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);
341 if (cld->cld_stopping || (rq_state & RQ_STOP)) {
342 spin_unlock(&config_list_lock);
347 cld->cld_lostlock = 1;
349 if (!(rq_state & RQ_RUNNING)) {
350 LASSERT(rq_state == 0);
351 rq_state = RQ_RUNNING | (later ? RQ_LATER : RQ_NOW);
352 spin_unlock(&config_list_lock);
353 rc = cfs_kernel_thread(mgc_requeue_thread, 0,
354 CLONE_VM | CLONE_FILES);
356 CERROR("log %s: cannot start requeue thread (%d),"
357 "no more log updates!\n", cld->cld_logname, rc);
358 /* Drop the ref, since the rq thread won't */
359 cld->cld_lostlock = 0;
365 rq_state |= later ? RQ_LATER : RQ_NOW;
366 spin_unlock(&config_list_lock);
367 cfs_waitq_signal(&rq_waitq);
373 /********************** class fns **********************/
375 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
376 struct vfsmount *mnt)
378 struct lvfs_run_ctxt saved;
379 struct lustre_sb_info *lsi = s2lsi(sb);
380 struct client_obd *cli = &obd->u.cli;
381 struct dentry *dentry;
387 LASSERT(lsi->lsi_srv_mnt == mnt);
389 /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
390 down(&cli->cl_mgc_sem);
392 cleanup_group_info();
394 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
395 if (IS_ERR(obd->obd_fsops)) {
396 up(&cli->cl_mgc_sem);
397 CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd),
398 PTR_ERR(obd->obd_fsops));
399 RETURN(PTR_ERR(obd->obd_fsops));
402 cli->cl_mgc_vfsmnt = mnt;
403 fsfilt_setup(obd, mnt->mnt_sb);
405 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
406 obd->obd_lvfs_ctxt.pwdmnt = mnt;
407 obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
408 obd->obd_lvfs_ctxt.fs = get_ds();
410 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
411 dentry = lookup_one_len(MOUNT_CONFIGS_DIR, current->fs->pwd,
412 strlen(MOUNT_CONFIGS_DIR));
413 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
414 if (IS_ERR(dentry)) {
415 err = PTR_ERR(dentry);
416 CERROR("cannot lookup %s directory: rc = %d\n",
417 MOUNT_CONFIGS_DIR, err);
420 cli->cl_mgc_configs_dir = dentry;
422 /* We take an obd ref to insure that we can't get to mgc_cleanup
423 without calling mgc_fs_cleanup first. */
424 class_incref(obd, "mgc_fs", obd);
426 label = fsfilt_get_label(obd, mnt->mnt_sb);
428 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
430 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
434 fsfilt_put_ops(obd->obd_fsops);
435 obd->obd_fsops = NULL;
436 cli->cl_mgc_vfsmnt = NULL;
437 up(&cli->cl_mgc_sem);
441 static int mgc_fs_cleanup(struct obd_device *obd)
443 struct client_obd *cli = &obd->u.cli;
447 LASSERT(cli->cl_mgc_vfsmnt != NULL);
449 if (cli->cl_mgc_configs_dir != NULL) {
450 struct lvfs_run_ctxt saved;
451 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
452 l_dput(cli->cl_mgc_configs_dir);
453 cli->cl_mgc_configs_dir = NULL;
454 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
455 class_decref(obd, "mgc_fs", obd);
