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)
129 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
130 atomic_read(&cld->cld_refcount));
131 if (atomic_dec_and_test(&cld->cld_refcount)) {
132 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
133 class_export_put(cld->cld_mgcexp);
134 spin_lock(&config_list_lock);
135 list_del(&cld->cld_list_chain);
136 spin_unlock(&config_list_lock);
137 OBD_FREE(cld->cld_logname, strlen(cld->cld_logname) + 1);
138 if (cld->cld_cfg.cfg_instance != NULL)
139 OBD_FREE(cld->cld_cfg.cfg_instance,
140 strlen(cld->cld_cfg.cfg_instance) + 1);
141 OBD_FREE(cld, sizeof(*cld));
146 /* Find a config log by name */
147 static struct config_llog_data *config_log_find(char *logname,
148 struct config_llog_instance *cfg)
150 struct list_head *tmp;
151 struct config_llog_data *cld;
152 char *logid = logname;
153 int match_instance = 0;
156 if (cfg && cfg->cfg_instance) {
158 logid = cfg->cfg_instance;
161 CERROR("No log specified\n");
162 RETURN(ERR_PTR(-EINVAL));
165 spin_lock(&config_list_lock);
166 list_for_each(tmp, &config_llog_list) {
167 cld = list_entry(tmp, struct config_llog_data, cld_list_chain);
168 if (match_instance && cld->cld_cfg.cfg_instance &&
169 strcmp(logid, cld->cld_cfg.cfg_instance) == 0)
171 if (!match_instance &&
172 strcmp(logid, cld->cld_logname) == 0)
175 spin_unlock(&config_list_lock);
177 CDEBUG(D_CONFIG, "can't get log %s\n", logid);
178 RETURN(ERR_PTR(-ENOENT));
180 atomic_inc(&cld->cld_refcount);
181 spin_unlock(&config_list_lock);
185 /* Add this log to our list of active logs.
186 We have one active log per "mount" - client instance or servername.
187 Each instance may be at a different point in the log. */
188 static int config_log_add(char *logname, struct config_llog_instance *cfg,
189 struct super_block *sb)
191 struct config_llog_data *cld;
192 struct lustre_sb_info *lsi = s2lsi(sb);
196 CDEBUG(D_MGC, "adding config log %s:%s\n", logname, cfg->cfg_instance);
198 OBD_ALLOC(cld, sizeof(*cld));
201 OBD_ALLOC(cld->cld_logname, strlen(logname) + 1);
202 if (!cld->cld_logname) {
203 OBD_FREE(cld, sizeof(*cld));
206 strcpy(cld->cld_logname, logname);
208 cld->cld_cfg.cfg_last_idx = 0;
209 cld->cld_cfg.cfg_flags = 0;
210 cld->cld_cfg.cfg_sb = sb;
211 atomic_set(&cld->cld_refcount, 1);
213 /* Keep the mgc around until we are done */
214 cld->cld_mgcexp = class_export_get(lsi->lsi_mgc->obd_self_export);
216 if (cfg->cfg_instance != NULL) {
217 OBD_ALLOC(cld->cld_cfg.cfg_instance,
218 strlen(cfg->cfg_instance) + 1);
219 strcpy(cld->cld_cfg.cfg_instance, cfg->cfg_instance);
221 rc = mgc_logname2resid(logname, &cld->cld_resid);
222 spin_lock(&config_list_lock);
223 list_add(&cld->cld_list_chain, &config_llog_list);
224 spin_unlock(&config_list_lock);
234 /* Stop watching for updates on this log. */
235 static int config_log_end(char *logname, struct config_llog_instance *cfg)
237 struct config_llog_data *cld;
241 cld = config_log_find(logname, cfg);
243 RETURN(PTR_ERR(cld));
244 /* drop the ref from the find */
247 cld->cld_stopping = 1;
248 /* drop the start ref */
250 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
255 /* reenqueue any lost locks */
256 #define RQ_RUNNING 0x1
260 static int rq_state = 0;
261 static cfs_waitq_t rq_waitq;
263 static int mgc_process_log(struct obd_device *mgc,
264 struct config_llog_data *cld);
265 static int mgc_requeue_add(struct config_llog_data *cld, int later);
267 static int mgc_requeue_thread(void *data)
269 struct l_wait_info lwi_now, lwi_later;
270 struct config_llog_data *cld, *n;
271 char name[] = "ll_cfg_requeue";
275 ptlrpc_daemonize(name);
277 CDEBUG(D_MGC, "Starting requeue thread\n");
279 lwi_later = LWI_TIMEOUT(60 * HZ, NULL, NULL);
280 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP), &lwi_later);
