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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mgc/mgc_request.c
34 * Author: Nathan Rutman <nathan@clusterfs.com>
37 #define DEBUG_SUBSYSTEM S_MGC
38 #define D_MGC D_CONFIG /*|D_WARNING*/
40 #include <linux/module.h>
41 #include <linux/kthread.h>
43 #include <dt_object.h>
44 #include <lprocfs_status.h>
45 #include <lustre_dlm.h>
46 #include <lustre_disk.h>
47 #include <lustre_log.h>
48 #include <lustre_nodemap.h>
49 #include <lustre_swab.h>
50 #include <obd_class.h>
52 #include "mgc_internal.h"
54 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
59 if (len > sizeof(resname)) {
60 CERROR("name too long: %s\n", name);
64 CERROR("missing name: %s\n", name);
67 memcpy(&resname, name, len);
69 /* Always use the same endianness for the resid */
70 memset(res_id, 0, sizeof(*res_id));
71 res_id->name[0] = cpu_to_le64(resname);
72 /* XXX: unfortunately, sptlprc and config llog share one lock */
75 case CONFIG_T_SPTLRPC:
78 case CONFIG_T_RECOVER:
80 case CONFIG_T_NODEMAP:
86 res_id->name[1] = cpu_to_le64(resname);
87 CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
88 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
92 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
94 /* fsname is at most 8 chars long, maybe contain "-".
95 * e.g. "lustre", "SUN-000" */
96 return mgc_name2resid(fsname, strlen(fsname), res_id, type);
98 EXPORT_SYMBOL(mgc_fsname2resid);
100 static int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id,
106 /* logname consists of "fsname-nodetype".
107 * e.g. "lustre-MDT0001", "SUN-000-client"
108 * there is an exception: llog "params" */
109 name_end = strrchr(logname, '-');
111 len = strlen(logname);
113 len = name_end - logname;
114 return mgc_name2resid(logname, len, res_id, type);
117 /********************** config llog list **********************/
118 static struct list_head config_llog_list = LIST_HEAD_INIT(config_llog_list);
119 static DEFINE_SPINLOCK(config_list_lock);
121 /* Take a reference to a config log */
122 static int config_log_get(struct config_llog_data *cld)
125 atomic_inc(&cld->cld_refcount);
126 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
127 atomic_read(&cld->cld_refcount));
131 /* Drop a reference to a config log. When no longer referenced,
132 we can free the config log data */
133 static void config_log_put(struct config_llog_data *cld)
137 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
138 atomic_read(&cld->cld_refcount));
139 LASSERT(atomic_read(&cld->cld_refcount) > 0);
141 /* spinlock to make sure no item with 0 refcount in the list */
142 if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
143 list_del(&cld->cld_list_chain);
144 spin_unlock(&config_list_lock);
146 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
148 if (cld->cld_recover)
149 config_log_put(cld->cld_recover);
151 config_log_put(cld->cld_params);
152 if (cld->cld_nodemap)
153 config_log_put(cld->cld_nodemap);
154 if (cld->cld_sptlrpc)
155 config_log_put(cld->cld_sptlrpc);
156 if (cld_is_sptlrpc(cld))
157 sptlrpc_conf_log_stop(cld->cld_logname);
159 class_export_put(cld->cld_mgcexp);
160 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
166 /* Find a config log by name */
168 struct config_llog_data *config_log_find(char *logname,
169 struct config_llog_instance *cfg)
171 struct config_llog_data *cld;
172 struct config_llog_data *found = NULL;
176 LASSERT(logname != NULL);
178 instance = cfg ? cfg->cfg_instance : NULL;
179 spin_lock(&config_list_lock);
180 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
181 /* check if instance equals */
182 if (instance != cld->cld_cfg.cfg_instance)
185 /* instance may be NULL, should check name */
186 if (strcmp(logname, cld->cld_logname) == 0) {
188 config_log_get(found);
192 spin_unlock(&config_list_lock);
197 struct config_llog_data *do_config_log_add(struct obd_device *obd,
200 struct config_llog_instance *cfg,
201 struct super_block *sb)
203 struct config_llog_data *cld;
208 CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
209 cfg ? cfg->cfg_instance : NULL);
211 OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
213 RETURN(ERR_PTR(-ENOMEM));
215 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
217 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
221 strcpy(cld->cld_logname, logname);
225 cld->cld_cfg.cfg_callback = class_config_llog_handler;
226 mutex_init(&cld->cld_lock);
227 cld->cld_cfg.cfg_last_idx = 0;
228 cld->cld_cfg.cfg_flags = 0;
229 cld->cld_cfg.cfg_sb = sb;
230 cld->cld_type = type;
231 atomic_set(&cld->cld_refcount, 1);
233 /* Keep the mgc around until we are done */
234 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
236 if (cld_is_sptlrpc(cld)) {
237 sptlrpc_conf_log_start(logname);
