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/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mgc/mgc_request.c
34 * Author: Nathan Rutman <nathan@clusterfs.com>
37 #define DEBUG_SUBSYSTEM S_MGC
38 #define D_MGC D_CONFIG /*|D_WARNING*/
40 #include <linux/module.h>
41 #include <linux/kthread.h>
42 #include <linux/random.h>
44 #include <dt_object.h>
45 #include <lprocfs_status.h>
46 #include <lustre_dlm.h>
47 #include <lustre_disk.h>
48 #include <lustre_log.h>
49 #include <lustre_nodemap.h>
50 #include <lustre_swab.h>
51 #include <obd_class.h>
52 #include <lustre_barrier.h>
54 #include "mgc_internal.h"
56 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
61 if (len > sizeof(resname)) {
62 CERROR("name too long: %s\n", name);
66 CERROR("missing name: %s\n", name);
69 memcpy(&resname, name, len);
71 /* Always use the same endianness for the resid */
72 memset(res_id, 0, sizeof(*res_id));
73 res_id->name[0] = cpu_to_le64(resname);
74 /* XXX: unfortunately, sptlprc and config llog share one lock */
77 case CONFIG_T_SPTLRPC:
80 case CONFIG_T_RECOVER:
82 case CONFIG_T_NODEMAP:
83 case CONFIG_T_BARRIER:
89 res_id->name[1] = cpu_to_le64(resname);
90 CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
91 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
95 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
97 /* fsname is at most 8 chars long, maybe contain "-".
98 * e.g. "lustre", "SUN-000" */
99 return mgc_name2resid(fsname, strlen(fsname), res_id, type);
101 EXPORT_SYMBOL(mgc_fsname2resid);
103 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
108 /* logname consists of "fsname-nodetype".
109 * e.g. "lustre-MDT0001", "SUN-000-client"
110 * there is an exception: llog "params" */
111 name_end = strrchr(logname, '-');
113 len = strlen(logname);
115 len = name_end - logname;
116 return mgc_name2resid(logname, len, res_id, type);
118 EXPORT_SYMBOL(mgc_logname2resid);
120 /********************** config llog list **********************/
121 static LIST_HEAD(config_llog_list);
122 static DEFINE_SPINLOCK(config_list_lock); /* protects config_llog_list */
124 /* Take a reference to a config log */
125 static int config_log_get(struct config_llog_data *cld)
128 atomic_inc(&cld->cld_refcount);
129 CDEBUG(D_INFO, "log %s (%p) refs %d\n", cld->cld_logname, cld,
130 atomic_read(&cld->cld_refcount));
134 /* Drop a reference to a config log. When no longer referenced,
135 we can free the config log data */
136 static void config_log_put(struct config_llog_data *cld)
143 CDEBUG(D_INFO, "log %s(%p) refs %d\n", cld->cld_logname, cld,
144 atomic_read(&cld->cld_refcount));
145 LASSERT(atomic_read(&cld->cld_refcount) > 0);
147 /* spinlock to make sure no item with 0 refcount in the list */
148 if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
149 list_del(&cld->cld_list_chain);
150 spin_unlock(&config_list_lock);
152 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
154 config_log_put(cld->cld_barrier);
155 config_log_put(cld->cld_recover);
156 config_log_put(cld->cld_params);
157 config_log_put(cld->cld_nodemap);
158 config_log_put(cld->cld_sptlrpc);
159 if (cld_is_sptlrpc(cld))
160 sptlrpc_conf_log_stop(cld->cld_logname);
162 class_export_put(cld->cld_mgcexp);
163 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
169 /* Find a config log by name */
171 struct config_llog_data *config_log_find(char *logname,
172 struct config_llog_instance *cfg)
174 struct config_llog_data *cld;
175 struct config_llog_data *found = NULL;
176 unsigned long cfg_instance;
179 LASSERT(logname != NULL);
181 cfg_instance = cfg ? cfg->cfg_instance : 0;
182 spin_lock(&config_list_lock);
183 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
184 /* check if cfg_instance is the one we want */
185 if (cfg_instance != cld->cld_cfg.cfg_instance)
188 /* instance may be NULL, should check name */
189 if (strcmp(logname, cld->cld_logname) == 0) {
191 config_log_get(found);
195 spin_unlock(&config_list_lock);
200 struct config_llog_data *do_config_log_add(struct obd_device *obd,
203 struct config_llog_instance *cfg,
204 struct super_block *sb)
206 struct config_llog_data *cld;
211 CDEBUG(D_MGC, "do adding config log %s-%016lx\n", logname,
212 cfg ? cfg->cfg_instance : 0);
214 OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
216 RETURN(ERR_PTR(-ENOMEM));
218 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
220 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
224 strcpy(cld->cld_logname, logname);
228 cld->cld_cfg.cfg_callback = class_config_llog_handler;
229 mutex_init(&cld->cld_lock);
230 cld->cld_cfg.cfg_last_idx = 0;
231 cld->cld_cfg.cfg_flags = 0;
232 cld->cld_cfg.cfg_sb = sb;
233 cld->cld_type = type;
234 atomic_set(&cld->cld_refcount, 1);
236 /* Keep the mgc around until we are done */
237 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
239 if (cld_is_sptlrpc(cld))
240 sptlrpc_conf_log_start(logname);
242 spin_lock(&config_list_lock);
243 list_add(&cld->cld_list_chain, &config_llog_list);
244 spin_unlock(&config_list_lock);
246 if (cld_is_sptlrpc(cld) || cld_is_nodemap(cld) || cld_is_barrier(cld)) {
247 rc = mgc_process_log(obd, cld);
248 if (rc && rc != -ENOENT)
249 CERROR("%s: failed processing log, type %d: rc = %d\n",
250 obd->obd_name, type, rc);
256 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
258 struct config_llog_instance *cfg,
259 struct super_block *sb)
261 struct config_llog_instance lcfg = *cfg;
262 struct lustre_sb_info *lsi = s2lsi(sb);
263 struct config_llog_data *cld;
269 /* for osp-on-ost, see lustre_start_osp() */
270 if (IS_MDT(lsi) && lcfg.cfg_instance)
273 /* We have to use different llog for clients and MDTs for DNE,
274 * where only clients are notified if one of DNE server restarts.
276 LASSERT(strlen(fsname) < sizeof(logname) / 2);
277 strncpy(logname, fsname, sizeof(logname));
278 if (IS_SERVER(lsi)) { /* mdt */
279 LASSERT(lcfg.cfg_instance == 0);
280 lcfg.cfg_instance = ll_get_cfg_instance(sb);
281 strncat(logname, "-mdtir", sizeof(logname));
283 LASSERT(lcfg.cfg_instance != 0);
284 strncat(logname, "-cliir", sizeof(logname));
287 cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
291 static struct config_llog_data *config_log_find_or_add(struct obd_device *obd,
292 char *logname, struct super_block *sb, int type,
293 struct config_llog_instance *cfg)
295 struct config_llog_instance lcfg = *cfg;
296 struct config_llog_data *cld;
298 /* Note class_config_llog_handler() depends on getting "obd" back */
299 lcfg.cfg_instance = sb ? ll_get_cfg_instance(sb) : (unsigned long)obd;
301 cld = config_log_find(logname, &lcfg);
302 if (unlikely(cld != NULL))
305 return do_config_log_add(obd, logname, type, &lcfg, sb);
308 /** Add this log to the list of active logs watched by an MGC.
