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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mgc/mgc_request.c
38 * Author: Nathan Rutman <nathan@clusterfs.com>
41 #define DEBUG_SUBSYSTEM S_MGC
42 #define D_MGC D_CONFIG /*|D_WARNING*/
44 #include <linux/module.h>
45 #include "../include/obd_class.h"
46 #include "../include/lustre_dlm.h"
47 #include "../include/lprocfs_status.h"
48 #include "../include/lustre_log.h"
49 #include "../include/lustre_disk.h"
50 #include "../include/dt_object.h"
52 #include "mgc_internal.h"
54 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
59 if (len > sizeof(resname)) {
60 CERROR("name too long: %s\n", name);
64 CERROR("missing name: %s\n", name);
67 memcpy(&resname, name, len);
69 /* Always use the same endianness for the resid */
70 memset(res_id, 0, sizeof(*res_id));
71 res_id->name[0] = cpu_to_le64(resname);
72 /* XXX: unfortunately, sptlprc and config llog share one lock */
75 case CONFIG_T_SPTLRPC:
78 case CONFIG_T_RECOVER:
85 res_id->name[1] = cpu_to_le64(resname);
86 CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
87 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
91 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
93 /* fsname is at most 8 chars long, maybe contain "-".
94 * e.g. "lustre", "SUN-000" */
95 return mgc_name2resid(fsname, strlen(fsname), res_id, type);
97 EXPORT_SYMBOL(mgc_fsname2resid);
99 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
104 /* logname consists of "fsname-nodetype".
105 * e.g. "lustre-MDT0001", "SUN-000-client"
106 * there is an exception: llog "params" */
107 name_end = strrchr(logname, '-');
109 len = strlen(logname);
111 len = name_end - logname;
112 return mgc_name2resid(logname, len, res_id, type);
115 /********************** config llog list **********************/
116 static LIST_HEAD(config_llog_list);
117 static DEFINE_SPINLOCK(config_list_lock);
119 /* Take a reference to a config log */
120 static int config_log_get(struct config_llog_data *cld)
122 atomic_inc(&cld->cld_refcount);
123 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
124 atomic_read(&cld->cld_refcount));
128 /* Drop a reference to a config log. When no longer referenced,
129 we can free the config log data */
130 static void config_log_put(struct config_llog_data *cld)
132 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
133 atomic_read(&cld->cld_refcount));
134 LASSERT(atomic_read(&cld->cld_refcount) > 0);
136 /* spinlock to make sure no item with 0 refcount in the list */
137 if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
138 list_del(&cld->cld_list_chain);
139 spin_unlock(&config_list_lock);
141 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
143 if (cld->cld_recover)
144 config_log_put(cld->cld_recover);
145 if (cld->cld_sptlrpc)
146 config_log_put(cld->cld_sptlrpc);
148 config_log_put(cld->cld_params);
149 if (cld_is_sptlrpc(cld))
150 sptlrpc_conf_log_stop(cld->cld_logname);
152 class_export_put(cld->cld_mgcexp);
153 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
157 /* Find a config log by name */
159 struct config_llog_data *config_log_find(char *logname,
160 struct config_llog_instance *cfg)
162 struct config_llog_data *cld;
163 struct config_llog_data *found = NULL;
166 LASSERT(logname != NULL);
168 instance = cfg ? cfg->cfg_instance : NULL;
169 spin_lock(&config_list_lock);
170 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
171 /* check if instance equals */
172 if (instance != cld->cld_cfg.cfg_instance)
175 /* instance may be NULL, should check name */
176 if (strcmp(logname, cld->cld_logname) == 0) {
182 atomic_inc(&found->cld_refcount);
183 LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0);
185 spin_unlock(&config_list_lock);
190 struct config_llog_data *do_config_log_add(struct obd_device *obd,
193 struct config_llog_instance *cfg,
194 struct super_block *sb)
196 struct config_llog_data *cld;
199 CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
200 cfg ? cfg->cfg_instance : NULL);
202 OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
204 return ERR_PTR(-ENOMEM);
206 strcpy(cld->cld_logname, logname);
210 cld->cld_cfg.cfg_callback = class_config_llog_handler;
211 mutex_init(&cld->cld_lock);
212 cld->cld_cfg.cfg_last_idx = 0;
213 cld->cld_cfg.cfg_flags = 0;
214 cld->cld_cfg.cfg_sb = sb;
215 cld->cld_type = type;
216 atomic_set(&cld->cld_refcount, 1);
218 /* Keep the mgc around until we are done */
219 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
221 if (cld_is_sptlrpc(cld)) {
222 sptlrpc_conf_log_start(logname);
223 cld->cld_cfg.cfg_obdname = obd->obd_name;
226 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
228 spin_lock(&config_list_lock);
229 list_add(&cld->cld_list_chain, &config_llog_list);
230 spin_unlock(&config_list_lock);
237 if (cld_is_sptlrpc(cld)) {
238 rc = mgc_process_log(obd, cld);
239 if (rc && rc != -ENOENT)
240 CERROR("failed processing sptlrpc log: %d\n", rc);
246 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
248 struct config_llog_instance *cfg,
249 struct super_block *sb)
251 struct config_llog_instance lcfg = *cfg;
252 struct lustre_sb_info *lsi = s2lsi(sb);
253 struct config_llog_data *cld;
259 /* for osp-on-ost, see lustre_start_osp() */
260 if (IS_MDT(lsi) && lcfg.cfg_instance)
263 /* we have to use different llog for clients and mdts for cmd
264 * where only clients are notified if one of cmd server restarts */
265 LASSERT(strlen(fsname) < sizeof(logname) / 2);
266 strcpy(logname, fsname);
267 if (IS_SERVER(lsi)) { /* mdt */
268 LASSERT(lcfg.cfg_instance == NULL);
269 lcfg.cfg_instance = sb;
270 strcat(logname, "-mdtir");
272 LASSERT(lcfg.cfg_instance != NULL);
273 strcat(logname, "-cliir");
276 cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
280 static struct config_llog_data *config_params_log_add(struct obd_device *obd,
281 struct config_llog_instance *cfg, struct super_block *sb)
283 struct config_llog_instance lcfg = *cfg;
284 struct config_llog_data *cld;
286 lcfg.cfg_instance = sb;
288 cld = do_config_log_add(obd, PARAMS_FILENAME, CONFIG_T_PARAMS,
294 /** Add this log to the list of active logs watched by an MGC.
