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 static unsigned int mds_cpu_bind = 1;
  79 module_param(mds_cpu_bind, uint, 0444);
  80 MODULE_PARM_DESC(mds_cpu_bind,
  81                  "bind MDS threads to particular CPU partitions");
  82
  83 int mds_max_io_threads = 512;
  84 module_param(mds_max_io_threads, int, 0444);
  85 MODULE_PARM_DESC(mds_max_io_threads,
  86                  "maximum number of MDS IO service threads");
  87
  88 static unsigned int mds_io_cpu_bind = 1;
  89 module_param(mds_io_cpu_bind, uint, 0444);
  90 MODULE_PARM_DESC(mds_io_cpu_bind,
  91                  "bind MDS IO threads to particular CPU partitions");
  92
  93 static char *mds_io_num_cpts;
  94 module_param(mds_io_num_cpts, charp, 0444);
  95 MODULE_PARM_DESC(mds_io_num_cpts,
  96                  "CPU partitions MDS IO threads should run on");
  97
  98 static struct cfs_cpt_table *mdt_io_cptable;
  99
 100 static char *mds_num_cpts;
 101 module_param(mds_num_cpts, charp, 0444);
 102 MODULE_PARM_DESC(mds_num_cpts, "CPU partitions MDS threads should run on");
 103
 104 static unsigned long mds_rdpg_num_threads;
 105 module_param(mds_rdpg_num_threads, ulong, 0444);
 106 MODULE_PARM_DESC(mds_rdpg_num_threads,
 107                  "number of MDS readpage service threads to start");
 108
 109 static unsigned int mds_rdpg_cpu_bind = 1;
 110 module_param(mds_rdpg_cpu_bind, uint, 0444);
 111 MODULE_PARM_DESC(mds_rdpg_cpu_bind,
 112                  "bind MDS readpage threads to particular CPU partitions");
 113
 114 static char *mds_rdpg_num_cpts;
 115 module_param(mds_rdpg_num_cpts, charp, 0444);
 116 MODULE_PARM_DESC(mds_rdpg_num_cpts,
 117                  "CPU partitions MDS readpage threads should run on");
 118
 119 /* NB: these two should be removed along with setattr service in the future */
 120 static unsigned long mds_attr_num_threads;
 121 module_param(mds_attr_num_threads, ulong, 0444);
 122 MODULE_PARM_DESC(mds_attr_num_threads,
 123                  "number of MDS setattr service threads to start");
 124
 125 static unsigned int mds_attr_cpu_bind = 1;
 126 module_param(mds_attr_cpu_bind, uint, 0444);
 127 MODULE_PARM_DESC(mds_attr_cpu_bind,
 128                  "bind MDS setattr threads to particular CPU partitions");
 129
 130 static char *mds_attr_num_cpts;
 131 module_param(mds_attr_num_cpts, charp, 0444);
 132 MODULE_PARM_DESC(mds_attr_num_cpts,
 133                  "CPU partitions MDS setattr threads should run on");
 134
 135 /* device init/fini methods */
 136 static void mds_stop_ptlrpc_service(struct mds_device *m)
 137 {
 138         ENTRY;
 139
 140         mutex_lock(&m->mds_health_mutex);
 141         if (m->mds_regular_service != NULL) {
 142                 ptlrpc_unregister_service(m->mds_regular_service);
 143                 m->mds_regular_service = NULL;
 144         }
 145         if (m->mds_readpage_service != NULL) {
 146                 ptlrpc_unregister_service(m->mds_readpage_service);
 147                 m->mds_readpage_service = NULL;
 148         }
 149         if (m->mds_out_service != NULL) {
 150                 ptlrpc_unregister_service(m->mds_out_service);
 151                 m->mds_out_service = NULL;
 152         }
 153         if (m->mds_setattr_service != NULL) {
 154                 ptlrpc_unregister_service(m->mds_setattr_service);
 155                 m->mds_setattr_service = NULL;
 156         }
 157         if (m->mds_mdsc_service != NULL) {
 158                 ptlrpc_unregister_service(m->mds_mdsc_service);
 159                 m->mds_mdsc_service = NULL;
 160         }
 161         if (m->mds_mdss_service != NULL) {
