lustre/mgc/mgc_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/mgc/mgc_request.c
  37  *
  38  * Author: Nathan Rutman <nathan@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_MGC
  42 #define D_MGC D_CONFIG /*|D_WARNING*/
  43
  44 #include <linux/module.h>
  45 #include <linux/kthread.h>
  46
  47 #include <dt_object.h>
  48 #include <lprocfs_status.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_disk.h>
  51 #include <lustre_log.h>
  52 #include <lustre_nodemap.h>
  53 #include <lustre_swab.h>
  54 #include <obd_class.h>
  55
  56 #include "mgc_internal.h"
  57
  58 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
  59                           int type)
  60 {
  61         __u64 resname = 0;
  62
  63         if (len > sizeof(resname)) {
  64                 CERROR("name too long: %s\n", name);
  65                 return -EINVAL;
  66         }
  67         if (len <= 0) {
  68                 CERROR("missing name: %s\n", name);
  69                 return -EINVAL;
  70         }
  71         memcpy(&resname, name, len);
  72
  73         /* Always use the same endianness for the resid */
  74         memset(res_id, 0, sizeof(*res_id));
  75         res_id->name[0] = cpu_to_le64(resname);
  76         /* XXX: unfortunately, sptlprc and config llog share one lock */
  77         switch(type) {
  78         case CONFIG_T_CONFIG:
  79         case CONFIG_T_SPTLRPC:
  80                 resname = 0;
  81                 break;
  82         case CONFIG_T_RECOVER:
  83         case CONFIG_T_PARAMS:
  84         case CONFIG_T_NODEMAP:
  85                 resname = type;
  86                 break;
  87         default:
  88                 LBUG();
  89         }
  90         res_id->name[1] = cpu_to_le64(resname);
  91         CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
  92                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  93         return 0;
  94 }
  95
  96 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
  97 {
  98         /* fsname is at most 8 chars long, maybe contain "-".
  99          * e.g. "lustre", "SUN-000" */
 100         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
 101 }
 102 EXPORT_SYMBOL(mgc_fsname2resid);
 103
 104 static int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id,
 105                              int type)
 106 {
 107         char *name_end;
 108         int len;
 109
 110         /* logname consists of "fsname-nodetype".
 111          * e.g. "lustre-MDT0001", "SUN-000-client"
 112          * there is an exception: llog "params" */
 113         name_end = strrchr(logname, '-');
 114         if (!name_end)
 115                 len = strlen(logname);
 116         else
 117                 len = name_end - logname;
 118         return mgc_name2resid(logname, len, res_id, type);
 119 }
 120
 121 /********************** config llog list **********************/
 122 static struct list_head config_llog_list = LIST_HEAD_INIT(config_llog_list);
 123 static DEFINE_SPINLOCK(config_list_lock);
 124
 125 /* Take a reference to a config log */
 126 static int config_log_get(struct config_llog_data *cld)
 127 {
 128         ENTRY;
 129         atomic_inc(&cld->cld_refcount);
 130         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 131                 atomic_read(&cld->cld_refcount));
 132         RETURN(0);
 133 }
 134
 135 /* Drop a reference to a config log.  When no longer referenced,
 136    we can free the config log data */
 137 static void config_log_put(struct config_llog_data *cld)
 138 {
 139         ENTRY;
 140
 141         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 142                 atomic_read(&cld->cld_refcount));
 143         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 144
 145         /* spinlock to make sure no item with 0 refcount in the list */
 146         if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 147                 list_del(&cld->cld_list_chain);
 148                 spin_unlock(&config_list_lock);
 149
 150                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 151
 152                 if (cld->cld_recover)
 153                         config_log_put(cld->cld_recover);
 154                 if (cld->cld_sptlrpc)
 155                         config_log_put(cld->cld_sptlrpc);
 156                 if (cld->cld_params)
 157                         config_log_put(cld->cld_params);
 158                 if (cld->cld_nodemap)
 159                         config_log_put(cld->cld_nodemap);
 160                 if (cld_is_sptlrpc(cld))
 161                         sptlrpc_conf_log_stop(cld->cld_logname);
 162
 163                 class_export_put(cld->cld_mgcexp);
 164                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 165         }
 166
 167         EXIT;
 168 }
 169
 170 /* Find a config log by name */
 171 static
 172 struct config_llog_data *config_log_find(char *logname,
 173                                          struct config_llog_instance *cfg)
 174 {
 175         struct config_llog_data *cld;
 176         struct config_llog_data *found = NULL;
 177         void *                   instance;
 178         ENTRY;
 179
 180         LASSERT(logname != NULL);
 181
 182         instance = cfg ? cfg->cfg_instance : NULL;
 183         spin_lock(&config_list_lock);
 184         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 185                 /* check if instance equals */
 186                 if (instance != cld->cld_cfg.cfg_instance)
 187                         continue;
 188
 189                 /* instance may be NULL, should check name */
 190                 if (strcmp(logname, cld->cld_logname) == 0) {
 191                         found = cld;
 192                         break;
 193                 }
 194         }
 195         if (found) {
 196                 atomic_inc(&found->cld_refcount);
 197                 LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0);
 198         }
 199         spin_unlock(&config_list_lock);
 200         RETURN(found);
 201 }
 202
 203 static
 204 struct config_llog_data *do_config_log_add(struct obd_device *obd,
 205                                            char *logname,
 206                                            int type,
 207                                            struct config_llog_instance *cfg,
 208                                            struct super_block *sb)
 209 {
 210         struct config_llog_data *cld;
 211         int rc;
 212
 213         ENTRY;
 214
 215         CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
 216                cfg ? cfg->cfg_instance : NULL);
 217
 218         OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
 219         if (!cld)
 220                 RETURN(ERR_PTR(-ENOMEM));
 221
 222         strcpy(cld->cld_logname, logname);
 223         if (cfg)
 224                 cld->cld_cfg = *cfg;
 225         else
 226                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 227         mutex_init(&cld->cld_lock);
 228         cld->cld_cfg.cfg_last_idx = 0;
 229         cld->cld_cfg.cfg_flags = 0;
 230         cld->cld_cfg.cfg_sb = sb;
 231         cld->cld_type = type;
 232         atomic_set(&cld->cld_refcount, 1);
 233
 234         /* Keep the mgc around until we are done */
 235         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 236
 237         if (cld_is_sptlrpc(cld)) {
 238                 sptlrpc_conf_log_start(logname);
 239                 cld->cld_cfg.cfg_obdname = obd->obd_name;
 240         }
 241
 242         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 243
 244         spin_lock(&config_list_lock);
 245         list_add(&cld->cld_list_chain, &config_llog_list);
 246         spin_unlock(&config_list_lock);
 247
 248         if (rc) {
 249                 config_log_put(cld);
 250                 RETURN(ERR_PTR(rc));
 251         }
 252
 253         if (cld_is_sptlrpc(cld) || cld_is_nodemap(cld)) {
 254                 rc = mgc_process_log(obd, cld);
 255                 if (rc && rc != -ENOENT)
 256                         CERROR("%s: failed processing log, type %d: rc = %d\n",
 257                                obd->obd_name, type, rc);
 258         }
 259
 260         RETURN(cld);
 261 }
 262
 263 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
 264         char *fsname,
 265         struct config_llog_instance *cfg,
 266         struct super_block *sb)
 267 {
 268         struct config_llog_instance lcfg = *cfg;
 269         struct lustre_sb_info *lsi = s2lsi(sb);
 270         struct config_llog_data *cld;
 271         char logname[32];
 272
 273         if (IS_OST(lsi))
 274                 return NULL;
 275
 276         /* for osp-on-ost, see lustre_start_osp() */
 277         if (IS_MDT(lsi) && lcfg.cfg_instance)
 278                 return NULL;
 279
 280         /* we have to use different llog for clients and mdts for cmd
 281          * where only clients are notified if one of cmd server restarts */
 282         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 283         strcpy(logname, fsname);
 284         if (IS_SERVER(lsi)) { /* mdt */
 285                 LASSERT(lcfg.cfg_instance == NULL);
 286                 lcfg.cfg_instance = sb;
 287                 strcat(logname, "-mdtir");
 288         } else {
 289                 LASSERT(lcfg.cfg_instance != NULL);
 290                 strcat(logname, "-cliir");
 291         }
 292
 293         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
 294         return cld;
 295 }
 296
 297 static struct config_llog_data *config_params_log_add(struct obd_device *obd,
 298         struct config_llog_instance *cfg, struct super_block *sb)
 299 {
 300         struct config_llog_instance     lcfg = *cfg;
 301         struct config_llog_data         *cld;
 302
 303         lcfg.cfg_instance = sb;
 304
 305         cld = do_config_log_add(obd, PARAMS_FILENAME, CONFIG_T_PARAMS,
 306                                 &lcfg, sb);
 307
 308         return cld;
 309 }
 310
 311 /** Add this log to the list of active logs watched by an MGC.
 312  * Active means we're watching for updates.
 313  * We have one active log per "mount" - client instance or servername.
 314  * Each instance may be at a different point in the log.
