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 /* this timeout represents how many seconds MGC should wait before
 623  * requeue config and recover lock to the MGS. We need to randomize this
 624  * in order to not flood the MGS.
 625  */
 626 #define MGC_TIMEOUT_MIN_SECONDS   5
 627 #define MGC_TIMEOUT_RAND_CENTISEC 500
 628
 629 static int mgc_requeue_thread(void *data)
 630 {
 631         int rc = 0;
 632         bool first = true;
 633         ENTRY;
 634
 635         CDEBUG(D_MGC, "Starting requeue thread\n");
 636
 637         /* Keep trying failed locks periodically */
 638         spin_lock(&config_list_lock);
 639         rq_state |= RQ_RUNNING;
 640         while (!(rq_state & RQ_STOP)) {
 641                 struct config_llog_data *cld, *cld_prev;
 642                 int rand = prandom_u32_max(MGC_TIMEOUT_RAND_CENTISEC);
 643                 int to;
 644
 645                 /* Any new or requeued lostlocks will change the state */
 646                 rq_state &= ~(RQ_NOW | RQ_LATER);
 647                 spin_unlock(&config_list_lock);
 648
 649                 if (first) {
 650                         first = false;
 651                         complete(&rq_start);
 652                 }
 653
 654                 /* Always wait a few seconds to allow the server who
 655                  * caused the lock revocation to finish its setup, plus some
 656                  * random so everyone doesn't try to reconnect at once.
 657                  */
 658                 to = cfs_time_seconds(MGC_TIMEOUT_MIN_SECONDS * 100 + rand);
 659                 /* rand is centi-seconds */
 660                 wait_event_idle_timeout(rq_waitq,
 661                                         rq_state & (RQ_STOP | RQ_PRECLEANUP),
 662                                         to/100);
 663
 664                 /*
 665                  * iterate & processing through the list. for each cld, process
 666                  * its depending sptlrpc cld firstly (if any) and then itself.
 667                  *
 668                  * it's guaranteed any item in the list must have
 669                  * reference > 0; and if cld_lostlock is set, at
 670                  * least one reference is taken by the previous enqueue.
 671                  */
 672                 cld_prev = NULL;
 673
 674                 spin_lock(&config_list_lock);
 675                 rq_state &= ~RQ_PRECLEANUP;
 676                 list_for_each_entry(cld, &config_llog_list,
 677                                     cld_list_chain) {
 678                         if (!cld->cld_lostlock || cld->cld_stopping)
 679                                 continue;
 680
 681                         /* hold reference to avoid being freed during
 682                          * subsequent processing. */
 683                         config_log_get(cld);
 684                         cld->cld_lostlock = 0;
 685                         spin_unlock(&config_list_lock);
 686
 687                         config_log_put(cld_prev);
 688                         cld_prev = cld;
 689
 690                         if (likely(!(rq_state & RQ_STOP))) {
 691                                 do_requeue(cld);
 692                                 spin_lock(&config_list_lock);
 693                         } else {
 694                                 spin_lock(&config_list_lock);
 695                                 break;
 696                         }
 697                 }
 698                 spin_unlock(&config_list_lock);
 699                 config_log_put(cld_prev);
 700
 701                 /* Wait a bit to see if anyone else needs a requeue */
 702                 wait_event_idle(rq_waitq, rq_state & (RQ_NOW | RQ_STOP));
 703                 spin_lock(&config_list_lock);
 704         }
 705
 706         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 707         rq_state &= ~RQ_RUNNING;
 708         spin_unlock(&config_list_lock);
 709
 710         complete(&rq_exit);
 711
 712         CDEBUG(D_MGC, "Ending requeue thread\n");
 713         RETURN(rc);
 714 }
 715
 716 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 717    We are responsible for dropping the config log reference from here on out. */
 718 static void mgc_requeue_add(struct config_llog_data *cld)
 719 {
 720         bool wakeup = false;
 721         ENTRY;
 722
 723         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 724                 cld->cld_logname, atomic_read(&cld->cld_refcount),
 725                 cld->cld_stopping, rq_state);
 726         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 727
 728         /* lets cancel an existent lock to mark cld as "lostlock" */
 729         CDEBUG(D_INFO, "lockh %#llx\n", cld->cld_lockh.cookie);
 730         if (!ldlm_lock_addref_try(&cld->cld_lockh, LCK_CR))
 731                 ldlm_lock_decref_and_cancel(&cld->cld_lockh, LCK_CR);
 732
 733         mutex_lock(&cld->cld_lock);
 734         spin_lock(&config_list_lock);
 735         if (!(rq_state & RQ_STOP) && !cld->cld_stopping) {
 736                 cld->cld_lostlock = 1;
 737                 rq_state |= RQ_NOW;
 738                 wakeup = true;
 739         }
 740         spin_unlock(&config_list_lock);
 741         mutex_unlock(&cld->cld_lock);
 742         if (wakeup)
 743                 wake_up(&rq_waitq);
 744
 745         EXIT;
 746 }
 747
 748 /********************** class fns **********************/
 749 static int mgc_local_llog_init(const struct lu_env *env,
 750                                struct obd_device *obd,
 751                                struct obd_device *disk)
 752 {
 753         struct llog_ctxt        *ctxt;
 754         int                      rc;
 755
 756         ENTRY;
 757
 758         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
 759                         &llog_osd_ops);
 760         if (rc)
 761                 RETURN(rc);
 762
 763         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 764         LASSERT(ctxt);
 765         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
 766         llog_ctxt_put(ctxt);
 767
 768         RETURN(0);
 769 }
 770
 771 static int mgc_local_llog_fini(const struct lu_env *env,
 772                                struct obd_device *obd)
 773 {
 774         struct llog_ctxt *ctxt;
 775
 776         ENTRY;
 777
 778         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 779         llog_cleanup(env, ctxt);
 780
 781         RETURN(0);
 782 }
 783
 784 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
 785                         struct super_block *sb)
 786 {
 787         struct lustre_sb_info   *lsi = s2lsi(sb);
 788         struct client_obd       *cli = &obd->u.cli;
 789         struct lu_fid            rfid, fid;
 790         struct dt_object        *root, *dto;
 791         int                      rc = 0;
 792
 793         ENTRY;
 794
 795         LASSERT(lsi);
 796         LASSERT(lsi->lsi_dt_dev);
 797
 798         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
 799         mutex_lock(&cli->cl_mgc_mutex);
 800
 801         /* Setup the configs dir */
 802         fid.f_seq = FID_SEQ_LOCAL_NAME;
 803         fid.f_oid = 1;
 804         fid.f_ver = 0;
 805         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
 806                                     &cli->cl_mgc_los);
 807         if (rc)
 808                 RETURN(rc);
 809
 810         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
 811         if (rc)
 812                 GOTO(out_los, rc);
 813
 814         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
 815                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
 816         if (unlikely(IS_ERR(root)))
 817                 GOTO(out_los, rc = PTR_ERR(root));
 818
 819         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
 820                                         MOUNT_CONFIGS_DIR,
 821                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
 822         dt_object_put_nocache(env, root);
 823         if (IS_ERR(dto))
 824                 GOTO(out_los, rc = PTR_ERR(dto));
 825
 826         cli->cl_mgc_configs_dir = dto;
 827
 828         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
 829         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
 830         if (rc)
