lustre/mgc/mgc_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/mgc/mgc_request.c
  33  *
  34  * Author: Nathan Rutman <nathan@clusterfs.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_MGC
  38 #define D_MGC D_CONFIG /*|D_WARNING*/
  39
  40 #include <linux/module.h>
  41 #include <linux/kthread.h>
  42 #include <linux/random.h>
  43
  44 #include <dt_object.h>
  45 #include <lprocfs_status.h>
  46 #include <lustre_dlm.h>
  47 #include <lustre_disk.h>
  48 #include <lustre_log.h>
  49 #include <lustre_nodemap.h>
  50 #include <lustre_swab.h>
  51 #include <obd_class.h>
  52 #include <lustre_barrier.h>
  53
  54 #include "mgc_internal.h"
  55
  56 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
  57                           int type)
  58 {
  59         __u64 resname = 0;
  60
  61         if (len > sizeof(resname)) {
  62                 CERROR("name too long: %s\n", name);
  63                 return -EINVAL;
  64         }
  65         if (len <= 0) {
  66                 CERROR("missing name: %s\n", name);
  67                 return -EINVAL;
  68         }
  69         memcpy(&resname, name, len);
  70
  71         /* Always use the same endianness for the resid */
  72         memset(res_id, 0, sizeof(*res_id));
  73         res_id->name[0] = cpu_to_le64(resname);
  74         /* XXX: unfortunately, sptlprc and config llog share one lock */
  75         switch(type) {
  76         case CONFIG_T_CONFIG:
  77         case CONFIG_T_SPTLRPC:
  78                 resname = 0;
  79                 break;
  80         case CONFIG_T_RECOVER:
  81         case CONFIG_T_PARAMS:
  82         case CONFIG_T_NODEMAP:
  83         case CONFIG_T_BARRIER:
  84                 resname = type;
  85                 break;
  86         default:
  87                 LBUG();
  88         }
  89         res_id->name[1] = cpu_to_le64(resname);
  90         CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
  91                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  92         return 0;
  93 }
  94
  95 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
  96 {
  97         /* fsname is at most 8 chars long, maybe contain "-".
  98          * e.g. "lustre", "SUN-000" */
  99         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
 100 }
 101 EXPORT_SYMBOL(mgc_fsname2resid);
 102
 103 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
 104 {
 105         char *name_end;
 106         int len;
 107
 108         /* logname consists of "fsname-nodetype".
 109          * e.g. "lustre-MDT0001", "SUN-000-client"
 110          * there is an exception: llog "params" */
 111         name_end = strrchr(logname, '-');
 112         if (!name_end)
 113                 len = strlen(logname);
 114         else
 115                 len = name_end - logname;
 116         return mgc_name2resid(logname, len, res_id, type);
 117 }
 118 EXPORT_SYMBOL(mgc_logname2resid);
 119
 120 /********************** config llog list **********************/
 121 static struct list_head config_llog_list = LIST_HEAD_INIT(config_llog_list);
 122 static DEFINE_SPINLOCK(config_list_lock);       /* protects config_llog_list */
 123
 124 /* Take a reference to a config log */
 125 static int config_log_get(struct config_llog_data *cld)
 126 {
 127         ENTRY;
 128         atomic_inc(&cld->cld_refcount);
 129         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 130                 atomic_read(&cld->cld_refcount));
 131         RETURN(0);
 132 }
 133
 134 /* Drop a reference to a config log.  When no longer referenced,
 135    we can free the config log data */
 136 static void config_log_put(struct config_llog_data *cld)
 137 {
 138         ENTRY;
 139
 140         if (unlikely(!cld))
 141                 RETURN_EXIT;
 142
 143         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 144                 atomic_read(&cld->cld_refcount));
 145         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 146
 147         /* spinlock to make sure no item with 0 refcount in the list */
 148         if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 149                 list_del(&cld->cld_list_chain);
 150                 spin_unlock(&config_list_lock);
 151
 152                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 153
 154                 config_log_put(cld->cld_barrier);
 155                 config_log_put(cld->cld_recover);
 156                 config_log_put(cld->cld_params);
 157                 config_log_put(cld->cld_nodemap);
 158                 config_log_put(cld->cld_sptlrpc);
 159                 if (cld_is_sptlrpc(cld))
 160                         sptlrpc_conf_log_stop(cld->cld_logname);
 161
 162                 class_export_put(cld->cld_mgcexp);
 163                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 164         }
 165
 166         EXIT;
 167 }
 168
 169 /* Find a config log by name */
 170 static
 171 struct config_llog_data *config_log_find(char *logname,
 172                                          struct config_llog_instance *cfg)
 173 {
 174         struct config_llog_data *cld;
 175         struct config_llog_data *found = NULL;
 176         unsigned long cfg_instance;
 177
 178         ENTRY;
 179         LASSERT(logname != NULL);
 180
 181         cfg_instance = cfg ? cfg->cfg_instance : 0;
 182         spin_lock(&config_list_lock);
 183         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 184                 /* check if cfg_instance is the one we want */
 185                 if (cfg_instance != cld->cld_cfg.cfg_instance)
 186                         continue;
 187
 188                 /* instance may be NULL, should check name */
 189                 if (strcmp(logname, cld->cld_logname) == 0) {
 190                         found = cld;
 191                         config_log_get(found);
 192                         break;
 193                 }
 194         }
 195         spin_unlock(&config_list_lock);
 196         RETURN(found);
 197 }
 198
 199 static
 200 struct config_llog_data *do_config_log_add(struct obd_device *obd,
 201                                            char *logname,
 202                                            int type,
 203                                            struct config_llog_instance *cfg,
 204                                            struct super_block *sb)
 205 {
 206         struct config_llog_data *cld;
 207         int rc;
 208
 209         ENTRY;
 210
 211         CDEBUG(D_MGC, "do adding config log %s-%016lx\n", logname,
 212                cfg ? cfg->cfg_instance : 0);
 213
 214         OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
 215         if (!cld)
 216                 RETURN(ERR_PTR(-ENOMEM));
 217
 218         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 219         if (rc) {
 220                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 221                 RETURN(ERR_PTR(rc));
 222         }
 223
 224         strcpy(cld->cld_logname, logname);
 225         if (cfg)
 226                 cld->cld_cfg = *cfg;
 227         else
 228                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 229         mutex_init(&cld->cld_lock);
 230         cld->cld_cfg.cfg_last_idx = 0;
 231         cld->cld_cfg.cfg_flags = 0;
 232         cld->cld_cfg.cfg_sb = sb;
 233         cld->cld_type = type;
 234         atomic_set(&cld->cld_refcount, 1);
 235
 236         /* Keep the mgc around until we are done */
 237         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 238
 239         if (cld_is_sptlrpc(cld))
 240                 sptlrpc_conf_log_start(logname);
 241
 242         spin_lock(&config_list_lock);
 243         list_add(&cld->cld_list_chain, &config_llog_list);
 244         spin_unlock(&config_list_lock);
 245
 246         if (cld_is_sptlrpc(cld) || cld_is_nodemap(cld) || cld_is_barrier(cld)) {
 247                 rc = mgc_process_log(obd, cld);
 248                 if (rc && rc != -ENOENT)
 249                         CERROR("%s: failed processing log, type %d: rc = %d\n",
 250                                obd->obd_name, type, rc);
 251         }
 252
 253         RETURN(cld);
 254 }
 255
 256 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
 257                                         char *fsname,
 258                                         struct config_llog_instance *cfg,
 259                                         struct super_block *sb)
 260 {
 261         struct config_llog_instance lcfg = *cfg;
 262         struct lustre_sb_info *lsi = s2lsi(sb);
 263         struct config_llog_data *cld;
 264         char logname[32];
 265
 266         if (IS_OST(lsi))
 267                 return NULL;
 268
 269         /* for osp-on-ost, see lustre_start_osp() */
 270         if (IS_MDT(lsi) && lcfg.cfg_instance)
 271                 return NULL;
 272
 273         /* We have to use different llog for clients and MDTs for DNE,
 274          * where only clients are notified if one of DNE server restarts.
 275          */
 276         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 277         strncpy(logname, fsname, sizeof(logname));
 278         if (IS_SERVER(lsi)) { /* mdt */
 279                 LASSERT(lcfg.cfg_instance == 0);
 280                 lcfg.cfg_instance = ll_get_cfg_instance(sb);
 281                 strncat(logname, "-mdtir", sizeof(logname));
 282         } else {
 283                 LASSERT(lcfg.cfg_instance != 0);
 284                 strncat(logname, "-cliir", sizeof(logname));
 285         }
 286
 287         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
 288         return cld;
 289 }
 290
 291 static struct config_llog_data *config_log_find_or_add(struct obd_device *obd,
 292                                 char *logname, struct super_block *sb, int type,
 293                                 struct config_llog_instance *cfg)
 294 {
 295         struct config_llog_instance lcfg = *cfg;
 296         struct config_llog_data *cld;
 297
 298         /* Note class_config_llog_handler() depends on getting "obd" back */
 299         lcfg.cfg_instance = sb ? ll_get_cfg_instance(sb) : (unsigned long)obd;
 300
 301         cld = config_log_find(logname, &lcfg);
 302         if (unlikely(cld != NULL))
 303                 return cld;
 304
 305         return do_config_log_add(obd, logname, type, &lcfg, sb);
 306 }
 307
 308 /** Add this log to the list of active logs watched by an MGC.
 309  * Active means we're watching for updates.
 310  * We have one active log per "mount" - client instance or servername.
 311  * Each instance may be at a different point in the log.
