lustre/mgc/mgc_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   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.
   9  *
  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).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/mgc/mgc_request.c
  37  *
  38  * Author: Nathan Rutman <nathan@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_MGC
  42 #define D_MGC D_CONFIG /*|D_WARNING*/
  43
  44 #include <linux/module.h>
  45 #include <obd_class.h>
  46 #include <lustre_dlm.h>
  47 #include <lprocfs_status.h>
  48 #include <lustre_log.h>
  49 #include <lustre_disk.h>
  50 #include <dt_object.h>
  51
  52 #include "mgc_internal.h"
  53
  54 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
  55                           int type)
  56 {
  57         __u64 resname = 0;
  58
  59         if (len > sizeof(resname)) {
  60                 CERROR("name too long: %s\n", name);
  61                 return -EINVAL;
  62         }
  63         if (len <= 0) {
  64                 CERROR("missing name: %s\n", name);
  65                 return -EINVAL;
  66         }
  67         memcpy(&resname, name, len);
  68
  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 */
  73         switch(type) {
  74         case CONFIG_T_CONFIG:
  75         case CONFIG_T_SPTLRPC:
  76                 resname = 0;
  77                 break;
  78         case CONFIG_T_RECOVER:
  79         case CONFIG_T_PARAMS:
  80                 resname = type;
  81                 break;
  82         default:
  83                 LBUG();
  84         }
  85         res_id->name[1] = cpu_to_le64(resname);
  86         CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
  87                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  88         return 0;
  89 }
  90
  91 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
  92 {
  93         /* fsname is at most 8 chars long, maybe contain "-".
  94          * e.g. "lustre", "SUN-000" */
  95         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
  96 }
  97 EXPORT_SYMBOL(mgc_fsname2resid);
  98
  99 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
 100 {
 101         char *name_end;
 102         int len;
 103
 104         /* logname consists of "fsname-nodetype".
 105          * e.g. "lustre-MDT0001", "SUN-000-client"
 106          * there is an exception: llog "params" */
 107         name_end = strrchr(logname, '-');
 108         if (!name_end)
 109                 len = strlen(logname);
 110         else
 111                 len = name_end - logname;
 112         return mgc_name2resid(logname, len, res_id, type);
 113 }
 114
 115 /********************** config llog list **********************/
 116 static CFS_LIST_HEAD(config_llog_list);
 117 static DEFINE_SPINLOCK(config_list_lock);
 118
 119 /* Take a reference to a config log */
 120 static int config_log_get(struct config_llog_data *cld)
 121 {
 122         ENTRY;
 123         atomic_inc(&cld->cld_refcount);
 124         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 125                 atomic_read(&cld->cld_refcount));
 126         RETURN(0);
 127 }
 128
 129 /* Drop a reference to a config log.  When no longer referenced,
 130    we can free the config log data */
 131 static void config_log_put(struct config_llog_data *cld)
 132 {
 133         ENTRY;
 134
 135         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 136                 atomic_read(&cld->cld_refcount));
 137         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 138
 139         /* spinlock to make sure no item with 0 refcount in the list */
 140         if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 141                 cfs_list_del(&cld->cld_list_chain);
 142                 spin_unlock(&config_list_lock);
 143
 144                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 145
 146                 if (cld->cld_recover)
 147                         config_log_put(cld->cld_recover);
 148                 if (cld->cld_sptlrpc)
 149                         config_log_put(cld->cld_sptlrpc);
 150                 if (cld->cld_params)
 151                         config_log_put(cld->cld_params);
 152                 if (cld_is_sptlrpc(cld))
 153                         sptlrpc_conf_log_stop(cld->cld_logname);
 154
 155                 class_export_put(cld->cld_mgcexp);
 156                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 157         }
 158
 159         EXIT;
 160 }
 161
 162 /* Find a config log by name */
 163 static
 164 struct config_llog_data *config_log_find(char *logname,
 165                                          struct config_llog_instance *cfg)
 166 {
 167         struct config_llog_data *cld;
 168         struct config_llog_data *found = NULL;
 169         void *                   instance;
 170         ENTRY;
 171
 172         LASSERT(logname != NULL);
 173
 174         instance = cfg ? cfg->cfg_instance : NULL;
 175         spin_lock(&config_list_lock);
 176         cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 177                 /* check if instance equals */
 178                 if (instance != cld->cld_cfg.cfg_instance)
 179                         continue;
 180
 181                 /* instance may be NULL, should check name */
 182                 if (strcmp(logname, cld->cld_logname) == 0) {
 183                         found = cld;
 184                         break;
 185                 }
 186         }
 187         if (found) {
 188                 atomic_inc(&found->cld_refcount);
 189                 LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0);
 190         }
 191         spin_unlock(&config_list_lock);
 192         RETURN(found);
 193 }
 194
 195 static
 196 struct config_llog_data *do_config_log_add(struct obd_device *obd,
 197                                            char *logname,
 198                                            int type,
 199                                            struct config_llog_instance *cfg,
 200                                            struct super_block *sb)
 201 {
 202         struct config_llog_data *cld;
 203         int                      rc;
 204         ENTRY;
 205
 206         CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
 207                cfg ? cfg->cfg_instance : 0);
 208
 209         OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
 210         if (!cld)
 211                 RETURN(ERR_PTR(-ENOMEM));
 212
 213         strcpy(cld->cld_logname, logname);
 214         if (cfg)
 215                 cld->cld_cfg = *cfg;
 216         else
 217                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 218         mutex_init(&cld->cld_lock);
 219         cld->cld_cfg.cfg_last_idx = 0;
 220         cld->cld_cfg.cfg_flags = 0;
 221         cld->cld_cfg.cfg_sb = sb;
 222         cld->cld_type = type;
 223         atomic_set(&cld->cld_refcount, 1);
 224
 225         /* Keep the mgc around until we are done */
 226         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 227
 228         if (cld_is_sptlrpc(cld)) {
 229                 sptlrpc_conf_log_start(logname);
 230                 cld->cld_cfg.cfg_obdname = obd->obd_name;
 231         }
 232
 233         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 234
 235         spin_lock(&config_list_lock);
 236         cfs_list_add(&cld->cld_list_chain, &config_llog_list);
 237         spin_unlock(&config_list_lock);
 238
 239         if (rc) {
 240                 config_log_put(cld);
 241                 RETURN(ERR_PTR(rc));
 242         }
 243
 244         if (cld_is_sptlrpc(cld)) {
 245                 rc = mgc_process_log(obd, cld);
 246                 if (rc && rc != -ENOENT)
 247                         CERROR("failed processing sptlrpc log: %d\n", rc);
 248         }
 249
 250         RETURN(cld);
 251 }
 252
 253 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
 254         char *fsname,
 255         struct config_llog_instance *cfg,
 256         struct super_block *sb)
 257 {
 258         struct config_llog_instance lcfg = *cfg;
 259         struct lustre_sb_info *lsi = s2lsi(sb);
 260         struct config_llog_data *cld;
 261         char logname[32];
 262
 263         if (IS_OST(lsi))
 264                 return NULL;
 265
 266         /* for osp-on-ost, see lustre_start_osp() */
 267         if (IS_MDT(lsi) && lcfg.cfg_instance)
 268                 return NULL;
 269
 270         /* we have to use different llog for clients and mdts for cmd
 271          * where only clients are notified if one of cmd server restarts */
 272         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 273         strcpy(logname, fsname);
 274         if (IS_SERVER(lsi)) { /* mdt */
 275                 LASSERT(lcfg.cfg_instance == NULL);
 276                 lcfg.cfg_instance = sb;
 277                 strcat(logname, "-mdtir");
 278         } else {
 279                 LASSERT(lcfg.cfg_instance != NULL);
 280                 strcat(logname, "-cliir");
 281         }
 282
 283         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
 284         return cld;
 285 }
 286
 287 static struct config_llog_data *config_params_log_add(struct obd_device *obd,
 288         struct config_llog_instance *cfg, struct super_block *sb)
 289 {
 290         struct config_llog_instance     lcfg = *cfg;
 291         struct config_llog_data         *cld;
 292
 293         lcfg.cfg_instance = sb;
 294
 295         cld = do_config_log_add(obd, PARAMS_FILENAME, CONFIG_T_PARAMS,
 296                                 &lcfg, sb);
 297
 298         return cld;
 299 }
 300
 301 /** Add this log to the list of active logs watched by an MGC.
