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 struct list_head config_llog_list = LIST_HEAD_INIT(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                 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         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         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("%s: sptlrpc log name not correct, %s: "
 369                                "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
 370                         config_log_put(cld);
 371                         RETURN(-EINVAL);
 372                 }
 373                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 374                 if (IS_ERR(recover_cld))
 375                         GOTO(out_err3, rc = PTR_ERR(recover_cld));
 376                 cld->cld_recover = recover_cld;
 377         }
 378
 379         RETURN(0);
 380
 381 out_err3:
 382         config_log_put(cld);
 383
 384 out_err2:
 385         config_log_put(params_cld);
 386
 387 out_err1:
 388         config_log_put(sptlrpc_cld);
 389
 390 out_err:
 391         RETURN(rc);
 392 }
 393
 394 DEFINE_MUTEX(llog_process_lock);
 395
 396 /** Stop watching for updates on this log.
 397  */
 398 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 399 {
 400         struct config_llog_data *cld;
 401         struct config_llog_data *cld_sptlrpc = NULL;
 402         struct config_llog_data *cld_params = NULL;
 403         struct config_llog_data *cld_recover = NULL;
 404         int rc = 0;
 405         ENTRY;
 406
 407         cld = config_log_find(logname, cfg);
 408         if (cld == NULL)
 409                 RETURN(-ENOENT);
 410
 411         mutex_lock(&cld->cld_lock);
 412         /*
 413          * if cld_stopping is set, it means we didn't start the log thus
 414          * not owning the start ref. this can happen after previous umount:
 415          * the cld still hanging there waiting for lock cancel, and we
 416          * remount again but failed in the middle and call log_end without
 417          * calling start_log.
 418          */
 419         if (unlikely(cld->cld_stopping)) {
 420                 mutex_unlock(&cld->cld_lock);
 421                 /* drop the ref from the find */
 422                 config_log_put(cld);
 423                 RETURN(rc);
 424         }
 425
 426         cld->cld_stopping = 1;
 427
 428         cld_recover = cld->cld_recover;
 429         cld->cld_recover = NULL;
 430         mutex_unlock(&cld->cld_lock);
 431
 432         if (cld_recover) {
 433                 mutex_lock(&cld_recover->cld_lock);
 434                 cld_recover->cld_stopping = 1;
 435                 mutex_unlock(&cld_recover->cld_lock);
 436                 config_log_put(cld_recover);
 437         }
 438
 439         spin_lock(&config_list_lock);
 440         cld_sptlrpc = cld->cld_sptlrpc;
 441         cld->cld_sptlrpc = NULL;
 442         cld_params = cld->cld_params;
 443         cld->cld_params = NULL;
 444         spin_unlock(&config_list_lock);
 445
 446         if (cld_sptlrpc)
 447                 config_log_put(cld_sptlrpc);
 448
 449         if (cld_params) {
 450                 mutex_lock(&cld_params->cld_lock);
 451                 cld_params->cld_stopping = 1;
 452                 mutex_unlock(&cld_params->cld_lock);
 453                 config_log_put(cld_params);
 454         }
 455
 456         /* drop the ref from the find */
 457         config_log_put(cld);
 458         /* drop the start ref */
 459         config_log_put(cld);
 460
 461         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 462                rc);
 463         RETURN(rc);
 464 }
 465
 466 #ifdef LPROCFS
 467 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 468 {
 469         struct obd_device       *obd = data;
 470         struct obd_import       *imp = obd->u.cli.cl_import;
 471         struct obd_connect_data *ocd = &imp->imp_connect_data;
 472         struct config_llog_data *cld;
 473         ENTRY;
 474
 475         seq_printf(m, "imperative_recovery: %s\n",
 476                    OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 477         seq_printf(m, "client_state:\n");
 478
 479         spin_lock(&config_list_lock);
 480         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 481                 if (cld->cld_recover == NULL)
 482                         continue;
 483                 seq_printf(m,  "    - { client: %s, nidtbl_version: %u }\n",
 484                                cld->cld_logname,
 485                                cld->cld_recover->cld_cfg.cfg_last_idx);
 486         }
 487         spin_unlock(&config_list_lock);
 488
 489         RETURN(0);
 490 }
 491 #endif
 492
 493 /* reenqueue any lost locks */
 494 #define RQ_RUNNING 0x1
 495 #define RQ_NOW     0x2
 496 #define RQ_LATER   0x4
 497 #define RQ_STOP    0x8
 498 static int                    rq_state = 0;
 499 static wait_queue_head_t      rq_waitq;
 500 static DECLARE_COMPLETION(rq_exit);
 501
 502 static void do_requeue(struct config_llog_data *cld)
 503 {
 504         int rc = 0;
 505         ENTRY;
 506
 507         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 508
 509         /* Do not run mgc_process_log on a disconnected export or an
 510          * export which is being disconnected. Take the client
 511          * semaphore to make the check non-racy. */
 512         down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 513         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 514                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 515                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 516                 if (rc && rc != -ENOENT)
 517                         CERROR("failed processing log: %d\n", rc);
 518         } else {
 519                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 520                        cld->cld_logname);
 521         }
 522         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 523
 524         EXIT;
 525 }
 526
 527 /* this timeout represents how many seconds MGC should wait before
 528  * requeue config and recover lock to the MGS. We need to randomize this
 529  * in order to not flood the MGS.
 530  */
 531 #define MGC_TIMEOUT_MIN_SECONDS   5
 532 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
 533
 534 static int mgc_requeue_thread(void *data)
 535 {
 536         int rc = 0;
 537         ENTRY;
 538
 539         CDEBUG(D_MGC, "Starting requeue thread\n");
 540
 541         /* Keep trying failed locks periodically */
 542         spin_lock(&config_list_lock);
 543         rq_state |= RQ_RUNNING;
 544         while (1) {
 545                 struct l_wait_info lwi;
 546                 struct config_llog_data *cld, *cld_prev;
 547                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
 548                 int stopped = !!(rq_state & RQ_STOP);
 549                 int to;
 550
 551                 /* Any new or requeued lostlocks will change the state */
 552                 rq_state &= ~(RQ_NOW | RQ_LATER);
 553                 spin_unlock(&config_list_lock);
 554
 555                 /* Always wait a few seconds to allow the server who
 556                    caused the lock revocation to finish its setup, plus some
 557                    random so everyone doesn't try to reconnect at once. */
 558                 to = MGC_TIMEOUT_MIN_SECONDS * HZ;
 559                 to += rand * HZ / 100; /* rand is centi-seconds */
 560                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 561                 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi);
 562
 563                 /*
 564                  * iterate & processing through the list. for each cld, process
 565                  * its depending sptlrpc cld firstly (if any) and then itself.
