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         cfs_atomic_inc(&cld->cld_refcount);
 124         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 125                cfs_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                cfs_atomic_read(&cld->cld_refcount));
 137         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 138
 139         /* spinlock to make sure no item with 0 refcount in the list */
 140         if (cfs_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                 cfs_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         cfs_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 int lprocfs_mgc_rd_ir_state(char *page, char **start, off_t off,
 466                             int count, int *eof, 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         int rc = 0;
 473         ENTRY;
 474
 475         rc = snprintf(page, count, "imperative_recovery: %s\n",
 476                       OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 477         rc += snprintf(page + rc, count - rc, "client_state:\n");
 478
 479         spin_lock(&config_list_lock);
 480         cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 481                 if (cld->cld_recover == NULL)
 482                         continue;
 483                 rc += snprintf(page + rc, count - rc,
 484                                "    - { client: %s, nidtbl_version: %u }\n",
 485                                cld->cld_logname,
 486                                cld->cld_recover->cld_cfg.cfg_last_idx);
 487         }
 488         spin_unlock(&config_list_lock);
 489
 490         RETURN(rc);
 491 }
 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         ENTRY;
 505         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 506
 507         /* Do not run mgc_process_log on a disconnected export or an
 508            export which is being disconnected. Take the client
 509            semaphore to make the check non-racy. */
 510         down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 511         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 512                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 513                 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 514         } else {
 515                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 516                        cld->cld_logname);
 517         }
 518         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 519
 520         EXIT;
 521 }
 522
 523 /* this timeout represents how many seconds MGC should wait before
 524  * requeue config and recover lock to the MGS. We need to randomize this
 525  * in order to not flood the MGS.
 526  */
 527 #define MGC_TIMEOUT_MIN_SECONDS   5
 528 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
 529
 530 static int mgc_requeue_thread(void *data)
 531 {
 532         int rc = 0;
 533         ENTRY;
 534
 535         CDEBUG(D_MGC, "Starting requeue thread\n");
 536
 537         /* Keep trying failed locks periodically */
 538         spin_lock(&config_list_lock);
 539         rq_state |= RQ_RUNNING;
 540         while (1) {
 541                 struct l_wait_info lwi;
 542                 struct config_llog_data *cld, *cld_prev;
 543                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
 544                 int stopped = !!(rq_state & RQ_STOP);
 545                 int to;
 546
 547                 /* Any new or requeued lostlocks will change the state */
 548                 rq_state &= ~(RQ_NOW | RQ_LATER);
 549                 spin_unlock(&config_list_lock);
 550
 551                 /* Always wait a few seconds to allow the server who
 552                    caused the lock revocation to finish its setup, plus some
 553                    random so everyone doesn't try to reconnect at once. */
 554                 to = MGC_TIMEOUT_MIN_SECONDS * HZ;
 555                 to += rand * HZ / 100; /* rand is centi-seconds */
 556                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 557                 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi);
 558
 559                 /*
 560                  * iterate & processing through the list. for each cld, process
 561                  * its depending sptlrpc cld firstly (if any) and then itself.
 562                  *
 563                  * it's guaranteed any item in the list must have
 564                  * reference > 0; and if cld_lostlock is set, at
 565                  * least one reference is taken by the previous enqueue.
 566                  */
 567                 cld_prev = NULL;
 568
 569                 spin_lock(&config_list_lock);
 570                 cfs_list_for_each_entry(cld, &config_llog_list,
 571                                         cld_list_chain) {
 572                         if (!cld->cld_lostlock)
 573                                 continue;
 574
 575                         spin_unlock(&config_list_lock);
 576
 577                         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 578
 579                         /* Whether we enqueued again or not in mgc_process_log,
 580                          * we're done with the ref from the old enqueue */
 581                         if (cld_prev)
 582                                 config_log_put(cld_prev);
 583                         cld_prev = cld;
 584
 585                         cld->cld_lostlock = 0;
 586                         if (likely(!stopped))
 587                                 do_requeue(cld);
 588
 589                         spin_lock(&config_list_lock);
 590                 }
 591                 spin_unlock(&config_list_lock);
 592                 if (cld_prev)
 593                         config_log_put(cld_prev);
 594
 595                 /* break after scanning the list so that we can drop
 596                  * refcount to losing lock clds */
 597                 if (unlikely(stopped)) {
 598                         spin_lock(&config_list_lock);
 599                         break;
 600                 }
 601
 602                 /* Wait a bit to see if anyone else needs a requeue */
 603                 lwi = (struct l_wait_info) { 0 };
 604                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 605                              &lwi);
 606                 spin_lock(&config_list_lock);
 607         }
 608         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 609         rq_state &= ~RQ_RUNNING;
 610         spin_unlock(&config_list_lock);
 611
 612         complete(&rq_exit);
 613
 614         CDEBUG(D_MGC, "Ending requeue thread\n");
 615         RETURN(rc);
 616 }
 617
 618 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 619    We are responsible for dropping the config log reference from here on out. */
 620 static void mgc_requeue_add(struct config_llog_data *cld)
 621 {
 622         ENTRY;
 623
 624         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 625                cld->cld_logname, cfs_atomic_read(&cld->cld_refcount),
 626                cld->cld_stopping, rq_state);
 627         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 628
 629         mutex_lock(&cld->cld_lock);
 630         if (cld->cld_stopping || cld->cld_lostlock) {
