lustre/mgc/mgc_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/mgc/mgc_request.c
  37  *
  38  * Author: Nathan Rutman <nathan@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_MGC
  42 #define D_MGC D_CONFIG /*|D_WARNING*/
  43
  44 #ifdef __KERNEL__
  45 # include <linux/module.h>
  46 # include <linux/pagemap.h>
  47 # include <linux/miscdevice.h>
  48 # include <linux/init.h>
  49 #else
  50 # include <liblustre.h>
  51 #endif
  52
  53 #include <obd_class.h>
  54 #include <lustre_dlm.h>
  55 #include <lprocfs_status.h>
  56 #include <lustre_log.h>
  57 #include <lustre_fsfilt.h>
  58 #include <lustre_disk.h>
  59 #include "mgc_internal.h"
  60
  61 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
  62                           int type)
  63 {
  64         __u64 resname = 0;
  65
  66         if (len > 8) {
  67                 CERROR("name too long: %s\n", name);
  68                 return -EINVAL;
  69         }
  70         if (len <= 0) {
  71                 CERROR("missing name: %s\n", name);
  72                 return -EINVAL;
  73         }
  74         memcpy(&resname, name, len);
  75
  76         /* Always use the same endianness for the resid */
  77         memset(res_id, 0, sizeof(*res_id));
  78         res_id->name[0] = cpu_to_le64(resname);
  79         /* XXX: unfortunately, sptlprc and config llog share one lock */
  80         switch(type) {
  81         case CONFIG_T_CONFIG:
  82         case CONFIG_T_SPTLRPC:
  83                 resname = 0;
  84                 break;
  85         case CONFIG_T_RECOVER:
  86                 resname = type;
  87                 break;
  88         default:
  89                 LBUG();
  90         }
  91         res_id->name[1] = cpu_to_le64(resname);
  92         CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
  93                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  94         return 0;
  95 }
  96
  97 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
  98 {
  99         /* fsname is at most 8 chars long, maybe contain "-".
 100          * e.g. "lustre", "SUN-000" */
 101         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
 102 }
 103 EXPORT_SYMBOL(mgc_fsname2resid);
 104
 105 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
 106 {
 107         char *name_end;
 108         int len;
 109
 110         /* logname consists of "fsname-nodetype".
 111          * e.g. "lustre-MDT0001", "SUN-000-client" */
 112         name_end = strrchr(logname, '-');
 113         LASSERT(name_end);
 114         len = name_end - logname;
 115         return mgc_name2resid(logname, len, res_id, type);
 116 }
 117
 118 /********************** config llog list **********************/
 119 static CFS_LIST_HEAD(config_llog_list);
 120 static DEFINE_SPINLOCK(config_list_lock);
 121
 122 /* Take a reference to a config log */
 123 static int config_log_get(struct config_llog_data *cld)
 124 {
 125         ENTRY;
 126         cfs_atomic_inc(&cld->cld_refcount);
 127         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 128                cfs_atomic_read(&cld->cld_refcount));
 129         RETURN(0);
 130 }
 131
 132 /* Drop a reference to a config log.  When no longer referenced,
 133    we can free the config log data */
 134 static void config_log_put(struct config_llog_data *cld)
 135 {
 136         ENTRY;
 137
 138         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 139                cfs_atomic_read(&cld->cld_refcount));
 140         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 141
 142         /* spinlock to make sure no item with 0 refcount in the list */
 143         if (cfs_atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 144                 cfs_list_del(&cld->cld_list_chain);
 145                 spin_unlock(&config_list_lock);
 146
 147                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 148
 149                 if (cld->cld_recover)
 150                         config_log_put(cld->cld_recover);
 151                 if (cld->cld_sptlrpc)
 152                         config_log_put(cld->cld_sptlrpc);
 153                 if (cld_is_sptlrpc(cld))
 154                         sptlrpc_conf_log_stop(cld->cld_logname);
 155
 156                 class_export_put(cld->cld_mgcexp);
 157                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
 158         }
 159
 160         EXIT;
 161 }
 162
 163 /* Find a config log by name */
 164 static
 165 struct config_llog_data *config_log_find(char *logname,
 166                                          struct config_llog_instance *cfg)
 167 {
 168         struct config_llog_data *cld;
 169         struct config_llog_data *found = NULL;
 170         void *                   instance;
 171         ENTRY;
 172
 173         LASSERT(logname != NULL);
 174
 175         instance = cfg ? cfg->cfg_instance : NULL;
 176         spin_lock(&config_list_lock);
 177         cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 178                 /* check if instance equals */
 179                 if (instance != cld->cld_cfg.cfg_instance)
 180                         continue;
 181
 182                 /* instance may be NULL, should check name */
 183                 if (strcmp(logname, cld->cld_logname) == 0) {
 184                         found = cld;
 185                         break;
 186                 }
 187         }
 188         if (found) {
 189                 cfs_atomic_inc(&found->cld_refcount);
 190                 LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0);
 191         }
 192         spin_unlock(&config_list_lock);
 193         RETURN(found);
 194 }
 195
 196 static
 197 struct config_llog_data *do_config_log_add(struct obd_device *obd,
 198                                            char *logname,
 199                                            int type,
 200                                            struct config_llog_instance *cfg,
 201                                            struct super_block *sb)
 202 {
 203         struct config_llog_data *cld;
 204         int                      rc;
 205         ENTRY;
 206
 207         CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
 208                cfg ? cfg->cfg_instance : 0);
 209
 210         OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
 211         if (!cld)
 212                 RETURN(ERR_PTR(-ENOMEM));
 213
 214         strcpy(cld->cld_logname, logname);
 215         if (cfg)
 216                 cld->cld_cfg = *cfg;
 217         else
 218                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 219         mutex_init(&cld->cld_lock);
 220         cld->cld_cfg.cfg_last_idx = 0;
 221         cld->cld_cfg.cfg_flags = 0;
 222         cld->cld_cfg.cfg_sb = sb;
 223         cld->cld_type = type;
 224         cfs_atomic_set(&cld->cld_refcount, 1);
 225
 226         /* Keep the mgc around until we are done */
 227         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 228
 229         if (cld_is_sptlrpc(cld)) {
 230                 sptlrpc_conf_log_start(logname);
 231                 cld->cld_cfg.cfg_obdname = obd->obd_name;
 232         }
 233
 234         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 235
 236         spin_lock(&config_list_lock);
 237         cfs_list_add(&cld->cld_list_chain, &config_llog_list);
 238         spin_unlock(&config_list_lock);
 239
 240         if (rc) {
 241                 config_log_put(cld);
 242                 RETURN(ERR_PTR(rc));
 243         }
 244
 245         if (cld_is_sptlrpc(cld)) {
 246                 rc = mgc_process_log(obd, cld);
 247                 if (rc && rc != -ENOENT)
 248                         CERROR("failed processing sptlrpc log: %d\n", rc);
 249         }
 250
 251         RETURN(cld);
 252 }
 253
 254 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
 255         char *fsname,
 256         struct config_llog_instance *cfg,
 257         struct super_block *sb)
 258 {
 259         struct config_llog_instance lcfg = *cfg;
 260         struct lustre_sb_info *lsi = s2lsi(sb);
 261         struct config_llog_data *cld;
 262         char logname[32];
 263
 264         if (IS_OST(lsi))
 265                 return NULL;
 266
 267         /* for osp-on-ost, see lustre_start_osp() */
 268         if (IS_MDT(lsi) && lcfg.cfg_instance)
 269                 return NULL;
 270
 271         /* we have to use different llog for clients and mdts for cmd
 272          * where only clients are notified if one of cmd server restarts */
 273         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 274         strcpy(logname, fsname);
 275         if (IS_SERVER(lsi)) { /* mdt */
 276                 LASSERT(lcfg.cfg_instance == NULL);
 277                 lcfg.cfg_instance = sb;
 278                 strcat(logname, "-mdtir");
 279         } else {
 280                 LASSERT(lcfg.cfg_instance != NULL);
 281                 strcat(logname, "-cliir");
 282         }
 283
 284         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
 285         return cld;
 286 }
 287
 288
 289 /** Add this log to the list of active logs watched by an MGC.
 290  * Active means we're watching for updates.
 291  * We have one active log per "mount" - client instance or servername.
 292  * Each instance may be at a different point in the log.
