1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * lustre/mgs/mgs_handler.c
5 * Lustre Management Server (mgs) request handler
7 * Copyright (C) 2006 Cluster File Systems, Inc.
8 * Author: Nathan Rutman <nathan@clusterfs.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
27 # define EXPORT_SYMTAB
29 #define DEBUG_SUBSYSTEM S_MGS
30 #define D_MGS D_CONFIG/*|D_WARNING*/
33 # include <linux/module.h>
34 # include <linux/pagemap.h>
35 # include <linux/miscdevice.h>
36 # include <linux/init.h>
38 # include <liblustre.h>
41 #include <obd_class.h>
42 #include <lustre_dlm.h>
43 #include <lprocfs_status.h>
44 #include <lustre_fsfilt.h>
45 #include <lustre_commit_confd.h>
46 #include <lustre_disk.h>
47 #include "mgs_internal.h"
50 /* Establish a connection to the MGS.*/
51 static int mgs_connect(struct lustre_handle *conn, struct obd_device *obd,
52 struct obd_uuid *cluuid, struct obd_connect_data *data)
54 struct obd_export *exp;
58 if (!conn || !obd || !cluuid)
61 rc = class_connect(conn, obd, cluuid);
64 exp = class_conn2export(conn);
68 data->ocd_connect_flags &= MGS_CONNECT_SUPPORTED;
69 exp->exp_connect_flags = data->ocd_connect_flags;
70 data->ocd_version = LUSTRE_VERSION_CODE;
74 class_disconnect(exp);
76 class_export_put(exp);
82 static int mgs_disconnect(struct obd_export *exp)
88 class_export_get(exp);
90 /* Disconnect early so that clients can't keep using export */
91 rc = class_disconnect(exp);
92 ldlm_cancel_locks_for_export(exp);
94 /* complete all outstanding replies */
95 spin_lock(&exp->exp_lock);
96 while (!list_empty(&exp->exp_outstanding_replies)) {
97 struct ptlrpc_reply_state *rs =
98 list_entry(exp->exp_outstanding_replies.next,
99 struct ptlrpc_reply_state, rs_exp_list);
100 struct ptlrpc_service *svc = rs->rs_service;
102 spin_lock(&svc->srv_lock);
103 list_del_init(&rs->rs_exp_list);
104 ptlrpc_schedule_difficult_reply(rs);
105 spin_unlock(&svc->srv_lock);
107 spin_unlock(&exp->exp_lock);
109 class_export_put(exp);
113 static int mgs_cleanup(struct obd_device *obd);
114 static int mgs_handle(struct ptlrpc_request *req);
116 /* Start the MGS obd */
117 static int mgs_setup(struct obd_device *obd, obd_count len, void *buf)
119 struct lprocfs_static_vars lvars;
120 struct mgs_obd *mgs = &obd->u.mgs;
121 struct lustre_mount_info *lmi;
122 struct lustre_sb_info *lsi;
123 struct vfsmount *mnt;
127 CDEBUG(D_CONFIG, "Starting MGS\n");
130 lmi = server_get_mount(obd->obd_name);
132 RETURN(rc = -EINVAL);
135 lsi = s2lsi(lmi->lmi_sb);
136 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
137 if (IS_ERR(obd->obd_fsops))
138 GOTO(err_put, rc = PTR_ERR(obd->obd_fsops));
140 /* namespace for mgs llog */
141 obd->obd_namespace = ldlm_namespace_new("MGS", LDLM_NAMESPACE_SERVER);
142 if (obd->obd_namespace == NULL) {
144 GOTO(err_ops, rc = -ENOMEM);
148 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
149 "mgs_ldlm_client", &obd->obd_ldlm_client);
151 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
153 rc = mgs_fs_setup(obd, mnt);
155 CERROR("%s: MGS filesystem method init failed: rc = %d\n",
160 rc = llog_start_commit_thread();
164 rc = llog_setup(obd, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
169 /* No recovery for MGC's */
170 obd->obd_replayable = 0;
172 /* Internal mgs setup */
173 mgs_init_fsdb_list(obd);
174 sema_init(&mgs->mgs_sem, 1);
176 /* Start the service threads */
178 ptlrpc_init_svc(MGS_NBUFS, MGS_BUFSIZE, MGS_MAXREQSIZE,
179 MGS_MAXREPSIZE, MGS_REQUEST_PORTAL,
180 MGC_REPLY_PORTAL, MGS_SERVICE_WATCHDOG_TIMEOUT,
181 mgs_handle, LUSTRE_MGS_NAME,
182 obd->obd_proc_entry, NULL,
183 MGS_THREADS_AUTO_MIN, MGS_THREADS_AUTO_MAX,
186 if (!mgs->mgs_service) {
187 CERROR("failed to start service\n");
188 GOTO(err_fs, rc = -ENOMEM);
191 rc = ptlrpc_start_threads(obd, mgs->mgs_service);
193 GOTO(err_thread, rc);
196 lprocfs_init_vars(mgs, &lvars);
197 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
198 lproc_mgs_setup(obd);
201 ping_evictor_start();
203 LCONSOLE_INFO("MGS %s started\n", obd->obd_name);
208 ptlrpc_unregister_service(mgs->mgs_service);
210 /* No extra cleanup needed for llog_init_commit_thread() */
213 ldlm_namespace_free(obd->obd_namespace, 0);