458 cli->cl_mgc_vfsmnt = NULL;
460 fsfilt_put_ops(obd->obd_fsops);
462 up(&cli->cl_mgc_sem);
467 static atomic_t mgc_count = ATOMIC_INIT(0);
468 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
474 case OBD_CLEANUP_EARLY:
476 case OBD_CLEANUP_EXPORTS:
477 if (atomic_dec_and_test(&mgc_count)) {
478 /* Kick the requeue waitq - cld's should all be
480 spin_lock(&config_list_lock);
482 spin_unlock(&config_list_lock);
483 cfs_waitq_signal(&rq_waitq);
485 rc = obd_llog_finish(obd, 0);
487 CERROR("failed to cleanup llogging subsystems\n");
493 static int mgc_cleanup(struct obd_device *obd)
495 struct client_obd *cli = &obd->u.cli;
499 LASSERT(cli->cl_mgc_vfsmnt == NULL);
501 /* COMPAT_146 - old config logs may have added profiles we don't
503 if (obd->obd_type->typ_refcnt <= 1)
504 /* Only for the last mgc */
505 class_del_profiles();
507 lprocfs_obd_cleanup(obd);
510 rc = client_obd_cleanup(obd);
514 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
516 struct lprocfs_static_vars lvars;
522 rc = client_obd_setup(obd, lcfg);
524 GOTO(err_decref, rc);
526 rc = obd_llog_init(obd, &obd->obd_olg, obd, 0, NULL, NULL);
528 CERROR("failed to setup llogging subsystems\n");
529 GOTO(err_cleanup, rc);
532 lprocfs_mgc_init_vars(&lvars);
533 lprocfs_obd_setup(obd, lvars.obd_vars);
535 spin_lock(&config_list_lock);
536 atomic_inc(&mgc_count);
537 if (atomic_read(&mgc_count) == 1) {
538 rq_state &= ~RQ_STOP;
539 cfs_waitq_init(&rq_waitq);
541 spin_unlock(&config_list_lock);
546 client_obd_cleanup(obd);
552 /* based on ll_mdc_blocking_ast */
553 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
554 void *data, int flag)
556 struct lustre_handle lockh;
557 struct config_llog_data *cld = (struct config_llog_data *)data;
562 case LDLM_CB_BLOCKING:
563 /* mgs wants the lock, give it up... */
564 LDLM_DEBUG(lock, "MGC blocking CB");
565 ldlm_lock2handle(lock, &lockh);
566 rc = ldlm_cli_cancel(&lockh);
568 case LDLM_CB_CANCELING: {
569 /* We've given up the lock, prepare ourselves to update. */
570 LDLM_DEBUG(lock, "MGC cancel CB");
572 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
573 lock->l_resource->lr_name.name[0],
574 (char *)&lock->l_resource->lr_name.name[0]);
577 CERROR("missing data, won't requeue\n");
580 /* Are we done with this log? */
581 if (cld->cld_stopping) {
582 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
587 /* Make sure not to re-enqueue when the mgc is stopping
588 (we get called from client_disconnect_export) */
589 if (!lock->l_conn_export ||
590 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
591 CDEBUG(D_MGC, "log %s: disconnecting, won't requeue\n",
596 /* Did we fail to get the lock? */
597 if (lock->l_req_mode != lock->l_granted_mode) {
598 CDEBUG(D_MGC, "log %s: original grant failed, will "
599 "requeue later\n", cld->cld_logname);
600 /* Try to re-enqueue later */
601 rc = mgc_requeue_add(cld, 1);
605 rc = mgc_requeue_add(cld, 0);
