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, 2016, 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
  43 #include <dt_object.h>
  44 #include <lprocfs_status.h>
  45 #include <lustre_dlm.h>
  46 #include <lustre_disk.h>
  47 #include <lustre_log.h>
  48 #include <lustre_nodemap.h>
  49 #include <lustre_swab.h>
  50 #include <obd_class.h>
  51 #include <lustre_barrier.h>
  52
  53 #include "mgc_internal.h"
  54
  55 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
  56                           int type)
  57 {
  58         __u64 resname = 0;
  59
  60         if (len > sizeof(resname)) {
  61                 CERROR("name too long: %s\n", name);
  62                 return -EINVAL;
  63         }
  64         if (len <= 0) {
  65                 CERROR("missing name: %s\n", name);
  66                 return -EINVAL;
  67         }
  68         memcpy(&resname, name, len);
  69
  70         /* Always use the same endianness for the resid */
  71         memset(res_id, 0, sizeof(*res_id));
  72         res_id->name[0] = cpu_to_le64(resname);
  73         /* XXX: unfortunately, sptlprc and config llog share one lock */
  74         switch(type) {
  75         case CONFIG_T_CONFIG:
  76         case CONFIG_T_SPTLRPC:
  77                 resname = 0;
  78                 break;
  79         case CONFIG_T_RECOVER:
  80         case CONFIG_T_PARAMS:
  81         case CONFIG_T_NODEMAP:
  82         case CONFIG_T_BARRIER:
  83                 resname = type;
  84                 break;
  85         default:
  86                 LBUG();
  87         }
  88         res_id->name[1] = cpu_to_le64(resname);
  89         CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
  90                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  91         return 0;
  92 }
  93
  94 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
  95 {
  96         /* fsname is at most 8 chars long, maybe contain "-".
  97          * e.g. "lustre", "SUN-000" */
  98         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
  99 }
 100 EXPORT_SYMBOL(mgc_fsname2resid);
 101
 102 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
 103 {
 104         char *name_end;
 105         int len;
 106
 107         /* logname consists of "fsname-nodetype".
 108          * e.g. "lustre-MDT0001", "SUN-000-client"
 109          * there is an exception: llog "params" */
 110         name_end = strrchr(logname, '-');
 111         if (!name_end)
 112                 len = strlen(logname);
 113         else
 114                 len = name_end - logname;
 115         return mgc_name2resid(logname, len, res_id, type);
 116 }
 117 EXPORT_SYMBOL(mgc_logname2resid);
 118
 119 /********************** config llog list **********************/
 120 static struct list_head config_llog_list = LIST_HEAD_INIT(config_llog_list);
 121 static DEFINE_SPINLOCK(config_list_lock);
 122
 123 /* Take a reference to a config log */
 124 static int config_log_get(struct config_llog_data *cld)
 125 {
 126         ENTRY;
 127         atomic_inc(&cld->cld_refcount);
 128         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 129                 atomic_read(&cld->cld_refcount));
 130         RETURN(0);
 131 }
 132
 133 /* Drop a reference to a config log.  When no longer referenced,
 134    we can free the config log data */
 135 static void config_log_put(struct config_llog_data *cld)
 136 {
 137         ENTRY;
 138
 139         if (unlikely(!cld))
 140                 RETURN_EXIT;
 141
 142         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 143                 atomic_read(&cld->cld_refcount));
 144         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 145
 146         /* spinlock to make sure no item with 0 refcount in the list */
 147         if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 148                 list_del(&cld->cld_list_chain);
 149                 spin_unlock(&config_list_lock);
 150
 151                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 152
 153                 config_log_put(cld->cld_barrier);
 154                 config_log_put(cld->cld_recover);
 155                 config_log_put(cld->cld_params);
 156                 config_log_put(cld->cld_nodemap);
 157                 config_log_put(cld->cld_sptlrpc);
 158                 if (cld_is_sptlrpc(cld))
 159                         sptlrpc_conf_log_stop(cld->cld_logname);
 160
 161                 class_export_put(cld->cld_mgcexp);
 162                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 163         }
 164
 165         EXIT;
 166 }
 167
 168 /* Find a config log by name */
 169 static
 170 struct config_llog_data *config_log_find(char *logname,
 171                                          struct config_llog_instance *cfg)
 172 {
 173         struct config_llog_data *cld;
 174         struct config_llog_data *found = NULL;
 175         void *                   instance;
 176         ENTRY;
 177
 178         LASSERT(logname != NULL);
 179
 180         instance = cfg ? cfg->cfg_instance : NULL;
 181         spin_lock(&config_list_lock);
 182         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 183                 /* check if instance equals */
 184                 if (instance != cld->cld_cfg.cfg_instance)
 185                         continue;
 186
 187                 /* instance may be NULL, should check name */
 188                 if (strcmp(logname, cld->cld_logname) == 0) {
 189                         found = cld;
 190                         config_log_get(found);
 191                         break;
 192                 }
 193         }
 194         spin_unlock(&config_list_lock);
 195         RETURN(found);
 196 }
 197
 198 static
 199 struct config_llog_data *do_config_log_add(struct obd_device *obd,
 200                                            char *logname,
 201                                            int type,
 202                                            struct config_llog_instance *cfg,
 203                                            struct super_block *sb)
 204 {
 205         struct config_llog_data *cld;
 206         int rc;
 207
 208         ENTRY;
 209
 210         CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
 211                cfg ? cfg->cfg_instance : NULL);
 212
 213         OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
 214         if (!cld)
 215                 RETURN(ERR_PTR(-ENOMEM));
 216
 217         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 218         if (rc) {
 219                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 220                 RETURN(ERR_PTR(rc));
 221         }
 222
 223         strcpy(cld->cld_logname, logname);
 224         if (cfg)
 225                 cld->cld_cfg = *cfg;
 226         else
 227                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 228         mutex_init(&cld->cld_lock);
 229         cld->cld_cfg.cfg_last_idx = 0;
 230         cld->cld_cfg.cfg_flags = 0;
 231         cld->cld_cfg.cfg_sb = sb;
 232         cld->cld_type = type;
 233         atomic_set(&cld->cld_refcount, 1);
 234
 235         /* Keep the mgc around until we are done */
 236         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 237
 238         if (cld_is_sptlrpc(cld)) {
 239                 sptlrpc_conf_log_start(logname);
 240                 cld->cld_cfg.cfg_obdname = obd->obd_name;
 241         }
 242
 243         spin_lock(&config_list_lock);
 244         list_add(&cld->cld_list_chain, &config_llog_list);
 245         spin_unlock(&config_list_lock);
 246
 247         if (cld_is_sptlrpc(cld) || cld_is_nodemap(cld) || cld_is_barrier(cld)) {
 248                 rc = mgc_process_log(obd, cld);
 249                 if (rc && rc != -ENOENT)
 250                         CERROR("%s: failed processing log, type %d: rc = %d\n",
 251                                obd->obd_name, type, rc);
 252         }
 253
 254         RETURN(cld);
 255 }
 256
 257 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
 258         char *fsname,
 259         struct config_llog_instance *cfg,
 260         struct super_block *sb)
 261 {
 262         struct config_llog_instance lcfg = *cfg;
 263         struct lustre_sb_info *lsi = s2lsi(sb);