 302  * Active means we're watching for updates.
 303  * We have one active log per "mount" - client instance or servername.
 304  * Each instance may be at a different point in the log.
 305  */
 306 static int config_log_add(struct obd_device *obd, char *logname,
 307                           struct config_llog_instance *cfg,
 308                           struct super_block *sb)
 309 {
 310         struct lustre_sb_info   *lsi = s2lsi(sb);
 311         struct config_llog_data *cld;
 312         struct config_llog_data *sptlrpc_cld;
 313         struct config_llog_data *params_cld;
 314         char                    seclogname[32];
 315         char                    *ptr;
 316         int                     rc;
 317         ENTRY;
 318
 319         CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
 320
 321         /*
 322          * for each regular log, the depended sptlrpc log name is
 323          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 324          */
 325         ptr = strrchr(logname, '-');
 326         if (ptr == NULL || ptr - logname > 8) {
 327                 CERROR("logname %s is too long\n", logname);
 328                 RETURN(-EINVAL);
 329         }
 330
 331         memcpy(seclogname, logname, ptr - logname);
 332         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 333
 334         sptlrpc_cld = config_log_find(seclogname, NULL);
 335         if (sptlrpc_cld == NULL) {
 336                 sptlrpc_cld = do_config_log_add(obd, seclogname,
 337                                                 CONFIG_T_SPTLRPC, NULL, NULL);
 338                 if (IS_ERR(sptlrpc_cld)) {
 339                         CERROR("can't create sptlrpc log: %s\n", seclogname);
 340                         GOTO(out_err, rc = PTR_ERR(sptlrpc_cld));
 341                 }
 342         }
 343         params_cld = config_params_log_add(obd, cfg, sb);
 344         if (IS_ERR(params_cld)) {
 345                 rc = PTR_ERR(params_cld);
 346                 CERROR("%s: can't create params log: rc = %d\n",
 347                        obd->obd_name, rc);
 348                 GOTO(out_err1, rc);
 349         }
 350
 351         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 352         if (IS_ERR(cld)) {
 353                 CERROR("can't create log: %s\n", logname);
 354                 GOTO(out_err2, rc = PTR_ERR(cld));
 355         }
 356
 357         cld->cld_sptlrpc = sptlrpc_cld;
 358         cld->cld_params = params_cld;
 359
 360         LASSERT(lsi->lsi_lmd);
 361         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
 362                 struct config_llog_data *recover_cld;
 363                 ptr = strrchr(seclogname, '-');
 364                 if (ptr != NULL) {
 365                         *ptr = 0;
 366                 }
 367                 else {
 368                         CERROR("sptlrpc log name not correct: %s", seclogname);
 369                         config_log_put(cld);
 370                         RETURN(-EINVAL);
 371                 }
 372                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 373                 if (IS_ERR(recover_cld))
 374                         GOTO(out_err3, rc = PTR_ERR(recover_cld));
 375                 cld->cld_recover = recover_cld;
 376         }
 377
 378         RETURN(0);
 379
 380 out_err3:
 381         config_log_put(cld);
 382
 383 out_err2:
 384         config_log_put(params_cld);
 385
 386 out_err1:
 387         config_log_put(sptlrpc_cld);
 388
 389 out_err:
 390         RETURN(rc);
 391 }
 392
 393 DEFINE_MUTEX(llog_process_lock);
 394
 395 /** Stop watching for updates on this log.
 396  */
 397 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 398 {
 399         struct config_llog_data *cld;
 400         struct config_llog_data *cld_sptlrpc = NULL;
 401         struct config_llog_data *cld_params = NULL;
 402         struct config_llog_data *cld_recover = NULL;
 403         int rc = 0;
 404         ENTRY;
 405
 406         cld = config_log_find(logname, cfg);
 407         if (cld == NULL)
 408                 RETURN(-ENOENT);
 409
 410         mutex_lock(&cld->cld_lock);
 411         /*
 412          * if cld_stopping is set, it means we didn't start the log thus
 413          * not owning the start ref. this can happen after previous umount:
 414          * the cld still hanging there waiting for lock cancel, and we
 415          * remount again but failed in the middle and call log_end without
 416          * calling start_log.
 417          */
 418         if (unlikely(cld->cld_stopping)) {
 419                 mutex_unlock(&cld->cld_lock);
 420                 /* drop the ref from the find */
 421                 config_log_put(cld);
 422                 RETURN(rc);
 423         }
 424
 425         cld->cld_stopping = 1;
 426
 427         cld_recover = cld->cld_recover;
 428         cld->cld_recover = NULL;
 429         mutex_unlock(&cld->cld_lock);
 430
 431         if (cld_recover) {
 432                 mutex_lock(&cld_recover->cld_lock);
 433                 cld_recover->cld_stopping = 1;
 434                 mutex_unlock(&cld_recover->cld_lock);
 435                 config_log_put(cld_recover);
 436         }
 437
 438         spin_lock(&config_list_lock);
 439         cld_sptlrpc = cld->cld_sptlrpc;
 440         cld->cld_sptlrpc = NULL;
 441         cld_params = cld->cld_params;
 442         cld->cld_params = NULL;
 443         spin_unlock(&config_list_lock);
 444
 445         if (cld_sptlrpc)
 446                 config_log_put(cld_sptlrpc);
 447
 448         if (cld_params) {
 449                 mutex_lock(&cld_params->cld_lock);
 450                 cld_params->cld_stopping = 1;
 451                 mutex_unlock(&cld_params->cld_lock);
 452                 config_log_put(cld_params);
 453         }
 454
 455         /* drop the ref from the find */
 456         config_log_put(cld);
 457         /* drop the start ref */
 458         config_log_put(cld);
 459
 460         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 461                rc);
 462         RETURN(rc);
 463 }
 464
 465 #ifdef LPROCFS
 466 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 467 {
 468         struct obd_device       *obd = data;
 469         struct obd_import       *imp = obd->u.cli.cl_import;
 470         struct obd_connect_data *ocd = &imp->imp_connect_data;
 471         struct config_llog_data *cld;
 472         ENTRY;
 473
 474         seq_printf(m, "imperative_recovery: %s\n",
 475                    OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 476         seq_printf(m, "client_state:\n");
 477
 478         spin_lock(&config_list_lock);
 479         cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 480                 if (cld->cld_recover == NULL)
 481                         continue;
 482                 seq_printf(m,  "    - { client: %s, nidtbl_version: %u }\n",
 483                                cld->cld_logname,
 484                                cld->cld_recover->cld_cfg.cfg_last_idx);
 485         }
 486         spin_unlock(&config_list_lock);
 487
 488         RETURN(0);
 489 }
 490 #endif
 491
 492 /* reenqueue any lost locks */
 493 #define RQ_RUNNING 0x1
 494 #define RQ_NOW     0x2
 495 #define RQ_LATER   0x4
 496 #define RQ_STOP    0x8
 497 static int                    rq_state = 0;
 498 static wait_queue_head_t      rq_waitq;
 499 static DECLARE_COMPLETION(rq_exit);
 500
 501 static void do_requeue(struct config_llog_data *cld)
 502 {
 503         int rc = 0;
 504         ENTRY;
 505
 506         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 507
 508         /* Do not run mgc_process_log on a disconnected export or an
 509          * export which is being disconnected. Take the client
 510          * semaphore to make the check non-racy. */
 511         down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 512         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 513                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 514                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 515                 if (rc && rc != -ENOENT)
 516                         CERROR("failed processing log: %d\n", rc);
 517         } else {
 518                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 519                        cld->cld_logname);
 520         }
 521         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 522
 523         EXIT;
 524 }
 525
 526 /* this timeout represents how many seconds MGC should wait before
 527  * requeue config and recover lock to the MGS. We need to randomize this
 528  * in order to not flood the MGS.
