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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/mgc/mgc_request.c
  33  *
  34  * Author: Nathan Rutman <nathan@clusterfs.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_MGC
  38 #define D_MGC D_CONFIG /*|D_WARNING*/
  39
  40 #include <linux/module.h>
  41
  42 #include <lprocfs_status.h>
  43 #include <lustre_dlm.h>
  44 #include <lustre_disk.h>
  45 #include <lustre_log.h>
  46 #include <lustre_swab.h>
  47 #include <obd_class.h>
  48
  49 #include "mgc_internal.h"
  50
  51 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
  52                           int type)
  53 {
  54         __u64 resname = 0;
  55
  56         if (len > sizeof(resname)) {
  57                 CERROR("name too long: %s\n", name);
  58                 return -EINVAL;
  59         }
  60         if (len <= 0) {
  61                 CERROR("missing name: %s\n", name);
  62                 return -EINVAL;
  63         }
  64         memcpy(&resname, name, len);
  65
  66         /* Always use the same endianness for the resid */
  67         memset(res_id, 0, sizeof(*res_id));
  68         res_id->name[0] = cpu_to_le64(resname);
  69         /* XXX: unfortunately, sptlprc and config llog share one lock */
  70         switch (type) {
  71         case CONFIG_T_CONFIG:
  72         case CONFIG_T_SPTLRPC:
  73                 resname = 0;
  74                 break;
  75         case CONFIG_T_RECOVER:
  76         case CONFIG_T_PARAMS:
  77                 resname = type;
  78                 break;
  79         default:
  80                 LBUG();
  81         }
  82         res_id->name[1] = cpu_to_le64(resname);
  83         CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
  84                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
  85         return 0;
  86 }
  87
  88 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
  89 {
  90         /* fsname is at most 8 chars long, maybe contain "-".
  91          * e.g. "lustre", "SUN-000"
  92          */
  93         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
  94 }
  95 EXPORT_SYMBOL(mgc_fsname2resid);
  96
  97 static int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
  98 {
  99         char *name_end;
 100         int len;
 101
 102         /* logname consists of "fsname-nodetype".
 103          * e.g. "lustre-MDT0001", "SUN-000-client"
 104          * there is an exception: llog "params"
 105          */
 106         name_end = strrchr(logname, '-');
 107         if (!name_end)
 108                 len = strlen(logname);
 109         else
 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  */
 130 static void config_log_put(struct config_llog_data *cld)
 131 {
 132         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
 133                atomic_read(&cld->cld_refcount));
 134         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 135
 136         /* spinlock to make sure no item with 0 refcount in the list */
 137         if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
 138                 list_del(&cld->cld_list_chain);
 139                 spin_unlock(&config_list_lock);
 140
 141                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
 142
 143                 if (cld->cld_recover)
 144                         config_log_put(cld->cld_recover);
 145                 if (cld->cld_params)
 146                         config_log_put(cld->cld_params);
 147                 if (cld->cld_sptlrpc)
 148                         config_log_put(cld->cld_sptlrpc);
 149                 if (cld_is_sptlrpc(cld))
 150                         sptlrpc_conf_log_stop(cld->cld_logname);
 151
 152                 class_export_put(cld->cld_mgcexp);
 153                 kfree(cld);
 154         }
 155 }
 156
 157 /* Find a config log by name */
 158 static
 159 struct config_llog_data *config_log_find(char *logname,
 160                                          struct config_llog_instance *cfg)
 161 {
 162         struct config_llog_data *cld;
 163         struct config_llog_data *found = NULL;
 164         void *instance;
 165
 166         LASSERT(logname);
 167
 168         instance = cfg ? cfg->cfg_instance : NULL;
 169         spin_lock(&config_list_lock);
 170         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 171                 /* check if instance equals */
 172                 if (instance != cld->cld_cfg.cfg_instance)
 173                         continue;
 174
 175                 /* instance may be NULL, should check name */
 176                 if (strcmp(logname, cld->cld_logname) == 0) {
 177                         found = cld;
 178                         config_log_get(found);
 179                         break;
 180                 }
 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 : NULL);
 198
 199         cld = kzalloc(sizeof(*cld) + strlen(logname) + 1, GFP_NOFS);
 200         if (!cld)
 201                 return ERR_PTR(-ENOMEM);
 202
 203         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
 204         if (rc) {
 205                 kfree(cld);
 206                 return ERR_PTR(rc);
 207         }
 208
 209         strcpy(cld->cld_logname, logname);
 210         if (cfg)
 211                 cld->cld_cfg = *cfg;
 212         else
 213                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
 214         mutex_init(&cld->cld_lock);
 215         cld->cld_cfg.cfg_last_idx = 0;
 216         cld->cld_cfg.cfg_flags = 0;
 217         cld->cld_cfg.cfg_sb = sb;
 218         cld->cld_type = type;
 219         atomic_set(&cld->cld_refcount, 1);
 220
 221         /* Keep the mgc around until we are done */
 222         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
 223
 224         if (cld_is_sptlrpc(cld)) {
 225                 sptlrpc_conf_log_start(logname);
 226                 cld->cld_cfg.cfg_obdname = obd->obd_name;
 227         }
 228
 229         spin_lock(&config_list_lock);
 230         list_add(&cld->cld_list_chain, &config_llog_list);
 231         spin_unlock(&config_list_lock);
 232
 233         if (cld_is_sptlrpc(cld)) {
 234                 rc = mgc_process_log(obd, cld);
 235                 if (rc && rc != -ENOENT)
 236                         CERROR("failed processing sptlrpc log: %d\n", rc);
 237         }
 238
 239         return cld;
 240 }
 241
 242 static struct config_llog_data *
 243 config_recover_log_add(struct obd_device *obd, char *fsname,
 244                        struct config_llog_instance *cfg,
 245                        struct super_block *sb)
 246 {
 247         struct config_llog_instance lcfg = *cfg;
 248         struct config_llog_data *cld;
 249         char logname[32];
 250
 251         /* we have to use different llog for clients and mdts for cmd
 252          * where only clients are notified if one of cmd server restarts
 253          */
 254         LASSERT(strlen(fsname) < sizeof(logname) / 2);
 255         strcpy(logname, fsname);
 256         LASSERT(lcfg.cfg_instance);
 257         strcat(logname, "-cliir");
 258
 259         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
 260         return cld;
 261 }
 262
 263 static struct config_llog_data *
 264 config_params_log_add(struct obd_device *obd,
 265                       struct config_llog_instance *cfg, struct super_block *sb)
 266 {
 267         struct config_llog_instance     lcfg = *cfg;
 268         struct config_llog_data         *cld;
 269
 270         lcfg.cfg_instance = sb;
 271
 272         cld = do_config_log_add(obd, PARAMS_FILENAME, CONFIG_T_PARAMS,
 273                                 &lcfg, sb);
 274
 275         return cld;
 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 struct config_llog_data *
 284 config_log_add(struct obd_device *obd, char *logname,
 285                struct config_llog_instance *cfg, 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         struct config_llog_data *params_cld;
 291         struct config_llog_data *recover_cld = NULL;
 292         char                    seclogname[32];
 293         char                    *ptr;
 294         int                     rc;
 295
 296         CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
 297
 298         /*
