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 #ifdef __KERNEL__
  45 # include <linux/module.h>
  46 # include <linux/pagemap.h>
  47 # include <linux/miscdevice.h>
  48 # include <linux/init.h>
  49 #else
  50 # include <liblustre.h>
  51 #endif
  52
  53 #include <obd_class.h>
  54 #include <lustre_dlm.h>
  55 #include <lprocfs_status.h>
  56 #include <lustre_log.h>
  57 #include <lustre_fsfilt.h>
  58 #include <lustre_disk.h>
  59 #include "mgc_internal.h"
  60
  61 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
  62                           int type)
  63 {
  64         __u64 resname = 0;
  65
  66         if (len > 8) {
  67                 CERROR("name too long: %s\n", name);
  68                 return -EINVAL;
  69         }
  70         if (len <= 0) {
  71                 CERROR("missing name: %s\n", name);
  72                 return -EINVAL;
  73         }
  74         memcpy(&resname, name, len);
  75
  76         /* Always use the same endianness for the resid */
  77         memset(res_id, 0, sizeof(*res_id));
  78         res_id->name[0] = cpu_to_le64(resname);
  79         /* XXX: unfortunately, sptlprc and config llog share one lock */
  80         switch(type) {
  81         case CONFIG_T_CONFIG:
  82         case CONFIG_T_SPTLRPC:
  83                 resname = 0;
  84                 break;
  85         case CONFIG_T_RECOVER:
  86                 resname = type;
  87                 break;
  88         default:
  89                 LBUG();
  90         }
  91         res_id->name[1] = cpu_to_le64(resname);
  92         CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
  93                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  94         return 0;
  95 }
  96
  97 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
  98 {
  99         /* fsname is at most 8 chars long, maybe contain "-".
 100          * e.g. "lustre", "SUN-000" */
 101         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
 102 }
 103 EXPORT_SYMBOL(mgc_fsname2resid);
 104
 105 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
 106 {
 107         char *name_end;
 108         int len;
 109
 110         /* logname consists of "fsname-nodetype".
 111          * e.g. "lustre-MDT0001", "SUN-000-client" */
 112         name_end = strrchr(logname, '-');
 113         LASSERT(name_end);
 114         len = name_end - logname;
 115         return mgc_name2resid(logname, len, res_id, type);
 116 }
 117
 118 /********************** config llog list **********************/
 119 static CFS_LIST_HEAD(config_llog_list);
 120 static DEFINE_SPINLOCK(config_list_lock);
 121
 122 /* Take a reference to a config log */
 123 static int config_log_get(struct config_llog_data *cld)
 124 {
 125         ENTRY;
 126         cfs_atomic_inc(&cld->cld_refcount);
 127         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 128                cfs_atomic_read(&cld->cld_refcount));
 129         RETURN(0);
 130 }
 131
 132 /* Drop a reference to a config log.  When no longer referenced,
 133    we can free the config log data */
 134 static void config_log_put(struct config_llog_data *cld)
 135 {
 136         ENTRY;
 137
 138         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 139                cfs_atomic_read(&cld->cld_refcount));
 140         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 141
 142         /* spinlock to make sure no item with 0 refcount in the list */
 143         if (cfs_atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 144                 cfs_list_del(&cld->cld_list_chain);
 145                 spin_unlock(&config_list_lock);
 146
 147                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 148
 149                 if (cld->cld_recover)
 150                         config_log_put(cld->cld_recover);
 151                 if (cld->cld_sptlrpc)
 152                         config_log_put(cld->cld_sptlrpc);
 153                 if (cld_is_sptlrpc(cld))
 154                         sptlrpc_conf_log_stop(cld->cld_logname);
 155
 156                 class_export_put(cld->cld_mgcexp);
 157                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 158         }
 159
 160         EXIT;
 161 }
 162
 163 /* Find a config log by name */
 164 static
 165 struct config_llog_data *config_log_find(char *logname,
 166                                          struct config_llog_instance *cfg)
 167 {
 168         struct config_llog_data *cld;
 169         struct config_llog_data *found = NULL;
 170         void *                   instance;
 171         ENTRY;
 172
 173         LASSERT(logname != NULL);
 174
 175         instance = cfg ? cfg->cfg_instance : NULL;
 176         spin_lock(&config_list_lock);
 177         cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 178                 /* check if instance equals */
 179                 if (instance != cld->cld_cfg.cfg_instance)
 180                         continue;
 181
 182                 /* instance may be NULL, should check name */
 183                 if (strcmp(logname, cld->cld_logname) == 0) {
 184                         found = cld;
 185                         break;
 186                 }
 187         }
 188         if (found) {
 189                 cfs_atomic_inc(&found->cld_refcount);
 190                 LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0);
 191         }
 192         spin_unlock(&config_list_lock);
 193         RETURN(found);
 194 }
 195
 196 static
 197 struct config_llog_data *do_config_log_add(struct obd_device *obd,
 198                                            char *logname,
 199                                            int type,
 200                                            struct config_llog_instance *cfg,
 201                                            struct super_block *sb)
 202 {
 203         struct config_llog_data *cld;
 204         int                      rc;
 205         ENTRY;
 206
 207         CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
 208                cfg ? cfg->cfg_instance : 0);
 209
 210         OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
 211         if (!cld)
 212                 RETURN(ERR_PTR(-ENOMEM));
 213
 214         strcpy(cld->cld_logname, logname);
 215         if (cfg)
 216                 cld->cld_cfg = *cfg;
 217         else
 218                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 219         mutex_init(&cld->cld_lock);
 220         cld->cld_cfg.cfg_last_idx = 0;
 221         cld->cld_cfg.cfg_flags = 0;
 222         cld->cld_cfg.cfg_sb = sb;
 223         cld->cld_type = type;
 224         cfs_atomic_set(&cld->cld_refcount, 1);
 225
 226         /* Keep the mgc around until we are done */
 227         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 228
 229         if (cld_is_sptlrpc(cld)) {
 230                 sptlrpc_conf_log_start(logname);
 231                 cld->cld_cfg.cfg_obdname = obd->obd_name;
 232         }
 233
 234         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 235
 236         spin_lock(&config_list_lock);
 237         cfs_list_add(&cld->cld_list_chain, &config_llog_list);
 238         spin_unlock(&config_list_lock);
 239
 240         if (rc) {
 241                 config_log_put(cld);
 242                 RETURN(ERR_PTR(rc));
 243         }
 244
 245         if (cld_is_sptlrpc(cld)) {
 246                 rc = mgc_process_log(obd, cld);
 247                 if (rc && rc != -ENOENT)
 248                         CERROR("failed processing sptlrpc log: %d\n", rc);
 249         }
 250
 251         RETURN(cld);
 252 }
 253
 254 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
 255         char *fsname,
 256         struct config_llog_instance *cfg,
 257         struct super_block *sb)
 258 {
 259         struct config_llog_instance lcfg = *cfg;
 260         struct lustre_sb_info *lsi = s2lsi(sb);
 261         struct config_llog_data *cld;
 262         char logname[32];
 263
 264         if (IS_OST(lsi))
 265                 return NULL;
 266
 267         /* for osp-on-ost, see lustre_start_osp() */
 268         if (IS_MDT(lsi) && lcfg.cfg_instance)
 269                 return NULL;
 270
 271         /* we have to use different llog for clients and mdts for cmd
 272          * where only clients are notified if one of cmd server restarts */
 273         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 274         strcpy(logname, fsname);
 275         if (IS_SERVER(lsi)) { /* mdt */
 276                 LASSERT(lcfg.cfg_instance == NULL);
 277                 lcfg.cfg_instance = sb;
 278                 strcat(logname, "-mdtir");
 279         } else {
 280                 LASSERT(lcfg.cfg_instance != NULL);
 281                 strcat(logname, "-cliir");
 282         }
 283
 284         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
 285         return cld;
 286 }
 287
 288
 289 /** Add this log to the list of active logs watched by an MGC.
 290  * Active means we're watching for updates.
 291  * We have one active log per "mount" - client instance or servername.
 292  * Each instance may be at a different point in the log.
