lustre/mgc/mgc_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lustre/mgc/mgc_request.c
  32  *
  33  * Author: Nathan Rutman <nathan@clusterfs.com>
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_MGC
  37 #define D_MGC D_CONFIG /*|D_WARNING*/
  38
  39 #include <linux/module.h>
  40 #include <linux/kthread.h>
  41 #include <linux/random.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                           enum mgs_cfg_type 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 MGS_CFG_T_CONFIG:
  76         case MGS_CFG_T_SPTLRPC:
  77                 resname = 0;
  78                 break;
  79         case MGS_CFG_T_RECOVER:
  80         case MGS_CFG_T_PARAMS:
  81         case MGS_CFG_T_NODEMAP:
  82         case MGS_CFG_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,
  95                      enum mgs_cfg_type type)
  96 {
  97         /* fsname is at most 8 chars long, maybe contain "-".
  98          * e.g. "lustre", "SUN-000" */
  99         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
 100 }
 101 EXPORT_SYMBOL(mgc_fsname2resid);
 102
 103 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id,
 104                       enum mgs_cfg_type type)
 105 {
 106         char *name_end;
 107         int len;
 108
 109         /* logname consists of "fsname-nodetype".
 110          * e.g. "lustre-MDT0001", "SUN-000-client"
 111          * there is an exception: llog "params" */
 112         name_end = strrchr(logname, '-');
 113         if (!name_end)
 114                 len = strlen(logname);
 115         else
 116                 len = name_end - logname;
 117         return mgc_name2resid(logname, len, res_id, type);
 118 }
 119 EXPORT_SYMBOL(mgc_logname2resid);
 120
 121 /********************** config llog list **********************/
 122 static LIST_HEAD(config_llog_list);
 123 static DEFINE_SPINLOCK(config_list_lock);       /* protects config_llog_list */
 124
 125 /* Take a reference to a config log */
 126 static int config_log_get(struct config_llog_data *cld)
 127 {
 128         ENTRY;
 129         atomic_inc(&cld->cld_refcount);
 130         CDEBUG(D_INFO, "log %s (%p) refs %d\n", cld->cld_logname, cld,
 131                 atomic_read(&cld->cld_refcount));
 132         RETURN(0);
 133 }
 134
 135 /* Drop a reference to a config log.  When no longer referenced,
 136    we can free the config log data */
 137 static void config_log_put(struct config_llog_data *cld)
 138 {
 139         ENTRY;
 140
 141         if (unlikely(!cld))
 142                 RETURN_EXIT;
 143
 144         CDEBUG(D_INFO, "log %s(%p) refs %d\n", cld->cld_logname, cld,
 145                 atomic_read(&cld->cld_refcount));
 146         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 147
 148         /* spinlock to make sure no item with 0 refcount in the list */
 149         if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 150                 list_del(&cld->cld_list_chain);
 151                 spin_unlock(&config_list_lock);
 152
 153                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 154
 155                 config_log_put(cld->cld_barrier);
 156                 config_log_put(cld->cld_recover);
 157                 config_log_put(cld->cld_params);
 158                 config_log_put(cld->cld_nodemap);
 159                 config_log_put(cld->cld_sptlrpc);
 160                 if (cld_is_sptlrpc(cld)) {
 161                         cld->cld_stopping = 1;
 162                         sptlrpc_conf_log_stop(cld->cld_logname);
 163                 }
 164
 165                 class_export_put(cld->cld_mgcexp);
 166                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 167         }
 168
 169         EXIT;
 170 }
 171
 172 /* Find a config log by name */
 173 static
 174 struct config_llog_data *config_log_find(char *logname,
 175                                          struct config_llog_instance *cfg)
 176 {
 177         struct config_llog_data *cld;
 178         struct config_llog_data *found = NULL;
 179         unsigned long cfg_instance;
 180
 181         ENTRY;
 182         LASSERT(logname != NULL);
 183
 184         cfg_instance = cfg ? cfg->cfg_instance : 0;
 185         spin_lock(&config_list_lock);
 186         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 187                 /* check if cfg_instance is the one we want */
 188                 if (cfg_instance != cld->cld_cfg.cfg_instance)
 189                         continue;
 190
 191                 /* instance may be NULL, should check name */
 192                 if (strcmp(logname, cld->cld_logname) == 0) {
 193                         found = cld;
 194                         config_log_get(found);
 195                         break;
 196                 }
 197         }
 198         spin_unlock(&config_list_lock);
 199         RETURN(found);
 200 }
 201
 202 static
 203 struct config_llog_data *do_config_log_add(struct obd_device *obd,
 204                                            char *logname,
 205                                            enum mgs_cfg_type type,
 206                                            struct config_llog_instance *cfg,
 207                                            struct super_block *sb)
 208 {
 209         struct config_llog_data *cld;
 210         int rc;
 211
 212         ENTRY;
 213
 214         CDEBUG(D_MGC, "do adding config log %s-%016lx\n", logname,
 215                cfg ? cfg->cfg_instance : 0);
 216
 217         OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
 218         if (!cld)
 219                 RETURN(ERR_PTR(-ENOMEM));
 220
 221         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 222         if (rc) {
 223                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 224                 RETURN(ERR_PTR(rc));
 225         }
 226
 227         strcpy(cld->cld_logname, logname);
 228         if (cfg)
 229                 cld->cld_cfg = *cfg;
 230         else
 231                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 232         mutex_init(&cld->cld_lock);
 233         cld->cld_cfg.cfg_last_idx = 0;
 234         cld->cld_cfg.cfg_flags = 0;
 235         cld->cld_cfg.cfg_sb = sb;
 236         cld->cld_type = type;
 237         atomic_set(&cld->cld_refcount, 1);
 238
 239         /* Keep the mgc around until we are done */
 240         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 241
 242         if (cld_is_sptlrpc(cld))
 243                 sptlrpc_conf_log_start(logname);
 244
 245         spin_lock(&config_list_lock);
 246         list_add(&cld->cld_list_chain, &config_llog_list);
 247         spin_unlock(&config_list_lock);
 248
 249         if (cld_is_sptlrpc(cld) || cld_is_nodemap(cld) || cld_is_barrier(cld)) {
 250                 rc = mgc_process_log(obd, cld);
 251                 if (rc && rc != -ENOENT)
 252                         CERROR("%s: failed processing log, type %d: rc = %d\n",
 253                                obd->obd_name, type, rc);
 254         }
 255
 256         RETURN(cld);
 257 }
 258
 259 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
 260                                         char *fsname,
 261                                         struct config_llog_instance *cfg,
 262                                         struct super_block *sb)
 263 {
 264         struct config_llog_instance lcfg = *cfg;
 265         struct lustre_sb_info *lsi = s2lsi(sb);
 266         struct config_llog_data *cld;
 267         char logname[32];
 268
 269         if (IS_OST(lsi))
 270                 return NULL;
 271
 272         /* for osp-on-ost, see lustre_start_osp() */
 273         if (IS_MDT(lsi) && lcfg.cfg_instance)
 274                 return NULL;
 275
 276         /* We have to use different llog for clients and MDTs for DNE,
 277          * where only clients are notified if one of DNE server restarts.
 278          */
 279         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 280         strncpy(logname, fsname, sizeof(logname));
 281         if (IS_SERVER(lsi)) { /* mdt */
 282                 LASSERT(lcfg.cfg_instance == 0);
 283                 lcfg.cfg_instance = ll_get_cfg_instance(sb);
 284                 strncat(logname, "-mdtir", sizeof(logname));
 285         } else {
 286                 LASSERT(lcfg.cfg_instance != 0);
 287                 strncat(logname, "-cliir", sizeof(logname));
 288         }
 289
 290         cld = do_config_log_add(obd, logname, MGS_CFG_T_RECOVER, &lcfg, sb);
 291         return cld;
 292 }
 293
 294 static struct config_llog_data *
 295 config_log_find_or_add(struct obd_device *obd, char *logname,
 296                        struct super_block *sb, enum mgs_cfg_type type,
 297                        struct config_llog_instance *cfg)
 298 {
 299         struct config_llog_instance lcfg = *cfg;
 300         struct config_llog_data *cld;
 301
 302         /* Note class_config_llog_handler() depends on getting "obd" back */
 303         lcfg.cfg_instance = sb ? ll_get_cfg_instance(sb) : (unsigned long)obd;
 304
 305         cld = config_log_find(logname, &lcfg);
 306         if (unlikely(cld != NULL))
 307                 return cld;
 308
 309         return do_config_log_add(obd, logname, type, &lcfg, sb);
 310 }
 311
 312 /** Add this log to the list of active logs watched by an MGC.
 313  * Active means we're watching for updates.
 314  * We have one active log per "mount" - client instance or servername.
 315  * Each instance may be at a different point in the log.