 315  */
 316 static int config_log_add(struct obd_device *obd, char *logname,
 317                           struct config_llog_instance *cfg,
 318                           struct super_block *sb)
 319 {
 320         struct lustre_sb_info   *lsi = s2lsi(sb);
 321         struct config_llog_data *cld;
 322         struct config_llog_data *sptlrpc_cld;
 323         struct config_llog_data *params_cld;
 324         struct config_llog_data *nodemap_cld;
 325         char                    seclogname[32];
 326         char                    *ptr;
 327         int                     rc;
 328         ENTRY;
 329
 330         CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
 331
 332         /*
 333          * for each regular log, the depended sptlrpc log name is
 334          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 335          */
 336         ptr = strrchr(logname, '-');
 337         if (ptr == NULL || ptr - logname > 8) {
 338                 CERROR("logname %s is too long\n", logname);
 339                 RETURN(-EINVAL);
 340         }
 341
 342         memcpy(seclogname, logname, ptr - logname);
 343         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 344
 345         sptlrpc_cld = config_log_find(seclogname, NULL);
 346         if (sptlrpc_cld == NULL) {
 347                 sptlrpc_cld = do_config_log_add(obd, seclogname,
 348                                                 CONFIG_T_SPTLRPC, NULL, NULL);
 349                 if (IS_ERR(sptlrpc_cld)) {
 350                         CERROR("can't create sptlrpc log: %s\n", seclogname);
 351                         GOTO(out, rc = PTR_ERR(sptlrpc_cld));
 352                 }
 353         }
 354
 355         nodemap_cld = config_log_find(LUSTRE_NODEMAP_NAME, NULL);
 356         if (!nodemap_cld && IS_SERVER(lsi) && !IS_MGS(lsi)) {
 357                 nodemap_cld = do_config_log_add(obd, LUSTRE_NODEMAP_NAME,
 358                                                 CONFIG_T_NODEMAP, NULL, NULL);
 359                 if (IS_ERR(nodemap_cld)) {
 360                         rc = PTR_ERR(nodemap_cld);
 361                         CERROR("%s: cannot create nodemap log: rc = %d\n",
 362                                obd->obd_name, rc);
 363                         GOTO(out_sptlrpc, rc);
 364                 }
 365         }
 366
 367         params_cld = config_params_log_add(obd, cfg, sb);
 368         if (IS_ERR(params_cld)) {
 369                 rc = PTR_ERR(params_cld);
 370                 CERROR("%s: can't create params log: rc = %d\n",
 371                        obd->obd_name, rc);
 372                 GOTO(out_nodemap, rc);
 373         }
 374
 375         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 376         if (IS_ERR(cld)) {
 377                 CERROR("can't create log: %s\n", logname);
 378                 GOTO(out_params, rc = PTR_ERR(cld));
 379         }
 380
 381         LASSERT(lsi->lsi_lmd);
 382         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
 383                 struct config_llog_data *recover_cld;
 384                 ptr = strrchr(seclogname, '-');
 385                 if (ptr != NULL) {
 386                         *ptr = 0;
 387                 }
 388                 else {
 389                         CERROR("%s: sptlrpc log name not correct, %s: "
 390                                "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
 391                         config_log_put(cld);
 392                         RETURN(-EINVAL);
 393                 }
 394                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 395                 if (IS_ERR(recover_cld))
 396                         GOTO(out_cld, rc = PTR_ERR(recover_cld));
 397                 cld->cld_recover = recover_cld;
 398         }
 399
 400         cld->cld_sptlrpc = sptlrpc_cld;
 401         cld->cld_params = params_cld;
 402         cld->cld_nodemap = nodemap_cld;
 403
 404         RETURN(0);
 405
 406 out_cld:
 407         config_log_put(cld);
 408
 409 out_params:
 410         config_log_put(params_cld);
 411
 412 out_nodemap:
 413         config_log_put(nodemap_cld);
 414
 415 out_sptlrpc:
 416         config_log_put(sptlrpc_cld);
 417
 418 out:
 419         return rc;
 420 }
 421
 422 DEFINE_MUTEX(llog_process_lock);
 423
 424 /** Stop watching for updates on this log.
 425  */
 426 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 427 {
 428         struct config_llog_data *cld;
 429         struct config_llog_data *cld_sptlrpc = NULL;
 430         struct config_llog_data *cld_params = NULL;
 431         struct config_llog_data *cld_recover = NULL;
 432         struct config_llog_data *cld_nodemap = NULL;
 433         int rc = 0;
 434
 435         ENTRY;
 436
 437         cld = config_log_find(logname, cfg);
 438         if (cld == NULL)
 439                 RETURN(-ENOENT);
 440
 441         mutex_lock(&cld->cld_lock);
 442         /*
 443          * if cld_stopping is set, it means we didn't start the log thus
 444          * not owning the start ref. this can happen after previous umount:
 445          * the cld still hanging there waiting for lock cancel, and we
 446          * remount again but failed in the middle and call log_end without
 447          * calling start_log.
 448          */
 449         if (unlikely(cld->cld_stopping)) {
 450                 mutex_unlock(&cld->cld_lock);
 451                 /* drop the ref from the find */
 452                 config_log_put(cld);
 453                 RETURN(rc);
 454         }
 455
 456         cld->cld_stopping = 1;
 457
 458         cld_recover = cld->cld_recover;
 459         cld->cld_recover = NULL;
 460         mutex_unlock(&cld->cld_lock);
 461
 462         if (cld_recover) {
 463                 mutex_lock(&cld_recover->cld_lock);
 464                 cld_recover->cld_stopping = 1;
 465                 mutex_unlock(&cld_recover->cld_lock);
 466                 config_log_put(cld_recover);
 467         }
 468
 469         spin_lock(&config_list_lock);
 470         cld_sptlrpc = cld->cld_sptlrpc;
 471         cld->cld_sptlrpc = NULL;
 472         cld_params = cld->cld_params;
 473         cld->cld_params = NULL;
 474         cld_nodemap = cld->cld_nodemap;
 475         cld->cld_nodemap = NULL;
 476         spin_unlock(&config_list_lock);
 477
 478         if (cld_sptlrpc)
 479                 config_log_put(cld_sptlrpc);
 480
 481         if (cld_params) {
 482                 mutex_lock(&cld_params->cld_lock);
 483                 cld_params->cld_stopping = 1;
 484                 mutex_unlock(&cld_params->cld_lock);
 485                 config_log_put(cld_params);
 486         }
 487
 488         if (cld_nodemap) {
 489                 mutex_lock(&cld_nodemap->cld_lock);
 490                 cld_nodemap->cld_stopping = 1;
 491                 mutex_unlock(&cld_nodemap->cld_lock);
 492                 config_log_put(cld_nodemap);
 493         }
 494
 495         /* drop the ref from the find */
 496         config_log_put(cld);
 497         /* drop the start ref */
 498         config_log_put(cld);
 499
 500         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 501                rc);
 502         RETURN(rc);
 503 }
 504
 505 #ifdef CONFIG_PROC_FS
 506 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 507 {
 508         struct obd_device       *obd = data;
 509         struct obd_import       *imp;
 510         struct obd_connect_data *ocd;
 511         struct config_llog_data *cld;
 512         ENTRY;
 513
 514         LASSERT(obd != NULL);
 515         LPROCFS_CLIMP_CHECK(obd);
 516         imp = obd->u.cli.cl_import;
 517         ocd = &imp->imp_connect_data;
 518
 519         seq_printf(m, "imperative_recovery: %s\n",
 520                    OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 521         seq_printf(m, "client_state:\n");
 522
 523         spin_lock(&config_list_lock);
 524         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 525                 if (cld->cld_recover == NULL)
 526                         continue;
 527                 seq_printf(m,  "    - { client: %s, nidtbl_version: %u }\n",
 528                            cld->cld_logname,
 529                            cld->cld_recover->cld_cfg.cfg_last_idx);
 530         }
 531         spin_unlock(&config_list_lock);
 532
 533         LPROCFS_CLIMP_EXIT(obd);
 534         RETURN(0);
 535 }
 536 #endif
 537
 538 /* reenqueue any lost locks */
 539 #define RQ_RUNNING      0x1
 540 #define RQ_NOW          0x2
 541 #define RQ_LATER        0x4
 542 #define RQ_STOP         0x8
 543 #define RQ_PRECLEANUP   0x10
 544 static int                    rq_state = 0;
 545 static wait_queue_head_t      rq_waitq;
 546 static DECLARE_COMPLETION(rq_exit);
 547 static DECLARE_COMPLETION(rq_start);
 548
 549 static void do_requeue(struct config_llog_data *cld)
 550 {
 551         int rc = 0;
 552         ENTRY;
 553
 554         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 555
 556         /*
 557          * Do not run mgc_process_log on a disconnected export or an
 558          * export which is being disconnected. Take the client
 559          * semaphore to make the check non-racy.
 560          */
 561         down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
 562                          OBD_CLI_SEM_MGC);
 563         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 564                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 565                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 566                 if (rc && rc != -ENOENT)
 567                         CERROR("failed processing log: %d\n", rc);
 568         } else {
 569                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 570                        cld->cld_logname);
 571         }
 572         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 573
 574         EXIT;
 575 }
 576
 577 /* this timeout represents how many seconds MGC should wait before
 578  * requeue config and recover lock to the MGS. We need to randomize this
 579  * in order to not flood the MGS.