309 * Active means we're watching for updates.
310 * We have one active log per "mount" - client instance or servername.
311 * Each instance may be at a different point in the log.
313 static struct config_llog_data *
314 config_log_add(struct obd_device *obd, char *logname,
315 struct config_llog_instance *cfg, struct super_block *sb)
317 struct lustre_sb_info *lsi = s2lsi(sb);
318 struct config_llog_data *cld = NULL;
319 struct config_llog_data *sptlrpc_cld = NULL;
320 struct config_llog_data *params_cld = NULL;
321 struct config_llog_data *nodemap_cld = NULL;
322 struct config_llog_data *barrier_cld = NULL;
329 CDEBUG(D_MGC, "add config log %s-%016lx\n", logname,
333 * for each regular log, the depended sptlrpc log name is
334 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
336 ptr = strrchr(logname, '-');
337 if (ptr == NULL || ptr - logname > 8) {
338 CERROR("logname %s is too long\n", logname);
339 RETURN(ERR_PTR(-EINVAL));
342 memcpy(seclogname, logname, ptr - logname);
343 strcpy(seclogname + (ptr - logname), "-sptlrpc");
345 if (cfg->cfg_sub_clds & CONFIG_SUB_SPTLRPC) {
346 sptlrpc_cld = config_log_find_or_add(obd, seclogname, NULL,
347 CONFIG_T_SPTLRPC, cfg);
348 if (IS_ERR(sptlrpc_cld)) {
349 CERROR("%s: can't create sptlrpc log %s: rc = %ld\n",
350 obd->obd_name, seclogname, PTR_ERR(sptlrpc_cld));
355 if (!IS_MGS(lsi) && cfg->cfg_sub_clds & CONFIG_SUB_NODEMAP) {
356 nodemap_cld = config_log_find_or_add(obd, LUSTRE_NODEMAP_NAME,
357 NULL, CONFIG_T_NODEMAP,
359 if (IS_ERR(nodemap_cld)) {
360 rc = PTR_ERR(nodemap_cld);
361 CERROR("%s: cannot create nodemap log: rc = %d\n",
363 GOTO(out_sptlrpc, rc);
367 if (cfg->cfg_sub_clds & CONFIG_SUB_PARAMS) {
368 params_cld = config_log_find_or_add(obd, PARAMS_FILENAME, sb,
369 CONFIG_T_PARAMS, cfg);
370 if (IS_ERR(params_cld)) {
371 rc = PTR_ERR(params_cld);
372 CERROR("%s: can't create params log: rc = %d\n",
374 GOTO(out_nodemap, rc);
378 if (IS_MDT(s2lsi(sb)) && cfg->cfg_sub_clds & CONFIG_SUB_BARRIER) {
379 snprintf(seclogname + (ptr - logname), sizeof(seclogname) - 1,
380 "-%s", BARRIER_FILENAME);
381 barrier_cld = config_log_find_or_add(obd, seclogname, sb,
382 CONFIG_T_BARRIER, cfg);
383 if (IS_ERR(barrier_cld)) {
384 rc = PTR_ERR(barrier_cld);
385 CERROR("%s: can't create barrier log: rc = %d\n",
387 GOTO(out_params, rc);
391 cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
394 CERROR("%s: can't create log: rc = %d\n",
396 GOTO(out_barrier, rc = PTR_ERR(cld));
399 LASSERT(lsi->lsi_lmd);
400 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR) &&
401 cfg->cfg_sub_clds & CONFIG_SUB_RECOVER) {
402 struct config_llog_data *recover_cld;
404 ptr = strrchr(seclogname, '-');
408 CERROR("%s: sptlrpc log name not correct, %s: "
409 "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
410 GOTO(out_cld, rc = -EINVAL);
413 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
414 if (IS_ERR(recover_cld)) {
415 rc = PTR_ERR(recover_cld);
416 CERROR("%s: can't create recover log: rc = %d\n",
421 mutex_lock(&cld->cld_lock);
423 cld->cld_recover = recover_cld;
427 mutex_lock(&cld->cld_lock);
428 cld->cld_params = params_cld;
429 cld->cld_barrier = barrier_cld;
430 cld->cld_nodemap = nodemap_cld;
431 cld->cld_sptlrpc = sptlrpc_cld;
432 mutex_unlock(&cld->cld_lock);
439 config_log_put(barrier_cld);
441 config_log_put(params_cld);
443 config_log_put(nodemap_cld);
445 config_log_put(sptlrpc_cld);
450 DEFINE_MUTEX(llog_process_lock);
452 static inline void config_mark_cld_stop_nolock(struct config_llog_data *cld)
456 spin_lock(&config_list_lock);
457 cld->cld_stopping = 1;
458 spin_unlock(&config_list_lock);
460 CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
461 if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
462 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
465 static inline void config_mark_cld_stop(struct config_llog_data *cld)
468 mutex_lock(&cld->cld_lock);
469 config_mark_cld_stop_nolock(cld);
470 mutex_unlock(&cld->cld_lock);
474 /** Stop watching for updates on this log.
476 static int config_log_end(char *logname, struct config_llog_instance *cfg)
478 struct config_llog_data *cld;
479 struct config_llog_data *cld_sptlrpc = NULL;
480 struct config_llog_data *cld_params = NULL;
481 struct config_llog_data *cld_recover = NULL;
482 struct config_llog_data *cld_nodemap = NULL;
483 struct config_llog_data *cld_barrier = NULL;
488 cld = config_log_find(logname, cfg);
492 mutex_lock(&cld->cld_lock);
494 * if cld_stopping is set, it means we didn't start the log thus
495 * not owning the start ref. this can happen after previous umount:
496 * the cld still hanging there waiting for lock cancel, and we
497 * remount again but failed in the middle and call log_end without
500 if (unlikely(cld->cld_stopping)) {
501 mutex_unlock(&cld->cld_lock);
502 /* drop the ref from the find */
507 cld_recover = cld->cld_recover;
508 cld->cld_recover = NULL;
509 cld_params = cld->cld_params;
510 cld->cld_params = NULL;
511 cld_nodemap = cld->cld_nodemap;
512 cld->cld_nodemap = NULL;
513 cld_barrier = cld->cld_barrier;
514 cld->cld_barrier = NULL;
515 cld_sptlrpc = cld->cld_sptlrpc;
516 cld->cld_sptlrpc = NULL;
518 config_mark_cld_stop_nolock(cld);
519 mutex_unlock(&cld->cld_lock);
521 config_mark_cld_stop(cld_recover);
522 config_mark_cld_stop(cld_params);
523 config_mark_cld_stop(cld_barrier);
524 config_mark_cld_stop(cld_sptlrpc);
526 config_log_put(cld_params);
527 config_log_put(cld_recover);
528 /* don't set cld_stopping on nm lock as other targets may be active */
529 config_log_put(cld_nodemap);
530 config_log_put(cld_barrier);
531 config_log_put(cld_sptlrpc);
533 /* drop the ref from the find */
535 /* drop the start ref */
538 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
543 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
545 struct obd_device *obd = data;
546 struct obd_import *imp;
547 struct obd_connect_data *ocd;
548 struct config_llog_data *cld;
552 LPROCFS_CLIMP_CHECK(obd);
553 imp = obd->u.cli.cl_import;
554 ocd = &imp->imp_connect_data;
556 seq_printf(m, "imperative_recovery: %s\n",
557 OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
558 seq_printf(m, "client_state:\n");
560 spin_lock(&config_list_lock);
561 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
562 if (cld->cld_recover == NULL)
564 seq_printf(m, " - { client: %s, nidtbl_version: %u }\n",
566 cld->cld_recover->cld_cfg.cfg_last_idx);
568 spin_unlock(&config_list_lock);
570 LPROCFS_CLIMP_EXIT(obd);
574 /* reenqueue any lost locks */
575 #define RQ_RUNNING 0x1
579 #define RQ_PRECLEANUP 0x10
580 static int rq_state = 0;
581 static wait_queue_head_t rq_waitq;
582 static DECLARE_COMPLETION(rq_exit);
583 static DECLARE_COMPLETION(rq_start);
585 static void do_requeue(struct config_llog_data *cld)
590 LASSERT(atomic_read(&cld->cld_refcount) > 0);
593 * Do not run mgc_process_log on a disconnected export or an
594 * export which is being disconnected. Take the client
595 * semaphore to make the check non-racy.