 316  */
 317 static struct config_llog_data *
 318 config_log_add(struct obd_device *obd, char *logname,
 319                struct config_llog_instance *cfg, struct super_block *sb)
 320 {
 321         struct lustre_sb_info *lsi = s2lsi(sb);
 322         struct config_llog_data *cld = NULL;
 323         struct config_llog_data *sptlrpc_cld = NULL;
 324         struct config_llog_data *params_cld = NULL;
 325         struct config_llog_data *nodemap_cld = NULL;
 326         struct config_llog_data *barrier_cld = NULL;
 327         char seclogname[32];
 328         char *ptr;
 329         int rc;
 330         bool locked = false;
 331         ENTRY;
 332
 333         CDEBUG(D_MGC, "add config log %s-%016lx\n", logname,
 334                cfg->cfg_instance);
 335
 336         /*
 337          * for each regular log, the depended sptlrpc log name is
 338          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 339          */
 340         ptr = strrchr(logname, '-');
 341         if (ptr == NULL || ptr - logname > 8) {
 342                 CERROR("logname %s is too long\n", logname);
 343                 RETURN(ERR_PTR(-EINVAL));
 344         }
 345
 346         memcpy(seclogname, logname, ptr - logname);
 347         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 348
 349         if (cfg->cfg_sub_clds & CONFIG_SUB_SPTLRPC) {
 350                 sptlrpc_cld = config_log_find_or_add(obd, seclogname, NULL,
 351                                                      MGS_CFG_T_SPTLRPC, cfg);
 352                 if (IS_ERR(sptlrpc_cld)) {
 353                         CERROR("%s: can't create sptlrpc log %s: rc = %ld\n",
 354                                obd->obd_name, seclogname, PTR_ERR(sptlrpc_cld));
 355                         RETURN(sptlrpc_cld);
 356                 }
 357         }
 358
 359         if (!IS_MGS(lsi) && cfg->cfg_sub_clds & CONFIG_SUB_NODEMAP) {
 360                 nodemap_cld = config_log_find_or_add(obd, LUSTRE_NODEMAP_NAME,
 361                                                      NULL, MGS_CFG_T_NODEMAP,
 362                                                      cfg);
 363                 if (IS_ERR(nodemap_cld)) {
 364                         rc = PTR_ERR(nodemap_cld);
 365                         CERROR("%s: cannot create nodemap log: rc = %d\n",
 366                                obd->obd_name, rc);
 367                         GOTO(out_sptlrpc, rc);
 368                 }
 369         }
 370
 371         if (cfg->cfg_sub_clds & CONFIG_SUB_PARAMS) {
 372                 params_cld = config_log_find_or_add(obd, PARAMS_FILENAME, sb,
 373                                                     MGS_CFG_T_PARAMS, cfg);
 374                 if (IS_ERR(params_cld)) {
 375                         rc = PTR_ERR(params_cld);
 376                         CERROR("%s: can't create params log: rc = %d\n",
 377                                obd->obd_name, rc);
 378                         GOTO(out_nodemap, rc);
 379                 }
 380         }
 381
 382         if (IS_MDT(s2lsi(sb)) && cfg->cfg_sub_clds & CONFIG_SUB_BARRIER) {
 383                 snprintf(seclogname + (ptr - logname), sizeof(seclogname) - 1,
 384                          "-%s", BARRIER_FILENAME);
 385                 barrier_cld = config_log_find_or_add(obd, seclogname, sb,
 386                                                      MGS_CFG_T_BARRIER, cfg);
 387                 if (IS_ERR(barrier_cld)) {
 388                         rc = PTR_ERR(barrier_cld);
 389                         CERROR("%s: can't create barrier log: rc = %d\n",
 390                                obd->obd_name, rc);
 391                         GOTO(out_params, rc);
 392                 }
 393         }
 394
 395         cld = do_config_log_add(obd, logname, MGS_CFG_T_CONFIG, cfg, sb);
 396         if (IS_ERR(cld)) {
 397                 rc = PTR_ERR(cld);
 398                 CERROR("%s: can't create log: rc = %d\n",
 399                        obd->obd_name, rc);
 400                 GOTO(out_barrier, rc = PTR_ERR(cld));
 401         }
 402
 403         LASSERT(lsi->lsi_lmd);
 404         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR) &&
 405             cfg->cfg_sub_clds & CONFIG_SUB_RECOVER) {
 406                 struct config_llog_data *recover_cld;
 407
 408                 ptr = strrchr(seclogname, '-');
 409                 if (ptr != NULL) {
 410                         *ptr = 0;
 411                 } else {
 412                         CERROR("%s: sptlrpc log name not correct, %s: "
 413                                "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
 414                         GOTO(out_cld, rc = -EINVAL);
 415                 }
 416
 417                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 418                 if (IS_ERR(recover_cld)) {
 419                         rc = PTR_ERR(recover_cld);
 420                         CERROR("%s: can't create recover log: rc = %d\n",
 421                                obd->obd_name, rc);
 422                         GOTO(out_cld, rc);
 423                 }
 424
 425                 mutex_lock(&cld->cld_lock);
 426                 locked = true;
 427                 cld->cld_recover = recover_cld;
 428         }
 429
 430         if (!locked)
 431                 mutex_lock(&cld->cld_lock);
 432         cld->cld_params = params_cld;
 433         cld->cld_barrier = barrier_cld;
 434         cld->cld_nodemap = nodemap_cld;
 435         cld->cld_sptlrpc = sptlrpc_cld;
 436         mutex_unlock(&cld->cld_lock);
 437
 438         RETURN(cld);
 439
 440 out_cld:
 441         config_log_put(cld);
 442 out_barrier:
 443         config_log_put(barrier_cld);
 444 out_params:
 445         config_log_put(params_cld);
 446 out_nodemap:
 447         config_log_put(nodemap_cld);
 448 out_sptlrpc:
 449         config_log_put(sptlrpc_cld);
 450
 451         return ERR_PTR(rc);
 452 }
 453
 454 DEFINE_MUTEX(llog_process_lock);
 455
 456 static inline void config_mark_cld_stop_nolock(struct config_llog_data *cld)
 457 {
 458         ENTRY;
 459
 460         spin_lock(&config_list_lock);
 461         cld->cld_stopping = 1;
 462         spin_unlock(&config_list_lock);
 463
 464         CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
 465         if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
 466                 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
 467 }
 468
 469 static inline void config_mark_cld_stop(struct config_llog_data *cld)
 470 {
 471         if (cld) {
 472                 mutex_lock(&cld->cld_lock);
 473                 config_mark_cld_stop_nolock(cld);
 474                 mutex_unlock(&cld->cld_lock);
 475         }
 476 }
 477
 478 /** Stop watching for updates on this log.
 479  */
 480 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 481 {
 482         struct config_llog_data *cld;
 483         struct config_llog_data *cld_sptlrpc = NULL;
 484         struct config_llog_data *cld_params = NULL;
 485         struct config_llog_data *cld_recover = NULL;
 486         struct config_llog_data *cld_nodemap = NULL;
 487         struct config_llog_data *cld_barrier = NULL;
 488         int rc = 0;
 489
 490         ENTRY;
 491
 492         cld = config_log_find(logname, cfg);
 493         if (cld == NULL)
 494                 RETURN(-ENOENT);
 495
 496         mutex_lock(&cld->cld_lock);
 497         /*
 498          * if cld_stopping is set, it means we didn't start the log thus
 499          * not owning the start ref. this can happen after previous umount:
 500          * the cld still hanging there waiting for lock cancel, and we
 501          * remount again but failed in the middle and call log_end without
 502          * calling start_log.
 503          */
 504         if (unlikely(cld->cld_stopping)) {
 505                 mutex_unlock(&cld->cld_lock);
 506                 /* drop the ref from the find */
 507                 config_log_put(cld);
 508                 RETURN(rc);
 509         }
 510
 511         cld_recover = cld->cld_recover;
 512         cld->cld_recover = NULL;
 513         cld_params = cld->cld_params;
 514         cld->cld_params = NULL;
 515         cld_nodemap = cld->cld_nodemap;
 516         cld->cld_nodemap = NULL;
 517         cld_barrier = cld->cld_barrier;
 518         cld->cld_barrier = NULL;
 519         cld_sptlrpc = cld->cld_sptlrpc;
 520         cld->cld_sptlrpc = NULL;
 521
 522         config_mark_cld_stop_nolock(cld);
 523         mutex_unlock(&cld->cld_lock);
 524
 525         config_mark_cld_stop(cld_recover);
 526         config_log_put(cld_recover);
 527         config_mark_cld_stop(cld_params);
 528         config_log_put(cld_params);
 529         config_mark_cld_stop(cld_barrier);
 530         config_log_put(cld_barrier);
 531         /* don't explicitly set cld_stopping on sptlrpc lock here, as other
 532          * targets may be active, it will be done in config_log_put if necessary
 533          */
 534         config_log_put(cld_sptlrpc);
 535         /* don't set cld_stopping on nm lock as other targets may be active */
 536         config_log_put(cld_nodemap);
 537
 538         /* drop the ref from the find */
 539         config_log_put(cld);
 540         /* drop the start ref */
 541         config_log_put(cld);
 542
 543         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 544                rc);
 545         RETURN(rc);
 546 }
 547
 548 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 549 {
 550         struct obd_device       *obd = data;
 551         struct obd_import       *imp;
 552         struct obd_connect_data *ocd;
 553         struct config_llog_data *cld;
 554         int rc = 0;
 555
 556         ENTRY;
 557         LASSERT(obd);
 558         with_imp_locked(obd, imp, rc) {
 559                 ocd = &imp->imp_connect_data;
 560
 561                 seq_printf(m, "imperative_recovery: %s\n",
 562                            OCD_HAS_FLAG(ocd, IMP_RECOV) ?
 563                            "ENABLED" : "DISABLED");
 564         }
 565         if (rc)
 566                 RETURN(rc);
 567
 568         seq_printf(m, "client_state:\n");
 569
 570         spin_lock(&config_list_lock);
 571         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 572                 if (cld->cld_recover == NULL)
 573                         continue;
 574                 seq_printf(m,  "    - { client: %s, nidtbl_version: %u }\n",
 575                            cld->cld_logname,
 576                            cld->cld_recover->cld_cfg.cfg_last_idx);
 577         }
 578         spin_unlock(&config_list_lock);
 579
 580         RETURN(0);
 581 }
 582
 583 /* reenqueue any lost locks */
 584 #define RQ_RUNNING      0x1
 585 #define RQ_NOW          0x2
 586 #define RQ_LATER        0x4
 587 #define RQ_STOP         0x8
 588 #define RQ_PRECLEANUP   0x10
 589 static int                    rq_state = 0;
 590 static wait_queue_head_t      rq_waitq;
 591 static DECLARE_COMPLETION(rq_exit);
 592 static DECLARE_COMPLETION(rq_start);
 593
 594 static void do_requeue(struct config_llog_data *cld)
 595 {
 596         int rc = 0;
 597         ENTRY;
 598
 599         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 600
 601         /*
 602          * Do not run mgc_process_log on a disconnected export or an
 603          * export which is being disconnected. Take the client
 604          * semaphore to make the check non-racy.
