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.
32 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * lustre/mgc/mgc_request.c
41 * Author: Nathan Rutman <nathan@clusterfs.com>
45 # define EXPORT_SYMTAB
47 #define DEBUG_SUBSYSTEM S_MGC
48 #define D_MGC D_CONFIG /*|D_WARNING*/
51 # include <linux/module.h>
52 # include <linux/pagemap.h>
53 # include <linux/miscdevice.h>
54 # include <linux/init.h>
56 # include <liblustre.h>
59 #include <obd_class.h>
60 #include <lustre_dlm.h>
61 #include <lprocfs_status.h>
62 #include <lustre_log.h>
63 #include <lustre_fsfilt.h>
64 #include <lustre_disk.h>
65 #include "mgc_internal.h"
67 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id)
72 CERROR("name too long: %s\n", name);
76 CERROR("missing name: %s\n", name);
79 memcpy(&resname, name, len);
81 memset(res_id, 0, sizeof(*res_id));
83 /* Always use the same endianness for the resid */
84 res_id->name[0] = cpu_to_le64(resname);
85 CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
86 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
90 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id)
92 /* fsname is at most 8 chars long, maybe contain "-".
93 * e.g. "lustre", "SUN-000" */
94 return mgc_name2resid(fsname, strlen(fsname), res_id);
96 EXPORT_SYMBOL(mgc_fsname2resid);
98 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id)
103 /* logname consists of "fsname-nodetype".
104 * e.g. "lustre-MDT0001", "SUN-000-client" */
105 name_end = strrchr(logname, '-');
107 len = name_end - logname;
108 return mgc_name2resid(logname, len, res_id);
111 /********************** config llog list **********************/
112 static CFS_LIST_HEAD(config_llog_list);
113 static cfs_spinlock_t config_list_lock = CFS_SPIN_LOCK_UNLOCKED;
115 /* Take a reference to a config log */
116 static int config_log_get(struct config_llog_data *cld)
119 if (cld->cld_stopping)
121 cfs_atomic_inc(&cld->cld_refcount);
122 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
123 cfs_atomic_read(&cld->cld_refcount));
127 /* Drop a reference to a config log. When no longer referenced,
128 we can free the config log data */
129 static void config_log_put(struct config_llog_data *cld)
133 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
134 cfs_atomic_read(&cld->cld_refcount));
135 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
137 /* spinlock to make sure no item with 0 refcount in the list */
138 cfs_spin_lock(&config_list_lock);
139 if (unlikely(cfs_atomic_dec_and_test(&cld->cld_refcount))) {
140 cfs_list_del(&cld->cld_list_chain);
141 cfs_spin_unlock(&config_list_lock);
143 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
145 if (cld->cld_sptlrpc)
146 config_log_put(cld->cld_sptlrpc);
147 if (cld->cld_is_sptlrpc)
148 sptlrpc_conf_log_stop(cld->cld_logname);
150 class_export_put(cld->cld_mgcexp);
151 OBD_FREE(cld->cld_logname, strlen(cld->cld_logname) + 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 struct config_llog_data *found = NULL;
170 LASSERT(logname != NULL);
172 instance = cfg ? cfg->cfg_instance : NULL;
