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 spinlock_t       config_list_lock = 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         atomic_inc(&cld->cld_refcount);
 119         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 120                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                atomic_read(&cld->cld_refcount));
 132         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 133
 134         /* spinlock to make sure no item with 0 refcount in the list */
 135         spin_lock(&config_list_lock);
 136         if (unlikely(atomic_dec_and_test(&cld->cld_refcount))) {
 137                 list_del(&cld->cld_list_chain);
 138                 spin_unlock(&config_list_lock);
 139
 140                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 141
 142                 if (cld->cld_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                 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         spin_lock(&config_list_lock);
 180         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         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         atomic_inc(&cld->cld_refcount);
 194         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         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         spin_lock(&config_list_lock);
 247         list_add(&cld->cld_list_chain, &config_llog_list);
 248         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 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         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                 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         up(&llog_process_lock);
 344
 345         spin_lock(&config_list_lock);
 346         cld_sptlrpc = cld->cld_sptlrpc;
 347         cld->cld_sptlrpc = NULL;
 348         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(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         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         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 * 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         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                 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 * HZ + (ll_rand() & 0xff) * (HZ / 100),
 421                                       NULL, NULL);
 422                 l_wait_event(rq_waitq, rq_state & RQ_STOP, &lwi_now);
 423
 424                 /*
 425                  * iterate & processing through the list. for each cld, process
 426                  * its depending sptlrpc cld firstly (if any) and then itself.
 427                  *
 428                  * it's guaranteed any item in the list must have
 429                  * reference > 0; and if cld_lostlock is set, at
 430                  * least one reference is taken by the previous enqueue.
 431                  *
 432                  * Note: releasing a cld might lead to itself and its depended
 433                  * sptlrpc cld be unlinked from the list. to safely iterate
 434                  * we need to take a reference on next cld before processing.
 435                  */
 436                 cld_prev = NULL;
 437
 438                 spin_lock(&config_list_lock);
 439                 list_for_each_entry_safe(cld, cld_next, &config_llog_list,
 440                                          cld_list_chain) {
 441                         if (cld->cld_list_chain.next != &config_llog_list)
 442                                 atomic_inc(&cld_next->cld_refcount);
 443
 444                         if (cld->cld_lostlock) {
 445                                 if (cld->cld_sptlrpc &&
 446                                     cld->cld_sptlrpc->cld_lostlock) {
 447                                         cld->cld_sptlrpc->cld_lostlock = 0;
 448
 449                                         spin_unlock(&config_list_lock);
 450                                         do_requeue(cld->cld_sptlrpc);
 451                                         spin_lock(&config_list_lock);
 452                                         LASSERT(cld->cld_lostlock);
 453                                 }
 454
 455                                 cld->cld_lostlock = 0;
 456
 457                                 spin_unlock(&config_list_lock);
 458                                 do_requeue(cld);
 459                                 spin_lock(&config_list_lock);
 460                         }
 461
 462
 463                         if (cld_prev) {
 464                                 spin_unlock(&config_list_lock);
 465                                 config_log_put(cld_prev);
 466                                 spin_lock(&config_list_lock);
 467                         }
 468
 469                         cld_prev = cld_next;
 470                 }
 471                 spin_unlock(&config_list_lock);
 472
 473                 /* Wait a bit to see if anyone else needs a requeue */
 474                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 475                              &lwi_later);
 476                 spin_lock(&config_list_lock);
 477         }
 478         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 479         rq_state &= ~RQ_RUNNING;
 480         spin_unlock(&config_list_lock);
 481
 482         CDEBUG(D_MGC, "Ending requeue thread\n");
 483         RETURN(rc);
 484 }
 485
