1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * lustre/mgc/mgc_request.c
41 * Author: Nathan Rutman <nathan@clusterfs.com>
45 # define EXPORT_SYMTAB
47 #define DEBUG_SUBSYSTEM S_MGC
48 #define D_MGC D_CONFIG /*|D_WARNING*/
51 # include <linux/module.h>
52 # include <linux/pagemap.h>
53 # include <linux/miscdevice.h>
54 # include <linux/init.h>
56 # include <liblustre.h>
59 #include <obd_class.h>
60 #include <lustre_dlm.h>
61 #include <lprocfs_status.h>
62 #include <lustre_log.h>
63 #include <lustre_fsfilt.h>
64 #include <lustre_disk.h>
65 #include "mgc_internal.h"
67 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
73 CERROR("name too long: %s\n", name);
77 CERROR("missing name: %s\n", name);
80 memcpy(&resname, name, len);
82 /* Always use the same endianness for the resid */
83 memset(res_id, 0, sizeof(*res_id));
84 res_id->name[0] = cpu_to_le64(resname);
85 /* XXX: unfortunately, sptlprc and config llog share one lock */
88 case CONFIG_T_SPTLRPC:
91 case CONFIG_T_RECOVER:
97 res_id->name[1] = cpu_to_le64(resname);
98 CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
99 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
103 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
105 /* fsname is at most 8 chars long, maybe contain "-".
106 * e.g. "lustre", "SUN-000" */
107 return mgc_name2resid(fsname, strlen(fsname), res_id, type);
109 EXPORT_SYMBOL(mgc_fsname2resid);
111 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
116 /* logname consists of "fsname-nodetype".
117 * e.g. "lustre-MDT0001", "SUN-000-client" */
118 name_end = strrchr(logname, '-');
120 len = name_end - logname;
121 return mgc_name2resid(logname, len, res_id, type);
124 /********************** config llog list **********************/
125 static CFS_LIST_HEAD(config_llog_list);
126 static cfs_spinlock_t config_list_lock = CFS_SPIN_LOCK_UNLOCKED;
128 /* Take a reference to a config log */
129 static int config_log_get(struct config_llog_data *cld)
132 cfs_atomic_inc(&cld->cld_refcount);
133 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
134 cfs_atomic_read(&cld->cld_refcount));
138 /* Drop a reference to a config log. When no longer referenced,
139 we can free the config log data */
140 static void config_log_put(struct config_llog_data *cld)
144 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
145 cfs_atomic_read(&cld->cld_refcount));
146 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
148 /* spinlock to make sure no item with 0 refcount in the list */
149 if (cfs_atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
150 cfs_list_del(&cld->cld_list_chain);
151 cfs_spin_unlock(&config_list_lock);
153 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
155 if (cld->cld_recover)
156 config_log_put(cld->cld_recover);
157 if (cld->cld_sptlrpc)
158 config_log_put(cld->cld_sptlrpc);
159 if (cld_is_sptlrpc(cld))
160 sptlrpc_conf_log_stop(cld->cld_logname);
162 class_export_put(cld->cld_mgcexp);
163 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
169 /* Find a config log by name */
171 struct config_llog_data *config_log_find(char *logname,
172 struct config_llog_instance *cfg)
174 struct config_llog_data *cld;
175 struct config_llog_data *found = NULL;
179 LASSERT(logname != NULL);
181 instance = cfg ? cfg->cfg_instance : NULL;
182 cfs_spin_lock(&config_list_lock);
183 cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
184 /* check if instance equals */
185 if (instance != cld->cld_cfg.cfg_instance)
188 /* instance may be NULL, should check name */
189 if (strcmp(logname, cld->cld_logname) == 0) {
195 cfs_atomic_inc(&found->cld_refcount);
196 LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0);
198 cfs_spin_unlock(&config_list_lock);
203 struct config_llog_data *do_config_log_add(struct obd_device *obd,
206 struct config_llog_instance *cfg,
207 struct super_block *sb)
209 struct config_llog_data *cld;
213 CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
214 cfg ? cfg->cfg_instance : 0);
216 OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
218 RETURN(ERR_PTR(-ENOMEM));
220 strcpy(cld->cld_logname, logname);
223 cfs_mutex_init(&cld->cld_lock);
224 cld->cld_cfg.cfg_last_idx = 0;
225 cld->cld_cfg.cfg_flags = 0;
226 cld->cld_cfg.cfg_sb = sb;
227 cld->cld_type = type;
228 cfs_atomic_set(&cld->cld_refcount, 1);
230 /* Keep the mgc around until we are done */
231 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
233 if (cld_is_sptlrpc(cld)) {
234 sptlrpc_conf_log_start(logname);
235 cld->cld_cfg.cfg_obdname = obd->obd_name;
238 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
240 cfs_spin_lock(&config_list_lock);
241 cfs_list_add(&cld->cld_list_chain, &config_llog_list);
242 cfs_spin_unlock(&config_list_lock);
249 if (cld_is_sptlrpc(cld)) {
250 rc = mgc_process_log(obd, cld);
252 CERROR("failed processing sptlrpc log: %d\n", rc);
258 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
260 struct config_llog_instance *cfg,
261 struct super_block *sb)
263 struct config_llog_instance lcfg = *cfg;
264 struct lustre_sb_info *lsi = s2lsi(sb);
265 struct config_llog_data *cld;
268 if ((lsi->lsi_flags & LSI_SERVER) && IS_OST(lsi->lsi_ldd))
271 /* we have to use different llog for clients and mdts for cmd
272 * where only clients are notified if one of cmd server restarts */
273 LASSERT(strlen(fsname) < sizeof(logname) / 2);
274 strcpy(logname, fsname);
275 if (lsi->lsi_flags & LSI_SERVER) { /* mdt */
276 LASSERT(lcfg.cfg_instance == NULL);
277 lcfg.cfg_instance = sb;
278 strcat(logname, "-mdtir");
280 LASSERT(lcfg.cfg_instance != NULL);
281 strcat(logname, "-cliir");
284 cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
289 /** Add this log to the list of active logs watched by an MGC.