173 cfs_spin_lock(&config_list_lock);
174 cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
175 /* check if instance equals */
176 if (instance != cld->cld_cfg.cfg_instance)
179 /* instance may be NULL, should check name */
180 if (strcmp(logname, cld->cld_logname) == 0) {
186 cfs_atomic_inc(&found->cld_refcount);
187 LASSERT(found->cld_stopping == 0 || found->cld_is_sptlrpc == 0);
189 cfs_spin_unlock(&config_list_lock);
194 struct config_llog_data *do_config_log_add(struct obd_device *obd,
196 unsigned int is_sptlrpc,
197 struct config_llog_instance *cfg,
198 struct super_block *sb)
200 struct config_llog_data *cld;
204 CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
205 cfg ? cfg->cfg_instance : 0);
207 OBD_ALLOC(cld, sizeof(*cld));
209 RETURN(ERR_PTR(-ENOMEM));
210 OBD_ALLOC(cld->cld_logname, strlen(logname) + 1);
211 if (!cld->cld_logname) {
212 OBD_FREE(cld, sizeof(*cld));
213 RETURN(ERR_PTR(-ENOMEM));
215 strcpy(cld->cld_logname, logname);
218 cfs_mutex_init(&cld->cld_lock);
219 cld->cld_cfg.cfg_last_idx = 0;
220 cld->cld_cfg.cfg_flags = 0;
221 cld->cld_cfg.cfg_sb = sb;
222 cld->cld_is_sptlrpc = is_sptlrpc;
223 cfs_atomic_set(&cld->cld_refcount, 1);
225 /* Keep the mgc around until we are done */
226 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
229 sptlrpc_conf_log_start(logname);
230 cld->cld_cfg.cfg_obdname = obd->obd_name;
233 rc = mgc_logname2resid(logname, &cld->cld_resid);
235 cfs_spin_lock(&config_list_lock);
236 cfs_list_add(&cld->cld_list_chain, &config_llog_list);
237 cfs_spin_unlock(&config_list_lock);
245 rc = mgc_process_log(obd, cld);
247 CERROR("failed processing sptlrpc log: %d\n", rc);
253 /** Add this log to the list of active logs watched by an MGC.
254 * Active means we're watching for updates.
255 * We have one active log per "mount" - client instance or servername.
256 * Each instance may be at a different point in the log.
258 static int config_log_add(struct obd_device *obd, char *logname,
259 struct config_llog_instance *cfg,
260 struct super_block *sb)
262 struct config_llog_data *cld, *sptlrpc_cld;
267 CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
270 * for each regular log, the depended sptlrpc log name is
271 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
273 ptr = strrchr(logname, '-');
274 if (ptr == NULL || ptr - logname > 8) {
275 CERROR("logname %s is too long\n", logname);
279 memcpy(seclogname, logname, ptr - logname);
280 strcpy(seclogname + (ptr - logname), "-sptlrpc");
282 sptlrpc_cld = config_log_find(seclogname, NULL);
283 if (sptlrpc_cld == NULL) {
284 sptlrpc_cld = do_config_log_add(obd, seclogname, 1, NULL, NULL);
285 if (IS_ERR(sptlrpc_cld)) {
286 CERROR("can't create sptlrpc log: %s\n", seclogname);
287 RETURN(PTR_ERR(sptlrpc_cld));
291 cld = do_config_log_add(obd, logname, 0, cfg, sb);
293 CERROR("can't create log: %s\n", logname);
294 config_log_put(sptlrpc_cld);
295 RETURN(PTR_ERR(cld));
298 cld->cld_sptlrpc = sptlrpc_cld;
303 CFS_DECLARE_MUTEX(llog_process_lock);
305 /** Stop watching for updates on this log.
