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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/mgc/mgc_request.c
  33  *
  34  * Author: Nathan Rutman <nathan@clusterfs.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_MGC
  38 #define D_MGC D_CONFIG /*|D_WARNING*/
  39
  40 #include <linux/module.h>
  41 #include <linux/kthread.h>
  42 #include <linux/random.h>
  43
  44 #include <dt_object.h>
  45 #include <lprocfs_status.h>
  46 #include <lustre_dlm.h>
  47 #include <lustre_disk.h>
  48 #include <lustre_log.h>
  49 #include <lustre_nodemap.h>
  50 #include <lustre_swab.h>
  51 #include <obd_class.h>
  52 #include <lustre_barrier.h>
  53
  54 #include "mgc_internal.h"
  55
  56 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
  57                           int type)
  58 {
  59         __u64 resname = 0;
  60
  61         if (len > sizeof(resname)) {
  62                 CERROR("name too long: %s\n", name);
  63                 return -EINVAL;
  64         }
  65         if (len <= 0) {
  66                 CERROR("missing name: %s\n", name);
  67                 return -EINVAL;
  68         }
  69         memcpy(&resname, name, len);
  70
  71         /* Always use the same endianness for the resid */
  72         memset(res_id, 0, sizeof(*res_id));
  73         res_id->name[0] = cpu_to_le64(resname);
  74         /* XXX: unfortunately, sptlprc and config llog share one lock */
  75         switch(type) {
  76         case CONFIG_T_CONFIG:
  77         case CONFIG_T_SPTLRPC:
  78                 resname = 0;
  79                 break;
  80         case CONFIG_T_RECOVER:
  81         case CONFIG_T_PARAMS:
  82         case CONFIG_T_NODEMAP:
  83         case CONFIG_T_BARRIER:
  84                 resname = type;
  85                 break;
  86         default:
  87                 LBUG();
  88         }
  89         res_id->name[1] = cpu_to_le64(resname);
  90         CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
  91                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  92         return 0;
  93 }
  94
  95 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
  96 {
  97         /* fsname is at most 8 chars long, maybe contain "-".
  98          * e.g. "lustre", "SUN-000" */
  99         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
 100 }
 101 EXPORT_SYMBOL(mgc_fsname2resid);
 102
 103 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
 104 {
 105         char *name_end;
 106         int len;
 107
 108         /* logname consists of "fsname-nodetype".
 109          * e.g. "lustre-MDT0001", "SUN-000-client"
 110          * there is an exception: llog "params" */
 111         name_end = strrchr(logname, '-');
 112         if (!name_end)
 113                 len = strlen(logname);
 114         else
 115                 len = name_end - logname;
 116         return mgc_name2resid(logname, len, res_id, type);
 117 }
 118 EXPORT_SYMBOL(mgc_logname2resid);
 119
 120 /********************** config llog list **********************/
 121 static LIST_HEAD(config_llog_list);
 122 static DEFINE_SPINLOCK(config_list_lock);       /* protects config_llog_list */
 123
 124 /* Take a reference to a config log */
 125 static int config_log_get(struct config_llog_data *cld)
 126 {
 127         ENTRY;
 128         atomic_inc(&cld->cld_refcount);
 129         CDEBUG(D_INFO, "log %s (%p) refs %d\n", cld->cld_logname, cld,
 130                 atomic_read(&cld->cld_refcount));
 131         RETURN(0);
 132 }
 133
 134 /* Drop a reference to a config log.  When no longer referenced,
 135    we can free the config log data */
 136 static void config_log_put(struct config_llog_data *cld)
 137 {
 138         ENTRY;
 139
 140         if (unlikely(!cld))
 141                 RETURN_EXIT;
 142
 143         CDEBUG(D_INFO, "log %s(%p) refs %d\n", cld->cld_logname, cld,
 144                 atomic_read(&cld->cld_refcount));
 145         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 146
 147         /* spinlock to make sure no item with 0 refcount in the list */
 148         if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 149                 list_del(&cld->cld_list_chain);
 150                 spin_unlock(&config_list_lock);
 151
 152                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 153
 154                 config_log_put(cld->cld_barrier);
 155                 config_log_put(cld->cld_recover);
 156                 config_log_put(cld->cld_params);
 157                 config_log_put(cld->cld_nodemap);
 158                 config_log_put(cld->cld_sptlrpc);
 159                 if (cld_is_sptlrpc(cld)) {
 160                         cld->cld_stopping = 1;
 161                         sptlrpc_conf_log_stop(cld->cld_logname);
 162                 }
 163
 164                 class_export_put(cld->cld_mgcexp);
 165                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 166         }
 167
 168         EXIT;
 169 }
 170
 171 /* Find a config log by name */
 172 static
 173 struct config_llog_data *config_log_find(char *logname,
 174                                          struct config_llog_instance *cfg)
 175 {
 176         struct config_llog_data *cld;
 177         struct config_llog_data *found = NULL;
 178         unsigned long cfg_instance;
 179
 180         ENTRY;
 181         LASSERT(logname != NULL);
 182
 183         cfg_instance = cfg ? cfg->cfg_instance : 0;
 184         spin_lock(&config_list_lock);
 185         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 186                 /* check if cfg_instance is the one we want */
 187                 if (cfg_instance != cld->cld_cfg.cfg_instance)
 188                         continue;
 189
 190                 /* instance may be NULL, should check name */
 191                 if (strcmp(logname, cld->cld_logname) == 0) {
 192                         found = cld;
 193                         config_log_get(found);
 194                         break;
 195                 }
 196         }
 197         spin_unlock(&config_list_lock);
 198         RETURN(found);
 199 }
 200
 201 static
 202 struct config_llog_data *do_config_log_add(struct obd_device *obd,
 203                                            char *logname,
 204                                            int type,
 205                                            struct config_llog_instance *cfg,
 206                                            struct super_block *sb)
 207 {
 208         struct config_llog_data *cld;
 209         int rc;
 210
 211         ENTRY;
 212
 213         CDEBUG(D_MGC, "do adding config log %s-%016lx\n", logname,
 214                cfg ? cfg->cfg_instance : 0);
 215
 216         OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
 217         if (!cld)
 218                 RETURN(ERR_PTR(-ENOMEM));
 219
 220         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 221         if (rc) {
 222                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 223                 RETURN(ERR_PTR(rc));
 224         }
 225
 226         strcpy(cld->cld_logname, logname);
 227         if (cfg)
 228                 cld->cld_cfg = *cfg;
 229         else
 230                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 231         mutex_init(&cld->cld_lock);
 232         cld->cld_cfg.cfg_last_idx = 0;
 233         cld->cld_cfg.cfg_flags = 0;
 234         cld->cld_cfg.cfg_sb = sb;
 235         cld->cld_type = type;
 236         atomic_set(&cld->cld_refcount, 1);
 237
 238         /* Keep the mgc around until we are done */
 239         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 240
 241         if (cld_is_sptlrpc(cld))
 242                 sptlrpc_conf_log_start(logname);
 243
 244         spin_lock(&config_list_lock);
 245         list_add(&cld->cld_list_chain, &config_llog_list);
 246         spin_unlock(&config_list_lock);
 247
 248         if (cld_is_sptlrpc(cld) || cld_is_nodemap(cld) || cld_is_barrier(cld)) {
 249                 rc = mgc_process_log(obd, cld);
 250                 if (rc && rc != -ENOENT)
 251                         CERROR("%s: failed processing log, type %d: rc = %d\n",
 252                                obd->obd_name, type, rc);
 253         }
 254
 255         RETURN(cld);
 256 }
 257
 258 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
 259                                         char *fsname,
 260                                         struct config_llog_instance *cfg,
 261                                         struct super_block *sb)
 262 {
 263         struct config_llog_instance lcfg = *cfg;
 264         struct lustre_sb_info *lsi = s2lsi(sb);
 265         struct config_llog_data *cld;
 266         char logname[32];
 267
 268         if (IS_OST(lsi))
 269                 return NULL;
 270
 271         /* for osp-on-ost, see lustre_start_osp() */
 272         if (IS_MDT(lsi) && lcfg.cfg_instance)
 273                 return NULL;
 274
 275         /* We have to use different llog for clients and MDTs for DNE,
 276          * where only clients are notified if one of DNE server restarts.
 277          */
 278         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 279         strncpy(logname, fsname, sizeof(logname));
 280         if (IS_SERVER(lsi)) { /* mdt */
 281                 LASSERT(lcfg.cfg_instance == 0);
 282                 lcfg.cfg_instance = ll_get_cfg_instance(sb);
 283                 strncat(logname, "-mdtir", sizeof(logname));
 284         } else {
 285                 LASSERT(lcfg.cfg_instance != 0);
 286                 strncat(logname, "-cliir", sizeof(logname));
 287         }
 288
 289         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
 290         return cld;
 291 }
 292
 293 static struct config_llog_data *config_log_find_or_add(struct obd_device *obd,
 294                                 char *logname, struct super_block *sb, int type,
 295                                 struct config_llog_instance *cfg)
 296 {
 297         struct config_llog_instance lcfg = *cfg;
 298         struct config_llog_data *cld;
 299
 300         /* Note class_config_llog_handler() depends on getting "obd" back */
 301         lcfg.cfg_instance = sb ? ll_get_cfg_instance(sb) : (unsigned long)obd;
 302
 303         cld = config_log_find(logname, &lcfg);
 304         if (unlikely(cld != NULL))
 305                 return cld;
 306
 307         return do_config_log_add(obd, logname, type, &lcfg, sb);
 308 }
 309
 310 /** Add this log to the list of active logs watched by an MGC.
 311  * Active means we're watching for updates.
 312  * We have one active log per "mount" - client instance or servername.
 313  * Each instance may be at a different point in the log.