 293  */
 294 static int config_log_add(struct obd_device *obd, char *logname,
 295                           struct config_llog_instance *cfg,
 296                           struct super_block *sb)
 297 {
 298         struct lustre_sb_info *lsi = s2lsi(sb);
 299         struct config_llog_data *cld;
 300         struct config_llog_data *sptlrpc_cld;
 301         char                     seclogname[32];
 302         char                    *ptr;
 303         ENTRY;
 304
 305         CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
 306
 307         /*
 308          * for each regular log, the depended sptlrpc log name is
 309          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 310          */
 311         ptr = strrchr(logname, '-');
 312         if (ptr == NULL || ptr - logname > 8) {
 313                 CERROR("logname %s is too long\n", logname);
 314                 RETURN(-EINVAL);
 315         }
 316
 317         memcpy(seclogname, logname, ptr - logname);
 318         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 319
 320         sptlrpc_cld = config_log_find(seclogname, NULL);
 321         if (sptlrpc_cld == NULL) {
 322                 sptlrpc_cld = do_config_log_add(obd, seclogname,
 323                                                 CONFIG_T_SPTLRPC, NULL, NULL);
 324                 if (IS_ERR(sptlrpc_cld)) {
 325                         CERROR("can't create sptlrpc log: %s\n", seclogname);
 326                         RETURN(PTR_ERR(sptlrpc_cld));
 327                 }
 328         }
 329
 330         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 331         if (IS_ERR(cld)) {
 332                 CERROR("can't create log: %s\n", logname);
 333                 config_log_put(sptlrpc_cld);
 334                 RETURN(PTR_ERR(cld));
 335         }
 336
 337         cld->cld_sptlrpc = sptlrpc_cld;
 338
 339         LASSERT(lsi->lsi_lmd);
 340         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
 341                 struct config_llog_data *recover_cld;
 342                 *strrchr(seclogname, '-') = 0;
 343                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 344                 if (IS_ERR(recover_cld)) {
 345                         config_log_put(cld);
 346                         RETURN(PTR_ERR(recover_cld));
 347                 }
 348                 cld->cld_recover = recover_cld;
 349         }
 350
 351         RETURN(0);
 352 }
 353
 354 DEFINE_MUTEX(llog_process_lock);
 355
 356 /** Stop watching for updates on this log.
 357  */
 358 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 359 {
 360         struct config_llog_data *cld;
 361         struct config_llog_data *cld_sptlrpc = NULL;
 362         struct config_llog_data *cld_recover = NULL;
 363         int rc = 0;
 364         ENTRY;
 365
 366         cld = config_log_find(logname, cfg);
 367         if (cld == NULL)
 368                 RETURN(-ENOENT);
 369
 370         mutex_lock(&cld->cld_lock);
 371         /*
 372          * if cld_stopping is set, it means we didn't start the log thus
 373          * not owning the start ref. this can happen after previous umount:
 374          * the cld still hanging there waiting for lock cancel, and we
 375          * remount again but failed in the middle and call log_end without
 376          * calling start_log.
 377          */
 378         if (unlikely(cld->cld_stopping)) {
 379                 mutex_unlock(&cld->cld_lock);
 380                 /* drop the ref from the find */
 381                 config_log_put(cld);
 382                 RETURN(rc);
 383         }
 384
 385         cld->cld_stopping = 1;
 386
 387         cld_recover = cld->cld_recover;
 388         cld->cld_recover = NULL;
 389         mutex_unlock(&cld->cld_lock);
 390
 391         if (cld_recover) {
 392                 mutex_lock(&cld_recover->cld_lock);
 393                 cld_recover->cld_stopping = 1;
 394                 mutex_unlock(&cld_recover->cld_lock);
 395                 config_log_put(cld_recover);
 396         }
 397
 398         spin_lock(&config_list_lock);
 399         cld_sptlrpc = cld->cld_sptlrpc;
 400         cld->cld_sptlrpc = NULL;
 401         spin_unlock(&config_list_lock);
 402
 403         if (cld_sptlrpc)
 404                 config_log_put(cld_sptlrpc);
 405
 406         /* drop the ref from the find */
 407         config_log_put(cld);
 408         /* drop the start ref */
 409         config_log_put(cld);
 410
 411         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 412                rc);
 413         RETURN(rc);
 414 }
 415
 416 int lprocfs_mgc_rd_ir_state(char *page, char **start, off_t off,
 417                             int count, int *eof, void *data)
 418 {
 419         struct obd_device       *obd = data;
 420         struct obd_import       *imp = obd->u.cli.cl_import;
 421         struct obd_connect_data *ocd = &imp->imp_connect_data;
 422         struct config_llog_data *cld;
 423         int rc = 0;
 424         ENTRY;
 425
 426         rc = snprintf(page, count, "imperative_recovery: %s\n",
 427                       OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 428         rc += snprintf(page + rc, count - rc, "client_state:\n");
 429
 430         spin_lock(&config_list_lock);
 431         cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 432                 if (cld->cld_recover == NULL)
 433                         continue;
 434                 rc += snprintf(page + rc, count - rc,
 435                                "    - { client: %s, nidtbl_version: %u }\n",
 436                                cld->cld_logname,
 437                                cld->cld_recover->cld_cfg.cfg_last_idx);
 438         }
 439         spin_unlock(&config_list_lock);
 440
 441         RETURN(rc);
 442 }
 443
 444 /* reenqueue any lost locks */
 445 #define RQ_RUNNING 0x1
 446 #define RQ_NOW     0x2
 447 #define RQ_LATER   0x4
 448 #define RQ_STOP    0x8
 449 static int                    rq_state = 0;
 450 static cfs_waitq_t            rq_waitq;
 451 static DECLARE_COMPLETION(rq_exit);
 452
 453 static void do_requeue(struct config_llog_data *cld)
 454 {
 455         ENTRY;
 456         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 457
 458         /* Do not run mgc_process_log on a disconnected export or an
 459            export which is being disconnected. Take the client
 460            semaphore to make the check non-racy. */
 461         down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 462         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 463                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 464                 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 465         } else {
 466                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 467                        cld->cld_logname);
 468         }
 469         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 470
 471         EXIT;
 472 }
 473
 474 /* this timeout represents how many seconds MGC should wait before
 475  * requeue config and recover lock to the MGS. We need to randomize this
 476  * in order to not flood the MGS.
 477  */
 478 #define MGC_TIMEOUT_MIN_SECONDS   5
 479 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
 480
 481 static int mgc_requeue_thread(void *data)
 482 {
 483         char name[] = "ll_cfg_requeue";
 484         int rc = 0;
 485         ENTRY;
 486
 487         cfs_daemonize(name);
 488
 489         CDEBUG(D_MGC, "Starting requeue thread\n");
 490
 491         /* Keep trying failed locks periodically */
 492         spin_lock(&config_list_lock);
 493         rq_state |= RQ_RUNNING;
 494         while (1) {
 495                 struct l_wait_info lwi;
 496                 struct config_llog_data *cld, *cld_prev;
 497                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
 498                 int stopped = !!(rq_state & RQ_STOP);
 499                 int to;
 500
 501                 /* Any new or requeued lostlocks will change the state */
 502                 rq_state &= ~(RQ_NOW | RQ_LATER);
 503                 spin_unlock(&config_list_lock);
 504
 505                 /* Always wait a few seconds to allow the server who
 506                    caused the lock revocation to finish its setup, plus some
 507                    random so everyone doesn't try to reconnect at once. */
 508                 to = MGC_TIMEOUT_MIN_SECONDS * CFS_HZ;
 509                 to += rand * CFS_HZ / 100; /* rand is centi-seconds */
 510                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 511                 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi);
 512
 513                 /*
 514                  * iterate & processing through the list. for each cld, process
 515                  * its depending sptlrpc cld firstly (if any) and then itself.
 516                  *
 517                  * it's guaranteed any item in the list must have
 518                  * reference > 0; and if cld_lostlock is set, at
 519                  * least one reference is taken by the previous enqueue.