597 down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
599 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
600 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
601 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
602 if (rc && rc != -ENOENT)
603 CERROR("failed processing log: %d\n", rc);
605 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
608 up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
613 /* this timeout represents how many seconds MGC should wait before
614 * requeue config and recover lock to the MGS. We need to randomize this
615 * in order to not flood the MGS.
617 #define MGC_TIMEOUT_MIN_SECONDS 5
618 #define MGC_TIMEOUT_RAND_CENTISEC 500
620 static int mgc_requeue_thread(void *data)
626 CDEBUG(D_MGC, "Starting requeue thread\n");
628 /* Keep trying failed locks periodically */
629 spin_lock(&config_list_lock);
630 rq_state |= RQ_RUNNING;
631 while (!(rq_state & RQ_STOP)) {
632 struct config_llog_data *cld, *cld_prev;
633 int rand = prandom_u32_max(MGC_TIMEOUT_RAND_CENTISEC);
636 /* Any new or requeued lostlocks will change the state */
637 rq_state &= ~(RQ_NOW | RQ_LATER);
638 spin_unlock(&config_list_lock);
645 /* Always wait a few seconds to allow the server who
646 * caused the lock revocation to finish its setup, plus some
647 * random so everyone doesn't try to reconnect at once.
649 to = cfs_time_seconds(MGC_TIMEOUT_MIN_SECONDS * 100 + rand);
650 /* rand is centi-seconds */
651 wait_event_idle_timeout(rq_waitq,
652 rq_state & (RQ_STOP | RQ_PRECLEANUP),
656 * iterate & processing through the list. for each cld, process
657 * its depending sptlrpc cld firstly (if any) and then itself.
659 * it's guaranteed any item in the list must have
660 * reference > 0; and if cld_lostlock is set, at
661 * least one reference is taken by the previous enqueue.
665 spin_lock(&config_list_lock);
666 rq_state &= ~RQ_PRECLEANUP;
667 list_for_each_entry(cld, &config_llog_list,
669 if (!cld->cld_lostlock || cld->cld_stopping)
672 /* hold reference to avoid being freed during
673 * subsequent processing. */
675 cld->cld_lostlock = 0;
676 spin_unlock(&config_list_lock);
678 config_log_put(cld_prev);
681 if (likely(!(rq_state & RQ_STOP))) {
683 spin_lock(&config_list_lock);
685 spin_lock(&config_list_lock);
689 spin_unlock(&config_list_lock);
690 config_log_put(cld_prev);
692 /* Wait a bit to see if anyone else needs a requeue */
693 wait_event_idle(rq_waitq, rq_state & (RQ_NOW | RQ_STOP));
694 spin_lock(&config_list_lock);
697 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
698 rq_state &= ~RQ_RUNNING;
699 spin_unlock(&config_list_lock);
703 CDEBUG(D_MGC, "Ending requeue thread\n");
707 /* Add a cld to the list to requeue. Start the requeue thread if needed.
708 We are responsible for dropping the config log reference from here on out. */
709 static void mgc_requeue_add(struct config_llog_data *cld)
714 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
715 cld->cld_logname, atomic_read(&cld->cld_refcount),
716 cld->cld_stopping, rq_state);
717 LASSERT(atomic_read(&cld->cld_refcount) > 0);
719 /* lets cancel an existent lock to mark cld as "lostlock" */
720 CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
721 if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
722 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
724 mutex_lock(&cld->cld_lock);
725 spin_lock(&config_list_lock);
726 if (!(rq_state & RQ_STOP) && !cld->cld_stopping) {
727 cld->cld_lostlock = 1;
731 spin_unlock(&config_list_lock);
732 mutex_unlock(&cld->cld_lock);
739 /********************** class fns **********************/
740 static int mgc_local_llog_init(const struct lu_env *env,
741 struct obd_device *obd,
742 struct obd_device *disk)
744 struct llog_ctxt *ctxt;
749 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
754 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
756 ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
762 static int mgc_local_llog_fini(const struct lu_env *env,
763 struct obd_device *obd)
765 struct llog_ctxt *ctxt;
769 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
770 llog_cleanup(env, ctxt);
775 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
776 struct super_block *sb)
778 struct lustre_sb_info *lsi = s2lsi(sb);
779 struct client_obd *cli = &obd->u.cli;
780 struct lu_fid rfid, fid;
781 struct dt_object *root, *dto;
787 LASSERT(lsi->lsi_dt_dev);
789 /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
790 mutex_lock(&cli->cl_mgc_mutex);
792 /* Setup the configs dir */
793 fid.f_seq = FID_SEQ_LOCAL_NAME;
796 rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
801 rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
805 root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
806 &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
807 if (unlikely(IS_ERR(root)))
808 GOTO(out_los, rc = PTR_ERR(root));
810 dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
812 S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
813 dt_object_put_nocache(env, root);
815 GOTO(out_los, rc = PTR_ERR(dto));
817 cli->cl_mgc_configs_dir = dto;
819 LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
820 rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
824 /* We take an obd ref to insure that we can't get to mgc_cleanup
825 * without calling mgc_fs_cleanup first. */
826 class_incref(obd, "mgc_fs", obd);
828 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
832 dt_object_put(env, cli->cl_mgc_configs_dir);
833 cli->cl_mgc_configs_dir = NULL;
837 local_oid_storage_fini(env, cli->cl_mgc_los);
838 cli->cl_mgc_los = NULL;
839 mutex_unlock(&cli->cl_mgc_mutex);
844 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
846 struct client_obd *cli = &obd->u.cli;
849 LASSERT(cli->cl_mgc_los != NULL);
851 mgc_local_llog_fini(env, obd);
853 dt_object_put_nocache(env, cli->cl_mgc_configs_dir);
854 cli->cl_mgc_configs_dir = NULL;
856 local_oid_storage_fini(env, cli->cl_mgc_los);
857 cli->cl_mgc_los = NULL;
859 class_decref(obd, "mgc_fs", obd);
860 mutex_unlock(&cli->cl_mgc_mutex);
865 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
867 struct llog_ctxt *ctxt;
872 /* setup only remote ctxt, the local disk context is switched per each
873 * filesystem during mgc_fs_setup() */
874 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
879 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
882 llog_initiator_connect(ctxt);
888 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
890 struct llog_ctxt *ctxt;
894 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
896 llog_cleanup(env, ctxt);
902 static atomic_t mgc_count = ATOMIC_INIT(0);
903 static int mgc_precleanup(struct obd_device *obd)
909 if (atomic_dec_and_test(&mgc_count)) {