290 * Active means we're watching for updates.
291 * We have one active log per "mount" - client instance or servername.
292 * Each instance may be at a different point in the log.
294 static int config_log_add(struct obd_device *obd, char *logname,
295 struct config_llog_instance *cfg,
296 struct super_block *sb)
298 struct lustre_sb_info *lsi = s2lsi(sb);
299 struct config_llog_data *cld;
300 struct config_llog_data *sptlrpc_cld;
305 CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
308 * for each regular log, the depended sptlrpc log name is
309 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
311 ptr = strrchr(logname, '-');
312 if (ptr == NULL || ptr - logname > 8) {
313 CERROR("logname %s is too long\n", logname);
317 memcpy(seclogname, logname, ptr - logname);
318 strcpy(seclogname + (ptr - logname), "-sptlrpc");
320 sptlrpc_cld = config_log_find(seclogname, NULL);
321 if (sptlrpc_cld == NULL) {
322 sptlrpc_cld = do_config_log_add(obd, seclogname,
323 CONFIG_T_SPTLRPC, NULL, NULL);
324 if (IS_ERR(sptlrpc_cld)) {
325 CERROR("can't create sptlrpc log: %s\n", seclogname);
326 RETURN(PTR_ERR(sptlrpc_cld));
330 cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
332 CERROR("can't create log: %s\n", logname);
333 config_log_put(sptlrpc_cld);
334 RETURN(PTR_ERR(cld));
337 cld->cld_sptlrpc = sptlrpc_cld;
339 LASSERT(lsi->lsi_lmd);
340 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
341 struct config_llog_data *recover_cld;
342 *strrchr(seclogname, '-') = 0;
343 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
344 if (IS_ERR(recover_cld)) {
346 RETURN(PTR_ERR(recover_cld));
348 cld->cld_recover = recover_cld;
354 CFS_DECLARE_MUTEX(llog_process_lock);
356 /** Stop watching for updates on this log.
358 static int config_log_end(char *logname, struct config_llog_instance *cfg)
360 struct config_llog_data *cld;
361 struct config_llog_data *cld_sptlrpc = NULL;
362 struct config_llog_data *cld_recover = NULL;
366 cld = config_log_find(logname, cfg);
370 cfs_mutex_lock(&cld->cld_lock);
372 * if cld_stopping is set, it means we didn't start the log thus
373 * not owning the start ref. this can happen after previous umount:
374 * the cld still hanging there waiting for lock cancel, and we
375 * remount again but failed in the middle and call log_end without
378 if (unlikely(cld->cld_stopping)) {
379 cfs_mutex_unlock(&cld->cld_lock);
380 /* drop the ref from the find */
385 cld->cld_stopping = 1;
387 cld_recover = cld->cld_recover;
388 cld->cld_recover = NULL;
389 cfs_mutex_unlock(&cld->cld_lock);
392 cfs_mutex_lock(&cld_recover->cld_lock);
393 cld_recover->cld_stopping = 1;
394 cfs_mutex_unlock(&cld_recover->cld_lock);
395 config_log_put(cld_recover);
398 cfs_spin_lock(&config_list_lock);
399 cld_sptlrpc = cld->cld_sptlrpc;
400 cld->cld_sptlrpc = NULL;
401 cfs_spin_unlock(&config_list_lock);
404 config_log_put(cld_sptlrpc);
406 /* drop the ref from the find */
408 /* drop the start ref */
411 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
416 /* reenqueue any lost locks */
417 #define RQ_RUNNING 0x1
421 static int rq_state = 0;
422 static cfs_waitq_t rq_waitq;
423 static CFS_DECLARE_COMPLETION(rq_exit);
425 static void do_requeue(struct config_llog_data *cld)
428 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
430 /* Do not run mgc_process_log on a disconnected export or an
431 export which is being disconnected. Take the client
432 semaphore to make the check non-racy. */
433 cfs_down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
434 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
435 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
436 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
438 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
441 cfs_up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
446 /* this timeout represents how many seconds MGC should wait before
447 * requeue config and recover lock to the MGS. We need to randomize this
448 * in order to not flood the MGS.
450 #define MGC_TIMEOUT_MIN_SECONDS 5
451 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
453 static int mgc_requeue_thread(void *data)
455 char name[] = "ll_cfg_requeue";
461 CDEBUG(D_MGC, "Starting requeue thread\n");
463 /* Keep trying failed locks periodically */
464 cfs_spin_lock(&config_list_lock);
465 rq_state |= RQ_RUNNING;
467 struct l_wait_info lwi;
468 struct config_llog_data *cld, *cld_prev;
469 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
470 int stopped = !!(rq_state & RQ_STOP);
473 /* Any new or requeued lostlocks will change the state */
474 rq_state &= ~(RQ_NOW | RQ_LATER);
475 cfs_spin_unlock(&config_list_lock);
477 /* Always wait a few seconds to allow the server who
478 caused the lock revocation to finish its setup, plus some
479 random so everyone doesn't try to reconnect at once. */
480 to = MGC_TIMEOUT_MIN_SECONDS * CFS_HZ;
481 to += rand * CFS_HZ / 100; /* rand is centi-seconds */
482 lwi = LWI_TIMEOUT(to, NULL, NULL);
483 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi);
486 * iterate & processing through the list. for each cld, process
487 * its depending sptlrpc cld firstly (if any) and then itself.