282 /* Keep trying failed locks periodically */
283 spin_lock(&config_list_lock);
284 while (rq_state & (RQ_NOW | RQ_LATER)) {
285 /* Any new or requeued lostlocks will change the state */
286 rq_state &= ~(RQ_NOW | RQ_LATER);
287 spin_unlock(&config_list_lock);
289 /* Always wait a few seconds to allow the server who
290 caused the lock revocation to finish its setup, plus some
291 random so everyone doesn't try to reconnect at once. */
292 lwi_now = LWI_TIMEOUT(3 * HZ + (ll_rand() & 0xff) * (HZ / 100),
294 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi_now);
296 spin_lock(&config_list_lock);
297 list_for_each_entry_safe(cld, n, &config_llog_list,
299 spin_unlock(&config_list_lock);
300 if (cld->cld_lostlock) {
301 CDEBUG(D_MGC, "updating log %s\n",
303 cld->cld_lostlock = 0;
304 rc = mgc_process_log(cld->cld_mgcexp->exp_obd,
306 /* Whether we enqueued again or not in
307 mgc_process_log, we're done with the ref
308 from the old enqueue */
311 spin_lock(&config_list_lock);
313 spin_unlock(&config_list_lock);
315 /* Wait a bit to see if anyone else needs a requeue */
316 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
318 spin_lock(&config_list_lock);
320 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
321 rq_state &= ~RQ_RUNNING;
322 spin_unlock(&config_list_lock);
324 CDEBUG(D_MGC, "Ending requeue thread\n");
328 /* Add a cld to the list to requeue. Start the requeue thread if needed.
329 We are responsible for dropping the config log reference from here on out. */
330 static int mgc_requeue_add(struct config_llog_data *cld, int later)
334 CDEBUG(D_INFO, "log %s: requeue (l=%d r=%d sp=%d st=%x)\n",
335 cld->cld_logname, later, atomic_read(&cld->cld_refcount),
336 cld->cld_stopping, rq_state);
338 /* Hold lock for rq_state */
339 spin_lock(&config_list_lock);
340 cld->cld_lostlock = 1;
342 if (cld->cld_stopping || (rq_state & RQ_STOP)) {
343 spin_unlock(&config_list_lock);
348 if (!(rq_state & RQ_RUNNING)) {
349 LASSERT(rq_state == 0);
350 rq_state = RQ_RUNNING | (later ? RQ_LATER : RQ_NOW);
351 spin_unlock(&config_list_lock);
352 rc = cfs_kernel_thread(mgc_requeue_thread, 0,
353 CLONE_VM | CLONE_FILES);
355 CERROR("log %s: cannot start requeue thread (%d),"
356 "no more log updates!\n", cld->cld_logname, rc);
357 /* Drop the ref, since the rq thread won't */
358 cld->cld_lostlock = 0;
364 rq_state |= later ? RQ_LATER : RQ_NOW;
365 spin_unlock(&config_list_lock);
366 cfs_waitq_signal(&rq_waitq);
372 /********************** class fns **********************/
374 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
375 struct vfsmount *mnt)
377 struct lvfs_run_ctxt saved;
378 struct lustre_sb_info *lsi = s2lsi(sb);
379 struct client_obd *cli = &obd->u.cli;
380 struct dentry *dentry;
386 LASSERT(lsi->lsi_srv_mnt == mnt);
388 /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
389 down(&cli->cl_mgc_sem);
391 cleanup_group_info();
393 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
394 if (IS_ERR(obd->obd_fsops)) {
395 up(&cli->cl_mgc_sem);
396 CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd),
397 PTR_ERR(obd->obd_fsops));
398 RETURN(PTR_ERR(obd->obd_fsops));
401 cli->cl_mgc_vfsmnt = mnt;
402 fsfilt_setup(obd, mnt->mnt_sb);
404 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
405 obd->obd_lvfs_ctxt.pwdmnt = mnt;
406 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
407 obd->obd_lvfs_ctxt.fs = get_ds();
409 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
410 dentry = lookup_one_len(MOUNT_CONFIGS_DIR, current->fs->pwd,
411 strlen(MOUNT_CONFIGS_DIR));
412 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
413 if (IS_ERR(dentry)) {
414 err = PTR_ERR(dentry);
415 CERROR("cannot lookup %s directory: rc = %d\n",
416 MOUNT_CONFIGS_DIR, err);
