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.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
38 * lustre/mgc/mgc_request.c
40 * Author: Nathan Rutman <nathan@clusterfs.com>
44 # define EXPORT_SYMTAB
46 #define DEBUG_SUBSYSTEM S_MGC
47 #define D_MGC D_CONFIG /*|D_WARNING*/
50 # include <linux/module.h>
51 # include <linux/pagemap.h>
52 # include <linux/miscdevice.h>
53 # include <linux/init.h>
55 # include <liblustre.h>
58 #include <obd_class.h>
59 #include <lustre_dlm.h>
60 #include <lprocfs_status.h>
61 #include <lustre_log.h>
62 #include <lustre_fsfilt.h>
63 #include <lustre_disk.h>
64 #include "mgc_internal.h"
66 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
72 CERROR("name too long: %s\n", name);
76 CERROR("missing name: %s\n", name);
79 memcpy(&resname, name, len);
81 /* Always use the same endianness for the resid */
82 memset(res_id, 0, sizeof(*res_id));
83 res_id->name[0] = cpu_to_le64(resname);
84 /* XXX: unfortunately, sptlprc and config llog share one lock */
87 case CONFIG_T_SPTLRPC:
90 case CONFIG_T_RECOVER:
96 res_id->name[1] = cpu_to_le64(resname);
97 CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
98 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
102 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
104 /* fsname is at most 8 chars long, maybe contain "-".
105 * e.g. "lustre", "SUN-000" */
106 return mgc_name2resid(fsname, strlen(fsname), res_id, type);
108 EXPORT_SYMBOL(mgc_fsname2resid);
110 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
115 /* logname consists of "fsname-nodetype".
116 * e.g. "lustre-MDT0001", "SUN-000-client" */
117 name_end = strrchr(logname, '-');
119 len = name_end - logname;
120 return mgc_name2resid(logname, len, res_id, type);
123 /********************** config llog list **********************/
124 static CFS_LIST_HEAD(config_llog_list);
125 static cfs_spinlock_t config_list_lock = CFS_SPIN_LOCK_UNLOCKED;
127 /* Take a reference to a config log */
128 static int config_log_get(struct config_llog_data *cld)
131 cfs_atomic_inc(&cld->cld_refcount);
132 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
133 cfs_atomic_read(&cld->cld_refcount));
137 /* Drop a reference to a config log. When no longer referenced,
138 we can free the config log data */
139 static void config_log_put(struct config_llog_data *cld)
143 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
144 cfs_atomic_read(&cld->cld_refcount));
145 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
147 /* spinlock to make sure no item with 0 refcount in the list */
148 if (cfs_atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
149 cfs_list_del(&cld->cld_list_chain);
150 cfs_spin_unlock(&config_list_lock);
152 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
154 if (cld->cld_recover)
155 config_log_put(cld->cld_recover);
156 if (cld->cld_sptlrpc)
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 cfs_spin_lock(&config_list_lock);
182 cfs_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) {
194 cfs_atomic_inc(&found->cld_refcount);
195 LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0);
197 cfs_spin_unlock(&config_list_lock);
202 struct config_llog_data *do_config_log_add(struct obd_device *obd,
205 struct config_llog_instance *cfg,
206 struct super_block *sb)
208 struct config_llog_data *cld;
212 CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
213 cfg ? cfg->cfg_instance : 0);
215 OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
217 RETURN(ERR_PTR(-ENOMEM));
219 strcpy(cld->cld_logname, logname);
222 cfs_mutex_init(&cld->cld_lock);
223 cld->cld_cfg.cfg_last_idx = 0;
224 cld->cld_cfg.cfg_flags = 0;
225 cld->cld_cfg.cfg_sb = sb;
226 cld->cld_type = type;
227 cfs_atomic_set(&cld->cld_refcount, 1);
229 /* Keep the mgc around until we are done */
230 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
232 if (cld_is_sptlrpc(cld)) {
233 sptlrpc_conf_log_start(logname);
234 cld->cld_cfg.cfg_obdname = obd->obd_name;
237 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
239 cfs_spin_lock(&config_list_lock);
240 cfs_list_add(&cld->cld_list_chain, &config_llog_list);
241 cfs_spin_unlock(&config_list_lock);
248 if (cld_is_sptlrpc(cld)) {
249 rc = mgc_process_log(obd, cld);
251 CERROR("failed processing sptlrpc log: %d\n", 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;
267 if ((lsi->lsi_flags & LSI_SERVER) && !IS_MDT(lsi->lsi_ldd))
270 /* we have to use different llog for clients and mdts for cmd
271 * where only clients are notified if one of cmd server restarts */
272 LASSERT(strlen(fsname) < sizeof(logname) / 2);
273 strcpy(logname, fsname);
274 if (lsi->lsi_flags & LSI_SERVER) { /* mdt */
275 LASSERT(lcfg.cfg_instance == NULL);
276 lcfg.cfg_instance = sb;
277 strcat(logname, "-mdtir");
279 LASSERT(lcfg.cfg_instance != NULL);
280 strcat(logname, "-cliir");
283 cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
288 /** Add this log to the list of active logs watched by an MGC.
289 * Active means we're watching for updates.
290 * We have one active log per "mount" - client instance or servername.
291 * Each instance may be at a different point in the log.
