1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mgc/mgc_request.c
38 * Author: Nathan Rutman <nathan@clusterfs.com>
42 # define EXPORT_SYMTAB
44 #define DEBUG_SUBSYSTEM S_MGC
45 #define D_MGC D_CONFIG /*|D_WARNING*/
48 # include <linux/module.h>
49 # include <linux/pagemap.h>
50 # include <linux/miscdevice.h>
51 # include <linux/init.h>
53 # include <liblustre.h>
56 #include <obd_class.h>
57 #include <lustre_dlm.h>
58 #include <lprocfs_status.h>
59 #include <lustre_log.h>
60 #include <lustre_fsfilt.h>
61 #include <lustre_disk.h>
62 #include "mgc_internal.h"
64 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id)
69 CERROR("name too long: %s\n", name);
73 CERROR("missing name: %s\n", name);
76 memcpy(&resname, name, len);
78 memset(res_id, 0, sizeof(*res_id));
80 /* Always use the same endianness for the resid */
81 res_id->name[0] = cpu_to_le64(resname);
82 CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
83 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
87 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id)
89 /* fsname is at most 8 chars long, maybe contain "-".
90 * e.g. "lustre", "SUN-000" */
91 return mgc_name2resid(fsname, strlen(fsname), res_id);
93 EXPORT_SYMBOL(mgc_fsname2resid);
95 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id)
100 /* logname consists of "fsname-nodetype".
101 * e.g. "lustre-MDT0001", "SUN-000-client" */
102 name_end = strrchr(logname, '-');
104 len = name_end - logname;
105 return mgc_name2resid(logname, len, res_id);
108 /********************** config llog list **********************/
109 static CFS_LIST_HEAD(config_llog_list);
110 static spinlock_t config_list_lock = 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 atomic_inc(&cld->cld_refcount);
119 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
120 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 atomic_read(&cld->cld_refcount));
132 LASSERT(atomic_read(&cld->cld_refcount) > 0);
134 /* spinlock to make sure no item with 0 refcount in the list */
135 spin_lock(&config_list_lock);
136 if (unlikely(atomic_dec_and_test(&cld->cld_refcount))) {
137 list_del(&cld->cld_list_chain);
138 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 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 spin_lock(&config_list_lock);
180 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 spin_unlock(&config_list_lock);
190 CDEBUG(D_CONFIG, "can't get log %s\n", logid);
191 RETURN(ERR_PTR(-ENOENT));
193 atomic_inc(&cld->cld_refcount);
194 spin_unlock(&config_list_lock);
195 LASSERT(cld->cld_stopping == 0 || cld->cld_is_sptlrpc == 0);
200 struct config_llog_data *do_config_log_add(struct obd_device *obd,
202 unsigned int is_sptlrpc,
203 struct config_llog_instance *cfg,
204 struct super_block *sb)
206 struct config_llog_data *cld;
210 CDEBUG(D_MGC, "do adding config log %s:%s\n", logname,
211 cfg ? cfg->cfg_instance : "NULL");
213 OBD_ALLOC(cld, sizeof(*cld));
215 RETURN(ERR_PTR(-ENOMEM));
216 OBD_ALLOC(cld->cld_logname, strlen(logname) + 1);
217 if (!cld->cld_logname) {
218 OBD_FREE(cld, sizeof(*cld));
219 RETURN(ERR_PTR(-ENOMEM));
221 strcpy(cld->cld_logname, logname);
224 cld->cld_cfg.cfg_last_idx = 0;
225 cld->cld_cfg.cfg_flags = 0;
226 cld->cld_cfg.cfg_sb = sb;
227 cld->cld_is_sptlrpc = is_sptlrpc;
228 atomic_set(&cld->cld_refcount, 1);
230 /* Keep the mgc around until we are done */
231 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
233 if (cfg && cfg->cfg_instance != NULL) {
234 OBD_ALLOC(cld->cld_cfg.cfg_instance,
235 strlen(cfg->cfg_instance) + 1);
236 strcpy(cld->cld_cfg.cfg_instance, cfg->cfg_instance);
240 sptlrpc_conf_log_start(logname);
241 cld->cld_cfg.cfg_obdname = obd->obd_name;
244 rc = mgc_logname2resid(logname, &cld->cld_resid);
246 spin_lock(&config_list_lock);
247 list_add(&cld->cld_list_chain, &config_llog_list);
248 spin_unlock(&config_list_lock);
256 rc = mgc_process_log(obd, cld);
258 CERROR("failed processing sptlrpc log: %d\n", rc);
265 * Add this log to our list of active logs.