307 static int config_log_end(char *logname, struct config_llog_instance *cfg)
309 struct config_llog_data *cld, *cld_sptlrpc = NULL;
313 cld = config_log_find(logname, cfg);
317 cfs_mutex_lock(&cld->cld_lock);
319 * if cld_stopping is set, it means we didn't start the log thus
320 * not owning the start ref. this can happen after previous umount:
321 * the cld still hanging there waiting for lock cancel, and we
322 * remount again but failed in the middle and call log_end without
325 if (unlikely(cld->cld_stopping)) {
326 cfs_mutex_unlock(&cld->cld_lock);
327 /* drop the ref from the find */
332 cld->cld_stopping = 1;
333 cfs_mutex_unlock(&cld->cld_lock);
335 cfs_spin_lock(&config_list_lock);
336 cld_sptlrpc = cld->cld_sptlrpc;
337 cld->cld_sptlrpc = NULL;
338 cfs_spin_unlock(&config_list_lock);
341 config_log_put(cld_sptlrpc);
343 /* drop the ref from the find */
345 /* drop the start ref */
348 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
353 /* reenqueue any lost locks */
354 #define RQ_RUNNING 0x1
358 static int rq_state = 0;
359 static cfs_waitq_t rq_waitq;
361 static int mgc_requeue_add(struct config_llog_data *cld, int later);
363 static void do_requeue(struct config_llog_data *cld)
365 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
367 /* Do not run mgc_process_log on a disconnected export or an
368 export which is being disconnected. Take the client
369 semaphore to make the check non-racy. */
370 cfs_down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
371 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
372 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
373 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
375 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
378 cfs_up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
380 /* Whether we enqueued again or not in mgc_process_log, we're done
381 * with the ref from the old enqueue */
385 static int mgc_requeue_thread(void *data)
387 struct l_wait_info lwi_now, lwi_later;
388 struct config_llog_data *cld, *cld_next, *cld_prev;
389 char name[] = "ll_cfg_requeue";
395 CDEBUG(D_MGC, "Starting requeue thread\n");
397 lwi_later = LWI_TIMEOUT(60 * CFS_HZ, NULL, NULL);
398 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP), &lwi_later);
400 /* Keep trying failed locks periodically */
401 cfs_spin_lock(&config_list_lock);
402 while (rq_state & (RQ_NOW | RQ_LATER)) {
403 /* Any new or requeued lostlocks will change the state */
404 rq_state &= ~(RQ_NOW | RQ_LATER);
405 cfs_spin_unlock(&config_list_lock);
407 /* Always wait a few seconds to allow the server who
408 caused the lock revocation to finish its setup, plus some
409 random so everyone doesn't try to reconnect at once. */
410 lwi_now = LWI_TIMEOUT(3 * CFS_HZ + (cfs_rand() & 0xff) * \
413 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi_now);
416 * iterate & processing through the list. for each cld, process
417 * its depending sptlrpc cld firstly (if any) and then itself.
419 * it's guaranteed any item in the list must have
420 * reference > 0; and if cld_lostlock is set, at
421 * least one reference is taken by the previous enqueue.
423 * Note: releasing a cld might lead to itself and its depended
424 * sptlrpc cld be unlinked from the list. to safely iterate
425 * we need to take a reference on next cld before processing.
429 cfs_spin_lock(&config_list_lock);
430 cfs_list_for_each_entry_safe(cld, cld_next, &config_llog_list,
432 if (cld->cld_list_chain.next != &config_llog_list)
433 cfs_atomic_inc(&cld_next->cld_refcount);
435 if (cld->cld_lostlock) {
436 if (cld->cld_sptlrpc &&
437 cld->cld_sptlrpc->cld_lostlock) {
438 cld->cld_sptlrpc->cld_lostlock = 0;
440 cfs_spin_unlock(&config_list_lock);
441 do_requeue(cld->cld_sptlrpc);
442 cfs_spin_lock(&config_list_lock);
443 LASSERT(cld->cld_lostlock);
446 cld->cld_lostlock = 0;
448 cfs_spin_unlock(&config_list_lock);
450 cfs_spin_lock(&config_list_lock);
455 cfs_spin_unlock(&config_list_lock);
456 config_log_put(cld_prev);
457 cfs_spin_lock(&config_list_lock);
462 cfs_spin_unlock(&config_list_lock);
464 /* Wait a bit to see if anyone else needs a requeue */
465 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
467 cfs_spin_lock(&config_list_lock);
469 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
470 rq_state &= ~RQ_RUNNING;
471 cfs_spin_unlock(&config_list_lock);
473 CDEBUG(D_MGC, "Ending requeue thread\n");
477 /* Add a cld to the list to requeue. Start the requeue thread if needed.