 580  */
 581 #define MGC_TIMEOUT_MIN_SECONDS   5
 582 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
 583
 584 static int mgc_requeue_thread(void *data)
 585 {
 586         int rc = 0;
 587         bool first = true;
 588         ENTRY;
 589
 590         CDEBUG(D_MGC, "Starting requeue thread\n");
 591
 592         /* Keep trying failed locks periodically */
 593         spin_lock(&config_list_lock);
 594         rq_state |= RQ_RUNNING;
 595         while (1) {
 596                 struct l_wait_info lwi;
 597                 struct config_llog_data *cld, *cld_prev;
 598                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
 599                 int stopped = !!(rq_state & RQ_STOP);
 600                 int to;
 601
 602                 /* Any new or requeued lostlocks will change the state */
 603                 rq_state &= ~(RQ_NOW | RQ_LATER);
 604                 spin_unlock(&config_list_lock);
 605
 606                 if (first) {
 607                         first = false;
 608                         complete(&rq_start);
 609                 }
 610
 611                 /* Always wait a few seconds to allow the server who
 612                    caused the lock revocation to finish its setup, plus some
 613                    random so everyone doesn't try to reconnect at once. */
 614                 to = msecs_to_jiffies(MGC_TIMEOUT_MIN_SECONDS * MSEC_PER_SEC);
 615                 /* rand is centi-seconds */
 616                 to += msecs_to_jiffies(rand * MSEC_PER_SEC / 100);
 617                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 618                 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
 619                              &lwi);
 620
 621                 /*
 622                  * iterate & processing through the list. for each cld, process
 623                  * its depending sptlrpc cld firstly (if any) and then itself.
 624                  *
 625                  * it's guaranteed any item in the list must have
 626                  * reference > 0; and if cld_lostlock is set, at
 627                  * least one reference is taken by the previous enqueue.
 628                  */
 629                 cld_prev = NULL;
 630
 631                 spin_lock(&config_list_lock);
 632                 rq_state &= ~RQ_PRECLEANUP;
 633                 list_for_each_entry(cld, &config_llog_list,
 634                                         cld_list_chain) {
 635                         if (!cld->cld_lostlock)
 636                                 continue;
 637
 638                         spin_unlock(&config_list_lock);
 639
 640                         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 641
 642                         /* Whether we enqueued again or not in mgc_process_log,
 643                          * we're done with the ref from the old enqueue */
 644                         if (cld_prev)
 645                                 config_log_put(cld_prev);
 646                         cld_prev = cld;
 647
 648                         cld->cld_lostlock = 0;
 649                         if (likely(!stopped))
 650                                 do_requeue(cld);
 651
 652                         spin_lock(&config_list_lock);
 653                 }
 654                 spin_unlock(&config_list_lock);
 655                 if (cld_prev)
 656                         config_log_put(cld_prev);
 657
 658                 /* break after scanning the list so that we can drop
 659                  * refcount to losing lock clds */
 660                 if (unlikely(stopped)) {
 661                         spin_lock(&config_list_lock);
 662                         break;
 663                 }
 664
 665                 /* Wait a bit to see if anyone else needs a requeue */
 666                 lwi = (struct l_wait_info) { 0 };
 667                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 668                              &lwi);
 669                 spin_lock(&config_list_lock);
 670         }
 671         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 672         rq_state &= ~RQ_RUNNING;
 673         spin_unlock(&config_list_lock);
 674
 675         complete(&rq_exit);
 676
 677         CDEBUG(D_MGC, "Ending requeue thread\n");
 678         RETURN(rc);
 679 }
 680
 681 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 682    We are responsible for dropping the config log reference from here on out. */
 683 static void mgc_requeue_add(struct config_llog_data *cld)
 684 {
 685         ENTRY;
 686
 687         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 688                 cld->cld_logname, atomic_read(&cld->cld_refcount),
 689                 cld->cld_stopping, rq_state);
 690         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 691
 692         mutex_lock(&cld->cld_lock);
 693         if (cld->cld_stopping || cld->cld_lostlock) {
 694                 mutex_unlock(&cld->cld_lock);
 695                 RETURN_EXIT;
 696         }
 697         /* this refcount will be released in mgc_requeue_thread. */
 698         config_log_get(cld);
 699         cld->cld_lostlock = 1;
 700         mutex_unlock(&cld->cld_lock);
 701
 702         /* Hold lock for rq_state */
 703         spin_lock(&config_list_lock);
 704         if (rq_state & RQ_STOP) {
 705                 spin_unlock(&config_list_lock);
 706                 cld->cld_lostlock = 0;
 707                 config_log_put(cld);
 708         } else {
 709                 rq_state |= RQ_NOW;
 710                 spin_unlock(&config_list_lock);
 711                 wake_up(&rq_waitq);
 712         }
 713         EXIT;
 714 }
 715
 716 /********************** class fns **********************/
 717 static int mgc_local_llog_init(const struct lu_env *env,
 718                                struct obd_device *obd,
 719                                struct obd_device *disk)
 720 {
 721         struct llog_ctxt        *ctxt;
 722         int                      rc;
 723
 724         ENTRY;
 725
 726         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
 727                         &llog_osd_ops);
 728         if (rc)
 729                 RETURN(rc);
 730
 731         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 732         LASSERT(ctxt);
 733         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
 734         llog_ctxt_put(ctxt);
 735
 736         RETURN(0);
 737 }
 738
 739 static int mgc_local_llog_fini(const struct lu_env *env,
 740                                struct obd_device *obd)
 741 {
 742         struct llog_ctxt *ctxt;
 743
 744         ENTRY;
 745
 746         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 747         llog_cleanup(env, ctxt);
 748
 749         RETURN(0);
 750 }
 751
 752 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
 753                         struct super_block *sb)
 754 {
 755         struct lustre_sb_info   *lsi = s2lsi(sb);
 756         struct client_obd       *cli = &obd->u.cli;
 757         struct lu_fid            rfid, fid;
 758         struct dt_object        *root, *dto;
 759         int                      rc = 0;
 760
 761         ENTRY;
 762
 763         LASSERT(lsi);
 764         LASSERT(lsi->lsi_dt_dev);
 765
 766         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
 767         mutex_lock(&cli->cl_mgc_mutex);
 768
 769         /* Setup the configs dir */
 770         fid.f_seq = FID_SEQ_LOCAL_NAME;
 771         fid.f_oid = 1;
 772         fid.f_ver = 0;
 773         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
 774                                     &cli->cl_mgc_los);
 775         if (rc)
 776                 RETURN(rc);
 777
 778         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
 779         if (rc)
 780                 GOTO(out_los, rc);
 781
 782         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
 783                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
 784         if (unlikely(IS_ERR(root)))
 785                 GOTO(out_los, rc = PTR_ERR(root));
 786
 787         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
 788                                         MOUNT_CONFIGS_DIR,
 789                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
 790         lu_object_put_nocache(env, &root->do_lu);
 791         if (IS_ERR(dto))
 792                 GOTO(out_los, rc = PTR_ERR(dto));
 793
 794         cli->cl_mgc_configs_dir = dto;
 795
 796         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
 797         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
 798         if (rc)
 799                 GOTO(out_llog, rc);
 800
 801         /* We take an obd ref to insure that we can't get to mgc_cleanup
 802          * without calling mgc_fs_cleanup first. */
 803         class_incref(obd, "mgc_fs", obd);
 804
 805         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 806         EXIT;
 807 out_llog:
 808         if (rc) {
 809                 lu_object_put(env, &cli->cl_mgc_configs_dir->do_lu);
 810                 cli->cl_mgc_configs_dir = NULL;
 811         }
 812 out_los:
 813         if (rc < 0) {
 814                 local_oid_storage_fini(env, cli->cl_mgc_los);
 815                 cli->cl_mgc_los = NULL;
 816                 mutex_unlock(&cli->cl_mgc_mutex);
 817         }
 818         return rc;
 819 }
 820
 821 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
 822 {
 823         struct client_obd       *cli = &obd->u.cli;
 824         ENTRY;
 825
 826         LASSERT(cli->cl_mgc_los != NULL);
 827
 828         mgc_local_llog_fini(env, obd);
 829
 830         lu_object_put_nocache(env, &cli->cl_mgc_configs_dir->do_lu);
 831         cli->cl_mgc_configs_dir = NULL;
 832
 833         local_oid_storage_fini(env, cli->cl_mgc_los);
 834         cli->cl_mgc_los = NULL;
 835
 836         class_decref(obd, "mgc_fs", obd);
 837         mutex_unlock(&cli->cl_mgc_mutex);
 838
 839         RETURN(0);
 840 }
 841
 842 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 843 {
 844         struct llog_ctxt        *ctxt;
 845         int                      rc;
 846
 847         ENTRY;
 848
 849         /* setup only remote ctxt, the local disk context is switched per each
 850          * filesystem during mgc_fs_setup() */
 851         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 852                         &llog_client_ops);
 853         if (rc)
 854                 RETURN(rc);
 855
 856         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 857         LASSERT(ctxt);
 858
 859         llog_initiator_connect(ctxt);
 860         llog_ctxt_put(ctxt);
 861
 862         RETURN(0);
 863 }
 864
 865 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 866 {
 867         struct llog_ctxt *ctxt;
 868
 869         ENTRY;
 870
 871         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 872         if (ctxt)
 873                 llog_cleanup(env, ctxt);
 874
 875         RETURN(0);
 876 }
 877
 878
 879 static atomic_t mgc_count = ATOMIC_INIT(0);
 880 static int mgc_precleanup(struct obd_device *obd)
 881 {
 882         int     rc = 0;
 883         int     temp;
 884         ENTRY;
 885
 886         if (atomic_dec_and_test(&mgc_count)) {
 887                 LASSERT(rq_state & RQ_RUNNING);
 888                 /* stop requeue thread */
 889                 temp = RQ_STOP;
 890         } else {