 299          * for each regular log, the depended sptlrpc log name is
 300          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
 301          */
 302         ptr = strrchr(logname, '-');
 303         if (!ptr || ptr - logname > 8) {
 304                 CERROR("logname %s is too long\n", logname);
 305                 return ERR_PTR(-EINVAL);
 306         }
 307
 308         memcpy(seclogname, logname, ptr - logname);
 309         strcpy(seclogname + (ptr - logname), "-sptlrpc");
 310
 311         sptlrpc_cld = config_log_find(seclogname, NULL);
 312         if (!sptlrpc_cld) {
 313                 sptlrpc_cld = do_config_log_add(obd, seclogname,
 314                                                 CONFIG_T_SPTLRPC, NULL, NULL);
 315                 if (IS_ERR(sptlrpc_cld)) {
 316                         CERROR("can't create sptlrpc log: %s\n", seclogname);
 317                         rc = PTR_ERR(sptlrpc_cld);
 318                         goto out_err;
 319                 }
 320         }
 321         params_cld = config_params_log_add(obd, cfg, sb);
 322         if (IS_ERR(params_cld)) {
 323                 rc = PTR_ERR(params_cld);
 324                 CERROR("%s: can't create params log: rc = %d\n",
 325                        obd->obd_name, rc);
 326                 goto out_sptlrpc;
 327         }
 328
 329         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
 330         if (IS_ERR(cld)) {
 331                 CERROR("can't create log: %s\n", logname);
 332                 rc = PTR_ERR(cld);
 333                 goto out_params;
 334         }
 335
 336         LASSERT(lsi->lsi_lmd);
 337         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
 338                 ptr = strrchr(seclogname, '-');
 339                 if (ptr) {
 340                         *ptr = 0;
 341                 } else {
 342                         CERROR("%s: sptlrpc log name not correct, %s: rc = %d\n",
 343                                obd->obd_name, seclogname, -EINVAL);
 344                         rc = -EINVAL;
 345                         goto out_cld;
 346                 }
 347                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
 348                 if (IS_ERR(recover_cld)) {
 349                         rc = PTR_ERR(recover_cld);
 350                         goto out_cld;
 351                 }
 352         }
 353
 354         mutex_lock(&cld->cld_lock);
 355         cld->cld_recover = recover_cld;
 356         cld->cld_params = params_cld;
 357         cld->cld_sptlrpc = sptlrpc_cld;
 358         mutex_unlock(&cld->cld_lock);
 359
 360         return cld;
 361
 362 out_cld:
 363         config_log_put(cld);
 364
 365 out_params:
 366         config_log_put(params_cld);
 367
 368 out_sptlrpc:
 369         config_log_put(sptlrpc_cld);
 370
 371 out_err:
 372         return ERR_PTR(rc);
 373 }
 374
 375 static DEFINE_MUTEX(llog_process_lock);
 376
 377 static inline void config_mark_cld_stop(struct config_llog_data *cld)
 378 {
 379         mutex_lock(&cld->cld_lock);
 380         spin_lock(&config_list_lock);
 381         cld->cld_stopping = 1;
 382         spin_unlock(&config_list_lock);
 383         mutex_unlock(&cld->cld_lock);
 384 }
 385
 386 /** Stop watching for updates on this log.
 387  */
 388 static int config_log_end(char *logname, struct config_llog_instance *cfg)
 389 {
 390         struct config_llog_data *cld;
 391         struct config_llog_data *cld_sptlrpc = NULL;
 392         struct config_llog_data *cld_params = NULL;
 393         struct config_llog_data *cld_recover = NULL;
 394         int rc = 0;
 395
 396         cld = config_log_find(logname, cfg);
 397         if (!cld)
 398                 return -ENOENT;
 399
 400         mutex_lock(&cld->cld_lock);
 401         /*
 402          * if cld_stopping is set, it means we didn't start the log thus
 403          * not owning the start ref. this can happen after previous umount:
 404          * the cld still hanging there waiting for lock cancel, and we
 405          * remount again but failed in the middle and call log_end without
 406          * calling start_log.
 407          */
 408         if (unlikely(cld->cld_stopping)) {
 409                 mutex_unlock(&cld->cld_lock);
 410                 /* drop the ref from the find */
 411                 config_log_put(cld);
 412                 return rc;
 413         }
 414
 415         spin_lock(&config_list_lock);
 416         cld->cld_stopping = 1;
 417         spin_unlock(&config_list_lock);
 418
 419         cld_recover = cld->cld_recover;
 420         cld->cld_recover = NULL;
 421
 422         cld_params = cld->cld_params;
 423         cld->cld_params = NULL;
 424         cld_sptlrpc = cld->cld_sptlrpc;
 425         cld->cld_sptlrpc = NULL;
 426         mutex_unlock(&cld->cld_lock);
 427
 428         if (cld_recover) {
 429                 config_mark_cld_stop(cld_recover);
 430                 config_log_put(cld_recover);
 431         }
 432
 433         if (cld_params) {
 434                 config_mark_cld_stop(cld_params);
 435                 config_log_put(cld_params);
 436         }
 437
 438         if (cld_sptlrpc)
 439                 config_log_put(cld_sptlrpc);
 440
 441         /* drop the ref from the find */
 442         config_log_put(cld);
 443         /* drop the start ref */
 444         config_log_put(cld);
 445
 446         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
 447                rc);
 448         return rc;
 449 }
 450
 451 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
 452 {
 453         struct obd_device       *obd = data;
 454         struct obd_import       *imp;
 455         struct obd_connect_data *ocd;
 456         struct config_llog_data *cld;
 457         int rc;
 458
 459         rc = lprocfs_climp_check(obd);
 460         if (rc)
 461                 return rc;
 462
 463         imp = obd->u.cli.cl_import;
 464         ocd = &imp->imp_connect_data;
 465
 466         seq_printf(m, "imperative_recovery: %s\n",
 467                    OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
 468         seq_printf(m, "client_state:\n");
 469
 470         spin_lock(&config_list_lock);
 471         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 472                 if (!cld->cld_recover)
 473                         continue;
 474                 seq_printf(m, "    - { client: %s, nidtbl_version: %u }\n",
 475                            cld->cld_logname,
 476                            cld->cld_recover->cld_cfg.cfg_last_idx);
 477         }
 478         spin_unlock(&config_list_lock);
 479
 480         up_read(&obd->u.cli.cl_sem);
 481         return 0;
 482 }
 483
 484 /* reenqueue any lost locks */
 485 #define RQ_RUNNING 0x1
 486 #define RQ_NOW     0x2
 487 #define RQ_LATER   0x4
 488 #define RQ_STOP    0x8
 489 #define RQ_PRECLEANUP  0x10
 490 static int rq_state;
 491 static wait_queue_head_t            rq_waitq;
 492 static DECLARE_COMPLETION(rq_exit);
 493 static DECLARE_COMPLETION(rq_start);
 494
 495 static void do_requeue(struct config_llog_data *cld)
 496 {
 497         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 498
 499         /* Do not run mgc_process_log on a disconnected export or an
 500          * export which is being disconnected. Take the client
 501          * semaphore to make the check non-racy.
 502          */
 503         down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
 504                          OBD_CLI_SEM_MGC);
 505
 506         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
 507                 int rc;
 508
 509                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
 510                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
 511                 if (rc && rc != -ENOENT)
 512                         CERROR("failed processing log: %d\n", rc);
 513         } else {
 514                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
 515                        cld->cld_logname);
 516         }
 517         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
 518 }
 519
 520 /* this timeout represents how many seconds MGC should wait before
 521  * requeue config and recover lock to the MGS. We need to randomize this
 522  * in order to not flood the MGS.