478 We are responsible for dropping the config log reference from here on out. */
479 static int mgc_requeue_add(struct config_llog_data *cld, int later)
483 CDEBUG(D_INFO, "log %s: requeue (l=%d r=%d sp=%d st=%x)\n",
484 cld->cld_logname, later, cfs_atomic_read(&cld->cld_refcount),
485 cld->cld_stopping, rq_state);
486 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
488 /* Hold lock for rq_state */
489 cfs_spin_lock(&config_list_lock);
491 if (cld->cld_stopping || (rq_state & RQ_STOP)) {
492 cld->cld_lostlock = 0;
493 cfs_spin_unlock(&config_list_lock);
498 cld->cld_lostlock = 1;
500 if (!(rq_state & RQ_RUNNING)) {
501 LASSERT(rq_state == 0);
502 rq_state = RQ_RUNNING | (later ? RQ_LATER : RQ_NOW);
503 cfs_spin_unlock(&config_list_lock);
504 rc = cfs_create_thread(mgc_requeue_thread, NULL,
507 CERROR("log %s: cannot start requeue thread (%d),"
508 "no more log updates!\n", cld->cld_logname, rc);
509 /* Drop the ref, since the rq thread won't */
510 cld->cld_lostlock = 0;
516 rq_state |= later ? RQ_LATER : RQ_NOW;
517 cfs_spin_unlock(&config_list_lock);
518 cfs_waitq_signal(&rq_waitq);
524 /********************** class fns **********************/
526 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
527 struct vfsmount *mnt)
529 struct lvfs_run_ctxt saved;
530 struct lustre_sb_info *lsi = s2lsi(sb);
531 struct client_obd *cli = &obd->u.cli;
532 struct dentry *dentry;
538 LASSERT(lsi->lsi_srv_mnt == mnt);
540 /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
541 cfs_down(&cli->cl_mgc_sem);
543 cfs_cleanup_group_info();
545 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
546 if (IS_ERR(obd->obd_fsops)) {
547 cfs_up(&cli->cl_mgc_sem);
548 CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd),
549 PTR_ERR(obd->obd_fsops));
550 RETURN(PTR_ERR(obd->obd_fsops));
553 cli->cl_mgc_vfsmnt = mnt;
554 fsfilt_setup(obd, mnt->mnt_sb);
556 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
557 obd->obd_lvfs_ctxt.pwdmnt = mnt;
558 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
559 obd->obd_lvfs_ctxt.fs = get_ds();
561 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
562 dentry = ll_lookup_one_len(MOUNT_CONFIGS_DIR, cfs_fs_pwd(current->fs),
563 strlen(MOUNT_CONFIGS_DIR));
564 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
565 if (IS_ERR(dentry)) {
566 err = PTR_ERR(dentry);
567 CERROR("cannot lookup %s directory: rc = %d\n",
568 MOUNT_CONFIGS_DIR, err);
571 cli->cl_mgc_configs_dir = dentry;
573 /* We take an obd ref to insure that we can't get to mgc_cleanup
574 without calling mgc_fs_cleanup first. */
575 class_incref(obd, "mgc_fs", obd);
577 label = fsfilt_get_label(obd, mnt->mnt_sb);
579 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
581 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
585 fsfilt_put_ops(obd->obd_fsops);
586 obd->obd_fsops = NULL;
587 cli->cl_mgc_vfsmnt = NULL;
588 cfs_up(&cli->cl_mgc_sem);
592 static int mgc_fs_cleanup(struct obd_device *obd)
594 struct client_obd *cli = &obd->u.cli;
598 LASSERT(cli->cl_mgc_vfsmnt != NULL);
600 if (cli->cl_mgc_configs_dir != NULL) {
601 struct lvfs_run_ctxt saved;
602 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
603 l_dput(cli->cl_mgc_configs_dir);