 314  */
 315 static struct config_llog_data *
 316 config_log_add(struct obd_device *obd, char *logname,
 317                struct config_llog_instance *cfg, struct super_block *sb)
 318 {
 319         struct lustre_sb_info *lsi = s2lsi(sb);
 320         struct config_llog_data *cld = NULL;
 321         struct config_llog_data *sptlrpc_cld = NULL;
 322         struct config_llog_data *params_cld = NULL;
 323         struct config_llog_data *nodemap_cld = NULL;
 324         struct config_llog_data *barrier_cld = NULL;
 325         char seclogname[32];
 326         char *ptr;
 327         int rc;
 328         bool locked = false;
 329         ENTRY;
 330
 331         CDEBUG(D_MGC, "add config log %s-%016lx\n", logname,
 332                cfg->cfg_instance);
 333
 334         /*
 335          * for each regular log, the depended sptlrpc log name is
 336          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 337          */
 338         ptr = strrchr(logname, '-');
 339         if (ptr == NULL || ptr - logname > 8) {
 340                 CERROR("logname %s is too long\n", logname);
 341                 RETURN(ERR_PTR(-EINVAL));
 342         }
 343
 344         memcpy(seclogname, logname, ptr - logname);
 345         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 346
 347         if (cfg->cfg_sub_clds & CONFIG_SUB_SPTLRPC) {
 348                 sptlrpc_cld = config_log_find_or_add(obd, seclogname, NULL,
 349                                                      CONFIG_T_SPTLRPC, cfg);
 350                 if (IS_ERR(sptlrpc_cld)) {
 351                         CERROR("%s: can't create sptlrpc log %s: rc = %ld\n",
 352                                obd->obd_name, seclogname, PTR_ERR(sptlrpc_cld));
 353                         RETURN(sptlrpc_cld);
 354                 }
 355         }
 356
 357         if (!IS_MGS(lsi) && cfg->cfg_sub_clds & CONFIG_SUB_NODEMAP) {
 358                 nodemap_cld = config_log_find_or_add(obd, LUSTRE_NODEMAP_NAME,
 359                                                      NULL, CONFIG_T_NODEMAP,
 360                                                      cfg);
 361                 if (IS_ERR(nodemap_cld)) {
 362                         rc = PTR_ERR(nodemap_cld);
 363                         CERROR("%s: cannot create nodemap log: rc = %d\n",
 364                                obd->obd_name, rc);
 365                         GOTO(out_sptlrpc, rc);
 366                 }
 367         }
 368
 369         if (cfg->cfg_sub_clds & CONFIG_SUB_PARAMS) {
 370                 params_cld = config_log_find_or_add(obd, PARAMS_FILENAME, sb,
 371                                                     CONFIG_T_PARAMS, cfg);
 372                 if (IS_ERR(params_cld)) {
 373                         rc = PTR_ERR(params_cld);
 374                         CERROR("%s: can't create params log: rc = %d\n",
 375                                obd->obd_name, rc);
 376                         GOTO(out_nodemap, rc);
 377                 }
 378         }
 379
 380         if (IS_MDT(s2lsi(sb)) && cfg->cfg_sub_clds & CONFIG_SUB_BARRIER) {
 381                 snprintf(seclogname + (ptr - logname), sizeof(seclogname) - 1,
 382                          "-%s", BARRIER_FILENAME);
 383                 barrier_cld = config_log_find_or_add(obd, seclogname, sb,
 384                                                      CONFIG_T_BARRIER, cfg);
 385                 if (IS_ERR(barrier_cld)) {
 386                         rc = PTR_ERR(barrier_cld);
 387                         CERROR("%s: can't create barrier log: rc = %d\n",
 388                                obd->obd_name, rc);
 389                         GOTO(out_params, rc);
 390                 }
 391         }
 392
 393         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 394         if (IS_ERR(cld)) {
 395                 rc = PTR_ERR(cld);
 396                 CERROR("%s: can't create log: rc = %d\n",
 397                        obd->obd_name, rc);
 398                 GOTO(out_barrier, rc = PTR_ERR(cld));
 399         }
 400
 401         LASSERT(lsi->lsi_lmd);
 402         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR) &&
 403             cfg->cfg_sub_clds & CONFIG_SUB_RECOVER) {
 404                 struct config_llog_data *recover_cld;
 405
 406                 ptr = strrchr(seclogname, '-');
 407                 if (ptr != NULL) {
 408                         *ptr = 0;
 409                 } else {
 410                         CERROR("%s: sptlrpc log name not correct, %s: "
 411                                "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
 412                         GOTO(out_cld, rc = -EINVAL);
 413                 }
 414
 415                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 416                 if (IS_ERR(recover_cld)) {
 417                         rc = PTR_ERR(recover_cld);
 418                         CERROR("%s: can't create recover log: rc = %d\n",
 419                                obd->obd_name, rc);
 420                         GOTO(out_cld, rc);
 421                 }
 422
 423                 mutex_lock(&cld->cld_lock);
 424                 locked = true;
 425                 cld->cld_recover = recover_cld;
 426         }
 427
 428         if (!locked)
 429                 mutex_lock(&cld->cld_lock);
 430         cld->cld_params = params_cld;
 431         cld->cld_barrier = barrier_cld;
 432         cld->cld_nodemap = nodemap_cld;
 433         cld->cld_sptlrpc = sptlrpc_cld;
 434         mutex_unlock(&cld->cld_lock);
 435
 436         RETURN(cld);
 437
 438 out_cld:
 439         config_log_put(cld);
 440 out_barrier:
 441         config_log_put(barrier_cld);
 442 out_params:
 443         config_log_put(params_cld);
 444 out_nodemap:
 445         config_log_put(nodemap_cld);
 446 out_sptlrpc:
 447         config_log_put(sptlrpc_cld);
 448
 449         return ERR_PTR(rc);
 450 }
 451
 452 DEFINE_MUTEX(llog_process_lock);
 453
 454 static inline void config_mark_cld_stop_nolock(struct config_llog_data *cld)
 455 {
 456         ENTRY;
 457
 458         spin_lock(&config_list_lock);
 459         cld->cld_stopping = 1;
 460         spin_unlock(&config_list_lock);
 461
 462         CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
 463         if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
 464                 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
 465 }
 466
 467 static inline void config_mark_cld_stop(struct config_llog_data *cld)
 468 {
 469         if (cld) {
 470                 mutex_lock(&cld->cld_lock);
 471                 config_mark_cld_stop_nolock(cld);
 472                 mutex_unlock(&cld->cld_lock);
 473         }
 474 }
 475
 476 /** Stop watching for updates on this log.
 477  */
 478 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 479 {
 480         struct config_llog_data *cld;
 481         struct config_llog_data *cld_sptlrpc = NULL;
 482         struct config_llog_data *cld_params = NULL;
 483         struct config_llog_data *cld_recover = NULL;
 484         struct config_llog_data *cld_nodemap = NULL;
 485         struct config_llog_data *cld_barrier = NULL;
 486         int rc = 0;
 487
 488         ENTRY;
 489
 490         cld = config_log_find(logname, cfg);
 491         if (cld == NULL)
 492                 RETURN(-ENOENT);
 493
 494         mutex_lock(&cld->cld_lock);
 495         /*
 496          * if cld_stopping is set, it means we didn't start the log thus
 497          * not owning the start ref. this can happen after previous umount:
 498          * the cld still hanging there waiting for lock cancel, and we
 499          * remount again but failed in the middle and call log_end without
 500          * calling start_log.
 501          */
 502         if (unlikely(cld->cld_stopping)) {
 503                 mutex_unlock(&cld->cld_lock);
 504                 /* drop the ref from the find */
 505                 config_log_put(cld);
 506                 RETURN(rc);
 507         }
 508
 509         cld_recover = cld->cld_recover;
 510         cld->cld_recover = NULL;
 511         cld_params = cld->cld_params;
 512         cld->cld_params = NULL;
 513         cld_nodemap = cld->cld_nodemap;
 514         cld->cld_nodemap = NULL;
 515         cld_barrier = cld->cld_barrier;
 516         cld->cld_barrier = NULL;
 517         cld_sptlrpc = cld->cld_sptlrpc;
 518         cld->cld_sptlrpc = NULL;
 519
 520         config_mark_cld_stop_nolock(cld);
 521         mutex_unlock(&cld->cld_lock);
 522
 523         config_mark_cld_stop(cld_recover);
 524         config_log_put(cld_recover);
 525         config_mark_cld_stop(cld_params);
 526         config_log_put(cld_params);
 527         config_mark_cld_stop(cld_barrier);
 528         config_log_put(cld_barrier);
 529         /* don't explicitly set cld_stopping on sptlrpc lock here, as other
 530          * targets may be active, it will be done in config_log_put if necessary
 531          */
 532         config_log_put(cld_sptlrpc);
 533         /* don't set cld_stopping on nm lock as other targets may be active */
 534         config_log_put(cld_nodemap);
 535
 536         /* drop the ref from the find */
 537         config_log_put(cld);
 538         /* drop the start ref */
 539         config_log_put(cld);
 540
 541         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 542                rc);
 543         RETURN(rc);
 544 }
 545
 546 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 547 {
 548         struct obd_device       *obd = data;
 549         struct obd_import       *imp;
 550         struct obd_connect_data *ocd;
 551         struct config_llog_data *cld;
 552         int rc = 0;
 553
 554         ENTRY;
 555         LASSERT(obd);
 556         with_imp_locked(obd, imp, rc) {
 557                 ocd = &imp->imp_connect_data;
 558
 559                 seq_printf(m, "imperative_recovery: %s\n",
 560                            OCD_HAS_FLAG(ocd, IMP_RECOV) ?
 561                            "ENABLED" : "DISABLED");
 562         }
 563         if (rc)
 564                 RETURN(rc);
 565
 566         seq_printf(m, "client_state:\n");
 567
 568         spin_lock(&config_list_lock);
 569         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 570                 if (cld->cld_recover == NULL)
 571                         continue;
 572                 seq_printf(m,  "    - { client: %s, nidtbl_version: %u }\n",
 573                            cld->cld_logname,
 574                            cld->cld_recover->cld_cfg.cfg_last_idx);
 575         }
 576         spin_unlock(&config_list_lock);
 577
 578         RETURN(0);
 579 }
 580
 581 /* reenqueue any lost locks */
 582 #define RQ_RUNNING      0x1
 583 #define RQ_NOW          0x2
 584 #define RQ_LATER        0x4
 585 #define RQ_STOP         0x8
 586 #define RQ_PRECLEANUP   0x10
 587 static int                    rq_state = 0;
 588 static wait_queue_head_t      rq_waitq;
 589 static DECLARE_COMPLETION(rq_exit);
 590 static DECLARE_COMPLETION(rq_start);
 591
 592 static void do_requeue(struct config_llog_data *cld)
 593 {
 594         int rc = 0;
 595         ENTRY;
 596
 597         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 598
 599         /*
 600          * Do not run mgc_process_log on a disconnected export or an
 601          * export which is being disconnected. Take the client
 602          * semaphore to make the check non-racy.