489 * it's guaranteed any item in the list must have
490 * reference > 0; and if cld_lostlock is set, at
491 * least one reference is taken by the previous enqueue.
495 cfs_spin_lock(&config_list_lock);
496 cfs_list_for_each_entry(cld, &config_llog_list,
498 if (!cld->cld_lostlock)
501 cfs_spin_unlock(&config_list_lock);
503 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
505 /* Whether we enqueued again or not in mgc_process_log,
506 * we're done with the ref from the old enqueue */
508 config_log_put(cld_prev);
511 cld->cld_lostlock = 0;
512 if (likely(!stopped))
515 cfs_spin_lock(&config_list_lock);
517 cfs_spin_unlock(&config_list_lock);
519 config_log_put(cld_prev);
521 /* break after scanning the list so that we can drop
522 * refcount to losing lock clds */
523 if (unlikely(stopped)) {
524 cfs_spin_lock(&config_list_lock);
528 /* Wait a bit to see if anyone else needs a requeue */
529 lwi = (struct l_wait_info) { 0 };
530 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
532 cfs_spin_lock(&config_list_lock);
534 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
535 rq_state &= ~RQ_RUNNING;
536 cfs_spin_unlock(&config_list_lock);
538 cfs_complete(&rq_exit);
540 CDEBUG(D_MGC, "Ending requeue thread\n");
544 /* Add a cld to the list to requeue. Start the requeue thread if needed.
545 We are responsible for dropping the config log reference from here on out. */
546 static void mgc_requeue_add(struct config_llog_data *cld)
550 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
551 cld->cld_logname, cfs_atomic_read(&cld->cld_refcount),
552 cld->cld_stopping, rq_state);
553 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
555 cfs_mutex_lock(&cld->cld_lock);
556 if (cld->cld_stopping || cld->cld_lostlock) {
557 cfs_mutex_unlock(&cld->cld_lock);
560 /* this refcount will be released in mgc_requeue_thread. */
562 cld->cld_lostlock = 1;
563 cfs_mutex_unlock(&cld->cld_lock);
565 /* Hold lock for rq_state */
566 cfs_spin_lock(&config_list_lock);
567 if (rq_state & RQ_STOP) {
568 cfs_spin_unlock(&config_list_lock);
569 cld->cld_lostlock = 0;
573 cfs_spin_unlock(&config_list_lock);
574 cfs_waitq_signal(&rq_waitq);
579 /********************** class fns **********************/
581 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
582 struct vfsmount *mnt)
584 struct lvfs_run_ctxt saved;
585 struct lustre_sb_info *lsi = s2lsi(sb);
586 struct client_obd *cli = &obd->u.cli;
587 struct dentry *dentry;
593 LASSERT(lsi->lsi_srv_mnt == mnt);
595 /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
596 cfs_down(&cli->cl_mgc_sem);
598 cfs_cleanup_group_info();
600 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
601 if (IS_ERR(obd->obd_fsops)) {
602 cfs_up(&cli->cl_mgc_sem);
603 CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd),
604 PTR_ERR(obd->obd_fsops));
605 RETURN(PTR_ERR(obd->obd_fsops));
608 cli->cl_mgc_vfsmnt = mnt;
609 fsfilt_setup(obd, mnt->mnt_sb);
611 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
612 obd->obd_lvfs_ctxt.pwdmnt = mnt;
613 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
614 obd->obd_lvfs_ctxt.fs = get_ds();
616 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
617 dentry = ll_lookup_one_len(MOUNT_CONFIGS_DIR, cfs_fs_pwd(current->fs),
618 strlen(MOUNT_CONFIGS_DIR));
619 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
620 if (IS_ERR(dentry)) {
621 err = PTR_ERR(dentry);
622 CERROR("cannot lookup %s directory: rc = %d\n",
623 MOUNT_CONFIGS_DIR, err);
626 cli->cl_mgc_configs_dir = dentry;
628 /* We take an obd ref to insure that we can't get to mgc_cleanup
629 without calling mgc_fs_cleanup first. */
630 class_incref(obd, "mgc_fs", obd);
632 label = fsfilt_get_label(obd, mnt->mnt_sb);
634 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
636 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
640 fsfilt_put_ops(obd->obd_fsops);
641 obd->obd_fsops = NULL;
642 cli->cl_mgc_vfsmnt = NULL;
643 cfs_up(&cli->cl_mgc_sem);
647 static int mgc_fs_cleanup(struct obd_device *obd)
649 struct client_obd *cli = &obd->u.cli;
653 LASSERT(cli->cl_mgc_vfsmnt != NULL);
655 if (cli->cl_mgc_configs_dir != NULL) {
656 struct lvfs_run_ctxt saved;
657 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
658 l_dput(cli->cl_mgc_configs_dir);
659 cli->cl_mgc_configs_dir = NULL;
660 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
661 class_decref(obd, "mgc_fs", obd);
664 cli->cl_mgc_vfsmnt = NULL;
666 fsfilt_put_ops(obd->obd_fsops);
668 cfs_up(&cli->cl_mgc_sem);
673 static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0);
674 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
680 case OBD_CLEANUP_EARLY:
682 case OBD_CLEANUP_EXPORTS:
683 if (cfs_atomic_dec_and_test(&mgc_count)) {
685 /* stop requeue thread */
686 cfs_spin_lock(&config_list_lock);
687 running = rq_state & RQ_RUNNING;
690 cfs_spin_unlock(&config_list_lock);
692 cfs_waitq_signal(&rq_waitq);
693 cfs_wait_for_completion(&rq_exit);
696 obd_cleanup_client_import(obd);
697 rc = obd_llog_finish(obd, 0);
699 CERROR("failed to cleanup llogging subsystems\n");
705 static int mgc_cleanup(struct obd_device *obd)
707 struct client_obd *cli = &obd->u.cli;
711 LASSERT(cli->cl_mgc_vfsmnt == NULL);
713 /* COMPAT_146 - old config logs may have added profiles we don't
715 if (obd->obd_type->typ_refcnt <= 1)
716 /* Only for the last mgc */
717 class_del_profiles();
719 lprocfs_obd_cleanup(obd);
