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 OBD_FREE(cld, sizeof(*cld));
151 cfs_spin_unlock(&config_list_lock);
157 /* Find a config log by name */
159 struct config_llog_data *config_log_find(char *logname,
160 struct config_llog_instance *cfg)
162 struct config_llog_data *cld;
163 struct config_llog_data *found = NULL;
167 LASSERT(logname != NULL);
169 instance = cfg ? cfg->cfg_instance : NULL;
170 cfs_spin_lock(&config_list_lock);
171 cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
172 /* check if instance equals */
173 if (instance != cld->cld_cfg.cfg_instance)
176 /* instance may be NULL, should check name */
177 if (strcmp(logname, cld->cld_logname) == 0) {
183 cfs_atomic_inc(&found->cld_refcount);
184 LASSERT(found->cld_stopping == 0 || found->cld_is_sptlrpc == 0);
186 cfs_spin_unlock(&config_list_lock);
191 struct config_llog_data *do_config_log_add(struct obd_device *obd,
193 unsigned int is_sptlrpc,
194 struct config_llog_instance *cfg,
195 struct super_block *sb)
197 struct config_llog_data *cld;
201 CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
202 cfg ? cfg->cfg_instance : 0);
204 OBD_ALLOC(cld, sizeof(*cld));
206 RETURN(ERR_PTR(-ENOMEM));
207 OBD_ALLOC(cld->cld_logname, strlen(logname) + 1);
208 if (!cld->cld_logname) {
209 OBD_FREE(cld, sizeof(*cld));
210 RETURN(ERR_PTR(-ENOMEM));
212 strcpy(cld->cld_logname, logname);
215 cfs_mutex_init(&cld->cld_lock);
216 cld->cld_cfg.cfg_last_idx = 0;
217 cld->cld_cfg.cfg_flags = 0;
218 cld->cld_cfg.cfg_sb = sb;
219 cld->cld_is_sptlrpc = is_sptlrpc;
220 cfs_atomic_set(&cld->cld_refcount, 1);
222 /* Keep the mgc around until we are done */
223 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
226 sptlrpc_conf_log_start(logname);
227 cld->cld_cfg.cfg_obdname = obd->obd_name;
230 rc = mgc_logname2resid(logname, &cld->cld_resid);
232 cfs_spin_lock(&config_list_lock);
233 cfs_list_add(&cld->cld_list_chain, &config_llog_list);
234 cfs_spin_unlock(&config_list_lock);
242 rc = mgc_process_log(obd, cld);
244 CERROR("failed processing sptlrpc log: %d\n", rc);
250 /** Add this log to the list of active logs watched by an MGC.
251 * Active means we're watching for updates.
252 * We have one active log per "mount" - client instance or servername.
253 * Each instance may be at a different point in the log.
255 static int config_log_add(struct obd_device *obd, char *logname,
256 struct config_llog_instance *cfg,
257 struct super_block *sb)
259 struct config_llog_data *cld, *sptlrpc_cld;
264 CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
267 * for each regular log, the depended sptlrpc log name is
268 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
270 ptr = strrchr(logname, '-');
271 if (ptr == NULL || ptr - logname > 8) {
272 CERROR("logname %s is too long\n", logname);
276 memcpy(seclogname, logname, ptr - logname);
277 strcpy(seclogname + (ptr - logname), "-sptlrpc");
279 sptlrpc_cld = config_log_find(seclogname, NULL);
280 if (sptlrpc_cld == NULL) {
281 sptlrpc_cld = do_config_log_add(obd, seclogname, 1, NULL, NULL);
282 if (IS_ERR(sptlrpc_cld)) {
283 CERROR("can't create sptlrpc log: %s\n", seclogname);
284 RETURN(PTR_ERR(sptlrpc_cld));
288 cld = do_config_log_add(obd, logname, 0, cfg, sb);
290 CERROR("can't create log: %s\n", logname);
291 config_log_put(sptlrpc_cld);
292 RETURN(PTR_ERR(cld));
295 cld->cld_sptlrpc = sptlrpc_cld;
300 CFS_DECLARE_MUTEX(llog_process_lock);
302 /** Stop watching for updates on this log.
