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 = NULL;
323 cld = config_log_find(logname, cfg);
325 RETURN(PTR_ERR(cld));
327 down(&llog_process_lock);
329 * if cld_stopping is set, it means we didn't start the log thus
330 * not owning the start ref. this can happen after previous umount:
331 * the cld still hanging there waiting for lock cancel, and we
332 * remount again but failed in the middle and call log_end without
335 if (unlikely(cld->cld_stopping)) {
336 up(&llog_process_lock);
337 /* drop the ref from the find */
342 cld->cld_stopping = 1;
343 up(&llog_process_lock);
345 spin_lock(&config_list_lock);
346 cld_sptlrpc = cld->cld_sptlrpc;
347 cld->cld_sptlrpc = NULL;
348 spin_unlock(&config_list_lock);
351 config_log_put(cld_sptlrpc);
353 /* drop the ref from the find */
355 /* drop the start ref */
358 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
363 /* reenqueue any lost locks */
364 #define RQ_RUNNING 0x1
368 static int rq_state = 0;
369 static cfs_waitq_t rq_waitq;
371 static int mgc_requeue_add(struct config_llog_data *cld, int later);
373 static void do_requeue(struct config_llog_data *cld)
375 LASSERT(atomic_read(&cld->cld_refcount) > 0);
377 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
378 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
379 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
381 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
385 /* Whether we enqueued again or not in mgc_process_log, we're done
386 * with the ref from the old enqueue */
390 static int mgc_requeue_thread(void *data)
392 struct l_wait_info lwi_now, lwi_later;
393 struct config_llog_data *cld, *cld_next, *cld_prev;
394 char name[] = "ll_cfg_requeue";
398 ptlrpc_daemonize(name);
400 CDEBUG(D_MGC, "Starting requeue thread\n");
402 lwi_later = LWI_TIMEOUT(60 * HZ, NULL, NULL);
403 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP), &lwi_later);
405 /* Keep trying failed locks periodically */
406 spin_lock(&config_list_lock);
407 while (rq_state & (RQ_NOW | RQ_LATER)) {
408 /* Any new or requeued lostlocks will change the state */
409 rq_state &= ~(RQ_NOW | RQ_LATER);
410 spin_unlock(&config_list_lock);
412 /* Always wait a few seconds to allow the server who
413 caused the lock revocation to finish its setup, plus some
414 random so everyone doesn't try to reconnect at once. */
415 lwi_now = LWI_TIMEOUT(3 * HZ + (ll_rand() & 0xff) * (HZ / 100),
417 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi_now);
420 * iterate & processing through the list. for each cld, process
421 * its depending sptlrpc cld firstly (if any) and then itself.
423 * it's guaranteed any item in the list must have
424 * reference > 0; and if cld_lostlock is set, at
425 * least one reference is taken by the previous enqueue.
427 * Note: releasing a cld might lead to itself and its depended
428 * sptlrpc cld be unlinked from the list. to safely iterate
429 * we need to take a reference on next cld before processing.
433 spin_lock(&config_list_lock);
434 list_for_each_entry_safe(cld, cld_next, &config_llog_list,
436 if (cld->cld_list_chain.next != &config_llog_list)
437 atomic_inc(&cld_next->cld_refcount);
439 if (cld->cld_lostlock) {
440 if (cld->cld_sptlrpc &&
441 cld->cld_sptlrpc->cld_lostlock) {
442 cld->cld_sptlrpc->cld_lostlock = 0;
444 spin_unlock(&config_list_lock);
445 do_requeue(cld->cld_sptlrpc);
446 spin_lock(&config_list_lock);
447 LASSERT(cld->cld_lostlock);
450 cld->cld_lostlock = 0;
452 spin_unlock(&config_list_lock);
454 spin_lock(&config_list_lock);
459 spin_unlock(&config_list_lock);
460 config_log_put(cld_prev);
461 spin_lock(&config_list_lock);
466 spin_unlock(&config_list_lock);
468 /* Wait a bit to see if anyone else needs a requeue */
469 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
471 spin_lock(&config_list_lock);
473 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
474 rq_state &= ~RQ_RUNNING;
475 spin_unlock(&config_list_lock);
477 CDEBUG(D_MGC, "Ending requeue thread\n");
481 /* Add a cld to the list to requeue. Start the requeue thread if needed.