 312  */
 313 static struct config_llog_data *
 314 config_log_add(struct obd_device *obd, char *logname,
 315                struct config_llog_instance *cfg, struct super_block *sb)
 316 {
 317         struct lustre_sb_info *lsi = s2lsi(sb);
 318         struct config_llog_data *cld = NULL;
 319         struct config_llog_data *sptlrpc_cld = NULL;
 320         struct config_llog_data *params_cld = NULL;
 321         struct config_llog_data *nodemap_cld = NULL;
 322         struct config_llog_data *barrier_cld = NULL;
 323         char seclogname[32];
 324         char *ptr;
 325         int rc;
 326         bool locked = false;
 327         ENTRY;
 328
 329         CDEBUG(D_MGC, "add config log %s-%016lx\n", logname,
 330                cfg->cfg_instance);
 331
 332         /*
 333          * for each regular log, the depended sptlrpc log name is
 334          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 335          */
 336         ptr = strrchr(logname, '-');
 337         if (ptr == NULL || ptr - logname > 8) {
 338                 CERROR("logname %s is too long\n", logname);
 339                 RETURN(ERR_PTR(-EINVAL));
 340         }
 341
 342         memcpy(seclogname, logname, ptr - logname);
 343         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 344
 345         if (cfg->cfg_sub_clds & CONFIG_SUB_SPTLRPC) {
 346                 sptlrpc_cld = config_log_find_or_add(obd, seclogname, NULL,
 347                                                      CONFIG_T_SPTLRPC, cfg);
 348                 if (IS_ERR(sptlrpc_cld)) {
 349                         CERROR("%s: can't create sptlrpc log %s: rc = %ld\n",
 350                                obd->obd_name, seclogname, PTR_ERR(sptlrpc_cld));
 351                         RETURN(sptlrpc_cld);
 352                 }
 353         }
 354
 355         if (!IS_MGS(lsi) && cfg->cfg_sub_clds & CONFIG_SUB_NODEMAP) {
 356                 nodemap_cld = config_log_find_or_add(obd, LUSTRE_NODEMAP_NAME,
 357                                                      NULL, CONFIG_T_NODEMAP,
 358                                                      cfg);
 359                 if (IS_ERR(nodemap_cld)) {
 360                         rc = PTR_ERR(nodemap_cld);
 361                         CERROR("%s: cannot create nodemap log: rc = %d\n",
 362                                obd->obd_name, rc);
 363                         GOTO(out_sptlrpc, rc);
 364                 }
 365         }
 366
 367         if (cfg->cfg_sub_clds & CONFIG_SUB_PARAMS) {
 368                 params_cld = config_log_find_or_add(obd, PARAMS_FILENAME, sb,
 369                                                     CONFIG_T_PARAMS, cfg);
 370                 if (IS_ERR(params_cld)) {
 371                         rc = PTR_ERR(params_cld);
 372                         CERROR("%s: can't create params log: rc = %d\n",
 373                                obd->obd_name, rc);
 374                         GOTO(out_nodemap, rc);
 375                 }
 376         }
 377
 378         if (IS_MDT(s2lsi(sb)) && cfg->cfg_sub_clds & CONFIG_SUB_BARRIER) {
 379                 snprintf(seclogname + (ptr - logname), sizeof(seclogname) - 1,
 380                          "-%s", BARRIER_FILENAME);
 381                 barrier_cld = config_log_find_or_add(obd, seclogname, sb,
 382                                                      CONFIG_T_BARRIER, cfg);
 383                 if (IS_ERR(barrier_cld)) {
 384                         rc = PTR_ERR(barrier_cld);
 385                         CERROR("%s: can't create barrier log: rc = %d\n",
 386                                obd->obd_name, rc);
 387                         GOTO(out_params, rc);
 388                 }
 389         }
 390
 391         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 392         if (IS_ERR(cld)) {
 393                 rc = PTR_ERR(cld);
 394                 CERROR("%s: can't create log: rc = %d\n",
 395                        obd->obd_name, rc);
 396                 GOTO(out_barrier, rc = PTR_ERR(cld));
 397         }
 398
 399         LASSERT(lsi->lsi_lmd);
 400         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR) &&
 401             cfg->cfg_sub_clds & CONFIG_SUB_RECOVER) {
 402                 struct config_llog_data *recover_cld;
 403
 404                 ptr = strrchr(seclogname, '-');
 405                 if (ptr != NULL) {
 406                         *ptr = 0;
 407                 } else {
 408                         CERROR("%s: sptlrpc log name not correct, %s: "
 409                                "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
 410                         GOTO(out_cld, rc = -EINVAL);
 411                 }
 412
 413                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 414                 if (IS_ERR(recover_cld)) {
 415                         rc = PTR_ERR(recover_cld);
 416                         CERROR("%s: can't create recover log: rc = %d\n",
 417                                obd->obd_name, rc);
 418                         GOTO(out_cld, rc);
 419                 }
 420
 421                 mutex_lock(&cld->cld_lock);
 422                 locked = true;
 423                 cld->cld_recover = recover_cld;
 424         }
 425
 426         if (!locked)
 427                 mutex_lock(&cld->cld_lock);
 428         cld->cld_params = params_cld;
 429         cld->cld_barrier = barrier_cld;
 430         cld->cld_nodemap = nodemap_cld;
 431         cld->cld_sptlrpc = sptlrpc_cld;
 432         mutex_unlock(&cld->cld_lock);
 433
 434         RETURN(cld);
 435
 436 out_cld:
 437         config_log_put(cld);
 438 out_barrier:
 439         config_log_put(barrier_cld);
 440 out_params:
 441         config_log_put(params_cld);
 442 out_nodemap:
 443         config_log_put(nodemap_cld);
 444 out_sptlrpc:
 445         config_log_put(sptlrpc_cld);
 446
 447         return ERR_PTR(rc);
 448 }
 449
 450 static inline void config_mark_cld_stop(struct config_llog_data *cld)
 451 {
 452         if (cld) {
 453                 mutex_lock(&cld->cld_lock);
 454                 spin_lock(&config_list_lock);
 455                 cld->cld_stopping = 1;
 456                 spin_unlock(&config_list_lock);
 457                 mutex_unlock(&cld->cld_lock);
 458         }
 459 }
 460
 461 /** Stop watching for updates on this log.
 462  */
 463 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 464 {
 465         struct config_llog_data *cld;
 466         struct config_llog_data *cld_sptlrpc = NULL;
 467         struct config_llog_data *cld_params = NULL;
 468         struct config_llog_data *cld_recover = NULL;
 469         struct config_llog_data *cld_nodemap = NULL;
 470         struct config_llog_data *cld_barrier = NULL;
 471         int rc = 0;
 472
 473         ENTRY;
 474
 475         cld = config_log_find(logname, cfg);
 476         if (cld == NULL)
 477                 RETURN(-ENOENT);
 478
 479         mutex_lock(&cld->cld_lock);
 480         /*
 481          * if cld_stopping is set, it means we didn't start the log thus
 482          * not owning the start ref. this can happen after previous umount:
 483          * the cld still hanging there waiting for lock cancel, and we
 484          * remount again but failed in the middle and call log_end without
 485          * calling start_log.
 486          */
 487         if (unlikely(cld->cld_stopping)) {
 488                 mutex_unlock(&cld->cld_lock);
 489                 /* drop the ref from the find */
 490                 config_log_put(cld);
 491                 RETURN(rc);
 492         }
 493
 494         spin_lock(&config_list_lock);
 495         cld->cld_stopping = 1;
 496         spin_unlock(&config_list_lock);
 497
 498         cld_recover = cld->cld_recover;
 499         cld->cld_recover = NULL;
 500         cld_params = cld->cld_params;
 501         cld->cld_params = NULL;
 502         cld_nodemap = cld->cld_nodemap;
 503         cld->cld_nodemap = NULL;
 504         cld_barrier = cld->cld_barrier;
 505         cld->cld_barrier = NULL;
 506         cld_sptlrpc = cld->cld_sptlrpc;
 507         cld->cld_sptlrpc = NULL;
 508         mutex_unlock(&cld->cld_lock);
 509
 510         config_mark_cld_stop(cld_recover);
 511         config_log_put(cld_recover);
 512
 513         config_mark_cld_stop(cld_params);
 514         config_log_put(cld_params);
 515
 516         /* don't set cld_stopping on nm lock as other targets may be active */
 517         config_log_put(cld_nodemap);
 518
 519         if (cld_barrier) {
 520                 mutex_lock(&cld_barrier->cld_lock);
 521                 cld_barrier->cld_stopping = 1;
 522                 mutex_unlock(&cld_barrier->cld_lock);
 523                 config_log_put(cld_barrier);
 524         }
 525
 526         config_log_put(cld_sptlrpc);
 527
 528         /* drop the ref from the find */
 529         config_log_put(cld);
 530         /* drop the start ref */
 531         config_log_put(cld);
 532
 533         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 534                rc);
 535         RETURN(rc);
 536 }
 537
 538 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 539 {
 540         struct obd_device       *obd = data;
 541         struct obd_import       *imp;
 542         struct obd_connect_data *ocd;
 543         struct config_llog_data *cld;
 544
 545         ENTRY;
 546         LASSERT(obd);
 547         LPROCFS_CLIMP_CHECK(obd);
 548         imp = obd->u.cli.cl_import;
 549         ocd = &imp->imp_connect_data;
 550
 551         seq_printf(m, "imperative_recovery: %s\n",
 552                    OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 553         seq_printf(m, "client_state:\n");
 554
 555         spin_lock(&config_list_lock);
 556         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 557                 if (cld->cld_recover == NULL)
 558                         continue;
 559                 seq_printf(m,  "    - { client: %s, nidtbl_version: %u }\n",
 560                            cld->cld_logname,
 561                            cld->cld_recover->cld_cfg.cfg_last_idx);
 562         }
 563         spin_unlock(&config_list_lock);
 564
 565         LPROCFS_CLIMP_EXIT(obd);
 566         RETURN(0);
 567 }
 568
 569 /* reenqueue any lost locks */
 570 #define RQ_RUNNING      0x1
 571 #define RQ_NOW          0x2
 572 #define RQ_LATER        0x4
 573 #define RQ_STOP         0x8
 574 #define RQ_PRECLEANUP   0x10
 575 static int                    rq_state = 0;
 576 static wait_queue_head_t      rq_waitq;
 577 static DECLARE_COMPLETION(rq_exit);
 578 static DECLARE_COMPLETION(rq_start);
 579
 580 static void do_requeue(struct config_llog_data *cld)
 581 {
 582         int rc = 0;
 583         ENTRY;
 584
 585         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 586
 587         /*
 588          * Do not run mgc_process_log on a disconnected export or an
 589          * export which is being disconnected. Take the client
 590          * semaphore to make the check non-racy.