304 static int config_log_end(char *logname, struct config_llog_instance *cfg)
306 struct config_llog_data *cld, *cld_sptlrpc = NULL;
310 cld = config_log_find(logname, cfg);
314 cfs_mutex_lock(&cld->cld_lock);
316 * if cld_stopping is set, it means we didn't start the log thus
317 * not owning the start ref. this can happen after previous umount:
318 * the cld still hanging there waiting for lock cancel, and we
319 * remount again but failed in the middle and call log_end without
322 if (unlikely(cld->cld_stopping)) {
323 cfs_mutex_unlock(&cld->cld_lock);
324 /* drop the ref from the find */
329 cld->cld_stopping = 1;
330 cfs_mutex_unlock(&cld->cld_lock);
332 cfs_spin_lock(&config_list_lock);
333 cld_sptlrpc = cld->cld_sptlrpc;
334 cld->cld_sptlrpc = NULL;
335 cfs_spin_unlock(&config_list_lock);
338 config_log_put(cld_sptlrpc);
340 /* drop the ref from the find */
342 /* drop the start ref */
345 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
350 /* reenqueue any lost locks */
351 #define RQ_RUNNING 0x1
355 static int rq_state = 0;
356 static cfs_waitq_t rq_waitq;
358 static int mgc_requeue_add(struct config_llog_data *cld, int later);
360 static void do_requeue(struct config_llog_data *cld)
362 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
364 /* Do not run mgc_process_log on a disconnected export or an
365 export which is being disconnected. Take the client
366 semaphore to make the check non-racy. */
367 cfs_down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
368 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
369 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
370 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
372 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
375 cfs_up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
377 /* Whether we enqueued again or not in mgc_process_log, we're done
378 * with the ref from the old enqueue */
382 static int mgc_requeue_thread(void *data)
384 struct l_wait_info lwi_now, lwi_later;
385 struct config_llog_data *cld, *cld_next, *cld_prev;
386 char name[] = "ll_cfg_requeue";
392 CDEBUG(D_MGC, "Starting requeue thread\n");
394 lwi_later = LWI_TIMEOUT(60 * CFS_HZ, NULL, NULL);
395 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP), &lwi_later);
397 /* Keep trying failed locks periodically */
398 cfs_spin_lock(&config_list_lock);
399 while (rq_state & (RQ_NOW | RQ_LATER)) {
400 /* Any new or requeued lostlocks will change the state */
401 rq_state &= ~(RQ_NOW | RQ_LATER);
402 cfs_spin_unlock(&config_list_lock);
404 /* Always wait a few seconds to allow the server who
405 caused the lock revocation to finish its setup, plus some
406 random so everyone doesn't try to reconnect at once. */
407 lwi_now = LWI_TIMEOUT(3 * CFS_HZ + (cfs_rand() & 0xff) * \
410 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi_now);
413 * iterate & processing through the list. for each cld, process
414 * its depending sptlrpc cld firstly (if any) and then itself.
416 * it's guaranteed any item in the list must have
417 * reference > 0; and if cld_lostlock is set, at
418 * least one reference is taken by the previous enqueue.
420 * Note: releasing a cld might lead to itself and its depended
421 * sptlrpc cld be unlinked from the list. to safely iterate
422 * we need to take a reference on next cld before processing.
426 cfs_spin_lock(&config_list_lock);
427 cfs_list_for_each_entry_safe(cld, cld_next, &config_llog_list,
429 if (cld->cld_list_chain.next != &config_llog_list)
430 cfs_atomic_inc(&cld_next->cld_refcount);
432 if (cld->cld_lostlock) {
433 if (cld->cld_sptlrpc &&
434 cld->cld_sptlrpc->cld_lostlock) {
435 cld->cld_sptlrpc->cld_lostlock = 0;
437 cfs_spin_unlock(&config_list_lock);
438 do_requeue(cld->cld_sptlrpc);
439 cfs_spin_lock(&config_list_lock);
440 LASSERT(cld->cld_lostlock);
443 cld->cld_lostlock = 0;
445 cfs_spin_unlock(&config_list_lock);
447 cfs_spin_lock(&config_list_lock);
452 cfs_spin_unlock(&config_list_lock);
453 config_log_put(cld_prev);
454 cfs_spin_lock(&config_list_lock);
459 cfs_spin_unlock(&config_list_lock);
461 /* Wait a bit to see if anyone else needs a requeue */
462 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
464 cfs_spin_lock(&config_list_lock);
466 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
467 rq_state &= ~RQ_RUNNING;
468 cfs_spin_unlock(&config_list_lock);
470 CDEBUG(D_MGC, "Ending requeue thread\n");
474 /* Add a cld to the list to requeue. Start the requeue thread if needed.