 631                 mutex_unlock(&cld->cld_lock);
 632                 RETURN_EXIT;
 633         }
 634         /* this refcount will be released in mgc_requeue_thread. */
 635         config_log_get(cld);
 636         cld->cld_lostlock = 1;
 637         mutex_unlock(&cld->cld_lock);
 638
 639         /* Hold lock for rq_state */
 640         spin_lock(&config_list_lock);
 641         if (rq_state & RQ_STOP) {
 642                 spin_unlock(&config_list_lock);
 643                 cld->cld_lostlock = 0;
 644                 config_log_put(cld);
 645         } else {
 646                 rq_state |= RQ_NOW;
 647                 spin_unlock(&config_list_lock);
 648                 wake_up(&rq_waitq);
 649         }
 650         EXIT;
 651 }
 652
 653 /********************** class fns **********************/
 654 static int mgc_local_llog_init(const struct lu_env *env,
 655                                struct obd_device *obd,
 656                                struct obd_device *disk)
 657 {
 658         struct llog_ctxt        *ctxt;
 659         int                      rc;
 660
 661         ENTRY;
 662
 663         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
 664                         &llog_osd_ops);
 665         if (rc)
 666                 RETURN(rc);
 667
 668         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 669         LASSERT(ctxt);
 670         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
 671         llog_ctxt_put(ctxt);
 672
 673         RETURN(0);
 674 }
 675
 676 static int mgc_local_llog_fini(const struct lu_env *env,
 677                                struct obd_device *obd)
 678 {
 679         struct llog_ctxt *ctxt;
 680
 681         ENTRY;
 682
 683         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 684         llog_cleanup(env, ctxt);
 685
 686         RETURN(0);
 687 }
 688
 689 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb)
 690 {
 691         struct lustre_sb_info   *lsi = s2lsi(sb);
 692         struct client_obd       *cli = &obd->u.cli;
 693         struct lu_fid            rfid, fid;
 694         struct dt_object        *root, *dto;
 695         struct lu_env           *env;
 696         int                      rc = 0;
 697
 698         ENTRY;
 699
 700         LASSERT(lsi);
 701         LASSERT(lsi->lsi_dt_dev);
 702
 703         OBD_ALLOC_PTR(env);
 704         if (env == NULL)
 705                 RETURN(-ENOMEM);
 706
 707         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
 708         mutex_lock(&cli->cl_mgc_mutex);
 709
 710         cfs_cleanup_group_info();
 711
 712         /* Setup the configs dir */
 713         rc = lu_env_init(env, LCT_MG_THREAD);
 714         if (rc)
 715                 GOTO(out_err, rc);
 716
 717         fid.f_seq = FID_SEQ_LOCAL_NAME;
 718         fid.f_oid = 1;
 719         fid.f_ver = 0;
 720         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
 721                                     &cli->cl_mgc_los);
 722         if (rc)
 723                 GOTO(out_env, rc);
 724
 725         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
 726         if (rc)
 727                 GOTO(out_env, rc);
 728
 729         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
 730                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
 731         if (unlikely(IS_ERR(root)))
 732                 GOTO(out_los, rc = PTR_ERR(root));
 733
 734         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
 735                                         MOUNT_CONFIGS_DIR,
 736                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
 737         lu_object_put_nocache(env, &root->do_lu);
 738         if (IS_ERR(dto))
 739                 GOTO(out_los, rc = PTR_ERR(dto));
 740
 741         cli->cl_mgc_configs_dir = dto;
 742
 743         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
 744         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
 745         if (rc)
 746                 GOTO(out_llog, rc);
 747
 748         /* We take an obd ref to insure that we can't get to mgc_cleanup
 749          * without calling mgc_fs_cleanup first. */
 750         class_incref(obd, "mgc_fs", obd);
 751
 752         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 753         EXIT;
 754 out_llog:
 755         if (rc) {
 756                 lu_object_put(env, &cli->cl_mgc_configs_dir->do_lu);
 757                 cli->cl_mgc_configs_dir = NULL;
 758         }
 759 out_los:
 760         if (rc < 0) {
 761                 local_oid_storage_fini(env, cli->cl_mgc_los);
 762                 cli->cl_mgc_los = NULL;
 763                 mutex_unlock(&cli->cl_mgc_mutex);
 764         }
 765 out_env:
 766         lu_env_fini(env);
 767 out_err:
 768         OBD_FREE_PTR(env);
 769         return rc;
 770 }
 771
 772 static int mgc_fs_cleanup(struct obd_device *obd)
 773 {
 774         struct lu_env            env;
 775         struct client_obd       *cli = &obd->u.cli;
 776         int                      rc;
 777
 778         ENTRY;
 779
 780         LASSERT(cli->cl_mgc_los != NULL);
 781
 782         rc = lu_env_init(&env, LCT_MG_THREAD);
 783         if (rc)
 784                 GOTO(unlock, rc);
 785
 786         mgc_local_llog_fini(&env, obd);
 787
 788         lu_object_put_nocache(&env, &cli->cl_mgc_configs_dir->do_lu);
 789         cli->cl_mgc_configs_dir = NULL;
 790
 791         local_oid_storage_fini(&env, cli->cl_mgc_los);
 792         cli->cl_mgc_los = NULL;
 793         lu_env_fini(&env);
 794
 795 unlock:
 796         class_decref(obd, "mgc_fs", obd);
 797         mutex_unlock(&cli->cl_mgc_mutex);
 798
 799         RETURN(0);
 800 }
 801
 802 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 803 {
 804         struct llog_ctxt        *ctxt;
 805         int                      rc;
 806
 807         ENTRY;
 808
 809         /* setup only remote ctxt, the local disk context is switched per each
 810          * filesystem during mgc_fs_setup() */
 811         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 812                         &llog_client_ops);
 813         if (rc)
 814                 RETURN(rc);
 815
 816         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 817         LASSERT(ctxt);
 818
 819         llog_initiator_connect(ctxt);
 820         llog_ctxt_put(ctxt);
 821
 822         RETURN(0);
 823 }
 824
 825 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 826 {
 827         struct llog_ctxt *ctxt;
 828
 829         ENTRY;
 830
 831         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 832         if (ctxt)
 833                 llog_cleanup(env, ctxt);
 834
 835         RETURN(0);
 836 }
 837
 838
 839 static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0);
 840 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 841 {
 842         int rc = 0;
 843         ENTRY;
 844
 845         switch (stage) {
 846         case OBD_CLEANUP_EARLY:
 847                 break;
 848         case OBD_CLEANUP_EXPORTS:
 849                 if (cfs_atomic_dec_and_test(&mgc_count)) {
 850                         int running;
 851                         /* stop requeue thread */
 852                         spin_lock(&config_list_lock);
 853                         running = rq_state & RQ_RUNNING;
