1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
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>
42 # define EXPORT_SYMTAB
44 #define DEBUG_SUBSYSTEM S_MGC
45 #define D_MGC D_CONFIG /*|D_WARNING*/
48 # include <linux/module.h>
49 # include <linux/pagemap.h>
50 # include <linux/miscdevice.h>
51 # include <linux/init.h>
53 # include <liblustre.h>
56 #include <obd_class.h>
57 #include <lustre_dlm.h>
58 #include <lprocfs_status.h>
59 #include <lustre_log.h>
60 #include <lustre_fsfilt.h>
61 #include <lustre_disk.h>
62 #include "mgc_internal.h"
64 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id)
69 CERROR("name too long: %s\n", name);
73 CERROR("missing name: %s\n", name);
76 memcpy(&resname, name, len);
78 memset(res_id, 0, sizeof(*res_id));
80 /* Always use the same endianness for the resid */
81 res_id->name[0] = cpu_to_le64(resname);
82 CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
83 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
87 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id)
89 /* fsname is at most 8 chars long, maybe contain "-".
90 * e.g. "lustre", "SUN-000" */
91 return mgc_name2resid(fsname, strlen(fsname), res_id);
93 EXPORT_SYMBOL(mgc_fsname2resid);
95 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id)
100 /* logname consists of "fsname-nodetype".
101 * e.g. "lustre-MDT0001", "SUN-000-client" */
102 name_end = strrchr(logname, '-');
104 len = name_end - logname;
105 return mgc_name2resid(logname, len, res_id);
108 /********************** config llog list **********************/
109 static CFS_LIST_HEAD(config_llog_list);
110 static cfs_spinlock_t config_list_lock = CFS_SPIN_LOCK_UNLOCKED;
112 /* Take a reference to a config log */
113 static int config_log_get(struct config_llog_data *cld)
116 if (cld->cld_stopping)
118 cfs_atomic_inc(&cld->cld_refcount);
119 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
120 cfs_atomic_read(&cld->cld_refcount));
124 /* Drop a reference to a config log. When no longer referenced,
125 we can free the config log data */
126 static void config_log_put(struct config_llog_data *cld)
130 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
131 cfs_atomic_read(&cld->cld_refcount));
132 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
134 /* spinlock to make sure no item with 0 refcount in the list */
135 cfs_spin_lock(&config_list_lock);
136 if (unlikely(cfs_atomic_dec_and_test(&cld->cld_refcount))) {
137 cfs_list_del(&cld->cld_list_chain);
138 cfs_spin_unlock(&config_list_lock);
140 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
142 if (cld->cld_sptlrpc)
143 config_log_put(cld->cld_sptlrpc);
144 if (cld->cld_is_sptlrpc)
145 sptlrpc_conf_log_stop(cld->cld_logname);
147 class_export_put(cld->cld_mgcexp);
148 OBD_FREE(cld->cld_logname, strlen(cld->cld_logname) + 1);
149 if (cld->cld_cfg.cfg_instance != NULL)
150 OBD_FREE(cld->cld_cfg.cfg_instance,
151 strlen(cld->cld_cfg.cfg_instance) + 1);
152 OBD_FREE(cld, sizeof(*cld));
154 cfs_spin_unlock(&config_list_lock);
160 /* Find a config log by name */
162 struct config_llog_data *config_log_find(char *logname,
163 struct config_llog_instance *cfg)
165 struct config_llog_data *cld;
166 char *logid = logname;
167 int match_instance = 0;
170 if (cfg && cfg->cfg_instance) {
172 logid = cfg->cfg_instance;
175 CERROR("No log specified\n");