 566                  *
 567                  * it's guaranteed any item in the list must have
 568                  * reference > 0; and if cld_lostlock is set, at
 569                  * least one reference is taken by the previous enqueue.
 570                  */
 571                 cld_prev = NULL;
 572
 573                 spin_lock(&config_list_lock);
 574                 list_for_each_entry(cld, &config_llog_list,
 575                                         cld_list_chain) {
 576                         if (!cld->cld_lostlock)
 577                                 continue;
 578
 579                         spin_unlock(&config_list_lock);
 580
 581                         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 582
 583                         /* Whether we enqueued again or not in mgc_process_log,
 584                          * we're done with the ref from the old enqueue */
 585                         if (cld_prev)
 586                                 config_log_put(cld_prev);
 587                         cld_prev = cld;
 588
 589                         cld->cld_lostlock = 0;
 590                         if (likely(!stopped))
 591                                 do_requeue(cld);
 592
 593                         spin_lock(&config_list_lock);
 594                 }
 595                 spin_unlock(&config_list_lock);
 596                 if (cld_prev)
 597                         config_log_put(cld_prev);
 598
 599                 /* break after scanning the list so that we can drop
 600                  * refcount to losing lock clds */
 601                 if (unlikely(stopped)) {
 602                         spin_lock(&config_list_lock);
 603                         break;
 604                 }
 605
 606                 /* Wait a bit to see if anyone else needs a requeue */
 607                 lwi = (struct l_wait_info) { 0 };
 608                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 609                              &lwi);
 610                 spin_lock(&config_list_lock);
 611         }
 612         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 613         rq_state &= ~RQ_RUNNING;
 614         spin_unlock(&config_list_lock);
 615
 616         complete(&rq_exit);
 617
 618         CDEBUG(D_MGC, "Ending requeue thread\n");
 619         RETURN(rc);
 620 }
 621
 622 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 623    We are responsible for dropping the config log reference from here on out. */
 624 static void mgc_requeue_add(struct config_llog_data *cld)
 625 {
 626         ENTRY;
 627
 628         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 629                 cld->cld_logname, atomic_read(&cld->cld_refcount),
 630                 cld->cld_stopping, rq_state);
 631         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 632
 633         mutex_lock(&cld->cld_lock);
 634         if (cld->cld_stopping || cld->cld_lostlock) {
 635                 mutex_unlock(&cld->cld_lock);
 636                 RETURN_EXIT;
 637         }
 638         /* this refcount will be released in mgc_requeue_thread. */
 639         config_log_get(cld);
 640         cld->cld_lostlock = 1;
 641         mutex_unlock(&cld->cld_lock);
 642
 643         /* Hold lock for rq_state */
 644         spin_lock(&config_list_lock);
 645         if (rq_state & RQ_STOP) {
 646                 spin_unlock(&config_list_lock);
 647                 cld->cld_lostlock = 0;
 648                 config_log_put(cld);
 649         } else {
 650                 rq_state |= RQ_NOW;
 651                 spin_unlock(&config_list_lock);
 652                 wake_up(&rq_waitq);
 653         }
 654         EXIT;
 655 }
 656
 657 /********************** class fns **********************/
 658 static int mgc_local_llog_init(const struct lu_env *env,
 659                                struct obd_device *obd,
 660                                struct obd_device *disk)
 661 {
 662         struct llog_ctxt        *ctxt;
 663         int                      rc;
 664
 665         ENTRY;
 666
 667         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
 668                         &llog_osd_ops);
 669         if (rc)
 670                 RETURN(rc);
 671
 672         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 673         LASSERT(ctxt);
 674         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
 675         llog_ctxt_put(ctxt);
 676
 677         RETURN(0);
 678 }
 679
 680 static int mgc_local_llog_fini(const struct lu_env *env,
 681                                struct obd_device *obd)
 682 {
 683         struct llog_ctxt *ctxt;
 684
 685         ENTRY;
 686
 687         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 688         llog_cleanup(env, ctxt);
 689
 690         RETURN(0);
 691 }
 692
 693 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb)
 694 {
 695         struct lustre_sb_info   *lsi = s2lsi(sb);
 696         struct client_obd       *cli = &obd->u.cli;
 697         struct lu_fid            rfid, fid;
 698         struct dt_object        *root, *dto;
 699         struct lu_env           *env;
 700         int                      rc = 0;
 701
 702         ENTRY;
 703
 704         LASSERT(lsi);
 705         LASSERT(lsi->lsi_dt_dev);
 706
 707         OBD_ALLOC_PTR(env);
 708         if (env == NULL)
 709                 RETURN(-ENOMEM);
 710
 711         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
 712         mutex_lock(&cli->cl_mgc_mutex);
 713
 714         /* Setup the configs dir */
 715         rc = lu_env_init(env, LCT_MG_THREAD);
 716         if (rc)
 717                 GOTO(out_err, rc);
 718
 719         fid.f_seq = FID_SEQ_LOCAL_NAME;
 720         fid.f_oid = 1;
 721         fid.f_ver = 0;
 722         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
 723                                     &cli->cl_mgc_los);
 724         if (rc)
 725                 GOTO(out_env, rc);
 726
 727         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
 728         if (rc)
 729                 GOTO(out_env, rc);
 730
 731         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
 732                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
 733         if (unlikely(IS_ERR(root)))
 734                 GOTO(out_los, rc = PTR_ERR(root));
 735
 736         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
 737                                         MOUNT_CONFIGS_DIR,
 738                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
 739         lu_object_put_nocache(env, &root->do_lu);
 740         if (IS_ERR(dto))
 741                 GOTO(out_los, rc = PTR_ERR(dto));
 742
 743         cli->cl_mgc_configs_dir = dto;
 744
 745         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
 746         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
 747         if (rc)
 748                 GOTO(out_llog, rc);
 749
 750         /* We take an obd ref to insure that we can't get to mgc_cleanup
 751          * without calling mgc_fs_cleanup first. */
 752         class_incref(obd, "mgc_fs", obd);
 753
 754         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 755         EXIT;
 756 out_llog:
 757         if (rc) {
 758                 lu_object_put(env, &cli->cl_mgc_configs_dir->do_lu);
 759                 cli->cl_mgc_configs_dir = NULL;
 760         }
 761 out_los:
 762         if (rc < 0) {
 763                 local_oid_storage_fini(env, cli->cl_mgc_los);
 764                 cli->cl_mgc_los = NULL;
 765                 mutex_unlock(&cli->cl_mgc_mutex);
 766         }
