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         cld = config_log_find(logname, &lcfg);
 300         if (unlikely(cld != NULL))
 301                 return cld;
 302
 303         return do_config_log_add(obd, logname, type, &lcfg, sb);
 304 }
 305
 306 /** Add this log to the list of active logs watched by an MGC.
 307  * Active means we're watching for updates.
 308  * We have one active log per "mount" - client instance or servername.
 309  * Each instance may be at a different point in the log.
 310  */
 311 static struct config_llog_data *
 312 config_log_add(struct obd_device *obd, char *logname,
 313                struct config_llog_instance *cfg, struct super_block *sb)
 314 {
 315         struct lustre_sb_info *lsi = s2lsi(sb);
 316         struct config_llog_data *cld = NULL;
 317         struct config_llog_data *sptlrpc_cld = NULL;
 318         struct config_llog_data *params_cld = NULL;
 319         struct config_llog_data *nodemap_cld = NULL;
 320         struct config_llog_data *barrier_cld = NULL;
 321         char seclogname[32];
 322         char *ptr;
 323         int rc;
 324         bool locked = false;
 325         ENTRY;
 326
 327         CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
 328
 329         /*
 330          * for each regular log, the depended sptlrpc log name is
 331          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 332          */
 333         ptr = strrchr(logname, '-');
 334         if (ptr == NULL || ptr - logname > 8) {
 335                 CERROR("logname %s is too long\n", logname);
 336                 RETURN(ERR_PTR(-EINVAL));
 337         }
 338
 339         memcpy(seclogname, logname, ptr - logname);
 340         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 341
 342         if (cfg->cfg_sub_clds & CONFIG_SUB_SPTLRPC) {
 343                 sptlrpc_cld = config_log_find_or_add(obd, seclogname, NULL,
 344                                                      CONFIG_T_SPTLRPC, cfg);
 345                 if (IS_ERR(sptlrpc_cld)) {
 346                         CERROR("%s: can't create sptlrpc log %s: rc = %ld\n",
 347                                obd->obd_name, seclogname, PTR_ERR(sptlrpc_cld));
 348                         RETURN(sptlrpc_cld);
 349                 }
 350         }
 351
 352         if (!IS_MGS(lsi) && cfg->cfg_sub_clds & CONFIG_SUB_NODEMAP) {
 353                 nodemap_cld = config_log_find_or_add(obd, LUSTRE_NODEMAP_NAME,
 354                                                      NULL, CONFIG_T_NODEMAP,
 355                                                      cfg);
 356                 if (IS_ERR(nodemap_cld)) {
 357                         rc = PTR_ERR(nodemap_cld);
 358                         CERROR("%s: cannot create nodemap log: rc = %d\n",
 359                                obd->obd_name, rc);
 360                         GOTO(out_sptlrpc, rc);
 361                 }
 362         }
 363
 364         if (cfg->cfg_sub_clds & CONFIG_SUB_PARAMS) {
 365                 params_cld = config_log_find_or_add(obd, PARAMS_FILENAME, sb,
 366                                                     CONFIG_T_PARAMS, cfg);
 367                 if (IS_ERR(params_cld)) {
 368                         rc = PTR_ERR(params_cld);
 369                         CERROR("%s: can't create params log: rc = %d\n",
 370                                obd->obd_name, rc);
 371                         GOTO(out_nodemap, rc);
 372                 }
 373         }
 374
 375         if (IS_MDT(s2lsi(sb)) && cfg->cfg_sub_clds & CONFIG_SUB_BARRIER) {
 376                 snprintf(seclogname + (ptr - logname), sizeof(seclogname) - 1,
 377                          "-%s", BARRIER_FILENAME);
 378                 barrier_cld = config_log_find_or_add(obd, seclogname, sb,
 379                                                      CONFIG_T_BARRIER, cfg);
 380                 if (IS_ERR(barrier_cld)) {
 381                         rc = PTR_ERR(barrier_cld);
 382                         CERROR("%s: can't create barrier log: rc = %d\n",
 383                                obd->obd_name, rc);
 384                         GOTO(out_params, rc);
 385                 }
 386         }
 387
 388         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 389         if (IS_ERR(cld)) {
 390                 rc = PTR_ERR(cld);
 391                 CERROR("%s: can't create log: rc = %d\n",
 392                        obd->obd_name, rc);
 393                 GOTO(out_barrier, rc = PTR_ERR(cld));
 394         }
 395
 396         LASSERT(lsi->lsi_lmd);
 397         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR) &&
 398             cfg->cfg_sub_clds & CONFIG_SUB_RECOVER) {
 399                 struct config_llog_data *recover_cld;
 400
 401                 ptr = strrchr(seclogname, '-');
 402                 if (ptr != NULL) {
 403                         *ptr = 0;
 404                 } else {
 405                         CERROR("%s: sptlrpc log name not correct, %s: "
 406                                "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
 407                         GOTO(out_cld, rc = -EINVAL);
 408                 }
 409
 410                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 411                 if (IS_ERR(recover_cld)) {
 412                         rc = PTR_ERR(recover_cld);
 413                         CERROR("%s: can't create recover log: rc = %d\n",
 414                                obd->obd_name, rc);
 415                         GOTO(out_cld, rc);
 416                 }
 417
 418                 mutex_lock(&cld->cld_lock);
 419                 locked = true;
 420                 cld->cld_recover = recover_cld;
 421         }
 422
 423         if (!locked)
 424                 mutex_lock(&cld->cld_lock);
 425         cld->cld_params = params_cld;
 426         cld->cld_barrier = barrier_cld;
 427         cld->cld_nodemap = nodemap_cld;
 428         cld->cld_sptlrpc = sptlrpc_cld;
 429         mutex_unlock(&cld->cld_lock);
 430
 431         RETURN(cld);
 432
 433 out_cld:
 434         config_log_put(cld);
 435 out_barrier:
 436         config_log_put(barrier_cld);
 437 out_params:
 438         config_log_put(params_cld);
 439 out_nodemap:
 440         config_log_put(nodemap_cld);
 441 out_sptlrpc:
 442         config_log_put(sptlrpc_cld);
 443
 444         return ERR_PTR(rc);
 445 }
 446
 447 DEFINE_MUTEX(llog_process_lock);
 448
 449 static inline void config_mark_cld_stop(struct config_llog_data *cld)
 450 {
 451         if (cld) {
 452                 mutex_lock(&cld->cld_lock);
 453                 spin_lock(&config_list_lock);
 454                 cld->cld_stopping = 1;
 455                 spin_unlock(&config_list_lock);
 456                 mutex_unlock(&cld->cld_lock);
 457         }
 458 }
 459
 460 /** Stop watching for updates on this log.
 461  */
 462 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 463 {
 464         struct config_llog_data *cld;
 465         struct config_llog_data *cld_sptlrpc = NULL;
 466         struct config_llog_data *cld_params = NULL;
 467         struct config_llog_data *cld_recover = NULL;
 468         struct config_llog_data *cld_nodemap = NULL;
 469         struct config_llog_data *cld_barrier = NULL;
 470         int rc = 0;
 471
 472         ENTRY;
 473
 474         cld = config_log_find(logname, cfg);
 475         if (cld == NULL)
 476                 RETURN(-ENOENT);
 477
 478         mutex_lock(&cld->cld_lock);
 479         /*
 480          * if cld_stopping is set, it means we didn't start the log thus
 481          * not owning the start ref. this can happen after previous umount:
 482          * the cld still hanging there waiting for lock cancel, and we
 483          * remount again but failed in the middle and call log_end without
 484          * calling start_log.