 523  */
 524 #define MGC_TIMEOUT_MIN_SECONDS   5
 525 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
 526
 527 static int mgc_requeue_thread(void *data)
 528 {
 529         bool first = true;
 530
 531         CDEBUG(D_MGC, "Starting requeue thread\n");
 532
 533         /* Keep trying failed locks periodically */
 534         spin_lock(&config_list_lock);
 535         rq_state |= RQ_RUNNING;
 536         while (!(rq_state & RQ_STOP)) {
 537                 struct l_wait_info lwi;
 538                 struct config_llog_data *cld, *cld_prev;
 539                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
 540                 int to;
 541
 542                 /* Any new or requeued lostlocks will change the state */
 543                 rq_state &= ~(RQ_NOW | RQ_LATER);
 544                 spin_unlock(&config_list_lock);
 545
 546                 if (first) {
 547                         first = false;
 548                         complete(&rq_start);
 549                 }
 550
 551                 /* Always wait a few seconds to allow the server who
 552                  * caused the lock revocation to finish its setup, plus some
 553                  * random so everyone doesn't try to reconnect at once.
 554                  */
 555                 to = msecs_to_jiffies(MGC_TIMEOUT_MIN_SECONDS * MSEC_PER_SEC);
 556                 /* rand is centi-seconds */
 557                 to += msecs_to_jiffies(rand * MSEC_PER_SEC / 100);
 558                 lwi = LWI_TIMEOUT(to, NULL, NULL);
 559                 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
 560                              &lwi);
 561
 562                 /*
 563                  * iterate & processing through the list. for each cld, process
 564                  * its depending sptlrpc cld firstly (if any) and then itself.
 565                  *
 566                  * it's guaranteed any item in the list must have
 567                  * reference > 0; and if cld_lostlock is set, at
 568                  * least one reference is taken by the previous enqueue.
 569                  */
 570                 cld_prev = NULL;
 571
 572                 spin_lock(&config_list_lock);
 573                 rq_state &= ~RQ_PRECLEANUP;
 574                 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
 575                         if (!cld->cld_lostlock || cld->cld_stopping)
 576                                 continue;
 577
 578                         /*
 579                          * hold reference to avoid being freed during
 580                          * subsequent processing.
 581                          */
 582                         config_log_get(cld);
 583                         cld->cld_lostlock = 0;
 584                         spin_unlock(&config_list_lock);
 585
 586                         if (cld_prev)
 587                                 config_log_put(cld_prev);
 588                         cld_prev = cld;
 589
 590                         if (likely(!(rq_state & RQ_STOP))) {
 591                                 do_requeue(cld);
 592                                 spin_lock(&config_list_lock);
 593                         } else {
 594                                 spin_lock(&config_list_lock);
 595                                 break;
 596                         }
 597                 }
 598                 spin_unlock(&config_list_lock);
 599                 if (cld_prev)
 600                         config_log_put(cld_prev);
 601
 602                 /* Wait a bit to see if anyone else needs a requeue */
 603                 lwi = (struct l_wait_info) { 0 };
 604                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
 605                              &lwi);
 606                 spin_lock(&config_list_lock);
 607         }
 608
 609         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
 610         rq_state &= ~RQ_RUNNING;
 611         spin_unlock(&config_list_lock);
 612
 613         complete(&rq_exit);
 614
 615         CDEBUG(D_MGC, "Ending requeue thread\n");
 616         return 0;
 617 }
 618
 619 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
 620  * We are responsible for dropping the config log reference from here on out.
 621  */
 622 static void mgc_requeue_add(struct config_llog_data *cld)
 623 {
 624         bool wakeup = false;
 625
 626         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
 627                cld->cld_logname, atomic_read(&cld->cld_refcount),
 628                cld->cld_stopping, rq_state);
 629         LASSERT(atomic_read(&cld->cld_refcount) > 0);
 630
 631         mutex_lock(&cld->cld_lock);
 632         spin_lock(&config_list_lock);
 633         if (!(rq_state & RQ_STOP) && !cld->cld_stopping && !cld->cld_lostlock) {
 634                 cld->cld_lostlock = 1;
 635                 rq_state |= RQ_NOW;
 636                 wakeup = true;
 637         }
 638         spin_unlock(&config_list_lock);
 639         mutex_unlock(&cld->cld_lock);
 640         if (wakeup)
 641                 wake_up(&rq_waitq);
 642 }
 643
 644 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
 645 {
 646         struct llog_ctxt        *ctxt;
 647         int                      rc;
 648
 649         /* setup only remote ctxt, the local disk context is switched per each
 650          * filesystem during mgc_fs_setup()
 651          */
 652         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
 653                         &llog_client_ops);
 654         if (rc)
 655                 return rc;
 656
 657         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 658         LASSERT(ctxt);
 659
 660         llog_initiator_connect(ctxt);
 661         llog_ctxt_put(ctxt);
 662
 663         return 0;
 664 }
 665
 666 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
 667 {
 668         struct llog_ctxt *ctxt;
 669
 670         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
 671         if (ctxt)
 672                 llog_cleanup(env, ctxt);
 673
 674         return 0;
 675 }
 676
 677 static atomic_t mgc_count = ATOMIC_INIT(0);
 678 static int mgc_precleanup(struct obd_device *obd)
 679 {
 680         int rc = 0;
 681         int temp;
 682
 683         if (atomic_dec_and_test(&mgc_count)) {
 684                 LASSERT(rq_state & RQ_RUNNING);
 685                 /* stop requeue thread */
 686                 temp = RQ_STOP;
 687         } else {
 688                 /* wakeup requeue thread to clean our cld */
 689                 temp = RQ_NOW | RQ_PRECLEANUP;
 690         }
 691
 692         spin_lock(&config_list_lock);
 693         rq_state |= temp;
 694         spin_unlock(&config_list_lock);
 695         wake_up(&rq_waitq);
 696
 697         if (temp & RQ_STOP)
 698                 wait_for_completion(&rq_exit);
 699         obd_cleanup_client_import(obd);
 700
 701         rc = mgc_llog_fini(NULL, obd);
 702         if (rc)
 703                 CERROR("failed to cleanup llogging subsystems\n");
 704
 705         return rc;
 706 }
 707
 708 static int mgc_cleanup(struct obd_device *obd)
 709 {
 710         /* COMPAT_146 - old config logs may have added profiles we don't
 711          * know about
 712          */
 713         if (obd->obd_type->typ_refcnt <= 1)
 714                 /* Only for the last mgc */
 715                 class_del_profiles();
 716
 717         lprocfs_obd_cleanup(obd);
 718         ptlrpcd_decref();
 719
 720         return client_obd_cleanup(obd);
 721 }
 722
 723 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 724 {
 725         struct lprocfs_static_vars lvars = { NULL };
 726         struct task_struct *task;