 854                         if (running)
 855                                 rq_state |= RQ_STOP;
 856                         spin_unlock(&config_list_lock);
 857                         if (running) {
 858                                 wake_up(&rq_waitq);
 859                                 wait_for_completion(&rq_exit);
 860                         }
 861                 }
 862                 obd_cleanup_client_import(obd);
 863                 rc = mgc_llog_fini(NULL, obd);
 864                 if (rc != 0)
 865                         CERROR("failed to cleanup llogging subsystems\n");
 866                 break;
 867         }
 868         RETURN(rc);
 869 }
 870
 871 static int mgc_cleanup(struct obd_device *obd)
 872 {
 873         int rc;
 874         ENTRY;
 875
 876         /* COMPAT_146 - old config logs may have added profiles we don't
 877            know about */
 878         if (obd->obd_type->typ_refcnt <= 1)
 879                 /* Only for the last mgc */
 880                 class_del_profiles();
 881
 882         lprocfs_obd_cleanup(obd);
 883         ptlrpcd_decref();
 884
 885         rc = client_obd_cleanup(obd);
 886         RETURN(rc);
 887 }
 888
 889 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 890 {
 891         struct lprocfs_static_vars lvars;
 892         struct task_struct         *task;
 893         int                        rc;
 894         ENTRY;
 895
 896         rc = ptlrpcd_addref();
 897         if (rc < 0)
 898                 RETURN(rc);
 899
 900         rc = client_obd_setup(obd, lcfg);
 901         if (rc)
 902                 GOTO(err_decref, rc);
 903
 904         rc = mgc_llog_init(NULL, obd);
 905         if (rc) {
 906                 CERROR("failed to setup llogging subsystems\n");
 907                 GOTO(err_cleanup, rc);
 908         }
 909
 910         lprocfs_mgc_init_vars(&lvars);
 911         lprocfs_obd_setup(obd, lvars.obd_vars);
 912         sptlrpc_lprocfs_cliobd_attach(obd);
 913
 914         if (cfs_atomic_inc_return(&mgc_count) == 1) {
 915                 rq_state = 0;
 916                 init_waitqueue_head(&rq_waitq);
 917
 918                 /* start requeue thread */
 919                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 920                 if (IS_ERR(task)) {
 921                         rc = PTR_ERR(task);
 922                         CERROR("%s: cannot start requeue thread: rc = %d; "
 923                                "no more log updates\n",
 924                                obd->obd_name, rc);
 925                         GOTO(err_cleanup, rc);
 926                 }
 927                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 928                 rc = 0;
 929         }
 930
 931         RETURN(rc);
 932
 933 err_cleanup:
 934         client_obd_cleanup(obd);
 935 err_decref:
 936         ptlrpcd_decref();
 937         RETURN(rc);
 938 }
 939
 940 /* based on ll_mdc_blocking_ast */
 941 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 942                             void *data, int flag)
 943 {
 944         struct lustre_handle lockh;
 945         struct config_llog_data *cld = (struct config_llog_data *)data;
 946         int rc = 0;
 947         ENTRY;
 948
 949         switch (flag) {
 950         case LDLM_CB_BLOCKING:
 951                 /* mgs wants the lock, give it up... */
 952                 LDLM_DEBUG(lock, "MGC blocking CB");
 953                 ldlm_lock2handle(lock, &lockh);
 954                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 955                 break;
 956         case LDLM_CB_CANCELING:
 957                 /* We've given up the lock, prepare ourselves to update. */
 958                 LDLM_DEBUG(lock, "MGC cancel CB");
 959
 960                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
 961                        PLDLMRES(lock->l_resource),
 962                        (char *)&lock->l_resource->lr_name.name[0]);
 963
 964                 if (!cld) {
 965                         CDEBUG(D_INFO, "missing data, won't requeue\n");
 966                         break;
 967                 }
 968
 969                 /* held at mgc_process_log(). */
 970                 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 971                 /* Are we done with this log? */
 972                 if (cld->cld_stopping) {
 973                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
 974                                cld->cld_logname);
 975                         config_log_put(cld);
 976                         break;
 977                 }
 978                 /* Make sure not to re-enqueue when the mgc is stopping
 979                    (we get called from client_disconnect_export) */
 980                 if (!lock->l_conn_export ||
 981                     !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
 982                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
 983                                cld->cld_logname);
 984                         config_log_put(cld);
 985                         break;
 986                 }
 987
 988                 /* Re-enqueue now */
 989                 mgc_requeue_add(cld);
 990                 config_log_put(cld);
 991                 break;
 992         default:
 993                 LBUG();
 994         }
 995
 996         RETURN(rc);
 997 }
 998
 999 /* Not sure where this should go... */
1000 #define  MGC_ENQUEUE_LIMIT 50
1001 #define  MGC_TARGET_REG_LIMIT 10
1002 #define  MGC_SEND_PARAM_LIMIT 10
1003
1004 /* Send parameter to MGS*/
1005 static int mgc_set_mgs_param(struct obd_export *exp,
1006                              struct mgs_send_param *msp)
1007 {
1008         struct ptlrpc_request *req;
1009         struct mgs_send_param *req_msp, *rep_msp;
1010         int rc;
1011         ENTRY;
1012
1013         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1014                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1015                                         MGS_SET_INFO);
1016         if (!req)
1017                 RETURN(-ENOMEM);
1018
1019         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1020         if (!req_msp) {
1021                 ptlrpc_req_finished(req);
1022                 RETURN(-ENOMEM);
1023         }
1024
1025         memcpy(req_msp, msp, sizeof(*req_msp));
1026         ptlrpc_request_set_replen(req);
1027
1028         /* Limit how long we will wait for the enqueue to complete */
1029         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1030         rc = ptlrpc_queue_wait(req);
1031         if (!rc) {
1032                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1033                 memcpy(msp, rep_msp, sizeof(*rep_msp));
1034         }
1035
1036         ptlrpc_req_finished(req);
1037
1038         RETURN(rc);
1039 }
1040
1041 /* Take a config lock so we can get cancel notifications */
1042 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
1043                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
1044                        __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
1045                        void *data, __u32 lvb_len, void *lvb_swabber,
1046                        struct lustre_handle *lockh)
1047 {
