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, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mgc/mgc_request.c
38 * Author: Nathan Rutman <nathan@clusterfs.com>
41 #define DEBUG_SUBSYSTEM S_MGC
42 #define D_MGC D_CONFIG /*|D_WARNING*/
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 DEFINE_SPINLOCK(config_list_lock);
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 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 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 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);
218 cld->cld_cfg.cfg_callback = class_config_llog_handler;
219 mutex_init(&cld->cld_lock);
220 cld->cld_cfg.cfg_last_idx = 0;
221 cld->cld_cfg.cfg_flags = 0;
222 cld->cld_cfg.cfg_sb = sb;
223 cld->cld_type = type;
224 cfs_atomic_set(&cld->cld_refcount, 1);
226 /* Keep the mgc around until we are done */
227 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
229 if (cld_is_sptlrpc(cld)) {
230 sptlrpc_conf_log_start(logname);
231 cld->cld_cfg.cfg_obdname = obd->obd_name;
234 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
236 spin_lock(&config_list_lock);
237 cfs_list_add(&cld->cld_list_chain, &config_llog_list);
238 spin_unlock(&config_list_lock);
245 if (cld_is_sptlrpc(cld)) {
246 rc = mgc_process_log(obd, cld);
247 if (rc && rc != -ENOENT)
248 CERROR("failed processing sptlrpc log: %d\n", rc);
254 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
256 struct config_llog_instance *cfg,
257 struct super_block *sb)
259 struct config_llog_instance lcfg = *cfg;
260 struct lustre_sb_info *lsi = s2lsi(sb);
261 struct config_llog_data *cld;
267 /* for osp-on-ost, see lustre_start_osp() */
268 if (IS_MDT(lsi) && lcfg.cfg_instance)
271 /* we have to use different llog for clients and mdts for cmd
272 * where only clients are notified if one of cmd server restarts */
273 LASSERT(strlen(fsname) < sizeof(logname) / 2);
274 strcpy(logname, fsname);
275 if (IS_SERVER(lsi)) { /* mdt */
276 LASSERT(lcfg.cfg_instance == NULL);
277 lcfg.cfg_instance = sb;
278 strcat(logname, "-mdtir");
280 LASSERT(lcfg.cfg_instance != NULL);
281 strcat(logname, "-cliir");
284 cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
289 /** Add this log to the list of active logs watched by an MGC.
290 * Active means we're watching for updates.
291 * We have one active log per "mount" - client instance or servername.
292 * Each instance may be at a different point in the log.
294 static int config_log_add(struct obd_device *obd, char *logname,
295 struct config_llog_instance *cfg,
296 struct super_block *sb)
298 struct lustre_sb_info *lsi = s2lsi(sb);
299 struct config_llog_data *cld;
300 struct config_llog_data *sptlrpc_cld;
305 CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
308 * for each regular log, the depended sptlrpc log name is
309 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
311 ptr = strrchr(logname, '-');
312 if (ptr == NULL || ptr - logname > 8) {
313 CERROR("logname %s is too long\n", logname);
317 memcpy(seclogname, logname, ptr - logname);
318 strcpy(seclogname + (ptr - logname), "-sptlrpc");
320 sptlrpc_cld = config_log_find(seclogname, NULL);
321 if (sptlrpc_cld == NULL) {
322 sptlrpc_cld = do_config_log_add(obd, seclogname,
323 CONFIG_T_SPTLRPC, NULL, NULL);
324 if (IS_ERR(sptlrpc_cld)) {
325 CERROR("can't create sptlrpc log: %s\n", seclogname);
326 RETURN(PTR_ERR(sptlrpc_cld));
330 cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
332 CERROR("can't create log: %s\n", logname);
333 config_log_put(sptlrpc_cld);
334 RETURN(PTR_ERR(cld));
337 cld->cld_sptlrpc = sptlrpc_cld;
339 LASSERT(lsi->lsi_lmd);
340 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
341 struct config_llog_data *recover_cld;
342 *strrchr(seclogname, '-') = 0;
343 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
344 if (IS_ERR(recover_cld)) {
346 RETURN(PTR_ERR(recover_cld));
348 cld->cld_recover = recover_cld;
354 DEFINE_MUTEX(llog_process_lock);
356 /** Stop watching for updates on this log.