 485          */
 486         if (unlikely(cld->cld_stopping)) {
 487                 mutex_unlock(&cld->cld_lock);
 488                 /* drop the ref from the find */
 489                 config_log_put(cld);
 490                 RETURN(rc);
 491         }
 492
 493         spin_lock(&config_list_lock);
 494         cld->cld_stopping = 1;
 495         spin_unlock(&config_list_lock);
 496
 497         cld_recover = cld->cld_recover;
 498         cld->cld_recover = NULL;
 499         cld_params = cld->cld_params;
 500         cld->cld_params = NULL;
 501         cld_nodemap = cld->cld_nodemap;
 502         cld->cld_nodemap = NULL;
 503         cld_barrier = cld->cld_barrier;
 504         cld->cld_barrier = NULL;
 505         cld_sptlrpc = cld->cld_sptlrpc;
 506         cld->cld_sptlrpc = NULL;
 507         mutex_unlock(&cld->cld_lock);
 508
 509         config_mark_cld_stop(cld_recover);
 510         config_log_put(cld_recover);
 511
 512         config_mark_cld_stop(cld_params);
 513         config_log_put(cld_params);
 514
 515         /* don't set cld_stopping on nm lock as other targets may be active */
 516         config_log_put(cld_nodemap);
 517
 518         if (cld_barrier) {
 519                 mutex_lock(&cld_barrier->cld_lock);
 520                 cld_barrier->cld_stopping = 1;
 521                 mutex_unlock(&cld_barrier->cld_lock);
 522                 config_log_put(cld_barrier);
 523         }
 524
 525         config_log_put(cld_sptlrpc);
 526
 527         /* drop the ref from the find */
 528         config_log_put(cld);
 529         /* drop the start ref */
 530         config_log_put(cld);
 531
 532         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 533                rc);
 534         RETURN(rc);
 535 }
 536
 537 #ifdef CONFIG_PROC_FS
 538 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 539 {
 540         struct obd_device       *obd = data;
 541         struct obd_import       *imp;
 542         struct obd_connect_data *ocd;
 543         struct config_llog_data *cld;
 544         ENTRY;
 545
 546         LASSERT(obd != NULL);
 547         LPROCFS_CLIMP_CHECK(obd);
 548         imp = obd->u.cli.cl_import;
 549         ocd = &imp->imp_connect_data;
 550
 551         seq_printf(m, "imperative_recovery: %s\n",
 552                    OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 553         seq_printf(m, "client_state:\n");
 554
 555         spin_lock(&config_list_lock);
 556         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 557                 if (cld->cld_recover == NULL)
 558                         continue;
 559                 seq_printf(m,  "    - { client: %s, nidtbl_version: %u }\n",
 560                            cld->cld_logname,
 561                            cld->cld_recover->cld_cfg.cfg_last_idx);
 562         }
 563         spin_unlock(&config_list_lock);
 564
 565         LPROCFS_CLIMP_EXIT(obd);
 566         RETURN(0);
 567 }
 568 #endif
 569
 570 /* reenqueue any lost locks */
 571 #define RQ_RUNNING      0x1
 572 #define RQ_NOW          0x2
 573 #define RQ_LATER        0x4
 574 #define RQ_STOP         0x8
 575 #define RQ_PRECLEANUP   0x10
 576 static int                    rq_state = 0;
 577 static wait_queue_head_t      rq_waitq;
 578 static DECLARE_COMPLETION(rq_exit);
 579 static DECLARE_COMPLETION(rq_start);
 580
 581 static void do_requeue(struct config_llog_data *cld)
 582 {
 583         int rc = 0;
 584         ENTRY;
 585
 586         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 587
 588         /*
 589          * Do not run mgc_process_log on a disconnected export or an
 590          * export which is being disconnected. Take the client
 591          * semaphore to make the check non-racy.
 592          */
 593         down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
 594                          OBD_CLI_SEM_MGC);
 595         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 596                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 597                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 598                 if (rc && rc != -ENOENT)
 599                         CERROR("failed processing log: %d\n", rc);
 600         } else {
 601                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 602                        cld->cld_logname);
 603         }
 604         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 605
 606         EXIT;
 607 }
 608
 609 /* this timeout represents how many seconds MGC should wait before
 610  * requeue config and recover lock to the MGS. We need to randomize this
 611  * in order to not flood the MGS.
 612  */
 613 #define MGC_TIMEOUT_MIN_SECONDS   5
 614 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
 615
 616 static int mgc_requeue_thread(void *data)
 617 {
 618         int rc = 0;
 619         bool first = true;
 620         ENTRY;
 621
 622         CDEBUG(D_MGC, "Starting requeue thread\n");
 623
 624         /* Keep trying failed locks periodically */
 625         spin_lock(&config_list_lock);
 626         rq_state |= RQ_RUNNING;
 627         while (!(rq_state & RQ_STOP)) {
 628                 struct l_wait_info lwi;
 629                 struct config_llog_data *cld, *cld_prev;
 630                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
 631                 int to;
 632
 633                 /* Any new or requeued lostlocks will change the state */
 634                 rq_state &= ~(RQ_NOW | RQ_LATER);
 635                 spin_unlock(&config_list_lock);
 636
 637                 if (first) {
 638                         first = false;
 639                         complete(&rq_start);
 640                 }
 641
 642                 /* Always wait a few seconds to allow the server who
 643                    caused the lock revocation to finish its setup, plus some
 644                    random so everyone doesn't try to reconnect at once. */
 645                 to = msecs_to_jiffies(MGC_TIMEOUT_MIN_SECONDS * MSEC_PER_SEC);
 646                 /* rand is centi-seconds */
 647                 to += msecs_to_jiffies(rand * MSEC_PER_SEC / 100);
 648                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 649                 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
 650                              &lwi);
 651
 652                 /*
 653                  * iterate & processing through the list. for each cld, process
 654                  * its depending sptlrpc cld firstly (if any) and then itself.
 655                  *
 656                  * it's guaranteed any item in the list must have
 657                  * reference > 0; and if cld_lostlock is set, at
 658                  * least one reference is taken by the previous enqueue.
 659                  */
 660                 cld_prev = NULL;
 661
 662                 spin_lock(&config_list_lock);
 663                 rq_state &= ~RQ_PRECLEANUP;
 664                 list_for_each_entry(cld, &config_llog_list,
 665                                     cld_list_chain) {
 666                         if (!cld->cld_lostlock || cld->cld_stopping)
 667                                 continue;
 668
 669                         /* hold reference to avoid being freed during
 670                          * subsequent processing. */
 671                         config_log_get(cld);
 672                         cld->cld_lostlock = 0;
 673                         spin_unlock(&config_list_lock);
 674
 675                         config_log_put(cld_prev);
 676                         cld_prev = cld;
 677
 678                         if (likely(!(rq_state & RQ_STOP))) {
 679                                 do_requeue(cld);
 680                                 spin_lock(&config_list_lock);
 681                         } else {
 682                                 spin_lock(&config_list_lock);
 683                                 break;
 684                         }
 685                 }
 686                 spin_unlock(&config_list_lock);
 687                 config_log_put(cld_prev);
 688
 689                 /* Wait a bit to see if anyone else needs a requeue */
 690                 lwi = (struct l_wait_info) { 0 };
 691                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 692                              &lwi);
 693                 spin_lock(&config_list_lock);
 694         }
 695
 696         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 697         rq_state &= ~RQ_RUNNING;
 698         spin_unlock(&config_list_lock);
 699
 700         complete(&rq_exit);
 701
 702         CDEBUG(D_MGC, "Ending requeue thread\n");
 703         RETURN(rc);
 704 }
 705
 706 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 707    We are responsible for dropping the config log reference from here on out. */
 708 static void mgc_requeue_add(struct config_llog_data *cld)
 709 {
 710         bool wakeup = false;
 711         ENTRY;
 712
 713         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 714                 cld->cld_logname, atomic_read(&cld->cld_refcount),
 715                 cld->cld_stopping, rq_state);
 716         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 717
 718         mutex_lock(&cld->cld_lock);
 719         spin_lock(&config_list_lock);
 720         if (!(rq_state & RQ_STOP) && !cld->cld_stopping && !cld->cld_lostlock) {
 721                 cld->cld_lostlock = 1;
 722                 rq_state |= RQ_NOW;
 723                 wakeup = true;
 724         }
 725         spin_unlock(&config_list_lock);
 726         mutex_unlock(&cld->cld_lock);
 727         if (wakeup)
 728                 wake_up(&rq_waitq);
 729
 730         EXIT;
 731 }
 732
 733 /********************** class fns **********************/
 734 static int mgc_local_llog_init(const struct lu_env *env,
 735                                struct obd_device *obd,
 736                                struct obd_device *disk)
 737 {
 738         struct llog_ctxt        *ctxt;
 739         int                      rc;
 740
 741         ENTRY;
 742
 743         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