 529  */
 530 #define MGC_TIMEOUT_MIN_SECONDS   5
 531 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
 532
 533 static int mgc_requeue_thread(void *data)
 534 {
 535         int rc = 0;
 536         ENTRY;
 537
 538         CDEBUG(D_MGC, "Starting requeue thread\n");
 539
 540         /* Keep trying failed locks periodically */
 541         spin_lock(&config_list_lock);
 542         rq_state |= RQ_RUNNING;
 543         while (1) {
 544                 struct l_wait_info lwi;
 545                 struct config_llog_data *cld, *cld_prev;
 546                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
 547                 int stopped = !!(rq_state & RQ_STOP);
 548                 int to;
 549
 550                 /* Any new or requeued lostlocks will change the state */
 551                 rq_state &= ~(RQ_NOW | RQ_LATER);
 552                 spin_unlock(&config_list_lock);
 553
 554                 /* Always wait a few seconds to allow the server who
 555                    caused the lock revocation to finish its setup, plus some
 556                    random so everyone doesn't try to reconnect at once. */
 557                 to = MGC_TIMEOUT_MIN_SECONDS * HZ;
 558                 to += rand * HZ / 100; /* rand is centi-seconds */
 559                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 560                 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi);
 561
 562                 /*
 563                  * iterate & processing through the list. for each cld, process
 564                  * its depending sptlrpc cld firstly (if any) and then itself.
 565                  *
 566                  * it's guaranteed any item in the list must have
 567                  * reference > 0; and if cld_lostlock is set, at
 568                  * least one reference is taken by the previous enqueue.
 569                  */
 570                 cld_prev = NULL;
 571
 572                 spin_lock(&config_list_lock);
 573                 cfs_list_for_each_entry(cld, &config_llog_list,
 574                                         cld_list_chain) {
 575                         if (!cld->cld_lostlock)
 576                                 continue;
 577
 578                         spin_unlock(&config_list_lock);
 579
 580                         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 581
 582                         /* Whether we enqueued again or not in mgc_process_log,
 583                          * we're done with the ref from the old enqueue */
 584                         if (cld_prev)
 585                                 config_log_put(cld_prev);
 586                         cld_prev = cld;
 587
 588                         cld->cld_lostlock = 0;
 589                         if (likely(!stopped))
 590                                 do_requeue(cld);
 591
 592                         spin_lock(&config_list_lock);
 593                 }
 594                 spin_unlock(&config_list_lock);
 595                 if (cld_prev)
 596                         config_log_put(cld_prev);
 597
 598                 /* break after scanning the list so that we can drop
 599                  * refcount to losing lock clds */
 600                 if (unlikely(stopped)) {
 601                         spin_lock(&config_list_lock);
 602                         break;
 603                 }
 604
 605                 /* Wait a bit to see if anyone else needs a requeue */
 606                 lwi = (struct l_wait_info) { 0 };
 607                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 608                              &lwi);
 609                 spin_lock(&config_list_lock);
 610         }
 611         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 612         rq_state &= ~RQ_RUNNING;
 613         spin_unlock(&config_list_lock);
 614
 615         complete(&rq_exit);
 616
 617         CDEBUG(D_MGC, "Ending requeue thread\n");
 618         RETURN(rc);
 619 }
 620
 621 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 622    We are responsible for dropping the config log reference from here on out. */
 623 static void mgc_requeue_add(struct config_llog_data *cld)
 624 {
 625         ENTRY;
 626
 627         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 628                 cld->cld_logname, atomic_read(&cld->cld_refcount),
 629                 cld->cld_stopping, rq_state);
 630         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 631
 632         mutex_lock(&cld->cld_lock);
 633         if (cld->cld_stopping || cld->cld_lostlock) {
 634                 mutex_unlock(&cld->cld_lock);
 635                 RETURN_EXIT;
 636         }
 637         /* this refcount will be released in mgc_requeue_thread. */
 638         config_log_get(cld);
 639         cld->cld_lostlock = 1;
 640         mutex_unlock(&cld->cld_lock);
 641
 642         /* Hold lock for rq_state */
 643         spin_lock(&config_list_lock);
 644         if (rq_state & RQ_STOP) {
 645                 spin_unlock(&config_list_lock);
 646                 cld->cld_lostlock = 0;
 647                 config_log_put(cld);
 648         } else {
 649                 rq_state |= RQ_NOW;
 650                 spin_unlock(&config_list_lock);
 651                 wake_up(&rq_waitq);
 652         }
 653         EXIT;
 654 }
 655
 656 /********************** class fns **********************/
 657 static int mgc_local_llog_init(const struct lu_env *env,
 658                                struct obd_device *obd,
 659                                struct obd_device *disk)
 660 {
 661         struct llog_ctxt        *ctxt;
 662         int                      rc;
 663
 664         ENTRY;
 665
 666         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
 667                         &llog_osd_ops);
 668         if (rc)
 669                 RETURN(rc);
 670
 671         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 672         LASSERT(ctxt);
 673         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
 674         llog_ctxt_put(ctxt);
 675
 676         RETURN(0);
 677 }
 678
 679 static int mgc_local_llog_fini(const struct lu_env *env,
 680                                struct obd_device *obd)
 681 {
 682         struct llog_ctxt *ctxt;
 683
 684         ENTRY;
 685
 686         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 687         llog_cleanup(env, ctxt);
 688
 689         RETURN(0);
 690 }
 691
 692 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb)
 693 {
 694         struct lustre_sb_info   *lsi = s2lsi(sb);
 695         struct client_obd       *cli = &obd->u.cli;
 696         struct lu_fid            rfid, fid;
 697         struct dt_object        *root, *dto;
 698         struct lu_env           *env;
 699         int                      rc = 0;
 700
 701         ENTRY;
 702
 703         LASSERT(lsi);
 704         LASSERT(lsi->lsi_dt_dev);
 705
 706         OBD_ALLOC_PTR(env);
 707         if (env == NULL)
 708                 RETURN(-ENOMEM);
 709
 710         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
 711         mutex_lock(&cli->cl_mgc_mutex);
 712
 713         cfs_cleanup_group_info();
 714
 715         /* Setup the configs dir */
 716         rc = lu_env_init(env, LCT_MG_THREAD);
 717         if (rc)
 718                 GOTO(out_err, rc);
 719
 720         fid.f_seq = FID_SEQ_LOCAL_NAME;
 721         fid.f_oid = 1;
 722         fid.f_ver = 0;
 723         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
 724                                     &cli->cl_mgc_los);
 725         if (rc)
 726                 GOTO(out_env, rc);
 727
 728         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
 729         if (rc)
 730                 GOTO(out_env, rc);
 731
 732         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
 733                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
 734         if (unlikely(IS_ERR(root)))
 735                 GOTO(out_los, rc = PTR_ERR(root));
 736
 737         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
 738                                         MOUNT_CONFIGS_DIR,
 739                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
 740         lu_object_put_nocache(env, &root->do_lu);
 741         if (IS_ERR(dto))
 742                 GOTO(out_los, rc = PTR_ERR(dto));
 743
 744         cli->cl_mgc_configs_dir = dto;
 745
 746         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
 747         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
 748         if (rc)
 749                 GOTO(out_llog, rc);
 750
 751         /* We take an obd ref to insure that we can't get to mgc_cleanup
 752          * without calling mgc_fs_cleanup first. */
 753         class_incref(obd, "mgc_fs", obd);
 754
