kernel/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.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, 2012, 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 "../include/obd_class.h"
  46 #include "../include/lustre_dlm.h"
  47 #include "../include/lprocfs_status.h"
  48 #include "../include/lustre_log.h"
  49 #include "../include/lustre_disk.h"
  50
  51 #include "mgc_internal.h"
  52
  53 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
  54                           int type)
  55 {
  56         __u64 resname = 0;
  57
  58         if (len > sizeof(resname)) {
  59                 CERROR("name too long: %s\n", name);
  60                 return -EINVAL;
  61         }
  62         if (len <= 0) {
  63                 CERROR("missing name: %s\n", name);
  64                 return -EINVAL;
  65         }
  66         memcpy(&resname, name, len);
  67
  68         /* Always use the same endianness for the resid */
  69         memset(res_id, 0, sizeof(*res_id));
  70         res_id->name[0] = cpu_to_le64(resname);
  71         /* XXX: unfortunately, sptlprc and config llog share one lock */
  72         switch (type) {
  73         case CONFIG_T_CONFIG:
  74         case CONFIG_T_SPTLRPC:
  75                 resname = 0;
  76                 break;
  77         case CONFIG_T_RECOVER:
  78         case CONFIG_T_PARAMS:
  79                 resname = type;
  80                 break;
  81         default:
  82                 LBUG();
  83         }
  84         res_id->name[1] = cpu_to_le64(resname);
  85         CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
  86                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  87         return 0;
  88 }
  89
  90 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
  91 {
  92         /* fsname is at most 8 chars long, maybe contain "-".
  93          * e.g. "lustre", "SUN-000" */
  94         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
  95 }
  96 EXPORT_SYMBOL(mgc_fsname2resid);
  97
  98 static int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
  99 {
 100         char *name_end;
 101         int len;
 102
 103         /* logname consists of "fsname-nodetype".
 104          * e.g. "lustre-MDT0001", "SUN-000-client"
 105          * there is an exception: llog "params" */
 106         name_end = strrchr(logname, '-');
 107         if (!name_end)
 108                 len = strlen(logname);
 109         else
 110                 len = name_end - logname;
 111         return mgc_name2resid(logname, len, res_id, type);
 112 }
 113
 114 /********************** config llog list **********************/
 115 static LIST_HEAD(config_llog_list);
 116 static DEFINE_SPINLOCK(config_list_lock);
 117
 118 /* Take a reference to a config log */
 119 static int config_log_get(struct config_llog_data *cld)
 120 {
 121         atomic_inc(&cld->cld_refcount);
 122         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 123                atomic_read(&cld->cld_refcount));
 124         return 0;
 125 }
 126
 127 /* Drop a reference to a config log.  When no longer referenced,
 128    we can free the config log data */
 129 static void config_log_put(struct config_llog_data *cld)
 130 {
 131         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 132                atomic_read(&cld->cld_refcount));
 133         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 134
 135         /* spinlock to make sure no item with 0 refcount in the list */
 136         if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 137                 list_del(&cld->cld_list_chain);
 138                 spin_unlock(&config_list_lock);
 139
 140                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 141
 142                 if (cld->cld_recover)
 143                         config_log_put(cld->cld_recover);
 144                 if (cld->cld_sptlrpc)
 145                         config_log_put(cld->cld_sptlrpc);
 146                 if (cld->cld_params)
 147                         config_log_put(cld->cld_params);
 148                 if (cld_is_sptlrpc(cld))
 149                         sptlrpc_conf_log_stop(cld->cld_logname);
 150
 151                 class_export_put(cld->cld_mgcexp);
 152                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 153         }
 154 }
 155
 156 /* Find a config log by name */
 157 static
 158 struct config_llog_data *config_log_find(char *logname,
 159                                          struct config_llog_instance *cfg)
 160 {
 161         struct config_llog_data *cld;
 162         struct config_llog_data *found = NULL;
 163         void *instance;
 164
 165         LASSERT(logname != NULL);
 166
 167         instance = cfg ? cfg->cfg_instance : NULL;
 168         spin_lock(&config_list_lock);
 169         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 170                 /* check if instance equals */
 171                 if (instance != cld->cld_cfg.cfg_instance)
 172                         continue;
 173
 174                 /* instance may be NULL, should check name */
 175                 if (strcmp(logname, cld->cld_logname) == 0) {
 176                         found = cld;
 177                         break;
 178                 }
 179         }
 180         if (found) {
 181                 atomic_inc(&found->cld_refcount);
 182                 LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0);
 183         }
 184         spin_unlock(&config_list_lock);
 185         return found;
 186 }
 187
 188 static
 189 struct config_llog_data *do_config_log_add(struct obd_device *obd,
 190                                            char *logname,
 191                                            int type,
 192                                            struct config_llog_instance *cfg,
 193                                            struct super_block *sb)
 194 {
 195         struct config_llog_data *cld;
 196         int                   rc;
 197
 198         CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
 199                cfg ? cfg->cfg_instance : NULL);
 200
 201         OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
 202         if (!cld)
 203                 return ERR_PTR(-ENOMEM);
 204
 205         strcpy(cld->cld_logname, logname);
 206         if (cfg)
 207                 cld->cld_cfg = *cfg;
 208         else
 209                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 210         mutex_init(&cld->cld_lock);
 211         cld->cld_cfg.cfg_last_idx = 0;
 212         cld->cld_cfg.cfg_flags = 0;
 213         cld->cld_cfg.cfg_sb = sb;
 214         cld->cld_type = type;
 215         atomic_set(&cld->cld_refcount, 1);
 216
 217         /* Keep the mgc around until we are done */
 218         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 219
 220         if (cld_is_sptlrpc(cld)) {
 221                 sptlrpc_conf_log_start(logname);
 222                 cld->cld_cfg.cfg_obdname = obd->obd_name;
 223         }
 224
 225         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 226
 227         spin_lock(&config_list_lock);
 228         list_add(&cld->cld_list_chain, &config_llog_list);
 229         spin_unlock(&config_list_lock);
 230
 231         if (rc) {
 232                 config_log_put(cld);
 233                 return ERR_PTR(rc);
 234         }
 235
 236         if (cld_is_sptlrpc(cld)) {
 237                 rc = mgc_process_log(obd, cld);
 238                 if (rc && rc != -ENOENT)
 239                         CERROR("failed processing sptlrpc log: %d\n", rc);
 240         }
 241
 242         return cld;
 243 }
 244
 245 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
 246         char *fsname,
 247         struct config_llog_instance *cfg,
 248         struct super_block *sb)
 249 {
 250         struct config_llog_instance lcfg = *cfg;
 251         struct lustre_sb_info *lsi = s2lsi(sb);
 252         struct config_llog_data *cld;
 253         char logname[32];
 254
 255         if (IS_OST(lsi))
 256                 return NULL;
 257
 258         /* for osp-on-ost, see lustre_start_osp() */
 259         if (IS_MDT(lsi) && lcfg.cfg_instance)
 260                 return NULL;
 261
 262         /* we have to use different llog for clients and mdts for cmd
 263          * where only clients are notified if one of cmd server restarts */
 264         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 265         strcpy(logname, fsname);
 266         if (IS_SERVER(lsi)) { /* mdt */
 267                 LASSERT(lcfg.cfg_instance == NULL);
 268                 lcfg.cfg_instance = sb;
 269                 strcat(logname, "-mdtir");
 270         } else {
 271                 LASSERT(lcfg.cfg_instance != NULL);
 272                 strcat(logname, "-cliir");
 273         }
 274
 275         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
 276         return cld;
 277 }
 278
 279 static struct config_llog_data *config_params_log_add(struct obd_device *obd,
 280         struct config_llog_instance *cfg, struct super_block *sb)
 281 {
 282         struct config_llog_instance     lcfg = *cfg;
 283         struct config_llog_data         *cld;
 284
 285         lcfg.cfg_instance = sb;
 286
 287         cld = do_config_log_add(obd, PARAMS_FILENAME, CONFIG_T_PARAMS,
 288                                 &lcfg, sb);
 289
 290         return cld;
 291 }
 292
 293 /** Add this log to the list of active logs watched by an MGC.