475 We are responsible for dropping the config log reference from here on out. */
476 static int mgc_requeue_add(struct config_llog_data *cld, int later)
480 CDEBUG(D_INFO, "log %s: requeue (l=%d r=%d sp=%d st=%x)\n",
481 cld->cld_logname, later, cfs_atomic_read(&cld->cld_refcount),
482 cld->cld_stopping, rq_state);
483 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
485 /* Hold lock for rq_state */
486 cfs_spin_lock(&config_list_lock);
488 if (cld->cld_stopping || (rq_state & RQ_STOP)) {
489 cld->cld_lostlock = 0;
490 cfs_spin_unlock(&config_list_lock);
495 cld->cld_lostlock = 1;
497 if (!(rq_state & RQ_RUNNING)) {
498 LASSERT(rq_state == 0);
499 rq_state = RQ_RUNNING | (later ? RQ_LATER : RQ_NOW);
500 cfs_spin_unlock(&config_list_lock);
501 rc = cfs_create_thread(mgc_requeue_thread, NULL,
504 CERROR("log %s: cannot start requeue thread (%d),"
505 "no more log updates!\n", cld->cld_logname, rc);
506 /* Drop the ref, since the rq thread won't */
507 cld->cld_lostlock = 0;
513 rq_state |= later ? RQ_LATER : RQ_NOW;
514 cfs_spin_unlock(&config_list_lock);
515 cfs_waitq_signal(&rq_waitq);
521 /********************** class fns **********************/
523 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
524 struct vfsmount *mnt)
526 struct lvfs_run_ctxt saved;
527 struct lustre_sb_info *lsi = s2lsi(sb);
528 struct client_obd *cli = &obd->u.cli;
529 struct dentry *dentry;
535 LASSERT(lsi->lsi_srv_mnt == mnt);
537 /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
538 cfs_down(&cli->cl_mgc_sem);
540 cfs_cleanup_group_info();
542 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
543 if (IS_ERR(obd->obd_fsops)) {
544 cfs_up(&cli->cl_mgc_sem);
545 CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd),
546 PTR_ERR(obd->obd_fsops));
547 RETURN(PTR_ERR(obd->obd_fsops));
550 cli->cl_mgc_vfsmnt = mnt;
551 fsfilt_setup(obd, mnt->mnt_sb);
553 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
554 obd->obd_lvfs_ctxt.pwdmnt = mnt;
555 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
556 obd->obd_lvfs_ctxt.fs = get_ds();
558 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
559 dentry = ll_lookup_one_len(MOUNT_CONFIGS_DIR, cfs_fs_pwd(current->fs),
560 strlen(MOUNT_CONFIGS_DIR));
561 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
562 if (IS_ERR(dentry)) {
563 err = PTR_ERR(dentry);
564 CERROR("cannot lookup %s directory: rc = %d\n",
565 MOUNT_CONFIGS_DIR, err);
568 cli->cl_mgc_configs_dir = dentry;
570 /* We take an obd ref to insure that we can't get to mgc_cleanup
571 without calling mgc_fs_cleanup first. */
572 class_incref(obd, "mgc_fs", obd);
574 label = fsfilt_get_label(obd, mnt->mnt_sb);
576 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
578 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
582 fsfilt_put_ops(obd->obd_fsops);
583 obd->obd_fsops = NULL;
584 cli->cl_mgc_vfsmnt = NULL;
585 cfs_up(&cli->cl_mgc_sem);
589 static int mgc_fs_cleanup(struct obd_device *obd)
591 struct client_obd *cli = &obd->u.cli;
595 LASSERT(cli->cl_mgc_vfsmnt != NULL);
597 if (cli->cl_mgc_configs_dir != NULL) {
598 struct lvfs_run_ctxt saved;
599 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
600 l_dput(cli->cl_mgc_configs_dir);
