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, 2014, 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 <linux/kthread.h>
  46 #include <obd_class.h>
  47 #include <lustre_dlm.h>
  48 #include <lprocfs_status.h>
  49 #include <lustre_log.h>
  50 #include <lustre_disk.h>
  51 #include <dt_object.h>
  52
  53 #include "mgc_internal.h"
  54
  55 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
  56                           int type)
  57 {
  58         __u64 resname = 0;
  59
  60         if (len > sizeof(resname)) {
  61                 CERROR("name too long: %s\n", name);
  62                 return -EINVAL;
  63         }
  64         if (len <= 0) {
  65                 CERROR("missing name: %s\n", name);
  66                 return -EINVAL;
  67         }
  68         memcpy(&resname, name, len);
  69
  70         /* Always use the same endianness for the resid */
  71         memset(res_id, 0, sizeof(*res_id));
  72         res_id->name[0] = cpu_to_le64(resname);
  73         /* XXX: unfortunately, sptlprc and config llog share one lock */
  74         switch(type) {
  75         case CONFIG_T_CONFIG:
  76         case CONFIG_T_SPTLRPC:
  77                 resname = 0;
  78                 break;
  79         case CONFIG_T_RECOVER:
  80         case CONFIG_T_PARAMS:
  81                 resname = type;
  82                 break;
  83         default:
  84                 LBUG();
  85         }
  86         res_id->name[1] = cpu_to_le64(resname);
  87         CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
  88                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  89         return 0;
  90 }
  91
  92 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
  93 {
  94         /* fsname is at most 8 chars long, maybe contain "-".
  95          * e.g. "lustre", "SUN-000" */
  96         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
  97 }
  98 EXPORT_SYMBOL(mgc_fsname2resid);
  99
 100 static int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id,
 101                              int type)
 102 {
 103         char *name_end;
 104         int len;
 105
 106         /* logname consists of "fsname-nodetype".
 107          * e.g. "lustre-MDT0001", "SUN-000-client"
 108          * there is an exception: llog "params" */
 109         name_end = strrchr(logname, '-');
 110         if (!name_end)
 111                 len = strlen(logname);
 112         else
 113                 len = name_end - logname;
 114         return mgc_name2resid(logname, len, res_id, type);
 115 }
 116
 117 /********************** config llog list **********************/
 118 static struct list_head config_llog_list = LIST_HEAD_INIT(config_llog_list);
 119 static DEFINE_SPINLOCK(config_list_lock);
 120
 121 /* Take a reference to a config log */
 122 static int config_log_get(struct config_llog_data *cld)
 123 {
 124         ENTRY;
 125         atomic_inc(&cld->cld_refcount);
 126         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 127                 atomic_read(&cld->cld_refcount));
 128         RETURN(0);
 129 }
 130
 131 /* Drop a reference to a config log.  When no longer referenced,
 132    we can free the config log data */
 133 static void config_log_put(struct config_llog_data *cld)
 134 {
 135         ENTRY;
 136
 137         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 138                 atomic_read(&cld->cld_refcount));
 139         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 140
 141         /* spinlock to make sure no item with 0 refcount in the list */
 142         if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 143                 list_del(&cld->cld_list_chain);
 144                 spin_unlock(&config_list_lock);
 145
 146                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 147
 148                 if (cld->cld_recover)
 149                         config_log_put(cld->cld_recover);
 150                 if (cld->cld_sptlrpc)
 151                         config_log_put(cld->cld_sptlrpc);
 152                 if (cld->cld_params)
 153                         config_log_put(cld->cld_params);
 154                 if (cld_is_sptlrpc(cld))
 155                         sptlrpc_conf_log_stop(cld->cld_logname);
 156
 157                 class_export_put(cld->cld_mgcexp);
 158                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 159         }
 160
 161         EXIT;
 162 }
 163
 164 /* Find a config log by name */
 165 static
 166 struct config_llog_data *config_log_find(char *logname,
 167                                          struct config_llog_instance *cfg)
 168 {
 169         struct config_llog_data *cld;
 170         struct config_llog_data *found = NULL;
 171         void *                   instance;
 172         ENTRY;
 173
 174         LASSERT(logname != NULL);
 175
 176         instance = cfg ? cfg->cfg_instance : NULL;
 177         spin_lock(&config_list_lock);
 178         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 179                 /* check if instance equals */
 180                 if (instance != cld->cld_cfg.cfg_instance)
 181                         continue;
 182
 183                 /* instance may be NULL, should check name */
 184                 if (strcmp(logname, cld->cld_logname) == 0) {
 185                         found = cld;
 186                         break;
 187                 }
 188         }
 189         if (found) {
 190                 atomic_inc(&found->cld_refcount);
 191                 LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0);
 192         }
 193         spin_unlock(&config_list_lock);
 194         RETURN(found);
 195 }
 196
 197 static
 198 struct config_llog_data *do_config_log_add(struct obd_device *obd,
 199                                            char *logname,
 200                                            int type,
 201                                            struct config_llog_instance *cfg,
 202                                            struct super_block *sb)
 203 {
 204         struct config_llog_data *cld;
 205         int                      rc;
 206         ENTRY;
 207
 208         CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
 209                cfg ? cfg->cfg_instance : NULL);
 210
 211         OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
 212         if (!cld)
 213                 RETURN(ERR_PTR(-ENOMEM));
 214
 215         strcpy(cld->cld_logname, logname);
 216         if (cfg)
 217                 cld->cld_cfg = *cfg;
 218         else
 219                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 220         mutex_init(&cld->cld_lock);
 221         cld->cld_cfg.cfg_last_idx = 0;
 222         cld->cld_cfg.cfg_flags = 0;
 223         cld->cld_cfg.cfg_sb = sb;
 224         cld->cld_type = type;
 225         atomic_set(&cld->cld_refcount, 1);
 226
 227         /* Keep the mgc around until we are done */
 228         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 229
 230         if (cld_is_sptlrpc(cld)) {
 231                 sptlrpc_conf_log_start(logname);
 232                 cld->cld_cfg.cfg_obdname = obd->obd_name;
 233         }
 234
 235         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 236
 237         spin_lock(&config_list_lock);
 238         list_add(&cld->cld_list_chain, &config_llog_list);
 239         spin_unlock(&config_list_lock);
 240
 241         if (rc) {
 242                 config_log_put(cld);
 243                 RETURN(ERR_PTR(rc));
 244         }
 245
 246         if (cld_is_sptlrpc(cld)) {
 247                 rc = mgc_process_log(obd, cld);
 248                 if (rc && rc != -ENOENT)
 249                         CERROR("failed processing sptlrpc log: %d\n", rc);
 250         }
 251
 252         RETURN(cld);
 253 }
 254
 255 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
 256         char *fsname,
 257         struct config_llog_instance *cfg,
 258         struct super_block *sb)
 259 {
 260         struct config_llog_instance lcfg = *cfg;
 261         struct lustre_sb_info *lsi = s2lsi(sb);
 262         struct config_llog_data *cld;
 263         char logname[32];
 264
 265         if (IS_OST(lsi))
 266                 return NULL;
 267
 268         /* for osp-on-ost, see lustre_start_osp() */
 269         if (IS_MDT(lsi) && lcfg.cfg_instance)
 270                 return NULL;
 271
 272         /* we have to use different llog for clients and mdts for cmd
 273          * where only clients are notified if one of cmd server restarts */
 274         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 275         strcpy(logname, fsname);
 276         if (IS_SERVER(lsi)) { /* mdt */
 277                 LASSERT(lcfg.cfg_instance == NULL);
 278                 lcfg.cfg_instance = sb;
 279                 strcat(logname, "-mdtir");
 280         } else {
 281                 LASSERT(lcfg.cfg_instance != NULL);
 282                 strcat(logname, "-cliir");
 283         }
 284
 285         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
 286         return cld;
 287 }
 288
 289 static struct config_llog_data *config_params_log_add(struct obd_device *obd,
 290         struct config_llog_instance *cfg, struct super_block *sb)
 291 {
 292         struct config_llog_instance     lcfg = *cfg;
 293         struct config_llog_data         *cld;
 294
 295         lcfg.cfg_instance = sb;
 296
 297         cld = do_config_log_add(obd, PARAMS_FILENAME, CONFIG_T_PARAMS,
 298                                 &lcfg, sb);
 299
 300         return cld;
 301 }
 302
 303 /** Add this log to the list of active logs watched by an MGC.
 304  * Active means we're watching for updates.
 305  * We have one active log per "mount" - client instance or servername.
 306  * Each instance may be at a different point in the log.
