drivers/staging/lustre/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, 2015, 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 "../include/obd_class.h"
  42 #include "../include/lustre_dlm.h"
  43 #include "../include/lprocfs_status.h"
  44 #include "../include/lustre_log.h"
  45 #include "../include/lustre_disk.h"
  46
  47 #include "mgc_internal.h"
  48
  49 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
  50                           int type)
  51 {
  52         __u64 resname = 0;
  53
  54         if (len > sizeof(resname)) {
  55                 CERROR("name too long: %s\n", name);
  56                 return -EINVAL;
  57         }
  58         if (len <= 0) {
  59                 CERROR("missing name: %s\n", name);
  60                 return -EINVAL;
  61         }
  62         memcpy(&resname, name, len);
  63
  64         /* Always use the same endianness for the resid */
  65         memset(res_id, 0, sizeof(*res_id));
  66         res_id->name[0] = cpu_to_le64(resname);
  67         /* XXX: unfortunately, sptlprc and config llog share one lock */
  68         switch (type) {
  69         case CONFIG_T_CONFIG:
  70         case CONFIG_T_SPTLRPC:
  71                 resname = 0;
  72                 break;
  73         case CONFIG_T_RECOVER:
  74         case CONFIG_T_PARAMS:
  75                 resname = type;
  76                 break;
  77         default:
  78                 LBUG();
  79         }
  80         res_id->name[1] = cpu_to_le64(resname);
  81         CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
  82                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  83         return 0;
  84 }
  85
  86 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
  87 {
  88         /* fsname is at most 8 chars long, maybe contain "-".
  89          * e.g. "lustre", "SUN-000"
  90          */
  91         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
  92 }
  93 EXPORT_SYMBOL(mgc_fsname2resid);
  94
  95 static int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
  96 {
  97         char *name_end;
  98         int len;
  99
 100         /* logname consists of "fsname-nodetype".
 101          * e.g. "lustre-MDT0001", "SUN-000-client"
 102          * there is an exception: llog "params"
 103          */
 104         name_end = strrchr(logname, '-');
 105         if (!name_end)
 106                 len = strlen(logname);
 107         else
 108                 len = name_end - logname;
 109         return mgc_name2resid(logname, len, res_id, type);
 110 }
 111
 112 /********************** config llog list **********************/
 113 static LIST_HEAD(config_llog_list);
 114 static DEFINE_SPINLOCK(config_list_lock);
 115
 116 /* Take a reference to a config log */
 117 static int config_log_get(struct config_llog_data *cld)
 118 {
 119         atomic_inc(&cld->cld_refcount);
 120         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 121                atomic_read(&cld->cld_refcount));
 122         return 0;
 123 }
 124
 125 /* Drop a reference to a config log.  When no longer referenced,
 126  * we can free the config log data
 127  */
 128 static void config_log_put(struct config_llog_data *cld)
 129 {
 130         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 131                atomic_read(&cld->cld_refcount));
 132         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 133
 134         /* spinlock to make sure no item with 0 refcount in the list */
 135         if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 136                 list_del(&cld->cld_list_chain);
 137                 spin_unlock(&config_list_lock);
 138
 139                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 140
 141                 if (cld->cld_recover)
 142                         config_log_put(cld->cld_recover);
 143                 if (cld->cld_sptlrpc)
 144                         config_log_put(cld->cld_sptlrpc);
 145                 if (cld->cld_params)
 146                         config_log_put(cld->cld_params);
 147                 if (cld_is_sptlrpc(cld))
 148                         sptlrpc_conf_log_stop(cld->cld_logname);
 149
 150                 class_export_put(cld->cld_mgcexp);
 151                 kfree(cld);
 152         }
 153 }
 154
 155 /* Find a config log by name */
 156 static
 157 struct config_llog_data *config_log_find(char *logname,
 158                                          struct config_llog_instance *cfg)
 159 {
 160         struct config_llog_data *cld;
 161         struct config_llog_data *found = NULL;
 162         void *instance;
 163
 164         LASSERT(logname);
 165
 166         instance = cfg ? cfg->cfg_instance : NULL;
 167         spin_lock(&config_list_lock);
 168         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 169                 /* check if instance equals */
 170                 if (instance != cld->cld_cfg.cfg_instance)
 171                         continue;
 172
 173                 /* instance may be NULL, should check name */
 174                 if (strcmp(logname, cld->cld_logname) == 0) {
 175                         found = cld;
 176                         break;
 177                 }
 178         }
 179         if (found) {
 180                 atomic_inc(&found->cld_refcount);
 181                 LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0);
 182         }
 183         spin_unlock(&config_list_lock);
 184         return found;
 185 }
 186
 187 static
 188 struct config_llog_data *do_config_log_add(struct obd_device *obd,
 189                                            char *logname,
 190                                            int type,
 191                                            struct config_llog_instance *cfg,
 192                                            struct super_block *sb)
 193 {
 194         struct config_llog_data *cld;
 195         int                   rc;
 196
 197         CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
 198                cfg ? cfg->cfg_instance : NULL);
 199
 200         cld = kzalloc(sizeof(*cld) + strlen(logname) + 1, GFP_NOFS);
 201         if (!cld)
 202                 return ERR_PTR(-ENOMEM);
 203
 204         strcpy(cld->cld_logname, logname);
 205         if (cfg)
 206                 cld->cld_cfg = *cfg;
 207         else
 208                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 209         mutex_init(&cld->cld_lock);
 210         cld->cld_cfg.cfg_last_idx = 0;
 211         cld->cld_cfg.cfg_flags = 0;
 212         cld->cld_cfg.cfg_sb = sb;
 213         cld->cld_type = type;
 214         atomic_set(&cld->cld_refcount, 1);
 215
 216         /* Keep the mgc around until we are done */
 217         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 218
 219         if (cld_is_sptlrpc(cld)) {
 220                 sptlrpc_conf_log_start(logname);
 221                 cld->cld_cfg.cfg_obdname = obd->obd_name;
 222         }
 223
 224         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 225
 226         spin_lock(&config_list_lock);
 227         list_add(&cld->cld_list_chain, &config_llog_list);
 228         spin_unlock(&config_list_lock);
 229
 230         if (rc) {
 231                 config_log_put(cld);
 232                 return ERR_PTR(rc);
 233         }
 234
 235         if (cld_is_sptlrpc(cld)) {
 236                 rc = mgc_process_log(obd, cld);
 237                 if (rc && rc != -ENOENT)
 238                         CERROR("failed processing sptlrpc log: %d\n", rc);
 239         }
 240
 241         return cld;
 242 }
 243
 244 static struct config_llog_data *
 245 config_recover_log_add(struct obd_device *obd, char *fsname,
 246                        struct config_llog_instance *cfg,
 247                        struct super_block *sb)
 248 {
 249         struct config_llog_instance lcfg = *cfg;
 250         struct config_llog_data *cld;
 251         char logname[32];
 252
 253         /* we have to use different llog for clients and mdts for cmd
 254          * where only clients are notified if one of cmd server restarts
 255          */
 256         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 257         strcpy(logname, fsname);
 258         LASSERT(lcfg.cfg_instance);
 259         strcat(logname, "-cliir");
 260
 261         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
 262         return cld;
 263 }
 264
 265 static struct config_llog_data *
 266 config_params_log_add(struct obd_device *obd,
 267                       struct config_llog_instance *cfg, struct super_block *sb)
 268 {
 269         struct config_llog_instance     lcfg = *cfg;
 270         struct config_llog_data         *cld;
 271
 272         lcfg.cfg_instance = sb;
 273
 274         cld = do_config_log_add(obd, PARAMS_FILENAME, CONFIG_T_PARAMS,
 275                                 &lcfg, sb);
 276
 277         return cld;
 278 }
 279
 280 /** Add this log to the list of active logs watched by an MGC.
