lustre/mgc/mgc_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/mgc/mgc_request.c
  33  *
  34  * Author: Nathan Rutman <nathan@clusterfs.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_MGC
  38 #define D_MGC D_CONFIG /*|D_WARNING*/
  39
  40 #include <linux/module.h>
  41 #include <linux/kthread.h>
  42 #include <linux/random.h>
  43
  44 #include <dt_object.h>
  45 #include <lprocfs_status.h>
  46 #include <lustre_dlm.h>
  47 #include <lustre_disk.h>
  48 #include <lustre_log.h>
  49 #include <lustre_nodemap.h>
  50 #include <lustre_swab.h>
  51 #include <obd_class.h>
  52 #include <lustre_barrier.h>
  53
  54 #include "mgc_internal.h"
  55
  56 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
  57                           int type)
  58 {
  59         __u64 resname = 0;
  60
  61         if (len > sizeof(resname)) {
  62                 CERROR("name too long: %s\n", name);
  63                 return -EINVAL;
  64         }
  65         if (len <= 0) {
  66                 CERROR("missing name: %s\n", name);
  67                 return -EINVAL;
  68         }
  69         memcpy(&resname, name, len);
  70
  71         /* Always use the same endianness for the resid */
  72         memset(res_id, 0, sizeof(*res_id));
  73         res_id->name[0] = cpu_to_le64(resname);
  74         /* XXX: unfortunately, sptlprc and config llog share one lock */
  75         switch(type) {
  76         case CONFIG_T_CONFIG:
  77         case CONFIG_T_SPTLRPC:
  78                 resname = 0;
  79                 break;
  80         case CONFIG_T_RECOVER:
  81         case CONFIG_T_PARAMS:
  82         case CONFIG_T_NODEMAP:
  83         case CONFIG_T_BARRIER:
  84                 resname = type;
  85                 break;
  86         default:
  87                 LBUG();
  88         }
  89         res_id->name[1] = cpu_to_le64(resname);
  90         CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
  91                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  92         return 0;
  93 }
  94
  95 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
  96 {
  97         /* fsname is at most 8 chars long, maybe contain "-".
  98          * e.g. "lustre", "SUN-000" */
  99         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
 100 }
 101 EXPORT_SYMBOL(mgc_fsname2resid);
 102
 103 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
 104 {
 105         char *name_end;
 106         int len;
 107
 108         /* logname consists of "fsname-nodetype".
 109          * e.g. "lustre-MDT0001", "SUN-000-client"
 110          * there is an exception: llog "params" */
 111         name_end = strrchr(logname, '-');
 112         if (!name_end)
 113                 len = strlen(logname);
 114         else
 115                 len = name_end - logname;
 116         return mgc_name2resid(logname, len, res_id, type);
 117 }
 118 EXPORT_SYMBOL(mgc_logname2resid);
 119
 120 /********************** config llog list **********************/
 121 static LIST_HEAD(config_llog_list);
 122 static DEFINE_SPINLOCK(config_list_lock);       /* protects config_llog_list */
 123
 124 /* Take a reference to a config log */
 125 static int config_log_get(struct config_llog_data *cld)
 126 {
 127         ENTRY;
 128         atomic_inc(&cld->cld_refcount);
 129         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 130                 atomic_read(&cld->cld_refcount));
 131         RETURN(0);
 132 }
 133
 134 /* Drop a reference to a config log.  When no longer referenced,
 135    we can free the config log data */
 136 static void config_log_put(struct config_llog_data *cld)
 137 {
 138         ENTRY;
 139
 140         if (unlikely(!cld))
 141                 RETURN_EXIT;
 142
 143         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 144                 atomic_read(&cld->cld_refcount));
 145         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 146
 147         /* spinlock to make sure no item with 0 refcount in the list */
 148         if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 149                 list_del(&cld->cld_list_chain);
 150                 spin_unlock(&config_list_lock);
 151
 152                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 153
 154                 config_log_put(cld->cld_barrier);
 155                 config_log_put(cld->cld_recover);
 156                 config_log_put(cld->cld_params);
 157                 config_log_put(cld->cld_nodemap);
 158                 config_log_put(cld->cld_sptlrpc);
 159                 if (cld_is_sptlrpc(cld))
 160                         sptlrpc_conf_log_stop(cld->cld_logname);
 161
 162                 class_export_put(cld->cld_mgcexp);
 163                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 164         }
 165
 166         EXIT;
 167 }
 168
 169 /* Find a config log by name */
 170 static
 171 struct config_llog_data *config_log_find(char *logname,
 172                                          struct config_llog_instance *cfg)
 173 {
 174         struct config_llog_data *cld;
 175         struct config_llog_data *found = NULL;
 176         unsigned long cfg_instance;
 177
 178         ENTRY;
 179         LASSERT(logname != NULL);
 180
 181         cfg_instance = cfg ? cfg->cfg_instance : 0;
 182         spin_lock(&config_list_lock);
 183         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 184                 /* check if cfg_instance is the one we want */
 185                 if (cfg_instance != cld->cld_cfg.cfg_instance)
 186                         continue;
 187
 188                 /* instance may be NULL, should check name */
 189                 if (strcmp(logname, cld->cld_logname) == 0) {
 190                         found = cld;
 191                         config_log_get(found);
 192                         break;
 193                 }
 194         }
 195         spin_unlock(&config_list_lock);
 196         RETURN(found);
 197 }
 198
 199 static
 200 struct config_llog_data *do_config_log_add(struct obd_device *obd,
 201                                            char *logname,
 202                                            int type,
 203                                            struct config_llog_instance *cfg,
 204                                            struct super_block *sb)
 205 {
 206         struct config_llog_data *cld;
 207         int rc;
 208
 209         ENTRY;
 210
 211         CDEBUG(D_MGC, "do adding config log %s-%016lx\n", logname,
 212                cfg ? cfg->cfg_instance : 0);
 213
 214         OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
 215         if (!cld)
 216                 RETURN(ERR_PTR(-ENOMEM));
 217
 218         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 219         if (rc) {
 220                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 221                 RETURN(ERR_PTR(rc));
 222         }
 223
 224         strcpy(cld->cld_logname, logname);
 225         if (cfg)
 226                 cld->cld_cfg = *cfg;
 227         else
 228                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 229         mutex_init(&cld->cld_lock);
 230         cld->cld_cfg.cfg_last_idx = 0;
 231         cld->cld_cfg.cfg_flags = 0;
 232         cld->cld_cfg.cfg_sb = sb;
 233         cld->cld_type = type;
 234         atomic_set(&cld->cld_refcount, 1);
 235
 236         /* Keep the mgc around until we are done */
 237         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 238
 239         if (cld_is_sptlrpc(cld))
 240                 sptlrpc_conf_log_start(logname);
 241
 242         spin_lock(&config_list_lock);
 243         list_add(&cld->cld_list_chain, &config_llog_list);
 244         spin_unlock(&config_list_lock);
 245
 246         if (cld_is_sptlrpc(cld) || cld_is_nodemap(cld) || cld_is_barrier(cld)) {
 247                 rc = mgc_process_log(obd, cld);
 248                 if (rc && rc != -ENOENT)
 249                         CERROR("%s: failed processing log, type %d: rc = %d\n",
 250                                obd->obd_name, type, rc);
 251         }
 252
 253         RETURN(cld);
 254 }
 255
 256 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
 257                                         char *fsname,
 258                                         struct config_llog_instance *cfg,
 259                                         struct super_block *sb)
 260 {
 261         struct config_llog_instance lcfg = *cfg;
 262         struct lustre_sb_info *lsi = s2lsi(sb);
 263         struct config_llog_data *cld;
 264         char logname[32];
 265
 266         if (IS_OST(lsi))
 267                 return NULL;
 268
 269         /* for osp-on-ost, see lustre_start_osp() */
 270         if (IS_MDT(lsi) && lcfg.cfg_instance)
 271                 return NULL;
 272
 273         /* We have to use different llog for clients and MDTs for DNE,
 274          * where only clients are notified if one of DNE server restarts.
 275          */
 276         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 277         strncpy(logname, fsname, sizeof(logname));
 278         if (IS_SERVER(lsi)) { /* mdt */
 279                 LASSERT(lcfg.cfg_instance == 0);
 280                 lcfg.cfg_instance = ll_get_cfg_instance(sb);
 281                 strncat(logname, "-mdtir", sizeof(logname));
 282         } else {
 283                 LASSERT(lcfg.cfg_instance != 0);
 284                 strncat(logname, "-cliir", sizeof(logname));
 285         }
 286
 287         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
 288         return cld;
 289 }
 290
 291 static struct config_llog_data *config_log_find_or_add(struct obd_device *obd,
 292                                 char *logname, struct super_block *sb, int type,
 293                                 struct config_llog_instance *cfg)
 294 {
 295         struct config_llog_instance lcfg = *cfg;
 296         struct config_llog_data *cld;
 297
 298         /* Note class_config_llog_handler() depends on getting "obd" back */
 299         lcfg.cfg_instance = sb ? ll_get_cfg_instance(sb) : (unsigned long)obd;
 300
 301         cld = config_log_find(logname, &lcfg);
 302         if (unlikely(cld != NULL))
 303                 return cld;
 304
 305         return do_config_log_add(obd, logname, type, &lcfg, sb);
 306 }
 307
 308 /** Add this log to the list of active logs watched by an MGC.
 309  * Active means we're watching for updates.
 310  * We have one active log per "mount" - client instance or servername.
 311  * Each instance may be at a different point in the log.