722 rc = client_obd_cleanup(obd);
726 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
728 struct lprocfs_static_vars lvars;
734 rc = client_obd_setup(obd, lcfg);
736 GOTO(err_decref, rc);
738 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
740 CERROR("failed to setup llogging subsystems\n");
741 GOTO(err_cleanup, rc);
744 lprocfs_mgc_init_vars(&lvars);
745 lprocfs_obd_setup(obd, lvars.obd_vars);
746 sptlrpc_lprocfs_cliobd_attach(obd);
748 if (cfs_atomic_inc_return(&mgc_count) == 1) {
750 cfs_waitq_init(&rq_waitq);
752 /* start requeue thread */
753 rc = cfs_create_thread(mgc_requeue_thread, NULL,
756 CERROR("%s: Cannot start requeue thread (%d),"
757 "no more log updates!\n",
759 GOTO(err_cleanup, rc);
761 /* rc is the pid of mgc_requeue_thread. */
768 client_obd_cleanup(obd);
774 /* based on ll_mdc_blocking_ast */
775 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
776 void *data, int flag)
778 struct lustre_handle lockh;
779 struct config_llog_data *cld = (struct config_llog_data *)data;
784 case LDLM_CB_BLOCKING:
785 /* mgs wants the lock, give it up... */
786 LDLM_DEBUG(lock, "MGC blocking CB");
787 ldlm_lock2handle(lock, &lockh);
788 rc = ldlm_cli_cancel(&lockh);
790 case LDLM_CB_CANCELING:
791 /* We've given up the lock, prepare ourselves to update. */
792 LDLM_DEBUG(lock, "MGC cancel CB");
794 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
795 lock->l_resource->lr_name.name[0],
796 (char *)&lock->l_resource->lr_name.name[0]);
799 CDEBUG(D_INFO, "missing data, won't requeue\n");
803 /* held at mgc_process_log(). */
804 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
805 /* Are we done with this log? */
806 if (cld->cld_stopping) {
807 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
812 /* Make sure not to re-enqueue when the mgc is stopping
813 (we get called from client_disconnect_export) */
814 if (!lock->l_conn_export ||
815 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
816 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
823 mgc_requeue_add(cld);
833 /* Not sure where this should go... */
834 #define MGC_ENQUEUE_LIMIT 50
835 #define MGC_TARGET_REG_LIMIT 10
836 #define MGC_SEND_PARAM_LIMIT 10
838 /* Send parameter to MGS*/
839 static int mgc_set_mgs_param(struct obd_export *exp,
840 struct mgs_send_param *msp)
842 struct ptlrpc_request *req;
843 struct mgs_send_param *req_msp, *rep_msp;
847 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
848 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
853 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
855 ptlrpc_req_finished(req);
859 memcpy(req_msp, msp, sizeof(*req_msp));
860 ptlrpc_request_set_replen(req);
862 /* Limit how long we will wait for the enqueue to complete */
863 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
864 rc = ptlrpc_queue_wait(req);
866 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
867 memcpy(msp, rep_msp, sizeof(*rep_msp));
870 ptlrpc_req_finished(req);
875 /* Take a config lock so we can get cancel notifications */
876 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
877 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
878 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
879 void *data, __u32 lvb_len, void *lvb_swabber,
880 struct lustre_handle *lockh)
882 struct config_llog_data *cld = (struct config_llog_data *)data;
883 struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
884 ldlm_completion_ast, NULL, NULL, NULL };
885 struct ptlrpc_request *req;
886 int short_limit = cld_is_sptlrpc(cld);
890 CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
891 cld->cld_resid.name[0]);
893 /* We need a callback for every lockholder, so don't try to
894 ldlm_lock_match (see rev 1.1.2.11.2.47) */
895 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
896 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
900 ptlrpc_request_set_replen(req);
902 /* check if this is server or client */
903 if (cld->cld_cfg.cfg_sb) {
904 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
905 if (lsi && (lsi->lsi_flags & LSI_SERVER))
908 /* Limit how long we will wait for the enqueue to complete */
909 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
910 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
912 /* A failed enqueue should still call the mgc_blocking_ast,
913 where it will be requeued if needed ("grant failed"). */
914 ptlrpc_req_finished(req);
918 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
919 __u32 mode, struct lustre_handle *lockh)
923 ldlm_lock_decref(lockh, mode);
928 static void mgc_notify_active(struct obd_device *unused)
930 /* wakeup mgc_requeue_thread to requeue mgc lock */
931 cfs_spin_lock(&config_list_lock);
933 cfs_spin_unlock(&config_list_lock);
934 cfs_waitq_signal(&rq_waitq);
936 /* TODO: Help the MGS rebuild nidtbl. -jay */
939 /* Send target_reg message to MGS */
940 static int mgc_target_register(struct obd_export *exp,
941 struct mgs_target_info *mti)
943 struct ptlrpc_request *req;
944 struct mgs_target_info *req_mti, *rep_mti;
948 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
949 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
954 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
956 ptlrpc_req_finished(req);
960 memcpy(req_mti, mti, sizeof(*req_mti));
961 ptlrpc_request_set_replen(req);
962 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