482 We are responsible for dropping the config log reference from here on out. */
483 static int mgc_requeue_add(struct config_llog_data *cld, int later)
487 CDEBUG(D_INFO, "log %s: requeue (l=%d r=%d sp=%d st=%x)\n",
488 cld->cld_logname, later, atomic_read(&cld->cld_refcount),
489 cld->cld_stopping, rq_state);
490 LASSERT(atomic_read(&cld->cld_refcount) > 0);
492 /* Hold lock for rq_state */
493 spin_lock(&config_list_lock);
495 if (cld->cld_stopping || (rq_state & RQ_STOP)) {
496 cld->cld_lostlock = 0;
497 spin_unlock(&config_list_lock);
502 cld->cld_lostlock = 1;
504 if (!(rq_state & RQ_RUNNING)) {
505 LASSERT(rq_state == 0);
506 rq_state = RQ_RUNNING | (later ? RQ_LATER : RQ_NOW);
507 spin_unlock(&config_list_lock);
508 rc = cfs_kernel_thread(mgc_requeue_thread, 0,
509 CLONE_VM | CLONE_FILES);
511 CERROR("log %s: cannot start requeue thread (%d),"
512 "no more log updates!\n", cld->cld_logname, rc);
513 /* Drop the ref, since the rq thread won't */
514 cld->cld_lostlock = 0;
520 rq_state |= later ? RQ_LATER : RQ_NOW;
521 spin_unlock(&config_list_lock);
522 cfs_waitq_signal(&rq_waitq);
528 /********************** class fns **********************/
530 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
531 struct vfsmount *mnt)
533 struct lvfs_run_ctxt saved;
534 struct lustre_sb_info *lsi = s2lsi(sb);
535 struct client_obd *cli = &obd->u.cli;
536 struct dentry *dentry;
542 LASSERT(lsi->lsi_srv_mnt == mnt);
544 /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
545 down(&cli->cl_mgc_sem);
547 cleanup_group_info();
549 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
550 if (IS_ERR(obd->obd_fsops)) {
551 up(&cli->cl_mgc_sem);
552 CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd),
553 PTR_ERR(obd->obd_fsops));
554 RETURN(PTR_ERR(obd->obd_fsops));
557 cli->cl_mgc_vfsmnt = mnt;
558 fsfilt_setup(obd, mnt->mnt_sb);
560 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
561 obd->obd_lvfs_ctxt.pwdmnt = mnt;
562 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
563 obd->obd_lvfs_ctxt.fs = get_ds();
565 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
566 dentry = lookup_one_len(MOUNT_CONFIGS_DIR, current->fs->pwd,
567 strlen(MOUNT_CONFIGS_DIR));
568 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
569 if (IS_ERR(dentry)) {
570 err = PTR_ERR(dentry);
571 CERROR("cannot lookup %s directory: rc = %d\n",
572 MOUNT_CONFIGS_DIR, err);
575 cli->cl_mgc_configs_dir = dentry;
577 /* We take an obd ref to insure that we can't get to mgc_cleanup
578 without calling mgc_fs_cleanup first. */
579 class_incref(obd, "mgc_fs", obd);
581 label = fsfilt_get_label(obd, mnt->mnt_sb);
583 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
585 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
589 fsfilt_put_ops(obd->obd_fsops);
590 obd->obd_fsops = NULL;
591 cli->cl_mgc_vfsmnt = NULL;
592 up(&cli->cl_mgc_sem);
596 static int mgc_fs_cleanup(struct obd_device *obd)
598 struct client_obd *cli = &obd->u.cli;
602 LASSERT(cli->cl_mgc_vfsmnt != NULL);
604 if (cli->cl_mgc_configs_dir != NULL) {
605 struct lvfs_run_ctxt saved;
606 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