910 LASSERT(rq_state & RQ_RUNNING);
911 /* stop requeue thread */
914 /* wakeup requeue thread to clean our cld */
915 temp = RQ_NOW | RQ_PRECLEANUP;
918 spin_lock(&config_list_lock);
920 spin_unlock(&config_list_lock);
924 wait_for_completion(&rq_exit);
925 obd_cleanup_client_import(obd);
927 rc = mgc_llog_fini(NULL, obd);
929 CERROR("failed to cleanup llogging subsystems\n");
934 static int mgc_cleanup(struct obd_device *obd)
939 /* COMPAT_146 - old config logs may have added profiles we don't
941 if (atomic_read(&obd->obd_type->typ_refcnt) <= 1)
942 /* Only for the last mgc */
943 class_del_profiles();
945 lprocfs_obd_cleanup(obd);
948 rc = client_obd_cleanup(obd);
952 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
954 struct task_struct *task;
958 rc = ptlrpcd_addref();
962 rc = client_obd_setup(obd, lcfg);
964 GOTO(err_decref, rc);
966 rc = mgc_llog_init(NULL, obd);
968 CERROR("failed to setup llogging subsystems\n");
969 GOTO(err_cleanup, rc);
972 rc = mgc_tunables_init(obd);
976 if (atomic_inc_return(&mgc_count) == 1) {
978 init_waitqueue_head(&rq_waitq);
980 /* start requeue thread */
981 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
984 CERROR("%s: cannot start requeue thread: rc = %d; "
985 "no more log updates\n",
989 /* rc is the task_struct pointer of mgc_requeue_thread. */
991 wait_for_completion(&rq_start);
997 lprocfs_obd_cleanup(obd);
999 client_obd_cleanup(obd);
1005 /* based on ll_mdc_blocking_ast */
1006 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1007 void *data, int flag)
1009 struct lustre_handle lockh;
1010 struct config_llog_data *cld = (struct config_llog_data *)data;
1015 case LDLM_CB_BLOCKING:
1016 /* mgs wants the lock, give it up... */
1017 LDLM_DEBUG(lock, "MGC blocking CB");
1018 ldlm_lock2handle(lock, &lockh);
1019 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
1021 case LDLM_CB_CANCELING:
1022 /* We've given up the lock, prepare ourselves to update. */
1023 LDLM_DEBUG(lock, "MGC cancel CB");
1025 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
1026 PLDLMRES(lock->l_resource),
1027 (char *)&lock->l_resource->lr_name.name[0]);
1030 CDEBUG(D_INFO, "missing data, won't requeue\n");
1034 /* held at mgc_process_log(). */
1035 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1037 lock->l_ast_data = NULL;
1038 cld->cld_lockh.cookie = 0;
1039 /* Are we done with this log? */
1040 if (cld->cld_stopping) {
1041 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1043 config_log_put(cld);
1046 /* Make sure not to re-enqueue when the mgc is stopping
1047 (we get called from client_disconnect_export) */
1048 if (lock->l_conn_export == NULL ||
1049 lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1050 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1052 config_log_put(cld);
1056 /* Re-enqueue now */
1057 mgc_requeue_add(cld);
1058 config_log_put(cld);
1067 /* Not sure where this should go... */
1068 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1069 * that we can have a chance to try the secondary MGS if any. */
1070 #define MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1072 #define MGC_TARGET_REG_LIMIT 10
1073 #define MGC_TARGET_REG_LIMIT_MAX RECONNECT_DELAY_MAX
1074 #define MGC_SEND_PARAM_LIMIT 10
1076 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1077 /* Send parameter to MGS*/
1078 static int mgc_set_mgs_param(struct obd_export *exp,
1079 struct mgs_send_param *msp)
1081 struct ptlrpc_request *req;
1082 struct mgs_send_param *req_msp, *rep_msp;
1086 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1087 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1092 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1094 ptlrpc_req_finished(req);
1098 memcpy(req_msp, msp, sizeof(*req_msp));
1099 ptlrpc_request_set_replen(req);
1101 /* Limit how long we will wait for the enqueue to complete */
1102 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1103 rc = ptlrpc_queue_wait(req);
1105 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1106 memcpy(msp, rep_msp, sizeof(*rep_msp));
1109 ptlrpc_req_finished(req);
1115 /* Take a config lock so we can get cancel notifications */
1116 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1117 union ldlm_policy_data *policy, enum ldlm_mode mode,
1118 __u64 *flags, ldlm_glimpse_callback glimpse_callback,
1119 void *data, __u32 lvb_len, void *lvb_swabber,
1120 struct lustre_handle *lockh)
1122 struct config_llog_data *cld = (struct config_llog_data *)data;
1123 struct ldlm_enqueue_info einfo = {
1126 .ei_cb_bl = mgc_blocking_ast,
1127 .ei_cb_cp = ldlm_completion_ast,
1128 .ei_cb_gl = glimpse_callback,
1130 struct ptlrpc_request *req;
1131 int short_limit = cld_is_sptlrpc(cld);
1135 CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1136 cld->cld_resid.name[0]);
1138 /* We need a callback for every lockholder, so don't try to
1139 ldlm_lock_match (see rev 1.1.2.11.2.47) */
1140 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1141 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1146 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1147 ptlrpc_request_set_replen(req);
1149 /* check if this is server or client */
1150 if (cld->cld_cfg.cfg_sb) {
1151 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1152 if (lsi && IS_SERVER(lsi))
1155 /* Limit how long we will wait for the enqueue to complete */
1156 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1157 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1158 NULL, 0, LVB_T_NONE, lockh, 0);
1159 /* A failed enqueue should still call the mgc_blocking_ast,
1160 where it will be requeued if needed ("grant failed"). */
1161 ptlrpc_req_finished(req);
1165 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1166 struct lustre_handle *lockh)
1170 ldlm_lock_decref(lockh, mode);
1175 static void mgc_notify_active(struct obd_device *unused)
1177 /* wakeup mgc_requeue_thread to requeue mgc lock */
1178 spin_lock(&config_list_lock);
1180 spin_unlock(&config_list_lock);
1183 /* TODO: Help the MGS rebuild nidtbl. -jay */
1186 /* Send target_reg message to MGS */
1187 static int mgc_target_register(struct obd_export *exp,
1188 struct mgs_target_info *mti)
1190 struct ptlrpc_request *req;
1191 struct mgs_target_info *req_mti, *rep_mti;
1195 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1196 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1201 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1203 ptlrpc_req_finished(req);
1207 memcpy(req_mti, mti, sizeof(*req_mti));
1208 ptlrpc_request_set_replen(req);
1209 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1210 /* Limit how long we will wait for the enqueue to complete */