238 cld->cld_cfg.cfg_obdname = obd->obd_name;
241 spin_lock(&config_list_lock);
242 list_add(&cld->cld_list_chain, &config_llog_list);
243 spin_unlock(&config_list_lock);
245 if (cld_is_sptlrpc(cld) || cld_is_nodemap(cld)) {
246 rc = mgc_process_log(obd, cld);
247 if (rc && rc != -ENOENT)
248 CERROR("%s: failed processing log, type %d: rc = %d\n",
249 obd->obd_name, type, rc);
255 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
257 struct config_llog_instance *cfg,
258 struct super_block *sb)
260 struct config_llog_instance lcfg = *cfg;
261 struct lustre_sb_info *lsi = s2lsi(sb);
262 struct config_llog_data *cld;
268 /* for osp-on-ost, see lustre_start_osp() */
269 if (IS_MDT(lsi) && lcfg.cfg_instance)
272 /* we have to use different llog for clients and mdts for cmd
273 * where only clients are notified if one of cmd server restarts */
274 LASSERT(strlen(fsname) < sizeof(logname) / 2);
275 strcpy(logname, fsname);
276 if (IS_SERVER(lsi)) { /* mdt */
277 LASSERT(lcfg.cfg_instance == NULL);
278 lcfg.cfg_instance = sb;
279 strcat(logname, "-mdtir");
281 LASSERT(lcfg.cfg_instance != NULL);
282 strcat(logname, "-cliir");
285 cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
289 static struct config_llog_data *config_log_find_or_add(struct obd_device *obd,
290 char *logname, struct super_block *sb, int type,
291 struct config_llog_instance *cfg)
293 struct config_llog_instance lcfg = *cfg;
294 struct config_llog_data *cld;
296 lcfg.cfg_instance = sb != NULL ? (void *)sb : (void *)obd;
297 cld = config_log_find(logname, &lcfg);
298 if (unlikely(cld != NULL))
301 return do_config_log_add(obd, logname, type, &lcfg, sb);
304 /** Add this log to the list of active logs watched by an MGC.
305 * Active means we're watching for updates.
306 * We have one active log per "mount" - client instance or servername.
307 * Each instance may be at a different point in the log.
309 static struct config_llog_data *
310 config_log_add(struct obd_device *obd, char *logname,
311 struct config_llog_instance *cfg, struct super_block *sb)
313 struct lustre_sb_info *lsi = s2lsi(sb);
314 struct config_llog_data *cld;
315 struct config_llog_data *sptlrpc_cld = NULL;
316 struct config_llog_data *params_cld = NULL;
317 struct config_llog_data *nodemap_cld = NULL;
324 CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
327 * for each regular log, the depended sptlrpc log name is
328 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
330 ptr = strrchr(logname, '-');
331 if (ptr == NULL || ptr - logname > 8) {
332 CERROR("logname %s is too long\n", logname);
333 RETURN(ERR_PTR(-EINVAL));
336 memcpy(seclogname, logname, ptr - logname);
337 strcpy(seclogname + (ptr - logname), "-sptlrpc");
339 if (cfg->cfg_sub_clds & CONFIG_T_SPTLRPC) {
340 sptlrpc_cld = config_log_find_or_add(obd, seclogname, NULL,
341 CONFIG_T_SPTLRPC, cfg);
342 if (IS_ERR(sptlrpc_cld)) {
343 CERROR("can't create sptlrpc log: %s\n", seclogname);
344 GOTO(out, rc = PTR_ERR(sptlrpc_cld));
348 if (!IS_MGS(lsi) && cfg->cfg_sub_clds & CONFIG_T_NODEMAP) {
349 nodemap_cld = config_log_find_or_add(obd, LUSTRE_NODEMAP_NAME,
350 NULL, CONFIG_T_NODEMAP,
352 if (IS_ERR(nodemap_cld)) {
353 rc = PTR_ERR(nodemap_cld);
354 CERROR("%s: cannot create nodemap log: rc = %d\n",
356 GOTO(out_sptlrpc, rc);
360 if (cfg->cfg_sub_clds & CONFIG_T_PARAMS) {
361 params_cld = config_log_find_or_add(obd, PARAMS_FILENAME, sb,
362 CONFIG_T_PARAMS, cfg);
363 if (IS_ERR(params_cld)) {
364 rc = PTR_ERR(params_cld);
365 CERROR("%s: can't create params log: rc = %d\n",
367 GOTO(out_nodemap, rc);
371 cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
373 CERROR("can't create log: %s\n", logname);
374 GOTO(out_params, rc = PTR_ERR(cld));
377 LASSERT(lsi->lsi_lmd);
378 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR) &&
379 cfg->cfg_sub_clds & CONFIG_T_RECOVER) {
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 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1043 /* Send parameter to MGS*/
1044 static int mgc_set_mgs_param(struct obd_export *exp,
1045 struct mgs_send_param *msp)
1047 struct ptlrpc_request *req;
1048 struct mgs_send_param *req_msp, *rep_msp;
1052 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1053 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1058 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1060 ptlrpc_req_finished(req);
1064 memcpy(req_msp, msp, sizeof(*req_msp));
1065 ptlrpc_request_set_replen(req);
1067 /* Limit how long we will wait for the enqueue to complete */
1068 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1069 rc = ptlrpc_queue_wait(req);
1071 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1072 memcpy(msp, rep_msp, sizeof(*rep_msp));
1075 ptlrpc_req_finished(req);