614 CERROR("%s CB failed %d:\n", flag == LDLM_CB_BLOCKING ?
615 "blocking" : "cancel", rc);
616 LDLM_ERROR(lock, "MGC ast");
621 /* Send parameter to MGS*/
622 static int mgc_set_mgs_param(struct obd_export *exp,
623 struct mgs_send_param *msp)
625 struct ptlrpc_request *req;
626 struct mgs_send_param *req_msp, *rep_msp;
627 int size[] = { sizeof(struct ptlrpc_body), sizeof(*req_msp) };
628 __u32 rep_size[] = { sizeof(struct ptlrpc_body), sizeof(*msp) };
632 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_MGS_VERSION,
633 MGS_SET_INFO, 2, size, NULL);
637 req_msp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*req_msp));
641 memcpy(req_msp, msp, sizeof(*req_msp));
642 ptlrpc_req_set_repsize(req, 2, rep_size);
643 rc = ptlrpc_queue_wait(req);
645 rep_msp = lustre_swab_repbuf(req, REPLY_REC_OFF,
646 sizeof(*rep_msp), NULL);
647 memcpy(msp, rep_msp, sizeof(*rep_msp));
650 ptlrpc_req_finished(req);
655 /* Take a config lock so we can get cancel notifications */
656 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
657 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
658 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
659 void *data, __u32 lvb_len, void *lvb_swabber,
660 struct lustre_handle *lockh)
662 struct config_llog_data *cld = (struct config_llog_data *)data;
663 struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
664 ldlm_completion_ast, NULL, NULL, data};
669 CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
670 cld->cld_resid.name[0]);
672 /* We can only drop this config log ref when we drop the lock */
673 if (config_log_get(cld))
674 RETURN(ELDLM_LOCK_ABORTED);
676 /* We need a callback for every lockholder, so don't try to
677 ldlm_lock_match (see rev 1.1.2.11.2.47) */
679 rc = ldlm_cli_enqueue(exp, NULL, &einfo, &cld->cld_resid,
680 NULL, flags, NULL, 0, NULL, lockh, 0);
681 /* A failed enqueue should still call the mgc_blocking_ast,
682 where it will be requeued if needed ("grant failed"). */
687 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
688 __u32 mode, struct lustre_handle *lockh)
692 ldlm_lock_decref(lockh, mode);
698 static int mgc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
699 void *karg, void *uarg)
701 struct obd_device *obd = exp->exp_obd;
702 struct obd_ioctl_data *data = karg;
703 struct llog_ctxt *ctxt;
704 struct lvfs_run_ctxt saved;
708 if (!try_module_get(THIS_MODULE)) {
709 CERROR("Can't get module. Is it alive?");
713 /* REPLicator context */
714 case OBD_IOC_PARSE: {
715 CERROR("MGC parsing llog %s\n", data->ioc_inlbuf1);
716 ctxt = llog_get_context(exp->exp_obd, LLOG_CONFIG_REPL_CTXT);
717 rc = class_config_parse_llog(ctxt, data->ioc_inlbuf1, NULL);
721 case OBD_IOC_LLOG_INFO:
722 case OBD_IOC_LLOG_PRINT: {
723 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
724 rc = llog_ioctl(ctxt, cmd, data);
729 /* ORIGinator context */
730 case OBD_IOC_DUMP_LOG: {
731 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
732 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
733 rc = class_config_dump_llog(ctxt, data->ioc_inlbuf1, NULL);
734 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
741 CERROR("mgc_ioctl(): unrecognised ioctl %#x\n", cmd);
742 GOTO(out, rc = -ENOTTY);
745 module_put(THIS_MODULE);
751 /* Send target_reg message to MGS */
752 static int mgc_target_register(struct obd_export *exp,
753 struct mgs_target_info *mti)
755 struct ptlrpc_request *req;
756 struct mgs_target_info *req_mti, *rep_mti;
760 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
761 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
766 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
767 memcpy(req_mti, mti, sizeof(*req_mti));
769 ptlrpc_request_set_replen(req);
771 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
773 rc = ptlrpc_queue_wait(req);
775 rep_mti = req_capsule_server_get(&req->rq_pill,
776 &RMF_MGS_TARGET_INFO);
777 memcpy(mti, rep_mti, sizeof(*rep_mti));
778 CDEBUG(D_MGC, "register %s got index = %d\n",
779 mti->mti_svname, mti->mti_stripe_index);
781 ptlrpc_req_finished(req);
786 int mgc_set_info_async(struct obd_export *exp, obd_count keylen,
787 void *key, obd_count vallen, void *val,