 486 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 487    We are responsible for dropping the config log reference from here on out. */
 488 static int mgc_requeue_add(struct config_llog_data *cld, int later)
 489 {
 490         int rc = 0;
 491
 492         CDEBUG(D_INFO, "log %s: requeue (l=%d r=%d sp=%d st=%x)\n",
 493                cld->cld_logname, later, atomic_read(&cld->cld_refcount),
 494                cld->cld_stopping, rq_state);
 495         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 496
 497         /* Hold lock for rq_state */
 498         spin_lock(&config_list_lock);
 499
 500         if (cld->cld_stopping || (rq_state & RQ_STOP)) {
 501                 cld->cld_lostlock = 0;
 502                 spin_unlock(&config_list_lock);
 503                 config_log_put(cld);
 504                 RETURN(0);
 505         }
 506
 507         cld->cld_lostlock = 1;
 508
 509         if (!(rq_state & RQ_RUNNING)) {
 510                 LASSERT(rq_state == 0);
 511                 rq_state = RQ_RUNNING | (later ? RQ_LATER : RQ_NOW);
 512                 spin_unlock(&config_list_lock);
 513                 rc = cfs_kernel_thread(mgc_requeue_thread, 0,
 514                                        CLONE_VM | CLONE_FILES);
 515                 if (rc < 0) {
 516                         CERROR("log %s: cannot start requeue thread (%d),"
 517                                "no more log updates!\n", cld->cld_logname, rc);
 518                         /* Drop the ref, since the rq thread won't */
 519                         cld->cld_lostlock = 0;
 520                         config_log_put(cld);
 521                         rq_state = 0;
 522                         RETURN(rc);
 523                 }
 524         } else {
 525                 rq_state |= later ? RQ_LATER : RQ_NOW;
 526                 spin_unlock(&config_list_lock);
 527                 cfs_waitq_signal(&rq_waitq);
 528         }
 529
 530         RETURN(0);
 531 }
 532
 533 /********************** class fns **********************/
 534
 535 static int mgc_fs_setup(struct obd_device *obd, struct super_block *sb,
 536                         struct vfsmount *mnt)
 537 {
 538         struct lvfs_run_ctxt saved;
 539         struct lustre_sb_info *lsi = s2lsi(sb);
 540         struct client_obd *cli = &obd->u.cli;
 541         struct dentry *dentry;
 542         char *label;
 543         int err = 0;
 544         ENTRY;
 545
 546         LASSERT(lsi);
 547         LASSERT(lsi->lsi_srv_mnt == mnt);
 548
 549         /* The mgc fs exclusion sem. Only one fs can be setup at a time. */
 550         down(&cli->cl_mgc_sem);
 551
 552         cleanup_group_info();
 553
 554         obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
 555         if (IS_ERR(obd->obd_fsops)) {
 556                 up(&cli->cl_mgc_sem);
 557                 CERROR("No fstype %s rc=%ld\n", MT_STR(lsi->lsi_ldd),
 558                        PTR_ERR(obd->obd_fsops));
 559                 RETURN(PTR_ERR(obd->obd_fsops));
 560         }
 561
 562         cli->cl_mgc_vfsmnt = mnt;
 563         fsfilt_setup(obd, mnt->mnt_sb);
 564
 565         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 566         obd->obd_lvfs_ctxt.pwdmnt = mnt;
 567         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
 568         obd->obd_lvfs_ctxt.fs = get_ds();
 569
 570         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 571         dentry = lookup_one_len(MOUNT_CONFIGS_DIR, cfs_fs_pwd(current->fs),
 572                                 strlen(MOUNT_CONFIGS_DIR));
 573         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 574         if (IS_ERR(dentry)) {
 575                 err = PTR_ERR(dentry);
 576                 CERROR("cannot lookup %s directory: rc = %d\n",
 577                        MOUNT_CONFIGS_DIR, err);
 578                 GOTO(err_ops, err);
 579         }
 580         cli->cl_mgc_configs_dir = dentry;
 581
 582         /* We take an obd ref to insure that we can't get to mgc_cleanup
 583            without calling mgc_fs_cleanup first. */
 584         class_incref(obd, "mgc_fs", obd);
 585
 586         label = fsfilt_get_label(obd, mnt->mnt_sb);
 587         if (label)
 588                 CDEBUG(D_MGC, "MGC using disk labelled=%s\n", label);
 589
 590         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
 591         RETURN(0);
 592
 593 err_ops:
 594         fsfilt_put_ops(obd->obd_fsops);
 595         obd->obd_fsops = NULL;
 596         cli->cl_mgc_vfsmnt = NULL;
 597         up(&cli->cl_mgc_sem);
 598         RETURN(err);
 599 }
 600
 601 static int mgc_fs_cleanup(struct obd_device *obd)
 602 {
 603         struct client_obd *cli = &obd->u.cli;
 604         int rc = 0;
 605         ENTRY;
 606
 607         LASSERT(cli->cl_mgc_vfsmnt != NULL);
 608
 609         if (cli->cl_mgc_configs_dir != NULL) {
 610                 struct lvfs_run_ctxt saved;
 611                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 612                 l_dput(cli->cl_mgc_configs_dir);
 613                 cli->cl_mgc_configs_dir = NULL;
 614                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 615                 class_decref(obd, "mgc_fs", obd);
 616         }
 617
 618         cli->cl_mgc_vfsmnt = NULL;
 619         if (obd->obd_fsops)
 620                 fsfilt_put_ops(obd->obd_fsops);
 621
 622         up(&cli->cl_mgc_sem);
 623
 624         RETURN(rc);
 625 }
 626