1081 /* Take a config lock so we can get cancel notifications */
1082 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1083 union ldlm_policy_data *policy, enum ldlm_mode mode,
1084 __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
1085 void *data, __u32 lvb_len, void *lvb_swabber,
1086 struct lustre_handle *lockh)
1088 struct config_llog_data *cld = (struct config_llog_data *)data;
1089 struct ldlm_enqueue_info einfo = {
1092 .ei_cb_bl = mgc_blocking_ast,
1093 .ei_cb_cp = ldlm_completion_ast,
1095 struct ptlrpc_request *req;
1096 int short_limit = cld_is_sptlrpc(cld);
1100 CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1101 cld->cld_resid.name[0]);
1103 /* We need a callback for every lockholder, so don't try to
1104 ldlm_lock_match (see rev 1.1.2.11.2.47) */
1105 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1106 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1111 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1112 ptlrpc_request_set_replen(req);
1114 /* check if this is server or client */
1115 if (cld->cld_cfg.cfg_sb) {
1116 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1117 if (lsi && IS_SERVER(lsi))
1120 /* Limit how long we will wait for the enqueue to complete */
1121 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1122 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1123 NULL, 0, LVB_T_NONE, lockh, 0);
1124 /* A failed enqueue should still call the mgc_blocking_ast,
1125 where it will be requeued if needed ("grant failed"). */
1126 ptlrpc_req_finished(req);
1130 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1131 struct lustre_handle *lockh)
1135 ldlm_lock_decref(lockh, mode);
1140 static void mgc_notify_active(struct obd_device *unused)
1142 /* wakeup mgc_requeue_thread to requeue mgc lock */
1143 spin_lock(&config_list_lock);
1145 spin_unlock(&config_list_lock);
1148 /* TODO: Help the MGS rebuild nidtbl. -jay */
1151 /* Send target_reg message to MGS */
1152 static int mgc_target_register(struct obd_export *exp,
1153 struct mgs_target_info *mti)
1155 struct ptlrpc_request *req;
1156 struct mgs_target_info *req_mti, *rep_mti;
1160 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1161 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1166 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1168 ptlrpc_req_finished(req);
1172 memcpy(req_mti, mti, sizeof(*req_mti));
1173 ptlrpc_request_set_replen(req);
1174 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1175 /* Limit how long we will wait for the enqueue to complete */
1176 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1178 rc = ptlrpc_queue_wait(req);
1180 rep_mti = req_capsule_server_get(&req->rq_pill,
1181 &RMF_MGS_TARGET_INFO);
1182 memcpy(mti, rep_mti, sizeof(*rep_mti));
1183 CDEBUG(D_MGC, "register %s got index = %d\n",
1184 mti->mti_svname, mti->mti_stripe_index);
1186 ptlrpc_req_finished(req);
1191 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1192 u32 keylen, void *key,
1193 u32 vallen, void *val,
1194 struct ptlrpc_request_set *set)
1199 /* Turn off initial_recov after we try all backup servers once */
1200 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1201 struct obd_import *imp = class_exp2cliimp(exp);
1203 if (vallen != sizeof(int))
1205 value = *(int *)val;
1206 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1207 imp->imp_obd->obd_name, value,
1208 imp->imp_deactive, imp->imp_invalid,
1209 imp->imp_replayable, imp->imp_obd->obd_replayable,
1210 ptlrpc_import_state_name(imp->imp_state));
1211 /* Resurrect if we previously died */
1212 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1213 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1214 ptlrpc_reconnect_import(imp);
1217 /* FIXME move this to mgc_process_config */
1218 if (KEY_IS(KEY_REGISTER_TARGET)) {
1219 struct mgs_target_info *mti;
1220 if (vallen != sizeof(struct mgs_target_info))
1222 mti = (struct mgs_target_info *)val;
1223 CDEBUG(D_MGC, "register_target %s %#x\n",
1224 mti->mti_svname, mti->mti_flags);
1225 rc = mgc_target_register(exp, mti);
1228 if (KEY_IS(KEY_SET_FS)) {
1229 struct super_block *sb = (struct super_block *)val;
1231 if (vallen != sizeof(struct super_block))
1234 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1237 if (KEY_IS(KEY_CLEAR_FS)) {
1240 rc = mgc_fs_cleanup(env, exp->exp_obd);
1243 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1244 if (KEY_IS(KEY_SET_INFO)) {
1245 struct mgs_send_param *msp;
1247 msp = (struct mgs_send_param *)val;
1248 rc = mgc_set_mgs_param(exp, msp);
1252 if (KEY_IS(KEY_MGSSEC)) {
1253 struct client_obd *cli = &exp->exp_obd->u.cli;
1254 struct sptlrpc_flavor flvr;
1257 * empty string means using current flavor, if which haven't
1258 * been set yet, set it as null.
