4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mgc/mgc_request.c
38 * Author: Nathan Rutman <nathan@clusterfs.com>
41 #define DEBUG_SUBSYSTEM S_MGC
42 #define D_MGC D_CONFIG /*|D_WARNING*/
45 # include <linux/module.h>
46 # include <linux/pagemap.h>
47 # include <linux/miscdevice.h>
48 # include <linux/init.h>
50 # include <liblustre.h>
53 #include <obd_class.h>
54 #include <lustre_dlm.h>
55 #include <lprocfs_status.h>
56 #include <lustre_log.h>
57 #include <lustre_fsfilt.h>
58 #include <lustre_disk.h>
59 #include "mgc_internal.h"
61 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
67 CERROR("name too long: %s\n", name);
71 CERROR("missing name: %s\n", name);
74 memcpy(&resname, name, len);
76 /* Always use the same endianness for the resid */
77 memset(res_id, 0, sizeof(*res_id));
78 res_id->name[0] = cpu_to_le64(resname);
79 /* XXX: unfortunately, sptlprc and config llog share one lock */
82 case CONFIG_T_SPTLRPC:
85 case CONFIG_T_RECOVER:
91 res_id->name[1] = cpu_to_le64(resname);
92 CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
93 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
97 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
99 /* fsname is at most 8 chars long, maybe contain "-".
100 * e.g. "lustre", "SUN-000" */
101 return mgc_name2resid(fsname, strlen(fsname), res_id, type);
103 EXPORT_SYMBOL(mgc_fsname2resid);
105 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
110 /* logname consists of "fsname-nodetype".
111 * e.g. "lustre-MDT0001", "SUN-000-client" */
112 name_end = strrchr(logname, '-');
114 len = name_end - logname;
115 return mgc_name2resid(logname, len, res_id, type);
118 /********************** config llog list **********************/
119 static CFS_LIST_HEAD(config_llog_list);
120 static cfs_spinlock_t config_list_lock = CFS_SPIN_LOCK_UNLOCKED;
122 /* Take a reference to a config log */
123 static int config_log_get(struct config_llog_data *cld)
126 cfs_atomic_inc(&cld->cld_refcount);
127 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
128 cfs_atomic_read(&cld->cld_refcount));
132 /* Drop a reference to a config log. When no longer referenced,
133 we can free the config log data */
134 static void config_log_put(struct config_llog_data *cld)
138 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
139 cfs_atomic_read(&cld->cld_refcount));
140 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
142 /* spinlock to make sure no item with 0 refcount in the list */
143 if (cfs_atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
144 cfs_list_del(&cld->cld_list_chain);
145 cfs_spin_unlock(&config_list_lock);
147 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
149 if (cld->cld_recover)
150 config_log_put(cld->cld_recover);
151 if (cld->cld_sptlrpc)
152 config_log_put(cld->cld_sptlrpc);
153 if (cld_is_sptlrpc(cld))
154 sptlrpc_conf_log_stop(cld->cld_logname);
156 class_export_put(cld->cld_mgcexp);
157 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
163 /* Find a config log by name */
165 struct config_llog_data *config_log_find(char *logname,
166 struct config_llog_instance *cfg)
168 struct config_llog_data *cld;
169 struct config_llog_data *found = NULL;
173 LASSERT(logname != NULL);
175 instance = cfg ? cfg->cfg_instance : NULL;
176 cfs_spin_lock(&config_list_lock);
177 cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
178 /* check if instance equals */
179 if (instance != cld->cld_cfg.cfg_instance)
182 /* instance may be NULL, should check name */
183 if (strcmp(logname, cld->cld_logname) == 0) {
189 cfs_atomic_inc(&found->cld_refcount);
190 LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0);
192 cfs_spin_unlock(&config_list_lock);
197 struct config_llog_data *do_config_log_add(struct obd_device *obd,
200 struct config_llog_instance *cfg,
201 struct super_block *sb)
203 struct config_llog_data *cld;
207 CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
208 cfg ? cfg->cfg_instance : 0);
210 OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
212 RETURN(ERR_PTR(-ENOMEM));
214 strcpy(cld->cld_logname, logname);
217 cfs_mutex_init(&cld->cld_lock);
218 cld->cld_cfg.cfg_last_idx = 0;
219 cld->cld_cfg.cfg_flags = 0;
220 cld->cld_cfg.cfg_sb = sb;
221 cld->cld_type = type;
222 cfs_atomic_set(&cld->cld_refcount, 1);
224 /* Keep the mgc around until we are done */
225 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
227 if (cld_is_sptlrpc(cld)) {
228 sptlrpc_conf_log_start(logname);
229 cld->cld_cfg.cfg_obdname = obd->obd_name;
232 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
234 cfs_spin_lock(&config_list_lock);
235 cfs_list_add(&cld->cld_list_chain, &config_llog_list);
236 cfs_spin_unlock(&config_list_lock);
243 if (cld_is_sptlrpc(cld)) {
244 rc = mgc_process_log(obd, cld);
246 CERROR("failed processing sptlrpc log: %d\n", rc);
252 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
254 struct config_llog_instance *cfg,
255 struct super_block *sb)
257 struct config_llog_instance lcfg = *cfg;
258 struct lustre_sb_info *lsi = s2lsi(sb);
259 struct config_llog_data *cld;
262 if ((lsi->lsi_flags & LSI_SERVER) && IS_OST(lsi->lsi_ldd))
265 /* we have to use different llog for clients and mdts for cmd
266 * where only clients are notified if one of cmd server restarts */
267 LASSERT(strlen(fsname) < sizeof(logname) / 2);
268 strcpy(logname, fsname);
269 if (lsi->lsi_flags & LSI_SERVER) { /* mdt */
270 LASSERT(lcfg.cfg_instance == NULL);
271 lcfg.cfg_instance = sb;
272 strcat(logname, "-mdtir");
274 LASSERT(lcfg.cfg_instance != NULL);
275 strcat(logname, "-cliir");
278 cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
283 /** Add this log to the list of active logs watched by an MGC.