607 l_dput(cli->cl_mgc_configs_dir);
608 cli->cl_mgc_configs_dir = NULL;
609 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
610 class_decref(obd, "mgc_fs", obd);
613 cli->cl_mgc_vfsmnt = NULL;
615 fsfilt_put_ops(obd->obd_fsops);
617 up(&cli->cl_mgc_sem);
622 static atomic_t mgc_count = ATOMIC_INIT(0);
623 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
629 case OBD_CLEANUP_EARLY:
631 case OBD_CLEANUP_EXPORTS:
632 if (atomic_dec_and_test(&mgc_count)) {
633 /* Kick the requeue waitq - cld's should all be
635 spin_lock(&config_list_lock);
637 spin_unlock(&config_list_lock);
638 cfs_waitq_signal(&rq_waitq);
640 rc = obd_llog_finish(obd, 0);
642 CERROR("failed to cleanup llogging subsystems\n");
648 static int mgc_cleanup(struct obd_device *obd)
650 struct client_obd *cli = &obd->u.cli;
654 LASSERT(cli->cl_mgc_vfsmnt == NULL);
656 /* COMPAT_146 - old config logs may have added profiles we don't
658 if (obd->obd_type->typ_refcnt <= 1)
659 /* Only for the last mgc */
660 class_del_profiles();
662 lprocfs_obd_cleanup(obd);
665 rc = client_obd_cleanup(obd);
669 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
671 struct lprocfs_static_vars lvars;
677 rc = client_obd_setup(obd, lcfg);
679 GOTO(err_decref, rc);
681 rc = obd_llog_init(obd, &obd->obd_olg, obd, 0, NULL, NULL);
683 CERROR("failed to setup llogging subsystems\n");
684 GOTO(err_cleanup, rc);
687 lprocfs_mgc_init_vars(&lvars);
688 lprocfs_obd_setup(obd, lvars.obd_vars);
689 sptlrpc_lprocfs_cliobd_attach(obd);
691 spin_lock(&config_list_lock);
692 atomic_inc(&mgc_count);
693 if (atomic_read(&mgc_count) == 1) {
694 rq_state &= ~RQ_STOP;
695 cfs_waitq_init(&rq_waitq);
697 spin_unlock(&config_list_lock);
702 client_obd_cleanup(obd);
708 /* based on ll_mdc_blocking_ast */
709 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
710 void *data, int flag)
712 struct lustre_handle lockh;
713 struct config_llog_data *cld = (struct config_llog_data *)data;
718 case LDLM_CB_BLOCKING:
719 /* mgs wants the lock, give it up... */
720 LDLM_DEBUG(lock, "MGC blocking CB");
721 ldlm_lock2handle(lock, &lockh);
722 rc = ldlm_cli_cancel(&lockh);
724 case LDLM_CB_CANCELING: {
725 /* We've given up the lock, prepare ourselves to update. */
726 LDLM_DEBUG(lock, "MGC cancel CB");
728 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
729 lock->l_resource->lr_name.name[0],
730 (char *)&lock->l_resource->lr_name.name[0]);
733 CERROR("missing data, won't requeue\n");
736 /* Are we done with this log? */
737 if (cld->cld_stopping) {
738 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
743 /* Make sure not to re-enqueue when the mgc is stopping
744 (we get called from client_disconnect_export) */
745 if (!lock->l_conn_export ||
746 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
747 CDEBUG(D_MGC, "log %s: disconnecting, won't requeue\n",
752 /* Did we fail to get the lock? */
753 if (lock->l_req_mode != lock->l_granted_mode &&
754 !cld->cld_is_sptlrpc) {
755 CDEBUG(D_MGC, "log %s: original grant failed, will "
756 "requeue later\n", cld->cld_logname);
757 /* Try to re-enqueue later */
758 rc = mgc_requeue_add(cld, 1);
762 rc = mgc_requeue_add(cld, 0);
