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>
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 struct list_head config_llog_list = LIST_HEAD_INIT(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 refs %d\n", cld->cld_logname,
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 refs %d\n", cld->cld_logname,
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 sptlrpc_conf_log_stop(cld->cld_logname);
161 class_export_put(cld->cld_mgcexp);
162 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
168 /* Find a config log by name */
170 struct config_llog_data *config_log_find(char *logname,
171 struct config_llog_instance *cfg)
173 struct config_llog_data *cld;
174 struct config_llog_data *found = NULL;
178 LASSERT(logname != NULL);
180 instance = cfg ? cfg->cfg_instance : NULL;
181 spin_lock(&config_list_lock);
182 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
183 /* check if instance equals */
184 if (instance != cld->cld_cfg.cfg_instance)
187 /* instance may be NULL, should check name */
188 if (strcmp(logname, cld->cld_logname) == 0) {
190 config_log_get(found);
194 spin_unlock(&config_list_lock);
199 struct config_llog_data *do_config_log_add(struct obd_device *obd,
202 struct config_llog_instance *cfg,
203 struct super_block *sb)
205 struct config_llog_data *cld;
210 CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
211 cfg ? cfg->cfg_instance : NULL);
213 OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
215 RETURN(ERR_PTR(-ENOMEM));
217 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
219 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
223 strcpy(cld->cld_logname, logname);
227 cld->cld_cfg.cfg_callback = class_config_llog_handler;
228 mutex_init(&cld->cld_lock);
229 cld->cld_cfg.cfg_last_idx = 0;
230 cld->cld_cfg.cfg_flags = 0;
231 cld->cld_cfg.cfg_sb = sb;
232 cld->cld_type = type;
233 atomic_set(&cld->cld_refcount, 1);
235 /* Keep the mgc around until we are done */
236 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
238 if (cld_is_sptlrpc(cld)) {
239 sptlrpc_conf_log_start(logname);
240 cld->cld_cfg.cfg_obdname = obd->obd_name;
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 cmd
275 * where only clients are notified if one of cmd server restarts */
276 LASSERT(strlen(fsname) < sizeof(logname) / 2);
277 strcpy(logname, fsname);
278 if (IS_SERVER(lsi)) { /* mdt */
279 LASSERT(lcfg.cfg_instance == NULL);
280 lcfg.cfg_instance = sb;
281 strcat(logname, "-mdtir");
283 LASSERT(lcfg.cfg_instance != NULL);
284 strcat(logname, "-cliir");
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 lcfg.cfg_instance = sb != NULL ? (void *)sb : (void *)obd;
300 if (type == CONFIG_T_SPTLRPC)
301 lcfg.cfg_instance = NULL;
303 cld = config_log_find(logname, &lcfg);
304 if (unlikely(cld != NULL))
307 return do_config_log_add(obd, logname, type, &lcfg, sb);
310 /** Add this log to the list of active logs watched by an MGC.
311 * Active means we're watching for updates.
312 * We have one active log per "mount" - client instance or servername.
313 * Each instance may be at a different point in the log.
315 static struct config_llog_data *
316 config_log_add(struct obd_device *obd, char *logname,
317 struct config_llog_instance *cfg, struct super_block *sb)
319 struct lustre_sb_info *lsi = s2lsi(sb);
320 struct config_llog_data *cld = NULL;
321 struct config_llog_data *sptlrpc_cld = NULL;
322 struct config_llog_data *params_cld = NULL;
323 struct config_llog_data *nodemap_cld = NULL;
324 struct config_llog_data *barrier_cld = NULL;
331 CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
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(struct config_llog_data *cld)
456 mutex_lock(&cld->cld_lock);
457 spin_lock(&config_list_lock);
458 cld->cld_stopping = 1;
459 spin_unlock(&config_list_lock);
460 mutex_unlock(&cld->cld_lock);
464 /** Stop watching for updates on this log.
466 static int config_log_end(char *logname, struct config_llog_instance *cfg)
468 struct config_llog_data *cld;
469 struct config_llog_data *cld_sptlrpc = NULL;
470 struct config_llog_data *cld_params = NULL;
471 struct config_llog_data *cld_recover = NULL;
472 struct config_llog_data *cld_nodemap = NULL;
473 struct config_llog_data *cld_barrier = NULL;
478 cld = config_log_find(logname, cfg);
482 mutex_lock(&cld->cld_lock);
484 * if cld_stopping is set, it means we didn't start the log thus
485 * not owning the start ref. this can happen after previous umount:
486 * the cld still hanging there waiting for lock cancel, and we
487 * remount again but failed in the middle and call log_end without
490 if (unlikely(cld->cld_stopping)) {
491 mutex_unlock(&cld->cld_lock);
492 /* drop the ref from the find */
497 spin_lock(&config_list_lock);
498 cld->cld_stopping = 1;
499 spin_unlock(&config_list_lock);
501 cld_recover = cld->cld_recover;
502 cld->cld_recover = NULL;
503 cld_params = cld->cld_params;
504 cld->cld_params = NULL;
505 cld_nodemap = cld->cld_nodemap;
506 cld->cld_nodemap = NULL;
507 cld_barrier = cld->cld_barrier;
508 cld->cld_barrier = NULL;
509 cld_sptlrpc = cld->cld_sptlrpc;
510 cld->cld_sptlrpc = NULL;
511 mutex_unlock(&cld->cld_lock);
513 config_mark_cld_stop(cld_recover);
514 config_log_put(cld_recover);
516 config_mark_cld_stop(cld_params);
517 config_log_put(cld_params);
519 /* don't set cld_stopping on nm lock as other targets may be active */