358 static int config_log_end(char *logname, struct config_llog_instance *cfg)
360 struct config_llog_data *cld;
361 struct config_llog_data *cld_sptlrpc = NULL;
362 struct config_llog_data *cld_recover = NULL;
366 cld = config_log_find(logname, cfg);
370 mutex_lock(&cld->cld_lock);
372 * if cld_stopping is set, it means we didn't start the log thus
373 * not owning the start ref. this can happen after previous umount:
374 * the cld still hanging there waiting for lock cancel, and we
375 * remount again but failed in the middle and call log_end without
378 if (unlikely(cld->cld_stopping)) {
379 mutex_unlock(&cld->cld_lock);
380 /* drop the ref from the find */
385 cld->cld_stopping = 1;
387 cld_recover = cld->cld_recover;
388 cld->cld_recover = NULL;
389 mutex_unlock(&cld->cld_lock);
392 mutex_lock(&cld_recover->cld_lock);
393 cld_recover->cld_stopping = 1;
394 mutex_unlock(&cld_recover->cld_lock);
395 config_log_put(cld_recover);
398 spin_lock(&config_list_lock);
399 cld_sptlrpc = cld->cld_sptlrpc;
400 cld->cld_sptlrpc = NULL;
401 spin_unlock(&config_list_lock);
404 config_log_put(cld_sptlrpc);
406 /* drop the ref from the find */
408 /* drop the start ref */
411 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
416 int lprocfs_mgc_rd_ir_state(char *page, char **start, off_t off,
417 int count, int *eof, void *data)
419 struct obd_device *obd = data;
420 struct obd_import *imp = obd->u.cli.cl_import;
421 struct obd_connect_data *ocd = &imp->imp_connect_data;
422 struct config_llog_data *cld;
426 rc = snprintf(page, count, "imperative_recovery: %s\n",
427 OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
428 rc += snprintf(page + rc, count - rc, "client_state:\n");
430 spin_lock(&config_list_lock);
431 cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
432 if (cld->cld_recover == NULL)
434 rc += snprintf(page + rc, count - rc,
435 " - { client: %s, nidtbl_version: %u }\n",
437 cld->cld_recover->cld_cfg.cfg_last_idx);
439 spin_unlock(&config_list_lock);
444 /* reenqueue any lost locks */
445 #define RQ_RUNNING 0x1
449 static int rq_state = 0;
450 static cfs_waitq_t rq_waitq;
451 static DECLARE_COMPLETION(rq_exit);
453 static void do_requeue(struct config_llog_data *cld)
456 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
458 /* Do not run mgc_process_log on a disconnected export or an
459 export which is being disconnected. Take the client
460 semaphore to make the check non-racy. */
461 down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
462 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
463 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
464 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
466 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
469 up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
474 /* this timeout represents how many seconds MGC should wait before
475 * requeue config and recover lock to the MGS. We need to randomize this
476 * in order to not flood the MGS.
478 #define MGC_TIMEOUT_MIN_SECONDS 5
479 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
481 static int mgc_requeue_thread(void *data)
486 CDEBUG(D_MGC, "Starting requeue thread\n");
488 /* Keep trying failed locks periodically */
489 spin_lock(&config_list_lock);
490 rq_state |= RQ_RUNNING;
492 struct l_wait_info lwi;
493 struct config_llog_data *cld, *cld_prev;
494 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
495 int stopped = !!(rq_state & RQ_STOP);
498 /* Any new or requeued lostlocks will change the state */
499 rq_state &= ~(RQ_NOW | RQ_LATER);
500 spin_unlock(&config_list_lock);
502 /* Always wait a few seconds to allow the server who
503 caused the lock revocation to finish its setup, plus some
504 random so everyone doesn't try to reconnect at once. */
505 to = MGC_TIMEOUT_MIN_SECONDS * CFS_HZ;
506 to += rand * CFS_HZ / 100; /* rand is centi-seconds */
507 lwi = LWI_TIMEOUT(to, NULL, NULL);
508 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi);
511 * iterate & processing through the list. for each cld, process
512 * its depending sptlrpc cld firstly (if any) and then itself.
514 * it's guaranteed any item in the list must have
515 * reference > 0; and if cld_lostlock is set, at
516 * least one reference is taken by the previous enqueue.
520 spin_lock(&config_list_lock);
521 cfs_list_for_each_entry(cld, &config_llog_list,
523 if (!cld->cld_lostlock)
526 spin_unlock(&config_list_lock);
528 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
530 /* Whether we enqueued again or not in mgc_process_log,
531 * we're done with the ref from the old enqueue */
533 config_log_put(cld_prev);
536 cld->cld_lostlock = 0;
537 if (likely(!stopped))
540 spin_lock(&config_list_lock);
542 spin_unlock(&config_list_lock);
544 config_log_put(cld_prev);
546 /* break after scanning the list so that we can drop
547 * refcount to losing lock clds */
548 if (unlikely(stopped)) {
549 spin_lock(&config_list_lock);
553 /* Wait a bit to see if anyone else needs a requeue */
554 lwi = (struct l_wait_info) { 0 };
555 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
557 spin_lock(&config_list_lock);
559 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
560 rq_state &= ~RQ_RUNNING;
561 spin_unlock(&config_list_lock);
565 CDEBUG(D_MGC, "Ending requeue thread\n");
569 /* Add a cld to the list to requeue. Start the requeue thread if needed.