 627 static atomic_t mgc_count = ATOMIC_INIT(0);
 628 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 629 {
 630         int rc = 0;
 631         ENTRY;
 632
 633         switch (stage) {
 634         case OBD_CLEANUP_EARLY:
 635                 break;
 636         case OBD_CLEANUP_EXPORTS:
 637                 if (atomic_dec_and_test(&mgc_count)) {
 638                         /* Kick the requeue waitq - cld's should all be
 639                            stopping */
 640                         spin_lock(&config_list_lock);
 641                         rq_state |= RQ_STOP;
 642                         spin_unlock(&config_list_lock);
 643                         cfs_waitq_signal(&rq_waitq);
 644                 }
 645                 rc = obd_llog_finish(obd, 0);
 646                 if (rc != 0)
 647                         CERROR("failed to cleanup llogging subsystems\n");
 648                 break;
 649         }
 650         RETURN(rc);
 651 }
 652
 653 static int mgc_cleanup(struct obd_device *obd)
 654 {
 655         struct client_obd *cli = &obd->u.cli;
 656         int rc;
 657         ENTRY;
 658
 659         LASSERT(cli->cl_mgc_vfsmnt == NULL);
 660
 661         /* COMPAT_146 - old config logs may have added profiles we don't
 662            know about */
 663         if (obd->obd_type->typ_refcnt <= 1)
 664                 /* Only for the last mgc */
 665                 class_del_profiles();
 666
 667         lprocfs_obd_cleanup(obd);
 668         ptlrpcd_decref();
 669
 670         rc = client_obd_cleanup(obd);
 671         RETURN(rc);
 672 }
 673
 674 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 675 {
 676         struct lprocfs_static_vars lvars;
 677         int rc;
 678         ENTRY;
 679
 680         ptlrpcd_addref();
 681
 682         rc = client_obd_setup(obd, lcfg);
 683         if (rc)
 684                 GOTO(err_decref, rc);
 685
 686         rc = obd_llog_init(obd, &obd->obd_olg, obd, NULL);
 687         if (rc) {
 688                 CERROR("failed to setup llogging subsystems\n");
 689                 GOTO(err_cleanup, rc);
 690         }
 691
 692         lprocfs_mgc_init_vars(&lvars);
 693         lprocfs_obd_setup(obd, lvars.obd_vars);
 694         sptlrpc_lprocfs_cliobd_attach(obd);
 695
 696         spin_lock(&config_list_lock);
 697         atomic_inc(&mgc_count);
 698         if (atomic_read(&mgc_count) == 1) {
 699                 rq_state &= ~RQ_STOP;
 700                 cfs_waitq_init(&rq_waitq);
 701         }
 702         spin_unlock(&config_list_lock);
 703
 704         RETURN(rc);
 705
 706 err_cleanup:
 707         client_obd_cleanup(obd);
 708 err_decref:
 709         ptlrpcd_decref();
 710         RETURN(rc);
 711 }
 712
 713 /* based on ll_mdc_blocking_ast */
 714 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 715                             void *data, int flag)
 716 {
 717         struct lustre_handle lockh;
 718         struct config_llog_data *cld = (struct config_llog_data *)data;
 719         int rc = 0;
 720         ENTRY;
 721
 722         switch (flag) {
 723         case LDLM_CB_BLOCKING:
 724                 /* mgs wants the lock, give it up... */
 725                 LDLM_DEBUG(lock, "MGC blocking CB");
 726                 ldlm_lock2handle(lock, &lockh);
 727                 rc = ldlm_cli_cancel(&lockh);
 728                 break;
 729         case LDLM_CB_CANCELING: {
 730                 /* We've given up the lock, prepare ourselves to update. */
 731                 LDLM_DEBUG(lock, "MGC cancel CB");
 732
 733                 CDEBUG(D_MGC, "Lock res "LPX64" (%.8s)\n",
 734                        lock->l_resource->lr_name.name[0],
 735                        (char *)&lock->l_resource->lr_name.name[0]);
 736
 737                 if (!cld) {
 738                         CERROR("missing data, won't requeue\n");
 739                         break;
 740                 }
 741                 /* Are we done with this log? */
 742                 if (cld->cld_stopping) {
 743                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
 744                                cld->cld_logname);
 745                         config_log_put(cld);
 746                         break;
 747                 }
 748                 /* Make sure not to re-enqueue when the mgc is stopping
 749                    (we get called from client_disconnect_export) */
 750                 if (!lock->l_conn_export ||
 751                     !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
 752                         CDEBUG(D_MGC, "log %s: disconnecting, won't requeue\n",
 753                                cld->cld_logname);
 754                         config_log_put(cld);
 755                         break;
 756                 }
 757                 /* Did we fail to get the lock? */
 758                 if (lock->l_req_mode != lock->l_granted_mode &&
 759                     !cld->cld_is_sptlrpc) {
 760                         CDEBUG(D_MGC, "log %s: original grant failed, will "
 761                                "requeue later\n", cld->cld_logname);
 762                         /* Try to re-enqueue later */
 763                         rc = mgc_requeue_add(cld, 1);
 764                         break;
 765                 }
 766                 /* Re-enqueue now */
 767                 rc = mgc_requeue_add(cld, 0);
 768                 break;
 769         }
 770         default:
 771                 LBUG();
 772         }
 773
 774
 775         if (rc) {
 776                 CERROR("%s CB failed %d:\n", flag == LDLM_CB_BLOCKING ?
 777                        "blocking" : "cancel", rc);
 778                 LDLM_ERROR(lock, "MGC ast");
 779         }
 780         RETURN(rc);
 781 }
 782
 783 /* Send parameter to MGS*/
 784 static int mgc_set_mgs_param(struct obd_export *exp,
 785                              struct mgs_send_param *msp)
 786 {
 787         struct ptlrpc_request *req;
 788         struct mgs_send_param *req_msp, *rep_msp;
 789         int rc;
 790         ENTRY;
 791
 792         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 793                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
 794                                         MGS_SET_INFO);
 795         if (!req)
 796                 RETURN(-ENOMEM);
 797
 798         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 799         if (!req_msp) {
 800                 ptlrpc_req_finished(req);
 801                 RETURN(-ENOMEM);
 802         }
 803
 804         memcpy(req_msp, msp, sizeof(*req_msp));
 805         ptlrpc_request_set_replen(req);
 806
 807         rc = ptlrpc_queue_wait(req);
 808         if (!rc) {
 809                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 810                 memcpy(msp, rep_msp, sizeof(*rep_msp));
 811         }
 812
 813         ptlrpc_req_finished(req);
 814
 815         RETURN(rc);
 816 }
 817
 818 /* Take a config lock so we can get cancel notifications */
 819 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
 820                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
 821                        int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
 822                        void *data, __u32 lvb_len, void *lvb_swabber,
 823                        struct lustre_handle *lockh)
 824 {
 825         struct config_llog_data *cld = (struct config_llog_data *)data;
 826         struct ldlm_enqueue_info einfo = { type, mode, mgc_blocking_ast,
 827                          ldlm_completion_ast, NULL, NULL, data};
 828
 829         int rc;
 830         ENTRY;
 831
 832         CDEBUG(D_MGC, "Enqueue for %s (res "LPX64")\n", cld->cld_logname,
 833                cld->cld_resid.name[0]);
 834
 835         /* We can only drop this config log ref when we drop the lock */
 836         if (config_log_get(cld))
 837                 RETURN(ELDLM_LOCK_ABORTED);
 838
 839         /* We need a callback for every lockholder, so don't try to
 840            ldlm_lock_match (see rev 1.1.2.11.2.47) */
 841
 842         rc = ldlm_cli_enqueue(exp, NULL, &einfo, &cld->cld_resid, NULL, flags,
 843                               NULL, 0, lockh, 0);
 844         /* A failed enqueue should still call the mgc_blocking_ast,
 845            where it will be requeued if needed ("grant failed"). */
 846
 847         RETURN(rc);
 848 }
 849
 850 static int mgc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
 851                       __u32 mode, struct lustre_handle *lockh)
 852 {
 853         ENTRY;
 854
 855         ldlm_lock_decref(lockh, mode);
 856
 857         RETURN(0);
 858 }
 859
 860 #if 0
 861 static int mgc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
 862                          void *karg, void *uarg)
 863 {
 864         struct obd_device *obd = exp->exp_obd;
 865         struct obd_ioctl_data *data = karg;
 866         struct llog_ctxt *ctxt;
 867         struct lvfs_run_ctxt saved;
 868         int rc;
 869         ENTRY;
 870
 871         if (!try_module_get(THIS_MODULE)) {
 872                 CERROR("Can't get module. Is it alive?");
 873                 return -EINVAL;
 874         }
 875         switch (cmd) {
 876         /* REPLicator context */
 877         case OBD_IOC_PARSE: {
 878                 CERROR("MGC parsing llog %s\n", data->ioc_inlbuf1);
 879                 ctxt = llog_get_context(exp->exp_obd, LLOG_CONFIG_REPL_CTXT);
 880                 rc = class_config_parse_llog(ctxt, data->ioc_inlbuf1, NULL);
 881                 GOTO(out, rc);
 882         }
 883 #ifdef __KERNEL__
 884         case OBD_IOC_LLOG_INFO:
 885         case OBD_IOC_LLOG_PRINT: {
 886                 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 887                 rc = llog_ioctl(ctxt, cmd, data);
 888
 889                 GOTO(out, rc);
 890         }
 891 #endif
 892         /* ORIGinator context */
 893         case OBD_IOC_DUMP_LOG: {
 894                 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
 895                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 896                 rc = class_config_dump_llog(ctxt, data->ioc_inlbuf1, NULL);
 897                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 898                 if (rc)
 899                         RETURN(rc);
 900
 901                 GOTO(out, rc);
 902         }
 903         default:
 904                 CERROR("mgc_ioctl(): unrecognised ioctl %#x\n", cmd);
 905                 GOTO(out, rc = -ENOTTY);
 906         }
 907 out:
 908         module_put(THIS_MODULE);
 909
 910         return rc;
 911 }
 912 #endif
 913
 914 /* Send target_reg message to MGS */
 915 static int mgc_target_register(struct obd_export *exp,
 916                                struct mgs_target_info *mti)
 917 {
 918         struct ptlrpc_request  *req;
 919         struct mgs_target_info *req_mti, *rep_mti;
 920         int                     rc;
 921         ENTRY;
 922
 923         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 924                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
 925                                         MGS_TARGET_REG);
 926         if (req == NULL)
 927                 RETURN(-ENOMEM);
 928
 929         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
 930         if (!req_mti) {
 931                 ptlrpc_req_finished(req);
 932                 RETURN(-ENOMEM);
 933         }
 934
 935         memcpy(req_mti, mti, sizeof(*req_mti));
 936         ptlrpc_request_set_replen(req);
 937         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
 938
 939         rc = ptlrpc_queue_wait(req);
 940         if (!rc) {