1260 * if flavor has been set previously, check the asking flavor
1261 * must match the existing one.
1264 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1270 rc = sptlrpc_parse_flavor(val, &flvr);
1272 CERROR("invalid sptlrpc flavor %s to MGS\n",
1278 * caller already hold a mutex
1280 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1281 cli->cl_flvr_mgc = flvr;
1282 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1283 sizeof(flvr)) != 0) {
1286 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1288 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1289 "currently %s is in use\n",
1299 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1300 __u32 keylen, void *key, __u32 *vallen, void *val)
1304 if (KEY_IS(KEY_CONN_DATA)) {
1305 struct obd_import *imp = class_exp2cliimp(exp);
1306 struct obd_connect_data *data = val;
1308 if (*vallen == sizeof(*data)) {
1309 *data = imp->imp_connect_data;
1317 static int mgc_import_event(struct obd_device *obd,
1318 struct obd_import *imp,
1319 enum obd_import_event event)
1323 LASSERT(imp->imp_obd == obd);
1324 CDEBUG(D_MGC, "import event %#x\n", event);
1327 case IMP_EVENT_DISCON:
1328 /* MGC imports should not wait for recovery */
1329 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1330 ptlrpc_pinger_ir_down();
1332 case IMP_EVENT_INACTIVE:
1334 case IMP_EVENT_INVALIDATE: {
1335 struct ldlm_namespace *ns = obd->obd_namespace;
1336 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1339 case IMP_EVENT_ACTIVE:
1340 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1341 /* Clearing obd_no_recov allows us to continue pinging */
1342 obd->obd_no_recov = 0;
1343 mgc_notify_active(obd);
1344 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1345 ptlrpc_pinger_ir_up();
1349 case IMP_EVENT_DEACTIVATE:
1350 case IMP_EVENT_ACTIVATE:
1353 CERROR("Unknown import event %#x\n", event);
1360 CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1361 CONFIG_READ_NRPAGES = 4
1364 static int mgc_apply_recover_logs(struct obd_device *mgc,
1365 struct config_llog_data *cld,
1367 void *data, int datalen, bool mne_swab)
1369 struct config_llog_instance *cfg = &cld->cld_cfg;
1370 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1371 struct mgs_nidtbl_entry *entry;
1372 struct lustre_cfg *lcfg;
1373 struct lustre_cfg_bufs bufs;
1374 u64 prev_version = 0;
1383 LASSERT(cfg->cfg_instance != NULL);
1384 LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1386 OBD_ALLOC(inst, PAGE_SIZE);
1390 if (!IS_SERVER(lsi)) {
1391 pos = snprintf(inst, PAGE_SIZE, "%p", cfg->cfg_instance);
1392 if (pos >= PAGE_SIZE) {
1393 OBD_FREE(inst, PAGE_SIZE);
1397 LASSERT(IS_MDT(lsi));
1398 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1401 OBD_FREE(inst, PAGE_SIZE);
1409 bufsz = PAGE_SIZE - pos;
1411 while (datalen > 0) {
1412 int entry_len = sizeof(*entry);
1414 struct obd_device *obd;
1421 if (datalen < sizeof(*entry))
1424 entry = (typeof(entry))(data + off);
1427 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1429 if (entry->mne_nid_count == 0) /* at least one nid entry */
1431 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1434 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1435 if (datalen < entry_len) /* must have entry_len at least */
1438 /* Keep this swab for normal mixed endian handling. LU-1644 */
1440 lustre_swab_mgs_nidtbl_entry(entry);
1441 if (entry->mne_length > PAGE_SIZE) {
1442 CERROR("MNE too large (%u)\n", entry->mne_length);
1446 if (entry->mne_length < entry_len)
1449 off += entry->mne_length;
1450 datalen -= entry->mne_length;
1454 if (entry->mne_version > max_version) {
1455 CERROR("entry index(%lld) is over max_index(%lld)\n",
1456 entry->mne_version, max_version);
1460 if (prev_version >= entry->mne_version) {
1461 CERROR("index unsorted, prev %lld, now %lld\n",
1462 prev_version, entry->mne_version);
1465 prev_version = entry->mne_version;
1468 * Write a string with format "nid::instance" to
1469 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1472 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1473 memset(buf, 0, bufsz);
1477 /* lustre-OST0001-osc-<instance #> */
1478 strcpy(obdname, cld->cld_logname);
1479 cname = strrchr(obdname, '-');
1480 if (cname == NULL) {
1481 CERROR("mgc %s: invalid logname %s\n",
1482 mgc->obd_name, obdname);
1486 pos = cname - obdname;
1488 pos += sprintf(obdname + pos, "-%s%04x",
1489 is_ost ? "OST" : "MDT", entry->mne_index);
1491 cname = is_ost ? "osc" : "mdc",
1492 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1493 lustre_cfg_bufs_reset(&bufs, obdname);
1495 /* find the obd by obdname */
1496 obd = class_name2obd(obdname);
1498 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1499 mgc->obd_name, obdname);