266 * We have one active log per "mount" - client instance or servername.
267 * Each instance may be at a different point in the log.
269 static int config_log_add(struct obd_device *obd, char *logname,
270 struct config_llog_instance *cfg,
271 struct super_block *sb)
273 struct config_llog_data *cld, *sptlrpc_cld;
278 CDEBUG(D_MGC, "adding config log %s:%s\n", logname, cfg->cfg_instance);
281 * for each regular log, the depended sptlrpc log name is
282 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
284 ptr = strrchr(logname, '-');
285 if (ptr == NULL || ptr - logname > 8) {
286 CERROR("logname %s is too long\n", logname);
290 memcpy(seclogname, logname, ptr - logname);
291 strcpy(seclogname + (ptr - logname), "-sptlrpc");
293 sptlrpc_cld = config_log_find(seclogname, NULL);
294 if (IS_ERR(sptlrpc_cld)) {
295 sptlrpc_cld = do_config_log_add(obd, seclogname, 1, NULL, NULL);
296 if (IS_ERR(sptlrpc_cld)) {
297 CERROR("can't create sptlrpc log: %s\n", seclogname);
298 RETURN(PTR_ERR(sptlrpc_cld));
302 cld = do_config_log_add(obd, logname, 0, cfg, sb);
304 CERROR("can't create log: %s\n", logname);
305 config_log_put(sptlrpc_cld);
306 RETURN(PTR_ERR(cld));
309 cld->cld_sptlrpc = sptlrpc_cld;
314 DECLARE_MUTEX(llog_process_lock);
316 /* Stop watching for updates on this log. */
317 static int config_log_end(char *logname, struct config_llog_instance *cfg)
319 struct config_llog_data *cld, *cld_sptlrpc;
323 cld = config_log_find(logname, cfg);
325 RETURN(PTR_ERR(cld));
326 /* drop the ref from the find */
329 down(&llog_process_lock);
330 cld->cld_stopping = 1;
331 up(&llog_process_lock);
333 spin_lock(&config_list_lock);
334 cld_sptlrpc = cld->cld_sptlrpc;
335 cld->cld_sptlrpc = NULL;
336 spin_unlock(&config_list_lock);
339 config_log_put(cld_sptlrpc);
341 /* drop the start ref */
343 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
348 /* reenqueue any lost locks */
349 #define RQ_RUNNING 0x1
353 static int rq_state = 0;
354 static cfs_waitq_t rq_waitq;
356 static int mgc_requeue_add(struct config_llog_data *cld, int later);
358 static void do_requeue(struct config_llog_data *cld)
360 LASSERT(atomic_read(&cld->cld_refcount) > 0);
362 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
363 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
364 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
366 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
370 /* Whether we enqueued again or not in mgc_process_log, we're done
371 * with the ref from the old enqueue */
375 static int mgc_requeue_thread(void *data)
377 struct l_wait_info lwi_now, lwi_later;
378 struct config_llog_data *cld, *cld_next, *cld_prev;
379 char name[] = "ll_cfg_requeue";
383 ptlrpc_daemonize(name);
385 CDEBUG(D_MGC, "Starting requeue thread\n");
387 lwi_later = LWI_TIMEOUT(60 * HZ, NULL, NULL);
388 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP), &lwi_later);
390 /* Keep trying failed locks periodically */
391 spin_lock(&config_list_lock);
392 while (rq_state & (RQ_NOW | RQ_LATER)) {
393 /* Any new or requeued lostlocks will change the state */
394 rq_state &= ~(RQ_NOW | RQ_LATER);
395 spin_unlock(&config_list_lock);
397 /* Always wait a few seconds to allow the server who
398 caused the lock revocation to finish its setup, plus some
399 random so everyone doesn't try to reconnect at once. */
400 lwi_now = LWI_TIMEOUT(3 * HZ + (ll_rand() & 0xff) * (HZ / 100),
402 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi_now);
405 * iterate & processing through the list. for each cld, process
406 * its depending sptlrpc cld firstly (if any) and then itself.
408 * it's guaranteed any item in the list must have
409 * reference > 0; and if cld_lostlock is set, at
410 * least one reference is taken by the previous enqueue.