176 RETURN(ERR_PTR(-EINVAL));
179 cfs_spin_lock(&config_list_lock);
180 cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
181 if (match_instance && cld->cld_cfg.cfg_instance &&
182 strcmp(logid, cld->cld_cfg.cfg_instance) == 0)
184 if (!match_instance &&
185 strcmp(logid, cld->cld_logname) == 0)
188 cfs_spin_unlock(&config_list_lock);
190 CDEBUG(D_CONFIG, "can't get log %s\n", logid);
191 RETURN(ERR_PTR(-ENOENT));
193 cfs_atomic_inc(&cld->cld_refcount);
194 cfs_spin_unlock(&config_list_lock);
195 LASSERT(cld->cld_stopping == 0 || cld->cld_is_sptlrpc == 0);
200 struct config_llog_data *do_config_log_add(struct obd_device *obd,
202 unsigned int is_sptlrpc,
203 struct config_llog_instance *cfg,
204 struct super_block *sb)
206 struct config_llog_data *cld;
210 CDEBUG(D_MGC, "do adding config log %s:%s\n", logname,
211 cfg ? cfg->cfg_instance : "NULL");
213 OBD_ALLOC(cld, sizeof(*cld));
215 RETURN(ERR_PTR(-ENOMEM));
216 OBD_ALLOC(cld->cld_logname, strlen(logname) + 1);
217 if (!cld->cld_logname) {
218 OBD_FREE(cld, sizeof(*cld));
219 RETURN(ERR_PTR(-ENOMEM));
221 strcpy(cld->cld_logname, logname);
224 cld->cld_cfg.cfg_last_idx = 0;
225 cld->cld_cfg.cfg_flags = 0;
226 cld->cld_cfg.cfg_sb = sb;
227 cld->cld_is_sptlrpc = is_sptlrpc;
228 cfs_atomic_set(&cld->cld_refcount, 1);
230 /* Keep the mgc around until we are done */
231 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
233 if (cfg && cfg->cfg_instance != NULL) {
234 OBD_ALLOC(cld->cld_cfg.cfg_instance,
235 strlen(cfg->cfg_instance) + 1);
236 strcpy(cld->cld_cfg.cfg_instance, cfg->cfg_instance);
240 sptlrpc_conf_log_start(logname);
241 cld->cld_cfg.cfg_obdname = obd->obd_name;
244 rc = mgc_logname2resid(logname, &cld->cld_resid);
246 cfs_spin_lock(&config_list_lock);
247 cfs_list_add(&cld->cld_list_chain, &config_llog_list);
248 cfs_spin_unlock(&config_list_lock);
256 rc = mgc_process_log(obd, cld);
258 CERROR("failed processing sptlrpc log: %d\n", rc);
264 /** Add this log to the list of active logs watched by an MGC.
265 * Active means we're watching for updates.
266 * We have one active log per "mount" - client instance or servername.
267 * Each instance may be at a different point in the log.
269 static int config_log_add(struct obd_device *obd, char *logname,
270 struct config_llog_instance *cfg,
271 struct super_block *sb)
273 struct config_llog_data *cld, *sptlrpc_cld;
278 CDEBUG(D_MGC, "adding config log %s:%s\n", logname, cfg->cfg_instance);
281 * for each regular log, the depended sptlrpc log name is
282 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
284 ptr = strrchr(logname, '-');
285 if (ptr == NULL || ptr - logname > 8) {
286 CERROR("logname %s is too long\n", logname);
290 memcpy(seclogname, logname, ptr - logname);
291 strcpy(seclogname + (ptr - logname), "-sptlrpc");
293 sptlrpc_cld = config_log_find(seclogname, NULL);
294 if (IS_ERR(sptlrpc_cld)) {
295 sptlrpc_cld = do_config_log_add(obd, seclogname, 1, NULL, NULL);
296 if (IS_ERR(sptlrpc_cld)) {
297 CERROR("can't create sptlrpc log: %s\n", seclogname);
298 RETURN(PTR_ERR(sptlrpc_cld));
302 cld = do_config_log_add(obd, logname, 0, cfg, sb);
304 CERROR("can't create log: %s\n", logname);
305 config_log_put(sptlrpc_cld);
306 RETURN(PTR_ERR(cld));
309 cld->cld_sptlrpc = sptlrpc_cld;
314 CFS_DECLARE_MUTEX(llog_process_lock);
316 /** Stop watching for updates on this log.