 520                  */
 521                 cld_prev = NULL;
 522
 523                 spin_lock(&config_list_lock);
 524                 cfs_list_for_each_entry(cld, &config_llog_list,
 525                                         cld_list_chain) {
 526                         if (!cld->cld_lostlock)
 527                                 continue;
 528
 529                         spin_unlock(&config_list_lock);
 530
 531                         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 532
 533                         /* Whether we enqueued again or not in mgc_process_log,
 534                          * we're done with the ref from the old enqueue */
 535                         if (cld_prev)
 536                                 config_log_put(cld_prev);
 537                         cld_prev = cld;
 538
 539                         cld->cld_lostlock = 0;
 540                         if (likely(!stopped))
 541                                 do_requeue(cld);
 542
 543                         spin_lock(&config_list_lock);
 544                 }
 545                 spin_unlock(&config_list_lock);
 546                 if (cld_prev)
 547                         config_log_put(cld_prev);
 548
 549                 /* break after scanning the list so that we can drop
 550                  * refcount to losing lock clds */
 551                 if (unlikely(stopped)) {
 552                         spin_lock(&config_list_lock);
 553                         break;
 554                 }
 555
 556                 /* Wait a bit to see if anyone else needs a requeue */
 557                 lwi = (struct l_wait_info) { 0 };
 558                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 559                              &lwi);
 560                 spin_lock(&config_list_lock);
 561         }
 562         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 563         rq_state &= ~RQ_RUNNING;
 564         spin_unlock(&config_list_lock);
 565
 566         complete(&rq_exit);
 567
 568         CDEBUG(D_MGC, "Ending requeue thread\n");
 569         RETURN(rc);
 570 }
 571
 572 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 573    We are responsible for dropping the config log reference from here on out. */
 574 static void mgc_requeue_add(struct config_llog_data *cld)
 575 {
 576         ENTRY;
 577
 578         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 579                cld->cld_logname, cfs_atomic_read(&cld->cld_refcount),
 580                cld->cld_stopping, rq_state);
 581         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 582
 583         mutex_lock(&cld->cld_lock);
 584         if (cld->cld_stopping || cld->cld_lostlock) {
 585                 mutex_unlock(&cld->cld_lock);
 586                 RETURN_EXIT;
 587         }
 588         /* this refcount will be released in mgc_requeue_thread. */
 589         config_log_get(cld);
 590         cld->cld_lostlock = 1;
 591         mutex_unlock(&cld->cld_lock);
 592
 593         /* Hold lock for rq_state */
 594         spin_lock(&config_list_lock);
 595         if (rq_state & RQ_STOP) {
 596                 spin_unlock(&config_list_lock);
 597                 cld->cld_lostlock = 0;
 598                 config_log_put(cld);
 599         } else {
 600                 rq_state |= RQ_NOW;
 601                 spin_unlock(&config_list_lock);
 602                 cfs_waitq_signal(&rq_waitq);
 603         }
 604         EXIT;
 605 }
 606
 607 /********************** class fns **********************/
 608
 609 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
 610                         struct vfsmount *mnt)
 611 {
 612         struct lvfs_run_ctxt saved;
 613         struct lustre_sb_info *lsi = s2lsi(sb);
 614         struct client_obd *cli = &obd->u.cli;
 615         struct dentry *dentry;
 616         char *label;
 617         int err = 0;
 618         ENTRY;
 619
 620         LASSERT(lsi);
 621         LASSERT(lsi->lsi_srv_mnt == mnt);
 622
 623         /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
 624         down(&cli->cl_mgc_sem);
 625
 626         cfs_cleanup_group_info();
 627
 628         obd->obd_fsops = fsfilt_get_ops(lsi->lsi_fstype);
 629         if (IS_ERR(obd->obd_fsops)) {
 630                 up(&cli->cl_mgc_sem);
 631                 CERROR("%s: No fstype %s: rc = %ld\n", lsi->lsi_fstype,
 632                        obd->obd_name, PTR_ERR(obd->obd_fsops));
 633                 RETURN(PTR_ERR(obd->obd_fsops));
 634         }
 635
 636         cli->cl_mgc_vfsmnt = mnt;
 637         err = fsfilt_setup(obd, mnt->mnt_sb);
 638         if (err)
 639                 GOTO(err_ops, err);
 640
 641         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 642         obd->obd_lvfs_ctxt.pwdmnt = mnt;
 643         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
 644         obd->obd_lvfs_ctxt.fs = get_ds();
 645
 646         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 647         dentry = ll_lookup_one_len(MOUNT_CONFIGS_DIR, cfs_fs_pwd(current->fs),
 648                                    strlen(MOUNT_CONFIGS_DIR));
 649         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 650         if (IS_ERR(dentry)) {
 651                 err = PTR_ERR(dentry);
 652                 CERROR("cannot lookup %s directory: rc = %d\n",
 653                        MOUNT_CONFIGS_DIR, err);
 654                 GOTO(err_ops, err);
 655         }
 656         cli->cl_mgc_configs_dir = dentry;
 657
 658         /* We take an obd ref to insure that we can't get to mgc_cleanup
 659            without calling mgc_fs_cleanup first. */
 660         class_incref(obd, "mgc_fs", obd);
 661
 662         label = fsfilt_get_label(obd, mnt->mnt_sb);
 663         if (label)
 664                 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
 665
 666         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 667         RETURN(0);
 668
 669 err_ops:
 670         fsfilt_put_ops(obd->obd_fsops);
 671         obd->obd_fsops = NULL;
 672         cli->cl_mgc_vfsmnt = NULL;
 673         up(&cli->cl_mgc_sem);
 674         RETURN(err);
 675 }
 676
 677 static int mgc_fs_cleanup(struct obd_device *obd)
 678 {
 679         struct client_obd *cli = &obd->u.cli;
 680         int rc = 0;
 681         ENTRY;
 682
 683         LASSERT(cli->cl_mgc_vfsmnt != NULL);
 684
 685         if (cli->cl_mgc_configs_dir != NULL) {
 686                 struct lvfs_run_ctxt saved;
 687                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 688                 l_dput(cli->cl_mgc_configs_dir);
 689                 cli->cl_mgc_configs_dir = NULL;
 690                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 691                 class_decref(obd, "mgc_fs", obd);
 692         }
 693
 694         cli->cl_mgc_vfsmnt = NULL;
 695         if (obd->obd_fsops)
 696                 fsfilt_put_ops(obd->obd_fsops);
 697
 698         up(&cli->cl_mgc_sem);
 699
 700         RETURN(rc);
 701 }
 702
 703 static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0);
 704 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 705 {
 706         int rc = 0;
 707         ENTRY;
 708
 709         switch (stage) {
 710         case OBD_CLEANUP_EARLY:
 711                 break;
 712         case OBD_CLEANUP_EXPORTS:
 713                 if (cfs_atomic_dec_and_test(&mgc_count)) {
 714                         int running;
 715                         /* stop requeue thread */
 716                         spin_lock(&config_list_lock);
 717                         running = rq_state & RQ_RUNNING;
 718                         if (running)
 719                                 rq_state |= RQ_STOP;
 720                         spin_unlock(&config_list_lock);
 721                         if (running) {
 722                                 cfs_waitq_signal(&rq_waitq);
 723                                 wait_for_completion(&rq_exit);
 724                         }
 725                 }
 726                 obd_cleanup_client_import(obd);
 727                 rc = obd_llog_finish(obd, 0);
 728                 if (rc != 0)
 729                         CERROR("failed to cleanup llogging subsystems\n");
 730                 break;
 731         }
 732         RETURN(rc);
 733 }
 734
 735 static int mgc_cleanup(struct obd_device *obd)
 736 {
 737         struct client_obd *cli = &obd->u.cli;
 738         int rc;
 739         ENTRY;
 740
 741         LASSERT(cli->cl_mgc_vfsmnt == NULL);
 742
 743         /* COMPAT_146 - old config logs may have added profiles we don't
 744            know about */
 745         if (obd->obd_type->typ_refcnt <= 1)
 746                 /* Only for the last mgc */
 747                 class_del_profiles();
 748
 749         lprocfs_obd_cleanup(obd);
 750         ptlrpcd_decref();
 751
 752         rc = client_obd_cleanup(obd);
 753         RETURN(rc);
 754 }
 755
 756 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 757 {
 758         struct lprocfs_static_vars lvars;
 759         int rc;
 760         ENTRY;
 761
 762         ptlrpcd_addref();
 763
 764         rc = client_obd_setup(obd, lcfg);
 765         if (rc)
 766                 GOTO(err_decref, rc);
 767
 768         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
 769         if (rc) {
 770                 CERROR("failed to setup llogging subsystems\n");
 771                 GOTO(err_cleanup, rc);
 772         }
 773
 774         lprocfs_mgc_init_vars(&lvars);
 775         lprocfs_obd_setup(obd, lvars.obd_vars);
 776         sptlrpc_lprocfs_cliobd_attach(obd);
 777
 778         if (cfs_atomic_inc_return(&mgc_count) == 1) {
 779                 rq_state = 0;
 780                 cfs_waitq_init(&rq_waitq);
 781
 782                 /* start requeue thread */
 783                 rc = cfs_create_thread(mgc_requeue_thread, NULL,
 784                                        CFS_DAEMON_FLAGS);
 785                 if (rc < 0) {
 786                         CERROR("%s: Cannot start requeue thread (%d),"
 787                                "no more log updates!\n",
 788                                obd->obd_name, rc);
 789                         GOTO(err_cleanup, rc);
 790                 }
 791                 /* rc is the pid of mgc_requeue_thread. */
 792                 rc = 0;
 793         }
 794
 795         RETURN(rc);
 796
 797 err_cleanup:
 798         client_obd_cleanup(obd);
 799 err_decref:
 800         ptlrpcd_decref();
 801         RETURN(rc);
 802 }
 803
 804 /* based on ll_mdc_blocking_ast */
 805 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 806                             void *data, int flag)
 807 {
 808         struct lustre_handle lockh;
 809         struct config_llog_data *cld = (struct config_llog_data *)data;
 810         int rc = 0;
 811         ENTRY;
 812
 813         switch (flag) {
 814         case LDLM_CB_BLOCKING:
 815                 /* mgs wants the lock, give it up... */
 816                 LDLM_DEBUG(lock, "MGC blocking CB");
 817                 ldlm_lock2handle(lock, &lockh);
 818                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 819                 break;