601 cli->cl_mgc_configs_dir = NULL;
602 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
603 class_decref(obd, "mgc_fs", obd);
606 cli->cl_mgc_vfsmnt = NULL;
608 fsfilt_put_ops(obd->obd_fsops);
610 cfs_up(&cli->cl_mgc_sem);
615 static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0);
616 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
622 case OBD_CLEANUP_EARLY:
624 case OBD_CLEANUP_EXPORTS:
625 if (cfs_atomic_dec_and_test(&mgc_count)) {
626 /* Kick the requeue waitq - cld's should all be
628 cfs_spin_lock(&config_list_lock);
630 cfs_spin_unlock(&config_list_lock);
631 cfs_waitq_signal(&rq_waitq);
633 obd_cleanup_client_import(obd);
634 rc = obd_llog_finish(obd, 0);
636 CERROR("failed to cleanup llogging subsystems\n");
642 static int mgc_cleanup(struct obd_device *obd)
644 struct client_obd *cli = &obd->u.cli;
648 LASSERT(cli->cl_mgc_vfsmnt == NULL);
650 /* COMPAT_146 - old config logs may have added profiles we don't
652 if (obd->obd_type->typ_refcnt <= 1)
653 /* Only for the last mgc */
654 class_del_profiles();
656 lprocfs_obd_cleanup(obd);
659 rc = client_obd_cleanup(obd);
663 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
665 struct lprocfs_static_vars lvars;
671 rc = client_obd_setup(obd, lcfg);
673 GOTO(err_decref, rc);
675 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
677 CERROR("failed to setup llogging subsystems\n");
678 GOTO(err_cleanup, rc);
681 lprocfs_mgc_init_vars(&lvars);
682 lprocfs_obd_setup(obd, lvars.obd_vars);
683 sptlrpc_lprocfs_cliobd_attach(obd);
685 cfs_spin_lock(&config_list_lock);
686 cfs_atomic_inc(&mgc_count);
687 if (cfs_atomic_read(&mgc_count) == 1) {
688 rq_state &= ~RQ_STOP;
689 cfs_waitq_init(&rq_waitq);
691 cfs_spin_unlock(&config_list_lock);
696 client_obd_cleanup(obd);
702 /* based on ll_mdc_blocking_ast */
703 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
704 void *data, int flag)
706 struct lustre_handle lockh;
707 struct config_llog_data *cld = (struct config_llog_data *)data;
712 case LDLM_CB_BLOCKING:
713 /* mgs wants the lock, give it up... */
714 LDLM_DEBUG(lock, "MGC blocking CB");
715 ldlm_lock2handle(lock, &lockh);
716 rc = ldlm_cli_cancel(&lockh);
718 case LDLM_CB_CANCELING: {
719 /* We've given up the lock, prepare ourselves to update. */
720 LDLM_DEBUG(lock, "MGC cancel CB");
722 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
723 lock->l_resource->lr_name.name[0],
724 (char *)&lock->l_resource->lr_name.name[0]);
727 CERROR("missing data, won't requeue\n");
730 /* Are we done with this log? */
731 if (cld->cld_stopping) {
732 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
737 /* Make sure not to re-enqueue when the mgc is stopping
738 (we get called from client_disconnect_export) */
739 if (!lock->l_conn_export ||
740 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
741 CDEBUG(D_MGC, "log %s: disconnecting, won't requeue\n",
746 /* Did we fail to get the lock? */
747 if (lock->l_req_mode != lock->l_granted_mode &&
748 !cld->cld_is_sptlrpc) {
749 CDEBUG(D_MGC, "log %s: original grant failed, will "
750 "requeue later\n", cld->cld_logname);
751 /* Try to re-enqueue later */
752 rc = mgc_requeue_add(cld, 1);
756 rc = mgc_requeue_add(cld, 0);