1501 /* this is a safe race, when the ost is starting up...*/
1505 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1508 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1511 down_read(&obd->u.cli.cl_sem);
1512 if (obd->u.cli.cl_import == NULL) {
1513 /* client does not connect to the OST yet */
1514 up_read(&obd->u.cli.cl_sem);
1519 /* iterate all nids to find one */
1520 /* find uuid by nid */
1522 for (i = 0; i < entry->mne_nid_count; i++) {
1523 rc = client_import_find_conn(obd->u.cli.cl_import,
1525 (struct obd_uuid *)uuid);
1530 up_read(&obd->u.cli.cl_sem);
1532 CERROR("mgc: cannot find uuid by nid %s\n",
1533 libcfs_nid2str(entry->u.nids[0]));
1537 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1538 uuid, libcfs_nid2str(entry->u.nids[0]));
1540 pos += strlen(uuid);
1541 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1542 LASSERT(pos < bufsz);
1544 lustre_cfg_bufs_set_string(&bufs, 1, params);
1546 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1552 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1553 prev_version, max_version, obdname, params);
1555 rc = class_process_config(lcfg);
1556 lustre_cfg_free(lcfg);
1558 CDEBUG(D_INFO, "process config for %s error %d\n",
1561 /* continue, even one with error */
1564 OBD_FREE(inst, PAGE_SIZE);
1569 * This function is called if this client was notified for target restarting
1570 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1573 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1574 struct config_llog_data *cld)
1576 struct ptlrpc_connection *mgc_conn;
1577 struct ptlrpc_request *req = NULL;
1578 struct config_llog_instance *cfg = &cld->cld_cfg;
1579 struct mgs_config_body *body;
1580 struct mgs_config_res *res;
1581 struct nodemap_config *new_config = NULL;
1582 struct lu_nodemap *recent_nodemap = NULL;
1583 struct ptlrpc_bulk_desc *desc;
1584 struct page **pages = NULL;
1585 __u64 config_read_offset = 0;
1586 __u8 nodemap_cur_pass = 0;
1589 bool mne_swab = false;
1595 mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1597 /* don't need to get local config */
1598 if (cld_is_nodemap(cld) &&
1599 (LNET_NETTYP(LNET_NIDNET(mgc_conn->c_peer.nid)) == LOLND))
1602 /* allocate buffer for bulk transfer.
1603 * if this is the first time for this mgs to read logs,
1604 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1605 * once; otherwise, it only reads increment of logs, this should be
1606 * small and CONFIG_READ_NRPAGES will be used.
1608 nrpages = CONFIG_READ_NRPAGES;
1609 if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1610 nrpages = CONFIG_READ_NRPAGES_INIT;
1612 OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1614 GOTO(out, rc = -ENOMEM);
1616 for (i = 0; i < nrpages; i++) {
1617 pages[i] = alloc_page(GFP_KERNEL);
1618 if (pages[i] == NULL)
1619 GOTO(out, rc = -ENOMEM);
1623 #ifdef HAVE_SERVER_SUPPORT
1624 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1625 new_config = nodemap_config_alloc();
1626 if (IS_ERR(new_config)) {
1627 rc = PTR_ERR(new_config);
1633 LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1634 LASSERT(mutex_is_locked(&cld->cld_lock));
1635 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1636 &RQF_MGS_CONFIG_READ);
1638 GOTO(out, rc = -ENOMEM);
1640 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1645 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1646 LASSERT(body != NULL);
1647 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1648 if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1649 >= sizeof(body->mcb_name))
1650 GOTO(out, rc = -E2BIG);
1651 if (cld_is_nodemap(cld))
1652 body->mcb_offset = config_read_offset;
1654 body->mcb_offset = cfg->cfg_last_idx + 1;
1655 body->mcb_type = cld->cld_type;
1656 body->mcb_bits = PAGE_SHIFT;
1657 body->mcb_units = nrpages;
1658 body->mcb_nm_cur_pass = nodemap_cur_pass;
1660 /* allocate bulk transfer descriptor */
1661 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1662 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1664 &ptlrpc_bulk_kiov_pin_ops);
1666 GOTO(out, rc = -ENOMEM);
1668 for (i = 0; i < nrpages; i++)
1669 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1672 ptlrpc_request_set_replen(req);
1673 rc = ptlrpc_queue_wait(req);
1677 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1679 GOTO(out, rc = -EPROTO);
1681 if (cld_is_nodemap(cld)) {
1682 config_read_offset = res->mcr_offset;
1683 eof = config_read_offset == II_END_OFF;
1684 nodemap_cur_pass = res->mcr_nm_cur_pass;
1686 if (res->mcr_size < res->mcr_offset)
1687 GOTO(out, rc = -EINVAL);
1689 /* always update the index even though it might have errors with
1690 * handling the recover logs
1692 cfg->cfg_last_idx = res->mcr_offset;
1693 eof = res->mcr_offset == res->mcr_size;
