lustre/mdt/mdt_mds.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,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License version 2 for more details.  A copy is
  14  * included in the COPYING file that accompanied this code.
  15
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  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) 2013, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lustre/mdt/mdt_mds.c
  32  *
  33  * Lustre Metadata Service Layer
  34  *
  35  * Author: Di Wang <di.wang@whamcloud.com>
  36  **/
  37
  38 #define DEBUG_SUBSYSTEM S_MDS
  39
  40 #include <linux/module.h>
  41
  42 #include <obd_support.h>
  43 /* struct ptlrpc_request */
  44 #include <lustre_net.h>
  45 /* struct obd_export */
  46 #include <lustre_export.h>
  47 /* struct obd_device */
  48 #include <obd.h>
  49 /* lu2dt_dev() */
  50 #include <dt_object.h>
  51 #include <lustre_mds.h>
  52 #include "mdt_internal.h"
  53 #include <lustre_quota.h>
  54 #include <lustre_acl.h>
  55 #include <uapi/linux/lustre/lustre_param.h>
  56
  57 struct mds_device {
  58         /* super-class */
  59         struct md_device         mds_md_dev;
  60         struct ptlrpc_service   *mds_regular_service;
  61         struct ptlrpc_service   *mds_readpage_service;
  62         struct ptlrpc_service   *mds_out_service;
  63         struct ptlrpc_service   *mds_setattr_service;
  64         struct ptlrpc_service   *mds_mdsc_service;
  65         struct ptlrpc_service   *mds_mdss_service;
  66         struct ptlrpc_service   *mds_fld_service;
  67         struct ptlrpc_service   *mds_io_service;
  68         struct mutex             mds_health_mutex;
  69 };
  70
  71 /*
  72  *  * Initialized in mds_mod_init().
  73  *   */
  74 static unsigned long mds_num_threads;
  75 module_param(mds_num_threads, ulong, 0444);
  76 MODULE_PARM_DESC(mds_num_threads, "number of MDS service threads to start");
  77
  78 int mds_max_io_threads = 512;
  79 module_param(mds_max_io_threads, int, 0444);
  80 MODULE_PARM_DESC(mds_max_io_threads, "maximum number of MDS IO service threads");
  81
  82 static char *mds_num_cpts;
  83 module_param(mds_num_cpts, charp, 0444);
  84 MODULE_PARM_DESC(mds_num_cpts, "CPU partitions MDS threads should run on");
  85
  86 static unsigned long mds_rdpg_num_threads;
  87 module_param(mds_rdpg_num_threads, ulong, 0444);
  88 MODULE_PARM_DESC(mds_rdpg_num_threads,
  89                  "number of MDS readpage service threads to start");
  90
  91 static char *mds_rdpg_num_cpts;
  92 module_param(mds_rdpg_num_cpts, charp, 0444);
  93 MODULE_PARM_DESC(mds_rdpg_num_cpts,
  94                  "CPU partitions MDS readpage threads should run on");
  95
  96 /* NB: these two should be removed along with setattr service in the future */
  97 static unsigned long mds_attr_num_threads;
  98 module_param(mds_attr_num_threads, ulong, 0444);
  99 MODULE_PARM_DESC(mds_attr_num_threads,
 100                  "number of MDS setattr service threads to start");
 101
 102 static char *mds_attr_num_cpts;
 103 module_param(mds_attr_num_cpts, charp, 0444);
 104 MODULE_PARM_DESC(mds_attr_num_cpts,
 105                  "CPU partitions MDS setattr threads should run on");
 106
 107 /* device init/fini methods */
 108 static void mds_stop_ptlrpc_service(struct mds_device *m)
 109 {
 110         ENTRY;
 111
 112         mutex_lock(&m->mds_health_mutex);
 113         if (m->mds_regular_service != NULL) {
 114                 ptlrpc_unregister_service(m->mds_regular_service);
 115                 m->mds_regular_service = NULL;
 116         }
 117         if (m->mds_readpage_service != NULL) {
 118                 ptlrpc_unregister_service(m->mds_readpage_service);
 119                 m->mds_readpage_service = NULL;
 120         }
 121         if (m->mds_out_service != NULL) {
 122                 ptlrpc_unregister_service(m->mds_out_service);
 123                 m->mds_out_service = NULL;