214 obd->obd_namespace = NULL;
216 fsfilt_put_ops(obd->obd_fsops);
218 server_put_mount(obd->obd_name, mnt);
223 static int mgs_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
229 case OBD_CLEANUP_EARLY:
230 case OBD_CLEANUP_EXPORTS:
232 case OBD_CLEANUP_SELF_EXP:
233 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
234 rc = obd_llog_finish(obd, 0);
236 case OBD_CLEANUP_OBD:
242 static int mgs_ldlm_nsfree(void *data)
244 struct ldlm_namespace *ns = (struct ldlm_namespace *)data;
248 ptlrpc_daemonize("ll_mgs_nsfree");
249 rc = ldlm_namespace_free(ns, 1 /* obd_force should always be on */);
253 static int mgs_cleanup(struct obd_device *obd)
255 struct mgs_obd *mgs = &obd->u.mgs;
260 if (mgs->mgs_sb == NULL)
263 ptlrpc_unregister_service(mgs->mgs_service);
265 mgs_cleanup_fsdb_list(obd);
267 lprocfs_obd_cleanup(obd);
268 mgs->mgs_proc_live = NULL;
272 server_put_mount(obd->obd_name, mgs->mgs_vfsmnt);
275 /* Free the namespace in it's own thread, so that if the
276 ldlm_cancel_handler put the last mgs obd ref, we won't
278 cfs_kernel_thread(mgs_ldlm_nsfree, obd->obd_namespace,
279 CLONE_VM | CLONE_FILES);
281 fsfilt_put_ops(obd->obd_fsops);
283 LCONSOLE_INFO("%s has stopped.\n", obd->obd_name);
287 /* similar to filter_prepare_destroy */
288 static int mgs_get_cfg_lock(struct obd_device *obd, char *fsname,
289 struct lustre_handle *lockh)
291 struct ldlm_res_id res_id;
295 rc = mgc_logname2resid(fsname, &res_id);
297 rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id,
298 LDLM_PLAIN, NULL, LCK_EX,
299 &flags, ldlm_blocking_ast,
300 ldlm_completion_ast, NULL,
301 fsname, 0, NULL, lockh);
303 CERROR("can't take cfg lock for %s (%d)\n", fsname, rc);
308 static int mgs_put_cfg_lock(struct lustre_handle *lockh)
311 ldlm_lock_decref(lockh, LCK_EX);
318 static int mgs_check_target(struct obd_device *obd, struct mgs_target_info *mti)
323 rc = mgs_check_index(obd, mti);
325 LCONSOLE_ERROR("%s claims to have registered, but this MGS "
326 "does not know about it. Assuming writeconf.\n",
328 mti->mti_flags |= LDD_F_WRITECONF;
330 } else if (rc == -1) {
331 LCONSOLE_ERROR("Client log %s-client has disappeared! "
332 "Regenerating all logs.\n",
334 mti->mti_flags |= LDD_F_WRITECONF;
337 /* Index is correctly marked as used */
339 /* If the logs don't contain the mti_nids then add
340 them as failover nids */
341 rc = mgs_check_failnid(obd, mti);
347 /* Called whenever a target starts up. Flags indicate first connect, etc. */
348 static int mgs_handle_target_reg(struct ptlrpc_request *req)
350 struct obd_device *obd = req->rq_export->exp_obd;
351 struct lustre_handle lockh;
352 struct mgs_target_info *mti, *rep_mti;
353 int rep_size[] = { sizeof(struct ptlrpc_body), sizeof(*mti) };
357 mti = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*mti),
358 lustre_swab_mgs_target_info);
360 if (!(mti->mti_flags & (LDD_F_WRITECONF | LDD_F_UPGRADE14 |
362 /* We're just here as a startup ping. */
363 CDEBUG(D_MGS, "Server %s is running on %s\n",
364 mti->mti_svname, obd_export_nid2str(req->rq_export));
365 rc = mgs_check_target(obd, mti);
366 /* above will set appropriate mti flags */
368 /* Nothing wrong, or fatal error */
369 GOTO(out_nolock, rc);
372 /* Revoke the config lock to make sure nobody is reading. */
373 /* Although actually I think it should be alright if
374 someone was reading while we were updating the logs - if we
375 revoke at the end they will just update from where they left off. */
376 lockrc = mgs_get_cfg_lock(obd, mti->mti_fsname, &lockh);
377 if (lockrc != ELDLM_OK) {
378 LCONSOLE_ERROR("%s: Can't signal other nodes to update "
379 "their configuration (%d). Updating local logs "
380 "anyhow; you might have to manually restart "
381 "other nodes to get the latest configuration.\n",
382 obd->obd_name, lockrc);
385 /* Log writing contention is handled by the fsdb_sem */
387 if (mti->mti_flags & LDD_F_WRITECONF) {
388 if (mti->mti_flags & LDD_F_SV_TYPE_MDT) {
389 rc = mgs_erase_logs(obd, mti->mti_fsname);
390 LCONSOLE_WARN("%s: Logs for fs %s were removed by user "
391 "request. All servers must be restarted "
392 "in order to regenerate the logs."
