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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/mgc/mgc_request.c
  37  *
  38  * Author: Nathan Rutman <nathan@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_MGC
  42 #define D_MGC D_CONFIG /*|D_WARNING*/
  43
  44 #include <linux/module.h>
  45 #include <linux/kthread.h>
  46
  47 #include <dt_object.h>
  48 #include <lprocfs_status.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_disk.h>
  51 #include <lustre_log.h>
  52 #include <lustre_swab.h>
  53 #include <obd_class.h>
  54
  55 #include "mgc_internal.h"
  56
  57 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
  58                           int type)
  59 {
  60         __u64 resname = 0;
  61
  62         if (len > sizeof(resname)) {
  63                 CERROR("name too long: %s\n", name);
  64                 return -EINVAL;
  65         }
  66         if (len <= 0) {
  67                 CERROR("missing name: %s\n", name);
  68                 return -EINVAL;
  69         }
  70         memcpy(&resname, name, len);
  71
  72         /* Always use the same endianness for the resid */
  73         memset(res_id, 0, sizeof(*res_id));
  74         res_id->name[0] = cpu_to_le64(resname);
  75         /* XXX: unfortunately, sptlprc and config llog share one lock */
  76         switch(type) {
  77         case CONFIG_T_CONFIG:
  78         case CONFIG_T_SPTLRPC:
  79                 resname = 0;
  80                 break;
  81         case CONFIG_T_RECOVER:
  82         case CONFIG_T_PARAMS:
  83                 resname = type;
  84                 break;
  85         default:
  86                 LBUG();
  87         }
  88         res_id->name[1] = cpu_to_le64(resname);
  89         CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
  90                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  91         return 0;
  92 }
  93
  94 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
  95 {
  96         /* fsname is at most 8 chars long, maybe contain "-".
  97          * e.g. "lustre", "SUN-000" */
  98         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
  99 }
 100 EXPORT_SYMBOL(mgc_fsname2resid);
 101
 102 static int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id,
 103                              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
 119 /********************** config llog list **********************/
 120 static struct list_head config_llog_list = LIST_HEAD_INIT(config_llog_list);
 121 static DEFINE_SPINLOCK(config_list_lock);
 122
 123 /* Take a reference to a config log */
 124 static int config_log_get(struct config_llog_data *cld)
 125 {
 126         ENTRY;
 127         atomic_inc(&cld->cld_refcount);
 128         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 129                 atomic_read(&cld->cld_refcount));
 130         RETURN(0);
 131 }
 132
 133 /* Drop a reference to a config log.  When no longer referenced,
 134    we can free the config log data */
 135 static void config_log_put(struct config_llog_data *cld)
 136 {
 137         ENTRY;
 138
 139         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 140                 atomic_read(&cld->cld_refcount));
 141         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 142
 143         /* spinlock to make sure no item with 0 refcount in the list */
 144         if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 145                 list_del(&cld->cld_list_chain);
 146                 spin_unlock(&config_list_lock);
 147
 148                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 149
 150                 if (cld->cld_recover)
 151                         config_log_put(cld->cld_recover);
 152                 if (cld->cld_sptlrpc)
 153                         config_log_put(cld->cld_sptlrpc);
 154                 if (cld->cld_params)
 155                         config_log_put(cld->cld_params);
 156                 if (cld_is_sptlrpc(cld))
 157                         sptlrpc_conf_log_stop(cld->cld_logname);
 158
 159                 class_export_put(cld->cld_mgcexp);
 160                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 161         }
 162
 163         EXIT;
 164 }
 165
 166 /* Find a config log by name */
 167 static
 168 struct config_llog_data *config_log_find(char *logname,
 169                                          struct config_llog_instance *cfg)
 170 {
 171         struct config_llog_data *cld;
 172         struct config_llog_data *found = NULL;
 173         void *                   instance;
 174         ENTRY;
 175
 176         LASSERT(logname != NULL);
 177
 178         instance = cfg ? cfg->cfg_instance : NULL;
 179         spin_lock(&config_list_lock);
 180         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 181                 /* check if instance equals */
 182                 if (instance != cld->cld_cfg.cfg_instance)
 183                         continue;
 184
 185                 /* instance may be NULL, should check name */
 186                 if (strcmp(logname, cld->cld_logname) == 0) {
 187                         found = cld;
 188                         break;
 189                 }
 190         }
 191         if (found) {
 192                 atomic_inc(&found->cld_refcount);
 193                 LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0);
 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         ENTRY;
 209
 210         CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
 211                cfg ? cfg->cfg_instance : NULL);
 212
 213         OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
 214         if (!cld)
 215                 RETURN(ERR_PTR(-ENOMEM));
 216
 217         strcpy(cld->cld_logname, logname);
 218         if (cfg)
 219                 cld->cld_cfg = *cfg;
 220         else
 221                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 222         mutex_init(&cld->cld_lock);
 223         cld->cld_cfg.cfg_last_idx = 0;
 224         cld->cld_cfg.cfg_flags = 0;
 225         cld->cld_cfg.cfg_sb = sb;
 226         cld->cld_type = type;
 227         atomic_set(&cld->cld_refcount, 1);
 228
 229         /* Keep the mgc around until we are done */
 230         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 231
 232         if (cld_is_sptlrpc(cld)) {
 233                 sptlrpc_conf_log_start(logname);
 234                 cld->cld_cfg.cfg_obdname = obd->obd_name;
 235         }
 236
 237         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 238
 239         spin_lock(&config_list_lock);
 240         list_add(&cld->cld_list_chain, &config_llog_list);
 241         spin_unlock(&config_list_lock);
 242
 243         if (rc) {
 244                 config_log_put(cld);
 245                 RETURN(ERR_PTR(rc));
 246         }
 247
 248         if (cld_is_sptlrpc(cld)) {
 249                 rc = mgc_process_log(obd, cld);
 250                 if (rc && rc != -ENOENT)
 251                         CERROR("failed processing sptlrpc log: %d\n", rc);
 252         }
 253
 254         RETURN(cld);
 255 }
 256
 257 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
 258         char *fsname,
 259         struct config_llog_instance *cfg,
 260         struct super_block *sb)
 261 {
 262         struct config_llog_instance lcfg = *cfg;
 263         struct lustre_sb_info *lsi = s2lsi(sb);
 264         struct config_llog_data *cld;
 265         char logname[32];
 266
 267         if (IS_OST(lsi))
 268                 return NULL;
 269
 270         /* for osp-on-ost, see lustre_start_osp() */
 271         if (IS_MDT(lsi) && lcfg.cfg_instance)
 272                 return NULL;
 273
 274         /* we have to use different llog for clients and mdts for cmd
 275          * where only clients are notified if one of cmd server restarts */
 276         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 277         strcpy(logname, fsname);
 278         if (IS_SERVER(lsi)) { /* mdt */
 279                 LASSERT(lcfg.cfg_instance == NULL);
 280                 lcfg.cfg_instance = sb;
 281                 strcat(logname, "-mdtir");
 282         } else {
 283                 LASSERT(lcfg.cfg_instance != NULL);
 284                 strcat(logname, "-cliir");
 285         }
 286
 287         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
 288         return cld;
 289 }
 290
 291 static struct config_llog_data *config_params_log_add(struct obd_device *obd,
 292         struct config_llog_instance *cfg, struct super_block *sb)
 293 {
 294         struct config_llog_instance     lcfg = *cfg;
 295         struct config_llog_data         *cld;
 296
 297         lcfg.cfg_instance = sb;
 298
 299         cld = do_config_log_add(obd, PARAMS_FILENAME, CONFIG_T_PARAMS,
 300                                 &lcfg, sb);
 301
 302         return cld;
 303 }
 304
 305 /** Add this log to the list of active logs watched by an MGC.
 306  * Active means we're watching for updates.
 307  * We have one active log per "mount" - client instance or servername.
 308  * Each instance may be at a different point in the log.
