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, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lustre/mgc/mgc_request.c
33 * Author: Nathan Rutman <nathan@clusterfs.com>
36 #define DEBUG_SUBSYSTEM S_MGC
37 #define D_MGC D_CONFIG /*|D_WARNING*/
39 #include <linux/module.h>
40 #include <linux/kthread.h>
41 #include <linux/random.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>
51 #include <lustre_barrier.h>
53 #include "mgc_internal.h"
55 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
56 enum mgs_cfg_type type)
60 if (len > sizeof(resname)) {
61 CERROR("name too long: %s\n", name);
65 CERROR("missing name: %s\n", name);
68 memcpy(&resname, name, len);
70 /* Always use the same endianness for the resid */
71 memset(res_id, 0, sizeof(*res_id));
72 res_id->name[0] = cpu_to_le64(resname);
73 /* XXX: unfortunately, sptlprc and config llog share one lock */
75 case MGS_CFG_T_CONFIG:
76 case MGS_CFG_T_SPTLRPC:
79 case MGS_CFG_T_RECOVER:
80 case MGS_CFG_T_PARAMS:
81 case MGS_CFG_T_NODEMAP:
82 case MGS_CFG_T_BARRIER:
88 res_id->name[1] = cpu_to_le64(resname);
89 CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
90 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
94 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id,
95 enum mgs_cfg_type type)
97 /* fsname is at most 8 chars long, maybe contain "-".
98 * e.g. "lustre", "SUN-000" */
99 return mgc_name2resid(fsname, strlen(fsname), res_id, type);
101 EXPORT_SYMBOL(mgc_fsname2resid);
103 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id,
104 enum mgs_cfg_type type)
109 /* logname consists of "fsname-nodetype".
110 * e.g. "lustre-MDT0001", "SUN-000-client"
111 * there is an exception: llog "params" */
112 name_end = strrchr(logname, '-');
114 len = strlen(logname);
116 len = name_end - logname;
117 return mgc_name2resid(logname, len, res_id, type);
119 EXPORT_SYMBOL(mgc_logname2resid);
121 /********************** config llog list **********************/
122 static LIST_HEAD(config_llog_list);
123 static DEFINE_SPINLOCK(config_list_lock); /* protects config_llog_list */
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 (%p) refs %d\n", cld->cld_logname, cld,
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)
144 CDEBUG(D_INFO, "log %s(%p) refs %d\n", cld->cld_logname, cld,
145 atomic_read(&cld->cld_refcount));
146 LASSERT(atomic_read(&cld->cld_refcount) > 0);
148 /* spinlock to make sure no item with 0 refcount in the list */
149 if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
150 list_del(&cld->cld_list_chain);
151 spin_unlock(&config_list_lock);
153 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
155 config_log_put(cld->cld_barrier);
156 config_log_put(cld->cld_recover);
157 config_log_put(cld->cld_params);
158 config_log_put(cld->cld_nodemap);
159 config_log_put(cld->cld_sptlrpc);
160 if (cld_is_sptlrpc(cld)) {
161 cld->cld_stopping = 1;
162 sptlrpc_conf_log_stop(cld->cld_logname);
165 class_export_put(cld->cld_mgcexp);
166 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
172 /* Find a config log by name */
174 struct config_llog_data *config_log_find(char *logname,
175 struct config_llog_instance *cfg)
177 struct config_llog_data *cld;
178 struct config_llog_data *found = NULL;
179 unsigned long cfg_instance;
182 LASSERT(logname != NULL);
184 cfg_instance = cfg ? cfg->cfg_instance : 0;
185 spin_lock(&config_list_lock);
186 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
187 /* check if cfg_instance is the one we want */
188 if (cfg_instance != cld->cld_cfg.cfg_instance)
191 /* instance may be NULL, should check name */
192 if (strcmp(logname, cld->cld_logname) == 0) {
194 config_log_get(found);
198 spin_unlock(&config_list_lock);
203 struct config_llog_data *do_config_log_add(struct obd_device *obd,
205 enum mgs_cfg_type type,
206 struct config_llog_instance *cfg,
207 struct super_block *sb)
209 struct config_llog_data *cld;
214 CDEBUG(D_MGC, "do adding config log %s-%016lx\n", logname,
215 cfg ? cfg->cfg_instance : 0);
217 OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
219 RETURN(ERR_PTR(-ENOMEM));
221 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
223 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
227 strcpy(cld->cld_logname, logname);
231 cld->cld_cfg.cfg_callback = class_config_llog_handler;
232 mutex_init(&cld->cld_lock);
233 cld->cld_cfg.cfg_last_idx = 0;
234 cld->cld_cfg.cfg_flags = 0;
235 cld->cld_cfg.cfg_sb = sb;
236 cld->cld_type = type;
237 atomic_set(&cld->cld_refcount, 1);
239 /* Keep the mgc around until we are done */
240 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
242 if (cld_is_sptlrpc(cld))
243 sptlrpc_conf_log_start(logname);
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) || cld_is_barrier(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 config_llog_data *cld;
268 if (IS_OST(s2lsi(sb)))
271 /* for osp-on-ost, see lustre_start_osp() */
272 if (IS_MDT(s2lsi(sb)) && lcfg.cfg_instance)
275 /* We have to use different llog for clients and MDTs for DNE,
276 * where only clients are notified if one of DNE server restarts.
278 LASSERT(strlen(fsname) < sizeof(logname) / 2);
279 strncpy(logname, fsname, sizeof(logname));
280 if (IS_SERVER(s2lsi(sb))) { /* mdt */
281 LASSERT(lcfg.cfg_instance == 0);
282 lcfg.cfg_instance = ll_get_cfg_instance(sb);
283 strncat(logname, "-mdtir", sizeof(logname));
285 LASSERT(lcfg.cfg_instance != 0);
286 strncat(logname, "-cliir", sizeof(logname));
289 cld = do_config_log_add(obd, logname, MGS_CFG_T_RECOVER, &lcfg, sb);
293 static struct config_llog_data *
294 config_log_find_or_add(struct obd_device *obd, char *logname,
295 struct super_block *sb, enum mgs_cfg_type type,
296 struct config_llog_instance *cfg)
298 struct config_llog_instance lcfg = *cfg;
299 struct config_llog_data *cld;
301 /* Note class_config_llog_handler() depends on getting "obd" back */
302 lcfg.cfg_instance = sb ? ll_get_cfg_instance(sb) : (unsigned long)obd;
304 cld = config_log_find(logname, &lcfg);
305 if (unlikely(cld != NULL))
308 return do_config_log_add(obd, logname, type, &lcfg, sb);
311 /** Add this log to the list of active logs watched by an MGC.
312 * Active means we're watching for updates.
313 * We have one active log per "mount" - client instance or servername.
314 * Each instance may be at a different point in the log.