 162                 ptlrpc_unregister_service(m->mds_mdss_service);
 163                 m->mds_mdss_service = NULL;
 164         }
 165         if (m->mds_fld_service != NULL) {
 166                 ptlrpc_unregister_service(m->mds_fld_service);
 167                 m->mds_fld_service = NULL;
 168         }
 169         if (m->mds_io_service != NULL) {
 170                 ptlrpc_unregister_service(m->mds_io_service);
 171                 m->mds_io_service = NULL;
 172         }
 173         mutex_unlock(&m->mds_health_mutex);
 174
 175         if (mdt_io_cptable != NULL) {
 176                 cfs_cpt_table_free(mdt_io_cptable);
 177                 mdt_io_cptable = NULL;
 178         }
 179
 180         EXIT;
 181 }
 182
 183 static int mds_start_ptlrpc_service(struct mds_device *m)
 184 {
 185         static struct ptlrpc_service_conf conf;
 186         struct obd_device *obd = m->mds_md_dev.md_lu_dev.ld_obd;
 187         nodemask_t *mask;
 188         int rc = 0;
 189
 190         ENTRY;
 191
 192         conf = (typeof(conf)) {
 193                 .psc_name               = LUSTRE_MDT_NAME,
 194                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 195                 .psc_buf                = {
 196                         .bc_nbufs               = MDS_NBUFS,
 197                         .bc_buf_size            = MDS_REG_BUFSIZE,
 198                         .bc_req_max_size        = MDS_REG_MAXREQSIZE,
 199                         .bc_rep_max_size        = MDS_REG_MAXREPSIZE,
 200                         .bc_req_portal          = MDS_REQUEST_PORTAL,
 201                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 202                 },
 203                 /*
 204                  * We'd like to have a mechanism to set this on a per-device
 205                  * basis, but alas...
 206                  */
 207                 .psc_thr                = {
 208                         .tc_thr_name            = LUSTRE_MDT_NAME,
 209                         .tc_thr_factor          = MDS_THR_FACTOR,
 210                         .tc_nthrs_init          = MDS_NTHRS_INIT,
 211                         .tc_nthrs_base          = MDS_NTHRS_BASE,
 212                         .tc_nthrs_max           = MDS_NTHRS_MAX,
 213                         .tc_nthrs_user          = mds_num_threads,
 214                         .tc_cpu_bind            = mds_cpu_bind,
 215                         .tc_ctx_tags            = LCT_MD_THREAD,
 216                 },
 217                 .psc_cpt                = {
 218                         .cc_pattern             = mds_num_cpts,
 219                         .cc_affinity            = true,
 220                 },
 221                 .psc_ops                = {
 222                         .so_req_handler         = tgt_request_handle,
 223                         .so_req_printer         = target_print_req,
 224                         .so_hpreq_handler       = ptlrpc_hpreq_handler,
 225                 },
 226         };
 227         m->mds_regular_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 228                                                          obd->obd_debugfs_entry);
 229         if (IS_ERR(m->mds_regular_service)) {
 230                 rc = PTR_ERR(m->mds_regular_service);
 231                 CERROR("failed to start regular mdt service: %d\n", rc);
 232                 m->mds_regular_service = NULL;
 233
 234                 RETURN(rc);
 235         }
 236
 237         /*
 238          * readpage service configuration. Parameters have to be adjusted,
 239          * ideally.