 891                 /* wakeup requeue thread to clean our cld */
 892                 temp = RQ_NOW | RQ_PRECLEANUP;
 893         }
 894
 895         spin_lock(&config_list_lock);
 896         rq_state |= temp;
 897         spin_unlock(&config_list_lock);
 898         wake_up(&rq_waitq);
 899
 900         if (temp & RQ_STOP)
 901                 wait_for_completion(&rq_exit);
 902         obd_cleanup_client_import(obd);
 903
 904         rc = mgc_llog_fini(NULL, obd);
 905         if (rc != 0)
 906                 CERROR("failed to cleanup llogging subsystems\n");
 907
 908         RETURN(rc);
 909 }
 910
 911 static int mgc_cleanup(struct obd_device *obd)
 912 {
 913         int rc;
 914         ENTRY;
 915
 916         /* COMPAT_146 - old config logs may have added profiles we don't
 917            know about */
 918         if (obd->obd_type->typ_refcnt <= 1)
 919                 /* Only for the last mgc */
 920                 class_del_profiles();
 921
 922         lprocfs_obd_cleanup(obd);
 923         ptlrpcd_decref();
 924
 925         rc = client_obd_cleanup(obd);
 926         RETURN(rc);
 927 }
 928
 929 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 930 {
 931         struct task_struct      *task;
 932         int                      rc;
 933         ENTRY;
 934
 935         rc = ptlrpcd_addref();
 936         if (rc < 0)
 937                 RETURN(rc);
 938
 939         rc = client_obd_setup(obd, lcfg);
 940         if (rc)
 941                 GOTO(err_decref, rc);
 942
 943         rc = mgc_llog_init(NULL, obd);
 944         if (rc) {
 945                 CERROR("failed to setup llogging subsystems\n");
 946                 GOTO(err_cleanup, rc);
 947         }
 948
 949 #ifdef CONFIG_PROC_FS
 950         obd->obd_vars = lprocfs_mgc_obd_vars;
 951         lprocfs_obd_setup(obd);
 952 #endif
 953         sptlrpc_lprocfs_cliobd_attach(obd);
 954
 955         if (atomic_inc_return(&mgc_count) == 1) {
 956                 rq_state = 0;
 957                 init_waitqueue_head(&rq_waitq);
 958
 959                 /* start requeue thread */
 960                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 961                 if (IS_ERR(task)) {
 962                         rc = PTR_ERR(task);
 963                         CERROR("%s: cannot start requeue thread: rc = %d; "
 964                                "no more log updates\n",
 965                                obd->obd_name, rc);
 966                         GOTO(err_cleanup, rc);
 967                 }
 968                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 969                 rc = 0;
 970                 wait_for_completion(&rq_start);
 971         }
 972
 973         RETURN(rc);
 974
 975 err_cleanup:
 976         client_obd_cleanup(obd);
 977 err_decref:
 978         ptlrpcd_decref();
 979         RETURN(rc);
 980 }
 981
 982 /* based on ll_mdc_blocking_ast */
 983 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 984                             void *data, int flag)
 985 {
 986         struct lustre_handle lockh;
 987         struct config_llog_data *cld = (struct config_llog_data *)data;
 988         int rc = 0;
 989         ENTRY;
 990
 991         switch (flag) {
 992         case LDLM_CB_BLOCKING:
 993                 /* mgs wants the lock, give it up... */
 994                 LDLM_DEBUG(lock, "MGC blocking CB");
 995                 ldlm_lock2handle(lock, &lockh);
 996                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 997                 break;
 998         case LDLM_CB_CANCELING:
 999                 /* We've given up the lock, prepare ourselves to update. */
1000                 LDLM_DEBUG(lock, "MGC cancel CB");
1001
1002                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
1003                        PLDLMRES(lock->l_resource),
1004                        (char *)&lock->l_resource->lr_name.name[0]);
1005
1006                 if (!cld) {
1007                         CDEBUG(D_INFO, "missing data, won't requeue\n");
1008                         break;
1009                 }
1010
1011                 /* held at mgc_process_log(). */
1012                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1013                 /* Are we done with this log? */
1014                 if (cld->cld_stopping) {
1015                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1016                                 cld->cld_logname);
1017                         config_log_put(cld);
1018                         break;
1019                 }
1020                 /* Make sure not to re-enqueue when the mgc is stopping
1021                    (we get called from client_disconnect_export) */
1022                 if (lock->l_conn_export == NULL ||
1023                     lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1024                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1025                                 cld->cld_logname);
1026                         config_log_put(cld);
1027                         break;
1028                 }
1029
1030                 /* Re-enqueue now */
1031                 mgc_requeue_add(cld);
1032                 config_log_put(cld);
1033                 break;
1034         default:
1035                 LBUG();
1036         }
1037
1038         RETURN(rc);
1039 }
1040
1041 /* Not sure where this should go... */
1042 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1043  * that we can have a chance to try the secondary MGS if any. */
1044 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1045                                 + PING_INTERVAL)
1046 #define  MGC_TARGET_REG_LIMIT 10
1047 #define  MGC_SEND_PARAM_LIMIT 10
1048
1049 /* Send parameter to MGS*/
1050 static int mgc_set_mgs_param(struct obd_export *exp,
1051                              struct mgs_send_param *msp)
1052 {
1053         struct ptlrpc_request *req;
1054         struct mgs_send_param *req_msp, *rep_msp;
1055         int rc;
1056         ENTRY;
1057
1058         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1059                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1060                                         MGS_SET_INFO);
1061         if (!req)
1062                 RETURN(-ENOMEM);
1063
1064         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1065         if (!req_msp) {
1066                 ptlrpc_req_finished(req);
1067                 RETURN(-ENOMEM);
1068         }
1069
1070         memcpy(req_msp, msp, sizeof(*req_msp));
1071         ptlrpc_request_set_replen(req);
1072
1073         /* Limit how long we will wait for the enqueue to complete */
1074         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1075         rc = ptlrpc_queue_wait(req);
1076         if (!rc) {
1077                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1078                 memcpy(msp, rep_msp, sizeof(*rep_msp));
1079         }
1080
1081         ptlrpc_req_finished(req);
1082
1083         RETURN(rc);
1084 }
1085
1086 /* Take a config lock so we can get cancel notifications */
1087 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1088                        union ldlm_policy_data *policy, enum ldlm_mode mode,
1089                        __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
1090                        void *data, __u32 lvb_len, void *lvb_swabber,
1091                        struct lustre_handle *lockh)
1092 {
1093         struct config_llog_data *cld = (struct config_llog_data *)data;
1094         struct ldlm_enqueue_info einfo = {
1095                 .ei_type        = type,
1096                 .ei_mode        = mode,
1097                 .ei_cb_bl       = mgc_blocking_ast,
1098                 .ei_cb_cp       = ldlm_completion_ast,
1099         };
1100         struct ptlrpc_request *req;
1101         int short_limit = cld_is_sptlrpc(cld);
1102         int rc;
1103         ENTRY;
1104
1105         CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
1106                cld->cld_resid.name[0]);
1107
1108         /* We need a callback for every lockholder, so don't try to
1109            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1110         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1111                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1112                                         LDLM_ENQUEUE);
1113         if (req == NULL)
1114                 RETURN(-ENOMEM);
1115
1116         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1117         ptlrpc_request_set_replen(req);
1118
1119         /* check if this is server or client */
1120         if (cld->cld_cfg.cfg_sb) {
1121                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1122                 if (lsi && IS_SERVER(lsi))
1123                         short_limit = 1;
1124         }
1125         /* Limit how long we will wait for the enqueue to complete */
1126         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1127         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1128                               NULL, 0, LVB_T_NONE, lockh, 0);
1129         /* A failed enqueue should still call the mgc_blocking_ast,
1130            where it will be requeued if needed ("grant failed"). */
1131         ptlrpc_req_finished(req);
1132         RETURN(rc);
1133 }
1134
1135 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1136                       struct lustre_handle *lockh)
1137 {
1138         ENTRY;
1139
1140         ldlm_lock_decref(lockh, mode);
1141
1142         RETURN(0);
1143 }
1144
1145 static void mgc_notify_active(struct obd_device *unused)
1146 {
1147         /* wakeup mgc_requeue_thread to requeue mgc lock */
1148         spin_lock(&config_list_lock);
1149         rq_state |= RQ_NOW;
1150         spin_unlock(&config_list_lock);
1151         wake_up(&rq_waitq);
1152
1153         /* TODO: Help the MGS rebuild nidtbl. -jay */
1154 }
1155
1156 /* Send target_reg message to MGS */
1157 static int mgc_target_register(struct obd_export *exp,
1158                                struct mgs_target_info *mti)
1159 {
1160         struct ptlrpc_request  *req;
1161         struct mgs_target_info *req_mti, *rep_mti;
1162         int                     rc;
1163         ENTRY;
1164
1165         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1166                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1167                                         MGS_TARGET_REG);
1168         if (req == NULL)
1169                 RETURN(-ENOMEM);
1170
1171         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1172         if (!req_mti) {
1173                 ptlrpc_req_finished(req);
1174                 RETURN(-ENOMEM);
1175         }
1176
1177         memcpy(req_mti, mti, sizeof(*req_mti));
1178         ptlrpc_request_set_replen(req);
1179         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1180         /* Limit how long we will wait for the enqueue to complete */
1181         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1182
1183         rc = ptlrpc_queue_wait(req);
1184         if (!rc) {
1185                 rep_mti = req_capsule_server_get(&req->rq_pill,
1186                                                  &RMF_MGS_TARGET_INFO);
1187                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1188                 CDEBUG(D_MGC, "register %s got index = %d\n",