419 cli->cl_mgc_configs_dir = dentry;
421 /* We take an obd ref to insure that we can't get to mgc_cleanup
422 without calling mgc_fs_cleanup first. */
425 label = fsfilt_get_label(obd, mnt->mnt_sb);
427 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
429 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
433 fsfilt_put_ops(obd->obd_fsops);
434 obd->obd_fsops = NULL;
435 cli->cl_mgc_vfsmnt = NULL;
436 up(&cli->cl_mgc_sem);
440 static int mgc_fs_cleanup(struct obd_device *obd)
442 struct client_obd *cli = &obd->u.cli;
446 LASSERT(cli->cl_mgc_vfsmnt != NULL);
448 if (cli->cl_mgc_configs_dir != NULL) {
449 struct lvfs_run_ctxt saved;
450 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
451 l_dput(cli->cl_mgc_configs_dir);
452 cli->cl_mgc_configs_dir = NULL;
453 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
457 cli->cl_mgc_vfsmnt = NULL;
459 fsfilt_put_ops(obd->obd_fsops);
461 up(&cli->cl_mgc_sem);
466 static atomic_t mgc_count = ATOMIC_INIT(0);
467 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
473 case OBD_CLEANUP_EARLY:
475 case OBD_CLEANUP_EXPORTS:
476 if (atomic_dec_and_test(&mgc_count)) {
477 /* Kick the requeue waitq - cld's should all be
479 spin_lock(&config_list_lock);
481 spin_unlock(&config_list_lock);
482 cfs_waitq_signal(&rq_waitq);
484 rc = obd_llog_finish(obd, 0);
486 CERROR("failed to cleanup llogging subsystems\n");
492 static int mgc_cleanup(struct obd_device *obd)
494 struct client_obd *cli = &obd->u.cli;
498 LASSERT(cli->cl_mgc_vfsmnt == NULL);
500 /* COMPAT_146 - old config logs may have added profiles we don't
502 if (obd->obd_type->typ_refcnt <= 1)
503 /* Only for the last mgc */
504 class_del_profiles();
506 lprocfs_obd_cleanup(obd);
509 rc = client_obd_cleanup(obd);
513 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
515 struct lprocfs_static_vars lvars;
521 rc = client_obd_setup(obd, lcfg);
523 GOTO(err_decref, rc);
525 rc = obd_llog_init(obd, &obd->obd_olg, obd, 0, NULL, NULL);
527 CERROR("failed to setup llogging subsystems\n");
528 GOTO(err_cleanup, rc);
531 lprocfs_mgc_init_vars(&lvars);
532 lprocfs_obd_setup(obd, lvars.obd_vars);
534 spin_lock(&config_list_lock);
535 atomic_inc(&mgc_count);
536 if (atomic_read(&mgc_count) == 1) {
537 rq_state &= ~RQ_STOP;
538 cfs_waitq_init(&rq_waitq);
540 spin_unlock(&config_list_lock);
545 client_obd_cleanup(obd);
551 /* based on ll_mdc_blocking_ast */
552 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
553 void *data, int flag)
555 struct lustre_handle lockh;
556 struct config_llog_data *cld = (struct config_llog_data *)data;
561 case LDLM_CB_BLOCKING:
562 /* mgs wants the lock, give it up... */
563 LDLM_DEBUG(lock, "MGC blocking CB");
564 ldlm_lock2handle(lock, &lockh);
565 rc = ldlm_cli_cancel(&lockh);
567 case LDLM_CB_CANCELING: {
568 /* We've given up the lock, prepare ourselves to update. */
569 LDLM_DEBUG(lock, "MGC cancel CB");
571 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
572 lock->l_resource->lr_name.name[0],
573 (char *)&lock->l_resource->lr_name.name[0]);
576 CERROR("missing data, won't requeue\n");
579 /* Are we done with this log? */
580 if (cld->cld_stopping) {
581 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
586 /* Make sure not to re-enqueue when the mgc is stopping
587 (we get called from client_disconnect_export) */
588 if (!lock->l_conn_export ||
589 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
590 CDEBUG(D_MGC, "log %s: disconnecting, won't requeue\n",
595 /* Did we fail to get the lock? */
596 if (lock->l_req_mode != lock->l_granted_mode) {
597 CDEBUG(D_MGC, "log %s: original grant failed, will "
598 "requeue later\n", cld->cld_logname);
599 /* Try to re-enqueue later */
600 rc = mgc_requeue_add(cld, 1);
604 rc = mgc_requeue_add(cld, 0);