 281  * Active means we're watching for updates.
 282  * We have one active log per "mount" - client instance or servername.
 283  * Each instance may be at a different point in the log.
 284  */
 285 static int config_log_add(struct obd_device *obd, char *logname,
 286                           struct config_llog_instance *cfg,
 287                           struct super_block *sb)
 288 {
 289         struct lustre_sb_info *lsi = s2lsi(sb);
 290         struct config_llog_data *cld;
 291         struct config_llog_data *sptlrpc_cld;
 292         struct config_llog_data *params_cld;
 293         char                    seclogname[32];
 294         char                    *ptr;
 295         int                     rc;
 296
 297         CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
 298
 299         /*
 300          * for each regular log, the depended sptlrpc log name is
 301          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 302          */
 303         ptr = strrchr(logname, '-');
 304         if (!ptr || ptr - logname > 8) {
 305                 CERROR("logname %s is too long\n", logname);
 306                 return -EINVAL;
 307         }
 308
 309         memcpy(seclogname, logname, ptr - logname);
 310         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 311
 312         sptlrpc_cld = config_log_find(seclogname, NULL);
 313         if (!sptlrpc_cld) {
 314                 sptlrpc_cld = do_config_log_add(obd, seclogname,
 315                                                 CONFIG_T_SPTLRPC, NULL, NULL);
 316                 if (IS_ERR(sptlrpc_cld)) {
 317                         CERROR("can't create sptlrpc log: %s\n", seclogname);
 318                         rc = PTR_ERR(sptlrpc_cld);
 319                         goto out_err;
 320                 }
 321         }
 322         params_cld = config_params_log_add(obd, cfg, sb);
 323         if (IS_ERR(params_cld)) {
 324                 rc = PTR_ERR(params_cld);
 325                 CERROR("%s: can't create params log: rc = %d\n",
 326                        obd->obd_name, rc);
 327                 goto out_err1;
 328         }
 329
 330         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 331         if (IS_ERR(cld)) {
 332                 CERROR("can't create log: %s\n", logname);
 333                 rc = PTR_ERR(cld);
 334                 goto out_err2;
 335         }
 336
 337         cld->cld_sptlrpc = sptlrpc_cld;
 338         cld->cld_params = params_cld;
 339
 340         LASSERT(lsi->lsi_lmd);
 341         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
 342                 struct config_llog_data *recover_cld;
 343
 344                 ptr = strrchr(seclogname, '-');
 345                 if (ptr) {
 346                         *ptr = 0;
 347                 } else {
 348                         CERROR("%s: sptlrpc log name not correct, %s: rc = %d\n",
 349                                obd->obd_name, seclogname, -EINVAL);
 350                         config_log_put(cld);
 351                         return -EINVAL;
 352                 }
 353                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 354                 if (IS_ERR(recover_cld)) {
 355                         rc = PTR_ERR(recover_cld);
 356                         goto out_err3;
 357                 }
 358                 cld->cld_recover = recover_cld;
 359         }
 360
 361         return 0;
 362
 363 out_err3:
 364         config_log_put(cld);
 365
 366 out_err2:
 367         config_log_put(params_cld);
 368
 369 out_err1:
 370         config_log_put(sptlrpc_cld);
 371
 372 out_err:
 373         return rc;
 374 }
 375
 376 DEFINE_MUTEX(llog_process_lock);
 377
 378 /** Stop watching for updates on this log.
 379  */
 380 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 381 {
 382         struct config_llog_data *cld;
 383         struct config_llog_data *cld_sptlrpc = NULL;
 384         struct config_llog_data *cld_params = NULL;
 385         struct config_llog_data *cld_recover = NULL;
 386         int rc = 0;
 387
 388         cld = config_log_find(logname, cfg);
 389         if (!cld)
 390                 return -ENOENT;
 391
 392         mutex_lock(&cld->cld_lock);
 393         /*
 394          * if cld_stopping is set, it means we didn't start the log thus
 395          * not owning the start ref. this can happen after previous umount:
 396          * the cld still hanging there waiting for lock cancel, and we
 397          * remount again but failed in the middle and call log_end without
 398          * calling start_log.
 399          */
 400         if (unlikely(cld->cld_stopping)) {
 401                 mutex_unlock(&cld->cld_lock);
 402                 /* drop the ref from the find */
 403                 config_log_put(cld);
 404                 return rc;
 405         }
 406
 407         cld->cld_stopping = 1;
 408
 409         cld_recover = cld->cld_recover;
 410         cld->cld_recover = NULL;
 411         mutex_unlock(&cld->cld_lock);
 412
 413         if (cld_recover) {
 414                 mutex_lock(&cld_recover->cld_lock);
 415                 cld_recover->cld_stopping = 1;
 416                 mutex_unlock(&cld_recover->cld_lock);
 417                 config_log_put(cld_recover);
 418         }
 419
 420         spin_lock(&config_list_lock);
 421         cld_sptlrpc = cld->cld_sptlrpc;
 422         cld->cld_sptlrpc = NULL;
 423         cld_params = cld->cld_params;
 424         cld->cld_params = NULL;
 425         spin_unlock(&config_list_lock);
 426
 427         if (cld_sptlrpc)
 428                 config_log_put(cld_sptlrpc);
 429
 430         if (cld_params) {
 431                 mutex_lock(&cld_params->cld_lock);
 432                 cld_params->cld_stopping = 1;
 433                 mutex_unlock(&cld_params->cld_lock);
 434                 config_log_put(cld_params);
 435         }
 436
 437         /* drop the ref from the find */
 438         config_log_put(cld);
 439         /* drop the start ref */
 440         config_log_put(cld);
 441
 442         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 443                rc);
 444         return rc;
 445 }
 446
 447 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 448 {
 449         struct obd_device       *obd = data;
 450         struct obd_import       *imp;
 451         struct obd_connect_data *ocd;
 452         struct config_llog_data *cld;
 453         int rc;
 454
 455         rc = lprocfs_climp_check(obd);
 456         if (rc)
 457                 return rc;
 458
 459         imp = obd->u.cli.cl_import;
 460         ocd = &imp->imp_connect_data;
 461
 462         seq_printf(m, "imperative_recovery: %s\n",
 463                    OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 464         seq_printf(m, "client_state:\n");
 465
 466         spin_lock(&config_list_lock);
 467         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 468                 if (!cld->cld_recover)
 469                         continue;
 470                 seq_printf(m, "    - { client: %s, nidtbl_version: %u }\n",
 471                            cld->cld_logname,
 472                            cld->cld_recover->cld_cfg.cfg_last_idx);
 473         }
 474         spin_unlock(&config_list_lock);
 475
 476         up_read(&obd->u.cli.cl_sem);
 477         return 0;
 478 }
 479
 480 /* reenqueue any lost locks */
 481 #define RQ_RUNNING 0x1
 482 #define RQ_NOW     0x2
 483 #define RQ_LATER   0x4
 484 #define RQ_STOP    0x8
 485 #define RQ_PRECLEANUP  0x10
 486 static int rq_state;
 487 static wait_queue_head_t            rq_waitq;
 488 static DECLARE_COMPLETION(rq_exit);
 489 static DECLARE_COMPLETION(rq_start);
 490
 491 static void do_requeue(struct config_llog_data *cld)
 492 {
 493         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 494
 495         /* Do not run mgc_process_log on a disconnected export or an
 496          * export which is being disconnected. Take the client
 497          * semaphore to make the check non-racy.