1048         struct config_llog_data *cld = (struct config_llog_data *)data;
1049         struct ldlm_enqueue_info einfo = {
1050                 .ei_type        = type,
1051                 .ei_mode        = mode,
1052                 .ei_cb_bl       = mgc_blocking_ast,
1053                 .ei_cb_cp       = ldlm_completion_ast,
1054         };
1055         struct ptlrpc_request *req;
1056         int short_limit = cld_is_sptlrpc(cld);
1057         int rc;
1058         ENTRY;
1059
1060         CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
1061                cld->cld_resid.name[0]);
1062
1063         /* We need a callback for every lockholder, so don't try to
1064            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1065         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1066                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1067                                         LDLM_ENQUEUE);
1068         if (req == NULL)
1069                 RETURN(-ENOMEM);
1070
1071         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1072         ptlrpc_request_set_replen(req);
1073
1074         /* check if this is server or client */
1075         if (cld->cld_cfg.cfg_sb) {
1076                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1077                 if (lsi && IS_SERVER(lsi))
1078                         short_limit = 1;
1079         }
1080         /* Limit how long we will wait for the enqueue to complete */
1081         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1082         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1083                               NULL, 0, LVB_T_NONE, lockh, 0);
1084         /* A failed enqueue should still call the mgc_blocking_ast,
1085            where it will be requeued if needed ("grant failed"). */
1086         ptlrpc_req_finished(req);
1087         RETURN(rc);
1088 }
1089
1090 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
1091                       __u32 mode, struct lustre_handle *lockh)
1092 {
1093         ENTRY;
1094
1095         ldlm_lock_decref(lockh, mode);
1096
1097         RETURN(0);
1098 }
1099
1100 static void mgc_notify_active(struct obd_device *unused)
1101 {
1102         /* wakeup mgc_requeue_thread to requeue mgc lock */
1103         spin_lock(&config_list_lock);
1104         rq_state |= RQ_NOW;
1105         spin_unlock(&config_list_lock);
1106         wake_up(&rq_waitq);
1107
1108         /* TODO: Help the MGS rebuild nidtbl. -jay */
1109 }
1110
1111 /* Send target_reg message to MGS */
1112 static int mgc_target_register(struct obd_export *exp,
1113                                struct mgs_target_info *mti)
1114 {
1115         struct ptlrpc_request  *req;
1116         struct mgs_target_info *req_mti, *rep_mti;
1117         int                     rc;
1118         ENTRY;
1119
1120         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1121                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1122                                         MGS_TARGET_REG);
1123         if (req == NULL)
1124                 RETURN(-ENOMEM);
1125
1126         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1127         if (!req_mti) {
1128                 ptlrpc_req_finished(req);
1129                 RETURN(-ENOMEM);
1130         }
1131
1132         memcpy(req_mti, mti, sizeof(*req_mti));
1133         ptlrpc_request_set_replen(req);
1134         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1135         /* Limit how long we will wait for the enqueue to complete */
1136         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1137
1138         rc = ptlrpc_queue_wait(req);
1139         if (!rc) {
1140                 rep_mti = req_capsule_server_get(&req->rq_pill,
1141                                                  &RMF_MGS_TARGET_INFO);
1142                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1143                 CDEBUG(D_MGC, "register %s got index = %d\n",
1144                        mti->mti_svname, mti->mti_stripe_index);
1145         }
1146         ptlrpc_req_finished(req);
1147
1148         RETURN(rc);
1149 }
1150
1151 int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1152                        obd_count keylen, void *key, obd_count vallen,
1153                        void *val, struct ptlrpc_request_set *set)
1154 {
1155         int rc = -EINVAL;
1156         ENTRY;
1157
1158         /* Turn off initial_recov after we try all backup servers once */
1159         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1160                 struct obd_import *imp = class_exp2cliimp(exp);
1161                 int value;
1162                 if (vallen != sizeof(int))
1163                         RETURN(-EINVAL);
1164                 value = *(int *)val;
1165                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1166                        imp->imp_obd->obd_name, value,
1167                        imp->imp_deactive, imp->imp_invalid,
1168                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1169                        ptlrpc_import_state_name(imp->imp_state));
1170                 /* Resurrect if we previously died */
1171                 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1172                      imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1173                         ptlrpc_reconnect_import(imp);
1174                 RETURN(0);
1175         }
1176         /* FIXME move this to mgc_process_config */
1177         if (KEY_IS(KEY_REGISTER_TARGET)) {
1178                 struct mgs_target_info *mti;
1179                 if (vallen != sizeof(struct mgs_target_info))
1180                         RETURN(-EINVAL);
1181                 mti = (struct mgs_target_info *)val;
1182                 CDEBUG(D_MGC, "register_target %s %#x\n",
1183                        mti->mti_svname, mti->mti_flags);
1184                 rc =  mgc_target_register(exp, mti);
1185                 RETURN(rc);
1186         }
1187         if (KEY_IS(KEY_SET_FS)) {
1188                 struct super_block *sb = (struct super_block *)val;
1189
1190                 if (vallen != sizeof(struct super_block))
1191                         RETURN(-EINVAL);
1192
1193                 rc = mgc_fs_setup(exp->exp_obd, sb);
1194                 RETURN(rc);
1195         }
1196         if (KEY_IS(KEY_CLEAR_FS)) {
1197                 if (vallen != 0)
1198                         RETURN(-EINVAL);
1199                 rc = mgc_fs_cleanup(exp->exp_obd);
1200                 RETURN(rc);
1201         }
1202         if (KEY_IS(KEY_SET_INFO)) {
1203                 struct mgs_send_param *msp;
1204
1205                 msp = (struct mgs_send_param *)val;
1206                 rc =  mgc_set_mgs_param(exp, msp);
1207                 RETURN(rc);
1208         }
1209         if (KEY_IS(KEY_MGSSEC)) {
1210                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1211                 struct sptlrpc_flavor  flvr;
1212
1213                 /*
1214                  * empty string means using current flavor, if which haven't
1215                  * been set yet, set it as null.
1216                  *
1217                  * if flavor has been set previously, check the asking flavor
1218                  * must match the existing one.