316 static struct config_llog_data *
317 config_log_add(struct obd_device *obd, char *logname,
318 struct config_llog_instance *cfg, struct super_block *sb)
320 struct lustre_sb_info *lsi = s2lsi(sb);
321 struct config_llog_data *cld = NULL;
322 struct config_llog_data *sptlrpc_cld = NULL;
323 struct config_llog_data *params_cld = NULL;
324 struct config_llog_data *nodemap_cld = NULL;
325 struct config_llog_data *barrier_cld = NULL;
332 CDEBUG(D_MGC, "add config log %s-%016lx\n", logname,
336 * for each regular log, the depended sptlrpc log name is
337 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
339 ptr = strrchr(logname, '-');
340 if (ptr == NULL || ptr - logname > 8) {
341 CERROR("logname %s is too long\n", logname);
342 RETURN(ERR_PTR(-EINVAL));
345 memcpy(seclogname, logname, ptr - logname);
346 strcpy(seclogname + (ptr - logname), "-sptlrpc");
348 if (cfg->cfg_sub_clds & CONFIG_SUB_SPTLRPC) {
349 sptlrpc_cld = config_log_find_or_add(obd, seclogname, NULL,
350 MGS_CFG_T_SPTLRPC, cfg);
351 if (IS_ERR(sptlrpc_cld)) {
352 CERROR("%s: can't create sptlrpc log %s: rc = %ld\n",
353 obd->obd_name, seclogname, PTR_ERR(sptlrpc_cld));
358 if (!IS_MGS(lsi) && cfg->cfg_sub_clds & CONFIG_SUB_NODEMAP) {
359 nodemap_cld = config_log_find_or_add(obd, LUSTRE_NODEMAP_NAME,
360 NULL, MGS_CFG_T_NODEMAP,
362 if (IS_ERR(nodemap_cld)) {
363 rc = PTR_ERR(nodemap_cld);
364 CERROR("%s: cannot create nodemap log: rc = %d\n",
366 GOTO(out_sptlrpc, rc);
370 if (cfg->cfg_sub_clds & CONFIG_SUB_PARAMS) {
371 params_cld = config_log_find_or_add(obd, PARAMS_FILENAME, sb,
372 MGS_CFG_T_PARAMS, cfg);
373 if (IS_ERR(params_cld)) {
374 rc = PTR_ERR(params_cld);
375 CERROR("%s: can't create params log: rc = %d\n",
377 GOTO(out_nodemap, rc);
381 if (IS_MDT(s2lsi(sb)) && cfg->cfg_sub_clds & CONFIG_SUB_BARRIER) {
382 snprintf(seclogname + (ptr - logname), sizeof(seclogname) - 1,
383 "-%s", BARRIER_FILENAME);
384 barrier_cld = config_log_find_or_add(obd, seclogname, sb,
385 MGS_CFG_T_BARRIER, cfg);
386 if (IS_ERR(barrier_cld)) {
387 rc = PTR_ERR(barrier_cld);
388 CERROR("%s: can't create barrier log: rc = %d\n",
390 GOTO(out_params, rc);
394 cld = do_config_log_add(obd, logname, MGS_CFG_T_CONFIG, cfg, sb);
397 CERROR("%s: can't create log: rc = %d\n",
399 GOTO(out_barrier, rc = PTR_ERR(cld));
402 LASSERT(lsi->lsi_lmd);
403 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR) &&
404 cfg->cfg_sub_clds & CONFIG_SUB_RECOVER) {
405 struct config_llog_data *recover_cld;
407 ptr = strrchr(seclogname, '-');
411 CERROR("%s: sptlrpc log name not correct, %s: "
412 "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
413 GOTO(out_cld, rc = -EINVAL);
416 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
417 if (IS_ERR(recover_cld)) {
418 rc = PTR_ERR(recover_cld);
419 CERROR("%s: can't create recover log: rc = %d\n",
424 mutex_lock(&cld->cld_lock);
426 cld->cld_recover = recover_cld;
430 mutex_lock(&cld->cld_lock);
431 cld->cld_params = params_cld;
432 cld->cld_barrier = barrier_cld;
433 cld->cld_nodemap = nodemap_cld;
434 cld->cld_sptlrpc = sptlrpc_cld;
435 mutex_unlock(&cld->cld_lock);
442 config_log_put(barrier_cld);
444 config_log_put(params_cld);
446 config_log_put(nodemap_cld);
448 config_log_put(sptlrpc_cld);
453 DEFINE_MUTEX(llog_process_lock);
455 static inline void config_mark_cld_stop_nolock(struct config_llog_data *cld)
459 spin_lock(&config_list_lock);
460 cld->cld_stopping = 1;
461 spin_unlock(&config_list_lock);
463 CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
464 if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
465 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
468 static inline void config_mark_cld_stop(struct config_llog_data *cld)
471 mutex_lock(&cld->cld_lock);
472 config_mark_cld_stop_nolock(cld);
473 mutex_unlock(&cld->cld_lock);
477 /** Stop watching for updates on this log.
479 static int config_log_end(char *logname, struct config_llog_instance *cfg)
481 struct config_llog_data *cld;
482 struct config_llog_data *cld_sptlrpc = NULL;
483 struct config_llog_data *cld_params = NULL;
484 struct config_llog_data *cld_recover = NULL;
485 struct config_llog_data *cld_nodemap = NULL;
486 struct config_llog_data *cld_barrier = NULL;
491 cld = config_log_find(logname, cfg);
495 mutex_lock(&cld->cld_lock);
497 * if cld_stopping is set, it means we didn't start the log thus
498 * not owning the start ref. this can happen after previous umount:
499 * the cld still hanging there waiting for lock cancel, and we
500 * remount again but failed in the middle and call log_end without
503 if (unlikely(cld->cld_stopping)) {
504 mutex_unlock(&cld->cld_lock);
505 /* drop the ref from the find */
510 cld_recover = cld->cld_recover;
511 cld->cld_recover = NULL;
512 cld_params = cld->cld_params;
513 cld->cld_params = NULL;
514 cld_nodemap = cld->cld_nodemap;
515 cld->cld_nodemap = NULL;
516 cld_barrier = cld->cld_barrier;
517 cld->cld_barrier = NULL;
518 cld_sptlrpc = cld->cld_sptlrpc;
519 cld->cld_sptlrpc = NULL;
521 config_mark_cld_stop_nolock(cld);
522 mutex_unlock(&cld->cld_lock);
524 config_mark_cld_stop(cld_recover);
525 config_log_put(cld_recover);
526 config_mark_cld_stop(cld_params);
527 config_log_put(cld_params);
528 config_mark_cld_stop(cld_barrier);
529 config_log_put(cld_barrier);
530 /* don't explicitly set cld_stopping on sptlrpc lock here, as other
531 * targets may be active, it will be done in config_log_put if necessary
533 config_log_put(cld_sptlrpc);
534 /* don't set cld_stopping on nm lock as other targets may be active */
535 config_log_put(cld_nodemap);
537 /* drop the ref from the find */
539 /* drop the start ref */
542 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
547 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
549 struct obd_device *obd = data;
550 struct obd_import *imp;
551 struct obd_connect_data *ocd;
552 struct config_llog_data *cld;
557 with_imp_locked(obd, imp, rc) {
558 ocd = &imp->imp_connect_data;
560 seq_printf(m, "imperative_recovery: %s\n",
561 OCD_HAS_FLAG(ocd, IMP_RECOV) ?
562 "ENABLED" : "DISABLED");
567 seq_printf(m, "client_state:\n");
569 spin_lock(&config_list_lock);
570 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
571 if (cld->cld_recover == NULL)
573 seq_printf(m, " - { client: %s, nidtbl_version: %u }\n",
575 cld->cld_recover->cld_cfg.cfg_last_idx);
577 spin_unlock(&config_list_lock);
582 /* reenqueue any lost locks */
583 #define RQ_RUNNING 0x1
587 #define RQ_PRECLEANUP 0x10
588 static int rq_state = 0;
589 static wait_queue_head_t rq_waitq;
590 static DECLARE_COMPLETION(rq_exit);
591 static DECLARE_COMPLETION(rq_start);
593 static void do_requeue(struct config_llog_data *cld)
598 LASSERT(atomic_read(&cld->cld_refcount) > 0);
601 * Do not run mgc_process_log on a disconnected export or an
602 * export which is being disconnected. Take the client
603 * semaphore to make the check non-racy.