 941                 rep_mti = req_capsule_server_get(&req->rq_pill,
 942                                                  &RMF_MGS_TARGET_INFO);
 943                 memcpy(mti, rep_mti, sizeof(*rep_mti));
 944                 CDEBUG(D_MGC, "register %s got index = %d\n",
 945                        mti->mti_svname, mti->mti_stripe_index);
 946         }
 947         ptlrpc_req_finished(req);
 948
 949         RETURN(rc);
 950 }
 951
 952 int mgc_set_info_async(struct obd_export *exp, obd_count keylen,
 953                        void *key, obd_count vallen, void *val,
 954                        struct ptlrpc_request_set *set)
 955 {
 956         struct obd_import *imp = class_exp2cliimp(exp);
 957         int rc = -EINVAL;
 958         ENTRY;
 959
 960         /* Try to "recover" the initial connection; i.e. retry */
 961         if (KEY_IS(KEY_INIT_RECOV)) {
 962                 if (vallen != sizeof(int))
 963                         RETURN(-EINVAL);
 964                 spin_lock(&imp->imp_lock);
 965                 imp->imp_initial_recov = *(int *)val;
 966                 spin_unlock(&imp->imp_lock);
 967                 CDEBUG(D_HA, "%s: set imp_initial_recov = %d\n",
 968                        exp->exp_obd->obd_name, imp->imp_initial_recov);
 969                 RETURN(0);
 970         }
 971         /* Turn off initial_recov after we try all backup servers once */
 972         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
 973                 int value;
 974                 if (vallen != sizeof(int))
 975                         RETURN(-EINVAL);
 976                 value = *(int *)val;
 977                 spin_lock(&imp->imp_lock);
 978                 imp->imp_initial_recov_bk = value > 0;
 979                 /* Even after the initial connection, give up all comms if
 980                    nobody answers the first time. */
 981                 imp->imp_recon_bk = 1;
 982                 spin_unlock(&imp->imp_lock);
 983                 CDEBUG(D_MGC, "InitRecov %s %d/%d:d%d:i%d:r%d:or%d:%s\n",
 984                        imp->imp_obd->obd_name, value, imp->imp_initial_recov,
 985                        imp->imp_deactive, imp->imp_invalid,
 986                        imp->imp_replayable, imp->imp_obd->obd_replayable,
 987                        ptlrpc_import_state_name(imp->imp_state));
 988                 /* Resurrect if we previously died */
 989                 if (imp->imp_invalid || value > 1)
 990                         ptlrpc_reconnect_import(imp);
 991                 RETURN(0);
 992         }
 993         /* FIXME move this to mgc_process_config */
 994         if (KEY_IS(KEY_REGISTER_TARGET)) {
 995                 struct mgs_target_info *mti;
 996                 if (vallen != sizeof(struct mgs_target_info))
 997                         RETURN(-EINVAL);
 998                 mti = (struct mgs_target_info *)val;
 999                 CDEBUG(D_MGC, "register_target %s %#x\n",
1000                        mti->mti_svname, mti->mti_flags);
1001                 rc =  mgc_target_register(exp, mti);
1002                 RETURN(rc);
1003         }
1004         if (KEY_IS(KEY_SET_FS)) {
1005                 struct super_block *sb = (struct super_block *)val;
1006                 struct lustre_sb_info *lsi;
1007                 if (vallen != sizeof(struct super_block))
1008                         RETURN(-EINVAL);
1009                 lsi = s2lsi(sb);
1010                 rc = mgc_fs_setup(exp->exp_obd, sb, lsi->lsi_srv_mnt);
1011                 if (rc) {
1012                         CERROR("set_fs got %d\n", rc);
1013                 }
1014                 RETURN(rc);
1015         }
1016         if (KEY_IS(KEY_CLEAR_FS)) {
1017                 if (vallen != 0)
1018                         RETURN(-EINVAL);
1019                 rc = mgc_fs_cleanup(exp->exp_obd);
1020                 if (rc) {
1021                         CERROR("clear_fs got %d\n", rc);
1022                 }
1023                 RETURN(rc);
1024         }
1025         if (KEY_IS(KEY_SET_INFO)) {
1026                 struct mgs_send_param *msp;
1027
1028                 msp = (struct mgs_send_param *)val;
1029                 rc =  mgc_set_mgs_param(exp, msp);
1030                 RETURN(rc);
1031         }
1032         if (KEY_IS(KEY_MGSSEC)) {
1033                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1034                 struct sptlrpc_flavor  flvr;
1035
1036                 /*
1037                  * empty string means using current flavor, if which haven't
1038                  * been set yet, set it as null.
1039                  *
1040                  * if flavor has been set previously, check the asking flavor
1041                  * must match the existing one.
1042                  */
1043                 if (vallen == 0) {
1044                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1045                                 RETURN(0);
1046                         val = "null";
1047                         vallen = 4;
1048                 }
1049
1050                 rc = sptlrpc_parse_flavor(val, &flvr);
1051                 if (rc) {
1052                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1053                                (char *) val);
1054                         RETURN(rc);
1055                 }
1056
1057                 /*
1058                  * caller already hold a mutex
1059                  */
1060                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1061                         cli->cl_flvr_mgc = flvr;
1062                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1063                                   sizeof(flvr)) != 0) {
1064                         char    str[20];
1065
1066                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1067                                             str, sizeof(str));
1068                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1069                                        "currently %s is in use\n",
1070                                        (char *) val, str);
1071                         rc = -EPERM;
1072                 }
1073                 RETURN(rc);
1074         }
1075
1076         RETURN(rc);