 603          */
 604         down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
 605                          OBD_CLI_SEM_MGC);
 606         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 607                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 608                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 609                 if (rc && rc != -ENOENT)
 610                         CERROR("failed processing log: %d\n", rc);
 611         } else {
 612                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 613                        cld->cld_logname);
 614         }
 615         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 616
 617         EXIT;
 618 }
 619
 620 /* this timeout represents how many seconds MGC should wait before
 621  * requeue config and recover lock to the MGS. We need to randomize this
 622  * in order to not flood the MGS.
 623  */
 624 #define MGC_TIMEOUT_MIN_SECONDS   5
 625 #define MGC_TIMEOUT_RAND_CENTISEC 500
 626
 627 static int mgc_requeue_thread(void *data)
 628 {
 629         int rc = 0;
 630         bool first = true;
 631         ENTRY;
 632
 633         CDEBUG(D_MGC, "Starting requeue thread\n");
 634
 635         /* Keep trying failed locks periodically */
 636         spin_lock(&config_list_lock);
 637         rq_state |= RQ_RUNNING;
 638         while (!(rq_state & RQ_STOP)) {
 639                 struct config_llog_data *cld, *cld_prev;
 640                 int rand = prandom_u32_max(MGC_TIMEOUT_RAND_CENTISEC);
 641                 int to;
 642
 643                 /* Any new or requeued lostlocks will change the state */
 644                 rq_state &= ~(RQ_NOW | RQ_LATER);
 645                 spin_unlock(&config_list_lock);
 646
 647                 if (first) {
 648                         first = false;
 649                         complete(&rq_start);
 650                 }
 651
 652                 /* Always wait a few seconds to allow the server who
 653                  * caused the lock revocation to finish its setup, plus some
 654                  * random so everyone doesn't try to reconnect at once.
 655                  */
 656                 to = cfs_time_seconds(MGC_TIMEOUT_MIN_SECONDS * 100 + rand);
 657                 /* rand is centi-seconds */
 658                 wait_event_idle_timeout(rq_waitq,
 659                                         rq_state & (RQ_STOP | RQ_PRECLEANUP),
 660                                         to/100);
 661
 662                 /*
 663                  * iterate & processing through the list. for each cld, process
 664                  * its depending sptlrpc cld firstly (if any) and then itself.
 665                  *
 666                  * it's guaranteed any item in the list must have
 667                  * reference > 0; and if cld_lostlock is set, at
 668                  * least one reference is taken by the previous enqueue.
 669                  */
 670                 cld_prev = NULL;
 671
 672                 spin_lock(&config_list_lock);
 673                 rq_state &= ~RQ_PRECLEANUP;
 674                 list_for_each_entry(cld, &config_llog_list,
 675                                     cld_list_chain) {
 676                         if (!cld->cld_lostlock || cld->cld_stopping)
 677                                 continue;
 678
 679                         /* hold reference to avoid being freed during
 680                          * subsequent processing. */
 681                         config_log_get(cld);
 682                         cld->cld_lostlock = 0;
 683                         spin_unlock(&config_list_lock);
 684
 685                         config_log_put(cld_prev);
 686                         cld_prev = cld;
 687
 688                         if (likely(!(rq_state & RQ_STOP))) {
 689                                 do_requeue(cld);
 690                                 spin_lock(&config_list_lock);
 691                         } else {
 692                                 spin_lock(&config_list_lock);
 693                                 break;
 694                         }
 695                 }
 696                 spin_unlock(&config_list_lock);
 697                 config_log_put(cld_prev);
 698
 699                 /* Wait a bit to see if anyone else needs a requeue */
 700                 wait_event_idle(rq_waitq, rq_state & (RQ_NOW | RQ_STOP));
 701                 spin_lock(&config_list_lock);
 702         }
 703
 704         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 705         rq_state &= ~RQ_RUNNING;
 706         spin_unlock(&config_list_lock);
 707
 708         complete(&rq_exit);
 709
 710         CDEBUG(D_MGC, "Ending requeue thread\n");
 711         RETURN(rc);
 712 }
 713
 714 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 715    We are responsible for dropping the config log reference from here on out. */
 716 static void mgc_requeue_add(struct config_llog_data *cld)
 717 {
 718         bool wakeup = false;
 719         ENTRY;
 720
 721         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 722                 cld->cld_logname, atomic_read(&cld->cld_refcount),
 723                 cld->cld_stopping, rq_state);
 724         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 725
 726         /* lets cancel an existent lock to mark cld as "lostlock" */
 727         CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
 728         if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
 729                 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
 730
 731         mutex_lock(&cld->cld_lock);
 732         spin_lock(&config_list_lock);
 733         if (!(rq_state & RQ_STOP) && !cld->cld_stopping) {
 734                 cld->cld_lostlock = 1;
 735                 rq_state |= RQ_NOW;
 736                 wakeup = true;
 737         }
 738         spin_unlock(&config_list_lock);
 739         mutex_unlock(&cld->cld_lock);
 740         if (wakeup)
 741                 wake_up(&rq_waitq);
 742
 743         EXIT;
 744 }
 745
 746 /********************** class fns **********************/
 747 static int mgc_local_llog_init(const struct lu_env *env,
 748                                struct obd_device *obd,
 749                                struct obd_device *disk)
 750 {
 751         struct llog_ctxt        *ctxt;
 752         int                      rc;
 753
 754         ENTRY;
 755
 756         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
 757                         &llog_osd_ops);
 758         if (rc)
 759                 RETURN(rc);
 760
 761         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 762         LASSERT(ctxt);
 763         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
 764         llog_ctxt_put(ctxt);
 765
 766         RETURN(0);
 767 }
 768
 769 static int mgc_local_llog_fini(const struct lu_env *env,
 770                                struct obd_device *obd)
 771 {
 772         struct llog_ctxt *ctxt;
 773
 774         ENTRY;
 775
 776         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 777         llog_cleanup(env, ctxt);
 778
 779         RETURN(0);
 780 }
 781
 782 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
 783                         struct super_block *sb)
 784 {
 785         struct lustre_sb_info   *lsi = s2lsi(sb);
 786         struct client_obd       *cli = &obd->u.cli;
 787         struct lu_fid            rfid, fid;
 788         struct dt_object        *root, *dto;
 789         int                      rc = 0;
 790
 791         ENTRY;
 792
 793         LASSERT(lsi);
 794         LASSERT(lsi->lsi_dt_dev);
 795
 796         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
 797         mutex_lock(&cli->cl_mgc_mutex);
 798
 799         /* Setup the configs dir */
 800         fid.f_seq = FID_SEQ_LOCAL_NAME;
 801         fid.f_oid = 1;
 802         fid.f_ver = 0;
 803         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
 804                                     &cli->cl_mgc_los);
 805         if (rc)
 806                 RETURN(rc);
 807
 808         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
 809         if (rc)
 810                 GOTO(out_los, rc);
 811
 812         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
 813                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
 814         if (unlikely(IS_ERR(root)))
 815                 GOTO(out_los, rc = PTR_ERR(root));
 816
 817         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
 818                                         MOUNT_CONFIGS_DIR,
 819                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
 820         dt_object_put_nocache(env, root);
 821         if (IS_ERR(dto))
 822                 GOTO(out_los, rc = PTR_ERR(dto));
 823
 824         cli->cl_mgc_configs_dir = dto;
 825
 826         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
 827         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
 828         if (rc)
 829                 GOTO(out_llog, rc);
 830
 831         /* We take an obd ref to insure that we can't get to mgc_cleanup
 832          * without calling mgc_fs_cleanup first. */
 833         class_incref(obd, "mgc_fs", obd);
 834
 835         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 836         EXIT;
 837 out_llog:
 838         if (rc) {
 839                 dt_object_put(env, cli->cl_mgc_configs_dir);
 840                 cli->cl_mgc_configs_dir = NULL;
 841         }
 842 out_los:
 843         if (rc < 0) {
 844                 local_oid_storage_fini(env, cli->cl_mgc_los);
 845                 cli->cl_mgc_los = NULL;
 846                 mutex_unlock(&cli->cl_mgc_mutex);
 847         }
 848         return rc;
 849 }
 850
 851 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
 852 {
 853         struct client_obd       *cli = &obd->u.cli;
 854         ENTRY;
 855
 856         LASSERT(cli->cl_mgc_los != NULL);
 857
 858         mgc_local_llog_fini(env, obd);
 859
 860         dt_object_put_nocache(env, cli->cl_mgc_configs_dir);
 861         cli->cl_mgc_configs_dir = NULL;
 862
 863         local_oid_storage_fini(env, cli->cl_mgc_los);
 864         cli->cl_mgc_los = NULL;
 865
 866         class_decref(obd, "mgc_fs", obd);
 867         mutex_unlock(&cli->cl_mgc_mutex);
 868
 869         RETURN(0);
 870 }
 871
 872 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 873 {
 874         struct llog_ctxt        *ctxt;
 875         int                      rc;
 876
 877         ENTRY;
 878
 879         /* setup only remote ctxt, the local disk context is switched per each
 880          * filesystem during mgc_fs_setup() */
 881         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 882                         &llog_client_ops);
 883         if (rc)
 884                 RETURN(rc);
 885
 886         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 887         LASSERT(ctxt);
 888
 889         llog_initiator_connect(ctxt);
 890         llog_ctxt_put(ctxt);
 891
 892         RETURN(0);
 893 }
 894
 895 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 896 {
 897         struct llog_ctxt *ctxt;
 898
 899         ENTRY;
 900
 901         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 902         if (ctxt)
 903                 llog_cleanup(env, ctxt);
 904
 905         RETURN(0);
 906 }
 907
 908
 909 static atomic_t mgc_count = ATOMIC_INIT(0);
 910 static int mgc_precleanup(struct obd_device *obd)
 911 {
 912         int     rc = 0;
 913         int     temp;
 914         ENTRY;
 915
 916         if (atomic_dec_and_test(&mgc_count)) {
 917                 LASSERT(rq_state & RQ_RUNNING);
 918                 /* stop requeue thread */
 919                 temp = RQ_STOP;
 920         } else {
 921                 /* wakeup requeue thread to clean our cld */