 591          */
 592         down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
 593                          OBD_CLI_SEM_MGC);
 594         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 595                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 596                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 597                 if (rc && rc != -ENOENT)
 598                         CERROR("failed processing log: %d\n", rc);
 599         } else {
 600                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 601                        cld->cld_logname);
 602         }
 603         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 604
 605         EXIT;
 606 }
 607
 608 /* this timeout represents how many seconds MGC should wait before
 609  * requeue config and recover lock to the MGS. We need to randomize this
 610  * in order to not flood the MGS.
 611  */
 612 #define MGC_TIMEOUT_MIN_SECONDS   5
 613 #define MGC_TIMEOUT_RAND_CENTISEC 500
 614
 615 static int mgc_requeue_thread(void *data)
 616 {
 617         int rc = 0;
 618         bool first = true;
 619         ENTRY;
 620
 621         CDEBUG(D_MGC, "Starting requeue thread\n");
 622
 623         /* Keep trying failed locks periodically */
 624         spin_lock(&config_list_lock);
 625         rq_state |= RQ_RUNNING;
 626         while (!(rq_state & RQ_STOP)) {
 627                 struct l_wait_info lwi;
 628                 struct config_llog_data *cld, *cld_prev;
 629                 int rand = prandom_u32_max(MGC_TIMEOUT_RAND_CENTISEC);
 630                 int to;
 631
 632                 /* Any new or requeued lostlocks will change the state */
 633                 rq_state &= ~(RQ_NOW | RQ_LATER);
 634                 spin_unlock(&config_list_lock);
 635
 636                 if (first) {
 637                         first = false;
 638                         complete(&rq_start);
 639                 }
 640
 641                 /* Always wait a few seconds to allow the server who
 642                    caused the lock revocation to finish its setup, plus some
 643                    random so everyone doesn't try to reconnect at once. */
 644                 to = cfs_time_seconds(MGC_TIMEOUT_MIN_SECONDS);
 645                 /* rand is centi-seconds */
 646                 to += cfs_time_seconds(rand) / 100;
 647                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 648                 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
 649                              &lwi);
 650
 651                 /*
 652                  * iterate & processing through the list. for each cld, process
 653                  * its depending sptlrpc cld firstly (if any) and then itself.
 654                  *
 655                  * it's guaranteed any item in the list must have
 656                  * reference > 0; and if cld_lostlock is set, at
 657                  * least one reference is taken by the previous enqueue.
 658                  */
 659                 cld_prev = NULL;
 660
 661                 spin_lock(&config_list_lock);
 662                 rq_state &= ~RQ_PRECLEANUP;
 663                 list_for_each_entry(cld, &config_llog_list,
 664                                     cld_list_chain) {
 665                         if (!cld->cld_lostlock || cld->cld_stopping)
 666                                 continue;
 667
 668                         /* hold reference to avoid being freed during
 669                          * subsequent processing. */
 670                         config_log_get(cld);
 671                         cld->cld_lostlock = 0;
 672                         spin_unlock(&config_list_lock);
 673
 674                         config_log_put(cld_prev);
 675                         cld_prev = cld;
 676
 677                         if (likely(!(rq_state & RQ_STOP))) {
 678                                 do_requeue(cld);
 679                                 spin_lock(&config_list_lock);
 680                         } else {
 681                                 spin_lock(&config_list_lock);
 682                                 break;
 683                         }
 684                 }
 685                 spin_unlock(&config_list_lock);
 686                 config_log_put(cld_prev);
 687
 688                 /* Wait a bit to see if anyone else needs a requeue */
 689                 lwi = (struct l_wait_info) { 0 };
 690                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 691                              &lwi);
 692                 spin_lock(&config_list_lock);
 693         }
 694
 695         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 696         rq_state &= ~RQ_RUNNING;
 697         spin_unlock(&config_list_lock);
 698
 699         complete(&rq_exit);
 700
 701         CDEBUG(D_MGC, "Ending requeue thread\n");
 702         RETURN(rc);
 703 }
 704
 705 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 706    We are responsible for dropping the config log reference from here on out. */
 707 static void mgc_requeue_add(struct config_llog_data *cld)
 708 {
 709         bool wakeup = false;
 710         ENTRY;
 711
 712         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 713                 cld->cld_logname, atomic_read(&cld->cld_refcount),
 714                 cld->cld_stopping, rq_state);
 715         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 716
 717         mutex_lock(&cld->cld_lock);
 718         spin_lock(&config_list_lock);
 719         if (!(rq_state & RQ_STOP) && !cld->cld_stopping && !cld->cld_lostlock) {
 720                 cld->cld_lostlock = 1;
 721                 rq_state |= RQ_NOW;
 722                 wakeup = true;
 723         }
 724         spin_unlock(&config_list_lock);
 725         mutex_unlock(&cld->cld_lock);
 726         if (wakeup)
 727                 wake_up(&rq_waitq);
 728
 729         EXIT;
 730 }
 731
 732 /********************** class fns **********************/
 733 static int mgc_local_llog_init(const struct lu_env *env,
 734                                struct obd_device *obd,
 735                                struct obd_device *disk)
 736 {
 737         struct llog_ctxt        *ctxt;
 738         int                      rc;
 739
 740         ENTRY;
 741
 742         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
 743                         &llog_osd_ops);
 744         if (rc)
 745                 RETURN(rc);
 746
 747         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 748         LASSERT(ctxt);
 749         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
 750         llog_ctxt_put(ctxt);
 751
 752         RETURN(0);
 753 }
 754
 755 static int mgc_local_llog_fini(const struct lu_env *env,
 756                                struct obd_device *obd)
 757 {
 758         struct llog_ctxt *ctxt;
 759
 760         ENTRY;
 761
 762         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 763         llog_cleanup(env, ctxt);
 764
 765         RETURN(0);
 766 }
 767
 768 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
 769                         struct super_block *sb)
 770 {
 771         struct lustre_sb_info   *lsi = s2lsi(sb);
 772         struct client_obd       *cli = &obd->u.cli;
 773         struct lu_fid            rfid, fid;
 774         struct dt_object        *root, *dto;
 775         int                      rc = 0;
 776
 777         ENTRY;
 778
 779         LASSERT(lsi);
 780         LASSERT(lsi->lsi_dt_dev);
 781
 782         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
 783         mutex_lock(&cli->cl_mgc_mutex);
 784
 785         /* Setup the configs dir */
 786         fid.f_seq = FID_SEQ_LOCAL_NAME;
 787         fid.f_oid = 1;
 788         fid.f_ver = 0;
 789         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
 790                                     &cli->cl_mgc_los);
 791         if (rc)
 792                 RETURN(rc);
 793
 794         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
 795         if (rc)
 796                 GOTO(out_los, rc);
 797
 798         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
 799                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
 800         if (unlikely(IS_ERR(root)))
 801                 GOTO(out_los, rc = PTR_ERR(root));
 802
 803         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
 804                                         MOUNT_CONFIGS_DIR,
 805                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
 806         dt_object_put_nocache(env, root);
 807         if (IS_ERR(dto))
 808                 GOTO(out_los, rc = PTR_ERR(dto));
 809
 810         cli->cl_mgc_configs_dir = dto;
 811
 812         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
 813         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
 814         if (rc)
 815                 GOTO(out_llog, rc);
 816
 817         /* We take an obd ref to insure that we can't get to mgc_cleanup
 818          * without calling mgc_fs_cleanup first. */
 819         class_incref(obd, "mgc_fs", obd);
 820
 821         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 822         EXIT;
 823 out_llog:
 824         if (rc) {
 825                 dt_object_put(env, cli->cl_mgc_configs_dir);
 826                 cli->cl_mgc_configs_dir = NULL;
 827         }
 828 out_los:
 829         if (rc < 0) {
 830                 local_oid_storage_fini(env, cli->cl_mgc_los);
 831                 cli->cl_mgc_los = NULL;
 832                 mutex_unlock(&cli->cl_mgc_mutex);
 833         }
 834         return rc;
 835 }
 836
 837 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
 838 {
 839         struct client_obd       *cli = &obd->u.cli;
 840         ENTRY;
 841
 842         LASSERT(cli->cl_mgc_los != NULL);
 843
 844         mgc_local_llog_fini(env, obd);
 845
 846         dt_object_put_nocache(env, cli->cl_mgc_configs_dir);
 847         cli->cl_mgc_configs_dir = NULL;
 848
 849         local_oid_storage_fini(env, cli->cl_mgc_los);
 850         cli->cl_mgc_los = NULL;
 851
 852         class_decref(obd, "mgc_fs", obd);
 853         mutex_unlock(&cli->cl_mgc_mutex);
 854
 855         RETURN(0);
 856 }
 857
 858 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 859 {
 860         struct llog_ctxt        *ctxt;
 861         int                      rc;
 862
 863         ENTRY;
 864
 865         /* setup only remote ctxt, the local disk context is switched per each
 866          * filesystem during mgc_fs_setup() */
 867         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 868                         &llog_client_ops);
 869         if (rc)
 870                 RETURN(rc);
 871
 872         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 873         LASSERT(ctxt);
 874
 875         llog_initiator_connect(ctxt);
 876         llog_ctxt_put(ctxt);
 877
 878         RETURN(0);
 879 }
 880
 881 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 882 {
 883         struct llog_ctxt *ctxt;
 884
 885         ENTRY;
 886
 887         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 888         if (ctxt)
 889                 llog_cleanup(env, ctxt);
 890
 891         RETURN(0);
 892 }
 893
 894
 895 static atomic_t mgc_count = ATOMIC_INIT(0);
 896 static int mgc_precleanup(struct obd_device *obd)
 897 {
 898         int     rc = 0;
 899         int     temp;
 900         ENTRY;
 901
 902         if (atomic_dec_and_test(&mgc_count)) {
 903                 LASSERT(rq_state & RQ_RUNNING);
 904                 /* stop requeue thread */
 905                 temp = RQ_STOP;
 906         } else {
 907                 /* wakeup requeue thread to clean our cld */
