4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
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>
41 #define DEBUG_SUBSYSTEM S_MGC
42 #define D_MGC D_CONFIG /*|D_WARNING*/
44 #include <linux/module.h>
45 #include <linux/kthread.h>
47 #include <dt_object.h>
48 #include <lprocfs_status.h>
49 #include <lustre_dlm.h>
50 #include <lustre_disk.h>
51 #include <lustre_log.h>
52 #include <lustre_nodemap.h>
53 #include <lustre_swab.h>
54 #include <obd_class.h>
56 #include "mgc_internal.h"
58 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
63 if (len > sizeof(resname)) {
64 CERROR("name too long: %s\n", name);
68 CERROR("missing name: %s\n", name);
71 memcpy(&resname, name, len);
73 /* Always use the same endianness for the resid */
74 memset(res_id, 0, sizeof(*res_id));
75 res_id->name[0] = cpu_to_le64(resname);
76 /* XXX: unfortunately, sptlprc and config llog share one lock */
79 case CONFIG_T_SPTLRPC:
82 case CONFIG_T_RECOVER:
84 case CONFIG_T_NODEMAP:
90 res_id->name[1] = cpu_to_le64(resname);
91 CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
92 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
96 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
98 /* fsname is at most 8 chars long, maybe contain "-".
99 * e.g. "lustre", "SUN-000" */
100 return mgc_name2resid(fsname, strlen(fsname), res_id, type);
102 EXPORT_SYMBOL(mgc_fsname2resid);
104 static int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id,
110 /* logname consists of "fsname-nodetype".
111 * e.g. "lustre-MDT0001", "SUN-000-client"
112 * there is an exception: llog "params" */
113 name_end = strrchr(logname, '-');
115 len = strlen(logname);
117 len = name_end - logname;
118 return mgc_name2resid(logname, len, res_id, type);
121 /********************** config llog list **********************/
122 static struct list_head config_llog_list = LIST_HEAD_INIT(config_llog_list);
123 static DEFINE_SPINLOCK(config_list_lock);
125 /* Take a reference to a config log */
126 static int config_log_get(struct config_llog_data *cld)
129 atomic_inc(&cld->cld_refcount);
130 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
131 atomic_read(&cld->cld_refcount));
135 /* Drop a reference to a config log. When no longer referenced,
136 we can free the config log data */
137 static void config_log_put(struct config_llog_data *cld)
141 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
142 atomic_read(&cld->cld_refcount));
143 LASSERT(atomic_read(&cld->cld_refcount) > 0);
145 /* spinlock to make sure no item with 0 refcount in the list */
146 if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
147 list_del(&cld->cld_list_chain);
148 spin_unlock(&config_list_lock);
150 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
152 if (cld->cld_recover)
153 config_log_put(cld->cld_recover);
155 config_log_put(cld->cld_params);
156 if (cld->cld_nodemap)
157 config_log_put(cld->cld_nodemap);
158 if (cld->cld_sptlrpc)
159 config_log_put(cld->cld_sptlrpc);
160 if (cld_is_sptlrpc(cld))
161 sptlrpc_conf_log_stop(cld->cld_logname);
163 class_export_put(cld->cld_mgcexp);
164 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
170 /* Find a config log by name */
172 struct config_llog_data *config_log_find(char *logname,
173 struct config_llog_instance *cfg)
175 struct config_llog_data *cld;
176 struct config_llog_data *found = NULL;
180 LASSERT(logname != NULL);
182 instance = cfg ? cfg->cfg_instance : NULL;
183 spin_lock(&config_list_lock);
184 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
185 /* check if instance equals */
186 if (instance != cld->cld_cfg.cfg_instance)
189 /* instance may be NULL, should check name */
190 if (strcmp(logname, cld->cld_logname) == 0) {
192 config_log_get(found);
196 spin_unlock(&config_list_lock);
201 struct config_llog_data *do_config_log_add(struct obd_device *obd,
204 struct config_llog_instance *cfg,
205 struct super_block *sb)
207 struct config_llog_data *cld;
212 CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
213 cfg ? cfg->cfg_instance : NULL);
215 OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
217 RETURN(ERR_PTR(-ENOMEM));
219 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
221 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
225 strcpy(cld->cld_logname, logname);
229 cld->cld_cfg.cfg_callback = class_config_llog_handler;
230 mutex_init(&cld->cld_lock);
231 cld->cld_cfg.cfg_last_idx = 0;
232 cld->cld_cfg.cfg_flags = 0;
233 cld->cld_cfg.cfg_sb = sb;
234 cld->cld_type = type;
235 atomic_set(&cld->cld_refcount, 1);
237 /* Keep the mgc around until we are done */
238 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
240 if (cld_is_sptlrpc(cld)) {
241 sptlrpc_conf_log_start(logname);
242 cld->cld_cfg.cfg_obdname = obd->obd_name;
245 spin_lock(&config_list_lock);
246 list_add(&cld->cld_list_chain, &config_llog_list);
247 spin_unlock(&config_list_lock);
249 if (cld_is_sptlrpc(cld) || cld_is_nodemap(cld)) {
250 rc = mgc_process_log(obd, cld);
251 if (rc && rc != -ENOENT)
252 CERROR("%s: failed processing log, type %d: rc = %d\n",
253 obd->obd_name, type, rc);
259 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
261 struct config_llog_instance *cfg,
262 struct super_block *sb)
264 struct config_llog_instance lcfg = *cfg;
265 struct lustre_sb_info *lsi = s2lsi(sb);
266 struct config_llog_data *cld;
272 /* for osp-on-ost, see lustre_start_osp() */
273 if (IS_MDT(lsi) && lcfg.cfg_instance)
276 /* we have to use different llog for clients and mdts for cmd
277 * where only clients are notified if one of cmd server restarts */
278 LASSERT(strlen(fsname) < sizeof(logname) / 2);
279 strcpy(logname, fsname);
280 if (IS_SERVER(lsi)) { /* mdt */
281 LASSERT(lcfg.cfg_instance == NULL);
282 lcfg.cfg_instance = sb;
283 strcat(logname, "-mdtir");
285 LASSERT(lcfg.cfg_instance != NULL);
286 strcat(logname, "-cliir");
289 cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
293 static struct config_llog_data *config_params_log_add(struct obd_device *obd,
294 struct config_llog_instance *cfg, struct super_block *sb)
296 struct config_llog_instance lcfg = *cfg;
297 struct config_llog_data *cld;
299 lcfg.cfg_instance = sb;
301 cld = do_config_log_add(obd, PARAMS_FILENAME, CONFIG_T_PARAMS,
307 /** Add this log to the list of active logs watched by an MGC.