 293  */
 294 static int config_log_add(struct obd_device *obd, char *logname,
 295                           struct config_llog_instance *cfg,
 296                           struct super_block *sb)
 297 {
 298         struct lustre_sb_info *lsi = s2lsi(sb);
 299         struct config_llog_data *cld;
 300         struct config_llog_data *sptlrpc_cld;
 301         char                     seclogname[32];
 302         char                    *ptr;
 303         ENTRY;
 304
 305         CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
 306
 307         /*
 308          * for each regular log, the depended sptlrpc log name is
 309          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 310          */
 311         ptr = strrchr(logname, '-');
 312         if (ptr == NULL || ptr - logname > 8) {
 313                 CERROR("logname %s is too long\n", logname);
 314                 RETURN(-EINVAL);
 315         }
 316
 317         memcpy(seclogname, logname, ptr - logname);
 318         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 319
 320         sptlrpc_cld = config_log_find(seclogname, NULL);
 321         if (sptlrpc_cld == NULL) {
 322                 sptlrpc_cld = do_config_log_add(obd, seclogname,
 323                                                 CONFIG_T_SPTLRPC, NULL, NULL);
 324                 if (IS_ERR(sptlrpc_cld)) {
 325                         CERROR("can't create sptlrpc log: %s\n", seclogname);
 326                         RETURN(PTR_ERR(sptlrpc_cld));
 327                 }
 328         }
 329
 330         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 331         if (IS_ERR(cld)) {
 332                 CERROR("can't create log: %s\n", logname);
 333                 config_log_put(sptlrpc_cld);
 334                 RETURN(PTR_ERR(cld));
 335         }
 336
 337         cld->cld_sptlrpc = sptlrpc_cld;
 338
 339         LASSERT(lsi->lsi_lmd);
 340         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
 341                 struct config_llog_data *recover_cld;
 342                 *strrchr(seclogname, '-') = 0;
 343                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 344                 if (IS_ERR(recover_cld)) {
 345                         config_log_put(cld);
 346                         RETURN(PTR_ERR(recover_cld));
 347                 }
 348                 cld->cld_recover = recover_cld;
 349         }
 350
 351         RETURN(0);
 352 }
 353
 354 DEFINE_MUTEX(llog_process_lock);
 355
 356 /** Stop watching for updates on this log.
 357  */
 358 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 359 {
 360         struct config_llog_data *cld;
 361         struct config_llog_data *cld_sptlrpc = NULL;
 362         struct config_llog_data *cld_recover = NULL;
 363         int rc = 0;
 364         ENTRY;
 365
 366         cld = config_log_find(logname, cfg);
 367         if (cld == NULL)
 368                 RETURN(-ENOENT);
 369
 370         mutex_lock(&cld->cld_lock);
 371         /*
 372          * if cld_stopping is set, it means we didn't start the log thus
 373          * not owning the start ref. this can happen after previous umount:
 374          * the cld still hanging there waiting for lock cancel, and we
 375          * remount again but failed in the middle and call log_end without
 376          * calling start_log.
 377          */
 378         if (unlikely(cld->cld_stopping)) {
 379                 mutex_unlock(&cld->cld_lock);
 380                 /* drop the ref from the find */
 381                 config_log_put(cld);
 382                 RETURN(rc);
 383         }
 384
 385         cld->cld_stopping = 1;
 386
 387         cld_recover = cld->cld_recover;
 388         cld->cld_recover = NULL;
 389         mutex_unlock(&cld->cld_lock);
 390
 391         if (cld_recover) {
 392                 mutex_lock(&cld_recover->cld_lock);
 393                 cld_recover->cld_stopping = 1;
 394                 mutex_unlock(&cld_recover->cld_lock);
 395                 config_log_put(cld_recover);
 396         }
 397
 398         spin_lock(&config_list_lock);
 399         cld_sptlrpc = cld->cld_sptlrpc;
 400         cld->cld_sptlrpc = NULL;
 401         spin_unlock(&config_list_lock);
 402
 403         if (cld_sptlrpc)
 404                 config_log_put(cld_sptlrpc);
 405
 406         /* drop the ref from the find */
 407         config_log_put(cld);
 408         /* drop the start ref */
 409         config_log_put(cld);
 410
 411         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 412                rc);
 413         RETURN(rc);
 414 }
 415
 416 int lprocfs_mgc_rd_ir_state(char *page, char **start, off_t off,
 417                             int count, int *eof, void *data)
 418 {
 419         struct obd_device       *obd = data;
 420         struct obd_import       *imp = obd->u.cli.cl_import;
 421         struct obd_connect_data *ocd = &imp->imp_connect_data;
 422         struct config_llog_data *cld;
 423         int rc = 0;
 424         ENTRY;
 425
 426         rc = snprintf(page, count, "imperative_recovery: %s\n",
 427                       OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 428         rc += snprintf(page + rc, count - rc, "client_state:\n");
 429
 430         spin_lock(&config_list_lock);
 431         cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 432                 if (cld->cld_recover == NULL)
 433                         continue;
 434                 rc += snprintf(page + rc, count - rc,
 435                                "    - { client: %s, nidtbl_version: %u }\n",
 436                                cld->cld_logname,
 437                                cld->cld_recover->cld_cfg.cfg_last_idx);
 438         }
 439         spin_unlock(&config_list_lock);
 440
 441         RETURN(rc);
 442 }
 443
 444 /* reenqueue any lost locks */
 445 #define RQ_RUNNING 0x1
 446 #define RQ_NOW     0x2
 447 #define RQ_LATER   0x4
 448 #define RQ_STOP    0x8
 449 static int                    rq_state = 0;
 450 static cfs_waitq_t            rq_waitq;
 451 static DECLARE_COMPLETION(rq_exit);
 452
 453 static void do_requeue(struct config_llog_data *cld)
 454 {
 455         ENTRY;
 456         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 457
 458         /* Do not run mgc_process_log on a disconnected export or an
 459            export which is being disconnected. Take the client
 460            semaphore to make the check non-racy. */
 461         down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 462         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 463                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 464                 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 465         } else {
 466                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 467                        cld->cld_logname);
 468         }
 469         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 470
 471         EXIT;
 472 }
 473
 474 /* this timeout represents how many seconds MGC should wait before
 475  * requeue config and recover lock to the MGS. We need to randomize this
 476  * in order to not flood the MGS.
 477  */
 478 #define MGC_TIMEOUT_MIN_SECONDS   5
 479 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
 480
 481 static int mgc_requeue_thread(void *data)
 482 {
 483         char name[] = "ll_cfg_requeue";
 484         int rc = 0;
 485         ENTRY;
 486
 487         cfs_daemonize(name);
 488
 489         CDEBUG(D_MGC, "Starting requeue thread\n");
 490
 491         /* Keep trying failed locks periodically */
 492         spin_lock(&config_list_lock);
 493         rq_state |= RQ_RUNNING;
 494         while (1) {
 495                 struct l_wait_info lwi;
 496                 struct config_llog_data *cld, *cld_prev;
 497                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
 498                 int stopped = !!(rq_state & RQ_STOP);
 499                 int to;
 500
 501                 /* Any new or requeued lostlocks will change the state */
 502                 rq_state &= ~(RQ_NOW | RQ_LATER);
 503                 spin_unlock(&config_list_lock);
 504
 505                 /* Always wait a few seconds to allow the server who
 506                    caused the lock revocation to finish its setup, plus some
 507                    random so everyone doesn't try to reconnect at once. */
 508                 to = MGC_TIMEOUT_MIN_SECONDS * CFS_HZ;
 509                 to += rand * CFS_HZ / 100; /* rand is centi-seconds */
 510                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 511                 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi);
 512
 513                 /*
 514                  * iterate & processing through the list. for each cld, process
 515                  * its depending sptlrpc cld firstly (if any) and then itself.
 516                  *
 517                  * it's guaranteed any item in the list must have
 518                  * reference > 0; and if cld_lostlock is set, at
 519                  * least one reference is taken by the previous enqueue.