 312  */
 313 static struct config_llog_data *
 314 config_log_add(struct obd_device *obd, char *logname,
 315                struct config_llog_instance *cfg, struct super_block *sb)
 316 {
 317         struct lustre_sb_info *lsi = s2lsi(sb);
 318         struct config_llog_data *cld = NULL;
 319         struct config_llog_data *sptlrpc_cld = NULL;
 320         struct config_llog_data *params_cld = NULL;
 321         struct config_llog_data *nodemap_cld = NULL;
 322         struct config_llog_data *barrier_cld = NULL;
 323         char seclogname[32];
 324         char *ptr;
 325         int rc;
 326         bool locked = false;
 327         ENTRY;
 328
 329         CDEBUG(D_MGC, "add config log %s-%016lx\n", logname,
 330                cfg->cfg_instance);
 331
 332         /*
 333          * for each regular log, the depended sptlrpc log name is
 334          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 335          */
 336         ptr = strrchr(logname, '-');
 337         if (ptr == NULL || ptr - logname > 8) {
 338                 CERROR("logname %s is too long\n", logname);
 339                 RETURN(ERR_PTR(-EINVAL));
 340         }
 341
 342         memcpy(seclogname, logname, ptr - logname);
 343         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 344
 345         if (cfg->cfg_sub_clds & CONFIG_SUB_SPTLRPC) {
 346                 sptlrpc_cld = config_log_find_or_add(obd, seclogname, NULL,
 347                                                      CONFIG_T_SPTLRPC, cfg);
 348                 if (IS_ERR(sptlrpc_cld)) {
 349                         CERROR("%s: can't create sptlrpc log %s: rc = %ld\n",
 350                                obd->obd_name, seclogname, PTR_ERR(sptlrpc_cld));
 351                         RETURN(sptlrpc_cld);
 352                 }
 353         }
 354
 355         if (!IS_MGS(lsi) && cfg->cfg_sub_clds & CONFIG_SUB_NODEMAP) {
 356                 nodemap_cld = config_log_find_or_add(obd, LUSTRE_NODEMAP_NAME,
 357                                                      NULL, CONFIG_T_NODEMAP,
 358                                                      cfg);
 359                 if (IS_ERR(nodemap_cld)) {
 360                         rc = PTR_ERR(nodemap_cld);
 361                         CERROR("%s: cannot create nodemap log: rc = %d\n",
 362                                obd->obd_name, rc);
 363                         GOTO(out_sptlrpc, rc);
 364                 }
 365         }
 366
 367         if (cfg->cfg_sub_clds & CONFIG_SUB_PARAMS) {
 368                 params_cld = config_log_find_or_add(obd, PARAMS_FILENAME, sb,
 369                                                     CONFIG_T_PARAMS, cfg);
 370                 if (IS_ERR(params_cld)) {
 371                         rc = PTR_ERR(params_cld);
 372                         CERROR("%s: can't create params log: rc = %d\n",
 373                                obd->obd_name, rc);
 374                         GOTO(out_nodemap, rc);
 375                 }
 376         }
 377
 378         if (IS_MDT(s2lsi(sb)) && cfg->cfg_sub_clds & CONFIG_SUB_BARRIER) {
 379                 snprintf(seclogname + (ptr - logname), sizeof(seclogname) - 1,
 380                          "-%s", BARRIER_FILENAME);
 381                 barrier_cld = config_log_find_or_add(obd, seclogname, sb,
 382                                                      CONFIG_T_BARRIER, cfg);
 383                 if (IS_ERR(barrier_cld)) {
 384                         rc = PTR_ERR(barrier_cld);
 385                         CERROR("%s: can't create barrier log: rc = %d\n",
 386                                obd->obd_name, rc);
 387                         GOTO(out_params, rc);
 388                 }
 389         }
 390
 391         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 392         if (IS_ERR(cld)) {
 393                 rc = PTR_ERR(cld);
 394                 CERROR("%s: can't create log: rc = %d\n",
 395                        obd->obd_name, rc);
 396                 GOTO(out_barrier, rc = PTR_ERR(cld));
 397         }
 398
 399         LASSERT(lsi->lsi_lmd);
 400         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR) &&
 401             cfg->cfg_sub_clds & CONFIG_SUB_RECOVER) {
 402                 struct config_llog_data *recover_cld;
 403
 404                 ptr = strrchr(seclogname, '-');
 405                 if (ptr != NULL) {
 406                         *ptr = 0;
 407                 } else {
 408                         CERROR("%s: sptlrpc log name not correct, %s: "
 409                                "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
 410                         GOTO(out_cld, rc = -EINVAL);
 411                 }
 412
 413                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 414                 if (IS_ERR(recover_cld)) {
 415                         rc = PTR_ERR(recover_cld);
 416                         CERROR("%s: can't create recover log: rc = %d\n",
 417                                obd->obd_name, rc);
 418                         GOTO(out_cld, rc);
 419                 }
 420
 421                 mutex_lock(&cld->cld_lock);
 422                 locked = true;
 423                 cld->cld_recover = recover_cld;
 424         }
 425
 426         if (!locked)
 427                 mutex_lock(&cld->cld_lock);
 428         cld->cld_params = params_cld;
 429         cld->cld_barrier = barrier_cld;
 430         cld->cld_nodemap = nodemap_cld;
 431         cld->cld_sptlrpc = sptlrpc_cld;
 432         mutex_unlock(&cld->cld_lock);
 433
 434         RETURN(cld);
 435
 436 out_cld:
 437         config_log_put(cld);
 438 out_barrier:
 439         config_log_put(barrier_cld);
 440 out_params:
 441         config_log_put(params_cld);
 442 out_nodemap:
 443         config_log_put(nodemap_cld);
 444 out_sptlrpc:
 445         config_log_put(sptlrpc_cld);
 446
 447         return ERR_PTR(rc);
 448 }
 449
 450 static inline void config_mark_cld_stop(struct config_llog_data *cld)
 451 {
 452         if (cld) {
 453                 mutex_lock(&cld->cld_lock);
 454                 spin_lock(&config_list_lock);
 455                 cld->cld_stopping = 1;
 456                 spin_unlock(&config_list_lock);
 457                 mutex_unlock(&cld->cld_lock);
 458         }
 459 }
 460
 461 /** Stop watching for updates on this log.
 462  */
 463 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 464 {
 465         struct config_llog_data *cld;
 466         struct config_llog_data *cld_sptlrpc = NULL;
 467         struct config_llog_data *cld_params = NULL;
 468         struct config_llog_data *cld_recover = NULL;
 469         struct config_llog_data *cld_nodemap = NULL;
 470         struct config_llog_data *cld_barrier = NULL;
 471         int rc = 0;
 472
 473         ENTRY;
 474
 475         cld = config_log_find(logname, cfg);
 476         if (cld == NULL)
 477                 RETURN(-ENOENT);
 478
 479         mutex_lock(&cld->cld_lock);
 480         /*
 481          * if cld_stopping is set, it means we didn't start the log thus
 482          * not owning the start ref. this can happen after previous umount:
 483          * the cld still hanging there waiting for lock cancel, and we
 484          * remount again but failed in the middle and call log_end without
 485          * calling start_log.
 486          */
 487         if (unlikely(cld->cld_stopping)) {
 488                 mutex_unlock(&cld->cld_lock);
 489                 /* drop the ref from the find */
 490                 config_log_put(cld);
 491                 RETURN(rc);
 492         }
 493
 494         spin_lock(&config_list_lock);
 495         cld->cld_stopping = 1;
 496         spin_unlock(&config_list_lock);
 497
 498         cld_recover = cld->cld_recover;
 499         cld->cld_recover = NULL;
 500         cld_params = cld->cld_params;
 501         cld->cld_params = NULL;
 502         cld_nodemap = cld->cld_nodemap;
 503         cld->cld_nodemap = NULL;
 504         cld_barrier = cld->cld_barrier;
 505         cld->cld_barrier = NULL;
 506         cld_sptlrpc = cld->cld_sptlrpc;
 507         cld->cld_sptlrpc = NULL;
 508         mutex_unlock(&cld->cld_lock);
 509
 510         config_mark_cld_stop(cld_recover);
 511         config_log_put(cld_recover);
 512
 513         config_mark_cld_stop(cld_params);
 514         config_log_put(cld_params);
 515
 516         /* don't set cld_stopping on nm lock as other targets may be active */
 517         config_log_put(cld_nodemap);
 518
 519         if (cld_barrier) {
 520                 mutex_lock(&cld_barrier->cld_lock);
 521                 cld_barrier->cld_stopping = 1;
 522                 mutex_unlock(&cld_barrier->cld_lock);
 523                 config_log_put(cld_barrier);
 524         }
 525
 526         config_log_put(cld_sptlrpc);
 527
 528         /* drop the ref from the find */
 529         config_log_put(cld);
 530         /* drop the start ref */
 531         config_log_put(cld);
 532
 533         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 534                rc);
 535         RETURN(rc);
 536 }
 537
 538 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 539 {
 540         struct obd_device       *obd = data;
 541         struct obd_import       *imp;
 542         struct obd_connect_data *ocd;
 543         struct config_llog_data *cld;
 544
 545         ENTRY;
 546         LASSERT(obd);
 547         LPROCFS_CLIMP_CHECK(obd);
 548         imp = obd->u.cli.cl_import;
 549         ocd = &imp->imp_connect_data;
 550
 551         seq_printf(m, "imperative_recovery: %s\n",
 552                    OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 553         seq_printf(m, "client_state:\n");
 554
 555         spin_lock(&config_list_lock);
 556         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 557                 if (cld->cld_recover == NULL)
 558                         continue;
 559                 seq_printf(m,  "    - { client: %s, nidtbl_version: %u }\n",
 560                            cld->cld_logname,
 561                            cld->cld_recover->cld_cfg.cfg_last_idx);
 562         }
 563         spin_unlock(&config_list_lock);
 564
 565         LPROCFS_CLIMP_EXIT(obd);
 566         RETURN(0);
 567 }
 568
 569 /* reenqueue any lost locks */
 570 #define RQ_RUNNING      0x1
 571 #define RQ_NOW          0x2
 572 #define RQ_LATER        0x4
 573 #define RQ_STOP         0x8
 574 #define RQ_PRECLEANUP   0x10
 575 static int                    rq_state = 0;
 576 static wait_queue_head_t      rq_waitq;
 577 static DECLARE_COMPLETION(rq_exit);
 578 static DECLARE_COMPLETION(rq_start);
 579
 580 static void do_requeue(struct config_llog_data *cld)
 581 {
 582         int rc = 0;
 583         ENTRY;
 584
 585         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 586
 587         /*
 588          * Do not run mgc_process_log on a disconnected export or an
 589          * export which is being disconnected. Take the client
 590          * semaphore to make the check non-racy.