 820         case LDLM_CB_CANCELING:
 821                 /* We've given up the lock, prepare ourselves to update. */
 822                 LDLM_DEBUG(lock, "MGC cancel CB");
 823
 824                 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
 825                        lock->l_resource->lr_name.name[0],
 826                        (char *)&lock->l_resource->lr_name.name[0]);
 827
 828                 if (!cld) {
 829                         CDEBUG(D_INFO, "missing data, won't requeue\n");
 830                         break;
 831                 }
 832
 833                 /* held at mgc_process_log(). */
 834                 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 835                 /* Are we done with this log? */
 836                 if (cld->cld_stopping) {
 837                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
 838                                cld->cld_logname);
 839                         config_log_put(cld);
 840                         break;
 841                 }
 842                 /* Make sure not to re-enqueue when the mgc is stopping
 843                    (we get called from client_disconnect_export) */
 844                 if (!lock->l_conn_export ||
 845                     !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
 846                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
 847                                cld->cld_logname);
 848                         config_log_put(cld);
 849                         break;
 850                 }
 851
 852                 /* Re-enqueue now */
 853                 mgc_requeue_add(cld);
 854                 config_log_put(cld);
 855                 break;
 856         default:
 857                 LBUG();
 858         }
 859
 860         RETURN(rc);
 861 }
 862
 863 /* Not sure where this should go... */
 864 #define  MGC_ENQUEUE_LIMIT 50
 865 #define  MGC_TARGET_REG_LIMIT 10
 866 #define  MGC_SEND_PARAM_LIMIT 10
 867
 868 /* Send parameter to MGS*/
 869 static int mgc_set_mgs_param(struct obd_export *exp,
 870                              struct mgs_send_param *msp)
 871 {
 872         struct ptlrpc_request *req;
 873         struct mgs_send_param *req_msp, *rep_msp;
 874         int rc;
 875         ENTRY;
 876
 877         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 878                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
 879                                         MGS_SET_INFO);
 880         if (!req)
 881                 RETURN(-ENOMEM);
 882
 883         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 884         if (!req_msp) {
 885                 ptlrpc_req_finished(req);
 886                 RETURN(-ENOMEM);
 887         }
 888
 889         memcpy(req_msp, msp, sizeof(*req_msp));
 890         ptlrpc_request_set_replen(req);
 891
 892         /* Limit how long we will wait for the enqueue to complete */
 893         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
 894         rc = ptlrpc_queue_wait(req);
 895         if (!rc) {
 896                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 897                 memcpy(msp, rep_msp, sizeof(*rep_msp));
 898         }
 899
 900         ptlrpc_req_finished(req);
 901
 902         RETURN(rc);
 903 }
 904
 905 /* Take a config lock so we can get cancel notifications */
 906 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
 907                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
 908                        __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
 909                        void *data, __u32 lvb_len, void *lvb_swabber,
 910                        struct lustre_handle *lockh)
 911 {
 912         struct config_llog_data *cld = (struct config_llog_data *)data;
 913         struct ldlm_enqueue_info einfo = {
 914                 .ei_type        = type,
 915                 .ei_mode        = mode,
 916                 .ei_cb_bl       = mgc_blocking_ast,
 917                 .ei_cb_cp       = ldlm_completion_ast,
 918         };
 919         struct ptlrpc_request *req;
 920         int short_limit = cld_is_sptlrpc(cld);
 921         int rc;
 922         ENTRY;
 923
 924         CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
 925                cld->cld_resid.name[0]);
 926
 927         /* We need a callback for every lockholder, so don't try to
 928            ldlm_lock_match (see rev 1.1.2.11.2.47) */
 929         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 930                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
 931                                         LDLM_ENQUEUE);
 932         if (req == NULL)
 933                 RETURN(-ENOMEM);
 934
 935         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
 936         ptlrpc_request_set_replen(req);
 937
 938         /* check if this is server or client */
 939         if (cld->cld_cfg.cfg_sb) {
 940                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
 941                 if (lsi && IS_SERVER(lsi))
 942                         short_limit = 1;
 943         }
 944         /* Limit how long we will wait for the enqueue to complete */
 945         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
 946         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
 947                               NULL, 0, LVB_T_NONE, lockh, 0);
 948         /* A failed enqueue should still call the mgc_blocking_ast,
 949            where it will be requeued if needed ("grant failed"). */
 950         ptlrpc_req_finished(req);
 951         RETURN(rc);
 952 }
 953
 954 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
 955                       __u32 mode, struct lustre_handle *lockh)
 956 {
 957         ENTRY;
 958
 959         ldlm_lock_decref(lockh, mode);
 960
 961         RETURN(0);
 962 }
 963
 964 static void mgc_notify_active(struct obd_device *unused)
 965 {
 966         /* wakeup mgc_requeue_thread to requeue mgc lock */
 967         spin_lock(&config_list_lock);
 968         rq_state |= RQ_NOW;
 969         spin_unlock(&config_list_lock);
 970         cfs_waitq_signal(&rq_waitq);
 971
 972         /* TODO: Help the MGS rebuild nidtbl. -jay */
 973 }
 974
 975 /* Send target_reg message to MGS */
 976 static int mgc_target_register(struct obd_export *exp,
 977                                struct mgs_target_info *mti)
 978 {
 979         struct ptlrpc_request  *req;
 980         struct mgs_target_info *req_mti, *rep_mti;
 981         int                     rc;
 982         ENTRY;
 983
 984         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 985                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
 986                                         MGS_TARGET_REG);
 987         if (req == NULL)
 988                 RETURN(-ENOMEM);
 989
 990         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
 991         if (!req_mti) {
 992                 ptlrpc_req_finished(req);
 993                 RETURN(-ENOMEM);
 994         }
 995
 996         memcpy(req_mti, mti, sizeof(*req_mti));
 997         ptlrpc_request_set_replen(req);
 998         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
 999         /* Limit how long we will wait for the enqueue to complete */
1000         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1001
1002         rc = ptlrpc_queue_wait(req);
1003         if (!rc) {
1004                 rep_mti = req_capsule_server_get(&req->rq_pill,
1005                                                  &RMF_MGS_TARGET_INFO);
1006                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1007                 CDEBUG(D_MGC, "register %s got index = %d\n",
1008                        mti->mti_svname, mti->mti_stripe_index);
1009         }
1010         ptlrpc_req_finished(req);
1011
1012         RETURN(rc);
1013 }
1014
1015 int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1016                        obd_count keylen, void *key, obd_count vallen,
1017                        void *val, struct ptlrpc_request_set *set)
1018 {
1019         int rc = -EINVAL;
1020         ENTRY;
1021
1022         /* Turn off initial_recov after we try all backup servers once */
1023         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1024                 struct obd_import *imp = class_exp2cliimp(exp);
1025                 int value;
1026                 if (vallen != sizeof(int))
1027                         RETURN(-EINVAL);
1028                 value = *(int *)val;
1029                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1030                        imp->imp_obd->obd_name, value,
1031                        imp->imp_deactive, imp->imp_invalid,
1032                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1033                        ptlrpc_import_state_name(imp->imp_state));
1034                 /* Resurrect if we previously died */
1035                 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1036                      imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1037                         ptlrpc_reconnect_import(imp);
1038                 RETURN(0);
1039         }
1040         /* FIXME move this to mgc_process_config */
1041         if (KEY_IS(KEY_REGISTER_TARGET)) {
1042                 struct mgs_target_info *mti;
1043                 if (vallen != sizeof(struct mgs_target_info))
1044                         RETURN(-EINVAL);
1045                 mti = (struct mgs_target_info *)val;
1046                 CDEBUG(D_MGC, "register_target %s %#x\n",
1047                        mti->mti_svname, mti->mti_flags);
1048                 rc =  mgc_target_register(exp, mti);
1049                 RETURN(rc);
1050         }
1051         if (KEY_IS(KEY_SET_FS)) {
1052                 struct super_block *sb = (struct super_block *)val;
1053                 struct lustre_sb_info *lsi;
1054                 if (vallen != sizeof(struct super_block))
1055                         RETURN(-EINVAL);
1056                 lsi = s2lsi(sb);
1057                 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
1058                 if (rc) {
1059                         CERROR("set_fs got %d\n", rc);
1060                 }
1061                 RETURN(rc);
1062         }
1063         if (KEY_IS(KEY_CLEAR_FS)) {
1064                 if (vallen != 0)
1065                         RETURN(-EINVAL);
1066                 rc = mgc_fs_cleanup(exp->exp_obd);
1067                 if (rc) {
1068                         CERROR("clear_fs got %d\n", rc);
1069                 }
1070                 RETURN(rc);
1071         }
1072         if (KEY_IS(KEY_SET_INFO)) {
1073                 struct mgs_send_param *msp;
1074
1075                 msp = (struct mgs_send_param *)val;
1076                 rc =  mgc_set_mgs_param(exp, msp);
1077                 RETURN(rc);
1078         }
1079         if (KEY_IS(KEY_MGSSEC)) {
1080                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1081                 struct sptlrpc_flavor  flvr;
1082
1083                 /*
1084                  * empty string means using current flavor, if which haven't
1085                  * been set yet, set it as null.