308 * Active means we're watching for updates.
309 * We have one active log per "mount" - client instance or servername.
310 * Each instance may be at a different point in the log.
312 static struct config_llog_data *
313 config_log_add(struct obd_device *obd, char *logname,
314 struct config_llog_instance *cfg, struct super_block *sb)
316 struct lustre_sb_info *lsi = s2lsi(sb);
317 struct config_llog_data *cld;
318 struct config_llog_data *sptlrpc_cld;
319 struct config_llog_data *params_cld;
320 struct config_llog_data *nodemap_cld;
327 CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
330 * for each regular log, the depended sptlrpc log name is
331 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
333 ptr = strrchr(logname, '-');
334 if (ptr == NULL || ptr - logname > 8) {
335 CERROR("logname %s is too long\n", logname);
336 RETURN(ERR_PTR(-EINVAL));
339 memcpy(seclogname, logname, ptr - logname);
340 strcpy(seclogname + (ptr - logname), "-sptlrpc");
342 sptlrpc_cld = config_log_find(seclogname, NULL);
343 if (sptlrpc_cld == NULL) {
344 sptlrpc_cld = do_config_log_add(obd, seclogname,
345 CONFIG_T_SPTLRPC, NULL, NULL);
346 if (IS_ERR(sptlrpc_cld)) {
347 CERROR("can't create sptlrpc log: %s\n", seclogname);
348 GOTO(out, rc = PTR_ERR(sptlrpc_cld));
352 nodemap_cld = config_log_find(LUSTRE_NODEMAP_NAME, NULL);
353 if (!nodemap_cld && IS_SERVER(lsi) && !IS_MGS(lsi)) {
354 nodemap_cld = do_config_log_add(obd, LUSTRE_NODEMAP_NAME,
355 CONFIG_T_NODEMAP, NULL, NULL);
356 if (IS_ERR(nodemap_cld)) {
357 rc = PTR_ERR(nodemap_cld);
358 CERROR("%s: cannot create nodemap log: rc = %d\n",
360 GOTO(out_sptlrpc, rc);
364 params_cld = config_params_log_add(obd, cfg, sb);
365 if (IS_ERR(params_cld)) {
366 rc = PTR_ERR(params_cld);
367 CERROR("%s: can't create params log: rc = %d\n",
369 GOTO(out_nodemap, rc);
372 cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
374 CERROR("can't create log: %s\n", logname);
375 GOTO(out_params, rc = PTR_ERR(cld));
378 LASSERT(lsi->lsi_lmd);
379 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
380 struct config_llog_data *recover_cld;
382 ptr = strrchr(seclogname, '-');
386 CERROR("%s: sptlrpc log name not correct, %s: "
387 "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
388 GOTO(out_cld, rc = -EINVAL);
391 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
392 if (IS_ERR(recover_cld))
393 GOTO(out_cld, rc = PTR_ERR(recover_cld));
395 mutex_lock(&cld->cld_lock);
397 cld->cld_recover = recover_cld;
401 mutex_lock(&cld->cld_lock);
402 cld->cld_params = params_cld;
403 cld->cld_nodemap = nodemap_cld;
404 cld->cld_sptlrpc = sptlrpc_cld;
405 mutex_unlock(&cld->cld_lock);
413 config_log_put(params_cld);
416 config_log_put(nodemap_cld);
419 config_log_put(sptlrpc_cld);
425 DEFINE_MUTEX(llog_process_lock);
427 static inline void config_mark_cld_stop(struct config_llog_data *cld)
429 mutex_lock(&cld->cld_lock);
430 spin_lock(&config_list_lock);
431 cld->cld_stopping = 1;
432 spin_unlock(&config_list_lock);
433 mutex_unlock(&cld->cld_lock);
436 /** Stop watching for updates on this log.
438 static int config_log_end(char *logname, struct config_llog_instance *cfg)
440 struct config_llog_data *cld;
441 struct config_llog_data *cld_sptlrpc = NULL;
442 struct config_llog_data *cld_params = NULL;
443 struct config_llog_data *cld_recover = NULL;
444 struct config_llog_data *cld_nodemap = NULL;
449 cld = config_log_find(logname, cfg);
453 mutex_lock(&cld->cld_lock);
455 * if cld_stopping is set, it means we didn't start the log thus
456 * not owning the start ref. this can happen after previous umount:
457 * the cld still hanging there waiting for lock cancel, and we
458 * remount again but failed in the middle and call log_end without
461 if (unlikely(cld->cld_stopping)) {
462 mutex_unlock(&cld->cld_lock);
463 /* drop the ref from the find */
468 spin_lock(&config_list_lock);
469 cld->cld_stopping = 1;
470 spin_unlock(&config_list_lock);
472 cld_recover = cld->cld_recover;
473 cld->cld_recover = NULL;
474 cld_params = cld->cld_params;
475 cld->cld_params = NULL;
476 cld_nodemap = cld->cld_nodemap;
477 cld->cld_nodemap = NULL;
478 cld_sptlrpc = cld->cld_sptlrpc;
479 cld->cld_sptlrpc = NULL;
480 mutex_unlock(&cld->cld_lock);
483 config_mark_cld_stop(cld_recover);
484 config_log_put(cld_recover);
488 config_mark_cld_stop(cld_params);
489 config_log_put(cld_params);
492 /* don't set cld_stopping on nm lock as other targets may be active */
494 config_log_put(cld_nodemap);
497 config_log_put(cld_sptlrpc);
499 /* drop the ref from the find */
501 /* drop the start ref */
504 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
509 #ifdef CONFIG_PROC_FS
510 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
512 struct obd_device *obd = data;
513 struct obd_import *imp;
514 struct obd_connect_data *ocd;
515 struct config_llog_data *cld;
518 LASSERT(obd != NULL);
519 LPROCFS_CLIMP_CHECK(obd);
520 imp = obd->u.cli.cl_import;
521 ocd = &imp->imp_connect_data;
523 seq_printf(m, "imperative_recovery: %s\n",
524 OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
525 seq_printf(m, "client_state:\n");
527 spin_lock(&config_list_lock);
528 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
529 if (cld->cld_recover == NULL)
531 seq_printf(m, " - { client: %s, nidtbl_version: %u }\n",
533 cld->cld_recover->cld_cfg.cfg_last_idx);
535 spin_unlock(&config_list_lock);
537 LPROCFS_CLIMP_EXIT(obd);
542 /* reenqueue any lost locks */
543 #define RQ_RUNNING 0x1
547 #define RQ_PRECLEANUP 0x10
548 static int rq_state = 0;
549 static wait_queue_head_t rq_waitq;
550 static DECLARE_COMPLETION(rq_exit);
551 static DECLARE_COMPLETION(rq_start);
553 static void do_requeue(struct config_llog_data *cld)
558 LASSERT(atomic_read(&cld->cld_refcount) > 0);
561 * Do not run mgc_process_log on a disconnected export or an
562 * export which is being disconnected. Take the client
563 * semaphore to make the check non-racy.