284 * Active means we're watching for updates.
285 * We have one active log per "mount" - client instance or servername.
286 * Each instance may be at a different point in the log.
288 static int config_log_add(struct obd_device *obd, char *logname,
289 struct config_llog_instance *cfg,
290 struct super_block *sb)
292 struct lustre_sb_info *lsi = s2lsi(sb);
293 struct config_llog_data *cld;
294 struct config_llog_data *sptlrpc_cld;
299 CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
302 * for each regular log, the depended sptlrpc log name is
303 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
305 ptr = strrchr(logname, '-');
306 if (ptr == NULL || ptr - logname > 8) {
307 CERROR("logname %s is too long\n", logname);
311 memcpy(seclogname, logname, ptr - logname);
312 strcpy(seclogname + (ptr - logname), "-sptlrpc");
314 sptlrpc_cld = config_log_find(seclogname, NULL);
315 if (sptlrpc_cld == NULL) {
316 sptlrpc_cld = do_config_log_add(obd, seclogname,
317 CONFIG_T_SPTLRPC, NULL, NULL);
318 if (IS_ERR(sptlrpc_cld)) {
319 CERROR("can't create sptlrpc log: %s\n", seclogname);
320 RETURN(PTR_ERR(sptlrpc_cld));
324 cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
326 CERROR("can't create log: %s\n", logname);
327 config_log_put(sptlrpc_cld);
328 RETURN(PTR_ERR(cld));
331 cld->cld_sptlrpc = sptlrpc_cld;
333 LASSERT(lsi->lsi_lmd);
334 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
335 struct config_llog_data *recover_cld;
336 *strrchr(seclogname, '-') = 0;
337 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
338 if (IS_ERR(recover_cld)) {
340 RETURN(PTR_ERR(recover_cld));
342 cld->cld_recover = recover_cld;
348 CFS_DEFINE_MUTEX(llog_process_lock);
350 /** Stop watching for updates on this log.
352 static int config_log_end(char *logname, struct config_llog_instance *cfg)
354 struct config_llog_data *cld;
355 struct config_llog_data *cld_sptlrpc = NULL;
356 struct config_llog_data *cld_recover = NULL;
360 cld = config_log_find(logname, cfg);
364 cfs_mutex_lock(&cld->cld_lock);
366 * if cld_stopping is set, it means we didn't start the log thus
367 * not owning the start ref. this can happen after previous umount:
368 * the cld still hanging there waiting for lock cancel, and we
369 * remount again but failed in the middle and call log_end without
372 if (unlikely(cld->cld_stopping)) {
373 cfs_mutex_unlock(&cld->cld_lock);
374 /* drop the ref from the find */
379 cld->cld_stopping = 1;
381 cld_recover = cld->cld_recover;
382 cld->cld_recover = NULL;
383 cfs_mutex_unlock(&cld->cld_lock);
386 cfs_mutex_lock(&cld_recover->cld_lock);
387 cld_recover->cld_stopping = 1;
388 cfs_mutex_unlock(&cld_recover->cld_lock);
389 config_log_put(cld_recover);
392 cfs_spin_lock(&config_list_lock);
393 cld_sptlrpc = cld->cld_sptlrpc;
394 cld->cld_sptlrpc = NULL;
395 cfs_spin_unlock(&config_list_lock);
398 config_log_put(cld_sptlrpc);
400 /* drop the ref from the find */
402 /* drop the start ref */
405 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
410 int lprocfs_mgc_rd_ir_state(char *page, char **start, off_t off,
411 int count, int *eof, void *data)
413 struct obd_device *obd = data;
414 struct obd_import *imp = obd->u.cli.cl_import;
415 struct obd_connect_data *ocd = &imp->imp_connect_data;
416 struct config_llog_data *cld;
420 rc = snprintf(page, count, "imperative_recovery: %s\n",
421 OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ON" : "OFF");
422 rc += snprintf(page + rc, count - rc, "client_state:\n");
424 cfs_spin_lock(&config_list_lock);
425 cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
426 if (cld->cld_recover == NULL)
428 rc += snprintf(page + rc, count - rc,
429 " - { client: %s, nidtbl_version: %u }\n",
431 cld->cld_recover->cld_cfg.cfg_last_idx);
433 cfs_spin_unlock(&config_list_lock);
438 /* reenqueue any lost locks */
439 #define RQ_RUNNING 0x1
443 static int rq_state = 0;
444 static cfs_waitq_t rq_waitq;
445 static CFS_DECLARE_COMPLETION(rq_exit);
447 static void do_requeue(struct config_llog_data *cld)
450 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
452 /* Do not run mgc_process_log on a disconnected export or an
453 export which is being disconnected. Take the client
454 semaphore to make the check non-racy. */
455 cfs_down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
456 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
457 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
458 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
460 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
463 cfs_up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
468 /* this timeout represents how many seconds MGC should wait before
469 * requeue config and recover lock to the MGS. We need to randomize this
470 * in order to not flood the MGS.
472 #define MGC_TIMEOUT_MIN_SECONDS 5
473 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
475 static int mgc_requeue_thread(void *data)
477 char name[] = "ll_cfg_requeue";
483 CDEBUG(D_MGC, "Starting requeue thread\n");
485 /* Keep trying failed locks periodically */
486 cfs_spin_lock(&config_list_lock);
487 rq_state |= RQ_RUNNING;
489 struct l_wait_info lwi;
490 struct config_llog_data *cld, *cld_prev;
491 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
492 int stopped = !!(rq_state & RQ_STOP);
495 /* Any new or requeued lostlocks will change the state */
496 rq_state &= ~(RQ_NOW | RQ_LATER);
497 cfs_spin_unlock(&config_list_lock);
499 /* Always wait a few seconds to allow the server who
500 caused the lock revocation to finish its setup, plus some
501 random so everyone doesn't try to reconnect at once. */
502 to = MGC_TIMEOUT_MIN_SECONDS * CFS_HZ;
503 to += rand * CFS_HZ / 100; /* rand is centi-seconds */
504 lwi = LWI_TIMEOUT(to, NULL, NULL);
505 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi);
508 * iterate & processing through the list. for each cld, process
509 * its depending sptlrpc cld firstly (if any) and then itself.