 767 out_env:
 768         lu_env_fini(env);
 769 out_err:
 770         OBD_FREE_PTR(env);
 771         return rc;
 772 }
 773
 774 static int mgc_fs_cleanup(struct obd_device *obd)
 775 {
 776         struct lu_env            env;
 777         struct client_obd       *cli = &obd->u.cli;
 778         int                      rc;
 779
 780         ENTRY;
 781
 782         LASSERT(cli->cl_mgc_los != NULL);
 783
 784         rc = lu_env_init(&env, LCT_MG_THREAD);
 785         if (rc)
 786                 GOTO(unlock, rc);
 787
 788         mgc_local_llog_fini(&env, obd);
 789
 790         lu_object_put_nocache(&env, &cli->cl_mgc_configs_dir->do_lu);
 791         cli->cl_mgc_configs_dir = NULL;
 792
 793         local_oid_storage_fini(&env, cli->cl_mgc_los);
 794         cli->cl_mgc_los = NULL;
 795         lu_env_fini(&env);
 796
 797 unlock:
 798         class_decref(obd, "mgc_fs", obd);
 799         mutex_unlock(&cli->cl_mgc_mutex);
 800
 801         RETURN(0);
 802 }
 803
 804 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 805 {
 806         struct llog_ctxt        *ctxt;
 807         int                      rc;
 808
 809         ENTRY;
 810
 811         /* setup only remote ctxt, the local disk context is switched per each
 812          * filesystem during mgc_fs_setup() */
 813         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 814                         &llog_client_ops);
 815         if (rc)
 816                 RETURN(rc);
 817
 818         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 819         LASSERT(ctxt);
 820
 821         llog_initiator_connect(ctxt);
 822         llog_ctxt_put(ctxt);
 823
 824         RETURN(0);
 825 }
 826
 827 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 828 {
 829         struct llog_ctxt *ctxt;
 830
 831         ENTRY;
 832
 833         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 834         if (ctxt)
 835                 llog_cleanup(env, ctxt);
 836
 837         RETURN(0);
 838 }
 839
 840
 841 static atomic_t mgc_count = ATOMIC_INIT(0);
 842 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 843 {
 844         int rc = 0;
 845         ENTRY;
 846
 847         switch (stage) {
 848         case OBD_CLEANUP_EARLY:
 849                 break;
 850         case OBD_CLEANUP_EXPORTS:
 851                 if (atomic_dec_and_test(&mgc_count)) {
 852                         int running;
 853                         /* stop requeue thread */
 854                         spin_lock(&config_list_lock);
 855                         running = rq_state & RQ_RUNNING;
 856                         if (running)
 857                                 rq_state |= RQ_STOP;
 858                         spin_unlock(&config_list_lock);
 859                         if (running) {
 860                                 wake_up(&rq_waitq);
 861                                 wait_for_completion(&rq_exit);
 862                         }
 863                 }
 864                 obd_cleanup_client_import(obd);
 865                 rc = mgc_llog_fini(NULL, obd);
 866                 if (rc != 0)
 867                         CERROR("failed to cleanup llogging subsystems\n");
 868                 break;
 869         }
 870         RETURN(rc);
 871 }
 872
 873 static int mgc_cleanup(struct obd_device *obd)
 874 {
 875         int rc;
 876         ENTRY;
 877
 878         /* COMPAT_146 - old config logs may have added profiles we don't
 879            know about */
 880         if (obd->obd_type->typ_refcnt <= 1)
 881                 /* Only for the last mgc */
 882                 class_del_profiles();
 883
 884         lprocfs_obd_cleanup(obd);
 885         ptlrpcd_decref();
 886
 887         rc = client_obd_cleanup(obd);
 888         RETURN(rc);
 889 }
 890
 891 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 892 {
 893         struct task_struct      *task;
 894         int                      rc;
 895         ENTRY;
 896
 897         rc = ptlrpcd_addref();
 898         if (rc < 0)
 899                 RETURN(rc);
 900
 901         rc = client_obd_setup(obd, lcfg);
 902         if (rc)
 903                 GOTO(err_decref, rc);
 904
 905         rc = mgc_llog_init(NULL, obd);
 906         if (rc) {
 907                 CERROR("failed to setup llogging subsystems\n");
 908                 GOTO(err_cleanup, rc);
 909         }
 910
 911 #ifdef LPROCFS
 912         obd->obd_vars = lprocfs_mgc_obd_vars;
 913         lprocfs_seq_obd_setup(obd);
 914 #endif
 915         sptlrpc_lprocfs_cliobd_attach(obd);
 916
 917         if (atomic_inc_return(&mgc_count) == 1) {
 918                 rq_state = 0;
 919                 init_waitqueue_head(&rq_waitq);
 920
 921                 /* start requeue thread */
 922                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 923                 if (IS_ERR(task)) {
 924                         rc = PTR_ERR(task);
 925                         CERROR("%s: cannot start requeue thread: rc = %d; "
 926                                "no more log updates\n",
 927                                obd->obd_name, rc);
 928                         GOTO(err_cleanup, rc);
 929                 }
 930                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 931                 rc = 0;
 932         }
 933
 934         RETURN(rc);
 935
 936 err_cleanup:
 937         client_obd_cleanup(obd);
 938 err_decref:
 939         ptlrpcd_decref();
 940         RETURN(rc);
 941 }
 942
 943 /* based on ll_mdc_blocking_ast */
 944 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 945                             void *data, int flag)
 946 {
 947         struct lustre_handle lockh;
 948         struct config_llog_data *cld = (struct config_llog_data *)data;
 949         int rc = 0;
 950         ENTRY;
 951
 952         switch (flag) {
 953         case LDLM_CB_BLOCKING:
 954                 /* mgs wants the lock, give it up... */
 955                 LDLM_DEBUG(lock, "MGC blocking CB");
 956                 ldlm_lock2handle(lock, &lockh);
 957                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 958                 break;
 959         case LDLM_CB_CANCELING:
 960                 /* We've given up the lock, prepare ourselves to update. */
 961                 LDLM_DEBUG(lock, "MGC cancel CB");
 962
 963                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
 964                        PLDLMRES(lock->l_resource),
 965                        (char *)&lock->l_resource->lr_name.name[0]);
 966
 967                 if (!cld) {
 968                         CDEBUG(D_INFO, "missing data, won't requeue\n");
 969                         break;
 970                 }
 971
 972                 /* held at mgc_process_log(). */
 973                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
 974                 /* Are we done with this log? */
 975                 if (cld->cld_stopping) {
 976                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
 977                                 cld->cld_logname);
 978                         config_log_put(cld);
 979                         break;
 980                 }
 981                 /* Make sure not to re-enqueue when the mgc is stopping
 982                    (we get called from client_disconnect_export) */
 983                 if (!lock->l_conn_export ||
 984                     !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
 985                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
 986                                 cld->cld_logname);
 987                         config_log_put(cld);