412 * Note: releasing a cld might lead to itself and its depended
413 * sptlrpc cld be unlinked from the list. to safely iterate
414 * we need to take a reference on next cld before processing.
418 spin_lock(&config_list_lock);
419 list_for_each_entry_safe(cld, cld_next, &config_llog_list,
421 if (cld->cld_list_chain.next != &config_llog_list)
422 atomic_inc(&cld_next->cld_refcount);
424 if (cld->cld_lostlock) {
425 if (cld->cld_sptlrpc &&
426 cld->cld_sptlrpc->cld_lostlock) {
427 cld->cld_sptlrpc->cld_lostlock = 0;
429 spin_unlock(&config_list_lock);
430 do_requeue(cld->cld_sptlrpc);
431 spin_lock(&config_list_lock);
432 LASSERT(cld->cld_lostlock);
435 cld->cld_lostlock = 0;
437 spin_unlock(&config_list_lock);
439 spin_lock(&config_list_lock);
444 spin_unlock(&config_list_lock);
445 config_log_put(cld_prev);
446 spin_lock(&config_list_lock);
451 spin_unlock(&config_list_lock);
453 /* Wait a bit to see if anyone else needs a requeue */
454 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
456 spin_lock(&config_list_lock);
458 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
459 rq_state &= ~RQ_RUNNING;
460 spin_unlock(&config_list_lock);
462 CDEBUG(D_MGC, "Ending requeue thread\n");
466 /* Add a cld to the list to requeue. Start the requeue thread if needed.
467 We are responsible for dropping the config log reference from here on out. */
468 static int mgc_requeue_add(struct config_llog_data *cld, int later)
472 CDEBUG(D_INFO, "log %s: requeue (l=%d r=%d sp=%d st=%x)\n",
473 cld->cld_logname, later, atomic_read(&cld->cld_refcount),
474 cld->cld_stopping, rq_state);
475 LASSERT(atomic_read(&cld->cld_refcount) > 0);
477 /* Hold lock for rq_state */
478 spin_lock(&config_list_lock);
480 if (cld->cld_stopping || (rq_state & RQ_STOP)) {
481 cld->cld_lostlock = 0;
482 spin_unlock(&config_list_lock);
487 cld->cld_lostlock = 1;
489 if (!(rq_state & RQ_RUNNING)) {
490 LASSERT(rq_state == 0);
491 rq_state = RQ_RUNNING | (later ? RQ_LATER : RQ_NOW);
492 spin_unlock(&config_list_lock);
493 rc = cfs_kernel_thread(mgc_requeue_thread, 0,
494 CLONE_VM | CLONE_FILES);
496 CERROR("log %s: cannot start requeue thread (%d),"
497 "no more log updates!\n", cld->cld_logname, rc);
498 /* Drop the ref, since the rq thread won't */
499 cld->cld_lostlock = 0;
505 rq_state |= later ? RQ_LATER : RQ_NOW;
506 spin_unlock(&config_list_lock);
507 cfs_waitq_signal(&rq_waitq);
513 /********************** class fns **********************/
515 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
516 struct vfsmount *mnt)
518 struct lvfs_run_ctxt saved;
519 struct lustre_sb_info *lsi = s2lsi(sb);
520 struct client_obd *cli = &obd->u.cli;
521 struct dentry *dentry;
527 LASSERT(lsi->lsi_srv_mnt == mnt);
529 /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
530 down(&cli->cl_mgc_sem);
532 cleanup_group_info();
534 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
535 if (IS_ERR(obd->obd_fsops)) {
536 up(&cli->cl_mgc_sem);
537 CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd),
538 PTR_ERR(obd->obd_fsops));
539 RETURN(PTR_ERR(obd->obd_fsops));
542 cli->cl_mgc_vfsmnt = mnt;
543 fsfilt_setup(obd, mnt->mnt_sb);
545 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
546 obd->obd_lvfs_ctxt.pwdmnt = mnt;
547 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
548 obd->obd_lvfs_ctxt.fs = get_ds();
550 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
551 dentry = lookup_one_len(MOUNT_CONFIGS_DIR, current->fs->pwd,
552 strlen(MOUNT_CONFIGS_DIR));
553 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
554 if (IS_ERR(dentry)) {
555 err = PTR_ERR(dentry);
556 CERROR("cannot lookup %s directory: rc = %d\n",
557 MOUNT_CONFIGS_DIR, err);
560 cli->cl_mgc_configs_dir = dentry;
562 /* We take an obd ref to insure that we can't get to mgc_cleanup