 240          */
 241         memset(&conf, 0, sizeof(conf));
 242         conf = (typeof(conf)) {
 243                 .psc_name               = LUSTRE_MDT_NAME "_readpage",
 244                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 245                 .psc_buf                = {
 246                         .bc_nbufs               = MDS_NBUFS,
 247                         .bc_buf_size            = MDS_BUFSIZE,
 248                         .bc_req_max_size        = MDS_MAXREQSIZE,
 249                         .bc_rep_max_size        = MDS_MAXREPSIZE,
 250                         .bc_req_portal          = MDS_READPAGE_PORTAL,
 251                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 252                 },
 253                 .psc_thr                = {
 254                         .tc_thr_name            = LUSTRE_MDT_NAME "_rdpg",
 255                         .tc_thr_factor          = MDS_RDPG_THR_FACTOR,
 256                         .tc_nthrs_init          = MDS_RDPG_NTHRS_INIT,
 257                         .tc_nthrs_base          = MDS_RDPG_NTHRS_BASE,
 258                         .tc_nthrs_max           = MDS_RDPG_NTHRS_MAX,
 259                         .tc_nthrs_user          = mds_rdpg_num_threads,
 260                         .tc_cpu_bind            = mds_rdpg_cpu_bind,
 261                         .tc_ctx_tags            = LCT_MD_THREAD,
 262                 },
 263                 .psc_cpt                = {
 264                         .cc_pattern             = mds_rdpg_num_cpts,
 265                         .cc_affinity            = true,
 266                 },
 267                 .psc_ops                = {
 268                         .so_req_handler         = tgt_request_handle,
 269                         .so_req_printer         = target_print_req,
 270                 },
 271         };
 272         m->mds_readpage_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 273                                                           obd->obd_debugfs_entry);
 274         if (IS_ERR(m->mds_readpage_service)) {
 275                 rc = PTR_ERR(m->mds_readpage_service);
 276                 CERROR("failed to start readpage service: %d\n", rc);
 277                 m->mds_readpage_service = NULL;
 278
 279                 GOTO(err_mds_svc, rc);
 280         }
 281
 282         /*
 283          * setattr service configuration.
 284          *
 285          * XXX To keep the compatibility with old client(< 2.2), we need to
 286          * preserve this portal for a certain time, it should be removed
 287          * eventually. LU-617.
 288          */
 289         memset(&conf, 0, sizeof(conf));
 290         conf = (typeof(conf)) {
 291                 .psc_name               = LUSTRE_MDT_NAME "_setattr",
 292                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 293                 .psc_buf                = {
 294                         .bc_nbufs               = MDS_NBUFS,
 295                         .bc_buf_size            = MDS_BUFSIZE,
 296                         .bc_req_max_size        = MDS_MAXREQSIZE,
 297                         .bc_rep_max_size        = MDS_LOV_MAXREPSIZE,
 298                         .bc_req_portal          = MDS_SETATTR_PORTAL,
 299                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 300                 },
 301                 .psc_thr                = {
 302                         .tc_thr_name            = LUSTRE_MDT_NAME "_attr",
 303                         .tc_thr_factor          = MDS_SETA_THR_FACTOR,
 304                         .tc_nthrs_init          = MDS_SETA_NTHRS_INIT,
 305                         .tc_nthrs_base          = MDS_SETA_NTHRS_BASE,
 306                         .tc_nthrs_max           = MDS_SETA_NTHRS_MAX,
 307                         .tc_nthrs_user          = mds_attr_num_threads,
 308                         .tc_cpu_bind            = mds_attr_cpu_bind,
 309                         .tc_ctx_tags            = LCT_MD_THREAD,
 310                 },
 311                 .psc_cpt                = {
 312                         .cc_pattern             = mds_attr_num_cpts,
 313                         .cc_affinity            = true,
 314                 },
 315                 .psc_ops                = {
 316                         .so_req_handler         = tgt_request_handle,
 317                         .so_req_printer         = target_print_req,
 318                         .so_hpreq_handler       = NULL,
 319                 },
 320         };
 321         m->mds_setattr_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 322                                                          obd->obd_debugfs_entry);
 323         if (IS_ERR(m->mds_setattr_service)) {
 324                 rc = PTR_ERR(m->mds_setattr_service);
 325                 CERROR("failed to start setattr service: %d\n", rc);
 326                 m->mds_setattr_service = NULL;
 327
 328                 GOTO(err_mds_svc, rc);
 329         }
 330
 331         /* Object update service */
 332         conf = (typeof(conf)) {
 333                 .psc_name               = LUSTRE_MDT_NAME "_out",
 334                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 335                 .psc_buf                = {
 336                         .bc_nbufs               = MDS_NBUFS,
 337                         .bc_buf_size            = OUT_BUFSIZE,
 338                         .bc_req_max_size        = OUT_MAXREQSIZE,
 339                         .bc_rep_max_size        = OUT_MAXREPSIZE,
 340                         .bc_req_portal          = OUT_PORTAL,
 341                         .bc_rep_portal          = OSC_REPLY_PORTAL,
 342                 },
 343                 /*
 344                  * We'd like to have a mechanism to set this on a per-device
 345                  * basis, but alas...