 988                         break;
 989                 }
 990
 991                 /* Re-enqueue now */
 992                 mgc_requeue_add(cld);
 993                 config_log_put(cld);
 994                 break;
 995         default:
 996                 LBUG();
 997         }
 998
 999         RETURN(rc);
1000 }
1001
1002 /* Not sure where this should go... */
1003 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1004  * that we can have a chance to try the secondary MGS if any. */
1005 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1006                                 + PING_INTERVAL)
1007 #define  MGC_TARGET_REG_LIMIT 10
1008 #define  MGC_SEND_PARAM_LIMIT 10
1009
1010 /* Send parameter to MGS*/
1011 static int mgc_set_mgs_param(struct obd_export *exp,
1012                              struct mgs_send_param *msp)
1013 {
1014         struct ptlrpc_request *req;
1015         struct mgs_send_param *req_msp, *rep_msp;
1016         int rc;
1017         ENTRY;
1018
1019         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1020                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1021                                         MGS_SET_INFO);
1022         if (!req)
1023                 RETURN(-ENOMEM);
1024
1025         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1026         if (!req_msp) {
1027                 ptlrpc_req_finished(req);
1028                 RETURN(-ENOMEM);
1029         }
1030
1031         memcpy(req_msp, msp, sizeof(*req_msp));
1032         ptlrpc_request_set_replen(req);
1033
1034         /* Limit how long we will wait for the enqueue to complete */
1035         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1036         rc = ptlrpc_queue_wait(req);
1037         if (!rc) {
1038                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1039                 memcpy(msp, rep_msp, sizeof(*rep_msp));
1040         }
1041
1042         ptlrpc_req_finished(req);
1043
1044         RETURN(rc);
1045 }
1046
1047 /* Take a config lock so we can get cancel notifications */
1048 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
1049                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
1050                        __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
1051                        void *data, __u32 lvb_len, void *lvb_swabber,
1052                        struct lustre_handle *lockh)
1053 {
1054         struct config_llog_data *cld = (struct config_llog_data *)data;
1055         struct ldlm_enqueue_info einfo = {
1056                 .ei_type        = type,
1057                 .ei_mode        = mode,
1058                 .ei_cb_bl       = mgc_blocking_ast,
1059                 .ei_cb_cp       = ldlm_completion_ast,
1060         };
1061         struct ptlrpc_request *req;
1062         int short_limit = cld_is_sptlrpc(cld);
1063         int rc;
1064         ENTRY;
1065
1066         CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
1067                cld->cld_resid.name[0]);
1068
1069         /* We need a callback for every lockholder, so don't try to
1070            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1071         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1072                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1073                                         LDLM_ENQUEUE);
1074         if (req == NULL)
1075                 RETURN(-ENOMEM);
1076
1077         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1078         ptlrpc_request_set_replen(req);
1079
1080         /* check if this is server or client */
1081         if (cld->cld_cfg.cfg_sb) {
1082                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1083                 if (lsi && IS_SERVER(lsi))
1084                         short_limit = 1;
1085         }
1086         /* Limit how long we will wait for the enqueue to complete */
1087         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1088         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1089                               NULL, 0, LVB_T_NONE, lockh, 0);
1090         /* A failed enqueue should still call the mgc_blocking_ast,
1091            where it will be requeued if needed ("grant failed"). */
1092         ptlrpc_req_finished(req);
1093         RETURN(rc);
1094 }
1095
1096 static int mgc_cancel(struct obd_export *exp, ldlm_mode_t mode,
1097                       struct lustre_handle *lockh)
1098 {
1099         ENTRY;
1100
1101         ldlm_lock_decref(lockh, mode);
1102
1103         RETURN(0);
1104 }
1105
1106 static void mgc_notify_active(struct obd_device *unused)
1107 {
1108         /* wakeup mgc_requeue_thread to requeue mgc lock */
1109         spin_lock(&config_list_lock);
1110         rq_state |= RQ_NOW;
1111         spin_unlock(&config_list_lock);
1112         wake_up(&rq_waitq);
1113
1114         /* TODO: Help the MGS rebuild nidtbl. -jay */
1115 }
1116
1117 /* Send target_reg message to MGS */
1118 static int mgc_target_register(struct obd_export *exp,
1119                                struct mgs_target_info *mti)
1120 {
1121         struct ptlrpc_request  *req;
1122         struct mgs_target_info *req_mti, *rep_mti;
1123         int                     rc;
1124         ENTRY;
1125
1126         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1127                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1128                                         MGS_TARGET_REG);
1129         if (req == NULL)
1130                 RETURN(-ENOMEM);
1131
1132         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1133         if (!req_mti) {
1134                 ptlrpc_req_finished(req);
1135                 RETURN(-ENOMEM);
1136         }
1137
1138         memcpy(req_mti, mti, sizeof(*req_mti));
1139         ptlrpc_request_set_replen(req);
1140         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1141         /* Limit how long we will wait for the enqueue to complete */
1142         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1143
1144         rc = ptlrpc_queue_wait(req);
1145         if (!rc) {
1146                 rep_mti = req_capsule_server_get(&req->rq_pill,
1147                                                  &RMF_MGS_TARGET_INFO);
1148                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1149                 CDEBUG(D_MGC, "register %s got index = %d\n",
1150                        mti->mti_svname, mti->mti_stripe_index);
1151         }
1152         ptlrpc_req_finished(req);
1153
1154         RETURN(rc);
1155 }
1156
1157 int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1158                        obd_count keylen, void *key, obd_count vallen,
1159                        void *val, struct ptlrpc_request_set *set)
1160 {
1161         int rc = -EINVAL;
1162         ENTRY;
1163
1164         /* Turn off initial_recov after we try all backup servers once */
1165         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1166                 struct obd_import *imp = class_exp2cliimp(exp);
1167                 int value;
1168                 if (vallen != sizeof(int))
1169                         RETURN(-EINVAL);
1170                 value = *(int *)val;
1171                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1172                        imp->imp_obd->obd_name, value,
1173                        imp->imp_deactive, imp->imp_invalid,
1174                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1175                        ptlrpc_import_state_name(imp->imp_state));
1176                 /* Resurrect if we previously died */