1695 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1696 res->mcr_offset, eof == false);
1699 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1701 GOTO(out, rc = ealen);
1703 if (ealen > nrpages << PAGE_SHIFT)
1704 GOTO(out, rc = -EINVAL);
1706 if (ealen == 0) { /* no logs transferred */
1707 #ifdef HAVE_SERVER_SUPPORT
1708 /* config changed since first read RPC */
1709 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1710 recent_nodemap = NULL;
1711 nodemap_config_dealloc(new_config);
1714 CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1716 /* setting eof to false, we request config again */
1726 mne_swab = !!ptlrpc_rep_need_swab(req);
1727 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1728 /* This import flag means the server did an extra swab of IR MNE
1729 * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1730 if (unlikely(req->rq_import->imp_need_mne_swab))
1731 mne_swab = !mne_swab;
1734 /* When a nodemap config is received, we build a new nodemap config,
1735 * with new nodemap structs. We keep track of the most recently added
1736 * nodemap since the config is read ordered by nodemap_id, and so it
1737 * is likely that the next record will be related. Because access to
1738 * the nodemaps is single threaded until the nodemap_config is active,
1739 * we don't need to reference count with recent_nodemap, though
1740 * recent_nodemap should be set to NULL when the nodemap_config
1741 * is either destroyed or set active.
1743 for (i = 0; i < nrpages && ealen > 0; i++) {
1747 ptr = kmap(pages[i]);
1748 if (cld_is_nodemap(cld))
1749 rc2 = nodemap_process_idx_pages(new_config, ptr,
1752 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1759 CWARN("%s: error processing %s log %s: rc = %d\n",
1761 cld_is_nodemap(cld) ? "nodemap" : "recovery",
1764 GOTO(out, rc = rc2);
1772 ptlrpc_req_finished(req);
1776 if (rc == 0 && !eof)
1779 #ifdef HAVE_SERVER_SUPPORT
1780 if (new_config != NULL) {
1781 /* recent_nodemap cannot be used after set_active/dealloc */
1783 nodemap_config_set_active_mgc(new_config);
1785 nodemap_config_dealloc(new_config);
1790 for (i = 0; i < nrpages; i++) {
1791 if (pages[i] == NULL)
1793 __free_page(pages[i]);
1795 OBD_FREE(pages, sizeof(*pages) * nrpages);
1800 /* Copy a remote log locally */
1801 static int mgc_llog_local_copy(const struct lu_env *env,
1802 struct obd_device *obd,
1803 struct llog_ctxt *rctxt,
1804 struct llog_ctxt *lctxt, char *logname)
1812 * - copy it to backup using llog_backup()
1813 * - copy remote llog to logname using llog_backup()
1814 * - if failed then move bakup to logname again
1817 OBD_ALLOC(temp_log, strlen(logname) + 2);
1820 sprintf(temp_log, "%sT", logname);
1822 /* make a copy of local llog at first */
1823 rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1824 if (rc < 0 && rc != -ENOENT)
1826 /* copy remote llog to the local copy */
1827 rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1828 if (rc == -ENOENT) {
1829 /* no remote llog, delete local one too */
1830 llog_erase(env, lctxt, NULL, logname);
1831 } else if (rc < 0) {
1832 /* error during backup, get local one back from the copy */
1833 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1835 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1836 obd->obd_name, logname, rc);
1838 llog_erase(env, lctxt, NULL, temp_log);
1839 OBD_FREE(temp_log, strlen(logname) + 2);
1843 /* local_only means it cannot get remote llogs */
1844 static int mgc_process_cfg_log(struct obd_device *mgc,
1845 struct config_llog_data *cld, int local_only)
1847 struct llog_ctxt *ctxt, *lctxt = NULL;
1848 struct client_obd *cli = &mgc->u.cli;
1849 struct lustre_sb_info *lsi = NULL;
1851 bool sptlrpc_started = false;
1857 LASSERT(mutex_is_locked(&cld->cld_lock));
1860 * local copy of sptlrpc log is controlled elsewhere, don't try to
1863 if (cld_is_sptlrpc(cld) && local_only)
1866 if (cld->cld_cfg.cfg_sb)
1867 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1873 rc = lu_env_init(env, LCT_MG_THREAD);
1877 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1880 lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1882 /* Copy the setup log locally if we can. Don't mess around if we're
1883 * running an MGS though (logs are already local). */
1884 if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1885 cli->cl_mgc_configs_dir != NULL &&
1886 lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1889 /* Only try to copy log if we have the lock. */
1890 rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1892 if (local_only || rc) {
1893 if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1894 llog_is_empty(env, lctxt, cld->cld_logname)) {
1895 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1896 "log %s and no local copy."