 124         }
 125         if (m->mds_setattr_service != NULL) {
 126                 ptlrpc_unregister_service(m->mds_setattr_service);
 127                 m->mds_setattr_service = NULL;
 128         }
 129         if (m->mds_mdsc_service != NULL) {
 130                 ptlrpc_unregister_service(m->mds_mdsc_service);
 131                 m->mds_mdsc_service = NULL;
 132         }
 133         if (m->mds_mdss_service != NULL) {
 134                 ptlrpc_unregister_service(m->mds_mdss_service);
 135                 m->mds_mdss_service = NULL;
 136         }
 137         if (m->mds_fld_service != NULL) {
 138                 ptlrpc_unregister_service(m->mds_fld_service);
 139                 m->mds_fld_service = NULL;
 140         }
 141         if (m->mds_io_service != NULL) {
 142                 ptlrpc_unregister_service(m->mds_io_service);
 143                 m->mds_io_service = NULL;
 144         }
 145         mutex_unlock(&m->mds_health_mutex);
 146
 147         EXIT;
 148 }
 149
 150 static int mds_start_ptlrpc_service(struct mds_device *m)
 151 {
 152         static struct ptlrpc_service_conf conf;
 153         struct obd_device *obd = m->mds_md_dev.md_lu_dev.ld_obd;
 154         struct proc_dir_entry *procfs_entry;
 155         int rc = 0;
 156         ENTRY;
 157
 158         procfs_entry = obd->obd_proc_entry;
 159         LASSERT(procfs_entry != NULL);
 160
 161         conf = (typeof(conf)) {
 162                 .psc_name               = LUSTRE_MDT_NAME,
 163                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 164                 .psc_buf                = {
 165                         .bc_nbufs               = MDS_NBUFS,
 166                         .bc_buf_size            = MDS_REG_BUFSIZE,
 167                         .bc_req_max_size        = MDS_REG_MAXREQSIZE,
 168                         .bc_rep_max_size        = MDS_REG_MAXREPSIZE,
 169                         .bc_req_portal          = MDS_REQUEST_PORTAL,
 170                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 171                 },
 172                 /*
 173                  * We'd like to have a mechanism to set this on a per-device
 174                  * basis, but alas...
 175                  */
 176                 .psc_thr                = {
 177                         .tc_thr_name            = LUSTRE_MDT_NAME,
 178                         .tc_thr_factor          = MDS_THR_FACTOR,
 179                         .tc_nthrs_init          = MDS_NTHRS_INIT,
 180                         .tc_nthrs_base          = MDS_NTHRS_BASE,
 181                         .tc_nthrs_max           = MDS_NTHRS_MAX,
 182                         .tc_nthrs_user          = mds_num_threads,
 183                         .tc_cpu_affinity        = 1,
 184                         .tc_ctx_tags            = LCT_MD_THREAD,
 185                 },
 186                 .psc_cpt                = {
 187                         .cc_pattern             = mds_num_cpts,
 188                 },
 189                 .psc_ops                = {
 190                         .so_req_handler         = tgt_request_handle,
 191                         .so_req_printer         = target_print_req,
 192                         .so_hpreq_handler       = ptlrpc_hpreq_handler,
 193                 },
 194         };
 195         m->mds_regular_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 196                                                          procfs_entry);
 197         if (IS_ERR(m->mds_regular_service)) {
 198                 rc = PTR_ERR(m->mds_regular_service);
 199                 CERROR("failed to start regular mdt service: %d\n", rc);
 200                 m->mds_regular_service = NULL;
 201
 202                 RETURN(rc);
 203         }
 204
 205         /*
 206          * readpage service configuration. Parameters have to be adjusted,
 207          * ideally.