 520                  */
 521                 cld_prev = NULL;
 522
 523                 spin_lock(&config_list_lock);
 524                 cfs_list_for_each_entry(cld, &config_llog_list,
 525                                         cld_list_chain) {
 526                         if (!cld->cld_lostlock)
 527                                 continue;
 528
 529                         spin_unlock(&config_list_lock);
 530
 531                         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 532
 533                         /* Whether we enqueued again or not in mgc_process_log,
 534                          * we're done with the ref from the old enqueue */
 535                         if (cld_prev)
 536                                 config_log_put(cld_prev);
 537                         cld_prev = cld;
 538
 539                         cld->cld_lostlock = 0;
 540                         if (likely(!stopped))
 541                                 do_requeue(cld);
 542
 543                         spin_lock(&config_list_lock);
 544                 }
 545                 spin_unlock(&config_list_lock);
 546                 if (cld_prev)
 547                         config_log_put(cld_prev);
 548
 549                 /* break after scanning the list so that we can drop
 550                  * refcount to losing lock clds */
 551                 if (unlikely(stopped)) {
 552                         spin_lock(&config_list_lock);
 553                         break;
 554                 }
 555
 556                 /* Wait a bit to see if anyone else needs a requeue */
 557                 lwi = (struct l_wait_info) { 0 };
 558                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 559                              &lwi);
 560                 spin_lock(&config_list_lock);
 561         }
 562         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 563         rq_state &= ~RQ_RUNNING;
 564         spin_unlock(&config_list_lock);
 565
 566         complete(&rq_exit);
 567
 568         CDEBUG(D_MGC, "Ending requeue thread\n");
 569         RETURN(rc);
 570 }
 571
 572 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 573    We are responsible for dropping the config log reference from here on out. */
 574 static void mgc_requeue_add(struct config_llog_data *cld)
 575 {
 576         ENTRY;
 577
 578         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 579                cld->cld_logname, cfs_atomic_read(&cld->cld_refcount),
 580                cld->cld_stopping, rq_state);
 581         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 582
 583         mutex_lock(&cld->cld_lock);
 584         if (cld->cld_stopping || cld->cld_lostlock) {
 585                 mutex_unlock(&cld->cld_lock);
 586                 RETURN_EXIT;
 587         }
 588         /* this refcount will be released in mgc_requeue_thread. */
 589         config_log_get(cld);
 590         cld->cld_lostlock = 1;
 591         mutex_unlock(&cld->cld_lock);
 592
 593         /* Hold lock for rq_state */
 594         spin_lock(&config_list_lock);
 595         if (rq_state & RQ_STOP) {
 596                 spin_unlock(&config_list_lock);
 597                 cld->cld_lostlock = 0;
 598                 config_log_put(cld);
 599         } else {
 600                 rq_state |= RQ_NOW;
 601                 spin_unlock(&config_list_lock);
 602                 cfs_waitq_signal(&rq_waitq);
 603         }
 604         EXIT;
 605 }
 606
 607 /********************** class fns **********************/
 608
 609 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
 610                         struct vfsmount *mnt)
 611 {
 612         struct lvfs_run_ctxt saved;
 613         struct lustre_sb_info *lsi = s2lsi(sb);
 614         struct client_obd *cli = &obd->u.cli;
 615         struct dentry *dentry;
 616         char *label;
 617         int err = 0;
 618         ENTRY;
 619
 620         LASSERT(lsi);
 621         LASSERT(lsi->lsi_srv_mnt == mnt);
 622
 623         /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
 624         down(&cli->cl_mgc_sem);
 625
 626         cfs_cleanup_group_info();
 627
 628         obd->obd_fsops = fsfilt_get_ops(lsi->lsi_fstype);
 629         if (IS_ERR(obd->obd_fsops)) {
 630                 up(&cli->cl_mgc_sem);
 631                 CERROR("%s: No fstype %s: rc = %ld\n", lsi->lsi_fstype,
 632                        obd->obd_name, PTR_ERR(obd->obd_fsops));
 633                 RETURN(PTR_ERR(obd->obd_fsops));
 634         }
 635
 636         cli->cl_mgc_vfsmnt = mnt;
 637         err = fsfilt_setup(obd, mnt->mnt_sb);
 638         if (err)
 639                 GOTO(err_ops, err);
 640
 641         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 642         obd->obd_lvfs_ctxt.pwdmnt = mnt;
 643         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
 644         obd->obd_lvfs_ctxt.fs = get_ds();
 645
 646         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 647         dentry = ll_lookup_one_len(MOUNT_CONFIGS_DIR, cfs_fs_pwd(current->fs),
 648                                    strlen(MOUNT_CONFIGS_DIR));
 649         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 650         if (IS_ERR(dentry)) {
 651                 err = PTR_ERR(dentry);
 652                 CERROR("cannot lookup %s directory: rc = %d\n",
 653                        MOUNT_CONFIGS_DIR, err);
 654                 GOTO(err_ops, err);
 655         }
 656         cli->cl_mgc_configs_dir = dentry;
 657
 658         /* We take an obd ref to insure that we can't get to mgc_cleanup
 659            without calling mgc_fs_cleanup first. */
 660         class_incref(obd, "mgc_fs", obd);
 661
 662         label = fsfilt_get_label(obd, mnt->mnt_sb);
 663         if (label)
 664                 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
 665
 666         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 667         RETURN(0);
 668
 669 err_ops:
 670         fsfilt_put_ops(obd->obd_fsops);
 671         obd->obd_fsops = NULL;
 672         cli->cl_mgc_vfsmnt = NULL;
 673         up(&cli->cl_mgc_sem);
 674         RETURN(err);
 675 }
 676
 677 static int mgc_fs_cleanup(struct obd_device *obd)
 678 {
 679         struct client_obd *cli = &obd->u.cli;
 680         int rc = 0;
 681         ENTRY;
 682
 683         LASSERT(cli->cl_mgc_vfsmnt != NULL);
 684
 685         if (cli->cl_mgc_configs_dir != NULL) {
 686                 struct lvfs_run_ctxt saved;
 687                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 688                 l_dput(cli->cl_mgc_configs_dir);
 689                 cli->cl_mgc_configs_dir = NULL;
 690                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 691                 class_decref(obd, "mgc_fs", obd);
 692         }
 693
 694         cli->cl_mgc_vfsmnt = NULL;
 695         if (obd->obd_fsops)
 696                 fsfilt_put_ops(obd->obd_fsops);
 697
 698         up(&cli->cl_mgc_sem);
 699
 700         RETURN(rc);
 701 }
 702
 703 static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0);
 704 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 705 {
 706         int rc = 0;
 707         ENTRY;
 708
 709         switch (stage) {
 710         case OBD_CLEANUP_EARLY:
 711                 break;
 712         case OBD_CLEANUP_EXPORTS:
 713                 if (cfs_atomic_dec_and_test(&mgc_count)) {
 714                         int running;
 715                         /* stop requeue thread */
 716                         spin_lock(&config_list_lock);
 717                         running = rq_state & RQ_RUNNING;
 718                         if (running)
 719                                 rq_state |= RQ_STOP;
 720                         spin_unlock(&config_list_lock);
 721                         if (running) {
 722                                 cfs_waitq_signal(&rq_waitq);
 723                                 wait_for_completion(&rq_exit);
 724                         }
 725                 }
 726                 obd_cleanup_client_import(obd);
 727                 rc = obd_llog_finish(obd, 0);
 728                 if (rc != 0)
 729                         CERROR("failed to cleanup llogging subsystems\n");
 730                 break;
 731         }
 732         RETURN(rc);
 733 }
 734
 735 static int mgc_cleanup(struct obd_device *obd)
 736 {
 737         struct client_obd *cli = &obd->u.cli;
 738         int rc;
 739         ENTRY;
 740
 741         LASSERT(cli->cl_mgc_vfsmnt == NULL);
 742
 743         /* COMPAT_146 - old config logs may have added profiles we don't
 744            know about */
 745         if (obd->obd_type->typ_refcnt <= 1)
 746                 /* Only for the last mgc */
 747                 class_del_profiles();
 748
 749         lprocfs_obd_cleanup(obd);
 750         ptlrpcd_decref();
 751
 752         rc = client_obd_cleanup(obd);
 753         RETURN(rc);
 754 }
 755
 756 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 757 {
 758         struct lprocfs_static_vars lvars;
 759         int rc;
 760         ENTRY;
 761
 762         ptlrpcd_addref();
 763
 764         rc = client_obd_setup(obd, lcfg);
 765         if (rc)
 766                 GOTO(err_decref, rc);
 767
 768         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
 769         if (rc) {
 770                 CERROR("failed to setup llogging subsystems\n");
 771                 GOTO(err_cleanup, rc);
 772         }
 773
 774         lprocfs_mgc_init_vars(&lvars);
 775         lprocfs_obd_setup(obd, lvars.obd_vars);
 776         sptlrpc_lprocfs_cliobd_attach(obd);
 777
 778         if (cfs_atomic_inc_return(&mgc_count) == 1) {
 779                 rq_state = 0;
 780                 cfs_waitq_init(&rq_waitq);
 781
 782                 /* start requeue thread */
 783                 rc = cfs_create_thread(mgc_requeue_thread, NULL,
 784                                        CFS_DAEMON_FLAGS);
 785                 if (rc < 0) {
 786                         CERROR("%s: Cannot start requeue thread (%d),"
 787                                "no more log updates!\n",
 788                                obd->obd_name, rc);
 789                         GOTO(err_cleanup, rc);
 790                 }
 791                 /* rc is the pid of mgc_requeue_thread. */
 792                 rc = 0;
 793         }
 794
 795         RETURN(rc);
 796
 797 err_cleanup:
 798         client_obd_cleanup(obd);
 799 err_decref:
 800         ptlrpcd_decref();
 801         RETURN(rc);
 802 }
 803
 804 /* based on ll_mdc_blocking_ast */
 805 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 806                             void *data, int flag)
 807 {
 808         struct lustre_handle lockh;
 809         struct config_llog_data *cld = (struct config_llog_data *)data;
 810         int rc = 0;
 811         ENTRY;
 812
 813         switch (flag) {
 814         case LDLM_CB_BLOCKING:
 815                 /* mgs wants the lock, give it up... */
 816                 LDLM_DEBUG(lock, "MGC blocking CB");
 817                 ldlm_lock2handle(lock, &lockh);