1211 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1213 /* if the target needs to regenerate the config log in MGS, it's better
1214 * to use some longer limit to let MGC have time to change connection to
1215 * another MGS (or try again with the same MGS) for the target (server)
1216 * will fail and exit if the request expired due to delay limit. */
1217 if (mti->mti_flags & (LDD_F_UPDATE | LDD_F_NEED_INDEX))
1218 req->rq_delay_limit = MGC_TARGET_REG_LIMIT_MAX;
1220 rc = ptlrpc_queue_wait(req);
1222 rep_mti = req_capsule_server_get(&req->rq_pill,
1223 &RMF_MGS_TARGET_INFO);
1224 memcpy(mti, rep_mti, sizeof(*rep_mti));
1225 CDEBUG(D_MGC, "register %s got index = %d\n",
1226 mti->mti_svname, mti->mti_stripe_index);
1228 ptlrpc_req_finished(req);
1233 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1234 u32 keylen, void *key,
1235 u32 vallen, void *val,
1236 struct ptlrpc_request_set *set)
1241 /* Turn off initial_recov after we try all backup servers once */
1242 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1243 struct obd_import *imp = class_exp2cliimp(exp);
1245 if (vallen != sizeof(int))
1247 value = *(int *)val;
1248 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1249 imp->imp_obd->obd_name, value,
1250 imp->imp_deactive, imp->imp_invalid,
1251 imp->imp_replayable, imp->imp_obd->obd_replayable,
1252 ptlrpc_import_state_name(imp->imp_state));
1253 /* Resurrect the import immediately if
1254 * 1. we previously got disconnected,
1255 * 2. value > 1 (at the same node with MGS)
1257 if (imp->imp_state == LUSTRE_IMP_DISCON || value > 1)
1258 ptlrpc_reconnect_import(imp);
1263 /* FIXME move this to mgc_process_config */
1264 if (KEY_IS(KEY_REGISTER_TARGET)) {
1265 struct mgs_target_info *mti;
1266 if (vallen != sizeof(struct mgs_target_info))
1268 mti = (struct mgs_target_info *)val;
1269 CDEBUG(D_MGC, "register_target %s %#x\n",
1270 mti->mti_svname, mti->mti_flags);
1271 rc = mgc_target_register(exp, mti);
1274 if (KEY_IS(KEY_SET_FS)) {
1275 struct super_block *sb = (struct super_block *)val;
1277 if (vallen != sizeof(struct super_block))
1280 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1283 if (KEY_IS(KEY_CLEAR_FS)) {
1286 rc = mgc_fs_cleanup(env, exp->exp_obd);
1289 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1290 if (KEY_IS(KEY_SET_INFO)) {
1291 struct mgs_send_param *msp;
1293 msp = (struct mgs_send_param *)val;
1294 rc = mgc_set_mgs_param(exp, msp);
1298 if (KEY_IS(KEY_MGSSEC)) {
1299 struct client_obd *cli = &exp->exp_obd->u.cli;
1300 struct sptlrpc_flavor flvr;
1303 * empty string means using current flavor, if which haven't
1304 * been set yet, set it as null.
1306 * if flavor has been set previously, check the asking flavor
1307 * must match the existing one.
1310 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1316 rc = sptlrpc_parse_flavor(val, &flvr);
1318 CERROR("invalid sptlrpc flavor %s to MGS\n",
1324 * caller already hold a mutex
1326 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1327 cli->cl_flvr_mgc = flvr;
1328 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1329 sizeof(flvr)) != 0) {
1332 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1334 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1335 "currently %s is in use\n",
1345 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1346 __u32 keylen, void *key, __u32 *vallen, void *val)
1350 if (KEY_IS(KEY_CONN_DATA)) {
1351 struct obd_import *imp = class_exp2cliimp(exp);
1352 struct obd_connect_data *data = val;
1354 if (*vallen == sizeof(*data)) {
1355 *data = imp->imp_connect_data;
1363 static int mgc_import_event(struct obd_device *obd,
1364 struct obd_import *imp,
1365 enum obd_import_event event)
1369 LASSERT(imp->imp_obd == obd);
1370 CDEBUG(D_MGC, "import event %#x\n", event);
1373 case IMP_EVENT_DISCON:
1374 /* MGC imports should not wait for recovery */
1375 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1376 ptlrpc_pinger_ir_down();
1378 case IMP_EVENT_INACTIVE:
1380 case IMP_EVENT_INVALIDATE: {
1381 struct ldlm_namespace *ns = obd->obd_namespace;
1382 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1385 case IMP_EVENT_ACTIVE:
1386 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1387 /* Clearing obd_no_recov allows us to continue pinging */
1388 obd->obd_no_recov = 0;
1389 mgc_notify_active(obd);
1390 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1391 ptlrpc_pinger_ir_up();
1395 case IMP_EVENT_DEACTIVATE:
1396 case IMP_EVENT_ACTIVATE:
1399 CERROR("Unknown import event %#x\n", event);
1406 CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1407 CONFIG_READ_NRPAGES = 4
1410 static int mgc_apply_recover_logs(struct obd_device *mgc,
1411 struct config_llog_data *cld,
1413 void *data, int datalen, bool mne_swab)
1415 struct config_llog_instance *cfg = &cld->cld_cfg;
1416 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1417 struct mgs_nidtbl_entry *entry;
1418 struct lustre_cfg *lcfg;
1419 struct lustre_cfg_bufs bufs;
1420 u64 prev_version = 0;
1429 LASSERT(cfg->cfg_instance != 0);
1430 LASSERT(ll_get_cfg_instance(cfg->cfg_sb) == cfg->cfg_instance);
1432 OBD_ALLOC(inst, PAGE_SIZE);
1436 if (!IS_SERVER(lsi)) {
1437 pos = snprintf(inst, PAGE_SIZE, "%016lx", cfg->cfg_instance);
1438 if (pos >= PAGE_SIZE) {
1439 OBD_FREE(inst, PAGE_SIZE);
1442 #ifdef HAVE_SERVER_SUPPORT
1444 LASSERT(IS_MDT(lsi));
1445 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1448 OBD_FREE(inst, PAGE_SIZE);
1452 #endif /* HAVE_SERVER_SUPPORT */
1457 bufsz = PAGE_SIZE - pos;
1459 while (datalen > 0) {
1460 int entry_len = sizeof(*entry);
1462 struct obd_device *obd;
1469 if (datalen < sizeof(*entry))
1472 entry = (typeof(entry))(data + off);
1475 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1477 if (entry->mne_nid_count == 0) /* at least one nid entry */
1479 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1482 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1483 if (datalen < entry_len) /* must have entry_len at least */
1486 /* Keep this swab for normal mixed endian handling. LU-1644 */
1488 lustre_swab_mgs_nidtbl_entry(entry);
1489 if (entry->mne_length > PAGE_SIZE) {
1490 CERROR("MNE too large (%u)\n", entry->mne_length);
1494 if (entry->mne_length < entry_len)
1497 off += entry->mne_length;
1498 datalen -= entry->mne_length;
1502 if (entry->mne_version > max_version) {
1503 CERROR("entry index(%lld) is over max_index(%lld)\n",
1504 entry->mne_version, max_version);
1508 if (prev_version >= entry->mne_version) {
1509 CERROR("index unsorted, prev %lld, now %lld\n",
1510 prev_version, entry->mne_version);
1513 prev_version = entry->mne_version;
1516 * Write a string with format "nid::instance" to