565 down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
567 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
568 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
569 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
570 if (rc && rc != -ENOENT)
571 CERROR("failed processing log: %d\n", rc);
573 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
576 up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
581 /* this timeout represents how many seconds MGC should wait before
582 * requeue config and recover lock to the MGS. We need to randomize this
583 * in order to not flood the MGS.
585 #define MGC_TIMEOUT_MIN_SECONDS 5
586 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
588 static int mgc_requeue_thread(void *data)
594 CDEBUG(D_MGC, "Starting requeue thread\n");
596 /* Keep trying failed locks periodically */
597 spin_lock(&config_list_lock);
598 rq_state |= RQ_RUNNING;
599 while (!(rq_state & RQ_STOP)) {
600 struct l_wait_info lwi;
601 struct config_llog_data *cld, *cld_prev;
602 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
605 /* Any new or requeued lostlocks will change the state */
606 rq_state &= ~(RQ_NOW | RQ_LATER);
607 spin_unlock(&config_list_lock);
614 /* Always wait a few seconds to allow the server who
615 caused the lock revocation to finish its setup, plus some
616 random so everyone doesn't try to reconnect at once. */
617 to = msecs_to_jiffies(MGC_TIMEOUT_MIN_SECONDS * MSEC_PER_SEC);
618 /* rand is centi-seconds */
619 to += msecs_to_jiffies(rand * MSEC_PER_SEC / 100);
620 lwi = LWI_TIMEOUT(to, NULL, NULL);
621 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
625 * iterate & processing through the list. for each cld, process
626 * its depending sptlrpc cld firstly (if any) and then itself.
628 * it's guaranteed any item in the list must have
629 * reference > 0; and if cld_lostlock is set, at
630 * least one reference is taken by the previous enqueue.
634 spin_lock(&config_list_lock);
635 rq_state &= ~RQ_PRECLEANUP;
636 list_for_each_entry(cld, &config_llog_list,
638 if (!cld->cld_lostlock || cld->cld_stopping)
641 /* hold reference to avoid being freed during
642 * subsequent processing. */
644 cld->cld_lostlock = 0;
645 spin_unlock(&config_list_lock);
648 config_log_put(cld_prev);
651 if (likely(!(rq_state & RQ_STOP))) {
653 spin_lock(&config_list_lock);
655 spin_lock(&config_list_lock);
659 spin_unlock(&config_list_lock);
661 config_log_put(cld_prev);
663 /* Wait a bit to see if anyone else needs a requeue */
664 lwi = (struct l_wait_info) { 0 };
665 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
667 spin_lock(&config_list_lock);
670 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
671 rq_state &= ~RQ_RUNNING;
672 spin_unlock(&config_list_lock);
676 CDEBUG(D_MGC, "Ending requeue thread\n");
680 /* Add a cld to the list to requeue. Start the requeue thread if needed.
681 We are responsible for dropping the config log reference from here on out. */
682 static void mgc_requeue_add(struct config_llog_data *cld)
687 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
688 cld->cld_logname, atomic_read(&cld->cld_refcount),
689 cld->cld_stopping, rq_state);
690 LASSERT(atomic_read(&cld->cld_refcount) > 0);
692 mutex_lock(&cld->cld_lock);
693 spin_lock(&config_list_lock);
694 if (!(rq_state & RQ_STOP) && !cld->cld_stopping && !cld->cld_lostlock) {
695 cld->cld_lostlock = 1;
699 spin_unlock(&config_list_lock);
700 mutex_unlock(&cld->cld_lock);
707 /********************** class fns **********************/
708 static int mgc_local_llog_init(const struct lu_env *env,
709 struct obd_device *obd,
710 struct obd_device *disk)
712 struct llog_ctxt *ctxt;
717 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
722 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
724 ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
730 static int mgc_local_llog_fini(const struct lu_env *env,
731 struct obd_device *obd)
733 struct llog_ctxt *ctxt;
737 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
738 llog_cleanup(env, ctxt);
743 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
744 struct super_block *sb)
746 struct lustre_sb_info *lsi = s2lsi(sb);
747 struct client_obd *cli = &obd->u.cli;
748 struct lu_fid rfid, fid;
749 struct dt_object *root, *dto;
755 LASSERT(lsi->lsi_dt_dev);
757 /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
758 mutex_lock(&cli->cl_mgc_mutex);
760 /* Setup the configs dir */
761 fid.f_seq = FID_SEQ_LOCAL_NAME;
764 rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
769 rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
773 root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
774 &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
775 if (unlikely(IS_ERR(root)))
776 GOTO(out_los, rc = PTR_ERR(root));
778 dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
780 S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
781 lu_object_put_nocache(env, &root->do_lu);
783 GOTO(out_los, rc = PTR_ERR(dto));
785 cli->cl_mgc_configs_dir = dto;
787 LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
788 rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
792 /* We take an obd ref to insure that we can't get to mgc_cleanup
793 * without calling mgc_fs_cleanup first. */
794 class_incref(obd, "mgc_fs", obd);
796 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
800 lu_object_put(env, &cli->cl_mgc_configs_dir->do_lu);
801 cli->cl_mgc_configs_dir = NULL;
805 local_oid_storage_fini(env, cli->cl_mgc_los);
806 cli->cl_mgc_los = NULL;
807 mutex_unlock(&cli->cl_mgc_mutex);
812 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
814 struct client_obd *cli = &obd->u.cli;
817 LASSERT(cli->cl_mgc_los != NULL);
819 mgc_local_llog_fini(env, obd);
821 lu_object_put_nocache(env, &cli->cl_mgc_configs_dir->do_lu);
822 cli->cl_mgc_configs_dir = NULL;
824 local_oid_storage_fini(env, cli->cl_mgc_los);
825 cli->cl_mgc_los = NULL;
827 class_decref(obd, "mgc_fs", obd);
828 mutex_unlock(&cli->cl_mgc_mutex);
833 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
835 struct llog_ctxt *ctxt;
840 /* setup only remote ctxt, the local disk context is switched per each
841 * filesystem during mgc_fs_setup() */
842 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
847 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
850 llog_initiator_connect(ctxt);
856 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
858 struct llog_ctxt *ctxt;
862 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
864 llog_cleanup(env, ctxt);
870 static atomic_t mgc_count = ATOMIC_INIT(0);