 264         struct config_llog_data *cld;
 265         char logname[32];
 266
 267         if (IS_OST(lsi))
 268                 return NULL;
 269
 270         /* for osp-on-ost, see lustre_start_osp() */
 271         if (IS_MDT(lsi) && lcfg.cfg_instance)
 272                 return NULL;
 273
 274         /* we have to use different llog for clients and mdts for cmd
 275          * where only clients are notified if one of cmd server restarts */
 276         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 277         strcpy(logname, fsname);
 278         if (IS_SERVER(lsi)) { /* mdt */
 279                 LASSERT(lcfg.cfg_instance == NULL);
 280                 lcfg.cfg_instance = sb;
 281                 strcat(logname, "-mdtir");
 282         } else {
 283                 LASSERT(lcfg.cfg_instance != NULL);
 284                 strcat(logname, "-cliir");
 285         }
 286
 287         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
 288         return cld;
 289 }
 290
 291 static struct config_llog_data *config_log_find_or_add(struct obd_device *obd,
 292                                 char *logname, struct super_block *sb, int type,
 293                                 struct config_llog_instance *cfg)
 294 {
 295         struct config_llog_instance     lcfg = *cfg;
 296         struct config_llog_data         *cld;
 297
 298         lcfg.cfg_instance = sb != NULL ? (void *)sb : (void *)obd;
 299
 300         if (type == CONFIG_T_SPTLRPC)
 301                 lcfg.cfg_instance = NULL;
 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, "adding config log %s:%p\n", logname, cfg->cfg_instance);
 332
 333         /*
 334          * for each regular log, the depended sptlrpc log name is
 335          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 336          */
 337         ptr = strrchr(logname, '-');
 338         if (ptr == NULL || ptr - logname > 8) {
 339                 CERROR("logname %s is too long\n", logname);
 340                 RETURN(ERR_PTR(-EINVAL));
 341         }
 342
 343         memcpy(seclogname, logname, ptr - logname);
 344         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 345
 346         if (cfg->cfg_sub_clds & CONFIG_SUB_SPTLRPC) {
 347                 sptlrpc_cld = config_log_find_or_add(obd, seclogname, NULL,
 348                                                      CONFIG_T_SPTLRPC, cfg);
 349                 if (IS_ERR(sptlrpc_cld)) {
 350                         CERROR("%s: can't create sptlrpc log %s: rc = %ld\n",
 351                                obd->obd_name, seclogname, PTR_ERR(sptlrpc_cld));
 352                         RETURN(sptlrpc_cld);
 353                 }
 354         }
 355
 356         if (!IS_MGS(lsi) && cfg->cfg_sub_clds & CONFIG_SUB_NODEMAP) {
 357                 nodemap_cld = config_log_find_or_add(obd, LUSTRE_NODEMAP_NAME,
 358                                                      NULL, CONFIG_T_NODEMAP,
 359                                                      cfg);
 360                 if (IS_ERR(nodemap_cld)) {
 361                         rc = PTR_ERR(nodemap_cld);
 362                         CERROR("%s: cannot create nodemap log: rc = %d\n",
 363                                obd->obd_name, rc);
 364                         GOTO(out_sptlrpc, rc);
 365                 }
 366         }
 367
 368         if (cfg->cfg_sub_clds & CONFIG_SUB_PARAMS) {
 369                 params_cld = config_log_find_or_add(obd, PARAMS_FILENAME, sb,
 370                                                     CONFIG_T_PARAMS, cfg);
 371                 if (IS_ERR(params_cld)) {
 372                         rc = PTR_ERR(params_cld);
 373                         CERROR("%s: can't create params log: rc = %d\n",
 374                                obd->obd_name, rc);
 375                         GOTO(out_nodemap, rc);
 376                 }
 377         }
 378
 379         if (IS_MDT(s2lsi(sb)) && cfg->cfg_sub_clds & CONFIG_SUB_BARRIER) {
 380                 snprintf(seclogname + (ptr - logname), sizeof(seclogname) - 1,
 381                          "-%s", BARRIER_FILENAME);
 382                 barrier_cld = config_log_find_or_add(obd, seclogname, sb,
 383                                                      CONFIG_T_BARRIER, cfg);
 384                 if (IS_ERR(barrier_cld)) {
 385                         rc = PTR_ERR(barrier_cld);
 386                         CERROR("%s: can't create barrier log: rc = %d\n",
 387                                obd->obd_name, rc);
 388                         GOTO(out_params, rc);
 389                 }
 390         }
 391
 392         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 393         if (IS_ERR(cld)) {
 394                 rc = PTR_ERR(cld);
 395                 CERROR("%s: can't create log: rc = %d\n",
 396                        obd->obd_name, rc);
 397                 GOTO(out_barrier, rc = PTR_ERR(cld));
 398         }
 399
 400         LASSERT(lsi->lsi_lmd);
 401         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR) &&
 402             cfg->cfg_sub_clds & CONFIG_SUB_RECOVER) {
 403                 struct config_llog_data *recover_cld;
 404
 405                 ptr = strrchr(seclogname, '-');
 406                 if (ptr != NULL) {
 407                         *ptr = 0;
 408                 } else {
 409                         CERROR("%s: sptlrpc log name not correct, %s: "
 410                                "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
 411                         GOTO(out_cld, rc = -EINVAL);
 412                 }
 413
 414                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 415                 if (IS_ERR(recover_cld)) {
 416                         rc = PTR_ERR(recover_cld);
 417                         CERROR("%s: can't create recover log: rc = %d\n",
 418                                obd->obd_name, rc);
 419                         GOTO(out_cld, rc);
 420                 }
 421
 422                 mutex_lock(&cld->cld_lock);
 423                 locked = true;
 424                 cld->cld_recover = recover_cld;
 425         }
 426
 427         if (!locked)
 428                 mutex_lock(&cld->cld_lock);
 429         cld->cld_params = params_cld;
 430         cld->cld_barrier = barrier_cld;
 431         cld->cld_nodemap = nodemap_cld;
 432         cld->cld_sptlrpc = sptlrpc_cld;
 433         mutex_unlock(&cld->cld_lock);
 434
 435         RETURN(cld);
 436
 437 out_cld:
 438         config_log_put(cld);
 439 out_barrier:
 440         config_log_put(barrier_cld);
 441 out_params:
 442         config_log_put(params_cld);
 443 out_nodemap:
 444         config_log_put(nodemap_cld);
 445 out_sptlrpc:
 446         config_log_put(sptlrpc_cld);
 447
 448         return ERR_PTR(rc);
 449 }
 450
 451 DEFINE_MUTEX(llog_process_lock);
 452
 453 static inline void config_mark_cld_stop(struct config_llog_data *cld)
 454 {
 455         if (cld) {
 456                 mutex_lock(&cld->cld_lock);
 457                 spin_lock(&config_list_lock);
 458                 cld->cld_stopping = 1;
 459                 spin_unlock(&config_list_lock);
 460                 mutex_unlock(&cld->cld_lock);
 461         }
 462 }
 463
 464 /** Stop watching for updates on this log.
 465  */
 466 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 467 {
 468         struct config_llog_data *cld;
 469         struct config_llog_data *cld_sptlrpc = NULL;
 470         struct config_llog_data *cld_params = NULL;
 471         struct config_llog_data *cld_recover = NULL;
 472         struct config_llog_data *cld_nodemap = NULL;
 473         struct config_llog_data *cld_barrier = NULL;
 474         int rc = 0;
 475
 476         ENTRY;
 477
 478         cld = config_log_find(logname, cfg);
 479         if (cld == NULL)
 480                 RETURN(-ENOENT);
 481
 482         mutex_lock(&cld->cld_lock);
 483         /*
 484          * if cld_stopping is set, it means we didn't start the log thus
 485          * not owning the start ref. this can happen after previous umount:
 486          * the cld still hanging there waiting for lock cancel, and we
 487          * remount again but failed in the middle and call log_end without
 488          * calling start_log.