511 * it's guaranteed any item in the list must have
512 * reference > 0; and if cld_lostlock is set, at
513 * least one reference is taken by the previous enqueue.
517 cfs_spin_lock(&config_list_lock);
518 cfs_list_for_each_entry(cld, &config_llog_list,
520 if (!cld->cld_lostlock)
523 cfs_spin_unlock(&config_list_lock);
525 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
527 /* Whether we enqueued again or not in mgc_process_log,
528 * we're done with the ref from the old enqueue */
530 config_log_put(cld_prev);
533 cld->cld_lostlock = 0;
534 if (likely(!stopped))
537 cfs_spin_lock(&config_list_lock);
539 cfs_spin_unlock(&config_list_lock);
541 config_log_put(cld_prev);
543 /* break after scanning the list so that we can drop
544 * refcount to losing lock clds */
545 if (unlikely(stopped)) {
546 cfs_spin_lock(&config_list_lock);
550 /* Wait a bit to see if anyone else needs a requeue */
551 lwi = (struct l_wait_info) { 0 };
552 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
554 cfs_spin_lock(&config_list_lock);
556 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
557 rq_state &= ~RQ_RUNNING;
558 cfs_spin_unlock(&config_list_lock);
560 cfs_complete(&rq_exit);
562 CDEBUG(D_MGC, "Ending requeue thread\n");
566 /* Add a cld to the list to requeue. Start the requeue thread if needed.
567 We are responsible for dropping the config log reference from here on out. */
568 static void mgc_requeue_add(struct config_llog_data *cld)
572 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
573 cld->cld_logname, cfs_atomic_read(&cld->cld_refcount),
574 cld->cld_stopping, rq_state);
575 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
577 cfs_mutex_lock(&cld->cld_lock);
578 if (cld->cld_stopping || cld->cld_lostlock) {
579 cfs_mutex_unlock(&cld->cld_lock);
582 /* this refcount will be released in mgc_requeue_thread. */
584 cld->cld_lostlock = 1;
585 cfs_mutex_unlock(&cld->cld_lock);
587 /* Hold lock for rq_state */
588 cfs_spin_lock(&config_list_lock);
589 if (rq_state & RQ_STOP) {
590 cfs_spin_unlock(&config_list_lock);
591 cld->cld_lostlock = 0;
595 cfs_spin_unlock(&config_list_lock);
596 cfs_waitq_signal(&rq_waitq);
601 /********************** class fns **********************/
603 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
604 struct vfsmount *mnt)
606 struct lvfs_run_ctxt saved;
607 struct lustre_sb_info *lsi = s2lsi(sb);
608 struct client_obd *cli = &obd->u.cli;
609 struct dentry *dentry;
615 LASSERT(lsi->lsi_srv_mnt == mnt);
617 /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
618 cfs_down(&cli->cl_mgc_sem);
620 cfs_cleanup_group_info();
622 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
623 if (IS_ERR(obd->obd_fsops)) {
624 cfs_up(&cli->cl_mgc_sem);
625 CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd),
626 PTR_ERR(obd->obd_fsops));
627 RETURN(PTR_ERR(obd->obd_fsops));
630 cli->cl_mgc_vfsmnt = mnt;
631 err = fsfilt_setup(obd, mnt->mnt_sb);
635 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
636 obd->obd_lvfs_ctxt.pwdmnt = mnt;
637 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
638 obd->obd_lvfs_ctxt.fs = get_ds();
640 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
641 dentry = ll_lookup_one_len(MOUNT_CONFIGS_DIR, cfs_fs_pwd(current->fs),
642 strlen(MOUNT_CONFIGS_DIR));
643 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
644 if (IS_ERR(dentry)) {
645 err = PTR_ERR(dentry);
646 CERROR("cannot lookup %s directory: rc = %d\n",
647 MOUNT_CONFIGS_DIR, err);
650 cli->cl_mgc_configs_dir = dentry;
652 /* We take an obd ref to insure that we can't get to mgc_cleanup
653 without calling mgc_fs_cleanup first. */
654 class_incref(obd, "mgc_fs", obd);
656 label = fsfilt_get_label(obd, mnt->mnt_sb);
658 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
660 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
664 fsfilt_put_ops(obd->obd_fsops);
665 obd->obd_fsops = NULL;
666 cli->cl_mgc_vfsmnt = NULL;
667 cfs_up(&cli->cl_mgc_sem);
671 static int mgc_fs_cleanup(struct obd_device *obd)
673 struct client_obd *cli = &obd->u.cli;
677 LASSERT(cli->cl_mgc_vfsmnt != NULL);
679 if (cli->cl_mgc_configs_dir != NULL) {
680 struct lvfs_run_ctxt saved;
681 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
682 l_dput(cli->cl_mgc_configs_dir);
683 cli->cl_mgc_configs_dir = NULL;
684 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
685 class_decref(obd, "mgc_fs", obd);
688 cli->cl_mgc_vfsmnt = NULL;
690 fsfilt_put_ops(obd->obd_fsops);
692 cfs_up(&cli->cl_mgc_sem);
697 static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0);
698 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
704 case OBD_CLEANUP_EARLY:
706 case OBD_CLEANUP_EXPORTS:
707 if (cfs_atomic_dec_and_test(&mgc_count)) {
709 /* stop requeue thread */
710 cfs_spin_lock(&config_list_lock);
711 running = rq_state & RQ_RUNNING;
714 cfs_spin_unlock(&config_list_lock);
716 cfs_waitq_signal(&rq_waitq);
717 cfs_wait_for_completion(&rq_exit);
720 obd_cleanup_client_import(obd);
721 rc = obd_llog_finish(obd, 0);
723 CERROR("failed to cleanup llogging subsystems\n");
729 static int mgc_cleanup(struct obd_device *obd)
731 struct client_obd *cli = &obd->u.cli;
735 LASSERT(cli->cl_mgc_vfsmnt == NULL);
737 /* COMPAT_146 - old config logs may have added profiles we don't
739 if (obd->obd_type->typ_refcnt <= 1)
740 /* Only for the last mgc */
741 class_del_profiles();
743 lprocfs_obd_cleanup(obd);
746 rc = client_obd_cleanup(obd);
750 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
752 struct lprocfs_static_vars lvars;
758 rc = client_obd_setup(obd, lcfg);
760 GOTO(err_decref, rc);
762 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