 755         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 756         EXIT;
 757 out_llog:
 758         if (rc) {
 759                 lu_object_put(env, &cli->cl_mgc_configs_dir->do_lu);
 760                 cli->cl_mgc_configs_dir = NULL;
 761         }
 762 out_los:
 763         if (rc < 0) {
 764                 local_oid_storage_fini(env, cli->cl_mgc_los);
 765                 cli->cl_mgc_los = NULL;
 766                 mutex_unlock(&cli->cl_mgc_mutex);
 767         }
 768 out_env:
 769         lu_env_fini(env);
 770 out_err:
 771         OBD_FREE_PTR(env);
 772         return rc;
 773 }
 774
 775 static int mgc_fs_cleanup(struct obd_device *obd)
 776 {
 777         struct lu_env            env;
 778         struct client_obd       *cli = &obd->u.cli;
 779         int                      rc;
 780
 781         ENTRY;
 782
 783         LASSERT(cli->cl_mgc_los != NULL);
 784
 785         rc = lu_env_init(&env, LCT_MG_THREAD);
 786         if (rc)
 787                 GOTO(unlock, rc);
 788
 789         mgc_local_llog_fini(&env, obd);
 790
 791         lu_object_put_nocache(&env, &cli->cl_mgc_configs_dir->do_lu);
 792         cli->cl_mgc_configs_dir = NULL;
 793
 794         local_oid_storage_fini(&env, cli->cl_mgc_los);
 795         cli->cl_mgc_los = NULL;
 796         lu_env_fini(&env);
 797
 798 unlock:
 799         class_decref(obd, "mgc_fs", obd);
 800         mutex_unlock(&cli->cl_mgc_mutex);
 801
 802         RETURN(0);
 803 }
 804
 805 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 806 {
 807         struct llog_ctxt        *ctxt;
 808         int                      rc;
 809
 810         ENTRY;
 811
 812         /* setup only remote ctxt, the local disk context is switched per each
 813          * filesystem during mgc_fs_setup() */
 814         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 815                         &llog_client_ops);
 816         if (rc)
 817                 RETURN(rc);
 818
 819         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 820         LASSERT(ctxt);
 821
 822         llog_initiator_connect(ctxt);
 823         llog_ctxt_put(ctxt);
 824
 825         RETURN(0);
 826 }
 827
 828 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 829 {
 830         struct llog_ctxt *ctxt;
 831
 832         ENTRY;
 833
 834         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 835         if (ctxt)
 836                 llog_cleanup(env, ctxt);
 837
 838         RETURN(0);
 839 }
 840
 841
 842 static atomic_t mgc_count = ATOMIC_INIT(0);
 843 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 844 {
 845         int rc = 0;
 846         ENTRY;
 847
 848         switch (stage) {
 849         case OBD_CLEANUP_EARLY:
 850                 break;
 851         case OBD_CLEANUP_EXPORTS:
 852                 if (atomic_dec_and_test(&mgc_count)) {
 853                         int running;
 854                         /* stop requeue thread */
 855                         spin_lock(&config_list_lock);
 856                         running = rq_state & RQ_RUNNING;
 857                         if (running)
 858                                 rq_state |= RQ_STOP;
 859                         spin_unlock(&config_list_lock);
 860                         if (running) {
 861                                 wake_up(&rq_waitq);
 862                                 wait_for_completion(&rq_exit);
 863                         }
 864                 }
 865                 obd_cleanup_client_import(obd);
 866                 rc = mgc_llog_fini(NULL, obd);
 867                 if (rc != 0)
 868                         CERROR("failed to cleanup llogging subsystems\n");
 869                 break;
 870         }
 871         RETURN(rc);
 872 }
 873
 874 static int mgc_cleanup(struct obd_device *obd)
 875 {
 876         int rc;
 877         ENTRY;
 878
 879         /* COMPAT_146 - old config logs may have added profiles we don't
 880            know about */
 881         if (obd->obd_type->typ_refcnt <= 1)
 882                 /* Only for the last mgc */
 883                 class_del_profiles();
 884
 885         lprocfs_obd_cleanup(obd);
 886         ptlrpcd_decref();
 887
 888         rc = client_obd_cleanup(obd);
 889         RETURN(rc);
 890 }
 891
 892 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 893 {
 894         struct task_struct      *task;
 895         int                      rc;
 896         ENTRY;
 897
 898         rc = ptlrpcd_addref();
 899         if (rc < 0)
 900                 RETURN(rc);
 901
 902         rc = client_obd_setup(obd, lcfg);
 903         if (rc)
 904                 GOTO(err_decref, rc);
 905
 906         rc = mgc_llog_init(NULL, obd);
 907         if (rc) {
 908                 CERROR("failed to setup llogging subsystems\n");
 909                 GOTO(err_cleanup, rc);
 910         }
 911
 912 #ifdef LPROCFS
 913         obd->obd_vars = lprocfs_mgc_obd_vars;
 914         lprocfs_seq_obd_setup(obd);
 915 #endif
 916         sptlrpc_lprocfs_cliobd_attach(obd);
 917
 918         if (atomic_inc_return(&mgc_count) == 1) {
 919                 rq_state = 0;
 920                 init_waitqueue_head(&rq_waitq);
 921
 922                 /* start requeue thread */
 923                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 924                 if (IS_ERR(task)) {
 925                         rc = PTR_ERR(task);
 926                         CERROR("%s: cannot start requeue thread: rc = %d; "
 927                                "no more log updates\n",
 928                                obd->obd_name, rc);
 929                         GOTO(err_cleanup, rc);
 930                 }
 931                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 932                 rc = 0;
 933         }
 934
 935         RETURN(rc);
 936
 937 err_cleanup:
 938         client_obd_cleanup(obd);
 939 err_decref:
 940         ptlrpcd_decref();
 941         RETURN(rc);
 942 }
 943
 944 /* based on ll_mdc_blocking_ast */
 945 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 946                             void *data, int flag)
 947 {
 948         struct lustre_handle lockh;
 949         struct config_llog_data *cld = (struct config_llog_data *)data;
 950         int rc = 0;
 951         ENTRY;
 952
 953         switch (flag) {
 954         case LDLM_CB_BLOCKING:
 955                 /* mgs wants the lock, give it up... */
 956                 LDLM_DEBUG(lock, "MGC blocking CB");
 957                 ldlm_lock2handle(lock, &lockh);
 958                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 959                 break;
 960         case LDLM_CB_CANCELING:
 961                 /* We've given up the lock, prepare ourselves to update. */
 962                 LDLM_DEBUG(lock, "MGC cancel CB");
 963
 964                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
 965                        PLDLMRES(lock->l_resource),
 966                        (char *)&lock->l_resource->lr_name.name[0]);
 967
 968                 if (!cld) {
 969                         CDEBUG(D_INFO, "missing data, won't requeue\n");
 970                         break;
 971                 }
 972
 973                 /* held at mgc_process_log(). */
 974                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
 975                 /* Are we done with this log? */
 976                 if (cld->cld_stopping) {
 977                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
 978                                 cld->cld_logname);
 979                         config_log_put(cld);
 980                         break;
 981                 }
 982                 /* Make sure not to re-enqueue when the mgc is stopping
 983                    (we get called from client_disconnect_export) */
 984                 if (!lock->l_conn_export ||
 985                     !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
 986                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
 987                                 cld->cld_logname);
 988                         config_log_put(cld);
 989                         break;
 990                 }
 991
 992                 /* Re-enqueue now */
 993                 mgc_requeue_add(cld);
 994                 config_log_put(cld);
 995                 break;
 996         default:
 997                 LBUG();
 998         }
 999
1000         RETURN(rc);
1001 }
1002
1003 /* Not sure where this should go... */
1004 #define  MGC_ENQUEUE_LIMIT 50
1005 #define  MGC_TARGET_REG_LIMIT 10
1006 #define  MGC_SEND_PARAM_LIMIT 10
1007
1008 /* Send parameter to MGS*/
1009 static int mgc_set_mgs_param(struct obd_export *exp,
1010                              struct mgs_send_param *msp)
1011 {
1012         struct ptlrpc_request *req;
1013         struct mgs_send_param *req_msp, *rep_msp;