 744                         &llog_osd_ops);
 745         if (rc)
 746                 RETURN(rc);
 747
 748         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 749         LASSERT(ctxt);
 750         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
 751         llog_ctxt_put(ctxt);
 752
 753         RETURN(0);
 754 }
 755
 756 static int mgc_local_llog_fini(const struct lu_env *env,
 757                                struct obd_device *obd)
 758 {
 759         struct llog_ctxt *ctxt;
 760
 761         ENTRY;
 762
 763         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 764         llog_cleanup(env, ctxt);
 765
 766         RETURN(0);
 767 }
 768
 769 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
 770                         struct super_block *sb)
 771 {
 772         struct lustre_sb_info   *lsi = s2lsi(sb);
 773         struct client_obd       *cli = &obd->u.cli;
 774         struct lu_fid            rfid, fid;
 775         struct dt_object        *root, *dto;
 776         int                      rc = 0;
 777
 778         ENTRY;
 779
 780         LASSERT(lsi);
 781         LASSERT(lsi->lsi_dt_dev);
 782
 783         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
 784         mutex_lock(&cli->cl_mgc_mutex);
 785
 786         /* Setup the configs dir */
 787         fid.f_seq = FID_SEQ_LOCAL_NAME;
 788         fid.f_oid = 1;
 789         fid.f_ver = 0;
 790         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
 791                                     &cli->cl_mgc_los);
 792         if (rc)
 793                 RETURN(rc);
 794
 795         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
 796         if (rc)
 797                 GOTO(out_los, rc);
 798
 799         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
 800                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
 801         if (unlikely(IS_ERR(root)))
 802                 GOTO(out_los, rc = PTR_ERR(root));
 803
 804         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
 805                                         MOUNT_CONFIGS_DIR,
 806                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
 807         dt_object_put_nocache(env, root);
 808         if (IS_ERR(dto))
 809                 GOTO(out_los, rc = PTR_ERR(dto));
 810
 811         cli->cl_mgc_configs_dir = dto;
 812
 813         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
 814         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
 815         if (rc)
 816                 GOTO(out_llog, rc);
 817
 818         /* We take an obd ref to insure that we can't get to mgc_cleanup
 819          * without calling mgc_fs_cleanup first. */
 820         class_incref(obd, "mgc_fs", obd);
 821
 822         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 823         EXIT;
 824 out_llog:
 825         if (rc) {
 826                 dt_object_put(env, cli->cl_mgc_configs_dir);
 827                 cli->cl_mgc_configs_dir = NULL;
 828         }
 829 out_los:
 830         if (rc < 0) {
 831                 local_oid_storage_fini(env, cli->cl_mgc_los);
 832                 cli->cl_mgc_los = NULL;
 833                 mutex_unlock(&cli->cl_mgc_mutex);
 834         }
 835         return rc;
 836 }
 837
 838 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
 839 {
 840         struct client_obd       *cli = &obd->u.cli;
 841         ENTRY;
 842
 843         LASSERT(cli->cl_mgc_los != NULL);
 844
 845         mgc_local_llog_fini(env, obd);
 846
 847         dt_object_put_nocache(env, cli->cl_mgc_configs_dir);
 848         cli->cl_mgc_configs_dir = NULL;
 849
 850         local_oid_storage_fini(env, cli->cl_mgc_los);
 851         cli->cl_mgc_los = NULL;
 852
 853         class_decref(obd, "mgc_fs", obd);
 854         mutex_unlock(&cli->cl_mgc_mutex);
 855
 856         RETURN(0);
 857 }
 858
 859 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 860 {
 861         struct llog_ctxt        *ctxt;
 862         int                      rc;
 863
 864         ENTRY;
 865
 866         /* setup only remote ctxt, the local disk context is switched per each
 867          * filesystem during mgc_fs_setup() */
 868         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 869                         &llog_client_ops);
 870         if (rc)
 871                 RETURN(rc);
 872
 873         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 874         LASSERT(ctxt);
 875
 876         llog_initiator_connect(ctxt);
 877         llog_ctxt_put(ctxt);
 878
 879         RETURN(0);
 880 }
 881
 882 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 883 {
 884         struct llog_ctxt *ctxt;
 885
 886         ENTRY;
 887
 888         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 889         if (ctxt)
 890                 llog_cleanup(env, ctxt);
 891
 892         RETURN(0);
 893 }
 894
 895
 896 static atomic_t mgc_count = ATOMIC_INIT(0);
 897 static int mgc_precleanup(struct obd_device *obd)
 898 {
 899         int     rc = 0;
 900         int     temp;
 901         ENTRY;
 902
 903         if (atomic_dec_and_test(&mgc_count)) {
 904                 LASSERT(rq_state & RQ_RUNNING);
 905                 /* stop requeue thread */
 906                 temp = RQ_STOP;
 907         } else {
 908                 /* wakeup requeue thread to clean our cld */
 909                 temp = RQ_NOW | RQ_PRECLEANUP;
 910         }
 911
 912         spin_lock(&config_list_lock);
 913         rq_state |= temp;
 914         spin_unlock(&config_list_lock);
 915         wake_up(&rq_waitq);
 916
 917         if (temp & RQ_STOP)
 918                 wait_for_completion(&rq_exit);
 919         obd_cleanup_client_import(obd);
 920
 921         rc = mgc_llog_fini(NULL, obd);
 922         if (rc != 0)
 923                 CERROR("failed to cleanup llogging subsystems\n");
 924
 925         RETURN(rc);
 926 }
 927
 928 static int mgc_cleanup(struct obd_device *obd)
 929 {
 930         int rc;
 931         ENTRY;
 932
 933         /* COMPAT_146 - old config logs may have added profiles we don't
 934            know about */
 935         if (obd->obd_type->typ_refcnt <= 1)
 936                 /* Only for the last mgc */
 937                 class_del_profiles();
 938
 939         lprocfs_obd_cleanup(obd);
 940         ptlrpcd_decref();
 941
 942         rc = client_obd_cleanup(obd);
 943         RETURN(rc);
 944 }
 945
 946 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 947 {
 948         struct task_struct      *task;
 949         int                      rc;
 950         ENTRY;
 951
 952         rc = ptlrpcd_addref();
 953         if (rc < 0)
 954                 RETURN(rc);
 955
 956         rc = client_obd_setup(obd, lcfg);
 957         if (rc)
 958                 GOTO(err_decref, rc);
 959
 960         rc = mgc_llog_init(NULL, obd);
 961         if (rc) {
 962                 CERROR("failed to setup llogging subsystems\n");
 963                 GOTO(err_cleanup, rc);
 964         }
 965
 966 #ifdef CONFIG_PROC_FS
 967         obd->obd_vars = lprocfs_mgc_obd_vars;
 968         lprocfs_obd_setup(obd);
 969 #endif
 970         sptlrpc_lprocfs_cliobd_attach(obd);
 971
 972         if (atomic_inc_return(&mgc_count) == 1) {
 973                 rq_state = 0;
 974                 init_waitqueue_head(&rq_waitq);
 975
 976                 /* start requeue thread */
 977                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 978                 if (IS_ERR(task)) {
 979                         rc = PTR_ERR(task);
 980                         CERROR("%s: cannot start requeue thread: rc = %d; "
 981                                "no more log updates\n",
 982                                obd->obd_name, rc);
 983                         GOTO(err_cleanup, rc);
 984                 }
 985                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 986                 rc = 0;
 987                 wait_for_completion(&rq_start);
 988         }
 989
 990         RETURN(rc);
 991
 992 err_cleanup:
 993         client_obd_cleanup(obd);
 994 err_decref:
 995         ptlrpcd_decref();
 996         RETURN(rc);
 997 }
 998
 999 /* based on ll_mdc_blocking_ast */
1000 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1001                             void *data, int flag)
1002 {
1003         struct lustre_handle lockh;
1004         struct config_llog_data *cld = (struct config_llog_data *)data;
1005         int rc = 0;
1006         ENTRY;
1007
1008         switch (flag) {
1009         case LDLM_CB_BLOCKING:
1010                 /* mgs wants the lock, give it up... */
1011                 LDLM_DEBUG(lock, "MGC blocking CB");
1012                 ldlm_lock2handle(lock, &lockh);
1013                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
1014                 break;
1015         case LDLM_CB_CANCELING:
1016                 /* We've given up the lock, prepare ourselves to update. */
1017                 LDLM_DEBUG(lock, "MGC cancel CB");
1018
1019                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
1020                        PLDLMRES(lock->l_resource),
1021                        (char *)&lock->l_resource->lr_name.name[0]);
1022
1023                 if (!cld) {
1024                         CDEBUG(D_INFO, "missing data, won't requeue\n");
1025                         break;
1026                 }
1027
1028                 /* held at mgc_process_log(). */
1029                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1030
1031                 lock->l_ast_data = NULL;
1032                 /* Are we done with this log? */
1033                 if (cld->cld_stopping) {
1034                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1035                                 cld->cld_logname);
1036                         config_log_put(cld);
1037                         break;
1038                 }
1039                 /* Make sure not to re-enqueue when the mgc is stopping
1040                    (we get called from client_disconnect_export) */
1041                 if (lock->l_conn_export == NULL ||