771 CERROR("%s CB failed %d:\n", flag == LDLM_CB_BLOCKING ?
772 "blocking" : "cancel", rc);
773 LDLM_ERROR(lock, "MGC ast");
778 /* Send parameter to MGS*/
779 static int mgc_set_mgs_param(struct obd_export *exp,
780 struct mgs_send_param *msp)
782 struct ptlrpc_request *req;
783 struct mgs_send_param *req_msp, *rep_msp;
784 int size[] = { sizeof(struct ptlrpc_body), sizeof(*req_msp) };
785 __u32 rep_size[] = { sizeof(struct ptlrpc_body), sizeof(*msp) };
789 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_MGS_VERSION,
790 MGS_SET_INFO, 2, size, NULL);
794 req_msp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*req_msp));
796 ptlrpc_req_finished(req);
800 memcpy(req_msp, msp, sizeof(*req_msp));
801 ptlrpc_req_set_repsize(req, 2, rep_size);
803 rc = ptlrpc_queue_wait(req);
805 rep_msp = lustre_swab_repbuf(req, REPLY_REC_OFF,
806 sizeof(*rep_msp), NULL);
807 memcpy(msp, rep_msp, sizeof(*rep_msp));
810 ptlrpc_req_finished(req);
815 /* Take a config lock so we can get cancel notifications */
816 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
817 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
818 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
819 void *data, __u32 lvb_len, void *lvb_swabber,
820 struct lustre_handle *lockh)
822 struct config_llog_data *cld = (struct config_llog_data *)data;
823 struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
824 ldlm_completion_ast, NULL, NULL, data};
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) */
839 rc = ldlm_cli_enqueue(exp, NULL, &einfo, &cld->cld_resid,
840 NULL, flags, NULL, 0, NULL, lockh, 0);
841 /* A failed enqueue should still call the mgc_blocking_ast,
842 where it will be requeued if needed ("grant failed"). */
847 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
848 __u32 mode, struct lustre_handle *lockh)
852 ldlm_lock_decref(lockh, mode);
858 static int mgc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
859 void *karg, void *uarg)
861 struct obd_device *obd = exp->exp_obd;
862 struct obd_ioctl_data *data = karg;
863 struct llog_ctxt *ctxt;
864 struct lvfs_run_ctxt saved;
868 if (!try_module_get(THIS_MODULE)) {
869 CERROR("Can't get module. Is it alive?");
873 /* REPLicator context */
874 case OBD_IOC_PARSE: {
875 CERROR("MGC parsing llog %s\n", data->ioc_inlbuf1);
876 ctxt = llog_get_context(exp->exp_obd, LLOG_CONFIG_REPL_CTXT);
877 rc = class_config_parse_llog(ctxt, data->ioc_inlbuf1, NULL);
881 case OBD_IOC_LLOG_INFO:
882 case OBD_IOC_LLOG_PRINT: {
883 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
884 rc = llog_ioctl(ctxt, cmd, data);
889 /* ORIGinator context */
890 case OBD_IOC_DUMP_LOG: {
891 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
892 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
893 rc = class_config_dump_llog(ctxt, data->ioc_inlbuf1, NULL);
894 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
901 CERROR("mgc_ioctl(): unrecognised ioctl %#x\n", cmd);
902 GOTO(out, rc = -ENOTTY);
905 module_put(THIS_MODULE);
911 /* Send target_reg message to MGS */
912 static int mgc_target_register(struct obd_export *exp,
913 struct mgs_target_info *mti)
915 struct ptlrpc_request *req;
916 struct mgs_target_info *req_mti, *rep_mti;
920 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
921 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
926 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
928 ptlrpc_req_finished(req);
932 memcpy(req_mti, mti, sizeof(*req_mti));
933 ptlrpc_request_set_replen(req);
934 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
936 rc = ptlrpc_queue_wait(req);
938 rep_mti = req_capsule_server_get(&req->rq_pill,
939 &RMF_MGS_TARGET_INFO);
940 memcpy(mti, rep_mti, sizeof(*rep_mti));
941 CDEBUG(D_MGC, "register %s got index = %d\n",
942 mti->mti_svname, mti->mti_stripe_index);
944 ptlrpc_req_finished(req);
949 int mgc_set_info_async(struct obd_export *exp, obd_count keylen,
950 void *key, obd_count vallen, void *val,
951 struct ptlrpc_request_set *set)
953 struct obd_import *imp = class_exp2cliimp(exp);
957 /* Try to "recover" the initial connection; i.e. retry */
958 if (KEY_IS(KEY_INIT_RECOV)) {
959 if (vallen != sizeof(int))
961 spin_lock(&imp->imp_lock);
962 imp->imp_initial_recov = *(int *)val;
963 spin_unlock(&imp->imp_lock);
964 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
965 exp->exp_obd->obd_name, imp->imp_initial_recov);
968 /* Turn off initial_recov after we try all backup servers once */
969 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
971 if (vallen != sizeof(int))