 831                 GOTO(out_llog, rc);
 832
 833         /* We take an obd ref to insure that we can't get to mgc_cleanup
 834          * without calling mgc_fs_cleanup first. */
 835         class_incref(obd, "mgc_fs", obd);
 836
 837         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 838         EXIT;
 839 out_llog:
 840         if (rc) {
 841                 dt_object_put(env, cli->cl_mgc_configs_dir);
 842                 cli->cl_mgc_configs_dir = NULL;
 843         }
 844 out_los:
 845         if (rc < 0) {
 846                 local_oid_storage_fini(env, cli->cl_mgc_los);
 847                 cli->cl_mgc_los = NULL;
 848                 mutex_unlock(&cli->cl_mgc_mutex);
 849         }
 850         return rc;
 851 }
 852
 853 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
 854 {
 855         struct client_obd       *cli = &obd->u.cli;
 856         ENTRY;
 857
 858         LASSERT(cli->cl_mgc_los != NULL);
 859
 860         mgc_local_llog_fini(env, obd);
 861
 862         dt_object_put_nocache(env, cli->cl_mgc_configs_dir);
 863         cli->cl_mgc_configs_dir = NULL;
 864
 865         local_oid_storage_fini(env, cli->cl_mgc_los);
 866         cli->cl_mgc_los = NULL;
 867
 868         class_decref(obd, "mgc_fs", obd);
 869         mutex_unlock(&cli->cl_mgc_mutex);
 870
 871         RETURN(0);
 872 }
 873
 874 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 875 {
 876         struct llog_ctxt        *ctxt;
 877         int                      rc;
 878
 879         ENTRY;
 880
 881         /* setup only remote ctxt, the local disk context is switched per each
 882          * filesystem during mgc_fs_setup() */
 883         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 884                         &llog_client_ops);
 885         if (rc)
 886                 RETURN(rc);
 887
 888         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 889         LASSERT(ctxt);
 890
 891         llog_initiator_connect(ctxt);
 892         llog_ctxt_put(ctxt);
 893
 894         RETURN(0);
 895 }
 896
 897 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 898 {
 899         struct llog_ctxt *ctxt;
 900
 901         ENTRY;
 902
 903         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 904         if (ctxt)
 905                 llog_cleanup(env, ctxt);
 906
 907         RETURN(0);
 908 }
 909
 910
 911 static atomic_t mgc_count = ATOMIC_INIT(0);
 912 static int mgc_precleanup(struct obd_device *obd)
 913 {
 914         int     rc = 0;
 915         int     temp;
 916         ENTRY;
 917
 918         if (atomic_dec_and_test(&mgc_count)) {
 919                 LASSERT(rq_state & RQ_RUNNING);
 920                 /* stop requeue thread */
 921                 temp = RQ_STOP;
 922         } else {
 923                 /* wakeup requeue thread to clean our cld */
 924                 temp = RQ_NOW | RQ_PRECLEANUP;
 925         }
 926
 927         spin_lock(&config_list_lock);
 928         rq_state |= temp;
 929         spin_unlock(&config_list_lock);
 930         wake_up(&rq_waitq);
 931
 932         if (temp & RQ_STOP)
 933                 wait_for_completion(&rq_exit);
 934         obd_cleanup_client_import(obd);
 935
 936         rc = mgc_llog_fini(NULL, obd);
 937         if (rc != 0)
 938                 CERROR("failed to cleanup llogging subsystems\n");
 939
 940         RETURN(rc);
 941 }
 942
 943 static int mgc_cleanup(struct obd_device *obd)
 944 {
 945         int rc;
 946         ENTRY;
 947
 948         /* COMPAT_146 - old config logs may have added profiles we don't
 949            know about */
 950         if (atomic_read(&obd->obd_type->typ_refcnt) <= 1)
 951                 /* Only for the last mgc */
 952                 class_del_profiles();
 953
 954         lprocfs_obd_cleanup(obd);
 955         ptlrpcd_decref();
 956
 957         rc = client_obd_cleanup(obd);
 958         RETURN(rc);
 959 }
 960
 961 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 962 {
 963         struct task_struct      *task;
 964         int                      rc;
 965         ENTRY;
 966
 967         rc = ptlrpcd_addref();
 968         if (rc < 0)
 969                 RETURN(rc);
 970
 971         rc = client_obd_setup(obd, lcfg);
 972         if (rc)
 973                 GOTO(err_decref, rc);
 974
 975         rc = mgc_llog_init(NULL, obd);
 976         if (rc) {
 977                 CERROR("failed to setup llogging subsystems\n");
 978                 GOTO(err_cleanup, rc);
 979         }
 980
 981         rc = mgc_tunables_init(obd);
 982         if (rc)
 983                 GOTO(err_sysfs, rc);
 984
 985         if (atomic_inc_return(&mgc_count) == 1) {
 986                 rq_state = 0;
 987                 init_waitqueue_head(&rq_waitq);
 988
 989                 /* start requeue thread */
 990                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 991                 if (IS_ERR(task)) {
 992                         rc = PTR_ERR(task);
 993                         CERROR("%s: cannot start requeue thread: rc = %d; "
 994                                "no more log updates\n",
 995                                obd->obd_name, rc);
 996                         GOTO(err_sysfs, rc);
 997                 }
 998                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 999                 rc = 0;
1000                 wait_for_completion(&rq_start);
1001         }
1002
1003         RETURN(rc);
1004
1005 err_sysfs:
1006         lprocfs_obd_cleanup(obd);
1007 err_cleanup:
1008         client_obd_cleanup(obd);
1009 err_decref:
1010         ptlrpcd_decref();
1011         RETURN(rc);
1012 }
1013
1014 /* based on ll_mdc_blocking_ast */
1015 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1016                             void *data, int flag)
1017 {
1018         struct lustre_handle lockh;
1019         struct config_llog_data *cld = (struct config_llog_data *)data;
1020         int rc = 0;
1021         ENTRY;
1022
1023         switch (flag) {
1024         case LDLM_CB_BLOCKING:
1025                 /* mgs wants the lock, give it up... */
1026                 LDLM_DEBUG(lock, "MGC blocking CB");
1027                 ldlm_lock2handle(lock, &lockh);
1028                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
1029                 break;
1030         case LDLM_CB_CANCELING:
1031                 /* We've given up the lock, prepare ourselves to update. */
1032                 LDLM_DEBUG(lock, "MGC cancel CB");
1033
1034                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
1035                        PLDLMRES(lock->l_resource),
1036                        (char *)&lock->l_resource->lr_name.name[0]);
1037
1038                 if (!cld) {
1039                         CDEBUG(D_INFO, "missing data, won't requeue\n");
1040                         break;
1041                 }
1042
1043                 /* held at mgc_process_log(). */
1044                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1045
1046                 lock->l_ast_data = NULL;
1047                 cld->cld_lockh.cookie = 0;
1048                 /* Are we done with this log? */
1049                 if (cld->cld_stopping) {
1050                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1051                                 cld->cld_logname);
1052                         config_log_put(cld);
1053                         break;
1054                 }
1055                 /* Make sure not to re-enqueue when the mgc is stopping
1056                    (we get called from client_disconnect_export) */
1057                 if (lock->l_conn_export == NULL ||
1058                     lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1059                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1060                                 cld->cld_logname);
1061                         config_log_put(cld);
1062                         break;
1063                 }
1064
1065                 /* Re-enqueue now */
1066                 mgc_requeue_add(cld);
1067                 config_log_put(cld);
1068                 break;
1069         default:
1070                 LBUG();
1071         }
1072
1073         RETURN(rc);
1074 }
1075
1076 /* Not sure where this should go... */
1077 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1078  * that we can have a chance to try the secondary MGS if any. */