563 without calling mgc_fs_cleanup first. */
564 class_incref(obd, "mgc_fs", obd);
566 label = fsfilt_get_label(obd, mnt->mnt_sb);
568 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
570 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
574 fsfilt_put_ops(obd->obd_fsops);
575 obd->obd_fsops = NULL;
576 cli->cl_mgc_vfsmnt = NULL;
577 up(&cli->cl_mgc_sem);
581 static int mgc_fs_cleanup(struct obd_device *obd)
583 struct client_obd *cli = &obd->u.cli;
587 LASSERT(cli->cl_mgc_vfsmnt != NULL);
589 if (cli->cl_mgc_configs_dir != NULL) {
590 struct lvfs_run_ctxt saved;
591 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
592 l_dput(cli->cl_mgc_configs_dir);
593 cli->cl_mgc_configs_dir = NULL;
594 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
595 class_decref(obd, "mgc_fs", obd);
598 cli->cl_mgc_vfsmnt = NULL;
600 fsfilt_put_ops(obd->obd_fsops);
602 up(&cli->cl_mgc_sem);
607 static atomic_t mgc_count = ATOMIC_INIT(0);
608 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
614 case OBD_CLEANUP_EARLY:
616 case OBD_CLEANUP_EXPORTS:
617 if (atomic_dec_and_test(&mgc_count)) {
618 /* Kick the requeue waitq - cld's should all be
620 spin_lock(&config_list_lock);
622 spin_unlock(&config_list_lock);
623 cfs_waitq_signal(&rq_waitq);
625 rc = obd_llog_finish(obd, 0);
627 CERROR("failed to cleanup llogging subsystems\n");
633 static int mgc_cleanup(struct obd_device *obd)
635 struct client_obd *cli = &obd->u.cli;
639 LASSERT(cli->cl_mgc_vfsmnt == NULL);
641 /* COMPAT_146 - old config logs may have added profiles we don't
643 if (obd->obd_type->typ_refcnt <= 1)
644 /* Only for the last mgc */
645 class_del_profiles();
647 lprocfs_obd_cleanup(obd);
650 rc = client_obd_cleanup(obd);
654 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
656 struct lprocfs_static_vars lvars;
662 rc = client_obd_setup(obd, lcfg);
664 GOTO(err_decref, rc);
666 rc = obd_llog_init(obd, &obd->obd_olg, obd, 0, NULL, NULL);
668 CERROR("failed to setup llogging subsystems\n");
669 GOTO(err_cleanup, rc);
672 lprocfs_mgc_init_vars(&lvars);
673 lprocfs_obd_setup(obd, lvars.obd_vars);
675 spin_lock(&config_list_lock);
676 atomic_inc(&mgc_count);
677 if (atomic_read(&mgc_count) == 1) {
678 rq_state &= ~RQ_STOP;
679 cfs_waitq_init(&rq_waitq);
681 spin_unlock(&config_list_lock);
686 client_obd_cleanup(obd);
692 /* based on ll_mdc_blocking_ast */
693 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
694 void *data, int flag)
696 struct lustre_handle lockh;
697 struct config_llog_data *cld = (struct config_llog_data *)data;
702 case LDLM_CB_BLOCKING:
703 /* mgs wants the lock, give it up... */
704 LDLM_DEBUG(lock, "MGC blocking CB");
705 ldlm_lock2handle(lock, &lockh);
706 rc = ldlm_cli_cancel(&lockh);
708 case LDLM_CB_CANCELING: {
709 /* We've given up the lock, prepare ourselves to update. */
710 LDLM_DEBUG(lock, "MGC cancel CB");
712 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
713 lock->l_resource->lr_name.name[0],
714 (char *)&lock->l_resource->lr_name.name[0]);
717 CERROR("missing data, won't requeue\n");
720 /* Are we done with this log? */
721 if (cld->cld_stopping) {
722 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
727 /* Make sure not to re-enqueue when the mgc is stopping
728 (we get called from client_disconnect_export) */
729 if (!lock->l_conn_export ||
730 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
731 CDEBUG(D_MGC, "log %s: disconnecting, won't requeue\n",
736 /* Did we fail to get the lock? */
737 if (lock->l_req_mode != lock->l_granted_mode &&
738 !cld->cld_is_sptlrpc) {
739 CDEBUG(D_MGC, "log %s: original grant failed, will "
740 "requeue later\n", cld->cld_logname);
741 /* Try to re-enqueue later */
742 rc = mgc_requeue_add(cld, 1);
746 rc = mgc_requeue_add(cld, 0);