293 static int config_log_add(struct obd_device *obd, char *logname,
294 struct config_llog_instance *cfg,
295 struct super_block *sb)
297 struct lustre_sb_info *lsi = s2lsi(sb);
298 struct config_llog_data *cld;
299 struct config_llog_data *sptlrpc_cld;
304 CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
307 * for each regular log, the depended sptlrpc log name is
308 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
310 ptr = strrchr(logname, '-');
311 if (ptr == NULL || ptr - logname > 8) {
312 CERROR("logname %s is too long\n", logname);
316 memcpy(seclogname, logname, ptr - logname);
317 strcpy(seclogname + (ptr - logname), "-sptlrpc");
319 sptlrpc_cld = config_log_find(seclogname, NULL);
320 if (sptlrpc_cld == NULL) {
321 sptlrpc_cld = do_config_log_add(obd, seclogname,
322 CONFIG_T_SPTLRPC, NULL, NULL);
323 if (IS_ERR(sptlrpc_cld)) {
324 CERROR("can't create sptlrpc log: %s\n", seclogname);
325 RETURN(PTR_ERR(sptlrpc_cld));
329 cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
331 CERROR("can't create log: %s\n", logname);
332 config_log_put(sptlrpc_cld);
333 RETURN(PTR_ERR(cld));
336 cld->cld_sptlrpc = sptlrpc_cld;
338 LASSERT(lsi->lsi_lmd);
339 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
340 struct config_llog_data *recover_cld;
341 *strrchr(seclogname, '-') = 0;
342 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
343 if (IS_ERR(recover_cld)) {
345 RETURN(PTR_ERR(recover_cld));
347 cld->cld_recover = recover_cld;
353 CFS_DEFINE_MUTEX(llog_process_lock);
355 /** Stop watching for updates on this log.
357 static int config_log_end(char *logname, struct config_llog_instance *cfg)
359 struct config_llog_data *cld;
360 struct config_llog_data *cld_sptlrpc = NULL;
361 struct config_llog_data *cld_recover = NULL;
365 cld = config_log_find(logname, cfg);
369 cfs_mutex_lock(&cld->cld_lock);
371 * if cld_stopping is set, it means we didn't start the log thus
372 * not owning the start ref. this can happen after previous umount:
373 * the cld still hanging there waiting for lock cancel, and we
374 * remount again but failed in the middle and call log_end without
377 if (unlikely(cld->cld_stopping)) {
378 cfs_mutex_unlock(&cld->cld_lock);
379 /* drop the ref from the find */
384 cld->cld_stopping = 1;
386 cld_recover = cld->cld_recover;
387 cld->cld_recover = NULL;
388 cfs_mutex_unlock(&cld->cld_lock);
391 cfs_mutex_lock(&cld_recover->cld_lock);
392 cld_recover->cld_stopping = 1;
393 cfs_mutex_unlock(&cld_recover->cld_lock);
394 config_log_put(cld_recover);
397 cfs_spin_lock(&config_list_lock);
398 cld_sptlrpc = cld->cld_sptlrpc;
399 cld->cld_sptlrpc = NULL;
400 cfs_spin_unlock(&config_list_lock);
403 config_log_put(cld_sptlrpc);
405 /* drop the ref from the find */
407 /* drop the start ref */
410 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
415 int lprocfs_mgc_rd_ir_state(char *page, char **start, off_t off,
416 int count, int *eof, void *data)
418 struct obd_device *obd = data;
419 struct obd_import *imp = obd->u.cli.cl_import;
420 struct obd_connect_data *ocd = &imp->imp_connect_data;
421 struct config_llog_data *cld;
425 rc = snprintf(page, count, "imperative_recovery: %s\n",
426 OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ON" : "OFF");
427 rc += snprintf(page + rc, count - rc, "client_state:\n");
429 cfs_spin_lock(&config_list_lock);
430 cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
431 if (cld->cld_recover == NULL)
433 rc += snprintf(page + rc, count - rc,
434 " - { client: %s, nidtbl_version: %u }\n",
436 cld->cld_recover->cld_cfg.cfg_last_idx);
438 cfs_spin_unlock(&config_list_lock);
443 /* reenqueue any lost locks */
444 #define RQ_RUNNING 0x1
448 static int rq_state = 0;
449 static cfs_waitq_t rq_waitq;
450 static CFS_DECLARE_COMPLETION(rq_exit);
452 static void do_requeue(struct config_llog_data *cld)
455 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
457 /* Do not run mgc_process_log on a disconnected export or an
458 export which is being disconnected. Take the client
459 semaphore to make the check non-racy. */
460 cfs_down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
461 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
462 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
463 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
465 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
468 cfs_up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
473 /* this timeout represents how many seconds MGC should wait before
474 * requeue config and recover lock to the MGS. We need to randomize this
475 * in order to not flood the MGS.
477 #define MGC_TIMEOUT_MIN_SECONDS 5
478 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
480 static int mgc_requeue_thread(void *data)
482 char name[] = "ll_cfg_requeue";
488 CDEBUG(D_MGC, "Starting requeue thread\n");
490 /* Keep trying failed locks periodically */
491 cfs_spin_lock(&config_list_lock);
492 rq_state |= RQ_RUNNING;
494 struct l_wait_info lwi;
495 struct config_llog_data *cld, *cld_prev;
496 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
497 int stopped = !!(rq_state & RQ_STOP);
500 /* Any new or requeued lostlocks will change the state */
501 rq_state &= ~(RQ_NOW | RQ_LATER);
502 cfs_spin_unlock(&config_list_lock);
504 /* Always wait a few seconds to allow the server who
505 caused the lock revocation to finish its setup, plus some
506 random so everyone doesn't try to reconnect at once. */
507 to = MGC_TIMEOUT_MIN_SECONDS * CFS_HZ;
508 to += rand * CFS_HZ / 100; /* rand is centi-seconds */
509 lwi = LWI_TIMEOUT(to, NULL, NULL);
510 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi);
513 * iterate & processing through the list. for each cld, process
514 * its depending sptlrpc cld firstly (if any) and then itself.
516 * it's guaranteed any item in the list must have
517 * reference > 0; and if cld_lostlock is set, at
518 * least one reference is taken by the previous enqueue.