605 down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
607 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
608 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
609 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
610 if (rc && rc != -ENOENT)
611 CERROR("failed processing log: %d\n", rc);
613 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
616 up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
621 static int mgc_requeue_thread(void *data)
627 CDEBUG(D_MGC, "Starting requeue thread\n");
629 /* Keep trying failed locks periodically */
630 spin_lock(&config_list_lock);
631 rq_state |= RQ_RUNNING;
632 while (!(rq_state & RQ_STOP)) {
633 struct config_llog_data *cld, *cld_prev;
636 /* Any new or requeued lostlocks will change the state */
637 rq_state &= ~(RQ_NOW | RQ_LATER);
638 spin_unlock(&config_list_lock);
645 /* Always wait a few seconds to allow the server who
646 * caused the lock revocation to finish its setup, plus some
647 * random so everyone doesn't try to reconnect at once.
649 to = mgc_requeue_timeout_min == 0 ? 1 : mgc_requeue_timeout_min;
650 to = cfs_time_seconds(mgc_requeue_timeout_min) +
651 prandom_u32_max(cfs_time_seconds(to));
652 wait_event_idle_timeout(rq_waitq,
653 rq_state & (RQ_STOP | RQ_PRECLEANUP), to);
656 * iterate & processing through the list. for each cld, process
657 * its depending sptlrpc cld firstly (if any) and then itself.
659 * it's guaranteed any item in the list must have
660 * reference > 0; and if cld_lostlock is set, at
661 * least one reference is taken by the previous enqueue.
665 spin_lock(&config_list_lock);
666 rq_state &= ~RQ_PRECLEANUP;
667 list_for_each_entry(cld, &config_llog_list,
669 if (!cld->cld_lostlock || cld->cld_stopping)
672 /* hold reference to avoid being freed during
673 * subsequent processing. */
675 cld->cld_lostlock = 0;
676 spin_unlock(&config_list_lock);
678 config_log_put(cld_prev);
681 if (likely(!(rq_state & RQ_STOP))) {
683 spin_lock(&config_list_lock);
685 spin_lock(&config_list_lock);
689 spin_unlock(&config_list_lock);
690 config_log_put(cld_prev);
692 /* Wait a bit to see if anyone else needs a requeue */
693 wait_event_idle(rq_waitq, rq_state & (RQ_NOW | RQ_STOP));
694 spin_lock(&config_list_lock);
697 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
698 rq_state &= ~RQ_RUNNING;
699 spin_unlock(&config_list_lock);
703 CDEBUG(D_MGC, "Ending requeue thread\n");
707 /* Add a cld to the list to requeue. Start the requeue thread if needed.
708 We are responsible for dropping the config log reference from here on out. */
709 static void mgc_requeue_add(struct config_llog_data *cld)
714 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
715 cld->cld_logname, atomic_read(&cld->cld_refcount),
716 cld->cld_stopping, rq_state);
717 LASSERT(atomic_read(&cld->cld_refcount) > 0);
719 /* lets cancel an existent lock to mark cld as "lostlock" */
720 CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
721 if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
722 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
724 mutex_lock(&cld->cld_lock);
725 spin_lock(&config_list_lock);
726 if (!(rq_state & RQ_STOP) && !cld->cld_stopping) {
727 cld->cld_lostlock = 1;
731 spin_unlock(&config_list_lock);
732 mutex_unlock(&cld->cld_lock);
739 /********************** class fns **********************/
740 #ifdef HAVE_SERVER_SUPPORT
741 static int mgc_local_llog_init(const struct lu_env *env,
742 struct obd_device *obd,
743 struct obd_device *disk)
745 struct llog_ctxt *ctxt;
750 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
755 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
757 ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
763 static int mgc_local_llog_fini(const struct lu_env *env,
764 struct obd_device *obd)
766 struct llog_ctxt *ctxt;
770 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
771 llog_cleanup(env, ctxt);
776 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
777 struct super_block *sb)
779 struct lustre_sb_info *lsi = s2lsi(sb);
780 struct client_obd *cli = &obd->u.cli;
781 struct lu_fid rfid, fid;
782 struct dt_object *root, *dto;
788 LASSERT(lsi->lsi_dt_dev);
790 /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
791 mutex_lock(&cli->cl_mgc_mutex);
793 /* Setup the configs dir */
794 fid.f_seq = FID_SEQ_LOCAL_NAME;
797 rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
802 rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
806 root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
807 &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
808 if (unlikely(IS_ERR(root)))
809 GOTO(out_los, rc = PTR_ERR(root));
811 dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
813 S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
814 dt_object_put_nocache(env, root);
816 GOTO(out_los, rc = PTR_ERR(dto));
818 cli->cl_mgc_configs_dir = dto;
820 LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
821 rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
825 /* We take an obd ref to insure that we can't get to mgc_cleanup
826 * without calling mgc_fs_cleanup first. */
827 class_incref(obd, "mgc_fs", obd);
829 /* We keep the cl_mgc_sem until mgc_fs_cleanup */
833 dt_object_put(env, cli->cl_mgc_configs_dir);
834 cli->cl_mgc_configs_dir = NULL;
838 local_oid_storage_fini(env, cli->cl_mgc_los);
840 cli->cl_mgc_los = NULL;
841 mutex_unlock(&cli->cl_mgc_mutex);
846 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
848 struct client_obd *cli = &obd->u.cli;
851 LASSERT(cli->cl_mgc_los != NULL);
853 mgc_local_llog_fini(env, obd);
855 dt_object_put_nocache(env, cli->cl_mgc_configs_dir);
856 cli->cl_mgc_configs_dir = NULL;
858 local_oid_storage_fini(env, cli->cl_mgc_los);
859 cli->cl_mgc_los = NULL;
861 class_decref(obd, "mgc_fs", obd);
862 mutex_unlock(&cli->cl_mgc_mutex);
866 #endif /* HAVE_SERVER_SUPPORT */
868 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
870 struct llog_ctxt *ctxt;
875 /* setup only remote ctxt, the local disk context is switched per each
876 * filesystem during mgc_fs_setup() */
877 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
882 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
885 llog_initiator_connect(ctxt);
891 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
893 struct llog_ctxt *ctxt;
897 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
899 llog_cleanup(env, ctxt);
905 static atomic_t mgc_count = ATOMIC_INIT(0);
906 static int mgc_precleanup(struct obd_device *obd)
912 if (atomic_dec_and_test(&mgc_count)) {
913 LASSERT(rq_state & RQ_RUNNING);
914 /* stop requeue thread */
917 /* wakeup requeue thread to clean our cld */
918 temp = RQ_NOW | RQ_PRECLEANUP;
921 spin_lock(&config_list_lock);
923 spin_unlock(&config_list_lock);
927 wait_for_completion(&rq_exit);
928 obd_cleanup_client_import(obd);
930 rc = mgc_llog_fini(NULL, obd);
932 CERROR("failed to cleanup llogging subsystems\n");
937 static int mgc_cleanup(struct obd_device *obd)
942 /* COMPAT_146 - old config logs may have added profiles we don't
944 if (atomic_read(&obd->obd_type->typ_refcnt) <= 1)
945 /* Only for the last mgc */
946 class_del_profiles();
948 lprocfs_obd_cleanup(obd);
951 rc = client_obd_cleanup(obd);
955 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