755 CERROR("%s CB failed %d:\n", flag == LDLM_CB_BLOCKING ?
756 "blocking" : "cancel", rc);
757 LDLM_ERROR(lock, "MGC ast");
762 /* Send parameter to MGS*/
763 static int mgc_set_mgs_param(struct obd_export *exp,
764 struct mgs_send_param *msp)
766 struct ptlrpc_request *req;
767 struct mgs_send_param *req_msp, *rep_msp;
768 int size[] = { sizeof(struct ptlrpc_body), sizeof(*req_msp) };
769 __u32 rep_size[] = { sizeof(struct ptlrpc_body), sizeof(*msp) };
773 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_MGS_VERSION,
774 MGS_SET_INFO, 2, size, NULL);
778 req_msp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*req_msp));
782 memcpy(req_msp, msp, sizeof(*req_msp));
783 ptlrpc_req_set_repsize(req, 2, rep_size);
784 rc = ptlrpc_queue_wait(req);
786 rep_msp = lustre_swab_repbuf(req, REPLY_REC_OFF,
787 sizeof(*rep_msp), NULL);
788 memcpy(msp, rep_msp, sizeof(*rep_msp));
791 ptlrpc_req_finished(req);
796 /* Take a config lock so we can get cancel notifications */
797 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
798 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
799 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
800 void *data, __u32 lvb_len, void *lvb_swabber,
801 struct lustre_handle *lockh)
803 struct config_llog_data *cld = (struct config_llog_data *)data;
804 struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
805 ldlm_completion_ast, NULL, NULL, data};
810 CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
811 cld->cld_resid.name[0]);
813 /* We can only drop this config log ref when we drop the lock */
814 if (config_log_get(cld))
815 RETURN(ELDLM_LOCK_ABORTED);
817 /* We need a callback for every lockholder, so don't try to
818 ldlm_lock_match (see rev 1.1.2.11.2.47) */
820 rc = ldlm_cli_enqueue(exp, NULL, &einfo, &cld->cld_resid,
821 NULL, flags, NULL, 0, NULL, lockh, 0);
822 /* A failed enqueue should still call the mgc_blocking_ast,
823 where it will be requeued if needed ("grant failed"). */
828 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
829 __u32 mode, struct lustre_handle *lockh)
833 ldlm_lock_decref(lockh, mode);
839 static int mgc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
840 void *karg, void *uarg)
842 struct obd_device *obd = exp->exp_obd;
843 struct obd_ioctl_data *data = karg;
844 struct llog_ctxt *ctxt;
845 struct lvfs_run_ctxt saved;
849 if (!try_module_get(THIS_MODULE)) {
850 CERROR("Can't get module. Is it alive?");
854 /* REPLicator context */
855 case OBD_IOC_PARSE: {
856 CERROR("MGC parsing llog %s\n", data->ioc_inlbuf1);
857 ctxt = llog_get_context(exp->exp_obd, LLOG_CONFIG_REPL_CTXT);
858 rc = class_config_parse_llog(ctxt, data->ioc_inlbuf1, NULL);
862 case OBD_IOC_LLOG_INFO:
863 case OBD_IOC_LLOG_PRINT: {
864 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
865 rc = llog_ioctl(ctxt, cmd, data);
870 /* ORIGinator context */
871 case OBD_IOC_DUMP_LOG: {
872 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
873 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
874 rc = class_config_dump_llog(ctxt, data->ioc_inlbuf1, NULL);
875 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
882 CERROR("mgc_ioctl(): unrecognised ioctl %#x\n", cmd);
883 GOTO(out, rc = -ENOTTY);
886 module_put(THIS_MODULE);
892 /* Send target_reg message to MGS */
893 static int mgc_target_register(struct obd_export *exp,
894 struct mgs_target_info *mti)
896 struct ptlrpc_request *req;
897 struct mgs_target_info *req_mti, *rep_mti;
901 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
902 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
907 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
908 memcpy(req_mti, mti, sizeof(*req_mti));
910 ptlrpc_request_set_replen(req);
912 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
914 rc = ptlrpc_queue_wait(req);
916 rep_mti = req_capsule_server_get(&req->rq_pill,
917 &RMF_MGS_TARGET_INFO);
918 memcpy(mti, rep_mti, sizeof(*rep_mti));
919 CDEBUG(D_MGC, "register %s got index = %d\n",
920 mti->mti_svname, mti->mti_stripe_index);
922 ptlrpc_req_finished(req);
927 int mgc_set_info_async(struct obd_export *exp, obd_count keylen,
928 void *key, obd_count vallen, void *val,
929 struct ptlrpc_request_set *set)
931 struct obd_import *imp = class_exp2cliimp(exp);
935 /* Try to "recover" the initial connection; i.e. retry */