 498          */
 499         down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
 500                          OBD_CLI_SEM_MGC);
 501
 502         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 503                 int rc;
 504
 505                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 506                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 507                 if (rc && rc != -ENOENT)
 508                         CERROR("failed processing log: %d\n", rc);
 509         } else {
 510                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 511                        cld->cld_logname);
 512         }
 513         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 514 }
 515
 516 /* this timeout represents how many seconds MGC should wait before
 517  * requeue config and recover lock to the MGS. We need to randomize this
 518  * in order to not flood the MGS.
 519  */
 520 #define MGC_TIMEOUT_MIN_SECONDS   5
 521 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
 522
 523 static int mgc_requeue_thread(void *data)
 524 {
 525         bool first = true;
 526
 527         CDEBUG(D_MGC, "Starting requeue thread\n");
 528
 529         /* Keep trying failed locks periodically */
 530         spin_lock(&config_list_lock);
 531         rq_state |= RQ_RUNNING;
 532         while (1) {
 533                 struct l_wait_info lwi;
 534                 struct config_llog_data *cld, *cld_prev;
 535                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
 536                 int stopped = !!(rq_state & RQ_STOP);
 537                 int to;
 538
 539                 /* Any new or requeued lostlocks will change the state */
 540                 rq_state &= ~(RQ_NOW | RQ_LATER);
 541                 spin_unlock(&config_list_lock);
 542
 543                 if (first) {
 544                         first = false;
 545                         complete(&rq_start);
 546                 }
 547
 548                 /* Always wait a few seconds to allow the server who
 549                  * caused the lock revocation to finish its setup, plus some
 550                  * random so everyone doesn't try to reconnect at once.
 551                  */
 552                 to = msecs_to_jiffies(MGC_TIMEOUT_MIN_SECONDS * MSEC_PER_SEC);
 553                 /* rand is centi-seconds */
 554                 to += msecs_to_jiffies(rand * MSEC_PER_SEC / 100);
 555                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 556                 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
 557                              &lwi);
 558
 559                 /*
 560                  * iterate & processing through the list. for each cld, process
 561                  * its depending sptlrpc cld firstly (if any) and then itself.
 562                  *
 563                  * it's guaranteed any item in the list must have
 564                  * reference > 0; and if cld_lostlock is set, at
 565                  * least one reference is taken by the previous enqueue.
 566                  */
 567                 cld_prev = NULL;
 568
 569                 spin_lock(&config_list_lock);
 570                 rq_state &= ~RQ_PRECLEANUP;
 571                 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 572                         if (!cld->cld_lostlock)
 573                                 continue;
 574
 575                         spin_unlock(&config_list_lock);
 576
 577                         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 578
 579                         /* Whether we enqueued again or not in mgc_process_log,
 580                          * we're done with the ref from the old enqueue
 581                          */
 582                         if (cld_prev)
 583                                 config_log_put(cld_prev);
 584                         cld_prev = cld;
 585
 586                         cld->cld_lostlock = 0;
 587                         if (likely(!stopped))
 588                                 do_requeue(cld);
 589
 590                         spin_lock(&config_list_lock);
 591                 }
 592                 spin_unlock(&config_list_lock);
 593                 if (cld_prev)
 594                         config_log_put(cld_prev);
 595
 596                 /* break after scanning the list so that we can drop
 597                  * refcount to losing lock clds
 598                  */
 599                 if (unlikely(stopped)) {
 600                         spin_lock(&config_list_lock);
 601                         break;
 602                 }
 603
 604                 /* Wait a bit to see if anyone else needs a requeue */
 605                 lwi = (struct l_wait_info) { 0 };
 606                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 607                              &lwi);
 608                 spin_lock(&config_list_lock);
 609         }
 610         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 611         rq_state &= ~RQ_RUNNING;
 612         spin_unlock(&config_list_lock);
 613
 614         complete(&rq_exit);
 615
 616         CDEBUG(D_MGC, "Ending requeue thread\n");
 617         return 0;
 618 }
 619
 620 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 621  * We are responsible for dropping the config log reference from here on out.