295 * Active means we're watching for updates.
296 * We have one active log per "mount" - client instance or servername.
297 * Each instance may be at a different point in the log.
299 static int config_log_add(struct obd_device *obd, char *logname,
300 struct config_llog_instance *cfg,
301 struct super_block *sb)
303 struct lustre_sb_info *lsi = s2lsi(sb);
304 struct config_llog_data *cld;
305 struct config_llog_data *sptlrpc_cld;
306 struct config_llog_data *params_cld;
311 CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
314 * for each regular log, the depended sptlrpc log name is
315 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
317 ptr = strrchr(logname, '-');
318 if (ptr == NULL || ptr - logname > 8) {
319 CERROR("logname %s is too long\n", logname);
323 memcpy(seclogname, logname, ptr - logname);
324 strcpy(seclogname + (ptr - logname), "-sptlrpc");
326 sptlrpc_cld = config_log_find(seclogname, NULL);
327 if (sptlrpc_cld == NULL) {
328 sptlrpc_cld = do_config_log_add(obd, seclogname,
329 CONFIG_T_SPTLRPC, NULL, NULL);
330 if (IS_ERR(sptlrpc_cld)) {
331 CERROR("can't create sptlrpc log: %s\n", seclogname);
332 rc = PTR_ERR(sptlrpc_cld);
336 params_cld = config_params_log_add(obd, cfg, sb);
337 if (IS_ERR(params_cld)) {
338 rc = PTR_ERR(params_cld);
339 CERROR("%s: can't create params log: rc = %d\n",
344 cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
346 CERROR("can't create log: %s\n", logname);
351 cld->cld_sptlrpc = sptlrpc_cld;
352 cld->cld_params = params_cld;
354 LASSERT(lsi->lsi_lmd);
355 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
356 struct config_llog_data *recover_cld;
357 *strrchr(seclogname, '-') = 0;
358 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
359 if (IS_ERR(recover_cld)) {
360 rc = PTR_ERR(recover_cld);
363 cld->cld_recover = recover_cld;
372 config_log_put(params_cld);
375 config_log_put(sptlrpc_cld);
381 DEFINE_MUTEX(llog_process_lock);
383 /** Stop watching for updates on this log.
385 static int config_log_end(char *logname, struct config_llog_instance *cfg)
387 struct config_llog_data *cld;
388 struct config_llog_data *cld_sptlrpc = NULL;
389 struct config_llog_data *cld_params = NULL;
390 struct config_llog_data *cld_recover = NULL;
393 cld = config_log_find(logname, cfg);
397 mutex_lock(&cld->cld_lock);
399 * if cld_stopping is set, it means we didn't start the log thus
400 * not owning the start ref. this can happen after previous umount:
401 * the cld still hanging there waiting for lock cancel, and we
402 * remount again but failed in the middle and call log_end without
405 if (unlikely(cld->cld_stopping)) {
406 mutex_unlock(&cld->cld_lock);
407 /* drop the ref from the find */
412 cld->cld_stopping = 1;
414 cld_recover = cld->cld_recover;
415 cld->cld_recover = NULL;
416 mutex_unlock(&cld->cld_lock);
419 mutex_lock(&cld_recover->cld_lock);
420 cld_recover->cld_stopping = 1;
421 mutex_unlock(&cld_recover->cld_lock);
422 config_log_put(cld_recover);
425 spin_lock(&config_list_lock);
426 cld_sptlrpc = cld->cld_sptlrpc;
427 cld->cld_sptlrpc = NULL;
428 cld_params = cld->cld_params;
429 cld->cld_params = NULL;
430 spin_unlock(&config_list_lock);
433 config_log_put(cld_sptlrpc);
436 mutex_lock(&cld_params->cld_lock);
437 cld_params->cld_stopping = 1;
438 mutex_unlock(&cld_params->cld_lock);
439 config_log_put(cld_params);
442 /* drop the ref from the find */
444 /* drop the start ref */
447 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
452 #if defined (CONFIG_PROC_FS)
453 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
455 struct obd_device *obd = data;
456 struct obd_import *imp = obd->u.cli.cl_import;
457 struct obd_connect_data *ocd = &imp->imp_connect_data;
458 struct config_llog_data *cld;
460 seq_printf(m, "imperative_recovery: %s\n",
461 OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
462 seq_printf(m, "client_state:\n");
464 spin_lock(&config_list_lock);
465 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
466 if (cld->cld_recover == NULL)
468 seq_printf(m, " - { client: %s, nidtbl_version: %u }\n",
470 cld->cld_recover->cld_cfg.cfg_last_idx);
472 spin_unlock(&config_list_lock);
478 /* reenqueue any lost locks */
479 #define RQ_RUNNING 0x1
483 static int rq_state = 0;
484 static wait_queue_head_t rq_waitq;
485 static DECLARE_COMPLETION(rq_exit);
487 static void do_requeue(struct config_llog_data *cld)
489 LASSERT(atomic_read(&cld->cld_refcount) > 0);
491 /* Do not run mgc_process_log on a disconnected export or an
492 export which is being disconnected. Take the client
493 semaphore to make the check non-racy. */
494 down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
495 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
496 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
497 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
499 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
502 up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
505 /* this timeout represents how many seconds MGC should wait before
506 * requeue config and recover lock to the MGS. We need to randomize this
507 * in order to not flood the MGS.