1042                     lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1043                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1044                                 cld->cld_logname);
1045                         config_log_put(cld);
1046                         break;
1047                 }
1048
1049                 /* Re-enqueue now */
1050                 mgc_requeue_add(cld);
1051                 config_log_put(cld);
1052                 break;
1053         default:
1054                 LBUG();
1055         }
1056
1057         RETURN(rc);
1058 }
1059
1060 /* Not sure where this should go... */
1061 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1062  * that we can have a chance to try the secondary MGS if any. */
1063 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1064                                 + PING_INTERVAL)
1065 #define  MGC_TARGET_REG_LIMIT 10
1066 #define  MGC_SEND_PARAM_LIMIT 10
1067
1068 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1069 /* Send parameter to MGS*/
1070 static int mgc_set_mgs_param(struct obd_export *exp,
1071                              struct mgs_send_param *msp)
1072 {
1073         struct ptlrpc_request *req;
1074         struct mgs_send_param *req_msp, *rep_msp;
1075         int rc;
1076         ENTRY;
1077
1078         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1079                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1080                                         MGS_SET_INFO);
1081         if (!req)
1082                 RETURN(-ENOMEM);
1083
1084         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1085         if (!req_msp) {
1086                 ptlrpc_req_finished(req);
1087                 RETURN(-ENOMEM);
1088         }
1089
1090         memcpy(req_msp, msp, sizeof(*req_msp));
1091         ptlrpc_request_set_replen(req);
1092
1093         /* Limit how long we will wait for the enqueue to complete */
1094         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1095         rc = ptlrpc_queue_wait(req);
1096         if (!rc) {
1097                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1098                 memcpy(msp, rep_msp, sizeof(*rep_msp));
1099         }
1100
1101         ptlrpc_req_finished(req);
1102
1103         RETURN(rc);
1104 }
1105 #endif
1106
1107 /* Take a config lock so we can get cancel notifications */
1108 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1109                        union ldlm_policy_data *policy, enum ldlm_mode mode,
1110                        __u64 *flags, ldlm_glimpse_callback glimpse_callback,
1111                        void *data, __u32 lvb_len, void *lvb_swabber,
1112                        struct lustre_handle *lockh)
1113 {
1114         struct config_llog_data *cld = (struct config_llog_data *)data;
1115         struct ldlm_enqueue_info einfo = {
1116                 .ei_type        = type,
1117                 .ei_mode        = mode,
1118                 .ei_cb_bl       = mgc_blocking_ast,
1119                 .ei_cb_cp       = ldlm_completion_ast,
1120                 .ei_cb_gl       = glimpse_callback,
1121         };
1122         struct ptlrpc_request *req;
1123         int short_limit = cld_is_sptlrpc(cld);
1124         int rc;
1125         ENTRY;
1126
1127         CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1128                cld->cld_resid.name[0]);
1129
1130         /* We need a callback for every lockholder, so don't try to
1131            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1132         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1133                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1134                                         LDLM_ENQUEUE);
1135         if (req == NULL)
1136                 RETURN(-ENOMEM);
1137
1138         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1139         ptlrpc_request_set_replen(req);
1140
1141         /* check if this is server or client */
1142         if (cld->cld_cfg.cfg_sb) {
1143                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1144                 if (lsi && IS_SERVER(lsi))
1145                         short_limit = 1;
1146         }
1147         /* Limit how long we will wait for the enqueue to complete */
1148         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1149         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1150                               NULL, 0, LVB_T_NONE, lockh, 0);
1151         /* A failed enqueue should still call the mgc_blocking_ast,
1152            where it will be requeued if needed ("grant failed"). */
1153         ptlrpc_req_finished(req);
1154         RETURN(rc);
1155 }
1156
1157 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1158                       struct lustre_handle *lockh)
1159 {
1160         ENTRY;
1161
1162         ldlm_lock_decref(lockh, mode);
1163
1164         RETURN(0);
1165 }
1166
1167 static void mgc_notify_active(struct obd_device *unused)
1168 {
1169         /* wakeup mgc_requeue_thread to requeue mgc lock */
1170         spin_lock(&config_list_lock);
1171         rq_state |= RQ_NOW;
1172         spin_unlock(&config_list_lock);
1173         wake_up(&rq_waitq);
1174
1175         /* TODO: Help the MGS rebuild nidtbl. -jay */
1176 }
1177
1178 /* Send target_reg message to MGS */
1179 static int mgc_target_register(struct obd_export *exp,
1180                                struct mgs_target_info *mti)
1181 {
1182         struct ptlrpc_request  *req;
1183         struct mgs_target_info *req_mti, *rep_mti;
1184         int                     rc;
1185         ENTRY;
1186
1187         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1188                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1189                                         MGS_TARGET_REG);
1190         if (req == NULL)
1191                 RETURN(-ENOMEM);
1192
1193         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1194         if (!req_mti) {
1195                 ptlrpc_req_finished(req);
1196                 RETURN(-ENOMEM);
1197         }
1198
1199         memcpy(req_mti, mti, sizeof(*req_mti));
1200         ptlrpc_request_set_replen(req);
1201         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1202         /* Limit how long we will wait for the enqueue to complete */
1203         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1204
1205         rc = ptlrpc_queue_wait(req);
1206         if (!rc) {
1207                 rep_mti = req_capsule_server_get(&req->rq_pill,
1208                                                  &RMF_MGS_TARGET_INFO);
1209                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1210                 CDEBUG(D_MGC, "register %s got index = %d\n",
1211                        mti->mti_svname, mti->mti_stripe_index);
1212         }
1213         ptlrpc_req_finished(req);
1214
1215         RETURN(rc);
1216 }
1217
1218 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1219                               u32 keylen, void *key,
1220                               u32 vallen, void *val,
1221                               struct ptlrpc_request_set *set)
1222 {
1223         int rc = -EINVAL;
1224         ENTRY;
1225
1226         /* Turn off initial_recov after we try all backup servers once */
1227         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1228                 struct obd_import *imp = class_exp2cliimp(exp);
1229                 int value;
1230                 if (vallen != sizeof(int))
1231                         RETURN(-EINVAL);
1232                 value = *(int *)val;
1233                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1234                        imp->imp_obd->obd_name, value,
1235                        imp->imp_deactive, imp->imp_invalid,
1236                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1237                        ptlrpc_import_state_name(imp->imp_state));
1238                 /* Resurrect if we previously died */
1239                 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1240                      imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1241                         ptlrpc_reconnect_import(imp);
1242                 RETURN(0);
1243         }
1244         /* FIXME move this to mgc_process_config */
1245         if (KEY_IS(KEY_REGISTER_TARGET)) {
1246                 struct mgs_target_info *mti;
1247                 if (vallen != sizeof(struct mgs_target_info))
1248                         RETURN(-EINVAL);
1249                 mti = (struct mgs_target_info *)val;
1250                 CDEBUG(D_MGC, "register_target %s %#x\n",
1251                        mti->mti_svname, mti->mti_flags);
1252                 rc =  mgc_target_register(exp, mti);
1253                 RETURN(rc);
1254         }
1255         if (KEY_IS(KEY_SET_FS)) {
1256                 struct super_block *sb = (struct super_block *)val;
1257
1258                 if (vallen != sizeof(struct super_block))
1259                         RETURN(-EINVAL);
1260
1261                 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1262                 RETURN(rc);
1263         }
1264         if (KEY_IS(KEY_CLEAR_FS)) {
1265                 if (vallen != 0)
1266                         RETURN(-EINVAL);
1267                 rc = mgc_fs_cleanup(env, exp->exp_obd);
1268                 RETURN(rc);
1269         }
1270 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1271         if (KEY_IS(KEY_SET_INFO)) {
1272                 struct mgs_send_param *msp;
1273
1274                 msp = (struct mgs_send_param *)val;
1275                 rc =  mgc_set_mgs_param(exp, msp);
1276                 RETURN(rc);
1277         }
1278 #endif
1279         if (KEY_IS(KEY_MGSSEC)) {
1280                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1281                 struct sptlrpc_flavor  flvr;
1282
1283                 /*
1284                  * empty string means using current flavor, if which haven't
1285                  * been set yet, set it as null.
1286                  *
1287                  * if flavor has been set previously, check the asking flavor
1288                  * must match the existing one.