1219                  */
1220                 if (vallen == 0) {
1221                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1222                                 RETURN(0);
1223                         val = "null";
1224                         vallen = 4;
1225                 }
1226
1227                 rc = sptlrpc_parse_flavor(val, &flvr);
1228                 if (rc) {
1229                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1230                                (char *) val);
1231                         RETURN(rc);
1232                 }
1233
1234                 /*
1235                  * caller already hold a mutex
1236                  */
1237                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1238                         cli->cl_flvr_mgc = flvr;
1239                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1240                                   sizeof(flvr)) != 0) {
1241                         char    str[20];
1242
1243                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1244                                             str, sizeof(str));
1245                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1246                                        "currently %s is in use\n",
1247                                        (char *) val, str);
1248                         rc = -EPERM;
1249                 }
1250                 RETURN(rc);
1251         }
1252
1253         RETURN(rc);
1254 }
1255
1256 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1257                         __u32 keylen, void *key, __u32 *vallen, void *val,
1258                         struct lov_stripe_md *unused)
1259 {
1260         int rc = -EINVAL;
1261
1262         if (KEY_IS(KEY_CONN_DATA)) {
1263                 struct obd_import *imp = class_exp2cliimp(exp);
1264                 struct obd_connect_data *data = val;
1265
1266                 if (*vallen == sizeof(*data)) {
1267                         *data = imp->imp_connect_data;
1268                         rc = 0;
1269                 }
1270         }
1271
1272         return rc;
1273 }
1274
1275 static int mgc_import_event(struct obd_device *obd,
1276                             struct obd_import *imp,
1277                             enum obd_import_event event)
1278 {
1279         int rc = 0;
1280
1281         LASSERT(imp->imp_obd == obd);
1282         CDEBUG(D_MGC, "import event %#x\n", event);
1283
1284         switch (event) {
1285         case IMP_EVENT_DISCON:
1286                 /* MGC imports should not wait for recovery */
1287                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1288                         ptlrpc_pinger_ir_down();
1289                 break;
1290         case IMP_EVENT_INACTIVE:
1291                 break;
1292         case IMP_EVENT_INVALIDATE: {
1293                 struct ldlm_namespace *ns = obd->obd_namespace;
1294                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1295                 break;
1296         }
1297         case IMP_EVENT_ACTIVE:
1298                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1299                 /* Clearing obd_no_recov allows us to continue pinging */
1300                 obd->obd_no_recov = 0;
1301                 mgc_notify_active(obd);
1302                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1303                         ptlrpc_pinger_ir_up();
1304                 break;
1305         case IMP_EVENT_OCD:
1306                 break;
1307         case IMP_EVENT_DEACTIVATE:
1308         case IMP_EVENT_ACTIVATE:
1309                 break;
1310         default:
1311                 CERROR("Unknown import event %#x\n", event);
1312                 LBUG();
1313         }
1314         RETURN(rc);
1315 }
1316
1317 enum {
1318         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_CACHE_SHIFT),
1319         CONFIG_READ_NRPAGES      = 4
1320 };
1321
1322 static int mgc_apply_recover_logs(struct obd_device *mgc,
1323                                   struct config_llog_data *cld,
1324                                   __u64 max_version,
1325                                   void *data, int datalen, bool mne_swab)
1326 {
1327         struct config_llog_instance *cfg = &cld->cld_cfg;
1328         struct lustre_sb_info       *lsi = s2lsi(cfg->cfg_sb);
1329         struct mgs_nidtbl_entry *entry;
1330         struct lustre_cfg       *lcfg;
1331         struct lustre_cfg_bufs   bufs;
1332         u64   prev_version = 0;
1333         char *inst;
1334         char *buf;
1335         int   bufsz;
1336         int   pos;
1337         int   rc  = 0;
1338         int   off = 0;
1339         ENTRY;
1340
1341         LASSERT(cfg->cfg_instance != NULL);
1342         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1343
1344         OBD_ALLOC(inst, PAGE_CACHE_SIZE);
1345         if (inst == NULL)
1346                 RETURN(-ENOMEM);
1347
1348         if (!IS_SERVER(lsi)) {
1349                 pos = snprintf(inst, PAGE_CACHE_SIZE, "%p", cfg->cfg_instance);
1350                 if (pos >= PAGE_CACHE_SIZE) {
1351                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1352                         return -E2BIG;
1353                 }
1354         } else {
1355                 LASSERT(IS_MDT(lsi));
1356                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1357                                         PAGE_CACHE_SIZE);
1358                 if (rc) {
1359                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1360                         RETURN(-EINVAL);
1361                 }
1362                 pos = strlen(inst);
1363         }
1364
1365         ++pos;
1366         buf   = inst + pos;
1367         bufsz = PAGE_CACHE_SIZE - pos;
1368
1369         while (datalen > 0) {
1370                 int   entry_len = sizeof(*entry);
1371                 int   is_ost;
1372                 struct obd_device *obd;
1373                 char *obdname;
1374                 char *cname;
1375                 char *params;
1376                 char *uuid;
1377
1378                 rc = -EINVAL;
1379                 if (datalen < sizeof(*entry))
1380                         break;
1381
1382                 entry = (typeof(entry))(data + off);
1383
1384                 /* sanity check */
1385                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1386                         break;
1387                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1388                         break;
1389                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1390                         break;
1391
1392                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1393                 if (datalen < entry_len) /* must have entry_len at least */
1394                         break;
1395
1396                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1397                 if (mne_swab)
1398                         lustre_swab_mgs_nidtbl_entry(entry);
1399                 if (entry->mne_length > PAGE_CACHE_SIZE) {
1400                         CERROR("MNE too large (%u)\n", entry->mne_length);
1401                         break;
1402                 }
1403
1404                 if (entry->mne_length < entry_len)
1405                         break;
1406
1407                 off     += entry->mne_length;
1408                 datalen -= entry->mne_length;
1409                 if (datalen < 0)
1410                         break;
1411
1412                 if (entry->mne_version > max_version) {
1413                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1414                                entry->mne_version, max_version);
1415                         break;
1416                 }
1417
1418                 if (prev_version >= entry->mne_version) {
1419                         CERROR("index unsorted, prev %lld, now %lld\n",
1420                                prev_version, entry->mne_version);
1421                         break;
1422                 }
1423                 prev_version = entry->mne_version;
1424
1425                 /*
1426                  * Write a string with format "nid::instance" to