 922                 temp = RQ_NOW | RQ_PRECLEANUP;
 923         }
 924
 925         spin_lock(&config_list_lock);
 926         rq_state |= temp;
 927         spin_unlock(&config_list_lock);
 928         wake_up(&rq_waitq);
 929
 930         if (temp & RQ_STOP)
 931                 wait_for_completion(&rq_exit);
 932         obd_cleanup_client_import(obd);
 933
 934         rc = mgc_llog_fini(NULL, obd);
 935         if (rc != 0)
 936                 CERROR("failed to cleanup llogging subsystems\n");
 937
 938         RETURN(rc);
 939 }
 940
 941 static int mgc_cleanup(struct obd_device *obd)
 942 {
 943         int rc;
 944         ENTRY;
 945
 946         /* COMPAT_146 - old config logs may have added profiles we don't
 947            know about */
 948         if (atomic_read(&obd->obd_type->typ_refcnt) <= 1)
 949                 /* Only for the last mgc */
 950                 class_del_profiles();
 951
 952         lprocfs_obd_cleanup(obd);
 953         ptlrpcd_decref();
 954
 955         rc = client_obd_cleanup(obd);
 956         RETURN(rc);
 957 }
 958
 959 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 960 {
 961         struct task_struct      *task;
 962         int                      rc;
 963         ENTRY;
 964
 965         rc = ptlrpcd_addref();
 966         if (rc < 0)
 967                 RETURN(rc);
 968
 969         rc = client_obd_setup(obd, lcfg);
 970         if (rc)
 971                 GOTO(err_decref, rc);
 972
 973         rc = mgc_llog_init(NULL, obd);
 974         if (rc) {
 975                 CERROR("failed to setup llogging subsystems\n");
 976                 GOTO(err_cleanup, rc);
 977         }
 978
 979         rc = mgc_tunables_init(obd);
 980         if (rc)
 981                 GOTO(err_sysfs, rc);
 982
 983         if (atomic_inc_return(&mgc_count) == 1) {
 984                 rq_state = 0;
 985                 init_waitqueue_head(&rq_waitq);
 986
 987                 /* start requeue thread */
 988                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 989                 if (IS_ERR(task)) {
 990                         rc = PTR_ERR(task);
 991                         CERROR("%s: cannot start requeue thread: rc = %d; "
 992                                "no more log updates\n",
 993                                obd->obd_name, rc);
 994                         GOTO(err_sysfs, rc);
 995                 }
 996                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 997                 rc = 0;
 998                 wait_for_completion(&rq_start);
 999         }
1000
1001         RETURN(rc);
1002
1003 err_sysfs:
1004         lprocfs_obd_cleanup(obd);
1005 err_cleanup:
1006         client_obd_cleanup(obd);
1007 err_decref:
1008         ptlrpcd_decref();
1009         RETURN(rc);
1010 }
1011
1012 /* based on ll_mdc_blocking_ast */
1013 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1014                             void *data, int flag)
1015 {
1016         struct lustre_handle lockh;
1017         struct config_llog_data *cld = (struct config_llog_data *)data;
1018         int rc = 0;
1019         ENTRY;
1020
1021         switch (flag) {
1022         case LDLM_CB_BLOCKING:
1023                 /* mgs wants the lock, give it up... */
1024                 LDLM_DEBUG(lock, "MGC blocking CB");
1025                 ldlm_lock2handle(lock, &lockh);
1026                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
1027                 break;
1028         case LDLM_CB_CANCELING:
1029                 /* We've given up the lock, prepare ourselves to update. */
1030                 LDLM_DEBUG(lock, "MGC cancel CB");
1031
1032                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
1033                        PLDLMRES(lock->l_resource),
1034                        (char *)&lock->l_resource->lr_name.name[0]);
1035
1036                 if (!cld) {
1037                         CDEBUG(D_INFO, "missing data, won't requeue\n");
1038                         break;
1039                 }
1040
1041                 /* held at mgc_process_log(). */
1042                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1043
1044                 lock->l_ast_data = NULL;
1045                 cld->cld_lockh.cookie = 0;
1046                 /* Are we done with this log? */
1047                 if (cld->cld_stopping) {
1048                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1049                                 cld->cld_logname);
1050                         config_log_put(cld);
1051                         break;
1052                 }
1053                 /* Make sure not to re-enqueue when the mgc is stopping
1054                    (we get called from client_disconnect_export) */
1055                 if (lock->l_conn_export == NULL ||
1056                     lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1057                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1058                                 cld->cld_logname);
1059                         config_log_put(cld);
1060                         break;
1061                 }
1062
1063                 /* Re-enqueue now */
1064                 mgc_requeue_add(cld);
1065                 config_log_put(cld);
1066                 break;
1067         default:
1068                 LBUG();
1069         }
1070
1071         RETURN(rc);
1072 }
1073
1074 /* Not sure where this should go... */
1075 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1076  * that we can have a chance to try the secondary MGS if any. */
1077 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1078                                 + PING_INTERVAL)
1079 #define  MGC_TARGET_REG_LIMIT 10
1080 #define  MGC_TARGET_REG_LIMIT_MAX RECONNECT_DELAY_MAX
1081 #define  MGC_SEND_PARAM_LIMIT 10
1082
1083 /* Take a config lock so we can get cancel notifications */
1084 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1085                        union ldlm_policy_data *policy, enum ldlm_mode mode,
1086                        __u64 *flags, ldlm_glimpse_callback glimpse_callback,
1087                        void *data, __u32 lvb_len, void *lvb_swabber,
1088                        struct lustre_handle *lockh)
1089 {
1090         struct config_llog_data *cld = (struct config_llog_data *)data;
1091         struct ldlm_enqueue_info einfo = {
1092                 .ei_type        = type,
1093                 .ei_mode        = mode,
1094                 .ei_cb_bl       = mgc_blocking_ast,
1095                 .ei_cb_cp       = ldlm_completion_ast,
1096                 .ei_cb_gl       = glimpse_callback,
1097         };
1098         struct ptlrpc_request *req;
1099         int short_limit = cld_is_sptlrpc(cld);
1100         int rc;
1101         ENTRY;
1102
1103         if (!exp)
1104                 RETURN(-EBADR);
1105
1106         CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1107                cld->cld_resid.name[0]);
1108
1109         /* We need a callback for every lockholder, so don't try to
1110            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1111         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1112                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1113                                         LDLM_ENQUEUE);
1114         if (req == NULL)
1115                 RETURN(-ENOMEM);
1116
1117         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1118         ptlrpc_request_set_replen(req);
1119
1120         /* check if this is server or client */
1121         if (cld->cld_cfg.cfg_sb) {
1122                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1123                 if (lsi && IS_SERVER(lsi))
1124                         short_limit = 1;
1125         }
1126         /* Limit how long we will wait for the enqueue to complete */
1127         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1128         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1129                               NULL, 0, LVB_T_NONE, lockh, 0);
1130         /* A failed enqueue should still call the mgc_blocking_ast,
1131            where it will be requeued if needed ("grant failed"). */
1132         ptlrpc_req_finished(req);
1133         RETURN(rc);
1134 }
1135
1136 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1137                       struct lustre_handle *lockh)
1138 {
1139         ENTRY;
1140
1141         ldlm_lock_decref(lockh, mode);
1142
1143         RETURN(0);
1144 }
1145
1146 static void mgc_notify_active(struct obd_device *unused)
1147 {
1148         /* wakeup mgc_requeue_thread to requeue mgc lock */
1149         spin_lock(&config_list_lock);
1150         rq_state |= RQ_NOW;
1151         spin_unlock(&config_list_lock);
1152         wake_up(&rq_waitq);
1153
1154         /* TODO: Help the MGS rebuild nidtbl. -jay */
1155 }
1156
1157 /* Send target_reg message to MGS */
1158 static int mgc_target_register(struct obd_export *exp,
1159                                struct mgs_target_info *mti)
1160 {
1161         struct ptlrpc_request  *req;
1162         struct mgs_target_info *req_mti, *rep_mti;
1163         int                     rc;
1164         ENTRY;
1165
1166         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1167                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1168                                         MGS_TARGET_REG);
1169         if (req == NULL)
1170                 RETURN(-ENOMEM);
1171
1172         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1173         if (!req_mti) {
1174                 ptlrpc_req_finished(req);
1175                 RETURN(-ENOMEM);
1176         }
1177
1178         memcpy(req_mti, mti, sizeof(*req_mti));
1179         ptlrpc_request_set_replen(req);
1180         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1181         /* Limit how long we will wait for the enqueue to complete */
1182         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1183
1184         /* if the target needs to regenerate the config log in MGS, it's better
1185          * to use some longer limit to let MGC have time to change connection to
1186          * another MGS (or try again with the same MGS) for the target (server)
1187          * will fail and exit if the request expired due to delay limit. */
1188         if (mti->mti_flags & (LDD_F_UPDATE | LDD_F_NEED_INDEX))
1189                 req->rq_delay_limit = MGC_TARGET_REG_LIMIT_MAX;
1190
1191         rc = ptlrpc_queue_wait(req);
1192         if (!rc) {
1193                 rep_mti = req_capsule_server_get(&req->rq_pill,
1194                                                  &RMF_MGS_TARGET_INFO);
1195                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1196                 CDEBUG(D_MGC, "register %s got index = %d\n",
1197                        mti->mti_svname, mti->mti_stripe_index);
1198         }
1199         ptlrpc_req_finished(req);
1200
1201         RETURN(rc);
1202 }
1203
1204 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1205                               u32 keylen, void *key,
1206                               u32 vallen, void *val,
1207                               struct ptlrpc_request_set *set)
1208 {
1209         int rc = -EINVAL;
1210         ENTRY;
1211
1212         /* Turn off initial_recov after we try all backup servers once */
1213         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1214                 struct obd_import *imp = class_exp2cliimp(exp);
1215                 int value;
1216                 if (vallen != sizeof(int))
1217                         RETURN(-EINVAL);
1218                 value = *(int *)val;
1219                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1220                        imp->imp_obd->obd_name, value,