936 if (KEY_IS(KEY_INIT_RECOV)) {
937 if (vallen != sizeof(int))
939 spin_lock(&imp->imp_lock);
940 imp->imp_initial_recov = *(int *)val;
941 spin_unlock(&imp->imp_lock);
942 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
943 exp->exp_obd->obd_name, imp->imp_initial_recov);
946 /* Turn off initial_recov after we try all backup servers once */
947 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
949 if (vallen != sizeof(int))
952 spin_lock(&imp->imp_lock);
953 imp->imp_initial_recov_bk = value > 0;
954 /* Even after the initial connection, give up all comms if
955 nobody answers the first time. */
956 imp->imp_recon_bk = 1;
957 spin_unlock(&imp->imp_lock);
958 CDEBUG(D_MGC, "InitRecov %s %d/%d:d%d:i%d:r%d:or%d:%s\n",
959 imp->imp_obd->obd_name, value, imp->imp_initial_recov,
960 imp->imp_deactive, imp->imp_invalid,
961 imp->imp_replayable, imp->imp_obd->obd_replayable,
962 ptlrpc_import_state_name(imp->imp_state));
963 /* Resurrect if we previously died */
964 if (imp->imp_invalid || value > 1)
965 ptlrpc_reconnect_import(imp);
968 /* FIXME move this to mgc_process_config */
969 if (KEY_IS(KEY_REGISTER_TARGET)) {
970 struct mgs_target_info *mti;
971 if (vallen != sizeof(struct mgs_target_info))
973 mti = (struct mgs_target_info *)val;
974 CDEBUG(D_MGC, "register_target %s %#x\n",
975 mti->mti_svname, mti->mti_flags);
976 rc = mgc_target_register(exp, mti);
979 if (KEY_IS(KEY_SET_FS)) {
980 struct super_block *sb = (struct super_block *)val;
981 struct lustre_sb_info *lsi;
982 if (vallen != sizeof(struct super_block))
985 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
987 CERROR("set_fs got %d\n", rc);
991 if (KEY_IS(KEY_CLEAR_FS)) {
994 rc = mgc_fs_cleanup(exp->exp_obd);
996 CERROR("clear_fs got %d\n", rc);
1000 if (KEY_IS(KEY_SET_INFO)) {
1001 struct mgs_send_param *msp;
1003 msp = (struct mgs_send_param *)val;
1004 rc = mgc_set_mgs_param(exp, msp);
1011 static int mgc_import_event(struct obd_device *obd,
1012 struct obd_import *imp,
1013 enum obd_import_event event)
1017 LASSERT(imp->imp_obd == obd);
1018 CDEBUG(D_MGC, "import event %#x\n", event);
1021 case IMP_EVENT_DISCON:
1022 /* MGC imports should not wait for recovery */
1024 case IMP_EVENT_INACTIVE:
1026 case IMP_EVENT_INVALIDATE: {
1027 struct ldlm_namespace *ns = obd->obd_namespace;
1028 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1031 case IMP_EVENT_ACTIVE:
1032 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
1033 /* Clearing obd_no_recov allows us to continue pinging */
1034 obd->obd_no_recov = 0;
1039 CERROR("Unknown import event %#x\n", event);
1045 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1046 struct obd_device *tgt, int count,
1047 struct llog_catid *logid, struct obd_uuid *uuid)
1049 struct llog_ctxt *ctxt;
1053 LASSERT(olg == &obd->obd_olg);
1055 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
1060 rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
1063 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1064 llog_initiator_connect(ctxt);
1065 llog_ctxt_put(ctxt);
1067 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1075 static int mgc_llog_finish(struct obd_device *obd, int count)
1077 struct llog_ctxt *ctxt;
1078 int rc = 0, rc2 = 0;
1081 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1083 rc = llog_cleanup(ctxt);
1085 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1087 rc2 = llog_cleanup(ctxt);
1095 /* identical to mgs_log_is_empty */
1096 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1099 struct lvfs_run_ctxt saved;
1100 struct llog_handle *llh;
1103 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1104 rc = llog_create(ctxt, &llh, NULL, name);
1106 llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL);
1107 rc = llog_get_size(llh);
1110 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1111 /* header is record 1 */
1115 static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec,
1118 struct llog_rec_hdr local_rec = *rec;
1119 struct llog_handle *local_llh = (struct llog_handle *)data;
1120 char *cfg_buf = (char*) (rec + 1);
1121 struct lustre_cfg *lcfg;
1125 /* Append all records */
1126 local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
1127 rc = llog_write_rec(local_llh, &local_rec, NULL, 0,
1128 (void *)cfg_buf, -1);