520 config_log_put(cld_nodemap);
523 mutex_lock(&cld_barrier->cld_lock);
524 cld_barrier->cld_stopping = 1;
525 mutex_unlock(&cld_barrier->cld_lock);
526 config_log_put(cld_barrier);
529 config_log_put(cld_sptlrpc);
531 /* drop the ref from the find */
533 /* drop the start ref */
536 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
541 #ifdef CONFIG_PROC_FS
542 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
544 struct obd_device *obd = data;
545 struct obd_import *imp;
546 struct obd_connect_data *ocd;
547 struct config_llog_data *cld;
550 LASSERT(obd != NULL);
551 LPROCFS_CLIMP_CHECK(obd);
552 imp = obd->u.cli.cl_import;
553 ocd = &imp->imp_connect_data;
555 seq_printf(m, "imperative_recovery: %s\n",
556 OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
557 seq_printf(m, "client_state:\n");
559 spin_lock(&config_list_lock);
560 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
561 if (cld->cld_recover == NULL)
563 seq_printf(m, " - { client: %s, nidtbl_version: %u }\n",
565 cld->cld_recover->cld_cfg.cfg_last_idx);
567 spin_unlock(&config_list_lock);
569 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 0x1ff /* ~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 l_wait_info lwi;
633 struct config_llog_data *cld, *cld_prev;
634 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
637 /* Any new or requeued lostlocks will change the state */
638 rq_state &= ~(RQ_NOW | RQ_LATER);
639 spin_unlock(&config_list_lock);
646 /* Always wait a few seconds to allow the server who
647 caused the lock revocation to finish its setup, plus some
648 random so everyone doesn't try to reconnect at once. */
649 to = msecs_to_jiffies(MGC_TIMEOUT_MIN_SECONDS * MSEC_PER_SEC);
650 /* rand is centi-seconds */
651 to += msecs_to_jiffies(rand * MSEC_PER_SEC / 100);
652 lwi = LWI_TIMEOUT(to, NULL, NULL);
653 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
657 * iterate & processing through the list. for each cld, process
658 * its depending sptlrpc cld firstly (if any) and then itself.
660 * it's guaranteed any item in the list must have
661 * reference > 0; and if cld_lostlock is set, at
662 * least one reference is taken by the previous enqueue.
666 spin_lock(&config_list_lock);
667 rq_state &= ~RQ_PRECLEANUP;
668 list_for_each_entry(cld, &config_llog_list,
670 if (!cld->cld_lostlock || cld->cld_stopping)
673 /* hold reference to avoid being freed during
674 * subsequent processing. */
676 cld->cld_lostlock = 0;
677 spin_unlock(&config_list_lock);
679 config_log_put(cld_prev);
682 if (likely(!(rq_state & RQ_STOP))) {
684 spin_lock(&config_list_lock);
686 spin_lock(&config_list_lock);
690 spin_unlock(&config_list_lock);
691 config_log_put(cld_prev);
693 /* Wait a bit to see if anyone else needs a requeue */
694 lwi = (struct l_wait_info) { 0 };
695 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
697 spin_lock(&config_list_lock);
700 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
701 rq_state &= ~RQ_RUNNING;
702 spin_unlock(&config_list_lock);
706 CDEBUG(D_MGC, "Ending requeue thread\n");
710 /* Add a cld to the list to requeue. Start the requeue thread if needed.
711 We are responsible for dropping the config log reference from here on out. */
712 static void mgc_requeue_add(struct config_llog_data *cld)
717 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
718 cld->cld_logname, atomic_read(&cld->cld_refcount),
719 cld->cld_stopping, rq_state);
720 LASSERT(atomic_read(&cld->cld_refcount) > 0);
722 mutex_lock(&cld->cld_lock);
723 spin_lock(&config_list_lock);
724 if (!(rq_state & RQ_STOP) && !cld->cld_stopping && !cld->cld_lostlock) {
725 cld->cld_lostlock = 1;
729 spin_unlock(&config_list_lock);
730 mutex_unlock(&cld->cld_lock);
737 /********************** class fns **********************/
738 static int mgc_local_llog_init(const struct lu_env *env,
739 struct obd_device *obd,
740 struct obd_device *disk)
742 struct llog_ctxt *ctxt;
747 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
752 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
754 ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
760 static int mgc_local_llog_fini(const struct lu_env *env,
761 struct obd_device *obd)
763 struct llog_ctxt *ctxt;
767 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
768 llog_cleanup(env, ctxt);
773 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
774 struct super_block *sb)
776 struct lustre_sb_info *lsi = s2lsi(sb);
777 struct client_obd *cli = &obd->u.cli;
778 struct lu_fid rfid, fid;
779 struct dt_object *root, *dto;
785 LASSERT(lsi->lsi_dt_dev);
787 /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
788 mutex_lock(&cli->cl_mgc_mutex);
790 /* Setup the configs dir */
791 fid.f_seq = FID_SEQ_LOCAL_NAME;
794 rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
799 rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
803 root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
804 &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
805 if (unlikely(IS_ERR(root)))
806 GOTO(out_los, rc = PTR_ERR(root));
808 dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
810 S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
811 dt_object_put_nocache(env, root);
813 GOTO(out_los, rc = PTR_ERR(dto));
815 cli->cl_mgc_configs_dir = dto;
817 LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
818 rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