 346                  */
 347                 .psc_thr                = {
 348                         .tc_thr_name            = LUSTRE_MDT_NAME "_out",
 349                         .tc_thr_factor          = MDS_THR_FACTOR,
 350                         .tc_nthrs_init          = MDS_NTHRS_INIT,
 351                         .tc_nthrs_base          = MDS_NTHRS_BASE,
 352                         .tc_nthrs_max           = MDS_NTHRS_MAX,
 353                         .tc_nthrs_user          = mds_num_threads,
 354                         .tc_cpu_bind            = mds_cpu_bind,
 355                         .tc_ctx_tags            = LCT_MD_THREAD |
 356                                                   LCT_DT_THREAD,
 357                 },
 358                 .psc_cpt                = {
 359                         .cc_pattern             = mds_num_cpts,
 360                         .cc_affinity            = true,
 361                 },
 362                 .psc_ops                = {
 363                         .so_req_handler         = tgt_request_handle,
 364                         .so_req_printer         = target_print_req,
 365                         .so_hpreq_handler       = NULL,
 366                 },
 367         };
 368         m->mds_out_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 369                                                      obd->obd_debugfs_entry);
 370         if (IS_ERR(m->mds_out_service)) {
 371                 rc = PTR_ERR(m->mds_out_service);
 372                 CERROR("failed to start out service: %d\n", rc);
 373                 m->mds_out_service = NULL;
 374                 GOTO(err_mds_svc, rc);
 375         }
 376
 377         /*
 378          * sequence controller service configuration
 379          */
 380         memset(&conf, 0, sizeof(conf));
 381         conf = (typeof(conf)) {
 382                 .psc_name               = LUSTRE_MDT_NAME "_seqs",
 383                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 384                 .psc_buf                = {
 385                         .bc_nbufs               = MDS_NBUFS,
 386                         .bc_buf_size            = SEQ_BUFSIZE,
 387                         .bc_req_max_size        = SEQ_MAXREQSIZE,
 388                         .bc_rep_max_size        = SEQ_MAXREPSIZE,
 389                         .bc_req_portal          = SEQ_CONTROLLER_PORTAL,
 390                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 391                 },
 392                 .psc_thr                = {
 393                         .tc_thr_name            = LUSTRE_MDT_NAME "_seqs",
 394                         .tc_nthrs_init          = MDS_OTHR_NTHRS_INIT,
 395                         .tc_nthrs_max           = MDS_OTHR_NTHRS_MAX,
 396                         .tc_ctx_tags            = LCT_MD_THREAD,
 397                 },
 398                 .psc_ops                = {
 399                         .so_req_handler         = tgt_request_handle,
 400                         .so_req_printer         = target_print_req,
 401                         .so_hpreq_handler       = NULL,
 402                 },
 403         };
 404         m->mds_mdsc_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 405                                                       obd->obd_debugfs_entry);
 406         if (IS_ERR(m->mds_mdsc_service)) {
 407                 rc = PTR_ERR(m->mds_mdsc_service);
 408                 CERROR("failed to start seq controller service: %d\n", rc);
 409                 m->mds_mdsc_service = NULL;
 410
 411                 GOTO(err_mds_svc, rc);
 412         }
 413
 414         /*
 415          * metadata sequence server service configuration
 416          */
 417         memset(&conf, 0, sizeof(conf));
 418         conf = (typeof(conf)) {
 419                 .psc_name               = LUSTRE_MDT_NAME "_seqm",