509 #define MGC_TIMEOUT_MIN_SECONDS 5
510 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
512 static int mgc_requeue_thread(void *data)
516 CDEBUG(D_MGC, "Starting requeue thread\n");
518 /* Keep trying failed locks periodically */
519 spin_lock(&config_list_lock);
520 rq_state |= RQ_RUNNING;
522 struct l_wait_info lwi;
523 struct config_llog_data *cld, *cld_prev;
524 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
525 int stopped = !!(rq_state & RQ_STOP);
528 /* Any new or requeued lostlocks will change the state */
529 rq_state &= ~(RQ_NOW | RQ_LATER);
530 spin_unlock(&config_list_lock);
532 /* Always wait a few seconds to allow the server who
533 caused the lock revocation to finish its setup, plus some
534 random so everyone doesn't try to reconnect at once. */
535 to = MGC_TIMEOUT_MIN_SECONDS * HZ;
536 to += rand * HZ / 100; /* rand is centi-seconds */
537 lwi = LWI_TIMEOUT(to, NULL, NULL);
538 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi);
541 * iterate & processing through the list. for each cld, process
542 * its depending sptlrpc cld firstly (if any) and then itself.
544 * it's guaranteed any item in the list must have
545 * reference > 0; and if cld_lostlock is set, at
546 * least one reference is taken by the previous enqueue.
550 spin_lock(&config_list_lock);
551 list_for_each_entry(cld, &config_llog_list,
553 if (!cld->cld_lostlock)
556 spin_unlock(&config_list_lock);
558 LASSERT(atomic_read(&cld->cld_refcount) > 0);
560 /* Whether we enqueued again or not in mgc_process_log,
561 * we're done with the ref from the old enqueue */
563 config_log_put(cld_prev);
566 cld->cld_lostlock = 0;
567 if (likely(!stopped))
570 spin_lock(&config_list_lock);
572 spin_unlock(&config_list_lock);
574 config_log_put(cld_prev);
576 /* break after scanning the list so that we can drop
577 * refcount to losing lock clds */
578 if (unlikely(stopped)) {
579 spin_lock(&config_list_lock);
583 /* Wait a bit to see if anyone else needs a requeue */
584 lwi = (struct l_wait_info) { 0 };
585 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
587 spin_lock(&config_list_lock);
589 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
590 rq_state &= ~RQ_RUNNING;
591 spin_unlock(&config_list_lock);
595 CDEBUG(D_MGC, "Ending requeue thread\n");
599 /* Add a cld to the list to requeue. Start the requeue thread if needed.
600 We are responsible for dropping the config log reference from here on out. */
601 static void mgc_requeue_add(struct config_llog_data *cld)
603 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
604 cld->cld_logname, atomic_read(&cld->cld_refcount),
605 cld->cld_stopping, rq_state);
606 LASSERT(atomic_read(&cld->cld_refcount) > 0);
608 mutex_lock(&cld->cld_lock);
609 if (cld->cld_stopping || cld->cld_lostlock) {
610 mutex_unlock(&cld->cld_lock);
613 /* this refcount will be released in mgc_requeue_thread. */
615 cld->cld_lostlock = 1;
616 mutex_unlock(&cld->cld_lock);
618 /* Hold lock for rq_state */
619 spin_lock(&config_list_lock);
620 if (rq_state & RQ_STOP) {
621 spin_unlock(&config_list_lock);
622 cld->cld_lostlock = 0;
626 spin_unlock(&config_list_lock);
631 /********************** class fns **********************/
632 static int mgc_local_llog_init(const struct lu_env *env,
633 struct obd_device *obd,
634 struct obd_device *disk)
636 struct llog_ctxt *ctxt;
639 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
644 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
646 ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
652 static int mgc_local_llog_fini(const struct lu_env *env,
653 struct obd_device *obd)
655 struct llog_ctxt *ctxt;
657 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
658 llog_cleanup(env, ctxt);
663 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb)
665 struct lustre_sb_info *lsi = s2lsi(sb);
666 struct client_obd *cli = &obd->u.cli;
667 struct lu_fid rfid, fid;
668 struct dt_object *root, *dto;
673 LASSERT(lsi->lsi_dt_dev);
679 /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
680 mutex_lock(&cli->cl_mgc_mutex);
682 cfs_cleanup_group_info();
684 /* Setup the configs dir */
685 rc = lu_env_init(env, LCT_MG_THREAD);
689 fid.f_seq = FID_SEQ_LOCAL_NAME;
692 rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
697 rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
701 root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
702 &cli->cl_mgc_los->los_dev->dd_lu_dev);
703 if (unlikely(IS_ERR(root))) {
708 dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
710 S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
711 lu_object_put_nocache(env, &root->do_lu);
717 cli->cl_mgc_configs_dir = dto;
719 LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
720 rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
724 /* We take an obd ref to insure that we can't get to mgc_cleanup
725 * without calling mgc_fs_cleanup first. */
726 class_incref(obd, "mgc_fs", obd);