871 static int mgc_precleanup(struct obd_device *obd)
877 if (atomic_dec_and_test(&mgc_count)) {
878 LASSERT(rq_state & RQ_RUNNING);
879 /* stop requeue thread */
882 /* wakeup requeue thread to clean our cld */
883 temp = RQ_NOW | RQ_PRECLEANUP;
886 spin_lock(&config_list_lock);
888 spin_unlock(&config_list_lock);
892 wait_for_completion(&rq_exit);
893 obd_cleanup_client_import(obd);
895 rc = mgc_llog_fini(NULL, obd);
897 CERROR("failed to cleanup llogging subsystems\n");
902 static int mgc_cleanup(struct obd_device *obd)
907 /* COMPAT_146 - old config logs may have added profiles we don't
909 if (obd->obd_type->typ_refcnt <= 1)
910 /* Only for the last mgc */
911 class_del_profiles();
913 lprocfs_obd_cleanup(obd);
916 rc = client_obd_cleanup(obd);
920 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
922 struct task_struct *task;
926 rc = ptlrpcd_addref();
930 rc = client_obd_setup(obd, lcfg);
932 GOTO(err_decref, rc);
934 rc = mgc_llog_init(NULL, obd);
936 CERROR("failed to setup llogging subsystems\n");
937 GOTO(err_cleanup, rc);
940 #ifdef CONFIG_PROC_FS
941 obd->obd_vars = lprocfs_mgc_obd_vars;
942 lprocfs_obd_setup(obd);
944 sptlrpc_lprocfs_cliobd_attach(obd);
946 if (atomic_inc_return(&mgc_count) == 1) {
948 init_waitqueue_head(&rq_waitq);
950 /* start requeue thread */
951 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
954 CERROR("%s: cannot start requeue thread: rc = %d; "
955 "no more log updates\n",
957 GOTO(err_cleanup, rc);
959 /* rc is the task_struct pointer of mgc_requeue_thread. */
961 wait_for_completion(&rq_start);
967 client_obd_cleanup(obd);
973 /* based on ll_mdc_blocking_ast */
974 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
975 void *data, int flag)
977 struct lustre_handle lockh;
978 struct config_llog_data *cld = (struct config_llog_data *)data;
983 case LDLM_CB_BLOCKING:
984 /* mgs wants the lock, give it up... */
985 LDLM_DEBUG(lock, "MGC blocking CB");
986 ldlm_lock2handle(lock, &lockh);
987 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
989 case LDLM_CB_CANCELING:
990 /* We've given up the lock, prepare ourselves to update. */
991 LDLM_DEBUG(lock, "MGC cancel CB");
993 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
994 PLDLMRES(lock->l_resource),
995 (char *)&lock->l_resource->lr_name.name[0]);
998 CDEBUG(D_INFO, "missing data, won't requeue\n");
1002 /* held at mgc_process_log(). */
1003 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1005 lock->l_ast_data = NULL;
1006 /* Are we done with this log? */
1007 if (cld->cld_stopping) {
1008 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1010 config_log_put(cld);
1013 /* Make sure not to re-enqueue when the mgc is stopping
1014 (we get called from client_disconnect_export) */
1015 if (lock->l_conn_export == NULL ||
1016 lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1017 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1019 config_log_put(cld);
1023 /* Re-enqueue now */
1024 mgc_requeue_add(cld);
1025 config_log_put(cld);
1034 /* Not sure where this should go... */
1035 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1036 * that we can have a chance to try the secondary MGS if any. */
1037 #define MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1039 #define MGC_TARGET_REG_LIMIT 10
1040 #define MGC_SEND_PARAM_LIMIT 10
1042 /* Send parameter to MGS*/
1043 static int mgc_set_mgs_param(struct obd_export *exp,
1044 struct mgs_send_param *msp)
1046 struct ptlrpc_request *req;
1047 struct mgs_send_param *req_msp, *rep_msp;
1051 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1052 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1057 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1059 ptlrpc_req_finished(req);
1063 memcpy(req_msp, msp, sizeof(*req_msp));
1064 ptlrpc_request_set_replen(req);
1066 /* Limit how long we will wait for the enqueue to complete */
1067 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1068 rc = ptlrpc_queue_wait(req);
1070 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1071 memcpy(msp, rep_msp, sizeof(*rep_msp));
1074 ptlrpc_req_finished(req);
1079 /* Take a config lock so we can get cancel notifications */
1080 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1081 union ldlm_policy_data *policy, enum ldlm_mode mode,
1082 __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
1083 void *data, __u32 lvb_len, void *lvb_swabber,
1084 struct lustre_handle *lockh)
1086 struct config_llog_data *cld = (struct config_llog_data *)data;
1087 struct ldlm_enqueue_info einfo = {
1090 .ei_cb_bl = mgc_blocking_ast,
1091 .ei_cb_cp = ldlm_completion_ast,
1093 struct ptlrpc_request *req;
1094 int short_limit = cld_is_sptlrpc(cld);
1098 CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1099 cld->cld_resid.name[0]);
1101 /* We need a callback for every lockholder, so don't try to
1102 ldlm_lock_match (see rev 1.1.2.11.2.47) */
1103 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1104 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1109 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1110 ptlrpc_request_set_replen(req);
1112 /* check if this is server or client */
1113 if (cld->cld_cfg.cfg_sb) {
1114 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1115 if (lsi && IS_SERVER(lsi))
1118 /* Limit how long we will wait for the enqueue to complete */
1119 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1120 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1121 NULL, 0, LVB_T_NONE, lockh, 0);
1122 /* A failed enqueue should still call the mgc_blocking_ast,
1123 where it will be requeued if needed ("grant failed"). */
1124 ptlrpc_req_finished(req);
1128 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1129 struct lustre_handle *lockh)
1133 ldlm_lock_decref(lockh, mode);
1138 static void mgc_notify_active(struct obd_device *unused)
1140 /* wakeup mgc_requeue_thread to requeue mgc lock */
1141 spin_lock(&config_list_lock);
1143 spin_unlock(&config_list_lock);
1146 /* TODO: Help the MGS rebuild nidtbl. -jay */
1149 /* Send target_reg message to MGS */
1150 static int mgc_target_register(struct obd_export *exp,
1151 struct mgs_target_info *mti)
1153 struct ptlrpc_request *req;
1154 struct mgs_target_info *req_mti, *rep_mti;
1158 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1159 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1164 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1166 ptlrpc_req_finished(req);
1170 memcpy(req_mti, mti, sizeof(*req_mti));
1171 ptlrpc_request_set_replen(req);