1077 }
1078
1079 static int mgc_import_event(struct obd_device *obd,
1080                             struct obd_import *imp,
1081                             enum obd_import_event event)
1082 {
1083         int rc = 0;
1084
1085         LASSERT(imp->imp_obd == obd);
1086         CDEBUG(D_MGC, "import event %#x\n", event);
1087
1088         switch (event) {
1089         case IMP_EVENT_DISCON:
1090                 /* MGC imports should not wait for recovery */
1091                 break;
1092         case IMP_EVENT_INACTIVE:
1093                 break;
1094         case IMP_EVENT_INVALIDATE: {
1095                 struct ldlm_namespace *ns = obd->obd_namespace;
1096                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1097                 break;
1098         }
1099         case IMP_EVENT_ACTIVE:
1100                 LCONSOLE_WARN("%s: Reactivating import\n", obd->obd_name);
1101                 /* Clearing obd_no_recov allows us to continue pinging */
1102                 obd->obd_no_recov = 0;
1103                 break;
1104         case IMP_EVENT_OCD:
1105                 break;
1106         default:
1107                 CERROR("Unknown import event %#x\n", event);
1108                 LBUG();
1109         }
1110         RETURN(rc);
1111 }
1112
1113 static int mgc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
1114                          struct obd_device *tgt, int *index)
1115 {
1116         struct llog_ctxt *ctxt;
1117         int rc;
1118         ENTRY;
1119
1120         LASSERT(olg == &obd->obd_olg);
1121
1122         rc = llog_setup(obd, olg, LLOG_CONFIG_ORIG_CTXT, tgt, 0, NULL,
1123                         &llog_lvfs_ops);
1124         if (rc)
1125                 RETURN(rc);
1126
1127         rc = llog_setup(obd, olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
1128                         &llog_client_ops);
1129         if (rc == 0) {
1130                 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1131                 if (!ctxt) {
1132                         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1133                         if (ctxt)
1134                                 llog_cleanup(ctxt);
1135                         RETURN(-ENODEV);
1136                 }
1137                 llog_initiator_connect(ctxt);
1138                 llog_ctxt_put(ctxt);
1139         } else {
1140                 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1141                 if (ctxt)
1142                         llog_cleanup(ctxt);
1143         }
1144
1145         RETURN(rc);
1146 }
1147
1148 static int mgc_llog_finish(struct obd_device *obd, int count)
1149 {
1150         struct llog_ctxt *ctxt;
1151         int rc = 0, rc2 = 0;
1152         ENTRY;
1153
1154         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
1155         if (ctxt)
1156                 rc = llog_cleanup(ctxt);
1157
1158         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
1159         if (ctxt)
1160                 rc2 = llog_cleanup(ctxt);
1161
1162         if (!rc)
1163                 rc = rc2;
1164
1165         RETURN(rc);
1166 }
1167
1168 /* identical to mgs_log_is_empty */
1169 static int mgc_llog_is_empty(struct obd_device *obd, struct llog_ctxt *ctxt,
1170                             char *name)
1171 {
1172         struct lvfs_run_ctxt saved;
1173         struct llog_handle *llh;
1174         int rc = 0;
1175
1176         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1177         rc = llog_create(ctxt, &llh, NULL, name);
1178         if (rc == 0) {
1179                 llog_init_handle(llh, LLOG_F_IS_PLAIN, NULL);
1180                 rc = llog_get_size(llh);
1181                 llog_close(llh);
1182         }
1183         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1184         /* header is record 1 */
1185         return(rc <= 1);
1186 }
1187
1188 static int mgc_copy_handler(struct llog_handle *llh, struct llog_rec_hdr *rec,
1189                             void *data)
1190 {
1191         struct llog_rec_hdr local_rec = *rec;
1192         struct llog_handle *local_llh = (struct llog_handle *)data;
1193         char *cfg_buf = (char*) (rec + 1);
1194         struct lustre_cfg *lcfg;
1195         int rc = 0;
1196         ENTRY;
1197
1198         /* Append all records */
1199         local_rec.lrh_len -= sizeof(*rec) + sizeof(struct llog_rec_tail);
1200         rc = llog_write_rec(local_llh, &local_rec, NULL, 0,
1201                             (void *)cfg_buf, -1);
1202
1203         lcfg = (struct lustre_cfg *)cfg_buf;
1204         CDEBUG(D_INFO, "idx=%d, rc=%d, len=%d, cmd %x %s %s\n",
1205                rec->lrh_index, rc, rec->lrh_len, lcfg->lcfg_command,
1206                lustre_cfg_string(lcfg, 0), lustre_cfg_string(lcfg, 1));
1207
1208         RETURN(rc);
1209 }
1210
1211 /* Copy a remote log locally */
1212 static int mgc_copy_llog(struct obd_device *obd, struct llog_ctxt *rctxt,
1213                          struct llog_ctxt *lctxt, char *logname)
1214 {
1215         struct llog_handle *local_llh, *remote_llh;
1216         struct obd_uuid *uuid;
1217         char *temp_log;
1218         int rc, rc2;
1219         ENTRY;
1220
1221         /* Write new log to a temp name, then vfs_rename over logname
1222            upon successful completion. */
1223
1224         OBD_ALLOC(temp_log, strlen(logname) + 1);
1225         if (!temp_log)
1226                 RETURN(-ENOMEM);
1227         sprintf(temp_log, "%sT", logname);
1228
1229         /* Make sure there's no old temp log */
1230         rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1231         if (rc)
1232                 GOTO(out, rc);
1233         rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, NULL);
1234         if (rc)
1235                 GOTO(out, rc);
1236         rc = llog_destroy(local_llh);
1237         llog_free_handle(local_llh);
1238         if (rc)
1239                 GOTO(out, rc);
1240
1241         /* open local log */
1242         rc = llog_create(lctxt, &local_llh, NULL, temp_log);
1243         if (rc)
1244                 GOTO(out, rc);
1245
1246         /* set the log header uuid for fun */
1247         OBD_ALLOC_PTR(uuid);
1248         obd_str2uuid(uuid, logname);