 622  */
 623 static void mgc_requeue_add(struct config_llog_data *cld)
 624 {
 625         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 626                cld->cld_logname, atomic_read(&cld->cld_refcount),
 627                cld->cld_stopping, rq_state);
 628         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 629
 630         mutex_lock(&cld->cld_lock);
 631         if (cld->cld_stopping || cld->cld_lostlock) {
 632                 mutex_unlock(&cld->cld_lock);
 633                 return;
 634         }
 635         /* this refcount will be released in mgc_requeue_thread. */
 636         config_log_get(cld);
 637         cld->cld_lostlock = 1;
 638         mutex_unlock(&cld->cld_lock);
 639
 640         /* Hold lock for rq_state */
 641         spin_lock(&config_list_lock);
 642         if (rq_state & RQ_STOP) {
 643                 spin_unlock(&config_list_lock);
 644                 cld->cld_lostlock = 0;
 645                 config_log_put(cld);
 646         } else {
 647                 rq_state |= RQ_NOW;
 648                 spin_unlock(&config_list_lock);
 649                 wake_up(&rq_waitq);
 650         }
 651 }
 652
 653 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 654 {
 655         struct llog_ctxt        *ctxt;
 656         int                      rc;
 657
 658         /* setup only remote ctxt, the local disk context is switched per each
 659          * filesystem during mgc_fs_setup()
 660          */
 661         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 662                         &llog_client_ops);
 663         if (rc)
 664                 return rc;
 665
 666         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 667         LASSERT(ctxt);
 668
 669         llog_initiator_connect(ctxt);
 670         llog_ctxt_put(ctxt);
 671
 672         return 0;
 673 }
 674
 675 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 676 {
 677         struct llog_ctxt *ctxt;
 678
 679         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 680         if (ctxt)
 681                 llog_cleanup(env, ctxt);
 682
 683         return 0;
 684 }
 685
 686 static atomic_t mgc_count = ATOMIC_INIT(0);
 687 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 688 {
 689         int rc = 0;
 690         int temp;
 691
 692         switch (stage) {
 693         case OBD_CLEANUP_EARLY:
 694                 break;
 695         case OBD_CLEANUP_EXPORTS:
 696                 if (atomic_dec_and_test(&mgc_count)) {
 697                         LASSERT(rq_state & RQ_RUNNING);
 698                         /* stop requeue thread */
 699                         temp = RQ_STOP;
 700                 } else {
 701                         /* wakeup requeue thread to clean our cld */
 702                         temp = RQ_NOW | RQ_PRECLEANUP;
 703                 }
 704                 spin_lock(&config_list_lock);
 705                 rq_state |= temp;
 706                 spin_unlock(&config_list_lock);
 707                 wake_up(&rq_waitq);
 708                 if (temp & RQ_STOP)
 709                         wait_for_completion(&rq_exit);
 710                 obd_cleanup_client_import(obd);
 711                 rc = mgc_llog_fini(NULL, obd);
 712                 if (rc != 0)
 713                         CERROR("failed to cleanup llogging subsystems\n");
 714                 break;
 715         }
 716         return rc;
 717 }
 718
 719 static int mgc_cleanup(struct obd_device *obd)
 720 {
 721         /* COMPAT_146 - old config logs may have added profiles we don't
 722          * know about
 723          */
 724         if (obd->obd_type->typ_refcnt <= 1)
 725                 /* Only for the last mgc */
 726                 class_del_profiles();
 727
 728         lprocfs_obd_cleanup(obd);
 729         ptlrpcd_decref();
 730
 731         return client_obd_cleanup(obd);
 732 }
 733
 734 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 735 {
 736         struct lprocfs_static_vars lvars = { NULL };
 737         struct task_struct *task;
 738         int rc;
 739
 740         rc = ptlrpcd_addref();
 741         if (rc < 0)
 742                 goto err_noref;
 743
 744         rc = client_obd_setup(obd, lcfg);
 745         if (rc)
 746                 goto err_decref;
 747
 748         rc = mgc_llog_init(NULL, obd);
 749         if (rc) {
 750                 CERROR("failed to setup llogging subsystems\n");
 751                 goto err_cleanup;
 752         }
 753
 754         lprocfs_mgc_init_vars(&lvars);
 755         lprocfs_obd_setup(obd, lvars.obd_vars, lvars.sysfs_vars);
 756         sptlrpc_lprocfs_cliobd_attach(obd);
 757
 758         if (atomic_inc_return(&mgc_count) == 1) {
 759                 rq_state = 0;
 760                 init_waitqueue_head(&rq_waitq);
 761
 762                 /* start requeue thread */
 763                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 764                 if (IS_ERR(task)) {
 765                         rc = PTR_ERR(task);
 766                         CERROR("%s: cannot start requeue thread: rc = %d; no more log updates\n",
 767                                obd->obd_name, rc);
 768                         goto err_cleanup;
 769                 }
 770                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 771                 rc = 0;
 772                 wait_for_completion(&rq_start);
 773         }
 774
 775         return rc;
 776
 777 err_cleanup:
 778         client_obd_cleanup(obd);
 779 err_decref:
 780         ptlrpcd_decref();
 781 err_noref:
 782         return rc;
 783 }
 784
 785 /* based on ll_mdc_blocking_ast */
 786 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 787                             void *data, int flag)
 788 {
 789         struct lustre_handle lockh;
 790         struct config_llog_data *cld = data;
 791         int rc = 0;
 792
 793         switch (flag) {
 794         case LDLM_CB_BLOCKING:
 795                 /* mgs wants the lock, give it up... */
 796                 LDLM_DEBUG(lock, "MGC blocking CB");
 797                 ldlm_lock2handle(lock, &lockh);
 798                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 799                 break;
 800         case LDLM_CB_CANCELING:
 801                 /* We've given up the lock, prepare ourselves to update. */
 802                 LDLM_DEBUG(lock, "MGC cancel CB");
 803
 804                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
 805                        PLDLMRES(lock->l_resource),
 806                        (char *)&lock->l_resource->lr_name.name[0]);
 807
 808                 if (!cld) {
 809                         CDEBUG(D_INFO, "missing data, won't requeue\n");
 810                         break;
 811                 }
 812
 813                 /* held at mgc_process_log(). */
 814                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
 815                 /* Are we done with this log? */
 816                 if (cld->cld_stopping) {
 817                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
 818                                cld->cld_logname);
 819                         config_log_put(cld);
 820                         break;
 821                 }
 822                 /* Make sure not to re-enqueue when the mgc is stopping
 823                  * (we get called from client_disconnect_export)
 824                  */
 825                 if (!lock->l_conn_export ||
 826                     !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
 827                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
 828                                cld->cld_logname);
 829                         config_log_put(cld);
 830                         break;
 831                 }
 832
 833                 /* Re-enqueue now */
 834                 mgc_requeue_add(cld);
 835                 config_log_put(cld);
 836                 break;
 837         default:
 838                 LBUG();
 839         }
 840
 841         return rc;
 842 }
 843
 844 /* Not sure where this should go... */
 845 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
 846  * that we can have a chance to try the secondary MGS if any.
 847  */
 848 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
 849                                 + PING_INTERVAL)
 850 #define  MGC_TARGET_REG_LIMIT 10
 851 #define  MGC_SEND_PARAM_LIMIT 10
 852
 853 /* Send parameter to MGS*/
 854 static int mgc_set_mgs_param(struct obd_export *exp,
 855                              struct mgs_send_param *msp)
 856 {
 857         struct ptlrpc_request *req;
 858         struct mgs_send_param *req_msp, *rep_msp;
 859         int rc;
 860
 861         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 862                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
 863                                         MGS_SET_INFO);
 864         if (!req)
 865                 return -ENOMEM;
 866
 867         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 868         if (!req_msp) {
 869                 ptlrpc_req_finished(req);
 870                 return -ENOMEM;
 871         }
 872
 873         memcpy(req_msp, msp, sizeof(*req_msp));
 874         ptlrpc_request_set_replen(req);
 875
 876         /* Limit how long we will wait for the enqueue to complete */
 877         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
 878         rc = ptlrpc_queue_wait(req);
 879         if (!rc) {
 880                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 881                 memcpy(msp, rep_msp, sizeof(*rep_msp));
 882         }
 883
 884         ptlrpc_req_finished(req);
 885
 886         return rc;
 887 }
 888
 889 /* Take a config lock so we can get cancel notifications */
 890 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
 891                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
 892                        __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
 893                        void *data, __u32 lvb_len, void *lvb_swabber,
 894                        struct lustre_handle *lockh)
 895 {
 896         struct config_llog_data *cld = data;
 897         struct ldlm_enqueue_info einfo = {
 898                 .ei_type        = type,
 899                 .ei_mode        = mode,
 900                 .ei_cb_bl       = mgc_blocking_ast,
 901                 .ei_cb_cp       = ldlm_completion_ast,
 902         };
 903         struct ptlrpc_request *req;
 904         int short_limit = cld_is_sptlrpc(cld);
 905         int rc;
 906
 907         CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
 908                cld->cld_resid.name[0]);
 909
 910         /* We need a callback for every lockholder, so don't try to
 911          * ldlm_lock_match (see rev 1.1.2.11.2.47)
 912          */
 913         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 914                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
 915                                         LDLM_ENQUEUE);
 916         if (!req)
 917                 return -ENOMEM;
 918
 919         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
 920         ptlrpc_request_set_replen(req);
 921
 922         /* Limit how long we will wait for the enqueue to complete */
 923         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
 924         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
 925                               NULL, 0, LVB_T_NONE, lockh, 0);
 926         /* A failed enqueue should still call the mgc_blocking_ast,
 927          * where it will be requeued if needed ("grant failed").