1189                        mti->mti_svname, mti->mti_stripe_index);
1190         }
1191         ptlrpc_req_finished(req);
1192
1193         RETURN(rc);
1194 }
1195
1196 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1197                               u32 keylen, void *key,
1198                               u32 vallen, void *val,
1199                               struct ptlrpc_request_set *set)
1200 {
1201         int rc = -EINVAL;
1202         ENTRY;
1203
1204         /* Turn off initial_recov after we try all backup servers once */
1205         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1206                 struct obd_import *imp = class_exp2cliimp(exp);
1207                 int value;
1208                 if (vallen != sizeof(int))
1209                         RETURN(-EINVAL);
1210                 value = *(int *)val;
1211                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1212                        imp->imp_obd->obd_name, value,
1213                        imp->imp_deactive, imp->imp_invalid,
1214                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1215                        ptlrpc_import_state_name(imp->imp_state));
1216                 /* Resurrect if we previously died */
1217                 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1218                      imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1219                         ptlrpc_reconnect_import(imp);
1220                 RETURN(0);
1221         }
1222         /* FIXME move this to mgc_process_config */
1223         if (KEY_IS(KEY_REGISTER_TARGET)) {
1224                 struct mgs_target_info *mti;
1225                 if (vallen != sizeof(struct mgs_target_info))
1226                         RETURN(-EINVAL);
1227                 mti = (struct mgs_target_info *)val;
1228                 CDEBUG(D_MGC, "register_target %s %#x\n",
1229                        mti->mti_svname, mti->mti_flags);
1230                 rc =  mgc_target_register(exp, mti);
1231                 RETURN(rc);
1232         }
1233         if (KEY_IS(KEY_SET_FS)) {
1234                 struct super_block *sb = (struct super_block *)val;
1235
1236                 if (vallen != sizeof(struct super_block))
1237                         RETURN(-EINVAL);
1238
1239                 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1240                 RETURN(rc);
1241         }
1242         if (KEY_IS(KEY_CLEAR_FS)) {
1243                 if (vallen != 0)
1244                         RETURN(-EINVAL);
1245                 rc = mgc_fs_cleanup(env, exp->exp_obd);
1246                 RETURN(rc);
1247         }
1248         if (KEY_IS(KEY_SET_INFO)) {
1249                 struct mgs_send_param *msp;
1250
1251                 msp = (struct mgs_send_param *)val;
1252                 rc =  mgc_set_mgs_param(exp, msp);
1253                 RETURN(rc);
1254         }
1255         if (KEY_IS(KEY_MGSSEC)) {
1256                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1257                 struct sptlrpc_flavor  flvr;
1258
1259                 /*
1260                  * empty string means using current flavor, if which haven't
1261                  * been set yet, set it as null.
1262                  *
1263                  * if flavor has been set previously, check the asking flavor
1264                  * must match the existing one.
1265                  */
1266                 if (vallen == 0) {
1267                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1268                                 RETURN(0);
1269                         val = "null";
1270                         vallen = 4;
1271                 }
1272
1273                 rc = sptlrpc_parse_flavor(val, &flvr);
1274                 if (rc) {
1275                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1276                                (char *) val);
1277                         RETURN(rc);
1278                 }
1279
1280                 /*
1281                  * caller already hold a mutex
1282                  */
1283                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1284                         cli->cl_flvr_mgc = flvr;
1285                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1286                                   sizeof(flvr)) != 0) {
1287                         char    str[20];
1288
1289                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1290                                             str, sizeof(str));
1291                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1292                                        "currently %s is in use\n",
1293                                        (char *) val, str);
1294                         rc = -EPERM;
1295                 }
1296                 RETURN(rc);
1297         }
1298
1299         RETURN(rc);
1300 }
1301
1302 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1303                         __u32 keylen, void *key, __u32 *vallen, void *val)
1304 {
1305         int rc = -EINVAL;
1306
1307         if (KEY_IS(KEY_CONN_DATA)) {
1308                 struct obd_import *imp = class_exp2cliimp(exp);
1309                 struct obd_connect_data *data = val;
1310
1311                 if (*vallen == sizeof(*data)) {
1312                         *data = imp->imp_connect_data;
1313                         rc = 0;
1314                 }
1315         }
1316
1317         return rc;
1318 }
1319
1320 static int mgc_import_event(struct obd_device *obd,
1321                             struct obd_import *imp,
1322                             enum obd_import_event event)
1323 {
1324         int rc = 0;
1325
1326         LASSERT(imp->imp_obd == obd);
1327         CDEBUG(D_MGC, "import event %#x\n", event);
1328
1329         switch (event) {
1330         case IMP_EVENT_DISCON:
1331                 /* MGC imports should not wait for recovery */
1332                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1333                         ptlrpc_pinger_ir_down();
1334                 break;
1335         case IMP_EVENT_INACTIVE:
1336                 break;
1337         case IMP_EVENT_INVALIDATE: {
1338                 struct ldlm_namespace *ns = obd->obd_namespace;
1339                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1340                 break;
1341         }
1342         case IMP_EVENT_ACTIVE:
1343                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1344                 /* Clearing obd_no_recov allows us to continue pinging */
1345                 obd->obd_no_recov = 0;
1346                 mgc_notify_active(obd);
1347                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1348                         ptlrpc_pinger_ir_up();
1349                 break;
1350         case IMP_EVENT_OCD:
1351                 break;
1352         case IMP_EVENT_DEACTIVATE:
1353         case IMP_EVENT_ACTIVATE:
1354                 break;
1355         default:
1356                 CERROR("Unknown import event %#x\n", event);
1357                 LBUG();
1358         }
1359         RETURN(rc);
1360 }
1361
1362 enum {
1363         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_CACHE_SHIFT),
1364         CONFIG_READ_NRPAGES      = 4
1365 };
1366
1367 static int mgc_apply_recover_logs(struct obd_device *mgc,
1368                                   struct config_llog_data *cld,
1369                                   __u64 max_version,
1370                                   void *data, int datalen, bool mne_swab)
1371 {
1372         struct config_llog_instance *cfg = &cld->cld_cfg;
1373         struct lustre_sb_info       *lsi = s2lsi(cfg->cfg_sb);
1374         struct mgs_nidtbl_entry *entry;
1375         struct lustre_cfg       *lcfg;
1376         struct lustre_cfg_bufs   bufs;
1377         u64   prev_version = 0;
1378         char *inst;
1379         char *buf;
1380         int   bufsz;
1381         int   pos;
1382         int   rc  = 0;
1383         int   off = 0;
1384         ENTRY;
1385
1386         LASSERT(cfg->cfg_instance != NULL);
1387         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1388
1389         OBD_ALLOC(inst, PAGE_CACHE_SIZE);
1390         if (inst == NULL)
1391                 RETURN(-ENOMEM);
1392
1393         if (!IS_SERVER(lsi)) {
1394                 pos = snprintf(inst, PAGE_CACHE_SIZE, "%p", cfg->cfg_instance);
1395                 if (pos >= PAGE_CACHE_SIZE) {
1396                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1397                         return -E2BIG;
1398                 }
1399         } else {
1400                 LASSERT(IS_MDT(lsi));
1401                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1402                                         PAGE_CACHE_SIZE);
1403                 if (rc) {
1404                         OBD_FREE(inst, PAGE_CACHE_SIZE);
1405                         RETURN(-EINVAL);
1406                 }
1407                 pos = strlen(inst);
1408         }
1409
1410         ++pos;
1411         buf   = inst + pos;
1412         bufsz = PAGE_CACHE_SIZE - pos;
1413
1414         while (datalen > 0) {
1415                 int   entry_len = sizeof(*entry);
1416                 int   is_ost, i;
1417                 struct obd_device *obd;
1418                 char *obdname;
1419                 char *cname;
1420                 char *params;
1421                 char *uuid;
1422
1423                 rc = -EINVAL;
1424                 if (datalen < sizeof(*entry))
1425                         break;
1426
1427                 entry = (typeof(entry))(data + off);
1428
1429                 /* sanity check */
1430                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1431                         break;
1432                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1433                         break;
1434                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1435                         break;
1436
1437                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1438                 if (datalen < entry_len) /* must have entry_len at least */
1439                         break;
1440
1441                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1442                 if (mne_swab)
1443                         lustre_swab_mgs_nidtbl_entry(entry);
1444                 if (entry->mne_length > PAGE_CACHE_SIZE) {
1445                         CERROR("MNE too large (%u)\n", entry->mne_length);
1446                         break;
1447                 }
1448
1449                 if (entry->mne_length < entry_len)
1450                         break;
1451
1452                 off     += entry->mne_length;
1453                 datalen -= entry->mne_length;
1454                 if (datalen < 0)
1455                         break;
1456
1457                 if (entry->mne_version > max_version) {
1458                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1459                                entry->mne_version, max_version);
1460                         break;
1461                 }
1462
1463                 if (prev_version >= entry->mne_version) {
1464                         CERROR("index unsorted, prev %lld, now %lld\n",
1465                                prev_version, entry->mne_version);
1466                         break;
1467                 }