974 spin_lock(&imp->imp_lock);
975 imp->imp_initial_recov_bk = value > 0;
976 /* Even after the initial connection, give up all comms if
977 nobody answers the first time. */
978 imp->imp_recon_bk = 1;
979 spin_unlock(&imp->imp_lock);
980 CDEBUG(D_MGC, "InitRecov %s %d/%d:d%d:i%d:r%d:or%d:%s\n",
981 imp->imp_obd->obd_name, value, imp->imp_initial_recov,
982 imp->imp_deactive, imp->imp_invalid,
983 imp->imp_replayable, imp->imp_obd->obd_replayable,
984 ptlrpc_import_state_name(imp->imp_state));
985 /* Resurrect if we previously died */
986 if (imp->imp_invalid || value > 1)
987 ptlrpc_reconnect_import(imp);
990 /* FIXME move this to mgc_process_config */
991 if (KEY_IS(KEY_REGISTER_TARGET)) {
992 struct mgs_target_info *mti;
993 if (vallen != sizeof(struct mgs_target_info))
995 mti = (struct mgs_target_info *)val;
996 CDEBUG(D_MGC, "register_target %s %#x\n",
997 mti->mti_svname, mti->mti_flags);
998 rc = mgc_target_register(exp, mti);
1001 if (KEY_IS(KEY_SET_FS)) {
1002 struct super_block *sb = (struct super_block *)val;
1003 struct lustre_sb_info *lsi;
1004 if (vallen != sizeof(struct super_block))
1007 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
1009 CERROR("set_fs got %d\n", rc);
1013 if (KEY_IS(KEY_CLEAR_FS)) {
1016 rc = mgc_fs_cleanup(exp->exp_obd);
1018 CERROR("clear_fs got %d\n", rc);
1022 if (KEY_IS(KEY_SET_INFO)) {
1023 struct mgs_send_param *msp;
1025 msp = (struct mgs_send_param *)val;
1026 rc = mgc_set_mgs_param(exp, msp);
1029 if (KEY_IS(KEY_MGSSEC)) {
1030 struct client_obd *cli = &exp->exp_obd->u.cli;
1031 struct sptlrpc_flavor flvr;
1034 * empty string means using current flavor, if which haven't
1035 * been set yet, set it as null.
1037 * if flavor has been set previously, check the asking flavor
1038 * must match the existing one.
1041 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1047 rc = sptlrpc_parse_flavor(val, &flvr);
1049 CERROR("invalid sptlrpc flavor %s to MGS\n",
1055 * caller already hold a mutex
1057 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1058 cli->cl_flvr_mgc = flvr;
1059 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1060 sizeof(flvr)) != 0) {
1063 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1065 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1066 "currently %s is in use\n",
1076 static int mgc_import_event(struct obd_device *obd,
1077 struct obd_import *imp,
1078 enum obd_import_event event)
1082 LASSERT(imp->imp_obd == obd);
1083 CDEBUG(D_MGC, "import event %#x\n", event);
1086 case IMP_EVENT_DISCON:
1087 /* MGC imports should not wait for recovery */
1089 case IMP_EVENT_INACTIVE:
1091 case IMP_EVENT_INVALIDATE: {
1092 struct ldlm_namespace *ns = obd->obd_namespace;
1093 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1096 case IMP_EVENT_ACTIVE:
1097 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
1098 /* Clearing obd_no_recov allows us to continue pinging */
1099 obd->obd_no_recov = 0;
1104 CERROR("Unknown import event %#x\n", event);
1110 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1111 struct obd_device *tgt, int count,
1112 struct llog_catid *logid, struct obd_uuid *uuid)
1114 struct llog_ctxt *ctxt;
1118 LASSERT(olg == &obd->obd_olg);
1120 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
1125 rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
1128 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1130 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1135 llog_initiator_connect(ctxt);
1136 llog_ctxt_put(ctxt);
1138 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1146 static int mgc_llog_finish(struct obd_device *obd, int count)
1148 struct llog_ctxt *ctxt;
1149 int rc = 0, rc2 = 0;
1152 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1154 rc = llog_cleanup(ctxt);
1156 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1158 rc2 = llog_cleanup(ctxt);
1166 /* identical to mgs_log_is_empty */
1167 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1170 struct lvfs_run_ctxt saved;
1171 struct llog_handle *llh;
1174 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1175 rc = llog_create(ctxt, &llh, NULL, name);
1177 llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL);
1178 rc = llog_get_size(llh);
1181 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1182 /* header is record 1 */
1186 static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec,
1189 struct llog_rec_hdr local_rec = *rec;
1190 struct llog_handle *local_llh = (struct llog_handle *)data;
1191 char *cfg_buf = (char*) (rec + 1);