 591          */
 592         down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
 593                          OBD_CLI_SEM_MGC);
 594         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 595                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 596                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 597                 if (rc && rc != -ENOENT)
 598                         CERROR("failed processing log: %d\n", rc);
 599         } else {
 600                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 601                        cld->cld_logname);
 602         }
 603         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 604
 605         EXIT;
 606 }
 607
 608 /* this timeout represents how many seconds MGC should wait before
 609  * requeue config and recover lock to the MGS. We need to randomize this
 610  * in order to not flood the MGS.
 611  */
 612 #define MGC_TIMEOUT_MIN_SECONDS   5
 613 #define MGC_TIMEOUT_RAND_CENTISEC 500
 614
 615 static int mgc_requeue_thread(void *data)
 616 {
 617         int rc = 0;
 618         bool first = true;
 619         ENTRY;
 620
 621         CDEBUG(D_MGC, "Starting requeue thread\n");
 622
 623         /* Keep trying failed locks periodically */
 624         spin_lock(&config_list_lock);
 625         rq_state |= RQ_RUNNING;
 626         while (!(rq_state & RQ_STOP)) {
 627                 struct l_wait_info lwi;
 628                 struct config_llog_data *cld, *cld_prev;
 629                 int rand = prandom_u32_max(MGC_TIMEOUT_RAND_CENTISEC);
 630                 int to;
 631
 632                 /* Any new or requeued lostlocks will change the state */
 633                 rq_state &= ~(RQ_NOW | RQ_LATER);
 634                 spin_unlock(&config_list_lock);
 635
 636                 if (first) {
 637                         first = false;
 638                         complete(&rq_start);
 639                 }
 640
 641                 /* Always wait a few seconds to allow the server who
 642                    caused the lock revocation to finish its setup, plus some
 643                    random so everyone doesn't try to reconnect at once. */
 644                 to = cfs_time_seconds(MGC_TIMEOUT_MIN_SECONDS * 100 + rand);
 645                 /* rand is centi-seconds */
 646                 lwi = LWI_TIMEOUT(to / 100, NULL, NULL);
 647                 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
 648                              &lwi);
 649
 650                 /*
 651                  * iterate & processing through the list. for each cld, process
 652                  * its depending sptlrpc cld firstly (if any) and then itself.
 653                  *
 654                  * it's guaranteed any item in the list must have
 655                  * reference > 0; and if cld_lostlock is set, at
 656                  * least one reference is taken by the previous enqueue.
 657                  */
 658                 cld_prev = NULL;
 659
 660                 spin_lock(&config_list_lock);
 661                 rq_state &= ~RQ_PRECLEANUP;
 662                 list_for_each_entry(cld, &config_llog_list,
 663                                     cld_list_chain) {
 664                         if (!cld->cld_lostlock || cld->cld_stopping)
 665                                 continue;
 666
 667                         /* hold reference to avoid being freed during
 668                          * subsequent processing. */
 669                         config_log_get(cld);
 670                         cld->cld_lostlock = 0;
 671                         spin_unlock(&config_list_lock);
 672
 673                         config_log_put(cld_prev);
 674                         cld_prev = cld;
 675
 676                         if (likely(!(rq_state & RQ_STOP))) {
 677                                 do_requeue(cld);
 678                                 spin_lock(&config_list_lock);
 679                         } else {
 680                                 spin_lock(&config_list_lock);
 681                                 break;
 682                         }
 683                 }
 684                 spin_unlock(&config_list_lock);
 685                 config_log_put(cld_prev);
 686
 687                 /* Wait a bit to see if anyone else needs a requeue */
 688                 wait_event_idle(rq_waitq, rq_state & (RQ_NOW | RQ_STOP));
 689                 spin_lock(&config_list_lock);
 690         }
 691
 692         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 693         rq_state &= ~RQ_RUNNING;
 694         spin_unlock(&config_list_lock);
 695
 696         complete(&rq_exit);
 697
 698         CDEBUG(D_MGC, "Ending requeue thread\n");
 699         RETURN(rc);
 700 }
 701
 702 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 703    We are responsible for dropping the config log reference from here on out. */
 704 static void mgc_requeue_add(struct config_llog_data *cld)
 705 {
 706         bool wakeup = false;
 707         ENTRY;
 708
 709         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 710                 cld->cld_logname, atomic_read(&cld->cld_refcount),
 711                 cld->cld_stopping, rq_state);
 712         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 713
 714         mutex_lock(&cld->cld_lock);
 715         spin_lock(&config_list_lock);
 716         if (!(rq_state & RQ_STOP) && !cld->cld_stopping && !cld->cld_lostlock) {
 717                 cld->cld_lostlock = 1;
 718                 rq_state |= RQ_NOW;
 719                 wakeup = true;
 720         }
 721         spin_unlock(&config_list_lock);
 722         mutex_unlock(&cld->cld_lock);
 723         if (wakeup)
 724                 wake_up(&rq_waitq);
 725
 726         EXIT;
 727 }
 728
 729 /********************** class fns **********************/
 730 static int mgc_local_llog_init(const struct lu_env *env,
 731                                struct obd_device *obd,
 732                                struct obd_device *disk)
 733 {
 734         struct llog_ctxt        *ctxt;
 735         int                      rc;
 736
 737         ENTRY;
 738
 739         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
 740                         &llog_osd_ops);
 741         if (rc)
 742                 RETURN(rc);
 743
 744         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 745         LASSERT(ctxt);
 746         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
 747         llog_ctxt_put(ctxt);
 748
 749         RETURN(0);
 750 }
 751
 752 static int mgc_local_llog_fini(const struct lu_env *env,
 753                                struct obd_device *obd)
 754 {
 755         struct llog_ctxt *ctxt;
 756
 757         ENTRY;
 758
 759         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 760         llog_cleanup(env, ctxt);
 761
 762         RETURN(0);
 763 }
 764
 765 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
 766                         struct super_block *sb)
 767 {
 768         struct lustre_sb_info   *lsi = s2lsi(sb);
 769         struct client_obd       *cli = &obd->u.cli;
 770         struct lu_fid            rfid, fid;
 771         struct dt_object        *root, *dto;
 772         int                      rc = 0;
 773
 774         ENTRY;
 775
 776         LASSERT(lsi);
 777         LASSERT(lsi->lsi_dt_dev);
 778
 779         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
 780         mutex_lock(&cli->cl_mgc_mutex);
 781
 782         /* Setup the configs dir */
 783         fid.f_seq = FID_SEQ_LOCAL_NAME;
 784         fid.f_oid = 1;
 785         fid.f_ver = 0;
 786         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
 787                                     &cli->cl_mgc_los);
 788         if (rc)
 789                 RETURN(rc);
 790
 791         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
 792         if (rc)
 793                 GOTO(out_los, rc);
 794
 795         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
 796                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
 797         if (unlikely(IS_ERR(root)))
 798                 GOTO(out_los, rc = PTR_ERR(root));
 799
 800         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
 801                                         MOUNT_CONFIGS_DIR,
 802                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
 803         dt_object_put_nocache(env, root);
 804         if (IS_ERR(dto))
 805                 GOTO(out_los, rc = PTR_ERR(dto));
 806
 807         cli->cl_mgc_configs_dir = dto;
 808
 809         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
 810         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
 811         if (rc)
 812                 GOTO(out_llog, rc);
 813
 814         /* We take an obd ref to insure that we can't get to mgc_cleanup
 815          * without calling mgc_fs_cleanup first. */
 816         class_incref(obd, "mgc_fs", obd);
 817
 818         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 819         EXIT;
 820 out_llog:
 821         if (rc) {
 822                 dt_object_put(env, cli->cl_mgc_configs_dir);
 823                 cli->cl_mgc_configs_dir = NULL;
 824         }
 825 out_los:
 826         if (rc < 0) {
 827                 local_oid_storage_fini(env, cli->cl_mgc_los);
 828                 cli->cl_mgc_los = NULL;
 829                 mutex_unlock(&cli->cl_mgc_mutex);
 830         }
 831         return rc;
 832 }
 833
 834 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
 835 {
 836         struct client_obd       *cli = &obd->u.cli;
 837         ENTRY;
 838
 839         LASSERT(cli->cl_mgc_los != NULL);
 840
 841         mgc_local_llog_fini(env, obd);
 842
 843         dt_object_put_nocache(env, cli->cl_mgc_configs_dir);
 844         cli->cl_mgc_configs_dir = NULL;
 845
 846         local_oid_storage_fini(env, cli->cl_mgc_los);
 847         cli->cl_mgc_los = NULL;
 848
 849         class_decref(obd, "mgc_fs", obd);
 850         mutex_unlock(&cli->cl_mgc_mutex);
 851
 852         RETURN(0);
 853 }
 854
 855 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 856 {
 857         struct llog_ctxt        *ctxt;
 858         int                      rc;
 859
 860         ENTRY;
 861
 862         /* setup only remote ctxt, the local disk context is switched per each
 863          * filesystem during mgc_fs_setup() */
 864         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 865                         &llog_client_ops);
 866         if (rc)
 867                 RETURN(rc);
 868
 869         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 870         LASSERT(ctxt);
 871
 872         llog_initiator_connect(ctxt);
 873         llog_ctxt_put(ctxt);
 874
 875         RETURN(0);
 876 }
 877
 878 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 879 {
 880         struct llog_ctxt *ctxt;
 881
 882         ENTRY;
 883
 884         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 885         if (ctxt)
 886                 llog_cleanup(env, ctxt);
 887
 888         RETURN(0);
 889 }
 890
 891
 892 static atomic_t mgc_count = ATOMIC_INIT(0);
 893 static int mgc_precleanup(struct obd_device *obd)
 894 {
 895         int     rc = 0;
 896         int     temp;
 897         ENTRY;
 898
 899         if (atomic_dec_and_test(&mgc_count)) {
 900                 LASSERT(rq_state & RQ_RUNNING);
 901                 /* stop requeue thread */
 902                 temp = RQ_STOP;
 903         } else {
 904                 /* wakeup requeue thread to clean our cld */
 905                 temp = RQ_NOW | RQ_PRECLEANUP;
 906         }
 907