 309  */
 310 static int config_log_add(struct obd_device *obd, char *logname,
 311                           struct config_llog_instance *cfg,
 312                           struct super_block *sb)
 313 {
 314         struct lustre_sb_info   *lsi = s2lsi(sb);
 315         struct config_llog_data *cld;
 316         struct config_llog_data *sptlrpc_cld;
 317         struct config_llog_data *params_cld;
 318         char                    seclogname[32];
 319         char                    *ptr;
 320         int                     rc;
 321         ENTRY;
 322
 323         CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
 324
 325         /*
 326          * for each regular log, the depended sptlrpc log name is
 327          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 328          */
 329         ptr = strrchr(logname, '-');
 330         if (ptr == NULL || ptr - logname > 8) {
 331                 CERROR("logname %s is too long\n", logname);
 332                 RETURN(-EINVAL);
 333         }
 334
 335         memcpy(seclogname, logname, ptr - logname);
 336         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 337
 338         sptlrpc_cld = config_log_find(seclogname, NULL);
 339         if (sptlrpc_cld == NULL) {
 340                 sptlrpc_cld = do_config_log_add(obd, seclogname,
 341                                                 CONFIG_T_SPTLRPC, NULL, NULL);
 342                 if (IS_ERR(sptlrpc_cld)) {
 343                         CERROR("can't create sptlrpc log: %s\n", seclogname);
 344                         GOTO(out_err, rc = PTR_ERR(sptlrpc_cld));
 345                 }
 346         }
 347         params_cld = config_params_log_add(obd, cfg, sb);
 348         if (IS_ERR(params_cld)) {
 349                 rc = PTR_ERR(params_cld);
 350                 CERROR("%s: can't create params log: rc = %d\n",
 351                        obd->obd_name, rc);
 352                 GOTO(out_err1, rc);
 353         }
 354
 355         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 356         if (IS_ERR(cld)) {
 357                 CERROR("can't create log: %s\n", logname);
 358                 GOTO(out_err2, rc = PTR_ERR(cld));
 359         }
 360
 361         cld->cld_sptlrpc = sptlrpc_cld;
 362         cld->cld_params = params_cld;
 363
 364         LASSERT(lsi->lsi_lmd);
 365         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
 366                 struct config_llog_data *recover_cld;
 367                 ptr = strrchr(seclogname, '-');
 368                 if (ptr != NULL) {
 369                         *ptr = 0;
 370                 }
 371                 else {
 372                         CERROR("%s: sptlrpc log name not correct, %s: "
 373                                "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
 374                         config_log_put(cld);
 375                         RETURN(-EINVAL);
 376                 }
 377                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 378                 if (IS_ERR(recover_cld))
 379                         GOTO(out_err3, rc = PTR_ERR(recover_cld));
 380                 cld->cld_recover = recover_cld;
 381         }
 382
 383         RETURN(0);
 384
 385 out_err3:
 386         config_log_put(cld);
 387
 388 out_err2:
 389         config_log_put(params_cld);
 390
 391 out_err1:
 392         config_log_put(sptlrpc_cld);
 393
 394 out_err:
 395         RETURN(rc);
 396 }
 397
 398 DEFINE_MUTEX(llog_process_lock);
 399
 400 /** Stop watching for updates on this log.
 401  */
 402 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 403 {
 404         struct config_llog_data *cld;
 405         struct config_llog_data *cld_sptlrpc = NULL;
 406         struct config_llog_data *cld_params = NULL;
 407         struct config_llog_data *cld_recover = NULL;
 408         int rc = 0;
 409         ENTRY;
 410
 411         cld = config_log_find(logname, cfg);
 412         if (cld == NULL)
 413                 RETURN(-ENOENT);
 414
 415         mutex_lock(&cld->cld_lock);
 416         /*
 417          * if cld_stopping is set, it means we didn't start the log thus
 418          * not owning the start ref. this can happen after previous umount:
 419          * the cld still hanging there waiting for lock cancel, and we
 420          * remount again but failed in the middle and call log_end without
 421          * calling start_log.
 422          */
 423         if (unlikely(cld->cld_stopping)) {
 424                 mutex_unlock(&cld->cld_lock);
 425                 /* drop the ref from the find */
 426                 config_log_put(cld);
 427                 RETURN(rc);
 428         }
 429
 430         cld->cld_stopping = 1;
 431
 432         cld_recover = cld->cld_recover;
 433         cld->cld_recover = NULL;
 434         mutex_unlock(&cld->cld_lock);
 435
 436         if (cld_recover) {
 437                 mutex_lock(&cld_recover->cld_lock);
 438                 cld_recover->cld_stopping = 1;
 439                 mutex_unlock(&cld_recover->cld_lock);
 440                 config_log_put(cld_recover);
 441         }
 442
 443         spin_lock(&config_list_lock);
 444         cld_sptlrpc = cld->cld_sptlrpc;
 445         cld->cld_sptlrpc = NULL;
 446         cld_params = cld->cld_params;
 447         cld->cld_params = NULL;
 448         spin_unlock(&config_list_lock);
 449
 450         if (cld_sptlrpc)
 451                 config_log_put(cld_sptlrpc);
 452
 453         if (cld_params) {
 454                 mutex_lock(&cld_params->cld_lock);
 455                 cld_params->cld_stopping = 1;
 456                 mutex_unlock(&cld_params->cld_lock);
 457                 config_log_put(cld_params);
 458         }
 459
 460         /* drop the ref from the find */
 461         config_log_put(cld);
 462         /* drop the start ref */
 463         config_log_put(cld);
 464
 465         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 466                rc);
 467         RETURN(rc);
 468 }
 469
 470 #ifdef CONFIG_PROC_FS
 471 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 472 {
 473         struct obd_device       *obd = data;
 474         struct obd_import       *imp;
 475         struct obd_connect_data *ocd;
 476         struct config_llog_data *cld;
 477         ENTRY;
 478
 479         LASSERT(obd != NULL);
 480         LPROCFS_CLIMP_CHECK(obd);
 481         imp = obd->u.cli.cl_import;
 482         ocd = &imp->imp_connect_data;
 483
 484         seq_printf(m, "imperative_recovery: %s\n",
 485                    OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 486         seq_printf(m, "client_state:\n");
 487
 488         spin_lock(&config_list_lock);
 489         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 490                 if (cld->cld_recover == NULL)
 491                         continue;
 492                 seq_printf(m,  "    - { client: %s, nidtbl_version: %u }\n",
 493                            cld->cld_logname,
 494                            cld->cld_recover->cld_cfg.cfg_last_idx);
 495         }
 496         spin_unlock(&config_list_lock);
 497
 498         LPROCFS_CLIMP_EXIT(obd);
 499         RETURN(0);
 500 }
 501 #endif
 502
 503 /* reenqueue any lost locks */
 504 #define RQ_RUNNING      0x1
 505 #define RQ_NOW          0x2
 506 #define RQ_LATER        0x4
 507 #define RQ_STOP         0x8
 508 #define RQ_PRECLEANUP   0x10
 509 static int                    rq_state = 0;
 510 static wait_queue_head_t      rq_waitq;
 511 static DECLARE_COMPLETION(rq_exit);
 512 static DECLARE_COMPLETION(rq_start);
 513
 514 static void do_requeue(struct config_llog_data *cld)
 515 {
 516         int rc = 0;
 517         ENTRY;
 518
 519         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 520
 521         /* Do not run mgc_process_log on a disconnected export or an
 522          * export which is being disconnected. Take the client
 523          * semaphore to make the check non-racy. */
 524         down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 525         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 526                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 527                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 528                 if (rc && rc != -ENOENT)
 529                         CERROR("failed processing log: %d\n", rc);
 530         } else {
 531                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 532                        cld->cld_logname);
 533         }
 534         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 535
 536         EXIT;
 537 }
 538
 539 /* this timeout represents how many seconds MGC should wait before
 540  * requeue config and recover lock to the MGS. We need to randomize this
 541  * in order to not flood the MGS.
 542  */
 543 #define MGC_TIMEOUT_MIN_SECONDS   5
 544 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
 545
 546 static int mgc_requeue_thread(void *data)
 547 {
 548         int rc = 0;
 549         bool first = true;
 550         ENTRY;
 551
 552         CDEBUG(D_MGC, "Starting requeue thread\n");
 553
 554         /* Keep trying failed locks periodically */
 555         spin_lock(&config_list_lock);
 556         rq_state |= RQ_RUNNING;
 557         while (1) {
 558                 struct l_wait_info lwi;
 559                 struct config_llog_data *cld, *cld_prev;
 560                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
 561                 int stopped = !!(rq_state & RQ_STOP);
 562                 int to;
 563
 564                 /* Any new or requeued lostlocks will change the state */
 565                 rq_state &= ~(RQ_NOW | RQ_LATER);
 566                 spin_unlock(&config_list_lock);
 567
 568                 if (first) {
 569                         first = false;
 570                         complete(&rq_start);
 571                 }
 572
 573                 /* Always wait a few seconds to allow the server who
 574                    caused the lock revocation to finish its setup, plus some
 575                    random so everyone doesn't try to reconnect at once. */
 576                 to = msecs_to_jiffies(MGC_TIMEOUT_MIN_SECONDS * MSEC_PER_SEC);
 577                 /* rand is centi-seconds */
 578                 to += msecs_to_jiffies(rand * MSEC_PER_SEC / 100);
 579                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 580                 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
 581                              &lwi);
 582
 583                 /*
 584                  * iterate & processing through the list. for each cld, process
 585                  * its depending sptlrpc cld firstly (if any) and then itself.