 489          */
 490         if (unlikely(cld->cld_stopping)) {
 491                 mutex_unlock(&cld->cld_lock);
 492                 /* drop the ref from the find */
 493                 config_log_put(cld);
 494                 RETURN(rc);
 495         }
 496
 497         spin_lock(&config_list_lock);
 498         cld->cld_stopping = 1;
 499         spin_unlock(&config_list_lock);
 500
 501         cld_recover = cld->cld_recover;
 502         cld->cld_recover = NULL;
 503         cld_params = cld->cld_params;
 504         cld->cld_params = NULL;
 505         cld_nodemap = cld->cld_nodemap;
 506         cld->cld_nodemap = NULL;
 507         cld_barrier = cld->cld_barrier;
 508         cld->cld_barrier = NULL;
 509         cld_sptlrpc = cld->cld_sptlrpc;
 510         cld->cld_sptlrpc = NULL;
 511         mutex_unlock(&cld->cld_lock);
 512
 513         config_mark_cld_stop(cld_recover);
 514         config_log_put(cld_recover);
 515
 516         config_mark_cld_stop(cld_params);
 517         config_log_put(cld_params);
 518
 519         /* don't set cld_stopping on nm lock as other targets may be active */
 520         config_log_put(cld_nodemap);
 521
 522         if (cld_barrier) {
 523                 mutex_lock(&cld_barrier->cld_lock);
 524                 cld_barrier->cld_stopping = 1;
 525                 mutex_unlock(&cld_barrier->cld_lock);
 526                 config_log_put(cld_barrier);
 527         }
 528
 529         config_log_put(cld_sptlrpc);
 530
 531         /* drop the ref from the find */
 532         config_log_put(cld);
 533         /* drop the start ref */
 534         config_log_put(cld);
 535
 536         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 537                rc);
 538         RETURN(rc);
 539 }
 540
 541 #ifdef CONFIG_PROC_FS
 542 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 543 {
 544         struct obd_device       *obd = data;
 545         struct obd_import       *imp;
 546         struct obd_connect_data *ocd;
 547         struct config_llog_data *cld;
 548         ENTRY;
 549
 550         LASSERT(obd != NULL);
 551         LPROCFS_CLIMP_CHECK(obd);
 552         imp = obd->u.cli.cl_import;
 553         ocd = &imp->imp_connect_data;
 554
 555         seq_printf(m, "imperative_recovery: %s\n",
 556                    OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 557         seq_printf(m, "client_state:\n");
 558
 559         spin_lock(&config_list_lock);
 560         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 561                 if (cld->cld_recover == NULL)
 562                         continue;
 563                 seq_printf(m,  "    - { client: %s, nidtbl_version: %u }\n",
 564                            cld->cld_logname,
 565                            cld->cld_recover->cld_cfg.cfg_last_idx);
 566         }
 567         spin_unlock(&config_list_lock);
 568
 569         LPROCFS_CLIMP_EXIT(obd);
 570         RETURN(0);
 571 }
 572 #endif
 573
 574 /* reenqueue any lost locks */
 575 #define RQ_RUNNING      0x1
 576 #define RQ_NOW          0x2
 577 #define RQ_LATER        0x4
 578 #define RQ_STOP         0x8
 579 #define RQ_PRECLEANUP   0x10
 580 static int                    rq_state = 0;
 581 static wait_queue_head_t      rq_waitq;
 582 static DECLARE_COMPLETION(rq_exit);
 583 static DECLARE_COMPLETION(rq_start);
 584
 585 static void do_requeue(struct config_llog_data *cld)
 586 {
 587         int rc = 0;
 588         ENTRY;
 589
 590         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 591
 592         /*
 593          * Do not run mgc_process_log on a disconnected export or an
 594          * export which is being disconnected. Take the client
 595          * semaphore to make the check non-racy.
 596          */
 597         down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
 598                          OBD_CLI_SEM_MGC);
 599         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 600                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 601                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 602                 if (rc && rc != -ENOENT)
 603                         CERROR("failed processing log: %d\n", rc);
 604         } else {
 605                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 606                        cld->cld_logname);
 607         }
 608         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 609
 610         EXIT;
 611 }
 612
 613 /* this timeout represents how many seconds MGC should wait before
 614  * requeue config and recover lock to the MGS. We need to randomize this
 615  * in order to not flood the MGS.
 616  */
 617 #define MGC_TIMEOUT_MIN_SECONDS   5
 618 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
 619
 620 static int mgc_requeue_thread(void *data)
 621 {
 622         int rc = 0;
 623         bool first = true;
 624         ENTRY;
 625
 626         CDEBUG(D_MGC, "Starting requeue thread\n");
 627
 628         /* Keep trying failed locks periodically */
 629         spin_lock(&config_list_lock);
 630         rq_state |= RQ_RUNNING;
 631         while (!(rq_state & RQ_STOP)) {
 632                 struct l_wait_info lwi;
 633                 struct config_llog_data *cld, *cld_prev;
 634                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
 635                 int to;
 636
 637                 /* Any new or requeued lostlocks will change the state */
 638                 rq_state &= ~(RQ_NOW | RQ_LATER);
 639                 spin_unlock(&config_list_lock);
 640
 641                 if (first) {
 642                         first = false;
 643                         complete(&rq_start);
 644                 }
 645
 646                 /* Always wait a few seconds to allow the server who
 647                    caused the lock revocation to finish its setup, plus some
 648                    random so everyone doesn't try to reconnect at once. */
 649                 to = msecs_to_jiffies(MGC_TIMEOUT_MIN_SECONDS * MSEC_PER_SEC);
 650                 /* rand is centi-seconds */
 651                 to += msecs_to_jiffies(rand * MSEC_PER_SEC / 100);
 652                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 653                 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
 654                              &lwi);
 655
 656                 /*
 657                  * iterate & processing through the list. for each cld, process
 658                  * its depending sptlrpc cld firstly (if any) and then itself.
 659                  *
 660                  * it's guaranteed any item in the list must have
 661                  * reference > 0; and if cld_lostlock is set, at
 662                  * least one reference is taken by the previous enqueue.
 663                  */
 664                 cld_prev = NULL;
 665
 666                 spin_lock(&config_list_lock);
 667                 rq_state &= ~RQ_PRECLEANUP;
 668                 list_for_each_entry(cld, &config_llog_list,
 669                                     cld_list_chain) {
 670                         if (!cld->cld_lostlock || cld->cld_stopping)
 671                                 continue;
 672
 673                         /* hold reference to avoid being freed during
 674                          * subsequent processing. */
 675                         config_log_get(cld);
 676                         cld->cld_lostlock = 0;
 677                         spin_unlock(&config_list_lock);
 678
 679                         config_log_put(cld_prev);
 680                         cld_prev = cld;
 681
 682                         if (likely(!(rq_state & RQ_STOP))) {
 683                                 do_requeue(cld);
 684                                 spin_lock(&config_list_lock);
 685                         } else {
 686                                 spin_lock(&config_list_lock);
 687                                 break;
 688                         }
 689                 }
 690                 spin_unlock(&config_list_lock);
 691                 config_log_put(cld_prev);
 692
 693                 /* Wait a bit to see if anyone else needs a requeue */
 694                 lwi = (struct l_wait_info) { 0 };
 695                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 696                              &lwi);
 697                 spin_lock(&config_list_lock);
 698         }
 699
 700         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 701         rq_state &= ~RQ_RUNNING;
 702         spin_unlock(&config_list_lock);
 703
 704         complete(&rq_exit);
 705
 706         CDEBUG(D_MGC, "Ending requeue thread\n");
 707         RETURN(rc);
 708 }
 709
 710 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 711    We are responsible for dropping the config log reference from here on out. */
 712 static void mgc_requeue_add(struct config_llog_data *cld)
 713 {
 714         bool wakeup = false;
 715         ENTRY;
 716
 717         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 718                 cld->cld_logname, atomic_read(&cld->cld_refcount),
 719                 cld->cld_stopping, rq_state);
 720         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 721
 722         mutex_lock(&cld->cld_lock);
 723         spin_lock(&config_list_lock);
 724         if (!(rq_state & RQ_STOP) && !cld->cld_stopping && !cld->cld_lostlock) {
 725                 cld->cld_lostlock = 1;
 726                 rq_state |= RQ_NOW;
 727                 wakeup = true;
 728         }
 729         spin_unlock(&config_list_lock);
 730         mutex_unlock(&cld->cld_lock);
 731         if (wakeup)
 732                 wake_up(&rq_waitq);
 733
 734         EXIT;
 735 }
 736
 737 /********************** class fns **********************/
 738 static int mgc_local_llog_init(const struct lu_env *env,
 739                                struct obd_device *obd,
 740                                struct obd_device *disk)
 741 {
 742         struct llog_ctxt        *ctxt;
 743         int                      rc;
 744
 745         ENTRY;
 746
 747         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