393 "\n", obd->obd_name, mti->mti_fsname);
394 } else if (mti->mti_flags & LDD_F_SV_TYPE_OST) {
395 rc = mgs_erase_log(obd, mti->mti_svname);
396 LCONSOLE_WARN("%s: Regenerating %s log by user "
398 obd->obd_name, mti->mti_svname);
400 mti->mti_flags |= LDD_F_UPDATE;
401 /* Erased logs means start from scratch. */
402 mti->mti_flags &= ~LDD_F_UPGRADE14;
406 if (mti->mti_flags & LDD_F_UPGRADE14) {
407 rc = mgs_upgrade_sv_14(obd, mti);
409 CERROR("Can't upgrade from 1.4 (%d)\n", rc);
413 /* We're good to go */
414 mti->mti_flags |= LDD_F_UPDATE;
418 if (mti->mti_flags & LDD_F_UPDATE) {
419 CDEBUG(D_MGS, "updating %s, index=%d\n", mti->mti_svname,
420 mti->mti_stripe_index);
422 /* create or update the target log
423 and update the client/mdt logs */
424 rc = mgs_write_log_target(obd, mti);
426 CERROR("Failed to write %s log (%d)\n",
427 mti->mti_svname, rc);
431 mti->mti_flags &= ~(LDD_F_VIRGIN | LDD_F_UPDATE |
432 LDD_F_NEED_INDEX | LDD_F_WRITECONF);
433 mti->mti_flags |= LDD_F_REWRITE_LDD;
437 /* done with log update */
438 if (lockrc == ELDLM_OK)
439 mgs_put_cfg_lock(&lockh);
441 CDEBUG(D_MGS, "replying with %s, index=%d, rc=%d\n", mti->mti_svname,
442 mti->mti_stripe_index, rc);
443 lustre_pack_reply(req, 2, rep_size, NULL);
444 /* send back the whole mti in the reply */
445 rep_mti = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
447 memcpy(rep_mti, mti, sizeof(*rep_mti));
449 /* Flush logs to disk */
450 fsfilt_sync(obd, obd->u.mgs.mgs_sb);
454 int mgs_handle(struct ptlrpc_request *req)
456 int fail = OBD_FAIL_MGS_ALL_REPLY_NET;
460 OBD_FAIL_TIMEOUT(OBD_FAIL_MGS_SLOW_REQUEST_NET, 2);
461 OBD_FAIL_RETURN(OBD_FAIL_MGS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
463 LASSERT(current->journal_info == NULL);
464 opc = lustre_msg_get_opc(req->rq_reqmsg);
465 if (opc != MGS_CONNECT) {
466 if (req->rq_export == NULL) {
467 CERROR("lustre_mgs: operation %d on unconnected MGS\n",
469 req->rq_status = -ENOTCONN;
470 GOTO(out, rc = -ENOTCONN);
476 DEBUG_REQ(D_MGS, req, "connect");
477 rc = target_handle_connect(req, mgs_handle);
478 if (!rc && (lustre_msg_get_conn_cnt(req->rq_reqmsg) > 1))
479 /* Make clients trying to reconnect after a MGS restart
480 happy; also requires obd_replayable */
481 lustre_msg_add_op_flags(req->rq_repmsg,
482 MSG_CONNECT_RECONNECT);
485 DEBUG_REQ(D_MGS, req, "disconnect");
486 rc = target_handle_disconnect(req);
487 req->rq_status = rc; /* superfluous? */
490 DEBUG_REQ(D_MGS, req, "target add\n");
491 rc = mgs_handle_target_reg(req);
494 DEBUG_REQ(D_MGS, req, "target del\n");
495 //rc = mgs_handle_target_del(req);
499 DEBUG_REQ(D_MGS, req, "enqueue");
500 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
501 ldlm_server_blocking_ast, NULL);
503 case LDLM_BL_CALLBACK:
504 case LDLM_CP_CALLBACK:
505 DEBUG_REQ(D_MGS, req, "callback");
506 CERROR("callbacks should not happen on MGS\n");
511 DEBUG_REQ(D_INFO, req, "ping");
512 rc = target_handle_ping(req);
515 DEBUG_REQ(D_MGS, req, "log cancel\n");
516 rc = -ENOTSUPP; /* la la la */
519 case LLOG_ORIGIN_HANDLE_CREATE:
520 DEBUG_REQ(D_MGS, req, "llog_init");
521 rc = llog_origin_handle_create(req);
523 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
524 DEBUG_REQ(D_MGS, req, "llog next block");
525 rc = llog_origin_handle_next_block(req);
527 case LLOG_ORIGIN_HANDLE_READ_HEADER:
528 DEBUG_REQ(D_MGS, req, "llog read header");
529 rc = llog_origin_handle_read_header(req);
531 case LLOG_ORIGIN_HANDLE_CLOSE:
532 DEBUG_REQ(D_MGS, req, "llog close");
533 rc = llog_origin_handle_close(req);
536 DEBUG_REQ(D_MGS, req, "llog catinfo");
537 rc = llog_catinfo(req);
540 req->rq_status = -ENOTSUPP;
541 rc = ptlrpc_error(req);
545 LASSERT(current->journal_info == NULL);
548 CERROR("MGS handle cmd=%d rc=%d\n", opc, rc);
551 target_send_reply(req, rc, fail);
555 static inline int mgs_destroy_export(struct obd_export *exp)
559 target_destroy_export(exp);
564 /* from mdt_iocontrol */
565 int mgs_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
566 void *karg, void *uarg)
568 struct obd_device *obd = exp->exp_obd;
569 struct obd_ioctl_data *data = karg;
570 struct lvfs_run_ctxt saved;
574 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
578 case OBD_IOC_PARAM: {
579 struct lustre_handle lockh;
580 struct lustre_cfg *lcfg;
581 struct llog_rec_hdr rec;
582 char fsname[MTI_NAME_MAXLEN];
585 rec.lrh_len = llog_data_len(data->ioc_plen1);
587 if (data->ioc_type == LUSTRE_CFG_TYPE) {
588 rec.lrh_type = OBD_CFG_REC;
590 CERROR("unknown cfg record type:%d \n", data->ioc_type);
594 OBD_ALLOC(lcfg, data->ioc_plen1);
597 rc = copy_from_user(lcfg, data->ioc_pbuf1, data->ioc_plen1);
601 if (lcfg->lcfg_bufcount < 1)
602 GOTO(out_free, rc = -EINVAL);
604 rc = mgs_setparam(obd, lcfg, fsname);
606 CERROR("setparam err %d\n", rc);
610 /* Revoke lock so everyone updates. Should be alright if
611 someone was already reading while we were updating the logs,
612 so we don't really need to hold the lock while we're
615 lockrc = mgs_get_cfg_lock(obd, fsname, &lockh);
616 if (lockrc != ELDLM_OK)
617 CERROR("lock error %d for fs %s\n", lockrc,
620 mgs_put_cfg_lock(&lockh);
624 OBD_FREE(lcfg, data->ioc_plen1);
628 case OBD_IOC_DUMP_LOG: {
629 struct llog_ctxt *ctxt =
630 llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
631 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
632 rc = class_config_dump_llog(ctxt, data->ioc_inlbuf1, NULL);
633 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
640 case OBD_IOC_LLOG_CHECK:
641 case OBD_IOC_LLOG_INFO:
642 case OBD_IOC_LLOG_PRINT: {
643 struct llog_ctxt *ctxt =
644 llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
646 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
647 rc = llog_ioctl(ctxt, cmd, data);
648 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
654 CDEBUG(D_INFO, "unknown command %x\n", cmd);
660 /* use obd ops to offer management infrastructure */
661 static struct obd_ops mgs_obd_ops = {
662 .o_owner = THIS_MODULE,
663 .o_connect = mgs_connect,
664 .o_disconnect = mgs_disconnect,
665 .o_setup = mgs_setup,
666 .o_precleanup = mgs_precleanup,
667 .o_cleanup = mgs_cleanup,
668 .o_destroy_export = mgs_destroy_export,
669 .o_iocontrol = mgs_iocontrol,
672 static int __init mgs_init(void)
674 struct lprocfs_static_vars lvars;
676 lprocfs_init_vars(mgs, &lvars);
677 class_register_type(&mgs_obd_ops, lvars.module_vars, LUSTRE_MGS_NAME);
682 static void /*__exit*/ mgs_exit(void)
684 class_unregister_type(LUSTRE_MGS_NAME);
687 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
688 MODULE_DESCRIPTION("Lustre Management Server (MGS)");
689 MODULE_LICENSE("GPL");
691 module_init(mgs_init);
692 module_exit(mgs_exit);