604 cli->cl_mgc_configs_dir = NULL;
605 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
606 class_decref(obd, "mgc_fs", obd);
609 cli->cl_mgc_vfsmnt = NULL;
611 fsfilt_put_ops(obd->obd_fsops);
613 cfs_up(&cli->cl_mgc_sem);
618 static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0);
619 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
625 case OBD_CLEANUP_EARLY:
627 case OBD_CLEANUP_EXPORTS:
628 if (cfs_atomic_dec_and_test(&mgc_count)) {
629 /* Kick the requeue waitq - cld's should all be
631 cfs_spin_lock(&config_list_lock);
633 cfs_spin_unlock(&config_list_lock);
634 cfs_waitq_signal(&rq_waitq);
636 obd_cleanup_client_import(obd);
637 rc = obd_llog_finish(obd, 0);
639 CERROR("failed to cleanup llogging subsystems\n");
645 static int mgc_cleanup(struct obd_device *obd)
647 struct client_obd *cli = &obd->u.cli;
651 LASSERT(cli->cl_mgc_vfsmnt == NULL);
653 /* COMPAT_146 - old config logs may have added profiles we don't
655 if (obd->obd_type->typ_refcnt <= 1)
656 /* Only for the last mgc */
657 class_del_profiles();
659 lprocfs_obd_cleanup(obd);
662 rc = client_obd_cleanup(obd);
666 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
668 struct lprocfs_static_vars lvars;
674 rc = client_obd_setup(obd, lcfg);
676 GOTO(err_decref, rc);
678 rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
680 CERROR("failed to setup llogging subsystems\n");
681 GOTO(err_cleanup, rc);
684 lprocfs_mgc_init_vars(&lvars);
685 lprocfs_obd_setup(obd, lvars.obd_vars);
686 sptlrpc_lprocfs_cliobd_attach(obd);
688 cfs_spin_lock(&config_list_lock);
689 cfs_atomic_inc(&mgc_count);
690 if (cfs_atomic_read(&mgc_count) == 1) {
691 rq_state &= ~RQ_STOP;
692 cfs_waitq_init(&rq_waitq);
694 cfs_spin_unlock(&config_list_lock);
699 client_obd_cleanup(obd);
705 /* based on ll_mdc_blocking_ast */
706 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
707 void *data, int flag)
709 struct lustre_handle lockh;
710 struct config_llog_data *cld = (struct config_llog_data *)data;
715 case LDLM_CB_BLOCKING:
716 /* mgs wants the lock, give it up... */
717 LDLM_DEBUG(lock, "MGC blocking CB");
718 ldlm_lock2handle(lock, &lockh);
719 rc = ldlm_cli_cancel(&lockh);
721 case LDLM_CB_CANCELING: {
722 /* We've given up the lock, prepare ourselves to update. */
723 LDLM_DEBUG(lock, "MGC cancel CB");
725 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
726 lock->l_resource->lr_name.name[0],
727 (char *)&lock->l_resource->lr_name.name[0]);
730 CERROR("missing data, won't requeue\n");
733 /* Are we done with this log? */
734 if (cld->cld_stopping) {
735 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
740 /* Make sure not to re-enqueue when the mgc is stopping
741 (we get called from client_disconnect_export) */
742 if (!lock->l_conn_export ||
743 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
744 CDEBUG(D_MGC, "log %s: disconnecting, won't requeue\n",
749 /* Did we fail to get the lock? */
750 if (lock->l_req_mode != lock->l_granted_mode &&
751 !cld->cld_is_sptlrpc) {
752 CDEBUG(D_MGC, "log %s: original grant failed, will "
753 "requeue later\n", cld->cld_logname);
754 /* Try to re-enqueue later */
755 rc = mgc_requeue_add(cld, 1);
759 rc = mgc_requeue_add(cld, 0);