 928          */
 929         ptlrpc_req_finished(req);
 930         return rc;
 931 }
 932
 933 static void mgc_notify_active(struct obd_device *unused)
 934 {
 935         /* wakeup mgc_requeue_thread to requeue mgc lock */
 936         spin_lock(&config_list_lock);
 937         rq_state |= RQ_NOW;
 938         spin_unlock(&config_list_lock);
 939         wake_up(&rq_waitq);
 940
 941         /* TODO: Help the MGS rebuild nidtbl. -jay */
 942 }
 943
 944 /* Send target_reg message to MGS */
 945 static int mgc_target_register(struct obd_export *exp,
 946                                struct mgs_target_info *mti)
 947 {
 948         struct ptlrpc_request  *req;
 949         struct mgs_target_info *req_mti, *rep_mti;
 950         int                  rc;
 951
 952         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 953                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
 954                                         MGS_TARGET_REG);
 955         if (!req)
 956                 return -ENOMEM;
 957
 958         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
 959         if (!req_mti) {
 960                 ptlrpc_req_finished(req);
 961                 return -ENOMEM;
 962         }
 963
 964         memcpy(req_mti, mti, sizeof(*req_mti));
 965         ptlrpc_request_set_replen(req);
 966         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
 967         /* Limit how long we will wait for the enqueue to complete */
 968         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
 969
 970         rc = ptlrpc_queue_wait(req);
 971         if (!rc) {
 972                 rep_mti = req_capsule_server_get(&req->rq_pill,
 973                                                  &RMF_MGS_TARGET_INFO);
 974                 memcpy(mti, rep_mti, sizeof(*rep_mti));
 975                 CDEBUG(D_MGC, "register %s got index = %d\n",
 976                        mti->mti_svname, mti->mti_stripe_index);
 977         }
 978         ptlrpc_req_finished(req);
 979
 980         return rc;
 981 }
 982
 983 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
 984                               u32 keylen, void *key, u32 vallen,
 985                               void *val, struct ptlrpc_request_set *set)
 986 {
 987         int rc = -EINVAL;
 988
 989         /* Turn off initial_recov after we try all backup servers once */
 990         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
 991                 struct obd_import *imp = class_exp2cliimp(exp);
 992                 int value;
 993
 994                 if (vallen != sizeof(int))
 995                         return -EINVAL;
 996                 value = *(int *)val;
 997                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
 998                        imp->imp_obd->obd_name, value,
 999                        imp->imp_deactive, imp->imp_invalid,
1000                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1001                        ptlrpc_import_state_name(imp->imp_state));
1002                 /* Resurrect if we previously died */
1003                 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1004                      imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1005                         ptlrpc_reconnect_import(imp);
1006                 return 0;
1007         }
1008         if (KEY_IS(KEY_SET_INFO)) {
1009                 struct mgs_send_param *msp;
1010
1011                 msp = val;
1012                 rc =  mgc_set_mgs_param(exp, msp);
1013                 return rc;
1014         }
1015         if (KEY_IS(KEY_MGSSEC)) {
1016                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1017                 struct sptlrpc_flavor  flvr;
1018
1019                 /*
1020                  * empty string means using current flavor, if which haven't
1021                  * been set yet, set it as null.
1022                  *
1023                  * if flavor has been set previously, check the asking flavor
1024                  * must match the existing one.
1025                  */
1026                 if (vallen == 0) {
1027                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1028                                 return 0;
1029                         val = "null";
1030                         vallen = 4;
1031                 }
1032
1033                 rc = sptlrpc_parse_flavor(val, &flvr);
1034                 if (rc) {
1035                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1036                                (char *)val);
1037                         return rc;
1038                 }
1039
1040                 /*
1041                  * caller already hold a mutex
1042                  */
1043                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1044                         cli->cl_flvr_mgc = flvr;
1045                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1046                                   sizeof(flvr)) != 0) {
1047                         char    str[20];
1048
1049                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1050                                             str, sizeof(str));
1051                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but currently %s is in use\n",
1052                                        (char *)val, str);
1053                         rc = -EPERM;
1054                 }
1055                 return rc;
1056         }
1057
1058         return rc;
1059 }
1060
1061 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1062                         __u32 keylen, void *key, __u32 *vallen, void *val,
1063                         struct lov_stripe_md *unused)
1064 {
1065         int rc = -EINVAL;
1066
1067         if (KEY_IS(KEY_CONN_DATA)) {
1068                 struct obd_import *imp = class_exp2cliimp(exp);
1069                 struct obd_connect_data *data = val;
1070
1071                 if (*vallen == sizeof(*data)) {
1072                         *data = imp->imp_connect_data;
1073                         rc = 0;
1074                 }
1075         }
1076
1077         return rc;
1078 }
1079
1080 static int mgc_import_event(struct obd_device *obd,
1081                             struct obd_import *imp,
1082                             enum obd_import_event event)
1083 {
1084         LASSERT(imp->imp_obd == obd);
1085         CDEBUG(D_MGC, "import event %#x\n", event);
1086
1087         switch (event) {
1088         case IMP_EVENT_DISCON:
1089                 /* MGC imports should not wait for recovery */
1090                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1091                         ptlrpc_pinger_ir_down();
1092                 break;
1093         case IMP_EVENT_INACTIVE:
1094                 break;
1095         case IMP_EVENT_INVALIDATE: {
1096                 struct ldlm_namespace *ns = obd->obd_namespace;
1097
1098                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1099                 break;
1100         }
1101         case IMP_EVENT_ACTIVE:
1102                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1103                 /* Clearing obd_no_recov allows us to continue pinging */
1104                 obd->obd_no_recov = 0;
1105                 mgc_notify_active(obd);
1106                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1107                         ptlrpc_pinger_ir_up();
1108                 break;
1109         case IMP_EVENT_OCD:
1110                 break;
1111         case IMP_EVENT_DEACTIVATE:
1112         case IMP_EVENT_ACTIVATE:
1113                 break;
1114         default:
1115                 CERROR("Unknown import event %#x\n", event);
1116                 LBUG();
1117         }
1118         return 0;
1119 }
1120
1121 enum {
1122         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1123         CONFIG_READ_NRPAGES      = 4
1124 };
1125
1126 static int mgc_apply_recover_logs(struct obd_device *mgc,
1127                                   struct config_llog_data *cld,
1128                                   __u64 max_version,
1129                                   void *data, int datalen, bool mne_swab)
1130 {
1131         struct config_llog_instance *cfg = &cld->cld_cfg;
1132         struct mgs_nidtbl_entry *entry;
1133         struct lustre_cfg       *lcfg;
1134         struct lustre_cfg_bufs   bufs;