822 /* We take an obd ref to insure that we can't get to mgc_cleanup
823 * without calling mgc_fs_cleanup first. */
824 class_incref(obd, "mgc_fs", obd);
826 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
830 dt_object_put(env, cli->cl_mgc_configs_dir);
831 cli->cl_mgc_configs_dir = NULL;
835 local_oid_storage_fini(env, cli->cl_mgc_los);
836 cli->cl_mgc_los = NULL;
837 mutex_unlock(&cli->cl_mgc_mutex);
842 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
844 struct client_obd *cli = &obd->u.cli;
847 LASSERT(cli->cl_mgc_los != NULL);
849 mgc_local_llog_fini(env, obd);
851 dt_object_put_nocache(env, cli->cl_mgc_configs_dir);
852 cli->cl_mgc_configs_dir = NULL;
854 local_oid_storage_fini(env, cli->cl_mgc_los);
855 cli->cl_mgc_los = NULL;
857 class_decref(obd, "mgc_fs", obd);
858 mutex_unlock(&cli->cl_mgc_mutex);
863 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
865 struct llog_ctxt *ctxt;
870 /* setup only remote ctxt, the local disk context is switched per each
871 * filesystem during mgc_fs_setup() */
872 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
877 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
880 llog_initiator_connect(ctxt);
886 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
888 struct llog_ctxt *ctxt;
892 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
894 llog_cleanup(env, ctxt);
900 static atomic_t mgc_count = ATOMIC_INIT(0);
901 static int mgc_precleanup(struct obd_device *obd)
907 if (atomic_dec_and_test(&mgc_count)) {
908 LASSERT(rq_state & RQ_RUNNING);
909 /* stop requeue thread */
912 /* wakeup requeue thread to clean our cld */
913 temp = RQ_NOW | RQ_PRECLEANUP;
916 spin_lock(&config_list_lock);
918 spin_unlock(&config_list_lock);
922 wait_for_completion(&rq_exit);
923 obd_cleanup_client_import(obd);
925 rc = mgc_llog_fini(NULL, obd);
927 CERROR("failed to cleanup llogging subsystems\n");
932 static int mgc_cleanup(struct obd_device *obd)
937 /* COMPAT_146 - old config logs may have added profiles we don't
939 if (obd->obd_type->typ_refcnt <= 1)
940 /* Only for the last mgc */
941 class_del_profiles();
943 lprocfs_obd_cleanup(obd);
946 rc = client_obd_cleanup(obd);
950 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
952 struct task_struct *task;
956 rc = ptlrpcd_addref();
960 rc = client_obd_setup(obd, lcfg);
962 GOTO(err_decref, rc);
964 rc = mgc_llog_init(NULL, obd);
966 CERROR("failed to setup llogging subsystems\n");
967 GOTO(err_cleanup, rc);
970 #ifdef CONFIG_PROC_FS
971 obd->obd_vars = lprocfs_mgc_obd_vars;
972 lprocfs_obd_setup(obd, true);
974 sptlrpc_lprocfs_cliobd_attach(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",
987 GOTO(err_cleanup, rc);
989 /* rc is the task_struct pointer of mgc_requeue_thread. */
991 wait_for_completion(&rq_start);
997 client_obd_cleanup(obd);
1003 /* based on ll_mdc_blocking_ast */
1004 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1005 void *data, int flag)
1007 struct lustre_handle lockh;
1008 struct config_llog_data *cld = (struct config_llog_data *)data;
1013 case LDLM_CB_BLOCKING:
1014 /* mgs wants the lock, give it up... */
1015 LDLM_DEBUG(lock, "MGC blocking CB");
1016 ldlm_lock2handle(lock, &lockh);
1017 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
1019 case LDLM_CB_CANCELING:
1020 /* We've given up the lock, prepare ourselves to update. */
1021 LDLM_DEBUG(lock, "MGC cancel CB");
1023 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
1024 PLDLMRES(lock->l_resource),
1025 (char *)&lock->l_resource->lr_name.name[0]);
1028 CDEBUG(D_INFO, "missing data, won't requeue\n");
1032 /* held at mgc_process_log(). */
1033 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1035 lock->l_ast_data = NULL;
1036 /* Are we done with this log? */
1037 if (cld->cld_stopping) {
1038 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1040 config_log_put(cld);
1043 /* Make sure not to re-enqueue when the mgc is stopping
1044 (we get called from client_disconnect_export) */
1045 if (lock->l_conn_export == NULL ||
1046 lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1047 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1049 config_log_put(cld);
1053 /* Re-enqueue now */
1054 mgc_requeue_add(cld);
1055 config_log_put(cld);
1064 /* Not sure where this should go... */
1065 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1066 * that we can have a chance to try the secondary MGS if any. */
1067 #define MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1069 #define MGC_TARGET_REG_LIMIT 10
1070 #define MGC_TARGET_REG_LIMIT_MAX RECONNECT_DELAY_MAX
1071 #define MGC_SEND_PARAM_LIMIT 10
1073 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1074 /* Send parameter to MGS*/
1075 static int mgc_set_mgs_param(struct obd_export *exp,
1076 struct mgs_send_param *msp)
1078 struct ptlrpc_request *req;
1079 struct mgs_send_param *req_msp, *rep_msp;
1083 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1084 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1089 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1091 ptlrpc_req_finished(req);
1095 memcpy(req_msp, msp, sizeof(*req_msp));
1096 ptlrpc_request_set_replen(req);
1098 /* Limit how long we will wait for the enqueue to complete */
1099 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1100 rc = ptlrpc_queue_wait(req);
1102 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1103 memcpy(msp, rep_msp, sizeof(*rep_msp));