522 cfs_spin_lock(&config_list_lock);
523 cfs_list_for_each_entry(cld, &config_llog_list,
525 if (!cld->cld_lostlock)
528 cfs_spin_unlock(&config_list_lock);
530 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
532 /* Whether we enqueued again or not in mgc_process_log,
533 * we're done with the ref from the old enqueue */
535 config_log_put(cld_prev);
538 cld->cld_lostlock = 0;
539 if (likely(!stopped))
542 cfs_spin_lock(&config_list_lock);
544 cfs_spin_unlock(&config_list_lock);
546 config_log_put(cld_prev);
548 /* break after scanning the list so that we can drop
549 * refcount to losing lock clds */
550 if (unlikely(stopped)) {
551 cfs_spin_lock(&config_list_lock);
555 /* Wait a bit to see if anyone else needs a requeue */
556 lwi = (struct l_wait_info) { 0 };
557 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
559 cfs_spin_lock(&config_list_lock);
561 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
562 rq_state &= ~RQ_RUNNING;
563 cfs_spin_unlock(&config_list_lock);
565 cfs_complete(&rq_exit);
567 CDEBUG(D_MGC, "Ending requeue thread\n");
571 /* Add a cld to the list to requeue. Start the requeue thread if needed.
572 We are responsible for dropping the config log reference from here on out. */
573 static void mgc_requeue_add(struct config_llog_data *cld)
577 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
578 cld->cld_logname, cfs_atomic_read(&cld->cld_refcount),
579 cld->cld_stopping, rq_state);
580 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
582 cfs_mutex_lock(&cld->cld_lock);
583 if (cld->cld_stopping || cld->cld_lostlock) {
584 cfs_mutex_unlock(&cld->cld_lock);
587 /* this refcount will be released in mgc_requeue_thread. */
589 cld->cld_lostlock = 1;
590 cfs_mutex_unlock(&cld->cld_lock);
592 /* Hold lock for rq_state */
593 cfs_spin_lock(&config_list_lock);
594 if (rq_state & RQ_STOP) {
595 cfs_spin_unlock(&config_list_lock);
596 cld->cld_lostlock = 0;
600 cfs_spin_unlock(&config_list_lock);
601 cfs_waitq_signal(&rq_waitq);
606 /********************** class fns **********************/
608 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
609 struct vfsmount *mnt)
611 struct lvfs_run_ctxt saved;
612 struct lustre_sb_info *lsi = s2lsi(sb);
613 struct client_obd *cli = &obd->u.cli;
614 struct dentry *dentry;
620 LASSERT(lsi->lsi_srv_mnt == mnt);
622 /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
623 cfs_down(&cli->cl_mgc_sem);
625 cfs_cleanup_group_info();
627 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
628 if (IS_ERR(obd->obd_fsops)) {
629 cfs_up(&cli->cl_mgc_sem);
630 CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd),
631 PTR_ERR(obd->obd_fsops));
632 RETURN(PTR_ERR(obd->obd_fsops));
635 cli->cl_mgc_vfsmnt = mnt;
636 err = fsfilt_setup(obd, mnt->mnt_sb);
640 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
641 obd->obd_lvfs_ctxt.pwdmnt = mnt;
642 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
643 obd->obd_lvfs_ctxt.fs = get_ds();
645 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
646 dentry = ll_lookup_one_len(MOUNT_CONFIGS_DIR, cfs_fs_pwd(current->fs),
647 strlen(MOUNT_CONFIGS_DIR));
648 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
649 if (IS_ERR(dentry)) {
650 err = PTR_ERR(dentry);
651 CERROR("cannot lookup %s directory: rc = %d\n",
652 MOUNT_CONFIGS_DIR, err);
655 cli->cl_mgc_configs_dir = dentry;
657 /* We take an obd ref to insure that we can't get to mgc_cleanup
658 without calling mgc_fs_cleanup first. */
659 class_incref(obd, "mgc_fs", obd);
661 label = fsfilt_get_label(obd, mnt->mnt_sb);
663 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
665 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
669 fsfilt_put_ops(obd->obd_fsops);
670 obd->obd_fsops = NULL;
671 cli->cl_mgc_vfsmnt = NULL;
672 cfs_up(&cli->cl_mgc_sem);
676 static int mgc_fs_cleanup(struct obd_device *obd)
678 struct client_obd *cli = &obd->u.cli;
682 LASSERT(cli->cl_mgc_vfsmnt != NULL);
684 if (cli->cl_mgc_configs_dir != NULL) {
685 struct lvfs_run_ctxt saved;
686 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
687 l_dput(cli->cl_mgc_configs_dir);
688 cli->cl_mgc_configs_dir = NULL;
689 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
690 class_decref(obd, "mgc_fs", obd);
693 cli->cl_mgc_vfsmnt = NULL;
695 fsfilt_put_ops(obd->obd_fsops);
697 cfs_up(&cli->cl_mgc_sem);
702 static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0);
703 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
709 case OBD_CLEANUP_EARLY:
711 case OBD_CLEANUP_EXPORTS:
712 if (cfs_atomic_dec_and_test(&mgc_count)) {
714 /* stop requeue thread */
715 cfs_spin_lock(&config_list_lock);
716 running = rq_state & RQ_RUNNING;
719 cfs_spin_unlock(&config_list_lock);
721 cfs_waitq_signal(&rq_waitq);
722 cfs_wait_for_completion(&rq_exit);
725 obd_cleanup_client_import(obd);
726 rc = obd_llog_finish(obd, 0);
728 CERROR("failed to cleanup llogging subsystems\n");
734 static int mgc_cleanup(struct obd_device *obd)
736 struct client_obd *cli = &obd->u.cli;
740 LASSERT(cli->cl_mgc_vfsmnt == NULL);
742 /* COMPAT_146 - old config logs may have added profiles we don't
744 if (obd->obd_type->typ_refcnt <= 1)
745 /* Only for the last mgc */
746 class_del_profiles();
748 lprocfs_obd_cleanup(obd);
751 rc = client_obd_cleanup(obd);
755 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
757 struct lprocfs_static_vars lvars;
763 rc = client_obd_setup(obd, lcfg);
765 GOTO(err_decref, rc);
767 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
769 CERROR("failed to setup llogging subsystems\n");
770 GOTO(err_cleanup, rc);
773 lprocfs_mgc_init_vars(&lvars);