 908                 temp = RQ_NOW | RQ_PRECLEANUP;
 909         }
 910
 911         spin_lock(&config_list_lock);
 912         rq_state |= temp;
 913         spin_unlock(&config_list_lock);
 914         wake_up(&rq_waitq);
 915
 916         if (temp & RQ_STOP)
 917                 wait_for_completion(&rq_exit);
 918         obd_cleanup_client_import(obd);
 919
 920         rc = mgc_llog_fini(NULL, obd);
 921         if (rc != 0)
 922                 CERROR("failed to cleanup llogging subsystems\n");
 923
 924         RETURN(rc);
 925 }
 926
 927 static int mgc_cleanup(struct obd_device *obd)
 928 {
 929         int rc;
 930         ENTRY;
 931
 932         /* COMPAT_146 - old config logs may have added profiles we don't
 933            know about */
 934         if (obd->obd_type->typ_refcnt <= 1)
 935                 /* Only for the last mgc */
 936                 class_del_profiles();
 937
 938         lprocfs_obd_cleanup(obd);
 939         ptlrpcd_decref();
 940
 941         rc = client_obd_cleanup(obd);
 942         RETURN(rc);
 943 }
 944
 945 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 946 {
 947         struct task_struct      *task;
 948         int                      rc;
 949         ENTRY;
 950
 951         rc = ptlrpcd_addref();
 952         if (rc < 0)
 953                 RETURN(rc);
 954
 955         rc = client_obd_setup(obd, lcfg);
 956         if (rc)
 957                 GOTO(err_decref, rc);
 958
 959         rc = mgc_llog_init(NULL, obd);
 960         if (rc) {
 961                 CERROR("failed to setup llogging subsystems\n");
 962                 GOTO(err_cleanup, rc);
 963         }
 964
 965         rc = mgc_tunables_init(obd);
 966         if (rc)
 967                 GOTO(err_sysfs, rc);
 968
 969         if (atomic_inc_return(&mgc_count) == 1) {
 970                 rq_state = 0;
 971                 init_waitqueue_head(&rq_waitq);
 972
 973                 /* start requeue thread */
 974                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 975                 if (IS_ERR(task)) {
 976                         rc = PTR_ERR(task);
 977                         CERROR("%s: cannot start requeue thread: rc = %d; "
 978                                "no more log updates\n",
 979                                obd->obd_name, rc);
 980                         GOTO(err_sysfs, rc);
 981                 }
 982                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 983                 rc = 0;
 984                 wait_for_completion(&rq_start);
 985         }
 986
 987         RETURN(rc);
 988
 989 err_sysfs:
 990         lprocfs_obd_cleanup(obd);
 991 err_cleanup:
 992         client_obd_cleanup(obd);
 993 err_decref:
 994         ptlrpcd_decref();
 995         RETURN(rc);
 996 }
 997
 998 /* based on ll_mdc_blocking_ast */
 999 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1000                             void *data, int flag)
1001 {
1002         struct lustre_handle lockh;
1003         struct config_llog_data *cld = (struct config_llog_data *)data;
1004         int rc = 0;
1005         ENTRY;
1006
1007         switch (flag) {
1008         case LDLM_CB_BLOCKING:
1009                 /* mgs wants the lock, give it up... */
1010                 LDLM_DEBUG(lock, "MGC blocking CB");
1011                 ldlm_lock2handle(lock, &lockh);
1012                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
1013                 break;
1014         case LDLM_CB_CANCELING:
1015                 /* We've given up the lock, prepare ourselves to update. */
1016                 LDLM_DEBUG(lock, "MGC cancel CB");
1017
1018                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
1019                        PLDLMRES(lock->l_resource),
1020                        (char *)&lock->l_resource->lr_name.name[0]);
1021
1022                 if (!cld) {
1023                         CDEBUG(D_INFO, "missing data, won't requeue\n");
1024                         break;
1025                 }
1026
1027                 /* held at mgc_process_log(). */
1028                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1029
1030                 lock->l_ast_data = NULL;
1031                 /* Are we done with this log? */
1032                 if (cld->cld_stopping) {
1033                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1034                                 cld->cld_logname);
1035                         config_log_put(cld);
1036                         break;
1037                 }
1038                 /* Make sure not to re-enqueue when the mgc is stopping
1039                    (we get called from client_disconnect_export) */
1040                 if (lock->l_conn_export == NULL ||
1041                     lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1042                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1043                                 cld->cld_logname);
1044                         config_log_put(cld);
1045                         break;
1046                 }
1047
1048                 /* Re-enqueue now */
1049                 mgc_requeue_add(cld);
1050                 config_log_put(cld);
1051                 break;
1052         default:
1053                 LBUG();
1054         }
1055
1056         RETURN(rc);
1057 }
1058
1059 /* Not sure where this should go... */
1060 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1061  * that we can have a chance to try the secondary MGS if any. */
1062 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1063                                 + PING_INTERVAL)
1064 #define  MGC_TARGET_REG_LIMIT 10
1065 #define  MGC_TARGET_REG_LIMIT_MAX RECONNECT_DELAY_MAX
1066 #define  MGC_SEND_PARAM_LIMIT 10
1067
1068 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1069 /* Send parameter to MGS*/
1070 static int mgc_set_mgs_param(struct obd_export *exp,
1071                              struct mgs_send_param *msp)
1072 {
1073         struct ptlrpc_request *req;
1074         struct mgs_send_param *req_msp, *rep_msp;
1075         int rc;
1076         ENTRY;
1077
1078         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1079                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1080                                         MGS_SET_INFO);
1081         if (!req)
1082                 RETURN(-ENOMEM);
1083
1084         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1085         if (!req_msp) {
1086                 ptlrpc_req_finished(req);
1087                 RETURN(-ENOMEM);
1088         }
1089
1090         memcpy(req_msp, msp, sizeof(*req_msp));
1091         ptlrpc_request_set_replen(req);
1092
1093         /* Limit how long we will wait for the enqueue to complete */
1094         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1095         rc = ptlrpc_queue_wait(req);
1096         if (!rc) {
1097                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1098                 memcpy(msp, rep_msp, sizeof(*rep_msp));
1099         }
1100
1101         ptlrpc_req_finished(req);
1102
1103         RETURN(rc);
1104 }
1105 #endif
1106
1107 /* Take a config lock so we can get cancel notifications */
1108 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1109                        union ldlm_policy_data *policy, enum ldlm_mode mode,
1110                        __u64 *flags, ldlm_glimpse_callback glimpse_callback,
1111                        void *data, __u32 lvb_len, void *lvb_swabber,
1112                        struct lustre_handle *lockh)
1113 {
1114         struct config_llog_data *cld = (struct config_llog_data *)data;
1115         struct ldlm_enqueue_info einfo = {
1116                 .ei_type        = type,
1117                 .ei_mode        = mode,
1118                 .ei_cb_bl       = mgc_blocking_ast,
1119                 .ei_cb_cp       = ldlm_completion_ast,
1120                 .ei_cb_gl       = glimpse_callback,
1121         };
1122         struct ptlrpc_request *req;
1123         int short_limit = cld_is_sptlrpc(cld);
1124         int rc;
1125         ENTRY;
1126
1127         CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1128                cld->cld_resid.name[0]);
1129
1130         /* We need a callback for every lockholder, so don't try to
1131            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1132         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1133                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1134                                         LDLM_ENQUEUE);
1135         if (req == NULL)
1136                 RETURN(-ENOMEM);
1137
1138         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1139         ptlrpc_request_set_replen(req);
1140
1141         /* check if this is server or client */
1142         if (cld->cld_cfg.cfg_sb) {
1143                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1144                 if (lsi && IS_SERVER(lsi))
1145                         short_limit = 1;
1146         }
1147         /* Limit how long we will wait for the enqueue to complete */
1148         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1149         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1150                               NULL, 0, LVB_T_NONE, lockh, 0);
1151         /* A failed enqueue should still call the mgc_blocking_ast,
1152            where it will be requeued if needed ("grant failed"). */
1153         ptlrpc_req_finished(req);
1154         RETURN(rc);
1155 }
1156
1157 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1158                       struct lustre_handle *lockh)
1159 {
1160         ENTRY;
1161
1162         ldlm_lock_decref(lockh, mode);
1163
1164         RETURN(0);
1165 }
1166
1167 static void mgc_notify_active(struct obd_device *unused)
1168 {
1169         /* wakeup mgc_requeue_thread to requeue mgc lock */
1170         spin_lock(&config_list_lock);
1171         rq_state |= RQ_NOW;
1172         spin_unlock(&config_list_lock);
1173         wake_up(&rq_waitq);
1174
1175         /* TODO: Help the MGS rebuild nidtbl. -jay */
1176 }
1177
1178 /* Send target_reg message to MGS */
1179 static int mgc_target_register(struct obd_export *exp,
1180                                struct mgs_target_info *mti)
1181 {
1182         struct ptlrpc_request  *req;
1183         struct mgs_target_info *req_mti, *rep_mti;
1184         int                     rc;
1185         ENTRY;
1186
1187         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1188                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1189                                         MGS_TARGET_REG);
1190         if (req == NULL)
1191                 RETURN(-ENOMEM);
1192
1193         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1194         if (!req_mti) {
1195                 ptlrpc_req_finished(req);
1196                 RETURN(-ENOMEM);
1197         }
1198
1199         memcpy(req_mti, mti, sizeof(*req_mti));
1200         ptlrpc_request_set_replen(req);
1201         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1202         /* Limit how long we will wait for the enqueue to complete */
1203         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1204
1205         /* if the target needs to regenerate the config log in MGS, it's better
1206          * to use some longer limit to let MGC have time to change connection to
1207          * another MGS (or try again with the same MGS) for the target (server)
1208          * will fail and exit if the request expired due to delay limit. */
1209         if (mti->mti_flags & (LDD_F_UPDATE | LDD_F_NEED_INDEX))
1210                 req->rq_delay_limit = MGC_TARGET_REG_LIMIT_MAX;