 727         int rc;
 728
 729         rc = ptlrpcd_addref();
 730         if (rc < 0)
 731                 goto err_noref;
 732
 733         rc = client_obd_setup(obd, lcfg);
 734         if (rc)
 735                 goto err_decref;
 736
 737         rc = mgc_llog_init(NULL, obd);
 738         if (rc) {
 739                 CERROR("failed to setup llogging subsystems\n");
 740                 goto err_cleanup;
 741         }
 742
 743         lprocfs_mgc_init_vars(&lvars);
 744         lprocfs_obd_setup(obd, lvars.obd_vars, lvars.sysfs_vars);
 745         sptlrpc_lprocfs_cliobd_attach(obd);
 746
 747         if (atomic_inc_return(&mgc_count) == 1) {
 748                 rq_state = 0;
 749                 init_waitqueue_head(&rq_waitq);
 750
 751                 /* start requeue thread */
 752                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
 753                 if (IS_ERR(task)) {
 754                         rc = PTR_ERR(task);
 755                         CERROR("%s: cannot start requeue thread: rc = %d; no more log updates\n",
 756                                obd->obd_name, rc);
 757                         goto err_cleanup;
 758                 }
 759                 /* rc is the task_struct pointer of mgc_requeue_thread. */
 760                 rc = 0;
 761                 wait_for_completion(&rq_start);
 762         }
 763
 764         return rc;
 765
 766 err_cleanup:
 767         client_obd_cleanup(obd);
 768 err_decref:
 769         ptlrpcd_decref();
 770 err_noref:
 771         return rc;
 772 }
 773
 774 /* based on ll_mdc_blocking_ast */
 775 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 776                             void *data, int flag)
 777 {
 778         struct lustre_handle lockh;
 779         struct config_llog_data *cld = data;
 780         int rc = 0;
 781
 782         switch (flag) {
 783         case LDLM_CB_BLOCKING:
 784                 /* mgs wants the lock, give it up... */
 785                 LDLM_DEBUG(lock, "MGC blocking CB");
 786                 ldlm_lock2handle(lock, &lockh);
 787                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 788                 break;
 789         case LDLM_CB_CANCELING:
 790                 /* We've given up the lock, prepare ourselves to update. */
 791                 LDLM_DEBUG(lock, "MGC cancel CB");
 792
 793                 CDEBUG(D_MGC, "Lock res " DLDLMRES " (%.8s)\n",
 794                        PLDLMRES(lock->l_resource),
 795                        (char *)&lock->l_resource->lr_name.name[0]);
 796
 797                 if (!cld) {
 798                         CDEBUG(D_INFO, "missing data, won't requeue\n");
 799                         break;
 800                 }
 801
 802                 /* held at mgc_process_log(). */
 803                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
 804
 805                 lock->l_ast_data = NULL;
 806                 /* Are we done with this log? */
 807                 if (cld->cld_stopping) {
 808                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
 809                                cld->cld_logname);
 810                         config_log_put(cld);
 811                         break;
 812                 }
 813                 /* Make sure not to re-enqueue when the mgc is stopping
 814                  * (we get called from client_disconnect_export)
 815                  */
 816                 if (!lock->l_conn_export ||
 817                     !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
 818                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
 819                                cld->cld_logname);
 820                         config_log_put(cld);
 821                         break;
 822                 }
 823
 824                 /* Re-enqueue now */
 825                 mgc_requeue_add(cld);
 826                 config_log_put(cld);
 827                 break;
 828         default:
 829                 LBUG();
 830         }
 831
 832         return rc;
 833 }
 834
 835 /* Not sure where this should go... */
 836 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
 837  * that we can have a chance to try the secondary MGS if any.
 838  */
 839 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
 840                                 + PING_INTERVAL)
 841 #define  MGC_TARGET_REG_LIMIT 10
 842 #define  MGC_SEND_PARAM_LIMIT 10
 843
 844 /* Send parameter to MGS*/
 845 static int mgc_set_mgs_param(struct obd_export *exp,
 846                              struct mgs_send_param *msp)
 847 {
 848         struct ptlrpc_request *req;
 849         struct mgs_send_param *req_msp, *rep_msp;
 850         int rc;
 851
 852         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 853                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
 854                                         MGS_SET_INFO);
 855         if (!req)
 856                 return -ENOMEM;
 857
 858         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 859         if (!req_msp) {
 860                 ptlrpc_req_finished(req);
 861                 return -ENOMEM;
 862         }
 863
 864         memcpy(req_msp, msp, sizeof(*req_msp));
 865         ptlrpc_request_set_replen(req);
 866
 867         /* Limit how long we will wait for the enqueue to complete */
 868         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
 869         rc = ptlrpc_queue_wait(req);
 870         if (!rc) {
 871                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
 872                 memcpy(msp, rep_msp, sizeof(*rep_msp));
 873         }
 874
 875         ptlrpc_req_finished(req);
 876
 877         return rc;
 878 }
 879
 880 /* Take a config lock so we can get cancel notifications */
 881 static int mgc_enqueue(struct obd_export *exp, __u32 type,
 882                        union ldlm_policy_data *policy, __u32 mode,
 883                        __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
 884                        void *data, __u32 lvb_len, void *lvb_swabber,
 885                        struct lustre_handle *lockh)
 886 {
 887         struct config_llog_data *cld = data;
 888         struct ldlm_enqueue_info einfo = {
 889                 .ei_type        = type,
 890                 .ei_mode        = mode,
 891                 .ei_cb_bl       = mgc_blocking_ast,
 892                 .ei_cb_cp       = ldlm_completion_ast,
 893         };
 894         struct ptlrpc_request *req;
 895         int short_limit = cld_is_sptlrpc(cld);
 896         int rc;
 897
 898         CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
 899                cld->cld_resid.name[0]);
 900
 901         /* We need a callback for every lockholder, so don't try to
 902          * ldlm_lock_match (see rev 1.1.2.11.2.47)
 903          */
 904         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 905                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
 906                                         LDLM_ENQUEUE);
 907         if (!req)
 908                 return -ENOMEM;
 909
 910         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
 911         ptlrpc_request_set_replen(req);
 912
 913         /* Limit how long we will wait for the enqueue to complete */
 914         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
 915         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
 916                               NULL, 0, LVB_T_NONE, lockh, 0);
 917         /* A failed enqueue should still call the mgc_blocking_ast,
 918          * where it will be requeued if needed ("grant failed").
