd, cld->cld_recover);
2226 struct config_llog_data *cir;
2228 mutex_lock(&cld->cld_lock);
2229 cir = cld->cld_recover;
2230 cld->cld_recover = NULL;
2231 mutex_unlock(&cld->cld_lock);
2232 config_log_put(cir);
2236 CERROR("Cannot process recover llog %d\n", rc);
2239 if (rc == 0 && cld->cld_params != NULL) {
2240 rc = mgc_process_log(obd, cld->cld_params);
2241 if (rc == -ENOENT) {
2242 CDEBUG(D_MGC, "There is no params "
2243 "config file yet\n");
2246 /* params log is optional */
2248 CERROR("%s: can't process params llog: rc = %d\n",
2254 case LCFG_LOG_END: {
2255 logname = lustre_cfg_string(lcfg, 1);
2257 if (lcfg->lcfg_bufcount >= 2)
2258 cfg = (struct config_llog_instance *)lustre_cfg_buf(
2260 rc = config_log_end(logname, cfg);
2264 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2265 GOTO(out, rc = -EINVAL);
2273 static const struct obd_ops mgc_obd_ops = {
2274 .o_owner = THIS_MODULE,
2275 .o_setup = mgc_setup,
2276 .o_precleanup = mgc_precleanup,
2277 .o_cleanup = mgc_cleanup,
2278 .o_add_conn = client_import_add_conn,
2279 .o_del_conn = client_import_del_conn,
2280 .o_connect = client_connect_import,
2281 .o_disconnect = client_disconnect_export,
2282 .o_set_info_async = mgc_set_info_async,
2283 .o_get_info = mgc_get_info,
2284 .o_import_event = mgc_import_event,
2285 .o_process_config = mgc_process_config,
2288 static int __init mgc_init(void)
2290 return class_register_type(&mgc_obd_ops, NULL, false,
2291 LUSTRE_MGC_NAME, NULL);
2294 static void __exit mgc_exit(void)
2296 class_unregister_type(LUSTRE_MGC_NAME);
2299 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2300 MODULE_DESCRIPTION("Lustre Management Client");
2301 MODULE_VERSION(LUSTRE_VERSION_STRING);
2302 MODULE_LICENSE("GPL");
2304 module_init(mgc_init);
2305 module_exit(mgc_exit);