1211
1212         rc = ptlrpc_queue_wait(req);
1213         if (!rc) {
1214                 rep_mti = req_capsule_server_get(&req->rq_pill,
1215                                                  &RMF_MGS_TARGET_INFO);
1216                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1217                 CDEBUG(D_MGC, "register %s got index = %d\n",
1218                        mti->mti_svname, mti->mti_stripe_index);
1219         }
1220         ptlrpc_req_finished(req);
1221
1222         RETURN(rc);
1223 }
1224
1225 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1226                               u32 keylen, void *key,
1227                               u32 vallen, void *val,
1228                               struct ptlrpc_request_set *set)
1229 {
1230         int rc = -EINVAL;
1231         ENTRY;
1232
1233         /* Turn off initial_recov after we try all backup servers once */
1234         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1235                 struct obd_import *imp = class_exp2cliimp(exp);
1236                 int value;
1237                 if (vallen != sizeof(int))
1238                         RETURN(-EINVAL);
1239                 value = *(int *)val;
1240                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1241                        imp->imp_obd->obd_name, value,
1242                        imp->imp_deactive, imp->imp_invalid,
1243                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1244                        ptlrpc_import_state_name(imp->imp_state));
1245                 /* Resurrect the import immediately if
1246                  * 1. we previously got disconnected,
1247                  * 2. value > 1 (at the same node with MGS)
1248                  * */
1249                 if (imp->imp_state == LUSTRE_IMP_DISCON || value > 1)
1250                         ptlrpc_reconnect_import(imp);
1251
1252                 RETURN(0);
1253         }
1254
1255         /* FIXME move this to mgc_process_config */
1256         if (KEY_IS(KEY_REGISTER_TARGET)) {
1257                 struct mgs_target_info *mti;
1258                 if (vallen != sizeof(struct mgs_target_info))
1259                         RETURN(-EINVAL);
1260                 mti = (struct mgs_target_info *)val;
1261                 CDEBUG(D_MGC, "register_target %s %#x\n",
1262                        mti->mti_svname, mti->mti_flags);
1263                 rc =  mgc_target_register(exp, mti);
1264                 RETURN(rc);
1265         }
1266         if (KEY_IS(KEY_SET_FS)) {
1267                 struct super_block *sb = (struct super_block *)val;
1268
1269                 if (vallen != sizeof(struct super_block))
1270                         RETURN(-EINVAL);
1271
1272                 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1273                 RETURN(rc);
1274         }
1275         if (KEY_IS(KEY_CLEAR_FS)) {
1276                 if (vallen != 0)
1277                         RETURN(-EINVAL);
1278                 rc = mgc_fs_cleanup(env, exp->exp_obd);
1279                 RETURN(rc);
1280         }
1281 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1282         if (KEY_IS(KEY_SET_INFO)) {
1283                 struct mgs_send_param *msp;
1284
1285                 msp = (struct mgs_send_param *)val;
1286                 rc =  mgc_set_mgs_param(exp, msp);
1287                 RETURN(rc);
1288         }
1289 #endif
1290         if (KEY_IS(KEY_MGSSEC)) {
1291                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1292                 struct sptlrpc_flavor  flvr;
1293
1294                 /*
1295                  * empty string means using current flavor, if which haven't
1296                  * been set yet, set it as null.
1297                  *
1298                  * if flavor has been set previously, check the asking flavor
1299                  * must match the existing one.
1300                  */
1301                 if (vallen == 0) {
1302                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1303                                 RETURN(0);
1304                         val = "null";
1305                         vallen = 4;
1306                 }
1307
1308                 rc = sptlrpc_parse_flavor(val, &flvr);
1309                 if (rc) {
1310                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1311                                (char *) val);
1312                         RETURN(rc);
1313                 }
1314
1315                 /*
1316                  * caller already hold a mutex
1317                  */
1318                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1319                         cli->cl_flvr_mgc = flvr;
1320                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1321                                   sizeof(flvr)) != 0) {
1322                         char    str[20];
1323
1324                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1325                                             str, sizeof(str));
1326                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1327                                        "currently %s is in use\n",
1328                                        (char *) val, str);
1329                         rc = -EPERM;
1330                 }
1331                 RETURN(rc);
1332         }
1333
1334         RETURN(rc);
1335 }
1336
1337 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1338                         __u32 keylen, void *key, __u32 *vallen, void *val)
1339 {
1340         int rc = -EINVAL;
1341
1342         if (KEY_IS(KEY_CONN_DATA)) {
1343                 struct obd_import *imp = class_exp2cliimp(exp);
1344                 struct obd_connect_data *data = val;
1345
1346                 if (*vallen == sizeof(*data)) {
1347                         *data = imp->imp_connect_data;
1348                         rc = 0;
1349                 }
1350         }
1351
1352         return rc;
1353 }
1354
1355 static int mgc_import_event(struct obd_device *obd,
1356                             struct obd_import *imp,
1357                             enum obd_import_event event)
1358 {
1359         int rc = 0;
1360
1361         LASSERT(imp->imp_obd == obd);
1362         CDEBUG(D_MGC, "import event %#x\n", event);
1363
1364         switch (event) {
1365         case IMP_EVENT_DISCON:
1366                 /* MGC imports should not wait for recovery */
1367                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1368                         ptlrpc_pinger_ir_down();
1369                 break;
1370         case IMP_EVENT_INACTIVE:
1371                 break;
1372         case IMP_EVENT_INVALIDATE: {
1373                 struct ldlm_namespace *ns = obd->obd_namespace;
1374                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1375                 break;
1376         }
1377         case IMP_EVENT_ACTIVE:
1378                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1379                 /* Clearing obd_no_recov allows us to continue pinging */
1380                 obd->obd_no_recov = 0;
1381                 mgc_notify_active(obd);
1382                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1383                         ptlrpc_pinger_ir_up();
1384                 break;
1385         case IMP_EVENT_OCD:
1386                 break;
1387         case IMP_EVENT_DEACTIVATE:
1388         case IMP_EVENT_ACTIVATE:
1389                 break;
1390         default:
1391                 CERROR("Unknown import event %#x\n", event);
1392                 LBUG();
1393         }
1394         RETURN(rc);
1395 }
1396
1397 enum {
1398         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1399         CONFIG_READ_NRPAGES      = 4
1400 };
1401
1402 static int mgc_apply_recover_logs(struct obd_device *mgc,
1403                                   struct config_llog_data *cld,
1404                                   __u64 max_version,
1405                                   void *data, int datalen, bool mne_swab)
1406 {
1407         struct config_llog_instance *cfg = &cld->cld_cfg;
1408         struct lustre_sb_info *lsi = s2lsi(cfg->cfg_sb);
1409         struct mgs_nidtbl_entry *entry;
1410         struct lustre_cfg *lcfg;
1411         struct lustre_cfg_bufs bufs;
1412         u64 prev_version = 0;
1413         char *inst;
1414         char *buf;
1415         int bufsz;
1416         int pos = 0;
1417         int rc  = 0;
1418         int off = 0;
1419
1420         ENTRY;
1421         LASSERT(cfg->cfg_instance != 0);
1422         LASSERT(ll_get_cfg_instance(cfg->cfg_sb) == cfg->cfg_instance);
1423
1424         OBD_ALLOC(inst, PAGE_SIZE);
1425         if (inst == NULL)
1426                 RETURN(-ENOMEM);
1427
1428         if (!IS_SERVER(lsi)) {
1429                 pos = snprintf(inst, PAGE_SIZE, "%016lx", cfg->cfg_instance);
1430                 if (pos >= PAGE_SIZE) {
1431                         OBD_FREE(inst, PAGE_SIZE);
1432                         return -E2BIG;
1433                 }
1434 #ifdef HAVE_SERVER_SUPPORT
1435         } else {
1436                 LASSERT(IS_MDT(lsi));
1437                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1438                                         PAGE_SIZE);
1439                 if (rc) {
1440                         OBD_FREE(inst, PAGE_SIZE);
1441                         RETURN(-EINVAL);
1442                 }
1443                 pos = strlen(inst);
1444 #endif /* HAVE_SERVER_SUPPORT */
1445         }
1446
1447         ++pos;
1448         buf   = inst + pos;
1449         bufsz = PAGE_SIZE - pos;
1450
1451         while (datalen > 0) {
1452                 int   entry_len = sizeof(*entry);
1453                 int   is_ost, i;
1454                 struct obd_device *obd;
1455                 char *obdname;
1456                 char *cname;
1457                 char *params;
1458                 char *uuid;
1459
1460                 rc = -EINVAL;
1461                 if (datalen < sizeof(*entry))
1462                         break;
1463
1464                 entry = (typeof(entry))(data + off);
1465
1466                 /* sanity check */
1467                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1468                         break;
1469                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1470                         break;
1471                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1472                         break;
1473
1474                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1475                 if (datalen < entry_len) /* must have entry_len at least */
1476                         break;
1477
1478                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1479                 if (mne_swab)
1480                         lustre_swab_mgs_nidtbl_entry(entry);
1481                 if (entry->mne_length > PAGE_SIZE) {
1482                         CERROR("MNE too large (%u)\n", entry->mne_length);
1483                         break;
1484                 }
1485
1486                 if (entry->mne_length < entry_len)
1487                         break;
1488
1489                 off     += entry->mne_length;
1490                 datalen -= entry->mne_length;
1491                 if (datalen < 0)
1492                         break;
1493
1494                 if (entry->mne_version > max_version) {
1495                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1496                                entry->mne_version, max_version);
1497                         break;
1498                 }
1499
1500                 if (prev_version >= entry->mne_version) {