 605          */
 606         down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
 607                          OBD_CLI_SEM_MGC);
 608         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 609                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 610                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 611                 if (rc && rc != -ENOENT)
 612                         CERROR("failed processing log: %d\n", rc);
 613         } else {
 614                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 615                        cld->cld_logname);
 616         }
 617         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 618
 619         EXIT;
 620 }
 621
 622 static int mgc_requeue_thread(void *data)
 623 {
 624         int rc = 0;
 625         bool first = true;
 626         ENTRY;
 627
 628         CDEBUG(D_MGC, "Starting requeue thread\n");
 629
 630         /* Keep trying failed locks periodically */
 631         spin_lock(&config_list_lock);
 632         rq_state |= RQ_RUNNING;
 633         while (!(rq_state & RQ_STOP)) {
 634                 struct config_llog_data *cld, *cld_prev;
 635                 int to;
 636
 637                 /* Any new or requeued lostlocks will change the state */
 638                 rq_state &= ~(RQ_NOW | RQ_LATER);
 639                 spin_unlock(&config_list_lock);
 640
 641                 if (first) {
 642                         first = false;
 643                         complete(&rq_start);
 644                 }
 645
 646                 /* Always wait a few seconds to allow the server who
 647                  * caused the lock revocation to finish its setup, plus some
 648                  * random so everyone doesn't try to reconnect at once.
 649                  */
 650                 to = mgc_requeue_timeout_min == 0 ? 1 : mgc_requeue_timeout_min;
 651                 to = cfs_time_seconds(mgc_requeue_timeout_min) +
 652                         prandom_u32_max(cfs_time_seconds(to));
 653                 wait_event_idle_timeout(rq_waitq,
 654                                         rq_state & (RQ_STOP | RQ_PRECLEANUP), to);
 655
 656                 /*
 657                  * iterate & processing through the list. for each cld, process
 658                  * its depending sptlrpc cld firstly (if any) and then itself.
 659                  *
 660                  * it's guaranteed any item in the list must have
 661                  * reference > 0; and if cld_lostlock is set, at
 662                  * least one reference is taken by the previous enqueue.
 663                  */
 664                 cld_prev = NULL;
 665
 666                 spin_lock(&config_list_lock);
 667                 rq_state &= ~RQ_PRECLEANUP;
 668                 list_for_each_entry(cld, &config_llog_list,
 669                                     cld_list_chain) {
 670                         if (!cld->cld_lostlock || cld->cld_stopping)
 671                                 continue;
 672
 673                         /* hold reference to avoid being freed during
 674                          * subsequent processing. */
 675                         config_log_get(cld);
 676                         cld->cld_lostlock = 0;
 677                         spin_unlock(&config_list_lock);
 678
 679                         config_log_put(cld_prev);
 680                         cld_prev = cld;
 681
 682                         if (likely(!(rq_state & RQ_STOP))) {
 683                                 do_requeue(cld);
 684                                 spin_lock(&config_list_lock);
 685                         } else {
 686                                 spin_lock(&config_list_lock);
 687                                 break;
 688                         }
 689                 }
 690                 spin_unlock(&config_list_lock);
 691                 config_log_put(cld_prev);
 692
 693                 /* Wait a bit to see if anyone else needs a requeue */
 694                 wait_event_idle(rq_waitq, rq_state & (RQ_NOW | RQ_STOP));
 695                 spin_lock(&config_list_lock);
 696         }
 697
 698         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 699         rq_state &= ~RQ_RUNNING;
 700         spin_unlock(&config_list_lock);
 701
 702         complete(&rq_exit);
 703
 704         CDEBUG(D_MGC, "Ending requeue thread\n");
 705         RETURN(rc);
 706 }
 707
 708 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 709    We are responsible for dropping the config log reference from here on out. */
 710 static void mgc_requeue_add(struct config_llog_data *cld)
 711 {
 712         bool wakeup = false;
 713         ENTRY;
 714
 715         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 716                 cld->cld_logname, atomic_read(&cld->cld_refcount),
 717                 cld->cld_stopping, rq_state);
 718         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 719
 720         /* lets cancel an existent lock to mark cld as "lostlock" */
 721         CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
 722         if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
 723                 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
 724
 725         mutex_lock(&cld->cld_lock);
 726         spin_lock(&config_list_lock);
 727         if (!(rq_state & RQ_STOP) && !cld->cld_stopping) {
 728                 cld->cld_lostlock = 1;
 729                 rq_state |= RQ_NOW;
 730                 wakeup = true;
 731         }
 732         spin_unlock(&config_list_lock);
 733         mutex_unlock(&cld->cld_lock);
 734         if (wakeup)
 735                 wake_up(&rq_waitq);
 736
 737         EXIT;
 738 }
 739
 740 /********************** class fns **********************/
 741 static int mgc_local_llog_init(const struct lu_env *env,
 742                                struct obd_device *obd,
 743                                struct obd_device *disk)
 744 {
 745         struct llog_ctxt        *ctxt;
 746         int                      rc;
 747
 748         ENTRY;
 749
 750         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
 751                         &llog_osd_ops);
 752         if (rc)
 753                 RETURN(rc);
 754
 755         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 756         LASSERT(ctxt);
 757         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
 758         llog_ctxt_put(ctxt);
 759
 760         RETURN(0);
 761 }
 762
 763 static int mgc_local_llog_fini(const struct lu_env *env,
 764                                struct obd_device *obd)
 765 {
 766         struct llog_ctxt *ctxt;
 767
 768         ENTRY;
 769
 770         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 771         llog_cleanup(env, ctxt);
 772
 773         RETURN(0);
 774 }
 775
 776 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
 777                         struct super_block *sb)
 778 {
 779         struct lustre_sb_info   *lsi = s2lsi(sb);
 780         struct client_obd       *cli = &obd->u.cli;
 781         struct lu_fid            rfid, fid;
 782         struct dt_object        *root, *dto;
 783         int                      rc = 0;
 784
 785         ENTRY;
 786
 787         LASSERT(lsi);
 788         LASSERT(lsi->lsi_dt_dev);
 789
 790         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
 791         mutex_lock(&cli->cl_mgc_mutex);
 792
 793         /* Setup the configs dir */
 794         fid.f_seq = FID_SEQ_LOCAL_NAME;
 795         fid.f_oid = 1;
 796         fid.f_ver = 0;
 797         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
 798                                     &cli->cl_mgc_los);
 799         if (rc)
 800                 RETURN(rc);
 801
 802         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
 803         if (rc)
 804                 GOTO(out_los, rc);
 805
 806         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
 807                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
 808         if (unlikely(IS_ERR(root)))
 809                 GOTO(out_los, rc = PTR_ERR(root));
 810
 811         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
 812                                         MOUNT_CONFIGS_DIR,
 813                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
 814         dt_object_put_nocache(env, root);
 815         if (IS_ERR(dto))
 816                 GOTO(out_los, rc = PTR_ERR(dto));
 817
 818         cli->cl_mgc_configs_dir = dto;
 819
 820         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
 821         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
 822         if (rc)
 823                 GOTO(out_llog, rc);
 824
 825         /* We take an obd ref to insure that we can't get to mgc_cleanup
 826          * without calling mgc_fs_cleanup first. */
 827         class_incref(obd, "mgc_fs", obd);
 828
 829         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 830         EXIT;
 831 out_llog:
 832         if (rc) {
 833                 dt_object_put(env, cli->cl_mgc_configs_dir);
 834                 cli->cl_mgc_configs_dir = NULL;
 835         }
 836 out_los:
 837         if (rc < 0) {
 838                 local_oid_storage_fini(env, cli->cl_mgc_los);
 839                 cli->cl_mgc_los = NULL;
 840                 mutex_unlock(&cli->cl_mgc_mutex);
 841         }
 842         return rc;
 843 }
 844
 845 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
 846 {
 847         struct client_obd       *cli = &obd->u.cli;
 848         ENTRY;
 849
 850         LASSERT(cli->cl_mgc_los != NULL);
 851
 852         mgc_local_llog_fini(env, obd);
 853
 854         dt_object_put_nocache(env, cli->cl_mgc_configs_dir);
 855         cli->cl_mgc_configs_dir = NULL;
 856
 857         local_oid_storage_fini(env, cli->cl_mgc_los);
 858         cli->cl_mgc_los = NULL;
 859
 860         class_decref(obd, "mgc_fs", obd);
 861         mutex_unlock(&cli->cl_mgc_mutex);
 862
 863         RETURN(0);
 864 }
 865
 866 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 867 {
 868         struct llog_ctxt        *ctxt;
 869         int                      rc;
 870
 871         ENTRY;
 872
 873         /* setup only remote ctxt, the local disk context is switched per each
 874          * filesystem during mgc_fs_setup() */
 875         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 876                         &llog_client_ops);
 877         if (rc)
 878                 RETURN(rc);
 879
 880         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 881         LASSERT(ctxt);
 882
 883         llog_initiator_connect(ctxt);
 884         llog_ctxt_put(ctxt);
 885
 886         RETURN(0);
 887 }
 888
 889 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 890 {
 891         struct llog_ctxt *ctxt;
 892
 893         ENTRY;
 894
 895         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 896         if (ctxt)
 897                 llog_cleanup(env, ctxt);
 898
 899         RETURN(0);
 900 }
 901
 902
 903 static atomic_t mgc_count = ATOMIC_INIT(0);
 904 static int mgc_precleanup(struct obd_device *obd)
 905 {
 906         int     rc = 0;
 907         int     temp;
 908         ENTRY;
 909
 910         if (atomic_dec_and_test(&mgc_count)) {
 911                 LASSERT(rq_state & RQ_RUNNING);
 912                 /* stop requeue thread */
 913                 temp = RQ_STOP;
 914         } else {
 915                 /* wakeup requeue thread to clean our cld */
 916                 temp = RQ_NOW | RQ_PRECLEANUP;
 917         }
 918
 919         spin_lock(&config_list_lock);
 920         rq_state |= temp;
 921         spin_unlock(&config_list_lock);
 922         wake_up(&rq_waitq);
 923
 924         if (temp & RQ_STOP)
 925                 wait_for_completion(&rq_exit);
 926         obd_cleanup_client_import(obd);
 927
 928         rc = mgc_llog_fini(NULL, obd);
 929         if (rc != 0)
 930                 CERROR("failed to cleanup llogging subsystems\n");
 931
 932         RETURN(rc);
 933 }
 934
 935 static int mgc_cleanup(struct obd_device *obd)
 936 {
 937         int rc;
 938         ENTRY;
 939
 940         /* COMPAT_146 - old config logs may have added profiles we don't
 941            know about */
 942         if (atomic_read(&obd->obd_type->typ_refcnt) <= 1)
 943                 /* Only for the last mgc */
 944                 class_del_profiles();
 945
 946         lprocfs_obd_cleanup(obd);
 947         ptlrpcd_decref();
 948
 949         rc = client_obd_cleanup(obd);
 950         RETURN(rc);
 951 }
 952
 953 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 954 {
 955         struct task_struct      *task;