1079 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1080                                 + PING_INTERVAL)
1081 #define  MGC_TARGET_REG_LIMIT 10
1082 #define  MGC_TARGET_REG_LIMIT_MAX RECONNECT_DELAY_MAX
1083 #define  MGC_SEND_PARAM_LIMIT 10
1084
1085 /* Take a config lock so we can get cancel notifications */
1086 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1087                        union ldlm_policy_data *policy, enum ldlm_mode mode,
1088                        __u64 *flags, ldlm_glimpse_callback glimpse_callback,
1089                        void *data, __u32 lvb_len, void *lvb_swabber,
1090                        struct lustre_handle *lockh)
1091 {
1092         struct config_llog_data *cld = (struct config_llog_data *)data;
1093         struct ldlm_enqueue_info einfo = {
1094                 .ei_type        = type,
1095                 .ei_mode        = mode,
1096                 .ei_cb_bl       = mgc_blocking_ast,
1097                 .ei_cb_cp       = ldlm_completion_ast,
1098                 .ei_cb_gl       = glimpse_callback,
1099         };
1100         struct ptlrpc_request *req;
1101         int short_limit = cld_is_sptlrpc(cld);
1102         int rc;
1103         ENTRY;
1104
1105         if (!exp)
1106                 RETURN(-EBADR);
1107
1108         CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1109                cld->cld_resid.name[0]);
1110
1111         /* We need a callback for every lockholder, so don't try to
1112            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1113         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1114                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1115                                         LDLM_ENQUEUE);
1116         if (req == NULL)
1117                 RETURN(-ENOMEM);
1118
1119         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1120         ptlrpc_request_set_replen(req);
1121
1122         /* check if this is server or client */
1123         if (cld->cld_cfg.cfg_sb) {
1124                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1125                 if (lsi && IS_SERVER(lsi))
1126                         short_limit = 1;
1127         }
1128         /* Limit how long we will wait for the enqueue to complete */
1129         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1130         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1131                               NULL, 0, LVB_T_NONE, lockh, 0);
1132         /* A failed enqueue should still call the mgc_blocking_ast,
1133            where it will be requeued if needed ("grant failed"). */
1134         ptlrpc_req_finished(req);
1135         RETURN(rc);
1136 }
1137
1138 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1139                       struct lustre_handle *lockh)
1140 {
1141         ENTRY;
1142
1143         ldlm_lock_decref(lockh, mode);
1144
1145         RETURN(0);
1146 }
1147
1148 static void mgc_notify_active(struct obd_device *unused)
1149 {
1150         /* wakeup mgc_requeue_thread to requeue mgc lock */
1151         spin_lock(&config_list_lock);
1152         rq_state |= RQ_NOW;
1153         spin_unlock(&config_list_lock);
1154         wake_up(&rq_waitq);
1155
1156         /* TODO: Help the MGS rebuild nidtbl. -jay */
1157 }
1158
1159 /* Send target_reg message to MGS */
1160 static int mgc_target_register(struct obd_export *exp,
1161                                struct mgs_target_info *mti)
1162 {
1163         struct ptlrpc_request  *req;
1164         struct mgs_target_info *req_mti, *rep_mti;
1165         int                     rc;
1166         ENTRY;
1167
1168         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1169                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1170                                         MGS_TARGET_REG);
1171         if (req == NULL)
1172                 RETURN(-ENOMEM);
1173
1174         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1175         if (!req_mti) {
1176                 ptlrpc_req_finished(req);
1177                 RETURN(-ENOMEM);
1178         }
1179
1180         memcpy(req_mti, mti, sizeof(*req_mti));
1181         ptlrpc_request_set_replen(req);
1182         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1183         /* Limit how long we will wait for the enqueue to complete */
1184         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1185
1186         /* if the target needs to regenerate the config log in MGS, it's better
1187          * to use some longer limit to let MGC have time to change connection to
1188          * another MGS (or try again with the same MGS) for the target (server)
1189          * will fail and exit if the request expired due to delay limit. */
1190         if (mti->mti_flags & (LDD_F_UPDATE | LDD_F_NEED_INDEX))
1191                 req->rq_delay_limit = MGC_TARGET_REG_LIMIT_MAX;
1192
1193         rc = ptlrpc_queue_wait(req);
1194         if (!rc) {
1195                 rep_mti = req_capsule_server_get(&req->rq_pill,
1196                                                  &RMF_MGS_TARGET_INFO);
1197                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1198                 CDEBUG(D_MGC, "register %s got index = %d\n",
1199                        mti->mti_svname, mti->mti_stripe_index);
1200         }
1201         ptlrpc_req_finished(req);
1202
1203         RETURN(rc);
1204 }
1205
1206 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1207                               u32 keylen, void *key,
1208                               u32 vallen, void *val,
1209                               struct ptlrpc_request_set *set)
1210 {
1211         int rc = -EINVAL;
1212         ENTRY;
1213
1214         /* Turn off initial_recov after we try all backup servers once */
1215         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1216                 struct obd_import *imp = class_exp2cliimp(exp);
1217                 int value;
1218                 if (vallen != sizeof(int))
1219                         RETURN(-EINVAL);
1220                 value = *(int *)val;
1221                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1222                        imp->imp_obd->obd_name, value,
1223                        imp->imp_deactive, imp->imp_invalid,
1224                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1225                        ptlrpc_import_state_name(imp->imp_state));
1226                 /* Resurrect the import immediately if
1227                  * 1. we previously got disconnected,
1228                  * 2. value > 1 (at the same node with MGS)
1229                  * */
1230                 if (imp->imp_state == LUSTRE_IMP_DISCON || value > 1)
1231                         ptlrpc_reconnect_import(imp);
1232
1233                 RETURN(0);
1234         }
1235
1236         /* FIXME move this to mgc_process_config */
1237         if (KEY_IS(KEY_REGISTER_TARGET)) {
1238                 struct mgs_target_info *mti;
1239                 if (vallen != sizeof(struct mgs_target_info))
1240                         RETURN(-EINVAL);
1241                 mti = (struct mgs_target_info *)val;
1242                 CDEBUG(D_MGC, "register_target %s %#x\n",
1243                        mti->mti_svname, mti->mti_flags);
1244                 rc =  mgc_target_register(exp, mti);
1245                 RETURN(rc);
1246         }
1247         if (KEY_IS(KEY_SET_FS)) {
1248                 struct super_block *sb = (struct super_block *)val;
1249
1250                 if (vallen != sizeof(struct super_block))
1251                         RETURN(-EINVAL);
1252
1253                 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1254                 RETURN(rc);
1255         }
1256         if (KEY_IS(KEY_CLEAR_FS)) {
1257                 if (vallen != 0)
1258                         RETURN(-EINVAL);
1259                 rc = mgc_fs_cleanup(env, exp->exp_obd);
1260                 RETURN(rc);
1261         }
1262         if (KEY_IS(KEY_MGSSEC)) {
1263                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1264                 struct sptlrpc_flavor  flvr;
1265
1266                 /*
1267                  * empty string means using current flavor, if which haven't
1268                  * been set yet, set it as null.