 294  * Active means we're watching for updates.
 295  * We have one active log per "mount" - client instance or servername.
 296  * Each instance may be at a different point in the log.
 297  */
 298 static int config_log_add(struct obd_device *obd, char *logname,
 299                           struct config_llog_instance *cfg,
 300                           struct super_block *sb)
 301 {
 302         struct lustre_sb_info *lsi = s2lsi(sb);
 303         struct config_llog_data *cld;
 304         struct config_llog_data *sptlrpc_cld;
 305         struct config_llog_data *params_cld;
 306         char                    seclogname[32];
 307         char                    *ptr;
 308         int                     rc;
 309
 310         CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
 311
 312         /*
 313          * for each regular log, the depended sptlrpc log name is
 314          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 315          */
 316         ptr = strrchr(logname, '-');
 317         if (ptr == NULL || ptr - logname > 8) {
 318                 CERROR("logname %s is too long\n", logname);
 319                 return -EINVAL;
 320         }
 321
 322         memcpy(seclogname, logname, ptr - logname);
 323         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 324
 325         sptlrpc_cld = config_log_find(seclogname, NULL);
 326         if (sptlrpc_cld == NULL) {
 327                 sptlrpc_cld = do_config_log_add(obd, seclogname,
 328                                                 CONFIG_T_SPTLRPC, NULL, NULL);
 329                 if (IS_ERR(sptlrpc_cld)) {
 330                         CERROR("can't create sptlrpc log: %s\n", seclogname);
 331                         rc = PTR_ERR(sptlrpc_cld);
 332                         goto out_err;
 333                 }
 334         }
 335         params_cld = config_params_log_add(obd, cfg, sb);
 336         if (IS_ERR(params_cld)) {
 337                 rc = PTR_ERR(params_cld);
 338                 CERROR("%s: can't create params log: rc = %d\n",
 339                        obd->obd_name, rc);
 340                 goto out_err1;
 341         }
 342
 343         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 344         if (IS_ERR(cld)) {
 345                 CERROR("can't create log: %s\n", logname);
 346                 rc = PTR_ERR(cld);
 347                 goto out_err2;
 348         }
 349
 350         cld->cld_sptlrpc = sptlrpc_cld;
 351         cld->cld_params = params_cld;
 352
 353         LASSERT(lsi->lsi_lmd);
 354         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
 355                 struct config_llog_data *recover_cld;
 356                 *strrchr(seclogname, '-') = 0;
 357                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 358                 if (IS_ERR(recover_cld)) {
 359                         rc = PTR_ERR(recover_cld);
 360                         goto out_err3;
 361                 }
 362                 cld->cld_recover = recover_cld;
 363         }
 364
 365         return 0;
 366
 367 out_err3:
 368         config_log_put(cld);
 369
 370 out_err2:
 371         config_log_put(params_cld);
 372
 373 out_err1:
 374         config_log_put(sptlrpc_cld);
 375
 376 out_err:
 377         return rc;
 378 }
 379
 380 DEFINE_MUTEX(llog_process_lock);
 381
 382 /** Stop watching for updates on this log.
 383  */
 384 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 385 {
 386         struct config_llog_data *cld;
 387         struct config_llog_data *cld_sptlrpc = NULL;
 388         struct config_llog_data *cld_params = NULL;
 389         struct config_llog_data *cld_recover = NULL;
 390         int rc = 0;
 391
 392         cld = config_log_find(logname, cfg);
 393         if (cld == NULL)
 394                 return -ENOENT;
 395
 396         mutex_lock(&cld->cld_lock);
 397         /*
 398          * if cld_stopping is set, it means we didn't start the log thus
 399          * not owning the start ref. this can happen after previous umount:
 400          * the cld still hanging there waiting for lock cancel, and we
 401          * remount again but failed in the middle and call log_end without
 402          * calling start_log.
 403          */
 404         if (unlikely(cld->cld_stopping)) {
 405                 mutex_unlock(&cld->cld_lock);
 406                 /* drop the ref from the find */
 407                 config_log_put(cld);
 408                 return rc;
 409         }
 410
 411         cld->cld_stopping = 1;
 412
 413         cld_recover = cld->cld_recover;
 414         cld->cld_recover = NULL;
 415         mutex_unlock(&cld->cld_lock);
 416
 417         if (cld_recover) {
 418                 mutex_lock(&cld_recover->cld_lock);
 419                 cld_recover->cld_stopping = 1;
 420                 mutex_unlock(&cld_recover->cld_lock);
 421                 config_log_put(cld_recover);
 422         }
 423
 424         spin_lock(&config_list_lock);
 425         cld_sptlrpc = cld->cld_sptlrpc;
 426         cld->cld_sptlrpc = NULL;
 427         cld_params = cld->cld_params;
 428         cld->cld_params = NULL;
 429         spin_unlock(&config_list_lock);
 430
 431         if (cld_sptlrpc)
 432                 config_log_put(cld_sptlrpc);
 433
 434         if (cld_params) {
 435                 mutex_lock(&cld_params->cld_lock);
 436                 cld_params->cld_stopping = 1;
 437                 mutex_unlock(&cld_params->cld_lock);
 438                 config_log_put(cld_params);
 439         }
 440
 441         /* drop the ref from the find */
 442         config_log_put(cld);
 443         /* drop the start ref */
 444         config_log_put(cld);
 445
 446         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 447                rc);
 448         return rc;
 449 }
 450
 451 #if defined (CONFIG_PROC_FS)
 452 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 453 {
 454         struct obd_device       *obd = data;
 455         struct obd_import       *imp;
 456         struct obd_connect_data *ocd;
 457         struct config_llog_data *cld;
 458
 459         LPROCFS_CLIMP_CHECK(obd);
 460         imp = obd->u.cli.cl_import;
 461         ocd = &imp->imp_connect_data;
 462
 463         seq_printf(m, "imperative_recovery: %s\n",
 464                       OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 465         seq_printf(m, "client_state:\n");
 466
 467         spin_lock(&config_list_lock);
 468         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 469                 if (cld->cld_recover == NULL)
 470                         continue;
 471                 seq_printf(m,  "    - { client: %s, nidtbl_version: %u }\n",
 472                                cld->cld_logname,
 473                                cld->cld_recover->cld_cfg.cfg_last_idx);
 474         }
 475         spin_unlock(&config_list_lock);
 476
 477         LPROCFS_CLIMP_EXIT(obd);
 478         return 0;
 479 }
 480 #endif
 481
 482 /* reenqueue any lost locks */
 483 #define RQ_RUNNING 0x1
 484 #define RQ_NOW     0x2
 485 #define RQ_LATER   0x4
 486 #define RQ_STOP    0x8
 487 #define RQ_PRECLEANUP  0x10
 488 static int rq_state;
 489 static wait_queue_head_t            rq_waitq;
 490 static DECLARE_COMPLETION(rq_exit);
 491 static DECLARE_COMPLETION(rq_start);
 492
 493 static void do_requeue(struct config_llog_data *cld)
 494 {
 495         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 496
 497         /* Do not run mgc_process_log on a disconnected export or an
 498            export which is being disconnected. Take the client
 499            semaphore to make the check non-racy. */
 500         down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 501         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 502                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 503                 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 504         } else {
 505                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 506                        cld->cld_logname);
 507         }
 508         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 509 }
 510
 511 /* this timeout represents how many seconds MGC should wait before
 512  * requeue config and recover lock to the MGS. We need to randomize this
 513  * in order to not flood the MGS.