1106 ptlrpc_req_finished(req);
1112 /* Take a config lock so we can get cancel notifications */
1113 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1114 union ldlm_policy_data *policy, enum ldlm_mode mode,
1115 __u64 *flags, ldlm_glimpse_callback glimpse_callback,
1116 void *data, __u32 lvb_len, void *lvb_swabber,
1117 struct lustre_handle *lockh)
1119 struct config_llog_data *cld = (struct config_llog_data *)data;
1120 struct ldlm_enqueue_info einfo = {
1123 .ei_cb_bl = mgc_blocking_ast,
1124 .ei_cb_cp = ldlm_completion_ast,
1125 .ei_cb_gl = glimpse_callback,
1127 struct ptlrpc_request *req;
1128 int short_limit = cld_is_sptlrpc(cld);
1132 CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1133 cld->cld_resid.name[0]);
1135 /* We need a callback for every lockholder, so don't try to
1136 ldlm_lock_match (see rev 1.1.2.11.2.47) */
1137 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1138 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1143 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1144 ptlrpc_request_set_replen(req);
1146 /* check if this is server or client */
1147 if (cld->cld_cfg.cfg_sb) {
1148 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1149 if (lsi && IS_SERVER(lsi))
1152 /* Limit how long we will wait for the enqueue to complete */
1153 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1154 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1155 NULL, 0, LVB_T_NONE, lockh, 0);
1156 /* A failed enqueue should still call the mgc_blocking_ast,
1157 where it will be requeued if needed ("grant failed"). */
1158 ptlrpc_req_finished(req);
1162 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1163 struct lustre_handle *lockh)
1167 ldlm_lock_decref(lockh, mode);
1172 static void mgc_notify_active(struct obd_device *unused)
1174 /* wakeup mgc_requeue_thread to requeue mgc lock */
1175 spin_lock(&config_list_lock);
1177 spin_unlock(&config_list_lock);
1180 /* TODO: Help the MGS rebuild nidtbl. -jay */
1183 /* Send target_reg message to MGS */
1184 static int mgc_target_register(struct obd_export *exp,
1185 struct mgs_target_info *mti)
1187 struct ptlrpc_request *req;
1188 struct mgs_target_info *req_mti, *rep_mti;
1192 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1193 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1198 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1200 ptlrpc_req_finished(req);
1204 memcpy(req_mti, mti, sizeof(*req_mti));
1205 ptlrpc_request_set_replen(req);
1206 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1207 /* Limit how long we will wait for the enqueue to complete */
1208 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1210 /* if the target needs to regenerate the config log in MGS, it's better
1211 * to use some longer limit to let MGC have time to change connection to
1212 * another MGS (or try again with the same MGS) for the target (server)
1213 * will fail and exit if the request expired due to delay limit. */
1214 if (mti->mti_flags & (LDD_F_UPDATE | LDD_F_NEED_INDEX))
1215 req->rq_delay_limit = MGC_TARGET_REG_LIMIT_MAX;
1217 rc = ptlrpc_queue_wait(req);
1219 rep_mti = req_capsule_server_get(&req->rq_pill,
1220 &RMF_MGS_TARGET_INFO);
1221 memcpy(mti, rep_mti, sizeof(*rep_mti));
1222 CDEBUG(D_MGC, "register %s got index = %d\n",
1223 mti->mti_svname, mti->mti_stripe_index);
1225 ptlrpc_req_finished(req);
1230 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1231 u32 keylen, void *key,
1232 u32 vallen, void *val,
1233 struct ptlrpc_request_set *set)
1238 /* Turn off initial_recov after we try all backup servers once */
1239 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1240 struct obd_import *imp = class_exp2cliimp(exp);
1242 if (vallen != sizeof(int))
1244 value = *(int *)val;
1245 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1246 imp->imp_obd->obd_name, value,
1247 imp->imp_deactive, imp->imp_invalid,
1248 imp->imp_replayable, imp->imp_obd->obd_replayable,
1249 ptlrpc_import_state_name(imp->imp_state));
1250 /* Resurrect the import immediately if
1251 * 1. we previously got disconnected,
1252 * 2. value > 1 (at the same node with MGS)
1254 if (imp->imp_state == LUSTRE_IMP_DISCON || value > 1)
1255 ptlrpc_reconnect_import(imp);
1260 /* FIXME move this to mgc_process_config */
1261 if (KEY_IS(KEY_REGISTER_TARGET)) {
1262 struct mgs_target_info *mti;
1263 if (vallen != sizeof(struct mgs_target_info))
1265 mti = (struct mgs_target_info *)val;
1266 CDEBUG(D_MGC, "register_target %s %#x\n",
1267 mti->mti_svname, mti->mti_flags);
1268 rc = mgc_target_register(exp, mti);
1271 if (KEY_IS(KEY_SET_FS)) {
1272 struct super_block *sb = (struct super_block *)val;
1274 if (vallen != sizeof(struct super_block))
1277 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1280 if (KEY_IS(KEY_CLEAR_FS)) {
1283 rc = mgc_fs_cleanup(env, exp->exp_obd);
1286 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1287 if (KEY_IS(KEY_SET_INFO)) {
1288 struct mgs_send_param *msp;
1290 msp = (struct mgs_send_param *)val;
1291 rc = mgc_set_mgs_param(exp, msp);
1295 if (KEY_IS(KEY_MGSSEC)) {
1296 struct client_obd *cli = &exp->exp_obd->u.cli;
1297 struct sptlrpc_flavor flvr;
1300 * empty string means using current flavor, if which haven't
1301 * been set yet, set it as null.
1303 * if flavor has been set previously, check the asking flavor
1304 * must match the existing one.