788 struct ptlrpc_request_set *set)
790 struct obd_import *imp = class_exp2cliimp(exp);
794 /* Try to "recover" the initial connection; i.e. retry */
795 if (KEY_IS(KEY_INIT_RECOV)) {
796 if (vallen != sizeof(int))
798 spin_lock(&imp->imp_lock);
799 imp->imp_initial_recov = *(int *)val;
800 spin_unlock(&imp->imp_lock);
801 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
802 exp->exp_obd->obd_name, imp->imp_initial_recov);
805 /* Turn off initial_recov after we try all backup servers once */
806 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
808 if (vallen != sizeof(int))
811 spin_lock(&imp->imp_lock);
812 imp->imp_initial_recov_bk = value > 0;
813 /* Even after the initial connection, give up all comms if
814 nobody answers the first time. */
815 imp->imp_recon_bk = 1;
816 spin_unlock(&imp->imp_lock);
817 CDEBUG(D_MGC, "InitRecov %s %d/%d:d%d:i%d:r%d:or%d:%s\n",
818 imp->imp_obd->obd_name, value, imp->imp_initial_recov,
819 imp->imp_deactive, imp->imp_invalid,
820 imp->imp_replayable, imp->imp_obd->obd_replayable,
821 ptlrpc_import_state_name(imp->imp_state));
822 /* Resurrect if we previously died */
823 if (imp->imp_invalid || value > 1)
824 ptlrpc_reconnect_import(imp);
827 /* FIXME move this to mgc_process_config */
828 if (KEY_IS(KEY_REGISTER_TARGET)) {
829 struct mgs_target_info *mti;
830 if (vallen != sizeof(struct mgs_target_info))
832 mti = (struct mgs_target_info *)val;
833 CDEBUG(D_MGC, "register_target %s %#x\n",
834 mti->mti_svname, mti->mti_flags);
835 rc = mgc_target_register(exp, mti);
838 if (KEY_IS(KEY_SET_FS)) {
839 struct super_block *sb = (struct super_block *)val;
840 struct lustre_sb_info *lsi;
841 if (vallen != sizeof(struct super_block))
844 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
846 CERROR("set_fs got %d\n", rc);
850 if (KEY_IS(KEY_CLEAR_FS)) {
853 rc = mgc_fs_cleanup(exp->exp_obd);
855 CERROR("clear_fs got %d\n", rc);
859 if (KEY_IS(KEY_SET_INFO)) {
860 struct mgs_send_param *msp;
862 msp = (struct mgs_send_param *)val;
863 rc = mgc_set_mgs_param(exp, msp);
870 static int mgc_import_event(struct obd_device *obd,
871 struct obd_import *imp,
872 enum obd_import_event event)
876 LASSERT(imp->imp_obd == obd);
877 CDEBUG(D_MGC, "import event %#x\n", event);
880 case IMP_EVENT_DISCON:
881 /* MGC imports should not wait for recovery */
883 case IMP_EVENT_INACTIVE:
885 case IMP_EVENT_INVALIDATE: {
886 struct ldlm_namespace *ns = obd->obd_namespace;
887 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
890 case IMP_EVENT_ACTIVE:
891 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
892 /* Clearing obd_no_recov allows us to continue pinging */
893 obd->obd_no_recov = 0;
898 CERROR("Unknown import event %#x\n", event);
904 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
905 struct obd_device *tgt, int count,
906 struct llog_catid *logid, struct obd_uuid *uuid)
908 struct llog_ctxt *ctxt;
912 LASSERT(olg == &obd->obd_olg);
914 rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
919 rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
922 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
923 llog_initiator_connect(ctxt);
930 static int mgc_llog_finish(struct obd_device *obd, int count)
932 struct llog_ctxt *ctxt;
936 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
938 rc = llog_cleanup(ctxt);
940 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
942 rc2 = llog_cleanup(ctxt);
950 /* identical to mgs_log_is_empty */
951 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
954 struct lvfs_run_ctxt saved;
955 struct llog_handle *llh;
958 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
959 rc = llog_create(ctxt, &llh, NULL, name);
961 llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL);
962 rc = llog_get_size(llh);
965 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
966 /* header is record 1 */
970 static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec,
973 struct llog_rec_hdr local_rec = *rec;
974 struct llog_handle *local_llh = (struct llog_handle *)data;
975 char *cfg_buf = (char*) (rec + 1);
976 struct lustre_cfg *lcfg;
980 /* Append all records */