1249         rc = llog_init_handle(local_llh, LLOG_F_IS_PLAIN, uuid);
1250         OBD_FREE_PTR(uuid);
1251         if (rc)
1252                 GOTO(out_closel, rc);
1253
1254         /* open remote log */
1255         rc = llog_create(rctxt, &remote_llh, NULL, logname);
1256         if (rc)
1257                 GOTO(out_closel, rc);
1258         rc = llog_init_handle(remote_llh, LLOG_F_IS_PLAIN, NULL);
1259         if (rc)
1260                 GOTO(out_closer, rc);
1261
1262         /* Copy remote log */
1263         rc = llog_process(remote_llh, mgc_copy_handler,(void *)local_llh, NULL);
1264
1265 out_closer:
1266         rc2 = llog_close(remote_llh);
1267         if (!rc)
1268                 rc = rc2;
1269 out_closel:
1270         rc2 = llog_close(local_llh);
1271         if (!rc)
1272                 rc = rc2;
1273
1274         /* We've copied the remote log to the local temp log, now
1275            replace the old local log with the temp log. */
1276         if (!rc) {
1277                 struct client_obd *cli = &obd->u.cli;
1278                 LASSERT(cli);
1279                 LASSERT(cli->cl_mgc_configs_dir);
1280                 rc = lustre_rename(cli->cl_mgc_configs_dir, cli->cl_mgc_vfsmnt,
1281                                    temp_log, logname);
1282         }
1283         CDEBUG(D_MGC, "Copied remote log %s (%d)\n", logname, rc);
1284 out:
1285         if (rc)
1286                 CERROR("Failed to copy remote log %s (%d)\n", logname, rc);
1287         OBD_FREE(temp_log, strlen(logname) + 1);
1288         RETURN(rc);
1289 }
1290
1291 /* Get a config log from the MGS and process it.
1292    This func is called for both clients and servers. */
1293 int mgc_process_log(struct obd_device *mgc,
1294                     struct config_llog_data *cld)
1295 {
1296         struct llog_ctxt *ctxt, *lctxt;
1297         struct lustre_handle lockh;
1298         struct client_obd *cli = &mgc->u.cli;
1299         struct lvfs_run_ctxt saved;
1300         struct lustre_sb_info *lsi = NULL;
1301         int rc = 0, rcl, flags = 0, must_pop = 0;
1302         ENTRY;
1303
1304         LASSERT(cld);
1305
1306         /* I don't want multiple processes running process_log at once --
1307            sounds like badness.  It actually might be fine, as long as
1308            we're not trying to update from the same log
1309            simultaneously (in which case we should use a per-log sem.) */
1310         down(&llog_process_lock);
1311
1312         if (cld->cld_stopping) {
1313                 up(&llog_process_lock);
1314                 RETURN(0);
1315         }
1316
1317         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1318
1319         if (cld->cld_cfg.cfg_sb)
1320                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1321
1322         CDEBUG(D_MGC, "Process log %s:%s from %d\n", cld->cld_logname,
1323                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1324
1325         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1326         if (!ctxt) {
1327                 CERROR("missing llog context\n");
1328                 up(&llog_process_lock);
1329                 RETURN(-EINVAL);
1330         }
1331
1332         /* Get the cfg lock on the llog */
1333         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1334                           LCK_CR, &flags, NULL, NULL, NULL,
1335                           cld, 0, NULL, &lockh);
1336         if (rcl)
1337                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1338
1339         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1340
1341         /*
1342          * local copy of sptlrpc log is controlled elsewhere, don't try to
1343          * read it up here.
1344          */
1345         if (rcl && cld->cld_is_sptlrpc)
1346                 goto out_pop;
1347
1348         /* Copy the setup log locally if we can. Don't mess around if we're
1349            running an MGS though (logs are already local). */
1350         if (lctxt && lsi && (lsi->lsi_flags & LSI_SERVER) &&
1351             (lsi->lsi_srv_mnt == cli->cl_mgc_vfsmnt) &&
1352             !IS_MGS(lsi->lsi_ldd)) {
1353                 push_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1354                 must_pop++;
1355                 if (rcl == 0)
1356                         /* Only try to copy log if we have the lock. */
1357                         rc = mgc_copy_llog(mgc, ctxt, lctxt, cld->cld_logname);
1358                 if (rcl || rc) {
1359                         if (mgc_llog_is_empty(mgc, lctxt, cld->cld_logname)) {
1360                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1361                                                    "log %s and no local copy."
1362                                                    "\n", cld->cld_logname);
1363                                 GOTO(out_pop, rc = -ENOTCONN);
1364                         }
1365                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1366                                "copy for now, will try to update later.\n",
1367                                cld->cld_logname);
1368                 }
1369                 /* Now, whether we copied or not, start using the local llog.
1370                    If we failed to copy, we'll start using whatever the old
1371                    log has. */
1372                 llog_ctxt_put(ctxt);
1373                 ctxt = lctxt;
1374         }
1375
1376         if (cld->cld_is_sptlrpc)
1377                 sptlrpc_conf_log_update_begin(cld->cld_logname);
1378
1379         /* logname and instance info should be the same, so use our
1380            copy of the instance for the update.  The cfg_last_idx will
1381            be updated here. */
1382         rc = class_config_parse_llog(ctxt, cld->cld_logname, &cld->cld_cfg);
1383 out_pop:
1384         llog_ctxt_put(ctxt);
1385         if (ctxt != lctxt)
1386                 llog_ctxt_put(lctxt);
1387         if (must_pop)
1388                 pop_ctxt(&saved, &mgc->obd_lvfs_ctxt, NULL);
1389
1390         /*
1391          * update settings on existing OBDs. doing it inside
1392          * of llog_process_lock so no device is attaching/detaching
1393          * in parallel.