570 We are responsible for dropping the config log reference from here on out. */
571 static void mgc_requeue_add(struct config_llog_data *cld)
575 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
576 cld->cld_logname, cfs_atomic_read(&cld->cld_refcount),
577 cld->cld_stopping, rq_state);
578 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
580 mutex_lock(&cld->cld_lock);
581 if (cld->cld_stopping || cld->cld_lostlock) {
582 mutex_unlock(&cld->cld_lock);
585 /* this refcount will be released in mgc_requeue_thread. */
587 cld->cld_lostlock = 1;
588 mutex_unlock(&cld->cld_lock);
590 /* Hold lock for rq_state */
591 spin_lock(&config_list_lock);
592 if (rq_state & RQ_STOP) {
593 spin_unlock(&config_list_lock);
594 cld->cld_lostlock = 0;
598 spin_unlock(&config_list_lock);
599 cfs_waitq_signal(&rq_waitq);
604 /********************** class fns **********************/
606 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
607 struct vfsmount *mnt)
609 struct lvfs_run_ctxt saved;
610 struct lustre_sb_info *lsi = s2lsi(sb);
611 struct client_obd *cli = &obd->u.cli;
612 struct dentry *dentry;
618 LASSERT(lsi->lsi_srv_mnt == mnt);
620 /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
621 down(&cli->cl_mgc_sem);
623 cfs_cleanup_group_info();
625 obd->obd_fsops = fsfilt_get_ops(lsi->lsi_fstype);
626 if (IS_ERR(obd->obd_fsops)) {
627 up(&cli->cl_mgc_sem);
628 CERROR("%s: No fstype %s: rc = %ld\n", lsi->lsi_fstype,
629 obd->obd_name, PTR_ERR(obd->obd_fsops));
630 RETURN(PTR_ERR(obd->obd_fsops));
633 cli->cl_mgc_vfsmnt = mnt;
634 err = fsfilt_setup(obd, mnt->mnt_sb);
638 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
639 obd->obd_lvfs_ctxt.pwdmnt = mnt;
640 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
641 obd->obd_lvfs_ctxt.fs = get_ds();
643 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
644 dentry = ll_lookup_one_len(MOUNT_CONFIGS_DIR, cfs_fs_pwd(current->fs),
645 strlen(MOUNT_CONFIGS_DIR));
646 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
647 if (IS_ERR(dentry)) {
648 err = PTR_ERR(dentry);
649 CERROR("cannot lookup %s directory: rc = %d\n",
650 MOUNT_CONFIGS_DIR, err);
653 cli->cl_mgc_configs_dir = dentry;
655 /* We take an obd ref to insure that we can't get to mgc_cleanup
656 without calling mgc_fs_cleanup first. */
657 class_incref(obd, "mgc_fs", obd);
659 label = fsfilt_get_label(obd, mnt->mnt_sb);
661 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
663 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
667 fsfilt_put_ops(obd->obd_fsops);
668 obd->obd_fsops = NULL;
669 cli->cl_mgc_vfsmnt = NULL;
670 up(&cli->cl_mgc_sem);
674 static int mgc_fs_cleanup(struct obd_device *obd)
676 struct client_obd *cli = &obd->u.cli;
680 LASSERT(cli->cl_mgc_vfsmnt != NULL);
682 if (cli->cl_mgc_configs_dir != NULL) {
683 struct lvfs_run_ctxt saved;
684 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
685 l_dput(cli->cl_mgc_configs_dir);
686 cli->cl_mgc_configs_dir = NULL;
687 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
688 class_decref(obd, "mgc_fs", obd);
691 cli->cl_mgc_vfsmnt = NULL;
693 fsfilt_put_ops(obd->obd_fsops);
695 up(&cli->cl_mgc_sem);
700 static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0);
701 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
707 case OBD_CLEANUP_EARLY:
709 case OBD_CLEANUP_EXPORTS:
710 if (cfs_atomic_dec_and_test(&mgc_count)) {
712 /* stop requeue thread */
713 spin_lock(&config_list_lock);
714 running = rq_state & RQ_RUNNING;
717 spin_unlock(&config_list_lock);
719 cfs_waitq_signal(&rq_waitq);
720 wait_for_completion(&rq_exit);
723 obd_cleanup_client_import(obd);
724 rc = obd_llog_finish(obd, 0);
726 CERROR("failed to cleanup llogging subsystems\n");
732 static int mgc_cleanup(struct obd_device *obd)
734 struct client_obd *cli = &obd->u.cli;
738 LASSERT(cli->cl_mgc_vfsmnt == NULL);
740 /* COMPAT_146 - old config logs may have added profiles we don't
742 if (obd->obd_type->typ_refcnt <= 1)
743 /* Only for the last mgc */
744 class_del_profiles();
746 lprocfs_obd_cleanup(obd);
749 rc = client_obd_cleanup(obd);
753 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
755 struct lprocfs_static_vars lvars;
761 rc = client_obd_setup(obd, lcfg);
763 GOTO(err_decref, rc);
765 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
767 CERROR("failed to setup llogging subsystems\n");
768 GOTO(err_cleanup, rc);
771 lprocfs_mgc_init_vars(&lvars);
772 lprocfs_obd_setup(obd, lvars.obd_vars);
773 sptlrpc_lprocfs_cliobd_attach(obd);
775 if (cfs_atomic_inc_return(&mgc_count) == 1) {
777 cfs_waitq_init(&rq_waitq);
779 /* start requeue thread */
780 rc = PTR_ERR(kthread_run(mgc_requeue_thread, NULL,
782 if (IS_ERR_VALUE(rc)) {
783 CERROR("%s: Cannot start requeue thread (%d),"
784 "no more log updates!\n",
786 GOTO(err_cleanup, rc);
788 /* rc is the task_struct pointer of mgc_requeue_thread. */
795 client_obd_cleanup(obd);
801 /* based on ll_mdc_blocking_ast */
802 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
803 void *data, int flag)
805 struct lustre_handle lockh;
806 struct config_llog_data *cld = (struct config_llog_data *)data;
811 case LDLM_CB_BLOCKING:
812 /* mgs wants the lock, give it up... */
813 LDLM_DEBUG(lock, "MGC blocking CB");
814 ldlm_lock2handle(lock, &lockh);
815 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
817 case LDLM_CB_CANCELING:
818 /* We've given up the lock, prepare ourselves to update. */
819 LDLM_DEBUG(lock, "MGC cancel CB");
821 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
822 lock->l_resource->lr_name.name[0],
823 (char *)&lock->l_resource->lr_name.name[0]);
826 CDEBUG(D_INFO, "missing data, won't requeue\n");
830 /* held at mgc_process_log(). */
831 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
832 /* Are we done with this log? */
833 if (cld->cld_stopping) {
834 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
839 /* Make sure not to re-enqueue when the mgc is stopping