 514  */
 515 #define MGC_TIMEOUT_MIN_SECONDS   5
 516 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
 517
 518 static int mgc_requeue_thread(void *data)
 519 {
 520         bool first = true;
 521
 522         CDEBUG(D_MGC, "Starting requeue thread\n");
 523
 524         /* Keep trying failed locks periodically */
 525         spin_lock(&config_list_lock);
 526         rq_state |= RQ_RUNNING;
 527         while (1) {
 528                 struct l_wait_info lwi;
 529                 struct config_llog_data *cld, *cld_prev;
 530                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
 531                 int stopped = !!(rq_state & RQ_STOP);
 532                 int to;
 533
 534                 /* Any new or requeued lostlocks will change the state */
 535                 rq_state &= ~(RQ_NOW | RQ_LATER);
 536                 spin_unlock(&config_list_lock);
 537
 538                 if (first) {
 539                         first = false;
 540                         complete(&rq_start);
 541                 }
 542
 543                 /* Always wait a few seconds to allow the server who
 544                    caused the lock revocation to finish its setup, plus some
 545                    random so everyone doesn't try to reconnect at once. */
 546                 to = MGC_TIMEOUT_MIN_SECONDS * HZ;
 547                 to += rand * HZ / 100; /* rand is centi-seconds */
 548                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 549                 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
 550                              &lwi);
 551
 552                 /*
 553                  * iterate & processing through the list. for each cld, process
 554                  * its depending sptlrpc cld firstly (if any) and then itself.
 555                  *
 556                  * it's guaranteed any item in the list must have
 557                  * reference > 0; and if cld_lostlock is set, at
 558                  * least one reference is taken by the previous enqueue.
 559                  */
 560                 cld_prev = NULL;
 561
 562                 spin_lock(&config_list_lock);
 563                 rq_state &= ~RQ_PRECLEANUP;
 564                 list_for_each_entry(cld, &config_llog_list,
 565                                         cld_list_chain) {
 566                         if (!cld->cld_lostlock)
 567                                 continue;
 568
 569                         spin_unlock(&config_list_lock);
 570
 571                         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 572
 573                         /* Whether we enqueued again or not in mgc_process_log,
 574                          * we're done with the ref from the old enqueue */
 575                         if (cld_prev)
 576                                 config_log_put(cld_prev);
 577                         cld_prev = cld;
 578
 579                         cld->cld_lostlock = 0;
 580                         if (likely(!stopped))
 581                                 do_requeue(cld);
 582
 583                         spin_lock(&config_list_lock);
 584                 }
 585                 spin_unlock(&config_list_lock);
 586                 if (cld_prev)
 587                         config_log_put(cld_prev);
 588
 589                 /* break after scanning the list so that we can drop
 590                  * refcount to losing lock clds */
 591                 if (unlikely(stopped)) {
 592                         spin_lock(&config_list_lock);
 593                         break;
 594                 }
 595
 596                 /* Wait a bit to see if anyone else needs a requeue */
 597                 lwi = (struct l_wait_info) { 0 };
 598                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 599                              &lwi);
 600                 spin_lock(&config_list_lock);
 601         }
 602         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 603         rq_state &= ~RQ_RUNNING;
 604         spin_unlock(&config_list_lock);
 605
 606         complete(&rq_exit);
 607
 608         CDEBUG(D_MGC, "Ending requeue thread\n");
 609         return 0;
 610 }
 611
 612 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 613    We are responsible for dropping the config log reference from here on out. */
 614 static void mgc_requeue_add(struct config_llog_data *cld)
 615 {
 616         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 617                cld->cld_logname, atomic_read(&cld->cld_refcount),
 618                cld->cld_stopping, rq_state);
 619         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 620
 621         mutex_lock(&cld->cld_lock);
 622         if (cld->cld_stopping || cld->cld_lostlock) {
 623                 mutex_unlock(&cld->cld_lock);
 624                 return;
 625         }
 626         /* this refcount will be released in mgc_requeue_thread. */
 627         config_log_get(cld);
 628         cld->cld_lostlock = 1;
 629         mutex_unlock(&cld->cld_lock);
 630
 631         /* Hold lock for rq_state */
 632         spin_lock(&config_list_lock);
 633         if (rq_state & RQ_STOP) {
 634                 spin_unlock(&config_list_lock);
 635                 cld->cld_lostlock = 0;
 636                 config_log_put(cld);
 637         } else {
 638                 rq_state |= RQ_NOW;
 639                 spin_unlock(&config_list_lock);
 640                 wake_up(&rq_waitq);
 641         }
 642 }
 643
 644 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 645 {
 646         struct llog_ctxt        *ctxt;
 647         int                      rc;
 648
 649         /* setup only remote ctxt, the local disk context is switched per each
 650          * filesystem during mgc_fs_setup() */
 651         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 652                         &llog_client_ops);
 653         if (rc)
 654                 return rc;
 655
 656         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 657         LASSERT(ctxt);
 658
 659         llog_initiator_connect(ctxt);
 660         llog_ctxt_put(ctxt);
 661
 662         return 0;
 663 }
 664
 665 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 666 {
 667         struct llog_ctxt *ctxt;
 668
 669         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 670         if (ctxt)
 671                 llog_cleanup(env, ctxt);
 672
 673         return 0;
 674 }
 675
 676 static atomic_t mgc_count = ATOMIC_INIT(0);
 677 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 678 {
 679         int rc = 0;
 680         int temp;
 681
 682         switch (stage) {
 683         case OBD_CLEANUP_EARLY:
 684                 break;
 685         case OBD_CLEANUP_EXPORTS:
 686                 if (atomic_dec_and_test(&mgc_count)) {
 687                         LASSERT(rq_state & RQ_RUNNING);
 688                         /* stop requeue thread */
 689                         temp = RQ_STOP;
 690                 } else {
 691                         /* wakeup requeue thread to clean our cld */
 692                         temp = RQ_NOW | RQ_PRECLEANUP;
 693                 }
 694                 spin_lock(&config_list_lock);
 695                 rq_state |= temp;
 696                 spin_unlock(&config_list_lock);
 697                 wake_up(&rq_waitq);
 698                 if (temp & RQ_STOP)
 699                         wait_for_completion(&rq_exit);
 700                 obd_cleanup_client_import(obd);
 701                 rc = mgc_llog_fini(NULL, obd);
 702                 if (rc != 0)
 703                         CERROR("failed to cleanup llogging subsystems\n");
 704                 break;
 705         }
 706         return rc;
 707 }
 708
 709 static int mgc_cleanup(struct obd_device *obd)
 710 {
 711         /* COMPAT_146 - old config logs may have added profiles we don't
 712            know about */
 713         if (obd->obd_type->typ_refcnt <= 1)
 714                 /* Only for the last mgc */
 715                 class_del_profiles();
 716
 717         lprocfs_obd_cleanup(obd);
 718         ptlrpcd_decref();
 719
 720         return client_obd_cleanup(obd);
 721 }
 722
 723 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 724 {