 919          */
 920         ptlrpc_req_finished(req);
 921         return rc;
 922 }
 923
 924 static void mgc_notify_active(struct obd_device *unused)
 925 {
 926         /* wakeup mgc_requeue_thread to requeue mgc lock */
 927         spin_lock(&config_list_lock);
 928         rq_state |= RQ_NOW;
 929         spin_unlock(&config_list_lock);
 930         wake_up(&rq_waitq);
 931
 932         /* TODO: Help the MGS rebuild nidtbl. -jay */
 933 }
 934
 935 /* Send target_reg message to MGS */
 936 static int mgc_target_register(struct obd_export *exp,
 937                                struct mgs_target_info *mti)
 938 {
 939         struct ptlrpc_request  *req;
 940         struct mgs_target_info *req_mti, *rep_mti;
 941         int                  rc;
 942
 943         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 944                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
 945                                         MGS_TARGET_REG);
 946         if (!req)
 947                 return -ENOMEM;
 948
 949         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
 950         if (!req_mti) {
 951                 ptlrpc_req_finished(req);
 952                 return -ENOMEM;
 953         }
 954
 955         memcpy(req_mti, mti, sizeof(*req_mti));
 956         ptlrpc_request_set_replen(req);
 957         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
 958         /* Limit how long we will wait for the enqueue to complete */
 959         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
 960
 961         rc = ptlrpc_queue_wait(req);
 962         if (!rc) {
 963                 rep_mti = req_capsule_server_get(&req->rq_pill,
 964                                                  &RMF_MGS_TARGET_INFO);
 965                 memcpy(mti, rep_mti, sizeof(*rep_mti));
 966                 CDEBUG(D_MGC, "register %s got index = %d\n",
 967                        mti->mti_svname, mti->mti_stripe_index);
 968         }
 969         ptlrpc_req_finished(req);
 970
 971         return rc;
 972 }
 973
 974 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
 975                               u32 keylen, void *key, u32 vallen,
 976                               void *val, struct ptlrpc_request_set *set)
 977 {
 978         int rc = -EINVAL;
 979
 980         /* Turn off initial_recov after we try all backup servers once */
 981         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
 982                 struct obd_import *imp = class_exp2cliimp(exp);
 983                 int value;
 984
 985                 if (vallen != sizeof(int))
 986                         return -EINVAL;
 987                 value = *(int *)val;
 988                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
 989                        imp->imp_obd->obd_name, value,
 990                        imp->imp_deactive, imp->imp_invalid,
 991                        imp->imp_replayable, imp->imp_obd->obd_replayable,
 992                        ptlrpc_import_state_name(imp->imp_state));
 993                 /* Resurrect if we previously died */
 994                 if ((imp->imp_state != LUSTRE_IMP_FULL &&
 995                      imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
 996                         ptlrpc_reconnect_import(imp);
 997                 return 0;
 998         }
 999         if (KEY_IS(KEY_SET_INFO)) {
1000                 struct mgs_send_param *msp;
1001
1002                 msp = val;
1003                 rc =  mgc_set_mgs_param(exp, msp);
1004                 return rc;
1005         }
1006         if (KEY_IS(KEY_MGSSEC)) {
1007                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1008                 struct sptlrpc_flavor  flvr;
1009
1010                 /*
1011                  * empty string means using current flavor, if which haven't
1012                  * been set yet, set it as null.
1013                  *
1014                  * if flavor has been set previously, check the asking flavor
1015                  * must match the existing one.
1016                  */
1017                 if (vallen == 0) {
1018                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1019                                 return 0;
1020                         val = "null";
1021                         vallen = 4;
1022                 }
1023
1024                 rc = sptlrpc_parse_flavor(val, &flvr);
1025                 if (rc) {
1026                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1027                                (char *)val);
1028                         return rc;
1029                 }
1030
1031                 /*
1032                  * caller already hold a mutex
1033                  */
1034                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1035                         cli->cl_flvr_mgc = flvr;
1036                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1037                                   sizeof(flvr)) != 0) {
1038                         char    str[20];
1039
1040                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1041                                             str, sizeof(str));
1042                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but currently %s is in use\n",
1043                                        (char *)val, str);
1044                         rc = -EPERM;
1045                 }
1046                 return rc;
1047         }
1048
1049         return rc;
1050 }
1051
1052 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1053                         __u32 keylen, void *key, __u32 *vallen, void *val)
1054 {
1055         int rc = -EINVAL;
1056
1057         if (KEY_IS(KEY_CONN_DATA)) {
1058                 struct obd_import *imp = class_exp2cliimp(exp);
1059                 struct obd_connect_data *data = val;
1060
1061                 if (*vallen == sizeof(*data)) {
1062                         *data = imp->imp_connect_data;
1063                         rc = 0;
1064                 }
1065         }
1066
1067         return rc;
1068 }
1069
1070 static int mgc_import_event(struct obd_device *obd,
1071                             struct obd_import *imp,
1072                             enum obd_import_event event)
1073 {
1074         LASSERT(imp->imp_obd == obd);
1075         CDEBUG(D_MGC, "import event %#x\n", event);
1076
1077         switch (event) {
1078         case IMP_EVENT_DISCON:
1079                 /* MGC imports should not wait for recovery */
1080                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1081                         ptlrpc_pinger_ir_down();
1082                 break;
1083         case IMP_EVENT_INACTIVE:
1084                 break;
1085         case IMP_EVENT_INVALIDATE: {
1086                 struct ldlm_namespace *ns = obd->obd_namespace;
1087
1088                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1089                 break;
1090         }
1091         case IMP_EVENT_ACTIVE:
1092                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1093                 /* Clearing obd_no_recov allows us to continue pinging */
1094                 obd->obd_no_recov = 0;
1095                 mgc_notify_active(obd);
1096                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1097                         ptlrpc_pinger_ir_up();
1098                 break;
1099         case IMP_EVENT_OCD:
1100                 break;
1101         case IMP_EVENT_DEACTIVATE:
1102         case IMP_EVENT_ACTIVATE:
1103                 break;
1104         default:
1105                 CERROR("Unknown import event %#x\n", event);
1106                 LBUG();
1107         }
1108         return 0;
1109 }
1110
1111 enum {
1112         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1113         CONFIG_READ_NRPAGES      = 4
1114 };
1115
1116 static int mgc_apply_recover_logs(struct obd_device *mgc,
1117                                   struct config_llog_data *cld,
1118                                   __u64 max_version,
1119                                   void *data, int datalen, bool mne_swab)
1120 {
1121         struct config_llog_instance *cfg = &cld->cld_cfg;
1122         struct mgs_nidtbl_entry *entry;
1123         struct lustre_cfg       *lcfg;
1124         struct lustre_cfg_bufs   bufs;
1125         u64   prev_version = 0;
1126         char *inst;
1127         char *buf;
1128         int   bufsz;
1129         int   pos;
1130         int   rc  = 0;
1131         int   off = 0;
1132
1133         LASSERT(cfg->cfg_instance);
1134         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1135
1136         inst = kzalloc(PAGE_SIZE, GFP_KERNEL);
1137         if (!inst)
1138                 return -ENOMEM;
1139
1140         pos = snprintf(inst, PAGE_SIZE, "%p", cfg->cfg_instance);
1141         if (pos >= PAGE_SIZE) {
1142                 kfree(inst);
1143                 return -E2BIG;
1144         }
1145
1146         ++pos;
1147         buf   = inst + pos;
