1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
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 cfs_mutex_init(&cld->cld_lock);
225 cld->cld_cfg.cfg_last_idx = 0;
226 cld->cld_cfg.cfg_flags = 0;
227 cld->cld_cfg.cfg_sb = sb;
228 cld->cld_is_sptlrpc = is_sptlrpc;
229 cfs_atomic_set(&cld->cld_refcount, 1);
231 /* Keep the mgc around until we are done */
232 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
234 if (cfg && cfg->cfg_instance != NULL) {
235 OBD_ALLOC(cld->cld_cfg.cfg_instance,
236 strlen(cfg->cfg_instance) + 1);
237 strcpy(cld->cld_cfg.cfg_instance, cfg->cfg_instance);
241 sptlrpc_conf_log_start(logname);
242 cld->cld_cfg.cfg_obdname = obd->obd_name;
245 rc = mgc_logname2resid(logname, &cld->cld_resid);
247 cfs_spin_lock(&config_list_lock);
248 cfs_list_add(&cld->cld_list_chain, &config_llog_list);
249 cfs_spin_unlock(&config_list_lock);
257 rc = mgc_process_log(obd, cld);
259 CERROR("failed processing sptlrpc log: %d\n", rc);
265 /** Add this log to the list of active logs watched by an MGC.
266 * Active means we're watching for updates.
267 * We have one active log per "mount" - client instance or servername.
268 * Each instance may be at a different point in the log.
270 static int config_log_add(struct obd_device *obd, char *logname,
271 struct config_llog_instance *cfg,
272 struct super_block *sb)
274 struct config_llog_data *cld, *sptlrpc_cld;
279 CDEBUG(D_MGC, "adding config log %s:%s\n", logname, cfg->cfg_instance);
282 * for each regular log, the depended sptlrpc log name is
283 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
285 ptr = strrchr(logname, '-');
286 if (ptr == NULL || ptr - logname > 8) {
287 CERROR("logname %s is too long\n", logname);
291 memcpy(seclogname, logname, ptr - logname);
292 strcpy(seclogname + (ptr - logname), "-sptlrpc");
294 sptlrpc_cld = config_log_find(seclogname, NULL);
295 if (IS_ERR(sptlrpc_cld)) {
296 sptlrpc_cld = do_config_log_add(obd, seclogname, 1, NULL, NULL);
297 if (IS_ERR(sptlrpc_cld)) {
298 CERROR("can't create sptlrpc log: %s\n", seclogname);
299 RETURN(PTR_ERR(sptlrpc_cld));
303 cld = do_config_log_add(obd, logname, 0, cfg, sb);
305 CERROR("can't create log: %s\n", logname);
306 config_log_put(sptlrpc_cld);
307 RETURN(PTR_ERR(cld));
310 cld->cld_sptlrpc = sptlrpc_cld;
315 CFS_DECLARE_MUTEX(llog_process_lock);
317 /** Stop watching for updates on this log.