 307  */
 308 static int config_log_add(struct obd_device *obd, char *logname,
 309                           struct config_llog_instance *cfg,
 310                           struct super_block *sb)
 311 {
 312         struct lustre_sb_info   *lsi = s2lsi(sb);
 313         struct config_llog_data *cld;
 314         struct config_llog_data *sptlrpc_cld;
 315         struct config_llog_data *params_cld;
 316         char                    seclogname[32];
 317         char                    *ptr;
 318         int                     rc;
 319         ENTRY;
 320
 321         CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
 322
 323         /*
 324          * for each regular log, the depended sptlrpc log name is
 325          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 326          */
 327         ptr = strrchr(logname, '-');
 328         if (ptr == NULL || ptr - logname > 8) {
 329                 CERROR("logname %s is too long\n", logname);
 330                 RETURN(-EINVAL);
 331         }
 332
 333         memcpy(seclogname, logname, ptr - logname);
 334         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 335
 336         sptlrpc_cld = config_log_find(seclogname, NULL);
 337         if (sptlrpc_cld == NULL) {
 338                 sptlrpc_cld = do_config_log_add(obd, seclogname,
 339                                                 CONFIG_T_SPTLRPC, NULL, NULL);
 340                 if (IS_ERR(sptlrpc_cld)) {
 341                         CERROR("can't create sptlrpc log: %s\n", seclogname);
 342                         GOTO(out_err, rc = PTR_ERR(sptlrpc_cld));
 343                 }
 344         }
 345         params_cld = config_params_log_add(obd, cfg, sb);
 346         if (IS_ERR(params_cld)) {
 347                 rc = PTR_ERR(params_cld);
 348                 CERROR("%s: can't create params log: rc = %d\n",
 349                        obd->obd_name, rc);
 350                 GOTO(out_err1, rc);
 351         }
 352
 353         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 354         if (IS_ERR(cld)) {
 355                 CERROR("can't create log: %s\n", logname);
 356                 GOTO(out_err2, rc = PTR_ERR(cld));
 357         }
 358
 359         cld->cld_sptlrpc = sptlrpc_cld;
 360         cld->cld_params = params_cld;
 361
 362         LASSERT(lsi->lsi_lmd);
 363         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
 364                 struct config_llog_data *recover_cld;
 365                 ptr = strrchr(seclogname, '-');
 366                 if (ptr != NULL) {
 367                         *ptr = 0;
 368                 }
 369                 else {
 370                         CERROR("%s: sptlrpc log name not correct, %s: "
 371                                "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
 372                         config_log_put(cld);
 373                         RETURN(-EINVAL);
 374                 }
 375                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 376                 if (IS_ERR(recover_cld))
 377                         GOTO(out_err3, rc = PTR_ERR(recover_cld));
 378                 cld->cld_recover = recover_cld;
 379         }
 380
 381         RETURN(0);
 382
 383 out_err3:
 384         config_log_put(cld);
 385
 386 out_err2:
 387         config_log_put(params_cld);
 388
 389 out_err1:
 390         config_log_put(sptlrpc_cld);
 391
 392 out_err:
 393         RETURN(rc);
 394 }
 395
 396 DEFINE_MUTEX(llog_process_lock);
 397
 398 /** Stop watching for updates on this log.
 399  */
 400 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 401 {
 402         struct config_llog_data *cld;
 403         struct config_llog_data *cld_sptlrpc = NULL;
 404         struct config_llog_data *cld_params = NULL;
 405         struct config_llog_data *cld_recover = NULL;
 406         int rc = 0;
 407         ENTRY;
 408
 409         cld = config_log_find(logname, cfg);
 410         if (cld == NULL)
 411                 RETURN(-ENOENT);
 412
 413         mutex_lock(&cld->cld_lock);
 414         /*
 415          * if cld_stopping is set, it means we didn't start the log thus
 416          * not owning the start ref. this can happen after previous umount:
 417          * the cld still hanging there waiting for lock cancel, and we
 418          * remount again but failed in the middle and call log_end without
 419          * calling start_log.
 420          */
 421         if (unlikely(cld->cld_stopping)) {
 422                 mutex_unlock(&cld->cld_lock);
 423                 /* drop the ref from the find */
 424                 config_log_put(cld);
 425                 RETURN(rc);
 426         }
 427
 428         cld->cld_stopping = 1;
 429
 430         cld_recover = cld->cld_recover;
 431         cld->cld_recover = NULL;
 432         mutex_unlock(&cld->cld_lock);
 433
 434         if (cld_recover) {
 435                 mutex_lock(&cld_recover->cld_lock);
 436                 cld_recover->cld_stopping = 1;
 437                 mutex_unlock(&cld_recover->cld_lock);
 438                 config_log_put(cld_recover);
 439         }
 440
 441         spin_lock(&config_list_lock);
 442         cld_sptlrpc = cld->cld_sptlrpc;
 443         cld->cld_sptlrpc = NULL;
 444         cld_params = cld->cld_params;
 445         cld->cld_params = NULL;
 446         spin_unlock(&config_list_lock);
 447
 448         if (cld_sptlrpc)
 449                 config_log_put(cld_sptlrpc);
 450
 451         if (cld_params) {
 452                 mutex_lock(&cld_params->cld_lock);
 453                 cld_params->cld_stopping = 1;
 454                 mutex_unlock(&cld_params->cld_lock);
 455                 config_log_put(cld_params);
 456         }
 457
 458         /* drop the ref from the find */
 459         config_log_put(cld);
 460         /* drop the start ref */
 461         config_log_put(cld);
 462
 463         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 464                rc);
 465         RETURN(rc);
 466 }
 467
 468 #ifdef CONFIG_PROC_FS
 469 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 470 {
 471         struct obd_device       *obd = data;
 472         struct obd_import       *imp;
 473         struct obd_connect_data *ocd;
 474         struct config_llog_data *cld;
 475         ENTRY;
 476
 477         LASSERT(obd != NULL);
 478         LPROCFS_CLIMP_CHECK(obd);
 479         imp = obd->u.cli.cl_import;
 480         ocd = &imp->imp_connect_data;
 481
 482         seq_printf(m, "imperative_recovery: %s\n",
 483                    OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 484         seq_printf(m, "client_state:\n");
 485
 486         spin_lock(&config_list_lock);
 487         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 488                 if (cld->cld_recover == NULL)
 489                         continue;
 490                 seq_printf(m,  "    - { client: %s, nidtbl_version: %u }\n",
 491                                cld->cld_logname,
 492                                cld->cld_recover->cld_cfg.cfg_last_idx);
 493         }
 494         spin_unlock(&config_list_lock);
 495
 496         LPROCFS_CLIMP_EXIT(obd);
 497         RETURN(0);
 498 }
 499 #endif
 500
 501 /* reenqueue any lost locks */
 502 #define RQ_RUNNING      0x1
 503 #define RQ_NOW          0x2
 504 #define RQ_LATER        0x4
 505 #define RQ_STOP         0x8
 506 #define RQ_PRECLEANUP   0x10
 507 static int                    rq_state = 0;
 508 static wait_queue_head_t      rq_waitq;
 509 static DECLARE_COMPLETION(rq_exit);
 510 static DECLARE_COMPLETION(rq_start);
 511
 512 static void do_requeue(struct config_llog_data *cld)
 513 {
 514         int rc = 0;
 515         ENTRY;
 516
 517         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 518
 519         /* Do not run mgc_process_log on a disconnected export or an
 520          * export which is being disconnected. Take the client
 521          * semaphore to make the check non-racy. */
 522         down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 523         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 524                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 525                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 526                 if (rc && rc != -ENOENT)
 527                         CERROR("failed processing log: %d\n", rc);
 528         } else {
 529                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 530                        cld->cld_logname);
 531         }
 532         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 533
 534         EXIT;
 535 }
 536
 537 /* this timeout represents how many seconds MGC should wait before
 538  * requeue config and recover lock to the MGS. We need to randomize this
 539  * in order to not flood the MGS.
 540  */
 541 #define MGC_TIMEOUT_MIN_SECONDS   5
 542 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
 543
 544 static int mgc_requeue_thread(void *data)
 545 {
 546         int rc = 0;
 547         bool first = true;
 548         ENTRY;
 549
 550         CDEBUG(D_MGC, "Starting requeue thread\n");
 551
 552         /* Keep trying failed locks periodically */
 553         spin_lock(&config_list_lock);
 554         rq_state |= RQ_RUNNING;
 555         while (1) {
 556                 struct l_wait_info lwi;
 557                 struct config_llog_data *cld, *cld_prev;
 558                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
 559                 int stopped = !!(rq_state & RQ_STOP);
 560                 int to;
 561
 562                 /* Any new or requeued lostlocks will change the state */
 563                 rq_state &= ~(RQ_NOW | RQ_LATER);
 564                 spin_unlock(&config_list_lock);
 565
 566                 if (first) {
 567                         first = false;
 568                         complete(&rq_start);
 569                 }
 570
 571                 /* Always wait a few seconds to allow the server who
 572                    caused the lock revocation to finish its setup, plus some
 573                    random so everyone doesn't try to reconnect at once. */
 574                 to = msecs_to_jiffies(MGC_TIMEOUT_MIN_SECONDS * MSEC_PER_SEC);
 575                 /* rand is centi-seconds */
 576                 to += msecs_to_jiffies(rand * MSEC_PER_SEC / 100);
 577                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 578                 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
 579                              &lwi);
 580
 581                 /*
 582                  * iterate & processing through the list. for each cld, process
 583                  * its depending sptlrpc cld firstly (if any) and then itself.
