lustre/mgc/mgc_request.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  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 #ifndef EXPORT_SYMTAB
  42 # define EXPORT_SYMTAB
  43 #endif
  44 #define DEBUG_SUBSYSTEM S_MGC
  45 #define D_MGC D_CONFIG /*|D_WARNING*/
  46
  47 #ifdef __KERNEL__
  48 # include <linux/module.h>
  49 # include <linux/pagemap.h>
  50 # include <linux/miscdevice.h>
  51 # include <linux/init.h>
  52 #else
  53 # include <liblustre.h>
  54 #endif
  55
  56 #include <obd_class.h>
  57 #include <lustre_dlm.h>
  58 #include <lprocfs_status.h>
  59 #include <lustre_log.h>
  60 #include <lustre_fsfilt.h>
  61 #include <lustre_disk.h>
  62 #include "mgc_internal.h"
  63
  64 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id)
  65 {
  66         __u64 resname = 0;
  67
  68         if (len > 8) {
  69                 CERROR("name too long: %s\n", name);
  70                 return -EINVAL;
  71         }
  72         if (len <= 0) {
  73                 CERROR("missing name: %s\n", name);
  74                 return -EINVAL;
  75         }
  76         memcpy(&resname, name, len);
  77
  78         memset(res_id, 0, sizeof(*res_id));
  79
  80         /* Always use the same endianness for the resid */
  81         res_id->name[0] = cpu_to_le64(resname);
  82         CDEBUG(D_MGC, "log %s to resid "LPX64"/"LPX64" (%.8s)\n", name,
  83                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  84         return 0;
  85 }
  86
  87 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id)
  88 {
  89         /* fsname is at most 8 chars long, maybe contain "-".
  90          * e.g. "lustre", "SUN-000" */
  91         return mgc_name2resid(fsname, strlen(fsname), res_id);
  92 }
  93 EXPORT_SYMBOL(mgc_fsname2resid);
  94
  95 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id)
  96 {
  97         char *name_end;
  98         int len;
  99
 100         /* logname consists of "fsname-nodetype".
 101          * e.g. "lustre-MDT0001", "SUN-000-client" */
 102         name_end = strrchr(logname, '-');
 103         LASSERT(name_end);
 104         len = name_end - logname;
 105         return mgc_name2resid(logname, len, res_id);
 106 }
 107
 108 /********************** config llog list **********************/
 109 static CFS_LIST_HEAD(config_llog_list);
 110 static cfs_spinlock_t       config_list_lock = CFS_SPIN_LOCK_UNLOCKED;
 111
 112 /* Take a reference to a config log */
 113 static int config_log_get(struct config_llog_data *cld)
 114 {
 115         ENTRY;
 116         if (cld->cld_stopping)
 117                 RETURN(1);
 118         cfs_atomic_inc(&cld->cld_refcount);
 119         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 120                cfs_atomic_read(&cld->cld_refcount));
 121         RETURN(0);
 122 }
 123
 124 /* Drop a reference to a config log.  When no longer referenced,
 125    we can free the config log data */
 126 static void config_log_put(struct config_llog_data *cld)
 127 {
 128         ENTRY;
 129
 130         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 131                cfs_atomic_read(&cld->cld_refcount));
 132         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 133
 134         /* spinlock to make sure no item with 0 refcount in the list */
 135         cfs_spin_lock(&config_list_lock);
 136         if (unlikely(cfs_atomic_dec_and_test(&cld->cld_refcount))) {
 137                 cfs_list_del(&cld->cld_list_chain);
 138                 cfs_spin_unlock(&config_list_lock);
 139
 140                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 141
 142                 if (cld->cld_sptlrpc)
 143                         config_log_put(cld->cld_sptlrpc);
 144                 if (cld->cld_is_sptlrpc)
 145                         sptlrpc_conf_log_stop(cld->cld_logname);
 146
 147                 class_export_put(cld->cld_mgcexp);
 148                 OBD_FREE(cld->cld_logname, strlen(cld->cld_logname) + 1);
 149                 if (cld->cld_cfg.cfg_instance != NULL)
 150                         OBD_FREE(cld->cld_cfg.cfg_instance,
 151                                  strlen(cld->cld_cfg.cfg_instance) + 1);
 152                 OBD_FREE(cld, sizeof(*cld));
 153         } else {
 154                 cfs_spin_unlock(&config_list_lock);
 155         }
 156
 157         EXIT;
 158 }
 159
 160 /* Find a config log by name */
 161 static
 162 struct config_llog_data *config_log_find(char *logname,
 163                                          struct config_llog_instance *cfg)
 164 {
 165         struct config_llog_data *cld;
 166         char *logid = logname;
 167         int match_instance = 0;
 168         ENTRY;
 169
 170         if (cfg && cfg->cfg_instance) {
 171                 match_instance++;
 172                 logid = cfg->cfg_instance;
 173         }
 174         if (!logid) {
 175                 CERROR("No log specified\n");
 176                 RETURN(ERR_PTR(-EINVAL));
 177         }
 178
 179         cfs_spin_lock(&config_list_lock);
 180         cfs_list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 181                 if (match_instance && cld->cld_cfg.cfg_instance &&
 182                     strcmp(logid, cld->cld_cfg.cfg_instance) == 0)
 183                         goto out_found;
 184                 if (!match_instance &&
 185                     strcmp(logid, cld->cld_logname) == 0)
 186                         goto out_found;
 187         }
 188         cfs_spin_unlock(&config_list_lock);
 189
 190         CDEBUG(D_CONFIG, "can't get log %s\n", logid);
 191         RETURN(ERR_PTR(-ENOENT));
 192 out_found:
 193         cfs_atomic_inc(&cld->cld_refcount);
 194         cfs_spin_unlock(&config_list_lock);
 195         LASSERT(cld->cld_stopping == 0 || cld->cld_is_sptlrpc == 0);
 196         RETURN(cld);
 197 }
 198
 199 static
 200 struct config_llog_data *do_config_log_add(struct obd_device *obd,
 201                                            char *logname,
 202                                            unsigned int is_sptlrpc,
 203                                            struct config_llog_instance *cfg,
 204                                            struct super_block *sb)
 205 {
 206         struct config_llog_data *cld;
 207         int                      rc;
 208         ENTRY;
 209
 210         CDEBUG(D_MGC, "do adding config log %s:%s\n", logname,
 211                cfg ? cfg->cfg_instance : "NULL");
 212
 213         OBD_ALLOC(cld, sizeof(*cld));
 214         if (!cld)
 215                 RETURN(ERR_PTR(-ENOMEM));
 216         OBD_ALLOC(cld->cld_logname, strlen(logname) + 1);
 217         if (!cld->cld_logname) {
 218                 OBD_FREE(cld, sizeof(*cld));
 219                 RETURN(ERR_PTR(-ENOMEM));
 220         }
 221         strcpy(cld->cld_logname, logname);
 222         if (cfg)
 223                 cld->cld_cfg = *cfg;
 224         cld->cld_cfg.cfg_last_idx = 0;
 225         cld->cld_cfg.cfg_flags = 0;
 226         cld->cld_cfg.cfg_sb = sb;
 227         cld->cld_is_sptlrpc = is_sptlrpc;
 228         cfs_atomic_set(&cld->cld_refcount, 1);
 229
 230         /* Keep the mgc around until we are done */
 231         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 232
 233         if (cfg && cfg->cfg_instance != NULL) {
 234                 OBD_ALLOC(cld->cld_cfg.cfg_instance,
 235                           strlen(cfg->cfg_instance) + 1);
 236                 strcpy(cld->cld_cfg.cfg_instance, cfg->cfg_instance);
 237         }
 238
 239         if (is_sptlrpc) {
 240                 sptlrpc_conf_log_start(logname);
 241                 cld->cld_cfg.cfg_obdname = obd->obd_name;
 242         }
 243
 244         rc = mgc_logname2resid(logname, &cld->cld_resid);
 245
 246         cfs_spin_lock(&config_list_lock);
 247         cfs_list_add(&cld->cld_list_chain, &config_llog_list);
 248         cfs_spin_unlock(&config_list_lock);
 249
 250         if (rc) {
 251                 config_log_put(cld);
 252                 RETURN(ERR_PTR(rc));
 253         }
 254
 255         if (is_sptlrpc) {
 256                 rc = mgc_process_log(obd, cld);
 257                 if (rc)
 258                         CERROR("failed processing sptlrpc log: %d\n", rc);
 259         }
 260
 261         RETURN(cld);
 262 }
 263
 264 /**
 265  * Add this log to our list of active logs.