319 static int config_log_end(char *logname, struct config_llog_instance *cfg)
321 struct config_llog_data *cld, *cld_sptlrpc = NULL;
325 cld = config_log_find(logname, cfg);
327 RETURN(PTR_ERR(cld));
329 cfs_mutex_lock(&cld->cld_lock);
331 * if cld_stopping is set, it means we didn't start the log thus
332 * not owning the start ref. this can happen after previous umount:
333 * the cld still hanging there waiting for lock cancel, and we
334 * remount again but failed in the middle and call log_end without
337 if (unlikely(cld->cld_stopping)) {
338 cfs_mutex_unlock(&cld->cld_lock);
339 /* drop the ref from the find */
344 cld->cld_stopping = 1;
345 cfs_mutex_unlock(&cld->cld_lock);
347 cfs_spin_lock(&config_list_lock);
348 cld_sptlrpc = cld->cld_sptlrpc;
349 cld->cld_sptlrpc = NULL;
350 cfs_spin_unlock(&config_list_lock);
353 config_log_put(cld_sptlrpc);
355 /* drop the ref from the find */
357 /* drop the start ref */
360 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
365 /* reenqueue any lost locks */
366 #define RQ_RUNNING 0x1
370 static int rq_state = 0;
371 static cfs_waitq_t rq_waitq;
373 static int mgc_requeue_add(struct config_llog_data *cld, int later);
375 static void do_requeue(struct config_llog_data *cld)
377 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
379 /* Do not run mgc_process_log on a disconnected export or an
380 export which is being disconnected. Take the client
381 semaphore to make the check non-racy. */
382 cfs_down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
383 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
384 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
385 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
387 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
390 cfs_up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
392 /* Whether we enqueued again or not in mgc_process_log, we're done
393 * with the ref from the old enqueue */
397 static int mgc_requeue_thread(void *data)
399 struct l_wait_info lwi_now, lwi_later;
400 struct config_llog_data *cld, *cld_next, *cld_prev;
401 char name[] = "ll_cfg_requeue";
407 CDEBUG(D_MGC, "Starting requeue thread\n");
409 lwi_later = LWI_TIMEOUT(60 * CFS_HZ, NULL, NULL);
410 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP), &lwi_later);
412 /* Keep trying failed locks periodically */
413 cfs_spin_lock(&config_list_lock);
414 while (rq_state & (RQ_NOW | RQ_LATER)) {
415 /* Any new or requeued lostlocks will change the state */
416 rq_state &= ~(RQ_NOW | RQ_LATER);
417 cfs_spin_unlock(&config_list_lock);
419 /* Always wait a few seconds to allow the server who
420 caused the lock revocation to finish its setup, plus some
421 random so everyone doesn't try to reconnect at once. */
422 lwi_now = LWI_TIMEOUT(3 * CFS_HZ + (cfs_rand() & 0xff) * \
425 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi_now);
428 * iterate & processing through the list. for each cld, process
429 * its depending sptlrpc cld firstly (if any) and then itself.
431 * it's guaranteed any item in the list must have
432 * reference > 0; and if cld_lostlock is set, at
433 * least one reference is taken by the previous enqueue.
435 * Note: releasing a cld might lead to itself and its depended
436 * sptlrpc cld be unlinked from the list. to safely iterate
437 * we need to take a reference on next cld before processing.
441 cfs_spin_lock(&config_list_lock);
442 cfs_list_for_each_entry_safe(cld, cld_next, &config_llog_list,
444 if (cld->cld_list_chain.next != &config_llog_list)
445 cfs_atomic_inc(&cld_next->cld_refcount);
447 if (cld->cld_lostlock) {
448 if (cld->cld_sptlrpc &&
449 cld->cld_sptlrpc->cld_lostlock) {
450 cld->cld_sptlrpc->cld_lostlock = 0;
452 cfs_spin_unlock(&config_list_lock);
453 do_requeue(cld->cld_sptlrpc);
454 cfs_spin_lock(&config_list_lock);
455 LASSERT(cld->cld_lostlock);
458 cld->cld_lostlock = 0;
460 cfs_spin_unlock(&config_list_lock);
462 cfs_spin_lock(&config_list_lock);
467 cfs_spin_unlock(&config_list_lock);
468 config_log_put(cld_prev);
469 cfs_spin_lock(&config_list_lock);
474 cfs_spin_unlock(&config_list_lock);
476 /* Wait a bit to see if anyone else needs a requeue */
477 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
479 cfs_spin_lock(&config_list_lock);
481 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
482 rq_state &= ~RQ_RUNNING;
483 cfs_spin_unlock(&config_list_lock);
485 CDEBUG(D_MGC, "Ending requeue thread\n");
489 /* Add a cld to the list to requeue. Start the requeue thread if needed.