 584                  *
 585                  * it's guaranteed any item in the list must have
 586                  * reference > 0; and if cld_lostlock is set, at
 587                  * least one reference is taken by the previous enqueue.
 588                  */
 589                 cld_prev = NULL;
 590
 591                 spin_lock(&config_list_lock);
 592                 rq_state &= ~RQ_PRECLEANUP;
 593                 list_for_each_entry(cld, &config_llog_list,
 594                                         cld_list_chain) {
 595                         if (!cld->cld_lostlock)
 596                                 continue;
 597
 598                         spin_unlock(&config_list_lock);
 599
 600                         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 601
 602                         /* Whether we enqueued again or not in mgc_process_log,
 603                          * we're done with the ref from the old enqueue */
 604                         if (cld_prev)
 605                                 config_log_put(cld_prev);
 606                         cld_prev = cld;
 607
 608                         cld->cld_lostlock = 0;
 609                         if (likely(!stopped))
 610                                 do_requeue(cld);
 611
 612                         spin_lock(&config_list_lock);
 613                 }
 614                 spin_unlock(&config_list_lock);
 615                 if (cld_prev)
 616                         config_log_put(cld_prev);
 617
 618                 /* break after scanning the list so that we can drop
 619                  * refcount to losing lock clds */
 620                 if (unlikely(stopped)) {
 621                         spin_lock(&config_list_lock);
 622                         break;
 623                 }
 624
 625                 /* Wait a bit to see if anyone else needs a requeue */
 626                 lwi = (struct l_wait_info) { 0 };
 627                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 628                              &lwi);
 629                 spin_lock(&config_list_lock);
 630         }
 631         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 632         rq_state &= ~RQ_RUNNING;
 633         spin_unlock(&config_list_lock);
 634
 635         complete(&rq_exit);
 636
 637         CDEBUG(D_MGC, "Ending requeue thread\n");
 638         RETURN(rc);
 639 }
 640
 641 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 642    We are responsible for dropping the config log reference from here on out. */
 643 static void mgc_requeue_add(struct config_llog_data *cld)
 644 {
 645         ENTRY;
 646
 647         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 648                 cld->cld_logname, atomic_read(&cld->cld_refcount),
 649                 cld->cld_stopping, rq_state);
 650         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 651
 652         mutex_lock(&cld->cld_lock);
 653         if (cld->cld_stopping || cld->cld_lostlock) {
 654                 mutex_unlock(&cld->cld_lock);
 655                 RETURN_EXIT;
 656         }
 657         /* this refcount will be released in mgc_requeue_thread. */
 658         config_log_get(cld);
 659         cld->cld_lostlock = 1;
 660         mutex_unlock(&cld->cld_lock);
 661
 662         /* Hold lock for rq_state */
 663         spin_lock(&config_list_lock);
 664         if (rq_state & RQ_STOP) {
 665                 spin_unlock(&config_list_lock);
 666                 cld->cld_lostlock = 0;
 667                 config_log_put(cld);
 668         } else {
 669                 rq_state |= RQ_NOW;
 670                 spin_unlock(&config_list_lock);
 671                 wake_up(&rq_waitq);
 672         }
 673         EXIT;
 674 }
 675
 676 /********************** class fns **********************/
 677 static int mgc_local_llog_init(const struct lu_env *env,
 678                                struct obd_device *obd,
 679                                struct obd_device *disk)
 680 {
 681         struct llog_ctxt        *ctxt;
 682         int                      rc;
 683
 684         ENTRY;
 685
 686         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
 687                         &llog_osd_ops);
 688         if (rc)
 689                 RETURN(rc);
 690
 691         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 692         LASSERT(ctxt);
 693         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
 694         llog_ctxt_put(ctxt);
 695
 696         RETURN(0);
 697 }
 698
 699 static int mgc_local_llog_fini(const struct lu_env *env,
 700                                struct obd_device *obd)
 701 {
 702         struct llog_ctxt *ctxt;
 703
 704         ENTRY;
 705
 706         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 707         llog_cleanup(env, ctxt);
 708
 709         RETURN(0);
 710 }
 711
 712 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb)
 713 {
 714         struct lustre_sb_info   *lsi = s2lsi(sb);
 715         struct client_obd       *cli = &obd->u.cli;
 716         struct lu_fid            rfid, fid;
 717         struct dt_object        *root, *dto;
 718         struct lu_env           *env;
 719         int                      rc = 0;
 720
 721         ENTRY;
 722
 723         LASSERT(lsi);
 724         LASSERT(lsi->lsi_dt_dev);
 725
 726         OBD_ALLOC_PTR(env);
 727         if (env == NULL)
 728                 RETURN(-ENOMEM);
 729
 730         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
 731         mutex_lock(&cli->cl_mgc_mutex);
 732
 733         /* Setup the configs dir */
 734         rc = lu_env_init(env, LCT_MG_THREAD);
 735         if (rc)
 736                 GOTO(out_err, rc);
 737
 738         fid.f_seq = FID_SEQ_LOCAL_NAME;
 739         fid.f_oid = 1;
 740         fid.f_ver = 0;
 741         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
 742                                     &cli->cl_mgc_los);
 743         if (rc)
 744                 GOTO(out_env, rc);
 745
 746         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
 747         if (rc)
 748                 GOTO(out_env, rc);
 749
 750         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
 751                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
 752         if (unlikely(IS_ERR(root)))
 753                 GOTO(out_los, rc = PTR_ERR(root));
 754
 755         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
 756                                         MOUNT_CONFIGS_DIR,
 757                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
 758         lu_object_put_nocache(env, &root->do_lu);
 759         if (IS_ERR(dto))
 760                 GOTO(out_los, rc = PTR_ERR(dto));
 761
 762         cli->cl_mgc_configs_dir = dto;
 763
 764         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
 765         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
 766         if (rc)
 767                 GOTO(out_llog, rc);
 768
 769         /* We take an obd ref to insure that we can't get to mgc_cleanup
 770          * without calling mgc_fs_cleanup first. */
 771         class_incref(obd, "mgc_fs", obd);
 772
 773         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 774         EXIT;
 775 out_llog:
 776         if (rc) {
 777                 lu_object_put(env, &cli->cl_mgc_configs_dir->do_lu);
 778                 cli->cl_mgc_configs_dir = NULL;
 779         }
 780 out_los:
 781         if (rc < 0) {
 782                 local_oid_storage_fini(env, cli->cl_mgc_los);
 783                 cli->cl_mgc_los = NULL;
 784                 mutex_unlock(&cli->cl_mgc_mutex);
 785         }
 786 out_env:
 787         lu_env_fini(env);
 788 out_err:
 789         OBD_FREE_PTR(env);
 790         return rc;
 791 }
 792
 793 static int mgc_fs_cleanup(struct obd_device *obd)
 794 {
 795         struct lu_env            env;
 796         struct client_obd       *cli = &obd->u.cli;
 797         int                      rc;
 798
 799         ENTRY;
 800
 801         LASSERT(cli->cl_mgc_los != NULL);
 802
 803         rc = lu_env_init(&env, LCT_MG_THREAD);
 804         if (rc)
 805                 GOTO(unlock, rc);
 806
 807         mgc_local_llog_fini(&env, obd);
 808
 809         lu_object_put_nocache(&env, &cli->cl_mgc_configs_dir->do_lu);
 810         cli->cl_mgc_configs_dir = NULL;
 811
 812         local_oid_storage_fini(&env, cli->cl_mgc_los);
 813         cli->cl_mgc_los = NULL;
 814         lu_env_fini(&env);
 815
 816 unlock:
 817         class_decref(obd, "mgc_fs", obd);
 818         mutex_unlock(&cli->cl_mgc_mutex);
 819
 820         RETURN(0);
 821 }
 822
 823 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 824 {
 825         struct llog_ctxt        *ctxt;
 826         int                      rc;
 827
 828         ENTRY;
 829
 830         /* setup only remote ctxt, the local disk context is switched per each
 831          * filesystem during mgc_fs_setup() */
 832         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 833                         &llog_client_ops);
 834         if (rc)
 835                 RETURN(rc);
 836
 837         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 838         LASSERT(ctxt);
 839
 840         llog_initiator_connect(ctxt);
 841         llog_ctxt_put(ctxt);
 842
 843         RETURN(0);
 844 }
 845
 846 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 847 {
 848         struct llog_ctxt *ctxt;
 849
 850         ENTRY;
 851
 852         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 853         if (ctxt)
 854                 llog_cleanup(env, ctxt);
 855
 856         RETURN(0);
 857 }