1086                  *
1087                  * if flavor has been set previously, check the asking flavor
1088                  * must match the existing one.
1089                  */
1090                 if (vallen == 0) {
1091                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1092                                 RETURN(0);
1093                         val = "null";
1094                         vallen = 4;
1095                 }
1096
1097                 rc = sptlrpc_parse_flavor(val, &flvr);
1098                 if (rc) {
1099                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1100                                (char *) val);
1101                         RETURN(rc);
1102                 }
1103
1104                 /*
1105                  * caller already hold a mutex
1106                  */
1107                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1108                         cli->cl_flvr_mgc = flvr;
1109                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1110                                   sizeof(flvr)) != 0) {
1111                         char    str[20];
1112
1113                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1114                                             str, sizeof(str));
1115                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1116                                        "currently %s is in use\n",
1117                                        (char *) val, str);
1118                         rc = -EPERM;
1119                 }
1120                 RETURN(rc);
1121         }
1122
1123         RETURN(rc);
1124 }
1125
1126 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1127                         __u32 keylen, void *key, __u32 *vallen, void *val,
1128                         struct lov_stripe_md *unused)
1129 {
1130         int rc = -EINVAL;
1131
1132         if (KEY_IS(KEY_CONN_DATA)) {
1133                 struct obd_import *imp = class_exp2cliimp(exp);
1134                 struct obd_connect_data *data = val;
1135
1136                 if (*vallen == sizeof(*data)) {
1137                         *data = imp->imp_connect_data;
1138                         rc = 0;
1139                 }
1140         }
1141
1142         return rc;
1143 }
1144
1145 static int mgc_import_event(struct obd_device *obd,
1146                             struct obd_import *imp,
1147                             enum obd_import_event event)
1148 {
1149         int rc = 0;
1150
1151         LASSERT(imp->imp_obd == obd);
1152         CDEBUG(D_MGC, "import event %#x\n", event);
1153
1154         switch (event) {
1155         case IMP_EVENT_DISCON:
1156                 /* MGC imports should not wait for recovery */
1157                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1158                         ptlrpc_pinger_ir_down();
1159                 break;
1160         case IMP_EVENT_INACTIVE:
1161                 break;
1162         case IMP_EVENT_INVALIDATE: {
1163                 struct ldlm_namespace *ns = obd->obd_namespace;
1164                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1165                 break;
1166         }
1167         case IMP_EVENT_ACTIVE:
1168                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1169                 /* Clearing obd_no_recov allows us to continue pinging */
1170                 obd->obd_no_recov = 0;
1171                 mgc_notify_active(obd);
1172                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1173                         ptlrpc_pinger_ir_up();
1174                 break;
1175         case IMP_EVENT_OCD:
1176                 break;
1177         case IMP_EVENT_DEACTIVATE:
1178         case IMP_EVENT_ACTIVATE:
1179                 break;
1180         default:
1181                 CERROR("Unknown import event %#x\n", event);
1182                 LBUG();
1183         }
1184         RETURN(rc);
1185 }
1186
1187 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1188                          struct obd_device *tgt, int *index)
1189 {
1190         struct llog_ctxt *ctxt;
1191         int rc;
1192         ENTRY;
1193
1194         LASSERT(olg == &obd->obd_olg);
1195
1196 #ifdef HAVE_LDISKFS_OSD
1197         rc = llog_setup(NULL, obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt,
1198                         &llog_lvfs_ops);
1199         if (rc)
1200                 RETURN(rc);
1201 #endif
1202
1203         rc = llog_setup(NULL, obd, olg, LLOG_CONFIG_REPL_CTXT, tgt,
1204                         &llog_client_ops);
1205         if (rc)
1206                 GOTO(out, rc);
1207
1208         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_REPL_CTXT);
1209         if (!ctxt)
1210                 GOTO(out, rc = -ENODEV);
1211
1212         llog_initiator_connect(ctxt);
1213         llog_ctxt_put(ctxt);
1214
1215         RETURN(0);
1216 out:
1217         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1218         if (ctxt)
1219                 llog_cleanup(NULL, ctxt);
1220         RETURN(rc);
1221 }
1222
1223 static int mgc_llog_finish(struct obd_device *obd, int count)
1224 {
1225         struct llog_ctxt *ctxt;
1226
1227         ENTRY;
1228
1229         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1230         if (ctxt)
1231                 llog_cleanup(NULL, ctxt);
1232
1233         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1234         if (ctxt)
1235                 llog_cleanup(NULL, ctxt);
1236         RETURN(0);
1237 }
1238
1239 enum {
1240         CONFIG_READ_NRPAGES_INIT = 1 << (20 - CFS_PAGE_SHIFT),
1241         CONFIG_READ_NRPAGES      = 4
1242 };
1243
1244 static int mgc_apply_recover_logs(struct obd_device *mgc,
1245                                   struct config_llog_data *cld,
1246                                   __u64 max_version,
1247                                   void *data, int datalen, bool mne_swab)
1248 {
1249         struct config_llog_instance *cfg = &cld->cld_cfg;
1250         struct lustre_sb_info       *lsi = s2lsi(cfg->cfg_sb);
1251         struct mgs_nidtbl_entry *entry;
1252         struct lustre_cfg       *lcfg;
1253         struct lustre_cfg_bufs   bufs;
1254         u64   prev_version = 0;
1255         char *inst;
1256         char *buf;
1257         int   bufsz;
1258         int   pos;
1259         int   rc  = 0;
1260         int   off = 0;
1261         ENTRY;
1262
1263         LASSERT(cfg->cfg_instance != NULL);
1264         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1265
1266         OBD_ALLOC(inst, CFS_PAGE_SIZE);
1267         if (inst == NULL)
1268                 RETURN(-ENOMEM);
1269
1270         if (!IS_SERVER(lsi)) {
1271                 pos = snprintf(inst, CFS_PAGE_SIZE, "%p", cfg->cfg_instance);
1272                 if (pos >= CFS_PAGE_SIZE) {
1273                         OBD_FREE(inst, CFS_PAGE_SIZE);
1274                         return -E2BIG;
1275                 }
1276         } else {
1277                 LASSERT(IS_MDT(lsi));
1278                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1279                                         CFS_PAGE_SIZE);
1280                 if (rc) {
1281                         OBD_FREE(inst, CFS_PAGE_SIZE);
1282                         RETURN(-EINVAL);
1283                 }
1284                 pos = strlen(inst);
1285         }
1286
1287         ++pos;
1288         buf   = inst + pos;
1289         bufsz = CFS_PAGE_SIZE - pos;
1290
1291         while (datalen > 0) {