1172 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1173 /* Limit how long we will wait for the enqueue to complete */
1174 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1176 rc = ptlrpc_queue_wait(req);
1178 rep_mti = req_capsule_server_get(&req->rq_pill,
1179 &RMF_MGS_TARGET_INFO);
1180 memcpy(mti, rep_mti, sizeof(*rep_mti));
1181 CDEBUG(D_MGC, "register %s got index = %d\n",
1182 mti->mti_svname, mti->mti_stripe_index);
1184 ptlrpc_req_finished(req);
1189 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1190 u32 keylen, void *key,
1191 u32 vallen, void *val,
1192 struct ptlrpc_request_set *set)
1197 /* Turn off initial_recov after we try all backup servers once */
1198 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1199 struct obd_import *imp = class_exp2cliimp(exp);
1201 if (vallen != sizeof(int))
1203 value = *(int *)val;
1204 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1205 imp->imp_obd->obd_name, value,
1206 imp->imp_deactive, imp->imp_invalid,
1207 imp->imp_replayable, imp->imp_obd->obd_replayable,
1208 ptlrpc_import_state_name(imp->imp_state));
1209 /* Resurrect if we previously died */
1210 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1211 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1212 ptlrpc_reconnect_import(imp);
1215 /* FIXME move this to mgc_process_config */
1216 if (KEY_IS(KEY_REGISTER_TARGET)) {
1217 struct mgs_target_info *mti;
1218 if (vallen != sizeof(struct mgs_target_info))
1220 mti = (struct mgs_target_info *)val;
1221 CDEBUG(D_MGC, "register_target %s %#x\n",
1222 mti->mti_svname, mti->mti_flags);
1223 rc = mgc_target_register(exp, mti);
1226 if (KEY_IS(KEY_SET_FS)) {
1227 struct super_block *sb = (struct super_block *)val;
1229 if (vallen != sizeof(struct super_block))
1232 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1235 if (KEY_IS(KEY_CLEAR_FS)) {
1238 rc = mgc_fs_cleanup(env, exp->exp_obd);
1241 if (KEY_IS(KEY_SET_INFO)) {
1242 struct mgs_send_param *msp;
1244 msp = (struct mgs_send_param *)val;
1245 rc = mgc_set_mgs_param(exp, msp);
1248 if (KEY_IS(KEY_MGSSEC)) {
1249 struct client_obd *cli = &exp->exp_obd->u.cli;
1250 struct sptlrpc_flavor flvr;
1253 * empty string means using current flavor, if which haven't
1254 * been set yet, set it as null.
1256 * if flavor has been set previously, check the asking flavor
1257 * must match the existing one.
1260 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1266 rc = sptlrpc_parse_flavor(val, &flvr);
1268 CERROR("invalid sptlrpc flavor %s to MGS\n",
1274 * caller already hold a mutex
1276 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1277 cli->cl_flvr_mgc = flvr;
1278 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1279 sizeof(flvr)) != 0) {
1282 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1284 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1285 "currently %s is in use\n",
1295 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1296 __u32 keylen, void *key, __u32 *vallen, void *val)
1300 if (KEY_IS(KEY_CONN_DATA)) {
1301 struct obd_import *imp = class_exp2cliimp(exp);
1302 struct obd_connect_data *data = val;
1304 if (*vallen == sizeof(*data)) {
1305 *data = imp->imp_connect_data;
1313 static int mgc_import_event(struct obd_device *obd,
1314 struct obd_import *imp,
1315 enum obd_import_event event)
1319 LASSERT(imp->imp_obd == obd);
1320 CDEBUG(D_MGC, "import event %#x\n", event);
1323 case IMP_EVENT_DISCON:
1324 /* MGC imports should not wait for recovery */
1325 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1326 ptlrpc_pinger_ir_down();
1328 case IMP_EVENT_INACTIVE:
1330 case IMP_EVENT_INVALIDATE: {
1331 struct ldlm_namespace *ns = obd->obd_namespace;
1332 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1335 case IMP_EVENT_ACTIVE:
1336 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1337 /* Clearing obd_no_recov allows us to continue pinging */
1338 obd->obd_no_recov = 0;
1339 mgc_notify_active(obd);
1340 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1341 ptlrpc_pinger_ir_up();
1345 case IMP_EVENT_DEACTIVATE:
1346 case IMP_EVENT_ACTIVATE:
1349 CERROR("Unknown import event %#x\n", event);
1356 CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_CACHE_SHIFT),
1357 CONFIG_READ_NRPAGES = 4
1360 static int mgc_apply_recover_logs(struct obd_device *mgc,
1361 struct config_llog_data *cld,
1363 void *data, int datalen, bool mne_swab)
1365 struct config_llog_instance *cfg = &cld->cld_cfg;
1366 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1367 struct mgs_nidtbl_entry *entry;
1368 struct lustre_cfg *lcfg;
1369 struct lustre_cfg_bufs bufs;
1370 u64 prev_version = 0;
1379 LASSERT(cfg->cfg_instance != NULL);
1380 LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1382 OBD_ALLOC(inst, PAGE_CACHE_SIZE);
1386 if (!IS_SERVER(lsi)) {
1387 pos = snprintf(inst, PAGE_CACHE_SIZE, "%p", cfg->cfg_instance);
1388 if (pos >= PAGE_CACHE_SIZE) {
1389 OBD_FREE(inst, PAGE_CACHE_SIZE);
1393 LASSERT(IS_MDT(lsi));
1394 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1397 OBD_FREE(inst, PAGE_CACHE_SIZE);
1405 bufsz = PAGE_CACHE_SIZE - pos;
1407 while (datalen > 0) {
1408 int entry_len = sizeof(*entry);
1410 struct obd_device *obd;
1417 if (datalen < sizeof(*entry))
1420 entry = (typeof(entry))(data + off);
1423 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1425 if (entry->mne_nid_count == 0) /* at least one nid entry */
1427 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1430 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1431 if (datalen < entry_len) /* must have entry_len at least */
1434 /* Keep this swab for normal mixed endian handling. LU-1644 */
1436 lustre_swab_mgs_nidtbl_entry(entry);
1437 if (entry->mne_length > PAGE_CACHE_SIZE) {
1438 CERROR("MNE too large (%u)\n", entry->mne_length);
1442 if (entry->mne_length < entry_len)
1445 off += entry->mne_length;
1446 datalen -= entry->mne_length;
1450 if (entry->mne_version > max_version) {
1451 CERROR("entry index(%lld) is over max_index(%lld)\n",
1452 entry->mne_version, max_version);
1456 if (prev_version >= entry->mne_version) {
1457 CERROR("index unsorted, prev %lld, now %lld\n",
1458 prev_version, entry->mne_version);
1461 prev_version = entry->mne_version;
1464 * Write a string with format "nid::instance" to
1465 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1468 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1469 memset(buf, 0, bufsz);
1473 /* lustre-OST0001-osc-<instance #> */
1474 strcpy(obdname, cld->cld_logname);
1475 cname = strrchr(obdname, '-');
1476 if (cname == NULL) {
1477 CERROR("mgc %s: invalid logname %s\n",