1427                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1428                  */
1429
1430                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1431                 memset(buf, 0, bufsz);
1432                 obdname = buf;
1433                 pos = 0;
1434
1435                 /* lustre-OST0001-osc-<instance #> */
1436                 strcpy(obdname, cld->cld_logname);
1437                 cname = strrchr(obdname, '-');
1438                 if (cname == NULL) {
1439                         CERROR("mgc %s: invalid logname %s\n",
1440                                mgc->obd_name, obdname);
1441                         break;
1442                 }
1443
1444                 pos = cname - obdname;
1445                 obdname[pos] = 0;
1446                 pos += sprintf(obdname + pos, "-%s%04x",
1447                                   is_ost ? "OST" : "MDT", entry->mne_index);
1448
1449                 cname = is_ost ? "osc" : "mdc",
1450                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1451                 lustre_cfg_bufs_reset(&bufs, obdname);
1452
1453                 /* find the obd by obdname */
1454                 obd = class_name2obd(obdname);
1455                 if (obd == NULL) {
1456                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1457                                mgc->obd_name, obdname);
1458                         rc = 0;
1459                         /* this is a safe race, when the ost is starting up...*/
1460                         continue;
1461                 }
1462
1463                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1464                 ++pos;
1465                 params = buf + pos;
1466                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1467                 uuid = buf + pos;
1468
1469                 down_read(&obd->u.cli.cl_sem);
1470                 if (obd->u.cli.cl_import == NULL) {
1471                         /* client does not connect to the OST yet */
1472                         up_read(&obd->u.cli.cl_sem);
1473                         rc = 0;
1474                         continue;
1475                 }
1476
1477                 /* TODO: iterate all nids to find one */
1478                 /* find uuid by nid */
1479                 rc = client_import_find_conn(obd->u.cli.cl_import,
1480                                              entry->u.nids[0],
1481                                              (struct obd_uuid *)uuid);
1482                 up_read(&obd->u.cli.cl_sem);
1483                 if (rc < 0) {
1484                         CERROR("mgc: cannot find uuid by nid %s\n",
1485                                libcfs_nid2str(entry->u.nids[0]));
1486                         break;
1487                 }
1488
1489                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1490                        uuid, libcfs_nid2str(entry->u.nids[0]));
1491
1492                 pos += strlen(uuid);
1493                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1494                 LASSERT(pos < bufsz);
1495
1496                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1497
1498                 rc = -ENOMEM;
1499                 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1500                 if (lcfg == NULL) {
1501                         CERROR("mgc: cannot allocate memory\n");
1502                         break;
1503                 }
1504
1505                 CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n",
1506                        prev_version, max_version, obdname, params);
1507
1508                 rc = class_process_config(lcfg);
1509                 lustre_cfg_free(lcfg);
1510                 if (rc)
1511                         CDEBUG(D_INFO, "process config for %s error %d\n",
1512                                obdname, rc);
1513
1514                 /* continue, even one with error */
1515         }
1516
1517         OBD_FREE(inst, PAGE_CACHE_SIZE);
1518         RETURN(rc);
1519 }
1520
1521 /**
1522  * This function is called if this client was notified for target restarting
1523  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1524  */
1525 static int mgc_process_recover_log(struct obd_device *obd,
1526                                    struct config_llog_data *cld)
1527 {
1528         struct ptlrpc_request *req = NULL;
1529         struct config_llog_instance *cfg = &cld->cld_cfg;
1530         struct mgs_config_body *body;
1531         struct mgs_config_res  *res;
1532         struct ptlrpc_bulk_desc *desc;
1533         struct page **pages;
1534         int nrpages;
1535         bool eof = true;
1536         bool mne_swab = false;
1537         int i;
1538         int ealen;
1539         int rc;
1540         ENTRY;
1541
1542         /* allocate buffer for bulk transfer.
1543          * if this is the first time for this mgs to read logs,
1544          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1545          * once; otherwise, it only reads increment of logs, this should be
1546          * small and CONFIG_READ_NRPAGES will be used.
1547          */
1548         nrpages = CONFIG_READ_NRPAGES;
1549         if (cfg->cfg_last_idx == 0) /* the first time */
1550                 nrpages = CONFIG_READ_NRPAGES_INIT;
1551
1552         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1553         if (pages == NULL)
1554                 GOTO(out, rc = -ENOMEM);
1555
1556         for (i = 0; i < nrpages; i++) {
1557                 pages[i] = alloc_page(GFP_IOFS);
1558                 if (pages[i] == NULL)
1559                         GOTO(out, rc = -ENOMEM);
1560         }
1561
1562 again:
1563         LASSERT(cld_is_recover(cld));
1564         LASSERT(mutex_is_locked(&cld->cld_lock));
1565         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1566                                    &RQF_MGS_CONFIG_READ);
1567         if (req == NULL)
1568                 GOTO(out, rc = -ENOMEM);
1569
1570         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1571         if (rc)
1572                 GOTO(out, rc);
1573
1574         /* pack request */
1575         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1576         LASSERT(body != NULL);
1577         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1578         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1579             >= sizeof(body->mcb_name))
1580                 GOTO(out, rc = -E2BIG);
1581         body->mcb_offset = cfg->cfg_last_idx + 1;
1582         body->mcb_type   = cld->cld_type;
1583         body->mcb_bits   = PAGE_CACHE_SHIFT;
1584         body->mcb_units  = nrpages;
1585
1586         /* allocate bulk transfer descriptor */
1587         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1, BULK_PUT_SINK,
1588                                     MGS_BULK_PORTAL);
1589         if (desc == NULL)
1590                 GOTO(out, rc = -ENOMEM);
1591
1592         for (i = 0; i < nrpages; i++)
1593                 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
1594
1595         ptlrpc_request_set_replen(req);
1596         rc = ptlrpc_queue_wait(req);
1597         if (rc)
1598                 GOTO(out, rc);
1599
1600         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1601         if (res->mcr_size < res->mcr_offset)
1602                 GOTO(out, rc = -EINVAL);
1603
1604         /* always update the index even though it might have errors with
1605          * handling the recover logs */
1606         cfg->cfg_last_idx = res->mcr_offset;
1607         eof = res->mcr_offset == res->mcr_size;
1608
1609         CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n",
1610                res->mcr_offset, eof == false);
1611
1612         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1613         if (ealen < 0)
1614                 GOTO(out, rc = ealen);
1615
1616         if (ealen > nrpages << PAGE_CACHE_SHIFT)
1617                 GOTO(out, rc = -EINVAL);
1618
1619         if (ealen == 0) { /* no logs transferred */
1620                 if (!eof)
1621                         rc = -EINVAL;
1622                 GOTO(out, rc);
1623         }
1624
1625         mne_swab = !!ptlrpc_rep_need_swab(req);