1221                        imp->imp_deactive, imp->imp_invalid,
1222                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1223                        ptlrpc_import_state_name(imp->imp_state));
1224                 /* Resurrect the import immediately if
1225                  * 1. we previously got disconnected,
1226                  * 2. value > 1 (at the same node with MGS)
1227                  * */
1228                 if (imp->imp_state == LUSTRE_IMP_DISCON || value > 1)
1229                         ptlrpc_reconnect_import(imp);
1230
1231                 RETURN(0);
1232         }
1233
1234         /* FIXME move this to mgc_process_config */
1235         if (KEY_IS(KEY_REGISTER_TARGET)) {
1236                 struct mgs_target_info *mti;
1237                 if (vallen != sizeof(struct mgs_target_info))
1238                         RETURN(-EINVAL);
1239                 mti = (struct mgs_target_info *)val;
1240                 CDEBUG(D_MGC, "register_target %s %#x\n",
1241                        mti->mti_svname, mti->mti_flags);
1242                 rc =  mgc_target_register(exp, mti);
1243                 RETURN(rc);
1244         }
1245         if (KEY_IS(KEY_SET_FS)) {
1246                 struct super_block *sb = (struct super_block *)val;
1247
1248                 if (vallen != sizeof(struct super_block))
1249                         RETURN(-EINVAL);
1250
1251                 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1252                 RETURN(rc);
1253         }
1254         if (KEY_IS(KEY_CLEAR_FS)) {
1255                 if (vallen != 0)
1256                         RETURN(-EINVAL);
1257                 rc = mgc_fs_cleanup(env, exp->exp_obd);
1258                 RETURN(rc);
1259         }
1260         if (KEY_IS(KEY_MGSSEC)) {
1261                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1262                 struct sptlrpc_flavor  flvr;
1263
1264                 /*
1265                  * empty string means using current flavor, if which haven't
1266                  * been set yet, set it as null.
1267                  *
1268                  * if flavor has been set previously, check the asking flavor
1269                  * must match the existing one.
1270                  */
1271                 if (vallen == 0) {
1272                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1273                                 RETURN(0);
1274                         val = "null";
1275                         vallen = 4;
1276                 }
1277
1278                 rc = sptlrpc_parse_flavor(val, &flvr);
1279                 if (rc) {
1280                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1281                                (char *) val);
1282                         RETURN(rc);
1283                 }
1284
1285                 /*
1286                  * caller already hold a mutex
1287                  */
1288                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1289                         cli->cl_flvr_mgc = flvr;
1290                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1291                                   sizeof(flvr)) != 0) {
1292                         char    str[20];
1293
1294                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1295                                             str, sizeof(str));
1296                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1297                                        "currently %s is in use\n",
1298                                        (char *) val, str);
1299                         rc = -EPERM;
1300                 }
1301                 RETURN(rc);
1302         }
1303
1304         RETURN(rc);
1305 }
1306
1307 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1308                         __u32 keylen, void *key, __u32 *vallen, void *val)
1309 {
1310         int rc = -EINVAL;
1311
1312         if (KEY_IS(KEY_CONN_DATA)) {
1313                 struct obd_import *imp = class_exp2cliimp(exp);
1314                 struct obd_connect_data *data = val;
1315
1316                 if (*vallen == sizeof(*data)) {
1317                         *data = imp->imp_connect_data;
1318                         rc = 0;
1319                 }
1320         }
1321
1322         return rc;
1323 }
1324
1325 static int mgc_import_event(struct obd_device *obd,
1326                             struct obd_import *imp,
1327                             enum obd_import_event event)
1328 {
1329         int rc = 0;
1330
1331         LASSERT(imp->imp_obd == obd);
1332         CDEBUG(D_MGC, "import event %#x\n", event);
1333
1334         switch (event) {
1335         case IMP_EVENT_DISCON:
1336                 /* MGC imports should not wait for recovery */
1337                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1338                         ptlrpc_pinger_ir_down();
1339                 break;
1340         case IMP_EVENT_INACTIVE:
1341                 break;
1342         case IMP_EVENT_INVALIDATE: {
1343                 struct ldlm_namespace *ns = obd->obd_namespace;
1344                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1345                 break;
1346         }
1347         case IMP_EVENT_ACTIVE:
1348                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1349                 /* Clearing obd_no_recov allows us to continue pinging */
1350                 obd->obd_no_recov = 0;
1351                 mgc_notify_active(obd);
1352                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1353                         ptlrpc_pinger_ir_up();
1354                 break;
1355         case IMP_EVENT_OCD:
1356                 break;
1357         case IMP_EVENT_DEACTIVATE:
1358         case IMP_EVENT_ACTIVATE:
1359                 break;
1360         default:
1361                 CERROR("Unknown import event %#x\n", event);
1362                 LBUG();
1363         }
1364         RETURN(rc);
1365 }
1366
1367 enum {
1368         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1369         CONFIG_READ_NRPAGES      = 4
1370 };
1371
1372 static int mgc_apply_recover_logs(struct obd_device *mgc,
1373                                   struct config_llog_data *cld,
1374                                   __u64 max_version,
1375                                   void *data, int datalen, bool mne_swab)
1376 {
1377         struct config_llog_instance *cfg = &cld->cld_cfg;
1378         struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1379         struct mgs_nidtbl_entry *entry;
1380         struct lustre_cfg *lcfg;
1381         struct lustre_cfg_bufs bufs;
1382         u64 prev_version = 0;
1383         char *inst;
1384         char *buf;
1385         int bufsz;
1386         int pos = 0;
1387         int rc  = 0;
1388         int off = 0;
1389         unsigned long dynamic_nids;
1390
1391         ENTRY;
1392         LASSERT(cfg->cfg_instance != 0);
1393         LASSERT(ll_get_cfg_instance(cfg->cfg_sb) == cfg->cfg_instance);
1394
1395         /* get dynamic nids setting */
1396         dynamic_nids = mgc->obd_dynamic_nids;
1397
1398         OBD_ALLOC(inst, PAGE_SIZE);
1399         if (inst == NULL)
1400                 RETURN(-ENOMEM);
1401
1402         if (!IS_SERVER(lsi)) {
1403                 pos = snprintf(inst, PAGE_SIZE, "%016lx", cfg->cfg_instance);
1404                 if (pos >= PAGE_SIZE) {
1405                         OBD_FREE(inst, PAGE_SIZE);
1406                         return -E2BIG;
1407                 }
1408 #ifdef HAVE_SERVER_SUPPORT
1409         } else {
1410                 LASSERT(IS_MDT(lsi));
1411                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1412                                         PAGE_SIZE);
1413                 if (rc) {
1414                         OBD_FREE(inst, PAGE_SIZE);
1415                         RETURN(-EINVAL);
1416                 }
1417                 pos = strlen(inst);
1418 #endif /* HAVE_SERVER_SUPPORT */
1419         }
1420
1421         ++pos;
1422         buf   = inst + pos;
1423         bufsz = PAGE_SIZE - pos;
1424
1425         while (datalen > 0) {
1426                 int   entry_len = sizeof(*entry);
1427                 int   is_ost;
1428                 struct obd_device *obd;
1429                 char *obdname;
1430                 char *cname;
1431                 char *params;
1432                 char *uuid;
1433
1434                 rc = -EINVAL;
1435                 if (datalen < sizeof(*entry))
1436                         break;
1437
1438                 entry = (typeof(entry))(data + off);
1439
1440                 /* sanity check */
1441                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1442                         break;
1443                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1444                         break;
1445                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1446                         break;
1447
1448                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1449                 if (datalen < entry_len) /* must have entry_len at least */
1450                         break;
1451
1452                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1453                 if (mne_swab)
1454                         lustre_swab_mgs_nidtbl_entry(entry);
1455                 if (entry->mne_length > PAGE_SIZE) {
1456                         CERROR("MNE too large (%u)\n", entry->mne_length);
1457                         break;
1458                 }
1459
1460                 if (entry->mne_length < entry_len)
1461                         break;
1462
1463                 off     += entry->mne_length;
1464                 datalen -= entry->mne_length;
1465                 if (datalen < 0)
1466                         break;
1467
1468                 if (entry->mne_version > max_version) {
1469                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1470                                entry->mne_version, max_version);
1471                         break;
1472                 }
1473
1474                 if (prev_version >= entry->mne_version) {
1475                         CERROR("index unsorted, prev %lld, now %lld\n",
1476                                prev_version, entry->mne_version);
1477                         break;
1478                 }
1479                 prev_version = entry->mne_version;
1480
1481                 /*
1482                  * Write a string with format "nid::instance" to
1483                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1484                  */
1485
1486                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1487                 memset(buf, 0, bufsz);
1488                 obdname = buf;
1489                 pos = 0;
1490
1491                 /* lustre-OST0001-osc-<instance #> */
1492                 strcpy(obdname, cld->cld_logname);
1493                 cname = strrchr(obdname, '-');
1494                 if (cname == NULL) {
1495                         CERROR("mgc %s: invalid logname %s\n",
1496                                mgc->obd_name, obdname);
1497                         break;
1498                 }
1499
1500                 pos = cname - obdname;
1501                 obdname[pos] = 0;
1502                 pos += sprintf(obdname + pos, "-%s%04x",
1503                                   is_ost ? "OST" : "MDT", entry->mne_index);
1504
1505                 cname = is_ost ? "osc" : "mdc",
1506                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1507                 lustre_cfg_bufs_reset(&bufs, obdname);
1508
1509                 /* find the obd by obdname */
1510                 obd = class_name2obd(obdname);
1511                 if (obd == NULL) {
1512                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1513                                mgc->obd_name, obdname);