613 CERROR("%s CB failed %d:\n", flag == LDLM_CB_BLOCKING ?
614 "blocking" : "cancel", rc);
615 LDLM_ERROR(lock, "MGC ast");
620 /* Send parameter to MGS*/
621 static int mgc_set_mgs_param(struct obd_export *exp,
622 struct mgs_send_param *msp)
624 struct ptlrpc_request *req;
625 struct mgs_send_param *req_msp, *rep_msp;
626 int size[] = { sizeof(struct ptlrpc_body), sizeof(*req_msp) };
627 __u32 rep_size[] = { sizeof(struct ptlrpc_body), sizeof(*msp) };
631 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_MGS_VERSION,
632 MGS_SET_INFO, 2, size, NULL);
636 req_msp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*req_msp));
640 memcpy(req_msp, msp, sizeof(*req_msp));
641 ptlrpc_req_set_repsize(req, 2, rep_size);
642 rc = ptlrpc_queue_wait(req);
644 rep_msp = lustre_swab_repbuf(req, REPLY_REC_OFF,
645 sizeof(*rep_msp), NULL);
646 memcpy(msp, rep_msp, sizeof(*rep_msp));
649 ptlrpc_req_finished(req);
654 /* Take a config lock so we can get cancel notifications */
655 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
656 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
657 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
658 void *data, __u32 lvb_len, void *lvb_swabber,
659 struct lustre_handle *lockh)
661 struct config_llog_data *cld = (struct config_llog_data *)data;
662 struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
663 ldlm_completion_ast, NULL, data};
668 CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
669 cld->cld_resid.name[0]);
671 /* We can only drop this config log ref when we drop the lock */
672 if (config_log_get(cld))
673 RETURN(ELDLM_LOCK_ABORTED);
675 /* We need a callback for every lockholder, so don't try to
676 ldlm_lock_match (see rev 1.1.2.11.2.47) */
678 rc = ldlm_cli_enqueue(exp, NULL, &einfo, &cld->cld_resid,
679 NULL, flags, NULL, 0, NULL, lockh, 0);
680 /* A failed enqueue should still call the mgc_blocking_ast,
681 where it will be requeued if needed ("grant failed"). */
686 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
687 __u32 mode, struct lustre_handle *lockh)
691 ldlm_lock_decref(lockh, mode);
697 static int mgc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
698 void *karg, void *uarg)
700 struct obd_device *obd = exp->exp_obd;
701 struct obd_ioctl_data *data = karg;
702 struct llog_ctxt *ctxt;
703 struct lvfs_run_ctxt saved;
707 if (!try_module_get(THIS_MODULE)) {
708 CERROR("Can't get module. Is it alive?");
712 /* REPLicator context */
713 case OBD_IOC_PARSE: {
714 CERROR("MGC parsing llog %s\n", data->ioc_inlbuf1);
715 ctxt = llog_get_context(exp->exp_obd, LLOG_CONFIG_REPL_CTXT);
716 rc = class_config_parse_llog(ctxt, data->ioc_inlbuf1, NULL);
720 case OBD_IOC_LLOG_INFO:
721 case OBD_IOC_LLOG_PRINT: {
722 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
723 rc = llog_ioctl(ctxt, cmd, data);
728 /* ORIGinator context */
729 case OBD_IOC_DUMP_LOG: {
730 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
731 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
732 rc = class_config_dump_llog(ctxt, data->ioc_inlbuf1, NULL);
733 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
740 CERROR("mgc_ioctl(): unrecognised ioctl %#x\n", cmd);
741 GOTO(out, rc = -ENOTTY);
744 module_put(THIS_MODULE);
750 /* Send target_reg message to MGS */
751 static int mgc_target_register(struct obd_export *exp,
752 struct mgs_target_info *mti)
754 struct ptlrpc_request *req;
755 struct mgs_target_info *req_mti, *rep_mti;
759 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
760 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
765 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
766 memcpy(req_mti, mti, sizeof(*req_mti));
768 ptlrpc_request_set_replen(req);
770 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
772 rc = ptlrpc_queue_wait(req);
774 rep_mti = req_capsule_server_get(&req->rq_pill,
775 &RMF_MGS_TARGET_INFO);
776 memcpy(mti, rep_mti, sizeof(*rep_mti));
777 CDEBUG(D_MGC, "register %s got index = %d\n",
778 mti->mti_svname, mti->mti_stripe_index);
780 ptlrpc_req_finished(req);
785 int mgc_set_info_async(struct obd_export *exp, obd_count keylen,
786 void *key, obd_count vallen, void *val,