1135         u64   prev_version = 0;
1136         char *inst;
1137         char *buf;
1138         int   bufsz;
1139         int   pos;
1140         int   rc  = 0;
1141         int   off = 0;
1142
1143         LASSERT(cfg->cfg_instance);
1144         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1145
1146         inst = kzalloc(PAGE_SIZE, GFP_KERNEL);
1147         if (!inst)
1148                 return -ENOMEM;
1149
1150         pos = snprintf(inst, PAGE_SIZE, "%p", cfg->cfg_instance);
1151         if (pos >= PAGE_SIZE) {
1152                 kfree(inst);
1153                 return -E2BIG;
1154         }
1155
1156         ++pos;
1157         buf   = inst + pos;
1158         bufsz = PAGE_SIZE - pos;
1159
1160         while (datalen > 0) {
1161                 int   entry_len = sizeof(*entry);
1162                 int is_ost, i;
1163                 struct obd_device *obd;
1164                 char *obdname;
1165                 char *cname;
1166                 char *params;
1167                 char *uuid;
1168
1169                 rc = -EINVAL;
1170                 if (datalen < sizeof(*entry))
1171                         break;
1172
1173                 entry = (typeof(entry))(data + off);
1174
1175                 /* sanity check */
1176                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1177                         break;
1178                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1179                         break;
1180                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1181                         break;
1182
1183                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1184                 if (datalen < entry_len) /* must have entry_len at least */
1185                         break;
1186
1187                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1188                 if (mne_swab)
1189                         lustre_swab_mgs_nidtbl_entry(entry);
1190                 if (entry->mne_length > PAGE_SIZE) {
1191                         CERROR("MNE too large (%u)\n", entry->mne_length);
1192                         break;
1193                 }
1194
1195                 if (entry->mne_length < entry_len)
1196                         break;
1197
1198                 off     += entry->mne_length;
1199                 datalen -= entry->mne_length;
1200                 if (datalen < 0)
1201                         break;
1202
1203                 if (entry->mne_version > max_version) {
1204                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1205                                entry->mne_version, max_version);
1206                         break;
1207                 }
1208
1209                 if (prev_version >= entry->mne_version) {
1210                         CERROR("index unsorted, prev %lld, now %lld\n",
1211                                prev_version, entry->mne_version);
1212                         break;
1213                 }
1214                 prev_version = entry->mne_version;
1215
1216                 /*
1217                  * Write a string with format "nid::instance" to
1218                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1219                  */
1220
1221                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1222                 memset(buf, 0, bufsz);
1223                 obdname = buf;
1224                 pos = 0;
1225
1226                 /* lustre-OST0001-osc-<instance #> */
1227                 strcpy(obdname, cld->cld_logname);
1228                 cname = strrchr(obdname, '-');
1229                 if (!cname) {
1230                         CERROR("mgc %s: invalid logname %s\n",
1231                                mgc->obd_name, obdname);
1232                         break;
1233                 }
1234
1235                 pos = cname - obdname;
1236                 obdname[pos] = 0;
1237                 pos += sprintf(obdname + pos, "-%s%04x",
1238                                   is_ost ? "OST" : "MDT", entry->mne_index);
1239
1240                 cname = is_ost ? "osc" : "mdc";
1241                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1242                 lustre_cfg_bufs_reset(&bufs, obdname);
1243
1244                 /* find the obd by obdname */
1245                 obd = class_name2obd(obdname);
1246                 if (!obd) {
1247                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1248                                mgc->obd_name, obdname);
1249                         rc = 0;
1250                         /* this is a safe race, when the ost is starting up...*/
1251                         continue;
1252                 }
1253
1254                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1255                 ++pos;
1256                 params = buf + pos;
1257                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1258                 uuid = buf + pos;
1259
1260                 down_read(&obd->u.cli.cl_sem);
1261                 if (!obd->u.cli.cl_import) {
1262                         /* client does not connect to the OST yet */
1263                         up_read(&obd->u.cli.cl_sem);
1264                         rc = 0;
1265                         continue;
1266                 }
1267
1268                 /* iterate all nids to find one */
1269                 /* find uuid by nid */
1270                 rc = -ENOENT;
1271                 for (i = 0; i < entry->mne_nid_count; i++) {
1272                         rc = client_import_find_conn(obd->u.cli.cl_import,
1273                                                      entry->u.nids[0],
1274                                                      (struct obd_uuid *)uuid);
1275                         if (!rc)
1276                                 break;
1277                 }
1278
1279                 up_read(&obd->u.cli.cl_sem);
1280                 if (rc < 0) {
1281                         CERROR("mgc: cannot find uuid by nid %s\n",
1282                                libcfs_nid2str(entry->u.nids[0]));
1283                         break;
1284                 }
1285
1286                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1287                        uuid, libcfs_nid2str(entry->u.nids[0]));
1288
1289                 pos += strlen(uuid);
1290                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1291                 LASSERT(pos < bufsz);
1292
1293                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1294
1295                 rc = -ENOMEM;
1296                 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1297                 if (IS_ERR(lcfg)) {
1298                         CERROR("mgc: cannot allocate memory\n");
1299                         break;
1300                 }
1301
1302                 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1303                        prev_version, max_version, obdname, params);
1304
1305                 rc = class_process_config(lcfg);
1306                 lustre_cfg_free(lcfg);
1307                 if (rc)
1308                         CDEBUG(D_INFO, "process config for %s error %d\n",
1309                                obdname, rc);
1310
1311                 /* continue, even one with error */
1312         }
1313
1314         kfree(inst);
1315         return rc;
1316 }
1317
1318 /**
1319  * This function is called if this client was notified for target restarting
1320  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1321  */
1322 static int mgc_process_recover_log(struct obd_device *obd,
1323                                    struct config_llog_data *cld)
1324 {
1325         struct ptlrpc_request *req = NULL;
1326         struct config_llog_instance *cfg = &cld->cld_cfg;
1327         struct mgs_config_body *body;
1328         struct mgs_config_res  *res;
1329         struct ptlrpc_bulk_desc *desc;
1330         struct page **pages;
1331         int nrpages;
1332         bool eof = true;
1333         bool mne_swab;
1334         int i;
1335         int ealen;
1336         int rc;
1337
1338         /* allocate buffer for bulk transfer.
1339          * if this is the first time for this mgs to read logs,
1340          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1341          * once; otherwise, it only reads increment of logs, this should be
1342          * small and CONFIG_READ_NRPAGES will be used.