1468                 prev_version = entry->mne_version;
1469
1470                 /*
1471                  * Write a string with format "nid::instance" to
1472                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1473                  */
1474
1475                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1476                 memset(buf, 0, bufsz);
1477                 obdname = buf;
1478                 pos = 0;
1479
1480                 /* lustre-OST0001-osc-<instance #> */
1481                 strcpy(obdname, cld->cld_logname);
1482                 cname = strrchr(obdname, '-');
1483                 if (cname == NULL) {
1484                         CERROR("mgc %s: invalid logname %s\n",
1485                                mgc->obd_name, obdname);
1486                         break;
1487                 }
1488
1489                 pos = cname - obdname;
1490                 obdname[pos] = 0;
1491                 pos += sprintf(obdname + pos, "-%s%04x",
1492                                   is_ost ? "OST" : "MDT", entry->mne_index);
1493
1494                 cname = is_ost ? "osc" : "mdc",
1495                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1496                 lustre_cfg_bufs_reset(&bufs, obdname);
1497
1498                 /* find the obd by obdname */
1499                 obd = class_name2obd(obdname);
1500                 if (obd == NULL) {
1501                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1502                                mgc->obd_name, obdname);
1503                         rc = 0;
1504                         /* this is a safe race, when the ost is starting up...*/
1505                         continue;
1506                 }
1507
1508                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1509                 ++pos;
1510                 params = buf + pos;
1511                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1512                 uuid = buf + pos;
1513
1514                 down_read(&obd->u.cli.cl_sem);
1515                 if (obd->u.cli.cl_import == NULL) {
1516                         /* client does not connect to the OST yet */
1517                         up_read(&obd->u.cli.cl_sem);
1518                         rc = 0;
1519                         continue;
1520                 }
1521
1522                 /* iterate all nids to find one */
1523                 /* find uuid by nid */
1524                 rc = -ENOENT;
1525                 for (i = 0; i < entry->mne_nid_count; i++) {
1526                         rc = client_import_find_conn(obd->u.cli.cl_import,
1527                                                      entry->u.nids[i],
1528                                                      (struct obd_uuid *)uuid);
1529                         if (rc == 0)
1530                                 break;
1531                 }
1532
1533                 up_read(&obd->u.cli.cl_sem);
1534                 if (rc < 0) {
1535                         CERROR("mgc: cannot find uuid by nid %s\n",
1536                                libcfs_nid2str(entry->u.nids[0]));
1537                         break;
1538                 }
1539
1540                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1541                        uuid, libcfs_nid2str(entry->u.nids[0]));
1542
1543                 pos += strlen(uuid);
1544                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1545                 LASSERT(pos < bufsz);
1546
1547                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1548
1549                 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1550                 if (lcfg == NULL) {
1551                         rc = -ENOMEM;
1552                         break;
1553                 }
1554
1555                 CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n",
1556                        prev_version, max_version, obdname, params);
1557
1558                 rc = class_process_config(lcfg);
1559                 lustre_cfg_free(lcfg);
1560                 if (rc)
1561                         CDEBUG(D_INFO, "process config for %s error %d\n",
1562                                obdname, rc);
1563
1564                 /* continue, even one with error */
1565         }
1566
1567         OBD_FREE(inst, PAGE_CACHE_SIZE);
1568         RETURN(rc);
1569 }
1570
1571 /**
1572  * This function is called if this client was notified for target restarting
1573  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1574  * nodemap logs.
1575  */
1576 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1577                                            struct config_llog_data *cld)
1578 {
1579         struct ptlrpc_request *req = NULL;
1580         struct config_llog_instance *cfg = &cld->cld_cfg;
1581         struct mgs_config_body *body;
1582         struct mgs_config_res *res;
1583         struct nodemap_config *new_config = NULL;
1584         struct lu_nodemap *recent_nodemap = NULL;
1585         struct ptlrpc_bulk_desc *desc;
1586         struct page **pages;
1587         __u64 config_read_offset = 0;
1588         int nrpages;
1589         bool eof = true;
1590         bool mne_swab = false;
1591         int i;
1592         int ealen;
1593         int rc;
1594
1595         ENTRY;
1596
1597         /* allocate buffer for bulk transfer.
1598          * if this is the first time for this mgs to read logs,
1599          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1600          * once; otherwise, it only reads increment of logs, this should be
1601          * small and CONFIG_READ_NRPAGES will be used.
1602          */
1603         nrpages = CONFIG_READ_NRPAGES;
1604         if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1605                 nrpages = CONFIG_READ_NRPAGES_INIT;
1606
1607         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1608         if (pages == NULL)
1609                 GOTO(out, rc = -ENOMEM);
1610
1611         for (i = 0; i < nrpages; i++) {
1612                 pages[i] = alloc_page(GFP_KERNEL);
1613                 if (pages[i] == NULL)
1614                         GOTO(out, rc = -ENOMEM);
1615         }
1616
1617 again:
1618 #ifdef HAVE_SERVER_SUPPORT
1619         if (cld_is_nodemap(cld) && config_read_offset == 0) {
1620                 new_config = nodemap_config_alloc();
1621                 if (IS_ERR(new_config)) {
1622                         rc = PTR_ERR(new_config);
1623                         new_config = NULL;
1624                         GOTO(out, rc);
1625                 }
1626         }
1627 #endif
1628         LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1629         LASSERT(mutex_is_locked(&cld->cld_lock));
1630         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1631                                    &RQF_MGS_CONFIG_READ);
1632         if (req == NULL)
1633                 GOTO(out, rc = -ENOMEM);
1634
1635         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1636         if (rc)
1637                 GOTO(out, rc);
1638
1639         /* pack request */
1640         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1641         LASSERT(body != NULL);
1642         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1643         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1644             >= sizeof(body->mcb_name))
1645                 GOTO(out, rc = -E2BIG);
1646         if (cld_is_nodemap(cld))
1647                 body->mcb_offset = config_read_offset;
1648         else
1649                 body->mcb_offset = cfg->cfg_last_idx + 1;
1650         body->mcb_type   = cld->cld_type;
1651         body->mcb_bits   = PAGE_CACHE_SHIFT;
1652         body->mcb_units  = nrpages;
1653
1654         /* allocate bulk transfer descriptor */
1655         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1656                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1657                                     MGS_BULK_PORTAL,
1658                                     &ptlrpc_bulk_kiov_pin_ops);
1659         if (desc == NULL)
1660                 GOTO(out, rc = -ENOMEM);
1661
1662         for (i = 0; i < nrpages; i++)
1663                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1664                                                  PAGE_CACHE_SIZE);
1665
1666         ptlrpc_request_set_replen(req);
1667         rc = ptlrpc_queue_wait(req);
1668         if (rc)
1669                 GOTO(out, rc);
1670
1671         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1672         if (!res)
1673                 GOTO(out, rc = -EPROTO);
1674
1675         if (cld_is_nodemap(cld)) {
1676                 config_read_offset = res->mcr_offset;
1677                 eof = config_read_offset == II_END_OFF;
1678         } else {
1679                 if (res->mcr_size < res->mcr_offset)
1680                         GOTO(out, rc = -EINVAL);
1681
1682                 /* always update the index even though it might have errors with
1683                  * handling the recover logs
1684                  */
1685                 cfg->cfg_last_idx = res->mcr_offset;
1686                 eof = res->mcr_offset == res->mcr_size;
1687
1688                 CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n",
1689                        res->mcr_offset, eof == false);
1690         }
1691
1692         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1693         if (ealen < 0)
1694                 GOTO(out, rc = ealen);
1695
1696         if (ealen > nrpages << PAGE_CACHE_SHIFT)
1697                 GOTO(out, rc = -EINVAL);
1698
1699         if (ealen == 0) { /* no logs transferred */
1700 #ifdef HAVE_SERVER_SUPPORT
1701                 /* config changed since first read RPC */
1702                 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1703                         recent_nodemap = NULL;
1704                         nodemap_config_dealloc(new_config);
1705                         new_config = NULL;
1706
1707                         CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1708
1709                         /* setting eof to false, we request config again */
1710                         eof = false;
1711                         GOTO(out, rc = 0);
1712                 }
1713 #endif
1714                 if (!eof)
1715                         rc = -EINVAL;
1716                 GOTO(out, rc);
1717         }
1718
1719         mne_swab = !!ptlrpc_rep_need_swab(req);
1720 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1721         /* This import flag means the server did an extra swab of IR MNE
1722          * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1723         if (unlikely(req->rq_import->imp_need_mne_swab))
1724                 mne_swab = !mne_swab;
1725 #endif
1726
1727         /* When a nodemap config is received, we build a new nodemap config,
1728          * with new nodemap structs. We keep track of the most recently added
1729          * nodemap since the config is read ordered by nodemap_id, and so it
1730          * is likely that the next record will be related. Because access to
1731          * the nodemaps is single threaded until the nodemap_config is active,
1732          * we don't need to reference count with recent_nodemap, though
1733          * recent_nodemap should be set to NULL when the nodemap_config
1734          * is either destroyed or set active.