765 CERROR("%s CB failed %d:\n", flag == LDLM_CB_BLOCKING ?
766 "blocking" : "cancel", rc);
767 LDLM_ERROR(lock, "MGC ast");
772 /* Not sure where this should go... */
773 #define MGC_ENQUEUE_LIMIT 50
774 #define MGC_TARGET_REG_LIMIT 10
775 #define MGC_SEND_PARAM_LIMIT 10
777 /* Send parameter to MGS*/
778 static int mgc_set_mgs_param(struct obd_export *exp,
779 struct mgs_send_param *msp)
781 struct ptlrpc_request *req;
782 struct mgs_send_param *req_msp, *rep_msp;
786 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
787 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
792 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
794 ptlrpc_req_finished(req);
798 memcpy(req_msp, msp, sizeof(*req_msp));
799 ptlrpc_request_set_replen(req);
801 /* Limit how long we will wait for the enqueue to complete */
802 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
803 rc = ptlrpc_queue_wait(req);
805 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
806 memcpy(msp, rep_msp, sizeof(*rep_msp));
809 ptlrpc_req_finished(req);
814 /* Take a config lock so we can get cancel notifications */
815 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
816 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
817 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
818 void *data, __u32 lvb_len, void *lvb_swabber,
819 struct lustre_handle *lockh)
821 struct config_llog_data *cld = (struct config_llog_data *)data;
822 struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
823 ldlm_completion_ast, NULL, NULL, data};
824 struct ptlrpc_request *req;
825 int short_limit = cld->cld_is_sptlrpc;
829 CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
830 cld->cld_resid.name[0]);
832 /* We can only drop this config log ref when we drop the lock */
833 if (config_log_get(cld))
834 RETURN(ELDLM_LOCK_ABORTED);
836 /* We need a callback for every lockholder, so don't try to
837 ldlm_lock_match (see rev 1.1.2.11.2.47) */
838 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
839 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
843 ptlrpc_request_set_replen(req);
844 /* check if this is server or client */
845 if (cld->cld_cfg.cfg_sb) {
846 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
847 if (lsi && (lsi->lsi_flags & LSI_SERVER))
850 /* Limit how long we will wait for the enqueue to complete */
851 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
852 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
854 /* A failed enqueue should still call the mgc_blocking_ast,
855 where it will be requeued if needed ("grant failed"). */
856 ptlrpc_req_finished(req);
860 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
861 __u32 mode, struct lustre_handle *lockh)
865 ldlm_lock_decref(lockh, mode);
870 /* Send target_reg message to MGS */
871 static int mgc_target_register(struct obd_export *exp,
872 struct mgs_target_info *mti)
874 struct ptlrpc_request *req;
875 struct mgs_target_info *req_mti, *rep_mti;
879 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
880 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
885 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
887 ptlrpc_req_finished(req);
891 memcpy(req_mti, mti, sizeof(*req_mti));
892 ptlrpc_request_set_replen(req);
893 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
894 /* Limit how long we will wait for the enqueue to complete */
895 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
897 rc = ptlrpc_queue_wait(req);
899 rep_mti = req_capsule_server_get(&req->rq_pill,
900 &RMF_MGS_TARGET_INFO);
901 memcpy(mti, rep_mti, sizeof(*rep_mti));
902 CDEBUG(D_MGC, "register %s got index = %d\n",
903 mti->mti_svname, mti->mti_stripe_index);
905 ptlrpc_req_finished(req);
910 int mgc_set_info_async(struct obd_export *exp, obd_count keylen,
911 void *key, obd_count vallen, void *val,
912 struct ptlrpc_request_set *set)
917 /* Turn off initial_recov after we try all backup servers once */
918 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
919 struct obd_import *imp = class_exp2cliimp(exp);
921 if (vallen != sizeof(int))
924 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
925 imp->imp_obd->obd_name, value,
926 imp->imp_deactive, imp->imp_invalid,
927 imp->imp_replayable, imp->imp_obd->obd_replayable,
928 ptlrpc_import_state_name(imp->imp_state));
929 /* Resurrect if we previously died */
930 if ((imp->imp_state != LUSTRE_IMP_FULL &&