1014         int rc;
1015         ENTRY;
1016
1017         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1018                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1019                                         MGS_SET_INFO);
1020         if (!req)
1021                 RETURN(-ENOMEM);
1022
1023         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1024         if (!req_msp) {
1025                 ptlrpc_req_finished(req);
1026                 RETURN(-ENOMEM);
1027         }
1028
1029         memcpy(req_msp, msp, sizeof(*req_msp));
1030         ptlrpc_request_set_replen(req);
1031
1032         /* Limit how long we will wait for the enqueue to complete */
1033         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1034         rc = ptlrpc_queue_wait(req);
1035         if (!rc) {
1036                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1037                 memcpy(msp, rep_msp, sizeof(*rep_msp));
1038         }
1039
1040         ptlrpc_req_finished(req);
1041
1042         RETURN(rc);
1043 }
1044
1045 /* Take a config lock so we can get cancel notifications */
1046 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
1047                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
1048                        __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
1049                        void *data, __u32 lvb_len, void *lvb_swabber,
1050                        struct lustre_handle *lockh)
1051 {
1052         struct config_llog_data *cld = (struct config_llog_data *)data;
1053         struct ldlm_enqueue_info einfo = {
1054                 .ei_type        = type,
1055                 .ei_mode        = mode,
1056                 .ei_cb_bl       = mgc_blocking_ast,
1057                 .ei_cb_cp       = ldlm_completion_ast,
1058         };
1059         struct ptlrpc_request *req;
1060         int short_limit = cld_is_sptlrpc(cld);
1061         int rc;
1062         ENTRY;
1063
1064         CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
1065                cld->cld_resid.name[0]);
1066
1067         /* We need a callback for every lockholder, so don't try to
1068            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1069         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1070                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1071                                         LDLM_ENQUEUE);
1072         if (req == NULL)
1073                 RETURN(-ENOMEM);
1074
1075         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1076         ptlrpc_request_set_replen(req);
1077
1078         /* check if this is server or client */
1079         if (cld->cld_cfg.cfg_sb) {
1080                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1081                 if (lsi && IS_SERVER(lsi))
1082                         short_limit = 1;
1083         }
1084         /* Limit how long we will wait for the enqueue to complete */
1085         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1086         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1087                               NULL, 0, LVB_T_NONE, lockh, 0);
1088         /* A failed enqueue should still call the mgc_blocking_ast,
1089            where it will be requeued if needed ("grant failed"). */
1090         ptlrpc_req_finished(req);
1091         RETURN(rc);
1092 }
1093
1094 static int mgc_cancel(struct obd_export *exp, ldlm_mode_t mode,
1095                       struct lustre_handle *lockh)
1096 {
1097         ENTRY;
1098
1099         ldlm_lock_decref(lockh, mode);
1100
1101         RETURN(0);
1102 }
1103
1104 static void mgc_notify_active(struct obd_device *unused)
1105 {
1106         /* wakeup mgc_requeue_thread to requeue mgc lock */
1107         spin_lock(&config_list_lock);
1108         rq_state |= RQ_NOW;
1109         spin_unlock(&config_list_lock);
1110         wake_up(&rq_waitq);
1111
1112         /* TODO: Help the MGS rebuild nidtbl. -jay */
1113 }
1114
1115 /* Send target_reg message to MGS */
1116 static int mgc_target_register(struct obd_export *exp,
1117                                struct mgs_target_info *mti)
1118 {
1119         struct ptlrpc_request  *req;
1120         struct mgs_target_info *req_mti, *rep_mti;
1121         int                     rc;
1122         ENTRY;
1123
1124         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1125                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1126                                         MGS_TARGET_REG);
1127         if (req == NULL)
1128                 RETURN(-ENOMEM);
1129
1130         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1131         if (!req_mti) {
1132                 ptlrpc_req_finished(req);
1133                 RETURN(-ENOMEM);
1134         }
1135
1136         memcpy(req_mti, mti, sizeof(*req_mti));
1137         ptlrpc_request_set_replen(req);
1138         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1139         /* Limit how long we will wait for the enqueue to complete */
1140         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1141
1142         rc = ptlrpc_queue_wait(req);
1143         if (!rc) {
1144                 rep_mti = req_capsule_server_get(&req->rq_pill,
1145                                                  &RMF_MGS_TARGET_INFO);
1146                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1147                 CDEBUG(D_MGC, "register %s got index = %d\n",
1148                        mti->mti_svname, mti->mti_stripe_index);
1149         }
1150         ptlrpc_req_finished(req);
1151
1152         RETURN(rc);
1153 }
1154
1155 int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1156                        obd_count keylen, void *key, obd_count vallen,
1157                        void *val, struct ptlrpc_request_set *set)
1158 {
1159         int rc = -EINVAL;
1160         ENTRY;
1161
1162         /* Turn off initial_recov after we try all backup servers once */
1163         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1164                 struct obd_import *imp = class_exp2cliimp(exp);
1165                 int value;
1166                 if (vallen != sizeof(int))
1167                         RETURN(-EINVAL);
1168                 value = *(int *)val;
1169                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1170                        imp->imp_obd->obd_name, value,
1171                        imp->imp_deactive, imp->imp_invalid,
1172                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1173                        ptlrpc_import_state_name(imp->imp_state));
1174                 /* Resurrect if we previously died */
1175                 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1176                      imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1177                         ptlrpc_reconnect_import(imp);
1178                 RETURN(0);
1179         }
1180         /* FIXME move this to mgc_process_config */
1181         if (KEY_IS(KEY_REGISTER_TARGET)) {
1182                 struct mgs_target_info *mti;
1183                 if (vallen != sizeof(struct mgs_target_info))
1184                         RETURN(-EINVAL);
1185                 mti = (struct mgs_target_info *)val;
1186                 CDEBUG(D_MGC, "register_target %s %#x\n",
1187                        mti->mti_svname, mti->mti_flags);
1188                 rc =  mgc_target_register(exp, mti);
1189                 RETURN(rc);
1190         }
1191         if (KEY_IS(KEY_SET_FS)) {
1192                 struct super_block *sb = (struct super_block *)val;
1193
1194                 if (vallen != sizeof(struct super_block))
1195                         RETURN(-EINVAL);
1196
1197                 rc = mgc_fs_setup(exp->exp_obd, sb);
1198                 RETURN(rc);
1199         }
1200         if (KEY_IS(KEY_CLEAR_FS)) {
1201                 if (vallen != 0)
1202                         RETURN(-EINVAL);
1203                 rc = mgc_fs_cleanup(exp->exp_obd);
1204                 RETURN(rc);
1205         }
1206         if (KEY_IS(KEY_SET_INFO)) {
1207                 struct mgs_send_param *msp;
1208
1209                 msp = (struct mgs_send_param *)val;
1210                 rc =  mgc_set_mgs_param(exp, msp);
1211                 RETURN(rc);
1212         }
1213         if (KEY_IS(KEY_MGSSEC)) {
1214                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1215                 struct sptlrpc_flavor  flvr;
1216
1217                 /*
1218                  * empty string means using current flavor, if which haven't
1219                  * been set yet, set it as null.
1220                  *
1221                  * if flavor has been set previously, check the asking flavor
1222                  * must match the existing one.