768 CERROR("%s CB failed %d:\n", flag == LDLM_CB_BLOCKING ?
769 "blocking" : "cancel", rc);
770 LDLM_ERROR(lock, "MGC ast");
775 /* Not sure where this should go... */
776 #define MGC_ENQUEUE_LIMIT 50
777 #define MGC_TARGET_REG_LIMIT 10
778 #define MGC_SEND_PARAM_LIMIT 10
780 /* Send parameter to MGS*/
781 static int mgc_set_mgs_param(struct obd_export *exp,
782 struct mgs_send_param *msp)
784 struct ptlrpc_request *req;
785 struct mgs_send_param *req_msp, *rep_msp;
789 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
790 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
795 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
797 ptlrpc_req_finished(req);
801 memcpy(req_msp, msp, sizeof(*req_msp));
802 ptlrpc_request_set_replen(req);
804 /* Limit how long we will wait for the enqueue to complete */
805 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
806 rc = ptlrpc_queue_wait(req);
808 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
809 memcpy(msp, rep_msp, sizeof(*rep_msp));
812 ptlrpc_req_finished(req);
817 /* Take a config lock so we can get cancel notifications */
818 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
819 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
820 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
821 void *data, __u32 lvb_len, void *lvb_swabber,
822 struct lustre_handle *lockh)
824 struct config_llog_data *cld = (struct config_llog_data *)data;
825 struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
826 ldlm_completion_ast, NULL, NULL, data};
827 struct ptlrpc_request *req;
828 int short_limit = cld->cld_is_sptlrpc;
832 CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
833 cld->cld_resid.name[0]);
835 /* We can only drop this config log ref when we drop the lock */
836 if (config_log_get(cld))
837 RETURN(ELDLM_LOCK_ABORTED);
839 /* We need a callback for every lockholder, so don't try to
840 ldlm_lock_match (see rev 1.1.2.11.2.47) */
841 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
842 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
846 ptlrpc_request_set_replen(req);
847 /* check if this is server or client */
848 if (cld->cld_cfg.cfg_sb) {
849 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
850 if (lsi && (lsi->lsi_flags & LSI_SERVER))
853 /* Limit how long we will wait for the enqueue to complete */
854 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
855 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
857 /* A failed enqueue should still call the mgc_blocking_ast,
858 where it will be requeued if needed ("grant failed"). */
859 ptlrpc_req_finished(req);
863 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
864 __u32 mode, struct lustre_handle *lockh)
868 ldlm_lock_decref(lockh, mode);
873 /* Send target_reg message to MGS */
874 static int mgc_target_register(struct obd_export *exp,
875 struct mgs_target_info *mti)
877 struct ptlrpc_request *req;
878 struct mgs_target_info *req_mti, *rep_mti;
882 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
883 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
888 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
890 ptlrpc_req_finished(req);
894 memcpy(req_mti, mti, sizeof(*req_mti));
895 ptlrpc_request_set_replen(req);
896 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
897 /* Limit how long we will wait for the enqueue to complete */
898 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
900 rc = ptlrpc_queue_wait(req);
902 rep_mti = req_capsule_server_get(&req->rq_pill,
903 &RMF_MGS_TARGET_INFO);
904 memcpy(mti, rep_mti, sizeof(*rep_mti));
905 CDEBUG(D_MGC, "register %s got index = %d\n",
906 mti->mti_svname, mti->mti_stripe_index);
908 ptlrpc_req_finished(req);
913 int mgc_set_info_async(struct obd_export *exp, obd_count keylen,
914 void *key, obd_count vallen, void *val,
915 struct ptlrpc_request_set *set)
920 /* Turn off initial_recov after we try all backup servers once */
921 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
922 struct obd_import *imp = class_exp2cliimp(exp);
924 if (vallen != sizeof(int))
927 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
928 imp->imp_obd->obd_name, value,
929 imp->imp_deactive, imp->imp_invalid,
930 imp->imp_replayable, imp->imp_obd->obd_replayable,
931 ptlrpc_import_state_name(imp->imp_state));
932 /* Resurrect if we previously died */
933 if ((imp->imp_state != LUSTRE_IMP_FULL &&
934 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
935 ptlrpc_reconnect_import(imp);
938 /* FIXME move this to mgc_process_config */
939 if (KEY_IS(KEY_REGISTER_TARGET)) {
940 struct mgs_target_info *mti;
941 if (vallen != sizeof(struct mgs_target_info))
943 mti = (struct mgs_target_info *)val;
944 CDEBUG(D_MGC, "register_target %s %#x\n",
945 mti->mti_svname, mti->mti_flags);
946 rc = mgc_target_register(exp, mti);
949 if (KEY_IS(KEY_SET_FS)) {
950 struct super_block *sb = (struct super_block *)val;
951 struct lustre_sb_info *lsi;
952 if (vallen != sizeof(struct super_block))
955 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
957 CERROR("set_fs got %d\n", rc);
961 if (KEY_IS(KEY_CLEAR_FS)) {
964 rc = mgc_fs_cleanup(exp->exp_obd);
966 CERROR("clear_fs got %d\n", rc);
970 if (KEY_IS(KEY_SET_INFO)) {
971 struct mgs_send_param *msp;
973 msp = (struct mgs_send_param *)val;
974 rc = mgc_set_mgs_param(exp, msp);
977 if (KEY_IS(KEY_MGSSEC)) {
978 struct client_obd *cli = &exp->exp_obd->u.cli;
979 struct sptlrpc_flavor flvr;
982 * empty string means using current flavor, if which haven't
983 * been set yet, set it as null.