318 static int config_log_end(char *logname, struct config_llog_instance *cfg)
320 struct config_llog_data *cld, *cld_sptlrpc = NULL;
324 cld = config_log_find(logname, cfg);
326 RETURN(PTR_ERR(cld));
328 cfs_down(&llog_process_lock);
330 * if cld_stopping is set, it means we didn't start the log thus
331 * not owning the start ref. this can happen after previous umount:
332 * the cld still hanging there waiting for lock cancel, and we
333 * remount again but failed in the middle and call log_end without
336 if (unlikely(cld->cld_stopping)) {
337 cfs_up(&llog_process_lock);
338 /* drop the ref from the find */
343 cld->cld_stopping = 1;
344 cfs_up(&llog_process_lock);
346 cfs_spin_lock(&config_list_lock);
347 cld_sptlrpc = cld->cld_sptlrpc;
348 cld->cld_sptlrpc = NULL;
349 cfs_spin_unlock(&config_list_lock);
352 config_log_put(cld_sptlrpc);
354 /* drop the ref from the find */
356 /* drop the start ref */
359 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
364 /* reenqueue any lost locks */
365 #define RQ_RUNNING 0x1
369 static int rq_state = 0;
370 static cfs_waitq_t rq_waitq;
372 static int mgc_requeue_add(struct config_llog_data *cld, int later);
374 static void do_requeue(struct config_llog_data *cld)
376 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
378 /* Do not run mgc_process_log on a disconnected export or an
379 export which is being disconnected. Take the client
380 semaphore to make the check non-racy. */
381 cfs_down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
382 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
383 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
384 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
386 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
389 cfs_up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
391 /* Whether we enqueued again or not in mgc_process_log, we're done
392 * with the ref from the old enqueue */
396 static int mgc_requeue_thread(void *data)
398 struct l_wait_info lwi_now, lwi_later;
399 struct config_llog_data *cld, *cld_next, *cld_prev;
400 char name[] = "ll_cfg_requeue";
406 CDEBUG(D_MGC, "Starting requeue thread\n");
408 lwi_later = LWI_TIMEOUT(60 * CFS_HZ, NULL, NULL);
409 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP), &lwi_later);
411 /* Keep trying failed locks periodically */
412 cfs_spin_lock(&config_list_lock);
413 while (rq_state & (RQ_NOW | RQ_LATER)) {
414 /* Any new or requeued lostlocks will change the state */
415 rq_state &= ~(RQ_NOW | RQ_LATER);
416 cfs_spin_unlock(&config_list_lock);
418 /* Always wait a few seconds to allow the server who
419 caused the lock revocation to finish its setup, plus some
420 random so everyone doesn't try to reconnect at once. */
421 lwi_now = LWI_TIMEOUT(3 * CFS_HZ + (ll_rand() & 0xff) * \
424 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi_now);
427 * iterate & processing through the list. for each cld, process
428 * its depending sptlrpc cld firstly (if any) and then itself.
430 * it's guaranteed any item in the list must have
431 * reference > 0; and if cld_lostlock is set, at
432 * least one reference is taken by the previous enqueue.
434 * Note: releasing a cld might lead to itself and its depended
435 * sptlrpc cld be unlinked from the list. to safely iterate
436 * we need to take a reference on next cld before processing.
440 cfs_spin_lock(&config_list_lock);
441 cfs_list_for_each_entry_safe(cld, cld_next, &config_llog_list,
443 if (cld->cld_list_chain.next != &config_llog_list)
444 cfs_atomic_inc(&cld_next->cld_refcount);
446 if (cld->cld_lostlock) {
447 if (cld->cld_sptlrpc &&
448 cld->cld_sptlrpc->cld_lostlock) {
449 cld->cld_sptlrpc->cld_lostlock = 0;
451 cfs_spin_unlock(&config_list_lock);
452 do_requeue(cld->cld_sptlrpc);
453 cfs_spin_lock(&config_list_lock);
454 LASSERT(cld->cld_lostlock);
457 cld->cld_lostlock = 0;
459 cfs_spin_unlock(&config_list_lock);
461 cfs_spin_lock(&config_list_lock);
466 cfs_spin_unlock(&config_list_lock);
467 config_log_put(cld_prev);
468 cfs_spin_lock(&config_list_lock);
473 cfs_spin_unlock(&config_list_lock);
475 /* Wait a bit to see if anyone else needs a requeue */
476 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
478 cfs_spin_lock(&config_list_lock);
480 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
481 rq_state &= ~RQ_RUNNING;
482 cfs_spin_unlock(&config_list_lock);
484 CDEBUG(D_MGC, "Ending requeue thread\n");
488 /* Add a cld to the list to requeue. Start the requeue thread if needed.