957 struct task_struct *task;
961 rc = ptlrpcd_addref();
965 rc = client_obd_setup(obd, lcfg);
967 GOTO(err_decref, rc);
969 rc = mgc_llog_init(NULL, obd);
971 CERROR("failed to setup llogging subsystems\n");
972 GOTO(err_cleanup, rc);
975 rc = mgc_tunables_init(obd);
979 if (atomic_inc_return(&mgc_count) == 1) {
981 init_waitqueue_head(&rq_waitq);
983 /* start requeue thread */
984 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
987 CERROR("%s: cannot start requeue thread: rc = %d; "
988 "no more log updates\n",
992 /* rc is the task_struct pointer of mgc_requeue_thread. */
994 wait_for_completion(&rq_start);
1000 lprocfs_obd_cleanup(obd);
1002 client_obd_cleanup(obd);
1008 /* based on ll_mdc_blocking_ast */
1009 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1010 void *data, int flag)
1012 struct lustre_handle lockh;
1013 struct config_llog_data *cld = (struct config_llog_data *)data;
1018 case LDLM_CB_BLOCKING:
1019 /* mgs wants the lock, give it up... */
1020 LDLM_DEBUG(lock, "MGC blocking CB");
1021 ldlm_lock2handle(lock, &lockh);
1022 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
1024 case LDLM_CB_CANCELING:
1025 /* We've given up the lock, prepare ourselves to update. */
1026 LDLM_DEBUG(lock, "MGC cancel CB");
1028 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
1029 PLDLMRES(lock->l_resource),
1030 (char *)&lock->l_resource->lr_name.name[0]);
1033 CDEBUG(D_INFO, "missing data, won't requeue\n");
1037 /* held at mgc_process_log(). */
1038 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1040 lock->l_ast_data = NULL;
1041 cld->cld_lockh.cookie = 0;
1042 /* Are we done with this log? */
1043 if (cld->cld_stopping) {
1044 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1046 config_log_put(cld);
1049 /* Make sure not to re-enqueue when the mgc is stopping
1050 (we get called from client_disconnect_export) */
1051 if (lock->l_conn_export == NULL ||
1052 lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1053 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1055 config_log_put(cld);
1059 /* Re-enqueue now */
1060 mgc_requeue_add(cld);
1061 config_log_put(cld);
1070 /* Not sure where this should go... */
1071 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1072 * that we can have a chance to try the secondary MGS if any. */
1073 #define MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1075 #define MGC_TARGET_REG_LIMIT 10
1076 #define MGC_TARGET_REG_LIMIT_MAX RECONNECT_DELAY_MAX
1077 #define MGC_SEND_PARAM_LIMIT 10
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, ldlm_glimpse_callback glimpse_callback,
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,
1092 .ei_cb_gl = glimpse_callback,
1094 struct ptlrpc_request *req;
1095 int short_limit = cld_is_sptlrpc(cld);
1102 CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1103 cld->cld_resid.name[0]);
1105 /* We need a callback for every lockholder, so don't try to
1106 ldlm_lock_match (see rev 1.1.2.11.2.47) */
1107 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1108 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1113 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1114 ptlrpc_request_set_replen(req);
1116 /* check if this is server or client */
1117 if (cld->cld_cfg.cfg_sb &&
1118 IS_SERVER(s2lsi(cld->cld_cfg.cfg_sb)))
1121 /* Limit how long we will wait for the enqueue to complete */
1122 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1123 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1124 NULL, 0, LVB_T_NONE, lockh, 0);
1125 /* A failed enqueue should still call the mgc_blocking_ast,
1126 where it will be requeued if needed ("grant failed"). */
1127 ptlrpc_req_finished(req);
1131 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1132 struct lustre_handle *lockh)
1136 ldlm_lock_decref(lockh, mode);
1141 static void mgc_notify_active(struct obd_device *unused)
1143 /* wakeup mgc_requeue_thread to requeue mgc lock */
1144 spin_lock(&config_list_lock);
1146 spin_unlock(&config_list_lock);
1149 /* TODO: Help the MGS rebuild nidtbl. -jay */
1152 #ifdef HAVE_SERVER_SUPPORT
1153 /* Send target_reg message to MGS */
1154 static int mgc_target_register(struct obd_export *exp,
1155 struct mgs_target_info *mti)
1157 struct ptlrpc_request *req;
1158 struct mgs_target_info *req_mti, *rep_mti;
1162 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1163 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1168 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1170 ptlrpc_req_finished(req);
1174 memcpy(req_mti, mti, sizeof(*req_mti));
1175 ptlrpc_request_set_replen(req);
1176 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1177 /* Limit how long we will wait for the enqueue to complete */
1178 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1180 /* if the target needs to regenerate the config log in MGS, it's better
1181 * to use some longer limit to let MGC have time to change connection to
1182 * another MGS (or try again with the same MGS) for the target (server)
1183 * will fail and exit if the request expired due to delay limit. */
1184 if (mti->mti_flags & (LDD_F_UPDATE | LDD_F_NEED_INDEX))
1185 req->rq_delay_limit = MGC_TARGET_REG_LIMIT_MAX;
1187 rc = ptlrpc_queue_wait(req);
1188 if (ptlrpc_client_replied(req)) {
1189 rep_mti = req_capsule_server_get(&req->rq_pill,
1190 &RMF_MGS_TARGET_INFO);
1192 memcpy(mti, rep_mti, sizeof(*rep_mti));
1195 CDEBUG(D_MGC, "register %s got index = %d\n",
1196 mti->mti_svname, mti->mti_stripe_index);
1198 ptlrpc_req_finished(req);
1202 #endif /* HAVE_SERVER_SUPPORT */
1204 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1205 u32 keylen, void *key,
1206 u32 vallen, void *val,
1207 struct ptlrpc_request_set *set)
1212 /* Turn off initial_recov after we try all backup servers once */
1213 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1214 struct obd_import *imp = class_exp2cliimp(exp);
1216 if (vallen != sizeof(int))
1218 value = *(int *)val;
1219 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1220 imp->imp_obd->obd_name, value,
1221 imp->imp_deactive, imp->imp_invalid,
1222 imp->imp_replayable, imp->imp_obd->obd_replayable,
1223 ptlrpc_import_state_name(imp->imp_state));
1224 /* Resurrect the import immediately if
1225 * 1. we previously got disconnected,
1226 * 2. value > 1 (at the same node with MGS)
1228 if (imp->imp_state == LUSTRE_IMP_DISCON || value > 1)
1229 ptlrpc_reconnect_import(imp);
1234 #ifdef HAVE_SERVER_SUPPORT
1235 /* FIXME move this to mgc_process_config */
1236 if (KEY_IS(KEY_REGISTER_TARGET)) {
1237 struct mgs_target_info *mti;
1239 if (vallen != sizeof(struct mgs_target_info))
1241 mti = (struct mgs_target_info *)val;
1242 CDEBUG(D_MGC, "register_target %s %#x\n",
1243 mti->mti_svname, mti->mti_flags);
1244 rc = mgc_target_register(exp, mti);
1247 if (KEY_IS(KEY_SET_FS)) {
1248 struct super_block *sb = (struct super_block *)val;
1250 if (vallen != sizeof(struct super_block))
1253 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1256 if (KEY_IS(KEY_CLEAR_FS)) {
1259 rc = mgc_fs_cleanup(env, exp->exp_obd);
1263 if (KEY_IS(KEY_MGSSEC)) {
1264 struct client_obd *cli = &exp->exp_obd->u.cli;
1265 struct sptlrpc_flavor flvr;
1268 * empty string means using current flavor, if which haven't
1269 * been set yet, set it as null.