490 We are responsible for dropping the config log reference from here on out. */
491 static int mgc_requeue_add(struct config_llog_data *cld, int later)
495 CDEBUG(D_INFO, "log %s: requeue (l=%d r=%d sp=%d st=%x)\n",
496 cld->cld_logname, later, cfs_atomic_read(&cld->cld_refcount),
497 cld->cld_stopping, rq_state);
498 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
500 /* Hold lock for rq_state */
501 cfs_spin_lock(&config_list_lock);
503 if (cld->cld_stopping || (rq_state & RQ_STOP)) {
504 cld->cld_lostlock = 0;
505 cfs_spin_unlock(&config_list_lock);
510 cld->cld_lostlock = 1;
512 if (!(rq_state & RQ_RUNNING)) {
513 LASSERT(rq_state == 0);
514 rq_state = RQ_RUNNING | (later ? RQ_LATER : RQ_NOW);
515 cfs_spin_unlock(&config_list_lock);
516 rc = cfs_kernel_thread(mgc_requeue_thread, 0,
517 CLONE_VM | CLONE_FILES);
519 CERROR("log %s: cannot start requeue thread (%d),"
520 "no more log updates!\n", cld->cld_logname, rc);
521 /* Drop the ref, since the rq thread won't */
522 cld->cld_lostlock = 0;
528 rq_state |= later ? RQ_LATER : RQ_NOW;
529 cfs_spin_unlock(&config_list_lock);
530 cfs_waitq_signal(&rq_waitq);
536 /********************** class fns **********************/
538 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
539 struct vfsmount *mnt)
541 struct lvfs_run_ctxt saved;
542 struct lustre_sb_info *lsi = s2lsi(sb);
543 struct client_obd *cli = &obd->u.cli;
544 struct dentry *dentry;
550 LASSERT(lsi->lsi_srv_mnt == mnt);
552 /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
553 cfs_down(&cli->cl_mgc_sem);
555 cfs_cleanup_group_info();
557 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
558 if (IS_ERR(obd->obd_fsops)) {
559 cfs_up(&cli->cl_mgc_sem);
560 CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd),
561 PTR_ERR(obd->obd_fsops));
562 RETURN(PTR_ERR(obd->obd_fsops));
565 cli->cl_mgc_vfsmnt = mnt;
566 fsfilt_setup(obd, mnt->mnt_sb);
568 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
569 obd->obd_lvfs_ctxt.pwdmnt = mnt;
570 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
571 obd->obd_lvfs_ctxt.fs = get_ds();
573 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
574 dentry = lookup_one_len(MOUNT_CONFIGS_DIR, cfs_fs_pwd(current->fs),
575 strlen(MOUNT_CONFIGS_DIR));
576 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
577 if (IS_ERR(dentry)) {
578 err = PTR_ERR(dentry);
579 CERROR("cannot lookup %s directory: rc = %d\n",
580 MOUNT_CONFIGS_DIR, err);
583 cli->cl_mgc_configs_dir = dentry;
585 /* We take an obd ref to insure that we can't get to mgc_cleanup
586 without calling mgc_fs_cleanup first. */
587 class_incref(obd, "mgc_fs", obd);
589 label = fsfilt_get_label(obd, mnt->mnt_sb);
591 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
593 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
597 fsfilt_put_ops(obd->obd_fsops);
598 obd->obd_fsops = NULL;
599 cli->cl_mgc_vfsmnt = NULL;
600 cfs_up(&cli->cl_mgc_sem);
604 static int mgc_fs_cleanup(struct obd_device *obd)
606 struct client_obd *cli = &obd->u.cli;
610 LASSERT(cli->cl_mgc_vfsmnt != NULL);
612 if (cli->cl_mgc_configs_dir != NULL) {
613 struct lvfs_run_ctxt saved;
614 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