 858
 859
 860 static atomic_t mgc_count = ATOMIC_INIT(0);
 861 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 862 {
 863         int rc = 0;
 864         int temp;
 865         ENTRY;
 866
 867         switch (stage) {
 868         case OBD_CLEANUP_EARLY:
 869                 break;
 870         case OBD_CLEANUP_EXPORTS:
 871                 if (atomic_dec_and_test(&mgc_count)) {
 872                         LASSERT(rq_state & RQ_RUNNING);
 873                         /* stop requeue thread */
 874                         temp = RQ_STOP;
 875                 } else {
 876                         /* wakeup requeue thread to clean our cld */
 877                         temp = RQ_NOW | RQ_PRECLEANUP;
 878                 }
 879                 spin_lock(&config_list_lock);
 880                 rq_state |= temp;
 881                 spin_unlock(&config_list_lock);
 882                 wake_up(&rq_waitq);
 883                 if (temp & RQ_STOP)
 884                         wait_for_completion(&rq_exit);
 885                 obd_cleanup_client_import(obd);
 886                 rc = mgc_llog_fini(NULL, obd);
 887                 if (rc != 0)
 888                         CERROR("failed to cleanup llogging subsystems\n");
 889                 break;
 890         }
 891         RETURN(rc);
 892 }
 893
 894 static int mgc_cleanup(struct obd_device *obd)
 895 {
 896         int rc;
 897         ENTRY;
 898
 899         /* COMPAT_146 - old config logs may have added profiles we don't
 900            know about */
 901         if (obd->obd_type->typ_refcnt <= 1)
 902                 /* Only for the last mgc */
 903                 class_del_profiles();
 904
 905         lprocfs_obd_cleanup(obd);
 906         ptlrpcd_decref();
 907
 908         rc = client_obd_cleanup(obd);
 909         RETURN(rc);
 910 }
 911
 912 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 913 {
 914         struct task_struct      *task;
 915         int                      rc;
 916         ENTRY;
 917
 918         rc = ptlrpcd_addref();
 919         if (rc < 0)
 920                 RETURN(rc);
 921
 922         rc = client_obd_setup(obd, lcfg);
 923         if (rc)
 924                 GOTO(err_decref, rc);
 925
 926         rc = mgc_llog_init(NULL, obd);
 927         if (rc) {
 928                 CERROR("failed to setup llogging subsystems\n");
 929                 GOTO(err_cleanup, rc);
 930         }
 931
 932 #ifdef CONFIG_PROC_FS
 933         obd->obd_vars = lprocfs_mgc_obd_vars;
 934         lprocfs_obd_setup(obd);
 935 #endif
 936         sptlrpc_lprocfs_cliobd_attach(obd);
 937
 938         if (atomic_inc_return(&mgc_count) == 1) {
 939                 rq_state = 0;
 940                 init_waitqueue_head(&rq_waitq);
 941
 942                 /* start requeue thread */
 943                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 944                 if (IS_ERR(task)) {
 945                         rc = PTR_ERR(task);
 946                         CERROR("%s: cannot start requeue thread: rc = %d; "
 947                                "no more log updates\n",
 948                                obd->obd_name, rc);
 949                         GOTO(err_cleanup, rc);
 950                 }
 951                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 952                 rc = 0;
 953                 wait_for_completion(&rq_start);
 954         }
 955
 956         RETURN(rc);
 957
 958 err_cleanup:
 959         client_obd_cleanup(obd);
 960 err_decref:
 961         ptlrpcd_decref();
 962         RETURN(rc);
 963 }
 964
 965 /* based on ll_mdc_blocking_ast */
 966 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 967                             void *data, int flag)
 968 {
 969         struct lustre_handle lockh;
 970         struct config_llog_data *cld = (struct config_llog_data *)data;
 971         int rc = 0;
 972         ENTRY;
 973
 974         switch (flag) {
 975         case LDLM_CB_BLOCKING:
 976                 /* mgs wants the lock, give it up... */
 977                 LDLM_DEBUG(lock, "MGC blocking CB");
 978                 ldlm_lock2handle(lock, &lockh);
 979                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 980                 break;
 981         case LDLM_CB_CANCELING:
 982                 /* We've given up the lock, prepare ourselves to update. */
 983                 LDLM_DEBUG(lock, "MGC cancel CB");
 984
 985                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
 986                        PLDLMRES(lock->l_resource),
 987                        (char *)&lock->l_resource->lr_name.name[0]);
 988
 989                 if (!cld) {
 990                         CDEBUG(D_INFO, "missing data, won't requeue\n");
 991                         break;
 992                 }
 993
 994                 /* held at mgc_process_log(). */
 995                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
 996                 /* Are we done with this log? */
 997                 if (cld->cld_stopping) {
 998                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
 999                                 cld->cld_logname);
1000                         config_log_put(cld);
1001                         break;
1002                 }
1003                 /* Make sure not to re-enqueue when the mgc is stopping
1004                    (we get called from client_disconnect_export) */
1005                 if (lock->l_conn_export == NULL ||
1006                     lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1007                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1008                                 cld->cld_logname);
1009                         config_log_put(cld);
1010                         break;
1011                 }
1012
1013                 /* Re-enqueue now */
1014                 mgc_requeue_add(cld);
1015                 config_log_put(cld);
1016                 break;
1017         default:
1018                 LBUG();
1019         }
1020
1021         RETURN(rc);
1022 }
1023
1024 /* Not sure where this should go... */
1025 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1026  * that we can have a chance to try the secondary MGS if any. */
1027 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1028                                 + PING_INTERVAL)
1029 #define  MGC_TARGET_REG_LIMIT 10
1030 #define  MGC_SEND_PARAM_LIMIT 10
1031
1032 /* Send parameter to MGS*/
1033 static int mgc_set_mgs_param(struct obd_export *exp,
1034                              struct mgs_send_param *msp)
1035 {
1036         struct ptlrpc_request *req;
1037         struct mgs_send_param *req_msp, *rep_msp;
1038         int rc;
1039         ENTRY;
1040
1041         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1042                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1043                                         MGS_SET_INFO);
1044         if (!req)
1045                 RETURN(-ENOMEM);
1046
1047         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1048         if (!req_msp) {
1049                 ptlrpc_req_finished(req);
1050                 RETURN(-ENOMEM);
1051         }
1052
1053         memcpy(req_msp, msp, sizeof(*req_msp));
1054         ptlrpc_request_set_replen(req);
1055
1056         /* Limit how long we will wait for the enqueue to complete */
1057         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1058         rc = ptlrpc_queue_wait(req);
1059         if (!rc) {
1060                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1061                 memcpy(msp, rep_msp, sizeof(*rep_msp));
1062         }
1063
1064         ptlrpc_req_finished(req);
1065
1066         RETURN(rc);
1067 }
1068
1069 /* Take a config lock so we can get cancel notifications */
1070 static int mgc_enqueue(struct obd_export *exp, __u32 type,
1071                        ldlm_policy_data_t *policy, __u32 mode,
1072                        __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
1073                        void *data, __u32 lvb_len, void *lvb_swabber,
1074                        struct lustre_handle *lockh)
1075 {
1076         struct config_llog_data *cld = (struct config_llog_data *)data;
1077         struct ldlm_enqueue_info einfo = {
1078                 .ei_type        = type,
1079                 .ei_mode        = mode,
1080                 .ei_cb_bl       = mgc_blocking_ast,
1081                 .ei_cb_cp       = ldlm_completion_ast,
1082         };
1083         struct ptlrpc_request *req;
1084         int short_limit = cld_is_sptlrpc(cld);
1085         int rc;
1086         ENTRY;
1087
1088         CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
1089                cld->cld_resid.name[0]);
1090
1091         /* We need a callback for every lockholder, so don't try to
1092            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1093         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1094                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1095                                         LDLM_ENQUEUE);
1096         if (req == NULL)
1097                 RETURN(-ENOMEM);
1098
1099         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1100         ptlrpc_request_set_replen(req);
1101
1102         /* check if this is server or client */
1103         if (cld->cld_cfg.cfg_sb) {
1104                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1105                 if (lsi && IS_SERVER(lsi))
1106                         short_limit = 1;
1107         }
1108         /* Limit how long we will wait for the enqueue to complete */