1271 * if flavor has been set previously, check the asking flavor
1272 * must match the existing one.
1275 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1281 rc = sptlrpc_parse_flavor(val, &flvr);
1283 CERROR("invalid sptlrpc flavor %s to MGS\n",
1289 * caller already hold a mutex
1291 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1292 cli->cl_flvr_mgc = flvr;
1293 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1294 sizeof(flvr)) != 0) {
1297 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1299 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1300 "currently %s is in use\n",
1310 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1311 __u32 keylen, void *key, __u32 *vallen, void *val)
1315 if (KEY_IS(KEY_CONN_DATA)) {
1316 struct obd_import *imp = class_exp2cliimp(exp);
1317 struct obd_connect_data *data = val;
1319 if (*vallen == sizeof(*data)) {
1320 *data = imp->imp_connect_data;
1328 static int mgc_import_event(struct obd_device *obd,
1329 struct obd_import *imp,
1330 enum obd_import_event event)
1334 LASSERT(imp->imp_obd == obd);
1335 CDEBUG(D_MGC, "import event %#x\n", event);
1338 case IMP_EVENT_DISCON:
1339 /* MGC imports should not wait for recovery */
1340 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1341 ptlrpc_pinger_ir_down();
1343 case IMP_EVENT_INACTIVE:
1345 case IMP_EVENT_INVALIDATE: {
1346 struct ldlm_namespace *ns = obd->obd_namespace;
1347 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1350 case IMP_EVENT_ACTIVE:
1351 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1352 /* Clearing obd_no_recov allows us to continue pinging */
1353 obd->obd_no_recov = 0;
1354 mgc_notify_active(obd);
1355 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1356 ptlrpc_pinger_ir_up();
1360 case IMP_EVENT_DEACTIVATE:
1361 case IMP_EVENT_ACTIVATE:
1364 CERROR("Unknown import event %#x\n", event);
1371 CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1372 CONFIG_READ_NRPAGES = 4
1375 static int mgc_apply_recover_logs(struct obd_device *mgc,
1376 struct config_llog_data *cld,
1378 void *data, int datalen, bool mne_swab)
1380 struct config_llog_instance *cfg = &cld->cld_cfg;
1381 struct mgs_nidtbl_entry *entry;
1382 struct lustre_cfg *lcfg;
1383 struct lustre_cfg_bufs bufs;
1384 u64 prev_version = 0;
1385 char inst[MTI_NAME_MAXLEN + 1];
1391 unsigned long dynamic_nids;
1394 LASSERT(cfg->cfg_instance != 0);
1395 LASSERT(ll_get_cfg_instance(cfg->cfg_sb) == cfg->cfg_instance);
1397 /* get dynamic nids setting */
1398 dynamic_nids = mgc->obd_dynamic_nids;
1400 if (!IS_SERVER(s2lsi(cfg->cfg_sb))) {
1401 pos = snprintf(inst, sizeof(inst), "%016lx", cfg->cfg_instance);
1402 if (pos >= PAGE_SIZE)
1404 #ifdef HAVE_SERVER_SUPPORT
1406 struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1408 LASSERT(IS_MDT(lsi));
1409 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1413 #endif /* HAVE_SERVER_SUPPORT */
1416 OBD_ALLOC(buf, PAGE_SIZE);
1422 while (datalen > 0) {
1423 int entry_len = sizeof(*entry);
1425 struct obd_device *obd;
1426 struct obd_import *imp;
1433 if (datalen < sizeof(*entry))
1436 entry = (typeof(entry))(data + off);
1439 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1441 if (entry->mne_nid_count == 0) /* at least one nid entry */
1443 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1446 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1447 if (datalen < entry_len) /* must have entry_len at least */
1450 /* Keep this swab for normal mixed endian handling. LU-1644 */
1452 lustre_swab_mgs_nidtbl_entry(entry);
1453 if (entry->mne_length > PAGE_SIZE) {
1454 CERROR("MNE too large (%u)\n", entry->mne_length);
1458 if (entry->mne_length < entry_len)
1461 off += entry->mne_length;
1462 datalen -= entry->mne_length;
1466 if (entry->mne_version > max_version) {
1467 CERROR("entry index(%lld) is over max_index(%lld)\n",
1468 entry->mne_version, max_version);
1472 if (prev_version >= entry->mne_version) {
1473 CERROR("index unsorted, prev %lld, now %lld\n",
1474 prev_version, entry->mne_version);
1477 prev_version = entry->mne_version;
1480 * Write a string with format "nid::instance" to
1481 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1484 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1485 memset(buf, 0, bufsz);
1489 /* lustre-OST0001-osc-<instance #> */
1490 strcpy(obdname, cld->cld_logname);
1491 cname = strrchr(obdname, '-');
1492 if (cname == NULL) {
1493 CERROR("mgc %s: invalid logname %s\n",
1494 mgc->obd_name, obdname);
1498 pos = cname - obdname;
1500 pos += sprintf(obdname + pos, "-%s%04x",
1501 is_ost ? "OST" : "MDT", entry->mne_index);
1503 cname = is_ost ? "osc" : "mdc",
1504 pos += snprintf(obdname + pos, bufsz, "-%s-%s", cname, inst);
1505 lustre_cfg_bufs_reset(&bufs, obdname);
1507 /* find the obd by obdname */
1508 obd = class_name2obd(obdname);
1510 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1511 mgc->obd_name, obdname);
1513 /* this is a safe race, when the ost is starting up...*/
1517 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1520 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1523 with_imp_locked(obd, imp, rc) {
1524 /* iterate all nids to find one */
1525 /* find uuid by nid */
1526 /* create import entries if they don't exist */
1527 rc = client_import_add_nids_to_conn(
1528 imp, entry->u.nids, entry->mne_nid_count,
1529 (struct obd_uuid *)uuid);
1531 if (rc == -ENOENT && dynamic_nids) {
1532 /* create a new connection for this import */
1534 libcfs_nid2str(entry->u.nids[0]);
1535 int prim_nid_len = strlen(primary_nid) + 1;
1536 struct obd_uuid server_uuid;
1538 if (prim_nid_len > UUID_MAX)
1540 strncpy(server_uuid.uuid, primary_nid,
1543 CDEBUG(D_INFO, "Adding a connection for %s\n",
1546 rc = client_import_dyn_add_conn(
1547 imp, &server_uuid, entry->u.nids[0], 1);
1549 CERROR("%s: Failed to add new connection with NID '%s' to import: rc = %d\n",
1550 obd->obd_name, primary_nid, rc);
1553 rc = client_import_add_nids_to_conn(
1555 entry->mne_nid_count,
1556 (struct obd_uuid *)uuid);
1558 CERROR("%s: failed to lookup UUID: rc = %d\n",
1565 if (rc == -ENODEV) {
1566 /* client does not connect to the OST yet */
1571 if (rc < 0 && rc != -ENOSPC) {
1572 CERROR("mgc: cannot find UUID by nid '%s': rc = %d\n",
1573 libcfs_nid2str(entry->u.nids[0]), rc);
1577 CDEBUG(D_INFO, "Found UUID '%s' by NID '%s'\n",
1578 uuid, libcfs_nid2str(entry->u.nids[0]));
1580 pos += strlen(uuid);
1581 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1582 LASSERT(pos < bufsz);
1584 lustre_cfg_bufs_set_string(&bufs, 1, params);
1586 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1592 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1594 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1595 prev_version, max_version, obdname, params);
1597 rc = class_process_config(lcfg);
1598 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1599 lcfg->lcfg_buflens));
1601 CDEBUG(D_INFO, "process config for %s error %d\n",
1604 /* continue, even one with error */
1607 OBD_FREE(buf, PAGE_SIZE);
1613 * This function is called if this client was notified for target restarting
1614 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1617 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1618 struct config_llog_data *cld)
1620 struct ptlrpc_connection *mgc_conn;
1621 struct ptlrpc_request *req = NULL;
1622 struct config_llog_instance *cfg = &cld->cld_cfg;
1623 struct mgs_config_body *body;
1624 struct mgs_config_res *res;
1625 struct nodemap_config *new_config = NULL;
1626 struct lu_nodemap *recent_nodemap = NULL;
1627 struct ptlrpc_bulk_desc *desc;
1628 struct page **pages = NULL;
1629 __u64 config_read_offset = 0;
1630 __u8 nodemap_cur_pass = 0;
1633 bool mne_swab = false;
1639 mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1641 /* don't need to get local config */
1642 if (cld_is_nodemap(cld) &&
1643 LNetIsPeerLocal(&mgc_conn->c_peer.nid))
1646 /* allocate buffer for bulk transfer.