 725         struct lprocfs_static_vars lvars;
 726         int rc;
 727
 728         ptlrpcd_addref();
 729
 730         rc = client_obd_setup(obd, lcfg);
 731         if (rc)
 732                 goto err_decref;
 733
 734         rc = mgc_llog_init(NULL, obd);
 735         if (rc) {
 736                 CERROR("failed to setup llogging subsystems\n");
 737                 goto err_cleanup;
 738         }
 739
 740         lprocfs_mgc_init_vars(&lvars);
 741         lprocfs_obd_setup(obd, lvars.obd_vars);
 742         sptlrpc_lprocfs_cliobd_attach(obd);
 743
 744         if (atomic_inc_return(&mgc_count) == 1) {
 745                 rq_state = 0;
 746                 init_waitqueue_head(&rq_waitq);
 747
 748                 /* start requeue thread */
 749                 rc = PTR_ERR(kthread_run(mgc_requeue_thread, NULL,
 750                                              "ll_cfg_requeue"));
 751                 if (IS_ERR_VALUE(rc)) {
 752                         CERROR("%s: Cannot start requeue thread (%d),no more log updates!\n",
 753                                obd->obd_name, rc);
 754                         goto err_cleanup;
 755                 }
 756                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 757                 rc = 0;
 758                 wait_for_completion(&rq_start);
 759         }
 760
 761         return rc;
 762
 763 err_cleanup:
 764         client_obd_cleanup(obd);
 765 err_decref:
 766         ptlrpcd_decref();
 767         return rc;
 768 }
 769
 770 /* based on ll_mdc_blocking_ast */
 771 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 772                             void *data, int flag)
 773 {
 774         struct lustre_handle lockh;
 775         struct config_llog_data *cld = (struct config_llog_data *)data;
 776         int rc = 0;
 777
 778         switch (flag) {
 779         case LDLM_CB_BLOCKING:
 780                 /* mgs wants the lock, give it up... */
 781                 LDLM_DEBUG(lock, "MGC blocking CB");
 782                 ldlm_lock2handle(lock, &lockh);
 783                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 784                 break;
 785         case LDLM_CB_CANCELING:
 786                 /* We've given up the lock, prepare ourselves to update. */
 787                 LDLM_DEBUG(lock, "MGC cancel CB");
 788
 789                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
 790                        PLDLMRES(lock->l_resource),
 791                        (char *)&lock->l_resource->lr_name.name[0]);
 792
 793                 if (!cld) {
 794                         CDEBUG(D_INFO, "missing data, won't requeue\n");
 795                         break;
 796                 }
 797
 798                 /* held at mgc_process_log(). */
 799                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
 800                 /* Are we done with this log? */
 801                 if (cld->cld_stopping) {
 802                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
 803                                cld->cld_logname);
 804                         config_log_put(cld);
 805                         break;
 806                 }
 807                 /* Make sure not to re-enqueue when the mgc is stopping
 808                    (we get called from client_disconnect_export) */
 809                 if (!lock->l_conn_export ||
 810                     !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
 811                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
 812                                cld->cld_logname);
 813                         config_log_put(cld);
 814                         break;
 815                 }
 816
 817                 /* Re-enqueue now */
 818                 mgc_requeue_add(cld);
 819                 config_log_put(cld);
 820                 break;
 821         default:
 822                 LBUG();
 823         }
 824
 825         return rc;
 826 }
 827
 828 /* Not sure where this should go... */
 829 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
 830  * that we can have a chance to try the secondary MGS if any. */
 831 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
 832                                 + PING_INTERVAL)
 833 #define  MGC_TARGET_REG_LIMIT 10
 834 #define  MGC_SEND_PARAM_LIMIT 10
 835
 836 /* Send parameter to MGS*/
 837 static int mgc_set_mgs_param(struct obd_export *exp,
 838                              struct mgs_send_param *msp)
 839 {
 840         struct ptlrpc_request *req;
 841         struct mgs_send_param *req_msp, *rep_msp;
 842         int rc;
 843
 844         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 845                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
 846                                         MGS_SET_INFO);
 847         if (!req)
 848                 return -ENOMEM;
 849
 850         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 851         if (!req_msp) {
 852                 ptlrpc_req_finished(req);
 853                 return -ENOMEM;
 854         }
 855
 856         memcpy(req_msp, msp, sizeof(*req_msp));
 857         ptlrpc_request_set_replen(req);
 858
 859         /* Limit how long we will wait for the enqueue to complete */
 860         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
 861         rc = ptlrpc_queue_wait(req);
 862         if (!rc) {
 863                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 864                 memcpy(msp, rep_msp, sizeof(*rep_msp));
 865         }
 866
 867         ptlrpc_req_finished(req);
 868
 869         return rc;
 870 }
 871
 872 /* Take a config lock so we can get cancel notifications */
 873 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
 874                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
 875                        __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
 876                        void *data, __u32 lvb_len, void *lvb_swabber,
 877                        struct lustre_handle *lockh)
 878 {
 879         struct config_llog_data *cld = (struct config_llog_data *)data;
 880         struct ldlm_enqueue_info einfo = {
 881                 .ei_type        = type,
 882                 .ei_mode        = mode,
 883                 .ei_cb_bl       = mgc_blocking_ast,
 884                 .ei_cb_cp       = ldlm_completion_ast,
 885         };
 886         struct ptlrpc_request *req;
 887         int short_limit = cld_is_sptlrpc(cld);
 888         int rc;
 889
 890         CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
 891                cld->cld_resid.name[0]);
 892
 893         /* We need a callback for every lockholder, so don't try to
 894            ldlm_lock_match (see rev 1.1.2.11.2.47) */
 895         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 896                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
 897                                         LDLM_ENQUEUE);
 898         if (req == NULL)
 899                 return -ENOMEM;
 900
 901         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
 902         ptlrpc_request_set_replen(req);
 903
 904         /* check if this is server or client */
 905         if (cld->cld_cfg.cfg_sb) {
 906                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
 907                 if (lsi && IS_SERVER(lsi))
 908                         short_limit = 1;
 909         }
 910         /* Limit how long we will wait for the enqueue to complete */
 911         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
 912         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
 913                               NULL, 0, LVB_T_NONE, lockh, 0);
 914         /* A failed enqueue should still call the mgc_blocking_ast,
 915            where it will be requeued if needed ("grant failed"). */
 916         ptlrpc_req_finished(req);
 917         return rc;
 918 }
 919
 920 static void mgc_notify_active(struct obd_device *unused)
 921 {
 922         /* wakeup mgc_requeue_thread to requeue mgc lock */
 923         spin_lock(&config_list_lock);
 924         rq_state |= RQ_NOW;
 925         spin_unlock(&config_list_lock);
 926         wake_up(&rq_waitq);
 927
 928         /* TODO: Help the MGS rebuild nidtbl. -jay */
 929 }
 930