728 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
731 lu_object_put(env, &cli->cl_mgc_configs_dir->do_lu);
732 cli->cl_mgc_configs_dir = NULL;
736 local_oid_storage_fini(env, cli->cl_mgc_los);
737 cli->cl_mgc_los = NULL;
738 mutex_unlock(&cli->cl_mgc_mutex);
747 static int mgc_fs_cleanup(struct obd_device *obd)
750 struct client_obd *cli = &obd->u.cli;
753 LASSERT(cli->cl_mgc_los != NULL);
755 rc = lu_env_init(&env, LCT_MG_THREAD);
759 mgc_local_llog_fini(&env, obd);
761 lu_object_put_nocache(&env, &cli->cl_mgc_configs_dir->do_lu);
762 cli->cl_mgc_configs_dir = NULL;
764 local_oid_storage_fini(&env, cli->cl_mgc_los);
765 cli->cl_mgc_los = NULL;
769 class_decref(obd, "mgc_fs", obd);
770 mutex_unlock(&cli->cl_mgc_mutex);
775 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
777 struct llog_ctxt *ctxt;
780 /* setup only remote ctxt, the local disk context is switched per each
781 * filesystem during mgc_fs_setup() */
782 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
787 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
790 llog_initiator_connect(ctxt);
796 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
798 struct llog_ctxt *ctxt;
800 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
802 llog_cleanup(env, ctxt);
807 static atomic_t mgc_count = ATOMIC_INIT(0);
808 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
813 case OBD_CLEANUP_EARLY:
815 case OBD_CLEANUP_EXPORTS:
816 if (atomic_dec_and_test(&mgc_count)) {
818 /* stop requeue thread */
819 spin_lock(&config_list_lock);
820 running = rq_state & RQ_RUNNING;
823 spin_unlock(&config_list_lock);
826 wait_for_completion(&rq_exit);
829 obd_cleanup_client_import(obd);
830 rc = mgc_llog_fini(NULL, obd);
832 CERROR("failed to cleanup llogging subsystems\n");
838 static int mgc_cleanup(struct obd_device *obd)
842 /* COMPAT_146 - old config logs may have added profiles we don't
844 if (obd->obd_type->typ_refcnt <= 1)
845 /* Only for the last mgc */
846 class_del_profiles();
848 lprocfs_obd_cleanup(obd);
851 rc = client_obd_cleanup(obd);
855 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
857 struct lprocfs_static_vars lvars;
862 rc = client_obd_setup(obd, lcfg);
866 rc = mgc_llog_init(NULL, obd);
868 CERROR("failed to setup llogging subsystems\n");
872 lprocfs_mgc_init_vars(&lvars);
873 lprocfs_obd_setup(obd, lvars.obd_vars);
874 sptlrpc_lprocfs_cliobd_attach(obd);
876 if (atomic_inc_return(&mgc_count) == 1) {
878 init_waitqueue_head(&rq_waitq);
880 /* start requeue thread */
881 rc = PTR_ERR(kthread_run(mgc_requeue_thread, NULL,
883 if (IS_ERR_VALUE(rc)) {
884 CERROR("%s: Cannot start requeue thread (%d),"
885 "no more log updates!\n",
889 /* rc is the task_struct pointer of mgc_requeue_thread. */
896 client_obd_cleanup(obd);
902 /* based on ll_mdc_blocking_ast */
903 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
904 void *data, int flag)
906 struct lustre_handle lockh;
907 struct config_llog_data *cld = (struct config_llog_data *)data;
911 case LDLM_CB_BLOCKING:
912 /* mgs wants the lock, give it up... */
913 LDLM_DEBUG(lock, "MGC blocking CB");
914 ldlm_lock2handle(lock, &lockh);
915 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
917 case LDLM_CB_CANCELING:
918 /* We've given up the lock, prepare ourselves to update. */
919 LDLM_DEBUG(lock, "MGC cancel CB");
921 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
922 PLDLMRES(lock->l_resource),
923 (char *)&lock->l_resource->lr_name.name[0]);
926 CDEBUG(D_INFO, "missing data, won't requeue\n");
930 /* held at mgc_process_log(). */
931 LASSERT(atomic_read(&cld->cld_refcount) > 0);
932 /* Are we done with this log? */
933 if (cld->cld_stopping) {
934 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
939 /* Make sure not to re-enqueue when the mgc is stopping
940 (we get called from client_disconnect_export) */
941 if (!lock->l_conn_export ||
942 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
943 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
950 mgc_requeue_add(cld);
960 /* Not sure where this should go... */
961 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
962 * that we can have a chance to try the secondary MGS if any. */
963 #define MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
965 #define MGC_TARGET_REG_LIMIT 10
966 #define MGC_SEND_PARAM_LIMIT 10
968 /* Send parameter to MGS*/
969 static int mgc_set_mgs_param(struct obd_export *exp,
970 struct mgs_send_param *msp)
972 struct ptlrpc_request *req;
973 struct mgs_send_param *req_msp, *rep_msp;
976 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
977 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
982 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
984 ptlrpc_req_finished(req);
988 memcpy(req_msp, msp, sizeof(*req_msp));
989 ptlrpc_request_set_replen(req);
991 /* Limit how long we will wait for the enqueue to complete */
992 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
993 rc = ptlrpc_queue_wait(req);
995 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