 586                  *
 587                  * it's guaranteed any item in the list must have
 588                  * reference > 0; and if cld_lostlock is set, at
 589                  * least one reference is taken by the previous enqueue.
 590                  */
 591                 cld_prev = NULL;
 592
 593                 spin_lock(&config_list_lock);
 594                 rq_state &= ~RQ_PRECLEANUP;
 595                 list_for_each_entry(cld, &config_llog_list,
 596                                         cld_list_chain) {
 597                         if (!cld->cld_lostlock)
 598                                 continue;
 599
 600                         spin_unlock(&config_list_lock);
 601
 602                         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 603
 604                         /* Whether we enqueued again or not in mgc_process_log,
 605                          * we're done with the ref from the old enqueue */
 606                         if (cld_prev)
 607                                 config_log_put(cld_prev);
 608                         cld_prev = cld;
 609
 610                         cld->cld_lostlock = 0;
 611                         if (likely(!stopped))
 612                                 do_requeue(cld);
 613
 614                         spin_lock(&config_list_lock);
 615                 }
 616                 spin_unlock(&config_list_lock);
 617                 if (cld_prev)
 618                         config_log_put(cld_prev);
 619
 620                 /* break after scanning the list so that we can drop
 621                  * refcount to losing lock clds */
 622                 if (unlikely(stopped)) {
 623                         spin_lock(&config_list_lock);
 624                         break;
 625                 }
 626
 627                 /* Wait a bit to see if anyone else needs a requeue */
 628                 lwi = (struct l_wait_info) { 0 };
 629                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 630                              &lwi);
 631                 spin_lock(&config_list_lock);
 632         }
 633         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 634         rq_state &= ~RQ_RUNNING;
 635         spin_unlock(&config_list_lock);
 636
 637         complete(&rq_exit);
 638
 639         CDEBUG(D_MGC, "Ending requeue thread\n");
 640         RETURN(rc);
 641 }
 642
 643 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 644    We are responsible for dropping the config log reference from here on out. */
 645 static void mgc_requeue_add(struct config_llog_data *cld)
 646 {
 647         ENTRY;
 648
 649         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 650                 cld->cld_logname, atomic_read(&cld->cld_refcount),
 651                 cld->cld_stopping, rq_state);
 652         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 653
 654         mutex_lock(&cld->cld_lock);
 655         if (cld->cld_stopping || cld->cld_lostlock) {
 656                 mutex_unlock(&cld->cld_lock);
 657                 RETURN_EXIT;
 658         }
 659         /* this refcount will be released in mgc_requeue_thread. */
 660         config_log_get(cld);
 661         cld->cld_lostlock = 1;
 662         mutex_unlock(&cld->cld_lock);
 663
 664         /* Hold lock for rq_state */
 665         spin_lock(&config_list_lock);
 666         if (rq_state & RQ_STOP) {
 667                 spin_unlock(&config_list_lock);
 668                 cld->cld_lostlock = 0;
 669                 config_log_put(cld);
 670         } else {
 671                 rq_state |= RQ_NOW;
 672                 spin_unlock(&config_list_lock);
 673                 wake_up(&rq_waitq);
 674         }
 675         EXIT;
 676 }
 677
 678 /********************** class fns **********************/
 679 static int mgc_local_llog_init(const struct lu_env *env,
 680                                struct obd_device *obd,
 681                                struct obd_device *disk)
 682 {
 683         struct llog_ctxt        *ctxt;
 684         int                      rc;
 685
 686         ENTRY;
 687
 688         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
 689                         &llog_osd_ops);
 690         if (rc)
 691                 RETURN(rc);
 692
 693         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 694         LASSERT(ctxt);
 695         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
 696         llog_ctxt_put(ctxt);
 697
 698         RETURN(0);
 699 }
 700
 701 static int mgc_local_llog_fini(const struct lu_env *env,
 702                                struct obd_device *obd)
 703 {
 704         struct llog_ctxt *ctxt;
 705
 706         ENTRY;
 707
 708         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 709         llog_cleanup(env, ctxt);
 710
 711         RETURN(0);
 712 }
 713
 714 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb)
 715 {
 716         struct lustre_sb_info   *lsi = s2lsi(sb);
 717         struct client_obd       *cli = &obd->u.cli;
 718         struct lu_fid            rfid, fid;
 719         struct dt_object        *root, *dto;
 720         struct lu_env           *env;
 721         int                      rc = 0;
 722
 723         ENTRY;
 724
 725         LASSERT(lsi);
 726         LASSERT(lsi->lsi_dt_dev);
 727
 728         OBD_ALLOC_PTR(env);
 729         if (env == NULL)
 730                 RETURN(-ENOMEM);
 731
 732         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
 733         mutex_lock(&cli->cl_mgc_mutex);
 734
 735         /* Setup the configs dir */
 736         rc = lu_env_init(env, LCT_MG_THREAD);
 737         if (rc)
 738                 GOTO(out_err, rc);
 739
 740         fid.f_seq = FID_SEQ_LOCAL_NAME;
 741         fid.f_oid = 1;
 742         fid.f_ver = 0;
 743         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
 744                                     &cli->cl_mgc_los);
 745         if (rc)
 746                 GOTO(out_env, rc);
 747
 748         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
 749         if (rc)
 750                 GOTO(out_env, rc);
 751
 752         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
 753                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
 754         if (unlikely(IS_ERR(root)))
 755                 GOTO(out_los, rc = PTR_ERR(root));
 756
 757         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
 758                                         MOUNT_CONFIGS_DIR,
 759                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
 760         lu_object_put_nocache(env, &root->do_lu);
 761         if (IS_ERR(dto))
 762                 GOTO(out_los, rc = PTR_ERR(dto));
 763
 764         cli->cl_mgc_configs_dir = dto;
 765
 766         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
 767         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
 768         if (rc)
 769                 GOTO(out_llog, rc);
 770
 771         /* We take an obd ref to insure that we can't get to mgc_cleanup
 772          * without calling mgc_fs_cleanup first. */
 773         class_incref(obd, "mgc_fs", obd);
 774
 775         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 776         EXIT;
 777 out_llog:
 778         if (rc) {
 779                 lu_object_put(env, &cli->cl_mgc_configs_dir->do_lu);
 780                 cli->cl_mgc_configs_dir = NULL;
 781         }
 782 out_los:
 783         if (rc < 0) {
 784                 local_oid_storage_fini(env, cli->cl_mgc_los);
 785                 cli->cl_mgc_los = NULL;
 786                 mutex_unlock(&cli->cl_mgc_mutex);
 787         }
 788 out_env:
 789         lu_env_fini(env);
 790 out_err:
 791         OBD_FREE_PTR(env);
 792         return rc;
 793 }
 794
 795 static int mgc_fs_cleanup(struct obd_device *obd)
 796 {
 797         struct lu_env            env;
 798         struct client_obd       *cli = &obd->u.cli;
 799         int                      rc;
 800
 801         ENTRY;
 802
 803         LASSERT(cli->cl_mgc_los != NULL);
 804
 805         rc = lu_env_init(&env, LCT_MG_THREAD);
 806         if (rc)
 807                 GOTO(unlock, rc);
 808
 809         mgc_local_llog_fini(&env, obd);
 810
 811         lu_object_put_nocache(&env, &cli->cl_mgc_configs_dir->do_lu);
 812         cli->cl_mgc_configs_dir = NULL;
 813
 814         local_oid_storage_fini(&env, cli->cl_mgc_los);
 815         cli->cl_mgc_los = NULL;
 816         lu_env_fini(&env);
 817
 818 unlock:
 819         class_decref(obd, "mgc_fs", obd);
 820         mutex_unlock(&cli->cl_mgc_mutex);
 821
 822         RETURN(0);
 823 }
 824
 825 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 826 {
 827         struct llog_ctxt        *ctxt;
 828         int                      rc;
 829
 830         ENTRY;
 831
 832         /* setup only remote ctxt, the local disk context is switched per each
 833          * filesystem during mgc_fs_setup() */
 834         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 835                         &llog_client_ops);
 836         if (rc)
 837                 RETURN(rc);
 838
 839         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 840         LASSERT(ctxt);
 841
 842         llog_initiator_connect(ctxt);
 843         llog_ctxt_put(ctxt);
 844
 845         RETURN(0);
 846 }
 847
 848 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 849 {
 850         struct llog_ctxt *ctxt;
 851
 852         ENTRY;
 853
 854         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 855         if (ctxt)