 266  * We have one active log per "mount" - client instance or servername.
 267  * Each instance may be at a different point in the log.
 268  */
 269 static int config_log_add(struct obd_device *obd, char *logname,
 270                           struct config_llog_instance *cfg,
 271                           struct super_block *sb)
 272 {
 273         struct config_llog_data *cld, *sptlrpc_cld;
 274         char                     seclogname[20];
 275         char                    *ptr;
 276         ENTRY;
 277
 278         CDEBUG(D_MGC, "adding config log %s:%s\n", logname, cfg->cfg_instance);
 279
 280         /*
 281          * for each regular log, the depended sptlrpc log name is
 282          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 283          */
 284         ptr = strrchr(logname, '-');
 285         if (ptr == NULL || ptr - logname > 8) {
 286                 CERROR("logname %s is too long\n", logname);
 287                 RETURN(-EINVAL);
 288         }
 289
 290         memcpy(seclogname, logname, ptr - logname);
 291         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 292
 293         sptlrpc_cld = config_log_find(seclogname, NULL);
 294         if (IS_ERR(sptlrpc_cld)) {
 295                 sptlrpc_cld = do_config_log_add(obd, seclogname, 1, NULL, NULL);
 296                 if (IS_ERR(sptlrpc_cld)) {
 297                         CERROR("can't create sptlrpc log: %s\n", seclogname);
 298                         RETURN(PTR_ERR(sptlrpc_cld));
 299                 }
 300         }
 301
 302         cld = do_config_log_add(obd, logname, 0, cfg, sb);
 303         if (IS_ERR(cld)) {
 304                 CERROR("can't create log: %s\n", logname);
 305                 config_log_put(sptlrpc_cld);
 306                 RETURN(PTR_ERR(cld));
 307         }
 308
 309         cld->cld_sptlrpc = sptlrpc_cld;
 310
 311         RETURN(0);
 312 }
 313
 314 CFS_DECLARE_MUTEX(llog_process_lock);
 315
 316 /* Stop watching for updates on this log. */
 317 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 318 {
 319         struct config_llog_data *cld, *cld_sptlrpc = NULL;
 320         int rc = 0;
 321         ENTRY;
 322
 323         cld = config_log_find(logname, cfg);
 324         if (IS_ERR(cld))
 325                 RETURN(PTR_ERR(cld));
 326
 327         cfs_down(&llog_process_lock);
 328         /*
 329          * if cld_stopping is set, it means we didn't start the log thus
 330          * not owning the start ref. this can happen after previous umount:
 331          * the cld still hanging there waiting for lock cancel, and we
 332          * remount again but failed in the middle and call log_end without
 333          * calling start_log.
 334          */
 335         if (unlikely(cld->cld_stopping)) {
 336                 cfs_up(&llog_process_lock);
 337                 /* drop the ref from the find */
 338                 config_log_put(cld);
 339                 RETURN(rc);
 340         }
 341
 342         cld->cld_stopping = 1;
 343         cfs_up(&llog_process_lock);
 344
 345         cfs_spin_lock(&config_list_lock);
 346         cld_sptlrpc = cld->cld_sptlrpc;
 347         cld->cld_sptlrpc = NULL;
 348         cfs_spin_unlock(&config_list_lock);
 349
 350         if (cld_sptlrpc)
 351                 config_log_put(cld_sptlrpc);
 352
 353         /* drop the ref from the find */
 354         config_log_put(cld);
 355         /* drop the start ref */
 356         config_log_put(cld);
 357
 358         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 359                rc);
 360         RETURN(rc);
 361 }
 362
 363 /* reenqueue any lost locks */
 364 #define RQ_RUNNING 0x1
 365 #define RQ_NOW     0x2
 366 #define RQ_LATER   0x4
 367 #define RQ_STOP    0x8
 368 static int rq_state = 0;
 369 static cfs_waitq_t rq_waitq;
 370
 371 static int mgc_requeue_add(struct config_llog_data *cld, int later);
 372
 373 static void do_requeue(struct config_llog_data *cld)
 374 {
 375         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 376
 377         /* Do not run mgc_process_log on a disconnected export or an
 378            export which is being disconnected. Take the client
 379            semaphore to make the check non-racy. */
 380         cfs_down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 381         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 382                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 383                 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 384         } else {
 385                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 386                        cld->cld_logname);
 387         }
 388         cfs_up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 389
 390         /* Whether we enqueued again or not in mgc_process_log, we're done
 391          * with the ref from the old enqueue */
 392         config_log_put(cld);
 393 }
 394
 395 static int mgc_requeue_thread(void *data)
 396 {
 397         struct l_wait_info lwi_now, lwi_later;
 398         struct config_llog_data *cld, *cld_next, *cld_prev;
 399         char name[] = "ll_cfg_requeue";
 400         int rc = 0;
 401         ENTRY;
 402
 403         cfs_daemonize(name);
 404
 405         CDEBUG(D_MGC, "Starting requeue thread\n");
 406
 407         lwi_later = LWI_TIMEOUT(60 * CFS_HZ, NULL, NULL);
 408         l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP), &lwi_later);
 409
 410         /* Keep trying failed locks periodically */
 411         cfs_spin_lock(&config_list_lock);
 412         while (rq_state & (RQ_NOW | RQ_LATER)) {
 413                 /* Any new or requeued lostlocks will change the state */
 414                 rq_state &= ~(RQ_NOW | RQ_LATER);
 415                 cfs_spin_unlock(&config_list_lock);
 416
 417                 /* Always wait a few seconds to allow the server who
 418                    caused the lock revocation to finish its setup, plus some
 419                    random so everyone doesn't try to reconnect at once. */
 420                 lwi_now = LWI_TIMEOUT(3 * CFS_HZ + (ll_rand() & 0xff) * \
 421                                       (CFS_HZ / 100),
 422                                       NULL, NULL);
 423                 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi_now);
 424
 425                 /*
 426                  * iterate & processing through the list. for each cld, process
 427                  * its depending sptlrpc cld firstly (if any) and then itself.