 748                         &llog_osd_ops);
 749         if (rc)
 750                 RETURN(rc);
 751
 752         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 753         LASSERT(ctxt);
 754         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
 755         llog_ctxt_put(ctxt);
 756
 757         RETURN(0);
 758 }
 759
 760 static int mgc_local_llog_fini(const struct lu_env *env,
 761                                struct obd_device *obd)
 762 {
 763         struct llog_ctxt *ctxt;
 764
 765         ENTRY;
 766
 767         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 768         llog_cleanup(env, ctxt);
 769
 770         RETURN(0);
 771 }
 772
 773 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
 774                         struct super_block *sb)
 775 {
 776         struct lustre_sb_info   *lsi = s2lsi(sb);
 777         struct client_obd       *cli = &obd->u.cli;
 778         struct lu_fid            rfid, fid;
 779         struct dt_object        *root, *dto;
 780         int                      rc = 0;
 781
 782         ENTRY;
 783
 784         LASSERT(lsi);
 785         LASSERT(lsi->lsi_dt_dev);
 786
 787         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
 788         mutex_lock(&cli->cl_mgc_mutex);
 789
 790         /* Setup the configs dir */
 791         fid.f_seq = FID_SEQ_LOCAL_NAME;
 792         fid.f_oid = 1;
 793         fid.f_ver = 0;
 794         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
 795                                     &cli->cl_mgc_los);
 796         if (rc)
 797                 RETURN(rc);
 798
 799         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
 800         if (rc)
 801                 GOTO(out_los, rc);
 802
 803         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
 804                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
 805         if (unlikely(IS_ERR(root)))
 806                 GOTO(out_los, rc = PTR_ERR(root));
 807
 808         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
 809                                         MOUNT_CONFIGS_DIR,
 810                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
 811         dt_object_put_nocache(env, root);
 812         if (IS_ERR(dto))
 813                 GOTO(out_los, rc = PTR_ERR(dto));
 814
 815         cli->cl_mgc_configs_dir = dto;
 816
 817         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
 818         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
 819         if (rc)
 820                 GOTO(out_llog, rc);
 821
 822         /* We take an obd ref to insure that we can't get to mgc_cleanup
 823          * without calling mgc_fs_cleanup first. */
 824         class_incref(obd, "mgc_fs", obd);
 825
 826         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 827         EXIT;
 828 out_llog:
 829         if (rc) {
 830                 dt_object_put(env, cli->cl_mgc_configs_dir);
 831                 cli->cl_mgc_configs_dir = NULL;
 832         }
 833 out_los:
 834         if (rc < 0) {
 835                 local_oid_storage_fini(env, cli->cl_mgc_los);
 836                 cli->cl_mgc_los = NULL;
 837                 mutex_unlock(&cli->cl_mgc_mutex);
 838         }
 839         return rc;
 840 }
 841
 842 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
 843 {
 844         struct client_obd       *cli = &obd->u.cli;
 845         ENTRY;
 846
 847         LASSERT(cli->cl_mgc_los != NULL);
 848
 849         mgc_local_llog_fini(env, obd);
 850
 851         dt_object_put_nocache(env, cli->cl_mgc_configs_dir);
 852         cli->cl_mgc_configs_dir = NULL;
 853
 854         local_oid_storage_fini(env, cli->cl_mgc_los);
 855         cli->cl_mgc_los = NULL;
 856
 857         class_decref(obd, "mgc_fs", obd);
 858         mutex_unlock(&cli->cl_mgc_mutex);
 859
 860         RETURN(0);
 861 }
 862
 863 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 864 {
 865         struct llog_ctxt        *ctxt;
 866         int                      rc;
 867
 868         ENTRY;
 869
 870         /* setup only remote ctxt, the local disk context is switched per each
 871          * filesystem during mgc_fs_setup() */
 872         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 873                         &llog_client_ops);
 874         if (rc)
 875                 RETURN(rc);
 876
 877         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 878         LASSERT(ctxt);
 879
 880         llog_initiator_connect(ctxt);
 881         llog_ctxt_put(ctxt);
 882
 883         RETURN(0);
 884 }
 885
 886 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 887 {
 888         struct llog_ctxt *ctxt;
 889
 890         ENTRY;
 891
 892         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 893         if (ctxt)
 894                 llog_cleanup(env, ctxt);
 895
 896         RETURN(0);
 897 }
 898
 899
 900 static atomic_t mgc_count = ATOMIC_INIT(0);
 901 static int mgc_precleanup(struct obd_device *obd)
 902 {
 903         int     rc = 0;
 904         int     temp;
 905         ENTRY;
 906
 907         if (atomic_dec_and_test(&mgc_count)) {
 908                 LASSERT(rq_state & RQ_RUNNING);
 909                 /* stop requeue thread */
 910                 temp = RQ_STOP;
 911         } else {
 912                 /* wakeup requeue thread to clean our cld */
 913                 temp = RQ_NOW | RQ_PRECLEANUP;
 914         }
 915
 916         spin_lock(&config_list_lock);
 917         rq_state |= temp;
 918         spin_unlock(&config_list_lock);
 919         wake_up(&rq_waitq);
 920
 921         if (temp & RQ_STOP)
 922                 wait_for_completion(&rq_exit);
 923         obd_cleanup_client_import(obd);
 924
 925         rc = mgc_llog_fini(NULL, obd);
 926         if (rc != 0)
 927                 CERROR("failed to cleanup llogging subsystems\n");
 928
 929         RETURN(rc);
 930 }
 931
 932 static int mgc_cleanup(struct obd_device *obd)
 933 {
 934         int rc;
 935         ENTRY;
 936
 937         /* COMPAT_146 - old config logs may have added profiles we don't
 938            know about */
 939         if (obd->obd_type->typ_refcnt <= 1)
 940                 /* Only for the last mgc */
 941                 class_del_profiles();
 942
 943         lprocfs_obd_cleanup(obd);
 944         ptlrpcd_decref();
 945
 946         rc = client_obd_cleanup(obd);
 947         RETURN(rc);
 948 }
 949
 950 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 951 {
 952         struct task_struct      *task;
 953         int                      rc;
 954         ENTRY;
 955
 956         rc = ptlrpcd_addref();
 957         if (rc < 0)
 958                 RETURN(rc);
 959
 960         rc = client_obd_setup(obd, lcfg);
 961         if (rc)
 962                 GOTO(err_decref, rc);
 963
 964         rc = mgc_llog_init(NULL, obd);
 965         if (rc) {
 966                 CERROR("failed to setup llogging subsystems\n");
 967                 GOTO(err_cleanup, rc);
 968         }
 969
 970 #ifdef CONFIG_PROC_FS
 971         obd->obd_vars = lprocfs_mgc_obd_vars;
 972         lprocfs_obd_setup(obd, true);
 973 #endif
 974         sptlrpc_lprocfs_cliobd_attach(obd);
 975
 976         if (atomic_inc_return(&mgc_count) == 1) {
 977                 rq_state = 0;
 978                 init_waitqueue_head(&rq_waitq);
 979
 980                 /* start requeue thread */
 981                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 982                 if (IS_ERR(task)) {
 983                         rc = PTR_ERR(task);
 984                         CERROR("%s: cannot start requeue thread: rc = %d; "
 985                                "no more log updates\n",
 986                                obd->obd_name, rc);
 987                         GOTO(err_cleanup, rc);
 988                 }
 989                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 990                 rc = 0;
 991                 wait_for_completion(&rq_start);
 992         }
 993
 994         RETURN(rc);
 995
 996 err_cleanup:
 997         client_obd_cleanup(obd);
 998 err_decref:
 999         ptlrpcd_decref();
1000         RETURN(rc);
1001 }
1002
1003 /* based on ll_mdc_blocking_ast */
1004 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1005                             void *data, int flag)
1006 {
1007         struct lustre_handle lockh;
1008         struct config_llog_data *cld = (struct config_llog_data *)data;
1009         int rc = 0;
1010         ENTRY;
1011
1012         switch (flag) {
1013         case LDLM_CB_BLOCKING:
1014                 /* mgs wants the lock, give it up... */
1015                 LDLM_DEBUG(lock, "MGC blocking CB");
1016                 ldlm_lock2handle(lock, &lockh);
1017                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
1018                 break;
1019         case LDLM_CB_CANCELING:
1020                 /* We've given up the lock, prepare ourselves to update. */
1021                 LDLM_DEBUG(lock, "MGC cancel CB");
1022
1023                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
1024                        PLDLMRES(lock->l_resource),
1025                        (char *)&lock->l_resource->lr_name.name[0]);
1026
1027                 if (!cld) {
1028                         CDEBUG(D_INFO, "missing data, won't requeue\n");
1029                         break;
1030                 }
1031
1032                 /* held at mgc_process_log(). */
1033                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1034
1035                 lock->l_ast_data = NULL;
1036                 /* Are we done with this log? */
1037                 if (cld->cld_stopping) {
1038                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1039                                 cld->cld_logname);
1040                         config_log_put(cld);
1041                         break;
1042                 }
1043                 /* Make sure not to re-enqueue when the mgc is stopping
1044                    (we get called from client_disconnect_export) */
1045                 if (lock->l_conn_export == NULL ||