 856                 llog_cleanup(env, ctxt);
 857
 858         RETURN(0);
 859 }
 860
 861
 862 static atomic_t mgc_count = ATOMIC_INIT(0);
 863 static int mgc_precleanup(struct obd_device *obd)
 864 {
 865         int     rc = 0;
 866         int     temp;
 867         ENTRY;
 868
 869         if (atomic_dec_and_test(&mgc_count)) {
 870                 LASSERT(rq_state & RQ_RUNNING);
 871                 /* stop requeue thread */
 872                 temp = RQ_STOP;
 873         } else {
 874                 /* wakeup requeue thread to clean our cld */
 875                 temp = RQ_NOW | RQ_PRECLEANUP;
 876         }
 877
 878         spin_lock(&config_list_lock);
 879         rq_state |= temp;
 880         spin_unlock(&config_list_lock);
 881         wake_up(&rq_waitq);
 882
 883         if (temp & RQ_STOP)
 884                 wait_for_completion(&rq_exit);
 885         obd_cleanup_client_import(obd);
 886
 887         rc = mgc_llog_fini(NULL, obd);
 888         if (rc != 0)
 889                 CERROR("failed to cleanup llogging subsystems\n");
 890
 891         RETURN(rc);
 892 }
 893
 894 static int mgc_cleanup(struct obd_device *obd)
 895 {
 896         int rc;
 897         ENTRY;
 898
 899         /* COMPAT_146 - old config logs may have added profiles we don't
 900            know about */
 901         if (obd->obd_type->typ_refcnt <= 1)
 902                 /* Only for the last mgc */
 903                 class_del_profiles();
 904
 905         lprocfs_obd_cleanup(obd);
 906         ptlrpcd_decref();
 907
 908         rc = client_obd_cleanup(obd);
 909         RETURN(rc);
 910 }
 911
 912 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 913 {
 914         struct task_struct      *task;
 915         int                      rc;
 916         ENTRY;
 917
 918         rc = ptlrpcd_addref();
 919         if (rc < 0)
 920                 RETURN(rc);
 921
 922         rc = client_obd_setup(obd, lcfg);
 923         if (rc)
 924                 GOTO(err_decref, rc);
 925
 926         rc = mgc_llog_init(NULL, obd);
 927         if (rc) {
 928                 CERROR("failed to setup llogging subsystems\n");
 929                 GOTO(err_cleanup, rc);
 930         }
 931
 932 #ifdef CONFIG_PROC_FS
 933         obd->obd_vars = lprocfs_mgc_obd_vars;
 934         lprocfs_obd_setup(obd);
 935 #endif
 936         sptlrpc_lprocfs_cliobd_attach(obd);
 937
 938         if (atomic_inc_return(&mgc_count) == 1) {
 939                 rq_state = 0;
 940                 init_waitqueue_head(&rq_waitq);
 941
 942                 /* start requeue thread */
 943                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 944                 if (IS_ERR(task)) {
 945                         rc = PTR_ERR(task);
 946                         CERROR("%s: cannot start requeue thread: rc = %d; "
 947                                "no more log updates\n",
 948                                obd->obd_name, rc);
 949                         GOTO(err_cleanup, rc);
 950                 }
 951                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 952                 rc = 0;
 953                 wait_for_completion(&rq_start);
 954         }
 955
 956         RETURN(rc);
 957
 958 err_cleanup:
 959         client_obd_cleanup(obd);
 960 err_decref:
 961         ptlrpcd_decref();
 962         RETURN(rc);
 963 }
 964
 965 /* based on ll_mdc_blocking_ast */
 966 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 967                             void *data, int flag)
 968 {
 969         struct lustre_handle lockh;
 970         struct config_llog_data *cld = (struct config_llog_data *)data;
 971         int rc = 0;
 972         ENTRY;
 973
 974         switch (flag) {
 975         case LDLM_CB_BLOCKING:
 976                 /* mgs wants the lock, give it up... */
 977                 LDLM_DEBUG(lock, "MGC blocking CB");
 978                 ldlm_lock2handle(lock, &lockh);
 979                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 980                 break;
 981         case LDLM_CB_CANCELING:
 982                 /* We've given up the lock, prepare ourselves to update. */
 983                 LDLM_DEBUG(lock, "MGC cancel CB");
 984
 985                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
 986                        PLDLMRES(lock->l_resource),
 987                        (char *)&lock->l_resource->lr_name.name[0]);
 988
 989                 if (!cld) {
 990                         CDEBUG(D_INFO, "missing data, won't requeue\n");
 991                         break;
 992                 }
 993
 994                 /* held at mgc_process_log(). */
 995                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
 996                 /* Are we done with this log? */
 997                 if (cld->cld_stopping) {
 998                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
 999                                 cld->cld_logname);
1000                         config_log_put(cld);
1001                         break;
1002                 }
1003                 /* Make sure not to re-enqueue when the mgc is stopping
1004                    (we get called from client_disconnect_export) */
1005                 if (lock->l_conn_export == NULL ||
1006                     lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1007                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1008                                 cld->cld_logname);
1009                         config_log_put(cld);
1010                         break;
1011                 }
1012
1013                 /* Re-enqueue now */
1014                 mgc_requeue_add(cld);
1015                 config_log_put(cld);
1016                 break;
1017         default:
1018                 LBUG();
1019         }
1020
1021         RETURN(rc);
1022 }
1023
1024 /* Not sure where this should go... */
1025 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1026  * that we can have a chance to try the secondary MGS if any. */
1027 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1028                                 + PING_INTERVAL)
1029 #define  MGC_TARGET_REG_LIMIT 10
1030 #define  MGC_SEND_PARAM_LIMIT 10
1031
1032 /* Send parameter to MGS*/
1033 static int mgc_set_mgs_param(struct obd_export *exp,
1034                              struct mgs_send_param *msp)
1035 {
1036         struct ptlrpc_request *req;
1037         struct mgs_send_param *req_msp, *rep_msp;
1038         int rc;
1039         ENTRY;
1040
1041         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1042                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1043                                         MGS_SET_INFO);
1044         if (!req)
1045                 RETURN(-ENOMEM);
1046
1047         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1048         if (!req_msp) {
1049                 ptlrpc_req_finished(req);
1050                 RETURN(-ENOMEM);
1051         }
1052
1053         memcpy(req_msp, msp, sizeof(*req_msp));
1054         ptlrpc_request_set_replen(req);
1055
1056         /* Limit how long we will wait for the enqueue to complete */
1057         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1058         rc = ptlrpc_queue_wait(req);
1059         if (!rc) {
1060                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1061                 memcpy(msp, rep_msp, sizeof(*rep_msp));
1062         }
1063
1064         ptlrpc_req_finished(req);
1065
1066         RETURN(rc);
1067 }
1068
1069 /* Take a config lock so we can get cancel notifications */
1070 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1071                        union ldlm_policy_data *policy, enum ldlm_mode mode,
1072                        __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
1073                        void *data, __u32 lvb_len, void *lvb_swabber,
1074                        struct lustre_handle *lockh)
1075 {
1076         struct config_llog_data *cld = (struct config_llog_data *)data;
1077         struct ldlm_enqueue_info einfo = {
1078                 .ei_type        = type,
1079                 .ei_mode        = mode,
1080                 .ei_cb_bl       = mgc_blocking_ast,
1081                 .ei_cb_cp       = ldlm_completion_ast,
1082         };
1083         struct ptlrpc_request *req;
1084         int short_limit = cld_is_sptlrpc(cld);
1085         int rc;
1086         ENTRY;
1087
1088         CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
1089                cld->cld_resid.name[0]);
1090
1091         /* We need a callback for every lockholder, so don't try to
1092            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1093         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1094                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1095                                         LDLM_ENQUEUE);
1096         if (req == NULL)
1097                 RETURN(-ENOMEM);
1098
1099         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1100         ptlrpc_request_set_replen(req);
1101
1102         /* check if this is server or client */
1103         if (cld->cld_cfg.cfg_sb) {
1104                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1105                 if (lsi && IS_SERVER(lsi))
1106                         short_limit = 1;
1107         }
1108         /* Limit how long we will wait for the enqueue to complete */
1109         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1110         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1111                               NULL, 0, LVB_T_NONE, lockh, 0);