 931 /* Send target_reg message to MGS */
 932 static int mgc_target_register(struct obd_export *exp,
 933                                struct mgs_target_info *mti)
 934 {
 935         struct ptlrpc_request  *req;
 936         struct mgs_target_info *req_mti, *rep_mti;
 937         int                  rc;
 938
 939         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 940                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
 941                                         MGS_TARGET_REG);
 942         if (req == NULL)
 943                 return -ENOMEM;
 944
 945         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
 946         if (!req_mti) {
 947                 ptlrpc_req_finished(req);
 948                 return -ENOMEM;
 949         }
 950
 951         memcpy(req_mti, mti, sizeof(*req_mti));
 952         ptlrpc_request_set_replen(req);
 953         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
 954         /* Limit how long we will wait for the enqueue to complete */
 955         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
 956
 957         rc = ptlrpc_queue_wait(req);
 958         if (!rc) {
 959                 rep_mti = req_capsule_server_get(&req->rq_pill,
 960                                                  &RMF_MGS_TARGET_INFO);
 961                 memcpy(mti, rep_mti, sizeof(*rep_mti));
 962                 CDEBUG(D_MGC, "register %s got index = %d\n",
 963                        mti->mti_svname, mti->mti_stripe_index);
 964         }
 965         ptlrpc_req_finished(req);
 966
 967         return rc;
 968 }
 969
 970 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
 971                        u32 keylen, void *key, u32 vallen,
 972                        void *val, struct ptlrpc_request_set *set)
 973 {
 974         int rc = -EINVAL;
 975
 976         /* Turn off initial_recov after we try all backup servers once */
 977         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
 978                 struct obd_import *imp = class_exp2cliimp(exp);
 979                 int value;
 980                 if (vallen != sizeof(int))
 981                         return -EINVAL;
 982                 value = *(int *)val;
 983                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
 984                        imp->imp_obd->obd_name, value,
 985                        imp->imp_deactive, imp->imp_invalid,
 986                        imp->imp_replayable, imp->imp_obd->obd_replayable,
 987                        ptlrpc_import_state_name(imp->imp_state));
 988                 /* Resurrect if we previously died */
 989                 if ((imp->imp_state != LUSTRE_IMP_FULL &&
 990                      imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
 991                         ptlrpc_reconnect_import(imp);
 992                 return 0;
 993         }
 994         if (KEY_IS(KEY_SET_INFO)) {
 995                 struct mgs_send_param *msp;
 996
 997                 msp = (struct mgs_send_param *)val;
 998                 rc =  mgc_set_mgs_param(exp, msp);
 999                 return rc;
1000         }
1001         if (KEY_IS(KEY_MGSSEC)) {
1002                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1003                 struct sptlrpc_flavor  flvr;
1004
1005                 /*
1006                  * empty string means using current flavor, if which haven't
1007                  * been set yet, set it as null.
1008                  *
1009                  * if flavor has been set previously, check the asking flavor
1010                  * must match the existing one.
1011                  */
1012                 if (vallen == 0) {
1013                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1014                                 return 0;
1015                         val = "null";
1016                         vallen = 4;
1017                 }
1018
1019                 rc = sptlrpc_parse_flavor(val, &flvr);
1020                 if (rc) {
1021                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1022                                (char *) val);
1023                         return rc;
1024                 }
1025
1026                 /*
1027                  * caller already hold a mutex
1028                  */
1029                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1030                         cli->cl_flvr_mgc = flvr;
1031                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1032                                   sizeof(flvr)) != 0) {
1033                         char    str[20];
1034
1035                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1036                                             str, sizeof(str));
1037                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but currently %s is in use\n",
1038                                        (char *) val, str);
1039                         rc = -EPERM;
1040                 }
1041                 return rc;
1042         }
1043
1044         return rc;
1045 }
1046
1047 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1048                         __u32 keylen, void *key, __u32 *vallen, void *val,
1049                         struct lov_stripe_md *unused)
1050 {
1051         int rc = -EINVAL;
1052
1053         if (KEY_IS(KEY_CONN_DATA)) {
1054                 struct obd_import *imp = class_exp2cliimp(exp);
1055                 struct obd_connect_data *data = val;
1056
1057                 if (*vallen == sizeof(*data)) {
1058                         *data = imp->imp_connect_data;
1059                         rc = 0;
1060                 }
1061         }
1062
1063         return rc;
1064 }
1065
1066 static int mgc_import_event(struct obd_device *obd,
1067                             struct obd_import *imp,
1068                             enum obd_import_event event)
1069 {
1070         LASSERT(imp->imp_obd == obd);
1071         CDEBUG(D_MGC, "import event %#x\n", event);
1072
1073         switch (event) {
1074         case IMP_EVENT_DISCON:
1075                 /* MGC imports should not wait for recovery */
1076                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1077                         ptlrpc_pinger_ir_down();
1078                 break;
1079         case IMP_EVENT_INACTIVE:
1080                 break;
1081         case IMP_EVENT_INVALIDATE: {
1082                 struct ldlm_namespace *ns = obd->obd_namespace;
1083                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1084                 break;
1085         }
1086         case IMP_EVENT_ACTIVE:
1087                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1088                 /* Clearing obd_no_recov allows us to continue pinging */
1089                 obd->obd_no_recov = 0;
1090                 mgc_notify_active(obd);
1091                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1092                         ptlrpc_pinger_ir_up();
1093                 break;
1094         case IMP_EVENT_OCD:
1095                 break;
1096         case IMP_EVENT_DEACTIVATE:
1097         case IMP_EVENT_ACTIVATE:
1098                 break;
1099         default:
1100                 CERROR("Unknown import event %#x\n", event);
1101                 LBUG();
1102         }
1103         return 0;
1104 }
1105
1106 enum {
1107         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_CACHE_SHIFT),
1108         CONFIG_READ_NRPAGES      = 4
1109 };
1110
1111 static int mgc_apply_recover_logs(struct obd_device *mgc,
1112                                   struct config_llog_data *cld,
1113                                   __u64 max_version,
1114                                   void *data, int datalen, bool mne_swab)
1115 {
1116         struct config_llog_instance *cfg = &cld->cld_cfg;
1117         struct lustre_sb_info       *lsi = s2lsi(cfg->cfg_sb);
1118         struct mgs_nidtbl_entry *entry;
1119         struct lustre_cfg       *lcfg;
1120         struct lustre_cfg_bufs   bufs;
1121         u64   prev_version = 0;
1122         char *inst;
1123         char *buf;
1124         int   bufsz;
1125         int   pos;
1126         int   rc  = 0;
1127         int   off = 0;
1128
1129         LASSERT(cfg->cfg_instance != NULL);