615 l_dput(cli->cl_mgc_configs_dir);
616 cli->cl_mgc_configs_dir = NULL;
617 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
618 class_decref(obd, "mgc_fs", obd);
621 cli->cl_mgc_vfsmnt = NULL;
623 fsfilt_put_ops(obd->obd_fsops);
625 cfs_up(&cli->cl_mgc_sem);
630 static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0);
631 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
637 case OBD_CLEANUP_EARLY:
639 case OBD_CLEANUP_EXPORTS:
640 if (cfs_atomic_dec_and_test(&mgc_count)) {
641 /* Kick the requeue waitq - cld's should all be
643 cfs_spin_lock(&config_list_lock);
645 cfs_spin_unlock(&config_list_lock);
646 cfs_waitq_signal(&rq_waitq);
648 rc = obd_llog_finish(obd, 0);
650 CERROR("failed to cleanup llogging subsystems\n");
656 static int mgc_cleanup(struct obd_device *obd)
658 struct client_obd *cli = &obd->u.cli;
662 LASSERT(cli->cl_mgc_vfsmnt == NULL);
664 /* COMPAT_146 - old config logs may have added profiles we don't
666 if (obd->obd_type->typ_refcnt <= 1)
667 /* Only for the last mgc */
668 class_del_profiles();
670 lprocfs_obd_cleanup(obd);
673 rc = client_obd_cleanup(obd);
677 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
679 struct lprocfs_static_vars lvars;
685 rc = client_obd_setup(obd, lcfg);
687 GOTO(err_decref, rc);
689 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
691 CERROR("failed to setup llogging subsystems\n");
692 GOTO(err_cleanup, rc);
695 lprocfs_mgc_init_vars(&lvars);
696 lprocfs_obd_setup(obd, lvars.obd_vars);
697 sptlrpc_lprocfs_cliobd_attach(obd);
699 cfs_spin_lock(&config_list_lock);
700 cfs_atomic_inc(&mgc_count);
701 if (cfs_atomic_read(&mgc_count) == 1) {
702 rq_state &= ~RQ_STOP;
703 cfs_waitq_init(&rq_waitq);
705 cfs_spin_unlock(&config_list_lock);
710 client_obd_cleanup(obd);
716 /* based on ll_mdc_blocking_ast */
717 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
718 void *data, int flag)
720 struct lustre_handle lockh;
721 struct config_llog_data *cld = (struct config_llog_data *)data;
726 case LDLM_CB_BLOCKING:
727 /* mgs wants the lock, give it up... */
728 LDLM_DEBUG(lock, "MGC blocking CB");
729 ldlm_lock2handle(lock, &lockh);
730 rc = ldlm_cli_cancel(&lockh);
732 case LDLM_CB_CANCELING: {
733 /* We've given up the lock, prepare ourselves to update. */
734 LDLM_DEBUG(lock, "MGC cancel CB");
736 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
737 lock->l_resource->lr_name.name[0],
738 (char *)&lock->l_resource->lr_name.name[0]);
741 CERROR("missing data, won't requeue\n");
744 /* Are we done with this log? */
745 if (cld->cld_stopping) {
746 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
751 /* Make sure not to re-enqueue when the mgc is stopping
752 (we get called from client_disconnect_export) */
753 if (!lock->l_conn_export ||
754 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
755 CDEBUG(D_MGC, "log %s: disconnecting, won't requeue\n",
760 /* Did we fail to get the lock? */
761 if (lock->l_req_mode != lock->l_granted_mode &&
762 !cld->cld_is_sptlrpc) {
763 CDEBUG(D_MGC, "log %s: original grant failed, will "
764 "requeue later\n", cld->cld_logname);
765 /* Try to re-enqueue later */
766 rc = mgc_requeue_add(cld, 1);
770 rc = mgc_requeue_add(cld, 0);