1289                  */
1290                 if (vallen == 0) {
1291                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1292                                 RETURN(0);
1293                         val = "null";
1294                         vallen = 4;
1295                 }
1296
1297                 rc = sptlrpc_parse_flavor(val, &flvr);
1298                 if (rc) {
1299                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1300                                (char *) val);
1301                         RETURN(rc);
1302                 }
1303
1304                 /*
1305                  * caller already hold a mutex
1306                  */
1307                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1308                         cli->cl_flvr_mgc = flvr;
1309                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1310                                   sizeof(flvr)) != 0) {
1311                         char    str[20];
1312
1313                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1314                                             str, sizeof(str));
1315                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1316                                        "currently %s is in use\n",
1317                                        (char *) val, str);
1318                         rc = -EPERM;
1319                 }
1320                 RETURN(rc);
1321         }
1322
1323         RETURN(rc);
1324 }
1325
1326 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1327                         __u32 keylen, void *key, __u32 *vallen, void *val)
1328 {
1329         int rc = -EINVAL;
1330
1331         if (KEY_IS(KEY_CONN_DATA)) {
1332                 struct obd_import *imp = class_exp2cliimp(exp);
1333                 struct obd_connect_data *data = val;
1334
1335                 if (*vallen == sizeof(*data)) {
1336                         *data = imp->imp_connect_data;
1337                         rc = 0;
1338                 }
1339         }
1340
1341         return rc;
1342 }
1343
1344 static int mgc_import_event(struct obd_device *obd,
1345                             struct obd_import *imp,
1346                             enum obd_import_event event)
1347 {
1348         int rc = 0;
1349
1350         LASSERT(imp->imp_obd == obd);
1351         CDEBUG(D_MGC, "import event %#x\n", event);
1352
1353         switch (event) {
1354         case IMP_EVENT_DISCON:
1355                 /* MGC imports should not wait for recovery */
1356                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1357                         ptlrpc_pinger_ir_down();
1358                 break;
1359         case IMP_EVENT_INACTIVE:
1360                 break;
1361         case IMP_EVENT_INVALIDATE: {
1362                 struct ldlm_namespace *ns = obd->obd_namespace;
1363                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1364                 break;
1365         }
1366         case IMP_EVENT_ACTIVE:
1367                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1368                 /* Clearing obd_no_recov allows us to continue pinging */
1369                 obd->obd_no_recov = 0;
1370                 mgc_notify_active(obd);
1371                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1372                         ptlrpc_pinger_ir_up();
1373                 break;
1374         case IMP_EVENT_OCD:
1375                 break;
1376         case IMP_EVENT_DEACTIVATE:
1377         case IMP_EVENT_ACTIVATE:
1378                 break;
1379         default:
1380                 CERROR("Unknown import event %#x\n", event);
1381                 LBUG();
1382         }
1383         RETURN(rc);
1384 }
1385
1386 enum {
1387         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1388         CONFIG_READ_NRPAGES      = 4
1389 };
1390
1391 static int mgc_apply_recover_logs(struct obd_device *mgc,
1392                                   struct config_llog_data *cld,
1393                                   __u64 max_version,
1394                                   void *data, int datalen, bool mne_swab)
1395 {
1396         struct config_llog_instance *cfg = &cld->cld_cfg;
1397         struct lustre_sb_info       *lsi = s2lsi(cfg->cfg_sb);
1398         struct mgs_nidtbl_entry *entry;
1399         struct lustre_cfg       *lcfg;
1400         struct lustre_cfg_bufs   bufs;
1401         u64   prev_version = 0;
1402         char *inst;
1403         char *buf;
1404         int   bufsz;
1405         int   pos;
1406         int   rc  = 0;
1407         int   off = 0;
1408         ENTRY;
1409
1410         LASSERT(cfg->cfg_instance != NULL);
1411         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1412
1413         OBD_ALLOC(inst, PAGE_SIZE);
1414         if (inst == NULL)
1415                 RETURN(-ENOMEM);
1416
1417         if (!IS_SERVER(lsi)) {
1418                 pos = snprintf(inst, PAGE_SIZE, "%p", cfg->cfg_instance);
1419                 if (pos >= PAGE_SIZE) {
1420                         OBD_FREE(inst, PAGE_SIZE);
1421                         return -E2BIG;
1422                 }
1423         } else {
1424                 LASSERT(IS_MDT(lsi));
1425                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1426                                         PAGE_SIZE);
1427                 if (rc) {
1428                         OBD_FREE(inst, PAGE_SIZE);
1429                         RETURN(-EINVAL);
1430                 }
1431                 pos = strlen(inst);
1432         }
1433
1434         ++pos;
1435         buf   = inst + pos;
1436         bufsz = PAGE_SIZE - pos;
1437
1438         while (datalen > 0) {
1439                 int   entry_len = sizeof(*entry);
1440                 int   is_ost, i;
1441                 struct obd_device *obd;
1442                 char *obdname;
1443                 char *cname;
1444                 char *params;
1445                 char *uuid;
1446
1447                 rc = -EINVAL;
1448                 if (datalen < sizeof(*entry))
1449                         break;
1450
1451                 entry = (typeof(entry))(data + off);
1452
1453                 /* sanity check */
1454                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1455                         break;
1456                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1457                         break;
1458                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1459                         break;
1460
1461                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1462                 if (datalen < entry_len) /* must have entry_len at least */
1463                         break;
1464
1465                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1466                 if (mne_swab)
1467                         lustre_swab_mgs_nidtbl_entry(entry);
1468                 if (entry->mne_length > PAGE_SIZE) {
1469                         CERROR("MNE too large (%u)\n", entry->mne_length);
1470                         break;
1471                 }
1472
1473                 if (entry->mne_length < entry_len)
1474                         break;
1475
1476                 off     += entry->mne_length;
1477                 datalen -= entry->mne_length;
1478                 if (datalen < 0)
1479                         break;
1480
1481                 if (entry->mne_version > max_version) {
1482                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1483                                entry->mne_version, max_version);
1484                         break;
1485                 }
1486
1487                 if (prev_version >= entry->mne_version) {
1488                         CERROR("index unsorted, prev %lld, now %lld\n",
1489                                prev_version, entry->mne_version);
1490                         break;
1491                 }
1492                 prev_version = entry->mne_version;
1493
1494                 /*
1495                  * Write a string with format "nid::instance" to
1496                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1497                  */
1498
1499                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1500                 memset(buf, 0, bufsz);
1501                 obdname = buf;
1502                 pos = 0;
1503
1504                 /* lustre-OST0001-osc-<instance #> */
1505                 strcpy(obdname, cld->cld_logname);
1506                 cname = strrchr(obdname, '-');
1507                 if (cname == NULL) {
1508                         CERROR("mgc %s: invalid logname %s\n",
1509                                mgc->obd_name, obdname);
1510                         break;
1511                 }
1512
1513                 pos = cname - obdname;
1514                 obdname[pos] = 0;
1515                 pos += sprintf(obdname + pos, "-%s%04x",
1516                                   is_ost ? "OST" : "MDT", entry->mne_index);
1517
1518                 cname = is_ost ? "osc" : "mdc",
1519                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1520                 lustre_cfg_bufs_reset(&bufs, obdname);
1521
1522                 /* find the obd by obdname */
1523                 obd = class_name2obd(obdname);
1524                 if (obd == NULL) {
1525                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1526                                mgc->obd_name, obdname);
1527                         rc = 0;
1528                         /* this is a safe race, when the ost is starting up...*/
1529                         continue;
1530                 }
1531
1532                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1533                 ++pos;
1534                 params = buf + pos;
1535                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1536                 uuid = buf + pos;
1537
1538                 down_read(&obd->u.cli.cl_sem);
1539                 if (obd->u.cli.cl_import == NULL) {
1540                         /* client does not connect to the OST yet */
1541                         up_read(&obd->u.cli.cl_sem);
1542                         rc = 0;
1543                         continue;
1544                 }
1545
1546                 /* iterate all nids to find one */
1547                 /* find uuid by nid */
1548                 rc = -ENOENT;
1549                 for (i = 0; i < entry->mne_nid_count; i++) {
1550                         rc = client_import_find_conn(obd->u.cli.cl_import,
1551                                                      entry->u.nids[i],
1552                                                      (struct obd_uuid *)uuid);
1553                         if (rc == 0)
1554                                 break;
1555                 }
1556
1557                 up_read(&obd->u.cli.cl_sem);
1558                 if (rc < 0) {
1559                         CERROR("mgc: cannot find uuid by nid %s\n",
1560                                libcfs_nid2str(entry->u.nids[0]));
1561                         break;
1562                 }
1563
1564                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1565                        uuid, libcfs_nid2str(entry->u.nids[0]));
1566
1567                 pos += strlen(uuid);
1568                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1569                 LASSERT(pos < bufsz);
1570
1571                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1572
1573                 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1574                 if (lcfg == NULL) {
1575                         rc = -ENOMEM;
1576                         break;
1577                 }
1578
1579                 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1580                        prev_version, max_version, obdname, params);
1581
1582                 rc = class_process_config(lcfg);
1583                 lustre_cfg_free(lcfg);
1584                 if (rc)
1585                         CDEBUG(D_INFO, "process config for %s error %d\n",
1586                                obdname, rc);
1587
1588                 /* continue, even one with error */
1589         }
1590
1591         OBD_FREE(inst, PAGE_SIZE);
1592         RETURN(rc);
1593 }
1594
1595 /**
1596  * This function is called if this client was notified for target restarting
1597  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1598  * nodemap logs.