1109         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1110         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1111                               NULL, 0, LVB_T_NONE, lockh, 0);
1112         /* A failed enqueue should still call the mgc_blocking_ast,
1113            where it will be requeued if needed ("grant failed"). */
1114         ptlrpc_req_finished(req);
1115         RETURN(rc);
1116 }
1117
1118 static int mgc_cancel(struct obd_export *exp, ldlm_mode_t mode,
1119                       struct lustre_handle *lockh)
1120 {
1121         ENTRY;
1122
1123         ldlm_lock_decref(lockh, mode);
1124
1125         RETURN(0);
1126 }
1127
1128 static void mgc_notify_active(struct obd_device *unused)
1129 {
1130         /* wakeup mgc_requeue_thread to requeue mgc lock */
1131         spin_lock(&config_list_lock);
1132         rq_state |= RQ_NOW;
1133         spin_unlock(&config_list_lock);
1134         wake_up(&rq_waitq);
1135
1136         /* TODO: Help the MGS rebuild nidtbl. -jay */
1137 }
1138
1139 /* Send target_reg message to MGS */
1140 static int mgc_target_register(struct obd_export *exp,
1141                                struct mgs_target_info *mti)
1142 {
1143         struct ptlrpc_request  *req;
1144         struct mgs_target_info *req_mti, *rep_mti;
1145         int                     rc;
1146         ENTRY;
1147
1148         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1149                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1150                                         MGS_TARGET_REG);
1151         if (req == NULL)
1152                 RETURN(-ENOMEM);
1153
1154         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1155         if (!req_mti) {
1156                 ptlrpc_req_finished(req);
1157                 RETURN(-ENOMEM);
1158         }
1159
1160         memcpy(req_mti, mti, sizeof(*req_mti));
1161         ptlrpc_request_set_replen(req);
1162         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1163         /* Limit how long we will wait for the enqueue to complete */
1164         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1165
1166         rc = ptlrpc_queue_wait(req);
1167         if (!rc) {
1168                 rep_mti = req_capsule_server_get(&req->rq_pill,
1169                                                  &RMF_MGS_TARGET_INFO);
1170                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1171                 CDEBUG(D_MGC, "register %s got index = %d\n",
1172                        mti->mti_svname, mti->mti_stripe_index);
1173         }
1174         ptlrpc_req_finished(req);
1175
1176         RETURN(rc);
1177 }
1178
1179 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1180                               u32 keylen, void *key,
1181                               u32 vallen, void *val,
1182                               struct ptlrpc_request_set *set)
1183 {
1184         int rc = -EINVAL;
1185         ENTRY;
1186
1187         /* Turn off initial_recov after we try all backup servers once */
1188         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1189                 struct obd_import *imp = class_exp2cliimp(exp);
1190                 int value;
1191                 if (vallen != sizeof(int))
1192                         RETURN(-EINVAL);
1193                 value = *(int *)val;
1194                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1195                        imp->imp_obd->obd_name, value,
1196                        imp->imp_deactive, imp->imp_invalid,
1197                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1198                        ptlrpc_import_state_name(imp->imp_state));
1199                 /* Resurrect if we previously died */
1200                 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1201                      imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1202                         ptlrpc_reconnect_import(imp);
1203                 RETURN(0);
1204         }
1205         /* FIXME move this to mgc_process_config */
1206         if (KEY_IS(KEY_REGISTER_TARGET)) {
1207                 struct mgs_target_info *mti;
1208                 if (vallen != sizeof(struct mgs_target_info))
1209                         RETURN(-EINVAL);
1210                 mti = (struct mgs_target_info *)val;
1211                 CDEBUG(D_MGC, "register_target %s %#x\n",
1212                        mti->mti_svname, mti->mti_flags);
1213                 rc =  mgc_target_register(exp, mti);
1214                 RETURN(rc);
1215         }
1216         if (KEY_IS(KEY_SET_FS)) {
1217                 struct super_block *sb = (struct super_block *)val;
1218
1219                 if (vallen != sizeof(struct super_block))
1220                         RETURN(-EINVAL);
1221
1222                 rc = mgc_fs_setup(exp->exp_obd, sb);
1223                 RETURN(rc);
1224         }
1225         if (KEY_IS(KEY_CLEAR_FS)) {
1226                 if (vallen != 0)
1227                         RETURN(-EINVAL);
1228                 rc = mgc_fs_cleanup(exp->exp_obd);
1229                 RETURN(rc);
1230         }
1231         if (KEY_IS(KEY_SET_INFO)) {
1232                 struct mgs_send_param *msp;
1233
1234                 msp = (struct mgs_send_param *)val;
1235                 rc =  mgc_set_mgs_param(exp, msp);
1236                 RETURN(rc);
1237         }
1238         if (KEY_IS(KEY_MGSSEC)) {
1239                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1240                 struct sptlrpc_flavor  flvr;
1241
1242                 /*
1243                  * empty string means using current flavor, if which haven't
1244                  * been set yet, set it as null.
1245                  *
1246                  * if flavor has been set previously, check the asking flavor
1247                  * must match the existing one.
1248                  */
1249                 if (vallen == 0) {
1250                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1251                                 RETURN(0);
1252                         val = "null";
1253                         vallen = 4;
1254                 }
1255
1256                 rc = sptlrpc_parse_flavor(val, &flvr);
1257                 if (rc) {
1258                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1259                                (char *) val);
1260                         RETURN(rc);
1261                 }
1262
1263                 /*
1264                  * caller already hold a mutex
1265                  */
1266                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1267                         cli->cl_flvr_mgc = flvr;
1268                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1269                                   sizeof(flvr)) != 0) {
1270                         char    str[20];
1271
1272                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1273                                             str, sizeof(str));
1274                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1275                                        "currently %s is in use\n",
1276                                        (char *) val, str);
1277                         rc = -EPERM;
1278                 }
1279                 RETURN(rc);
1280         }
1281
1282         RETURN(rc);
1283 }
1284
1285 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1286                         __u32 keylen, void *key, __u32 *vallen, void *val)
1287 {
1288         int rc = -EINVAL;
1289
1290         if (KEY_IS(KEY_CONN_DATA)) {
1291                 struct obd_import *imp = class_exp2cliimp(exp);
1292                 struct obd_connect_data *data = val;
1293
1294                 if (*vallen == sizeof(*data)) {
1295                         *data = imp->imp_connect_data;
1296                         rc = 0;
1297                 }
1298         }
1299
1300         return rc;
1301 }
1302
1303 static int mgc_import_event(struct obd_device *obd,
1304                             struct obd_import *imp,
1305                             enum obd_import_event event)
1306 {
1307         int rc = 0;
1308
1309         LASSERT(imp->imp_obd == obd);
1310         CDEBUG(D_MGC, "import event %#x\n", event);
1311
1312         switch (event) {
1313         case IMP_EVENT_DISCON:
1314                 /* MGC imports should not wait for recovery */
1315                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1316                         ptlrpc_pinger_ir_down();
1317                 break;
1318         case IMP_EVENT_INACTIVE:
1319                 break;
1320         case IMP_EVENT_INVALIDATE: {
1321                 struct ldlm_namespace *ns = obd->obd_namespace;
1322                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1323                 break;
1324         }
1325         case IMP_EVENT_ACTIVE:
1326                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1327                 /* Clearing obd_no_recov allows us to continue pinging */
1328                 obd->obd_no_recov = 0;
1329                 mgc_notify_active(obd);
1330                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1331                         ptlrpc_pinger_ir_up();
1332                 break;
1333         case IMP_EVENT_OCD:
1334                 break;
1335         case IMP_EVENT_DEACTIVATE:
1336         case IMP_EVENT_ACTIVATE:
1337                 break;
1338         default:
1339                 CERROR("Unknown import event %#x\n", event);
1340                 LBUG();
1341         }
1342         RETURN(rc);
1343 }
1344
1345 enum {
1346         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_CACHE_SHIFT),