779 CERROR("%s CB failed %d:\n", flag == LDLM_CB_BLOCKING ?
780 "blocking" : "cancel", rc);
781 LDLM_ERROR(lock, "MGC ast");
786 /* Not sure where this should go... */
787 #define MGC_ENQUEUE_LIMIT 50
788 #define MGC_TARGET_REG_LIMIT 10
789 #define MGC_SEND_PARAM_LIMIT 10
791 /* Send parameter to MGS*/
792 static int mgc_set_mgs_param(struct obd_export *exp,
793 struct mgs_send_param *msp)
795 struct ptlrpc_request *req;
796 struct mgs_send_param *req_msp, *rep_msp;
800 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
801 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
806 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
808 ptlrpc_req_finished(req);
812 memcpy(req_msp, msp, sizeof(*req_msp));
813 ptlrpc_request_set_replen(req);
815 /* Limit how long we will wait for the enqueue to complete */
816 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
817 rc = ptlrpc_queue_wait(req);
819 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
820 memcpy(msp, rep_msp, sizeof(*rep_msp));
823 ptlrpc_req_finished(req);
828 /* Take a config lock so we can get cancel notifications */
829 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
830 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
831 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
832 void *data, __u32 lvb_len, void *lvb_swabber,
833 struct lustre_handle *lockh)
835 struct config_llog_data *cld = (struct config_llog_data *)data;
836 struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
837 ldlm_completion_ast, NULL, NULL, data};
838 struct ptlrpc_request *req;
839 int short_limit = cld->cld_is_sptlrpc;
843 CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
844 cld->cld_resid.name[0]);
846 /* We can only drop this config log ref when we drop the lock */
847 if (config_log_get(cld))
848 RETURN(ELDLM_LOCK_ABORTED);
850 /* We need a callback for every lockholder, so don't try to
851 ldlm_lock_match (see rev 1.1.2.11.2.47) */
852 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
853 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
857 ptlrpc_request_set_replen(req);
858 /* check if this is server or client */
859 if (cld->cld_cfg.cfg_sb) {
860 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
861 if (lsi && (lsi->lsi_flags & LSI_SERVER))
864 /* Limit how long we will wait for the enqueue to complete */
865 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
866 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
868 /* A failed enqueue should still call the mgc_blocking_ast,
869 where it will be requeued if needed ("grant failed"). */
870 ptlrpc_req_finished(req);
874 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
875 __u32 mode, struct lustre_handle *lockh)
879 ldlm_lock_decref(lockh, mode);
884 /* Send target_reg message to MGS */
885 static int mgc_target_register(struct obd_export *exp,
886 struct mgs_target_info *mti)
888 struct ptlrpc_request *req;
889 struct mgs_target_info *req_mti, *rep_mti;
893 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
894 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
899 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
901 ptlrpc_req_finished(req);
905 memcpy(req_mti, mti, sizeof(*req_mti));
906 ptlrpc_request_set_replen(req);
907 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
908 /* Limit how long we will wait for the enqueue to complete */
909 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
911 rc = ptlrpc_queue_wait(req);
913 rep_mti = req_capsule_server_get(&req->rq_pill,
914 &RMF_MGS_TARGET_INFO);
915 memcpy(mti, rep_mti, sizeof(*rep_mti));
916 CDEBUG(D_MGC, "register %s got index = %d\n",
917 mti->mti_svname, mti->mti_stripe_index);
919 ptlrpc_req_finished(req);
924 int mgc_set_info_async(struct obd_export *exp, obd_count keylen,
925 void *key, obd_count vallen, void *val,
926 struct ptlrpc_request_set *set)
931 /* Turn off initial_recov after we try all backup servers once */
932 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
933 struct obd_import *imp = class_exp2cliimp(exp);
935 if (vallen != sizeof(int))
938 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
939 imp->imp_obd->obd_name, value,
940 imp->imp_deactive, imp->imp_invalid,
941 imp->imp_replayable, imp->imp_obd->obd_replayable,
942 ptlrpc_import_state_name(imp->imp_state));
943 /* Resurrect if we previously died */
944 if ((imp->imp_state != LUSTRE_IMP_FULL &&
945 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
946 ptlrpc_reconnect_import(imp);
949 /* FIXME move this to mgc_process_config */
950 if (KEY_IS(KEY_REGISTER_TARGET)) {
951 struct mgs_target_info *mti;
952 if (vallen != sizeof(struct mgs_target_info))
954 mti = (struct mgs_target_info *)val;
955 CDEBUG(D_MGC, "register_target %s %#x\n",
956 mti->mti_svname, mti->mti_flags);
957 rc = mgc_target_register(exp, mti);
960 if (KEY_IS(KEY_SET_FS)) {
961 struct super_block *sb = (struct super_block *)val;
962 struct lustre_sb_info *lsi;
963 if (vallen != sizeof(struct super_block))
966 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
968 CERROR("set_fs got %d\n", rc);
972 if (KEY_IS(KEY_CLEAR_FS)) {
975 rc = mgc_fs_cleanup(exp->exp_obd);
977 CERROR("clear_fs got %d\n", rc);
981 if (KEY_IS(KEY_SET_INFO)) {
982 struct mgs_send_param *msp;
984 msp = (struct mgs_send_param *)val;
985 rc = mgc_set_mgs_param(exp, msp);
988 if (KEY_IS(KEY_MGSSEC)) {
989 struct client_obd *cli = &exp->exp_obd->u.cli;
990 struct sptlrpc_flavor flvr;
993 * empty string means using current flavor, if which haven't
994 * been set yet, set it as null.