931 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
932 ptlrpc_reconnect_import(imp);
935 /* FIXME move this to mgc_process_config */
936 if (KEY_IS(KEY_REGISTER_TARGET)) {
937 struct mgs_target_info *mti;
938 if (vallen != sizeof(struct mgs_target_info))
940 mti = (struct mgs_target_info *)val;
941 CDEBUG(D_MGC, "register_target %s %#x\n",
942 mti->mti_svname, mti->mti_flags);
943 rc = mgc_target_register(exp, mti);
946 if (KEY_IS(KEY_SET_FS)) {
947 struct super_block *sb = (struct super_block *)val;
948 struct lustre_sb_info *lsi;
949 if (vallen != sizeof(struct super_block))
952 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
954 CERROR("set_fs got %d\n", rc);
958 if (KEY_IS(KEY_CLEAR_FS)) {
961 rc = mgc_fs_cleanup(exp->exp_obd);
963 CERROR("clear_fs got %d\n", rc);
967 if (KEY_IS(KEY_SET_INFO)) {
968 struct mgs_send_param *msp;
970 msp = (struct mgs_send_param *)val;
971 rc = mgc_set_mgs_param(exp, msp);
974 if (KEY_IS(KEY_MGSSEC)) {
975 struct client_obd *cli = &exp->exp_obd->u.cli;
976 struct sptlrpc_flavor flvr;
979 * empty string means using current flavor, if which haven't
980 * been set yet, set it as null.
982 * if flavor has been set previously, check the asking flavor
983 * must match the existing one.
986 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
992 rc = sptlrpc_parse_flavor(val, &flvr);
994 CERROR("invalid sptlrpc flavor %s to MGS\n",
1000 * caller already hold a mutex
1002 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1003 cli->cl_flvr_mgc = flvr;
1004 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1005 sizeof(flvr)) != 0) {
1008 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1010 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1011 "currently %s is in use\n",
1021 static int mgc_import_event(struct obd_device *obd,
1022 struct obd_import *imp,
1023 enum obd_import_event event)
1027 LASSERT(imp->imp_obd == obd);
1028 CDEBUG(D_MGC, "import event %#x\n", event);
1031 case IMP_EVENT_DISCON:
1032 /* MGC imports should not wait for recovery */
1034 case IMP_EVENT_INACTIVE:
1036 case IMP_EVENT_INVALIDATE: {
1037 struct ldlm_namespace *ns = obd->obd_namespace;
1038 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1041 case IMP_EVENT_ACTIVE:
1042 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
1043 /* Clearing obd_no_recov allows us to continue pinging */
1044 obd->obd_no_recov = 0;
1048 case IMP_EVENT_DEACTIVATE:
1049 case IMP_EVENT_ACTIVATE:
1052 CERROR("Unknown import event %#x\n", event);
1058 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1059 struct obd_device *tgt, int *index)
1061 struct llog_ctxt *ctxt;
1065 LASSERT(olg == &obd->obd_olg);
1067 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
1072 rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
1075 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1077 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1082 llog_initiator_connect(ctxt);
1083 llog_ctxt_put(ctxt);
1085 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1093 static int mgc_llog_finish(struct obd_device *obd, int count)
1095 struct llog_ctxt *ctxt;
1096 int rc = 0, rc2 = 0;
1099 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1101 rc = llog_cleanup(ctxt);
1103 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1105 rc2 = llog_cleanup(ctxt);
1113 /* identical to mgs_log_is_empty */
1114 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1117 struct lvfs_run_ctxt saved;
1118 struct llog_handle *llh;
1121 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1122 rc = llog_create(ctxt, &llh, NULL, name);
1124 llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL);
1125 rc = llog_get_size(llh);
1128 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1129 /* header is record 1 */
1133 static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec,
1136 struct llog_rec_hdr local_rec = *rec;
1137 struct llog_handle *local_llh = (struct llog_handle *)data;
1138 char *cfg_buf = (char*) (rec + 1);
1139 struct lustre_cfg *lcfg;
1143 /* Append all records */
1144 local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
1145 rc = llog_write_rec(local_llh, &local_rec, NULL, 0,
1146 (void *)cfg_buf, -1);
1148 lcfg = (struct lustre_cfg *)cfg_buf;
1149 CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
1150 rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