787 struct ptlrpc_request_set *set)
789 struct obd_import *imp = class_exp2cliimp(exp);
793 /* Try to "recover" the initial connection; i.e. retry */
794 if (KEY_IS(KEY_INIT_RECOV)) {
795 if (vallen != sizeof(int))
797 spin_lock(&imp->imp_lock);
798 imp->imp_initial_recov = *(int *)val;
799 spin_unlock(&imp->imp_lock);
800 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
801 exp->exp_obd->obd_name, imp->imp_initial_recov);
804 /* Turn off initial_recov after we try all backup servers once */
805 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
807 if (vallen != sizeof(int))
810 spin_lock(&imp->imp_lock);
811 imp->imp_initial_recov_bk = value > 0;
812 /* Even after the initial connection, give up all comms if
813 nobody answers the first time. */
814 imp->imp_recon_bk = 1;
815 spin_unlock(&imp->imp_lock);
816 CDEBUG(D_MGC, "InitRecov %s %d/%d:d%d:i%d:r%d:or%d:%s\n",
817 imp->imp_obd->obd_name, value, imp->imp_initial_recov,
818 imp->imp_deactive, imp->imp_invalid,
819 imp->imp_replayable, imp->imp_obd->obd_replayable,
820 ptlrpc_import_state_name(imp->imp_state));
821 /* Resurrect if we previously died */
822 if (imp->imp_invalid || value > 1)
823 ptlrpc_reconnect_import(imp);
826 /* FIXME move this to mgc_process_config */
827 if (KEY_IS(KEY_REGISTER_TARGET)) {
828 struct mgs_target_info *mti;
829 if (vallen != sizeof(struct mgs_target_info))
831 mti = (struct mgs_target_info *)val;
832 CDEBUG(D_MGC, "register_target %s %#x\n",
833 mti->mti_svname, mti->mti_flags);
834 rc = mgc_target_register(exp, mti);
837 if (KEY_IS(KEY_SET_FS)) {
838 struct super_block *sb = (struct super_block *)val;
839 struct lustre_sb_info *lsi;
840 if (vallen != sizeof(struct super_block))
843 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
845 CERROR("set_fs got %d\n", rc);
849 if (KEY_IS(KEY_CLEAR_FS)) {
852 rc = mgc_fs_cleanup(exp->exp_obd);
854 CERROR("clear_fs got %d\n", rc);
858 if (KEY_IS(KEY_SET_INFO)) {
859 struct mgs_send_param *msp;
861 msp = (struct mgs_send_param *)val;
862 rc = mgc_set_mgs_param(exp, msp);
869 static int mgc_import_event(struct obd_device *obd,
870 struct obd_import *imp,
871 enum obd_import_event event)
875 LASSERT(imp->imp_obd == obd);
876 CDEBUG(D_MGC, "import event %#x\n", event);
879 case IMP_EVENT_DISCON:
880 /* MGC imports should not wait for recovery */
882 case IMP_EVENT_INACTIVE:
884 case IMP_EVENT_INVALIDATE: {
885 struct ldlm_namespace *ns = obd->obd_namespace;
886 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
889 case IMP_EVENT_ACTIVE:
890 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
891 /* Clearing obd_no_recov allows us to continue pinging */
892 obd->obd_no_recov = 0;
897 CERROR("Unknown import event %#x\n", event);
903 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
904 struct obd_device *tgt, int count,
905 struct llog_catid *logid, struct obd_uuid *uuid)
907 struct llog_ctxt *ctxt;
911 LASSERT(olg == &obd->obd_olg);
913 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
918 rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
921 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
922 llog_initiator_connect(ctxt);
929 static int mgc_llog_finish(struct obd_device *obd, int count)
931 struct llog_ctxt *ctxt;
935 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
937 rc = llog_cleanup(ctxt);
939 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
941 rc2 = llog_cleanup(ctxt);
949 /* identical to mgs_log_is_empty */
950 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
953 struct lvfs_run_ctxt saved;
954 struct llog_handle *llh;
957 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
958 rc = llog_create(ctxt, &llh, NULL, name);
960 llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL);
961 rc = llog_get_size(llh);
964 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
965 /* header is record 1 */
969 static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec,
972 struct llog_rec_hdr local_rec = *rec;
973 struct llog_handle *local_llh = (struct llog_handle *)data;
974 char *cfg_buf = (char*) (rec + 1);
975 struct lustre_cfg *lcfg;
979 /* Append all records */
980 local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