1501                         CERROR("index unsorted, prev %lld, now %lld\n",
1502                                prev_version, entry->mne_version);
1503                         break;
1504                 }
1505                 prev_version = entry->mne_version;
1506
1507                 /*
1508                  * Write a string with format "nid::instance" to
1509                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1510                  */
1511
1512                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1513                 memset(buf, 0, bufsz);
1514                 obdname = buf;
1515                 pos = 0;
1516
1517                 /* lustre-OST0001-osc-<instance #> */
1518                 strcpy(obdname, cld->cld_logname);
1519                 cname = strrchr(obdname, '-');
1520                 if (cname == NULL) {
1521                         CERROR("mgc %s: invalid logname %s\n",
1522                                mgc->obd_name, obdname);
1523                         break;
1524                 }
1525
1526                 pos = cname - obdname;
1527                 obdname[pos] = 0;
1528                 pos += sprintf(obdname + pos, "-%s%04x",
1529                                   is_ost ? "OST" : "MDT", entry->mne_index);
1530
1531                 cname = is_ost ? "osc" : "mdc",
1532                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1533                 lustre_cfg_bufs_reset(&bufs, obdname);
1534
1535                 /* find the obd by obdname */
1536                 obd = class_name2obd(obdname);
1537                 if (obd == NULL) {
1538                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1539                                mgc->obd_name, obdname);
1540                         rc = 0;
1541                         /* this is a safe race, when the ost is starting up...*/
1542                         continue;
1543                 }
1544
1545                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1546                 ++pos;
1547                 params = buf + pos;
1548                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1549                 uuid = buf + pos;
1550
1551                 down_read(&obd->u.cli.cl_sem);
1552                 if (obd->u.cli.cl_import == NULL) {
1553                         /* client does not connect to the OST yet */
1554                         up_read(&obd->u.cli.cl_sem);
1555                         rc = 0;
1556                         continue;
1557                 }
1558
1559                 /* iterate all nids to find one */
1560                 /* find uuid by nid */
1561                 rc = -ENOENT;
1562                 for (i = 0; i < entry->mne_nid_count; i++) {
1563                         rc = client_import_find_conn(obd->u.cli.cl_import,
1564                                                      entry->u.nids[i],
1565                                                      (struct obd_uuid *)uuid);
1566                         if (rc == 0)
1567                                 break;
1568                 }
1569
1570                 up_read(&obd->u.cli.cl_sem);
1571                 if (rc < 0) {
1572                         CERROR("mgc: cannot find uuid by nid %s\n",
1573                                libcfs_nid2str(entry->u.nids[0]));
1574                         break;
1575                 }
1576
1577                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1578                        uuid, libcfs_nid2str(entry->u.nids[0]));
1579
1580                 pos += strlen(uuid);
1581                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1582                 LASSERT(pos < bufsz);
1583
1584                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1585
1586                 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1587                                                bufs.lcfg_buflen));
1588                 if (!lcfg) {
1589                         rc = -ENOMEM;
1590                         break;
1591                 }
1592                 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1593
1594                 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1595                        prev_version, max_version, obdname, params);
1596
1597                 rc = class_process_config(lcfg);
1598                 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1599                                               lcfg->lcfg_buflens));
1600                 if (rc)
1601                         CDEBUG(D_INFO, "process config for %s error %d\n",
1602                                obdname, rc);
1603
1604                 /* continue, even one with error */
1605         }
1606
1607         OBD_FREE(inst, PAGE_SIZE);
1608         RETURN(rc);
1609 }
1610
1611 /**
1612  * This function is called if this client was notified for target restarting
1613  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1614  * nodemap logs.
1615  */
1616 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1617                                            struct config_llog_data *cld)
1618 {
1619         struct ptlrpc_connection *mgc_conn;
1620         struct ptlrpc_request *req = NULL;
1621         struct config_llog_instance *cfg = &cld->cld_cfg;
1622         struct mgs_config_body *body;
1623         struct mgs_config_res *res;
1624         struct nodemap_config *new_config = NULL;
1625         struct lu_nodemap *recent_nodemap = NULL;
1626         struct ptlrpc_bulk_desc *desc;
1627         struct page **pages = NULL;
1628         __u64 config_read_offset = 0;
1629         __u8 nodemap_cur_pass = 0;
1630         int nrpages = 0;
1631         bool eof = true;
1632         bool mne_swab = false;
1633         int i;
1634         int ealen;
1635         int rc;
1636         ENTRY;
1637
1638         mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1639
1640         /* don't need to get local config */
1641         if (cld_is_nodemap(cld) &&
1642             (LNET_NETTYP(LNET_NIDNET(mgc_conn->c_peer.nid)) == LOLND))
1643                 GOTO(out, rc = 0);
1644
1645         /* allocate buffer for bulk transfer.
1646          * if this is the first time for this mgs to read logs,
1647          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1648          * once; otherwise, it only reads increment of logs, this should be
1649          * small and CONFIG_READ_NRPAGES will be used.
1650          */
1651         nrpages = CONFIG_READ_NRPAGES;
1652         if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1653                 nrpages = CONFIG_READ_NRPAGES_INIT;
1654
1655         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1656         if (pages == NULL)
1657                 GOTO(out, rc = -ENOMEM);
1658
1659         for (i = 0; i < nrpages; i++) {
1660                 pages[i] = alloc_page(GFP_KERNEL);
1661                 if (pages[i] == NULL)
1662                         GOTO(out, rc = -ENOMEM);
1663         }
1664
1665 again:
1666 #ifdef HAVE_SERVER_SUPPORT
1667         if (cld_is_nodemap(cld) && config_read_offset == 0) {
1668                 new_config = nodemap_config_alloc();
1669                 if (IS_ERR(new_config)) {
1670                         rc = PTR_ERR(new_config);
1671                         new_config = NULL;
1672                         GOTO(out, rc);
1673                 }
1674         }
1675 #endif
1676         LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1677         LASSERT(mutex_is_locked(&cld->cld_lock));
1678         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1679                                    &RQF_MGS_CONFIG_READ);
1680         if (req == NULL)
1681                 GOTO(out, rc = -ENOMEM);
1682
1683         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1684         if (rc)
1685                 GOTO(out, rc);
1686
1687         /* pack request */
1688         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1689         LASSERT(body != NULL);
1690         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1691         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1692             >= sizeof(body->mcb_name))
1693                 GOTO(out, rc = -E2BIG);
1694         if (cld_is_nodemap(cld))
1695                 body->mcb_offset = config_read_offset;
1696         else
1697                 body->mcb_offset = cfg->cfg_last_idx + 1;
1698         body->mcb_type   = cld->cld_type;
1699         body->mcb_bits   = PAGE_SHIFT;
1700         body->mcb_units  = nrpages;
1701         body->mcb_nm_cur_pass = nodemap_cur_pass;
1702
1703         /* allocate bulk transfer descriptor */
1704         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1705                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1706                                     MGS_BULK_PORTAL,
1707                                     &ptlrpc_bulk_kiov_pin_ops);
1708         if (desc == NULL)
1709                 GOTO(out, rc = -ENOMEM);
1710
1711         for (i = 0; i < nrpages; i++)
1712                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1713                                                  PAGE_SIZE);
1714
1715         ptlrpc_request_set_replen(req);
1716         rc = ptlrpc_queue_wait(req);
1717         if (rc)
1718                 GOTO(out, rc);
1719
1720         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1721         if (!res)
1722                 GOTO(out, rc = -EPROTO);
1723
1724         if (cld_is_nodemap(cld)) {
1725                 config_read_offset = res->mcr_offset;
1726                 eof = config_read_offset == II_END_OFF;
1727                 nodemap_cur_pass = res->mcr_nm_cur_pass;
1728         } else {
1729                 if (res->mcr_size < res->mcr_offset)
1730                         GOTO(out, rc = -EINVAL);
1731
1732                 /* always update the index even though it might have errors with
1733                  * handling the recover logs
1734                  */
1735                 cfg->cfg_last_idx = res->mcr_offset;
1736                 eof = res->mcr_offset == res->mcr_size;
1737
1738                 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1739                        res->mcr_offset, eof == false);
1740         }
1741
1742         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1743         if (ealen < 0)
1744                 GOTO(out, rc = ealen);
1745
1746         if (ealen > nrpages << PAGE_SHIFT)
1747                 GOTO(out, rc = -EINVAL);
1748
1749         if (ealen == 0) { /* no logs transferred */
1750 #ifdef HAVE_SERVER_SUPPORT
1751                 /* config changed since first read RPC */
1752                 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1753                         CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1754                         GOTO(out, rc = -EAGAIN);
1755                 }
1756 #endif
1757                 if (!eof)
1758                         rc = -EINVAL;
1759                 GOTO(out, rc);
1760         }
1761
1762         mne_swab = ptlrpc_rep_need_swab(req);
1763
1764         /* When a nodemap config is received, we build a new nodemap config,
1765          * with new nodemap structs. We keep track of the most recently added
1766          * nodemap since the config is read ordered by nodemap_id, and so it
1767          * is likely that the next record will be related. Because access to
1768          * the nodemaps is single threaded until the nodemap_config is active,
1769          * we don't need to reference count with recent_nodemap, though
1770          * recent_nodemap should be set to NULL when the nodemap_config
1771          * is either destroyed or set active.