963 /* Limit how long we will wait for the enqueue to complete */
964 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
966 rc = ptlrpc_queue_wait(req);
968 rep_mti = req_capsule_server_get(&req->rq_pill,
969 &RMF_MGS_TARGET_INFO);
970 memcpy(mti, rep_mti, sizeof(*rep_mti));
971 CDEBUG(D_MGC, "register %s got index = %d\n",
972 mti->mti_svname, mti->mti_stripe_index);
974 ptlrpc_req_finished(req);
979 int mgc_set_info_async(struct obd_export *exp, obd_count keylen,
980 void *key, obd_count vallen, void *val,
981 struct ptlrpc_request_set *set)
986 /* Turn off initial_recov after we try all backup servers once */
987 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
988 struct obd_import *imp = class_exp2cliimp(exp);
990 if (vallen != sizeof(int))
993 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
994 imp->imp_obd->obd_name, value,
995 imp->imp_deactive, imp->imp_invalid,
996 imp->imp_replayable, imp->imp_obd->obd_replayable,
997 ptlrpc_import_state_name(imp->imp_state));
998 /* Resurrect if we previously died */
999 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1000 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1001 ptlrpc_reconnect_import(imp);
1004 /* FIXME move this to mgc_process_config */
1005 if (KEY_IS(KEY_REGISTER_TARGET)) {
1006 struct mgs_target_info *mti;
1007 if (vallen != sizeof(struct mgs_target_info))
1009 mti = (struct mgs_target_info *)val;
1010 CDEBUG(D_MGC, "register_target %s %#x\n",
1011 mti->mti_svname, mti->mti_flags);
1012 rc = mgc_target_register(exp, mti);
1015 if (KEY_IS(KEY_SET_FS)) {
1016 struct super_block *sb = (struct super_block *)val;
1017 struct lustre_sb_info *lsi;
1018 if (vallen != sizeof(struct super_block))
1021 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
1023 CERROR("set_fs got %d\n", rc);
1027 if (KEY_IS(KEY_CLEAR_FS)) {
1030 rc = mgc_fs_cleanup(exp->exp_obd);
1032 CERROR("clear_fs got %d\n", rc);
1036 if (KEY_IS(KEY_SET_INFO)) {
1037 struct mgs_send_param *msp;
1039 msp = (struct mgs_send_param *)val;
1040 rc = mgc_set_mgs_param(exp, msp);
1043 if (KEY_IS(KEY_MGSSEC)) {
1044 struct client_obd *cli = &exp->exp_obd->u.cli;
1045 struct sptlrpc_flavor flvr;
1048 * empty string means using current flavor, if which haven't
1049 * been set yet, set it as null.
1051 * if flavor has been set previously, check the asking flavor
1052 * must match the existing one.
1055 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1061 rc = sptlrpc_parse_flavor(val, &flvr);
1063 CERROR("invalid sptlrpc flavor %s to MGS\n",
1069 * caller already hold a mutex
1071 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1072 cli->cl_flvr_mgc = flvr;
1073 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1074 sizeof(flvr)) != 0) {
1077 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1079 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1080 "currently %s is in use\n",
1090 static int mgc_get_info(struct obd_export *exp, __u32 keylen, void *key,
1091 __u32 *vallen, void *val, struct lov_stripe_md *unused)
1095 if (KEY_IS(KEY_CONN_DATA)) {
1096 struct obd_import *imp = class_exp2cliimp(exp);
1097 struct obd_connect_data *data = val;
1099 if (*vallen == sizeof(*data)) {
1100 *data = imp->imp_connect_data;
1108 static int mgc_import_event(struct obd_device *obd,
1109 struct obd_import *imp,
1110 enum obd_import_event event)
1114 LASSERT(imp->imp_obd == obd);
1115 CDEBUG(D_MGC, "import event %#x\n", event);
1118 case IMP_EVENT_DISCON:
1119 /* MGC imports should not wait for recovery */
1121 case IMP_EVENT_INACTIVE:
1123 case IMP_EVENT_INVALIDATE: {
1124 struct ldlm_namespace *ns = obd->obd_namespace;
1125 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1128 case IMP_EVENT_ACTIVE:
1129 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
1130 /* Clearing obd_no_recov allows us to continue pinging */
1131 obd->obd_no_recov = 0;
1132 mgc_notify_active(obd);
1136 case IMP_EVENT_DEACTIVATE:
1137 case IMP_EVENT_ACTIVATE:
1140 CERROR("Unknown import event %#x\n", event);
1146 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1147 struct obd_device *tgt, int *index)
1149 struct llog_ctxt *ctxt;
1153 LASSERT(olg == &obd->obd_olg);
1155 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
1160 rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
1163 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1165 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1170 llog_initiator_connect(ctxt);
1171 llog_ctxt_put(ctxt);
1173 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1181 static int mgc_llog_finish(struct obd_device *obd, int count)
1183 struct llog_ctxt *ctxt;
1184 int rc = 0, rc2 = 0;
1187 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1189 rc = llog_cleanup(ctxt);
1191 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1193 rc2 = llog_cleanup(ctxt);
1202 CONFIG_READ_NRPAGES_INIT = 1 << (20 - CFS_PAGE_SHIFT),
1203 CONFIG_READ_NRPAGES = 4
1206 static int mgc_apply_recover_logs(struct obd_device *obd,
1207 struct config_llog_data *cld,
1209 void *data, int datalen)
1211 struct config_llog_instance *cfg = &cld->cld_cfg;
1212 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1213 struct mgs_nidtbl_entry *entry;
1214 struct lustre_cfg *lcfg;
1215 struct lustre_cfg_bufs bufs;
1216 u64 prev_version = 0;
1219 int bufsz = CFS_PAGE_SIZE;
1224 OBD_ALLOC(buf, CFS_PAGE_SIZE);
1228 LASSERT(cfg->cfg_instance != NULL);