 428                  *
 429                  * it's guaranteed any item in the list must have
 430                  * reference > 0; and if cld_lostlock is set, at
 431                  * least one reference is taken by the previous enqueue.
 432                  *
 433                  * Note: releasing a cld might lead to itself and its depended
 434                  * sptlrpc cld be unlinked from the list. to safely iterate
 435                  * we need to take a reference on next cld before processing.
 436                  */
 437                 cld_prev = NULL;
 438
 439                 cfs_spin_lock(&config_list_lock);
 440                 cfs_list_for_each_entry_safe(cld, cld_next, &config_llog_list,
 441                                              cld_list_chain) {
 442                         if (cld->cld_list_chain.next != &config_llog_list)
 443                                 cfs_atomic_inc(&cld_next->cld_refcount);
 444
 445                         if (cld->cld_lostlock) {
 446                                 if (cld->cld_sptlrpc &&
 447                                     cld->cld_sptlrpc->cld_lostlock) {
 448                                         cld->cld_sptlrpc->cld_lostlock = 0;
 449
 450                                         cfs_spin_unlock(&config_list_lock);
 451                                         do_requeue(cld->cld_sptlrpc);
 452                                         cfs_spin_lock(&config_list_lock);
 453                                         LASSERT(cld->cld_lostlock);
 454                                 }
 455
 456                                 cld->cld_lostlock = 0;
 457
 458                                 cfs_spin_unlock(&config_list_lock);
 459                                 do_requeue(cld);
 460                                 cfs_spin_lock(&config_list_lock);
 461                         }
 462
 463
 464                         if (cld_prev) {
 465                                 cfs_spin_unlock(&config_list_lock);
 466                                 config_log_put(cld_prev);
 467                                 cfs_spin_lock(&config_list_lock);
 468                         }
 469
 470                         cld_prev = cld_next;
 471                 }
 472                 cfs_spin_unlock(&config_list_lock);
 473
 474                 /* Wait a bit to see if anyone else needs a requeue */
 475                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 476                              &lwi_later);
 477                 cfs_spin_lock(&config_list_lock);
 478         }
 479         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 480         rq_state &= ~RQ_RUNNING;
 481         cfs_spin_unlock(&config_list_lock);
 482
 483         CDEBUG(D_MGC, "Ending requeue thread\n");
 484         RETURN(rc);
 485 }
 486
 487 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 488    We are responsible for dropping the config log reference from here on out. */
 489 static int mgc_requeue_add(struct config_llog_data *cld, int later)
 490 {
 491         int rc = 0;
 492
 493         CDEBUG(D_INFO, "log %s: requeue (l=%d r=%d sp=%d st=%x)\n",
 494                cld->cld_logname, later, cfs_atomic_read(&cld->cld_refcount),
 495                cld->cld_stopping, rq_state);
 496         LASSERT(cfs_atomic_read(&cld->cld_refcount) > 0);
 497
 498         /* Hold lock for rq_state */
 499         cfs_spin_lock(&config_list_lock);
 500
 501         if (cld->cld_stopping || (rq_state & RQ_STOP)) {
 502                 cld->cld_lostlock = 0;
 503                 cfs_spin_unlock(&config_list_lock);
 504                 config_log_put(cld);
 505                 RETURN(0);
 506         }
 507
 508         cld->cld_lostlock = 1;
 509
 510         if (!(rq_state & RQ_RUNNING)) {
 511                 LASSERT(rq_state == 0);
 512                 rq_state = RQ_RUNNING | (later ? RQ_LATER : RQ_NOW);
 513                 cfs_spin_unlock(&config_list_lock);
 514                 rc = cfs_kernel_thread(mgc_requeue_thread, 0,
 515                                        CLONE_VM | CLONE_FILES);
 516                 if (rc < 0) {
 517                         CERROR("log %s: cannot start requeue thread (%d),"
 518                                "no more log updates!\n", cld->cld_logname, rc);
 519                         /* Drop the ref, since the rq thread won't */
 520                         cld->cld_lostlock = 0;
 521                         config_log_put(cld);
 522                         rq_state = 0;
 523                         RETURN(rc);
 524                 }
 525         } else {
 526                 rq_state |= later ? RQ_LATER : RQ_NOW;
 527                 cfs_spin_unlock(&config_list_lock);
 528                 cfs_waitq_signal(&rq_waitq);
 529         }
 530
 531         RETURN(0);
 532 }
 533
 534 /********************** class fns **********************/
 535
 536 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
 537                         struct vfsmount *mnt)
 538 {
 539         struct lvfs_run_ctxt saved;
 540         struct lustre_sb_info *lsi = s2lsi(sb);
 541         struct client_obd *cli = &obd->u.cli;
 542         struct dentry *dentry;
 543         char *label;
 544         int err = 0;
 545         ENTRY;
 546
 547         LASSERT(lsi);
 548         LASSERT(lsi->lsi_srv_mnt == mnt);
 549
 550         /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
 551         cfs_down(&cli->cl_mgc_sem);
 552
 553         cfs_cleanup_group_info();
 554
 555         obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
 556         if (IS_ERR(obd->obd_fsops)) {
 557                 cfs_up(&cli->cl_mgc_sem);
 558                 CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd),
 559                        PTR_ERR(obd->obd_fsops));
 560                 RETURN(PTR_ERR(obd->obd_fsops));
 561         }
 562
 563         cli->cl_mgc_vfsmnt = mnt;
 564         fsfilt_setup(obd, mnt->mnt_sb);
 565
 566         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 567         obd->obd_lvfs_ctxt.pwdmnt = mnt;
 568         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
 569         obd->obd_lvfs_ctxt.fs = get_ds();
 570
 571         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 572         dentry = lookup_one_len(MOUNT_CONFIGS_DIR, cfs_fs_pwd(current->fs),
 573                                 strlen(MOUNT_CONFIGS_DIR));
 574         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 575         if (IS_ERR(dentry)) {
 576                 err = PTR_ERR(dentry);
 577                 CERROR("cannot lookup %s directory: rc = %d\n",
 578                        MOUNT_CONFIGS_DIR, err);
 579                 GOTO(err_ops, err);
 580         }
 581         cli->cl_mgc_configs_dir = dentry;
 582
 583         /* We take an obd ref to insure that we can't get to mgc_cleanup
 584            without calling mgc_fs_cleanup first. */