489 We are responsible for dropping the config log reference from here on out. */
490 static int mgc_requeue_add(struct config_llog_data *cld, int later)
494 CDEBUG(D_INFO, "log %s: requeue (l=%d r=%d sp=%d st=%x)\n",
495 cld->cld_logname, later, cfs_atomic_read(&cld->cld_refcount),
496 cld->cld_stopping, rq_state);
497 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
499 /* Hold lock for rq_state */
500 cfs_spin_lock(&config_list_lock);
502 if (cld->cld_stopping || (rq_state & RQ_STOP)) {
503 cld->cld_lostlock = 0;
504 cfs_spin_unlock(&config_list_lock);
509 cld->cld_lostlock = 1;
511 if (!(rq_state & RQ_RUNNING)) {
512 LASSERT(rq_state == 0);
513 rq_state = RQ_RUNNING | (later ? RQ_LATER : RQ_NOW);
514 cfs_spin_unlock(&config_list_lock);
515 rc = cfs_kernel_thread(mgc_requeue_thread, 0,
516 CLONE_VM | CLONE_FILES);
518 CERROR("log %s: cannot start requeue thread (%d),"
519 "no more log updates!\n", cld->cld_logname, rc);
520 /* Drop the ref, since the rq thread won't */
521 cld->cld_lostlock = 0;
527 rq_state |= later ? RQ_LATER : RQ_NOW;
528 cfs_spin_unlock(&config_list_lock);
529 cfs_waitq_signal(&rq_waitq);
535 /********************** class fns **********************/
537 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
538 struct vfsmount *mnt)
540 struct lvfs_run_ctxt saved;
541 struct lustre_sb_info *lsi = s2lsi(sb);
542 struct client_obd *cli = &obd->u.cli;
543 struct dentry *dentry;
549 LASSERT(lsi->lsi_srv_mnt == mnt);
551 /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
552 cfs_down(&cli->cl_mgc_sem);
554 cfs_cleanup_group_info();
556 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
557 if (IS_ERR(obd->obd_fsops)) {
558 cfs_up(&cli->cl_mgc_sem);
559 CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd),
560 PTR_ERR(obd->obd_fsops));
561 RETURN(PTR_ERR(obd->obd_fsops));
564 cli->cl_mgc_vfsmnt = mnt;
565 fsfilt_setup(obd, mnt->mnt_sb);
567 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
568 obd->obd_lvfs_ctxt.pwdmnt = mnt;
569 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
570 obd->obd_lvfs_ctxt.fs = get_ds();
572 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
573 dentry = lookup_one_len(MOUNT_CONFIGS_DIR, cfs_fs_pwd(current->fs),
574 strlen(MOUNT_CONFIGS_DIR));
575 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
576 if (IS_ERR(dentry)) {
577 err = PTR_ERR(dentry);
578 CERROR("cannot lookup %s directory: rc = %d\n",
579 MOUNT_CONFIGS_DIR, err);
582 cli->cl_mgc_configs_dir = dentry;
584 /* We take an obd ref to insure that we can't get to mgc_cleanup
585 without calling mgc_fs_cleanup first. */
586 class_incref(obd, "mgc_fs", obd);
588 label = fsfilt_get_label(obd, mnt->mnt_sb);
590 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
592 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
596 fsfilt_put_ops(obd->obd_fsops);
597 obd->obd_fsops = NULL;
598 cli->cl_mgc_vfsmnt = NULL;
599 cfs_up(&cli->cl_mgc_sem);
603 static int mgc_fs_cleanup(struct obd_device *obd)
605 struct client_obd *cli = &obd->u.cli;
609 LASSERT(cli->cl_mgc_vfsmnt != NULL);
611 if (cli->cl_mgc_configs_dir != NULL) {
612 struct lvfs_run_ctxt saved;
613 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
614 l_dput(cli->cl_mgc_configs_dir);
615 cli->cl_mgc_configs_dir = NULL;
616 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
617 class_decref(obd, "mgc_fs", obd);
620 cli->cl_mgc_vfsmnt = NULL;
622 fsfilt_put_ops(obd->obd_fsops);
624 cfs_up(&cli->cl_mgc_sem);
629 static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0);
630 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
636 case OBD_CLEANUP_EARLY:
638 case OBD_CLEANUP_EXPORTS:
639 if (cfs_atomic_dec_and_test(&mgc_count)) {
640 /* Kick the requeue waitq - cld's should all be
642 cfs_spin_lock(&config_list_lock);
644 cfs_spin_unlock(&config_list_lock);
645 cfs_waitq_signal(&rq_waitq);
647 rc = obd_llog_finish(obd, 0);
649 CERROR("failed to cleanup llogging subsystems\n");
655 static int mgc_cleanup(struct obd_device *obd)
657 struct client_obd *cli = &obd->u.cli;
661 LASSERT(cli->cl_mgc_vfsmnt == NULL);
663 /* COMPAT_146 - old config logs may have added profiles we don't
665 if (obd->obd_type->typ_refcnt <= 1)
666 /* Only for the last mgc */
667 class_del_profiles();
669 lprocfs_obd_cleanup(obd);
672 rc = client_obd_cleanup(obd);
676 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
678 struct lprocfs_static_vars lvars;
684 rc = client_obd_setup(obd, lcfg);
686 GOTO(err_decref, rc);
688 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
690 CERROR("failed to setup llogging subsystems\n");
691 GOTO(err_cleanup, rc);
694 lprocfs_mgc_init_vars(&lvars);
695 lprocfs_obd_setup(obd, lvars.obd_vars);
696 sptlrpc_lprocfs_cliobd_attach(obd);
698 cfs_spin_lock(&config_list_lock);
699 cfs_atomic_inc(&mgc_count);
700 if (cfs_atomic_read(&mgc_count) == 1) {