1229 LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1231 if (!(lsi->lsi_flags & LSI_SERVER)) {
1232 pos = sprintf(inst, "%p", cfg->cfg_instance);
1234 LASSERT(IS_MDT(lsi->lsi_ldd));
1235 pos = sprintf(inst, "MDT%04x", lsi->lsi_ldd->ldd_svindex);
1240 while (datalen > 0) {
1241 int entry_len = sizeof(*entry);
1243 struct obd_device *obd;
1250 if (datalen < sizeof(*entry))
1253 entry = (typeof(entry))(data + off);
1256 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1258 if (entry->mne_nid_count == 0) /* at least one nid entry */
1260 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1263 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1264 if (datalen < entry_len) /* must have entry_len at least */
1267 lustre_swab_mgs_nidtbl_entry(entry);
1268 LASSERT(entry->mne_length <= CFS_PAGE_SIZE);
1269 if (entry->mne_length < entry_len)
1272 off += entry->mne_length;
1273 datalen -= entry->mne_length;
1277 if (entry->mne_version > max_version) {
1278 CERROR("entry index(%lld) is over max_index(%lld)\n",
1279 entry->mne_version, max_version);
1283 if (prev_version >= entry->mne_version) {
1284 CERROR("index unsorted, prev %lld, now %lld\n",
1285 prev_version, entry->mne_version);
1288 prev_version = entry->mne_version;
1291 * Write a string with format "nid::instance" to
1292 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1295 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1296 memset(buf, 0, bufsz);
1300 /* lustre-OST0001-osc-<instance #> */
1301 strcpy(obdname, cld->cld_logname);
1302 cname = strrchr(obdname, '-');
1303 if (cname == NULL) {
1304 CERROR("mgc: invalid logname %s\n", obdname);
1308 pos = cname - obdname;
1310 pos += sprintf(obdname + pos, "-%s%04x",
1311 is_ost ? "OST" : "MDT", entry->mne_index);
1313 cname = is_ost ? "osc" : "mdc",
1314 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1315 lustre_cfg_bufs_reset(&bufs, obdname);
1317 /* find the obd by obdname */
1318 obd = class_name2obd(obdname);
1320 CDEBUG(D_INFO, "mgc: cannot find obdname %s\n",
1323 /* this is a safe race, when the ost is starting up...*/
1327 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1330 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1333 /* TODO: iterate all nids to find one */
1334 /* find uuid by nid */
1335 rc = client_import_find_conn(obd->u.cli.cl_import,
1337 (struct obd_uuid *)uuid);
1339 CERROR("mgc: cannot find uuid by nid %s\n",
1340 libcfs_nid2str(entry->u.nids[0]));
1344 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1345 uuid, libcfs_nid2str(entry->u.nids[0]));
1347 pos += strlen(uuid);
1348 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1349 LASSERT(pos < bufsz);
1351 lustre_cfg_bufs_set_string(&bufs, 1, params);
1354 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1356 CERROR("mgc: cannot allocate memory\n");
1360 CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n",
1361 prev_version, max_version, obdname, params);
1363 rc = class_process_config(lcfg);
1364 lustre_cfg_free(lcfg);
1366 CDEBUG(D_INFO, "process config for %s error %d\n",
1369 /* continue, even one with error */
1372 OBD_FREE(inst, CFS_PAGE_SIZE);
1377 * This function is called if this client was notified for target restarting
1378 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1380 static int mgc_process_recover_log(struct obd_device *obd,
1381 struct config_llog_data *cld)
1383 struct ptlrpc_request *req = NULL;
1384 struct config_llog_instance *cfg = &cld->cld_cfg;
1385 struct mgs_config_body *body;
1386 struct mgs_config_res *res;
1387 struct ptlrpc_bulk_desc *desc;
1396 /* allocate buffer for bulk transfer.
1397 * if this is the first time for this mgs to read logs,
1398 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1399 * once; otherwise, it only reads increment of logs, this should be
1400 * small and CONFIG_READ_NRPAGES will be used.
1402 nrpages = CONFIG_READ_NRPAGES;
1403 if (cfg->cfg_last_idx == 0) /* the first time */
1404 nrpages = CONFIG_READ_NRPAGES_INIT;
1406 OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1408 GOTO(out, rc = -ENOMEM);
1410 for (i = 0; i < nrpages; i++) {
1411 pages[i] = cfs_alloc_page(CFS_ALLOC_STD);
1412 if (pages[i] == NULL)
1413 GOTO(out, rc = -ENOMEM);
1417 LASSERT(cld_is_recover(cld));
1418 LASSERT(cfs_mutex_is_locked(&cld->cld_lock));
1419 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1420 &RQF_MGS_CONFIG_READ);
1422 GOTO(out, rc = -ENOMEM);
1424 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1429 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1430 LASSERT(body != NULL);
1431 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1432 strncpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name));
1433 body->mcb_offset = cfg->cfg_last_idx + 1;
1434 body->mcb_type = cld->cld_type;
1435 body->mcb_bits = CFS_PAGE_SHIFT;
1436 body->mcb_units = nrpages;
1438 /* allocate bulk transfer descriptor */
1439 desc = ptlrpc_prep_bulk_imp(req, nrpages, BULK_PUT_SINK,
1442 GOTO(out, rc = -ENOMEM);
1444 for (i = 0; i < nrpages; i++)
1445 ptlrpc_prep_bulk_page(desc, pages[i], 0, CFS_PAGE_SIZE);
1447 ptlrpc_request_set_replen(req);
1448 rc = ptlrpc_queue_wait(req);
1452 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1453 if (res->mcr_size < res->mcr_offset)
1454 GOTO(out, rc = -EINVAL);
1456 /* always update the index even though it might have errors with
1457 * handling the recover logs */
1458 cfg->cfg_last_idx = res->mcr_offset;