1151 lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
1156 /* Copy a remote log locally */
1157 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1158 struct llog_ctxt *lctxt, char *logname)
1160 struct llog_handle *local_llh, *remote_llh;
1161 struct obd_uuid *uuid;
1166 /* Write new log to a temp name, then vfs_rename over logname
1167 upon successful completion. */
1169 OBD_ALLOC(temp_log, strlen(logname) + 1);
1172 sprintf(temp_log, "%sT", logname);
1174 /* Make sure there's no old temp log */
1175 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1178 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL);
1181 rc = llog_destroy(local_llh);
1182 llog_free_handle(local_llh);
1186 /* open local log */
1187 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1191 /* set the log header uuid for fun */
1192 OBD_ALLOC_PTR(uuid);
1193 obd_str2uuid(uuid, logname);
1194 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid);
1197 GOTO(out_closel, rc);
1199 /* open remote log */
1200 rc = llog_create(rctxt, &remote_llh, NULL, logname);
1202 GOTO(out_closel, rc);
1203 rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL);
1205 GOTO(out_closer, rc);
1207 /* Copy remote log */
1208 rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL);
1211 rc2 = llog_close(remote_llh);
1215 rc2 = llog_close(local_llh);
1219 /* We've copied the remote log to the local temp log, now
1220 replace the old local log with the temp log. */
1222 struct client_obd *cli = &obd->u.cli;
1224 LASSERT(cli->cl_mgc_configs_dir);
1225 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1228 CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1231 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1232 OBD_FREE(temp_log, strlen(logname) + 1);
1236 /** Get a config log from the MGS and process it.
1237 * This func is called for both clients and servers.
1238 * Copy the log locally before parsing it if appropriate (non-MGS server)
1240 int mgc_process_log(struct obd_device *mgc,
1241 struct config_llog_data *cld)
1243 struct llog_ctxt *ctxt, *lctxt;
1244 struct lustre_handle lockh;
1245 struct client_obd *cli = &mgc->u.cli;
1246 struct lvfs_run_ctxt *saved_ctxt;
1247 struct lustre_sb_info *lsi = NULL;
1248 int rc = 0, rcl, flags = 0, must_pop = 0;
1253 /* I don't want multiple processes running process_log at once --
1254 sounds like badness. It actually might be fine, as long as
1255 we're not trying to update from the same log
1256 simultaneously (in which case we should use a per-log sem.) */
1257 cfs_mutex_lock(&cld->cld_lock);
1259 if (cld->cld_stopping) {
1260 cfs_mutex_unlock(&cld->cld_lock);
1264 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1266 if (cld->cld_cfg.cfg_sb)
1267 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1269 CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1270 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1272 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1274 CERROR("missing llog context\n");
1275 cfs_mutex_unlock(&cld->cld_lock);
1279 OBD_ALLOC_PTR(saved_ctxt);
1280 if (saved_ctxt == NULL) {
1281 cfs_mutex_unlock(&cld->cld_lock);
1285 /* Get the cfg lock on the llog */
1286 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1287 LCK_CR, &flags, NULL, NULL, NULL,
1288 cld, 0, NULL, &lockh);
1290 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1292 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1295 * local copy of sptlrpc log is controlled elsewhere, don't try to
1298 if (rcl && cld->cld_is_sptlrpc)
1301 /* Copy the setup log locally if we can. Don't mess around if we're
1302 running an MGS though (logs are already local). */
1303 if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
1304 (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1305 !IS_MGS(lsi->lsi_ldd)) {
1306 push_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1309 /* Only try to copy log if we have the lock. */
1310 rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1312 if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1313 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1314 "log %s and no local copy."
1315 "\n", cld->cld_logname);
1316 GOTO(out_pop, rc = -ENOTCONN);
1318 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1319 "copy for now, will try to update later.\n",
1322 /* Now, whether we copied or not, start using the local llog.