1130         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1131
1132         OBD_ALLOC(inst, PAGE_CACHE_SIZE);
1133         if (inst == NULL)
1134                 return -ENOMEM;
1135
1136         if (!IS_SERVER(lsi)) {
1137                 pos = snprintf(inst, PAGE_CACHE_SIZE, "%p", cfg->cfg_instance);
1138                 if (pos >= PAGE_CACHE_SIZE) {
1139                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1140                         return -E2BIG;
1141                 }
1142         } else {
1143                 LASSERT(IS_MDT(lsi));
1144                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1145                                         PAGE_CACHE_SIZE);
1146                 if (rc) {
1147                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1148                         return -EINVAL;
1149                 }
1150                 pos = strlen(inst);
1151         }
1152
1153         ++pos;
1154         buf   = inst + pos;
1155         bufsz = PAGE_CACHE_SIZE - pos;
1156
1157         while (datalen > 0) {
1158                 int   entry_len = sizeof(*entry);
1159                 int   is_ost;
1160                 struct obd_device *obd;
1161                 char *obdname;
1162                 char *cname;
1163                 char *params;
1164                 char *uuid;
1165
1166                 rc = -EINVAL;
1167                 if (datalen < sizeof(*entry))
1168                         break;
1169
1170                 entry = (typeof(entry))(data + off);
1171
1172                 /* sanity check */
1173                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1174                         break;
1175                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1176                         break;
1177                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1178                         break;
1179
1180                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1181                 if (datalen < entry_len) /* must have entry_len at least */
1182                         break;
1183
1184                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1185                 if (mne_swab)
1186                         lustre_swab_mgs_nidtbl_entry(entry);
1187                 if (entry->mne_length > PAGE_CACHE_SIZE) {
1188                         CERROR("MNE too large (%u)\n", entry->mne_length);
1189                         break;
1190                 }
1191
1192                 if (entry->mne_length < entry_len)
1193                         break;
1194
1195                 off     += entry->mne_length;
1196                 datalen -= entry->mne_length;
1197                 if (datalen < 0)
1198                         break;
1199
1200                 if (entry->mne_version > max_version) {
1201                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1202                                entry->mne_version, max_version);
1203                         break;
1204                 }
1205
1206                 if (prev_version >= entry->mne_version) {
1207                         CERROR("index unsorted, prev %lld, now %lld\n",
1208                                prev_version, entry->mne_version);
1209                         break;
1210                 }
1211                 prev_version = entry->mne_version;
1212
1213                 /*
1214                  * Write a string with format "nid::instance" to
1215                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1216                  */
1217
1218                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1219                 memset(buf, 0, bufsz);
1220                 obdname = buf;
1221                 pos = 0;
1222
1223                 /* lustre-OST0001-osc-<instance #> */
1224                 strcpy(obdname, cld->cld_logname);
1225                 cname = strrchr(obdname, '-');
1226                 if (cname == NULL) {
1227                         CERROR("mgc %s: invalid logname %s\n",
1228                                mgc->obd_name, obdname);
1229                         break;
1230                 }
1231
1232                 pos = cname - obdname;
1233                 obdname[pos] = 0;
1234                 pos += sprintf(obdname + pos, "-%s%04x",
1235                                   is_ost ? "OST" : "MDT", entry->mne_index);
1236
1237                 cname = is_ost ? "osc" : "mdc",
1238                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1239                 lustre_cfg_bufs_reset(&bufs, obdname);
1240
1241                 /* find the obd by obdname */
1242                 obd = class_name2obd(obdname);
1243                 if (obd == NULL) {
1244                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1245                                mgc->obd_name, obdname);
1246                         rc = 0;
1247                         /* this is a safe race, when the ost is starting up...*/
1248                         continue;
1249                 }
1250
1251                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1252                 ++pos;
1253                 params = buf + pos;
1254                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1255                 uuid = buf + pos;
1256
1257                 down_read(&obd->u.cli.cl_sem);
1258                 if (obd->u.cli.cl_import == NULL) {
1259                         /* client does not connect to the OST yet */
1260                         up_read(&obd->u.cli.cl_sem);
1261                         rc = 0;
1262                         continue;
1263                 }
1264
1265                 /* TODO: iterate all nids to find one */
1266                 /* find uuid by nid */
1267                 rc = client_import_find_conn(obd->u.cli.cl_import,
1268                                              entry->u.nids[0],
1269                                              (struct obd_uuid *)uuid);
1270                 up_read(&obd->u.cli.cl_sem);
1271                 if (rc < 0) {
1272                         CERROR("mgc: cannot find uuid by nid %s\n",
1273                                libcfs_nid2str(entry->u.nids[0]));
1274                         break;
1275                 }
1276
1277                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1278                        uuid, libcfs_nid2str(entry->u.nids[0]));
1279
1280                 pos += strlen(uuid);
1281                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1282                 LASSERT(pos < bufsz);
1283
1284                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1285
1286                 rc = -ENOMEM;
1287                 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1288                 if (lcfg == NULL) {
1289                         CERROR("mgc: cannot allocate memory\n");
1290                         break;
1291                 }
1292
1293                 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1294                        prev_version, max_version, obdname, params);
1295
1296                 rc = class_process_config(lcfg);
1297                 lustre_cfg_free(lcfg);
1298                 if (rc)
1299                         CDEBUG(D_INFO, "process config for %s error %d\n",
1300                                obdname, rc);
1301
1302                 /* continue, even one with error */
1303         }
1304
1305         OBD_FREE(inst, PAGE_CACHE_SIZE);
1306         return rc;
1307 }
1308
1309 /**
1310  * This function is called if this client was notified for target restarting
1311  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1312  */
1313 static int mgc_process_recover_log(struct obd_device *obd,
1314                                    struct config_llog_data *cld)
1315 {
1316         struct ptlrpc_request *req = NULL;
1317         struct config_llog_instance *cfg = &cld->cld_cfg;
1318         struct mgs_config_body *body;
1319         struct mgs_config_res  *res;
1320         struct ptlrpc_bulk_desc *desc;
1321         struct page **pages;
1322         int nrpages;
1323         bool eof = true;
1324         bool mne_swab = false;
1325         int i;
1326         int ealen;
1327         int rc;
1328
1329         /* allocate buffer for bulk transfer.