764 CERROR("failed to setup llogging subsystems\n");
765 GOTO(err_cleanup, rc);
768 lprocfs_mgc_init_vars(&lvars);
769 lprocfs_obd_setup(obd, lvars.obd_vars);
770 sptlrpc_lprocfs_cliobd_attach(obd);
772 if (cfs_atomic_inc_return(&mgc_count) == 1) {
774 cfs_waitq_init(&rq_waitq);
776 /* start requeue thread */
777 rc = cfs_create_thread(mgc_requeue_thread, NULL,
780 CERROR("%s: Cannot start requeue thread (%d),"
781 "no more log updates!\n",
783 GOTO(err_cleanup, rc);
785 /* rc is the pid of mgc_requeue_thread. */
792 client_obd_cleanup(obd);
798 /* based on ll_mdc_blocking_ast */
799 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
800 void *data, int flag)
802 struct lustre_handle lockh;
803 struct config_llog_data *cld = (struct config_llog_data *)data;
808 case LDLM_CB_BLOCKING:
809 /* mgs wants the lock, give it up... */
810 LDLM_DEBUG(lock, "MGC blocking CB");
811 ldlm_lock2handle(lock, &lockh);
812 rc = ldlm_cli_cancel(&lockh);
814 case LDLM_CB_CANCELING:
815 /* We've given up the lock, prepare ourselves to update. */
816 LDLM_DEBUG(lock, "MGC cancel CB");
818 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
819 lock->l_resource->lr_name.name[0],
820 (char *)&lock->l_resource->lr_name.name[0]);
823 CDEBUG(D_INFO, "missing data, won't requeue\n");
827 /* held at mgc_process_log(). */
828 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
829 /* Are we done with this log? */
830 if (cld->cld_stopping) {
831 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
836 /* Make sure not to re-enqueue when the mgc is stopping
837 (we get called from client_disconnect_export) */
838 if (!lock->l_conn_export ||
839 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
840 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
847 mgc_requeue_add(cld);
857 /* Not sure where this should go... */
858 #define MGC_ENQUEUE_LIMIT 50
859 #define MGC_TARGET_REG_LIMIT 10
860 #define MGC_SEND_PARAM_LIMIT 10
862 /* Send parameter to MGS*/
863 static int mgc_set_mgs_param(struct obd_export *exp,
864 struct mgs_send_param *msp)
866 struct ptlrpc_request *req;
867 struct mgs_send_param *req_msp, *rep_msp;
871 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
872 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
877 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
879 ptlrpc_req_finished(req);
883 memcpy(req_msp, msp, sizeof(*req_msp));
884 ptlrpc_request_set_replen(req);
886 /* Limit how long we will wait for the enqueue to complete */
887 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
888 rc = ptlrpc_queue_wait(req);
890 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
891 memcpy(msp, rep_msp, sizeof(*rep_msp));
894 ptlrpc_req_finished(req);
899 /* Take a config lock so we can get cancel notifications */
900 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
901 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
902 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
903 void *data, __u32 lvb_len, void *lvb_swabber,
904 struct lustre_handle *lockh)
906 struct config_llog_data *cld = (struct config_llog_data *)data;
907 struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
908 ldlm_completion_ast, NULL, NULL, NULL };
909 struct ptlrpc_request *req;
910 int short_limit = cld_is_sptlrpc(cld);
914 CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
915 cld->cld_resid.name[0]);
917 /* We need a callback for every lockholder, so don't try to
918 ldlm_lock_match (see rev 1.1.2.11.2.47) */
919 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
920 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
924 ptlrpc_request_set_replen(req);
926 /* check if this is server or client */
927 if (cld->cld_cfg.cfg_sb) {
928 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
929 if (lsi && (lsi->lsi_flags & LSI_SERVER))
932 /* Limit how long we will wait for the enqueue to complete */
933 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
934 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
936 /* A failed enqueue should still call the mgc_blocking_ast,
937 where it will be requeued if needed ("grant failed"). */
938 ptlrpc_req_finished(req);
942 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
943 __u32 mode, struct lustre_handle *lockh)
947 ldlm_lock_decref(lockh, mode);
952 static void mgc_notify_active(struct obd_device *unused)
954 /* wakeup mgc_requeue_thread to requeue mgc lock */
955 cfs_spin_lock(&config_list_lock);
957 cfs_spin_unlock(&config_list_lock);
958 cfs_waitq_signal(&rq_waitq);
960 /* TODO: Help the MGS rebuild nidtbl. -jay */
963 /* Send target_reg message to MGS */
964 static int mgc_target_register(struct obd_export *exp,
965 struct mgs_target_info *mti)
967 struct ptlrpc_request *req;
968 struct mgs_target_info *req_mti, *rep_mti;
972 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
973 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
978 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
980 ptlrpc_req_finished(req);
984 memcpy(req_mti, mti, sizeof(*req_mti));
985 ptlrpc_request_set_replen(req);
986 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
987 /* Limit how long we will wait for the enqueue to complete */
988 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
990 rc = ptlrpc_queue_wait(req);
992 rep_mti = req_capsule_server_get(&req->rq_pill,
993 &RMF_MGS_TARGET_INFO);
994 memcpy(mti, rep_mti, sizeof(*rep_mti));
995 CDEBUG(D_MGC, "register %s got index = %d\n",
996 mti->mti_svname, mti->mti_stripe_index);
998 ptlrpc_req_finished(req);
1003 int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1004 obd_count keylen, void *key, obd_count vallen,