1735          */
1736         for (i = 0; i < nrpages && ealen > 0; i++) {
1737                 int rc2;
1738                 union lu_page   *ptr;
1739
1740                 ptr = kmap(pages[i]);
1741                 if (cld_is_nodemap(cld))
1742                         rc2 = nodemap_process_idx_pages(new_config, ptr,
1743                                                        &recent_nodemap);
1744                 else
1745                         rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1746                                                      ptr,
1747                                                      min_t(int, ealen,
1748                                                            PAGE_CACHE_SIZE),
1749                                                      mne_swab);
1750                 kunmap(pages[i]);
1751                 if (rc2 < 0) {
1752                         CWARN("%s: error processing %s log %s: rc = %d\n",
1753                               obd->obd_name,
1754                               cld_is_nodemap(cld) ? "nodemap" : "recovery",
1755                               cld->cld_logname,
1756                               rc2);
1757                         break;
1758                 }
1759
1760                 ealen -= PAGE_CACHE_SIZE;
1761         }
1762
1763 out:
1764         if (req) {
1765                 ptlrpc_req_finished(req);
1766                 req = NULL;
1767         }
1768
1769         if (rc == 0 && !eof)
1770                 goto again;
1771
1772 #ifdef HAVE_SERVER_SUPPORT
1773         if (new_config != NULL) {
1774                 /* recent_nodemap cannot be used after set_active/dealloc */
1775                 if (rc == 0)
1776                         nodemap_config_set_active_mgc(new_config);
1777                 else
1778                         nodemap_config_dealloc(new_config);
1779         }
1780 #endif
1781
1782         if (pages) {
1783                 for (i = 0; i < nrpages; i++) {
1784                         if (pages[i] == NULL)
1785                                 break;
1786                         __free_page(pages[i]);
1787                 }
1788                 OBD_FREE(pages, sizeof(*pages) * nrpages);
1789         }
1790         return rc;
1791 }
1792
1793 /* Copy a remote log locally */
1794 static int mgc_llog_local_copy(const struct lu_env *env,
1795                                struct obd_device *obd,
1796                                struct llog_ctxt *rctxt,
1797                                struct llog_ctxt *lctxt, char *logname)
1798 {
1799         char    *temp_log;
1800         int      rc;
1801
1802         ENTRY;
1803
1804         /*
1805          * - copy it to backup using llog_backup()
1806          * - copy remote llog to logname using llog_backup()
1807          * - if failed then move bakup to logname again
1808          */
1809
1810         OBD_ALLOC(temp_log, strlen(logname) + 2);
1811         if (!temp_log)
1812                 RETURN(-ENOMEM);
1813         sprintf(temp_log, "%sT", logname);
1814
1815         /* make a copy of local llog at first */
1816         rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1817         if (rc < 0 && rc != -ENOENT)
1818                 GOTO(out, rc);
1819         /* copy remote llog to the local copy */
1820         rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1821         if (rc == -ENOENT) {
1822                 /* no remote llog, delete local one too */
1823                 llog_erase(env, lctxt, NULL, logname);
1824         } else if (rc < 0) {
1825                 /* error during backup, get local one back from the copy */
1826                 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1827 out:
1828                 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1829                        obd->obd_name, logname, rc);
1830         }
1831         llog_erase(env, lctxt, NULL, temp_log);
1832         OBD_FREE(temp_log, strlen(logname) + 2);
1833         return rc;
1834 }
1835
1836 /* local_only means it cannot get remote llogs */
1837 static int mgc_process_cfg_log(struct obd_device *mgc,
1838                                struct config_llog_data *cld, int local_only)
1839 {
1840         struct llog_ctxt        *ctxt, *lctxt = NULL;
1841         struct client_obd       *cli = &mgc->u.cli;
1842         struct lustre_sb_info   *lsi = NULL;
1843         int                      rc = 0;
1844         bool                     sptlrpc_started = false;
1845         struct lu_env           *env;
1846
1847         ENTRY;
1848
1849         LASSERT(cld);
1850         LASSERT(mutex_is_locked(&cld->cld_lock));
1851
1852         /*
1853          * local copy of sptlrpc log is controlled elsewhere, don't try to
1854          * read it up here.
1855          */
1856         if (cld_is_sptlrpc(cld) && local_only)
1857                 RETURN(0);
1858
1859         if (cld->cld_cfg.cfg_sb)
1860                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1861
1862         OBD_ALLOC_PTR(env);
1863         if (env == NULL)
1864                 RETURN(-ENOMEM);
1865
1866         rc = lu_env_init(env, LCT_MG_THREAD);
1867         if (rc)
1868                 GOTO(out_free, rc);
1869
1870         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1871         LASSERT(ctxt);
1872
1873         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1874
1875         /* Copy the setup log locally if we can. Don't mess around if we're
1876          * running an MGS though (logs are already local). */
1877         if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1878             cli->cl_mgc_configs_dir != NULL &&
1879             lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1880             lsi->lsi_dt_dev) {
1881                 if (!local_only)
1882                         /* Only try to copy log if we have the lock. */
1883                         rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1884                                                  cld->cld_logname);
1885                 if (local_only || rc) {
1886                         if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1887                             llog_is_empty(env, lctxt, cld->cld_logname)) {
1888                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1889                                                    "log %s and no local copy."
1890                                                    "\n", cld->cld_logname);
1891                                 GOTO(out_pop, rc = -ENOENT);
1892                         }
1893                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1894                                "copy for now, will try to update later.\n",
1895                                cld->cld_logname);
1896                         rc = 0;
1897                 }
1898                 /* Now, whether we copied or not, start using the local llog.
1899                  * If we failed to copy, we'll start using whatever the old
1900                  * log has. */
1901                 llog_ctxt_put(ctxt);
1902                 ctxt = lctxt;
1903                 lctxt = NULL;
1904         } else {
1905                 if (local_only) /* no local log at client side */
1906                         GOTO(out_pop, rc = -EIO);
1907         }
1908
1909         if (cld_is_sptlrpc(cld)) {
1910                 sptlrpc_conf_log_update_begin(cld->cld_logname);
1911                 sptlrpc_started = true;
1912         }
1913
1914         /* logname and instance info should be the same, so use our
1915          * copy of the instance for the update.  The cfg_last_idx will
1916          * be updated here. */
1917         rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1918                                      &cld->cld_cfg);
1919         EXIT;
1920
1921 out_pop:
1922         __llog_ctxt_put(env, ctxt);
1923         if (lctxt)
1924                 __llog_ctxt_put(env, lctxt);
1925
1926         /*
1927          * update settings on existing OBDs. doing it inside
1928          * of llog_process_lock so no device is attaching/detaching
1929          * in parallel.
1930          * the logname must be <fsname>-sptlrpc
1931          */
1932         if (sptlrpc_started) {
1933                 LASSERT(cld_is_sptlrpc(cld));
1934                 sptlrpc_conf_log_update_end(cld->cld_logname);
1935                 class_notify_sptlrpc_conf(cld->cld_logname,
1936                                           strlen(cld->cld_logname) -
1937                                           strlen("-sptlrpc"));
1938         }
1939
1940         lu_env_fini(env);
1941 out_free:
1942         OBD_FREE_PTR(env);
1943         return rc;
1944 }
1945
1946 static bool mgc_import_in_recovery(struct obd_import *imp)
1947 {
1948         bool in_recovery = true;
1949
1950         spin_lock(&imp->imp_lock);
1951         if (imp->imp_state == LUSTRE_IMP_FULL ||
1952             imp->imp_state == LUSTRE_IMP_CLOSED)
1953                 in_recovery = false;
1954         spin_unlock(&imp->imp_lock);
1955
1956         return in_recovery;
1957 }
1958
1959 /**
1960  * Get a configuration log from the MGS and process it.
1961  *
1962  * This function is called for both clients and servers to process the
1963  * configuration log from the MGS.  The MGC enqueues a DLM lock on the
1964  * log from the MGS, and if the lock gets revoked the MGC will be notified
1965  * by the lock cancellation callback that the config log has changed,
1966  * and will enqueue another MGS lock on it, and then continue processing
1967  * the new additions to the end of the log.
1968  *
1969  * Since the MGC import is not replayable, if the import is being evicted
1970  * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
1971  * the log until recovery is finished or the import is closed.
1972  *
1973  * Make a local copy of the log before parsing it if appropriate (non-MGS
1974  * server) so that the server can start even when the MGS is down.