1897 "\n", cld->cld_logname);
1898 GOTO(out_pop, rc = -ENOENT);
1900 CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1901 "copy for now, will try to update later.\n",
1905 /* Now, whether we copied or not, start using the local llog.
1906 * If we failed to copy, we'll start using whatever the old
1908 llog_ctxt_put(ctxt);
1912 if (local_only) /* no local log at client side */
1913 GOTO(out_pop, rc = -EIO);
1916 if (cld_is_sptlrpc(cld)) {
1917 sptlrpc_conf_log_update_begin(cld->cld_logname);
1918 sptlrpc_started = true;
1921 /* logname and instance info should be the same, so use our
1922 * copy of the instance for the update. The cfg_last_idx will
1923 * be updated here. */
1924 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1929 __llog_ctxt_put(env, ctxt);
1931 __llog_ctxt_put(env, lctxt);
1934 * update settings on existing OBDs. doing it inside
1935 * of llog_process_lock so no device is attaching/detaching
1937 * the logname must be <fsname>-sptlrpc
1939 if (sptlrpc_started) {
1940 LASSERT(cld_is_sptlrpc(cld));
1941 sptlrpc_conf_log_update_end(cld->cld_logname);
1942 class_notify_sptlrpc_conf(cld->cld_logname,
1943 strlen(cld->cld_logname) -
1944 strlen("-sptlrpc"));
1953 static bool mgc_import_in_recovery(struct obd_import *imp)
1955 bool in_recovery = true;
1957 spin_lock(&imp->imp_lock);
1958 if (imp->imp_state == LUSTRE_IMP_FULL ||
1959 imp->imp_state == LUSTRE_IMP_CLOSED)
1960 in_recovery = false;
1961 spin_unlock(&imp->imp_lock);
1967 * Get a configuration log from the MGS and process it.
1969 * This function is called for both clients and servers to process the
1970 * configuration log from the MGS. The MGC enqueues a DLM lock on the
1971 * log from the MGS, and if the lock gets revoked the MGC will be notified
1972 * by the lock cancellation callback that the config log has changed,
1973 * and will enqueue another MGS lock on it, and then continue processing
1974 * the new additions to the end of the log.
1976 * Since the MGC import is not replayable, if the import is being evicted
1977 * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
1978 * the log until recovery is finished or the import is closed.
1980 * Make a local copy of the log before parsing it if appropriate (non-MGS
1981 * server) so that the server can start even when the MGS is down.
1983 * There shouldn't be multiple processes running process_log at once --
1984 * sounds like badness. It actually might be fine, as long as they're not
1985 * trying to update from the same log simultaneously, in which case we
1986 * should use a per-log semaphore instead of cld_lock.
1988 * \param[in] mgc MGC device by which to fetch the configuration log
1989 * \param[in] cld log processing state (stored in lock callback data)
1991 * \retval 0 on success
1992 * \retval negative errno on failure
1994 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1996 struct lustre_handle lockh = { 0 };
1997 __u64 flags = LDLM_FL_NO_LRU;
2002 LASSERT(cld != NULL);
2004 /* I don't want multiple processes running process_log at once --
2005 sounds like badness. It actually might be fine, as long as
2006 we're not trying to update from the same log
2007 simultaneously (in which case we should use a per-log sem.) */
2009 mutex_lock(&cld->cld_lock);
2010 if (cld->cld_stopping) {
2011 mutex_unlock(&cld->cld_lock);
2015 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2017 CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
2018 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2020 /* Get the cfg lock on the llog */
2021 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2022 LCK_CR, &flags, NULL, NULL, NULL,
2023 cld, 0, NULL, &lockh);
2025 /* Get the cld, it will be released in mgc_blocking_ast. */
2026 config_log_get(cld);
2027 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2030 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2032 if (rcl == -ESHUTDOWN &&
2033 atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2034 struct obd_import *imp;
2035 struct l_wait_info lwi;
2036 int secs = cfs_time_seconds(obd_timeout);
2038 mutex_unlock(&cld->cld_lock);
2039 imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2041 /* Let's force the pinger, and wait the import to be
2042 * connected, note: since mgc import is non-replayable,
2043 * and even the import state is disconnected, it does
2044 * not mean the "recovery" is stopped, so we will keep
2045 * waitting until timeout or the import state is
2047 ptlrpc_pinger_force(imp);
2049 lwi = LWI_TIMEOUT(secs, NULL, NULL);
2050 l_wait_event(imp->imp_recovery_waitq,
2051 !mgc_import_in_recovery(imp), &lwi);
2053 if (imp->imp_state == LUSTRE_IMP_FULL) {
2057 mutex_lock(&cld->cld_lock);
2058 spin_lock(&config_list_lock);
2059 cld->cld_lostlock = 1;
2060 spin_unlock(&config_list_lock);
2063 /* mark cld_lostlock so that it will requeue
2064 * after MGC becomes available. */
2065 spin_lock(&config_list_lock);
2066 cld->cld_lostlock = 1;
2067 spin_unlock(&config_list_lock);
2071 if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2073 rc = mgc_process_recover_nodemap_log(mgc, cld);
2074 else if (cld_is_nodemap(cld))
2077 if (cld_is_recover(cld) && rc) {
2079 CERROR("%s: recover log %s failed, not fatal: rc = %d\n",
2080 mgc->obd_name, cld->cld_logname, rc);
2081 spin_lock(&config_list_lock);
2082 cld->cld_lostlock = 1;
2083 spin_unlock(&config_list_lock);
2085 rc = 0; /* this is not a fatal error for recover log */
2088 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2091 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2092 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2094 mutex_unlock(&cld->cld_lock);
2096 /* Now drop the lock so MGS can revoke it */
2098 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2100 CERROR("Can't drop cfg lock: %d\n", rcl);
2107 /** Called from lustre_process_log.