1307 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1313 rc = sptlrpc_parse_flavor(val, &flvr);
1315 CERROR("invalid sptlrpc flavor %s to MGS\n",
1321 * caller already hold a mutex
1323 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1324 cli->cl_flvr_mgc = flvr;
1325 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1326 sizeof(flvr)) != 0) {
1329 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1331 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1332 "currently %s is in use\n",
1342 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1343 __u32 keylen, void *key, __u32 *vallen, void *val)
1347 if (KEY_IS(KEY_CONN_DATA)) {
1348 struct obd_import *imp = class_exp2cliimp(exp);
1349 struct obd_connect_data *data = val;
1351 if (*vallen == sizeof(*data)) {
1352 *data = imp->imp_connect_data;
1360 static int mgc_import_event(struct obd_device *obd,
1361 struct obd_import *imp,
1362 enum obd_import_event event)
1366 LASSERT(imp->imp_obd == obd);
1367 CDEBUG(D_MGC, "import event %#x\n", event);
1370 case IMP_EVENT_DISCON:
1371 /* MGC imports should not wait for recovery */
1372 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1373 ptlrpc_pinger_ir_down();
1375 case IMP_EVENT_INACTIVE:
1377 case IMP_EVENT_INVALIDATE: {
1378 struct ldlm_namespace *ns = obd->obd_namespace;
1379 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1382 case IMP_EVENT_ACTIVE:
1383 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1384 /* Clearing obd_no_recov allows us to continue pinging */
1385 obd->obd_no_recov = 0;
1386 mgc_notify_active(obd);
1387 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1388 ptlrpc_pinger_ir_up();
1392 case IMP_EVENT_DEACTIVATE:
1393 case IMP_EVENT_ACTIVATE:
1396 CERROR("Unknown import event %#x\n", event);
1403 CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1404 CONFIG_READ_NRPAGES = 4
1407 static int mgc_apply_recover_logs(struct obd_device *mgc,
1408 struct config_llog_data *cld,
1410 void *data, int datalen, bool mne_swab)
1412 struct config_llog_instance *cfg = &cld->cld_cfg;
1413 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1414 struct mgs_nidtbl_entry *entry;
1415 struct lustre_cfg *lcfg;
1416 struct lustre_cfg_bufs bufs;
1417 u64 prev_version = 0;
1426 LASSERT(cfg->cfg_instance != NULL);
1427 LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1429 OBD_ALLOC(inst, PAGE_SIZE);
1433 if (!IS_SERVER(lsi)) {
1434 pos = snprintf(inst, PAGE_SIZE, "%p", cfg->cfg_instance);
1435 if (pos >= PAGE_SIZE) {
1436 OBD_FREE(inst, PAGE_SIZE);
1440 LASSERT(IS_MDT(lsi));
1441 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1444 OBD_FREE(inst, PAGE_SIZE);
1452 bufsz = PAGE_SIZE - pos;
1454 while (datalen > 0) {
1455 int entry_len = sizeof(*entry);
1457 struct obd_device *obd;
1464 if (datalen < sizeof(*entry))
1467 entry = (typeof(entry))(data + off);
1470 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1472 if (entry->mne_nid_count == 0) /* at least one nid entry */
1474 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1477 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1478 if (datalen < entry_len) /* must have entry_len at least */
1481 /* Keep this swab for normal mixed endian handling. LU-1644 */
1483 lustre_swab_mgs_nidtbl_entry(entry);
1484 if (entry->mne_length > PAGE_SIZE) {
1485 CERROR("MNE too large (%u)\n", entry->mne_length);
1489 if (entry->mne_length < entry_len)
1492 off += entry->mne_length;
1493 datalen -= entry->mne_length;
1497 if (entry->mne_version > max_version) {
1498 CERROR("entry index(%lld) is over max_index(%lld)\n",
1499 entry->mne_version, max_version);
1503 if (prev_version >= entry->mne_version) {
1504 CERROR("index unsorted, prev %lld, now %lld\n",
1505 prev_version, entry->mne_version);
1508 prev_version = entry->mne_version;
1511 * Write a string with format "nid::instance" to
1512 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1515 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1516 memset(buf, 0, bufsz);
1520 /* lustre-OST0001-osc-<instance #> */
1521 strcpy(obdname, cld->cld_logname);
1522 cname = strrchr(obdname, '-');
1523 if (cname == NULL) {
1524 CERROR("mgc %s: invalid logname %s\n",
1525 mgc->obd_name, obdname);
1529 pos = cname - obdname;
1531 pos += sprintf(obdname + pos, "-%s%04x",
1532 is_ost ? "OST" : "MDT", entry->mne_index);
1534 cname = is_ost ? "osc" : "mdc",
1535 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1536 lustre_cfg_bufs_reset(&bufs, obdname);
1538 /* find the obd by obdname */
1539 obd = class_name2obd(obdname);
1541 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1542 mgc->obd_name, obdname);
1544 /* this is a safe race, when the ost is starting up...*/
1548 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1551 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1554 down_read(&obd->u.cli.cl_sem);
1555 if (obd->u.cli.cl_import == NULL) {
1556 /* client does not connect to the OST yet */
1557 up_read(&obd->u.cli.cl_sem);
1562 /* iterate all nids to find one */
1563 /* find uuid by nid */
1565 for (i = 0; i < entry->mne_nid_count; i++) {
1566 rc = client_import_find_conn(obd->u.cli.cl_import,
1568 (struct obd_uuid *)uuid);
1573 up_read(&obd->u.cli.cl_sem);
1575 CERROR("mgc: cannot find uuid by nid %s\n",
1576 libcfs_nid2str(entry->u.nids[0]));
1580 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1581 uuid, libcfs_nid2str(entry->u.nids[0]));
1583 pos += strlen(uuid);
1584 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1585 LASSERT(pos < bufsz);
1587 lustre_cfg_bufs_set_string(&bufs, 1, params);
1589 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1595 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1597 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1598 prev_version, max_version, obdname, params);
1600 rc = class_process_config(lcfg);
1601 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1602 lcfg->lcfg_buflens));
1604 CDEBUG(D_INFO, "process config for %s error %d\n",
1607 /* continue, even one with error */
1610 OBD_FREE(inst, PAGE_SIZE);
1615 * This function is called if this client was notified for target restarting
1616 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1619 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1620 struct config_llog_data *cld)
1622 struct ptlrpc_connection *mgc_conn;
1623 struct ptlrpc_request *req = NULL;
1624 struct config_llog_instance *cfg = &cld->cld_cfg;
1625 struct mgs_config_body *body;
1626 struct mgs_config_res *res;
1627 struct nodemap_config *new_config = NULL;
1628 struct lu_nodemap *recent_nodemap = NULL;
1629 struct ptlrpc_bulk_desc *desc;
1630 struct page **pages = NULL;
1631 __u64 config_read_offset = 0;
1632 __u8 nodemap_cur_pass = 0;
1635 bool mne_swab = false;
1641 mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1643 /* don't need to get local config */
1644 if (cld_is_nodemap(cld) &&
1645 (LNET_NETTYP(LNET_NIDNET(mgc_conn->c_peer.nid)) == LOLND))
1648 /* allocate buffer for bulk transfer.