701 rq_state &= ~RQ_STOP;
702 cfs_waitq_init(&rq_waitq);
704 cfs_spin_unlock(&config_list_lock);
709 client_obd_cleanup(obd);
715 /* based on ll_mdc_blocking_ast */
716 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
717 void *data, int flag)
719 struct lustre_handle lockh;
720 struct config_llog_data *cld = (struct config_llog_data *)data;
725 case LDLM_CB_BLOCKING:
726 /* mgs wants the lock, give it up... */
727 LDLM_DEBUG(lock, "MGC blocking CB");
728 ldlm_lock2handle(lock, &lockh);
729 rc = ldlm_cli_cancel(&lockh);
731 case LDLM_CB_CANCELING: {
732 /* We've given up the lock, prepare ourselves to update. */
733 LDLM_DEBUG(lock, "MGC cancel CB");
735 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
736 lock->l_resource->lr_name.name[0],
737 (char *)&lock->l_resource->lr_name.name[0]);
740 CERROR("missing data, won't requeue\n");
743 /* Are we done with this log? */
744 if (cld->cld_stopping) {
745 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
750 /* Make sure not to re-enqueue when the mgc is stopping
751 (we get called from client_disconnect_export) */
752 if (!lock->l_conn_export ||
753 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
754 CDEBUG(D_MGC, "log %s: disconnecting, won't requeue\n",
759 /* Did we fail to get the lock? */
760 if (lock->l_req_mode != lock->l_granted_mode &&
761 !cld->cld_is_sptlrpc) {
762 CDEBUG(D_MGC, "log %s: original grant failed, will "
763 "requeue later\n", cld->cld_logname);
764 /* Try to re-enqueue later */
765 rc = mgc_requeue_add(cld, 1);
769 rc = mgc_requeue_add(cld, 0);
778 CERROR("%s CB failed %d:\n", flag == LDLM_CB_BLOCKING ?
779 "blocking" : "cancel", rc);
780 LDLM_ERROR(lock, "MGC ast");
785 /* Not sure where this should go... */
786 #define MGC_ENQUEUE_LIMIT 50
787 #define MGC_TARGET_REG_LIMIT 10
788 #define MGC_SEND_PARAM_LIMIT 10
790 /* Send parameter to MGS*/
791 static int mgc_set_mgs_param(struct obd_export *exp,
792 struct mgs_send_param *msp)
794 struct ptlrpc_request *req;
795 struct mgs_send_param *req_msp, *rep_msp;
799 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
800 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
805 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
807 ptlrpc_req_finished(req);
811 memcpy(req_msp, msp, sizeof(*req_msp));
812 ptlrpc_request_set_replen(req);
814 /* Limit how long we will wait for the enqueue to complete */
815 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
816 rc = ptlrpc_queue_wait(req);
818 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
819 memcpy(msp, rep_msp, sizeof(*rep_msp));
822 ptlrpc_req_finished(req);
827 /* Take a config lock so we can get cancel notifications */
828 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
829 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
830 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
831 void *data, __u32 lvb_len, void *lvb_swabber,
832 struct lustre_handle *lockh)
834 struct config_llog_data *cld = (struct config_llog_data *)data;
835 struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
836 ldlm_completion_ast, NULL, NULL, data};
837 struct ptlrpc_request *req;
838 int short_limit = cld->cld_is_sptlrpc;
842 CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
843 cld->cld_resid.name[0]);
845 /* We can only drop this config log ref when we drop the lock */
846 if (config_log_get(cld))
847 RETURN(ELDLM_LOCK_ABORTED);
849 /* We need a callback for every lockholder, so don't try to
850 ldlm_lock_match (see rev 1.1.2.11.2.47) */
851 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
852 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
856 ptlrpc_request_set_replen(req);
857 /* check if this is server or client */
858 if (cld->cld_cfg.cfg_sb) {
859 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
860 if (lsi && (lsi->lsi_flags & LSI_SERVER))
863 /* Limit how long we will wait for the enqueue to complete */
864 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
865 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
867 /* A failed enqueue should still call the mgc_blocking_ast,
868 where it will be requeued if needed ("grant failed"). */
869 ptlrpc_req_finished(req);
873 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
874 __u32 mode, struct lustre_handle *lockh)
878 ldlm_lock_decref(lockh, mode);
883 /* Send target_reg message to MGS */
884 static int mgc_target_register(struct obd_export *exp,
885 struct mgs_target_info *mti)
887 struct ptlrpc_request *req;
888 struct mgs_target_info *req_mti, *rep_mti;
892 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
893 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
898 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
900 ptlrpc_req_finished(req);
904 memcpy(req_mti, mti, sizeof(*req_mti));
905 ptlrpc_request_set_replen(req);