840 (we get called from client_disconnect_export) */
841 if (!lock->l_conn_export ||
842 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
843 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
850 mgc_requeue_add(cld);
860 /* Not sure where this should go... */
861 #define MGC_ENQUEUE_LIMIT 50
862 #define MGC_TARGET_REG_LIMIT 10
863 #define MGC_SEND_PARAM_LIMIT 10
865 /* Send parameter to MGS*/
866 static int mgc_set_mgs_param(struct obd_export *exp,
867 struct mgs_send_param *msp)
869 struct ptlrpc_request *req;
870 struct mgs_send_param *req_msp, *rep_msp;
874 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
875 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
880 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
882 ptlrpc_req_finished(req);
886 memcpy(req_msp, msp, sizeof(*req_msp));
887 ptlrpc_request_set_replen(req);
889 /* Limit how long we will wait for the enqueue to complete */
890 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
891 rc = ptlrpc_queue_wait(req);
893 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
894 memcpy(msp, rep_msp, sizeof(*rep_msp));
897 ptlrpc_req_finished(req);
902 /* Take a config lock so we can get cancel notifications */
903 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
904 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
905 __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
906 void *data, __u32 lvb_len, void *lvb_swabber,
907 struct lustre_handle *lockh)
909 struct config_llog_data *cld = (struct config_llog_data *)data;
910 struct ldlm_enqueue_info einfo = {
913 .ei_cb_bl = mgc_blocking_ast,
914 .ei_cb_cp = ldlm_completion_ast,
916 struct ptlrpc_request *req;
917 int short_limit = cld_is_sptlrpc(cld);
921 CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
922 cld->cld_resid.name[0]);
924 /* We need a callback for every lockholder, so don't try to
925 ldlm_lock_match (see rev 1.1.2.11.2.47) */
926 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
927 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
932 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
933 ptlrpc_request_set_replen(req);
935 /* check if this is server or client */
936 if (cld->cld_cfg.cfg_sb) {
937 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
938 if (lsi && IS_SERVER(lsi))
941 /* Limit how long we will wait for the enqueue to complete */
942 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
943 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
944 NULL, 0, LVB_T_NONE, lockh, 0);
945 /* A failed enqueue should still call the mgc_blocking_ast,
946 where it will be requeued if needed ("grant failed"). */
947 ptlrpc_req_finished(req);
951 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
952 __u32 mode, struct lustre_handle *lockh)
956 ldlm_lock_decref(lockh, mode);
961 static void mgc_notify_active(struct obd_device *unused)
963 /* wakeup mgc_requeue_thread to requeue mgc lock */
964 spin_lock(&config_list_lock);
966 spin_unlock(&config_list_lock);
967 cfs_waitq_signal(&rq_waitq);
969 /* TODO: Help the MGS rebuild nidtbl. -jay */
972 /* Send target_reg message to MGS */
973 static int mgc_target_register(struct obd_export *exp,
974 struct mgs_target_info *mti)
976 struct ptlrpc_request *req;
977 struct mgs_target_info *req_mti, *rep_mti;
981 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
982 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
987 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
989 ptlrpc_req_finished(req);
993 memcpy(req_mti, mti, sizeof(*req_mti));
994 ptlrpc_request_set_replen(req);
995 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
996 /* Limit how long we will wait for the enqueue to complete */
997 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
999 rc = ptlrpc_queue_wait(req);
1001 rep_mti = req_capsule_server_get(&req->rq_pill,
1002 &RMF_MGS_TARGET_INFO);
1003 memcpy(mti, rep_mti, sizeof(*rep_mti));
1004 CDEBUG(D_MGC, "register %s got index = %d\n",
1005 mti->mti_svname, mti->mti_stripe_index);
1007 ptlrpc_req_finished(req);
1012 int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1013 obd_count keylen, void *key, obd_count vallen,
1014 void *val, struct ptlrpc_request_set *set)
1019 /* Turn off initial_recov after we try all backup servers once */
1020 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1021 struct obd_import *imp = class_exp2cliimp(exp);
1023 if (vallen != sizeof(int))
1025 value = *(int *)val;
1026 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1027 imp->imp_obd->obd_name, value,
1028 imp->imp_deactive, imp->imp_invalid,
1029 imp->imp_replayable, imp->imp_obd->obd_replayable,
1030 ptlrpc_import_state_name(imp->imp_state));
1031 /* Resurrect if we previously died */
1032 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1033 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1034 ptlrpc_reconnect_import(imp);
1037 /* FIXME move this to mgc_process_config */
1038 if (KEY_IS(KEY_REGISTER_TARGET)) {
1039 struct mgs_target_info *mti;
1040 if (vallen != sizeof(struct mgs_target_info))
1042 mti = (struct mgs_target_info *)val;
1043 CDEBUG(D_MGC, "register_target %s %#x\n",
1044 mti->mti_svname, mti->mti_flags);
1045 rc = mgc_target_register(exp, mti);
1048 if (KEY_IS(KEY_SET_FS)) {
1049 struct super_block *sb = (struct super_block *)val;
1050 struct lustre_sb_info *lsi;
1051 if (vallen != sizeof(struct super_block))
1054 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
1056 CERROR("set_fs got %d\n", rc);
1060 if (KEY_IS(KEY_CLEAR_FS)) {
1063 rc = mgc_fs_cleanup(exp->exp_obd);
1065 CERROR("clear_fs got %d\n", rc);
1069 if (KEY_IS(KEY_SET_INFO)) {
1070 struct mgs_send_param *msp;
1072 msp = (struct mgs_send_param *)val;
1073 rc = mgc_set_mgs_param(exp, msp);
1076 if (KEY_IS(KEY_MGSSEC)) {
1077 struct client_obd *cli = &exp->exp_obd->u.cli;
1078 struct sptlrpc_flavor flvr;
1081 * empty string means using current flavor, if which haven't
1082 * been set yet, set it as null.