 420                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 421                 .psc_buf                = {
 422                         .bc_nbufs               = MDS_NBUFS,
 423                         .bc_buf_size            = SEQ_BUFSIZE,
 424                         .bc_req_max_size        = SEQ_MAXREQSIZE,
 425                         .bc_rep_max_size        = SEQ_MAXREPSIZE,
 426                         .bc_req_portal          = SEQ_METADATA_PORTAL,
 427                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 428                 },
 429                 .psc_thr                = {
 430                         .tc_thr_name            = LUSTRE_MDT_NAME "_seqm",
 431                         .tc_nthrs_init          = MDS_OTHR_NTHRS_INIT,
 432                         .tc_nthrs_max           = MDS_OTHR_NTHRS_MAX,
 433                         .tc_ctx_tags            = LCT_MD_THREAD | LCT_DT_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_mdss_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 442                                                       obd->obd_debugfs_entry);
 443         if (IS_ERR(m->mds_mdss_service)) {
 444                 rc = PTR_ERR(m->mds_mdss_service);
 445                 CERROR("failed to start metadata seq server service: %d\n", rc);
 446                 m->mds_mdss_service = NULL;
 447
 448                 GOTO(err_mds_svc, rc);
 449         }
 450
 451         /* FLD service start */
 452         memset(&conf, 0, sizeof(conf));
 453         conf = (typeof(conf)) {
 454                 .psc_name            = LUSTRE_MDT_NAME "_fld",
 455                 .psc_watchdog_factor = MDT_SERVICE_WATCHDOG_FACTOR,
 456                 .psc_buf                = {
 457                         .bc_nbufs               = MDS_NBUFS,
 458                         .bc_buf_size            = FLD_BUFSIZE,
 459                         .bc_req_max_size        = FLD_MAXREQSIZE,
 460                         .bc_rep_max_size        = FLD_MAXREPSIZE,
 461                         .bc_req_portal          = FLD_REQUEST_PORTAL,
 462                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 463                 },
 464                 .psc_thr                = {
 465                         .tc_thr_name            = LUSTRE_MDT_NAME "_fld",
 466                         .tc_nthrs_init          = MDS_OTHR_NTHRS_INIT,
 467                         .tc_nthrs_max           = MDS_OTHR_NTHRS_MAX,
 468                         .tc_ctx_tags            = LCT_DT_THREAD | LCT_MD_THREAD,
 469                 },
 470                 .psc_ops                = {
 471                         .so_req_handler         = tgt_request_handle,
 472                         .so_req_printer         = target_print_req,
 473                         .so_hpreq_handler       = NULL,
 474                 },
 475         };
 476         m->mds_fld_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 477                                                      obd->obd_debugfs_entry);
 478         if (IS_ERR(m->mds_fld_service)) {
 479                 rc = PTR_ERR(m->mds_fld_service);
 480                 CERROR("failed to start fld service: %d\n", rc);
 481                 m->mds_fld_service = NULL;
 482
 483                 GOTO(err_mds_svc, rc);
 484         }
 485
 486
 487         mask = cfs_cpt_nodemask(cfs_cpt_table, CFS_CPT_ANY);
 488         /* event CPT feature is disabled in libcfs level by set partition
 489          * number to 1, we still want to set node affinity for io service */
 490         if (cfs_cpt_number(cfs_cpt_table) == 1 && nodes_weight(*mask) > 1) {
 491                 int cpt = 0;
 492                 int i;
 493
 494                 mdt_io_cptable = cfs_cpt_table_alloc(nodes_weight(*mask));
 495                 for_each_node_mask(i, *mask) {
 496                         if (mdt_io_cptable == NULL) {
 497                                 CWARN("MDS failed to create CPT table\n");