 585         class_incref(obd, "mgc_fs", obd);
 586
 587         label = fsfilt_get_label(obd, mnt->mnt_sb);
 588         if (label)
 589                 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
 590
 591         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 592         RETURN(0);
 593
 594 err_ops:
 595         fsfilt_put_ops(obd->obd_fsops);
 596         obd->obd_fsops = NULL;
 597         cli->cl_mgc_vfsmnt = NULL;
 598         cfs_up(&cli->cl_mgc_sem);
 599         RETURN(err);
 600 }
 601
 602 static int mgc_fs_cleanup(struct obd_device *obd)
 603 {
 604         struct client_obd *cli = &obd->u.cli;
 605         int rc = 0;
 606         ENTRY;
 607
 608         LASSERT(cli->cl_mgc_vfsmnt != NULL);
 609
 610         if (cli->cl_mgc_configs_dir != NULL) {
 611                 struct lvfs_run_ctxt saved;
 612                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 613                 l_dput(cli->cl_mgc_configs_dir);
 614                 cli->cl_mgc_configs_dir = NULL;
 615                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 616                 class_decref(obd, "mgc_fs", obd);
 617         }
 618
 619         cli->cl_mgc_vfsmnt = NULL;
 620         if (obd->obd_fsops)
 621                 fsfilt_put_ops(obd->obd_fsops);
 622
 623         cfs_up(&cli->cl_mgc_sem);
 624
 625         RETURN(rc);
 626 }
 627
 628 static cfs_atomic_t mgc_count = CFS_ATOMIC_INIT(0);
 629 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 630 {
 631         int rc = 0;
 632         ENTRY;
 633
 634         switch (stage) {
 635         case OBD_CLEANUP_EARLY:
 636                 break;
 637         case OBD_CLEANUP_EXPORTS:
 638                 if (cfs_atomic_dec_and_test(&mgc_count)) {
 639                         /* Kick the requeue waitq - cld's should all be
 640                            stopping */
 641                         cfs_spin_lock(&config_list_lock);
 642                         rq_state |= RQ_STOP;
 643                         cfs_spin_unlock(&config_list_lock);
 644                         cfs_waitq_signal(&rq_waitq);
 645                 }
 646                 rc = obd_llog_finish(obd, 0);
 647                 if (rc != 0)
 648                         CERROR("failed to cleanup llogging subsystems\n");
 649                 break;
 650         }
 651         RETURN(rc);
 652 }
 653
 654 static int mgc_cleanup(struct obd_device *obd)
 655 {
 656         struct client_obd *cli = &obd->u.cli;
 657         int rc;
 658         ENTRY;
 659
 660         LASSERT(cli->cl_mgc_vfsmnt == NULL);
 661
 662         /* COMPAT_146 - old config logs may have added profiles we don't
 663            know about */
 664         if (obd->obd_type->typ_refcnt <= 1)
 665                 /* Only for the last mgc */
 666                 class_del_profiles();
 667
 668         lprocfs_obd_cleanup(obd);
 669         ptlrpcd_decref();
 670
 671         rc = client_obd_cleanup(obd);
 672         RETURN(rc);
 673 }
 674
 675 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 676 {
 677         struct lprocfs_static_vars lvars;
 678         int rc;
 679         ENTRY;
 680
 681         ptlrpcd_addref();
 682
 683         rc = client_obd_setup(obd, lcfg);
 684         if (rc)
 685                 GOTO(err_decref, rc);
 686
 687         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
 688         if (rc) {
 689                 CERROR("failed to setup llogging subsystems\n");
 690                 GOTO(err_cleanup, rc);
 691         }
 692
 693         lprocfs_mgc_init_vars(&lvars);
 694         lprocfs_obd_setup(obd, lvars.obd_vars);
 695         sptlrpc_lprocfs_cliobd_attach(obd);
 696
 697         cfs_spin_lock(&config_list_lock);
 698         cfs_atomic_inc(&mgc_count);
 699         if (cfs_atomic_read(&mgc_count) == 1) {
 700                 rq_state &= ~RQ_STOP;
 701                 cfs_waitq_init(&rq_waitq);
 702         }
 703         cfs_spin_unlock(&config_list_lock);
 704
 705         RETURN(rc);
 706
 707 err_cleanup:
 708         client_obd_cleanup(obd);
 709 err_decref:
 710         ptlrpcd_decref();
 711         RETURN(rc);
 712 }
 713
 714 /* based on ll_mdc_blocking_ast */
 715 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 716                             void *data, int flag)
 717 {
 718         struct lustre_handle lockh;
 719         struct config_llog_data *cld = (struct config_llog_data *)data;
 720         int rc = 0;
 721         ENTRY;
 722
 723         switch (flag) {
 724         case LDLM_CB_BLOCKING:
 725                 /* mgs wants the lock, give it up... */
 726                 LDLM_DEBUG(lock, "MGC blocking CB");
 727                 ldlm_lock2handle(lock, &lockh);
 728                 rc = ldlm_cli_cancel(&lockh);
 729                 break;
 730         case LDLM_CB_CANCELING: {
 731                 /* We've given up the lock, prepare ourselves to update. */
 732                 LDLM_DEBUG(lock, "MGC cancel CB");
 733
 734                 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
 735                        lock->l_resource->lr_name.name[0],
 736                        (char *)&lock->l_resource->lr_name.name[0]);
 737
 738                 if (!cld) {
 739                         CERROR("missing data, won't requeue\n");
 740                         break;
 741                 }
 742                 /* Are we done with this log? */
 743                 if (cld->cld_stopping) {
 744                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
 745                                cld->cld_logname);
 746                         config_log_put(cld);
 747                         break;
 748                 }
 749                 /* Make sure not to re-enqueue when the mgc is stopping
 750                    (we get called from client_disconnect_export) */
 751                 if (!lock->l_conn_export ||
 752                     !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
 753                         CDEBUG(D_MGC, "log %s: disconnecting, won't requeue\n",
 754                                cld->cld_logname);
 755                         config_log_put(cld);
 756                         break;
 757                 }
 758                 /* Did we fail to get the lock? */
 759                 if (lock->l_req_mode != lock->l_granted_mode &&
 760                     !cld->cld_is_sptlrpc) {
 761                         CDEBUG(D_MGC, "log %s: original grant failed, will "
 762                                "requeue later\n", cld->cld_logname);
 763                         /* Try to re-enqueue later */
 764                         rc = mgc_requeue_add(cld, 1);
 765                         break;
 766                 }
 767                 /* Re-enqueue now */
 768                 rc = mgc_requeue_add(cld, 0);
 769                 break;
 770         }
 771         default:
 772                 LBUG();
 773         }
 774
 775
 776         if (rc) {
 777                 CERROR("%s CB failed %d:\n", flag == LDLM_CB_BLOCKING ?
 778                        "blocking" : "cancel", rc);
 779                 LDLM_ERROR(lock, "MGC ast");
 780         }
 781         RETURN(rc);
 782 }
 783
 784 /* Send parameter to MGS*/
 785 static int mgc_set_mgs_param(struct obd_export *exp,
 786                              struct mgs_send_param *msp)
 787 {
 788         struct ptlrpc_request *req;
 789         struct mgs_send_param *req_msp, *rep_msp;
 790         int rc;
 791         ENTRY;
 792
 793         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 794                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
 795                                         MGS_SET_INFO);
 796         if (!req)
 797                 RETURN(-ENOMEM);
 798
 799         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 800         if (!req_msp) {
 801                 ptlrpc_req_finished(req);
 802                 RETURN(-ENOMEM);
 803         }
 804
 805         memcpy(req_msp, msp, sizeof(*req_msp));
 806         ptlrpc_request_set_replen(req);
 807
 808         rc = ptlrpc_queue_wait(req);
 809         if (!rc) {
 810                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 811                 memcpy(msp, rep_msp, sizeof(*rep_msp));
 812         }
 813
 814         ptlrpc_req_finished(req);
 815
 816         RETURN(rc);
 817 }
 818
 819 /* Take a config lock so we can get cancel notifications */
 820 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
 821                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
 822                        int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
 823                        void *data, __u32 lvb_len, void *lvb_swabber,
 824                        struct lustre_handle *lockh)
 825 {
 826         struct config_llog_data *cld = (struct config_llog_data *)data;
 827         struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
 828                          ldlm_completion_ast, NULL, NULL, data};
 829
 830         int rc;
 831         ENTRY;
 832
 833         CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
 834                cld->cld_resid.name[0]);
 835
 836         /* We can only drop this config log ref when we drop the lock */
 837         if (config_log_get(cld))
 838                 RETURN(ELDLM_LOCK_ABORTED);
 839
 840         /* We need a callback for every lockholder, so don't try to
 841            ldlm_lock_match (see rev 1.1.2.11.2.47) */
 842
 843         rc = ldlm_cli_enqueue(exp, NULL, &einfo, &cld->cld_resid, NULL, flags,
 844                               NULL, 0, lockh, 0);
 845         /* A failed enqueue should still call the mgc_blocking_ast,
 846            where it will be requeued if needed ("grant failed"). */
 847
 848         RETURN(rc);
 849 }
 850
 851 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
 852                       __u32 mode, struct lustre_handle *lockh)
 853 {
 854         ENTRY;
 855
 856         ldlm_lock_decref(lockh, mode);
 857
 858         RETURN(0);
 859 }
 860
 861 #if 0
 862 static int mgc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
 863                          void *karg, void *uarg)
 864 {
 865         struct obd_device *obd = exp->exp_obd;
 866         struct obd_ioctl_data *data = karg;
 867         struct llog_ctxt *ctxt;
 868         struct lvfs_run_ctxt saved;
 869         int rc;
 870         ENTRY;
 871
 872         if (!cfs_try_module_get(THIS_MODULE)) {
 873                 CERROR("Can't get module. Is it alive?");
 874                 return -EINVAL;
 875         }
 876         switch (cmd) {
 877         /* REPLicator context */
 878         case OBD_IOC_PARSE: {
 879                 CERROR("MGC parsing llog %s\n", data->ioc_inlbuf1);
 880                 ctxt = llog_get_context(exp->exp_obd, LLOG_CONFIG_REPL_CTXT);
 881                 rc = class_config_parse_llog(ctxt, data->ioc_inlbuf1, NULL);
 882                 GOTO(out, rc);
 883         }
 884 #ifdef __KERNEL__
 885         case OBD_IOC_LLOG_INFO:
 886         case OBD_IOC_LLOG_PRINT: {
 887                 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 888                 rc = llog_ioctl(ctxt, cmd, data);
 889
 890                 GOTO(out, rc);
 891         }
 892 #endif
 893         /* ORIGinator context */
 894         case OBD_IOC_DUMP_LOG: {
 895                 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 896                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 897                 rc = class_config_dump_llog(ctxt, data->ioc_inlbuf1, NULL);
 898                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 899                 if (rc)
 900                         RETURN(rc);
 901
 902                 GOTO(out, rc);
 903         }
 904         default:
 905                 CERROR("mgc_ioctl(): unrecognised ioctl %#x\n", cmd);
 906                 GOTO(out, rc = -ENOTTY);
 907         }
 908 out:
 909         cfs_module_put(THIS_MODULE);
 910
 911         return rc;
 912 }
 913 #endif
 914
 915 /* Send target_reg message to MGS */
 916 static int mgc_target_register(struct obd_export *exp,
 917                                struct mgs_target_info *mti)
 918 {
 919         struct ptlrpc_request  *req;
 920         struct mgs_target_info *req_mti, *rep_mti;
 921         int                     rc;
 922         ENTRY;
 923
 924         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 925                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
 926                                         MGS_TARGET_REG);
 927         if (req == NULL)
 928                 RETURN(-ENOMEM);
 929
 930         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
 931         if (!req_mti) {
 932                 ptlrpc_req_finished(req);
 933                 RETURN(-ENOMEM);
 934         }
 935
 936         memcpy(req_mti, mti, sizeof(*req_mti));
 937         ptlrpc_request_set_replen(req);
 938         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
 939
 940         rc = ptlrpc_queue_wait(req);
 941         if (!rc) {
 942                 rep_mti = req_capsule_server_get(&req->rq_pill,
 943                                                  &RMF_MGS_TARGET_INFO);
 944                 memcpy(mti, rep_mti, sizeof(*rep_mti));
 945                 CDEBUG(D_MGC, "register %s got index = %d\n",
 946                        mti->mti_svname, mti->mti_stripe_index);
 947         }
 948         ptlrpc_req_finished(req);
 949
 950         RETURN(rc);
 951 }
 952
 953 int mgc_set_info_async(struct obd_export *exp, obd_count keylen,
 954                        void *key, obd_count vallen, void *val,
 955                        struct ptlrpc_request_set *set)
 956 {
 957         struct obd_import *imp = class_exp2cliimp(exp);
 958         int rc = -EINVAL;
 959         ENTRY;
 960
 961         /* Try to "recover" the initial connection; i.e. retry */
 962         if (KEY_IS(KEY_INIT_RECOV)) {
 963                 if (vallen != sizeof(int))
 964                         RETURN(-EINVAL);
 965                 cfs_spin_lock(&imp->imp_lock);
 966                 imp->imp_initial_recov = *(int *)val;
 967                 cfs_spin_unlock(&imp->imp_lock);
 968                 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