774 lprocfs_obd_setup(obd, lvars.obd_vars);
775 sptlrpc_lprocfs_cliobd_attach(obd);
777 if (cfs_atomic_inc_return(&mgc_count) == 1) {
779 cfs_waitq_init(&rq_waitq);
781 /* start requeue thread */
782 rc = cfs_create_thread(mgc_requeue_thread, NULL,
785 CERROR("%s: Cannot start requeue thread (%d),"
786 "no more log updates!\n",
788 GOTO(err_cleanup, rc);
790 /* rc is the pid of mgc_requeue_thread. */
797 client_obd_cleanup(obd);
803 /* based on ll_mdc_blocking_ast */
804 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
805 void *data, int flag)
807 struct lustre_handle lockh;
808 struct config_llog_data *cld = (struct config_llog_data *)data;
813 case LDLM_CB_BLOCKING:
814 /* mgs wants the lock, give it up... */
815 LDLM_DEBUG(lock, "MGC blocking CB");
816 ldlm_lock2handle(lock, &lockh);
817 rc = ldlm_cli_cancel(&lockh);
819 case LDLM_CB_CANCELING:
820 /* We've given up the lock, prepare ourselves to update. */
821 LDLM_DEBUG(lock, "MGC cancel CB");
823 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
824 lock->l_resource->lr_name.name[0],
825 (char *)&lock->l_resource->lr_name.name[0]);
828 CDEBUG(D_INFO, "missing data, won't requeue\n");
832 /* held at mgc_process_log(). */
833 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
834 /* Are we done with this log? */
835 if (cld->cld_stopping) {
836 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
841 /* Make sure not to re-enqueue when the mgc is stopping
842 (we get called from client_disconnect_export) */
843 if (!lock->l_conn_export ||
844 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
845 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
852 mgc_requeue_add(cld);
862 /* Not sure where this should go... */
863 #define MGC_ENQUEUE_LIMIT 50
864 #define MGC_TARGET_REG_LIMIT 10
865 #define MGC_SEND_PARAM_LIMIT 10
867 /* Send parameter to MGS*/
868 static int mgc_set_mgs_param(struct obd_export *exp,
869 struct mgs_send_param *msp)
871 struct ptlrpc_request *req;
872 struct mgs_send_param *req_msp, *rep_msp;
876 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
877 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
882 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
884 ptlrpc_req_finished(req);
888 memcpy(req_msp, msp, sizeof(*req_msp));
889 ptlrpc_request_set_replen(req);
891 /* Limit how long we will wait for the enqueue to complete */
892 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
893 rc = ptlrpc_queue_wait(req);
895 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
896 memcpy(msp, rep_msp, sizeof(*rep_msp));
899 ptlrpc_req_finished(req);
904 /* Take a config lock so we can get cancel notifications */
905 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
906 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
907 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
908 void *data, __u32 lvb_len, void *lvb_swabber,
909 struct lustre_handle *lockh)
911 struct config_llog_data *cld = (struct config_llog_data *)data;
912 struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
913 ldlm_completion_ast, NULL, NULL, NULL };
914 struct ptlrpc_request *req;
915 int short_limit = cld_is_sptlrpc(cld);
919 CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
920 cld->cld_resid.name[0]);
922 /* We need a callback for every lockholder, so don't try to
923 ldlm_lock_match (see rev 1.1.2.11.2.47) */
924 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
925 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
929 ptlrpc_request_set_replen(req);
931 /* check if this is server or client */
932 if (cld->cld_cfg.cfg_sb) {
933 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
934 if (lsi && (lsi->lsi_flags & LSI_SERVER))
937 /* Limit how long we will wait for the enqueue to complete */
938 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
939 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
941 /* A failed enqueue should still call the mgc_blocking_ast,
942 where it will be requeued if needed ("grant failed"). */
943 ptlrpc_req_finished(req);
947 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
948 __u32 mode, struct lustre_handle *lockh)
952 ldlm_lock_decref(lockh, mode);
957 static void mgc_notify_active(struct obd_device *unused)
959 /* wakeup mgc_requeue_thread to requeue mgc lock */
960 cfs_spin_lock(&config_list_lock);
962 cfs_spin_unlock(&config_list_lock);
963 cfs_waitq_signal(&rq_waitq);
965 /* TODO: Help the MGS rebuild nidtbl. -jay */
968 /* Send target_reg message to MGS */
969 static int mgc_target_register(struct obd_export *exp,
970 struct mgs_target_info *mti)
972 struct ptlrpc_request *req;
973 struct mgs_target_info *req_mti, *rep_mti;
977 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
978 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
983 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
985 ptlrpc_req_finished(req);
989 memcpy(req_mti, mti, sizeof(*req_mti));
990 ptlrpc_request_set_replen(req);
991 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
992 /* Limit how long we will wait for the enqueue to complete */
993 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
995 rc = ptlrpc_queue_wait(req);
997 rep_mti = req_capsule_server_get(&req->rq_pill,
998 &RMF_MGS_TARGET_INFO);
999 memcpy(mti, rep_mti, sizeof(*rep_mti));
1000 CDEBUG(D_MGC, "register %s got index = %d\n",
1001 mti->mti_svname, mti->mti_stripe_index);
1003 ptlrpc_req_finished(req);
1008 int mgc_set_info_async(struct obd_export *exp, obd_count keylen,
1009 void *key, obd_count vallen, void *val,
1010 struct ptlrpc_request_set *set)