1112         /* A failed enqueue should still call the mgc_blocking_ast,
1113            where it will be requeued if needed ("grant failed"). */
1114         ptlrpc_req_finished(req);
1115         RETURN(rc);
1116 }
1117
1118 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1119                       struct lustre_handle *lockh)
1120 {
1121         ENTRY;
1122
1123         ldlm_lock_decref(lockh, mode);
1124
1125         RETURN(0);
1126 }
1127
1128 static void mgc_notify_active(struct obd_device *unused)
1129 {
1130         /* wakeup mgc_requeue_thread to requeue mgc lock */
1131         spin_lock(&config_list_lock);
1132         rq_state |= RQ_NOW;
1133         spin_unlock(&config_list_lock);
1134         wake_up(&rq_waitq);
1135
1136         /* TODO: Help the MGS rebuild nidtbl. -jay */
1137 }
1138
1139 /* Send target_reg message to MGS */
1140 static int mgc_target_register(struct obd_export *exp,
1141                                struct mgs_target_info *mti)
1142 {
1143         struct ptlrpc_request  *req;
1144         struct mgs_target_info *req_mti, *rep_mti;
1145         int                     rc;
1146         ENTRY;
1147
1148         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1149                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1150                                         MGS_TARGET_REG);
1151         if (req == NULL)
1152                 RETURN(-ENOMEM);
1153
1154         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1155         if (!req_mti) {
1156                 ptlrpc_req_finished(req);
1157                 RETURN(-ENOMEM);
1158         }
1159
1160         memcpy(req_mti, mti, sizeof(*req_mti));
1161         ptlrpc_request_set_replen(req);
1162         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1163         /* Limit how long we will wait for the enqueue to complete */
1164         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1165
1166         rc = ptlrpc_queue_wait(req);
1167         if (!rc) {
1168                 rep_mti = req_capsule_server_get(&req->rq_pill,
1169                                                  &RMF_MGS_TARGET_INFO);
1170                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1171                 CDEBUG(D_MGC, "register %s got index = %d\n",
1172                        mti->mti_svname, mti->mti_stripe_index);
1173         }
1174         ptlrpc_req_finished(req);
1175
1176         RETURN(rc);
1177 }
1178
1179 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1180                               u32 keylen, void *key,
1181                               u32 vallen, void *val,
1182                               struct ptlrpc_request_set *set)
1183 {
1184         int rc = -EINVAL;
1185         ENTRY;
1186
1187         /* Turn off initial_recov after we try all backup servers once */
1188         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1189                 struct obd_import *imp = class_exp2cliimp(exp);
1190                 int value;
1191                 if (vallen != sizeof(int))
1192                         RETURN(-EINVAL);
1193                 value = *(int *)val;
1194                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1195                        imp->imp_obd->obd_name, value,
1196                        imp->imp_deactive, imp->imp_invalid,
1197                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1198                        ptlrpc_import_state_name(imp->imp_state));
1199                 /* Resurrect if we previously died */
1200                 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1201                      imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1202                         ptlrpc_reconnect_import(imp);
1203                 RETURN(0);
1204         }
1205         /* FIXME move this to mgc_process_config */
1206         if (KEY_IS(KEY_REGISTER_TARGET)) {
1207                 struct mgs_target_info *mti;
1208                 if (vallen != sizeof(struct mgs_target_info))
1209                         RETURN(-EINVAL);
1210                 mti = (struct mgs_target_info *)val;
1211                 CDEBUG(D_MGC, "register_target %s %#x\n",
1212                        mti->mti_svname, mti->mti_flags);
1213                 rc =  mgc_target_register(exp, mti);
1214                 RETURN(rc);
1215         }
1216         if (KEY_IS(KEY_SET_FS)) {
1217                 struct super_block *sb = (struct super_block *)val;
1218
1219                 if (vallen != sizeof(struct super_block))
1220                         RETURN(-EINVAL);
1221
1222                 rc = mgc_fs_setup(exp->exp_obd, sb);
1223                 RETURN(rc);
1224         }
1225         if (KEY_IS(KEY_CLEAR_FS)) {
1226                 if (vallen != 0)
1227                         RETURN(-EINVAL);
1228                 rc = mgc_fs_cleanup(exp->exp_obd);
1229                 RETURN(rc);
1230         }
1231         if (KEY_IS(KEY_SET_INFO)) {
1232                 struct mgs_send_param *msp;
1233
1234                 msp = (struct mgs_send_param *)val;
1235                 rc =  mgc_set_mgs_param(exp, msp);
1236                 RETURN(rc);
1237         }
1238         if (KEY_IS(KEY_MGSSEC)) {
1239                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1240                 struct sptlrpc_flavor  flvr;
1241
1242                 /*
1243                  * empty string means using current flavor, if which haven't
1244                  * been set yet, set it as null.
1245                  *
1246                  * if flavor has been set previously, check the asking flavor
1247                  * must match the existing one.
1248                  */
1249                 if (vallen == 0) {
1250                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1251                                 RETURN(0);
1252                         val = "null";
1253                         vallen = 4;
1254                 }
1255
1256                 rc = sptlrpc_parse_flavor(val, &flvr);
1257                 if (rc) {
1258                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1259                                (char *) val);
1260                         RETURN(rc);
1261                 }
1262
1263                 /*
1264                  * caller already hold a mutex
1265                  */
1266                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1267                         cli->cl_flvr_mgc = flvr;
1268                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1269                                   sizeof(flvr)) != 0) {
1270                         char    str[20];
1271
1272                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1273                                             str, sizeof(str));
1274                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1275                                        "currently %s is in use\n",
1276                                        (char *) val, str);
1277                         rc = -EPERM;
1278                 }
1279                 RETURN(rc);
1280         }
1281
1282         RETURN(rc);
1283 }
1284
1285 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1286                         __u32 keylen, void *key, __u32 *vallen, void *val)
1287 {
1288         int rc = -EINVAL;
1289
1290         if (KEY_IS(KEY_CONN_DATA)) {
1291                 struct obd_import *imp = class_exp2cliimp(exp);
1292                 struct obd_connect_data *data = val;
1293
1294                 if (*vallen == sizeof(*data)) {
1295                         *data = imp->imp_connect_data;
1296                         rc = 0;
1297                 }
1298         }
1299
1300         return rc;
1301 }
1302
1303 static int mgc_import_event(struct obd_device *obd,
1304                             struct obd_import *imp,
1305                             enum obd_import_event event)
1306 {
1307         int rc = 0;
1308
1309         LASSERT(imp->imp_obd == obd);
1310         CDEBUG(D_MGC, "import event %#x\n", event);
1311
1312         switch (event) {
1313         case IMP_EVENT_DISCON:
1314                 /* MGC imports should not wait for recovery */
1315                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1316                         ptlrpc_pinger_ir_down();
1317                 break;
1318         case IMP_EVENT_INACTIVE:
1319                 break;
1320         case IMP_EVENT_INVALIDATE: {
1321                 struct ldlm_namespace *ns = obd->obd_namespace;
1322                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1323                 break;
1324         }
1325         case IMP_EVENT_ACTIVE:
1326                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1327                 /* Clearing obd_no_recov allows us to continue pinging */
1328                 obd->obd_no_recov = 0;
1329                 mgc_notify_active(obd);
1330                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1331                         ptlrpc_pinger_ir_up();
1332                 break;
1333         case IMP_EVENT_OCD:
1334                 break;
1335         case IMP_EVENT_DEACTIVATE:
1336         case IMP_EVENT_ACTIVATE:
1337                 break;
1338         default:
1339                 CERROR("Unknown import event %#x\n", event);
1340                 LBUG();
1341         }
1342         RETURN(rc);
1343 }
1344
1345 enum {
1346         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_CACHE_SHIFT),
1347         CONFIG_READ_NRPAGES      = 4