981 local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
982 rc = llog_write_rec(local_llh, &local_rec, NULL, 0,
983 (void *)cfg_buf, -1);
985 lcfg = (struct lustre_cfg *)cfg_buf;
986 CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
987 rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
988 lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
993 /* Copy a remote log locally */
994 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
995 struct llog_ctxt *lctxt, char *logname)
997 struct llog_handle *local_llh, *remote_llh;
998 struct obd_uuid *uuid;
1003 /* Write new log to a temp name, then vfs_rename over logname
1004 upon successful completion. */
1006 OBD_ALLOC(temp_log, strlen(logname) + 1);
1009 sprintf(temp_log, "%sT", logname);
1011 /* Make sure there's no old temp log */
1012 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1015 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL);
1018 rc = llog_destroy(local_llh);
1019 llog_free_handle(local_llh);
1023 /* open local log */
1024 rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1028 /* set the log header uuid for fun */
1029 OBD_ALLOC_PTR(uuid);
1030 obd_str2uuid(uuid, logname);
1031 rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid);
1034 GOTO(out_closel, rc);
1036 /* open remote log */
1037 rc = llog_create(rctxt, &remote_llh, NULL, logname);
1039 GOTO(out_closel, rc);
1040 rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL);
1042 GOTO(out_closer, rc);
1044 /* Copy remote log */
1045 rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL);
1048 rc2 = llog_close(remote_llh);
1052 rc2 = llog_close(local_llh);
1056 /* We've copied the remote log to the local temp log, now
1057 replace the old local log with the temp log. */
1059 struct client_obd *cli = &obd->u.cli;
1061 LASSERT(cli->cl_mgc_configs_dir);
1062 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1065 CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1068 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1069 OBD_FREE(temp_log, strlen(logname) + 1);
1073 DECLARE_MUTEX(llog_process_lock);
1075 /* Get a config log from the MGS and process it.
1076 This func is called for both clients and servers. */
1077 static int mgc_process_log(struct obd_device *mgc,
1078 struct config_llog_data *cld)
1080 struct llog_ctxt *ctxt, *lctxt;
1081 struct lustre_handle lockh;
1082 struct client_obd *cli = &mgc->u.cli;
1083 struct lvfs_run_ctxt saved;
1084 struct lustre_sb_info *lsi;
1085 int rc = 0, rcl, flags = 0, must_pop = 0;
1088 if (!cld || !cld->cld_cfg.cfg_sb) {
1089 /* This should never happen */
1090 CERROR("Missing cld, aborting log update\n");
1093 if (cld->cld_stopping)
1096 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1098 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1100 CDEBUG(D_MGC, "Process log %s:%s from %d\n", cld->cld_logname,
1101 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1103 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1105 CERROR("missing llog context\n");
1109 /* I don't want mutliple processes running process_log at once --
1110 sounds like badness. It actually might be fine, as long as
1111 we're not trying to update from the same log
1112 simultaneously (in which case we should use a per-log sem.) */
1113 down(&llog_process_lock);
1115 /* Get the cfg lock on the llog */
1116 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1117 LCK_CR, &flags, NULL, NULL, NULL,
1118 cld, 0, NULL, &lockh);
1120 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1122 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1124 /* Copy the setup log locally if we can. Don't mess around if we're
1125 running an MGS though (logs are already local). */
1126 if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
1127 (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1128 !IS_MGS(lsi->lsi_ldd)) {
1129 push_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1132 /* Only try to copy log if we have the lock. */
1133 rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1135 if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1136 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1137 "log %s and no local copy."