1347         CONFIG_READ_NRPAGES      = 4
1348 };
1349
1350 static int mgc_apply_recover_logs(struct obd_device *mgc,
1351                                   struct config_llog_data *cld,
1352                                   __u64 max_version,
1353                                   void *data, int datalen, bool mne_swab)
1354 {
1355         struct config_llog_instance *cfg = &cld->cld_cfg;
1356         struct lustre_sb_info       *lsi = s2lsi(cfg->cfg_sb);
1357         struct mgs_nidtbl_entry *entry;
1358         struct lustre_cfg       *lcfg;
1359         struct lustre_cfg_bufs   bufs;
1360         u64   prev_version = 0;
1361         char *inst;
1362         char *buf;
1363         int   bufsz;
1364         int   pos;
1365         int   rc  = 0;
1366         int   off = 0;
1367         ENTRY;
1368
1369         LASSERT(cfg->cfg_instance != NULL);
1370         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1371
1372         OBD_ALLOC(inst, PAGE_CACHE_SIZE);
1373         if (inst == NULL)
1374                 RETURN(-ENOMEM);
1375
1376         if (!IS_SERVER(lsi)) {
1377                 pos = snprintf(inst, PAGE_CACHE_SIZE, "%p", cfg->cfg_instance);
1378                 if (pos >= PAGE_CACHE_SIZE) {
1379                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1380                         return -E2BIG;
1381                 }
1382         } else {
1383                 LASSERT(IS_MDT(lsi));
1384                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1385                                         PAGE_CACHE_SIZE);
1386                 if (rc) {
1387                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1388                         RETURN(-EINVAL);
1389                 }
1390                 pos = strlen(inst);
1391         }
1392
1393         ++pos;
1394         buf   = inst + pos;
1395         bufsz = PAGE_CACHE_SIZE - pos;
1396
1397         while (datalen > 0) {
1398                 int   entry_len = sizeof(*entry);
1399                 int   is_ost, i;
1400                 struct obd_device *obd;
1401                 char *obdname;
1402                 char *cname;
1403                 char *params;
1404                 char *uuid;
1405
1406                 rc = -EINVAL;
1407                 if (datalen < sizeof(*entry))
1408                         break;
1409
1410                 entry = (typeof(entry))(data + off);
1411
1412                 /* sanity check */
1413                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1414                         break;
1415                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1416                         break;
1417                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1418                         break;
1419
1420                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1421                 if (datalen < entry_len) /* must have entry_len at least */
1422                         break;
1423
1424                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1425                 if (mne_swab)
1426                         lustre_swab_mgs_nidtbl_entry(entry);
1427                 if (entry->mne_length > PAGE_CACHE_SIZE) {
1428                         CERROR("MNE too large (%u)\n", entry->mne_length);
1429                         break;
1430                 }
1431
1432                 if (entry->mne_length < entry_len)
1433                         break;
1434
1435                 off     += entry->mne_length;
1436                 datalen -= entry->mne_length;
1437                 if (datalen < 0)
1438                         break;
1439
1440                 if (entry->mne_version > max_version) {
1441                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1442                                entry->mne_version, max_version);
1443                         break;
1444                 }
1445
1446                 if (prev_version >= entry->mne_version) {
1447                         CERROR("index unsorted, prev %lld, now %lld\n",
1448                                prev_version, entry->mne_version);
1449                         break;
1450                 }
1451                 prev_version = entry->mne_version;
1452
1453                 /*
1454                  * Write a string with format "nid::instance" to
1455                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1456                  */
1457
1458                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1459                 memset(buf, 0, bufsz);
1460                 obdname = buf;
1461                 pos = 0;
1462
1463                 /* lustre-OST0001-osc-<instance #> */
1464                 strcpy(obdname, cld->cld_logname);
1465                 cname = strrchr(obdname, '-');
1466                 if (cname == NULL) {
1467                         CERROR("mgc %s: invalid logname %s\n",
1468                                mgc->obd_name, obdname);
1469                         break;
1470                 }
1471
1472                 pos = cname - obdname;
1473                 obdname[pos] = 0;
1474                 pos += sprintf(obdname + pos, "-%s%04x",
1475                                   is_ost ? "OST" : "MDT", entry->mne_index);
1476
1477                 cname = is_ost ? "osc" : "mdc",
1478                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1479                 lustre_cfg_bufs_reset(&bufs, obdname);
1480
1481                 /* find the obd by obdname */
1482                 obd = class_name2obd(obdname);
1483                 if (obd == NULL) {
1484                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1485                                mgc->obd_name, obdname);
1486                         rc = 0;
1487                         /* this is a safe race, when the ost is starting up...*/
1488                         continue;
1489                 }
1490
1491                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1492                 ++pos;
1493                 params = buf + pos;
1494                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1495                 uuid = buf + pos;
1496
1497                 down_read(&obd->u.cli.cl_sem);
1498                 if (obd->u.cli.cl_import == NULL) {
1499                         /* client does not connect to the OST yet */
1500                         up_read(&obd->u.cli.cl_sem);
1501                         rc = 0;
1502                         continue;
1503                 }
1504
1505                 /* iterate all nids to find one */
1506                 /* find uuid by nid */
1507                 rc = -ENOENT;
1508                 for (i = 0; i < entry->mne_nid_count; i++) {
1509                         rc = client_import_find_conn(obd->u.cli.cl_import,
1510                                                      entry->u.nids[i],
1511                                                      (struct obd_uuid *)uuid);
1512                         if (rc == 0)
1513                                 break;
1514                 }
1515
1516                 up_read(&obd->u.cli.cl_sem);
1517                 if (rc < 0) {
1518                         CERROR("mgc: cannot find uuid by nid %s\n",
1519                                libcfs_nid2str(entry->u.nids[0]));
1520                         break;
1521                 }
1522
1523                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1524                        uuid, libcfs_nid2str(entry->u.nids[0]));
1525
1526                 pos += strlen(uuid);
1527                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1528                 LASSERT(pos < bufsz);
1529
1530                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1531
1532                 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1533                 if (lcfg == NULL) {
1534                         rc = -ENOMEM;
1535                         break;
1536                 }
1537
1538                 CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n",
1539                        prev_version, max_version, obdname, params);
1540
1541                 rc = class_process_config(lcfg);
1542                 lustre_cfg_free(lcfg);
1543                 if (rc)
1544                         CDEBUG(D_INFO, "process config for %s error %d\n",
1545                                obdname, rc);
1546
1547                 /* continue, even one with error */
1548         }
1549
1550         OBD_FREE(inst, PAGE_CACHE_SIZE);
1551         RETURN(rc);
1552 }
1553
1554 /**
1555  * This function is called if this client was notified for target restarting
1556  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1557  */
1558 static int mgc_process_recover_log(struct obd_device *obd,
1559                                    struct config_llog_data *cld)
1560 {
1561         struct ptlrpc_request *req = NULL;
1562         struct config_llog_instance *cfg = &cld->cld_cfg;
1563         struct mgs_config_body *body;
1564         struct mgs_config_res  *res;
1565         struct ptlrpc_bulk_desc *desc;
1566         struct page **pages;
1567         int nrpages;
1568         bool eof = true;
1569         bool mne_swab = false;
1570         int i;
1571         int ealen;
1572         int rc;
1573         ENTRY;
1574
1575         /* allocate buffer for bulk transfer.
1576          * if this is the first time for this mgs to read logs,
1577          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1578          * once; otherwise, it only reads increment of logs, this should be
1579          * small and CONFIG_READ_NRPAGES will be used.