1348 };
1349
1350 static int mgc_apply_recover_logs(struct obd_device *mgc,
1351                                   struct config_llog_data *cld,
1352                                   __u64 max_version,
1353                                   void *data, int datalen, bool mne_swab)
1354 {
1355         struct config_llog_instance *cfg = &cld->cld_cfg;
1356         struct lustre_sb_info       *lsi = s2lsi(cfg->cfg_sb);
1357         struct mgs_nidtbl_entry *entry;
1358         struct lustre_cfg       *lcfg;
1359         struct lustre_cfg_bufs   bufs;
1360         u64   prev_version = 0;
1361         char *inst;
1362         char *buf;
1363         int   bufsz;
1364         int   pos;
1365         int   rc  = 0;
1366         int   off = 0;
1367         ENTRY;
1368
1369         LASSERT(cfg->cfg_instance != NULL);
1370         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1371
1372         OBD_ALLOC(inst, PAGE_CACHE_SIZE);
1373         if (inst == NULL)
1374                 RETURN(-ENOMEM);
1375
1376         if (!IS_SERVER(lsi)) {
1377                 pos = snprintf(inst, PAGE_CACHE_SIZE, "%p", cfg->cfg_instance);
1378                 if (pos >= PAGE_CACHE_SIZE) {
1379                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1380                         return -E2BIG;
1381                 }
1382         } else {
1383                 LASSERT(IS_MDT(lsi));
1384                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1385                                         PAGE_CACHE_SIZE);
1386                 if (rc) {
1387                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1388                         RETURN(-EINVAL);
1389                 }
1390                 pos = strlen(inst);
1391         }
1392
1393         ++pos;
1394         buf   = inst + pos;
1395         bufsz = PAGE_CACHE_SIZE - pos;
1396
1397         while (datalen > 0) {
1398                 int   entry_len = sizeof(*entry);
1399                 int   is_ost, i;
1400                 struct obd_device *obd;
1401                 char *obdname;
1402                 char *cname;
1403                 char *params;
1404                 char *uuid;
1405
1406                 rc = -EINVAL;
1407                 if (datalen < sizeof(*entry))
1408                         break;
1409
1410                 entry = (typeof(entry))(data + off);
1411
1412                 /* sanity check */
1413                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1414                         break;
1415                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1416                         break;
1417                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1418                         break;
1419
1420                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1421                 if (datalen < entry_len) /* must have entry_len at least */
1422                         break;
1423
1424                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1425                 if (mne_swab)
1426                         lustre_swab_mgs_nidtbl_entry(entry);
1427                 if (entry->mne_length > PAGE_CACHE_SIZE) {
1428                         CERROR("MNE too large (%u)\n", entry->mne_length);
1429                         break;
1430                 }
1431
1432                 if (entry->mne_length < entry_len)
1433                         break;
1434
1435                 off     += entry->mne_length;
1436                 datalen -= entry->mne_length;
1437                 if (datalen < 0)
1438                         break;
1439
1440                 if (entry->mne_version > max_version) {
1441                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1442                                entry->mne_version, max_version);
1443                         break;
1444                 }
1445
1446                 if (prev_version >= entry->mne_version) {
1447                         CERROR("index unsorted, prev %lld, now %lld\n",
1448                                prev_version, entry->mne_version);
1449                         break;
1450                 }
1451                 prev_version = entry->mne_version;
1452
1453                 /*
1454                  * Write a string with format "nid::instance" to
1455                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1456                  */
1457
1458                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1459                 memset(buf, 0, bufsz);
1460                 obdname = buf;
1461                 pos = 0;
1462
1463                 /* lustre-OST0001-osc-<instance #> */
1464                 strcpy(obdname, cld->cld_logname);
1465                 cname = strrchr(obdname, '-');
1466                 if (cname == NULL) {
1467                         CERROR("mgc %s: invalid logname %s\n",
1468                                mgc->obd_name, obdname);
1469                         break;
1470                 }
1471
1472                 pos = cname - obdname;
1473                 obdname[pos] = 0;
1474                 pos += sprintf(obdname + pos, "-%s%04x",
1475                                   is_ost ? "OST" : "MDT", entry->mne_index);
1476
1477                 cname = is_ost ? "osc" : "mdc",
1478                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1479                 lustre_cfg_bufs_reset(&bufs, obdname);
1480
1481                 /* find the obd by obdname */
1482                 obd = class_name2obd(obdname);
1483                 if (obd == NULL) {
1484                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1485                                mgc->obd_name, obdname);
1486                         rc = 0;
1487                         /* this is a safe race, when the ost is starting up...*/
1488                         continue;
1489                 }
1490
1491                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1492                 ++pos;
1493                 params = buf + pos;
1494                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1495                 uuid = buf + pos;
1496
1497                 down_read(&obd->u.cli.cl_sem);
1498                 if (obd->u.cli.cl_import == NULL) {
1499                         /* client does not connect to the OST yet */
1500                         up_read(&obd->u.cli.cl_sem);
1501                         rc = 0;
1502                         continue;
1503                 }
1504
1505                 /* iterate all nids to find one */
1506                 /* find uuid by nid */
1507                 rc = -ENOENT;
1508                 for (i = 0; i < entry->mne_nid_count; i++) {
1509                         rc = client_import_find_conn(obd->u.cli.cl_import,
1510                                                      entry->u.nids[i],
1511                                                      (struct obd_uuid *)uuid);
1512                         if (rc == 0)
1513                                 break;
1514                 }
1515
1516                 up_read(&obd->u.cli.cl_sem);
1517                 if (rc < 0) {
1518                         CERROR("mgc: cannot find uuid by nid %s\n",
1519                                libcfs_nid2str(entry->u.nids[0]));
1520                         break;
1521                 }
1522
1523                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1524                        uuid, libcfs_nid2str(entry->u.nids[0]));
1525
1526                 pos += strlen(uuid);
1527                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1528                 LASSERT(pos < bufsz);
1529
1530                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1531
1532                 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1533                 if (lcfg == NULL) {
1534                         rc = -ENOMEM;
1535                         break;
1536                 }
1537
1538                 CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n",
1539                        prev_version, max_version, obdname, params);
1540
1541                 rc = class_process_config(lcfg);
1542                 lustre_cfg_free(lcfg);
1543                 if (rc)
1544                         CDEBUG(D_INFO, "process config for %s error %d\n",
1545                                obdname, rc);
1546
1547                 /* continue, even one with error */
1548         }
1549
1550         OBD_FREE(inst, PAGE_CACHE_SIZE);
1551         RETURN(rc);
1552 }
1553
1554 /**
1555  * This function is called if this client was notified for target restarting
1556  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1557  */
1558 static int mgc_process_recover_log(struct obd_device *obd,
1559                                    struct config_llog_data *cld)
1560 {
1561         struct ptlrpc_request *req = NULL;
1562         struct config_llog_instance *cfg = &cld->cld_cfg;
1563         struct mgs_config_body *body;
1564         struct mgs_config_res  *res;
1565         struct ptlrpc_bulk_desc *desc;
1566         struct page **pages;
1567         int nrpages;
1568         bool eof = true;
1569         bool mne_swab = false;
1570         int i;
1571         int ealen;
1572         int rc;
1573         ENTRY;
1574
1575         /* allocate buffer for bulk transfer.
1576          * if this is the first time for this mgs to read logs,
1577          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1578          * once; otherwise, it only reads increment of logs, this should be
1579          * small and CONFIG_READ_NRPAGES will be used.