 908         spin_lock(&config_list_lock);
 909         rq_state |= temp;
 910         spin_unlock(&config_list_lock);
 911         wake_up(&rq_waitq);
 912
 913         if (temp & RQ_STOP)
 914                 wait_for_completion(&rq_exit);
 915         obd_cleanup_client_import(obd);
 916
 917         rc = mgc_llog_fini(NULL, obd);
 918         if (rc != 0)
 919                 CERROR("failed to cleanup llogging subsystems\n");
 920
 921         RETURN(rc);
 922 }
 923
 924 static int mgc_cleanup(struct obd_device *obd)
 925 {
 926         int rc;
 927         ENTRY;
 928
 929         /* COMPAT_146 - old config logs may have added profiles we don't
 930            know about */
 931         if (atomic_read(&obd->obd_type->typ_refcnt) <= 1)
 932                 /* Only for the last mgc */
 933                 class_del_profiles();
 934
 935         lprocfs_obd_cleanup(obd);
 936         ptlrpcd_decref();
 937
 938         rc = client_obd_cleanup(obd);
 939         RETURN(rc);
 940 }
 941
 942 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 943 {
 944         struct task_struct      *task;
 945         int                      rc;
 946         ENTRY;
 947
 948         rc = ptlrpcd_addref();
 949         if (rc < 0)
 950                 RETURN(rc);
 951
 952         rc = client_obd_setup(obd, lcfg);
 953         if (rc)
 954                 GOTO(err_decref, rc);
 955
 956         rc = mgc_llog_init(NULL, obd);
 957         if (rc) {
 958                 CERROR("failed to setup llogging subsystems\n");
 959                 GOTO(err_cleanup, rc);
 960         }
 961
 962         rc = mgc_tunables_init(obd);
 963         if (rc)
 964                 GOTO(err_sysfs, rc);
 965
 966         if (atomic_inc_return(&mgc_count) == 1) {
 967                 rq_state = 0;
 968                 init_waitqueue_head(&rq_waitq);
 969
 970                 /* start requeue thread */
 971                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 972                 if (IS_ERR(task)) {
 973                         rc = PTR_ERR(task);
 974                         CERROR("%s: cannot start requeue thread: rc = %d; "
 975                                "no more log updates\n",
 976                                obd->obd_name, rc);
 977                         GOTO(err_sysfs, rc);
 978                 }
 979                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 980                 rc = 0;
 981                 wait_for_completion(&rq_start);
 982         }
 983
 984         RETURN(rc);
 985
 986 err_sysfs:
 987         lprocfs_obd_cleanup(obd);
 988 err_cleanup:
 989         client_obd_cleanup(obd);
 990 err_decref:
 991         ptlrpcd_decref();
 992         RETURN(rc);
 993 }
 994
 995 /* based on ll_mdc_blocking_ast */
 996 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 997                             void *data, int flag)
 998 {
 999         struct lustre_handle lockh;
1000         struct config_llog_data *cld = (struct config_llog_data *)data;
1001         int rc = 0;
1002         ENTRY;
1003
1004         switch (flag) {
1005         case LDLM_CB_BLOCKING:
1006                 /* mgs wants the lock, give it up... */
1007                 LDLM_DEBUG(lock, "MGC blocking CB");
1008                 ldlm_lock2handle(lock, &lockh);
1009                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
1010                 break;
1011         case LDLM_CB_CANCELING:
1012                 /* We've given up the lock, prepare ourselves to update. */
1013                 LDLM_DEBUG(lock, "MGC cancel CB");
1014
1015                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
1016                        PLDLMRES(lock->l_resource),
1017                        (char *)&lock->l_resource->lr_name.name[0]);
1018
1019                 if (!cld) {
1020                         CDEBUG(D_INFO, "missing data, won't requeue\n");
1021                         break;
1022                 }
1023
1024                 /* held at mgc_process_log(). */
1025                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1026
1027                 lock->l_ast_data = NULL;
1028                 /* Are we done with this log? */
1029                 if (cld->cld_stopping) {
1030                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1031                                 cld->cld_logname);
1032                         config_log_put(cld);
1033                         break;
1034                 }
1035                 /* Make sure not to re-enqueue when the mgc is stopping
1036                    (we get called from client_disconnect_export) */
1037                 if (lock->l_conn_export == NULL ||
1038                     lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1039                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1040                                 cld->cld_logname);
1041                         config_log_put(cld);
1042                         break;
1043                 }
1044
1045                 /* Re-enqueue now */
1046                 mgc_requeue_add(cld);
1047                 config_log_put(cld);
1048                 break;
1049         default:
1050                 LBUG();
1051         }
1052
1053         RETURN(rc);
1054 }
1055
1056 /* Not sure where this should go... */
1057 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1058  * that we can have a chance to try the secondary MGS if any. */
1059 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1060                                 + PING_INTERVAL)
1061 #define  MGC_TARGET_REG_LIMIT 10
1062 #define  MGC_TARGET_REG_LIMIT_MAX RECONNECT_DELAY_MAX
1063 #define  MGC_SEND_PARAM_LIMIT 10
1064
1065 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1066 /* Send parameter to MGS*/
1067 static int mgc_set_mgs_param(struct obd_export *exp,
1068                              struct mgs_send_param *msp)
1069 {
1070         struct ptlrpc_request *req;
1071         struct mgs_send_param *req_msp, *rep_msp;
1072         int rc;
1073         ENTRY;
1074
1075         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1076                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1077                                         MGS_SET_INFO);
1078         if (!req)
1079                 RETURN(-ENOMEM);
1080
1081         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1082         if (!req_msp) {
1083                 ptlrpc_req_finished(req);
1084                 RETURN(-ENOMEM);
1085         }
1086
1087         memcpy(req_msp, msp, sizeof(*req_msp));
1088         ptlrpc_request_set_replen(req);
1089
1090         /* Limit how long we will wait for the enqueue to complete */
1091         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1092         rc = ptlrpc_queue_wait(req);
1093         if (!rc) {
1094                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1095                 memcpy(msp, rep_msp, sizeof(*rep_msp));
1096         }
1097
1098         ptlrpc_req_finished(req);
1099
1100         RETURN(rc);
1101 }
1102 #endif
1103
1104 /* Take a config lock so we can get cancel notifications */
1105 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1106                        union ldlm_policy_data *policy, enum ldlm_mode mode,
1107                        __u64 *flags, ldlm_glimpse_callback glimpse_callback,
1108                        void *data, __u32 lvb_len, void *lvb_swabber,
1109                        struct lustre_handle *lockh)
1110 {
1111         struct config_llog_data *cld = (struct config_llog_data *)data;
1112         struct ldlm_enqueue_info einfo = {
1113                 .ei_type        = type,
1114                 .ei_mode        = mode,
1115                 .ei_cb_bl       = mgc_blocking_ast,
1116                 .ei_cb_cp       = ldlm_completion_ast,
1117                 .ei_cb_gl       = glimpse_callback,
1118         };
1119         struct ptlrpc_request *req;
1120         int short_limit = cld_is_sptlrpc(cld);
1121         int rc;
1122         ENTRY;
1123
1124         CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1125                cld->cld_resid.name[0]);
1126
1127         /* We need a callback for every lockholder, so don't try to
1128            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1129         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1130                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1131                                         LDLM_ENQUEUE);
1132         if (req == NULL)
1133                 RETURN(-ENOMEM);
1134
1135         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1136         ptlrpc_request_set_replen(req);
1137
1138         /* check if this is server or client */
1139         if (cld->cld_cfg.cfg_sb) {
1140                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1141                 if (lsi && IS_SERVER(lsi))
1142                         short_limit = 1;
1143         }
1144         /* Limit how long we will wait for the enqueue to complete */
1145         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1146         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1147                               NULL, 0, LVB_T_NONE, lockh, 0);
1148         /* A failed enqueue should still call the mgc_blocking_ast,
1149            where it will be requeued if needed ("grant failed"). */
1150         ptlrpc_req_finished(req);
1151         RETURN(rc);
1152 }
1153
1154 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1155                       struct lustre_handle *lockh)
1156 {
1157         ENTRY;
1158
1159         ldlm_lock_decref(lockh, mode);
1160
1161         RETURN(0);
1162 }
1163
1164 static void mgc_notify_active(struct obd_device *unused)
1165 {
1166         /* wakeup mgc_requeue_thread to requeue mgc lock */
1167         spin_lock(&config_list_lock);
1168         rq_state |= RQ_NOW;
1169         spin_unlock(&config_list_lock);
1170         wake_up(&rq_waitq);
1171
1172         /* TODO: Help the MGS rebuild nidtbl. -jay */
1173 }
1174
1175 /* Send target_reg message to MGS */
1176 static int mgc_target_register(struct obd_export *exp,
1177                                struct mgs_target_info *mti)
1178 {
1179         struct ptlrpc_request  *req;
1180         struct mgs_target_info *req_mti, *rep_mti;
1181         int                     rc;
1182         ENTRY;
1183
1184         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1185                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1186                                         MGS_TARGET_REG);
1187         if (req == NULL)
1188                 RETURN(-ENOMEM);
1189
1190         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1191         if (!req_mti) {
1192                 ptlrpc_req_finished(req);
1193                 RETURN(-ENOMEM);
1194         }
1195
1196         memcpy(req_mti, mti, sizeof(*req_mti));
1197         ptlrpc_request_set_replen(req);
1198         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1199         /* Limit how long we will wait for the enqueue to complete */
1200         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1201
1202         /* if the target needs to regenerate the config log in MGS, it's better
1203          * to use some longer limit to let MGC have time to change connection to
1204          * another MGS (or try again with the same MGS) for the target (server)
1205          * will fail and exit if the request expired due to delay limit. */
1206         if (mti->mti_flags & (LDD_F_UPDATE | LDD_F_NEED_INDEX))
1207                 req->rq_delay_limit = MGC_TARGET_REG_LIMIT_MAX;
1208
1209         rc = ptlrpc_queue_wait(req);