1084 * if flavor has been set previously, check the asking flavor
1085 * must match the existing one.
1088 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1094 rc = sptlrpc_parse_flavor(val, &flvr);
1096 CERROR("invalid sptlrpc flavor %s to MGS\n",
1102 * caller already hold a mutex
1104 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1105 cli->cl_flvr_mgc = flvr;
1106 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1107 sizeof(flvr)) != 0) {
1110 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1112 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1113 "currently %s is in use\n",
1123 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1124 __u32 keylen, void *key, __u32 *vallen, void *val,
1125 struct lov_stripe_md *unused)
1129 if (KEY_IS(KEY_CONN_DATA)) {
1130 struct obd_import *imp = class_exp2cliimp(exp);
1131 struct obd_connect_data *data = val;
1133 if (*vallen == sizeof(*data)) {
1134 *data = imp->imp_connect_data;
1142 static int mgc_import_event(struct obd_device *obd,
1143 struct obd_import *imp,
1144 enum obd_import_event event)
1148 LASSERT(imp->imp_obd == obd);
1149 CDEBUG(D_MGC, "import event %#x\n", event);
1152 case IMP_EVENT_DISCON:
1153 /* MGC imports should not wait for recovery */
1154 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1155 ptlrpc_pinger_ir_down();
1157 case IMP_EVENT_INACTIVE:
1159 case IMP_EVENT_INVALIDATE: {
1160 struct ldlm_namespace *ns = obd->obd_namespace;
1161 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1164 case IMP_EVENT_ACTIVE:
1165 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1166 /* Clearing obd_no_recov allows us to continue pinging */
1167 obd->obd_no_recov = 0;
1168 mgc_notify_active(obd);
1169 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1170 ptlrpc_pinger_ir_up();
1174 case IMP_EVENT_DEACTIVATE:
1175 case IMP_EVENT_ACTIVATE:
1178 CERROR("Unknown import event %#x\n", event);
1184 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1185 struct obd_device *tgt, int *index)
1187 struct llog_ctxt *ctxt;
1191 LASSERT(olg == &obd->obd_olg);
1193 #ifdef HAVE_LDISKFS_OSD
1194 rc = llog_setup(NULL, obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt,
1200 rc = llog_setup(NULL, obd, olg, LLOG_CONFIG_REPL_CTXT, tgt,
1205 ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_REPL_CTXT);
1207 GOTO(out, rc = -ENODEV);
1209 llog_initiator_connect(ctxt);
1210 llog_ctxt_put(ctxt);
1214 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1216 llog_cleanup(NULL, ctxt);
1220 static int mgc_llog_finish(struct obd_device *obd, int count)
1222 struct llog_ctxt *ctxt;
1226 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1228 llog_cleanup(NULL, ctxt);
1230 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1232 llog_cleanup(NULL, ctxt);
1237 CONFIG_READ_NRPAGES_INIT = 1 << (20 - CFS_PAGE_SHIFT),
1238 CONFIG_READ_NRPAGES = 4
1241 static int mgc_apply_recover_logs(struct obd_device *mgc,
1242 struct config_llog_data *cld,
1244 void *data, int datalen, bool mne_swab)
1246 struct config_llog_instance *cfg = &cld->cld_cfg;
1247 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1248 struct mgs_nidtbl_entry *entry;
1249 struct lustre_cfg *lcfg;
1250 struct lustre_cfg_bufs bufs;
1251 u64 prev_version = 0;
1260 LASSERT(cfg->cfg_instance != NULL);
1261 LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1263 OBD_ALLOC(inst, CFS_PAGE_SIZE);
1267 if (!IS_SERVER(lsi)) {
1268 pos = snprintf(inst, CFS_PAGE_SIZE, "%p", cfg->cfg_instance);
1269 if (pos >= CFS_PAGE_SIZE) {
1270 OBD_FREE(inst, CFS_PAGE_SIZE);
1274 LASSERT(IS_MDT(lsi));
1275 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1278 OBD_FREE(inst, CFS_PAGE_SIZE);
1286 bufsz = CFS_PAGE_SIZE - pos;
1288 while (datalen > 0) {
1289 int entry_len = sizeof(*entry);
1291 struct obd_device *obd;
1298 if (datalen < sizeof(*entry))
1301 entry = (typeof(entry))(data + off);
1304 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1306 if (entry->mne_nid_count == 0) /* at least one nid entry */
1308 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1311 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1312 if (datalen < entry_len) /* must have entry_len at least */
1315 /* Keep this swab for normal mixed endian handling. LU-1644 */
1317 lustre_swab_mgs_nidtbl_entry(entry);
1318 if (entry->mne_length > CFS_PAGE_SIZE) {
1319 CERROR("MNE too large (%u)\n", entry->mne_length);
1323 if (entry->mne_length < entry_len)
1326 off += entry->mne_length;
1327 datalen -= entry->mne_length;
1331 if (entry->mne_version > max_version) {
1332 CERROR("entry index(%lld) is over max_index(%lld)\n",
1333 entry->mne_version, max_version);
1337 if (prev_version >= entry->mne_version) {
1338 CERROR("index unsorted, prev %lld, now %lld\n",
1339 prev_version, entry->mne_version);
1342 prev_version = entry->mne_version;
1345 * Write a string with format "nid::instance" to
1346 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1349 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1350 memset(buf, 0, bufsz);
1354 /* lustre-OST0001-osc-<instance #> */
1355 strcpy(obdname, cld->cld_logname);
1356 cname = strrchr(obdname, '-');
1357 if (cname == NULL) {
1358 CERROR("mgc %s: invalid logname %s\n",
1359 mgc->obd_name, obdname);
1363 pos = cname - obdname;