1177                 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1178                      imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1179                         ptlrpc_reconnect_import(imp);
1180                 RETURN(0);
1181         }
1182         /* FIXME move this to mgc_process_config */
1183         if (KEY_IS(KEY_REGISTER_TARGET)) {
1184                 struct mgs_target_info *mti;
1185                 if (vallen != sizeof(struct mgs_target_info))
1186                         RETURN(-EINVAL);
1187                 mti = (struct mgs_target_info *)val;
1188                 CDEBUG(D_MGC, "register_target %s %#x\n",
1189                        mti->mti_svname, mti->mti_flags);
1190                 rc =  mgc_target_register(exp, mti);
1191                 RETURN(rc);
1192         }
1193         if (KEY_IS(KEY_SET_FS)) {
1194                 struct super_block *sb = (struct super_block *)val;
1195
1196                 if (vallen != sizeof(struct super_block))
1197                         RETURN(-EINVAL);
1198
1199                 rc = mgc_fs_setup(exp->exp_obd, sb);
1200                 RETURN(rc);
1201         }
1202         if (KEY_IS(KEY_CLEAR_FS)) {
1203                 if (vallen != 0)
1204                         RETURN(-EINVAL);
1205                 rc = mgc_fs_cleanup(exp->exp_obd);
1206                 RETURN(rc);
1207         }
1208         if (KEY_IS(KEY_SET_INFO)) {
1209                 struct mgs_send_param *msp;
1210
1211                 msp = (struct mgs_send_param *)val;
1212                 rc =  mgc_set_mgs_param(exp, msp);
1213                 RETURN(rc);
1214         }
1215         if (KEY_IS(KEY_MGSSEC)) {
1216                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1217                 struct sptlrpc_flavor  flvr;
1218
1219                 /*
1220                  * empty string means using current flavor, if which haven't
1221                  * been set yet, set it as null.
1222                  *
1223                  * if flavor has been set previously, check the asking flavor
1224                  * must match the existing one.
1225                  */
1226                 if (vallen == 0) {
1227                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1228                                 RETURN(0);
1229                         val = "null";
1230                         vallen = 4;
1231                 }
1232
1233                 rc = sptlrpc_parse_flavor(val, &flvr);
1234                 if (rc) {
1235                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1236                                (char *) val);
1237                         RETURN(rc);
1238                 }
1239
1240                 /*
1241                  * caller already hold a mutex
1242                  */
1243                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1244                         cli->cl_flvr_mgc = flvr;
1245                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1246                                   sizeof(flvr)) != 0) {
1247                         char    str[20];
1248
1249                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1250                                             str, sizeof(str));
1251                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1252                                        "currently %s is in use\n",
1253                                        (char *) val, str);
1254                         rc = -EPERM;
1255                 }
1256                 RETURN(rc);
1257         }
1258
1259         RETURN(rc);
1260 }
1261
1262 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1263                         __u32 keylen, void *key, __u32 *vallen, void *val,
1264                         struct lov_stripe_md *unused)
1265 {
1266         int rc = -EINVAL;
1267
1268         if (KEY_IS(KEY_CONN_DATA)) {
1269                 struct obd_import *imp = class_exp2cliimp(exp);
1270                 struct obd_connect_data *data = val;
1271
1272                 if (*vallen == sizeof(*data)) {
1273                         *data = imp->imp_connect_data;
1274                         rc = 0;
1275                 }
1276         }
1277
1278         return rc;
1279 }
1280
1281 static int mgc_import_event(struct obd_device *obd,
1282                             struct obd_import *imp,
1283                             enum obd_import_event event)
1284 {
1285         int rc = 0;
1286
1287         LASSERT(imp->imp_obd == obd);
1288         CDEBUG(D_MGC, "import event %#x\n", event);
1289
1290         switch (event) {
1291         case IMP_EVENT_DISCON:
1292                 /* MGC imports should not wait for recovery */
1293                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1294                         ptlrpc_pinger_ir_down();
1295                 break;
1296         case IMP_EVENT_INACTIVE:
1297                 break;
1298         case IMP_EVENT_INVALIDATE: {
1299                 struct ldlm_namespace *ns = obd->obd_namespace;
1300                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1301                 break;
1302         }
1303         case IMP_EVENT_ACTIVE:
1304                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1305                 /* Clearing obd_no_recov allows us to continue pinging */
1306                 obd->obd_no_recov = 0;
1307                 mgc_notify_active(obd);
1308                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1309                         ptlrpc_pinger_ir_up();
1310                 break;
1311         case IMP_EVENT_OCD:
1312                 break;
1313         case IMP_EVENT_DEACTIVATE:
1314         case IMP_EVENT_ACTIVATE:
1315                 break;
1316         default:
1317                 CERROR("Unknown import event %#x\n", event);
1318                 LBUG();
1319         }
1320         RETURN(rc);
1321 }
1322
1323 enum {
1324         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_CACHE_SHIFT),
1325         CONFIG_READ_NRPAGES      = 4
1326 };
1327
1328 static int mgc_apply_recover_logs(struct obd_device *mgc,
1329                                   struct config_llog_data *cld,
1330                                   __u64 max_version,
1331                                   void *data, int datalen, bool mne_swab)
1332 {
1333         struct config_llog_instance *cfg = &cld->cld_cfg;
1334         struct lustre_sb_info       *lsi = s2lsi(cfg->cfg_sb);
1335         struct mgs_nidtbl_entry *entry;
1336         struct lustre_cfg       *lcfg;
1337         struct lustre_cfg_bufs   bufs;
1338         u64   prev_version = 0;
1339         char *inst;
1340         char *buf;
1341         int   bufsz;
1342         int   pos;
1343         int   rc  = 0;
1344         int   off = 0;
1345         ENTRY;
1346
1347         LASSERT(cfg->cfg_instance != NULL);
1348         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1349
1350         OBD_ALLOC(inst, PAGE_CACHE_SIZE);
1351         if (inst == NULL)
1352                 RETURN(-ENOMEM);
1353
1354         if (!IS_SERVER(lsi)) {
1355                 pos = snprintf(inst, PAGE_CACHE_SIZE, "%p", cfg->cfg_instance);
1356                 if (pos >= PAGE_CACHE_SIZE) {
1357                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1358                         return -E2BIG;
1359                 }
1360         } else {
1361                 LASSERT(IS_MDT(lsi));
1362                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1363                                         PAGE_CACHE_SIZE);