 969                        exp->exp_obd->obd_name, imp->imp_initial_recov);
 970                 RETURN(0);
 971         }
 972         /* Turn off initial_recov after we try all backup servers once */
 973         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
 974                 int value;
 975                 if (vallen != sizeof(int))
 976                         RETURN(-EINVAL);
 977                 value = *(int *)val;
 978                 cfs_spin_lock(&imp->imp_lock);
 979                 imp->imp_initial_recov_bk = value > 0;
 980                 /* Even after the initial connection, give up all comms if
 981                    nobody answers the first time. */
 982                 imp->imp_recon_bk = 1;
 983                 cfs_spin_unlock(&imp->imp_lock);
 984                 CDEBUG(D_MGC, "InitRecov %s %d/%d:d%d:i%d:r%d:or%d:%s\n",
 985                        imp->imp_obd->obd_name, value, imp->imp_initial_recov,
 986                        imp->imp_deactive, imp->imp_invalid,
 987                        imp->imp_replayable, imp->imp_obd->obd_replayable,
 988                        ptlrpc_import_state_name(imp->imp_state));
 989                 /* Resurrect if we previously died */
 990                 if (imp->imp_invalid || value > 1)
 991                         ptlrpc_reconnect_import(imp);
 992                 RETURN(0);
 993         }
 994         /* FIXME move this to mgc_process_config */
 995         if (KEY_IS(KEY_REGISTER_TARGET)) {
 996                 struct mgs_target_info *mti;
 997                 if (vallen != sizeof(struct mgs_target_info))
 998                         RETURN(-EINVAL);
 999                 mti = (struct mgs_target_info *)val;
1000                 CDEBUG(D_MGC, "register_target %s %#x\n",
1001                        mti->mti_svname, mti->mti_flags);
1002                 rc =  mgc_target_register(exp, mti);
1003                 RETURN(rc);
1004         }
1005         if (KEY_IS(KEY_SET_FS)) {
1006                 struct super_block *sb = (struct super_block *)val;
1007                 struct lustre_sb_info *lsi;
1008                 if (vallen != sizeof(struct super_block))
1009                         RETURN(-EINVAL);
1010                 lsi = s2lsi(sb);
1011                 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
1012                 if (rc) {
1013                         CERROR("set_fs got %d\n", rc);
1014                 }
1015                 RETURN(rc);
1016         }
1017         if (KEY_IS(KEY_CLEAR_FS)) {
1018                 if (vallen != 0)
1019                         RETURN(-EINVAL);
1020                 rc = mgc_fs_cleanup(exp->exp_obd);
1021                 if (rc) {
1022                         CERROR("clear_fs got %d\n", rc);
1023                 }
1024                 RETURN(rc);
1025         }
1026         if (KEY_IS(KEY_SET_INFO)) {
1027                 struct mgs_send_param *msp;
1028
1029                 msp = (struct mgs_send_param *)val;
1030                 rc =  mgc_set_mgs_param(exp, msp);
1031                 RETURN(rc);
1032         }
1033         if (KEY_IS(KEY_MGSSEC)) {
1034                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1035                 struct sptlrpc_flavor  flvr;
1036
1037                 /*
1038                  * empty string means using current flavor, if which haven't
1039                  * been set yet, set it as null.
1040                  *
1041                  * if flavor has been set previously, check the asking flavor
1042                  * must match the existing one.
1043                  */
1044                 if (vallen == 0) {
1045                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1046                                 RETURN(0);
1047                         val = "null";
1048                         vallen = 4;
1049                 }
1050
1051                 rc = sptlrpc_parse_flavor(val, &flvr);
1052                 if (rc) {
1053                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1054                                (char *) val);
1055                         RETURN(rc);
1056                 }
1057
1058                 /*
1059                  * caller already hold a mutex
1060                  */
1061                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1062                         cli->cl_flvr_mgc = flvr;
1063                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1064                                   sizeof(flvr)) != 0) {
1065                         char    str[20];
1066
1067                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1068                                             str, sizeof(str));
1069                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1070                                        "currently %s is in use\n",
1071                                        (char *) val, str);
1072                         rc = -EPERM;
1073                 }
1074                 RETURN(rc);
1075         }
1076
1077         RETURN(rc);
1078 }
1079
1080 static int mgc_import_event(struct obd_device *obd,
1081                             struct obd_import *imp,
1082                             enum obd_import_event event)
1083 {
1084         int rc = 0;
1085
1086         LASSERT(imp->imp_obd == obd);
1087         CDEBUG(D_MGC, "import event %#x\n", event);
1088
1089         switch (event) {
1090         case IMP_EVENT_DISCON:
1091                 /* MGC imports should not wait for recovery */
1092                 break;
1093         case IMP_EVENT_INACTIVE:
1094                 break;
1095         case IMP_EVENT_INVALIDATE: {
1096                 struct ldlm_namespace *ns = obd->obd_namespace;
1097                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1098                 break;
1099         }
1100         case IMP_EVENT_ACTIVE:
1101                 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
1102                 /* Clearing obd_no_recov allows us to continue pinging */
1103                 obd->obd_no_recov = 0;
1104                 break;
1105         case IMP_EVENT_OCD:
1106                 break;
1107         default:
1108                 CERROR("Unknown import event %#x\n", event);
1109                 LBUG();
1110         }
1111         RETURN(rc);
1112 }
1113
1114 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1115                          struct obd_device *tgt, int *index)
1116 {
1117         struct llog_ctxt *ctxt;
1118         int rc;
1119         ENTRY;
1120
1121         LASSERT(olg == &obd->obd_olg);
1122
1123         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
1124                         &llog_lvfs_ops);
1125         if (rc)
1126                 RETURN(rc);
1127
1128         rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
1129                         &llog_client_ops);
1130         if (rc == 0) {
1131                 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1132                 if (!ctxt) {
1133                         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1134                         if (ctxt)
1135                                 llog_cleanup(ctxt);
1136                         RETURN(-ENODEV);
1137                 }
1138                 llog_initiator_connect(ctxt);
1139                 llog_ctxt_put(ctxt);
1140         } else {
1141                 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1142                 if (ctxt)
