Pileus Git - ~andy/linux/blob - drivers/staging/lustre/lustre/mgc/mgc_request.c

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