1192 struct lustre_cfg *lcfg;
1196 /* Append all records */
1197 local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
1198 rc = llog_write_rec(local_llh, &local_rec, NULL, 0,
1199 (void *)cfg_buf, -1);
1201 lcfg = (struct lustre_cfg *)cfg_buf;
1202 CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
1203 rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
1204 lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
1209 /* Copy a remote log locally */
1210 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1211 struct llog_ctxt *lctxt, char *logname)
1213 struct llog_handle *local_llh, *remote_llh;
1214 struct obd_uuid *uuid;
1219 /* Write new log to a temp name, then vfs_rename over logname
1220 upon successful completion. */
1222 OBD_ALLOC(temp_log, strlen(logname) + 1);
1225 sprintf(temp_log, "%sT", logname);
1227 /* Make sure there's no old temp log */
1228 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1231 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL);
1234 rc = llog_destroy(local_llh);
1235 llog_free_handle(local_llh);
1239 /* open local log */
1240 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1244 /* set the log header uuid for fun */
1245 OBD_ALLOC_PTR(uuid);
1246 obd_str2uuid(uuid, logname);
1247 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid);
1250 GOTO(out_closel, rc);
1252 /* open remote log */
1253 rc = llog_create(rctxt, &remote_llh, NULL, logname);
1255 GOTO(out_closel, rc);
1256 rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL);
1258 GOTO(out_closer, rc);
1260 /* Copy remote log */
1261 rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL);
1264 rc2 = llog_close(remote_llh);
1268 rc2 = llog_close(local_llh);
1272 /* We've copied the remote log to the local temp log, now
1273 replace the old local log with the temp log. */
1275 struct client_obd *cli = &obd->u.cli;
1277 LASSERT(cli->cl_mgc_configs_dir);
1278 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1281 CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1284 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1285 OBD_FREE(temp_log, strlen(logname) + 1);
1289 /* Get a config log from the MGS and process it.
1290 This func is called for both clients and servers. */
1291 int mgc_process_log(struct obd_device *mgc,
1292 struct config_llog_data *cld)
1294 struct llog_ctxt *ctxt, *lctxt;
1295 struct lustre_handle lockh;
1296 struct client_obd *cli = &mgc->u.cli;
1297 struct lvfs_run_ctxt saved;
1298 struct lustre_sb_info *lsi = NULL;
1299 int rc = 0, rcl, flags = 0, must_pop = 0;
1304 /* I don't want mutliple processes running process_log at once --
1305 sounds like badness. It actually might be fine, as long as
1306 we're not trying to update from the same log
1307 simultaneously (in which case we should use a per-log sem.) */
1308 down(&llog_process_lock);
1310 if (cld->cld_stopping) {
1311 up(&llog_process_lock);
1315 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1317 if (cld->cld_cfg.cfg_sb)
1318 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1320 CDEBUG(D_MGC, "Process log %s:%s from %d\n", cld->cld_logname,
1321 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1323 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1325 CERROR("missing llog context\n");
1326 up(&llog_process_lock);
1330 /* Get the cfg lock on the llog */
1331 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1332 LCK_CR, &flags, NULL, NULL, NULL,
1333 cld, 0, NULL, &lockh);
1335 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1337 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1340 * local copy of sptlrpc log is controlled elsewhere, don't try to
1343 if (rcl && cld->cld_is_sptlrpc)
1346 /* Copy the setup log locally if we can. Don't mess around if we're
1347 running an MGS though (logs are already local). */
1348 if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
1349 (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1350 !IS_MGS(lsi->lsi_ldd)) {
1351 push_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1354 /* Only try to copy log if we have the lock. */
1355 rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1357 if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1358 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1359 "log %s and no local copy."
1360 "\n", cld->cld_logname);
1361 GOTO(out_pop, rc = -ENOTCONN);
1363 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1364 "copy for now, will try to update later.\n",
1367 /* Now, whether we copied or not, start using the local llog.