1365 pos += sprintf(obdname + pos, "-%s%04x",
1366 is_ost ? "OST" : "MDT", entry->mne_index);
1368 cname = is_ost ? "osc" : "mdc",
1369 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1370 lustre_cfg_bufs_reset(&bufs, obdname);
1372 /* find the obd by obdname */
1373 obd = class_name2obd(obdname);
1375 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1376 mgc->obd_name, obdname);
1378 /* this is a safe race, when the ost is starting up...*/
1382 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1385 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1388 down_read(&obd->u.cli.cl_sem);
1389 if (obd->u.cli.cl_import == NULL) {
1390 /* client does not connect to the OST yet */
1391 up_read(&obd->u.cli.cl_sem);
1396 /* TODO: iterate all nids to find one */
1397 /* find uuid by nid */
1398 rc = client_import_find_conn(obd->u.cli.cl_import,
1400 (struct obd_uuid *)uuid);
1401 up_read(&obd->u.cli.cl_sem);
1403 CERROR("mgc: cannot find uuid by nid %s\n",
1404 libcfs_nid2str(entry->u.nids[0]));
1408 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1409 uuid, libcfs_nid2str(entry->u.nids[0]));
1411 pos += strlen(uuid);
1412 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1413 LASSERT(pos < bufsz);
1415 lustre_cfg_bufs_set_string(&bufs, 1, params);
1418 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1420 CERROR("mgc: cannot allocate memory\n");
1424 CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n",
1425 prev_version, max_version, obdname, params);
1427 rc = class_process_config(lcfg);
1428 lustre_cfg_free(lcfg);
1430 CDEBUG(D_INFO, "process config for %s error %d\n",
1433 /* continue, even one with error */
1436 OBD_FREE(inst, CFS_PAGE_SIZE);
1441 * This function is called if this client was notified for target restarting
1442 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1444 static int mgc_process_recover_log(struct obd_device *obd,
1445 struct config_llog_data *cld)
1447 struct ptlrpc_request *req = NULL;
1448 struct config_llog_instance *cfg = &cld->cld_cfg;
1449 struct mgs_config_body *body;
1450 struct mgs_config_res *res;
1451 struct ptlrpc_bulk_desc *desc;
1455 bool mne_swab = false;
1461 /* allocate buffer for bulk transfer.
1462 * if this is the first time for this mgs to read logs,
1463 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1464 * once; otherwise, it only reads increment of logs, this should be
1465 * small and CONFIG_READ_NRPAGES will be used.
1467 nrpages = CONFIG_READ_NRPAGES;
1468 if (cfg->cfg_last_idx == 0) /* the first time */
1469 nrpages = CONFIG_READ_NRPAGES_INIT;
1471 OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1473 GOTO(out, rc = -ENOMEM);
1475 for (i = 0; i < nrpages; i++) {
1476 pages[i] = cfs_alloc_page(CFS_ALLOC_STD);
1477 if (pages[i] == NULL)
1478 GOTO(out, rc = -ENOMEM);
1482 LASSERT(cld_is_recover(cld));
1483 LASSERT(mutex_is_locked(&cld->cld_lock));
1484 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1485 &RQF_MGS_CONFIG_READ);
1487 GOTO(out, rc = -ENOMEM);
1489 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1494 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1495 LASSERT(body != NULL);
1496 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1497 if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1498 >= sizeof(body->mcb_name))
1499 GOTO(out, rc = -E2BIG);
1500 body->mcb_offset = cfg->cfg_last_idx + 1;
1501 body->mcb_type = cld->cld_type;
1502 body->mcb_bits = CFS_PAGE_SHIFT;
1503 body->mcb_units = nrpages;
1505 /* allocate bulk transfer descriptor */
1506 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1, BULK_PUT_SINK,
1509 GOTO(out, rc = -ENOMEM);
1511 for (i = 0; i < nrpages; i++)
1512 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, CFS_PAGE_SIZE);
1514 ptlrpc_request_set_replen(req);
1515 rc = ptlrpc_queue_wait(req);
1519 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1520 if (res->mcr_size < res->mcr_offset)
1521 GOTO(out, rc = -EINVAL);
1523 /* always update the index even though it might have errors with
1524 * handling the recover logs */
1525 cfg->cfg_last_idx = res->mcr_offset;
1526 eof = res->mcr_offset == res->mcr_size;
1528 CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n",
1529 res->mcr_offset, eof == false);
1531 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1533 GOTO(out, rc = ealen);
1535 if (ealen > nrpages << CFS_PAGE_SHIFT)
1536 GOTO(out, rc = -EINVAL);
1538 if (ealen == 0) { /* no logs transferred */
1544 mne_swab = !!ptlrpc_rep_need_swab(req);
1545 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1546 /* This import flag means the server did an extra swab of IR MNE
1547 * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1548 if (unlikely(req->rq_import->imp_need_mne_swab))
1549 mne_swab = !mne_swab;
1551 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1554 for (i = 0; i < nrpages && ealen > 0; i++) {
1558 ptr = cfs_kmap(pages[i]);
1559 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1560 min_t(int, ealen, CFS_PAGE_SIZE),
1562 cfs_kunmap(pages[i]);
1564 CWARN("Process recover log %s error %d\n",
1565 cld->cld_logname, rc2);
1569 ealen -= CFS_PAGE_SIZE;
1574 ptlrpc_req_finished(req);
1576 if (rc == 0 && !eof)
1580 for (i = 0; i < nrpages; i++) {
1581 if (pages[i] == NULL)
1583 cfs_free_page(pages[i]);