1015 /* Turn off initial_recov after we try all backup servers once */
1016 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1017 struct obd_import *imp = class_exp2cliimp(exp);
1019 if (vallen != sizeof(int))
1021 value = *(int *)val;
1022 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1023 imp->imp_obd->obd_name, value,
1024 imp->imp_deactive, imp->imp_invalid,
1025 imp->imp_replayable, imp->imp_obd->obd_replayable,
1026 ptlrpc_import_state_name(imp->imp_state));
1027 /* Resurrect if we previously died */
1028 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1029 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1030 ptlrpc_reconnect_import(imp);
1033 /* FIXME move this to mgc_process_config */
1034 if (KEY_IS(KEY_REGISTER_TARGET)) {
1035 struct mgs_target_info *mti;
1036 if (vallen != sizeof(struct mgs_target_info))
1038 mti = (struct mgs_target_info *)val;
1039 CDEBUG(D_MGC, "register_target %s %#x\n",
1040 mti->mti_svname, mti->mti_flags);
1041 rc = mgc_target_register(exp, mti);
1044 if (KEY_IS(KEY_SET_FS)) {
1045 struct super_block *sb = (struct super_block *)val;
1046 struct lustre_sb_info *lsi;
1047 if (vallen != sizeof(struct super_block))
1050 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
1052 CERROR("set_fs got %d\n", rc);
1056 if (KEY_IS(KEY_CLEAR_FS)) {
1059 rc = mgc_fs_cleanup(exp->exp_obd);
1061 CERROR("clear_fs got %d\n", rc);
1065 if (KEY_IS(KEY_SET_INFO)) {
1066 struct mgs_send_param *msp;
1068 msp = (struct mgs_send_param *)val;
1069 rc = mgc_set_mgs_param(exp, msp);
1072 if (KEY_IS(KEY_MGSSEC)) {
1073 struct client_obd *cli = &exp->exp_obd->u.cli;
1074 struct sptlrpc_flavor flvr;
1077 * empty string means using current flavor, if which haven't
1078 * been set yet, set it as null.
1080 * if flavor has been set previously, check the asking flavor
1081 * must match the existing one.
1084 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1090 rc = sptlrpc_parse_flavor(val, &flvr);
1092 CERROR("invalid sptlrpc flavor %s to MGS\n",
1098 * caller already hold a mutex
1100 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1101 cli->cl_flvr_mgc = flvr;
1102 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1103 sizeof(flvr)) != 0) {
1106 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1108 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1109 "currently %s is in use\n",
1119 static int mgc_get_info(struct obd_export *exp, __u32 keylen, void *key,
1120 __u32 *vallen, void *val, struct lov_stripe_md *unused)
1124 if (KEY_IS(KEY_CONN_DATA)) {
1125 struct obd_import *imp = class_exp2cliimp(exp);
1126 struct obd_connect_data *data = val;
1128 if (*vallen == sizeof(*data)) {
1129 *data = imp->imp_connect_data;
1137 static int mgc_import_event(struct obd_device *obd,
1138 struct obd_import *imp,
1139 enum obd_import_event event)
1143 LASSERT(imp->imp_obd == obd);
1144 CDEBUG(D_MGC, "import event %#x\n", event);
1147 case IMP_EVENT_DISCON:
1148 /* MGC imports should not wait for recovery */
1150 case IMP_EVENT_INACTIVE:
1152 case IMP_EVENT_INVALIDATE: {
1153 struct ldlm_namespace *ns = obd->obd_namespace;
1154 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1157 case IMP_EVENT_ACTIVE:
1158 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
1159 /* Clearing obd_no_recov allows us to continue pinging */
1160 obd->obd_no_recov = 0;
1161 mgc_notify_active(obd);
1165 case IMP_EVENT_DEACTIVATE:
1166 case IMP_EVENT_ACTIVATE:
1169 CERROR("Unknown import event %#x\n", event);
1175 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1176 struct obd_device *tgt, int *index)
1178 struct llog_ctxt *ctxt;
1182 LASSERT(olg == &obd->obd_olg);
1184 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
1189 rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
1192 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1194 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1199 llog_initiator_connect(ctxt);
1200 llog_ctxt_put(ctxt);
1202 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1210 static int mgc_llog_finish(struct obd_device *obd, int count)
1212 struct llog_ctxt *ctxt;
1213 int rc = 0, rc2 = 0;
1216 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1218 rc = llog_cleanup(ctxt);
1220 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1222 rc2 = llog_cleanup(ctxt);
1231 CONFIG_READ_NRPAGES_INIT = 1 << (20 - CFS_PAGE_SHIFT),
1232 CONFIG_READ_NRPAGES = 4
1235 static int mgc_apply_recover_logs(struct obd_device *mgc,
1236 struct config_llog_data *cld,
1238 void *data, int datalen)
1240 struct config_llog_instance *cfg = &cld->cld_cfg;
1241 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1242 struct mgs_nidtbl_entry *entry;
1243 struct lustre_cfg *lcfg;
1244 struct lustre_cfg_bufs bufs;
1245 u64 prev_version = 0;
1254 LASSERT(cfg->cfg_instance != NULL);
1255 LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1257 OBD_ALLOC(inst, CFS_PAGE_SIZE);
1261 if (!(lsi->lsi_flags & LSI_SERVER)) {
1262 pos = sprintf(inst, "%p", cfg->cfg_instance);
1264 LASSERT(IS_MDT(lsi->lsi_ldd));
1265 pos = sprintf(inst, "MDT%04x", lsi->lsi_ldd->ldd_svindex);
1270 bufsz = CFS_PAGE_SIZE - pos;
1272 while (datalen > 0) {
1273 int entry_len = sizeof(*entry);
1275 struct obd_device *obd;
1282 if (datalen < sizeof(*entry))
1285 entry = (typeof(entry))(data + off);
1288 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1290 if (entry->mne_nid_count == 0) /* at least one nid entry */
1292 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1295 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1296 if (datalen < entry_len) /* must have entry_len at least */
1299 lustre_swab_mgs_nidtbl_entry(entry);