906 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
907 /* Limit how long we will wait for the enqueue to complete */
908 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
910 rc = ptlrpc_queue_wait(req);
912 rep_mti = req_capsule_server_get(&req->rq_pill,
913 &RMF_MGS_TARGET_INFO);
914 memcpy(mti, rep_mti, sizeof(*rep_mti));
915 CDEBUG(D_MGC, "register %s got index = %d\n",
916 mti->mti_svname, mti->mti_stripe_index);
918 ptlrpc_req_finished(req);
923 int mgc_set_info_async(struct obd_export *exp, obd_count keylen,
924 void *key, obd_count vallen, void *val,
925 struct ptlrpc_request_set *set)
930 /* Turn off initial_recov after we try all backup servers once */
931 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
932 struct obd_import *imp = class_exp2cliimp(exp);
934 if (vallen != sizeof(int))
937 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
938 imp->imp_obd->obd_name, value,
939 imp->imp_deactive, imp->imp_invalid,
940 imp->imp_replayable, imp->imp_obd->obd_replayable,
941 ptlrpc_import_state_name(imp->imp_state));
942 /* Resurrect if we previously died */
943 if ((imp->imp_state != LUSTRE_IMP_FULL &&
944 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
945 ptlrpc_reconnect_import(imp);
948 /* FIXME move this to mgc_process_config */
949 if (KEY_IS(KEY_REGISTER_TARGET)) {
950 struct mgs_target_info *mti;
951 if (vallen != sizeof(struct mgs_target_info))
953 mti = (struct mgs_target_info *)val;
954 CDEBUG(D_MGC, "register_target %s %#x\n",
955 mti->mti_svname, mti->mti_flags);
956 rc = mgc_target_register(exp, mti);
959 if (KEY_IS(KEY_SET_FS)) {
960 struct super_block *sb = (struct super_block *)val;
961 struct lustre_sb_info *lsi;
962 if (vallen != sizeof(struct super_block))
965 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
967 CERROR("set_fs got %d\n", rc);
971 if (KEY_IS(KEY_CLEAR_FS)) {
974 rc = mgc_fs_cleanup(exp->exp_obd);
976 CERROR("clear_fs got %d\n", rc);
980 if (KEY_IS(KEY_SET_INFO)) {
981 struct mgs_send_param *msp;
983 msp = (struct mgs_send_param *)val;
984 rc = mgc_set_mgs_param(exp, msp);
987 if (KEY_IS(KEY_MGSSEC)) {
988 struct client_obd *cli = &exp->exp_obd->u.cli;
989 struct sptlrpc_flavor flvr;
992 * empty string means using current flavor, if which haven't
993 * been set yet, set it as null.
995 * if flavor has been set previously, check the asking flavor
996 * must match the existing one.
999 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1005 rc = sptlrpc_parse_flavor(val, &flvr);
1007 CERROR("invalid sptlrpc flavor %s to MGS\n",
1013 * caller already hold a mutex
1015 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1016 cli->cl_flvr_mgc = flvr;
1017 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1018 sizeof(flvr)) != 0) {
1021 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1023 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1024 "currently %s is in use\n",
1034 static int mgc_import_event(struct obd_device *obd,
1035 struct obd_import *imp,
1036 enum obd_import_event event)
1040 LASSERT(imp->imp_obd == obd);
1041 CDEBUG(D_MGC, "import event %#x\n", event);
1044 case IMP_EVENT_DISCON:
1045 /* MGC imports should not wait for recovery */
1047 case IMP_EVENT_INACTIVE:
1049 case IMP_EVENT_INVALIDATE: {
1050 struct ldlm_namespace *ns = obd->obd_namespace;
1051 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1054 case IMP_EVENT_ACTIVE:
1055 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
1056 /* Clearing obd_no_recov allows us to continue pinging */
1057 obd->obd_no_recov = 0;
1062 CERROR("Unknown import event %#x\n", event);
1068 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1069 struct obd_device *tgt, int *index)
1071 struct llog_ctxt *ctxt;
1075 LASSERT(olg == &obd->obd_olg);
1077 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
1082 rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
1085 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1087 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1092 llog_initiator_connect(ctxt);
1093 llog_ctxt_put(ctxt);
1095 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1103 static int mgc_llog_finish(struct obd_device *obd, int count)
1105 struct llog_ctxt *ctxt;
1106 int rc = 0, rc2 = 0;
1109 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1111 rc = llog_cleanup(ctxt);
1113 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1115 rc2 = llog_cleanup(ctxt);
1123 /* identical to mgs_log_is_empty */
1124 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1127 struct lvfs_run_ctxt saved;
1128 struct llog_handle *llh;
1131 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1132 rc = llog_create(ctxt, &llh, NULL, name);
1134 llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL);
1135 rc = llog_get_size(llh);
1138 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1139 /* header is record 1 */