1292                 int   entry_len = sizeof(*entry);
1293                 int   is_ost;
1294                 struct obd_device *obd;
1295                 char *obdname;
1296                 char *cname;
1297                 char *params;
1298                 char *uuid;
1299
1300                 rc = -EINVAL;
1301                 if (datalen < sizeof(*entry))
1302                         break;
1303
1304                 entry = (typeof(entry))(data + off);
1305
1306                 /* sanity check */
1307                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1308                         break;
1309                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1310                         break;
1311                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1312                         break;
1313
1314                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1315                 if (datalen < entry_len) /* must have entry_len at least */
1316                         break;
1317
1318                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1319                 if (mne_swab)
1320                         lustre_swab_mgs_nidtbl_entry(entry);
1321                 if (entry->mne_length > CFS_PAGE_SIZE) {
1322                         CERROR("MNE too large (%u)\n", entry->mne_length);
1323                         break;
1324                 }
1325
1326                 if (entry->mne_length < entry_len)
1327                         break;
1328
1329                 off     += entry->mne_length;
1330                 datalen -= entry->mne_length;
1331                 if (datalen < 0)
1332                         break;
1333
1334                 if (entry->mne_version > max_version) {
1335                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1336                                entry->mne_version, max_version);
1337                         break;
1338                 }
1339
1340                 if (prev_version >= entry->mne_version) {
1341                         CERROR("index unsorted, prev %lld, now %lld\n",
1342                                prev_version, entry->mne_version);
1343                         break;
1344                 }
1345                 prev_version = entry->mne_version;
1346
1347                 /*
1348                  * Write a string with format "nid::instance" to
1349                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1350                  */
1351
1352                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1353                 memset(buf, 0, bufsz);
1354                 obdname = buf;
1355                 pos = 0;
1356
1357                 /* lustre-OST0001-osc-<instance #> */
1358                 strcpy(obdname, cld->cld_logname);
1359                 cname = strrchr(obdname, '-');
1360                 if (cname == NULL) {
1361                         CERROR("mgc %s: invalid logname %s\n",
1362                                mgc->obd_name, obdname);
1363                         break;
1364                 }
1365
1366                 pos = cname - obdname;
1367                 obdname[pos] = 0;
1368                 pos += sprintf(obdname + pos, "-%s%04x",
1369                                   is_ost ? "OST" : "MDT", entry->mne_index);
1370
1371                 cname = is_ost ? "osc" : "mdc",
1372                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1373                 lustre_cfg_bufs_reset(&bufs, obdname);
1374
1375                 /* find the obd by obdname */
1376                 obd = class_name2obd(obdname);
1377                 if (obd == NULL) {
1378                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1379                                mgc->obd_name, obdname);
1380                         rc = 0;
1381                         /* this is a safe race, when the ost is starting up...*/
1382                         continue;
1383                 }
1384
1385                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1386                 ++pos;
1387                 params = buf + pos;
1388                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1389                 uuid = buf + pos;
1390
1391                 down_read(&obd->u.cli.cl_sem);
1392                 if (obd->u.cli.cl_import == NULL) {
1393                         /* client does not connect to the OST yet */
1394                         up_read(&obd->u.cli.cl_sem);
1395                         rc = 0;
1396                         continue;
1397                 }
1398
1399                 /* TODO: iterate all nids to find one */
1400                 /* find uuid by nid */
1401                 rc = client_import_find_conn(obd->u.cli.cl_import,
1402                                              entry->u.nids[0],
1403                                              (struct obd_uuid *)uuid);
1404                 up_read(&obd->u.cli.cl_sem);
1405                 if (rc < 0) {
1406                         CERROR("mgc: cannot find uuid by nid %s\n",
1407                                libcfs_nid2str(entry->u.nids[0]));
1408                         break;
1409                 }
1410
1411                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1412                        uuid, libcfs_nid2str(entry->u.nids[0]));
1413
1414                 pos += strlen(uuid);
1415                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1416                 LASSERT(pos < bufsz);
1417
1418                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1419
1420                 rc = -ENOMEM;
1421                 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1422                 if (lcfg == NULL) {
1423                         CERROR("mgc: cannot allocate memory\n");
1424                         break;
1425                 }
1426
1427                 CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n",
1428                        prev_version, max_version, obdname, params);
1429
1430                 rc = class_process_config(lcfg);
1431                 lustre_cfg_free(lcfg);
1432                 if (rc)
1433                         CDEBUG(D_INFO, "process config for %s error %d\n",
1434                                obdname, rc);
1435
1436                 /* continue, even one with error */
1437         }
1438
1439         OBD_FREE(inst, CFS_PAGE_SIZE);
1440         RETURN(rc);
1441 }
1442
1443 /**
1444  * This function is called if this client was notified for target restarting
1445  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1446  */
1447 static int mgc_process_recover_log(struct obd_device *obd,
1448                                    struct config_llog_data *cld)
1449 {
1450         struct ptlrpc_request *req = NULL;
1451         struct config_llog_instance *cfg = &cld->cld_cfg;
1452         struct mgs_config_body *body;
1453         struct mgs_config_res  *res;
1454         struct ptlrpc_bulk_desc *desc;
1455         cfs_page_t **pages;
1456         int nrpages;
1457         bool eof = true;
1458         bool mne_swab = false;
1459         int i;
1460         int ealen;
1461         int rc;
1462         ENTRY;
1463
1464         /* allocate buffer for bulk transfer.
1465          * if this is the first time for this mgs to read logs,
1466          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1467          * once; otherwise, it only reads increment of logs, this should be
1468          * small and CONFIG_READ_NRPAGES will be used.