1005 void *val, struct ptlrpc_request_set *set)
1010 /* Turn off initial_recov after we try all backup servers once */
1011 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1012 struct obd_import *imp = class_exp2cliimp(exp);
1014 if (vallen != sizeof(int))
1016 value = *(int *)val;
1017 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1018 imp->imp_obd->obd_name, value,
1019 imp->imp_deactive, imp->imp_invalid,
1020 imp->imp_replayable, imp->imp_obd->obd_replayable,
1021 ptlrpc_import_state_name(imp->imp_state));
1022 /* Resurrect if we previously died */
1023 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1024 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1025 ptlrpc_reconnect_import(imp);
1028 /* FIXME move this to mgc_process_config */
1029 if (KEY_IS(KEY_REGISTER_TARGET)) {
1030 struct mgs_target_info *mti;
1031 if (vallen != sizeof(struct mgs_target_info))
1033 mti = (struct mgs_target_info *)val;
1034 CDEBUG(D_MGC, "register_target %s %#x\n",
1035 mti->mti_svname, mti->mti_flags);
1036 rc = mgc_target_register(exp, mti);
1039 if (KEY_IS(KEY_SET_FS)) {
1040 struct super_block *sb = (struct super_block *)val;
1041 struct lustre_sb_info *lsi;
1042 if (vallen != sizeof(struct super_block))
1045 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
1047 CERROR("set_fs got %d\n", rc);
1051 if (KEY_IS(KEY_CLEAR_FS)) {
1054 rc = mgc_fs_cleanup(exp->exp_obd);
1056 CERROR("clear_fs got %d\n", rc);
1060 if (KEY_IS(KEY_SET_INFO)) {
1061 struct mgs_send_param *msp;
1063 msp = (struct mgs_send_param *)val;
1064 rc = mgc_set_mgs_param(exp, msp);
1067 if (KEY_IS(KEY_MGSSEC)) {
1068 struct client_obd *cli = &exp->exp_obd->u.cli;
1069 struct sptlrpc_flavor flvr;
1072 * empty string means using current flavor, if which haven't
1073 * been set yet, set it as null.
1075 * if flavor has been set previously, check the asking flavor
1076 * must match the existing one.
1079 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1085 rc = sptlrpc_parse_flavor(val, &flvr);
1087 CERROR("invalid sptlrpc flavor %s to MGS\n",
1093 * caller already hold a mutex
1095 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1096 cli->cl_flvr_mgc = flvr;
1097 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1098 sizeof(flvr)) != 0) {
1101 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1103 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1104 "currently %s is in use\n",
1114 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1115 __u32 keylen, void *key, __u32 *vallen, void *val,
1116 struct lov_stripe_md *unused)
1120 if (KEY_IS(KEY_CONN_DATA)) {
1121 struct obd_import *imp = class_exp2cliimp(exp);
1122 struct obd_connect_data *data = val;
1124 if (*vallen == sizeof(*data)) {
1125 *data = imp->imp_connect_data;
1133 static int mgc_import_event(struct obd_device *obd,
1134 struct obd_import *imp,
1135 enum obd_import_event event)
1139 LASSERT(imp->imp_obd == obd);
1140 CDEBUG(D_MGC, "import event %#x\n", event);
1143 case IMP_EVENT_DISCON:
1144 /* MGC imports should not wait for recovery */
1146 case IMP_EVENT_INACTIVE:
1148 case IMP_EVENT_INVALIDATE: {
1149 struct ldlm_namespace *ns = obd->obd_namespace;
1150 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1153 case IMP_EVENT_ACTIVE:
1154 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
1155 /* Clearing obd_no_recov allows us to continue pinging */
1156 obd->obd_no_recov = 0;
1157 mgc_notify_active(obd);
1161 case IMP_EVENT_DEACTIVATE:
1162 case IMP_EVENT_ACTIVATE:
1165 CERROR("Unknown import event %#x\n", event);
1171 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1172 struct obd_device *tgt, int *index)
1174 struct llog_ctxt *ctxt;
1178 LASSERT(olg == &obd->obd_olg);
1180 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
1185 rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
1188 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1190 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1195 llog_initiator_connect(ctxt);
1196 llog_ctxt_put(ctxt);
1198 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1206 static int mgc_llog_finish(struct obd_device *obd, int count)
1208 struct llog_ctxt *ctxt;
1209 int rc = 0, rc2 = 0;
1212 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1214 rc = llog_cleanup(ctxt);
1216 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1218 rc2 = llog_cleanup(ctxt);
1227 CONFIG_READ_NRPAGES_INIT = 1 << (20 - CFS_PAGE_SHIFT),
1228 CONFIG_READ_NRPAGES = 4
1231 static int mgc_apply_recover_logs(struct obd_device *mgc,
1232 struct config_llog_data *cld,
1234 void *data, int datalen, int need_swab)
1236 struct config_llog_instance *cfg = &cld->cld_cfg;
1237 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1238 struct mgs_nidtbl_entry *entry;
1239 struct lustre_cfg *lcfg;
1240 struct lustre_cfg_bufs bufs;
1241 u64 prev_version = 0;
1250 LASSERT(cfg->cfg_instance != NULL);
1251 LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1253 OBD_ALLOC(inst, CFS_PAGE_SIZE);
1257 if (!(lsi->lsi_flags & LSI_SERVER)) {
1258 pos = sprintf(inst, "%p", cfg->cfg_instance);
1260 LASSERT(IS_MDT(lsi->lsi_ldd));
1261 pos = sprintf(inst, "MDT%04x", lsi->lsi_ldd->ldd_svindex);
1266 bufsz = CFS_PAGE_SIZE - pos;
1268 while (datalen > 0) {
1269 int entry_len = sizeof(*entry);
1271 struct obd_device *obd;
1278 if (datalen < sizeof(*entry))
1281 entry = (typeof(entry))(data + off);
1284 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1286 if (entry->mne_nid_count == 0) /* at least one nid entry */
1288 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1291 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1292 if (datalen < entry_len) /* must have entry_len at least */