1975  *
1976  * There shouldn't be multiple processes running process_log at once --
1977  * sounds like badness.  It actually might be fine, as long as they're not
1978  * trying to update from the same log simultaneously, in which case we
1979  * should use a per-log semaphore instead of cld_lock.
1980  *
1981  * \param[in] mgc       MGC device by which to fetch the configuration log
1982  * \param[in] cld       log processing state (stored in lock callback data)
1983  *
1984  * \retval              0 on success
1985  * \retval              negative errno on failure
1986  */
1987 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1988 {
1989         struct lustre_handle lockh = { 0 };
1990         __u64 flags = LDLM_FL_NO_LRU;
1991         int rc = 0, rcl;
1992         bool retry = false;
1993         ENTRY;
1994
1995         LASSERT(cld != NULL);
1996
1997         /* I don't want multiple processes running process_log at once --
1998            sounds like badness.  It actually might be fine, as long as
1999            we're not trying to update from the same log
2000            simultaneously (in which case we should use a per-log sem.) */
2001 restart:
2002         mutex_lock(&cld->cld_lock);
2003         if (cld->cld_stopping) {
2004                 mutex_unlock(&cld->cld_lock);
2005                 RETURN(0);
2006         }
2007
2008         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2009
2010         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
2011                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2012
2013         /* Get the cfg lock on the llog */
2014         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2015                           LCK_CR, &flags, NULL, NULL, NULL,
2016                           cld, 0, NULL, &lockh);
2017         if (rcl == 0) {
2018                 /* Get the cld, it will be released in mgc_blocking_ast. */
2019                 config_log_get(cld);
2020                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2021                 LASSERT(rc == 0);
2022         } else {
2023                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2024
2025                 if (rcl == -ESHUTDOWN &&
2026                     atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2027                         struct obd_import *imp;
2028                         struct l_wait_info lwi;
2029                         int secs = cfs_time_seconds(obd_timeout);
2030
2031                         mutex_unlock(&cld->cld_lock);
2032                         imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2033
2034                         /* Let's force the pinger, and wait the import to be
2035                          * connected, note: since mgc import is non-replayable,
2036                          * and even the import state is disconnected, it does
2037                          * not mean the "recovery" is stopped, so we will keep
2038                          * waitting until timeout or the import state is
2039                          * FULL or closed */
2040                         ptlrpc_pinger_force(imp);
2041
2042                         lwi = LWI_TIMEOUT(secs, NULL, NULL);
2043                         l_wait_event(imp->imp_recovery_waitq,
2044                                      !mgc_import_in_recovery(imp), &lwi);
2045
2046                         if (imp->imp_state == LUSTRE_IMP_FULL) {
2047                                 retry = true;
2048                                 goto restart;
2049                         } else {
2050                                 mutex_lock(&cld->cld_lock);
2051                                 cld->cld_lostlock = 1;
2052                         }
2053                 } else {
2054                         /* mark cld_lostlock so that it will requeue
2055                          * after MGC becomes available. */
2056                         cld->cld_lostlock = 1;
2057                 }
2058                 /* Get extra reference, it will be put in requeue thread */
2059                 config_log_get(cld);
2060         }
2061
2062
2063         if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2064                 if (!rcl)
2065                         rc = mgc_process_recover_nodemap_log(mgc, cld);
2066                 else if (cld_is_nodemap(cld))
2067                         rc = rcl;
2068
2069                 if (cld_is_recover(cld) && rc) {
2070                         if (!rcl) {
2071                                 CERROR("%s: recover log %s failed, not fatal: rc = %d\n",
2072                                        mgc->obd_name, cld->cld_logname, rc);
2073                                 cld->cld_lostlock = 1;
2074                         }
2075                         rc = 0; /* this is not a fatal error for recover log */
2076                 }
2077         } else {
2078                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2079         }
2080
2081         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2082                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2083
2084         mutex_unlock(&cld->cld_lock);
2085
2086         /* Now drop the lock so MGS can revoke it */
2087         if (!rcl) {
2088                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2089                 if (rcl)
2090                         CERROR("Can't drop cfg lock: %d\n", rcl);
2091         }
2092
2093         RETURN(rc);
2094 }
2095
2096
2097 /** Called from lustre_process_log.
2098  * LCFG_LOG_START gets the config log from the MGS, processes it to start
2099  * any services, and adds it to the list logs to watch (follow).
2100  */
2101 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2102 {
2103         struct lustre_cfg *lcfg = buf;
2104         struct config_llog_instance *cfg = NULL;
2105         char *logname;
2106         int rc = 0;
2107         ENTRY;
2108
2109         switch(lcfg->lcfg_command) {
2110         case LCFG_LOV_ADD_OBD: {
2111                 /* Overloading this cfg command: register a new target */
2112                 struct mgs_target_info *mti;
2113
2114                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2115                     sizeof(struct mgs_target_info))
2116                         GOTO(out, rc = -EINVAL);
2117
2118                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2119                 CDEBUG(D_MGC, "add_target %s %#x\n",
2120                        mti->mti_svname, mti->mti_flags);
2121                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2122                 break;
2123         }
2124         case LCFG_LOV_DEL_OBD:
2125                 /* Unregister has no meaning at the moment. */
2126                 CERROR("lov_del_obd unimplemented\n");
2127                 rc = -ENOSYS;
2128                 break;
2129         case LCFG_SPTLRPC_CONF: {
2130                 rc = sptlrpc_process_config(lcfg);
2131                 break;
2132         }
2133         case LCFG_LOG_START: {
2134                 struct config_llog_data *cld;
2135                 struct super_block *sb;
2136
2137                 logname = lustre_cfg_string(lcfg, 1);
2138                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2139                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2140
2141                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2142                        cfg->cfg_last_idx);
2143
2144                 /* We're only called through here on the initial mount */
2145                 rc = config_log_add(obd, logname, cfg, sb);
2146                 if (rc)
2147                         break;
2148                 cld = config_log_find(logname, cfg);
2149                 if (cld == NULL) {
2150                         rc = -ENOENT;
2151                         break;
2152                 }
2153
2154                 /* COMPAT_146 */
2155                 /* FIXME only set this for old logs!  Right now this forces
2156                    us to always skip the "inside markers" check */
2157                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
2158
2159                 rc = mgc_process_log(obd, cld);
2160                 if (rc == 0 && cld->cld_recover != NULL) {
2161                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2162                                          imp_connect_data, IMP_RECOV)) {
2163                                 rc = mgc_process_log(obd, cld->cld_recover);
2164                         } else {
2165                                 struct config_llog_data *cir = cld->cld_recover;
2166                                 cld->cld_recover = NULL;
2167                                 config_log_put(cir);
2168                         }
2169                         if (rc)
2170                                 CERROR("Cannot process recover llog %d\n", rc);
2171                 }
2172
2173                 if (rc == 0 && cld->cld_params != NULL) {
2174                         rc = mgc_process_log(obd, cld->cld_params);
2175                         if (rc == -ENOENT) {
2176                                 CDEBUG(D_MGC, "There is no params "
2177                                               "config file yet\n");
2178                                 rc = 0;
2179                         }
2180                         /* params log is optional */
2181                         if (rc)
2182                                 CERROR("%s: can't process params llog: rc = %d\n",
2183                                        obd->obd_name, rc);
2184                 }
2185                 config_log_put(cld);
2186
2187                 break;
2188         }
2189         case LCFG_LOG_END: {
2190                 logname = lustre_cfg_string(lcfg, 1);
2191
2192                 if (lcfg->lcfg_bufcount >= 2)
2193                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
2194                                 lcfg, 2);
2195                 rc = config_log_end(logname, cfg);
2196                 break;
2197         }
2198         default: {
2199                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2200                 GOTO(out, rc = -EINVAL);
2201
2202         }
2203         }
2204 out:
2205         RETURN(rc);
2206 }
2207
2208 static struct obd_ops mgc_obd_ops = {
2209         .o_owner        = THIS_MODULE,
2210         .o_setup        = mgc_setup,
2211         .o_precleanup   = mgc_precleanup,
2212         .o_cleanup      = mgc_cleanup,
2213         .o_add_conn     = client_import_add_conn,
2214         .o_del_conn     = client_import_del_conn,
2215         .o_connect      = client_connect_import,
2216         .o_disconnect   = client_disconnect_export,
2217         .o_set_info_async = mgc_set_info_async,
2218         .o_get_info       = mgc_get_info,
2219         .o_import_event = mgc_import_event,
2220         .o_process_config = mgc_process_config,
2221 };
2222
2223 static int __init mgc_init(void)
2224 {
2225         return class_register_type(&mgc_obd_ops, NULL, true, NULL,
2226                                    LUSTRE_MGC_NAME, NULL);
2227 }
2228
2229 static void __exit mgc_exit(void)
2230 {
2231         class_unregister_type(LUSTRE_MGC_NAME);
2232 }
2233
2234 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2235 MODULE_DESCRIPTION("Lustre Management Client");
2236 MODULE_VERSION(LUSTRE_VERSION_STRING);
2237 MODULE_LICENSE("GPL");
2238
2239 module_init(mgc_init);
2240 module_exit(mgc_exit);