1599  */
1600 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1601                                            struct config_llog_data *cld)
1602 {
1603         struct ptlrpc_connection *mgc_conn;
1604         struct ptlrpc_request *req = NULL;
1605         struct config_llog_instance *cfg = &cld->cld_cfg;
1606         struct mgs_config_body *body;
1607         struct mgs_config_res *res;
1608         struct nodemap_config *new_config = NULL;
1609         struct lu_nodemap *recent_nodemap = NULL;
1610         struct ptlrpc_bulk_desc *desc;
1611         struct page **pages = NULL;
1612         __u64 config_read_offset = 0;
1613         __u8 nodemap_cur_pass = 0;
1614         int nrpages = 0;
1615         bool eof = true;
1616         bool mne_swab = false;
1617         int i;
1618         int ealen;
1619         int rc;
1620         ENTRY;
1621
1622         mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1623
1624         /* don't need to get local config */
1625         if (cld_is_nodemap(cld) &&
1626             (LNET_NETTYP(LNET_NIDNET(mgc_conn->c_peer.nid)) == LOLND))
1627                 GOTO(out, rc = 0);
1628
1629         /* allocate buffer for bulk transfer.
1630          * if this is the first time for this mgs to read logs,
1631          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1632          * once; otherwise, it only reads increment of logs, this should be
1633          * small and CONFIG_READ_NRPAGES will be used.
1634          */
1635         nrpages = CONFIG_READ_NRPAGES;
1636         if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1637                 nrpages = CONFIG_READ_NRPAGES_INIT;
1638
1639         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1640         if (pages == NULL)
1641                 GOTO(out, rc = -ENOMEM);
1642
1643         for (i = 0; i < nrpages; i++) {
1644                 pages[i] = alloc_page(GFP_KERNEL);
1645                 if (pages[i] == NULL)
1646                         GOTO(out, rc = -ENOMEM);
1647         }
1648
1649 again:
1650 #ifdef HAVE_SERVER_SUPPORT
1651         if (cld_is_nodemap(cld) && config_read_offset == 0) {
1652                 new_config = nodemap_config_alloc();
1653                 if (IS_ERR(new_config)) {
1654                         rc = PTR_ERR(new_config);
1655                         new_config = NULL;
1656                         GOTO(out, rc);
1657                 }
1658         }
1659 #endif
1660         LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1661         LASSERT(mutex_is_locked(&cld->cld_lock));
1662         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1663                                    &RQF_MGS_CONFIG_READ);
1664         if (req == NULL)
1665                 GOTO(out, rc = -ENOMEM);
1666
1667         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1668         if (rc)
1669                 GOTO(out, rc);
1670
1671         /* pack request */
1672         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1673         LASSERT(body != NULL);
1674         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1675         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1676             >= sizeof(body->mcb_name))
1677                 GOTO(out, rc = -E2BIG);
1678         if (cld_is_nodemap(cld))
1679                 body->mcb_offset = config_read_offset;
1680         else
1681                 body->mcb_offset = cfg->cfg_last_idx + 1;
1682         body->mcb_type   = cld->cld_type;
1683         body->mcb_bits   = PAGE_SHIFT;
1684         body->mcb_units  = nrpages;
1685         body->mcb_nm_cur_pass = nodemap_cur_pass;
1686
1687         /* allocate bulk transfer descriptor */
1688         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1689                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1690                                     MGS_BULK_PORTAL,
1691                                     &ptlrpc_bulk_kiov_pin_ops);
1692         if (desc == NULL)
1693                 GOTO(out, rc = -ENOMEM);
1694
1695         for (i = 0; i < nrpages; i++)
1696                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1697                                                  PAGE_SIZE);
1698
1699         ptlrpc_request_set_replen(req);
1700         rc = ptlrpc_queue_wait(req);
1701         if (rc)
1702                 GOTO(out, rc);
1703
1704         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1705         if (!res)
1706                 GOTO(out, rc = -EPROTO);
1707
1708         if (cld_is_nodemap(cld)) {
1709                 config_read_offset = res->mcr_offset;
1710                 eof = config_read_offset == II_END_OFF;
1711                 nodemap_cur_pass = res->mcr_nm_cur_pass;
1712         } else {
1713                 if (res->mcr_size < res->mcr_offset)
1714                         GOTO(out, rc = -EINVAL);
1715
1716                 /* always update the index even though it might have errors with
1717                  * handling the recover logs
1718                  */
1719                 cfg->cfg_last_idx = res->mcr_offset;
1720                 eof = res->mcr_offset == res->mcr_size;
1721
1722                 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1723                        res->mcr_offset, eof == false);
1724         }
1725
1726         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1727         if (ealen < 0)
1728                 GOTO(out, rc = ealen);
1729
1730         if (ealen > nrpages << PAGE_SHIFT)
1731                 GOTO(out, rc = -EINVAL);
1732
1733         if (ealen == 0) { /* no logs transferred */
1734 #ifdef HAVE_SERVER_SUPPORT
1735                 /* config changed since first read RPC */
1736                 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1737                         recent_nodemap = NULL;
1738                         nodemap_config_dealloc(new_config);
1739                         new_config = NULL;
1740
1741                         CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1742
1743                         /* setting eof to false, we request config again */
1744                         eof = false;
1745                         GOTO(out, rc = 0);
1746                 }
1747 #endif
1748                 if (!eof)
1749                         rc = -EINVAL;
1750                 GOTO(out, rc);
1751         }
1752
1753         mne_swab = !!ptlrpc_rep_need_swab(req);
1754 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1755         /* This import flag means the server did an extra swab of IR MNE
1756          * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1757         if (unlikely(req->rq_import->imp_need_mne_swab))
1758                 mne_swab = !mne_swab;
1759 #endif
1760
1761         /* When a nodemap config is received, we build a new nodemap config,
1762          * with new nodemap structs. We keep track of the most recently added
1763          * nodemap since the config is read ordered by nodemap_id, and so it
1764          * is likely that the next record will be related. Because access to
1765          * the nodemaps is single threaded until the nodemap_config is active,
1766          * we don't need to reference count with recent_nodemap, though
1767          * recent_nodemap should be set to NULL when the nodemap_config
1768          * is either destroyed or set active.