 818                 rc = ldlm_cli_cancel(&lockh);
 819                 break;
 820         case LDLM_CB_CANCELING:
 821                 /* We've given up the lock, prepare ourselves to update. */
 822                 LDLM_DEBUG(lock, "MGC cancel CB");
 823
 824                 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
 825                        lock->l_resource->lr_name.name[0],
 826                        (char *)&lock->l_resource->lr_name.name[0]);
 827
 828                 if (!cld) {
 829                         CDEBUG(D_INFO, "missing data, won't requeue\n");
 830                         break;
 831                 }
 832
 833                 /* held at mgc_process_log(). */
 834                 LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 835                 /* Are we done with this log? */
 836                 if (cld->cld_stopping) {
 837                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
 838                                cld->cld_logname);
 839                         config_log_put(cld);
 840                         break;
 841                 }
 842                 /* Make sure not to re-enqueue when the mgc is stopping
 843                    (we get called from client_disconnect_export) */
 844                 if (!lock->l_conn_export ||
 845                     !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
 846                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
 847                                cld->cld_logname);
 848                         config_log_put(cld);
 849                         break;
 850                 }
 851
 852                 /* Re-enqueue now */
 853                 mgc_requeue_add(cld);
 854                 config_log_put(cld);
 855                 break;
 856         default:
 857                 LBUG();
 858         }
 859
 860         RETURN(rc);
 861 }
 862
 863 /* Not sure where this should go... */
 864 #define  MGC_ENQUEUE_LIMIT 50
 865 #define  MGC_TARGET_REG_LIMIT 10
 866 #define  MGC_SEND_PARAM_LIMIT 10
 867
 868 /* Send parameter to MGS*/
 869 static int mgc_set_mgs_param(struct obd_export *exp,
 870                              struct mgs_send_param *msp)
 871 {
 872         struct ptlrpc_request *req;
 873         struct mgs_send_param *req_msp, *rep_msp;
 874         int rc;
 875         ENTRY;
 876
 877         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 878                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
 879                                         MGS_SET_INFO);
 880         if (!req)
 881                 RETURN(-ENOMEM);
 882
 883         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 884         if (!req_msp) {
 885                 ptlrpc_req_finished(req);
 886                 RETURN(-ENOMEM);
 887         }
 888
 889         memcpy(req_msp, msp, sizeof(*req_msp));
 890         ptlrpc_request_set_replen(req);
 891
 892         /* Limit how long we will wait for the enqueue to complete */
 893         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
 894         rc = ptlrpc_queue_wait(req);
 895         if (!rc) {
 896                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 897                 memcpy(msp, rep_msp, sizeof(*rep_msp));
 898         }
 899
 900         ptlrpc_req_finished(req);
 901
 902         RETURN(rc);
 903 }
 904
 905 /* Take a config lock so we can get cancel notifications */
 906 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
 907                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
 908                        __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
 909                        void *data, __u32 lvb_len, void *lvb_swabber,
 910                        struct lustre_handle *lockh)
 911 {
 912         struct config_llog_data *cld = (struct config_llog_data *)data;
 913         struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
 914                          ldlm_completion_ast, NULL, NULL, NULL };
 915         struct ptlrpc_request *req;
 916         int short_limit = cld_is_sptlrpc(cld);
 917         int rc;
 918         ENTRY;
 919
 920         CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
 921                cld->cld_resid.name[0]);
 922
 923         /* We need a callback for every lockholder, so don't try to
 924            ldlm_lock_match (see rev 1.1.2.11.2.47) */
 925         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 926                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
 927                                         LDLM_ENQUEUE);
 928         if (req == NULL)
 929                 RETURN(-ENOMEM);
 930         ptlrpc_request_set_replen(req);
 931
 932         /* check if this is server or client */
 933         if (cld->cld_cfg.cfg_sb) {
 934                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
 935                 if (lsi && IS_SERVER(lsi))
 936                         short_limit = 1;
 937         }
 938         /* Limit how long we will wait for the enqueue to complete */
 939         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
 940         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
 941                               NULL, 0, lockh, 0);
 942         /* A failed enqueue should still call the mgc_blocking_ast,
 943            where it will be requeued if needed ("grant failed"). */
 944         ptlrpc_req_finished(req);
 945         RETURN(rc);
 946 }
 947
 948 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
 949                       __u32 mode, struct lustre_handle *lockh)
 950 {
 951         ENTRY;
 952
 953         ldlm_lock_decref(lockh, mode);
 954
 955         RETURN(0);
 956 }
 957
 958 static void mgc_notify_active(struct obd_device *unused)
 959 {
 960         /* wakeup mgc_requeue_thread to requeue mgc lock */
 961         spin_lock(&config_list_lock);
 962         rq_state |= RQ_NOW;
 963         spin_unlock(&config_list_lock);
 964         cfs_waitq_signal(&rq_waitq);
 965
 966         /* TODO: Help the MGS rebuild nidtbl. -jay */
 967 }
 968
 969 /* Send target_reg message to MGS */
 970 static int mgc_target_register(struct obd_export *exp,
 971                                struct mgs_target_info *mti)
 972 {
 973         struct ptlrpc_request  *req;
 974         struct mgs_target_info *req_mti, *rep_mti;
 975         int                     rc;
 976         ENTRY;
 977
 978         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 979                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
 980                                         MGS_TARGET_REG);
 981         if (req == NULL)
 982                 RETURN(-ENOMEM);
 983
 984         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
 985         if (!req_mti) {
 986                 ptlrpc_req_finished(req);
 987                 RETURN(-ENOMEM);
 988         }
 989
 990         memcpy(req_mti, mti, sizeof(*req_mti));
 991         ptlrpc_request_set_replen(req);
 992         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
 993         /* Limit how long we will wait for the enqueue to complete */
 994         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
 995
 996         rc = ptlrpc_queue_wait(req);
 997         if (!rc) {
 998                 rep_mti = req_capsule_server_get(&req->rq_pill,
 999                                                  &RMF_MGS_TARGET_INFO);
1000                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1001                 CDEBUG(D_MGC, "register %s got index = %d\n",
1002                        mti->mti_svname, mti->mti_stripe_index);
1003         }
1004         ptlrpc_req_finished(req);
1005
1006         RETURN(rc);
1007 }
1008
1009 int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1010                        obd_count keylen, void *key, obd_count vallen,
1011                        void *val, struct ptlrpc_request_set *set)
1012 {
1013         int rc = -EINVAL;
1014         ENTRY;
1015
1016         /* Turn off initial_recov after we try all backup servers once */
1017         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1018                 struct obd_import *imp = class_exp2cliimp(exp);
1019                 int value;
1020                 if (vallen != sizeof(int))
1021                         RETURN(-EINVAL);
1022                 value = *(int *)val;
1023                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1024                        imp->imp_obd->obd_name, value,
1025                        imp->imp_deactive, imp->imp_invalid,
1026                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1027                        ptlrpc_import_state_name(imp->imp_state));
1028                 /* Resurrect if we previously died */
1029                 if ((imp->imp_state != LUSTRE_IMP_FULL &&
1030                      imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
1031                         ptlrpc_reconnect_import(imp);
1032                 RETURN(0);
1033         }
1034         /* FIXME move this to mgc_process_config */
1035         if (KEY_IS(KEY_REGISTER_TARGET)) {
1036                 struct mgs_target_info *mti;
1037                 if (vallen != sizeof(struct mgs_target_info))
1038                         RETURN(-EINVAL);
1039                 mti = (struct mgs_target_info *)val;
1040                 CDEBUG(D_MGC, "register_target %s %#x\n",
1041                        mti->mti_svname, mti->mti_flags);
1042                 rc =  mgc_target_register(exp, mti);
1043                 RETURN(rc);
1044         }
1045         if (KEY_IS(KEY_SET_FS)) {
1046                 struct super_block *sb = (struct super_block *)val;
1047                 struct lustre_sb_info *lsi;
1048                 if (vallen != sizeof(struct super_block))
1049                         RETURN(-EINVAL);
1050                 lsi = s2lsi(sb);
1051                 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
1052                 if (rc) {
1053                         CERROR("set_fs got %d\n", rc);
1054                 }
1055                 RETURN(rc);
1056         }
1057         if (KEY_IS(KEY_CLEAR_FS)) {
1058                 if (vallen != 0)
1059                         RETURN(-EINVAL);
1060                 rc = mgc_fs_cleanup(exp->exp_obd);
1061                 if (rc) {
1062                         CERROR("clear_fs got %d\n", rc);
1063                 }
1064                 RETURN(rc);
1065         }
1066         if (KEY_IS(KEY_SET_INFO)) {
1067                 struct mgs_send_param *msp;
1068
1069                 msp = (struct mgs_send_param *)val;
1070                 rc =  mgc_set_mgs_param(exp, msp);
1071                 RETURN(rc);
1072         }
1073         if (KEY_IS(KEY_MGSSEC)) {
1074                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1075                 struct sptlrpc_flavor  flvr;
1076
1077                 /*
1078                  * empty string means using current flavor, if which haven't
1079                  * been set yet, set it as null.