996 * if flavor has been set previously, check the asking flavor
997 * must match the existing one.
1000 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1006 rc = sptlrpc_parse_flavor(val, &flvr);
1008 CERROR("invalid sptlrpc flavor %s to MGS\n",
1014 * caller already hold a mutex
1016 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1017 cli->cl_flvr_mgc = flvr;
1018 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1019 sizeof(flvr)) != 0) {
1022 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1024 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1025 "currently %s is in use\n",
1035 static int mgc_import_event(struct obd_device *obd,
1036 struct obd_import *imp,
1037 enum obd_import_event event)
1041 LASSERT(imp->imp_obd == obd);
1042 CDEBUG(D_MGC, "import event %#x\n", event);
1045 case IMP_EVENT_DISCON:
1046 /* MGC imports should not wait for recovery */
1048 case IMP_EVENT_INACTIVE:
1050 case IMP_EVENT_INVALIDATE: {
1051 struct ldlm_namespace *ns = obd->obd_namespace;
1052 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1055 case IMP_EVENT_ACTIVE:
1056 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
1057 /* Clearing obd_no_recov allows us to continue pinging */
1058 obd->obd_no_recov = 0;
1063 CERROR("Unknown import event %#x\n", event);
1069 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1070 struct obd_device *tgt, int *index)
1072 struct llog_ctxt *ctxt;
1076 LASSERT(olg == &obd->obd_olg);
1078 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
1083 rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
1086 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1088 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1093 llog_initiator_connect(ctxt);
1094 llog_ctxt_put(ctxt);
1096 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1104 static int mgc_llog_finish(struct obd_device *obd, int count)
1106 struct llog_ctxt *ctxt;
1107 int rc = 0, rc2 = 0;
1110 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1112 rc = llog_cleanup(ctxt);
1114 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1116 rc2 = llog_cleanup(ctxt);
1124 /* identical to mgs_log_is_empty */
1125 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1128 struct lvfs_run_ctxt saved;
1129 struct llog_handle *llh;
1132 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1133 rc = llog_create(ctxt, &llh, NULL, name);
1135 llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL);
1136 rc = llog_get_size(llh);
1139 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1140 /* header is record 1 */
1144 static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec,
1147 struct llog_rec_hdr local_rec = *rec;
1148 struct llog_handle *local_llh = (struct llog_handle *)data;
1149 char *cfg_buf = (char*) (rec + 1);
1150 struct lustre_cfg *lcfg;
1154 /* Append all records */
1155 local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
1156 rc = llog_write_rec(local_llh, &local_rec, NULL, 0,
1157 (void *)cfg_buf, -1);
1159 lcfg = (struct lustre_cfg *)cfg_buf;
1160 CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
1161 rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
1162 lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
1167 /* Copy a remote log locally */
1168 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1169 struct llog_ctxt *lctxt, char *logname)
1171 struct llog_handle *local_llh, *remote_llh;
1172 struct obd_uuid *uuid;
1177 /* Write new log to a temp name, then vfs_rename over logname
1178 upon successful completion. */
1180 OBD_ALLOC(temp_log, strlen(logname) + 1);
1183 sprintf(temp_log, "%sT", logname);
1185 /* Make sure there's no old temp log */