1647 * if this is the first time for this mgs to read logs,
1648 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1649 * once; otherwise, it only reads increment of logs, this should be
1650 * small and CONFIG_READ_NRPAGES will be used.
1652 nrpages = CONFIG_READ_NRPAGES;
1653 if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1654 nrpages = CONFIG_READ_NRPAGES_INIT;
1656 OBD_ALLOC_PTR_ARRAY_LARGE(pages, nrpages);
1658 GOTO(out, rc = -ENOMEM);
1660 for (i = 0; i < nrpages; i++) {
1661 pages[i] = alloc_page(GFP_KERNEL);
1662 if (pages[i] == NULL)
1663 GOTO(out, rc = -ENOMEM);
1667 #ifdef HAVE_SERVER_SUPPORT
1668 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1669 new_config = nodemap_config_alloc();
1670 if (IS_ERR(new_config)) {
1671 rc = PTR_ERR(new_config);
1677 LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1678 LASSERT(mutex_is_locked(&cld->cld_lock));
1679 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1680 &RQF_MGS_CONFIG_READ);
1682 GOTO(out, rc = -ENOMEM);
1684 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1689 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1690 LASSERT(body != NULL);
1691 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1692 if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1693 >= sizeof(body->mcb_name))
1694 GOTO(out, rc = -E2BIG);
1695 if (cld_is_nodemap(cld))
1696 body->mcb_offset = config_read_offset;
1698 body->mcb_offset = cfg->cfg_last_idx + 1;
1699 body->mcb_type = cld->cld_type;
1700 body->mcb_bits = PAGE_SHIFT;
1701 body->mcb_units = nrpages;
1702 body->mcb_nm_cur_pass = nodemap_cur_pass;
1704 /* allocate bulk transfer descriptor */
1705 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1706 PTLRPC_BULK_PUT_SINK,
1708 &ptlrpc_bulk_kiov_pin_ops);
1710 GOTO(out, rc = -ENOMEM);
1712 for (i = 0; i < nrpages; i++)
1713 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1716 ptlrpc_request_set_replen(req);
1717 rc = ptlrpc_queue_wait(req);
1721 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1723 GOTO(out, rc = -EPROTO);
1725 if (cld_is_nodemap(cld)) {
1726 config_read_offset = res->mcr_offset;
1727 eof = config_read_offset == II_END_OFF;
1728 nodemap_cur_pass = res->mcr_nm_cur_pass;
1730 if (res->mcr_size < res->mcr_offset)
1731 GOTO(out, rc = -EINVAL);
1733 /* always update the index even though it might have errors with
1734 * handling the recover logs
1736 cfg->cfg_last_idx = res->mcr_offset;
1737 eof = res->mcr_offset == res->mcr_size;
1739 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1740 res->mcr_offset, eof == false);
1743 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1745 GOTO(out, rc = ealen);
1747 if (ealen > nrpages << PAGE_SHIFT)
1748 GOTO(out, rc = -EINVAL);
1750 if (ealen == 0) { /* no logs transferred */
1751 #ifdef HAVE_SERVER_SUPPORT
1752 /* config changed since first read RPC */
1753 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1754 CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1755 GOTO(out, rc = -EAGAIN);
1763 mne_swab = req_capsule_rep_need_swab(&req->rq_pill);
1765 /* When a nodemap config is received, we build a new nodemap config,
1766 * with new nodemap structs. We keep track of the most recently added
1767 * nodemap since the config is read ordered by nodemap_id, and so it
1768 * is likely that the next record will be related. Because access to
1769 * the nodemaps is single threaded until the nodemap_config is active,
1770 * we don't need to reference count with recent_nodemap, though
1771 * recent_nodemap should be set to NULL when the nodemap_config
1772 * is either destroyed or set active.