1080                  *
1081                  * if flavor has been set previously, check the asking flavor
1082                  * must match the existing one.
1083                  */
1084                 if (vallen == 0) {
1085                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1086                                 RETURN(0);
1087                         val = "null";
1088                         vallen = 4;
1089                 }
1090
1091                 rc = sptlrpc_parse_flavor(val, &flvr);
1092                 if (rc) {
1093                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1094                                (char *) val);
1095                         RETURN(rc);
1096                 }
1097
1098                 /*
1099                  * caller already hold a mutex
1100                  */
1101                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1102                         cli->cl_flvr_mgc = flvr;
1103                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1104                                   sizeof(flvr)) != 0) {
1105                         char    str[20];
1106
1107                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1108                                             str, sizeof(str));
1109                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1110                                        "currently %s is in use\n",
1111                                        (char *) val, str);
1112                         rc = -EPERM;
1113                 }
1114                 RETURN(rc);
1115         }
1116
1117         RETURN(rc);
1118 }
1119
1120 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1121                         __u32 keylen, void *key, __u32 *vallen, void *val,
1122                         struct lov_stripe_md *unused)
1123 {
1124         int rc = -EINVAL;
1125
1126         if (KEY_IS(KEY_CONN_DATA)) {
1127                 struct obd_import *imp = class_exp2cliimp(exp);
1128                 struct obd_connect_data *data = val;
1129
1130                 if (*vallen == sizeof(*data)) {
1131                         *data = imp->imp_connect_data;
1132                         rc = 0;
1133                 }
1134         }
1135
1136         return rc;
1137 }
1138
1139 static int mgc_import_event(struct obd_device *obd,
1140                             struct obd_import *imp,
1141                             enum obd_import_event event)
1142 {
1143         int rc = 0;
1144
1145         LASSERT(imp->imp_obd == obd);
1146         CDEBUG(D_MGC, "import event %#x\n", event);
1147
1148         switch (event) {
1149         case IMP_EVENT_DISCON:
1150                 /* MGC imports should not wait for recovery */
1151                 break;
1152         case IMP_EVENT_INACTIVE:
1153                 break;
1154         case IMP_EVENT_INVALIDATE: {
1155                 struct ldlm_namespace *ns = obd->obd_namespace;
1156                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1157                 break;
1158         }
1159         case IMP_EVENT_ACTIVE:
1160                 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
1161                 /* Clearing obd_no_recov allows us to continue pinging */
1162                 obd->obd_no_recov = 0;
1163                 mgc_notify_active(obd);
1164                 break;
1165         case IMP_EVENT_OCD:
1166                 break;
1167         case IMP_EVENT_DEACTIVATE:
1168         case IMP_EVENT_ACTIVATE:
1169                 break;
1170         default:
1171                 CERROR("Unknown import event %#x\n", event);
1172                 LBUG();
1173         }
1174         RETURN(rc);
1175 }
1176
1177 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1178                          struct obd_device *tgt, int *index)
1179 {
1180         struct llog_ctxt *ctxt;
1181         int rc;
1182         ENTRY;
1183
1184         LASSERT(olg == &obd->obd_olg);
1185
1186 #ifdef HAVE_LDISKFS_OSD
1187         rc = llog_setup(NULL, obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt,
1188                         &llog_lvfs_ops);
1189         if (rc)
1190                 RETURN(rc);
1191 #endif
1192
1193         rc = llog_setup(NULL, obd, olg, LLOG_CONFIG_REPL_CTXT, tgt,
1194                         &llog_client_ops);
1195         if (rc)
1196                 GOTO(out, rc);
1197
1198         ctxt = llog_group_get_ctxt(olg, LLOG_CONFIG_REPL_CTXT);
1199         if (!ctxt)
1200                 GOTO(out, rc = -ENODEV);
1201
1202         llog_initiator_connect(ctxt);
1203         llog_ctxt_put(ctxt);
1204
1205         RETURN(0);
1206 out:
1207         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1208         if (ctxt)
1209                 llog_cleanup(NULL, ctxt);
1210         RETURN(rc);
1211 }
1212
1213 static int mgc_llog_finish(struct obd_device *obd, int count)
1214 {
1215         struct llog_ctxt *ctxt;
1216
1217         ENTRY;
1218
1219         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1220         if (ctxt)
1221                 llog_cleanup(NULL, ctxt);
1222
1223         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1224         if (ctxt)
1225                 llog_cleanup(NULL, ctxt);
1226         RETURN(0);
1227 }
1228
1229 enum {
1230         CONFIG_READ_NRPAGES_INIT = 1 << (20 - CFS_PAGE_SHIFT),
1231         CONFIG_READ_NRPAGES      = 4
1232 };
1233
1234 static int mgc_apply_recover_logs(struct obd_device *mgc,
1235                                   struct config_llog_data *cld,
1236                                   __u64 max_version,
1237                                   void *data, int datalen, bool mne_swab)
1238 {
1239         struct config_llog_instance *cfg = &cld->cld_cfg;
1240         struct lustre_sb_info       *lsi = s2lsi(cfg->cfg_sb);
1241         struct mgs_nidtbl_entry *entry;
1242         struct lustre_cfg       *lcfg;
1243         struct lustre_cfg_bufs   bufs;
1244         u64   prev_version = 0;
1245         char *inst;
1246         char *buf;
1247         int   bufsz;
1248         int   pos;
1249         int   rc  = 0;
1250         int   off = 0;
1251         ENTRY;
1252
1253         LASSERT(cfg->cfg_instance != NULL);
1254         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1255
1256         OBD_ALLOC(inst, CFS_PAGE_SIZE);
1257         if (inst == NULL)
1258                 RETURN(-ENOMEM);
1259
1260         if (!IS_SERVER(lsi)) {
1261                 pos = sprintf(inst, "%p", cfg->cfg_instance);
1262         } else {
1263                 LASSERT(IS_MDT(lsi));
1264                 rc = server_name2svname(lsi->lsi_svname, inst, NULL);
1265                 if (rc)
1266                         RETURN(-EINVAL);
1267                 pos = strlen(inst);
1268         }
1269
1270         ++pos;
1271         buf   = inst + pos;
1272         bufsz = CFS_PAGE_SIZE - pos;
1273
1274         while (datalen > 0) {
1275                 int   entry_len = sizeof(*entry);
1276                 int   is_ost;
1277                 struct obd_device *obd;
1278                 char *obdname;
1279                 char *cname;
1280                 char *params;
1281                 char *uuid;
1282
1283                 rc = -EINVAL;
1284                 if (datalen < sizeof(*entry))
1285                         break;
1286
1287                 entry = (typeof(entry))(data + off);
1288
1289                 /* sanity check */
1290                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1291                         break;
1292                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1293                         break;
1294                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1295                         break;
1296
1297                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1298                 if (datalen < entry_len) /* must have entry_len at least */
1299                         break;
1300
1301                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1302                 if (mne_swab)
1303                         lustre_swab_mgs_nidtbl_entry(entry);
1304                 if (entry->mne_length > CFS_PAGE_SIZE) {
1305                         CERROR("MNE too large (%u)\n", entry->mne_length);
1306                         break;
1307                 }
1308
1309                 if (entry->mne_length < entry_len)
1310                         break;
1311
1312                 off     += entry->mne_length;
1313                 datalen -= entry->mne_length;
1314                 if (datalen < 0)
1315                         break;
1316
1317                 if (entry->mne_version > max_version) {
1318                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1319                                entry->mne_version, max_version);
1320                         break;
1321                 }
1322
1323                 if (prev_version >= entry->mne_version) {
1324                         CERROR("index unsorted, prev %lld, now %lld\n",