1626 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1627         /* This import flag means the server did an extra swab of IR MNE
1628          * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1629         if (unlikely(req->rq_import->imp_need_mne_swab))
1630                 mne_swab = !mne_swab;
1631 #else
1632 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1633 #endif
1634
1635         for (i = 0; i < nrpages && ealen > 0; i++) {
1636                 int rc2;
1637                 void *ptr;
1638
1639                 ptr = kmap(pages[i]);
1640                 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1641                                              min_t(int, ealen, PAGE_CACHE_SIZE),
1642                                              mne_swab);
1643                 kunmap(pages[i]);
1644                 if (rc2 < 0) {
1645                         CWARN("Process recover log %s error %d\n",
1646                               cld->cld_logname, rc2);
1647                         break;
1648                 }
1649
1650                 ealen -= PAGE_CACHE_SIZE;
1651         }
1652
1653 out:
1654         if (req)
1655                 ptlrpc_req_finished(req);
1656
1657         if (rc == 0 && !eof)
1658                 goto again;
1659
1660         if (pages) {
1661                 for (i = 0; i < nrpages; i++) {
1662                         if (pages[i] == NULL)
1663                                 break;
1664                         __free_page(pages[i]);
1665                 }
1666                 OBD_FREE(pages, sizeof(*pages) * nrpages);
1667         }
1668         return rc;
1669 }
1670
1671 /* Copy a remote log locally */
1672 static int mgc_llog_local_copy(const struct lu_env *env,
1673                                struct obd_device *obd,
1674                                struct llog_ctxt *rctxt,
1675                                struct llog_ctxt *lctxt, char *logname)
1676 {
1677         char    *temp_log;
1678         int      rc;
1679
1680         ENTRY;
1681
1682         /*
1683          * - copy it to backup using llog_backup()
1684          * - copy remote llog to logname using llog_backup()
1685          * - if failed then move bakup to logname again
1686          */
1687
1688         OBD_ALLOC(temp_log, strlen(logname) + 1);
1689         if (!temp_log)
1690                 RETURN(-ENOMEM);
1691         sprintf(temp_log, "%sT", logname);
1692
1693         /* make a copy of local llog at first */
1694         rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1695         if (rc < 0 && rc != -ENOENT)
1696                 GOTO(out, rc);
1697         /* copy remote llog to the local copy */
1698         rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1699         if (rc == -ENOENT) {
1700                 /* no remote llog, delete local one too */
1701                 llog_erase(env, lctxt, NULL, logname);
1702         } else if (rc < 0) {
1703                 /* error during backup, get local one back from the copy */
1704                 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1705 out:
1706                 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1707                        obd->obd_name, logname, rc);
1708         }
1709         llog_erase(env, lctxt, NULL, temp_log);
1710         OBD_FREE(temp_log, strlen(logname) + 1);
1711         return rc;
1712 }
1713
1714 /* local_only means it cannot get remote llogs */
1715 static int mgc_process_cfg_log(struct obd_device *mgc,
1716                                struct config_llog_data *cld, int local_only)
1717 {
1718         struct llog_ctxt        *ctxt, *lctxt = NULL;
1719         struct client_obd       *cli = &mgc->u.cli;
1720         struct lustre_sb_info   *lsi = NULL;
1721         int                      rc = 0;
1722         bool                     sptlrpc_started = false;
1723         struct lu_env           *env;
1724
1725         ENTRY;
1726
1727         LASSERT(cld);
1728         LASSERT(mutex_is_locked(&cld->cld_lock));
1729
1730         /*
1731          * local copy of sptlrpc log is controlled elsewhere, don't try to
1732          * read it up here.
1733          */
1734         if (cld_is_sptlrpc(cld) && local_only)
1735                 RETURN(0);
1736
1737         if (cld->cld_cfg.cfg_sb)
1738                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1739
1740         OBD_ALLOC_PTR(env);
1741         if (env == NULL)
1742                 RETURN(-ENOMEM);
1743
1744         rc = lu_env_init(env, LCT_MG_THREAD);
1745         if (rc)
1746                 GOTO(out_free, rc);
1747
1748         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1749         LASSERT(ctxt);
1750
1751         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1752
1753         /* Copy the setup log locally if we can. Don't mess around if we're
1754          * running an MGS though (logs are already local). */
1755         if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1756             cli->cl_mgc_configs_dir != NULL &&
1757             lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1758             lsi->lsi_dt_dev) {
1759                 if (!local_only)
1760                         /* Only try to copy log if we have the lock. */
1761                         rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1762                                                  cld->cld_logname);
1763                 if (local_only || rc) {
1764                         if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1765                             llog_is_empty(env, lctxt, cld->cld_logname)) {
1766                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1767                                                    "log %s and no local copy."
1768                                                    "\n", cld->cld_logname);
1769                                 GOTO(out_pop, rc = -ENOENT);
1770                         }
1771                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1772                                "copy for now, will try to update later.\n",
1773                                cld->cld_logname);
1774                         rc = 0;
1775                 }
1776                 /* Now, whether we copied or not, start using the local llog.
1777                  * If we failed to copy, we'll start using whatever the old
1778                  * log has. */
1779                 llog_ctxt_put(ctxt);
1780                 ctxt = lctxt;
1781                 lctxt = NULL;
1782         } else {
1783                 if (local_only) /* no local log at client side */
1784                         GOTO(out_pop, rc = -EIO);
1785         }
1786
1787         if (cld_is_sptlrpc(cld)) {
1788                 sptlrpc_conf_log_update_begin(cld->cld_logname);
1789                 sptlrpc_started = true;
1790         }
1791
1792         /* logname and instance info should be the same, so use our
1793          * copy of the instance for the update.  The cfg_last_idx will
1794          * be updated here. */
1795         rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1796                                      &cld->cld_cfg);
1797         EXIT;
1798
1799 out_pop:
1800         __llog_ctxt_put(env, ctxt);
1801         if (lctxt)
1802                 __llog_ctxt_put(env, lctxt);
1803
1804         /*
1805          * update settings on existing OBDs. doing it inside
1806          * of llog_process_lock so no device is attaching/detaching
1807          * in parallel.
1808          * the logname must be <fsname>-sptlrpc
1809          */
1810         if (sptlrpc_started) {
1811                 LASSERT(cld_is_sptlrpc(cld));
1812                 sptlrpc_conf_log_update_end(cld->cld_logname);
1813                 class_notify_sptlrpc_conf(cld->cld_logname,
1814                                           strlen(cld->cld_logname) -
1815                                           strlen("-sptlrpc"));
1816         }
1817
1818         lu_env_fini(env);
1819 out_free:
1820         OBD_FREE_PTR(env);
1821         return rc;
1822 }
1823
1824 /** Get a config log from the MGS and process it.
1825  * This func is called for both clients and servers.