1769          */
1770         for (i = 0; i < nrpages && ealen > 0; i++) {
1771                 int rc2;
1772                 union lu_page   *ptr;
1773
1774                 ptr = kmap(pages[i]);
1775                 if (cld_is_nodemap(cld))
1776                         rc2 = nodemap_process_idx_pages(new_config, ptr,
1777                                                        &recent_nodemap);
1778                 else
1779                         rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1780                                                      ptr,
1781                                                      min_t(int, ealen,
1782                                                            PAGE_SIZE),
1783                                                      mne_swab);
1784                 kunmap(pages[i]);
1785                 if (rc2 < 0) {
1786                         CWARN("%s: error processing %s log %s: rc = %d\n",
1787                               obd->obd_name,
1788                               cld_is_nodemap(cld) ? "nodemap" : "recovery",
1789                               cld->cld_logname,
1790                               rc2);
1791                         GOTO(out, rc = rc2);
1792                 }
1793
1794                 ealen -= PAGE_SIZE;
1795         }
1796
1797 out:
1798         if (req) {
1799                 ptlrpc_req_finished(req);
1800                 req = NULL;
1801         }
1802
1803         if (rc == 0 && !eof)
1804                 goto again;
1805
1806 #ifdef HAVE_SERVER_SUPPORT
1807         if (new_config != NULL) {
1808                 /* recent_nodemap cannot be used after set_active/dealloc */
1809                 if (rc == 0)
1810                         nodemap_config_set_active_mgc(new_config);
1811                 else
1812                         nodemap_config_dealloc(new_config);
1813         }
1814 #endif
1815
1816         if (pages) {
1817                 for (i = 0; i < nrpages; i++) {
1818                         if (pages[i] == NULL)
1819                                 break;
1820                         __free_page(pages[i]);
1821                 }
1822                 OBD_FREE(pages, sizeof(*pages) * nrpages);
1823         }
1824         return rc;
1825 }
1826
1827 static int mgc_barrier_glimpse_ast(struct ldlm_lock *lock, void *data)
1828 {
1829         struct config_llog_data *cld = lock->l_ast_data;
1830         int rc;
1831         ENTRY;
1832
1833         if (cld->cld_stopping)
1834                 RETURN(-ENODEV);
1835
1836         rc = barrier_handler(s2lsi(cld->cld_cfg.cfg_sb)->lsi_dt_dev,
1837                              (struct ptlrpc_request *)data);
1838
1839         RETURN(rc);
1840 }
1841
1842 /* Copy a remote log locally */
1843 static int mgc_llog_local_copy(const struct lu_env *env,
1844                                struct obd_device *obd,
1845                                struct llog_ctxt *rctxt,
1846                                struct llog_ctxt *lctxt, char *logname)
1847 {
1848         char    *temp_log;
1849         int      rc;
1850
1851         ENTRY;
1852
1853         /*
1854          * - copy it to backup using llog_backup()
1855          * - copy remote llog to logname using llog_backup()
1856          * - if failed then move bakup to logname again
1857          */
1858
1859         OBD_ALLOC(temp_log, strlen(logname) + 2);
1860         if (!temp_log)
1861                 RETURN(-ENOMEM);
1862         sprintf(temp_log, "%sT", logname);
1863
1864         /* make a copy of local llog at first */
1865         rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1866         if (rc < 0 && rc != -ENOENT)
1867                 GOTO(out, rc);
1868         /* copy remote llog to the local copy */
1869         rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1870         if (rc == -ENOENT) {
1871                 /* no remote llog, delete local one too */
1872                 llog_erase(env, lctxt, NULL, logname);
1873         } else if (rc < 0) {
1874                 /* error during backup, get local one back from the copy */
1875                 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1876 out:
1877                 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1878                        obd->obd_name, logname, rc);
1879         }
1880         llog_erase(env, lctxt, NULL, temp_log);
1881         OBD_FREE(temp_log, strlen(logname) + 2);
1882         return rc;
1883 }
1884
1885 /* local_only means it cannot get remote llogs */
1886 static int mgc_process_cfg_log(struct obd_device *mgc,
1887                                struct config_llog_data *cld, int local_only)
1888 {
1889         struct llog_ctxt        *ctxt, *lctxt = NULL;
1890         struct client_obd       *cli = &mgc->u.cli;
1891         struct lustre_sb_info   *lsi = NULL;
1892         int                      rc = 0;
1893         struct lu_env           *env;
1894
1895         ENTRY;
1896
1897         LASSERT(cld);
1898         LASSERT(mutex_is_locked(&cld->cld_lock));
1899
1900         if (cld->cld_cfg.cfg_sb)
1901                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1902
1903         OBD_ALLOC_PTR(env);
1904         if (env == NULL)
1905                 RETURN(-ENOMEM);
1906
1907         rc = lu_env_init(env, LCT_MG_THREAD);
1908         if (rc)
1909                 GOTO(out_free, rc);
1910
1911         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1912         LASSERT(ctxt);
1913
1914         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1915
1916         /* Copy the setup log locally if we can. Don't mess around if we're
1917          * running an MGS though (logs are already local). */
1918         if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1919             cli->cl_mgc_configs_dir != NULL &&
1920             lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1921             lsi->lsi_dt_dev) {
1922                 if (!local_only && !lsi->lsi_dt_dev->dd_rdonly)
1923                         /* Only try to copy log if we have the lock. */
1924                         rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1925                                                  cld->cld_logname);
1926                 if (local_only || rc) {
1927                         if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1928                             llog_is_empty(env, lctxt, cld->cld_logname)) {
1929                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1930                                                    "log %s and no local copy."
1931                                                    "\n", cld->cld_logname);
1932                                 GOTO(out_pop, rc = -ENOENT);
1933                         }
1934                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1935                                "copy for now, will try to update later.\n",
1936                                cld->cld_logname);
1937                         rc = 0;
1938                 }
1939                 /* Now, whether we copied or not, start using the local llog.
1940                  * If we failed to copy, we'll start using whatever the old
1941                  * log has. */
1942                 llog_ctxt_put(ctxt);
1943                 ctxt = lctxt;
1944                 lctxt = NULL;
1945         } else {
1946                 if (local_only) /* no local log at client side */
1947                         GOTO(out_pop, rc = -EIO);
1948         }
1949
1950         rc = -EAGAIN;
1951         if (lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1952             lsi->lsi_dt_dev->dd_rdonly) {
1953                 struct llog_ctxt *rctxt;
1954
1955                 /* Under readonly mode, we may have no local copy or local
1956                  * copy is incomplete, so try to use remote llog firstly. */
1957                 rctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1958                 LASSERT(rctxt);
1959
1960                 rc = class_config_parse_llog(env, rctxt, cld->cld_logname,
1961                                              &cld->cld_cfg);
1962                 llog_ctxt_put(rctxt);
1963         }
1964
1965         if (rc && rc != -ENOENT)
1966                 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1967                                              &cld->cld_cfg);
1968
1969         /*
1970          * update settings on existing OBDs. doing it inside
1971          * of llog_process_lock so no device is attaching/detaching
1972          * in parallel.
1973          * the logname must be <fsname>-sptlrpc
1974          */
1975         if (rc == 0 && cld_is_sptlrpc(cld))
1976                 class_notify_sptlrpc_conf(cld->cld_logname,
1977                                           strlen(cld->cld_logname) -
1978                                           strlen("-sptlrpc"));
1979         EXIT;
1980
1981 out_pop:
1982         __llog_ctxt_put(env, ctxt);
1983         if (lctxt)
1984                 __llog_ctxt_put(env, lctxt);
1985
1986         lu_env_fini(env);
1987 out_free:
1988         OBD_FREE_PTR(env);
1989         return rc;
1990 }
1991
1992 static bool mgc_import_in_recovery(struct obd_import *imp)
1993 {
1994         bool in_recovery = true;
1995
1996         spin_lock(&imp->imp_lock);
1997         if (imp->imp_state == LUSTRE_IMP_FULL ||
1998             imp->imp_state == LUSTRE_IMP_CLOSED)
1999                 in_recovery = false;
2000         spin_unlock(&imp->imp_lock);
2001
2002         return in_recovery;
2003 }
2004
2005 /**
2006  * Get a configuration log from the MGS and process it.
2007  *
2008  * This function is called for both clients and servers to process the
2009  * configuration log from the MGS.  The MGC enqueues a DLM lock on the
2010  * log from the MGS, and if the lock gets revoked the MGC will be notified
2011  * by the lock cancellation callback that the config log has changed,
2012  * and will enqueue another MGS lock on it, and then continue processing
2013  * the new additions to the end of the log.
2014  *
2015  * Since the MGC import is not replayable, if the import is being evicted
2016  * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
2017  * the log until recovery is finished or the import is closed.
2018  *
2019  * Make a local copy of the log before parsing it if appropriate (non-MGS
2020  * server) so that the server can start even when the MGS is down.
2021  *
2022  * There shouldn't be multiple processes running process_log at once --
2023  * sounds like badness.  It actually might be fine, as long as they're not
2024  * trying to update from the same log simultaneously, in which case we
2025  * should use a per-log semaphore instead of cld_lock.