1325                                prev_version, entry->mne_version);
1326                         break;
1327                 }
1328                 prev_version = entry->mne_version;
1329
1330                 /*
1331                  * Write a string with format "nid::instance" to
1332                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1333                  */
1334
1335                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1336                 memset(buf, 0, bufsz);
1337                 obdname = buf;
1338                 pos = 0;
1339
1340                 /* lustre-OST0001-osc-<instance #> */
1341                 strcpy(obdname, cld->cld_logname);
1342                 cname = strrchr(obdname, '-');
1343                 if (cname == NULL) {
1344                         CERROR("mgc %s: invalid logname %s\n",
1345                                mgc->obd_name, obdname);
1346                         break;
1347                 }
1348
1349                 pos = cname - obdname;
1350                 obdname[pos] = 0;
1351                 pos += sprintf(obdname + pos, "-%s%04x",
1352                                   is_ost ? "OST" : "MDT", entry->mne_index);
1353
1354                 cname = is_ost ? "osc" : "mdc",
1355                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1356                 lustre_cfg_bufs_reset(&bufs, obdname);
1357
1358                 /* find the obd by obdname */
1359                 obd = class_name2obd(obdname);
1360                 if (obd == NULL) {
1361                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1362                                mgc->obd_name, obdname);
1363                         rc = 0;
1364                         /* this is a safe race, when the ost is starting up...*/
1365                         continue;
1366                 }
1367
1368                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1369                 ++pos;
1370                 params = buf + pos;
1371                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1372                 uuid = buf + pos;
1373
1374                 down_read(&obd->u.cli.cl_sem);
1375                 if (obd->u.cli.cl_import == NULL) {
1376                         /* client does not connect to the OST yet */
1377                         up_read(&obd->u.cli.cl_sem);
1378                         rc = 0;
1379                         continue;
1380                 }
1381
1382                 /* TODO: iterate all nids to find one */
1383                 /* find uuid by nid */
1384                 rc = client_import_find_conn(obd->u.cli.cl_import,
1385                                              entry->u.nids[0],
1386                                              (struct obd_uuid *)uuid);
1387                 up_read(&obd->u.cli.cl_sem);
1388                 if (rc < 0) {
1389                         CERROR("mgc: cannot find uuid by nid %s\n",
1390                                libcfs_nid2str(entry->u.nids[0]));
1391                         break;
1392                 }
1393
1394                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1395                        uuid, libcfs_nid2str(entry->u.nids[0]));
1396
1397                 pos += strlen(uuid);
1398                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1399                 LASSERT(pos < bufsz);
1400
1401                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1402
1403                 rc = -ENOMEM;
1404                 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1405                 if (lcfg == NULL) {
1406                         CERROR("mgc: cannot allocate memory\n");
1407                         break;
1408                 }
1409
1410                 CDEBUG(D_INFO, "ir apply logs "LPD64"/"LPD64" for %s -> %s\n",
1411                        prev_version, max_version, obdname, params);
1412
1413                 rc = class_process_config(lcfg);
1414                 lustre_cfg_free(lcfg);
1415                 if (rc)
1416                         CDEBUG(D_INFO, "process config for %s error %d\n",
1417                                obdname, rc);
1418
1419                 /* continue, even one with error */
1420         }
1421
1422         OBD_FREE(inst, CFS_PAGE_SIZE);
1423         RETURN(rc);
1424 }
1425
1426 /**
1427  * This function is called if this client was notified for target restarting
1428  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1429  */
1430 static int mgc_process_recover_log(struct obd_device *obd,
1431                                    struct config_llog_data *cld)
1432 {
1433         struct ptlrpc_request *req = NULL;
1434         struct config_llog_instance *cfg = &cld->cld_cfg;
1435         struct mgs_config_body *body;
1436         struct mgs_config_res  *res;
1437         struct ptlrpc_bulk_desc *desc;
1438         cfs_page_t **pages;
1439         int nrpages;
1440         bool eof = true;
1441         bool mne_swab = false;
1442         int i;
1443         int ealen;
1444         int rc;
1445         ENTRY;
1446
1447         /* allocate buffer for bulk transfer.
1448          * if this is the first time for this mgs to read logs,
1449          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1450          * once; otherwise, it only reads increment of logs, this should be
1451          * small and CONFIG_READ_NRPAGES will be used.
1452          */
1453         nrpages = CONFIG_READ_NRPAGES;
1454         if (cfg->cfg_last_idx == 0) /* the first time */
1455                 nrpages = CONFIG_READ_NRPAGES_INIT;
1456
1457         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1458         if (pages == NULL)
1459                 GOTO(out, rc = -ENOMEM);
1460
1461         for (i = 0; i < nrpages; i++) {
1462                 pages[i] = cfs_alloc_page(CFS_ALLOC_STD);
1463                 if (pages[i] == NULL)
1464                         GOTO(out, rc = -ENOMEM);
1465         }
1466
1467 again:
1468         LASSERT(cld_is_recover(cld));
1469         LASSERT(mutex_is_locked(&cld->cld_lock));
1470         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1471                                    &RQF_MGS_CONFIG_READ);
1472         if (req == NULL)
1473                 GOTO(out, rc = -ENOMEM);
1474
1475         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1476         if (rc)
1477                 GOTO(out, rc);
1478
1479         /* pack request */
1480         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1481         LASSERT(body != NULL);
1482         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1483         strncpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name));
1484         body->mcb_offset = cfg->cfg_last_idx + 1;
1485         body->mcb_type   = cld->cld_type;
1486         body->mcb_bits   = CFS_PAGE_SHIFT;
1487         body->mcb_units  = nrpages;
1488
1489         /* allocate bulk transfer descriptor */
1490         desc = ptlrpc_prep_bulk_imp(req, nrpages, BULK_PUT_SINK,
1491                                     MGS_BULK_PORTAL);
1492         if (desc == NULL)
1493                 GOTO(out, rc = -ENOMEM);
1494
1495         for (i = 0; i < nrpages; i++)
1496                 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, CFS_PAGE_SIZE);
1497
1498         ptlrpc_request_set_replen(req);
1499         rc = ptlrpc_queue_wait(req);
1500         if (rc)
1501                 GOTO(out, rc);
1502
1503         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1504         if (res->mcr_size < res->mcr_offset)
1505                 GOTO(out, rc = -EINVAL);
1506
1507         /* always update the index even though it might have errors with
1508          * handling the recover logs */
1509         cfg->cfg_last_idx = res->mcr_offset;
1510         eof = res->mcr_offset == res->mcr_size;
1511
1512         CDEBUG(D_INFO, "Latest version "LPD64", more %d.\n",
1513                res->mcr_offset, eof == false);
1514
1515         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1516         if (ealen < 0)
1517                 GOTO(out, rc = ealen);
1518
1519         if (ealen > nrpages << CFS_PAGE_SHIFT)
1520                 GOTO(out, rc = -EINVAL);
1521
1522         if (ealen == 0) { /* no logs transferred */
1523                 if (!eof)
1524                         rc = -EINVAL;
1525                 GOTO(out, rc);
1526         }
1527
1528         mne_swab = !!ptlrpc_rep_need_swab(req);
1529 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1530         /* This import flag means the server did an extra swab of IR MNE
1531          * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1532         if (unlikely(req->rq_import->imp_need_mne_swab))
1533                 mne_swab = !mne_swab;
1534 #else
1535 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1536 #endif
1537
1538         for (i = 0; i < nrpages && ealen > 0; i++) {
1539                 int rc2;