985 * if flavor has been set previously, check the asking flavor
986 * must match the existing one.
989 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
995 rc = sptlrpc_parse_flavor(val, &flvr);
997 CERROR("invalid sptlrpc flavor %s to MGS\n",
1003 * caller already hold a mutex
1005 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1006 cli->cl_flvr_mgc = flvr;
1007 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1008 sizeof(flvr)) != 0) {
1011 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1013 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1014 "currently %s is in use\n",
1024 static int mgc_import_event(struct obd_device *obd,
1025 struct obd_import *imp,
1026 enum obd_import_event event)
1030 LASSERT(imp->imp_obd == obd);
1031 CDEBUG(D_MGC, "import event %#x\n", event);
1034 case IMP_EVENT_DISCON:
1035 /* MGC imports should not wait for recovery */
1037 case IMP_EVENT_INACTIVE:
1039 case IMP_EVENT_INVALIDATE: {
1040 struct ldlm_namespace *ns = obd->obd_namespace;
1041 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1044 case IMP_EVENT_ACTIVE:
1045 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
1046 /* Clearing obd_no_recov allows us to continue pinging */
1047 obd->obd_no_recov = 0;
1051 case IMP_EVENT_DEACTIVATE:
1052 case IMP_EVENT_ACTIVATE:
1055 CERROR("Unknown import event %#x\n", event);
1061 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1062 struct obd_device *tgt, int *index)
1064 struct llog_ctxt *ctxt;
1068 LASSERT(olg == &obd->obd_olg);
1070 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
1075 rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
1078 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1080 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1085 llog_initiator_connect(ctxt);
1086 llog_ctxt_put(ctxt);
1088 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1096 static int mgc_llog_finish(struct obd_device *obd, int count)
1098 struct llog_ctxt *ctxt;
1099 int rc = 0, rc2 = 0;
1102 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1104 rc = llog_cleanup(ctxt);
1106 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1108 rc2 = llog_cleanup(ctxt);
1116 /* identical to mgs_log_is_empty */
1117 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1120 struct lvfs_run_ctxt saved;
1121 struct llog_handle *llh;
1124 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1125 rc = llog_create(ctxt, &llh, NULL, name);
1127 llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL);
1128 rc = llog_get_size(llh);
1131 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1132 /* header is record 1 */
1136 static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec,
1139 struct llog_rec_hdr local_rec = *rec;
1140 struct llog_handle *local_llh = (struct llog_handle *)data;
1141 char *cfg_buf = (char*) (rec + 1);
1142 struct lustre_cfg *lcfg;
1146 /* Append all records */
1147 local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
1148 rc = llog_write_rec(local_llh, &local_rec, NULL, 0,
1149 (void *)cfg_buf, -1);
1151 lcfg = (struct lustre_cfg *)cfg_buf;
1152 CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
1153 rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
1154 lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
1159 /* Copy a remote log locally */
1160 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1161 struct llog_ctxt *lctxt, char *logname)
1163 struct llog_handle *local_llh, *remote_llh;
1164 struct obd_uuid *uuid;
1169 /* Write new log to a temp name, then vfs_rename over logname
1170 upon successful completion. */
1172 OBD_ALLOC(temp_log, strlen(logname) + 1);
1175 sprintf(temp_log, "%sT", logname);
1177 /* Make sure there's no old temp log */
1178 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1181 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL);
1184 rc = llog_destroy(local_llh);
1185 llog_free_handle(local_llh);
1189 /* open local log */
1190 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1194 /* set the log header uuid for fun */
1195 OBD_ALLOC_PTR(uuid);
1196 obd_str2uuid(uuid, logname);
1197 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid);
1200 GOTO(out_closel, rc);
1202 /* open remote log */
1203 rc = llog_create(rctxt, &remote_llh, NULL, logname);
1205 GOTO(out_closel, rc);
1206 rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL);
1208 GOTO(out_closer, rc);
1210 /* Copy remote log */
1211 rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL);
1214 rc2 = llog_close(remote_llh);
1218 rc2 = llog_close(local_llh);
1222 /* We've copied the remote log to the local temp log, now
1223 replace the old local log with the temp log. */
1225 struct client_obd *cli = &obd->u.cli;
1227 LASSERT(cli->cl_mgc_configs_dir);
1228 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1231 CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1234 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1235 OBD_FREE(temp_log, strlen(logname) + 1);
1239 /** Get a config log from the MGS and process it.