1517 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1520 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1521 memset(buf, 0, bufsz);
1525 /* lustre-OST0001-osc-<instance #> */
1526 strcpy(obdname, cld->cld_logname);
1527 cname = strrchr(obdname, '-');
1528 if (cname == NULL) {
1529 CERROR("mgc %s: invalid logname %s\n",
1530 mgc->obd_name, obdname);
1534 pos = cname - obdname;
1536 pos += sprintf(obdname + pos, "-%s%04x",
1537 is_ost ? "OST" : "MDT", entry->mne_index);
1539 cname = is_ost ? "osc" : "mdc",
1540 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1541 lustre_cfg_bufs_reset(&bufs, obdname);
1543 /* find the obd by obdname */
1544 obd = class_name2obd(obdname);
1546 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1547 mgc->obd_name, obdname);
1549 /* this is a safe race, when the ost is starting up...*/
1553 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1556 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1559 down_read(&obd->u.cli.cl_sem);
1560 if (obd->u.cli.cl_import == NULL) {
1561 /* client does not connect to the OST yet */
1562 up_read(&obd->u.cli.cl_sem);
1567 /* iterate all nids to find one */
1568 /* find uuid by nid */
1570 for (i = 0; i < entry->mne_nid_count; i++) {
1571 rc = client_import_find_conn(obd->u.cli.cl_import,
1573 (struct obd_uuid *)uuid);
1578 up_read(&obd->u.cli.cl_sem);
1580 CERROR("mgc: cannot find uuid by nid %s\n",
1581 libcfs_nid2str(entry->u.nids[0]));
1585 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1586 uuid, libcfs_nid2str(entry->u.nids[0]));
1588 pos += strlen(uuid);
1589 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1590 LASSERT(pos < bufsz);
1592 lustre_cfg_bufs_set_string(&bufs, 1, params);
1594 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1600 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1602 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1603 prev_version, max_version, obdname, params);
1605 rc = class_process_config(lcfg);
1606 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1607 lcfg->lcfg_buflens));
1609 CDEBUG(D_INFO, "process config for %s error %d\n",
1612 /* continue, even one with error */
1615 OBD_FREE(inst, PAGE_SIZE);
1620 * This function is called if this client was notified for target restarting
1621 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1624 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1625 struct config_llog_data *cld)
1627 struct ptlrpc_connection *mgc_conn;
1628 struct ptlrpc_request *req = NULL;
1629 struct config_llog_instance *cfg = &cld->cld_cfg;
1630 struct mgs_config_body *body;
1631 struct mgs_config_res *res;
1632 struct nodemap_config *new_config = NULL;
1633 struct lu_nodemap *recent_nodemap = NULL;
1634 struct ptlrpc_bulk_desc *desc;
1635 struct page **pages = NULL;
1636 __u64 config_read_offset = 0;
1637 __u8 nodemap_cur_pass = 0;
1640 bool mne_swab = false;
1646 mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1648 /* don't need to get local config */
1649 if (cld_is_nodemap(cld) &&
1650 (LNET_NETTYP(LNET_NIDNET(mgc_conn->c_peer.nid)) == LOLND))
1653 /* allocate buffer for bulk transfer.
1654 * if this is the first time for this mgs to read logs,
1655 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1656 * once; otherwise, it only reads increment of logs, this should be
1657 * small and CONFIG_READ_NRPAGES will be used.
1659 nrpages = CONFIG_READ_NRPAGES;
1660 if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1661 nrpages = CONFIG_READ_NRPAGES_INIT;
1663 OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1665 GOTO(out, rc = -ENOMEM);
1667 for (i = 0; i < nrpages; i++) {
1668 pages[i] = alloc_page(GFP_KERNEL);
1669 if (pages[i] == NULL)
1670 GOTO(out, rc = -ENOMEM);
1674 #ifdef HAVE_SERVER_SUPPORT
1675 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1676 new_config = nodemap_config_alloc();
1677 if (IS_ERR(new_config)) {
1678 rc = PTR_ERR(new_config);
1684 LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1685 LASSERT(mutex_is_locked(&cld->cld_lock));
1686 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1687 &RQF_MGS_CONFIG_READ);
1689 GOTO(out, rc = -ENOMEM);
1691 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1696 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1697 LASSERT(body != NULL);
1698 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1699 if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1700 >= sizeof(body->mcb_name))
1701 GOTO(out, rc = -E2BIG);
1702 if (cld_is_nodemap(cld))
1703 body->mcb_offset = config_read_offset;
1705 body->mcb_offset = cfg->cfg_last_idx + 1;
1706 body->mcb_type = cld->cld_type;
1707 body->mcb_bits = PAGE_SHIFT;
1708 body->mcb_units = nrpages;
1709 body->mcb_nm_cur_pass = nodemap_cur_pass;
1711 /* allocate bulk transfer descriptor */
1712 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1713 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1715 &ptlrpc_bulk_kiov_pin_ops);
1717 GOTO(out, rc = -ENOMEM);
1719 for (i = 0; i < nrpages; i++)
1720 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1723 ptlrpc_request_set_replen(req);
1724 rc = ptlrpc_queue_wait(req);
1728 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1730 GOTO(out, rc = -EPROTO);
1732 if (cld_is_nodemap(cld)) {
1733 config_read_offset = res->mcr_offset;
1734 eof = config_read_offset == II_END_OFF;
1735 nodemap_cur_pass = res->mcr_nm_cur_pass;
1737 if (res->mcr_size < res->mcr_offset)
1738 GOTO(out, rc = -EINVAL);
1740 /* always update the index even though it might have errors with
1741 * handling the recover logs
1743 cfg->cfg_last_idx = res->mcr_offset;
1744 eof = res->mcr_offset == res->mcr_size;
1746 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1747 res->mcr_offset, eof == false);
1750 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1752 GOTO(out, rc = ealen);
1754 if (ealen > nrpages << PAGE_SHIFT)
1755 GOTO(out, rc = -EINVAL);
1757 if (ealen == 0) { /* no logs transferred */
1758 #ifdef HAVE_SERVER_SUPPORT
1759 /* config changed since first read RPC */
1760 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1761 CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1762 GOTO(out, rc = -EAGAIN);
1770 mne_swab = ptlrpc_rep_need_swab(req);
1772 /* When a nodemap config is received, we build a new nodemap config,
1773 * with new nodemap structs. We keep track of the most recently added
1774 * nodemap since the config is read ordered by nodemap_id, and so it
1775 * is likely that the next record will be related. Because access to
1776 * the nodemaps is single threaded until the nodemap_config is active,
1777 * we don't need to reference count with recent_nodemap, though
1778 * recent_nodemap should be set to NULL when the nodemap_config
1779 * is either destroyed or set active.