1585 OBD_FREE(pages, sizeof(*pages) * nrpages);
1590 #ifdef HAVE_LDISKFS_OSD
1593 * XXX: mgc_copy_llog() does not support osd-based llogs yet
1596 /* identical to mgs_log_is_empty */
1597 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1600 struct lvfs_run_ctxt saved;
1601 struct llog_handle *llh;
1604 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1605 rc = llog_open(NULL, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1607 rc = llog_init_handle(NULL, llh, LLOG_F_IS_PLAIN, NULL);
1609 rc = llog_get_size(llh);
1610 llog_close(NULL, llh);
1611 } else if (rc == -ENOENT) {
1614 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1615 /* header is record 1 */
1619 /* Copy a remote log locally */
1620 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1621 struct llog_ctxt *lctxt, char *logname)
1623 struct llog_handle *local_llh, *remote_llh;
1624 struct obd_uuid *uuid;
1629 /* Write new log to a temp name, then vfs_rename over logname
1630 upon successful completion. */
1632 OBD_ALLOC(temp_log, strlen(logname) + 1);
1635 sprintf(temp_log, "%sT", logname);
1637 /* Make sure there's no old temp log */
1638 rc = llog_erase(NULL, lctxt, NULL, temp_log);
1639 if (rc < 0 && rc != -ENOENT)
1642 /* open local log */
1643 rc = llog_open_create(NULL, lctxt, &local_llh, NULL, temp_log);
1647 /* set the log header uuid for fun */
1648 OBD_ALLOC_PTR(uuid);
1649 obd_str2uuid(uuid, logname);
1650 rc = llog_init_handle(NULL, local_llh, LLOG_F_IS_PLAIN, uuid);
1653 GOTO(out_closel, rc);
1655 /* open remote log */
1656 rc = llog_open(NULL, rctxt, &remote_llh, NULL, logname,
1661 GOTO(out_closel, rc);
1664 rc = llog_init_handle(NULL, remote_llh, LLOG_F_IS_PLAIN, NULL);
1666 GOTO(out_closer, rc);
1668 /* Copy remote log */
1669 rc = llog_process(NULL, remote_llh, llog_copy_handler,
1670 (void *)local_llh, NULL);
1673 rc2 = llog_close(NULL, remote_llh);
1677 rc2 = llog_close(NULL, local_llh);
1681 /* We've copied the remote log to the local temp log, now
1682 replace the old local log with the temp log. */
1684 struct client_obd *cli = &obd->u.cli;
1687 LASSERT(cli->cl_mgc_configs_dir);
1688 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1691 CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1694 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1695 OBD_FREE(temp_log, strlen(logname) + 1);
1700 /* local_only means it cannot get remote llogs */
1701 static int mgc_process_cfg_log(struct obd_device *mgc,
1702 struct config_llog_data *cld,
1705 struct llog_ctxt *ctxt, *lctxt = NULL;
1706 #ifdef HAVE_LDISKFS_OSD
1707 struct client_obd *cli = &mgc->u.cli;
1709 struct lvfs_run_ctxt *saved_ctxt;
1710 struct lustre_sb_info *lsi = NULL;
1711 int rc = 0, must_pop = 0;
1712 bool sptlrpc_started = false;
1717 LASSERT(mutex_is_locked(&cld->cld_lock));
1720 * local copy of sptlrpc log is controlled elsewhere, don't try to
1723 if (cld_is_sptlrpc(cld) && local_only)
1726 if (cld->cld_cfg.cfg_sb)
1727 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1729 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1731 CERROR("missing llog context\n");
1735 OBD_ALLOC_PTR(saved_ctxt);
1736 if (saved_ctxt == NULL)
1739 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1741 #ifdef HAVE_LDISKFS_OSD
1743 * XXX: at the moment mgc_copy_llog() works with lvfs-based llogs
1745 /* Copy the setup log locally if we can. Don't mess around if we're
1746 running an MGS though (logs are already local). */
1747 if (lctxt && lsi && IS_SERVER(lsi) &&
1748 (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1749 !IS_MGS(lsi) && lsi->lsi_srv_mnt) {
1750 push_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1753 /* Only try to copy log if we have the lock. */
1754 rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1755 if (local_only || rc) {
1756 if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1757 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1758 "log %s and no local copy."
1759 "\n", cld->cld_logname);
1760 GOTO(out_pop, rc = -ENOTCONN);
1762 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1763 "copy for now, will try to update later.\n",
1766 /* Now, whether we copied or not, start using the local llog.
1767 If we failed to copy, we'll start using whatever the old
1769 llog_ctxt_put(ctxt);
1774 if (local_only) { /* no local log at client side */
1775 GOTO(out_pop, rc = -EIO);
1778 if (cld_is_sptlrpc(cld)) {
1779 sptlrpc_conf_log_update_begin(cld->cld_logname);
1780 sptlrpc_started = true;
1783 /* logname and instance info should be the same, so use our
1784 copy of the instance for the update. The cfg_last_idx will
1786 rc = class_config_parse_llog(NULL, ctxt, cld->cld_logname,
1791 llog_ctxt_put(ctxt);
1793 llog_ctxt_put(lctxt);
1795 pop_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1797 OBD_FREE_PTR(saved_ctxt);
1799 * update settings on existing OBDs. doing it inside
1800 * of llog_process_lock so no device is attaching/detaching
1802 * the logname must be <fsname>-sptlrpc
1804 if (sptlrpc_started) {
1805 LASSERT(cld_is_sptlrpc(cld));
1806 sptlrpc_conf_log_update_end(cld->cld_logname);
1807 class_notify_sptlrpc_conf(cld->cld_logname,
1808 strlen(cld->cld_logname) -
1809 strlen("-sptlrpc"));
1815 /** Get a config log from the MGS and process it.