1364                 if (rc) {
1365                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1366                         RETURN(-EINVAL);
1367                 }
1368                 pos = strlen(inst);
1369         }
1370
1371         ++pos;
1372         buf   = inst + pos;
1373         bufsz = PAGE_CACHE_SIZE - pos;
1374
1375         while (datalen > 0) {
1376                 int   entry_len = sizeof(*entry);
1377                 int   is_ost;
1378                 struct obd_device *obd;
1379                 char *obdname;
1380                 char *cname;
1381                 char *params;
1382                 char *uuid;
1383
1384                 rc = -EINVAL;
1385                 if (datalen < sizeof(*entry))
1386                         break;
1387
1388                 entry = (typeof(entry))(data + off);
1389
1390                 /* sanity check */
1391                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1392                         break;
1393                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1394                         break;
1395                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1396                         break;
1397
1398                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1399                 if (datalen < entry_len) /* must have entry_len at least */
1400                         break;
1401
1402                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1403                 if (mne_swab)
1404                         lustre_swab_mgs_nidtbl_entry(entry);
1405                 if (entry->mne_length > PAGE_CACHE_SIZE) {
1406                         CERROR("MNE too large (%u)\n", entry->mne_length);
1407                         break;
1408                 }
1409
1410                 if (entry->mne_length < entry_len)
1411                         break;
1412
1413                 off     += entry->mne_length;
1414                 datalen -= entry->mne_length;
1415                 if (datalen < 0)
1416                         break;
1417
1418                 if (entry->mne_version > max_version) {
1419                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1420                                entry->mne_version, max_version);
1421                         break;
1422                 }
1423
1424                 if (prev_version >= entry->mne_version) {
1425                         CERROR("index unsorted, prev %lld, now %lld\n",
1426                                prev_version, entry->mne_version);
1427                         break;
1428                 }
1429                 prev_version = entry->mne_version;
1430
1431                 /*
1432                  * Write a string with format "nid::instance" to
1433                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1434                  */
1435
1436                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1437                 memset(buf, 0, bufsz);
1438                 obdname = buf;
1439                 pos = 0;
1440
1441                 /* lustre-OST0001-osc-<instance #> */
1442                 strcpy(obdname, cld->cld_logname);
1443                 cname = strrchr(obdname, '-');
1444                 if (cname == NULL) {
1445                         CERROR("mgc %s: invalid logname %s\n",
1446                                mgc->obd_name, obdname);
1447                         break;
1448                 }
1449
1450                 pos = cname - obdname;
1451                 obdname[pos] = 0;
1452                 pos += sprintf(obdname + pos, "-%s%04x",
1453                                   is_ost ? "OST" : "MDT", entry->mne_index);
1454
1455                 cname = is_ost ? "osc" : "mdc",
1456                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1457                 lustre_cfg_bufs_reset(&bufs, obdname);
1458
1459                 /* find the obd by obdname */
1460                 obd = class_name2obd(obdname);
1461                 if (obd == NULL) {
1462                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1463                                mgc->obd_name, obdname);
1464                         rc = 0;
1465                         /* this is a safe race, when the ost is starting up...*/
1466                         continue;
1467                 }
1468
1469                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1470                 ++pos;
1471                 params = buf + pos;
1472                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1473                 uuid = buf + pos;
1474
1475                 down_read(&obd->u.cli.cl_sem);
1476                 if (obd->u.cli.cl_import == NULL) {
1477                         /* client does not connect to the OST yet */
1478                         up_read(&obd->u.cli.cl_sem);
1479                         rc = 0;
1480                         continue;
1481                 }
1482
1483                 /* TODO: iterate all nids to find one */
1484                 /* find uuid by nid */
1485                 rc = client_import_find_conn(obd->u.cli.cl_import,
1486                                              entry->u.nids[0],
1487                                              (struct obd_uuid *)uuid);
1488                 up_read(&obd->u.cli.cl_sem);
1489                 if (rc < 0) {
1490                         CERROR("mgc: cannot find uuid by nid %s\n",
1491                                libcfs_nid2str(entry->u.nids[0]));
1492                         break;
1493                 }
1494
1495                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1496                        uuid, libcfs_nid2str(entry->u.nids[0]));
1497
1498                 pos += strlen(uuid);
1499                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1500                 LASSERT(pos < bufsz);
1501
1502                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1503
1504                 rc = -ENOMEM;
1505                 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1506                 if (lcfg == NULL) {
1507                         CERROR("mgc: cannot allocate memory\n");
1508                         break;
1509                 }
1510
1511                 CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n",
1512                        prev_version, max_version, obdname, params);
1513
1514                 rc = class_process_config(lcfg);
1515                 lustre_cfg_free(lcfg);
1516                 if (rc)
1517                         CDEBUG(D_INFO, "process config for %s error %d\n",
1518                                obdname, rc);
1519
1520                 /* continue, even one with error */
1521         }
1522
1523         OBD_FREE(inst, PAGE_CACHE_SIZE);
1524         RETURN(rc);
1525 }
1526
1527 /**
1528  * This function is called if this client was notified for target restarting
1529  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1530  */
1531 static int mgc_process_recover_log(struct obd_device *obd,
1532                                    struct config_llog_data *cld)
1533 {
1534         struct ptlrpc_request *req = NULL;
1535         struct config_llog_instance *cfg = &cld->cld_cfg;
1536         struct mgs_config_body *body;
1537         struct mgs_config_res  *res;
1538         struct ptlrpc_bulk_desc *desc;
1539         struct page **pages;
1540         int nrpages;
1541         bool eof = true;
1542         bool mne_swab = false;
1543         int i;
1544         int ealen;
1545         int rc;
1546         ENTRY;
1547
1548         /* allocate buffer for bulk transfer.