1210         if (!rc) {
1211                 rep_mti = req_capsule_server_get(&req->rq_pill,
1212                                                  &RMF_MGS_TARGET_INFO);
1213                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1214                 CDEBUG(D_MGC, "register %s got index = %d\n",
1215                        mti->mti_svname, mti->mti_stripe_index);
1216         }
1217         ptlrpc_req_finished(req);
1218
1219         RETURN(rc);
1220 }
1221
1222 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1223                               u32 keylen, void *key,
1224                               u32 vallen, void *val,
1225                               struct ptlrpc_request_set *set)
1226 {
1227         int rc = -EINVAL;
1228         ENTRY;
1229
1230         /* Turn off initial_recov after we try all backup servers once */
1231         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1232                 struct obd_import *imp = class_exp2cliimp(exp);
1233                 int value;
1234                 if (vallen != sizeof(int))
1235                         RETURN(-EINVAL);
1236                 value = *(int *)val;
1237                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1238                        imp->imp_obd->obd_name, value,
1239                        imp->imp_deactive, imp->imp_invalid,
1240                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1241                        ptlrpc_import_state_name(imp->imp_state));
1242                 /* Resurrect the import immediately if
1243                  * 1. we previously got disconnected,
1244                  * 2. value > 1 (at the same node with MGS)
1245                  * */
1246                 if (imp->imp_state == LUSTRE_IMP_DISCON || value > 1)
1247                         ptlrpc_reconnect_import(imp);
1248
1249                 RETURN(0);
1250         }
1251
1252         /* FIXME move this to mgc_process_config */
1253         if (KEY_IS(KEY_REGISTER_TARGET)) {
1254                 struct mgs_target_info *mti;
1255                 if (vallen != sizeof(struct mgs_target_info))
1256                         RETURN(-EINVAL);
1257                 mti = (struct mgs_target_info *)val;
1258                 CDEBUG(D_MGC, "register_target %s %#x\n",
1259                        mti->mti_svname, mti->mti_flags);
1260                 rc =  mgc_target_register(exp, mti);
1261                 RETURN(rc);
1262         }
1263         if (KEY_IS(KEY_SET_FS)) {
1264                 struct super_block *sb = (struct super_block *)val;
1265
1266                 if (vallen != sizeof(struct super_block))
1267                         RETURN(-EINVAL);
1268
1269                 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1270                 RETURN(rc);
1271         }
1272         if (KEY_IS(KEY_CLEAR_FS)) {
1273                 if (vallen != 0)
1274                         RETURN(-EINVAL);
1275                 rc = mgc_fs_cleanup(env, exp->exp_obd);
1276                 RETURN(rc);
1277         }
1278 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1279         if (KEY_IS(KEY_SET_INFO)) {
1280                 struct mgs_send_param *msp;
1281
1282                 msp = (struct mgs_send_param *)val;
1283                 rc =  mgc_set_mgs_param(exp, msp);
1284                 RETURN(rc);
1285         }
1286 #endif
1287         if (KEY_IS(KEY_MGSSEC)) {
1288                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1289                 struct sptlrpc_flavor  flvr;
1290
1291                 /*
1292                  * empty string means using current flavor, if which haven't
1293                  * been set yet, set it as null.
1294                  *
1295                  * if flavor has been set previously, check the asking flavor
1296                  * must match the existing one.
1297                  */
1298                 if (vallen == 0) {
1299                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1300                                 RETURN(0);
1301                         val = "null";
1302                         vallen = 4;
1303                 }
1304
1305                 rc = sptlrpc_parse_flavor(val, &flvr);
1306                 if (rc) {
1307                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1308                                (char *) val);
1309                         RETURN(rc);
1310                 }
1311
1312                 /*
1313                  * caller already hold a mutex
1314                  */
1315                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1316                         cli->cl_flvr_mgc = flvr;
1317                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1318                                   sizeof(flvr)) != 0) {
1319                         char    str[20];
1320
1321                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1322                                             str, sizeof(str));
1323                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1324                                        "currently %s is in use\n",
1325                                        (char *) val, str);
1326                         rc = -EPERM;
1327                 }
1328                 RETURN(rc);
1329         }
1330
1331         RETURN(rc);
1332 }
1333
1334 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1335                         __u32 keylen, void *key, __u32 *vallen, void *val)
1336 {
1337         int rc = -EINVAL;
1338
1339         if (KEY_IS(KEY_CONN_DATA)) {
1340                 struct obd_import *imp = class_exp2cliimp(exp);
1341                 struct obd_connect_data *data = val;
1342
1343                 if (*vallen == sizeof(*data)) {
1344                         *data = imp->imp_connect_data;
1345                         rc = 0;
1346                 }
1347         }
1348
1349         return rc;
1350 }
1351
1352 static int mgc_import_event(struct obd_device *obd,
1353                             struct obd_import *imp,
1354                             enum obd_import_event event)
1355 {
1356         int rc = 0;
1357
1358         LASSERT(imp->imp_obd == obd);
1359         CDEBUG(D_MGC, "import event %#x\n", event);
1360
1361         switch (event) {
1362         case IMP_EVENT_DISCON:
1363                 /* MGC imports should not wait for recovery */
1364                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1365                         ptlrpc_pinger_ir_down();
1366                 break;
1367         case IMP_EVENT_INACTIVE:
1368                 break;
1369         case IMP_EVENT_INVALIDATE: {
1370                 struct ldlm_namespace *ns = obd->obd_namespace;
1371                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1372                 break;
1373         }
1374         case IMP_EVENT_ACTIVE:
1375                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1376                 /* Clearing obd_no_recov allows us to continue pinging */
1377                 obd->obd_no_recov = 0;
1378                 mgc_notify_active(obd);
1379                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1380                         ptlrpc_pinger_ir_up();
1381                 break;
1382         case IMP_EVENT_OCD:
1383                 break;
1384         case IMP_EVENT_DEACTIVATE:
1385         case IMP_EVENT_ACTIVATE:
1386                 break;
1387         default:
1388                 CERROR("Unknown import event %#x\n", event);
1389                 LBUG();
1390         }
1391         RETURN(rc);
1392 }
1393
1394 enum {
1395         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1396         CONFIG_READ_NRPAGES      = 4
1397 };
1398
1399 static int mgc_apply_recover_logs(struct obd_device *mgc,
1400                                   struct config_llog_data *cld,
1401                                   __u64 max_version,
1402                                   void *data, int datalen, bool mne_swab)
1403 {
1404         struct config_llog_instance *cfg = &cld->cld_cfg;
1405         struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1406         struct mgs_nidtbl_entry *entry;
1407         struct lustre_cfg *lcfg;
1408         struct lustre_cfg_bufs bufs;
1409         u64 prev_version = 0;
1410         char *inst;
1411         char *buf;
1412         int bufsz;
1413         int pos = 0;
1414         int rc  = 0;
1415         int off = 0;
1416
1417         ENTRY;
1418         LASSERT(cfg->cfg_instance != 0);
1419         LASSERT(ll_get_cfg_instance(cfg->cfg_sb) == cfg->cfg_instance);
1420
1421         OBD_ALLOC(inst, PAGE_SIZE);
1422         if (inst == NULL)
1423                 RETURN(-ENOMEM);
1424
1425         if (!IS_SERVER(lsi)) {
1426                 pos = snprintf(inst, PAGE_SIZE, "%016lx", cfg->cfg_instance);
1427                 if (pos >= PAGE_SIZE) {
1428                         OBD_FREE(inst, PAGE_SIZE);
1429                         return -E2BIG;
1430                 }
1431 #ifdef HAVE_SERVER_SUPPORT
1432         } else {
1433                 LASSERT(IS_MDT(lsi));
1434                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1435                                         PAGE_SIZE);
1436                 if (rc) {
1437                         OBD_FREE(inst, PAGE_SIZE);
1438                         RETURN(-EINVAL);
1439                 }
1440                 pos = strlen(inst);
1441 #endif /* HAVE_SERVER_SUPPORT */
1442         }
1443
1444         ++pos;
1445         buf   = inst + pos;
1446         bufsz = PAGE_SIZE - pos;
1447
1448         while (datalen > 0) {
1449                 int   entry_len = sizeof(*entry);
1450                 int   is_ost, i;
1451                 struct obd_device *obd;
1452                 char *obdname;
1453                 char *cname;
1454                 char *params;
1455                 char *uuid;
1456
1457                 rc = -EINVAL;
1458                 if (datalen < sizeof(*entry))
1459                         break;
1460
1461                 entry = (typeof(entry))(data + off);
1462
1463                 /* sanity check */
1464                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1465                         break;
1466                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1467                         break;
1468                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1469                         break;
1470
1471                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1472                 if (datalen < entry_len) /* must have entry_len at least */
1473                         break;
1474
1475                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1476                 if (mne_swab)
1477                         lustre_swab_mgs_nidtbl_entry(entry);
1478                 if (entry->mne_length > PAGE_SIZE) {
1479                         CERROR("MNE too large (%u)\n", entry->mne_length);
1480                         break;
1481                 }
1482
1483                 if (entry->mne_length < entry_len)
1484                         break;
1485
1486                 off     += entry->mne_length;
1487                 datalen -= entry->mne_length;
1488                 if (datalen < 0)
1489                         break;
1490
1491                 if (entry->mne_version > max_version) {
1492                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1493                                entry->mne_version, max_version);
1494                         break;
1495                 }
1496
1497                 if (prev_version >= entry->mne_version) {
1498                         CERROR("index unsorted, prev %lld, now %lld\n",