981 rc = llog_write_rec(local_llh, &local_rec, NULL, 0,
982 (void *)cfg_buf, -1);
984 lcfg = (struct lustre_cfg *)cfg_buf;
985 CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
986 rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
987 lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
992 /* Copy a remote log locally */
993 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
994 struct llog_ctxt *lctxt, char *logname)
996 struct llog_handle *local_llh, *remote_llh;
997 struct obd_uuid *uuid;
1002 /* Write new log to a temp name, then vfs_rename over logname
1003 upon successful completion. */
1005 OBD_ALLOC(temp_log, strlen(logname) + 1);
1008 sprintf(temp_log, "%sT", logname);
1010 /* Make sure there's no old temp log */
1011 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1014 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL);
1017 rc = llog_destroy(local_llh);
1018 llog_free_handle(local_llh);
1022 /* open local log */
1023 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1027 /* set the log header uuid for fun */
1028 OBD_ALLOC_PTR(uuid);
1029 obd_str2uuid(uuid, logname);
1030 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid);
1033 GOTO(out_closel, rc);
1035 /* open remote log */
1036 rc = llog_create(rctxt, &remote_llh, NULL, logname);
1038 GOTO(out_closel, rc);
1039 rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL);
1041 GOTO(out_closer, rc);
1043 /* Copy remote log */
1044 rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL);
1047 rc2 = llog_close(remote_llh);
1051 rc2 = llog_close(local_llh);
1055 /* We've copied the remote log to the local temp log, now
1056 replace the old local log with the temp log. */
1058 struct client_obd *cli = &obd->u.cli;
1060 LASSERT(cli->cl_mgc_configs_dir);
1061 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1064 CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1067 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1068 OBD_FREE(temp_log, strlen(logname) + 1);
1072 DECLARE_MUTEX(llog_process_lock);
1074 /* Get a config log from the MGS and process it.
1075 This func is called for both clients and servers. */
1076 static int mgc_process_log(struct obd_device *mgc,
1077 struct config_llog_data *cld)
1079 struct llog_ctxt *ctxt, *lctxt;
1080 struct lustre_handle lockh;
1081 struct client_obd *cli = &mgc->u.cli;
1082 struct lvfs_run_ctxt saved;
1083 struct lustre_sb_info *lsi;
1084 int rc = 0, rcl, flags = 0, must_pop = 0;
1087 if (!cld || !cld->cld_cfg.cfg_sb) {
1088 /* This should never happen */
1089 CERROR("Missing cld, aborting log update\n");
1092 if (cld->cld_stopping)
1095 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1097 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1099 CDEBUG(D_MGC, "Process log %s:%s from %d\n", cld->cld_logname,
1100 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1102 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1104 CERROR("missing llog context\n");
1108 /* I don't want mutliple processes running process_log at once --
1109 sounds like badness. It actually might be fine, as long as
1110 we're not trying to update from the same log
1111 simultaneously (in which case we should use a per-log sem.) */
1112 down(&llog_process_lock);
1114 /* Get the cfg lock on the llog */
1115 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1116 LCK_CR, &flags, NULL, NULL, NULL,
1117 cld, 0, NULL, &lockh);
1119 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1121 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1123 /* Copy the setup log locally if we can. Don't mess around if we're
1124 running an MGS though (logs are already local). */
1125 if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
1126 (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1127 !IS_MGS(lsi->lsi_ldd)) {
1128 push_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1131 /* Only try to copy log if we have the lock. */
1132 rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1134 if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1135 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1136 "log %s and no local copy."