1816 * This func is called for both clients and servers.
1817 * Copy the log locally before parsing it if appropriate (non-MGS server)
1819 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1821 struct lustre_handle lockh = { 0 };
1822 __u64 flags = LDLM_FL_NO_LRU;
1828 /* I don't want multiple processes running process_log at once --
1829 sounds like badness. It actually might be fine, as long as
1830 we're not trying to update from the same log
1831 simultaneously (in which case we should use a per-log sem.) */
1832 mutex_lock(&cld->cld_lock);
1833 if (cld->cld_stopping) {
1834 mutex_unlock(&cld->cld_lock);
1838 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1840 CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1841 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1843 /* Get the cfg lock on the llog */
1844 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1845 LCK_CR, &flags, NULL, NULL, NULL,
1846 cld, 0, NULL, &lockh);
1848 /* Get the cld, it will be released in mgc_blocking_ast. */
1849 config_log_get(cld);
1850 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1853 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1855 /* mark cld_lostlock so that it will requeue
1856 * after MGC becomes available. */
1857 cld->cld_lostlock = 1;
1858 /* Get extra reference, it will be put in requeue thread */
1859 config_log_get(cld);
1863 if (cld_is_recover(cld)) {
1864 rc = 0; /* this is not a fatal error for recover log */
1866 rc = mgc_process_recover_log(mgc, cld);
1868 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1871 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1872 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1874 mutex_unlock(&cld->cld_lock);
1876 /* Now drop the lock so MGS can revoke it */
1878 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1881 CERROR("Can't drop cfg lock: %d\n", rcl);
1888 /** Called from lustre_process_log.
1889 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1890 * any services, and adds it to the list logs to watch (follow).
1892 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1894 struct lustre_cfg *lcfg = buf;
1895 struct config_llog_instance *cfg = NULL;
1900 switch(lcfg->lcfg_command) {
1901 case LCFG_LOV_ADD_OBD: {
1902 /* Overloading this cfg command: register a new target */
1903 struct mgs_target_info *mti;
1905 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1906 sizeof(struct mgs_target_info))
1907 GOTO(out, rc = -EINVAL);
1909 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1910 CDEBUG(D_MGC, "add_target %s %#x\n",
1911 mti->mti_svname, mti->mti_flags);
1912 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1915 case LCFG_LOV_DEL_OBD:
1916 /* Unregister has no meaning at the moment. */
1917 CERROR("lov_del_obd unimplemented\n");
1920 case LCFG_SPTLRPC_CONF: {
1921 rc = sptlrpc_process_config(lcfg);
1924 case LCFG_LOG_START: {
1925 struct config_llog_data *cld;
1926 struct super_block *sb;
1928 logname = lustre_cfg_string(lcfg, 1);
1929 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1930 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1932 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1935 /* We're only called through here on the initial mount */
1936 rc = config_log_add(obd, logname, cfg, sb);
1939 cld = config_log_find(logname, cfg);
1946 /* FIXME only set this for old logs! Right now this forces
1947 us to always skip the "inside markers" check */
1948 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1950 rc = mgc_process_log(obd, cld);
1951 if (rc == 0 && cld->cld_recover != NULL) {
1952 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1953 imp_connect_data, IMP_RECOV)) {
1954 rc = mgc_process_log(obd, cld->cld_recover);
1956 struct config_llog_data *cir = cld->cld_recover;
1957 cld->cld_recover = NULL;
1958 config_log_put(cir);
1961 CERROR("Cannot process recover llog %d\n", rc);
1963 config_log_put(cld);
1967 case LCFG_LOG_END: {
1968 logname = lustre_cfg_string(lcfg, 1);
1970 if (lcfg->lcfg_bufcount >= 2)
1971 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1973 rc = config_log_end(logname, cfg);
1977 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1978 GOTO(out, rc = -EINVAL);
1986 struct obd_ops mgc_obd_ops = {
1987 .o_owner = THIS_MODULE,
1988 .o_setup = mgc_setup,
1989 .o_precleanup = mgc_precleanup,
1990 .o_cleanup = mgc_cleanup,
1991 .o_add_conn = client_import_add_conn,
1992 .o_del_conn = client_import_del_conn,
1993 .o_connect = client_connect_import,
1994 .o_disconnect = client_disconnect_export,
1995 //.o_enqueue = mgc_enqueue,
1996 .o_cancel = mgc_cancel,
1997 //.o_iocontrol = mgc_iocontrol,
1998 .o_set_info_async = mgc_set_info_async,
1999 .o_get_info = mgc_get_info,
2000 .o_import_event = mgc_import_event,
2001 .o_llog_init = mgc_llog_init,
2002 .o_llog_finish = mgc_llog_finish,
2003 .o_process_config = mgc_process_config,
2006 int __init mgc_init(void)
2008 return class_register_type(&mgc_obd_ops, NULL, NULL,
2009 LUSTRE_MGC_NAME, NULL);
2013 static void /*__exit*/ mgc_exit(void)
2015 class_unregister_type(LUSTRE_MGC_NAME);
2018 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2019 MODULE_DESCRIPTION("Lustre Management Client");
2020 MODULE_LICENSE("GPL");
2022 module_init(mgc_init);
2023 module_exit(mgc_exit);