1394          * the logname must be <fsname>-sptlrpc
1395          */
1396         if (cld->cld_is_sptlrpc && rcl == 0) {
1397                 sptlrpc_conf_log_update_end(cld->cld_logname);
1398                 class_notify_sptlrpc_conf(cld->cld_logname,
1399                                           strlen(cld->cld_logname) -
1400                                           strlen("-sptlrpc"));
1401         }
1402
1403         /* Now drop the lock so MGS can revoke it */
1404         if (!rcl) {
1405                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, NULL,
1406                                  LCK_CR, &lockh);
1407                 if (rcl)
1408                         CERROR("Can't drop cfg lock: %d\n", rcl);
1409         }
1410
1411         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1412                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1413
1414         up(&llog_process_lock);
1415
1416         RETURN(rc);
1417 }
1418
1419 static int mgc_process_config(struct obd_device *obd, obd_count len, void *buf)
1420 {
1421         struct lustre_cfg *lcfg = buf;
1422         int cmd;
1423         int rc = 0;
1424         ENTRY;
1425
1426         switch(cmd = lcfg->lcfg_command) {
1427         case LCFG_LOV_ADD_OBD: {
1428                 /* Add any new target, not just osts */
1429                 struct mgs_target_info *mti;
1430
1431                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1432                     sizeof(struct mgs_target_info))
1433                         GOTO(out, rc = -EINVAL);
1434
1435                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1436                 CDEBUG(D_MGC, "add_target %s %#x\n",
1437                        mti->mti_svname, mti->mti_flags);
1438                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1439                 break;
1440         }
1441         case LCFG_LOV_DEL_OBD:
1442                 /* Remove target from the fs? */
1443                 /* FIXME */
1444                 CERROR("lov_del_obd unimplemented\n");
1445                 rc = -ENOSYS;
1446                 break;
1447         case LCFG_SPTLRPC_CONF: {
1448                 rc = sptlrpc_process_config(lcfg);
1449                 break;
1450         }
1451         case LCFG_LOG_START: {
1452                 struct config_llog_data *cld;
1453                 struct config_llog_instance *cfg;
1454                 struct super_block *sb;
1455                 char *logname = lustre_cfg_string(lcfg, 1);
1456                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1457                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1458
1459                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1460                        cfg->cfg_last_idx);
1461
1462                 /* We're only called through here on the initial mount */
1463                 rc = config_log_add(obd, logname, cfg, sb);
1464                 if (rc)
1465                         break;
1466                 cld = config_log_find(logname, cfg);
1467                 if (IS_ERR(cld)) {
1468                         rc = PTR_ERR(cld);
1469                         break;
1470                 }
1471
1472                 /* COMPAT_146 */
1473                 /* FIXME only set this for old logs!  Right now this forces
1474                    us to always skip the "inside markers" check */
1475                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1476
1477                 rc = mgc_process_log(obd, cld);
1478                 config_log_put(cld);
1479
1480                 break;
1481         }
1482         case LCFG_LOG_END: {
1483                 struct config_llog_instance *cfg = NULL;
1484                 char *logname = lustre_cfg_string(lcfg, 1);
1485                 if (lcfg->lcfg_bufcount >= 2)
1486                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
1487                                 lcfg, 2);
1488                 rc = config_log_end(logname, cfg);
1489                 break;
1490         }
1491         default: {
1492                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1493                 GOTO(out, rc = -EINVAL);
1494
1495         }
1496         }
1497 out:
1498         RETURN(rc);
1499 }
1500
1501 struct obd_ops mgc_obd_ops = {
1502         .o_owner        = THIS_MODULE,
1503         .o_setup        = mgc_setup,
1504         .o_precleanup   = mgc_precleanup,
1505         .o_cleanup      = mgc_cleanup,
1506         .o_add_conn     = client_import_add_conn,
1507         .o_del_conn     = client_import_del_conn,
1508         .o_connect      = client_connect_import,
1509         .o_disconnect   = client_disconnect_export,
1510         //.o_enqueue      = mgc_enqueue,
1511         .o_cancel       = mgc_cancel,
1512         //.o_iocontrol    = mgc_iocontrol,
1513         .o_set_info_async = mgc_set_info_async,
1514         .o_import_event = mgc_import_event,
1515         .o_llog_init    = mgc_llog_init,
1516         .o_llog_finish  = mgc_llog_finish,
1517         .o_process_config = mgc_process_config,
1518 };
1519
1520 int __init mgc_init(void)
1521 {
1522         return class_register_type(&mgc_obd_ops, NULL, NULL,
1523                                    LUSTRE_MGC_NAME, NULL);
1524 }
1525
1526 #ifdef __KERNEL__
1527 static void /*__exit*/ mgc_exit(void)
1528 {
1529         class_unregister_type(LUSTRE_MGC_NAME);
1530 }
1531
1532 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1533 MODULE_DESCRIPTION("Lustre Management Client");
1534 MODULE_LICENSE("GPL");
1535
1536 module_init(mgc_init);
1537 module_exit(mgc_exit);
1538 #endif