 956         int                      rc;
 957         ENTRY;
 958
 959         rc = ptlrpcd_addref();
 960         if (rc < 0)
 961                 RETURN(rc);
 962
 963         rc = client_obd_setup(obd, lcfg);
 964         if (rc)
 965                 GOTO(err_decref, rc);
 966
 967         rc = mgc_llog_init(NULL, obd);
 968         if (rc) {
 969                 CERROR("failed to setup llogging subsystems\n");
 970                 GOTO(err_cleanup, rc);
 971         }
 972
 973         rc = mgc_tunables_init(obd);
 974         if (rc)
 975                 GOTO(err_sysfs, rc);
 976
 977         if (atomic_inc_return(&mgc_count) == 1) {
 978                 rq_state = 0;
 979                 init_waitqueue_head(&rq_waitq);
 980
 981                 /* start requeue thread */
 982                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 983                 if (IS_ERR(task)) {
 984                         rc = PTR_ERR(task);
 985                         CERROR("%s: cannot start requeue thread: rc = %d; "
 986                                "no more log updates\n",
 987                                obd->obd_name, rc);
 988                         GOTO(err_sysfs, rc);
 989                 }
 990                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 991                 rc = 0;
 992                 wait_for_completion(&rq_start);
 993         }
 994
 995         RETURN(rc);
 996
 997 err_sysfs:
 998         lprocfs_obd_cleanup(obd);
 999 err_cleanup:
1000         client_obd_cleanup(obd);
1001 err_decref:
1002         ptlrpcd_decref();
1003         RETURN(rc);
1004 }
1005
1006 /* based on ll_mdc_blocking_ast */
1007 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1008                             void *data, int flag)
1009 {
1010         struct lustre_handle lockh;
1011         struct config_llog_data *cld = (struct config_llog_data *)data;
1012         int rc = 0;
1013         ENTRY;
1014
1015         switch (flag) {
1016         case LDLM_CB_BLOCKING:
1017                 /* mgs wants the lock, give it up... */
1018                 LDLM_DEBUG(lock, "MGC blocking CB");
1019                 ldlm_lock2handle(lock, &lockh);
1020                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
1021                 break;
1022         case LDLM_CB_CANCELING:
1023                 /* We've given up the lock, prepare ourselves to update. */
1024                 LDLM_DEBUG(lock, "MGC cancel CB");
1025
1026                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
1027                        PLDLMRES(lock->l_resource),
1028                        (char *)&lock->l_resource->lr_name.name[0]);
1029
1030                 if (!cld) {
1031                         CDEBUG(D_INFO, "missing data, won't requeue\n");
1032                         break;
1033                 }
1034
1035                 /* held at mgc_process_log(). */
1036                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1037
1038                 lock->l_ast_data = NULL;
1039                 cld->cld_lockh.cookie = 0;
1040                 /* Are we done with this log? */
1041                 if (cld->cld_stopping) {
1042                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1043                                 cld->cld_logname);
1044                         config_log_put(cld);
1045                         break;
1046                 }
1047                 /* Make sure not to re-enqueue when the mgc is stopping
1048                    (we get called from client_disconnect_export) */
1049                 if (lock->l_conn_export == NULL ||
1050                     lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1051                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1052                                 cld->cld_logname);
1053                         config_log_put(cld);
1054                         break;
1055                 }
1056
1057                 /* Re-enqueue now */
1058                 mgc_requeue_add(cld);
1059                 config_log_put(cld);
1060                 break;
1061         default:
1062                 LBUG();
1063         }
1064
1065         RETURN(rc);
1066 }
1067
1068 /* Not sure where this should go... */
1069 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1070  * that we can have a chance to try the secondary MGS if any. */
1071 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1072                                 + PING_INTERVAL)
1073 #define  MGC_TARGET_REG_LIMIT 10
1074 #define  MGC_TARGET_REG_LIMIT_MAX RECONNECT_DELAY_MAX
1075 #define  MGC_SEND_PARAM_LIMIT 10
1076
1077 /* Take a config lock so we can get cancel notifications */
1078 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1079                        union ldlm_policy_data *policy, enum ldlm_mode mode,
1080                        __u64 *flags, ldlm_glimpse_callback glimpse_callback,
1081                        void *data, __u32 lvb_len, void *lvb_swabber,
1082                        struct lustre_handle *lockh)
1083 {
1084         struct config_llog_data *cld = (struct config_llog_data *)data;
1085         struct ldlm_enqueue_info einfo = {
1086                 .ei_type        = type,
1087                 .ei_mode        = mode,
1088                 .ei_cb_bl       = mgc_blocking_ast,
1089                 .ei_cb_cp       = ldlm_completion_ast,
1090                 .ei_cb_gl       = glimpse_callback,
1091         };
1092         struct ptlrpc_request *req;
1093         int short_limit = cld_is_sptlrpc(cld);
1094         int rc;
1095         ENTRY;
1096
1097         if (!exp)
1098                 RETURN(-EBADR);
1099
1100         CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1101                cld->cld_resid.name[0]);
1102
1103         /* We need a callback for every lockholder, so don't try to
1104            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1105         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1106                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1107                                         LDLM_ENQUEUE);
1108         if (req == NULL)
1109                 RETURN(-ENOMEM);
1110
1111         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1112         ptlrpc_request_set_replen(req);
1113
1114         /* check if this is server or client */
1115         if (cld->cld_cfg.cfg_sb) {
1116                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1117                 if (lsi && IS_SERVER(lsi))
1118                         short_limit = 1;
1119         }
1120         /* Limit how long we will wait for the enqueue to complete */
1121         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1122         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1123                               NULL, 0, LVB_T_NONE, lockh, 0);
1124         /* A failed enqueue should still call the mgc_blocking_ast,
1125            where it will be requeued if needed ("grant failed"). */
1126         ptlrpc_req_finished(req);
1127         RETURN(rc);
1128 }
1129
1130 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1131                       struct lustre_handle *lockh)
1132 {
1133         ENTRY;
1134
1135         ldlm_lock_decref(lockh, mode);
1136
1137         RETURN(0);
1138 }
1139
1140 static void mgc_notify_active(struct obd_device *unused)
1141 {
1142         /* wakeup mgc_requeue_thread to requeue mgc lock */
1143         spin_lock(&config_list_lock);
1144         rq_state |= RQ_NOW;
1145         spin_unlock(&config_list_lock);
1146         wake_up(&rq_waitq);
1147
1148         /* TODO: Help the MGS rebuild nidtbl. -jay */
1149 }
1150
1151 /* Send target_reg message to MGS */
1152 static int mgc_target_register(struct obd_export *exp,
1153                                struct mgs_target_info *mti)
1154 {
1155         struct ptlrpc_request  *req;
1156         struct mgs_target_info *req_mti, *rep_mti;
1157         int                     rc;
1158         ENTRY;
1159
1160         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1161                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1162                                         MGS_TARGET_REG);
1163         if (req == NULL)
1164                 RETURN(-ENOMEM);
1165
1166         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1167         if (!req_mti) {
1168                 ptlrpc_req_finished(req);
1169                 RETURN(-ENOMEM);
1170         }
1171
1172         memcpy(req_mti, mti, sizeof(*req_mti));
1173         ptlrpc_request_set_replen(req);
1174         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1175         /* Limit how long we will wait for the enqueue to complete */
1176         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1177
1178         /* if the target needs to regenerate the config log in MGS, it's better
1179          * to use some longer limit to let MGC have time to change connection to
1180          * another MGS (or try again with the same MGS) for the target (server)
1181          * will fail and exit if the request expired due to delay limit. */
1182         if (mti->mti_flags & (LDD_F_UPDATE | LDD_F_NEED_INDEX))
1183                 req->rq_delay_limit = MGC_TARGET_REG_LIMIT_MAX;
1184
1185         rc = ptlrpc_queue_wait(req);
1186         if (!rc) {
1187                 rep_mti = req_capsule_server_get(&req->rq_pill,
1188                                                  &RMF_MGS_TARGET_INFO);
1189                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1190                 CDEBUG(D_MGC, "register %s got index = %d\n",
1191                        mti->mti_svname, mti->mti_stripe_index);
1192         }
1193         ptlrpc_req_finished(req);
1194
1195         RETURN(rc);
1196 }
1197
1198 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1199                               u32 keylen, void *key,
1200                               u32 vallen, void *val,
1201                               struct ptlrpc_request_set *set)
1202 {
1203         int rc = -EINVAL;
1204         ENTRY;
1205
1206         /* Turn off initial_recov after we try all backup servers once */
1207         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1208                 struct obd_import *imp = class_exp2cliimp(exp);
1209                 int value;
1210                 if (vallen != sizeof(int))
1211                         RETURN(-EINVAL);
1212                 value = *(int *)val;
1213                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1214                        imp->imp_obd->obd_name, value,
1215                        imp->imp_deactive, imp->imp_invalid,
1216                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1217                        ptlrpc_import_state_name(imp->imp_state));
1218                 /* Resurrect the import immediately if
1219                  * 1. we previously got disconnected,
1220                  * 2. value > 1 (at the same node with MGS)
1221                  * */
1222                 if (imp->imp_state == LUSTRE_IMP_DISCON || value > 1)
1223                         ptlrpc_reconnect_import(imp);
1224
1225                 RETURN(0);
1226         }
1227
1228         /* FIXME move this to mgc_process_config */
1229         if (KEY_IS(KEY_REGISTER_TARGET)) {
1230                 struct mgs_target_info *mti;
1231                 if (vallen != sizeof(struct mgs_target_info))
1232                         RETURN(-EINVAL);
1233                 mti = (struct mgs_target_info *)val;
1234                 CDEBUG(D_MGC, "register_target %s %#x\n",
1235                        mti->mti_svname, mti->mti_flags);
1236                 rc =  mgc_target_register(exp, mti);
1237                 RETURN(rc);
1238         }
1239         if (KEY_IS(KEY_SET_FS)) {
1240                 struct super_block *sb = (struct super_block *)val;
1241
1242                 if (vallen != sizeof(struct super_block))
1243                         RETURN(-EINVAL);
1244
1245                 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1246                 RETURN(rc);
1247         }
1248         if (KEY_IS(KEY_CLEAR_FS)) {
1249                 if (vallen != 0)
1250                         RETURN(-EINVAL);
1251                 rc = mgc_fs_cleanup(env, exp->exp_obd);
1252                 RETURN(rc);
1253         }
1254         if (KEY_IS(KEY_MGSSEC)) {
1255                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1256                 struct sptlrpc_flavor  flvr;
1257
1258                 /*
1259                  * empty string means using current flavor, if which haven't
1260                  * been set yet, set it as null.