1186 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1189 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL);
1192 rc = llog_destroy(local_llh);
1193 llog_free_handle(local_llh);
1197 /* open local log */
1198 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1202 /* set the log header uuid for fun */
1203 OBD_ALLOC_PTR(uuid);
1204 obd_str2uuid(uuid, logname);
1205 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid);
1208 GOTO(out_closel, rc);
1210 /* open remote log */
1211 rc = llog_create(rctxt, &remote_llh, NULL, logname);
1213 GOTO(out_closel, rc);
1214 rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL);
1216 GOTO(out_closer, rc);
1218 /* Copy remote log */
1219 rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL);
1222 rc2 = llog_close(remote_llh);
1226 rc2 = llog_close(local_llh);
1230 /* We've copied the remote log to the local temp log, now
1231 replace the old local log with the temp log. */
1233 struct client_obd *cli = &obd->u.cli;
1235 LASSERT(cli->cl_mgc_configs_dir);
1236 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1239 CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1242 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1243 OBD_FREE(temp_log, strlen(logname) + 1);
1247 /** Get a config log from the MGS and process it.
1248 * This func is called for both clients and servers.
1249 * Copy the log locally before parsing it if appropriate (non-MGS server)
1251 int mgc_process_log(struct obd_device *mgc,
1252 struct config_llog_data *cld)
1254 struct llog_ctxt *ctxt, *lctxt;
1255 struct lustre_handle lockh;
1256 struct client_obd *cli = &mgc->u.cli;
1257 struct lvfs_run_ctxt saved;
1258 struct lustre_sb_info *lsi = NULL;
1259 int rc = 0, rcl, flags = 0, must_pop = 0;
1264 /* I don't want multiple processes running process_log at once --
1265 sounds like badness. It actually might be fine, as long as
1266 we're not trying to update from the same log
1267 simultaneously (in which case we should use a per-log sem.) */
1268 cfs_mutex_lock(&cld->cld_lock);
1270 if (cld->cld_stopping) {
1271 cfs_mutex_unlock(&cld->cld_lock);
1275 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1277 if (cld->cld_cfg.cfg_sb)
1278 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1280 CDEBUG(D_MGC, "Process log %s:%s from %d\n", cld->cld_logname,
1281 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1283 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1285 CERROR("missing llog context\n");
1286 cfs_mutex_unlock(&cld->cld_lock);
1290 /* Get the cfg lock on the llog */
1291 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1292 LCK_CR, &flags, NULL, NULL, NULL,
1293 cld, 0, NULL, &lockh);
1295 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1297 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1300 * local copy of sptlrpc log is controlled elsewhere, don't try to
1303 if (rcl && cld->cld_is_sptlrpc)
1306 /* Copy the setup log locally if we can. Don't mess around if we're
1307 running an MGS though (logs are already local). */
1308 if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
1309 (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1310 !IS_MGS(lsi->lsi_ldd)) {
1311 push_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1314 /* Only try to copy log if we have the lock. */
1315 rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1317 if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1318 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1319 "log %s and no local copy."
1320 "\n", cld->cld_logname);
1321 GOTO(out_pop, rc = -ENOTCONN);
1323 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1324 "copy for now, will try to update later.\n",
1327 /* Now, whether we copied or not, start using the local llog.