1772          */
1773         for (i = 0; i < nrpages && ealen > 0; i++) {
1774                 int rc2;
1775                 union lu_page   *ptr;
1776
1777                 ptr = kmap(pages[i]);
1778                 if (cld_is_nodemap(cld))
1779                         rc2 = nodemap_process_idx_pages(new_config, ptr,
1780                                                        &recent_nodemap);
1781                 else
1782                         rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1783                                                      ptr,
1784                                                      min_t(int, ealen,
1785                                                            PAGE_SIZE),
1786                                                      mne_swab);
1787                 kunmap(pages[i]);
1788                 if (rc2 < 0) {
1789                         CWARN("%s: error processing %s log %s: rc = %d\n",
1790                               obd->obd_name,
1791                               cld_is_nodemap(cld) ? "nodemap" : "recovery",
1792                               cld->cld_logname,
1793                               rc2);
1794                         GOTO(out, rc = rc2);
1795                 }
1796
1797                 ealen -= PAGE_SIZE;
1798         }
1799
1800 out:
1801         if (req) {
1802                 ptlrpc_req_finished(req);
1803                 req = NULL;
1804         }
1805
1806         if (rc == 0 && !eof)
1807                 goto again;
1808
1809 #ifdef HAVE_SERVER_SUPPORT
1810         if (new_config != NULL) {
1811                 /* recent_nodemap cannot be used after set_active/dealloc */
1812                 if (rc == 0)
1813                         nodemap_config_set_active_mgc(new_config);
1814                 else
1815                         nodemap_config_dealloc(new_config);
1816         }
1817 #endif
1818
1819         if (pages) {
1820                 for (i = 0; i < nrpages; i++) {
1821                         if (pages[i] == NULL)
1822                                 break;
1823                         __free_page(pages[i]);
1824                 }
1825                 OBD_FREE(pages, sizeof(*pages) * nrpages);
1826         }
1827         return rc;
1828 }
1829
1830 static int mgc_barrier_glimpse_ast(struct ldlm_lock *lock, void *data)
1831 {
1832         struct config_llog_data *cld = lock->l_ast_data;
1833         int rc;
1834         ENTRY;
1835
1836         if (cld->cld_stopping)
1837                 RETURN(-ENODEV);
1838
1839         rc = barrier_handler(s2lsi(cld->cld_cfg.cfg_sb)->lsi_dt_dev,
1840                              (struct ptlrpc_request *)data);
1841
1842         RETURN(rc);
1843 }
1844
1845 /* Copy a remote log locally */
1846 static int mgc_llog_local_copy(const struct lu_env *env,
1847                                struct obd_device *obd,
1848                                struct llog_ctxt *rctxt,
1849                                struct llog_ctxt *lctxt, char *logname)
1850 {
1851         char    *temp_log;
1852         int      rc;
1853
1854         ENTRY;
1855
1856         /*
1857          * - copy it to backup using llog_backup()
1858          * - copy remote llog to logname using llog_backup()
1859          * - if failed then move bakup to logname again
1860          */
1861
1862         OBD_ALLOC(temp_log, strlen(logname) + 2);
1863         if (!temp_log)
1864                 RETURN(-ENOMEM);
1865         sprintf(temp_log, "%sT", logname);
1866
1867         /* make a copy of local llog at first */
1868         rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1869         if (rc < 0 && rc != -ENOENT)
1870                 GOTO(out, rc);
1871         /* copy remote llog to the local copy */
1872         rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1873         if (rc == -ENOENT) {
1874                 /* no remote llog, delete local one too */
1875                 llog_erase(env, lctxt, NULL, logname);
1876         } else if (rc < 0) {
1877                 /* error during backup, get local one back from the copy */
1878                 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1879 out:
1880                 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1881                        obd->obd_name, logname, rc);
1882         }
1883         llog_erase(env, lctxt, NULL, temp_log);
1884         OBD_FREE(temp_log, strlen(logname) + 2);
1885         return rc;
1886 }
1887
1888 /* local_only means it cannot get remote llogs */
1889 static int mgc_process_cfg_log(struct obd_device *mgc,
1890                                struct config_llog_data *cld, int local_only)
1891 {
1892         struct llog_ctxt        *ctxt, *lctxt = NULL;
1893         struct client_obd       *cli = &mgc->u.cli;
1894         struct lustre_sb_info   *lsi = NULL;
1895         int                      rc = 0;
1896         struct lu_env           *env;
1897
1898         ENTRY;
1899
1900         LASSERT(cld);
1901         LASSERT(mutex_is_locked(&cld->cld_lock));
1902
1903         if (cld->cld_cfg.cfg_sb)
1904                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1905
1906         OBD_ALLOC_PTR(env);
1907         if (env == NULL)
1908                 RETURN(-ENOMEM);
1909
1910         rc = lu_env_init(env, LCT_MG_THREAD);
1911         if (rc)
1912                 GOTO(out_free, rc);
1913
1914         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1915         LASSERT(ctxt);
1916
1917         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1918
1919         /* Copy the setup log locally if we can. Don't mess around if we're
1920          * running an MGS though (logs are already local). */
1921         if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1922             cli->cl_mgc_configs_dir != NULL &&
1923             lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1924             lsi->lsi_dt_dev) {
1925                 if (!local_only && !lsi->lsi_dt_dev->dd_rdonly)
1926                         /* Only try to copy log if we have the lock. */
1927                         rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1928                                                  cld->cld_logname);
1929                 if (local_only || rc) {
1930                         if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1931                             llog_is_empty(env, lctxt, cld->cld_logname)) {
1932                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1933                                                    "log %s and no local copy."
1934                                                    "\n", cld->cld_logname);
1935                                 GOTO(out_pop, rc = -ENOENT);
1936                         }
1937                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1938                                "copy for now, will try to update later.\n",
1939                                cld->cld_logname);
1940                         rc = 0;
1941                 }
1942                 /* Now, whether we copied or not, start using the local llog.
1943                  * If we failed to copy, we'll start using whatever the old
1944                  * log has. */
1945                 llog_ctxt_put(ctxt);
1946                 ctxt = lctxt;
1947                 lctxt = NULL;
1948         } else {
1949                 if (local_only) /* no local log at client side */
1950                         GOTO(out_pop, rc = -EIO);
1951         }
1952
1953         rc = -EAGAIN;
1954         if (lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1955             lsi->lsi_dt_dev->dd_rdonly) {
1956                 struct llog_ctxt *rctxt;
1957
1958                 /* Under readonly mode, we may have no local copy or local
1959                  * copy is incomplete, so try to use remote llog firstly. */
1960                 rctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1961                 LASSERT(rctxt);
1962
1963                 rc = class_config_parse_llog(env, rctxt, cld->cld_logname,
1964                                              &cld->cld_cfg);
1965                 llog_ctxt_put(rctxt);
1966         }
1967
1968         if (rc && rc != -ENOENT)
1969                 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1970                                              &cld->cld_cfg);
1971
1972         /*
1973          * update settings on existing OBDs.
1974          * the logname must be <fsname>-sptlrpc
1975          */
1976         if (rc == 0 && cld_is_sptlrpc(cld))
1977                 class_notify_sptlrpc_conf(cld->cld_logname,
1978                                           strlen(cld->cld_logname) -
1979                                           strlen("-sptlrpc"));
1980         EXIT;
1981
1982 out_pop:
1983         __llog_ctxt_put(env, ctxt);
1984         if (lctxt)
1985                 __llog_ctxt_put(env, lctxt);
1986
1987         lu_env_fini(env);
1988 out_free:
1989         OBD_FREE_PTR(env);
1990         return rc;
1991 }
1992
1993 static bool mgc_import_in_recovery(struct obd_import *imp)
1994 {
1995         bool in_recovery = true;
1996
1997         spin_lock(&imp->imp_lock);
1998         if (imp->imp_state == LUSTRE_IMP_FULL ||
1999             imp->imp_state == LUSTRE_IMP_CLOSED)
2000                 in_recovery = false;
2001         spin_unlock(&imp->imp_lock);
2002
2003         return in_recovery;
2004 }
2005
2006 /**
2007  * Get a configuration log from the MGS and process it.
2008  *
2009  * This function is called for both clients and servers to process the
2010  * configuration log from the MGS.  The MGC enqueues a DLM lock on the
2011  * log from the MGS, and if the lock gets revoked the MGC will be notified
2012  * by the lock cancellation callback that the config log has changed,
2013  * and will enqueue another MGS lock on it, and then continue processing
2014  * the new additions to the end of the log.
2015  *
2016  * Since the MGC import is not replayable, if the import is being evicted
2017  * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
2018  * the log until recovery is finished or the import is closed.
2019  *
2020  * Make a local copy of the log before parsing it if appropriate (non-MGS
2021  * server) so that the server can start even when the MGS is down.