1046                     lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1047                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1048                                 cld->cld_logname);
1049                         config_log_put(cld);
1050                         break;
1051                 }
1052
1053                 /* Re-enqueue now */
1054                 mgc_requeue_add(cld);
1055                 config_log_put(cld);
1056                 break;
1057         default:
1058                 LBUG();
1059         }
1060
1061         RETURN(rc);
1062 }
1063
1064 /* Not sure where this should go... */
1065 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1066  * that we can have a chance to try the secondary MGS if any. */
1067 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1068                                 + PING_INTERVAL)
1069 #define  MGC_TARGET_REG_LIMIT 10
1070 #define  MGC_SEND_PARAM_LIMIT 10
1071
1072 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1073 /* Send parameter to MGS*/
1074 static int mgc_set_mgs_param(struct obd_export *exp,
1075                              struct mgs_send_param *msp)
1076 {
1077         struct ptlrpc_request *req;
1078         struct mgs_send_param *req_msp, *rep_msp;
1079         int rc;
1080         ENTRY;
1081
1082         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1083                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1084                                         MGS_SET_INFO);
1085         if (!req)
1086                 RETURN(-ENOMEM);
1087
1088         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1089         if (!req_msp) {
1090                 ptlrpc_req_finished(req);
1091                 RETURN(-ENOMEM);
1092         }
1093
1094         memcpy(req_msp, msp, sizeof(*req_msp));
1095         ptlrpc_request_set_replen(req);
1096
1097         /* Limit how long we will wait for the enqueue to complete */
1098         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1099         rc = ptlrpc_queue_wait(req);
1100         if (!rc) {
1101                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1102                 memcpy(msp, rep_msp, sizeof(*rep_msp));
1103         }
1104
1105         ptlrpc_req_finished(req);
1106
1107         RETURN(rc);
1108 }
1109 #endif
1110
1111 /* Take a config lock so we can get cancel notifications */
1112 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1113                        union ldlm_policy_data *policy, enum ldlm_mode mode,
1114                        __u64 *flags, ldlm_glimpse_callback glimpse_callback,
1115                        void *data, __u32 lvb_len, void *lvb_swabber,
1116                        struct lustre_handle *lockh)
1117 {
1118         struct config_llog_data *cld = (struct config_llog_data *)data;
1119         struct ldlm_enqueue_info einfo = {
1120                 .ei_type        = type,
1121                 .ei_mode        = mode,
1122                 .ei_cb_bl       = mgc_blocking_ast,
1123                 .ei_cb_cp       = ldlm_completion_ast,
1124                 .ei_cb_gl       = glimpse_callback,
1125         };
1126         struct ptlrpc_request *req;
1127         int short_limit = cld_is_sptlrpc(cld);
1128         int rc;
1129         ENTRY;
1130
1131         CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1132                cld->cld_resid.name[0]);
1133
1134         /* We need a callback for every lockholder, so don't try to
1135            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1136         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1137                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1138                                         LDLM_ENQUEUE);
1139         if (req == NULL)
1140                 RETURN(-ENOMEM);
1141
1142         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1143         ptlrpc_request_set_replen(req);
1144
1145         /* check if this is server or client */
1146         if (cld->cld_cfg.cfg_sb) {
1147                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1148                 if (lsi && IS_SERVER(lsi))
1149                         short_limit = 1;
1150         }
1151         /* Limit how long we will wait for the enqueue to complete */
1152         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1153         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1154                               NULL, 0, LVB_T_NONE, lockh, 0);
1155         /* A failed enqueue should still call the mgc_blocking_ast,
1156            where it will be requeued if needed ("grant failed"). */
1157         ptlrpc_req_finished(req);
1158         RETURN(rc);
1159 }
1160
1161 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1162                       struct lustre_handle *lockh)
1163 {
1164         ENTRY;
1165
1166         ldlm_lock_decref(lockh, mode);
1167
1168         RETURN(0);
1169 }
1170
1171 static void mgc_notify_active(struct obd_device *unused)
1172 {
1173         /* wakeup mgc_requeue_thread to requeue mgc lock */
1174         spin_lock(&config_list_lock);
1175         rq_state |= RQ_NOW;
1176         spin_unlock(&config_list_lock);
1177         wake_up(&rq_waitq);
1178
1179         /* TODO: Help the MGS rebuild nidtbl. -jay */
1180 }
1181
1182 /* Send target_reg message to MGS */
1183 static int mgc_target_register(struct obd_export *exp,
1184                                struct mgs_target_info *mti)
1185 {
1186         struct ptlrpc_request  *req;
1187         struct mgs_target_info *req_mti, *rep_mti;
1188         int                     rc;
1189         ENTRY;
1190
1191         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1192                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1193                                         MGS_TARGET_REG);
1194         if (req == NULL)
1195                 RETURN(-ENOMEM);
1196
1197         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1198         if (!req_mti) {
1199                 ptlrpc_req_finished(req);
1200                 RETURN(-ENOMEM);
1201         }
1202
1203         memcpy(req_mti, mti, sizeof(*req_mti));
1204         ptlrpc_request_set_replen(req);
1205         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1206         /* Limit how long we will wait for the enqueue to complete */
1207         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1208
1209         rc = ptlrpc_queue_wait(req);
1210         if (!rc) {
1211                 rep_mti = req_capsule_server_get(&req->rq_pill,
1212                                                  &RMF_MGS_TARGET_INFO);
1213                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1214                 CDEBUG(D_MGC, "register %s got index = %d\n",
1215                        mti->mti_svname, mti->mti_stripe_index);
1216         }
1217         ptlrpc_req_finished(req);
1218
1219         RETURN(rc);
1220 }
1221
1222 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1223                               u32 keylen, void *key,
1224                               u32 vallen, void *val,
1225                               struct ptlrpc_request_set *set)
1226 {
1227         int rc = -EINVAL;
1228         ENTRY;
1229
1230         /* Turn off initial_recov after we try all backup servers once */
1231         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1232                 struct obd_import *imp = class_exp2cliimp(exp);
1233                 int value;
1234                 if (vallen != sizeof(int))
1235                         RETURN(-EINVAL);
1236                 value = *(int *)val;
1237                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1238                        imp->imp_obd->obd_name, value,
1239                        imp->imp_deactive, imp->imp_invalid,
1240                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1241                        ptlrpc_import_state_name(imp->imp_state));
1242                 /* Resurrect if we previously died */
1243                 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1244                      imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1245                         ptlrpc_reconnect_import(imp);
1246                 RETURN(0);
1247         }
1248         /* FIXME move this to mgc_process_config */
1249         if (KEY_IS(KEY_REGISTER_TARGET)) {
1250                 struct mgs_target_info *mti;
1251                 if (vallen != sizeof(struct mgs_target_info))
1252                         RETURN(-EINVAL);
1253                 mti = (struct mgs_target_info *)val;
1254                 CDEBUG(D_MGC, "register_target %s %#x\n",
1255                        mti->mti_svname, mti->mti_flags);
1256                 rc =  mgc_target_register(exp, mti);
1257                 RETURN(rc);
1258         }
1259         if (KEY_IS(KEY_SET_FS)) {
1260                 struct super_block *sb = (struct super_block *)val;
1261
1262                 if (vallen != sizeof(struct super_block))
1263                         RETURN(-EINVAL);
1264
1265                 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1266                 RETURN(rc);
1267         }
1268         if (KEY_IS(KEY_CLEAR_FS)) {
1269                 if (vallen != 0)
1270                         RETURN(-EINVAL);
1271                 rc = mgc_fs_cleanup(env, exp->exp_obd);
1272                 RETURN(rc);
1273         }
1274 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1275         if (KEY_IS(KEY_SET_INFO)) {
1276                 struct mgs_send_param *msp;
1277
1278                 msp = (struct mgs_send_param *)val;
1279                 rc =  mgc_set_mgs_param(exp, msp);
1280                 RETURN(rc);
1281         }
1282 #endif
1283         if (KEY_IS(KEY_MGSSEC)) {
1284                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1285                 struct sptlrpc_flavor  flvr;
1286
1287                 /*
1288                  * empty string means using current flavor, if which haven't
1289                  * been set yet, set it as null.
1290                  *
1291                  * if flavor has been set previously, check the asking flavor
1292                  * must match the existing one.