1138 "\n", cld->cld_logname);
1139 GOTO(out_pop, rc = -ENOTCONN);
1141 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1142 "copy for now, will try to update later.\n",
1145 /* Now, whether we copied or not, start using the local llog.
1146 If we failed to copy, we'll start using whatever the old
1148 llog_ctxt_put(ctxt);
1152 /* logname and instance info should be the same, so use our
1153 copy of the instance for the update. The cfg_last_idx will
1155 rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg);
1157 llog_ctxt_put(ctxt);
1159 llog_ctxt_put(lctxt);
1161 pop_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1163 /* Now drop the lock so MGS can revoke it */
1165 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1168 CERROR("Can't drop cfg lock: %d\n", rcl);
1171 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1172 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1174 up(&llog_process_lock);
1179 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1181 struct lustre_cfg *lcfg = buf;
1186 switch(cmd = lcfg->lcfg_command) {
1187 case LCFG_LOV_ADD_OBD: {
1188 /* Add any new target, not just osts */
1189 struct mgs_target_info *mti;
1191 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1192 sizeof(struct mgs_target_info))
1193 GOTO(out, rc = -EINVAL);
1195 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1196 CDEBUG(D_MGC, "add_target %s %#x\n",
1197 mti->mti_svname, mti->mti_flags);
1198 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1201 case LCFG_LOV_DEL_OBD:
1202 /* Remove target from the fs? */
1204 CERROR("lov_del_obd unimplemented\n");
1207 case LCFG_LOG_START: {
1208 struct config_llog_data *cld;
1209 struct config_llog_instance *cfg;
1210 struct super_block *sb;
1211 char *logname = lustre_cfg_string(lcfg, 1);
1212 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1213 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1215 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1218 /* We're only called through here on the initial mount */
1219 rc = config_log_add(logname, cfg, sb);
1222 cld = config_log_find(logname, cfg);
1229 /* FIXME only set this for old logs! Right now this forces
1230 us to always skip the "inside markers" check */
1231 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1233 rc = mgc_process_log(obd, cld);
1234 config_log_put(cld);
1238 case LCFG_LOG_END: {
1239 struct config_llog_instance *cfg = NULL;
1240 char *logname = lustre_cfg_string(lcfg, 1);
1241 if (lcfg->lcfg_bufcount >= 2)
1242 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1244 rc = config_log_end(logname, cfg);
1248 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1249 GOTO(out, rc = -EINVAL);
1257 struct obd_ops mgc_obd_ops = {
1258 .o_owner = THIS_MODULE,
1259 .o_setup = mgc_setup,
1260 .o_precleanup = mgc_precleanup,
1261 .o_cleanup = mgc_cleanup,
1262 .o_add_conn = client_import_add_conn,
1263 .o_del_conn = client_import_del_conn,
1264 .o_connect = client_connect_import,
1265 .o_disconnect = client_disconnect_export,
1266 //.o_enqueue = mgc_enqueue,
1267 .o_cancel = mgc_cancel,
1268 //.o_iocontrol = mgc_iocontrol,
1269 .o_set_info_async = mgc_set_info_async,
1270 .o_import_event = mgc_import_event,
1271 .o_llog_init = mgc_llog_init,
1272 .o_llog_finish = mgc_llog_finish,
1273 .o_process_config = mgc_process_config,
1276 int __init mgc_init(void)
1278 return class_register_type(&mgc_obd_ops, NULL, NULL,
1279 LUSTRE_MGC_NAME, NULL);
1283 static void /*__exit*/ mgc_exit(void)
1285 class_unregister_type(LUSTRE_MGC_NAME);
1288 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1289 MODULE_DESCRIPTION("Lustre Management Client");
1290 MODULE_LICENSE("GPL");
1292 module_init(mgc_init);
1293 module_exit(mgc_exit);