1296 lustre_swab_mgs_nidtbl_entry(entry);
1297 LASSERT(entry->mne_length <= CFS_PAGE_SIZE);
1298 if (entry->mne_length < entry_len)
1301 off += entry->mne_length;
1302 datalen -= entry->mne_length;
1306 if (entry->mne_version > max_version) {
1307 CERROR("entry index(%lld) is over max_index(%lld)\n",
1308 entry->mne_version, max_version);
1312 if (prev_version >= entry->mne_version) {
1313 CERROR("index unsorted, prev %lld, now %lld\n",
1314 prev_version, entry->mne_version);
1317 prev_version = entry->mne_version;
1320 * Write a string with format "nid::instance" to
1321 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1324 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1325 memset(buf, 0, bufsz);
1329 /* lustre-OST0001-osc-<instance #> */
1330 strcpy(obdname, cld->cld_logname);
1331 cname = strrchr(obdname, '-');
1332 if (cname == NULL) {
1333 CERROR("mgc %s: invalid logname %s\n",
1334 mgc->obd_name, obdname);
1338 pos = cname - obdname;
1340 pos += sprintf(obdname + pos, "-%s%04x",
1341 is_ost ? "OST" : "MDT", entry->mne_index);
1343 cname = is_ost ? "osc" : "mdc",
1344 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1345 lustre_cfg_bufs_reset(&bufs, obdname);
1347 /* find the obd by obdname */
1348 obd = class_name2obd(obdname);
1350 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1351 mgc->obd_name, obdname);
1353 /* this is a safe race, when the ost is starting up...*/
1357 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1360 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1363 /* TODO: iterate all nids to find one */
1364 /* find uuid by nid */
1365 rc = client_import_find_conn(obd->u.cli.cl_import,
1367 (struct obd_uuid *)uuid);
1369 CERROR("mgc: cannot find uuid by nid %s\n",
1370 libcfs_nid2str(entry->u.nids[0]));
1374 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1375 uuid, libcfs_nid2str(entry->u.nids[0]));
1377 pos += strlen(uuid);
1378 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1379 LASSERT(pos < bufsz);
1381 lustre_cfg_bufs_set_string(&bufs, 1, params);
1384 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1386 CERROR("mgc: cannot allocate memory\n");
1390 CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n",
1391 prev_version, max_version, obdname, params);
1393 rc = class_process_config(lcfg);
1394 lustre_cfg_free(lcfg);
1396 CDEBUG(D_INFO, "process config for %s error %d\n",
1399 /* continue, even one with error */
1402 OBD_FREE(inst, CFS_PAGE_SIZE);
1407 * This function is called if this client was notified for target restarting
1408 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1410 static int mgc_process_recover_log(struct obd_device *obd,
1411 struct config_llog_data *cld)
1413 struct ptlrpc_request *req = NULL;
1414 struct config_llog_instance *cfg = &cld->cld_cfg;
1415 struct mgs_config_body *body;
1416 struct mgs_config_res *res;
1417 struct ptlrpc_bulk_desc *desc;
1426 /* allocate buffer for bulk transfer.
1427 * if this is the first time for this mgs to read logs,
1428 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1429 * once; otherwise, it only reads increment of logs, this should be
1430 * small and CONFIG_READ_NRPAGES will be used.
1432 nrpages = CONFIG_READ_NRPAGES;
1433 if (cfg->cfg_last_idx == 0) /* the first time */
1434 nrpages = CONFIG_READ_NRPAGES_INIT;
1436 OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1438 GOTO(out, rc = -ENOMEM);
1440 for (i = 0; i < nrpages; i++) {
1441 pages[i] = cfs_alloc_page(CFS_ALLOC_STD);
1442 if (pages[i] == NULL)
1443 GOTO(out, rc = -ENOMEM);
1447 LASSERT(cld_is_recover(cld));
1448 LASSERT(cfs_mutex_is_locked(&cld->cld_lock));
1449 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1450 &RQF_MGS_CONFIG_READ);
1452 GOTO(out, rc = -ENOMEM);
1454 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1459 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1460 LASSERT(body != NULL);
1461 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1462 strncpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name));
1463 body->mcb_offset = cfg->cfg_last_idx + 1;
1464 body->mcb_type = cld->cld_type;
1465 body->mcb_bits = CFS_PAGE_SHIFT;
1466 body->mcb_units = nrpages;
1468 /* allocate bulk transfer descriptor */
1469 desc = ptlrpc_prep_bulk_imp(req, nrpages, BULK_PUT_SINK,
1472 GOTO(out, rc = -ENOMEM);
1474 for (i = 0; i < nrpages; i++)
1475 ptlrpc_prep_bulk_page(desc, pages[i], 0, CFS_PAGE_SIZE);
1477 ptlrpc_request_set_replen(req);
1478 rc = ptlrpc_queue_wait(req);
1482 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1483 if (res->mcr_size < res->mcr_offset)
1484 GOTO(out, rc = -EINVAL);
1486 /* always update the index even though it might have errors with
1487 * handling the recover logs */
1488 cfg->cfg_last_idx = res->mcr_offset;
1489 eof = res->mcr_offset == res->mcr_size;
1491 CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n",
1492 res->mcr_offset, eof == false);
1494 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1496 GOTO(out, rc = ealen);
1498 if (ealen > nrpages << CFS_PAGE_SHIFT)
1499 GOTO(out, rc = -EINVAL);
1501 if (ealen == 0) { /* no logs transferred */
1507 for (i = 0; i < nrpages && ealen > 0; i++) {
1511 ptr = cfs_kmap(pages[i]);
1512 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1513 min_t(int, ealen, CFS_PAGE_SIZE),
1514 ptlrpc_rep_need_swab(req));
1515 cfs_kunmap(pages[i]);
1517 CWARN("Process recover log %s error %d\n",
1518 cld->cld_logname, rc2);
1522 ealen -= CFS_PAGE_SIZE;
1527 ptlrpc_req_finished(req);
1529 if (rc == 0 && !eof)
1533 for (i = 0; i < nrpages; i++) {
1534 if (pages[i] == NULL)
1536 cfs_free_page(pages[i]);
1538 OBD_FREE(pages, sizeof(*pages) * nrpages);