1499                                prev_version, entry->mne_version);
1500                         break;
1501                 }
1502                 prev_version = entry->mne_version;
1503
1504                 /*
1505                  * Write a string with format "nid::instance" to
1506                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1507                  */
1508
1509                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1510                 memset(buf, 0, bufsz);
1511                 obdname = buf;
1512                 pos = 0;
1513
1514                 /* lustre-OST0001-osc-<instance #> */
1515                 strcpy(obdname, cld->cld_logname);
1516                 cname = strrchr(obdname, '-');
1517                 if (cname == NULL) {
1518                         CERROR("mgc %s: invalid logname %s\n",
1519                                mgc->obd_name, obdname);
1520                         break;
1521                 }
1522
1523                 pos = cname - obdname;
1524                 obdname[pos] = 0;
1525                 pos += sprintf(obdname + pos, "-%s%04x",
1526                                   is_ost ? "OST" : "MDT", entry->mne_index);
1527
1528                 cname = is_ost ? "osc" : "mdc",
1529                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1530                 lustre_cfg_bufs_reset(&bufs, obdname);
1531
1532                 /* find the obd by obdname */
1533                 obd = class_name2obd(obdname);
1534                 if (obd == NULL) {
1535                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1536                                mgc->obd_name, obdname);
1537                         rc = 0;
1538                         /* this is a safe race, when the ost is starting up...*/
1539                         continue;
1540                 }
1541
1542                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1543                 ++pos;
1544                 params = buf + pos;
1545                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1546                 uuid = buf + pos;
1547
1548                 down_read(&obd->u.cli.cl_sem);
1549                 if (obd->u.cli.cl_import == NULL) {
1550                         /* client does not connect to the OST yet */
1551                         up_read(&obd->u.cli.cl_sem);
1552                         rc = 0;
1553                         continue;
1554                 }
1555
1556                 /* iterate all nids to find one */
1557                 /* find uuid by nid */
1558                 rc = -ENOENT;
1559                 for (i = 0; i < entry->mne_nid_count; i++) {
1560                         rc = client_import_find_conn(obd->u.cli.cl_import,
1561                                                      entry->u.nids[i],
1562                                                      (struct obd_uuid *)uuid);
1563                         if (rc == 0)
1564                                 break;
1565                 }
1566
1567                 up_read(&obd->u.cli.cl_sem);
1568                 if (rc < 0) {
1569                         CERROR("mgc: cannot find uuid by nid %s\n",
1570                                libcfs_nid2str(entry->u.nids[0]));
1571                         break;
1572                 }
1573
1574                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1575                        uuid, libcfs_nid2str(entry->u.nids[0]));
1576
1577                 pos += strlen(uuid);
1578                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1579                 LASSERT(pos < bufsz);
1580
1581                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1582
1583                 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1584                                                bufs.lcfg_buflen));
1585                 if (!lcfg) {
1586                         rc = -ENOMEM;
1587                         break;
1588                 }
1589                 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1590
1591                 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1592                        prev_version, max_version, obdname, params);
1593
1594                 rc = class_process_config(lcfg);
1595                 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1596                                               lcfg->lcfg_buflens));
1597                 if (rc)
1598                         CDEBUG(D_INFO, "process config for %s error %d\n",
1599                                obdname, rc);
1600
1601                 /* continue, even one with error */
1602         }
1603
1604         OBD_FREE(inst, PAGE_SIZE);
1605         RETURN(rc);
1606 }
1607
1608 /**
1609  * This function is called if this client was notified for target restarting
1610  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1611  * nodemap logs.
1612  */
1613 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1614                                            struct config_llog_data *cld)
1615 {
1616         struct ptlrpc_connection *mgc_conn;
1617         struct ptlrpc_request *req = NULL;
1618         struct config_llog_instance *cfg = &cld->cld_cfg;
1619         struct mgs_config_body *body;
1620         struct mgs_config_res *res;
1621         struct nodemap_config *new_config = NULL;
1622         struct lu_nodemap *recent_nodemap = NULL;
1623         struct ptlrpc_bulk_desc *desc;
1624         struct page **pages = NULL;
1625         __u64 config_read_offset = 0;
1626         __u8 nodemap_cur_pass = 0;
1627         int nrpages = 0;
1628         bool eof = true;
1629         bool mne_swab = false;
1630         int i;
1631         int ealen;
1632         int rc;
1633         ENTRY;
1634
1635         mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1636
1637         /* don't need to get local config */
1638         if (cld_is_nodemap(cld) &&
1639             (LNET_NETTYP(LNET_NIDNET(mgc_conn->c_peer.nid)) == LOLND))
1640                 GOTO(out, rc = 0);
1641
1642         /* allocate buffer for bulk transfer.
1643          * if this is the first time for this mgs to read logs,
1644          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1645          * once; otherwise, it only reads increment of logs, this should be
1646          * small and CONFIG_READ_NRPAGES will be used.
1647          */
1648         nrpages = CONFIG_READ_NRPAGES;
1649         if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1650                 nrpages = CONFIG_READ_NRPAGES_INIT;
1651
1652         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1653         if (pages == NULL)
1654                 GOTO(out, rc = -ENOMEM);
1655
1656         for (i = 0; i < nrpages; i++) {
1657                 pages[i] = alloc_page(GFP_KERNEL);
1658                 if (pages[i] == NULL)
1659                         GOTO(out, rc = -ENOMEM);
1660         }
1661
1662 again:
1663 #ifdef HAVE_SERVER_SUPPORT
1664         if (cld_is_nodemap(cld) && config_read_offset == 0) {
1665                 new_config = nodemap_config_alloc();
1666                 if (IS_ERR(new_config)) {
1667                         rc = PTR_ERR(new_config);
1668                         new_config = NULL;
1669                         GOTO(out, rc);
1670                 }
1671         }
1672 #endif
1673         LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1674         LASSERT(mutex_is_locked(&cld->cld_lock));
1675         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1676                                    &RQF_MGS_CONFIG_READ);
1677         if (req == NULL)
1678                 GOTO(out, rc = -ENOMEM);
1679
1680         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1681         if (rc)
1682                 GOTO(out, rc);
1683
1684         /* pack request */
1685         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1686         LASSERT(body != NULL);
1687         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1688         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1689             >= sizeof(body->mcb_name))
1690                 GOTO(out, rc = -E2BIG);
1691         if (cld_is_nodemap(cld))
1692                 body->mcb_offset = config_read_offset;
1693         else
1694                 body->mcb_offset = cfg->cfg_last_idx + 1;
1695         body->mcb_type   = cld->cld_type;
1696         body->mcb_bits   = PAGE_SHIFT;
1697         body->mcb_units  = nrpages;
1698         body->mcb_nm_cur_pass = nodemap_cur_pass;
1699
1700         /* allocate bulk transfer descriptor */
1701         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1702                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1703                                     MGS_BULK_PORTAL,
1704                                     &ptlrpc_bulk_kiov_pin_ops);
1705         if (desc == NULL)
1706                 GOTO(out, rc = -ENOMEM);
1707
1708         for (i = 0; i < nrpages; i++)
1709                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1710                                                  PAGE_SIZE);
1711
1712         ptlrpc_request_set_replen(req);
1713         rc = ptlrpc_queue_wait(req);
1714         if (rc)
1715                 GOTO(out, rc);
1716
1717         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1718         if (!res)
1719                 GOTO(out, rc = -EPROTO);
1720
1721         if (cld_is_nodemap(cld)) {
1722                 config_read_offset = res->mcr_offset;
1723                 eof = config_read_offset == II_END_OFF;
1724                 nodemap_cur_pass = res->mcr_nm_cur_pass;
1725         } else {
1726                 if (res->mcr_size < res->mcr_offset)
1727                         GOTO(out, rc = -EINVAL);
1728
1729                 /* always update the index even though it might have errors with
1730                  * handling the recover logs
1731                  */
1732                 cfg->cfg_last_idx = res->mcr_offset;
1733                 eof = res->mcr_offset == res->mcr_size;
1734
1735                 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1736                        res->mcr_offset, eof == false);
1737         }
1738
1739         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1740         if (ealen < 0)
1741                 GOTO(out, rc = ealen);
1742
1743         if (ealen > nrpages << PAGE_SHIFT)
1744                 GOTO(out, rc = -EINVAL);
1745
1746         if (ealen == 0) { /* no logs transferred */
1747 #ifdef HAVE_SERVER_SUPPORT
1748                 /* config changed since first read RPC */
1749                 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1750                         CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1751                         GOTO(out, rc = -EAGAIN);
1752                 }
1753 #endif
1754                 if (!eof)
1755                         rc = -EINVAL;
1756                 GOTO(out, rc);
1757         }
1758
1759         mne_swab = ptlrpc_rep_need_swab(req);
1760
1761         /* When a nodemap config is received, we build a new nodemap config,
1762          * with new nodemap structs. We keep track of the most recently added
1763          * nodemap since the config is read ordered by nodemap_id, and so it
1764          * is likely that the next record will be related. Because access to
1765          * the nodemaps is single threaded until the nodemap_config is active,
1766          * we don't need to reference count with recent_nodemap, though
1767          * recent_nodemap should be set to NULL when the nodemap_config
1768          * is either destroyed or set active.