1540                 void *ptr;
1541
1542                 ptr = cfs_kmap(pages[i]);
1543                 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1544                                              min_t(int, ealen, CFS_PAGE_SIZE),
1545                                              mne_swab);
1546                 cfs_kunmap(pages[i]);
1547                 if (rc2 < 0) {
1548                         CWARN("Process recover log %s error %d\n",
1549                               cld->cld_logname, rc2);
1550                         break;
1551                 }
1552
1553                 ealen -= CFS_PAGE_SIZE;
1554         }
1555
1556 out:
1557         if (req)
1558                 ptlrpc_req_finished(req);
1559
1560         if (rc == 0 && !eof)
1561                 goto again;
1562
1563         if (pages) {
1564                 for (i = 0; i < nrpages; i++) {
1565                         if (pages[i] == NULL)
1566                                 break;
1567                         cfs_free_page(pages[i]);
1568                 }
1569                 OBD_FREE(pages, sizeof(*pages) * nrpages);
1570         }
1571         return rc;
1572 }
1573
1574 #ifdef HAVE_LDISKFS_OSD
1575
1576 /*
1577  * XXX: mgc_copy_llog() does not support osd-based llogs yet
1578  */
1579
1580 /* identical to mgs_log_is_empty */
1581 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1582                              char *name)
1583 {
1584         struct lvfs_run_ctxt     saved;
1585         struct llog_handle      *llh;
1586         int                      rc = 0;
1587
1588         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1589         rc = llog_open(NULL, ctxt, &llh, NULL, name, LLOG_OPEN_EXISTS);
1590         if (rc == 0) {
1591                 llog_init_handle(NULL, llh, LLOG_F_IS_PLAIN, NULL);
1592                 rc = llog_get_size(llh);
1593                 llog_close(NULL, llh);
1594         } else if (rc == -ENOENT) {
1595                 rc = 0;
1596         }
1597         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1598         /* header is record 1 */
1599         return (rc <= 1);
1600 }
1601
1602 static int mgc_copy_handler(const struct lu_env *env, struct llog_handle *llh,
1603                             struct llog_rec_hdr *rec, void *data)
1604 {
1605         struct llog_rec_hdr local_rec = *rec;
1606         struct llog_handle *local_llh = (struct llog_handle *)data;
1607         char *cfg_buf = (char*) (rec + 1);
1608         struct lustre_cfg *lcfg;
1609         int rc = 0;
1610         ENTRY;
1611
1612         /* Append all records */
1613         local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
1614         rc = llog_write(env, local_llh, &local_rec, NULL, 0,
1615                         (void *)cfg_buf, -1);
1616
1617         lcfg = (struct lustre_cfg *)cfg_buf;
1618         CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
1619                rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
1620                lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
1621
1622         RETURN(rc);
1623 }
1624
1625 /* Copy a remote log locally */
1626 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1627                          struct llog_ctxt *lctxt, char *logname)
1628 {
1629         struct llog_handle *local_llh, *remote_llh;
1630         struct obd_uuid *uuid;
1631         char *temp_log;
1632         int rc, rc2;
1633         ENTRY;
1634
1635         /* Write new log to a temp name, then vfs_rename over logname
1636            upon successful completion. */
1637
1638         OBD_ALLOC(temp_log, strlen(logname) + 1);
1639         if (!temp_log)
1640                 RETURN(-ENOMEM);
1641         sprintf(temp_log, "%sT", logname);
1642
1643         /* Make sure there's no old temp log */
1644         rc = llog_erase(NULL, lctxt, NULL, temp_log);
1645         if (rc < 0 && rc != -ENOENT)
1646                 GOTO(out, rc);
1647
1648         /* open local log */
1649         rc = llog_open_create(NULL, lctxt, &local_llh, NULL, temp_log);
1650         if (rc)
1651                 GOTO(out, rc);
1652
1653         /* set the log header uuid for fun */
1654         OBD_ALLOC_PTR(uuid);
1655         obd_str2uuid(uuid, logname);
1656         rc = llog_init_handle(NULL, local_llh, LLOG_F_IS_PLAIN, uuid);
1657         OBD_FREE_PTR(uuid);
1658         if (rc)
1659                 GOTO(out_closel, rc);
1660
1661         /* open remote log */
1662         rc = llog_open(NULL, rctxt, &remote_llh, NULL, logname,
1663                        LLOG_OPEN_EXISTS);
1664         if (rc < 0) {
1665                 if (rc == -ENOENT)
1666                         rc = 0;
1667                 GOTO(out_closel, rc);
1668         }
1669
1670         rc = llog_init_handle(NULL, remote_llh, LLOG_F_IS_PLAIN, NULL);
1671         if (rc)
1672                 GOTO(out_closer, rc);
1673
1674         /* Copy remote log */
1675         rc = llog_process(NULL, remote_llh, mgc_copy_handler,
1676                           (void *)local_llh, NULL);
1677
1678 out_closer:
1679         rc2 = llog_close(NULL, remote_llh);
1680         if (!rc)
1681                 rc = rc2;
1682 out_closel:
1683         rc2 = llog_close(NULL, local_llh);
1684         if (!rc)
1685                 rc = rc2;
1686
1687         /* We've copied the remote log to the local temp log, now
1688            replace the old local log with the temp log. */
1689         if (rc == 0) {
1690                 struct client_obd *cli = &obd->u.cli;
1691
1692                 LASSERT(cli);
1693                 LASSERT(cli->cl_mgc_configs_dir);
1694                 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1695                                    temp_log, logname);
1696         }
1697         CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1698 out:
1699         if (rc)
1700                 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1701         OBD_FREE(temp_log, strlen(logname) + 1);
1702         RETURN(rc);
1703 }
1704 #endif
1705
1706 /* local_only means it cannot get remote llogs */
1707 static int mgc_process_cfg_log(struct obd_device *mgc,
1708                                struct config_llog_data *cld,
1709                                int local_only)
1710 {
1711         struct llog_ctxt *ctxt, *lctxt = NULL;
1712 #ifdef HAVE_LDISKFS_OSD
1713         struct client_obd *cli = &mgc->u.cli;
1714 #endif
1715         struct lvfs_run_ctxt *saved_ctxt;
1716         struct lustre_sb_info *lsi = NULL;
1717         int rc = 0, must_pop = 0;
1718         bool sptlrpc_started = false;
1719
1720         ENTRY;
1721
1722         LASSERT(cld);
1723         LASSERT(mutex_is_locked(&cld->cld_lock));
1724
1725         /*
1726          * local copy of sptlrpc log is controlled elsewhere, don't try to
1727          * read it up here.
1728          */
1729         if (cld_is_sptlrpc(cld) && local_only)
1730                 RETURN(0);
1731
1732         if (cld->cld_cfg.cfg_sb)
1733                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1734
1735         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1736         if (!ctxt) {
1737                 CERROR("missing llog context\n");
1738                 RETURN(-EINVAL);
1739         }
1740
1741         OBD_ALLOC_PTR(saved_ctxt);
1742         if (saved_ctxt == NULL)
1743                 RETURN(-ENOMEM);
1744
1745         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1746
1747 #ifdef HAVE_LDISKFS_OSD
1748         /*
1749          * XXX: at the moment mgc_copy_llog() works with lvfs-based llogs
1750          */
1751         /* Copy the setup log locally if we can. Don't mess around if we're
1752            running an MGS though (logs are already local). */
1753         if (lctxt && lsi && IS_SERVER(lsi) &&
1754             (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1755             !IS_MGS(lsi) && lsi->lsi_srv_mnt) {
1756                 push_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1757                 must_pop++;
1758                 if (!local_only)
1759                         /* Only try to copy log if we have the lock. */
1760                         rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1761                 if (local_only || rc) {
1762                         if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1763                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1764                                                    "log %s and no local copy."
1765                                                    "\n", cld->cld_logname);
1766                                 GOTO(out_pop, rc = -ENOTCONN);
1767                         }
1768                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1769                                "copy for now, will try to update later.\n",
1770                                cld->cld_logname);
1771                 }
1772                 /* Now, whether we copied or not, start using the local llog.