996 memcpy(msp, rep_msp, sizeof(*rep_msp));
999 ptlrpc_req_finished(req);
1004 /* Take a config lock so we can get cancel notifications */
1005 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
1006 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
1007 __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
1008 void *data, __u32 lvb_len, void *lvb_swabber,
1009 struct lustre_handle *lockh)
1011 struct config_llog_data *cld = (struct config_llog_data *)data;
1012 struct ldlm_enqueue_info einfo = {
1015 .ei_cb_bl = mgc_blocking_ast,
1016 .ei_cb_cp = ldlm_completion_ast,
1018 struct ptlrpc_request *req;
1019 int short_limit = cld_is_sptlrpc(cld);
1022 CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1023 cld->cld_resid.name[0]);
1025 /* We need a callback for every lockholder, so don't try to
1026 ldlm_lock_match (see rev 1.1.2.11.2.47) */
1027 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1028 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1033 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1034 ptlrpc_request_set_replen(req);
1036 /* check if this is server or client */
1037 if (cld->cld_cfg.cfg_sb) {
1038 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1039 if (lsi && IS_SERVER(lsi))
1042 /* Limit how long we will wait for the enqueue to complete */
1043 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1044 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1045 NULL, 0, LVB_T_NONE, lockh, 0);
1046 /* A failed enqueue should still call the mgc_blocking_ast,
1047 where it will be requeued if needed ("grant failed"). */
1048 ptlrpc_req_finished(req);
1052 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
1053 __u32 mode, struct lustre_handle *lockh)
1055 ldlm_lock_decref(lockh, mode);
1060 static void mgc_notify_active(struct obd_device *unused)
1062 /* wakeup mgc_requeue_thread to requeue mgc lock */
1063 spin_lock(&config_list_lock);
1065 spin_unlock(&config_list_lock);
1068 /* TODO: Help the MGS rebuild nidtbl. -jay */
1071 /* Send target_reg message to MGS */
1072 static int mgc_target_register(struct obd_export *exp,
1073 struct mgs_target_info *mti)
1075 struct ptlrpc_request *req;
1076 struct mgs_target_info *req_mti, *rep_mti;
1079 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1080 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1085 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1087 ptlrpc_req_finished(req);
1091 memcpy(req_mti, mti, sizeof(*req_mti));
1092 ptlrpc_request_set_replen(req);
1093 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1094 /* Limit how long we will wait for the enqueue to complete */
1095 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1097 rc = ptlrpc_queue_wait(req);
1099 rep_mti = req_capsule_server_get(&req->rq_pill,
1100 &RMF_MGS_TARGET_INFO);
1101 memcpy(mti, rep_mti, sizeof(*rep_mti));
1102 CDEBUG(D_MGC, "register %s got index = %d\n",
1103 mti->mti_svname, mti->mti_stripe_index);
1105 ptlrpc_req_finished(req);
1110 int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1111 u32 keylen, void *key, u32 vallen,
1112 void *val, struct ptlrpc_request_set *set)
1116 /* Turn off initial_recov after we try all backup servers once */
1117 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1118 struct obd_import *imp = class_exp2cliimp(exp);
1120 if (vallen != sizeof(int))
1122 value = *(int *)val;
1123 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1124 imp->imp_obd->obd_name, value,
1125 imp->imp_deactive, imp->imp_invalid,
1126 imp->imp_replayable, imp->imp_obd->obd_replayable,
1127 ptlrpc_import_state_name(imp->imp_state));
1128 /* Resurrect if we previously died */
1129 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1130 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1131 ptlrpc_reconnect_import(imp);
1134 /* FIXME move this to mgc_process_config */
1135 if (KEY_IS(KEY_REGISTER_TARGET)) {
1136 struct mgs_target_info *mti;
1137 if (vallen != sizeof(struct mgs_target_info))
1139 mti = (struct mgs_target_info *)val;
1140 CDEBUG(D_MGC, "register_target %s %#x\n",
1141 mti->mti_svname, mti->mti_flags);
1142 rc = mgc_target_register(exp, mti);
1145 if (KEY_IS(KEY_SET_FS)) {
1146 struct super_block *sb = (struct super_block *)val;
1148 if (vallen != sizeof(struct super_block))
1151 rc = mgc_fs_setup(exp->exp_obd, sb);
1153 CERROR("set_fs got %d\n", rc);
1157 if (KEY_IS(KEY_CLEAR_FS)) {
1160 rc = mgc_fs_cleanup(exp->exp_obd);
1162 CERROR("clear_fs got %d\n", rc);
1166 if (KEY_IS(KEY_SET_INFO)) {
1167 struct mgs_send_param *msp;
1169 msp = (struct mgs_send_param *)val;
1170 rc = mgc_set_mgs_param(exp, msp);
1173 if (KEY_IS(KEY_MGSSEC)) {
1174 struct client_obd *cli = &exp->exp_obd->u.cli;
1175 struct sptlrpc_flavor flvr;
1178 * empty string means using current flavor, if which haven't
1179 * been set yet, set it as null.
1181 * if flavor has been set previously, check the asking flavor
1182 * must match the existing one.