1769          */
1770         for (i = 0; i < nrpages && ealen > 0; i++) {
1771                 int rc2;
1772                 union lu_page   *ptr;
1773
1774                 ptr = kmap(pages[i]);
1775                 if (cld_is_nodemap(cld))
1776                         rc2 = nodemap_process_idx_pages(new_config, ptr,
1777                                                        &recent_nodemap);
1778                 else
1779                         rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1780                                                      ptr,
1781                                                      min_t(int, ealen,
1782                                                            PAGE_SIZE),
1783                                                      mne_swab);
1784                 kunmap(pages[i]);
1785                 if (rc2 < 0) {
1786                         CWARN("%s: error processing %s log %s: rc = %d\n",
1787                               obd->obd_name,
1788                               cld_is_nodemap(cld) ? "nodemap" : "recovery",
1789                               cld->cld_logname,
1790                               rc2);
1791                         GOTO(out, rc = rc2);
1792                 }
1793
1794                 ealen -= PAGE_SIZE;
1795         }
1796
1797 out:
1798         if (req) {
1799                 ptlrpc_req_finished(req);
1800                 req = NULL;
1801         }
1802
1803         if (rc == 0 && !eof)
1804                 goto again;
1805
1806 #ifdef HAVE_SERVER_SUPPORT
1807         if (new_config != NULL) {
1808                 /* recent_nodemap cannot be used after set_active/dealloc */
1809                 if (rc == 0)
1810                         nodemap_config_set_active_mgc(new_config);
1811                 else
1812                         nodemap_config_dealloc(new_config);
1813         }
1814 #endif
1815
1816         if (pages) {
1817                 for (i = 0; i < nrpages; i++) {
1818                         if (pages[i] == NULL)
1819                                 break;
1820                         __free_page(pages[i]);
1821                 }
1822                 OBD_FREE(pages, sizeof(*pages) * nrpages);
1823         }
1824         return rc;
1825 }
1826
1827 static int mgc_barrier_glimpse_ast(struct ldlm_lock *lock, void *data)
1828 {
1829         struct config_llog_data *cld = lock->l_ast_data;
1830         int rc;
1831         ENTRY;
1832
1833         if (cld->cld_stopping)
1834                 RETURN(-ENODEV);
1835
1836         rc = barrier_handler(s2lsi(cld->cld_cfg.cfg_sb)->lsi_dt_dev,
1837                              (struct ptlrpc_request *)data);
1838
1839         RETURN(rc);
1840 }
1841
1842 /* Copy a remote log locally */
1843 static int mgc_llog_local_copy(const struct lu_env *env,
1844                                struct obd_device *obd,
1845                                struct llog_ctxt *rctxt,
1846                                struct llog_ctxt *lctxt, char *logname)
1847 {
1848         char    *temp_log;
1849         int      rc;
1850
1851         ENTRY;
1852
1853         /*
1854          * - copy it to backup using llog_backup()
1855          * - copy remote llog to logname using llog_backup()
1856          * - if failed then move bakup to logname again
1857          */
1858
1859         OBD_ALLOC(temp_log, strlen(logname) + 2);
1860         if (!temp_log)
1861                 RETURN(-ENOMEM);
1862         sprintf(temp_log, "%sT", logname);
1863
1864         /* make a copy of local llog at first */
1865         rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1866         if (rc < 0 && rc != -ENOENT)
1867                 GOTO(out, rc);
1868         /* copy remote llog to the local copy */
1869         rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1870         if (rc == -ENOENT) {
1871                 /* no remote llog, delete local one too */
1872                 llog_erase(env, lctxt, NULL, logname);
1873         } else if (rc < 0) {
1874                 /* error during backup, get local one back from the copy */
1875                 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1876 out:
1877                 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1878                        obd->obd_name, logname, rc);
1879         }
1880         llog_erase(env, lctxt, NULL, temp_log);
1881         OBD_FREE(temp_log, strlen(logname) + 2);
1882         return rc;
1883 }
1884
1885 /* local_only means it cannot get remote llogs */
1886 static int mgc_process_cfg_log(struct obd_device *mgc,
1887                                struct config_llog_data *cld, int local_only)
1888 {
1889         struct llog_ctxt        *ctxt, *lctxt = NULL;
1890         struct client_obd       *cli = &mgc->u.cli;
1891         struct lustre_sb_info   *lsi = NULL;
1892         int                      rc = 0;
1893         struct lu_env           *env;
1894
1895         ENTRY;
1896
1897         LASSERT(cld);
1898         LASSERT(mutex_is_locked(&cld->cld_lock));
1899
1900         if (cld->cld_cfg.cfg_sb)
1901                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1902
1903         OBD_ALLOC_PTR(env);
1904         if (env == NULL)
1905                 RETURN(-ENOMEM);
1906
1907         rc = lu_env_init(env, LCT_MG_THREAD);
1908         if (rc)
1909                 GOTO(out_free, rc);
1910
1911         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1912         LASSERT(ctxt);
1913
1914         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1915
1916         /* Copy the setup log locally if we can. Don't mess around if we're
1917          * running an MGS though (logs are already local). */
1918         if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1919             cli->cl_mgc_configs_dir != NULL &&
1920             lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1921             lsi->lsi_dt_dev) {
1922                 if (!local_only && !lsi->lsi_dt_dev->dd_rdonly)
1923                         /* Only try to copy log if we have the lock. */
1924                         rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1925                                                  cld->cld_logname);
1926                 if (local_only || rc) {
1927                         if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1928                             llog_is_empty(env, lctxt, cld->cld_logname)) {
1929                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1930                                                    "log %s and no local copy."
1931                                                    "\n", cld->cld_logname);
1932                                 GOTO(out_pop, rc = -ENOENT);
1933                         }
1934                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1935                                "copy for now, will try to update later.\n",
1936                                cld->cld_logname);
1937                         rc = 0;
1938                 }
1939                 /* Now, whether we copied or not, start using the local llog.
1940                  * If we failed to copy, we'll start using whatever the old
1941                  * log has. */
1942                 llog_ctxt_put(ctxt);
1943                 ctxt = lctxt;
1944                 lctxt = NULL;
1945         } else {
1946                 if (local_only) /* no local log at client side */
1947                         GOTO(out_pop, rc = -EIO);
1948         }
1949
1950         rc = -EAGAIN;
1951         if (lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1952             lsi->lsi_dt_dev->dd_rdonly) {
1953                 struct llog_ctxt *rctxt;
1954
1955                 /* Under readonly mode, we may have no local copy or local
1956                  * copy is incomplete, so try to use remote llog firstly. */
1957                 rctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1958                 LASSERT(rctxt);
1959
1960                 rc = class_config_parse_llog(env, rctxt, cld->cld_logname,
1961                                              &cld->cld_cfg);
1962                 llog_ctxt_put(rctxt);
1963         }
1964
1965         if (rc && rc != -ENOENT)
1966                 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1967                                              &cld->cld_cfg);
1968
1969         /*
1970          * update settings on existing OBDs.
1971          * the logname must be <fsname>-sptlrpc
1972          */
1973         if (rc == 0 && cld_is_sptlrpc(cld))
1974                 class_notify_sptlrpc_conf(cld->cld_logname,
1975                                           strlen(cld->cld_logname) -
1976                                           strlen("-sptlrpc"));
1977         EXIT;
1978
1979 out_pop:
1980         __llog_ctxt_put(env, ctxt);
1981         if (lctxt)
1982                 __llog_ctxt_put(env, lctxt);
1983
1984         lu_env_fini(env);
1985 out_free:
1986         OBD_FREE_PTR(env);
1987         return rc;
1988 }
1989
1990 static bool mgc_import_in_recovery(struct obd_import *imp)
1991 {
1992         bool in_recovery = true;
1993
1994         spin_lock(&imp->imp_lock);
1995         if (imp->imp_state == LUSTRE_IMP_FULL ||
1996             imp->imp_state == LUSTRE_IMP_CLOSED)
1997                 in_recovery = false;
1998         spin_unlock(&imp->imp_lock);
1999
2000         return in_recovery;
2001 }
2002
2003 /**
2004  * Get a configuration log from the MGS and process it.
2005  *
2006  * This function is called for both clients and servers to process the
2007  * configuration log from the MGS.  The MGC enqueues a DLM lock on the
2008  * log from the MGS, and if the lock gets revoked the MGC will be notified
2009  * by the lock cancellation callback that the config log has changed,
2010  * and will enqueue another MGS lock on it, and then continue processing
2011  * the new additions to the end of the log.
2012  *
2013  * Since the MGC import is not replayable, if the import is being evicted
2014  * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
2015  * the log until recovery is finished or the import is closed.
2016  *
2017  * Make a local copy of the log before parsing it if appropriate (non-MGS
2018  * server) so that the server can start even when the MGS is down.