1261                  *
1262                  * if flavor has been set previously, check the asking flavor
1263                  * must match the existing one.
1264                  */
1265                 if (vallen == 0) {
1266                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1267                                 RETURN(0);
1268                         val = "null";
1269                         vallen = 4;
1270                 }
1271
1272                 rc = sptlrpc_parse_flavor(val, &flvr);
1273                 if (rc) {
1274                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1275                                (char *) val);
1276                         RETURN(rc);
1277                 }
1278
1279                 /*
1280                  * caller already hold a mutex
1281                  */
1282                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1283                         cli->cl_flvr_mgc = flvr;
1284                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1285                                   sizeof(flvr)) != 0) {
1286                         char    str[20];
1287
1288                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1289                                             str, sizeof(str));
1290                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1291                                        "currently %s is in use\n",
1292                                        (char *) val, str);
1293                         rc = -EPERM;
1294                 }
1295                 RETURN(rc);
1296         }
1297
1298         RETURN(rc);
1299 }
1300
1301 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1302                         __u32 keylen, void *key, __u32 *vallen, void *val)
1303 {
1304         int rc = -EINVAL;
1305
1306         if (KEY_IS(KEY_CONN_DATA)) {
1307                 struct obd_import *imp = class_exp2cliimp(exp);
1308                 struct obd_connect_data *data = val;
1309
1310                 if (*vallen == sizeof(*data)) {
1311                         *data = imp->imp_connect_data;
1312                         rc = 0;
1313                 }
1314         }
1315
1316         return rc;
1317 }
1318
1319 static int mgc_import_event(struct obd_device *obd,
1320                             struct obd_import *imp,
1321                             enum obd_import_event event)
1322 {
1323         int rc = 0;
1324
1325         LASSERT(imp->imp_obd == obd);
1326         CDEBUG(D_MGC, "import event %#x\n", event);
1327
1328         switch (event) {
1329         case IMP_EVENT_DISCON:
1330                 /* MGC imports should not wait for recovery */
1331                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1332                         ptlrpc_pinger_ir_down();
1333                 break;
1334         case IMP_EVENT_INACTIVE:
1335                 break;
1336         case IMP_EVENT_INVALIDATE: {
1337                 struct ldlm_namespace *ns = obd->obd_namespace;
1338                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1339                 break;
1340         }
1341         case IMP_EVENT_ACTIVE:
1342                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1343                 /* Clearing obd_no_recov allows us to continue pinging */
1344                 obd->obd_no_recov = 0;
1345                 mgc_notify_active(obd);
1346                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1347                         ptlrpc_pinger_ir_up();
1348                 break;
1349         case IMP_EVENT_OCD:
1350                 break;
1351         case IMP_EVENT_DEACTIVATE:
1352         case IMP_EVENT_ACTIVATE:
1353                 break;
1354         default:
1355                 CERROR("Unknown import event %#x\n", event);
1356                 LBUG();
1357         }
1358         RETURN(rc);
1359 }
1360
1361 enum {
1362         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1363         CONFIG_READ_NRPAGES      = 4
1364 };
1365
1366 static int mgc_apply_recover_logs(struct obd_device *mgc,
1367                                   struct config_llog_data *cld,
1368                                   __u64 max_version,
1369                                   void *data, int datalen, bool mne_swab)
1370 {
1371         struct config_llog_instance *cfg = &cld->cld_cfg;
1372         struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1373         struct mgs_nidtbl_entry *entry;
1374         struct lustre_cfg *lcfg;
1375         struct lustre_cfg_bufs bufs;
1376         u64 prev_version = 0;
1377         char *inst;
1378         char *buf;
1379         int bufsz;
1380         int pos = 0;
1381         int rc  = 0;
1382         int off = 0;
1383         unsigned long dynamic_nids;
1384
1385         ENTRY;
1386         LASSERT(cfg->cfg_instance != 0);
1387         LASSERT(ll_get_cfg_instance(cfg->cfg_sb) == cfg->cfg_instance);
1388
1389         /* get dynamic nids setting */
1390         dynamic_nids = mgc->obd_dynamic_nids;
1391
1392         OBD_ALLOC(inst, PAGE_SIZE);
1393         if (inst == NULL)
1394                 RETURN(-ENOMEM);
1395
1396         if (!IS_SERVER(lsi)) {
1397                 pos = snprintf(inst, PAGE_SIZE, "%016lx", cfg->cfg_instance);
1398                 if (pos >= PAGE_SIZE) {
1399                         OBD_FREE(inst, PAGE_SIZE);
1400                         return -E2BIG;
1401                 }
1402 #ifdef HAVE_SERVER_SUPPORT
1403         } else {
1404                 LASSERT(IS_MDT(lsi));
1405                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1406                                         PAGE_SIZE);
1407                 if (rc) {
1408                         OBD_FREE(inst, PAGE_SIZE);
1409                         RETURN(-EINVAL);
1410                 }
1411                 pos = strlen(inst);
1412 #endif /* HAVE_SERVER_SUPPORT */
1413         }
1414
1415         ++pos;
1416         buf   = inst + pos;
1417         bufsz = PAGE_SIZE - pos;
1418
1419         while (datalen > 0) {
1420                 int   entry_len = sizeof(*entry);
1421                 int   is_ost;
1422                 struct obd_device *obd;
1423                 struct obd_import *imp;
1424                 char *obdname;
1425                 char *cname;
1426                 char *params;
1427                 char *uuid;
1428
1429                 rc = -EINVAL;
1430                 if (datalen < sizeof(*entry))
1431                         break;
1432
1433                 entry = (typeof(entry))(data + off);
1434
1435                 /* sanity check */
1436                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1437                         break;
1438                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1439                         break;
1440                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1441                         break;
1442
1443                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1444                 if (datalen < entry_len) /* must have entry_len at least */
1445                         break;
1446
1447                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1448                 if (mne_swab)
1449                         lustre_swab_mgs_nidtbl_entry(entry);
1450                 if (entry->mne_length > PAGE_SIZE) {
1451                         CERROR("MNE too large (%u)\n", entry->mne_length);
1452                         break;
1453                 }
1454
1455                 if (entry->mne_length < entry_len)
1456                         break;
1457
1458                 off     += entry->mne_length;
1459                 datalen -= entry->mne_length;
1460                 if (datalen < 0)
1461                         break;
1462
1463                 if (entry->mne_version > max_version) {
1464                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1465                                entry->mne_version, max_version);
1466                         break;
1467                 }
1468
1469                 if (prev_version >= entry->mne_version) {
1470                         CERROR("index unsorted, prev %lld, now %lld\n",
1471                                prev_version, entry->mne_version);
1472                         break;
1473                 }
1474                 prev_version = entry->mne_version;
1475
1476                 /*
1477                  * Write a string with format "nid::instance" to
1478                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1479                  */
1480
1481                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1482                 memset(buf, 0, bufsz);
1483                 obdname = buf;
1484                 pos = 0;
1485
1486                 /* lustre-OST0001-osc-<instance #> */
1487                 strcpy(obdname, cld->cld_logname);
1488                 cname = strrchr(obdname, '-');
1489                 if (cname == NULL) {
1490                         CERROR("mgc %s: invalid logname %s\n",
1491                                mgc->obd_name, obdname);
1492                         break;
1493                 }
1494
1495                 pos = cname - obdname;
1496                 obdname[pos] = 0;
1497                 pos += sprintf(obdname + pos, "-%s%04x",
1498                                is_ost ? "OST" : "MDT", entry->mne_index);
1499
1500                 cname = is_ost ? "osc" : "mdc",
1501                         pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1502                 lustre_cfg_bufs_reset(&bufs, obdname);
1503
1504                 /* find the obd by obdname */
1505                 obd = class_name2obd(obdname);
1506                 if (obd == NULL) {
1507                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1508                                mgc->obd_name, obdname);
1509                         rc = 0;
1510                         /* this is a safe race, when the ost is starting up...*/
1511                         continue;
1512                 }
1513
1514                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1515                 ++pos;
1516                 params = buf + pos;
1517                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1518                 uuid = buf + pos;
1519
1520                 with_imp_locked(obd, imp, rc) {
1521                         /* iterate all nids to find one */
1522                         /* find uuid by nid */
1523                         /* create import entries if they don't exist */
1524                         rc = client_import_add_nids_to_conn(
1525                                 imp, entry->u.nids, entry->mne_nid_count,
1526                                 (struct obd_uuid *)uuid);