1543 /* identical to mgs_log_is_empty */
1544 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1547 struct lvfs_run_ctxt saved;
1548 struct llog_handle *llh;
1551 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1552 rc = llog_create(ctxt, &llh, NULL, name);
1554 llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL);
1555 rc = llog_get_size(llh);
1558 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1559 /* header is record 1 */
1563 static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec,
1566 struct llog_rec_hdr local_rec = *rec;
1567 struct llog_handle *local_llh = (struct llog_handle *)data;
1568 char *cfg_buf = (char*) (rec + 1);
1569 struct lustre_cfg *lcfg;
1573 /* Append all records */
1574 local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
1575 rc = llog_write_rec(local_llh, &local_rec, NULL, 0,
1576 (void *)cfg_buf, -1);
1578 lcfg = (struct lustre_cfg *)cfg_buf;
1579 CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
1580 rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
1581 lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
1586 /* Copy a remote log locally */
1587 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1588 struct llog_ctxt *lctxt, char *logname)
1590 struct llog_handle *local_llh, *remote_llh;
1591 struct obd_uuid *uuid;
1596 /* Write new log to a temp name, then vfs_rename over logname
1597 upon successful completion. */
1599 OBD_ALLOC(temp_log, strlen(logname) + 1);
1602 sprintf(temp_log, "%sT", logname);
1604 /* Make sure there's no old temp log */
1605 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1608 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL);
1611 rc = llog_destroy(local_llh);
1612 llog_free_handle(local_llh);
1616 /* open local log */
1617 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1621 /* set the log header uuid for fun */
1622 OBD_ALLOC_PTR(uuid);
1623 obd_str2uuid(uuid, logname);
1624 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid);
1627 GOTO(out_closel, rc);
1629 /* open remote log */
1630 rc = llog_create(rctxt, &remote_llh, NULL, logname);
1632 GOTO(out_closel, rc);
1633 rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL);
1635 GOTO(out_closer, rc);
1637 /* Copy remote log */
1638 rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL);
1641 rc2 = llog_close(remote_llh);
1645 rc2 = llog_close(local_llh);
1649 /* We've copied the remote log to the local temp log, now
1650 replace the old local log with the temp log. */
1652 struct client_obd *cli = &obd->u.cli;
1654 LASSERT(cli->cl_mgc_configs_dir);
1655 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1658 CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1661 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1662 OBD_FREE(temp_log, strlen(logname) + 1);
1666 /* local_only means it cannot get remote llogs */
1667 static int mgc_process_cfg_log(struct obd_device *mgc,
1668 struct config_llog_data *cld,
1671 struct llog_ctxt *ctxt, *lctxt = NULL;
1672 struct client_obd *cli = &mgc->u.cli;
1673 struct lvfs_run_ctxt *saved_ctxt;
1674 struct lustre_sb_info *lsi = NULL;
1675 int rc = 0, must_pop = 0;
1676 bool sptlrpc_started = false;
1681 LASSERT(cfs_mutex_is_locked(&cld->cld_lock));
1684 * local copy of sptlrpc log is controlled elsewhere, don't try to
1687 if (cld_is_sptlrpc(cld) && local_only)
1690 if (cld->cld_cfg.cfg_sb)
1691 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1693 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1695 CERROR("missing llog context\n");
1699 OBD_ALLOC_PTR(saved_ctxt);
1700 if (saved_ctxt == NULL)
1703 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1705 /* Copy the setup log locally if we can. Don't mess around if we're
1706 running an MGS though (logs are already local). */
1707 if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
1708 (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1709 !IS_MGS(lsi->lsi_ldd) && lsi->lsi_lmd->lmd_osd_type == NULL) {
1710 push_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1713 /* Only try to copy log if we have the lock. */
1714 rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1715 if (local_only || rc) {
1716 if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1717 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1718 "log %s and no local copy."
1719 "\n", cld->cld_logname);
1720 GOTO(out_pop, rc = -ENOTCONN);
1722 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1723 "copy for now, will try to update later.\n",
1726 /* Now, whether we copied or not, start using the local llog.
1727 If we failed to copy, we'll start using whatever the old
1729 llog_ctxt_put(ctxt);
1732 } else if (local_only) { /* no local log at client side */
1733 GOTO(out_pop, rc = -EIO);
1736 if (cld_is_sptlrpc(cld)) {
1737 sptlrpc_conf_log_update_begin(cld->cld_logname);
1738 sptlrpc_started = true;
1741 /* logname and instance info should be the same, so use our
1742 copy of the instance for the update. The cfg_last_idx will
1744 rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg);
1748 llog_ctxt_put(ctxt);
1750 llog_ctxt_put(lctxt);
1752 pop_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1754 OBD_FREE_PTR(saved_ctxt);
1756 * update settings on existing OBDs. doing it inside
1757 * of llog_process_lock so no device is attaching/detaching
1759 * the logname must be <fsname>-sptlrpc
1761 if (sptlrpc_started) {
1762 LASSERT(cld_is_sptlrpc(cld));
1763 sptlrpc_conf_log_update_end(cld->cld_logname);
1764 class_notify_sptlrpc_conf(cld->cld_logname,
1765 strlen(cld->cld_logname) -
1766 strlen("-sptlrpc"));
1772 /** Get a config log from the MGS and process it.