1328 If we failed to copy, we'll start using whatever the old
1330 llog_ctxt_put(ctxt);
1334 if (cld->cld_is_sptlrpc)
1335 sptlrpc_conf_log_update_begin(cld->cld_logname);
1337 /* logname and instance info should be the same, so use our
1338 copy of the instance for the update. The cfg_last_idx will
1340 if (rcl == 0 || lctxt == ctxt)
1341 rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg);
1343 llog_ctxt_put(ctxt);
1345 llog_ctxt_put(lctxt);
1347 pop_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1350 * update settings on existing OBDs. doing it inside
1351 * of llog_process_lock so no device is attaching/detaching
1353 * the logname must be <fsname>-sptlrpc
1355 if (cld->cld_is_sptlrpc && rcl == 0) {
1356 sptlrpc_conf_log_update_end(cld->cld_logname);
1357 class_notify_sptlrpc_conf(cld->cld_logname,
1358 strlen(cld->cld_logname) -
1359 strlen("-sptlrpc"));
1362 /* Now drop the lock so MGS can revoke it */
1364 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1367 CERROR("Can't drop cfg lock: %d\n", rcl);
1370 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1371 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1373 cfs_mutex_unlock(&cld->cld_lock);
1378 /** Called from lustre_process_log.
1379 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1380 * any services, and adds it to the list logs to watch (follow).
1382 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1384 struct lustre_cfg *lcfg = buf;
1389 switch(cmd = lcfg->lcfg_command) {
1390 case LCFG_LOV_ADD_OBD: {
1391 /* Overloading this cfg command: register a new target */
1392 struct mgs_target_info *mti;
1394 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1395 sizeof(struct mgs_target_info))
1396 GOTO(out, rc = -EINVAL);
1398 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1399 CDEBUG(D_MGC, "add_target %s %#x\n",
1400 mti->mti_svname, mti->mti_flags);
1401 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1404 case LCFG_LOV_DEL_OBD:
1405 /* Unregister has no meaning at the moment. */
1406 CERROR("lov_del_obd unimplemented\n");
1409 case LCFG_SPTLRPC_CONF: {
1410 rc = sptlrpc_process_config(lcfg);
1413 case LCFG_LOG_START: {
1414 struct config_llog_data *cld;
1415 struct config_llog_instance *cfg;
1416 struct super_block *sb;
1417 char *logname = lustre_cfg_string(lcfg, 1);
1418 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1419 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1421 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1424 /* We're only called through here on the initial mount */
1425 rc = config_log_add(obd, logname, cfg, sb);
1428 cld = config_log_find(logname, cfg);
1435 /* FIXME only set this for old logs! Right now this forces
1436 us to always skip the "inside markers" check */
1437 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1439 rc = mgc_process_log(obd, cld);
1440 config_log_put(cld);
1444 case LCFG_LOG_END: {
1445 struct config_llog_instance *cfg = NULL;
1446 char *logname = lustre_cfg_string(lcfg, 1);
1447 if (lcfg->lcfg_bufcount >= 2)
1448 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1450 rc = config_log_end(logname, cfg);
1454 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1455 GOTO(out, rc = -EINVAL);
1463 struct obd_ops mgc_obd_ops = {
1464 .o_owner = THIS_MODULE,
1465 .o_setup = mgc_setup,
1466 .o_precleanup = mgc_precleanup,
1467 .o_cleanup = mgc_cleanup,
1468 .o_add_conn = client_import_add_conn,
1469 .o_del_conn = client_import_del_conn,
1470 .o_connect = client_connect_import,
1471 .o_disconnect = client_disconnect_export,
1472 //.o_enqueue = mgc_enqueue,
1473 .o_cancel = mgc_cancel,
1474 //.o_iocontrol = mgc_iocontrol,
1475 .o_set_info_async = mgc_set_info_async,
1476 .o_import_event = mgc_import_event,
1477 .o_llog_init = mgc_llog_init,
1478 .o_llog_finish = mgc_llog_finish,
1479 .o_process_config = mgc_process_config,
1482 int __init mgc_init(void)
1484 return class_register_type(&mgc_obd_ops, NULL, NULL,
1485 LUSTRE_MGC_NAME, NULL);
1489 static void /*__exit*/ mgc_exit(void)
1491 class_unregister_type(LUSTRE_MGC_NAME);
1494 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1495 MODULE_DESCRIPTION("Lustre Management Client");
1496 MODULE_LICENSE("GPL");
1498 module_init(mgc_init);
1499 module_exit(mgc_exit);