1143 static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec,
1146 struct llog_rec_hdr local_rec = *rec;
1147 struct llog_handle *local_llh = (struct llog_handle *)data;
1148 char *cfg_buf = (char*) (rec + 1);
1149 struct lustre_cfg *lcfg;
1153 /* Append all records */
1154 local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
1155 rc = llog_write_rec(local_llh, &local_rec, NULL, 0,
1156 (void *)cfg_buf, -1);
1158 lcfg = (struct lustre_cfg *)cfg_buf;
1159 CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
1160 rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
1161 lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
1166 /* Copy a remote log locally */
1167 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1168 struct llog_ctxt *lctxt, char *logname)
1170 struct llog_handle *local_llh, *remote_llh;
1171 struct obd_uuid *uuid;
1176 /* Write new log to a temp name, then vfs_rename over logname
1177 upon successful completion. */
1179 OBD_ALLOC(temp_log, strlen(logname) + 1);
1182 sprintf(temp_log, "%sT", logname);
1184 /* Make sure there's no old temp log */
1185 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1188 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL);
1191 rc = llog_destroy(local_llh);
1192 llog_free_handle(local_llh);
1196 /* open local log */
1197 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1201 /* set the log header uuid for fun */
1202 OBD_ALLOC_PTR(uuid);
1203 obd_str2uuid(uuid, logname);
1204 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid);
1207 GOTO(out_closel, rc);
1209 /* open remote log */
1210 rc = llog_create(rctxt, &remote_llh, NULL, logname);
1212 GOTO(out_closel, rc);
1213 rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL);
1215 GOTO(out_closer, rc);
1217 /* Copy remote log */
1218 rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL);
1221 rc2 = llog_close(remote_llh);
1225 rc2 = llog_close(local_llh);
1229 /* We've copied the remote log to the local temp log, now
1230 replace the old local log with the temp log. */
1232 struct client_obd *cli = &obd->u.cli;
1234 LASSERT(cli->cl_mgc_configs_dir);
1235 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1238 CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1241 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1242 OBD_FREE(temp_log, strlen(logname) + 1);
1246 /** Get a config log from the MGS and process it.
1247 * This func is called for both clients and servers.
1248 * Copy the log locally before parsing it if appropriate (non-MGS server)
1250 int mgc_process_log(struct obd_device *mgc,
1251 struct config_llog_data *cld)
1253 struct llog_ctxt *ctxt, *lctxt;
1254 struct lustre_handle lockh;
1255 struct client_obd *cli = &mgc->u.cli;
1256 struct lvfs_run_ctxt saved;
1257 struct lustre_sb_info *lsi = NULL;
1258 int rc = 0, rcl, flags = 0, must_pop = 0;
1263 /* I don't want multiple processes running process_log at once --
1264 sounds like badness. It actually might be fine, as long as
1265 we're not trying to update from the same log
1266 simultaneously (in which case we should use a per-log sem.) */
1267 cfs_down(&llog_process_lock);
1269 if (cld->cld_stopping) {
1270 cfs_up(&llog_process_lock);
1274 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1276 if (cld->cld_cfg.cfg_sb)
1277 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1279 CDEBUG(D_MGC, "Process log %s:%s from %d\n", cld->cld_logname,
1280 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1282 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1284 CERROR("missing llog context\n");
1285 cfs_up(&llog_process_lock);
1289 /* Get the cfg lock on the llog */
1290 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1291 LCK_CR, &flags, NULL, NULL, NULL,
1292 cld, 0, NULL, &lockh);
1294 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1296 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1299 * local copy of sptlrpc log is controlled elsewhere, don't try to
1302 if (rcl && cld->cld_is_sptlrpc)
1305 /* Copy the setup log locally if we can. Don't mess around if we're
1306 running an MGS though (logs are already local). */
1307 if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
1308 (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1309 !IS_MGS(lsi->lsi_ldd)) {
1310 push_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1313 /* Only try to copy log if we have the lock. */
1314 rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1316 if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1317 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1318 "log %s and no local copy."
1319 "\n", cld->cld_logname);
1320 GOTO(out_pop, rc = -ENOTCONN);
1322 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1323 "copy for now, will try to update later.\n",
1326 /* Now, whether we copied or not, start using the local llog.