1580          */
1581         nrpages = CONFIG_READ_NRPAGES;
1582         if (cfg->cfg_last_idx == 0) /* the first time */
1583                 nrpages = CONFIG_READ_NRPAGES_INIT;
1584
1585         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1586         if (pages == NULL)
1587                 GOTO(out, rc = -ENOMEM);
1588
1589         for (i = 0; i < nrpages; i++) {
1590                 pages[i] = alloc_page(GFP_KERNEL);
1591                 if (pages[i] == NULL)
1592                         GOTO(out, rc = -ENOMEM);
1593         }
1594
1595 again:
1596         LASSERT(cld_is_recover(cld));
1597         LASSERT(mutex_is_locked(&cld->cld_lock));
1598         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1599                                    &RQF_MGS_CONFIG_READ);
1600         if (req == NULL)
1601                 GOTO(out, rc = -ENOMEM);
1602
1603         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1604         if (rc)
1605                 GOTO(out, rc);
1606
1607         /* pack request */
1608         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1609         LASSERT(body != NULL);
1610         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1611         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1612             >= sizeof(body->mcb_name))
1613                 GOTO(out, rc = -E2BIG);
1614         body->mcb_offset = cfg->cfg_last_idx + 1;
1615         body->mcb_type   = cld->cld_type;
1616         body->mcb_bits   = PAGE_CACHE_SHIFT;
1617         body->mcb_units  = nrpages;
1618
1619         /* allocate bulk transfer descriptor */
1620         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1621                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1622                                     MGS_BULK_PORTAL,
1623                                     &ptlrpc_bulk_kiov_pin_ops);
1624         if (desc == NULL)
1625                 GOTO(out, rc = -ENOMEM);
1626
1627         for (i = 0; i < nrpages; i++)
1628                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1629                                                  PAGE_CACHE_SIZE);
1630
1631         ptlrpc_request_set_replen(req);
1632         rc = ptlrpc_queue_wait(req);
1633         if (rc)
1634                 GOTO(out, rc);
1635
1636         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1637         if (res->mcr_size < res->mcr_offset)
1638                 GOTO(out, rc = -EINVAL);
1639
1640         /* always update the index even though it might have errors with
1641          * handling the recover logs */
1642         cfg->cfg_last_idx = res->mcr_offset;
1643         eof = res->mcr_offset == res->mcr_size;
1644
1645         CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n",
1646                res->mcr_offset, eof == false);
1647
1648         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1649         if (ealen < 0)
1650                 GOTO(out, rc = ealen);
1651
1652         if (ealen > nrpages << PAGE_CACHE_SHIFT)
1653                 GOTO(out, rc = -EINVAL);
1654
1655         if (ealen == 0) { /* no logs transferred */
1656                 if (!eof)
1657                         rc = -EINVAL;
1658                 GOTO(out, rc);
1659         }
1660
1661         mne_swab = !!ptlrpc_rep_need_swab(req);
1662 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1663         /* This import flag means the server did an extra swab of IR MNE
1664          * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1665         if (unlikely(req->rq_import->imp_need_mne_swab))
1666                 mne_swab = !mne_swab;
1667 #endif
1668
1669         for (i = 0; i < nrpages && ealen > 0; i++) {
1670                 int rc2;
1671                 void *ptr;
1672
1673                 ptr = kmap(pages[i]);
1674                 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1675                                              min_t(int, ealen, PAGE_CACHE_SIZE),
1676                                              mne_swab);
1677                 kunmap(pages[i]);
1678                 if (rc2 < 0) {
1679                         CWARN("Process recover log %s error %d\n",
1680                               cld->cld_logname, rc2);
1681                         break;
1682                 }
1683
1684                 ealen -= PAGE_CACHE_SIZE;
1685         }
1686
1687 out:
1688         if (req)
1689                 ptlrpc_req_finished(req);
1690
1691         if (rc == 0 && !eof)
1692                 goto again;
1693
1694         if (pages) {
1695                 for (i = 0; i < nrpages; i++) {
1696                         if (pages[i] == NULL)
1697                                 break;
1698                         __free_page(pages[i]);
1699                 }
1700                 OBD_FREE(pages, sizeof(*pages) * nrpages);
1701         }
1702         return rc;
1703 }
1704
1705 /* Copy a remote log locally */
1706 static int mgc_llog_local_copy(const struct lu_env *env,
1707                                struct obd_device *obd,
1708                                struct llog_ctxt *rctxt,
1709                                struct llog_ctxt *lctxt, char *logname)
1710 {
1711         char    *temp_log;
1712         int      rc;
1713
1714         ENTRY;
1715
1716         /*
1717          * - copy it to backup using llog_backup()
1718          * - copy remote llog to logname using llog_backup()
1719          * - if failed then move bakup to logname again
1720          */
1721
1722         OBD_ALLOC(temp_log, strlen(logname) + 2);
1723         if (!temp_log)
1724                 RETURN(-ENOMEM);
1725         sprintf(temp_log, "%sT", logname);
1726
1727         /* make a copy of local llog at first */
1728         rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1729         if (rc < 0 && rc != -ENOENT)
1730                 GOTO(out, rc);
1731         /* copy remote llog to the local copy */
1732         rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1733         if (rc == -ENOENT) {
1734                 /* no remote llog, delete local one too */
1735                 llog_erase(env, lctxt, NULL, logname);
1736         } else if (rc < 0) {
1737                 /* error during backup, get local one back from the copy */
1738                 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1739 out:
1740                 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1741                        obd->obd_name, logname, rc);
1742         }
1743         llog_erase(env, lctxt, NULL, temp_log);
1744         OBD_FREE(temp_log, strlen(logname) + 2);
1745         return rc;
1746 }
1747
1748 /* local_only means it cannot get remote llogs */
1749 static int mgc_process_cfg_log(struct obd_device *mgc,
1750                                struct config_llog_data *cld, int local_only)
1751 {
1752         struct llog_ctxt        *ctxt, *lctxt = NULL;
1753         struct client_obd       *cli = &mgc->u.cli;
1754         struct lustre_sb_info   *lsi = NULL;
1755         int                      rc = 0;
1756         bool                     sptlrpc_started = false;
1757         struct lu_env           *env;
1758
1759         ENTRY;
1760
1761         LASSERT(cld);
1762         LASSERT(mutex_is_locked(&cld->cld_lock));
1763
1764         /*
1765          * local copy of sptlrpc log is controlled elsewhere, don't try to
1766          * read it up here.
1767          */
1768         if (cld_is_sptlrpc(cld) && local_only)
1769                 RETURN(0);
1770
1771         if (cld->cld_cfg.cfg_sb)
1772                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1773
1774         OBD_ALLOC_PTR(env);
1775         if (env == NULL)
1776                 RETURN(-ENOMEM);
1777
1778         rc = lu_env_init(env, LCT_MG_THREAD);
1779         if (rc)
1780                 GOTO(out_free, rc);
1781
1782         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1783         LASSERT(ctxt);
1784
1785         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1786
1787         /* Copy the setup log locally if we can. Don't mess around if we're
1788          * running an MGS though (logs are already local). */
1789         if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1790             cli->cl_mgc_configs_dir != NULL &&
1791             lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1792             lsi->lsi_dt_dev) {
1793                 if (!local_only)
1794                         /* Only try to copy log if we have the lock. */
1795                         rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1796                                                  cld->cld_logname);
1797                 if (local_only || rc) {
1798                         if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1799                             llog_is_empty(env, lctxt, cld->cld_logname)) {
1800                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1801                                                    "log %s and no local copy."
1802                                                    "\n", cld->cld_logname);
1803                                 GOTO(out_pop, rc = -ENOENT);
1804                         }
1805                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1806                                "copy for now, will try to update later.\n",
1807                                cld->cld_logname);
1808                         rc = 0;
1809                 }
1810                 /* Now, whether we copied or not, start using the local llog.
1811                  * If we failed to copy, we'll start using whatever the old
1812                  * log has. */
1813                 llog_ctxt_put(ctxt);
1814                 ctxt = lctxt;
1815                 lctxt = NULL;
1816         } else {
1817                 if (local_only) /* no local log at client side */
1818                         GOTO(out_pop, rc = -EIO);
1819         }
1820
1821         if (cld_is_sptlrpc(cld)) {
1822                 sptlrpc_conf_log_update_begin(cld->cld_logname);
1823                 sptlrpc_started = true;
1824         }
1825
1826         /* logname and instance info should be the same, so use our
1827          * copy of the instance for the update.  The cfg_last_idx will
1828          * be updated here. */
1829         rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1830                                      &cld->cld_cfg);
1831         EXIT;
1832
1833 out_pop:
1834         __llog_ctxt_put(env, ctxt);
1835         if (lctxt)
1836                 __llog_ctxt_put(env, lctxt);
1837
1838         /*
1839          * update settings on existing OBDs. doing it inside
1840          * of llog_process_lock so no device is attaching/detaching
1841          * in parallel.
1842          * the logname must be <fsname>-sptlrpc
1843          */
1844         if (sptlrpc_started) {
1845                 LASSERT(cld_is_sptlrpc(cld));
1846                 sptlrpc_conf_log_update_end(cld->cld_logname);
1847                 class_notify_sptlrpc_conf(cld->cld_logname,
1848                                           strlen(cld->cld_logname) -
1849                                           strlen("-sptlrpc"));
1850         }
1851
1852         lu_env_fini(env);
1853 out_free:
1854         OBD_FREE_PTR(env);
1855         return rc;
1856 }
1857
1858 static bool mgc_import_in_recovery(struct obd_import *imp)
1859 {
1860         bool in_recovery = true;
1861
1862         spin_lock(&imp->imp_lock);
1863         if (imp->imp_state == LUSTRE_IMP_FULL ||
1864             imp->imp_state == LUSTRE_IMP_CLOSED)
1865                 in_recovery = false;
1866         spin_unlock(&imp->imp_lock);
1867
1868         return in_recovery;
1869 }
1870
1871 /**
1872  * Get a configuration log from the MGS and process it.
1873  *
1874  * This function is called for both clients and servers to process the
1875  * configuration log from the MGS.  The MGC enqueues a DLM lock on the
1876  * log from the MGS, and if the lock gets revoked the MGC will be notified
1877  * by the lock cancellation callback that the config log has changed,
1878  * and will enqueue another MGS lock on it, and then continue processing
1879  * the new additions to the end of the log.
1880  *
1881  * Since the MGC import is not replayable, if the import is being evicted
1882  * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
1883  * the log until recovery is finished or the import is closed.
1884  *
1885  * Make a local copy of the log before parsing it if appropriate (non-MGS
1886  * server) so that the server can start even when the MGS is down.
1887  *
1888  * There shouldn't be multiple processes running process_log at once --
1889  * sounds like badness.  It actually might be fine, as long as they're not
1890  * trying to update from the same log simultaneously, in which case we
1891  * should use a per-log semaphore instead of cld_lock.