1323 If we failed to copy, we'll start using whatever the old
1325 llog_ctxt_put(ctxt);
1329 if (cld->cld_is_sptlrpc)
1330 sptlrpc_conf_log_update_begin(cld->cld_logname);
1332 /* logname and instance info should be the same, so use our
1333 copy of the instance for the update. The cfg_last_idx will
1335 if (rcl == 0 || lctxt == ctxt)
1336 rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg);
1338 llog_ctxt_put(ctxt);
1340 llog_ctxt_put(lctxt);
1342 pop_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1344 OBD_FREE_PTR(saved_ctxt);
1346 * update settings on existing OBDs. doing it inside
1347 * of llog_process_lock so no device is attaching/detaching
1349 * the logname must be <fsname>-sptlrpc
1351 if (cld->cld_is_sptlrpc && rcl == 0) {
1352 sptlrpc_conf_log_update_end(cld->cld_logname);
1353 class_notify_sptlrpc_conf(cld->cld_logname,
1354 strlen(cld->cld_logname) -
1355 strlen("-sptlrpc"));
1358 /* Now drop the lock so MGS can revoke it */
1360 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1363 CERROR("Can't drop cfg lock: %d\n", rcl);
1366 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1367 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1369 cfs_mutex_unlock(&cld->cld_lock);
1374 /** Called from lustre_process_log.
1375 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1376 * any services, and adds it to the list logs to watch (follow).
1378 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1380 struct lustre_cfg *lcfg = buf;
1381 struct config_llog_instance *cfg = NULL;
1386 switch(lcfg->lcfg_command) {
1387 case LCFG_LOV_ADD_OBD: {
1388 /* Overloading this cfg command: register a new target */
1389 struct mgs_target_info *mti;
1391 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1392 sizeof(struct mgs_target_info))
1393 GOTO(out, rc = -EINVAL);
1395 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1396 CDEBUG(D_MGC, "add_target %s %#x\n",
1397 mti->mti_svname, mti->mti_flags);
1398 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1401 case LCFG_LOV_DEL_OBD:
1402 /* Unregister has no meaning at the moment. */
1403 CERROR("lov_del_obd unimplemented\n");
1406 case LCFG_SPTLRPC_CONF: {
1407 rc = sptlrpc_process_config(lcfg);
1410 case LCFG_LOG_START: {
1411 struct config_llog_data *cld;
1412 struct super_block *sb;
1414 logname = lustre_cfg_string(lcfg, 1);
1415 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1416 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1418 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1421 /* We're only called through here on the initial mount */
1422 rc = config_log_add(obd, logname, cfg, sb);
1425 cld = config_log_find(logname, cfg);
1432 /* FIXME only set this for old logs! Right now this forces
1433 us to always skip the "inside markers" check */
1434 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1436 rc = mgc_process_log(obd, cld);
1437 config_log_put(cld);
1441 case LCFG_LOG_END: {
1442 logname = lustre_cfg_string(lcfg, 1);
1444 if (lcfg->lcfg_bufcount >= 2)
1445 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1447 rc = config_log_end(logname, cfg);
1451 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1452 GOTO(out, rc = -EINVAL);
1460 struct obd_ops mgc_obd_ops = {
1461 .o_owner = THIS_MODULE,
1462 .o_setup = mgc_setup,
1463 .o_precleanup = mgc_precleanup,
1464 .o_cleanup = mgc_cleanup,
1465 .o_add_conn = client_import_add_conn,
1466 .o_del_conn = client_import_del_conn,
1467 .o_connect = client_connect_import,
1468 .o_disconnect = client_disconnect_export,
1469 //.o_enqueue = mgc_enqueue,
1470 .o_cancel = mgc_cancel,
1471 //.o_iocontrol = mgc_iocontrol,
1472 .o_set_info_async = mgc_set_info_async,
1473 .o_import_event = mgc_import_event,
1474 .o_llog_init = mgc_llog_init,
1475 .o_llog_finish = mgc_llog_finish,
1476 .o_process_config = mgc_process_config,
1479 int __init mgc_init(void)
1481 return class_register_type(&mgc_obd_ops, NULL, NULL,
1482 LUSTRE_MGC_NAME, NULL);
1486 static void /*__exit*/ mgc_exit(void)
1488 class_unregister_type(LUSTRE_MGC_NAME);
1491 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1492 MODULE_DESCRIPTION("Lustre Management Client");
1493 MODULE_LICENSE("GPL");
1495 module_init(mgc_init);
1496 module_exit(mgc_exit);