 208          */
 209         memset(&conf, 0, sizeof(conf));
 210         conf = (typeof(conf)) {
 211                 .psc_name               = LUSTRE_MDT_NAME "_readpage",
 212                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 213                 .psc_buf                = {
 214                         .bc_nbufs               = MDS_NBUFS,
 215                         .bc_buf_size            = MDS_BUFSIZE,
 216                         .bc_req_max_size        = MDS_MAXREQSIZE,
 217                         .bc_rep_max_size        = MDS_MAXREPSIZE,
 218                         .bc_req_portal          = MDS_READPAGE_PORTAL,
 219                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 220                 },
 221                 .psc_thr                = {
 222                         .tc_thr_name            = LUSTRE_MDT_NAME "_rdpg",
 223                         .tc_thr_factor          = MDS_RDPG_THR_FACTOR,
 224                         .tc_nthrs_init          = MDS_RDPG_NTHRS_INIT,
 225                         .tc_nthrs_base          = MDS_RDPG_NTHRS_BASE,
 226                         .tc_nthrs_max           = MDS_RDPG_NTHRS_MAX,
 227                         .tc_nthrs_user          = mds_rdpg_num_threads,
 228                         .tc_cpu_affinity        = 1,
 229                         .tc_ctx_tags            = LCT_MD_THREAD,
 230                 },
 231                 .psc_cpt                = {
 232                         .cc_pattern             = mds_rdpg_num_cpts,
 233                 },
 234                 .psc_ops                = {
 235                         .so_req_handler         = tgt_request_handle,
 236                         .so_req_printer         = target_print_req,
 237                 },
 238         };
 239         m->mds_readpage_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 240                                                           procfs_entry);
 241         if (IS_ERR(m->mds_readpage_service)) {
 242                 rc = PTR_ERR(m->mds_readpage_service);
 243                 CERROR("failed to start readpage service: %d\n", rc);
 244                 m->mds_readpage_service = NULL;
 245
 246                 GOTO(err_mds_svc, rc);
 247         }
 248
 249         /*
 250          * setattr service configuration.
 251          *
 252          * XXX To keep the compatibility with old client(< 2.2), we need to
 253          * preserve this portal for a certain time, it should be removed
 254          * eventually. LU-617.
 255          */
 256         memset(&conf, 0, sizeof(conf));
 257         conf = (typeof(conf)) {
 258                 .psc_name               = LUSTRE_MDT_NAME "_setattr",
 259                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 260                 .psc_buf                = {
 261                         .bc_nbufs               = MDS_NBUFS,
 262                         .bc_buf_size            = MDS_BUFSIZE,
 263                         .bc_req_max_size        = MDS_MAXREQSIZE,
 264                         .bc_rep_max_size        = MDS_LOV_MAXREPSIZE,
 265                         .bc_req_portal          = MDS_SETATTR_PORTAL,
 266                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 267                 },
 268                 .psc_thr                = {
 269                         .tc_thr_name            = LUSTRE_MDT_NAME "_attr",
 270                         .tc_thr_factor          = MDS_SETA_THR_FACTOR,
 271                         .tc_nthrs_init          = MDS_SETA_NTHRS_INIT,
 272                         .tc_nthrs_base          = MDS_SETA_NTHRS_BASE,
 273                         .tc_nthrs_max           = MDS_SETA_NTHRS_MAX,
 274                         .tc_nthrs_user          = mds_attr_num_threads,
 275                         .tc_cpu_affinity        = 1,
 276                         .tc_ctx_tags            = LCT_MD_THREAD,
 277                 },
 278                 .psc_cpt                = {
 279                         .cc_pattern             = mds_attr_num_cpts,
 280                 },
 281                 .psc_ops                = {
 282                         .so_req_handler         = tgt_request_handle,
 283                         .so_req_printer         = target_print_req,
 284                         .so_hpreq_handler       = NULL,
 285                 },
 286         };
 287         m->mds_setattr_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 288                                                          procfs_entry);
 289         if (IS_ERR(m->mds_setattr_service)) {
 290                 rc = PTR_ERR(m->mds_setattr_service);
 291                 CERROR("failed to start setattr service: %d\n", rc);
 292                 m->mds_setattr_service = NULL;
 293
 294                 GOTO(err_mds_svc, rc);
 295         }
 296
 297         /* Object update service */
 298         conf = (typeof(conf)) {
 299                 .psc_name               = LUSTRE_MDT_NAME "_out",
 300                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 301                 .psc_buf                = {
 302                         .bc_nbufs               = MDS_NBUFS,
 303                         .bc_buf_size            = OUT_BUFSIZE,
 304                         .bc_req_max_size        = OUT_MAXREQSIZE,
 305                         .bc_rep_max_size        = OUT_MAXREPSIZE,
 306                         .bc_req_portal          = OUT_PORTAL,
 307                         .bc_rep_portal          = OSC_REPLY_PORTAL,
 308                 },
 309                 /*
 310                  * We'd like to have a mechanism to set this on a per-device
 311                  * basis, but alas...