1300 LASSERT(entry->mne_length <= CFS_PAGE_SIZE);
1301 if (entry->mne_length < entry_len)
1304 off += entry->mne_length;
1305 datalen -= entry->mne_length;
1309 if (entry->mne_version > max_version) {
1310 CERROR("entry index(%lld) is over max_index(%lld)\n",
1311 entry->mne_version, max_version);
1315 if (prev_version >= entry->mne_version) {
1316 CERROR("index unsorted, prev %lld, now %lld\n",
1317 prev_version, entry->mne_version);
1320 prev_version = entry->mne_version;
1323 * Write a string with format "nid::instance" to
1324 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1327 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1328 memset(buf, 0, bufsz);
1332 /* lustre-OST0001-osc-<instance #> */
1333 strcpy(obdname, cld->cld_logname);
1334 cname = strrchr(obdname, '-');
1335 if (cname == NULL) {
1336 CERROR("mgc %s: invalid logname %s\n",
1337 mgc->obd_name, obdname);
1341 pos = cname - obdname;
1343 pos += sprintf(obdname + pos, "-%s%04x",
1344 is_ost ? "OST" : "MDT", entry->mne_index);
1346 cname = is_ost ? "osc" : "mdc",
1347 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1348 lustre_cfg_bufs_reset(&bufs, obdname);
1350 /* find the obd by obdname */
1351 obd = class_name2obd(obdname);
1353 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1354 mgc->obd_name, obdname);
1356 /* this is a safe race, when the ost is starting up...*/
1360 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1363 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1366 /* TODO: iterate all nids to find one */
1367 /* find uuid by nid */
1368 rc = client_import_find_conn(obd->u.cli.cl_import,
1370 (struct obd_uuid *)uuid);
1372 CERROR("mgc: cannot find uuid by nid %s\n",
1373 libcfs_nid2str(entry->u.nids[0]));
1377 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1378 uuid, libcfs_nid2str(entry->u.nids[0]));
1380 pos += strlen(uuid);
1381 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1382 LASSERT(pos < bufsz);
1384 lustre_cfg_bufs_set_string(&bufs, 1, params);
1387 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1389 CERROR("mgc: cannot allocate memory\n");
1393 CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n",
1394 prev_version, max_version, obdname, params);
1396 rc = class_process_config(lcfg);
1397 lustre_cfg_free(lcfg);
1399 CDEBUG(D_INFO, "process config for %s error %d\n",
1402 /* continue, even one with error */
1405 OBD_FREE(inst, CFS_PAGE_SIZE);
1410 * This function is called if this client was notified for target restarting
1411 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1413 static int mgc_process_recover_log(struct obd_device *obd,
1414 struct config_llog_data *cld)
1416 struct ptlrpc_request *req = NULL;
1417 struct config_llog_instance *cfg = &cld->cld_cfg;
1418 struct mgs_config_body *body;
1419 struct mgs_config_res *res;
1420 struct ptlrpc_bulk_desc *desc;
1429 /* allocate buffer for bulk transfer.
1430 * if this is the first time for this mgs to read logs,
1431 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1432 * once; otherwise, it only reads increment of logs, this should be
1433 * small and CONFIG_READ_NRPAGES will be used.
1435 nrpages = CONFIG_READ_NRPAGES;
1436 if (cfg->cfg_last_idx == 0) /* the first time */
1437 nrpages = CONFIG_READ_NRPAGES_INIT;
1439 OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1441 GOTO(out, rc = -ENOMEM);
1443 for (i = 0; i < nrpages; i++) {
1444 pages[i] = cfs_alloc_page(CFS_ALLOC_STD);
1445 if (pages[i] == NULL)
1446 GOTO(out, rc = -ENOMEM);
1450 LASSERT(cld_is_recover(cld));
1451 LASSERT(cfs_mutex_is_locked(&cld->cld_lock));
1452 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1453 &RQF_MGS_CONFIG_READ);
1455 GOTO(out, rc = -ENOMEM);
1457 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1462 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1463 LASSERT(body != NULL);
1464 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1465 strncpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name));
1466 body->mcb_offset = cfg->cfg_last_idx + 1;
1467 body->mcb_type = cld->cld_type;
1468 body->mcb_bits = CFS_PAGE_SHIFT;
1469 body->mcb_units = nrpages;
1471 /* allocate bulk transfer descriptor */
1472 desc = ptlrpc_prep_bulk_imp(req, nrpages, BULK_PUT_SINK,
1475 GOTO(out, rc = -ENOMEM);
1477 for (i = 0; i < nrpages; i++)
1478 ptlrpc_prep_bulk_page(desc, pages[i], 0, CFS_PAGE_SIZE);
1480 ptlrpc_request_set_replen(req);
1481 rc = ptlrpc_queue_wait(req);
1485 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1486 if (res->mcr_size < res->mcr_offset)
1487 GOTO(out, rc = -EINVAL);
1489 /* always update the index even though it might have errors with
1490 * handling the recover logs */
1491 cfg->cfg_last_idx = res->mcr_offset;
1492 eof = res->mcr_offset == res->mcr_size;
1494 CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n",
1495 res->mcr_offset, eof == false);
1497 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1499 GOTO(out, rc = ealen);
1501 if (ealen > nrpages << CFS_PAGE_SHIFT)
1502 GOTO(out, rc = -EINVAL);
1504 if (ealen == 0) { /* no logs transferred */
1510 for (i = 0; i < nrpages && ealen > 0; i++) {
1514 ptr = cfs_kmap(pages[i]);
1515 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1516 min_t(int, ealen, CFS_PAGE_SIZE));
1517 cfs_kunmap(pages[i]);
1519 CWARN("Process recover log %s error %d\n",
1520 cld->cld_logname, rc2);
1524 ealen -= CFS_PAGE_SIZE;
1529 ptlrpc_req_finished(req);
1531 if (rc == 0 && !eof)
1535 for (i = 0; i < nrpages; i++) {
1536 if (pages[i] == NULL)
1538 cfs_free_page(pages[i]);
1540 OBD_FREE(pages, sizeof(*pages) * nrpages);