1269                  *
1270                  * if flavor has been set previously, check the asking flavor
1271                  * must match the existing one.
1272                  */
1273                 if (vallen == 0) {
1274                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1275                                 RETURN(0);
1276                         val = "null";
1277                         vallen = 4;
1278                 }
1279
1280                 rc = sptlrpc_parse_flavor(val, &flvr);
1281                 if (rc) {
1282                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1283                                (char *) val);
1284                         RETURN(rc);
1285                 }
1286
1287                 /*
1288                  * caller already hold a mutex
1289                  */
1290                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1291                         cli->cl_flvr_mgc = flvr;
1292                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1293                                   sizeof(flvr)) != 0) {
1294                         char    str[20];
1295
1296                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1297                                             str, sizeof(str));
1298                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1299                                        "currently %s is in use\n",
1300                                        (char *) val, str);
1301                         rc = -EPERM;
1302                 }
1303                 RETURN(rc);
1304         }
1305
1306         RETURN(rc);
1307 }
1308
1309 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1310                         __u32 keylen, void *key, __u32 *vallen, void *val)
1311 {
1312         int rc = -EINVAL;
1313
1314         if (KEY_IS(KEY_CONN_DATA)) {
1315                 struct obd_import *imp = class_exp2cliimp(exp);
1316                 struct obd_connect_data *data = val;
1317
1318                 if (*vallen == sizeof(*data)) {
1319                         *data = imp->imp_connect_data;
1320                         rc = 0;
1321                 }
1322         }
1323
1324         return rc;
1325 }
1326
1327 static int mgc_import_event(struct obd_device *obd,
1328                             struct obd_import *imp,
1329                             enum obd_import_event event)
1330 {
1331         int rc = 0;
1332
1333         LASSERT(imp->imp_obd == obd);
1334         CDEBUG(D_MGC, "import event %#x\n", event);
1335
1336         switch (event) {
1337         case IMP_EVENT_DISCON:
1338                 /* MGC imports should not wait for recovery */
1339                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1340                         ptlrpc_pinger_ir_down();
1341                 break;
1342         case IMP_EVENT_INACTIVE:
1343                 break;
1344         case IMP_EVENT_INVALIDATE: {
1345                 struct ldlm_namespace *ns = obd->obd_namespace;
1346                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1347                 break;
1348         }
1349         case IMP_EVENT_ACTIVE:
1350                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1351                 /* Clearing obd_no_recov allows us to continue pinging */
1352                 obd->obd_no_recov = 0;
1353                 mgc_notify_active(obd);
1354                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1355                         ptlrpc_pinger_ir_up();
1356                 break;
1357         case IMP_EVENT_OCD:
1358                 break;
1359         case IMP_EVENT_DEACTIVATE:
1360         case IMP_EVENT_ACTIVATE:
1361                 break;
1362         default:
1363                 CERROR("Unknown import event %#x\n", event);
1364                 LBUG();
1365         }
1366         RETURN(rc);
1367 }
1368
1369 enum {
1370         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1371         CONFIG_READ_NRPAGES      = 4
1372 };
1373
1374 static int mgc_apply_recover_logs(struct obd_device *mgc,
1375                                   struct config_llog_data *cld,
1376                                   __u64 max_version,
1377                                   void *data, int datalen, bool mne_swab)
1378 {
1379         struct config_llog_instance *cfg = &cld->cld_cfg;
1380         struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1381         struct mgs_nidtbl_entry *entry;
1382         struct lustre_cfg *lcfg;
1383         struct lustre_cfg_bufs bufs;
1384         u64 prev_version = 0;
1385         char *inst;
1386         char *buf;
1387         int bufsz;
1388         int pos = 0;
1389         int rc  = 0;
1390         int off = 0;
1391         unsigned long dynamic_nids;
1392
1393         ENTRY;
1394         LASSERT(cfg->cfg_instance != 0);
1395         LASSERT(ll_get_cfg_instance(cfg->cfg_sb) == cfg->cfg_instance);
1396
1397         /* get dynamic nids setting */
1398         dynamic_nids = mgc->obd_dynamic_nids;
1399
1400         OBD_ALLOC(inst, PAGE_SIZE);
1401         if (inst == NULL)
1402                 RETURN(-ENOMEM);
1403
1404         if (!IS_SERVER(lsi)) {
1405                 pos = snprintf(inst, PAGE_SIZE, "%016lx", cfg->cfg_instance);
1406                 if (pos >= PAGE_SIZE) {
1407                         OBD_FREE(inst, PAGE_SIZE);
1408                         return -E2BIG;
1409                 }
1410 #ifdef HAVE_SERVER_SUPPORT
1411         } else {
1412                 LASSERT(IS_MDT(lsi));
1413                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1414                                         PAGE_SIZE);
1415                 if (rc) {
1416                         OBD_FREE(inst, PAGE_SIZE);
1417                         RETURN(-EINVAL);
1418                 }
1419                 pos = strlen(inst);
1420 #endif /* HAVE_SERVER_SUPPORT */
1421         }
1422
1423         ++pos;
1424         buf   = inst + pos;
1425         bufsz = PAGE_SIZE - pos;
1426
1427         while (datalen > 0) {
1428                 int   entry_len = sizeof(*entry);
1429                 int   is_ost;
1430                 struct obd_device *obd;
1431                 struct obd_import *imp;
1432                 char *obdname;
1433                 char *cname;
1434                 char *params;
1435                 char *uuid;
1436
1437                 rc = -EINVAL;
1438                 if (datalen < sizeof(*entry))
1439                         break;
1440
1441                 entry = (typeof(entry))(data + off);
1442
1443                 /* sanity check */
1444                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1445                         break;
1446                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1447                         break;
1448                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1449                         break;
1450
1451                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1452                 if (datalen < entry_len) /* must have entry_len at least */
1453                         break;
1454
1455                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1456                 if (mne_swab)
1457                         lustre_swab_mgs_nidtbl_entry(entry);
1458                 if (entry->mne_length > PAGE_SIZE) {
1459                         CERROR("MNE too large (%u)\n", entry->mne_length);
1460                         break;
1461                 }
1462
1463                 if (entry->mne_length < entry_len)
1464                         break;
1465
1466                 off     += entry->mne_length;
1467                 datalen -= entry->mne_length;
1468                 if (datalen < 0)
1469                         break;
1470
1471                 if (entry->mne_version > max_version) {
1472                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1473                                entry->mne_version, max_version);
1474                         break;
1475                 }
1476
1477                 if (prev_version >= entry->mne_version) {