2022  *
2023  * There shouldn't be multiple processes running process_log at once --
2024  * sounds like badness.  It actually might be fine, as long as they're not
2025  * trying to update from the same log simultaneously, in which case we
2026  * should use a per-log semaphore instead of cld_lock.
2027  *
2028  * \param[in] mgc       MGC device by which to fetch the configuration log
2029  * \param[in] cld       log processing state (stored in lock callback data)
2030  *
2031  * \retval              0 on success
2032  * \retval              negative errno on failure
2033  */
2034 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
2035 {
2036         struct lustre_handle lockh = { 0 };
2037         __u64 flags = LDLM_FL_NO_LRU;
2038         int rc = 0, rcl;
2039         bool retry = false;
2040         ENTRY;
2041
2042         LASSERT(cld != NULL);
2043
2044         /* I don't want multiple processes running process_log at once --
2045            sounds like badness.  It actually might be fine, as long as
2046            we're not trying to update from the same log
2047            simultaneously (in which case we should use a per-log sem.) */
2048 restart:
2049         mutex_lock(&cld->cld_lock);
2050         if (cld->cld_stopping) {
2051                 mutex_unlock(&cld->cld_lock);
2052                 RETURN(0);
2053         }
2054
2055         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2056
2057         CDEBUG(D_MGC, "Process log %s-%016lx from %d\n", cld->cld_logname,
2058                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2059
2060         /* Get the cfg lock on the llog */
2061         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2062                           LCK_CR, &flags,
2063                           cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
2064                           cld, 0, NULL, &lockh);
2065         if (rcl == 0) {
2066                 /* Get the cld, it will be released in mgc_blocking_ast. */
2067                 config_log_get(cld);
2068                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2069                 LASSERT(rc == 0);
2070         } else {
2071                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2072
2073                 if (rcl == -ESHUTDOWN &&
2074                     atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2075                         struct obd_import *imp;
2076                         struct l_wait_info lwi;
2077                         int secs = cfs_time_seconds(obd_timeout);
2078
2079                         mutex_unlock(&cld->cld_lock);
2080                         imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2081
2082                         /* Let's force the pinger, and wait the import to be
2083                          * connected, note: since mgc import is non-replayable,
2084                          * and even the import state is disconnected, it does
2085                          * not mean the "recovery" is stopped, so we will keep
2086                          * waitting until timeout or the import state is
2087                          * FULL or closed */
2088                         ptlrpc_pinger_force(imp);
2089
2090                         lwi = LWI_TIMEOUT(secs, NULL, NULL);
2091                         l_wait_event(imp->imp_recovery_waitq,
2092                                      !mgc_import_in_recovery(imp), &lwi);
2093
2094                         if (imp->imp_state == LUSTRE_IMP_FULL) {
2095                                 retry = true;
2096                                 goto restart;
2097                         } else {
2098                                 mutex_lock(&cld->cld_lock);
2099                                 /* unlock/lock mutex, so check stopping again */
2100                                 if (cld->cld_stopping) {
2101                                         mutex_unlock(&cld->cld_lock);
2102                                         RETURN(0);
2103                                 }
2104                                 spin_lock(&config_list_lock);
2105                                 cld->cld_lostlock = 1;
2106                                 spin_unlock(&config_list_lock);
2107                         }
2108                 } else {
2109                         /* mark cld_lostlock so that it will requeue
2110                          * after MGC becomes available. */
2111                         spin_lock(&config_list_lock);
2112                         cld->cld_lostlock = 1;
2113                         spin_unlock(&config_list_lock);
2114                 }
2115         }
2116
2117         if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2118                 if (!rcl)
2119                         rc = mgc_process_recover_nodemap_log(mgc, cld);
2120                 else if (cld_is_nodemap(cld))
2121                         rc = rcl;
2122
2123                 if (cld_is_recover(cld) && rc) {
2124                         if (!rcl) {
2125                                 CERROR("%s: recover log %s failed, not fatal: rc = %d\n",
2126                                        mgc->obd_name, cld->cld_logname, rc);
2127                                 spin_lock(&config_list_lock);
2128                                 cld->cld_lostlock = 1;
2129                                 spin_unlock(&config_list_lock);
2130                         }
2131                         rc = 0; /* this is not a fatal error for recover log */
2132                 }
2133         } else if (!cld_is_barrier(cld)) {
2134                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2135         }
2136
2137         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2138                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2139
2140         mutex_unlock(&cld->cld_lock);
2141
2142         /* Now drop the lock so MGS can revoke it */
2143         if (!rcl) {
2144                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2145                 if (rcl)
2146                         CERROR("Can't drop cfg lock: %d\n", rcl);
2147         }
2148
2149         /* requeue nodemap lock immediately if transfer was interrupted */
2150         if (cld_is_nodemap(cld) && rc == -EAGAIN) {
2151                 mgc_requeue_add(cld);
2152                 rc = 0;
2153         }
2154
2155         RETURN(rc);
2156 }
2157
2158
2159 /** Called from lustre_process_log.
2160  * LCFG_LOG_START gets the config log from the MGS, processes it to start
2161  * any services, and adds it to the list logs to watch (follow).
2162  */
2163 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2164 {
2165         struct lustre_cfg *lcfg = buf;
2166         struct config_llog_instance *cfg = NULL;
2167         char *logname;
2168         int rc = 0;
2169         ENTRY;
2170
2171         switch(lcfg->lcfg_command) {
2172         case LCFG_LOV_ADD_OBD: {
2173                 /* Overloading this cfg command: register a new target */
2174                 struct mgs_target_info *mti;
2175
2176                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2177                     sizeof(struct mgs_target_info))
2178                         GOTO(out, rc = -EINVAL);
2179
2180                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2181                 CDEBUG(D_MGC, "add_target %s %#x\n",
2182                        mti->mti_svname, mti->mti_flags);
2183                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2184                 break;
2185         }
2186         case LCFG_LOV_DEL_OBD:
2187                 /* Unregister has no meaning at the moment. */
2188                 CERROR("lov_del_obd unimplemented\n");
2189                 rc = -ENOSYS;
2190                 break;
2191         case LCFG_SPTLRPC_CONF: {
2192                 rc = sptlrpc_process_config(lcfg);
2193                 break;
2194         }
2195         case LCFG_LOG_START: {
2196                 struct config_llog_data *cld;
2197                 struct super_block *sb;
2198
2199                 logname = lustre_cfg_string(lcfg, 1);
2200                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2201                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2202
2203                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2204                        cfg->cfg_last_idx);
2205
2206                 /* We're only called through here on the initial mount */
2207                 cld = config_log_add(obd, logname, cfg, sb);
2208                 if (IS_ERR(cld)) {
2209                         rc = PTR_ERR(cld);
2210                         break;
2211                 }
2212
2213                 rc = mgc_process_log(obd, cld);
2214                 if (rc == 0 && cld->cld_recover != NULL) {
2215                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2216                                          imp_connect_data, IMP_RECOV)) {
2217                                 rc = mgc_process_log(obd, cld->cld_recover);
2218                         } else {
2219                                 struct config_llog_data *cir;
2220
2221                                 mutex_lock(&cld->cld_lock);
2222                                 cir = cld->cld_recover;
2223                                 cld->cld_recover = NULL;
2224                                 mutex_unlock(&cld->cld_lock);
2225                                 config_log_put(cir);
2226                         }
2227
2228                         if (rc)
2229                                 CERROR("Cannot process recover llog %d\n", rc);
2230                 }
2231
2232                 if (rc == 0 && cld->cld_params != NULL) {
2233                         rc = mgc_process_log(obd, cld->cld_params);
2234                         if (rc == -ENOENT) {
2235                                 CDEBUG(D_MGC, "There is no params "
2236                                               "config file yet\n");
2237                                 rc = 0;
2238                         }
2239                         /* params log is optional */
2240                         if (rc)
2241                                 CERROR("%s: can't process params llog: rc = %d\n",
2242                                        obd->obd_name, rc);
2243                 }
2244
2245                 break;
2246         }
2247         case LCFG_LOG_END: {
2248                 logname = lustre_cfg_string(lcfg, 1);
2249
2250                 if (lcfg->lcfg_bufcount >= 2)
2251                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
2252                                 lcfg, 2);
2253                 rc = config_log_end(logname, cfg);
2254                 break;
2255         }
2256         default: {
2257                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2258                 GOTO(out, rc = -EINVAL);
2259
2260         }
2261         }
2262 out:
2263         RETURN(rc);
2264 }
2265
2266 static struct obd_ops mgc_obd_ops = {
2267         .o_owner        = THIS_MODULE,
2268         .o_setup        = mgc_setup,
2269         .o_precleanup   = mgc_precleanup,
2270         .o_cleanup      = mgc_cleanup,
2271         .o_add_conn     = client_import_add_conn,
2272         .o_del_conn     = client_import_del_conn,
2273         .o_connect      = client_connect_import,
2274         .o_disconnect   = client_disconnect_export,
2275         .o_set_info_async = mgc_set_info_async,
2276         .o_get_info       = mgc_get_info,
2277         .o_import_event = mgc_import_event,
2278         .o_process_config = mgc_process_config,
2279 };
2280
2281 static int __init mgc_init(void)
2282 {
2283         return class_register_type(&mgc_obd_ops, NULL, false, NULL,
2284                                    LUSTRE_MGC_NAME, NULL);
2285 }
2286
2287 static void __exit mgc_exit(void)
2288 {
2289         class_unregister_type(LUSTRE_MGC_NAME);
2290 }
2291
2292 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2293 MODULE_DESCRIPTION("Lustre Management Client");
2294 MODULE_VERSION(LUSTRE_VERSION_STRING);
2295 MODULE_LICENSE("GPL");
2296
2297 module_init(mgc_init);
2298 module_exit(mgc_exit);