1137 "\n", cld->cld_logname);
1138 GOTO(out_pop, rc = -ENOTCONN);
1140 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1141 "copy for now, will try to update later.\n",
1144 /* Now, whether we copied or not, start using the local llog.
1145 If we failed to copy, we'll start using whatever the old
1147 llog_ctxt_put(ctxt);
1151 /* logname and instance info should be the same, so use our
1152 copy of the instance for the update. The cfg_last_idx will
1154 rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg);
1156 llog_ctxt_put(ctxt);
1158 llog_ctxt_put(lctxt);
1160 pop_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1162 /* Now drop the lock so MGS can revoke it */
1164 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1167 CERROR("Can't drop cfg lock: %d\n", rcl);
1170 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1171 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1173 up(&llog_process_lock);
1178 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1180 struct lustre_cfg *lcfg = buf;
1185 switch(cmd = lcfg->lcfg_command) {
1186 case LCFG_LOV_ADD_OBD: {
1187 /* Add any new target, not just osts */
1188 struct mgs_target_info *mti;
1190 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1191 sizeof(struct mgs_target_info))
1192 GOTO(out, rc = -EINVAL);
1194 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1195 CDEBUG(D_MGC, "add_target %s %#x\n",
1196 mti->mti_svname, mti->mti_flags);
1197 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1200 case LCFG_LOV_DEL_OBD:
1201 /* Remove target from the fs? */
1203 CERROR("lov_del_obd unimplemented\n");
1206 case LCFG_LOG_START: {
1207 struct config_llog_data *cld;
1208 struct config_llog_instance *cfg;
1209 struct super_block *sb;
1210 char *logname = lustre_cfg_string(lcfg, 1);
1211 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1212 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1214 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1217 /* We're only called through here on the initial mount */
1218 rc = config_log_add(logname, cfg, sb);
1221 cld = config_log_find(logname, cfg);
1228 /* FIXME only set this for old logs! Right now this forces
1229 us to always skip the "inside markers" check */
1230 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1232 rc = mgc_process_log(obd, cld);
1233 config_log_put(cld);
1237 case LCFG_LOG_END: {
1238 struct config_llog_instance *cfg = NULL;
1239 char *logname = lustre_cfg_string(lcfg, 1);
1240 if (lcfg->lcfg_bufcount >= 2)
1241 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1243 rc = config_log_end(logname, cfg);
1247 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1248 GOTO(out, rc = -EINVAL);
1256 struct obd_ops mgc_obd_ops = {
1257 .o_owner = THIS_MODULE,
1258 .o_setup = mgc_setup,
1259 .o_precleanup = mgc_precleanup,
1260 .o_cleanup = mgc_cleanup,
1261 .o_add_conn = client_import_add_conn,
1262 .o_del_conn = client_import_del_conn,
1263 .o_connect = client_connect_import,
1264 .o_disconnect = client_disconnect_export,
1265 //.o_enqueue = mgc_enqueue,
1266 .o_cancel = mgc_cancel,
1267 //.o_iocontrol = mgc_iocontrol,
1268 .o_set_info_async = mgc_set_info_async,
1269 .o_import_event = mgc_import_event,
1270 .o_llog_init = mgc_llog_init,
1271 .o_llog_finish = mgc_llog_finish,
1272 .o_process_config = mgc_process_config,
1275 int __init mgc_init(void)
1277 return class_register_type(&mgc_obd_ops, NULL, NULL,
1278 LUSTRE_MGC_NAME, NULL);
1282 static void /*__exit*/ mgc_exit(void)
1284 class_unregister_type(LUSTRE_MGC_NAME);
1287 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1288 MODULE_DESCRIPTION("Lustre Management Client");
1289 MODULE_LICENSE("GPL");
1291 module_init(mgc_init);
1292 module_exit(mgc_exit);