1478 mgc->obd_name, obdname);
1482 pos = cname - obdname;
1484 pos += sprintf(obdname + pos, "-%s%04x",
1485 is_ost ? "OST" : "MDT", entry->mne_index);
1487 cname = is_ost ? "osc" : "mdc",
1488 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1489 lustre_cfg_bufs_reset(&bufs, obdname);
1491 /* find the obd by obdname */
1492 obd = class_name2obd(obdname);
1494 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1495 mgc->obd_name, obdname);
1497 /* this is a safe race, when the ost is starting up...*/
1501 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1504 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1507 down_read(&obd->u.cli.cl_sem);
1508 if (obd->u.cli.cl_import == NULL) {
1509 /* client does not connect to the OST yet */
1510 up_read(&obd->u.cli.cl_sem);
1515 /* iterate all nids to find one */
1516 /* find uuid by nid */
1518 for (i = 0; i < entry->mne_nid_count; i++) {
1519 rc = client_import_find_conn(obd->u.cli.cl_import,
1521 (struct obd_uuid *)uuid);
1526 up_read(&obd->u.cli.cl_sem);
1528 CERROR("mgc: cannot find uuid by nid %s\n",
1529 libcfs_nid2str(entry->u.nids[0]));
1533 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1534 uuid, libcfs_nid2str(entry->u.nids[0]));
1536 pos += strlen(uuid);
1537 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1538 LASSERT(pos < bufsz);
1540 lustre_cfg_bufs_set_string(&bufs, 1, params);
1542 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1548 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1549 prev_version, max_version, obdname, params);
1551 rc = class_process_config(lcfg);
1552 lustre_cfg_free(lcfg);
1554 CDEBUG(D_INFO, "process config for %s error %d\n",
1557 /* continue, even one with error */
1560 OBD_FREE(inst, PAGE_CACHE_SIZE);
1565 * This function is called if this client was notified for target restarting
1566 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1569 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1570 struct config_llog_data *cld)
1572 struct ptlrpc_connection *mgc_conn;
1573 struct ptlrpc_request *req = NULL;
1574 struct config_llog_instance *cfg = &cld->cld_cfg;
1575 struct mgs_config_body *body;
1576 struct mgs_config_res *res;
1577 struct nodemap_config *new_config = NULL;
1578 struct lu_nodemap *recent_nodemap = NULL;
1579 struct ptlrpc_bulk_desc *desc;
1580 struct page **pages = NULL;
1581 __u64 config_read_offset = 0;
1584 bool mne_swab = false;
1590 mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1592 /* don't need to get local config */
1593 if (cld_is_nodemap(cld) &&
1594 (LNET_NETTYP(LNET_NIDNET(mgc_conn->c_peer.nid)) == LOLND))
1597 /* allocate buffer for bulk transfer.
1598 * if this is the first time for this mgs to read logs,
1599 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1600 * once; otherwise, it only reads increment of logs, this should be
1601 * small and CONFIG_READ_NRPAGES will be used.
1603 nrpages = CONFIG_READ_NRPAGES;
1604 if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1605 nrpages = CONFIG_READ_NRPAGES_INIT;
1607 OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1609 GOTO(out, rc = -ENOMEM);
1611 for (i = 0; i < nrpages; i++) {
1612 pages[i] = alloc_page(GFP_KERNEL);
1613 if (pages[i] == NULL)
1614 GOTO(out, rc = -ENOMEM);
1618 #ifdef HAVE_SERVER_SUPPORT
1619 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1620 new_config = nodemap_config_alloc();
1621 if (IS_ERR(new_config)) {
1622 rc = PTR_ERR(new_config);
1628 LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1629 LASSERT(mutex_is_locked(&cld->cld_lock));
1630 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1631 &RQF_MGS_CONFIG_READ);
1633 GOTO(out, rc = -ENOMEM);
1635 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1640 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1641 LASSERT(body != NULL);
1642 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1643 if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1644 >= sizeof(body->mcb_name))
1645 GOTO(out, rc = -E2BIG);
1646 if (cld_is_nodemap(cld))
1647 body->mcb_offset = config_read_offset;
1649 body->mcb_offset = cfg->cfg_last_idx + 1;
1650 body->mcb_type = cld->cld_type;
1651 body->mcb_bits = PAGE_CACHE_SHIFT;
1652 body->mcb_units = nrpages;
1654 /* allocate bulk transfer descriptor */
1655 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1656 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1658 &ptlrpc_bulk_kiov_pin_ops);
1660 GOTO(out, rc = -ENOMEM);
1662 for (i = 0; i < nrpages; i++)
1663 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1666 ptlrpc_request_set_replen(req);
1667 rc = ptlrpc_queue_wait(req);
1671 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1673 GOTO(out, rc = -EPROTO);
1675 if (cld_is_nodemap(cld)) {
1676 config_read_offset = res->mcr_offset;
1677 eof = config_read_offset == II_END_OFF;
1679 if (res->mcr_size < res->mcr_offset)
1680 GOTO(out, rc = -EINVAL);
1682 /* always update the index even though it might have errors with
1683 * handling the recover logs
1685 cfg->cfg_last_idx = res->mcr_offset;
1686 eof = res->mcr_offset == res->mcr_size;
1688 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1689 res->mcr_offset, eof == false);
1692 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1694 GOTO(out, rc = ealen);
1696 if (ealen > nrpages << PAGE_CACHE_SHIFT)
1697 GOTO(out, rc = -EINVAL);
1699 if (ealen == 0) { /* no logs transferred */
1700 #ifdef HAVE_SERVER_SUPPORT
1701 /* config changed since first read RPC */
1702 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1703 recent_nodemap = NULL;
1704 nodemap_config_dealloc(new_config);
1707 CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1709 /* setting eof to false, we request config again */
1719 mne_swab = !!ptlrpc_rep_need_swab(req);
1720 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1721 /* This import flag means the server did an extra swab of IR MNE
1722 * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1723 if (unlikely(req->rq_import->imp_need_mne_swab))
1724 mne_swab = !mne_swab;
1727 /* When a nodemap config is received, we build a new nodemap config,
1728 * with new nodemap structs. We keep track of the most recently added
1729 * nodemap since the config is read ordered by nodemap_id, and so it
1730 * is likely that the next record will be related. Because access to
1731 * the nodemaps is single threaded until the nodemap_config is active,
1732 * we don't need to reference count with recent_nodemap, though
1733 * recent_nodemap should be set to NULL when the nodemap_config
1734 * is either destroyed or set active.