1545 /* identical to mgs_log_is_empty */
1546 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1549 struct lvfs_run_ctxt saved;
1550 struct llog_handle *llh;
1553 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1554 rc = llog_create(ctxt, &llh, NULL, name);
1556 llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL);
1557 rc = llog_get_size(llh);
1560 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1561 /* header is record 1 */
1565 static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec,
1568 struct llog_rec_hdr local_rec = *rec;
1569 struct llog_handle *local_llh = (struct llog_handle *)data;
1570 char *cfg_buf = (char*) (rec + 1);
1571 struct lustre_cfg *lcfg;
1575 /* Append all records */
1576 local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
1577 rc = llog_write_rec(local_llh, &local_rec, NULL, 0,
1578 (void *)cfg_buf, -1);
1580 lcfg = (struct lustre_cfg *)cfg_buf;
1581 CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
1582 rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
1583 lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
1588 /* Copy a remote log locally */
1589 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1590 struct llog_ctxt *lctxt, char *logname)
1592 struct llog_handle *local_llh, *remote_llh;
1593 struct obd_uuid *uuid;
1598 /* Write new log to a temp name, then vfs_rename over logname
1599 upon successful completion. */
1601 OBD_ALLOC(temp_log, strlen(logname) + 1);
1604 sprintf(temp_log, "%sT", logname);
1606 /* Make sure there's no old temp log */
1607 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1610 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL);
1613 rc = llog_destroy(local_llh);
1614 llog_free_handle(local_llh);
1618 /* open local log */
1619 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1623 /* set the log header uuid for fun */
1624 OBD_ALLOC_PTR(uuid);
1625 obd_str2uuid(uuid, logname);
1626 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid);
1629 GOTO(out_closel, rc);
1631 /* open remote log */
1632 rc = llog_create(rctxt, &remote_llh, NULL, logname);
1634 GOTO(out_closel, rc);
1635 rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL);
1637 GOTO(out_closer, rc);
1639 /* Copy remote log */
1640 rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL);
1643 rc2 = llog_close(remote_llh);
1647 rc2 = llog_close(local_llh);
1651 /* We've copied the remote log to the local temp log, now
1652 replace the old local log with the temp log. */
1654 struct client_obd *cli = &obd->u.cli;
1656 LASSERT(cli->cl_mgc_configs_dir);
1657 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1660 CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1663 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1664 OBD_FREE(temp_log, strlen(logname) + 1);
1668 /* local_only means it cannot get remote llogs */
1669 static int mgc_process_cfg_log(struct obd_device *mgc,
1670 struct config_llog_data *cld,
1673 struct llog_ctxt *ctxt, *lctxt = NULL;
1674 struct client_obd *cli = &mgc->u.cli;
1675 struct lvfs_run_ctxt *saved_ctxt;
1676 struct lustre_sb_info *lsi = NULL;
1677 int rc = 0, must_pop = 0;
1678 bool sptlrpc_started = false;
1683 LASSERT(cfs_mutex_is_locked(&cld->cld_lock));
1686 * local copy of sptlrpc log is controlled elsewhere, don't try to
1689 if (cld_is_sptlrpc(cld) && local_only)
1692 if (cld->cld_cfg.cfg_sb)
1693 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1695 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1697 CERROR("missing llog context\n");
1701 OBD_ALLOC_PTR(saved_ctxt);
1702 if (saved_ctxt == NULL)
1705 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1707 /* Copy the setup log locally if we can. Don't mess around if we're
1708 running an MGS though (logs are already local). */
1709 if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
1710 (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1711 !IS_MGS(lsi->lsi_ldd)) {
1712 push_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1715 /* Only try to copy log if we have the lock. */
1716 rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1717 if (local_only || rc) {
1718 if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1719 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1720 "log %s and no local copy."
1721 "\n", cld->cld_logname);
1722 GOTO(out_pop, rc = -ENOTCONN);
1724 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1725 "copy for now, will try to update later.\n",
1728 /* Now, whether we copied or not, start using the local llog.
1729 If we failed to copy, we'll start using whatever the old
1731 llog_ctxt_put(ctxt);
1734 } else if (local_only) { /* no local log at client side */
1735 GOTO(out_pop, rc = -EIO);
1738 if (cld_is_sptlrpc(cld)) {
1739 sptlrpc_conf_log_update_begin(cld->cld_logname);
1740 sptlrpc_started = true;
1743 /* logname and instance info should be the same, so use our
1744 copy of the instance for the update. The cfg_last_idx will
1746 rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg);
1750 llog_ctxt_put(ctxt);
1752 llog_ctxt_put(lctxt);
1754 pop_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1756 OBD_FREE_PTR(saved_ctxt);
1758 * update settings on existing OBDs. doing it inside
1759 * of llog_process_lock so no device is attaching/detaching
1761 * the logname must be <fsname>-sptlrpc
1763 if (sptlrpc_started) {
1764 LASSERT(cld_is_sptlrpc(cld));
1765 sptlrpc_conf_log_update_end(cld->cld_logname);
1766 class_notify_sptlrpc_conf(cld->cld_logname,
1767 strlen(cld->cld_logname) -
1768 strlen("-sptlrpc"));
1774 /** Get a config log from the MGS and process it.