1185 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1191 rc = sptlrpc_parse_flavor(val, &flvr);
1193 CERROR("invalid sptlrpc flavor %s to MGS\n",
1199 * caller already hold a mutex
1201 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1202 cli->cl_flvr_mgc = flvr;
1203 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1204 sizeof(flvr)) != 0) {
1207 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1209 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1210 "currently %s is in use\n",
1220 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1221 __u32 keylen, void *key, __u32 *vallen, void *val,
1222 struct lov_stripe_md *unused)
1226 if (KEY_IS(KEY_CONN_DATA)) {
1227 struct obd_import *imp = class_exp2cliimp(exp);
1228 struct obd_connect_data *data = val;
1230 if (*vallen == sizeof(*data)) {
1231 *data = imp->imp_connect_data;
1239 static int mgc_import_event(struct obd_device *obd,
1240 struct obd_import *imp,
1241 enum obd_import_event event)
1245 LASSERT(imp->imp_obd == obd);
1246 CDEBUG(D_MGC, "import event %#x\n", event);
1249 case IMP_EVENT_DISCON:
1250 /* MGC imports should not wait for recovery */
1251 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1252 ptlrpc_pinger_ir_down();
1254 case IMP_EVENT_INACTIVE:
1256 case IMP_EVENT_INVALIDATE: {
1257 struct ldlm_namespace *ns = obd->obd_namespace;
1258 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1261 case IMP_EVENT_ACTIVE:
1262 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1263 /* Clearing obd_no_recov allows us to continue pinging */
1264 obd->obd_no_recov = 0;
1265 mgc_notify_active(obd);
1266 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1267 ptlrpc_pinger_ir_up();
1271 case IMP_EVENT_DEACTIVATE:
1272 case IMP_EVENT_ACTIVATE:
1275 CERROR("Unknown import event %#x\n", event);
1282 CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_CACHE_SHIFT),
1283 CONFIG_READ_NRPAGES = 4
1286 static int mgc_apply_recover_logs(struct obd_device *mgc,
1287 struct config_llog_data *cld,
1289 void *data, int datalen, bool mne_swab)
1291 struct config_llog_instance *cfg = &cld->cld_cfg;
1292 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1293 struct mgs_nidtbl_entry *entry;
1294 struct lustre_cfg *lcfg;
1295 struct lustre_cfg_bufs bufs;
1296 u64 prev_version = 0;
1304 LASSERT(cfg->cfg_instance != NULL);
1305 LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1307 OBD_ALLOC(inst, PAGE_CACHE_SIZE);
1311 if (!IS_SERVER(lsi)) {
1312 pos = snprintf(inst, PAGE_CACHE_SIZE, "%p", cfg->cfg_instance);
1313 if (pos >= PAGE_CACHE_SIZE) {
1314 OBD_FREE(inst, PAGE_CACHE_SIZE);
1318 LASSERT(IS_MDT(lsi));
1319 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1322 OBD_FREE(inst, PAGE_CACHE_SIZE);
1330 bufsz = PAGE_CACHE_SIZE - pos;
1332 while (datalen > 0) {
1333 int entry_len = sizeof(*entry);
1335 struct obd_device *obd;
1342 if (datalen < sizeof(*entry))
1345 entry = (typeof(entry))(data + off);
1348 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1350 if (entry->mne_nid_count == 0) /* at least one nid entry */
1352 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1355 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1356 if (datalen < entry_len) /* must have entry_len at least */
1359 /* Keep this swab for normal mixed endian handling. LU-1644 */
1361 lustre_swab_mgs_nidtbl_entry(entry);
1362 if (entry->mne_length > PAGE_CACHE_SIZE) {
1363 CERROR("MNE too large (%u)\n", entry->mne_length);
1367 if (entry->mne_length < entry_len)
1370 off += entry->mne_length;
1371 datalen -= entry->mne_length;
1375 if (entry->mne_version > max_version) {
1376 CERROR("entry index(%lld) is over max_index(%lld)\n",
1377 entry->mne_version, max_version);
1381 if (prev_version >= entry->mne_version) {
1382 CERROR("index unsorted, prev %lld, now %lld\n",
1383 prev_version, entry->mne_version);
1386 prev_version = entry->mne_version;
1389 * Write a string with format "nid::instance" to
1390 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1393 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1394 memset(buf, 0, bufsz);
1398 /* lustre-OST0001-osc-<instance #> */
1399 strcpy(obdname, cld->cld_logname);
1400 cname = strrchr(obdname, '-');
1401 if (cname == NULL) {
1402 CERROR("mgc %s: invalid logname %s\n",
1403 mgc->obd_name, obdname);
1407 pos = cname - obdname;
1409 pos += sprintf(obdname + pos, "-%s%04x",
1410 is_ost ? "OST" : "MDT", entry->mne_index);
1412 cname = is_ost ? "osc" : "mdc",
1413 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1414 lustre_cfg_bufs_reset(&bufs, obdname);
1416 /* find the obd by obdname */
1417 obd = class_name2obd(obdname);
1419 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1420 mgc->obd_name, obdname);
1422 /* this is a safe race, when the ost is starting up...*/
1426 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1429 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1432 down_read(&obd->u.cli.cl_sem);
1433 if (obd->u.cli.cl_import == NULL) {
1434 /* client does not connect to the OST yet */
1435 up_read(&obd->u.cli.cl_sem);
1440 /* TODO: iterate all nids to find one */
1441 /* find uuid by nid */
1442 rc = client_import_find_conn(obd->u.cli.cl_import,
1444 (struct obd_uuid *)uuid);
1445 up_read(&obd->u.cli.cl_sem);
1447 CERROR("mgc: cannot find uuid by nid %s\n",
1448 libcfs_nid2str(entry->u.nids[0]));
1452 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1453 uuid, libcfs_nid2str(entry->u.nids[0]));
1455 pos += strlen(uuid);
1456 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1457 LASSERT(pos < bufsz);
1459 lustre_cfg_bufs_set_string(&bufs, 1, params);
1462 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1464 CERROR("mgc: cannot allocate memory\n");
1468 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1469 prev_version, max_version, obdname, params);
1471 rc = class_process_config(lcfg);
1472 lustre_cfg_free(lcfg);
1474 CDEBUG(D_INFO, "process config for %s error %d\n",
1477 /* continue, even one with error */
1480 OBD_FREE(inst, PAGE_CACHE_SIZE);
1485 * This function is called if this client was notified for target restarting
1486 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1488 static int mgc_process_recover_log(struct obd_device *obd,
1489 struct config_llog_data *cld)
1491 struct ptlrpc_request *req = NULL;
1492 struct config_llog_instance *cfg = &cld->cld_cfg;
1493 struct mgs_config_body *body;
1494 struct mgs_config_res *res;
1495 struct ptlrpc_bulk_desc *desc;
1496 struct page **pages;
1499 bool mne_swab = false;
1504 /* allocate buffer for bulk transfer.
1505 * if this is the first time for this mgs to read logs,
1506 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1507 * once; otherwise, it only reads increment of logs, this should be
1508 * small and CONFIG_READ_NRPAGES will be used.