1773 * This func is called for both clients and servers.
1774 * Copy the log locally before parsing it if appropriate (non-MGS server)
1776 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1778 struct lustre_handle lockh = { 0 };
1779 int rc = 0, rcl, flags = LDLM_FL_NO_LRU;
1784 /* I don't want multiple processes running process_log at once --
1785 sounds like badness. It actually might be fine, as long as
1786 we're not trying to update from the same log
1787 simultaneously (in which case we should use a per-log sem.) */
1788 cfs_mutex_lock(&cld->cld_lock);
1789 if (cld->cld_stopping) {
1790 cfs_mutex_unlock(&cld->cld_lock);
1794 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1796 CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1797 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1799 /* Get the cfg lock on the llog */
1800 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1801 LCK_CR, &flags, NULL, NULL, NULL,
1802 cld, 0, NULL, &lockh);
1804 /* Get the cld, it will be released in mgc_blocking_ast. */
1805 config_log_get(cld);
1806 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1809 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1811 /* mark cld_lostlock so that it will requeue
1812 * after MGC becomes available. */
1813 cld->cld_lostlock = 1;
1814 /* Get extra reference, it will be put in requeue thread */
1815 config_log_get(cld);
1819 if (cld_is_recover(cld)) {
1820 rc = 0; /* this is not a fatal error for recover log */
1822 rc = mgc_process_recover_log(mgc, cld);
1824 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1827 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1828 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1830 cfs_mutex_unlock(&cld->cld_lock);
1832 /* Now drop the lock so MGS can revoke it */
1834 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1837 CERROR("Can't drop cfg lock: %d\n", rcl);
1844 /** Called from lustre_process_log.
1845 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1846 * any services, and adds it to the list logs to watch (follow).
1848 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1850 struct lustre_cfg *lcfg = buf;
1851 struct config_llog_instance *cfg = NULL;
1856 switch(lcfg->lcfg_command) {
1857 case LCFG_LOV_ADD_OBD: {
1858 /* Overloading this cfg command: register a new target */
1859 struct mgs_target_info *mti;
1861 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1862 sizeof(struct mgs_target_info))
1863 GOTO(out, rc = -EINVAL);
1865 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1866 CDEBUG(D_MGC, "add_target %s %#x\n",
1867 mti->mti_svname, mti->mti_flags);
1868 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1871 case LCFG_LOV_DEL_OBD:
1872 /* Unregister has no meaning at the moment. */
1873 CERROR("lov_del_obd unimplemented\n");
1876 case LCFG_SPTLRPC_CONF: {
1877 rc = sptlrpc_process_config(lcfg);
1880 case LCFG_LOG_START: {
1881 struct config_llog_data *cld;
1882 struct super_block *sb;
1884 logname = lustre_cfg_string(lcfg, 1);
1885 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1886 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1888 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1891 /* We're only called through here on the initial mount */
1892 rc = config_log_add(obd, logname, cfg, sb);
1895 cld = config_log_find(logname, cfg);
1902 /* FIXME only set this for old logs! Right now this forces
1903 us to always skip the "inside markers" check */
1904 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1906 rc = mgc_process_log(obd, cld);
1907 if (rc == 0 && cld->cld_recover != NULL) {
1908 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1909 imp_connect_data, IMP_RECOV)) {
1910 rc = mgc_process_log(obd, cld->cld_recover);
1912 struct config_llog_data *cir = cld->cld_recover;
1913 cld->cld_recover = NULL;
1914 config_log_put(cir);
1917 CERROR("Cannot process recover llog %d\n", rc);
1919 config_log_put(cld);
1923 case LCFG_LOG_END: {
1924 logname = lustre_cfg_string(lcfg, 1);
1926 if (lcfg->lcfg_bufcount >= 2)
1927 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1929 rc = config_log_end(logname, cfg);
1933 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1934 GOTO(out, rc = -EINVAL);
1942 struct obd_ops mgc_obd_ops = {
1943 .o_owner = THIS_MODULE,
1944 .o_setup = mgc_setup,
1945 .o_precleanup = mgc_precleanup,
1946 .o_cleanup = mgc_cleanup,
1947 .o_add_conn = client_import_add_conn,
1948 .o_del_conn = client_import_del_conn,
1949 .o_connect = client_connect_import,
1950 .o_disconnect = client_disconnect_export,
1951 //.o_enqueue = mgc_enqueue,
1952 .o_cancel = mgc_cancel,
1953 //.o_iocontrol = mgc_iocontrol,
1954 .o_set_info_async = mgc_set_info_async,
1955 .o_get_info = mgc_get_info,
1956 .o_import_event = mgc_import_event,
1957 .o_llog_init = mgc_llog_init,
1958 .o_llog_finish = mgc_llog_finish,
1959 .o_process_config = mgc_process_config,
1962 int __init mgc_init(void)
1964 return class_register_type(&mgc_obd_ops, NULL, NULL,
1965 LUSTRE_MGC_NAME, NULL);
1969 static void /*__exit*/ mgc_exit(void)
1971 class_unregister_type(LUSTRE_MGC_NAME);
1974 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1975 MODULE_DESCRIPTION("Lustre Management Client");
1976 MODULE_LICENSE("GPL");
1978 module_init(mgc_init);
1979 module_exit(mgc_exit);