1892  *
1893  * \param[in] mgc       MGC device by which to fetch the configuration log
1894  * \param[in] cld       log processing state (stored in lock callback data)
1895  *
1896  * \retval              0 on success
1897  * \retval              negative errno on failure
1898  */
1899 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1900 {
1901         struct lustre_handle lockh = { 0 };
1902         __u64 flags = LDLM_FL_NO_LRU;
1903         int rc = 0, rcl;
1904         bool retry = false;
1905         ENTRY;
1906
1907         LASSERT(cld != NULL);
1908
1909         /* I don't want multiple processes running process_log at once --
1910            sounds like badness.  It actually might be fine, as long as
1911            we're not trying to update from the same log
1912            simultaneously (in which case we should use a per-log sem.) */
1913 restart:
1914         mutex_lock(&cld->cld_lock);
1915         if (cld->cld_stopping) {
1916                 mutex_unlock(&cld->cld_lock);
1917                 RETURN(0);
1918         }
1919
1920         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1921
1922         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1923                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1924
1925         /* Get the cfg lock on the llog */
1926         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
1927                           LCK_CR, &flags, NULL, NULL, NULL,
1928                           cld, 0, NULL, &lockh);
1929         if (rcl == 0) {
1930                 /* Get the cld, it will be released in mgc_blocking_ast. */
1931                 config_log_get(cld);
1932                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1933                 LASSERT(rc == 0);
1934         } else {
1935                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1936
1937                 if (rcl == -ESHUTDOWN &&
1938                     atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
1939                         struct obd_import *imp;
1940                         struct l_wait_info lwi;
1941                         int secs = cfs_time_seconds(obd_timeout);
1942
1943                         mutex_unlock(&cld->cld_lock);
1944                         imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
1945
1946                         /* Let's force the pinger, and wait the import to be
1947                          * connected, note: since mgc import is non-replayable,
1948                          * and even the import state is disconnected, it does
1949                          * not mean the "recovery" is stopped, so we will keep
1950                          * waitting until timeout or the import state is
1951                          * FULL or closed */
1952                         ptlrpc_pinger_force(imp);
1953
1954                         lwi = LWI_TIMEOUT(secs, NULL, NULL);
1955                         l_wait_event(imp->imp_recovery_waitq,
1956                                      !mgc_import_in_recovery(imp), &lwi);
1957
1958                         if (imp->imp_state == LUSTRE_IMP_FULL) {
1959                                 retry = true;
1960                                 goto restart;
1961                         } else {
1962                                 mutex_lock(&cld->cld_lock);
1963                                 cld->cld_lostlock = 1;
1964                         }
1965                 } else {
1966                         /* mark cld_lostlock so that it will requeue
1967                          * after MGC becomes available. */
1968                         cld->cld_lostlock = 1;
1969                 }
1970                 /* Get extra reference, it will be put in requeue thread */
1971                 config_log_get(cld);
1972         }
1973
1974
1975         if (cld_is_recover(cld)) {
1976                 rc = 0; /* this is not a fatal error for recover log */
1977                 if (rcl == 0) {
1978                         rc = mgc_process_recover_log(mgc, cld);
1979                         if (rc != 0) {
1980                                 CERROR("%s: recover log %s failed: rc = %d"
1981                                        "not fatal.\n", mgc->obd_name,
1982                                        cld->cld_logname, rc);
1983                                 rc = 0;
1984                                 cld->cld_lostlock = 1;
1985                         }
1986                 }
1987         } else {
1988                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1989         }
1990
1991         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1992                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1993
1994         mutex_unlock(&cld->cld_lock);
1995
1996         /* Now drop the lock so MGS can revoke it */
1997         if (!rcl) {
1998                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
1999                 if (rcl)
2000                         CERROR("Can't drop cfg lock: %d\n", rcl);
2001         }
2002
2003         RETURN(rc);
2004 }
2005
2006
2007 /** Called from lustre_process_log.
2008  * LCFG_LOG_START gets the config log from the MGS, processes it to start
2009  * any services, and adds it to the list logs to watch (follow).
2010  */
2011 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2012 {
2013         struct lustre_cfg *lcfg = buf;
2014         struct config_llog_instance *cfg = NULL;
2015         char *logname;
2016         int rc = 0;
2017         ENTRY;
2018
2019         switch(lcfg->lcfg_command) {
2020         case LCFG_LOV_ADD_OBD: {
2021                 /* Overloading this cfg command: register a new target */
2022                 struct mgs_target_info *mti;
2023
2024                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2025                     sizeof(struct mgs_target_info))
2026                         GOTO(out, rc = -EINVAL);
2027
2028                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2029                 CDEBUG(D_MGC, "add_target %s %#x\n",
2030                        mti->mti_svname, mti->mti_flags);
2031                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2032                 break;
2033         }
2034         case LCFG_LOV_DEL_OBD:
2035                 /* Unregister has no meaning at the moment. */
2036                 CERROR("lov_del_obd unimplemented\n");
2037                 rc = -ENOSYS;
2038                 break;
2039         case LCFG_SPTLRPC_CONF: {
2040                 rc = sptlrpc_process_config(lcfg);
2041                 break;
2042         }
2043         case LCFG_LOG_START: {
2044                 struct config_llog_data *cld;
2045                 struct super_block *sb;
2046
2047                 logname = lustre_cfg_string(lcfg, 1);
2048                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2049                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2050
2051                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2052                        cfg->cfg_last_idx);
2053
2054                 /* We're only called through here on the initial mount */
2055                 rc = config_log_add(obd, logname, cfg, sb);
2056                 if (rc)
2057                         break;
2058                 cld = config_log_find(logname, cfg);
2059                 if (cld == NULL) {
2060                         rc = -ENOENT;
2061                         break;
2062                 }
2063
2064                 /* COMPAT_146 */
2065                 /* FIXME only set this for old logs!  Right now this forces
2066                    us to always skip the "inside markers" check */
2067                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
2068
2069                 rc = mgc_process_log(obd, cld);
2070                 if (rc == 0 && cld->cld_recover != NULL) {
2071                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2072                                          imp_connect_data, IMP_RECOV)) {
2073                                 rc = mgc_process_log(obd, cld->cld_recover);
2074                         } else {
2075                                 struct config_llog_data *cir = cld->cld_recover;
2076                                 cld->cld_recover = NULL;
2077                                 config_log_put(cir);
2078                         }
2079                         if (rc)
2080                                 CERROR("Cannot process recover llog %d\n", rc);
2081                 }
2082
2083                 if (rc == 0 && cld->cld_params != NULL) {
2084                         rc = mgc_process_log(obd, cld->cld_params);
2085                         if (rc == -ENOENT) {
2086                                 CDEBUG(D_MGC, "There is no params "
2087                                               "config file yet\n");
2088                                 rc = 0;
2089                         }
2090                         /* params log is optional */
2091                         if (rc)
2092                                 CERROR("%s: can't process params llog: rc = %d\n",
2093                                        obd->obd_name, rc);
2094                 }
2095                 config_log_put(cld);
2096
2097                 break;
2098         }
2099         case LCFG_LOG_END: {
2100                 logname = lustre_cfg_string(lcfg, 1);
2101
2102                 if (lcfg->lcfg_bufcount >= 2)
2103                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
2104                                 lcfg, 2);
2105                 rc = config_log_end(logname, cfg);
2106                 break;
2107         }
2108         default: {
2109                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2110                 GOTO(out, rc = -EINVAL);
2111
2112         }
2113         }
2114 out:
2115         RETURN(rc);
2116 }
2117
2118 static struct obd_ops mgc_obd_ops = {
2119         .o_owner        = THIS_MODULE,
2120         .o_setup        = mgc_setup,
2121         .o_precleanup   = mgc_precleanup,
2122         .o_cleanup      = mgc_cleanup,
2123         .o_add_conn     = client_import_add_conn,
2124         .o_del_conn     = client_import_del_conn,
2125         .o_connect      = client_connect_import,
2126         .o_disconnect   = client_disconnect_export,
2127         .o_set_info_async = mgc_set_info_async,
2128         .o_get_info       = mgc_get_info,
2129         .o_import_event = mgc_import_event,
2130         .o_process_config = mgc_process_config,
2131 };
2132
2133 static int __init mgc_init(void)
2134 {
2135         return class_register_type(&mgc_obd_ops, NULL, true, NULL,
2136                                    LUSTRE_MGC_NAME, NULL);
2137 }
2138
2139 static void __exit mgc_exit(void)
2140 {
2141         class_unregister_type(LUSTRE_MGC_NAME);
2142 }
2143
2144 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2145 MODULE_DESCRIPTION("Lustre Management Client");
2146 MODULE_VERSION(LUSTRE_VERSION_STRING);
2147 MODULE_LICENSE("GPL");
2148
2149 module_init(mgc_init);
2150 module_exit(mgc_exit);