1527
1528                         if (rc == -ENOENT && dynamic_nids) {
1529                                 /* create a new connection for this import */
1530                                 char *primary_nid =
1531                                         libcfs_nid2str(entry->u.nids[0]);
1532                                 int prim_nid_len = strlen(primary_nid) + 1;
1533                                 struct obd_uuid server_uuid;
1534
1535                                 if (prim_nid_len > UUID_MAX)
1536                                         goto fail;
1537                                 strncpy(server_uuid.uuid, primary_nid,
1538                                         prim_nid_len);
1539
1540                                 CDEBUG(D_INFO, "Adding a connection for %s\n",
1541                                        primary_nid);
1542
1543                                 rc = client_import_dyn_add_conn(
1544                                         imp, &server_uuid, entry->u.nids[0], 1);
1545                                 if (rc < 0) {
1546                                         CERROR("%s: Failed to add new connection with NID '%s' to import: rc = %d\n",
1547                                                obd->obd_name, primary_nid, rc);
1548                                         goto fail;
1549                                 }
1550                                 rc = client_import_add_nids_to_conn(
1551                                         imp, entry->u.nids,
1552                                         entry->mne_nid_count,
1553                                         (struct obd_uuid *)uuid);
1554                                 if (rc < 0) {
1555                                         CERROR("%s: failed to lookup UUID: rc = %d\n",
1556                                                obd->obd_name, rc);
1557                                         goto fail;
1558                                 }
1559                         }
1560 fail:;
1561                 }
1562                 if (rc == -ENODEV) {
1563                         /* client does not connect to the OST yet */
1564                         rc = 0;
1565                         continue;
1566                 }
1567
1568                 if (rc < 0 && rc != -ENOSPC) {
1569                         CERROR("mgc: cannot find UUID by nid '%s': rc = %d\n",
1570                                libcfs_nid2str(entry->u.nids[0]), rc);
1571                         break;
1572                 }
1573
1574                 CDEBUG(D_INFO, "Found UUID '%s' by NID '%s'\n",
1575                        uuid, libcfs_nid2str(entry->u.nids[0]));
1576
1577                 pos += strlen(uuid);
1578                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1579                 LASSERT(pos < bufsz);
1580
1581                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1582
1583                 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1584                                                bufs.lcfg_buflen));
1585                 if (!lcfg) {
1586                         rc = -ENOMEM;
1587                         break;
1588                 }
1589                 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1590
1591                 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1592                        prev_version, max_version, obdname, params);
1593
1594                 rc = class_process_config(lcfg);
1595                 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1596                                               lcfg->lcfg_buflens));
1597                 if (rc)
1598                         CDEBUG(D_INFO, "process config for %s error %d\n",
1599                                obdname, rc);
1600
1601                 /* continue, even one with error */
1602         }
1603
1604         OBD_FREE(inst, PAGE_SIZE);
1605         RETURN(rc);
1606 }
1607
1608 /**
1609  * This function is called if this client was notified for target restarting
1610  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1611  * nodemap logs.
1612  */
1613 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1614                                            struct config_llog_data *cld)
1615 {
1616         struct ptlrpc_connection *mgc_conn;
1617         struct ptlrpc_request *req = NULL;
1618         struct config_llog_instance *cfg = &cld->cld_cfg;
1619         struct mgs_config_body *body;
1620         struct mgs_config_res *res;
1621         struct nodemap_config *new_config = NULL;
1622         struct lu_nodemap *recent_nodemap = NULL;
1623         struct ptlrpc_bulk_desc *desc;
1624         struct page **pages = NULL;
1625         __u64 config_read_offset = 0;
1626         __u8 nodemap_cur_pass = 0;
1627         int nrpages = 0;
1628         bool eof = true;
1629         bool mne_swab = false;
1630         int i;
1631         int ealen;
1632         int rc;
1633         ENTRY;
1634
1635         mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1636
1637         /* don't need to get local config */
1638         if (cld_is_nodemap(cld) && LNetIsPeerLocal(mgc_conn->c_peer.nid))
1639                 GOTO(out, rc = 0);
1640
1641         /* allocate buffer for bulk transfer.
1642          * if this is the first time for this mgs to read logs,
1643          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1644          * once; otherwise, it only reads increment of logs, this should be
1645          * small and CONFIG_READ_NRPAGES will be used.
1646          */
1647         nrpages = CONFIG_READ_NRPAGES;
1648         if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1649                 nrpages = CONFIG_READ_NRPAGES_INIT;
1650
1651         OBD_ALLOC_PTR_ARRAY_LARGE(pages, nrpages);
1652         if (pages == NULL)
1653                 GOTO(out, rc = -ENOMEM);
1654
1655         for (i = 0; i < nrpages; i++) {
1656                 pages[i] = alloc_page(GFP_KERNEL);
1657                 if (pages[i] == NULL)
1658                         GOTO(out, rc = -ENOMEM);
1659         }
1660
1661 again:
1662 #ifdef HAVE_SERVER_SUPPORT
1663         if (cld_is_nodemap(cld) && config_read_offset == 0) {
1664                 new_config = nodemap_config_alloc();
1665                 if (IS_ERR(new_config)) {
1666                         rc = PTR_ERR(new_config);
1667                         new_config = NULL;
1668                         GOTO(out, rc);
1669                 }
1670         }
1671 #endif
1672         LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1673         LASSERT(mutex_is_locked(&cld->cld_lock));
1674         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1675                                    &RQF_MGS_CONFIG_READ);
1676         if (req == NULL)
1677                 GOTO(out, rc = -ENOMEM);
1678
1679         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1680         if (rc)
1681                 GOTO(out, rc);
1682
1683         /* pack request */
1684         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1685         LASSERT(body != NULL);
1686         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1687         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1688             >= sizeof(body->mcb_name))
1689                 GOTO(out, rc = -E2BIG);
1690         if (cld_is_nodemap(cld))
1691                 body->mcb_offset = config_read_offset;
1692         else
1693                 body->mcb_offset = cfg->cfg_last_idx + 1;
1694         body->mcb_type   = cld->cld_type;
1695         body->mcb_bits   = PAGE_SHIFT;
1696         body->mcb_units  = nrpages;
1697         body->mcb_nm_cur_pass = nodemap_cur_pass;
1698
1699         /* allocate bulk transfer descriptor */
1700         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1701                                     PTLRPC_BULK_PUT_SINK,
1702                                     MGS_BULK_PORTAL,
1703                                     &ptlrpc_bulk_kiov_pin_ops);
1704         if (desc == NULL)
1705                 GOTO(out, rc = -ENOMEM);
1706
1707         for (i = 0; i < nrpages; i++)
1708                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1709                                                  PAGE_SIZE);
1710
1711         ptlrpc_request_set_replen(req);
1712         rc = ptlrpc_queue_wait(req);
1713         if (rc)
1714                 GOTO(out, rc);
1715
1716         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1717         if (!res)
1718                 GOTO(out, rc = -EPROTO);
1719
1720         if (cld_is_nodemap(cld)) {
1721                 config_read_offset = res->mcr_offset;
1722                 eof = config_read_offset == II_END_OFF;
1723                 nodemap_cur_pass = res->mcr_nm_cur_pass;
1724         } else {
1725                 if (res->mcr_size < res->mcr_offset)
1726                         GOTO(out, rc = -EINVAL);
1727
1728                 /* always update the index even though it might have errors with
1729                  * handling the recover logs
1730                  */
1731                 cfg->cfg_last_idx = res->mcr_offset;
1732                 eof = res->mcr_offset == res->mcr_size;
1733
1734                 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1735                        res->mcr_offset, eof == false);
1736         }
1737
1738         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1739         if (ealen < 0)
1740                 GOTO(out, rc = ealen);
1741
1742         if (ealen > nrpages << PAGE_SHIFT)
1743                 GOTO(out, rc = -EINVAL);
1744
1745         if (ealen == 0) { /* no logs transferred */
1746 #ifdef HAVE_SERVER_SUPPORT
1747                 /* config changed since first read RPC */
1748                 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1749                         CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1750                         GOTO(out, rc = -EAGAIN);
1751                 }
1752 #endif
1753                 if (!eof)
1754                         rc = -EINVAL;
1755                 GOTO(out, rc);
1756         }
1757
1758         mne_swab = req_capsule_rep_need_swab(&req->rq_pill);
1759
1760         /* When a nodemap config is received, we build a new nodemap config,
1761          * with new nodemap structs. We keep track of the most recently added
1762          * nodemap since the config is read ordered by nodemap_id, and so it
1763          * is likely that the next record will be related. Because access to
1764          * the nodemaps is single threaded until the nodemap_config is active,
1765          * we don't need to reference count with recent_nodemap, though
1766          * recent_nodemap should be set to NULL when the nodemap_config
1767          * is either destroyed or set active.