1478                         CERROR("index unsorted, prev %lld, now %lld\n",
1479                                prev_version, entry->mne_version);
1480                         break;
1481                 }
1482                 prev_version = entry->mne_version;
1483
1484                 /*
1485                  * Write a string with format "nid::instance" to
1486                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1487                  */
1488
1489                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1490                 memset(buf, 0, bufsz);
1491                 obdname = buf;
1492                 pos = 0;
1493
1494                 /* lustre-OST0001-osc-<instance #> */
1495                 strcpy(obdname, cld->cld_logname);
1496                 cname = strrchr(obdname, '-');
1497                 if (cname == NULL) {
1498                         CERROR("mgc %s: invalid logname %s\n",
1499                                mgc->obd_name, obdname);
1500                         break;
1501                 }
1502
1503                 pos = cname - obdname;
1504                 obdname[pos] = 0;
1505                 pos += sprintf(obdname + pos, "-%s%04x",
1506                                is_ost ? "OST" : "MDT", entry->mne_index);
1507
1508                 cname = is_ost ? "osc" : "mdc",
1509                         pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1510                 lustre_cfg_bufs_reset(&bufs, obdname);
1511
1512                 /* find the obd by obdname */
1513                 obd = class_name2obd(obdname);
1514                 if (obd == NULL) {
1515                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1516                                mgc->obd_name, obdname);
1517                         rc = 0;
1518                         /* this is a safe race, when the ost is starting up...*/
1519                         continue;
1520                 }
1521
1522                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1523                 ++pos;
1524                 params = buf + pos;
1525                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1526                 uuid = buf + pos;
1527
1528                 with_imp_locked(obd, imp, rc) {
1529                         /* iterate all nids to find one */
1530                         /* find uuid by nid */
1531                         /* create import entries if they don't exist */
1532                         rc = client_import_add_nids_to_conn(
1533                                 imp, entry->u.nids, entry->mne_nid_count,
1534                                 (struct obd_uuid *)uuid);
1535
1536                         if (rc == -ENOENT && dynamic_nids) {
1537                                 /* create a new connection for this import */
1538                                 char *primary_nid =
1539                                         libcfs_nid2str(entry->u.nids[0]);
1540                                 int prim_nid_len = strlen(primary_nid) + 1;
1541                                 struct obd_uuid server_uuid;
1542
1543                                 if (prim_nid_len > UUID_MAX)
1544                                         goto fail;
1545                                 strncpy(server_uuid.uuid, primary_nid,
1546                                         prim_nid_len);
1547
1548                                 CDEBUG(D_INFO, "Adding a connection for %s\n",
1549                                        primary_nid);
1550
1551                                 rc = client_import_dyn_add_conn(
1552                                         imp, &server_uuid, entry->u.nids[0], 1);
1553                                 if (rc < 0) {
1554                                         CERROR("%s: Failed to add new connection with NID '%s' to import: rc = %d\n",
1555                                                obd->obd_name, primary_nid, rc);
1556                                         goto fail;
1557                                 }
1558                                 rc = client_import_add_nids_to_conn(
1559                                         imp, entry->u.nids,
1560                                         entry->mne_nid_count,
1561                                         (struct obd_uuid *)uuid);
1562                                 if (rc < 0) {
1563                                         CERROR("%s: failed to lookup UUID: rc = %d\n",
1564                                                obd->obd_name, rc);
1565                                         goto fail;
1566                                 }
1567                         }
1568 fail:;
1569                 }
1570                 if (rc == -ENODEV) {
1571                         /* client does not connect to the OST yet */
1572                         rc = 0;
1573                         continue;
1574                 }
1575
1576                 if (rc < 0 && rc != -ENOSPC) {
1577                         CERROR("mgc: cannot find UUID by nid '%s': rc = %d\n",
1578                                libcfs_nid2str(entry->u.nids[0]), rc);
1579                         break;
1580                 }
1581
1582                 CDEBUG(D_INFO, "Found UUID '%s' by NID '%s'\n",
1583                        uuid, libcfs_nid2str(entry->u.nids[0]));
1584
1585                 pos += strlen(uuid);
1586                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1587                 LASSERT(pos < bufsz);
1588
1589                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1590
1591                 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1592                                                bufs.lcfg_buflen));
1593                 if (!lcfg) {
1594                         rc = -ENOMEM;
1595                         break;
1596                 }
1597                 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1598
1599                 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1600                        prev_version, max_version, obdname, params);
1601
1602                 rc = class_process_config(lcfg);
1603                 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1604                                               lcfg->lcfg_buflens));
1605                 if (rc)
1606                         CDEBUG(D_INFO, "process config for %s error %d\n",
1607                                obdname, rc);
1608
1609                 /* continue, even one with error */
1610         }
1611
1612         OBD_FREE(inst, PAGE_SIZE);
1613         RETURN(rc);
1614 }
1615
1616 /**
1617  * This function is called if this client was notified for target restarting
1618  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1619  * nodemap logs.
1620  */
1621 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1622                                            struct config_llog_data *cld)
1623 {
1624         struct ptlrpc_connection *mgc_conn;
1625         struct ptlrpc_request *req = NULL;
1626         struct config_llog_instance *cfg = &cld->cld_cfg;
1627         struct mgs_config_body *body;
1628         struct mgs_config_res *res;
1629         struct nodemap_config *new_config = NULL;
1630         struct lu_nodemap *recent_nodemap = NULL;
1631         struct ptlrpc_bulk_desc *desc;
1632         struct page **pages = NULL;
1633         __u64 config_read_offset = 0;
1634         __u8 nodemap_cur_pass = 0;
1635         int nrpages = 0;
1636         bool eof = true;
1637         bool mne_swab = false;
1638         int i;
1639         int ealen;
1640         int rc;
1641         ENTRY;
1642
1643         mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1644
1645         /* don't need to get local config */
1646         if (cld_is_nodemap(cld) && LNetIsPeerLocal(mgc_conn->c_peer.nid))
1647                 GOTO(out, rc = 0);
1648
1649         /* allocate buffer for bulk transfer.
1650          * if this is the first time for this mgs to read logs,
1651          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1652          * once; otherwise, it only reads increment of logs, this should be
1653          * small and CONFIG_READ_NRPAGES will be used.