1826  * Copy the log locally before parsing it if appropriate (non-MGS server)
1827  */
1828 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1829 {
1830         struct lustre_handle lockh = { 0 };
1831         __u64 flags = LDLM_FL_NO_LRU;
1832         int rc = 0, rcl;
1833         ENTRY;
1834
1835         LASSERT(cld);
1836
1837         /* I don't want multiple processes running process_log at once --
1838            sounds like badness.  It actually might be fine, as long as
1839            we're not trying to update from the same log
1840            simultaneously (in which case we should use a per-log sem.) */
1841         mutex_lock(&cld->cld_lock);
1842         if (cld->cld_stopping) {
1843                 mutex_unlock(&cld->cld_lock);
1844                 RETURN(0);
1845         }
1846
1847         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1848
1849         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1850                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1851
1852         /* Get the cfg lock on the llog */
1853         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1854                           LCK_CR, &flags, NULL, NULL, NULL,
1855                           cld, 0, NULL, &lockh);
1856         if (rcl == 0) {
1857                 /* Get the cld, it will be released in mgc_blocking_ast. */
1858                 config_log_get(cld);
1859                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1860                 LASSERT(rc == 0);
1861         } else {
1862                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1863
1864                 /* mark cld_lostlock so that it will requeue
1865                  * after MGC becomes available. */
1866                 cld->cld_lostlock = 1;
1867                 /* Get extra reference, it will be put in requeue thread */
1868                 config_log_get(cld);
1869         }
1870
1871
1872         if (cld_is_recover(cld)) {
1873                 rc = 0; /* this is not a fatal error for recover log */
1874                 if (rcl == 0)
1875                         rc = mgc_process_recover_log(mgc, cld);
1876         } else {
1877                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1878         }
1879
1880         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1881                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1882
1883         mutex_unlock(&cld->cld_lock);
1884
1885         /* Now drop the lock so MGS can revoke it */
1886         if (!rcl) {
1887                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1888                                  LCK_CR, &lockh);
1889                 if (rcl)
1890                         CERROR("Can't drop cfg lock: %d\n", rcl);
1891         }
1892
1893         RETURN(rc);
1894 }
1895
1896
1897 /** Called from lustre_process_log.
1898  * LCFG_LOG_START gets the config log from the MGS, processes it to start
1899  * any services, and adds it to the list logs to watch (follow).
1900  */
1901 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1902 {
1903         struct lustre_cfg *lcfg = buf;
1904         struct config_llog_instance *cfg = NULL;
1905         char *logname;
1906         int rc = 0;
1907         ENTRY;
1908
1909         switch(lcfg->lcfg_command) {
1910         case LCFG_LOV_ADD_OBD: {
1911                 /* Overloading this cfg command: register a new target */
1912                 struct mgs_target_info *mti;
1913
1914                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1915                     sizeof(struct mgs_target_info))
1916                         GOTO(out, rc = -EINVAL);
1917
1918                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1919                 CDEBUG(D_MGC, "add_target %s %#x\n",
1920                        mti->mti_svname, mti->mti_flags);
1921                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1922                 break;
1923         }
1924         case LCFG_LOV_DEL_OBD:
1925                 /* Unregister has no meaning at the moment. */
1926                 CERROR("lov_del_obd unimplemented\n");
1927                 rc = -ENOSYS;
1928                 break;
1929         case LCFG_SPTLRPC_CONF: {
1930                 rc = sptlrpc_process_config(lcfg);
1931                 break;
1932         }
1933         case LCFG_LOG_START: {
1934                 struct config_llog_data *cld;
1935                 struct super_block *sb;
1936
1937                 logname = lustre_cfg_string(lcfg, 1);
1938                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1939                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1940
1941                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1942                        cfg->cfg_last_idx);
1943
1944                 /* We're only called through here on the initial mount */
1945                 rc = config_log_add(obd, logname, cfg, sb);
1946                 if (rc)
1947                         break;
1948                 cld = config_log_find(logname, cfg);
1949                 if (cld == NULL) {
1950                         rc = -ENOENT;
1951                         break;
1952                 }
1953
1954                 /* COMPAT_146 */
1955                 /* FIXME only set this for old logs!  Right now this forces
1956                    us to always skip the "inside markers" check */
1957                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1958
1959                 rc = mgc_process_log(obd, cld);
1960                 if (rc == 0 && cld->cld_recover != NULL) {
1961                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1962                                          imp_connect_data, IMP_RECOV)) {
1963                                 rc = mgc_process_log(obd, cld->cld_recover);
1964                         } else {
1965                                 struct config_llog_data *cir = cld->cld_recover;
1966                                 cld->cld_recover = NULL;
1967                                 config_log_put(cir);
1968                         }
1969                         if (rc)
1970                                 CERROR("Cannot process recover llog %d\n", rc);
1971                 }
1972
1973                 if (rc == 0 && cld->cld_params != NULL) {
1974                         rc = mgc_process_log(obd, cld->cld_params);
1975                         if (rc == -ENOENT) {
1976                                 CDEBUG(D_MGC, "There is no params "
1977                                               "config file yet\n");
1978                                 rc = 0;
1979                         }
1980                         /* params log is optional */
1981                         if (rc)
1982                                 CERROR("%s: can't process params llog: rc = %d\n",
1983                                        obd->obd_name, rc);
1984                 }
1985                 config_log_put(cld);
1986
1987                 break;
1988         }
1989         case LCFG_LOG_END: {
1990                 logname = lustre_cfg_string(lcfg, 1);
1991
1992                 if (lcfg->lcfg_bufcount >= 2)
1993                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
1994                                 lcfg, 2);
1995                 rc = config_log_end(logname, cfg);
1996                 break;
1997         }
1998         default: {
1999                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2000                 GOTO(out, rc = -EINVAL);
2001
2002         }
2003         }
2004 out:
2005         RETURN(rc);
2006 }
2007
2008 struct obd_ops mgc_obd_ops = {
2009         .o_owner        = THIS_MODULE,
2010         .o_setup        = mgc_setup,
2011         .o_precleanup   = mgc_precleanup,
2012         .o_cleanup      = mgc_cleanup,
2013         .o_add_conn     = client_import_add_conn,
2014         .o_del_conn     = client_import_del_conn,
2015         .o_connect      = client_connect_import,
2016         .o_disconnect   = client_disconnect_export,
2017         //.o_enqueue      = mgc_enqueue,
2018         .o_cancel       = mgc_cancel,
2019         //.o_iocontrol    = mgc_iocontrol,
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, NULL,
2029                                    LUSTRE_MGC_NAME, NULL);
2030 }
2031
2032 #ifdef __KERNEL__
2033 static void /*__exit*/ mgc_exit(void)
2034 {
2035         class_unregister_type(LUSTRE_MGC_NAME);
2036 }
2037
2038 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2039 MODULE_DESCRIPTION("Lustre Management Client");
2040 MODULE_LICENSE("GPL");
2041
2042 module_init(mgc_init);
2043 module_exit(mgc_exit);
2044 #endif