1768          */
1769         for (i = 0; i < nrpages && ealen > 0; i++) {
1770                 int rc2;
1771                 union lu_page   *ptr;
1772
1773                 ptr = kmap(pages[i]);
1774                 if (cld_is_nodemap(cld))
1775                         rc2 = nodemap_process_idx_pages(new_config, ptr,
1776                                                        &recent_nodemap);
1777                 else
1778                         rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1779                                                      ptr,
1780                                                      min_t(int, ealen,
1781                                                            PAGE_SIZE),
1782                                                      mne_swab);
1783                 kunmap(pages[i]);
1784                 if (rc2 < 0) {
1785                         CWARN("%s: error processing %s log %s: rc = %d\n",
1786                               obd->obd_name,
1787                               cld_is_nodemap(cld) ? "nodemap" : "recovery",
1788                               cld->cld_logname,
1789                               rc2);
1790                         GOTO(out, rc = rc2);
1791                 }
1792
1793                 ealen -= PAGE_SIZE;
1794         }
1795
1796 out:
1797         if (req) {
1798                 ptlrpc_req_finished(req);
1799                 req = NULL;
1800         }
1801
1802         if (rc == 0 && !eof)
1803                 goto again;
1804
1805 #ifdef HAVE_SERVER_SUPPORT
1806         if (new_config != NULL) {
1807                 /* recent_nodemap cannot be used after set_active/dealloc */
1808                 if (rc == 0)
1809                         nodemap_config_set_active_mgc(new_config);
1810                 else
1811                         nodemap_config_dealloc(new_config);
1812         }
1813 #endif
1814
1815         if (pages) {
1816                 for (i = 0; i < nrpages; i++) {
1817                         if (pages[i] == NULL)
1818                                 break;
1819                         __free_page(pages[i]);
1820                 }
1821                 OBD_FREE_PTR_ARRAY_LARGE(pages, nrpages);
1822         }
1823         return rc;
1824 }
1825
1826 static int mgc_barrier_glimpse_ast(struct ldlm_lock *lock, void *data)
1827 {
1828         struct config_llog_data *cld = lock->l_ast_data;
1829         int rc;
1830         ENTRY;
1831
1832         if (cld->cld_stopping)
1833                 RETURN(-ENODEV);
1834
1835         rc = barrier_handler(s2lsi(cld->cld_cfg.cfg_sb)->lsi_dt_dev,
1836                              (struct ptlrpc_request *)data);
1837
1838         RETURN(rc);
1839 }
1840
1841 /* Copy a remote log locally */
1842 static int mgc_llog_local_copy(const struct lu_env *env,
1843                                struct obd_device *obd,
1844                                struct llog_ctxt *rctxt,
1845                                struct llog_ctxt *lctxt, char *logname)
1846 {
1847         char    *temp_log;
1848         int      rc;
1849
1850         ENTRY;
1851
1852         /*
1853          * - copy it to backup using llog_backup()
1854          * - copy remote llog to logname using llog_backup()
1855          * - if failed then move bakup to logname again
1856          */
1857
1858         OBD_ALLOC(temp_log, strlen(logname) + 2);
1859         if (!temp_log)
1860                 RETURN(-ENOMEM);
1861         sprintf(temp_log, "%sT", logname);
1862
1863         /* make a copy of local llog at first */
1864         rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1865         if (rc < 0 && rc != -ENOENT)
1866                 GOTO(out, rc);
1867         /* copy remote llog to the local copy */
1868         rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1869         if (rc == -ENOENT) {
1870                 /* no remote llog, delete local one too */
1871                 llog_erase(env, lctxt, NULL, logname);
1872         } else if (rc < 0) {
1873                 /* error during backup, get local one back from the copy */
1874                 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1875 out:
1876                 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1877                        obd->obd_name, logname, rc);
1878         }
1879         llog_erase(env, lctxt, NULL, temp_log);
1880         OBD_FREE(temp_log, strlen(logname) + 2);
1881         return rc;
1882 }
1883
1884 /* local_only means it cannot get remote llogs */
1885 static int mgc_process_cfg_log(struct obd_device *mgc,
1886                                struct config_llog_data *cld, int local_only)
1887 {
1888         struct llog_ctxt        *ctxt, *lctxt = NULL;
1889         struct client_obd       *cli = &mgc->u.cli;
1890         struct lustre_sb_info   *lsi = NULL;
1891         int                      rc = 0;
1892         struct lu_env           *env;
1893
1894         ENTRY;
1895
1896         LASSERT(cld);
1897         LASSERT(mutex_is_locked(&cld->cld_lock));
1898
1899         if (cld->cld_cfg.cfg_sb)
1900                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1901
1902         OBD_ALLOC_PTR(env);
1903         if (env == NULL)
1904                 RETURN(-ENOMEM);
1905
1906         rc = lu_env_init(env, LCT_MG_THREAD);
1907         if (rc)
1908                 GOTO(out_free, rc);
1909
1910         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1911         LASSERT(ctxt);
1912
1913         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1914
1915         /* Copy the setup log locally if we can. Don't mess around if we're
1916          * running an MGS though (logs are already local). */
1917         if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1918             cli->cl_mgc_configs_dir != NULL &&
1919             lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1920             lsi->lsi_dt_dev) {
1921                 if (!local_only && !lsi->lsi_dt_dev->dd_rdonly) {
1922                         /* Only try to copy log if we have the lock. */
1923                         CDEBUG(D_INFO, "%s: copy local log %s\n",
1924                                mgc->obd_name, cld->cld_logname);
1925
1926                         rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1927                                                  cld->cld_logname);
1928                         if (!rc)
1929                                 lsi->lsi_flags &= ~LDD_F_NO_LOCAL_LOGS;
1930                 }
1931                 if (local_only || rc) {
1932                         if (unlikely(lsi->lsi_flags & LDD_F_NO_LOCAL_LOGS)
1933                             || rc) {
1934                                 CWARN("%s: local log %s are not valid and/or remote logs are not accessbile rc = %d\n",
1935                                       mgc->obd_name, cld->cld_logname, rc);
1936                                 GOTO(out_pop, rc = -EIO);
1937                         }
1938
1939                         if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1940                             llog_is_empty(env, lctxt, cld->cld_logname)) {
1941                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS log %s and no local copy.\n",
1942                                                    cld->cld_logname);
1943                                 GOTO(out_pop, rc = -ENOENT);
1944                         }
1945                         CDEBUG(D_MGC, "%s: Failed to get MGS log %s, using local copy for now, will try to update later.\n",
1946                                mgc->obd_name, cld->cld_logname);
1947                         rc = 0;
1948                 }
1949                 /* Now, whether we copied or not, start using the local llog.
1950                  * If we failed to copy, we'll start using whatever the old
1951                  * log has. */
1952                 llog_ctxt_put(ctxt);
1953                 ctxt = lctxt;
1954                 lctxt = NULL;
1955         } else {
1956                 if (local_only) /* no local log at client side */
1957                         GOTO(out_pop, rc = -EIO);
1958         }
1959
1960         rc = -EAGAIN;
1961         if (lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1962             lsi->lsi_dt_dev->dd_rdonly) {
1963                 struct llog_ctxt *rctxt;
1964
1965                 /* Under readonly mode, we may have no local copy or local
1966                  * copy is incomplete, so try to use remote llog firstly. */
1967                 rctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1968                 LASSERT(rctxt);
1969
1970                 rc = class_config_parse_llog(env, rctxt, cld->cld_logname,
1971                                              &cld->cld_cfg);
1972                 llog_ctxt_put(rctxt);
1973         }
1974
1975         if (rc && rc != -ENOENT)
1976                 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1977                                              &cld->cld_cfg);
1978
1979         /*
1980          * update settings on existing OBDs.
1981          * the logname must be <fsname>-sptlrpc
1982          */
1983         if (rc == 0 && cld_is_sptlrpc(cld))
1984                 class_notify_sptlrpc_conf(cld->cld_logname,
1985                                           strlen(cld->cld_logname) -
1986                                           strlen("-sptlrpc"));
1987         EXIT;
1988
1989 out_pop:
1990         __llog_ctxt_put(env, ctxt);
1991         if (lctxt)
1992                 __llog_ctxt_put(env, lctxt);
1993
1994         lu_env_fini(env);
1995 out_free:
1996         OBD_FREE_PTR(env);
1997         return rc;
1998 }
1999
2000 static bool mgc_import_in_recovery(struct obd_import *imp)
2001 {
2002         bool in_recovery = true;
2003
2004         spin_lock(&imp->imp_lock);
2005         if (imp->imp_state == LUSTRE_IMP_FULL ||
2006             imp->imp_state == LUSTRE_IMP_CLOSED)
2007                 in_recovery = false;
2008         spin_unlock(&imp->imp_lock);
2009
2010         return in_recovery;
2011 }
2012
2013 /**
2014  * Get a configuration log from the MGS and process it.
2015  *
2016  * This function is called for both clients and servers to process the
2017  * configuration log from the MGS.  The MGC enqueues a DLM lock on the
2018  * log from the MGS, and if the lock gets revoked the MGC will be notified
2019  * by the lock cancellation callback that the config log has changed,
2020  * and will enqueue another MGS lock on it, and then continue processing
2021  * the new additions to the end of the log.
2022  *
2023  * Since the MGC import is not replayable, if the import is being evicted
2024  * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
2025  * the log until recovery is finished or the import is closed.
2026  *
2027  * Make a local copy of the log before parsing it if appropriate (non-MGS
2028  * server) so that the server can start even when the MGS is down.
2029  *
2030  * There shouldn't be multiple processes running process_log at once --
2031  * sounds like badness.  It actually might be fine, as long as they're not
2032  * trying to update from the same log simultaneously, in which case we
2033  * should use a per-log semaphore instead of cld_lock.