1649 * if this is the first time for this mgs to read logs,
1650 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1651 * once; otherwise, it only reads increment of logs, this should be
1652 * small and CONFIG_READ_NRPAGES will be used.
1654 nrpages = CONFIG_READ_NRPAGES;
1655 if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1656 nrpages = CONFIG_READ_NRPAGES_INIT;
1658 OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1660 GOTO(out, rc = -ENOMEM);
1662 for (i = 0; i < nrpages; i++) {
1663 pages[i] = alloc_page(GFP_KERNEL);
1664 if (pages[i] == NULL)
1665 GOTO(out, rc = -ENOMEM);
1669 #ifdef HAVE_SERVER_SUPPORT
1670 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1671 new_config = nodemap_config_alloc();
1672 if (IS_ERR(new_config)) {
1673 rc = PTR_ERR(new_config);
1679 LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1680 LASSERT(mutex_is_locked(&cld->cld_lock));
1681 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1682 &RQF_MGS_CONFIG_READ);
1684 GOTO(out, rc = -ENOMEM);
1686 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1691 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1692 LASSERT(body != NULL);
1693 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1694 if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1695 >= sizeof(body->mcb_name))
1696 GOTO(out, rc = -E2BIG);
1697 if (cld_is_nodemap(cld))
1698 body->mcb_offset = config_read_offset;
1700 body->mcb_offset = cfg->cfg_last_idx + 1;
1701 body->mcb_type = cld->cld_type;
1702 body->mcb_bits = PAGE_SHIFT;
1703 body->mcb_units = nrpages;
1704 body->mcb_nm_cur_pass = nodemap_cur_pass;
1706 /* allocate bulk transfer descriptor */
1707 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1708 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1710 &ptlrpc_bulk_kiov_pin_ops);
1712 GOTO(out, rc = -ENOMEM);
1714 for (i = 0; i < nrpages; i++)
1715 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1718 ptlrpc_request_set_replen(req);
1719 rc = ptlrpc_queue_wait(req);
1723 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1725 GOTO(out, rc = -EPROTO);
1727 if (cld_is_nodemap(cld)) {
1728 config_read_offset = res->mcr_offset;
1729 eof = config_read_offset == II_END_OFF;
1730 nodemap_cur_pass = res->mcr_nm_cur_pass;
1732 if (res->mcr_size < res->mcr_offset)
1733 GOTO(out, rc = -EINVAL);
1735 /* always update the index even though it might have errors with
1736 * handling the recover logs
1738 cfg->cfg_last_idx = res->mcr_offset;
1739 eof = res->mcr_offset == res->mcr_size;
1741 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1742 res->mcr_offset, eof == false);
1745 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1747 GOTO(out, rc = ealen);
1749 if (ealen > nrpages << PAGE_SHIFT)
1750 GOTO(out, rc = -EINVAL);
1752 if (ealen == 0) { /* no logs transferred */
1753 #ifdef HAVE_SERVER_SUPPORT
1754 /* config changed since first read RPC */
1755 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1756 CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1757 GOTO(out, rc = -EAGAIN);
1765 mne_swab = !!ptlrpc_rep_need_swab(req);
1766 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1767 /* This import flag means the server did an extra swab of IR MNE
1768 * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1769 if (unlikely(req->rq_import->imp_need_mne_swab))
1770 mne_swab = !mne_swab;
1773 /* When a nodemap config is received, we build a new nodemap config,
1774 * with new nodemap structs. We keep track of the most recently added
1775 * nodemap since the config is read ordered by nodemap_id, and so it
1776 * is likely that the next record will be related. Because access to
1777 * the nodemaps is single threaded until the nodemap_config is active,
1778 * we don't need to reference count with recent_nodemap, though
1779 * recent_nodemap should be set to NULL when the nodemap_config
1780 * is either destroyed or set active.
1782 for (i = 0; i < nrpages && ealen > 0; i++) {
1786 ptr = kmap(pages[i]);
1787 if (cld_is_nodemap(cld))
1788 rc2 = nodemap_process_idx_pages(new_config, ptr,
1791 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1798 CWARN("%s: error processing %s log %s: rc = %d\n",
1800 cld_is_nodemap(cld) ? "nodemap" : "recovery",
1803 GOTO(out, rc = rc2);
1811 ptlrpc_req_finished(req);
1815 if (rc == 0 && !eof)
1818 #ifdef HAVE_SERVER_SUPPORT
1819 if (new_config != NULL) {
1820 /* recent_nodemap cannot be used after set_active/dealloc */
1822 nodemap_config_set_active_mgc(new_config);
1824 nodemap_config_dealloc(new_config);
1829 for (i = 0; i < nrpages; i++) {
1830 if (pages[i] == NULL)
1832 __free_page(pages[i]);
1834 OBD_FREE(pages, sizeof(*pages) * nrpages);
1839 static int mgc_barrier_glimpse_ast(struct ldlm_lock *lock, void *data)
1841 struct config_llog_data *cld = lock->l_ast_data;
1845 if (cld->cld_stopping)
1848 rc = barrier_handler(s2lsi(cld->cld_cfg.cfg_sb)->lsi_dt_dev,
1849 (struct ptlrpc_request *)data);
1854 /* Copy a remote log locally */
1855 static int mgc_llog_local_copy(const struct lu_env *env,
1856 struct obd_device *obd,
1857 struct llog_ctxt *rctxt,
1858 struct llog_ctxt *lctxt, char *logname)
1866 * - copy it to backup using llog_backup()
1867 * - copy remote llog to logname using llog_backup()
1868 * - if failed then move bakup to logname again
1871 OBD_ALLOC(temp_log, strlen(logname) + 2);
1874 sprintf(temp_log, "%sT", logname);
1876 /* make a copy of local llog at first */
1877 rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1878 if (rc < 0 && rc != -ENOENT)
1880 /* copy remote llog to the local copy */
1881 rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1882 if (rc == -ENOENT) {
1883 /* no remote llog, delete local one too */
1884 llog_erase(env, lctxt, NULL, logname);
1885 } else if (rc < 0) {
1886 /* error during backup, get local one back from the copy */
1887 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1889 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1890 obd->obd_name, logname, rc);
1892 llog_erase(env, lctxt, NULL, temp_log);
1893 OBD_FREE(temp_log, strlen(logname) + 2);
1897 /* local_only means it cannot get remote llogs */
1898 static int mgc_process_cfg_log(struct obd_device *mgc,
1899 struct config_llog_data *cld, int local_only)
1901 struct llog_ctxt *ctxt, *lctxt = NULL;
1902 struct client_obd *cli = &mgc->u.cli;
1903 struct lustre_sb_info *lsi = NULL;
1910 LASSERT(mutex_is_locked(&cld->cld_lock));
1912 if (cld->cld_cfg.cfg_sb)
1913 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1919 rc = lu_env_init(env, LCT_MG_THREAD);
1923 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1926 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1928 /* Copy the setup log locally if we can. Don't mess around if we're
1929 * running an MGS though (logs are already local). */
1930 if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1931 cli->cl_mgc_configs_dir != NULL &&
1932 lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1934 if (!local_only && !lsi->lsi_dt_dev->dd_rdonly)
1935 /* Only try to copy log if we have the lock. */
1936 rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1938 if (local_only || rc) {
1939 if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1940 llog_is_empty(env, lctxt, cld->cld_logname)) {
1941 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1942 "log %s and no local copy."