1223                  */
1224                 if (vallen == 0) {
1225                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1226                                 RETURN(0);
1227                         val = "null";
1228                         vallen = 4;
1229                 }
1230
1231                 rc = sptlrpc_parse_flavor(val, &flvr);
1232                 if (rc) {
1233                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1234                                (char *) val);
1235                         RETURN(rc);
1236                 }
1237
1238                 /*
1239                  * caller already hold a mutex
1240                  */
1241                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1242                         cli->cl_flvr_mgc = flvr;
1243                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1244                                   sizeof(flvr)) != 0) {
1245                         char    str[20];
1246
1247                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1248                                             str, sizeof(str));
1249                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1250                                        "currently %s is in use\n",
1251                                        (char *) val, str);
1252                         rc = -EPERM;
1253                 }
1254                 RETURN(rc);
1255         }
1256
1257         RETURN(rc);
1258 }
1259
1260 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1261                         __u32 keylen, void *key, __u32 *vallen, void *val,
1262                         struct lov_stripe_md *unused)
1263 {
1264         int rc = -EINVAL;
1265
1266         if (KEY_IS(KEY_CONN_DATA)) {
1267                 struct obd_import *imp = class_exp2cliimp(exp);
1268                 struct obd_connect_data *data = val;
1269
1270                 if (*vallen == sizeof(*data)) {
1271                         *data = imp->imp_connect_data;
1272                         rc = 0;
1273                 }
1274         }
1275
1276         return rc;
1277 }
1278
1279 static int mgc_import_event(struct obd_device *obd,
1280                             struct obd_import *imp,
1281                             enum obd_import_event event)
1282 {
1283         int rc = 0;
1284
1285         LASSERT(imp->imp_obd == obd);
1286         CDEBUG(D_MGC, "import event %#x\n", event);
1287
1288         switch (event) {
1289         case IMP_EVENT_DISCON:
1290                 /* MGC imports should not wait for recovery */
1291                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1292                         ptlrpc_pinger_ir_down();
1293                 break;
1294         case IMP_EVENT_INACTIVE:
1295                 break;
1296         case IMP_EVENT_INVALIDATE: {
1297                 struct ldlm_namespace *ns = obd->obd_namespace;
1298                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1299                 break;
1300         }
1301         case IMP_EVENT_ACTIVE:
1302                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1303                 /* Clearing obd_no_recov allows us to continue pinging */
1304                 obd->obd_no_recov = 0;
1305                 mgc_notify_active(obd);
1306                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1307                         ptlrpc_pinger_ir_up();
1308                 break;
1309         case IMP_EVENT_OCD:
1310                 break;
1311         case IMP_EVENT_DEACTIVATE:
1312         case IMP_EVENT_ACTIVATE:
1313                 break;
1314         default:
1315                 CERROR("Unknown import event %#x\n", event);
1316                 LBUG();
1317         }
1318         RETURN(rc);
1319 }
1320
1321 enum {
1322         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_CACHE_SHIFT),
1323         CONFIG_READ_NRPAGES      = 4
1324 };
1325
1326 static int mgc_apply_recover_logs(struct obd_device *mgc,
1327                                   struct config_llog_data *cld,
1328                                   __u64 max_version,
1329                                   void *data, int datalen, bool mne_swab)
1330 {
1331         struct config_llog_instance *cfg = &cld->cld_cfg;
1332         struct lustre_sb_info       *lsi = s2lsi(cfg->cfg_sb);
1333         struct mgs_nidtbl_entry *entry;
1334         struct lustre_cfg       *lcfg;
1335         struct lustre_cfg_bufs   bufs;
1336         u64   prev_version = 0;
1337         char *inst;
1338         char *buf;
1339         int   bufsz;
1340         int   pos;
1341         int   rc  = 0;
1342         int   off = 0;
1343         ENTRY;
1344
1345         LASSERT(cfg->cfg_instance != NULL);
1346         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1347
1348         OBD_ALLOC(inst, PAGE_CACHE_SIZE);
1349         if (inst == NULL)
1350                 RETURN(-ENOMEM);
1351
1352         if (!IS_SERVER(lsi)) {
1353                 pos = snprintf(inst, PAGE_CACHE_SIZE, "%p", cfg->cfg_instance);
1354                 if (pos >= PAGE_CACHE_SIZE) {
1355                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1356                         return -E2BIG;
1357                 }
1358         } else {
1359                 LASSERT(IS_MDT(lsi));
1360                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1361                                         PAGE_CACHE_SIZE);
1362                 if (rc) {
1363                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1364                         RETURN(-EINVAL);
1365                 }
1366                 pos = strlen(inst);
1367         }
1368
1369         ++pos;
1370         buf   = inst + pos;
1371         bufsz = PAGE_CACHE_SIZE - pos;
1372
1373         while (datalen > 0) {
1374                 int   entry_len = sizeof(*entry);
1375                 int   is_ost;
1376                 struct obd_device *obd;
1377                 char *obdname;
1378                 char *cname;
1379                 char *params;
1380                 char *uuid;
1381
1382                 rc = -EINVAL;
1383                 if (datalen < sizeof(*entry))
1384                         break;
1385
1386                 entry = (typeof(entry))(data + off);
1387
1388                 /* sanity check */
1389                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1390                         break;
1391                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1392                         break;
1393                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1394                         break;
1395
1396                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1397                 if (datalen < entry_len) /* must have entry_len at least */
1398                         break;
1399
1400                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1401                 if (mne_swab)
1402                         lustre_swab_mgs_nidtbl_entry(entry);
1403                 if (entry->mne_length > PAGE_CACHE_SIZE) {
1404                         CERROR("MNE too large (%u)\n", entry->mne_length);
1405                         break;
1406                 }
1407
1408                 if (entry->mne_length < entry_len)
1409                         break;
1410
1411                 off     += entry->mne_length;
1412                 datalen -= entry->mne_length;
1413                 if (datalen < 0)
1414                         break;
1415
1416                 if (entry->mne_version > max_version) {
1417                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1418                                entry->mne_version, max_version);
1419                         break;
1420                 }
1421
1422                 if (prev_version >= entry->mne_version) {
1423                         CERROR("index unsorted, prev %lld, now %lld\n",
1424                                prev_version, entry->mne_version);
1425                         break;
1426                 }
1427                 prev_version = entry->mne_version;
1428
1429                 /*
1430                  * Write a string with format "nid::instance" to
1431                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1432                  */
1433
1434                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1435                 memset(buf, 0, bufsz);
1436                 obdname = buf;
1437                 pos = 0;
1438
1439                 /* lustre-OST0001-osc-<instance #> */
1440                 strcpy(obdname, cld->cld_logname);
1441                 cname = strrchr(obdname, '-');
1442                 if (cname == NULL) {
1443                         CERROR("mgc %s: invalid logname %s\n",
1444                                mgc->obd_name, obdname);
1445                         break;
1446                 }
1447
1448                 pos = cname - obdname;
1449                 obdname[pos] = 0;
1450                 pos += sprintf(obdname + pos, "-%s%04x",
1451                                   is_ost ? "OST" : "MDT", entry->mne_index);
1452
1453                 cname = is_ost ? "osc" : "mdc",
1454                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1455                 lustre_cfg_bufs_reset(&bufs, obdname);
1456
1457                 /* find the obd by obdname */
1458                 obd = class_name2obd(obdname);
1459                 if (obd == NULL) {
1460                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1461                                mgc->obd_name, obdname);
1462                         rc = 0;
1463                         /* this is a safe race, when the ost is starting up...*/
1464                         continue;
1465                 }
1466
1467                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1468                 ++pos;
1469                 params = buf + pos;
1470                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1471                 uuid = buf + pos;
1472
1473                 down_read(&obd->u.cli.cl_sem);
1474                 if (obd->u.cli.cl_import == NULL) {
1475                         /* client does not connect to the OST yet */
1476                         up_read(&obd->u.cli.cl_sem);
1477                         rc = 0;
1478                         continue;
1479                 }
1480
1481                 /* TODO: iterate all nids to find one */
1482                 /* find uuid by nid */
1483                 rc = client_import_find_conn(obd->u.cli.cl_import,
1484                                              entry->u.nids[0],
1485                                              (struct obd_uuid *)uuid);
1486                 up_read(&obd->u.cli.cl_sem);
1487                 if (rc < 0) {
1488                         CERROR("mgc: cannot find uuid by nid %s\n",
1489                                libcfs_nid2str(entry->u.nids[0]));
1490                         break;
1491                 }
1492
1493                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1494                        uuid, libcfs_nid2str(entry->u.nids[0]));
1495
1496                 pos += strlen(uuid);
1497                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1498                 LASSERT(pos < bufsz);
1499
1500                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1501
1502                 rc = -ENOMEM;
1503                 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1504                 if (lcfg == NULL) {
1505                         CERROR("mgc: cannot allocate memory\n");
1506                         break;
1507                 }
1508
1509                 CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n",
1510                        prev_version, max_version, obdname, params);
1511
1512                 rc = class_process_config(lcfg);
1513                 lustre_cfg_free(lcfg);
1514                 if (rc)
1515                         CDEBUG(D_INFO, "process config for %s error %d\n",
1516                                obdname, rc);
1517
1518                 /* continue, even one with error */
1519         }
1520
1521         OBD_FREE(inst, PAGE_CACHE_SIZE);
1522         RETURN(rc);
1523 }
1524
1525 /**
1526  * This function is called if this client was notified for target restarting
1527  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1528  */
1529 static int mgc_process_recover_log(struct obd_device *obd,
1530                                    struct config_llog_data *cld)
1531 {
1532         struct ptlrpc_request *req = NULL;
1533         struct config_llog_instance *cfg = &cld->cld_cfg;
1534         struct mgs_config_body *body;
1535         struct mgs_config_res  *res;
1536         struct ptlrpc_bulk_desc *desc;
1537         struct page **pages;
1538         int nrpages;
1539         bool eof = true;
1540         bool mne_swab = false;
1541         int i;
1542         int ealen;
1543         int rc;
1544         ENTRY;
1545
1546         /* allocate buffer for bulk transfer.
1547          * if this is the first time for this mgs to read logs,
1548          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1549          * once; otherwise, it only reads increment of logs, this should be
1550          * small and CONFIG_READ_NRPAGES will be used.