1514                         rc = 0;
1515                         /* this is a safe race, when the ost is starting up...*/
1516                         continue;
1517                 }
1518
1519                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1520                 ++pos;
1521                 params = buf + pos;
1522                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1523                 uuid = buf + pos;
1524
1525                 down_read(&obd->u.cli.cl_sem);
1526                 if (obd->u.cli.cl_import == NULL) {
1527                         /* client does not connect to the OST yet */
1528                         up_read(&obd->u.cli.cl_sem);
1529                         rc = 0;
1530                         continue;
1531                 }
1532
1533                 /* iterate all nids to find one */
1534                 /* find uuid by nid */
1535                 /* create import entries if they don't exist */
1536                 rc = -ENOENT;
1537                 rc = client_import_add_nids_to_conn(obd->u.cli.cl_import,
1538                                                     entry->u.nids,
1539                                                     entry->mne_nid_count,
1540                                                     (struct obd_uuid *)uuid);
1541
1542                 if (rc == -ENOENT && dynamic_nids) {
1543                         /* create a new connection for this import */
1544                         char *primary_nid = libcfs_nid2str(entry->u.nids[0]);
1545                         int prim_nid_len = strlen(primary_nid) + 1;
1546                         struct obd_uuid server_uuid;
1547
1548                         if (prim_nid_len > UUID_MAX)
1549                                 goto fail;
1550                         strncpy(server_uuid.uuid, primary_nid, prim_nid_len);
1551
1552                         CDEBUG(D_INFO, "Adding a connection for %s\n",
1553                                primary_nid);
1554
1555                         rc = client_import_dyn_add_conn(obd->u.cli.cl_import,
1556                                                         &server_uuid,
1557                                                         entry->u.nids[0], 1);
1558                         if (rc < 0) {
1559                                 CERROR("%s: Failed to add new connection with NID '%s' to import: rc = %d\n",
1560                                        obd->obd_name, primary_nid, rc);
1561                                 goto fail;
1562                         }
1563                         rc = client_import_add_nids_to_conn(obd->u.cli.cl_import,
1564                                                         entry->u.nids,
1565                                                         entry->mne_nid_count,
1566                                                         (struct obd_uuid *)uuid);
1567                         if (rc < 0) {
1568                                 CERROR("%s: failed to lookup UUID: rc = %d\n",
1569                                        obd->obd_name, rc);
1570                                 goto fail;
1571                         }
1572                 }
1573
1574 fail:
1575                 up_read(&obd->u.cli.cl_sem);
1576                 if (rc < 0 && rc != -ENOSPC) {
1577                         CERROR("mgc: cannot find UUID by nid '%s': rc = %d\n",
1578                                libcfs_nid2str(entry->u.nids[0]), rc);
1579                         break;
1580                 }
1581
1582                 CDEBUG(D_INFO, "Found UUID '%s' by NID '%s'\n",
1583                        uuid, libcfs_nid2str(entry->u.nids[0]));
1584
1585                 pos += strlen(uuid);
1586                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1587                 LASSERT(pos < bufsz);
1588
1589                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1590
1591                 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1592                                                bufs.lcfg_buflen));
1593                 if (!lcfg) {
1594                         rc = -ENOMEM;
1595                         break;
1596                 }
1597                 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1598
1599                 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1600                        prev_version, max_version, obdname, params);
1601
1602                 rc = class_process_config(lcfg);
1603                 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1604                                               lcfg->lcfg_buflens));
1605                 if (rc)
1606                         CDEBUG(D_INFO, "process config for %s error %d\n",
1607                                obdname, rc);
1608
1609                 /* continue, even one with error */
1610         }
1611
1612         OBD_FREE(inst, PAGE_SIZE);
1613         RETURN(rc);
1614 }
1615
1616 /**
1617  * This function is called if this client was notified for target restarting
1618  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1619  * nodemap logs.
1620  */
1621 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1622                                            struct config_llog_data *cld)
1623 {
1624         struct ptlrpc_connection *mgc_conn;
1625         struct ptlrpc_request *req = NULL;
1626         struct config_llog_instance *cfg = &cld->cld_cfg;
1627         struct mgs_config_body *body;
1628         struct mgs_config_res *res;
1629         struct nodemap_config *new_config = NULL;
1630         struct lu_nodemap *recent_nodemap = NULL;
1631         struct ptlrpc_bulk_desc *desc;
1632         struct page **pages = NULL;
1633         __u64 config_read_offset = 0;
1634         __u8 nodemap_cur_pass = 0;
1635         int nrpages = 0;
1636         bool eof = true;
1637         bool mne_swab = false;
1638         int i;
1639         int ealen;
1640         int rc;
1641         ENTRY;
1642
1643         mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1644
1645         /* don't need to get local config */
1646         if (cld_is_nodemap(cld) && LNetIsPeerLocal(mgc_conn->c_peer.nid))
1647                 GOTO(out, rc = 0);
1648
1649         /* allocate buffer for bulk transfer.
1650          * if this is the first time for this mgs to read logs,
1651          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1652          * once; otherwise, it only reads increment of logs, this should be
1653          * small and CONFIG_READ_NRPAGES will be used.
1654          */
1655         nrpages = CONFIG_READ_NRPAGES;
1656         if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1657                 nrpages = CONFIG_READ_NRPAGES_INIT;
1658
1659         OBD_ALLOC_PTR_ARRAY(pages, nrpages);
1660         if (pages == NULL)
1661                 GOTO(out, rc = -ENOMEM);
1662
1663         for (i = 0; i < nrpages; i++) {
1664                 pages[i] = alloc_page(GFP_KERNEL);
1665                 if (pages[i] == NULL)
1666                         GOTO(out, rc = -ENOMEM);
1667         }
1668
1669 again:
1670 #ifdef HAVE_SERVER_SUPPORT
1671         if (cld_is_nodemap(cld) && config_read_offset == 0) {
1672                 new_config = nodemap_config_alloc();
1673                 if (IS_ERR(new_config)) {
1674                         rc = PTR_ERR(new_config);
1675                         new_config = NULL;
1676                         GOTO(out, rc);
1677                 }
1678         }
1679 #endif
1680         LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1681         LASSERT(mutex_is_locked(&cld->cld_lock));
1682         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1683                                    &RQF_MGS_CONFIG_READ);
1684         if (req == NULL)
1685                 GOTO(out, rc = -ENOMEM);
1686
1687         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1688         if (rc)
1689                 GOTO(out, rc);
1690
1691         /* pack request */
1692         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1693         LASSERT(body != NULL);
1694         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1695         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1696             >= sizeof(body->mcb_name))
1697                 GOTO(out, rc = -E2BIG);
1698         if (cld_is_nodemap(cld))
1699                 body->mcb_offset = config_read_offset;
1700         else
1701                 body->mcb_offset = cfg->cfg_last_idx + 1;
1702         body->mcb_type   = cld->cld_type;
1703         body->mcb_bits   = PAGE_SHIFT;
1704         body->mcb_units  = nrpages;
1705         body->mcb_nm_cur_pass = nodemap_cur_pass;
1706
1707         /* allocate bulk transfer descriptor */
1708         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1709                                     PTLRPC_BULK_PUT_SINK,
1710                                     MGS_BULK_PORTAL,
1711                                     &ptlrpc_bulk_kiov_pin_ops);
1712         if (desc == NULL)
1713                 GOTO(out, rc = -ENOMEM);
1714
1715         for (i = 0; i < nrpages; i++)
1716                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1717                                                  PAGE_SIZE);
1718
1719         ptlrpc_request_set_replen(req);
1720         rc = ptlrpc_queue_wait(req);
1721         if (rc)
1722                 GOTO(out, rc);
1723
1724         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1725         if (!res)
1726                 GOTO(out, rc = -EPROTO);
1727
1728         if (cld_is_nodemap(cld)) {
1729                 config_read_offset = res->mcr_offset;
1730                 eof = config_read_offset == II_END_OFF;
1731                 nodemap_cur_pass = res->mcr_nm_cur_pass;
1732         } else {
1733                 if (res->mcr_size < res->mcr_offset)
1734                         GOTO(out, rc = -EINVAL);
1735
1736                 /* always update the index even though it might have errors with
1737                  * handling the recover logs
1738                  */
1739                 cfg->cfg_last_idx = res->mcr_offset;
1740                 eof = res->mcr_offset == res->mcr_size;
1741
1742                 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1743                        res->mcr_offset, eof == false);
1744         }
1745
1746         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1747         if (ealen < 0)
1748                 GOTO(out, rc = ealen);
1749
1750         if (ealen > nrpages << PAGE_SHIFT)
1751                 GOTO(out, rc = -EINVAL);
1752
1753         if (ealen == 0) { /* no logs transferred */
1754 #ifdef HAVE_SERVER_SUPPORT
1755                 /* config changed since first read RPC */
1756                 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1757                         CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1758                         GOTO(out, rc = -EAGAIN);
1759                 }
1760 #endif
1761                 if (!eof)
1762                         rc = -EINVAL;
1763                 GOTO(out, rc);
1764         }
1765
1766         mne_swab = ptlrpc_rep_need_swab(req);
1767
1768         /* When a nodemap config is received, we build a new nodemap config,
1769          * with new nodemap structs. We keep track of the most recently added
1770          * nodemap since the config is read ordered by nodemap_id, and so it
1771          * is likely that the next record will be related. Because access to
1772          * the nodemaps is single threaded until the nodemap_config is active,
1773          * we don't need to reference count with recent_nodemap, though
1774          * recent_nodemap should be set to NULL when the nodemap_config
1775          * is either destroyed or set active.