1327 If we failed to copy, we'll start using whatever the old
1329 llog_ctxt_put(ctxt);
1333 if (cld->cld_is_sptlrpc)
1334 sptlrpc_conf_log_update_begin(cld->cld_logname);
1336 /* logname and instance info should be the same, so use our
1337 copy of the instance for the update. The cfg_last_idx will
1339 if (rcl == 0 || lctxt == ctxt)
1340 rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg);
1342 llog_ctxt_put(ctxt);
1344 llog_ctxt_put(lctxt);
1346 pop_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1349 * update settings on existing OBDs. doing it inside
1350 * of llog_process_lock so no device is attaching/detaching
1352 * the logname must be <fsname>-sptlrpc
1354 if (cld->cld_is_sptlrpc && rcl == 0) {
1355 sptlrpc_conf_log_update_end(cld->cld_logname);
1356 class_notify_sptlrpc_conf(cld->cld_logname,
1357 strlen(cld->cld_logname) -
1358 strlen("-sptlrpc"));
1361 /* Now drop the lock so MGS can revoke it */
1363 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1366 CERROR("Can't drop cfg lock: %d\n", rcl);
1369 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1370 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1372 cfs_up(&llog_process_lock);
1377 /** Called from lustre_process_log.
1378 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1379 * any services, and adds it to the list logs to watch (follow).
1381 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1383 struct lustre_cfg *lcfg = buf;
1388 switch(cmd = lcfg->lcfg_command) {
1389 case LCFG_LOV_ADD_OBD: {
1390 /* Overloading this cfg command: register a new target */
1391 struct mgs_target_info *mti;
1393 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1394 sizeof(struct mgs_target_info))
1395 GOTO(out, rc = -EINVAL);
1397 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1398 CDEBUG(D_MGC, "add_target %s %#x\n",
1399 mti->mti_svname, mti->mti_flags);
1400 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1403 case LCFG_LOV_DEL_OBD:
1404 /* Unregister has no meaning at the moment. */
1405 CERROR("lov_del_obd unimplemented\n");
1408 case LCFG_SPTLRPC_CONF: {
1409 rc = sptlrpc_process_config(lcfg);
1412 case LCFG_LOG_START: {
1413 struct config_llog_data *cld;
1414 struct config_llog_instance *cfg;
1415 struct super_block *sb;
1416 char *logname = lustre_cfg_string(lcfg, 1);
1417 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1418 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1420 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1423 /* We're only called through here on the initial mount */
1424 rc = config_log_add(obd, logname, cfg, sb);
1427 cld = config_log_find(logname, cfg);
1434 /* FIXME only set this for old logs! Right now this forces
1435 us to always skip the "inside markers" check */
1436 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1438 rc = mgc_process_log(obd, cld);
1439 config_log_put(cld);
1443 case LCFG_LOG_END: {
1444 struct config_llog_instance *cfg = NULL;
1445 char *logname = lustre_cfg_string(lcfg, 1);
1446 if (lcfg->lcfg_bufcount >= 2)
1447 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1449 rc = config_log_end(logname, cfg);
1453 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1454 GOTO(out, rc = -EINVAL);
1462 struct obd_ops mgc_obd_ops = {
1463 .o_owner = THIS_MODULE,
1464 .o_setup = mgc_setup,
1465 .o_precleanup = mgc_precleanup,
1466 .o_cleanup = mgc_cleanup,
1467 .o_add_conn = client_import_add_conn,
1468 .o_del_conn = client_import_del_conn,
1469 .o_connect = client_connect_import,
1470 .o_disconnect = client_disconnect_export,
1471 //.o_enqueue = mgc_enqueue,
1472 .o_cancel = mgc_cancel,
1473 //.o_iocontrol = mgc_iocontrol,
1474 .o_set_info_async = mgc_set_info_async,
1475 .o_import_event = mgc_import_event,
1476 .o_llog_init = mgc_llog_init,
1477 .o_llog_finish = mgc_llog_finish,
1478 .o_process_config = mgc_process_config,
1481 int __init mgc_init(void)
1483 return class_register_type(&mgc_obd_ops, NULL, NULL,
1484 LUSTRE_MGC_NAME, NULL);
1488 static void /*__exit*/ mgc_exit(void)
1490 class_unregister_type(LUSTRE_MGC_NAME);
1493 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1494 MODULE_DESCRIPTION("Lustre Management Client");
1495 MODULE_LICENSE("GPL");
1497 module_init(mgc_init);
1498 module_exit(mgc_exit);