1330          * if this is the first time for this mgs to read logs,
1331          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1332          * once; otherwise, it only reads increment of logs, this should be
1333          * small and CONFIG_READ_NRPAGES will be used.
1334          */
1335         nrpages = CONFIG_READ_NRPAGES;
1336         if (cfg->cfg_last_idx == 0) /* the first time */
1337                 nrpages = CONFIG_READ_NRPAGES_INIT;
1338
1339         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1340         if (pages == NULL) {
1341                 rc = -ENOMEM;
1342                 goto out;
1343         }
1344
1345         for (i = 0; i < nrpages; i++) {
1346                 pages[i] = alloc_page(GFP_IOFS);
1347                 if (pages[i] == NULL) {
1348                         rc = -ENOMEM;
1349                         goto out;
1350                 }
1351         }
1352
1353 again:
1354         LASSERT(cld_is_recover(cld));
1355         LASSERT(mutex_is_locked(&cld->cld_lock));
1356         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1357                                    &RQF_MGS_CONFIG_READ);
1358         if (req == NULL) {
1359                 rc = -ENOMEM;
1360                 goto out;
1361         }
1362
1363         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1364         if (rc)
1365                 goto out;
1366
1367         /* pack request */
1368         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1369         LASSERT(body != NULL);
1370         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1371         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1372             >= sizeof(body->mcb_name)) {
1373                 rc = -E2BIG;
1374                 goto out;
1375         }
1376         body->mcb_offset = cfg->cfg_last_idx + 1;
1377         body->mcb_type   = cld->cld_type;
1378         body->mcb_bits   = PAGE_CACHE_SHIFT;
1379         body->mcb_units  = nrpages;
1380
1381         /* allocate bulk transfer descriptor */
1382         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1, BULK_PUT_SINK,
1383                                     MGS_BULK_PORTAL);
1384         if (desc == NULL) {
1385                 rc = -ENOMEM;
1386                 goto out;
1387         }
1388
1389         for (i = 0; i < nrpages; i++)
1390                 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
1391
1392         ptlrpc_request_set_replen(req);
1393         rc = ptlrpc_queue_wait(req);
1394         if (rc)
1395                 goto out;
1396
1397         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1398         if (res->mcr_size < res->mcr_offset) {
1399                 rc = -EINVAL;
1400                 goto out;
1401         }
1402
1403         /* always update the index even though it might have errors with
1404          * handling the recover logs */
1405         cfg->cfg_last_idx = res->mcr_offset;
1406         eof = res->mcr_offset == res->mcr_size;
1407
1408         CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1409                res->mcr_offset, eof == false);
1410
1411         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1412         if (ealen < 0) {
1413                 rc = ealen;
1414                 goto out;
1415         }
1416
1417         if (ealen > nrpages << PAGE_CACHE_SHIFT) {
1418                 rc = -EINVAL;
1419                 goto out;
1420         }
1421
1422         if (ealen == 0) { /* no logs transferred */
1423                 if (!eof)
1424                         rc = -EINVAL;
1425                 goto out;
1426         }
1427
1428         mne_swab = !!ptlrpc_rep_need_swab(req);
1429 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1430         /* This import flag means the server did an extra swab of IR MNE
1431          * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1432         if (unlikely(req->rq_import->imp_need_mne_swab))
1433                 mne_swab = !mne_swab;
1434 #else
1435 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1436 #endif
1437
1438         for (i = 0; i < nrpages && ealen > 0; i++) {
1439                 int rc2;
1440                 void *ptr;
1441
1442                 ptr = kmap(pages[i]);
1443                 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1444                                              min_t(int, ealen, PAGE_CACHE_SIZE),
1445                                              mne_swab);
1446                 kunmap(pages[i]);
1447                 if (rc2 < 0) {
1448                         CWARN("Process recover log %s error %d\n",
1449                               cld->cld_logname, rc2);
1450                         break;
1451                 }
1452
1453                 ealen -= PAGE_CACHE_SIZE;
1454         }
1455
1456 out:
1457         if (req)
1458                 ptlrpc_req_finished(req);
1459
1460         if (rc == 0 && !eof)
1461                 goto again;
1462
1463         if (pages) {
1464                 for (i = 0; i < nrpages; i++) {
1465                         if (pages[i] == NULL)
1466                                 break;
1467                         __free_page(pages[i]);
1468                 }
1469                 OBD_FREE(pages, sizeof(*pages) * nrpages);
1470         }
1471         return rc;
1472 }
1473
1474 /* local_only means it cannot get remote llogs */
1475 static int mgc_process_cfg_log(struct obd_device *mgc,
1476                                struct config_llog_data *cld, int local_only)
1477 {
1478         struct llog_ctxt        *ctxt;
1479         struct lustre_sb_info   *lsi = NULL;
1480         int                      rc = 0;
1481         bool                     sptlrpc_started = false;
1482         struct lu_env           *env;
1483
1484         LASSERT(cld);
1485         LASSERT(mutex_is_locked(&cld->cld_lock));
1486
1487         /*
1488          * local copy of sptlrpc log is controlled elsewhere, don't try to
1489          * read it up here.
1490          */
1491         if (cld_is_sptlrpc(cld) && local_only)
1492                 return 0;
1493
1494         if (cld->cld_cfg.cfg_sb)
1495                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1496
1497         OBD_ALLOC_PTR(env);
1498         if (env == NULL)
1499                 return -ENOMEM;
1500
1501         rc = lu_env_init(env, LCT_MG_THREAD);
1502         if (rc)
1503                 goto out_free;
1504
1505         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1506         LASSERT(ctxt);
1507
1508         if (local_only) /* no local log at client side */ {
1509                 rc = -EIO;
1510                 goto out_pop;
1511         }
1512
1513         if (cld_is_sptlrpc(cld)) {
1514                 sptlrpc_conf_log_update_begin(cld->cld_logname);
1515                 sptlrpc_started = true;
1516         }
1517
1518         /* logname and instance info should be the same, so use our
1519          * copy of the instance for the update.  The cfg_last_idx will
1520          * be updated here. */
1521         rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1522                                      &cld->cld_cfg);
1523
1524 out_pop:
1525         __llog_ctxt_put(env, ctxt);
1526
1527         /*
1528          * update settings on existing OBDs. doing it inside
1529          * of llog_process_lock so no device is attaching/detaching
1530          * in parallel.
1531          * the logname must be <fsname>-sptlrpc
1532          */
1533         if (sptlrpc_started) {
1534                 LASSERT(cld_is_sptlrpc(cld));
1535                 sptlrpc_conf_log_update_end(cld->cld_logname);
1536                 class_notify_sptlrpc_conf(cld->cld_logname,
1537                                           strlen(cld->cld_logname) -
1538                                           strlen("-sptlrpc"));
1539         }
1540
1541         lu_env_fini(env);
1542 out_free:
1543         OBD_FREE_PTR(env);
1544         return rc;
1545 }
1546
1547 /** Get a config log from the MGS and process it.
1548  * This func is called for both clients and servers.