1943 "\n", cld->cld_logname);
1944 GOTO(out_pop, rc = -ENOENT);
1946 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1947 "copy for now, will try to update later.\n",
1951 /* Now, whether we copied or not, start using the local llog.
1952 * If we failed to copy, we'll start using whatever the old
1954 llog_ctxt_put(ctxt);
1958 if (local_only) /* no local log at client side */
1959 GOTO(out_pop, rc = -EIO);
1963 if (lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1964 lsi->lsi_dt_dev->dd_rdonly) {
1965 struct llog_ctxt *rctxt;
1967 /* Under readonly mode, we may have no local copy or local
1968 * copy is incomplete, so try to use remote llog firstly. */
1969 rctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1972 rc = class_config_parse_llog(env, rctxt, cld->cld_logname,
1974 llog_ctxt_put(rctxt);
1977 if (rc && rc != -ENOENT)
1978 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1982 * update settings on existing OBDs. doing it inside
1983 * of llog_process_lock so no device is attaching/detaching
1985 * the logname must be <fsname>-sptlrpc
1987 if (rc == 0 && cld_is_sptlrpc(cld))
1988 class_notify_sptlrpc_conf(cld->cld_logname,
1989 strlen(cld->cld_logname) -
1990 strlen("-sptlrpc"));
1994 __llog_ctxt_put(env, ctxt);
1996 __llog_ctxt_put(env, lctxt);
2004 static bool mgc_import_in_recovery(struct obd_import *imp)
2006 bool in_recovery = true;
2008 spin_lock(&imp->imp_lock);
2009 if (imp->imp_state == LUSTRE_IMP_FULL ||
2010 imp->imp_state == LUSTRE_IMP_CLOSED)
2011 in_recovery = false;
2012 spin_unlock(&imp->imp_lock);
2018 * Get a configuration log from the MGS and process it.
2020 * This function is called for both clients and servers to process the
2021 * configuration log from the MGS. The MGC enqueues a DLM lock on the
2022 * log from the MGS, and if the lock gets revoked the MGC will be notified
2023 * by the lock cancellation callback that the config log has changed,
2024 * and will enqueue another MGS lock on it, and then continue processing
2025 * the new additions to the end of the log.
2027 * Since the MGC import is not replayable, if the import is being evicted
2028 * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
2029 * the log until recovery is finished or the import is closed.
2031 * Make a local copy of the log before parsing it if appropriate (non-MGS
2032 * server) so that the server can start even when the MGS is down.
2034 * There shouldn't be multiple processes running process_log at once --
2035 * sounds like badness. It actually might be fine, as long as they're not
2036 * trying to update from the same log simultaneously, in which case we
2037 * should use a per-log semaphore instead of cld_lock.