2108 * LCFG_LOG_START gets the config log from the MGS, processes it to start
2109 * any services, and adds it to the list logs to watch (follow).
2111 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2113 struct lustre_cfg *lcfg = buf;
2114 struct config_llog_instance *cfg = NULL;
2119 switch(lcfg->lcfg_command) {
2120 case LCFG_LOV_ADD_OBD: {
2121 /* Overloading this cfg command: register a new target */
2122 struct mgs_target_info *mti;
2124 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2125 sizeof(struct mgs_target_info))
2126 GOTO(out, rc = -EINVAL);
2128 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2129 CDEBUG(D_MGC, "add_target %s %#x\n",
2130 mti->mti_svname, mti->mti_flags);
2131 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2134 case LCFG_LOV_DEL_OBD:
2135 /* Unregister has no meaning at the moment. */
2136 CERROR("lov_del_obd unimplemented\n");
2139 case LCFG_SPTLRPC_CONF: {
2140 rc = sptlrpc_process_config(lcfg);
2143 case LCFG_LOG_START: {
2144 struct config_llog_data *cld;
2145 struct super_block *sb;
2147 logname = lustre_cfg_string(lcfg, 1);
2148 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2149 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2151 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2154 /* We're only called through here on the initial mount */
2155 cld = config_log_add(obd, logname, cfg, sb);
2162 /* FIXME only set this for old logs! Right now this forces
2163 us to always skip the "inside markers" check */
2164 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
2166 rc = mgc_process_log(obd, cld);
2167 if (rc == 0 && cld->cld_recover != NULL) {
2168 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2169 imp_connect_data, IMP_RECOV)) {
2170 rc = mgc_process_log(obd, cld->cld_recover);
2172 struct config_llog_data *cir;
2174 mutex_lock(&cld->cld_lock);
2175 cir = cld->cld_recover;
2176 cld->cld_recover = NULL;
2177 mutex_unlock(&cld->cld_lock);
2178 config_log_put(cir);
2182 CERROR("Cannot process recover llog %d\n", rc);
2185 if (rc == 0 && cld->cld_params != NULL) {
2186 rc = mgc_process_log(obd, cld->cld_params);
2187 if (rc == -ENOENT) {
2188 CDEBUG(D_MGC, "There is no params "
2189 "config file yet\n");
2192 /* params log is optional */
2194 CERROR("%s: can't process params llog: rc = %d\n",
2200 case LCFG_LOG_END: {
2201 logname = lustre_cfg_string(lcfg, 1);
2203 if (lcfg->lcfg_bufcount >= 2)
2204 cfg = (struct config_llog_instance *)lustre_cfg_buf(
2206 rc = config_log_end(logname, cfg);
2210 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2211 GOTO(out, rc = -EINVAL);
2219 static struct obd_ops mgc_obd_ops = {
2220 .o_owner = THIS_MODULE,
2221 .o_setup = mgc_setup,
2222 .o_precleanup = mgc_precleanup,
2223 .o_cleanup = mgc_cleanup,
2224 .o_add_conn = client_import_add_conn,
2225 .o_del_conn = client_import_del_conn,
2226 .o_connect = client_connect_import,
2227 .o_disconnect = client_disconnect_export,
2228 .o_set_info_async = mgc_set_info_async,
2229 .o_get_info = mgc_get_info,
2230 .o_import_event = mgc_import_event,
2231 .o_process_config = mgc_process_config,
2234 static int __init mgc_init(void)
2236 return class_register_type(&mgc_obd_ops, NULL, true, NULL,
2237 LUSTRE_MGC_NAME, NULL);
2240 static void __exit mgc_exit(void)
2242 class_unregister_type(LUSTRE_MGC_NAME);
2245 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2246 MODULE_DESCRIPTION("Lustre Management Client");
2247 MODULE_VERSION(LUSTRE_VERSION_STRING);
2248 MODULE_LICENSE("GPL");
2250 module_init(mgc_init);
2251 module_exit(mgc_exit);