2019  *
2020  * There shouldn't be multiple processes running process_log at once --
2021  * sounds like badness.  It actually might be fine, as long as they're not
2022  * trying to update from the same log simultaneously, in which case we
2023  * should use a per-log semaphore instead of cld_lock.
2024  *
2025  * \param[in] mgc       MGC device by which to fetch the configuration log
2026  * \param[in] cld       log processing state (stored in lock callback data)
2027  *
2028  * \retval              0 on success
2029  * \retval              negative errno on failure
2030  */
2031 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
2032 {
2033         struct lustre_handle lockh = { 0 };
2034         __u64 flags = LDLM_FL_NO_LRU;
2035         int rc = 0, rcl;
2036         bool retry = false;
2037         ENTRY;
2038
2039         LASSERT(cld != NULL);
2040
2041         /* I don't want multiple processes running process_log at once --
2042            sounds like badness.  It actually might be fine, as long as
2043            we're not trying to update from the same log
2044            simultaneously (in which case we should use a per-log sem.) */
2045 restart:
2046         mutex_lock(&cld->cld_lock);
2047         if (cld->cld_stopping) {
2048                 mutex_unlock(&cld->cld_lock);
2049                 RETURN(0);
2050         }
2051
2052         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2053
2054         CDEBUG(D_MGC, "Process log %s-%016lx from %d\n", cld->cld_logname,
2055                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2056
2057         /* Get the cfg lock on the llog */
2058         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2059                           LCK_CR, &flags,
2060                           cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
2061                           cld, 0, NULL, &lockh);
2062         if (rcl == 0) {
2063                 /* Get the cld, it will be released in mgc_blocking_ast. */
2064                 config_log_get(cld);
2065                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2066                 LASSERT(rc == 0);
2067         } else {
2068                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2069
2070                 if (rcl == -ESHUTDOWN &&
2071                     atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2072                         struct obd_import *imp;
2073                         struct l_wait_info lwi;
2074                         long timeout = cfs_time_seconds(obd_timeout);
2075
2076                         mutex_unlock(&cld->cld_lock);
2077                         imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2078
2079                         /* Let's force the pinger, and wait the import to be
2080                          * connected, note: since mgc import is non-replayable,
2081                          * and even the import state is disconnected, it does
2082                          * not mean the "recovery" is stopped, so we will keep
2083                          * waitting until timeout or the import state is
2084                          * FULL or closed */
2085                         ptlrpc_pinger_force(imp);
2086
2087                         lwi = LWI_TIMEOUT(timeout, NULL, NULL);
2088                         l_wait_event(imp->imp_recovery_waitq,
2089                                      !mgc_import_in_recovery(imp), &lwi);
2090
2091                         if (imp->imp_state == LUSTRE_IMP_FULL) {
2092                                 retry = true;
2093                                 goto restart;
2094                         } else {
2095                                 mutex_lock(&cld->cld_lock);
2096                                 /* unlock/lock mutex, so check stopping again */
2097                                 if (cld->cld_stopping) {
2098                                         mutex_unlock(&cld->cld_lock);
2099                                         RETURN(0);
2100                                 }
2101                                 spin_lock(&config_list_lock);
2102                                 cld->cld_lostlock = 1;
2103                                 spin_unlock(&config_list_lock);
2104                         }
2105                 } else {
2106                         /* mark cld_lostlock so that it will requeue
2107                          * after MGC becomes available. */
2108                         spin_lock(&config_list_lock);
2109                         cld->cld_lostlock = 1;
2110                         spin_unlock(&config_list_lock);
2111                 }
2112         }
2113
2114         if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2115                 if (!rcl)
2116                         rc = mgc_process_recover_nodemap_log(mgc, cld);
2117                 else if (cld_is_nodemap(cld))
2118                         rc = rcl;
2119
2120                 if (cld_is_recover(cld) && rc) {
2121                         if (!rcl) {
2122                                 CERROR("%s: recover log %s failed, not fatal: rc = %d\n",
2123                                        mgc->obd_name, cld->cld_logname, rc);
2124                                 spin_lock(&config_list_lock);
2125                                 cld->cld_lostlock = 1;
2126                                 spin_unlock(&config_list_lock);
2127                         }
2128                         rc = 0; /* this is not a fatal error for recover log */
2129                 }
2130         } else if (!cld_is_barrier(cld)) {
2131                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2132         }
2133
2134         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2135                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2136
2137         mutex_unlock(&cld->cld_lock);
2138
2139         /* Now drop the lock so MGS can revoke it */
2140         if (!rcl) {
2141                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2142                 if (rcl)
2143                         CERROR("Can't drop cfg lock: %d\n", rcl);
2144         }
2145
2146         /* requeue nodemap lock immediately if transfer was interrupted */
2147         if (cld_is_nodemap(cld) && rc == -EAGAIN) {
2148                 mgc_requeue_add(cld);
2149                 rc = 0;
2150         }
2151
2152         RETURN(rc);
2153 }
2154
2155
2156 /** Called from lustre_process_log.
2157  * LCFG_LOG_START gets the config log from the MGS, processes it to start
2158  * any services, and adds it to the list logs to watch (follow).
2159  */
2160 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2161 {
2162         struct lustre_cfg *lcfg = buf;
2163         struct config_llog_instance *cfg = NULL;
2164         char *logname;
2165         int rc = 0;
2166         ENTRY;
2167
2168         switch(lcfg->lcfg_command) {
2169         case LCFG_LOV_ADD_OBD: {
2170                 /* Overloading this cfg command: register a new target */
2171                 struct mgs_target_info *mti;
2172
2173                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2174                     sizeof(struct mgs_target_info))
2175                         GOTO(out, rc = -EINVAL);
2176
2177                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2178                 CDEBUG(D_MGC, "add_target %s %#x\n",
2179                        mti->mti_svname, mti->mti_flags);
2180                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2181                 break;
2182         }
2183         case LCFG_LOV_DEL_OBD:
2184                 /* Unregister has no meaning at the moment. */
2185                 CERROR("lov_del_obd unimplemented\n");
2186                 rc = -ENOSYS;
2187                 break;
2188         case LCFG_SPTLRPC_CONF: {
2189                 rc = sptlrpc_process_config(lcfg);
2190                 break;
2191         }
2192         case LCFG_LOG_START: {
2193                 struct config_llog_data *cld;
2194                 struct super_block *sb;
2195
2196                 logname = lustre_cfg_string(lcfg, 1);
2197                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2198                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2199
2200                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2201                        cfg->cfg_last_idx);
2202
2203                 /* We're only called through here on the initial mount */
2204                 cld = config_log_add(obd, logname, cfg, sb);
2205                 if (IS_ERR(cld)) {
2206                         rc = PTR_ERR(cld);
2207                         break;
2208                 }
2209
2210                 rc = mgc_process_log(obd, cld);
2211                 if (rc == 0 && cld->cld_recover != NULL) {
2212                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2213                                          imp_connect_data, IMP_RECOV)) {
2214                                 rc = mgc_process_log(obd, cld->cld_recover);
2215                         } else {
2216                                 struct config_llog_data *cir;
2217
2218                                 mutex_lock(&cld->cld_lock);
2219                                 cir = cld->cld_recover;
2220                                 cld->cld_recover = NULL;
2221                                 mutex_unlock(&cld->cld_lock);
2222                                 config_log_put(cir);
2223                         }
2224
2225                         if (rc)
2226                                 CERROR("Cannot process recover llog %d\n", rc);
2227                 }
2228
2229                 if (rc == 0 && cld->cld_params != NULL) {
2230                         rc = mgc_process_log(obd, cld->cld_params);
2231                         if (rc == -ENOENT) {
2232                                 CDEBUG(D_MGC, "There is no params "
2233                                               "config file yet\n");
2234                                 rc = 0;
2235                         }
2236                         /* params log is optional */
2237                         if (rc)
2238                                 CERROR("%s: can't process params llog: rc = %d\n",
2239                                        obd->obd_name, rc);
2240                 }
2241
2242                 break;
2243         }
2244         case LCFG_LOG_END: {
2245                 logname = lustre_cfg_string(lcfg, 1);
2246
2247                 if (lcfg->lcfg_bufcount >= 2)
2248                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
2249                                 lcfg, 2);
2250                 rc = config_log_end(logname, cfg);
2251                 break;
2252         }
2253         default: {
2254                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2255                 GOTO(out, rc = -EINVAL);
2256
2257         }
2258         }
2259 out:
2260         RETURN(rc);
2261 }
2262
2263 static struct obd_ops mgc_obd_ops = {
2264         .o_owner        = THIS_MODULE,
2265         .o_setup        = mgc_setup,
2266         .o_precleanup   = mgc_precleanup,
2267         .o_cleanup      = mgc_cleanup,
2268         .o_add_conn     = client_import_add_conn,
2269         .o_del_conn     = client_import_del_conn,
2270         .o_connect      = client_connect_import,
2271         .o_disconnect   = client_disconnect_export,
2272         .o_set_info_async = mgc_set_info_async,
2273         .o_get_info       = mgc_get_info,
2274         .o_import_event = mgc_import_event,
2275         .o_process_config = mgc_process_config,
2276 };
2277
2278 static int __init mgc_init(void)
2279 {
2280         return class_register_type(&mgc_obd_ops, NULL, false, NULL,
2281                                    LUSTRE_MGC_NAME, NULL);
2282 }
2283
2284 static void __exit mgc_exit(void)
2285 {
2286         class_unregister_type(LUSTRE_MGC_NAME);
2287 }
2288
2289 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2290 MODULE_DESCRIPTION("Lustre Management Client");
2291 MODULE_VERSION(LUSTRE_VERSION_STRING);
2292 MODULE_LICENSE("GPL");
2293
2294 module_init(mgc_init);
2295 module_exit(mgc_exit);