 312                  */
 313                 .psc_thr                = {
 314                         .tc_thr_name            = LUSTRE_MDT_NAME "_out",
 315                         .tc_thr_factor          = MDS_THR_FACTOR,
 316                         .tc_nthrs_init          = MDS_NTHRS_INIT,
 317                         .tc_nthrs_base          = MDS_NTHRS_BASE,
 318                         .tc_nthrs_max           = MDS_NTHRS_MAX,
 319                         .tc_nthrs_user          = mds_num_threads,
 320                         .tc_cpu_affinity        = 1,
 321                         .tc_ctx_tags            = LCT_MD_THREAD |
 322                                                   LCT_DT_THREAD,
 323                 },
 324                 .psc_cpt                = {
 325                         .cc_pattern             = mds_num_cpts,
 326                 },
 327                 .psc_ops                = {
 328                         .so_req_handler         = tgt_request_handle,
 329                         .so_req_printer         = target_print_req,
 330                         .so_hpreq_handler       = NULL,
 331                 },
 332         };
 333         m->mds_out_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 334                                                      procfs_entry);
 335         if (IS_ERR(m->mds_out_service)) {
 336                 rc = PTR_ERR(m->mds_out_service);
 337                 CERROR("failed to start out service: %d\n", rc);
 338                 m->mds_out_service = NULL;
 339                 GOTO(err_mds_svc, rc);
 340         }
 341
 342         /*
 343          * sequence controller service configuration
 344          */
 345         memset(&conf, 0, sizeof(conf));
 346         conf = (typeof(conf)) {
 347                 .psc_name               = LUSTRE_MDT_NAME "_seqs",
 348                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 349                 .psc_buf                = {
 350                         .bc_nbufs               = MDS_NBUFS,
 351                         .bc_buf_size            = SEQ_BUFSIZE,
 352                         .bc_req_max_size        = SEQ_MAXREQSIZE,
 353                         .bc_rep_max_size        = SEQ_MAXREPSIZE,
 354                         .bc_req_portal          = SEQ_CONTROLLER_PORTAL,
 355                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 356                 },
 357                 .psc_thr                = {
 358                         .tc_thr_name            = LUSTRE_MDT_NAME "_seqs",
 359                         .tc_nthrs_init          = MDS_OTHR_NTHRS_INIT,
 360                         .tc_nthrs_max           = MDS_OTHR_NTHRS_MAX,
 361                         .tc_ctx_tags            = LCT_MD_THREAD,
 362                 },
 363                 .psc_ops                = {
 364                         .so_req_handler         = tgt_request_handle,
 365                         .so_req_printer         = target_print_req,
 366                         .so_hpreq_handler       = NULL,
 367                 },
 368         };
 369         m->mds_mdsc_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 370                                                       procfs_entry);
 371         if (IS_ERR(m->mds_mdsc_service)) {
 372                 rc = PTR_ERR(m->mds_mdsc_service);
 373                 CERROR("failed to start seq controller service: %d\n", rc);
 374                 m->mds_mdsc_service = NULL;
 375
 376                 GOTO(err_mds_svc, rc);
 377         }
 378
 379         /*
 380          * metadata sequence server service configuration
 381          */
 382         memset(&conf, 0, sizeof(conf));
 383         conf = (typeof(conf)) {
 384                 .psc_name               = LUSTRE_MDT_NAME "_seqm",
 385                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 386                 .psc_buf                = {
 387                         .bc_nbufs               = MDS_NBUFS,
 388                         .bc_buf_size            = SEQ_BUFSIZE,
 389                         .bc_req_max_size        = SEQ_MAXREQSIZE,
 390                         .bc_rep_max_size        = SEQ_MAXREPSIZE,
 391                         .bc_req_portal          = SEQ_METADATA_PORTAL,
 392                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 393                 },
 394                 .psc_thr                = {
 395                         .tc_thr_name            = LUSTRE_MDT_NAME "_seqm",
 396                         .tc_nthrs_init          = MDS_OTHR_NTHRS_INIT,
 397                         .tc_nthrs_max           = MDS_OTHR_NTHRS_MAX,
 398                         .tc_ctx_tags            = LCT_MD_THREAD | LCT_DT_THREAD