 498                                 break;
 499                         }
 500
 501                         rc = cfs_cpt_set_node(mdt_io_cptable, cpt++, i);
 502                         if (!rc) {
 503                                 CWARN("MDS Failed to set node %d for"
 504                                       "IO CPT table\n", i);
 505                                 cfs_cpt_table_free(mdt_io_cptable);
 506                                 mdt_io_cptable = NULL;
 507                                 break;
 508                         }
 509                 }
 510         }
 511
 512         memset(&conf, 0, sizeof(conf));
 513         conf = (typeof(conf)) {
 514                 .psc_name               = LUSTRE_MDT_NAME "_io",
 515                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 516                 .psc_buf                = {
 517                         .bc_nbufs               = OST_NBUFS,
 518                         .bc_buf_size            = OST_IO_BUFSIZE,
 519                         .bc_req_max_size        = OST_IO_MAXREQSIZE,
 520                         .bc_rep_max_size        = OST_IO_MAXREPSIZE,
 521                         .bc_req_portal          = MDS_IO_PORTAL,
 522                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 523                 },
 524                 .psc_thr                = {
 525                         .tc_thr_name            = LUSTRE_MDT_NAME "_io",
 526                         .tc_thr_factor          = OSS_THR_FACTOR,
 527                         .tc_nthrs_init          = OSS_NTHRS_INIT,
 528                         .tc_nthrs_base          = OSS_NTHRS_BASE,
 529                         .tc_nthrs_max           = mds_max_io_threads,
 530                         .tc_nthrs_user          = mds_num_threads,
 531                         .tc_cpu_bind            = mds_io_cpu_bind,
 532                         .tc_ctx_tags            = LCT_DT_THREAD | LCT_MD_THREAD,
 533                 },
 534                 .psc_cpt                = {
 535                         .cc_cptable             = mdt_io_cptable,
 536                         .cc_pattern             = mdt_io_cptable == NULL ?
 537                                                   mds_io_num_cpts : NULL,
 538                         .cc_affinity            = true,
 539                 },
 540                 .psc_ops                = {
 541                         .so_thr_init            = tgt_io_thread_init,
 542                         .so_thr_done            = tgt_io_thread_done,
 543                         .so_req_handler         = tgt_request_handle,
 544                         .so_req_printer         = target_print_req,
 545                         .so_hpreq_handler       = tgt_hpreq_handler,
 546                 },
 547         };
 548         m->mds_io_service = ptlrpc_register_service(&conf, &obd->obd_kset,
 549                                                     obd->obd_debugfs_entry);
 550         if (IS_ERR(m->mds_io_service)) {
 551                 rc = PTR_ERR(m->mds_io_service);
 552                 CERROR("failed to start MDT I/O service: %d\n", rc);
 553                 m->mds_io_service = NULL;
 554                 GOTO(err_mds_svc, rc);
 555         }
 556
 557         EXIT;
 558 err_mds_svc:
 559         if (rc)
 560                 mds_stop_ptlrpc_service(m);
 561
 562         return rc;
 563 }
 564
 565 static inline struct mds_device *mds_dev(struct lu_device *d)
 566 {
 567         return container_of0(d, struct mds_device, mds_md_dev.md_lu_dev);
 568 }
 569
 570 static struct lu_device *mds_device_fini(const struct lu_env *env,
 571                                          struct lu_device *d)
 572 {
 573         struct mds_device *m = mds_dev(d);
 574         struct obd_device *obd = d->ld_obd;
 575         ENTRY;
 576
 577         mds_stop_ptlrpc_service(m);
 578         lprocfs_obd_cleanup(obd);