1775 * This func is called for both clients and servers.
1776 * Copy the log locally before parsing it if appropriate (non-MGS server)
1778 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1780 struct lustre_handle lockh = { 0 };
1781 int rc = 0, rcl, flags = LDLM_FL_NO_LRU;
1786 /* I don't want multiple processes running process_log at once --
1787 sounds like badness. It actually might be fine, as long as
1788 we're not trying to update from the same log
1789 simultaneously (in which case we should use a per-log sem.) */
1790 cfs_mutex_lock(&cld->cld_lock);
1791 if (cld->cld_stopping) {
1792 cfs_mutex_unlock(&cld->cld_lock);
1796 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1798 CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1799 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1801 /* Get the cfg lock on the llog */
1802 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1803 LCK_CR, &flags, NULL, NULL, NULL,
1804 cld, 0, NULL, &lockh);
1806 /* Get the cld, it will be released in mgc_blocking_ast. */
1807 config_log_get(cld);
1808 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1811 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1813 /* mark cld_lostlock so that it will requeue
1814 * after MGC becomes available. */
1815 cld->cld_lostlock = 1;
1816 /* Get extra reference, it will be put in requeue thread */
1817 config_log_get(cld);
1821 if (cld_is_recover(cld)) {
1822 rc = 0; /* this is not a fatal error for recover log */
1824 rc = mgc_process_recover_log(mgc, cld);
1826 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1829 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1830 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1832 cfs_mutex_unlock(&cld->cld_lock);
1834 /* Now drop the lock so MGS can revoke it */
1836 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1839 CERROR("Can't drop cfg lock: %d\n", rcl);
1846 /** Called from lustre_process_log.
1847 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1848 * any services, and adds it to the list logs to watch (follow).
1850 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1852 struct lustre_cfg *lcfg = buf;
1853 struct config_llog_instance *cfg = NULL;
1858 switch(lcfg->lcfg_command) {
1859 case LCFG_LOV_ADD_OBD: {
1860 /* Overloading this cfg command: register a new target */
1861 struct mgs_target_info *mti;
1863 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1864 sizeof(struct mgs_target_info))
1865 GOTO(out, rc = -EINVAL);
1867 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1868 CDEBUG(D_MGC, "add_target %s %#x\n",
1869 mti->mti_svname, mti->mti_flags);
1870 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1873 case LCFG_LOV_DEL_OBD:
1874 /* Unregister has no meaning at the moment. */
1875 CERROR("lov_del_obd unimplemented\n");
1878 case LCFG_SPTLRPC_CONF: {
1879 rc = sptlrpc_process_config(lcfg);
1882 case LCFG_LOG_START: {
1883 struct config_llog_data *cld;
1884 struct super_block *sb;
1886 logname = lustre_cfg_string(lcfg, 1);
1887 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1888 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1890 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1893 /* We're only called through here on the initial mount */
1894 rc = config_log_add(obd, logname, cfg, sb);
1897 cld = config_log_find(logname, cfg);
1904 /* FIXME only set this for old logs! Right now this forces
1905 us to always skip the "inside markers" check */
1906 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1908 rc = mgc_process_log(obd, cld);
1909 if (rc == 0 && cld->cld_recover != NULL) {
1910 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1911 imp_connect_data, IMP_RECOV)) {
1912 rc = mgc_process_log(obd, cld->cld_recover);
1914 struct config_llog_data *cir = cld->cld_recover;
1915 cld->cld_recover = NULL;
1916 config_log_put(cir);
1919 CERROR("Cannot process recover llog %d\n", rc);
1921 config_log_put(cld);
1925 case LCFG_LOG_END: {
1926 logname = lustre_cfg_string(lcfg, 1);
1928 if (lcfg->lcfg_bufcount >= 2)
1929 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1931 rc = config_log_end(logname, cfg);
1935 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1936 GOTO(out, rc = -EINVAL);
1944 struct obd_ops mgc_obd_ops = {
1945 .o_owner = THIS_MODULE,
1946 .o_setup = mgc_setup,
1947 .o_precleanup = mgc_precleanup,
1948 .o_cleanup = mgc_cleanup,
1949 .o_add_conn = client_import_add_conn,
1950 .o_del_conn = client_import_del_conn,
1951 .o_connect = client_connect_import,
1952 .o_disconnect = client_disconnect_export,
1953 //.o_enqueue = mgc_enqueue,
1954 .o_cancel = mgc_cancel,
1955 //.o_iocontrol = mgc_iocontrol,
1956 .o_set_info_async = mgc_set_info_async,
1957 .o_get_info = mgc_get_info,
1958 .o_import_event = mgc_import_event,
1959 .o_llog_init = mgc_llog_init,
1960 .o_llog_finish = mgc_llog_finish,
1961 .o_process_config = mgc_process_config,
1964 int __init mgc_init(void)
1966 return class_register_type(&mgc_obd_ops, NULL, NULL,
1967 LUSTRE_MGC_NAME, NULL);
1971 static void /*__exit*/ mgc_exit(void)
1973 class_unregister_type(LUSTRE_MGC_NAME);
1976 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1977 MODULE_DESCRIPTION("Lustre Management Client");
1978 MODULE_LICENSE("GPL");
1980 module_init(mgc_init);
1981 module_exit(mgc_exit);