1148         bufsz = PAGE_SIZE - pos;
1149
1150         while (datalen > 0) {
1151                 int   entry_len = sizeof(*entry);
1152                 int is_ost, i;
1153                 struct obd_device *obd;
1154                 char *obdname;
1155                 char *cname;
1156                 char *params;
1157                 char *uuid;
1158                 size_t len;
1159
1160                 rc = -EINVAL;
1161                 if (datalen < sizeof(*entry))
1162                         break;
1163
1164                 entry = (typeof(entry))(data + off);
1165
1166                 /* sanity check */
1167                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1168                         break;
1169                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1170                         break;
1171                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1172                         break;
1173
1174                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1175                 if (datalen < entry_len) /* must have entry_len at least */
1176                         break;
1177
1178                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1179                 if (mne_swab)
1180                         lustre_swab_mgs_nidtbl_entry(entry);
1181                 if (entry->mne_length > PAGE_SIZE) {
1182                         CERROR("MNE too large (%u)\n", entry->mne_length);
1183                         break;
1184                 }
1185
1186                 if (entry->mne_length < entry_len)
1187                         break;
1188
1189                 off     += entry->mne_length;
1190                 datalen -= entry->mne_length;
1191                 if (datalen < 0)
1192                         break;
1193
1194                 if (entry->mne_version > max_version) {
1195                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1196                                entry->mne_version, max_version);
1197                         break;
1198                 }
1199
1200                 if (prev_version >= entry->mne_version) {
1201                         CERROR("index unsorted, prev %lld, now %lld\n",
1202                                prev_version, entry->mne_version);
1203                         break;
1204                 }
1205                 prev_version = entry->mne_version;
1206
1207                 /*
1208                  * Write a string with format "nid::instance" to
1209                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1210                  */
1211
1212                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1213                 memset(buf, 0, bufsz);
1214                 obdname = buf;
1215                 pos = 0;
1216
1217                 /* lustre-OST0001-osc-<instance #> */
1218                 strcpy(obdname, cld->cld_logname);
1219                 cname = strrchr(obdname, '-');
1220                 if (!cname) {
1221                         CERROR("mgc %s: invalid logname %s\n",
1222                                mgc->obd_name, obdname);
1223                         break;
1224                 }
1225
1226                 pos = cname - obdname;
1227                 obdname[pos] = 0;
1228                 pos += sprintf(obdname + pos, "-%s%04x",
1229                                   is_ost ? "OST" : "MDT", entry->mne_index);
1230
1231                 cname = is_ost ? "osc" : "mdc";
1232                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1233                 lustre_cfg_bufs_reset(&bufs, obdname);
1234
1235                 /* find the obd by obdname */
1236                 obd = class_name2obd(obdname);
1237                 if (!obd) {
1238                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1239                                mgc->obd_name, obdname);
1240                         rc = 0;
1241                         /* this is a safe race, when the ost is starting up...*/
1242                         continue;
1243                 }
1244
1245                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1246                 ++pos;
1247                 params = buf + pos;
1248                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1249                 uuid = buf + pos;
1250
1251                 down_read(&obd->u.cli.cl_sem);
1252                 if (!obd->u.cli.cl_import) {
1253                         /* client does not connect to the OST yet */
1254                         up_read(&obd->u.cli.cl_sem);
1255                         rc = 0;
1256                         continue;
1257                 }
1258
1259                 /* iterate all nids to find one */
1260                 /* find uuid by nid */
1261                 rc = -ENOENT;
1262                 for (i = 0; i < entry->mne_nid_count; i++) {
1263                         rc = client_import_find_conn(obd->u.cli.cl_import,
1264                                                      entry->u.nids[0],
1265                                                      (struct obd_uuid *)uuid);
1266                         if (!rc)
1267                                 break;
1268                 }
1269
1270                 up_read(&obd->u.cli.cl_sem);
1271                 if (rc < 0) {
1272                         CERROR("mgc: cannot find uuid by nid %s\n",
1273                                libcfs_nid2str(entry->u.nids[0]));
1274                         break;
1275                 }
1276
1277                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1278                        uuid, libcfs_nid2str(entry->u.nids[0]));
1279
1280                 pos += strlen(uuid);
1281                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1282                 LASSERT(pos < bufsz);
1283
1284                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1285
1286                 rc = -ENOMEM;
1287                 len = lustre_cfg_len(bufs.lcfg_bufcount, bufs.lcfg_buflen);
1288                 lcfg = kzalloc(len, GFP_NOFS);
1289                 if (!lcfg) {
1290                         rc = -ENOMEM;
1291                         break;
1292                 }
1293                 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1294
1295                 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1296                        prev_version, max_version, obdname, params);
1297
1298                 rc = class_process_config(lcfg);
1299                 kfree(lcfg);
1300                 if (rc)
1301                         CDEBUG(D_INFO, "process config for %s error %d\n",
1302                                obdname, rc);
1303
1304                 /* continue, even one with error */
1305         }
1306
1307         kfree(inst);
1308         return rc;
1309 }
1310
1311 /**
1312  * This function is called if this client was notified for target restarting
1313  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1314  */
1315 static int mgc_process_recover_log(struct obd_device *obd,
1316                                    struct config_llog_data *cld)
1317 {
1318         struct ptlrpc_request *req = NULL;
1319         struct config_llog_instance *cfg = &cld->cld_cfg;
1320         struct mgs_config_body *body;
1321         struct mgs_config_res  *res;
1322         struct ptlrpc_bulk_desc *desc;
1323         struct page **pages;
1324         int nrpages;
1325         bool eof = true;
1326         bool mne_swab;
1327         int i;
1328         int ealen;
1329         int rc;
1330
1331         /* allocate buffer for bulk transfer.
1332          * if this is the first time for this mgs to read logs,
1333          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1334          * once; otherwise, it only reads increment of logs, this should be
1335          * small and CONFIG_READ_NRPAGES will be used.
1336          */
1337         nrpages = CONFIG_READ_NRPAGES;
1338         if (cfg->cfg_last_idx == 0) /* the first time */
1339                 nrpages = CONFIG_READ_NRPAGES_INIT;
1340
1341         pages = kcalloc(nrpages, sizeof(*pages), GFP_KERNEL);
1342         if (!pages) {
1343                 rc = -ENOMEM;
1344                 goto out;
1345         }
1346
1347         for (i = 0; i < nrpages; i++) {
1348                 pages[i] = alloc_page(GFP_KERNEL);
1349                 if (!pages[i]) {
1350                         rc = -ENOMEM;
1351                         goto out;
1352                 }
1353         }
1354
1355 again:
1356         LASSERT(cld_is_recover(cld));
1357         LASSERT(mutex_is_locked(&cld->cld_lock));
1358         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1359                                    &RQF_MGS_CONFIG_READ);
1360         if (!req) {
1361                 rc = -ENOMEM;
1362                 goto out;
1363         }
1364
1365         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1366         if (rc)
1367                 goto out;
1368
1369         /* pack request */