 399                 },
 400                 .psc_ops                = {
 401                         .so_req_handler         = tgt_request_handle,
 402                         .so_req_printer         = target_print_req,
 403                         .so_hpreq_handler       = NULL,
 404                 },
 405         };
 406         m->mds_mdss_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 407                                                       procfs_entry);
 408         if (IS_ERR(m->mds_mdss_service)) {
 409                 rc = PTR_ERR(m->mds_mdss_service);
 410                 CERROR("failed to start metadata seq server service: %d\n", rc);
 411                 m->mds_mdss_service = NULL;
 412
 413                 GOTO(err_mds_svc, rc);
 414         }
 415
 416         /* FLD service start */
 417         memset(&conf, 0, sizeof(conf));
 418         conf = (typeof(conf)) {
 419                 .psc_name            = LUSTRE_MDT_NAME "_fld",
 420                 .psc_watchdog_factor = MDT_SERVICE_WATCHDOG_FACTOR,
 421                 .psc_buf                = {
 422                         .bc_nbufs               = MDS_NBUFS,
 423                         .bc_buf_size            = FLD_BUFSIZE,
 424                         .bc_req_max_size        = FLD_MAXREQSIZE,
 425                         .bc_rep_max_size        = FLD_MAXREPSIZE,
 426                         .bc_req_portal          = FLD_REQUEST_PORTAL,
 427                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 428                 },
 429                 .psc_thr                = {
 430                         .tc_thr_name            = LUSTRE_MDT_NAME "_fld",
 431                         .tc_nthrs_init          = MDS_OTHR_NTHRS_INIT,
 432                         .tc_nthrs_max           = MDS_OTHR_NTHRS_MAX,
 433                         .tc_ctx_tags            = LCT_DT_THREAD | LCT_MD_THREAD,
 434                 },
 435                 .psc_ops                = {
 436                         .so_req_handler         = tgt_request_handle,
 437                         .so_req_printer         = target_print_req,
 438                         .so_hpreq_handler       = NULL,
 439                 },
 440         };
 441         m->mds_fld_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 442                                                      procfs_entry);
 443         if (IS_ERR(m->mds_fld_service)) {
 444                 rc = PTR_ERR(m->mds_fld_service);
 445                 CERROR("failed to start fld service: %d\n", rc);
 446                 m->mds_fld_service = NULL;
 447
 448                 GOTO(err_mds_svc, rc);
 449         }
 450
 451         memset(&conf, 0, sizeof(conf));
 452         conf = (typeof(conf)) {
 453                 .psc_name               = LUSTRE_MDT_NAME "_io",
 454                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 455                 .psc_buf                = {
 456                         .bc_nbufs               = OST_NBUFS,
 457                         .bc_buf_size            = OST_IO_BUFSIZE,
 458                         .bc_req_max_size        = OST_IO_MAXREQSIZE,
 459                         .bc_rep_max_size        = OST_IO_MAXREPSIZE,
 460                         .bc_req_portal          = MDS_IO_PORTAL,
 461                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 462                 },
 463                 .psc_thr                = {
 464                         .tc_thr_name            = "ll_mdt_io",
 465                         .tc_thr_factor          = OSS_THR_FACTOR,
 466                         .tc_nthrs_init          = OSS_NTHRS_INIT,
 467                         .tc_nthrs_base          = OSS_NTHRS_BASE,
 468                         .tc_nthrs_max           = mds_max_io_threads,
 469                         .tc_cpu_affinity        = 1,
 470                         .tc_ctx_tags            = LCT_DT_THREAD | LCT_MD_THREAD,
 471                 },
 472                 .psc_ops                = {
 473                         .so_thr_init            = tgt_io_thread_init,
 474                         .so_thr_done            = tgt_io_thread_done,
 475                         .so_req_handler         = tgt_request_handle,
 476                         .so_req_printer         = target_print_req,
 477                 },
 478         };
 479         m->mds_io_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 480                                                     procfs_entry);
 481         if (IS_ERR(m->mds_io_service)) {
 482                 rc = PTR_ERR(m->mds_io_service);
 483                 CERROR("failed to start MDT I/O service: %d\n", rc);