1549          * if this is the first time for this mgs to read logs,
1550          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1551          * once; otherwise, it only reads increment of logs, this should be
1552          * small and CONFIG_READ_NRPAGES will be used.
1553          */
1554         nrpages = CONFIG_READ_NRPAGES;
1555         if (cfg->cfg_last_idx == 0) /* the first time */
1556                 nrpages = CONFIG_READ_NRPAGES_INIT;
1557
1558         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1559         if (pages == NULL)
1560                 GOTO(out, rc = -ENOMEM);
1561
1562         for (i = 0; i < nrpages; i++) {
1563                 pages[i] = alloc_page(GFP_IOFS);
1564                 if (pages[i] == NULL)
1565                         GOTO(out, rc = -ENOMEM);
1566         }
1567
1568 again:
1569         LASSERT(cld_is_recover(cld));
1570         LASSERT(mutex_is_locked(&cld->cld_lock));
1571         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1572                                    &RQF_MGS_CONFIG_READ);
1573         if (req == NULL)
1574                 GOTO(out, rc = -ENOMEM);
1575
1576         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1577         if (rc)
1578                 GOTO(out, rc);
1579
1580         /* pack request */
1581         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1582         LASSERT(body != NULL);
1583         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1584         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1585             >= sizeof(body->mcb_name))
1586                 GOTO(out, rc = -E2BIG);
1587         body->mcb_offset = cfg->cfg_last_idx + 1;
1588         body->mcb_type   = cld->cld_type;
1589         body->mcb_bits   = PAGE_CACHE_SHIFT;
1590         body->mcb_units  = nrpages;
1591
1592         /* allocate bulk transfer descriptor */
1593         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1, BULK_PUT_SINK,
1594                                     MGS_BULK_PORTAL);
1595         if (desc == NULL)
1596                 GOTO(out, rc = -ENOMEM);
1597
1598         for (i = 0; i < nrpages; i++)
1599                 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
1600
1601         ptlrpc_request_set_replen(req);
1602         rc = ptlrpc_queue_wait(req);
1603         if (rc)
1604                 GOTO(out, rc);
1605
1606         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1607         if (res->mcr_size < res->mcr_offset)
1608                 GOTO(out, rc = -EINVAL);
1609
1610         /* always update the index even though it might have errors with
1611          * handling the recover logs */
1612         cfg->cfg_last_idx = res->mcr_offset;
1613         eof = res->mcr_offset == res->mcr_size;
1614
1615         CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n",
1616                res->mcr_offset, eof == false);
1617
1618         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1619         if (ealen < 0)
1620                 GOTO(out, rc = ealen);
1621
1622         if (ealen > nrpages << PAGE_CACHE_SHIFT)
1623                 GOTO(out, rc = -EINVAL);
1624
1625         if (ealen == 0) { /* no logs transferred */
1626                 if (!eof)
1627                         rc = -EINVAL;
1628                 GOTO(out, rc);
1629         }
1630
1631         mne_swab = !!ptlrpc_rep_need_swab(req);
1632 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1633         /* This import flag means the server did an extra swab of IR MNE
1634          * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1635         if (unlikely(req->rq_import->imp_need_mne_swab))
1636                 mne_swab = !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 (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1769                             llog_is_empty(env, lctxt, cld->cld_logname)) {
1770                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1771                                                    "log %s and no local copy."
1772                                                    "\n", cld->cld_logname);
1773                                 GOTO(out_pop, rc = -ENOENT);
1774                         }
1775                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1776                                "copy for now, will try to update later.\n",
1777                                cld->cld_logname);
1778                         rc = 0;
1779                 }
1780                 /* Now, whether we copied or not, start using the local llog.
1781                  * If we failed to copy, we'll start using whatever the old
1782                  * log has. */
1783                 llog_ctxt_put(ctxt);
1784                 ctxt = lctxt;
1785                 lctxt = NULL;
1786         } else {
1787                 if (local_only) /* no local log at client side */
1788                         GOTO(out_pop, rc = -EIO);
1789         }
1790
1791         if (cld_is_sptlrpc(cld)) {
1792                 sptlrpc_conf_log_update_begin(cld->cld_logname);
1793                 sptlrpc_started = true;
1794         }
1795
1796         /* logname and instance info should be the same, so use our
1797          * copy of the instance for the update.  The cfg_last_idx will
1798          * be updated here. */
1799         rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1800                                      &cld->cld_cfg);
1801         EXIT;
1802
1803 out_pop:
1804         __llog_ctxt_put(env, ctxt);
1805         if (lctxt)
1806                 __llog_ctxt_put(env, lctxt);
1807
1808         /*
1809          * update settings on existing OBDs. doing it inside
1810          * of llog_process_lock so no device is attaching/detaching
1811          * in parallel.
1812          * the logname must be <fsname>-sptlrpc
1813          */
1814         if (sptlrpc_started) {
1815                 LASSERT(cld_is_sptlrpc(cld));
1816                 sptlrpc_conf_log_update_end(cld->cld_logname);
1817                 class_notify_sptlrpc_conf(cld->cld_logname,
1818                                           strlen(cld->cld_logname) -
1819                                           strlen("-sptlrpc"));
1820         }
1821
1822         lu_env_fini(env);
1823 out_free:
1824         OBD_FREE_PTR(env);
1825         return rc;
1826 }
1827
1828 /** Get a config log from the MGS and process it.
1829  * This func is called for both clients and servers.