1469          */
1470         nrpages = CONFIG_READ_NRPAGES;
1471         if (cfg->cfg_last_idx == 0) /* the first time */
1472                 nrpages = CONFIG_READ_NRPAGES_INIT;
1473
1474         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1475         if (pages == NULL)
1476                 GOTO(out, rc = -ENOMEM);
1477
1478         for (i = 0; i < nrpages; i++) {
1479                 pages[i] = cfs_alloc_page(CFS_ALLOC_STD);
1480                 if (pages[i] == NULL)
1481                         GOTO(out, rc = -ENOMEM);
1482         }
1483
1484 again:
1485         LASSERT(cld_is_recover(cld));
1486         LASSERT(mutex_is_locked(&cld->cld_lock));
1487         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1488                                    &RQF_MGS_CONFIG_READ);
1489         if (req == NULL)
1490                 GOTO(out, rc = -ENOMEM);
1491
1492         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1493         if (rc)
1494                 GOTO(out, rc);
1495
1496         /* pack request */
1497         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1498         LASSERT(body != NULL);
1499         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1500         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1501             >= sizeof(body->mcb_name))
1502                 GOTO(out, rc = -E2BIG);
1503         body->mcb_offset = cfg->cfg_last_idx + 1;
1504         body->mcb_type   = cld->cld_type;
1505         body->mcb_bits   = CFS_PAGE_SHIFT;
1506         body->mcb_units  = nrpages;
1507
1508         /* allocate bulk transfer descriptor */
1509         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1, BULK_PUT_SINK,
1510                                     MGS_BULK_PORTAL);
1511         if (desc == NULL)
1512                 GOTO(out, rc = -ENOMEM);
1513
1514         for (i = 0; i < nrpages; i++)
1515                 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, CFS_PAGE_SIZE);
1516
1517         ptlrpc_request_set_replen(req);
1518         rc = ptlrpc_queue_wait(req);
1519         if (rc)
1520                 GOTO(out, rc);
1521
1522         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1523         if (res->mcr_size < res->mcr_offset)
1524                 GOTO(out, rc = -EINVAL);
1525
1526         /* always update the index even though it might have errors with
1527          * handling the recover logs */
1528         cfg->cfg_last_idx = res->mcr_offset;
1529         eof = res->mcr_offset == res->mcr_size;
1530
1531         CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n",
1532                res->mcr_offset, eof == false);
1533
1534         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1535         if (ealen < 0)
1536                 GOTO(out, rc = ealen);
1537
1538         if (ealen > nrpages << CFS_PAGE_SHIFT)
1539                 GOTO(out, rc = -EINVAL);
1540
1541         if (ealen == 0) { /* no logs transferred */
1542                 if (!eof)
1543                         rc = -EINVAL;
1544                 GOTO(out, rc);
1545         }
1546
1547         mne_swab = !!ptlrpc_rep_need_swab(req);
1548 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1549         /* This import flag means the server did an extra swab of IR MNE
1550          * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1551         if (unlikely(req->rq_import->imp_need_mne_swab))
1552                 mne_swab = !mne_swab;
1553 #else
1554 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1555 #endif
1556
1557         for (i = 0; i < nrpages && ealen > 0; i++) {
1558                 int rc2;
1559                 void *ptr;
1560
1561                 ptr = cfs_kmap(pages[i]);
1562                 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1563                                              min_t(int, ealen, CFS_PAGE_SIZE),
1564                                              mne_swab);
1565                 cfs_kunmap(pages[i]);
1566                 if (rc2 < 0) {
1567                         CWARN("Process recover log %s error %d\n",
1568                               cld->cld_logname, rc2);
1569                         break;
1570                 }
1571
1572                 ealen -= CFS_PAGE_SIZE;
1573         }
1574
1575 out:
1576         if (req)
1577                 ptlrpc_req_finished(req);
1578
1579         if (rc == 0 && !eof)
1580                 goto again;
1581
1582         if (pages) {
1583                 for (i = 0; i < nrpages; i++) {
1584                         if (pages[i] == NULL)
1585                                 break;
1586                         cfs_free_page(pages[i]);
1587                 }
1588                 OBD_FREE(pages, sizeof(*pages) * nrpages);
1589         }
1590         return rc;
1591 }
1592
1593 #ifdef HAVE_LDISKFS_OSD
1594
1595 /*
1596  * XXX: mgc_copy_llog() does not support osd-based llogs yet
1597  */
1598
1599 /* identical to mgs_log_is_empty */
1600 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1601                              char *name)
1602 {
1603         struct lvfs_run_ctxt     saved;
1604         struct llog_handle      *llh;
1605         int                      rc = 0;
1606
1607         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1608         rc = llog_open(NULL, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1609         if (rc == 0) {
1610                 rc = llog_init_handle(NULL, llh, LLOG_F_IS_PLAIN, NULL);
1611                 if (rc == 0)
1612                         rc = llog_get_size(llh);
1613                 llog_close(NULL, llh);
1614         } else if (rc == -ENOENT) {
1615                 rc = 0;
1616         }
1617         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1618         /* header is record 1 */
1619         return (rc <= 1);
1620 }
1621
1622 /* Copy a remote log locally */
1623 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1624                          struct llog_ctxt *lctxt, char *logname)
1625 {
1626         struct llog_handle *local_llh, *remote_llh;
1627         struct obd_uuid *uuid;
1628         char *temp_log;
1629         int rc, rc2;
1630         ENTRY;
1631
1632         /* Write new log to a temp name, then vfs_rename over logname
1633            upon successful completion. */
1634
1635         OBD_ALLOC(temp_log, strlen(logname) + 1);
1636         if (!temp_log)
1637                 RETURN(-ENOMEM);
1638         sprintf(temp_log, "%sT", logname);
1639
1640         /* Make sure there's no old temp log */
1641         rc = llog_erase(NULL, lctxt, NULL, temp_log);
1642         if (rc < 0 && rc != -ENOENT)
1643                 GOTO(out, rc);
1644
1645         /* open local log */
1646         rc = llog_open_create(NULL, lctxt, &local_llh, NULL, temp_log);
1647         if (rc)
1648                 GOTO(out, rc);
1649
1650         /* set the log header uuid for fun */
1651         OBD_ALLOC_PTR(uuid);
1652         obd_str2uuid(uuid, logname);
1653         rc = llog_init_handle(NULL, local_llh, LLOG_F_IS_PLAIN, uuid);
1654         OBD_FREE_PTR(uuid);
1655         if (rc)
1656                 GOTO(out_closel, rc);
1657
1658         /* open remote log */
1659         rc = llog_open(NULL, rctxt, &remote_llh, NULL, logname,
1660                        LLOG_OPEN_EXISTS);
1661         if (rc < 0) {
1662                 if (rc == -ENOENT)
1663                         rc = 0;
1664                 GOTO(out_closel, rc);
1665         }
1666
1667         rc = llog_init_handle(NULL, remote_llh, LLOG_F_IS_PLAIN, NULL);
1668         if (rc)
1669                 GOTO(out_closer, rc);
1670
1671         /* Copy remote log */
1672         rc = llog_process(NULL, remote_llh, llog_copy_handler,
1673                           (void *)local_llh, NULL);
1674
1675 out_closer:
1676         rc2 = llog_close(NULL, remote_llh);
1677         if (!rc)
1678                 rc = rc2;
1679 out_closel:
1680         rc2 = llog_close(NULL, local_llh);
1681         if (!rc)
1682                 rc = rc2;
1683
1684         /* We've copied the remote log to the local temp log, now
1685            replace the old local log with the temp log. */
1686         if (rc == 0) {
1687                 struct client_obd *cli = &obd->u.cli;
1688
1689                 LASSERT(cli);
1690                 LASSERT(cli->cl_mgc_configs_dir);
1691                 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1692                                    temp_log, logname);
1693         }
1694         CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1695 out:
1696         if (rc)
1697                 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1698         OBD_FREE(temp_log, strlen(logname) + 1);
1699         RETURN(rc);
1700 }
1701 #endif
1702
1703 /* local_only means it cannot get remote llogs */
1704 static int mgc_process_cfg_log(struct obd_device *mgc,
1705                                struct config_llog_data *cld,
1706                                int local_only)
1707 {
1708         struct llog_ctxt *ctxt, *lctxt = NULL;
1709 #ifdef HAVE_LDISKFS_OSD
1710         struct client_obd *cli = &mgc->u.cli;
1711 #endif
1712         struct lvfs_run_ctxt *saved_ctxt;
1713         struct lustre_sb_info *lsi = NULL;
1714         int rc = 0, must_pop = 0;
1715         bool sptlrpc_started = false;
1716
1717         ENTRY;
1718
1719         LASSERT(cld);
1720         LASSERT(mutex_is_locked(&cld->cld_lock));
1721
1722         /*
1723          * local copy of sptlrpc log is controlled elsewhere, don't try to
1724          * read it up here.
1725          */
1726         if (cld_is_sptlrpc(cld) && local_only)
1727                 RETURN(0);
1728
1729         if (cld->cld_cfg.cfg_sb)
1730                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1731
1732         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1733         if (!ctxt) {
1734                 CERROR("missing llog context\n");
1735                 RETURN(-EINVAL);
1736         }
1737
1738         OBD_ALLOC_PTR(saved_ctxt);
1739         if (saved_ctxt == NULL)
1740                 RETURN(-ENOMEM);
1741
1742         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1743
1744 #ifdef HAVE_LDISKFS_OSD
1745         /*
1746          * XXX: at the moment mgc_copy_llog() works with lvfs-based llogs
1747          */
1748         /* Copy the setup log locally if we can. Don't mess around if we're
1749            running an MGS though (logs are already local). */
1750         if (lctxt && lsi && IS_SERVER(lsi) &&
1751             (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1752             !IS_MGS(lsi) && lsi->lsi_srv_mnt) {
1753                 push_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1754                 must_pop++;
1755                 if (!local_only)
1756                         /* Only try to copy log if we have the lock. */
1757                         rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1758                 if (local_only || rc) {
1759                         if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1760                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1761                                                    "log %s and no local copy."
1762                                                    "\n", cld->cld_logname);
1763                                 GOTO(out_pop, rc = -ENOTCONN);
1764                         }
1765                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1766                                "copy for now, will try to update later.\n",
1767                                cld->cld_logname);
1768                 }
1769                 /* Now, whether we copied or not, start using the local llog.