1293                  */
1294                 if (vallen == 0) {
1295                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1296                                 RETURN(0);
1297                         val = "null";
1298                         vallen = 4;
1299                 }
1300
1301                 rc = sptlrpc_parse_flavor(val, &flvr);
1302                 if (rc) {
1303                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1304                                (char *) val);
1305                         RETURN(rc);
1306                 }
1307
1308                 /*
1309                  * caller already hold a mutex
1310                  */
1311                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1312                         cli->cl_flvr_mgc = flvr;
1313                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1314                                   sizeof(flvr)) != 0) {
1315                         char    str[20];
1316
1317                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1318                                             str, sizeof(str));
1319                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1320                                        "currently %s is in use\n",
1321                                        (char *) val, str);
1322                         rc = -EPERM;
1323                 }
1324                 RETURN(rc);
1325         }
1326
1327         RETURN(rc);
1328 }
1329
1330 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1331                         __u32 keylen, void *key, __u32 *vallen, void *val)
1332 {
1333         int rc = -EINVAL;
1334
1335         if (KEY_IS(KEY_CONN_DATA)) {
1336                 struct obd_import *imp = class_exp2cliimp(exp);
1337                 struct obd_connect_data *data = val;
1338
1339                 if (*vallen == sizeof(*data)) {
1340                         *data = imp->imp_connect_data;
1341                         rc = 0;
1342                 }
1343         }
1344
1345         return rc;
1346 }
1347
1348 static int mgc_import_event(struct obd_device *obd,
1349                             struct obd_import *imp,
1350                             enum obd_import_event event)
1351 {
1352         int rc = 0;
1353
1354         LASSERT(imp->imp_obd == obd);
1355         CDEBUG(D_MGC, "import event %#x\n", event);
1356
1357         switch (event) {
1358         case IMP_EVENT_DISCON:
1359                 /* MGC imports should not wait for recovery */
1360                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1361                         ptlrpc_pinger_ir_down();
1362                 break;
1363         case IMP_EVENT_INACTIVE:
1364                 break;
1365         case IMP_EVENT_INVALIDATE: {
1366                 struct ldlm_namespace *ns = obd->obd_namespace;
1367                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1368                 break;
1369         }
1370         case IMP_EVENT_ACTIVE:
1371                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1372                 /* Clearing obd_no_recov allows us to continue pinging */
1373                 obd->obd_no_recov = 0;
1374                 mgc_notify_active(obd);
1375                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1376                         ptlrpc_pinger_ir_up();
1377                 break;
1378         case IMP_EVENT_OCD:
1379                 break;
1380         case IMP_EVENT_DEACTIVATE:
1381         case IMP_EVENT_ACTIVATE:
1382                 break;
1383         default:
1384                 CERROR("Unknown import event %#x\n", event);
1385                 LBUG();
1386         }
1387         RETURN(rc);
1388 }
1389
1390 enum {
1391         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1392         CONFIG_READ_NRPAGES      = 4
1393 };
1394
1395 static int mgc_apply_recover_logs(struct obd_device *mgc,
1396                                   struct config_llog_data *cld,
1397                                   __u64 max_version,
1398                                   void *data, int datalen, bool mne_swab)
1399 {
1400         struct config_llog_instance *cfg = &cld->cld_cfg;
1401         struct lustre_sb_info       *lsi = s2lsi(cfg->cfg_sb);
1402         struct mgs_nidtbl_entry *entry;
1403         struct lustre_cfg       *lcfg;
1404         struct lustre_cfg_bufs   bufs;
1405         u64   prev_version = 0;
1406         char *inst;
1407         char *buf;
1408         int   bufsz;
1409         int   pos;
1410         int   rc  = 0;
1411         int   off = 0;
1412         ENTRY;
1413
1414         LASSERT(cfg->cfg_instance != NULL);
1415         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1416
1417         OBD_ALLOC(inst, PAGE_SIZE);
1418         if (inst == NULL)
1419                 RETURN(-ENOMEM);
1420
1421         if (!IS_SERVER(lsi)) {
1422                 pos = snprintf(inst, PAGE_SIZE, "%p", cfg->cfg_instance);
1423                 if (pos >= PAGE_SIZE) {
1424                         OBD_FREE(inst, PAGE_SIZE);
1425                         return -E2BIG;
1426                 }
1427         } else {
1428                 LASSERT(IS_MDT(lsi));
1429                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1430                                         PAGE_SIZE);
1431                 if (rc) {
1432                         OBD_FREE(inst, PAGE_SIZE);
1433                         RETURN(-EINVAL);
1434                 }
1435                 pos = strlen(inst);
1436         }
1437
1438         ++pos;
1439         buf   = inst + pos;
1440         bufsz = PAGE_SIZE - pos;
1441
1442         while (datalen > 0) {
1443                 int   entry_len = sizeof(*entry);
1444                 int   is_ost, i;
1445                 struct obd_device *obd;
1446                 char *obdname;
1447                 char *cname;
1448                 char *params;
1449                 char *uuid;
1450
1451                 rc = -EINVAL;
1452                 if (datalen < sizeof(*entry))
1453                         break;
1454
1455                 entry = (typeof(entry))(data + off);
1456
1457                 /* sanity check */
1458                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1459                         break;
1460                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1461                         break;
1462                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1463                         break;
1464
1465                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1466                 if (datalen < entry_len) /* must have entry_len at least */
1467                         break;
1468
1469                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1470                 if (mne_swab)
1471                         lustre_swab_mgs_nidtbl_entry(entry);
1472                 if (entry->mne_length > PAGE_SIZE) {
1473                         CERROR("MNE too large (%u)\n", entry->mne_length);
1474                         break;
1475                 }
1476
1477                 if (entry->mne_length < entry_len)
1478                         break;
1479
1480                 off     += entry->mne_length;
1481                 datalen -= entry->mne_length;
1482                 if (datalen < 0)
1483                         break;
1484
1485                 if (entry->mne_version > max_version) {
1486                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1487                                entry->mne_version, max_version);
1488                         break;
1489                 }
1490
1491                 if (prev_version >= entry->mne_version) {
1492                         CERROR("index unsorted, prev %lld, now %lld\n",
1493                                prev_version, entry->mne_version);
1494                         break;
1495                 }
1496                 prev_version = entry->mne_version;
1497
1498                 /*
1499                  * Write a string with format "nid::instance" to
1500                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1501                  */
1502
1503                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1504                 memset(buf, 0, bufsz);
1505                 obdname = buf;
1506                 pos = 0;
1507
1508                 /* lustre-OST0001-osc-<instance #> */
1509                 strcpy(obdname, cld->cld_logname);
1510                 cname = strrchr(obdname, '-');
1511                 if (cname == NULL) {
1512                         CERROR("mgc %s: invalid logname %s\n",
1513                                mgc->obd_name, obdname);
1514                         break;
1515                 }
1516
1517                 pos = cname - obdname;
1518                 obdname[pos] = 0;
1519                 pos += sprintf(obdname + pos, "-%s%04x",
1520                                   is_ost ? "OST" : "MDT", entry->mne_index);
1521
1522                 cname = is_ost ? "osc" : "mdc",
1523                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1524                 lustre_cfg_bufs_reset(&bufs, obdname);
1525
1526                 /* find the obd by obdname */
1527                 obd = class_name2obd(obdname);
1528                 if (obd == NULL) {
1529                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1530                                mgc->obd_name, obdname);
1531                         rc = 0;
1532                         /* this is a safe race, when the ost is starting up...*/
1533                         continue;
1534                 }
1535
1536                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1537                 ++pos;
1538                 params = buf + pos;
1539                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1540                 uuid = buf + pos;
1541
1542                 down_read(&obd->u.cli.cl_sem);
1543                 if (obd->u.cli.cl_import == NULL) {
1544                         /* client does not connect to the OST yet */
1545                         up_read(&obd->u.cli.cl_sem);
1546                         rc = 0;
1547                         continue;
1548                 }
1549
1550                 /* iterate all nids to find one */
1551                 /* find uuid by nid */
1552                 rc = -ENOENT;
1553                 for (i = 0; i < entry->mne_nid_count; i++) {
1554                         rc = client_import_find_conn(obd->u.cli.cl_import,
1555                                                      entry->u.nids[i],
1556                                                      (struct obd_uuid *)uuid);
1557                         if (rc == 0)
1558                                 break;
1559                 }
1560
1561                 up_read(&obd->u.cli.cl_sem);
1562                 if (rc < 0) {
1563                         CERROR("mgc: cannot find uuid by nid %s\n",
1564                                libcfs_nid2str(entry->u.nids[0]));
1565                         break;
1566                 }
1567
1568                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1569                        uuid, libcfs_nid2str(entry->u.nids[0]));
1570
1571                 pos += strlen(uuid);
1572                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1573                 LASSERT(pos < bufsz);
1574
1575                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1576
1577                 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1578                                                bufs.lcfg_buflen));
1579                 if (!lcfg) {
1580                         rc = -ENOMEM;
1581                         break;
1582                 }
1583                 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1584
1585                 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1586                        prev_version, max_version, obdname, params);
1587
1588                 rc = class_process_config(lcfg);
1589                 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1590                                               lcfg->lcfg_buflens));
1591                 if (rc)
1592                         CDEBUG(D_INFO, "process config for %s error %d\n",
1593                                obdname, rc);
1594
1595                 /* continue, even one with error */
1596         }
1597
1598         OBD_FREE(inst, PAGE_SIZE);
1599         RETURN(rc);
1600 }
1601
1602 /**
1603  * This function is called if this client was notified for target restarting
1604  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1605  * nodemap logs.