1510 nrpages = CONFIG_READ_NRPAGES;
1511 if (cfg->cfg_last_idx == 0) /* the first time */
1512 nrpages = CONFIG_READ_NRPAGES_INIT;
1514 OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1515 if (pages == NULL) {
1520 for (i = 0; i < nrpages; i++) {
1521 pages[i] = alloc_page(GFP_IOFS);
1522 if (pages[i] == NULL) {
1529 LASSERT(cld_is_recover(cld));
1530 LASSERT(mutex_is_locked(&cld->cld_lock));
1531 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1532 &RQF_MGS_CONFIG_READ);
1538 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1543 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1544 LASSERT(body != NULL);
1545 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1546 if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1547 >= sizeof(body->mcb_name)) {
1551 body->mcb_offset = cfg->cfg_last_idx + 1;
1552 body->mcb_type = cld->cld_type;
1553 body->mcb_bits = PAGE_CACHE_SHIFT;
1554 body->mcb_units = nrpages;
1556 /* allocate bulk transfer descriptor */
1557 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1, BULK_PUT_SINK,
1564 for (i = 0; i < nrpages; i++)
1565 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
1567 ptlrpc_request_set_replen(req);
1568 rc = ptlrpc_queue_wait(req);
1572 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1573 if (res->mcr_size < res->mcr_offset) {
1578 /* always update the index even though it might have errors with
1579 * handling the recover logs */
1580 cfg->cfg_last_idx = res->mcr_offset;
1581 eof = res->mcr_offset == res->mcr_size;
1583 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1584 res->mcr_offset, eof == false);
1586 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1592 if (ealen > nrpages << PAGE_CACHE_SHIFT) {
1597 if (ealen == 0) { /* no logs transferred */
1603 mne_swab = !!ptlrpc_rep_need_swab(req);
1604 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1605 /* This import flag means the server did an extra swab of IR MNE
1606 * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1607 if (unlikely(req->rq_import->imp_need_mne_swab))
1608 mne_swab = !mne_swab;
1610 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1613 for (i = 0; i < nrpages && ealen > 0; i++) {
1617 ptr = kmap(pages[i]);
1618 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1619 min_t(int, ealen, PAGE_CACHE_SIZE),
1623 CWARN("Process recover log %s error %d\n",
1624 cld->cld_logname, rc2);
1628 ealen -= PAGE_CACHE_SIZE;
1633 ptlrpc_req_finished(req);
1635 if (rc == 0 && !eof)
1639 for (i = 0; i < nrpages; i++) {
1640 if (pages[i] == NULL)
1642 __free_page(pages[i]);
1644 OBD_FREE(pages, sizeof(*pages) * nrpages);
1649 static int mgc_llog_local_copy(const struct lu_env *env,
1650 struct obd_device *obd,
1651 struct llog_ctxt *rctxt,
1652 struct llog_ctxt *lctxt, char *logname)
1660 * - copy it to backup using llog_backup()
1661 * - copy remote llog to logname using llog_backup()
1662 * - if failed then move backup to logname again
1665 OBD_ALLOC(temp_log, strlen(logname) + 1);
1668 sprintf(temp_log, "%sT", logname);
1670 /* make a copy of local llog at first */
1671 rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1672 if (rc < 0 && rc != -ENOENT)
1674 /* copy remote llog to the local copy */
1675 rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1676 if (rc == -ENOENT) {
1677 /* no remote llog, delete local one too */
1678 llog_erase(env, lctxt, NULL, logname);
1679 } else if (rc < 0) {
1680 /* error during backup, get local one back from the copy */
1681 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1683 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1684 obd->obd_name, logname, rc);
1686 llog_erase(env, lctxt, NULL, temp_log);
1687 OBD_FREE(temp_log, strlen(logname) + 1);
1691 /* local_only means it cannot get remote llogs */
1692 static int mgc_process_cfg_log(struct obd_device *mgc,
1693 struct config_llog_data *cld, int local_only)
1695 struct llog_ctxt *ctxt, *lctxt = NULL;
1696 struct dt_object *cl_mgc_dir = mgc->u.cli.cl_mgc_configs_dir;
1697 struct lustre_sb_info *lsi = NULL;
1699 bool sptlrpc_started = false;
1703 LASSERT(mutex_is_locked(&cld->cld_lock));
1706 * local copy of sptlrpc log is controlled elsewhere, don't try to
1709 if (cld_is_sptlrpc(cld) && local_only)
1712 if (cld->cld_cfg.cfg_sb)
1713 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1719 rc = lu_env_init(env, LCT_MG_THREAD);
1723 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1726 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1728 /* Copy the setup log locally if we can. Don't mess around if we're
1729 * running an MGS though (logs are already local). */
1730 if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1731 cl_mgc_dir != NULL &&
1732 lu2dt_dev(cl_mgc_dir->do_lu.lo_dev) == lsi->lsi_dt_dev) {
1734 /* Only try to copy log if we have the lock. */
1735 rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1737 if (local_only || rc) {
1738 if (llog_is_empty(env, lctxt, cld->cld_logname)) {
1739 LCONSOLE_ERROR_MSG(0x13a,
1740 "Failed to get MGS log %s and no local copy.\n",
1746 "Failed to get MGS log %s, using local copy for now, will try to update later.\n",
1749 /* Now, whether we copied or not, start using the local llog.
1750 * If we failed to copy, we'll start using whatever the old
1752 llog_ctxt_put(ctxt);
1756 if (local_only) /* no local log at client side */ {
1762 if (cld_is_sptlrpc(cld)) {
1763 sptlrpc_conf_log_update_begin(cld->cld_logname);
1764 sptlrpc_started = true;
1767 /* logname and instance info should be the same, so use our
1768 * copy of the instance for the update. The cfg_last_idx will
1769 * be updated here. */
1770 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1774 __llog_ctxt_put(env, ctxt);
1776 __llog_ctxt_put(env, lctxt);
1779 * update settings on existing OBDs. doing it inside
1780 * of llog_process_lock so no device is attaching/detaching
1782 * the logname must be <fsname>-sptlrpc
1784 if (sptlrpc_started) {
1785 LASSERT(cld_is_sptlrpc(cld));
1786 sptlrpc_conf_log_update_end(cld->cld_logname);
1787 class_notify_sptlrpc_conf(cld->cld_logname,
1788 strlen(cld->cld_logname) -
1789 strlen("-sptlrpc"));
1798 /** Get a config log from the MGS and process it.