1770                    If we failed to copy, we'll start using whatever the old
1771                    log has. */
1772                 llog_ctxt_put(ctxt);
1773                 ctxt = lctxt;
1774                 lctxt = NULL;
1775         } else
1776 #endif
1777                 if (local_only) { /* no local log at client side */
1778                 GOTO(out_pop, rc = -EIO);
1779         }
1780
1781         if (cld_is_sptlrpc(cld)) {
1782                 sptlrpc_conf_log_update_begin(cld->cld_logname);
1783                 sptlrpc_started = true;
1784         }
1785
1786         /* logname and instance info should be the same, so use our
1787            copy of the instance for the update.  The cfg_last_idx will
1788            be updated here. */
1789         rc = class_config_parse_llog(NULL, ctxt, cld->cld_logname,
1790                                      &cld->cld_cfg);
1791         EXIT;
1792
1793 out_pop:
1794         llog_ctxt_put(ctxt);
1795         if (lctxt)
1796                 llog_ctxt_put(lctxt);
1797         if (must_pop)
1798                 pop_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1799
1800         OBD_FREE_PTR(saved_ctxt);
1801         /*
1802          * update settings on existing OBDs. doing it inside
1803          * of llog_process_lock so no device is attaching/detaching
1804          * in parallel.
1805          * the logname must be <fsname>-sptlrpc
1806          */
1807         if (sptlrpc_started) {
1808                 LASSERT(cld_is_sptlrpc(cld));
1809                 sptlrpc_conf_log_update_end(cld->cld_logname);
1810                 class_notify_sptlrpc_conf(cld->cld_logname,
1811                                           strlen(cld->cld_logname) -
1812                                           strlen("-sptlrpc"));
1813         }
1814
1815         RETURN(rc);
1816 }
1817
1818 /** Get a config log from the MGS and process it.
1819  * This func is called for both clients and servers.
1820  * Copy the log locally before parsing it if appropriate (non-MGS server)
1821  */
1822 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1823 {
1824         struct lustre_handle lockh = { 0 };
1825         __u64 flags = LDLM_FL_NO_LRU;
1826         int rc = 0, rcl;
1827         ENTRY;
1828
1829         LASSERT(cld);
1830
1831         /* I don't want multiple processes running process_log at once --
1832            sounds like badness.  It actually might be fine, as long as
1833            we're not trying to update from the same log
1834            simultaneously (in which case we should use a per-log sem.) */
1835         mutex_lock(&cld->cld_lock);
1836         if (cld->cld_stopping) {
1837                 mutex_unlock(&cld->cld_lock);
1838                 RETURN(0);
1839         }
1840
1841         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1842
1843         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1844                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1845
1846         /* Get the cfg lock on the llog */
1847         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1848                           LCK_CR, &flags, NULL, NULL, NULL,
1849                           cld, 0, NULL, &lockh);
1850         if (rcl == 0) {
1851                 /* Get the cld, it will be released in mgc_blocking_ast. */
1852                 config_log_get(cld);
1853                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1854                 LASSERT(rc == 0);
1855         } else {
1856                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1857
1858                 /* mark cld_lostlock so that it will requeue
1859                  * after MGC becomes available. */
1860                 cld->cld_lostlock = 1;
1861                 /* Get extra reference, it will be put in requeue thread */
1862                 config_log_get(cld);
1863         }
1864
1865
1866         if (cld_is_recover(cld)) {
1867                 rc = 0; /* this is not a fatal error for recover log */
1868                 if (rcl == 0)
1869                         rc = mgc_process_recover_log(mgc, cld);
1870         } else {
1871                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1872         }
1873
1874         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1875                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1876
1877         mutex_unlock(&cld->cld_lock);
1878
1879         /* Now drop the lock so MGS can revoke it */
1880         if (!rcl) {
1881                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1882                                  LCK_CR, &lockh);
1883                 if (rcl)
1884                         CERROR("Can't drop cfg lock: %d\n", rcl);
1885         }
1886
1887         RETURN(rc);
1888 }
1889
1890
1891 /** Called from lustre_process_log.
1892  * LCFG_LOG_START gets the config log from the MGS, processes it to start
1893  * any services, and adds it to the list logs to watch (follow).
1894  */
1895 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1896 {
1897         struct lustre_cfg *lcfg = buf;
1898         struct config_llog_instance *cfg = NULL;
1899         char *logname;
1900         int rc = 0;
1901         ENTRY;
1902
1903         switch(lcfg->lcfg_command) {
1904         case LCFG_LOV_ADD_OBD: {
1905                 /* Overloading this cfg command: register a new target */
1906                 struct mgs_target_info *mti;
1907
1908                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1909                     sizeof(struct mgs_target_info))
1910                         GOTO(out, rc = -EINVAL);
1911
1912                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1913                 CDEBUG(D_MGC, "add_target %s %#x\n",
1914                        mti->mti_svname, mti->mti_flags);
1915                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1916                 break;
1917         }
1918         case LCFG_LOV_DEL_OBD:
1919                 /* Unregister has no meaning at the moment. */
1920                 CERROR("lov_del_obd unimplemented\n");
1921                 rc = -ENOSYS;
1922                 break;
1923         case LCFG_SPTLRPC_CONF: {
1924                 rc = sptlrpc_process_config(lcfg);
1925                 break;
1926         }
1927         case LCFG_LOG_START: {
1928                 struct config_llog_data *cld;
1929                 struct super_block *sb;
1930
1931                 logname = lustre_cfg_string(lcfg, 1);
1932                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1933                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1934
1935                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1936                        cfg->cfg_last_idx);
1937
1938                 /* We're only called through here on the initial mount */
1939                 rc = config_log_add(obd, logname, cfg, sb);
1940                 if (rc)
1941                         break;
1942                 cld = config_log_find(logname, cfg);
1943                 if (cld == NULL) {
1944                         rc = -ENOENT;
1945                         break;
1946                 }
1947
1948                 /* COMPAT_146 */
1949                 /* FIXME only set this for old logs!  Right now this forces
1950                    us to always skip the "inside markers" check */
1951                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1952
1953                 rc = mgc_process_log(obd, cld);
1954                 if (rc == 0 && cld->cld_recover != NULL) {
1955                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1956                                          imp_connect_data, IMP_RECOV)) {
1957                                 rc = mgc_process_log(obd, cld->cld_recover);
1958                         } else {
1959                                 struct config_llog_data *cir = cld->cld_recover;
1960                                 cld->cld_recover = NULL;
1961                                 config_log_put(cir);
1962                         }
1963                         if (rc)
1964                                 CERROR("Cannot process recover llog %d\n", rc);
1965                 }
1966                 config_log_put(cld);
1967
1968                 break;
1969         }
1970         case LCFG_LOG_END: {
1971                 logname = lustre_cfg_string(lcfg, 1);
1972
1973                 if (lcfg->lcfg_bufcount >= 2)
1974                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
1975                                 lcfg, 2);
1976                 rc = config_log_end(logname, cfg);
1977                 break;
1978         }
1979         default: {
1980                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1981                 GOTO(out, rc = -EINVAL);
1982
1983         }
1984         }
1985 out:
1986         RETURN(rc);
1987 }
1988
1989 struct obd_ops mgc_obd_ops = {
1990         .o_owner        = THIS_MODULE,
1991         .o_setup        = mgc_setup,
1992         .o_precleanup   = mgc_precleanup,
1993         .o_cleanup      = mgc_cleanup,
1994         .o_add_conn     = client_import_add_conn,
1995         .o_del_conn     = client_import_del_conn,
1996         .o_connect      = client_connect_import,
1997         .o_disconnect   = client_disconnect_export,
1998         //.o_enqueue      = mgc_enqueue,
1999         .o_cancel       = mgc_cancel,
2000         //.o_iocontrol    = mgc_iocontrol,
2001         .o_set_info_async = mgc_set_info_async,
2002         .o_get_info       = mgc_get_info,
2003         .o_import_event = mgc_import_event,
2004         .o_llog_init    = mgc_llog_init,
2005         .o_llog_finish  = mgc_llog_finish,
2006         .o_process_config = mgc_process_config,
2007 };
2008
2009 int __init mgc_init(void)
2010 {
2011         return class_register_type(&mgc_obd_ops, NULL, NULL,
2012                                    LUSTRE_MGC_NAME, NULL);
2013 }
2014
2015 #ifdef __KERNEL__
2016 static void /*__exit*/ mgc_exit(void)
2017 {
2018         class_unregister_type(LUSTRE_MGC_NAME);
2019 }
2020
2021 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2022 MODULE_DESCRIPTION("Lustre Management Client");
2023 MODULE_LICENSE("GPL");
2024
2025 module_init(mgc_init);
2026 module_exit(mgc_exit);
2027 #endif