1781 for (i = 0; i < nrpages && ealen > 0; i++) {
1785 ptr = kmap(pages[i]);
1786 if (cld_is_nodemap(cld))
1787 rc2 = nodemap_process_idx_pages(new_config, ptr,
1790 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1797 CWARN("%s: error processing %s log %s: rc = %d\n",
1799 cld_is_nodemap(cld) ? "nodemap" : "recovery",
1802 GOTO(out, rc = rc2);
1810 ptlrpc_req_finished(req);
1814 if (rc == 0 && !eof)
1817 #ifdef HAVE_SERVER_SUPPORT
1818 if (new_config != NULL) {
1819 /* recent_nodemap cannot be used after set_active/dealloc */
1821 nodemap_config_set_active_mgc(new_config);
1823 nodemap_config_dealloc(new_config);
1828 for (i = 0; i < nrpages; i++) {
1829 if (pages[i] == NULL)
1831 __free_page(pages[i]);
1833 OBD_FREE(pages, sizeof(*pages) * nrpages);
1838 static int mgc_barrier_glimpse_ast(struct ldlm_lock *lock, void *data)
1840 struct config_llog_data *cld = lock->l_ast_data;
1844 if (cld->cld_stopping)
1847 rc = barrier_handler(s2lsi(cld->cld_cfg.cfg_sb)->lsi_dt_dev,
1848 (struct ptlrpc_request *)data);
1853 /* Copy a remote log locally */
1854 static int mgc_llog_local_copy(const struct lu_env *env,
1855 struct obd_device *obd,
1856 struct llog_ctxt *rctxt,
1857 struct llog_ctxt *lctxt, char *logname)
1865 * - copy it to backup using llog_backup()
1866 * - copy remote llog to logname using llog_backup()
1867 * - if failed then move bakup to logname again
1870 OBD_ALLOC(temp_log, strlen(logname) + 2);
1873 sprintf(temp_log, "%sT", logname);
1875 /* make a copy of local llog at first */
1876 rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1877 if (rc < 0 && rc != -ENOENT)
1879 /* copy remote llog to the local copy */
1880 rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1881 if (rc == -ENOENT) {
1882 /* no remote llog, delete local one too */
1883 llog_erase(env, lctxt, NULL, logname);
1884 } else if (rc < 0) {
1885 /* error during backup, get local one back from the copy */
1886 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1888 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1889 obd->obd_name, logname, rc);
1891 llog_erase(env, lctxt, NULL, temp_log);
1892 OBD_FREE(temp_log, strlen(logname) + 2);
1896 /* local_only means it cannot get remote llogs */
1897 static int mgc_process_cfg_log(struct obd_device *mgc,
1898 struct config_llog_data *cld, int local_only)
1900 struct llog_ctxt *ctxt, *lctxt = NULL;
1901 struct client_obd *cli = &mgc->u.cli;
1902 struct lustre_sb_info *lsi = NULL;
1909 LASSERT(mutex_is_locked(&cld->cld_lock));
1911 if (cld->cld_cfg.cfg_sb)
1912 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1918 rc = lu_env_init(env, LCT_MG_THREAD);
1922 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1925 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1927 /* Copy the setup log locally if we can. Don't mess around if we're
1928 * running an MGS though (logs are already local). */
1929 if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1930 cli->cl_mgc_configs_dir != NULL &&
1931 lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1933 if (!local_only && !lsi->lsi_dt_dev->dd_rdonly)
1934 /* Only try to copy log if we have the lock. */
1935 rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1937 if (local_only || rc) {
1938 if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1939 llog_is_empty(env, lctxt, cld->cld_logname)) {
1940 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1941 "log %s and no local copy."
1942 "\n", cld->cld_logname);
1943 GOTO(out_pop, rc = -ENOENT);
1945 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1946 "copy for now, will try to update later.\n",
1950 /* Now, whether we copied or not, start using the local llog.
1951 * If we failed to copy, we'll start using whatever the old
1953 llog_ctxt_put(ctxt);
1957 if (local_only) /* no local log at client side */
1958 GOTO(out_pop, rc = -EIO);
1962 if (lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1963 lsi->lsi_dt_dev->dd_rdonly) {
1964 struct llog_ctxt *rctxt;
1966 /* Under readonly mode, we may have no local copy or local
1967 * copy is incomplete, so try to use remote llog firstly. */
1968 rctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1971 rc = class_config_parse_llog(env, rctxt, cld->cld_logname,
1973 llog_ctxt_put(rctxt);
1976 if (rc && rc != -ENOENT)
1977 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1981 * update settings on existing OBDs.
1982 * the logname must be <fsname>-sptlrpc
1984 if (rc == 0 && cld_is_sptlrpc(cld))
1985 class_notify_sptlrpc_conf(cld->cld_logname,
1986 strlen(cld->cld_logname) -
1987 strlen("-sptlrpc"));
1991 __llog_ctxt_put(env, ctxt);
1993 __llog_ctxt_put(env, lctxt);
2001 static bool mgc_import_in_recovery(struct obd_import *imp)
2003 bool in_recovery = true;
2005 spin_lock(&imp->imp_lock);
2006 if (imp->imp_state == LUSTRE_IMP_FULL ||
2007 imp->imp_state == LUSTRE_IMP_CLOSED)
2008 in_recovery = false;
2009 spin_unlock(&imp->imp_lock);
2015 * Get a configuration log from the MGS and process it.
2017 * This function is called for both clients and servers to process the
2018 * configuration log from the MGS. The MGC enqueues a DLM lock on the
2019 * log from the MGS, and if the lock gets revoked the MGC will be notified
2020 * by the lock cancellation callback that the config log has changed,
2021 * and will enqueue another MGS lock on it, and then continue processing
2022 * the new additions to the end of the log.
2024 * Since the MGC import is not replayable, if the import is being evicted
2025 * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
2026 * the log until recovery is finished or the import is closed.
2028 * Make a local copy of the log before parsing it if appropriate (non-MGS
2029 * server) so that the server can start even when the MGS is down.
2031 * There shouldn't be multiple processes running process_log at once --
2032 * sounds like badness. It actually might be fine, as long as they're not
2033 * trying to update from the same log simultaneously, in which case we
2034 * should use a per-log semaphore instead of cld_lock.