1240 * This func is called for both clients and servers.
1241 * Copy the log locally before parsing it if appropriate (non-MGS server)
1243 int mgc_process_log(struct obd_device *mgc,
1244 struct config_llog_data *cld)
1246 struct llog_ctxt *ctxt, *lctxt;
1247 struct lustre_handle lockh;
1248 struct client_obd *cli = &mgc->u.cli;
1249 struct lvfs_run_ctxt *saved_ctxt;
1250 struct lustre_sb_info *lsi = NULL;
1251 int rc = 0, rcl, flags = 0, must_pop = 0;
1256 /* I don't want multiple processes running process_log at once --
1257 sounds like badness. It actually might be fine, as long as
1258 we're not trying to update from the same log
1259 simultaneously (in which case we should use a per-log sem.) */
1260 cfs_mutex_lock(&cld->cld_lock);
1262 if (cld->cld_stopping) {
1263 cfs_mutex_unlock(&cld->cld_lock);
1267 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1269 if (cld->cld_cfg.cfg_sb)
1270 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1272 CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1273 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1275 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1277 CERROR("missing llog context\n");
1278 cfs_mutex_unlock(&cld->cld_lock);
1282 OBD_ALLOC_PTR(saved_ctxt);
1283 if (saved_ctxt == NULL) {
1284 cfs_mutex_unlock(&cld->cld_lock);
1288 /* Get the cfg lock on the llog */
1289 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1290 LCK_CR, &flags, NULL, NULL, NULL,
1291 cld, 0, NULL, &lockh);
1293 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1295 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1298 * local copy of sptlrpc log is controlled elsewhere, don't try to
1301 if (rcl && cld->cld_is_sptlrpc)
1304 /* Copy the setup log locally if we can. Don't mess around if we're
1305 running an MGS though (logs are already local). */
1306 if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
1307 (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1308 !IS_MGS(lsi->lsi_ldd)) {
1309 push_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1312 /* Only try to copy log if we have the lock. */
1313 rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1315 if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1316 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1317 "log %s and no local copy."
1318 "\n", cld->cld_logname);
1319 GOTO(out_pop, rc = -ENOTCONN);
1321 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1322 "copy for now, will try to update later.\n",
1325 /* Now, whether we copied or not, start using the local llog.
1326 If we failed to copy, we'll start using whatever the old
1328 llog_ctxt_put(ctxt);
1332 if (cld->cld_is_sptlrpc)
1333 sptlrpc_conf_log_update_begin(cld->cld_logname);
1335 /* logname and instance info should be the same, so use our
1336 copy of the instance for the update. The cfg_last_idx will
1338 if (rcl == 0 || lctxt == ctxt)
1339 rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg);
1341 llog_ctxt_put(ctxt);
1343 llog_ctxt_put(lctxt);
1345 pop_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1347 OBD_FREE_PTR(saved_ctxt);
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_mutex_unlock(&cld->cld_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;
1384 struct config_llog_instance *cfg = NULL;
1389 switch(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 super_block *sb;
1417 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 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);