1774 for (i = 0; i < nrpages && ealen > 0; i++) {
1778 ptr = kmap(pages[i]);
1779 if (cld_is_nodemap(cld))
1780 rc2 = nodemap_process_idx_pages(new_config, ptr,
1783 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1790 CWARN("%s: error processing %s log %s: rc = %d\n",
1792 cld_is_nodemap(cld) ? "nodemap" : "recovery",
1795 GOTO(out, rc = rc2);
1803 ptlrpc_req_finished(req);
1807 if (rc == 0 && !eof)
1810 #ifdef HAVE_SERVER_SUPPORT
1811 if (new_config != NULL) {
1812 /* recent_nodemap cannot be used after set_active/dealloc */
1814 nodemap_config_set_active_mgc(new_config);
1816 nodemap_config_dealloc(new_config);
1821 for (i = 0; i < nrpages; i++) {
1822 if (pages[i] == NULL)
1824 __free_page(pages[i]);
1826 OBD_FREE_PTR_ARRAY_LARGE(pages, nrpages);
1831 #ifdef HAVE_SERVER_SUPPORT
1832 static int mgc_barrier_glimpse_ast(struct ldlm_lock *lock, void *data)
1834 struct config_llog_data *cld = lock->l_ast_data;
1838 if (cld->cld_stopping)
1841 rc = barrier_handler(s2lsi(cld->cld_cfg.cfg_sb)->lsi_dt_dev,
1842 (struct ptlrpc_request *)data);
1847 /* Copy a remote log locally */
1848 static int mgc_llog_local_copy(const struct lu_env *env,
1849 struct obd_device *obd,
1850 struct llog_ctxt *rctxt,
1851 struct llog_ctxt *lctxt, char *logname)
1859 * - copy it to backup using llog_backup()
1860 * - copy remote llog to logname using llog_backup()
1861 * - if failed then move bakup to logname again
1864 OBD_ALLOC(temp_log, strlen(logname) + 2);
1867 sprintf(temp_log, "%sT", logname);
1869 /* make a copy of local llog at first */
1870 rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1871 if (rc < 0 && rc != -ENOENT)
1873 /* copy remote llog to the local copy */
1874 rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1875 if (rc == -ENOENT) {
1876 /* no remote llog, delete local one too */
1877 llog_erase(env, lctxt, NULL, logname);
1878 } else if (rc < 0) {
1879 /* error during backup, get local one back from the copy */
1880 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1882 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1883 obd->obd_name, logname, rc);
1885 llog_erase(env, lctxt, NULL, temp_log);
1886 OBD_FREE(temp_log, strlen(logname) + 2);
1891 mgc_process_server_cfg_log(struct lu_env *env, struct llog_ctxt **ctxt,
1892 struct lustre_sb_info *lsi, struct obd_device *mgc,
1893 struct config_llog_data *cld, int local_only)
1895 struct llog_ctxt *lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1896 struct client_obd *cli = &mgc->u.cli;
1899 /* Copy the setup log locally if we can. Don't mess around if we're
1900 * running an MGS though (logs are already local). */
1901 if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1902 cli->cl_mgc_configs_dir != NULL &&
1903 lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1905 if (!local_only && !lsi->lsi_dt_dev->dd_rdonly) {
1906 /* Only try to copy log if we have the lock. */
1907 CDEBUG(D_INFO, "%s: copy local log %s\n",
1908 mgc->obd_name, cld->cld_logname);
1910 rc = mgc_llog_local_copy(env, mgc, *ctxt, lctxt,
1913 lsi->lsi_flags &= ~LDD_F_NO_LOCAL_LOGS;
1915 if (local_only || rc) {
1916 if (unlikely(lsi->lsi_flags & LDD_F_NO_LOCAL_LOGS) ||
1918 CWARN("%s: local log %s are not valid and/or remote logs are not accessbile rc = %d\n",
1919 mgc->obd_name, cld->cld_logname, rc);
1920 GOTO(out_pop, rc = -EIO);
1923 if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1924 llog_is_empty(env, lctxt, cld->cld_logname)) {
1925 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS log %s and no local copy.\n",
1927 GOTO(out_pop, rc = -ENOENT);
1929 CDEBUG(D_MGC, "%s: Failed to get MGS log %s, using local copy for now, will try to update later.\n",
1930 mgc->obd_name, cld->cld_logname);
1933 /* Now, whether we copied or not, start using the local llog.
1934 * If we failed to copy, we'll start using whatever the old
1936 llog_ctxt_put(*ctxt);
1939 } else if (local_only) { /* no local log at client side */
1940 GOTO(out_pop, rc = -EIO);
1944 if (lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1945 lsi->lsi_dt_dev->dd_rdonly) {
1946 struct llog_ctxt *rctxt;
1948 /* Under readonly mode, we may have no local copy or local
1949 * copy is incomplete, so try to use remote llog firstly. */
1950 rctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1953 rc = class_config_parse_llog(env, rctxt, cld->cld_logname,
1955 llog_ctxt_put(rctxt);
1959 __llog_ctxt_put(env, lctxt);
1962 #else /* !HAVE_SERVER_SUPPORT */
1963 #define mgc_barrier_glimpse_ast NULL
1964 #endif /* HAVE_SERVER_SUPPORT */
1966 /* local_only means it cannot get remote llogs */
1967 static int mgc_process_cfg_log(struct obd_device *mgc,
1968 struct config_llog_data *cld, int local_only)
1970 struct llog_ctxt *ctxt;
1971 struct lustre_sb_info *lsi = NULL;
1977 LASSERT(mutex_is_locked(&cld->cld_lock));
1979 if (cld->cld_cfg.cfg_sb)
1980 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1986 rc = lu_env_init(env, LCT_MG_THREAD);
1990 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1992 #ifdef HAVE_SERVER_SUPPORT
1993 rc = mgc_process_server_cfg_log(env, &ctxt, lsi, mgc, cld,
1995 if (rc == -EIO && local_only)
1997 if (rc && rc != -ENOENT)
2002 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
2006 * update settings on existing OBDs.
2007 * the logname must be <fsname>-sptlrpc
2009 if (rc == 0 && cld_is_sptlrpc(cld))
2010 class_notify_sptlrpc_conf(cld->cld_logname,
2011 strlen(cld->cld_logname) -
2012 strlen("-sptlrpc"));
2016 __llog_ctxt_put(env, ctxt);
2023 static bool mgc_import_in_recovery(struct obd_import *imp)
2025 bool in_recovery = true;
2027 spin_lock(&imp->imp_lock);
2028 if (imp->imp_state == LUSTRE_IMP_FULL ||
2029 imp->imp_state == LUSTRE_IMP_CLOSED)
2030 in_recovery = false;
2031 spin_unlock(&imp->imp_lock);
2037 * Get a configuration log from the MGS and process it.
2039 * This function is called for both clients and servers to process the
2040 * configuration log from the MGS. The MGC enqueues a DLM lock on the
2041 * log from the MGS, and if the lock gets revoked the MGC will be notified
2042 * by the lock cancellation callback that the config log has changed,
2043 * and will enqueue another MGS lock on it, and then continue processing
2044 * the new additions to the end of the log.
2046 * Since the MGC import is not replayable, if the import is being evicted
2047 * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
2048 * the log until recovery is finished or the import is closed.
2050 * Make a local copy of the log before parsing it if appropriate (non-MGS
2051 * server) so that the server can start even when the MGS is down.
2053 * There shouldn't be multiple processes running process_log at once --
2054 * sounds like badness. It actually might be fine, as long as they're not
2055 * trying to update from the same log simultaneously, in which case we
2056 * should use a per-log semaphore instead of cld_lock.