1343          */
1344         nrpages = CONFIG_READ_NRPAGES;
1345         if (cfg->cfg_last_idx == 0) /* the first time */
1346                 nrpages = CONFIG_READ_NRPAGES_INIT;
1347
1348         pages = kcalloc(nrpages, sizeof(*pages), GFP_KERNEL);
1349         if (!pages) {
1350                 rc = -ENOMEM;
1351                 goto out;
1352         }
1353
1354         for (i = 0; i < nrpages; i++) {
1355                 pages[i] = alloc_page(GFP_KERNEL);
1356                 if (!pages[i]) {
1357                         rc = -ENOMEM;
1358                         goto out;
1359                 }
1360         }
1361
1362 again:
1363         LASSERT(cld_is_recover(cld));
1364         LASSERT(mutex_is_locked(&cld->cld_lock));
1365         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1366                                    &RQF_MGS_CONFIG_READ);
1367         if (!req) {
1368                 rc = -ENOMEM;
1369                 goto out;
1370         }
1371
1372         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1373         if (rc)
1374                 goto out;
1375
1376         /* pack request */
1377         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1378         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1379         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1380             >= sizeof(body->mcb_name)) {
1381                 rc = -E2BIG;
1382                 goto out;
1383         }
1384         body->mcb_offset = cfg->cfg_last_idx + 1;
1385         body->mcb_type   = cld->cld_type;
1386         body->mcb_bits   = PAGE_SHIFT;
1387         body->mcb_units  = nrpages;
1388
1389         /* allocate bulk transfer descriptor */
1390         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1, BULK_PUT_SINK,
1391                                     MGS_BULK_PORTAL);
1392         if (!desc) {
1393                 rc = -ENOMEM;
1394                 goto out;
1395         }
1396
1397         for (i = 0; i < nrpages; i++)
1398                 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_SIZE);
1399
1400         ptlrpc_request_set_replen(req);
1401         rc = ptlrpc_queue_wait(req);
1402         if (rc)
1403                 goto out;
1404
1405         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1406         if (res->mcr_size < res->mcr_offset) {
1407                 rc = -EINVAL;
1408                 goto out;
1409         }
1410
1411         /* always update the index even though it might have errors with
1412          * handling the recover logs
1413          */
1414         cfg->cfg_last_idx = res->mcr_offset;
1415         eof = res->mcr_offset == res->mcr_size;
1416
1417         CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1418                res->mcr_offset, eof == false);
1419
1420         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1421         if (ealen < 0) {
1422                 rc = ealen;
1423                 goto out;
1424         }
1425
1426         if (ealen > nrpages << PAGE_SHIFT) {
1427                 rc = -EINVAL;
1428                 goto out;
1429         }
1430
1431         if (ealen == 0) { /* no logs transferred */
1432                 if (!eof)
1433                         rc = -EINVAL;
1434                 goto out;
1435         }
1436
1437         mne_swab = !!ptlrpc_rep_need_swab(req);
1438 #if OBD_OCD_VERSION(3, 0, 53, 0) > LUSTRE_VERSION_CODE
1439         /* This import flag means the server did an extra swab of IR MNE
1440          * records (fixed in LU-1252), reverse it here if needed. LU-1644
1441          */
1442         if (unlikely(req->rq_import->imp_need_mne_swab))
1443                 mne_swab = !mne_swab;
1444 #endif
1445
1446         for (i = 0; i < nrpages && ealen > 0; i++) {
1447                 int rc2;
1448                 void *ptr;
1449
1450                 ptr = kmap(pages[i]);
1451                 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1452                                              min_t(int, ealen, PAGE_SIZE),
1453                                              mne_swab);
1454                 kunmap(pages[i]);
1455                 if (rc2 < 0) {
1456                         CWARN("Process recover log %s error %d\n",
1457                               cld->cld_logname, rc2);
1458                         break;
1459                 }
1460
1461                 ealen -= PAGE_SIZE;
1462         }
1463
1464 out:
1465         if (req)
1466                 ptlrpc_req_finished(req);
1467
1468         if (rc == 0 && !eof)
1469                 goto again;
1470
1471         if (pages) {
1472                 for (i = 0; i < nrpages; i++) {
1473                         if (!pages[i])
1474                                 break;
1475                         __free_page(pages[i]);
1476                 }
1477                 kfree(pages);
1478         }
1479         return rc;
1480 }
1481
1482 /* local_only means it cannot get remote llogs */
1483 static int mgc_process_cfg_log(struct obd_device *mgc,
1484                                struct config_llog_data *cld, int local_only)
1485 {
1486         struct llog_ctxt        *ctxt;
1487         struct lustre_sb_info   *lsi = NULL;
1488         int                      rc = 0;
1489         bool                     sptlrpc_started = false;
1490         struct lu_env           *env;
1491
1492         LASSERT(cld);
1493         LASSERT(mutex_is_locked(&cld->cld_lock));
1494
1495         /*
1496          * local copy of sptlrpc log is controlled elsewhere, don't try to
1497          * read it up here.
1498          */
1499         if (cld_is_sptlrpc(cld) && local_only)
1500                 return 0;
1501
1502         if (cld->cld_cfg.cfg_sb)
1503                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1504
1505         env = kzalloc(sizeof(*env), GFP_KERNEL);
1506         if (!env)
1507                 return -ENOMEM;
1508
1509         rc = lu_env_init(env, LCT_MG_THREAD);
1510         if (rc)
1511                 goto out_free;
1512
1513         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1514         LASSERT(ctxt);
1515
1516         if (local_only) /* no local log at client side */ {
1517                 rc = -EIO;
1518                 goto out_pop;
1519         }
1520
1521         if (cld_is_sptlrpc(cld)) {
1522                 sptlrpc_conf_log_update_begin(cld->cld_logname);
1523                 sptlrpc_started = true;
1524         }
1525
1526         /* logname and instance info should be the same, so use our
1527          * copy of the instance for the update.  The cfg_last_idx will
1528          * be updated here.
1529          */
1530         rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1531                                      &cld->cld_cfg);
1532
1533 out_pop:
1534         __llog_ctxt_put(env, ctxt);
1535
1536         /*
1537          * update settings on existing OBDs. doing it inside
1538          * of llog_process_lock so no device is attaching/detaching
1539          * in parallel.
1540          * the logname must be <fsname>-sptlrpc
1541          */
1542         if (sptlrpc_started) {
1543                 LASSERT(cld_is_sptlrpc(cld));
1544                 sptlrpc_conf_log_update_end(cld->cld_logname);
1545                 class_notify_sptlrpc_conf(cld->cld_logname,
1546                                           strlen(cld->cld_logname) -
1547                                           strlen("-sptlrpc"));
1548         }
1549
1550         lu_env_fini(env);
1551 out_free:
1552         kfree(env);
1553         return rc;
1554 }
1555
1556 /** Get a config log from the MGS and process it.
1557  * This func is called for both clients and servers.