1370         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1371         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1372         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1373             >= sizeof(body->mcb_name)) {
1374                 rc = -E2BIG;
1375                 goto out;
1376         }
1377         body->mcb_offset = cfg->cfg_last_idx + 1;
1378         body->mcb_type   = cld->cld_type;
1379         body->mcb_bits   = PAGE_SHIFT;
1380         body->mcb_units  = nrpages;
1381
1382         /* allocate bulk transfer descriptor */
1383         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1384                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1385                                     MGS_BULK_PORTAL,
1386                                     &ptlrpc_bulk_kiov_pin_ops);
1387         if (!desc) {
1388                 rc = -ENOMEM;
1389                 goto out;
1390         }
1391
1392         for (i = 0; i < nrpages; i++)
1393                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0, PAGE_SIZE);
1394
1395         ptlrpc_request_set_replen(req);
1396         rc = ptlrpc_queue_wait(req);
1397         if (rc)
1398                 goto out;
1399
1400         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1401         if (res->mcr_size < res->mcr_offset) {
1402                 rc = -EINVAL;
1403                 goto out;
1404         }
1405
1406         /* always update the index even though it might have errors with
1407          * handling the recover logs
1408          */
1409         cfg->cfg_last_idx = res->mcr_offset;
1410         eof = res->mcr_offset == res->mcr_size;
1411
1412         CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1413                res->mcr_offset, eof == false);
1414
1415         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1416         if (ealen < 0) {
1417                 rc = ealen;
1418                 goto out;
1419         }
1420
1421         if (ealen > nrpages << PAGE_SHIFT) {
1422                 rc = -EINVAL;
1423                 goto out;
1424         }
1425
1426         if (ealen == 0) { /* no logs transferred */
1427                 if (!eof)
1428                         rc = -EINVAL;
1429                 goto out;
1430         }
1431
1432         mne_swab = !!ptlrpc_rep_need_swab(req);
1433 #if OBD_OCD_VERSION(3, 0, 53, 0) > LUSTRE_VERSION_CODE
1434         /* This import flag means the server did an extra swab of IR MNE
1435          * records (fixed in LU-1252), reverse it here if needed. LU-1644
1436          */
1437         if (unlikely(req->rq_import->imp_need_mne_swab))
1438                 mne_swab = !mne_swab;
1439 #endif
1440
1441         for (i = 0; i < nrpages && ealen > 0; i++) {
1442                 int rc2;
1443                 void *ptr;
1444
1445                 ptr = kmap(pages[i]);
1446                 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1447                                              min_t(int, ealen, PAGE_SIZE),
1448                                              mne_swab);
1449                 kunmap(pages[i]);
1450                 if (rc2 < 0) {
1451                         CWARN("Process recover log %s error %d\n",
1452                               cld->cld_logname, rc2);
1453                         break;
1454                 }
1455
1456                 ealen -= PAGE_SIZE;
1457         }
1458
1459 out:
1460         if (req)
1461                 ptlrpc_req_finished(req);
1462
1463         if (rc == 0 && !eof)
1464                 goto again;
1465
1466         if (pages) {
1467                 for (i = 0; i < nrpages; i++) {
1468                         if (!pages[i])
1469                                 break;
1470                         __free_page(pages[i]);
1471                 }
1472                 kfree(pages);
1473         }
1474         return rc;
1475 }
1476
1477 /* local_only means it cannot get remote llogs */
1478 static int mgc_process_cfg_log(struct obd_device *mgc,
1479                                struct config_llog_data *cld, int local_only)
1480 {
1481         struct llog_ctxt        *ctxt;
1482         struct lustre_sb_info   *lsi = NULL;
1483         int                      rc = 0;
1484         bool                     sptlrpc_started = false;
1485         struct lu_env           *env;
1486
1487         LASSERT(cld);
1488         LASSERT(mutex_is_locked(&cld->cld_lock));
1489
1490         /*
1491          * local copy of sptlrpc log is controlled elsewhere, don't try to
1492          * read it up here.
1493          */
1494         if (cld_is_sptlrpc(cld) && local_only)
1495                 return 0;
1496
1497         if (cld->cld_cfg.cfg_sb)
1498                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1499
1500         env = kzalloc(sizeof(*env), GFP_KERNEL);
1501         if (!env)
1502                 return -ENOMEM;
1503
1504         rc = lu_env_init(env, LCT_MG_THREAD);
1505         if (rc)
1506                 goto out_free;
1507
1508         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1509         LASSERT(ctxt);
1510
1511         if (local_only) /* no local log at client side */ {
1512                 rc = -EIO;
1513                 goto out_pop;
1514         }
1515
1516         if (cld_is_sptlrpc(cld)) {
1517                 sptlrpc_conf_log_update_begin(cld->cld_logname);
1518                 sptlrpc_started = true;
1519         }
1520
1521         /* logname and instance info should be the same, so use our
1522          * copy of the instance for the update.  The cfg_last_idx will
1523          * be updated here.
1524          */
1525         rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1526                                      &cld->cld_cfg);
1527
1528 out_pop:
1529         __llog_ctxt_put(env, ctxt);
1530
1531         /*
1532          * update settings on existing OBDs. doing it inside
1533          * of llog_process_lock so no device is attaching/detaching
1534          * in parallel.
1535          * the logname must be <fsname>-sptlrpc
1536          */
1537         if (sptlrpc_started) {
1538                 LASSERT(cld_is_sptlrpc(cld));
1539                 sptlrpc_conf_log_update_end(cld->cld_logname);
1540                 class_notify_sptlrpc_conf(cld->cld_logname,
1541                                           strlen(cld->cld_logname) -
1542                                           strlen("-sptlrpc"));
1543         }
1544
1545         lu_env_fini(env);
1546 out_free:
1547         kfree(env);
1548         return rc;
1549 }
1550
1551 static bool mgc_import_in_recovery(struct obd_import *imp)
1552 {
1553         bool in_recovery = true;
1554
1555         spin_lock(&imp->imp_lock);
1556         if (imp->imp_state == LUSTRE_IMP_FULL ||
1557             imp->imp_state == LUSTRE_IMP_CLOSED)
1558                 in_recovery = false;
1559         spin_unlock(&imp->imp_lock);
1560
1561         return in_recovery;
1562 }
1563
1564 /**
1565  * Get a configuration log from the MGS and process it.
1566  *
1567  * This function is called for both clients and servers to process the
1568  * configuration log from the MGS.  The MGC enqueues a DLM lock on the
1569  * log from the MGS, and if the lock gets revoked the MGC will be notified
1570  * by the lock cancellation callback that the config log has changed,
1571  * and will enqueue another MGS lock on it, and then continue processing
1572  * the new additions to the end of the log.
1573  *
1574  * Since the MGC import is not replayable, if the import is being evicted
1575  * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
1576  * the log until recovery is finished or the import is closed.
1577  *
1578  * Make a local copy of the log before parsing it if appropriate (non-MGS
1579  * server) so that the server can start even when the MGS is down.
1580  *
1581  * There shouldn't be multiple processes running process_log at once --
1582  * sounds like badness.  It actually might be fine, as long as they're not
1583  * trying to update from the same log simultaneously, in which case we
1584  * should use a per-log semaphore instead of cld_lock.
1585  *
1586  * \param[in] mgc       MGC device by which to fetch the configuration log
1587  * \param[in] cld       log processing state (stored in lock callback data)
1588  *
1589  * \retval              0 on success
1590  * \retval              negative errno on failure
1591  */
1592 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1593 {
1594         struct lustre_handle lockh = { 0 };
1595         __u64 flags = LDLM_FL_NO_LRU;
1596         bool retry = false;
1597         int rc = 0, rcl;
1598
1599         LASSERT(cld);
1600
1601         /* I don't want multiple processes running process_log at once --
1602          * sounds like badness.  It actually might be fine, as long as
1603          * we're not trying to update from the same log
1604          * simultaneously (in which case we should use a per-log sem.)