1143                         llog_cleanup(ctxt);
1144         }
1145
1146         RETURN(rc);
1147 }
1148
1149 static int mgc_llog_finish(struct obd_device *obd, int count)
1150 {
1151         struct llog_ctxt *ctxt;
1152         int rc = 0, rc2 = 0;
1153         ENTRY;
1154
1155         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1156         if (ctxt)
1157                 rc = llog_cleanup(ctxt);
1158
1159         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1160         if (ctxt)
1161                 rc2 = llog_cleanup(ctxt);
1162
1163         if (!rc)
1164                 rc = rc2;
1165
1166         RETURN(rc);
1167 }
1168
1169 /* identical to mgs_log_is_empty */
1170 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1171                             char *name)
1172 {
1173         struct lvfs_run_ctxt saved;
1174         struct llog_handle *llh;
1175         int rc = 0;
1176
1177         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1178         rc = llog_create(ctxt, &llh, NULL, name);
1179         if (rc == 0) {
1180                 llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL);
1181                 rc = llog_get_size(llh);
1182                 llog_close(llh);
1183         }
1184         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1185         /* header is record 1 */
1186         return(rc <= 1);
1187 }
1188
1189 static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec,
1190                             void *data)
1191 {
1192         struct llog_rec_hdr local_rec = *rec;
1193         struct llog_handle *local_llh = (struct llog_handle *)data;
1194         char *cfg_buf = (char*) (rec + 1);
1195         struct lustre_cfg *lcfg;
1196         int rc = 0;
1197         ENTRY;
1198
1199         /* Append all records */
1200         local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
1201         rc = llog_write_rec(local_llh, &local_rec, NULL, 0,
1202                             (void *)cfg_buf, -1);
1203
1204         lcfg = (struct lustre_cfg *)cfg_buf;
1205         CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
1206                rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
1207                lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
1208
1209         RETURN(rc);
1210 }
1211
1212 /* Copy a remote log locally */
1213 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1214                          struct llog_ctxt *lctxt, char *logname)
1215 {
1216         struct llog_handle *local_llh, *remote_llh;
1217         struct obd_uuid *uuid;
1218         char *temp_log;
1219         int rc, rc2;
1220         ENTRY;
1221
1222         /* Write new log to a temp name, then vfs_rename over logname
1223            upon successful completion. */
1224
1225         OBD_ALLOC(temp_log, strlen(logname) + 1);
1226         if (!temp_log)
1227                 RETURN(-ENOMEM);
1228         sprintf(temp_log, "%sT", logname);
1229
1230         /* Make sure there's no old temp log */
1231         rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1232         if (rc)
1233                 GOTO(out, rc);
1234         rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL);
1235         if (rc)
1236                 GOTO(out, rc);
1237         rc = llog_destroy(local_llh);
1238         llog_free_handle(local_llh);
1239         if (rc)
1240                 GOTO(out, rc);
1241
1242         /* open local log */
1243         rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1244         if (rc)
1245                 GOTO(out, rc);
1246
1247         /* set the log header uuid for fun */
1248         OBD_ALLOC_PTR(uuid);
1249         obd_str2uuid(uuid, logname);
1250         rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid);
1251         OBD_FREE_PTR(uuid);
1252         if (rc)
1253                 GOTO(out_closel, rc);
1254
1255         /* open remote log */
1256         rc = llog_create(rctxt, &remote_llh, NULL, logname);
1257         if (rc)
1258                 GOTO(out_closel, rc);
1259         rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL);
1260         if (rc)
1261                 GOTO(out_closer, rc);
1262
1263         /* Copy remote log */
1264         rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL);
1265
1266 out_closer:
1267         rc2 = llog_close(remote_llh);
1268         if (!rc)
1269                 rc = rc2;
1270 out_closel:
1271         rc2 = llog_close(local_llh);
1272         if (!rc)
1273                 rc = rc2;
1274
1275         /* We've copied the remote log to the local temp log, now
1276            replace the old local log with the temp log. */
1277         if (!rc) {
1278                 struct client_obd *cli = &obd->u.cli;
1279                 LASSERT(cli);
1280                 LASSERT(cli->cl_mgc_configs_dir);
1281                 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1282                                    temp_log, logname);
1283         }
1284         CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1285 out:
1286         if (rc)
1287                 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1288         OBD_FREE(temp_log, strlen(logname) + 1);
1289         RETURN(rc);
1290 }
1291
1292 /* Get a config log from the MGS and process it.
1293    This func is called for both clients and servers. */
1294 int mgc_process_log(struct obd_device *mgc,
1295                     struct config_llog_data *cld)
1296 {
1297         struct llog_ctxt *ctxt, *lctxt;
1298         struct lustre_handle lockh;
1299         struct client_obd *cli = &mgc->u.cli;
1300         struct lvfs_run_ctxt saved;
1301         struct lustre_sb_info *lsi = NULL;
1302         int rc = 0, rcl, flags = 0, must_pop = 0;
1303         ENTRY;
1304
1305         LASSERT(cld);
1306
1307         /* I don't want multiple processes running process_log at once --
1308            sounds like badness.  It actually might be fine, as long as
1309            we're not trying to update from the same log
1310            simultaneously (in which case we should use a per-log sem.) */
1311         cfs_down(&llog_process_lock);
1312
1313         if (cld->cld_stopping) {
1314                 cfs_up(&llog_process_lock);
1315                 RETURN(0);
1316         }
1317
1318         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1319
1320         if (cld->cld_cfg.cfg_sb)
1321                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1322
1323         CDEBUG(D_MGC, "Process log %s:%s from %d\n", cld->cld_logname,
1324                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1325
1326         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1327         if (!ctxt) {
1328                 CERROR("missing llog context\n");
1329                 cfs_up(&llog_process_lock);
1330                 RETURN(-EINVAL);
1331         }
1332
1333         /* Get the cfg lock on the llog */
1334         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1335                           LCK_CR, &flags, NULL, NULL, NULL,
1336                           cld, 0, NULL, &lockh);
1337         if (rcl)
1338                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1339
1340         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1341
1342         /*
1343          * local copy of sptlrpc log is controlled elsewhere, don't try to
1344          * read it up here.