1776          */
1777         for (i = 0; i < nrpages && ealen > 0; i++) {
1778                 int rc2;
1779                 union lu_page   *ptr;
1780
1781                 ptr = kmap(pages[i]);
1782                 if (cld_is_nodemap(cld))
1783                         rc2 = nodemap_process_idx_pages(new_config, ptr,
1784                                                        &recent_nodemap);
1785                 else
1786                         rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1787                                                      ptr,
1788                                                      min_t(int, ealen,
1789                                                            PAGE_SIZE),
1790                                                      mne_swab);
1791                 kunmap(pages[i]);
1792                 if (rc2 < 0) {
1793                         CWARN("%s: error processing %s log %s: rc = %d\n",
1794                               obd->obd_name,
1795                               cld_is_nodemap(cld) ? "nodemap" : "recovery",
1796                               cld->cld_logname,
1797                               rc2);
1798                         GOTO(out, rc = rc2);
1799                 }
1800
1801                 ealen -= PAGE_SIZE;
1802         }
1803
1804 out:
1805         if (req) {
1806                 ptlrpc_req_finished(req);
1807                 req = NULL;
1808         }
1809
1810         if (rc == 0 && !eof)
1811                 goto again;
1812
1813 #ifdef HAVE_SERVER_SUPPORT
1814         if (new_config != NULL) {
1815                 /* recent_nodemap cannot be used after set_active/dealloc */
1816                 if (rc == 0)
1817                         nodemap_config_set_active_mgc(new_config);
1818                 else
1819                         nodemap_config_dealloc(new_config);
1820         }
1821 #endif
1822
1823         if (pages) {
1824                 for (i = 0; i < nrpages; i++) {
1825                         if (pages[i] == NULL)
1826                                 break;
1827                         __free_page(pages[i]);
1828                 }
1829                 OBD_FREE_PTR_ARRAY(pages, nrpages);
1830         }
1831         return rc;
1832 }
1833
1834 static int mgc_barrier_glimpse_ast(struct ldlm_lock *lock, void *data)
1835 {
1836         struct config_llog_data *cld = lock->l_ast_data;
1837         int rc;
1838         ENTRY;
1839
1840         if (cld->cld_stopping)
1841                 RETURN(-ENODEV);
1842
1843         rc = barrier_handler(s2lsi(cld->cld_cfg.cfg_sb)->lsi_dt_dev,
1844                              (struct ptlrpc_request *)data);
1845
1846         RETURN(rc);
1847 }
1848
1849 /* Copy a remote log locally */
1850 static int mgc_llog_local_copy(const struct lu_env *env,
1851                                struct obd_device *obd,
1852                                struct llog_ctxt *rctxt,
1853                                struct llog_ctxt *lctxt, char *logname)
1854 {
1855         char    *temp_log;
1856         int      rc;
1857
1858         ENTRY;
1859
1860         /*
1861          * - copy it to backup using llog_backup()
1862          * - copy remote llog to logname using llog_backup()
1863          * - if failed then move bakup to logname again
1864          */
1865
1866         OBD_ALLOC(temp_log, strlen(logname) + 2);
1867         if (!temp_log)
1868                 RETURN(-ENOMEM);
1869         sprintf(temp_log, "%sT", logname);
1870
1871         /* make a copy of local llog at first */
1872         rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1873         if (rc < 0 && rc != -ENOENT)
1874                 GOTO(out, rc);
1875         /* copy remote llog to the local copy */
1876         rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1877         if (rc == -ENOENT) {
1878                 /* no remote llog, delete local one too */
1879                 llog_erase(env, lctxt, NULL, logname);
1880         } else if (rc < 0) {
1881                 /* error during backup, get local one back from the copy */
1882                 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1883 out:
1884                 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1885                        obd->obd_name, logname, rc);
1886         }
1887         llog_erase(env, lctxt, NULL, temp_log);
1888         OBD_FREE(temp_log, strlen(logname) + 2);
1889         return rc;
1890 }
1891
1892 /* local_only means it cannot get remote llogs */
1893 static int mgc_process_cfg_log(struct obd_device *mgc,
1894                                struct config_llog_data *cld, int local_only)
1895 {
1896         struct llog_ctxt        *ctxt, *lctxt = NULL;
1897         struct client_obd       *cli = &mgc->u.cli;
1898         struct lustre_sb_info   *lsi = NULL;
1899         int                      rc = 0;
1900         struct lu_env           *env;
1901
1902         ENTRY;
1903
1904         LASSERT(cld);
1905         LASSERT(mutex_is_locked(&cld->cld_lock));
1906
1907         if (cld->cld_cfg.cfg_sb)
1908                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1909
1910         OBD_ALLOC_PTR(env);
1911         if (env == NULL)
1912                 RETURN(-ENOMEM);
1913
1914         rc = lu_env_init(env, LCT_MG_THREAD);
1915         if (rc)
1916                 GOTO(out_free, rc);
1917
1918         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1919         LASSERT(ctxt);
1920
1921         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1922
1923         /* Copy the setup log locally if we can. Don't mess around if we're
1924          * running an MGS though (logs are already local). */
1925         if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1926             cli->cl_mgc_configs_dir != NULL &&
1927             lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1928             lsi->lsi_dt_dev) {
1929                 if (!local_only && !lsi->lsi_dt_dev->dd_rdonly)
1930                         /* Only try to copy log if we have the lock. */
1931                         rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1932                                                  cld->cld_logname);
1933                 if (local_only || rc) {
1934                         if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1935                             llog_is_empty(env, lctxt, cld->cld_logname)) {
1936                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1937                                                    "log %s and no local copy."
1938                                                    "\n", cld->cld_logname);
1939                                 GOTO(out_pop, rc = -ENOENT);
1940                         }
1941                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1942                                "copy for now, will try to update later.\n",
1943                                cld->cld_logname);
1944                         rc = 0;
1945                 }
1946                 /* Now, whether we copied or not, start using the local llog.
1947                  * If we failed to copy, we'll start using whatever the old
1948                  * log has. */
1949                 llog_ctxt_put(ctxt);
1950                 ctxt = lctxt;
1951                 lctxt = NULL;
1952         } else {
1953                 if (local_only) /* no local log at client side */
1954                         GOTO(out_pop, rc = -EIO);
1955         }
1956
1957         rc = -EAGAIN;
1958         if (lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1959             lsi->lsi_dt_dev->dd_rdonly) {
1960                 struct llog_ctxt *rctxt;
1961
1962                 /* Under readonly mode, we may have no local copy or local
1963                  * copy is incomplete, so try to use remote llog firstly. */
1964                 rctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1965                 LASSERT(rctxt);
1966
1967                 rc = class_config_parse_llog(env, rctxt, cld->cld_logname,
1968                                              &cld->cld_cfg);
1969                 llog_ctxt_put(rctxt);
1970         }
1971
1972         if (rc && rc != -ENOENT)
1973                 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1974                                              &cld->cld_cfg);
1975
1976         /*
1977          * update settings on existing OBDs.
1978          * the logname must be <fsname>-sptlrpc
1979          */
1980         if (rc == 0 && cld_is_sptlrpc(cld))
1981                 class_notify_sptlrpc_conf(cld->cld_logname,
1982                                           strlen(cld->cld_logname) -
1983                                           strlen("-sptlrpc"));
1984         EXIT;
1985
1986 out_pop:
1987         __llog_ctxt_put(env, ctxt);
1988         if (lctxt)
1989                 __llog_ctxt_put(env, lctxt);
1990
1991         lu_env_fini(env);
1992 out_free:
1993         OBD_FREE_PTR(env);
1994         return rc;
1995 }
1996
1997 static bool mgc_import_in_recovery(struct obd_import *imp)
1998 {
1999         bool in_recovery = true;
2000
2001         spin_lock(&imp->imp_lock);
2002         if (imp->imp_state == LUSTRE_IMP_FULL ||
2003             imp->imp_state == LUSTRE_IMP_CLOSED)
2004                 in_recovery = false;
2005         spin_unlock(&imp->imp_lock);
2006
2007         return in_recovery;
2008 }
2009
2010 /**
2011  * Get a configuration log from the MGS and process it.
2012  *
2013  * This function is called for both clients and servers to process the
2014  * configuration log from the MGS.  The MGC enqueues a DLM lock on the
2015  * log from the MGS, and if the lock gets revoked the MGC will be notified
2016  * by the lock cancellation callback that the config log has changed,
2017  * and will enqueue another MGS lock on it, and then continue processing
2018  * the new additions to the end of the log.
2019  *
2020  * Since the MGC import is not replayable, if the import is being evicted
2021  * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
2022  * the log until recovery is finished or the import is closed.