1799 * This func is called for both clients and servers.
1800 * Copy the log locally before parsing it if appropriate (non-MGS server)
1802 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1804 struct lustre_handle lockh = { 0 };
1805 __u64 flags = LDLM_FL_NO_LRU;
1810 /* I don't want multiple processes running process_log at once --
1811 sounds like badness. It actually might be fine, as long as
1812 we're not trying to update from the same log
1813 simultaneously (in which case we should use a per-log sem.) */
1814 mutex_lock(&cld->cld_lock);
1815 if (cld->cld_stopping) {
1816 mutex_unlock(&cld->cld_lock);
1820 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1822 CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1823 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1825 /* Get the cfg lock on the llog */
1826 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1827 LCK_CR, &flags, NULL, NULL, NULL,
1828 cld, 0, NULL, &lockh);
1830 /* Get the cld, it will be released in mgc_blocking_ast. */
1831 config_log_get(cld);
1832 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1835 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1837 /* mark cld_lostlock so that it will requeue
1838 * after MGC becomes available. */
1839 cld->cld_lostlock = 1;
1840 /* Get extra reference, it will be put in requeue thread */
1841 config_log_get(cld);
1845 if (cld_is_recover(cld)) {
1846 rc = 0; /* this is not a fatal error for recover log */
1848 rc = mgc_process_recover_log(mgc, cld);
1850 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1853 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1854 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1856 mutex_unlock(&cld->cld_lock);
1858 /* Now drop the lock so MGS can revoke it */
1860 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1863 CERROR("Can't drop cfg lock: %d\n", rcl);
1870 /** Called from lustre_process_log.
1871 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1872 * any services, and adds it to the list logs to watch (follow).
1874 static int mgc_process_config(struct obd_device *obd, u32 len, void *buf)
1876 struct lustre_cfg *lcfg = buf;
1877 struct config_llog_instance *cfg = NULL;
1881 switch (lcfg->lcfg_command) {
1882 case LCFG_LOV_ADD_OBD: {
1883 /* Overloading this cfg command: register a new target */
1884 struct mgs_target_info *mti;
1886 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1887 sizeof(struct mgs_target_info)) {
1892 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1893 CDEBUG(D_MGC, "add_target %s %#x\n",
1894 mti->mti_svname, mti->mti_flags);
1895 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1898 case LCFG_LOV_DEL_OBD:
1899 /* Unregister has no meaning at the moment. */
1900 CERROR("lov_del_obd unimplemented\n");
1903 case LCFG_SPTLRPC_CONF: {
1904 rc = sptlrpc_process_config(lcfg);
1907 case LCFG_LOG_START: {
1908 struct config_llog_data *cld;
1909 struct super_block *sb;
1911 logname = lustre_cfg_string(lcfg, 1);
1912 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1913 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1915 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1918 /* We're only called through here on the initial mount */
1919 rc = config_log_add(obd, logname, cfg, sb);
1922 cld = config_log_find(logname, cfg);
1929 /* FIXME only set this for old logs! Right now this forces
1930 us to always skip the "inside markers" check */
1931 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1933 rc = mgc_process_log(obd, cld);
1934 if (rc == 0 && cld->cld_recover != NULL) {
1935 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1936 imp_connect_data, IMP_RECOV)) {
1937 rc = mgc_process_log(obd, cld->cld_recover);
1939 struct config_llog_data *cir = cld->cld_recover;
1940 cld->cld_recover = NULL;
1941 config_log_put(cir);
1944 CERROR("Cannot process recover llog %d\n", rc);
1947 if (rc == 0 && cld->cld_params != NULL) {
1948 rc = mgc_process_log(obd, cld->cld_params);
1949 if (rc == -ENOENT) {
1951 "There is no params config file yet\n");
1954 /* params log is optional */
1957 "%s: can't process params llog: rc = %d\n",
1960 config_log_put(cld);
1964 case LCFG_LOG_END: {
1965 logname = lustre_cfg_string(lcfg, 1);
1967 if (lcfg->lcfg_bufcount >= 2)
1968 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1970 rc = config_log_end(logname, cfg);
1974 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1984 struct obd_ops mgc_obd_ops = {
1985 .o_owner = THIS_MODULE,
1986 .o_setup = mgc_setup,
1987 .o_precleanup = mgc_precleanup,
1988 .o_cleanup = mgc_cleanup,
1989 .o_add_conn = client_import_add_conn,
1990 .o_del_conn = client_import_del_conn,
1991 .o_connect = client_connect_import,
1992 .o_disconnect = client_disconnect_export,
1993 /* .o_enqueue = mgc_enqueue, */
1994 .o_cancel = mgc_cancel,
1995 /* .o_iocontrol = mgc_iocontrol, */
1996 .o_set_info_async = mgc_set_info_async,
1997 .o_get_info = mgc_get_info,
1998 .o_import_event = mgc_import_event,
1999 .o_process_config = mgc_process_config,
2002 int __init mgc_init(void)
2004 return class_register_type(&mgc_obd_ops, NULL, NULL,
2005 LUSTRE_MGC_NAME, NULL);
2008 static void /*__exit*/ mgc_exit(void)
2010 class_unregister_type(LUSTRE_MGC_NAME);
2013 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2014 MODULE_DESCRIPTION("Lustre Management Client");
2015 MODULE_LICENSE("GPL");
2017 module_init(mgc_init);
2018 module_exit(mgc_exit);