1580          */
1581         nrpages = CONFIG_READ_NRPAGES;
1582         if (cfg->cfg_last_idx == 0) /* the first time */
1583                 nrpages = CONFIG_READ_NRPAGES_INIT;
1584
1585         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1586         if (pages == NULL)
1587                 GOTO(out, rc = -ENOMEM);
1588
1589         for (i = 0; i < nrpages; i++) {
1590                 pages[i] = alloc_page(GFP_IOFS);
1591                 if (pages[i] == NULL)
1592                         GOTO(out, rc = -ENOMEM);
1593         }
1594
1595 again:
1596         LASSERT(cld_is_recover(cld));
1597         LASSERT(mutex_is_locked(&cld->cld_lock));
1598         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1599                                    &RQF_MGS_CONFIG_READ);
1600         if (req == NULL)
1601                 GOTO(out, rc = -ENOMEM);
1602
1603         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1604         if (rc)
1605                 GOTO(out, rc);
1606
1607         /* pack request */
1608         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1609         LASSERT(body != NULL);
1610         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1611         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1612             >= sizeof(body->mcb_name))
1613                 GOTO(out, rc = -E2BIG);
1614         body->mcb_offset = cfg->cfg_last_idx + 1;
1615         body->mcb_type   = cld->cld_type;
1616         body->mcb_bits   = PAGE_CACHE_SHIFT;
1617         body->mcb_units  = nrpages;
1618
1619         /* allocate bulk transfer descriptor */
1620         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1, BULK_PUT_SINK,
1621                                     MGS_BULK_PORTAL);
1622         if (desc == NULL)
1623                 GOTO(out, rc = -ENOMEM);
1624
1625         for (i = 0; i < nrpages; i++)
1626                 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
1627
1628         ptlrpc_request_set_replen(req);
1629         rc = ptlrpc_queue_wait(req);
1630         if (rc)
1631                 GOTO(out, rc);
1632
1633         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1634         if (res->mcr_size < res->mcr_offset)
1635                 GOTO(out, rc = -EINVAL);
1636
1637         /* always update the index even though it might have errors with
1638          * handling the recover logs */
1639         cfg->cfg_last_idx = res->mcr_offset;
1640         eof = res->mcr_offset == res->mcr_size;
1641
1642         CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n",
1643                res->mcr_offset, eof == false);
1644
1645         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1646         if (ealen < 0)
1647                 GOTO(out, rc = ealen);
1648
1649         if (ealen > nrpages << PAGE_CACHE_SHIFT)
1650                 GOTO(out, rc = -EINVAL);
1651
1652         if (ealen == 0) { /* no logs transferred */
1653                 if (!eof)
1654                         rc = -EINVAL;
1655                 GOTO(out, rc);
1656         }
1657
1658         mne_swab = !!ptlrpc_rep_need_swab(req);
1659 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1660         /* This import flag means the server did an extra swab of IR MNE
1661          * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1662         if (unlikely(req->rq_import->imp_need_mne_swab))
1663                 mne_swab = !mne_swab;
1664 #endif
1665
1666         for (i = 0; i < nrpages && ealen > 0; i++) {
1667                 int rc2;
1668                 void *ptr;
1669
1670                 ptr = kmap(pages[i]);
1671                 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1672                                              min_t(int, ealen, PAGE_CACHE_SIZE),
1673                                              mne_swab);
1674                 kunmap(pages[i]);
1675                 if (rc2 < 0) {
1676                         CWARN("Process recover log %s error %d\n",
1677                               cld->cld_logname, rc2);
1678                         break;
1679                 }
1680
1681                 ealen -= PAGE_CACHE_SIZE;
1682         }
1683
1684 out:
1685         if (req)
1686                 ptlrpc_req_finished(req);
1687
1688         if (rc == 0 && !eof)
1689                 goto again;
1690
1691         if (pages) {
1692                 for (i = 0; i < nrpages; i++) {
1693                         if (pages[i] == NULL)
1694                                 break;
1695                         __free_page(pages[i]);
1696                 }
1697                 OBD_FREE(pages, sizeof(*pages) * nrpages);
1698         }
1699         return rc;
1700 }
1701
1702 /* Copy a remote log locally */
1703 static int mgc_llog_local_copy(const struct lu_env *env,
1704                                struct obd_device *obd,
1705                                struct llog_ctxt *rctxt,
1706                                struct llog_ctxt *lctxt, char *logname)
1707 {
1708         char    *temp_log;
1709         int      rc;
1710
1711         ENTRY;
1712
1713         /*
1714          * - copy it to backup using llog_backup()
1715          * - copy remote llog to logname using llog_backup()
1716          * - if failed then move bakup to logname again
1717          */
1718
1719         OBD_ALLOC(temp_log, strlen(logname) + 2);
1720         if (!temp_log)
1721                 RETURN(-ENOMEM);
1722         sprintf(temp_log, "%sT", logname);
1723
1724         /* make a copy of local llog at first */
1725         rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1726         if (rc < 0 && rc != -ENOENT)
1727                 GOTO(out, rc);
1728         /* copy remote llog to the local copy */
1729         rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1730         if (rc == -ENOENT) {
1731                 /* no remote llog, delete local one too */
1732                 llog_erase(env, lctxt, NULL, logname);
1733         } else if (rc < 0) {
1734                 /* error during backup, get local one back from the copy */
1735                 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1736 out:
1737                 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1738                        obd->obd_name, logname, rc);
1739         }
1740         llog_erase(env, lctxt, NULL, temp_log);
1741         OBD_FREE(temp_log, strlen(logname) + 2);
1742         return rc;
1743 }
1744
1745 /* local_only means it cannot get remote llogs */
1746 static int mgc_process_cfg_log(struct obd_device *mgc,
1747                                struct config_llog_data *cld, int local_only)
1748 {
1749         struct llog_ctxt        *ctxt, *lctxt = NULL;
1750         struct client_obd       *cli = &mgc->u.cli;
1751         struct lustre_sb_info   *lsi = NULL;
1752         int                      rc = 0;
1753         bool                     sptlrpc_started = false;
1754         struct lu_env           *env;
1755
1756         ENTRY;
1757
1758         LASSERT(cld);
1759         LASSERT(mutex_is_locked(&cld->cld_lock));
1760
1761         /*
1762          * local copy of sptlrpc log is controlled elsewhere, don't try to
1763          * read it up here.
1764          */
1765         if (cld_is_sptlrpc(cld) && local_only)
1766                 RETURN(0);
1767
1768         if (cld->cld_cfg.cfg_sb)
1769                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1770
1771         OBD_ALLOC_PTR(env);
1772         if (env == NULL)
1773                 RETURN(-ENOMEM);
1774
1775         rc = lu_env_init(env, LCT_MG_THREAD);
1776         if (rc)
1777                 GOTO(out_free, rc);
1778
1779         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1780         LASSERT(ctxt);
1781
1782         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1783
1784         /* Copy the setup log locally if we can. Don't mess around if we're
1785          * running an MGS though (logs are already local). */
1786         if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1787             cli->cl_mgc_configs_dir != NULL &&
1788             lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1789             lsi->lsi_dt_dev) {
1790                 if (!local_only)
1791                         /* Only try to copy log if we have the lock. */
1792                         rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1793                                                  cld->cld_logname);
1794                 if (local_only || rc) {
1795                         if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1796                             llog_is_empty(env, lctxt, cld->cld_logname)) {
1797                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1798                                                    "log %s and no local copy."
1799                                                    "\n", cld->cld_logname);
1800                                 GOTO(out_pop, rc = -ENOENT);
1801                         }
1802                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1803                                "copy for now, will try to update later.\n",
1804                                cld->cld_logname);
1805                         rc = 0;
1806                 }
1807                 /* Now, whether we copied or not, start using the local llog.
1808                  * If we failed to copy, we'll start using whatever the old
1809                  * log has. */
1810                 llog_ctxt_put(ctxt);
1811                 ctxt = lctxt;
1812                 lctxt = NULL;
1813         } else {
1814                 if (local_only) /* no local log at client side */
1815                         GOTO(out_pop, rc = -EIO);
1816         }
1817
1818         if (cld_is_sptlrpc(cld)) {
1819                 sptlrpc_conf_log_update_begin(cld->cld_logname);
1820                 sptlrpc_started = true;
1821         }
1822
1823         /* logname and instance info should be the same, so use our
1824          * copy of the instance for the update.  The cfg_last_idx will
1825          * be updated here. */
1826         rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1827                                      &cld->cld_cfg);
1828         EXIT;
1829
1830 out_pop:
1831         __llog_ctxt_put(env, ctxt);
1832         if (lctxt)
1833                 __llog_ctxt_put(env, lctxt);
1834
1835         /*
1836          * update settings on existing OBDs. doing it inside
1837          * of llog_process_lock so no device is attaching/detaching
1838          * in parallel.
1839          * the logname must be <fsname>-sptlrpc
1840          */
1841         if (sptlrpc_started) {
1842                 LASSERT(cld_is_sptlrpc(cld));
1843                 sptlrpc_conf_log_update_end(cld->cld_logname);
1844                 class_notify_sptlrpc_conf(cld->cld_logname,
1845                                           strlen(cld->cld_logname) -
1846                                           strlen("-sptlrpc"));
1847         }
1848
1849         lu_env_fini(env);
1850 out_free:
1851         OBD_FREE_PTR(env);
1852         return rc;
1853 }
1854
1855 static bool mgc_import_in_recovery(struct obd_import *imp)
1856 {
1857         bool in_recovery = true;
1858
1859         spin_lock(&imp->imp_lock);
1860         if (imp->imp_state == LUSTRE_IMP_FULL ||
1861             imp->imp_state == LUSTRE_IMP_CLOSED)
1862                 in_recovery = false;
1863         spin_unlock(&imp->imp_lock);
1864
1865         return in_recovery;
1866 }
1867
1868 /**
1869  * Get a configuration log from the MGS and process it.
1870  *
1871  * This function is called for both clients and servers to process the
1872  * configuration log from the MGS.  The MGC enqueues a DLM lock on the
1873  * log from the MGS, and if the lock gets revoked the MGC will be notified
1874  * by the lock cancellation callback that the config log has changed,
1875  * and will enqueue another MGS lock on it, and then continue processing
1876  * the new additions to the end of the log.
1877  *
1878  * Since the MGC import is not replayable, if the import is being evicted
1879  * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
1880  * the log until recovery is finished or the import is closed.
1881  *
1882  * Make a local copy of the log before parsing it if appropriate (non-MGS
1883  * server) so that the server can start even when the MGS is down.