2058 * \param[in] mgc MGC device by which to fetch the configuration log
2059 * \param[in] cld log processing state (stored in lock callback data)
2061 * \retval 0 on success
2062 * \retval negative errno on failure
2064 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
2066 struct lustre_handle lockh = { 0 };
2067 __u64 flags = LDLM_FL_NO_LRU;
2072 LASSERT(cld != NULL);
2074 /* I don't want multiple processes running process_log at once --
2075 sounds like badness. It actually might be fine, as long as
2076 we're not trying to update from the same log
2077 simultaneously (in which case we should use a per-log sem.) */
2079 mutex_lock(&cld->cld_lock);
2080 if (cld->cld_stopping) {
2081 mutex_unlock(&cld->cld_lock);
2085 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2087 CDEBUG(D_MGC, "Process log %s-%016lx from %d\n", cld->cld_logname,
2088 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2090 /* Get the cfg lock on the llog */
2091 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2093 cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
2094 cld, 0, NULL, &lockh);
2096 /* Get the cld, it will be released in mgc_blocking_ast. */
2097 config_log_get(cld);
2098 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2099 LASSERT(!lustre_handle_is_used(&cld->cld_lockh));
2101 cld->cld_lockh = lockh;
2103 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2104 cld->cld_lockh.cookie = 0;
2106 if (rcl == -ESHUTDOWN &&
2107 atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2108 struct obd_import *imp;
2109 long timeout = cfs_time_seconds(obd_timeout);
2111 mutex_unlock(&cld->cld_lock);
2112 imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2114 /* Let's force the pinger, and wait the import to be
2115 * connected, note: since mgc import is non-replayable,
2116 * and even the import state is disconnected, it does
2117 * not mean the "recovery" is stopped, so we will keep
2118 * waitting until timeout or the import state is
2120 ptlrpc_pinger_force(imp);
2122 wait_event_idle_timeout(imp->imp_recovery_waitq,
2123 !mgc_import_in_recovery(imp),
2126 if (imp->imp_state == LUSTRE_IMP_FULL) {
2130 mutex_lock(&cld->cld_lock);
2131 /* unlock/lock mutex, so check stopping again */
2132 if (cld->cld_stopping) {
2133 mutex_unlock(&cld->cld_lock);
2136 spin_lock(&config_list_lock);
2137 cld->cld_lostlock = 1;
2138 spin_unlock(&config_list_lock);
2141 /* mark cld_lostlock so that it will requeue
2142 * after MGC becomes available. */
2143 spin_lock(&config_list_lock);
2144 cld->cld_lostlock = 1;
2145 spin_unlock(&config_list_lock);
2149 if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2151 rc = mgc_process_recover_nodemap_log(mgc, cld);
2152 else if (cld_is_nodemap(cld))
2155 } else if (!cld_is_barrier(cld)) {
2156 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2159 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2160 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2162 /* Now drop the lock so MGS can revoke it */
2164 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2166 CERROR("Can't drop cfg lock: %d\n", rcl);
2168 mutex_unlock(&cld->cld_lock);
2170 /* requeue nodemap lock immediately if transfer was interrupted */
2171 if ((cld_is_nodemap(cld) && rc == -EAGAIN) ||
2172 (cld_is_recover(cld) && rc)) {
2173 if (cld_is_recover(cld))
2174 CWARN("%s: IR log %s failed, not fatal: rc = %d\n",
2175 mgc->obd_name, cld->cld_logname, rc);
2176 mgc_requeue_add(cld);
2184 /** Called from lustre_process_log.
2185 * LCFG_LOG_START gets the config log from the MGS, processes it to start
2186 * any services, and adds it to the list logs to watch (follow).
2188 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2190 struct lustre_cfg *lcfg = buf;
2191 struct config_llog_instance *cfg = NULL;
2196 switch(lcfg->lcfg_command) {
2197 #ifdef HAVE_SERVER_SUPPORT
2198 case LCFG_LOV_ADD_OBD: {
2199 /* Overloading this cfg command: register a new target */
2200 struct mgs_target_info *mti;
2202 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2203 sizeof(struct mgs_target_info))
2204 GOTO(out, rc = -EINVAL);
2206 mti = lustre_cfg_buf(lcfg, 1);
2207 CDEBUG(D_MGC, "add_target %s %#x\n",
2208 mti->mti_svname, mti->mti_flags);
2209 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2212 case LCFG_LOV_DEL_OBD:
2213 /* Unregister has no meaning at the moment. */
2214 CERROR("lov_del_obd unimplemented\n");
2218 case LCFG_SPTLRPC_CONF: {
2219 rc = sptlrpc_process_config(lcfg);
2222 case LCFG_LOG_START: {
2223 struct config_llog_data *cld;
2224 struct super_block *sb;
2226 logname = lustre_cfg_string(lcfg, 1);
2227 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2228 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2230 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2233 /* We're only called through here on the initial mount */
2234 cld = config_log_add(obd, logname, cfg, sb);
2240 rc = mgc_process_log(obd, cld);
2241 if (rc == 0 && cld->cld_recover != NULL) {
2242 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2243 imp_connect_data, IMP_RECOV)) {
2244 rc = mgc_process_log(obd, cld->cld_recover);
2246 struct config_llog_data *cir;
2248 mutex_lock(&cld->cld_lock);
2249 cir = cld->cld_recover;
2250 cld->cld_recover = NULL;
2251 mutex_unlock(&cld->cld_lock);
2252 config_log_put(cir);
2256 CERROR("Cannot process recover llog %d\n", rc);
2259 if (rc == 0 && cld->cld_params != NULL) {
2260 rc = mgc_process_log(obd, cld->cld_params);
2261 if (rc == -ENOENT) {
2262 CDEBUG(D_MGC, "There is no params "
2263 "config file yet\n");
2266 /* params log is optional */
2268 CERROR("%s: can't process params llog: rc = %d\n",
2274 case LCFG_LOG_END: {
2275 logname = lustre_cfg_string(lcfg, 1);
2277 if (lcfg->lcfg_bufcount >= 2)
2278 cfg = (struct config_llog_instance *)lustre_cfg_buf(
2280 rc = config_log_end(logname, cfg);
2284 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2285 GOTO(out, rc = -EINVAL);
2293 static const struct obd_ops mgc_obd_ops = {
2294 .o_owner = THIS_MODULE,
2295 .o_setup = mgc_setup,
2296 .o_precleanup = mgc_precleanup,
2297 .o_cleanup = mgc_cleanup,
2298 .o_add_conn = client_import_add_conn,
2299 .o_del_conn = client_import_del_conn,
2300 .o_connect = client_connect_import,
2301 .o_disconnect = client_disconnect_export,
2302 .o_set_info_async = mgc_set_info_async,
2303 .o_get_info = mgc_get_info,
2304 .o_import_event = mgc_import_event,
2305 .o_process_config = mgc_process_config,
2308 static int mgc_param_requeue_timeout_min_set(const char *val,
2309 cfs_kernel_param_arg_t *kp)
2314 rc = kstrtouint(val, 0, &num);
2320 mgc_requeue_timeout_min = num;
2325 static const struct kernel_param_ops param_ops_requeue_timeout_min = {
2326 .set = mgc_param_requeue_timeout_min_set,
2327 .get = param_get_uint,
2330 #define param_check_requeue_timeout_min(name, p) \
2331 __param_check(name, p, unsigned int)
2333 unsigned int mgc_requeue_timeout_min = MGC_TIMEOUT_MIN_SECONDS;
2334 #ifdef HAVE_KERNEL_PARAM_OPS
2335 module_param(mgc_requeue_timeout_min, requeue_timeout_min, 0644);
2337 module_param_call(mgc_requeue_timeout_min, mgc_param_requeue_timeout_min_set,
2338 param_get_uint, ¶m_ops_requeue_timeout_min, 0644);
2340 MODULE_PARM_DESC(mgc_requeue_timeout_min, "Minimal requeue time to refresh logs");
2342 static int __init mgc_init(void)
2344 return class_register_type(&mgc_obd_ops, NULL, false,
2345 LUSTRE_MGC_NAME, NULL);
2348 static void __exit mgc_exit(void)
2350 class_unregister_type(LUSTRE_MGC_NAME);
2353 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2354 MODULE_DESCRIPTION("Lustre Management Client");
2355 MODULE_VERSION(LUSTRE_VERSION_STRING);
2356 MODULE_LICENSE("GPL");
2358 module_init(mgc_init);
2359 module_exit(mgc_exit);