1654          */
1655         nrpages = CONFIG_READ_NRPAGES;
1656         if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1657                 nrpages = CONFIG_READ_NRPAGES_INIT;
1658
1659         OBD_ALLOC_PTR_ARRAY_LARGE(pages, nrpages);
1660         if (pages == NULL)
1661                 GOTO(out, rc = -ENOMEM);
1662
1663         for (i = 0; i < nrpages; i++) {
1664                 pages[i] = alloc_page(GFP_KERNEL);
1665                 if (pages[i] == NULL)
1666                         GOTO(out, rc = -ENOMEM);
1667         }
1668
1669 again:
1670 #ifdef HAVE_SERVER_SUPPORT
1671         if (cld_is_nodemap(cld) && config_read_offset == 0) {
1672                 new_config = nodemap_config_alloc();
1673                 if (IS_ERR(new_config)) {
1674                         rc = PTR_ERR(new_config);
1675                         new_config = NULL;
1676                         GOTO(out, rc);
1677                 }
1678         }
1679 #endif
1680         LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1681         LASSERT(mutex_is_locked(&cld->cld_lock));
1682         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1683                                    &RQF_MGS_CONFIG_READ);
1684         if (req == NULL)
1685                 GOTO(out, rc = -ENOMEM);
1686
1687         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1688         if (rc)
1689                 GOTO(out, rc);
1690
1691         /* pack request */
1692         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1693         LASSERT(body != NULL);
1694         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1695         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1696             >= sizeof(body->mcb_name))
1697                 GOTO(out, rc = -E2BIG);
1698         if (cld_is_nodemap(cld))
1699                 body->mcb_offset = config_read_offset;
1700         else
1701                 body->mcb_offset = cfg->cfg_last_idx + 1;
1702         body->mcb_type   = cld->cld_type;
1703         body->mcb_bits   = PAGE_SHIFT;
1704         body->mcb_units  = nrpages;
1705         body->mcb_nm_cur_pass = nodemap_cur_pass;
1706
1707         /* allocate bulk transfer descriptor */
1708         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1709                                     PTLRPC_BULK_PUT_SINK,
1710                                     MGS_BULK_PORTAL,
1711                                     &ptlrpc_bulk_kiov_pin_ops);
1712         if (desc == NULL)
1713                 GOTO(out, rc = -ENOMEM);
1714
1715         for (i = 0; i < nrpages; i++)
1716                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1717                                                  PAGE_SIZE);
1718
1719         ptlrpc_request_set_replen(req);
1720         rc = ptlrpc_queue_wait(req);
1721         if (rc)
1722                 GOTO(out, rc);
1723
1724         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1725         if (!res)
1726                 GOTO(out, rc = -EPROTO);
1727
1728         if (cld_is_nodemap(cld)) {
1729                 config_read_offset = res->mcr_offset;
1730                 eof = config_read_offset == II_END_OFF;
1731                 nodemap_cur_pass = res->mcr_nm_cur_pass;
1732         } else {
1733                 if (res->mcr_size < res->mcr_offset)
1734                         GOTO(out, rc = -EINVAL);
1735
1736                 /* always update the index even though it might have errors with
1737                  * handling the recover logs
1738                  */
1739                 cfg->cfg_last_idx = res->mcr_offset;
1740                 eof = res->mcr_offset == res->mcr_size;
1741
1742                 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1743                        res->mcr_offset, eof == false);
1744         }
1745
1746         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1747         if (ealen < 0)
1748                 GOTO(out, rc = ealen);
1749
1750         if (ealen > nrpages << PAGE_SHIFT)
1751                 GOTO(out, rc = -EINVAL);
1752
1753         if (ealen == 0) { /* no logs transferred */
1754 #ifdef HAVE_SERVER_SUPPORT
1755                 /* config changed since first read RPC */
1756                 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1757                         CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1758                         GOTO(out, rc = -EAGAIN);
1759                 }
1760 #endif
1761                 if (!eof)
1762                         rc = -EINVAL;
1763                 GOTO(out, rc);
1764         }
1765
1766         mne_swab = req_capsule_rep_need_swab(&req->rq_pill);
1767
1768         /* When a nodemap config is received, we build a new nodemap config,
1769          * with new nodemap structs. We keep track of the most recently added
1770          * nodemap since the config is read ordered by nodemap_id, and so it
1771          * is likely that the next record will be related. Because access to
1772          * the nodemaps is single threaded until the nodemap_config is active,
1773          * we don't need to reference count with recent_nodemap, though
1774          * recent_nodemap should be set to NULL when the nodemap_config
1775          * is either destroyed or set active.
1776          */
1777         for (i = 0; i < nrpages && ealen > 0; i++) {
1778                 int rc2;
1779                 union lu_page   *ptr;
1780
1781                 ptr = kmap(pages[i]);
1782                 if (cld_is_nodemap(cld))
1783                         rc2 = nodemap_process_idx_pages(new_config, ptr,
1784                                                        &recent_nodemap);
1785                 else
1786                         rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1787                                                      ptr,
1788                                                      min_t(int, ealen,
1789                                                            PAGE_SIZE),
1790                                                      mne_swab);
1791                 kunmap(pages[i]);
1792                 if (rc2 < 0) {
1793                         CWARN("%s: error processing %s log %s: rc = %d\n",
1794                               obd->obd_name,
1795                               cld_is_nodemap(cld) ? "nodemap" : "recovery",
1796                               cld->cld_logname,
1797                               rc2);
1798                         GOTO(out, rc = rc2);
1799                 }
1800
1801                 ealen -= PAGE_SIZE;
1802         }
1803
1804 out:
1805         if (req) {
1806                 ptlrpc_req_finished(req);
1807                 req = NULL;
1808         }
1809
1810         if (rc == 0 && !eof)
1811                 goto again;
1812
1813 #ifdef HAVE_SERVER_SUPPORT
1814         if (new_config != NULL) {
1815                 /* recent_nodemap cannot be used after set_active/dealloc */
1816                 if (rc == 0)
1817                         nodemap_config_set_active_mgc(new_config);
1818                 else
1819                         nodemap_config_dealloc(new_config);
1820         }
1821 #endif
1822
1823         if (pages) {
1824                 for (i = 0; i < nrpages; i++) {
1825                         if (pages[i] == NULL)
1826                                 break;
1827                         __free_page(pages[i]);
1828                 }
1829                 OBD_FREE_PTR_ARRAY_LARGE(pages, nrpages);
1830         }
1831         return rc;
1832 }
1833
1834 static int mgc_barrier_glimpse_ast(struct ldlm_lock *lock, void *data)
1835 {
1836         struct config_llog_data *cld = lock->l_ast_data;
1837         int rc;
1838         ENTRY;
1839
1840         if (cld->cld_stopping)
1841                 RETURN(-ENODEV);
1842
1843         rc = barrier_handler(s2lsi(cld->cld_cfg.cfg_sb)->lsi_dt_dev,
1844                              (struct ptlrpc_request *)data);
1845
1846         RETURN(rc);
1847 }
1848
1849 /* Copy a remote log locally */
1850 static int mgc_llog_local_copy(const struct lu_env *env,
1851                                struct obd_device *obd,
1852                                struct llog_ctxt *rctxt,
1853                                struct llog_ctxt *lctxt, char *logname)
1854 {
1855         char    *temp_log;
1856         int      rc;
1857
1858         ENTRY;
1859
1860         /*
1861          * - copy it to backup using llog_backup()
1862          * - copy remote llog to logname using llog_backup()
1863          * - if failed then move bakup to logname again
1864          */
1865
1866         OBD_ALLOC(temp_log, strlen(logname) + 2);
1867         if (!temp_log)
1868                 RETURN(-ENOMEM);
1869         sprintf(temp_log, "%sT", logname);
1870
1871         /* make a copy of local llog at first */
1872         rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1873         if (rc < 0 && rc != -ENOENT)
1874                 GOTO(out, rc);
1875         /* copy remote llog to the local copy */
1876         rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1877         if (rc == -ENOENT) {
1878                 /* no remote llog, delete local one too */
1879                 llog_erase(env, lctxt, NULL, logname);
1880         } else if (rc < 0) {
1881                 /* error during backup, get local one back from the copy */
1882                 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1883 out:
1884                 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1885                        obd->obd_name, logname, rc);
1886         }
1887         llog_erase(env, lctxt, NULL, temp_log);
1888         OBD_FREE(temp_log, strlen(logname) + 2);
1889         return rc;
1890 }
1891
1892 /* local_only means it cannot get remote llogs */
1893 static int mgc_process_cfg_log(struct obd_device *mgc,
1894                                struct config_llog_data *cld, int local_only)
1895 {
1896         struct llog_ctxt        *ctxt, *lctxt = NULL;
1897         struct client_obd       *cli = &mgc->u.cli;
1898         struct lustre_sb_info   *lsi = NULL;
1899         int                      rc = 0;
1900         struct lu_env           *env;
1901
1902         ENTRY;
1903
1904         LASSERT(cld);
1905         LASSERT(mutex_is_locked(&cld->cld_lock));
1906
1907         if (cld->cld_cfg.cfg_sb)
1908                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1909
1910         OBD_ALLOC_PTR(env);
1911         if (env == NULL)
1912                 RETURN(-ENOMEM);
1913
1914         rc = lu_env_init(env, LCT_MG_THREAD);
1915         if (rc)
1916                 GOTO(out_free, rc);
1917
1918         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1919         LASSERT(ctxt);
1920
1921         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1922
1923         /* Copy the setup log locally if we can. Don't mess around if we're
1924          * running an MGS though (logs are already local). */
1925         if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1926             cli->cl_mgc_configs_dir != NULL &&
1927             lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1928             lsi->lsi_dt_dev) {
1929                 if (!local_only && !lsi->lsi_dt_dev->dd_rdonly) {
1930                         /* Only try to copy log if we have the lock. */
1931                         CDEBUG(D_INFO, "%s: copy local log %s\n",
1932                                mgc->obd_name, cld->cld_logname);
1933
1934                         rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1935                                                  cld->cld_logname);
1936                         if (!rc)
1937                                 lsi->lsi_flags &= ~LDD_F_NO_LOCAL_LOGS;
1938                 }
1939                 if (local_only || rc) {
1940                         if (unlikely(lsi->lsi_flags & LDD_F_NO_LOCAL_LOGS)
1941                             || rc) {
1942                                 CWARN("%s: local log %s are not valid and/or remote logs are not accessbile rc = %d\n",
1943                                       mgc->obd_name, cld->cld_logname, rc);
1944                                 GOTO(out_pop, rc = -EIO);
1945                         }
1946
1947                         if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1948                             llog_is_empty(env, lctxt, cld->cld_logname)) {
1949                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS log %s and no local copy.\n",
1950                                                    cld->cld_logname);
1951                                 GOTO(out_pop, rc = -ENOENT);
1952                         }
1953                         CDEBUG(D_MGC, "%s: Failed to get MGS log %s, using local copy for now, will try to update later.\n",
1954                                mgc->obd_name, cld->cld_logname);
1955                         rc = 0;
1956                 }
1957                 /* Now, whether we copied or not, start using the local llog.