2023  *
2024  * Make a local copy of the log before parsing it if appropriate (non-MGS
2025  * server) so that the server can start even when the MGS is down.
2026  *
2027  * There shouldn't be multiple processes running process_log at once --
2028  * sounds like badness.  It actually might be fine, as long as they're not
2029  * trying to update from the same log simultaneously, in which case we
2030  * should use a per-log semaphore instead of cld_lock.
2031  *
2032  * \param[in] mgc       MGC device by which to fetch the configuration log
2033  * \param[in] cld       log processing state (stored in lock callback data)
2034  *
2035  * \retval              0 on success
2036  * \retval              negative errno on failure
2037  */
2038 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
2039 {
2040         struct lustre_handle lockh = { 0 };
2041         __u64 flags = LDLM_FL_NO_LRU;
2042         int rc = 0, rcl;
2043         bool retry = false;
2044         ENTRY;
2045
2046         LASSERT(cld != NULL);
2047
2048         /* I don't want multiple processes running process_log at once --
2049            sounds like badness.  It actually might be fine, as long as
2050            we're not trying to update from the same log
2051            simultaneously (in which case we should use a per-log sem.) */
2052 restart:
2053         mutex_lock(&cld->cld_lock);
2054         if (cld->cld_stopping) {
2055                 mutex_unlock(&cld->cld_lock);
2056                 RETURN(0);
2057         }
2058
2059         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2060
2061         CDEBUG(D_MGC, "Process log %s-%016lx from %d\n", cld->cld_logname,
2062                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2063
2064         /* Get the cfg lock on the llog */
2065         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2066                           LCK_CR, &flags,
2067                           cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
2068                           cld, 0, NULL, &lockh);
2069         if (rcl == 0) {
2070                 /* Get the cld, it will be released in mgc_blocking_ast. */
2071                 config_log_get(cld);
2072                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2073                 LASSERT(!lustre_handle_is_used(&cld->cld_lockh));
2074                 LASSERT(rc == 0);
2075                 cld->cld_lockh = lockh;
2076         } else {
2077                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2078                 cld->cld_lockh.cookie = 0;
2079
2080                 if (rcl == -ESHUTDOWN &&
2081                     atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2082                         struct obd_import *imp;
2083                         long timeout = cfs_time_seconds(obd_timeout);
2084
2085                         mutex_unlock(&cld->cld_lock);
2086                         imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2087
2088                         /* Let's force the pinger, and wait the import to be
2089                          * connected, note: since mgc import is non-replayable,
2090                          * and even the import state is disconnected, it does
2091                          * not mean the "recovery" is stopped, so we will keep
2092                          * waitting until timeout or the import state is
2093                          * FULL or closed */
2094                         ptlrpc_pinger_force(imp);
2095
2096                         wait_event_idle_timeout(imp->imp_recovery_waitq,
2097                                                 !mgc_import_in_recovery(imp),
2098                                                 timeout);
2099
2100                         if (imp->imp_state == LUSTRE_IMP_FULL) {
2101                                 retry = true;
2102                                 goto restart;
2103                         } else {
2104                                 mutex_lock(&cld->cld_lock);
2105                                 /* unlock/lock mutex, so check stopping again */
2106                                 if (cld->cld_stopping) {
2107                                         mutex_unlock(&cld->cld_lock);
2108                                         RETURN(0);
2109                                 }
2110                                 spin_lock(&config_list_lock);
2111                                 cld->cld_lostlock = 1;
2112                                 spin_unlock(&config_list_lock);
2113                         }
2114                 } else {
2115                         /* mark cld_lostlock so that it will requeue
2116                          * after MGC becomes available. */
2117                         spin_lock(&config_list_lock);
2118                         cld->cld_lostlock = 1;
2119                         spin_unlock(&config_list_lock);
2120                 }
2121         }
2122
2123         if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2124                 if (!rcl)
2125                         rc = mgc_process_recover_nodemap_log(mgc, cld);
2126                 else if (cld_is_nodemap(cld))
2127                         rc = rcl;
2128
2129         } else if (!cld_is_barrier(cld)) {
2130                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2131         }
2132
2133         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2134                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2135
2136         /* Now drop the lock so MGS can revoke it */
2137         if (!rcl) {
2138                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2139                 if (rcl)
2140                         CERROR("Can't drop cfg lock: %d\n", rcl);
2141         }
2142         mutex_unlock(&cld->cld_lock);
2143
2144         /* requeue nodemap lock immediately if transfer was interrupted */
2145         if ((cld_is_nodemap(cld) && rc == -EAGAIN) ||
2146             (cld_is_recover(cld) && rc)) {
2147                 if (cld_is_recover(cld))
2148                         CWARN("%s: IR log %s failed, not fatal: rc = %d\n",
2149                               mgc->obd_name, cld->cld_logname, rc);
2150                 mgc_requeue_add(cld);
2151                 rc = 0;
2152         }
2153
2154         RETURN(rc);
2155 }
2156
2157
2158 /** Called from lustre_process_log.
2159  * LCFG_LOG_START gets the config log from the MGS, processes it to start
2160  * any services, and adds it to the list logs to watch (follow).
2161  */
2162 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2163 {
2164         struct lustre_cfg *lcfg = buf;
2165         struct config_llog_instance *cfg = NULL;
2166         char *logname;
2167         int rc = 0;
2168         ENTRY;
2169
2170         switch(lcfg->lcfg_command) {
2171         case LCFG_LOV_ADD_OBD: {
2172                 /* Overloading this cfg command: register a new target */
2173                 struct mgs_target_info *mti;
2174
2175                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2176                     sizeof(struct mgs_target_info))
2177                         GOTO(out, rc = -EINVAL);
2178
2179                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2180                 CDEBUG(D_MGC, "add_target %s %#x\n",
2181                        mti->mti_svname, mti->mti_flags);
2182                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2183                 break;
2184         }
2185         case LCFG_LOV_DEL_OBD:
2186                 /* Unregister has no meaning at the moment. */
2187                 CERROR("lov_del_obd unimplemented\n");
2188                 rc = -ENOSYS;
2189                 break;
2190         case LCFG_SPTLRPC_CONF: {
2191                 rc = sptlrpc_process_config(lcfg);
2192                 break;
2193         }
2194         case LCFG_LOG_START: {
2195                 struct config_llog_data *cld;
2196                 struct super_block *sb;
2197
2198                 logname = lustre_cfg_string(lcfg, 1);
2199                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2200                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2201
2202                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2203                        cfg->cfg_last_idx);
2204
2205                 /* We're only called through here on the initial mount */
2206                 cld = config_log_add(obd, logname, cfg, sb);
2207                 if (IS_ERR(cld)) {
2208                         rc = PTR_ERR(cld);
2209                         break;
2210                 }
2211
2212                 rc = mgc_process_log(obd, cld);
2213                 if (rc == 0 && cld->cld_recover != NULL) {
2214                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2215                                          imp_connect_data, IMP_RECOV)) {
2216                                 rc = mgc_process_log(obd, cld->cld_recover);
2217                         } else {
2218                                 struct config_llog_data *cir;
2219
2220                                 mutex_lock(&cld->cld_lock);
2221                                 cir = cld->cld_recover;
2222                                 cld->cld_recover = NULL;
2223                                 mutex_unlock(&cld->cld_lock);
2224                                 config_log_put(cir);
2225                         }
2226
2227                         if (rc)
2228                                 CERROR("Cannot process recover llog %d\n", rc);
2229                 }
2230
2231                 if (rc == 0 && cld->cld_params != NULL) {
2232                         rc = mgc_process_log(obd, cld->cld_params);
2233                         if (rc == -ENOENT) {
2234                                 CDEBUG(D_MGC, "There is no params "
2235                                               "config file yet\n");
2236                                 rc = 0;
2237                         }
2238                         /* params log is optional */
2239                         if (rc)
2240                                 CERROR("%s: can't process params llog: rc = %d\n",
2241                                        obd->obd_name, rc);
2242                 }
2243
2244                 break;
2245         }
2246         case LCFG_LOG_END: {
2247                 logname = lustre_cfg_string(lcfg, 1);
2248
2249                 if (lcfg->lcfg_bufcount >= 2)
2250                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
2251                                 lcfg, 2);
2252                 rc = config_log_end(logname, cfg);
2253                 break;
2254         }
2255         default: {
2256                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2257                 GOTO(out, rc = -EINVAL);
2258
2259         }
2260         }
2261 out:
2262         RETURN(rc);
2263 }
2264
2265 static const struct obd_ops mgc_obd_ops = {
2266         .o_owner        = THIS_MODULE,
2267         .o_setup        = mgc_setup,
2268         .o_precleanup   = mgc_precleanup,
2269         .o_cleanup      = mgc_cleanup,
2270         .o_add_conn     = client_import_add_conn,
2271         .o_del_conn     = client_import_del_conn,
2272         .o_connect      = client_connect_import,
2273         .o_disconnect   = client_disconnect_export,
2274         .o_set_info_async = mgc_set_info_async,
2275         .o_get_info       = mgc_get_info,
2276         .o_import_event = mgc_import_event,
2277         .o_process_config = mgc_process_config,
2278 };
2279
2280 static int __init mgc_init(void)
2281 {
2282         return class_register_type(&mgc_obd_ops, NULL, false, NULL,
2283                                    LUSTRE_MGC_NAME, NULL);
2284 }
2285
2286 static void __exit mgc_exit(void)
2287 {
2288         class_unregister_type(LUSTRE_MGC_NAME);
2289 }
2290
2291 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2292 MODULE_DESCRIPTION("Lustre Management Client");
2293 MODULE_VERSION(LUSTRE_VERSION_STRING);
2294 MODULE_LICENSE("GPL");
2295
2296 module_init(mgc_init);
2297 module_exit(mgc_exit);