1606  */
1607 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1608                                            struct config_llog_data *cld)
1609 {
1610         struct ptlrpc_connection *mgc_conn;
1611         struct ptlrpc_request *req = NULL;
1612         struct config_llog_instance *cfg = &cld->cld_cfg;
1613         struct mgs_config_body *body;
1614         struct mgs_config_res *res;
1615         struct nodemap_config *new_config = NULL;
1616         struct lu_nodemap *recent_nodemap = NULL;
1617         struct ptlrpc_bulk_desc *desc;
1618         struct page **pages = NULL;
1619         __u64 config_read_offset = 0;
1620         __u8 nodemap_cur_pass = 0;
1621         int nrpages = 0;
1622         bool eof = true;
1623         bool mne_swab = false;
1624         int i;
1625         int ealen;
1626         int rc;
1627         ENTRY;
1628
1629         mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1630
1631         /* don't need to get local config */
1632         if (cld_is_nodemap(cld) &&
1633             (LNET_NETTYP(LNET_NIDNET(mgc_conn->c_peer.nid)) == LOLND))
1634                 GOTO(out, rc = 0);
1635
1636         /* allocate buffer for bulk transfer.
1637          * if this is the first time for this mgs to read logs,
1638          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1639          * once; otherwise, it only reads increment of logs, this should be
1640          * small and CONFIG_READ_NRPAGES will be used.
1641          */
1642         nrpages = CONFIG_READ_NRPAGES;
1643         if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1644                 nrpages = CONFIG_READ_NRPAGES_INIT;
1645
1646         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1647         if (pages == NULL)
1648                 GOTO(out, rc = -ENOMEM);
1649
1650         for (i = 0; i < nrpages; i++) {
1651                 pages[i] = alloc_page(GFP_KERNEL);
1652                 if (pages[i] == NULL)
1653                         GOTO(out, rc = -ENOMEM);
1654         }
1655
1656 again:
1657 #ifdef HAVE_SERVER_SUPPORT
1658         if (cld_is_nodemap(cld) && config_read_offset == 0) {
1659                 new_config = nodemap_config_alloc();
1660                 if (IS_ERR(new_config)) {
1661                         rc = PTR_ERR(new_config);
1662                         new_config = NULL;
1663                         GOTO(out, rc);
1664                 }
1665         }
1666 #endif
1667         LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1668         LASSERT(mutex_is_locked(&cld->cld_lock));
1669         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1670                                    &RQF_MGS_CONFIG_READ);
1671         if (req == NULL)
1672                 GOTO(out, rc = -ENOMEM);
1673
1674         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1675         if (rc)
1676                 GOTO(out, rc);
1677
1678         /* pack request */
1679         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1680         LASSERT(body != NULL);
1681         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1682         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1683             >= sizeof(body->mcb_name))
1684                 GOTO(out, rc = -E2BIG);
1685         if (cld_is_nodemap(cld))
1686                 body->mcb_offset = config_read_offset;
1687         else
1688                 body->mcb_offset = cfg->cfg_last_idx + 1;
1689         body->mcb_type   = cld->cld_type;
1690         body->mcb_bits   = PAGE_SHIFT;
1691         body->mcb_units  = nrpages;
1692         body->mcb_nm_cur_pass = nodemap_cur_pass;
1693
1694         /* allocate bulk transfer descriptor */
1695         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1696                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1697                                     MGS_BULK_PORTAL,
1698                                     &ptlrpc_bulk_kiov_pin_ops);
1699         if (desc == NULL)
1700                 GOTO(out, rc = -ENOMEM);
1701
1702         for (i = 0; i < nrpages; i++)
1703                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1704                                                  PAGE_SIZE);
1705
1706         ptlrpc_request_set_replen(req);
1707         rc = ptlrpc_queue_wait(req);
1708         if (rc)
1709                 GOTO(out, rc);
1710
1711         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1712         if (!res)
1713                 GOTO(out, rc = -EPROTO);
1714
1715         if (cld_is_nodemap(cld)) {
1716                 config_read_offset = res->mcr_offset;
1717                 eof = config_read_offset == II_END_OFF;
1718                 nodemap_cur_pass = res->mcr_nm_cur_pass;
1719         } else {
1720                 if (res->mcr_size < res->mcr_offset)
1721                         GOTO(out, rc = -EINVAL);
1722
1723                 /* always update the index even though it might have errors with
1724                  * handling the recover logs
1725                  */
1726                 cfg->cfg_last_idx = res->mcr_offset;
1727                 eof = res->mcr_offset == res->mcr_size;
1728
1729                 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1730                        res->mcr_offset, eof == false);
1731         }
1732
1733         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1734         if (ealen < 0)
1735                 GOTO(out, rc = ealen);
1736
1737         if (ealen > nrpages << PAGE_SHIFT)
1738                 GOTO(out, rc = -EINVAL);
1739
1740         if (ealen == 0) { /* no logs transferred */
1741 #ifdef HAVE_SERVER_SUPPORT
1742                 /* config changed since first read RPC */
1743                 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1744                         recent_nodemap = NULL;
1745                         nodemap_config_dealloc(new_config);
1746                         new_config = NULL;
1747
1748                         CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1749
1750                         /* setting eof to false, we request config again */
1751                         eof = false;
1752                         GOTO(out, rc = 0);
1753                 }
1754 #endif
1755                 if (!eof)
1756                         rc = -EINVAL;
1757                 GOTO(out, rc);
1758         }
1759
1760         mne_swab = !!ptlrpc_rep_need_swab(req);
1761 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1762         /* This import flag means the server did an extra swab of IR MNE
1763          * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1764         if (unlikely(req->rq_import->imp_need_mne_swab))
1765                 mne_swab = !mne_swab;
1766 #endif
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(pages, sizeof(*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. doing it inside
1978          * of llog_process_lock so no device is attaching/detaching
1979          * in parallel.
1980          * the logname must be <fsname>-sptlrpc
1981          */
1982         if (rc == 0 && cld_is_sptlrpc(cld))
1983                 class_notify_sptlrpc_conf(cld->cld_logname,
1984                                           strlen(cld->cld_logname) -
1985                                           strlen("-sptlrpc"));
1986         EXIT;
1987
1988 out_pop:
1989         __llog_ctxt_put(env, ctxt);
1990         if (lctxt)
1991                 __llog_ctxt_put(env, lctxt);
1992
1993         lu_env_fini(env);
1994 out_free:
1995         OBD_FREE_PTR(env);
1996         return rc;
1997 }
1998
1999 static bool mgc_import_in_recovery(struct obd_import *imp)
2000 {
2001         bool in_recovery = true;
2002
2003         spin_lock(&imp->imp_lock);
2004         if (imp->imp_state == LUSTRE_IMP_FULL ||
2005             imp->imp_state == LUSTRE_IMP_CLOSED)
2006                 in_recovery = false;
2007         spin_unlock(&imp->imp_lock);
2008
2009         return in_recovery;
2010 }
2011
2012 /**
2013  * Get a configuration log from the MGS and process it.
2014  *
2015  * This function is called for both clients and servers to process the
2016  * configuration log from the MGS.  The MGC enqueues a DLM lock on the
2017  * log from the MGS, and if the lock gets revoked the MGC will be notified
2018  * by the lock cancellation callback that the config log has changed,
2019  * and will enqueue another MGS lock on it, and then continue processing
2020  * the new additions to the end of the log.
2021  *
2022  * Since the MGC import is not replayable, if the import is being evicted
2023  * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
2024  * the log until recovery is finished or the import is closed.
2025  *
2026  * Make a local copy of the log before parsing it if appropriate (non-MGS
2027  * server) so that the server can start even when the MGS is down.
2028  *
2029  * There shouldn't be multiple processes running process_log at once --
2030  * sounds like badness.  It actually might be fine, as long as they're not
2031  * trying to update from the same log simultaneously, in which case we
2032  * should use a per-log semaphore instead of cld_lock.