1958                  * If we failed to copy, we'll start using whatever the old
1959                  * log has. */
1960                 llog_ctxt_put(ctxt);
1961                 ctxt = lctxt;
1962                 lctxt = NULL;
1963         } else {
1964                 if (local_only) /* no local log at client side */
1965                         GOTO(out_pop, rc = -EIO);
1966         }
1967
1968         rc = -EAGAIN;
1969         if (lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1970             lsi->lsi_dt_dev->dd_rdonly) {
1971                 struct llog_ctxt *rctxt;
1972
1973                 /* Under readonly mode, we may have no local copy or local
1974                  * copy is incomplete, so try to use remote llog firstly. */
1975                 rctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1976                 LASSERT(rctxt);
1977
1978                 rc = class_config_parse_llog(env, rctxt, cld->cld_logname,
1979                                              &cld->cld_cfg);
1980                 llog_ctxt_put(rctxt);
1981         }
1982
1983         if (rc && rc != -ENOENT)
1984                 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1985                                              &cld->cld_cfg);
1986
1987         /*
1988          * update settings on existing OBDs.
1989          * the logname must be <fsname>-sptlrpc
1990          */
1991         if (rc == 0 && cld_is_sptlrpc(cld))
1992                 class_notify_sptlrpc_conf(cld->cld_logname,
1993                                           strlen(cld->cld_logname) -
1994                                           strlen("-sptlrpc"));
1995         EXIT;
1996
1997 out_pop:
1998         __llog_ctxt_put(env, ctxt);
1999         if (lctxt)
2000                 __llog_ctxt_put(env, lctxt);
2001
2002         lu_env_fini(env);
2003 out_free:
2004         OBD_FREE_PTR(env);
2005         return rc;
2006 }
2007
2008 static bool mgc_import_in_recovery(struct obd_import *imp)
2009 {
2010         bool in_recovery = true;
2011
2012         spin_lock(&imp->imp_lock);
2013         if (imp->imp_state == LUSTRE_IMP_FULL ||
2014             imp->imp_state == LUSTRE_IMP_CLOSED)
2015                 in_recovery = false;
2016         spin_unlock(&imp->imp_lock);
2017
2018         return in_recovery;
2019 }
2020
2021 /**
2022  * Get a configuration log from the MGS and process it.
2023  *
2024  * This function is called for both clients and servers to process the
2025  * configuration log from the MGS.  The MGC enqueues a DLM lock on the
2026  * log from the MGS, and if the lock gets revoked the MGC will be notified
2027  * by the lock cancellation callback that the config log has changed,
2028  * and will enqueue another MGS lock on it, and then continue processing
2029  * the new additions to the end of the log.
2030  *
2031  * Since the MGC import is not replayable, if the import is being evicted
2032  * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
2033  * the log until recovery is finished or the import is closed.
2034  *
2035  * Make a local copy of the log before parsing it if appropriate (non-MGS
2036  * server) so that the server can start even when the MGS is down.
2037  *
2038  * There shouldn't be multiple processes running process_log at once --
2039  * sounds like badness.  It actually might be fine, as long as they're not
2040  * trying to update from the same log simultaneously, in which case we
2041  * should use a per-log semaphore instead of cld_lock.