1130 lcfg = (struct lustre_cfg *)cfg_buf;
1131 CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
1132 rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
1133 lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
1138 /* Copy a remote log locally */
1139 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1140 struct llog_ctxt *lctxt, char *logname)
1142 struct llog_handle *local_llh, *remote_llh;
1143 struct obd_uuid *uuid;
1148 /* Write new log to a temp name, then vfs_rename over logname
1149 upon successful completion. */
1151 OBD_ALLOC(temp_log, strlen(logname) + 1);
1154 sprintf(temp_log, "%sT", logname);
1156 /* Make sure there's no old temp log */
1157 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1160 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL);
1163 rc = llog_destroy(local_llh);
1164 llog_free_handle(local_llh);
1168 /* open local log */
1169 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1173 /* set the log header uuid for fun */
1174 OBD_ALLOC_PTR(uuid);
1175 obd_str2uuid(uuid, logname);
1176 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid);
1179 GOTO(out_closel, rc);
1181 /* open remote log */
1182 rc = llog_create(rctxt, &remote_llh, NULL, logname);
1184 GOTO(out_closel, rc);
1185 rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL);
1187 GOTO(out_closer, rc);
1189 /* Copy remote log */
1190 rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL);
1193 rc2 = llog_close(remote_llh);
1197 rc2 = llog_close(local_llh);
1201 /* We've copied the remote log to the local temp log, now
1202 replace the old local log with the temp log. */
1204 struct client_obd *cli = &obd->u.cli;
1206 LASSERT(cli->cl_mgc_configs_dir);
1207 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1210 CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1213 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1214 OBD_FREE(temp_log, strlen(logname) + 1);
1218 /* Get a config log from the MGS and process it.
1219 This func is called for both clients and servers. */
1220 int mgc_process_log(struct obd_device *mgc,
1221 struct config_llog_data *cld)
1223 struct llog_ctxt *ctxt, *lctxt;
1224 struct lustre_handle lockh;
1225 struct client_obd *cli = &mgc->u.cli;
1226 struct lvfs_run_ctxt saved;
1227 struct lustre_sb_info *lsi = NULL;
1228 int rc = 0, rcl, flags = 0, must_pop = 0;
1233 /* I don't want mutliple processes running process_log at once --
1234 sounds like badness. It actually might be fine, as long as
1235 we're not trying to update from the same log
1236 simultaneously (in which case we should use a per-log sem.) */
1237 down(&llog_process_lock);
1239 if (cld->cld_stopping) {
1240 up(&llog_process_lock);
1244 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1246 if (cld->cld_cfg.cfg_sb)
1247 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1249 CDEBUG(D_MGC, "Process log %s:%s from %d\n", cld->cld_logname,
1250 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1252 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1254 CERROR("missing llog context\n");
1255 up(&llog_process_lock);
1259 /* Get the cfg lock on the llog */
1260 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1261 LCK_CR, &flags, NULL, NULL, NULL,
1262 cld, 0, NULL, &lockh);
1264 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1266 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1269 * local copy of sptlrpc log is controlled elsewhere, don't try to
1272 if (rcl && cld->cld_is_sptlrpc)
1275 /* Copy the setup log locally if we can. Don't mess around if we're
1276 running an MGS though (logs are already local). */
1277 if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
1278 (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1279 !IS_MGS(lsi->lsi_ldd)) {
1280 push_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1283 /* Only try to copy log if we have the lock. */
1284 rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1286 if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1287 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1288 "log %s and no local copy."
1289 "\n", cld->cld_logname);
1290 GOTO(out_pop, rc = -ENOTCONN);
1292 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1293 "copy for now, will try to update later.\n",
1296 /* Now, whether we copied or not, start using the local llog.