1551          */
1552         nrpages = CONFIG_READ_NRPAGES;
1553         if (cfg->cfg_last_idx == 0) /* the first time */
1554                 nrpages = CONFIG_READ_NRPAGES_INIT;
1555
1556         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1557         if (pages == NULL)
1558                 GOTO(out, rc = -ENOMEM);
1559
1560         for (i = 0; i < nrpages; i++) {
1561                 pages[i] = alloc_page(GFP_IOFS);
1562                 if (pages[i] == NULL)
1563                         GOTO(out, rc = -ENOMEM);
1564         }
1565
1566 again:
1567         LASSERT(cld_is_recover(cld));
1568         LASSERT(mutex_is_locked(&cld->cld_lock));
1569         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1570                                    &RQF_MGS_CONFIG_READ);
1571         if (req == NULL)
1572                 GOTO(out, rc = -ENOMEM);
1573
1574         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1575         if (rc)
1576                 GOTO(out, rc);
1577
1578         /* pack request */
1579         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1580         LASSERT(body != NULL);
1581         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1582         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1583             >= sizeof(body->mcb_name))
1584                 GOTO(out, rc = -E2BIG);
1585         body->mcb_offset = cfg->cfg_last_idx + 1;
1586         body->mcb_type   = cld->cld_type;
1587         body->mcb_bits   = PAGE_CACHE_SHIFT;
1588         body->mcb_units  = nrpages;
1589
1590         /* allocate bulk transfer descriptor */
1591         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1, BULK_PUT_SINK,
1592                                     MGS_BULK_PORTAL);
1593         if (desc == NULL)
1594                 GOTO(out, rc = -ENOMEM);
1595
1596         for (i = 0; i < nrpages; i++)
1597                 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
1598
1599         ptlrpc_request_set_replen(req);
1600         rc = ptlrpc_queue_wait(req);
1601         if (rc)
1602                 GOTO(out, rc);
1603
1604         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1605         if (res->mcr_size < res->mcr_offset)
1606                 GOTO(out, rc = -EINVAL);
1607
1608         /* always update the index even though it might have errors with
1609          * handling the recover logs */
1610         cfg->cfg_last_idx = res->mcr_offset;
1611         eof = res->mcr_offset == res->mcr_size;
1612
1613         CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n",
1614                res->mcr_offset, eof == false);
1615
1616         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1617         if (ealen < 0)
1618                 GOTO(out, rc = ealen);
1619
1620         if (ealen > nrpages << PAGE_CACHE_SHIFT)
1621                 GOTO(out, rc = -EINVAL);
1622
1623         if (ealen == 0) { /* no logs transferred */
1624                 if (!eof)
1625                         rc = -EINVAL;
1626                 GOTO(out, rc);
1627         }
1628
1629         mne_swab = !!ptlrpc_rep_need_swab(req);
1630 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1631         /* This import flag means the server did an extra swab of IR MNE
1632          * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1633         if (unlikely(req->rq_import->imp_need_mne_swab))
1634                 mne_swab = !mne_swab;
1635 #else
1636 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1637 #endif
1638
1639         for (i = 0; i < nrpages && ealen > 0; i++) {
1640                 int rc2;
1641                 void *ptr;
1642
1643                 ptr = kmap(pages[i]);
1644                 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1645                                              min_t(int, ealen, PAGE_CACHE_SIZE),
1646                                              mne_swab);
1647                 kunmap(pages[i]);
1648                 if (rc2 < 0) {
1649                         CWARN("Process recover log %s error %d\n",
1650                               cld->cld_logname, rc2);
1651                         break;
1652                 }
1653
1654                 ealen -= PAGE_CACHE_SIZE;
1655         }
1656
1657 out:
1658         if (req)
1659                 ptlrpc_req_finished(req);
1660
1661         if (rc == 0 && !eof)
1662                 goto again;
1663
1664         if (pages) {
1665                 for (i = 0; i < nrpages; i++) {
1666                         if (pages[i] == NULL)
1667                                 break;
1668                         __free_page(pages[i]);
1669                 }
1670                 OBD_FREE(pages, sizeof(*pages) * nrpages);
1671         }
1672         return rc;
1673 }
1674
1675 /* Copy a remote log locally */
1676 static int mgc_llog_local_copy(const struct lu_env *env,
1677                                struct obd_device *obd,
1678                                struct llog_ctxt *rctxt,
1679                                struct llog_ctxt *lctxt, char *logname)
1680 {
1681         char    *temp_log;
1682         int      rc;
1683
1684         ENTRY;
1685
1686         /*
1687          * - copy it to backup using llog_backup()
1688          * - copy remote llog to logname using llog_backup()
1689          * - if failed then move bakup to logname again
1690          */
1691
1692         OBD_ALLOC(temp_log, strlen(logname) + 1);
1693         if (!temp_log)
1694                 RETURN(-ENOMEM);
1695         sprintf(temp_log, "%sT", logname);
1696
1697         /* make a copy of local llog at first */
1698         rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1699         if (rc < 0 && rc != -ENOENT)
1700                 GOTO(out, rc);
1701         /* copy remote llog to the local copy */
1702         rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1703         if (rc == -ENOENT) {
1704                 /* no remote llog, delete local one too */
1705                 llog_erase(env, lctxt, NULL, logname);
1706         } else if (rc < 0) {
1707                 /* error during backup, get local one back from the copy */
1708                 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1709 out:
1710                 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1711                        obd->obd_name, logname, rc);
1712         }
1713         llog_erase(env, lctxt, NULL, temp_log);
1714         OBD_FREE(temp_log, strlen(logname) + 1);
1715         return rc;
1716 }
1717
1718 /* local_only means it cannot get remote llogs */
1719 static int mgc_process_cfg_log(struct obd_device *mgc,
1720                                struct config_llog_data *cld, int local_only)
1721 {
1722         struct llog_ctxt        *ctxt, *lctxt = NULL;
1723         struct client_obd       *cli = &mgc->u.cli;
1724         struct lustre_sb_info   *lsi = NULL;
1725         int                      rc = 0;
1726         bool                     sptlrpc_started = false;
1727         struct lu_env           *env;
1728
1729         ENTRY;
1730
1731         LASSERT(cld);
1732         LASSERT(mutex_is_locked(&cld->cld_lock));
1733
1734         /*
1735          * local copy of sptlrpc log is controlled elsewhere, don't try to
1736          * read it up here.
1737          */
1738         if (cld_is_sptlrpc(cld) && local_only)
1739                 RETURN(0);
1740
1741         if (cld->cld_cfg.cfg_sb)
1742                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1743
1744         OBD_ALLOC_PTR(env);
1745         if (env == NULL)
1746                 RETURN(-ENOMEM);
1747
1748         rc = lu_env_init(env, LCT_MG_THREAD);
1749         if (rc)
1750                 GOTO(out_free, rc);
1751
1752         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1753         LASSERT(ctxt);
1754
1755         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1756
1757         /* Copy the setup log locally if we can. Don't mess around if we're
1758          * running an MGS though (logs are already local). */
1759         if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1760             cli->cl_mgc_configs_dir != NULL &&
1761             lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1762             lsi->lsi_dt_dev) {
1763                 if (!local_only)
1764                         /* Only try to copy log if we have the lock. */
1765                         rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1766                                                  cld->cld_logname);
1767                 if (local_only || rc) {
1768                         if (llog_is_empty(env, lctxt, cld->cld_logname)) {
1769                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1770                                                    "log %s and no local copy."
1771                                                    "\n", cld->cld_logname);
1772                                 GOTO(out_pop, rc = -ENOENT);
1773                         }
1774                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1775                                "copy for now, will try to update later.\n",
1776                                cld->cld_logname);
1777                 }
1778                 /* Now, whether we copied or not, start using the local llog.
1779                  * If we failed to copy, we'll start using whatever the old
1780                  * log has. */
1781                 llog_ctxt_put(ctxt);
1782                 ctxt = lctxt;
1783                 lctxt = NULL;
1784         } else {
1785                 if (local_only) /* no local log at client side */
1786                         GOTO(out_pop, rc = -EIO);
1787         }
1788
1789         if (cld_is_sptlrpc(cld)) {
1790                 sptlrpc_conf_log_update_begin(cld->cld_logname);
1791                 sptlrpc_started = true;
1792         }
1793
1794         /* logname and instance info should be the same, so use our
1795          * copy of the instance for the update.  The cfg_last_idx will
1796          * be updated here. */
1797         rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1798                                      &cld->cld_cfg);
1799         EXIT;
1800
1801 out_pop:
1802         __llog_ctxt_put(env, ctxt);
1803         if (lctxt)
1804                 __llog_ctxt_put(env, lctxt);
1805
1806         /*
1807          * update settings on existing OBDs. doing it inside
1808          * of llog_process_lock so no device is attaching/detaching
1809          * in parallel.