 484                 m->mds_io_service = NULL;
 485                 GOTO(err_mds_svc, rc);
 486         }
 487
 488         EXIT;
 489 err_mds_svc:
 490         if (rc)
 491                 mds_stop_ptlrpc_service(m);
 492
 493         return rc;
 494 }
 495
 496 static inline struct mds_device *mds_dev(struct lu_device *d)
 497 {
 498         return container_of0(d, struct mds_device, mds_md_dev.md_lu_dev);
 499 }
 500
 501 static struct lu_device *mds_device_fini(const struct lu_env *env,
 502                                          struct lu_device *d)
 503 {
 504         struct mds_device *m = mds_dev(d);
 505         struct obd_device *obd = d->ld_obd;
 506         ENTRY;
 507
 508         mds_stop_ptlrpc_service(m);
 509         lprocfs_obd_cleanup(obd);
 510         RETURN(NULL);
 511 }
 512
 513 static struct lu_device *mds_device_free(const struct lu_env *env,
 514                                          struct lu_device *d)
 515 {
 516         struct mds_device *m = mds_dev(d);
 517         ENTRY;
 518
 519         md_device_fini(&m->mds_md_dev);
 520         OBD_FREE_PTR(m);
 521         RETURN(NULL);
 522 }
 523
 524 static struct lu_device *mds_device_alloc(const struct lu_env *env,
 525                                           struct lu_device_type *t,
 526                                           struct lustre_cfg *cfg)
 527 {
 528         struct mds_device        *m;
 529         struct obd_device        *obd;
 530         struct lu_device          *l;
 531         int rc;
 532
 533         OBD_ALLOC_PTR(m);
 534         if (m == NULL)
 535                 return ERR_PTR(-ENOMEM);
 536
 537         md_device_init(&m->mds_md_dev, t);
 538         l = &m->mds_md_dev.md_lu_dev;
 539
 540         obd = class_name2obd(lustre_cfg_string(cfg, 0));
 541         LASSERT(obd != NULL);
 542
 543         l->ld_obd = obd;
 544         /* set this lu_device to obd, because error handling need it */
 545         obd->obd_lu_dev = l;
 546
 547         rc = lprocfs_obd_setup(obd, true);
 548         if (rc != 0) {
 549                 mds_device_free(env, l);
 550                 l = ERR_PTR(rc);
 551                 return l;
 552         }
 553
 554         mutex_init(&m->mds_health_mutex);
 555
 556         rc = mds_start_ptlrpc_service(m);
 557         if (rc != 0) {
 558                 lprocfs_obd_cleanup(obd);
 559                 mds_device_free(env, l);
 560                 l = ERR_PTR(rc);
 561                 return l;
 562         }
 563         return l;
 564 }
 565
 566 /* type constructor/destructor: mdt_type_init, mdt_type_fini */
 567 LU_TYPE_INIT_FINI(mds, &mdt_thread_key);
 568
 569 static struct lu_device_type_operations mds_device_type_ops = {
 570         .ldto_init = mds_type_init,
 571         .ldto_fini = mds_type_fini,
 572
 573         .ldto_start = mds_type_start,
 574         .ldto_stop  = mds_type_stop,
 575
 576         .ldto_device_alloc = mds_device_alloc,
 577         .ldto_device_free  = mds_device_free,
 578         .ldto_device_fini  = mds_device_fini
 579 };
 580
 581 static struct lu_device_type mds_device_type = {
 582         .ldt_tags     = LU_DEVICE_MD,
 583         .ldt_name     = LUSTRE_MDS_NAME,
 584         .ldt_ops      = &mds_device_type_ops,
 585         .ldt_ctx_tags = LCT_MD_THREAD
 586 };
 587
 588 static int mds_health_check(const struct lu_env *env, struct obd_device *obd)
 589 {
 590         struct mds_device *mds = mds_dev(obd->obd_lu_dev);
 591         int rc = 0;
 592
 593
 594         mutex_lock(&mds->mds_health_mutex);
 595         rc |= ptlrpc_service_health_check(mds->mds_regular_service);
 596         rc |= ptlrpc_service_health_check(mds->mds_readpage_service);
 597         rc |= ptlrpc_service_health_check(mds->mds_out_service);
 598         rc |= ptlrpc_service_health_check(mds->mds_setattr_service);
 599         rc |= ptlrpc_service_health_check(mds->mds_mdsc_service);
 600         rc |= ptlrpc_service_health_check(mds->mds_mdss_service);
 601         rc |= ptlrpc_service_health_check(mds->mds_fld_service);
 602         rc |= ptlrpc_service_health_check(mds->mds_io_service);
 603         mutex_unlock(&mds->mds_health_mutex);
 604
 605         return rc != 0 ? 1 : 0;
 606 }
 607
 608 static struct obd_ops mds_obd_device_ops = {
 609         .o_owner           = THIS_MODULE,
 610         .o_health_check    = mds_health_check,
 611 };
 612
 613 int mds_mod_init(void)
 614 {
 615         return class_register_type(&mds_obd_device_ops, NULL, true, NULL,
 616                                    LUSTRE_MDS_NAME, &mds_device_type);
 617 }
 618
 619 void mds_mod_exit(void)
 620 {
 621         class_unregister_type(LUSTRE_MDS_NAME);
 622 }