1297 If we failed to copy, we'll start using whatever the old
1299 llog_ctxt_put(ctxt);
1303 if (cld->cld_is_sptlrpc)
1304 sptlrpc_conf_log_update_begin(cld->cld_logname);
1306 /* logname and instance info should be the same, so use our
1307 copy of the instance for the update. The cfg_last_idx will
1309 rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg);
1311 llog_ctxt_put(ctxt);
1313 llog_ctxt_put(lctxt);
1315 pop_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1318 * update settings on existing OBDs. doing it inside
1319 * of llog_process_lock so no device is attaching/detaching
1321 * the logname must be <fsname>-sptlrpc
1323 if (cld->cld_is_sptlrpc && rcl == 0) {
1324 sptlrpc_conf_log_update_end(cld->cld_logname);
1325 class_notify_sptlrpc_conf(cld->cld_logname,
1326 strlen(cld->cld_logname) -
1327 strlen("-sptlrpc"));
1330 /* Now drop the lock so MGS can revoke it */
1332 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1335 CERROR("Can't drop cfg lock: %d\n", rcl);
1338 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1339 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1341 up(&llog_process_lock);
1346 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1348 struct lustre_cfg *lcfg = buf;
1353 switch(cmd = lcfg->lcfg_command) {
1354 case LCFG_LOV_ADD_OBD: {
1355 /* Add any new target, not just osts */
1356 struct mgs_target_info *mti;
1358 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1359 sizeof(struct mgs_target_info))
1360 GOTO(out, rc = -EINVAL);
1362 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1363 CDEBUG(D_MGC, "add_target %s %#x\n",
1364 mti->mti_svname, mti->mti_flags);
1365 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1368 case LCFG_LOV_DEL_OBD:
1369 /* Remove target from the fs? */
1371 CERROR("lov_del_obd unimplemented\n");
1374 case LCFG_SPTLRPC_CONF: {
1375 rc = sptlrpc_process_config(lcfg);
1378 case LCFG_LOG_START: {
1379 struct config_llog_data *cld;
1380 struct config_llog_instance *cfg;
1381 struct super_block *sb;
1382 char *logname = lustre_cfg_string(lcfg, 1);
1383 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1384 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1386 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1389 /* We're only called through here on the initial mount */
1390 rc = config_log_add(obd, logname, cfg, sb);
1393 cld = config_log_find(logname, cfg);
1400 /* FIXME only set this for old logs! Right now this forces
1401 us to always skip the "inside markers" check */
1402 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1404 rc = mgc_process_log(obd, cld);
1405 config_log_put(cld);
1409 case LCFG_LOG_END: {
1410 struct config_llog_instance *cfg = NULL;
1411 char *logname = lustre_cfg_string(lcfg, 1);
1412 if (lcfg->lcfg_bufcount >= 2)
1413 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1415 rc = config_log_end(logname, cfg);
1419 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1420 GOTO(out, rc = -EINVAL);
1428 struct obd_ops mgc_obd_ops = {
1429 .o_owner = THIS_MODULE,
1430 .o_setup = mgc_setup,
1431 .o_precleanup = mgc_precleanup,
1432 .o_cleanup = mgc_cleanup,
1433 .o_add_conn = client_import_add_conn,
1434 .o_del_conn = client_import_del_conn,
1435 .o_connect = client_connect_import,
1436 .o_disconnect = client_disconnect_export,
1437 //.o_enqueue = mgc_enqueue,
1438 .o_cancel = mgc_cancel,
1439 //.o_iocontrol = mgc_iocontrol,
1440 .o_set_info_async = mgc_set_info_async,
1441 .o_import_event = mgc_import_event,
1442 .o_llog_init = mgc_llog_init,
1443 .o_llog_finish = mgc_llog_finish,
1444 .o_process_config = mgc_process_config,
1447 int __init mgc_init(void)
1449 return class_register_type(&mgc_obd_ops, NULL, NULL,
1450 LUSTRE_MGC_NAME, NULL);
1454 static void /*__exit*/ mgc_exit(void)
1456 class_unregister_type(LUSTRE_MGC_NAME);
1459 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1460 MODULE_DESCRIPTION("Lustre Management Client");
1461 MODULE_LICENSE("GPL");
1463 module_init(mgc_init);
1464 module_exit(mgc_exit);