1459 eof = res->mcr_offset == res->mcr_size;
1461 CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n",
1462 res->mcr_offset, eof == false);
1464 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1466 GOTO(out, rc = ealen);
1468 if (ealen > nrpages << CFS_PAGE_SHIFT)
1469 GOTO(out, rc = -EINVAL);
1471 if (ealen == 0) { /* no logs transferred */
1477 for (i = 0; i < nrpages && ealen > 0; i++) {
1481 ptr = cfs_kmap(pages[i]);
1482 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1483 min_t(int, ealen, CFS_PAGE_SIZE));
1484 cfs_kunmap(pages[i]);
1486 CWARN("Process recover log %s error %d\n",
1487 cld->cld_logname, rc2);
1491 ealen -= CFS_PAGE_SIZE;
1496 ptlrpc_req_finished(req);
1498 if (rc == 0 && !eof)
1502 for (i = 0; i < nrpages; i++) {
1503 if (pages[i] == NULL)
1505 cfs_free_page(pages[i]);
1507 OBD_FREE(pages, sizeof(*pages) * nrpages);
1512 /* identical to mgs_log_is_empty */
1513 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1516 struct lvfs_run_ctxt saved;
1517 struct llog_handle *llh;
1520 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1521 rc = llog_create(ctxt, &llh, NULL, name);
1523 llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL);
1524 rc = llog_get_size(llh);
1527 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1528 /* header is record 1 */
1532 static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec,
1535 struct llog_rec_hdr local_rec = *rec;
1536 struct llog_handle *local_llh = (struct llog_handle *)data;
1537 char *cfg_buf = (char*) (rec + 1);
1538 struct lustre_cfg *lcfg;
1542 /* Append all records */
1543 local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
1544 rc = llog_write_rec(local_llh, &local_rec, NULL, 0,
1545 (void *)cfg_buf, -1);
1547 lcfg = (struct lustre_cfg *)cfg_buf;
1548 CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
1549 rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
1550 lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
1555 /* Copy a remote log locally */
1556 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1557 struct llog_ctxt *lctxt, char *logname)
1559 struct llog_handle *local_llh, *remote_llh;
1560 struct obd_uuid *uuid;
1565 /* Write new log to a temp name, then vfs_rename over logname
1566 upon successful completion. */
1568 OBD_ALLOC(temp_log, strlen(logname) + 1);
1571 sprintf(temp_log, "%sT", logname);
1573 /* Make sure there's no old temp log */
1574 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1577 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL);
1580 rc = llog_destroy(local_llh);
1581 llog_free_handle(local_llh);
1585 /* open local log */
1586 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1590 /* set the log header uuid for fun */
1591 OBD_ALLOC_PTR(uuid);
1592 obd_str2uuid(uuid, logname);
1593 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid);
1596 GOTO(out_closel, rc);
1598 /* open remote log */
1599 rc = llog_create(rctxt, &remote_llh, NULL, logname);
1601 GOTO(out_closel, rc);
1602 rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL);
1604 GOTO(out_closer, rc);
1606 /* Copy remote log */
1607 rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL);
1610 rc2 = llog_close(remote_llh);
1614 rc2 = llog_close(local_llh);
1618 /* We've copied the remote log to the local temp log, now
1619 replace the old local log with the temp log. */
1621 struct client_obd *cli = &obd->u.cli;
1623 LASSERT(cli->cl_mgc_configs_dir);
1624 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1627 CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1630 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1631 OBD_FREE(temp_log, strlen(logname) + 1);
1635 /* local_only means it cannot get remote llogs */
1636 static int mgc_process_cfg_log(struct obd_device *mgc,
1637 struct config_llog_data *cld,
1640 struct llog_ctxt *ctxt, *lctxt = NULL;
1641 struct client_obd *cli = &mgc->u.cli;
1642 struct lvfs_run_ctxt *saved_ctxt;
1643 struct lustre_sb_info *lsi = NULL;
1644 int rc = 0, must_pop = 0;
1645 bool sptlrpc_started = false;
1650 LASSERT(cfs_mutex_is_locked(&cld->cld_lock));
1653 * local copy of sptlrpc log is controlled elsewhere, don't try to
1656 if (cld_is_sptlrpc(cld) && local_only)
1659 if (cld->cld_cfg.cfg_sb)
1660 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1662 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1664 CERROR("missing llog context\n");
1668 OBD_ALLOC_PTR(saved_ctxt);
1669 if (saved_ctxt == NULL)
1672 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1674 /* Copy the setup log locally if we can. Don't mess around if we're
1675 running an MGS though (logs are already local). */
1676 if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
1677 (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1678 !IS_MGS(lsi->lsi_ldd)) {
1679 push_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1682 /* Only try to copy log if we have the lock. */
1683 rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1684 if (local_only || rc) {
1685 if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1686 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1687 "log %s and no local copy."
1688 "\n", cld->cld_logname);
1689 GOTO(out_pop, rc = -ENOTCONN);
1691 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1692 "copy for now, will try to update later.\n",
1695 /* Now, whether we copied or not, start using the local llog.