1830  * Copy the log locally before parsing it if appropriate (non-MGS server)
1831  */
1832 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1833 {
1834         struct lustre_handle lockh = { 0 };
1835         __u64 flags = LDLM_FL_NO_LRU;
1836         int rc = 0, rcl;
1837         ENTRY;
1838
1839         LASSERT(cld);
1840
1841         /* I don't want multiple processes running process_log at once --
1842            sounds like badness.  It actually might be fine, as long as
1843            we're not trying to update from the same log
1844            simultaneously (in which case we should use a per-log sem.) */
1845         mutex_lock(&cld->cld_lock);
1846         if (cld->cld_stopping) {
1847                 mutex_unlock(&cld->cld_lock);
1848                 RETURN(0);
1849         }
1850
1851         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1852
1853         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1854                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1855
1856         /* Get the cfg lock on the llog */
1857         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1858                           LCK_CR, &flags, NULL, NULL, NULL,
1859                           cld, 0, NULL, &lockh);
1860         if (rcl == 0) {
1861                 /* Get the cld, it will be released in mgc_blocking_ast. */
1862                 config_log_get(cld);
1863                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1864                 LASSERT(rc == 0);
1865         } else {
1866                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1867
1868                 /* mark cld_lostlock so that it will requeue
1869                  * after MGC becomes available. */
1870                 cld->cld_lostlock = 1;
1871                 /* Get extra reference, it will be put in requeue thread */
1872                 config_log_get(cld);
1873         }
1874
1875
1876         if (cld_is_recover(cld)) {
1877                 rc = 0; /* this is not a fatal error for recover log */
1878                 if (rcl == 0)
1879                         rc = mgc_process_recover_log(mgc, cld);
1880         } else {
1881                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1882         }
1883
1884         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1885                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1886
1887         mutex_unlock(&cld->cld_lock);
1888
1889         /* Now drop the lock so MGS can revoke it */
1890         if (!rcl) {
1891                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
1892                 if (rcl)
1893                         CERROR("Can't drop cfg lock: %d\n", rcl);
1894         }
1895
1896         RETURN(rc);
1897 }
1898
1899
1900 /** Called from lustre_process_log.
1901  * LCFG_LOG_START gets the config log from the MGS, processes it to start
1902  * any services, and adds it to the list logs to watch (follow).
1903  */
1904 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1905 {
1906         struct lustre_cfg *lcfg = buf;
1907         struct config_llog_instance *cfg = NULL;
1908         char *logname;
1909         int rc = 0;
1910         ENTRY;
1911
1912         switch(lcfg->lcfg_command) {
1913         case LCFG_LOV_ADD_OBD: {
1914                 /* Overloading this cfg command: register a new target */
1915                 struct mgs_target_info *mti;
1916
1917                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1918                     sizeof(struct mgs_target_info))
1919                         GOTO(out, rc = -EINVAL);
1920
1921                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1922                 CDEBUG(D_MGC, "add_target %s %#x\n",
1923                        mti->mti_svname, mti->mti_flags);
1924                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1925                 break;
1926         }
1927         case LCFG_LOV_DEL_OBD:
1928                 /* Unregister has no meaning at the moment. */
1929                 CERROR("lov_del_obd unimplemented\n");
1930                 rc = -ENOSYS;
1931                 break;
1932         case LCFG_SPTLRPC_CONF: {
1933                 rc = sptlrpc_process_config(lcfg);
1934                 break;
1935         }
1936         case LCFG_LOG_START: {
1937                 struct config_llog_data *cld;
1938                 struct super_block *sb;
1939
1940                 logname = lustre_cfg_string(lcfg, 1);
1941                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1942                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1943
1944                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1945                        cfg->cfg_last_idx);
1946
1947                 /* We're only called through here on the initial mount */
1948                 rc = config_log_add(obd, logname, cfg, sb);
1949                 if (rc)
1950                         break;
1951                 cld = config_log_find(logname, cfg);
1952                 if (cld == NULL) {
1953                         rc = -ENOENT;
1954                         break;
1955                 }
1956
1957                 /* COMPAT_146 */
1958                 /* FIXME only set this for old logs!  Right now this forces
1959                    us to always skip the "inside markers" check */
1960                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1961
1962                 rc = mgc_process_log(obd, cld);
1963                 if (rc == 0 && cld->cld_recover != NULL) {
1964                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1965                                          imp_connect_data, IMP_RECOV)) {
1966                                 rc = mgc_process_log(obd, cld->cld_recover);
1967                         } else {
1968                                 struct config_llog_data *cir = cld->cld_recover;
1969                                 cld->cld_recover = NULL;
1970                                 config_log_put(cir);
1971                         }
1972                         if (rc)
1973                                 CERROR("Cannot process recover llog %d\n", rc);
1974                 }
1975
1976                 if (rc == 0 && cld->cld_params != NULL) {
1977                         rc = mgc_process_log(obd, cld->cld_params);
1978                         if (rc == -ENOENT) {
1979                                 CDEBUG(D_MGC, "There is no params "
1980                                               "config file yet\n");
1981                                 rc = 0;
1982                         }
1983                         /* params log is optional */
1984                         if (rc)
1985                                 CERROR("%s: can't process params llog: rc = %d\n",
1986                                        obd->obd_name, rc);
1987                 }
1988                 config_log_put(cld);
1989
1990                 break;
1991         }
1992         case LCFG_LOG_END: {
1993                 logname = lustre_cfg_string(lcfg, 1);
1994
1995                 if (lcfg->lcfg_bufcount >= 2)
1996                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
1997                                 lcfg, 2);
1998                 rc = config_log_end(logname, cfg);
1999                 break;
2000         }
2001         default: {
2002                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2003                 GOTO(out, rc = -EINVAL);
2004
2005         }
2006         }
2007 out:
2008         RETURN(rc);
2009 }
2010
2011 struct obd_ops mgc_obd_ops = {
2012         .o_owner        = THIS_MODULE,
2013         .o_setup        = mgc_setup,
2014         .o_precleanup   = mgc_precleanup,
2015         .o_cleanup      = mgc_cleanup,
2016         .o_add_conn     = client_import_add_conn,
2017         .o_del_conn     = client_import_del_conn,
2018         .o_connect      = client_connect_import,
2019         .o_disconnect   = client_disconnect_export,
2020         .o_set_info_async = mgc_set_info_async,
2021         .o_get_info       = mgc_get_info,
2022         .o_import_event = mgc_import_event,
2023         .o_process_config = mgc_process_config,
2024 };
2025
2026 int __init mgc_init(void)
2027 {
2028         return class_register_type(&mgc_obd_ops, NULL, true, NULL,
2029 #ifndef HAVE_ONLY_PROCFS_SEQ
2030                                    NULL,
2031 #endif
2032                                    LUSTRE_MGC_NAME, NULL);
2033 }
2034
2035 static void /*__exit*/ mgc_exit(void)
2036 {
2037         class_unregister_type(LUSTRE_MGC_NAME);
2038 }
2039
2040 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2041 MODULE_DESCRIPTION("Lustre Management Client");
2042 MODULE_LICENSE("GPL");
2043
2044 module_init(mgc_init);
2045 module_exit(mgc_exit);