2036 * \param[in] mgc MGC device by which to fetch the configuration log
2037 * \param[in] cld log processing state (stored in lock callback data)
2039 * \retval 0 on success
2040 * \retval negative errno on failure
2042 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
2044 struct lustre_handle lockh = { 0 };
2045 __u64 flags = LDLM_FL_NO_LRU;
2050 LASSERT(cld != NULL);
2052 /* I don't want multiple processes running process_log at once --
2053 sounds like badness. It actually might be fine, as long as
2054 we're not trying to update from the same log
2055 simultaneously (in which case we should use a per-log sem.) */
2057 mutex_lock(&cld->cld_lock);
2058 if (cld->cld_stopping) {
2059 mutex_unlock(&cld->cld_lock);
2063 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2065 CDEBUG(D_MGC, "Process log %s-%016lx from %d\n", cld->cld_logname,
2066 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2068 /* Get the cfg lock on the llog */
2069 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2071 cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
2072 cld, 0, NULL, &lockh);
2074 /* Get the cld, it will be released in mgc_blocking_ast. */
2075 config_log_get(cld);
2076 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2077 LASSERT(!lustre_handle_is_used(&cld->cld_lockh));
2079 cld->cld_lockh = lockh;
2081 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2082 cld->cld_lockh.cookie = 0;
2084 if (rcl == -ESHUTDOWN &&
2085 atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2086 struct obd_import *imp;
2087 long timeout = cfs_time_seconds(obd_timeout);
2089 mutex_unlock(&cld->cld_lock);
2090 imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2092 /* Let's force the pinger, and wait the import to be
2093 * connected, note: since mgc import is non-replayable,
2094 * and even the import state is disconnected, it does
2095 * not mean the "recovery" is stopped, so we will keep
2096 * waitting until timeout or the import state is
2098 ptlrpc_pinger_force(imp);
2100 wait_event_idle_timeout(imp->imp_recovery_waitq,
2101 !mgc_import_in_recovery(imp),
2104 if (imp->imp_state == LUSTRE_IMP_FULL) {
2108 mutex_lock(&cld->cld_lock);
2109 /* unlock/lock mutex, so check stopping again */
2110 if (cld->cld_stopping) {
2111 mutex_unlock(&cld->cld_lock);
2114 spin_lock(&config_list_lock);
2115 cld->cld_lostlock = 1;
2116 spin_unlock(&config_list_lock);
2119 /* mark cld_lostlock so that it will requeue
2120 * after MGC becomes available. */
2121 spin_lock(&config_list_lock);
2122 cld->cld_lostlock = 1;
2123 spin_unlock(&config_list_lock);
2127 if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2129 rc = mgc_process_recover_nodemap_log(mgc, cld);
2130 else if (cld_is_nodemap(cld))
2133 } else if (!cld_is_barrier(cld)) {
2134 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2137 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2138 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2140 /* Now drop the lock so MGS can revoke it */
2142 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2144 CERROR("Can't drop cfg lock: %d\n", rcl);
2146 mutex_unlock(&cld->cld_lock);
2148 /* requeue nodemap lock immediately if transfer was interrupted */
2149 if ((cld_is_nodemap(cld) && rc == -EAGAIN) ||
2150 (cld_is_recover(cld) && rc)) {
2151 if (cld_is_recover(cld))
2152 CWARN("%s: IR log %s failed, not fatal: rc = %d\n",
2153 mgc->obd_name, cld->cld_logname, rc);
2154 mgc_requeue_add(cld);
2162 /** Called from lustre_process_log.
2163 * LCFG_LOG_START gets the config log from the MGS, processes it to start
2164 * any services, and adds it to the list logs to watch (follow).
2166 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2168 struct lustre_cfg *lcfg = buf;
2169 struct config_llog_instance *cfg = NULL;
2174 switch(lcfg->lcfg_command) {
2175 case LCFG_LOV_ADD_OBD: {
2176 /* Overloading this cfg command: register a new target */
2177 struct mgs_target_info *mti;
2179 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2180 sizeof(struct mgs_target_info))
2181 GOTO(out, rc = -EINVAL);
2183 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2184 CDEBUG(D_MGC, "add_target %s %#x\n",
2185 mti->mti_svname, mti->mti_flags);
2186 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2189 case LCFG_LOV_DEL_OBD:
2190 /* Unregister has no meaning at the moment. */
2191 CERROR("lov_del_obd unimplemented\n");
2194 case LCFG_SPTLRPC_CONF: {
2195 rc = sptlrpc_process_config(lcfg);
2198 case LCFG_LOG_START: {
2199 struct config_llog_data *cld;
2200 struct super_block *sb;
2202 logname = lustre_cfg_string(lcfg, 1);
2203 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2204 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2206 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2209 /* We're only called through here on the initial mount */
2210 cld = config_log_add(obd, logname, cfg, sb);
2216 rc = mgc_process_log(obd, cld);
2217 if (rc == 0 && cld->cld_recover != NULL) {
2218 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2219 imp_connect_data, IMP_RECOV)) {
2220 rc = mgc_process_log(obd, cld->cld_recover);
2222 struct config_llog_data *cir;
2224 mutex_lock(&cld->cld_lock);
2225 cir = cld->cld_recover;
2226 cld->cld_recover = NULL;
2227 mutex_unlock(&cld->cld_lock);
2228 config_log_put(cir);
2232 CERROR("Cannot process recover llog %d\n", rc);
2235 if (rc == 0 && cld->cld_params != NULL) {
2236 rc = mgc_process_log(obd, cld->cld_params);
2237 if (rc == -ENOENT) {
2238 CDEBUG(D_MGC, "There is no params "
2239 "config file yet\n");
2242 /* params log is optional */
2244 CERROR("%s: can't process params llog: rc = %d\n",
2250 case LCFG_LOG_END: {
2251 logname = lustre_cfg_string(lcfg, 1);
2253 if (lcfg->lcfg_bufcount >= 2)
2254 cfg = (struct config_llog_instance *)lustre_cfg_buf(
2256 rc = config_log_end(logname, cfg);
2260 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2261 GOTO(out, rc = -EINVAL);
2269 static const struct obd_ops mgc_obd_ops = {
2270 .o_owner = THIS_MODULE,
2271 .o_setup = mgc_setup,
2272 .o_precleanup = mgc_precleanup,
2273 .o_cleanup = mgc_cleanup,
2274 .o_add_conn = client_import_add_conn,
2275 .o_del_conn = client_import_del_conn,
2276 .o_connect = client_connect_import,
2277 .o_disconnect = client_disconnect_export,
2278 .o_set_info_async = mgc_set_info_async,
2279 .o_get_info = mgc_get_info,
2280 .o_import_event = mgc_import_event,
2281 .o_process_config = mgc_process_config,
2284 static int __init mgc_init(void)
2286 return class_register_type(&mgc_obd_ops, NULL, false, NULL,
2287 LUSTRE_MGC_NAME, NULL);
2290 static void __exit mgc_exit(void)
2292 class_unregister_type(LUSTRE_MGC_NAME);
2295 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2296 MODULE_DESCRIPTION("Lustre Management Client");
2297 MODULE_VERSION(LUSTRE_VERSION_STRING);
2298 MODULE_LICENSE("GPL");
2300 module_init(mgc_init);
2301 module_exit(mgc_exit);