2034  *
2035  * \param[in] mgc       MGC device by which to fetch the configuration log
2036  * \param[in] cld       log processing state (stored in lock callback data)
2037  *
2038  * \retval              0 on success
2039  * \retval              negative errno on failure
2040  */
2041 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
2042 {
2043         struct lustre_handle lockh = { 0 };
2044         __u64 flags = LDLM_FL_NO_LRU;
2045         int rc = 0, rcl;
2046         bool retry = false;
2047         ENTRY;
2048
2049         LASSERT(cld != NULL);
2050
2051         /* I don't want multiple processes running process_log at once --
2052            sounds like badness.  It actually might be fine, as long as
2053            we're not trying to update from the same log
2054            simultaneously (in which case we should use a per-log sem.) */
2055 restart:
2056         mutex_lock(&cld->cld_lock);
2057         if (cld->cld_stopping) {
2058                 mutex_unlock(&cld->cld_lock);
2059                 RETURN(0);
2060         }
2061
2062         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2063
2064         CDEBUG(D_MGC, "Process log %s-%016lx from %d\n", cld->cld_logname,
2065                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2066
2067         /* Get the cfg lock on the llog */
2068         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2069                           LCK_CR, &flags,
2070                           cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
2071                           cld, 0, NULL, &lockh);
2072         if (rcl == 0) {
2073                 /* Get the cld, it will be released in mgc_blocking_ast. */
2074                 config_log_get(cld);
2075                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2076                 LASSERT(!lustre_handle_is_used(&cld->cld_lockh));
2077                 LASSERT(rc == 0);
2078                 cld->cld_lockh = lockh;
2079         } else {
2080                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2081                 cld->cld_lockh.cookie = 0;
2082
2083                 if (rcl == -ESHUTDOWN &&
2084                     atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2085                         struct obd_import *imp;
2086                         long timeout = cfs_time_seconds(obd_timeout);
2087
2088                         mutex_unlock(&cld->cld_lock);
2089                         imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2090
2091                         /* Let's force the pinger, and wait the import to be
2092                          * connected, note: since mgc import is non-replayable,
2093                          * and even the import state is disconnected, it does
2094                          * not mean the "recovery" is stopped, so we will keep
2095                          * waitting until timeout or the import state is
2096                          * FULL or closed */
2097                         ptlrpc_pinger_force(imp);
2098
2099                         wait_event_idle_timeout(imp->imp_recovery_waitq,
2100                                                 !mgc_import_in_recovery(imp),
2101                                                 timeout);
2102
2103                         if (imp->imp_state == LUSTRE_IMP_FULL) {
2104                                 retry = true;
2105                                 goto restart;
2106                         } else {
2107                                 mutex_lock(&cld->cld_lock);
2108                                 /* unlock/lock mutex, so check stopping again */
2109                                 if (cld->cld_stopping) {
2110                                         mutex_unlock(&cld->cld_lock);
2111                                         RETURN(0);
2112                                 }
2113                                 spin_lock(&config_list_lock);
2114                                 cld->cld_lostlock = 1;
2115                                 spin_unlock(&config_list_lock);
2116                         }
2117                 } else {
2118                         /* mark cld_lostlock so that it will requeue
2119                          * after MGC becomes available. */
2120                         spin_lock(&config_list_lock);
2121                         cld->cld_lostlock = 1;
2122                         spin_unlock(&config_list_lock);
2123                 }
2124         }
2125
2126         if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2127                 if (!rcl)
2128                         rc = mgc_process_recover_nodemap_log(mgc, cld);
2129                 else if (cld_is_nodemap(cld))
2130                         rc = rcl;
2131
2132         } else if (!cld_is_barrier(cld)) {
2133                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2134         }
2135
2136         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2137                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2138
2139         /* Now drop the lock so MGS can revoke it */
2140         if (!rcl) {
2141                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2142                 if (rcl)
2143                         CERROR("Can't drop cfg lock: %d\n", rcl);
2144         }
2145         mutex_unlock(&cld->cld_lock);
2146
2147         /* requeue nodemap lock immediately if transfer was interrupted */
2148         if ((cld_is_nodemap(cld) && rc == -EAGAIN) ||
2149             (cld_is_recover(cld) && rc)) {
2150                 if (cld_is_recover(cld))
2151                         CWARN("%s: IR log %s failed, not fatal: rc = %d\n",
2152                               mgc->obd_name, cld->cld_logname, rc);
2153                 mgc_requeue_add(cld);
2154                 rc = 0;
2155         }
2156
2157         RETURN(rc);
2158 }
2159
2160
2161 /** Called from lustre_process_log.
2162  * LCFG_LOG_START gets the config log from the MGS, processes it to start
2163  * any services, and adds it to the list logs to watch (follow).
2164  */
2165 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2166 {
2167         struct lustre_cfg *lcfg = buf;
2168         struct config_llog_instance *cfg = NULL;
2169         char *logname;
2170         int rc = 0;
2171         ENTRY;
2172
2173         switch(lcfg->lcfg_command) {
2174         case LCFG_LOV_ADD_OBD: {
2175                 /* Overloading this cfg command: register a new target */
2176                 struct mgs_target_info *mti;
2177
2178                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2179                     sizeof(struct mgs_target_info))
2180                         GOTO(out, rc = -EINVAL);
2181
2182                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2183                 CDEBUG(D_MGC, "add_target %s %#x\n",
2184                        mti->mti_svname, mti->mti_flags);
2185                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2186                 break;
2187         }
2188         case LCFG_LOV_DEL_OBD:
2189                 /* Unregister has no meaning at the moment. */
2190                 CERROR("lov_del_obd unimplemented\n");
2191                 rc = -ENOSYS;
2192                 break;
2193         case LCFG_SPTLRPC_CONF: {
2194                 rc = sptlrpc_process_config(lcfg);
2195                 break;
2196         }
2197         case LCFG_LOG_START: {
2198                 struct config_llog_data *cld;
2199                 struct super_block *sb;
2200
2201                 logname = lustre_cfg_string(lcfg, 1);
2202                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2203                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2204
2205                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2206                        cfg->cfg_last_idx);
2207
2208                 /* We're only called through here on the initial mount */
2209                 cld = config_log_add(obd, logname, cfg, sb);
2210                 if (IS_ERR(cld)) {
2211                         rc = PTR_ERR(cld);
2212                         break;
2213                 }
2214
2215                 rc = mgc_process_log(obd, cld);
2216                 if (rc == 0 && cld->cld_recover != NULL) {
2217                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2218                                          imp_connect_data, IMP_RECOV)) {
2219                                 rc = mgc_process_log(obd, cld->cld_recover);
2220                         } else {
2221                                 struct config_llog_data *cir;
2222
2223                                 mutex_lock(&cld->cld_lock);
2224                                 cir = cld->cld_recover;
2225                                 cld->cld_recover = NULL;
2226                                 mutex_unlock(&cld->cld_lock);
2227                                 config_log_put(cir);
2228                         }
2229
2230                         if (rc)
2231                                 CERROR("Cannot process recover llog %d\n", rc);
2232                 }
2233
2234                 if (rc == 0 && cld->cld_params != NULL) {
2235                         rc = mgc_process_log(obd, cld->cld_params);
2236                         if (rc == -ENOENT) {
2237                                 CDEBUG(D_MGC, "There is no params "
2238                                               "config file yet\n");
2239                                 rc = 0;
2240                         }
2241                         /* params log is optional */
2242                         if (rc)
2243                                 CERROR("%s: can't process params llog: rc = %d\n",
2244                                        obd->obd_name, rc);
2245                 }
2246
2247                 break;
2248         }
2249         case LCFG_LOG_END: {
2250                 logname = lustre_cfg_string(lcfg, 1);
2251
2252                 if (lcfg->lcfg_bufcount >= 2)
2253                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
2254                                 lcfg, 2);
2255                 rc = config_log_end(logname, cfg);
2256                 break;
2257         }
2258         default: {
2259                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2260                 GOTO(out, rc = -EINVAL);
2261
2262         }
2263         }
2264 out:
2265         RETURN(rc);
2266 }
2267
2268 static const struct obd_ops mgc_obd_ops = {
2269         .o_owner        = THIS_MODULE,
2270         .o_setup        = mgc_setup,
2271         .o_precleanup   = mgc_precleanup,
2272         .o_cleanup      = mgc_cleanup,
2273         .o_add_conn     = client_import_add_conn,
2274         .o_del_conn     = client_import_del_conn,
2275         .o_connect      = client_connect_import,
2276         .o_disconnect   = client_disconnect_export,
2277         .o_set_info_async = mgc_set_info_async,
2278         .o_get_info       = mgc_get_info,
2279         .o_import_event = mgc_import_event,
2280         .o_process_config = mgc_process_config,
2281 };
2282
2283 static int mgc_param_requeue_timeout_min_set(const char *val,
2284                                      cfs_kernel_param_arg_t *kp)
2285 {
2286         int rc;
2287         unsigned int num;
2288
2289         rc = kstrtouint(val, 0, &num);
2290         if (rc < 0)
2291                 return rc;
2292         if (num > 120)
2293                 return -EINVAL;
2294
2295         mgc_requeue_timeout_min = num;
2296
2297         return 0;
2298 }
2299
2300 static const struct kernel_param_ops param_ops_requeue_timeout_min = {
2301         .set = mgc_param_requeue_timeout_min_set,
2302         .get = param_get_uint,
2303 };
2304
2305 #define param_check_requeue_timeout_min(name, p) \
2306                 __param_check(name, p, unsigned int)
2307
2308 unsigned int mgc_requeue_timeout_min = MGC_TIMEOUT_MIN_SECONDS;
2309 #ifdef HAVE_KERNEL_PARAM_OPS
2310 module_param(mgc_requeue_timeout_min, requeue_timeout_min, 0644);
2311 #else
2312 module_param_call(mgc_requeue_timeout_min, mgc_param_requeue_timeout_min_set,
2313                   param_get_uint, &param_ops_requeue_timeout_min, 0644);
2314 #endif
2315 MODULE_PARM_DESC(mgc_requeue_timeout_min, "Minimal requeue time to refresh logs");
2316
2317 static int __init mgc_init(void)
2318 {
2319         return class_register_type(&mgc_obd_ops, NULL, false,
2320                                    LUSTRE_MGC_NAME, NULL);
2321 }
2322
2323 static void __exit mgc_exit(void)
2324 {
2325         class_unregister_type(LUSTRE_MGC_NAME);
2326 }
2327
2328 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2329 MODULE_DESCRIPTION("Lustre Management Client");
2330 MODULE_VERSION(LUSTRE_VERSION_STRING);
2331 MODULE_LICENSE("GPL");
2332
2333 module_init(mgc_init);
2334 module_exit(mgc_exit);