2039 * \param[in] mgc MGC device by which to fetch the configuration log
2040 * \param[in] cld log processing state (stored in lock callback data)
2042 * \retval 0 on success
2043 * \retval negative errno on failure
2045 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
2047 struct lustre_handle lockh = { 0 };
2048 __u64 flags = LDLM_FL_NO_LRU;
2053 LASSERT(cld != NULL);
2055 /* I don't want multiple processes running process_log at once --
2056 sounds like badness. It actually might be fine, as long as
2057 we're not trying to update from the same log
2058 simultaneously (in which case we should use a per-log sem.) */
2060 mutex_lock(&cld->cld_lock);
2061 if (cld->cld_stopping) {
2062 mutex_unlock(&cld->cld_lock);
2066 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2068 CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
2069 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2071 /* Get the cfg lock on the llog */
2072 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2074 cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
2075 cld, 0, NULL, &lockh);
2077 /* Get the cld, it will be released in mgc_blocking_ast. */
2078 config_log_get(cld);
2079 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2082 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2084 if (rcl == -ESHUTDOWN &&
2085 atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2086 struct obd_import *imp;
2087 struct l_wait_info lwi;
2088 int secs = cfs_time_seconds(obd_timeout);
2090 mutex_unlock(&cld->cld_lock);
2091 imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2093 /* Let's force the pinger, and wait the import to be
2094 * connected, note: since mgc import is non-replayable,
2095 * and even the import state is disconnected, it does
2096 * not mean the "recovery" is stopped, so we will keep
2097 * waitting until timeout or the import state is
2099 ptlrpc_pinger_force(imp);
2101 lwi = LWI_TIMEOUT(secs, NULL, NULL);
2102 l_wait_event(imp->imp_recovery_waitq,
2103 !mgc_import_in_recovery(imp), &lwi);
2105 if (imp->imp_state == LUSTRE_IMP_FULL) {
2109 mutex_lock(&cld->cld_lock);
2110 spin_lock(&config_list_lock);
2111 cld->cld_lostlock = 1;
2112 spin_unlock(&config_list_lock);
2115 /* mark cld_lostlock so that it will requeue
2116 * after MGC becomes available. */
2117 spin_lock(&config_list_lock);
2118 cld->cld_lostlock = 1;
2119 spin_unlock(&config_list_lock);
2123 if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2125 rc = mgc_process_recover_nodemap_log(mgc, cld);
2126 else if (cld_is_nodemap(cld))
2129 if (cld_is_recover(cld) && rc) {
2131 CERROR("%s: recover log %s failed, not fatal: rc = %d\n",
2132 mgc->obd_name, cld->cld_logname, rc);
2133 spin_lock(&config_list_lock);
2134 cld->cld_lostlock = 1;
2135 spin_unlock(&config_list_lock);
2137 rc = 0; /* this is not a fatal error for recover log */
2139 } else if (!cld_is_barrier(cld)) {
2140 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2143 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2144 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2146 mutex_unlock(&cld->cld_lock);
2148 /* Now drop the lock so MGS can revoke it */
2150 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2152 CERROR("Can't drop cfg lock: %d\n", rcl);
2155 /* requeue nodemap lock immediately if transfer was interrupted */
2156 if (cld_is_nodemap(cld) && rc == -EAGAIN) {
2157 mgc_requeue_add(cld);
2165 /** Called from lustre_process_log.
2166 * LCFG_LOG_START gets the config log from the MGS, processes it to start
2167 * any services, and adds it to the list logs to watch (follow).
2169 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2171 struct lustre_cfg *lcfg = buf;
2172 struct config_llog_instance *cfg = NULL;
2177 switch(lcfg->lcfg_command) {
2178 case LCFG_LOV_ADD_OBD: {
2179 /* Overloading this cfg command: register a new target */
2180 struct mgs_target_info *mti;
2182 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2183 sizeof(struct mgs_target_info))
2184 GOTO(out, rc = -EINVAL);
2186 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2187 CDEBUG(D_MGC, "add_target %s %#x\n",
2188 mti->mti_svname, mti->mti_flags);
2189 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2192 case LCFG_LOV_DEL_OBD:
2193 /* Unregister has no meaning at the moment. */
2194 CERROR("lov_del_obd unimplemented\n");
2197 case LCFG_SPTLRPC_CONF: {
2198 rc = sptlrpc_process_config(lcfg);
2201 case LCFG_LOG_START: {
2202 struct config_llog_data *cld;
2203 struct super_block *sb;
2205 logname = lustre_cfg_string(lcfg, 1);
2206 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2207 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2209 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2212 /* We're only called through here on the initial mount */
2213 cld = config_log_add(obd, logname, cfg, sb);
2220 /* FIXME only set this for old logs! Right now this forces
2221 us to always skip the "inside markers" check */
2222 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
2224 rc = mgc_process_log(obd, cld);
2225 if (rc == 0 && cld->cld_recover != NULL) {
2226 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2227 imp_connect_data, IMP_RECOV)) {
2228 rc = mgc_process_log(obd, cld->cld_recover);
2230 struct config_llog_data *cir;
2232 mutex_lock(&cld->cld_lock);
2233 cir = cld->cld_recover;
2234 cld->cld_recover = NULL;
2235 mutex_unlock(&cld->cld_lock);
2236 config_log_put(cir);
2240 CERROR("Cannot process recover llog %d\n", rc);
2243 if (rc == 0 && cld->cld_params != NULL) {
2244 rc = mgc_process_log(obd, cld->cld_params);
2245 if (rc == -ENOENT) {
2246 CDEBUG(D_MGC, "There is no params "
2247 "config file yet\n");
2250 /* params log is optional */
2252 CERROR("%s: can't process params llog: rc = %d\n",
2258 case LCFG_LOG_END: {
2259 logname = lustre_cfg_string(lcfg, 1);
2261 if (lcfg->lcfg_bufcount >= 2)
2262 cfg = (struct config_llog_instance *)lustre_cfg_buf(
2264 rc = config_log_end(logname, cfg);
2268 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2269 GOTO(out, rc = -EINVAL);
2277 static struct obd_ops mgc_obd_ops = {
2278 .o_owner = THIS_MODULE,
2279 .o_setup = mgc_setup,
2280 .o_precleanup = mgc_precleanup,
2281 .o_cleanup = mgc_cleanup,
2282 .o_add_conn = client_import_add_conn,
2283 .o_del_conn = client_import_del_conn,
2284 .o_connect = client_connect_import,
2285 .o_disconnect = client_disconnect_export,
2286 .o_set_info_async = mgc_set_info_async,
2287 .o_get_info = mgc_get_info,
2288 .o_import_event = mgc_import_event,
2289 .o_process_config = mgc_process_config,
2292 static int __init mgc_init(void)
2294 return class_register_type(&mgc_obd_ops, NULL, true, NULL,
2295 LUSTRE_MGC_NAME, NULL);
2298 static void __exit mgc_exit(void)
2300 class_unregister_type(LUSTRE_MGC_NAME);
2303 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2304 MODULE_DESCRIPTION("Lustre Management Client");
2305 MODULE_VERSION(LUSTRE_VERSION_STRING);
2306 MODULE_LICENSE("GPL");
2308 module_init(mgc_init);
2309 module_exit(mgc_exit);