1736 for (i = 0; i < nrpages && ealen > 0; i++) {
1740 ptr = kmap(pages[i]);
1741 if (cld_is_nodemap(cld))
1742 rc2 = nodemap_process_idx_pages(new_config, ptr,
1745 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1752 CWARN("%s: error processing %s log %s: rc = %d\n",
1754 cld_is_nodemap(cld) ? "nodemap" : "recovery",
1760 ealen -= PAGE_CACHE_SIZE;
1765 ptlrpc_req_finished(req);
1769 if (rc == 0 && !eof)
1772 #ifdef HAVE_SERVER_SUPPORT
1773 if (new_config != NULL) {
1774 /* recent_nodemap cannot be used after set_active/dealloc */
1776 nodemap_config_set_active_mgc(new_config);
1778 nodemap_config_dealloc(new_config);
1783 for (i = 0; i < nrpages; i++) {
1784 if (pages[i] == NULL)
1786 __free_page(pages[i]);
1788 OBD_FREE(pages, sizeof(*pages) * nrpages);
1793 /* Copy a remote log locally */
1794 static int mgc_llog_local_copy(const struct lu_env *env,
1795 struct obd_device *obd,
1796 struct llog_ctxt *rctxt,
1797 struct llog_ctxt *lctxt, char *logname)
1805 * - copy it to backup using llog_backup()
1806 * - copy remote llog to logname using llog_backup()
1807 * - if failed then move bakup to logname again
1810 OBD_ALLOC(temp_log, strlen(logname) + 2);
1813 sprintf(temp_log, "%sT", logname);
1815 /* make a copy of local llog at first */
1816 rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1817 if (rc < 0 && rc != -ENOENT)
1819 /* copy remote llog to the local copy */
1820 rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1821 if (rc == -ENOENT) {
1822 /* no remote llog, delete local one too */
1823 llog_erase(env, lctxt, NULL, logname);
1824 } else if (rc < 0) {
1825 /* error during backup, get local one back from the copy */
1826 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1828 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1829 obd->obd_name, logname, rc);
1831 llog_erase(env, lctxt, NULL, temp_log);
1832 OBD_FREE(temp_log, strlen(logname) + 2);
1836 /* local_only means it cannot get remote llogs */
1837 static int mgc_process_cfg_log(struct obd_device *mgc,
1838 struct config_llog_data *cld, int local_only)
1840 struct llog_ctxt *ctxt, *lctxt = NULL;
1841 struct client_obd *cli = &mgc->u.cli;
1842 struct lustre_sb_info *lsi = NULL;
1844 bool sptlrpc_started = false;
1850 LASSERT(mutex_is_locked(&cld->cld_lock));
1853 * local copy of sptlrpc log is controlled elsewhere, don't try to
1856 if (cld_is_sptlrpc(cld) && local_only)
1859 if (cld->cld_cfg.cfg_sb)
1860 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1866 rc = lu_env_init(env, LCT_MG_THREAD);
1870 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1873 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1875 /* Copy the setup log locally if we can. Don't mess around if we're
1876 * running an MGS though (logs are already local). */
1877 if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1878 cli->cl_mgc_configs_dir != NULL &&
1879 lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1882 /* Only try to copy log if we have the lock. */
1883 rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1885 if (local_only || rc) {
1886 if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1887 llog_is_empty(env, lctxt, cld->cld_logname)) {
1888 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1889 "log %s and no local copy."
1890 "\n", cld->cld_logname);
1891 GOTO(out_pop, rc = -ENOENT);
1893 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1894 "copy for now, will try to update later.\n",
1898 /* Now, whether we copied or not, start using the local llog.
1899 * If we failed to copy, we'll start using whatever the old
1901 llog_ctxt_put(ctxt);
1905 if (local_only) /* no local log at client side */
1906 GOTO(out_pop, rc = -EIO);
1909 if (cld_is_sptlrpc(cld)) {
1910 sptlrpc_conf_log_update_begin(cld->cld_logname);
1911 sptlrpc_started = true;
1914 /* logname and instance info should be the same, so use our
1915 * copy of the instance for the update. The cfg_last_idx will
1916 * be updated here. */
1917 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1922 __llog_ctxt_put(env, ctxt);
1924 __llog_ctxt_put(env, lctxt);
1927 * update settings on existing OBDs. doing it inside
1928 * of llog_process_lock so no device is attaching/detaching
1930 * the logname must be <fsname>-sptlrpc
1932 if (sptlrpc_started) {
1933 LASSERT(cld_is_sptlrpc(cld));
1934 sptlrpc_conf_log_update_end(cld->cld_logname);
1935 class_notify_sptlrpc_conf(cld->cld_logname,
1936 strlen(cld->cld_logname) -
1937 strlen("-sptlrpc"));
1946 static bool mgc_import_in_recovery(struct obd_import *imp)
1948 bool in_recovery = true;
1950 spin_lock(&imp->imp_lock);
1951 if (imp->imp_state == LUSTRE_IMP_FULL ||
1952 imp->imp_state == LUSTRE_IMP_CLOSED)
1953 in_recovery = false;
1954 spin_unlock(&imp->imp_lock);
1960 * Get a configuration log from the MGS and process it.
1962 * This function is called for both clients and servers to process the
1963 * configuration log from the MGS. The MGC enqueues a DLM lock on the
1964 * log from the MGS, and if the lock gets revoked the MGC will be notified
1965 * by the lock cancellation callback that the config log has changed,
1966 * and will enqueue another MGS lock on it, and then continue processing
1967 * the new additions to the end of the log.
1969 * Since the MGC import is not replayable, if the import is being evicted
1970 * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
1971 * the log until recovery is finished or the import is closed.
1973 * Make a local copy of the log before parsing it if appropriate (non-MGS
1974 * server) so that the server can start even when the MGS is down.
1976 * There shouldn't be multiple processes running process_log at once --
1977 * sounds like badness. It actually might be fine, as long as they're not
1978 * trying to update from the same log simultaneously, in which case we
1979 * should use a per-log semaphore instead of cld_lock.