1368 If we failed to copy, we'll start using whatever the old
1370 llog_ctxt_put(ctxt);
1374 if (cld->cld_is_sptlrpc)
1375 sptlrpc_conf_log_update_begin(cld->cld_logname);
1377 /* logname and instance info should be the same, so use our
1378 copy of the instance for the update. The cfg_last_idx will
1380 rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg);
1382 llog_ctxt_put(ctxt);
1384 llog_ctxt_put(lctxt);
1386 pop_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1389 * update settings on existing OBDs. doing it inside
1390 * of llog_process_lock so no device is attaching/detaching
1392 * the logname must be <fsname>-sptlrpc
1394 if (cld->cld_is_sptlrpc && rcl == 0) {
1395 sptlrpc_conf_log_update_end(cld->cld_logname);
1396 class_notify_sptlrpc_conf(cld->cld_logname,
1397 strlen(cld->cld_logname) -
1398 strlen("-sptlrpc"));
1401 /* Now drop the lock so MGS can revoke it */
1403 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1406 CERROR("Can't drop cfg lock: %d\n", rcl);
1409 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1410 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1412 up(&llog_process_lock);
1417 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1419 struct lustre_cfg *lcfg = buf;
1424 switch(cmd = lcfg->lcfg_command) {
1425 case LCFG_LOV_ADD_OBD: {
1426 /* Add any new target, not just osts */
1427 struct mgs_target_info *mti;
1429 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1430 sizeof(struct mgs_target_info))
1431 GOTO(out, rc = -EINVAL);
1433 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1434 CDEBUG(D_MGC, "add_target %s %#x\n",
1435 mti->mti_svname, mti->mti_flags);
1436 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1439 case LCFG_LOV_DEL_OBD:
1440 /* Remove target from the fs? */
1442 CERROR("lov_del_obd unimplemented\n");
1445 case LCFG_SPTLRPC_CONF: {
1446 rc = sptlrpc_process_config(lcfg);
1449 case LCFG_LOG_START: {
1450 struct config_llog_data *cld;
1451 struct config_llog_instance *cfg;
1452 struct super_block *sb;
1453 char *logname = lustre_cfg_string(lcfg, 1);
1454 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1455 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1457 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1460 /* We're only called through here on the initial mount */
1461 rc = config_log_add(obd, logname, cfg, sb);
1464 cld = config_log_find(logname, cfg);
1471 /* FIXME only set this for old logs! Right now this forces
1472 us to always skip the "inside markers" check */
1473 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1475 rc = mgc_process_log(obd, cld);
1476 config_log_put(cld);
1480 case LCFG_LOG_END: {
1481 struct config_llog_instance *cfg = NULL;
1482 char *logname = lustre_cfg_string(lcfg, 1);
1483 if (lcfg->lcfg_bufcount >= 2)
1484 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1486 rc = config_log_end(logname, cfg);
1490 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1491 GOTO(out, rc = -EINVAL);
1499 struct obd_ops mgc_obd_ops = {
1500 .o_owner = THIS_MODULE,
1501 .o_setup = mgc_setup,
1502 .o_precleanup = mgc_precleanup,
1503 .o_cleanup = mgc_cleanup,
1504 .o_add_conn = client_import_add_conn,
1505 .o_del_conn = client_import_del_conn,
1506 .o_connect = client_connect_import,
1507 .o_disconnect = client_disconnect_export,
1508 //.o_enqueue = mgc_enqueue,
1509 .o_cancel = mgc_cancel,
1510 //.o_iocontrol = mgc_iocontrol,
1511 .o_set_info_async = mgc_set_info_async,
1512 .o_import_event = mgc_import_event,
1513 .o_llog_init = mgc_llog_init,
1514 .o_llog_finish = mgc_llog_finish,
1515 .o_process_config = mgc_process_config,
1518 int __init mgc_init(void)
1520 return class_register_type(&mgc_obd_ops, NULL, NULL,
1521 LUSTRE_MGC_NAME, NULL);
1525 static void /*__exit*/ mgc_exit(void)
1527 class_unregister_type(LUSTRE_MGC_NAME);
1530 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1531 MODULE_DESCRIPTION("Lustre Management Client");
1532 MODULE_LICENSE("GPL");
1534 module_init(mgc_init);
1535 module_exit(mgc_exit);