2033  *
2034  * \param[in] mgc       MGC device by which to fetch the configuration log
2035  * \param[in] cld       log processing state (stored in lock callback data)
2036  *
2037  * \retval              0 on success
2038  * \retval              negative errno on failure
2039  */
2040 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
2041 {
2042         struct lustre_handle lockh = { 0 };
2043         __u64 flags = LDLM_FL_NO_LRU;
2044         int rc = 0, rcl;
2045         bool retry = false;
2046         ENTRY;
2047
2048         LASSERT(cld != NULL);
2049
2050         /* I don't want multiple processes running process_log at once --
2051            sounds like badness.  It actually might be fine, as long as
2052            we're not trying to update from the same log
2053            simultaneously (in which case we should use a per-log sem.) */
2054 restart:
2055         mutex_lock(&cld->cld_lock);
2056         if (cld->cld_stopping) {
2057                 mutex_unlock(&cld->cld_lock);
2058                 RETURN(0);
2059         }
2060
2061         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2062
2063         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
2064                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2065
2066         /* Get the cfg lock on the llog */
2067         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2068                           LCK_CR, &flags,
2069                           cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
2070                           cld, 0, NULL, &lockh);
2071         if (rcl == 0) {
2072                 /* Get the cld, it will be released in mgc_blocking_ast. */
2073                 config_log_get(cld);
2074                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2075                 LASSERT(rc == 0);
2076         } else {
2077                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2078
2079                 if (rcl == -ESHUTDOWN &&
2080                     atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2081                         struct obd_import *imp;
2082                         struct l_wait_info lwi;
2083                         int secs = 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                         lwi = LWI_TIMEOUT(secs, NULL, NULL);
2097                         l_wait_event(imp->imp_recovery_waitq,
2098                                      !mgc_import_in_recovery(imp), &lwi);
2099
2100                         if (imp->imp_state == LUSTRE_IMP_FULL) {
2101                                 retry = true;
2102                                 goto restart;
2103                         } else {
2104                                 mutex_lock(&cld->cld_lock);
2105                                 spin_lock(&config_list_lock);
2106                                 cld->cld_lostlock = 1;
2107                                 spin_unlock(&config_list_lock);
2108                         }
2109                 } else {
2110                         /* mark cld_lostlock so that it will requeue
2111                          * after MGC becomes available. */
2112                         spin_lock(&config_list_lock);
2113                         cld->cld_lostlock = 1;
2114                         spin_unlock(&config_list_lock);
2115                 }
2116         }
2117
2118         if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2119                 if (!rcl)
2120                         rc = mgc_process_recover_nodemap_log(mgc, cld);
2121                 else if (cld_is_nodemap(cld))
2122                         rc = rcl;
2123
2124                 if (cld_is_recover(cld) && rc) {
2125                         if (!rcl) {
2126                                 CERROR("%s: recover log %s failed, not fatal: rc = %d\n",
2127                                        mgc->obd_name, cld->cld_logname, rc);
2128                                 spin_lock(&config_list_lock);
2129                                 cld->cld_lostlock = 1;
2130                                 spin_unlock(&config_list_lock);
2131                         }
2132                         rc = 0; /* this is not a fatal error for recover log */
2133                 }
2134         } else if (!cld_is_barrier(cld)) {
2135                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2136         }
2137
2138         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2139                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2140
2141         mutex_unlock(&cld->cld_lock);
2142
2143         /* Now drop the lock so MGS can revoke it */
2144         if (!rcl) {
2145                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2146                 if (rcl)
2147                         CERROR("Can't drop cfg lock: %d\n", rcl);
2148         }
2149
2150         RETURN(rc);
2151 }
2152
2153
2154 /** Called from lustre_process_log.
2155  * LCFG_LOG_START gets the config log from the MGS, processes it to start
2156  * any services, and adds it to the list logs to watch (follow).
2157  */
2158 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2159 {
2160         struct lustre_cfg *lcfg = buf;
2161         struct config_llog_instance *cfg = NULL;
2162         char *logname;
2163         int rc = 0;
2164         ENTRY;
2165
2166         switch(lcfg->lcfg_command) {
2167         case LCFG_LOV_ADD_OBD: {
2168                 /* Overloading this cfg command: register a new target */
2169                 struct mgs_target_info *mti;
2170
2171                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2172                     sizeof(struct mgs_target_info))
2173                         GOTO(out, rc = -EINVAL);
2174
2175                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2176                 CDEBUG(D_MGC, "add_target %s %#x\n",
2177                        mti->mti_svname, mti->mti_flags);
2178                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2179                 break;
2180         }
2181         case LCFG_LOV_DEL_OBD:
2182                 /* Unregister has no meaning at the moment. */
2183                 CERROR("lov_del_obd unimplemented\n");
2184                 rc = -ENOSYS;
2185                 break;
2186         case LCFG_SPTLRPC_CONF: {
2187                 rc = sptlrpc_process_config(lcfg);
2188                 break;
2189         }
2190         case LCFG_LOG_START: {
2191                 struct config_llog_data *cld;
2192                 struct super_block *sb;
2193
2194                 logname = lustre_cfg_string(lcfg, 1);
2195                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2196                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2197
2198                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2199                        cfg->cfg_last_idx);
2200
2201                 /* We're only called through here on the initial mount */
2202                 cld = config_log_add(obd, logname, cfg, sb);
2203                 if (IS_ERR(cld)) {
2204                         rc = PTR_ERR(cld);
2205                         break;
2206                 }
2207
2208                 /* COMPAT_146 */
2209                 /* FIXME only set this for old logs!  Right now this forces
2210                    us to always skip the "inside markers" check */
2211                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
2212
2213                 rc = mgc_process_log(obd, cld);
2214                 if (rc == 0 && cld->cld_recover != NULL) {
2215                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2216                                          imp_connect_data, IMP_RECOV)) {
2217                                 rc = mgc_process_log(obd, cld->cld_recover);
2218                         } else {
2219                                 struct config_llog_data *cir;
2220
2221                                 mutex_lock(&cld->cld_lock);
2222                                 cir = cld->cld_recover;
2223                                 cld->cld_recover = NULL;
2224                                 mutex_unlock(&cld->cld_lock);
2225                                 config_log_put(cir);
2226                         }
2227
2228                         if (rc)
2229                                 CERROR("Cannot process recover llog %d\n", rc);
2230                 }
2231
2232                 if (rc == 0 && cld->cld_params != NULL) {
2233                         rc = mgc_process_log(obd, cld->cld_params);
2234                         if (rc == -ENOENT) {
2235                                 CDEBUG(D_MGC, "There is no params "
2236                                               "config file yet\n");
2237                                 rc = 0;
2238                         }
2239                         /* params log is optional */
2240                         if (rc)
2241                                 CERROR("%s: can't process params llog: rc = %d\n",
2242                                        obd->obd_name, rc);
2243                 }
2244
2245                 break;
2246         }
2247         case LCFG_LOG_END: {
2248                 logname = lustre_cfg_string(lcfg, 1);
2249
2250                 if (lcfg->lcfg_bufcount >= 2)
2251                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
2252                                 lcfg, 2);
2253                 rc = config_log_end(logname, cfg);
2254                 break;
2255         }
2256         default: {
2257                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2258                 GOTO(out, rc = -EINVAL);
2259
2260         }
2261         }
2262 out:
2263         RETURN(rc);
2264 }
2265
2266 static struct obd_ops mgc_obd_ops = {
2267         .o_owner        = THIS_MODULE,
2268         .o_setup        = mgc_setup,
2269         .o_precleanup   = mgc_precleanup,
2270         .o_cleanup      = mgc_cleanup,
2271         .o_add_conn     = client_import_add_conn,
2272         .o_del_conn     = client_import_del_conn,
2273         .o_connect      = client_connect_import,
2274         .o_disconnect   = client_disconnect_export,
2275         .o_set_info_async = mgc_set_info_async,
2276         .o_get_info       = mgc_get_info,
2277         .o_import_event = mgc_import_event,
2278         .o_process_config = mgc_process_config,
2279 };
2280
2281 static int __init mgc_init(void)
2282 {
2283         return class_register_type(&mgc_obd_ops, NULL, true, NULL,
2284                                    LUSTRE_MGC_NAME, NULL);
2285 }
2286
2287 static void __exit mgc_exit(void)
2288 {
2289         class_unregister_type(LUSTRE_MGC_NAME);
2290 }
2291
2292 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2293 MODULE_DESCRIPTION("Lustre Management Client");
2294 MODULE_VERSION(LUSTRE_VERSION_STRING);
2295 MODULE_LICENSE("GPL");
2296
2297 module_init(mgc_init);
2298 module_exit(mgc_exit);