1773                    If we failed to copy, we'll start using whatever the old
1774                    log has. */
1775                 llog_ctxt_put(ctxt);
1776                 ctxt = lctxt;
1777                 lctxt = NULL;
1778         } else
1779 #endif
1780                 if (local_only) { /* no local log at client side */
1781                 GOTO(out_pop, rc = -EIO);
1782         }
1783
1784         if (cld_is_sptlrpc(cld)) {
1785                 sptlrpc_conf_log_update_begin(cld->cld_logname);
1786                 sptlrpc_started = true;
1787         }
1788
1789         /* logname and instance info should be the same, so use our
1790            copy of the instance for the update.  The cfg_last_idx will
1791            be updated here. */
1792         rc = class_config_parse_llog(NULL, ctxt, cld->cld_logname,
1793                                      &cld->cld_cfg);
1794         EXIT;
1795
1796 out_pop:
1797         llog_ctxt_put(ctxt);
1798         if (lctxt)
1799                 llog_ctxt_put(lctxt);
1800         if (must_pop)
1801                 pop_ctxt(saved_ctxt, &mgc->obd_lvfs_ctxt, NULL);
1802
1803         OBD_FREE_PTR(saved_ctxt);
1804         /*
1805          * update settings on existing OBDs. doing it inside
1806          * of llog_process_lock so no device is attaching/detaching
1807          * in parallel.
1808          * the logname must be <fsname>-sptlrpc
1809          */
1810         if (sptlrpc_started) {
1811                 LASSERT(cld_is_sptlrpc(cld));
1812                 sptlrpc_conf_log_update_end(cld->cld_logname);
1813                 class_notify_sptlrpc_conf(cld->cld_logname,
1814                                           strlen(cld->cld_logname) -
1815                                           strlen("-sptlrpc"));
1816         }
1817
1818         RETURN(rc);
1819 }
1820
1821 /** Get a config log from the MGS and process it.
1822  * This func is called for both clients and servers.
1823  * Copy the log locally before parsing it if appropriate (non-MGS server)
1824  */
1825 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1826 {
1827         struct lustre_handle lockh = { 0 };
1828         __u64 flags = LDLM_FL_NO_LRU;
1829         int rc = 0, rcl;
1830         ENTRY;
1831
1832         LASSERT(cld);
1833
1834         /* I don't want multiple processes running process_log at once --
1835            sounds like badness.  It actually might be fine, as long as
1836            we're not trying to update from the same log
1837            simultaneously (in which case we should use a per-log sem.) */
1838         mutex_lock(&cld->cld_lock);
1839         if (cld->cld_stopping) {
1840                 mutex_unlock(&cld->cld_lock);
1841                 RETURN(0);
1842         }
1843
1844         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1845
1846         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1847                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1848
1849         /* Get the cfg lock on the llog */
1850         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1851                           LCK_CR, &flags, NULL, NULL, NULL,
1852                           cld, 0, NULL, &lockh);
1853         if (rcl == 0) {
1854                 /* Get the cld, it will be released in mgc_blocking_ast. */
1855                 config_log_get(cld);
1856                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1857                 LASSERT(rc == 0);
1858         } else {
1859                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1860
1861                 /* mark cld_lostlock so that it will requeue
1862                  * after MGC becomes available. */
1863                 cld->cld_lostlock = 1;
1864                 /* Get extra reference, it will be put in requeue thread */
1865                 config_log_get(cld);
1866         }
1867
1868
1869         if (cld_is_recover(cld)) {
1870                 rc = 0; /* this is not a fatal error for recover log */
1871                 if (rcl == 0)
1872                         rc = mgc_process_recover_log(mgc, cld);
1873         } else {
1874                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1875         }
1876
1877         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1878                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1879
1880         mutex_unlock(&cld->cld_lock);
1881
1882         /* Now drop the lock so MGS can revoke it */
1883         if (!rcl) {
1884                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1885                                  LCK_CR, &lockh);
1886                 if (rcl)
1887                         CERROR("Can't drop cfg lock: %d\n", rcl);
1888         }
1889
1890         RETURN(rc);
1891 }
1892
1893
1894 /** Called from lustre_process_log.
1895  * LCFG_LOG_START gets the config log from the MGS, processes it to start
1896  * any services, and adds it to the list logs to watch (follow).
1897  */
1898 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1899 {
1900         struct lustre_cfg *lcfg = buf;
1901         struct config_llog_instance *cfg = NULL;
1902         char *logname;
1903         int rc = 0;
1904         ENTRY;
1905
1906         switch(lcfg->lcfg_command) {
1907         case LCFG_LOV_ADD_OBD: {
1908                 /* Overloading this cfg command: register a new target */
1909                 struct mgs_target_info *mti;
1910
1911                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1912                     sizeof(struct mgs_target_info))
1913                         GOTO(out, rc = -EINVAL);
1914
1915                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1916                 CDEBUG(D_MGC, "add_target %s %#x\n",
1917                        mti->mti_svname, mti->mti_flags);
1918                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1919                 break;
1920         }
1921         case LCFG_LOV_DEL_OBD:
1922                 /* Unregister has no meaning at the moment. */
1923                 CERROR("lov_del_obd unimplemented\n");
1924                 rc = -ENOSYS;
1925                 break;
1926         case LCFG_SPTLRPC_CONF: {
1927                 rc = sptlrpc_process_config(lcfg);
1928                 break;
1929         }
1930         case LCFG_LOG_START: {
1931                 struct config_llog_data *cld;
1932                 struct super_block *sb;
1933
1934                 logname = lustre_cfg_string(lcfg, 1);
1935                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1936                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1937
1938                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1939                        cfg->cfg_last_idx);
1940
1941                 /* We're only called through here on the initial mount */
1942                 rc = config_log_add(obd, logname, cfg, sb);
1943                 if (rc)
1944                         break;
1945                 cld = config_log_find(logname, cfg);
1946                 if (cld == NULL) {
1947                         rc = -ENOENT;
1948                         break;
1949                 }
1950
1951                 /* COMPAT_146 */
1952                 /* FIXME only set this for old logs!  Right now this forces
1953                    us to always skip the "inside markers" check */
1954                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1955
1956                 rc = mgc_process_log(obd, cld);
1957                 if (rc == 0 && cld->cld_recover != NULL) {
1958                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1959                                          imp_connect_data, IMP_RECOV)) {
1960                                 rc = mgc_process_log(obd, cld->cld_recover);
1961                         } else {
1962                                 struct config_llog_data *cir = cld->cld_recover;
1963                                 cld->cld_recover = NULL;
1964                                 config_log_put(cir);
1965                         }
1966                         if (rc)
1967                                 CERROR("Cannot process recover llog %d\n", rc);
1968                 }
1969                 config_log_put(cld);
1970
1971                 break;
1972         }
1973         case LCFG_LOG_END: {
1974                 logname = lustre_cfg_string(lcfg, 1);
1975
1976                 if (lcfg->lcfg_bufcount >= 2)
1977                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
1978                                 lcfg, 2);
1979                 rc = config_log_end(logname, cfg);
1980                 break;
1981         }
1982         default: {
1983                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1984                 GOTO(out, rc = -EINVAL);
1985
1986         }
1987         }
1988 out:
1989         RETURN(rc);
1990 }
1991
1992 struct obd_ops mgc_obd_ops = {
1993         .o_owner        = THIS_MODULE,
1994         .o_setup        = mgc_setup,
1995         .o_precleanup   = mgc_precleanup,
1996         .o_cleanup      = mgc_cleanup,
1997         .o_add_conn     = client_import_add_conn,
1998         .o_del_conn     = client_import_del_conn,
1999         .o_connect      = client_connect_import,
2000         .o_disconnect   = client_disconnect_export,
2001         //.o_enqueue      = mgc_enqueue,
2002         .o_cancel       = mgc_cancel,
2003         //.o_iocontrol    = mgc_iocontrol,
2004         .o_set_info_async = mgc_set_info_async,
2005         .o_get_info       = mgc_get_info,
2006         .o_import_event = mgc_import_event,
2007         .o_llog_init    = mgc_llog_init,
2008         .o_llog_finish  = mgc_llog_finish,
2009         .o_process_config = mgc_process_config,
2010 };
2011
2012 int __init mgc_init(void)
2013 {
2014         return class_register_type(&mgc_obd_ops, NULL, NULL,
2015                                    LUSTRE_MGC_NAME, NULL);
2016 }
2017
2018 #ifdef __KERNEL__
2019 static void /*__exit*/ mgc_exit(void)
2020 {
2021         class_unregister_type(LUSTRE_MGC_NAME);
2022 }
2023
2024 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2025 MODULE_DESCRIPTION("Lustre Management Client");
2026 MODULE_LICENSE("GPL");
2027
2028 module_init(mgc_init);
2029 module_exit(mgc_exit);
2030 #endif