1605          */
1606 restart:
1607         mutex_lock(&cld->cld_lock);
1608         if (cld->cld_stopping) {
1609                 mutex_unlock(&cld->cld_lock);
1610                 return 0;
1611         }
1612
1613         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1614
1615         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1616                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1617
1618         /* Get the cfg lock on the llog */
1619         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
1620                           LCK_CR, &flags, NULL, NULL, NULL,
1621                           cld, 0, NULL, &lockh);
1622         if (rcl == 0) {
1623                 /* Get the cld, it will be released in mgc_blocking_ast. */
1624                 config_log_get(cld);
1625                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1626                 LASSERT(rc == 0);
1627         } else {
1628                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1629
1630                 if (rcl == -ESHUTDOWN &&
1631                     atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
1632                         int secs = cfs_time_seconds(obd_timeout);
1633                         struct obd_import *imp;
1634                         struct l_wait_info lwi;
1635
1636                         mutex_unlock(&cld->cld_lock);
1637                         imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
1638
1639                         /*
1640                          * Let's force the pinger, and wait the import to be
1641                          * connected, note: since mgc import is non-replayable,
1642                          * and even the import state is disconnected, it does
1643                          * not mean the "recovery" is stopped, so we will keep
1644                          * waitting until timeout or the import state is
1645                          * FULL or closed
1646                          */
1647                         ptlrpc_pinger_force(imp);
1648
1649                         lwi = LWI_TIMEOUT(secs, NULL, NULL);
1650                         l_wait_event(imp->imp_recovery_waitq,
1651                                      !mgc_import_in_recovery(imp), &lwi);
1652
1653                         if (imp->imp_state == LUSTRE_IMP_FULL) {
1654                                 retry = true;
1655                                 goto restart;
1656                         } else {
1657                                 mutex_lock(&cld->cld_lock);
1658                                 spin_lock(&config_list_lock);
1659                                 cld->cld_lostlock = 1;
1660                                 spin_unlock(&config_list_lock);
1661                         }
1662                 } else {
1663                         /* mark cld_lostlock so that it will requeue
1664                          * after MGC becomes available.
1665                          */
1666                         spin_lock(&config_list_lock);
1667                         cld->cld_lostlock = 1;
1668                         spin_unlock(&config_list_lock);
1669                 }
1670         }
1671
1672         if (cld_is_recover(cld)) {
1673                 rc = 0; /* this is not a fatal error for recover log */
1674                 if (!rcl) {
1675                         rc = mgc_process_recover_log(mgc, cld);
1676                         if (rc) {
1677                                 CERROR("%s: recover log %s failed: rc = %d not fatal.\n",
1678                                        mgc->obd_name, cld->cld_logname, rc);
1679                                 rc = 0;
1680                                 spin_lock(&config_list_lock);
1681                                 cld->cld_lostlock = 1;
1682                                 spin_unlock(&config_list_lock);
1683                         }
1684                 }
1685         } else {
1686                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1687         }
1688
1689         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1690                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1691
1692         mutex_unlock(&cld->cld_lock);
1693
1694         /* Now drop the lock so MGS can revoke it */
1695         if (!rcl)
1696                 ldlm_lock_decref(&lockh, LCK_CR);
1697
1698         return rc;
1699 }
1700
1701 /** Called from lustre_process_log.
1702  * LCFG_LOG_START gets the config log from the MGS, processes it to start
1703  * any services, and adds it to the list logs to watch (follow).
1704  */
1705 static int mgc_process_config(struct obd_device *obd, u32 len, void *buf)
1706 {
1707         struct lustre_cfg *lcfg = buf;
1708         struct config_llog_instance *cfg = NULL;
1709         char *logname;
1710         int rc = 0;
1711
1712         switch (lcfg->lcfg_command) {
1713         case LCFG_LOV_ADD_OBD: {
1714                 /* Overloading this cfg command: register a new target */
1715                 struct mgs_target_info *mti;
1716
1717                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1718                     sizeof(struct mgs_target_info)) {
1719                         rc = -EINVAL;
1720                         goto out;
1721                 }
1722
1723                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1724                 CDEBUG(D_MGC, "add_target %s %#x\n",
1725                        mti->mti_svname, mti->mti_flags);
1726                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1727                 break;
1728         }
1729         case LCFG_LOV_DEL_OBD:
1730                 /* Unregister has no meaning at the moment. */
1731                 CERROR("lov_del_obd unimplemented\n");
1732                 rc = -ENOSYS;
1733                 break;
1734         case LCFG_SPTLRPC_CONF: {
1735                 rc = sptlrpc_process_config(lcfg);
1736                 break;
1737         }
1738         case LCFG_LOG_START: {
1739                 struct config_llog_data *cld;
1740                 struct super_block *sb;
1741
1742                 logname = lustre_cfg_string(lcfg, 1);
1743                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1744                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1745
1746                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1747                        cfg->cfg_last_idx);
1748
1749                 /* We're only called through here on the initial mount */
1750                 cld = config_log_add(obd, logname, cfg, sb);
1751                 if (IS_ERR(cld)) {
1752                         rc = PTR_ERR(cld);
1753                         break;
1754                 }
1755
1756                 /* COMPAT_146 */
1757                 /* FIXME only set this for old logs!  Right now this forces
1758                  * us to always skip the "inside markers" check
1759                  */
1760                 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1761
1762                 rc = mgc_process_log(obd, cld);
1763                 if (rc == 0 && cld->cld_recover) {
1764                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1765                                          imp_connect_data, IMP_RECOV)) {
1766                                 rc = mgc_process_log(obd, cld->cld_recover);
1767                         } else {
1768                                 struct config_llog_data *cir;
1769
1770                                 mutex_lock(&cld->cld_lock);
1771                                 cir = cld->cld_recover;
1772                                 cld->cld_recover = NULL;
1773                                 mutex_unlock(&cld->cld_lock);
1774                                 config_log_put(cir);
1775                         }
1776
1777                         if (rc)
1778                                 CERROR("Cannot process recover llog %d\n", rc);
1779                 }
1780
1781                 if (rc == 0 && cld->cld_params) {
1782                         rc = mgc_process_log(obd, cld->cld_params);
1783                         if (rc == -ENOENT) {
1784                                 CDEBUG(D_MGC,
1785                                        "There is no params config file yet\n");
1786                                 rc = 0;
1787                         }
1788                         /* params log is optional */
1789                         if (rc)
1790                                 CERROR(
1791                                        "%s: can't process params llog: rc = %d\n",
1792                                        obd->obd_name, rc);
1793                 }
1794
1795                 break;
1796         }
1797         case LCFG_LOG_END: {
1798                 logname = lustre_cfg_string(lcfg, 1);
1799
1800                 if (lcfg->lcfg_bufcount >= 2)
1801                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
1802                                 lcfg, 2);
1803                 rc = config_log_end(logname, cfg);
1804                 break;
1805         }
1806         default: {
1807                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1808                 rc = -EINVAL;
1809                 goto out;
1810         }
1811         }
1812 out:
1813         return rc;
1814 }
1815
1816 static struct obd_ops mgc_obd_ops = {
1817         .owner          = THIS_MODULE,
1818         .setup          = mgc_setup,
1819         .precleanup     = mgc_precleanup,
1820         .cleanup        = mgc_cleanup,
1821         .add_conn       = client_import_add_conn,
1822         .del_conn       = client_import_del_conn,
1823         .connect        = client_connect_import,
1824         .disconnect     = client_disconnect_export,
1825         .set_info_async = mgc_set_info_async,
1826         .get_info       = mgc_get_info,
1827         .import_event   = mgc_import_event,
1828         .process_config = mgc_process_config,
1829 };
1830
1831 static int __init mgc_init(void)
1832 {
1833         return class_register_type(&mgc_obd_ops, NULL,
1834                                    LUSTRE_MGC_NAME, NULL);
1835 }
1836
1837 static void /*__exit*/ mgc_exit(void)
1838 {
1839         class_unregister_type(LUSTRE_MGC_NAME);
1840 }
1841
1842 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1843 MODULE_DESCRIPTION("Lustre Management Client");
1844 MODULE_VERSION(LUSTRE_VERSION_STRING);
1845 MODULE_LICENSE("GPL");
1846
1847 module_init(mgc_init);
1848 module_exit(mgc_exit);