 579         RETURN(NULL);
 580 }
 581
 582 static struct lu_device *mds_device_free(const struct lu_env *env,
 583                                          struct lu_device *d)
 584 {
 585         struct mds_device *m = mds_dev(d);
 586         ENTRY;
 587
 588         md_device_fini(&m->mds_md_dev);
 589         OBD_FREE_PTR(m);
 590         RETURN(NULL);
 591 }
 592
 593 static struct lu_device *mds_device_alloc(const struct lu_env *env,
 594                                           struct lu_device_type *t,
 595                                           struct lustre_cfg *cfg)
 596 {
 597         struct mds_device        *m;
 598         struct obd_device        *obd;
 599         struct lu_device          *l;
 600         int rc;
 601
 602         OBD_ALLOC_PTR(m);
 603         if (m == NULL)
 604                 return ERR_PTR(-ENOMEM);
 605
 606         md_device_init(&m->mds_md_dev, t);
 607         l = &m->mds_md_dev.md_lu_dev;
 608
 609         obd = class_name2obd(lustre_cfg_string(cfg, 0));
 610         LASSERT(obd != NULL);
 611
 612         l->ld_obd = obd;
 613         /* set this lu_device to obd, because error handling need it */
 614         obd->obd_lu_dev = l;
 615
 616         rc = lprocfs_obd_setup(obd, true);
 617         if (rc != 0) {
 618                 mds_device_free(env, l);
 619                 l = ERR_PTR(rc);
 620                 return l;
 621         }
 622
 623         mutex_init(&m->mds_health_mutex);
 624
 625         rc = mds_start_ptlrpc_service(m);
 626         if (rc != 0) {
 627                 lprocfs_obd_cleanup(obd);
 628                 mds_device_free(env, l);
 629                 l = ERR_PTR(rc);
 630                 return l;
 631         }
 632         return l;
 633 }
 634
 635 /* type constructor/destructor: mdt_type_init, mdt_type_fini */
 636 LU_TYPE_INIT_FINI(mds, &mdt_thread_key);
 637
 638 static struct lu_device_type_operations mds_device_type_ops = {
 639         .ldto_init = mds_type_init,
 640         .ldto_fini = mds_type_fini,
 641
 642         .ldto_start = mds_type_start,
 643         .ldto_stop  = mds_type_stop,
 644
 645         .ldto_device_alloc = mds_device_alloc,
 646         .ldto_device_free  = mds_device_free,
 647         .ldto_device_fini  = mds_device_fini
 648 };
 649
 650 static struct lu_device_type mds_device_type = {
 651         .ldt_tags     = LU_DEVICE_MD,
 652         .ldt_name     = LUSTRE_MDS_NAME,
 653         .ldt_ops      = &mds_device_type_ops,
 654         .ldt_ctx_tags = LCT_MD_THREAD
 655 };
 656
 657 static int mds_health_check(const struct lu_env *env, struct obd_device *obd)
 658 {
 659         struct mds_device *mds = mds_dev(obd->obd_lu_dev);
 660         int rc = 0;
 661
 662
 663         mutex_lock(&mds->mds_health_mutex);
 664         rc |= ptlrpc_service_health_check(mds->mds_regular_service);
 665         rc |= ptlrpc_service_health_check(mds->mds_readpage_service);
 666         rc |= ptlrpc_service_health_check(mds->mds_out_service);
 667         rc |= ptlrpc_service_health_check(mds->mds_setattr_service);
 668         rc |= ptlrpc_service_health_check(mds->mds_mdsc_service);
 669         rc |= ptlrpc_service_health_check(mds->mds_mdss_service);
 670         rc |= ptlrpc_service_health_check(mds->mds_fld_service);
 671         rc |= ptlrpc_service_health_check(mds->mds_io_service);
 672         mutex_unlock(&mds->mds_health_mutex);
 673
 674         return rc != 0 ? 1 : 0;
 675 }
 676
 677 static struct obd_ops mds_obd_device_ops = {
 678         .o_owner           = THIS_MODULE,
 679         .o_health_check    = mds_health_check,
 680 };
 681
 682 int mds_mod_init(void)
 683 {
 684         return class_register_type(&mds_obd_device_ops, NULL, false, NULL,
 685                                    LUSTRE_MDS_NAME, &mds_device_type);
 686 }
 687
 688 void mds_mod_exit(void)
 689 {
 690         class_unregister_type(LUSTRE_MDS_NAME);
 691 }