1345          */
1346         if (rcl && cld->cld_is_sptlrpc)
1347                 goto out_pop;
1348
1349         /* Copy the setup log locally if we can. Don't mess around if we're
1350            running an MGS though (logs are already local). */
1351         if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
1352             (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1353             !IS_MGS(lsi->lsi_ldd)) {
1354                 push_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1355                 must_pop++;
1356                 if (rcl == 0)
1357                         /* Only try to copy log if we have the lock. */
1358                         rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1359                 if (rcl || rc) {
1360                         if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1361                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1362                                                    "log %s and no local copy."
1363                                                    "\n", cld->cld_logname);
1364                                 GOTO(out_pop, rc = -ENOTCONN);
1365                         }
1366                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1367                                "copy for now, will try to update later.\n",
1368                                cld->cld_logname);
1369                 }
1370                 /* Now, whether we copied or not, start using the local llog.
1371                    If we failed to copy, we'll start using whatever the old
1372                    log has. */
1373                 llog_ctxt_put(ctxt);
1374                 ctxt = lctxt;
1375         }
1376
1377         if (cld->cld_is_sptlrpc)
1378                 sptlrpc_conf_log_update_begin(cld->cld_logname);
1379
1380         /* logname and instance info should be the same, so use our
1381            copy of the instance for the update.  The cfg_last_idx will
1382            be updated here. */
1383         rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg);
1384 out_pop:
1385         llog_ctxt_put(ctxt);
1386         if (ctxt != lctxt)
1387                 llog_ctxt_put(lctxt);
1388         if (must_pop)
1389                 pop_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1390
1391         /*
1392          * update settings on existing OBDs. doing it inside
1393          * of llog_process_lock so no device is attaching/detaching
1394          * in parallel.
1395          * the logname must be <fsname>-sptlrpc
1396          */
1397         if (cld->cld_is_sptlrpc && rcl == 0) {
1398                 sptlrpc_conf_log_update_end(cld->cld_logname);
1399                 class_notify_sptlrpc_conf(cld->cld_logname,
1400                                           strlen(cld->cld_logname) -
1401                                           strlen("-sptlrpc"));
1402         }
1403
1404         /* Now drop the lock so MGS can revoke it */
1405         if (!rcl) {
1406                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1407                                  LCK_CR, &lockh);
1408                 if (rcl)
1409                         CERROR("Can't drop cfg lock: %d\n", rcl);
1410         }
1411
1412         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1413                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1414
1415         cfs_up(&llog_process_lock);
1416
1417         RETURN(rc);
1418 }
1419
1420 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1421 {
1422         struct lustre_cfg *lcfg = buf;
1423         int cmd;
1424         int rc = 0;
1425         ENTRY;
1426
1427         switch(cmd = lcfg->lcfg_command) {
1428         case LCFG_LOV_ADD_OBD: {
1429                 /* Add any new target, not just osts */
1430                 struct mgs_target_info *mti;
1431
1432                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1433                     sizeof(struct mgs_target_info))
1434                         GOTO(out, rc = -EINVAL);
1435
1436                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1437                 CDEBUG(D_MGC, "add_target %s %#x\n",
1438                        mti->mti_svname, mti->mti_flags);
1439                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1440                 break;
1441         }
1442         case LCFG_LOV_DEL_OBD:
1443                 /* Remove target from the fs? */
1444                 /* FIXME */
1445                 CERROR("lov_del_obd unimplemented\n");
1446                 rc = -ENOSYS;
1447                 break;
1448         case LCFG_SPTLRPC_CONF: {
1449                 rc = sptlrpc_process_config(lcfg);
1450                 break;
1451         }
1452         case LCFG_LOG_START: {
1453                 struct config_llog_data *cld;
1454                 struct config_llog_instance *cfg;
1455                 struct super_block *sb;
1456                 char *logname = lustre_cfg_string(lcfg, 1);
1457                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1458                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1459
1460                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1461                        cfg->cfg_last_idx);
1462
1463                 /* We're only called through here on the initial mount */
1464                 rc = config_log_add(obd, logname, cfg, sb);
1465                 if (rc)
1466                         break;
1467                 cld = config_log_find(logname, cfg);
1468                 if (IS_ERR(cld)) {
1469                         rc = PTR_ERR(cld);
1470                         break;
1471                 }
1472
1473                 /* COMPAT_146 */
1474                 /* FIXME only set this for old logs!  Right now this forces
1475                    us to always skip the "inside markers" check */
1476                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1477
1478                 rc = mgc_process_log(obd, cld);
1479                 config_log_put(cld);
1480
1481                 break;
1482         }
1483         case LCFG_LOG_END: {
1484                 struct config_llog_instance *cfg = NULL;
1485                 char *logname = lustre_cfg_string(lcfg, 1);
1486                 if (lcfg->lcfg_bufcount >= 2)
1487                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
1488                                 lcfg, 2);
1489                 rc = config_log_end(logname, cfg);
1490                 break;
1491         }
1492         default: {
1493                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1494                 GOTO(out, rc = -EINVAL);
1495
1496         }
1497         }
1498 out:
1499         RETURN(rc);
1500 }
1501
1502 struct obd_ops mgc_obd_ops = {
1503         .o_owner        = THIS_MODULE,
1504         .o_setup        = mgc_setup,
1505         .o_precleanup   = mgc_precleanup,
1506         .o_cleanup      = mgc_cleanup,
1507         .o_add_conn     = client_import_add_conn,
1508         .o_del_conn     = client_import_del_conn,
1509         .o_connect      = client_connect_import,
1510         .o_disconnect   = client_disconnect_export,
1511         //.o_enqueue      = mgc_enqueue,
1512         .o_cancel       = mgc_cancel,
1513         //.o_iocontrol    = mgc_iocontrol,
1514         .o_set_info_async = mgc_set_info_async,
1515         .o_import_event = mgc_import_event,
1516         .o_llog_init    = mgc_llog_init,
1517         .o_llog_finish  = mgc_llog_finish,
1518         .o_process_config = mgc_process_config,
1519 };
1520
1521 int __init mgc_init(void)
1522 {
1523         return class_register_type(&mgc_obd_ops, NULL, NULL,
1524                                    LUSTRE_MGC_NAME, NULL);
1525 }
1526
1527 #ifdef __KERNEL__
1528 static void /*__exit*/ mgc_exit(void)
1529 {
1530         class_unregister_type(LUSTRE_MGC_NAME);
1531 }
1532
1533 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1534 MODULE_DESCRIPTION("Lustre Management Client");
1535 MODULE_LICENSE("GPL");
1536
1537 module_init(mgc_init);
1538 module_exit(mgc_exit);
1539 #endif