1810          * the logname must be <fsname>-sptlrpc
1811          */
1812         if (sptlrpc_started) {
1813                 LASSERT(cld_is_sptlrpc(cld));
1814                 sptlrpc_conf_log_update_end(cld->cld_logname);
1815                 class_notify_sptlrpc_conf(cld->cld_logname,
1816                                           strlen(cld->cld_logname) -
1817                                           strlen("-sptlrpc"));
1818         }
1819
1820         lu_env_fini(env);
1821 out_free:
1822         OBD_FREE_PTR(env);
1823         return rc;
1824 }
1825
1826 /** Get a config log from the MGS and process it.
1827  * This func is called for both clients and servers.
1828  * Copy the log locally before parsing it if appropriate (non-MGS server)
1829  */
1830 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1831 {
1832         struct lustre_handle lockh = { 0 };
1833         __u64 flags = LDLM_FL_NO_LRU;
1834         int rc = 0, rcl;
1835         ENTRY;
1836
1837         LASSERT(cld);
1838
1839         /* I don't want multiple processes running process_log at once --
1840            sounds like badness.  It actually might be fine, as long as
1841            we're not trying to update from the same log
1842            simultaneously (in which case we should use a per-log sem.) */
1843         mutex_lock(&cld->cld_lock);
1844         if (cld->cld_stopping) {
1845                 mutex_unlock(&cld->cld_lock);
1846                 RETURN(0);
1847         }
1848
1849         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1850
1851         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1852                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1853
1854         /* Get the cfg lock on the llog */
1855         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1856                           LCK_CR, &flags, NULL, NULL, NULL,
1857                           cld, 0, NULL, &lockh);
1858         if (rcl == 0) {
1859                 /* Get the cld, it will be released in mgc_blocking_ast. */
1860                 config_log_get(cld);
1861                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1862                 LASSERT(rc == 0);
1863         } else {
1864                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1865
1866                 /* mark cld_lostlock so that it will requeue
1867                  * after MGC becomes available. */
1868                 cld->cld_lostlock = 1;
1869                 /* Get extra reference, it will be put in requeue thread */
1870                 config_log_get(cld);
1871         }
1872
1873
1874         if (cld_is_recover(cld)) {
1875                 rc = 0; /* this is not a fatal error for recover log */
1876                 if (rcl == 0)
1877                         rc = mgc_process_recover_log(mgc, cld);
1878         } else {
1879                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1880         }
1881
1882         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1883                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1884
1885         mutex_unlock(&cld->cld_lock);
1886
1887         /* Now drop the lock so MGS can revoke it */
1888         if (!rcl) {
1889                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
1890                 if (rcl)
1891                         CERROR("Can't drop cfg lock: %d\n", rcl);
1892         }
1893
1894         RETURN(rc);
1895 }
1896
1897
1898 /** Called from lustre_process_log.
1899  * LCFG_LOG_START gets the config log from the MGS, processes it to start
1900  * any services, and adds it to the list logs to watch (follow).
1901  */
1902 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1903 {
1904         struct lustre_cfg *lcfg = buf;
1905         struct config_llog_instance *cfg = NULL;
1906         char *logname;
1907         int rc = 0;
1908         ENTRY;
1909
1910         switch(lcfg->lcfg_command) {
1911         case LCFG_LOV_ADD_OBD: {
1912                 /* Overloading this cfg command: register a new target */
1913                 struct mgs_target_info *mti;
1914
1915                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1916                     sizeof(struct mgs_target_info))
1917                         GOTO(out, rc = -EINVAL);
1918
1919                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1920                 CDEBUG(D_MGC, "add_target %s %#x\n",
1921                        mti->mti_svname, mti->mti_flags);
1922                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1923                 break;
1924         }
1925         case LCFG_LOV_DEL_OBD:
1926                 /* Unregister has no meaning at the moment. */
1927                 CERROR("lov_del_obd unimplemented\n");
1928                 rc = -ENOSYS;
1929                 break;
1930         case LCFG_SPTLRPC_CONF: {
1931                 rc = sptlrpc_process_config(lcfg);
1932                 break;
1933         }
1934         case LCFG_LOG_START: {
1935                 struct config_llog_data *cld;
1936                 struct super_block *sb;
1937
1938                 logname = lustre_cfg_string(lcfg, 1);
1939                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1940                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1941
1942                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1943                        cfg->cfg_last_idx);
1944
1945                 /* We're only called through here on the initial mount */
1946                 rc = config_log_add(obd, logname, cfg, sb);
1947                 if (rc)
1948                         break;
1949                 cld = config_log_find(logname, cfg);
1950                 if (cld == NULL) {
1951                         rc = -ENOENT;
1952                         break;
1953                 }
1954
1955                 /* COMPAT_146 */
1956                 /* FIXME only set this for old logs!  Right now this forces
1957                    us to always skip the "inside markers" check */
1958                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1959
1960                 rc = mgc_process_log(obd, cld);
1961                 if (rc == 0 && cld->cld_recover != NULL) {
1962                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1963                                          imp_connect_data, IMP_RECOV)) {
1964                                 rc = mgc_process_log(obd, cld->cld_recover);
1965                         } else {
1966                                 struct config_llog_data *cir = cld->cld_recover;
1967                                 cld->cld_recover = NULL;
1968                                 config_log_put(cir);
1969                         }
1970                         if (rc)
1971                                 CERROR("Cannot process recover llog %d\n", rc);
1972                 }
1973
1974                 if (rc == 0 && cld->cld_params != NULL) {
1975                         rc = mgc_process_log(obd, cld->cld_params);
1976                         if (rc == -ENOENT) {
1977                                 CDEBUG(D_MGC, "There is no params "
1978                                               "config file yet\n");
1979                                 rc = 0;
1980                         }
1981                         /* params log is optional */
1982                         if (rc)
1983                                 CERROR("%s: can't process params llog: rc = %d\n",
1984                                        obd->obd_name, rc);
1985                 }
1986                 config_log_put(cld);
1987
1988                 break;
1989         }
1990         case LCFG_LOG_END: {
1991                 logname = lustre_cfg_string(lcfg, 1);
1992
1993                 if (lcfg->lcfg_bufcount >= 2)
1994                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
1995                                 lcfg, 2);
1996                 rc = config_log_end(logname, cfg);
1997                 break;
1998         }
1999         default: {
2000                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2001                 GOTO(out, rc = -EINVAL);
2002
2003         }
2004         }
2005 out:
2006         RETURN(rc);
2007 }
2008
2009 struct obd_ops mgc_obd_ops = {
2010         .o_owner        = THIS_MODULE,
2011         .o_setup        = mgc_setup,
2012         .o_precleanup   = mgc_precleanup,
2013         .o_cleanup      = mgc_cleanup,
2014         .o_add_conn     = client_import_add_conn,
2015         .o_del_conn     = client_import_del_conn,
2016         .o_connect      = client_connect_import,
2017         .o_disconnect   = client_disconnect_export,
2018         .o_set_info_async = mgc_set_info_async,
2019         .o_get_info       = mgc_get_info,
2020         .o_import_event = mgc_import_event,
2021         .o_process_config = mgc_process_config,
2022 };
2023
2024 int __init mgc_init(void)
2025 {
2026         return class_register_type(&mgc_obd_ops, NULL, true, NULL,
2027 #ifndef HAVE_ONLY_PROCFS_SEQ
2028                                    NULL,
2029 #endif
2030                                    LUSTRE_MGC_NAME, NULL);
2031 }
2032
2033 #ifdef __KERNEL__
2034 static void /*__exit*/ mgc_exit(void)
2035 {
2036         class_unregister_type(LUSTRE_MGC_NAME);
2037 }
2038
2039 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2040 MODULE_DESCRIPTION("Lustre Management Client");
2041 MODULE_LICENSE("GPL");
2042
2043 module_init(mgc_init);
2044 module_exit(mgc_exit);
2045 #endif