2026  *
2027  * \param[in] mgc       MGC device by which to fetch the configuration log
2028  * \param[in] cld       log processing state (stored in lock callback data)
2029  *
2030  * \retval              0 on success
2031  * \retval              negative errno on failure
2032  */
2033 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
2034 {
2035         struct lustre_handle lockh = { 0 };
2036         __u64 flags = LDLM_FL_NO_LRU;
2037         int rc = 0, rcl;
2038         bool retry = false;
2039         ENTRY;
2040
2041         LASSERT(cld != NULL);
2042
2043         /* I don't want multiple processes running process_log at once --
2044            sounds like badness.  It actually might be fine, as long as
2045            we're not trying to update from the same log
2046            simultaneously (in which case we should use a per-log sem.) */
2047 restart:
2048         mutex_lock(&cld->cld_lock);
2049         if (cld->cld_stopping) {
2050                 mutex_unlock(&cld->cld_lock);
2051                 RETURN(0);
2052         }
2053
2054         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2055
2056         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
2057                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2058
2059         /* Get the cfg lock on the llog */
2060         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2061                           LCK_CR, &flags,
2062                           cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
2063                           cld, 0, NULL, &lockh);
2064         if (rcl == 0) {
2065                 /* Get the cld, it will be released in mgc_blocking_ast. */
2066                 config_log_get(cld);
2067                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2068                 LASSERT(rc == 0);
2069         } else {
2070                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2071
2072                 if (rcl == -ESHUTDOWN &&
2073                     atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2074                         struct obd_import *imp;
2075                         struct l_wait_info lwi;
2076                         int secs = cfs_time_seconds(obd_timeout);
2077
2078                         mutex_unlock(&cld->cld_lock);
2079                         imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2080
2081                         /* Let's force the pinger, and wait the import to be
2082                          * connected, note: since mgc import is non-replayable,
2083                          * and even the import state is disconnected, it does
2084                          * not mean the "recovery" is stopped, so we will keep
2085                          * waitting until timeout or the import state is
2086                          * FULL or closed */
2087                         ptlrpc_pinger_force(imp);
2088
2089                         lwi = LWI_TIMEOUT(secs, NULL, NULL);
2090                         l_wait_event(imp->imp_recovery_waitq,
2091                                      !mgc_import_in_recovery(imp), &lwi);
2092
2093                         if (imp->imp_state == LUSTRE_IMP_FULL) {
2094                                 retry = true;
2095                                 goto restart;
2096                         } else {
2097                                 mutex_lock(&cld->cld_lock);
2098                                 spin_lock(&config_list_lock);
2099                                 cld->cld_lostlock = 1;
2100                                 spin_unlock(&config_list_lock);
2101                         }
2102                 } else {
2103                         /* mark cld_lostlock so that it will requeue
2104                          * after MGC becomes available. */
2105                         spin_lock(&config_list_lock);
2106                         cld->cld_lostlock = 1;
2107                         spin_unlock(&config_list_lock);
2108                 }
2109         }
2110
2111         if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2112                 if (!rcl)
2113                         rc = mgc_process_recover_nodemap_log(mgc, cld);
2114                 else if (cld_is_nodemap(cld))
2115                         rc = rcl;
2116
2117                 if (cld_is_recover(cld) && rc) {
2118                         if (!rcl) {
2119                                 CERROR("%s: recover log %s failed, not fatal: rc = %d\n",
2120                                        mgc->obd_name, cld->cld_logname, rc);
2121                                 spin_lock(&config_list_lock);
2122                                 cld->cld_lostlock = 1;
2123                                 spin_unlock(&config_list_lock);
2124                         }
2125                         rc = 0; /* this is not a fatal error for recover log */
2126                 }
2127         } else if (!cld_is_barrier(cld)) {
2128                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2129         }
2130
2131         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2132                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2133
2134         mutex_unlock(&cld->cld_lock);
2135
2136         /* Now drop the lock so MGS can revoke it */
2137         if (!rcl) {
2138                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2139                 if (rcl)
2140                         CERROR("Can't drop cfg lock: %d\n", rcl);
2141         }
2142
2143         RETURN(rc);
2144 }
2145
2146
2147 /** Called from lustre_process_log.
2148  * LCFG_LOG_START gets the config log from the MGS, processes it to start
2149  * any services, and adds it to the list logs to watch (follow).
2150  */
2151 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2152 {
2153         struct lustre_cfg *lcfg = buf;
2154         struct config_llog_instance *cfg = NULL;
2155         char *logname;
2156         int rc = 0;
2157         ENTRY;
2158
2159         switch(lcfg->lcfg_command) {
2160         case LCFG_LOV_ADD_OBD: {
2161                 /* Overloading this cfg command: register a new target */
2162                 struct mgs_target_info *mti;
2163
2164                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2165                     sizeof(struct mgs_target_info))
2166                         GOTO(out, rc = -EINVAL);
2167
2168                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2169                 CDEBUG(D_MGC, "add_target %s %#x\n",
2170                        mti->mti_svname, mti->mti_flags);
2171                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2172                 break;
2173         }
2174         case LCFG_LOV_DEL_OBD:
2175                 /* Unregister has no meaning at the moment. */
2176                 CERROR("lov_del_obd unimplemented\n");
2177                 rc = -ENOSYS;
2178                 break;
2179         case LCFG_SPTLRPC_CONF: {
2180                 rc = sptlrpc_process_config(lcfg);
2181                 break;
2182         }
2183         case LCFG_LOG_START: {
2184                 struct config_llog_data *cld;
2185                 struct super_block *sb;
2186
2187                 logname = lustre_cfg_string(lcfg, 1);
2188                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2189                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2190
2191                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2192                        cfg->cfg_last_idx);
2193
2194                 /* We're only called through here on the initial mount */
2195                 cld = config_log_add(obd, logname, cfg, sb);
2196                 if (IS_ERR(cld)) {
2197                         rc = PTR_ERR(cld);
2198                         break;
2199                 }
2200
2201                 /* COMPAT_146 */
2202                 /* FIXME only set this for old logs!  Right now this forces
2203                    us to always skip the "inside markers" check */
2204                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
2205
2206                 rc = mgc_process_log(obd, cld);
2207                 if (rc == 0 && cld->cld_recover != NULL) {
2208                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2209                                          imp_connect_data, IMP_RECOV)) {
2210                                 rc = mgc_process_log(obd, cld->cld_recover);
2211                         } else {
2212                                 struct config_llog_data *cir;
2213
2214                                 mutex_lock(&cld->cld_lock);
2215                                 cir = cld->cld_recover;
2216                                 cld->cld_recover = NULL;
2217                                 mutex_unlock(&cld->cld_lock);
2218                                 config_log_put(cir);
2219                         }
2220
2221                         if (rc)
2222                                 CERROR("Cannot process recover llog %d\n", rc);
2223                 }
2224
2225                 if (rc == 0 && cld->cld_params != NULL) {
2226                         rc = mgc_process_log(obd, cld->cld_params);
2227                         if (rc == -ENOENT) {
2228                                 CDEBUG(D_MGC, "There is no params "
2229                                               "config file yet\n");
2230                                 rc = 0;
2231                         }
2232                         /* params log is optional */
2233                         if (rc)
2234                                 CERROR("%s: can't process params llog: rc = %d\n",
2235                                        obd->obd_name, rc);
2236                 }
2237
2238                 break;
2239         }
2240         case LCFG_LOG_END: {
2241                 logname = lustre_cfg_string(lcfg, 1);
2242
2243                 if (lcfg->lcfg_bufcount >= 2)
2244                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
2245                                 lcfg, 2);
2246                 rc = config_log_end(logname, cfg);
2247                 break;
2248         }
2249         default: {
2250                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2251                 GOTO(out, rc = -EINVAL);
2252
2253         }
2254         }
2255 out:
2256         RETURN(rc);
2257 }
2258
2259 static struct obd_ops mgc_obd_ops = {
2260         .o_owner        = THIS_MODULE,
2261         .o_setup        = mgc_setup,
2262         .o_precleanup   = mgc_precleanup,
2263         .o_cleanup      = mgc_cleanup,
2264         .o_add_conn     = client_import_add_conn,
2265         .o_del_conn     = client_import_del_conn,
2266         .o_connect      = client_connect_import,
2267         .o_disconnect   = client_disconnect_export,
2268         .o_set_info_async = mgc_set_info_async,
2269         .o_get_info       = mgc_get_info,
2270         .o_import_event = mgc_import_event,
2271         .o_process_config = mgc_process_config,
2272 };
2273
2274 static int __init mgc_init(void)
2275 {
2276         return class_register_type(&mgc_obd_ops, NULL, true, NULL,
2277                                    LUSTRE_MGC_NAME, NULL);
2278 }
2279
2280 static void __exit mgc_exit(void)
2281 {
2282         class_unregister_type(LUSTRE_MGC_NAME);
2283 }
2284
2285 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2286 MODULE_DESCRIPTION("Lustre Management Client");
2287 MODULE_VERSION(LUSTRE_VERSION_STRING);
2288 MODULE_LICENSE("GPL");
2289
2290 module_init(mgc_init);
2291 module_exit(mgc_exit);