2042  *
2043  * \param[in] mgc       MGC device by which to fetch the configuration log
2044  * \param[in] cld       log processing state (stored in lock callback data)
2045  *
2046  * \retval              0 on success
2047  * \retval              negative errno on failure
2048  */
2049 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
2050 {
2051         struct lustre_handle lockh = { 0 };
2052         __u64 flags = LDLM_FL_NO_LRU;
2053         int rc = 0, rcl;
2054         bool retry = false;
2055         ENTRY;
2056
2057         LASSERT(cld != NULL);
2058
2059         /* I don't want multiple processes running process_log at once --
2060            sounds like badness.  It actually might be fine, as long as
2061            we're not trying to update from the same log
2062            simultaneously (in which case we should use a per-log sem.) */
2063 restart:
2064         mutex_lock(&cld->cld_lock);
2065         if (cld->cld_stopping) {
2066                 mutex_unlock(&cld->cld_lock);
2067                 RETURN(0);
2068         }
2069
2070         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2071
2072         CDEBUG(D_MGC, "Process log %s-%016lx from %d\n", cld->cld_logname,
2073                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2074
2075         /* Get the cfg lock on the llog */
2076         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2077                           LCK_CR, &flags,
2078                           cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
2079                           cld, 0, NULL, &lockh);
2080         if (rcl == 0) {
2081                 /* Get the cld, it will be released in mgc_blocking_ast. */
2082                 config_log_get(cld);
2083                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2084                 LASSERT(!lustre_handle_is_used(&cld->cld_lockh));
2085                 LASSERT(rc == 0);
2086                 cld->cld_lockh = lockh;
2087         } else {
2088                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2089                 cld->cld_lockh.cookie = 0;
2090
2091                 if (rcl == -ESHUTDOWN &&
2092                     atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2093                         struct obd_import *imp;
2094                         long timeout = cfs_time_seconds(obd_timeout);
2095
2096                         mutex_unlock(&cld->cld_lock);
2097                         imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2098
2099                         /* Let's force the pinger, and wait the import to be
2100                          * connected, note: since mgc import is non-replayable,
2101                          * and even the import state is disconnected, it does
2102                          * not mean the "recovery" is stopped, so we will keep
2103                          * waitting until timeout or the import state is
2104                          * FULL or closed */
2105                         ptlrpc_pinger_force(imp);
2106
2107                         wait_event_idle_timeout(imp->imp_recovery_waitq,
2108                                                 !mgc_import_in_recovery(imp),
2109                                                 timeout);
2110
2111                         if (imp->imp_state == LUSTRE_IMP_FULL) {
2112                                 retry = true;
2113                                 goto restart;
2114                         } else {
2115                                 mutex_lock(&cld->cld_lock);
2116                                 /* unlock/lock mutex, so check stopping again */
2117                                 if (cld->cld_stopping) {
2118                                         mutex_unlock(&cld->cld_lock);
2119                                         RETURN(0);
2120                                 }
2121                                 spin_lock(&config_list_lock);
2122                                 cld->cld_lostlock = 1;
2123                                 spin_unlock(&config_list_lock);
2124                         }
2125                 } else {
2126                         /* mark cld_lostlock so that it will requeue
2127                          * after MGC becomes available. */
2128                         spin_lock(&config_list_lock);
2129                         cld->cld_lostlock = 1;
2130                         spin_unlock(&config_list_lock);
2131                 }
2132         }
2133
2134         if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2135                 if (!rcl)
2136                         rc = mgc_process_recover_nodemap_log(mgc, cld);
2137                 else if (cld_is_nodemap(cld))
2138                         rc = rcl;
2139
2140         } else if (!cld_is_barrier(cld)) {
2141                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2142         }
2143
2144         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2145                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2146
2147         /* Now drop the lock so MGS can revoke it */
2148         if (!rcl) {
2149                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2150                 if (rcl)
2151                         CERROR("Can't drop cfg lock: %d\n", rcl);
2152         }
2153         mutex_unlock(&cld->cld_lock);
2154
2155         /* requeue nodemap lock immediately if transfer was interrupted */
2156         if ((cld_is_nodemap(cld) && rc == -EAGAIN) ||
2157             (cld_is_recover(cld) && rc)) {
2158                 if (cld_is_recover(cld))
2159                         CWARN("%s: IR log %s failed, not fatal: rc = %d\n",
2160                               mgc->obd_name, cld->cld_logname, rc);
2161                 mgc_requeue_add(cld);
2162                 rc = 0;
2163         }
2164
2165         RETURN(rc);
2166 }
2167
2168
2169 /** Called from lustre_process_log.
2170  * LCFG_LOG_START gets the config log from the MGS, processes it to start
2171  * any services, and adds it to the list logs to watch (follow).
2172  */
2173 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2174 {
2175         struct lustre_cfg *lcfg = buf;
2176         struct config_llog_instance *cfg = NULL;
2177         char *logname;
2178         int rc = 0;
2179         ENTRY;
2180
2181         switch(lcfg->lcfg_command) {
2182         case LCFG_LOV_ADD_OBD: {
2183                 /* Overloading this cfg command: register a new target */
2184                 struct mgs_target_info *mti;
2185
2186                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2187                     sizeof(struct mgs_target_info))
2188                         GOTO(out, rc = -EINVAL);
2189
2190                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2191                 CDEBUG(D_MGC, "add_target %s %#x\n",
2192                        mti->mti_svname, mti->mti_flags);
2193                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2194                 break;
2195         }
2196         case LCFG_LOV_DEL_OBD:
2197                 /* Unregister has no meaning at the moment. */
2198                 CERROR("lov_del_obd unimplemented\n");
2199                 rc = -ENOSYS;
2200                 break;
2201         case LCFG_SPTLRPC_CONF: {
2202                 rc = sptlrpc_process_config(lcfg);
2203                 break;
2204         }
2205         case LCFG_LOG_START: {
2206                 struct config_llog_data *cld;
2207                 struct super_block *sb;
2208
2209                 logname = lustre_cfg_string(lcfg, 1);
2210                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2211                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2212
2213                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2214                        cfg->cfg_last_idx);
2215
2216                 /* We're only called through here on the initial mount */
2217                 cld = config_log_add(obd, logname, cfg, sb);
2218                 if (IS_ERR(cld)) {
2219                         rc = PTR_ERR(cld);
2220                         break;
2221                 }
2222
2223                 rc = mgc_process_log(obd, cld);
2224                 if (rc == 0 && cld->cld_recover != NULL) {
2225                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2226                                          imp_connect_data, IMP_RECOV)) {
2227                                 rc = mgc_process_log(obd, cld->cld_recover);
2228                         } else {
2229                                 struct config_llog_data *cir;
2230
2231                                 mutex_lock(&cld->cld_lock);
2232                                 cir = cld->cld_recover;
2233                                 cld->cld_recover = NULL;
2234                                 mutex_unlock(&cld->cld_lock);
2235                                 config_log_put(cir);
2236                         }
2237
2238                         if (rc)
2239                                 CERROR("Cannot process recover llog %d\n", rc);
2240                 }
2241
2242                 if (rc == 0 && cld->cld_params != NULL) {
2243                         rc = mgc_process_log(obd, cld->cld_params);
2244                         if (rc == -ENOENT) {
2245                                 CDEBUG(D_MGC, "There is no params "
2246                                               "config file yet\n");
2247                                 rc = 0;
2248                         }
2249                         /* params log is optional */
2250                         if (rc)
2251                                 CERROR("%s: can't process params llog: rc = %d\n",
2252                                        obd->obd_name, rc);
2253                 }
2254
2255                 break;
2256         }
2257         case LCFG_LOG_END: {
2258                 logname = lustre_cfg_string(lcfg, 1);
2259
2260                 if (lcfg->lcfg_bufcount >= 2)
2261                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
2262                                 lcfg, 2);
2263                 rc = config_log_end(logname, cfg);
2264                 break;
2265         }
2266         default: {
2267                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2268                 GOTO(out, rc = -EINVAL);
2269
2270         }
2271         }
2272 out:
2273         RETURN(rc);
2274 }
2275
2276 static const struct obd_ops mgc_obd_ops = {
2277         .o_owner        = THIS_MODULE,
2278         .o_setup        = mgc_setup,
2279         .o_precleanup   = mgc_precleanup,
2280         .o_cleanup      = mgc_cleanup,
2281         .o_add_conn     = client_import_add_conn,
2282         .o_del_conn     = client_import_del_conn,
2283         .o_connect      = client_connect_import,
2284         .o_disconnect   = client_disconnect_export,
2285         .o_set_info_async = mgc_set_info_async,
2286         .o_get_info       = mgc_get_info,
2287         .o_import_event = mgc_import_event,
2288         .o_process_config = mgc_process_config,
2289 };
2290
2291 static int __init mgc_init(void)
2292 {
2293         return class_register_type(&mgc_obd_ops, NULL, false,
2294                                    LUSTRE_MGC_NAME, NULL);
2295 }
2296
2297 static void __exit mgc_exit(void)
2298 {
2299         class_unregister_type(LUSTRE_MGC_NAME);
2300 }
2301
2302 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2303 MODULE_DESCRIPTION("Lustre Management Client");
2304 MODULE_VERSION(LUSTRE_VERSION_STRING);
2305 MODULE_LICENSE("GPL");
2306
2307 module_init(mgc_init);
2308 module_exit(mgc_exit);