1549  * Copy the log locally before parsing it if appropriate (non-MGS server)
1550  */
1551 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1552 {
1553         struct lustre_handle lockh = { 0 };
1554         __u64 flags = LDLM_FL_NO_LRU;
1555         int rc = 0, rcl;
1556
1557         LASSERT(cld);
1558
1559         /* I don't want multiple processes running process_log at once --
1560            sounds like badness.  It actually might be fine, as long as
1561            we're not trying to update from the same log
1562            simultaneously (in which case we should use a per-log sem.) */
1563         mutex_lock(&cld->cld_lock);
1564         if (cld->cld_stopping) {
1565                 mutex_unlock(&cld->cld_lock);
1566                 return 0;
1567         }
1568
1569         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1570
1571         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1572                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1573
1574         /* Get the cfg lock on the llog */
1575         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1576                           LCK_CR, &flags, NULL, NULL, NULL,
1577                           cld, 0, NULL, &lockh);
1578         if (rcl == 0) {
1579                 /* Get the cld, it will be released in mgc_blocking_ast. */
1580                 config_log_get(cld);
1581                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1582                 LASSERT(rc == 0);
1583         } else {
1584                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1585
1586                 /* mark cld_lostlock so that it will requeue
1587                  * after MGC becomes available. */
1588                 cld->cld_lostlock = 1;
1589                 /* Get extra reference, it will be put in requeue thread */
1590                 config_log_get(cld);
1591         }
1592
1593
1594         if (cld_is_recover(cld)) {
1595                 rc = 0; /* this is not a fatal error for recover log */
1596                 if (rcl == 0)
1597                         rc = mgc_process_recover_log(mgc, cld);
1598         } else {
1599                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1600         }
1601
1602         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1603                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1604
1605         mutex_unlock(&cld->cld_lock);
1606
1607         /* Now drop the lock so MGS can revoke it */
1608         if (!rcl)
1609                 ldlm_lock_decref(&lockh, LCK_CR);
1610
1611         return rc;
1612 }
1613
1614
1615 /** Called from lustre_process_log.
1616  * LCFG_LOG_START gets the config log from the MGS, processes it to start
1617  * any services, and adds it to the list logs to watch (follow).
1618  */
1619 static int mgc_process_config(struct obd_device *obd, u32 len, void *buf)
1620 {
1621         struct lustre_cfg *lcfg = buf;
1622         struct config_llog_instance *cfg = NULL;
1623         char *logname;
1624         int rc = 0;
1625
1626         switch (lcfg->lcfg_command) {
1627         case LCFG_LOV_ADD_OBD: {
1628                 /* Overloading this cfg command: register a new target */
1629                 struct mgs_target_info *mti;
1630
1631                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1632                     sizeof(struct mgs_target_info)) {
1633                         rc = -EINVAL;
1634                         goto out;
1635                 }
1636
1637                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1638                 CDEBUG(D_MGC, "add_target %s %#x\n",
1639                        mti->mti_svname, mti->mti_flags);
1640                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1641                 break;
1642         }
1643         case LCFG_LOV_DEL_OBD:
1644                 /* Unregister has no meaning at the moment. */
1645                 CERROR("lov_del_obd unimplemented\n");
1646                 rc = -ENOSYS;
1647                 break;
1648         case LCFG_SPTLRPC_CONF: {
1649                 rc = sptlrpc_process_config(lcfg);
1650                 break;
1651         }
1652         case LCFG_LOG_START: {
1653                 struct config_llog_data *cld;
1654                 struct super_block *sb;
1655
1656                 logname = lustre_cfg_string(lcfg, 1);
1657                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1658                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1659
1660                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1661                        cfg->cfg_last_idx);
1662
1663                 /* We're only called through here on the initial mount */
1664                 rc = config_log_add(obd, logname, cfg, sb);
1665                 if (rc)
1666                         break;
1667                 cld = config_log_find(logname, cfg);
1668                 if (cld == NULL) {
1669                         rc = -ENOENT;
1670                         break;
1671                 }
1672
1673                 /* COMPAT_146 */
1674                 /* FIXME only set this for old logs!  Right now this forces
1675                    us to always skip the "inside markers" check */
1676                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1677
1678                 rc = mgc_process_log(obd, cld);
1679                 if (rc == 0 && cld->cld_recover != NULL) {
1680                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1681                                          imp_connect_data, IMP_RECOV)) {
1682                                 rc = mgc_process_log(obd, cld->cld_recover);
1683                         } else {
1684                                 struct config_llog_data *cir = cld->cld_recover;
1685                                 cld->cld_recover = NULL;
1686                                 config_log_put(cir);
1687                         }
1688                         if (rc)
1689                                 CERROR("Cannot process recover llog %d\n", rc);
1690                 }
1691
1692                 if (rc == 0 && cld->cld_params != NULL) {
1693                         rc = mgc_process_log(obd, cld->cld_params);
1694                         if (rc == -ENOENT) {
1695                                 CDEBUG(D_MGC,
1696                                        "There is no params config file yet\n");
1697                                 rc = 0;
1698                         }
1699                         /* params log is optional */
1700                         if (rc)
1701                                 CERROR(
1702                                        "%s: can't process params llog: rc = %d\n",
1703                                        obd->obd_name, rc);
1704                 }
1705                 config_log_put(cld);
1706
1707                 break;
1708         }
1709         case LCFG_LOG_END: {
1710                 logname = lustre_cfg_string(lcfg, 1);
1711
1712                 if (lcfg->lcfg_bufcount >= 2)
1713                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
1714                                 lcfg, 2);
1715                 rc = config_log_end(logname, cfg);
1716                 break;
1717         }
1718         default: {
1719                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1720                 rc = -EINVAL;
1721                 goto out;
1722
1723         }
1724         }
1725 out:
1726         return rc;
1727 }
1728
1729 struct obd_ops mgc_obd_ops = {
1730         .o_owner        = THIS_MODULE,
1731         .o_setup        = mgc_setup,
1732         .o_precleanup   = mgc_precleanup,
1733         .o_cleanup      = mgc_cleanup,
1734         .o_add_conn     = client_import_add_conn,
1735         .o_del_conn     = client_import_del_conn,
1736         .o_connect      = client_connect_import,
1737         .o_disconnect   = client_disconnect_export,
1738         /* .o_enqueue      = mgc_enqueue, */
1739         /* .o_iocontrol    = mgc_iocontrol, */
1740         .o_set_info_async = mgc_set_info_async,
1741         .o_get_info       = mgc_get_info,
1742         .o_import_event = mgc_import_event,
1743         .o_process_config = mgc_process_config,
1744 };
1745
1746 static int __init mgc_init(void)
1747 {
1748         return class_register_type(&mgc_obd_ops, NULL, NULL,
1749                                    LUSTRE_MGC_NAME, NULL);
1750 }
1751
1752 static void /*__exit*/ mgc_exit(void)
1753 {
1754         class_unregister_type(LUSTRE_MGC_NAME);
1755 }
1756
1757 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1758 MODULE_DESCRIPTION("Lustre Management Client");
1759 MODULE_LICENSE("GPL");
1760
1761 module_init(mgc_init);
1762 module_exit(mgc_exit);