1558  * Copy the log locally before parsing it if appropriate (non-MGS server)
1559  */
1560 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1561 {
1562         struct lustre_handle lockh = { 0 };
1563         __u64 flags = LDLM_FL_NO_LRU;
1564         int rc = 0, rcl;
1565
1566         LASSERT(cld);
1567
1568         /* I don't want multiple processes running process_log at once --
1569          * sounds like badness.  It actually might be fine, as long as
1570          * we're not trying to update from the same log
1571          * simultaneously (in which case we should use a per-log sem.)
1572          */
1573         mutex_lock(&cld->cld_lock);
1574         if (cld->cld_stopping) {
1575                 mutex_unlock(&cld->cld_lock);
1576                 return 0;
1577         }
1578
1579         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1580
1581         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1582                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1583
1584         /* Get the cfg lock on the llog */
1585         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1586                           LCK_CR, &flags, NULL, NULL, NULL,
1587                           cld, 0, NULL, &lockh);
1588         if (rcl == 0) {
1589                 /* Get the cld, it will be released in mgc_blocking_ast. */
1590                 config_log_get(cld);
1591                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1592                 LASSERT(rc == 0);
1593         } else {
1594                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1595
1596                 /* mark cld_lostlock so that it will requeue
1597                  * after MGC becomes available.
1598                  */
1599                 cld->cld_lostlock = 1;
1600                 /* Get extra reference, it will be put in requeue thread */
1601                 config_log_get(cld);
1602         }
1603
1604         if (cld_is_recover(cld)) {
1605                 rc = 0; /* this is not a fatal error for recover log */
1606                 if (rcl == 0)
1607                         rc = mgc_process_recover_log(mgc, cld);
1608         } else {
1609                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1610         }
1611
1612         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1613                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1614
1615         mutex_unlock(&cld->cld_lock);
1616
1617         /* Now drop the lock so MGS can revoke it */
1618         if (!rcl)
1619                 ldlm_lock_decref(&lockh, LCK_CR);
1620
1621         return rc;
1622 }
1623
1624 /** Called from lustre_process_log.
1625  * LCFG_LOG_START gets the config log from the MGS, processes it to start
1626  * any services, and adds it to the list logs to watch (follow).
1627  */
1628 static int mgc_process_config(struct obd_device *obd, u32 len, void *buf)
1629 {
1630         struct lustre_cfg *lcfg = buf;
1631         struct config_llog_instance *cfg = NULL;
1632         char *logname;
1633         int rc = 0;
1634
1635         switch (lcfg->lcfg_command) {
1636         case LCFG_LOV_ADD_OBD: {
1637                 /* Overloading this cfg command: register a new target */
1638                 struct mgs_target_info *mti;
1639
1640                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1641                     sizeof(struct mgs_target_info)) {
1642                         rc = -EINVAL;
1643                         goto out;
1644                 }
1645
1646                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1647                 CDEBUG(D_MGC, "add_target %s %#x\n",
1648                        mti->mti_svname, mti->mti_flags);
1649                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1650                 break;
1651         }
1652         case LCFG_LOV_DEL_OBD:
1653                 /* Unregister has no meaning at the moment. */
1654                 CERROR("lov_del_obd unimplemented\n");
1655                 rc = -ENOSYS;
1656                 break;
1657         case LCFG_SPTLRPC_CONF: {
1658                 rc = sptlrpc_process_config(lcfg);
1659                 break;
1660         }
1661         case LCFG_LOG_START: {
1662                 struct config_llog_data *cld;
1663                 struct super_block *sb;
1664
1665                 logname = lustre_cfg_string(lcfg, 1);
1666                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1667                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1668
1669                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1670                        cfg->cfg_last_idx);
1671
1672                 /* We're only called through here on the initial mount */
1673                 rc = config_log_add(obd, logname, cfg, sb);
1674                 if (rc)
1675                         break;
1676                 cld = config_log_find(logname, cfg);
1677                 if (!cld) {
1678                         rc = -ENOENT;
1679                         break;
1680                 }
1681
1682                 /* COMPAT_146 */
1683                 /* FIXME only set this for old logs!  Right now this forces
1684                  * us to always skip the "inside markers" check
1685                  */
1686                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1687
1688                 rc = mgc_process_log(obd, cld);
1689                 if (rc == 0 && cld->cld_recover) {
1690                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1691                                          imp_connect_data, IMP_RECOV)) {
1692                                 rc = mgc_process_log(obd, cld->cld_recover);
1693                         } else {
1694                                 struct config_llog_data *cir = cld->cld_recover;
1695
1696                                 cld->cld_recover = NULL;
1697                                 config_log_put(cir);
1698                         }
1699                         if (rc)
1700                                 CERROR("Cannot process recover llog %d\n", rc);
1701                 }
1702
1703                 if (rc == 0 && cld->cld_params) {
1704                         rc = mgc_process_log(obd, cld->cld_params);
1705                         if (rc == -ENOENT) {
1706                                 CDEBUG(D_MGC,
1707                                        "There is no params config file yet\n");
1708                                 rc = 0;
1709                         }
1710                         /* params log is optional */
1711                         if (rc)
1712                                 CERROR(
1713                                        "%s: can't process params llog: rc = %d\n",
1714                                        obd->obd_name, rc);
1715                 }
1716                 config_log_put(cld);
1717
1718                 break;
1719         }
1720         case LCFG_LOG_END: {
1721                 logname = lustre_cfg_string(lcfg, 1);
1722
1723                 if (lcfg->lcfg_bufcount >= 2)
1724                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
1725                                 lcfg, 2);
1726                 rc = config_log_end(logname, cfg);
1727                 break;
1728         }
1729         default: {
1730                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1731                 rc = -EINVAL;
1732                 goto out;
1733         }
1734         }
1735 out:
1736         return rc;
1737 }
1738
1739 static struct obd_ops mgc_obd_ops = {
1740         .owner          = THIS_MODULE,
1741         .setup          = mgc_setup,
1742         .precleanup     = mgc_precleanup,
1743         .cleanup        = mgc_cleanup,
1744         .add_conn       = client_import_add_conn,
1745         .del_conn       = client_import_del_conn,
1746         .connect        = client_connect_import,
1747         .disconnect     = client_disconnect_export,
1748         .set_info_async = mgc_set_info_async,
1749         .get_info       = mgc_get_info,
1750         .import_event   = mgc_import_event,
1751         .process_config = mgc_process_config,
1752 };
1753
1754 static int __init mgc_init(void)
1755 {
1756         return class_register_type(&mgc_obd_ops, NULL,
1757                                    LUSTRE_MGC_NAME, NULL);
1758 }
1759
1760 static void /*__exit*/ mgc_exit(void)
1761 {
1762         class_unregister_type(LUSTRE_MGC_NAME);
1763 }
1764
1765 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1766 MODULE_DESCRIPTION("Lustre Management Client");
1767 MODULE_VERSION(LUSTRE_VERSION_STRING);
1768 MODULE_LICENSE("GPL");
1769
1770 module_init(mgc_init);
1771 module_exit(mgc_exit);