1696 If we failed to copy, we'll start using whatever the old
1698 llog_ctxt_put(ctxt);
1701 } else if (local_only) { /* no local log at client side */
1702 GOTO(out_pop, rc = -EIO);
1705 if (cld_is_sptlrpc(cld)) {
1706 sptlrpc_conf_log_update_begin(cld->cld_logname);
1707 sptlrpc_started = true;
1710 /* logname and instance info should be the same, so use our
1711 copy of the instance for the update. The cfg_last_idx will
1713 rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg);
1717 llog_ctxt_put(ctxt);
1719 llog_ctxt_put(lctxt);
1721 pop_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1723 OBD_FREE_PTR(saved_ctxt);
1725 * update settings on existing OBDs. doing it inside
1726 * of llog_process_lock so no device is attaching/detaching
1728 * the logname must be <fsname>-sptlrpc
1730 if (sptlrpc_started) {
1731 LASSERT(cld_is_sptlrpc(cld));
1732 sptlrpc_conf_log_update_end(cld->cld_logname);
1733 class_notify_sptlrpc_conf(cld->cld_logname,
1734 strlen(cld->cld_logname) -
1735 strlen("-sptlrpc"));
1741 /** Get a config log from the MGS and process it.
1742 * This func is called for both clients and servers.
1743 * Copy the log locally before parsing it if appropriate (non-MGS server)
1745 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1747 struct lustre_handle lockh = { 0 };
1748 int rc = 0, rcl, flags = 0;
1753 /* I don't want multiple processes running process_log at once --
1754 sounds like badness. It actually might be fine, as long as
1755 we're not trying to update from the same log
1756 simultaneously (in which case we should use a per-log sem.) */
1757 cfs_mutex_lock(&cld->cld_lock);
1758 if (cld->cld_stopping) {
1759 cfs_mutex_unlock(&cld->cld_lock);
1763 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1765 CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1766 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1768 /* Get the cfg lock on the llog */
1769 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1770 LCK_CR, &flags, NULL, NULL, NULL,
1771 cld, 0, NULL, &lockh);
1773 /* Get the cld, it will be released in mgc_blocking_ast. */
1774 config_log_get(cld);
1775 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1778 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1780 /* mark cld_lostlock so that it will requeue
1781 * after MGC becomes available. */
1782 cld->cld_lostlock = 1;
1783 /* Get extra reference, it will be put in requeue thread */
1784 config_log_get(cld);
1788 if (cld_is_recover(cld)) {
1789 rc = 0; /* this is not a fatal error for recover log */
1791 rc = mgc_process_recover_log(mgc, cld);
1793 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1796 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1797 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1799 cfs_mutex_unlock(&cld->cld_lock);
1801 /* Now drop the lock so MGS can revoke it */
1803 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1806 CERROR("Can't drop cfg lock: %d\n", rcl);
1813 /** Called from lustre_process_log.
1814 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1815 * any services, and adds it to the list logs to watch (follow).
1817 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1819 struct lustre_cfg *lcfg = buf;
1820 struct config_llog_instance *cfg = NULL;
1825 switch(lcfg->lcfg_command) {
1826 case LCFG_LOV_ADD_OBD: {
1827 /* Overloading this cfg command: register a new target */
1828 struct mgs_target_info *mti;
1830 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1831 sizeof(struct mgs_target_info))
1832 GOTO(out, rc = -EINVAL);
1834 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1835 CDEBUG(D_MGC, "add_target %s %#x\n",
1836 mti->mti_svname, mti->mti_flags);
1837 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1840 case LCFG_LOV_DEL_OBD:
1841 /* Unregister has no meaning at the moment. */
1842 CERROR("lov_del_obd unimplemented\n");
1845 case LCFG_SPTLRPC_CONF: {
1846 rc = sptlrpc_process_config(lcfg);
1849 case LCFG_LOG_START: {
1850 struct config_llog_data *cld;
1851 struct super_block *sb;
1853 logname = lustre_cfg_string(lcfg, 1);
1854 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1855 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1857 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1860 /* We're only called through here on the initial mount */
1861 rc = config_log_add(obd, logname, cfg, sb);
1864 cld = config_log_find(logname, cfg);
1871 /* FIXME only set this for old logs! Right now this forces
1872 us to always skip the "inside markers" check */
1873 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1875 rc = mgc_process_log(obd, cld);
1876 if (rc == 0 && cld->cld_recover) {
1877 rc = mgc_process_log(obd, cld->cld_recover);
1879 CERROR("Cannot process recover llog %d\n", rc);
1881 config_log_put(cld);
1885 case LCFG_LOG_END: {
1886 logname = lustre_cfg_string(lcfg, 1);
1888 if (lcfg->lcfg_bufcount >= 2)
1889 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1891 rc = config_log_end(logname, cfg);
1895 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1896 GOTO(out, rc = -EINVAL);
1904 struct obd_ops mgc_obd_ops = {
1905 .o_owner = THIS_MODULE,
1906 .o_setup = mgc_setup,
1907 .o_precleanup = mgc_precleanup,
1908 .o_cleanup = mgc_cleanup,
1909 .o_add_conn = client_import_add_conn,
1910 .o_del_conn = client_import_del_conn,
1911 .o_connect = client_connect_import,
1912 .o_disconnect = client_disconnect_export,
1913 //.o_enqueue = mgc_enqueue,
1914 .o_cancel = mgc_cancel,
1915 //.o_iocontrol = mgc_iocontrol,
1916 .o_set_info_async = mgc_set_info_async,
1917 .o_get_info = mgc_get_info,
1918 .o_import_event = mgc_import_event,
1919 .o_llog_init = mgc_llog_init,
1920 .o_llog_finish = mgc_llog_finish,
1921 .o_process_config = mgc_process_config,
1924 int __init mgc_init(void)
1926 return class_register_type(&mgc_obd_ops, NULL, NULL,
1927 LUSTRE_MGC_NAME, NULL);
1931 static void /*__exit*/ mgc_exit(void)
1933 class_unregister_type(LUSTRE_MGC_NAME);
1936 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1937 MODULE_DESCRIPTION("Lustre Management Client");
1938 MODULE_LICENSE("GPL");
1940 module_init(mgc_init);
1941 module_exit(mgc_exit);