1981 * \param[in] mgc MGC device by which to fetch the configuration log
1982 * \param[in] cld log processing state (stored in lock callback data)
1984 * \retval 0 on success
1985 * \retval negative errno on failure
1987 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1989 struct lustre_handle lockh = { 0 };
1990 __u64 flags = LDLM_FL_NO_LRU;
1995 LASSERT(cld != NULL);
1997 /* I don't want multiple processes running process_log at once --
1998 sounds like badness. It actually might be fine, as long as
1999 we're not trying to update from the same log
2000 simultaneously (in which case we should use a per-log sem.) */
2002 mutex_lock(&cld->cld_lock);
2003 if (cld->cld_stopping) {
2004 mutex_unlock(&cld->cld_lock);
2008 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2010 CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
2011 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2013 /* Get the cfg lock on the llog */
2014 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2015 LCK_CR, &flags, NULL, NULL, NULL,
2016 cld, 0, NULL, &lockh);
2018 /* Get the cld, it will be released in mgc_blocking_ast. */
2019 config_log_get(cld);
2020 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2023 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2025 if (rcl == -ESHUTDOWN &&
2026 atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2027 struct obd_import *imp;
2028 struct l_wait_info lwi;
2029 int secs = cfs_time_seconds(obd_timeout);
2031 mutex_unlock(&cld->cld_lock);
2032 imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2034 /* Let's force the pinger, and wait the import to be
2035 * connected, note: since mgc import is non-replayable,
2036 * and even the import state is disconnected, it does
2037 * not mean the "recovery" is stopped, so we will keep
2038 * waitting until timeout or the import state is
2040 ptlrpc_pinger_force(imp);
2042 lwi = LWI_TIMEOUT(secs, NULL, NULL);
2043 l_wait_event(imp->imp_recovery_waitq,
2044 !mgc_import_in_recovery(imp), &lwi);
2046 if (imp->imp_state == LUSTRE_IMP_FULL) {
2050 mutex_lock(&cld->cld_lock);
2051 spin_lock(&config_list_lock);
2052 cld->cld_lostlock = 1;
2053 spin_unlock(&config_list_lock);
2056 /* mark cld_lostlock so that it will requeue
2057 * after MGC becomes available. */
2058 spin_lock(&config_list_lock);
2059 cld->cld_lostlock = 1;
2060 spin_unlock(&config_list_lock);
2064 if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2066 rc = mgc_process_recover_nodemap_log(mgc, cld);
2067 else if (cld_is_nodemap(cld))
2070 if (cld_is_recover(cld) && rc) {
2072 CERROR("%s: recover log %s failed, not fatal: rc = %d\n",
2073 mgc->obd_name, cld->cld_logname, rc);
2074 spin_lock(&config_list_lock);
2075 cld->cld_lostlock = 1;
2076 spin_unlock(&config_list_lock);
2078 rc = 0; /* this is not a fatal error for recover log */
2081 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2084 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2085 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2087 mutex_unlock(&cld->cld_lock);
2089 /* Now drop the lock so MGS can revoke it */
2091 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2093 CERROR("Can't drop cfg lock: %d\n", rcl);
2100 /** Called from lustre_process_log.
2101 * LCFG_LOG_START gets the config log from the MGS, processes it to start
2102 * any services, and adds it to the list logs to watch (follow).
2104 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2106 struct lustre_cfg *lcfg = buf;
2107 struct config_llog_instance *cfg = NULL;
2112 switch(lcfg->lcfg_command) {
2113 case LCFG_LOV_ADD_OBD: {
2114 /* Overloading this cfg command: register a new target */
2115 struct mgs_target_info *mti;
2117 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2118 sizeof(struct mgs_target_info))
2119 GOTO(out, rc = -EINVAL);
2121 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2122 CDEBUG(D_MGC, "add_target %s %#x\n",
2123 mti->mti_svname, mti->mti_flags);
2124 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2127 case LCFG_LOV_DEL_OBD:
2128 /* Unregister has no meaning at the moment. */
2129 CERROR("lov_del_obd unimplemented\n");
2132 case LCFG_SPTLRPC_CONF: {
2133 rc = sptlrpc_process_config(lcfg);
2136 case LCFG_LOG_START: {
2137 struct config_llog_data *cld;
2138 struct super_block *sb;
2140 logname = lustre_cfg_string(lcfg, 1);
2141 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2142 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2144 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2147 /* We're only called through here on the initial mount */
2148 cld = config_log_add(obd, logname, cfg, sb);
2155 /* FIXME only set this for old logs! Right now this forces
2156 us to always skip the "inside markers" check */
2157 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
2159 rc = mgc_process_log(obd, cld);
2160 if (rc == 0 && cld->cld_recover != NULL) {
2161 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2162 imp_connect_data, IMP_RECOV)) {
2163 rc = mgc_process_log(obd, cld->cld_recover);
2165 struct config_llog_data *cir;
2167 mutex_lock(&cld->cld_lock);
2168 cir = cld->cld_recover;
2169 cld->cld_recover = NULL;
2170 mutex_unlock(&cld->cld_lock);
2171 config_log_put(cir);
2175 CERROR("Cannot process recover llog %d\n", rc);
2178 if (rc == 0 && cld->cld_params != NULL) {
2179 rc = mgc_process_log(obd, cld->cld_params);
2180 if (rc == -ENOENT) {
2181 CDEBUG(D_MGC, "There is no params "
2182 "config file yet\n");
2185 /* params log is optional */
2187 CERROR("%s: can't process params llog: rc = %d\n",
2193 case LCFG_LOG_END: {
2194 logname = lustre_cfg_string(lcfg, 1);
2196 if (lcfg->lcfg_bufcount >= 2)
2197 cfg = (struct config_llog_instance *)lustre_cfg_buf(
2199 rc = config_log_end(logname, cfg);
2203 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2204 GOTO(out, rc = -EINVAL);
2212 static struct obd_ops mgc_obd_ops = {
2213 .o_owner = THIS_MODULE,
2214 .o_setup = mgc_setup,
2215 .o_precleanup = mgc_precleanup,
2216 .o_cleanup = mgc_cleanup,
2217 .o_add_conn = client_import_add_conn,
2218 .o_del_conn = client_import_del_conn,
2219 .o_connect = client_connect_import,
2220 .o_disconnect = client_disconnect_export,
2221 .o_set_info_async = mgc_set_info_async,
2222 .o_get_info = mgc_get_info,
2223 .o_import_event = mgc_import_event,
2224 .o_process_config = mgc_process_config,
2227 static int __init mgc_init(void)
2229 return class_register_type(&mgc_obd_ops, NULL, true, NULL,
2230 LUSTRE_MGC_NAME, NULL);
2233 static void __exit mgc_exit(void)
2235 class_unregister_type(LUSTRE_MGC_NAME);
2238 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2239 MODULE_DESCRIPTION("Lustre Management Client");
2240 MODULE_VERSION(LUSTRE_VERSION_STRING);
2241 MODULE_LICENSE("GPL");
2243 module_init(mgc_init);
2244 module_exit(mgc_exit);