1884  *
1885  * There shouldn't be multiple processes running process_log at once --
1886  * sounds like badness.  It actually might be fine, as long as they're not
1887  * trying to update from the same log simultaneously, in which case we
1888  * should use a per-log semaphore instead of cld_lock.
1889  *
1890  * \param[in] mgc       MGC device by which to fetch the configuration log
1891  * \param[in] cld       log processing state (stored in lock callback data)
1892  *
1893  * \retval              0 on success
1894  * \retval              negative errno on failure
1895  */
1896 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1897 {
1898         struct lustre_handle lockh = { 0 };
1899         __u64 flags = LDLM_FL_NO_LRU;
1900         int rc = 0, rcl;
1901         bool retry = false;
1902         ENTRY;
1903
1904         LASSERT(cld != NULL);
1905
1906         /* I don't want multiple processes running process_log at once --
1907            sounds like badness.  It actually might be fine, as long as
1908            we're not trying to update from the same log
1909            simultaneously (in which case we should use a per-log sem.) */
1910 restart:
1911         mutex_lock(&cld->cld_lock);
1912         if (cld->cld_stopping) {
1913                 mutex_unlock(&cld->cld_lock);
1914                 RETURN(0);
1915         }
1916
1917         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1918
1919         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1920                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1921
1922         /* Get the cfg lock on the llog */
1923         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
1924                           LCK_CR, &flags, NULL, NULL, NULL,
1925                           cld, 0, NULL, &lockh);
1926         if (rcl == 0) {
1927                 /* Get the cld, it will be released in mgc_blocking_ast. */
1928                 config_log_get(cld);
1929                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1930                 LASSERT(rc == 0);
1931         } else {
1932                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1933
1934                 if (rcl == -ESHUTDOWN &&
1935                     atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
1936                         struct obd_import *imp;
1937                         struct l_wait_info lwi;
1938                         int secs = cfs_time_seconds(obd_timeout);
1939
1940                         mutex_unlock(&cld->cld_lock);
1941                         imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
1942
1943                         /* Let's force the pinger, and wait the import to be
1944                          * connected, note: since mgc import is non-replayable,
1945                          * and even the import state is disconnected, it does
1946                          * not mean the "recovery" is stopped, so we will keep
1947                          * waitting until timeout or the import state is
1948                          * FULL or closed */
1949                         ptlrpc_pinger_force(imp);
1950
1951                         lwi = LWI_TIMEOUT(secs, NULL, NULL);
1952                         l_wait_event(imp->imp_recovery_waitq,
1953                                      !mgc_import_in_recovery(imp), &lwi);
1954
1955                         if (imp->imp_state == LUSTRE_IMP_FULL) {
1956                                 retry = true;
1957                                 goto restart;
1958                         } else {
1959                                 mutex_lock(&cld->cld_lock);
1960                                 cld->cld_lostlock = 1;
1961                         }
1962                 } else {
1963                         /* mark cld_lostlock so that it will requeue
1964                          * after MGC becomes available. */
1965                         cld->cld_lostlock = 1;
1966                 }
1967                 /* Get extra reference, it will be put in requeue thread */
1968                 config_log_get(cld);
1969         }
1970
1971
1972         if (cld_is_recover(cld)) {
1973                 rc = 0; /* this is not a fatal error for recover log */
1974                 if (rcl == 0)
1975                         rc = mgc_process_recover_log(mgc, cld);
1976         } else {
1977                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1978         }
1979
1980         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1981                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1982
1983         mutex_unlock(&cld->cld_lock);
1984
1985         /* Now drop the lock so MGS can revoke it */
1986         if (!rcl) {
1987                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
1988                 if (rcl)
1989                         CERROR("Can't drop cfg lock: %d\n", rcl);
1990         }
1991
1992         RETURN(rc);
1993 }
1994
1995
1996 /** Called from lustre_process_log.
1997  * LCFG_LOG_START gets the config log from the MGS, processes it to start
1998  * any services, and adds it to the list logs to watch (follow).
1999  */
2000 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2001 {
2002         struct lustre_cfg *lcfg = buf;
2003         struct config_llog_instance *cfg = NULL;
2004         char *logname;
2005         int rc = 0;
2006         ENTRY;
2007
2008         switch(lcfg->lcfg_command) {
2009         case LCFG_LOV_ADD_OBD: {
2010                 /* Overloading this cfg command: register a new target */
2011                 struct mgs_target_info *mti;
2012
2013                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2014                     sizeof(struct mgs_target_info))
2015                         GOTO(out, rc = -EINVAL);
2016
2017                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2018                 CDEBUG(D_MGC, "add_target %s %#x\n",
2019                        mti->mti_svname, mti->mti_flags);
2020                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2021                 break;
2022         }
2023         case LCFG_LOV_DEL_OBD:
2024                 /* Unregister has no meaning at the moment. */
2025                 CERROR("lov_del_obd unimplemented\n");
2026                 rc = -ENOSYS;
2027                 break;
2028         case LCFG_SPTLRPC_CONF: {
2029                 rc = sptlrpc_process_config(lcfg);
2030                 break;
2031         }
2032         case LCFG_LOG_START: {
2033                 struct config_llog_data *cld;
2034                 struct super_block *sb;
2035
2036                 logname = lustre_cfg_string(lcfg, 1);
2037                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2038                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2039
2040                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2041                        cfg->cfg_last_idx);
2042
2043                 /* We're only called through here on the initial mount */
2044                 rc = config_log_add(obd, logname, cfg, sb);
2045                 if (rc)
2046                         break;
2047                 cld = config_log_find(logname, cfg);
2048                 if (cld == NULL) {
2049                         rc = -ENOENT;
2050                         break;
2051                 }
2052
2053                 /* COMPAT_146 */
2054                 /* FIXME only set this for old logs!  Right now this forces
2055                    us to always skip the "inside markers" check */
2056                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
2057
2058                 rc = mgc_process_log(obd, cld);
2059                 if (rc == 0 && cld->cld_recover != NULL) {
2060                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2061                                          imp_connect_data, IMP_RECOV)) {
2062                                 rc = mgc_process_log(obd, cld->cld_recover);
2063                         } else {
2064                                 struct config_llog_data *cir = cld->cld_recover;
2065                                 cld->cld_recover = NULL;
2066                                 config_log_put(cir);
2067                         }
2068                         if (rc)
2069                                 CERROR("Cannot process recover llog %d\n", rc);
2070                 }
2071
2072                 if (rc == 0 && cld->cld_params != NULL) {
2073                         rc = mgc_process_log(obd, cld->cld_params);
2074                         if (rc == -ENOENT) {
2075                                 CDEBUG(D_MGC, "There is no params "
2076                                               "config file yet\n");
2077                                 rc = 0;
2078                         }
2079                         /* params log is optional */
2080                         if (rc)
2081                                 CERROR("%s: can't process params llog: rc = %d\n",
2082                                        obd->obd_name, rc);
2083                 }
2084                 config_log_put(cld);
2085
2086                 break;
2087         }
2088         case LCFG_LOG_END: {
2089                 logname = lustre_cfg_string(lcfg, 1);
2090
2091                 if (lcfg->lcfg_bufcount >= 2)
2092                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
2093                                 lcfg, 2);
2094                 rc = config_log_end(logname, cfg);
2095                 break;
2096         }
2097         default: {
2098                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2099                 GOTO(out, rc = -EINVAL);
2100
2101         }
2102         }
2103 out:
2104         RETURN(rc);
2105 }
2106
2107 static struct obd_ops mgc_obd_ops = {
2108         .o_owner        = THIS_MODULE,
2109         .o_setup        = mgc_setup,
2110         .o_precleanup   = mgc_precleanup,
2111         .o_cleanup      = mgc_cleanup,
2112         .o_add_conn     = client_import_add_conn,
2113         .o_del_conn     = client_import_del_conn,
2114         .o_connect      = client_connect_import,
2115         .o_disconnect   = client_disconnect_export,
2116         .o_set_info_async = mgc_set_info_async,
2117         .o_get_info       = mgc_get_info,
2118         .o_import_event = mgc_import_event,
2119         .o_process_config = mgc_process_config,
2120 };
2121
2122 static int __init mgc_init(void)
2123 {
2124         return class_register_type(&mgc_obd_ops, NULL, true, NULL,
2125                                    LUSTRE_MGC_NAME, NULL);
2126 }
2127
2128 static void /*__exit*/ mgc_exit(void)
2129 {
2130         class_unregister_type(LUSTRE_MGC_NAME);
2131 }
2132
2133 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2134 MODULE_DESCRIPTION("Lustre Management Client");
2135 MODULE_LICENSE("GPL");
2136
2137 module_init(mgc_init);
2138 module_exit(mgc_exit);